diff options
author | 2017-06-01 13:11:16 -0400 | |
---|---|---|
committer | 2017-06-01 18:08:25 +0000 | |
commit | 9b10f8ff0d163d5d076e7028a1a173f9c1f3b714 (patch) | |
tree | beec177f1320b93b8a4d48afe929a7c510f31963 /src/jumper/SkJumper_generated_win.S | |
parent | 97b10ac4847fba563834911f35235aaf0299d0c8 (diff) |
plumb y through to SkJumper
There'll still be a little more refactoring after this, but this is the
main thing we want to do.
This makes y available in a general-purpose register in pipeline stages,
just like x. Stages that need y (seed_shader and dither) can just use
it rather than pulling it off a context pointer. seed_shader loses its
context pointer, and dither's gets simpler.
Change-Id: Ic2d1e13b03fb45b73e308b38aafbb3a14c29cf7f
Reviewed-on: https://skia-review.googlesource.com/18383
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/jumper/SkJumper_generated_win.S')
-rw-r--r-- | src/jumper/SkJumper_generated_win.S | 11154 |
1 files changed, 5617 insertions, 5537 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 0543822f9e..33d209ea38 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -18,8 +18,9 @@ _sk_start_pipeline_hsw LABEL PROC DB 65,84 ; push %r12 DB 86 ; push %rsi DB 87 ; push %rdi + DB 85 ; push %rbp DB 83 ; push %rbx - DB 72,129,236,160,0,0,0 ; sub $0xa0,%rsp + DB 72,129,236,168,0,0,0 ; sub $0xa8,%rsp DB 197,120,41,188,36,144,0,0,0 ; vmovaps %xmm15,0x90(%rsp) DB 197,120,41,180,36,128,0,0,0 ; vmovaps %xmm14,0x80(%rsp) DB 197,120,41,108,36,112 ; vmovaps %xmm13,0x70(%rsp) @@ -30,19 +31,20 @@ _sk_start_pipeline_hsw LABEL PROC DB 197,120,41,68,36,32 ; vmovaps %xmm8,0x20(%rsp) DB 197,248,41,124,36,16 ; vmovaps %xmm7,0x10(%rsp) DB 197,248,41,52,36 ; vmovaps %xmm6,(%rsp) - DB 77,137,205 ; mov %r9,%r13 - DB 77,137,198 ; mov %r8,%r14 - DB 72,137,203 ; mov %rcx,%rbx - DB 72,137,214 ; mov %rdx,%rsi - DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,199 ; mov %rax,%r15 - DB 73,137,244 ; mov %rsi,%r12 - DB 72,141,67,8 ; lea 0x8(%rbx),%rax - DB 76,57,232 ; cmp %r13,%rax - DB 118,5 ; jbe 75 <_sk_start_pipeline_hsw+0x75> - DB 72,137,223 ; mov %rbx,%rdi - DB 235,65 ; jmp b6 <_sk_start_pipeline_hsw+0xb6> - DB 185,0,0,0,0 ; mov $0x0,%ecx + DB 76,137,195 ; mov %r8,%rbx + DB 73,137,214 ; mov %rdx,%r14 + DB 72,137,205 ; mov %rcx,%rbp + DB 76,139,188,36,16,1,0,0 ; mov 0x110(%rsp),%r15 + DB 76,137,206 ; mov %r9,%rsi + DB 72,173 ; lods %ds:(%rsi),%rax + DB 73,137,196 ; mov %rax,%r12 + DB 73,137,245 ; mov %rsi,%r13 + DB 72,141,69,8 ; lea 0x8(%rbp),%rax + DB 72,57,216 ; cmp %rbx,%rax + DB 118,5 ; jbe 7e <_sk_start_pipeline_hsw+0x7e> + DB 72,137,234 ; mov %rbp,%rdx + DB 235,69 ; jmp c3 <_sk_start_pipeline_hsw+0xc3> + DB 65,184,0,0,0,0 ; mov $0x0,%r8d DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 @@ -51,17 +53,18 @@ _sk_start_pipeline_hsw LABEL PROC DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 197,204,87,246 ; vxorps %ymm6,%ymm6,%ymm6 DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 - DB 72,137,223 ; mov %rbx,%rdi - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 65,255,215 ; callq *%r15 - DB 72,141,123,8 ; lea 0x8(%rbx),%rdi - DB 72,131,195,16 ; add $0x10,%rbx - DB 76,57,235 ; cmp %r13,%rbx - DB 72,137,251 ; mov %rdi,%rbx - DB 118,191 ; jbe 75 <_sk_start_pipeline_hsw+0x75> - DB 73,41,253 ; sub %rdi,%r13 - DB 116,44 ; je e7 <_sk_start_pipeline_hsw+0xe7> + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 72,137,234 ; mov %rbp,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 65,255,212 ; callq *%r12 + DB 72,141,85,8 ; lea 0x8(%rbp),%rdx + DB 72,131,197,16 ; add $0x10,%rbp + DB 72,57,221 ; cmp %rbx,%rbp + DB 72,137,213 ; mov %rdx,%rbp + DB 118,187 ; jbe 7e <_sk_start_pipeline_hsw+0x7e> + DB 72,41,211 ; sub %rdx,%rbx + DB 116,47 ; je f7 <_sk_start_pipeline_hsw+0xf7> DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 @@ -70,10 +73,11 @@ _sk_start_pipeline_hsw LABEL PROC DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 197,204,87,246 ; vxorps %ymm6,%ymm6,%ymm6 DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,233 ; mov %r13,%rcx - DB 65,255,215 ; callq *%r15 + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 76,137,241 ; mov %r14,%rcx + DB 73,137,216 ; mov %rbx,%r8 + DB 65,255,212 ; callq *%r12 DB 197,248,40,52,36 ; vmovaps (%rsp),%xmm6 DB 197,248,40,124,36,16 ; vmovaps 0x10(%rsp),%xmm7 DB 197,120,40,68,36,32 ; vmovaps 0x20(%rsp),%xmm8 @@ -84,8 +88,9 @@ _sk_start_pipeline_hsw LABEL PROC DB 197,120,40,108,36,112 ; vmovaps 0x70(%rsp),%xmm13 DB 197,120,40,180,36,128,0,0,0 ; vmovaps 0x80(%rsp),%xmm14 DB 197,120,40,188,36,144,0,0,0 ; vmovaps 0x90(%rsp),%xmm15 - DB 72,129,196,160,0,0,0 ; add $0xa0,%rsp + DB 72,129,196,168,0,0,0 ; add $0xa8,%rsp DB 91 ; pop %rbx + DB 93 ; pop %rbp DB 95 ; pop %rdi DB 94 ; pop %rsi DB 65,92 ; pop %r12 @@ -101,18 +106,18 @@ _sk_just_return_hsw LABEL PROC PUBLIC _sk_seed_shader_hsw _sk_seed_shader_hsw LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,249,110,199 ; vmovd %edi,%xmm0 + DB 197,249,110,194 ; vmovd %edx,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,61,73,0,0 ; vbroadcastss 0x493d(%rip),%ymm1 # 4a94 <_sk_callback_hsw+0x11a> + DB 196,226,125,24,13,42,73,0,0 ; vbroadcastss 0x492a(%rip),%ymm1 # 4a90 <_sk_callback_hsw+0x124> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 197,252,88,2 ; vaddps (%rdx),%ymm0,%ymm0 - DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 + DB 197,252,88,7 ; vaddps (%rdi),%ymm0,%ymm0 + DB 197,249,110,209 ; vmovd %ecx,%xmm2 + DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,33,73,0,0 ; vbroadcastss 0x4921(%rip),%ymm2 # 4a98 <_sk_callback_hsw+0x11e> + DB 196,226,125,24,21,10,73,0,0 ; vbroadcastss 0x490a(%rip),%ymm2 # 4a94 <_sk_callback_hsw+0x128> DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 @@ -123,19 +128,19 @@ _sk_seed_shader_hsw LABEL PROC PUBLIC _sk_dither_hsw _sk_dither_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,121,110,199 ; vmovd %edi,%xmm8 + DB 197,121,110,194 ; vmovd %edx,%xmm8 DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 - DB 197,61,254,66,32 ; vpaddd 0x20(%rdx),%ymm8,%ymm8 - DB 76,139,0 ; mov (%rax),%r8 - DB 196,66,125,88,8 ; vpbroadcastd (%r8),%ymm9 - DB 196,65,61,239,201 ; vpxor %ymm9,%ymm8,%ymm9 - DB 196,98,125,88,21,233,72,0,0 ; vpbroadcastd 0x48e9(%rip),%ymm10 # 4a9c <_sk_callback_hsw+0x122> + DB 197,61,254,71,32 ; vpaddd 0x20(%rdi),%ymm8,%ymm8 + DB 197,121,110,201 ; vmovd %ecx,%xmm9 + DB 196,66,125,88,201 ; vpbroadcastd %xmm9,%ymm9 + DB 196,65,53,239,200 ; vpxor %ymm8,%ymm9,%ymm9 + DB 196,98,125,88,21,209,72,0,0 ; vpbroadcastd 0x48d1(%rip),%ymm10 # 4a98 <_sk_callback_hsw+0x12c> DB 196,65,53,219,218 ; vpand %ymm10,%ymm9,%ymm11 DB 196,193,37,114,243,5 ; vpslld $0x5,%ymm11,%ymm11 DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,193,45,114,242,4 ; vpslld $0x4,%ymm10,%ymm10 - DB 196,98,125,88,37,206,72,0,0 ; vpbroadcastd 0x48ce(%rip),%ymm12 # 4aa0 <_sk_callback_hsw+0x126> - DB 196,98,125,88,45,201,72,0,0 ; vpbroadcastd 0x48c9(%rip),%ymm13 # 4aa4 <_sk_callback_hsw+0x12a> + DB 196,98,125,88,37,182,72,0,0 ; vpbroadcastd 0x48b6(%rip),%ymm12 # 4a9c <_sk_callback_hsw+0x130> + DB 196,98,125,88,45,177,72,0,0 ; vpbroadcastd 0x48b1(%rip),%ymm13 # 4aa0 <_sk_callback_hsw+0x134> DB 196,65,53,219,245 ; vpand %ymm13,%ymm9,%ymm14 DB 196,193,13,114,246,2 ; vpslld $0x2,%ymm14,%ymm14 DB 196,65,61,219,237 ; vpand %ymm13,%ymm8,%ymm13 @@ -146,14 +151,14 @@ _sk_dither_hsw LABEL PROC DB 196,193,61,114,208,2 ; vpsrld $0x2,%ymm8,%ymm8 DB 196,65,21,235,210 ; vpor %ymm10,%ymm13,%ymm10 DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 - DB 196,65,61,235,195 ; vpor %ymm11,%ymm8,%ymm8 - DB 196,65,13,235,201 ; vpor %ymm9,%ymm14,%ymm9 + DB 196,65,37,235,214 ; vpor %ymm14,%ymm11,%ymm10 + DB 196,65,61,235,194 ; vpor %ymm10,%ymm8,%ymm8 DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,123,72,0,0 ; vbroadcastss 0x487b(%rip),%ymm9 # 4aa8 <_sk_callback_hsw+0x12e> - DB 196,98,125,24,21,118,72,0,0 ; vbroadcastss 0x4876(%rip),%ymm10 # 4aac <_sk_callback_hsw+0x132> + DB 196,98,125,24,13,99,72,0,0 ; vbroadcastss 0x4863(%rip),%ymm9 # 4aa4 <_sk_callback_hsw+0x138> + DB 196,98,125,24,21,94,72,0,0 ; vbroadcastss 0x485e(%rip),%ymm10 # 4aa8 <_sk_callback_hsw+0x13c> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,64,8 ; vbroadcastss 0x8(%rax),%ymm8 + DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 DB 197,188,88,201 ; vaddps %ymm1,%ymm8,%ymm1 @@ -210,7 +215,7 @@ _sk_clear_hsw LABEL PROC PUBLIC _sk_srcatop_hsw _sk_srcatop_hsw LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,205,71,0,0 ; vbroadcastss 0x47cd(%rip),%ymm8 # 4ab0 <_sk_callback_hsw+0x136> + DB 196,98,125,24,5,182,71,0,0 ; vbroadcastss 0x47b6(%rip),%ymm8 # 4aac <_sk_callback_hsw+0x140> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,226,61,184,196 ; vfmadd231ps %ymm4,%ymm8,%ymm0 DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 @@ -224,7 +229,7 @@ _sk_srcatop_hsw LABEL PROC PUBLIC _sk_dstatop_hsw _sk_dstatop_hsw LABEL PROC - DB 196,98,125,24,5,160,71,0,0 ; vbroadcastss 0x47a0(%rip),%ymm8 # 4ab4 <_sk_callback_hsw+0x13a> + DB 196,98,125,24,5,137,71,0,0 ; vbroadcastss 0x4789(%rip),%ymm8 # 4ab0 <_sk_callback_hsw+0x144> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,226,101,184,196 ; vfmadd231ps %ymm4,%ymm3,%ymm0 @@ -257,7 +262,7 @@ _sk_dstin_hsw LABEL PROC PUBLIC _sk_srcout_hsw _sk_srcout_hsw LABEL PROC - DB 196,98,125,24,5,71,71,0,0 ; vbroadcastss 0x4747(%rip),%ymm8 # 4ab8 <_sk_callback_hsw+0x13e> + DB 196,98,125,24,5,48,71,0,0 ; vbroadcastss 0x4730(%rip),%ymm8 # 4ab4 <_sk_callback_hsw+0x148> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -268,7 +273,7 @@ _sk_srcout_hsw LABEL PROC PUBLIC _sk_dstout_hsw _sk_dstout_hsw LABEL PROC - DB 196,226,125,24,5,42,71,0,0 ; vbroadcastss 0x472a(%rip),%ymm0 # 4abc <_sk_callback_hsw+0x142> + DB 196,226,125,24,5,19,71,0,0 ; vbroadcastss 0x4713(%rip),%ymm0 # 4ab8 <_sk_callback_hsw+0x14c> DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3 DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0 DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1 @@ -279,7 +284,7 @@ _sk_dstout_hsw LABEL PROC PUBLIC _sk_srcover_hsw _sk_srcover_hsw LABEL PROC - DB 196,98,125,24,5,13,71,0,0 ; vbroadcastss 0x470d(%rip),%ymm8 # 4ac0 <_sk_callback_hsw+0x146> + DB 196,98,125,24,5,246,70,0,0 ; vbroadcastss 0x46f6(%rip),%ymm8 # 4abc <_sk_callback_hsw+0x150> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0 DB 196,194,85,184,200 ; vfmadd231ps %ymm8,%ymm5,%ymm1 @@ -290,7 +295,7 @@ _sk_srcover_hsw LABEL PROC PUBLIC _sk_dstover_hsw _sk_dstover_hsw LABEL PROC - DB 196,98,125,24,5,236,70,0,0 ; vbroadcastss 0x46ec(%rip),%ymm8 # 4ac4 <_sk_callback_hsw+0x14a> + DB 196,98,125,24,5,213,70,0,0 ; vbroadcastss 0x46d5(%rip),%ymm8 # 4ac0 <_sk_callback_hsw+0x154> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 DB 196,226,61,168,205 ; vfmadd213ps %ymm5,%ymm8,%ymm1 @@ -310,7 +315,7 @@ _sk_modulate_hsw LABEL PROC PUBLIC _sk_multiply_hsw _sk_multiply_hsw LABEL PROC - DB 196,98,125,24,5,183,70,0,0 ; vbroadcastss 0x46b7(%rip),%ymm8 # 4ac8 <_sk_callback_hsw+0x14e> + DB 196,98,125,24,5,160,70,0,0 ; vbroadcastss 0x46a0(%rip),%ymm8 # 4ac4 <_sk_callback_hsw+0x158> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -352,7 +357,7 @@ _sk_screen_hsw LABEL PROC PUBLIC _sk_xor__hsw _sk_xor__hsw LABEL PROC - DB 196,98,125,24,5,50,70,0,0 ; vbroadcastss 0x4632(%rip),%ymm8 # 4acc <_sk_callback_hsw+0x152> + DB 196,98,125,24,5,27,70,0,0 ; vbroadcastss 0x461b(%rip),%ymm8 # 4ac8 <_sk_callback_hsw+0x15c> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -384,7 +389,7 @@ _sk_darken_hsw LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,186,69,0,0 ; vbroadcastss 0x45ba(%rip),%ymm8 # 4ad0 <_sk_callback_hsw+0x156> + DB 196,98,125,24,5,163,69,0,0 ; vbroadcastss 0x45a3(%rip),%ymm8 # 4acc <_sk_callback_hsw+0x160> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -407,7 +412,7 @@ _sk_lighten_hsw LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,105,69,0,0 ; vbroadcastss 0x4569(%rip),%ymm8 # 4ad4 <_sk_callback_hsw+0x15a> + DB 196,98,125,24,5,82,69,0,0 ; vbroadcastss 0x4552(%rip),%ymm8 # 4ad0 <_sk_callback_hsw+0x164> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -433,7 +438,7 @@ _sk_difference_hsw LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,12,69,0,0 ; vbroadcastss 0x450c(%rip),%ymm8 # 4ad8 <_sk_callback_hsw+0x15e> + DB 196,98,125,24,5,245,68,0,0 ; vbroadcastss 0x44f5(%rip),%ymm8 # 4ad4 <_sk_callback_hsw+0x168> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -453,7 +458,7 @@ _sk_exclusion_hsw LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,202,68,0,0 ; vbroadcastss 0x44ca(%rip),%ymm8 # 4adc <_sk_callback_hsw+0x162> + DB 196,98,125,24,5,179,68,0,0 ; vbroadcastss 0x44b3(%rip),%ymm8 # 4ad8 <_sk_callback_hsw+0x16c> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -461,7 +466,7 @@ _sk_exclusion_hsw LABEL PROC PUBLIC _sk_colorburn_hsw _sk_colorburn_hsw LABEL PROC - DB 196,98,125,24,5,184,68,0,0 ; vbroadcastss 0x44b8(%rip),%ymm8 # 4ae0 <_sk_callback_hsw+0x166> + DB 196,98,125,24,5,161,68,0,0 ; vbroadcastss 0x44a1(%rip),%ymm8 # 4adc <_sk_callback_hsw+0x170> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 @@ -517,7 +522,7 @@ _sk_colorburn_hsw LABEL PROC PUBLIC _sk_colordodge_hsw _sk_colordodge_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,195,67,0,0 ; vbroadcastss 0x43c3(%rip),%ymm9 # 4ae4 <_sk_callback_hsw+0x16a> + DB 196,98,125,24,13,172,67,0,0 ; vbroadcastss 0x43ac(%rip),%ymm9 # 4ae0 <_sk_callback_hsw+0x174> DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9 @@ -568,7 +573,7 @@ _sk_colordodge_hsw LABEL PROC PUBLIC _sk_hardlight_hsw _sk_hardlight_hsw LABEL PROC - DB 196,98,125,24,5,228,66,0,0 ; vbroadcastss 0x42e4(%rip),%ymm8 # 4ae8 <_sk_callback_hsw+0x16e> + DB 196,98,125,24,5,205,66,0,0 ; vbroadcastss 0x42cd(%rip),%ymm8 # 4ae4 <_sk_callback_hsw+0x178> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -617,7 +622,7 @@ _sk_hardlight_hsw LABEL PROC PUBLIC _sk_overlay_hsw _sk_overlay_hsw LABEL PROC - DB 196,98,125,24,5,28,66,0,0 ; vbroadcastss 0x421c(%rip),%ymm8 # 4aec <_sk_callback_hsw+0x172> + DB 196,98,125,24,5,5,66,0,0 ; vbroadcastss 0x4205(%rip),%ymm8 # 4ae8 <_sk_callback_hsw+0x17c> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -677,10 +682,10 @@ _sk_softlight_hsw LABEL PROC DB 196,65,20,88,197 ; vaddps %ymm13,%ymm13,%ymm8 DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,66,61,168,192 ; vfmadd213ps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,29,35,65,0,0 ; vbroadcastss 0x4123(%rip),%ymm11 # 4af4 <_sk_callback_hsw+0x17a> + DB 196,98,125,24,29,12,65,0,0 ; vbroadcastss 0x410c(%rip),%ymm11 # 4af0 <_sk_callback_hsw+0x184> DB 196,65,20,88,227 ; vaddps %ymm11,%ymm13,%ymm12 DB 196,65,28,89,192 ; vmulps %ymm8,%ymm12,%ymm8 - DB 196,98,125,24,37,20,65,0,0 ; vbroadcastss 0x4114(%rip),%ymm12 # 4af8 <_sk_callback_hsw+0x17e> + DB 196,98,125,24,37,253,64,0,0 ; vbroadcastss 0x40fd(%rip),%ymm12 # 4af4 <_sk_callback_hsw+0x188> DB 196,66,21,184,196 ; vfmadd231ps %ymm12,%ymm13,%ymm8 DB 196,65,124,82,245 ; vrsqrtps %ymm13,%ymm14 DB 196,65,124,83,246 ; vrcpps %ymm14,%ymm14 @@ -690,7 +695,7 @@ _sk_softlight_hsw LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,215,64,0,0 ; vbroadcastss 0x40d7(%rip),%ymm8 # 4af0 <_sk_callback_hsw+0x176> + DB 196,98,125,24,5,192,64,0,0 ; vbroadcastss 0x40c0(%rip),%ymm8 # 4aec <_sk_callback_hsw+0x180> DB 196,65,60,92,237 ; vsubps %ymm13,%ymm8,%ymm13 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,98,125,168,235 ; vfmadd213ps %ymm3,%ymm0,%ymm13 @@ -803,11 +808,11 @@ _sk_hue_hsw LABEL PROC DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10 DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10 DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - DB 196,98,125,24,53,214,62,0,0 ; vbroadcastss 0x3ed6(%rip),%ymm14 # 4afc <_sk_callback_hsw+0x182> - DB 196,98,125,24,61,209,62,0,0 ; vbroadcastss 0x3ed1(%rip),%ymm15 # 4b00 <_sk_callback_hsw+0x186> + DB 196,98,125,24,53,191,62,0,0 ; vbroadcastss 0x3ebf(%rip),%ymm14 # 4af8 <_sk_callback_hsw+0x18c> + DB 196,98,125,24,61,186,62,0,0 ; vbroadcastss 0x3eba(%rip),%ymm15 # 4afc <_sk_callback_hsw+0x190> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13 - DB 196,226,125,24,5,194,62,0,0 ; vbroadcastss 0x3ec2(%rip),%ymm0 # 4b04 <_sk_callback_hsw+0x18a> + DB 196,226,125,24,5,171,62,0,0 ; vbroadcastss 0x3eab(%rip),%ymm0 # 4b00 <_sk_callback_hsw+0x194> DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13 DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10 DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -862,7 +867,7 @@ _sk_hue_hsw LABEL PROC DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9 DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,13,175,61,0,0 ; vbroadcastss 0x3daf(%rip),%ymm1 # 4b08 <_sk_callback_hsw+0x18e> + DB 196,226,125,24,13,152,61,0,0 ; vbroadcastss 0x3d98(%rip),%ymm1 # 4b04 <_sk_callback_hsw+0x198> DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11 @@ -916,11 +921,11 @@ _sk_saturation_hsw LABEL PROC DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10 DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10 DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - DB 196,98,125,24,53,192,60,0,0 ; vbroadcastss 0x3cc0(%rip),%ymm14 # 4b0c <_sk_callback_hsw+0x192> - DB 196,98,125,24,61,187,60,0,0 ; vbroadcastss 0x3cbb(%rip),%ymm15 # 4b10 <_sk_callback_hsw+0x196> + DB 196,98,125,24,53,169,60,0,0 ; vbroadcastss 0x3ca9(%rip),%ymm14 # 4b08 <_sk_callback_hsw+0x19c> + DB 196,98,125,24,61,164,60,0,0 ; vbroadcastss 0x3ca4(%rip),%ymm15 # 4b0c <_sk_callback_hsw+0x1a0> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13 - DB 196,226,125,24,5,172,60,0,0 ; vbroadcastss 0x3cac(%rip),%ymm0 # 4b14 <_sk_callback_hsw+0x19a> + DB 196,226,125,24,5,149,60,0,0 ; vbroadcastss 0x3c95(%rip),%ymm0 # 4b10 <_sk_callback_hsw+0x1a4> DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13 DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10 DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -975,7 +980,7 @@ _sk_saturation_hsw LABEL PROC DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9 DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,13,153,59,0,0 ; vbroadcastss 0x3b99(%rip),%ymm1 # 4b18 <_sk_callback_hsw+0x19e> + DB 196,226,125,24,13,130,59,0,0 ; vbroadcastss 0x3b82(%rip),%ymm1 # 4b14 <_sk_callback_hsw+0x1a8> DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11 @@ -1003,11 +1008,11 @@ _sk_color_hsw LABEL PROC DB 197,108,89,199 ; vmulps %ymm7,%ymm2,%ymm8 DB 197,116,89,215 ; vmulps %ymm7,%ymm1,%ymm10 DB 197,52,89,223 ; vmulps %ymm7,%ymm9,%ymm11 - DB 196,98,125,24,45,44,59,0,0 ; vbroadcastss 0x3b2c(%rip),%ymm13 # 4b1c <_sk_callback_hsw+0x1a2> - DB 196,98,125,24,53,39,59,0,0 ; vbroadcastss 0x3b27(%rip),%ymm14 # 4b20 <_sk_callback_hsw+0x1a6> + DB 196,98,125,24,45,21,59,0,0 ; vbroadcastss 0x3b15(%rip),%ymm13 # 4b18 <_sk_callback_hsw+0x1ac> + DB 196,98,125,24,53,16,59,0,0 ; vbroadcastss 0x3b10(%rip),%ymm14 # 4b1c <_sk_callback_hsw+0x1b0> DB 196,65,84,89,230 ; vmulps %ymm14,%ymm5,%ymm12 DB 196,66,93,184,229 ; vfmadd231ps %ymm13,%ymm4,%ymm12 - DB 196,98,125,24,61,24,59,0,0 ; vbroadcastss 0x3b18(%rip),%ymm15 # 4b24 <_sk_callback_hsw+0x1aa> + DB 196,98,125,24,61,1,59,0,0 ; vbroadcastss 0x3b01(%rip),%ymm15 # 4b20 <_sk_callback_hsw+0x1b4> DB 196,66,77,184,231 ; vfmadd231ps %ymm15,%ymm6,%ymm12 DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9 DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -1063,7 +1068,7 @@ _sk_color_hsw LABEL PROC DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1 DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8 DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9 - DB 196,226,125,24,5,250,57,0,0 ; vbroadcastss 0x39fa(%rip),%ymm0 # 4b28 <_sk_callback_hsw+0x1ae> + DB 196,226,125,24,5,227,57,0,0 ; vbroadcastss 0x39e3(%rip),%ymm0 # 4b24 <_sk_callback_hsw+0x1b8> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -1091,11 +1096,11 @@ _sk_luminosity_hsw LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,213 ; vmulps %ymm5,%ymm3,%ymm10 DB 197,100,89,222 ; vmulps %ymm6,%ymm3,%ymm11 - DB 196,98,125,24,45,141,57,0,0 ; vbroadcastss 0x398d(%rip),%ymm13 # 4b2c <_sk_callback_hsw+0x1b2> - DB 196,98,125,24,53,136,57,0,0 ; vbroadcastss 0x3988(%rip),%ymm14 # 4b30 <_sk_callback_hsw+0x1b6> + DB 196,98,125,24,45,118,57,0,0 ; vbroadcastss 0x3976(%rip),%ymm13 # 4b28 <_sk_callback_hsw+0x1bc> + DB 196,98,125,24,53,113,57,0,0 ; vbroadcastss 0x3971(%rip),%ymm14 # 4b2c <_sk_callback_hsw+0x1c0> DB 196,65,116,89,230 ; vmulps %ymm14,%ymm1,%ymm12 DB 196,66,109,184,229 ; vfmadd231ps %ymm13,%ymm2,%ymm12 - DB 196,98,125,24,61,121,57,0,0 ; vbroadcastss 0x3979(%rip),%ymm15 # 4b34 <_sk_callback_hsw+0x1ba> + DB 196,98,125,24,61,98,57,0,0 ; vbroadcastss 0x3962(%rip),%ymm15 # 4b30 <_sk_callback_hsw+0x1c4> DB 196,66,53,184,231 ; vfmadd231ps %ymm15,%ymm9,%ymm12 DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9 DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -1151,7 +1156,7 @@ _sk_luminosity_hsw LABEL PROC DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1 DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8 DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9 - DB 196,226,125,24,5,91,56,0,0 ; vbroadcastss 0x385b(%rip),%ymm0 # 4b38 <_sk_callback_hsw+0x1be> + DB 196,226,125,24,5,68,56,0,0 ; vbroadcastss 0x3844(%rip),%ymm0 # 4b34 <_sk_callback_hsw+0x1c8> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -1171,24 +1176,24 @@ _sk_luminosity_hsw LABEL PROC PUBLIC _sk_srcover_rgba_8888_hsw _sk_srcover_rgba_8888_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 DB 77,133,192 ; test %r8,%r8 - DB 15,133,180,0,0,0 ; jne 13ef <_sk_srcover_rgba_8888_hsw+0xcd> - DB 196,193,124,16,57 ; vmovups (%r9),%ymm7 - DB 197,196,84,37,88,59,0,0 ; vandps 0x3b58(%rip),%ymm7,%ymm4 # 4ea0 <_sk_callback_hsw+0x526> + DB 15,133,180,0,0,0 ; jne 1402 <_sk_srcover_rgba_8888_hsw+0xcd> + DB 196,193,124,16,58 ; vmovups (%r10),%ymm7 + DB 197,196,84,37,69,59,0,0 ; vandps 0x3b45(%rip),%ymm7,%ymm4 # 4ea0 <_sk_callback_hsw+0x534> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,69,0,45,107,59,0,0 ; vpshufb 0x3b6b(%rip),%ymm7,%ymm5 # 4ec0 <_sk_callback_hsw+0x546> + DB 196,226,69,0,45,88,59,0,0 ; vpshufb 0x3b58(%rip),%ymm7,%ymm5 # 4ec0 <_sk_callback_hsw+0x554> DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,69,0,53,126,59,0,0 ; vpshufb 0x3b7e(%rip),%ymm7,%ymm6 # 4ee0 <_sk_callback_hsw+0x566> + DB 196,226,69,0,53,107,59,0,0 ; vpshufb 0x3b6b(%rip),%ymm7,%ymm6 # 4ee0 <_sk_callback_hsw+0x574> DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 DB 197,197,114,215,24 ; vpsrld $0x18,%ymm7,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,196,55,0,0 ; vbroadcastss 0x37c4(%rip),%ymm8 # 4b3c <_sk_callback_hsw+0x1c2> + DB 196,98,125,24,5,173,55,0,0 ; vbroadcastss 0x37ad(%rip),%ymm8 # 4b38 <_sk_callback_hsw+0x1cc> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,187,55,0,0 ; vbroadcastss 0x37bb(%rip),%ymm9 # 4b40 <_sk_callback_hsw+0x1c6> + DB 196,98,125,24,13,164,55,0,0 ; vbroadcastss 0x37a4(%rip),%ymm9 # 4b3c <_sk_callback_hsw+0x1d0> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 @@ -1208,10 +1213,10 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 196,65,53,235,202 ; vpor %ymm10,%ymm9,%ymm9 DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8 DB 77,133,192 ; test %r8,%r8 - DB 117,53 ; jne 1418 <_sk_srcover_rgba_8888_hsw+0xf6> - DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) + DB 117,53 ; jne 142b <_sk_srcover_rgba_8888_hsw+0xf6> + DB 196,65,124,17,2 ; vmovups %ymm8,(%r10) DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx @@ -1220,8 +1225,8 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,224 ; vmovq %rax,%xmm4 DB 196,226,125,33,228 ; vpmovsxbd %xmm4,%ymm4 - DB 196,194,93,44,57 ; vmaskmovps (%r9),%ymm4,%ymm7 - DB 233,40,255,255,255 ; jmpq 1340 <_sk_srcover_rgba_8888_hsw+0x1e> + DB 196,194,93,44,58 ; vmaskmovps (%r10),%ymm4,%ymm7 + DB 233,40,255,255,255 ; jmpq 1353 <_sk_srcover_rgba_8888_hsw+0x1e> DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx DB 192,225,3 ; shl $0x3,%cl @@ -1229,8 +1234,8 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9 - DB 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) - DB 235,170 ; jmp 13e8 <_sk_srcover_rgba_8888_hsw+0xc6> + DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10) + DB 235,170 ; jmp 13fb <_sk_srcover_rgba_8888_hsw+0xc6> PUBLIC _sk_clamp_0_hsw _sk_clamp_0_hsw LABEL PROC @@ -1244,7 +1249,7 @@ _sk_clamp_0_hsw LABEL PROC PUBLIC _sk_clamp_1_hsw _sk_clamp_1_hsw LABEL PROC - DB 196,98,125,24,5,224,54,0,0 ; vbroadcastss 0x36e0(%rip),%ymm8 # 4b44 <_sk_callback_hsw+0x1ca> + DB 196,98,125,24,5,201,54,0,0 ; vbroadcastss 0x36c9(%rip),%ymm8 # 4b40 <_sk_callback_hsw+0x1d4> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 @@ -1254,7 +1259,7 @@ _sk_clamp_1_hsw LABEL PROC PUBLIC _sk_clamp_a_hsw _sk_clamp_a_hsw LABEL PROC - DB 196,98,125,24,5,195,54,0,0 ; vbroadcastss 0x36c3(%rip),%ymm8 # 4b48 <_sk_callback_hsw+0x1ce> + DB 196,98,125,24,5,172,54,0,0 ; vbroadcastss 0x36ac(%rip),%ymm8 # 4b44 <_sk_callback_hsw+0x1d8> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0 DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1 @@ -1326,7 +1331,7 @@ PUBLIC _sk_unpremul_hsw _sk_unpremul_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,11,54,0,0 ; vbroadcastss 0x360b(%rip),%ymm10 # 4b4c <_sk_callback_hsw+0x1d2> + DB 196,98,125,24,21,244,53,0,0 ; vbroadcastss 0x35f4(%rip),%ymm10 # 4b48 <_sk_callback_hsw+0x1dc> DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10 DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 @@ -1337,16 +1342,16 @@ _sk_unpremul_hsw LABEL PROC PUBLIC _sk_from_srgb_hsw _sk_from_srgb_hsw LABEL PROC - DB 196,98,125,24,5,236,53,0,0 ; vbroadcastss 0x35ec(%rip),%ymm8 # 4b50 <_sk_callback_hsw+0x1d6> + DB 196,98,125,24,5,213,53,0,0 ; vbroadcastss 0x35d5(%rip),%ymm8 # 4b4c <_sk_callback_hsw+0x1e0> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,222,53,0,0 ; vbroadcastss 0x35de(%rip),%ymm11 # 4b54 <_sk_callback_hsw+0x1da> - DB 196,98,125,24,37,217,53,0,0 ; vbroadcastss 0x35d9(%rip),%ymm12 # 4b58 <_sk_callback_hsw+0x1de> + DB 196,98,125,24,29,199,53,0,0 ; vbroadcastss 0x35c7(%rip),%ymm11 # 4b50 <_sk_callback_hsw+0x1e4> + DB 196,98,125,24,37,194,53,0,0 ; vbroadcastss 0x35c2(%rip),%ymm12 # 4b54 <_sk_callback_hsw+0x1e8> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13 - DB 196,98,125,24,53,202,53,0,0 ; vbroadcastss 0x35ca(%rip),%ymm14 # 4b5c <_sk_callback_hsw+0x1e2> + DB 196,98,125,24,53,179,53,0,0 ; vbroadcastss 0x35b3(%rip),%ymm14 # 4b58 <_sk_callback_hsw+0x1ec> DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13 - DB 196,98,125,24,21,192,53,0,0 ; vbroadcastss 0x35c0(%rip),%ymm10 # 4b60 <_sk_callback_hsw+0x1e6> + DB 196,98,125,24,21,169,53,0,0 ; vbroadcastss 0x35a9(%rip),%ymm10 # 4b5c <_sk_callback_hsw+0x1f0> DB 196,193,124,194,194,1 ; vcmpltps %ymm10,%ymm0,%ymm0 DB 196,195,21,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm13,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -1368,19 +1373,19 @@ _sk_from_srgb_hsw LABEL PROC PUBLIC _sk_to_srgb_hsw _sk_to_srgb_hsw LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,100,53,0,0 ; vbroadcastss 0x3564(%rip),%ymm8 # 4b64 <_sk_callback_hsw+0x1ea> + DB 196,98,125,24,5,77,53,0,0 ; vbroadcastss 0x354d(%rip),%ymm8 # 4b60 <_sk_callback_hsw+0x1f4> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,90,53,0,0 ; vbroadcastss 0x355a(%rip),%ymm11 # 4b68 <_sk_callback_hsw+0x1ee> - DB 196,98,125,24,37,85,53,0,0 ; vbroadcastss 0x3555(%rip),%ymm12 # 4b6c <_sk_callback_hsw+0x1f2> + DB 196,98,125,24,29,67,53,0,0 ; vbroadcastss 0x3543(%rip),%ymm11 # 4b64 <_sk_callback_hsw+0x1f8> + DB 196,98,125,24,37,62,53,0,0 ; vbroadcastss 0x353e(%rip),%ymm12 # 4b68 <_sk_callback_hsw+0x1fc> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,53,168,235 ; vfmadd213ps %ymm11,%ymm9,%ymm13 - DB 196,98,125,24,53,70,53,0,0 ; vbroadcastss 0x3546(%rip),%ymm14 # 4b70 <_sk_callback_hsw+0x1f6> + DB 196,98,125,24,53,47,53,0,0 ; vbroadcastss 0x352f(%rip),%ymm14 # 4b6c <_sk_callback_hsw+0x200> DB 196,66,53,168,238 ; vfmadd213ps %ymm14,%ymm9,%ymm13 - DB 196,98,125,24,61,60,53,0,0 ; vbroadcastss 0x353c(%rip),%ymm15 # 4b74 <_sk_callback_hsw+0x1fa> + DB 196,98,125,24,61,37,53,0,0 ; vbroadcastss 0x3525(%rip),%ymm15 # 4b70 <_sk_callback_hsw+0x204> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,20,89,201 ; vmulps %ymm9,%ymm13,%ymm9 - DB 196,98,125,24,45,40,53,0,0 ; vbroadcastss 0x3528(%rip),%ymm13 # 4b78 <_sk_callback_hsw+0x1fe> + DB 196,98,125,24,45,17,53,0,0 ; vbroadcastss 0x3511(%rip),%ymm13 # 4b74 <_sk_callback_hsw+0x208> DB 196,193,124,194,197,1 ; vcmpltps %ymm13,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -1412,26 +1417,26 @@ _sk_rgb_to_hsl_hsw LABEL PROC DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9 DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,157,52,0,0 ; vbroadcastss 0x349d(%rip),%ymm11 # 4b7c <_sk_callback_hsw+0x202> + DB 196,98,125,24,29,134,52,0,0 ; vbroadcastss 0x3486(%rip),%ymm11 # 4b78 <_sk_callback_hsw+0x20c> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 197,116,194,234,1 ; vcmpltps %ymm2,%ymm1,%ymm13 - DB 196,98,125,24,53,138,52,0,0 ; vbroadcastss 0x348a(%rip),%ymm14 # 4b80 <_sk_callback_hsw+0x206> + DB 196,98,125,24,53,115,52,0,0 ; vbroadcastss 0x3473(%rip),%ymm14 # 4b7c <_sk_callback_hsw+0x210> DB 196,65,4,87,255 ; vxorps %ymm15,%ymm15,%ymm15 DB 196,67,5,74,238,208 ; vblendvps %ymm13,%ymm14,%ymm15,%ymm13 DB 196,66,37,168,229 ; vfmadd213ps %ymm13,%ymm11,%ymm12 DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 DB 197,124,92,233 ; vsubps %ymm1,%ymm0,%ymm13 - DB 196,98,125,24,53,113,52,0,0 ; vbroadcastss 0x3471(%rip),%ymm14 # 4b88 <_sk_callback_hsw+0x20e> + DB 196,98,125,24,53,90,52,0,0 ; vbroadcastss 0x345a(%rip),%ymm14 # 4b84 <_sk_callback_hsw+0x218> DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13 - DB 196,98,125,24,53,95,52,0,0 ; vbroadcastss 0x345f(%rip),%ymm14 # 4b84 <_sk_callback_hsw+0x20a> + DB 196,98,125,24,53,72,52,0,0 ; vbroadcastss 0x3448(%rip),%ymm14 # 4b80 <_sk_callback_hsw+0x214> DB 196,194,37,168,214 ; vfmadd213ps %ymm14,%ymm11,%ymm2 DB 197,188,194,201,0 ; vcmpeqps %ymm1,%ymm8,%ymm1 DB 196,227,21,74,202,16 ; vblendvps %ymm1,%ymm2,%ymm13,%ymm1 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,195,117,74,196,0 ; vblendvps %ymm0,%ymm12,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,29,66,52,0,0 ; vbroadcastss 0x3442(%rip),%ymm11 # 4b90 <_sk_callback_hsw+0x216> + DB 196,98,125,24,29,43,52,0,0 ; vbroadcastss 0x342b(%rip),%ymm11 # 4b8c <_sk_callback_hsw+0x220> DB 196,193,116,89,211 ; vmulps %ymm11,%ymm1,%ymm2 DB 197,36,194,218,1 ; vcmpltps %ymm2,%ymm11,%ymm11 DB 196,65,12,92,224 ; vsubps %ymm8,%ymm14,%ymm12 @@ -1441,7 +1446,7 @@ _sk_rgb_to_hsl_hsw LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,199,128 ; vblendvps %ymm8,%ymm15,%ymm0,%ymm0 DB 196,195,117,74,207,128 ; vblendvps %ymm8,%ymm15,%ymm1,%ymm1 - DB 196,98,125,24,5,5,52,0,0 ; vbroadcastss 0x3405(%rip),%ymm8 # 4b8c <_sk_callback_hsw+0x212> + DB 196,98,125,24,5,238,51,0,0 ; vbroadcastss 0x33ee(%rip),%ymm8 # 4b88 <_sk_callback_hsw+0x21c> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1456,30 +1461,30 @@ _sk_hsl_to_rgb_hsw LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,233 ; vmovaps %ymm1,%ymm5 DB 197,252,40,224 ; vmovaps %ymm0,%ymm4 - DB 196,98,125,24,5,204,51,0,0 ; vbroadcastss 0x33cc(%rip),%ymm8 # 4b94 <_sk_callback_hsw+0x21a> + DB 196,98,125,24,5,181,51,0,0 ; vbroadcastss 0x33b5(%rip),%ymm8 # 4b90 <_sk_callback_hsw+0x224> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,84,89,210 ; vmulps %ymm2,%ymm5,%ymm10 DB 196,65,84,92,218 ; vsubps %ymm10,%ymm5,%ymm11 DB 196,67,45,74,203,144 ; vblendvps %ymm9,%ymm11,%ymm10,%ymm9 DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 - DB 196,98,125,24,13,175,51,0,0 ; vbroadcastss 0x33af(%rip),%ymm9 # 4b98 <_sk_callback_hsw+0x21e> + DB 196,98,125,24,13,152,51,0,0 ; vbroadcastss 0x3398(%rip),%ymm9 # 4b94 <_sk_callback_hsw+0x228> DB 196,66,109,170,202 ; vfmsub213ps %ymm10,%ymm2,%ymm9 - DB 196,98,125,24,29,165,51,0,0 ; vbroadcastss 0x33a5(%rip),%ymm11 # 4b9c <_sk_callback_hsw+0x222> + DB 196,98,125,24,29,142,51,0,0 ; vbroadcastss 0x338e(%rip),%ymm11 # 4b98 <_sk_callback_hsw+0x22c> DB 196,65,92,88,219 ; vaddps %ymm11,%ymm4,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,45,143,51,0,0 ; vbroadcastss 0x338f(%rip),%ymm13 # 4ba4 <_sk_callback_hsw+0x22a> + DB 196,98,125,24,45,120,51,0,0 ; vbroadcastss 0x3378(%rip),%ymm13 # 4ba0 <_sk_callback_hsw+0x234> DB 196,193,4,89,197 ; vmulps %ymm13,%ymm15,%ymm0 - DB 196,98,125,24,53,133,51,0,0 ; vbroadcastss 0x3385(%rip),%ymm14 # 4ba8 <_sk_callback_hsw+0x22e> + DB 196,98,125,24,53,110,51,0,0 ; vbroadcastss 0x336e(%rip),%ymm14 # 4ba4 <_sk_callback_hsw+0x238> DB 197,12,92,224 ; vsubps %ymm0,%ymm14,%ymm12 DB 196,66,37,168,225 ; vfmadd213ps %ymm9,%ymm11,%ymm12 - DB 196,226,125,24,29,107,51,0,0 ; vbroadcastss 0x336b(%rip),%ymm3 # 4ba0 <_sk_callback_hsw+0x226> + DB 196,226,125,24,29,84,51,0,0 ; vbroadcastss 0x3354(%rip),%ymm3 # 4b9c <_sk_callback_hsw+0x230> DB 196,193,100,194,255,2 ; vcmpleps %ymm15,%ymm3,%ymm7 DB 196,195,29,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm12,%ymm7 DB 196,65,60,194,231,2 ; vcmpleps %ymm15,%ymm8,%ymm12 DB 196,227,45,74,255,192 ; vblendvps %ymm12,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,37,86,51,0,0 ; vbroadcastss 0x3356(%rip),%ymm12 # 4bac <_sk_callback_hsw+0x232> + DB 196,98,125,24,37,63,51,0,0 ; vbroadcastss 0x333f(%rip),%ymm12 # 4ba8 <_sk_callback_hsw+0x23c> DB 196,65,28,194,255,2 ; vcmpleps %ymm15,%ymm12,%ymm15 DB 196,194,37,168,193 ; vfmadd213ps %ymm9,%ymm11,%ymm0 DB 196,99,125,74,255,240 ; vblendvps %ymm15,%ymm7,%ymm0,%ymm15 @@ -1495,7 +1500,7 @@ _sk_hsl_to_rgb_hsw LABEL PROC DB 197,156,194,192,2 ; vcmpleps %ymm0,%ymm12,%ymm0 DB 196,194,37,168,249 ; vfmadd213ps %ymm9,%ymm11,%ymm7 DB 196,227,69,74,201,0 ; vblendvps %ymm0,%ymm1,%ymm7,%ymm1 - DB 196,226,125,24,5,2,51,0,0 ; vbroadcastss 0x3302(%rip),%ymm0 # 4bb0 <_sk_callback_hsw+0x236> + DB 196,226,125,24,5,235,50,0,0 ; vbroadcastss 0x32eb(%rip),%ymm0 # 4bac <_sk_callback_hsw+0x240> DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 DB 196,227,125,8,224,1 ; vroundps $0x1,%ymm0,%ymm4 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 @@ -1536,36 +1541,38 @@ _sk_scale_1_float_hsw LABEL PROC PUBLIC _sk_scale_u8_hsw _sk_scale_u8_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,51 ; jne 1994 <_sk_scale_u8_hsw+0x43> + DB 117,51 ; jne 19a7 <_sk_scale_u8_hsw+0x43> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,60,50,0,0 ; vbroadcastss 0x323c(%rip),%ymm9 # 4bb4 <_sk_callback_hsw+0x23a> + DB 196,98,125,24,13,37,50,0,0 ; vbroadcastss 0x3225(%rip),%ymm9 # 4bb0 <_sk_callback_hsw+0x244> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 199c <_sk_scale_u8_hsw+0x4b> - DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,172 ; jmp 1965 <_sk_scale_u8_hsw+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 19b0 <_sk_scale_u8_hsw+0x4c> + DB 196,65,249,110,194 ; vmovq %r10,%xmm8 + DB 91 ; pop %rbx + DB 235,171 ; jmp 1978 <_sk_scale_u8_hsw+0x14> PUBLIC _sk_lerp_1_float_hsw _sk_lerp_1_float_hsw LABEL PROC @@ -1584,16 +1591,16 @@ _sk_lerp_1_float_hsw LABEL PROC PUBLIC _sk_lerp_u8_hsw _sk_lerp_u8_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,71 ; jne 1a3f <_sk_lerp_u8_hsw+0x57> + DB 117,71 ; jne 1a53 <_sk_lerp_u8_hsw+0x57> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,169,49,0,0 ; vbroadcastss 0x31a9(%rip),%ymm9 # 4bb8 <_sk_callback_hsw+0x23e> + DB 196,98,125,24,13,145,49,0,0 ; vbroadcastss 0x3191(%rip),%ymm9 # 4bb4 <_sk_callback_hsw+0x248> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -1604,43 +1611,45 @@ _sk_lerp_u8_hsw LABEL PROC DB 197,228,92,223 ; vsubps %ymm7,%ymm3,%ymm3 DB 196,226,61,168,223 ; vfmadd213ps %ymm7,%ymm8,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 1a47 <_sk_lerp_u8_hsw+0x5f> - DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,152 ; jmp 19fc <_sk_lerp_u8_hsw+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 1a5c <_sk_lerp_u8_hsw+0x60> + DB 196,65,249,110,194 ; vmovq %r10,%xmm8 + DB 91 ; pop %rbx + DB 235,151 ; jmp 1a10 <_sk_lerp_u8_hsw+0x14> PUBLIC _sk_lerp_565_hsw _sk_lerp_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,169,0,0,0 ; jne 1b1b <_sk_lerp_565_hsw+0xb7> - DB 196,65,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm8 + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,169,0,0,0 ; jne 1b30 <_sk_lerp_565_hsw+0xb7> + DB 196,65,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm8 DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 - DB 196,98,125,88,13,54,49,0,0 ; vpbroadcastd 0x3136(%rip),%ymm9 # 4bbc <_sk_callback_hsw+0x242> + DB 196,98,125,88,13,29,49,0,0 ; vpbroadcastd 0x311d(%rip),%ymm9 # 4bb8 <_sk_callback_hsw+0x24c> DB 196,65,61,219,201 ; vpand %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,39,49,0,0 ; vbroadcastss 0x3127(%rip),%ymm10 # 4bc0 <_sk_callback_hsw+0x246> + DB 196,98,125,24,21,14,49,0,0 ; vbroadcastss 0x310e(%rip),%ymm10 # 4bbc <_sk_callback_hsw+0x250> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,88,21,29,49,0,0 ; vpbroadcastd 0x311d(%rip),%ymm10 # 4bc4 <_sk_callback_hsw+0x24a> + DB 196,98,125,88,21,4,49,0,0 ; vpbroadcastd 0x3104(%rip),%ymm10 # 4bc0 <_sk_callback_hsw+0x254> DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,14,49,0,0 ; vbroadcastss 0x310e(%rip),%ymm11 # 4bc8 <_sk_callback_hsw+0x24e> + DB 196,98,125,24,29,245,48,0,0 ; vbroadcastss 0x30f5(%rip),%ymm11 # 4bc4 <_sk_callback_hsw+0x258> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,88,29,4,49,0,0 ; vpbroadcastd 0x3104(%rip),%ymm11 # 4bcc <_sk_callback_hsw+0x252> + DB 196,98,125,88,29,235,48,0,0 ; vpbroadcastd 0x30eb(%rip),%ymm11 # 4bc8 <_sk_callback_hsw+0x25c> DB 196,65,61,219,195 ; vpand %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,245,48,0,0 ; vbroadcastss 0x30f5(%rip),%ymm11 # 4bd0 <_sk_callback_hsw+0x256> + DB 196,98,125,24,29,220,48,0,0 ; vbroadcastss 0x30dc(%rip),%ymm11 # 4bcc <_sk_callback_hsw+0x260> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,226,53,168,196 ; vfmadd213ps %ymm4,%ymm9,%ymm0 @@ -1656,98 +1665,100 @@ _sk_lerp_565_hsw LABEL PROC DB 197,180,95,219 ; vmaxps %ymm3,%ymm9,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,68,255,255,255 ; ja 1a78 <_sk_lerp_565_hsw+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,77,0,0,0 ; lea 0x4d(%rip),%r9 # 1b8c <_sk_lerp_565_hsw+0x128> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 15,135,68,255,255,255 ; ja 1a8d <_sk_lerp_565_hsw+0x14> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,76,0,0,0 ; lea 0x4c(%rip),%r10 # 1ba0 <_sk_lerp_565_hsw+0x127> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 196,65,57,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 - DB 233,239,254,255,255 ; jmpq 1a78 <_sk_lerp_565_hsw+0x14> - DB 15,31,0 ; nopl (%rax) - DB 241 ; icebp + DB 196,65,57,196,68,83,12,6 ; vpinsrw $0x6,0xc(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,10,5 ; vpinsrw $0x5,0xa(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,8,4 ; vpinsrw $0x4,0x8(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,6,3 ; vpinsrw $0x3,0x6(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm8,%xmm8 + DB 233,239,254,255,255 ; jmpq 1a8d <_sk_lerp_565_hsw+0x14> + DB 102,144 ; xchg %ax,%ax + DB 242,255 ; repnz (bad) + DB 255 ; (bad) DB 255 ; (bad) + DB 234 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2001b94 <_sk_callback_hsw+0xffffffffe1ffd21a> + DB 255,226 ; jmpq *%rdx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 217,255 ; fcos + DB 218,255 ; (bad) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) - DB 255,201 ; dec %ecx + DB 255,202 ; dec %edx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 188 ; .byte 0xbc + DB 189 ; .byte 0xbd DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_tables_hsw _sk_load_tables_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 DB 77,133,192 ; test %r8,%r8 - DB 117,105 ; jne 1c26 <_sk_load_tables_hsw+0x7e> - DB 196,193,124,16,25 ; vmovups (%r9),%ymm3 - DB 197,228,84,13,54,51,0,0 ; vandps 0x3336(%rip),%ymm3,%ymm1 # 4f00 <_sk_callback_hsw+0x586> + DB 117,105 ; jne 1c3a <_sk_load_tables_hsw+0x7e> + DB 196,193,124,16,26 ; vmovups (%r10),%ymm3 + DB 197,228,84,13,34,51,0,0 ; vandps 0x3322(%rip),%ymm3,%ymm1 # 4f00 <_sk_callback_hsw+0x594> DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 DB 72,139,72,8 ; mov 0x8(%rax),%rcx - DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 DB 196,226,109,146,4,137 ; vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0 - DB 196,226,101,0,21,54,51,0,0 ; vpshufb 0x3336(%rip),%ymm3,%ymm2 # 4f20 <_sk_callback_hsw+0x5a6> + DB 196,226,101,0,21,34,51,0,0 ; vpshufb 0x3322(%rip),%ymm3,%ymm2 # 4f20 <_sk_callback_hsw+0x5b4> DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,194,53,146,12,145 ; vgatherdps %ymm9,(%r9,%ymm2,4),%ymm1 + DB 196,194,53,146,12,146 ; vgatherdps %ymm9,(%r10,%ymm2,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,98,101,0,13,62,51,0,0 ; vpshufb 0x333e(%rip),%ymm3,%ymm9 # 4f40 <_sk_callback_hsw+0x5c6> + DB 196,98,101,0,13,42,51,0,0 ; vpshufb 0x332a(%rip),%ymm3,%ymm9 # 4f40 <_sk_callback_hsw+0x5d4> DB 196,162,61,146,20,136 ; vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,186,47,0,0 ; vbroadcastss 0x2fba(%rip),%ymm8 # 4bd4 <_sk_callback_hsw+0x25a> + DB 196,98,125,24,5,162,47,0,0 ; vbroadcastss 0x2fa2(%rip),%ymm8 # 4bd0 <_sk_callback_hsw+0x264> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx DB 192,225,3 ; shl $0x3,%cl - DB 73,199,194,255,255,255,255 ; mov $0xffffffffffffffff,%r10 - DB 73,211,234 ; shr %cl,%r10 - DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 73,199,195,255,255,255,255 ; mov $0xffffffffffffffff,%r11 + DB 73,211,235 ; shr %cl,%r11 + DB 196,193,249,110,195 ; vmovq %r11,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 - DB 196,194,125,44,25 ; vmaskmovps (%r9),%ymm0,%ymm3 - DB 233,115,255,255,255 ; jmpq 1bc2 <_sk_load_tables_hsw+0x1a> + DB 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3 + DB 233,115,255,255,255 ; jmpq 1bd6 <_sk_load_tables_hsw+0x1a> PUBLIC _sk_load_tables_u16_be_hsw _sk_load_tables_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,201,0,0,0 ; jne 1d2e <_sk_load_tables_u16_be_hsw+0xdf> - DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8 - DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2 - DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3 - DB 196,1,122,111,76,72,48 ; vmovdqu 0x30(%r8,%r9,2),%xmm9 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,201,0,0,0 ; jne 1d42 <_sk_load_tables_u16_be_hsw+0xdf> + DB 196,1,121,16,4,81 ; vmovupd (%r9,%r10,2),%xmm8 + DB 196,129,121,16,84,81,16 ; vmovupd 0x10(%r9,%r10,2),%xmm2 + DB 196,129,121,16,92,81,32 ; vmovupd 0x20(%r9,%r10,2),%xmm3 + DB 196,1,122,111,76,81,48 ; vmovdqu 0x30(%r9,%r10,2),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 @@ -1759,18 +1770,18 @@ _sk_load_tables_u16_be_hsw LABEL PROC DB 197,185,108,200 ; vpunpcklqdq %xmm0,%xmm8,%xmm1 DB 197,185,109,208 ; vpunpckhqdq %xmm0,%xmm8,%xmm2 DB 197,49,108,195 ; vpunpcklqdq %xmm3,%xmm9,%xmm8 - DB 197,121,111,21,202,51,0,0 ; vmovdqa 0x33ca(%rip),%xmm10 # 5080 <_sk_callback_hsw+0x706> + DB 197,121,111,21,182,51,0,0 ; vmovdqa 0x33b6(%rip),%xmm10 # 5080 <_sk_callback_hsw+0x714> DB 196,193,113,219,194 ; vpand %xmm10,%xmm1,%xmm0 DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1 DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 - DB 196,194,29,146,4,136 ; vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0 + DB 196,194,29,146,4,137 ; vgatherdps %ymm12,(%r9,%ymm1,4),%ymm0 DB 196,193,105,219,202 ; vpand %xmm10,%xmm2,%xmm1 DB 196,226,125,51,209 ; vpmovzxwd %xmm1,%ymm2 DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 - DB 196,194,29,146,12,145 ; vgatherdps %ymm12,(%r9,%ymm2,4),%ymm1 + DB 196,194,29,146,12,146 ; vgatherdps %ymm12,(%r10,%ymm2,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax DB 196,193,57,219,210 ; vpand %xmm10,%xmm8,%xmm2 DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8 @@ -1781,48 +1792,48 @@ _sk_load_tables_u16_be_hsw LABEL PROC DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3 DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,179,46,0,0 ; vbroadcastss 0x2eb3(%rip),%ymm8 # 4bd8 <_sk_callback_hsw+0x25e> + DB 196,98,125,24,5,155,46,0,0 ; vbroadcastss 0x2e9b(%rip),%ymm8 # 4bd4 <_sk_callback_hsw+0x268> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8 + DB 196,1,123,16,4,81 ; vmovsd (%r9,%r10,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 1d94 <_sk_load_tables_u16_be_hsw+0x145> - DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 1d94 <_sk_load_tables_u16_be_hsw+0x145> - DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 1da1 <_sk_load_tables_u16_be_hsw+0x152> - DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 1da1 <_sk_load_tables_u16_be_hsw+0x152> - DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,9,255,255,255 ; je 1c80 <_sk_load_tables_u16_be_hsw+0x31> - DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,248,254,255,255 ; jb 1c80 <_sk_load_tables_u16_be_hsw+0x31> - DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9 - DB 233,236,254,255,255 ; jmpq 1c80 <_sk_load_tables_u16_be_hsw+0x31> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,85 ; je 1da8 <_sk_load_tables_u16_be_hsw+0x145> + DB 196,1,57,22,68,81,8 ; vmovhpd 0x8(%r9,%r10,2),%xmm8,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,72 ; jb 1da8 <_sk_load_tables_u16_be_hsw+0x145> + DB 196,129,123,16,84,81,16 ; vmovsd 0x10(%r9,%r10,2),%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 116,72 ; je 1db5 <_sk_load_tables_u16_be_hsw+0x152> + DB 196,129,105,22,84,81,24 ; vmovhpd 0x18(%r9,%r10,2),%xmm2,%xmm2 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,59 ; jb 1db5 <_sk_load_tables_u16_be_hsw+0x152> + DB 196,129,123,16,92,81,32 ; vmovsd 0x20(%r9,%r10,2),%xmm3 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 15,132,9,255,255,255 ; je 1c94 <_sk_load_tables_u16_be_hsw+0x31> + DB 196,129,97,22,92,81,40 ; vmovhpd 0x28(%r9,%r10,2),%xmm3,%xmm3 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 15,130,248,254,255,255 ; jb 1c94 <_sk_load_tables_u16_be_hsw+0x31> + DB 196,1,122,126,76,81,48 ; vmovq 0x30(%r9,%r10,2),%xmm9 + DB 233,236,254,255,255 ; jmpq 1c94 <_sk_load_tables_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,223,254,255,255 ; jmpq 1c80 <_sk_load_tables_u16_be_hsw+0x31> + DB 233,223,254,255,255 ; jmpq 1c94 <_sk_load_tables_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,214,254,255,255 ; jmpq 1c80 <_sk_load_tables_u16_be_hsw+0x31> + DB 233,214,254,255,255 ; jmpq 1c94 <_sk_load_tables_u16_be_hsw+0x31> PUBLIC _sk_load_tables_rgb_u16_be_hsw _sk_load_tables_rgb_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,193,0,0,0 ; jne 1e7d <_sk_load_tables_rgb_u16_be_hsw+0xd3> - DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0 - DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2 - DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1 - DB 196,129,122,111,92,72,32 ; vmovdqu 0x20(%r8,%r9,2),%xmm3 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,82 ; lea (%rdx,%rdx,2),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,193,0,0,0 ; jne 1e91 <_sk_load_tables_rgb_u16_be_hsw+0xd3> + DB 196,129,122,111,4,81 ; vmovdqu (%r9,%r10,2),%xmm0 + DB 196,129,122,111,84,81,12 ; vmovdqu 0xc(%r9,%r10,2),%xmm2 + DB 196,129,122,111,76,81,24 ; vmovdqu 0x18(%r9,%r10,2),%xmm1 + DB 196,129,122,111,92,81,32 ; vmovdqu 0x20(%r9,%r10,2),%xmm3 DB 197,225,115,219,4 ; vpsrldq $0x4,%xmm3,%xmm3 DB 197,185,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm8 DB 197,177,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm9 @@ -1839,344 +1850,314 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3 DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2 DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8 - DB 197,121,111,13,106,50,0,0 ; vmovdqa 0x326a(%rip),%xmm9 # 5090 <_sk_callback_hsw+0x716> + DB 197,121,111,13,86,50,0,0 ; vmovdqa 0x3256(%rip),%xmm9 # 5090 <_sk_callback_hsw+0x724> DB 196,193,97,219,193 ; vpand %xmm9,%xmm3,%xmm0 DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 - DB 196,194,45,146,4,136 ; vgatherdps %ymm10,(%r8,%ymm1,4),%ymm0 + DB 196,194,45,146,4,137 ; vgatherdps %ymm10,(%r9,%ymm1,4),%ymm0 DB 196,193,105,219,201 ; vpand %xmm9,%xmm2,%xmm1 DB 196,226,125,51,209 ; vpmovzxwd %xmm1,%ymm2 DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 - DB 196,194,45,146,12,145 ; vgatherdps %ymm10,(%r9,%ymm2,4),%ymm1 + DB 196,194,45,146,12,146 ; vgatherdps %ymm10,(%r10,%ymm2,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax DB 196,193,57,219,209 ; vpand %xmm9,%xmm8,%xmm2 DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8 DB 196,162,101,146,20,128 ; vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,97,45,0,0 ; vbroadcastss 0x2d61(%rip),%ymm3 # 4bdc <_sk_callback_hsw+0x262> - DB 255,224 ; jmpq *%rax - DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0 - DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 1e96 <_sk_load_tables_rgb_u16_be_hsw+0xec> - DB 233,90,255,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1 - DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 1ec5 <_sk_load_tables_rgb_u16_be_hsw+0x11b> - DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1 - DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 1eca <_sk_load_tables_rgb_u16_be_hsw+0x120> - DB 233,43,255,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 233,38,255,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1 - DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 1ef9 <_sk_load_tables_rgb_u16_be_hsw+0x14f> - DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1 - DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 1efe <_sk_load_tables_rgb_u16_be_hsw+0x154> - DB 233,247,254,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 233,242,254,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3 - DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 1f27 <_sk_load_tables_rgb_u16_be_hsw+0x17d> - DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3 - DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 - DB 233,201,254,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 233,196,254,255,255 ; jmpq 1df0 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 196,226,125,24,29,73,45,0,0 ; vbroadcastss 0x2d49(%rip),%ymm3 # 4bd8 <_sk_callback_hsw+0x26c> + DB 255,224 ; jmpq *%rax + DB 196,129,121,110,4,81 ; vmovd (%r9,%r10,2),%xmm0 + DB 196,129,121,196,68,81,4,2 ; vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,5 ; jne 1eaa <_sk_load_tables_rgb_u16_be_hsw+0xec> + DB 233,90,255,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 196,129,121,110,76,81,6 ; vmovd 0x6(%r9,%r10,2),%xmm1 + DB 196,1,113,196,68,81,10,2 ; vpinsrw $0x2,0xa(%r9,%r10,2),%xmm1,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,26 ; jb 1ed9 <_sk_load_tables_rgb_u16_be_hsw+0x11b> + DB 196,129,121,110,76,81,12 ; vmovd 0xc(%r9,%r10,2),%xmm1 + DB 196,129,113,196,84,81,16,2 ; vpinsrw $0x2,0x10(%r9,%r10,2),%xmm1,%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 117,10 ; jne 1ede <_sk_load_tables_rgb_u16_be_hsw+0x120> + DB 233,43,255,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,38,255,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 196,129,121,110,76,81,18 ; vmovd 0x12(%r9,%r10,2),%xmm1 + DB 196,1,113,196,76,81,22,2 ; vpinsrw $0x2,0x16(%r9,%r10,2),%xmm1,%xmm9 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,26 ; jb 1f0d <_sk_load_tables_rgb_u16_be_hsw+0x14f> + DB 196,129,121,110,76,81,24 ; vmovd 0x18(%r9,%r10,2),%xmm1 + DB 196,129,113,196,76,81,28,2 ; vpinsrw $0x2,0x1c(%r9,%r10,2),%xmm1,%xmm1 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 117,10 ; jne 1f12 <_sk_load_tables_rgb_u16_be_hsw+0x154> + DB 233,247,254,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,242,254,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 196,129,121,110,92,81,30 ; vmovd 0x1e(%r9,%r10,2),%xmm3 + DB 196,1,97,196,92,81,34,2 ; vpinsrw $0x2,0x22(%r9,%r10,2),%xmm3,%xmm11 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,20 ; jb 1f3b <_sk_load_tables_rgb_u16_be_hsw+0x17d> + DB 196,129,121,110,92,81,36 ; vmovd 0x24(%r9,%r10,2),%xmm3 + DB 196,129,97,196,92,81,40,2 ; vpinsrw $0x2,0x28(%r9,%r10,2),%xmm3,%xmm3 + DB 233,201,254,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,196,254,255,255 ; jmpq 1e04 <_sk_load_tables_rgb_u16_be_hsw+0x46> PUBLIC _sk_byte_tables_hsw _sk_byte_tables_hsw LABEL PROC - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,159,44,0,0 ; vbroadcastss 0x2c9f(%rip),%ymm8 # 4be0 <_sk_callback_hsw+0x266> + DB 196,98,125,24,5,145,44,0,0 ; vbroadcastss 0x2c91(%rip),%ymm8 # 4bdc <_sk_callback_hsw+0x270> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 68,137,197 ; mov %r8d,%ebp - DB 77,137,194 ; mov %r8,%r10 + DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 + DB 196,65,249,126,201 ; vmovq %xmm9,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 76,139,24 ; mov (%rax),%r11 + DB 196,131,121,32,4,19,0 ; vpinsrb $0x0,(%r11,%r10,1),%xmm0,%xmm0 + DB 196,67,249,22,202,1 ; vpextrq $0x1,%xmm9,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,121,32,20,11,1 ; vpinsrb $0x1,(%r11,%r9,1),%xmm0,%xmm10 + DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,199 ; mov %r8,%r15 - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 69,137,198 ; mov %r8d,%r14d - DB 77,137,196 ; mov %r8,%r12 - DB 73,193,236,32 ; shr $0x20,%r12 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,221 ; mov %ebx,%r13d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 76,139,8 ; mov (%rax),%r9 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 - DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp - DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp - DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,41 ; movzbl (%r9,%r13,1),%ebp - DB 196,227,121,32,197,4 ; vpinsrb $0x4,%ebp,%xmm0,%xmm0 - DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp - DB 196,227,121,32,197,5 ; vpinsrb $0x5,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp - DB 196,227,121,32,197,6 ; vpinsrb $0x6,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp - DB 196,227,121,32,197,7 ; vpinsrb $0x7,%ebp,%xmm0,%xmm0 - DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 - DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,240,43,0,0 ; vbroadcastss 0x2bf0(%rip),%ymm9 # 4be4 <_sk_callback_hsw+0x26a> - DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 196,99,125,57,200,1 ; vextracti128 $0x1,%ymm9,%xmm0 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,67,41,32,201,2 ; vpinsrb $0x2,%r9d,%xmm10,%xmm9 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,49,32,202,3 ; vpinsrb $0x3,%r10d,%xmm9,%xmm9 + DB 69,137,202 ; mov %r9d,%r10d + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,49,32,202,4 ; vpinsrb $0x4,%r10d,%xmm9,%xmm9 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,195,49,32,193,5 ; vpinsrb $0x5,%r9d,%xmm9,%xmm0 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,195,121,32,193,6 ; vpinsrb $0x6,%r9d,%xmm0,%xmm0 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,121,32,202,7 ; vpinsrb $0x7,%r10d,%xmm0,%xmm9 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 - DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,203 ; vmovq %xmm1,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,121,32,4,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm0,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,195,121,32,195,3 ; vpinsrb $0x3,%r11d,%xmm0,%xmm0 + DB 69,137,211 ; mov %r10d,%r11d + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,195,121,32,195,4 ; vpinsrb $0x4,%r11d,%xmm0,%xmm0 DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 - DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,5 ; vpinsrb $0x5,%r10d,%xmm0,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,193,249,126,207 ; vmovq %xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 - DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp - DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp - DB 196,227,113,32,205,4 ; vpinsrb $0x4,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,56 ; movzbl (%r8,%r15,1),%ebp - DB 196,227,113,32,205,5 ; vpinsrb $0x5,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp - DB 196,227,113,32,205,6 ; vpinsrb $0x6,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp - DB 196,227,113,32,205,7 ; vpinsrb $0x7,%ebp,%xmm1,%xmm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,6 ; vpinsrb $0x6,%r10d,%xmm0,%xmm0 + DB 71,15,182,12,25 ; movzbl (%r9,%r11,1),%r9d + DB 196,195,121,32,201,7 ; vpinsrb $0x7,%r9d,%xmm0,%xmm1 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 196,193,108,89,192 ; vmulps %ymm8,%ymm2,%ymm0 + DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,20,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm2 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,105,32,20,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm2,%xmm2 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,105,32,210,2 ; vpinsrb $0x2,%r10d,%xmm2,%xmm2 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,195,105,32,211,3 ; vpinsrb $0x3,%r11d,%xmm2,%xmm2 + DB 69,137,211 ; mov %r10d,%r11d + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,195,105,32,211,4 ; vpinsrb $0x4,%r11d,%xmm2,%xmm2 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,105,32,194,5 ; vpinsrb $0x5,%r10d,%xmm2,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,6 ; vpinsrb $0x6,%r10d,%xmm0,%xmm0 + DB 73,193,235,32 ; shr $0x20,%r11 + DB 71,15,182,12,25 ; movzbl (%r9,%r11,1),%r9d + DB 196,195,121,32,209,7 ; vpinsrb $0x7,%r9d,%xmm0,%xmm2 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 + DB 196,193,100,89,192 ; vmulps %ymm8,%ymm3,%ymm0 + DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 68,137,200 ; mov %r9d,%eax + DB 196,195,121,32,28,2,0 ; vpinsrb $0x0,(%r10,%rax,1),%xmm0,%xmm3 + DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,97,32,28,10,1 ; vpinsrb $0x1,(%r10,%r9,1),%xmm3,%xmm3 + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,195,97,32,217,2 ; vpinsrb $0x2,%r9d,%xmm3,%xmm3 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,227,97,32,216,3 ; vpinsrb $0x3,%eax,%xmm3,%xmm3 + DB 68,137,200 ; mov %r9d,%eax + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,227,97,32,216,4 ; vpinsrb $0x4,%eax,%xmm3,%xmm3 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 67,15,182,4,10 ; movzbl (%r10,%r9,1),%eax + DB 196,227,97,32,192,5 ; vpinsrb $0x5,%eax,%xmm3,%xmm0 + DB 68,137,216 ; mov %r11d,%eax + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,227,121,32,216,6 ; vpinsrb $0x6,%eax,%xmm0,%xmm3 + DB 73,193,235,32 ; shr $0x20,%r11 + DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax + DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0 + DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 + DB 196,98,125,24,5,54,42,0,0 ; vbroadcastss 0x2a36(%rip),%ymm8 # 4be0 <_sk_callback_hsw+0x274> + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 - DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 - DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,211 ; vmovq %xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,227,125,57,210,1 ; vextracti128 $0x1,%ymm2,%xmm2 - DB 196,195,249,22,211,1 ; vpextrq $0x1,%xmm2,%r11 - DB 69,137,222 ; mov %r11d,%r14d - DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,193,249,126,215 ; vmovq %xmm2,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 - DB 196,195,105,32,20,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2 - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp - DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2 - DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp - DB 196,227,105,32,213,4 ; vpinsrb $0x4,%ebp,%xmm2,%xmm2 - DB 67,15,182,44,56 ; movzbl (%r8,%r15,1),%ebp - DB 196,227,105,32,213,5 ; vpinsrb $0x5,%ebp,%xmm2,%xmm2 - DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp - DB 196,227,105,32,213,6 ; vpinsrb $0x6,%ebp,%xmm2,%xmm2 - DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp - DB 196,227,105,32,213,7 ; vpinsrb $0x7,%ebp,%xmm2,%xmm2 + DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 196,226,125,49,210 ; vpmovzxbd %xmm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 - DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 - DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 - DB 196,227,249,22,221,1 ; vpextrq $0x1,%xmm3,%rbp - DB 65,137,232 ; mov %ebp,%r8d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,219 ; vmovq %xmm3,%rbx - DB 65,137,217 ; mov %ebx,%r9d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,227,125,57,219,1 ; vextracti128 $0x1,%ymm3,%xmm3 - DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,222 ; vmovq %xmm3,%r14 - DB 69,137,247 ; mov %r14d,%r15d - DB 73,193,238,32 ; shr $0x20,%r14 - DB 196,163,121,32,28,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3 - DB 196,227,97,32,28,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3 - DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx - DB 196,227,97,32,219,2 ; vpinsrb $0x2,%ebx,%xmm3,%xmm3 - DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp - DB 196,227,97,32,221,3 ; vpinsrb $0x3,%ebp,%xmm3,%xmm3 - DB 66,15,182,44,56 ; movzbl (%rax,%r15,1),%ebp - DB 196,227,97,32,221,4 ; vpinsrb $0x4,%ebp,%xmm3,%xmm3 - DB 66,15,182,44,48 ; movzbl (%rax,%r14,1),%ebp - DB 196,227,97,32,221,5 ; vpinsrb $0x5,%ebp,%xmm3,%xmm3 - DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp - DB 196,227,97,32,221,6 ; vpinsrb $0x6,%ebp,%xmm3,%xmm3 - DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax + DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 196,227,97,32,216,7 ; vpinsrb $0x7,%eax,%xmm3,%xmm3 DB 196,226,125,49,219 ; vpmovzxbd %xmm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,193,100,89,217 ; vmulps %ymm9,%ymm3,%ymm3 + DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_byte_tables_rgb_hsw _sk_byte_tables_rgb_hsw LABEL PROC - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,139,64,24 ; mov 0x18(%rax),%r8d - DB 65,255,200 ; dec %r8d - DB 196,65,121,110,192 ; vmovd %r8d,%xmm8 + DB 68,139,72,24 ; mov 0x18(%rax),%r9d + DB 65,255,201 ; dec %r9d + DB 196,65,121,110,193 ; vmovd %r9d,%xmm8 DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 68,137,197 ; mov %r8d,%ebp - DB 77,137,194 ; mov %r8,%r10 + DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 + DB 196,65,249,126,201 ; vmovq %xmm9,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 76,139,24 ; mov (%rax),%r11 + DB 196,131,121,32,4,19,0 ; vpinsrb $0x0,(%r11,%r10,1),%xmm0,%xmm0 + DB 196,67,249,22,202,1 ; vpextrq $0x1,%xmm9,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,121,32,20,11,1 ; vpinsrb $0x1,(%r11,%r9,1),%xmm0,%xmm10 + DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,199 ; mov %r8,%r15 - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 69,137,198 ; mov %r8d,%r14d - DB 77,137,196 ; mov %r8,%r12 - DB 73,193,236,32 ; shr $0x20,%r12 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,221 ; mov %ebx,%r13d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 76,139,8 ; mov (%rax),%r9 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 - DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp - DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp - DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,41 ; movzbl (%r9,%r13,1),%ebp - DB 196,227,121,32,197,4 ; vpinsrb $0x4,%ebp,%xmm0,%xmm0 - DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp - DB 196,227,121,32,197,5 ; vpinsrb $0x5,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp - DB 196,227,121,32,197,6 ; vpinsrb $0x6,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp - DB 196,227,121,32,197,7 ; vpinsrb $0x7,%ebp,%xmm0,%xmm0 - DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 - DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,41,41,0,0 ; vbroadcastss 0x2929(%rip),%ymm9 # 4be8 <_sk_callback_hsw+0x26e> - DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 196,99,125,57,200,1 ; vextracti128 $0x1,%ymm9,%xmm0 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,67,41,32,201,2 ; vpinsrb $0x2,%r9d,%xmm10,%xmm9 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,49,32,202,3 ; vpinsrb $0x3,%r10d,%xmm9,%xmm9 + DB 69,137,202 ; mov %r9d,%r10d + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,49,32,202,4 ; vpinsrb $0x4,%r10d,%xmm9,%xmm9 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,195,49,32,193,5 ; vpinsrb $0x5,%r9d,%xmm9,%xmm0 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,195,121,32,193,6 ; vpinsrb $0x6,%r9d,%xmm0,%xmm0 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,121,32,202,7 ; vpinsrb $0x7,%r10d,%xmm0,%xmm9 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 - DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,203 ; vmovq %xmm1,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,121,32,4,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm0,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,195,121,32,195,3 ; vpinsrb $0x3,%r11d,%xmm0,%xmm0 + DB 69,137,211 ; mov %r10d,%r11d + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,195,121,32,195,4 ; vpinsrb $0x4,%r11d,%xmm0,%xmm0 DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 - DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,5 ; vpinsrb $0x5,%r10d,%xmm0,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,6 ; vpinsrb $0x6,%r10d,%xmm0,%xmm0 DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,193,249,126,207 ; vmovq %xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 - DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp - DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp - DB 196,227,113,32,205,4 ; vpinsrb $0x4,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,56 ; movzbl (%r8,%r15,1),%ebp - DB 196,227,113,32,205,5 ; vpinsrb $0x5,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp - DB 196,227,113,32,205,6 ; vpinsrb $0x6,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp - DB 196,227,113,32,205,7 ; vpinsrb $0x7,%ebp,%xmm1,%xmm1 + DB 71,15,182,12,25 ; movzbl (%r9,%r11,1),%r9d + DB 196,195,121,32,201,7 ; vpinsrb $0x7,%r9d,%xmm0,%xmm1 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 + DB 197,188,89,194 ; vmulps %ymm2,%ymm8,%ymm0 + DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 68,137,200 ; mov %r9d,%eax + DB 196,195,121,32,20,2,0 ; vpinsrb $0x0,(%r10,%rax,1),%xmm0,%xmm2 + DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,105,32,20,10,1 ; vpinsrb $0x1,(%r10,%r9,1),%xmm2,%xmm2 + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,195,105,32,209,2 ; vpinsrb $0x2,%r9d,%xmm2,%xmm2 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,227,105,32,208,3 ; vpinsrb $0x3,%eax,%xmm2,%xmm2 + DB 68,137,200 ; mov %r9d,%eax + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,227,105,32,208,4 ; vpinsrb $0x4,%eax,%xmm2,%xmm2 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 67,15,182,4,10 ; movzbl (%r10,%r9,1),%eax + DB 196,227,105,32,192,5 ; vpinsrb $0x5,%eax,%xmm2,%xmm0 + DB 68,137,216 ; mov %r11d,%eax + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,227,121,32,208,6 ; vpinsrb $0x6,%eax,%xmm0,%xmm2 + DB 73,193,235,32 ; shr $0x20,%r11 + DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax + DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0 + DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 + DB 196,98,125,24,5,34,40,0,0 ; vbroadcastss 0x2822(%rip),%ymm8 # 4be4 <_sk_callback_hsw+0x278> + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 - DB 72,139,64,16 ; mov 0x10(%rax),%rax - DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 - DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 - DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp - DB 65,137,232 ; mov %ebp,%r8d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,211 ; vmovq %xmm2,%rbx - DB 65,137,217 ; mov %ebx,%r9d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,227,125,57,210,1 ; vextracti128 $0x1,%ymm2,%xmm2 - DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,214 ; vmovq %xmm2,%r14 - DB 69,137,247 ; mov %r14d,%r15d - DB 73,193,238,32 ; shr $0x20,%r14 - DB 196,163,121,32,20,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2 - DB 196,227,105,32,20,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2 - DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx - DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2 - DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp - DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2 - DB 66,15,182,44,56 ; movzbl (%rax,%r15,1),%ebp - DB 196,227,105,32,213,4 ; vpinsrb $0x4,%ebp,%xmm2,%xmm2 - DB 66,15,182,44,48 ; movzbl (%rax,%r14,1),%ebp - DB 196,227,105,32,213,5 ; vpinsrb $0x5,%ebp,%xmm2,%xmm2 - DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp - DB 196,227,105,32,213,6 ; vpinsrb $0x6,%ebp,%xmm2,%xmm2 - DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax + DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 196,227,105,32,208,7 ; vpinsrb $0x7,%eax,%xmm2,%xmm2 DB 196,226,125,49,210 ; vpmovzxbd %xmm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_table_r_hsw _sk_table_r_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -2185,14 +2166,14 @@ _sk_table_r_hsw LABEL PROC DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,125,91,192 ; vcvtps2dq %ymm0,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,130,53,146,4,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm0 + DB 196,130,53,146,4,129 ; vgatherdps %ymm9,(%r9,%ymm8,4),%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_table_g_hsw _sk_table_g_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -2201,14 +2182,14 @@ _sk_table_g_hsw LABEL PROC DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,125,91,193 ; vcvtps2dq %ymm1,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,130,53,146,12,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm1 + DB 196,130,53,146,12,129 ; vgatherdps %ymm9,(%r9,%ymm8,4),%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_table_b_hsw _sk_table_b_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -2217,14 +2198,14 @@ _sk_table_b_hsw LABEL PROC DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 DB 197,125,91,194 ; vcvtps2dq %ymm2,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,130,53,146,20,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm2 + DB 196,130,53,146,20,129 ; vgatherdps %ymm9,(%r9,%ymm8,4),%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_table_a_hsw _sk_table_a_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -2233,7 +2214,7 @@ _sk_table_a_hsw LABEL PROC DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3 DB 197,125,91,195 ; vcvtps2dq %ymm3,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,130,53,146,28,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm3 + DB 196,130,53,146,28,129 ; vgatherdps %ymm9,(%r9,%ymm8,4),%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2250,33 +2231,33 @@ _sk_parametric_r_hsw LABEL PROC DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10 DB 196,226,125,24,0 ; vbroadcastss (%rax),%ymm0 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,220,38,0,0 ; vbroadcastss 0x26dc(%rip),%ymm12 # 4bec <_sk_callback_hsw+0x272> - DB 196,98,125,24,45,215,38,0,0 ; vbroadcastss 0x26d7(%rip),%ymm13 # 4bf0 <_sk_callback_hsw+0x276> + DB 196,98,125,24,37,252,38,0,0 ; vbroadcastss 0x26fc(%rip),%ymm12 # 4be8 <_sk_callback_hsw+0x27c> + DB 196,98,125,24,45,247,38,0,0 ; vbroadcastss 0x26f7(%rip),%ymm13 # 4bec <_sk_callback_hsw+0x280> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,205,38,0,0 ; vbroadcastss 0x26cd(%rip),%ymm13 # 4bf4 <_sk_callback_hsw+0x27a> + DB 196,98,125,24,45,237,38,0,0 ; vbroadcastss 0x26ed(%rip),%ymm13 # 4bf0 <_sk_callback_hsw+0x284> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,195,38,0,0 ; vbroadcastss 0x26c3(%rip),%ymm13 # 4bf8 <_sk_callback_hsw+0x27e> + DB 196,98,125,24,45,227,38,0,0 ; vbroadcastss 0x26e3(%rip),%ymm13 # 4bf4 <_sk_callback_hsw+0x288> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,185,38,0,0 ; vbroadcastss 0x26b9(%rip),%ymm11 # 4bfc <_sk_callback_hsw+0x282> + DB 196,98,125,24,29,217,38,0,0 ; vbroadcastss 0x26d9(%rip),%ymm11 # 4bf8 <_sk_callback_hsw+0x28c> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,175,38,0,0 ; vbroadcastss 0x26af(%rip),%ymm12 # 4c00 <_sk_callback_hsw+0x286> + DB 196,98,125,24,37,207,38,0,0 ; vbroadcastss 0x26cf(%rip),%ymm12 # 4bfc <_sk_callback_hsw+0x290> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,165,38,0,0 ; vbroadcastss 0x26a5(%rip),%ymm12 # 4c04 <_sk_callback_hsw+0x28a> + DB 196,98,125,24,37,197,38,0,0 ; vbroadcastss 0x26c5(%rip),%ymm12 # 4c00 <_sk_callback_hsw+0x294> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,134,38,0,0 ; vbroadcastss 0x2686(%rip),%ymm11 # 4c08 <_sk_callback_hsw+0x28e> + DB 196,98,125,24,29,166,38,0,0 ; vbroadcastss 0x26a6(%rip),%ymm11 # 4c04 <_sk_callback_hsw+0x298> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,124,38,0,0 ; vbroadcastss 0x267c(%rip),%ymm11 # 4c0c <_sk_callback_hsw+0x292> + DB 196,98,125,24,29,156,38,0,0 ; vbroadcastss 0x269c(%rip),%ymm11 # 4c08 <_sk_callback_hsw+0x29c> DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11 - DB 196,226,125,24,5,114,38,0,0 ; vbroadcastss 0x2672(%rip),%ymm0 # 4c10 <_sk_callback_hsw+0x296> + DB 196,226,125,24,5,146,38,0,0 ; vbroadcastss 0x2692(%rip),%ymm0 # 4c0c <_sk_callback_hsw+0x2a0> DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,104,38,0,0 ; vbroadcastss 0x2668(%rip),%ymm10 # 4c14 <_sk_callback_hsw+0x29a> + DB 196,98,125,24,21,136,38,0,0 ; vbroadcastss 0x2688(%rip),%ymm10 # 4c10 <_sk_callback_hsw+0x2a4> DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0 DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,21,91,38,0,0 ; vbroadcastss 0x265b(%rip),%ymm10 # 4c18 <_sk_callback_hsw+0x29e> + DB 196,98,125,24,21,123,38,0,0 ; vbroadcastss 0x267b(%rip),%ymm10 # 4c14 <_sk_callback_hsw+0x2a8> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2284,7 +2265,7 @@ _sk_parametric_r_hsw LABEL PROC DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,50,38,0,0 ; vbroadcastss 0x2632(%rip),%ymm8 # 4c1c <_sk_callback_hsw+0x2a2> + DB 196,98,125,24,5,82,38,0,0 ; vbroadcastss 0x2652(%rip),%ymm8 # 4c18 <_sk_callback_hsw+0x2ac> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2302,33 +2283,33 @@ _sk_parametric_g_hsw LABEL PROC DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10 DB 196,226,125,24,8 ; vbroadcastss (%rax),%ymm1 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,234,37,0,0 ; vbroadcastss 0x25ea(%rip),%ymm12 # 4c20 <_sk_callback_hsw+0x2a6> - DB 196,98,125,24,45,229,37,0,0 ; vbroadcastss 0x25e5(%rip),%ymm13 # 4c24 <_sk_callback_hsw+0x2aa> + DB 196,98,125,24,37,10,38,0,0 ; vbroadcastss 0x260a(%rip),%ymm12 # 4c1c <_sk_callback_hsw+0x2b0> + DB 196,98,125,24,45,5,38,0,0 ; vbroadcastss 0x2605(%rip),%ymm13 # 4c20 <_sk_callback_hsw+0x2b4> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,219,37,0,0 ; vbroadcastss 0x25db(%rip),%ymm13 # 4c28 <_sk_callback_hsw+0x2ae> + DB 196,98,125,24,45,251,37,0,0 ; vbroadcastss 0x25fb(%rip),%ymm13 # 4c24 <_sk_callback_hsw+0x2b8> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,209,37,0,0 ; vbroadcastss 0x25d1(%rip),%ymm13 # 4c2c <_sk_callback_hsw+0x2b2> + DB 196,98,125,24,45,241,37,0,0 ; vbroadcastss 0x25f1(%rip),%ymm13 # 4c28 <_sk_callback_hsw+0x2bc> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,199,37,0,0 ; vbroadcastss 0x25c7(%rip),%ymm11 # 4c30 <_sk_callback_hsw+0x2b6> + DB 196,98,125,24,29,231,37,0,0 ; vbroadcastss 0x25e7(%rip),%ymm11 # 4c2c <_sk_callback_hsw+0x2c0> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,189,37,0,0 ; vbroadcastss 0x25bd(%rip),%ymm12 # 4c34 <_sk_callback_hsw+0x2ba> + DB 196,98,125,24,37,221,37,0,0 ; vbroadcastss 0x25dd(%rip),%ymm12 # 4c30 <_sk_callback_hsw+0x2c4> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,179,37,0,0 ; vbroadcastss 0x25b3(%rip),%ymm12 # 4c38 <_sk_callback_hsw+0x2be> + DB 196,98,125,24,37,211,37,0,0 ; vbroadcastss 0x25d3(%rip),%ymm12 # 4c34 <_sk_callback_hsw+0x2c8> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,148,37,0,0 ; vbroadcastss 0x2594(%rip),%ymm11 # 4c3c <_sk_callback_hsw+0x2c2> + DB 196,98,125,24,29,180,37,0,0 ; vbroadcastss 0x25b4(%rip),%ymm11 # 4c38 <_sk_callback_hsw+0x2cc> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,138,37,0,0 ; vbroadcastss 0x258a(%rip),%ymm11 # 4c40 <_sk_callback_hsw+0x2c6> + DB 196,98,125,24,29,170,37,0,0 ; vbroadcastss 0x25aa(%rip),%ymm11 # 4c3c <_sk_callback_hsw+0x2d0> DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11 - DB 196,226,125,24,13,128,37,0,0 ; vbroadcastss 0x2580(%rip),%ymm1 # 4c44 <_sk_callback_hsw+0x2ca> + DB 196,226,125,24,13,160,37,0,0 ; vbroadcastss 0x25a0(%rip),%ymm1 # 4c40 <_sk_callback_hsw+0x2d4> DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,118,37,0,0 ; vbroadcastss 0x2576(%rip),%ymm10 # 4c48 <_sk_callback_hsw+0x2ce> + DB 196,98,125,24,21,150,37,0,0 ; vbroadcastss 0x2596(%rip),%ymm10 # 4c44 <_sk_callback_hsw+0x2d8> DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 196,98,125,24,21,105,37,0,0 ; vbroadcastss 0x2569(%rip),%ymm10 # 4c4c <_sk_callback_hsw+0x2d2> + DB 196,98,125,24,21,137,37,0,0 ; vbroadcastss 0x2589(%rip),%ymm10 # 4c48 <_sk_callback_hsw+0x2dc> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2336,7 +2317,7 @@ _sk_parametric_g_hsw LABEL PROC DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,64,37,0,0 ; vbroadcastss 0x2540(%rip),%ymm8 # 4c50 <_sk_callback_hsw+0x2d6> + DB 196,98,125,24,5,96,37,0,0 ; vbroadcastss 0x2560(%rip),%ymm8 # 4c4c <_sk_callback_hsw+0x2e0> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2354,33 +2335,33 @@ _sk_parametric_b_hsw LABEL PROC DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10 DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,248,36,0,0 ; vbroadcastss 0x24f8(%rip),%ymm12 # 4c54 <_sk_callback_hsw+0x2da> - DB 196,98,125,24,45,243,36,0,0 ; vbroadcastss 0x24f3(%rip),%ymm13 # 4c58 <_sk_callback_hsw+0x2de> + DB 196,98,125,24,37,24,37,0,0 ; vbroadcastss 0x2518(%rip),%ymm12 # 4c50 <_sk_callback_hsw+0x2e4> + DB 196,98,125,24,45,19,37,0,0 ; vbroadcastss 0x2513(%rip),%ymm13 # 4c54 <_sk_callback_hsw+0x2e8> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,233,36,0,0 ; vbroadcastss 0x24e9(%rip),%ymm13 # 4c5c <_sk_callback_hsw+0x2e2> + DB 196,98,125,24,45,9,37,0,0 ; vbroadcastss 0x2509(%rip),%ymm13 # 4c58 <_sk_callback_hsw+0x2ec> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,223,36,0,0 ; vbroadcastss 0x24df(%rip),%ymm13 # 4c60 <_sk_callback_hsw+0x2e6> + DB 196,98,125,24,45,255,36,0,0 ; vbroadcastss 0x24ff(%rip),%ymm13 # 4c5c <_sk_callback_hsw+0x2f0> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,213,36,0,0 ; vbroadcastss 0x24d5(%rip),%ymm11 # 4c64 <_sk_callback_hsw+0x2ea> + DB 196,98,125,24,29,245,36,0,0 ; vbroadcastss 0x24f5(%rip),%ymm11 # 4c60 <_sk_callback_hsw+0x2f4> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,203,36,0,0 ; vbroadcastss 0x24cb(%rip),%ymm12 # 4c68 <_sk_callback_hsw+0x2ee> + DB 196,98,125,24,37,235,36,0,0 ; vbroadcastss 0x24eb(%rip),%ymm12 # 4c64 <_sk_callback_hsw+0x2f8> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,193,36,0,0 ; vbroadcastss 0x24c1(%rip),%ymm12 # 4c6c <_sk_callback_hsw+0x2f2> + DB 196,98,125,24,37,225,36,0,0 ; vbroadcastss 0x24e1(%rip),%ymm12 # 4c68 <_sk_callback_hsw+0x2fc> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,162,36,0,0 ; vbroadcastss 0x24a2(%rip),%ymm11 # 4c70 <_sk_callback_hsw+0x2f6> + DB 196,98,125,24,29,194,36,0,0 ; vbroadcastss 0x24c2(%rip),%ymm11 # 4c6c <_sk_callback_hsw+0x300> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,152,36,0,0 ; vbroadcastss 0x2498(%rip),%ymm11 # 4c74 <_sk_callback_hsw+0x2fa> + DB 196,98,125,24,29,184,36,0,0 ; vbroadcastss 0x24b8(%rip),%ymm11 # 4c70 <_sk_callback_hsw+0x304> DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11 - DB 196,226,125,24,21,142,36,0,0 ; vbroadcastss 0x248e(%rip),%ymm2 # 4c78 <_sk_callback_hsw+0x2fe> + DB 196,226,125,24,21,174,36,0,0 ; vbroadcastss 0x24ae(%rip),%ymm2 # 4c74 <_sk_callback_hsw+0x308> DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,132,36,0,0 ; vbroadcastss 0x2484(%rip),%ymm10 # 4c7c <_sk_callback_hsw+0x302> + DB 196,98,125,24,21,164,36,0,0 ; vbroadcastss 0x24a4(%rip),%ymm10 # 4c78 <_sk_callback_hsw+0x30c> DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2 DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 196,98,125,24,21,119,36,0,0 ; vbroadcastss 0x2477(%rip),%ymm10 # 4c80 <_sk_callback_hsw+0x306> + DB 196,98,125,24,21,151,36,0,0 ; vbroadcastss 0x2497(%rip),%ymm10 # 4c7c <_sk_callback_hsw+0x310> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2388,7 +2369,7 @@ _sk_parametric_b_hsw LABEL PROC DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,78,36,0,0 ; vbroadcastss 0x244e(%rip),%ymm8 # 4c84 <_sk_callback_hsw+0x30a> + DB 196,98,125,24,5,110,36,0,0 ; vbroadcastss 0x246e(%rip),%ymm8 # 4c80 <_sk_callback_hsw+0x314> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2406,33 +2387,33 @@ _sk_parametric_a_hsw LABEL PROC DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10 DB 196,226,125,24,24 ; vbroadcastss (%rax),%ymm3 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,6,36,0,0 ; vbroadcastss 0x2406(%rip),%ymm12 # 4c88 <_sk_callback_hsw+0x30e> - DB 196,98,125,24,45,1,36,0,0 ; vbroadcastss 0x2401(%rip),%ymm13 # 4c8c <_sk_callback_hsw+0x312> + DB 196,98,125,24,37,38,36,0,0 ; vbroadcastss 0x2426(%rip),%ymm12 # 4c84 <_sk_callback_hsw+0x318> + DB 196,98,125,24,45,33,36,0,0 ; vbroadcastss 0x2421(%rip),%ymm13 # 4c88 <_sk_callback_hsw+0x31c> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,247,35,0,0 ; vbroadcastss 0x23f7(%rip),%ymm13 # 4c90 <_sk_callback_hsw+0x316> + DB 196,98,125,24,45,23,36,0,0 ; vbroadcastss 0x2417(%rip),%ymm13 # 4c8c <_sk_callback_hsw+0x320> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,237,35,0,0 ; vbroadcastss 0x23ed(%rip),%ymm13 # 4c94 <_sk_callback_hsw+0x31a> + DB 196,98,125,24,45,13,36,0,0 ; vbroadcastss 0x240d(%rip),%ymm13 # 4c90 <_sk_callback_hsw+0x324> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,227,35,0,0 ; vbroadcastss 0x23e3(%rip),%ymm11 # 4c98 <_sk_callback_hsw+0x31e> + DB 196,98,125,24,29,3,36,0,0 ; vbroadcastss 0x2403(%rip),%ymm11 # 4c94 <_sk_callback_hsw+0x328> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,217,35,0,0 ; vbroadcastss 0x23d9(%rip),%ymm12 # 4c9c <_sk_callback_hsw+0x322> + DB 196,98,125,24,37,249,35,0,0 ; vbroadcastss 0x23f9(%rip),%ymm12 # 4c98 <_sk_callback_hsw+0x32c> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,207,35,0,0 ; vbroadcastss 0x23cf(%rip),%ymm12 # 4ca0 <_sk_callback_hsw+0x326> + DB 196,98,125,24,37,239,35,0,0 ; vbroadcastss 0x23ef(%rip),%ymm12 # 4c9c <_sk_callback_hsw+0x330> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,176,35,0,0 ; vbroadcastss 0x23b0(%rip),%ymm11 # 4ca4 <_sk_callback_hsw+0x32a> + DB 196,98,125,24,29,208,35,0,0 ; vbroadcastss 0x23d0(%rip),%ymm11 # 4ca0 <_sk_callback_hsw+0x334> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,166,35,0,0 ; vbroadcastss 0x23a6(%rip),%ymm11 # 4ca8 <_sk_callback_hsw+0x32e> + DB 196,98,125,24,29,198,35,0,0 ; vbroadcastss 0x23c6(%rip),%ymm11 # 4ca4 <_sk_callback_hsw+0x338> DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11 - DB 196,226,125,24,29,156,35,0,0 ; vbroadcastss 0x239c(%rip),%ymm3 # 4cac <_sk_callback_hsw+0x332> + DB 196,226,125,24,29,188,35,0,0 ; vbroadcastss 0x23bc(%rip),%ymm3 # 4ca8 <_sk_callback_hsw+0x33c> DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,146,35,0,0 ; vbroadcastss 0x2392(%rip),%ymm10 # 4cb0 <_sk_callback_hsw+0x336> + DB 196,98,125,24,21,178,35,0,0 ; vbroadcastss 0x23b2(%rip),%ymm10 # 4cac <_sk_callback_hsw+0x340> DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3 DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 196,98,125,24,21,133,35,0,0 ; vbroadcastss 0x2385(%rip),%ymm10 # 4cb4 <_sk_callback_hsw+0x33a> + DB 196,98,125,24,21,165,35,0,0 ; vbroadcastss 0x23a5(%rip),%ymm10 # 4cb0 <_sk_callback_hsw+0x344> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2440,33 +2421,33 @@ _sk_parametric_a_hsw LABEL PROC DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,92,35,0,0 ; vbroadcastss 0x235c(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x33e> + DB 196,98,125,24,5,124,35,0,0 ; vbroadcastss 0x237c(%rip),%ymm8 # 4cb4 <_sk_callback_hsw+0x348> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_lab_to_xyz_hsw _sk_lab_to_xyz_hsw LABEL PROC - DB 196,98,125,24,5,78,35,0,0 ; vbroadcastss 0x234e(%rip),%ymm8 # 4cbc <_sk_callback_hsw+0x342> - DB 196,98,125,24,13,73,35,0,0 ; vbroadcastss 0x2349(%rip),%ymm9 # 4cc0 <_sk_callback_hsw+0x346> - DB 196,98,125,24,21,68,35,0,0 ; vbroadcastss 0x2344(%rip),%ymm10 # 4cc4 <_sk_callback_hsw+0x34a> + DB 196,98,125,24,5,110,35,0,0 ; vbroadcastss 0x236e(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x34c> + DB 196,98,125,24,13,105,35,0,0 ; vbroadcastss 0x2369(%rip),%ymm9 # 4cbc <_sk_callback_hsw+0x350> + DB 196,98,125,24,21,100,35,0,0 ; vbroadcastss 0x2364(%rip),%ymm10 # 4cc0 <_sk_callback_hsw+0x354> DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1 DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2 - DB 196,98,125,24,13,53,35,0,0 ; vbroadcastss 0x2335(%rip),%ymm9 # 4cc8 <_sk_callback_hsw+0x34e> + DB 196,98,125,24,13,85,35,0,0 ; vbroadcastss 0x2355(%rip),%ymm9 # 4cc4 <_sk_callback_hsw+0x358> DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9 - DB 196,226,125,24,5,43,35,0,0 ; vbroadcastss 0x232b(%rip),%ymm0 # 4ccc <_sk_callback_hsw+0x352> + DB 196,226,125,24,5,75,35,0,0 ; vbroadcastss 0x234b(%rip),%ymm0 # 4cc8 <_sk_callback_hsw+0x35c> DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 - DB 196,98,125,24,5,34,35,0,0 ; vbroadcastss 0x2322(%rip),%ymm8 # 4cd0 <_sk_callback_hsw+0x356> + DB 196,98,125,24,5,66,35,0,0 ; vbroadcastss 0x2342(%rip),%ymm8 # 4ccc <_sk_callback_hsw+0x360> DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8 - DB 196,98,125,24,13,24,35,0,0 ; vbroadcastss 0x2318(%rip),%ymm9 # 4cd4 <_sk_callback_hsw+0x35a> + DB 196,98,125,24,13,56,35,0,0 ; vbroadcastss 0x2338(%rip),%ymm9 # 4cd0 <_sk_callback_hsw+0x364> DB 196,98,109,172,200 ; vfnmadd213ps %ymm0,%ymm2,%ymm9 DB 196,193,60,89,200 ; vmulps %ymm8,%ymm8,%ymm1 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 - DB 196,226,125,24,21,5,35,0,0 ; vbroadcastss 0x2305(%rip),%ymm2 # 4cd8 <_sk_callback_hsw+0x35e> + DB 196,226,125,24,21,37,35,0,0 ; vbroadcastss 0x2325(%rip),%ymm2 # 4cd4 <_sk_callback_hsw+0x368> DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10 - DB 196,98,125,24,29,251,34,0,0 ; vbroadcastss 0x22fb(%rip),%ymm11 # 4cdc <_sk_callback_hsw+0x362> + DB 196,98,125,24,29,27,35,0,0 ; vbroadcastss 0x231b(%rip),%ymm11 # 4cd8 <_sk_callback_hsw+0x36c> DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,37,241,34,0,0 ; vbroadcastss 0x22f1(%rip),%ymm12 # 4ce0 <_sk_callback_hsw+0x366> + DB 196,98,125,24,37,17,35,0,0 ; vbroadcastss 0x2311(%rip),%ymm12 # 4cdc <_sk_callback_hsw+0x370> DB 196,65,60,89,196 ; vmulps %ymm12,%ymm8,%ymm8 DB 196,99,61,74,193,160 ; vblendvps %ymm10,%ymm1,%ymm8,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -2481,88 +2462,91 @@ _sk_lab_to_xyz_hsw LABEL PROC DB 196,65,52,88,203 ; vaddps %ymm11,%ymm9,%ymm9 DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 DB 196,227,53,74,208,32 ; vblendvps %ymm2,%ymm0,%ymm9,%ymm2 - DB 196,226,125,24,5,166,34,0,0 ; vbroadcastss 0x22a6(%rip),%ymm0 # 4ce4 <_sk_callback_hsw+0x36a> + DB 196,226,125,24,5,198,34,0,0 ; vbroadcastss 0x22c6(%rip),%ymm0 # 4ce0 <_sk_callback_hsw+0x374> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,157,34,0,0 ; vbroadcastss 0x229d(%rip),%ymm8 # 4ce8 <_sk_callback_hsw+0x36e> + DB 196,98,125,24,5,189,34,0,0 ; vbroadcastss 0x22bd(%rip),%ymm8 # 4ce4 <_sk_callback_hsw+0x378> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_load_a8_hsw _sk_load_a8_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,45 ; jne 2a91 <_sk_load_a8_hsw+0x3d> + DB 117,45 ; jne 2a6d <_sk_load_a8_hsw+0x3d> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,114,34,0,0 ; vbroadcastss 0x2272(%rip),%ymm1 # 4cec <_sk_callback_hsw+0x372> + DB 196,226,125,24,13,146,34,0,0 ; vbroadcastss 0x2292(%rip),%ymm1 # 4ce8 <_sk_callback_hsw+0x37c> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2a99 <_sk_load_a8_hsw+0x45> - DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,178 ; jmp 2a68 <_sk_load_a8_hsw+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 2a76 <_sk_load_a8_hsw+0x46> + DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 91 ; pop %rbx + DB 235,177 ; jmp 2a44 <_sk_load_a8_hsw+0x14> PUBLIC _sk_gather_a8_hsw _sk_gather_a8_hsw LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,199 ; vmovq %xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,4,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0 - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 196,193,249,126,196 ; vmovq %xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,131,121,32,4,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 + DB 196,131,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,227,121,32,192,3 ; vpinsrb $0x3,%eax,%xmm0,%xmm0 - DB 67,15,182,4,32 ; movzbl (%r8,%r12,1),%eax + DB 67,15,182,4,41 ; movzbl (%r9,%r13,1),%eax DB 196,227,121,32,192,4 ; vpinsrb $0x4,%eax,%xmm0,%xmm0 - DB 67,15,182,4,56 ; movzbl (%r8,%r15,1),%eax + DB 67,15,182,4,33 ; movzbl (%r9,%r12,1),%eax DB 196,227,121,32,192,5 ; vpinsrb $0x5,%eax,%xmm0,%xmm0 - DB 67,15,182,4,48 ; movzbl (%r8,%r14,1),%eax + DB 67,15,182,4,57 ; movzbl (%r9,%r15,1),%eax DB 196,227,121,32,192,6 ; vpinsrb $0x6,%eax,%xmm0,%xmm0 - DB 65,15,182,4,24 ; movzbl (%r8,%rbx,1),%eax + DB 65,15,182,4,25 ; movzbl (%r9,%rbx,1),%eax DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,125,33,0,0 ; vbroadcastss 0x217d(%rip),%ymm1 # 4cf0 <_sk_callback_hsw+0x376> + DB 196,226,125,24,13,154,33,0,0 ; vbroadcastss 0x219a(%rip),%ymm1 # 4cec <_sk_callback_hsw+0x380> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -2570,6 +2554,7 @@ _sk_gather_a8_hsw LABEL PROC DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -2577,143 +2562,146 @@ _sk_gather_a8_hsw LABEL PROC PUBLIC _sk_store_a8_hsw _sk_store_a8_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,88,33,0,0 ; vbroadcastss 0x2158(%rip),%ymm8 # 4cf4 <_sk_callback_hsw+0x37a> + DB 76,139,24 ; mov (%rax),%r11 + DB 196,98,125,24,5,115,33,0,0 ; vbroadcastss 0x2173(%rip),%ymm8 # 4cf0 <_sk_callback_hsw+0x384> DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8 DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 2bc5 <_sk_store_a8_hsw+0x37> - DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1) + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 2ba6 <_sk_store_a8_hsw+0x37> + DB 196,65,123,17,4,19 ; vmovsd %xmm8,(%r11,%rdx,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 2bc1 <_sk_store_a8_hsw+0x33> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 2ba2 <_sk_store_a8_hsw+0x33> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 2c28 <_sk_store_a8_hsw+0x9a> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,121,20,68,58,6,12 ; vpextrb $0xc,%xmm8,0x6(%r10,%rdi,1) - DB 196,67,121,20,68,58,5,10 ; vpextrb $0xa,%xmm8,0x5(%r10,%rdi,1) - DB 196,67,121,20,68,58,4,8 ; vpextrb $0x8,%xmm8,0x4(%r10,%rdi,1) - DB 196,67,121,20,68,58,3,6 ; vpextrb $0x6,%xmm8,0x3(%r10,%rdi,1) - DB 196,67,121,20,68,58,2,4 ; vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1) - DB 196,67,121,20,68,58,1,2 ; vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1) - DB 196,67,121,20,4,58,0 ; vpextrb $0x0,%xmm8,(%r10,%rdi,1) - DB 235,154 ; jmp 2bc1 <_sk_store_a8_hsw+0x33> - DB 144 ; nop - DB 246,255 ; idiv %bh + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 2c08 <_sk_store_a8_hsw+0x99> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,67,121,20,68,19,6,12 ; vpextrb $0xc,%xmm8,0x6(%r11,%rdx,1) + DB 196,67,121,20,68,19,5,10 ; vpextrb $0xa,%xmm8,0x5(%r11,%rdx,1) + DB 196,67,121,20,68,19,4,8 ; vpextrb $0x8,%xmm8,0x4(%r11,%rdx,1) + DB 196,67,121,20,68,19,3,6 ; vpextrb $0x6,%xmm8,0x3(%r11,%rdx,1) + DB 196,67,121,20,68,19,2,4 ; vpextrb $0x4,%xmm8,0x2(%r11,%rdx,1) + DB 196,67,121,20,68,19,1,2 ; vpextrb $0x2,%xmm8,0x1(%r11,%rdx,1) + DB 196,67,121,20,4,19,0 ; vpextrb $0x0,%xmm8,(%r11,%rdx,1) + DB 235,154 ; jmp 2ba2 <_sk_store_a8_hsw+0x33> + DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) - DB 238 ; out %al,(%dx) + DB 239 ; out %eax,(%dx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) + DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,214 ; callq *%rsi + DB 255,215 ; callq *%rdi DB 255 ; (bad) DB 255 ; (bad) - DB 255,206 ; dec %esi + DB 255,207 ; dec %edi DB 255 ; (bad) DB 255 ; (bad) - DB 255,198 ; inc %esi + DB 255,199 ; inc %edi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_g8_hsw _sk_load_g8_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,50 ; jne 2c86 <_sk_load_g8_hsw+0x42> + DB 117,50 ; jne 2c66 <_sk_load_g8_hsw+0x42> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,142,32,0,0 ; vbroadcastss 0x208e(%rip),%ymm1 # 4cf8 <_sk_callback_hsw+0x37e> + DB 196,226,125,24,13,170,32,0,0 ; vbroadcastss 0x20aa(%rip),%ymm1 # 4cf4 <_sk_callback_hsw+0x388> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,131,32,0,0 ; vbroadcastss 0x2083(%rip),%ymm3 # 4cfc <_sk_callback_hsw+0x382> - DB 76,137,193 ; mov %r8,%rcx + DB 196,226,125,24,29,159,32,0,0 ; vbroadcastss 0x209f(%rip),%ymm3 # 4cf8 <_sk_callback_hsw+0x38c> + DB 76,137,201 ; mov %r9,%rcx DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2c8e <_sk_load_g8_hsw+0x4a> - DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,173 ; jmp 2c58 <_sk_load_g8_hsw+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 2c6f <_sk_load_g8_hsw+0x4b> + DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 91 ; pop %rbx + DB 235,172 ; jmp 2c38 <_sk_load_g8_hsw+0x14> PUBLIC _sk_gather_g8_hsw _sk_gather_g8_hsw LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,199 ; vmovq %xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,4,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0 - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 196,193,249,126,196 ; vmovq %xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,131,121,32,4,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 + DB 196,131,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,227,121,32,192,3 ; vpinsrb $0x3,%eax,%xmm0,%xmm0 - DB 67,15,182,4,32 ; movzbl (%r8,%r12,1),%eax + DB 67,15,182,4,41 ; movzbl (%r9,%r13,1),%eax DB 196,227,121,32,192,4 ; vpinsrb $0x4,%eax,%xmm0,%xmm0 - DB 67,15,182,4,56 ; movzbl (%r8,%r15,1),%eax + DB 67,15,182,4,33 ; movzbl (%r9,%r12,1),%eax DB 196,227,121,32,192,5 ; vpinsrb $0x5,%eax,%xmm0,%xmm0 - DB 67,15,182,4,48 ; movzbl (%r8,%r14,1),%eax + DB 67,15,182,4,57 ; movzbl (%r9,%r15,1),%eax DB 196,227,121,32,192,6 ; vpinsrb $0x6,%eax,%xmm0,%xmm0 - DB 65,15,182,4,24 ; movzbl (%r8,%rbx,1),%eax + DB 65,15,182,4,25 ; movzbl (%r9,%rbx,1),%eax DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,152,31,0,0 ; vbroadcastss 0x1f98(%rip),%ymm1 # 4d00 <_sk_callback_hsw+0x386> + DB 196,226,125,24,13,177,31,0,0 ; vbroadcastss 0x1fb1(%rip),%ymm1 # 4cfc <_sk_callback_hsw+0x390> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,141,31,0,0 ; vbroadcastss 0x1f8d(%rip),%ymm3 # 4d04 <_sk_callback_hsw+0x38a> + DB 196,226,125,24,29,166,31,0,0 ; vbroadcastss 0x1fa6(%rip),%ymm3 # 4d00 <_sk_callback_hsw+0x394> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -2721,56 +2709,57 @@ _sk_gather_g8_hsw LABEL PROC PUBLIC _sk_gather_i8_hsw _sk_gather_i8_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,192 ; mov %rax,%r8 - DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 2d97 <_sk_gather_i8_hsw+0xf> - DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 2d99 <_sk_gather_i8_hsw+0x11> + DB 73,137,193 ; mov %rax,%r9 + DB 77,133,201 ; test %r9,%r9 + DB 116,5 ; je 2d7c <_sk_gather_i8_hsw+0xf> + DB 76,137,200 ; mov %r9,%rax + DB 235,2 ; jmp 2d7e <_sk_gather_i8_hsw+0x11> DB 72,173 ; lods %ds:(%rsi),%rax + DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 76,139,8 ; mov (%rax),%r9 + DB 76,139,16 ; mov (%rax),%r10 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,194 ; mov %eax,%r10d + DB 65,137,195 ; mov %eax,%r11d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,195 ; vmovq %xmm0,%r11 - DB 69,137,222 ; mov %r11d,%r14d - DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,223 ; mov %ebx,%r15d + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,196 ; vmovq %xmm0,%r12 - DB 69,137,229 ; mov %r12d,%r13d - DB 73,193,236,32 ; shr $0x20,%r12 - DB 196,131,121,32,4,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 - DB 196,131,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,17,2 ; vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0 - DB 196,195,121,32,4,1,3 ; vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0 - DB 196,131,121,32,4,41,4 ; vpinsrb $0x4,(%r9,%r13,1),%xmm0,%xmm0 - DB 196,131,121,32,4,33,5 ; vpinsrb $0x5,(%r9,%r12,1),%xmm0,%xmm0 - DB 196,131,121,32,4,57,6 ; vpinsrb $0x6,(%r9,%r15,1),%xmm0,%xmm0 - DB 196,195,121,32,4,25,7 ; vpinsrb $0x7,(%r9,%rbx,1),%xmm0,%xmm0 + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 + DB 69,137,252 ; mov %r15d,%r12d + DB 73,193,239,32 ; shr $0x20,%r15 + DB 196,193,249,126,197 ; vmovq %xmm0,%r13 + DB 68,137,237 ; mov %r13d,%ebp + DB 73,193,237,32 ; shr $0x20,%r13 + DB 196,131,121,32,4,50,0 ; vpinsrb $0x0,(%r10,%r14,1),%xmm0,%xmm0 + DB 196,195,121,32,4,26,1 ; vpinsrb $0x1,(%r10,%rbx,1),%xmm0,%xmm0 + DB 196,131,121,32,4,26,2 ; vpinsrb $0x2,(%r10,%r11,1),%xmm0,%xmm0 + DB 196,195,121,32,4,2,3 ; vpinsrb $0x3,(%r10,%rax,1),%xmm0,%xmm0 + DB 196,195,121,32,4,42,4 ; vpinsrb $0x4,(%r10,%rbp,1),%xmm0,%xmm0 + DB 196,131,121,32,4,42,5 ; vpinsrb $0x5,(%r10,%r13,1),%xmm0,%xmm0 + DB 196,131,121,32,4,34,6 ; vpinsrb $0x6,(%r10,%r12,1),%xmm0,%xmm0 + DB 196,131,121,32,4,58,7 ; vpinsrb $0x7,(%r10,%r15,1),%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 - DB 73,139,64,8 ; mov 0x8(%r8),%rax + DB 73,139,65,8 ; mov 0x8(%r9),%rax DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,226,117,144,28,128 ; vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 - DB 197,229,219,5,25,33,0,0 ; vpand 0x2119(%rip),%ymm3,%ymm0 # 4f60 <_sk_callback_hsw+0x5e6> + DB 197,229,219,5,51,33,0,0 ; vpand 0x2133(%rip),%ymm3,%ymm0 # 4f60 <_sk_callback_hsw+0x5f4> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,180,30,0,0 ; vbroadcastss 0x1eb4(%rip),%ymm8 # 4d08 <_sk_callback_hsw+0x38e> + DB 196,98,125,24,5,202,30,0,0 ; vbroadcastss 0x1eca(%rip),%ymm8 # 4d04 <_sk_callback_hsw+0x398> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,30,33,0,0 ; vpshufb 0x211e(%rip),%ymm3,%ymm1 # 4f80 <_sk_callback_hsw+0x606> + DB 196,226,101,0,13,56,33,0,0 ; vpshufb 0x2138(%rip),%ymm3,%ymm1 # 4f80 <_sk_callback_hsw+0x614> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,44,33,0,0 ; vpshufb 0x212c(%rip),%ymm3,%ymm2 # 4fa0 <_sk_callback_hsw+0x626> + DB 196,226,101,0,21,70,33,0,0 ; vpshufb 0x2146(%rip),%ymm3,%ymm2 # 4fa0 <_sk_callback_hsw+0x634> DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 @@ -2782,153 +2771,158 @@ _sk_gather_i8_hsw LABEL PROC DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_load_565_hsw _sk_load_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 117,114 ; jne 2f14 <_sk_load_565_hsw+0x7c> - DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 117,114 ; jne 2efb <_sk_load_565_hsw+0x7c> + DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,86,30,0,0 ; vpbroadcastd 0x1e56(%rip),%ymm0 # 4d0c <_sk_callback_hsw+0x392> + DB 196,226,125,88,5,107,30,0,0 ; vpbroadcastd 0x1e6b(%rip),%ymm0 # 4d08 <_sk_callback_hsw+0x39c> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,73,30,0,0 ; vbroadcastss 0x1e49(%rip),%ymm1 # 4d10 <_sk_callback_hsw+0x396> + DB 196,226,125,24,13,94,30,0,0 ; vbroadcastss 0x1e5e(%rip),%ymm1 # 4d0c <_sk_callback_hsw+0x3a0> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,64,30,0,0 ; vpbroadcastd 0x1e40(%rip),%ymm1 # 4d14 <_sk_callback_hsw+0x39a> + DB 196,226,125,88,13,85,30,0,0 ; vpbroadcastd 0x1e55(%rip),%ymm1 # 4d10 <_sk_callback_hsw+0x3a4> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,51,30,0,0 ; vbroadcastss 0x1e33(%rip),%ymm3 # 4d18 <_sk_callback_hsw+0x39e> + DB 196,226,125,24,29,72,30,0,0 ; vbroadcastss 0x1e48(%rip),%ymm3 # 4d14 <_sk_callback_hsw+0x3a8> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,42,30,0,0 ; vpbroadcastd 0x1e2a(%rip),%ymm3 # 4d1c <_sk_callback_hsw+0x3a2> + DB 196,226,125,88,29,63,30,0,0 ; vpbroadcastd 0x1e3f(%rip),%ymm3 # 4d18 <_sk_callback_hsw+0x3ac> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,29,30,0,0 ; vbroadcastss 0x1e1d(%rip),%ymm3 # 4d20 <_sk_callback_hsw+0x3a6> + DB 196,226,125,24,29,50,30,0,0 ; vbroadcastss 0x1e32(%rip),%ymm3 # 4d1c <_sk_callback_hsw+0x3b0> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,18,30,0,0 ; vbroadcastss 0x1e12(%rip),%ymm3 # 4d24 <_sk_callback_hsw+0x3aa> + DB 196,226,125,24,29,39,30,0,0 ; vbroadcastss 0x1e27(%rip),%ymm3 # 4d20 <_sk_callback_hsw+0x3b4> DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,128 ; ja 2ea8 <_sk_load_565_hsw+0x10> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2f7c <_sk_load_565_hsw+0xe4> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,128 ; ja 2e8f <_sk_load_565_hsw+0x10> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,74,0,0,0 ; lea 0x4a(%rip),%r10 # 2f64 <_sk_load_565_hsw+0xe5> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,193,121,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,44,255,255,255 ; jmpq 2ea8 <_sk_load_565_hsw+0x10> - DB 244 ; hlt - DB 255 ; (bad) + DB 196,193,121,196,68,83,12,6 ; vpinsrw $0x6,0xc(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,10,5 ; vpinsrw $0x5,0xa(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,8,4 ; vpinsrw $0x4,0x8(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,6,3 ; vpinsrw $0x3,0x6(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0 + DB 233,44,255,255,255 ; jmpq 2e8f <_sk_load_565_hsw+0x10> + DB 144 ; nop + DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 236 ; in (%dx),%al + DB 235,255 ; jmp 2f69 <_sk_load_565_hsw+0xea> DB 255 ; (bad) + DB 255,227 ; jmpq *%rbx DB 255 ; (bad) - DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) + DB 219,255 ; (bad) DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) + DB 255,211 ; callq *%rbx DB 255 ; (bad) - DB 255,212 ; callq *%rsp DB 255 ; (bad) + DB 255,203 ; dec %ebx DB 255 ; (bad) - DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax + DB 191 ; .byte 0xbf DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_gather_565_hsw _sk_gather_565_hsw LABEL PROC + DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,199 ; vmovq %xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 71,15,183,20,80 ; movzwl (%r8,%r10,2),%r10d - DB 71,15,183,28,88 ; movzwl (%r8,%r11,2),%r11d - DB 196,193,121,110,195 ; vmovd %r11d,%xmm0 - DB 196,193,121,196,194,1 ; vpinsrw $0x1,%r10d,%xmm0,%xmm0 - DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d - DB 196,193,121,196,193,2 ; vpinsrw $0x2,%r9d,%xmm0,%xmm0 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 196,193,249,126,196 ; vmovq %xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 71,15,183,28,89 ; movzwl (%r9,%r11,2),%r11d + DB 67,15,183,44,113 ; movzwl (%r9,%r14,2),%ebp + DB 197,249,110,197 ; vmovd %ebp,%xmm0 + DB 196,193,121,196,195,1 ; vpinsrw $0x1,%r11d,%xmm0,%xmm0 + DB 67,15,183,44,81 ; movzwl (%r9,%r10,2),%ebp + DB 197,249,196,197,2 ; vpinsrw $0x2,%ebp,%xmm0,%xmm0 + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 197,249,196,192,3 ; vpinsrw $0x3,%eax,%xmm0,%xmm0 - DB 67,15,183,4,96 ; movzwl (%r8,%r12,2),%eax + DB 67,15,183,4,105 ; movzwl (%r9,%r13,2),%eax DB 197,249,196,192,4 ; vpinsrw $0x4,%eax,%xmm0,%xmm0 - DB 67,15,183,4,120 ; movzwl (%r8,%r15,2),%eax + DB 67,15,183,4,97 ; movzwl (%r9,%r12,2),%eax DB 197,249,196,192,5 ; vpinsrw $0x5,%eax,%xmm0,%xmm0 - DB 67,15,183,4,112 ; movzwl (%r8,%r14,2),%eax + DB 67,15,183,4,121 ; movzwl (%r9,%r15,2),%eax DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0 - DB 65,15,183,4,88 ; movzwl (%r8,%rbx,2),%eax + DB 65,15,183,4,89 ; movzwl (%r9,%rbx,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,213,28,0,0 ; vpbroadcastd 0x1cd5(%rip),%ymm0 # 4d28 <_sk_callback_hsw+0x3ae> + DB 196,226,125,88,5,232,28,0,0 ; vpbroadcastd 0x1ce8(%rip),%ymm0 # 4d24 <_sk_callback_hsw+0x3b8> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,200,28,0,0 ; vbroadcastss 0x1cc8(%rip),%ymm1 # 4d2c <_sk_callback_hsw+0x3b2> + DB 196,226,125,24,13,219,28,0,0 ; vbroadcastss 0x1cdb(%rip),%ymm1 # 4d28 <_sk_callback_hsw+0x3bc> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,191,28,0,0 ; vpbroadcastd 0x1cbf(%rip),%ymm1 # 4d30 <_sk_callback_hsw+0x3b6> + DB 196,226,125,88,13,210,28,0,0 ; vpbroadcastd 0x1cd2(%rip),%ymm1 # 4d2c <_sk_callback_hsw+0x3c0> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,178,28,0,0 ; vbroadcastss 0x1cb2(%rip),%ymm3 # 4d34 <_sk_callback_hsw+0x3ba> + DB 196,226,125,24,29,197,28,0,0 ; vbroadcastss 0x1cc5(%rip),%ymm3 # 4d30 <_sk_callback_hsw+0x3c4> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,169,28,0,0 ; vpbroadcastd 0x1ca9(%rip),%ymm3 # 4d38 <_sk_callback_hsw+0x3be> + DB 196,226,125,88,29,188,28,0,0 ; vpbroadcastd 0x1cbc(%rip),%ymm3 # 4d34 <_sk_callback_hsw+0x3c8> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,156,28,0,0 ; vbroadcastss 0x1c9c(%rip),%ymm3 # 4d3c <_sk_callback_hsw+0x3c2> + DB 196,226,125,24,29,175,28,0,0 ; vbroadcastss 0x1caf(%rip),%ymm3 # 4d38 <_sk_callback_hsw+0x3cc> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,145,28,0,0 ; vbroadcastss 0x1c91(%rip),%ymm3 # 4d40 <_sk_callback_hsw+0x3c6> + DB 196,226,125,24,29,164,28,0,0 ; vbroadcastss 0x1ca4(%rip),%ymm3 # 4d3c <_sk_callback_hsw+0x3d0> DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_hsw _sk_store_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,126,28,0,0 ; vbroadcastss 0x1c7e(%rip),%ymm8 # 4d44 <_sk_callback_hsw+0x3ca> + DB 76,139,24 ; mov (%rax),%r11 + DB 196,98,125,24,5,142,28,0,0 ; vbroadcastss 0x1c8e(%rip),%ymm8 # 4d40 <_sk_callback_hsw+0x3d4> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,53,114,241,11 ; vpslld $0xb,%ymm9,%ymm9 - DB 196,98,125,24,21,105,28,0,0 ; vbroadcastss 0x1c69(%rip),%ymm10 # 4d48 <_sk_callback_hsw+0x3ce> + DB 196,98,125,24,21,121,28,0,0 ; vbroadcastss 0x1c79(%rip),%ymm10 # 4d44 <_sk_callback_hsw+0x3d8> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,45,114,242,5 ; vpslld $0x5,%ymm10,%ymm10 @@ -2938,29 +2932,29 @@ _sk_store_565_hsw LABEL PROC DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 311d <_sk_store_565_hsw+0x65> - DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 3119 <_sk_store_565_hsw+0x61> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 317c <_sk_store_565_hsw+0xc4> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,121,21,68,122,12,6 ; vpextrw $0x6,%xmm8,0xc(%r10,%rdi,2) - DB 196,67,121,21,68,122,10,5 ; vpextrw $0x5,%xmm8,0xa(%r10,%rdi,2) - DB 196,67,121,21,68,122,8,4 ; vpextrw $0x4,%xmm8,0x8(%r10,%rdi,2) - DB 196,67,121,21,68,122,6,3 ; vpextrw $0x3,%xmm8,0x6(%r10,%rdi,2) - DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) - DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) - DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2) - DB 235,159 ; jmp 3119 <_sk_store_565_hsw+0x61> + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3109 <_sk_store_565_hsw+0x65> + DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 3105 <_sk_store_565_hsw+0x61> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 3168 <_sk_store_565_hsw+0xc4> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,67,121,21,68,83,12,6 ; vpextrw $0x6,%xmm8,0xc(%r11,%rdx,2) + DB 196,67,121,21,68,83,10,5 ; vpextrw $0x5,%xmm8,0xa(%r11,%rdx,2) + DB 196,67,121,21,68,83,8,4 ; vpextrw $0x4,%xmm8,0x8(%r11,%rdx,2) + DB 196,67,121,21,68,83,6,3 ; vpextrw $0x3,%xmm8,0x6(%r11,%rdx,2) + DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2) + DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2) + DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2) + DB 235,159 ; jmp 3105 <_sk_store_565_hsw+0x61> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -2989,53 +2983,53 @@ _sk_store_565_hsw LABEL PROC PUBLIC _sk_load_4444_hsw _sk_load_4444_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,138,0,0,0 ; jne 3230 <_sk_load_4444_hsw+0x98> - DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,138,0,0,0 ; jne 321c <_sk_load_4444_hsw+0x98> + DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,146,27,0,0 ; vpbroadcastd 0x1b92(%rip),%ymm0 # 4d4c <_sk_callback_hsw+0x3d2> + DB 196,226,125,88,5,162,27,0,0 ; vpbroadcastd 0x1ba2(%rip),%ymm0 # 4d48 <_sk_callback_hsw+0x3dc> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,133,27,0,0 ; vbroadcastss 0x1b85(%rip),%ymm1 # 4d50 <_sk_callback_hsw+0x3d6> + DB 196,226,125,24,13,149,27,0,0 ; vbroadcastss 0x1b95(%rip),%ymm1 # 4d4c <_sk_callback_hsw+0x3e0> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,124,27,0,0 ; vpbroadcastd 0x1b7c(%rip),%ymm1 # 4d54 <_sk_callback_hsw+0x3da> + DB 196,226,125,88,13,140,27,0,0 ; vpbroadcastd 0x1b8c(%rip),%ymm1 # 4d50 <_sk_callback_hsw+0x3e4> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,111,27,0,0 ; vbroadcastss 0x1b6f(%rip),%ymm2 # 4d58 <_sk_callback_hsw+0x3de> + DB 196,226,125,24,21,127,27,0,0 ; vbroadcastss 0x1b7f(%rip),%ymm2 # 4d54 <_sk_callback_hsw+0x3e8> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,102,27,0,0 ; vpbroadcastd 0x1b66(%rip),%ymm2 # 4d5c <_sk_callback_hsw+0x3e2> + DB 196,226,125,88,21,118,27,0,0 ; vpbroadcastd 0x1b76(%rip),%ymm2 # 4d58 <_sk_callback_hsw+0x3ec> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,89,27,0,0 ; vbroadcastss 0x1b59(%rip),%ymm8 # 4d60 <_sk_callback_hsw+0x3e6> + DB 196,98,125,24,5,105,27,0,0 ; vbroadcastss 0x1b69(%rip),%ymm8 # 4d5c <_sk_callback_hsw+0x3f0> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,79,27,0,0 ; vpbroadcastd 0x1b4f(%rip),%ymm8 # 4d64 <_sk_callback_hsw+0x3ea> + DB 196,98,125,88,5,95,27,0,0 ; vpbroadcastd 0x1b5f(%rip),%ymm8 # 4d60 <_sk_callback_hsw+0x3f4> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,65,27,0,0 ; vbroadcastss 0x1b41(%rip),%ymm8 # 4d68 <_sk_callback_hsw+0x3ee> + DB 196,98,125,24,5,81,27,0,0 ; vbroadcastss 0x1b51(%rip),%ymm8 # 4d64 <_sk_callback_hsw+0x3f8> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,100,255,255,255 ; ja 31ac <_sk_load_4444_hsw+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 329c <_sk_load_4444_hsw+0x104> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 15,135,100,255,255,255 ; ja 3198 <_sk_load_4444_hsw+0x14> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,73,0,0,0 ; lea 0x49(%rip),%r10 # 3288 <_sk_load_4444_hsw+0x104> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,193,121,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,16,255,255,255 ; jmpq 31ac <_sk_load_4444_hsw+0x14> + DB 196,193,121,196,68,83,12,6 ; vpinsrw $0x6,0xc(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,10,5 ; vpinsrw $0x5,0xa(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,8,4 ; vpinsrw $0x4,0x8(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,6,3 ; vpinsrw $0x3,0x6(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0 + DB 233,16,255,255,255 ; jmpq 3198 <_sk_load_4444_hsw+0x14> DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) @@ -3062,79 +3056,83 @@ _sk_load_4444_hsw LABEL PROC PUBLIC _sk_gather_4444_hsw _sk_gather_4444_hsw LABEL PROC + DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,199 ; vmovq %xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 71,15,183,20,80 ; movzwl (%r8,%r10,2),%r10d - DB 71,15,183,28,88 ; movzwl (%r8,%r11,2),%r11d - DB 196,193,121,110,195 ; vmovd %r11d,%xmm0 - DB 196,193,121,196,194,1 ; vpinsrw $0x1,%r10d,%xmm0,%xmm0 - DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d - DB 196,193,121,196,193,2 ; vpinsrw $0x2,%r9d,%xmm0,%xmm0 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 196,193,249,126,196 ; vmovq %xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 71,15,183,28,89 ; movzwl (%r9,%r11,2),%r11d + DB 67,15,183,44,113 ; movzwl (%r9,%r14,2),%ebp + DB 197,249,110,197 ; vmovd %ebp,%xmm0 + DB 196,193,121,196,195,1 ; vpinsrw $0x1,%r11d,%xmm0,%xmm0 + DB 67,15,183,44,81 ; movzwl (%r9,%r10,2),%ebp + DB 197,249,196,197,2 ; vpinsrw $0x2,%ebp,%xmm0,%xmm0 + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 197,249,196,192,3 ; vpinsrw $0x3,%eax,%xmm0,%xmm0 - DB 67,15,183,4,96 ; movzwl (%r8,%r12,2),%eax + DB 67,15,183,4,105 ; movzwl (%r9,%r13,2),%eax DB 197,249,196,192,4 ; vpinsrw $0x4,%eax,%xmm0,%xmm0 - DB 67,15,183,4,120 ; movzwl (%r8,%r15,2),%eax + DB 67,15,183,4,97 ; movzwl (%r9,%r12,2),%eax DB 197,249,196,192,5 ; vpinsrw $0x5,%eax,%xmm0,%xmm0 - DB 67,15,183,4,112 ; movzwl (%r8,%r14,2),%eax + DB 67,15,183,4,121 ; movzwl (%r9,%r15,2),%eax DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0 - DB 65,15,183,4,88 ; movzwl (%r8,%rbx,2),%eax + DB 65,15,183,4,89 ; movzwl (%r9,%rbx,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,249,25,0,0 ; vpbroadcastd 0x19f9(%rip),%ymm0 # 4d6c <_sk_callback_hsw+0x3f2> + DB 196,226,125,88,5,8,26,0,0 ; vpbroadcastd 0x1a08(%rip),%ymm0 # 4d68 <_sk_callback_hsw+0x3fc> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,236,25,0,0 ; vbroadcastss 0x19ec(%rip),%ymm1 # 4d70 <_sk_callback_hsw+0x3f6> + DB 196,226,125,24,13,251,25,0,0 ; vbroadcastss 0x19fb(%rip),%ymm1 # 4d6c <_sk_callback_hsw+0x400> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,227,25,0,0 ; vpbroadcastd 0x19e3(%rip),%ymm1 # 4d74 <_sk_callback_hsw+0x3fa> + DB 196,226,125,88,13,242,25,0,0 ; vpbroadcastd 0x19f2(%rip),%ymm1 # 4d70 <_sk_callback_hsw+0x404> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,214,25,0,0 ; vbroadcastss 0x19d6(%rip),%ymm2 # 4d78 <_sk_callback_hsw+0x3fe> + DB 196,226,125,24,21,229,25,0,0 ; vbroadcastss 0x19e5(%rip),%ymm2 # 4d74 <_sk_callback_hsw+0x408> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,205,25,0,0 ; vpbroadcastd 0x19cd(%rip),%ymm2 # 4d7c <_sk_callback_hsw+0x402> + DB 196,226,125,88,21,220,25,0,0 ; vpbroadcastd 0x19dc(%rip),%ymm2 # 4d78 <_sk_callback_hsw+0x40c> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,192,25,0,0 ; vbroadcastss 0x19c0(%rip),%ymm8 # 4d80 <_sk_callback_hsw+0x406> + DB 196,98,125,24,5,207,25,0,0 ; vbroadcastss 0x19cf(%rip),%ymm8 # 4d7c <_sk_callback_hsw+0x410> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,182,25,0,0 ; vpbroadcastd 0x19b6(%rip),%ymm8 # 4d84 <_sk_callback_hsw+0x40a> + DB 196,98,125,88,5,197,25,0,0 ; vpbroadcastd 0x19c5(%rip),%ymm8 # 4d80 <_sk_callback_hsw+0x414> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,168,25,0,0 ; vbroadcastss 0x19a8(%rip),%ymm8 # 4d88 <_sk_callback_hsw+0x40e> + DB 196,98,125,24,5,183,25,0,0 ; vbroadcastss 0x19b7(%rip),%ymm8 # 4d84 <_sk_callback_hsw+0x418> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_store_4444_hsw _sk_store_4444_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,142,25,0,0 ; vbroadcastss 0x198e(%rip),%ymm8 # 4d8c <_sk_callback_hsw+0x412> + DB 76,139,24 ; mov (%rax),%r11 + DB 196,98,125,24,5,154,25,0,0 ; vbroadcastss 0x199a(%rip),%ymm8 # 4d88 <_sk_callback_hsw+0x41c> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,53,114,241,12 ; vpslld $0xc,%ymm9,%ymm9 @@ -3151,29 +3149,29 @@ _sk_store_4444_hsw LABEL PROC DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3461 <_sk_store_4444_hsw+0x71> - DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 345d <_sk_store_4444_hsw+0x6d> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 34c0 <_sk_store_4444_hsw+0xd0> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,121,21,68,122,12,6 ; vpextrw $0x6,%xmm8,0xc(%r10,%rdi,2) - DB 196,67,121,21,68,122,10,5 ; vpextrw $0x5,%xmm8,0xa(%r10,%rdi,2) - DB 196,67,121,21,68,122,8,4 ; vpextrw $0x4,%xmm8,0x8(%r10,%rdi,2) - DB 196,67,121,21,68,122,6,3 ; vpextrw $0x3,%xmm8,0x6(%r10,%rdi,2) - DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) - DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) - DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2) - DB 235,159 ; jmp 345d <_sk_store_4444_hsw+0x6d> + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3451 <_sk_store_4444_hsw+0x71> + DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 344d <_sk_store_4444_hsw+0x6d> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 34b0 <_sk_store_4444_hsw+0xd0> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,67,121,21,68,83,12,6 ; vpextrw $0x6,%xmm8,0xc(%r11,%rdx,2) + DB 196,67,121,21,68,83,10,5 ; vpextrw $0x5,%xmm8,0xa(%r11,%rdx,2) + DB 196,67,121,21,68,83,8,4 ; vpextrw $0x4,%xmm8,0x8(%r11,%rdx,2) + DB 196,67,121,21,68,83,6,3 ; vpextrw $0x3,%xmm8,0x6(%r11,%rdx,2) + DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2) + DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2) + DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2) + DB 235,159 ; jmp 344d <_sk_store_4444_hsw+0x6d> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -3201,28 +3199,28 @@ _sk_store_4444_hsw LABEL PROC PUBLIC _sk_load_8888_hsw _sk_load_8888_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 DB 77,133,192 ; test %r8,%r8 - DB 117,88 ; jne 3549 <_sk_load_8888_hsw+0x6d> - DB 196,193,124,16,25 ; vmovups (%r9),%ymm3 - DB 197,228,84,5,194,26,0,0 ; vandps 0x1ac2(%rip),%ymm3,%ymm0 # 4fc0 <_sk_callback_hsw+0x646> + DB 117,88 ; jne 3539 <_sk_load_8888_hsw+0x6d> + DB 196,193,124,16,26 ; vmovups (%r10),%ymm3 + DB 197,228,84,5,210,26,0,0 ; vandps 0x1ad2(%rip),%ymm3,%ymm0 # 4fc0 <_sk_callback_hsw+0x654> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,133,24,0,0 ; vbroadcastss 0x1885(%rip),%ymm8 # 4d90 <_sk_callback_hsw+0x416> + DB 196,98,125,24,5,145,24,0,0 ; vbroadcastss 0x1891(%rip),%ymm8 # 4d8c <_sk_callback_hsw+0x420> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,199,26,0,0 ; vpshufb 0x1ac7(%rip),%ymm3,%ymm1 # 4fe0 <_sk_callback_hsw+0x666> + DB 196,226,101,0,13,215,26,0,0 ; vpshufb 0x1ad7(%rip),%ymm3,%ymm1 # 4fe0 <_sk_callback_hsw+0x674> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,213,26,0,0 ; vpshufb 0x1ad5(%rip),%ymm3,%ymm2 # 5000 <_sk_callback_hsw+0x686> + DB 196,226,101,0,21,229,26,0,0 ; vpshufb 0x1ae5(%rip),%ymm3,%ymm2 # 5000 <_sk_callback_hsw+0x694> DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx @@ -3231,28 +3229,28 @@ _sk_load_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 - DB 196,194,125,44,25 ; vmaskmovps (%r9),%ymm0,%ymm3 - DB 235,135 ; jmp 34f6 <_sk_load_8888_hsw+0x1a> + DB 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3 + DB 235,135 ; jmp 34e6 <_sk_load_8888_hsw+0x1a> PUBLIC _sk_gather_8888_hsw _sk_gather_8888_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,194,117,144,28,128 ; vpgatherdd %ymm1,(%r8,%ymm0,4),%ymm3 - DB 197,229,219,5,131,26,0,0 ; vpand 0x1a83(%rip),%ymm3,%ymm0 # 5020 <_sk_callback_hsw+0x6a6> + DB 196,194,117,144,28,129 ; vpgatherdd %ymm1,(%r9,%ymm0,4),%ymm3 + DB 197,229,219,5,147,26,0,0 ; vpand 0x1a93(%rip),%ymm3,%ymm0 # 5020 <_sk_callback_hsw+0x6b4> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,234,23,0,0 ; vbroadcastss 0x17ea(%rip),%ymm8 # 4d94 <_sk_callback_hsw+0x41a> + DB 196,98,125,24,5,246,23,0,0 ; vbroadcastss 0x17f6(%rip),%ymm8 # 4d90 <_sk_callback_hsw+0x424> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,136,26,0,0 ; vpshufb 0x1a88(%rip),%ymm3,%ymm1 # 5040 <_sk_callback_hsw+0x6c6> + DB 196,226,101,0,13,152,26,0,0 ; vpshufb 0x1a98(%rip),%ymm3,%ymm1 # 5040 <_sk_callback_hsw+0x6d4> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,150,26,0,0 ; vpshufb 0x1a96(%rip),%ymm3,%ymm2 # 5060 <_sk_callback_hsw+0x6e6> + DB 196,226,101,0,21,166,26,0,0 ; vpshufb 0x1aa6(%rip),%ymm3,%ymm2 # 5060 <_sk_callback_hsw+0x6f4> DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 @@ -3263,11 +3261,11 @@ _sk_gather_8888_hsw LABEL PROC PUBLIC _sk_store_8888_hsw _sk_store_8888_hsw LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 - DB 196,98,125,24,5,154,23,0,0 ; vbroadcastss 0x179a(%rip),%ymm8 # 4d98 <_sk_callback_hsw+0x41e> + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 + DB 196,98,125,24,5,166,23,0,0 ; vbroadcastss 0x17a6(%rip),%ymm8 # 4d94 <_sk_callback_hsw+0x428> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -3283,10 +3281,10 @@ _sk_store_8888_hsw LABEL PROC DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 77,133,192 ; test %r8,%r8 - DB 117,12 ; jne 3658 <_sk_store_8888_hsw+0x73> - DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) + DB 117,12 ; jne 3648 <_sk_store_8888_hsw+0x73> + DB 196,65,124,17,2 ; vmovups %ymm8,(%r10) DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx @@ -3295,19 +3293,19 @@ _sk_store_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9 - DB 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) - DB 235,211 ; jmp 3651 <_sk_store_8888_hsw+0x6c> + DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10) + DB 235,211 ; jmp 3641 <_sk_store_8888_hsw+0x6c> PUBLIC _sk_load_f16_hsw _sk_load_f16_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 117,97 ; jne 36e9 <_sk_load_f16_hsw+0x6b> - DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 - DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 - DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 - DB 197,122,111,76,248,48 ; vmovdqu 0x30(%rax,%rdi,8),%xmm9 + DB 77,133,192 ; test %r8,%r8 + DB 117,97 ; jne 36d9 <_sk_load_f16_hsw+0x6b> + DB 197,121,16,4,208 ; vmovupd (%rax,%rdx,8),%xmm8 + DB 197,249,16,84,208,16 ; vmovupd 0x10(%rax,%rdx,8),%xmm2 + DB 197,249,16,92,208,32 ; vmovupd 0x20(%rax,%rdx,8),%xmm3 + DB 197,122,111,76,208,48 ; vmovdqu 0x30(%rax,%rdx,8),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 @@ -3326,37 +3324,37 @@ _sk_load_f16_hsw LABEL PROC DB 196,226,125,19,219 ; vcvtph2ps %xmm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 + DB 197,123,16,4,208 ; vmovsd (%rax,%rdx,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 3748 <_sk_load_f16_hsw+0xca> - DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 3748 <_sk_load_f16_hsw+0xca> - DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 3755 <_sk_load_f16_hsw+0xd7> - DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 3755 <_sk_load_f16_hsw+0xd7> - DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,114,255,255,255 ; je 369f <_sk_load_f16_hsw+0x21> - DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,98,255,255,255 ; jb 369f <_sk_load_f16_hsw+0x21> - DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,87,255,255,255 ; jmpq 369f <_sk_load_f16_hsw+0x21> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,79 ; je 3738 <_sk_load_f16_hsw+0xca> + DB 197,57,22,68,208,8 ; vmovhpd 0x8(%rax,%rdx,8),%xmm8,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,67 ; jb 3738 <_sk_load_f16_hsw+0xca> + DB 197,251,16,84,208,16 ; vmovsd 0x10(%rax,%rdx,8),%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 116,68 ; je 3745 <_sk_load_f16_hsw+0xd7> + DB 197,233,22,84,208,24 ; vmovhpd 0x18(%rax,%rdx,8),%xmm2,%xmm2 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,56 ; jb 3745 <_sk_load_f16_hsw+0xd7> + DB 197,251,16,92,208,32 ; vmovsd 0x20(%rax,%rdx,8),%xmm3 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 15,132,114,255,255,255 ; je 368f <_sk_load_f16_hsw+0x21> + DB 197,225,22,92,208,40 ; vmovhpd 0x28(%rax,%rdx,8),%xmm3,%xmm3 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 15,130,98,255,255,255 ; jb 368f <_sk_load_f16_hsw+0x21> + DB 197,122,126,76,208,48 ; vmovq 0x30(%rax,%rdx,8),%xmm9 + DB 233,87,255,255,255 ; jmpq 368f <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,74,255,255,255 ; jmpq 369f <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq 368f <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq 369f <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 368f <_sk_load_f16_hsw+0x21> PUBLIC _sk_gather_f16_hsw _sk_gather_f16_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 @@ -3364,9 +3362,9 @@ _sk_gather_f16_hsw LABEL PROC DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,194,237,144,28,192 ; vpgatherdq %ymm2,(%r8,%xmm0,8),%ymm3 + DB 196,194,237,144,28,193 ; vpgatherdq %ymm2,(%r9,%xmm0,8),%ymm3 DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 - DB 196,194,245,144,20,192 ; vpgatherdq %ymm1,(%r8,%xmm0,8),%ymm2 + DB 196,194,245,144,20,193 ; vpgatherdq %ymm1,(%r9,%xmm0,8),%ymm2 DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0 DB 196,227,125,57,209,1 ; vextracti128 $0x1,%ymm2,%xmm1 DB 197,97,97,192 ; vpunpcklwd %xmm0,%xmm3,%xmm8 @@ -3404,44 +3402,44 @@ _sk_store_f16_hsw LABEL PROC DB 196,65,25,106,209 ; vpunpckhdq %xmm9,%xmm12,%xmm10 DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9 DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,27 ; jne 384d <_sk_store_f16_hsw+0x65> - DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8) - DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8) - DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8) - DB 197,122,127,68,248,48 ; vmovdqu %xmm8,0x30(%rax,%rdi,8) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,241 ; je 3849 <_sk_store_f16_hsw+0x61> - DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,229 ; jb 3849 <_sk_store_f16_hsw+0x61> - DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je 3849 <_sk_store_f16_hsw+0x61> - DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8) - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,209 ; jb 3849 <_sk_store_f16_hsw+0x61> - DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je 3849 <_sk_store_f16_hsw+0x61> - DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8) - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,189 ; jb 3849 <_sk_store_f16_hsw+0x61> - DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp 3849 <_sk_store_f16_hsw+0x61> + DB 77,133,192 ; test %r8,%r8 + DB 117,27 ; jne 383d <_sk_store_f16_hsw+0x65> + DB 197,120,17,28,208 ; vmovups %xmm11,(%rax,%rdx,8) + DB 197,120,17,84,208,16 ; vmovups %xmm10,0x10(%rax,%rdx,8) + DB 197,120,17,76,208,32 ; vmovups %xmm9,0x20(%rax,%rdx,8) + DB 197,122,127,68,208,48 ; vmovdqu %xmm8,0x30(%rax,%rdx,8) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 197,121,214,28,208 ; vmovq %xmm11,(%rax,%rdx,8) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,241 ; je 3839 <_sk_store_f16_hsw+0x61> + DB 197,121,23,92,208,8 ; vmovhpd %xmm11,0x8(%rax,%rdx,8) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,229 ; jb 3839 <_sk_store_f16_hsw+0x61> + DB 197,121,214,84,208,16 ; vmovq %xmm10,0x10(%rax,%rdx,8) + DB 116,221 ; je 3839 <_sk_store_f16_hsw+0x61> + DB 197,121,23,84,208,24 ; vmovhpd %xmm10,0x18(%rax,%rdx,8) + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,209 ; jb 3839 <_sk_store_f16_hsw+0x61> + DB 197,121,214,76,208,32 ; vmovq %xmm9,0x20(%rax,%rdx,8) + DB 116,201 ; je 3839 <_sk_store_f16_hsw+0x61> + DB 197,121,23,76,208,40 ; vmovhpd %xmm9,0x28(%rax,%rdx,8) + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,189 ; jb 3839 <_sk_store_f16_hsw+0x61> + DB 197,121,214,68,208,48 ; vmovq %xmm8,0x30(%rax,%rdx,8) + DB 235,181 ; jmp 3839 <_sk_store_f16_hsw+0x61> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,204,0,0,0 ; jne 3976 <_sk_load_u16_be_hsw+0xe2> - DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 - DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 - DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 - DB 196,65,122,111,76,64,48 ; vmovdqu 0x30(%r8,%rax,2),%xmm9 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,204,0,0,0 ; jne 3966 <_sk_load_u16_be_hsw+0xe2> + DB 196,65,121,16,4,65 ; vmovupd (%r9,%rax,2),%xmm8 + DB 196,193,121,16,84,65,16 ; vmovupd 0x10(%r9,%rax,2),%xmm2 + DB 196,193,121,16,92,65,32 ; vmovupd 0x20(%r9,%rax,2),%xmm3 + DB 196,65,122,111,76,65,48 ; vmovdqu 0x30(%r9,%rax,2),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 @@ -3456,7 +3454,7 @@ _sk_load_u16_be_hsw LABEL PROC DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,21,145,20,0,0 ; vbroadcastss 0x1491(%rip),%ymm10 # 4d9c <_sk_callback_hsw+0x422> + DB 196,98,125,24,21,157,20,0,0 ; vbroadcastss 0x149d(%rip),%ymm10 # 4d98 <_sk_callback_hsw+0x42c> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -3481,44 +3479,44 @@ _sk_load_u16_be_hsw LABEL PROC DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 + DB 196,65,123,16,4,65 ; vmovsd (%r9,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 39dc <_sk_load_u16_be_hsw+0x148> - DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 39dc <_sk_load_u16_be_hsw+0x148> - DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 39e9 <_sk_load_u16_be_hsw+0x155> - DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 39e9 <_sk_load_u16_be_hsw+0x155> - DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,6,255,255,255 ; je 38c5 <_sk_load_u16_be_hsw+0x31> - DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,245,254,255,255 ; jb 38c5 <_sk_load_u16_be_hsw+0x31> - DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,233,254,255,255 ; jmpq 38c5 <_sk_load_u16_be_hsw+0x31> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,85 ; je 39cc <_sk_load_u16_be_hsw+0x148> + DB 196,65,57,22,68,65,8 ; vmovhpd 0x8(%r9,%rax,2),%xmm8,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,72 ; jb 39cc <_sk_load_u16_be_hsw+0x148> + DB 196,193,123,16,84,65,16 ; vmovsd 0x10(%r9,%rax,2),%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 116,72 ; je 39d9 <_sk_load_u16_be_hsw+0x155> + DB 196,193,105,22,84,65,24 ; vmovhpd 0x18(%r9,%rax,2),%xmm2,%xmm2 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,59 ; jb 39d9 <_sk_load_u16_be_hsw+0x155> + DB 196,193,123,16,92,65,32 ; vmovsd 0x20(%r9,%rax,2),%xmm3 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 15,132,6,255,255,255 ; je 38b5 <_sk_load_u16_be_hsw+0x31> + DB 196,193,97,22,92,65,40 ; vmovhpd 0x28(%r9,%rax,2),%xmm3,%xmm3 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 15,130,245,254,255,255 ; jb 38b5 <_sk_load_u16_be_hsw+0x31> + DB 196,65,122,126,76,65,48 ; vmovq 0x30(%r9,%rax,2),%xmm9 + DB 233,233,254,255,255 ; jmpq 38b5 <_sk_load_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,220,254,255,255 ; jmpq 38c5 <_sk_load_u16_be_hsw+0x31> + DB 233,220,254,255,255 ; jmpq 38b5 <_sk_load_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,211,254,255,255 ; jmpq 38c5 <_sk_load_u16_be_hsw+0x31> + DB 233,211,254,255,255 ; jmpq 38b5 <_sk_load_u16_be_hsw+0x31> PUBLIC _sk_load_rgb_u16_be_hsw _sk_load_rgb_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,204,0,0,0 ; jne 3ad0 <_sk_load_rgb_u16_be_hsw+0xde> - DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0 - DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2 - DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1 - DB 196,193,122,111,92,64,32 ; vmovdqu 0x20(%r8,%rax,2),%xmm3 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,82 ; lea (%rdx,%rdx,2),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,204,0,0,0 ; jne 3ac0 <_sk_load_rgb_u16_be_hsw+0xde> + DB 196,193,122,111,4,65 ; vmovdqu (%r9,%rax,2),%xmm0 + DB 196,193,122,111,84,65,12 ; vmovdqu 0xc(%r9,%rax,2),%xmm2 + DB 196,193,122,111,76,65,24 ; vmovdqu 0x18(%r9,%rax,2),%xmm1 + DB 196,193,122,111,92,65,32 ; vmovdqu 0x20(%r9,%rax,2),%xmm3 DB 197,225,115,219,4 ; vpsrldq $0x4,%xmm3,%xmm3 DB 197,185,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm8 DB 197,177,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm9 @@ -3538,7 +3536,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,21,34,19,0,0 ; vbroadcastss 0x1322(%rip),%ymm10 # 4da0 <_sk_callback_hsw+0x426> + DB 196,98,125,24,21,46,19,0,0 ; vbroadcastss 0x132e(%rip),%ymm10 # 4d9c <_sk_callback_hsw+0x430> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -3555,48 +3553,48 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,214,18,0,0 ; vbroadcastss 0x12d6(%rip),%ymm3 # 4da4 <_sk_callback_hsw+0x42a> - DB 255,224 ; jmpq *%rax - DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 - DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 3ae9 <_sk_load_rgb_u16_be_hsw+0xf7> - DB 233,79,255,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> - DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1 - DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 3b18 <_sk_load_rgb_u16_be_hsw+0x126> - DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1 - DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 3b1d <_sk_load_rgb_u16_be_hsw+0x12b> - DB 233,32,255,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,27,255,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> - DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1 - DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 3b4c <_sk_load_rgb_u16_be_hsw+0x15a> - DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1 - DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 3b51 <_sk_load_rgb_u16_be_hsw+0x15f> - DB 233,236,254,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,231,254,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> - DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3 - DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 3b7a <_sk_load_rgb_u16_be_hsw+0x188> - DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3 - DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 233,190,254,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,185,254,255,255 ; jmpq 3a38 <_sk_load_rgb_u16_be_hsw+0x46> + DB 196,226,125,24,29,226,18,0,0 ; vbroadcastss 0x12e2(%rip),%ymm3 # 4da0 <_sk_callback_hsw+0x434> + DB 255,224 ; jmpq *%rax + DB 196,193,121,110,4,65 ; vmovd (%r9,%rax,2),%xmm0 + DB 196,193,121,196,68,65,4,2 ; vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,5 ; jne 3ad9 <_sk_load_rgb_u16_be_hsw+0xf7> + DB 233,79,255,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> + DB 196,193,121,110,76,65,6 ; vmovd 0x6(%r9,%rax,2),%xmm1 + DB 196,65,113,196,68,65,10,2 ; vpinsrw $0x2,0xa(%r9,%rax,2),%xmm1,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,26 ; jb 3b08 <_sk_load_rgb_u16_be_hsw+0x126> + DB 196,193,121,110,76,65,12 ; vmovd 0xc(%r9,%rax,2),%xmm1 + DB 196,193,113,196,84,65,16,2 ; vpinsrw $0x2,0x10(%r9,%rax,2),%xmm1,%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 117,10 ; jne 3b0d <_sk_load_rgb_u16_be_hsw+0x12b> + DB 233,32,255,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,27,255,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> + DB 196,193,121,110,76,65,18 ; vmovd 0x12(%r9,%rax,2),%xmm1 + DB 196,65,113,196,76,65,22,2 ; vpinsrw $0x2,0x16(%r9,%rax,2),%xmm1,%xmm9 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,26 ; jb 3b3c <_sk_load_rgb_u16_be_hsw+0x15a> + DB 196,193,121,110,76,65,24 ; vmovd 0x18(%r9,%rax,2),%xmm1 + DB 196,193,113,196,76,65,28,2 ; vpinsrw $0x2,0x1c(%r9,%rax,2),%xmm1,%xmm1 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 117,10 ; jne 3b41 <_sk_load_rgb_u16_be_hsw+0x15f> + DB 233,236,254,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,231,254,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> + DB 196,193,121,110,92,65,30 ; vmovd 0x1e(%r9,%rax,2),%xmm3 + DB 196,65,97,196,92,65,34,2 ; vpinsrw $0x2,0x22(%r9,%rax,2),%xmm3,%xmm11 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,20 ; jb 3b6a <_sk_load_rgb_u16_be_hsw+0x188> + DB 196,193,121,110,92,65,36 ; vmovd 0x24(%r9,%rax,2),%xmm3 + DB 196,193,97,196,92,65,40,2 ; vpinsrw $0x2,0x28(%r9,%rax,2),%xmm3,%xmm3 + DB 233,190,254,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,185,254,255,255 ; jmpq 3a28 <_sk_load_rgb_u16_be_hsw+0x46> PUBLIC _sk_store_u16_be_hsw _sk_store_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 196,98,125,24,5,19,18,0,0 ; vbroadcastss 0x1213(%rip),%ymm8 # 4da8 <_sk_callback_hsw+0x42e> + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 196,98,125,24,5,31,18,0,0 ; vbroadcastss 0x121f(%rip),%ymm8 # 4da4 <_sk_callback_hsw+0x438> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10 @@ -3633,55 +3631,55 @@ _sk_store_u16_be_hsw LABEL PROC DB 196,65,25,106,209 ; vpunpckhdq %xmm9,%xmm12,%xmm10 DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 3c79 <_sk_store_u16_be_hsw+0xfa> - DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2) - DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2) - DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2) - DB 196,65,122,127,68,64,48 ; vmovdqu %xmm8,0x30(%r8,%rax,2) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 3c75 <_sk_store_u16_be_hsw+0xf6> - DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 3c75 <_sk_store_u16_be_hsw+0xf6> - DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je 3c75 <_sk_store_u16_be_hsw+0xf6> - DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2) - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 3c75 <_sk_store_u16_be_hsw+0xf6> - DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je 3c75 <_sk_store_u16_be_hsw+0xf6> - DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2) - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 3c75 <_sk_store_u16_be_hsw+0xf6> - DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp 3c75 <_sk_store_u16_be_hsw+0xf6> + DB 77,133,192 ; test %r8,%r8 + DB 117,31 ; jne 3c69 <_sk_store_u16_be_hsw+0xfa> + DB 196,65,120,17,28,65 ; vmovups %xmm11,(%r9,%rax,2) + DB 196,65,120,17,84,65,16 ; vmovups %xmm10,0x10(%r9,%rax,2) + DB 196,65,120,17,76,65,32 ; vmovups %xmm9,0x20(%r9,%rax,2) + DB 196,65,122,127,68,65,48 ; vmovdqu %xmm8,0x30(%r9,%rax,2) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 196,65,121,214,28,65 ; vmovq %xmm11,(%r9,%rax,2) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 3c65 <_sk_store_u16_be_hsw+0xf6> + DB 196,65,121,23,92,65,8 ; vmovhpd %xmm11,0x8(%r9,%rax,2) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 3c65 <_sk_store_u16_be_hsw+0xf6> + DB 196,65,121,214,84,65,16 ; vmovq %xmm10,0x10(%r9,%rax,2) + DB 116,218 ; je 3c65 <_sk_store_u16_be_hsw+0xf6> + DB 196,65,121,23,84,65,24 ; vmovhpd %xmm10,0x18(%r9,%rax,2) + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,205 ; jb 3c65 <_sk_store_u16_be_hsw+0xf6> + DB 196,65,121,214,76,65,32 ; vmovq %xmm9,0x20(%r9,%rax,2) + DB 116,196 ; je 3c65 <_sk_store_u16_be_hsw+0xf6> + DB 196,65,121,23,76,65,40 ; vmovhpd %xmm9,0x28(%r9,%rax,2) + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,183 ; jb 3c65 <_sk_store_u16_be_hsw+0xf6> + DB 196,65,121,214,68,65,48 ; vmovq %xmm8,0x30(%r9,%rax,2) + DB 235,174 ; jmp 3c65 <_sk_store_u16_be_hsw+0xf6> PUBLIC _sk_load_f32_hsw _sk_load_f32_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 119,110 ; ja 3d3d <_sk_load_f32_hsw+0x76> - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 3d68 <_sk_load_f32_hsw+0xa1> - DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax - DB 76,1,208 ; add %r10,%rax - DB 255,224 ; jmpq *%rax - DB 196,3,125,24,68,136,112,1 ; vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8 - DB 196,131,125,24,92,136,96,1 ; vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3 - DB 196,131,125,24,76,136,80,1 ; vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1 - DB 196,131,125,24,84,136,64,1 ; vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2 - DB 196,129,121,16,68,136,48 ; vmovupd 0x30(%r8,%r9,4),%xmm0 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 119,110 ; ja 3d2d <_sk_load_f32_hsw+0x76> + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,141,29,135,0,0,0 ; lea 0x87(%rip),%r11 # 3d58 <_sk_load_f32_hsw+0xa1> + DB 75,99,4,131 ; movslq (%r11,%r8,4),%rax + DB 76,1,216 ; add %r11,%rax + DB 255,224 ; jmpq *%rax + DB 196,3,125,24,68,145,112,1 ; vinsertf128 $0x1,0x70(%r9,%r10,4),%ymm0,%ymm8 + DB 196,131,125,24,92,145,96,1 ; vinsertf128 $0x1,0x60(%r9,%r10,4),%ymm0,%ymm3 + DB 196,131,125,24,76,145,80,1 ; vinsertf128 $0x1,0x50(%r9,%r10,4),%ymm0,%ymm1 + DB 196,131,125,24,84,145,64,1 ; vinsertf128 $0x1,0x40(%r9,%r10,4),%ymm0,%ymm2 + DB 196,129,121,16,68,145,48 ; vmovupd 0x30(%r9,%r10,4),%xmm0 DB 196,195,125,13,192,12 ; vblendpd $0xc,%ymm8,%ymm0,%ymm0 - DB 196,1,121,16,68,136,32 ; vmovupd 0x20(%r8,%r9,4),%xmm8 + DB 196,1,121,16,68,145,32 ; vmovupd 0x20(%r9,%r10,4),%xmm8 DB 196,99,61,13,203,12 ; vblendpd $0xc,%ymm3,%ymm8,%ymm9 - DB 196,129,121,16,92,136,16 ; vmovupd 0x10(%r8,%r9,4),%xmm3 + DB 196,129,121,16,92,145,16 ; vmovupd 0x10(%r9,%r10,4),%xmm3 DB 196,99,101,13,209,12 ; vblendpd $0xc,%ymm1,%ymm3,%ymm10 - DB 196,129,121,16,12,136 ; vmovupd (%r8,%r9,4),%xmm1 + DB 196,129,121,16,12,145 ; vmovupd (%r9,%r10,4),%xmm1 DB 196,227,117,13,202,12 ; vblendpd $0xc,%ymm2,%ymm1,%ymm1 DB 196,193,116,20,210 ; vunpcklps %ymm10,%ymm1,%ymm2 DB 196,193,116,21,218 ; vunpckhps %ymm10,%ymm1,%ymm3 @@ -3715,8 +3713,8 @@ _sk_load_f32_hsw LABEL PROC PUBLIC _sk_store_f32_hsw _sk_store_f32_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax DB 197,124,20,193 ; vunpcklps %ymm1,%ymm0,%ymm8 DB 197,124,21,217 ; vunpckhps %ymm1,%ymm0,%ymm11 DB 197,108,20,203 ; vunpcklps %ymm3,%ymm2,%ymm9 @@ -3725,36 +3723,36 @@ _sk_store_f32_hsw LABEL PROC DB 196,65,61,21,201 ; vunpckhpd %ymm9,%ymm8,%ymm9 DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 - DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 3df5 <_sk_store_f32_hsw+0x6d> + DB 77,133,192 ; test %r8,%r8 + DB 117,55 ; jne 3de5 <_sk_store_f32_hsw+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 DB 196,67,61,6,195,49 ; vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 - DB 196,65,125,17,36,128 ; vmovupd %ymm12,(%r8,%rax,4) - DB 196,65,125,17,108,128,32 ; vmovupd %ymm13,0x20(%r8,%rax,4) - DB 196,65,125,17,76,128,64 ; vmovupd %ymm9,0x40(%r8,%rax,4) - DB 196,65,125,17,68,128,96 ; vmovupd %ymm8,0x60(%r8,%rax,4) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 3df1 <_sk_store_f32_hsw+0x69> - DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 3df1 <_sk_store_f32_hsw+0x69> - DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 3df1 <_sk_store_f32_hsw+0x69> - DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 3df1 <_sk_store_f32_hsw+0x69> - DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 3df1 <_sk_store_f32_hsw+0x69> - DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 3df1 <_sk_store_f32_hsw+0x69> - DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 3df1 <_sk_store_f32_hsw+0x69> + DB 196,65,125,17,36,129 ; vmovupd %ymm12,(%r9,%rax,4) + DB 196,65,125,17,108,129,32 ; vmovupd %ymm13,0x20(%r9,%rax,4) + DB 196,65,125,17,76,129,64 ; vmovupd %ymm9,0x40(%r9,%rax,4) + DB 196,65,125,17,68,129,96 ; vmovupd %ymm8,0x60(%r9,%rax,4) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 196,65,121,17,20,129 ; vmovupd %xmm10,(%r9,%rax,4) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 3de1 <_sk_store_f32_hsw+0x69> + DB 196,65,121,17,76,129,16 ; vmovupd %xmm9,0x10(%r9,%rax,4) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 3de1 <_sk_store_f32_hsw+0x69> + DB 196,65,121,17,68,129,32 ; vmovupd %xmm8,0x20(%r9,%rax,4) + DB 116,218 ; je 3de1 <_sk_store_f32_hsw+0x69> + DB 196,65,121,17,92,129,48 ; vmovupd %xmm11,0x30(%r9,%rax,4) + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,205 ; jb 3de1 <_sk_store_f32_hsw+0x69> + DB 196,67,125,25,84,129,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r9,%rax,4) + DB 116,195 ; je 3de1 <_sk_store_f32_hsw+0x69> + DB 196,67,125,25,76,129,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r9,%rax,4) + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,181 ; jb 3de1 <_sk_store_f32_hsw+0x69> + DB 196,67,125,25,68,129,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r9,%rax,4) + DB 235,171 ; jmp 3de1 <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -3838,7 +3836,7 @@ PUBLIC _sk_clamp_x_1_hsw _sk_clamp_x_1_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,102,14,0,0 ; vbroadcastss 0xe66(%rip),%ymm8 # 4dac <_sk_callback_hsw+0x432> + DB 196,98,125,24,5,114,14,0,0 ; vbroadcastss 0xe72(%rip),%ymm8 # 4da8 <_sk_callback_hsw+0x43c> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3852,9 +3850,9 @@ _sk_repeat_x_1_hsw LABEL PROC PUBLIC _sk_mirror_x_1_hsw _sk_mirror_x_1_hsw LABEL PROC - DB 196,98,125,24,5,73,14,0,0 ; vbroadcastss 0xe49(%rip),%ymm8 # 4db0 <_sk_callback_hsw+0x436> + DB 196,98,125,24,5,85,14,0,0 ; vbroadcastss 0xe55(%rip),%ymm8 # 4dac <_sk_callback_hsw+0x440> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,63,14,0,0 ; vbroadcastss 0xe3f(%rip),%ymm9 # 4db4 <_sk_callback_hsw+0x43a> + DB 196,98,125,24,13,75,14,0,0 ; vbroadcastss 0xe4b(%rip),%ymm9 # 4db0 <_sk_callback_hsw+0x444> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -3868,11 +3866,11 @@ _sk_mirror_x_1_hsw LABEL PROC PUBLIC _sk_luminance_to_alpha_hsw _sk_luminance_to_alpha_hsw LABEL PROC - DB 196,226,125,24,29,15,14,0,0 ; vbroadcastss 0xe0f(%rip),%ymm3 # 4db8 <_sk_callback_hsw+0x43e> - DB 196,98,125,24,5,10,14,0,0 ; vbroadcastss 0xe0a(%rip),%ymm8 # 4dbc <_sk_callback_hsw+0x442> + DB 196,226,125,24,29,27,14,0,0 ; vbroadcastss 0xe1b(%rip),%ymm3 # 4db4 <_sk_callback_hsw+0x448> + DB 196,98,125,24,5,22,14,0,0 ; vbroadcastss 0xe16(%rip),%ymm8 # 4db8 <_sk_callback_hsw+0x44c> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1 - DB 196,226,125,24,29,251,13,0,0 ; vbroadcastss 0xdfb(%rip),%ymm3 # 4dc0 <_sk_callback_hsw+0x446> + DB 196,226,125,24,29,7,14,0,0 ; vbroadcastss 0xe07(%rip),%ymm3 # 4dbc <_sk_callback_hsw+0x450> DB 196,226,109,168,217 ; vfmadd213ps %ymm1,%ymm2,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -4028,62 +4026,63 @@ _sk_matrix_perspective_hsw LABEL PROC PUBLIC _sk_evenly_spaced_gradient_hsw _sk_evenly_spaced_gradient_hsw LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 77,137,202 ; mov %r9,%r10 - DB 73,255,202 ; dec %r10 - DB 120,7 ; js 427d <_sk_evenly_spaced_gradient_hsw+0x18> - DB 196,193,242,42,202 ; vcvtsi2ss %r10,%xmm1,%xmm1 - DB 235,22 ; jmp 4293 <_sk_evenly_spaced_gradient_hsw+0x2e> + DB 76,139,16 ; mov (%rax),%r10 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 DB 77,137,211 ; mov %r10,%r11 - DB 73,209,235 ; shr %r11 - DB 65,131,226,1 ; and $0x1,%r10d - DB 77,9,218 ; or %r11,%r10 - DB 196,193,242,42,202 ; vcvtsi2ss %r10,%xmm1,%xmm1 + DB 73,255,203 ; dec %r11 + DB 120,7 ; js 426e <_sk_evenly_spaced_gradient_hsw+0x19> + DB 196,193,242,42,203 ; vcvtsi2ss %r11,%xmm1,%xmm1 + DB 235,22 ; jmp 4284 <_sk_evenly_spaced_gradient_hsw+0x2f> + DB 76,137,219 ; mov %r11,%rbx + DB 72,209,235 ; shr %rbx + DB 65,131,227,1 ; and $0x1,%r11d + DB 73,9,219 ; or %rbx,%r11 + DB 196,193,242,42,203 ; vcvtsi2ss %r11,%xmm1,%xmm1 DB 197,242,88,201 ; vaddss %xmm1,%xmm1,%xmm1 DB 196,226,125,24,201 ; vbroadcastss %xmm1,%ymm1 DB 197,244,89,200 ; vmulps %ymm0,%ymm1,%ymm1 DB 197,126,91,217 ; vcvttps2dq %ymm1,%ymm11 - DB 73,131,249,8 ; cmp $0x8,%r9 - DB 119,70 ; ja 42ec <_sk_evenly_spaced_gradient_hsw+0x87> - DB 196,66,37,22,0 ; vpermps (%r8),%ymm11,%ymm8 - DB 76,139,64,40 ; mov 0x28(%rax),%r8 - DB 196,66,37,22,8 ; vpermps (%r8),%ymm11,%ymm9 - DB 76,139,64,16 ; mov 0x10(%rax),%r8 + DB 73,131,250,8 ; cmp $0x8,%r10 + DB 119,70 ; ja 42dd <_sk_evenly_spaced_gradient_hsw+0x88> + DB 196,66,37,22,1 ; vpermps (%r9),%ymm11,%ymm8 + DB 72,139,88,40 ; mov 0x28(%rax),%rbx + DB 196,98,37,22,11 ; vpermps (%rbx),%ymm11,%ymm9 + DB 72,139,88,16 ; mov 0x10(%rax),%rbx DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 196,194,37,22,8 ; vpermps (%r8),%ymm11,%ymm1 - DB 76,139,64,48 ; mov 0x30(%rax),%r8 - DB 196,66,37,22,16 ; vpermps (%r8),%ymm11,%ymm10 + DB 196,226,37,22,11 ; vpermps (%rbx),%ymm11,%ymm1 + DB 72,139,88,48 ; mov 0x30(%rax),%rbx + DB 196,98,37,22,19 ; vpermps (%rbx),%ymm11,%ymm10 DB 196,194,37,22,17 ; vpermps (%r9),%ymm11,%ymm2 - DB 76,139,64,56 ; mov 0x38(%rax),%r8 - DB 196,66,37,22,32 ; vpermps (%r8),%ymm11,%ymm12 - DB 76,139,64,32 ; mov 0x20(%rax),%r8 - DB 196,194,37,22,24 ; vpermps (%r8),%ymm11,%ymm3 + DB 72,139,88,56 ; mov 0x38(%rax),%rbx + DB 196,98,37,22,35 ; vpermps (%rbx),%ymm11,%ymm12 + DB 72,139,88,32 ; mov 0x20(%rax),%rbx + DB 196,226,37,22,27 ; vpermps (%rbx),%ymm11,%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 196,98,37,22,40 ; vpermps (%rax),%ymm11,%ymm13 - DB 235,110 ; jmp 435a <_sk_evenly_spaced_gradient_hsw+0xf5> + DB 235,110 ; jmp 434b <_sk_evenly_spaced_gradient_hsw+0xf6> DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,2,117,146,4,152 ; vgatherdps %ymm1,(%r8,%ymm11,4),%ymm8 - DB 76,139,64,40 ; mov 0x28(%rax),%r8 + DB 196,2,117,146,4,153 ; vgatherdps %ymm1,(%r9,%ymm11,4),%ymm8 + DB 72,139,88,40 ; mov 0x28(%rax),%rbx DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,2,117,146,12,152 ; vgatherdps %ymm1,(%r8,%ymm11,4),%ymm9 - DB 76,139,64,16 ; mov 0x10(%rax),%r8 + DB 196,34,117,146,12,155 ; vgatherdps %ymm1,(%rbx,%ymm11,4),%ymm9 + DB 72,139,88,16 ; mov 0x10(%rax),%rbx DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,130,109,146,12,152 ; vgatherdps %ymm2,(%r8,%ymm11,4),%ymm1 - DB 76,139,64,48 ; mov 0x30(%rax),%r8 + DB 196,162,109,146,12,155 ; vgatherdps %ymm2,(%rbx,%ymm11,4),%ymm1 + DB 72,139,88,48 ; mov 0x30(%rax),%rbx DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,2,109,146,20,152 ; vgatherdps %ymm2,(%r8,%ymm11,4),%ymm10 + DB 196,34,109,146,20,155 ; vgatherdps %ymm2,(%rbx,%ymm11,4),%ymm10 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 DB 196,130,101,146,20,153 ; vgatherdps %ymm3,(%r9,%ymm11,4),%ymm2 - DB 76,139,64,56 ; mov 0x38(%rax),%r8 + DB 72,139,88,56 ; mov 0x38(%rax),%rbx DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,2,101,146,36,152 ; vgatherdps %ymm3,(%r8,%ymm11,4),%ymm12 - DB 76,139,64,32 ; mov 0x20(%rax),%r8 + DB 196,34,101,146,36,155 ; vgatherdps %ymm3,(%rbx,%ymm11,4),%ymm12 + DB 72,139,88,32 ; mov 0x20(%rax),%rbx DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 - DB 196,130,21,146,28,152 ; vgatherdps %ymm13,(%r8,%ymm11,4),%ymm3 + DB 196,162,21,146,28,155 ; vgatherdps %ymm13,(%rbx,%ymm11,4),%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 196,34,13,146,44,152 ; vgatherdps %ymm14,(%rax,%ymm11,4),%ymm13 DB 196,66,125,168,193 ; vfmadd213ps %ymm9,%ymm0,%ymm8 @@ -4092,18 +4091,19 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC DB 196,194,125,168,221 ; vfmadd213ps %ymm13,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,124,41,192 ; vmovaps %ymm8,%ymm0 + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_gauss_a_to_rgba_hsw _sk_gauss_a_to_rgba_hsw LABEL PROC - DB 196,226,125,24,5,69,10,0,0 ; vbroadcastss 0xa45(%rip),%ymm0 # 4dc4 <_sk_callback_hsw+0x44a> - DB 196,226,125,24,13,64,10,0,0 ; vbroadcastss 0xa40(%rip),%ymm1 # 4dc8 <_sk_callback_hsw+0x44e> + DB 196,226,125,24,5,79,10,0,0 ; vbroadcastss 0xa4f(%rip),%ymm0 # 4dc0 <_sk_callback_hsw+0x454> + DB 196,226,125,24,13,74,10,0,0 ; vbroadcastss 0xa4a(%rip),%ymm1 # 4dc4 <_sk_callback_hsw+0x458> DB 196,226,101,168,200 ; vfmadd213ps %ymm0,%ymm3,%ymm1 - DB 196,226,125,24,5,54,10,0,0 ; vbroadcastss 0xa36(%rip),%ymm0 # 4dcc <_sk_callback_hsw+0x452> + DB 196,226,125,24,5,64,10,0,0 ; vbroadcastss 0xa40(%rip),%ymm0 # 4dc8 <_sk_callback_hsw+0x45c> DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0 - DB 196,226,125,24,13,44,10,0,0 ; vbroadcastss 0xa2c(%rip),%ymm1 # 4dd0 <_sk_callback_hsw+0x456> + DB 196,226,125,24,13,54,10,0,0 ; vbroadcastss 0xa36(%rip),%ymm1 # 4dcc <_sk_callback_hsw+0x460> DB 196,226,101,184,200 ; vfmadd231ps %ymm0,%ymm3,%ymm1 - DB 196,226,125,24,5,34,10,0,0 ; vbroadcastss 0xa22(%rip),%ymm0 # 4dd4 <_sk_callback_hsw+0x45a> + DB 196,226,125,24,5,44,10,0,0 ; vbroadcastss 0xa2c(%rip),%ymm0 # 4dd0 <_sk_callback_hsw+0x464> DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -4114,63 +4114,63 @@ _sk_gauss_a_to_rgba_hsw LABEL PROC PUBLIC _sk_gradient_hsw _sk_gradient_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 73,131,248,1 ; cmp $0x1,%r8 - DB 15,134,180,0,0,0 ; jbe 448a <_sk_gradient_hsw+0xc3> - DB 76,139,72,72 ; mov 0x48(%rax),%r9 + DB 76,139,8 ; mov (%rax),%r9 + DB 73,131,249,1 ; cmp $0x1,%r9 + DB 15,134,180,0,0,0 ; jbe 447c <_sk_gradient_hsw+0xc3> + DB 76,139,80,72 ; mov 0x48(%rax),%r10 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 - DB 65,186,1,0,0,0 ; mov $0x1,%r10d - DB 196,226,125,24,21,235,9,0,0 ; vbroadcastss 0x9eb(%rip),%ymm2 # 4dd8 <_sk_callback_hsw+0x45e> + DB 65,187,1,0,0,0 ; mov $0x1,%r11d + DB 196,226,125,24,21,245,9,0,0 ; vbroadcastss 0x9f5(%rip),%ymm2 # 4dd4 <_sk_callback_hsw+0x468> DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9 - DB 196,130,125,24,28,145 ; vbroadcastss (%r9,%r10,4),%ymm3 + DB 196,130,125,24,28,154 ; vbroadcastss (%r10,%r11,4),%ymm3 DB 197,228,194,216,2 ; vcmpleps %ymm0,%ymm3,%ymm3 DB 196,227,117,74,218,48 ; vblendvps %ymm3,%ymm2,%ymm1,%ymm3 DB 196,65,101,254,201 ; vpaddd %ymm9,%ymm3,%ymm9 - DB 73,255,194 ; inc %r10 - DB 77,57,208 ; cmp %r10,%r8 - DB 117,226 ; jne 43f2 <_sk_gradient_hsw+0x2b> - DB 76,139,72,8 ; mov 0x8(%rax),%r9 - DB 73,131,248,8 ; cmp $0x8,%r8 - DB 118,121 ; jbe 4493 <_sk_gradient_hsw+0xcc> + DB 73,255,195 ; inc %r11 + DB 77,57,217 ; cmp %r11,%r9 + DB 117,226 ; jne 43e4 <_sk_gradient_hsw+0x2b> + DB 76,139,80,8 ; mov 0x8(%rax),%r10 + DB 73,131,249,8 ; cmp $0x8,%r9 + DB 118,121 ; jbe 4485 <_sk_gradient_hsw+0xcc> DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,2,117,146,4,137 ; vgatherdps %ymm1,(%r9,%ymm9,4),%ymm8 - DB 76,139,64,40 ; mov 0x28(%rax),%r8 + DB 196,2,117,146,4,138 ; vgatherdps %ymm1,(%r10,%ymm9,4),%ymm8 + DB 76,139,72,40 ; mov 0x28(%rax),%r9 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,2,117,146,20,136 ; vgatherdps %ymm1,(%r8,%ymm9,4),%ymm10 - DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 + DB 196,2,117,146,20,137 ; vgatherdps %ymm1,(%r9,%ymm9,4),%ymm10 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,130,109,146,12,136 ; vgatherdps %ymm2,(%r8,%ymm9,4),%ymm1 - DB 76,139,64,48 ; mov 0x30(%rax),%r8 + DB 196,130,109,146,12,137 ; vgatherdps %ymm2,(%r9,%ymm9,4),%ymm1 + DB 76,139,72,48 ; mov 0x30(%rax),%r9 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,2,109,146,28,136 ; vgatherdps %ymm2,(%r8,%ymm9,4),%ymm11 + DB 196,2,109,146,28,137 ; vgatherdps %ymm2,(%r9,%ymm9,4),%ymm11 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,130,101,146,20,137 ; vgatherdps %ymm3,(%r9,%ymm9,4),%ymm2 - DB 76,139,64,56 ; mov 0x38(%rax),%r8 + DB 196,130,101,146,20,138 ; vgatherdps %ymm3,(%r10,%ymm9,4),%ymm2 + DB 76,139,72,56 ; mov 0x38(%rax),%r9 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,2,101,146,36,136 ; vgatherdps %ymm3,(%r8,%ymm9,4),%ymm12 - DB 76,139,64,32 ; mov 0x20(%rax),%r8 + DB 196,2,101,146,36,137 ; vgatherdps %ymm3,(%r9,%ymm9,4),%ymm12 + DB 76,139,72,32 ; mov 0x20(%rax),%r9 DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 - DB 196,130,21,146,28,136 ; vgatherdps %ymm13,(%r8,%ymm9,4),%ymm3 + DB 196,130,21,146,28,137 ; vgatherdps %ymm13,(%r9,%ymm9,4),%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 196,34,13,146,44,136 ; vgatherdps %ymm14,(%rax,%ymm9,4),%ymm13 - DB 235,77 ; jmp 44d7 <_sk_gradient_hsw+0x110> - DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 235,77 ; jmp 44c9 <_sk_gradient_hsw+0x110> + DB 76,139,80,8 ; mov 0x8(%rax),%r10 DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 - DB 196,66,53,22,1 ; vpermps (%r9),%ymm9,%ymm8 - DB 76,139,64,40 ; mov 0x28(%rax),%r8 - DB 196,66,53,22,16 ; vpermps (%r8),%ymm9,%ymm10 - DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 196,194,53,22,8 ; vpermps (%r8),%ymm9,%ymm1 - DB 76,139,64,48 ; mov 0x30(%rax),%r8 - DB 196,66,53,22,24 ; vpermps (%r8),%ymm9,%ymm11 - DB 196,194,53,22,17 ; vpermps (%r9),%ymm9,%ymm2 - DB 76,139,64,56 ; mov 0x38(%rax),%r8 - DB 196,66,53,22,32 ; vpermps (%r8),%ymm9,%ymm12 - DB 76,139,64,32 ; mov 0x20(%rax),%r8 - DB 196,194,53,22,24 ; vpermps (%r8),%ymm9,%ymm3 + DB 196,66,53,22,2 ; vpermps (%r10),%ymm9,%ymm8 + DB 76,139,72,40 ; mov 0x28(%rax),%r9 + DB 196,66,53,22,17 ; vpermps (%r9),%ymm9,%ymm10 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 + DB 196,194,53,22,9 ; vpermps (%r9),%ymm9,%ymm1 + DB 76,139,72,48 ; mov 0x30(%rax),%r9 + DB 196,66,53,22,25 ; vpermps (%r9),%ymm9,%ymm11 + DB 196,194,53,22,18 ; vpermps (%r10),%ymm9,%ymm2 + DB 76,139,72,56 ; mov 0x38(%rax),%r9 + DB 196,66,53,22,33 ; vpermps (%r9),%ymm9,%ymm12 + DB 76,139,72,32 ; mov 0x20(%rax),%r9 + DB 196,194,53,22,25 ; vpermps (%r9),%ymm9,%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 196,98,53,22,40 ; vpermps (%rax),%ymm9,%ymm13 DB 196,66,125,168,194 ; vfmadd213ps %ymm10,%ymm0,%ymm8 @@ -4211,24 +4211,24 @@ _sk_xy_to_unit_angle_hsw LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,106,8,0,0 ; vbroadcastss 0x86a(%rip),%ymm13 # 4ddc <_sk_callback_hsw+0x462> - DB 196,98,125,24,53,101,8,0,0 ; vbroadcastss 0x865(%rip),%ymm14 # 4de0 <_sk_callback_hsw+0x466> + DB 196,98,125,24,45,116,8,0,0 ; vbroadcastss 0x874(%rip),%ymm13 # 4dd8 <_sk_callback_hsw+0x46c> + DB 196,98,125,24,53,111,8,0,0 ; vbroadcastss 0x86f(%rip),%ymm14 # 4ddc <_sk_callback_hsw+0x470> DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14 - DB 196,98,125,24,45,91,8,0,0 ; vbroadcastss 0x85b(%rip),%ymm13 # 4de4 <_sk_callback_hsw+0x46a> + DB 196,98,125,24,45,101,8,0,0 ; vbroadcastss 0x865(%rip),%ymm13 # 4de0 <_sk_callback_hsw+0x474> DB 196,66,29,184,238 ; vfmadd231ps %ymm14,%ymm12,%ymm13 - DB 196,98,125,24,53,81,8,0,0 ; vbroadcastss 0x851(%rip),%ymm14 # 4de8 <_sk_callback_hsw+0x46e> + DB 196,98,125,24,53,91,8,0,0 ; vbroadcastss 0x85b(%rip),%ymm14 # 4de4 <_sk_callback_hsw+0x478> DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14 DB 196,65,36,89,222 ; vmulps %ymm14,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,60,8,0,0 ; vbroadcastss 0x83c(%rip),%ymm10 # 4dec <_sk_callback_hsw+0x472> + DB 196,98,125,24,21,70,8,0,0 ; vbroadcastss 0x846(%rip),%ymm10 # 4de8 <_sk_callback_hsw+0x47c> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,38,8,0,0 ; vbroadcastss 0x826(%rip),%ymm10 # 4df0 <_sk_callback_hsw+0x476> + DB 196,98,125,24,21,48,8,0,0 ; vbroadcastss 0x830(%rip),%ymm10 # 4dec <_sk_callback_hsw+0x480> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,16,8,0,0 ; vbroadcastss 0x810(%rip),%ymm10 # 4df4 <_sk_callback_hsw+0x47a> + DB 196,98,125,24,21,26,8,0,0 ; vbroadcastss 0x81a(%rip),%ymm10 # 4df0 <_sk_callback_hsw+0x484> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -4247,7 +4247,7 @@ _sk_xy_to_radius_hsw LABEL PROC PUBLIC _sk_save_xy_hsw _sk_save_xy_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,221,7,0,0 ; vbroadcastss 0x7dd(%rip),%ymm8 # 4df8 <_sk_callback_hsw+0x47e> + DB 196,98,125,24,5,231,7,0,0 ; vbroadcastss 0x7e7(%rip),%ymm8 # 4df4 <_sk_callback_hsw+0x488> DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9 DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 @@ -4277,9 +4277,9 @@ _sk_accumulate_hsw LABEL PROC PUBLIC _sk_bilinear_nx_hsw _sk_bilinear_nx_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,113,7,0,0 ; vbroadcastss 0x771(%rip),%ymm0 # 4dfc <_sk_callback_hsw+0x482> + DB 196,226,125,24,5,123,7,0,0 ; vbroadcastss 0x77b(%rip),%ymm0 # 4df8 <_sk_callback_hsw+0x48c> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,104,7,0,0 ; vbroadcastss 0x768(%rip),%ymm8 # 4e00 <_sk_callback_hsw+0x486> + DB 196,98,125,24,5,114,7,0,0 ; vbroadcastss 0x772(%rip),%ymm8 # 4dfc <_sk_callback_hsw+0x490> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -4288,7 +4288,7 @@ _sk_bilinear_nx_hsw LABEL PROC PUBLIC _sk_bilinear_px_hsw _sk_bilinear_px_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,80,7,0,0 ; vbroadcastss 0x750(%rip),%ymm0 # 4e04 <_sk_callback_hsw+0x48a> + DB 196,226,125,24,5,90,7,0,0 ; vbroadcastss 0x75a(%rip),%ymm0 # 4e00 <_sk_callback_hsw+0x494> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -4298,9 +4298,9 @@ _sk_bilinear_px_hsw LABEL PROC PUBLIC _sk_bilinear_ny_hsw _sk_bilinear_ny_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,52,7,0,0 ; vbroadcastss 0x734(%rip),%ymm1 # 4e08 <_sk_callback_hsw+0x48e> + DB 196,226,125,24,13,62,7,0,0 ; vbroadcastss 0x73e(%rip),%ymm1 # 4e04 <_sk_callback_hsw+0x498> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,42,7,0,0 ; vbroadcastss 0x72a(%rip),%ymm8 # 4e0c <_sk_callback_hsw+0x492> + DB 196,98,125,24,5,52,7,0,0 ; vbroadcastss 0x734(%rip),%ymm8 # 4e08 <_sk_callback_hsw+0x49c> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -4309,7 +4309,7 @@ _sk_bilinear_ny_hsw LABEL PROC PUBLIC _sk_bilinear_py_hsw _sk_bilinear_py_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,18,7,0,0 ; vbroadcastss 0x712(%rip),%ymm1 # 4e10 <_sk_callback_hsw+0x496> + DB 196,226,125,24,13,28,7,0,0 ; vbroadcastss 0x71c(%rip),%ymm1 # 4e0c <_sk_callback_hsw+0x4a0> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -4319,13 +4319,13 @@ _sk_bilinear_py_hsw LABEL PROC PUBLIC _sk_bicubic_n3x_hsw _sk_bicubic_n3x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,245,6,0,0 ; vbroadcastss 0x6f5(%rip),%ymm0 # 4e14 <_sk_callback_hsw+0x49a> + DB 196,226,125,24,5,255,6,0,0 ; vbroadcastss 0x6ff(%rip),%ymm0 # 4e10 <_sk_callback_hsw+0x4a4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,236,6,0,0 ; vbroadcastss 0x6ec(%rip),%ymm8 # 4e18 <_sk_callback_hsw+0x49e> + DB 196,98,125,24,5,246,6,0,0 ; vbroadcastss 0x6f6(%rip),%ymm8 # 4e14 <_sk_callback_hsw+0x4a8> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,221,6,0,0 ; vbroadcastss 0x6dd(%rip),%ymm10 # 4e1c <_sk_callback_hsw+0x4a2> - DB 196,98,125,24,29,216,6,0,0 ; vbroadcastss 0x6d8(%rip),%ymm11 # 4e20 <_sk_callback_hsw+0x4a6> + DB 196,98,125,24,21,231,6,0,0 ; vbroadcastss 0x6e7(%rip),%ymm10 # 4e18 <_sk_callback_hsw+0x4ac> + DB 196,98,125,24,29,226,6,0,0 ; vbroadcastss 0x6e2(%rip),%ymm11 # 4e1c <_sk_callback_hsw+0x4b0> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,36,89,193 ; vmulps %ymm9,%ymm11,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -4335,16 +4335,16 @@ _sk_bicubic_n3x_hsw LABEL PROC PUBLIC _sk_bicubic_n1x_hsw _sk_bicubic_n1x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,187,6,0,0 ; vbroadcastss 0x6bb(%rip),%ymm0 # 4e24 <_sk_callback_hsw+0x4aa> + DB 196,226,125,24,5,197,6,0,0 ; vbroadcastss 0x6c5(%rip),%ymm0 # 4e20 <_sk_callback_hsw+0x4b4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,178,6,0,0 ; vbroadcastss 0x6b2(%rip),%ymm8 # 4e28 <_sk_callback_hsw+0x4ae> + DB 196,98,125,24,5,188,6,0,0 ; vbroadcastss 0x6bc(%rip),%ymm8 # 4e24 <_sk_callback_hsw+0x4b8> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,168,6,0,0 ; vbroadcastss 0x6a8(%rip),%ymm9 # 4e2c <_sk_callback_hsw+0x4b2> - DB 196,98,125,24,21,163,6,0,0 ; vbroadcastss 0x6a3(%rip),%ymm10 # 4e30 <_sk_callback_hsw+0x4b6> + DB 196,98,125,24,13,178,6,0,0 ; vbroadcastss 0x6b2(%rip),%ymm9 # 4e28 <_sk_callback_hsw+0x4bc> + DB 196,98,125,24,21,173,6,0,0 ; vbroadcastss 0x6ad(%rip),%ymm10 # 4e2c <_sk_callback_hsw+0x4c0> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,153,6,0,0 ; vbroadcastss 0x699(%rip),%ymm9 # 4e34 <_sk_callback_hsw+0x4ba> + DB 196,98,125,24,13,163,6,0,0 ; vbroadcastss 0x6a3(%rip),%ymm9 # 4e30 <_sk_callback_hsw+0x4c4> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,143,6,0,0 ; vbroadcastss 0x68f(%rip),%ymm10 # 4e38 <_sk_callback_hsw+0x4be> + DB 196,98,125,24,21,153,6,0,0 ; vbroadcastss 0x699(%rip),%ymm10 # 4e34 <_sk_callback_hsw+0x4c8> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 DB 197,124,17,144,128,0,0,0 ; vmovups %ymm10,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -4353,14 +4353,14 @@ _sk_bicubic_n1x_hsw LABEL PROC PUBLIC _sk_bicubic_p1x_hsw _sk_bicubic_p1x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,119,6,0,0 ; vbroadcastss 0x677(%rip),%ymm8 # 4e3c <_sk_callback_hsw+0x4c2> + DB 196,98,125,24,5,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm8 # 4e38 <_sk_callback_hsw+0x4cc> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,105,6,0,0 ; vbroadcastss 0x669(%rip),%ymm10 # 4e40 <_sk_callback_hsw+0x4c6> - DB 196,98,125,24,29,100,6,0,0 ; vbroadcastss 0x664(%rip),%ymm11 # 4e44 <_sk_callback_hsw+0x4ca> + DB 196,98,125,24,21,115,6,0,0 ; vbroadcastss 0x673(%rip),%ymm10 # 4e3c <_sk_callback_hsw+0x4d0> + DB 196,98,125,24,29,110,6,0,0 ; vbroadcastss 0x66e(%rip),%ymm11 # 4e40 <_sk_callback_hsw+0x4d4> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,85,6,0,0 ; vbroadcastss 0x655(%rip),%ymm8 # 4e48 <_sk_callback_hsw+0x4ce> + DB 196,98,125,24,5,95,6,0,0 ; vbroadcastss 0x65f(%rip),%ymm8 # 4e44 <_sk_callback_hsw+0x4d8> DB 196,66,53,184,195 ; vfmadd231ps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -4369,12 +4369,12 @@ _sk_bicubic_p1x_hsw LABEL PROC PUBLIC _sk_bicubic_p3x_hsw _sk_bicubic_p3x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,61,6,0,0 ; vbroadcastss 0x63d(%rip),%ymm0 # 4e4c <_sk_callback_hsw+0x4d2> + DB 196,226,125,24,5,71,6,0,0 ; vbroadcastss 0x647(%rip),%ymm0 # 4e48 <_sk_callback_hsw+0x4dc> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,42,6,0,0 ; vbroadcastss 0x62a(%rip),%ymm10 # 4e50 <_sk_callback_hsw+0x4d6> - DB 196,98,125,24,29,37,6,0,0 ; vbroadcastss 0x625(%rip),%ymm11 # 4e54 <_sk_callback_hsw+0x4da> + DB 196,98,125,24,21,52,6,0,0 ; vbroadcastss 0x634(%rip),%ymm10 # 4e4c <_sk_callback_hsw+0x4e0> + DB 196,98,125,24,29,47,6,0,0 ; vbroadcastss 0x62f(%rip),%ymm11 # 4e50 <_sk_callback_hsw+0x4e4> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,52,89,195 ; vmulps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -4384,13 +4384,13 @@ _sk_bicubic_p3x_hsw LABEL PROC PUBLIC _sk_bicubic_n3y_hsw _sk_bicubic_n3y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,8,6,0,0 ; vbroadcastss 0x608(%rip),%ymm1 # 4e58 <_sk_callback_hsw+0x4de> + DB 196,226,125,24,13,18,6,0,0 ; vbroadcastss 0x612(%rip),%ymm1 # 4e54 <_sk_callback_hsw+0x4e8> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,254,5,0,0 ; vbroadcastss 0x5fe(%rip),%ymm8 # 4e5c <_sk_callback_hsw+0x4e2> + DB 196,98,125,24,5,8,6,0,0 ; vbroadcastss 0x608(%rip),%ymm8 # 4e58 <_sk_callback_hsw+0x4ec> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,239,5,0,0 ; vbroadcastss 0x5ef(%rip),%ymm10 # 4e60 <_sk_callback_hsw+0x4e6> - DB 196,98,125,24,29,234,5,0,0 ; vbroadcastss 0x5ea(%rip),%ymm11 # 4e64 <_sk_callback_hsw+0x4ea> + DB 196,98,125,24,21,249,5,0,0 ; vbroadcastss 0x5f9(%rip),%ymm10 # 4e5c <_sk_callback_hsw+0x4f0> + DB 196,98,125,24,29,244,5,0,0 ; vbroadcastss 0x5f4(%rip),%ymm11 # 4e60 <_sk_callback_hsw+0x4f4> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,36,89,193 ; vmulps %ymm9,%ymm11,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -4400,16 +4400,16 @@ _sk_bicubic_n3y_hsw LABEL PROC PUBLIC _sk_bicubic_n1y_hsw _sk_bicubic_n1y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,205,5,0,0 ; vbroadcastss 0x5cd(%rip),%ymm1 # 4e68 <_sk_callback_hsw+0x4ee> + DB 196,226,125,24,13,215,5,0,0 ; vbroadcastss 0x5d7(%rip),%ymm1 # 4e64 <_sk_callback_hsw+0x4f8> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,195,5,0,0 ; vbroadcastss 0x5c3(%rip),%ymm8 # 4e6c <_sk_callback_hsw+0x4f2> + DB 196,98,125,24,5,205,5,0,0 ; vbroadcastss 0x5cd(%rip),%ymm8 # 4e68 <_sk_callback_hsw+0x4fc> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,185,5,0,0 ; vbroadcastss 0x5b9(%rip),%ymm9 # 4e70 <_sk_callback_hsw+0x4f6> - DB 196,98,125,24,21,180,5,0,0 ; vbroadcastss 0x5b4(%rip),%ymm10 # 4e74 <_sk_callback_hsw+0x4fa> + DB 196,98,125,24,13,195,5,0,0 ; vbroadcastss 0x5c3(%rip),%ymm9 # 4e6c <_sk_callback_hsw+0x500> + DB 196,98,125,24,21,190,5,0,0 ; vbroadcastss 0x5be(%rip),%ymm10 # 4e70 <_sk_callback_hsw+0x504> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,170,5,0,0 ; vbroadcastss 0x5aa(%rip),%ymm9 # 4e78 <_sk_callback_hsw+0x4fe> + DB 196,98,125,24,13,180,5,0,0 ; vbroadcastss 0x5b4(%rip),%ymm9 # 4e74 <_sk_callback_hsw+0x508> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,160,5,0,0 ; vbroadcastss 0x5a0(%rip),%ymm10 # 4e7c <_sk_callback_hsw+0x502> + DB 196,98,125,24,21,170,5,0,0 ; vbroadcastss 0x5aa(%rip),%ymm10 # 4e78 <_sk_callback_hsw+0x50c> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 DB 197,124,17,144,160,0,0,0 ; vmovups %ymm10,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -4418,14 +4418,14 @@ _sk_bicubic_n1y_hsw LABEL PROC PUBLIC _sk_bicubic_p1y_hsw _sk_bicubic_p1y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,136,5,0,0 ; vbroadcastss 0x588(%rip),%ymm8 # 4e80 <_sk_callback_hsw+0x506> + DB 196,98,125,24,5,146,5,0,0 ; vbroadcastss 0x592(%rip),%ymm8 # 4e7c <_sk_callback_hsw+0x510> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,121,5,0,0 ; vbroadcastss 0x579(%rip),%ymm10 # 4e84 <_sk_callback_hsw+0x50a> - DB 196,98,125,24,29,116,5,0,0 ; vbroadcastss 0x574(%rip),%ymm11 # 4e88 <_sk_callback_hsw+0x50e> + DB 196,98,125,24,21,131,5,0,0 ; vbroadcastss 0x583(%rip),%ymm10 # 4e80 <_sk_callback_hsw+0x514> + DB 196,98,125,24,29,126,5,0,0 ; vbroadcastss 0x57e(%rip),%ymm11 # 4e84 <_sk_callback_hsw+0x518> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,101,5,0,0 ; vbroadcastss 0x565(%rip),%ymm8 # 4e8c <_sk_callback_hsw+0x512> + DB 196,98,125,24,5,111,5,0,0 ; vbroadcastss 0x56f(%rip),%ymm8 # 4e88 <_sk_callback_hsw+0x51c> DB 196,66,53,184,195 ; vfmadd231ps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -4434,12 +4434,12 @@ _sk_bicubic_p1y_hsw LABEL PROC PUBLIC _sk_bicubic_p3y_hsw _sk_bicubic_p3y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,77,5,0,0 ; vbroadcastss 0x54d(%rip),%ymm1 # 4e90 <_sk_callback_hsw+0x516> + DB 196,226,125,24,13,87,5,0,0 ; vbroadcastss 0x557(%rip),%ymm1 # 4e8c <_sk_callback_hsw+0x520> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,57,5,0,0 ; vbroadcastss 0x539(%rip),%ymm10 # 4e94 <_sk_callback_hsw+0x51a> - DB 196,98,125,24,29,52,5,0,0 ; vbroadcastss 0x534(%rip),%ymm11 # 4e98 <_sk_callback_hsw+0x51e> + DB 196,98,125,24,21,67,5,0,0 ; vbroadcastss 0x543(%rip),%ymm10 # 4e90 <_sk_callback_hsw+0x524> + DB 196,98,125,24,29,62,5,0,0 ; vbroadcastss 0x53e(%rip),%ymm11 # 4e94 <_sk_callback_hsw+0x528> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,52,89,195 ; vmulps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -4450,14 +4450,16 @@ PUBLIC _sk_callback_hsw _sk_callback_hsw LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,160,0,0,0 ; sub $0xa0,%rsp + DB 72,129,236,184,0,0,0 ; sub $0xb8,%rsp DB 197,252,17,188,36,128,0,0,0 ; vmovups %ymm7,0x80(%rsp) DB 197,252,17,116,36,96 ; vmovups %ymm6,0x60(%rsp) DB 197,252,17,108,36,64 ; vmovups %ymm5,0x40(%rsp) DB 197,252,17,100,36,32 ; vmovups %ymm4,0x20(%rsp) - DB 73,137,207 ; mov %rcx,%r15 - DB 73,137,214 ; mov %rdx,%r14 + DB 77,137,196 ; mov %r8,%r12 + DB 73,137,206 ; mov %rcx,%r14 + DB 73,137,215 ; mov %rdx,%r15 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,137,195 ; mov %rax,%rbx DB 197,252,20,225 ; vunpcklps %ymm1,%ymm0,%ymm4 @@ -4476,9 +4478,9 @@ _sk_callback_hsw LABEL PROC DB 197,253,17,107,40 ; vmovupd %ymm5,0x28(%rbx) DB 197,253,17,75,72 ; vmovupd %ymm1,0x48(%rbx) DB 197,253,17,67,104 ; vmovupd %ymm0,0x68(%rbx) - DB 77,133,255 ; test %r15,%r15 + DB 77,133,228 ; test %r12,%r12 DB 186,8,0,0,0 ; mov $0x8,%edx - DB 65,15,69,215 ; cmovne %r15d,%edx + DB 65,15,69,212 ; cmovne %r12d,%edx DB 72,137,217 ; mov %rbx,%rcx DB 197,248,119 ; vzeroupper DB 255,19 ; callq *(%rbx) @@ -4500,14 +4502,16 @@ _sk_callback_hsw LABEL PROC DB 197,213,20,211 ; vunpcklpd %ymm3,%ymm5,%ymm2 DB 197,213,21,219 ; vunpckhpd %ymm3,%ymm5,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,249 ; mov %r15,%rcx + DB 76,137,250 ; mov %r15,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 77,137,224 ; mov %r12,%r8 DB 197,252,16,100,36,32 ; vmovups 0x20(%rsp),%ymm4 DB 197,252,16,108,36,64 ; vmovups 0x40(%rsp),%ymm5 DB 197,252,16,116,36,96 ; vmovups 0x60(%rsp),%ymm6 DB 197,252,16,188,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm7 - DB 72,129,196,160,0,0,0 ; add $0xa0,%rsp + DB 72,129,196,184,0,0,0 ; add $0xb8,%rsp DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -4553,25 +4557,25 @@ ALIGN 4 DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4b45 <.literal4+0xb1> + DB 71,225,61 ; rex.RXB loope 4b41 <.literal4+0xb1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4b55 <.literal4+0xc1> + DB 71,225,61 ; rex.RXB loope 4b51 <.literal4+0xc1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4b65 <.literal4+0xd1> + DB 71,225,61 ; rex.RXB loope 4b61 <.literal4+0xd1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4b75 <.literal4+0xe1> + DB 71,225,61 ; rex.RXB loope 4b71 <.literal4+0xe1> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) @@ -4621,7 +4625,7 @@ ALIGN 4 DB 190,129,128,128,59 ; mov $0x3b808081,%esi DB 129,128,128,59,0,248,0,0,8,33 ; addl $0x21080000,-0x7ffc480(%rax) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4bcd <.literal4+0x139> + DB 224,7 ; loopne 4bc9 <.literal4+0x139> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -4637,10 +4641,10 @@ ALIGN 4 DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) DB 0,52,255 ; add %dh,(%rdi,%rdi,8) DB 255 ; (bad) - DB 127,0 ; jg 4bf4 <.literal4+0x160> + DB 127,0 ; jg 4bf0 <.literal4+0x160> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4c6d <.literal4+0x1d9> + DB 119,115 ; ja 4c69 <.literal4+0x1d9> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4654,10 +4658,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 4c28 <.literal4+0x194> + DB 127,0 ; jg 4c24 <.literal4+0x194> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4ca1 <.literal4+0x20d> + DB 119,115 ; ja 4c9d <.literal4+0x20d> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4671,10 +4675,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 4c5c <.literal4+0x1c8> + DB 127,0 ; jg 4c58 <.literal4+0x1c8> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4cd5 <.literal4+0x241> + DB 119,115 ; ja 4cd1 <.literal4+0x241> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4688,10 +4692,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 4c90 <.literal4+0x1fc> + DB 127,0 ; jg 4c8c <.literal4+0x1fc> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4d09 <.literal4+0x275> + DB 119,115 ; ja 4d05 <.literal4+0x275> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4704,7 +4708,7 @@ ALIGN 4 DB 0,75,0 ; add %cl,0x0(%rbx) DB 0,128,63,0,0,200 ; add %al,-0x37ffffc1(%rax) DB 66,0,0 ; rex.X add %al,(%rax) - DB 127,67 ; jg 4d07 <.literal4+0x273> + DB 127,67 ; jg 4d03 <.literal4+0x273> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -4716,10 +4720,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 4d27 <.literal4+0x293> + DB 118,63 ; jbe 4d23 <.literal4+0x293> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 4d3b <.literal4+0x2a7> + DB 127,67 ; jg 4d37 <.literal4+0x2a7> DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax) DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax) @@ -4728,7 +4732,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4d1d <.literal4+0x289> + DB 224,7 ; loopne 4d19 <.literal4+0x289> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -4740,7 +4744,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4d39 <.literal4+0x2a5> + DB 224,7 ; loopne 4d35 <.literal4+0x2a5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -4751,7 +4755,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 4d8e <.literal4+0x2fa> + DB 124,66 ; jl 4d8a <.literal4+0x2fa> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -4769,9 +4773,9 @@ ALIGN 4 DB 137,136,136,59,15,0 ; mov %ecx,0xf3b88(%rax) DB 0,0 ; add %al,(%rax) DB 137,136,136,61,0,0 ; mov %ecx,0x3d88(%rax) - DB 112,65 ; jo 4dd1 <.literal4+0x33d> + DB 112,65 ; jo 4dcd <.literal4+0x33d> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 4ddf <.literal4+0x34b> + DB 127,67 ; jg 4ddb <.literal4+0x34b> DB 128,0,128 ; addb $0x80,(%rax) DB 55 ; (bad) DB 128,0,128 ; addb $0x80,(%rax) @@ -4779,7 +4783,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 4df3 <.literal4+0x35f> + DB 127,71 ; jg 4def <.literal4+0x35f> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,191,0,0,0 ; add %al,0xbf(%rax) @@ -4887,16 +4891,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004ec8 <_sk_callback_hsw+0xa00054e> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004ec8 <_sk_callback_hsw+0xa00055c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004ed0 <_sk_callback_hsw+0x12000556> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004ed0 <_sk_callback_hsw+0x12000564> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004ed8 <_sk_callback_hsw+0x1a00055e> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004ed8 <_sk_callback_hsw+0x1a00056c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004ee0 <_sk_callback_hsw+0x3000566> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004ee0 <_sk_callback_hsw+0x3000574> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -4939,16 +4943,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004f28 <_sk_callback_hsw+0xa0005ae> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004f28 <_sk_callback_hsw+0xa0005bc> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004f30 <_sk_callback_hsw+0x120005b6> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004f30 <_sk_callback_hsw+0x120005c4> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004f38 <_sk_callback_hsw+0x1a0005be> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004f38 <_sk_callback_hsw+0x1a0005cc> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004f40 <_sk_callback_hsw+0x30005c6> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004f40 <_sk_callback_hsw+0x30005d4> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -4991,16 +4995,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004f88 <_sk_callback_hsw+0xa00060e> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004f88 <_sk_callback_hsw+0xa00061c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004f90 <_sk_callback_hsw+0x12000616> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004f90 <_sk_callback_hsw+0x12000624> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004f98 <_sk_callback_hsw+0x1a00061e> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004f98 <_sk_callback_hsw+0x1a00062c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004fa0 <_sk_callback_hsw+0x3000626> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004fa0 <_sk_callback_hsw+0x3000634> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5043,16 +5047,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004fe8 <_sk_callback_hsw+0xa00066e> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004fe8 <_sk_callback_hsw+0xa00067c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004ff0 <_sk_callback_hsw+0x12000676> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004ff0 <_sk_callback_hsw+0x12000684> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004ff8 <_sk_callback_hsw+0x1a00067e> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004ff8 <_sk_callback_hsw+0x1a00068c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005000 <_sk_callback_hsw+0x3000686> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005000 <_sk_callback_hsw+0x3000694> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5095,16 +5099,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005048 <_sk_callback_hsw+0xa0006ce> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005048 <_sk_callback_hsw+0xa0006dc> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005050 <_sk_callback_hsw+0x120006d6> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005050 <_sk_callback_hsw+0x120006e4> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005058 <_sk_callback_hsw+0x1a0006de> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005058 <_sk_callback_hsw+0x1a0006ec> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005060 <_sk_callback_hsw+0x30006e6> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005060 <_sk_callback_hsw+0x30006f4> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5157,8 +5161,9 @@ _sk_start_pipeline_avx LABEL PROC DB 65,84 ; push %r12 DB 86 ; push %rsi DB 87 ; push %rdi + DB 85 ; push %rbp DB 83 ; push %rbx - DB 72,129,236,160,0,0,0 ; sub $0xa0,%rsp + DB 72,129,236,168,0,0,0 ; sub $0xa8,%rsp DB 197,120,41,188,36,144,0,0,0 ; vmovaps %xmm15,0x90(%rsp) DB 197,120,41,180,36,128,0,0,0 ; vmovaps %xmm14,0x80(%rsp) DB 197,120,41,108,36,112 ; vmovaps %xmm13,0x70(%rsp) @@ -5169,19 +5174,20 @@ _sk_start_pipeline_avx LABEL PROC DB 197,120,41,68,36,32 ; vmovaps %xmm8,0x20(%rsp) DB 197,248,41,124,36,16 ; vmovaps %xmm7,0x10(%rsp) DB 197,248,41,52,36 ; vmovaps %xmm6,(%rsp) - DB 77,137,205 ; mov %r9,%r13 - DB 77,137,198 ; mov %r8,%r14 - DB 72,137,203 ; mov %rcx,%rbx - DB 72,137,214 ; mov %rdx,%rsi - DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,199 ; mov %rax,%r15 - DB 73,137,244 ; mov %rsi,%r12 - DB 72,141,67,8 ; lea 0x8(%rbx),%rax - DB 76,57,232 ; cmp %r13,%rax - DB 118,5 ; jbe 75 <_sk_start_pipeline_avx+0x75> - DB 72,137,223 ; mov %rbx,%rdi - DB 235,65 ; jmp b6 <_sk_start_pipeline_avx+0xb6> - DB 185,0,0,0,0 ; mov $0x0,%ecx + DB 76,137,195 ; mov %r8,%rbx + DB 73,137,214 ; mov %rdx,%r14 + DB 72,137,205 ; mov %rcx,%rbp + DB 76,139,188,36,16,1,0,0 ; mov 0x110(%rsp),%r15 + DB 76,137,206 ; mov %r9,%rsi + DB 72,173 ; lods %ds:(%rsi),%rax + DB 73,137,196 ; mov %rax,%r12 + DB 73,137,245 ; mov %rsi,%r13 + DB 72,141,69,8 ; lea 0x8(%rbp),%rax + DB 72,57,216 ; cmp %rbx,%rax + DB 118,5 ; jbe 7e <_sk_start_pipeline_avx+0x7e> + DB 72,137,234 ; mov %rbp,%rdx + DB 235,69 ; jmp c3 <_sk_start_pipeline_avx+0xc3> + DB 65,184,0,0,0,0 ; mov $0x0,%r8d DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 @@ -5190,17 +5196,18 @@ _sk_start_pipeline_avx LABEL PROC DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 197,204,87,246 ; vxorps %ymm6,%ymm6,%ymm6 DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 - DB 72,137,223 ; mov %rbx,%rdi - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 65,255,215 ; callq *%r15 - DB 72,141,123,8 ; lea 0x8(%rbx),%rdi - DB 72,131,195,16 ; add $0x10,%rbx - DB 76,57,235 ; cmp %r13,%rbx - DB 72,137,251 ; mov %rdi,%rbx - DB 118,191 ; jbe 75 <_sk_start_pipeline_avx+0x75> - DB 73,41,253 ; sub %rdi,%r13 - DB 116,44 ; je e7 <_sk_start_pipeline_avx+0xe7> + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 72,137,234 ; mov %rbp,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 65,255,212 ; callq *%r12 + DB 72,141,85,8 ; lea 0x8(%rbp),%rdx + DB 72,131,197,16 ; add $0x10,%rbp + DB 72,57,221 ; cmp %rbx,%rbp + DB 72,137,213 ; mov %rdx,%rbp + DB 118,187 ; jbe 7e <_sk_start_pipeline_avx+0x7e> + DB 72,41,211 ; sub %rdx,%rbx + DB 116,47 ; je f7 <_sk_start_pipeline_avx+0xf7> DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 @@ -5209,10 +5216,11 @@ _sk_start_pipeline_avx LABEL PROC DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 197,204,87,246 ; vxorps %ymm6,%ymm6,%ymm6 DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,233 ; mov %r13,%rcx - DB 65,255,215 ; callq *%r15 + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 76,137,241 ; mov %r14,%rcx + DB 73,137,216 ; mov %rbx,%r8 + DB 65,255,212 ; callq *%r12 DB 197,248,40,52,36 ; vmovaps (%rsp),%xmm6 DB 197,248,40,124,36,16 ; vmovaps 0x10(%rsp),%xmm7 DB 197,120,40,68,36,32 ; vmovaps 0x20(%rsp),%xmm8 @@ -5223,8 +5231,9 @@ _sk_start_pipeline_avx LABEL PROC DB 197,120,40,108,36,112 ; vmovaps 0x70(%rsp),%xmm13 DB 197,120,40,180,36,128,0,0,0 ; vmovaps 0x80(%rsp),%xmm14 DB 197,120,40,188,36,144,0,0,0 ; vmovaps 0x90(%rsp),%xmm15 - DB 72,129,196,160,0,0,0 ; add $0xa0,%rsp + DB 72,129,196,168,0,0,0 ; add $0xa8,%rsp DB 91 ; pop %rbx + DB 93 ; pop %rbp DB 95 ; pop %rdi DB 94 ; pop %rsi DB 65,92 ; pop %r12 @@ -5240,19 +5249,20 @@ _sk_just_return_avx LABEL PROC PUBLIC _sk_seed_shader_avx _sk_seed_shader_avx LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,249,110,199 ; vmovd %edi,%xmm0 + DB 197,249,110,194 ; vmovd %edx,%xmm0 DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,31,101,0,0 ; vbroadcastss 0x651f(%rip),%ymm1 # 667c <_sk_callback_avx+0x11c> + DB 196,226,125,24,13,112,101,0,0 ; vbroadcastss 0x6570(%rip),%ymm1 # 66dc <_sk_callback_avx+0x126> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 197,252,88,2 ; vaddps (%rdx),%ymm0,%ymm0 - DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 + DB 197,252,88,7 ; vaddps (%rdi),%ymm0,%ymm0 + DB 197,249,110,209 ; vmovd %ecx,%xmm2 + DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 + DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,3,101,0,0 ; vbroadcastss 0x6503(%rip),%ymm2 # 6680 <_sk_callback_avx+0x120> + DB 196,226,125,24,21,74,101,0,0 ; vbroadcastss 0x654a(%rip),%ymm2 # 66e0 <_sk_callback_avx+0x12a> DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 @@ -5263,17 +5273,18 @@ _sk_seed_shader_avx LABEL PROC PUBLIC _sk_dither_avx _sk_dither_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,124,16,66,32 ; vmovups 0x20(%rdx),%ymm8 + DB 197,124,16,71,32 ; vmovups 0x20(%rdi),%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 197,121,110,215 ; vmovd %edi,%xmm10 + DB 197,121,110,210 ; vmovd %edx,%xmm10 DB 196,65,121,112,210,0 ; vpshufd $0x0,%xmm10,%xmm10 DB 196,65,49,254,202 ; vpaddd %xmm10,%xmm9,%xmm9 DB 196,65,57,254,194 ; vpaddd %xmm10,%xmm8,%xmm8 DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - DB 76,139,0 ; mov (%rax),%r8 - DB 196,66,125,24,8 ; vbroadcastss (%r8),%ymm9 - DB 196,65,60,87,209 ; vxorps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,180,100,0,0 ; vbroadcastss 0x64b4(%rip),%ymm11 # 6684 <_sk_callback_avx+0x124> + DB 197,121,110,201 ; vmovd %ecx,%xmm9 + DB 196,65,121,112,201,0 ; vpshufd $0x0,%xmm9,%xmm9 + DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 + DB 196,65,52,87,208 ; vxorps %ymm8,%ymm9,%ymm10 + DB 196,98,125,24,29,243,100,0,0 ; vbroadcastss 0x64f3(%rip),%ymm11 # 66e4 <_sk_callback_avx+0x12e> DB 196,65,44,84,203 ; vandps %ymm11,%ymm10,%ymm9 DB 196,193,25,114,241,5 ; vpslld $0x5,%xmm9,%xmm12 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 @@ -5284,8 +5295,8 @@ _sk_dither_avx LABEL PROC DB 196,67,125,25,219,1 ; vextractf128 $0x1,%ymm11,%xmm11 DB 196,193,33,114,243,4 ; vpslld $0x4,%xmm11,%xmm11 DB 196,67,29,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm12,%ymm11 - DB 196,98,125,24,37,117,100,0,0 ; vbroadcastss 0x6475(%rip),%ymm12 # 6688 <_sk_callback_avx+0x128> - DB 196,98,125,24,45,112,100,0,0 ; vbroadcastss 0x6470(%rip),%ymm13 # 668c <_sk_callback_avx+0x12c> + DB 196,98,125,24,37,180,100,0,0 ; vbroadcastss 0x64b4(%rip),%ymm12 # 66e8 <_sk_callback_avx+0x132> + DB 196,98,125,24,45,175,100,0,0 ; vbroadcastss 0x64af(%rip),%ymm13 # 66ec <_sk_callback_avx+0x136> DB 196,65,44,84,245 ; vandps %ymm13,%ymm10,%ymm14 DB 196,193,1,114,246,2 ; vpslld $0x2,%xmm14,%xmm15 DB 196,67,125,25,246,1 ; vextractf128 $0x1,%ymm14,%xmm14 @@ -5308,15 +5319,15 @@ _sk_dither_avx LABEL PROC DB 196,67,29,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm12,%ymm8 DB 196,65,20,86,219 ; vorps %ymm11,%ymm13,%ymm11 DB 196,65,36,86,192 ; vorps %ymm8,%ymm11,%ymm8 + DB 196,65,52,86,206 ; vorps %ymm14,%ymm9,%ymm9 DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8 - DB 196,65,12,86,202 ; vorps %ymm10,%ymm14,%ymm9 - DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8 + DB 196,65,60,86,194 ; vorps %ymm10,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,219,99,0,0 ; vbroadcastss 0x63db(%rip),%ymm9 # 6690 <_sk_callback_avx+0x130> + DB 196,98,125,24,13,26,100,0,0 ; vbroadcastss 0x641a(%rip),%ymm9 # 66f0 <_sk_callback_avx+0x13a> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,209,99,0,0 ; vbroadcastss 0x63d1(%rip),%ymm9 # 6694 <_sk_callback_avx+0x134> + DB 196,98,125,24,13,16,100,0,0 ; vbroadcastss 0x6410(%rip),%ymm9 # 66f4 <_sk_callback_avx+0x13e> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,72,8 ; vbroadcastss 0x8(%rax),%ymm9 + DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 DB 197,188,88,201 ; vaddps %ymm1,%ymm8,%ymm1 @@ -5373,7 +5384,7 @@ _sk_clear_avx LABEL PROC PUBLIC _sk_srcatop_avx _sk_srcatop_avx LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,40,99,0,0 ; vbroadcastss 0x6328(%rip),%ymm8 # 6698 <_sk_callback_avx+0x138> + DB 196,98,125,24,5,104,99,0,0 ; vbroadcastss 0x6368(%rip),%ymm8 # 66f8 <_sk_callback_avx+0x142> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -5392,7 +5403,7 @@ _sk_srcatop_avx LABEL PROC PUBLIC _sk_dstatop_avx _sk_dstatop_avx LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 - DB 196,98,125,24,13,234,98,0,0 ; vbroadcastss 0x62ea(%rip),%ymm9 # 669c <_sk_callback_avx+0x13c> + DB 196,98,125,24,13,42,99,0,0 ; vbroadcastss 0x632a(%rip),%ymm9 # 66fc <_sk_callback_avx+0x146> DB 197,52,92,207 ; vsubps %ymm7,%ymm9,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 @@ -5428,7 +5439,7 @@ _sk_dstin_avx LABEL PROC PUBLIC _sk_srcout_avx _sk_srcout_avx LABEL PROC - DB 196,98,125,24,5,137,98,0,0 ; vbroadcastss 0x6289(%rip),%ymm8 # 66a0 <_sk_callback_avx+0x140> + DB 196,98,125,24,5,201,98,0,0 ; vbroadcastss 0x62c9(%rip),%ymm8 # 6700 <_sk_callback_avx+0x14a> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -5439,7 +5450,7 @@ _sk_srcout_avx LABEL PROC PUBLIC _sk_dstout_avx _sk_dstout_avx LABEL PROC - DB 196,226,125,24,5,108,98,0,0 ; vbroadcastss 0x626c(%rip),%ymm0 # 66a4 <_sk_callback_avx+0x144> + DB 196,226,125,24,5,172,98,0,0 ; vbroadcastss 0x62ac(%rip),%ymm0 # 6704 <_sk_callback_avx+0x14e> DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3 DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0 DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1 @@ -5450,7 +5461,7 @@ _sk_dstout_avx LABEL PROC PUBLIC _sk_srcover_avx _sk_srcover_avx LABEL PROC - DB 196,98,125,24,5,79,98,0,0 ; vbroadcastss 0x624f(%rip),%ymm8 # 66a8 <_sk_callback_avx+0x148> + DB 196,98,125,24,5,143,98,0,0 ; vbroadcastss 0x628f(%rip),%ymm8 # 6708 <_sk_callback_avx+0x152> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -5465,7 +5476,7 @@ _sk_srcover_avx LABEL PROC PUBLIC _sk_dstover_avx _sk_dstover_avx LABEL PROC - DB 196,98,125,24,5,34,98,0,0 ; vbroadcastss 0x6222(%rip),%ymm8 # 66ac <_sk_callback_avx+0x14c> + DB 196,98,125,24,5,98,98,0,0 ; vbroadcastss 0x6262(%rip),%ymm8 # 670c <_sk_callback_avx+0x156> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0 @@ -5489,7 +5500,7 @@ _sk_modulate_avx LABEL PROC PUBLIC _sk_multiply_avx _sk_multiply_avx LABEL PROC - DB 196,98,125,24,5,225,97,0,0 ; vbroadcastss 0x61e1(%rip),%ymm8 # 66b0 <_sk_callback_avx+0x150> + DB 196,98,125,24,5,33,98,0,0 ; vbroadcastss 0x6221(%rip),%ymm8 # 6710 <_sk_callback_avx+0x15a> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5543,7 +5554,7 @@ _sk_screen_avx LABEL PROC PUBLIC _sk_xor__avx _sk_xor__avx LABEL PROC - DB 196,98,125,24,5,48,97,0,0 ; vbroadcastss 0x6130(%rip),%ymm8 # 66b4 <_sk_callback_avx+0x154> + DB 196,98,125,24,5,112,97,0,0 ; vbroadcastss 0x6170(%rip),%ymm8 # 6714 <_sk_callback_avx+0x15e> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5578,7 +5589,7 @@ _sk_darken_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,176,96,0,0 ; vbroadcastss 0x60b0(%rip),%ymm8 # 66b8 <_sk_callback_avx+0x158> + DB 196,98,125,24,5,240,96,0,0 ; vbroadcastss 0x60f0(%rip),%ymm8 # 6718 <_sk_callback_avx+0x162> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5602,7 +5613,7 @@ _sk_lighten_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,92,96,0,0 ; vbroadcastss 0x605c(%rip),%ymm8 # 66bc <_sk_callback_avx+0x15c> + DB 196,98,125,24,5,156,96,0,0 ; vbroadcastss 0x609c(%rip),%ymm8 # 671c <_sk_callback_avx+0x166> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5629,7 +5640,7 @@ _sk_difference_avx LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,252,95,0,0 ; vbroadcastss 0x5ffc(%rip),%ymm8 # 66c0 <_sk_callback_avx+0x160> + DB 196,98,125,24,5,60,96,0,0 ; vbroadcastss 0x603c(%rip),%ymm8 # 6720 <_sk_callback_avx+0x16a> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5650,7 +5661,7 @@ _sk_exclusion_avx LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,183,95,0,0 ; vbroadcastss 0x5fb7(%rip),%ymm8 # 66c4 <_sk_callback_avx+0x164> + DB 196,98,125,24,5,247,95,0,0 ; vbroadcastss 0x5ff7(%rip),%ymm8 # 6724 <_sk_callback_avx+0x16e> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5659,7 +5670,7 @@ _sk_exclusion_avx LABEL PROC PUBLIC _sk_colorburn_avx _sk_colorburn_avx LABEL PROC - DB 196,98,125,24,5,162,95,0,0 ; vbroadcastss 0x5fa2(%rip),%ymm8 # 66c8 <_sk_callback_avx+0x168> + DB 196,98,125,24,5,226,95,0,0 ; vbroadcastss 0x5fe2(%rip),%ymm8 # 6728 <_sk_callback_avx+0x172> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 @@ -5719,7 +5730,7 @@ _sk_colorburn_avx LABEL PROC PUBLIC _sk_colordodge_avx _sk_colordodge_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,158,94,0,0 ; vbroadcastss 0x5e9e(%rip),%ymm9 # 66cc <_sk_callback_avx+0x16c> + DB 196,98,125,24,13,222,94,0,0 ; vbroadcastss 0x5ede(%rip),%ymm9 # 672c <_sk_callback_avx+0x176> DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9 @@ -5774,7 +5785,7 @@ _sk_colordodge_avx LABEL PROC PUBLIC _sk_hardlight_avx _sk_hardlight_avx LABEL PROC - DB 196,98,125,24,5,176,93,0,0 ; vbroadcastss 0x5db0(%rip),%ymm8 # 66d0 <_sk_callback_avx+0x170> + DB 196,98,125,24,5,240,93,0,0 ; vbroadcastss 0x5df0(%rip),%ymm8 # 6730 <_sk_callback_avx+0x17a> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5827,7 +5838,7 @@ _sk_hardlight_avx LABEL PROC PUBLIC _sk_overlay_avx _sk_overlay_avx LABEL PROC - DB 196,98,125,24,5,217,92,0,0 ; vbroadcastss 0x5cd9(%rip),%ymm8 # 66d4 <_sk_callback_avx+0x174> + DB 196,98,125,24,5,25,93,0,0 ; vbroadcastss 0x5d19(%rip),%ymm8 # 6734 <_sk_callback_avx+0x17e> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5892,10 +5903,10 @@ _sk_softlight_avx LABEL PROC DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,65,60,89,216 ; vmulps %ymm8,%ymm8,%ymm11 DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,29,204,91,0,0 ; vbroadcastss 0x5bcc(%rip),%ymm11 # 66dc <_sk_callback_avx+0x17c> + DB 196,98,125,24,29,12,92,0,0 ; vbroadcastss 0x5c0c(%rip),%ymm11 # 673c <_sk_callback_avx+0x186> DB 196,65,28,88,235 ; vaddps %ymm11,%ymm12,%ymm13 DB 196,65,20,89,192 ; vmulps %ymm8,%ymm13,%ymm8 - DB 196,98,125,24,45,189,91,0,0 ; vbroadcastss 0x5bbd(%rip),%ymm13 # 66e0 <_sk_callback_avx+0x180> + DB 196,98,125,24,45,253,91,0,0 ; vbroadcastss 0x5bfd(%rip),%ymm13 # 6740 <_sk_callback_avx+0x18a> DB 196,65,28,89,245 ; vmulps %ymm13,%ymm12,%ymm14 DB 196,65,12,88,192 ; vaddps %ymm8,%ymm14,%ymm8 DB 196,65,124,82,244 ; vrsqrtps %ymm12,%ymm14 @@ -5906,7 +5917,7 @@ _sk_softlight_avx LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,123,91,0,0 ; vbroadcastss 0x5b7b(%rip),%ymm8 # 66d8 <_sk_callback_avx+0x178> + DB 196,98,125,24,5,187,91,0,0 ; vbroadcastss 0x5bbb(%rip),%ymm8 # 6738 <_sk_callback_avx+0x182> DB 196,65,60,92,228 ; vsubps %ymm12,%ymm8,%ymm12 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 @@ -6033,12 +6044,12 @@ _sk_hue_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,69,89,0,0 ; vbroadcastss 0x5945(%rip),%ymm14 # 66e4 <_sk_callback_avx+0x184> + DB 196,98,125,24,53,133,89,0,0 ; vbroadcastss 0x5985(%rip),%ymm14 # 6744 <_sk_callback_avx+0x18e> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,59,89,0,0 ; vbroadcastss 0x593b(%rip),%ymm15 # 66e8 <_sk_callback_avx+0x188> + DB 196,98,125,24,61,123,89,0,0 ; vbroadcastss 0x597b(%rip),%ymm15 # 6748 <_sk_callback_avx+0x192> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,44,89,0,0 ; vbroadcastss 0x592c(%rip),%ymm0 # 66ec <_sk_callback_avx+0x18c> + DB 196,226,125,24,5,108,89,0,0 ; vbroadcastss 0x596c(%rip),%ymm0 # 674c <_sk_callback_avx+0x196> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -6099,7 +6110,7 @@ _sk_hue_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,5,88,0,0 ; vbroadcastss 0x5805(%rip),%ymm8 # 66f0 <_sk_callback_avx+0x190> + DB 196,98,125,24,5,69,88,0,0 ; vbroadcastss 0x5845(%rip),%ymm8 # 6750 <_sk_callback_avx+0x19a> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6156,12 +6167,12 @@ _sk_saturation_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,13,87,0,0 ; vbroadcastss 0x570d(%rip),%ymm14 # 66f4 <_sk_callback_avx+0x194> + DB 196,98,125,24,53,77,87,0,0 ; vbroadcastss 0x574d(%rip),%ymm14 # 6754 <_sk_callback_avx+0x19e> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,3,87,0,0 ; vbroadcastss 0x5703(%rip),%ymm15 # 66f8 <_sk_callback_avx+0x198> + DB 196,98,125,24,61,67,87,0,0 ; vbroadcastss 0x5743(%rip),%ymm15 # 6758 <_sk_callback_avx+0x1a2> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,244,86,0,0 ; vbroadcastss 0x56f4(%rip),%ymm0 # 66fc <_sk_callback_avx+0x19c> + DB 196,226,125,24,5,52,87,0,0 ; vbroadcastss 0x5734(%rip),%ymm0 # 675c <_sk_callback_avx+0x1a6> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -6222,7 +6233,7 @@ _sk_saturation_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,205,85,0,0 ; vbroadcastss 0x55cd(%rip),%ymm8 # 6700 <_sk_callback_avx+0x1a0> + DB 196,98,125,24,5,13,86,0,0 ; vbroadcastss 0x560d(%rip),%ymm8 # 6760 <_sk_callback_avx+0x1aa> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6251,12 +6262,12 @@ _sk_color_avx LABEL PROC DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp) DB 197,124,89,199 ; vmulps %ymm7,%ymm0,%ymm8 DB 197,116,89,207 ; vmulps %ymm7,%ymm1,%ymm9 - DB 196,98,125,24,45,93,85,0,0 ; vbroadcastss 0x555d(%rip),%ymm13 # 6704 <_sk_callback_avx+0x1a4> + DB 196,98,125,24,45,157,85,0,0 ; vbroadcastss 0x559d(%rip),%ymm13 # 6764 <_sk_callback_avx+0x1ae> DB 196,65,92,89,213 ; vmulps %ymm13,%ymm4,%ymm10 - DB 196,98,125,24,53,83,85,0,0 ; vbroadcastss 0x5553(%rip),%ymm14 # 6708 <_sk_callback_avx+0x1a8> + DB 196,98,125,24,53,147,85,0,0 ; vbroadcastss 0x5593(%rip),%ymm14 # 6768 <_sk_callback_avx+0x1b2> DB 196,65,84,89,222 ; vmulps %ymm14,%ymm5,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,68,85,0,0 ; vbroadcastss 0x5544(%rip),%ymm15 # 670c <_sk_callback_avx+0x1ac> + DB 196,98,125,24,61,132,85,0,0 ; vbroadcastss 0x5584(%rip),%ymm15 # 676c <_sk_callback_avx+0x1b6> DB 196,65,76,89,223 ; vmulps %ymm15,%ymm6,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -6319,7 +6330,7 @@ _sk_color_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,11,84,0,0 ; vbroadcastss 0x540b(%rip),%ymm0 # 6710 <_sk_callback_avx+0x1b0> + DB 196,226,125,24,5,75,84,0,0 ; vbroadcastss 0x544b(%rip),%ymm0 # 6770 <_sk_callback_avx+0x1ba> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,84,36,32 ; vmulps 0x20(%rsp),%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -6349,12 +6360,12 @@ _sk_luminosity_avx LABEL PROC DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 - DB 196,98,125,24,45,151,83,0,0 ; vbroadcastss 0x5397(%rip),%ymm13 # 6714 <_sk_callback_avx+0x1b4> + DB 196,98,125,24,45,215,83,0,0 ; vbroadcastss 0x53d7(%rip),%ymm13 # 6774 <_sk_callback_avx+0x1be> DB 196,65,108,89,213 ; vmulps %ymm13,%ymm2,%ymm10 - DB 196,98,125,24,53,141,83,0,0 ; vbroadcastss 0x538d(%rip),%ymm14 # 6718 <_sk_callback_avx+0x1b8> + DB 196,98,125,24,53,205,83,0,0 ; vbroadcastss 0x53cd(%rip),%ymm14 # 6778 <_sk_callback_avx+0x1c2> DB 196,65,116,89,222 ; vmulps %ymm14,%ymm1,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,126,83,0,0 ; vbroadcastss 0x537e(%rip),%ymm15 # 671c <_sk_callback_avx+0x1bc> + DB 196,98,125,24,61,190,83,0,0 ; vbroadcastss 0x53be(%rip),%ymm15 # 677c <_sk_callback_avx+0x1c6> DB 196,65,28,89,223 ; vmulps %ymm15,%ymm12,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -6417,7 +6428,7 @@ _sk_luminosity_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,69,82,0,0 ; vbroadcastss 0x5245(%rip),%ymm0 # 6720 <_sk_callback_avx+0x1c0> + DB 196,226,125,24,5,133,82,0,0 ; vbroadcastss 0x5285(%rip),%ymm0 # 6780 <_sk_callback_avx+0x1ca> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -6441,14 +6452,14 @@ _sk_luminosity_avx LABEL PROC PUBLIC _sk_srcover_rgba_8888_avx _sk_srcover_rgba_8888_avx LABEL PROC DB 72,131,236,16 ; sub $0x10,%rsp - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 DB 77,133,192 ; test %r8,%r8 - DB 15,133,47,1,0,0 ; jne 1675 <_sk_srcover_rgba_8888_avx+0x14c> - DB 196,193,124,16,57 ; vmovups (%r9),%ymm7 - DB 197,124,40,13,13,86,0,0 ; vmovaps 0x560d(%rip),%ymm9 # 6b60 <_sk_callback_avx+0x600> + DB 15,133,47,1,0,0 ; jne 1695 <_sk_srcover_rgba_8888_avx+0x14c> + DB 196,193,124,16,58 ; vmovups (%r10),%ymm7 + DB 197,124,40,13,77,86,0,0 ; vmovaps 0x564d(%rip),%ymm9 # 6bc0 <_sk_callback_avx+0x60a> DB 196,193,68,84,225 ; vandps %ymm9,%ymm7,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 197,209,114,215,8 ; vpsrld $0x8,%xmm7,%xmm5 @@ -6466,9 +6477,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7 DB 196,227,53,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm9,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,112,81,0,0 ; vbroadcastss 0x5170(%rip),%ymm8 # 6724 <_sk_callback_avx+0x1c4> + DB 196,98,125,24,5,176,81,0,0 ; vbroadcastss 0x51b0(%rip),%ymm8 # 6784 <_sk_callback_avx+0x1ce> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,103,81,0,0 ; vbroadcastss 0x5167(%rip),%ymm9 # 6728 <_sk_callback_avx+0x1c8> + DB 196,98,125,24,13,167,81,0,0 ; vbroadcastss 0x51a7(%rip),%ymm9 # 6788 <_sk_callback_avx+0x1d2> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,60,89,212 ; vmulps %ymm4,%ymm8,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 @@ -6501,10 +6512,10 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,65,53,86,202 ; vorpd %ymm10,%ymm9,%ymm9 DB 196,65,61,86,193 ; vorpd %ymm9,%ymm8,%ymm8 DB 77,133,192 ; test %r8,%r8 - DB 117,91 ; jne 16c0 <_sk_srcover_rgba_8888_avx+0x197> - DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) + DB 117,91 ; jne 16e0 <_sk_srcover_rgba_8888_avx+0x197> + DB 196,65,124,17,2 ; vmovups %ymm8,(%r10) DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 72,131,196,16 ; add $0x10,%rsp DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx @@ -6514,13 +6525,13 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,224 ; vmovq %rax,%xmm4 DB 196,226,121,48,228 ; vpmovzxbw %xmm4,%xmm4 - DB 196,226,89,0,45,19,84,0,0 ; vpshufb 0x5413(%rip),%xmm4,%xmm5 # 6ab0 <_sk_callback_avx+0x550> + DB 196,226,89,0,45,83,84,0,0 ; vpshufb 0x5453(%rip),%xmm4,%xmm5 # 6b10 <_sk_callback_avx+0x55a> DB 196,226,121,33,237 ; vpmovsxbd %xmm5,%xmm5 - DB 196,226,89,0,37,21,84,0,0 ; vpshufb 0x5415(%rip),%xmm4,%xmm4 # 6ac0 <_sk_callback_avx+0x560> + DB 196,226,89,0,37,85,84,0,0 ; vpshufb 0x5455(%rip),%xmm4,%xmm4 # 6b20 <_sk_callback_avx+0x56a> DB 196,226,121,33,228 ; vpmovsxbd %xmm4,%xmm4 DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 - DB 196,194,93,44,57 ; vmaskmovps (%r9),%ymm4,%ymm7 - DB 233,139,254,255,255 ; jmpq 154b <_sk_srcover_rgba_8888_avx+0x22> + DB 196,194,93,44,58 ; vmaskmovps (%r10),%ymm4,%ymm7 + DB 233,139,254,255,255 ; jmpq 156b <_sk_srcover_rgba_8888_avx+0x22> DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx DB 192,225,3 ; shl $0x3,%cl @@ -6528,13 +6539,13 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 - DB 196,98,49,0,21,200,83,0,0 ; vpshufb 0x53c8(%rip),%xmm9,%xmm10 # 6ab0 <_sk_callback_avx+0x550> + DB 196,98,49,0,21,8,84,0,0 ; vpshufb 0x5408(%rip),%xmm9,%xmm10 # 6b10 <_sk_callback_avx+0x55a> DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10 - DB 196,98,49,0,13,202,83,0,0 ; vpshufb 0x53ca(%rip),%xmm9,%xmm9 # 6ac0 <_sk_callback_avx+0x560> + DB 196,98,49,0,13,10,84,0,0 ; vpshufb 0x540a(%rip),%xmm9,%xmm9 # 6b20 <_sk_callback_avx+0x56a> DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) - DB 233,95,255,255,255 ; jmpq 166a <_sk_srcover_rgba_8888_avx+0x141> + DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10) + DB 233,95,255,255,255 ; jmpq 168a <_sk_srcover_rgba_8888_avx+0x141> PUBLIC _sk_clamp_0_avx _sk_clamp_0_avx LABEL PROC @@ -6548,7 +6559,7 @@ _sk_clamp_0_avx LABEL PROC PUBLIC _sk_clamp_1_avx _sk_clamp_1_avx LABEL PROC - DB 196,98,125,24,5,251,79,0,0 ; vbroadcastss 0x4ffb(%rip),%ymm8 # 672c <_sk_callback_avx+0x1cc> + DB 196,98,125,24,5,59,80,0,0 ; vbroadcastss 0x503b(%rip),%ymm8 # 678c <_sk_callback_avx+0x1d6> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 @@ -6558,7 +6569,7 @@ _sk_clamp_1_avx LABEL PROC PUBLIC _sk_clamp_a_avx _sk_clamp_a_avx LABEL PROC - DB 196,98,125,24,5,222,79,0,0 ; vbroadcastss 0x4fde(%rip),%ymm8 # 6730 <_sk_callback_avx+0x1d0> + DB 196,98,125,24,5,30,80,0,0 ; vbroadcastss 0x501e(%rip),%ymm8 # 6790 <_sk_callback_avx+0x1da> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0 DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1 @@ -6630,7 +6641,7 @@ PUBLIC _sk_unpremul_avx _sk_unpremul_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,38,79,0,0 ; vbroadcastss 0x4f26(%rip),%ymm10 # 6734 <_sk_callback_avx+0x1d4> + DB 196,98,125,24,21,102,79,0,0 ; vbroadcastss 0x4f66(%rip),%ymm10 # 6794 <_sk_callback_avx+0x1de> DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10 DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 @@ -6641,17 +6652,17 @@ _sk_unpremul_avx LABEL PROC PUBLIC _sk_from_srgb_avx _sk_from_srgb_avx LABEL PROC - DB 196,98,125,24,5,7,79,0,0 ; vbroadcastss 0x4f07(%rip),%ymm8 # 6738 <_sk_callback_avx+0x1d8> + DB 196,98,125,24,5,71,79,0,0 ; vbroadcastss 0x4f47(%rip),%ymm8 # 6798 <_sk_callback_avx+0x1e2> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,249,78,0,0 ; vbroadcastss 0x4ef9(%rip),%ymm11 # 673c <_sk_callback_avx+0x1dc> + DB 196,98,125,24,29,57,79,0,0 ; vbroadcastss 0x4f39(%rip),%ymm11 # 679c <_sk_callback_avx+0x1e6> DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12 - DB 196,98,125,24,45,239,78,0,0 ; vbroadcastss 0x4eef(%rip),%ymm13 # 6740 <_sk_callback_avx+0x1e0> + DB 196,98,125,24,45,47,79,0,0 ; vbroadcastss 0x4f2f(%rip),%ymm13 # 67a0 <_sk_callback_avx+0x1ea> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,224,78,0,0 ; vbroadcastss 0x4ee0(%rip),%ymm12 # 6744 <_sk_callback_avx+0x1e4> + DB 196,98,125,24,37,32,79,0,0 ; vbroadcastss 0x4f20(%rip),%ymm12 # 67a4 <_sk_callback_avx+0x1ee> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,53,214,78,0,0 ; vbroadcastss 0x4ed6(%rip),%ymm14 # 6748 <_sk_callback_avx+0x1e8> + DB 196,98,125,24,53,22,79,0,0 ; vbroadcastss 0x4f16(%rip),%ymm14 # 67a8 <_sk_callback_avx+0x1f2> DB 196,193,124,194,198,1 ; vcmpltps %ymm14,%ymm0,%ymm0 DB 196,195,45,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm10,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -6676,20 +6687,20 @@ _sk_from_srgb_avx LABEL PROC PUBLIC _sk_to_srgb_avx _sk_to_srgb_avx LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,107,78,0,0 ; vbroadcastss 0x4e6b(%rip),%ymm8 # 674c <_sk_callback_avx+0x1ec> + DB 196,98,125,24,5,171,78,0,0 ; vbroadcastss 0x4eab(%rip),%ymm8 # 67ac <_sk_callback_avx+0x1f6> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,97,78,0,0 ; vbroadcastss 0x4e61(%rip),%ymm11 # 6750 <_sk_callback_avx+0x1f0> + DB 196,98,125,24,29,161,78,0,0 ; vbroadcastss 0x4ea1(%rip),%ymm11 # 67b0 <_sk_callback_avx+0x1fa> DB 196,65,52,89,227 ; vmulps %ymm11,%ymm9,%ymm12 - DB 196,98,125,24,45,87,78,0,0 ; vbroadcastss 0x4e57(%rip),%ymm13 # 6754 <_sk_callback_avx+0x1f4> + DB 196,98,125,24,45,151,78,0,0 ; vbroadcastss 0x4e97(%rip),%ymm13 # 67b4 <_sk_callback_avx+0x1fe> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,52,89,228 ; vmulps %ymm12,%ymm9,%ymm12 - DB 196,98,125,24,53,72,78,0,0 ; vbroadcastss 0x4e48(%rip),%ymm14 # 6758 <_sk_callback_avx+0x1f8> + DB 196,98,125,24,53,136,78,0,0 ; vbroadcastss 0x4e88(%rip),%ymm14 # 67b8 <_sk_callback_avx+0x202> DB 196,65,28,88,230 ; vaddps %ymm14,%ymm12,%ymm12 - DB 196,98,125,24,61,62,78,0,0 ; vbroadcastss 0x4e3e(%rip),%ymm15 # 675c <_sk_callback_avx+0x1fc> + DB 196,98,125,24,61,126,78,0,0 ; vbroadcastss 0x4e7e(%rip),%ymm15 # 67bc <_sk_callback_avx+0x206> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 - DB 196,98,125,24,37,42,78,0,0 ; vbroadcastss 0x4e2a(%rip),%ymm12 # 6760 <_sk_callback_avx+0x200> + DB 196,98,125,24,37,106,78,0,0 ; vbroadcastss 0x4e6a(%rip),%ymm12 # 67c0 <_sk_callback_avx+0x20a> DB 196,193,124,194,196,1 ; vcmpltps %ymm12,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -6724,7 +6735,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9 DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,144,77,0,0 ; vbroadcastss 0x4d90(%rip),%ymm11 # 6764 <_sk_callback_avx+0x204> + DB 196,98,125,24,29,208,77,0,0 ; vbroadcastss 0x4dd0(%rip),%ymm11 # 67c4 <_sk_callback_avx+0x20e> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 196,65,28,89,227 ; vmulps %ymm11,%ymm12,%ymm12 @@ -6734,19 +6745,19 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 DB 196,193,116,89,203 ; vmulps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,105,77,0,0 ; vbroadcastss 0x4d69(%rip),%ymm11 # 6770 <_sk_callback_avx+0x210> + DB 196,98,125,24,29,169,77,0,0 ; vbroadcastss 0x4da9(%rip),%ymm11 # 67d0 <_sk_callback_avx+0x21a> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,87,77,0,0 ; vbroadcastss 0x4d57(%rip),%ymm11 # 676c <_sk_callback_avx+0x20c> + DB 196,98,125,24,29,151,77,0,0 ; vbroadcastss 0x4d97(%rip),%ymm11 # 67cc <_sk_callback_avx+0x216> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,227,117,74,202,224 ; vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,63,77,0,0 ; vbroadcastss 0x4d3f(%rip),%ymm2 # 6768 <_sk_callback_avx+0x208> + DB 196,226,125,24,21,127,77,0,0 ; vbroadcastss 0x4d7f(%rip),%ymm2 # 67c8 <_sk_callback_avx+0x212> DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 DB 196,227,117,74,194,0 ; vblendvps %ymm0,%ymm2,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,37,38,77,0,0 ; vbroadcastss 0x4d26(%rip),%ymm12 # 6778 <_sk_callback_avx+0x218> + DB 196,98,125,24,37,102,77,0,0 ; vbroadcastss 0x4d66(%rip),%ymm12 # 67d8 <_sk_callback_avx+0x222> DB 196,193,116,89,212 ; vmulps %ymm12,%ymm1,%ymm2 DB 197,28,194,226,1 ; vcmpltps %ymm2,%ymm12,%ymm12 DB 196,65,36,92,216 ; vsubps %ymm8,%ymm11,%ymm11 @@ -6756,7 +6767,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,198,128 ; vblendvps %ymm8,%ymm14,%ymm0,%ymm0 DB 196,195,117,74,206,128 ; vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - DB 196,98,125,24,5,233,76,0,0 ; vbroadcastss 0x4ce9(%rip),%ymm8 # 6774 <_sk_callback_avx+0x214> + DB 196,98,125,24,5,41,77,0,0 ; vbroadcastss 0x4d29(%rip),%ymm8 # 67d4 <_sk_callback_avx+0x21e> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6771,7 +6782,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,225 ; vmovaps %ymm1,%ymm4 DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 - DB 196,98,125,24,5,176,76,0,0 ; vbroadcastss 0x4cb0(%rip),%ymm8 # 677c <_sk_callback_avx+0x21c> + DB 196,98,125,24,5,240,76,0,0 ; vbroadcastss 0x4cf0(%rip),%ymm8 # 67dc <_sk_callback_avx+0x226> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,92,89,210 ; vmulps %ymm2,%ymm4,%ymm10 DB 196,65,92,92,218 ; vsubps %ymm10,%ymm4,%ymm11 @@ -6779,23 +6790,23 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 DB 197,108,88,202 ; vaddps %ymm2,%ymm2,%ymm9 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,29,138,76,0,0 ; vbroadcastss 0x4c8a(%rip),%ymm11 # 6780 <_sk_callback_avx+0x220> + DB 196,98,125,24,29,202,76,0,0 ; vbroadcastss 0x4cca(%rip),%ymm11 # 67e0 <_sk_callback_avx+0x22a> DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,37,116,76,0,0 ; vbroadcastss 0x4c74(%rip),%ymm12 # 6788 <_sk_callback_avx+0x228> + DB 196,98,125,24,37,180,76,0,0 ; vbroadcastss 0x4cb4(%rip),%ymm12 # 67e8 <_sk_callback_avx+0x232> DB 196,193,4,89,196 ; vmulps %ymm12,%ymm15,%ymm0 - DB 196,98,125,24,45,106,76,0,0 ; vbroadcastss 0x4c6a(%rip),%ymm13 # 678c <_sk_callback_avx+0x22c> + DB 196,98,125,24,45,170,76,0,0 ; vbroadcastss 0x4caa(%rip),%ymm13 # 67ec <_sk_callback_avx+0x236> DB 197,20,92,240 ; vsubps %ymm0,%ymm13,%ymm14 DB 196,65,36,89,246 ; vmulps %ymm14,%ymm11,%ymm14 DB 196,65,52,88,246 ; vaddps %ymm14,%ymm9,%ymm14 - DB 196,226,125,24,13,75,76,0,0 ; vbroadcastss 0x4c4b(%rip),%ymm1 # 6784 <_sk_callback_avx+0x224> + DB 196,226,125,24,13,139,76,0,0 ; vbroadcastss 0x4c8b(%rip),%ymm1 # 67e4 <_sk_callback_avx+0x22e> DB 196,193,116,194,255,2 ; vcmpleps %ymm15,%ymm1,%ymm7 DB 196,195,13,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm14,%ymm7 DB 196,65,60,194,247,2 ; vcmpleps %ymm15,%ymm8,%ymm14 DB 196,227,45,74,255,224 ; vblendvps %ymm14,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,53,54,76,0,0 ; vbroadcastss 0x4c36(%rip),%ymm14 # 6790 <_sk_callback_avx+0x230> + DB 196,98,125,24,53,118,76,0,0 ; vbroadcastss 0x4c76(%rip),%ymm14 # 67f0 <_sk_callback_avx+0x23a> DB 196,65,12,194,255,2 ; vcmpleps %ymm15,%ymm14,%ymm15 DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -6814,7 +6825,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,164,89,247 ; vmulps %ymm7,%ymm11,%ymm6 DB 197,180,88,246 ; vaddps %ymm6,%ymm9,%ymm6 DB 196,227,77,74,237,0 ; vblendvps %ymm0,%ymm5,%ymm6,%ymm5 - DB 196,226,125,24,5,216,75,0,0 ; vbroadcastss 0x4bd8(%rip),%ymm0 # 6794 <_sk_callback_avx+0x234> + DB 196,226,125,24,5,24,76,0,0 ; vbroadcastss 0x4c18(%rip),%ymm0 # 67f4 <_sk_callback_avx+0x23e> DB 197,228,88,192 ; vaddps %ymm0,%ymm3,%ymm0 DB 196,227,125,8,216,1 ; vroundps $0x1,%ymm0,%ymm3 DB 197,252,92,195 ; vsubps %ymm3,%ymm0,%ymm0 @@ -6857,39 +6868,41 @@ _sk_scale_1_float_avx LABEL PROC PUBLIC _sk_scale_u8_avx _sk_scale_u8_avx LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,68 ; jne 1cb9 <_sk_scale_u8_avx+0x54> + DB 117,68 ; jne 1cd9 <_sk_scale_u8_avx+0x54> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9 DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8 DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8 DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,251,74,0,0 ; vbroadcastss 0x4afb(%rip),%ymm9 # 6798 <_sk_callback_avx+0x238> + DB 196,98,125,24,13,59,75,0,0 ; vbroadcastss 0x4b3b(%rip),%ymm9 # 67f8 <_sk_callback_avx+0x242> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 1cc1 <_sk_scale_u8_avx+0x5c> - DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,155 ; jmp 1c79 <_sk_scale_u8_avx+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 1ce2 <_sk_scale_u8_avx+0x5d> + DB 196,65,249,110,194 ; vmovq %r10,%xmm8 + DB 91 ; pop %rbx + DB 235,154 ; jmp 1c99 <_sk_scale_u8_avx+0x14> PUBLIC _sk_lerp_1_float_avx _sk_lerp_1_float_avx LABEL PROC @@ -6912,19 +6925,19 @@ _sk_lerp_1_float_avx LABEL PROC PUBLIC _sk_lerp_u8_avx _sk_lerp_u8_avx LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,104 ; jne 1d95 <_sk_lerp_u8_avx+0x78> + DB 117,104 ; jne 1db6 <_sk_lerp_u8_avx+0x78> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9 DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8 DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8 DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,71,74,0,0 ; vbroadcastss 0x4a47(%rip),%ymm9 # 679c <_sk_callback_avx+0x23c> + DB 196,98,125,24,13,134,74,0,0 ; vbroadcastss 0x4a86(%rip),%ymm9 # 67fc <_sk_callback_avx+0x246> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 @@ -6939,46 +6952,48 @@ _sk_lerp_u8_avx LABEL PROC DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 197,228,88,223 ; vaddps %ymm7,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 1d9d <_sk_lerp_u8_avx+0x80> - DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 233,116,255,255,255 ; jmpq 1d31 <_sk_lerp_u8_avx+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 1dbf <_sk_lerp_u8_avx+0x81> + DB 196,65,249,110,194 ; vmovq %r10,%xmm8 + DB 91 ; pop %rbx + DB 233,115,255,255,255 ; jmpq 1d52 <_sk_lerp_u8_avx+0x14> PUBLIC _sk_lerp_565_avx _sk_lerp_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,208,0,0,0 ; jne 1e9b <_sk_lerp_565_avx+0xde> - DB 196,65,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm8 + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,208,0,0,0 ; jne 1ebd <_sk_lerp_565_avx+0xde> + DB 196,65,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - DB 196,98,125,24,13,177,73,0,0 ; vbroadcastss 0x49b1(%rip),%ymm9 # 67a0 <_sk_callback_avx+0x240> + DB 196,98,125,24,13,239,73,0,0 ; vbroadcastss 0x49ef(%rip),%ymm9 # 6800 <_sk_callback_avx+0x24a> DB 196,65,60,84,201 ; vandps %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,162,73,0,0 ; vbroadcastss 0x49a2(%rip),%ymm10 # 67a4 <_sk_callback_avx+0x244> + DB 196,98,125,24,21,224,73,0,0 ; vbroadcastss 0x49e0(%rip),%ymm10 # 6804 <_sk_callback_avx+0x24e> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,152,73,0,0 ; vbroadcastss 0x4998(%rip),%ymm10 # 67a8 <_sk_callback_avx+0x248> + DB 196,98,125,24,21,214,73,0,0 ; vbroadcastss 0x49d6(%rip),%ymm10 # 6808 <_sk_callback_avx+0x252> DB 196,65,60,84,210 ; vandps %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,137,73,0,0 ; vbroadcastss 0x4989(%rip),%ymm11 # 67ac <_sk_callback_avx+0x24c> + DB 196,98,125,24,29,199,73,0,0 ; vbroadcastss 0x49c7(%rip),%ymm11 # 680c <_sk_callback_avx+0x256> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,29,127,73,0,0 ; vbroadcastss 0x497f(%rip),%ymm11 # 67b0 <_sk_callback_avx+0x250> + DB 196,98,125,24,29,189,73,0,0 ; vbroadcastss 0x49bd(%rip),%ymm11 # 6810 <_sk_callback_avx+0x25a> DB 196,65,60,84,195 ; vandps %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,112,73,0,0 ; vbroadcastss 0x4970(%rip),%ymm11 # 67b4 <_sk_callback_avx+0x254> + DB 196,98,125,24,29,174,73,0,0 ; vbroadcastss 0x49ae(%rip),%ymm11 # 6814 <_sk_callback_avx+0x25e> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 @@ -7000,201 +7015,190 @@ _sk_lerp_565_avx LABEL PROC DB 197,180,95,219 ; vmaxps %ymm3,%ymm9,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,29,255,255,255 ; ja 1dd1 <_sk_lerp_565_avx+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,77,0,0,0 ; lea 0x4d(%rip),%r9 # 1f0c <_sk_lerp_565_avx+0x14f> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 15,135,29,255,255,255 ; ja 1df3 <_sk_lerp_565_avx+0x14> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 1f2c <_sk_lerp_565_avx+0x14d> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 196,65,57,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 - DB 196,65,57,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 - DB 233,200,254,255,255 ; jmpq 1dd1 <_sk_lerp_565_avx+0x14> - DB 15,31,0 ; nopl (%rax) - DB 241 ; icebp + DB 196,65,57,196,68,83,12,6 ; vpinsrw $0x6,0xc(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,10,5 ; vpinsrw $0x5,0xa(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,8,4 ; vpinsrw $0x4,0x8(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,6,3 ; vpinsrw $0x3,0x6(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm8,%xmm8 + DB 196,65,57,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm8,%xmm8 + DB 233,200,254,255,255 ; jmpq 1df3 <_sk_lerp_565_avx+0x14> + DB 144 ; nop + DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) + DB 235,255 ; jmp 1f31 <_sk_lerp_565_avx+0x152> DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2001f14 <_sk_callback_avx+0xffffffffe1ffb9b4> + DB 255,227 ; jmpq *%rbx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 217,255 ; fcos + DB 219,255 ; (bad) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,211 ; callq *%rbx DB 255 ; (bad) DB 255 ; (bad) - DB 255,201 ; dec %ecx + DB 255,203 ; dec %ebx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 188 ; .byte 0xbc + DB 190 ; .byte 0xbe DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_tables_avx _sk_load_tables_avx LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 83 ; push %rbx + DB 72,131,236,48 ; sub $0x30,%rsp + DB 197,252,17,60,36 ; vmovups %ymm7,(%rsp) + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 DB 77,133,192 ; test %r8,%r8 - DB 15,133,31,2,0,0 ; jne 2160 <_sk_load_tables_avx+0x238> - DB 196,65,124,16,17 ; vmovups (%r9),%ymm10 - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx - DB 80 ; push %rax - DB 197,124,40,13,39,76,0,0 ; vmovaps 0x4c27(%rip),%ymm9 # 6b80 <_sk_callback_avx+0x620> - DB 196,193,44,84,193 ; vandps %ymm9,%ymm10,%ymm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,203 ; mov %r9d,%r11d + DB 15,133,251,1,0,0 ; jne 2166 <_sk_load_tables_avx+0x21e> + DB 196,65,124,16,18 ; vmovups (%r10),%ymm10 + DB 197,124,40,13,104,76,0,0 ; vmovaps 0x4c68(%rip),%ymm9 # 6be0 <_sk_callback_avx+0x62a> + DB 196,193,44,84,201 ; vandps %ymm9,%ymm10,%ymm1 + DB 196,227,125,25,200,1 ; vextractf128 $0x1,%ymm1,%xmm0 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,218 ; mov %r11d,%r10d + DB 72,139,88,8 ; mov 0x8(%rax),%rbx + DB 196,161,122,16,20,147 ; vmovss (%rbx,%r10,4),%xmm2 DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 - DB 69,137,214 ; mov %r10d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 196,163,105,33,4,155,16 ; vinsertps $0x10,(%rbx,%r11,4),%xmm2,%xmm0 + DB 68,137,209 ; mov %r10d,%ecx + DB 196,227,121,33,4,139,32 ; vinsertps $0x20,(%rbx,%rcx,4),%xmm0,%xmm0 + DB 196,193,249,126,203 ; vmovq %xmm1,%r11 DB 73,193,234,32 ; shr $0x20,%r10 - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,223 ; mov %ebx,%r15d - DB 196,227,249,22,193,1 ; vpextrq $0x1,%xmm0,%rcx - DB 65,137,205 ; mov %ecx,%r13d + DB 196,35,121,33,44,147,48 ; vinsertps $0x30,(%rbx,%r10,4),%xmm0,%xmm13 + DB 68,137,217 ; mov %r11d,%ecx + DB 197,250,16,20,139 ; vmovss (%rbx,%rcx,4),%xmm2 + DB 196,227,249,22,201,1 ; vpextrq $0x1,%xmm1,%rcx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 196,163,105,33,12,155,16 ; vinsertps $0x10,(%rbx,%r11,4),%xmm2,%xmm1 + DB 65,137,202 ; mov %ecx,%r10d DB 72,193,233,32 ; shr $0x20,%rcx - DB 72,193,235,32 ; shr $0x20,%rbx - DB 72,139,104,8 ; mov 0x8(%rax),%rbp - DB 76,139,96,16 ; mov 0x10(%rax),%r12 - DB 196,161,122,16,68,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm0 - DB 196,227,121,33,68,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm0,%xmm0 - DB 196,163,121,33,68,173,0,32 ; vinsertps $0x20,0x0(%rbp,%r13,4),%xmm0,%xmm0 - DB 196,227,121,33,68,141,0,48 ; vinsertps $0x30,0x0(%rbp,%rcx,4),%xmm0,%xmm0 - DB 196,161,122,16,76,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm1 - DB 196,163,113,33,76,141,0,16 ; vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,76,181,0,32 ; vinsertps $0x20,0x0(%rbp,%r14,4),%xmm1,%xmm1 - DB 196,163,113,33,76,149,0,48 ; vinsertps $0x30,0x0(%rbp,%r10,4),%xmm1,%xmm1 - DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 196,193,113,114,210,8 ; vpsrld $0x8,%xmm10,%xmm1 + DB 196,163,113,33,12,147,32 ; vinsertps $0x20,(%rbx,%r10,4),%xmm1,%xmm1 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 + DB 196,99,113,33,36,139,48 ; vinsertps $0x30,(%rbx,%rcx,4),%xmm1,%xmm12 + DB 196,193,105,114,210,8 ; vpsrld $0x8,%xmm10,%xmm2 DB 196,67,125,25,208,1 ; vextractf128 $0x1,%ymm10,%xmm8 - DB 196,193,105,114,208,8 ; vpsrld $0x8,%xmm8,%xmm2 - DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,193,116,84,201 ; vandps %ymm9,%ymm1,%ymm1 - DB 196,193,249,126,201 ; vmovq %xmm1,%r9 - DB 69,137,203 ; mov %r9d,%r11d - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 - DB 69,137,214 ; mov %r10d,%r14d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 - DB 196,225,249,126,203 ; vmovq %xmm1,%rbx - DB 65,137,223 ; mov %ebx,%r15d - DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp - DB 137,233 ; mov %ebp,%ecx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,129,122,16,12,188 ; vmovss (%r12,%r15,4),%xmm1 - DB 196,195,113,33,12,156,16 ; vinsertps $0x10,(%r12,%rbx,4),%xmm1,%xmm1 - DB 196,193,122,16,20,140 ; vmovss (%r12,%rcx,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,193,122,16,20,172 ; vmovss (%r12,%rbp,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,129,122,16,20,156 ; vmovss (%r12,%r11,4),%xmm2 - DB 196,131,105,33,20,140,16 ; vinsertps $0x10,(%r12,%r9,4),%xmm2,%xmm2 - DB 196,129,122,16,28,180 ; vmovss (%r12,%r14,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,129,122,16,28,148 ; vmovss (%r12,%r10,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 - DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,193,105,114,210,16 ; vpsrld $0x10,%xmm10,%xmm2 - DB 196,193,97,114,208,16 ; vpsrld $0x10,%xmm8,%xmm3 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 - DB 196,193,108,84,209 ; vandps %ymm9,%ymm2,%ymm2 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 69,137,202 ; mov %r9d,%r10d - DB 196,227,249,22,209,1 ; vpextrq $0x1,%xmm2,%rcx - DB 65,137,203 ; mov %ecx,%r11d + DB 196,193,121,114,208,8 ; vpsrld $0x8,%xmm8,%xmm0 + DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 + DB 196,193,124,84,209 ; vandps %ymm9,%ymm0,%ymm2 + DB 196,227,125,25,208,1 ; vextractf128 $0x1,%ymm2,%xmm0 + DB 196,225,249,126,193 ; vmovq %xmm0,%rcx + DB 137,203 ; mov %ecx,%ebx + DB 196,193,122,16,12,154 ; vmovss (%r10,%rbx,4),%xmm1 + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx DB 72,193,233,32 ; shr $0x20,%rcx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 - DB 196,225,249,126,213 ; vmovq %xmm2,%rbp - DB 65,137,238 ; mov %ebp,%r14d + DB 196,67,113,33,52,138,16 ; vinsertps $0x10,(%r10,%rcx,4),%xmm1,%xmm14 + DB 137,217 ; mov %ebx,%ecx + DB 196,193,122,16,28,138 ; vmovss (%r10,%rcx,4),%xmm3 + DB 196,225,249,126,209 ; vmovq %xmm2,%rcx + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,122,16,12,154 ; vmovss (%r10,%rbx,4),%xmm1 + DB 137,203 ; mov %ecx,%ebx + DB 196,193,122,16,4,154 ; vmovss (%r10,%rbx,4),%xmm0 DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,223 ; mov %ebx,%r15d + DB 72,193,233,32 ; shr $0x20,%rcx + DB 196,67,121,33,28,138,16 ; vinsertps $0x10,(%r10,%rcx,4),%xmm0,%xmm11 + DB 137,217 ; mov %ebx,%ecx + DB 196,65,122,16,60,138 ; vmovss (%r10,%rcx,4),%xmm15 + DB 196,195,29,24,197,1 ; vinsertf128 $0x1,%xmm13,%ymm12,%ymm0 DB 72,193,235,32 ; shr $0x20,%rbx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,105,33,20,168,16 ; vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2 - DB 196,161,122,16,28,184 ; vmovss (%rax,%r15,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 196,227,9,33,219,32 ; vinsertps $0x20,%xmm3,%xmm14,%xmm3 + DB 196,227,97,33,249,48 ; vinsertps $0x30,%xmm1,%xmm3,%xmm7 + DB 196,65,122,16,52,154 ; vmovss (%r10,%rbx,4),%xmm14 + DB 72,139,64,24 ; mov 0x18(%rax),%rax + DB 196,193,97,114,210,16 ; vpsrld $0x10,%xmm10,%xmm3 + DB 196,193,105,114,208,16 ; vpsrld $0x10,%xmm8,%xmm2 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 + DB 196,65,108,84,201 ; vandps %ymm9,%ymm2,%ymm9 + DB 196,99,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm2 + DB 196,225,249,126,209 ; vmovq %xmm2,%rcx + DB 137,203 ; mov %ecx,%ebx DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,99,105,33,203,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm9 - DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 - DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 + DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx + DB 72,193,233,32 ; shr $0x20,%rcx + DB 196,99,97,33,36,136,16 ; vinsertps $0x10,(%rax,%rcx,4),%xmm3,%xmm12 + DB 137,217 ; mov %ebx,%ecx DB 197,250,16,28,136 ; vmovss (%rax,%rcx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,195,109,24,209,1 ; vinsertf128 $0x1,%xmm9,%ymm2,%ymm2 - DB 196,193,49,114,210,24 ; vpsrld $0x18,%xmm10,%xmm9 - DB 196,193,97,114,208,24 ; vpsrld $0x18,%xmm8,%xmm3 - DB 196,227,53,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 + DB 196,97,249,126,201 ; vmovq %xmm9,%rcx + DB 72,193,235,32 ; shr $0x20,%rbx + DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 + DB 137,203 ; mov %ecx,%ebx + DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 + DB 196,99,249,22,203,1 ; vpextrq $0x1,%xmm9,%rbx + DB 72,193,233,32 ; shr $0x20,%rcx + DB 196,99,113,33,12,136,16 ; vinsertps $0x10,(%rax,%rcx,4),%xmm1,%xmm9 + DB 137,217 ; mov %ebx,%ecx + DB 197,122,16,44,136 ; vmovss (%rax,%rcx,4),%xmm13 + DB 196,195,33,33,207,32 ; vinsertps $0x20,%xmm15,%xmm11,%xmm1 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 197,122,16,28,152 ; vmovss (%rax,%rbx,4),%xmm11 + DB 196,195,113,33,206,48 ; vinsertps $0x30,%xmm14,%xmm1,%xmm1 + DB 196,227,117,24,207,1 ; vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 + DB 196,227,25,33,219,32 ; vinsertps $0x20,%xmm3,%xmm12,%xmm3 + DB 196,227,97,33,210,48 ; vinsertps $0x30,%xmm2,%xmm3,%xmm2 + DB 196,195,49,33,221,32 ; vinsertps $0x20,%xmm13,%xmm9,%xmm3 + DB 196,195,97,33,219,48 ; vinsertps $0x30,%xmm11,%xmm3,%xmm3 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 + DB 196,193,97,114,210,24 ; vpsrld $0x18,%xmm10,%xmm3 + DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7 + DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,114,70,0,0 ; vbroadcastss 0x4672(%rip),%ymm8 # 67b8 <_sk_callback_avx+0x258> - DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 + DB 196,226,125,24,61,199,70,0,0 ; vbroadcastss 0x46c7(%rip),%ymm7 # 6818 <_sk_callback_avx+0x262> + DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx - DB 72,131,196,8 ; add $0x8,%rsp + DB 76,137,201 ; mov %r9,%rcx + DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7 + DB 72,131,196,48 ; add $0x30,%rsp DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx DB 192,225,3 ; shl $0x3,%cl - DB 73,199,194,255,255,255,255 ; mov $0xffffffffffffffff,%r10 - DB 73,211,234 ; shr %cl,%r10 - DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 73,199,195,255,255,255,255 ; mov $0xffffffffffffffff,%r11 + DB 73,211,235 ; shr %cl,%r11 + DB 196,193,249,110,195 ; vmovq %r11,%xmm0 DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0 - DB 196,226,121,0,13,72,73,0,0 ; vpshufb 0x4948(%rip),%xmm0,%xmm1 # 6ad0 <_sk_callback_avx+0x570> + DB 196,226,121,0,13,162,73,0,0 ; vpshufb 0x49a2(%rip),%xmm0,%xmm1 # 6b30 <_sk_callback_avx+0x57a> DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1 - DB 196,226,121,0,5,74,73,0,0 ; vpshufb 0x494a(%rip),%xmm0,%xmm0 # 6ae0 <_sk_callback_avx+0x580> + DB 196,226,121,0,5,164,73,0,0 ; vpshufb 0x49a4(%rip),%xmm0,%xmm0 # 6b40 <_sk_callback_avx+0x58a> DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 196,66,125,44,17 ; vmaskmovps (%r9),%ymm0,%ymm10 - DB 233,155,253,255,255 ; jmpq 1f46 <_sk_load_tables_avx+0x1e> + DB 196,66,125,44,18 ; vmaskmovps (%r10),%ymm0,%ymm10 + DB 233,191,253,255,255 ; jmpq 1f70 <_sk_load_tables_avx+0x28> PUBLIC _sk_load_tables_u16_be_avx _sk_load_tables_u16_be_avx LABEL PROC + DB 72,131,236,56 ; sub $0x38,%rsp DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,113,2,0,0 ; jne 2432 <_sk_load_tables_u16_be_avx+0x287> - DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8 - DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2 - DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3 - DB 196,1,122,111,76,72,48 ; vmovdqu 0x30(%r8,%r9,2),%xmm9 - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 197,252,17,60,36 ; vmovups %ymm7,(%rsp) + DB 15,133,87,2,0,0 ; jne 2427 <_sk_load_tables_u16_be_avx+0x276> + DB 196,1,121,16,4,81 ; vmovupd (%r9,%r10,2),%xmm8 + DB 196,129,121,16,84,81,16 ; vmovupd 0x10(%r9,%r10,2),%xmm2 + DB 196,129,121,16,92,81,32 ; vmovupd 0x20(%r9,%r10,2),%xmm3 + DB 196,1,122,111,76,81,48 ; vmovdqu 0x30(%r9,%r10,2),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 @@ -7202,168 +7206,159 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 197,121,97,202 ; vpunpcklwd %xmm2,%xmm0,%xmm9 DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8 DB 197,241,97,195 ; vpunpcklwd %xmm3,%xmm1,%xmm0 - DB 197,113,105,227 ; vpunpckhwd %xmm3,%xmm1,%xmm12 - DB 197,177,108,208 ; vpunpcklqdq %xmm0,%xmm9,%xmm2 - DB 197,177,109,200 ; vpunpckhqdq %xmm0,%xmm9,%xmm1 - DB 196,65,57,108,212 ; vpunpcklqdq %xmm12,%xmm8,%xmm10 - DB 197,121,111,29,211,72,0,0 ; vmovdqa 0x48d3(%rip),%xmm11 # 6af0 <_sk_callback_avx+0x590> - DB 196,193,105,219,195 ; vpand %xmm11,%xmm2,%xmm0 + DB 197,113,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm11 + DB 197,177,108,200 ; vpunpcklqdq %xmm0,%xmm9,%xmm1 + DB 197,49,109,224 ; vpunpckhqdq %xmm0,%xmm9,%xmm12 + DB 197,121,111,21,51,73,0,0 ; vmovdqa 0x4933(%rip),%xmm10 # 6b50 <_sk_callback_avx+0x59a> + DB 196,193,113,219,202 ; vpand %xmm10,%xmm1,%xmm1 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,193,121,105,209 ; vpunpckhwd %xmm9,%xmm0,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,208 ; vmovq %xmm2,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,198 ; mov %r8,%r14 - DB 73,193,238,32 ; shr $0x20,%r14 - DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,223 ; mov %ebx,%r15d - DB 72,193,235,30 ; shr $0x1e,%rbx - DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 - DB 69,137,229 ; mov %r12d,%r13d - DB 73,193,236,30 ; shr $0x1e,%r12 - DB 72,139,104,8 ; mov 0x8(%rax),%rbp - DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 196,161,122,16,68,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm0 - DB 196,163,121,33,68,181,0,16 ; vinsertps $0x10,0x0(%rbp,%r14,4),%xmm0,%xmm0 - DB 196,161,122,16,84,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 196,161,122,16,84,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 - DB 196,227,105,33,84,29,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,1),%xmm2,%xmm2 - DB 196,161,122,16,92,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,92,37,0 ; vmovss 0x0(%rbp,%r12,1),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 - DB 196,193,113,219,203 ; vpand %xmm11,%xmm1,%xmm1 DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2 - DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,211 ; vmovq %xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 76,139,88,8 ; mov 0x8(%rax),%r11 + DB 196,129,122,16,28,147 ; vmovss (%r11,%r10,4),%xmm3 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,97,33,44,139,16 ; vinsertps $0x10,(%r11,%r9,4),%xmm3,%xmm13 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,32 ; shr $0x20,%r10 DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 - DB 196,193,249,126,203 ; vmovq %xmm1,%r11 - DB 69,137,222 ; mov %r11d,%r14d + DB 196,129,122,16,28,139 ; vmovss (%r11,%r9,4),%xmm3 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,129,122,16,4,147 ; vmovss (%r11,%r10,4),%xmm0 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,129,122,16,20,147 ; vmovss (%r11,%r10,4),%xmm2 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,131,105,33,12,11,16 ; vinsertps $0x10,(%r11,%r9,1),%xmm2,%xmm1 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,129,122,16,20,139 ; vmovss (%r11,%r9,4),%xmm2 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 196,227,17,33,219,32 ; vinsertps $0x20,%xmm3,%xmm13,%xmm3 + DB 196,99,97,33,232,48 ; vinsertps $0x30,%xmm0,%xmm3,%xmm13 + DB 196,99,113,33,242,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm14 + DB 196,1,122,16,60,19 ; vmovss (%r11,%r10,1),%xmm15 + DB 196,193,25,219,210 ; vpand %xmm10,%xmm12,%xmm2 + DB 196,193,105,105,193 ; vpunpckhwd %xmm9,%xmm2,%xmm0 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,129,122,16,12,153 ; vmovss (%r9,%r11,4),%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,3,113,33,36,145,16 ; vinsertps $0x10,(%r9,%r10,4),%xmm1,%xmm12 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 196,226,121,51,194 ; vpmovzxwd %xmm2,%xmm0 + DB 196,129,122,16,20,145 ; vmovss (%r9,%r10,4),%xmm2 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 + DB 196,129,122,16,28,153 ; vmovss (%r9,%r11,4),%xmm3 + DB 69,137,211 ; mov %r10d,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,129,122,16,12,153 ; vmovss (%r9,%r11,4),%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 196,131,113,33,12,17,16 ; vinsertps $0x10,(%r9,%r10,1),%xmm1,%xmm1 + DB 69,137,218 ; mov %r11d,%r10d + DB 196,129,122,16,60,145 ; vmovss (%r9,%r10,4),%xmm7 + DB 196,195,9,33,199,48 ; vinsertps $0x30,%xmm15,%xmm14,%xmm0 + DB 196,65,57,108,243 ; vpunpcklqdq %xmm11,%xmm8,%xmm14 + DB 196,195,125,24,197,1 ; vinsertf128 $0x1,%xmm13,%ymm0,%ymm0 DB 73,193,235,30 ; shr $0x1e,%r11 - DB 196,195,249,22,207,1 ; vpextrq $0x1,%xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,30 ; shr $0x1e,%r15 - DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1 - DB 196,195,113,33,12,152,16 ; vinsertps $0x10,(%r8,%rbx,4),%xmm1,%xmm1 - DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,193,122,16,20,168 ; vmovss (%r8,%rbp,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,129,122,16,20,176 ; vmovss (%r8,%r14,4),%xmm2 - DB 196,131,105,33,20,24,16 ; vinsertps $0x10,(%r8,%r11,1),%xmm2,%xmm2 - DB 196,129,122,16,28,160 ; vmovss (%r8,%r12,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,129,122,16,28,56 ; vmovss (%r8,%r15,1),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 + DB 196,227,25,33,210,32 ; vinsertps $0x20,%xmm2,%xmm12,%xmm2 + DB 196,227,105,33,219,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm3 + DB 196,99,113,33,239,32 ; vinsertps $0x20,%xmm7,%xmm1,%xmm13 + DB 196,1,122,16,60,25 ; vmovss (%r9,%r11,1),%xmm15 DB 76,139,80,24 ; mov 0x18(%rax),%r10 - DB 196,193,41,219,211 ; vpand %xmm11,%xmm10,%xmm2 - DB 196,193,105,105,217 ; vpunpckhwd %xmm9,%xmm2,%xmm3 - DB 196,227,249,22,221,1 ; vpextrq $0x1,%xmm3,%rbp - DB 65,137,232 ; mov %ebp,%r8d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,219 ; vmovq %xmm3,%rbx - DB 65,137,217 ; mov %ebx,%r9d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,225,249,126,208 ; vmovq %xmm2,%rax - DB 65,137,195 ; mov %eax,%r11d - DB 72,193,232,30 ; shr $0x1e,%rax - DB 196,195,249,22,214,1 ; vpextrq $0x1,%xmm2,%r14 - DB 69,137,247 ; mov %r14d,%r15d - DB 73,193,238,30 ; shr $0x1e,%r14 - DB 196,129,122,16,20,138 ; vmovss (%r10,%r9,4),%xmm2 - DB 196,195,105,33,20,154,16 ; vinsertps $0x10,(%r10,%rbx,4),%xmm2,%xmm2 - DB 196,129,122,16,28,130 ; vmovss (%r10,%r8,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,193,122,16,28,170 ; vmovss (%r10,%rbp,4),%xmm3 - DB 196,99,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm10 - DB 196,129,122,16,28,154 ; vmovss (%r10,%r11,4),%xmm3 - DB 196,195,97,33,28,2,16 ; vinsertps $0x10,(%r10,%rax,1),%xmm3,%xmm3 - DB 196,129,122,16,20,186 ; vmovss (%r10,%r15,4),%xmm2 - DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 - DB 196,129,122,16,28,50 ; vmovss (%r10,%r14,1),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,195,109,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm2,%ymm2 - DB 196,193,57,109,220 ; vpunpckhqdq %xmm12,%xmm8,%xmm3 - DB 197,185,113,243,8 ; vpsllw $0x8,%xmm3,%xmm8 + DB 196,193,9,219,250 ; vpand %xmm10,%xmm14,%xmm7 + DB 196,193,65,105,209 ; vpunpckhwd %xmm9,%xmm7,%xmm2 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 68,137,200 ; mov %r9d,%eax + DB 196,193,122,16,12,130 ; vmovss (%r10,%rax,4),%xmm1 + DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,113,33,20,138,16 ; vinsertps $0x10,(%r10,%r9,4),%xmm1,%xmm10 + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,226,121,51,207 ; vpmovzxwd %xmm7,%xmm1 + DB 196,1,122,16,52,138 ; vmovss (%r10,%r9,4),%xmm14 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,65,122,16,36,130 ; vmovss (%r10,%rax,4),%xmm12 + DB 68,137,200 ; mov %r9d,%eax + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,193,122,16,20,130 ; vmovss (%r10,%rax,4),%xmm2 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 196,131,105,33,20,10,16 ; vinsertps $0x10,(%r10,%r9,1),%xmm2,%xmm2 + DB 68,137,216 ; mov %r11d,%eax + DB 196,193,122,16,60,130 ; vmovss (%r10,%rax,4),%xmm7 + DB 196,195,17,33,207,48 ; vinsertps $0x30,%xmm15,%xmm13,%xmm1 + DB 73,193,235,30 ; shr $0x1e,%r11 + DB 196,1,122,16,44,26 ; vmovss (%r10,%r11,1),%xmm13 + DB 196,227,117,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm1,%ymm1 + DB 196,195,41,33,222,32 ; vinsertps $0x20,%xmm14,%xmm10,%xmm3 + DB 196,195,97,33,220,48 ; vinsertps $0x30,%xmm12,%xmm3,%xmm3 + DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 + DB 196,195,105,33,213,48 ; vinsertps $0x30,%xmm13,%xmm2,%xmm2 + DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 + DB 196,193,57,109,219 ; vpunpckhqdq %xmm11,%xmm8,%xmm3 + DB 197,193,113,243,8 ; vpsllw $0x8,%xmm3,%xmm7 DB 197,225,113,211,8 ; vpsrlw $0x8,%xmm3,%xmm3 - DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3 - DB 196,65,97,105,193 ; vpunpckhwd %xmm9,%xmm3,%xmm8 + DB 197,193,235,219 ; vpor %xmm3,%xmm7,%xmm3 + DB 196,193,97,105,249 ; vpunpckhwd %xmm9,%xmm3,%xmm7 DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 - DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 + DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,157,67,0,0 ; vbroadcastss 0x439d(%rip),%ymm8 # 67bc <_sk_callback_avx+0x25c> - DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 + DB 196,226,125,24,61,6,68,0,0 ; vbroadcastss 0x4406(%rip),%ymm7 # 681c <_sk_callback_avx+0x266> + DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp + DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7 + DB 72,131,196,56 ; add $0x38,%rsp DB 255,224 ; jmpq *%rax - DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8 + DB 196,1,123,16,4,81 ; vmovsd (%r9,%r10,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 2498 <_sk_load_tables_u16_be_avx+0x2ed> - DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 2498 <_sk_load_tables_u16_be_avx+0x2ed> - DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 24a5 <_sk_load_tables_u16_be_avx+0x2fa> - DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 24a5 <_sk_load_tables_u16_be_avx+0x2fa> - DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,97,253,255,255 ; je 21dc <_sk_load_tables_u16_be_avx+0x31> - DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,80,253,255,255 ; jb 21dc <_sk_load_tables_u16_be_avx+0x31> - DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9 - DB 233,68,253,255,255 ; jmpq 21dc <_sk_load_tables_u16_be_avx+0x31> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,85 ; je 248d <_sk_load_tables_u16_be_avx+0x2dc> + DB 196,1,57,22,68,81,8 ; vmovhpd 0x8(%r9,%r10,2),%xmm8,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,72 ; jb 248d <_sk_load_tables_u16_be_avx+0x2dc> + DB 196,129,123,16,84,81,16 ; vmovsd 0x10(%r9,%r10,2),%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 116,72 ; je 249a <_sk_load_tables_u16_be_avx+0x2e9> + DB 196,129,105,22,84,81,24 ; vmovhpd 0x18(%r9,%r10,2),%xmm2,%xmm2 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,59 ; jb 249a <_sk_load_tables_u16_be_avx+0x2e9> + DB 196,129,123,16,92,81,32 ; vmovsd 0x20(%r9,%r10,2),%xmm3 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 15,132,123,253,255,255 ; je 21eb <_sk_load_tables_u16_be_avx+0x3a> + DB 196,129,97,22,92,81,40 ; vmovhpd 0x28(%r9,%r10,2),%xmm3,%xmm3 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 15,130,106,253,255,255 ; jb 21eb <_sk_load_tables_u16_be_avx+0x3a> + DB 196,1,122,126,76,81,48 ; vmovq 0x30(%r9,%r10,2),%xmm9 + DB 233,94,253,255,255 ; jmpq 21eb <_sk_load_tables_u16_be_avx+0x3a> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,55,253,255,255 ; jmpq 21dc <_sk_load_tables_u16_be_avx+0x31> + DB 233,81,253,255,255 ; jmpq 21eb <_sk_load_tables_u16_be_avx+0x3a> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,46,253,255,255 ; jmpq 21dc <_sk_load_tables_u16_be_avx+0x31> + DB 233,72,253,255,255 ; jmpq 21eb <_sk_load_tables_u16_be_avx+0x3a> PUBLIC _sk_load_tables_rgb_u16_be_avx _sk_load_tables_rgb_u16_be_avx LABEL PROC + DB 72,131,236,88 ; sub $0x58,%rsp DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,93,2,0,0 ; jne 271d <_sk_load_tables_rgb_u16_be_avx+0x26f> - DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0 - DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2 - DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1 - DB 196,129,122,111,92,72,32 ; vmovdqu 0x20(%r8,%r9,2),%xmm3 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,82 ; lea (%rdx,%rdx,2),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 197,252,17,124,36,32 ; vmovups %ymm7,0x20(%rsp) + DB 197,252,17,52,36 ; vmovups %ymm6,(%rsp) + DB 15,133,74,2,0,0 ; jne 270e <_sk_load_tables_rgb_u16_be_avx+0x26b> + DB 196,129,122,111,4,81 ; vmovdqu (%r9,%r10,2),%xmm0 + DB 196,129,122,111,84,81,12 ; vmovdqu 0xc(%r9,%r10,2),%xmm2 + DB 196,129,122,111,76,81,24 ; vmovdqu 0x18(%r9,%r10,2),%xmm1 + DB 196,129,122,111,92,81,32 ; vmovdqu 0x20(%r9,%r10,2),%xmm3 DB 197,225,115,219,4 ; vpsrldq $0x4,%xmm3,%xmm3 DB 197,185,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm8 DB 197,177,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm9 DB 197,161,115,217,6 ; vpsrldq $0x6,%xmm1,%xmm11 DB 197,169,115,219,6 ; vpsrldq $0x6,%xmm3,%xmm10 - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx DB 197,249,97,194 ; vpunpcklwd %xmm2,%xmm0,%xmm0 DB 196,193,57,97,209 ; vpunpcklwd %xmm9,%xmm8,%xmm2 DB 197,241,97,203 ; vpunpcklwd %xmm3,%xmm1,%xmm1 @@ -7371,434 +7366,400 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 197,121,97,194 ; vpunpcklwd %xmm2,%xmm0,%xmm8 DB 197,249,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm0 DB 197,241,97,211 ; vpunpcklwd %xmm3,%xmm1,%xmm2 - DB 197,241,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm3 - DB 197,185,108,202 ; vpunpcklqdq %xmm2,%xmm8,%xmm1 - DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2 - DB 197,121,108,195 ; vpunpcklqdq %xmm3,%xmm0,%xmm8 - DB 197,121,111,13,204,69,0,0 ; vmovdqa 0x45cc(%rip),%xmm9 # 6b00 <_sk_callback_avx+0x5a0> - DB 196,193,113,219,193 ; vpand %xmm9,%xmm1,%xmm0 + DB 197,241,105,203 ; vpunpckhwd %xmm3,%xmm1,%xmm1 + DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3 + DB 197,57,109,218 ; vpunpckhqdq %xmm2,%xmm8,%xmm11 + DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8 + DB 197,121,111,13,50,70,0,0 ; vmovdqa 0x4632(%rip),%xmm9 # 6b60 <_sk_callback_avx+0x5aa> + DB 196,193,97,219,193 ; vpand %xmm9,%xmm3,%xmm0 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 - DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 77,137,194 ; mov %r8,%r10 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 76,139,88,8 ; mov 0x8(%rax),%r11 + DB 196,129,122,16,20,147 ; vmovss (%r11,%r10,4),%xmm2 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,105,33,36,139,16 ; vinsertps $0x10,(%r11,%r9,4),%xmm2,%xmm12 + DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,198 ; mov %r8,%r14 - DB 73,193,238,32 ; shr $0x20,%r14 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,223 ; mov %ebx,%r15d - DB 72,193,235,30 ; shr $0x1e,%rbx - DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 - DB 69,137,229 ; mov %r12d,%r13d - DB 73,193,236,30 ; shr $0x1e,%r12 - DB 72,139,104,8 ; mov 0x8(%rax),%rbp - DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 196,161,122,16,68,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm0 - DB 196,163,121,33,68,181,0,16 ; vinsertps $0x10,0x0(%rbp,%r14,4),%xmm0,%xmm0 - DB 196,161,122,16,76,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm1 - DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 - DB 196,161,122,16,76,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm1 - DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 - DB 196,161,122,16,76,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm1 - DB 196,227,113,33,76,29,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,1),%xmm1,%xmm1 - DB 196,161,122,16,92,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm3 - DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 - DB 196,161,122,16,92,37,0 ; vmovss 0x0(%rbp,%r12,1),%xmm3 - DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 - DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 196,193,105,219,201 ; vpand %xmm9,%xmm2,%xmm1 - DB 196,193,113,105,210 ; vpunpckhwd %xmm10,%xmm1,%xmm2 - DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,211 ; vmovq %xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 - DB 196,193,249,126,203 ; vmovq %xmm1,%r11 - DB 69,137,222 ; mov %r11d,%r14d + DB 196,129,122,16,20,139 ; vmovss (%r11,%r9,4),%xmm2 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 196,129,122,16,12,147 ; vmovss (%r11,%r10,4),%xmm1 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,129,122,16,28,147 ; vmovss (%r11,%r10,4),%xmm3 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,131,97,33,28,11,16 ; vinsertps $0x10,(%r11,%r9,1),%xmm3,%xmm3 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,129,122,16,4,139 ; vmovss (%r11,%r9,4),%xmm0 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 196,227,25,33,210,32 ; vinsertps $0x20,%xmm2,%xmm12,%xmm2 + DB 196,227,105,33,201,48 ; vinsertps $0x30,%xmm1,%xmm2,%xmm1 + DB 196,129,122,16,20,19 ; vmovss (%r11,%r10,1),%xmm2 + DB 196,65,33,219,225 ; vpand %xmm9,%xmm11,%xmm12 + DB 196,65,25,105,218 ; vpunpckhwd %xmm10,%xmm12,%xmm11 + DB 196,65,249,126,218 ; vmovq %xmm11,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,1,122,16,44,153 ; vmovss (%r9,%r11,4),%xmm13 + DB 196,67,249,22,219,1 ; vpextrq $0x1,%xmm11,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,3,17,33,28,145,16 ; vinsertps $0x10,(%r9,%r10,4),%xmm13,%xmm11 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 196,66,121,51,244 ; vpmovzxwd %xmm12,%xmm14 + DB 196,1,122,16,44,145 ; vmovss (%r9,%r10,4),%xmm13 + DB 196,65,249,126,242 ; vmovq %xmm14,%r10 + DB 196,1,122,16,36,153 ; vmovss (%r9,%r11,4),%xmm12 + DB 69,137,211 ; mov %r10d,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,1,122,16,60,153 ; vmovss (%r9,%r11,4),%xmm15 + DB 196,67,249,22,243,1 ; vpextrq $0x1,%xmm14,%r11 + DB 196,3,1,33,52,17,16 ; vinsertps $0x10,(%r9,%r10,1),%xmm15,%xmm14 + DB 69,137,218 ; mov %r11d,%r10d + DB 196,1,122,16,60,145 ; vmovss (%r9,%r10,4),%xmm15 + DB 196,227,97,33,192,32 ; vinsertps $0x20,%xmm0,%xmm3,%xmm0 + DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 + DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 73,193,235,30 ; shr $0x1e,%r11 - DB 196,195,249,22,207,1 ; vpextrq $0x1,%xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,30 ; shr $0x1e,%r15 - DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1 - DB 196,195,113,33,12,152,16 ; vinsertps $0x10,(%r8,%rbx,4),%xmm1,%xmm1 - DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,193,122,16,20,168 ; vmovss (%r8,%rbp,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,129,122,16,20,176 ; vmovss (%r8,%r14,4),%xmm2 - DB 196,131,105,33,20,24,16 ; vinsertps $0x10,(%r8,%r11,1),%xmm2,%xmm2 - DB 196,129,122,16,28,160 ; vmovss (%r8,%r12,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,129,122,16,28,56 ; vmovss (%r8,%r15,1),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 + DB 196,129,122,16,52,25 ; vmovss (%r9,%r11,1),%xmm6 DB 76,139,80,24 ; mov 0x18(%rax),%r10 - DB 196,193,57,219,209 ; vpand %xmm9,%xmm8,%xmm2 - DB 196,193,105,105,218 ; vpunpckhwd %xmm10,%xmm2,%xmm3 - DB 196,227,249,22,221,1 ; vpextrq $0x1,%xmm3,%rbp - DB 65,137,232 ; mov %ebp,%r8d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,219 ; vmovq %xmm3,%rbx - DB 65,137,217 ; mov %ebx,%r9d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,225,249,126,208 ; vmovq %xmm2,%rax - DB 65,137,195 ; mov %eax,%r11d - DB 72,193,232,30 ; shr $0x1e,%rax - DB 196,195,249,22,214,1 ; vpextrq $0x1,%xmm2,%r14 - DB 69,137,247 ; mov %r14d,%r15d - DB 73,193,238,30 ; shr $0x1e,%r14 - DB 196,129,122,16,20,138 ; vmovss (%r10,%r9,4),%xmm2 - DB 196,195,105,33,20,154,16 ; vinsertps $0x10,(%r10,%rbx,4),%xmm2,%xmm2 - DB 196,129,122,16,28,130 ; vmovss (%r10,%r8,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,193,122,16,28,170 ; vmovss (%r10,%rbp,4),%xmm3 - DB 196,99,105,33,195,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm8 - DB 196,129,122,16,28,154 ; vmovss (%r10,%r11,4),%xmm3 - DB 196,195,97,33,28,2,16 ; vinsertps $0x10,(%r10,%rax,1),%xmm3,%xmm3 - DB 196,129,122,16,20,186 ; vmovss (%r10,%r15,4),%xmm2 - DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 - DB 196,129,122,16,28,50 ; vmovss (%r10,%r14,1),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,195,109,24,208,1 ; vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 + DB 196,65,57,219,193 ; vpand %xmm9,%xmm8,%xmm8 + DB 196,193,57,105,210 ; vpunpckhwd %xmm10,%xmm8,%xmm2 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 68,137,200 ; mov %r9d,%eax + DB 196,193,122,16,12,130 ; vmovss (%r10,%rax,4),%xmm1 + DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,113,33,12,138,16 ; vinsertps $0x10,(%r10,%r9,4),%xmm1,%xmm9 + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,194,121,51,200 ; vpmovzxwd %xmm8,%xmm1 + DB 196,1,122,16,4,138 ; vmovss (%r10,%r9,4),%xmm8 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,65,122,16,20,130 ; vmovss (%r10,%rax,4),%xmm10 + DB 68,137,200 ; mov %r9d,%eax + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,193,122,16,20,130 ; vmovss (%r10,%rax,4),%xmm2 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 196,131,105,33,20,10,16 ; vinsertps $0x10,(%r10,%r9,1),%xmm2,%xmm2 + DB 68,137,216 ; mov %r11d,%eax + DB 196,193,122,16,60,130 ; vmovss (%r10,%rax,4),%xmm7 + DB 196,195,33,33,205,32 ; vinsertps $0x20,%xmm13,%xmm11,%xmm1 + DB 73,193,235,30 ; shr $0x1e,%r11 + DB 196,1,122,16,28,26 ; vmovss (%r10,%r11,1),%xmm11 + DB 196,195,113,33,204,48 ; vinsertps $0x30,%xmm12,%xmm1,%xmm1 + DB 196,195,9,33,223,32 ; vinsertps $0x20,%xmm15,%xmm14,%xmm3 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 + DB 196,195,49,33,216,32 ; vinsertps $0x20,%xmm8,%xmm9,%xmm3 + DB 196,195,97,33,218,48 ; vinsertps $0x30,%xmm10,%xmm3,%xmm3 + DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 + DB 196,195,105,33,211,48 ; vinsertps $0x30,%xmm11,%xmm2,%xmm2 + DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,175,64,0,0 ; vbroadcastss 0x40af(%rip),%ymm3 # 67c0 <_sk_callback_avx+0x260> - DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp + DB 196,226,125,24,29,35,65,0,0 ; vbroadcastss 0x4123(%rip),%ymm3 # 6820 <_sk_callback_avx+0x26a> + DB 197,252,16,52,36 ; vmovups (%rsp),%ymm6 + DB 197,252,16,124,36,32 ; vmovups 0x20(%rsp),%ymm7 + DB 72,131,196,88 ; add $0x58,%rsp DB 255,224 ; jmpq *%rax - DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0 - DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 2736 <_sk_load_tables_rgb_u16_be_avx+0x288> - DB 233,190,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1 - DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 2765 <_sk_load_tables_rgb_u16_be_avx+0x2b7> - DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1 - DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 276a <_sk_load_tables_rgb_u16_be_avx+0x2bc> - DB 233,143,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,138,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1 - DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 2799 <_sk_load_tables_rgb_u16_be_avx+0x2eb> - DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1 - DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 279e <_sk_load_tables_rgb_u16_be_avx+0x2f0> - DB 233,91,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,86,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3 - DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 27c7 <_sk_load_tables_rgb_u16_be_avx+0x319> - DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3 - DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 - DB 233,45,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,40,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 196,129,121,110,4,81 ; vmovd (%r9,%r10,2),%xmm0 + DB 196,129,121,196,68,81,4,2 ; vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,5 ; jne 2727 <_sk_load_tables_rgb_u16_be_avx+0x284> + DB 233,209,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> + DB 196,129,121,110,76,81,6 ; vmovd 0x6(%r9,%r10,2),%xmm1 + DB 196,1,113,196,68,81,10,2 ; vpinsrw $0x2,0xa(%r9,%r10,2),%xmm1,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,26 ; jb 2756 <_sk_load_tables_rgb_u16_be_avx+0x2b3> + DB 196,129,121,110,76,81,12 ; vmovd 0xc(%r9,%r10,2),%xmm1 + DB 196,129,113,196,84,81,16,2 ; vpinsrw $0x2,0x10(%r9,%r10,2),%xmm1,%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 117,10 ; jne 275b <_sk_load_tables_rgb_u16_be_avx+0x2b8> + DB 233,162,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> + DB 233,157,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> + DB 196,129,121,110,76,81,18 ; vmovd 0x12(%r9,%r10,2),%xmm1 + DB 196,1,113,196,76,81,22,2 ; vpinsrw $0x2,0x16(%r9,%r10,2),%xmm1,%xmm9 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,26 ; jb 278a <_sk_load_tables_rgb_u16_be_avx+0x2e7> + DB 196,129,121,110,76,81,24 ; vmovd 0x18(%r9,%r10,2),%xmm1 + DB 196,129,113,196,76,81,28,2 ; vpinsrw $0x2,0x1c(%r9,%r10,2),%xmm1,%xmm1 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 117,10 ; jne 278f <_sk_load_tables_rgb_u16_be_avx+0x2ec> + DB 233,110,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> + DB 233,105,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> + DB 196,129,121,110,92,81,30 ; vmovd 0x1e(%r9,%r10,2),%xmm3 + DB 196,1,97,196,92,81,34,2 ; vpinsrw $0x2,0x22(%r9,%r10,2),%xmm3,%xmm11 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,20 ; jb 27b8 <_sk_load_tables_rgb_u16_be_avx+0x315> + DB 196,129,121,110,92,81,36 ; vmovd 0x24(%r9,%r10,2),%xmm3 + DB 196,129,97,196,92,81,40,2 ; vpinsrw $0x2,0x28(%r9,%r10,2),%xmm3,%xmm3 + DB 233,64,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> + DB 233,59,253,255,255 ; jmpq 24f8 <_sk_load_tables_rgb_u16_be_avx+0x55> PUBLIC _sk_byte_tables_avx _sk_byte_tables_avx LABEL PROC - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,227,63,0,0 ; vbroadcastss 0x3fe3(%rip),%ymm8 # 67c4 <_sk_callback_avx+0x264> + DB 196,98,125,24,5,92,64,0,0 ; vbroadcastss 0x405c(%rip),%ymm8 # 6824 <_sk_callback_avx+0x26e> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 + DB 196,65,249,126,201 ; vmovq %xmm9,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 76,139,24 ; mov (%rax),%r11 + DB 196,131,121,32,4,19,0 ; vpinsrb $0x0,(%r11,%r10,1),%xmm0,%xmm0 + DB 196,67,249,22,202,1 ; vpextrq $0x1,%xmm9,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,121,32,20,11,1 ; vpinsrb $0x1,(%r11,%r9,1),%xmm0,%xmm10 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,67,41,32,201,2 ; vpinsrb $0x2,%r9d,%xmm10,%xmm9 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,49,32,202,3 ; vpinsrb $0x3,%r10d,%xmm9,%xmm9 + DB 69,137,202 ; mov %r9d,%r10d + DB 196,3,121,32,20,19,0 ; vpinsrb $0x0,(%r11,%r10,1),%xmm0,%xmm10 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,41,32,4,11,1 ; vpinsrb $0x1,(%r11,%r9,1),%xmm10,%xmm0 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,121,32,210,3 ; vpinsrb $0x3,%r10d,%xmm0,%xmm10 + DB 196,193,116,89,192 ; vmulps %ymm8,%ymm1,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 68,137,197 ; mov %r8d,%ebp - DB 77,137,194 ; mov %r8,%r10 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,12,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,199 ; mov %r8,%r15 - DB 73,193,239,32 ; shr $0x20,%r15 + DB 196,131,113,32,12,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm1,%xmm1 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 69,137,198 ; mov %r8d,%r14d - DB 77,137,196 ; mov %r8,%r12 - DB 73,193,236,32 ; shr $0x20,%r12 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,221 ; mov %ebx,%r13d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 76,139,8 ; mov (%rax),%r9 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 - DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp - DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp - DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0 - DB 196,98,121,49,200 ; vpmovzxbd %xmm0,%xmm9 - DB 196,131,121,32,4,41,0 ; vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0 - DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 - DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp - DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp - DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0 - DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 - DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,49,63,0,0 ; vbroadcastss 0x3f31(%rip),%ymm9 # 67c8 <_sk_callback_avx+0x268> - DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 - DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,113,32,202,2 ; vpinsrb $0x2,%r10d,%xmm1,%xmm1 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,67,113,32,227,3 ; vpinsrb $0x3,%r11d,%xmm1,%xmm12 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,12,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,113,32,4,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm1,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 + DB 71,15,182,12,25 ; movzbl (%r9,%r11,1),%r9d + DB 196,67,121,32,233,3 ; vpinsrb $0x3,%r9d,%xmm0,%xmm13 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 196,193,108,89,200 ; vmulps %ymm8,%ymm2,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 - DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,203 ; vmovq %xmm1,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,20,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm2 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,105,32,20,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm2,%xmm2 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,105,32,210,2 ; vpinsrb $0x2,%r10d,%xmm2,%xmm2 + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,67,105,32,219,3 ; vpinsrb $0x3,%r11d,%xmm2,%xmm11 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,20,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm2 DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 - DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,105,32,12,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm2,%xmm1 + DB 69,137,218 ; mov %r11d,%r10d + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,113,32,202,2 ; vpinsrb $0x2,%r10d,%xmm1,%xmm1 DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,193,249,126,207 ; vmovq %xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 - DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp - DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1 - DB 196,98,121,49,209 ; vpmovzxbd %xmm1,%xmm10 - DB 196,131,121,32,12,32,0 ; vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1 - DB 196,131,113,32,12,56,1 ; vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1 - DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp - DB 196,227,113,32,205,2 ; vpinsrb $0x2,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp - DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1 - DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 - DB 196,227,45,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 + DB 71,15,182,12,25 ; movzbl (%r9,%r11,1),%r9d + DB 196,67,113,32,241,3 ; vpinsrb $0x3,%r9d,%xmm1,%xmm14 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 + DB 196,193,100,89,200 ; vmulps %ymm8,%ymm3,%ymm1 + DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 68,137,200 ; mov %r9d,%eax + DB 196,195,121,32,28,2,0 ; vpinsrb $0x0,(%r10,%rax,1),%xmm0,%xmm3 + DB 196,227,249,22,200,1 ; vpextrq $0x1,%xmm1,%rax + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,97,32,28,10,1 ; vpinsrb $0x1,(%r10,%r9,1),%xmm3,%xmm3 + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,195,97,32,217,2 ; vpinsrb $0x2,%r9d,%xmm3,%xmm3 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,99,97,32,192,3 ; vpinsrb $0x3,%eax,%xmm3,%xmm8 + DB 68,137,200 ; mov %r9d,%eax + DB 196,195,121,32,4,2,0 ; vpinsrb $0x0,(%r10,%rax,1),%xmm0,%xmm0 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,121,32,4,10,1 ; vpinsrb $0x1,(%r10,%r9,1),%xmm0,%xmm0 + DB 68,137,216 ; mov %r11d,%eax + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,99,121,32,248,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm15 + DB 73,193,235,32 ; shr $0x20,%r11 + DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax + DB 196,194,121,49,193 ; vpmovzxbd %xmm9,%xmm0 + DB 196,194,121,49,202 ; vpmovzxbd %xmm10,%xmm1 + DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 + DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 + DB 196,194,121,49,204 ; vpmovzxbd %xmm12,%xmm1 + DB 196,194,121,49,213 ; vpmovzxbd %xmm13,%xmm2 + DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 + DB 196,98,125,24,13,6,62,0,0 ; vbroadcastss 0x3e06(%rip),%ymm9 # 6828 <_sk_callback_avx+0x272> + DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 - DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 - DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,211 ; vmovq %xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 - DB 196,195,249,22,211,1 ; vpextrq $0x1,%xmm2,%r11 - DB 69,137,222 ; mov %r11d,%r14d - DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,193,249,126,215 ; vmovq %xmm2,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 - DB 196,195,105,32,20,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2 - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp - DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2 - DB 196,98,121,49,210 ; vpmovzxbd %xmm2,%xmm10 - DB 196,131,121,32,20,32,0 ; vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm2 - DB 196,131,105,32,20,56,1 ; vpinsrb $0x1,(%r8,%r15,1),%xmm2,%xmm2 - DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp - DB 196,227,105,32,213,2 ; vpinsrb $0x2,%ebp,%xmm2,%xmm2 - DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp - DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2 - DB 196,226,121,49,210 ; vpmovzxbd %xmm2,%xmm2 - DB 196,227,45,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 + DB 196,194,121,49,211 ; vpmovzxbd %xmm11,%xmm2 + DB 196,194,121,49,222 ; vpmovzxbd %xmm14,%xmm3 + DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 - DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 - DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 - DB 196,227,249,22,221,1 ; vpextrq $0x1,%xmm3,%rbp - DB 65,137,232 ; mov %ebp,%r8d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,219 ; vmovq %xmm3,%rbx - DB 65,137,217 ; mov %ebx,%r9d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,227,125,25,219,1 ; vextractf128 $0x1,%ymm3,%xmm3 - DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,222 ; vmovq %xmm3,%r14 - DB 69,137,247 ; mov %r14d,%r15d - DB 73,193,238,32 ; shr $0x20,%r14 - DB 196,163,121,32,28,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3 - DB 196,227,97,32,28,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3 - DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx - DB 196,227,97,32,219,2 ; vpinsrb $0x2,%ebx,%xmm3,%xmm3 - DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp - DB 196,227,97,32,221,3 ; vpinsrb $0x3,%ebp,%xmm3,%xmm3 - DB 196,98,121,49,195 ; vpmovzxbd %xmm3,%xmm8 - DB 196,163,121,32,28,56,0 ; vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm3 - DB 196,163,97,32,28,48,1 ; vpinsrb $0x1,(%rax,%r14,1),%xmm3,%xmm3 - DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp - DB 196,227,97,32,221,2 ; vpinsrb $0x2,%ebp,%xmm3,%xmm3 - DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax - DB 196,227,97,32,216,3 ; vpinsrb $0x3,%eax,%xmm3,%xmm3 + DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8 + DB 196,227,1,32,216,3 ; vpinsrb $0x3,%eax,%xmm15,%xmm3 DB 196,226,121,49,219 ; vpmovzxbd %xmm3,%xmm3 DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 196,193,100,89,217 ; vmulps %ymm9,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_byte_tables_rgb_avx _sk_byte_tables_rgb_avx LABEL PROC - DB 85 ; push %rbp - DB 65,87 ; push %r15 - DB 65,86 ; push %r14 - DB 65,85 ; push %r13 - DB 65,84 ; push %r12 - DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,139,64,24 ; mov 0x18(%rax),%r8d - DB 65,255,200 ; dec %r8d - DB 196,65,121,110,192 ; vmovd %r8d,%xmm8 + DB 68,139,72,24 ; mov 0x18(%rax),%r9d + DB 65,255,201 ; dec %r9d + DB 196,65,121,110,193 ; vmovd %r9d,%xmm8 DB 196,65,121,112,192,0 ; vpshufd $0x0,%xmm8,%xmm8 DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 68,137,197 ; mov %r8d,%ebp - DB 77,137,194 ; mov %r8,%r10 + DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 + DB 196,65,249,126,201 ; vmovq %xmm9,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 76,139,24 ; mov (%rax),%r11 + DB 196,131,121,32,4,19,0 ; vpinsrb $0x0,(%r11,%r10,1),%xmm0,%xmm0 + DB 196,67,249,22,202,1 ; vpextrq $0x1,%xmm9,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,3,121,32,20,11,1 ; vpinsrb $0x1,(%r11,%r9,1),%xmm0,%xmm10 + DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,199 ; mov %r8,%r15 - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 69,137,198 ; mov %r8d,%r14d - DB 77,137,196 ; mov %r8,%r12 - DB 73,193,236,32 ; shr $0x20,%r12 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,221 ; mov %ebx,%r13d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 76,139,8 ; mov (%rax),%r9 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 - DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp - DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp - DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0 - DB 196,98,121,49,200 ; vpmovzxbd %xmm0,%xmm9 - DB 196,131,121,32,4,41,0 ; vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0 - DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 - DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp - DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 - DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp - DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0 - DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 - DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,87,60,0,0 ; vbroadcastss 0x3c57(%rip),%ymm9 # 67cc <_sk_callback_avx+0x26c> - DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,67,41,32,201,2 ; vpinsrb $0x2,%r9d,%xmm10,%xmm9 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,49,32,202,3 ; vpinsrb $0x3,%r10d,%xmm9,%xmm9 + DB 69,137,202 ; mov %r9d,%r10d + DB 196,3,121,32,20,19,0 ; vpinsrb $0x0,(%r11,%r10,1),%xmm0,%xmm10 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,41,32,4,11,1 ; vpinsrb $0x1,(%r11,%r9,1),%xmm10,%xmm0 + DB 69,137,209 ; mov %r10d,%r9d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 71,15,182,12,11 ; movzbl (%r11,%r9,1),%r9d + DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 71,15,182,20,19 ; movzbl (%r11,%r10,1),%r10d + DB 196,67,121,32,218,3 ; vpinsrb $0x3,%r10d,%xmm0,%xmm11 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 - DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp - DB 65,137,233 ; mov %ebp,%r9d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,203 ; vmovq %xmm1,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,121,32,4,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm0,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 + DB 196,193,249,126,202 ; vmovq %xmm1,%r10 + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 196,67,121,32,211,3 ; vpinsrb $0x3,%r11d,%xmm0,%xmm10 + DB 69,137,211 ; mov %r10d,%r11d + DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 - DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,131,121,32,4,17,1 ; vpinsrb $0x1,(%r9,%r10,1),%xmm0,%xmm0 + DB 69,137,218 ; mov %r11d,%r10d + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 196,195,121,32,194,2 ; vpinsrb $0x2,%r10d,%xmm0,%xmm0 DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,193,249,126,207 ; vmovq %xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 - DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp - DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1 - DB 196,98,121,49,209 ; vpmovzxbd %xmm1,%xmm10 - DB 196,131,121,32,12,32,0 ; vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1 - DB 196,131,113,32,12,56,1 ; vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1 - DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp - DB 196,227,113,32,205,2 ; vpinsrb $0x2,%ebp,%xmm1,%xmm1 - DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp - DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1 - DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 - DB 196,227,45,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 + DB 71,15,182,12,25 ; movzbl (%r9,%r11,1),%r9d + DB 196,67,121,32,225,3 ; vpinsrb $0x3,%r9d,%xmm0,%xmm12 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 + DB 197,188,89,194 ; vmulps %ymm2,%ymm8,%ymm0 + DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 68,137,200 ; mov %r9d,%eax + DB 196,195,121,32,20,2,0 ; vpinsrb $0x0,(%r10,%rax,1),%xmm0,%xmm2 + DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,105,32,20,10,1 ; vpinsrb $0x1,(%r10,%r9,1),%xmm2,%xmm2 + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,195,105,32,209,2 ; vpinsrb $0x2,%r9d,%xmm2,%xmm2 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,99,105,32,192,3 ; vpinsrb $0x3,%eax,%xmm2,%xmm8 + DB 68,137,200 ; mov %r9d,%eax + DB 196,195,121,32,12,2,0 ; vpinsrb $0x0,(%r10,%rax,1),%xmm0,%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,131,113,32,4,10,1 ; vpinsrb $0x1,(%r10,%r9,1),%xmm1,%xmm0 + DB 68,137,216 ; mov %r11d,%eax + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 196,99,121,32,232,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm13 + DB 73,193,235,32 ; shr $0x20,%r11 + DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax + DB 196,194,121,49,193 ; vpmovzxbd %xmm9,%xmm0 + DB 196,194,121,49,203 ; vpmovzxbd %xmm11,%xmm1 + DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 + DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 + DB 196,98,125,24,13,231,59,0,0 ; vbroadcastss 0x3be7(%rip),%ymm9 # 682c <_sk_callback_avx+0x276> + DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 196,194,121,49,202 ; vpmovzxbd %xmm10,%xmm1 + DB 196,194,121,49,212 ; vpmovzxbd %xmm12,%xmm2 + DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 - DB 72,139,64,16 ; mov 0x10(%rax),%rax - DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 - DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 - DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp - DB 65,137,232 ; mov %ebp,%r8d - DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,225,249,126,211 ; vmovq %xmm2,%rbx - DB 65,137,217 ; mov %ebx,%r9d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 - DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,193,249,126,214 ; vmovq %xmm2,%r14 - DB 69,137,247 ; mov %r14d,%r15d - DB 73,193,238,32 ; shr $0x20,%r14 - DB 196,163,121,32,20,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2 - DB 196,227,105,32,20,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2 - DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx - DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2 - DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp - DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2 - DB 196,98,121,49,194 ; vpmovzxbd %xmm2,%xmm8 - DB 196,163,121,32,20,56,0 ; vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm2 - DB 196,163,105,32,20,48,1 ; vpinsrb $0x1,(%rax,%r14,1),%xmm2,%xmm2 - DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp - DB 196,227,105,32,213,2 ; vpinsrb $0x2,%ebp,%xmm2,%xmm2 - DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax - DB 196,227,105,32,208,3 ; vpinsrb $0x3,%eax,%xmm2,%xmm2 + DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8 + DB 196,227,17,32,208,3 ; vpinsrb $0x3,%eax,%xmm13,%xmm2 DB 196,226,121,49,210 ; vpmovzxbd %xmm2,%xmm2 DB 196,227,61,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm8,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 91 ; pop %rbx - DB 65,92 ; pop %r12 - DB 65,93 ; pop %r13 - DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 - DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_table_r_avx _sk_table_r_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -7808,34 +7769,35 @@ _sk_table_r_avx LABEL PROC DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,199 ; vmovq %xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,122,16,4,160 ; vmovss (%r8,%r12,4),%xmm0 - DB 196,3,121,33,4,184,16 ; vinsertps $0x10,(%r8,%r15,4),%xmm0,%xmm8 - DB 196,129,122,16,4,176 ; vmovss (%r8,%r14,4),%xmm0 + DB 196,193,249,126,196 ; vmovq %xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,129,122,16,4,169 ; vmovss (%r9,%r13,4),%xmm0 + DB 196,3,121,33,4,161,16 ; vinsertps $0x10,(%r9,%r12,4),%xmm0,%xmm8 + DB 196,129,122,16,4,185 ; vmovss (%r9,%r15,4),%xmm0 DB 196,99,57,33,192,32 ; vinsertps $0x20,%xmm0,%xmm8,%xmm8 - DB 196,193,122,16,4,152 ; vmovss (%r8,%rbx,4),%xmm0 + DB 196,193,122,16,4,153 ; vmovss (%r9,%rbx,4),%xmm0 DB 196,99,57,33,192,48 ; vinsertps $0x30,%xmm0,%xmm8,%xmm8 - DB 196,129,122,16,4,152 ; vmovss (%r8,%r11,4),%xmm0 - DB 196,3,121,33,12,144,16 ; vinsertps $0x10,(%r8,%r10,4),%xmm0,%xmm9 - DB 196,129,122,16,4,136 ; vmovss (%r8,%r9,4),%xmm0 + DB 196,129,122,16,4,177 ; vmovss (%r9,%r14,4),%xmm0 + DB 196,3,121,33,12,153,16 ; vinsertps $0x10,(%r9,%r11,4),%xmm0,%xmm9 + DB 196,129,122,16,4,145 ; vmovss (%r9,%r10,4),%xmm0 DB 196,99,49,33,200,32 ; vinsertps $0x20,%xmm0,%xmm9,%xmm9 - DB 196,193,122,16,4,128 ; vmovss (%r8,%rax,4),%xmm0 + DB 196,193,122,16,4,129 ; vmovss (%r9,%rax,4),%xmm0 DB 196,227,49,33,192,48 ; vinsertps $0x30,%xmm0,%xmm9,%xmm0 DB 196,195,125,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -7844,10 +7806,11 @@ PUBLIC _sk_table_g_avx _sk_table_g_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -7857,34 +7820,35 @@ _sk_table_g_avx LABEL PROC DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,227,249,22,200,1 ; vpextrq $0x1,%xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,203 ; vmovq %xmm1,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,207 ; vmovq %xmm1,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,122,16,12,160 ; vmovss (%r8,%r12,4),%xmm1 - DB 196,3,113,33,4,184,16 ; vinsertps $0x10,(%r8,%r15,4),%xmm1,%xmm8 - DB 196,129,122,16,12,176 ; vmovss (%r8,%r14,4),%xmm1 + DB 196,193,249,126,204 ; vmovq %xmm1,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,129,122,16,12,169 ; vmovss (%r9,%r13,4),%xmm1 + DB 196,3,113,33,4,161,16 ; vinsertps $0x10,(%r9,%r12,4),%xmm1,%xmm8 + DB 196,129,122,16,12,185 ; vmovss (%r9,%r15,4),%xmm1 DB 196,99,57,33,193,32 ; vinsertps $0x20,%xmm1,%xmm8,%xmm8 - DB 196,193,122,16,12,152 ; vmovss (%r8,%rbx,4),%xmm1 + DB 196,193,122,16,12,153 ; vmovss (%r9,%rbx,4),%xmm1 DB 196,99,57,33,193,48 ; vinsertps $0x30,%xmm1,%xmm8,%xmm8 - DB 196,129,122,16,12,152 ; vmovss (%r8,%r11,4),%xmm1 - DB 196,3,113,33,12,144,16 ; vinsertps $0x10,(%r8,%r10,4),%xmm1,%xmm9 - DB 196,129,122,16,12,136 ; vmovss (%r8,%r9,4),%xmm1 + DB 196,129,122,16,12,177 ; vmovss (%r9,%r14,4),%xmm1 + DB 196,3,113,33,12,153,16 ; vinsertps $0x10,(%r9,%r11,4),%xmm1,%xmm9 + DB 196,129,122,16,12,145 ; vmovss (%r9,%r10,4),%xmm1 DB 196,99,49,33,201,32 ; vinsertps $0x20,%xmm1,%xmm9,%xmm9 - DB 196,193,122,16,12,128 ; vmovss (%r8,%rax,4),%xmm1 + DB 196,193,122,16,12,129 ; vmovss (%r9,%rax,4),%xmm1 DB 196,227,49,33,201,48 ; vinsertps $0x30,%xmm1,%xmm9,%xmm1 DB 196,195,117,24,200,1 ; vinsertf128 $0x1,%xmm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -7893,10 +7857,11 @@ PUBLIC _sk_table_b_avx _sk_table_b_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -7906,34 +7871,35 @@ _sk_table_b_avx LABEL PROC DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,215 ; vmovq %xmm2,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,122,16,20,160 ; vmovss (%r8,%r12,4),%xmm2 - DB 196,3,105,33,4,184,16 ; vinsertps $0x10,(%r8,%r15,4),%xmm2,%xmm8 - DB 196,129,122,16,20,176 ; vmovss (%r8,%r14,4),%xmm2 + DB 196,193,249,126,212 ; vmovq %xmm2,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,129,122,16,20,169 ; vmovss (%r9,%r13,4),%xmm2 + DB 196,3,105,33,4,161,16 ; vinsertps $0x10,(%r9,%r12,4),%xmm2,%xmm8 + DB 196,129,122,16,20,185 ; vmovss (%r9,%r15,4),%xmm2 DB 196,99,57,33,194,32 ; vinsertps $0x20,%xmm2,%xmm8,%xmm8 - DB 196,193,122,16,20,152 ; vmovss (%r8,%rbx,4),%xmm2 + DB 196,193,122,16,20,153 ; vmovss (%r9,%rbx,4),%xmm2 DB 196,99,57,33,194,48 ; vinsertps $0x30,%xmm2,%xmm8,%xmm8 - DB 196,129,122,16,20,152 ; vmovss (%r8,%r11,4),%xmm2 - DB 196,3,105,33,12,144,16 ; vinsertps $0x10,(%r8,%r10,4),%xmm2,%xmm9 - DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2 + DB 196,129,122,16,20,177 ; vmovss (%r9,%r14,4),%xmm2 + DB 196,3,105,33,12,153,16 ; vinsertps $0x10,(%r9,%r11,4),%xmm2,%xmm9 + DB 196,129,122,16,20,145 ; vmovss (%r9,%r10,4),%xmm2 DB 196,99,49,33,202,32 ; vinsertps $0x20,%xmm2,%xmm9,%xmm9 - DB 196,193,122,16,20,128 ; vmovss (%r8,%rax,4),%xmm2 + DB 196,193,122,16,20,129 ; vmovss (%r9,%rax,4),%xmm2 DB 196,227,49,33,210,48 ; vinsertps $0x30,%xmm2,%xmm9,%xmm2 DB 196,195,109,24,208,1 ; vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -7942,10 +7908,11 @@ PUBLIC _sk_table_a_avx _sk_table_a_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 197,121,110,192 ; vmovd %eax,%xmm8 @@ -7955,34 +7922,35 @@ _sk_table_a_avx LABEL PROC DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,227,249,22,216,1 ; vpextrq $0x1,%xmm3,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,218 ; vmovq %xmm3,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,219 ; vmovq %xmm3,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,219,1 ; vextractf128 $0x1,%ymm3,%xmm3 DB 196,227,249,22,219,1 ; vpextrq $0x1,%xmm3,%rbx - DB 65,137,222 ; mov %ebx,%r14d + DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,249,126,223 ; vmovq %xmm3,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,122,16,28,160 ; vmovss (%r8,%r12,4),%xmm3 - DB 196,3,97,33,4,184,16 ; vinsertps $0x10,(%r8,%r15,4),%xmm3,%xmm8 - DB 196,129,122,16,28,176 ; vmovss (%r8,%r14,4),%xmm3 + DB 196,193,249,126,220 ; vmovq %xmm3,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,129,122,16,28,169 ; vmovss (%r9,%r13,4),%xmm3 + DB 196,3,97,33,4,161,16 ; vinsertps $0x10,(%r9,%r12,4),%xmm3,%xmm8 + DB 196,129,122,16,28,185 ; vmovss (%r9,%r15,4),%xmm3 DB 196,99,57,33,195,32 ; vinsertps $0x20,%xmm3,%xmm8,%xmm8 - DB 196,193,122,16,28,152 ; vmovss (%r8,%rbx,4),%xmm3 + DB 196,193,122,16,28,153 ; vmovss (%r9,%rbx,4),%xmm3 DB 196,99,57,33,195,48 ; vinsertps $0x30,%xmm3,%xmm8,%xmm8 - DB 196,129,122,16,28,152 ; vmovss (%r8,%r11,4),%xmm3 - DB 196,3,97,33,12,144,16 ; vinsertps $0x10,(%r8,%r10,4),%xmm3,%xmm9 - DB 196,129,122,16,28,136 ; vmovss (%r8,%r9,4),%xmm3 + DB 196,129,122,16,28,177 ; vmovss (%r9,%r14,4),%xmm3 + DB 196,3,97,33,12,153,16 ; vinsertps $0x10,(%r9,%r11,4),%xmm3,%xmm9 + DB 196,129,122,16,28,145 ; vmovss (%r9,%r10,4),%xmm3 DB 196,99,49,33,203,32 ; vinsertps $0x20,%xmm3,%xmm9,%xmm9 - DB 196,193,122,16,28,128 ; vmovss (%r8,%rax,4),%xmm3 + DB 196,193,122,16,28,129 ; vmovss (%r9,%rax,4),%xmm3 DB 196,227,49,33,219,48 ; vinsertps $0x30,%xmm3,%xmm9,%xmm3 DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -8002,36 +7970,36 @@ _sk_parametric_r_avx LABEL PROC DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,216 ; vcvtdq2ps %ymm0,%ymm11 - DB 196,98,125,24,37,181,55,0,0 ; vbroadcastss 0x37b5(%rip),%ymm12 # 67d0 <_sk_callback_avx+0x270> + DB 196,98,125,24,37,76,56,0,0 ; vbroadcastss 0x384c(%rip),%ymm12 # 6830 <_sk_callback_avx+0x27a> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,171,55,0,0 ; vbroadcastss 0x37ab(%rip),%ymm12 # 67d4 <_sk_callback_avx+0x274> + DB 196,98,125,24,37,66,56,0,0 ; vbroadcastss 0x3842(%rip),%ymm12 # 6834 <_sk_callback_avx+0x27e> DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,161,55,0,0 ; vbroadcastss 0x37a1(%rip),%ymm12 # 67d8 <_sk_callback_avx+0x278> + DB 196,98,125,24,37,56,56,0,0 ; vbroadcastss 0x3838(%rip),%ymm12 # 6838 <_sk_callback_avx+0x282> DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,151,55,0,0 ; vbroadcastss 0x3797(%rip),%ymm12 # 67dc <_sk_callback_avx+0x27c> + DB 196,98,125,24,37,46,56,0,0 ; vbroadcastss 0x382e(%rip),%ymm12 # 683c <_sk_callback_avx+0x286> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,141,55,0,0 ; vbroadcastss 0x378d(%rip),%ymm12 # 67e0 <_sk_callback_avx+0x280> + DB 196,98,125,24,37,36,56,0,0 ; vbroadcastss 0x3824(%rip),%ymm12 # 6840 <_sk_callback_avx+0x28a> DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,126,55,0,0 ; vbroadcastss 0x377e(%rip),%ymm12 # 67e4 <_sk_callback_avx+0x284> + DB 196,98,125,24,37,21,56,0,0 ; vbroadcastss 0x3815(%rip),%ymm12 # 6844 <_sk_callback_avx+0x28e> DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,116,55,0,0 ; vbroadcastss 0x3774(%rip),%ymm12 # 67e8 <_sk_callback_avx+0x288> + DB 196,98,125,24,37,11,56,0,0 ; vbroadcastss 0x380b(%rip),%ymm12 # 6848 <_sk_callback_avx+0x292> DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,88,55,0,0 ; vbroadcastss 0x3758(%rip),%ymm11 # 67ec <_sk_callback_avx+0x28c> + DB 196,98,125,24,29,239,55,0,0 ; vbroadcastss 0x37ef(%rip),%ymm11 # 684c <_sk_callback_avx+0x296> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,78,55,0,0 ; vbroadcastss 0x374e(%rip),%ymm11 # 67f0 <_sk_callback_avx+0x290> + DB 196,98,125,24,29,229,55,0,0 ; vbroadcastss 0x37e5(%rip),%ymm11 # 6850 <_sk_callback_avx+0x29a> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,63,55,0,0 ; vbroadcastss 0x373f(%rip),%ymm11 # 67f4 <_sk_callback_avx+0x294> + DB 196,98,125,24,29,214,55,0,0 ; vbroadcastss 0x37d6(%rip),%ymm11 # 6854 <_sk_callback_avx+0x29e> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,53,55,0,0 ; vbroadcastss 0x3735(%rip),%ymm11 # 67f8 <_sk_callback_avx+0x298> + DB 196,98,125,24,29,204,55,0,0 ; vbroadcastss 0x37cc(%rip),%ymm11 # 6858 <_sk_callback_avx+0x2a2> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,38,55,0,0 ; vbroadcastss 0x3726(%rip),%ymm10 # 67fc <_sk_callback_avx+0x29c> + DB 196,98,125,24,21,189,55,0,0 ; vbroadcastss 0x37bd(%rip),%ymm10 # 685c <_sk_callback_avx+0x2a6> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8039,7 +8007,7 @@ _sk_parametric_r_avx LABEL PROC DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,253,54,0,0 ; vbroadcastss 0x36fd(%rip),%ymm8 # 6800 <_sk_callback_avx+0x2a0> + DB 196,98,125,24,5,148,55,0,0 ; vbroadcastss 0x3794(%rip),%ymm8 # 6860 <_sk_callback_avx+0x2aa> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8059,36 +8027,36 @@ _sk_parametric_g_avx LABEL PROC DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,217 ; vcvtdq2ps %ymm1,%ymm11 - DB 196,98,125,24,37,174,54,0,0 ; vbroadcastss 0x36ae(%rip),%ymm12 # 6804 <_sk_callback_avx+0x2a4> + DB 196,98,125,24,37,69,55,0,0 ; vbroadcastss 0x3745(%rip),%ymm12 # 6864 <_sk_callback_avx+0x2ae> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,164,54,0,0 ; vbroadcastss 0x36a4(%rip),%ymm12 # 6808 <_sk_callback_avx+0x2a8> + DB 196,98,125,24,37,59,55,0,0 ; vbroadcastss 0x373b(%rip),%ymm12 # 6868 <_sk_callback_avx+0x2b2> DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,154,54,0,0 ; vbroadcastss 0x369a(%rip),%ymm12 # 680c <_sk_callback_avx+0x2ac> + DB 196,98,125,24,37,49,55,0,0 ; vbroadcastss 0x3731(%rip),%ymm12 # 686c <_sk_callback_avx+0x2b6> DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,144,54,0,0 ; vbroadcastss 0x3690(%rip),%ymm12 # 6810 <_sk_callback_avx+0x2b0> + DB 196,98,125,24,37,39,55,0,0 ; vbroadcastss 0x3727(%rip),%ymm12 # 6870 <_sk_callback_avx+0x2ba> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,134,54,0,0 ; vbroadcastss 0x3686(%rip),%ymm12 # 6814 <_sk_callback_avx+0x2b4> + DB 196,98,125,24,37,29,55,0,0 ; vbroadcastss 0x371d(%rip),%ymm12 # 6874 <_sk_callback_avx+0x2be> DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,119,54,0,0 ; vbroadcastss 0x3677(%rip),%ymm12 # 6818 <_sk_callback_avx+0x2b8> + DB 196,98,125,24,37,14,55,0,0 ; vbroadcastss 0x370e(%rip),%ymm12 # 6878 <_sk_callback_avx+0x2c2> DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,109,54,0,0 ; vbroadcastss 0x366d(%rip),%ymm12 # 681c <_sk_callback_avx+0x2bc> + DB 196,98,125,24,37,4,55,0,0 ; vbroadcastss 0x3704(%rip),%ymm12 # 687c <_sk_callback_avx+0x2c6> DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1 DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,81,54,0,0 ; vbroadcastss 0x3651(%rip),%ymm11 # 6820 <_sk_callback_avx+0x2c0> + DB 196,98,125,24,29,232,54,0,0 ; vbroadcastss 0x36e8(%rip),%ymm11 # 6880 <_sk_callback_avx+0x2ca> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,71,54,0,0 ; vbroadcastss 0x3647(%rip),%ymm11 # 6824 <_sk_callback_avx+0x2c4> + DB 196,98,125,24,29,222,54,0,0 ; vbroadcastss 0x36de(%rip),%ymm11 # 6884 <_sk_callback_avx+0x2ce> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,56,54,0,0 ; vbroadcastss 0x3638(%rip),%ymm11 # 6828 <_sk_callback_avx+0x2c8> + DB 196,98,125,24,29,207,54,0,0 ; vbroadcastss 0x36cf(%rip),%ymm11 # 6888 <_sk_callback_avx+0x2d2> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,46,54,0,0 ; vbroadcastss 0x362e(%rip),%ymm11 # 682c <_sk_callback_avx+0x2cc> + DB 196,98,125,24,29,197,54,0,0 ; vbroadcastss 0x36c5(%rip),%ymm11 # 688c <_sk_callback_avx+0x2d6> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,31,54,0,0 ; vbroadcastss 0x361f(%rip),%ymm10 # 6830 <_sk_callback_avx+0x2d0> + DB 196,98,125,24,21,182,54,0,0 ; vbroadcastss 0x36b6(%rip),%ymm10 # 6890 <_sk_callback_avx+0x2da> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8096,7 +8064,7 @@ _sk_parametric_g_avx LABEL PROC DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,246,53,0,0 ; vbroadcastss 0x35f6(%rip),%ymm8 # 6834 <_sk_callback_avx+0x2d4> + DB 196,98,125,24,5,141,54,0,0 ; vbroadcastss 0x368d(%rip),%ymm8 # 6894 <_sk_callback_avx+0x2de> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8116,36 +8084,36 @@ _sk_parametric_b_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,218 ; vcvtdq2ps %ymm2,%ymm11 - DB 196,98,125,24,37,167,53,0,0 ; vbroadcastss 0x35a7(%rip),%ymm12 # 6838 <_sk_callback_avx+0x2d8> + DB 196,98,125,24,37,62,54,0,0 ; vbroadcastss 0x363e(%rip),%ymm12 # 6898 <_sk_callback_avx+0x2e2> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,157,53,0,0 ; vbroadcastss 0x359d(%rip),%ymm12 # 683c <_sk_callback_avx+0x2dc> + DB 196,98,125,24,37,52,54,0,0 ; vbroadcastss 0x3634(%rip),%ymm12 # 689c <_sk_callback_avx+0x2e6> DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,147,53,0,0 ; vbroadcastss 0x3593(%rip),%ymm12 # 6840 <_sk_callback_avx+0x2e0> + DB 196,98,125,24,37,42,54,0,0 ; vbroadcastss 0x362a(%rip),%ymm12 # 68a0 <_sk_callback_avx+0x2ea> DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,137,53,0,0 ; vbroadcastss 0x3589(%rip),%ymm12 # 6844 <_sk_callback_avx+0x2e4> + DB 196,98,125,24,37,32,54,0,0 ; vbroadcastss 0x3620(%rip),%ymm12 # 68a4 <_sk_callback_avx+0x2ee> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,127,53,0,0 ; vbroadcastss 0x357f(%rip),%ymm12 # 6848 <_sk_callback_avx+0x2e8> + DB 196,98,125,24,37,22,54,0,0 ; vbroadcastss 0x3616(%rip),%ymm12 # 68a8 <_sk_callback_avx+0x2f2> DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,112,53,0,0 ; vbroadcastss 0x3570(%rip),%ymm12 # 684c <_sk_callback_avx+0x2ec> + DB 196,98,125,24,37,7,54,0,0 ; vbroadcastss 0x3607(%rip),%ymm12 # 68ac <_sk_callback_avx+0x2f6> DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,102,53,0,0 ; vbroadcastss 0x3566(%rip),%ymm12 # 6850 <_sk_callback_avx+0x2f0> + DB 196,98,125,24,37,253,53,0,0 ; vbroadcastss 0x35fd(%rip),%ymm12 # 68b0 <_sk_callback_avx+0x2fa> DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2 DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,74,53,0,0 ; vbroadcastss 0x354a(%rip),%ymm11 # 6854 <_sk_callback_avx+0x2f4> + DB 196,98,125,24,29,225,53,0,0 ; vbroadcastss 0x35e1(%rip),%ymm11 # 68b4 <_sk_callback_avx+0x2fe> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,64,53,0,0 ; vbroadcastss 0x3540(%rip),%ymm11 # 6858 <_sk_callback_avx+0x2f8> + DB 196,98,125,24,29,215,53,0,0 ; vbroadcastss 0x35d7(%rip),%ymm11 # 68b8 <_sk_callback_avx+0x302> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,49,53,0,0 ; vbroadcastss 0x3531(%rip),%ymm11 # 685c <_sk_callback_avx+0x2fc> + DB 196,98,125,24,29,200,53,0,0 ; vbroadcastss 0x35c8(%rip),%ymm11 # 68bc <_sk_callback_avx+0x306> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,39,53,0,0 ; vbroadcastss 0x3527(%rip),%ymm11 # 6860 <_sk_callback_avx+0x300> + DB 196,98,125,24,29,190,53,0,0 ; vbroadcastss 0x35be(%rip),%ymm11 # 68c0 <_sk_callback_avx+0x30a> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,24,53,0,0 ; vbroadcastss 0x3518(%rip),%ymm10 # 6864 <_sk_callback_avx+0x304> + DB 196,98,125,24,21,175,53,0,0 ; vbroadcastss 0x35af(%rip),%ymm10 # 68c4 <_sk_callback_avx+0x30e> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8153,7 +8121,7 @@ _sk_parametric_b_avx LABEL PROC DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,239,52,0,0 ; vbroadcastss 0x34ef(%rip),%ymm8 # 6868 <_sk_callback_avx+0x308> + DB 196,98,125,24,5,134,53,0,0 ; vbroadcastss 0x3586(%rip),%ymm8 # 68c8 <_sk_callback_avx+0x312> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8173,36 +8141,36 @@ _sk_parametric_a_avx LABEL PROC DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,219 ; vcvtdq2ps %ymm3,%ymm11 - DB 196,98,125,24,37,160,52,0,0 ; vbroadcastss 0x34a0(%rip),%ymm12 # 686c <_sk_callback_avx+0x30c> + DB 196,98,125,24,37,55,53,0,0 ; vbroadcastss 0x3537(%rip),%ymm12 # 68cc <_sk_callback_avx+0x316> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,150,52,0,0 ; vbroadcastss 0x3496(%rip),%ymm12 # 6870 <_sk_callback_avx+0x310> + DB 196,98,125,24,37,45,53,0,0 ; vbroadcastss 0x352d(%rip),%ymm12 # 68d0 <_sk_callback_avx+0x31a> DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,140,52,0,0 ; vbroadcastss 0x348c(%rip),%ymm12 # 6874 <_sk_callback_avx+0x314> + DB 196,98,125,24,37,35,53,0,0 ; vbroadcastss 0x3523(%rip),%ymm12 # 68d4 <_sk_callback_avx+0x31e> DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,130,52,0,0 ; vbroadcastss 0x3482(%rip),%ymm12 # 6878 <_sk_callback_avx+0x318> + DB 196,98,125,24,37,25,53,0,0 ; vbroadcastss 0x3519(%rip),%ymm12 # 68d8 <_sk_callback_avx+0x322> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,120,52,0,0 ; vbroadcastss 0x3478(%rip),%ymm12 # 687c <_sk_callback_avx+0x31c> + DB 196,98,125,24,37,15,53,0,0 ; vbroadcastss 0x350f(%rip),%ymm12 # 68dc <_sk_callback_avx+0x326> DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,105,52,0,0 ; vbroadcastss 0x3469(%rip),%ymm12 # 6880 <_sk_callback_avx+0x320> + DB 196,98,125,24,37,0,53,0,0 ; vbroadcastss 0x3500(%rip),%ymm12 # 68e0 <_sk_callback_avx+0x32a> DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,95,52,0,0 ; vbroadcastss 0x345f(%rip),%ymm12 # 6884 <_sk_callback_avx+0x324> + DB 196,98,125,24,37,246,52,0,0 ; vbroadcastss 0x34f6(%rip),%ymm12 # 68e4 <_sk_callback_avx+0x32e> DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3 DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,67,52,0,0 ; vbroadcastss 0x3443(%rip),%ymm11 # 6888 <_sk_callback_avx+0x328> + DB 196,98,125,24,29,218,52,0,0 ; vbroadcastss 0x34da(%rip),%ymm11 # 68e8 <_sk_callback_avx+0x332> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,57,52,0,0 ; vbroadcastss 0x3439(%rip),%ymm11 # 688c <_sk_callback_avx+0x32c> + DB 196,98,125,24,29,208,52,0,0 ; vbroadcastss 0x34d0(%rip),%ymm11 # 68ec <_sk_callback_avx+0x336> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,42,52,0,0 ; vbroadcastss 0x342a(%rip),%ymm11 # 6890 <_sk_callback_avx+0x330> + DB 196,98,125,24,29,193,52,0,0 ; vbroadcastss 0x34c1(%rip),%ymm11 # 68f0 <_sk_callback_avx+0x33a> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,32,52,0,0 ; vbroadcastss 0x3420(%rip),%ymm11 # 6894 <_sk_callback_avx+0x334> + DB 196,98,125,24,29,183,52,0,0 ; vbroadcastss 0x34b7(%rip),%ymm11 # 68f4 <_sk_callback_avx+0x33e> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,17,52,0,0 ; vbroadcastss 0x3411(%rip),%ymm10 # 6898 <_sk_callback_avx+0x338> + DB 196,98,125,24,21,168,52,0,0 ; vbroadcastss 0x34a8(%rip),%ymm10 # 68f8 <_sk_callback_avx+0x342> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8210,38 +8178,38 @@ _sk_parametric_a_avx LABEL PROC DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,232,51,0,0 ; vbroadcastss 0x33e8(%rip),%ymm8 # 689c <_sk_callback_avx+0x33c> + DB 196,98,125,24,5,127,52,0,0 ; vbroadcastss 0x347f(%rip),%ymm8 # 68fc <_sk_callback_avx+0x346> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_lab_to_xyz_avx _sk_lab_to_xyz_avx LABEL PROC - DB 196,98,125,24,5,218,51,0,0 ; vbroadcastss 0x33da(%rip),%ymm8 # 68a0 <_sk_callback_avx+0x340> + DB 196,98,125,24,5,113,52,0,0 ; vbroadcastss 0x3471(%rip),%ymm8 # 6900 <_sk_callback_avx+0x34a> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,208,51,0,0 ; vbroadcastss 0x33d0(%rip),%ymm8 # 68a4 <_sk_callback_avx+0x344> + DB 196,98,125,24,5,103,52,0,0 ; vbroadcastss 0x3467(%rip),%ymm8 # 6904 <_sk_callback_avx+0x34e> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,13,198,51,0,0 ; vbroadcastss 0x33c6(%rip),%ymm9 # 68a8 <_sk_callback_avx+0x348> + DB 196,98,125,24,13,93,52,0,0 ; vbroadcastss 0x345d(%rip),%ymm9 # 6908 <_sk_callback_avx+0x352> DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 196,193,108,88,209 ; vaddps %ymm9,%ymm2,%ymm2 - DB 196,98,125,24,5,178,51,0,0 ; vbroadcastss 0x33b2(%rip),%ymm8 # 68ac <_sk_callback_avx+0x34c> + DB 196,98,125,24,5,73,52,0,0 ; vbroadcastss 0x3449(%rip),%ymm8 # 690c <_sk_callback_avx+0x356> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,168,51,0,0 ; vbroadcastss 0x33a8(%rip),%ymm8 # 68b0 <_sk_callback_avx+0x350> + DB 196,98,125,24,5,63,52,0,0 ; vbroadcastss 0x343f(%rip),%ymm8 # 6910 <_sk_callback_avx+0x35a> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,158,51,0,0 ; vbroadcastss 0x339e(%rip),%ymm8 # 68b4 <_sk_callback_avx+0x354> + DB 196,98,125,24,5,53,52,0,0 ; vbroadcastss 0x3435(%rip),%ymm8 # 6914 <_sk_callback_avx+0x35e> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1 - DB 196,98,125,24,5,144,51,0,0 ; vbroadcastss 0x3390(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x358> + DB 196,98,125,24,5,39,52,0,0 ; vbroadcastss 0x3427(%rip),%ymm8 # 6918 <_sk_callback_avx+0x362> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,252,92,210 ; vsubps %ymm2,%ymm0,%ymm2 DB 197,116,89,193 ; vmulps %ymm1,%ymm1,%ymm8 DB 196,65,116,89,192 ; vmulps %ymm8,%ymm1,%ymm8 - DB 196,98,125,24,13,121,51,0,0 ; vbroadcastss 0x3379(%rip),%ymm9 # 68bc <_sk_callback_avx+0x35c> + DB 196,98,125,24,13,16,52,0,0 ; vbroadcastss 0x3410(%rip),%ymm9 # 691c <_sk_callback_avx+0x366> DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,110,51,0,0 ; vbroadcastss 0x336e(%rip),%ymm11 # 68c0 <_sk_callback_avx+0x360> + DB 196,98,125,24,29,5,52,0,0 ; vbroadcastss 0x3405(%rip),%ymm11 # 6920 <_sk_callback_avx+0x36a> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,37,100,51,0,0 ; vbroadcastss 0x3364(%rip),%ymm12 # 68c4 <_sk_callback_avx+0x364> + DB 196,98,125,24,37,251,51,0,0 ; vbroadcastss 0x33fb(%rip),%ymm12 # 6924 <_sk_callback_avx+0x36e> DB 196,193,116,89,204 ; vmulps %ymm12,%ymm1,%ymm1 DB 196,67,117,74,192,160 ; vblendvps %ymm10,%ymm8,%ymm1,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -8256,56 +8224,59 @@ _sk_lab_to_xyz_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,193,108,89,212 ; vmulps %ymm12,%ymm2,%ymm2 DB 196,227,109,74,208,144 ; vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - DB 196,226,125,24,5,26,51,0,0 ; vbroadcastss 0x331a(%rip),%ymm0 # 68c8 <_sk_callback_avx+0x368> + DB 196,226,125,24,5,177,51,0,0 ; vbroadcastss 0x33b1(%rip),%ymm0 # 6928 <_sk_callback_avx+0x372> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,17,51,0,0 ; vbroadcastss 0x3311(%rip),%ymm8 # 68cc <_sk_callback_avx+0x36c> + DB 196,98,125,24,5,168,51,0,0 ; vbroadcastss 0x33a8(%rip),%ymm8 # 692c <_sk_callback_avx+0x376> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_load_a8_avx _sk_load_a8_avx LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,62 ; jne 3612 <_sk_load_a8_avx+0x4e> + DB 117,62 ; jne 35db <_sk_load_a8_avx+0x4e> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1 DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,213,50,0,0 ; vbroadcastss 0x32d5(%rip),%ymm1 # 68d0 <_sk_callback_avx+0x370> + DB 196,226,125,24,13,108,51,0,0 ; vbroadcastss 0x336c(%rip),%ymm1 # 6930 <_sk_callback_avx+0x37a> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 - DB 76,137,193 ; mov %r8,%rcx + DB 76,137,201 ; mov %r9,%rcx DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 361a <_sk_load_a8_avx+0x56> - DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,161 ; jmp 35d8 <_sk_load_a8_avx+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 35e4 <_sk_load_a8_avx+0x57> + DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 91 ; pop %rbx + DB 235,160 ; jmp 35a1 <_sk_load_a8_avx+0x14> PUBLIC _sk_gather_a8_avx _sk_gather_a8_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 @@ -8316,35 +8287,35 @@ _sk_gather_a8_avx LABEL PROC DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,131,121,32,4,48,0 ; vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0 - DB 196,195,121,32,4,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0 - DB 67,15,182,28,32 ; movzbl (%r8,%r12,1),%ebx + DB 196,131,121,32,4,57,0 ; vpinsrb $0x0,(%r9,%r15,1),%xmm0,%xmm0 + DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 + DB 67,15,182,28,41 ; movzbl (%r9,%r13,1),%ebx DB 196,227,121,32,195,2 ; vpinsrb $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,182,28,56 ; movzbl (%r8,%r15,1),%ebx + DB 67,15,182,28,33 ; movzbl (%r9,%r12,1),%ebx DB 196,227,121,32,195,3 ; vpinsrb $0x3,%ebx,%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,131,121,32,12,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1 - DB 196,131,113,32,12,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx + DB 196,131,121,32,12,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm1 + DB 196,131,113,32,12,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm1,%xmm1 + DB 67,15,182,28,17 ; movzbl (%r9,%r10,1),%ebx DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,227,113,32,200,3 ; vpinsrb $0x3,%eax,%xmm1,%xmm1 DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,202,49,0,0 ; vbroadcastss 0x31ca(%rip),%ymm1 # 68d4 <_sk_callback_avx+0x374> + DB 196,226,125,24,13,94,50,0,0 ; vbroadcastss 0x325e(%rip),%ymm1 # 6934 <_sk_callback_avx+0x37e> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -8352,6 +8323,7 @@ _sk_gather_a8_avx LABEL PROC DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -8359,105 +8331,106 @@ _sk_gather_a8_avx LABEL PROC PUBLIC _sk_store_a8_avx _sk_store_a8_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,165,49,0,0 ; vbroadcastss 0x31a5(%rip),%ymm8 # 68d8 <_sk_callback_avx+0x378> + DB 76,139,24 ; mov (%rax),%r11 + DB 196,98,125,24,5,55,50,0,0 ; vbroadcastss 0x3237(%rip),%ymm8 # 6938 <_sk_callback_avx+0x382> DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8 DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 375c <_sk_store_a8_avx+0x37> - DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1) + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 372a <_sk_store_a8_avx+0x37> + DB 196,65,123,17,4,19 ; vmovsd %xmm8,(%r11,%rdx,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 3758 <_sk_store_a8_avx+0x33> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 3726 <_sk_store_a8_avx+0x33> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 37c0 <_sk_store_a8_avx+0x9b> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,121,20,68,58,6,12 ; vpextrb $0xc,%xmm8,0x6(%r10,%rdi,1) - DB 196,67,121,20,68,58,5,10 ; vpextrb $0xa,%xmm8,0x5(%r10,%rdi,1) - DB 196,67,121,20,68,58,4,8 ; vpextrb $0x8,%xmm8,0x4(%r10,%rdi,1) - DB 196,67,121,20,68,58,3,6 ; vpextrb $0x6,%xmm8,0x3(%r10,%rdi,1) - DB 196,67,121,20,68,58,2,4 ; vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1) - DB 196,67,121,20,68,58,1,2 ; vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1) - DB 196,67,121,20,4,58,0 ; vpextrb $0x0,%xmm8,(%r10,%rdi,1) - DB 235,154 ; jmp 3758 <_sk_store_a8_avx+0x33> - DB 102,144 ; xchg %ax,%ax - DB 245 ; cmc - DB 255 ; (bad) + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 378c <_sk_store_a8_avx+0x99> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,67,121,20,68,19,6,12 ; vpextrb $0xc,%xmm8,0x6(%r11,%rdx,1) + DB 196,67,121,20,68,19,5,10 ; vpextrb $0xa,%xmm8,0x5(%r11,%rdx,1) + DB 196,67,121,20,68,19,4,8 ; vpextrb $0x8,%xmm8,0x4(%r11,%rdx,1) + DB 196,67,121,20,68,19,3,6 ; vpextrb $0x6,%xmm8,0x3(%r11,%rdx,1) + DB 196,67,121,20,68,19,2,4 ; vpextrb $0x4,%xmm8,0x2(%r11,%rdx,1) + DB 196,67,121,20,68,19,1,2 ; vpextrb $0x2,%xmm8,0x1(%r11,%rdx,1) + DB 196,67,121,20,4,19,0 ; vpextrb $0x0,%xmm8,(%r11,%rdx,1) + DB 235,154 ; jmp 3726 <_sk_store_a8_avx+0x33> + DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) - DB 237 ; in (%dx),%eax + DB 239 ; out %eax,(%dx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,229 ; jmpq *%rbp + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 221,255 ; (bad) + DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,213 ; callq *%rbp + DB 255,215 ; callq *%rdi DB 255 ; (bad) DB 255 ; (bad) - DB 255,205 ; dec %ebp + DB 255,207 ; dec %edi DB 255 ; (bad) DB 255 ; (bad) - DB 255,197 ; inc %ebp + DB 255,199 ; inc %edi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_g8_avx _sk_load_g8_avx LABEL PROC - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,1,248 ; add %rdi,%rax + DB 72,1,208 ; add %rdx,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,67 ; jne 382f <_sk_load_g8_avx+0x53> + DB 117,67 ; jne 37fb <_sk_load_g8_avx+0x53> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1 DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,201,48,0,0 ; vbroadcastss 0x30c9(%rip),%ymm1 # 68dc <_sk_callback_avx+0x37c> + DB 196,226,125,24,13,93,49,0,0 ; vbroadcastss 0x315d(%rip),%ymm1 # 693c <_sk_callback_avx+0x386> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,190,48,0,0 ; vbroadcastss 0x30be(%rip),%ymm3 # 68e0 <_sk_callback_avx+0x380> - DB 76,137,193 ; mov %r8,%rcx + DB 196,226,125,24,29,82,49,0,0 ; vbroadcastss 0x3152(%rip),%ymm3 # 6940 <_sk_callback_avx+0x38a> + DB 76,137,201 ; mov %r9,%rcx DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax + DB 83 ; push %rbx DB 49,201 ; xor %ecx,%ecx - DB 77,137,194 ; mov %r8,%r10 - DB 69,49,201 ; xor %r9d,%r9d - DB 68,15,182,24 ; movzbl (%rax),%r11d + DB 77,137,195 ; mov %r8,%r11 + DB 69,49,210 ; xor %r10d,%r10d + DB 15,182,24 ; movzbl (%rax),%ebx DB 72,255,192 ; inc %rax - DB 73,211,227 ; shl %cl,%r11 - DB 77,9,217 ; or %r11,%r9 + DB 72,211,227 ; shl %cl,%rbx + DB 73,9,218 ; or %rbx,%r10 DB 72,131,193,8 ; add $0x8,%rcx - DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 3837 <_sk_load_g8_avx+0x5b> - DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,156 ; jmp 37f0 <_sk_load_g8_avx+0x14> + DB 73,255,203 ; dec %r11 + DB 117,235 ; jne 3804 <_sk_load_g8_avx+0x5c> + DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 91 ; pop %rbx + DB 235,155 ; jmp 37bc <_sk_load_g8_avx+0x14> PUBLIC _sk_gather_g8_avx _sk_gather_g8_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 @@ -8468,42 +8441,43 @@ _sk_gather_g8_avx LABEL PROC DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,131,121,32,4,48,0 ; vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0 - DB 196,195,121,32,4,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0 - DB 67,15,182,28,32 ; movzbl (%r8,%r12,1),%ebx + DB 196,131,121,32,4,57,0 ; vpinsrb $0x0,(%r9,%r15,1),%xmm0,%xmm0 + DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 + DB 67,15,182,28,41 ; movzbl (%r9,%r13,1),%ebx DB 196,227,121,32,195,2 ; vpinsrb $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,182,28,56 ; movzbl (%r8,%r15,1),%ebx + DB 67,15,182,28,33 ; movzbl (%r9,%r12,1),%ebx DB 196,227,121,32,195,3 ; vpinsrb $0x3,%ebx,%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,131,121,32,12,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1 - DB 196,131,113,32,12,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx + DB 196,131,121,32,12,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm1 + DB 196,131,113,32,12,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm1,%xmm1 + DB 67,15,182,28,17 ; movzbl (%r9,%r10,1),%ebx DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,227,113,32,200,3 ; vpinsrb $0x3,%eax,%xmm1,%xmm1 DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,189,47,0,0 ; vbroadcastss 0x2fbd(%rip),%ymm1 # 68e4 <_sk_callback_avx+0x384> + DB 196,226,125,24,13,78,48,0,0 ; vbroadcastss 0x304e(%rip),%ymm1 # 6944 <_sk_callback_avx+0x38e> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,178,47,0,0 ; vbroadcastss 0x2fb2(%rip),%ymm3 # 68e8 <_sk_callback_avx+0x388> + DB 196,226,125,24,29,67,48,0,0 ; vbroadcastss 0x3043(%rip),%ymm3 # 6948 <_sk_callback_avx+0x392> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -8511,18 +8485,19 @@ _sk_gather_g8_avx LABEL PROC PUBLIC _sk_gather_i8_avx _sk_gather_i8_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,192 ; mov %rax,%r8 - DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 3956 <_sk_gather_i8_avx+0xf> - DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 3958 <_sk_gather_i8_avx+0x11> + DB 73,137,193 ; mov %rax,%r9 + DB 77,133,201 ; test %r9,%r9 + DB 116,5 ; je 3927 <_sk_gather_i8_avx+0xf> + DB 76,137,200 ; mov %r9,%rax + DB 235,2 ; jmp 3929 <_sk_gather_i8_avx+0x11> DB 72,173 ; lods %ds:(%rsi),%rax + DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 76,139,8 ; mov (%rax),%r9 + DB 76,139,16 ; mov (%rax),%r10 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 @@ -8533,54 +8508,54 @@ _sk_gather_i8_avx LABEL PROC DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,194 ; mov %eax,%r10d + DB 65,137,195 ; mov %eax,%r11d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,211 ; vmovq %xmm2,%r11 - DB 69,137,222 ; mov %r11d,%r14d - DB 73,193,235,32 ; shr $0x20,%r11 - DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,223 ; mov %ebx,%r15d - DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 - DB 69,137,229 ; mov %r12d,%r13d - DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,131,121,32,4,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 - DB 196,131,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,17,2 ; vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0 - DB 196,195,121,32,4,1,3 ; vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0 + DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 + DB 196,193,249,126,199 ; vmovq %xmm0,%r15 + DB 69,137,252 ; mov %r15d,%r12d + DB 196,195,249,22,197,1 ; vpextrq $0x1,%xmm0,%r13 + DB 68,137,237 ; mov %r13d,%ebp + DB 73,193,237,32 ; shr $0x20,%r13 + DB 73,193,239,32 ; shr $0x20,%r15 + DB 196,131,121,32,4,50,0 ; vpinsrb $0x0,(%r10,%r14,1),%xmm0,%xmm0 + DB 196,195,121,32,4,26,1 ; vpinsrb $0x1,(%r10,%rbx,1),%xmm0,%xmm0 + DB 196,131,121,32,4,26,2 ; vpinsrb $0x2,(%r10,%r11,1),%xmm0,%xmm0 + DB 196,195,121,32,4,2,3 ; vpinsrb $0x3,(%r10,%rax,1),%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 - DB 196,193,249,126,195 ; vmovq %xmm0,%r11 - DB 196,131,121,32,4,57,0 ; vpinsrb $0x0,(%r9,%r15,1),%xmm0,%xmm0 - DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 - DB 196,131,121,32,4,41,2 ; vpinsrb $0x2,(%r9,%r13,1),%xmm0,%xmm0 - DB 196,131,121,32,4,33,3 ; vpinsrb $0x3,(%r9,%r12,1),%xmm0,%xmm0 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 196,193,249,126,198 ; vmovq %xmm0,%r14 + DB 196,131,121,32,4,34,0 ; vpinsrb $0x0,(%r10,%r12,1),%xmm0,%xmm0 + DB 196,131,121,32,4,58,1 ; vpinsrb $0x1,(%r10,%r15,1),%xmm0,%xmm0 + DB 196,195,121,32,4,42,2 ; vpinsrb $0x2,(%r10,%rbp,1),%xmm0,%xmm0 + DB 196,131,121,32,4,42,3 ; vpinsrb $0x3,(%r10,%r13,1),%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 73,139,88,8 ; mov 0x8(%r8),%rbx - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 73,193,233,30 ; shr $0x1e,%r9 + DB 73,139,105,8 ; mov 0x8(%r9),%rbp + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,217 ; mov %ebx,%r9d + DB 72,193,235,30 ; shr $0x1e,%rbx DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax - DB 65,137,198 ; mov %eax,%r14d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,30 ; shr $0x1e,%rax - DB 69,137,223 ; mov %r11d,%r15d + DB 69,137,247 ; mov %r14d,%r15d + DB 73,193,238,30 ; shr $0x1e,%r14 + DB 69,137,220 ; mov %r11d,%r12d DB 73,193,235,30 ; shr $0x1e,%r11 - DB 69,137,212 ; mov %r10d,%r12d - DB 73,193,234,30 ; shr $0x1e,%r10 - DB 196,161,121,110,4,131 ; vmovd (%rbx,%r8,4),%xmm0 - DB 196,163,121,34,4,11,1 ; vpinsrd $0x1,(%rbx,%r9,1),%xmm0,%xmm0 - DB 196,163,121,34,4,179,2 ; vpinsrd $0x2,(%rbx,%r14,4),%xmm0,%xmm0 - DB 196,99,121,34,4,3,3 ; vpinsrd $0x3,(%rbx,%rax,1),%xmm0,%xmm8 - DB 196,161,121,110,4,187 ; vmovd (%rbx,%r15,4),%xmm0 - DB 196,163,121,34,4,27,1 ; vpinsrd $0x1,(%rbx,%r11,1),%xmm0,%xmm0 - DB 196,163,121,34,4,163,2 ; vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0 - DB 196,163,121,34,28,19,3 ; vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3 + DB 196,161,121,110,68,141,0 ; vmovd 0x0(%rbp,%r9,4),%xmm0 + DB 196,227,121,34,68,29,0,1 ; vpinsrd $0x1,0x0(%rbp,%rbx,1),%xmm0,%xmm0 + DB 196,163,121,34,68,149,0,2 ; vpinsrd $0x2,0x0(%rbp,%r10,4),%xmm0,%xmm0 + DB 196,99,121,34,68,5,0,3 ; vpinsrd $0x3,0x0(%rbp,%rax,1),%xmm0,%xmm8 + DB 196,161,121,110,68,189,0 ; vmovd 0x0(%rbp,%r15,4),%xmm0 + DB 196,163,121,34,68,53,0,1 ; vpinsrd $0x1,0x0(%rbp,%r14,1),%xmm0,%xmm0 + DB 196,163,121,34,68,165,0,2 ; vpinsrd $0x2,0x0(%rbp,%r12,4),%xmm0,%xmm0 + DB 196,163,121,34,92,29,0,3 ; vpinsrd $0x3,0x0(%rbp,%r11,1),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,30,49,0,0 ; vmovaps 0x311e(%rip),%ymm10 # 6ba0 <_sk_callback_avx+0x640> + DB 197,124,40,21,164,49,0,0 ; vmovaps 0x31a4(%rip),%ymm10 # 6c00 <_sk_callback_avx+0x64a> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,88,46,0,0 ; vbroadcastss 0x2e58(%rip),%ymm9 # 68ec <_sk_callback_avx+0x38c> + DB 196,98,125,24,13,222,46,0,0 ; vbroadcastss 0x2ede(%rip),%ymm9 # 694c <_sk_callback_avx+0x396> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -8605,77 +8580,78 @@ _sk_gather_i8_avx LABEL PROC DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_load_565_avx _sk_load_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,128,0,0,0 ; jne 3b8c <_sk_load_565_avx+0x8e> - DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,128,0,0,0 ; jne 3b67 <_sk_load_565_avx+0x8e> + DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,194,45,0,0 ; vbroadcastss 0x2dc2(%rip),%ymm0 # 68f0 <_sk_callback_avx+0x390> + DB 196,226,125,24,5,71,46,0,0 ; vbroadcastss 0x2e47(%rip),%ymm0 # 6950 <_sk_callback_avx+0x39a> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,181,45,0,0 ; vbroadcastss 0x2db5(%rip),%ymm1 # 68f4 <_sk_callback_avx+0x394> + DB 196,226,125,24,13,58,46,0,0 ; vbroadcastss 0x2e3a(%rip),%ymm1 # 6954 <_sk_callback_avx+0x39e> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,172,45,0,0 ; vbroadcastss 0x2dac(%rip),%ymm1 # 68f8 <_sk_callback_avx+0x398> + DB 196,226,125,24,13,49,46,0,0 ; vbroadcastss 0x2e31(%rip),%ymm1 # 6958 <_sk_callback_avx+0x3a2> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,159,45,0,0 ; vbroadcastss 0x2d9f(%rip),%ymm3 # 68fc <_sk_callback_avx+0x39c> + DB 196,226,125,24,29,36,46,0,0 ; vbroadcastss 0x2e24(%rip),%ymm3 # 695c <_sk_callback_avx+0x3a6> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,150,45,0,0 ; vbroadcastss 0x2d96(%rip),%ymm3 # 6900 <_sk_callback_avx+0x3a0> + DB 196,226,125,24,29,27,46,0,0 ; vbroadcastss 0x2e1b(%rip),%ymm3 # 6960 <_sk_callback_avx+0x3aa> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,137,45,0,0 ; vbroadcastss 0x2d89(%rip),%ymm3 # 6904 <_sk_callback_avx+0x3a4> + DB 196,226,125,24,29,14,46,0,0 ; vbroadcastss 0x2e0e(%rip),%ymm3 # 6964 <_sk_callback_avx+0x3ae> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,126,45,0,0 ; vbroadcastss 0x2d7e(%rip),%ymm3 # 6908 <_sk_callback_avx+0x3a8> + DB 196,226,125,24,29,3,46,0,0 ; vbroadcastss 0x2e03(%rip),%ymm3 # 6968 <_sk_callback_avx+0x3b2> DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,110,255,255,255 ; ja 3b12 <_sk_load_565_avx+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 3bf8 <_sk_load_565_avx+0xfa> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 15,135,110,255,255,255 ; ja 3aed <_sk_load_565_avx+0x14> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,74,0,0,0 ; lea 0x4a(%rip),%r10 # 3bd4 <_sk_load_565_avx+0xfb> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,193,121,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,26,255,255,255 ; jmpq 3b12 <_sk_load_565_avx+0x14> - DB 244 ; hlt + DB 196,193,121,196,68,83,12,6 ; vpinsrw $0x6,0xc(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,10,5 ; vpinsrw $0x5,0xa(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,8,4 ; vpinsrw $0x4,0x8(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,6,3 ; vpinsrw $0x3,0x6(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0 + DB 233,26,255,255,255 ; jmpq 3aed <_sk_load_565_avx+0x14> + DB 144 ; nop + DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) + DB 235,255 ; jmp 3bd9 <_sk_load_565_avx+0x100> DB 255 ; (bad) - DB 236 ; in (%dx),%al + DB 255,227 ; jmpq *%rbx DB 255 ; (bad) DB 255 ; (bad) - DB 255,228 ; jmpq *%rsp - DB 255 ; (bad) DB 255 ; (bad) + DB 219,255 ; (bad) DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) + DB 255,211 ; callq *%rbx DB 255 ; (bad) - DB 255,212 ; callq *%rsp DB 255 ; (bad) + DB 255,203 ; dec %ebx DB 255 ; (bad) - DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax + DB 191 ; .byte 0xbf DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8685,10 +8661,11 @@ _sk_gather_565_avx LABEL PROC DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 @@ -8699,57 +8676,58 @@ _sk_gather_565_avx LABEL PROC DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,15,183,28,88 ; movzwl (%r8,%rbx,2),%ebx - DB 67,15,183,44,112 ; movzwl (%r8,%r14,2),%ebp + DB 65,15,183,28,89 ; movzwl (%r9,%rbx,2),%ebx + DB 67,15,183,44,121 ; movzwl (%r9,%r15,2),%ebp DB 197,249,110,197 ; vmovd %ebp,%xmm0 DB 197,249,196,195,1 ; vpinsrw $0x1,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,96 ; movzwl (%r8,%r12,2),%ebx + DB 67,15,183,28,105 ; movzwl (%r9,%r13,2),%ebx DB 197,249,196,195,2 ; vpinsrw $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,120 ; movzwl (%r8,%r15,2),%ebx + DB 67,15,183,28,97 ; movzwl (%r9,%r12,2),%ebx DB 197,249,196,195,3 ; vpinsrw $0x3,%ebx,%xmm0,%xmm0 - DB 67,15,183,44,88 ; movzwl (%r8,%r11,2),%ebp - DB 197,249,196,197,4 ; vpinsrw $0x4,%ebp,%xmm0,%xmm0 - DB 67,15,183,44,80 ; movzwl (%r8,%r10,2),%ebp + DB 67,15,183,28,113 ; movzwl (%r9,%r14,2),%ebx + DB 197,249,196,195,4 ; vpinsrw $0x4,%ebx,%xmm0,%xmm0 + DB 67,15,183,44,89 ; movzwl (%r9,%r11,2),%ebp DB 197,249,196,197,5 ; vpinsrw $0x5,%ebp,%xmm0,%xmm0 - DB 67,15,183,44,72 ; movzwl (%r8,%r9,2),%ebp + DB 67,15,183,44,81 ; movzwl (%r9,%r10,2),%ebp DB 197,249,196,197,6 ; vpinsrw $0x6,%ebp,%xmm0,%xmm0 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,30,44,0,0 ; vbroadcastss 0x2c1e(%rip),%ymm0 # 690c <_sk_callback_avx+0x3ac> + DB 196,226,125,24,5,160,44,0,0 ; vbroadcastss 0x2ca0(%rip),%ymm0 # 696c <_sk_callback_avx+0x3b6> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,17,44,0,0 ; vbroadcastss 0x2c11(%rip),%ymm1 # 6910 <_sk_callback_avx+0x3b0> + DB 196,226,125,24,13,147,44,0,0 ; vbroadcastss 0x2c93(%rip),%ymm1 # 6970 <_sk_callback_avx+0x3ba> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,8,44,0,0 ; vbroadcastss 0x2c08(%rip),%ymm1 # 6914 <_sk_callback_avx+0x3b4> + DB 196,226,125,24,13,138,44,0,0 ; vbroadcastss 0x2c8a(%rip),%ymm1 # 6974 <_sk_callback_avx+0x3be> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,251,43,0,0 ; vbroadcastss 0x2bfb(%rip),%ymm3 # 6918 <_sk_callback_avx+0x3b8> + DB 196,226,125,24,29,125,44,0,0 ; vbroadcastss 0x2c7d(%rip),%ymm3 # 6978 <_sk_callback_avx+0x3c2> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,242,43,0,0 ; vbroadcastss 0x2bf2(%rip),%ymm3 # 691c <_sk_callback_avx+0x3bc> + DB 196,226,125,24,29,116,44,0,0 ; vbroadcastss 0x2c74(%rip),%ymm3 # 697c <_sk_callback_avx+0x3c6> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,229,43,0,0 ; vbroadcastss 0x2be5(%rip),%ymm3 # 6920 <_sk_callback_avx+0x3c0> + DB 196,226,125,24,29,103,44,0,0 ; vbroadcastss 0x2c67(%rip),%ymm3 # 6980 <_sk_callback_avx+0x3ca> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,218,43,0,0 ; vbroadcastss 0x2bda(%rip),%ymm3 # 6924 <_sk_callback_avx+0x3c4> + DB 196,226,125,24,29,92,44,0,0 ; vbroadcastss 0x2c5c(%rip),%ymm3 # 6984 <_sk_callback_avx+0x3ce> DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 93 ; pop %rbp @@ -8758,15 +8736,15 @@ _sk_gather_565_avx LABEL PROC PUBLIC _sk_store_565_avx _sk_store_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,198,43,0,0 ; vbroadcastss 0x2bc6(%rip),%ymm8 # 6928 <_sk_callback_avx+0x3c8> + DB 76,139,24 ; mov (%rax),%r11 + DB 196,98,125,24,5,70,44,0,0 ; vbroadcastss 0x2c46(%rip),%ymm8 # 6988 <_sk_callback_avx+0x3d2> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,11 ; vpslld $0xb,%xmm9,%xmm10 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 DB 196,193,49,114,241,11 ; vpslld $0xb,%xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,98,125,24,21,159,43,0,0 ; vbroadcastss 0x2b9f(%rip),%ymm10 # 692c <_sk_callback_avx+0x3cc> + DB 196,98,125,24,21,31,44,0,0 ; vbroadcastss 0x2c1f(%rip),%ymm10 # 698c <_sk_callback_avx+0x3d6> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,33,114,242,5 ; vpslld $0x5,%xmm10,%xmm11 @@ -8779,29 +8757,29 @@ _sk_store_565_avx LABEL PROC DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3ddd <_sk_store_565_avx+0x89> - DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 3dd9 <_sk_store_565_avx+0x85> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3e3c <_sk_store_565_avx+0xe8> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,121,21,68,122,12,6 ; vpextrw $0x6,%xmm8,0xc(%r10,%rdi,2) - DB 196,67,121,21,68,122,10,5 ; vpextrw $0x5,%xmm8,0xa(%r10,%rdi,2) - DB 196,67,121,21,68,122,8,4 ; vpextrw $0x4,%xmm8,0x8(%r10,%rdi,2) - DB 196,67,121,21,68,122,6,3 ; vpextrw $0x3,%xmm8,0x6(%r10,%rdi,2) - DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) - DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) - DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2) - DB 235,159 ; jmp 3dd9 <_sk_store_565_avx+0x85> + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3dbd <_sk_store_565_avx+0x89> + DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 3db9 <_sk_store_565_avx+0x85> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 3e1c <_sk_store_565_avx+0xe8> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,67,121,21,68,83,12,6 ; vpextrw $0x6,%xmm8,0xc(%r11,%rdx,2) + DB 196,67,121,21,68,83,10,5 ; vpextrw $0x5,%xmm8,0xa(%r11,%rdx,2) + DB 196,67,121,21,68,83,8,4 ; vpextrw $0x4,%xmm8,0x8(%r11,%rdx,2) + DB 196,67,121,21,68,83,6,3 ; vpextrw $0x3,%xmm8,0x6(%r11,%rdx,2) + DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2) + DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2) + DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2) + DB 235,159 ; jmp 3db9 <_sk_store_565_avx+0x85> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -8830,56 +8808,56 @@ _sk_store_565_avx LABEL PROC PUBLIC _sk_load_4444_avx _sk_load_4444_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,152,0,0,0 ; jne 3efe <_sk_load_4444_avx+0xa6> - DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,152,0,0,0 ; jne 3ede <_sk_load_4444_avx+0xa6> + DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,168,42,0,0 ; vbroadcastss 0x2aa8(%rip),%ymm0 # 6930 <_sk_callback_avx+0x3d0> + DB 196,226,125,24,5,40,43,0,0 ; vbroadcastss 0x2b28(%rip),%ymm0 # 6990 <_sk_callback_avx+0x3da> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,155,42,0,0 ; vbroadcastss 0x2a9b(%rip),%ymm1 # 6934 <_sk_callback_avx+0x3d4> + DB 196,226,125,24,13,27,43,0,0 ; vbroadcastss 0x2b1b(%rip),%ymm1 # 6994 <_sk_callback_avx+0x3de> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,146,42,0,0 ; vbroadcastss 0x2a92(%rip),%ymm1 # 6938 <_sk_callback_avx+0x3d8> + DB 196,226,125,24,13,18,43,0,0 ; vbroadcastss 0x2b12(%rip),%ymm1 # 6998 <_sk_callback_avx+0x3e2> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,133,42,0,0 ; vbroadcastss 0x2a85(%rip),%ymm2 # 693c <_sk_callback_avx+0x3dc> + DB 196,226,125,24,21,5,43,0,0 ; vbroadcastss 0x2b05(%rip),%ymm2 # 699c <_sk_callback_avx+0x3e6> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,124,42,0,0 ; vbroadcastss 0x2a7c(%rip),%ymm2 # 6940 <_sk_callback_avx+0x3e0> + DB 196,226,125,24,21,252,42,0,0 ; vbroadcastss 0x2afc(%rip),%ymm2 # 69a0 <_sk_callback_avx+0x3ea> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,111,42,0,0 ; vbroadcastss 0x2a6f(%rip),%ymm8 # 6944 <_sk_callback_avx+0x3e4> + DB 196,98,125,24,5,239,42,0,0 ; vbroadcastss 0x2aef(%rip),%ymm8 # 69a4 <_sk_callback_avx+0x3ee> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,101,42,0,0 ; vbroadcastss 0x2a65(%rip),%ymm8 # 6948 <_sk_callback_avx+0x3e8> + DB 196,98,125,24,5,229,42,0,0 ; vbroadcastss 0x2ae5(%rip),%ymm8 # 69a8 <_sk_callback_avx+0x3f2> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,87,42,0,0 ; vbroadcastss 0x2a57(%rip),%ymm8 # 694c <_sk_callback_avx+0x3ec> + DB 196,98,125,24,5,215,42,0,0 ; vbroadcastss 0x2ad7(%rip),%ymm8 # 69ac <_sk_callback_avx+0x3f6> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,86,255,255,255 ; ja 3e6c <_sk_load_4444_avx+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 3f6c <_sk_load_4444_avx+0x114> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 15,135,86,255,255,255 ; ja 3e4c <_sk_load_4444_avx+0x14> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 3f4c <_sk_load_4444_avx+0x114> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,193,121,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 - DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,2,255,255,255 ; jmpq 3e6c <_sk_load_4444_avx+0x14> + DB 196,193,121,196,68,83,12,6 ; vpinsrw $0x6,0xc(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,10,5 ; vpinsrw $0x5,0xa(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,8,4 ; vpinsrw $0x4,0x8(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,6,3 ; vpinsrw $0x3,0x6(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0 + DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0 + DB 233,2,255,255,255 ; jmpq 3e4c <_sk_load_4444_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 242,255 ; repnz (bad) DB 255 ; (bad) @@ -8910,10 +8888,11 @@ _sk_gather_4444_avx LABEL PROC DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 @@ -8924,61 +8903,62 @@ _sk_gather_4444_avx LABEL PROC DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,15,183,28,88 ; movzwl (%r8,%rbx,2),%ebx - DB 67,15,183,44,112 ; movzwl (%r8,%r14,2),%ebp + DB 65,15,183,28,89 ; movzwl (%r9,%rbx,2),%ebx + DB 67,15,183,44,121 ; movzwl (%r9,%r15,2),%ebp DB 197,249,110,197 ; vmovd %ebp,%xmm0 DB 197,249,196,195,1 ; vpinsrw $0x1,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,96 ; movzwl (%r8,%r12,2),%ebx + DB 67,15,183,28,105 ; movzwl (%r9,%r13,2),%ebx DB 197,249,196,195,2 ; vpinsrw $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,120 ; movzwl (%r8,%r15,2),%ebx + DB 67,15,183,28,97 ; movzwl (%r9,%r12,2),%ebx DB 197,249,196,195,3 ; vpinsrw $0x3,%ebx,%xmm0,%xmm0 - DB 67,15,183,44,88 ; movzwl (%r8,%r11,2),%ebp - DB 197,249,196,197,4 ; vpinsrw $0x4,%ebp,%xmm0,%xmm0 - DB 67,15,183,44,80 ; movzwl (%r8,%r10,2),%ebp + DB 67,15,183,28,113 ; movzwl (%r9,%r14,2),%ebx + DB 197,249,196,195,4 ; vpinsrw $0x4,%ebx,%xmm0,%xmm0 + DB 67,15,183,44,89 ; movzwl (%r9,%r11,2),%ebp DB 197,249,196,197,5 ; vpinsrw $0x5,%ebp,%xmm0,%xmm0 - DB 67,15,183,44,72 ; movzwl (%r8,%r9,2),%ebp + DB 67,15,183,44,81 ; movzwl (%r9,%r10,2),%ebp DB 197,249,196,197,6 ; vpinsrw $0x6,%ebp,%xmm0,%xmm0 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,238,40,0,0 ; vbroadcastss 0x28ee(%rip),%ymm0 # 6950 <_sk_callback_avx+0x3f0> + DB 196,226,125,24,5,108,41,0,0 ; vbroadcastss 0x296c(%rip),%ymm0 # 69b0 <_sk_callback_avx+0x3fa> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,225,40,0,0 ; vbroadcastss 0x28e1(%rip),%ymm1 # 6954 <_sk_callback_avx+0x3f4> + DB 196,226,125,24,13,95,41,0,0 ; vbroadcastss 0x295f(%rip),%ymm1 # 69b4 <_sk_callback_avx+0x3fe> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,216,40,0,0 ; vbroadcastss 0x28d8(%rip),%ymm1 # 6958 <_sk_callback_avx+0x3f8> + DB 196,226,125,24,13,86,41,0,0 ; vbroadcastss 0x2956(%rip),%ymm1 # 69b8 <_sk_callback_avx+0x402> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,203,40,0,0 ; vbroadcastss 0x28cb(%rip),%ymm2 # 695c <_sk_callback_avx+0x3fc> + DB 196,226,125,24,21,73,41,0,0 ; vbroadcastss 0x2949(%rip),%ymm2 # 69bc <_sk_callback_avx+0x406> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,194,40,0,0 ; vbroadcastss 0x28c2(%rip),%ymm2 # 6960 <_sk_callback_avx+0x400> + DB 196,226,125,24,21,64,41,0,0 ; vbroadcastss 0x2940(%rip),%ymm2 # 69c0 <_sk_callback_avx+0x40a> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,181,40,0,0 ; vbroadcastss 0x28b5(%rip),%ymm8 # 6964 <_sk_callback_avx+0x404> + DB 196,98,125,24,5,51,41,0,0 ; vbroadcastss 0x2933(%rip),%ymm8 # 69c4 <_sk_callback_avx+0x40e> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,171,40,0,0 ; vbroadcastss 0x28ab(%rip),%ymm8 # 6968 <_sk_callback_avx+0x408> + DB 196,98,125,24,5,41,41,0,0 ; vbroadcastss 0x2929(%rip),%ymm8 # 69c8 <_sk_callback_avx+0x412> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,157,40,0,0 ; vbroadcastss 0x289d(%rip),%ymm8 # 696c <_sk_callback_avx+0x40c> + DB 196,98,125,24,5,27,41,0,0 ; vbroadcastss 0x291b(%rip),%ymm8 # 69cc <_sk_callback_avx+0x416> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 93 ; pop %rbp @@ -8987,8 +8967,8 @@ _sk_gather_4444_avx LABEL PROC PUBLIC _sk_store_4444_avx _sk_store_4444_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,130,40,0,0 ; vbroadcastss 0x2882(%rip),%ymm8 # 6970 <_sk_callback_avx+0x410> + DB 76,139,24 ; mov (%rax),%r11 + DB 196,98,125,24,5,254,40,0,0 ; vbroadcastss 0x28fe(%rip),%ymm8 # 69d0 <_sk_callback_avx+0x41a> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,12 ; vpslld $0xc,%xmm9,%xmm10 @@ -9014,29 +8994,29 @@ _sk_store_4444_avx LABEL PROC DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 4187 <_sk_store_4444_avx+0xa7> - DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 4183 <_sk_store_4444_avx+0xa3> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,66,0,0,0 ; lea 0x42(%rip),%r9 # 41e4 <_sk_store_4444_avx+0x104> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,121,21,68,122,12,6 ; vpextrw $0x6,%xmm8,0xc(%r10,%rdi,2) - DB 196,67,121,21,68,122,10,5 ; vpextrw $0x5,%xmm8,0xa(%r10,%rdi,2) - DB 196,67,121,21,68,122,8,4 ; vpextrw $0x4,%xmm8,0x8(%r10,%rdi,2) - DB 196,67,121,21,68,122,6,3 ; vpextrw $0x3,%xmm8,0x6(%r10,%rdi,2) - DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) - DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) - DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2) - DB 235,159 ; jmp 4183 <_sk_store_4444_avx+0xa3> + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 416b <_sk_store_4444_avx+0xa7> + DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 4167 <_sk_store_4444_avx+0xa3> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 41c8 <_sk_store_4444_avx+0x104> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,67,121,21,68,83,12,6 ; vpextrw $0x6,%xmm8,0xc(%r11,%rdx,2) + DB 196,67,121,21,68,83,10,5 ; vpextrw $0x5,%xmm8,0xa(%r11,%rdx,2) + DB 196,67,121,21,68,83,8,4 ; vpextrw $0x4,%xmm8,0x8(%r11,%rdx,2) + DB 196,67,121,21,68,83,6,3 ; vpextrw $0x3,%xmm8,0x6(%r11,%rdx,2) + DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2) + DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2) + DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2) + DB 235,159 ; jmp 4167 <_sk_store_4444_avx+0xa3> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -9063,17 +9043,17 @@ _sk_store_4444_avx LABEL PROC PUBLIC _sk_load_8888_avx _sk_load_8888_avx LABEL PROC DB 80 ; push %rax - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 DB 77,133,192 ; test %r8,%r8 - DB 15,133,139,0,0,0 ; jne 42a5 <_sk_load_8888_avx+0xa5> - DB 196,193,124,16,25 ; vmovups (%r9),%ymm3 - DB 197,124,40,21,153,41,0,0 ; vmovaps 0x2999(%rip),%ymm10 # 6bc0 <_sk_callback_avx+0x660> + DB 15,133,139,0,0,0 ; jne 4289 <_sk_load_8888_avx+0xa5> + DB 196,193,124,16,26 ; vmovups (%r10),%ymm3 + DB 197,124,40,21,21,42,0,0 ; vmovaps 0x2a15(%rip),%ymm10 # 6c20 <_sk_callback_avx+0x66a> DB 196,193,100,84,194 ; vandps %ymm10,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,59,39,0,0 ; vbroadcastss 0x273b(%rip),%ymm8 # 6974 <_sk_callback_avx+0x414> + DB 196,98,125,24,5,183,39,0,0 ; vbroadcastss 0x27b7(%rip),%ymm8 # 69d4 <_sk_callback_avx+0x41e> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,241,114,211,8 ; vpsrld $0x8,%xmm3,%xmm1 DB 196,195,125,25,217,1 ; vextractf128 $0x1,%ymm3,%xmm9 @@ -9094,8 +9074,8 @@ _sk_load_8888_avx LABEL PROC DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx - DB 65,88 ; pop %r8 + DB 76,137,201 ; mov %r9,%rcx + DB 65,89 ; pop %r9 DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx @@ -9104,22 +9084,23 @@ _sk_load_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0 - DB 196,226,121,0,13,67,40,0,0 ; vpshufb 0x2843(%rip),%xmm0,%xmm1 # 6b10 <_sk_callback_avx+0x5b0> + DB 196,226,121,0,13,191,40,0,0 ; vpshufb 0x28bf(%rip),%xmm0,%xmm1 # 6b70 <_sk_callback_avx+0x5ba> DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1 - DB 196,226,121,0,5,69,40,0,0 ; vpshufb 0x2845(%rip),%xmm0,%xmm0 # 6b20 <_sk_callback_avx+0x5c0> + DB 196,226,121,0,5,193,40,0,0 ; vpshufb 0x28c1(%rip),%xmm0,%xmm0 # 6b80 <_sk_callback_avx+0x5ca> DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 196,194,125,44,25 ; vmaskmovps (%r9),%ymm0,%ymm3 - DB 233,47,255,255,255 ; jmpq 421f <_sk_load_8888_avx+0x1f> + DB 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3 + DB 233,47,255,255,255 ; jmpq 4203 <_sk_load_8888_avx+0x1f> PUBLIC _sk_gather_8888_avx _sk_gather_8888_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 DB 197,249,110,80,16 ; vmovd 0x10(%rax),%xmm2 DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 @@ -9130,31 +9111,31 @@ _sk_gather_8888_avx LABEL PROC DB 196,227,125,25,194,1 ; vextractf128 $0x1,%ymm0,%xmm2 DB 197,241,254,202 ; vpaddd %xmm2,%xmm1,%xmm1 DB 196,225,249,126,200 ; vmovq %xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,121,110,4,176 ; vmovd (%r8,%r14,4),%xmm0 - DB 196,195,121,34,4,152,1 ; vpinsrd $0x1,(%r8,%rbx,4),%xmm0,%xmm0 - DB 196,131,121,34,4,160,2 ; vpinsrd $0x2,(%r8,%r12,4),%xmm0,%xmm0 - DB 196,3,121,34,4,184,3 ; vpinsrd $0x3,(%r8,%r15,4),%xmm0,%xmm8 - DB 196,129,121,110,4,136 ; vmovd (%r8,%r9,4),%xmm0 - DB 196,195,121,34,4,128,1 ; vpinsrd $0x1,(%r8,%rax,4),%xmm0,%xmm0 - DB 196,131,121,34,4,152,2 ; vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 - DB 196,131,121,34,28,144,3 ; vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,129,121,110,4,185 ; vmovd (%r9,%r15,4),%xmm0 + DB 196,195,121,34,4,153,1 ; vpinsrd $0x1,(%r9,%rbx,4),%xmm0,%xmm0 + DB 196,131,121,34,4,169,2 ; vpinsrd $0x2,(%r9,%r13,4),%xmm0,%xmm0 + DB 196,3,121,34,4,161,3 ; vpinsrd $0x3,(%r9,%r12,4),%xmm0,%xmm8 + DB 196,129,121,110,4,145 ; vmovd (%r9,%r10,4),%xmm0 + DB 196,195,121,34,4,129,1 ; vpinsrd $0x1,(%r9,%rax,4),%xmm0,%xmm0 + DB 196,131,121,34,4,177,2 ; vpinsrd $0x2,(%r9,%r14,4),%xmm0,%xmm0 + DB 196,131,121,34,28,153,3 ; vpinsrd $0x3,(%r9,%r11,4),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,62,40,0,0 ; vmovaps 0x283e(%rip),%ymm10 # 6be0 <_sk_callback_avx+0x680> + DB 197,124,40,21,184,40,0,0 ; vmovaps 0x28b8(%rip),%ymm10 # 6c40 <_sk_callback_avx+0x68a> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,196,37,0,0 ; vbroadcastss 0x25c4(%rip),%ymm9 # 6978 <_sk_callback_avx+0x418> + DB 196,98,125,24,13,62,38,0,0 ; vbroadcastss 0x263e(%rip),%ymm9 # 69d8 <_sk_callback_avx+0x422> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -9176,6 +9157,7 @@ _sk_gather_8888_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -9183,11 +9165,11 @@ _sk_gather_8888_avx LABEL PROC PUBLIC _sk_store_8888_avx _sk_store_8888_avx LABEL PROC DB 80 ; push %rax - DB 73,137,200 ; mov %rcx,%r8 + DB 73,137,201 ; mov %rcx,%r9 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,3,8 ; add (%rax),%r9 - DB 196,98,125,24,5,70,37,0,0 ; vbroadcastss 0x2546(%rip),%ymm8 # 697c <_sk_callback_avx+0x41c> + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,3,16 ; add (%rax),%r10 + DB 196,98,125,24,5,190,37,0,0 ; vbroadcastss 0x25be(%rip),%ymm8 # 69dc <_sk_callback_avx+0x426> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -9212,11 +9194,11 @@ _sk_store_8888_avx LABEL PROC DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8 DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 77,133,192 ; test %r8,%r8 - DB 117,14 ; jne 44c8 <_sk_store_8888_avx+0xac> - DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) + DB 117,14 ; jne 44b0 <_sk_store_8888_avx+0xac> + DB 196,65,124,17,2 ; vmovups %ymm8,(%r10) DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,193 ; mov %r8,%rcx - DB 65,88 ; pop %r8 + DB 76,137,201 ; mov %r9,%rcx + DB 65,89 ; pop %r9 DB 255,224 ; jmpq *%rax DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx @@ -9225,29 +9207,29 @@ _sk_store_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 - DB 196,98,49,0,21,64,38,0,0 ; vpshufb 0x2640(%rip),%xmm9,%xmm10 # 6b30 <_sk_callback_avx+0x5d0> + DB 196,98,49,0,21,184,38,0,0 ; vpshufb 0x26b8(%rip),%xmm9,%xmm10 # 6b90 <_sk_callback_avx+0x5da> DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10 - DB 196,98,49,0,13,66,38,0,0 ; vpshufb 0x2642(%rip),%xmm9,%xmm9 # 6b40 <_sk_callback_avx+0x5e0> + DB 196,98,49,0,13,186,38,0,0 ; vpshufb 0x26ba(%rip),%xmm9,%xmm9 # 6ba0 <_sk_callback_avx+0x5ea> DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) - DB 235,175 ; jmp 44bf <_sk_store_8888_avx+0xa3> + DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10) + DB 235,175 ; jmp 44a7 <_sk_store_8888_avx+0xa3> PUBLIC _sk_load_f16_avx _sk_load_f16_avx LABEL PROC DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx + DB 77,133,192 ; test %r8,%r8 DB 197,252,17,124,36,96 ; vmovups %ymm7,0x60(%rsp) DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) - DB 15,133,143,2,0,0 ; jne 47cb <_sk_load_f16_avx+0x2bb> - DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 - DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 - DB 197,249,16,76,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm1 - DB 197,122,111,76,248,48 ; vmovdqu 0x30(%rax,%rdi,8),%xmm9 + DB 15,133,143,2,0,0 ; jne 47b3 <_sk_load_f16_avx+0x2bb> + DB 197,121,16,4,208 ; vmovupd (%rax,%rdx,8),%xmm8 + DB 197,249,16,84,208,16 ; vmovupd 0x10(%rax,%rdx,8),%xmm2 + DB 197,249,16,76,208,32 ; vmovupd 0x20(%rax,%rdx,8),%xmm1 + DB 197,122,111,76,208,48 ; vmovdqu 0x30(%rax,%rdx,8),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,113,97,217 ; vpunpcklwd %xmm9,%xmm1,%xmm3 @@ -9261,13 +9243,13 @@ _sk_load_f16_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,98,125,24,37,235,35,0,0 ; vbroadcastss 0x23eb(%rip),%ymm12 # 6980 <_sk_callback_avx+0x420> + DB 196,98,125,24,37,99,36,0,0 ; vbroadcastss 0x2463(%rip),%ymm12 # 69e0 <_sk_callback_avx+0x42a> DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,215,35,0,0 ; vbroadcastss 0x23d7(%rip),%xmm11 # 6984 <_sk_callback_avx+0x424> + DB 196,98,121,24,29,79,36,0,0 ; vbroadcastss 0x244f(%rip),%xmm11 # 69e4 <_sk_callback_avx+0x42e> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,205,35,0,0 ; vbroadcastss 0x23cd(%rip),%xmm13 # 6988 <_sk_callback_avx+0x428> + DB 196,98,121,24,45,69,36,0,0 ; vbroadcastss 0x2445(%rip),%xmm13 # 69e8 <_sk_callback_avx+0x432> DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -9281,7 +9263,7 @@ _sk_load_f16_avx LABEL PROC DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0 DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 - DB 196,226,121,24,29,131,35,0,0 ; vbroadcastss 0x2383(%rip),%xmm3 # 698c <_sk_callback_avx+0x42c> + DB 196,226,121,24,29,251,35,0,0 ; vbroadcastss 0x23fb(%rip),%xmm3 # 69ec <_sk_callback_avx+0x436> DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -9371,46 +9353,47 @@ _sk_load_f16_avx LABEL PROC DB 197,252,16,124,36,96 ; vmovups 0x60(%rsp),%ymm7 DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax - DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 + DB 197,123,16,4,208 ; vmovsd (%rax,%rdx,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 482a <_sk_load_f16_avx+0x31a> - DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 482a <_sk_load_f16_avx+0x31a> - DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 4837 <_sk_load_f16_avx+0x327> - DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 4837 <_sk_load_f16_avx+0x327> - DB 197,251,16,76,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm1 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,68,253,255,255 ; je 4553 <_sk_load_f16_avx+0x43> - DB 197,241,22,76,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,52,253,255,255 ; jb 4553 <_sk_load_f16_avx+0x43> - DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,41,253,255,255 ; jmpq 4553 <_sk_load_f16_avx+0x43> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,79 ; je 4812 <_sk_load_f16_avx+0x31a> + DB 197,57,22,68,208,8 ; vmovhpd 0x8(%rax,%rdx,8),%xmm8,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,67 ; jb 4812 <_sk_load_f16_avx+0x31a> + DB 197,251,16,84,208,16 ; vmovsd 0x10(%rax,%rdx,8),%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 116,68 ; je 481f <_sk_load_f16_avx+0x327> + DB 197,233,22,84,208,24 ; vmovhpd 0x18(%rax,%rdx,8),%xmm2,%xmm2 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,56 ; jb 481f <_sk_load_f16_avx+0x327> + DB 197,251,16,76,208,32 ; vmovsd 0x20(%rax,%rdx,8),%xmm1 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 15,132,68,253,255,255 ; je 453b <_sk_load_f16_avx+0x43> + DB 197,241,22,76,208,40 ; vmovhpd 0x28(%rax,%rdx,8),%xmm1,%xmm1 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 15,130,52,253,255,255 ; jb 453b <_sk_load_f16_avx+0x43> + DB 197,122,126,76,208,48 ; vmovq 0x30(%rax,%rdx,8),%xmm9 + DB 233,41,253,255,255 ; jmpq 453b <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,28,253,255,255 ; jmpq 4553 <_sk_load_f16_avx+0x43> + DB 233,28,253,255,255 ; jmpq 453b <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 - DB 233,19,253,255,255 ; jmpq 4553 <_sk_load_f16_avx+0x43> + DB 233,19,253,255,255 ; jmpq 453b <_sk_load_f16_avx+0x43> PUBLIC _sk_gather_f16_avx _sk_gather_f16_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp + DB 72,129,236,144,0,0,0 ; sub $0x90,%rsp DB 197,252,17,124,36,96 ; vmovups %ymm7,0x60(%rsp) DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 @@ -9421,29 +9404,29 @@ _sk_gather_f16_avx LABEL PROC DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 DB 196,225,249,126,195 ; vmovq %xmm0,%rbx - DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 - DB 69,137,252 ; mov %r15d,%r12d - DB 73,193,239,32 ; shr $0x20,%r15 + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,122,126,4,216 ; vmovq (%r8,%rbx,8),%xmm0 - DB 196,129,122,126,12,240 ; vmovq (%r8,%r14,8),%xmm1 + DB 196,193,122,126,4,217 ; vmovq (%r9,%rbx,8),%xmm0 + DB 196,129,122,126,12,249 ; vmovq (%r9,%r15,8),%xmm1 DB 197,113,108,200 ; vpunpcklqdq %xmm0,%xmm1,%xmm9 - DB 196,129,122,126,12,248 ; vmovq (%r8,%r15,8),%xmm1 - DB 196,129,122,126,20,224 ; vmovq (%r8,%r12,8),%xmm2 + DB 196,129,122,126,12,225 ; vmovq (%r9,%r12,8),%xmm1 + DB 196,129,122,126,20,233 ; vmovq (%r9,%r13,8),%xmm2 DB 197,233,108,201 ; vpunpcklqdq %xmm1,%xmm2,%xmm1 - DB 196,129,122,126,20,208 ; vmovq (%r8,%r10,8),%xmm2 - DB 196,129,122,126,28,216 ; vmovq (%r8,%r11,8),%xmm3 + DB 196,129,122,126,20,217 ; vmovq (%r9,%r11,8),%xmm2 + DB 196,129,122,126,28,241 ; vmovq (%r9,%r14,8),%xmm3 DB 197,97,108,210 ; vpunpcklqdq %xmm2,%xmm3,%xmm10 - DB 196,65,122,126,4,192 ; vmovq (%r8,%rax,8),%xmm8 - DB 196,129,122,126,28,200 ; vmovq (%r8,%r9,8),%xmm3 + DB 196,65,122,126,4,193 ; vmovq (%r9,%rax,8),%xmm8 + DB 196,129,122,126,28,209 ; vmovq (%r9,%r10,8),%xmm3 DB 196,193,97,108,216 ; vpunpcklqdq %xmm8,%xmm3,%xmm3 DB 197,177,97,193 ; vpunpcklwd %xmm1,%xmm9,%xmm0 DB 197,177,105,201 ; vpunpckhwd %xmm1,%xmm9,%xmm1 @@ -9458,13 +9441,13 @@ _sk_gather_f16_avx LABEL PROC DB 197,249,105,210 ; vpunpckhwd %xmm2,%xmm0,%xmm2 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - DB 196,98,125,24,37,67,32,0,0 ; vbroadcastss 0x2043(%rip),%ymm12 # 6990 <_sk_callback_avx+0x430> + DB 196,98,125,24,37,185,32,0,0 ; vbroadcastss 0x20b9(%rip),%ymm12 # 69f0 <_sk_callback_avx+0x43a> DB 196,193,124,84,212 ; vandps %ymm12,%ymm0,%ymm2 DB 197,252,87,194 ; vxorps %ymm2,%ymm0,%ymm0 DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,47,32,0,0 ; vbroadcastss 0x202f(%rip),%xmm11 # 6994 <_sk_callback_avx+0x434> + DB 196,98,121,24,29,165,32,0,0 ; vbroadcastss 0x20a5(%rip),%xmm11 # 69f4 <_sk_callback_avx+0x43e> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,37,32,0,0 ; vbroadcastss 0x2025(%rip),%xmm13 # 6998 <_sk_callback_avx+0x438> + DB 196,98,121,24,45,155,32,0,0 ; vbroadcastss 0x209b(%rip),%xmm13 # 69f8 <_sk_callback_avx+0x442> DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -9478,7 +9461,7 @@ _sk_gather_f16_avx LABEL PROC DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 DB 197,252,86,194 ; vorps %ymm2,%ymm0,%ymm0 DB 196,227,125,25,194,1 ; vextractf128 $0x1,%ymm0,%xmm2 - DB 196,226,121,24,29,219,31,0,0 ; vbroadcastss 0x1fdb(%rip),%xmm3 # 699c <_sk_callback_avx+0x43c> + DB 196,226,121,24,29,81,32,0,0 ; vbroadcastss 0x2051(%rip),%xmm3 # 69fc <_sk_callback_avx+0x446> DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 @@ -9566,9 +9549,10 @@ _sk_gather_f16_avx LABEL PROC DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 DB 197,252,16,124,36,96 ; vmovups 0x60(%rsp),%ymm7 - DB 72,129,196,152,0,0,0 ; add $0x98,%rsp + DB 72,129,196,144,0,0,0 ; add $0x90,%rsp DB 91 ; pop %rbx DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -9580,12 +9564,12 @@ _sk_store_f16_avx LABEL PROC DB 197,252,17,180,36,128,0,0,0 ; vmovups %ymm6,0x80(%rsp) DB 197,252,17,108,36,96 ; vmovups %ymm5,0x60(%rsp) DB 197,252,17,100,36,64 ; vmovups %ymm4,0x40(%rsp) - DB 196,98,125,24,13,232,29,0,0 ; vbroadcastss 0x1de8(%rip),%ymm9 # 69a0 <_sk_callback_avx+0x440> + DB 196,98,125,24,13,92,30,0,0 ; vbroadcastss 0x1e5c(%rip),%ymm9 # 6a00 <_sk_callback_avx+0x44a> DB 196,65,124,84,209 ; vandps %ymm9,%ymm0,%ymm10 DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) DB 196,65,124,87,218 ; vxorps %ymm10,%ymm0,%ymm11 DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12 - DB 196,98,121,24,5,206,29,0,0 ; vbroadcastss 0x1dce(%rip),%xmm8 # 69a4 <_sk_callback_avx+0x444> + DB 196,98,121,24,5,66,30,0,0 ; vbroadcastss 0x1e42(%rip),%xmm8 # 6a04 <_sk_callback_avx+0x44e> DB 196,65,57,102,236 ; vpcmpgtd %xmm12,%xmm8,%xmm13 DB 196,65,57,102,243 ; vpcmpgtd %xmm11,%xmm8,%xmm14 DB 196,67,13,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 @@ -9595,7 +9579,7 @@ _sk_store_f16_avx LABEL PROC DB 196,67,13,24,242,1 ; vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11 DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12 - DB 196,98,125,24,21,149,29,0,0 ; vbroadcastss 0x1d95(%rip),%ymm10 # 69a8 <_sk_callback_avx+0x448> + DB 196,98,125,24,21,9,30,0,0 ; vbroadcastss 0x1e09(%rip),%ymm10 # 6a08 <_sk_callback_avx+0x452> DB 196,65,12,86,242 ; vorps %ymm10,%ymm14,%ymm14 DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15 DB 196,65,1,254,228 ; vpaddd %xmm12,%xmm15,%xmm12 @@ -9676,12 +9660,12 @@ _sk_store_f16_avx LABEL PROC DB 197,121,106,193 ; vpunpckhdq %xmm1,%xmm0,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 117,75 ; jne 4dfa <_sk_store_f16_avx+0x270> - DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8) - DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8) - DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8) - DB 197,122,127,68,248,48 ; vmovdqu %xmm8,0x30(%rax,%rdi,8) + DB 77,133,192 ; test %r8,%r8 + DB 117,75 ; jne 4de6 <_sk_store_f16_avx+0x270> + DB 197,120,17,28,208 ; vmovups %xmm11,(%rax,%rdx,8) + DB 197,120,17,84,208,16 ; vmovups %xmm10,0x10(%rax,%rdx,8) + DB 197,120,17,76,208,32 ; vmovups %xmm9,0x20(%rax,%rdx,8) + DB 197,122,127,68,208,48 ; vmovdqu %xmm8,0x30(%rax,%rdx,8) DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 DB 197,252,16,76,36,32 ; vmovups 0x20(%rsp),%ymm1 @@ -9691,36 +9675,36 @@ _sk_store_f16_avx LABEL PROC DB 197,252,16,188,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm7 DB 72,129,196,216,0,0,0 ; add $0xd8,%rsp DB 255,224 ; jmpq *%rax - DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,193 ; je 4dc6 <_sk_store_f16_avx+0x23c> - DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,181 ; jb 4dc6 <_sk_store_f16_avx+0x23c> - DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,173 ; je 4dc6 <_sk_store_f16_avx+0x23c> - DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8) - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,161 ; jb 4dc6 <_sk_store_f16_avx+0x23c> - DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,153 ; je 4dc6 <_sk_store_f16_avx+0x23c> - DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8) - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,141 ; jb 4dc6 <_sk_store_f16_avx+0x23c> - DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,133 ; jmp 4dc6 <_sk_store_f16_avx+0x23c> + DB 197,121,214,28,208 ; vmovq %xmm11,(%rax,%rdx,8) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,193 ; je 4db2 <_sk_store_f16_avx+0x23c> + DB 197,121,23,92,208,8 ; vmovhpd %xmm11,0x8(%rax,%rdx,8) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,181 ; jb 4db2 <_sk_store_f16_avx+0x23c> + DB 197,121,214,84,208,16 ; vmovq %xmm10,0x10(%rax,%rdx,8) + DB 116,173 ; je 4db2 <_sk_store_f16_avx+0x23c> + DB 197,121,23,84,208,24 ; vmovhpd %xmm10,0x18(%rax,%rdx,8) + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,161 ; jb 4db2 <_sk_store_f16_avx+0x23c> + DB 197,121,214,76,208,32 ; vmovq %xmm9,0x20(%rax,%rdx,8) + DB 116,153 ; je 4db2 <_sk_store_f16_avx+0x23c> + DB 197,121,23,76,208,40 ; vmovhpd %xmm9,0x28(%rax,%rdx,8) + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,141 ; jb 4db2 <_sk_store_f16_avx+0x23c> + DB 197,121,214,68,208,48 ; vmovq %xmm8,0x30(%rax,%rdx,8) + DB 235,133 ; jmp 4db2 <_sk_store_f16_avx+0x23c> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,253,0,0,0 ; jne 4f54 <_sk_load_u16_be_avx+0x113> - DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 - DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 - DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 - DB 196,65,122,111,76,64,48 ; vmovdqu 0x30(%r8,%rax,2),%xmm9 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,253,0,0,0 ; jne 4f40 <_sk_load_u16_be_avx+0x113> + DB 196,65,121,16,4,65 ; vmovupd (%r9,%rax,2),%xmm8 + DB 196,193,121,16,84,65,16 ; vmovupd 0x10(%r9,%rax,2),%xmm2 + DB 196,193,121,16,92,65,32 ; vmovupd 0x20(%r9,%rax,2),%xmm3 + DB 196,65,122,111,76,65,48 ; vmovdqu 0x30(%r9,%rax,2),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 @@ -9738,7 +9722,7 @@ _sk_load_u16_be_avx LABEL PROC DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,228,26,0,0 ; vbroadcastss 0x1ae4(%rip),%ymm11 # 69ac <_sk_callback_avx+0x44c> + DB 196,98,125,24,29,88,27,0,0 ; vbroadcastss 0x1b58(%rip),%ymm11 # 6a0c <_sk_callback_avx+0x456> DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,177,109,202 ; vpunpckhqdq %xmm2,%xmm9,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -9769,44 +9753,44 @@ _sk_load_u16_be_avx LABEL PROC DB 196,193,100,89,219 ; vmulps %ymm11,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 + DB 196,65,123,16,4,65 ; vmovsd (%r9,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 4fba <_sk_load_u16_be_avx+0x179> - DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 4fba <_sk_load_u16_be_avx+0x179> - DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 4fc7 <_sk_load_u16_be_avx+0x186> - DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 4fc7 <_sk_load_u16_be_avx+0x186> - DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,213,254,255,255 ; je 4e72 <_sk_load_u16_be_avx+0x31> - DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,196,254,255,255 ; jb 4e72 <_sk_load_u16_be_avx+0x31> - DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,184,254,255,255 ; jmpq 4e72 <_sk_load_u16_be_avx+0x31> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,85 ; je 4fa6 <_sk_load_u16_be_avx+0x179> + DB 196,65,57,22,68,65,8 ; vmovhpd 0x8(%r9,%rax,2),%xmm8,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,72 ; jb 4fa6 <_sk_load_u16_be_avx+0x179> + DB 196,193,123,16,84,65,16 ; vmovsd 0x10(%r9,%rax,2),%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 116,72 ; je 4fb3 <_sk_load_u16_be_avx+0x186> + DB 196,193,105,22,84,65,24 ; vmovhpd 0x18(%r9,%rax,2),%xmm2,%xmm2 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,59 ; jb 4fb3 <_sk_load_u16_be_avx+0x186> + DB 196,193,123,16,92,65,32 ; vmovsd 0x20(%r9,%rax,2),%xmm3 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 15,132,213,254,255,255 ; je 4e5e <_sk_load_u16_be_avx+0x31> + DB 196,193,97,22,92,65,40 ; vmovhpd 0x28(%r9,%rax,2),%xmm3,%xmm3 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 15,130,196,254,255,255 ; jb 4e5e <_sk_load_u16_be_avx+0x31> + DB 196,65,122,126,76,65,48 ; vmovq 0x30(%r9,%rax,2),%xmm9 + DB 233,184,254,255,255 ; jmpq 4e5e <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,171,254,255,255 ; jmpq 4e72 <_sk_load_u16_be_avx+0x31> + DB 233,171,254,255,255 ; jmpq 4e5e <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,162,254,255,255 ; jmpq 4e72 <_sk_load_u16_be_avx+0x31> + DB 233,162,254,255,255 ; jmpq 4e5e <_sk_load_u16_be_avx+0x31> PUBLIC _sk_load_rgb_u16_be_avx _sk_load_rgb_u16_be_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,243,0,0,0 ; jne 50d5 <_sk_load_rgb_u16_be_avx+0x105> - DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0 - DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2 - DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1 - DB 196,193,122,111,92,64,32 ; vmovdqu 0x20(%r8,%rax,2),%xmm3 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,82 ; lea (%rdx,%rdx,2),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,243,0,0,0 ; jne 50c1 <_sk_load_rgb_u16_be_avx+0x105> + DB 196,193,122,111,4,65 ; vmovdqu (%r9,%rax,2),%xmm0 + DB 196,193,122,111,84,65,12 ; vmovdqu 0xc(%r9,%rax,2),%xmm2 + DB 196,193,122,111,76,65,24 ; vmovdqu 0x18(%r9,%rax,2),%xmm1 + DB 196,193,122,111,92,65,32 ; vmovdqu 0x20(%r9,%rax,2),%xmm3 DB 197,225,115,219,4 ; vpsrldq $0x4,%xmm3,%xmm3 DB 197,185,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm8 DB 197,177,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm9 @@ -9829,7 +9813,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,68,25,0,0 ; vbroadcastss 0x1944(%rip),%ymm11 # 69b0 <_sk_callback_avx+0x450> + DB 196,98,125,24,29,184,25,0,0 ; vbroadcastss 0x19b8(%rip),%ymm11 # 6a10 <_sk_callback_avx+0x45a> DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -9850,48 +9834,48 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,225,24,0,0 ; vbroadcastss 0x18e1(%rip),%ymm3 # 69b4 <_sk_callback_avx+0x454> - DB 255,224 ; jmpq *%rax - DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 - DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 50ee <_sk_load_rgb_u16_be_avx+0x11e> - DB 233,40,255,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> - DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1 - DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 511d <_sk_load_rgb_u16_be_avx+0x14d> - DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1 - DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 5122 <_sk_load_rgb_u16_be_avx+0x152> - DB 233,249,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,244,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> - DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1 - DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 5151 <_sk_load_rgb_u16_be_avx+0x181> - DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1 - DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 5156 <_sk_load_rgb_u16_be_avx+0x186> - DB 233,197,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,192,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> - DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3 - DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 517f <_sk_load_rgb_u16_be_avx+0x1af> - DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3 - DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 233,151,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,146,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> + DB 196,226,125,24,29,85,25,0,0 ; vbroadcastss 0x1955(%rip),%ymm3 # 6a14 <_sk_callback_avx+0x45e> + DB 255,224 ; jmpq *%rax + DB 196,193,121,110,4,65 ; vmovd (%r9,%rax,2),%xmm0 + DB 196,193,121,196,68,65,4,2 ; vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,5 ; jne 50da <_sk_load_rgb_u16_be_avx+0x11e> + DB 233,40,255,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> + DB 196,193,121,110,76,65,6 ; vmovd 0x6(%r9,%rax,2),%xmm1 + DB 196,65,113,196,68,65,10,2 ; vpinsrw $0x2,0xa(%r9,%rax,2),%xmm1,%xmm8 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,26 ; jb 5109 <_sk_load_rgb_u16_be_avx+0x14d> + DB 196,193,121,110,76,65,12 ; vmovd 0xc(%r9,%rax,2),%xmm1 + DB 196,193,113,196,84,65,16,2 ; vpinsrw $0x2,0x10(%r9,%rax,2),%xmm1,%xmm2 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 117,10 ; jne 510e <_sk_load_rgb_u16_be_avx+0x152> + DB 233,249,254,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,244,254,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> + DB 196,193,121,110,76,65,18 ; vmovd 0x12(%r9,%rax,2),%xmm1 + DB 196,65,113,196,76,65,22,2 ; vpinsrw $0x2,0x16(%r9,%rax,2),%xmm1,%xmm9 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,26 ; jb 513d <_sk_load_rgb_u16_be_avx+0x181> + DB 196,193,121,110,76,65,24 ; vmovd 0x18(%r9,%rax,2),%xmm1 + DB 196,193,113,196,76,65,28,2 ; vpinsrw $0x2,0x1c(%r9,%rax,2),%xmm1,%xmm1 + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 117,10 ; jne 5142 <_sk_load_rgb_u16_be_avx+0x186> + DB 233,197,254,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,192,254,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> + DB 196,193,121,110,92,65,30 ; vmovd 0x1e(%r9,%rax,2),%xmm3 + DB 196,65,97,196,92,65,34,2 ; vpinsrw $0x2,0x22(%r9,%rax,2),%xmm3,%xmm11 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,20 ; jb 516b <_sk_load_rgb_u16_be_avx+0x1af> + DB 196,193,121,110,92,65,36 ; vmovd 0x24(%r9,%rax,2),%xmm3 + DB 196,193,97,196,92,65,40,2 ; vpinsrw $0x2,0x28(%r9,%rax,2),%xmm3,%xmm3 + DB 233,151,254,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,146,254,255,255 ; jmpq 5002 <_sk_load_rgb_u16_be_avx+0x46> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 196,98,125,24,5,30,24,0,0 ; vbroadcastss 0x181e(%rip),%ymm8 # 69b8 <_sk_callback_avx+0x458> + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 196,98,125,24,5,146,24,0,0 ; vbroadcastss 0x1892(%rip),%ymm8 # 6a18 <_sk_callback_avx+0x462> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10 @@ -9928,55 +9912,55 @@ _sk_store_u16_be_avx LABEL PROC DB 196,65,25,106,209 ; vpunpckhdq %xmm9,%xmm12,%xmm10 DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 527e <_sk_store_u16_be_avx+0xfa> - DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2) - DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2) - DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2) - DB 196,65,122,127,68,64,48 ; vmovdqu %xmm8,0x30(%r8,%rax,2) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 527a <_sk_store_u16_be_avx+0xf6> - DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 527a <_sk_store_u16_be_avx+0xf6> - DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je 527a <_sk_store_u16_be_avx+0xf6> - DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2) - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 527a <_sk_store_u16_be_avx+0xf6> - DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je 527a <_sk_store_u16_be_avx+0xf6> - DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2) - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 527a <_sk_store_u16_be_avx+0xf6> - DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp 527a <_sk_store_u16_be_avx+0xf6> + DB 77,133,192 ; test %r8,%r8 + DB 117,31 ; jne 526a <_sk_store_u16_be_avx+0xfa> + DB 196,65,120,17,28,65 ; vmovups %xmm11,(%r9,%rax,2) + DB 196,65,120,17,84,65,16 ; vmovups %xmm10,0x10(%r9,%rax,2) + DB 196,65,120,17,76,65,32 ; vmovups %xmm9,0x20(%r9,%rax,2) + DB 196,65,122,127,68,65,48 ; vmovdqu %xmm8,0x30(%r9,%rax,2) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 196,65,121,214,28,65 ; vmovq %xmm11,(%r9,%rax,2) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 5266 <_sk_store_u16_be_avx+0xf6> + DB 196,65,121,23,92,65,8 ; vmovhpd %xmm11,0x8(%r9,%rax,2) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 5266 <_sk_store_u16_be_avx+0xf6> + DB 196,65,121,214,84,65,16 ; vmovq %xmm10,0x10(%r9,%rax,2) + DB 116,218 ; je 5266 <_sk_store_u16_be_avx+0xf6> + DB 196,65,121,23,84,65,24 ; vmovhpd %xmm10,0x18(%r9,%rax,2) + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,205 ; jb 5266 <_sk_store_u16_be_avx+0xf6> + DB 196,65,121,214,76,65,32 ; vmovq %xmm9,0x20(%r9,%rax,2) + DB 116,196 ; je 5266 <_sk_store_u16_be_avx+0xf6> + DB 196,65,121,23,76,65,40 ; vmovhpd %xmm9,0x28(%r9,%rax,2) + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,183 ; jb 5266 <_sk_store_u16_be_avx+0xf6> + DB 196,65,121,214,68,65,48 ; vmovq %xmm8,0x30(%r9,%rax,2) + DB 235,174 ; jmp 5266 <_sk_store_u16_be_avx+0xf6> PUBLIC _sk_load_f32_avx _sk_load_f32_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 119,110 ; ja 5342 <_sk_load_f32_avx+0x76> - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 536c <_sk_load_f32_avx+0xa0> - DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax - DB 76,1,208 ; add %r10,%rax - DB 255,224 ; jmpq *%rax - DB 196,3,125,24,68,136,112,1 ; vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8 - DB 196,131,125,24,92,136,96,1 ; vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3 - DB 196,131,125,24,76,136,80,1 ; vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1 - DB 196,131,125,24,84,136,64,1 ; vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2 - DB 196,129,121,16,68,136,48 ; vmovupd 0x30(%r8,%r9,4),%xmm0 + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 119,110 ; ja 532e <_sk_load_f32_avx+0x76> + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 76,141,29,134,0,0,0 ; lea 0x86(%rip),%r11 # 5358 <_sk_load_f32_avx+0xa0> + DB 75,99,4,131 ; movslq (%r11,%r8,4),%rax + DB 76,1,216 ; add %r11,%rax + DB 255,224 ; jmpq *%rax + DB 196,3,125,24,68,145,112,1 ; vinsertf128 $0x1,0x70(%r9,%r10,4),%ymm0,%ymm8 + DB 196,131,125,24,92,145,96,1 ; vinsertf128 $0x1,0x60(%r9,%r10,4),%ymm0,%ymm3 + DB 196,131,125,24,76,145,80,1 ; vinsertf128 $0x1,0x50(%r9,%r10,4),%ymm0,%ymm1 + DB 196,131,125,24,84,145,64,1 ; vinsertf128 $0x1,0x40(%r9,%r10,4),%ymm0,%ymm2 + DB 196,129,121,16,68,145,48 ; vmovupd 0x30(%r9,%r10,4),%xmm0 DB 196,195,125,13,192,12 ; vblendpd $0xc,%ymm8,%ymm0,%ymm0 - DB 196,1,121,16,68,136,32 ; vmovupd 0x20(%r8,%r9,4),%xmm8 + DB 196,1,121,16,68,145,32 ; vmovupd 0x20(%r9,%r10,4),%xmm8 DB 196,99,61,13,203,12 ; vblendpd $0xc,%ymm3,%ymm8,%ymm9 - DB 196,129,121,16,92,136,16 ; vmovupd 0x10(%r8,%r9,4),%xmm3 + DB 196,129,121,16,92,145,16 ; vmovupd 0x10(%r9,%r10,4),%xmm3 DB 196,99,101,13,209,12 ; vblendpd $0xc,%ymm1,%ymm3,%ymm10 - DB 196,129,121,16,12,136 ; vmovupd (%r8,%r9,4),%xmm1 + DB 196,129,121,16,12,145 ; vmovupd (%r9,%r10,4),%xmm1 DB 196,227,117,13,202,12 ; vblendpd $0xc,%ymm2,%ymm1,%ymm1 DB 196,193,116,20,210 ; vunpcklps %ymm10,%ymm1,%ymm2 DB 196,193,116,21,218 ; vunpckhps %ymm10,%ymm1,%ymm3 @@ -10008,8 +9992,8 @@ _sk_load_f32_avx LABEL PROC PUBLIC _sk_store_f32_avx _sk_store_f32_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax DB 197,124,20,193 ; vunpcklps %ymm1,%ymm0,%ymm8 DB 197,124,21,217 ; vunpckhps %ymm1,%ymm0,%ymm11 DB 197,108,20,203 ; vunpcklps %ymm3,%ymm2,%ymm9 @@ -10018,36 +10002,36 @@ _sk_store_f32_avx LABEL PROC DB 196,65,61,21,201 ; vunpckhpd %ymm9,%ymm8,%ymm9 DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 - DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 53f9 <_sk_store_f32_avx+0x6d> + DB 77,133,192 ; test %r8,%r8 + DB 117,55 ; jne 53e5 <_sk_store_f32_avx+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 DB 196,67,61,6,195,49 ; vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 - DB 196,65,125,17,36,128 ; vmovupd %ymm12,(%r8,%rax,4) - DB 196,65,125,17,108,128,32 ; vmovupd %ymm13,0x20(%r8,%rax,4) - DB 196,65,125,17,76,128,64 ; vmovupd %ymm9,0x40(%r8,%rax,4) - DB 196,65,125,17,68,128,96 ; vmovupd %ymm8,0x60(%r8,%rax,4) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 53f5 <_sk_store_f32_avx+0x69> - DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 53f5 <_sk_store_f32_avx+0x69> - DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 53f5 <_sk_store_f32_avx+0x69> - DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) - DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 53f5 <_sk_store_f32_avx+0x69> - DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 53f5 <_sk_store_f32_avx+0x69> - DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) - DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 53f5 <_sk_store_f32_avx+0x69> - DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 53f5 <_sk_store_f32_avx+0x69> + DB 196,65,125,17,36,129 ; vmovupd %ymm12,(%r9,%rax,4) + DB 196,65,125,17,108,129,32 ; vmovupd %ymm13,0x20(%r9,%rax,4) + DB 196,65,125,17,76,129,64 ; vmovupd %ymm9,0x40(%r9,%rax,4) + DB 196,65,125,17,68,129,96 ; vmovupd %ymm8,0x60(%r9,%rax,4) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 196,65,121,17,20,129 ; vmovupd %xmm10,(%r9,%rax,4) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 53e1 <_sk_store_f32_avx+0x69> + DB 196,65,121,17,76,129,16 ; vmovupd %xmm9,0x10(%r9,%rax,4) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 53e1 <_sk_store_f32_avx+0x69> + DB 196,65,121,17,68,129,32 ; vmovupd %xmm8,0x20(%r9,%rax,4) + DB 116,218 ; je 53e1 <_sk_store_f32_avx+0x69> + DB 196,65,121,17,92,129,48 ; vmovupd %xmm11,0x30(%r9,%rax,4) + DB 73,131,248,5 ; cmp $0x5,%r8 + DB 114,205 ; jb 53e1 <_sk_store_f32_avx+0x69> + DB 196,67,125,25,84,129,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r9,%rax,4) + DB 116,195 ; je 53e1 <_sk_store_f32_avx+0x69> + DB 196,67,125,25,76,129,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r9,%rax,4) + DB 73,131,248,7 ; cmp $0x7,%r8 + DB 114,181 ; jb 53e1 <_sk_store_f32_avx+0x69> + DB 196,67,125,25,68,129,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r9,%rax,4) + DB 235,171 ; jmp 53e1 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -10137,7 +10121,7 @@ PUBLIC _sk_clamp_x_1_avx _sk_clamp_x_1_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,78,20,0,0 ; vbroadcastss 0x144e(%rip),%ymm8 # 69bc <_sk_callback_avx+0x45c> + DB 196,98,125,24,5,194,20,0,0 ; vbroadcastss 0x14c2(%rip),%ymm8 # 6a1c <_sk_callback_avx+0x466> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10151,9 +10135,9 @@ _sk_repeat_x_1_avx LABEL PROC PUBLIC _sk_mirror_x_1_avx _sk_mirror_x_1_avx LABEL PROC - DB 196,98,125,24,5,49,20,0,0 ; vbroadcastss 0x1431(%rip),%ymm8 # 69c0 <_sk_callback_avx+0x460> + DB 196,98,125,24,5,165,20,0,0 ; vbroadcastss 0x14a5(%rip),%ymm8 # 6a20 <_sk_callback_avx+0x46a> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,39,20,0,0 ; vbroadcastss 0x1427(%rip),%ymm9 # 69c4 <_sk_callback_avx+0x464> + DB 196,98,125,24,13,155,20,0,0 ; vbroadcastss 0x149b(%rip),%ymm9 # 6a24 <_sk_callback_avx+0x46e> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -10167,12 +10151,12 @@ _sk_mirror_x_1_avx LABEL PROC PUBLIC _sk_luminance_to_alpha_avx _sk_luminance_to_alpha_avx LABEL PROC - DB 196,226,125,24,29,247,19,0,0 ; vbroadcastss 0x13f7(%rip),%ymm3 # 69c8 <_sk_callback_avx+0x468> + DB 196,226,125,24,29,107,20,0,0 ; vbroadcastss 0x146b(%rip),%ymm3 # 6a28 <_sk_callback_avx+0x472> DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,29,238,19,0,0 ; vbroadcastss 0x13ee(%rip),%ymm3 # 69cc <_sk_callback_avx+0x46c> + DB 196,226,125,24,29,98,20,0,0 ; vbroadcastss 0x1462(%rip),%ymm3 # 6a2c <_sk_callback_avx+0x476> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,225,19,0,0 ; vbroadcastss 0x13e1(%rip),%ymm1 # 69d0 <_sk_callback_avx+0x470> + DB 196,226,125,24,13,85,20,0,0 ; vbroadcastss 0x1455(%rip),%ymm1 # 6a30 <_sk_callback_avx+0x47a> DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 DB 197,252,88,217 ; vaddps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -10379,141 +10363,141 @@ _sk_evenly_spaced_gradient_avx LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 72,139,104,8 ; mov 0x8(%rax),%rbp DB 72,255,203 ; dec %rbx - DB 120,7 ; js 5945 <_sk_evenly_spaced_gradient_avx+0x1f> + DB 120,7 ; js 5931 <_sk_evenly_spaced_gradient_avx+0x1f> DB 196,225,242,42,203 ; vcvtsi2ss %rbx,%xmm1,%xmm1 - DB 235,21 ; jmp 595a <_sk_evenly_spaced_gradient_avx+0x34> - DB 73,137,216 ; mov %rbx,%r8 - DB 73,209,232 ; shr %r8 + DB 235,21 ; jmp 5946 <_sk_evenly_spaced_gradient_avx+0x34> + DB 73,137,217 ; mov %rbx,%r9 + DB 73,209,233 ; shr %r9 DB 131,227,1 ; and $0x1,%ebx - DB 76,9,195 ; or %r8,%rbx + DB 76,9,203 ; or %r9,%rbx DB 196,225,242,42,203 ; vcvtsi2ss %rbx,%xmm1,%xmm1 DB 197,242,88,201 ; vaddss %xmm1,%xmm1,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 DB 197,244,89,200 ; vmulps %ymm0,%ymm1,%ymm1 DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 - DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,195,249,22,201,1 ; vpextrq $0x1,%xmm1,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,193,249,126,203 ; vmovq %xmm1,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 - DB 196,195,249,22,207,1 ; vpextrq $0x1,%xmm1,%r15 - DB 69,137,254 ; mov %r15d,%r14d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,193,249,126,205 ; vmovq %xmm1,%r13 - DB 69,137,236 ; mov %r13d,%r12d - DB 73,193,237,32 ; shr $0x20,%r13 - DB 196,161,122,16,76,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm1 - DB 196,163,113,33,76,173,0,16 ; vinsertps $0x10,0x0(%rbp,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 196,195,249,22,204,1 ; vpextrq $0x1,%xmm1,%r12 + DB 69,137,231 ; mov %r12d,%r15d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,221 ; mov %ebx,%r13d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,84,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm2 - DB 196,163,105,33,84,149,0,16 ; vinsertps $0x10,0x0(%rbp,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,92,133,0 ; vmovss 0x0(%rbp,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,99,109,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm8 - DB 72,139,88,40 ; mov 0x28(%rax),%rbx - DB 196,161,122,16,20,163 ; vmovss (%rbx,%r12,4),%xmm2 - DB 196,163,105,33,20,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm2,%xmm2 - DB 196,161,122,16,28,179 ; vmovss (%rbx,%r14,4),%xmm3 + DB 72,139,104,40 ; mov 0x28(%rax),%rbp + DB 196,161,122,16,84,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm2 + DB 196,227,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,92,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,187 ; vmovss (%rbx,%r15,4),%xmm3 + DB 196,161,122,16,92,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,155 ; vmovss (%rbx,%r11,4),%xmm3 - DB 196,163,97,33,28,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,12,139 ; vmovss (%rbx,%r9,4),%xmm1 + DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3 + DB 196,163,97,33,92,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm3,%xmm3 + DB 196,161,122,16,76,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm1 DB 196,227,97,33,201,32 ; vinsertps $0x20,%xmm1,%xmm3,%xmm1 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 DB 196,99,117,24,226,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm12 - DB 72,139,88,16 ; mov 0x10(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,28,179 ; vmovss (%rbx,%r14,4),%xmm3 + DB 72,139,104,16 ; mov 0x10(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,92,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm3 DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 - DB 196,161,122,16,28,187 ; vmovss (%rbx,%r15,4),%xmm3 + DB 196,161,122,16,92,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm3 DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 - DB 196,161,122,16,28,155 ; vmovss (%rbx,%r11,4),%xmm3 - DB 196,163,97,33,28,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,20,139 ; vmovss (%rbx,%r9,4),%xmm2 + DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3 + DB 196,163,97,33,92,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm3,%xmm3 + DB 196,161,122,16,84,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm2 DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,99,109,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm13 - DB 72,139,88,48 ; mov 0x30(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,48 ; mov 0x30(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,99,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm9 - DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,24 ; mov 0x18(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,99,109,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm10 - DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,56 ; mov 0x38(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,99,109,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm11 - DB 72,139,88,32 ; mov 0x20(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,32 ; mov 0x20(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 + DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,99,109,24,241,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm14 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 196,161,122,16,12,160 ; vmovss (%rax,%r12,4),%xmm1 - DB 196,163,113,33,12,168,16 ; vinsertps $0x10,(%rax,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 196,161,122,16,12,168 ; vmovss (%rax,%r13,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 DB 196,161,122,16,20,184 ; vmovss (%rax,%r15,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 196,161,122,16,20,160 ; vmovss (%rax,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 + DB 196,163,105,33,20,152,16 ; vinsertps $0x10,(%rax,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,227,109,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm3 DB 197,188,89,200 ; vmulps %ymm0,%ymm8,%ymm1 @@ -10536,18 +10520,18 @@ _sk_evenly_spaced_gradient_avx LABEL PROC PUBLIC _sk_gauss_a_to_rgba_avx _sk_gauss_a_to_rgba_avx LABEL PROC - DB 196,226,125,24,5,76,13,0,0 ; vbroadcastss 0xd4c(%rip),%ymm0 # 69d4 <_sk_callback_avx+0x474> + DB 196,226,125,24,5,144,13,0,0 ; vbroadcastss 0xd90(%rip),%ymm0 # 6a34 <_sk_callback_avx+0x47e> DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 196,226,125,24,13,67,13,0,0 ; vbroadcastss 0xd43(%rip),%ymm1 # 69d8 <_sk_callback_avx+0x478> + DB 196,226,125,24,13,135,13,0,0 ; vbroadcastss 0xd87(%rip),%ymm1 # 6a38 <_sk_callback_avx+0x482> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,54,13,0,0 ; vbroadcastss 0xd36(%rip),%ymm1 # 69dc <_sk_callback_avx+0x47c> + DB 196,226,125,24,13,122,13,0,0 ; vbroadcastss 0xd7a(%rip),%ymm1 # 6a3c <_sk_callback_avx+0x486> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,41,13,0,0 ; vbroadcastss 0xd29(%rip),%ymm1 # 69e0 <_sk_callback_avx+0x480> + DB 196,226,125,24,13,109,13,0,0 ; vbroadcastss 0xd6d(%rip),%ymm1 # 6a40 <_sk_callback_avx+0x48a> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,28,13,0,0 ; vbroadcastss 0xd1c(%rip),%ymm1 # 69e4 <_sk_callback_avx+0x484> + DB 196,226,125,24,13,96,13,0,0 ; vbroadcastss 0xd60(%rip),%ymm1 # 6a44 <_sk_callback_avx+0x48e> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -10563,16 +10547,18 @@ _sk_gradient_avx LABEL PROC DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx + DB 72,131,236,40 ; sub $0x28,%rsp + DB 197,252,17,60,36 ; vmovups %ymm7,(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 - DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,80 ; jb 5d45 <_sk_gradient_avx+0x69> + DB 73,131,249,2 ; cmp $0x2,%r9 + DB 114,80 ; jb 5d6a <_sk_gradient_avx+0x72> DB 72,139,88,72 ; mov 0x48(%rax),%rbx - DB 73,255,200 ; dec %r8 + DB 73,255,201 ; dec %r9 DB 72,131,195,4 ; add $0x4,%rbx DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 - DB 196,98,125,24,21,218,12,0,0 ; vbroadcastss 0xcda(%rip),%ymm10 # 69e8 <_sk_callback_avx+0x488> + DB 196,98,125,24,21,21,13,0,0 ; vbroadcastss 0xd15(%rip),%ymm10 # 6a48 <_sk_callback_avx+0x492> DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,98,125,24,3 ; vbroadcastss (%rbx),%ymm8 DB 197,60,194,192,2 ; vcmpleps %ymm0,%ymm8,%ymm8 @@ -10583,132 +10569,132 @@ _sk_gradient_avx LABEL PROC DB 197,185,254,201 ; vpaddd %xmm1,%xmm8,%xmm1 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 72,131,195,4 ; add $0x4,%rbx - DB 73,255,200 ; dec %r8 - DB 117,205 ; jne 5d12 <_sk_gradient_avx+0x36> - DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 73,255,201 ; dec %r9 + DB 117,205 ; jne 5d37 <_sk_gradient_avx+0x3f> + DB 196,195,249,22,201,1 ; vpextrq $0x1,%xmm1,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 196,193,249,126,203 ; vmovq %xmm1,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 - DB 196,195,249,22,207,1 ; vpextrq $0x1,%xmm1,%r15 - DB 69,137,254 ; mov %r15d,%r14d - DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,193,249,126,205 ; vmovq %xmm1,%r13 - DB 69,137,236 ; mov %r13d,%r12d - DB 73,193,237,32 ; shr $0x20,%r13 + DB 196,195,249,22,204,1 ; vpextrq $0x1,%xmm1,%r12 + DB 69,137,231 ; mov %r12d,%r15d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,221 ; mov %ebx,%r13d + DB 72,193,235,32 ; shr $0x20,%rbx DB 72,139,104,8 ; mov 0x8(%rax),%rbp - DB 72,139,88,16 ; mov 0x10(%rax),%rbx - DB 196,161,122,16,76,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm1 - DB 196,163,113,33,76,173,0,16 ; vinsertps $0x10,0x0(%rbp,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,84,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm2 - DB 196,163,105,33,84,149,0,16 ; vinsertps $0x10,0x0(%rbp,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,92,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,92,133,0 ; vmovss 0x0(%rbp,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,99,109,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm8 + DB 196,99,113,33,194,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm8 + DB 196,33,122,16,76,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm9 + DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3 + DB 196,35,97,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm3,%xmm10 + DB 196,33,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm11 + DB 196,33,122,16,100,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm12 DB 72,139,104,40 ; mov 0x28(%rax),%rbp - DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 - DB 196,163,105,33,84,173,0,16 ; vinsertps $0x10,0x0(%rbp,%r13,4),%xmm2,%xmm2 + DB 196,161,122,16,92,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm3 + DB 196,99,97,33,108,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm3,%xmm13 + DB 196,33,122,16,116,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm14 + DB 196,33,122,16,124,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm15 DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 196,163,97,33,124,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm3,%xmm7 + DB 196,161,122,16,76,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm1 + DB 196,161,122,16,84,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm2 + DB 72,139,104,16 ; mov 0x10(%rax),%rbp + DB 196,67,57,33,193,48 ; vinsertps $0x30,%xmm9,%xmm8,%xmm8 + DB 196,195,41,33,219,32 ; vinsertps $0x20,%xmm11,%xmm10,%xmm3 + DB 196,195,97,33,220,48 ; vinsertps $0x30,%xmm12,%xmm3,%xmm3 + DB 196,67,101,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm8 + DB 196,195,17,33,222,32 ; vinsertps $0x20,%xmm14,%xmm13,%xmm3 + DB 196,195,97,33,223,48 ; vinsertps $0x30,%xmm15,%xmm3,%xmm3 + DB 196,227,65,33,201,32 ; vinsertps $0x20,%xmm1,%xmm7,%xmm1 + DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 + DB 196,99,117,24,227,1 ; vinsertf128 $0x1,%xmm3,%ymm1,%ymm12 + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 DB 196,161,122,16,92,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,92,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm3 - DB 196,163,97,33,92,149,0,16 ; vinsertps $0x10,0x0(%rbp,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,76,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm1 - DB 196,227,97,33,201,32 ; vinsertps $0x20,%xmm1,%xmm3,%xmm1 - DB 196,161,122,16,92,133,0 ; vmovss 0x0(%rbp,%r8,4),%xmm3 - DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 - DB 196,99,117,24,226,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm12 - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,28,179 ; vmovss (%rbx,%r14,4),%xmm3 DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 - DB 196,161,122,16,28,187 ; vmovss (%rbx,%r15,4),%xmm3 + DB 196,161,122,16,92,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm3 DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 - DB 196,161,122,16,28,155 ; vmovss (%rbx,%r11,4),%xmm3 - DB 196,163,97,33,28,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,20,139 ; vmovss (%rbx,%r9,4),%xmm2 - DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,99,109,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm13 - DB 72,139,88,48 ; mov 0x30(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3 + DB 196,163,97,33,92,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm3,%xmm3 + DB 196,161,122,16,124,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm7 + DB 196,227,97,33,223,32 ; vinsertps $0x20,%xmm7,%xmm3,%xmm3 + DB 196,161,122,16,124,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm7 + DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 + DB 196,99,101,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm13 + DB 72,139,104,48 ; mov 0x30(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,124,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm7 + DB 196,227,113,33,207,32 ; vinsertps $0x20,%xmm7,%xmm1,%xmm1 + DB 196,161,122,16,124,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm7 + DB 196,227,113,33,207,48 ; vinsertps $0x30,%xmm7,%xmm1,%xmm1 + DB 196,161,122,16,124,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm7 + DB 196,163,65,33,124,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm7,%xmm7 + DB 196,161,122,16,84,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm2 + DB 196,227,65,33,210,32 ; vinsertps $0x20,%xmm2,%xmm7,%xmm2 + DB 196,161,122,16,124,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm7 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 DB 196,99,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm9 - DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,24 ; mov 0x18(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,124,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm7 + DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 + DB 196,161,122,16,124,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm7 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 DB 196,99,109,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm10 - DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,56 ; mov 0x38(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,161,122,16,84,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm2 + DB 196,163,105,33,84,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm2,%xmm2 + DB 196,161,122,16,124,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm7 + DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 + DB 196,161,122,16,124,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm7 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 DB 196,99,109,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm11 - DB 72,139,88,32 ; mov 0x20(%rax),%rbx - DB 196,161,122,16,12,163 ; vmovss (%rbx,%r12,4),%xmm1 - DB 196,163,113,33,12,171,16 ; vinsertps $0x10,(%rbx,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,179 ; vmovss (%rbx,%r14,4),%xmm2 + DB 72,139,104,32 ; mov 0x20(%rax),%rbp + DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1 + DB 196,227,113,33,76,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,187 ; vmovss (%rbx,%r15,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,155 ; vmovss (%rbx,%r11,4),%xmm2 - DB 196,163,105,33,20,147,16 ; vinsertps $0x10,(%rbx,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,139 ; vmovss (%rbx,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,131 ; vmovss (%rbx,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,99,109,24,241,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm14 + DB 196,161,122,16,84,165,0 ; vmovss 0x0(%rbp,%r12,4),%xmm2 + DB 196,161,122,16,124,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm7 + DB 196,35,65,33,116,157,0,16 ; vinsertps $0x10,0x0(%rbp,%r11,4),%xmm7,%xmm14 + DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3 + DB 196,161,122,16,124,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm7 + DB 196,99,113,33,250,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm15 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 196,161,122,16,12,160 ; vmovss (%rax,%r12,4),%xmm1 - DB 196,163,113,33,12,168,16 ; vinsertps $0x10,(%rax,%r13,4),%xmm1,%xmm1 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,184 ; vmovss (%rax,%r15,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,161,122,16,20,168 ; vmovss (%rax,%r13,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,227,9,33,219,32 ; vinsertps $0x20,%xmm3,%xmm14,%xmm3 + DB 196,161,122,16,12,184 ; vmovss (%rax,%r15,4),%xmm1 + DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 + DB 196,161,122,16,60,160 ; vmovss (%rax,%r12,4),%xmm7 + DB 196,67,101,24,247,1 ; vinsertf128 $0x1,%xmm15,%ymm3,%ymm14 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,163,97,33,28,152,16 ; vinsertps $0x10,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,105,33,201,32 ; vinsertps $0x20,%xmm1,%xmm2,%xmm1 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,113,33,207,48 ; vinsertps $0x30,%xmm7,%xmm1,%xmm1 + DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 + DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 DB 196,227,109,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm3 DB 197,188,89,200 ; vmulps %ymm0,%ymm8,%ymm1 DB 196,65,116,88,196 ; vaddps %ymm12,%ymm1,%ymm8 @@ -10720,6 +10706,8 @@ _sk_gradient_avx LABEL PROC DB 197,252,88,219 ; vaddps %ymm3,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,124,41,192 ; vmovaps %ymm8,%ymm0 + DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7 + DB 72,131,196,40 ; add $0x28,%rsp DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,93 ; pop %r13 @@ -10762,27 +10750,27 @@ _sk_xy_to_unit_angle_avx LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,254,8,0,0 ; vbroadcastss 0x8fe(%rip),%ymm13 # 69ec <_sk_callback_avx+0x48c> + DB 196,98,125,24,45,8,9,0,0 ; vbroadcastss 0x908(%rip),%ymm13 # 6a4c <_sk_callback_avx+0x496> DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,244,8,0,0 ; vbroadcastss 0x8f4(%rip),%ymm14 # 69f0 <_sk_callback_avx+0x490> + DB 196,98,125,24,53,254,8,0,0 ; vbroadcastss 0x8fe(%rip),%ymm14 # 6a50 <_sk_callback_avx+0x49a> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,229,8,0,0 ; vbroadcastss 0x8e5(%rip),%ymm14 # 69f4 <_sk_callback_avx+0x494> + DB 196,98,125,24,53,239,8,0,0 ; vbroadcastss 0x8ef(%rip),%ymm14 # 6a54 <_sk_callback_avx+0x49e> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,229 ; vmulps %ymm13,%ymm12,%ymm12 - DB 196,98,125,24,45,214,8,0,0 ; vbroadcastss 0x8d6(%rip),%ymm13 # 69f8 <_sk_callback_avx+0x498> + DB 196,98,125,24,45,224,8,0,0 ; vbroadcastss 0x8e0(%rip),%ymm13 # 6a58 <_sk_callback_avx+0x4a2> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,193,8,0,0 ; vbroadcastss 0x8c1(%rip),%ymm10 # 69fc <_sk_callback_avx+0x49c> + DB 196,98,125,24,21,203,8,0,0 ; vbroadcastss 0x8cb(%rip),%ymm10 # 6a5c <_sk_callback_avx+0x4a6> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,171,8,0,0 ; vbroadcastss 0x8ab(%rip),%ymm10 # 6a00 <_sk_callback_avx+0x4a0> + DB 196,98,125,24,21,181,8,0,0 ; vbroadcastss 0x8b5(%rip),%ymm10 # 6a60 <_sk_callback_avx+0x4aa> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,149,8,0,0 ; vbroadcastss 0x895(%rip),%ymm10 # 6a04 <_sk_callback_avx+0x4a4> + DB 196,98,125,24,21,159,8,0,0 ; vbroadcastss 0x89f(%rip),%ymm10 # 6a64 <_sk_callback_avx+0x4ae> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -10802,7 +10790,7 @@ _sk_xy_to_radius_avx LABEL PROC PUBLIC _sk_save_xy_avx _sk_save_xy_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,95,8,0,0 ; vbroadcastss 0x85f(%rip),%ymm8 # 6a08 <_sk_callback_avx+0x4a8> + DB 196,98,125,24,5,105,8,0,0 ; vbroadcastss 0x869(%rip),%ymm8 # 6a68 <_sk_callback_avx+0x4b2> DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9 DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 @@ -10835,9 +10823,9 @@ _sk_accumulate_avx LABEL PROC PUBLIC _sk_bilinear_nx_avx _sk_bilinear_nx_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,235,7,0,0 ; vbroadcastss 0x7eb(%rip),%ymm0 # 6a0c <_sk_callback_avx+0x4ac> + DB 196,226,125,24,5,245,7,0,0 ; vbroadcastss 0x7f5(%rip),%ymm0 # 6a6c <_sk_callback_avx+0x4b6> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,226,7,0,0 ; vbroadcastss 0x7e2(%rip),%ymm8 # 6a10 <_sk_callback_avx+0x4b0> + DB 196,98,125,24,5,236,7,0,0 ; vbroadcastss 0x7ec(%rip),%ymm8 # 6a70 <_sk_callback_avx+0x4ba> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -10846,7 +10834,7 @@ _sk_bilinear_nx_avx LABEL PROC PUBLIC _sk_bilinear_px_avx _sk_bilinear_px_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,202,7,0,0 ; vbroadcastss 0x7ca(%rip),%ymm0 # 6a14 <_sk_callback_avx+0x4b4> + DB 196,226,125,24,5,212,7,0,0 ; vbroadcastss 0x7d4(%rip),%ymm0 # 6a74 <_sk_callback_avx+0x4be> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -10856,9 +10844,9 @@ _sk_bilinear_px_avx LABEL PROC PUBLIC _sk_bilinear_ny_avx _sk_bilinear_ny_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,174,7,0,0 ; vbroadcastss 0x7ae(%rip),%ymm1 # 6a18 <_sk_callback_avx+0x4b8> + DB 196,226,125,24,13,184,7,0,0 ; vbroadcastss 0x7b8(%rip),%ymm1 # 6a78 <_sk_callback_avx+0x4c2> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,164,7,0,0 ; vbroadcastss 0x7a4(%rip),%ymm8 # 6a1c <_sk_callback_avx+0x4bc> + DB 196,98,125,24,5,174,7,0,0 ; vbroadcastss 0x7ae(%rip),%ymm8 # 6a7c <_sk_callback_avx+0x4c6> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -10867,7 +10855,7 @@ _sk_bilinear_ny_avx LABEL PROC PUBLIC _sk_bilinear_py_avx _sk_bilinear_py_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,140,7,0,0 ; vbroadcastss 0x78c(%rip),%ymm1 # 6a20 <_sk_callback_avx+0x4c0> + DB 196,226,125,24,13,150,7,0,0 ; vbroadcastss 0x796(%rip),%ymm1 # 6a80 <_sk_callback_avx+0x4ca> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -10877,14 +10865,14 @@ _sk_bilinear_py_avx LABEL PROC PUBLIC _sk_bicubic_n3x_avx _sk_bicubic_n3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,111,7,0,0 ; vbroadcastss 0x76f(%rip),%ymm0 # 6a24 <_sk_callback_avx+0x4c4> + DB 196,226,125,24,5,121,7,0,0 ; vbroadcastss 0x779(%rip),%ymm0 # 6a84 <_sk_callback_avx+0x4ce> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,102,7,0,0 ; vbroadcastss 0x766(%rip),%ymm8 # 6a28 <_sk_callback_avx+0x4c8> + DB 196,98,125,24,5,112,7,0,0 ; vbroadcastss 0x770(%rip),%ymm8 # 6a88 <_sk_callback_avx+0x4d2> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,87,7,0,0 ; vbroadcastss 0x757(%rip),%ymm10 # 6a2c <_sk_callback_avx+0x4cc> + DB 196,98,125,24,21,97,7,0,0 ; vbroadcastss 0x761(%rip),%ymm10 # 6a8c <_sk_callback_avx+0x4d6> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,77,7,0,0 ; vbroadcastss 0x74d(%rip),%ymm10 # 6a30 <_sk_callback_avx+0x4d0> + DB 196,98,125,24,21,87,7,0,0 ; vbroadcastss 0x757(%rip),%ymm10 # 6a90 <_sk_callback_avx+0x4da> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -10894,19 +10882,19 @@ _sk_bicubic_n3x_avx LABEL PROC PUBLIC _sk_bicubic_n1x_avx _sk_bicubic_n1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,48,7,0,0 ; vbroadcastss 0x730(%rip),%ymm0 # 6a34 <_sk_callback_avx+0x4d4> + DB 196,226,125,24,5,58,7,0,0 ; vbroadcastss 0x73a(%rip),%ymm0 # 6a94 <_sk_callback_avx+0x4de> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,39,7,0,0 ; vbroadcastss 0x727(%rip),%ymm8 # 6a38 <_sk_callback_avx+0x4d8> + DB 196,98,125,24,5,49,7,0,0 ; vbroadcastss 0x731(%rip),%ymm8 # 6a98 <_sk_callback_avx+0x4e2> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,29,7,0,0 ; vbroadcastss 0x71d(%rip),%ymm9 # 6a3c <_sk_callback_avx+0x4dc> + DB 196,98,125,24,13,39,7,0,0 ; vbroadcastss 0x727(%rip),%ymm9 # 6a9c <_sk_callback_avx+0x4e6> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,19,7,0,0 ; vbroadcastss 0x713(%rip),%ymm10 # 6a40 <_sk_callback_avx+0x4e0> + DB 196,98,125,24,21,29,7,0,0 ; vbroadcastss 0x71d(%rip),%ymm10 # 6aa0 <_sk_callback_avx+0x4ea> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,4,7,0,0 ; vbroadcastss 0x704(%rip),%ymm10 # 6a44 <_sk_callback_avx+0x4e4> + DB 196,98,125,24,21,14,7,0,0 ; vbroadcastss 0x70e(%rip),%ymm10 # 6aa4 <_sk_callback_avx+0x4ee> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,245,6,0,0 ; vbroadcastss 0x6f5(%rip),%ymm9 # 6a48 <_sk_callback_avx+0x4e8> + DB 196,98,125,24,13,255,6,0,0 ; vbroadcastss 0x6ff(%rip),%ymm9 # 6aa8 <_sk_callback_avx+0x4f2> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -10915,17 +10903,17 @@ _sk_bicubic_n1x_avx LABEL PROC PUBLIC _sk_bicubic_p1x_avx _sk_bicubic_p1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,221,6,0,0 ; vbroadcastss 0x6dd(%rip),%ymm8 # 6a4c <_sk_callback_avx+0x4ec> + DB 196,98,125,24,5,231,6,0,0 ; vbroadcastss 0x6e7(%rip),%ymm8 # 6aac <_sk_callback_avx+0x4f6> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,207,6,0,0 ; vbroadcastss 0x6cf(%rip),%ymm10 # 6a50 <_sk_callback_avx+0x4f0> + DB 196,98,125,24,21,217,6,0,0 ; vbroadcastss 0x6d9(%rip),%ymm10 # 6ab0 <_sk_callback_avx+0x4fa> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,197,6,0,0 ; vbroadcastss 0x6c5(%rip),%ymm11 # 6a54 <_sk_callback_avx+0x4f4> + DB 196,98,125,24,29,207,6,0,0 ; vbroadcastss 0x6cf(%rip),%ymm11 # 6ab4 <_sk_callback_avx+0x4fe> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,172,6,0,0 ; vbroadcastss 0x6ac(%rip),%ymm9 # 6a58 <_sk_callback_avx+0x4f8> + DB 196,98,125,24,13,182,6,0,0 ; vbroadcastss 0x6b6(%rip),%ymm9 # 6ab8 <_sk_callback_avx+0x502> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -10934,13 +10922,13 @@ _sk_bicubic_p1x_avx LABEL PROC PUBLIC _sk_bicubic_p3x_avx _sk_bicubic_p3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,148,6,0,0 ; vbroadcastss 0x694(%rip),%ymm0 # 6a5c <_sk_callback_avx+0x4fc> + DB 196,226,125,24,5,158,6,0,0 ; vbroadcastss 0x69e(%rip),%ymm0 # 6abc <_sk_callback_avx+0x506> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm10 # 6a60 <_sk_callback_avx+0x500> + DB 196,98,125,24,21,139,6,0,0 ; vbroadcastss 0x68b(%rip),%ymm10 # 6ac0 <_sk_callback_avx+0x50a> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,119,6,0,0 ; vbroadcastss 0x677(%rip),%ymm10 # 6a64 <_sk_callback_avx+0x504> + DB 196,98,125,24,21,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm10 # 6ac4 <_sk_callback_avx+0x50e> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -10950,14 +10938,14 @@ _sk_bicubic_p3x_avx LABEL PROC PUBLIC _sk_bicubic_n3y_avx _sk_bicubic_n3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,90,6,0,0 ; vbroadcastss 0x65a(%rip),%ymm1 # 6a68 <_sk_callback_avx+0x508> + DB 196,226,125,24,13,100,6,0,0 ; vbroadcastss 0x664(%rip),%ymm1 # 6ac8 <_sk_callback_avx+0x512> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,80,6,0,0 ; vbroadcastss 0x650(%rip),%ymm8 # 6a6c <_sk_callback_avx+0x50c> + DB 196,98,125,24,5,90,6,0,0 ; vbroadcastss 0x65a(%rip),%ymm8 # 6acc <_sk_callback_avx+0x516> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,65,6,0,0 ; vbroadcastss 0x641(%rip),%ymm10 # 6a70 <_sk_callback_avx+0x510> + DB 196,98,125,24,21,75,6,0,0 ; vbroadcastss 0x64b(%rip),%ymm10 # 6ad0 <_sk_callback_avx+0x51a> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,55,6,0,0 ; vbroadcastss 0x637(%rip),%ymm10 # 6a74 <_sk_callback_avx+0x514> + DB 196,98,125,24,21,65,6,0,0 ; vbroadcastss 0x641(%rip),%ymm10 # 6ad4 <_sk_callback_avx+0x51e> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -10967,19 +10955,19 @@ _sk_bicubic_n3y_avx LABEL PROC PUBLIC _sk_bicubic_n1y_avx _sk_bicubic_n1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,26,6,0,0 ; vbroadcastss 0x61a(%rip),%ymm1 # 6a78 <_sk_callback_avx+0x518> + DB 196,226,125,24,13,36,6,0,0 ; vbroadcastss 0x624(%rip),%ymm1 # 6ad8 <_sk_callback_avx+0x522> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm8 # 6a7c <_sk_callback_avx+0x51c> + DB 196,98,125,24,5,26,6,0,0 ; vbroadcastss 0x61a(%rip),%ymm8 # 6adc <_sk_callback_avx+0x526> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm9 # 6a80 <_sk_callback_avx+0x520> + DB 196,98,125,24,13,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm9 # 6ae0 <_sk_callback_avx+0x52a> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm10 # 6a84 <_sk_callback_avx+0x524> + DB 196,98,125,24,21,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm10 # 6ae4 <_sk_callback_avx+0x52e> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,237,5,0,0 ; vbroadcastss 0x5ed(%rip),%ymm10 # 6a88 <_sk_callback_avx+0x528> + DB 196,98,125,24,21,247,5,0,0 ; vbroadcastss 0x5f7(%rip),%ymm10 # 6ae8 <_sk_callback_avx+0x532> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,222,5,0,0 ; vbroadcastss 0x5de(%rip),%ymm9 # 6a8c <_sk_callback_avx+0x52c> + DB 196,98,125,24,13,232,5,0,0 ; vbroadcastss 0x5e8(%rip),%ymm9 # 6aec <_sk_callback_avx+0x536> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -10988,17 +10976,17 @@ _sk_bicubic_n1y_avx LABEL PROC PUBLIC _sk_bicubic_p1y_avx _sk_bicubic_p1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,198,5,0,0 ; vbroadcastss 0x5c6(%rip),%ymm8 # 6a90 <_sk_callback_avx+0x530> + DB 196,98,125,24,5,208,5,0,0 ; vbroadcastss 0x5d0(%rip),%ymm8 # 6af0 <_sk_callback_avx+0x53a> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,183,5,0,0 ; vbroadcastss 0x5b7(%rip),%ymm10 # 6a94 <_sk_callback_avx+0x534> + DB 196,98,125,24,21,193,5,0,0 ; vbroadcastss 0x5c1(%rip),%ymm10 # 6af4 <_sk_callback_avx+0x53e> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,173,5,0,0 ; vbroadcastss 0x5ad(%rip),%ymm11 # 6a98 <_sk_callback_avx+0x538> + DB 196,98,125,24,29,183,5,0,0 ; vbroadcastss 0x5b7(%rip),%ymm11 # 6af8 <_sk_callback_avx+0x542> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,148,5,0,0 ; vbroadcastss 0x594(%rip),%ymm9 # 6a9c <_sk_callback_avx+0x53c> + DB 196,98,125,24,13,158,5,0,0 ; vbroadcastss 0x59e(%rip),%ymm9 # 6afc <_sk_callback_avx+0x546> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11007,13 +10995,13 @@ _sk_bicubic_p1y_avx LABEL PROC PUBLIC _sk_bicubic_p3y_avx _sk_bicubic_p3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,124,5,0,0 ; vbroadcastss 0x57c(%rip),%ymm1 # 6aa0 <_sk_callback_avx+0x540> + DB 196,226,125,24,13,134,5,0,0 ; vbroadcastss 0x586(%rip),%ymm1 # 6b00 <_sk_callback_avx+0x54a> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,104,5,0,0 ; vbroadcastss 0x568(%rip),%ymm10 # 6aa4 <_sk_callback_avx+0x544> + DB 196,98,125,24,21,114,5,0,0 ; vbroadcastss 0x572(%rip),%ymm10 # 6b04 <_sk_callback_avx+0x54e> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,94,5,0,0 ; vbroadcastss 0x55e(%rip),%ymm10 # 6aa8 <_sk_callback_avx+0x548> + DB 196,98,125,24,21,104,5,0,0 ; vbroadcastss 0x568(%rip),%ymm10 # 6b08 <_sk_callback_avx+0x552> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -11024,14 +11012,16 @@ PUBLIC _sk_callback_avx _sk_callback_avx LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,160,0,0,0 ; sub $0xa0,%rsp + DB 72,129,236,184,0,0,0 ; sub $0xb8,%rsp DB 197,252,17,188,36,128,0,0,0 ; vmovups %ymm7,0x80(%rsp) DB 197,252,17,116,36,96 ; vmovups %ymm6,0x60(%rsp) DB 197,252,17,108,36,64 ; vmovups %ymm5,0x40(%rsp) DB 197,252,17,100,36,32 ; vmovups %ymm4,0x20(%rsp) - DB 73,137,207 ; mov %rcx,%r15 - DB 73,137,214 ; mov %rdx,%r14 + DB 77,137,196 ; mov %r8,%r12 + DB 73,137,206 ; mov %rcx,%r14 + DB 73,137,215 ; mov %rdx,%r15 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,137,195 ; mov %rax,%rbx DB 197,252,20,225 ; vunpcklps %ymm1,%ymm0,%ymm4 @@ -11050,9 +11040,9 @@ _sk_callback_avx LABEL PROC DB 197,253,17,107,40 ; vmovupd %ymm5,0x28(%rbx) DB 197,253,17,75,72 ; vmovupd %ymm1,0x48(%rbx) DB 197,253,17,67,104 ; vmovupd %ymm0,0x68(%rbx) - DB 77,133,255 ; test %r15,%r15 + DB 77,133,228 ; test %r12,%r12 DB 186,8,0,0,0 ; mov $0x8,%edx - DB 65,15,69,215 ; cmovne %r15d,%edx + DB 65,15,69,212 ; cmovne %r12d,%edx DB 72,137,217 ; mov %rbx,%rcx DB 197,248,119 ; vzeroupper DB 255,19 ; callq *(%rbx) @@ -11074,14 +11064,16 @@ _sk_callback_avx LABEL PROC DB 197,213,20,211 ; vunpcklpd %ymm3,%ymm5,%ymm2 DB 197,213,21,219 ; vunpckhpd %ymm3,%ymm5,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,249 ; mov %r15,%rcx + DB 76,137,250 ; mov %r15,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 77,137,224 ; mov %r12,%r8 DB 197,252,16,100,36,32 ; vmovups 0x20(%rsp),%ymm4 DB 197,252,16,108,36,64 ; vmovups 0x40(%rsp),%ymm5 DB 197,252,16,116,36,96 ; vmovups 0x60(%rsp),%ymm6 DB 197,252,16,188,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm7 - DB 72,129,196,160,0,0,0 ; add $0xa0,%rsp + DB 72,129,196,184,0,0,0 ; add $0xb8,%rsp DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -11127,25 +11119,25 @@ ALIGN 4 DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 672d <.literal4+0xb1> + DB 71,225,61 ; rex.RXB loope 678d <.literal4+0xb1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 673d <.literal4+0xc1> + DB 71,225,61 ; rex.RXB loope 679d <.literal4+0xc1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 674d <.literal4+0xd1> + DB 71,225,61 ; rex.RXB loope 67ad <.literal4+0xd1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 675d <.literal4+0xe1> + DB 71,225,61 ; rex.RXB loope 67bd <.literal4+0xe1> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) @@ -11197,7 +11189,7 @@ ALIGN 4 DB 190,129,128,128,59 ; mov $0x3b808081,%esi DB 129,128,128,59,0,248,0,0,8,33 ; addl $0x21080000,-0x7ffc480(%rax) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 67b1 <.literal4+0x135> + DB 224,7 ; loopne 6811 <.literal4+0x135> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11213,10 +11205,10 @@ ALIGN 4 DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) DB 0,52,255 ; add %dh,(%rdi,%rdi,8) DB 255 ; (bad) - DB 127,0 ; jg 67d8 <.literal4+0x15c> + DB 127,0 ; jg 6838 <.literal4+0x15c> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6851 <.literal4+0x1d5> + DB 119,115 ; ja 68b1 <.literal4+0x1d5> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11230,10 +11222,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 680c <.literal4+0x190> + DB 127,0 ; jg 686c <.literal4+0x190> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6885 <.literal4+0x209> + DB 119,115 ; ja 68e5 <.literal4+0x209> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11247,10 +11239,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 6840 <.literal4+0x1c4> + DB 127,0 ; jg 68a0 <.literal4+0x1c4> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 68b9 <.literal4+0x23d> + DB 119,115 ; ja 6919 <.literal4+0x23d> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11264,10 +11256,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 6874 <.literal4+0x1f8> + DB 127,0 ; jg 68d4 <.literal4+0x1f8> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 68ed <.literal4+0x271> + DB 119,115 ; ja 694d <.literal4+0x271> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11280,7 +11272,7 @@ ALIGN 4 DB 0,75,0 ; add %cl,0x0(%rbx) DB 0,128,63,0,0,200 ; add %al,-0x37ffffc1(%rax) DB 66,0,0 ; rex.X add %al,(%rax) - DB 127,67 ; jg 68eb <.literal4+0x26f> + DB 127,67 ; jg 694b <.literal4+0x26f> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -11292,10 +11284,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 690b <.literal4+0x28f> + DB 118,63 ; jbe 696b <.literal4+0x28f> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 691f <.literal4+0x2a3> + DB 127,67 ; jg 697f <.literal4+0x2a3> DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax) DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax) @@ -11304,7 +11296,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6901 <.literal4+0x285> + DB 224,7 ; loopne 6961 <.literal4+0x285> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11316,7 +11308,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 691d <.literal4+0x2a1> + DB 224,7 ; loopne 697d <.literal4+0x2a1> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11327,7 +11319,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 6972 <.literal4+0x2f6> + DB 124,66 ; jl 69d2 <.literal4+0x2f6> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -11345,9 +11337,9 @@ ALIGN 4 DB 137,136,136,59,15,0 ; mov %ecx,0xf3b88(%rax) DB 0,0 ; add %al,(%rax) DB 137,136,136,61,0,0 ; mov %ecx,0x3d88(%rax) - DB 112,65 ; jo 69b5 <.literal4+0x339> + DB 112,65 ; jo 6a15 <.literal4+0x339> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 69c3 <.literal4+0x347> + DB 127,67 ; jg 6a23 <.literal4+0x347> DB 0,128,0,0,0,0 ; add %al,0x0(%rax) DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) DB 0,0 ; add %al,(%rax) @@ -11363,7 +11355,7 @@ ALIGN 4 DB 0,128,55,0,0,128 ; add %al,-0x7fffffc9(%rax) DB 63 ; (bad) DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6a03 <.literal4+0x387> + DB 127,71 ; jg 6a63 <.literal4+0x387> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,191,0,0,0 ; add %al,0xbf(%rax) @@ -11632,8 +11624,9 @@ _sk_start_pipeline_sse41 LABEL PROC DB 65,84 ; push %r12 DB 86 ; push %rsi DB 87 ; push %rdi + DB 85 ; push %rbp DB 83 ; push %rbx - DB 72,129,236,160,0,0,0 ; sub $0xa0,%rsp + DB 72,129,236,168,0,0,0 ; sub $0xa8,%rsp DB 68,15,41,188,36,144,0,0,0 ; movaps %xmm15,0x90(%rsp) DB 68,15,41,180,36,128,0,0,0 ; movaps %xmm14,0x80(%rsp) DB 68,15,41,108,36,112 ; movaps %xmm13,0x70(%rsp) @@ -11644,19 +11637,20 @@ _sk_start_pipeline_sse41 LABEL PROC DB 68,15,41,68,36,32 ; movaps %xmm8,0x20(%rsp) DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp) DB 15,41,52,36 ; movaps %xmm6,(%rsp) - DB 77,137,205 ; mov %r9,%r13 - DB 77,137,198 ; mov %r8,%r14 - DB 72,137,203 ; mov %rcx,%rbx - DB 72,137,214 ; mov %rdx,%rsi - DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,199 ; mov %rax,%r15 - DB 73,137,244 ; mov %rsi,%r12 - DB 72,141,67,4 ; lea 0x4(%rbx),%rax - DB 76,57,232 ; cmp %r13,%rax - DB 118,5 ; jbe 73 <_sk_start_pipeline_sse41+0x73> - DB 72,137,223 ; mov %rbx,%rdi - DB 235,57 ; jmp ac <_sk_start_pipeline_sse41+0xac> - DB 185,0,0,0,0 ; mov $0x0,%ecx + DB 76,137,195 ; mov %r8,%rbx + DB 73,137,214 ; mov %rdx,%r14 + DB 72,137,205 ; mov %rcx,%rbp + DB 76,139,188,36,16,1,0,0 ; mov 0x110(%rsp),%r15 + DB 76,137,206 ; mov %r9,%rsi + DB 72,173 ; lods %ds:(%rsi),%rax + DB 73,137,196 ; mov %rax,%r12 + DB 73,137,245 ; mov %rsi,%r13 + DB 72,141,69,4 ; lea 0x4(%rbp),%rax + DB 72,57,216 ; cmp %rbx,%rax + DB 118,5 ; jbe 7c <_sk_start_pipeline_sse41+0x7c> + DB 72,137,234 ; mov %rbp,%rdx + DB 235,61 ; jmp b9 <_sk_start_pipeline_sse41+0xb9> + DB 65,184,0,0,0,0 ; mov $0x0,%r8d DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -11665,17 +11659,18 @@ _sk_start_pipeline_sse41 LABEL PROC DB 15,87,237 ; xorps %xmm5,%xmm5 DB 15,87,246 ; xorps %xmm6,%xmm6 DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 72,137,223 ; mov %rbx,%rdi - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 65,255,215 ; callq *%r15 - DB 72,141,123,4 ; lea 0x4(%rbx),%rdi - DB 72,131,195,8 ; add $0x8,%rbx - DB 76,57,235 ; cmp %r13,%rbx - DB 72,137,251 ; mov %rdi,%rbx - DB 118,199 ; jbe 73 <_sk_start_pipeline_sse41+0x73> - DB 73,41,253 ; sub %rdi,%r13 - DB 116,36 ; je d5 <_sk_start_pipeline_sse41+0xd5> + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 72,137,234 ; mov %rbp,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 65,255,212 ; callq *%r12 + DB 72,141,85,4 ; lea 0x4(%rbp),%rdx + DB 72,131,197,8 ; add $0x8,%rbp + DB 72,57,221 ; cmp %rbx,%rbp + DB 72,137,213 ; mov %rdx,%rbp + DB 118,195 ; jbe 7c <_sk_start_pipeline_sse41+0x7c> + DB 72,41,211 ; sub %rdx,%rbx + DB 116,39 ; je e5 <_sk_start_pipeline_sse41+0xe5> DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -11684,10 +11679,11 @@ _sk_start_pipeline_sse41 LABEL PROC DB 15,87,237 ; xorps %xmm5,%xmm5 DB 15,87,246 ; xorps %xmm6,%xmm6 DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,233 ; mov %r13,%rcx - DB 65,255,215 ; callq *%r15 + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 76,137,241 ; mov %r14,%rcx + DB 73,137,216 ; mov %rbx,%r8 + DB 65,255,212 ; callq *%r12 DB 15,40,52,36 ; movaps (%rsp),%xmm6 DB 15,40,124,36,16 ; movaps 0x10(%rsp),%xmm7 DB 68,15,40,68,36,32 ; movaps 0x20(%rsp),%xmm8 @@ -11698,8 +11694,9 @@ _sk_start_pipeline_sse41 LABEL PROC DB 68,15,40,108,36,112 ; movaps 0x70(%rsp),%xmm13 DB 68,15,40,180,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm14 DB 68,15,40,188,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm15 - DB 72,129,196,160,0,0,0 ; add $0xa0,%rsp + DB 72,129,196,168,0,0,0 ; add $0xa8,%rsp DB 91 ; pop %rbx + DB 93 ; pop %rbp DB 95 ; pop %rdi DB 94 ; pop %rsi DB 65,92 ; pop %r12 @@ -11714,20 +11711,19 @@ _sk_just_return_sse41 LABEL PROC PUBLIC _sk_seed_shader_sse41 _sk_seed_shader_sse41 LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,110,199 ; movd %edi,%xmm0 + DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 - DB 15,40,21,211,79,0,0 ; movaps 0x4fd3(%rip),%xmm2 # 5110 <_sk_callback_sse41+0xbd> + DB 15,40,21,20,80,0,0 ; movaps 0x5014(%rip),%xmm2 # 5160 <_sk_callback_sse41+0xd3> DB 15,88,202 ; addps %xmm2,%xmm1 - DB 15,16,2 ; movups (%rdx),%xmm0 + DB 15,16,7 ; movups (%rdi),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 - DB 102,15,110,8 ; movd (%rax),%xmm1 + DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,194,79,0,0 ; movaps 0x4fc2(%rip),%xmm2 # 5120 <_sk_callback_sse41+0xcd> + DB 15,40,21,3,80,0,0 ; movaps 0x5003(%rip),%xmm2 # 5170 <_sk_callback_sse41+0xe3> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -11738,22 +11734,21 @@ _sk_seed_shader_sse41 LABEL PROC PUBLIC _sk_dither_sse41 _sk_dither_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,68,15,110,199 ; movd %edi,%xmm8 + DB 102,68,15,110,194 ; movd %edx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 - DB 243,68,15,111,74,32 ; movdqu 0x20(%rdx),%xmm9 + DB 243,68,15,111,79,32 ; movdqu 0x20(%rdi),%xmm9 DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 - DB 76,139,0 ; mov (%rax),%r8 - DB 102,69,15,110,0 ; movd (%r8),%xmm8 + DB 102,68,15,110,193 ; movd %ecx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 - DB 102,68,15,111,21,141,79,0,0 ; movdqa 0x4f8d(%rip),%xmm10 # 5130 <_sk_callback_sse41+0xdd> + DB 102,68,15,111,21,209,79,0,0 ; movdqa 0x4fd1(%rip),%xmm10 # 5180 <_sk_callback_sse41+0xf3> DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,65,15,114,243,5 ; pslld $0x5,%xmm11 DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 102,65,15,114,242,4 ; pslld $0x4,%xmm10 - DB 102,68,15,111,37,121,79,0,0 ; movdqa 0x4f79(%rip),%xmm12 # 5140 <_sk_callback_sse41+0xed> - DB 102,68,15,111,45,128,79,0,0 ; movdqa 0x4f80(%rip),%xmm13 # 5150 <_sk_callback_sse41+0xfd> + DB 102,68,15,111,37,189,79,0,0 ; movdqa 0x4fbd(%rip),%xmm12 # 5190 <_sk_callback_sse41+0x103> + DB 102,68,15,111,45,196,79,0,0 ; movdqa 0x4fc4(%rip),%xmm13 # 51a0 <_sk_callback_sse41+0x113> DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 DB 102,69,15,219,245 ; pand %xmm13,%xmm14 DB 102,65,15,114,246,2 ; pslld $0x2,%xmm14 @@ -11765,13 +11760,13 @@ _sk_dither_sse41 LABEL PROC DB 102,65,15,114,209,2 ; psrld $0x2,%xmm9 DB 102,69,15,235,234 ; por %xmm10,%xmm13 DB 102,69,15,235,233 ; por %xmm9,%xmm13 - DB 102,69,15,235,235 ; por %xmm11,%xmm13 - DB 102,69,15,235,198 ; por %xmm14,%xmm8 - DB 102,69,15,235,197 ; por %xmm13,%xmm8 - DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,59,79,0,0 ; mulps 0x4f3b(%rip),%xmm8 # 5160 <_sk_callback_sse41+0x10d> - DB 68,15,88,5,67,79,0,0 ; addps 0x4f43(%rip),%xmm8 # 5170 <_sk_callback_sse41+0x11d> - DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 + DB 102,69,15,235,243 ; por %xmm11,%xmm14 + DB 102,69,15,235,245 ; por %xmm13,%xmm14 + DB 102,69,15,235,240 ; por %xmm8,%xmm14 + DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8 + DB 68,15,89,5,127,79,0,0 ; mulps 0x4f7f(%rip),%xmm8 # 51b0 <_sk_callback_sse41+0x123> + DB 68,15,88,5,135,79,0,0 ; addps 0x4f87(%rip),%xmm8 # 51c0 <_sk_callback_sse41+0x133> + DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,88,194 ; addps %xmm10,%xmm0 @@ -11837,7 +11832,7 @@ _sk_clear_sse41 LABEL PROC PUBLIC _sk_srcatop_sse41 _sk_srcatop_sse41 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,156,78,0,0 ; movaps 0x4e9c(%rip),%xmm8 # 5180 <_sk_callback_sse41+0x12d> + DB 68,15,40,5,225,78,0,0 ; movaps 0x4ee1(%rip),%xmm8 # 51d0 <_sk_callback_sse41+0x143> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -11860,7 +11855,7 @@ PUBLIC _sk_dstatop_sse41 _sk_dstatop_sse41 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 DB 68,15,89,196 ; mulps %xmm4,%xmm8 - DB 68,15,40,13,95,78,0,0 ; movaps 0x4e5f(%rip),%xmm9 # 5190 <_sk_callback_sse41+0x13d> + DB 68,15,40,13,164,78,0,0 ; movaps 0x4ea4(%rip),%xmm9 # 51e0 <_sk_callback_sse41+0x153> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -11901,7 +11896,7 @@ _sk_dstin_sse41 LABEL PROC PUBLIC _sk_srcout_sse41 _sk_srcout_sse41 LABEL PROC - DB 68,15,40,5,3,78,0,0 ; movaps 0x4e03(%rip),%xmm8 # 51a0 <_sk_callback_sse41+0x14d> + DB 68,15,40,5,72,78,0,0 ; movaps 0x4e48(%rip),%xmm8 # 51f0 <_sk_callback_sse41+0x163> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -11912,7 +11907,7 @@ _sk_srcout_sse41 LABEL PROC PUBLIC _sk_dstout_sse41 _sk_dstout_sse41 LABEL PROC - DB 68,15,40,5,243,77,0,0 ; movaps 0x4df3(%rip),%xmm8 # 51b0 <_sk_callback_sse41+0x15d> + DB 68,15,40,5,56,78,0,0 ; movaps 0x4e38(%rip),%xmm8 # 5200 <_sk_callback_sse41+0x173> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -11927,7 +11922,7 @@ _sk_dstout_sse41 LABEL PROC PUBLIC _sk_srcover_sse41 _sk_srcover_sse41 LABEL PROC - DB 68,15,40,5,214,77,0,0 ; movaps 0x4dd6(%rip),%xmm8 # 51c0 <_sk_callback_sse41+0x16d> + DB 68,15,40,5,27,78,0,0 ; movaps 0x4e1b(%rip),%xmm8 # 5210 <_sk_callback_sse41+0x183> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -11945,7 +11940,7 @@ _sk_srcover_sse41 LABEL PROC PUBLIC _sk_dstover_sse41 _sk_dstover_sse41 LABEL PROC - DB 68,15,40,5,170,77,0,0 ; movaps 0x4daa(%rip),%xmm8 # 51d0 <_sk_callback_sse41+0x17d> + DB 68,15,40,5,239,77,0,0 ; movaps 0x4def(%rip),%xmm8 # 5220 <_sk_callback_sse41+0x193> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -11969,7 +11964,7 @@ _sk_modulate_sse41 LABEL PROC PUBLIC _sk_multiply_sse41 _sk_multiply_sse41 LABEL PROC - DB 68,15,40,5,126,77,0,0 ; movaps 0x4d7e(%rip),%xmm8 # 51e0 <_sk_callback_sse41+0x18d> + DB 68,15,40,5,195,77,0,0 ; movaps 0x4dc3(%rip),%xmm8 # 5230 <_sk_callback_sse41+0x1a3> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -12038,7 +12033,7 @@ _sk_screen_sse41 LABEL PROC PUBLIC _sk_xor__sse41 _sk_xor__sse41 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,179,76,0,0 ; movaps 0x4cb3(%rip),%xmm3 # 51f0 <_sk_callback_sse41+0x19d> + DB 15,40,29,248,76,0,0 ; movaps 0x4cf8(%rip),%xmm3 # 5240 <_sk_callback_sse41+0x1b3> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -12084,7 +12079,7 @@ _sk_darken_sse41 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,95,209 ; maxps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,30,76,0,0 ; movaps 0x4c1e(%rip),%xmm2 # 5200 <_sk_callback_sse41+0x1ad> + DB 15,40,21,99,76,0,0 ; movaps 0x4c63(%rip),%xmm2 # 5250 <_sk_callback_sse41+0x1c3> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12116,7 +12111,7 @@ _sk_lighten_sse41 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,195,75,0,0 ; movaps 0x4bc3(%rip),%xmm2 # 5210 <_sk_callback_sse41+0x1bd> + DB 15,40,21,8,76,0,0 ; movaps 0x4c08(%rip),%xmm2 # 5260 <_sk_callback_sse41+0x1d3> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12151,7 +12146,7 @@ _sk_difference_sse41 LABEL PROC DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,93,75,0,0 ; movaps 0x4b5d(%rip),%xmm2 # 5220 <_sk_callback_sse41+0x1cd> + DB 15,40,21,162,75,0,0 ; movaps 0x4ba2(%rip),%xmm2 # 5270 <_sk_callback_sse41+0x1e3> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12177,7 +12172,7 @@ _sk_exclusion_sse41 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,29,75,0,0 ; movaps 0x4b1d(%rip),%xmm2 # 5230 <_sk_callback_sse41+0x1dd> + DB 15,40,21,98,75,0,0 ; movaps 0x4b62(%rip),%xmm2 # 5280 <_sk_callback_sse41+0x1f3> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12188,7 +12183,7 @@ _sk_exclusion_sse41 LABEL PROC PUBLIC _sk_colorburn_sse41 _sk_colorburn_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,16,75,0,0 ; movaps 0x4b10(%rip),%xmm10 # 5240 <_sk_callback_sse41+0x1ed> + DB 68,15,40,21,85,75,0,0 ; movaps 0x4b55(%rip),%xmm10 # 5290 <_sk_callback_sse41+0x203> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,203 ; movaps %xmm11,%xmm9 @@ -12268,7 +12263,7 @@ _sk_colorburn_sse41 LABEL PROC PUBLIC _sk_colordodge_sse41 _sk_colordodge_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,238,73,0,0 ; movaps 0x49ee(%rip),%xmm10 # 5250 <_sk_callback_sse41+0x1fd> + DB 68,15,40,21,51,74,0,0 ; movaps 0x4a33(%rip),%xmm10 # 52a0 <_sk_callback_sse41+0x213> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -12349,7 +12344,7 @@ _sk_hardlight_sse41 LABEL PROC DB 15,40,244 ; movaps %xmm4,%xmm6 DB 15,40,227 ; movaps %xmm3,%xmm4 DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 68,15,40,21,196,72,0,0 ; movaps 0x48c4(%rip),%xmm10 # 5260 <_sk_callback_sse41+0x20d> + DB 68,15,40,21,9,73,0,0 ; movaps 0x4909(%rip),%xmm10 # 52b0 <_sk_callback_sse41+0x223> DB 65,15,40,234 ; movaps %xmm10,%xmm5 DB 15,92,239 ; subps %xmm7,%xmm5 DB 15,40,197 ; movaps %xmm5,%xmm0 @@ -12431,7 +12426,7 @@ PUBLIC _sk_overlay_sse41 _sk_overlay_sse41 LABEL PROC DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 68,15,40,240 ; movaps %xmm0,%xmm14 - DB 68,15,40,21,166,71,0,0 ; movaps 0x47a6(%rip),%xmm10 # 5270 <_sk_callback_sse41+0x21d> + DB 68,15,40,21,235,71,0,0 ; movaps 0x47eb(%rip),%xmm10 # 52c0 <_sk_callback_sse41+0x233> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 65,15,40,195 ; movaps %xmm11,%xmm0 @@ -12515,7 +12510,7 @@ _sk_softlight_sse41 LABEL PROC DB 15,40,198 ; movaps %xmm6,%xmm0 DB 15,94,199 ; divps %xmm7,%xmm0 DB 65,15,84,193 ; andps %xmm9,%xmm0 - DB 15,40,13,121,70,0,0 ; movaps 0x4679(%rip),%xmm1 # 5280 <_sk_callback_sse41+0x22d> + DB 15,40,13,190,70,0,0 ; movaps 0x46be(%rip),%xmm1 # 52d0 <_sk_callback_sse41+0x243> DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 68,15,40,240 ; movaps %xmm0,%xmm14 @@ -12528,10 +12523,10 @@ _sk_softlight_sse41 LABEL PROC DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,89,210 ; mulps %xmm2,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 68,15,40,45,87,70,0,0 ; movaps 0x4657(%rip),%xmm13 # 5290 <_sk_callback_sse41+0x23d> + DB 68,15,40,45,156,70,0,0 ; movaps 0x469c(%rip),%xmm13 # 52e0 <_sk_callback_sse41+0x253> DB 69,15,88,245 ; addps %xmm13,%xmm14 DB 68,15,89,242 ; mulps %xmm2,%xmm14 - DB 68,15,40,37,87,70,0,0 ; movaps 0x4657(%rip),%xmm12 # 52a0 <_sk_callback_sse41+0x24d> + DB 68,15,40,37,156,70,0,0 ; movaps 0x469c(%rip),%xmm12 # 52f0 <_sk_callback_sse41+0x263> DB 69,15,89,252 ; mulps %xmm12,%xmm15 DB 69,15,88,254 ; addps %xmm14,%xmm15 DB 15,40,198 ; movaps %xmm6,%xmm0 @@ -12717,12 +12712,12 @@ _sk_hue_sse41 LABEL PROC DB 68,15,84,208 ; andps %xmm0,%xmm10 DB 15,84,200 ; andps %xmm0,%xmm1 DB 68,15,84,232 ; andps %xmm0,%xmm13 - DB 15,40,5,189,67,0,0 ; movaps 0x43bd(%rip),%xmm0 # 52b0 <_sk_callback_sse41+0x25d> + DB 15,40,5,2,68,0,0 ; movaps 0x4402(%rip),%xmm0 # 5300 <_sk_callback_sse41+0x273> DB 68,15,89,224 ; mulps %xmm0,%xmm12 - DB 15,40,21,194,67,0,0 ; movaps 0x43c2(%rip),%xmm2 # 52c0 <_sk_callback_sse41+0x26d> + DB 15,40,21,7,68,0,0 ; movaps 0x4407(%rip),%xmm2 # 5310 <_sk_callback_sse41+0x283> DB 15,89,250 ; mulps %xmm2,%xmm7 DB 65,15,88,252 ; addps %xmm12,%xmm7 - DB 68,15,40,53,195,67,0,0 ; movaps 0x43c3(%rip),%xmm14 # 52d0 <_sk_callback_sse41+0x27d> + DB 68,15,40,53,8,68,0,0 ; movaps 0x4408(%rip),%xmm14 # 5320 <_sk_callback_sse41+0x293> DB 68,15,40,252 ; movaps %xmm4,%xmm15 DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 68,15,88,255 ; addps %xmm7,%xmm15 @@ -12805,7 +12800,7 @@ _sk_hue_sse41 LABEL PROC DB 65,15,88,214 ; addps %xmm14,%xmm2 DB 15,40,196 ; movaps %xmm4,%xmm0 DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 - DB 68,15,40,13,136,66,0,0 ; movaps 0x4288(%rip),%xmm9 # 52e0 <_sk_callback_sse41+0x28d> + DB 68,15,40,13,205,66,0,0 ; movaps 0x42cd(%rip),%xmm9 # 5330 <_sk_callback_sse41+0x2a3> DB 65,15,40,225 ; movaps %xmm9,%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 15,40,68,36,48 ; movaps 0x30(%rsp),%xmm0 @@ -12899,14 +12894,14 @@ _sk_saturation_sse41 LABEL PROC DB 68,15,84,215 ; andps %xmm7,%xmm10 DB 68,15,84,223 ; andps %xmm7,%xmm11 DB 68,15,84,199 ; andps %xmm7,%xmm8 - DB 15,40,21,59,65,0,0 ; movaps 0x413b(%rip),%xmm2 # 52f0 <_sk_callback_sse41+0x29d> + DB 15,40,21,128,65,0,0 ; movaps 0x4180(%rip),%xmm2 # 5340 <_sk_callback_sse41+0x2b3> DB 15,40,221 ; movaps %xmm5,%xmm3 DB 15,89,218 ; mulps %xmm2,%xmm3 - DB 15,40,13,62,65,0,0 ; movaps 0x413e(%rip),%xmm1 # 5300 <_sk_callback_sse41+0x2ad> + DB 15,40,13,131,65,0,0 ; movaps 0x4183(%rip),%xmm1 # 5350 <_sk_callback_sse41+0x2c3> DB 15,40,254 ; movaps %xmm6,%xmm7 DB 15,89,249 ; mulps %xmm1,%xmm7 DB 15,88,251 ; addps %xmm3,%xmm7 - DB 68,15,40,45,61,65,0,0 ; movaps 0x413d(%rip),%xmm13 # 5310 <_sk_callback_sse41+0x2bd> + DB 68,15,40,45,130,65,0,0 ; movaps 0x4182(%rip),%xmm13 # 5360 <_sk_callback_sse41+0x2d3> DB 69,15,89,245 ; mulps %xmm13,%xmm14 DB 68,15,88,247 ; addps %xmm7,%xmm14 DB 65,15,40,218 ; movaps %xmm10,%xmm3 @@ -12987,7 +12982,7 @@ _sk_saturation_sse41 LABEL PROC DB 65,15,88,253 ; addps %xmm13,%xmm7 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 102,68,15,56,20,223 ; blendvps %xmm0,%xmm7,%xmm11 - DB 68,15,40,13,3,64,0,0 ; movaps 0x4003(%rip),%xmm9 # 5320 <_sk_callback_sse41+0x2cd> + DB 68,15,40,13,72,64,0,0 ; movaps 0x4048(%rip),%xmm9 # 5370 <_sk_callback_sse41+0x2e3> DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 68,15,92,204 ; subps %xmm4,%xmm9 DB 15,40,60,36 ; movaps (%rsp),%xmm7 @@ -13042,14 +13037,14 @@ _sk_color_sse41 LABEL PROC DB 15,40,231 ; movaps %xmm7,%xmm4 DB 68,15,89,244 ; mulps %xmm4,%xmm14 DB 15,89,204 ; mulps %xmm4,%xmm1 - DB 68,15,40,13,72,63,0,0 ; movaps 0x3f48(%rip),%xmm9 # 5330 <_sk_callback_sse41+0x2dd> + DB 68,15,40,13,141,63,0,0 ; movaps 0x3f8d(%rip),%xmm9 # 5380 <_sk_callback_sse41+0x2f3> DB 65,15,40,250 ; movaps %xmm10,%xmm7 DB 65,15,89,249 ; mulps %xmm9,%xmm7 - DB 68,15,40,21,72,63,0,0 ; movaps 0x3f48(%rip),%xmm10 # 5340 <_sk_callback_sse41+0x2ed> + DB 68,15,40,21,141,63,0,0 ; movaps 0x3f8d(%rip),%xmm10 # 5390 <_sk_callback_sse41+0x303> DB 65,15,40,219 ; movaps %xmm11,%xmm3 DB 65,15,89,218 ; mulps %xmm10,%xmm3 DB 15,88,223 ; addps %xmm7,%xmm3 - DB 68,15,40,29,69,63,0,0 ; movaps 0x3f45(%rip),%xmm11 # 5350 <_sk_callback_sse41+0x2fd> + DB 68,15,40,29,138,63,0,0 ; movaps 0x3f8a(%rip),%xmm11 # 53a0 <_sk_callback_sse41+0x313> DB 69,15,40,236 ; movaps %xmm12,%xmm13 DB 69,15,89,235 ; mulps %xmm11,%xmm13 DB 68,15,88,235 ; addps %xmm3,%xmm13 @@ -13134,7 +13129,7 @@ _sk_color_sse41 LABEL PROC DB 65,15,88,251 ; addps %xmm11,%xmm7 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 102,15,56,20,207 ; blendvps %xmm0,%xmm7,%xmm1 - DB 68,15,40,13,1,62,0,0 ; movaps 0x3e01(%rip),%xmm9 # 5360 <_sk_callback_sse41+0x30d> + DB 68,15,40,13,70,62,0,0 ; movaps 0x3e46(%rip),%xmm9 # 53b0 <_sk_callback_sse41+0x323> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,92,196 ; subps %xmm4,%xmm0 DB 68,15,89,192 ; mulps %xmm0,%xmm8 @@ -13186,13 +13181,13 @@ _sk_luminosity_sse41 LABEL PROC DB 69,15,89,216 ; mulps %xmm8,%xmm11 DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,89,205 ; mulps %xmm5,%xmm9 - DB 68,15,40,5,83,61,0,0 ; movaps 0x3d53(%rip),%xmm8 # 5370 <_sk_callback_sse41+0x31d> + DB 68,15,40,5,152,61,0,0 ; movaps 0x3d98(%rip),%xmm8 # 53c0 <_sk_callback_sse41+0x333> DB 65,15,89,192 ; mulps %xmm8,%xmm0 - DB 68,15,40,21,87,61,0,0 ; movaps 0x3d57(%rip),%xmm10 # 5380 <_sk_callback_sse41+0x32d> + DB 68,15,40,21,156,61,0,0 ; movaps 0x3d9c(%rip),%xmm10 # 53d0 <_sk_callback_sse41+0x343> DB 15,40,233 ; movaps %xmm1,%xmm5 DB 65,15,89,234 ; mulps %xmm10,%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 - DB 68,15,40,37,85,61,0,0 ; movaps 0x3d55(%rip),%xmm12 # 5390 <_sk_callback_sse41+0x33d> + DB 68,15,40,37,154,61,0,0 ; movaps 0x3d9a(%rip),%xmm12 # 53e0 <_sk_callback_sse41+0x353> DB 68,15,40,242 ; movaps %xmm2,%xmm14 DB 69,15,89,244 ; mulps %xmm12,%xmm14 DB 68,15,88,245 ; addps %xmm5,%xmm14 @@ -13277,7 +13272,7 @@ _sk_luminosity_sse41 LABEL PROC DB 65,15,88,244 ; addps %xmm12,%xmm6 DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 102,68,15,56,20,206 ; blendvps %xmm0,%xmm6,%xmm9 - DB 15,40,5,11,60,0,0 ; movaps 0x3c0b(%rip),%xmm0 # 53a0 <_sk_callback_sse41+0x34d> + DB 15,40,5,80,60,0,0 ; movaps 0x3c50(%rip),%xmm0 # 53f0 <_sk_callback_sse41+0x363> DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,92,215 ; subps %xmm7,%xmm2 DB 15,89,226 ; mulps %xmm2,%xmm4 @@ -13315,24 +13310,24 @@ PUBLIC _sk_srcover_rgba_8888_sse41 _sk_srcover_rgba_8888_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,217,0,0,0 ; jne 18ec <_sk_srcover_rgba_8888_sse41+0xe7> - DB 243,15,111,60,184 ; movdqu (%rax,%rdi,4),%xmm7 - DB 72,133,201 ; test %rcx,%rcx - DB 102,15,111,37,141,59,0,0 ; movdqa 0x3b8d(%rip),%xmm4 # 53b0 <_sk_callback_sse41+0x35d> + DB 77,133,192 ; test %r8,%r8 + DB 15,133,217,0,0,0 ; jne 18f7 <_sk_srcover_rgba_8888_sse41+0xe7> + DB 243,15,111,60,144 ; movdqu (%rax,%rdx,4),%xmm7 + DB 77,133,192 ; test %r8,%r8 + DB 102,15,111,37,210,59,0,0 ; movdqa 0x3bd2(%rip),%xmm4 # 5400 <_sk_callback_sse41+0x373> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,56,0,45,137,59,0,0 ; pshufb 0x3b89(%rip),%xmm5 # 53c0 <_sk_callback_sse41+0x36d> + DB 102,15,56,0,45,206,59,0,0 ; pshufb 0x3bce(%rip),%xmm5 # 5410 <_sk_callback_sse41+0x383> DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,56,0,53,137,59,0,0 ; pshufb 0x3b89(%rip),%xmm6 # 53d0 <_sk_callback_sse41+0x37d> + DB 102,15,56,0,53,206,59,0,0 ; pshufb 0x3bce(%rip),%xmm6 # 5420 <_sk_callback_sse41+0x393> DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 DB 102,15,114,215,24 ; psrld $0x18,%xmm7 DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 - DB 68,15,40,5,134,59,0,0 ; movaps 0x3b86(%rip),%xmm8 # 53e0 <_sk_callback_sse41+0x38d> + DB 68,15,40,5,203,59,0,0 ; movaps 0x3bcb(%rip),%xmm8 # 5430 <_sk_callback_sse41+0x3a3> DB 68,15,92,195 ; subps %xmm3,%xmm8 - DB 68,15,40,37,138,59,0,0 ; movaps 0x3b8a(%rip),%xmm12 # 53f0 <_sk_callback_sse41+0x39d> + DB 68,15,40,37,207,59,0,0 ; movaps 0x3bcf(%rip),%xmm12 # 5440 <_sk_callback_sse41+0x3b3> DB 65,15,89,196 ; mulps %xmm12,%xmm0 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -13358,40 +13353,40 @@ _sk_srcover_rgba_8888_sse41 LABEL PROC DB 102,15,114,240,24 ; pslld $0x18,%xmm0 DB 102,15,235,194 ; por %xmm2,%xmm0 DB 102,15,235,193 ; por %xmm1,%xmm0 - DB 117,89 ; jne 192c <_sk_srcover_rgba_8888_sse41+0x127> - DB 243,15,127,4,184 ; movdqu %xmm0,(%rax,%rdi,4) + DB 117,89 ; jne 1937 <_sk_srcover_rgba_8888_sse41+0x127> + DB 243,15,127,4,144 ; movdqu %xmm0,(%rax,%rdx,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 65,15,40,202 ; movaps %xmm10,%xmm1 DB 65,15,40,211 ; movaps %xmm11,%xmm2 DB 65,15,40,216 ; movaps %xmm8,%xmm3 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,35 ; je 1920 <_sk_srcover_rgba_8888_sse41+0x11b> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 1918 <_sk_srcover_rgba_8888_sse41+0x113> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 15,133,11,255,255,255 ; jne 1818 <_sk_srcover_rgba_8888_sse41+0x13> - DB 102,15,110,100,184,8 ; movd 0x8(%rax,%rdi,4),%xmm4 + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,35 ; je 192b <_sk_srcover_rgba_8888_sse41+0x11b> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 1923 <_sk_srcover_rgba_8888_sse41+0x113> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 15,133,11,255,255,255 ; jne 1823 <_sk_srcover_rgba_8888_sse41+0x13> + DB 102,15,110,100,144,8 ; movd 0x8(%rax,%rdx,4),%xmm4 DB 102,15,112,252,69 ; pshufd $0x45,%xmm4,%xmm7 - DB 102,15,58,34,124,184,4,1 ; pinsrd $0x1,0x4(%rax,%rdi,4),%xmm7 - DB 102,15,58,34,60,184,0 ; pinsrd $0x0,(%rax,%rdi,4),%xmm7 - DB 233,236,254,255,255 ; jmpq 1818 <_sk_srcover_rgba_8888_sse41+0x13> - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,28 ; je 1955 <_sk_srcover_rgba_8888_sse41+0x150> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,14 ; je 194d <_sk_srcover_rgba_8888_sse41+0x148> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,147 ; jne 18d8 <_sk_srcover_rgba_8888_sse41+0xd3> - DB 102,15,58,22,68,184,8,2 ; pextrd $0x2,%xmm0,0x8(%rax,%rdi,4) - DB 102,15,58,22,68,184,4,1 ; pextrd $0x1,%xmm0,0x4(%rax,%rdi,4) - DB 102,15,126,4,184 ; movd %xmm0,(%rax,%rdi,4) - DB 233,121,255,255,255 ; jmpq 18d8 <_sk_srcover_rgba_8888_sse41+0xd3> + DB 102,15,58,34,124,144,4,1 ; pinsrd $0x1,0x4(%rax,%rdx,4),%xmm7 + DB 102,15,58,34,60,144,0 ; pinsrd $0x0,(%rax,%rdx,4),%xmm7 + DB 233,236,254,255,255 ; jmpq 1823 <_sk_srcover_rgba_8888_sse41+0x13> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,28 ; je 1960 <_sk_srcover_rgba_8888_sse41+0x150> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,14 ; je 1958 <_sk_srcover_rgba_8888_sse41+0x148> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,147 ; jne 18e3 <_sk_srcover_rgba_8888_sse41+0xd3> + DB 102,15,58,22,68,144,8,2 ; pextrd $0x2,%xmm0,0x8(%rax,%rdx,4) + DB 102,15,58,22,68,144,4,1 ; pextrd $0x1,%xmm0,0x4(%rax,%rdx,4) + DB 102,15,126,4,144 ; movd %xmm0,(%rax,%rdx,4) + DB 233,121,255,255,255 ; jmpq 18e3 <_sk_srcover_rgba_8888_sse41+0xd3> PUBLIC _sk_clamp_0_sse41 _sk_clamp_0_sse41 LABEL PROC @@ -13405,7 +13400,7 @@ _sk_clamp_0_sse41 LABEL PROC PUBLIC _sk_clamp_1_sse41 _sk_clamp_1_sse41 LABEL PROC - DB 68,15,40,5,129,58,0,0 ; movaps 0x3a81(%rip),%xmm8 # 5400 <_sk_callback_sse41+0x3ad> + DB 68,15,40,5,198,58,0,0 ; movaps 0x3ac6(%rip),%xmm8 # 5450 <_sk_callback_sse41+0x3c3> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -13415,7 +13410,7 @@ _sk_clamp_1_sse41 LABEL PROC PUBLIC _sk_clamp_a_sse41 _sk_clamp_a_sse41 LABEL PROC - DB 15,93,29,118,58,0,0 ; minps 0x3a76(%rip),%xmm3 # 5410 <_sk_callback_sse41+0x3bd> + DB 15,93,29,187,58,0,0 ; minps 0x3abb(%rip),%xmm3 # 5460 <_sk_callback_sse41+0x3d3> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -13488,7 +13483,7 @@ _sk_premul_sse41 LABEL PROC PUBLIC _sk_unpremul_sse41 _sk_unpremul_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,225,57,0,0 ; movaps 0x39e1(%rip),%xmm9 # 5420 <_sk_callback_sse41+0x3cd> + DB 68,15,40,13,38,58,0,0 ; movaps 0x3a26(%rip),%xmm9 # 5470 <_sk_callback_sse41+0x3e3> DB 68,15,94,203 ; divps %xmm3,%xmm9 DB 68,15,194,195,4 ; cmpneqps %xmm3,%xmm8 DB 69,15,84,193 ; andps %xmm9,%xmm8 @@ -13500,20 +13495,20 @@ _sk_unpremul_sse41 LABEL PROC PUBLIC _sk_from_srgb_sse41 _sk_from_srgb_sse41 LABEL PROC - DB 68,15,40,29,204,57,0,0 ; movaps 0x39cc(%rip),%xmm11 # 5430 <_sk_callback_sse41+0x3dd> + DB 68,15,40,29,17,58,0,0 ; movaps 0x3a11(%rip),%xmm11 # 5480 <_sk_callback_sse41+0x3f3> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 DB 68,15,40,208 ; movaps %xmm0,%xmm10 DB 69,15,89,210 ; mulps %xmm10,%xmm10 - DB 68,15,40,37,196,57,0,0 ; movaps 0x39c4(%rip),%xmm12 # 5440 <_sk_callback_sse41+0x3ed> + DB 68,15,40,37,9,58,0,0 ; movaps 0x3a09(%rip),%xmm12 # 5490 <_sk_callback_sse41+0x403> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,196,57,0,0 ; movaps 0x39c4(%rip),%xmm13 # 5450 <_sk_callback_sse41+0x3fd> + DB 68,15,40,45,9,58,0,0 ; movaps 0x3a09(%rip),%xmm13 # 54a0 <_sk_callback_sse41+0x413> DB 69,15,88,197 ; addps %xmm13,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 - DB 68,15,40,53,196,57,0,0 ; movaps 0x39c4(%rip),%xmm14 # 5460 <_sk_callback_sse41+0x40d> + DB 68,15,40,53,9,58,0,0 ; movaps 0x3a09(%rip),%xmm14 # 54b0 <_sk_callback_sse41+0x423> DB 69,15,88,198 ; addps %xmm14,%xmm8 - DB 68,15,40,61,200,57,0,0 ; movaps 0x39c8(%rip),%xmm15 # 5470 <_sk_callback_sse41+0x41d> + DB 68,15,40,61,13,58,0,0 ; movaps 0x3a0d(%rip),%xmm15 # 54c0 <_sk_callback_sse41+0x433> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 68,15,40,209 ; movaps %xmm1,%xmm10 @@ -13555,22 +13550,22 @@ _sk_to_srgb_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,209 ; movaps %xmm1,%xmm2 DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 - DB 68,15,40,29,61,57,0,0 ; movaps 0x393d(%rip),%xmm11 # 5480 <_sk_callback_sse41+0x42d> + DB 68,15,40,29,130,57,0,0 ; movaps 0x3982(%rip),%xmm11 # 54d0 <_sk_callback_sse41+0x443> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 - DB 68,15,40,37,61,57,0,0 ; movaps 0x393d(%rip),%xmm12 # 5490 <_sk_callback_sse41+0x43d> + DB 68,15,40,37,130,57,0,0 ; movaps 0x3982(%rip),%xmm12 # 54e0 <_sk_callback_sse41+0x453> DB 69,15,40,248 ; movaps %xmm8,%xmm15 DB 69,15,89,252 ; mulps %xmm12,%xmm15 - DB 68,15,40,21,61,57,0,0 ; movaps 0x393d(%rip),%xmm10 # 54a0 <_sk_callback_sse41+0x44d> + DB 68,15,40,21,130,57,0,0 ; movaps 0x3982(%rip),%xmm10 # 54f0 <_sk_callback_sse41+0x463> DB 69,15,88,250 ; addps %xmm10,%xmm15 DB 69,15,89,248 ; mulps %xmm8,%xmm15 - DB 68,15,40,45,61,57,0,0 ; movaps 0x393d(%rip),%xmm13 # 54b0 <_sk_callback_sse41+0x45d> + DB 68,15,40,45,130,57,0,0 ; movaps 0x3982(%rip),%xmm13 # 5500 <_sk_callback_sse41+0x473> DB 69,15,88,253 ; addps %xmm13,%xmm15 - DB 68,15,40,53,65,57,0,0 ; movaps 0x3941(%rip),%xmm14 # 54c0 <_sk_callback_sse41+0x46d> + DB 68,15,40,53,134,57,0,0 ; movaps 0x3986(%rip),%xmm14 # 5510 <_sk_callback_sse41+0x483> DB 69,15,88,198 ; addps %xmm14,%xmm8 DB 69,15,83,192 ; rcpps %xmm8,%xmm8 DB 69,15,89,199 ; mulps %xmm15,%xmm8 - DB 68,15,40,61,61,57,0,0 ; movaps 0x393d(%rip),%xmm15 # 54d0 <_sk_callback_sse41+0x47d> + DB 68,15,40,61,130,57,0,0 ; movaps 0x3982(%rip),%xmm15 # 5520 <_sk_callback_sse41+0x493> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 68,15,82,202 ; rsqrtps %xmm2,%xmm9 @@ -13622,7 +13617,7 @@ _sk_rgb_to_hsl_sse41 LABEL PROC DB 68,15,93,226 ; minps %xmm2,%xmm12 DB 65,15,40,203 ; movaps %xmm11,%xmm1 DB 65,15,92,204 ; subps %xmm12,%xmm1 - DB 68,15,40,53,135,56,0,0 ; movaps 0x3887(%rip),%xmm14 # 54e0 <_sk_callback_sse41+0x48d> + DB 68,15,40,53,204,56,0,0 ; movaps 0x38cc(%rip),%xmm14 # 5530 <_sk_callback_sse41+0x4a3> DB 68,15,94,241 ; divps %xmm1,%xmm14 DB 69,15,40,211 ; movaps %xmm11,%xmm10 DB 69,15,194,208,0 ; cmpeqps %xmm8,%xmm10 @@ -13631,27 +13626,27 @@ _sk_rgb_to_hsl_sse41 LABEL PROC DB 65,15,89,198 ; mulps %xmm14,%xmm0 DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 68,15,194,250,1 ; cmpltps %xmm2,%xmm15 - DB 68,15,84,61,110,56,0,0 ; andps 0x386e(%rip),%xmm15 # 54f0 <_sk_callback_sse41+0x49d> + DB 68,15,84,61,179,56,0,0 ; andps 0x38b3(%rip),%xmm15 # 5540 <_sk_callback_sse41+0x4b3> DB 68,15,88,248 ; addps %xmm0,%xmm15 DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,194,193,0 ; cmpeqps %xmm9,%xmm0 DB 65,15,92,208 ; subps %xmm8,%xmm2 DB 65,15,89,214 ; mulps %xmm14,%xmm2 - DB 68,15,40,45,97,56,0,0 ; movaps 0x3861(%rip),%xmm13 # 5500 <_sk_callback_sse41+0x4ad> + DB 68,15,40,45,166,56,0,0 ; movaps 0x38a6(%rip),%xmm13 # 5550 <_sk_callback_sse41+0x4c3> DB 65,15,88,213 ; addps %xmm13,%xmm2 DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 69,15,89,198 ; mulps %xmm14,%xmm8 - DB 68,15,88,5,93,56,0,0 ; addps 0x385d(%rip),%xmm8 # 5510 <_sk_callback_sse41+0x4bd> + DB 68,15,88,5,162,56,0,0 ; addps 0x38a2(%rip),%xmm8 # 5560 <_sk_callback_sse41+0x4d3> DB 102,68,15,56,20,194 ; blendvps %xmm0,%xmm2,%xmm8 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 102,69,15,56,20,199 ; blendvps %xmm0,%xmm15,%xmm8 - DB 68,15,89,5,85,56,0,0 ; mulps 0x3855(%rip),%xmm8 # 5520 <_sk_callback_sse41+0x4cd> + DB 68,15,89,5,154,56,0,0 ; mulps 0x389a(%rip),%xmm8 # 5570 <_sk_callback_sse41+0x4e3> DB 69,15,40,203 ; movaps %xmm11,%xmm9 DB 69,15,194,204,4 ; cmpneqps %xmm12,%xmm9 DB 69,15,84,193 ; andps %xmm9,%xmm8 DB 69,15,92,235 ; subps %xmm11,%xmm13 DB 69,15,88,220 ; addps %xmm12,%xmm11 - DB 15,40,5,73,56,0,0 ; movaps 0x3849(%rip),%xmm0 # 5530 <_sk_callback_sse41+0x4dd> + DB 15,40,5,142,56,0,0 ; movaps 0x388e(%rip),%xmm0 # 5580 <_sk_callback_sse41+0x4f3> DB 65,15,40,211 ; movaps %xmm11,%xmm2 DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,194,194,1 ; cmpltps %xmm2,%xmm0 @@ -13672,7 +13667,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,100,36,32 ; movaps %xmm4,0x20(%rsp) DB 15,41,92,36,16 ; movaps %xmm3,0x10(%rsp) DB 68,15,40,208 ; movaps %xmm0,%xmm10 - DB 68,15,40,13,11,56,0,0 ; movaps 0x380b(%rip),%xmm9 # 5540 <_sk_callback_sse41+0x4ed> + DB 68,15,40,13,80,56,0,0 ; movaps 0x3850(%rip),%xmm9 # 5590 <_sk_callback_sse41+0x503> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,194,194,2 ; cmpleps %xmm2,%xmm0 DB 15,40,217 ; movaps %xmm1,%xmm3 @@ -13685,19 +13680,19 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,20,36 ; movaps %xmm2,(%rsp) DB 69,15,88,192 ; addps %xmm8,%xmm8 DB 68,15,92,197 ; subps %xmm5,%xmm8 - DB 68,15,40,53,231,55,0,0 ; movaps 0x37e7(%rip),%xmm14 # 5550 <_sk_callback_sse41+0x4fd> + DB 68,15,40,53,44,56,0,0 ; movaps 0x382c(%rip),%xmm14 # 55a0 <_sk_callback_sse41+0x513> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 102,65,15,58,8,198,1 ; roundps $0x1,%xmm14,%xmm0 DB 68,15,92,240 ; subps %xmm0,%xmm14 - DB 68,15,40,29,224,55,0,0 ; movaps 0x37e0(%rip),%xmm11 # 5560 <_sk_callback_sse41+0x50d> + DB 68,15,40,29,37,56,0,0 ; movaps 0x3825(%rip),%xmm11 # 55b0 <_sk_callback_sse41+0x523> DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 15,40,245 ; movaps %xmm5,%xmm6 DB 65,15,92,240 ; subps %xmm8,%xmm6 - DB 15,40,61,217,55,0,0 ; movaps 0x37d9(%rip),%xmm7 # 5570 <_sk_callback_sse41+0x51d> + DB 15,40,61,30,56,0,0 ; movaps 0x381e(%rip),%xmm7 # 55c0 <_sk_callback_sse41+0x533> DB 69,15,40,238 ; movaps %xmm14,%xmm13 DB 68,15,89,239 ; mulps %xmm7,%xmm13 - DB 15,40,29,218,55,0,0 ; movaps 0x37da(%rip),%xmm3 # 5580 <_sk_callback_sse41+0x52d> + DB 15,40,29,31,56,0,0 ; movaps 0x381f(%rip),%xmm3 # 55d0 <_sk_callback_sse41+0x543> DB 68,15,40,227 ; movaps %xmm3,%xmm12 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 68,15,89,230 ; mulps %xmm6,%xmm12 @@ -13707,7 +13702,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 68,15,40,253 ; movaps %xmm5,%xmm15 DB 102,69,15,56,20,252 ; blendvps %xmm0,%xmm12,%xmm15 - DB 68,15,40,37,185,55,0,0 ; movaps 0x37b9(%rip),%xmm12 # 5590 <_sk_callback_sse41+0x53d> + DB 68,15,40,37,254,55,0,0 ; movaps 0x37fe(%rip),%xmm12 # 55e0 <_sk_callback_sse41+0x553> DB 65,15,40,196 ; movaps %xmm12,%xmm0 DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 68,15,89,238 ; mulps %xmm6,%xmm13 @@ -13741,7 +13736,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 65,15,40,198 ; movaps %xmm14,%xmm0 DB 15,40,20,36 ; movaps (%rsp),%xmm2 DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 - DB 68,15,88,21,50,55,0,0 ; addps 0x3732(%rip),%xmm10 # 55a0 <_sk_callback_sse41+0x54d> + DB 68,15,88,21,119,55,0,0 ; addps 0x3777(%rip),%xmm10 # 55f0 <_sk_callback_sse41+0x563> DB 102,65,15,58,8,194,1 ; roundps $0x1,%xmm10,%xmm0 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 69,15,194,218,2 ; cmpleps %xmm10,%xmm11 @@ -13787,13 +13782,13 @@ _sk_scale_1_float_sse41 LABEL PROC PUBLIC _sk_scale_u8_sse41 _sk_scale_u8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,52 ; jne 1f4b <_sk_scale_u8_sse41+0x3e> - DB 102,69,15,56,49,4,57 ; pmovzxbd (%r9,%rdi,1),%xmm8 - DB 102,68,15,219,5,137,54,0,0 ; pand 0x3689(%rip),%xmm8 # 55b0 <_sk_callback_sse41+0x55d> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,52 ; jne 1f56 <_sk_scale_u8_sse41+0x3e> + DB 102,69,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm8 + DB 102,68,15,219,5,206,54,0,0 ; pand 0x36ce(%rip),%xmm8 # 5600 <_sk_callback_sse41+0x573> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,141,54,0,0 ; mulps 0x368d(%rip),%xmm8 # 55c0 <_sk_callback_sse41+0x56d> + DB 68,15,89,5,210,54,0,0 ; mulps 0x36d2(%rip),%xmm8 # 5610 <_sk_callback_sse41+0x583> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -13801,23 +13796,23 @@ _sk_scale_u8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,216 ; movaps %xmm8,%xmm3 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,42 ; je 1f87 <_sk_scale_u8_sse41+0x7a> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,23 ; je 1f7a <_sk_scale_u8_sse41+0x6d> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,181 ; jne 1f1e <_sk_scale_u8_sse41+0x11> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,42 ; je 1f92 <_sk_scale_u8_sse41+0x7a> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,23 ; je 1f85 <_sk_scale_u8_sse41+0x6d> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,181 ; jne 1f29 <_sk_scale_u8_sse41+0x11> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,68,15,58,34,192,1 ; pinsrd $0x1,%eax,%xmm8 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,68,15,58,34,192,0 ; pinsrd $0x0,%eax,%xmm8 - DB 235,137 ; jmp 1f1e <_sk_scale_u8_sse41+0x11> + DB 235,137 ; jmp 1f29 <_sk_scale_u8_sse41+0x11> PUBLIC _sk_lerp_1_float_sse41 _sk_lerp_1_float_sse41 LABEL PROC @@ -13842,13 +13837,13 @@ _sk_lerp_1_float_sse41 LABEL PROC PUBLIC _sk_lerp_u8_sse41 _sk_lerp_u8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,72 ; jne 201f <_sk_lerp_u8_sse41+0x52> - DB 102,69,15,56,49,4,57 ; pmovzxbd (%r9,%rdi,1),%xmm8 - DB 102,68,15,219,5,233,53,0,0 ; pand 0x35e9(%rip),%xmm8 # 55d0 <_sk_callback_sse41+0x57d> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,72 ; jne 202a <_sk_lerp_u8_sse41+0x52> + DB 102,69,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm8 + DB 102,68,15,219,5,46,54,0,0 ; pand 0x362e(%rip),%xmm8 # 5620 <_sk_callback_sse41+0x593> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,237,53,0,0 ; mulps 0x35ed(%rip),%xmm8 # 55e0 <_sk_callback_sse41+0x58d> + DB 68,15,89,5,50,54,0,0 ; mulps 0x3632(%rip),%xmm8 # 5630 <_sk_callback_sse41+0x5a3> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -13863,42 +13858,42 @@ _sk_lerp_u8_sse41 LABEL PROC DB 15,88,223 ; addps %xmm7,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,42 ; je 205b <_sk_lerp_u8_sse41+0x8e> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,23 ; je 204e <_sk_lerp_u8_sse41+0x81> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,161 ; jne 1fde <_sk_lerp_u8_sse41+0x11> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,42 ; je 2066 <_sk_lerp_u8_sse41+0x8e> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,23 ; je 2059 <_sk_lerp_u8_sse41+0x81> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,161 ; jne 1fe9 <_sk_lerp_u8_sse41+0x11> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,68,15,58,34,192,1 ; pinsrd $0x1,%eax,%xmm8 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,68,15,58,34,192,0 ; pinsrd $0x0,%eax,%xmm8 - DB 233,114,255,255,255 ; jmpq 1fde <_sk_lerp_u8_sse41+0x11> + DB 233,114,255,255,255 ; jmpq 1fe9 <_sk_lerp_u8_sse41+0x11> PUBLIC _sk_lerp_565_sse41 _sk_lerp_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,152,0,0,0 ; jne 2112 <_sk_lerp_565_sse41+0xa6> - DB 102,69,15,56,51,12,121 ; pmovzxwd (%r9,%rdi,2),%xmm9 - DB 102,68,15,111,5,102,53,0,0 ; movdqa 0x3566(%rip),%xmm8 # 55f0 <_sk_callback_sse41+0x59d> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,152,0,0,0 ; jne 211d <_sk_lerp_565_sse41+0xa6> + DB 102,69,15,56,51,12,82 ; pmovzxwd (%r10,%rdx,2),%xmm9 + DB 102,68,15,111,5,171,53,0,0 ; movdqa 0x35ab(%rip),%xmm8 # 5640 <_sk_callback_sse41+0x5b3> DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,101,53,0,0 ; mulps 0x3565(%rip),%xmm8 # 5600 <_sk_callback_sse41+0x5ad> - DB 102,68,15,111,21,108,53,0,0 ; movdqa 0x356c(%rip),%xmm10 # 5610 <_sk_callback_sse41+0x5bd> + DB 68,15,89,5,170,53,0,0 ; mulps 0x35aa(%rip),%xmm8 # 5650 <_sk_callback_sse41+0x5c3> + DB 102,68,15,111,21,177,53,0,0 ; movdqa 0x35b1(%rip),%xmm10 # 5660 <_sk_callback_sse41+0x5d3> DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 - DB 68,15,89,21,107,53,0,0 ; mulps 0x356b(%rip),%xmm10 # 5620 <_sk_callback_sse41+0x5cd> - DB 102,68,15,219,13,114,53,0,0 ; pand 0x3572(%rip),%xmm9 # 5630 <_sk_callback_sse41+0x5dd> + DB 68,15,89,21,176,53,0,0 ; mulps 0x35b0(%rip),%xmm10 # 5670 <_sk_callback_sse41+0x5e3> + DB 102,68,15,219,13,183,53,0,0 ; pand 0x35b7(%rip),%xmm9 # 5680 <_sk_callback_sse41+0x5f3> DB 69,15,91,201 ; cvtdq2ps %xmm9,%xmm9 - DB 68,15,89,13,118,53,0,0 ; mulps 0x3576(%rip),%xmm9 # 5640 <_sk_callback_sse41+0x5ed> + DB 68,15,89,13,187,53,0,0 ; mulps 0x35bb(%rip),%xmm9 # 5690 <_sk_callback_sse41+0x603> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -13920,107 +13915,110 @@ _sk_lerp_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,216 ; movaps %xmm8,%xmm3 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,46 ; je 2152 <_sk_lerp_565_sse41+0xe6> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,27 ; je 2145 <_sk_lerp_565_sse41+0xd9> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 15,133,77,255,255,255 ; jne 2081 <_sk_lerp_565_sse41+0x15> - DB 65,15,183,68,121,4 ; movzwl 0x4(%r9,%rdi,2),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,46 ; je 215d <_sk_lerp_565_sse41+0xe6> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,27 ; je 2150 <_sk_lerp_565_sse41+0xd9> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 15,133,77,255,255,255 ; jne 208c <_sk_lerp_565_sse41+0x15> + DB 65,15,183,68,82,4 ; movzwl 0x4(%r10,%rdx,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,200,69 ; pshufd $0x45,%xmm8,%xmm9 - DB 65,15,183,68,121,2 ; movzwl 0x2(%r9,%rdi,2),%eax + DB 65,15,183,68,82,2 ; movzwl 0x2(%r10,%rdx,2),%eax DB 102,68,15,58,34,200,1 ; pinsrd $0x1,%eax,%xmm9 - DB 65,15,183,4,121 ; movzwl (%r9,%rdi,2),%eax + DB 65,15,183,4,82 ; movzwl (%r10,%rdx,2),%eax DB 102,68,15,58,34,200,0 ; pinsrd $0x0,%eax,%xmm9 - DB 233,30,255,255,255 ; jmpq 2081 <_sk_lerp_565_sse41+0x15> + DB 233,30,255,255,255 ; jmpq 208c <_sk_lerp_565_sse41+0x15> PUBLIC _sk_load_tables_sse41 _sk_load_tables_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,20,1,0,0 ; jne 2285 <_sk_load_tables_sse41+0x122> - DB 243,69,15,111,4,184 ; movdqu (%r8,%rdi,4),%xmm8 + DB 76,139,8 ; mov (%rax),%r9 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,24,1,0,0 ; jne 2294 <_sk_load_tables_sse41+0x126> + DB 243,69,15,111,4,145 ; movdqu (%r9,%rdx,4),%xmm8 + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx - DB 102,15,111,5,206,52,0,0 ; movdqa 0x34ce(%rip),%xmm0 # 5650 <_sk_callback_sse41+0x5fd> + DB 102,15,111,5,17,53,0,0 ; movdqa 0x3511(%rip),%xmm0 # 56a0 <_sk_callback_sse41+0x613> DB 102,65,15,219,192 ; pand %xmm8,%xmm0 - DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 102,73,15,126,194 ; movq %xmm0,%r10 + DB 69,15,182,218 ; movzbl %r10b,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 69,15,182,241 ; movzbl %r9b,%r14d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 69,15,182,216 ; movzbl %r8b,%r11d - DB 73,193,232,30 ; shr $0x1e,%r8 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 - DB 102,66,15,58,33,4,11,16 ; insertps $0x10,(%rbx,%r9,1),%xmm0 - DB 102,66,15,58,33,4,155,32 ; insertps $0x20,(%rbx,%r11,4),%xmm0 - DB 102,66,15,58,33,4,3,48 ; insertps $0x30,(%rbx,%r8,1),%xmm0 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 + DB 102,66,15,58,33,4,19,16 ; insertps $0x10,(%rbx,%r10,1),%xmm0 + DB 102,66,15,58,33,4,179,32 ; insertps $0x20,(%rbx,%r14,4),%xmm0 + DB 102,66,15,58,33,4,11,48 ; insertps $0x30,(%rbx,%r9,1),%xmm0 DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 - DB 102,15,56,0,13,137,52,0,0 ; pshufb 0x3489(%rip),%xmm1 # 5660 <_sk_callback_sse41+0x60d> - DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 + DB 102,15,56,0,13,204,52,0,0 ; pshufb 0x34cc(%rip),%xmm1 # 56b0 <_sk_callback_sse41+0x623> + DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 68,15,182,203 ; movzbl %bl,%r9d + DB 68,15,182,211 ; movzbl %bl,%r10d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 69,15,182,208 ; movzbl %r8b,%r10d - DB 73,193,232,30 ; shr $0x1e,%r8 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 102,65,15,58,33,12,30,16 ; insertps $0x10,(%r14,%rbx,1),%xmm1 - DB 243,67,15,16,20,150 ; movss (%r14,%r10,4),%xmm2 + DB 69,15,182,217 ; movzbl %r9b,%r11d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 243,67,15,16,12,151 ; movss (%r15,%r10,4),%xmm1 + DB 102,65,15,58,33,12,31,16 ; insertps $0x10,(%r15,%rbx,1),%xmm1 + DB 243,67,15,16,20,159 ; movss (%r15,%r11,4),%xmm2 DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1 - DB 243,67,15,16,20,6 ; movss (%r14,%r8,1),%xmm2 + DB 243,67,15,16,20,15 ; movss (%r15,%r9,1),%xmm2 DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 - DB 76,139,64,24 ; mov 0x18(%rax),%r8 + DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 - DB 102,15,56,0,21,69,52,0,0 ; pshufb 0x3445(%rip),%xmm2 # 5670 <_sk_callback_sse41+0x61d> + DB 102,15,56,0,21,136,52,0,0 ; pshufb 0x3488(%rip),%xmm2 # 56c0 <_sk_callback_sse41+0x633> DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 68,15,182,200 ; movzbl %al,%r9d + DB 68,15,182,208 ; movzbl %al,%r10d DB 72,193,232,30 ; shr $0x1e,%rax - DB 68,15,182,211 ; movzbl %bl,%r10d + DB 68,15,182,219 ; movzbl %bl,%r11d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 - DB 102,65,15,58,33,20,0,16 ; insertps $0x10,(%r8,%rax,1),%xmm2 - DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 + DB 102,65,15,58,33,20,1,16 ; insertps $0x10,(%r9,%rax,1),%xmm2 + DB 243,67,15,16,28,153 ; movss (%r9,%r11,4),%xmm3 DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 - DB 243,65,15,16,28,24 ; movss (%r8,%rbx,1),%xmm3 + DB 243,65,15,16,28,25 ; movss (%r9,%rbx,1),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8 DB 65,15,91,216 ; cvtdq2ps %xmm8,%xmm3 - DB 15,89,29,2,52,0,0 ; mulps 0x3402(%rip),%xmm3 # 5680 <_sk_callback_sse41+0x62d> + DB 15,89,29,69,52,0,0 ; mulps 0x3445(%rip),%xmm3 # 56d0 <_sk_callback_sse41+0x643> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax - DB 65,137,201 ; mov %ecx,%r9d - DB 65,128,225,3 ; and $0x3,%r9b + DB 69,137,194 ; mov %r8d,%r10d + DB 65,128,226,3 ; and $0x3,%r10b DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,38 ; je 22bd <_sk_load_tables_sse41+0x15a> - DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 22b4 <_sk_load_tables_sse41+0x151> - DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,208,254,255,255 ; jne 2177 <_sk_load_tables_sse41+0x14> - DB 102,65,15,110,68,184,8 ; movd 0x8(%r8,%rdi,4),%xmm0 + DB 65,128,250,1 ; cmp $0x1,%r10b + DB 116,38 ; je 22cc <_sk_load_tables_sse41+0x15e> + DB 65,128,250,2 ; cmp $0x2,%r10b + DB 116,23 ; je 22c3 <_sk_load_tables_sse41+0x155> + DB 65,128,250,3 ; cmp $0x3,%r10b + DB 15,133,204,254,255,255 ; jne 2182 <_sk_load_tables_sse41+0x14> + DB 102,65,15,110,68,145,8 ; movd 0x8(%r9,%rdx,4),%xmm0 DB 102,68,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm8 - DB 102,69,15,58,34,68,184,4,1 ; pinsrd $0x1,0x4(%r8,%rdi,4),%xmm8 - DB 102,69,15,58,34,4,184,0 ; pinsrd $0x0,(%r8,%rdi,4),%xmm8 - DB 233,173,254,255,255 ; jmpq 2177 <_sk_load_tables_sse41+0x14> + DB 102,69,15,58,34,68,145,4,1 ; pinsrd $0x1,0x4(%r9,%rdx,4),%xmm8 + DB 102,69,15,58,34,4,145,0 ; pinsrd $0x0,(%r9,%rdx,4),%xmm8 + DB 233,169,254,255,255 ; jmpq 2182 <_sk_load_tables_sse41+0x14> PUBLIC _sk_load_tables_u16_be_sse41 _sk_load_tables_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,95,1,0,0 ; jne 243f <_sk_load_tables_u16_be_sse41+0x175> - DB 102,67,15,16,4,72 ; movupd (%r8,%r9,2),%xmm0 - DB 243,67,15,111,76,72,16 ; movdqu 0x10(%r8,%r9,2),%xmm1 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,99,1,0,0 ; jne 2452 <_sk_load_tables_u16_be_sse41+0x179> + DB 102,67,15,16,4,81 ; movupd (%r9,%r10,2),%xmm0 + DB 243,67,15,111,76,81,16 ; movdqu 0x10(%r9,%r10,2),%xmm1 + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 102,68,15,40,200 ; movapd %xmm0,%xmm9 @@ -14029,52 +14027,52 @@ _sk_load_tables_u16_be_sse41 LABEL PROC DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,68,15,111,5,123,51,0,0 ; movdqa 0x337b(%rip),%xmm8 # 5690 <_sk_callback_sse41+0x63d> + DB 102,68,15,111,5,186,51,0,0 ; movdqa 0x33ba(%rip),%xmm8 # 56e0 <_sk_callback_sse41+0x653> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 - DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 102,73,15,126,194 ; movq %xmm0,%r10 + DB 69,15,182,218 ; movzbl %r10b,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 69,15,182,241 ; movzbl %r9b,%r14d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 69,15,182,216 ; movzbl %r8b,%r11d - DB 73,193,232,30 ; shr $0x1e,%r8 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 - DB 102,66,15,58,33,4,11,16 ; insertps $0x10,(%rbx,%r9,1),%xmm0 - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 + DB 102,66,15,58,33,4,19,16 ; insertps $0x10,(%rbx,%r10,1),%xmm0 + DB 243,66,15,16,20,179 ; movss (%rbx,%r14,4),%xmm2 DB 102,15,58,33,194,32 ; insertps $0x20,%xmm2,%xmm0 - DB 243,66,15,16,20,3 ; movss (%rbx,%r8,1),%xmm2 + DB 243,66,15,16,20,11 ; movss (%rbx,%r9,1),%xmm2 DB 102,15,58,33,194,48 ; insertps $0x30,%xmm2,%xmm0 - DB 102,15,56,0,13,42,51,0,0 ; pshufb 0x332a(%rip),%xmm1 # 56a0 <_sk_callback_sse41+0x64d> + DB 102,15,56,0,13,105,51,0,0 ; pshufb 0x3369(%rip),%xmm1 # 56f0 <_sk_callback_sse41+0x663> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 - DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 + DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 68,15,182,203 ; movzbl %bl,%r9d + DB 68,15,182,211 ; movzbl %bl,%r10d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 69,15,182,208 ; movzbl %r8b,%r10d - DB 73,193,232,30 ; shr $0x1e,%r8 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 102,65,15,58,33,12,30,16 ; insertps $0x10,(%r14,%rbx,1),%xmm1 - DB 243,67,15,16,20,150 ; movss (%r14,%r10,4),%xmm2 + DB 69,15,182,217 ; movzbl %r9b,%r11d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 243,67,15,16,12,151 ; movss (%r15,%r10,4),%xmm1 + DB 102,65,15,58,33,12,31,16 ; insertps $0x10,(%r15,%rbx,1),%xmm1 + DB 243,67,15,16,20,159 ; movss (%r15,%r11,4),%xmm2 DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1 - DB 243,67,15,16,20,6 ; movss (%r14,%r8,1),%xmm2 + DB 243,67,15,16,20,15 ; movss (%r15,%r9,1),%xmm2 DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 - DB 76,139,64,24 ; mov 0x18(%rax),%r8 + DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 102,65,15,56,51,208 ; pmovzxwd %xmm8,%xmm2 DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 68,15,182,200 ; movzbl %al,%r9d + DB 68,15,182,208 ; movzbl %al,%r10d DB 72,193,232,30 ; shr $0x1e,%rax - DB 68,15,182,211 ; movzbl %bl,%r10d + DB 68,15,182,219 ; movzbl %bl,%r11d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 - DB 102,65,15,58,33,20,0,16 ; insertps $0x10,(%r8,%rax,1),%xmm2 - DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 + DB 102,65,15,58,33,20,1,16 ; insertps $0x10,(%r9,%rax,1),%xmm2 + DB 243,67,15,16,28,153 ; movss (%r9,%r11,4),%xmm3 DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 - DB 243,65,15,16,28,24 ; movss (%r8,%rbx,1),%xmm3 + DB 243,65,15,16,28,25 ; movss (%r9,%rbx,1),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3 DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 @@ -14083,199 +14081,204 @@ _sk_load_tables_u16_be_sse41 LABEL PROC DB 102,65,15,235,216 ; por %xmm8,%xmm3 DB 102,15,56,51,219 ; pmovzxwd %xmm3,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,120,50,0,0 ; mulps 0x3278(%rip),%xmm3 # 56b0 <_sk_callback_sse41+0x65d> + DB 15,89,29,183,50,0,0 ; mulps 0x32b7(%rip),%xmm3 # 5700 <_sk_callback_sse41+0x673> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax - DB 242,67,15,16,4,72 ; movsd (%r8,%r9,2),%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,13 ; jne 2458 <_sk_load_tables_u16_be_sse41+0x18e> + DB 242,67,15,16,4,81 ; movsd (%r9,%r10,2),%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,13 ; jne 246b <_sk_load_tables_u16_be_sse41+0x192> DB 243,15,126,192 ; movq %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,149,254,255,255 ; jmpq 22ed <_sk_load_tables_u16_be_sse41+0x23> - DB 102,67,15,22,68,72,8 ; movhpd 0x8(%r8,%r9,2),%xmm0 + DB 233,145,254,255,255 ; jmpq 22fc <_sk_load_tables_u16_be_sse41+0x23> + DB 102,67,15,22,68,81,8 ; movhpd 0x8(%r9,%r10,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 15,130,128,254,255,255 ; jb 22ed <_sk_load_tables_u16_be_sse41+0x23> - DB 243,67,15,126,76,72,16 ; movq 0x10(%r8,%r9,2),%xmm1 - DB 233,116,254,255,255 ; jmpq 22ed <_sk_load_tables_u16_be_sse41+0x23> + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 15,130,124,254,255,255 ; jb 22fc <_sk_load_tables_u16_be_sse41+0x23> + DB 243,67,15,126,76,81,16 ; movq 0x10(%r9,%r10,2),%xmm1 + DB 233,112,254,255,255 ; jmpq 22fc <_sk_load_tables_u16_be_sse41+0x23> PUBLIC _sk_load_tables_rgb_u16_be_sse41 _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,79,1,0,0 ; jne 25da <_sk_load_tables_rgb_u16_be_sse41+0x161> - DB 243,67,15,111,20,72 ; movdqu (%r8,%r9,2),%xmm2 - DB 243,67,15,111,76,72,8 ; movdqu 0x8(%r8,%r9,2),%xmm1 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,82 ; lea (%rdx,%rdx,2),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,83,1,0,0 ; jne 25f1 <_sk_load_tables_rgb_u16_be_sse41+0x165> + DB 243,67,15,111,20,81 ; movdqu (%r9,%r10,2),%xmm2 + DB 243,67,15,111,76,81,8 ; movdqu 0x8(%r9,%r10,2),%xmm1 DB 102,15,115,217,4 ; psrldq $0x4,%xmm1 DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 DB 102,65,15,115,217,6 ; psrldq $0x6,%xmm9 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 - DB 102,68,15,111,5,241,49,0,0 ; movdqa 0x31f1(%rip),%xmm8 # 56c0 <_sk_callback_sse41+0x66d> + DB 102,68,15,111,5,44,50,0,0 ; movdqa 0x322c(%rip),%xmm8 # 5710 <_sk_callback_sse41+0x683> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 - DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 102,73,15,126,194 ; movq %xmm0,%r10 + DB 69,15,182,218 ; movzbl %r10b,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 69,15,182,241 ; movzbl %r9b,%r14d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 69,15,182,216 ; movzbl %r8b,%r11d - DB 73,193,232,30 ; shr $0x1e,%r8 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 - DB 102,66,15,58,33,4,11,16 ; insertps $0x10,(%rbx,%r9,1),%xmm0 - DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 + DB 102,66,15,58,33,4,19,16 ; insertps $0x10,(%rbx,%r10,1),%xmm0 + DB 243,66,15,16,28,179 ; movss (%rbx,%r14,4),%xmm3 DB 102,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm0 - DB 243,66,15,16,28,3 ; movss (%rbx,%r8,1),%xmm3 + DB 243,66,15,16,28,11 ; movss (%rbx,%r9,1),%xmm3 DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0 - DB 102,15,56,0,13,160,49,0,0 ; pshufb 0x31a0(%rip),%xmm1 # 56d0 <_sk_callback_sse41+0x67d> + DB 102,15,56,0,13,219,49,0,0 ; pshufb 0x31db(%rip),%xmm1 # 5720 <_sk_callback_sse41+0x693> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 - DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 + DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 68,15,182,203 ; movzbl %bl,%r9d + DB 68,15,182,211 ; movzbl %bl,%r10d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 69,15,182,208 ; movzbl %r8b,%r10d - DB 73,193,232,30 ; shr $0x1e,%r8 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 102,65,15,58,33,12,30,16 ; insertps $0x10,(%r14,%rbx,1),%xmm1 - DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 69,15,182,217 ; movzbl %r9b,%r11d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 243,67,15,16,12,151 ; movss (%r15,%r10,4),%xmm1 + DB 102,65,15,58,33,12,31,16 ; insertps $0x10,(%r15,%rbx,1),%xmm1 + DB 243,67,15,16,28,159 ; movss (%r15,%r11,4),%xmm3 DB 102,15,58,33,203,32 ; insertps $0x20,%xmm3,%xmm1 - DB 243,67,15,16,28,6 ; movss (%r14,%r8,1),%xmm3 + DB 243,67,15,16,28,15 ; movss (%r15,%r9,1),%xmm3 DB 102,15,58,33,203,48 ; insertps $0x30,%xmm3,%xmm1 - DB 76,139,64,24 ; mov 0x18(%rax),%r8 + DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 102,65,15,105,209 ; punpckhwd %xmm9,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 DB 102,15,56,51,210 ; pmovzxwd %xmm2,%xmm2 DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 68,15,182,200 ; movzbl %al,%r9d + DB 68,15,182,208 ; movzbl %al,%r10d DB 72,193,232,30 ; shr $0x1e,%rax - DB 68,15,182,211 ; movzbl %bl,%r10d + DB 68,15,182,219 ; movzbl %bl,%r11d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 - DB 102,65,15,58,33,20,0,16 ; insertps $0x10,(%r8,%rax,1),%xmm2 - DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 + DB 102,65,15,58,33,20,1,16 ; insertps $0x10,(%r9,%rax,1),%xmm2 + DB 243,67,15,16,28,153 ; movss (%r9,%r11,4),%xmm3 DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 - DB 243,65,15,16,28,24 ; movss (%r8,%rbx,1),%xmm3 + DB 243,65,15,16,28,25 ; movss (%r9,%rbx,1),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,11,49,0,0 ; movaps 0x310b(%rip),%xmm3 # 56e0 <_sk_callback_sse41+0x68d> + DB 15,40,29,70,49,0,0 ; movaps 0x3146(%rip),%xmm3 # 5730 <_sk_callback_sse41+0x6a3> DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax - DB 102,67,15,110,20,72 ; movd (%r8,%r9,2),%xmm2 - DB 102,67,15,196,84,72,4,2 ; pinsrw $0x2,0x4(%r8,%r9,2),%xmm2 + DB 102,67,15,110,20,81 ; movd (%r9,%r10,2),%xmm2 + DB 102,67,15,196,84,81,4,2 ; pinsrw $0x2,0x4(%r9,%r10,2),%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,14 ; jne 2600 <_sk_load_tables_rgb_u16_be_sse41+0x187> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,14 ; jne 2617 <_sk_load_tables_rgb_u16_be_sse41+0x18b> DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 - DB 233,177,254,255,255 ; jmpq 24b1 <_sk_load_tables_rgb_u16_be_sse41+0x38> - DB 102,71,15,110,76,72,6 ; movd 0x6(%r8,%r9,2),%xmm9 - DB 102,71,15,196,76,72,10,2 ; pinsrw $0x2,0xa(%r8,%r9,2),%xmm9 + DB 233,173,254,255,255 ; jmpq 24c4 <_sk_load_tables_rgb_u16_be_sse41+0x38> + DB 102,71,15,110,76,81,6 ; movd 0x6(%r9,%r10,2),%xmm9 + DB 102,71,15,196,76,81,10,2 ; pinsrw $0x2,0xa(%r9,%r10,2),%xmm9 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,24 ; jb 2631 <_sk_load_tables_rgb_u16_be_sse41+0x1b8> - DB 102,67,15,110,76,72,12 ; movd 0xc(%r8,%r9,2),%xmm1 - DB 102,67,15,196,76,72,16,2 ; pinsrw $0x2,0x10(%r8,%r9,2),%xmm1 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,24 ; jb 2648 <_sk_load_tables_rgb_u16_be_sse41+0x1bc> + DB 102,67,15,110,76,81,12 ; movd 0xc(%r9,%r10,2),%xmm1 + DB 102,67,15,196,76,81,16,2 ; pinsrw $0x2,0x10(%r9,%r10,2),%xmm1 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,128,254,255,255 ; jmpq 24b1 <_sk_load_tables_rgb_u16_be_sse41+0x38> + DB 233,124,254,255,255 ; jmpq 24c4 <_sk_load_tables_rgb_u16_be_sse41+0x38> DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,119,254,255,255 ; jmpq 24b1 <_sk_load_tables_rgb_u16_be_sse41+0x38> + DB 233,115,254,255,255 ; jmpq 24c4 <_sk_load_tables_rgb_u16_be_sse41+0x38> PUBLIC _sk_byte_tables_sse41 _sk_byte_tables_sse41 LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,167,48,0,0 ; movaps 0x30a7(%rip),%xmm8 # 56f0 <_sk_callback_sse41+0x69d> + DB 68,15,40,5,222,48,0,0 ; movaps 0x30de(%rip),%xmm8 # 5740 <_sk_callback_sse41+0x6b3> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 - DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,139,56 ; mov (%rax),%r15 - DB 76,139,112,8 ; mov 0x8(%rax),%r14 - DB 102,67,15,58,32,4,31,0 ; pinsrb $0x0,(%r15,%r11,1),%xmm0 - DB 102,67,15,58,32,4,23,1 ; pinsrb $0x1,(%r15,%r10,1),%xmm0 - DB 67,15,182,28,15 ; movzbl (%r15,%r9,1),%ebx + DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 76,139,32 ; mov (%rax),%r12 + DB 76,139,120,8 ; mov 0x8(%rax),%r15 + DB 102,67,15,58,32,4,52,0 ; pinsrb $0x0,(%r12,%r14,1),%xmm0 + DB 102,67,15,58,32,4,28,1 ; pinsrb $0x1,(%r12,%r11,1),%xmm0 + DB 67,15,182,28,20 ; movzbl (%r12,%r10,1),%ebx DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0 - DB 67,15,182,28,7 ; movzbl (%r15,%r8,1),%ebx + DB 67,15,182,28,12 ; movzbl (%r12,%r9,1),%ebx DB 102,15,58,32,195,3 ; pinsrb $0x3,%ebx,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,88,48,0,0 ; movaps 0x3058(%rip),%xmm9 # 5700 <_sk_callback_sse41+0x6ad> + DB 68,15,40,13,143,48,0,0 ; movaps 0x308f(%rip),%xmm9 # 5750 <_sk_callback_sse41+0x6c3> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 DB 102,72,15,58,22,203,1 ; pextrq $0x1,%xmm1,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 69,137,202 ; mov %r9d,%r10d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 102,67,15,58,32,12,22,0 ; pinsrb $0x0,(%r14,%r10,1),%xmm1 - DB 102,67,15,58,32,12,14,1 ; pinsrb $0x1,(%r14,%r9,1),%xmm1 - DB 71,15,182,4,6 ; movzbl (%r14,%r8,1),%r8d - DB 102,65,15,58,32,200,2 ; pinsrb $0x2,%r8d,%xmm1 - DB 65,15,182,28,30 ; movzbl (%r14,%rbx,1),%ebx + DB 102,73,15,126,202 ; movq %xmm1,%r10 + DB 69,137,211 ; mov %r10d,%r11d + DB 73,193,234,32 ; shr $0x20,%r10 + DB 102,67,15,58,32,12,31,0 ; pinsrb $0x0,(%r15,%r11,1),%xmm1 + DB 102,67,15,58,32,12,23,1 ; pinsrb $0x1,(%r15,%r10,1),%xmm1 + DB 71,15,182,12,15 ; movzbl (%r15,%r9,1),%r9d + DB 102,65,15,58,32,201,2 ; pinsrb $0x2,%r9d,%xmm1 + DB 65,15,182,28,31 ; movzbl (%r15,%rbx,1),%ebx DB 102,15,58,32,203,3 ; pinsrb $0x3,%ebx,%xmm1 DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 76,139,80,16 ; mov 0x10(%rax),%r10 + DB 76,139,88,16 ; mov 0x10(%rax),%r11 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2 - DB 102,73,15,58,22,208,1 ; pextrq $0x1,%xmm2,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 + DB 102,73,15,58,22,209,1 ; pextrq $0x1,%xmm2,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 65,137,219 ; mov %ebx,%r11d + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,67,15,58,32,20,26,0 ; pinsrb $0x0,(%r10,%r11,1),%xmm2 - DB 102,65,15,58,32,20,26,1 ; pinsrb $0x1,(%r10,%rbx,1),%xmm2 - DB 67,15,182,28,10 ; movzbl (%r10,%r9,1),%ebx + DB 102,67,15,58,32,20,51,0 ; pinsrb $0x0,(%r11,%r14,1),%xmm2 + DB 102,65,15,58,32,20,27,1 ; pinsrb $0x1,(%r11,%rbx,1),%xmm2 + DB 67,15,182,28,19 ; movzbl (%r11,%r10,1),%ebx DB 102,15,58,32,211,2 ; pinsrb $0x2,%ebx,%xmm2 - DB 67,15,182,28,2 ; movzbl (%r10,%r8,1),%ebx + DB 67,15,182,28,11 ; movzbl (%r11,%r9,1),%ebx DB 102,15,58,32,211,3 ; pinsrb $0x3,%ebx,%xmm2 DB 102,15,56,49,210 ; pmovzxbd %xmm2,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 102,15,91,219 ; cvtps2dq %xmm3,%xmm3 DB 102,72,15,58,22,219,1 ; pextrq $0x1,%xmm3,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,32 ; shr $0x20,%rbx DB 102,72,15,126,216 ; movq %xmm3,%rax - DB 65,137,194 ; mov %eax,%r10d + DB 65,137,195 ; mov %eax,%r11d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,67,15,58,32,28,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm3 - DB 102,65,15,58,32,28,1,1 ; pinsrb $0x1,(%r9,%rax,1),%xmm3 - DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 102,67,15,58,32,28,26,0 ; pinsrb $0x0,(%r10,%r11,1),%xmm3 + DB 102,65,15,58,32,28,2,1 ; pinsrb $0x1,(%r10,%rax,1),%xmm3 + DB 67,15,182,4,10 ; movzbl (%r10,%r9,1),%eax DB 102,15,58,32,216,2 ; pinsrb $0x2,%eax,%xmm3 - DB 65,15,182,4,25 ; movzbl (%r9,%rbx,1),%eax + DB 65,15,182,4,26 ; movzbl (%r10,%rbx,1),%eax DB 102,15,58,32,216,3 ; pinsrb $0x3,%eax,%xmm3 DB 102,15,56,49,219 ; pmovzxbd %xmm3,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 DB 65,15,89,217 ; mulps %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -14284,80 +14287,83 @@ PUBLIC _sk_byte_tables_rgb_sse41 _sk_byte_tables_rgb_sse41 LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,139,64,24 ; mov 0x18(%rax),%r8d - DB 65,255,200 ; dec %r8d - DB 102,69,15,110,192 ; movd %r8d,%xmm8 + DB 68,139,72,24 ; mov 0x18(%rax),%r9d + DB 65,255,201 ; dec %r9d + DB 102,69,15,110,193 ; movd %r9d,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 - DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,198 ; mov %r8,%r14 - DB 73,193,238,32 ; shr $0x20,%r14 - DB 76,139,56 ; mov (%rax),%r15 - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 102,67,15,58,32,4,31,0 ; pinsrb $0x0,(%r15,%r11,1),%xmm0 - DB 102,67,15,58,32,4,55,1 ; pinsrb $0x1,(%r15,%r14,1),%xmm0 - DB 67,15,182,28,15 ; movzbl (%r15,%r9,1),%ebx + DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 77,137,203 ; mov %r9,%r11 + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,206 ; mov %r9d,%r14d + DB 77,137,207 ; mov %r9,%r15 + DB 73,193,239,32 ; shr $0x20,%r15 + DB 76,139,32 ; mov (%rax),%r12 + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 102,67,15,58,32,4,52,0 ; pinsrb $0x0,(%r12,%r14,1),%xmm0 + DB 102,67,15,58,32,4,60,1 ; pinsrb $0x1,(%r12,%r15,1),%xmm0 + DB 67,15,182,28,20 ; movzbl (%r12,%r10,1),%ebx DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0 - DB 67,15,182,28,23 ; movzbl (%r15,%r10,1),%ebx + DB 67,15,182,28,28 ; movzbl (%r12,%r11,1),%ebx DB 102,15,58,32,195,3 ; pinsrb $0x3,%ebx,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,212,46,0,0 ; movaps 0x2ed4(%rip),%xmm9 # 5710 <_sk_callback_sse41+0x6bd> + DB 68,15,40,13,7,47,0,0 ; movaps 0x2f07(%rip),%xmm9 # 5760 <_sk_callback_sse41+0x6d3> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 DB 102,72,15,58,22,203,1 ; pextrq $0x1,%xmm1,%rbx - DB 65,137,217 ; mov %ebx,%r9d + DB 65,137,218 ; mov %ebx,%r10d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,202 ; movq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,58,32,12,24,0 ; pinsrb $0x0,(%r8,%r11,1),%xmm1 - DB 102,67,15,58,32,12,16,1 ; pinsrb $0x1,(%r8,%r10,1),%xmm1 - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 102,65,15,58,32,201,2 ; pinsrb $0x2,%r9d,%xmm1 - DB 65,15,182,28,24 ; movzbl (%r8,%rbx,1),%ebx + DB 102,73,15,126,203 ; movq %xmm1,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,67,15,58,32,12,49,0 ; pinsrb $0x0,(%r9,%r14,1),%xmm1 + DB 102,67,15,58,32,12,25,1 ; pinsrb $0x1,(%r9,%r11,1),%xmm1 + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 102,65,15,58,32,202,2 ; pinsrb $0x2,%r10d,%xmm1 + DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx DB 102,15,58,32,203,3 ; pinsrb $0x3,%ebx,%xmm1 DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2 DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,32 ; shr $0x20,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 65,137,194 ; mov %eax,%r10d + DB 65,137,195 ; mov %eax,%r11d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,67,15,58,32,20,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm2 - DB 102,65,15,58,32,20,1,1 ; pinsrb $0x1,(%r9,%rax,1),%xmm2 - DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 102,67,15,58,32,20,26,0 ; pinsrb $0x0,(%r10,%r11,1),%xmm2 + DB 102,65,15,58,32,20,2,1 ; pinsrb $0x1,(%r10,%rax,1),%xmm2 + DB 67,15,182,4,10 ; movzbl (%r10,%r9,1),%eax DB 102,15,58,32,208,2 ; pinsrb $0x2,%eax,%xmm2 - DB 65,15,182,4,25 ; movzbl (%r9,%rbx,1),%eax + DB 65,15,182,4,26 ; movzbl (%r10,%rbx,1),%eax DB 102,15,58,32,208,3 ; pinsrb $0x3,%eax,%xmm2 DB 102,15,56,49,210 ; pmovzxbd %xmm2,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax PUBLIC _sk_table_r_sse41 _sk_table_r_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -14366,24 +14372,26 @@ _sk_table_r_sse41 LABEL PROC DB 68,15,89,192 ; mulps %xmm0,%xmm8 DB 102,65,15,91,192 ; cvtps2dq %xmm8,%xmm0 DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0 - DB 102,67,15,58,33,4,144,16 ; insertps $0x10,(%r8,%r10,4),%xmm0 - DB 243,71,15,16,4,136 ; movss (%r8,%r9,4),%xmm8 + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,65,15,16,4,153 ; movss (%r9,%rbx,4),%xmm0 + DB 102,67,15,58,33,4,153,16 ; insertps $0x10,(%r9,%r11,4),%xmm0 + DB 243,71,15,16,4,145 ; movss (%r9,%r10,4),%xmm8 DB 102,65,15,58,33,192,32 ; insertps $0x20,%xmm8,%xmm0 - DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,69,15,16,4,129 ; movss (%r9,%rax,4),%xmm8 DB 102,65,15,58,33,192,48 ; insertps $0x30,%xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_table_g_sse41 _sk_table_g_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -14392,24 +14400,26 @@ _sk_table_g_sse41 LABEL PROC DB 68,15,89,193 ; mulps %xmm1,%xmm8 DB 102,65,15,91,200 ; cvtps2dq %xmm8,%xmm1 DB 102,72,15,58,22,200,1 ; pextrq $0x1,%xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,202 ; movq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,67,15,16,12,152 ; movss (%r8,%r11,4),%xmm1 - DB 102,67,15,58,33,12,144,16 ; insertps $0x10,(%r8,%r10,4),%xmm1 - DB 243,71,15,16,4,136 ; movss (%r8,%r9,4),%xmm8 + DB 102,73,15,126,203 ; movq %xmm1,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,65,15,16,12,153 ; movss (%r9,%rbx,4),%xmm1 + DB 102,67,15,58,33,12,153,16 ; insertps $0x10,(%r9,%r11,4),%xmm1 + DB 243,71,15,16,4,145 ; movss (%r9,%r10,4),%xmm8 DB 102,65,15,58,33,200,32 ; insertps $0x20,%xmm8,%xmm1 - DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,69,15,16,4,129 ; movss (%r9,%rax,4),%xmm8 DB 102,65,15,58,33,200,48 ; insertps $0x30,%xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_table_b_sse41 _sk_table_b_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -14418,24 +14428,26 @@ _sk_table_b_sse41 LABEL PROC DB 68,15,89,194 ; mulps %xmm2,%xmm8 DB 102,65,15,91,208 ; cvtps2dq %xmm8,%xmm2 DB 102,72,15,58,22,208,1 ; pextrq $0x1,%xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,210 ; movq %xmm2,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2 - DB 102,67,15,58,33,20,144,16 ; insertps $0x10,(%r8,%r10,4),%xmm2 - DB 243,71,15,16,4,136 ; movss (%r8,%r9,4),%xmm8 + DB 102,73,15,126,211 ; movq %xmm2,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,65,15,16,20,153 ; movss (%r9,%rbx,4),%xmm2 + DB 102,67,15,58,33,20,153,16 ; insertps $0x10,(%r9,%r11,4),%xmm2 + DB 243,71,15,16,4,145 ; movss (%r9,%r10,4),%xmm8 DB 102,65,15,58,33,208,32 ; insertps $0x20,%xmm8,%xmm2 - DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,69,15,16,4,129 ; movss (%r9,%rax,4),%xmm8 DB 102,65,15,58,33,208,48 ; insertps $0x30,%xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_table_a_sse41 _sk_table_a_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -14444,18 +14456,19 @@ _sk_table_a_sse41 LABEL PROC DB 68,15,89,195 ; mulps %xmm3,%xmm8 DB 102,65,15,91,216 ; cvtps2dq %xmm8,%xmm3 DB 102,72,15,58,22,216,1 ; pextrq $0x1,%xmm3,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,218 ; movq %xmm3,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,67,15,16,28,152 ; movss (%r8,%r11,4),%xmm3 - DB 102,67,15,58,33,28,144,16 ; insertps $0x10,(%r8,%r10,4),%xmm3 - DB 243,71,15,16,4,136 ; movss (%r8,%r9,4),%xmm8 + DB 102,73,15,126,219 ; movq %xmm3,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,65,15,16,28,153 ; movss (%r9,%rbx,4),%xmm3 + DB 102,67,15,58,33,28,153,16 ; insertps $0x10,(%r9,%r11,4),%xmm3 + DB 243,71,15,16,4,145 ; movss (%r9,%r10,4),%xmm8 DB 102,65,15,58,33,216,32 ; insertps $0x20,%xmm8,%xmm3 - DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,69,15,16,4,129 ; movss (%r9,%rax,4),%xmm8 DB 102,65,15,58,33,216,48 ; insertps $0x30,%xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_parametric_r_sse41 @@ -14479,31 +14492,31 @@ _sk_parametric_r_sse41 LABEL PROC DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,194 ; cvtdq2ps %xmm10,%xmm8 - DB 68,15,89,5,38,44,0,0 ; mulps 0x2c26(%rip),%xmm8 # 5720 <_sk_callback_sse41+0x6cd> - DB 68,15,84,21,46,44,0,0 ; andps 0x2c2e(%rip),%xmm10 # 5730 <_sk_callback_sse41+0x6dd> - DB 68,15,86,21,54,44,0,0 ; orps 0x2c36(%rip),%xmm10 # 5740 <_sk_callback_sse41+0x6ed> - DB 68,15,88,5,62,44,0,0 ; addps 0x2c3e(%rip),%xmm8 # 5750 <_sk_callback_sse41+0x6fd> - DB 68,15,40,37,70,44,0,0 ; movaps 0x2c46(%rip),%xmm12 # 5760 <_sk_callback_sse41+0x70d> + DB 68,15,89,5,79,44,0,0 ; mulps 0x2c4f(%rip),%xmm8 # 5770 <_sk_callback_sse41+0x6e3> + DB 68,15,84,21,87,44,0,0 ; andps 0x2c57(%rip),%xmm10 # 5780 <_sk_callback_sse41+0x6f3> + DB 68,15,86,21,95,44,0,0 ; orps 0x2c5f(%rip),%xmm10 # 5790 <_sk_callback_sse41+0x703> + DB 68,15,88,5,103,44,0,0 ; addps 0x2c67(%rip),%xmm8 # 57a0 <_sk_callback_sse41+0x713> + DB 68,15,40,37,111,44,0,0 ; movaps 0x2c6f(%rip),%xmm12 # 57b0 <_sk_callback_sse41+0x723> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 68,15,88,21,70,44,0,0 ; addps 0x2c46(%rip),%xmm10 # 5770 <_sk_callback_sse41+0x71d> - DB 68,15,40,37,78,44,0,0 ; movaps 0x2c4e(%rip),%xmm12 # 5780 <_sk_callback_sse41+0x72d> + DB 68,15,88,21,111,44,0,0 ; addps 0x2c6f(%rip),%xmm10 # 57c0 <_sk_callback_sse41+0x733> + DB 68,15,40,37,119,44,0,0 ; movaps 0x2c77(%rip),%xmm12 # 57d0 <_sk_callback_sse41+0x743> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 DB 69,15,89,195 ; mulps %xmm11,%xmm8 DB 102,69,15,58,8,208,1 ; roundps $0x1,%xmm8,%xmm10 DB 69,15,40,216 ; movaps %xmm8,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,5,59,44,0,0 ; addps 0x2c3b(%rip),%xmm8 # 5790 <_sk_callback_sse41+0x73d> - DB 68,15,40,21,67,44,0,0 ; movaps 0x2c43(%rip),%xmm10 # 57a0 <_sk_callback_sse41+0x74d> + DB 68,15,88,5,100,44,0,0 ; addps 0x2c64(%rip),%xmm8 # 57e0 <_sk_callback_sse41+0x753> + DB 68,15,40,21,108,44,0,0 ; movaps 0x2c6c(%rip),%xmm10 # 57f0 <_sk_callback_sse41+0x763> DB 69,15,89,211 ; mulps %xmm11,%xmm10 DB 69,15,92,194 ; subps %xmm10,%xmm8 - DB 68,15,40,21,67,44,0,0 ; movaps 0x2c43(%rip),%xmm10 # 57b0 <_sk_callback_sse41+0x75d> + DB 68,15,40,21,108,44,0,0 ; movaps 0x2c6c(%rip),%xmm10 # 5800 <_sk_callback_sse41+0x773> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 68,15,40,29,71,44,0,0 ; movaps 0x2c47(%rip),%xmm11 # 57c0 <_sk_callback_sse41+0x76d> + DB 68,15,40,29,112,44,0,0 ; movaps 0x2c70(%rip),%xmm11 # 5810 <_sk_callback_sse41+0x783> DB 69,15,94,218 ; divps %xmm10,%xmm11 DB 69,15,88,216 ; addps %xmm8,%xmm11 - DB 68,15,89,29,71,44,0,0 ; mulps 0x2c47(%rip),%xmm11 # 57d0 <_sk_callback_sse41+0x77d> + DB 68,15,89,29,112,44,0,0 ; mulps 0x2c70(%rip),%xmm11 # 5820 <_sk_callback_sse41+0x793> DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 DB 243,68,15,16,64,20 ; movss 0x14(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 @@ -14511,7 +14524,7 @@ _sk_parametric_r_sse41 LABEL PROC DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,46,44,0,0 ; minps 0x2c2e(%rip),%xmm8 # 57e0 <_sk_callback_sse41+0x78d> + DB 68,15,93,5,87,44,0,0 ; minps 0x2c57(%rip),%xmm8 # 5830 <_sk_callback_sse41+0x7a3> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14539,31 +14552,31 @@ _sk_parametric_g_sse41 LABEL PROC DB 68,15,88,217 ; addps %xmm1,%xmm11 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 68,15,89,37,207,43,0,0 ; mulps 0x2bcf(%rip),%xmm12 # 57f0 <_sk_callback_sse41+0x79d> - DB 68,15,84,29,215,43,0,0 ; andps 0x2bd7(%rip),%xmm11 # 5800 <_sk_callback_sse41+0x7ad> - DB 68,15,86,29,223,43,0,0 ; orps 0x2bdf(%rip),%xmm11 # 5810 <_sk_callback_sse41+0x7bd> - DB 68,15,88,37,231,43,0,0 ; addps 0x2be7(%rip),%xmm12 # 5820 <_sk_callback_sse41+0x7cd> - DB 15,40,13,240,43,0,0 ; movaps 0x2bf0(%rip),%xmm1 # 5830 <_sk_callback_sse41+0x7dd> + DB 68,15,89,37,248,43,0,0 ; mulps 0x2bf8(%rip),%xmm12 # 5840 <_sk_callback_sse41+0x7b3> + DB 68,15,84,29,0,44,0,0 ; andps 0x2c00(%rip),%xmm11 # 5850 <_sk_callback_sse41+0x7c3> + DB 68,15,86,29,8,44,0,0 ; orps 0x2c08(%rip),%xmm11 # 5860 <_sk_callback_sse41+0x7d3> + DB 68,15,88,37,16,44,0,0 ; addps 0x2c10(%rip),%xmm12 # 5870 <_sk_callback_sse41+0x7e3> + DB 15,40,13,25,44,0,0 ; movaps 0x2c19(%rip),%xmm1 # 5880 <_sk_callback_sse41+0x7f3> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,88,29,240,43,0,0 ; addps 0x2bf0(%rip),%xmm11 # 5840 <_sk_callback_sse41+0x7ed> - DB 15,40,13,249,43,0,0 ; movaps 0x2bf9(%rip),%xmm1 # 5850 <_sk_callback_sse41+0x7fd> + DB 68,15,88,29,25,44,0,0 ; addps 0x2c19(%rip),%xmm11 # 5890 <_sk_callback_sse41+0x803> + DB 15,40,13,34,44,0,0 ; movaps 0x2c22(%rip),%xmm1 # 58a0 <_sk_callback_sse41+0x813> DB 65,15,94,203 ; divps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10 DB 69,15,40,220 ; movaps %xmm12,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,37,230,43,0,0 ; addps 0x2be6(%rip),%xmm12 # 5860 <_sk_callback_sse41+0x80d> - DB 15,40,13,239,43,0,0 ; movaps 0x2bef(%rip),%xmm1 # 5870 <_sk_callback_sse41+0x81d> + DB 68,15,88,37,15,44,0,0 ; addps 0x2c0f(%rip),%xmm12 # 58b0 <_sk_callback_sse41+0x823> + DB 15,40,13,24,44,0,0 ; movaps 0x2c18(%rip),%xmm1 # 58c0 <_sk_callback_sse41+0x833> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,40,21,239,43,0,0 ; movaps 0x2bef(%rip),%xmm10 # 5880 <_sk_callback_sse41+0x82d> + DB 68,15,40,21,24,44,0,0 ; movaps 0x2c18(%rip),%xmm10 # 58d0 <_sk_callback_sse41+0x843> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,13,244,43,0,0 ; movaps 0x2bf4(%rip),%xmm1 # 5890 <_sk_callback_sse41+0x83d> + DB 15,40,13,29,44,0,0 ; movaps 0x2c1d(%rip),%xmm1 # 58e0 <_sk_callback_sse41+0x853> DB 65,15,94,202 ; divps %xmm10,%xmm1 DB 65,15,88,204 ; addps %xmm12,%xmm1 - DB 15,89,13,245,43,0,0 ; mulps 0x2bf5(%rip),%xmm1 # 58a0 <_sk_callback_sse41+0x84d> + DB 15,89,13,30,44,0,0 ; mulps 0x2c1e(%rip),%xmm1 # 58f0 <_sk_callback_sse41+0x863> DB 102,68,15,91,209 ; cvtps2dq %xmm1,%xmm10 DB 243,15,16,72,20 ; movss 0x14(%rax),%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 @@ -14571,7 +14584,7 @@ _sk_parametric_g_sse41 LABEL PROC DB 102,65,15,56,20,201 ; blendvps %xmm0,%xmm9,%xmm1 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,95,200 ; maxps %xmm0,%xmm1 - DB 15,93,13,224,43,0,0 ; minps 0x2be0(%rip),%xmm1 # 58b0 <_sk_callback_sse41+0x85d> + DB 15,93,13,9,44,0,0 ; minps 0x2c09(%rip),%xmm1 # 5900 <_sk_callback_sse41+0x873> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14599,31 +14612,31 @@ _sk_parametric_b_sse41 LABEL PROC DB 68,15,88,218 ; addps %xmm2,%xmm11 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 68,15,89,37,129,43,0,0 ; mulps 0x2b81(%rip),%xmm12 # 58c0 <_sk_callback_sse41+0x86d> - DB 68,15,84,29,137,43,0,0 ; andps 0x2b89(%rip),%xmm11 # 58d0 <_sk_callback_sse41+0x87d> - DB 68,15,86,29,145,43,0,0 ; orps 0x2b91(%rip),%xmm11 # 58e0 <_sk_callback_sse41+0x88d> - DB 68,15,88,37,153,43,0,0 ; addps 0x2b99(%rip),%xmm12 # 58f0 <_sk_callback_sse41+0x89d> - DB 15,40,21,162,43,0,0 ; movaps 0x2ba2(%rip),%xmm2 # 5900 <_sk_callback_sse41+0x8ad> + DB 68,15,89,37,170,43,0,0 ; mulps 0x2baa(%rip),%xmm12 # 5910 <_sk_callback_sse41+0x883> + DB 68,15,84,29,178,43,0,0 ; andps 0x2bb2(%rip),%xmm11 # 5920 <_sk_callback_sse41+0x893> + DB 68,15,86,29,186,43,0,0 ; orps 0x2bba(%rip),%xmm11 # 5930 <_sk_callback_sse41+0x8a3> + DB 68,15,88,37,194,43,0,0 ; addps 0x2bc2(%rip),%xmm12 # 5940 <_sk_callback_sse41+0x8b3> + DB 15,40,21,203,43,0,0 ; movaps 0x2bcb(%rip),%xmm2 # 5950 <_sk_callback_sse41+0x8c3> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,88,29,162,43,0,0 ; addps 0x2ba2(%rip),%xmm11 # 5910 <_sk_callback_sse41+0x8bd> - DB 15,40,21,171,43,0,0 ; movaps 0x2bab(%rip),%xmm2 # 5920 <_sk_callback_sse41+0x8cd> + DB 68,15,88,29,203,43,0,0 ; addps 0x2bcb(%rip),%xmm11 # 5960 <_sk_callback_sse41+0x8d3> + DB 15,40,21,212,43,0,0 ; movaps 0x2bd4(%rip),%xmm2 # 5970 <_sk_callback_sse41+0x8e3> DB 65,15,94,211 ; divps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10 DB 69,15,40,220 ; movaps %xmm12,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,37,152,43,0,0 ; addps 0x2b98(%rip),%xmm12 # 5930 <_sk_callback_sse41+0x8dd> - DB 15,40,21,161,43,0,0 ; movaps 0x2ba1(%rip),%xmm2 # 5940 <_sk_callback_sse41+0x8ed> + DB 68,15,88,37,193,43,0,0 ; addps 0x2bc1(%rip),%xmm12 # 5980 <_sk_callback_sse41+0x8f3> + DB 15,40,21,202,43,0,0 ; movaps 0x2bca(%rip),%xmm2 # 5990 <_sk_callback_sse41+0x903> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,40,21,161,43,0,0 ; movaps 0x2ba1(%rip),%xmm10 # 5950 <_sk_callback_sse41+0x8fd> + DB 68,15,40,21,202,43,0,0 ; movaps 0x2bca(%rip),%xmm10 # 59a0 <_sk_callback_sse41+0x913> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,21,166,43,0,0 ; movaps 0x2ba6(%rip),%xmm2 # 5960 <_sk_callback_sse41+0x90d> + DB 15,40,21,207,43,0,0 ; movaps 0x2bcf(%rip),%xmm2 # 59b0 <_sk_callback_sse41+0x923> DB 65,15,94,210 ; divps %xmm10,%xmm2 DB 65,15,88,212 ; addps %xmm12,%xmm2 - DB 15,89,21,167,43,0,0 ; mulps 0x2ba7(%rip),%xmm2 # 5970 <_sk_callback_sse41+0x91d> + DB 15,89,21,208,43,0,0 ; mulps 0x2bd0(%rip),%xmm2 # 59c0 <_sk_callback_sse41+0x933> DB 102,68,15,91,210 ; cvtps2dq %xmm2,%xmm10 DB 243,15,16,80,20 ; movss 0x14(%rax),%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 @@ -14631,7 +14644,7 @@ _sk_parametric_b_sse41 LABEL PROC DB 102,65,15,56,20,209 ; blendvps %xmm0,%xmm9,%xmm2 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,95,208 ; maxps %xmm0,%xmm2 - DB 15,93,21,146,43,0,0 ; minps 0x2b92(%rip),%xmm2 # 5980 <_sk_callback_sse41+0x92d> + DB 15,93,21,187,43,0,0 ; minps 0x2bbb(%rip),%xmm2 # 59d0 <_sk_callback_sse41+0x943> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14659,31 +14672,31 @@ _sk_parametric_a_sse41 LABEL PROC DB 68,15,88,219 ; addps %xmm3,%xmm11 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 68,15,89,37,51,43,0,0 ; mulps 0x2b33(%rip),%xmm12 # 5990 <_sk_callback_sse41+0x93d> - DB 68,15,84,29,59,43,0,0 ; andps 0x2b3b(%rip),%xmm11 # 59a0 <_sk_callback_sse41+0x94d> - DB 68,15,86,29,67,43,0,0 ; orps 0x2b43(%rip),%xmm11 # 59b0 <_sk_callback_sse41+0x95d> - DB 68,15,88,37,75,43,0,0 ; addps 0x2b4b(%rip),%xmm12 # 59c0 <_sk_callback_sse41+0x96d> - DB 15,40,29,84,43,0,0 ; movaps 0x2b54(%rip),%xmm3 # 59d0 <_sk_callback_sse41+0x97d> + DB 68,15,89,37,92,43,0,0 ; mulps 0x2b5c(%rip),%xmm12 # 59e0 <_sk_callback_sse41+0x953> + DB 68,15,84,29,100,43,0,0 ; andps 0x2b64(%rip),%xmm11 # 59f0 <_sk_callback_sse41+0x963> + DB 68,15,86,29,108,43,0,0 ; orps 0x2b6c(%rip),%xmm11 # 5a00 <_sk_callback_sse41+0x973> + DB 68,15,88,37,116,43,0,0 ; addps 0x2b74(%rip),%xmm12 # 5a10 <_sk_callback_sse41+0x983> + DB 15,40,29,125,43,0,0 ; movaps 0x2b7d(%rip),%xmm3 # 5a20 <_sk_callback_sse41+0x993> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,88,29,84,43,0,0 ; addps 0x2b54(%rip),%xmm11 # 59e0 <_sk_callback_sse41+0x98d> - DB 15,40,29,93,43,0,0 ; movaps 0x2b5d(%rip),%xmm3 # 59f0 <_sk_callback_sse41+0x99d> + DB 68,15,88,29,125,43,0,0 ; addps 0x2b7d(%rip),%xmm11 # 5a30 <_sk_callback_sse41+0x9a3> + DB 15,40,29,134,43,0,0 ; movaps 0x2b86(%rip),%xmm3 # 5a40 <_sk_callback_sse41+0x9b3> DB 65,15,94,219 ; divps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10 DB 69,15,40,220 ; movaps %xmm12,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,37,74,43,0,0 ; addps 0x2b4a(%rip),%xmm12 # 5a00 <_sk_callback_sse41+0x9ad> - DB 15,40,29,83,43,0,0 ; movaps 0x2b53(%rip),%xmm3 # 5a10 <_sk_callback_sse41+0x9bd> + DB 68,15,88,37,115,43,0,0 ; addps 0x2b73(%rip),%xmm12 # 5a50 <_sk_callback_sse41+0x9c3> + DB 15,40,29,124,43,0,0 ; movaps 0x2b7c(%rip),%xmm3 # 5a60 <_sk_callback_sse41+0x9d3> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,40,21,83,43,0,0 ; movaps 0x2b53(%rip),%xmm10 # 5a20 <_sk_callback_sse41+0x9cd> + DB 68,15,40,21,124,43,0,0 ; movaps 0x2b7c(%rip),%xmm10 # 5a70 <_sk_callback_sse41+0x9e3> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,29,88,43,0,0 ; movaps 0x2b58(%rip),%xmm3 # 5a30 <_sk_callback_sse41+0x9dd> + DB 15,40,29,129,43,0,0 ; movaps 0x2b81(%rip),%xmm3 # 5a80 <_sk_callback_sse41+0x9f3> DB 65,15,94,218 ; divps %xmm10,%xmm3 DB 65,15,88,220 ; addps %xmm12,%xmm3 - DB 15,89,29,89,43,0,0 ; mulps 0x2b59(%rip),%xmm3 # 5a40 <_sk_callback_sse41+0x9ed> + DB 15,89,29,130,43,0,0 ; mulps 0x2b82(%rip),%xmm3 # 5a90 <_sk_callback_sse41+0xa03> DB 102,68,15,91,211 ; cvtps2dq %xmm3,%xmm10 DB 243,15,16,88,20 ; movss 0x14(%rax),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 @@ -14691,7 +14704,7 @@ _sk_parametric_a_sse41 LABEL PROC DB 102,65,15,56,20,217 ; blendvps %xmm0,%xmm9,%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,95,216 ; maxps %xmm0,%xmm3 - DB 15,93,29,68,43,0,0 ; minps 0x2b44(%rip),%xmm3 # 5a50 <_sk_callback_sse41+0x9fd> + DB 15,93,29,109,43,0,0 ; minps 0x2b6d(%rip),%xmm3 # 5aa0 <_sk_callback_sse41+0xa13> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14699,29 +14712,29 @@ _sk_parametric_a_sse41 LABEL PROC PUBLIC _sk_lab_to_xyz_sse41 _sk_lab_to_xyz_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,89,5,64,43,0,0 ; mulps 0x2b40(%rip),%xmm8 # 5a60 <_sk_callback_sse41+0xa0d> - DB 68,15,40,13,72,43,0,0 ; movaps 0x2b48(%rip),%xmm9 # 5a70 <_sk_callback_sse41+0xa1d> + DB 68,15,89,5,105,43,0,0 ; mulps 0x2b69(%rip),%xmm8 # 5ab0 <_sk_callback_sse41+0xa23> + DB 68,15,40,13,113,43,0,0 ; movaps 0x2b71(%rip),%xmm9 # 5ac0 <_sk_callback_sse41+0xa33> DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 15,40,5,77,43,0,0 ; movaps 0x2b4d(%rip),%xmm0 # 5a80 <_sk_callback_sse41+0xa2d> + DB 15,40,5,118,43,0,0 ; movaps 0x2b76(%rip),%xmm0 # 5ad0 <_sk_callback_sse41+0xa43> DB 15,88,200 ; addps %xmm0,%xmm1 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 68,15,88,5,75,43,0,0 ; addps 0x2b4b(%rip),%xmm8 # 5a90 <_sk_callback_sse41+0xa3d> - DB 68,15,89,5,83,43,0,0 ; mulps 0x2b53(%rip),%xmm8 # 5aa0 <_sk_callback_sse41+0xa4d> - DB 15,89,13,92,43,0,0 ; mulps 0x2b5c(%rip),%xmm1 # 5ab0 <_sk_callback_sse41+0xa5d> + DB 68,15,88,5,116,43,0,0 ; addps 0x2b74(%rip),%xmm8 # 5ae0 <_sk_callback_sse41+0xa53> + DB 68,15,89,5,124,43,0,0 ; mulps 0x2b7c(%rip),%xmm8 # 5af0 <_sk_callback_sse41+0xa63> + DB 15,89,13,133,43,0,0 ; mulps 0x2b85(%rip),%xmm1 # 5b00 <_sk_callback_sse41+0xa73> DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 15,89,21,97,43,0,0 ; mulps 0x2b61(%rip),%xmm2 # 5ac0 <_sk_callback_sse41+0xa6d> + DB 15,89,21,138,43,0,0 ; mulps 0x2b8a(%rip),%xmm2 # 5b10 <_sk_callback_sse41+0xa83> DB 69,15,40,208 ; movaps %xmm8,%xmm10 DB 68,15,92,210 ; subps %xmm2,%xmm10 DB 68,15,40,217 ; movaps %xmm1,%xmm11 DB 69,15,89,219 ; mulps %xmm11,%xmm11 DB 68,15,89,217 ; mulps %xmm1,%xmm11 - DB 68,15,40,13,85,43,0,0 ; movaps 0x2b55(%rip),%xmm9 # 5ad0 <_sk_callback_sse41+0xa7d> + DB 68,15,40,13,126,43,0,0 ; movaps 0x2b7e(%rip),%xmm9 # 5b20 <_sk_callback_sse41+0xa93> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 15,40,21,85,43,0,0 ; movaps 0x2b55(%rip),%xmm2 # 5ae0 <_sk_callback_sse41+0xa8d> + DB 15,40,21,126,43,0,0 ; movaps 0x2b7e(%rip),%xmm2 # 5b30 <_sk_callback_sse41+0xaa3> DB 15,88,202 ; addps %xmm2,%xmm1 - DB 68,15,40,37,90,43,0,0 ; movaps 0x2b5a(%rip),%xmm12 # 5af0 <_sk_callback_sse41+0xa9d> + DB 68,15,40,37,131,43,0,0 ; movaps 0x2b83(%rip),%xmm12 # 5b40 <_sk_callback_sse41+0xab3> DB 65,15,89,204 ; mulps %xmm12,%xmm1 DB 102,65,15,56,20,203 ; blendvps %xmm0,%xmm11,%xmm1 DB 69,15,40,216 ; movaps %xmm8,%xmm11 @@ -14740,8 +14753,8 @@ _sk_lab_to_xyz_sse41 LABEL PROC DB 65,15,89,212 ; mulps %xmm12,%xmm2 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 102,65,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm2 - DB 15,89,13,19,43,0,0 ; mulps 0x2b13(%rip),%xmm1 # 5b00 <_sk_callback_sse41+0xaad> - DB 15,89,21,28,43,0,0 ; mulps 0x2b1c(%rip),%xmm2 # 5b10 <_sk_callback_sse41+0xabd> + DB 15,89,13,60,43,0,0 ; mulps 0x2b3c(%rip),%xmm1 # 5b50 <_sk_callback_sse41+0xac3> + DB 15,89,21,69,43,0,0 ; mulps 0x2b45(%rip),%xmm2 # 5b60 <_sk_callback_sse41+0xad3> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,40,200 ; movaps %xmm8,%xmm1 @@ -14750,40 +14763,41 @@ _sk_lab_to_xyz_sse41 LABEL PROC PUBLIC _sk_load_a8_sse41 _sk_load_a8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,38 ; jne 302f <_sk_load_a8_sse41+0x30> - DB 102,65,15,56,49,4,57 ; pmovzxbd (%r9,%rdi,1),%xmm0 - DB 102,15,219,5,8,43,0,0 ; pand 0x2b08(%rip),%xmm0 # 5b20 <_sk_callback_sse41+0xacd> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,38 ; jne 3056 <_sk_load_a8_sse41+0x30> + DB 102,65,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm0 + DB 102,15,219,5,49,43,0,0 ; pand 0x2b31(%rip),%xmm0 # 5b70 <_sk_callback_sse41+0xae3> DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,14,43,0,0 ; mulps 0x2b0e(%rip),%xmm3 # 5b30 <_sk_callback_sse41+0xadd> + DB 15,89,29,55,43,0,0 ; mulps 0x2b37(%rip),%xmm3 # 5b80 <_sk_callback_sse41+0xaf3> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,39 ; je 3067 <_sk_load_a8_sse41+0x68> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 305b <_sk_load_a8_sse41+0x5c> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,196 ; jne 3010 <_sk_load_a8_sse41+0x11> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,39 ; je 308e <_sk_load_a8_sse41+0x68> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 3082 <_sk_load_a8_sse41+0x5c> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,196 ; jne 3037 <_sk_load_a8_sse41+0x11> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,15,58,34,192,1 ; pinsrd $0x1,%eax,%xmm0 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,15,58,34,192,0 ; pinsrd $0x0,%eax,%xmm0 - DB 235,156 ; jmp 3010 <_sk_load_a8_sse41+0x11> + DB 235,156 ; jmp 3037 <_sk_load_a8_sse41+0x11> PUBLIC _sk_gather_a8_sse41 _sk_gather_a8_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -14791,94 +14805,96 @@ _sk_gather_a8_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,58,32,4,24,0 ; pinsrb $0x0,(%r8,%r11,1),%xmm0 - DB 102,67,15,58,32,4,16,1 ; pinsrb $0x1,(%r8,%r10,1),%xmm0 - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 102,65,15,58,32,193,2 ; pinsrb $0x2,%r9d,%xmm0 - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,65,15,58,32,4,25,0 ; pinsrb $0x0,(%r9,%rbx,1),%xmm0 + DB 102,67,15,58,32,4,25,1 ; pinsrb $0x1,(%r9,%r11,1),%xmm0 + DB 67,15,182,28,17 ; movzbl (%r9,%r10,1),%ebx + DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0 + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 102,15,58,32,192,3 ; pinsrb $0x3,%eax,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,92,42,0,0 ; mulps 0x2a5c(%rip),%xmm3 # 5b40 <_sk_callback_sse41+0xaed> + DB 15,89,29,133,42,0,0 ; mulps 0x2a85(%rip),%xmm3 # 5b90 <_sk_callback_sse41+0xb03> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_a8_sse41 _sk_store_a8_sse41 LABEL PROC DB 72,131,236,4 ; sub $0x4,%rsp DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 68,15,40,5,76,42,0,0 ; movaps 0x2a4c(%rip),%xmm8 # 5b50 <_sk_callback_sse41+0xafd> + DB 76,139,16 ; mov (%rax),%r10 + DB 68,15,40,5,116,42,0,0 ; movaps 0x2a74(%rip),%xmm8 # 5ba0 <_sk_callback_sse41+0xb13> DB 68,15,89,195 ; mulps %xmm3,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,17 ; jne 312e <_sk_store_a8_sse41+0x3b> + DB 77,133,192 ; test %r8,%r8 + DB 117,17 ; jne 3156 <_sk_store_a8_sse41+0x3b> DB 102,68,15,126,192 ; movd %xmm8,%eax - DB 65,137,4,57 ; mov %eax,(%r9,%rdi,1) + DB 65,137,4,18 ; mov %eax,(%r10,%rdx,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 72,131,196,4 ; add $0x4,%rsp DB 255,224 ; jmpq *%rax DB 102,69,15,56,49,192 ; pmovzxbd %xmm8,%xmm8 - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,30 ; je 315f <_sk_store_a8_sse41+0x6c> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,15 ; je 3156 <_sk_store_a8_sse41+0x63> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,217 ; jne 3126 <_sk_store_a8_sse41+0x33> - DB 102,69,15,58,20,68,57,2,8 ; pextrb $0x8,%xmm8,0x2(%r9,%rdi,1) - DB 102,69,15,58,20,68,57,1,4 ; pextrb $0x4,%xmm8,0x1(%r9,%rdi,1) - DB 102,69,15,58,20,4,57,0 ; pextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,189 ; jmp 3126 <_sk_store_a8_sse41+0x33> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,30 ; je 3187 <_sk_store_a8_sse41+0x6c> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,15 ; je 317e <_sk_store_a8_sse41+0x63> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,217 ; jne 314e <_sk_store_a8_sse41+0x33> + DB 102,69,15,58,20,68,18,2,8 ; pextrb $0x8,%xmm8,0x2(%r10,%rdx,1) + DB 102,69,15,58,20,68,18,1,4 ; pextrb $0x4,%xmm8,0x1(%r10,%rdx,1) + DB 102,69,15,58,20,4,18,0 ; pextrb $0x0,%xmm8,(%r10,%rdx,1) + DB 235,189 ; jmp 314e <_sk_store_a8_sse41+0x33> PUBLIC _sk_load_g8_sse41 _sk_load_g8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,42 ; jne 319d <_sk_load_g8_sse41+0x34> - DB 102,65,15,56,49,4,57 ; pmovzxbd (%r9,%rdi,1),%xmm0 - DB 102,15,219,5,222,41,0,0 ; pand 0x29de(%rip),%xmm0 # 5b60 <_sk_callback_sse41+0xb0d> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,42 ; jne 31c5 <_sk_load_g8_sse41+0x34> + DB 102,65,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm0 + DB 102,15,219,5,6,42,0,0 ; pand 0x2a06(%rip),%xmm0 # 5bb0 <_sk_callback_sse41+0xb23> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,228,41,0,0 ; mulps 0x29e4(%rip),%xmm0 # 5b70 <_sk_callback_sse41+0xb1d> + DB 15,89,5,12,42,0,0 ; mulps 0x2a0c(%rip),%xmm0 # 5bc0 <_sk_callback_sse41+0xb33> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,235,41,0,0 ; movaps 0x29eb(%rip),%xmm3 # 5b80 <_sk_callback_sse41+0xb2d> + DB 15,40,29,19,42,0,0 ; movaps 0x2a13(%rip),%xmm3 # 5bd0 <_sk_callback_sse41+0xb43> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,39 ; je 31d5 <_sk_load_g8_sse41+0x6c> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 31c9 <_sk_load_g8_sse41+0x60> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,192 ; jne 317a <_sk_load_g8_sse41+0x11> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,39 ; je 31fd <_sk_load_g8_sse41+0x6c> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 31f1 <_sk_load_g8_sse41+0x60> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,192 ; jne 31a2 <_sk_load_g8_sse41+0x11> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,15,58,34,192,1 ; pinsrd $0x1,%eax,%xmm0 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,15,58,34,192,0 ; pinsrd $0x0,%eax,%xmm0 - DB 235,152 ; jmp 317a <_sk_load_g8_sse41+0x11> + DB 235,152 ; jmp 31a2 <_sk_load_g8_sse41+0x11> PUBLIC _sk_gather_g8_sse41 _sk_gather_g8_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -14886,37 +14902,39 @@ _sk_gather_g8_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,58,32,4,24,0 ; pinsrb $0x0,(%r8,%r11,1),%xmm0 - DB 102,67,15,58,32,4,16,1 ; pinsrb $0x1,(%r8,%r10,1),%xmm0 - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 102,65,15,58,32,193,2 ; pinsrb $0x2,%r9d,%xmm0 - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,65,15,58,32,4,25,0 ; pinsrb $0x0,(%r9,%rbx,1),%xmm0 + DB 102,67,15,58,32,4,25,1 ; pinsrb $0x1,(%r9,%r11,1),%xmm0 + DB 67,15,182,28,17 ; movzbl (%r9,%r10,1),%ebx + DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0 + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 102,15,58,32,192,3 ; pinsrb $0x3,%eax,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,62,41,0,0 ; mulps 0x293e(%rip),%xmm0 # 5b90 <_sk_callback_sse41+0xb3d> + DB 15,89,5,102,41,0,0 ; mulps 0x2966(%rip),%xmm0 # 5be0 <_sk_callback_sse41+0xb53> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,69,41,0,0 ; movaps 0x2945(%rip),%xmm3 # 5ba0 <_sk_callback_sse41+0xb4d> + DB 15,40,29,109,41,0,0 ; movaps 0x296d(%rip),%xmm3 # 5bf0 <_sk_callback_sse41+0xb63> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_gather_i8_sse41 _sk_gather_i8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,192 ; mov %rax,%r8 - DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 3272 <_sk_gather_i8_sse41+0xf> - DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 3274 <_sk_gather_i8_sse41+0x11> + DB 73,137,193 ; mov %rax,%r9 + DB 77,133,201 ; test %r9,%r9 + DB 116,5 ; je 329b <_sk_gather_i8_sse41+0xf> + DB 76,137,200 ; mov %r9,%rax + DB 235,2 ; jmp 329d <_sk_gather_i8_sse41+0x11> DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,86 ; push %r14 DB 83 ; push %rbx - DB 76,139,8 ; mov (%rax),%r9 + DB 76,139,16 ; mov (%rax),%r10 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -14924,38 +14942,38 @@ _sk_gather_i8_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax - DB 65,137,194 ; mov %eax,%r10d + DB 65,137,195 ; mov %eax,%r11d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,195 ; movq %xmm0,%r11 - DB 68,137,219 ; mov %r11d,%ebx - DB 73,193,235,32 ; shr $0x20,%r11 - DB 102,65,15,58,32,4,25,0 ; pinsrb $0x0,(%r9,%rbx,1),%xmm0 - DB 102,67,15,58,32,4,25,1 ; pinsrb $0x1,(%r9,%r11,1),%xmm0 - DB 102,67,15,58,32,4,17,2 ; pinsrb $0x2,(%r9,%r10,1),%xmm0 - DB 102,65,15,58,32,4,1,3 ; pinsrb $0x3,(%r9,%rax,1),%xmm0 + DB 102,72,15,126,195 ; movq %xmm0,%rbx + DB 65,137,222 ; mov %ebx,%r14d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 102,67,15,58,32,4,50,0 ; pinsrb $0x0,(%r10,%r14,1),%xmm0 + DB 102,65,15,58,32,4,26,1 ; pinsrb $0x1,(%r10,%rbx,1),%xmm0 + DB 102,67,15,58,32,4,26,2 ; pinsrb $0x2,(%r10,%r11,1),%xmm0 + DB 102,65,15,58,32,4,2,3 ; pinsrb $0x3,(%r10,%rax,1),%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 - DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 102,73,15,58,22,194,1 ; pextrq $0x1,%xmm0,%r10 DB 102,72,15,126,195 ; movq %xmm0,%rbx - DB 73,139,64,8 ; mov 0x8(%r8),%rax - DB 65,137,216 ; mov %ebx,%r8d + DB 73,139,65,8 ; mov 0x8(%r9),%rax + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 69,137,202 ; mov %r9d,%r10d - DB 73,193,233,30 ; shr $0x1e,%r9 - DB 102,66,15,110,28,128 ; movd (%rax,%r8,4),%xmm3 + DB 69,137,211 ; mov %r10d,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 102,66,15,110,28,136 ; movd (%rax,%r9,4),%xmm3 DB 102,15,58,34,28,24,1 ; pinsrd $0x1,(%rax,%rbx,1),%xmm3 - DB 102,66,15,58,34,28,144,2 ; pinsrd $0x2,(%rax,%r10,4),%xmm3 - DB 102,66,15,58,34,28,8,3 ; pinsrd $0x3,(%rax,%r9,1),%xmm3 - DB 102,15,111,5,155,40,0,0 ; movdqa 0x289b(%rip),%xmm0 # 5bb0 <_sk_callback_sse41+0xb5d> + DB 102,66,15,58,34,28,152,2 ; pinsrd $0x2,(%rax,%r11,4),%xmm3 + DB 102,66,15,58,34,28,16,3 ; pinsrd $0x3,(%rax,%r10,1),%xmm3 + DB 102,15,111,5,192,40,0,0 ; movdqa 0x28c0(%rip),%xmm0 # 5c00 <_sk_callback_sse41+0xb73> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,156,40,0,0 ; movaps 0x289c(%rip),%xmm8 # 5bc0 <_sk_callback_sse41+0xb6d> + DB 68,15,40,5,193,40,0,0 ; movaps 0x28c1(%rip),%xmm8 # 5c10 <_sk_callback_sse41+0xb83> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,155,40,0,0 ; pshufb 0x289b(%rip),%xmm1 # 5bd0 <_sk_callback_sse41+0xb7d> + DB 102,15,56,0,13,192,40,0,0 ; pshufb 0x28c0(%rip),%xmm1 # 5c20 <_sk_callback_sse41+0xb93> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,151,40,0,0 ; pshufb 0x2897(%rip),%xmm2 # 5be0 <_sk_callback_sse41+0xb8d> + DB 102,15,56,0,21,188,40,0,0 ; pshufb 0x28bc(%rip),%xmm2 # 5c30 <_sk_callback_sse41+0xba3> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -14963,51 +14981,53 @@ _sk_gather_i8_sse41 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx + DB 65,94 ; pop %r14 DB 255,224 ; jmpq *%rax PUBLIC _sk_load_565_sse41 _sk_load_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,80 ; jne 33bb <_sk_load_565_sse41+0x5a> - DB 102,65,15,56,51,20,121 ; pmovzxwd (%r9,%rdi,2),%xmm2 - DB 102,15,111,5,118,40,0,0 ; movdqa 0x2876(%rip),%xmm0 # 5bf0 <_sk_callback_sse41+0xb9d> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,80 ; jne 33e8 <_sk_load_565_sse41+0x5a> + DB 102,65,15,56,51,20,82 ; pmovzxwd (%r10,%rdx,2),%xmm2 + DB 102,15,111,5,153,40,0,0 ; movdqa 0x2899(%rip),%xmm0 # 5c40 <_sk_callback_sse41+0xbb3> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,120,40,0,0 ; mulps 0x2878(%rip),%xmm0 # 5c00 <_sk_callback_sse41+0xbad> - DB 102,15,111,13,128,40,0,0 ; movdqa 0x2880(%rip),%xmm1 # 5c10 <_sk_callback_sse41+0xbbd> + DB 15,89,5,155,40,0,0 ; mulps 0x289b(%rip),%xmm0 # 5c50 <_sk_callback_sse41+0xbc3> + DB 102,15,111,13,163,40,0,0 ; movdqa 0x28a3(%rip),%xmm1 # 5c60 <_sk_callback_sse41+0xbd3> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,130,40,0,0 ; mulps 0x2882(%rip),%xmm1 # 5c20 <_sk_callback_sse41+0xbcd> - DB 102,15,219,21,138,40,0,0 ; pand 0x288a(%rip),%xmm2 # 5c30 <_sk_callback_sse41+0xbdd> + DB 15,89,13,165,40,0,0 ; mulps 0x28a5(%rip),%xmm1 # 5c70 <_sk_callback_sse41+0xbe3> + DB 102,15,219,21,173,40,0,0 ; pand 0x28ad(%rip),%xmm2 # 5c80 <_sk_callback_sse41+0xbf3> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,144,40,0,0 ; mulps 0x2890(%rip),%xmm2 # 5c40 <_sk_callback_sse41+0xbed> + DB 15,89,21,179,40,0,0 ; mulps 0x28b3(%rip),%xmm2 # 5c90 <_sk_callback_sse41+0xc03> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,151,40,0,0 ; movaps 0x2897(%rip),%xmm3 # 5c50 <_sk_callback_sse41+0xbfd> + DB 15,40,29,186,40,0,0 ; movaps 0x28ba(%rip),%xmm3 # 5ca0 <_sk_callback_sse41+0xc13> DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,39 ; je 33f3 <_sk_load_565_sse41+0x92> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 33e7 <_sk_load_565_sse41+0x86> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,154 ; jne 3372 <_sk_load_565_sse41+0x11> - DB 65,15,183,68,121,4 ; movzwl 0x4(%r9,%rdi,2),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,39 ; je 3420 <_sk_load_565_sse41+0x92> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 3414 <_sk_load_565_sse41+0x86> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,154 ; jne 339f <_sk_load_565_sse41+0x11> + DB 65,15,183,68,82,4 ; movzwl 0x4(%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,208,69 ; pshufd $0x45,%xmm0,%xmm2 - DB 65,15,183,68,121,2 ; movzwl 0x2(%r9,%rdi,2),%eax + DB 65,15,183,68,82,2 ; movzwl 0x2(%r10,%rdx,2),%eax DB 102,15,58,34,208,1 ; pinsrd $0x1,%eax,%xmm2 - DB 65,15,183,4,121 ; movzwl (%r9,%rdi,2),%eax + DB 65,15,183,4,82 ; movzwl (%r10,%rdx,2),%eax DB 102,15,58,34,208,0 ; pinsrd $0x0,%eax,%xmm2 - DB 233,111,255,255,255 ; jmpq 3372 <_sk_load_565_sse41+0x11> + DB 233,111,255,255,255 ; jmpq 339f <_sk_load_565_sse41+0x11> PUBLIC _sk_gather_565_sse41 _sk_gather_565_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -15015,43 +15035,44 @@ _sk_gather_565_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,4,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm0 - DB 102,67,15,196,4,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm0 - DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d - DB 102,65,15,196,193,2 ; pinsrw $0x2,%r9d,%xmm0 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,65,15,196,4,89,0 ; pinsrw $0x0,(%r9,%rbx,2),%xmm0 + DB 102,67,15,196,4,89,1 ; pinsrw $0x1,(%r9,%r11,2),%xmm0 + DB 67,15,183,28,81 ; movzwl (%r9,%r10,2),%ebx + DB 102,15,196,195,2 ; pinsrw $0x2,%ebx,%xmm0 + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 102,15,196,192,3 ; pinsrw $0x3,%eax,%xmm0 DB 102,15,56,51,208 ; pmovzxwd %xmm0,%xmm2 - DB 102,15,111,5,243,39,0,0 ; movdqa 0x27f3(%rip),%xmm0 # 5c60 <_sk_callback_sse41+0xc0d> + DB 102,15,111,5,22,40,0,0 ; movdqa 0x2816(%rip),%xmm0 # 5cb0 <_sk_callback_sse41+0xc23> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,245,39,0,0 ; mulps 0x27f5(%rip),%xmm0 # 5c70 <_sk_callback_sse41+0xc1d> - DB 102,15,111,13,253,39,0,0 ; movdqa 0x27fd(%rip),%xmm1 # 5c80 <_sk_callback_sse41+0xc2d> + DB 15,89,5,24,40,0,0 ; mulps 0x2818(%rip),%xmm0 # 5cc0 <_sk_callback_sse41+0xc33> + DB 102,15,111,13,32,40,0,0 ; movdqa 0x2820(%rip),%xmm1 # 5cd0 <_sk_callback_sse41+0xc43> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,255,39,0,0 ; mulps 0x27ff(%rip),%xmm1 # 5c90 <_sk_callback_sse41+0xc3d> - DB 102,15,219,21,7,40,0,0 ; pand 0x2807(%rip),%xmm2 # 5ca0 <_sk_callback_sse41+0xc4d> + DB 15,89,13,34,40,0,0 ; mulps 0x2822(%rip),%xmm1 # 5ce0 <_sk_callback_sse41+0xc53> + DB 102,15,219,21,42,40,0,0 ; pand 0x282a(%rip),%xmm2 # 5cf0 <_sk_callback_sse41+0xc63> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,13,40,0,0 ; mulps 0x280d(%rip),%xmm2 # 5cb0 <_sk_callback_sse41+0xc5d> + DB 15,89,21,48,40,0,0 ; mulps 0x2830(%rip),%xmm2 # 5d00 <_sk_callback_sse41+0xc73> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,20,40,0,0 ; movaps 0x2814(%rip),%xmm3 # 5cc0 <_sk_callback_sse41+0xc6d> + DB 15,40,29,55,40,0,0 ; movaps 0x2837(%rip),%xmm3 # 5d10 <_sk_callback_sse41+0xc83> + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_sse41 _sk_store_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,21,40,0,0 ; movaps 0x2815(%rip),%xmm8 # 5cd0 <_sk_callback_sse41+0xc7d> + DB 68,15,40,5,55,40,0,0 ; movaps 0x2837(%rip),%xmm8 # 5d20 <_sk_callback_sse41+0xc93> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 DB 102,65,15,114,241,11 ; pslld $0xb,%xmm9 - DB 68,15,40,21,10,40,0,0 ; movaps 0x280a(%rip),%xmm10 # 5ce0 <_sk_callback_sse41+0xc8d> + DB 68,15,40,21,44,40,0,0 ; movaps 0x282c(%rip),%xmm10 # 5d30 <_sk_callback_sse41+0xca3> DB 68,15,89,209 ; mulps %xmm1,%xmm10 DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10 DB 102,65,15,114,242,5 ; pslld $0x5,%xmm10 @@ -15060,71 +15081,72 @@ _sk_store_565_sse41 LABEL PROC DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,69,15,86,194 ; orpd %xmm10,%xmm8 DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 350d <_sk_store_565_sse41+0x5f> - DB 242,68,15,17,4,120 ; movsd %xmm8,(%rax,%rdi,2) + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 353b <_sk_store_565_sse41+0x5f> + DB 242,68,15,17,4,80 ; movsd %xmm8,(%rax,%rdx,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,69,15,56,51,192 ; pmovzxwd %xmm8,%xmm8 - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,30 ; je 353e <_sk_store_565_sse41+0x90> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,15 ; je 3535 <_sk_store_565_sse41+0x87> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,221 ; jne 3509 <_sk_store_565_sse41+0x5b> - DB 102,68,15,58,21,68,120,4,4 ; pextrw $0x4,%xmm8,0x4(%rax,%rdi,2) - DB 102,68,15,58,21,68,120,2,2 ; pextrw $0x2,%xmm8,0x2(%rax,%rdi,2) - DB 102,68,15,58,21,4,120,0 ; pextrw $0x0,%xmm8,(%rax,%rdi,2) - DB 235,193 ; jmp 3509 <_sk_store_565_sse41+0x5b> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,30 ; je 356c <_sk_store_565_sse41+0x90> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,15 ; je 3563 <_sk_store_565_sse41+0x87> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,221 ; jne 3537 <_sk_store_565_sse41+0x5b> + DB 102,68,15,58,21,68,80,4,4 ; pextrw $0x4,%xmm8,0x4(%rax,%rdx,2) + DB 102,68,15,58,21,68,80,2,2 ; pextrw $0x2,%xmm8,0x2(%rax,%rdx,2) + DB 102,68,15,58,21,4,80,0 ; pextrw $0x0,%xmm8,(%rax,%rdx,2) + DB 235,193 ; jmp 3537 <_sk_store_565_sse41+0x5b> PUBLIC _sk_load_4444_sse41 _sk_load_4444_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,95 ; jne 35b1 <_sk_load_4444_sse41+0x69> - DB 102,65,15,56,51,28,121 ; pmovzxwd (%r9,%rdi,2),%xmm3 - DB 102,15,111,5,143,39,0,0 ; movdqa 0x278f(%rip),%xmm0 # 5cf0 <_sk_callback_sse41+0xc9d> + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,95 ; jne 35df <_sk_load_4444_sse41+0x69> + DB 102,65,15,56,51,28,82 ; pmovzxwd (%r10,%rdx,2),%xmm3 + DB 102,15,111,5,177,39,0,0 ; movdqa 0x27b1(%rip),%xmm0 # 5d40 <_sk_callback_sse41+0xcb3> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,145,39,0,0 ; mulps 0x2791(%rip),%xmm0 # 5d00 <_sk_callback_sse41+0xcad> - DB 102,15,111,13,153,39,0,0 ; movdqa 0x2799(%rip),%xmm1 # 5d10 <_sk_callback_sse41+0xcbd> + DB 15,89,5,179,39,0,0 ; mulps 0x27b3(%rip),%xmm0 # 5d50 <_sk_callback_sse41+0xcc3> + DB 102,15,111,13,187,39,0,0 ; movdqa 0x27bb(%rip),%xmm1 # 5d60 <_sk_callback_sse41+0xcd3> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,155,39,0,0 ; mulps 0x279b(%rip),%xmm1 # 5d20 <_sk_callback_sse41+0xccd> - DB 102,15,111,21,163,39,0,0 ; movdqa 0x27a3(%rip),%xmm2 # 5d30 <_sk_callback_sse41+0xcdd> + DB 15,89,13,189,39,0,0 ; mulps 0x27bd(%rip),%xmm1 # 5d70 <_sk_callback_sse41+0xce3> + DB 102,15,111,21,197,39,0,0 ; movdqa 0x27c5(%rip),%xmm2 # 5d80 <_sk_callback_sse41+0xcf3> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,165,39,0,0 ; mulps 0x27a5(%rip),%xmm2 # 5d40 <_sk_callback_sse41+0xced> - DB 102,15,219,29,173,39,0,0 ; pand 0x27ad(%rip),%xmm3 # 5d50 <_sk_callback_sse41+0xcfd> + DB 15,89,21,199,39,0,0 ; mulps 0x27c7(%rip),%xmm2 # 5d90 <_sk_callback_sse41+0xd03> + DB 102,15,219,29,207,39,0,0 ; pand 0x27cf(%rip),%xmm3 # 5da0 <_sk_callback_sse41+0xd13> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,179,39,0,0 ; mulps 0x27b3(%rip),%xmm3 # 5d60 <_sk_callback_sse41+0xd0d> + DB 15,89,29,213,39,0,0 ; mulps 0x27d5(%rip),%xmm3 # 5db0 <_sk_callback_sse41+0xd23> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,39 ; je 35e9 <_sk_load_4444_sse41+0xa1> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 35dd <_sk_load_4444_sse41+0x95> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,139 ; jne 3559 <_sk_load_4444_sse41+0x11> - DB 65,15,183,68,121,4 ; movzwl 0x4(%r9,%rdi,2),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,39 ; je 3617 <_sk_load_4444_sse41+0xa1> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 360b <_sk_load_4444_sse41+0x95> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,139 ; jne 3587 <_sk_load_4444_sse41+0x11> + DB 65,15,183,68,82,4 ; movzwl 0x4(%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 - DB 65,15,183,68,121,2 ; movzwl 0x2(%r9,%rdi,2),%eax + DB 65,15,183,68,82,2 ; movzwl 0x2(%r10,%rdx,2),%eax DB 102,15,58,34,216,1 ; pinsrd $0x1,%eax,%xmm3 - DB 65,15,183,4,121 ; movzwl (%r9,%rdi,2),%eax + DB 65,15,183,4,82 ; movzwl (%r10,%rdx,2),%eax DB 102,15,58,34,216,0 ; pinsrd $0x0,%eax,%xmm3 - DB 233,96,255,255,255 ; jmpq 3559 <_sk_load_4444_sse41+0x11> + DB 233,96,255,255,255 ; jmpq 3587 <_sk_load_4444_sse41+0x11> PUBLIC _sk_gather_4444_sse41 _sk_gather_4444_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -15132,41 +15154,42 @@ _sk_gather_4444_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,4,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm0 - DB 102,67,15,196,4,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm0 - DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d - DB 102,65,15,196,193,2 ; pinsrw $0x2,%r9d,%xmm0 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,65,15,196,4,89,0 ; pinsrw $0x0,(%r9,%rbx,2),%xmm0 + DB 102,67,15,196,4,89,1 ; pinsrw $0x1,(%r9,%r11,2),%xmm0 + DB 67,15,183,28,81 ; movzwl (%r9,%r10,2),%ebx + DB 102,15,196,195,2 ; pinsrw $0x2,%ebx,%xmm0 + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 102,15,196,192,3 ; pinsrw $0x3,%eax,%xmm0 DB 102,15,56,51,216 ; pmovzxwd %xmm0,%xmm3 - DB 102,15,111,5,13,39,0,0 ; movdqa 0x270d(%rip),%xmm0 # 5d70 <_sk_callback_sse41+0xd1d> + DB 102,15,111,5,47,39,0,0 ; movdqa 0x272f(%rip),%xmm0 # 5dc0 <_sk_callback_sse41+0xd33> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,15,39,0,0 ; mulps 0x270f(%rip),%xmm0 # 5d80 <_sk_callback_sse41+0xd2d> - DB 102,15,111,13,23,39,0,0 ; movdqa 0x2717(%rip),%xmm1 # 5d90 <_sk_callback_sse41+0xd3d> + DB 15,89,5,49,39,0,0 ; mulps 0x2731(%rip),%xmm0 # 5dd0 <_sk_callback_sse41+0xd43> + DB 102,15,111,13,57,39,0,0 ; movdqa 0x2739(%rip),%xmm1 # 5de0 <_sk_callback_sse41+0xd53> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,25,39,0,0 ; mulps 0x2719(%rip),%xmm1 # 5da0 <_sk_callback_sse41+0xd4d> - DB 102,15,111,21,33,39,0,0 ; movdqa 0x2721(%rip),%xmm2 # 5db0 <_sk_callback_sse41+0xd5d> + DB 15,89,13,59,39,0,0 ; mulps 0x273b(%rip),%xmm1 # 5df0 <_sk_callback_sse41+0xd63> + DB 102,15,111,21,67,39,0,0 ; movdqa 0x2743(%rip),%xmm2 # 5e00 <_sk_callback_sse41+0xd73> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,35,39,0,0 ; mulps 0x2723(%rip),%xmm2 # 5dc0 <_sk_callback_sse41+0xd6d> - DB 102,15,219,29,43,39,0,0 ; pand 0x272b(%rip),%xmm3 # 5dd0 <_sk_callback_sse41+0xd7d> + DB 15,89,21,69,39,0,0 ; mulps 0x2745(%rip),%xmm2 # 5e10 <_sk_callback_sse41+0xd83> + DB 102,15,219,29,77,39,0,0 ; pand 0x274d(%rip),%xmm3 # 5e20 <_sk_callback_sse41+0xd93> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,49,39,0,0 ; mulps 0x2731(%rip),%xmm3 # 5de0 <_sk_callback_sse41+0xd8d> + DB 15,89,29,83,39,0,0 ; mulps 0x2753(%rip),%xmm3 # 5e30 <_sk_callback_sse41+0xda3> DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_4444_sse41 _sk_store_4444_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,48,39,0,0 ; movaps 0x2730(%rip),%xmm8 # 5df0 <_sk_callback_sse41+0xd9d> + DB 68,15,40,5,81,39,0,0 ; movaps 0x2751(%rip),%xmm8 # 5e40 <_sk_callback_sse41+0xdb3> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -15185,43 +15208,43 @@ _sk_store_4444_sse41 LABEL PROC DB 102,69,15,86,193 ; orpd %xmm9,%xmm8 DB 102,69,15,86,194 ; orpd %xmm10,%xmm8 DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3726 <_sk_store_4444_sse41+0x73> - DB 242,68,15,17,4,120 ; movsd %xmm8,(%rax,%rdi,2) + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3755 <_sk_store_4444_sse41+0x73> + DB 242,68,15,17,4,80 ; movsd %xmm8,(%rax,%rdx,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,69,15,56,51,192 ; pmovzxwd %xmm8,%xmm8 - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,30 ; je 3757 <_sk_store_4444_sse41+0xa4> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,15 ; je 374e <_sk_store_4444_sse41+0x9b> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,221 ; jne 3722 <_sk_store_4444_sse41+0x6f> - DB 102,68,15,58,21,68,120,4,4 ; pextrw $0x4,%xmm8,0x4(%rax,%rdi,2) - DB 102,68,15,58,21,68,120,2,2 ; pextrw $0x2,%xmm8,0x2(%rax,%rdi,2) - DB 102,68,15,58,21,4,120,0 ; pextrw $0x0,%xmm8,(%rax,%rdi,2) - DB 235,193 ; jmp 3722 <_sk_store_4444_sse41+0x6f> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,30 ; je 3786 <_sk_store_4444_sse41+0xa4> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,15 ; je 377d <_sk_store_4444_sse41+0x9b> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,221 ; jne 3751 <_sk_store_4444_sse41+0x6f> + DB 102,68,15,58,21,68,80,4,4 ; pextrw $0x4,%xmm8,0x4(%rax,%rdx,2) + DB 102,68,15,58,21,68,80,2,2 ; pextrw $0x2,%xmm8,0x2(%rax,%rdx,2) + DB 102,68,15,58,21,4,80,0 ; pextrw $0x0,%xmm8,(%rax,%rdx,2) + DB 235,193 ; jmp 3751 <_sk_store_4444_sse41+0x6f> PUBLIC _sk_load_8888_sse41 _sk_load_8888_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 117,88 ; jne 37c3 <_sk_load_8888_sse41+0x62> - DB 243,15,111,28,184 ; movdqu (%rax,%rdi,4),%xmm3 - DB 102,15,111,5,136,38,0,0 ; movdqa 0x2688(%rip),%xmm0 # 5e00 <_sk_callback_sse41+0xdad> + DB 77,133,192 ; test %r8,%r8 + DB 117,88 ; jne 37f2 <_sk_load_8888_sse41+0x62> + DB 243,15,111,28,144 ; movdqu (%rax,%rdx,4),%xmm3 + DB 102,15,111,5,169,38,0,0 ; movdqa 0x26a9(%rip),%xmm0 # 5e50 <_sk_callback_sse41+0xdc3> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,137,38,0,0 ; movaps 0x2689(%rip),%xmm8 # 5e10 <_sk_callback_sse41+0xdbd> + DB 68,15,40,5,170,38,0,0 ; movaps 0x26aa(%rip),%xmm8 # 5e60 <_sk_callback_sse41+0xdd3> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,136,38,0,0 ; pshufb 0x2688(%rip),%xmm1 # 5e20 <_sk_callback_sse41+0xdcd> + DB 102,15,56,0,13,169,38,0,0 ; pshufb 0x26a9(%rip),%xmm1 # 5e70 <_sk_callback_sse41+0xde3> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,132,38,0,0 ; pshufb 0x2684(%rip),%xmm2 # 5e30 <_sk_callback_sse41+0xddd> + DB 102,15,56,0,21,165,38,0,0 ; pshufb 0x26a5(%rip),%xmm2 # 5e80 <_sk_callback_sse41+0xdf3> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -15229,25 +15252,26 @@ _sk_load_8888_sse41 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,31 ; je 37f3 <_sk_load_8888_sse41+0x92> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,17 ; je 37eb <_sk_load_8888_sse41+0x8a> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,144 ; jne 3770 <_sk_load_8888_sse41+0xf> - DB 102,15,110,68,184,8 ; movd 0x8(%rax,%rdi,4),%xmm0 + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,31 ; je 3822 <_sk_load_8888_sse41+0x92> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,17 ; je 381a <_sk_load_8888_sse41+0x8a> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,144 ; jne 379f <_sk_load_8888_sse41+0xf> + DB 102,15,110,68,144,8 ; movd 0x8(%rax,%rdx,4),%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 - DB 102,15,58,34,92,184,4,1 ; pinsrd $0x1,0x4(%rax,%rdi,4),%xmm3 - DB 102,15,58,34,28,184,0 ; pinsrd $0x0,(%rax,%rdi,4),%xmm3 - DB 233,113,255,255,255 ; jmpq 3770 <_sk_load_8888_sse41+0xf> + DB 102,15,58,34,92,144,4,1 ; pinsrd $0x1,0x4(%rax,%rdx,4),%xmm3 + DB 102,15,58,34,28,144,0 ; pinsrd $0x0,(%rax,%rdx,4),%xmm3 + DB 233,113,255,255,255 ; jmpq 379f <_sk_load_8888_sse41+0xf> PUBLIC _sk_gather_8888_sse41 _sk_gather_8888_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -15255,39 +15279,40 @@ _sk_gather_8888_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,126,192 ; movq %xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,58,22,194,1 ; pextrq $0x1,%xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,110,28,136 ; movd (%r8,%r9,4),%xmm3 - DB 102,65,15,58,34,28,128,1 ; pinsrd $0x1,(%r8,%rax,4),%xmm3 - DB 102,67,15,58,34,28,152,2 ; pinsrd $0x2,(%r8,%r11,4),%xmm3 - DB 102,67,15,58,34,28,144,3 ; pinsrd $0x3,(%r8,%r10,4),%xmm3 - DB 102,15,111,5,225,37,0,0 ; movdqa 0x25e1(%rip),%xmm0 # 5e40 <_sk_callback_sse41+0xded> + DB 102,73,15,58,22,195,1 ; pextrq $0x1,%xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,67,15,110,28,145 ; movd (%r9,%r10,4),%xmm3 + DB 102,65,15,58,34,28,129,1 ; pinsrd $0x1,(%r9,%rax,4),%xmm3 + DB 102,65,15,58,34,28,153,2 ; pinsrd $0x2,(%r9,%rbx,4),%xmm3 + DB 102,67,15,58,34,28,153,3 ; pinsrd $0x3,(%r9,%r11,4),%xmm3 + DB 102,15,111,5,1,38,0,0 ; movdqa 0x2601(%rip),%xmm0 # 5e90 <_sk_callback_sse41+0xe03> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,226,37,0,0 ; movaps 0x25e2(%rip),%xmm8 # 5e50 <_sk_callback_sse41+0xdfd> + DB 68,15,40,5,2,38,0,0 ; movaps 0x2602(%rip),%xmm8 # 5ea0 <_sk_callback_sse41+0xe13> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,225,37,0,0 ; pshufb 0x25e1(%rip),%xmm1 # 5e60 <_sk_callback_sse41+0xe0d> + DB 102,15,56,0,13,1,38,0,0 ; pshufb 0x2601(%rip),%xmm1 # 5eb0 <_sk_callback_sse41+0xe23> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,221,37,0,0 ; pshufb 0x25dd(%rip),%xmm2 # 5e70 <_sk_callback_sse41+0xe1d> + DB 102,15,56,0,21,253,37,0,0 ; pshufb 0x25fd(%rip),%xmm2 # 5ec0 <_sk_callback_sse41+0xe33> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_8888_sse41 _sk_store_8888_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,201,37,0,0 ; movaps 0x25c9(%rip),%xmm8 # 5e80 <_sk_callback_sse41+0xe2d> + DB 68,15,40,5,232,37,0,0 ; movaps 0x25e8(%rip),%xmm8 # 5ed0 <_sk_callback_sse41+0xe43> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -15305,32 +15330,32 @@ _sk_store_8888_sse41 LABEL PROC DB 102,65,15,114,240,24 ; pslld $0x18,%xmm8 DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 102,69,15,235,194 ; por %xmm10,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3917 <_sk_store_8888_sse41+0x6d> - DB 243,68,15,127,4,184 ; movdqu %xmm8,(%rax,%rdi,4) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,30 ; je 3942 <_sk_store_8888_sse41+0x98> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,15 ; je 3939 <_sk_store_8888_sse41+0x8f> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,227 ; jne 3913 <_sk_store_8888_sse41+0x69> - DB 102,68,15,58,22,68,184,8,2 ; pextrd $0x2,%xmm8,0x8(%rax,%rdi,4) - DB 102,68,15,58,22,68,184,4,1 ; pextrd $0x1,%xmm8,0x4(%rax,%rdi,4) - DB 102,68,15,126,4,184 ; movd %xmm8,(%rax,%rdi,4) - DB 235,201 ; jmp 3913 <_sk_store_8888_sse41+0x69> + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3948 <_sk_store_8888_sse41+0x6d> + DB 243,68,15,127,4,144 ; movdqu %xmm8,(%rax,%rdx,4) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,30 ; je 3973 <_sk_store_8888_sse41+0x98> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,15 ; je 396a <_sk_store_8888_sse41+0x8f> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,227 ; jne 3944 <_sk_store_8888_sse41+0x69> + DB 102,68,15,58,22,68,144,8,2 ; pextrd $0x2,%xmm8,0x8(%rax,%rdx,4) + DB 102,68,15,58,22,68,144,4,1 ; pextrd $0x1,%xmm8,0x4(%rax,%rdx,4) + DB 102,68,15,126,4,144 ; movd %xmm8,(%rax,%rdx,4) + DB 235,201 ; jmp 3944 <_sk_store_8888_sse41+0x69> PUBLIC _sk_load_f16_sse41 _sk_load_f16_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,60,1,0,0 ; jne 3a94 <_sk_load_f16_sse41+0x14a> - DB 102,15,16,4,248 ; movupd (%rax,%rdi,8),%xmm0 - DB 243,15,111,76,248,16 ; movdqu 0x10(%rax,%rdi,8),%xmm1 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,60,1,0,0 ; jne 3ac5 <_sk_load_f16_sse41+0x14a> + DB 102,15,16,4,208 ; movupd (%rax,%rdx,8),%xmm0 + DB 243,15,111,76,208,16 ; movdqu 0x10(%rax,%rdx,8),%xmm1 DB 102,68,15,40,200 ; movapd %xmm0,%xmm9 DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 @@ -15338,18 +15363,18 @@ _sk_load_f16_sse41 LABEL PROC DB 102,68,15,97,216 ; punpcklwd %xmm0,%xmm11 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1 - DB 102,68,15,111,5,1,37,0,0 ; movdqa 0x2501(%rip),%xmm8 # 5e90 <_sk_callback_sse41+0xe3d> + DB 102,68,15,111,5,32,37,0,0 ; movdqa 0x2520(%rip),%xmm8 # 5ee0 <_sk_callback_sse41+0xe53> DB 102,15,111,209 ; movdqa %xmm1,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 DB 102,15,239,202 ; pxor %xmm2,%xmm1 - DB 102,15,111,29,252,36,0,0 ; movdqa 0x24fc(%rip),%xmm3 # 5ea0 <_sk_callback_sse41+0xe4d> + DB 102,15,111,29,27,37,0,0 ; movdqa 0x251b(%rip),%xmm3 # 5ef0 <_sk_callback_sse41+0xe63> DB 102,15,114,242,16 ; pslld $0x10,%xmm2 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,56,63,195 ; pmaxud %xmm3,%xmm0 DB 102,15,118,193 ; pcmpeqd %xmm1,%xmm0 DB 102,15,114,241,13 ; pslld $0xd,%xmm1 DB 102,15,235,202 ; por %xmm2,%xmm1 - DB 102,68,15,111,21,232,36,0,0 ; movdqa 0x24e8(%rip),%xmm10 # 5eb0 <_sk_callback_sse41+0xe5d> + DB 102,68,15,111,21,7,37,0,0 ; movdqa 0x2507(%rip),%xmm10 # 5f00 <_sk_callback_sse41+0xe73> DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 DB 102,15,219,193 ; pand %xmm1,%xmm0 DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11 @@ -15390,23 +15415,24 @@ _sk_load_f16_sse41 LABEL PROC DB 102,65,15,219,217 ; pand %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 242,15,16,4,248 ; movsd (%rax,%rdi,8),%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,13 ; jne 3aac <_sk_load_f16_sse41+0x162> + DB 242,15,16,4,208 ; movsd (%rax,%rdx,8),%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,13 ; jne 3add <_sk_load_f16_sse41+0x162> DB 243,15,126,192 ; movq %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,183,254,255,255 ; jmpq 3963 <_sk_load_f16_sse41+0x19> - DB 102,15,22,68,248,8 ; movhpd 0x8(%rax,%rdi,8),%xmm0 + DB 233,183,254,255,255 ; jmpq 3994 <_sk_load_f16_sse41+0x19> + DB 102,15,22,68,208,8 ; movhpd 0x8(%rax,%rdx,8),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 15,130,163,254,255,255 ; jb 3963 <_sk_load_f16_sse41+0x19> - DB 243,15,126,76,248,16 ; movq 0x10(%rax,%rdi,8),%xmm1 - DB 233,152,254,255,255 ; jmpq 3963 <_sk_load_f16_sse41+0x19> + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 15,130,163,254,255,255 ; jb 3994 <_sk_load_f16_sse41+0x19> + DB 243,15,126,76,208,16 ; movq 0x10(%rax,%rdx,8),%xmm1 + DB 233,152,254,255,255 ; jmpq 3994 <_sk_load_f16_sse41+0x19> PUBLIC _sk_gather_f16_sse41 _sk_gather_f16_sse41 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -15414,16 +15440,16 @@ _sk_gather_f16_sse41 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,194 ; paddd %xmm2,%xmm0 DB 102,72,15,126,192 ; movq %xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,58,22,194,1 ; pextrq $0x1,%xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,67,15,126,4,208 ; movq (%r8,%r10,8),%xmm0 - DB 243,67,15,126,12,216 ; movq (%r8,%r11,8),%xmm1 + DB 102,73,15,58,22,195,1 ; pextrq $0x1,%xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,67,15,126,4,217 ; movq (%r9,%r11,8),%xmm0 + DB 243,65,15,126,12,217 ; movq (%r9,%rbx,8),%xmm1 DB 102,15,108,200 ; punpcklqdq %xmm0,%xmm1 - DB 243,65,15,126,4,192 ; movq (%r8,%rax,8),%xmm0 - DB 243,67,15,126,20,200 ; movq (%r8,%r9,8),%xmm2 + DB 243,65,15,126,4,193 ; movq (%r9,%rax,8),%xmm0 + DB 243,67,15,126,20,209 ; movq (%r9,%r10,8),%xmm2 DB 102,15,108,208 ; punpcklqdq %xmm0,%xmm2 DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9 @@ -15432,18 +15458,18 @@ _sk_gather_f16_sse41 LABEL PROC DB 102,68,15,97,218 ; punpcklwd %xmm2,%xmm11 DB 102,68,15,105,202 ; punpckhwd %xmm2,%xmm9 DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1 - DB 102,68,15,111,5,111,35,0,0 ; movdqa 0x236f(%rip),%xmm8 # 5ec0 <_sk_callback_sse41+0xe6d> + DB 102,68,15,111,5,141,35,0,0 ; movdqa 0x238d(%rip),%xmm8 # 5f10 <_sk_callback_sse41+0xe83> DB 102,15,111,209 ; movdqa %xmm1,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 DB 102,15,239,202 ; pxor %xmm2,%xmm1 - DB 102,15,111,29,106,35,0,0 ; movdqa 0x236a(%rip),%xmm3 # 5ed0 <_sk_callback_sse41+0xe7d> + DB 102,15,111,29,136,35,0,0 ; movdqa 0x2388(%rip),%xmm3 # 5f20 <_sk_callback_sse41+0xe93> DB 102,15,114,242,16 ; pslld $0x10,%xmm2 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,56,63,195 ; pmaxud %xmm3,%xmm0 DB 102,15,118,193 ; pcmpeqd %xmm1,%xmm0 DB 102,15,114,241,13 ; pslld $0xd,%xmm1 DB 102,15,235,202 ; por %xmm2,%xmm1 - DB 102,68,15,111,21,86,35,0,0 ; movdqa 0x2356(%rip),%xmm10 # 5ee0 <_sk_callback_sse41+0xe8d> + DB 102,68,15,111,21,116,35,0,0 ; movdqa 0x2374(%rip),%xmm10 # 5f30 <_sk_callback_sse41+0xea3> DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 DB 102,15,219,193 ; pand %xmm1,%xmm0 DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11 @@ -15483,23 +15509,24 @@ _sk_gather_f16_sse41 LABEL PROC DB 102,69,15,254,202 ; paddd %xmm10,%xmm9 DB 102,65,15,219,217 ; pand %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_f16_sse41 _sk_store_f16_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 102,68,15,111,21,140,34,0,0 ; movdqa 0x228c(%rip),%xmm10 # 5ef0 <_sk_callback_sse41+0xe9d> + DB 102,68,15,111,21,169,34,0,0 ; movdqa 0x22a9(%rip),%xmm10 # 5f40 <_sk_callback_sse41+0xeb3> DB 102,68,15,111,216 ; movdqa %xmm0,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 DB 102,69,15,239,235 ; pxor %xmm11,%xmm13 - DB 102,68,15,111,13,127,34,0,0 ; movdqa 0x227f(%rip),%xmm9 # 5f00 <_sk_callback_sse41+0xead> + DB 102,68,15,111,13,156,34,0,0 ; movdqa 0x229c(%rip),%xmm9 # 5f50 <_sk_callback_sse41+0xec3> DB 102,65,15,114,211,16 ; psrld $0x10,%xmm11 DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 102,68,15,111,37,112,34,0,0 ; movdqa 0x2270(%rip),%xmm12 # 5f10 <_sk_callback_sse41+0xebd> + DB 102,68,15,111,37,141,34,0,0 ; movdqa 0x228d(%rip),%xmm12 # 5f60 <_sk_callback_sse41+0xed3> DB 102,69,15,235,220 ; por %xmm12,%xmm11 DB 102,69,15,254,221 ; paddd %xmm13,%xmm11 DB 102,69,15,223,195 ; pandn %xmm11,%xmm8 @@ -15542,32 +15569,32 @@ _sk_store_f16_sse41 LABEL PROC DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,69,15,98,203 ; punpckldq %xmm11,%xmm9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,21 ; jne 3d96 <_sk_store_f16_sse41+0x140> - DB 68,15,17,12,248 ; movups %xmm9,(%rax,%rdi,8) + DB 77,133,192 ; test %r8,%r8 + DB 117,21 ; jne 3dc9 <_sk_store_f16_sse41+0x140> + DB 68,15,17,12,208 ; movups %xmm9,(%rax,%rdx,8) DB 102,69,15,106,195 ; punpckhdq %xmm11,%xmm8 - DB 243,68,15,127,68,248,16 ; movdqu %xmm8,0x10(%rax,%rdi,8) + DB 243,68,15,127,68,208,16 ; movdqu %xmm8,0x10(%rax,%rdx,8) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 102,68,15,214,12,248 ; movq %xmm9,(%rax,%rdi,8) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 3d92 <_sk_store_f16_sse41+0x13c> - DB 102,68,15,23,76,248,8 ; movhpd %xmm9,0x8(%rax,%rdi,8) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 3d92 <_sk_store_f16_sse41+0x13c> + DB 102,68,15,214,12,208 ; movq %xmm9,(%rax,%rdx,8) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 3dc5 <_sk_store_f16_sse41+0x13c> + DB 102,68,15,23,76,208,8 ; movhpd %xmm9,0x8(%rax,%rdx,8) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 3dc5 <_sk_store_f16_sse41+0x13c> DB 102,69,15,106,195 ; punpckhdq %xmm11,%xmm8 - DB 102,68,15,214,68,248,16 ; movq %xmm8,0x10(%rax,%rdi,8) - DB 235,213 ; jmp 3d92 <_sk_store_f16_sse41+0x13c> + DB 102,68,15,214,68,208,16 ; movq %xmm8,0x10(%rax,%rdx,8) + DB 235,213 ; jmp 3dc5 <_sk_store_f16_sse41+0x13c> PUBLIC _sk_load_u16_be_sse41 _sk_load_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,185,0,0,0 ; jne 3e8c <_sk_load_u16_be_sse41+0xcf> - DB 102,65,15,16,4,64 ; movupd (%r8,%rax,2),%xmm0 - DB 243,65,15,111,76,64,16 ; movdqu 0x10(%r8,%rax,2),%xmm1 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,185,0,0,0 ; jne 3ebf <_sk_load_u16_be_sse41+0xcf> + DB 102,65,15,16,4,65 ; movupd (%r9,%rax,2),%xmm0 + DB 243,65,15,111,76,65,16 ; movdqu 0x10(%r9,%rax,2),%xmm1 DB 102,15,40,208 ; movapd %xmm0,%xmm2 DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 @@ -15581,7 +15608,7 @@ _sk_load_u16_be_sse41 LABEL PROC DB 102,15,235,200 ; por %xmm0,%xmm1 DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,1,33,0,0 ; movaps 0x2101(%rip),%xmm8 # 5f20 <_sk_callback_sse41+0xecd> + DB 68,15,40,5,30,33,0,0 ; movaps 0x211e(%rip),%xmm8 # 5f70 <_sk_callback_sse41+0xee3> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -15607,28 +15634,28 @@ _sk_load_u16_be_sse41 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 242,65,15,16,4,64 ; movsd (%r8,%rax,2),%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,13 ; jne 3ea5 <_sk_load_u16_be_sse41+0xe8> + DB 242,65,15,16,4,65 ; movsd (%r9,%rax,2),%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,13 ; jne 3ed8 <_sk_load_u16_be_sse41+0xe8> DB 243,15,126,192 ; movq %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,59,255,255,255 ; jmpq 3de0 <_sk_load_u16_be_sse41+0x23> - DB 102,65,15,22,68,64,8 ; movhpd 0x8(%r8,%rax,2),%xmm0 + DB 233,59,255,255,255 ; jmpq 3e13 <_sk_load_u16_be_sse41+0x23> + DB 102,65,15,22,68,65,8 ; movhpd 0x8(%r9,%rax,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 15,130,38,255,255,255 ; jb 3de0 <_sk_load_u16_be_sse41+0x23> - DB 243,65,15,126,76,64,16 ; movq 0x10(%r8,%rax,2),%xmm1 - DB 233,26,255,255,255 ; jmpq 3de0 <_sk_load_u16_be_sse41+0x23> + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 15,130,38,255,255,255 ; jb 3e13 <_sk_load_u16_be_sse41+0x23> + DB 243,65,15,126,76,65,16 ; movq 0x10(%r9,%rax,2),%xmm1 + DB 233,26,255,255,255 ; jmpq 3e13 <_sk_load_u16_be_sse41+0x23> PUBLIC _sk_load_rgb_u16_be_sse41 _sk_load_rgb_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,170,0,0,0 ; jne 3f82 <_sk_load_rgb_u16_be_sse41+0xbc> - DB 243,65,15,111,20,64 ; movdqu (%r8,%rax,2),%xmm2 - DB 243,65,15,111,92,64,8 ; movdqu 0x8(%r8,%rax,2),%xmm3 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,82 ; lea (%rdx,%rdx,2),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,170,0,0,0 ; jne 3fb5 <_sk_load_rgb_u16_be_sse41+0xbc> + DB 243,65,15,111,20,65 ; movdqu (%r9,%rax,2),%xmm2 + DB 243,65,15,111,92,65,8 ; movdqu 0x8(%r9,%rax,2),%xmm3 DB 102,15,115,219,4 ; psrldq $0x4,%xmm3 DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 @@ -15646,7 +15673,7 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 102,15,235,200 ; por %xmm0,%xmm1 DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,249,31,0,0 ; movaps 0x1ff9(%rip),%xmm8 # 5f30 <_sk_callback_sse41+0xedd> + DB 68,15,40,5,22,32,0,0 ; movaps 0x2016(%rip),%xmm8 # 5f80 <_sk_callback_sse41+0xef3> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -15663,34 +15690,34 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,192,31,0,0 ; movaps 0x1fc0(%rip),%xmm3 # 5f40 <_sk_callback_sse41+0xeed> + DB 15,40,29,221,31,0,0 ; movaps 0x1fdd(%rip),%xmm3 # 5f90 <_sk_callback_sse41+0xf03> DB 255,224 ; jmpq *%rax - DB 102,65,15,110,20,64 ; movd (%r8,%rax,2),%xmm2 - DB 102,65,15,196,84,64,4,2 ; pinsrw $0x2,0x4(%r8,%rax,2),%xmm2 + DB 102,65,15,110,20,65 ; movd (%r9,%rax,2),%xmm2 + DB 102,65,15,196,84,65,4,2 ; pinsrw $0x2,0x4(%r9,%rax,2),%xmm2 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,13 ; jne 3fa7 <_sk_load_rgb_u16_be_sse41+0xe1> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,13 ; jne 3fda <_sk_load_rgb_u16_be_sse41+0xe1> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,85,255,255,255 ; jmpq 3efc <_sk_load_rgb_u16_be_sse41+0x36> - DB 102,65,15,110,68,64,6 ; movd 0x6(%r8,%rax,2),%xmm0 - DB 102,65,15,196,68,64,10,2 ; pinsrw $0x2,0xa(%r8,%rax,2),%xmm0 + DB 233,85,255,255,255 ; jmpq 3f2f <_sk_load_rgb_u16_be_sse41+0x36> + DB 102,65,15,110,68,65,6 ; movd 0x6(%r9,%rax,2),%xmm0 + DB 102,65,15,196,68,65,10,2 ; pinsrw $0x2,0xa(%r9,%rax,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,24 ; jb 3fd8 <_sk_load_rgb_u16_be_sse41+0x112> - DB 102,65,15,110,92,64,12 ; movd 0xc(%r8,%rax,2),%xmm3 - DB 102,65,15,196,92,64,16,2 ; pinsrw $0x2,0x10(%r8,%rax,2),%xmm3 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,24 ; jb 400b <_sk_load_rgb_u16_be_sse41+0x112> + DB 102,65,15,110,92,65,12 ; movd 0xc(%r9,%rax,2),%xmm3 + DB 102,65,15,196,92,65,16,2 ; pinsrw $0x2,0x10(%r9,%rax,2),%xmm3 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,36,255,255,255 ; jmpq 3efc <_sk_load_rgb_u16_be_sse41+0x36> + DB 233,36,255,255,255 ; jmpq 3f2f <_sk_load_rgb_u16_be_sse41+0x36> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 233,27,255,255,255 ; jmpq 3efc <_sk_load_rgb_u16_be_sse41+0x36> + DB 233,27,255,255,255 ; jmpq 3f2f <_sk_load_rgb_u16_be_sse41+0x36> PUBLIC _sk_store_u16_be_sse41 _sk_store_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 68,15,40,21,90,31,0,0 ; movaps 0x1f5a(%rip),%xmm10 # 5f50 <_sk_callback_sse41+0xefd> + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 68,15,40,21,119,31,0,0 ; movaps 0x1f77(%rip),%xmm10 # 5fa0 <_sk_callback_sse41+0xf13> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 @@ -15726,36 +15753,36 @@ _sk_store_u16_be_sse41 LABEL PROC DB 102,69,15,97,202 ; punpcklwd %xmm10,%xmm9 DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 DB 102,69,15,98,209 ; punpckldq %xmm9,%xmm10 - DB 72,133,201 ; test %rcx,%rcx - DB 117,21 ; jne 40c4 <_sk_store_u16_be_sse41+0xe3> - DB 69,15,17,20,64 ; movups %xmm10,(%r8,%rax,2) + DB 77,133,192 ; test %r8,%r8 + DB 117,21 ; jne 40f7 <_sk_store_u16_be_sse41+0xe3> + DB 69,15,17,20,65 ; movups %xmm10,(%r9,%rax,2) DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 - DB 243,69,15,127,68,64,16 ; movdqu %xmm8,0x10(%r8,%rax,2) + DB 243,69,15,127,68,65,16 ; movdqu %xmm8,0x10(%r9,%rax,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 102,69,15,214,20,64 ; movq %xmm10,(%r8,%rax,2) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 40c0 <_sk_store_u16_be_sse41+0xdf> - DB 102,69,15,23,84,64,8 ; movhpd %xmm10,0x8(%r8,%rax,2) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 40c0 <_sk_store_u16_be_sse41+0xdf> + DB 102,69,15,214,20,65 ; movq %xmm10,(%r9,%rax,2) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 40f3 <_sk_store_u16_be_sse41+0xdf> + DB 102,69,15,23,84,65,8 ; movhpd %xmm10,0x8(%r9,%rax,2) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 40f3 <_sk_store_u16_be_sse41+0xdf> DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 - DB 102,69,15,214,68,64,16 ; movq %xmm8,0x10(%r8,%rax,2) - DB 235,213 ; jmp 40c0 <_sk_store_u16_be_sse41+0xdf> + DB 102,69,15,214,68,65,16 ; movq %xmm8,0x10(%r9,%rax,2) + DB 235,213 ; jmp 40f3 <_sk_store_u16_be_sse41+0xdf> PUBLIC _sk_load_f32_sse41 _sk_load_f32_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 76,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r8 - DB 72,137,248 ; mov %rdi,%rax + DB 76,139,16 ; mov (%rax),%r10 + DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 + DB 72,137,208 ; mov %rdx,%rax DB 72,193,224,4 ; shl $0x4,%rax - DB 69,15,16,4,1 ; movups (%r9,%rax,1),%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,66 ; jne 414b <_sk_load_f32_sse41+0x60> - DB 67,15,16,68,129,16 ; movups 0x10(%r9,%r8,4),%xmm0 - DB 67,15,16,92,129,32 ; movups 0x20(%r9,%r8,4),%xmm3 - DB 71,15,16,76,129,48 ; movups 0x30(%r9,%r8,4),%xmm9 + DB 69,15,16,4,2 ; movups (%r10,%rax,1),%xmm8 + DB 77,133,192 ; test %r8,%r8 + DB 117,66 ; jne 417e <_sk_load_f32_sse41+0x60> + DB 67,15,16,68,138,16 ; movups 0x10(%r10,%r9,4),%xmm0 + DB 67,15,16,92,138,32 ; movups 0x20(%r10,%r9,4),%xmm3 + DB 71,15,16,76,138,48 ; movups 0x30(%r10,%r9,4),%xmm9 DB 65,15,40,208 ; movaps %xmm8,%xmm2 DB 15,20,208 ; unpcklps %xmm0,%xmm2 DB 15,40,203 ; movaps %xmm3,%xmm1 @@ -15771,25 +15798,25 @@ _sk_load_f32_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,8 ; jne 415d <_sk_load_f32_sse41+0x72> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,8 ; jne 4190 <_sk_load_f32_sse41+0x72> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 - DB 235,190 ; jmp 411b <_sk_load_f32_sse41+0x30> - DB 67,15,16,68,129,16 ; movups 0x10(%r9,%r8,4),%xmm0 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,8 ; jb 4171 <_sk_load_f32_sse41+0x86> - DB 67,15,16,92,129,32 ; movups 0x20(%r9,%r8,4),%xmm3 - DB 235,170 ; jmp 411b <_sk_load_f32_sse41+0x30> + DB 235,190 ; jmp 414e <_sk_load_f32_sse41+0x30> + DB 67,15,16,68,138,16 ; movups 0x10(%r10,%r9,4),%xmm0 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,8 ; jb 41a4 <_sk_load_f32_sse41+0x86> + DB 67,15,16,92,138,32 ; movups 0x20(%r10,%r9,4),%xmm3 + DB 235,170 ; jmp 414e <_sk_load_f32_sse41+0x30> DB 15,87,219 ; xorps %xmm3,%xmm3 - DB 235,165 ; jmp 411b <_sk_load_f32_sse41+0x30> + DB 235,165 ; jmp 414e <_sk_load_f32_sse41+0x30> PUBLIC _sk_store_f32_sse41 _sk_store_f32_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 76,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r8 - DB 72,137,248 ; mov %rdi,%rax + DB 76,139,16 ; mov (%rax),%r10 + DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 + DB 72,137,208 ; mov %rdx,%rax DB 72,193,224,4 ; shl $0x4,%rax DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 @@ -15804,22 +15831,22 @@ _sk_store_f32_sse41 LABEL PROC DB 69,15,18,193 ; movhlps %xmm9,%xmm8 DB 69,15,40,202 ; movaps %xmm10,%xmm9 DB 102,69,15,20,203 ; unpcklpd %xmm11,%xmm9 - DB 102,69,15,17,36,1 ; movupd %xmm12,(%r9,%rax,1) - DB 72,133,201 ; test %rcx,%rcx - DB 117,29 ; jne 41e8 <_sk_store_f32_sse41+0x72> + DB 102,69,15,17,36,2 ; movupd %xmm12,(%r10,%rax,1) + DB 77,133,192 ; test %r8,%r8 + DB 117,29 ; jne 421b <_sk_store_f32_sse41+0x72> DB 102,69,15,21,211 ; unpckhpd %xmm11,%xmm10 - DB 71,15,17,68,129,16 ; movups %xmm8,0x10(%r9,%r8,4) - DB 102,71,15,17,76,129,32 ; movupd %xmm9,0x20(%r9,%r8,4) - DB 102,71,15,17,84,129,48 ; movupd %xmm10,0x30(%r9,%r8,4) + DB 71,15,17,68,138,16 ; movups %xmm8,0x10(%r10,%r9,4) + DB 102,71,15,17,76,138,32 ; movupd %xmm9,0x20(%r10,%r9,4) + DB 102,71,15,17,84,138,48 ; movupd %xmm10,0x30(%r10,%r9,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,246 ; je 41e4 <_sk_store_f32_sse41+0x6e> - DB 71,15,17,68,129,16 ; movups %xmm8,0x10(%r9,%r8,4) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,234 ; jb 41e4 <_sk_store_f32_sse41+0x6e> - DB 102,71,15,17,76,129,32 ; movupd %xmm9,0x20(%r9,%r8,4) - DB 235,225 ; jmp 41e4 <_sk_store_f32_sse41+0x6e> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,246 ; je 4217 <_sk_store_f32_sse41+0x6e> + DB 71,15,17,68,138,16 ; movups %xmm8,0x10(%r10,%r9,4) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,234 ; jb 4217 <_sk_store_f32_sse41+0x6e> + DB 102,71,15,17,76,138,32 ; movupd %xmm9,0x20(%r10,%r9,4) + DB 235,225 ; jmp 4217 <_sk_store_f32_sse41+0x6e> PUBLIC _sk_clamp_x_sse41 _sk_clamp_x_sse41 LABEL PROC @@ -15917,7 +15944,7 @@ PUBLIC _sk_clamp_x_1_sse41 _sk_clamp_x_1_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,49,28,0,0 ; minps 0x1c31(%rip),%xmm8 # 5f60 <_sk_callback_sse41+0xf0d> + DB 68,15,93,5,78,28,0,0 ; minps 0x1c4e(%rip),%xmm8 # 5fb0 <_sk_callback_sse41+0xf23> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -15931,9 +15958,9 @@ _sk_repeat_x_1_sse41 LABEL PROC PUBLIC _sk_mirror_x_1_sse41 _sk_mirror_x_1_sse41 LABEL PROC - DB 68,15,40,5,34,28,0,0 ; movaps 0x1c22(%rip),%xmm8 # 5f70 <_sk_callback_sse41+0xf1d> + DB 68,15,40,5,63,28,0,0 ; movaps 0x1c3f(%rip),%xmm8 # 5fc0 <_sk_callback_sse41+0xf33> DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,13,38,28,0,0 ; movaps 0x1c26(%rip),%xmm9 # 5f80 <_sk_callback_sse41+0xf2d> + DB 68,15,40,13,67,28,0,0 ; movaps 0x1c43(%rip),%xmm9 # 5fd0 <_sk_callback_sse41+0xf43> DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 DB 69,15,88,201 ; addps %xmm9,%xmm9 @@ -15948,10 +15975,10 @@ _sk_mirror_x_1_sse41 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse41 _sk_luminance_to_alpha_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,5,28,0,0 ; mulps 0x1c05(%rip),%xmm0 # 5f90 <_sk_callback_sse41+0xf3d> - DB 15,89,13,14,28,0,0 ; mulps 0x1c0e(%rip),%xmm1 # 5fa0 <_sk_callback_sse41+0xf4d> + DB 15,89,5,34,28,0,0 ; mulps 0x1c22(%rip),%xmm0 # 5fe0 <_sk_callback_sse41+0xf53> + DB 15,89,13,43,28,0,0 ; mulps 0x1c2b(%rip),%xmm1 # 5ff0 <_sk_callback_sse41+0xf63> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,20,28,0,0 ; mulps 0x1c14(%rip),%xmm3 # 5fb0 <_sk_callback_sse41+0xf5d> + DB 15,89,29,49,28,0,0 ; mulps 0x1c31(%rip),%xmm3 # 6000 <_sk_callback_sse41+0xf73> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -16211,84 +16238,85 @@ _sk_matrix_perspective_sse41 LABEL PROC PUBLIC _sk_evenly_spaced_gradient_sse41 _sk_evenly_spaced_gradient_sse41 LABEL PROC + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,24 ; mov (%rax),%rbx DB 76,139,112,8 ; mov 0x8(%rax),%r14 DB 72,255,203 ; dec %rbx - DB 120,7 ; js 47d8 <_sk_evenly_spaced_gradient_sse41+0x18> + DB 120,7 ; js 480d <_sk_evenly_spaced_gradient_sse41+0x1a> DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 - DB 235,21 ; jmp 47ed <_sk_evenly_spaced_gradient_sse41+0x2d> - DB 73,137,216 ; mov %rbx,%r8 - DB 73,209,232 ; shr %r8 + DB 235,21 ; jmp 4822 <_sk_evenly_spaced_gradient_sse41+0x2f> + DB 73,137,217 ; mov %rbx,%r9 + DB 73,209,233 ; shr %r9 DB 131,227,1 ; and $0x1,%ebx - DB 76,9,195 ; or %r8,%rbx + DB 76,9,203 ; or %r9,%rbx DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 DB 243,15,88,201 ; addss %xmm1,%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 - DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,202 ; movq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,71,15,16,4,158 ; movss (%r14,%r11,4),%xmm8 - DB 102,71,15,58,33,4,150,16 ; insertps $0x10,(%r14,%r10,4),%xmm8 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 + DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 102,73,15,126,203 ; movq %xmm1,%r11 + DB 69,137,223 ; mov %r11d,%r15d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,71,15,16,4,190 ; movss (%r14,%r15,4),%xmm8 + DB 102,71,15,58,33,4,158,16 ; insertps $0x10,(%r14,%r11,4),%xmm8 + DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 DB 102,68,15,58,33,193,32 ; insertps $0x20,%xmm1,%xmm8 - DB 243,67,15,16,12,134 ; movss (%r14,%r8,4),%xmm1 + DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 DB 102,68,15,58,33,193,48 ; insertps $0x30,%xmm1,%xmm8 DB 72,139,88,40 ; mov 0x28(%rax),%rbx - DB 243,70,15,16,12,155 ; movss (%rbx,%r11,4),%xmm9 - DB 102,70,15,58,33,12,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm9 - DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,70,15,16,12,187 ; movss (%rbx,%r15,4),%xmm9 + DB 102,70,15,58,33,12,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm9 + DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 DB 102,68,15,58,33,201,32 ; insertps $0x20,%xmm1,%xmm9 - DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 DB 102,68,15,58,33,201,48 ; insertps $0x30,%xmm1,%xmm9 DB 72,139,88,16 ; mov 0x10(%rax),%rbx - DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 102,66,15,58,33,12,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm1 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,66,15,16,12,187 ; movss (%rbx,%r15,4),%xmm1 + DB 102,66,15,58,33,12,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm1 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 DB 72,139,88,48 ; mov 0x30(%rax),%rbx - DB 243,70,15,16,20,155 ; movss (%rbx,%r11,4),%xmm10 - DB 102,70,15,58,33,20,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm10 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,70,15,16,20,187 ; movss (%rbx,%r15,4),%xmm10 + DB 102,70,15,58,33,20,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm10 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 102,68,15,58,33,210,32 ; insertps $0x20,%xmm2,%xmm10 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 102,68,15,58,33,210,48 ; insertps $0x30,%xmm2,%xmm10 DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 102,66,15,58,33,20,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm2 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,66,15,16,20,187 ; movss (%rbx,%r15,4),%xmm2 + DB 102,66,15,58,33,20,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm2 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 243,70,15,16,28,155 ; movss (%rbx,%r11,4),%xmm11 - DB 102,70,15,58,33,28,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm11 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,70,15,16,28,187 ; movss (%rbx,%r15,4),%xmm11 + DB 102,70,15,58,33,28,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm11 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 DB 102,68,15,58,33,219,32 ; insertps $0x20,%xmm3,%xmm11 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 102,68,15,58,33,219,48 ; insertps $0x30,%xmm3,%xmm11 DB 72,139,88,32 ; mov 0x20(%rax),%rbx - DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 - DB 102,66,15,58,33,28,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm3 - DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 243,66,15,16,28,187 ; movss (%rbx,%r15,4),%xmm3 + DB 102,66,15,58,33,28,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm3 + DB 243,70,15,16,36,147 ; movss (%rbx,%r10,4),%xmm12 DB 102,65,15,58,33,220,32 ; insertps $0x20,%xmm12,%xmm3 - DB 243,70,15,16,36,131 ; movss (%rbx,%r8,4),%xmm12 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 DB 102,65,15,58,33,220,48 ; insertps $0x30,%xmm12,%xmm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 243,70,15,16,36,152 ; movss (%rax,%r11,4),%xmm12 - DB 102,70,15,58,33,36,144,16 ; insertps $0x10,(%rax,%r10,4),%xmm12 - DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 + DB 243,70,15,16,36,184 ; movss (%rax,%r15,4),%xmm12 + DB 102,70,15,58,33,36,152,16 ; insertps $0x10,(%rax,%r11,4),%xmm12 + DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 DB 102,69,15,58,33,229,32 ; insertps $0x20,%xmm13,%xmm12 - DB 243,70,15,16,44,128 ; movss (%rax,%r8,4),%xmm13 + DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 DB 102,69,15,58,33,229,48 ; insertps $0x30,%xmm13,%xmm12 DB 68,15,89,192 ; mulps %xmm0,%xmm8 DB 69,15,88,193 ; addps %xmm9,%xmm8 @@ -16302,19 +16330,20 @@ _sk_evenly_spaced_gradient_sse41 LABEL PROC DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax PUBLIC _sk_gauss_a_to_rgba_sse41 _sk_gauss_a_to_rgba_sse41 LABEL PROC - DB 15,40,5,39,22,0,0 ; movaps 0x1627(%rip),%xmm0 # 5fc0 <_sk_callback_sse41+0xf6d> + DB 15,40,5,64,22,0,0 ; movaps 0x1640(%rip),%xmm0 # 6010 <_sk_callback_sse41+0xf83> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,45,22,0,0 ; addps 0x162d(%rip),%xmm0 # 5fd0 <_sk_callback_sse41+0xf7d> + DB 15,88,5,70,22,0,0 ; addps 0x1646(%rip),%xmm0 # 6020 <_sk_callback_sse41+0xf93> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,51,22,0,0 ; addps 0x1633(%rip),%xmm0 # 5fe0 <_sk_callback_sse41+0xf8d> + DB 15,88,5,76,22,0,0 ; addps 0x164c(%rip),%xmm0 # 6030 <_sk_callback_sse41+0xfa3> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,57,22,0,0 ; addps 0x1639(%rip),%xmm0 # 5ff0 <_sk_callback_sse41+0xf9d> + DB 15,88,5,82,22,0,0 ; addps 0x1652(%rip),%xmm0 # 6040 <_sk_callback_sse41+0xfb3> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,63,22,0,0 ; addps 0x163f(%rip),%xmm0 # 6000 <_sk_callback_sse41+0xfad> + DB 15,88,5,88,22,0,0 ; addps 0x1658(%rip),%xmm0 # 6050 <_sk_callback_sse41+0xfc3> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -16323,87 +16352,88 @@ _sk_gauss_a_to_rgba_sse41 LABEL PROC PUBLIC _sk_gradient_sse41 _sk_gradient_sse41 LABEL PROC + DB 65,87 ; push %r15 + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,51 ; jb 4a10 <_sk_gradient_sse41+0x42> - DB 76,139,72,72 ; mov 0x48(%rax),%r9 - DB 73,255,200 ; dec %r8 - DB 73,131,193,4 ; add $0x4,%r9 + DB 73,131,249,2 ; cmp $0x2,%r9 + DB 114,50 ; jb 4a4b <_sk_gradient_sse41+0x46> + DB 72,139,88,72 ; mov 0x48(%rax),%rbx + DB 73,255,201 ; dec %r9 + DB 72,131,195,4 ; add $0x4,%rbx DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,21,29,22,0,0 ; movaps 0x161d(%rip),%xmm2 # 6010 <_sk_callback_sse41+0xfbd> - DB 243,65,15,16,25 ; movss (%r9),%xmm3 + DB 15,40,21,49,22,0,0 ; movaps 0x1631(%rip),%xmm2 # 6060 <_sk_callback_sse41+0xfd3> + DB 243,15,16,27 ; movss (%rbx),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 DB 15,84,218 ; andps %xmm2,%xmm3 DB 102,15,254,203 ; paddd %xmm3,%xmm1 - DB 73,131,193,4 ; add $0x4,%r9 - DB 73,255,200 ; dec %r8 - DB 117,227 ; jne 49f3 <_sk_gradient_sse41+0x25> - DB 65,86 ; push %r14 - DB 83 ; push %rbx - DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,202 ; movq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 72,131,195,4 ; add $0x4,%rbx + DB 73,255,201 ; dec %r9 + DB 117,228 ; jne 4a2f <_sk_gradient_sse41+0x2a> + DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 102,73,15,126,203 ; movq %xmm1,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,70,15,16,4,155 ; movss (%rbx,%r11,4),%xmm8 - DB 102,70,15,58,33,4,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm8 - DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,70,15,16,4,179 ; movss (%rbx,%r14,4),%xmm8 + DB 102,70,15,58,33,4,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm8 + DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 DB 102,68,15,58,33,193,32 ; insertps $0x20,%xmm1,%xmm8 - DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 DB 102,68,15,58,33,193,48 ; insertps $0x30,%xmm1,%xmm8 DB 72,139,88,40 ; mov 0x28(%rax),%rbx - DB 243,70,15,16,12,155 ; movss (%rbx,%r11,4),%xmm9 - DB 102,70,15,58,33,12,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm9 - DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,70,15,16,12,179 ; movss (%rbx,%r14,4),%xmm9 + DB 102,70,15,58,33,12,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm9 + DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 DB 102,68,15,58,33,201,32 ; insertps $0x20,%xmm1,%xmm9 - DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 DB 102,68,15,58,33,201,48 ; insertps $0x30,%xmm1,%xmm9 - DB 243,67,15,16,12,158 ; movss (%r14,%r11,4),%xmm1 - DB 102,67,15,58,33,12,150,16 ; insertps $0x10,(%r14,%r10,4),%xmm1 - DB 243,67,15,16,20,142 ; movss (%r14,%r9,4),%xmm2 + DB 243,67,15,16,12,183 ; movss (%r15,%r14,4),%xmm1 + DB 102,67,15,58,33,12,159,16 ; insertps $0x10,(%r15,%r11,4),%xmm1 + DB 243,67,15,16,20,151 ; movss (%r15,%r10,4),%xmm2 DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1 - DB 243,67,15,16,20,134 ; movss (%r14,%r8,4),%xmm2 + DB 243,67,15,16,20,143 ; movss (%r15,%r9,4),%xmm2 DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 DB 72,139,88,48 ; mov 0x30(%rax),%rbx - DB 243,70,15,16,20,155 ; movss (%rbx,%r11,4),%xmm10 - DB 102,70,15,58,33,20,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm10 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,70,15,16,20,179 ; movss (%rbx,%r14,4),%xmm10 + DB 102,70,15,58,33,20,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm10 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 102,68,15,58,33,210,32 ; insertps $0x20,%xmm2,%xmm10 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 102,68,15,58,33,210,48 ; insertps $0x30,%xmm2,%xmm10 DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 102,66,15,58,33,20,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm2 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,66,15,16,20,179 ; movss (%rbx,%r14,4),%xmm2 + DB 102,66,15,58,33,20,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm2 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 243,70,15,16,28,155 ; movss (%rbx,%r11,4),%xmm11 - DB 102,70,15,58,33,28,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm11 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,70,15,16,28,179 ; movss (%rbx,%r14,4),%xmm11 + DB 102,70,15,58,33,28,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm11 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 DB 102,68,15,58,33,219,32 ; insertps $0x20,%xmm3,%xmm11 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 102,68,15,58,33,219,48 ; insertps $0x30,%xmm3,%xmm11 DB 72,139,88,32 ; mov 0x20(%rax),%rbx - DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 - DB 102,66,15,58,33,28,147,16 ; insertps $0x10,(%rbx,%r10,4),%xmm3 - DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 243,66,15,16,28,179 ; movss (%rbx,%r14,4),%xmm3 + DB 102,66,15,58,33,28,155,16 ; insertps $0x10,(%rbx,%r11,4),%xmm3 + DB 243,70,15,16,36,147 ; movss (%rbx,%r10,4),%xmm12 DB 102,65,15,58,33,220,32 ; insertps $0x20,%xmm12,%xmm3 - DB 243,70,15,16,36,131 ; movss (%rbx,%r8,4),%xmm12 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 DB 102,65,15,58,33,220,48 ; insertps $0x30,%xmm12,%xmm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 243,70,15,16,36,152 ; movss (%rax,%r11,4),%xmm12 - DB 102,70,15,58,33,36,144,16 ; insertps $0x10,(%rax,%r10,4),%xmm12 - DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 + DB 243,70,15,16,36,176 ; movss (%rax,%r14,4),%xmm12 + DB 102,70,15,58,33,36,152,16 ; insertps $0x10,(%rax,%r11,4),%xmm12 + DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 DB 102,69,15,58,33,229,32 ; insertps $0x20,%xmm13,%xmm12 - DB 243,70,15,16,44,128 ; movss (%rax,%r8,4),%xmm13 + DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 DB 102,69,15,58,33,229,48 ; insertps $0x30,%xmm13,%xmm12 DB 68,15,89,192 ; mulps %xmm0,%xmm8 DB 69,15,88,193 ; addps %xmm9,%xmm8 @@ -16417,6 +16447,7 @@ _sk_gradient_sse41 LABEL PROC DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax PUBLIC _sk_evenly_spaced_2_stop_gradient_sse41 @@ -16467,26 +16498,26 @@ _sk_xy_to_unit_angle_sse41 LABEL PROC DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,40,236 ; movaps %xmm12,%xmm13 DB 69,15,89,237 ; mulps %xmm13,%xmm13 - DB 68,15,40,21,188,19,0,0 ; movaps 0x13bc(%rip),%xmm10 # 6020 <_sk_callback_sse41+0xfcd> + DB 68,15,40,21,210,19,0,0 ; movaps 0x13d2(%rip),%xmm10 # 6070 <_sk_callback_sse41+0xfe3> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,192,19,0,0 ; addps 0x13c0(%rip),%xmm10 # 6030 <_sk_callback_sse41+0xfdd> + DB 68,15,88,21,214,19,0,0 ; addps 0x13d6(%rip),%xmm10 # 6080 <_sk_callback_sse41+0xff3> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,196,19,0,0 ; addps 0x13c4(%rip),%xmm10 # 6040 <_sk_callback_sse41+0xfed> + DB 68,15,88,21,218,19,0,0 ; addps 0x13da(%rip),%xmm10 # 6090 <_sk_callback_sse41+0x1003> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,200,19,0,0 ; addps 0x13c8(%rip),%xmm10 # 6050 <_sk_callback_sse41+0xffd> + DB 68,15,88,21,222,19,0,0 ; addps 0x13de(%rip),%xmm10 # 60a0 <_sk_callback_sse41+0x1013> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 68,15,40,29,199,19,0,0 ; movaps 0x13c7(%rip),%xmm11 # 6060 <_sk_callback_sse41+0x100d> + DB 68,15,40,29,221,19,0,0 ; movaps 0x13dd(%rip),%xmm11 # 60b0 <_sk_callback_sse41+0x1023> DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 102,69,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm10 DB 69,15,194,200,1 ; cmpltps %xmm8,%xmm9 - DB 68,15,40,29,192,19,0,0 ; movaps 0x13c0(%rip),%xmm11 # 6070 <_sk_callback_sse41+0x101d> + DB 68,15,40,29,214,19,0,0 ; movaps 0x13d6(%rip),%xmm11 # 60c0 <_sk_callback_sse41+0x1033> DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 102,69,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm10 DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,194,192,1 ; cmpltps %xmm8,%xmm0 - DB 68,15,40,13,178,19,0,0 ; movaps 0x13b2(%rip),%xmm9 # 6080 <_sk_callback_sse41+0x102d> + DB 68,15,40,13,200,19,0,0 ; movaps 0x13c8(%rip),%xmm9 # 60d0 <_sk_callback_sse41+0x1043> DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 102,69,15,56,20,209 ; blendvps %xmm0,%xmm9,%xmm10 DB 69,15,194,194,7 ; cmpordps %xmm10,%xmm8 @@ -16508,7 +16539,7 @@ _sk_xy_to_radius_sse41 LABEL PROC PUBLIC _sk_save_xy_sse41 _sk_save_xy_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,134,19,0,0 ; movaps 0x1386(%rip),%xmm8 # 6090 <_sk_callback_sse41+0x103d> + DB 68,15,40,5,156,19,0,0 ; movaps 0x139c(%rip),%xmm8 # 60e0 <_sk_callback_sse41+0x1053> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -16548,8 +16579,8 @@ _sk_bilinear_nx_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,8,19,0,0 ; addps 0x1308(%rip),%xmm0 # 60a0 <_sk_callback_sse41+0x104d> - DB 68,15,40,13,16,19,0,0 ; movaps 0x1310(%rip),%xmm9 # 60b0 <_sk_callback_sse41+0x105d> + DB 15,88,5,30,19,0,0 ; addps 0x131e(%rip),%xmm0 # 60f0 <_sk_callback_sse41+0x1063> + DB 68,15,40,13,38,19,0,0 ; movaps 0x1326(%rip),%xmm9 # 6100 <_sk_callback_sse41+0x1073> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -16560,7 +16591,7 @@ _sk_bilinear_px_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,255,18,0,0 ; addps 0x12ff(%rip),%xmm0 # 60c0 <_sk_callback_sse41+0x106d> + DB 15,88,5,21,19,0,0 ; addps 0x1315(%rip),%xmm0 # 6110 <_sk_callback_sse41+0x1083> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16570,8 +16601,8 @@ _sk_bilinear_ny_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,241,18,0,0 ; addps 0x12f1(%rip),%xmm1 # 60d0 <_sk_callback_sse41+0x107d> - DB 68,15,40,13,249,18,0,0 ; movaps 0x12f9(%rip),%xmm9 # 60e0 <_sk_callback_sse41+0x108d> + DB 15,88,13,7,19,0,0 ; addps 0x1307(%rip),%xmm1 # 6120 <_sk_callback_sse41+0x1093> + DB 68,15,40,13,15,19,0,0 ; movaps 0x130f(%rip),%xmm9 # 6130 <_sk_callback_sse41+0x10a3> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -16582,7 +16613,7 @@ _sk_bilinear_py_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,231,18,0,0 ; addps 0x12e7(%rip),%xmm1 # 60f0 <_sk_callback_sse41+0x109d> + DB 15,88,13,253,18,0,0 ; addps 0x12fd(%rip),%xmm1 # 6140 <_sk_callback_sse41+0x10b3> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16592,13 +16623,13 @@ _sk_bicubic_n3x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,218,18,0,0 ; addps 0x12da(%rip),%xmm0 # 6100 <_sk_callback_sse41+0x10ad> - DB 68,15,40,13,226,18,0,0 ; movaps 0x12e2(%rip),%xmm9 # 6110 <_sk_callback_sse41+0x10bd> + DB 15,88,5,240,18,0,0 ; addps 0x12f0(%rip),%xmm0 # 6150 <_sk_callback_sse41+0x10c3> + DB 68,15,40,13,248,18,0,0 ; movaps 0x12f8(%rip),%xmm9 # 6160 <_sk_callback_sse41+0x10d3> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,222,18,0,0 ; mulps 0x12de(%rip),%xmm9 # 6120 <_sk_callback_sse41+0x10cd> - DB 68,15,88,13,230,18,0,0 ; addps 0x12e6(%rip),%xmm9 # 6130 <_sk_callback_sse41+0x10dd> + DB 68,15,89,13,244,18,0,0 ; mulps 0x12f4(%rip),%xmm9 # 6170 <_sk_callback_sse41+0x10e3> + DB 68,15,88,13,252,18,0,0 ; addps 0x12fc(%rip),%xmm9 # 6180 <_sk_callback_sse41+0x10f3> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -16609,16 +16640,16 @@ _sk_bicubic_n1x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,213,18,0,0 ; addps 0x12d5(%rip),%xmm0 # 6140 <_sk_callback_sse41+0x10ed> - DB 68,15,40,13,221,18,0,0 ; movaps 0x12dd(%rip),%xmm9 # 6150 <_sk_callback_sse41+0x10fd> + DB 15,88,5,235,18,0,0 ; addps 0x12eb(%rip),%xmm0 # 6190 <_sk_callback_sse41+0x1103> + DB 68,15,40,13,243,18,0,0 ; movaps 0x12f3(%rip),%xmm9 # 61a0 <_sk_callback_sse41+0x1113> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,225,18,0,0 ; movaps 0x12e1(%rip),%xmm8 # 6160 <_sk_callback_sse41+0x110d> + DB 68,15,40,5,247,18,0,0 ; movaps 0x12f7(%rip),%xmm8 # 61b0 <_sk_callback_sse41+0x1123> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,229,18,0,0 ; addps 0x12e5(%rip),%xmm8 # 6170 <_sk_callback_sse41+0x111d> + DB 68,15,88,5,251,18,0,0 ; addps 0x12fb(%rip),%xmm8 # 61c0 <_sk_callback_sse41+0x1133> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,233,18,0,0 ; addps 0x12e9(%rip),%xmm8 # 6180 <_sk_callback_sse41+0x112d> + DB 68,15,88,5,255,18,0,0 ; addps 0x12ff(%rip),%xmm8 # 61d0 <_sk_callback_sse41+0x1143> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,237,18,0,0 ; addps 0x12ed(%rip),%xmm8 # 6190 <_sk_callback_sse41+0x113d> + DB 68,15,88,5,3,19,0,0 ; addps 0x1303(%rip),%xmm8 # 61e0 <_sk_callback_sse41+0x1153> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16626,17 +16657,17 @@ _sk_bicubic_n1x_sse41 LABEL PROC PUBLIC _sk_bicubic_p1x_sse41 _sk_bicubic_p1x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,231,18,0,0 ; movaps 0x12e7(%rip),%xmm8 # 61a0 <_sk_callback_sse41+0x114d> + DB 68,15,40,5,253,18,0,0 ; movaps 0x12fd(%rip),%xmm8 # 61f0 <_sk_callback_sse41+0x1163> DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,72,64 ; movups 0x40(%rax),%xmm9 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,21,227,18,0,0 ; movaps 0x12e3(%rip),%xmm10 # 61b0 <_sk_callback_sse41+0x115d> + DB 68,15,40,21,249,18,0,0 ; movaps 0x12f9(%rip),%xmm10 # 6200 <_sk_callback_sse41+0x1173> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,231,18,0,0 ; addps 0x12e7(%rip),%xmm10 # 61c0 <_sk_callback_sse41+0x116d> + DB 68,15,88,21,253,18,0,0 ; addps 0x12fd(%rip),%xmm10 # 6210 <_sk_callback_sse41+0x1183> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,227,18,0,0 ; addps 0x12e3(%rip),%xmm10 # 61d0 <_sk_callback_sse41+0x117d> + DB 68,15,88,21,249,18,0,0 ; addps 0x12f9(%rip),%xmm10 # 6220 <_sk_callback_sse41+0x1193> DB 68,15,17,144,128,0,0,0 ; movups %xmm10,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16646,11 +16677,11 @@ _sk_bicubic_p3x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,214,18,0,0 ; addps 0x12d6(%rip),%xmm0 # 61e0 <_sk_callback_sse41+0x118d> + DB 15,88,5,236,18,0,0 ; addps 0x12ec(%rip),%xmm0 # 6230 <_sk_callback_sse41+0x11a3> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,214,18,0,0 ; mulps 0x12d6(%rip),%xmm8 # 61f0 <_sk_callback_sse41+0x119d> - DB 68,15,88,5,222,18,0,0 ; addps 0x12de(%rip),%xmm8 # 6200 <_sk_callback_sse41+0x11ad> + DB 68,15,89,5,236,18,0,0 ; mulps 0x12ec(%rip),%xmm8 # 6240 <_sk_callback_sse41+0x11b3> + DB 68,15,88,5,244,18,0,0 ; addps 0x12f4(%rip),%xmm8 # 6250 <_sk_callback_sse41+0x11c3> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -16661,13 +16692,13 @@ _sk_bicubic_n3y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,204,18,0,0 ; addps 0x12cc(%rip),%xmm1 # 6210 <_sk_callback_sse41+0x11bd> - DB 68,15,40,13,212,18,0,0 ; movaps 0x12d4(%rip),%xmm9 # 6220 <_sk_callback_sse41+0x11cd> + DB 15,88,13,226,18,0,0 ; addps 0x12e2(%rip),%xmm1 # 6260 <_sk_callback_sse41+0x11d3> + DB 68,15,40,13,234,18,0,0 ; movaps 0x12ea(%rip),%xmm9 # 6270 <_sk_callback_sse41+0x11e3> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,208,18,0,0 ; mulps 0x12d0(%rip),%xmm9 # 6230 <_sk_callback_sse41+0x11dd> - DB 68,15,88,13,216,18,0,0 ; addps 0x12d8(%rip),%xmm9 # 6240 <_sk_callback_sse41+0x11ed> + DB 68,15,89,13,230,18,0,0 ; mulps 0x12e6(%rip),%xmm9 # 6280 <_sk_callback_sse41+0x11f3> + DB 68,15,88,13,238,18,0,0 ; addps 0x12ee(%rip),%xmm9 # 6290 <_sk_callback_sse41+0x1203> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -16678,16 +16709,16 @@ _sk_bicubic_n1y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,198,18,0,0 ; addps 0x12c6(%rip),%xmm1 # 6250 <_sk_callback_sse41+0x11fd> - DB 68,15,40,13,206,18,0,0 ; movaps 0x12ce(%rip),%xmm9 # 6260 <_sk_callback_sse41+0x120d> + DB 15,88,13,220,18,0,0 ; addps 0x12dc(%rip),%xmm1 # 62a0 <_sk_callback_sse41+0x1213> + DB 68,15,40,13,228,18,0,0 ; movaps 0x12e4(%rip),%xmm9 # 62b0 <_sk_callback_sse41+0x1223> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,210,18,0,0 ; movaps 0x12d2(%rip),%xmm8 # 6270 <_sk_callback_sse41+0x121d> + DB 68,15,40,5,232,18,0,0 ; movaps 0x12e8(%rip),%xmm8 # 62c0 <_sk_callback_sse41+0x1233> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,214,18,0,0 ; addps 0x12d6(%rip),%xmm8 # 6280 <_sk_callback_sse41+0x122d> + DB 68,15,88,5,236,18,0,0 ; addps 0x12ec(%rip),%xmm8 # 62d0 <_sk_callback_sse41+0x1243> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,218,18,0,0 ; addps 0x12da(%rip),%xmm8 # 6290 <_sk_callback_sse41+0x123d> + DB 68,15,88,5,240,18,0,0 ; addps 0x12f0(%rip),%xmm8 # 62e0 <_sk_callback_sse41+0x1253> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,222,18,0,0 ; addps 0x12de(%rip),%xmm8 # 62a0 <_sk_callback_sse41+0x124d> + DB 68,15,88,5,244,18,0,0 ; addps 0x12f4(%rip),%xmm8 # 62f0 <_sk_callback_sse41+0x1263> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16695,17 +16726,17 @@ _sk_bicubic_n1y_sse41 LABEL PROC PUBLIC _sk_bicubic_p1y_sse41 _sk_bicubic_p1y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,216,18,0,0 ; movaps 0x12d8(%rip),%xmm8 # 62b0 <_sk_callback_sse41+0x125d> + DB 68,15,40,5,238,18,0,0 ; movaps 0x12ee(%rip),%xmm8 # 6300 <_sk_callback_sse41+0x1273> DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,72,96 ; movups 0x60(%rax),%xmm9 DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 68,15,40,21,211,18,0,0 ; movaps 0x12d3(%rip),%xmm10 # 62c0 <_sk_callback_sse41+0x126d> + DB 68,15,40,21,233,18,0,0 ; movaps 0x12e9(%rip),%xmm10 # 6310 <_sk_callback_sse41+0x1283> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,215,18,0,0 ; addps 0x12d7(%rip),%xmm10 # 62d0 <_sk_callback_sse41+0x127d> + DB 68,15,88,21,237,18,0,0 ; addps 0x12ed(%rip),%xmm10 # 6320 <_sk_callback_sse41+0x1293> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,211,18,0,0 ; addps 0x12d3(%rip),%xmm10 # 62e0 <_sk_callback_sse41+0x128d> + DB 68,15,88,21,233,18,0,0 ; addps 0x12e9(%rip),%xmm10 # 6330 <_sk_callback_sse41+0x12a3> DB 68,15,17,144,160,0,0,0 ; movups %xmm10,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16715,11 +16746,11 @@ _sk_bicubic_p3y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,197,18,0,0 ; addps 0x12c5(%rip),%xmm1 # 62f0 <_sk_callback_sse41+0x129d> + DB 15,88,13,219,18,0,0 ; addps 0x12db(%rip),%xmm1 # 6340 <_sk_callback_sse41+0x12b3> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,197,18,0,0 ; mulps 0x12c5(%rip),%xmm8 # 6300 <_sk_callback_sse41+0x12ad> - DB 68,15,88,5,205,18,0,0 ; addps 0x12cd(%rip),%xmm8 # 6310 <_sk_callback_sse41+0x12bd> + DB 68,15,89,5,219,18,0,0 ; mulps 0x12db(%rip),%xmm8 # 6350 <_sk_callback_sse41+0x12c3> + DB 68,15,88,5,227,18,0,0 ; addps 0x12e3(%rip),%xmm8 # 6360 <_sk_callback_sse41+0x12d3> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -16729,12 +16760,14 @@ PUBLIC _sk_callback_sse41 _sk_callback_sse41 LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,131,236,32 ; sub $0x20,%rsp + DB 72,131,236,40 ; sub $0x28,%rsp DB 68,15,40,197 ; movaps %xmm5,%xmm8 DB 68,15,40,204 ; movaps %xmm4,%xmm9 - DB 73,137,207 ; mov %rcx,%r15 - DB 73,137,214 ; mov %rdx,%r14 + DB 77,137,196 ; mov %r8,%r12 + DB 73,137,206 ; mov %rcx,%r14 + DB 73,137,215 ; mov %rdx,%r15 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,137,195 ; mov %rax,%rbx DB 15,40,224 ; movaps %xmm0,%xmm4 @@ -16753,9 +16786,9 @@ _sk_callback_sse41 LABEL PROC DB 15,17,107,24 ; movups %xmm5,0x18(%rbx) DB 102,15,17,91,40 ; movupd %xmm3,0x28(%rbx) DB 15,17,83,56 ; movups %xmm2,0x38(%rbx) - DB 77,133,255 ; test %r15,%r15 + DB 77,133,228 ; test %r12,%r12 DB 186,4,0,0,0 ; mov $0x4,%edx - DB 65,15,69,215 ; cmovne %r15d,%edx + DB 65,15,69,212 ; cmovne %r12d,%edx DB 72,137,217 ; mov %rbx,%rcx DB 255,19 ; callq *(%rbx) DB 72,139,131,136,0,0,0 ; mov 0x88(%rbx),%rax @@ -16776,12 +16809,14 @@ _sk_callback_sse41 LABEL PROC DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 15,18,220 ; movhlps %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,249 ; mov %r15,%rcx + DB 76,137,250 ; mov %r15,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 77,137,224 ; mov %r12,%r8 DB 65,15,40,225 ; movaps %xmm9,%xmm4 DB 65,15,40,232 ; movaps %xmm8,%xmm5 - DB 72,131,196,32 ; add $0x20,%rsp + DB 72,131,196,40 ; add $0x28,%rsp DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -16930,11 +16965,11 @@ ALIGN 16 DB 128,191,0,0,128,191,0 ; cmpb $0x0,-0x40800000(%rdi) DB 0,224 ; add %ah,%al DB 64,0,0 ; add %al,(%rax) - DB 224,64 ; loopne 52e8 <.literal16+0x1d8> + DB 224,64 ; loopne 5338 <.literal16+0x1d8> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 52ec <.literal16+0x1dc> + DB 224,64 ; loopne 533c <.literal16+0x1dc> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 52f0 <.literal16+0x1e0> + DB 224,64 ; loopne 5340 <.literal16+0x1e0> DB 154 ; (bad) DB 153 ; cltd DB 153 ; cltd @@ -16954,13 +16989,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5311 <.literal16+0x201> + DB 71,225,61 ; rex.RXB loope 5361 <.literal16+0x201> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5315 <.literal16+0x205> + DB 71,225,61 ; rex.RXB loope 5365 <.literal16+0x205> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5319 <.literal16+0x209> + DB 71,225,61 ; rex.RXB loope 5369 <.literal16+0x209> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 531d <.literal16+0x20d> + DB 71,225,61 ; rex.RXB loope 536d <.literal16+0x20d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -16985,13 +17020,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5351 <.literal16+0x241> + DB 71,225,61 ; rex.RXB loope 53a1 <.literal16+0x241> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5355 <.literal16+0x245> + DB 71,225,61 ; rex.RXB loope 53a5 <.literal16+0x245> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5359 <.literal16+0x249> + DB 71,225,61 ; rex.RXB loope 53a9 <.literal16+0x249> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 535d <.literal16+0x24d> + DB 71,225,61 ; rex.RXB loope 53ad <.literal16+0x24d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17016,13 +17051,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5391 <.literal16+0x281> + DB 71,225,61 ; rex.RXB loope 53e1 <.literal16+0x281> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5395 <.literal16+0x285> + DB 71,225,61 ; rex.RXB loope 53e5 <.literal16+0x285> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5399 <.literal16+0x289> + DB 71,225,61 ; rex.RXB loope 53e9 <.literal16+0x289> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 539d <.literal16+0x28d> + DB 71,225,61 ; rex.RXB loope 53ed <.literal16+0x28d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17047,13 +17082,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53d1 <.literal16+0x2c1> + DB 71,225,61 ; rex.RXB loope 5421 <.literal16+0x2c1> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53d5 <.literal16+0x2c5> + DB 71,225,61 ; rex.RXB loope 5425 <.literal16+0x2c5> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53d9 <.literal16+0x2c9> + DB 71,225,61 ; rex.RXB loope 5429 <.literal16+0x2c9> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53dd <.literal16+0x2cd> + DB 71,225,61 ; rex.RXB loope 542d <.literal16+0x2cd> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17070,10 +17105,10 @@ ALIGN 16 DB 0,1 ; add %al,(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0053c8 <_sk_callback_sse41+0xa000375> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005418 <_sk_callback_sse41+0xa00038b> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30053d0 <_sk_callback_sse41+0x300037d> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005420 <_sk_callback_sse41+0x3000393> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -17092,11 +17127,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) DB 67,0,0 ; rex.XB add %al,(%r8) - DB 127,67 ; jg 543b <.literal16+0x32b> + DB 127,67 ; jg 548b <.literal16+0x32b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 543f <.literal16+0x32f> + DB 127,67 ; jg 548f <.literal16+0x32f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5443 <.literal16+0x333> + DB 127,67 ; jg 5493 <.literal16+0x333> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17331,13 +17366,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5619 <.literal16+0x509> + DB 224,7 ; loopne 5669 <.literal16+0x509> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 561d <.literal16+0x50d> + DB 224,7 ; loopne 566d <.literal16+0x50d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5621 <.literal16+0x511> + DB 224,7 ; loopne 5671 <.literal16+0x511> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5625 <.literal16+0x515> + DB 224,7 ; loopne 5675 <.literal16+0x515> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -17371,10 +17406,10 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005668 <_sk_callback_sse41+0xa000615> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0056b8 <_sk_callback_sse41+0xa00062b> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005670 <_sk_callback_sse41+0x300061d> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30056c0 <_sk_callback_sse41+0x3000633> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -17429,11 +17464,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 573b <.literal16+0x62b> + DB 127,67 ; jg 578b <.literal16+0x62b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 573f <.literal16+0x62f> + DB 127,67 ; jg 578f <.literal16+0x62f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5743 <.literal16+0x633> + DB 127,67 ; jg 5793 <.literal16+0x633> DB 129,128,128,59,129,128,128,59,129,128; addl $0x80813b80,-0x7f7ec480(%rax) DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,129,128,128,59 ; addb $0x3b,-0x7f7f7ec5(%rax) @@ -17448,16 +17483,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5734 <.literal16+0x624> + DB 127,0 ; jg 5784 <.literal16+0x624> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5738 <.literal16+0x628> + DB 127,0 ; jg 5788 <.literal16+0x628> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 573c <.literal16+0x62c> + DB 127,0 ; jg 578c <.literal16+0x62c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5740 <.literal16+0x630> + DB 127,0 ; jg 5790 <.literal16+0x630> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17466,7 +17501,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 57c5 <.literal16+0x6b5> + DB 119,115 ; ja 5815 <.literal16+0x6b5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17477,7 +17512,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5729 <.literal16+0x619> + DB 117,191 ; jne 5779 <.literal16+0x619> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17489,7 +17524,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a3976a <_sk_callback_sse41+0xffffffffe9a34717> + DB 233,220,63,163,233 ; jmpq ffffffffe9a397ba <_sk_callback_sse41+0xffffffffe9a3472d> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -17544,16 +17579,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5804 <.literal16+0x6f4> + DB 127,0 ; jg 5854 <.literal16+0x6f4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5808 <.literal16+0x6f8> + DB 127,0 ; jg 5858 <.literal16+0x6f8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 580c <.literal16+0x6fc> + DB 127,0 ; jg 585c <.literal16+0x6fc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5810 <.literal16+0x700> + DB 127,0 ; jg 5860 <.literal16+0x700> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17562,7 +17597,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5895 <.literal16+0x785> + DB 119,115 ; ja 58e5 <.literal16+0x785> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17573,7 +17608,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 57f9 <.literal16+0x6e9> + DB 117,191 ; jne 5849 <.literal16+0x6e9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17585,7 +17620,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a3983a <_sk_callback_sse41+0xffffffffe9a347e7> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3988a <_sk_callback_sse41+0xffffffffe9a347fd> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -17640,16 +17675,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 58d4 <.literal16+0x7c4> + DB 127,0 ; jg 5924 <.literal16+0x7c4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 58d8 <.literal16+0x7c8> + DB 127,0 ; jg 5928 <.literal16+0x7c8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 58dc <.literal16+0x7cc> + DB 127,0 ; jg 592c <.literal16+0x7cc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 58e0 <.literal16+0x7d0> + DB 127,0 ; jg 5930 <.literal16+0x7d0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17658,7 +17693,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5965 <.literal16+0x855> + DB 119,115 ; ja 59b5 <.literal16+0x855> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17669,7 +17704,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 58c9 <.literal16+0x7b9> + DB 117,191 ; jne 5919 <.literal16+0x7b9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17681,7 +17716,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a3990a <_sk_callback_sse41+0xffffffffe9a348b7> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3995a <_sk_callback_sse41+0xffffffffe9a348cd> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -17736,16 +17771,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 59a4 <.literal16+0x894> + DB 127,0 ; jg 59f4 <.literal16+0x894> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 59a8 <.literal16+0x898> + DB 127,0 ; jg 59f8 <.literal16+0x898> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 59ac <.literal16+0x89c> + DB 127,0 ; jg 59fc <.literal16+0x89c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 59b0 <.literal16+0x8a0> + DB 127,0 ; jg 5a00 <.literal16+0x8a0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17754,7 +17789,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5a35 <.literal16+0x925> + DB 119,115 ; ja 5a85 <.literal16+0x925> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17765,7 +17800,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5999 <.literal16+0x889> + DB 117,191 ; jne 59e9 <.literal16+0x889> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17777,7 +17812,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a399da <_sk_callback_sse41+0xffffffffe9a34987> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39a2a <_sk_callback_sse41+0xffffffffe9a3499d> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -17828,13 +17863,13 @@ ALIGN 16 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 - DB 127,67 ; jg 5ab7 <.literal16+0x9a7> + DB 127,67 ; jg 5b07 <.literal16+0x9a7> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5abb <.literal16+0x9ab> + DB 127,67 ; jg 5b0b <.literal16+0x9ab> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5abf <.literal16+0x9af> + DB 127,67 ; jg 5b0f <.literal16+0x9af> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5ac3 <.literal16+0x9b3> + DB 127,67 ; jg 5b13 <.literal16+0x9b3> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -17881,16 +17916,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5b43 <.literal16+0xa33> + DB 118,63 ; jbe 5b93 <.literal16+0xa33> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5b47 <.literal16+0xa37> + DB 118,63 ; jbe 5b97 <.literal16+0xa37> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5b4b <.literal16+0xa3b> + DB 118,63 ; jbe 5b9b <.literal16+0xa3b> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5b4f <.literal16+0xa3f> + DB 118,63 ; jbe 5b9f <.literal16+0xa3f> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) @@ -17910,11 +17945,11 @@ ALIGN 16 DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5b9b <.literal16+0xa8b> + DB 127,67 ; jg 5beb <.literal16+0xa8b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5b9f <.literal16+0xa8f> + DB 127,67 ; jg 5bef <.literal16+0xa8f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5ba3 <.literal16+0xa93> + DB 127,67 ; jg 5bf3 <.literal16+0xa93> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -17951,7 +17986,7 @@ ALIGN 16 DB 5,255,255,255,9 ; add $0x9ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005be0 <_sk_callback_sse41+0x3000b8d> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005c30 <_sk_callback_sse41+0x3000ba3> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -17980,13 +18015,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5c19 <.literal16+0xb09> + DB 224,7 ; loopne 5c69 <.literal16+0xb09> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c1d <.literal16+0xb0d> + DB 224,7 ; loopne 5c6d <.literal16+0xb0d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c21 <.literal16+0xb11> + DB 224,7 ; loopne 5c71 <.literal16+0xb11> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c25 <.literal16+0xb15> + DB 224,7 ; loopne 5c75 <.literal16+0xb15> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -18032,13 +18067,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5c89 <.literal16+0xb79> + DB 224,7 ; loopne 5cd9 <.literal16+0xb79> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c8d <.literal16+0xb7d> + DB 224,7 ; loopne 5cdd <.literal16+0xb7d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c91 <.literal16+0xb81> + DB 224,7 ; loopne 5ce1 <.literal16+0xb81> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c95 <.literal16+0xb85> + DB 224,7 ; loopne 5ce5 <.literal16+0xb85> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -18076,13 +18111,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 5d26 <.literal16+0xc16> + DB 124,66 ; jl 5d76 <.literal16+0xc16> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 5d2a <.literal16+0xc1a> + DB 124,66 ; jl 5d7a <.literal16+0xc1a> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 5d2e <.literal16+0xc1e> + DB 124,66 ; jl 5d7e <.literal16+0xc1e> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 5d32 <.literal16+0xc22> + DB 124,66 ; jl 5d82 <.literal16+0xc22> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -18172,13 +18207,13 @@ ALIGN 16 DB 136,136,61,137,136,136 ; mov %cl,-0x777776c3(%rax) DB 61,137,136,136,61 ; cmp $0x3d888889,%eax DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5e35 <.literal16+0xd25> + DB 112,65 ; jo 5e85 <.literal16+0xd25> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5e39 <.literal16+0xd29> + DB 112,65 ; jo 5e89 <.literal16+0xd29> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5e3d <.literal16+0xd2d> + DB 112,65 ; jo 5e8d <.literal16+0xd2d> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5e41 <.literal16+0xd31> + DB 112,65 ; jo 5e91 <.literal16+0xd31> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -18193,7 +18228,7 @@ ALIGN 16 DB 5,255,255,255,9 ; add $0x9ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005e30 <_sk_callback_sse41+0x3000ddd> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005e80 <_sk_callback_sse41+0x3000df3> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -18220,7 +18255,7 @@ ALIGN 16 DB 5,255,255,255,9 ; add $0x9ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005e70 <_sk_callback_sse41+0x3000e1d> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005ec0 <_sk_callback_sse41+0x3000e33> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -18235,11 +18270,11 @@ ALIGN 16 DB 255,0 ; incl (%rax) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5ecb <.literal16+0xdbb> + DB 127,67 ; jg 5f1b <.literal16+0xdbb> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5ecf <.literal16+0xdbf> + DB 127,67 ; jg 5f1f <.literal16+0xdbf> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5ed3 <.literal16+0xdc3> + DB 127,67 ; jg 5f23 <.literal16+0xdc3> DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) @@ -18315,13 +18350,13 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 5f9b <.literal16+0xe8b> + DB 127,71 ; jg 5feb <.literal16+0xe8b> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 5f9f <.literal16+0xe8f> + DB 127,71 ; jg 5fef <.literal16+0xe8f> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 5fa3 <.literal16+0xe93> + DB 127,71 ; jg 5ff3 <.literal16+0xe93> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 5fa7 <.literal16+0xe97> + DB 127,71 ; jg 5ff7 <.literal16+0xe97> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -18367,10 +18402,10 @@ ALIGN 16 DB 61,152,221,147,61 ; cmp $0x3d93dd98,%eax DB 152 ; cwtl DB 221,147,61,45,16,17 ; fstl 0x11102d3d(%rbx) - DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc070da <_sk_callback_sse41+0x2dc02087> + DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc0712a <_sk_callback_sse41+0x2dc0209d> DB 17,192 ; adc %eax,%eax DB 45,16,17,192,18 ; sub $0x12c01110,%eax - DB 120,57 ; js 600c <.literal16+0xefc> + DB 120,57 ; js 605c <.literal16+0xefc> DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil @@ -18492,11 +18527,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6142 <.literal16+0x1032> + DB 62,114,28 ; jb,pt 6192 <.literal16+0x1032> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6146 <.literal16+0x1036> + DB 62,114,28 ; jb,pt 6196 <.literal16+0x1036> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 614a <.literal16+0x103a> + DB 62,114,28 ; jb,pt 619a <.literal16+0x103a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18540,7 +18575,7 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63efd5 <_sk_callback_sse41+0x3d639f82> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f025 <_sk_callback_sse41+0x3d639f98> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -18566,7 +18601,7 @@ ALIGN 16 DB 0,192 ; add %al,%al DB 63 ; (bad) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f015 <_sk_callback_sse41+0x3d639fc2> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f065 <_sk_callback_sse41+0x3d639fd8> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -18575,13 +18610,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 620e <.literal16+0x10fe> + DB 114,28 ; jb 625e <.literal16+0x10fe> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6212 <.literal16+0x1102> + DB 62,114,28 ; jb,pt 6262 <.literal16+0x1102> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6216 <.literal16+0x1106> + DB 62,114,28 ; jb,pt 6266 <.literal16+0x1106> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 621a <.literal16+0x110a> + DB 62,114,28 ; jb,pt 626a <.literal16+0x110a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18602,11 +18637,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6252 <.literal16+0x1142> + DB 62,114,28 ; jb,pt 62a2 <.literal16+0x1142> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6256 <.literal16+0x1146> + DB 62,114,28 ; jb,pt 62a6 <.literal16+0x1146> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 625a <.literal16+0x114a> + DB 62,114,28 ; jb,pt 62aa <.literal16+0x114a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18650,7 +18685,7 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f0e5 <_sk_callback_sse41+0x3d63a092> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f135 <_sk_callback_sse41+0x3d63a0a8> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -18676,7 +18711,7 @@ ALIGN 16 DB 0,192 ; add %al,%al DB 63 ; (bad) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f125 <_sk_callback_sse41+0x3d63a0d2> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f175 <_sk_callback_sse41+0x3d63a0e8> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -18685,13 +18720,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 631e <.literal16+0x120e> + DB 114,28 ; jb 636e <.literal16+0x120e> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6322 <_sk_callback_sse41+0x12cf> + DB 62,114,28 ; jb,pt 6372 <_sk_callback_sse41+0x12e5> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6326 <_sk_callback_sse41+0x12d3> + DB 62,114,28 ; jb,pt 6376 <_sk_callback_sse41+0x12e9> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 632a <_sk_callback_sse41+0x12d7> + DB 62,114,28 ; jb,pt 637a <_sk_callback_sse41+0x12ed> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18711,8 +18746,9 @@ _sk_start_pipeline_sse2 LABEL PROC DB 65,84 ; push %r12 DB 86 ; push %rsi DB 87 ; push %rdi + DB 85 ; push %rbp DB 83 ; push %rbx - DB 72,129,236,160,0,0,0 ; sub $0xa0,%rsp + DB 72,129,236,168,0,0,0 ; sub $0xa8,%rsp DB 68,15,41,188,36,144,0,0,0 ; movaps %xmm15,0x90(%rsp) DB 68,15,41,180,36,128,0,0,0 ; movaps %xmm14,0x80(%rsp) DB 68,15,41,108,36,112 ; movaps %xmm13,0x70(%rsp) @@ -18723,19 +18759,20 @@ _sk_start_pipeline_sse2 LABEL PROC DB 68,15,41,68,36,32 ; movaps %xmm8,0x20(%rsp) DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp) DB 15,41,52,36 ; movaps %xmm6,(%rsp) - DB 77,137,205 ; mov %r9,%r13 - DB 77,137,198 ; mov %r8,%r14 - DB 72,137,203 ; mov %rcx,%rbx - DB 72,137,214 ; mov %rdx,%rsi - DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,199 ; mov %rax,%r15 - DB 73,137,244 ; mov %rsi,%r12 - DB 72,141,67,4 ; lea 0x4(%rbx),%rax - DB 76,57,232 ; cmp %r13,%rax - DB 118,5 ; jbe 73 <_sk_start_pipeline_sse2+0x73> - DB 72,137,223 ; mov %rbx,%rdi - DB 235,57 ; jmp ac <_sk_start_pipeline_sse2+0xac> - DB 185,0,0,0,0 ; mov $0x0,%ecx + DB 76,137,195 ; mov %r8,%rbx + DB 73,137,214 ; mov %rdx,%r14 + DB 72,137,205 ; mov %rcx,%rbp + DB 76,139,188,36,16,1,0,0 ; mov 0x110(%rsp),%r15 + DB 76,137,206 ; mov %r9,%rsi + DB 72,173 ; lods %ds:(%rsi),%rax + DB 73,137,196 ; mov %rax,%r12 + DB 73,137,245 ; mov %rsi,%r13 + DB 72,141,69,4 ; lea 0x4(%rbp),%rax + DB 72,57,216 ; cmp %rbx,%rax + DB 118,5 ; jbe 7c <_sk_start_pipeline_sse2+0x7c> + DB 72,137,234 ; mov %rbp,%rdx + DB 235,61 ; jmp b9 <_sk_start_pipeline_sse2+0xb9> + DB 65,184,0,0,0,0 ; mov $0x0,%r8d DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -18744,17 +18781,18 @@ _sk_start_pipeline_sse2 LABEL PROC DB 15,87,237 ; xorps %xmm5,%xmm5 DB 15,87,246 ; xorps %xmm6,%xmm6 DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 72,137,223 ; mov %rbx,%rdi - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 65,255,215 ; callq *%r15 - DB 72,141,123,4 ; lea 0x4(%rbx),%rdi - DB 72,131,195,8 ; add $0x8,%rbx - DB 76,57,235 ; cmp %r13,%rbx - DB 72,137,251 ; mov %rdi,%rbx - DB 118,199 ; jbe 73 <_sk_start_pipeline_sse2+0x73> - DB 73,41,253 ; sub %rdi,%r13 - DB 116,36 ; je d5 <_sk_start_pipeline_sse2+0xd5> + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 72,137,234 ; mov %rbp,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 65,255,212 ; callq *%r12 + DB 72,141,85,4 ; lea 0x4(%rbp),%rdx + DB 72,131,197,8 ; add $0x8,%rbp + DB 72,57,221 ; cmp %rbx,%rbp + DB 72,137,213 ; mov %rdx,%rbp + DB 118,195 ; jbe 7c <_sk_start_pipeline_sse2+0x7c> + DB 72,41,211 ; sub %rdx,%rbx + DB 116,39 ; je e5 <_sk_start_pipeline_sse2+0xe5> DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -18763,10 +18801,11 @@ _sk_start_pipeline_sse2 LABEL PROC DB 15,87,237 ; xorps %xmm5,%xmm5 DB 15,87,246 ; xorps %xmm6,%xmm6 DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 76,137,230 ; mov %r12,%rsi - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,233 ; mov %r13,%rcx - DB 65,255,215 ; callq *%r15 + DB 76,137,255 ; mov %r15,%rdi + DB 76,137,238 ; mov %r13,%rsi + DB 76,137,241 ; mov %r14,%rcx + DB 73,137,216 ; mov %rbx,%r8 + DB 65,255,212 ; callq *%r12 DB 15,40,52,36 ; movaps (%rsp),%xmm6 DB 15,40,124,36,16 ; movaps 0x10(%rsp),%xmm7 DB 68,15,40,68,36,32 ; movaps 0x20(%rsp),%xmm8 @@ -18777,8 +18816,9 @@ _sk_start_pipeline_sse2 LABEL PROC DB 68,15,40,108,36,112 ; movaps 0x70(%rsp),%xmm13 DB 68,15,40,180,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm14 DB 68,15,40,188,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm15 - DB 72,129,196,160,0,0,0 ; add $0xa0,%rsp + DB 72,129,196,168,0,0,0 ; add $0xa8,%rsp DB 91 ; pop %rbx + DB 93 ; pop %rbp DB 95 ; pop %rdi DB 94 ; pop %rsi DB 65,92 ; pop %r12 @@ -18793,20 +18833,19 @@ _sk_just_return_sse2 LABEL PROC PUBLIC _sk_seed_shader_sse2 _sk_seed_shader_sse2 LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,110,199 ; movd %edi,%xmm0 + DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 - DB 15,40,21,51,85,0,0 ; movaps 0x5533(%rip),%xmm2 # 5670 <_sk_callback_sse2+0xca> + DB 15,40,21,84,85,0,0 ; movaps 0x5554(%rip),%xmm2 # 56a0 <_sk_callback_sse2+0xc6> DB 15,88,202 ; addps %xmm2,%xmm1 - DB 15,16,2 ; movups (%rdx),%xmm0 + DB 15,16,7 ; movups (%rdi),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 - DB 102,15,110,8 ; movd (%rax),%xmm1 + DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,34,85,0,0 ; movaps 0x5522(%rip),%xmm2 # 5680 <_sk_callback_sse2+0xda> + DB 15,40,21,67,85,0,0 ; movaps 0x5543(%rip),%xmm2 # 56b0 <_sk_callback_sse2+0xd6> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -18817,22 +18856,21 @@ _sk_seed_shader_sse2 LABEL PROC PUBLIC _sk_dither_sse2 _sk_dither_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,68,15,110,199 ; movd %edi,%xmm8 + DB 102,68,15,110,194 ; movd %edx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 - DB 243,68,15,111,74,32 ; movdqu 0x20(%rdx),%xmm9 + DB 243,68,15,111,79,32 ; movdqu 0x20(%rdi),%xmm9 DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 - DB 76,139,0 ; mov (%rax),%r8 - DB 102,69,15,110,0 ; movd (%r8),%xmm8 + DB 102,68,15,110,193 ; movd %ecx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 - DB 102,68,15,111,21,237,84,0,0 ; movdqa 0x54ed(%rip),%xmm10 # 5690 <_sk_callback_sse2+0xea> + DB 102,68,15,111,21,17,85,0,0 ; movdqa 0x5511(%rip),%xmm10 # 56c0 <_sk_callback_sse2+0xe6> DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,65,15,114,243,5 ; pslld $0x5,%xmm11 DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 102,65,15,114,242,4 ; pslld $0x4,%xmm10 - DB 102,68,15,111,37,217,84,0,0 ; movdqa 0x54d9(%rip),%xmm12 # 56a0 <_sk_callback_sse2+0xfa> - DB 102,68,15,111,45,224,84,0,0 ; movdqa 0x54e0(%rip),%xmm13 # 56b0 <_sk_callback_sse2+0x10a> + DB 102,68,15,111,37,253,84,0,0 ; movdqa 0x54fd(%rip),%xmm12 # 56d0 <_sk_callback_sse2+0xf6> + DB 102,68,15,111,45,4,85,0,0 ; movdqa 0x5504(%rip),%xmm13 # 56e0 <_sk_callback_sse2+0x106> DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 DB 102,69,15,219,245 ; pand %xmm13,%xmm14 DB 102,65,15,114,246,2 ; pslld $0x2,%xmm14 @@ -18844,13 +18882,13 @@ _sk_dither_sse2 LABEL PROC DB 102,65,15,114,209,2 ; psrld $0x2,%xmm9 DB 102,69,15,235,234 ; por %xmm10,%xmm13 DB 102,69,15,235,233 ; por %xmm9,%xmm13 - DB 102,69,15,235,235 ; por %xmm11,%xmm13 - DB 102,69,15,235,198 ; por %xmm14,%xmm8 - DB 102,69,15,235,197 ; por %xmm13,%xmm8 - DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,155,84,0,0 ; mulps 0x549b(%rip),%xmm8 # 56c0 <_sk_callback_sse2+0x11a> - DB 68,15,88,5,163,84,0,0 ; addps 0x54a3(%rip),%xmm8 # 56d0 <_sk_callback_sse2+0x12a> - DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 + DB 102,69,15,235,243 ; por %xmm11,%xmm14 + DB 102,69,15,235,245 ; por %xmm13,%xmm14 + DB 102,69,15,235,240 ; por %xmm8,%xmm14 + DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8 + DB 68,15,89,5,191,84,0,0 ; mulps 0x54bf(%rip),%xmm8 # 56f0 <_sk_callback_sse2+0x116> + DB 68,15,88,5,199,84,0,0 ; addps 0x54c7(%rip),%xmm8 # 5700 <_sk_callback_sse2+0x126> + DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,88,194 ; addps %xmm10,%xmm0 @@ -18916,7 +18954,7 @@ _sk_clear_sse2 LABEL PROC PUBLIC _sk_srcatop_sse2 _sk_srcatop_sse2 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,252,83,0,0 ; movaps 0x53fc(%rip),%xmm8 # 56e0 <_sk_callback_sse2+0x13a> + DB 68,15,40,5,33,84,0,0 ; movaps 0x5421(%rip),%xmm8 # 5710 <_sk_callback_sse2+0x136> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -18939,7 +18977,7 @@ PUBLIC _sk_dstatop_sse2 _sk_dstatop_sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 DB 68,15,89,196 ; mulps %xmm4,%xmm8 - DB 68,15,40,13,191,83,0,0 ; movaps 0x53bf(%rip),%xmm9 # 56f0 <_sk_callback_sse2+0x14a> + DB 68,15,40,13,228,83,0,0 ; movaps 0x53e4(%rip),%xmm9 # 5720 <_sk_callback_sse2+0x146> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -18980,7 +19018,7 @@ _sk_dstin_sse2 LABEL PROC PUBLIC _sk_srcout_sse2 _sk_srcout_sse2 LABEL PROC - DB 68,15,40,5,99,83,0,0 ; movaps 0x5363(%rip),%xmm8 # 5700 <_sk_callback_sse2+0x15a> + DB 68,15,40,5,136,83,0,0 ; movaps 0x5388(%rip),%xmm8 # 5730 <_sk_callback_sse2+0x156> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -18991,7 +19029,7 @@ _sk_srcout_sse2 LABEL PROC PUBLIC _sk_dstout_sse2 _sk_dstout_sse2 LABEL PROC - DB 68,15,40,5,83,83,0,0 ; movaps 0x5353(%rip),%xmm8 # 5710 <_sk_callback_sse2+0x16a> + DB 68,15,40,5,120,83,0,0 ; movaps 0x5378(%rip),%xmm8 # 5740 <_sk_callback_sse2+0x166> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -19006,7 +19044,7 @@ _sk_dstout_sse2 LABEL PROC PUBLIC _sk_srcover_sse2 _sk_srcover_sse2 LABEL PROC - DB 68,15,40,5,54,83,0,0 ; movaps 0x5336(%rip),%xmm8 # 5720 <_sk_callback_sse2+0x17a> + DB 68,15,40,5,91,83,0,0 ; movaps 0x535b(%rip),%xmm8 # 5750 <_sk_callback_sse2+0x176> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -19024,7 +19062,7 @@ _sk_srcover_sse2 LABEL PROC PUBLIC _sk_dstover_sse2 _sk_dstover_sse2 LABEL PROC - DB 68,15,40,5,10,83,0,0 ; movaps 0x530a(%rip),%xmm8 # 5730 <_sk_callback_sse2+0x18a> + DB 68,15,40,5,47,83,0,0 ; movaps 0x532f(%rip),%xmm8 # 5760 <_sk_callback_sse2+0x186> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -19048,7 +19086,7 @@ _sk_modulate_sse2 LABEL PROC PUBLIC _sk_multiply_sse2 _sk_multiply_sse2 LABEL PROC - DB 68,15,40,5,222,82,0,0 ; movaps 0x52de(%rip),%xmm8 # 5740 <_sk_callback_sse2+0x19a> + DB 68,15,40,5,3,83,0,0 ; movaps 0x5303(%rip),%xmm8 # 5770 <_sk_callback_sse2+0x196> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -19117,7 +19155,7 @@ _sk_screen_sse2 LABEL PROC PUBLIC _sk_xor__sse2 _sk_xor__sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,19,82,0,0 ; movaps 0x5213(%rip),%xmm3 # 5750 <_sk_callback_sse2+0x1aa> + DB 15,40,29,56,82,0,0 ; movaps 0x5238(%rip),%xmm3 # 5780 <_sk_callback_sse2+0x1a6> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -19163,7 +19201,7 @@ _sk_darken_sse2 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,95,209 ; maxps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,126,81,0,0 ; movaps 0x517e(%rip),%xmm2 # 5760 <_sk_callback_sse2+0x1ba> + DB 15,40,21,163,81,0,0 ; movaps 0x51a3(%rip),%xmm2 # 5790 <_sk_callback_sse2+0x1b6> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19195,7 +19233,7 @@ _sk_lighten_sse2 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,35,81,0,0 ; movaps 0x5123(%rip),%xmm2 # 5770 <_sk_callback_sse2+0x1ca> + DB 15,40,21,72,81,0,0 ; movaps 0x5148(%rip),%xmm2 # 57a0 <_sk_callback_sse2+0x1c6> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19230,7 +19268,7 @@ _sk_difference_sse2 LABEL PROC DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,189,80,0,0 ; movaps 0x50bd(%rip),%xmm2 # 5780 <_sk_callback_sse2+0x1da> + DB 15,40,21,226,80,0,0 ; movaps 0x50e2(%rip),%xmm2 # 57b0 <_sk_callback_sse2+0x1d6> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19256,7 +19294,7 @@ _sk_exclusion_sse2 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,125,80,0,0 ; movaps 0x507d(%rip),%xmm2 # 5790 <_sk_callback_sse2+0x1ea> + DB 15,40,21,162,80,0,0 ; movaps 0x50a2(%rip),%xmm2 # 57c0 <_sk_callback_sse2+0x1e6> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19267,7 +19305,7 @@ _sk_exclusion_sse2 LABEL PROC PUBLIC _sk_colorburn_sse2 _sk_colorburn_sse2 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,112,80,0,0 ; movaps 0x5070(%rip),%xmm10 # 57a0 <_sk_callback_sse2+0x1fa> + DB 68,15,40,21,149,80,0,0 ; movaps 0x5095(%rip),%xmm10 # 57d0 <_sk_callback_sse2+0x1f6> DB 69,15,40,202 ; movaps %xmm10,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,217 ; movaps %xmm9,%xmm11 @@ -19359,7 +19397,7 @@ _sk_colorburn_sse2 LABEL PROC PUBLIC _sk_colordodge_sse2 _sk_colordodge_sse2 LABEL PROC DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 68,15,40,21,38,79,0,0 ; movaps 0x4f26(%rip),%xmm10 # 57b0 <_sk_callback_sse2+0x20a> + DB 68,15,40,21,75,79,0,0 ; movaps 0x4f4b(%rip),%xmm10 # 57e0 <_sk_callback_sse2+0x206> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -19452,7 +19490,7 @@ _sk_hardlight_sse2 LABEL PROC DB 15,41,52,36 ; movaps %xmm6,(%rsp) DB 15,40,245 ; movaps %xmm5,%xmm6 DB 15,40,236 ; movaps %xmm4,%xmm5 - DB 68,15,40,29,216,77,0,0 ; movaps 0x4dd8(%rip),%xmm11 # 57c0 <_sk_callback_sse2+0x21a> + DB 68,15,40,29,253,77,0,0 ; movaps 0x4dfd(%rip),%xmm11 # 57f0 <_sk_callback_sse2+0x216> DB 69,15,40,211 ; movaps %xmm11,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,194 ; movaps %xmm10,%xmm8 @@ -19539,7 +19577,7 @@ PUBLIC _sk_overlay_sse2 _sk_overlay_sse2 LABEL PROC DB 68,15,40,193 ; movaps %xmm1,%xmm8 DB 68,15,40,232 ; movaps %xmm0,%xmm13 - DB 68,15,40,13,163,76,0,0 ; movaps 0x4ca3(%rip),%xmm9 # 57d0 <_sk_callback_sse2+0x22a> + DB 68,15,40,13,200,76,0,0 ; movaps 0x4cc8(%rip),%xmm9 # 5800 <_sk_callback_sse2+0x226> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -19629,7 +19667,7 @@ _sk_softlight_sse2 LABEL PROC DB 68,15,40,213 ; movaps %xmm5,%xmm10 DB 68,15,94,215 ; divps %xmm7,%xmm10 DB 69,15,84,212 ; andps %xmm12,%xmm10 - DB 68,15,40,13,93,75,0,0 ; movaps 0x4b5d(%rip),%xmm9 # 57e0 <_sk_callback_sse2+0x23a> + DB 68,15,40,13,130,75,0,0 ; movaps 0x4b82(%rip),%xmm9 # 5810 <_sk_callback_sse2+0x236> DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 69,15,92,250 ; subps %xmm10,%xmm15 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -19642,10 +19680,10 @@ _sk_softlight_sse2 LABEL PROC DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 DB 65,15,88,194 ; addps %xmm10,%xmm0 - DB 68,15,40,53,55,75,0,0 ; movaps 0x4b37(%rip),%xmm14 # 57f0 <_sk_callback_sse2+0x24a> + DB 68,15,40,53,92,75,0,0 ; movaps 0x4b5c(%rip),%xmm14 # 5820 <_sk_callback_sse2+0x246> DB 69,15,88,222 ; addps %xmm14,%xmm11 DB 68,15,89,216 ; mulps %xmm0,%xmm11 - DB 68,15,40,21,55,75,0,0 ; movaps 0x4b37(%rip),%xmm10 # 5800 <_sk_callback_sse2+0x25a> + DB 68,15,40,21,92,75,0,0 ; movaps 0x4b5c(%rip),%xmm10 # 5830 <_sk_callback_sse2+0x256> DB 69,15,89,234 ; mulps %xmm10,%xmm13 DB 69,15,88,235 ; addps %xmm11,%xmm13 DB 15,88,228 ; addps %xmm4,%xmm4 @@ -19790,7 +19828,7 @@ _sk_hue_sse2 LABEL PROC DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 68,15,89,211 ; mulps %xmm3,%xmm10 - DB 68,15,40,5,115,73,0,0 ; movaps 0x4973(%rip),%xmm8 # 5840 <_sk_callback_sse2+0x29a> + DB 68,15,40,5,152,73,0,0 ; movaps 0x4998(%rip),%xmm8 # 5870 <_sk_callback_sse2+0x296> DB 69,15,40,216 ; movaps %xmm8,%xmm11 DB 15,40,207 ; movaps %xmm7,%xmm1 DB 68,15,92,217 ; subps %xmm1,%xmm11 @@ -19838,12 +19876,12 @@ _sk_hue_sse2 LABEL PROC DB 69,15,84,206 ; andps %xmm14,%xmm9 DB 69,15,84,214 ; andps %xmm14,%xmm10 DB 65,15,84,214 ; andps %xmm14,%xmm2 - DB 68,15,40,61,128,72,0,0 ; movaps 0x4880(%rip),%xmm15 # 5810 <_sk_callback_sse2+0x26a> + DB 68,15,40,61,165,72,0,0 ; movaps 0x48a5(%rip),%xmm15 # 5840 <_sk_callback_sse2+0x266> DB 65,15,89,231 ; mulps %xmm15,%xmm4 - DB 15,40,5,133,72,0,0 ; movaps 0x4885(%rip),%xmm0 # 5820 <_sk_callback_sse2+0x27a> + DB 15,40,5,170,72,0,0 ; movaps 0x48aa(%rip),%xmm0 # 5850 <_sk_callback_sse2+0x276> DB 15,89,240 ; mulps %xmm0,%xmm6 DB 15,88,244 ; addps %xmm4,%xmm6 - DB 68,15,40,53,135,72,0,0 ; movaps 0x4887(%rip),%xmm14 # 5830 <_sk_callback_sse2+0x28a> + DB 68,15,40,53,172,72,0,0 ; movaps 0x48ac(%rip),%xmm14 # 5860 <_sk_callback_sse2+0x286> DB 68,15,40,239 ; movaps %xmm7,%xmm13 DB 69,15,89,238 ; mulps %xmm14,%xmm13 DB 68,15,88,238 ; addps %xmm6,%xmm13 @@ -20021,14 +20059,14 @@ _sk_saturation_sse2 LABEL PROC DB 68,15,84,211 ; andps %xmm3,%xmm10 DB 68,15,84,203 ; andps %xmm3,%xmm9 DB 15,84,195 ; andps %xmm3,%xmm0 - DB 68,15,40,5,20,70,0,0 ; movaps 0x4614(%rip),%xmm8 # 5850 <_sk_callback_sse2+0x2aa> + DB 68,15,40,5,57,70,0,0 ; movaps 0x4639(%rip),%xmm8 # 5880 <_sk_callback_sse2+0x2a6> DB 15,40,214 ; movaps %xmm6,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 - DB 15,40,13,22,70,0,0 ; movaps 0x4616(%rip),%xmm1 # 5860 <_sk_callback_sse2+0x2ba> + DB 15,40,13,59,70,0,0 ; movaps 0x463b(%rip),%xmm1 # 5890 <_sk_callback_sse2+0x2b6> DB 15,40,221 ; movaps %xmm5,%xmm3 DB 15,89,217 ; mulps %xmm1,%xmm3 DB 15,88,218 ; addps %xmm2,%xmm3 - DB 68,15,40,37,21,70,0,0 ; movaps 0x4615(%rip),%xmm12 # 5870 <_sk_callback_sse2+0x2ca> + DB 68,15,40,37,58,70,0,0 ; movaps 0x463a(%rip),%xmm12 # 58a0 <_sk_callback_sse2+0x2c6> DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 68,15,88,235 ; addps %xmm3,%xmm13 DB 65,15,40,210 ; movaps %xmm10,%xmm2 @@ -20073,7 +20111,7 @@ _sk_saturation_sse2 LABEL PROC DB 15,40,223 ; movaps %xmm7,%xmm3 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,221 ; mulps %xmm5,%xmm3 - DB 68,15,40,5,122,69,0,0 ; movaps 0x457a(%rip),%xmm8 # 5880 <_sk_callback_sse2+0x2da> + DB 68,15,40,5,159,69,0,0 ; movaps 0x459f(%rip),%xmm8 # 58b0 <_sk_callback_sse2+0x2d6> DB 65,15,40,224 ; movaps %xmm8,%xmm4 DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 15,88,253 ; addps %xmm5,%xmm7 @@ -20174,14 +20212,14 @@ _sk_color_sse2 LABEL PROC DB 68,15,40,213 ; movaps %xmm5,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,40,208 ; movaps %xmm8,%xmm2 - DB 68,15,40,45,18,68,0,0 ; movaps 0x4412(%rip),%xmm13 # 5890 <_sk_callback_sse2+0x2ea> + DB 68,15,40,45,55,68,0,0 ; movaps 0x4437(%rip),%xmm13 # 58c0 <_sk_callback_sse2+0x2e6> DB 68,15,40,198 ; movaps %xmm6,%xmm8 DB 69,15,89,197 ; mulps %xmm13,%xmm8 - DB 68,15,40,53,18,68,0,0 ; movaps 0x4412(%rip),%xmm14 # 58a0 <_sk_callback_sse2+0x2fa> + DB 68,15,40,53,55,68,0,0 ; movaps 0x4437(%rip),%xmm14 # 58d0 <_sk_callback_sse2+0x2f6> DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,89,198 ; mulps %xmm14,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,29,14,68,0,0 ; movaps 0x440e(%rip),%xmm11 # 58b0 <_sk_callback_sse2+0x30a> + DB 68,15,40,29,51,68,0,0 ; movaps 0x4433(%rip),%xmm11 # 58e0 <_sk_callback_sse2+0x306> DB 69,15,89,227 ; mulps %xmm11,%xmm12 DB 68,15,88,224 ; addps %xmm0,%xmm12 DB 65,15,40,193 ; movaps %xmm9,%xmm0 @@ -20189,7 +20227,7 @@ _sk_color_sse2 LABEL PROC DB 69,15,40,250 ; movaps %xmm10,%xmm15 DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 68,15,88,248 ; addps %xmm0,%xmm15 - DB 68,15,40,5,250,67,0,0 ; movaps 0x43fa(%rip),%xmm8 # 58c0 <_sk_callback_sse2+0x31a> + DB 68,15,40,5,31,68,0,0 ; movaps 0x441f(%rip),%xmm8 # 58f0 <_sk_callback_sse2+0x316> DB 65,15,40,224 ; movaps %xmm8,%xmm4 DB 15,92,226 ; subps %xmm2,%xmm4 DB 15,89,252 ; mulps %xmm4,%xmm7 @@ -20325,15 +20363,15 @@ _sk_luminosity_sse2 LABEL PROC DB 68,15,40,205 ; movaps %xmm5,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 DB 15,89,222 ; mulps %xmm6,%xmm3 - DB 68,15,40,37,12,66,0,0 ; movaps 0x420c(%rip),%xmm12 # 58d0 <_sk_callback_sse2+0x32a> + DB 68,15,40,37,49,66,0,0 ; movaps 0x4231(%rip),%xmm12 # 5900 <_sk_callback_sse2+0x326> DB 68,15,40,199 ; movaps %xmm7,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,12,66,0,0 ; movaps 0x420c(%rip),%xmm13 # 58e0 <_sk_callback_sse2+0x33a> + DB 68,15,40,45,49,66,0,0 ; movaps 0x4231(%rip),%xmm13 # 5910 <_sk_callback_sse2+0x336> DB 68,15,40,241 ; movaps %xmm1,%xmm14 DB 69,15,89,245 ; mulps %xmm13,%xmm14 DB 69,15,88,240 ; addps %xmm8,%xmm14 - DB 68,15,40,29,8,66,0,0 ; movaps 0x4208(%rip),%xmm11 # 58f0 <_sk_callback_sse2+0x34a> - DB 68,15,40,5,16,66,0,0 ; movaps 0x4210(%rip),%xmm8 # 5900 <_sk_callback_sse2+0x35a> + DB 68,15,40,29,45,66,0,0 ; movaps 0x422d(%rip),%xmm11 # 5920 <_sk_callback_sse2+0x346> + DB 68,15,40,5,53,66,0,0 ; movaps 0x4235(%rip),%xmm8 # 5930 <_sk_callback_sse2+0x356> DB 69,15,40,248 ; movaps %xmm8,%xmm15 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 68,15,92,248 ; subps %xmm0,%xmm15 @@ -20467,11 +20505,11 @@ PUBLIC _sk_srcover_rgba_8888_sse2 _sk_srcover_rgba_8888_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,227,0,0,0 ; jne 19ca <_sk_srcover_rgba_8888_sse2+0xf1> - DB 243,68,15,111,4,184 ; movdqu (%rax,%rdi,4),%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 102,15,111,53,24,64,0,0 ; movdqa 0x4018(%rip),%xmm6 # 5910 <_sk_callback_sse2+0x36a> + DB 77,133,192 ; test %r8,%r8 + DB 15,133,227,0,0,0 ; jne 19d5 <_sk_srcover_rgba_8888_sse2+0xf1> + DB 243,68,15,111,4,144 ; movdqu (%rax,%rdx,4),%xmm8 + DB 77,133,192 ; test %r8,%r8 + DB 102,15,111,53,61,64,0,0 ; movdqa 0x403d(%rip),%xmm6 # 5940 <_sk_callback_sse2+0x366> DB 102,65,15,111,224 ; movdqa %xmm8,%xmm4 DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 @@ -20485,9 +20523,9 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 15,91,247 ; cvtdq2ps %xmm7,%xmm6 DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8 DB 65,15,91,248 ; cvtdq2ps %xmm8,%xmm7 - DB 68,15,40,5,232,63,0,0 ; movaps 0x3fe8(%rip),%xmm8 # 5920 <_sk_callback_sse2+0x37a> + DB 68,15,40,5,13,64,0,0 ; movaps 0x400d(%rip),%xmm8 # 5950 <_sk_callback_sse2+0x376> DB 68,15,92,195 ; subps %xmm3,%xmm8 - DB 68,15,40,37,236,63,0,0 ; movaps 0x3fec(%rip),%xmm12 # 5930 <_sk_callback_sse2+0x38a> + DB 68,15,40,37,17,64,0,0 ; movaps 0x4011(%rip),%xmm12 # 5960 <_sk_callback_sse2+0x386> DB 65,15,89,196 ; mulps %xmm12,%xmm0 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -20513,46 +20551,46 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 102,15,114,240,24 ; pslld $0x18,%xmm0 DB 102,15,235,194 ; por %xmm2,%xmm0 DB 102,15,235,193 ; por %xmm1,%xmm0 - DB 117,106 ; jne 1a1b <_sk_srcover_rgba_8888_sse2+0x142> - DB 243,15,127,4,184 ; movdqu %xmm0,(%rax,%rdi,4) + DB 117,106 ; jne 1a26 <_sk_srcover_rgba_8888_sse2+0x142> + DB 243,15,127,4,144 ; movdqu %xmm0,(%rax,%rdx,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 65,15,40,202 ; movaps %xmm10,%xmm1 DB 65,15,40,211 ; movaps %xmm11,%xmm2 DB 65,15,40,216 ; movaps %xmm8,%xmm3 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,48 ; je 1a0c <_sk_srcover_rgba_8888_sse2+0x133> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,22 ; je 19f8 <_sk_srcover_rgba_8888_sse2+0x11f> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 15,133,1,255,255,255 ; jne 18ed <_sk_srcover_rgba_8888_sse2+0x14> - DB 102,15,110,100,184,8 ; movd 0x8(%rax,%rdi,4),%xmm4 + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,48 ; je 1a17 <_sk_srcover_rgba_8888_sse2+0x133> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,22 ; je 1a03 <_sk_srcover_rgba_8888_sse2+0x11f> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 15,133,1,255,255,255 ; jne 18f8 <_sk_srcover_rgba_8888_sse2+0x14> + DB 102,15,110,100,144,8 ; movd 0x8(%rax,%rdx,4),%xmm4 DB 102,68,15,112,196,69 ; pshufd $0x45,%xmm4,%xmm8 - DB 243,15,16,100,184,4 ; movss 0x4(%rax,%rdi,4),%xmm4 + DB 243,15,16,100,144,4 ; movss 0x4(%rax,%rdx,4),%xmm4 DB 65,15,198,224,0 ; shufps $0x0,%xmm8,%xmm4 DB 65,15,198,224,226 ; shufps $0xe2,%xmm8,%xmm4 DB 68,15,40,196 ; movaps %xmm4,%xmm8 - DB 243,15,16,36,184 ; movss (%rax,%rdi,4),%xmm4 + DB 243,15,16,36,144 ; movss (%rax,%rdx,4),%xmm4 DB 243,68,15,16,196 ; movss %xmm4,%xmm8 - DB 233,210,254,255,255 ; jmpq 18ed <_sk_srcover_rgba_8888_sse2+0x14> - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,34 ; je 1a4a <_sk_srcover_rgba_8888_sse2+0x171> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,17 ; je 1a3f <_sk_srcover_rgba_8888_sse2+0x166> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,130 ; jne 19b6 <_sk_srcover_rgba_8888_sse2+0xdd> + DB 233,210,254,255,255 ; jmpq 18f8 <_sk_srcover_rgba_8888_sse2+0x14> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,34 ; je 1a55 <_sk_srcover_rgba_8888_sse2+0x171> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,17 ; je 1a4a <_sk_srcover_rgba_8888_sse2+0x166> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,130 ; jne 19c1 <_sk_srcover_rgba_8888_sse2+0xdd> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,15,126,76,184,8 ; movd %xmm1,0x8(%rax,%rdi,4) + DB 102,15,126,76,144,8 ; movd %xmm1,0x8(%rax,%rdx,4) DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,15,126,76,184,4 ; movd %xmm1,0x4(%rax,%rdi,4) - DB 102,15,126,4,184 ; movd %xmm0,(%rax,%rdi,4) - DB 233,98,255,255,255 ; jmpq 19b6 <_sk_srcover_rgba_8888_sse2+0xdd> + DB 102,15,126,76,144,4 ; movd %xmm1,0x4(%rax,%rdx,4) + DB 102,15,126,4,144 ; movd %xmm0,(%rax,%rdx,4) + DB 233,98,255,255,255 ; jmpq 19c1 <_sk_srcover_rgba_8888_sse2+0xdd> PUBLIC _sk_clamp_0_sse2 _sk_clamp_0_sse2 LABEL PROC @@ -20566,7 +20604,7 @@ _sk_clamp_0_sse2 LABEL PROC PUBLIC _sk_clamp_1_sse2 _sk_clamp_1_sse2 LABEL PROC - DB 68,15,40,5,204,62,0,0 ; movaps 0x3ecc(%rip),%xmm8 # 5940 <_sk_callback_sse2+0x39a> + DB 68,15,40,5,241,62,0,0 ; movaps 0x3ef1(%rip),%xmm8 # 5970 <_sk_callback_sse2+0x396> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -20576,7 +20614,7 @@ _sk_clamp_1_sse2 LABEL PROC PUBLIC _sk_clamp_a_sse2 _sk_clamp_a_sse2 LABEL PROC - DB 15,93,29,193,62,0,0 ; minps 0x3ec1(%rip),%xmm3 # 5950 <_sk_callback_sse2+0x3aa> + DB 15,93,29,230,62,0,0 ; minps 0x3ee6(%rip),%xmm3 # 5980 <_sk_callback_sse2+0x3a6> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -20649,7 +20687,7 @@ _sk_premul_sse2 LABEL PROC PUBLIC _sk_unpremul_sse2 _sk_unpremul_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,44,62,0,0 ; movaps 0x3e2c(%rip),%xmm9 # 5960 <_sk_callback_sse2+0x3ba> + DB 68,15,40,13,81,62,0,0 ; movaps 0x3e51(%rip),%xmm9 # 5990 <_sk_callback_sse2+0x3b6> DB 68,15,94,203 ; divps %xmm3,%xmm9 DB 68,15,194,195,4 ; cmpneqps %xmm3,%xmm8 DB 69,15,84,193 ; andps %xmm9,%xmm8 @@ -20661,20 +20699,20 @@ _sk_unpremul_sse2 LABEL PROC PUBLIC _sk_from_srgb_sse2 _sk_from_srgb_sse2 LABEL PROC - DB 68,15,40,5,23,62,0,0 ; movaps 0x3e17(%rip),%xmm8 # 5970 <_sk_callback_sse2+0x3ca> + DB 68,15,40,5,60,62,0,0 ; movaps 0x3e3c(%rip),%xmm8 # 59a0 <_sk_callback_sse2+0x3c6> DB 68,15,40,232 ; movaps %xmm0,%xmm13 DB 69,15,89,232 ; mulps %xmm8,%xmm13 DB 68,15,40,216 ; movaps %xmm0,%xmm11 DB 69,15,89,219 ; mulps %xmm11,%xmm11 - DB 68,15,40,13,15,62,0,0 ; movaps 0x3e0f(%rip),%xmm9 # 5980 <_sk_callback_sse2+0x3da> + DB 68,15,40,13,52,62,0,0 ; movaps 0x3e34(%rip),%xmm9 # 59b0 <_sk_callback_sse2+0x3d6> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,241 ; mulps %xmm9,%xmm14 - DB 68,15,40,21,15,62,0,0 ; movaps 0x3e0f(%rip),%xmm10 # 5990 <_sk_callback_sse2+0x3ea> + DB 68,15,40,21,52,62,0,0 ; movaps 0x3e34(%rip),%xmm10 # 59c0 <_sk_callback_sse2+0x3e6> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 69,15,89,243 ; mulps %xmm11,%xmm14 - DB 68,15,40,29,15,62,0,0 ; movaps 0x3e0f(%rip),%xmm11 # 59a0 <_sk_callback_sse2+0x3fa> + DB 68,15,40,29,52,62,0,0 ; movaps 0x3e34(%rip),%xmm11 # 59d0 <_sk_callback_sse2+0x3f6> DB 69,15,88,243 ; addps %xmm11,%xmm14 - DB 68,15,40,37,19,62,0,0 ; movaps 0x3e13(%rip),%xmm12 # 59b0 <_sk_callback_sse2+0x40a> + DB 68,15,40,37,56,62,0,0 ; movaps 0x3e38(%rip),%xmm12 # 59e0 <_sk_callback_sse2+0x406> DB 65,15,194,196,1 ; cmpltps %xmm12,%xmm0 DB 68,15,84,232 ; andps %xmm0,%xmm13 DB 65,15,85,198 ; andnps %xmm14,%xmm0 @@ -20709,22 +20747,22 @@ _sk_from_srgb_sse2 LABEL PROC PUBLIC _sk_to_srgb_sse2 _sk_to_srgb_sse2 LABEL PROC DB 68,15,82,232 ; rsqrtps %xmm0,%xmm13 - DB 68,15,40,5,160,61,0,0 ; movaps 0x3da0(%rip),%xmm8 # 59c0 <_sk_callback_sse2+0x41a> + DB 68,15,40,5,197,61,0,0 ; movaps 0x3dc5(%rip),%xmm8 # 59f0 <_sk_callback_sse2+0x416> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,240 ; mulps %xmm8,%xmm14 - DB 68,15,40,13,160,61,0,0 ; movaps 0x3da0(%rip),%xmm9 # 59d0 <_sk_callback_sse2+0x42a> + DB 68,15,40,13,197,61,0,0 ; movaps 0x3dc5(%rip),%xmm9 # 5a00 <_sk_callback_sse2+0x426> DB 69,15,40,253 ; movaps %xmm13,%xmm15 DB 69,15,89,249 ; mulps %xmm9,%xmm15 - DB 68,15,40,21,160,61,0,0 ; movaps 0x3da0(%rip),%xmm10 # 59e0 <_sk_callback_sse2+0x43a> + DB 68,15,40,21,197,61,0,0 ; movaps 0x3dc5(%rip),%xmm10 # 5a10 <_sk_callback_sse2+0x436> DB 69,15,88,250 ; addps %xmm10,%xmm15 DB 69,15,89,253 ; mulps %xmm13,%xmm15 - DB 68,15,40,29,160,61,0,0 ; movaps 0x3da0(%rip),%xmm11 # 59f0 <_sk_callback_sse2+0x44a> + DB 68,15,40,29,197,61,0,0 ; movaps 0x3dc5(%rip),%xmm11 # 5a20 <_sk_callback_sse2+0x446> DB 69,15,88,251 ; addps %xmm11,%xmm15 - DB 68,15,40,37,164,61,0,0 ; movaps 0x3da4(%rip),%xmm12 # 5a00 <_sk_callback_sse2+0x45a> + DB 68,15,40,37,201,61,0,0 ; movaps 0x3dc9(%rip),%xmm12 # 5a30 <_sk_callback_sse2+0x456> DB 69,15,88,236 ; addps %xmm12,%xmm13 DB 69,15,83,237 ; rcpps %xmm13,%xmm13 DB 69,15,89,239 ; mulps %xmm15,%xmm13 - DB 68,15,40,61,160,61,0,0 ; movaps 0x3da0(%rip),%xmm15 # 5a10 <_sk_callback_sse2+0x46a> + DB 68,15,40,61,197,61,0,0 ; movaps 0x3dc5(%rip),%xmm15 # 5a40 <_sk_callback_sse2+0x466> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,197 ; andnps %xmm13,%xmm0 @@ -20772,7 +20810,7 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 68,15,93,218 ; minps %xmm2,%xmm11 DB 65,15,40,202 ; movaps %xmm10,%xmm1 DB 65,15,92,203 ; subps %xmm11,%xmm1 - DB 68,15,40,45,249,60,0,0 ; movaps 0x3cf9(%rip),%xmm13 # 5a20 <_sk_callback_sse2+0x47a> + DB 68,15,40,45,30,61,0,0 ; movaps 0x3d1e(%rip),%xmm13 # 5a50 <_sk_callback_sse2+0x476> DB 68,15,94,233 ; divps %xmm1,%xmm13 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,194,192,0 ; cmpeqps %xmm8,%xmm0 @@ -20781,30 +20819,30 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,40,241 ; movaps %xmm9,%xmm14 DB 68,15,194,242,1 ; cmpltps %xmm2,%xmm14 - DB 68,15,84,53,223,60,0,0 ; andps 0x3cdf(%rip),%xmm14 # 5a30 <_sk_callback_sse2+0x48a> + DB 68,15,84,53,4,61,0,0 ; andps 0x3d04(%rip),%xmm14 # 5a60 <_sk_callback_sse2+0x486> DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,40,250 ; movaps %xmm10,%xmm15 DB 69,15,194,249,0 ; cmpeqps %xmm9,%xmm15 DB 65,15,92,208 ; subps %xmm8,%xmm2 DB 65,15,89,213 ; mulps %xmm13,%xmm2 - DB 68,15,40,37,210,60,0,0 ; movaps 0x3cd2(%rip),%xmm12 # 5a40 <_sk_callback_sse2+0x49a> + DB 68,15,40,37,247,60,0,0 ; movaps 0x3cf7(%rip),%xmm12 # 5a70 <_sk_callback_sse2+0x496> DB 65,15,88,212 ; addps %xmm12,%xmm2 DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 69,15,89,197 ; mulps %xmm13,%xmm8 - DB 68,15,88,5,206,60,0,0 ; addps 0x3cce(%rip),%xmm8 # 5a50 <_sk_callback_sse2+0x4aa> + DB 68,15,88,5,243,60,0,0 ; addps 0x3cf3(%rip),%xmm8 # 5a80 <_sk_callback_sse2+0x4a6> DB 65,15,84,215 ; andps %xmm15,%xmm2 DB 69,15,85,248 ; andnps %xmm8,%xmm15 DB 68,15,86,250 ; orps %xmm2,%xmm15 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,199 ; andnps %xmm15,%xmm0 DB 65,15,86,198 ; orps %xmm14,%xmm0 - DB 15,89,5,191,60,0,0 ; mulps 0x3cbf(%rip),%xmm0 # 5a60 <_sk_callback_sse2+0x4ba> + DB 15,89,5,228,60,0,0 ; mulps 0x3ce4(%rip),%xmm0 # 5a90 <_sk_callback_sse2+0x4b6> DB 69,15,40,194 ; movaps %xmm10,%xmm8 DB 69,15,194,195,4 ; cmpneqps %xmm11,%xmm8 DB 65,15,84,192 ; andps %xmm8,%xmm0 DB 69,15,92,226 ; subps %xmm10,%xmm12 DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 68,15,40,13,178,60,0,0 ; movaps 0x3cb2(%rip),%xmm9 # 5a70 <_sk_callback_sse2+0x4ca> + DB 68,15,40,13,215,60,0,0 ; movaps 0x3cd7(%rip),%xmm9 # 5aa0 <_sk_callback_sse2+0x4c6> DB 65,15,40,210 ; movaps %xmm10,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 68,15,194,202,1 ; cmpltps %xmm2,%xmm9 @@ -20827,7 +20865,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,41,92,36,32 ; movaps %xmm3,0x20(%rsp) DB 68,15,40,218 ; movaps %xmm2,%xmm11 DB 15,40,240 ; movaps %xmm0,%xmm6 - DB 68,15,40,13,109,60,0,0 ; movaps 0x3c6d(%rip),%xmm9 # 5a80 <_sk_callback_sse2+0x4da> + DB 68,15,40,13,146,60,0,0 ; movaps 0x3c92(%rip),%xmm9 # 5ab0 <_sk_callback_sse2+0x4d6> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 69,15,194,211,2 ; cmpleps %xmm11,%xmm10 DB 15,40,193 ; movaps %xmm1,%xmm0 @@ -20844,28 +20882,28 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 69,15,88,211 ; addps %xmm11,%xmm10 DB 69,15,88,219 ; addps %xmm11,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 15,40,5,55,60,0,0 ; movaps 0x3c37(%rip),%xmm0 # 5a90 <_sk_callback_sse2+0x4ea> + DB 15,40,5,92,60,0,0 ; movaps 0x3c5c(%rip),%xmm0 # 5ac0 <_sk_callback_sse2+0x4e6> DB 15,88,198 ; addps %xmm6,%xmm0 DB 243,15,91,200 ; cvttps2dq %xmm0,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,40,216 ; movaps %xmm0,%xmm3 DB 15,194,217,1 ; cmpltps %xmm1,%xmm3 - DB 15,84,29,47,60,0,0 ; andps 0x3c2f(%rip),%xmm3 # 5aa0 <_sk_callback_sse2+0x4fa> + DB 15,84,29,84,60,0,0 ; andps 0x3c54(%rip),%xmm3 # 5ad0 <_sk_callback_sse2+0x4f6> DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,92,193 ; subps %xmm1,%xmm0 - DB 68,15,40,45,49,60,0,0 ; movaps 0x3c31(%rip),%xmm13 # 5ab0 <_sk_callback_sse2+0x50a> + DB 68,15,40,45,86,60,0,0 ; movaps 0x3c56(%rip),%xmm13 # 5ae0 <_sk_callback_sse2+0x506> DB 69,15,40,197 ; movaps %xmm13,%xmm8 DB 68,15,194,192,2 ; cmpleps %xmm0,%xmm8 DB 69,15,40,242 ; movaps %xmm10,%xmm14 DB 69,15,92,243 ; subps %xmm11,%xmm14 DB 65,15,40,217 ; movaps %xmm9,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 - DB 15,40,21,65,60,0,0 ; movaps 0x3c41(%rip),%xmm2 # 5ae0 <_sk_callback_sse2+0x53a> + DB 15,40,21,102,60,0,0 ; movaps 0x3c66(%rip),%xmm2 # 5b10 <_sk_callback_sse2+0x536> DB 68,15,40,250 ; movaps %xmm2,%xmm15 DB 68,15,194,248,2 ; cmpleps %xmm0,%xmm15 - DB 15,40,13,17,60,0,0 ; movaps 0x3c11(%rip),%xmm1 # 5ac0 <_sk_callback_sse2+0x51a> + DB 15,40,13,54,60,0,0 ; movaps 0x3c36(%rip),%xmm1 # 5af0 <_sk_callback_sse2+0x516> DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,40,45,23,60,0,0 ; movaps 0x3c17(%rip),%xmm5 # 5ad0 <_sk_callback_sse2+0x52a> + DB 15,40,45,60,60,0,0 ; movaps 0x3c3c(%rip),%xmm5 # 5b00 <_sk_callback_sse2+0x526> DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,92,224 ; subps %xmm0,%xmm4 DB 65,15,89,230 ; mulps %xmm14,%xmm4 @@ -20888,7 +20926,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,222 ; movaps %xmm6,%xmm3 DB 15,194,216,1 ; cmpltps %xmm0,%xmm3 - DB 15,84,29,140,59,0,0 ; andps 0x3b8c(%rip),%xmm3 # 5aa0 <_sk_callback_sse2+0x4fa> + DB 15,84,29,177,59,0,0 ; andps 0x3bb1(%rip),%xmm3 # 5ad0 <_sk_callback_sse2+0x4f6> DB 15,92,195 ; subps %xmm3,%xmm0 DB 68,15,40,230 ; movaps %xmm6,%xmm12 DB 68,15,92,224 ; subps %xmm0,%xmm12 @@ -20918,12 +20956,12 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,40,60,36 ; movaps (%rsp),%xmm7 DB 15,40,231 ; movaps %xmm7,%xmm4 DB 15,85,227 ; andnps %xmm3,%xmm4 - DB 15,88,53,101,59,0,0 ; addps 0x3b65(%rip),%xmm6 # 5af0 <_sk_callback_sse2+0x54a> + DB 15,88,53,138,59,0,0 ; addps 0x3b8a(%rip),%xmm6 # 5b20 <_sk_callback_sse2+0x546> DB 243,15,91,198 ; cvttps2dq %xmm6,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,222 ; movaps %xmm6,%xmm3 DB 15,194,216,1 ; cmpltps %xmm0,%xmm3 - DB 15,84,29,0,59,0,0 ; andps 0x3b00(%rip),%xmm3 # 5aa0 <_sk_callback_sse2+0x4fa> + DB 15,84,29,37,59,0,0 ; andps 0x3b25(%rip),%xmm3 # 5ad0 <_sk_callback_sse2+0x4f6> DB 15,92,195 ; subps %xmm3,%xmm0 DB 15,92,240 ; subps %xmm0,%xmm6 DB 15,89,206 ; mulps %xmm6,%xmm1 @@ -20978,15 +21016,15 @@ _sk_scale_1_float_sse2 LABEL PROC PUBLIC _sk_scale_u8_sse2 _sk_scale_u8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,61 ; jne 2097 <_sk_scale_u8_sse2+0x47> - DB 102,69,15,110,4,57 ; movd (%r9,%rdi,1),%xmm8 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,61 ; jne 20a2 <_sk_scale_u8_sse2+0x47> + DB 102,69,15,110,4,18 ; movd (%r10,%rdx,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,219,5,141,58,0,0 ; pand 0x3a8d(%rip),%xmm8 # 5b00 <_sk_callback_sse2+0x55a> + DB 102,68,15,219,5,178,58,0,0 ; pand 0x3ab2(%rip),%xmm8 # 5b30 <_sk_callback_sse2+0x556> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,145,58,0,0 ; mulps 0x3a91(%rip),%xmm8 # 5b10 <_sk_callback_sse2+0x56a> + DB 68,15,89,5,182,58,0,0 ; mulps 0x3ab6(%rip),%xmm8 # 5b40 <_sk_callback_sse2+0x566> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -20994,27 +21032,27 @@ _sk_scale_u8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,216 ; movaps %xmm8,%xmm3 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,54 ; je 20df <_sk_scale_u8_sse2+0x8f> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,23 ; je 20c6 <_sk_scale_u8_sse2+0x76> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,181 ; jne 206a <_sk_scale_u8_sse2+0x1a> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,54 ; je 20ea <_sk_scale_u8_sse2+0x8f> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,23 ; je 20d1 <_sk_scale_u8_sse2+0x76> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,181 ; jne 2075 <_sk_scale_u8_sse2+0x1a> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,68,15,110,200 ; movd %eax,%xmm9 DB 69,15,198,200,0 ; shufps $0x0,%xmm8,%xmm9 DB 69,15,198,200,226 ; shufps $0xe2,%xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,68,15,110,200 ; movd %eax,%xmm9 DB 243,69,15,16,193 ; movss %xmm9,%xmm8 - DB 233,119,255,255,255 ; jmpq 206a <_sk_scale_u8_sse2+0x1a> + DB 233,119,255,255,255 ; jmpq 2075 <_sk_scale_u8_sse2+0x1a> PUBLIC _sk_lerp_1_float_sse2 _sk_lerp_1_float_sse2 LABEL PROC @@ -21039,15 +21077,15 @@ _sk_lerp_1_float_sse2 LABEL PROC PUBLIC _sk_lerp_u8_sse2 _sk_lerp_u8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,81 ; jne 2186 <_sk_lerp_u8_sse2+0x5b> - DB 102,69,15,110,4,57 ; movd (%r9,%rdi,1),%xmm8 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,81 ; jne 2191 <_sk_lerp_u8_sse2+0x5b> + DB 102,69,15,110,4,18 ; movd (%r10,%rdx,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,219,5,210,57,0,0 ; pand 0x39d2(%rip),%xmm8 # 5b20 <_sk_callback_sse2+0x57a> + DB 102,68,15,219,5,247,57,0,0 ; pand 0x39f7(%rip),%xmm8 # 5b50 <_sk_callback_sse2+0x576> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,214,57,0,0 ; mulps 0x39d6(%rip),%xmm8 # 5b30 <_sk_callback_sse2+0x58a> + DB 68,15,89,5,251,57,0,0 ; mulps 0x39fb(%rip),%xmm8 # 5b60 <_sk_callback_sse2+0x586> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -21062,47 +21100,47 @@ _sk_lerp_u8_sse2 LABEL PROC DB 15,88,223 ; addps %xmm7,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,54 ; je 21ce <_sk_lerp_u8_sse2+0xa3> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,23 ; je 21b5 <_sk_lerp_u8_sse2+0x8a> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,161 ; jne 2145 <_sk_lerp_u8_sse2+0x1a> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,54 ; je 21d9 <_sk_lerp_u8_sse2+0xa3> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,23 ; je 21c0 <_sk_lerp_u8_sse2+0x8a> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,161 ; jne 2150 <_sk_lerp_u8_sse2+0x1a> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,68,15,110,200 ; movd %eax,%xmm9 DB 69,15,198,200,0 ; shufps $0x0,%xmm8,%xmm9 DB 69,15,198,200,226 ; shufps $0xe2,%xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,68,15,110,200 ; movd %eax,%xmm9 DB 243,69,15,16,193 ; movss %xmm9,%xmm8 - DB 233,99,255,255,255 ; jmpq 2145 <_sk_lerp_u8_sse2+0x1a> + DB 233,99,255,255,255 ; jmpq 2150 <_sk_lerp_u8_sse2+0x1a> PUBLIC _sk_lerp_565_sse2 _sk_lerp_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,156,0,0,0 ; jne 228c <_sk_lerp_565_sse2+0xaa> - DB 243,69,15,126,12,121 ; movq (%r9,%rdi,2),%xmm9 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,156,0,0,0 ; jne 2297 <_sk_lerp_565_sse2+0xaa> + DB 243,69,15,126,12,82 ; movq (%r10,%rdx,2),%xmm9 DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 - DB 102,68,15,111,5,60,57,0,0 ; movdqa 0x393c(%rip),%xmm8 # 5b40 <_sk_callback_sse2+0x59a> + DB 102,68,15,111,5,97,57,0,0 ; movdqa 0x3961(%rip),%xmm8 # 5b70 <_sk_callback_sse2+0x596> DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,59,57,0,0 ; mulps 0x393b(%rip),%xmm8 # 5b50 <_sk_callback_sse2+0x5aa> - DB 102,68,15,111,21,66,57,0,0 ; movdqa 0x3942(%rip),%xmm10 # 5b60 <_sk_callback_sse2+0x5ba> + DB 68,15,89,5,96,57,0,0 ; mulps 0x3960(%rip),%xmm8 # 5b80 <_sk_callback_sse2+0x5a6> + DB 102,68,15,111,21,103,57,0,0 ; movdqa 0x3967(%rip),%xmm10 # 5b90 <_sk_callback_sse2+0x5b6> DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 - DB 68,15,89,21,65,57,0,0 ; mulps 0x3941(%rip),%xmm10 # 5b70 <_sk_callback_sse2+0x5ca> - DB 102,68,15,219,13,72,57,0,0 ; pand 0x3948(%rip),%xmm9 # 5b80 <_sk_callback_sse2+0x5da> + DB 68,15,89,21,102,57,0,0 ; mulps 0x3966(%rip),%xmm10 # 5ba0 <_sk_callback_sse2+0x5c6> + DB 102,68,15,219,13,109,57,0,0 ; pand 0x396d(%rip),%xmm9 # 5bb0 <_sk_callback_sse2+0x5d6> DB 69,15,91,201 ; cvtdq2ps %xmm9,%xmm9 - DB 68,15,89,13,76,57,0,0 ; mulps 0x394c(%rip),%xmm9 # 5b90 <_sk_callback_sse2+0x5ea> + DB 68,15,89,13,113,57,0,0 ; mulps 0x3971(%rip),%xmm9 # 5bc0 <_sk_callback_sse2+0x5e6> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -21124,126 +21162,129 @@ _sk_lerp_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,216 ; movaps %xmm8,%xmm3 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,58 ; je 22d8 <_sk_lerp_565_sse2+0xf6> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,27 ; je 22bf <_sk_lerp_565_sse2+0xdd> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 15,133,77,255,255,255 ; jne 21fb <_sk_lerp_565_sse2+0x19> - DB 65,15,183,68,121,4 ; movzwl 0x4(%r9,%rdi,2),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,58 ; je 22e3 <_sk_lerp_565_sse2+0xf6> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,27 ; je 22ca <_sk_lerp_565_sse2+0xdd> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 15,133,77,255,255,255 ; jne 2206 <_sk_lerp_565_sse2+0x19> + DB 65,15,183,68,82,4 ; movzwl 0x4(%r10,%rdx,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,200,69 ; pshufd $0x45,%xmm8,%xmm9 - DB 65,15,183,68,121,2 ; movzwl 0x2(%r9,%rdi,2),%eax + DB 65,15,183,68,82,2 ; movzwl 0x2(%r10,%rdx,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 69,15,198,193,0 ; shufps $0x0,%xmm9,%xmm8 DB 69,15,198,193,226 ; shufps $0xe2,%xmm9,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 - DB 65,15,183,4,121 ; movzwl (%r9,%rdi,2),%eax + DB 65,15,183,4,82 ; movzwl (%r10,%rdx,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 243,69,15,16,200 ; movss %xmm8,%xmm9 - DB 233,15,255,255,255 ; jmpq 21fb <_sk_lerp_565_sse2+0x19> + DB 233,15,255,255,255 ; jmpq 2206 <_sk_lerp_565_sse2+0x19> PUBLIC _sk_load_tables_sse2 _sk_load_tables_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,32,1,0,0 ; jne 241a <_sk_load_tables_sse2+0x12e> - DB 243,69,15,111,12,184 ; movdqu (%r8,%rdi,4),%xmm9 + DB 76,139,8 ; mov (%rax),%r9 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,36,1,0,0 ; jne 2429 <_sk_load_tables_sse2+0x132> + DB 243,69,15,111,12,145 ; movdqu (%r9,%rdx,4),%xmm9 + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx - DB 102,68,15,111,5,148,56,0,0 ; movdqa 0x3894(%rip),%xmm8 # 5ba0 <_sk_callback_sse2+0x5fa> + DB 102,68,15,111,5,183,56,0,0 ; movdqa 0x38b7(%rip),%xmm8 # 5bd0 <_sk_callback_sse2+0x5f6> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,73,15,126,200 ; movq %xmm1,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 102,73,15,126,201 ; movq %xmm1,%r9 + DB 102,73,15,126,194 ; movq %xmm0,%r10 + DB 69,15,182,218 ; movzbl %r10b,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 69,15,182,241 ; movzbl %r9b,%r14d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 69,15,182,216 ; movzbl %r8b,%r11d - DB 73,193,232,30 ; shr $0x1e,%r8 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,12,11 ; movss (%rbx,%r9,1),%xmm1 - DB 243,66,15,16,4,3 ; movss (%rbx,%r8,1),%xmm0 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,66,15,16,12,19 ; movss (%rbx,%r10,1),%xmm1 + DB 243,66,15,16,4,11 ; movss (%rbx,%r9,1),%xmm0 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 + DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 + DB 243,66,15,16,20,179 ; movss (%rbx,%r14,4),%xmm2 DB 15,20,194 ; unpcklps %xmm2,%xmm0 DB 15,20,193 ; unpcklps %xmm1,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 DB 102,65,15,219,200 ; pand %xmm8,%xmm1 DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,73,15,126,208 ; movq %xmm2,%r8 + DB 102,73,15,126,209 ; movq %xmm2,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 68,15,182,203 ; movzbl %bl,%r9d + DB 68,15,182,211 ; movzbl %bl,%r10d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 69,15,182,208 ; movzbl %r8b,%r10d - DB 73,193,232,30 ; shr $0x1e,%r8 - DB 243,65,15,16,20,30 ; movss (%r14,%rbx,1),%xmm2 - DB 243,67,15,16,12,6 ; movss (%r14,%r8,1),%xmm1 + DB 69,15,182,217 ; movzbl %r9b,%r11d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 243,65,15,16,20,31 ; movss (%r15,%rbx,1),%xmm2 + DB 243,67,15,16,12,15 ; movss (%r15,%r9,1),%xmm1 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 243,67,15,16,12,151 ; movss (%r15,%r10,4),%xmm1 + DB 243,67,15,16,28,159 ; movss (%r15,%r11,4),%xmm3 DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 76,139,64,24 ; mov 0x18(%rax),%r8 + DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 102,65,15,111,209 ; movdqa %xmm9,%xmm2 DB 102,15,114,210,16 ; psrld $0x10,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 DB 102,72,15,126,219 ; movq %xmm3,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 68,15,182,200 ; movzbl %al,%r9d + DB 68,15,182,208 ; movzbl %al,%r10d DB 72,193,232,30 ; shr $0x1e,%rax - DB 68,15,182,211 ; movzbl %bl,%r10d + DB 68,15,182,219 ; movzbl %bl,%r11d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 243,69,15,16,4,0 ; movss (%r8,%rax,1),%xmm8 - DB 243,65,15,16,20,24 ; movss (%r8,%rbx,1),%xmm2 + DB 243,69,15,16,4,1 ; movss (%r9,%rax,1),%xmm8 + DB 243,65,15,16,20,25 ; movss (%r9,%rbx,1),%xmm2 DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 - DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 - DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 + DB 243,67,15,16,28,153 ; movss (%r9,%r11,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 DB 102,65,15,114,209,24 ; psrld $0x18,%xmm9 DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3 - DB 15,89,29,157,55,0,0 ; mulps 0x379d(%rip),%xmm3 # 5bb0 <_sk_callback_sse2+0x60a> + DB 15,89,29,192,55,0,0 ; mulps 0x37c0(%rip),%xmm3 # 5be0 <_sk_callback_sse2+0x606> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax - DB 65,137,201 ; mov %ecx,%r9d - DB 65,128,225,3 ; and $0x3,%r9b + DB 69,137,194 ; mov %r8d,%r10d + DB 65,128,226,3 ; and $0x3,%r10b DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 - DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,50 ; je 245e <_sk_load_tables_sse2+0x172> - DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 2449 <_sk_load_tables_sse2+0x15d> - DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,196,254,255,255 ; jne 2300 <_sk_load_tables_sse2+0x14> - DB 102,65,15,110,68,184,8 ; movd 0x8(%r8,%rdi,4),%xmm0 + DB 65,128,250,1 ; cmp $0x1,%r10b + DB 116,50 ; je 246d <_sk_load_tables_sse2+0x176> + DB 65,128,250,2 ; cmp $0x2,%r10b + DB 116,23 ; je 2458 <_sk_load_tables_sse2+0x161> + DB 65,128,250,3 ; cmp $0x3,%r10b + DB 15,133,192,254,255,255 ; jne 230b <_sk_load_tables_sse2+0x14> + DB 102,65,15,110,68,145,8 ; movd 0x8(%r9,%rdx,4),%xmm0 DB 102,68,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm9 - DB 243,65,15,16,68,184,4 ; movss 0x4(%r8,%rdi,4),%xmm0 + DB 243,65,15,16,68,145,4 ; movss 0x4(%r9,%rdx,4),%xmm0 DB 65,15,198,193,0 ; shufps $0x0,%xmm9,%xmm0 DB 65,15,198,193,226 ; shufps $0xe2,%xmm9,%xmm0 DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 243,65,15,16,4,184 ; movss (%r8,%rdi,4),%xmm0 + DB 243,65,15,16,4,145 ; movss (%r9,%rdx,4),%xmm0 DB 243,68,15,16,200 ; movss %xmm0,%xmm9 - DB 233,146,254,255,255 ; jmpq 2300 <_sk_load_tables_sse2+0x14> + DB 233,142,254,255,255 ; jmpq 230b <_sk_load_tables_sse2+0x14> PUBLIC _sk_load_tables_u16_be_sse2 _sk_load_tables_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,97,1,0,0 ; jne 25e5 <_sk_load_tables_u16_be_sse2+0x177> - DB 102,67,15,16,4,72 ; movupd (%r8,%r9,2),%xmm0 - DB 102,67,15,16,76,72,16 ; movupd 0x10(%r8,%r9,2),%xmm1 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,101,1,0,0 ; jne 25f8 <_sk_load_tables_u16_be_sse2+0x17b> + DB 102,67,15,16,4,81 ; movupd (%r9,%r10,2),%xmm0 + DB 102,67,15,16,76,81,16 ; movupd 0x10(%r9,%r10,2),%xmm1 + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 102,68,15,40,200 ; movapd %xmm0,%xmm9 @@ -21252,25 +21293,25 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,68,15,111,21,7,55,0,0 ; movdqa 0x3707(%rip),%xmm10 # 5bc0 <_sk_callback_sse2+0x61a> + DB 102,68,15,111,21,38,55,0,0 ; movdqa 0x3726(%rip),%xmm10 # 5bf0 <_sk_callback_sse2+0x616> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,194 ; pand %xmm10,%xmm0 DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,65,15,97,192 ; punpcklwd %xmm8,%xmm0 DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 - DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 69,15,182,200 ; movzbl %r8b,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,15,182,218 ; movzbl %r10b,%r11d - DB 73,193,234,30 ; shr $0x1e,%r10 + DB 102,73,15,126,217 ; movq %xmm3,%r9 + DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 69,15,182,243 ; movzbl %r11b,%r14d + DB 73,193,235,30 ; shr $0x1e,%r11 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,28,19 ; movss (%rbx,%r10,1),%xmm3 - DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,66,15,16,28,27 ; movss (%rbx,%r11,1),%xmm3 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,66,15,16,4,179 ; movss (%rbx,%r14,4),%xmm0 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,194 ; unpcklps %xmm2,%xmm0 DB 15,20,195 ; unpcklps %xmm3,%xmm0 DB 102,15,115,217,8 ; psrldq $0x8,%xmm1 @@ -21278,33 +21319,33 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 102,65,15,97,200 ; punpcklwd %xmm8,%xmm1 DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d + DB 68,15,182,203 ; movzbl %bl,%r9d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d - DB 73,193,233,30 ; shr $0x1e,%r9 - DB 243,67,15,16,20,14 ; movss (%r14,%r9,1),%xmm2 - DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 + DB 102,73,15,126,202 ; movq %xmm1,%r10 + DB 69,15,182,218 ; movzbl %r10b,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 243,67,15,16,20,23 ; movss (%r15,%r10,1),%xmm2 + DB 243,65,15,16,12,159 ; movss (%r15,%rbx,4),%xmm1 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 - DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 + DB 243,67,15,16,12,159 ; movss (%r15,%r11,4),%xmm1 + DB 243,67,15,16,28,143 ; movss (%r15,%r9,4),%xmm3 DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 102,69,15,97,208 ; punpcklwd %xmm8,%xmm10 DB 102,65,15,112,210,78 ; pshufd $0x4e,%xmm10,%xmm2 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d + DB 68,15,182,203 ; movzbl %bl,%r9d DB 72,193,235,32 ; shr $0x20,%rbx DB 102,76,15,126,208 ; movq %xmm10,%rax - DB 68,15,182,208 ; movzbl %al,%r10d + DB 68,15,182,216 ; movzbl %al,%r11d DB 72,193,232,30 ; shr $0x1e,%rax - DB 243,69,15,16,20,1 ; movss (%r9,%rax,1),%xmm10 - DB 243,65,15,16,20,153 ; movss (%r9,%rbx,4),%xmm2 + DB 243,69,15,16,20,2 ; movss (%r10,%rax,1),%xmm10 + DB 243,65,15,16,20,154 ; movss (%r10,%rbx,4),%xmm2 DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 - DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3 + DB 243,67,15,16,20,154 ; movss (%r10,%r11,4),%xmm2 + DB 243,67,15,16,28,138 ; movss (%r10,%r9,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,210 ; unpcklps %xmm10,%xmm2 DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3 @@ -21314,202 +21355,206 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 102,65,15,235,217 ; por %xmm9,%xmm3 DB 102,65,15,97,216 ; punpcklwd %xmm8,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,242,53,0,0 ; mulps 0x35f2(%rip),%xmm3 # 5bd0 <_sk_callback_sse2+0x62a> + DB 15,89,29,17,54,0,0 ; mulps 0x3611(%rip),%xmm3 # 5c00 <_sk_callback_sse2+0x626> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax - DB 242,67,15,16,4,72 ; movsd (%r8,%r9,2),%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,17 ; jne 2602 <_sk_load_tables_u16_be_sse2+0x194> + DB 242,67,15,16,4,81 ; movsd (%r9,%r10,2),%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,17 ; jne 2615 <_sk_load_tables_u16_be_sse2+0x198> DB 102,15,87,201 ; xorpd %xmm1,%xmm1 DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 233,143,254,255,255 ; jmpq 2491 <_sk_load_tables_u16_be_sse2+0x23> - DB 102,67,15,22,68,72,8 ; movhpd 0x8(%r8,%r9,2),%xmm0 + DB 233,139,254,255,255 ; jmpq 24a0 <_sk_load_tables_u16_be_sse2+0x23> + DB 102,67,15,22,68,81,8 ; movhpd 0x8(%r9,%r10,2),%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 15,130,122,254,255,255 ; jb 2491 <_sk_load_tables_u16_be_sse2+0x23> - DB 242,67,15,16,76,72,16 ; movsd 0x10(%r8,%r9,2),%xmm1 - DB 233,110,254,255,255 ; jmpq 2491 <_sk_load_tables_u16_be_sse2+0x23> + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 15,130,118,254,255,255 ; jb 24a0 <_sk_load_tables_u16_be_sse2+0x23> + DB 242,67,15,16,76,81,16 ; movsd 0x10(%r9,%r10,2),%xmm1 + DB 233,106,254,255,255 ; jmpq 24a0 <_sk_load_tables_u16_be_sse2+0x23> PUBLIC _sk_load_tables_rgb_u16_be_sse2 _sk_load_tables_rgb_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,80,1,0,0 ; jne 2785 <_sk_load_tables_rgb_u16_be_sse2+0x162> - DB 243,71,15,111,28,72 ; movdqu (%r8,%r9,2),%xmm11 - DB 243,67,15,111,76,72,8 ; movdqu 0x8(%r8,%r9,2),%xmm1 + DB 76,139,8 ; mov (%rax),%r9 + DB 76,141,20,82 ; lea (%rdx,%rdx,2),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,84,1,0,0 ; jne 279c <_sk_load_tables_rgb_u16_be_sse2+0x166> + DB 243,71,15,111,28,81 ; movdqu (%r9,%r10,2),%xmm11 + DB 243,67,15,111,76,81,8 ; movdqu 0x8(%r9,%r10,2),%xmm1 DB 102,15,115,217,4 ; psrldq $0x4,%xmm1 DB 102,69,15,111,211 ; movdqa %xmm11,%xmm10 DB 102,65,15,115,218,6 ; psrldq $0x6,%xmm10 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 102,68,15,97,217 ; punpcklwd %xmm1,%xmm11 DB 102,68,15,97,208 ; punpcklwd %xmm0,%xmm10 DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 DB 102,65,15,97,194 ; punpcklwd %xmm10,%xmm0 - DB 102,68,15,111,5,101,53,0,0 ; movdqa 0x3565(%rip),%xmm8 # 5be0 <_sk_callback_sse2+0x63a> + DB 102,68,15,111,5,128,53,0,0 ; movdqa 0x3580(%rip),%xmm8 # 5c10 <_sk_callback_sse2+0x636> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 - DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 69,15,182,200 ; movzbl %r8b,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,15,182,218 ; movzbl %r10b,%r11d - DB 73,193,234,30 ; shr $0x1e,%r10 + DB 102,73,15,126,217 ; movq %xmm3,%r9 + DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 69,15,182,243 ; movzbl %r11b,%r14d + DB 73,193,235,30 ; shr $0x1e,%r11 DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,28,19 ; movss (%rbx,%r10,1),%xmm3 - DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,66,15,16,28,27 ; movss (%rbx,%r11,1),%xmm3 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,66,15,16,4,179 ; movss (%rbx,%r14,4),%xmm0 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,194 ; unpcklps %xmm2,%xmm0 DB 15,20,195 ; unpcklps %xmm3,%xmm0 DB 102,65,15,219,200 ; pand %xmm8,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d + DB 68,15,182,203 ; movzbl %bl,%r9d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d - DB 73,193,233,30 ; shr $0x1e,%r9 - DB 243,67,15,16,20,14 ; movss (%r14,%r9,1),%xmm2 - DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 + DB 102,73,15,126,202 ; movq %xmm1,%r10 + DB 69,15,182,218 ; movzbl %r10b,%r11d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 243,67,15,16,20,23 ; movss (%r15,%r10,1),%xmm2 + DB 243,65,15,16,12,159 ; movss (%r15,%rbx,4),%xmm1 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 - DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 + DB 243,67,15,16,12,159 ; movss (%r15,%r11,4),%xmm1 + DB 243,67,15,16,28,143 ; movss (%r15,%r9,4),%xmm3 DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 DB 102,69,15,105,218 ; punpckhwd %xmm10,%xmm11 DB 102,69,15,219,216 ; pand %xmm8,%xmm11 DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11 DB 102,65,15,112,211,78 ; pshufd $0x4e,%xmm11,%xmm2 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d + DB 68,15,182,203 ; movzbl %bl,%r9d DB 72,193,235,32 ; shr $0x20,%rbx DB 102,76,15,126,216 ; movq %xmm11,%rax - DB 68,15,182,208 ; movzbl %al,%r10d + DB 68,15,182,216 ; movzbl %al,%r11d DB 72,193,232,30 ; shr $0x1e,%rax - DB 243,69,15,16,4,1 ; movss (%r9,%rax,1),%xmm8 - DB 243,65,15,16,20,153 ; movss (%r9,%rbx,4),%xmm2 + DB 243,69,15,16,4,2 ; movss (%r10,%rax,1),%xmm8 + DB 243,65,15,16,20,154 ; movss (%r10,%rbx,4),%xmm2 DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 - DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 - DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3 + DB 243,67,15,16,20,154 ; movss (%r10,%r11,4),%xmm2 + DB 243,67,15,16,28,138 ; movss (%r10,%r9,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,112,52,0,0 ; movaps 0x3470(%rip),%xmm3 # 5bf0 <_sk_callback_sse2+0x64a> + DB 15,40,29,139,52,0,0 ; movaps 0x348b(%rip),%xmm3 # 5c20 <_sk_callback_sse2+0x646> DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax - DB 102,71,15,110,28,72 ; movd (%r8,%r9,2),%xmm11 - DB 102,71,15,196,92,72,4,2 ; pinsrw $0x2,0x4(%r8,%r9,2),%xmm11 + DB 102,71,15,110,28,81 ; movd (%r9,%r10,2),%xmm11 + DB 102,71,15,196,92,81,4,2 ; pinsrw $0x2,0x4(%r9,%r10,2),%xmm11 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,14 ; jne 27ab <_sk_load_tables_rgb_u16_be_sse2+0x188> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,14 ; jne 27c2 <_sk_load_tables_rgb_u16_be_sse2+0x18c> DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 233,176,254,255,255 ; jmpq 265b <_sk_load_tables_rgb_u16_be_sse2+0x38> - DB 102,71,15,110,84,72,6 ; movd 0x6(%r8,%r9,2),%xmm10 - DB 102,71,15,196,84,72,10,2 ; pinsrw $0x2,0xa(%r8,%r9,2),%xmm10 + DB 233,172,254,255,255 ; jmpq 266e <_sk_load_tables_rgb_u16_be_sse2+0x38> + DB 102,71,15,110,84,81,6 ; movd 0x6(%r9,%r10,2),%xmm10 + DB 102,71,15,196,84,81,10,2 ; pinsrw $0x2,0xa(%r9,%r10,2),%xmm10 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,24 ; jb 27dc <_sk_load_tables_rgb_u16_be_sse2+0x1b9> - DB 102,67,15,110,76,72,12 ; movd 0xc(%r8,%r9,2),%xmm1 - DB 102,67,15,196,76,72,16,2 ; pinsrw $0x2,0x10(%r8,%r9,2),%xmm1 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,24 ; jb 27f3 <_sk_load_tables_rgb_u16_be_sse2+0x1bd> + DB 102,67,15,110,76,81,12 ; movd 0xc(%r9,%r10,2),%xmm1 + DB 102,67,15,196,76,81,16,2 ; pinsrw $0x2,0x10(%r9,%r10,2),%xmm1 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,127,254,255,255 ; jmpq 265b <_sk_load_tables_rgb_u16_be_sse2+0x38> + DB 233,123,254,255,255 ; jmpq 266e <_sk_load_tables_rgb_u16_be_sse2+0x38> DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,118,254,255,255 ; jmpq 265b <_sk_load_tables_rgb_u16_be_sse2+0x38> + DB 233,114,254,255,255 ; jmpq 266e <_sk_load_tables_rgb_u16_be_sse2+0x38> PUBLIC _sk_byte_tables_sse2 _sk_byte_tables_sse2 LABEL PROC DB 85 ; push %rbp + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,13,52,0,0 ; movaps 0x340d(%rip),%xmm8 # 5c00 <_sk_callback_sse2+0x65a> + DB 68,15,40,5,36,52,0,0 ; movaps 0x3424(%rip),%xmm8 # 5c30 <_sk_callback_sse2+0x656> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 - DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 77,137,203 ; mov %r9,%r11 + DB 73,193,235,32 ; shr $0x20,%r11 DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 - DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,198 ; mov %r8,%r14 - DB 73,193,238,32 ; shr $0x20,%r14 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,206 ; mov %r9d,%r14d + DB 77,137,207 ; mov %r9,%r15 + DB 73,193,239,32 ; shr $0x20,%r15 DB 72,139,24 ; mov (%rax),%rbx - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 70,15,182,28,27 ; movzbl (%rbx,%r11,1),%r11d - DB 66,15,182,44,51 ; movzbl (%rbx,%r14,1),%ebp + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 70,15,182,52,51 ; movzbl (%rbx,%r14,1),%r14d + DB 66,15,182,44,59 ; movzbl (%rbx,%r15,1),%ebp DB 193,229,8 ; shl $0x8,%ebp - DB 68,9,221 ; or %r11d,%ebp - DB 70,15,182,12,11 ; movzbl (%rbx,%r9,1),%r9d - DB 66,15,182,28,19 ; movzbl (%rbx,%r10,1),%ebx + DB 68,9,245 ; or %r14d,%ebp + DB 70,15,182,20,19 ; movzbl (%rbx,%r10,1),%r10d + DB 66,15,182,28,27 ; movzbl (%rbx,%r11,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,203 ; or %r9d,%ebx + DB 68,9,211 ; or %r10d,%ebx DB 102,15,196,195,0 ; pinsrw $0x0,%ebx,%xmm0 DB 102,15,196,197,1 ; pinsrw $0x1,%ebp,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,21,167,51,0,0 ; movaps 0x33a7(%rip),%xmm10 # 5c10 <_sk_callback_sse2+0x66a> + DB 68,15,40,21,190,51,0,0 ; movaps 0x33be(%rip),%xmm10 # 5c40 <_sk_callback_sse2+0x666> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 DB 102,72,15,126,205 ; movq %xmm1,%rbp - DB 65,137,233 ; mov %ebp,%r9d + DB 65,137,234 ; mov %ebp,%r10d DB 72,193,237,32 ; shr $0x20,%rbp DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1 DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 65,15,182,28,24 ; movzbl (%r8,%rbx,1),%ebx + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,211 ; or %r10d,%ebx - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp + DB 68,9,219 ; or %r11d,%ebx + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp DB 193,229,8 ; shl $0x8,%ebp - DB 68,9,205 ; or %r9d,%ebp + DB 68,9,213 ; or %r10d,%ebp DB 102,15,196,205,0 ; pinsrw $0x0,%ebp,%xmm1 DB 102,15,196,203,1 ; pinsrw $0x1,%ebx,%xmm1 DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,202 ; mulps %xmm10,%xmm1 - DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 76,139,80,16 ; mov 0x10(%rax),%r10 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,32 ; shr $0x20,%rbx DB 102,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm2 DB 102,72,15,126,213 ; movq %xmm2,%rbp - DB 65,137,234 ; mov %ebp,%r10d + DB 65,137,235 ; mov %ebp,%r11d DB 72,193,237,32 ; shr $0x20,%rbp - DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d - DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp + DB 71,15,182,28,26 ; movzbl (%r10,%r11,1),%r11d + DB 65,15,182,44,42 ; movzbl (%r10,%rbp,1),%ebp DB 193,229,8 ; shl $0x8,%ebp - DB 68,9,213 ; or %r10d,%ebp - DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d - DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx + DB 68,9,221 ; or %r11d,%ebp + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 65,15,182,28,26 ; movzbl (%r10,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,195 ; or %r8d,%ebx + DB 68,9,203 ; or %r9d,%ebx DB 102,15,196,211,0 ; pinsrw $0x0,%ebx,%xmm2 DB 102,15,196,213,1 ; pinsrw $0x1,%ebp,%xmm2 DB 102,65,15,96,209 ; punpcklbw %xmm9,%xmm2 @@ -21520,20 +21565,20 @@ _sk_byte_tables_sse2 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 102,15,91,219 ; cvtps2dq %xmm3,%xmm3 DB 102,72,15,126,221 ; movq %xmm3,%rbp - DB 65,137,232 ; mov %ebp,%r8d + DB 65,137,233 ; mov %ebp,%r9d DB 72,193,237,32 ; shr $0x20,%rbp DB 102,15,112,219,78 ; pshufd $0x4e,%xmm3,%xmm3 DB 102,72,15,126,219 ; movq %xmm3,%rbx - DB 65,137,217 ; mov %ebx,%r9d + DB 65,137,218 ; mov %ebx,%r10d DB 72,193,235,32 ; shr $0x20,%rbx - DB 70,15,182,12,8 ; movzbl (%rax,%r9,1),%r9d + DB 70,15,182,20,16 ; movzbl (%rax,%r10,1),%r10d DB 15,182,28,24 ; movzbl (%rax,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,203 ; or %r9d,%ebx - DB 70,15,182,4,0 ; movzbl (%rax,%r8,1),%r8d + DB 68,9,211 ; or %r10d,%ebx + DB 70,15,182,12,8 ; movzbl (%rax,%r9,1),%r9d DB 15,182,4,40 ; movzbl (%rax,%rbp,1),%eax DB 193,224,8 ; shl $0x8,%eax - DB 68,9,192 ; or %r8d,%eax + DB 68,9,200 ; or %r9d,%eax DB 102,15,196,216,0 ; pinsrw $0x0,%eax,%xmm3 DB 102,15,196,219,1 ; pinsrw $0x1,%ebx,%xmm3 DB 102,65,15,96,217 ; punpcklbw %xmm9,%xmm3 @@ -21543,66 +21588,68 @@ _sk_byte_tables_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_byte_tables_rgb_sse2 _sk_byte_tables_rgb_sse2 LABEL PROC DB 85 ; push %rbp + DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,139,64,24 ; mov 0x18(%rax),%r8d - DB 65,255,200 ; dec %r8d - DB 102,69,15,110,192 ; movd %r8d,%xmm8 + DB 68,139,72,24 ; mov 0x18(%rax),%r9d + DB 65,255,201 ; dec %r9d + DB 102,69,15,110,193 ; movd %r9d,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 - DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 77,137,203 ; mov %r9,%r11 + DB 73,193,235,32 ; shr $0x20,%r11 DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 - DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 69,137,195 ; mov %r8d,%r11d - DB 77,137,198 ; mov %r8,%r14 - DB 73,193,238,32 ; shr $0x20,%r14 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,206 ; mov %r9d,%r14d + DB 77,137,207 ; mov %r9,%r15 + DB 73,193,239,32 ; shr $0x20,%r15 DB 72,139,24 ; mov (%rax),%rbx - DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 70,15,182,28,27 ; movzbl (%rbx,%r11,1),%r11d - DB 66,15,182,44,51 ; movzbl (%rbx,%r14,1),%ebp + DB 76,139,72,8 ; mov 0x8(%rax),%r9 + DB 70,15,182,52,51 ; movzbl (%rbx,%r14,1),%r14d + DB 66,15,182,44,59 ; movzbl (%rbx,%r15,1),%ebp DB 193,229,8 ; shl $0x8,%ebp - DB 68,9,221 ; or %r11d,%ebp - DB 70,15,182,12,11 ; movzbl (%rbx,%r9,1),%r9d - DB 66,15,182,28,19 ; movzbl (%rbx,%r10,1),%ebx + DB 68,9,245 ; or %r14d,%ebp + DB 70,15,182,20,19 ; movzbl (%rbx,%r10,1),%r10d + DB 66,15,182,28,27 ; movzbl (%rbx,%r11,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,203 ; or %r9d,%ebx + DB 68,9,211 ; or %r10d,%ebx DB 102,15,196,195,0 ; pinsrw $0x0,%ebx,%xmm0 DB 102,15,196,197,1 ; pinsrw $0x1,%ebp,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,21,243,49,0,0 ; movaps 0x31f3(%rip),%xmm10 # 5c20 <_sk_callback_sse2+0x67a> + DB 68,15,40,21,6,50,0,0 ; movaps 0x3206(%rip),%xmm10 # 5c50 <_sk_callback_sse2+0x676> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 DB 102,72,15,126,205 ; movq %xmm1,%rbp - DB 65,137,233 ; mov %ebp,%r9d + DB 65,137,234 ; mov %ebp,%r10d DB 72,193,237,32 ; shr $0x20,%rbp DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1 DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 65,15,182,28,24 ; movzbl (%r8,%rbx,1),%ebx + DB 71,15,182,28,25 ; movzbl (%r9,%r11,1),%r11d + DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,211 ; or %r10d,%ebx - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp + DB 68,9,219 ; or %r11d,%ebx + DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d + DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp DB 193,229,8 ; shl $0x8,%ebp - DB 68,9,205 ; or %r9d,%ebp + DB 68,9,213 ; or %r10d,%ebp DB 102,15,196,205,0 ; pinsrw $0x0,%ebp,%xmm1 DB 102,15,196,203,1 ; pinsrw $0x1,%ebx,%xmm1 DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1 @@ -21613,20 +21660,20 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2 DB 102,72,15,126,213 ; movq %xmm2,%rbp - DB 65,137,232 ; mov %ebp,%r8d + DB 65,137,233 ; mov %ebp,%r9d DB 72,193,237,32 ; shr $0x20,%rbp DB 102,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm2 DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 65,137,217 ; mov %ebx,%r9d + DB 65,137,218 ; mov %ebx,%r10d DB 72,193,235,32 ; shr $0x20,%rbx - DB 70,15,182,12,8 ; movzbl (%rax,%r9,1),%r9d + DB 70,15,182,20,16 ; movzbl (%rax,%r10,1),%r10d DB 15,182,28,24 ; movzbl (%rax,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx - DB 68,9,203 ; or %r9d,%ebx - DB 70,15,182,4,0 ; movzbl (%rax,%r8,1),%r8d + DB 68,9,211 ; or %r10d,%ebx + DB 70,15,182,12,8 ; movzbl (%rax,%r9,1),%r9d DB 15,182,4,40 ; movzbl (%rax,%rbp,1),%eax DB 193,224,8 ; shl $0x8,%eax - DB 68,9,192 ; or %r8d,%eax + DB 68,9,200 ; or %r9d,%eax DB 102,15,196,208,0 ; pinsrw $0x0,%eax,%xmm2 DB 102,15,196,211,1 ; pinsrw $0x1,%ebx,%xmm2 DB 102,65,15,96,209 ; punpcklbw %xmm9,%xmm2 @@ -21636,13 +21683,15 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_table_r_sse2 _sk_table_r_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -21652,25 +21701,27 @@ _sk_table_r_sse2 LABEL PROC DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,112,192,78 ; pshufd $0x4e,%xmm8,%xmm0 DB 102,72,15,126,192 ; movq %xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,77,15,126,194 ; movq %xmm8,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,4,128 ; movss (%r8,%rax,4),%xmm0 + DB 102,77,15,126,195 ; movq %xmm8,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,71,15,16,4,153 ; movss (%r9,%r11,4),%xmm8 + DB 243,65,15,16,4,129 ; movss (%r9,%rax,4),%xmm0 DB 68,15,20,192 ; unpcklps %xmm0,%xmm8 - DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 243,65,15,16,4,153 ; movss (%r9,%rbx,4),%xmm0 + DB 243,71,15,16,12,145 ; movss (%r9,%r10,4),%xmm9 DB 65,15,20,193 ; unpcklps %xmm9,%xmm0 DB 65,15,20,192 ; unpcklps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_table_g_sse2 _sk_table_g_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -21680,25 +21731,27 @@ _sk_table_g_sse2 LABEL PROC DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,112,200,78 ; pshufd $0x4e,%xmm8,%xmm1 DB 102,72,15,126,200 ; movq %xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,77,15,126,194 ; movq %xmm8,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,12,128 ; movss (%r8,%rax,4),%xmm1 + DB 102,77,15,126,195 ; movq %xmm8,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,71,15,16,4,153 ; movss (%r9,%r11,4),%xmm8 + DB 243,65,15,16,12,129 ; movss (%r9,%rax,4),%xmm1 DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 - DB 243,67,15,16,12,152 ; movss (%r8,%r11,4),%xmm1 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 243,65,15,16,12,153 ; movss (%r9,%rbx,4),%xmm1 + DB 243,71,15,16,12,145 ; movss (%r9,%r10,4),%xmm9 DB 65,15,20,201 ; unpcklps %xmm9,%xmm1 DB 65,15,20,200 ; unpcklps %xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_table_b_sse2 _sk_table_b_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -21708,25 +21761,27 @@ _sk_table_b_sse2 LABEL PROC DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,112,208,78 ; pshufd $0x4e,%xmm8,%xmm2 DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,77,15,126,194 ; movq %xmm8,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,20,128 ; movss (%r8,%rax,4),%xmm2 + DB 102,77,15,126,195 ; movq %xmm8,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,71,15,16,4,153 ; movss (%r9,%r11,4),%xmm8 + DB 243,65,15,16,20,129 ; movss (%r9,%rax,4),%xmm2 DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 - DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 243,65,15,16,20,153 ; movss (%r9,%rbx,4),%xmm2 + DB 243,71,15,16,12,145 ; movss (%r9,%r10,4),%xmm9 DB 65,15,20,209 ; unpcklps %xmm9,%xmm2 DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_table_a_sse2 _sk_table_a_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 139,64,8 ; mov 0x8(%rax),%eax DB 255,200 ; dec %eax DB 102,68,15,110,192 ; movd %eax,%xmm8 @@ -21736,19 +21791,20 @@ _sk_table_a_sse2 LABEL PROC DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,112,216,78 ; pshufd $0x4e,%xmm8,%xmm3 DB 102,72,15,126,216 ; movq %xmm3,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,77,15,126,194 ; movq %xmm8,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,28,128 ; movss (%r8,%rax,4),%xmm3 + DB 102,77,15,126,195 ; movq %xmm8,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,71,15,16,4,153 ; movss (%r9,%r11,4),%xmm8 + DB 243,65,15,16,28,129 ; movss (%r9,%rax,4),%xmm3 DB 68,15,20,195 ; unpcklps %xmm3,%xmm8 - DB 243,67,15,16,28,152 ; movss (%r8,%r11,4),%xmm3 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 243,65,15,16,28,153 ; movss (%r9,%rbx,4),%xmm3 + DB 243,71,15,16,12,145 ; movss (%r9,%r10,4),%xmm9 DB 65,15,20,217 ; unpcklps %xmm9,%xmm3 DB 65,15,20,216 ; unpcklps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_parametric_r_sse2 @@ -21772,15 +21828,15 @@ _sk_parametric_r_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,49,47,0,0 ; mulps 0x2f31(%rip),%xmm9 # 5c30 <_sk_callback_sse2+0x68a> - DB 68,15,84,21,57,47,0,0 ; andps 0x2f39(%rip),%xmm10 # 5c40 <_sk_callback_sse2+0x69a> - DB 68,15,86,21,65,47,0,0 ; orps 0x2f41(%rip),%xmm10 # 5c50 <_sk_callback_sse2+0x6aa> - DB 68,15,88,13,73,47,0,0 ; addps 0x2f49(%rip),%xmm9 # 5c60 <_sk_callback_sse2+0x6ba> - DB 68,15,40,37,81,47,0,0 ; movaps 0x2f51(%rip),%xmm12 # 5c70 <_sk_callback_sse2+0x6ca> + DB 68,15,89,13,58,47,0,0 ; mulps 0x2f3a(%rip),%xmm9 # 5c60 <_sk_callback_sse2+0x686> + DB 68,15,84,21,66,47,0,0 ; andps 0x2f42(%rip),%xmm10 # 5c70 <_sk_callback_sse2+0x696> + DB 68,15,86,21,74,47,0,0 ; orps 0x2f4a(%rip),%xmm10 # 5c80 <_sk_callback_sse2+0x6a6> + DB 68,15,88,13,82,47,0,0 ; addps 0x2f52(%rip),%xmm9 # 5c90 <_sk_callback_sse2+0x6b6> + DB 68,15,40,37,90,47,0,0 ; movaps 0x2f5a(%rip),%xmm12 # 5ca0 <_sk_callback_sse2+0x6c6> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,81,47,0,0 ; addps 0x2f51(%rip),%xmm10 # 5c80 <_sk_callback_sse2+0x6da> - DB 68,15,40,37,89,47,0,0 ; movaps 0x2f59(%rip),%xmm12 # 5c90 <_sk_callback_sse2+0x6ea> + DB 68,15,88,21,90,47,0,0 ; addps 0x2f5a(%rip),%xmm10 # 5cb0 <_sk_callback_sse2+0x6d6> + DB 68,15,40,37,98,47,0,0 ; movaps 0x2f62(%rip),%xmm12 # 5cc0 <_sk_callback_sse2+0x6e6> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -21788,22 +21844,22 @@ _sk_parametric_r_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,67,47,0,0 ; movaps 0x2f43(%rip),%xmm10 # 5ca0 <_sk_callback_sse2+0x6fa> + DB 68,15,40,21,76,47,0,0 ; movaps 0x2f4c(%rip),%xmm10 # 5cd0 <_sk_callback_sse2+0x6f6> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,55,47,0,0 ; addps 0x2f37(%rip),%xmm9 # 5cb0 <_sk_callback_sse2+0x70a> - DB 68,15,40,37,63,47,0,0 ; movaps 0x2f3f(%rip),%xmm12 # 5cc0 <_sk_callback_sse2+0x71a> + DB 68,15,88,13,64,47,0,0 ; addps 0x2f40(%rip),%xmm9 # 5ce0 <_sk_callback_sse2+0x706> + DB 68,15,40,37,72,47,0,0 ; movaps 0x2f48(%rip),%xmm12 # 5cf0 <_sk_callback_sse2+0x716> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,63,47,0,0 ; movaps 0x2f3f(%rip),%xmm12 # 5cd0 <_sk_callback_sse2+0x72a> + DB 68,15,40,37,72,47,0,0 ; movaps 0x2f48(%rip),%xmm12 # 5d00 <_sk_callback_sse2+0x726> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,67,47,0,0 ; movaps 0x2f43(%rip),%xmm13 # 5ce0 <_sk_callback_sse2+0x73a> + DB 68,15,40,45,76,47,0,0 ; movaps 0x2f4c(%rip),%xmm13 # 5d10 <_sk_callback_sse2+0x736> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,67,47,0,0 ; mulps 0x2f43(%rip),%xmm13 # 5cf0 <_sk_callback_sse2+0x74a> + DB 68,15,89,45,76,47,0,0 ; mulps 0x2f4c(%rip),%xmm13 # 5d20 <_sk_callback_sse2+0x746> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -21837,15 +21893,15 @@ _sk_parametric_g_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,195,46,0,0 ; mulps 0x2ec3(%rip),%xmm9 # 5d00 <_sk_callback_sse2+0x75a> - DB 68,15,84,21,203,46,0,0 ; andps 0x2ecb(%rip),%xmm10 # 5d10 <_sk_callback_sse2+0x76a> - DB 68,15,86,21,211,46,0,0 ; orps 0x2ed3(%rip),%xmm10 # 5d20 <_sk_callback_sse2+0x77a> - DB 68,15,88,13,219,46,0,0 ; addps 0x2edb(%rip),%xmm9 # 5d30 <_sk_callback_sse2+0x78a> - DB 68,15,40,37,227,46,0,0 ; movaps 0x2ee3(%rip),%xmm12 # 5d40 <_sk_callback_sse2+0x79a> + DB 68,15,89,13,204,46,0,0 ; mulps 0x2ecc(%rip),%xmm9 # 5d30 <_sk_callback_sse2+0x756> + DB 68,15,84,21,212,46,0,0 ; andps 0x2ed4(%rip),%xmm10 # 5d40 <_sk_callback_sse2+0x766> + DB 68,15,86,21,220,46,0,0 ; orps 0x2edc(%rip),%xmm10 # 5d50 <_sk_callback_sse2+0x776> + DB 68,15,88,13,228,46,0,0 ; addps 0x2ee4(%rip),%xmm9 # 5d60 <_sk_callback_sse2+0x786> + DB 68,15,40,37,236,46,0,0 ; movaps 0x2eec(%rip),%xmm12 # 5d70 <_sk_callback_sse2+0x796> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,227,46,0,0 ; addps 0x2ee3(%rip),%xmm10 # 5d50 <_sk_callback_sse2+0x7aa> - DB 68,15,40,37,235,46,0,0 ; movaps 0x2eeb(%rip),%xmm12 # 5d60 <_sk_callback_sse2+0x7ba> + DB 68,15,88,21,236,46,0,0 ; addps 0x2eec(%rip),%xmm10 # 5d80 <_sk_callback_sse2+0x7a6> + DB 68,15,40,37,244,46,0,0 ; movaps 0x2ef4(%rip),%xmm12 # 5d90 <_sk_callback_sse2+0x7b6> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -21853,22 +21909,22 @@ _sk_parametric_g_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,213,46,0,0 ; movaps 0x2ed5(%rip),%xmm10 # 5d70 <_sk_callback_sse2+0x7ca> + DB 68,15,40,21,222,46,0,0 ; movaps 0x2ede(%rip),%xmm10 # 5da0 <_sk_callback_sse2+0x7c6> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,201,46,0,0 ; addps 0x2ec9(%rip),%xmm9 # 5d80 <_sk_callback_sse2+0x7da> - DB 68,15,40,37,209,46,0,0 ; movaps 0x2ed1(%rip),%xmm12 # 5d90 <_sk_callback_sse2+0x7ea> + DB 68,15,88,13,210,46,0,0 ; addps 0x2ed2(%rip),%xmm9 # 5db0 <_sk_callback_sse2+0x7d6> + DB 68,15,40,37,218,46,0,0 ; movaps 0x2eda(%rip),%xmm12 # 5dc0 <_sk_callback_sse2+0x7e6> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,209,46,0,0 ; movaps 0x2ed1(%rip),%xmm12 # 5da0 <_sk_callback_sse2+0x7fa> + DB 68,15,40,37,218,46,0,0 ; movaps 0x2eda(%rip),%xmm12 # 5dd0 <_sk_callback_sse2+0x7f6> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,213,46,0,0 ; movaps 0x2ed5(%rip),%xmm13 # 5db0 <_sk_callback_sse2+0x80a> + DB 68,15,40,45,222,46,0,0 ; movaps 0x2ede(%rip),%xmm13 # 5de0 <_sk_callback_sse2+0x806> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,213,46,0,0 ; mulps 0x2ed5(%rip),%xmm13 # 5dc0 <_sk_callback_sse2+0x81a> + DB 68,15,89,45,222,46,0,0 ; mulps 0x2ede(%rip),%xmm13 # 5df0 <_sk_callback_sse2+0x816> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -21902,15 +21958,15 @@ _sk_parametric_b_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,85,46,0,0 ; mulps 0x2e55(%rip),%xmm9 # 5dd0 <_sk_callback_sse2+0x82a> - DB 68,15,84,21,93,46,0,0 ; andps 0x2e5d(%rip),%xmm10 # 5de0 <_sk_callback_sse2+0x83a> - DB 68,15,86,21,101,46,0,0 ; orps 0x2e65(%rip),%xmm10 # 5df0 <_sk_callback_sse2+0x84a> - DB 68,15,88,13,109,46,0,0 ; addps 0x2e6d(%rip),%xmm9 # 5e00 <_sk_callback_sse2+0x85a> - DB 68,15,40,37,117,46,0,0 ; movaps 0x2e75(%rip),%xmm12 # 5e10 <_sk_callback_sse2+0x86a> + DB 68,15,89,13,94,46,0,0 ; mulps 0x2e5e(%rip),%xmm9 # 5e00 <_sk_callback_sse2+0x826> + DB 68,15,84,21,102,46,0,0 ; andps 0x2e66(%rip),%xmm10 # 5e10 <_sk_callback_sse2+0x836> + DB 68,15,86,21,110,46,0,0 ; orps 0x2e6e(%rip),%xmm10 # 5e20 <_sk_callback_sse2+0x846> + DB 68,15,88,13,118,46,0,0 ; addps 0x2e76(%rip),%xmm9 # 5e30 <_sk_callback_sse2+0x856> + DB 68,15,40,37,126,46,0,0 ; movaps 0x2e7e(%rip),%xmm12 # 5e40 <_sk_callback_sse2+0x866> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,117,46,0,0 ; addps 0x2e75(%rip),%xmm10 # 5e20 <_sk_callback_sse2+0x87a> - DB 68,15,40,37,125,46,0,0 ; movaps 0x2e7d(%rip),%xmm12 # 5e30 <_sk_callback_sse2+0x88a> + DB 68,15,88,21,126,46,0,0 ; addps 0x2e7e(%rip),%xmm10 # 5e50 <_sk_callback_sse2+0x876> + DB 68,15,40,37,134,46,0,0 ; movaps 0x2e86(%rip),%xmm12 # 5e60 <_sk_callback_sse2+0x886> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -21918,22 +21974,22 @@ _sk_parametric_b_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,103,46,0,0 ; movaps 0x2e67(%rip),%xmm10 # 5e40 <_sk_callback_sse2+0x89a> + DB 68,15,40,21,112,46,0,0 ; movaps 0x2e70(%rip),%xmm10 # 5e70 <_sk_callback_sse2+0x896> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,91,46,0,0 ; addps 0x2e5b(%rip),%xmm9 # 5e50 <_sk_callback_sse2+0x8aa> - DB 68,15,40,37,99,46,0,0 ; movaps 0x2e63(%rip),%xmm12 # 5e60 <_sk_callback_sse2+0x8ba> + DB 68,15,88,13,100,46,0,0 ; addps 0x2e64(%rip),%xmm9 # 5e80 <_sk_callback_sse2+0x8a6> + DB 68,15,40,37,108,46,0,0 ; movaps 0x2e6c(%rip),%xmm12 # 5e90 <_sk_callback_sse2+0x8b6> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,99,46,0,0 ; movaps 0x2e63(%rip),%xmm12 # 5e70 <_sk_callback_sse2+0x8ca> + DB 68,15,40,37,108,46,0,0 ; movaps 0x2e6c(%rip),%xmm12 # 5ea0 <_sk_callback_sse2+0x8c6> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,103,46,0,0 ; movaps 0x2e67(%rip),%xmm13 # 5e80 <_sk_callback_sse2+0x8da> + DB 68,15,40,45,112,46,0,0 ; movaps 0x2e70(%rip),%xmm13 # 5eb0 <_sk_callback_sse2+0x8d6> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,103,46,0,0 ; mulps 0x2e67(%rip),%xmm13 # 5e90 <_sk_callback_sse2+0x8ea> + DB 68,15,89,45,112,46,0,0 ; mulps 0x2e70(%rip),%xmm13 # 5ec0 <_sk_callback_sse2+0x8e6> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -21967,15 +22023,15 @@ _sk_parametric_a_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,231,45,0,0 ; mulps 0x2de7(%rip),%xmm9 # 5ea0 <_sk_callback_sse2+0x8fa> - DB 68,15,84,21,239,45,0,0 ; andps 0x2def(%rip),%xmm10 # 5eb0 <_sk_callback_sse2+0x90a> - DB 68,15,86,21,247,45,0,0 ; orps 0x2df7(%rip),%xmm10 # 5ec0 <_sk_callback_sse2+0x91a> - DB 68,15,88,13,255,45,0,0 ; addps 0x2dff(%rip),%xmm9 # 5ed0 <_sk_callback_sse2+0x92a> - DB 68,15,40,37,7,46,0,0 ; movaps 0x2e07(%rip),%xmm12 # 5ee0 <_sk_callback_sse2+0x93a> + DB 68,15,89,13,240,45,0,0 ; mulps 0x2df0(%rip),%xmm9 # 5ed0 <_sk_callback_sse2+0x8f6> + DB 68,15,84,21,248,45,0,0 ; andps 0x2df8(%rip),%xmm10 # 5ee0 <_sk_callback_sse2+0x906> + DB 68,15,86,21,0,46,0,0 ; orps 0x2e00(%rip),%xmm10 # 5ef0 <_sk_callback_sse2+0x916> + DB 68,15,88,13,8,46,0,0 ; addps 0x2e08(%rip),%xmm9 # 5f00 <_sk_callback_sse2+0x926> + DB 68,15,40,37,16,46,0,0 ; movaps 0x2e10(%rip),%xmm12 # 5f10 <_sk_callback_sse2+0x936> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,7,46,0,0 ; addps 0x2e07(%rip),%xmm10 # 5ef0 <_sk_callback_sse2+0x94a> - DB 68,15,40,37,15,46,0,0 ; movaps 0x2e0f(%rip),%xmm12 # 5f00 <_sk_callback_sse2+0x95a> + DB 68,15,88,21,16,46,0,0 ; addps 0x2e10(%rip),%xmm10 # 5f20 <_sk_callback_sse2+0x946> + DB 68,15,40,37,24,46,0,0 ; movaps 0x2e18(%rip),%xmm12 # 5f30 <_sk_callback_sse2+0x956> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -21983,22 +22039,22 @@ _sk_parametric_a_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,249,45,0,0 ; movaps 0x2df9(%rip),%xmm10 # 5f10 <_sk_callback_sse2+0x96a> + DB 68,15,40,21,2,46,0,0 ; movaps 0x2e02(%rip),%xmm10 # 5f40 <_sk_callback_sse2+0x966> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,237,45,0,0 ; addps 0x2ded(%rip),%xmm9 # 5f20 <_sk_callback_sse2+0x97a> - DB 68,15,40,37,245,45,0,0 ; movaps 0x2df5(%rip),%xmm12 # 5f30 <_sk_callback_sse2+0x98a> + DB 68,15,88,13,246,45,0,0 ; addps 0x2df6(%rip),%xmm9 # 5f50 <_sk_callback_sse2+0x976> + DB 68,15,40,37,254,45,0,0 ; movaps 0x2dfe(%rip),%xmm12 # 5f60 <_sk_callback_sse2+0x986> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,245,45,0,0 ; movaps 0x2df5(%rip),%xmm12 # 5f40 <_sk_callback_sse2+0x99a> + DB 68,15,40,37,254,45,0,0 ; movaps 0x2dfe(%rip),%xmm12 # 5f70 <_sk_callback_sse2+0x996> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,249,45,0,0 ; movaps 0x2df9(%rip),%xmm13 # 5f50 <_sk_callback_sse2+0x9aa> + DB 68,15,40,45,2,46,0,0 ; movaps 0x2e02(%rip),%xmm13 # 5f80 <_sk_callback_sse2+0x9a6> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,249,45,0,0 ; mulps 0x2df9(%rip),%xmm13 # 5f60 <_sk_callback_sse2+0x9ba> + DB 68,15,89,45,2,46,0,0 ; mulps 0x2e02(%rip),%xmm13 # 5f90 <_sk_callback_sse2+0x9b6> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -22013,29 +22069,29 @@ _sk_parametric_a_sse2 LABEL PROC PUBLIC _sk_lab_to_xyz_sse2 _sk_lab_to_xyz_sse2 LABEL PROC - DB 15,89,5,214,45,0,0 ; mulps 0x2dd6(%rip),%xmm0 # 5f70 <_sk_callback_sse2+0x9ca> - DB 68,15,40,5,222,45,0,0 ; movaps 0x2dde(%rip),%xmm8 # 5f80 <_sk_callback_sse2+0x9da> + DB 15,89,5,223,45,0,0 ; mulps 0x2ddf(%rip),%xmm0 # 5fa0 <_sk_callback_sse2+0x9c6> + DB 68,15,40,5,231,45,0,0 ; movaps 0x2de7(%rip),%xmm8 # 5fb0 <_sk_callback_sse2+0x9d6> DB 65,15,89,200 ; mulps %xmm8,%xmm1 - DB 68,15,40,13,226,45,0,0 ; movaps 0x2de2(%rip),%xmm9 # 5f90 <_sk_callback_sse2+0x9ea> + DB 68,15,40,13,235,45,0,0 ; movaps 0x2deb(%rip),%xmm9 # 5fc0 <_sk_callback_sse2+0x9e6> DB 65,15,88,201 ; addps %xmm9,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 65,15,88,209 ; addps %xmm9,%xmm2 - DB 15,88,5,223,45,0,0 ; addps 0x2ddf(%rip),%xmm0 # 5fa0 <_sk_callback_sse2+0x9fa> - DB 15,89,5,232,45,0,0 ; mulps 0x2de8(%rip),%xmm0 # 5fb0 <_sk_callback_sse2+0xa0a> - DB 15,89,13,241,45,0,0 ; mulps 0x2df1(%rip),%xmm1 # 5fc0 <_sk_callback_sse2+0xa1a> + DB 15,88,5,232,45,0,0 ; addps 0x2de8(%rip),%xmm0 # 5fd0 <_sk_callback_sse2+0x9f6> + DB 15,89,5,241,45,0,0 ; mulps 0x2df1(%rip),%xmm0 # 5fe0 <_sk_callback_sse2+0xa06> + DB 15,89,13,250,45,0,0 ; mulps 0x2dfa(%rip),%xmm1 # 5ff0 <_sk_callback_sse2+0xa16> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,21,247,45,0,0 ; mulps 0x2df7(%rip),%xmm2 # 5fd0 <_sk_callback_sse2+0xa2a> + DB 15,89,21,0,46,0,0 ; mulps 0x2e00(%rip),%xmm2 # 6000 <_sk_callback_sse2+0xa26> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 68,15,92,202 ; subps %xmm2,%xmm9 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 DB 68,15,89,225 ; mulps %xmm1,%xmm12 - DB 15,40,21,236,45,0,0 ; movaps 0x2dec(%rip),%xmm2 # 5fe0 <_sk_callback_sse2+0xa3a> + DB 15,40,21,245,45,0,0 ; movaps 0x2df5(%rip),%xmm2 # 6010 <_sk_callback_sse2+0xa36> DB 68,15,40,194 ; movaps %xmm2,%xmm8 DB 69,15,194,196,1 ; cmpltps %xmm12,%xmm8 - DB 68,15,40,21,235,45,0,0 ; movaps 0x2deb(%rip),%xmm10 # 5ff0 <_sk_callback_sse2+0xa4a> + DB 68,15,40,21,244,45,0,0 ; movaps 0x2df4(%rip),%xmm10 # 6020 <_sk_callback_sse2+0xa46> DB 65,15,88,202 ; addps %xmm10,%xmm1 - DB 68,15,40,29,239,45,0,0 ; movaps 0x2def(%rip),%xmm11 # 6000 <_sk_callback_sse2+0xa5a> + DB 68,15,40,29,248,45,0,0 ; movaps 0x2df8(%rip),%xmm11 # 6030 <_sk_callback_sse2+0xa56> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 69,15,84,224 ; andps %xmm8,%xmm12 DB 68,15,85,193 ; andnps %xmm1,%xmm8 @@ -22059,8 +22115,8 @@ _sk_lab_to_xyz_sse2 LABEL PROC DB 15,84,194 ; andps %xmm2,%xmm0 DB 65,15,85,209 ; andnps %xmm9,%xmm2 DB 15,86,208 ; orps %xmm0,%xmm2 - DB 68,15,89,5,159,45,0,0 ; mulps 0x2d9f(%rip),%xmm8 # 6010 <_sk_callback_sse2+0xa6a> - DB 15,89,21,168,45,0,0 ; mulps 0x2da8(%rip),%xmm2 # 6020 <_sk_callback_sse2+0xa7a> + DB 68,15,89,5,168,45,0,0 ; mulps 0x2da8(%rip),%xmm8 # 6040 <_sk_callback_sse2+0xa66> + DB 15,89,21,177,45,0,0 ; mulps 0x2db1(%rip),%xmm2 # 6050 <_sk_callback_sse2+0xa76> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -22068,46 +22124,48 @@ _sk_lab_to_xyz_sse2 LABEL PROC PUBLIC _sk_load_a8_sse2 _sk_load_a8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,45 ; jne 32b7 <_sk_load_a8_sse2+0x37> - DB 102,65,15,110,4,57 ; movd (%r9,%rdi,1),%xmm0 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,45 ; jne 32de <_sk_load_a8_sse2+0x37> + DB 102,65,15,110,4,18 ; movd (%r10,%rdx,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,5,144,45,0,0 ; pand 0x2d90(%rip),%xmm0 # 6030 <_sk_callback_sse2+0xa8a> + DB 102,15,219,5,153,45,0,0 ; pand 0x2d99(%rip),%xmm0 # 6060 <_sk_callback_sse2+0xa86> DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,150,45,0,0 ; mulps 0x2d96(%rip),%xmm3 # 6040 <_sk_callback_sse2+0xa9a> + DB 15,89,29,159,45,0,0 ; mulps 0x2d9f(%rip),%xmm3 # 6070 <_sk_callback_sse2+0xa96> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,48 ; je 32f8 <_sk_load_a8_sse2+0x78> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 32e3 <_sk_load_a8_sse2+0x63> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,196 ; jne 3298 <_sk_load_a8_sse2+0x18> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,48 ; je 331f <_sk_load_a8_sse2+0x78> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 330a <_sk_load_a8_sse2+0x63> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,196 ; jne 32bf <_sk_load_a8_sse2+0x18> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,15,110,200 ; movd %eax,%xmm1 DB 15,198,200,0 ; shufps $0x0,%xmm0,%xmm1 DB 15,198,200,226 ; shufps $0xe2,%xmm0,%xmm1 DB 15,40,193 ; movaps %xmm1,%xmm0 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,15,110,200 ; movd %eax,%xmm1 DB 243,15,16,193 ; movss %xmm1,%xmm0 - DB 235,145 ; jmp 3298 <_sk_load_a8_sse2+0x18> + DB 235,145 ; jmp 32bf <_sk_load_a8_sse2+0x18> PUBLIC _sk_gather_a8_sse2 _sk_gather_a8_sse2 LABEL PROC + DB 85 ; push %rbp + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22120,116 +22178,120 @@ _sk_gather_a8_sse2 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,72,15,126,192 ; movq %xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 71,15,182,28,24 ; movzbl (%r8,%r11,1),%r11d - DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 65,193,226,8 ; shl $0x8,%r10d - DB 69,9,218 ; or %r11d,%r10d - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp + DB 67,15,182,28,25 ; movzbl (%r9,%r11,1),%ebx + DB 193,227,8 ; shl $0x8,%ebx + DB 9,235 ; or %ebp,%ebx + DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 193,224,8 ; shl $0x8,%eax - DB 68,9,200 ; or %r9d,%eax + DB 9,232 ; or %ebp,%eax DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 - DB 102,65,15,196,194,1 ; pinsrw $0x1,%r10d,%xmm0 + DB 102,15,196,195,1 ; pinsrw $0x1,%ebx,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,180,44,0,0 ; mulps 0x2cb4(%rip),%xmm3 # 6050 <_sk_callback_sse2+0xaaa> + DB 15,89,29,191,44,0,0 ; mulps 0x2cbf(%rip),%xmm3 # 6080 <_sk_callback_sse2+0xaa6> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 91 ; pop %rbx + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_store_a8_sse2 _sk_store_a8_sse2 LABEL PROC DB 72,131,236,56 ; sub $0x38,%rsp DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 68,15,40,5,164,44,0,0 ; movaps 0x2ca4(%rip),%xmm8 # 6060 <_sk_callback_sse2+0xaba> + DB 76,139,16 ; mov (%rax),%r10 + DB 68,15,40,5,173,44,0,0 ; movaps 0x2cad(%rip),%xmm8 # 6090 <_sk_callback_sse2+0xab6> DB 68,15,89,195 ; mulps %xmm3,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,17 ; jne 33f1 <_sk_store_a8_sse2+0x46> + DB 77,133,192 ; test %r8,%r8 + DB 117,17 ; jne 3418 <_sk_store_a8_sse2+0x46> DB 102,68,15,126,192 ; movd %xmm8,%eax - DB 65,137,4,57 ; mov %eax,(%r9,%rdi,1) + DB 65,137,4,18 ; mov %eax,(%r10,%rdx,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 72,131,196,56 ; add $0x38,%rsp DB 255,224 ; jmpq *%rax DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,44 ; je 3434 <_sk_store_a8_sse2+0x89> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,22 ; je 3424 <_sk_store_a8_sse2+0x79> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,213 ; jne 33e9 <_sk_store_a8_sse2+0x3e> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,44 ; je 345b <_sk_store_a8_sse2+0x89> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,22 ; je 344b <_sk_store_a8_sse2+0x79> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,213 ; jne 3410 <_sk_store_a8_sse2+0x3e> DB 102,68,15,127,68,36,32 ; movdqa %xmm8,0x20(%rsp) DB 138,68,36,40 ; mov 0x28(%rsp),%al - DB 65,136,68,57,2 ; mov %al,0x2(%r9,%rdi,1) + DB 65,136,68,18,2 ; mov %al,0x2(%r10,%rdx,1) DB 102,68,15,127,68,36,16 ; movdqa %xmm8,0x10(%rsp) DB 138,68,36,20 ; mov 0x14(%rsp),%al - DB 65,136,68,57,1 ; mov %al,0x1(%r9,%rdi,1) + DB 65,136,68,18,1 ; mov %al,0x1(%r10,%rdx,1) DB 102,68,15,127,4,36 ; movdqa %xmm8,(%rsp) DB 138,4,36 ; mov (%rsp),%al - DB 65,136,4,57 ; mov %al,(%r9,%rdi,1) - DB 235,166 ; jmp 33e9 <_sk_store_a8_sse2+0x3e> + DB 65,136,4,18 ; mov %al,(%r10,%rdx,1) + DB 235,166 ; jmp 3410 <_sk_store_a8_sse2+0x3e> PUBLIC _sk_load_g8_sse2 _sk_load_g8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,49 ; jne 347e <_sk_load_g8_sse2+0x3b> - DB 102,65,15,110,4,57 ; movd (%r9,%rdi,1),%xmm0 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,49 ; jne 34a5 <_sk_load_g8_sse2+0x3b> + DB 102,65,15,110,4,18 ; movd (%r10,%rdx,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,5,13,44,0,0 ; pand 0x2c0d(%rip),%xmm0 # 6070 <_sk_callback_sse2+0xaca> + DB 102,15,219,5,22,44,0,0 ; pand 0x2c16(%rip),%xmm0 # 60a0 <_sk_callback_sse2+0xac6> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,19,44,0,0 ; mulps 0x2c13(%rip),%xmm0 # 6080 <_sk_callback_sse2+0xada> + DB 15,89,5,28,44,0,0 ; mulps 0x2c1c(%rip),%xmm0 # 60b0 <_sk_callback_sse2+0xad6> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,26,44,0,0 ; movaps 0x2c1a(%rip),%xmm3 # 6090 <_sk_callback_sse2+0xaea> + DB 15,40,29,35,44,0,0 ; movaps 0x2c23(%rip),%xmm3 # 60c0 <_sk_callback_sse2+0xae6> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,48 ; je 34bf <_sk_load_g8_sse2+0x7c> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 34aa <_sk_load_g8_sse2+0x67> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,192 ; jne 345b <_sk_load_g8_sse2+0x18> - DB 65,15,182,68,57,2 ; movzbl 0x2(%r9,%rdi,1),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,48 ; je 34e6 <_sk_load_g8_sse2+0x7c> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 34d1 <_sk_load_g8_sse2+0x67> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,192 ; jne 3482 <_sk_load_g8_sse2+0x18> + DB 65,15,182,68,18,2 ; movzbl 0x2(%r10,%rdx,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 - DB 65,15,182,68,57,1 ; movzbl 0x1(%r9,%rdi,1),%eax + DB 65,15,182,68,18,1 ; movzbl 0x1(%r10,%rdx,1),%eax DB 102,15,110,200 ; movd %eax,%xmm1 DB 15,198,200,0 ; shufps $0x0,%xmm0,%xmm1 DB 15,198,200,226 ; shufps $0xe2,%xmm0,%xmm1 DB 15,40,193 ; movaps %xmm1,%xmm0 - DB 65,15,182,4,57 ; movzbl (%r9,%rdi,1),%eax + DB 65,15,182,4,18 ; movzbl (%r10,%rdx,1),%eax DB 102,15,110,200 ; movd %eax,%xmm1 DB 243,15,16,193 ; movss %xmm1,%xmm0 - DB 235,141 ; jmp 345b <_sk_load_g8_sse2+0x18> + DB 235,141 ; jmp 3482 <_sk_load_g8_sse2+0x18> PUBLIC _sk_gather_g8_sse2 _sk_gather_g8_sse2 LABEL PROC + DB 85 ; push %rbp + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22242,45 +22304,48 @@ _sk_gather_g8_sse2 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,72,15,126,192 ; movq %xmm0,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 71,15,182,28,24 ; movzbl (%r8,%r11,1),%r11d - DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 65,193,226,8 ; shl $0x8,%r10d - DB 69,9,218 ; or %r11d,%r10d - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp + DB 67,15,182,28,25 ; movzbl (%r9,%r11,1),%ebx + DB 193,227,8 ; shl $0x8,%ebx + DB 9,235 ; or %ebp,%ebx + DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 193,224,8 ; shl $0x8,%eax - DB 68,9,200 ; or %r9d,%eax + DB 9,232 ; or %ebp,%eax DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 - DB 102,65,15,196,194,1 ; pinsrw $0x1,%r10d,%xmm0 + DB 102,15,196,195,1 ; pinsrw $0x1,%ebx,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,61,43,0,0 ; mulps 0x2b3d(%rip),%xmm0 # 60a0 <_sk_callback_sse2+0xafa> + DB 15,89,5,72,43,0,0 ; mulps 0x2b48(%rip),%xmm0 # 60d0 <_sk_callback_sse2+0xaf6> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,68,43,0,0 ; movaps 0x2b44(%rip),%xmm3 # 60b0 <_sk_callback_sse2+0xb0a> + DB 15,40,29,79,43,0,0 ; movaps 0x2b4f(%rip),%xmm3 # 60e0 <_sk_callback_sse2+0xb06> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 + DB 91 ; pop %rbx + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_gather_i8_sse2 _sk_gather_i8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 73,137,192 ; mov %rax,%r8 - DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 3583 <_sk_gather_i8_sse2+0xf> - DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 3585 <_sk_gather_i8_sse2+0x11> + DB 73,137,193 ; mov %rax,%r9 + DB 77,133,201 ; test %r9,%r9 + DB 116,5 ; je 35aa <_sk_gather_i8_sse2+0xf> + DB 76,137,200 ; mov %r9,%rax + DB 235,2 ; jmp 35ac <_sk_gather_i8_sse2+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 85 ; push %rbp + DB 65,86 ; push %r14 DB 83 ; push %rbx - DB 76,139,8 ; mov (%rax),%r9 + DB 76,139,16 ; mov (%rax),%r10 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22293,18 +22358,18 @@ _sk_gather_i8_sse2 LABEL PROC DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,72,15,126,192 ; movq %xmm0,%rax - DB 65,137,194 ; mov %eax,%r10d + DB 65,137,195 ; mov %eax,%r11d DB 72,193,232,32 ; shr $0x20,%rax DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 - DB 102,73,15,126,195 ; movq %xmm0,%r11 - DB 68,137,219 ; mov %r11d,%ebx - DB 73,193,235,32 ; shr $0x20,%r11 - DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp - DB 67,15,182,28,25 ; movzbl (%r9,%r11,1),%ebx + DB 102,72,15,126,195 ; movq %xmm0,%rbx + DB 65,137,222 ; mov %ebx,%r14d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 67,15,182,44,50 ; movzbl (%r10,%r14,1),%ebp + DB 65,15,182,28,26 ; movzbl (%r10,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx DB 9,235 ; or %ebp,%ebx - DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp - DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 67,15,182,44,26 ; movzbl (%r10,%r11,1),%ebp + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 9,232 ; or %ebp,%eax DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 @@ -22314,24 +22379,24 @@ _sk_gather_i8_sse2 LABEL PROC DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,72,15,126,200 ; movq %xmm1,%rax - DB 68,15,182,200 ; movzbl %al,%r9d + DB 68,15,182,208 ; movzbl %al,%r10d DB 72,193,232,32 ; shr $0x20,%rax DB 102,72,15,126,195 ; movq %xmm0,%rbx - DB 73,139,104,8 ; mov 0x8(%r8),%rbp - DB 68,15,182,195 ; movzbl %bl,%r8d + DB 73,139,105,8 ; mov 0x8(%r9),%rbp + DB 68,15,182,203 ; movzbl %bl,%r9d DB 72,193,235,30 ; shr $0x1e,%rbx DB 102,15,110,68,29,0 ; movd 0x0(%rbp,%rbx,1),%xmm0 DB 102,15,110,76,133,0 ; movd 0x0(%rbp,%rax,4),%xmm1 DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,70,15,110,76,133,0 ; movd 0x0(%rbp,%r8,4),%xmm9 - DB 102,66,15,110,76,141,0 ; movd 0x0(%rbp,%r9,4),%xmm1 + DB 102,70,15,110,76,141,0 ; movd 0x0(%rbp,%r9,4),%xmm9 + DB 102,66,15,110,76,149,0 ; movd 0x0(%rbp,%r10,4),%xmm1 DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,97,42,0,0 ; movdqa 0x2a61(%rip),%xmm2 # 60c0 <_sk_callback_sse2+0xb1a> + DB 102,15,111,21,104,42,0,0 ; movdqa 0x2a68(%rip),%xmm2 # 60f0 <_sk_callback_sse2+0xb16> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,93,42,0,0 ; movaps 0x2a5d(%rip),%xmm8 # 60d0 <_sk_callback_sse2+0xb2a> + DB 68,15,40,5,100,42,0,0 ; movaps 0x2a64(%rip),%xmm8 # 6100 <_sk_callback_sse2+0xb26> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -22348,57 +22413,59 @@ _sk_gather_i8_sse2 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx + DB 65,94 ; pop %r14 DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_load_565_sse2 _sk_load_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,83 ; jne 3712 <_sk_load_565_sse2+0x5d> - DB 243,65,15,126,20,121 ; movq (%r9,%rdi,2),%xmm2 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,83 ; jne 373d <_sk_load_565_sse2+0x5d> + DB 243,65,15,126,20,82 ; movq (%r10,%rdx,2),%xmm2 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,15,42,0,0 ; movdqa 0x2a0f(%rip),%xmm0 # 60e0 <_sk_callback_sse2+0xb3a> + DB 102,15,111,5,20,42,0,0 ; movdqa 0x2a14(%rip),%xmm0 # 6110 <_sk_callback_sse2+0xb36> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,17,42,0,0 ; mulps 0x2a11(%rip),%xmm0 # 60f0 <_sk_callback_sse2+0xb4a> - DB 102,15,111,13,25,42,0,0 ; movdqa 0x2a19(%rip),%xmm1 # 6100 <_sk_callback_sse2+0xb5a> + DB 15,89,5,22,42,0,0 ; mulps 0x2a16(%rip),%xmm0 # 6120 <_sk_callback_sse2+0xb46> + DB 102,15,111,13,30,42,0,0 ; movdqa 0x2a1e(%rip),%xmm1 # 6130 <_sk_callback_sse2+0xb56> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,27,42,0,0 ; mulps 0x2a1b(%rip),%xmm1 # 6110 <_sk_callback_sse2+0xb6a> - DB 102,15,219,21,35,42,0,0 ; pand 0x2a23(%rip),%xmm2 # 6120 <_sk_callback_sse2+0xb7a> + DB 15,89,13,32,42,0,0 ; mulps 0x2a20(%rip),%xmm1 # 6140 <_sk_callback_sse2+0xb66> + DB 102,15,219,21,40,42,0,0 ; pand 0x2a28(%rip),%xmm2 # 6150 <_sk_callback_sse2+0xb76> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,41,42,0,0 ; mulps 0x2a29(%rip),%xmm2 # 6130 <_sk_callback_sse2+0xb8a> + DB 15,89,21,46,42,0,0 ; mulps 0x2a2e(%rip),%xmm2 # 6160 <_sk_callback_sse2+0xb86> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,48,42,0,0 ; movaps 0x2a30(%rip),%xmm3 # 6140 <_sk_callback_sse2+0xb9a> + DB 15,40,29,53,42,0,0 ; movaps 0x2a35(%rip),%xmm3 # 6170 <_sk_callback_sse2+0xb96> DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,48 ; je 3753 <_sk_load_565_sse2+0x9e> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 373e <_sk_load_565_sse2+0x89> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,154 ; jne 36c9 <_sk_load_565_sse2+0x14> - DB 65,15,183,68,121,4 ; movzwl 0x4(%r9,%rdi,2),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,48 ; je 377e <_sk_load_565_sse2+0x9e> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 3769 <_sk_load_565_sse2+0x89> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,154 ; jne 36f4 <_sk_load_565_sse2+0x14> + DB 65,15,183,68,82,4 ; movzwl 0x4(%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,208,69 ; pshufd $0x45,%xmm0,%xmm2 - DB 65,15,183,68,121,2 ; movzwl 0x2(%r9,%rdi,2),%eax + DB 65,15,183,68,82,2 ; movzwl 0x2(%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 15,198,194,0 ; shufps $0x0,%xmm2,%xmm0 DB 15,198,194,226 ; shufps $0xe2,%xmm2,%xmm0 DB 15,40,208 ; movaps %xmm0,%xmm2 - DB 65,15,183,4,121 ; movzwl (%r9,%rdi,2),%eax + DB 65,15,183,4,82 ; movzwl (%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 243,15,16,208 ; movss %xmm0,%xmm2 - DB 233,100,255,255,255 ; jmpq 36c9 <_sk_load_565_sse2+0x14> + DB 233,100,255,255,255 ; jmpq 36f4 <_sk_load_565_sse2+0x14> PUBLIC _sk_gather_565_sse2 _sk_gather_565_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22412,44 +22479,45 @@ _sk_gather_565_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,72,15,126,200 ; movq %xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,20,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm2 - DB 102,67,15,196,20,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm2 - DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d - DB 102,65,15,196,209,2 ; pinsrw $0x2,%r9d,%xmm2 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,65,15,196,20,89,0 ; pinsrw $0x0,(%r9,%rbx,2),%xmm2 + DB 102,67,15,196,20,89,1 ; pinsrw $0x1,(%r9,%r11,2),%xmm2 + DB 67,15,183,28,81 ; movzwl (%r9,%r10,2),%ebx + DB 102,15,196,211,2 ; pinsrw $0x2,%ebx,%xmm2 + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 102,15,196,208,3 ; pinsrw $0x3,%eax,%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,101,41,0,0 ; movdqa 0x2965(%rip),%xmm0 # 6150 <_sk_callback_sse2+0xbaa> + DB 102,15,111,5,106,41,0,0 ; movdqa 0x296a(%rip),%xmm0 # 6180 <_sk_callback_sse2+0xba6> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,103,41,0,0 ; mulps 0x2967(%rip),%xmm0 # 6160 <_sk_callback_sse2+0xbba> - DB 102,15,111,13,111,41,0,0 ; movdqa 0x296f(%rip),%xmm1 # 6170 <_sk_callback_sse2+0xbca> + DB 15,89,5,108,41,0,0 ; mulps 0x296c(%rip),%xmm0 # 6190 <_sk_callback_sse2+0xbb6> + DB 102,15,111,13,116,41,0,0 ; movdqa 0x2974(%rip),%xmm1 # 61a0 <_sk_callback_sse2+0xbc6> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,113,41,0,0 ; mulps 0x2971(%rip),%xmm1 # 6180 <_sk_callback_sse2+0xbda> - DB 102,15,219,21,121,41,0,0 ; pand 0x2979(%rip),%xmm2 # 6190 <_sk_callback_sse2+0xbea> + DB 15,89,13,118,41,0,0 ; mulps 0x2976(%rip),%xmm1 # 61b0 <_sk_callback_sse2+0xbd6> + DB 102,15,219,21,126,41,0,0 ; pand 0x297e(%rip),%xmm2 # 61c0 <_sk_callback_sse2+0xbe6> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,127,41,0,0 ; mulps 0x297f(%rip),%xmm2 # 61a0 <_sk_callback_sse2+0xbfa> + DB 15,89,21,132,41,0,0 ; mulps 0x2984(%rip),%xmm2 # 61d0 <_sk_callback_sse2+0xbf6> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,134,41,0,0 ; movaps 0x2986(%rip),%xmm3 # 61b0 <_sk_callback_sse2+0xc0a> + DB 15,40,29,139,41,0,0 ; movaps 0x298b(%rip),%xmm3 # 61e0 <_sk_callback_sse2+0xc06> + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_sse2 _sk_store_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 68,15,40,5,135,41,0,0 ; movaps 0x2987(%rip),%xmm8 # 61c0 <_sk_callback_sse2+0xc1a> + DB 76,139,16 ; mov (%rax),%r10 + DB 68,15,40,5,139,41,0,0 ; movaps 0x298b(%rip),%xmm8 # 61f0 <_sk_callback_sse2+0xc16> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 DB 102,65,15,114,241,11 ; pslld $0xb,%xmm9 - DB 68,15,40,21,124,41,0,0 ; movaps 0x297c(%rip),%xmm10 # 61d0 <_sk_callback_sse2+0xc2a> + DB 68,15,40,21,128,41,0,0 ; movaps 0x2980(%rip),%xmm10 # 6200 <_sk_callback_sse2+0xc26> DB 68,15,89,209 ; mulps %xmm1,%xmm10 DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10 DB 102,65,15,114,242,5 ; pslld $0x5,%xmm10 @@ -22460,79 +22528,80 @@ _sk_store_565_sse2 LABEL PROC DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3896 <_sk_store_565_sse2+0x6a> - DB 242,69,15,17,4,121 ; movsd %xmm8,(%r9,%rdi,2) + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 38c2 <_sk_store_565_sse2+0x6a> + DB 242,69,15,17,4,82 ; movsd %xmm8,(%r10,%rdx,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,36 ; je 38cc <_sk_store_565_sse2+0xa0> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,18 ; je 38c0 <_sk_store_565_sse2+0x94> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,222 ; jne 3892 <_sk_store_565_sse2+0x66> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,36 ; je 38f8 <_sk_store_565_sse2+0xa0> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,18 ; je 38ec <_sk_store_565_sse2+0x94> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,222 ; jne 38be <_sk_store_565_sse2+0x66> DB 102,65,15,197,192,4 ; pextrw $0x4,%xmm8,%eax - DB 102,65,137,68,121,4 ; mov %ax,0x4(%r9,%rdi,2) + DB 102,65,137,68,82,4 ; mov %ax,0x4(%r10,%rdx,2) DB 102,65,15,197,192,2 ; pextrw $0x2,%xmm8,%eax - DB 102,65,137,68,121,2 ; mov %ax,0x2(%r9,%rdi,2) + DB 102,65,137,68,82,2 ; mov %ax,0x2(%r10,%rdx,2) DB 102,68,15,126,192 ; movd %xmm8,%eax - DB 102,65,137,4,121 ; mov %ax,(%r9,%rdi,2) - DB 235,186 ; jmp 3892 <_sk_store_565_sse2+0x66> + DB 102,65,137,4,82 ; mov %ax,(%r10,%rdx,2) + DB 235,186 ; jmp 38be <_sk_store_565_sse2+0x66> PUBLIC _sk_load_4444_sse2 _sk_load_4444_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,98 ; jne 3944 <_sk_load_4444_sse2+0x6c> - DB 243,65,15,126,28,121 ; movq (%r9,%rdi,2),%xmm3 + DB 76,139,16 ; mov (%rax),%r10 + DB 77,133,192 ; test %r8,%r8 + DB 117,98 ; jne 3970 <_sk_load_4444_sse2+0x6c> + DB 243,65,15,126,28,82 ; movq (%r10,%rdx,2),%xmm3 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,236,40,0,0 ; movdqa 0x28ec(%rip),%xmm0 # 61e0 <_sk_callback_sse2+0xc3a> + DB 102,15,111,5,240,40,0,0 ; movdqa 0x28f0(%rip),%xmm0 # 6210 <_sk_callback_sse2+0xc36> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,238,40,0,0 ; mulps 0x28ee(%rip),%xmm0 # 61f0 <_sk_callback_sse2+0xc4a> - DB 102,15,111,13,246,40,0,0 ; movdqa 0x28f6(%rip),%xmm1 # 6200 <_sk_callback_sse2+0xc5a> + DB 15,89,5,242,40,0,0 ; mulps 0x28f2(%rip),%xmm0 # 6220 <_sk_callback_sse2+0xc46> + DB 102,15,111,13,250,40,0,0 ; movdqa 0x28fa(%rip),%xmm1 # 6230 <_sk_callback_sse2+0xc56> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,248,40,0,0 ; mulps 0x28f8(%rip),%xmm1 # 6210 <_sk_callback_sse2+0xc6a> - DB 102,15,111,21,0,41,0,0 ; movdqa 0x2900(%rip),%xmm2 # 6220 <_sk_callback_sse2+0xc7a> + DB 15,89,13,252,40,0,0 ; mulps 0x28fc(%rip),%xmm1 # 6240 <_sk_callback_sse2+0xc66> + DB 102,15,111,21,4,41,0,0 ; movdqa 0x2904(%rip),%xmm2 # 6250 <_sk_callback_sse2+0xc76> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,2,41,0,0 ; mulps 0x2902(%rip),%xmm2 # 6230 <_sk_callback_sse2+0xc8a> - DB 102,15,219,29,10,41,0,0 ; pand 0x290a(%rip),%xmm3 # 6240 <_sk_callback_sse2+0xc9a> + DB 15,89,21,6,41,0,0 ; mulps 0x2906(%rip),%xmm2 # 6260 <_sk_callback_sse2+0xc86> + DB 102,15,219,29,14,41,0,0 ; pand 0x290e(%rip),%xmm3 # 6270 <_sk_callback_sse2+0xc96> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,16,41,0,0 ; mulps 0x2910(%rip),%xmm3 # 6250 <_sk_callback_sse2+0xcaa> + DB 15,89,29,20,41,0,0 ; mulps 0x2914(%rip),%xmm3 # 6280 <_sk_callback_sse2+0xca6> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,48 ; je 3985 <_sk_load_4444_sse2+0xad> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,21 ; je 3970 <_sk_load_4444_sse2+0x98> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,139 ; jne 38ec <_sk_load_4444_sse2+0x14> - DB 65,15,183,68,121,4 ; movzwl 0x4(%r9,%rdi,2),%eax + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,48 ; je 39b1 <_sk_load_4444_sse2+0xad> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,21 ; je 399c <_sk_load_4444_sse2+0x98> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,139 ; jne 3918 <_sk_load_4444_sse2+0x14> + DB 65,15,183,68,82,4 ; movzwl 0x4(%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 - DB 65,15,183,68,121,2 ; movzwl 0x2(%r9,%rdi,2),%eax + DB 65,15,183,68,82,2 ; movzwl 0x2(%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 15,198,195,0 ; shufps $0x0,%xmm3,%xmm0 DB 15,198,195,226 ; shufps $0xe2,%xmm3,%xmm0 DB 15,40,216 ; movaps %xmm0,%xmm3 - DB 65,15,183,4,121 ; movzwl (%r9,%rdi,2),%eax + DB 65,15,183,4,82 ; movzwl (%r10,%rdx,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 243,15,16,216 ; movss %xmm0,%xmm3 - DB 233,85,255,255,255 ; jmpq 38ec <_sk_load_4444_sse2+0x14> + DB 233,85,255,255,255 ; jmpq 3918 <_sk_load_4444_sse2+0x14> PUBLIC _sk_gather_4444_sse2 _sk_gather_4444_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22546,42 +22615,43 @@ _sk_gather_4444_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,72,15,126,200 ; movq %xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,28,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm3 - DB 102,67,15,196,28,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm3 - DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d - DB 102,65,15,196,217,2 ; pinsrw $0x2,%r9d,%xmm3 - DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,65,15,196,28,89,0 ; pinsrw $0x0,(%r9,%rbx,2),%xmm3 + DB 102,67,15,196,28,89,1 ; pinsrw $0x1,(%r9,%r11,2),%xmm3 + DB 67,15,183,28,81 ; movzwl (%r9,%r10,2),%ebx + DB 102,15,196,219,2 ; pinsrw $0x2,%ebx,%xmm3 + DB 65,15,183,4,65 ; movzwl (%r9,%rax,2),%eax DB 102,15,196,216,3 ; pinsrw $0x3,%eax,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,67,40,0,0 ; movdqa 0x2843(%rip),%xmm0 # 6260 <_sk_callback_sse2+0xcba> + DB 102,15,111,5,71,40,0,0 ; movdqa 0x2847(%rip),%xmm0 # 6290 <_sk_callback_sse2+0xcb6> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,69,40,0,0 ; mulps 0x2845(%rip),%xmm0 # 6270 <_sk_callback_sse2+0xcca> - DB 102,15,111,13,77,40,0,0 ; movdqa 0x284d(%rip),%xmm1 # 6280 <_sk_callback_sse2+0xcda> + DB 15,89,5,73,40,0,0 ; mulps 0x2849(%rip),%xmm0 # 62a0 <_sk_callback_sse2+0xcc6> + DB 102,15,111,13,81,40,0,0 ; movdqa 0x2851(%rip),%xmm1 # 62b0 <_sk_callback_sse2+0xcd6> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,79,40,0,0 ; mulps 0x284f(%rip),%xmm1 # 6290 <_sk_callback_sse2+0xcea> - DB 102,15,111,21,87,40,0,0 ; movdqa 0x2857(%rip),%xmm2 # 62a0 <_sk_callback_sse2+0xcfa> + DB 15,89,13,83,40,0,0 ; mulps 0x2853(%rip),%xmm1 # 62c0 <_sk_callback_sse2+0xce6> + DB 102,15,111,21,91,40,0,0 ; movdqa 0x285b(%rip),%xmm2 # 62d0 <_sk_callback_sse2+0xcf6> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,89,40,0,0 ; mulps 0x2859(%rip),%xmm2 # 62b0 <_sk_callback_sse2+0xd0a> - DB 102,15,219,29,97,40,0,0 ; pand 0x2861(%rip),%xmm3 # 62c0 <_sk_callback_sse2+0xd1a> + DB 15,89,21,93,40,0,0 ; mulps 0x285d(%rip),%xmm2 # 62e0 <_sk_callback_sse2+0xd06> + DB 102,15,219,29,101,40,0,0 ; pand 0x2865(%rip),%xmm3 # 62f0 <_sk_callback_sse2+0xd16> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,103,40,0,0 ; mulps 0x2867(%rip),%xmm3 # 62d0 <_sk_callback_sse2+0xd2a> + DB 15,89,29,107,40,0,0 ; mulps 0x286b(%rip),%xmm3 # 6300 <_sk_callback_sse2+0xd26> DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_4444_sse2 _sk_store_4444_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 68,15,40,5,102,40,0,0 ; movaps 0x2866(%rip),%xmm8 # 62e0 <_sk_callback_sse2+0xd3a> + DB 76,139,16 ; mov (%rax),%r10 + DB 68,15,40,5,105,40,0,0 ; movaps 0x2869(%rip),%xmm8 # 6310 <_sk_callback_sse2+0xd36> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -22602,40 +22672,40 @@ _sk_store_4444_sse2 LABEL PROC DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3aeb <_sk_store_4444_sse2+0x7e> - DB 242,69,15,17,4,121 ; movsd %xmm8,(%r9,%rdi,2) + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3b18 <_sk_store_4444_sse2+0x7e> + DB 242,69,15,17,4,82 ; movsd %xmm8,(%r10,%rdx,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,36 ; je 3b21 <_sk_store_4444_sse2+0xb4> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,18 ; je 3b15 <_sk_store_4444_sse2+0xa8> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,222 ; jne 3ae7 <_sk_store_4444_sse2+0x7a> + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,36 ; je 3b4e <_sk_store_4444_sse2+0xb4> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,18 ; je 3b42 <_sk_store_4444_sse2+0xa8> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,222 ; jne 3b14 <_sk_store_4444_sse2+0x7a> DB 102,65,15,197,192,4 ; pextrw $0x4,%xmm8,%eax - DB 102,65,137,68,121,4 ; mov %ax,0x4(%r9,%rdi,2) + DB 102,65,137,68,82,4 ; mov %ax,0x4(%r10,%rdx,2) DB 102,65,15,197,192,2 ; pextrw $0x2,%xmm8,%eax - DB 102,65,137,68,121,2 ; mov %ax,0x2(%r9,%rdi,2) + DB 102,65,137,68,82,2 ; mov %ax,0x2(%r10,%rdx,2) DB 102,68,15,126,192 ; movd %xmm8,%eax - DB 102,65,137,4,121 ; mov %ax,(%r9,%rdi,2) - DB 235,186 ; jmp 3ae7 <_sk_store_4444_sse2+0x7a> + DB 102,65,137,4,82 ; mov %ax,(%r10,%rdx,2) + DB 235,186 ; jmp 3b14 <_sk_store_4444_sse2+0x7a> PUBLIC _sk_load_8888_sse2 _sk_load_8888_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 117,98 ; jne 3b99 <_sk_load_8888_sse2+0x6c> - DB 243,68,15,111,12,184 ; movdqu (%rax,%rdi,4),%xmm9 - DB 102,15,111,21,171,39,0,0 ; movdqa 0x27ab(%rip),%xmm2 # 62f0 <_sk_callback_sse2+0xd4a> + DB 77,133,192 ; test %r8,%r8 + DB 117,98 ; jne 3bc6 <_sk_load_8888_sse2+0x6c> + DB 243,68,15,111,12,144 ; movdqu (%rax,%rdx,4),%xmm9 + DB 102,15,111,21,174,39,0,0 ; movdqa 0x27ae(%rip),%xmm2 # 6320 <_sk_callback_sse2+0xd46> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,167,39,0,0 ; movaps 0x27a7(%rip),%xmm8 # 6300 <_sk_callback_sse2+0xd5a> + DB 68,15,40,5,170,39,0,0 ; movaps 0x27aa(%rip),%xmm8 # 6330 <_sk_callback_sse2+0xd56> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -22652,29 +22722,30 @@ _sk_load_8888_sse2 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,44 ; je 3bd7 <_sk_load_8888_sse2+0xaa> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,18 ; je 3bc3 <_sk_load_8888_sse2+0x96> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,134 ; jne 3b3d <_sk_load_8888_sse2+0x10> - DB 102,15,110,68,184,8 ; movd 0x8(%rax,%rdi,4),%xmm0 + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,44 ; je 3c04 <_sk_load_8888_sse2+0xaa> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,18 ; je 3bf0 <_sk_load_8888_sse2+0x96> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,134 ; jne 3b6a <_sk_load_8888_sse2+0x10> + DB 102,15,110,68,144,8 ; movd 0x8(%rax,%rdx,4),%xmm0 DB 102,68,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm9 - DB 243,15,16,68,184,4 ; movss 0x4(%rax,%rdi,4),%xmm0 + DB 243,15,16,68,144,4 ; movss 0x4(%rax,%rdx,4),%xmm0 DB 65,15,198,193,0 ; shufps $0x0,%xmm9,%xmm0 DB 65,15,198,193,226 ; shufps $0xe2,%xmm9,%xmm0 DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 243,15,16,4,184 ; movss (%rax,%rdi,4),%xmm0 + DB 243,15,16,4,144 ; movss (%rax,%rdx,4),%xmm0 DB 243,68,15,16,200 ; movss %xmm0,%xmm9 - DB 233,87,255,255,255 ; jmpq 3b3d <_sk_load_8888_sse2+0x10> + DB 233,87,255,255,255 ; jmpq 3b6a <_sk_load_8888_sse2+0x10> PUBLIC _sk_gather_8888_sse2 _sk_gather_8888_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22688,23 +22759,23 @@ _sk_gather_8888_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,72,15,126,200 ; movq %xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,110,4,144 ; movd (%r8,%r10,4),%xmm0 - DB 102,65,15,110,12,128 ; movd (%r8,%rax,4),%xmm1 + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 102,67,15,110,4,153 ; movd (%r9,%r11,4),%xmm0 + DB 102,65,15,110,12,129 ; movd (%r9,%rax,4),%xmm1 DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,71,15,110,12,152 ; movd (%r8,%r11,4),%xmm9 - DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 + DB 102,69,15,110,12,153 ; movd (%r9,%rbx,4),%xmm9 + DB 102,67,15,110,12,145 ; movd (%r9,%r10,4),%xmm1 DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,169,38,0,0 ; movdqa 0x26a9(%rip),%xmm2 # 6310 <_sk_callback_sse2+0xd6a> + DB 102,15,111,21,171,38,0,0 ; movdqa 0x26ab(%rip),%xmm2 # 6340 <_sk_callback_sse2+0xd66> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,165,38,0,0 ; movaps 0x26a5(%rip),%xmm8 # 6320 <_sk_callback_sse2+0xd7a> + DB 68,15,40,5,167,38,0,0 ; movaps 0x26a7(%rip),%xmm8 # 6350 <_sk_callback_sse2+0xd76> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -22720,13 +22791,14 @@ _sk_gather_8888_sse2 LABEL PROC DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3 DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_8888_sse2 _sk_store_8888_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,104,38,0,0 ; movaps 0x2668(%rip),%xmm8 # 6330 <_sk_callback_sse2+0xd8a> + DB 68,15,40,5,105,38,0,0 ; movaps 0x2669(%rip),%xmm8 # 6360 <_sk_callback_sse2+0xd86> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -22744,34 +22816,34 @@ _sk_store_8888_sse2 LABEL PROC DB 102,65,15,114,240,24 ; pslld $0x18,%xmm8 DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 102,69,15,235,194 ; por %xmm10,%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3d28 <_sk_store_8888_sse2+0x6d> - DB 243,68,15,127,4,184 ; movdqu %xmm8,(%rax,%rdi,4) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,3 ; and $0x3,%r8b - DB 65,128,248,1 ; cmp $0x1,%r8b - DB 116,38 ; je 3d5b <_sk_store_8888_sse2+0xa0> - DB 65,128,248,2 ; cmp $0x2,%r8b - DB 116,19 ; je 3d4e <_sk_store_8888_sse2+0x93> - DB 65,128,248,3 ; cmp $0x3,%r8b - DB 117,227 ; jne 3d24 <_sk_store_8888_sse2+0x69> + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 3d57 <_sk_store_8888_sse2+0x6d> + DB 243,68,15,127,4,144 ; movdqu %xmm8,(%rax,%rdx,4) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,3 ; and $0x3,%r9b + DB 65,128,249,1 ; cmp $0x1,%r9b + DB 116,38 ; je 3d8a <_sk_store_8888_sse2+0xa0> + DB 65,128,249,2 ; cmp $0x2,%r9b + DB 116,19 ; je 3d7d <_sk_store_8888_sse2+0x93> + DB 65,128,249,3 ; cmp $0x3,%r9b + DB 117,227 ; jne 3d53 <_sk_store_8888_sse2+0x69> DB 102,69,15,112,200,78 ; pshufd $0x4e,%xmm8,%xmm9 - DB 102,68,15,126,76,184,8 ; movd %xmm9,0x8(%rax,%rdi,4) + DB 102,68,15,126,76,144,8 ; movd %xmm9,0x8(%rax,%rdx,4) DB 102,69,15,112,200,229 ; pshufd $0xe5,%xmm8,%xmm9 - DB 102,68,15,126,76,184,4 ; movd %xmm9,0x4(%rax,%rdi,4) - DB 102,68,15,126,4,184 ; movd %xmm8,(%rax,%rdi,4) - DB 235,193 ; jmp 3d24 <_sk_store_8888_sse2+0x69> + DB 102,68,15,126,76,144,4 ; movd %xmm9,0x4(%rax,%rdx,4) + DB 102,68,15,126,4,144 ; movd %xmm8,(%rax,%rdx,4) + DB 235,193 ; jmp 3d53 <_sk_store_8888_sse2+0x69> PUBLIC _sk_load_f16_sse2 _sk_load_f16_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,96,1,0,0 ; jne 3ed1 <_sk_load_f16_sse2+0x16e> - DB 102,15,16,4,248 ; movupd (%rax,%rdi,8),%xmm0 - DB 102,15,16,76,248,16 ; movupd 0x10(%rax,%rdi,8),%xmm1 + DB 77,133,192 ; test %r8,%r8 + DB 15,133,96,1,0,0 ; jne 3f00 <_sk_load_f16_sse2+0x16e> + DB 102,15,16,4,208 ; movupd (%rax,%rdx,8),%xmm0 + DB 102,15,16,76,208,16 ; movupd 0x10(%rax,%rdx,8),%xmm1 DB 102,68,15,40,192 ; movapd %xmm0,%xmm8 DB 102,68,15,97,193 ; punpcklwd %xmm1,%xmm8 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 @@ -22781,7 +22853,7 @@ _sk_load_f16_sse2 LABEL PROC DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 - DB 102,68,15,111,13,143,37,0,0 ; movdqa 0x258f(%rip),%xmm9 # 6340 <_sk_callback_sse2+0xd9a> + DB 102,68,15,111,13,144,37,0,0 ; movdqa 0x2590(%rip),%xmm9 # 6370 <_sk_callback_sse2+0xd96> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,193 ; pand %xmm9,%xmm0 DB 102,15,239,200 ; pxor %xmm0,%xmm1 @@ -22789,11 +22861,11 @@ _sk_load_f16_sse2 LABEL PROC DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 DB 102,68,15,235,232 ; por %xmm0,%xmm13 - DB 102,68,15,111,29,116,37,0,0 ; movdqa 0x2574(%rip),%xmm11 # 6350 <_sk_callback_sse2+0xdaa> + DB 102,68,15,111,29,117,37,0,0 ; movdqa 0x2575(%rip),%xmm11 # 6380 <_sk_callback_sse2+0xda6> DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,118,37,0,0 ; movdqa 0x2576(%rip),%xmm12 # 6360 <_sk_callback_sse2+0xdba> + DB 102,68,15,111,37,119,37,0,0 ; movdqa 0x2577(%rip),%xmm12 # 6390 <_sk_callback_sse2+0xdb6> DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,121,37,0,0 ; movdqa 0x2579(%rip),%xmm3 # 6370 <_sk_callback_sse2+0xdca> + DB 102,15,111,29,122,37,0,0 ; movdqa 0x257a(%rip),%xmm3 # 63a0 <_sk_callback_sse2+0xdc6> DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0 DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 @@ -22839,24 +22911,25 @@ _sk_load_f16_sse2 LABEL PROC DB 102,65,15,223,218 ; pandn %xmm10,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 242,15,16,4,248 ; movsd (%rax,%rdi,8),%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,17 ; jne 3eed <_sk_load_f16_sse2+0x18a> + DB 242,15,16,4,208 ; movsd (%rax,%rdx,8),%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,17 ; jne 3f1c <_sk_load_f16_sse2+0x18a> DB 102,15,87,201 ; xorpd %xmm1,%xmm1 DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 233,143,254,255,255 ; jmpq 3d7c <_sk_load_f16_sse2+0x19> - DB 102,15,22,68,248,8 ; movhpd 0x8(%rax,%rdi,8),%xmm0 + DB 233,143,254,255,255 ; jmpq 3dab <_sk_load_f16_sse2+0x19> + DB 102,15,22,68,208,8 ; movhpd 0x8(%rax,%rdx,8),%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 15,130,123,254,255,255 ; jb 3d7c <_sk_load_f16_sse2+0x19> - DB 242,15,16,76,248,16 ; movsd 0x10(%rax,%rdi,8),%xmm1 - DB 233,112,254,255,255 ; jmpq 3d7c <_sk_load_f16_sse2+0x19> + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 15,130,123,254,255,255 ; jb 3dab <_sk_load_f16_sse2+0x19> + DB 242,15,16,76,208,16 ; movsd 0x10(%rax,%rdx,8),%xmm1 + DB 233,112,254,255,255 ; jmpq 3dab <_sk_load_f16_sse2+0x19> PUBLIC _sk_gather_f16_sse2 _sk_gather_f16_sse2 LABEL PROC + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 @@ -22870,16 +22943,16 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,72,15,126,200 ; movq %xmm1,%rax - DB 65,137,193 ; mov %eax,%r9d + DB 65,137,194 ; mov %eax,%r10d DB 72,193,232,32 ; shr $0x20,%rax - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 243,67,15,126,4,208 ; movq (%r8,%r10,8),%xmm0 - DB 243,67,15,126,12,216 ; movq (%r8,%r11,8),%xmm1 + DB 102,73,15,126,195 ; movq %xmm0,%r11 + DB 68,137,219 ; mov %r11d,%ebx + DB 73,193,235,32 ; shr $0x20,%r11 + DB 243,67,15,126,4,217 ; movq (%r9,%r11,8),%xmm0 + DB 243,65,15,126,12,217 ; movq (%r9,%rbx,8),%xmm1 DB 102,15,108,200 ; punpcklqdq %xmm0,%xmm1 - DB 243,65,15,126,4,192 ; movq (%r8,%rax,8),%xmm0 - DB 243,67,15,126,20,200 ; movq (%r8,%r9,8),%xmm2 + DB 243,65,15,126,4,193 ; movq (%r9,%rax,8),%xmm0 + DB 243,67,15,126,20,209 ; movq (%r9,%r10,8),%xmm2 DB 102,15,108,208 ; punpcklqdq %xmm0,%xmm2 DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 DB 102,68,15,97,194 ; punpcklwd %xmm2,%xmm8 @@ -22890,7 +22963,7 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 - DB 102,68,15,111,13,204,35,0,0 ; movdqa 0x23cc(%rip),%xmm9 # 6380 <_sk_callback_sse2+0xdda> + DB 102,68,15,111,13,204,35,0,0 ; movdqa 0x23cc(%rip),%xmm9 # 63b0 <_sk_callback_sse2+0xdd6> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,193 ; pand %xmm9,%xmm0 DB 102,15,239,200 ; pxor %xmm0,%xmm1 @@ -22898,11 +22971,11 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 DB 102,68,15,235,232 ; por %xmm0,%xmm13 - DB 102,68,15,111,29,177,35,0,0 ; movdqa 0x23b1(%rip),%xmm11 # 6390 <_sk_callback_sse2+0xdea> + DB 102,68,15,111,29,177,35,0,0 ; movdqa 0x23b1(%rip),%xmm11 # 63c0 <_sk_callback_sse2+0xde6> DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,179,35,0,0 ; movdqa 0x23b3(%rip),%xmm12 # 63a0 <_sk_callback_sse2+0xdfa> + DB 102,68,15,111,37,179,35,0,0 ; movdqa 0x23b3(%rip),%xmm12 # 63d0 <_sk_callback_sse2+0xdf6> DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,182,35,0,0 ; movdqa 0x23b6(%rip),%xmm3 # 63b0 <_sk_callback_sse2+0xe0a> + DB 102,15,111,29,182,35,0,0 ; movdqa 0x23b6(%rip),%xmm3 # 63e0 <_sk_callback_sse2+0xe06> DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0 DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 @@ -22947,23 +23020,24 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,65,15,102,216 ; pcmpgtd %xmm8,%xmm3 DB 102,65,15,223,218 ; pandn %xmm10,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_store_f16_sse2 _sk_store_f16_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 102,68,15,111,21,222,34,0,0 ; movdqa 0x22de(%rip),%xmm10 # 63c0 <_sk_callback_sse2+0xe1a> + DB 102,68,15,111,21,221,34,0,0 ; movdqa 0x22dd(%rip),%xmm10 # 63f0 <_sk_callback_sse2+0xe16> DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 DB 102,69,15,219,226 ; pand %xmm10,%xmm12 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 - DB 102,68,15,111,13,209,34,0,0 ; movdqa 0x22d1(%rip),%xmm9 # 63d0 <_sk_callback_sse2+0xe2a> + DB 102,68,15,111,13,208,34,0,0 ; movdqa 0x22d0(%rip),%xmm9 # 6400 <_sk_callback_sse2+0xe26> DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 102,68,15,111,29,194,34,0,0 ; movdqa 0x22c2(%rip),%xmm11 # 63e0 <_sk_callback_sse2+0xe3a> + DB 102,68,15,111,29,193,34,0,0 ; movdqa 0x22c1(%rip),%xmm11 # 6410 <_sk_callback_sse2+0xe36> DB 102,69,15,235,227 ; por %xmm11,%xmm12 DB 102,69,15,254,229 ; paddd %xmm13,%xmm12 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 @@ -23014,32 +23088,32 @@ _sk_store_f16_sse2 LABEL PROC DB 102,69,15,97,225 ; punpcklwd %xmm9,%xmm12 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,69,15,98,204 ; punpckldq %xmm12,%xmm9 - DB 72,133,201 ; test %rcx,%rcx - DB 117,21 ; jne 4240 <_sk_store_f16_sse2+0x16c> - DB 68,15,17,12,248 ; movups %xmm9,(%rax,%rdi,8) + DB 77,133,192 ; test %r8,%r8 + DB 117,21 ; jne 4271 <_sk_store_f16_sse2+0x16c> + DB 68,15,17,12,208 ; movups %xmm9,(%rax,%rdx,8) DB 102,69,15,106,196 ; punpckhdq %xmm12,%xmm8 - DB 243,68,15,127,68,248,16 ; movdqu %xmm8,0x10(%rax,%rdi,8) + DB 243,68,15,127,68,208,16 ; movdqu %xmm8,0x10(%rax,%rdx,8) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 102,68,15,214,12,248 ; movq %xmm9,(%rax,%rdi,8) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 423c <_sk_store_f16_sse2+0x168> - DB 102,68,15,23,76,248,8 ; movhpd %xmm9,0x8(%rax,%rdi,8) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 423c <_sk_store_f16_sse2+0x168> + DB 102,68,15,214,12,208 ; movq %xmm9,(%rax,%rdx,8) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 426d <_sk_store_f16_sse2+0x168> + DB 102,68,15,23,76,208,8 ; movhpd %xmm9,0x8(%rax,%rdx,8) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 426d <_sk_store_f16_sse2+0x168> DB 102,69,15,106,196 ; punpckhdq %xmm12,%xmm8 - DB 102,68,15,214,68,248,16 ; movq %xmm8,0x10(%rax,%rdi,8) - DB 235,213 ; jmp 423c <_sk_store_f16_sse2+0x168> + DB 102,68,15,214,68,208,16 ; movq %xmm8,0x10(%rax,%rdx,8) + DB 235,213 ; jmp 426d <_sk_store_f16_sse2+0x168> PUBLIC _sk_load_u16_be_sse2 _sk_load_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,190,0,0,0 ; jne 433b <_sk_load_u16_be_sse2+0xd4> - DB 102,65,15,16,4,64 ; movupd (%r8,%rax,2),%xmm0 - DB 102,65,15,16,76,64,16 ; movupd 0x10(%r8,%rax,2),%xmm1 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,190,0,0,0 ; jne 436c <_sk_load_u16_be_sse2+0xd4> + DB 102,65,15,16,4,65 ; movupd (%r9,%rax,2),%xmm0 + DB 102,65,15,16,76,65,16 ; movupd 0x10(%r9,%rax,2),%xmm1 DB 102,15,40,208 ; movapd %xmm0,%xmm2 DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 @@ -23054,7 +23128,7 @@ _sk_load_u16_be_sse2 LABEL PROC DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 68,15,40,5,34,33,0,0 ; movaps 0x2122(%rip),%xmm8 # 63f0 <_sk_callback_sse2+0xe4a> + DB 68,15,40,5,33,33,0,0 ; movaps 0x2121(%rip),%xmm8 # 6420 <_sk_callback_sse2+0xe46> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -23080,29 +23154,29 @@ _sk_load_u16_be_sse2 LABEL PROC DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 242,65,15,16,4,64 ; movsd (%r8,%rax,2),%xmm0 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,17 ; jne 4358 <_sk_load_u16_be_sse2+0xf1> + DB 242,65,15,16,4,65 ; movsd (%r9,%rax,2),%xmm0 + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,17 ; jne 4389 <_sk_load_u16_be_sse2+0xf1> DB 102,15,87,201 ; xorpd %xmm1,%xmm1 DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 233,50,255,255,255 ; jmpq 428a <_sk_load_u16_be_sse2+0x23> - DB 102,65,15,22,68,64,8 ; movhpd 0x8(%r8,%rax,2),%xmm0 + DB 233,50,255,255,255 ; jmpq 42bb <_sk_load_u16_be_sse2+0x23> + DB 102,65,15,22,68,65,8 ; movhpd 0x8(%r9,%rax,2),%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 15,130,29,255,255,255 ; jb 428a <_sk_load_u16_be_sse2+0x23> - DB 242,65,15,16,76,64,16 ; movsd 0x10(%r8,%rax,2),%xmm1 - DB 233,17,255,255,255 ; jmpq 428a <_sk_load_u16_be_sse2+0x23> + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 15,130,29,255,255,255 ; jb 42bb <_sk_load_u16_be_sse2+0x23> + DB 242,65,15,16,76,65,16 ; movsd 0x10(%r9,%rax,2),%xmm1 + DB 233,17,255,255,255 ; jmpq 42bb <_sk_load_u16_be_sse2+0x23> PUBLIC _sk_load_rgb_u16_be_sse2 _sk_load_rgb_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,175,0,0,0 ; jne 443a <_sk_load_rgb_u16_be_sse2+0xc1> - DB 243,65,15,111,20,64 ; movdqu (%r8,%rax,2),%xmm2 - DB 243,65,15,111,92,64,8 ; movdqu 0x8(%r8,%rax,2),%xmm3 + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,82 ; lea (%rdx,%rdx,2),%rax + DB 77,133,192 ; test %r8,%r8 + DB 15,133,175,0,0,0 ; jne 446b <_sk_load_rgb_u16_be_sse2+0xc1> + DB 243,65,15,111,20,65 ; movdqu (%r9,%rax,2),%xmm2 + DB 243,65,15,111,92,65,8 ; movdqu 0x8(%r9,%rax,2),%xmm3 DB 102,15,115,219,4 ; psrldq $0x4,%xmm3 DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 @@ -23121,7 +23195,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,65,15,97,200 ; punpcklwd %xmm8,%xmm1 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 68,15,40,13,17,32,0,0 ; movaps 0x2011(%rip),%xmm9 # 6400 <_sk_callback_sse2+0xe5a> + DB 68,15,40,13,16,32,0,0 ; movaps 0x2010(%rip),%xmm9 # 6430 <_sk_callback_sse2+0xe56> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -23138,34 +23212,34 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,216,31,0,0 ; movaps 0x1fd8(%rip),%xmm3 # 6410 <_sk_callback_sse2+0xe6a> + DB 15,40,29,215,31,0,0 ; movaps 0x1fd7(%rip),%xmm3 # 6440 <_sk_callback_sse2+0xe66> DB 255,224 ; jmpq *%rax - DB 102,65,15,110,20,64 ; movd (%r8,%rax,2),%xmm2 - DB 102,65,15,196,84,64,4,2 ; pinsrw $0x2,0x4(%r8,%rax,2),%xmm2 + DB 102,65,15,110,20,65 ; movd (%r9,%rax,2),%xmm2 + DB 102,65,15,196,84,65,4,2 ; pinsrw $0x2,0x4(%r9,%rax,2),%xmm2 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,13 ; jne 445f <_sk_load_rgb_u16_be_sse2+0xe6> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,13 ; jne 4490 <_sk_load_rgb_u16_be_sse2+0xe6> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,80,255,255,255 ; jmpq 43af <_sk_load_rgb_u16_be_sse2+0x36> - DB 102,65,15,110,68,64,6 ; movd 0x6(%r8,%rax,2),%xmm0 - DB 102,65,15,196,68,64,10,2 ; pinsrw $0x2,0xa(%r8,%rax,2),%xmm0 + DB 233,80,255,255,255 ; jmpq 43e0 <_sk_load_rgb_u16_be_sse2+0x36> + DB 102,65,15,110,68,65,6 ; movd 0x6(%r9,%rax,2),%xmm0 + DB 102,65,15,196,68,65,10,2 ; pinsrw $0x2,0xa(%r9,%rax,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,24 ; jb 4490 <_sk_load_rgb_u16_be_sse2+0x117> - DB 102,65,15,110,92,64,12 ; movd 0xc(%r8,%rax,2),%xmm3 - DB 102,65,15,196,92,64,16,2 ; pinsrw $0x2,0x10(%r8,%rax,2),%xmm3 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,24 ; jb 44c1 <_sk_load_rgb_u16_be_sse2+0x117> + DB 102,65,15,110,92,65,12 ; movd 0xc(%r9,%rax,2),%xmm3 + DB 102,65,15,196,92,65,16,2 ; pinsrw $0x2,0x10(%r9,%rax,2),%xmm3 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,31,255,255,255 ; jmpq 43af <_sk_load_rgb_u16_be_sse2+0x36> + DB 233,31,255,255,255 ; jmpq 43e0 <_sk_load_rgb_u16_be_sse2+0x36> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 233,22,255,255,255 ; jmpq 43af <_sk_load_rgb_u16_be_sse2+0x36> + DB 233,22,255,255,255 ; jmpq 43e0 <_sk_load_rgb_u16_be_sse2+0x36> PUBLIC _sk_store_u16_be_sse2 _sk_store_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 68,15,40,21,114,31,0,0 ; movaps 0x1f72(%rip),%xmm10 # 6420 <_sk_callback_sse2+0xe7a> + DB 76,139,8 ; mov (%rax),%r9 + DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax + DB 68,15,40,21,113,31,0,0 ; movaps 0x1f71(%rip),%xmm10 # 6450 <_sk_callback_sse2+0xe76> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 @@ -23209,36 +23283,36 @@ _sk_store_u16_be_sse2 LABEL PROC DB 102,69,15,97,202 ; punpcklwd %xmm10,%xmm9 DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 DB 102,69,15,98,209 ; punpckldq %xmm9,%xmm10 - DB 72,133,201 ; test %rcx,%rcx - DB 117,21 ; jne 45a8 <_sk_store_u16_be_sse2+0x10f> - DB 69,15,17,20,64 ; movups %xmm10,(%r8,%rax,2) + DB 77,133,192 ; test %r8,%r8 + DB 117,21 ; jne 45d9 <_sk_store_u16_be_sse2+0x10f> + DB 69,15,17,20,65 ; movups %xmm10,(%r9,%rax,2) DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 - DB 243,69,15,127,68,64,16 ; movdqu %xmm8,0x10(%r8,%rax,2) + DB 243,69,15,127,68,65,16 ; movdqu %xmm8,0x10(%r9,%rax,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 102,69,15,214,20,64 ; movq %xmm10,(%r8,%rax,2) - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 45a4 <_sk_store_u16_be_sse2+0x10b> - DB 102,69,15,23,84,64,8 ; movhpd %xmm10,0x8(%r8,%rax,2) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 45a4 <_sk_store_u16_be_sse2+0x10b> + DB 102,69,15,214,20,65 ; movq %xmm10,(%r9,%rax,2) + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,240 ; je 45d5 <_sk_store_u16_be_sse2+0x10b> + DB 102,69,15,23,84,65,8 ; movhpd %xmm10,0x8(%r9,%rax,2) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,227 ; jb 45d5 <_sk_store_u16_be_sse2+0x10b> DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 - DB 102,69,15,214,68,64,16 ; movq %xmm8,0x10(%r8,%rax,2) - DB 235,213 ; jmp 45a4 <_sk_store_u16_be_sse2+0x10b> + DB 102,69,15,214,68,65,16 ; movq %xmm8,0x10(%r9,%rax,2) + DB 235,213 ; jmp 45d5 <_sk_store_u16_be_sse2+0x10b> PUBLIC _sk_load_f32_sse2 _sk_load_f32_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 76,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r8 - DB 72,137,248 ; mov %rdi,%rax + DB 76,139,16 ; mov (%rax),%r10 + DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 + DB 72,137,208 ; mov %rdx,%rax DB 72,193,224,4 ; shl $0x4,%rax - DB 69,15,16,4,1 ; movups (%r9,%rax,1),%xmm8 - DB 72,133,201 ; test %rcx,%rcx - DB 117,66 ; jne 462f <_sk_load_f32_sse2+0x60> - DB 67,15,16,68,129,16 ; movups 0x10(%r9,%r8,4),%xmm0 - DB 67,15,16,92,129,32 ; movups 0x20(%r9,%r8,4),%xmm3 - DB 71,15,16,76,129,48 ; movups 0x30(%r9,%r8,4),%xmm9 + DB 69,15,16,4,2 ; movups (%r10,%rax,1),%xmm8 + DB 77,133,192 ; test %r8,%r8 + DB 117,66 ; jne 4660 <_sk_load_f32_sse2+0x60> + DB 67,15,16,68,138,16 ; movups 0x10(%r10,%r9,4),%xmm0 + DB 67,15,16,92,138,32 ; movups 0x20(%r10,%r9,4),%xmm3 + DB 71,15,16,76,138,48 ; movups 0x30(%r10,%r9,4),%xmm9 DB 65,15,40,208 ; movaps %xmm8,%xmm2 DB 15,20,208 ; unpcklps %xmm0,%xmm2 DB 15,40,203 ; movaps %xmm3,%xmm1 @@ -23254,25 +23328,25 @@ _sk_load_f32_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,8 ; jne 4641 <_sk_load_f32_sse2+0x72> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 117,8 ; jne 4672 <_sk_load_f32_sse2+0x72> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 - DB 235,190 ; jmp 45ff <_sk_load_f32_sse2+0x30> - DB 67,15,16,68,129,16 ; movups 0x10(%r9,%r8,4),%xmm0 - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,8 ; jb 4655 <_sk_load_f32_sse2+0x86> - DB 67,15,16,92,129,32 ; movups 0x20(%r9,%r8,4),%xmm3 - DB 235,170 ; jmp 45ff <_sk_load_f32_sse2+0x30> + DB 235,190 ; jmp 4630 <_sk_load_f32_sse2+0x30> + DB 67,15,16,68,138,16 ; movups 0x10(%r10,%r9,4),%xmm0 + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,8 ; jb 4686 <_sk_load_f32_sse2+0x86> + DB 67,15,16,92,138,32 ; movups 0x20(%r10,%r9,4),%xmm3 + DB 235,170 ; jmp 4630 <_sk_load_f32_sse2+0x30> DB 15,87,219 ; xorps %xmm3,%xmm3 - DB 235,165 ; jmp 45ff <_sk_load_f32_sse2+0x30> + DB 235,165 ; jmp 4630 <_sk_load_f32_sse2+0x30> PUBLIC _sk_store_f32_sse2 _sk_store_f32_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,8 ; mov (%rax),%r9 - DB 76,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r8 - DB 72,137,248 ; mov %rdi,%rax + DB 76,139,16 ; mov (%rax),%r10 + DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 + DB 72,137,208 ; mov %rdx,%rax DB 72,193,224,4 ; shl $0x4,%rax DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 @@ -23287,22 +23361,22 @@ _sk_store_f32_sse2 LABEL PROC DB 69,15,18,193 ; movhlps %xmm9,%xmm8 DB 69,15,40,202 ; movaps %xmm10,%xmm9 DB 102,69,15,20,203 ; unpcklpd %xmm11,%xmm9 - DB 102,69,15,17,36,1 ; movupd %xmm12,(%r9,%rax,1) - DB 72,133,201 ; test %rcx,%rcx - DB 117,29 ; jne 46cc <_sk_store_f32_sse2+0x72> + DB 102,69,15,17,36,2 ; movupd %xmm12,(%r10,%rax,1) + DB 77,133,192 ; test %r8,%r8 + DB 117,29 ; jne 46fd <_sk_store_f32_sse2+0x72> DB 102,69,15,21,211 ; unpckhpd %xmm11,%xmm10 - DB 71,15,17,68,129,16 ; movups %xmm8,0x10(%r9,%r8,4) - DB 102,71,15,17,76,129,32 ; movupd %xmm9,0x20(%r9,%r8,4) - DB 102,71,15,17,84,129,48 ; movupd %xmm10,0x30(%r9,%r8,4) + DB 71,15,17,68,138,16 ; movups %xmm8,0x10(%r10,%r9,4) + DB 102,71,15,17,76,138,32 ; movupd %xmm9,0x20(%r10,%r9,4) + DB 102,71,15,17,84,138,48 ; movupd %xmm10,0x30(%r10,%r9,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,246 ; je 46c8 <_sk_store_f32_sse2+0x6e> - DB 71,15,17,68,129,16 ; movups %xmm8,0x10(%r9,%r8,4) - DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,234 ; jb 46c8 <_sk_store_f32_sse2+0x6e> - DB 102,71,15,17,76,129,32 ; movupd %xmm9,0x20(%r9,%r8,4) - DB 235,225 ; jmp 46c8 <_sk_store_f32_sse2+0x6e> + DB 73,131,248,1 ; cmp $0x1,%r8 + DB 116,246 ; je 46f9 <_sk_store_f32_sse2+0x6e> + DB 71,15,17,68,138,16 ; movups %xmm8,0x10(%r10,%r9,4) + DB 73,131,248,3 ; cmp $0x3,%r8 + DB 114,234 ; jb 46f9 <_sk_store_f32_sse2+0x6e> + DB 102,71,15,17,76,138,32 ; movupd %xmm9,0x20(%r10,%r9,4) + DB 235,225 ; jmp 46f9 <_sk_store_f32_sse2+0x6e> PUBLIC _sk_clamp_x_sse2 _sk_clamp_x_sse2 LABEL PROC @@ -23338,7 +23412,7 @@ _sk_repeat_x_sse2 LABEL PROC DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,227,28,0,0 ; andps 0x1ce3(%rip),%xmm9 # 6430 <_sk_callback_sse2+0xe8a> + DB 68,15,84,13,226,28,0,0 ; andps 0x1ce2(%rip),%xmm9 # 6460 <_sk_callback_sse2+0xe86> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,194 ; subps %xmm10,%xmm0 @@ -23355,7 +23429,7 @@ _sk_repeat_y_sse2 LABEL PROC DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,185,28,0,0 ; andps 0x1cb9(%rip),%xmm9 # 6440 <_sk_callback_sse2+0xe9a> + DB 68,15,84,13,184,28,0,0 ; andps 0x1cb8(%rip),%xmm9 # 6470 <_sk_callback_sse2+0xe96> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,202 ; subps %xmm10,%xmm1 @@ -23376,7 +23450,7 @@ _sk_mirror_x_sse2 LABEL PROC DB 243,69,15,91,218 ; cvttps2dq %xmm10,%xmm11 DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10 - DB 68,15,84,21,125,28,0,0 ; andps 0x1c7d(%rip),%xmm10 # 6450 <_sk_callback_sse2+0xeaa> + DB 68,15,84,21,124,28,0,0 ; andps 0x1c7c(%rip),%xmm10 # 6480 <_sk_callback_sse2+0xea6> DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 69,15,89,216 ; mulps %xmm8,%xmm11 @@ -23401,7 +23475,7 @@ _sk_mirror_y_sse2 LABEL PROC DB 243,69,15,91,218 ; cvttps2dq %xmm10,%xmm11 DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10 - DB 68,15,84,21,49,28,0,0 ; andps 0x1c31(%rip),%xmm10 # 6460 <_sk_callback_sse2+0xeba> + DB 68,15,84,21,48,28,0,0 ; andps 0x1c30(%rip),%xmm10 # 6490 <_sk_callback_sse2+0xeb6> DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 69,15,89,216 ; mulps %xmm8,%xmm11 @@ -23416,7 +23490,7 @@ PUBLIC _sk_clamp_x_1_sse2 _sk_clamp_x_1_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,17,28,0,0 ; minps 0x1c11(%rip),%xmm8 # 6470 <_sk_callback_sse2+0xeca> + DB 68,15,93,5,16,28,0,0 ; minps 0x1c10(%rip),%xmm8 # 64a0 <_sk_callback_sse2+0xec6> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -23427,7 +23501,7 @@ _sk_repeat_x_1_sse2 LABEL PROC DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,194,200,1 ; cmpltps %xmm8,%xmm9 - DB 68,15,84,13,255,27,0,0 ; andps 0x1bff(%rip),%xmm9 # 6480 <_sk_callback_sse2+0xeda> + DB 68,15,84,13,254,27,0,0 ; andps 0x1bfe(%rip),%xmm9 # 64b0 <_sk_callback_sse2+0xed6> DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 65,15,92,192 ; subps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax @@ -23435,14 +23509,14 @@ _sk_repeat_x_1_sse2 LABEL PROC PUBLIC _sk_mirror_x_1_sse2 _sk_mirror_x_1_sse2 LABEL PROC - DB 68,15,40,5,251,27,0,0 ; movaps 0x1bfb(%rip),%xmm8 # 6490 <_sk_callback_sse2+0xeea> + DB 68,15,40,5,250,27,0,0 ; movaps 0x1bfa(%rip),%xmm8 # 64c0 <_sk_callback_sse2+0xee6> DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,13,255,27,0,0 ; movaps 0x1bff(%rip),%xmm9 # 64a0 <_sk_callback_sse2+0xefa> + DB 68,15,40,13,254,27,0,0 ; movaps 0x1bfe(%rip),%xmm9 # 64d0 <_sk_callback_sse2+0xef6> DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,245,27,0,0 ; andps 0x1bf5(%rip),%xmm9 # 64b0 <_sk_callback_sse2+0xf0a> + DB 68,15,84,13,244,27,0,0 ; andps 0x1bf4(%rip),%xmm9 # 64e0 <_sk_callback_sse2+0xf06> DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,88,210 ; addps %xmm10,%xmm10 @@ -23456,10 +23530,10 @@ _sk_mirror_x_1_sse2 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse2 _sk_luminance_to_alpha_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,219,27,0,0 ; mulps 0x1bdb(%rip),%xmm0 # 64c0 <_sk_callback_sse2+0xf1a> - DB 15,89,13,228,27,0,0 ; mulps 0x1be4(%rip),%xmm1 # 64d0 <_sk_callback_sse2+0xf2a> + DB 15,89,5,218,27,0,0 ; mulps 0x1bda(%rip),%xmm0 # 64f0 <_sk_callback_sse2+0xf16> + DB 15,89,13,227,27,0,0 ; mulps 0x1be3(%rip),%xmm1 # 6500 <_sk_callback_sse2+0xf26> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,234,27,0,0 ; mulps 0x1bea(%rip),%xmm3 # 64e0 <_sk_callback_sse2+0xf3a> + DB 15,89,29,233,27,0,0 ; mulps 0x1be9(%rip),%xmm3 # 6510 <_sk_callback_sse2+0xf36> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -23725,86 +23799,86 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 76,139,112,8 ; mov 0x8(%rax),%r14 DB 72,255,203 ; dec %rbx - DB 120,7 ; js 4d32 <_sk_evenly_spaced_gradient_sse2+0x18> + DB 120,7 ; js 4d63 <_sk_evenly_spaced_gradient_sse2+0x18> DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 - DB 235,21 ; jmp 4d47 <_sk_evenly_spaced_gradient_sse2+0x2d> - DB 73,137,216 ; mov %rbx,%r8 - DB 73,209,232 ; shr %r8 + DB 235,21 ; jmp 4d78 <_sk_evenly_spaced_gradient_sse2+0x2d> + DB 73,137,217 ; mov %rbx,%r9 + DB 73,209,233 ; shr %r9 DB 131,227,1 ; and $0x1,%ebx - DB 76,9,195 ; or %r8,%rbx + DB 76,9,203 ; or %r9,%rbx DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 DB 243,15,88,201 ; addss %xmm1,%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,73,15,126,210 ; movq %xmm2,%r10 - DB 69,137,208 ; mov %r10d,%r8d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,73,15,126,203 ; movq %xmm1,%r11 + DB 102,73,15,126,211 ; movq %xmm2,%r11 DB 69,137,217 ; mov %r11d,%r9d DB 73,193,235,32 ; shr $0x20,%r11 - DB 243,67,15,16,12,158 ; movss (%r14,%r11,4),%xmm1 - DB 243,67,15,16,20,150 ; movss (%r14,%r10,4),%xmm2 + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 65,137,218 ; mov %ebx,%r10d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 + DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,71,15,16,4,142 ; movss (%r14,%r9,4),%xmm8 - DB 243,67,15,16,20,134 ; movss (%r14,%r8,4),%xmm2 + DB 243,71,15,16,4,150 ; movss (%r14,%r10,4),%xmm8 + DB 243,67,15,16,20,142 ; movss (%r14,%r9,4),%xmm2 DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 - DB 72,139,88,40 ; mov 0x28(%rax),%rbx - DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 76,139,112,40 ; mov 0x28(%rax),%r14 + DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 + DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,71,15,16,12,150 ; movss (%r14,%r10,4),%xmm9 + DB 243,67,15,16,20,142 ; movss (%r14,%r9,4),%xmm2 DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 - DB 72,139,88,16 ; mov 0x10(%rax),%rbx - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 + DB 76,139,112,16 ; mov 0x10(%rax),%r14 + DB 243,65,15,16,20,158 ; movss (%r14,%rbx,4),%xmm2 + DB 243,67,15,16,12,158 ; movss (%r14,%r11,4),%xmm1 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 72,139,88,48 ; mov 0x30(%rax),%rbx - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 76,139,112,48 ; mov 0x30(%rax),%r14 + DB 243,65,15,16,20,158 ; movss (%r14,%rbx,4),%xmm2 + DB 243,67,15,16,28,158 ; movss (%r14,%r11,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 243,70,15,16,20,139 ; movss (%rbx,%r9,4),%xmm10 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,71,15,16,20,150 ; movss (%r14,%r10,4),%xmm10 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 68,15,20,211 ; unpcklps %xmm3,%xmm10 DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 243,70,15,16,28,155 ; movss (%rbx,%r11,4),%xmm11 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 76,139,112,24 ; mov 0x18(%rax),%r14 + DB 243,69,15,16,28,158 ; movss (%r14,%rbx,4),%xmm11 + DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 68,15,20,218 ; unpcklps %xmm2,%xmm11 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,67,15,16,20,150 ; movss (%r14,%r10,4),%xmm2 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,211 ; unpcklps %xmm11,%xmm2 - DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 76,139,112,56 ; mov 0x38(%rax),%r14 + DB 243,69,15,16,36,158 ; movss (%r14,%rbx,4),%xmm12 + DB 243,67,15,16,28,158 ; movss (%r14,%r11,4),%xmm3 DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 - DB 243,70,15,16,28,139 ; movss (%rbx,%r9,4),%xmm11 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,71,15,16,28,150 ; movss (%r14,%r10,4),%xmm11 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 68,15,20,219 ; unpcklps %xmm3,%xmm11 DB 69,15,20,220 ; unpcklps %xmm12,%xmm11 - DB 72,139,88,32 ; mov 0x20(%rax),%rbx - DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 76,139,112,32 ; mov 0x20(%rax),%r14 + DB 243,69,15,16,36,158 ; movss (%r14,%rbx,4),%xmm12 + DB 243,67,15,16,28,158 ; movss (%r14,%r11,4),%xmm3 DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 - DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 + DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 243,71,15,16,44,142 ; movss (%r14,%r9,4),%xmm13 DB 65,15,20,221 ; unpcklps %xmm13,%xmm3 DB 65,15,20,220 ; unpcklps %xmm12,%xmm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 243,70,15,16,36,152 ; movss (%rax,%r11,4),%xmm12 - DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 + DB 243,68,15,16,36,152 ; movss (%rax,%rbx,4),%xmm12 + DB 243,70,15,16,44,152 ; movss (%rax,%r11,4),%xmm13 DB 69,15,20,229 ; unpcklps %xmm13,%xmm12 - DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 - DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 + DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 + DB 243,70,15,16,52,136 ; movss (%rax,%r9,4),%xmm14 DB 69,15,20,238 ; unpcklps %xmm14,%xmm13 DB 69,15,20,236 ; unpcklps %xmm12,%xmm13 DB 68,15,89,192 ; mulps %xmm0,%xmm8 @@ -23823,15 +23897,15 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC PUBLIC _sk_gauss_a_to_rgba_sse2 _sk_gauss_a_to_rgba_sse2 LABEL PROC - DB 15,40,5,29,22,0,0 ; movaps 0x161d(%rip),%xmm0 # 64f0 <_sk_callback_sse2+0xf4a> + DB 15,40,5,28,22,0,0 ; movaps 0x161c(%rip),%xmm0 # 6520 <_sk_callback_sse2+0xf46> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,35,22,0,0 ; addps 0x1623(%rip),%xmm0 # 6500 <_sk_callback_sse2+0xf5a> + DB 15,88,5,34,22,0,0 ; addps 0x1622(%rip),%xmm0 # 6530 <_sk_callback_sse2+0xf56> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,41,22,0,0 ; addps 0x1629(%rip),%xmm0 # 6510 <_sk_callback_sse2+0xf6a> + DB 15,88,5,40,22,0,0 ; addps 0x1628(%rip),%xmm0 # 6540 <_sk_callback_sse2+0xf66> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,47,22,0,0 ; addps 0x162f(%rip),%xmm0 # 6520 <_sk_callback_sse2+0xf7a> + DB 15,88,5,46,22,0,0 ; addps 0x162e(%rip),%xmm0 # 6550 <_sk_callback_sse2+0xf76> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,53,22,0,0 ; addps 0x1635(%rip),%xmm0 # 6530 <_sk_callback_sse2+0xf8a> + DB 15,88,5,52,22,0,0 ; addps 0x1634(%rip),%xmm0 # 6560 <_sk_callback_sse2+0xf86> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -23840,95 +23914,96 @@ _sk_gauss_a_to_rgba_sse2 LABEL PROC PUBLIC _sk_gradient_sse2 _sk_gradient_sse2 LABEL PROC + DB 65,87 ; push %r15 + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 + DB 76,139,8 ; mov (%rax),%r9 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,51 ; jb 4f4a <_sk_gradient_sse2+0x42> - DB 76,139,72,72 ; mov 0x48(%rax),%r9 - DB 73,255,200 ; dec %r8 - DB 73,131,193,4 ; add $0x4,%r9 + DB 73,131,249,2 ; cmp $0x2,%r9 + DB 114,50 ; jb 4f7f <_sk_gradient_sse2+0x46> + DB 72,139,88,72 ; mov 0x48(%rax),%rbx + DB 73,255,201 ; dec %r9 + DB 72,131,195,4 ; add $0x4,%rbx DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,21,19,22,0,0 ; movaps 0x1613(%rip),%xmm2 # 6540 <_sk_callback_sse2+0xf9a> - DB 243,65,15,16,25 ; movss (%r9),%xmm3 + DB 15,40,21,13,22,0,0 ; movaps 0x160d(%rip),%xmm2 # 6570 <_sk_callback_sse2+0xf96> + DB 243,15,16,27 ; movss (%rbx),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 DB 15,84,218 ; andps %xmm2,%xmm3 DB 102,15,254,203 ; paddd %xmm3,%xmm1 - DB 73,131,193,4 ; add $0x4,%r9 - DB 73,255,200 ; dec %r8 - DB 117,227 ; jne 4f2d <_sk_gradient_sse2+0x25> - DB 65,86 ; push %r14 - DB 83 ; push %rbx + DB 72,131,195,4 ; add $0x4,%rbx + DB 73,255,201 ; dec %r9 + DB 117,228 ; jne 4f63 <_sk_gradient_sse2+0x2a> DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,73,15,126,210 ; movq %xmm2,%r10 - DB 69,137,208 ; mov %r10d,%r8d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,73,15,126,203 ; movq %xmm1,%r11 + DB 102,73,15,126,211 ; movq %xmm2,%r11 DB 69,137,217 ; mov %r11d,%r9d DB 73,193,235,32 ; shr $0x20,%r11 - DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 65,137,218 ; mov %ebx,%r10d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 76,139,112,8 ; mov 0x8(%rax),%r14 + DB 76,139,120,16 ; mov 0x10(%rax),%r15 + DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 + DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,70,15,16,4,139 ; movss (%rbx,%r9,4),%xmm8 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,71,15,16,4,150 ; movss (%r14,%r10,4),%xmm8 + DB 243,67,15,16,20,142 ; movss (%r14,%r9,4),%xmm2 DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 - DB 72,139,88,40 ; mov 0x28(%rax),%rbx - DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 76,139,112,40 ; mov 0x28(%rax),%r14 + DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 + DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,71,15,16,12,150 ; movss (%r14,%r10,4),%xmm9 + DB 243,67,15,16,20,142 ; movss (%r14,%r9,4),%xmm2 DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 - DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 - DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 + DB 243,65,15,16,20,159 ; movss (%r15,%rbx,4),%xmm2 + DB 243,67,15,16,12,159 ; movss (%r15,%r11,4),%xmm1 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 + DB 243,67,15,16,12,151 ; movss (%r15,%r10,4),%xmm1 + DB 243,67,15,16,28,143 ; movss (%r15,%r9,4),%xmm3 DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 72,139,88,48 ; mov 0x30(%rax),%rbx - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 76,139,112,48 ; mov 0x30(%rax),%r14 + DB 243,65,15,16,20,158 ; movss (%r14,%rbx,4),%xmm2 + DB 243,67,15,16,28,158 ; movss (%r14,%r11,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 243,70,15,16,20,139 ; movss (%rbx,%r9,4),%xmm10 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,71,15,16,20,150 ; movss (%r14,%r10,4),%xmm10 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 68,15,20,211 ; unpcklps %xmm3,%xmm10 DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 243,70,15,16,28,155 ; movss (%rbx,%r11,4),%xmm11 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 76,139,112,24 ; mov 0x18(%rax),%r14 + DB 243,69,15,16,28,158 ; movss (%r14,%rbx,4),%xmm11 + DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 68,15,20,218 ; unpcklps %xmm2,%xmm11 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,67,15,16,20,150 ; movss (%r14,%r10,4),%xmm2 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,211 ; unpcklps %xmm11,%xmm2 - DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 76,139,112,56 ; mov 0x38(%rax),%r14 + DB 243,69,15,16,36,158 ; movss (%r14,%rbx,4),%xmm12 + DB 243,67,15,16,28,158 ; movss (%r14,%r11,4),%xmm3 DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 - DB 243,70,15,16,28,139 ; movss (%rbx,%r9,4),%xmm11 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,71,15,16,28,150 ; movss (%r14,%r10,4),%xmm11 + DB 243,67,15,16,28,142 ; movss (%r14,%r9,4),%xmm3 DB 68,15,20,219 ; unpcklps %xmm3,%xmm11 DB 69,15,20,220 ; unpcklps %xmm12,%xmm11 - DB 72,139,88,32 ; mov 0x20(%rax),%rbx - DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 76,139,112,32 ; mov 0x20(%rax),%r14 + DB 243,69,15,16,36,158 ; movss (%r14,%rbx,4),%xmm12 + DB 243,67,15,16,28,158 ; movss (%r14,%r11,4),%xmm3 DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 - DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 + DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 243,71,15,16,44,142 ; movss (%r14,%r9,4),%xmm13 DB 65,15,20,221 ; unpcklps %xmm13,%xmm3 DB 65,15,20,220 ; unpcklps %xmm12,%xmm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 243,70,15,16,36,152 ; movss (%rax,%r11,4),%xmm12 - DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 + DB 243,68,15,16,36,152 ; movss (%rax,%rbx,4),%xmm12 + DB 243,70,15,16,44,152 ; movss (%rax,%r11,4),%xmm13 DB 69,15,20,229 ; unpcklps %xmm13,%xmm12 - DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 - DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 + DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 + DB 243,70,15,16,52,136 ; movss (%rax,%r9,4),%xmm14 DB 69,15,20,238 ; unpcklps %xmm14,%xmm13 DB 69,15,20,236 ; unpcklps %xmm12,%xmm13 DB 68,15,89,192 ; mulps %xmm0,%xmm8 @@ -23943,6 +24018,7 @@ _sk_gradient_sse2 LABEL PROC DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 91 ; pop %rbx DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax PUBLIC _sk_evenly_spaced_2_stop_gradient_sse2 @@ -23993,29 +24069,29 @@ _sk_xy_to_unit_angle_sse2 LABEL PROC DB 69,15,94,220 ; divps %xmm12,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 - DB 68,15,40,45,210,19,0,0 ; movaps 0x13d2(%rip),%xmm13 # 6550 <_sk_callback_sse2+0xfaa> + DB 68,15,40,45,206,19,0,0 ; movaps 0x13ce(%rip),%xmm13 # 6580 <_sk_callback_sse2+0xfa6> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,214,19,0,0 ; addps 0x13d6(%rip),%xmm13 # 6560 <_sk_callback_sse2+0xfba> + DB 68,15,88,45,210,19,0,0 ; addps 0x13d2(%rip),%xmm13 # 6590 <_sk_callback_sse2+0xfb6> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,218,19,0,0 ; addps 0x13da(%rip),%xmm13 # 6570 <_sk_callback_sse2+0xfca> + DB 68,15,88,45,214,19,0,0 ; addps 0x13d6(%rip),%xmm13 # 65a0 <_sk_callback_sse2+0xfc6> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,222,19,0,0 ; addps 0x13de(%rip),%xmm13 # 6580 <_sk_callback_sse2+0xfda> + DB 68,15,88,45,218,19,0,0 ; addps 0x13da(%rip),%xmm13 # 65b0 <_sk_callback_sse2+0xfd6> DB 69,15,89,235 ; mulps %xmm11,%xmm13 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,40,21,221,19,0,0 ; movaps 0x13dd(%rip),%xmm10 # 6590 <_sk_callback_sse2+0xfea> + DB 68,15,40,21,217,19,0,0 ; movaps 0x13d9(%rip),%xmm10 # 65c0 <_sk_callback_sse2+0xfe6> DB 69,15,92,213 ; subps %xmm13,%xmm10 DB 69,15,84,209 ; andps %xmm9,%xmm10 DB 69,15,85,205 ; andnps %xmm13,%xmm9 DB 69,15,86,202 ; orps %xmm10,%xmm9 DB 68,15,194,192,1 ; cmpltps %xmm0,%xmm8 - DB 68,15,40,21,208,19,0,0 ; movaps 0x13d0(%rip),%xmm10 # 65a0 <_sk_callback_sse2+0xffa> + DB 68,15,40,21,204,19,0,0 ; movaps 0x13cc(%rip),%xmm10 # 65d0 <_sk_callback_sse2+0xff6> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,84,208 ; andps %xmm8,%xmm10 DB 69,15,85,193 ; andnps %xmm9,%xmm8 DB 69,15,86,194 ; orps %xmm10,%xmm8 DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 68,15,194,200,1 ; cmpltps %xmm0,%xmm9 - DB 68,15,40,21,191,19,0,0 ; movaps 0x13bf(%rip),%xmm10 # 65b0 <_sk_callback_sse2+0x100a> + DB 68,15,40,21,187,19,0,0 ; movaps 0x13bb(%rip),%xmm10 # 65e0 <_sk_callback_sse2+0x1006> DB 69,15,92,208 ; subps %xmm8,%xmm10 DB 69,15,84,209 ; andps %xmm9,%xmm10 DB 69,15,85,200 ; andnps %xmm8,%xmm9 @@ -24038,7 +24114,7 @@ _sk_xy_to_radius_sse2 LABEL PROC PUBLIC _sk_save_xy_sse2 _sk_save_xy_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,145,19,0,0 ; movaps 0x1391(%rip),%xmm8 # 65c0 <_sk_callback_sse2+0x101a> + DB 68,15,40,5,141,19,0,0 ; movaps 0x138d(%rip),%xmm8 # 65f0 <_sk_callback_sse2+0x1016> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -24046,7 +24122,7 @@ _sk_save_xy_sse2 LABEL PROC DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,40,217 ; movaps %xmm9,%xmm11 DB 69,15,194,218,1 ; cmpltps %xmm10,%xmm11 - DB 68,15,40,37,124,19,0,0 ; movaps 0x137c(%rip),%xmm12 # 65d0 <_sk_callback_sse2+0x102a> + DB 68,15,40,37,120,19,0,0 ; movaps 0x1378(%rip),%xmm12 # 6600 <_sk_callback_sse2+0x1026> DB 69,15,84,220 ; andps %xmm12,%xmm11 DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 69,15,92,202 ; subps %xmm10,%xmm9 @@ -24089,8 +24165,8 @@ _sk_bilinear_nx_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,245,18,0,0 ; addps 0x12f5(%rip),%xmm0 # 65e0 <_sk_callback_sse2+0x103a> - DB 68,15,40,13,253,18,0,0 ; movaps 0x12fd(%rip),%xmm9 # 65f0 <_sk_callback_sse2+0x104a> + DB 15,88,5,241,18,0,0 ; addps 0x12f1(%rip),%xmm0 # 6610 <_sk_callback_sse2+0x1036> + DB 68,15,40,13,249,18,0,0 ; movaps 0x12f9(%rip),%xmm9 # 6620 <_sk_callback_sse2+0x1046> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -24101,7 +24177,7 @@ _sk_bilinear_px_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,236,18,0,0 ; addps 0x12ec(%rip),%xmm0 # 6600 <_sk_callback_sse2+0x105a> + DB 15,88,5,232,18,0,0 ; addps 0x12e8(%rip),%xmm0 # 6630 <_sk_callback_sse2+0x1056> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -24111,8 +24187,8 @@ _sk_bilinear_ny_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,222,18,0,0 ; addps 0x12de(%rip),%xmm1 # 6610 <_sk_callback_sse2+0x106a> - DB 68,15,40,13,230,18,0,0 ; movaps 0x12e6(%rip),%xmm9 # 6620 <_sk_callback_sse2+0x107a> + DB 15,88,13,218,18,0,0 ; addps 0x12da(%rip),%xmm1 # 6640 <_sk_callback_sse2+0x1066> + DB 68,15,40,13,226,18,0,0 ; movaps 0x12e2(%rip),%xmm9 # 6650 <_sk_callback_sse2+0x1076> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -24123,7 +24199,7 @@ _sk_bilinear_py_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,212,18,0,0 ; addps 0x12d4(%rip),%xmm1 # 6630 <_sk_callback_sse2+0x108a> + DB 15,88,13,208,18,0,0 ; addps 0x12d0(%rip),%xmm1 # 6660 <_sk_callback_sse2+0x1086> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -24133,13 +24209,13 @@ _sk_bicubic_n3x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,199,18,0,0 ; addps 0x12c7(%rip),%xmm0 # 6640 <_sk_callback_sse2+0x109a> - DB 68,15,40,13,207,18,0,0 ; movaps 0x12cf(%rip),%xmm9 # 6650 <_sk_callback_sse2+0x10aa> + DB 15,88,5,195,18,0,0 ; addps 0x12c3(%rip),%xmm0 # 6670 <_sk_callback_sse2+0x1096> + DB 68,15,40,13,203,18,0,0 ; movaps 0x12cb(%rip),%xmm9 # 6680 <_sk_callback_sse2+0x10a6> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,203,18,0,0 ; mulps 0x12cb(%rip),%xmm9 # 6660 <_sk_callback_sse2+0x10ba> - DB 68,15,88,13,211,18,0,0 ; addps 0x12d3(%rip),%xmm9 # 6670 <_sk_callback_sse2+0x10ca> + DB 68,15,89,13,199,18,0,0 ; mulps 0x12c7(%rip),%xmm9 # 6690 <_sk_callback_sse2+0x10b6> + DB 68,15,88,13,207,18,0,0 ; addps 0x12cf(%rip),%xmm9 # 66a0 <_sk_callback_sse2+0x10c6> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -24150,16 +24226,16 @@ _sk_bicubic_n1x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,194,18,0,0 ; addps 0x12c2(%rip),%xmm0 # 6680 <_sk_callback_sse2+0x10da> - DB 68,15,40,13,202,18,0,0 ; movaps 0x12ca(%rip),%xmm9 # 6690 <_sk_callback_sse2+0x10ea> + DB 15,88,5,190,18,0,0 ; addps 0x12be(%rip),%xmm0 # 66b0 <_sk_callback_sse2+0x10d6> + DB 68,15,40,13,198,18,0,0 ; movaps 0x12c6(%rip),%xmm9 # 66c0 <_sk_callback_sse2+0x10e6> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,206,18,0,0 ; movaps 0x12ce(%rip),%xmm8 # 66a0 <_sk_callback_sse2+0x10fa> + DB 68,15,40,5,202,18,0,0 ; movaps 0x12ca(%rip),%xmm8 # 66d0 <_sk_callback_sse2+0x10f6> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,210,18,0,0 ; addps 0x12d2(%rip),%xmm8 # 66b0 <_sk_callback_sse2+0x110a> + DB 68,15,88,5,206,18,0,0 ; addps 0x12ce(%rip),%xmm8 # 66e0 <_sk_callback_sse2+0x1106> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,214,18,0,0 ; addps 0x12d6(%rip),%xmm8 # 66c0 <_sk_callback_sse2+0x111a> + DB 68,15,88,5,210,18,0,0 ; addps 0x12d2(%rip),%xmm8 # 66f0 <_sk_callback_sse2+0x1116> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,218,18,0,0 ; addps 0x12da(%rip),%xmm8 # 66d0 <_sk_callback_sse2+0x112a> + DB 68,15,88,5,214,18,0,0 ; addps 0x12d6(%rip),%xmm8 # 6700 <_sk_callback_sse2+0x1126> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -24167,17 +24243,17 @@ _sk_bicubic_n1x_sse2 LABEL PROC PUBLIC _sk_bicubic_p1x_sse2 _sk_bicubic_p1x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,212,18,0,0 ; movaps 0x12d4(%rip),%xmm8 # 66e0 <_sk_callback_sse2+0x113a> + DB 68,15,40,5,208,18,0,0 ; movaps 0x12d0(%rip),%xmm8 # 6710 <_sk_callback_sse2+0x1136> DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,72,64 ; movups 0x40(%rax),%xmm9 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,21,208,18,0,0 ; movaps 0x12d0(%rip),%xmm10 # 66f0 <_sk_callback_sse2+0x114a> + DB 68,15,40,21,204,18,0,0 ; movaps 0x12cc(%rip),%xmm10 # 6720 <_sk_callback_sse2+0x1146> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,212,18,0,0 ; addps 0x12d4(%rip),%xmm10 # 6700 <_sk_callback_sse2+0x115a> + DB 68,15,88,21,208,18,0,0 ; addps 0x12d0(%rip),%xmm10 # 6730 <_sk_callback_sse2+0x1156> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,208,18,0,0 ; addps 0x12d0(%rip),%xmm10 # 6710 <_sk_callback_sse2+0x116a> + DB 68,15,88,21,204,18,0,0 ; addps 0x12cc(%rip),%xmm10 # 6740 <_sk_callback_sse2+0x1166> DB 68,15,17,144,128,0,0,0 ; movups %xmm10,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -24187,11 +24263,11 @@ _sk_bicubic_p3x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,195,18,0,0 ; addps 0x12c3(%rip),%xmm0 # 6720 <_sk_callback_sse2+0x117a> + DB 15,88,5,191,18,0,0 ; addps 0x12bf(%rip),%xmm0 # 6750 <_sk_callback_sse2+0x1176> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,195,18,0,0 ; mulps 0x12c3(%rip),%xmm8 # 6730 <_sk_callback_sse2+0x118a> - DB 68,15,88,5,203,18,0,0 ; addps 0x12cb(%rip),%xmm8 # 6740 <_sk_callback_sse2+0x119a> + DB 68,15,89,5,191,18,0,0 ; mulps 0x12bf(%rip),%xmm8 # 6760 <_sk_callback_sse2+0x1186> + DB 68,15,88,5,199,18,0,0 ; addps 0x12c7(%rip),%xmm8 # 6770 <_sk_callback_sse2+0x1196> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -24202,13 +24278,13 @@ _sk_bicubic_n3y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,185,18,0,0 ; addps 0x12b9(%rip),%xmm1 # 6750 <_sk_callback_sse2+0x11aa> - DB 68,15,40,13,193,18,0,0 ; movaps 0x12c1(%rip),%xmm9 # 6760 <_sk_callback_sse2+0x11ba> + DB 15,88,13,181,18,0,0 ; addps 0x12b5(%rip),%xmm1 # 6780 <_sk_callback_sse2+0x11a6> + DB 68,15,40,13,189,18,0,0 ; movaps 0x12bd(%rip),%xmm9 # 6790 <_sk_callback_sse2+0x11b6> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,189,18,0,0 ; mulps 0x12bd(%rip),%xmm9 # 6770 <_sk_callback_sse2+0x11ca> - DB 68,15,88,13,197,18,0,0 ; addps 0x12c5(%rip),%xmm9 # 6780 <_sk_callback_sse2+0x11da> + DB 68,15,89,13,185,18,0,0 ; mulps 0x12b9(%rip),%xmm9 # 67a0 <_sk_callback_sse2+0x11c6> + DB 68,15,88,13,193,18,0,0 ; addps 0x12c1(%rip),%xmm9 # 67b0 <_sk_callback_sse2+0x11d6> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -24219,16 +24295,16 @@ _sk_bicubic_n1y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,179,18,0,0 ; addps 0x12b3(%rip),%xmm1 # 6790 <_sk_callback_sse2+0x11ea> - DB 68,15,40,13,187,18,0,0 ; movaps 0x12bb(%rip),%xmm9 # 67a0 <_sk_callback_sse2+0x11fa> + DB 15,88,13,175,18,0,0 ; addps 0x12af(%rip),%xmm1 # 67c0 <_sk_callback_sse2+0x11e6> + DB 68,15,40,13,183,18,0,0 ; movaps 0x12b7(%rip),%xmm9 # 67d0 <_sk_callback_sse2+0x11f6> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,191,18,0,0 ; movaps 0x12bf(%rip),%xmm8 # 67b0 <_sk_callback_sse2+0x120a> + DB 68,15,40,5,187,18,0,0 ; movaps 0x12bb(%rip),%xmm8 # 67e0 <_sk_callback_sse2+0x1206> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,195,18,0,0 ; addps 0x12c3(%rip),%xmm8 # 67c0 <_sk_callback_sse2+0x121a> + DB 68,15,88,5,191,18,0,0 ; addps 0x12bf(%rip),%xmm8 # 67f0 <_sk_callback_sse2+0x1216> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,199,18,0,0 ; addps 0x12c7(%rip),%xmm8 # 67d0 <_sk_callback_sse2+0x122a> + DB 68,15,88,5,195,18,0,0 ; addps 0x12c3(%rip),%xmm8 # 6800 <_sk_callback_sse2+0x1226> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,203,18,0,0 ; addps 0x12cb(%rip),%xmm8 # 67e0 <_sk_callback_sse2+0x123a> + DB 68,15,88,5,199,18,0,0 ; addps 0x12c7(%rip),%xmm8 # 6810 <_sk_callback_sse2+0x1236> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -24236,17 +24312,17 @@ _sk_bicubic_n1y_sse2 LABEL PROC PUBLIC _sk_bicubic_p1y_sse2 _sk_bicubic_p1y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,197,18,0,0 ; movaps 0x12c5(%rip),%xmm8 # 67f0 <_sk_callback_sse2+0x124a> + DB 68,15,40,5,193,18,0,0 ; movaps 0x12c1(%rip),%xmm8 # 6820 <_sk_callback_sse2+0x1246> DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,72,96 ; movups 0x60(%rax),%xmm9 DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 68,15,40,21,192,18,0,0 ; movaps 0x12c0(%rip),%xmm10 # 6800 <_sk_callback_sse2+0x125a> + DB 68,15,40,21,188,18,0,0 ; movaps 0x12bc(%rip),%xmm10 # 6830 <_sk_callback_sse2+0x1256> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,196,18,0,0 ; addps 0x12c4(%rip),%xmm10 # 6810 <_sk_callback_sse2+0x126a> + DB 68,15,88,21,192,18,0,0 ; addps 0x12c0(%rip),%xmm10 # 6840 <_sk_callback_sse2+0x1266> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,192,18,0,0 ; addps 0x12c0(%rip),%xmm10 # 6820 <_sk_callback_sse2+0x127a> + DB 68,15,88,21,188,18,0,0 ; addps 0x12bc(%rip),%xmm10 # 6850 <_sk_callback_sse2+0x1276> DB 68,15,17,144,160,0,0,0 ; movups %xmm10,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -24256,11 +24332,11 @@ _sk_bicubic_p3y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,178,18,0,0 ; addps 0x12b2(%rip),%xmm1 # 6830 <_sk_callback_sse2+0x128a> + DB 15,88,13,174,18,0,0 ; addps 0x12ae(%rip),%xmm1 # 6860 <_sk_callback_sse2+0x1286> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,178,18,0,0 ; mulps 0x12b2(%rip),%xmm8 # 6840 <_sk_callback_sse2+0x129a> - DB 68,15,88,5,186,18,0,0 ; addps 0x12ba(%rip),%xmm8 # 6850 <_sk_callback_sse2+0x12aa> + DB 68,15,89,5,174,18,0,0 ; mulps 0x12ae(%rip),%xmm8 # 6870 <_sk_callback_sse2+0x1296> + DB 68,15,88,5,182,18,0,0 ; addps 0x12b6(%rip),%xmm8 # 6880 <_sk_callback_sse2+0x12a6> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -24270,12 +24346,14 @@ PUBLIC _sk_callback_sse2 _sk_callback_sse2 LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,131,236,32 ; sub $0x20,%rsp + DB 72,131,236,40 ; sub $0x28,%rsp DB 68,15,40,197 ; movaps %xmm5,%xmm8 DB 68,15,40,204 ; movaps %xmm4,%xmm9 - DB 73,137,207 ; mov %rcx,%r15 - DB 73,137,214 ; mov %rdx,%r14 + DB 77,137,196 ; mov %r8,%r12 + DB 73,137,206 ; mov %rcx,%r14 + DB 73,137,215 ; mov %rdx,%r15 DB 72,173 ; lods %ds:(%rsi),%rax DB 72,137,195 ; mov %rax,%rbx DB 15,40,224 ; movaps %xmm0,%xmm4 @@ -24294,9 +24372,9 @@ _sk_callback_sse2 LABEL PROC DB 15,17,107,24 ; movups %xmm5,0x18(%rbx) DB 102,15,17,91,40 ; movupd %xmm3,0x28(%rbx) DB 15,17,83,56 ; movups %xmm2,0x38(%rbx) - DB 77,133,255 ; test %r15,%r15 + DB 77,133,228 ; test %r12,%r12 DB 186,4,0,0,0 ; mov $0x4,%edx - DB 65,15,69,215 ; cmovne %r15d,%edx + DB 65,15,69,212 ; cmovne %r12d,%edx DB 72,137,217 ; mov %rbx,%rcx DB 255,19 ; callq *(%rbx) DB 72,139,131,136,0,0,0 ; mov 0x88(%rbx),%rax @@ -24317,12 +24395,14 @@ _sk_callback_sse2 LABEL PROC DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 15,18,220 ; movhlps %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,137,242 ; mov %r14,%rdx - DB 76,137,249 ; mov %r15,%rcx + DB 76,137,250 ; mov %r15,%rdx + DB 76,137,241 ; mov %r14,%rcx + DB 77,137,224 ; mov %r12,%r8 DB 65,15,40,225 ; movaps %xmm9,%xmm4 DB 65,15,40,232 ; movaps %xmm8,%xmm5 - DB 72,131,196,32 ; add $0x20,%rsp + DB 72,131,196,40 ; add $0x28,%rsp DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax @@ -24471,11 +24551,11 @@ ALIGN 16 DB 128,191,0,0,128,191,0 ; cmpb $0x0,-0x40800000(%rdi) DB 0,224 ; add %ah,%al DB 64,0,0 ; add %al,(%rax) - DB 224,64 ; loopne 5848 <.literal16+0x1d8> + DB 224,64 ; loopne 5878 <.literal16+0x1d8> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 584c <.literal16+0x1dc> + DB 224,64 ; loopne 587c <.literal16+0x1dc> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 5850 <.literal16+0x1e0> + DB 224,64 ; loopne 5880 <.literal16+0x1e0> DB 154 ; (bad) DB 153 ; cltd DB 153 ; cltd @@ -24495,13 +24575,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5871 <.literal16+0x201> + DB 71,225,61 ; rex.RXB loope 58a1 <.literal16+0x201> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5875 <.literal16+0x205> + DB 71,225,61 ; rex.RXB loope 58a5 <.literal16+0x205> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5879 <.literal16+0x209> + DB 71,225,61 ; rex.RXB loope 58a9 <.literal16+0x209> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 587d <.literal16+0x20d> + DB 71,225,61 ; rex.RXB loope 58ad <.literal16+0x20d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24526,13 +24606,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58b1 <.literal16+0x241> + DB 71,225,61 ; rex.RXB loope 58e1 <.literal16+0x241> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58b5 <.literal16+0x245> + DB 71,225,61 ; rex.RXB loope 58e5 <.literal16+0x245> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58b9 <.literal16+0x249> + DB 71,225,61 ; rex.RXB loope 58e9 <.literal16+0x249> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58bd <.literal16+0x24d> + DB 71,225,61 ; rex.RXB loope 58ed <.literal16+0x24d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24557,13 +24637,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58f1 <.literal16+0x281> + DB 71,225,61 ; rex.RXB loope 5921 <.literal16+0x281> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58f5 <.literal16+0x285> + DB 71,225,61 ; rex.RXB loope 5925 <.literal16+0x285> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58f9 <.literal16+0x289> + DB 71,225,61 ; rex.RXB loope 5929 <.literal16+0x289> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 58fd <.literal16+0x28d> + DB 71,225,61 ; rex.RXB loope 592d <.literal16+0x28d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24588,13 +24668,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5931 <.literal16+0x2c1> + DB 71,225,61 ; rex.RXB loope 5961 <.literal16+0x2c1> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5935 <.literal16+0x2c5> + DB 71,225,61 ; rex.RXB loope 5965 <.literal16+0x2c5> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5939 <.literal16+0x2c9> + DB 71,225,61 ; rex.RXB loope 5969 <.literal16+0x2c9> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 593d <.literal16+0x2cd> + DB 71,225,61 ; rex.RXB loope 596d <.literal16+0x2cd> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24615,11 +24695,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) DB 67,0,0 ; rex.XB add %al,(%r8) - DB 127,67 ; jg 597b <.literal16+0x30b> + DB 127,67 ; jg 59ab <.literal16+0x30b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 597f <.literal16+0x30f> + DB 127,67 ; jg 59af <.literal16+0x30f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5983 <.literal16+0x313> + DB 127,67 ; jg 59b3 <.literal16+0x313> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24859,13 +24939,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5b69 <.literal16+0x4f9> + DB 224,7 ; loopne 5b99 <.literal16+0x4f9> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5b6d <.literal16+0x4fd> + DB 224,7 ; loopne 5b9d <.literal16+0x4fd> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5b71 <.literal16+0x501> + DB 224,7 ; loopne 5ba1 <.literal16+0x501> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5b75 <.literal16+0x505> + DB 224,7 ; loopne 5ba5 <.literal16+0x505> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -24930,11 +25010,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5c4b <.literal16+0x5db> + DB 127,67 ; jg 5c7b <.literal16+0x5db> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5c4f <.literal16+0x5df> + DB 127,67 ; jg 5c7f <.literal16+0x5df> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5c53 <.literal16+0x5e3> + DB 127,67 ; jg 5c83 <.literal16+0x5e3> DB 129,128,128,59,129,128,128,59,129,128; addl $0x80813b80,-0x7f7ec480(%rax) DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,129,128,128,59 ; addb $0x3b,-0x7f7f7ec5(%rax) @@ -24949,16 +25029,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5c44 <.literal16+0x5d4> + DB 127,0 ; jg 5c74 <.literal16+0x5d4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5c48 <.literal16+0x5d8> + DB 127,0 ; jg 5c78 <.literal16+0x5d8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5c4c <.literal16+0x5dc> + DB 127,0 ; jg 5c7c <.literal16+0x5dc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5c50 <.literal16+0x5e0> + DB 127,0 ; jg 5c80 <.literal16+0x5e0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -24967,7 +25047,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5cd5 <.literal16+0x665> + DB 119,115 ; ja 5d05 <.literal16+0x665> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -24978,7 +25058,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5c39 <.literal16+0x5c9> + DB 117,191 ; jne 5c69 <.literal16+0x5c9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -24990,7 +25070,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a39c7a <_sk_callback_sse2+0xffffffffe9a346d4> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39caa <_sk_callback_sse2+0xffffffffe9a346d0> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25044,16 +25124,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5d14 <.literal16+0x6a4> + DB 127,0 ; jg 5d44 <.literal16+0x6a4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5d18 <.literal16+0x6a8> + DB 127,0 ; jg 5d48 <.literal16+0x6a8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5d1c <.literal16+0x6ac> + DB 127,0 ; jg 5d4c <.literal16+0x6ac> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5d20 <.literal16+0x6b0> + DB 127,0 ; jg 5d50 <.literal16+0x6b0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25062,7 +25142,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5da5 <.literal16+0x735> + DB 119,115 ; ja 5dd5 <.literal16+0x735> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25073,7 +25153,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5d09 <.literal16+0x699> + DB 117,191 ; jne 5d39 <.literal16+0x699> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25085,7 +25165,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a39d4a <_sk_callback_sse2+0xffffffffe9a347a4> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39d7a <_sk_callback_sse2+0xffffffffe9a347a0> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25139,16 +25219,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5de4 <.literal16+0x774> + DB 127,0 ; jg 5e14 <.literal16+0x774> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5de8 <.literal16+0x778> + DB 127,0 ; jg 5e18 <.literal16+0x778> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5dec <.literal16+0x77c> + DB 127,0 ; jg 5e1c <.literal16+0x77c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5df0 <.literal16+0x780> + DB 127,0 ; jg 5e20 <.literal16+0x780> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25157,7 +25237,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5e75 <.literal16+0x805> + DB 119,115 ; ja 5ea5 <.literal16+0x805> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25168,7 +25248,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5dd9 <.literal16+0x769> + DB 117,191 ; jne 5e09 <.literal16+0x769> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25180,7 +25260,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a39e1a <_sk_callback_sse2+0xffffffffe9a34874> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39e4a <_sk_callback_sse2+0xffffffffe9a34870> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25234,16 +25314,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5eb4 <.literal16+0x844> + DB 127,0 ; jg 5ee4 <.literal16+0x844> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5eb8 <.literal16+0x848> + DB 127,0 ; jg 5ee8 <.literal16+0x848> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5ebc <.literal16+0x84c> + DB 127,0 ; jg 5eec <.literal16+0x84c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5ec0 <.literal16+0x850> + DB 127,0 ; jg 5ef0 <.literal16+0x850> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25252,7 +25332,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5f45 <.literal16+0x8d5> + DB 119,115 ; ja 5f75 <.literal16+0x8d5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25263,7 +25343,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5ea9 <.literal16+0x839> + DB 117,191 ; jne 5ed9 <.literal16+0x839> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25275,7 +25355,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a39eea <_sk_callback_sse2+0xffffffffe9a34944> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39f1a <_sk_callback_sse2+0xffffffffe9a34940> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25325,13 +25405,13 @@ ALIGN 16 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 - DB 127,67 ; jg 5fc7 <.literal16+0x957> + DB 127,67 ; jg 5ff7 <.literal16+0x957> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5fcb <.literal16+0x95b> + DB 127,67 ; jg 5ffb <.literal16+0x95b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5fcf <.literal16+0x95f> + DB 127,67 ; jg 5fff <.literal16+0x95f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5fd3 <.literal16+0x963> + DB 127,67 ; jg 6003 <.literal16+0x963> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -25378,16 +25458,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 6053 <.literal16+0x9e3> + DB 118,63 ; jbe 6083 <.literal16+0x9e3> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 6057 <.literal16+0x9e7> + DB 118,63 ; jbe 6087 <.literal16+0x9e7> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 605b <.literal16+0x9eb> + DB 118,63 ; jbe 608b <.literal16+0x9eb> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 605f <.literal16+0x9ef> + DB 118,63 ; jbe 608f <.literal16+0x9ef> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) @@ -25407,11 +25487,11 @@ ALIGN 16 DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 60ab <.literal16+0xa3b> + DB 127,67 ; jg 60db <.literal16+0xa3b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 60af <.literal16+0xa3f> + DB 127,67 ; jg 60df <.literal16+0xa3f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 60b3 <.literal16+0xa43> + DB 127,67 ; jg 60e3 <.literal16+0xa43> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -25459,13 +25539,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6109 <.literal16+0xa99> + DB 224,7 ; loopne 6139 <.literal16+0xa99> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 610d <.literal16+0xa9d> + DB 224,7 ; loopne 613d <.literal16+0xa9d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 6111 <.literal16+0xaa1> + DB 224,7 ; loopne 6141 <.literal16+0xaa1> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 6115 <.literal16+0xaa5> + DB 224,7 ; loopne 6145 <.literal16+0xaa5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -25511,13 +25591,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6179 <.literal16+0xb09> + DB 224,7 ; loopne 61a9 <.literal16+0xb09> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 617d <.literal16+0xb0d> + DB 224,7 ; loopne 61ad <.literal16+0xb0d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 6181 <.literal16+0xb11> + DB 224,7 ; loopne 61b1 <.literal16+0xb11> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 6185 <.literal16+0xb15> + DB 224,7 ; loopne 61b5 <.literal16+0xb15> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -25555,13 +25635,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 6216 <.literal16+0xba6> + DB 124,66 ; jl 6246 <.literal16+0xba6> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 621a <.literal16+0xbaa> + DB 124,66 ; jl 624a <.literal16+0xbaa> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 621e <.literal16+0xbae> + DB 124,66 ; jl 624e <.literal16+0xbae> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 6222 <.literal16+0xbb2> + DB 124,66 ; jl 6252 <.literal16+0xbb2> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -25651,13 +25731,13 @@ ALIGN 16 DB 136,136,61,137,136,136 ; mov %cl,-0x777776c3(%rax) DB 61,137,136,136,61 ; cmp $0x3d888889,%eax DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 6325 <.literal16+0xcb5> + DB 112,65 ; jo 6355 <.literal16+0xcb5> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 6329 <.literal16+0xcb9> + DB 112,65 ; jo 6359 <.literal16+0xcb9> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 632d <.literal16+0xcbd> + DB 112,65 ; jo 635d <.literal16+0xcbd> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 6331 <.literal16+0xcc1> + DB 112,65 ; jo 6361 <.literal16+0xcc1> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -25679,11 +25759,11 @@ ALIGN 16 DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,0,0,127,67 ; addb $0x43,0x7f00003b(%rax) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 637b <.literal16+0xd0b> + DB 127,67 ; jg 63ab <.literal16+0xd0b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 637f <.literal16+0xd0f> + DB 127,67 ; jg 63af <.literal16+0xd0f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 6383 <.literal16+0xd13> + DB 127,67 ; jg 63b3 <.literal16+0xd13> DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) @@ -25759,13 +25839,13 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 646b <.literal16+0xdfb> + DB 127,71 ; jg 649b <.literal16+0xdfb> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 646f <.literal16+0xdff> + DB 127,71 ; jg 649f <.literal16+0xdff> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6473 <.literal16+0xe03> + DB 127,71 ; jg 64a3 <.literal16+0xe03> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6477 <.literal16+0xe07> + DB 127,71 ; jg 64a7 <.literal16+0xe07> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -25842,10 +25922,10 @@ ALIGN 16 DB 61,152,221,147,61 ; cmp $0x3d93dd98,%eax DB 152 ; cwtl DB 221,147,61,45,16,17 ; fstl 0x11102d3d(%rbx) - DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc0760a <_sk_callback_sse2+0x2dc02064> + DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc0763a <_sk_callback_sse2+0x2dc02060> DB 17,192 ; adc %eax,%eax DB 45,16,17,192,18 ; sub $0x12c01110,%eax - DB 120,57 ; js 653c <.literal16+0xecc> + DB 120,57 ; js 656c <.literal16+0xecc> DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil @@ -25973,11 +26053,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6682 <.literal16+0x1012> + DB 62,114,28 ; jb,pt 66b2 <.literal16+0x1012> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6686 <.literal16+0x1016> + DB 62,114,28 ; jb,pt 66b6 <.literal16+0x1016> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 668a <.literal16+0x101a> + DB 62,114,28 ; jb,pt 66ba <.literal16+0x101a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26021,7 +26101,7 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f515 <_sk_callback_sse2+0x3d639f6f> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f545 <_sk_callback_sse2+0x3d639f6b> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -26047,7 +26127,7 @@ ALIGN 16 DB 0,192 ; add %al,%al DB 63 ; (bad) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f555 <_sk_callback_sse2+0x3d639faf> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f585 <_sk_callback_sse2+0x3d639fab> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -26056,13 +26136,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 674e <.literal16+0x10de> + DB 114,28 ; jb 677e <.literal16+0x10de> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6752 <.literal16+0x10e2> + DB 62,114,28 ; jb,pt 6782 <.literal16+0x10e2> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6756 <.literal16+0x10e6> + DB 62,114,28 ; jb,pt 6786 <.literal16+0x10e6> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 675a <.literal16+0x10ea> + DB 62,114,28 ; jb,pt 678a <.literal16+0x10ea> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26083,11 +26163,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6792 <.literal16+0x1122> + DB 62,114,28 ; jb,pt 67c2 <.literal16+0x1122> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6796 <.literal16+0x1126> + DB 62,114,28 ; jb,pt 67c6 <.literal16+0x1126> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 679a <.literal16+0x112a> + DB 62,114,28 ; jb,pt 67ca <.literal16+0x112a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26131,7 +26211,7 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f625 <_sk_callback_sse2+0x3d63a07f> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f655 <_sk_callback_sse2+0x3d63a07b> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -26157,7 +26237,7 @@ ALIGN 16 DB 0,192 ; add %al,%al DB 63 ; (bad) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f665 <_sk_callback_sse2+0x3d63a0bf> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f695 <_sk_callback_sse2+0x3d63a0bb> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -26166,13 +26246,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 685e <.literal16+0x11ee> + DB 114,28 ; jb 688e <.literal16+0x11ee> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6862 <_sk_callback_sse2+0x12bc> + DB 62,114,28 ; jb,pt 6892 <_sk_callback_sse2+0x12b8> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6866 <_sk_callback_sse2+0x12c0> + DB 62,114,28 ; jb,pt 6896 <_sk_callback_sse2+0x12bc> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 686a <_sk_callback_sse2+0x12c4> + DB 62,114,28 ; jb,pt 689a <_sk_callback_sse2+0x12c0> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) |