diff options
| author |  Mike Klein <mtklein@chromium.org> | 2017-05-01 14:22:10 -0400 |
|---|---|---|
| committer |  Skia Commit-Bot <skia-commit-bot@chromium.org> | 2017-05-01 18:52:53 +0000 |
| commit | fb11acdeefce2fe24e98dc4e4c1d002291604119 (patch) | |
| tree | 5f0c6ee439cca766fc2c7819ec1e04ee81b7452e /src | |
| parent | 7575bb1c389f96db4123783fcd717f3611b3a8eb (diff) | |
getting close on float constants
I think I'm now down to just the constants used in comparisons in
hsl_to_rgb... so weird. I don't get what's special about this code.
CQ_INCLUDE_TRYBOTS=skia.primary:Test-Win7-MSVC-Golo-CPU-AVX-x86_64-Release,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Release,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Release-SK_CPU_LIMIT_SSE41,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Release-SK_CPU_LIMIT_SSE2
Change-Id: I398139f00371bdbc45281a49bca56d02ba59cba9
Reviewed-on: https://skia-review.googlesource.com/14909
Reviewed-by: Mike Klein <mtklein@chromium.org>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src')
| -rw-r--r-- | src/jumper/SkJumper_generated.S | 3943 | ||||
| -rw-r--r-- | src/jumper/SkJumper_generated_win.S | 3944 | ||||
| -rw-r--r-- | src/jumper/SkJumper_stages.cpp | 18 |
3 files changed, 4061 insertions, 3844 deletions
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index 0932fd0be5..9aed8908d5 100644 --- a/src/jumper/SkJumper_generated.S +++ b/src/jumper/SkJumper_generated.S @@ -7347,14 +7347,14 @@ _sk_seed_shader_hsw: .byte 197,249,110,199 // vmovd %edi,%xmm0 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,125,60,0,0 // vbroadcastss 0x3c7d(%rip),%ymm1 # 3d40 <_sk_callback_hsw+0x125> + .byte 196,226,125,24,13,173,60,0,0 // vbroadcastss 0x3cad(%rip),%ymm1 # 3d70 <_sk_callback_hsw+0x125> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,21,97,60,0,0 // vbroadcastss 0x3c61(%rip),%ymm2 # 3d44 <_sk_callback_hsw+0x129> + .byte 196,226,125,24,21,145,60,0,0 // vbroadcastss 0x3c91(%rip),%ymm2 # 3d74 <_sk_callback_hsw+0x129> .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 @@ -7390,7 +7390,7 @@ HIDDEN _sk_srcatop_hsw FUNCTION(_sk_srcatop_hsw) _sk_srcatop_hsw: .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 - .byte 196,98,125,24,5,17,60,0,0 // vbroadcastss 0x3c11(%rip),%ymm8 # 3d48 <_sk_callback_hsw+0x12d> + .byte 196,98,125,24,5,65,60,0,0 // vbroadcastss 0x3c41(%rip),%ymm8 # 3d78 <_sk_callback_hsw+0x12d> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,226,61,184,196 // vfmadd231ps %ymm4,%ymm8,%ymm0 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 @@ -7406,7 +7406,7 @@ HIDDEN _sk_dstatop_hsw .globl _sk_dstatop_hsw FUNCTION(_sk_dstatop_hsw) _sk_dstatop_hsw: - .byte 196,98,125,24,5,228,59,0,0 // vbroadcastss 0x3be4(%rip),%ymm8 # 3d4c <_sk_callback_hsw+0x131> + .byte 196,98,125,24,5,20,60,0,0 // vbroadcastss 0x3c14(%rip),%ymm8 # 3d7c <_sk_callback_hsw+0x131> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 196,226,101,184,196 // vfmadd231ps %ymm4,%ymm3,%ymm0 @@ -7445,7 +7445,7 @@ HIDDEN _sk_srcout_hsw .globl _sk_srcout_hsw FUNCTION(_sk_srcout_hsw) _sk_srcout_hsw: - .byte 196,98,125,24,5,139,59,0,0 // vbroadcastss 0x3b8b(%rip),%ymm8 # 3d50 <_sk_callback_hsw+0x135> + .byte 196,98,125,24,5,187,59,0,0 // vbroadcastss 0x3bbb(%rip),%ymm8 # 3d80 <_sk_callback_hsw+0x135> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -7458,7 +7458,7 @@ HIDDEN _sk_dstout_hsw .globl _sk_dstout_hsw FUNCTION(_sk_dstout_hsw) _sk_dstout_hsw: - .byte 196,226,125,24,5,110,59,0,0 // vbroadcastss 0x3b6e(%rip),%ymm0 # 3d54 <_sk_callback_hsw+0x139> + .byte 196,226,125,24,5,158,59,0,0 // vbroadcastss 0x3b9e(%rip),%ymm0 # 3d84 <_sk_callback_hsw+0x139> .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 @@ -7471,7 +7471,7 @@ HIDDEN _sk_srcover_hsw .globl _sk_srcover_hsw FUNCTION(_sk_srcover_hsw) _sk_srcover_hsw: - .byte 196,98,125,24,5,81,59,0,0 // vbroadcastss 0x3b51(%rip),%ymm8 # 3d58 <_sk_callback_hsw+0x13d> + .byte 196,98,125,24,5,129,59,0,0 // vbroadcastss 0x3b81(%rip),%ymm8 # 3d88 <_sk_callback_hsw+0x13d> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0 .byte 196,194,85,184,200 // vfmadd231ps %ymm8,%ymm5,%ymm1 @@ -7484,7 +7484,7 @@ HIDDEN _sk_dstover_hsw .globl _sk_dstover_hsw FUNCTION(_sk_dstover_hsw) _sk_dstover_hsw: - .byte 196,98,125,24,5,48,59,0,0 // vbroadcastss 0x3b30(%rip),%ymm8 # 3d5c <_sk_callback_hsw+0x141> + .byte 196,98,125,24,5,96,59,0,0 // vbroadcastss 0x3b60(%rip),%ymm8 # 3d8c <_sk_callback_hsw+0x141> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 @@ -7508,7 +7508,7 @@ HIDDEN _sk_multiply_hsw .globl _sk_multiply_hsw FUNCTION(_sk_multiply_hsw) _sk_multiply_hsw: - .byte 196,98,125,24,5,251,58,0,0 // vbroadcastss 0x3afb(%rip),%ymm8 # 3d60 <_sk_callback_hsw+0x145> + .byte 196,98,125,24,5,43,59,0,0 // vbroadcastss 0x3b2b(%rip),%ymm8 # 3d90 <_sk_callback_hsw+0x145> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -7556,7 +7556,7 @@ HIDDEN _sk_xor__hsw .globl _sk_xor__hsw FUNCTION(_sk_xor__hsw) _sk_xor__hsw: - .byte 196,98,125,24,5,118,58,0,0 // vbroadcastss 0x3a76(%rip),%ymm8 # 3d64 <_sk_callback_hsw+0x149> + .byte 196,98,125,24,5,166,58,0,0 // vbroadcastss 0x3aa6(%rip),%ymm8 # 3d94 <_sk_callback_hsw+0x149> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -7590,7 +7590,7 @@ _sk_darken_hsw: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,254,57,0,0 // vbroadcastss 0x39fe(%rip),%ymm8 # 3d68 <_sk_callback_hsw+0x14d> + .byte 196,98,125,24,5,46,58,0,0 // vbroadcastss 0x3a2e(%rip),%ymm8 # 3d98 <_sk_callback_hsw+0x14d> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -7615,7 +7615,7 @@ _sk_lighten_hsw: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,173,57,0,0 // vbroadcastss 0x39ad(%rip),%ymm8 # 3d6c <_sk_callback_hsw+0x151> + .byte 196,98,125,24,5,221,57,0,0 // vbroadcastss 0x39dd(%rip),%ymm8 # 3d9c <_sk_callback_hsw+0x151> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -7643,7 +7643,7 @@ _sk_difference_hsw: .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,80,57,0,0 // vbroadcastss 0x3950(%rip),%ymm8 # 3d70 <_sk_callback_hsw+0x155> + .byte 196,98,125,24,5,128,57,0,0 // vbroadcastss 0x3980(%rip),%ymm8 # 3da0 <_sk_callback_hsw+0x155> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -7665,7 +7665,7 @@ _sk_exclusion_hsw: .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,14,57,0,0 // vbroadcastss 0x390e(%rip),%ymm8 # 3d74 <_sk_callback_hsw+0x159> + .byte 196,98,125,24,5,62,57,0,0 // vbroadcastss 0x393e(%rip),%ymm8 # 3da4 <_sk_callback_hsw+0x159> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -7675,7 +7675,7 @@ HIDDEN _sk_colorburn_hsw .globl _sk_colorburn_hsw FUNCTION(_sk_colorburn_hsw) _sk_colorburn_hsw: - .byte 196,98,125,24,5,252,56,0,0 // vbroadcastss 0x38fc(%rip),%ymm8 # 3d78 <_sk_callback_hsw+0x15d> + .byte 196,98,125,24,5,44,57,0,0 // vbroadcastss 0x392c(%rip),%ymm8 # 3da8 <_sk_callback_hsw+0x15d> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 @@ -7733,7 +7733,7 @@ HIDDEN _sk_colordodge_hsw FUNCTION(_sk_colordodge_hsw) _sk_colordodge_hsw: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 196,98,125,24,13,7,56,0,0 // vbroadcastss 0x3807(%rip),%ymm9 # 3d7c <_sk_callback_hsw+0x161> + .byte 196,98,125,24,13,55,56,0,0 // vbroadcastss 0x3837(%rip),%ymm9 # 3dac <_sk_callback_hsw+0x161> .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 @@ -7786,7 +7786,7 @@ HIDDEN _sk_hardlight_hsw .globl _sk_hardlight_hsw FUNCTION(_sk_hardlight_hsw) _sk_hardlight_hsw: - .byte 196,98,125,24,5,40,55,0,0 // vbroadcastss 0x3728(%rip),%ymm8 # 3d80 <_sk_callback_hsw+0x165> + .byte 196,98,125,24,5,88,55,0,0 // vbroadcastss 0x3758(%rip),%ymm8 # 3db0 <_sk_callback_hsw+0x165> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -7837,7 +7837,7 @@ HIDDEN _sk_overlay_hsw .globl _sk_overlay_hsw FUNCTION(_sk_overlay_hsw) _sk_overlay_hsw: - .byte 196,98,125,24,5,96,54,0,0 // vbroadcastss 0x3660(%rip),%ymm8 # 3d84 <_sk_callback_hsw+0x169> + .byte 196,98,125,24,5,144,54,0,0 // vbroadcastss 0x3690(%rip),%ymm8 # 3db4 <_sk_callback_hsw+0x169> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -7898,10 +7898,10 @@ _sk_softlight_hsw: .byte 196,65,20,88,197 // vaddps %ymm13,%ymm13,%ymm8 .byte 196,65,60,88,192 // vaddps %ymm8,%ymm8,%ymm8 .byte 196,66,61,168,192 // vfmadd213ps %ymm8,%ymm8,%ymm8 - .byte 196,98,125,24,29,107,53,0,0 // vbroadcastss 0x356b(%rip),%ymm11 # 3d8c <_sk_callback_hsw+0x171> + .byte 196,98,125,24,29,155,53,0,0 // vbroadcastss 0x359b(%rip),%ymm11 # 3dbc <_sk_callback_hsw+0x171> .byte 196,65,20,88,227 // vaddps %ymm11,%ymm13,%ymm12 .byte 196,65,28,89,192 // vmulps %ymm8,%ymm12,%ymm8 - .byte 196,98,125,24,37,92,53,0,0 // vbroadcastss 0x355c(%rip),%ymm12 # 3d90 <_sk_callback_hsw+0x175> + .byte 196,98,125,24,37,140,53,0,0 // vbroadcastss 0x358c(%rip),%ymm12 # 3dc0 <_sk_callback_hsw+0x175> .byte 196,66,21,184,196 // vfmadd231ps %ymm12,%ymm13,%ymm8 .byte 196,65,124,82,245 // vrsqrtps %ymm13,%ymm14 .byte 196,65,124,83,246 // vrcpps %ymm14,%ymm14 @@ -7911,7 +7911,7 @@ _sk_softlight_hsw: .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15 .byte 196,67,13,74,240,240 // vblendvps %ymm15,%ymm8,%ymm14,%ymm14 .byte 197,116,88,249 // vaddps %ymm1,%ymm1,%ymm15 - .byte 196,98,125,24,5,31,53,0,0 // vbroadcastss 0x351f(%rip),%ymm8 # 3d88 <_sk_callback_hsw+0x16d> + .byte 196,98,125,24,5,79,53,0,0 // vbroadcastss 0x354f(%rip),%ymm8 # 3db8 <_sk_callback_hsw+0x16d> .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13 .byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0 .byte 196,98,125,168,235 // vfmadd213ps %ymm3,%ymm0,%ymm13 @@ -8004,7 +8004,7 @@ HIDDEN _sk_clamp_1_hsw .globl _sk_clamp_1_hsw FUNCTION(_sk_clamp_1_hsw) _sk_clamp_1_hsw: - .byte 196,98,125,24,5,164,51,0,0 // vbroadcastss 0x33a4(%rip),%ymm8 # 3d94 <_sk_callback_hsw+0x179> + .byte 196,98,125,24,5,212,51,0,0 // vbroadcastss 0x33d4(%rip),%ymm8 # 3dc4 <_sk_callback_hsw+0x179> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 @@ -8016,7 +8016,7 @@ HIDDEN _sk_clamp_a_hsw .globl _sk_clamp_a_hsw FUNCTION(_sk_clamp_a_hsw) _sk_clamp_a_hsw: - .byte 196,98,125,24,5,135,51,0,0 // vbroadcastss 0x3387(%rip),%ymm8 # 3d98 <_sk_callback_hsw+0x17d> + .byte 196,98,125,24,5,183,51,0,0 // vbroadcastss 0x33b7(%rip),%ymm8 # 3dc8 <_sk_callback_hsw+0x17d> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 @@ -8102,7 +8102,7 @@ FUNCTION(_sk_unpremul_hsw) _sk_unpremul_hsw: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 - .byte 196,98,125,24,21,207,50,0,0 // vbroadcastss 0x32cf(%rip),%ymm10 # 3d9c <_sk_callback_hsw+0x181> + .byte 196,98,125,24,21,255,50,0,0 // vbroadcastss 0x32ff(%rip),%ymm10 # 3dcc <_sk_callback_hsw+0x181> .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 @@ -8115,16 +8115,16 @@ HIDDEN _sk_from_srgb_hsw .globl _sk_from_srgb_hsw FUNCTION(_sk_from_srgb_hsw) _sk_from_srgb_hsw: - .byte 196,98,125,24,5,176,50,0,0 // vbroadcastss 0x32b0(%rip),%ymm8 # 3da0 <_sk_callback_hsw+0x185> + .byte 196,98,125,24,5,224,50,0,0 // vbroadcastss 0x32e0(%rip),%ymm8 # 3dd0 <_sk_callback_hsw+0x185> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 - .byte 196,98,125,24,29,162,50,0,0 // vbroadcastss 0x32a2(%rip),%ymm11 # 3da4 <_sk_callback_hsw+0x189> - .byte 196,98,125,24,37,157,50,0,0 // vbroadcastss 0x329d(%rip),%ymm12 # 3da8 <_sk_callback_hsw+0x18d> + .byte 196,98,125,24,29,210,50,0,0 // vbroadcastss 0x32d2(%rip),%ymm11 # 3dd4 <_sk_callback_hsw+0x189> + .byte 196,98,125,24,37,205,50,0,0 // vbroadcastss 0x32cd(%rip),%ymm12 # 3dd8 <_sk_callback_hsw+0x18d> .byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13 .byte 196,66,125,168,235 // vfmadd213ps %ymm11,%ymm0,%ymm13 - .byte 196,98,125,24,53,142,50,0,0 // vbroadcastss 0x328e(%rip),%ymm14 # 3dac <_sk_callback_hsw+0x191> + .byte 196,98,125,24,53,190,50,0,0 // vbroadcastss 0x32be(%rip),%ymm14 # 3ddc <_sk_callback_hsw+0x191> .byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13 - .byte 196,98,125,24,21,132,50,0,0 // vbroadcastss 0x3284(%rip),%ymm10 # 3db0 <_sk_callback_hsw+0x195> + .byte 196,98,125,24,21,180,50,0,0 // vbroadcastss 0x32b4(%rip),%ymm10 # 3de0 <_sk_callback_hsw+0x195> .byte 196,193,124,194,194,1 // vcmpltps %ymm10,%ymm0,%ymm0 .byte 196,195,21,74,193,0 // vblendvps %ymm0,%ymm9,%ymm13,%ymm0 .byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9 @@ -8150,16 +8150,16 @@ _sk_to_srgb_hsw: .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 .byte 196,65,124,83,200 // vrcpps %ymm8,%ymm9 .byte 196,65,124,82,208 // vrsqrtps %ymm8,%ymm10 - .byte 196,98,125,24,5,30,50,0,0 // vbroadcastss 0x321e(%rip),%ymm8 # 3db4 <_sk_callback_hsw+0x199> + .byte 196,98,125,24,5,78,50,0,0 // vbroadcastss 0x324e(%rip),%ymm8 # 3de4 <_sk_callback_hsw+0x199> .byte 196,65,124,89,216 // vmulps %ymm8,%ymm0,%ymm11 - .byte 196,98,125,24,37,20,50,0,0 // vbroadcastss 0x3214(%rip),%ymm12 # 3db8 <_sk_callback_hsw+0x19d> - .byte 196,98,125,24,45,15,50,0,0 // vbroadcastss 0x320f(%rip),%ymm13 # 3dbc <_sk_callback_hsw+0x1a1> + .byte 196,98,125,24,37,68,50,0,0 // vbroadcastss 0x3244(%rip),%ymm12 # 3de8 <_sk_callback_hsw+0x19d> + .byte 196,98,125,24,45,63,50,0,0 // vbroadcastss 0x323f(%rip),%ymm13 # 3dec <_sk_callback_hsw+0x1a1> .byte 196,66,21,168,204 // vfmadd213ps %ymm12,%ymm13,%ymm9 - .byte 196,98,125,24,53,5,50,0,0 // vbroadcastss 0x3205(%rip),%ymm14 # 3dc0 <_sk_callback_hsw+0x1a5> + .byte 196,98,125,24,53,53,50,0,0 // vbroadcastss 0x3235(%rip),%ymm14 # 3df0 <_sk_callback_hsw+0x1a5> .byte 196,66,13,184,202 // vfmadd231ps %ymm10,%ymm14,%ymm9 - .byte 196,98,125,24,21,251,49,0,0 // vbroadcastss 0x31fb(%rip),%ymm10 # 3dc4 <_sk_callback_hsw+0x1a9> + .byte 196,98,125,24,21,43,50,0,0 // vbroadcastss 0x322b(%rip),%ymm10 # 3df4 <_sk_callback_hsw+0x1a9> .byte 196,65,44,93,201 // vminps %ymm9,%ymm10,%ymm9 - .byte 196,98,125,24,61,241,49,0,0 // vbroadcastss 0x31f1(%rip),%ymm15 # 3dc8 <_sk_callback_hsw+0x1ad> + .byte 196,98,125,24,61,33,50,0,0 // vbroadcastss 0x3221(%rip),%ymm15 # 3df8 <_sk_callback_hsw+0x1ad> .byte 196,193,124,194,199,1 // vcmpltps %ymm15,%ymm0,%ymm0 .byte 196,195,53,74,195,0 // vblendvps %ymm0,%ymm11,%ymm9,%ymm0 .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 @@ -8192,26 +8192,26 @@ _sk_rgb_to_hsl_hsw: .byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9 .byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9 .byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,29,107,49,0,0 // vbroadcastss 0x316b(%rip),%ymm11 # 3dcc <_sk_callback_hsw+0x1b1> + .byte 196,98,125,24,29,155,49,0,0 // vbroadcastss 0x319b(%rip),%ymm11 # 3dfc <_sk_callback_hsw+0x1b1> .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11 .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12 .byte 197,116,194,234,1 // vcmpltps %ymm2,%ymm1,%ymm13 - .byte 196,98,125,24,53,88,49,0,0 // vbroadcastss 0x3158(%rip),%ymm14 # 3dd0 <_sk_callback_hsw+0x1b5> + .byte 196,98,125,24,53,136,49,0,0 // vbroadcastss 0x3188(%rip),%ymm14 # 3e00 <_sk_callback_hsw+0x1b5> .byte 196,65,4,87,255 // vxorps %ymm15,%ymm15,%ymm15 .byte 196,67,5,74,238,208 // vblendvps %ymm13,%ymm14,%ymm15,%ymm13 .byte 196,66,37,168,229 // vfmadd213ps %ymm13,%ymm11,%ymm12 .byte 197,236,92,208 // vsubps %ymm0,%ymm2,%ymm2 .byte 197,124,92,233 // vsubps %ymm1,%ymm0,%ymm13 - .byte 196,98,125,24,53,63,49,0,0 // vbroadcastss 0x313f(%rip),%ymm14 # 3dd8 <_sk_callback_hsw+0x1bd> + .byte 196,98,125,24,53,111,49,0,0 // vbroadcastss 0x316f(%rip),%ymm14 # 3e08 <_sk_callback_hsw+0x1bd> .byte 196,66,37,168,238 // vfmadd213ps %ymm14,%ymm11,%ymm13 - .byte 196,98,125,24,53,45,49,0,0 // vbroadcastss 0x312d(%rip),%ymm14 # 3dd4 <_sk_callback_hsw+0x1b9> + .byte 196,98,125,24,53,93,49,0,0 // vbroadcastss 0x315d(%rip),%ymm14 # 3e04 <_sk_callback_hsw+0x1b9> .byte 196,194,37,168,214 // vfmadd213ps %ymm14,%ymm11,%ymm2 .byte 197,188,194,201,0 // vcmpeqps %ymm1,%ymm8,%ymm1 .byte 196,227,21,74,202,16 // vblendvps %ymm1,%ymm2,%ymm13,%ymm1 .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0 .byte 196,195,117,74,196,0 // vblendvps %ymm0,%ymm12,%ymm1,%ymm0 .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1 - .byte 196,98,125,24,29,16,49,0,0 // vbroadcastss 0x3110(%rip),%ymm11 # 3de0 <_sk_callback_hsw+0x1c5> + .byte 196,98,125,24,29,64,49,0,0 // vbroadcastss 0x3140(%rip),%ymm11 # 3e10 <_sk_callback_hsw+0x1c5> .byte 196,193,116,89,211 // vmulps %ymm11,%ymm1,%ymm2 .byte 197,36,194,218,1 // vcmpltps %ymm2,%ymm11,%ymm11 .byte 196,65,12,92,224 // vsubps %ymm8,%ymm14,%ymm12 @@ -8221,7 +8221,7 @@ _sk_rgb_to_hsl_hsw: .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 .byte 196,195,125,74,199,128 // vblendvps %ymm8,%ymm15,%ymm0,%ymm0 .byte 196,195,117,74,207,128 // vblendvps %ymm8,%ymm15,%ymm1,%ymm1 - .byte 196,98,125,24,5,211,48,0,0 // vbroadcastss 0x30d3(%rip),%ymm8 # 3ddc <_sk_callback_hsw+0x1c1> + .byte 196,98,125,24,5,3,49,0,0 // vbroadcastss 0x3103(%rip),%ymm8 # 3e0c <_sk_callback_hsw+0x1c1> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -8230,105 +8230,108 @@ HIDDEN _sk_hsl_to_rgb_hsw .globl _sk_hsl_to_rgb_hsw FUNCTION(_sk_hsl_to_rgb_hsw) _sk_hsl_to_rgb_hsw: - .byte 72,131,236,56 // sub $0x38,%rsp - .byte 197,252,17,60,36 // vmovups %ymm7,(%rsp) - .byte 197,252,17,116,36,224 // vmovups %ymm6,-0x20(%rsp) - .byte 197,252,17,108,36,192 // vmovups %ymm5,-0x40(%rsp) - .byte 197,252,17,100,36,160 // vmovups %ymm4,-0x60(%rsp) - .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) - .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 + .byte 72,131,236,120 // sub $0x78,%rsp + .byte 197,252,17,124,36,64 // vmovups %ymm7,0x40(%rsp) + .byte 197,252,17,116,36,32 // vmovups %ymm6,0x20(%rsp) + .byte 197,252,17,44,36 // vmovups %ymm5,(%rsp) + .byte 197,252,17,100,36,224 // vmovups %ymm4,-0x20(%rsp) + .byte 197,252,17,92,36,192 // vmovups %ymm3,-0x40(%rsp) + .byte 197,252,40,234 // vmovaps %ymm2,%ymm5 + .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 184,0,0,0,63 // mov $0x3f000000,%eax - .byte 197,249,110,200 // vmovd %eax,%xmm1 - .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 - .byte 196,193,108,194,200,1 // vcmpltps %ymm8,%ymm2,%ymm1 + .byte 197,249,110,192 // vmovd %eax,%xmm0 + .byte 196,98,125,88,192 // vpbroadcastd %xmm0,%ymm8 + .byte 196,193,84,194,192,1 // vcmpltps %ymm8,%ymm5,%ymm0 + .byte 196,98,125,24,21,188,48,0,0 // vbroadcastss 0x30bc(%rip),%ymm10 # 3e14 <_sk_callback_hsw+0x1c9> + .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp) + .byte 196,193,116,88,218 // vaddps %ymm10,%ymm1,%ymm3 + .byte 197,228,89,221 // vmulps %ymm5,%ymm3,%ymm3 + .byte 197,244,88,229 // vaddps %ymm5,%ymm1,%ymm4 + .byte 196,226,117,188,229 // vfnmadd231ps %ymm5,%ymm1,%ymm4 + .byte 196,99,93,74,203,0 // vblendvps %ymm0,%ymm3,%ymm4,%ymm9 + .byte 196,226,125,24,13,157,48,0,0 // vbroadcastss 0x309d(%rip),%ymm1 # 3e1c <_sk_callback_hsw+0x1d1> + .byte 197,236,88,241 // vaddps %ymm1,%ymm2,%ymm6 + .byte 65,184,0,0,0,0 // mov $0x0,%r8d .byte 184,0,0,128,63 // mov $0x3f800000,%eax - .byte 197,249,110,224 // vmovd %eax,%xmm4 - .byte 196,98,125,88,212 // vpbroadcastd %xmm4,%ymm10 - .byte 197,172,88,229 // vaddps %ymm5,%ymm10,%ymm4 - .byte 197,220,89,226 // vmulps %ymm2,%ymm4,%ymm4 - .byte 197,84,88,202 // vaddps %ymm2,%ymm5,%ymm9 - .byte 196,98,85,188,202 // vfnmadd231ps %ymm2,%ymm5,%ymm9 - .byte 196,99,53,74,204,16 // vblendvps %ymm1,%ymm4,%ymm9,%ymm9 - .byte 65,184,0,0,0,64 // mov $0x40000000,%r8d - .byte 184,171,170,170,62 // mov $0x3eaaaaab,%eax - .byte 197,249,110,200 // vmovd %eax,%xmm1 - .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 - .byte 197,148,88,224 // vaddps %ymm0,%ymm13,%ymm4 - .byte 184,0,0,0,0 // mov $0x0,%eax .byte 197,249,110,200 // vmovd %eax,%xmm1 .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12 - .byte 197,172,194,204,1 // vcmpltps %ymm4,%ymm10,%ymm1 - .byte 196,65,92,92,218 // vsubps %ymm10,%ymm4,%ymm11 - .byte 196,195,93,74,203,16 // vblendvps %ymm1,%ymm11,%ymm4,%ymm1 - .byte 196,65,92,194,220,1 // vcmpltps %ymm12,%ymm4,%ymm11 - .byte 197,44,88,244 // vaddps %ymm4,%ymm10,%ymm14 - .byte 196,195,117,74,206,176 // vblendvps %ymm11,%ymm14,%ymm1,%ymm1 + .byte 197,156,194,206,1 // vcmpltps %ymm6,%ymm12,%ymm1 + .byte 196,98,125,24,45,123,48,0,0 // vbroadcastss 0x307b(%rip),%ymm13 # 3e20 <_sk_callback_hsw+0x1d5> + .byte 196,193,76,88,221 // vaddps %ymm13,%ymm6,%ymm3 + .byte 196,227,77,74,203,16 // vblendvps %ymm1,%ymm3,%ymm6,%ymm1 .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 - .byte 196,66,109,170,217 // vfmsub213ps %ymm9,%ymm2,%ymm11 + .byte 196,98,125,88,251 // vpbroadcastd %xmm3,%ymm15 + .byte 196,193,76,194,223,1 // vcmpltps %ymm15,%ymm6,%ymm3 + .byte 196,193,76,88,226 // vaddps %ymm10,%ymm6,%ymm4 + .byte 196,227,117,74,196,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm0 + .byte 196,98,125,24,29,68,48,0,0 // vbroadcastss 0x3044(%rip),%ymm11 # 3e18 <_sk_callback_hsw+0x1cd> + .byte 196,66,85,170,217 // vfmsub213ps %ymm9,%ymm5,%ymm11 + .byte 196,193,52,92,203 // vsubps %ymm11,%ymm9,%ymm1 + .byte 196,226,125,24,29,61,48,0,0 // vbroadcastss 0x303d(%rip),%ymm3 # 3e24 <_sk_callback_hsw+0x1d9> + .byte 197,116,89,243 // vmulps %ymm3,%ymm1,%ymm14 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d - .byte 184,0,0,192,64 // mov $0x40c00000,%eax - .byte 197,249,110,216 // vmovd %eax,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 196,65,52,92,243 // vsubps %ymm11,%ymm9,%ymm14 - .byte 197,12,89,243 // vmulps %ymm3,%ymm14,%ymm14 .byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax - .byte 197,249,110,216 // vmovd %eax,%xmm3 - .byte 196,98,125,88,251 // vpbroadcastd %xmm3,%ymm15 - .byte 197,132,92,217 // vsubps %ymm1,%ymm15,%ymm3 - .byte 196,194,13,168,219 // vfmadd213ps %ymm11,%ymm14,%ymm3 - .byte 196,193,116,194,255,1 // vcmpltps %ymm15,%ymm1,%ymm7 - .byte 196,227,37,74,219,112 // vblendvps %ymm7,%ymm3,%ymm11,%ymm3 - .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 - .byte 196,195,101,74,249,112 // vblendvps %ymm7,%ymm9,%ymm3,%ymm7 - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 197,244,194,203,1 // vcmpltps %ymm3,%ymm1,%ymm1 - .byte 196,194,13,168,227 // vfmadd213ps %ymm11,%ymm14,%ymm4 - .byte 196,227,69,74,228,16 // vblendvps %ymm1,%ymm4,%ymm7,%ymm4 - .byte 197,172,194,200,1 // vcmpltps %ymm0,%ymm10,%ymm1 - .byte 196,193,124,92,250 // vsubps %ymm10,%ymm0,%ymm7 - .byte 196,227,125,74,207,16 // vblendvps %ymm1,%ymm7,%ymm0,%ymm1 - .byte 196,193,124,194,252,1 // vcmpltps %ymm12,%ymm0,%ymm7 - .byte 197,172,88,240 // vaddps %ymm0,%ymm10,%ymm6 - .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1 - .byte 197,132,92,241 // vsubps %ymm1,%ymm15,%ymm6 + .byte 197,249,110,200 // vmovd %eax,%xmm1 + .byte 196,226,125,88,225 // vpbroadcastd %xmm1,%ymm4 + .byte 196,226,125,24,29,32,48,0,0 // vbroadcastss 0x3020(%rip),%ymm3 # 3e28 <_sk_callback_hsw+0x1dd> + .byte 197,228,92,200 // vsubps %ymm0,%ymm3,%ymm1 + .byte 196,194,13,168,203 // vfmadd213ps %ymm11,%ymm14,%ymm1 + .byte 197,252,194,252,1 // vcmpltps %ymm4,%ymm0,%ymm7 + .byte 196,227,37,74,201,112 // vblendvps %ymm7,%ymm1,%ymm11,%ymm1 + .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7 + .byte 196,195,117,74,249,112 // vblendvps %ymm7,%ymm9,%ymm1,%ymm7 + .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 + .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 + .byte 197,252,194,193,1 // vcmpltps %ymm1,%ymm0,%ymm0 .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6 - .byte 196,193,116,194,255,1 // vcmpltps %ymm15,%ymm1,%ymm7 + .byte 196,227,69,74,198,0 // vblendvps %ymm0,%ymm6,%ymm7,%ymm0 + .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp) + .byte 197,156,194,194,1 // vcmpltps %ymm2,%ymm12,%ymm0 + .byte 196,193,108,88,253 // vaddps %ymm13,%ymm2,%ymm7 + .byte 196,227,109,74,199,0 // vblendvps %ymm0,%ymm7,%ymm2,%ymm0 + .byte 196,193,108,194,255,1 // vcmpltps %ymm15,%ymm2,%ymm7 + .byte 196,193,108,88,242 // vaddps %ymm10,%ymm2,%ymm6 + .byte 196,227,125,74,198,112 // vblendvps %ymm7,%ymm6,%ymm0,%ymm0 + .byte 197,228,92,240 // vsubps %ymm0,%ymm3,%ymm6 + .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6 + .byte 197,252,194,252,1 // vcmpltps %ymm4,%ymm0,%ymm7 .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6 - .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 + .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7 .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6 - .byte 197,244,194,203,1 // vcmpltps %ymm3,%ymm1,%ymm1 - .byte 196,193,124,92,253 // vsubps %ymm13,%ymm0,%ymm7 - .byte 196,194,13,168,195 // vfmadd213ps %ymm11,%ymm14,%ymm0 - .byte 196,227,77,74,200,16 // vblendvps %ymm1,%ymm0,%ymm6,%ymm1 - .byte 197,172,194,199,1 // vcmpltps %ymm7,%ymm10,%ymm0 - .byte 196,193,68,92,242 // vsubps %ymm10,%ymm7,%ymm6 - .byte 196,227,69,74,198,0 // vblendvps %ymm0,%ymm6,%ymm7,%ymm0 - .byte 196,193,68,194,244,1 // vcmpltps %ymm12,%ymm7,%ymm6 - .byte 197,44,88,215 // vaddps %ymm7,%ymm10,%ymm10 - .byte 196,195,125,74,194,96 // vblendvps %ymm6,%ymm10,%ymm0,%ymm0 + .byte 197,252,194,193,1 // vcmpltps %ymm1,%ymm0,%ymm0 + .byte 197,252,40,250 // vmovaps %ymm2,%ymm7 .byte 196,194,13,168,251 // vfmadd213ps %ymm11,%ymm14,%ymm7 - .byte 197,132,92,240 // vsubps %ymm0,%ymm15,%ymm6 - .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6 - .byte 196,65,124,194,215,1 // vcmpltps %ymm15,%ymm0,%ymm10 - .byte 196,227,37,74,246,160 // vblendvps %ymm10,%ymm6,%ymm11,%ymm6 - .byte 196,65,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm8 - .byte 196,195,77,74,241,128 // vblendvps %ymm8,%ymm9,%ymm6,%ymm6 - .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0 - .byte 196,227,77,74,223,0 // vblendvps %ymm0,%ymm7,%ymm6,%ymm3 + .byte 196,227,77,74,247,0 // vblendvps %ymm0,%ymm7,%ymm6,%ymm6 + .byte 196,226,125,24,5,134,47,0,0 // vbroadcastss 0x2f86(%rip),%ymm0 # 3e2c <_sk_callback_hsw+0x1e1> + .byte 197,236,88,192 // vaddps %ymm0,%ymm2,%ymm0 + .byte 197,156,194,208,1 // vcmpltps %ymm0,%ymm12,%ymm2 + .byte 196,193,124,88,253 // vaddps %ymm13,%ymm0,%ymm7 + .byte 196,227,125,74,215,32 // vblendvps %ymm2,%ymm7,%ymm0,%ymm2 + .byte 196,193,124,194,255,1 // vcmpltps %ymm15,%ymm0,%ymm7 + .byte 196,65,124,88,210 // vaddps %ymm10,%ymm0,%ymm10 + .byte 196,195,109,74,210,112 // vblendvps %ymm7,%ymm10,%ymm2,%ymm2 + .byte 196,194,13,168,195 // vfmadd213ps %ymm11,%ymm14,%ymm0 + .byte 197,228,92,218 // vsubps %ymm2,%ymm3,%ymm3 + .byte 196,194,13,168,219 // vfmadd213ps %ymm11,%ymm14,%ymm3 + .byte 197,236,194,228,1 // vcmpltps %ymm4,%ymm2,%ymm4 + .byte 196,227,37,74,219,64 // vblendvps %ymm4,%ymm3,%ymm11,%ymm3 + .byte 196,193,108,194,224,1 // vcmpltps %ymm8,%ymm2,%ymm4 + .byte 196,195,101,74,217,64 // vblendvps %ymm4,%ymm9,%ymm3,%ymm3 + .byte 197,236,194,201,1 // vcmpltps %ymm1,%ymm2,%ymm1 + .byte 196,227,101,74,208,16 // vblendvps %ymm1,%ymm0,%ymm3,%ymm2 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 - .byte 197,212,194,232,0 // vcmpeqps %ymm0,%ymm5,%ymm5 - .byte 196,227,93,74,194,80 // vblendvps %ymm5,%ymm2,%ymm4,%ymm0 - .byte 196,227,117,74,202,80 // vblendvps %ymm5,%ymm2,%ymm1,%ymm1 - .byte 196,227,101,74,210,80 // vblendvps %ymm5,%ymm2,%ymm3,%ymm2 + .byte 197,252,194,92,36,160,0 // vcmpeqps -0x60(%rsp),%ymm0,%ymm3 + .byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0 + .byte 196,227,125,74,197,48 // vblendvps %ymm3,%ymm5,%ymm0,%ymm0 + .byte 196,227,77,74,205,48 // vblendvps %ymm3,%ymm5,%ymm6,%ymm1 + .byte 196,227,109,74,213,48 // vblendvps %ymm3,%ymm5,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 197,252,16,92,36,128 // vmovups -0x80(%rsp),%ymm3 - .byte 197,252,16,100,36,160 // vmovups -0x60(%rsp),%ymm4 - .byte 197,252,16,108,36,192 // vmovups -0x40(%rsp),%ymm5 - .byte 197,252,16,116,36,224 // vmovups -0x20(%rsp),%ymm6 - .byte 197,252,16,60,36 // vmovups (%rsp),%ymm7 - .byte 72,131,196,56 // add $0x38,%rsp + .byte 197,252,16,92,36,192 // vmovups -0x40(%rsp),%ymm3 + .byte 197,252,16,100,36,224 // vmovups -0x20(%rsp),%ymm4 + .byte 197,252,16,44,36 // vmovups (%rsp),%ymm5 + .byte 197,252,16,116,36,32 // vmovups 0x20(%rsp),%ymm6 + .byte 197,252,16,124,36,64 // vmovups 0x40(%rsp),%ymm7 + .byte 72,131,196,120 // add $0x78,%rsp .byte 255,224 // jmpq *%rax HIDDEN _sk_scale_1_float_hsw @@ -8353,11 +8356,11 @@ _sk_scale_u8_hsw: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,51 // jne f70 <_sk_scale_u8_hsw+0x43> + .byte 117,51 // jne fa1 <_sk_scale_u8_hsw+0x43> .byte 197,122,126,0 // vmovq (%rax),%xmm8 .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,144,46,0,0 // vbroadcastss 0x2e90(%rip),%ymm9 # 3de4 <_sk_callback_hsw+0x1c9> + .byte 196,98,125,24,13,171,46,0,0 // vbroadcastss 0x2eab(%rip),%ymm9 # 3e30 <_sk_callback_hsw+0x1e5> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -8375,9 +8378,9 @@ _sk_scale_u8_hsw: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne f78 <_sk_scale_u8_hsw+0x4b> + .byte 117,234 // jne fa9 <_sk_scale_u8_hsw+0x4b> .byte 196,65,249,110,193 // vmovq %r9,%xmm8 - .byte 235,172 // jmp f41 <_sk_scale_u8_hsw+0x14> + .byte 235,172 // jmp f72 <_sk_scale_u8_hsw+0x14> HIDDEN _sk_lerp_1_float_hsw .globl _sk_lerp_1_float_hsw @@ -8405,11 +8408,11 @@ _sk_lerp_u8_hsw: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,71 // jne 101b <_sk_lerp_u8_hsw+0x57> + .byte 117,71 // jne 104c <_sk_lerp_u8_hsw+0x57> .byte 197,122,126,0 // vmovq (%rax),%xmm8 .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,253,45,0,0 // vbroadcastss 0x2dfd(%rip),%ymm9 # 3de8 <_sk_callback_hsw+0x1cd> + .byte 196,98,125,24,13,24,46,0,0 // vbroadcastss 0x2e18(%rip),%ymm9 # 3e34 <_sk_callback_hsw+0x1e9> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -8431,9 +8434,9 @@ _sk_lerp_u8_hsw: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 1023 <_sk_lerp_u8_hsw+0x5f> + .byte 117,234 // jne 1054 <_sk_lerp_u8_hsw+0x5f> .byte 196,65,249,110,193 // vmovq %r9,%xmm8 - .byte 235,152 // jmp fd8 <_sk_lerp_u8_hsw+0x14> + .byte 235,152 // jmp 1009 <_sk_lerp_u8_hsw+0x14> HIDDEN _sk_lerp_565_hsw .globl _sk_lerp_565_hsw @@ -8442,23 +8445,23 @@ _sk_lerp_565_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,149,0,0,0 // jne 10e3 <_sk_lerp_565_hsw+0xa3> + .byte 15,133,149,0,0,0 // jne 1114 <_sk_lerp_565_hsw+0xa3> .byte 196,193,122,111,28,122 // vmovdqu (%r10,%rdi,2),%xmm3 .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3 - .byte 196,98,125,88,5,138,45,0,0 // vpbroadcastd 0x2d8a(%rip),%ymm8 # 3dec <_sk_callback_hsw+0x1d1> + .byte 196,98,125,88,5,165,45,0,0 // vpbroadcastd 0x2da5(%rip),%ymm8 # 3e38 <_sk_callback_hsw+0x1ed> .byte 196,65,101,219,192 // vpand %ymm8,%ymm3,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,123,45,0,0 // vbroadcastss 0x2d7b(%rip),%ymm9 # 3df0 <_sk_callback_hsw+0x1d5> + .byte 196,98,125,24,13,150,45,0,0 // vbroadcastss 0x2d96(%rip),%ymm9 # 3e3c <_sk_callback_hsw+0x1f1> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,88,13,113,45,0,0 // vpbroadcastd 0x2d71(%rip),%ymm9 # 3df4 <_sk_callback_hsw+0x1d9> + .byte 196,98,125,88,13,140,45,0,0 // vpbroadcastd 0x2d8c(%rip),%ymm9 # 3e40 <_sk_callback_hsw+0x1f5> .byte 196,65,101,219,201 // vpand %ymm9,%ymm3,%ymm9 .byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9 - .byte 196,98,125,24,21,98,45,0,0 // vbroadcastss 0x2d62(%rip),%ymm10 # 3df8 <_sk_callback_hsw+0x1dd> + .byte 196,98,125,24,21,125,45,0,0 // vbroadcastss 0x2d7d(%rip),%ymm10 # 3e44 <_sk_callback_hsw+0x1f9> .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,88,21,88,45,0,0 // vpbroadcastd 0x2d58(%rip),%ymm10 # 3dfc <_sk_callback_hsw+0x1e1> + .byte 196,98,125,88,21,115,45,0,0 // vpbroadcastd 0x2d73(%rip),%ymm10 # 3e48 <_sk_callback_hsw+0x1fd> .byte 196,193,101,219,218 // vpand %ymm10,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,21,74,45,0,0 // vbroadcastss 0x2d4a(%rip),%ymm10 # 3e00 <_sk_callback_hsw+0x1e5> + .byte 196,98,125,24,21,101,45,0,0 // vbroadcastss 0x2d65(%rip),%ymm10 # 3e4c <_sk_callback_hsw+0x201> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -8467,16 +8470,16 @@ _sk_lerp_565_hsw: .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 .byte 196,226,101,168,214 // vfmadd213ps %ymm6,%ymm3,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,35,45,0,0 // vbroadcastss 0x2d23(%rip),%ymm3 # 3e04 <_sk_callback_hsw+0x1e9> + .byte 196,226,125,24,29,62,45,0,0 // vbroadcastss 0x2d3e(%rip),%ymm3 # 3e50 <_sk_callback_hsw+0x205> .byte 255,224 // jmpq *%rax .byte 65,137,200 // mov %ecx,%r8d .byte 65,128,224,7 // and $0x7,%r8b .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,89,255,255,255 // ja 1054 <_sk_lerp_565_hsw+0x14> + .byte 15,135,89,255,255,255 // ja 1085 <_sk_lerp_565_hsw+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 1150 <_sk_lerp_565_hsw+0x110> + .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 1180 <_sk_lerp_565_hsw+0x10f> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -8488,27 +8491,27 @@ _sk_lerp_565_hsw: .byte 196,193,97,196,92,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3 .byte 196,193,97,196,92,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3 .byte 196,193,97,196,28,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3 - .byte 233,5,255,255,255 // jmpq 1054 <_sk_lerp_565_hsw+0x14> - .byte 144 // nop - .byte 243,255 // repz (bad) + .byte 233,5,255,255,255 // jmpq 1085 <_sk_lerp_565_hsw+0x14> + .byte 244 // hlt .byte 255 // (bad) .byte 255 // (bad) - .byte 235,255 // jmp 1155 <_sk_lerp_565_hsw+0x115> .byte 255 // (bad) - .byte 255,227 // jmpq *%rbx + .byte 236 // in (%dx),%al .byte 255 // (bad) .byte 255 // (bad) + .byte 255,228 // jmpq *%rsp .byte 255 // (bad) - .byte 219,255 // (bad) .byte 255 // (bad) - .byte 255,211 // callq *%rbx .byte 255 // (bad) + .byte 220,255 // fdivr %st,%st(7) .byte 255 // (bad) - .byte 255,203 // dec %ebx + .byte 255,212 // callq *%rsp .byte 255 // (bad) .byte 255 // (bad) + .byte 255,204 // dec %esp .byte 255 // (bad) - .byte 191 // .byte 0xbf + .byte 255 // (bad) + .byte 255,192 // inc %eax .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -8522,23 +8525,23 @@ _sk_load_tables_hsw: .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 76,3,8 // add (%rax),%r9 .byte 77,133,192 // test %r8,%r8 - .byte 117,105 // jne 11ea <_sk_load_tables_hsw+0x7e> + .byte 117,105 // jne 121a <_sk_load_tables_hsw+0x7e> .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 - .byte 197,229,219,13,18,47,0,0 // vpand 0x2f12(%rip),%ymm3,%ymm1 # 40a0 <_sk_callback_hsw+0x485> + .byte 197,229,219,13,34,47,0,0 // vpand 0x2f22(%rip),%ymm3,%ymm1 # 40e0 <_sk_callback_hsw+0x495> .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 .byte 72,139,72,8 // mov 0x8(%rax),%rcx .byte 76,139,72,16 // mov 0x10(%rax),%r9 .byte 197,237,118,210 // vpcmpeqd %ymm2,%ymm2,%ymm2 .byte 196,226,109,146,4,137 // vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0 - .byte 196,226,101,0,21,18,47,0,0 // vpshufb 0x2f12(%rip),%ymm3,%ymm2 # 40c0 <_sk_callback_hsw+0x4a5> + .byte 196,226,101,0,21,34,47,0,0 // vpshufb 0x2f22(%rip),%ymm3,%ymm2 # 4100 <_sk_callback_hsw+0x4b5> .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 .byte 196,194,53,146,12,145 // vgatherdps %ymm9,(%r9,%ymm2,4),%ymm1 .byte 72,139,64,24 // mov 0x18(%rax),%rax - .byte 196,98,101,0,13,26,47,0,0 // vpshufb 0x2f1a(%rip),%ymm3,%ymm9 # 40e0 <_sk_callback_hsw+0x4c5> + .byte 196,98,101,0,13,42,47,0,0 // vpshufb 0x2f2a(%rip),%ymm3,%ymm9 # 4120 <_sk_callback_hsw+0x4d5> .byte 196,162,61,146,20,136 // vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,42,44,0,0 // vbroadcastss 0x2c2a(%rip),%ymm8 # 3e08 <_sk_callback_hsw+0x1ed> + .byte 196,98,125,24,5,70,44,0,0 // vbroadcastss 0x2c46(%rip),%ymm8 # 3e54 <_sk_callback_hsw+0x209> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,193 // mov %r8,%rcx @@ -8551,7 +8554,7 @@ _sk_load_tables_hsw: .byte 196,193,249,110,194 // vmovq %r10,%xmm0 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 - .byte 233,115,255,255,255 // jmpq 1186 <_sk_load_tables_hsw+0x1a> + .byte 233,115,255,255,255 // jmpq 11b6 <_sk_load_tables_hsw+0x1a> HIDDEN _sk_load_tables_u16_be_hsw .globl _sk_load_tables_u16_be_hsw @@ -8561,7 +8564,7 @@ _sk_load_tables_u16_be_hsw: .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,201,0,0,0 // jne 12f2 <_sk_load_tables_u16_be_hsw+0xdf> + .byte 15,133,201,0,0,0 // jne 1322 <_sk_load_tables_u16_be_hsw+0xdf> .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8 .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2 .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3 @@ -8577,7 +8580,7 @@ _sk_load_tables_u16_be_hsw: .byte 197,185,108,200 // vpunpcklqdq %xmm0,%xmm8,%xmm1 .byte 197,185,109,208 // vpunpckhqdq %xmm0,%xmm8,%xmm2 .byte 197,49,108,195 // vpunpcklqdq %xmm3,%xmm9,%xmm8 - .byte 197,121,111,21,166,47,0,0 // vmovdqa 0x2fa6(%rip),%xmm10 # 4220 <_sk_callback_hsw+0x605> + .byte 197,121,111,21,182,47,0,0 // vmovdqa 0x2fb6(%rip),%xmm10 # 4260 <_sk_callback_hsw+0x615> .byte 196,193,113,219,194 // vpand %xmm10,%xmm1,%xmm0 .byte 196,226,125,51,200 // vpmovzxwd %xmm0,%ymm1 .byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11 @@ -8599,36 +8602,36 @@ _sk_load_tables_u16_be_hsw: .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3 .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,35,43,0,0 // vbroadcastss 0x2b23(%rip),%ymm8 # 3e0c <_sk_callback_hsw+0x1f1> + .byte 196,98,125,24,5,63,43,0,0 // vbroadcastss 0x2b3f(%rip),%ymm8 # 3e58 <_sk_callback_hsw+0x20d> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 1358 <_sk_load_tables_u16_be_hsw+0x145> + .byte 116,85 // je 1388 <_sk_load_tables_u16_be_hsw+0x145> .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 1358 <_sk_load_tables_u16_be_hsw+0x145> + .byte 114,72 // jb 1388 <_sk_load_tables_u16_be_hsw+0x145> .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 1365 <_sk_load_tables_u16_be_hsw+0x152> + .byte 116,72 // je 1395 <_sk_load_tables_u16_be_hsw+0x152> .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 1365 <_sk_load_tables_u16_be_hsw+0x152> + .byte 114,59 // jb 1395 <_sk_load_tables_u16_be_hsw+0x152> .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,9,255,255,255 // je 1244 <_sk_load_tables_u16_be_hsw+0x31> + .byte 15,132,9,255,255,255 // je 1274 <_sk_load_tables_u16_be_hsw+0x31> .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,248,254,255,255 // jb 1244 <_sk_load_tables_u16_be_hsw+0x31> + .byte 15,130,248,254,255,255 // jb 1274 <_sk_load_tables_u16_be_hsw+0x31> .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9 - .byte 233,236,254,255,255 // jmpq 1244 <_sk_load_tables_u16_be_hsw+0x31> + .byte 233,236,254,255,255 // jmpq 1274 <_sk_load_tables_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,223,254,255,255 // jmpq 1244 <_sk_load_tables_u16_be_hsw+0x31> + .byte 233,223,254,255,255 // jmpq 1274 <_sk_load_tables_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,214,254,255,255 // jmpq 1244 <_sk_load_tables_u16_be_hsw+0x31> + .byte 233,214,254,255,255 // jmpq 1274 <_sk_load_tables_u16_be_hsw+0x31> HIDDEN _sk_load_tables_rgb_u16_be_hsw .globl _sk_load_tables_rgb_u16_be_hsw @@ -8638,7 +8641,7 @@ _sk_load_tables_rgb_u16_be_hsw: .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,193,0,0,0 // jne 1441 <_sk_load_tables_rgb_u16_be_hsw+0xd3> + .byte 15,133,193,0,0,0 // jne 1471 <_sk_load_tables_rgb_u16_be_hsw+0xd3> .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0 .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2 .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1 @@ -8659,7 +8662,7 @@ _sk_load_tables_rgb_u16_be_hsw: .byte 197,185,108,218 // vpunpcklqdq %xmm2,%xmm8,%xmm3 .byte 197,185,109,210 // vpunpckhqdq %xmm2,%xmm8,%xmm2 .byte 197,121,108,193 // vpunpcklqdq %xmm1,%xmm0,%xmm8 - .byte 197,121,111,13,70,46,0,0 // vmovdqa 0x2e46(%rip),%xmm9 # 4230 <_sk_callback_hsw+0x615> + .byte 197,121,111,13,86,46,0,0 // vmovdqa 0x2e56(%rip),%xmm9 # 4270 <_sk_callback_hsw+0x625> .byte 196,193,97,219,193 // vpand %xmm9,%xmm3,%xmm0 .byte 196,226,125,51,200 // vpmovzxwd %xmm0,%ymm1 .byte 197,229,118,219 // vpcmpeqd %ymm3,%ymm3,%ymm3 @@ -8676,41 +8679,41 @@ _sk_load_tables_rgb_u16_be_hsw: .byte 196,98,125,51,194 // vpmovzxwd %xmm2,%ymm8 .byte 196,162,101,146,20,128 // vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,209,41,0,0 // vbroadcastss 0x29d1(%rip),%ymm3 # 3e10 <_sk_callback_hsw+0x1f5> + .byte 196,226,125,24,29,237,41,0,0 // vbroadcastss 0x29ed(%rip),%ymm3 # 3e5c <_sk_callback_hsw+0x211> .byte 255,224 // jmpq *%rax .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0 .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 145a <_sk_load_tables_rgb_u16_be_hsw+0xec> - .byte 233,90,255,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 117,5 // jne 148a <_sk_load_tables_rgb_u16_be_hsw+0xec> + .byte 233,90,255,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1 .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 1489 <_sk_load_tables_rgb_u16_be_hsw+0x11b> + .byte 114,26 // jb 14b9 <_sk_load_tables_rgb_u16_be_hsw+0x11b> .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1 .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 148e <_sk_load_tables_rgb_u16_be_hsw+0x120> - .byte 233,43,255,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> - .byte 233,38,255,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 14be <_sk_load_tables_rgb_u16_be_hsw+0x120> + .byte 233,43,255,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 233,38,255,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1 .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 14bd <_sk_load_tables_rgb_u16_be_hsw+0x14f> + .byte 114,26 // jb 14ed <_sk_load_tables_rgb_u16_be_hsw+0x14f> .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1 .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 14c2 <_sk_load_tables_rgb_u16_be_hsw+0x154> - .byte 233,247,254,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> - .byte 233,242,254,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 14f2 <_sk_load_tables_rgb_u16_be_hsw+0x154> + .byte 233,247,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 233,242,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3 .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 14eb <_sk_load_tables_rgb_u16_be_hsw+0x17d> + .byte 114,20 // jb 151b <_sk_load_tables_rgb_u16_be_hsw+0x17d> .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3 .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 - .byte 233,201,254,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> - .byte 233,196,254,255,255 // jmpq 13b4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 233,201,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> + .byte 233,196,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46> HIDDEN _sk_byte_tables_hsw .globl _sk_byte_tables_hsw @@ -8723,7 +8726,7 @@ _sk_byte_tables_hsw: .byte 65,84 // push %r12 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,15,41,0,0 // vbroadcastss 0x290f(%rip),%ymm8 # 3e14 <_sk_callback_hsw+0x1f9> + .byte 196,98,125,24,5,43,41,0,0 // vbroadcastss 0x292b(%rip),%ymm8 # 3e60 <_sk_callback_hsw+0x215> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 @@ -8760,7 +8763,7 @@ _sk_byte_tables_hsw: .byte 196,227,121,32,197,7 // vpinsrb $0x7,%ebp,%xmm0,%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,96,40,0,0 // vbroadcastss 0x2860(%rip),%ymm9 # 3e18 <_sk_callback_hsw+0x1fd> + .byte 196,98,125,24,13,124,40,0,0 // vbroadcastss 0x287c(%rip),%ymm9 # 3e64 <_sk_callback_hsw+0x219> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 @@ -8921,7 +8924,7 @@ _sk_byte_tables_rgb_hsw: .byte 196,227,121,32,197,7 // vpinsrb $0x7,%ebp,%xmm0,%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,153,37,0,0 // vbroadcastss 0x2599(%rip),%ymm9 # 3e1c <_sk_callback_hsw+0x201> + .byte 196,98,125,24,13,181,37,0,0 // vbroadcastss 0x25b5(%rip),%ymm9 # 3e68 <_sk_callback_hsw+0x21d> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 @@ -9084,33 +9087,33 @@ _sk_parametric_r_hsw: .byte 196,66,125,168,211 // vfmadd213ps %ymm11,%ymm0,%ymm10 .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 196,98,125,24,37,76,35,0,0 // vbroadcastss 0x234c(%rip),%ymm12 # 3e20 <_sk_callback_hsw+0x205> - .byte 196,98,125,24,45,71,35,0,0 // vbroadcastss 0x2347(%rip),%ymm13 # 3e24 <_sk_callback_hsw+0x209> + .byte 196,98,125,24,37,104,35,0,0 // vbroadcastss 0x2368(%rip),%ymm12 # 3e6c <_sk_callback_hsw+0x221> + .byte 196,98,125,24,45,99,35,0,0 // vbroadcastss 0x2363(%rip),%ymm13 # 3e70 <_sk_callback_hsw+0x225> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,61,35,0,0 // vbroadcastss 0x233d(%rip),%ymm13 # 3e28 <_sk_callback_hsw+0x20d> + .byte 196,98,125,24,45,89,35,0,0 // vbroadcastss 0x2359(%rip),%ymm13 # 3e74 <_sk_callback_hsw+0x229> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,51,35,0,0 // vbroadcastss 0x2333(%rip),%ymm13 # 3e2c <_sk_callback_hsw+0x211> + .byte 196,98,125,24,45,79,35,0,0 // vbroadcastss 0x234f(%rip),%ymm13 # 3e78 <_sk_callback_hsw+0x22d> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,41,35,0,0 // vbroadcastss 0x2329(%rip),%ymm11 # 3e30 <_sk_callback_hsw+0x215> + .byte 196,98,125,24,29,69,35,0,0 // vbroadcastss 0x2345(%rip),%ymm11 # 3e7c <_sk_callback_hsw+0x231> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,31,35,0,0 // vbroadcastss 0x231f(%rip),%ymm12 # 3e34 <_sk_callback_hsw+0x219> + .byte 196,98,125,24,37,59,35,0,0 // vbroadcastss 0x233b(%rip),%ymm12 # 3e80 <_sk_callback_hsw+0x235> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,21,35,0,0 // vbroadcastss 0x2315(%rip),%ymm12 # 3e38 <_sk_callback_hsw+0x21d> + .byte 196,98,125,24,37,49,35,0,0 // vbroadcastss 0x2331(%rip),%ymm12 # 3e84 <_sk_callback_hsw+0x239> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10 - .byte 196,98,125,24,29,246,34,0,0 // vbroadcastss 0x22f6(%rip),%ymm11 # 3e3c <_sk_callback_hsw+0x221> + .byte 196,98,125,24,29,18,35,0,0 // vbroadcastss 0x2312(%rip),%ymm11 # 3e88 <_sk_callback_hsw+0x23d> .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,236,34,0,0 // vbroadcastss 0x22ec(%rip),%ymm11 # 3e40 <_sk_callback_hsw+0x225> + .byte 196,98,125,24,29,8,35,0,0 // vbroadcastss 0x2308(%rip),%ymm11 # 3e8c <_sk_callback_hsw+0x241> .byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11 - .byte 196,226,125,24,5,226,34,0,0 // vbroadcastss 0x22e2(%rip),%ymm0 # 3e44 <_sk_callback_hsw+0x229> + .byte 196,226,125,24,5,254,34,0,0 // vbroadcastss 0x22fe(%rip),%ymm0 # 3e90 <_sk_callback_hsw+0x245> .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 - .byte 196,98,125,24,21,216,34,0,0 // vbroadcastss 0x22d8(%rip),%ymm10 # 3e48 <_sk_callback_hsw+0x22d> + .byte 196,98,125,24,21,244,34,0,0 // vbroadcastss 0x22f4(%rip),%ymm10 # 3e94 <_sk_callback_hsw+0x249> .byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0 .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 - .byte 196,98,125,24,21,203,34,0,0 // vbroadcastss 0x22cb(%rip),%ymm10 # 3e4c <_sk_callback_hsw+0x231> + .byte 196,98,125,24,21,231,34,0,0 // vbroadcastss 0x22e7(%rip),%ymm10 # 3e98 <_sk_callback_hsw+0x24d> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -9118,7 +9121,7 @@ _sk_parametric_r_hsw: .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,162,34,0,0 // vbroadcastss 0x22a2(%rip),%ymm8 # 3e50 <_sk_callback_hsw+0x235> + .byte 196,98,125,24,5,190,34,0,0 // vbroadcastss 0x22be(%rip),%ymm8 # 3e9c <_sk_callback_hsw+0x251> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9138,33 +9141,33 @@ _sk_parametric_g_hsw: .byte 196,66,117,168,211 // vfmadd213ps %ymm11,%ymm1,%ymm10 .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 196,98,125,24,37,90,34,0,0 // vbroadcastss 0x225a(%rip),%ymm12 # 3e54 <_sk_callback_hsw+0x239> - .byte 196,98,125,24,45,85,34,0,0 // vbroadcastss 0x2255(%rip),%ymm13 # 3e58 <_sk_callback_hsw+0x23d> + .byte 196,98,125,24,37,118,34,0,0 // vbroadcastss 0x2276(%rip),%ymm12 # 3ea0 <_sk_callback_hsw+0x255> + .byte 196,98,125,24,45,113,34,0,0 // vbroadcastss 0x2271(%rip),%ymm13 # 3ea4 <_sk_callback_hsw+0x259> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,75,34,0,0 // vbroadcastss 0x224b(%rip),%ymm13 # 3e5c <_sk_callback_hsw+0x241> + .byte 196,98,125,24,45,103,34,0,0 // vbroadcastss 0x2267(%rip),%ymm13 # 3ea8 <_sk_callback_hsw+0x25d> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,65,34,0,0 // vbroadcastss 0x2241(%rip),%ymm13 # 3e60 <_sk_callback_hsw+0x245> + .byte 196,98,125,24,45,93,34,0,0 // vbroadcastss 0x225d(%rip),%ymm13 # 3eac <_sk_callback_hsw+0x261> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,55,34,0,0 // vbroadcastss 0x2237(%rip),%ymm11 # 3e64 <_sk_callback_hsw+0x249> + .byte 196,98,125,24,29,83,34,0,0 // vbroadcastss 0x2253(%rip),%ymm11 # 3eb0 <_sk_callback_hsw+0x265> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,45,34,0,0 // vbroadcastss 0x222d(%rip),%ymm12 # 3e68 <_sk_callback_hsw+0x24d> + .byte 196,98,125,24,37,73,34,0,0 // vbroadcastss 0x2249(%rip),%ymm12 # 3eb4 <_sk_callback_hsw+0x269> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,35,34,0,0 // vbroadcastss 0x2223(%rip),%ymm12 # 3e6c <_sk_callback_hsw+0x251> + .byte 196,98,125,24,37,63,34,0,0 // vbroadcastss 0x223f(%rip),%ymm12 # 3eb8 <_sk_callback_hsw+0x26d> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10 - .byte 196,98,125,24,29,4,34,0,0 // vbroadcastss 0x2204(%rip),%ymm11 # 3e70 <_sk_callback_hsw+0x255> + .byte 196,98,125,24,29,32,34,0,0 // vbroadcastss 0x2220(%rip),%ymm11 # 3ebc <_sk_callback_hsw+0x271> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,250,33,0,0 // vbroadcastss 0x21fa(%rip),%ymm11 # 3e74 <_sk_callback_hsw+0x259> + .byte 196,98,125,24,29,22,34,0,0 // vbroadcastss 0x2216(%rip),%ymm11 # 3ec0 <_sk_callback_hsw+0x275> .byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11 - .byte 196,226,125,24,13,240,33,0,0 // vbroadcastss 0x21f0(%rip),%ymm1 # 3e78 <_sk_callback_hsw+0x25d> + .byte 196,226,125,24,13,12,34,0,0 // vbroadcastss 0x220c(%rip),%ymm1 # 3ec4 <_sk_callback_hsw+0x279> .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 - .byte 196,98,125,24,21,230,33,0,0 // vbroadcastss 0x21e6(%rip),%ymm10 # 3e7c <_sk_callback_hsw+0x261> + .byte 196,98,125,24,21,2,34,0,0 // vbroadcastss 0x2202(%rip),%ymm10 # 3ec8 <_sk_callback_hsw+0x27d> .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 - .byte 196,98,125,24,21,217,33,0,0 // vbroadcastss 0x21d9(%rip),%ymm10 # 3e80 <_sk_callback_hsw+0x265> + .byte 196,98,125,24,21,245,33,0,0 // vbroadcastss 0x21f5(%rip),%ymm10 # 3ecc <_sk_callback_hsw+0x281> .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -9172,7 +9175,7 @@ _sk_parametric_g_hsw: .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,5,176,33,0,0 // vbroadcastss 0x21b0(%rip),%ymm8 # 3e84 <_sk_callback_hsw+0x269> + .byte 196,98,125,24,5,204,33,0,0 // vbroadcastss 0x21cc(%rip),%ymm8 # 3ed0 <_sk_callback_hsw+0x285> .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9192,33 +9195,33 @@ _sk_parametric_b_hsw: .byte 196,66,109,168,211 // vfmadd213ps %ymm11,%ymm2,%ymm10 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 196,98,125,24,37,104,33,0,0 // vbroadcastss 0x2168(%rip),%ymm12 # 3e88 <_sk_callback_hsw+0x26d> - .byte 196,98,125,24,45,99,33,0,0 // vbroadcastss 0x2163(%rip),%ymm13 # 3e8c <_sk_callback_hsw+0x271> + .byte 196,98,125,24,37,132,33,0,0 // vbroadcastss 0x2184(%rip),%ymm12 # 3ed4 <_sk_callback_hsw+0x289> + .byte 196,98,125,24,45,127,33,0,0 // vbroadcastss 0x217f(%rip),%ymm13 # 3ed8 <_sk_callback_hsw+0x28d> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,89,33,0,0 // vbroadcastss 0x2159(%rip),%ymm13 # 3e90 <_sk_callback_hsw+0x275> + .byte 196,98,125,24,45,117,33,0,0 // vbroadcastss 0x2175(%rip),%ymm13 # 3edc <_sk_callback_hsw+0x291> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,79,33,0,0 // vbroadcastss 0x214f(%rip),%ymm13 # 3e94 <_sk_callback_hsw+0x279> + .byte 196,98,125,24,45,107,33,0,0 // vbroadcastss 0x216b(%rip),%ymm13 # 3ee0 <_sk_callback_hsw+0x295> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,69,33,0,0 // vbroadcastss 0x2145(%rip),%ymm11 # 3e98 <_sk_callback_hsw+0x27d> + .byte 196,98,125,24,29,97,33,0,0 // vbroadcastss 0x2161(%rip),%ymm11 # 3ee4 <_sk_callback_hsw+0x299> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,59,33,0,0 // vbroadcastss 0x213b(%rip),%ymm12 # 3e9c <_sk_callback_hsw+0x281> + .byte 196,98,125,24,37,87,33,0,0 // vbroadcastss 0x2157(%rip),%ymm12 # 3ee8 <_sk_callback_hsw+0x29d> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,49,33,0,0 // vbroadcastss 0x2131(%rip),%ymm12 # 3ea0 <_sk_callback_hsw+0x285> + .byte 196,98,125,24,37,77,33,0,0 // vbroadcastss 0x214d(%rip),%ymm12 # 3eec <_sk_callback_hsw+0x2a1> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10 - .byte 196,98,125,24,29,18,33,0,0 // vbroadcastss 0x2112(%rip),%ymm11 # 3ea4 <_sk_callback_hsw+0x289> + .byte 196,98,125,24,29,46,33,0,0 // vbroadcastss 0x212e(%rip),%ymm11 # 3ef0 <_sk_callback_hsw+0x2a5> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,8,33,0,0 // vbroadcastss 0x2108(%rip),%ymm11 # 3ea8 <_sk_callback_hsw+0x28d> + .byte 196,98,125,24,29,36,33,0,0 // vbroadcastss 0x2124(%rip),%ymm11 # 3ef4 <_sk_callback_hsw+0x2a9> .byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11 - .byte 196,226,125,24,21,254,32,0,0 // vbroadcastss 0x20fe(%rip),%ymm2 # 3eac <_sk_callback_hsw+0x291> + .byte 196,226,125,24,21,26,33,0,0 // vbroadcastss 0x211a(%rip),%ymm2 # 3ef8 <_sk_callback_hsw+0x2ad> .byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2 - .byte 196,98,125,24,21,244,32,0,0 // vbroadcastss 0x20f4(%rip),%ymm10 # 3eb0 <_sk_callback_hsw+0x295> + .byte 196,98,125,24,21,16,33,0,0 // vbroadcastss 0x2110(%rip),%ymm10 # 3efc <_sk_callback_hsw+0x2b1> .byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2 .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 - .byte 196,98,125,24,21,231,32,0,0 // vbroadcastss 0x20e7(%rip),%ymm10 # 3eb4 <_sk_callback_hsw+0x299> + .byte 196,98,125,24,21,3,33,0,0 // vbroadcastss 0x2103(%rip),%ymm10 # 3f00 <_sk_callback_hsw+0x2b5> .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -9226,7 +9229,7 @@ _sk_parametric_b_hsw: .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,190,32,0,0 // vbroadcastss 0x20be(%rip),%ymm8 # 3eb8 <_sk_callback_hsw+0x29d> + .byte 196,98,125,24,5,218,32,0,0 // vbroadcastss 0x20da(%rip),%ymm8 # 3f04 <_sk_callback_hsw+0x2b9> .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9246,33 +9249,33 @@ _sk_parametric_a_hsw: .byte 196,66,101,168,211 // vfmadd213ps %ymm11,%ymm3,%ymm10 .byte 196,226,125,24,24 // vbroadcastss (%rax),%ymm3 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 196,98,125,24,37,118,32,0,0 // vbroadcastss 0x2076(%rip),%ymm12 # 3ebc <_sk_callback_hsw+0x2a1> - .byte 196,98,125,24,45,113,32,0,0 // vbroadcastss 0x2071(%rip),%ymm13 # 3ec0 <_sk_callback_hsw+0x2a5> + .byte 196,98,125,24,37,146,32,0,0 // vbroadcastss 0x2092(%rip),%ymm12 # 3f08 <_sk_callback_hsw+0x2bd> + .byte 196,98,125,24,45,141,32,0,0 // vbroadcastss 0x208d(%rip),%ymm13 # 3f0c <_sk_callback_hsw+0x2c1> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,103,32,0,0 // vbroadcastss 0x2067(%rip),%ymm13 # 3ec4 <_sk_callback_hsw+0x2a9> + .byte 196,98,125,24,45,131,32,0,0 // vbroadcastss 0x2083(%rip),%ymm13 # 3f10 <_sk_callback_hsw+0x2c5> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,93,32,0,0 // vbroadcastss 0x205d(%rip),%ymm13 # 3ec8 <_sk_callback_hsw+0x2ad> + .byte 196,98,125,24,45,121,32,0,0 // vbroadcastss 0x2079(%rip),%ymm13 # 3f14 <_sk_callback_hsw+0x2c9> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,83,32,0,0 // vbroadcastss 0x2053(%rip),%ymm11 # 3ecc <_sk_callback_hsw+0x2b1> + .byte 196,98,125,24,29,111,32,0,0 // vbroadcastss 0x206f(%rip),%ymm11 # 3f18 <_sk_callback_hsw+0x2cd> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,73,32,0,0 // vbroadcastss 0x2049(%rip),%ymm12 # 3ed0 <_sk_callback_hsw+0x2b5> + .byte 196,98,125,24,37,101,32,0,0 // vbroadcastss 0x2065(%rip),%ymm12 # 3f1c <_sk_callback_hsw+0x2d1> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,63,32,0,0 // vbroadcastss 0x203f(%rip),%ymm12 # 3ed4 <_sk_callback_hsw+0x2b9> + .byte 196,98,125,24,37,91,32,0,0 // vbroadcastss 0x205b(%rip),%ymm12 # 3f20 <_sk_callback_hsw+0x2d5> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10 - .byte 196,98,125,24,29,32,32,0,0 // vbroadcastss 0x2020(%rip),%ymm11 # 3ed8 <_sk_callback_hsw+0x2bd> + .byte 196,98,125,24,29,60,32,0,0 // vbroadcastss 0x203c(%rip),%ymm11 # 3f24 <_sk_callback_hsw+0x2d9> .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,22,32,0,0 // vbroadcastss 0x2016(%rip),%ymm11 # 3edc <_sk_callback_hsw+0x2c1> + .byte 196,98,125,24,29,50,32,0,0 // vbroadcastss 0x2032(%rip),%ymm11 # 3f28 <_sk_callback_hsw+0x2dd> .byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11 - .byte 196,226,125,24,29,12,32,0,0 // vbroadcastss 0x200c(%rip),%ymm3 # 3ee0 <_sk_callback_hsw+0x2c5> + .byte 196,226,125,24,29,40,32,0,0 // vbroadcastss 0x2028(%rip),%ymm3 # 3f2c <_sk_callback_hsw+0x2e1> .byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3 - .byte 196,98,125,24,21,2,32,0,0 // vbroadcastss 0x2002(%rip),%ymm10 # 3ee4 <_sk_callback_hsw+0x2c9> + .byte 196,98,125,24,21,30,32,0,0 // vbroadcastss 0x201e(%rip),%ymm10 # 3f30 <_sk_callback_hsw+0x2e5> .byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3 .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 - .byte 196,98,125,24,21,245,31,0,0 // vbroadcastss 0x1ff5(%rip),%ymm10 # 3ee8 <_sk_callback_hsw+0x2cd> + .byte 196,98,125,24,21,17,32,0,0 // vbroadcastss 0x2011(%rip),%ymm10 # 3f34 <_sk_callback_hsw+0x2e9> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -9280,7 +9283,7 @@ _sk_parametric_a_hsw: .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 - .byte 196,98,125,24,5,204,31,0,0 // vbroadcastss 0x1fcc(%rip),%ymm8 # 3eec <_sk_callback_hsw+0x2d1> + .byte 196,98,125,24,5,232,31,0,0 // vbroadcastss 0x1fe8(%rip),%ymm8 # 3f38 <_sk_callback_hsw+0x2ed> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9289,26 +9292,26 @@ HIDDEN _sk_lab_to_xyz_hsw .globl _sk_lab_to_xyz_hsw FUNCTION(_sk_lab_to_xyz_hsw) _sk_lab_to_xyz_hsw: - .byte 196,98,125,24,5,190,31,0,0 // vbroadcastss 0x1fbe(%rip),%ymm8 # 3ef0 <_sk_callback_hsw+0x2d5> - .byte 196,98,125,24,13,185,31,0,0 // vbroadcastss 0x1fb9(%rip),%ymm9 # 3ef4 <_sk_callback_hsw+0x2d9> - .byte 196,98,125,24,21,180,31,0,0 // vbroadcastss 0x1fb4(%rip),%ymm10 # 3ef8 <_sk_callback_hsw+0x2dd> + .byte 196,98,125,24,5,218,31,0,0 // vbroadcastss 0x1fda(%rip),%ymm8 # 3f3c <_sk_callback_hsw+0x2f1> + .byte 196,98,125,24,13,213,31,0,0 // vbroadcastss 0x1fd5(%rip),%ymm9 # 3f40 <_sk_callback_hsw+0x2f5> + .byte 196,98,125,24,21,208,31,0,0 // vbroadcastss 0x1fd0(%rip),%ymm10 # 3f44 <_sk_callback_hsw+0x2f9> .byte 196,194,53,168,202 // vfmadd213ps %ymm10,%ymm9,%ymm1 .byte 196,194,53,168,210 // vfmadd213ps %ymm10,%ymm9,%ymm2 - .byte 196,98,125,24,13,165,31,0,0 // vbroadcastss 0x1fa5(%rip),%ymm9 # 3efc <_sk_callback_hsw+0x2e1> + .byte 196,98,125,24,13,193,31,0,0 // vbroadcastss 0x1fc1(%rip),%ymm9 # 3f48 <_sk_callback_hsw+0x2fd> .byte 196,66,125,184,200 // vfmadd231ps %ymm8,%ymm0,%ymm9 - .byte 196,226,125,24,5,155,31,0,0 // vbroadcastss 0x1f9b(%rip),%ymm0 # 3f00 <_sk_callback_hsw+0x2e5> + .byte 196,226,125,24,5,183,31,0,0 // vbroadcastss 0x1fb7(%rip),%ymm0 # 3f4c <_sk_callback_hsw+0x301> .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 - .byte 196,98,125,24,5,146,31,0,0 // vbroadcastss 0x1f92(%rip),%ymm8 # 3f04 <_sk_callback_hsw+0x2e9> + .byte 196,98,125,24,5,174,31,0,0 // vbroadcastss 0x1fae(%rip),%ymm8 # 3f50 <_sk_callback_hsw+0x305> .byte 196,98,117,168,192 // vfmadd213ps %ymm0,%ymm1,%ymm8 - .byte 196,98,125,24,13,136,31,0,0 // vbroadcastss 0x1f88(%rip),%ymm9 # 3f08 <_sk_callback_hsw+0x2ed> + .byte 196,98,125,24,13,164,31,0,0 // vbroadcastss 0x1fa4(%rip),%ymm9 # 3f54 <_sk_callback_hsw+0x309> .byte 196,98,109,172,200 // vfnmadd213ps %ymm0,%ymm2,%ymm9 .byte 196,193,60,89,200 // vmulps %ymm8,%ymm8,%ymm1 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 - .byte 196,226,125,24,21,117,31,0,0 // vbroadcastss 0x1f75(%rip),%ymm2 # 3f0c <_sk_callback_hsw+0x2f1> + .byte 196,226,125,24,21,145,31,0,0 // vbroadcastss 0x1f91(%rip),%ymm2 # 3f58 <_sk_callback_hsw+0x30d> .byte 197,108,194,209,1 // vcmpltps %ymm1,%ymm2,%ymm10 - .byte 196,98,125,24,29,107,31,0,0 // vbroadcastss 0x1f6b(%rip),%ymm11 # 3f10 <_sk_callback_hsw+0x2f5> + .byte 196,98,125,24,29,135,31,0,0 // vbroadcastss 0x1f87(%rip),%ymm11 # 3f5c <_sk_callback_hsw+0x311> .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 - .byte 196,98,125,24,37,97,31,0,0 // vbroadcastss 0x1f61(%rip),%ymm12 # 3f14 <_sk_callback_hsw+0x2f9> + .byte 196,98,125,24,37,125,31,0,0 // vbroadcastss 0x1f7d(%rip),%ymm12 # 3f60 <_sk_callback_hsw+0x315> .byte 196,65,60,89,196 // vmulps %ymm12,%ymm8,%ymm8 .byte 196,99,61,74,193,160 // vblendvps %ymm10,%ymm1,%ymm8,%ymm8 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1 @@ -9323,9 +9326,9 @@ _sk_lab_to_xyz_hsw: .byte 196,65,52,88,203 // vaddps %ymm11,%ymm9,%ymm9 .byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9 .byte 196,227,53,74,208,32 // vblendvps %ymm2,%ymm0,%ymm9,%ymm2 - .byte 196,226,125,24,5,22,31,0,0 // vbroadcastss 0x1f16(%rip),%ymm0 # 3f18 <_sk_callback_hsw+0x2fd> + .byte 196,226,125,24,5,50,31,0,0 // vbroadcastss 0x1f32(%rip),%ymm0 # 3f64 <_sk_callback_hsw+0x319> .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,13,31,0,0 // vbroadcastss 0x1f0d(%rip),%ymm8 # 3f1c <_sk_callback_hsw+0x301> + .byte 196,98,125,24,5,41,31,0,0 // vbroadcastss 0x1f29(%rip),%ymm8 # 3f68 <_sk_callback_hsw+0x31d> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9339,11 +9342,11 @@ _sk_load_a8_hsw: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,45 // jne 2055 <_sk_load_a8_hsw+0x3d> + .byte 117,45 // jne 2085 <_sk_load_a8_hsw+0x3d> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,226,30,0,0 // vbroadcastss 0x1ee2(%rip),%ymm1 # 3f20 <_sk_callback_hsw+0x305> + .byte 196,226,125,24,13,254,30,0,0 // vbroadcastss 0x1efe(%rip),%ymm1 # 3f6c <_sk_callback_hsw+0x321> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -9360,9 +9363,9 @@ _sk_load_a8_hsw: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 205d <_sk_load_a8_hsw+0x45> + .byte 117,234 // jne 208d <_sk_load_a8_hsw+0x45> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,178 // jmp 202c <_sk_load_a8_hsw+0x14> + .byte 235,178 // jmp 205c <_sk_load_a8_hsw+0x14> HIDDEN _sk_gather_a8_hsw .globl _sk_gather_a8_hsw @@ -9408,7 +9411,7 @@ _sk_gather_a8_hsw: .byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,237,29,0,0 // vbroadcastss 0x1ded(%rip),%ymm1 # 3f24 <_sk_callback_hsw+0x309> + .byte 196,226,125,24,13,9,30,0,0 // vbroadcastss 0x1e09(%rip),%ymm1 # 3f70 <_sk_callback_hsw+0x325> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -9426,14 +9429,14 @@ FUNCTION(_sk_store_a8_hsw) _sk_store_a8_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,200,29,0,0 // vbroadcastss 0x1dc8(%rip),%ymm8 # 3f28 <_sk_callback_hsw+0x30d> + .byte 196,98,125,24,5,228,29,0,0 // vbroadcastss 0x1de4(%rip),%ymm8 # 3f74 <_sk_callback_hsw+0x329> .byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 2189 <_sk_store_a8_hsw+0x37> + .byte 117,10 // jne 21b9 <_sk_store_a8_hsw+0x37> .byte 196,65,123,17,4,58 // vmovsd %xmm8,(%r10,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9441,10 +9444,10 @@ _sk_store_a8_hsw: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 2185 <_sk_store_a8_hsw+0x33> + .byte 119,236 // ja 21b5 <_sk_store_a8_hsw+0x33> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 21ec <_sk_store_a8_hsw+0x9a> + .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 221c <_sk_store_a8_hsw+0x9a> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -9455,7 +9458,7 @@ _sk_store_a8_hsw: .byte 196,67,121,20,68,58,2,4 // vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1) .byte 196,67,121,20,68,58,1,2 // vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1) .byte 196,67,121,20,4,58,0 // vpextrb $0x0,%xmm8,(%r10,%rdi,1) - .byte 235,154 // jmp 2185 <_sk_store_a8_hsw+0x33> + .byte 235,154 // jmp 21b5 <_sk_store_a8_hsw+0x33> .byte 144 // nop .byte 246,255 // idiv %bh .byte 255 // (bad) @@ -9489,14 +9492,14 @@ _sk_load_g8_hsw: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,50 // jne 224a <_sk_load_g8_hsw+0x42> + .byte 117,50 // jne 227a <_sk_load_g8_hsw+0x42> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,254,28,0,0 // vbroadcastss 0x1cfe(%rip),%ymm1 # 3f2c <_sk_callback_hsw+0x311> + .byte 196,226,125,24,13,26,29,0,0 // vbroadcastss 0x1d1a(%rip),%ymm1 # 3f78 <_sk_callback_hsw+0x32d> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,243,28,0,0 // vbroadcastss 0x1cf3(%rip),%ymm3 # 3f30 <_sk_callback_hsw+0x315> + .byte 196,226,125,24,29,15,29,0,0 // vbroadcastss 0x1d0f(%rip),%ymm3 # 3f7c <_sk_callback_hsw+0x331> .byte 76,137,193 // mov %r8,%rcx .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 @@ -9510,9 +9513,9 @@ _sk_load_g8_hsw: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 2252 <_sk_load_g8_hsw+0x4a> + .byte 117,234 // jne 2282 <_sk_load_g8_hsw+0x4a> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,173 // jmp 221c <_sk_load_g8_hsw+0x14> + .byte 235,173 // jmp 224c <_sk_load_g8_hsw+0x14> HIDDEN _sk_gather_g8_hsw .globl _sk_gather_g8_hsw @@ -9558,10 +9561,10 @@ _sk_gather_g8_hsw: .byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,8,28,0,0 // vbroadcastss 0x1c08(%rip),%ymm1 # 3f34 <_sk_callback_hsw+0x319> + .byte 196,226,125,24,13,36,28,0,0 // vbroadcastss 0x1c24(%rip),%ymm1 # 3f80 <_sk_callback_hsw+0x335> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,253,27,0,0 // vbroadcastss 0x1bfd(%rip),%ymm3 # 3f38 <_sk_callback_hsw+0x31d> + .byte 196,226,125,24,29,25,28,0,0 // vbroadcastss 0x1c19(%rip),%ymm3 # 3f84 <_sk_callback_hsw+0x339> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 91 // pop %rbx @@ -9577,9 +9580,9 @@ _sk_gather_i8_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 235b <_sk_gather_i8_hsw+0xf> + .byte 116,5 // je 238b <_sk_gather_i8_hsw+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 235d <_sk_gather_i8_hsw+0x11> + .byte 235,2 // jmp 238d <_sk_gather_i8_hsw+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -9617,14 +9620,14 @@ _sk_gather_i8_hsw: .byte 73,139,64,8 // mov 0x8(%r8),%rax .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 .byte 196,226,117,144,28,128 // vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 - .byte 197,229,219,5,245,28,0,0 // vpand 0x1cf5(%rip),%ymm3,%ymm0 # 4100 <_sk_callback_hsw+0x4e5> + .byte 197,229,219,5,5,29,0,0 // vpand 0x1d05(%rip),%ymm3,%ymm0 # 4140 <_sk_callback_hsw+0x4f5> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,36,27,0,0 // vbroadcastss 0x1b24(%rip),%ymm8 # 3f3c <_sk_callback_hsw+0x321> + .byte 196,98,125,24,5,64,27,0,0 // vbroadcastss 0x1b40(%rip),%ymm8 # 3f88 <_sk_callback_hsw+0x33d> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,226,101,0,13,250,28,0,0 // vpshufb 0x1cfa(%rip),%ymm3,%ymm1 # 4120 <_sk_callback_hsw+0x505> + .byte 196,226,101,0,13,10,29,0,0 // vpshufb 0x1d0a(%rip),%ymm3,%ymm1 # 4160 <_sk_callback_hsw+0x515> .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,226,101,0,21,8,29,0,0 // vpshufb 0x1d08(%rip),%ymm3,%ymm2 # 4140 <_sk_callback_hsw+0x525> + .byte 196,226,101,0,21,24,29,0,0 // vpshufb 0x1d18(%rip),%ymm3,%ymm2 # 4180 <_sk_callback_hsw+0x535> .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 @@ -9645,35 +9648,35 @@ _sk_load_565_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 117,114 // jne 24d8 <_sk_load_565_hsw+0x7c> + .byte 117,114 // jne 2508 <_sk_load_565_hsw+0x7c> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 - .byte 196,226,125,88,5,198,26,0,0 // vpbroadcastd 0x1ac6(%rip),%ymm0 # 3f40 <_sk_callback_hsw+0x325> + .byte 196,226,125,88,5,226,26,0,0 // vpbroadcastd 0x1ae2(%rip),%ymm0 # 3f8c <_sk_callback_hsw+0x341> .byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,185,26,0,0 // vbroadcastss 0x1ab9(%rip),%ymm1 # 3f44 <_sk_callback_hsw+0x329> + .byte 196,226,125,24,13,213,26,0,0 // vbroadcastss 0x1ad5(%rip),%ymm1 # 3f90 <_sk_callback_hsw+0x345> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,176,26,0,0 // vpbroadcastd 0x1ab0(%rip),%ymm1 # 3f48 <_sk_callback_hsw+0x32d> + .byte 196,226,125,88,13,204,26,0,0 // vpbroadcastd 0x1acc(%rip),%ymm1 # 3f94 <_sk_callback_hsw+0x349> .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,163,26,0,0 // vbroadcastss 0x1aa3(%rip),%ymm3 # 3f4c <_sk_callback_hsw+0x331> + .byte 196,226,125,24,29,191,26,0,0 // vbroadcastss 0x1abf(%rip),%ymm3 # 3f98 <_sk_callback_hsw+0x34d> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,88,29,154,26,0,0 // vpbroadcastd 0x1a9a(%rip),%ymm3 # 3f50 <_sk_callback_hsw+0x335> + .byte 196,226,125,88,29,182,26,0,0 // vpbroadcastd 0x1ab6(%rip),%ymm3 # 3f9c <_sk_callback_hsw+0x351> .byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,141,26,0,0 // vbroadcastss 0x1a8d(%rip),%ymm3 # 3f54 <_sk_callback_hsw+0x339> + .byte 196,226,125,24,29,169,26,0,0 // vbroadcastss 0x1aa9(%rip),%ymm3 # 3fa0 <_sk_callback_hsw+0x355> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,130,26,0,0 // vbroadcastss 0x1a82(%rip),%ymm3 # 3f58 <_sk_callback_hsw+0x33d> + .byte 196,226,125,24,29,158,26,0,0 // vbroadcastss 0x1a9e(%rip),%ymm3 # 3fa4 <_sk_callback_hsw+0x359> .byte 255,224 // jmpq *%rax .byte 65,137,200 // mov %ecx,%r8d .byte 65,128,224,7 // and $0x7,%r8b .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,128 // ja 246c <_sk_load_565_hsw+0x10> + .byte 119,128 // ja 249c <_sk_load_565_hsw+0x10> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2540 <_sk_load_565_hsw+0xe4> + .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2570 <_sk_load_565_hsw+0xe4> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -9685,7 +9688,7 @@ _sk_load_565_hsw: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,44,255,255,255 // jmpq 246c <_sk_load_565_hsw+0x10> + .byte 233,44,255,255,255 // jmpq 249c <_sk_load_565_hsw+0x10> .byte 244 // hlt .byte 255 // (bad) .byte 255 // (bad) @@ -9755,23 +9758,23 @@ _sk_gather_565_hsw: .byte 65,15,183,4,88 // movzwl (%r8,%rbx,2),%eax .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 - .byte 196,226,125,88,5,69,25,0,0 // vpbroadcastd 0x1945(%rip),%ymm0 # 3f5c <_sk_callback_hsw+0x341> + .byte 196,226,125,88,5,97,25,0,0 // vpbroadcastd 0x1961(%rip),%ymm0 # 3fa8 <_sk_callback_hsw+0x35d> .byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,56,25,0,0 // vbroadcastss 0x1938(%rip),%ymm1 # 3f60 <_sk_callback_hsw+0x345> + .byte 196,226,125,24,13,84,25,0,0 // vbroadcastss 0x1954(%rip),%ymm1 # 3fac <_sk_callback_hsw+0x361> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,47,25,0,0 // vpbroadcastd 0x192f(%rip),%ymm1 # 3f64 <_sk_callback_hsw+0x349> + .byte 196,226,125,88,13,75,25,0,0 // vpbroadcastd 0x194b(%rip),%ymm1 # 3fb0 <_sk_callback_hsw+0x365> .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,34,25,0,0 // vbroadcastss 0x1922(%rip),%ymm3 # 3f68 <_sk_callback_hsw+0x34d> + .byte 196,226,125,24,29,62,25,0,0 // vbroadcastss 0x193e(%rip),%ymm3 # 3fb4 <_sk_callback_hsw+0x369> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,88,29,25,25,0,0 // vpbroadcastd 0x1919(%rip),%ymm3 # 3f6c <_sk_callback_hsw+0x351> + .byte 196,226,125,88,29,53,25,0,0 // vpbroadcastd 0x1935(%rip),%ymm3 # 3fb8 <_sk_callback_hsw+0x36d> .byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,12,25,0,0 // vbroadcastss 0x190c(%rip),%ymm3 # 3f70 <_sk_callback_hsw+0x355> + .byte 196,226,125,24,29,40,25,0,0 // vbroadcastss 0x1928(%rip),%ymm3 # 3fbc <_sk_callback_hsw+0x371> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,1,25,0,0 // vbroadcastss 0x1901(%rip),%ymm3 # 3f74 <_sk_callback_hsw+0x359> + .byte 196,226,125,24,29,29,25,0,0 // vbroadcastss 0x191d(%rip),%ymm3 # 3fc0 <_sk_callback_hsw+0x375> .byte 91 // pop %rbx .byte 65,92 // pop %r12 .byte 65,94 // pop %r14 @@ -9784,11 +9787,11 @@ FUNCTION(_sk_store_565_hsw) _sk_store_565_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,238,24,0,0 // vbroadcastss 0x18ee(%rip),%ymm8 # 3f78 <_sk_callback_hsw+0x35d> + .byte 196,98,125,24,5,10,25,0,0 // vbroadcastss 0x190a(%rip),%ymm8 # 3fc4 <_sk_callback_hsw+0x379> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,53,114,241,11 // vpslld $0xb,%ymm9,%ymm9 - .byte 196,98,125,24,21,217,24,0,0 // vbroadcastss 0x18d9(%rip),%ymm10 # 3f7c <_sk_callback_hsw+0x361> + .byte 196,98,125,24,21,245,24,0,0 // vbroadcastss 0x18f5(%rip),%ymm10 # 3fc8 <_sk_callback_hsw+0x37d> .byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 .byte 196,193,45,114,242,5 // vpslld $0x5,%ymm10,%ymm10 @@ -9799,7 +9802,7 @@ _sk_store_565_hsw: .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 26e1 <_sk_store_565_hsw+0x65> + .byte 117,10 // jne 2711 <_sk_store_565_hsw+0x65> .byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9807,9 +9810,9 @@ _sk_store_565_hsw: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 26dd <_sk_store_565_hsw+0x61> + .byte 119,236 // ja 270d <_sk_store_565_hsw+0x61> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2740 <_sk_store_565_hsw+0xc4> + .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2770 <_sk_store_565_hsw+0xc4> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -9820,7 +9823,7 @@ _sk_store_565_hsw: .byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) .byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) .byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2) - .byte 235,159 // jmp 26dd <_sk_store_565_hsw+0x61> + .byte 235,159 // jmp 270d <_sk_store_565_hsw+0x61> .byte 102,144 // xchg %ax,%ax .byte 245 // cmc .byte 255 // (bad) @@ -9853,28 +9856,28 @@ _sk_load_4444_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,138,0,0,0 // jne 27f4 <_sk_load_4444_hsw+0x98> + .byte 15,133,138,0,0,0 // jne 2824 <_sk_load_4444_hsw+0x98> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 196,226,125,51,216 // vpmovzxwd %xmm0,%ymm3 - .byte 196,226,125,88,5,2,24,0,0 // vpbroadcastd 0x1802(%rip),%ymm0 # 3f80 <_sk_callback_hsw+0x365> + .byte 196,226,125,88,5,30,24,0,0 // vpbroadcastd 0x181e(%rip),%ymm0 # 3fcc <_sk_callback_hsw+0x381> .byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,245,23,0,0 // vbroadcastss 0x17f5(%rip),%ymm1 # 3f84 <_sk_callback_hsw+0x369> + .byte 196,226,125,24,13,17,24,0,0 // vbroadcastss 0x1811(%rip),%ymm1 # 3fd0 <_sk_callback_hsw+0x385> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,236,23,0,0 // vpbroadcastd 0x17ec(%rip),%ymm1 # 3f88 <_sk_callback_hsw+0x36d> + .byte 196,226,125,88,13,8,24,0,0 // vpbroadcastd 0x1808(%rip),%ymm1 # 3fd4 <_sk_callback_hsw+0x389> .byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,223,23,0,0 // vbroadcastss 0x17df(%rip),%ymm2 # 3f8c <_sk_callback_hsw+0x371> + .byte 196,226,125,24,21,251,23,0,0 // vbroadcastss 0x17fb(%rip),%ymm2 # 3fd8 <_sk_callback_hsw+0x38d> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,88,21,214,23,0,0 // vpbroadcastd 0x17d6(%rip),%ymm2 # 3f90 <_sk_callback_hsw+0x375> + .byte 196,226,125,88,21,242,23,0,0 // vpbroadcastd 0x17f2(%rip),%ymm2 # 3fdc <_sk_callback_hsw+0x391> .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,201,23,0,0 // vbroadcastss 0x17c9(%rip),%ymm8 # 3f94 <_sk_callback_hsw+0x379> + .byte 196,98,125,24,5,229,23,0,0 // vbroadcastss 0x17e5(%rip),%ymm8 # 3fe0 <_sk_callback_hsw+0x395> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,88,5,191,23,0,0 // vpbroadcastd 0x17bf(%rip),%ymm8 # 3f98 <_sk_callback_hsw+0x37d> + .byte 196,98,125,88,5,219,23,0,0 // vpbroadcastd 0x17db(%rip),%ymm8 # 3fe4 <_sk_callback_hsw+0x399> .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,177,23,0,0 // vbroadcastss 0x17b1(%rip),%ymm8 # 3f9c <_sk_callback_hsw+0x381> + .byte 196,98,125,24,5,205,23,0,0 // vbroadcastss 0x17cd(%rip),%ymm8 # 3fe8 <_sk_callback_hsw+0x39d> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9883,9 +9886,9 @@ _sk_load_4444_hsw: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,100,255,255,255 // ja 2770 <_sk_load_4444_hsw+0x14> + .byte 15,135,100,255,255,255 // ja 27a0 <_sk_load_4444_hsw+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2860 <_sk_load_4444_hsw+0x104> + .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2890 <_sk_load_4444_hsw+0x104> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -9897,7 +9900,7 @@ _sk_load_4444_hsw: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,16,255,255,255 // jmpq 2770 <_sk_load_4444_hsw+0x14> + .byte 233,16,255,255,255 // jmpq 27a0 <_sk_load_4444_hsw+0x14> .byte 244 // hlt .byte 255 // (bad) .byte 255 // (bad) @@ -9967,25 +9970,25 @@ _sk_gather_4444_hsw: .byte 65,15,183,4,88 // movzwl (%r8,%rbx,2),%eax .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0 .byte 196,226,125,51,216 // vpmovzxwd %xmm0,%ymm3 - .byte 196,226,125,88,5,105,22,0,0 // vpbroadcastd 0x1669(%rip),%ymm0 # 3fa0 <_sk_callback_hsw+0x385> + .byte 196,226,125,88,5,133,22,0,0 // vpbroadcastd 0x1685(%rip),%ymm0 # 3fec <_sk_callback_hsw+0x3a1> .byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,92,22,0,0 // vbroadcastss 0x165c(%rip),%ymm1 # 3fa4 <_sk_callback_hsw+0x389> + .byte 196,226,125,24,13,120,22,0,0 // vbroadcastss 0x1678(%rip),%ymm1 # 3ff0 <_sk_callback_hsw+0x3a5> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,83,22,0,0 // vpbroadcastd 0x1653(%rip),%ymm1 # 3fa8 <_sk_callback_hsw+0x38d> + .byte 196,226,125,88,13,111,22,0,0 // vpbroadcastd 0x166f(%rip),%ymm1 # 3ff4 <_sk_callback_hsw+0x3a9> .byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,70,22,0,0 // vbroadcastss 0x1646(%rip),%ymm2 # 3fac <_sk_callback_hsw+0x391> + .byte 196,226,125,24,21,98,22,0,0 // vbroadcastss 0x1662(%rip),%ymm2 # 3ff8 <_sk_callback_hsw+0x3ad> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,88,21,61,22,0,0 // vpbroadcastd 0x163d(%rip),%ymm2 # 3fb0 <_sk_callback_hsw+0x395> + .byte 196,226,125,88,21,89,22,0,0 // vpbroadcastd 0x1659(%rip),%ymm2 # 3ffc <_sk_callback_hsw+0x3b1> .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,48,22,0,0 // vbroadcastss 0x1630(%rip),%ymm8 # 3fb4 <_sk_callback_hsw+0x399> + .byte 196,98,125,24,5,76,22,0,0 // vbroadcastss 0x164c(%rip),%ymm8 # 4000 <_sk_callback_hsw+0x3b5> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,88,5,38,22,0,0 // vpbroadcastd 0x1626(%rip),%ymm8 # 3fb8 <_sk_callback_hsw+0x39d> + .byte 196,98,125,88,5,66,22,0,0 // vpbroadcastd 0x1642(%rip),%ymm8 # 4004 <_sk_callback_hsw+0x3b9> .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,24,22,0,0 // vbroadcastss 0x1618(%rip),%ymm8 # 3fbc <_sk_callback_hsw+0x3a1> + .byte 196,98,125,24,5,52,22,0,0 // vbroadcastss 0x1634(%rip),%ymm8 # 4008 <_sk_callback_hsw+0x3bd> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx @@ -10000,7 +10003,7 @@ FUNCTION(_sk_store_4444_hsw) _sk_store_4444_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,254,21,0,0 // vbroadcastss 0x15fe(%rip),%ymm8 # 3fc0 <_sk_callback_hsw+0x3a5> + .byte 196,98,125,24,5,26,22,0,0 // vbroadcastss 0x161a(%rip),%ymm8 # 400c <_sk_callback_hsw+0x3c1> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,53,114,241,12 // vpslld $0xc,%ymm9,%ymm9 @@ -10018,7 +10021,7 @@ _sk_store_4444_hsw: .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 2a25 <_sk_store_4444_hsw+0x71> + .byte 117,10 // jne 2a55 <_sk_store_4444_hsw+0x71> .byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -10026,9 +10029,9 @@ _sk_store_4444_hsw: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 2a21 <_sk_store_4444_hsw+0x6d> + .byte 119,236 // ja 2a51 <_sk_store_4444_hsw+0x6d> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2a84 <_sk_store_4444_hsw+0xd0> + .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2ab4 <_sk_store_4444_hsw+0xd0> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -10039,7 +10042,7 @@ _sk_store_4444_hsw: .byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) .byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) .byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2) - .byte 235,159 // jmp 2a21 <_sk_store_4444_hsw+0x6d> + .byte 235,159 // jmp 2a51 <_sk_store_4444_hsw+0x6d> .byte 102,144 // xchg %ax,%ax .byte 245 // cmc .byte 255 // (bad) @@ -10074,16 +10077,16 @@ _sk_load_8888_hsw: .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 76,3,8 // add (%rax),%r9 .byte 77,133,192 // test %r8,%r8 - .byte 117,88 // jne 2b0d <_sk_load_8888_hsw+0x6d> + .byte 117,88 // jne 2b3d <_sk_load_8888_hsw+0x6d> .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 - .byte 197,229,219,5,158,22,0,0 // vpand 0x169e(%rip),%ymm3,%ymm0 # 4160 <_sk_callback_hsw+0x545> + .byte 197,229,219,5,174,22,0,0 // vpand 0x16ae(%rip),%ymm3,%ymm0 # 41a0 <_sk_callback_hsw+0x555> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,245,20,0,0 // vbroadcastss 0x14f5(%rip),%ymm8 # 3fc4 <_sk_callback_hsw+0x3a9> + .byte 196,98,125,24,5,17,21,0,0 // vbroadcastss 0x1511(%rip),%ymm8 # 4010 <_sk_callback_hsw+0x3c5> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,226,101,0,13,163,22,0,0 // vpshufb 0x16a3(%rip),%ymm3,%ymm1 # 4180 <_sk_callback_hsw+0x565> + .byte 196,226,101,0,13,179,22,0,0 // vpshufb 0x16b3(%rip),%ymm3,%ymm1 # 41c0 <_sk_callback_hsw+0x575> .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,226,101,0,21,177,22,0,0 // vpshufb 0x16b1(%rip),%ymm3,%ymm2 # 41a0 <_sk_callback_hsw+0x585> + .byte 196,226,101,0,21,193,22,0,0 // vpshufb 0x16c1(%rip),%ymm3,%ymm2 # 41e0 <_sk_callback_hsw+0x595> .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 @@ -10100,7 +10103,7 @@ _sk_load_8888_hsw: .byte 196,225,249,110,192 // vmovq %rax,%xmm0 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 - .byte 235,135 // jmp 2aba <_sk_load_8888_hsw+0x1a> + .byte 235,135 // jmp 2aea <_sk_load_8888_hsw+0x1a> HIDDEN _sk_gather_8888_hsw .globl _sk_gather_8888_hsw @@ -10115,14 +10118,14 @@ _sk_gather_8888_hsw: .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 .byte 196,194,117,144,28,128 // vpgatherdd %ymm1,(%r8,%ymm0,4),%ymm3 - .byte 197,229,219,5,95,22,0,0 // vpand 0x165f(%rip),%ymm3,%ymm0 # 41c0 <_sk_callback_hsw+0x5a5> + .byte 197,229,219,5,111,22,0,0 // vpand 0x166f(%rip),%ymm3,%ymm0 # 4200 <_sk_callback_hsw+0x5b5> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,90,20,0,0 // vbroadcastss 0x145a(%rip),%ymm8 # 3fc8 <_sk_callback_hsw+0x3ad> + .byte 196,98,125,24,5,118,20,0,0 // vbroadcastss 0x1476(%rip),%ymm8 # 4014 <_sk_callback_hsw+0x3c9> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,226,101,0,13,100,22,0,0 // vpshufb 0x1664(%rip),%ymm3,%ymm1 # 41e0 <_sk_callback_hsw+0x5c5> + .byte 196,226,101,0,13,116,22,0,0 // vpshufb 0x1674(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x5d5> .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,226,101,0,21,114,22,0,0 // vpshufb 0x1672(%rip),%ymm3,%ymm2 # 4200 <_sk_callback_hsw+0x5e5> + .byte 196,226,101,0,21,130,22,0,0 // vpshufb 0x1682(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x5f5> .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 @@ -10139,7 +10142,7 @@ _sk_store_8888_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 76,3,8 // add (%rax),%r9 - .byte 196,98,125,24,5,10,20,0,0 // vbroadcastss 0x140a(%rip),%ymm8 # 3fcc <_sk_callback_hsw+0x3b1> + .byte 196,98,125,24,5,38,20,0,0 // vbroadcastss 0x1426(%rip),%ymm8 # 4018 <_sk_callback_hsw+0x3cd> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,65,116,89,208 // vmulps %ymm8,%ymm1,%ymm10 @@ -10155,7 +10158,7 @@ _sk_store_8888_hsw: .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8 .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8 .byte 77,133,192 // test %r8,%r8 - .byte 117,12 // jne 2c1c <_sk_store_8888_hsw+0x73> + .byte 117,12 // jne 2c4c <_sk_store_8888_hsw+0x73> .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,193 // mov %r8,%rcx @@ -10168,7 +10171,7 @@ _sk_store_8888_hsw: .byte 196,97,249,110,200 // vmovq %rax,%xmm9 .byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9 .byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9) - .byte 235,211 // jmp 2c15 <_sk_store_8888_hsw+0x6c> + .byte 235,211 // jmp 2c45 <_sk_store_8888_hsw+0x6c> HIDDEN _sk_load_f16_hsw .globl _sk_load_f16_hsw @@ -10177,7 +10180,7 @@ _sk_load_f16_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 117,97 // jne 2cad <_sk_load_f16_hsw+0x6b> + .byte 117,97 // jne 2cdd <_sk_load_f16_hsw+0x6b> .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -10203,29 +10206,29 @@ _sk_load_f16_hsw: .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,79 // je 2d0c <_sk_load_f16_hsw+0xca> + .byte 116,79 // je 2d3c <_sk_load_f16_hsw+0xca> .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,67 // jb 2d0c <_sk_load_f16_hsw+0xca> + .byte 114,67 // jb 2d3c <_sk_load_f16_hsw+0xca> .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,68 // je 2d19 <_sk_load_f16_hsw+0xd7> + .byte 116,68 // je 2d49 <_sk_load_f16_hsw+0xd7> .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,56 // jb 2d19 <_sk_load_f16_hsw+0xd7> + .byte 114,56 // jb 2d49 <_sk_load_f16_hsw+0xd7> .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,114,255,255,255 // je 2c63 <_sk_load_f16_hsw+0x21> + .byte 15,132,114,255,255,255 // je 2c93 <_sk_load_f16_hsw+0x21> .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,98,255,255,255 // jb 2c63 <_sk_load_f16_hsw+0x21> + .byte 15,130,98,255,255,255 // jb 2c93 <_sk_load_f16_hsw+0x21> .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 - .byte 233,87,255,255,255 // jmpq 2c63 <_sk_load_f16_hsw+0x21> + .byte 233,87,255,255,255 // jmpq 2c93 <_sk_load_f16_hsw+0x21> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,74,255,255,255 // jmpq 2c63 <_sk_load_f16_hsw+0x21> + .byte 233,74,255,255,255 // jmpq 2c93 <_sk_load_f16_hsw+0x21> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,65,255,255,255 // jmpq 2c63 <_sk_load_f16_hsw+0x21> + .byte 233,65,255,255,255 // jmpq 2c93 <_sk_load_f16_hsw+0x21> HIDDEN _sk_gather_f16_hsw .globl _sk_gather_f16_hsw @@ -10283,7 +10286,7 @@ _sk_store_f16_hsw: .byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9 .byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,27 // jne 2e11 <_sk_store_f16_hsw+0x65> + .byte 117,27 // jne 2e41 <_sk_store_f16_hsw+0x65> .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8) .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8) .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8) @@ -10292,22 +10295,22 @@ _sk_store_f16_hsw: .byte 255,224 // jmpq *%rax .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,241 // je 2e0d <_sk_store_f16_hsw+0x61> + .byte 116,241 // je 2e3d <_sk_store_f16_hsw+0x61> .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,229 // jb 2e0d <_sk_store_f16_hsw+0x61> + .byte 114,229 // jb 2e3d <_sk_store_f16_hsw+0x61> .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) - .byte 116,221 // je 2e0d <_sk_store_f16_hsw+0x61> + .byte 116,221 // je 2e3d <_sk_store_f16_hsw+0x61> .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,209 // jb 2e0d <_sk_store_f16_hsw+0x61> + .byte 114,209 // jb 2e3d <_sk_store_f16_hsw+0x61> .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) - .byte 116,201 // je 2e0d <_sk_store_f16_hsw+0x61> + .byte 116,201 // je 2e3d <_sk_store_f16_hsw+0x61> .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,189 // jb 2e0d <_sk_store_f16_hsw+0x61> + .byte 114,189 // jb 2e3d <_sk_store_f16_hsw+0x61> .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) - .byte 235,181 // jmp 2e0d <_sk_store_f16_hsw+0x61> + .byte 235,181 // jmp 2e3d <_sk_store_f16_hsw+0x61> HIDDEN _sk_load_u16_be_hsw .globl _sk_load_u16_be_hsw @@ -10317,7 +10320,7 @@ _sk_load_u16_be_hsw: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,204,0,0,0 // jne 2f3a <_sk_load_u16_be_hsw+0xe2> + .byte 15,133,204,0,0,0 // jne 2f6a <_sk_load_u16_be_hsw+0xe2> .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -10336,7 +10339,7 @@ _sk_load_u16_be_hsw: .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,21,1,17,0,0 // vbroadcastss 0x1101(%rip),%ymm10 # 3fd0 <_sk_callback_hsw+0x3b5> + .byte 196,98,125,24,21,29,17,0,0 // vbroadcastss 0x111d(%rip),%ymm10 # 401c <_sk_callback_hsw+0x3d1> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -10364,29 +10367,29 @@ _sk_load_u16_be_hsw: .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 2fa0 <_sk_load_u16_be_hsw+0x148> + .byte 116,85 // je 2fd0 <_sk_load_u16_be_hsw+0x148> .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 2fa0 <_sk_load_u16_be_hsw+0x148> + .byte 114,72 // jb 2fd0 <_sk_load_u16_be_hsw+0x148> .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 2fad <_sk_load_u16_be_hsw+0x155> + .byte 116,72 // je 2fdd <_sk_load_u16_be_hsw+0x155> .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 2fad <_sk_load_u16_be_hsw+0x155> + .byte 114,59 // jb 2fdd <_sk_load_u16_be_hsw+0x155> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,6,255,255,255 // je 2e89 <_sk_load_u16_be_hsw+0x31> + .byte 15,132,6,255,255,255 // je 2eb9 <_sk_load_u16_be_hsw+0x31> .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,245,254,255,255 // jb 2e89 <_sk_load_u16_be_hsw+0x31> + .byte 15,130,245,254,255,255 // jb 2eb9 <_sk_load_u16_be_hsw+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,233,254,255,255 // jmpq 2e89 <_sk_load_u16_be_hsw+0x31> + .byte 233,233,254,255,255 // jmpq 2eb9 <_sk_load_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,220,254,255,255 // jmpq 2e89 <_sk_load_u16_be_hsw+0x31> + .byte 233,220,254,255,255 // jmpq 2eb9 <_sk_load_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,211,254,255,255 // jmpq 2e89 <_sk_load_u16_be_hsw+0x31> + .byte 233,211,254,255,255 // jmpq 2eb9 <_sk_load_u16_be_hsw+0x31> HIDDEN _sk_load_rgb_u16_be_hsw .globl _sk_load_rgb_u16_be_hsw @@ -10396,7 +10399,7 @@ _sk_load_rgb_u16_be_hsw: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,204,0,0,0 // jne 3094 <_sk_load_rgb_u16_be_hsw+0xde> + .byte 15,133,204,0,0,0 // jne 30c4 <_sk_load_rgb_u16_be_hsw+0xde> .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -10420,7 +10423,7 @@ _sk_load_rgb_u16_be_hsw: .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,21,146,15,0,0 // vbroadcastss 0xf92(%rip),%ymm10 # 3fd4 <_sk_callback_hsw+0x3b9> + .byte 196,98,125,24,21,174,15,0,0 // vbroadcastss 0xfae(%rip),%ymm10 # 4020 <_sk_callback_hsw+0x3d5> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -10437,41 +10440,41 @@ _sk_load_rgb_u16_be_hsw: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,70,15,0,0 // vbroadcastss 0xf46(%rip),%ymm3 # 3fd8 <_sk_callback_hsw+0x3bd> + .byte 196,226,125,24,29,98,15,0,0 // vbroadcastss 0xf62(%rip),%ymm3 # 4024 <_sk_callback_hsw+0x3d9> .byte 255,224 // jmpq *%rax .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 30ad <_sk_load_rgb_u16_be_hsw+0xf7> - .byte 233,79,255,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,5 // jne 30dd <_sk_load_rgb_u16_be_hsw+0xf7> + .byte 233,79,255,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 30dc <_sk_load_rgb_u16_be_hsw+0x126> + .byte 114,26 // jb 310c <_sk_load_rgb_u16_be_hsw+0x126> .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 30e1 <_sk_load_rgb_u16_be_hsw+0x12b> - .byte 233,32,255,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,27,255,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 3111 <_sk_load_rgb_u16_be_hsw+0x12b> + .byte 233,32,255,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,27,255,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 3110 <_sk_load_rgb_u16_be_hsw+0x15a> + .byte 114,26 // jb 3140 <_sk_load_rgb_u16_be_hsw+0x15a> .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 3115 <_sk_load_rgb_u16_be_hsw+0x15f> - .byte 233,236,254,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,231,254,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 3145 <_sk_load_rgb_u16_be_hsw+0x15f> + .byte 233,236,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,231,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 313e <_sk_load_rgb_u16_be_hsw+0x188> + .byte 114,20 // jb 316e <_sk_load_rgb_u16_be_hsw+0x188> .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - .byte 233,190,254,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,185,254,255,255 // jmpq 2ffc <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,190,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,185,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46> HIDDEN _sk_store_u16_be_hsw .globl _sk_store_u16_be_hsw @@ -10480,7 +10483,7 @@ _sk_store_u16_be_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax - .byte 196,98,125,24,5,131,14,0,0 // vbroadcastss 0xe83(%rip),%ymm8 # 3fdc <_sk_callback_hsw+0x3c1> + .byte 196,98,125,24,5,159,14,0,0 // vbroadcastss 0xe9f(%rip),%ymm8 # 4028 <_sk_callback_hsw+0x3dd> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 @@ -10518,7 +10521,7 @@ _sk_store_u16_be_hsw: .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,31 // jne 323d <_sk_store_u16_be_hsw+0xfa> + .byte 117,31 // jne 326d <_sk_store_u16_be_hsw+0xfa> .byte 196,65,120,17,28,64 // vmovups %xmm11,(%r8,%rax,2) .byte 196,65,120,17,84,64,16 // vmovups %xmm10,0x10(%r8,%rax,2) .byte 196,65,120,17,76,64,32 // vmovups %xmm9,0x20(%r8,%rax,2) @@ -10527,22 +10530,22 @@ _sk_store_u16_be_hsw: .byte 255,224 // jmpq *%rax .byte 196,65,121,214,28,64 // vmovq %xmm11,(%r8,%rax,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 116,240 // je 3269 <_sk_store_u16_be_hsw+0xf6> .byte 196,65,121,23,92,64,8 // vmovhpd %xmm11,0x8(%r8,%rax,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 114,227 // jb 3269 <_sk_store_u16_be_hsw+0xf6> .byte 196,65,121,214,84,64,16 // vmovq %xmm10,0x10(%r8,%rax,2) - .byte 116,218 // je 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 116,218 // je 3269 <_sk_store_u16_be_hsw+0xf6> .byte 196,65,121,23,84,64,24 // vmovhpd %xmm10,0x18(%r8,%rax,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 114,205 // jb 3269 <_sk_store_u16_be_hsw+0xf6> .byte 196,65,121,214,76,64,32 // vmovq %xmm9,0x20(%r8,%rax,2) - .byte 116,196 // je 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 116,196 // je 3269 <_sk_store_u16_be_hsw+0xf6> .byte 196,65,121,23,76,64,40 // vmovhpd %xmm9,0x28(%r8,%rax,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 114,183 // jb 3269 <_sk_store_u16_be_hsw+0xf6> .byte 196,65,121,214,68,64,48 // vmovq %xmm8,0x30(%r8,%rax,2) - .byte 235,174 // jmp 3239 <_sk_store_u16_be_hsw+0xf6> + .byte 235,174 // jmp 3269 <_sk_store_u16_be_hsw+0xf6> HIDDEN _sk_load_f32_hsw .globl _sk_load_f32_hsw @@ -10550,10 +10553,10 @@ FUNCTION(_sk_load_f32_hsw) _sk_load_f32_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 119,110 // ja 3301 <_sk_load_f32_hsw+0x76> + .byte 119,110 // ja 3331 <_sk_load_f32_hsw+0x76> .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 - .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 332c <_sk_load_f32_hsw+0xa1> + .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 335c <_sk_load_f32_hsw+0xa1> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -10614,7 +10617,7 @@ _sk_store_f32_hsw: .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 .byte 72,133,201 // test %rcx,%rcx - .byte 117,55 // jne 33b9 <_sk_store_f32_hsw+0x6d> + .byte 117,55 // jne 33e9 <_sk_store_f32_hsw+0x6d> .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -10627,22 +10630,22 @@ _sk_store_f32_hsw: .byte 255,224 // jmpq *%rax .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 33b5 <_sk_store_f32_hsw+0x69> + .byte 116,240 // je 33e5 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 33b5 <_sk_store_f32_hsw+0x69> + .byte 114,227 // jb 33e5 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 33b5 <_sk_store_f32_hsw+0x69> + .byte 116,218 // je 33e5 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 33b5 <_sk_store_f32_hsw+0x69> + .byte 114,205 // jb 33e5 <_sk_store_f32_hsw+0x69> .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - .byte 116,195 // je 33b5 <_sk_store_f32_hsw+0x69> + .byte 116,195 // je 33e5 <_sk_store_f32_hsw+0x69> .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,181 // jb 33b5 <_sk_store_f32_hsw+0x69> + .byte 114,181 // jb 33e5 <_sk_store_f32_hsw+0x69> .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - .byte 235,171 // jmp 33b5 <_sk_store_f32_hsw+0x69> + .byte 235,171 // jmp 33e5 <_sk_store_f32_hsw+0x69> HIDDEN _sk_clamp_x_hsw .globl _sk_clamp_x_hsw @@ -10752,11 +10755,11 @@ HIDDEN _sk_luminance_to_alpha_hsw .globl _sk_luminance_to_alpha_hsw FUNCTION(_sk_luminance_to_alpha_hsw) _sk_luminance_to_alpha_hsw: - .byte 196,226,125,24,29,157,10,0,0 // vbroadcastss 0xa9d(%rip),%ymm3 # 3fe0 <_sk_callback_hsw+0x3c5> - .byte 196,98,125,24,5,152,10,0,0 // vbroadcastss 0xa98(%rip),%ymm8 # 3fe4 <_sk_callback_hsw+0x3c9> + .byte 196,226,125,24,29,185,10,0,0 // vbroadcastss 0xab9(%rip),%ymm3 # 402c <_sk_callback_hsw+0x3e1> + .byte 196,98,125,24,5,180,10,0,0 // vbroadcastss 0xab4(%rip),%ymm8 # 4030 <_sk_callback_hsw+0x3e5> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 196,226,125,184,203 // vfmadd231ps %ymm3,%ymm0,%ymm1 - .byte 196,226,125,24,29,137,10,0,0 // vbroadcastss 0xa89(%rip),%ymm3 # 3fe8 <_sk_callback_hsw+0x3cd> + .byte 196,226,125,24,29,165,10,0,0 // vbroadcastss 0xaa5(%rip),%ymm3 # 4034 <_sk_callback_hsw+0x3e9> .byte 196,226,109,168,217 // vfmadd213ps %ymm1,%ymm2,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -10901,7 +10904,7 @@ _sk_linear_gradient_hsw: .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 .byte 76,139,0 // mov (%rax),%r8 .byte 77,133,192 // test %r8,%r8 - .byte 15,132,143,0,0,0 // je 3837 <_sk_linear_gradient_hsw+0xb5> + .byte 15,132,143,0,0,0 // je 3867 <_sk_linear_gradient_hsw+0xb5> .byte 72,139,64,8 // mov 0x8(%rax),%rax .byte 72,131,192,32 // add $0x20,%rax .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 @@ -10928,8 +10931,8 @@ _sk_linear_gradient_hsw: .byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9 .byte 72,131,192,36 // add $0x24,%rax .byte 73,255,200 // dec %r8 - .byte 117,140 // jne 37c1 <_sk_linear_gradient_hsw+0x3f> - .byte 235,17 // jmp 3848 <_sk_linear_gradient_hsw+0xc6> + .byte 117,140 // jne 37f1 <_sk_linear_gradient_hsw+0x3f> + .byte 235,17 // jmp 3878 <_sk_linear_gradient_hsw+0xc6> .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 @@ -10968,7 +10971,7 @@ HIDDEN _sk_save_xy_hsw FUNCTION(_sk_save_xy_hsw) _sk_save_xy_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,48,7,0,0 // vbroadcastss 0x730(%rip),%ymm8 # 3fec <_sk_callback_hsw+0x3d1> + .byte 196,98,125,24,5,76,7,0,0 // vbroadcastss 0x74c(%rip),%ymm8 # 4038 <_sk_callback_hsw+0x3ed> .byte 196,65,124,88,200 // vaddps %ymm8,%ymm0,%ymm9 .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 @@ -11002,9 +11005,9 @@ HIDDEN _sk_bilinear_nx_hsw FUNCTION(_sk_bilinear_nx_hsw) _sk_bilinear_nx_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,196,6,0,0 // vbroadcastss 0x6c4(%rip),%ymm0 # 3ff0 <_sk_callback_hsw+0x3d5> + .byte 196,226,125,24,5,224,6,0,0 // vbroadcastss 0x6e0(%rip),%ymm0 # 403c <_sk_callback_hsw+0x3f1> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,187,6,0,0 // vbroadcastss 0x6bb(%rip),%ymm8 # 3ff4 <_sk_callback_hsw+0x3d9> + .byte 196,98,125,24,5,215,6,0,0 // vbroadcastss 0x6d7(%rip),%ymm8 # 4040 <_sk_callback_hsw+0x3f5> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -11015,7 +11018,7 @@ HIDDEN _sk_bilinear_px_hsw FUNCTION(_sk_bilinear_px_hsw) _sk_bilinear_px_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,163,6,0,0 // vbroadcastss 0x6a3(%rip),%ymm0 # 3ff8 <_sk_callback_hsw+0x3dd> + .byte 196,226,125,24,5,191,6,0,0 // vbroadcastss 0x6bf(%rip),%ymm0 # 4044 <_sk_callback_hsw+0x3f9> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -11027,9 +11030,9 @@ HIDDEN _sk_bilinear_ny_hsw FUNCTION(_sk_bilinear_ny_hsw) _sk_bilinear_ny_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,135,6,0,0 // vbroadcastss 0x687(%rip),%ymm1 # 3ffc <_sk_callback_hsw+0x3e1> + .byte 196,226,125,24,13,163,6,0,0 // vbroadcastss 0x6a3(%rip),%ymm1 # 4048 <_sk_callback_hsw+0x3fd> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,125,6,0,0 // vbroadcastss 0x67d(%rip),%ymm8 # 4000 <_sk_callback_hsw+0x3e5> + .byte 196,98,125,24,5,153,6,0,0 // vbroadcastss 0x699(%rip),%ymm8 # 404c <_sk_callback_hsw+0x401> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -11040,7 +11043,7 @@ HIDDEN _sk_bilinear_py_hsw FUNCTION(_sk_bilinear_py_hsw) _sk_bilinear_py_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,101,6,0,0 // vbroadcastss 0x665(%rip),%ymm1 # 4004 <_sk_callback_hsw+0x3e9> + .byte 196,226,125,24,13,129,6,0,0 // vbroadcastss 0x681(%rip),%ymm1 # 4050 <_sk_callback_hsw+0x405> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -11052,13 +11055,13 @@ HIDDEN _sk_bicubic_n3x_hsw FUNCTION(_sk_bicubic_n3x_hsw) _sk_bicubic_n3x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,72,6,0,0 // vbroadcastss 0x648(%rip),%ymm0 # 4008 <_sk_callback_hsw+0x3ed> + .byte 196,226,125,24,5,100,6,0,0 // vbroadcastss 0x664(%rip),%ymm0 # 4054 <_sk_callback_hsw+0x409> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,63,6,0,0 // vbroadcastss 0x63f(%rip),%ymm8 # 400c <_sk_callback_hsw+0x3f1> + .byte 196,98,125,24,5,91,6,0,0 // vbroadcastss 0x65b(%rip),%ymm8 # 4058 <_sk_callback_hsw+0x40d> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,48,6,0,0 // vbroadcastss 0x630(%rip),%ymm10 # 4010 <_sk_callback_hsw+0x3f5> - .byte 196,98,125,24,29,43,6,0,0 // vbroadcastss 0x62b(%rip),%ymm11 # 4014 <_sk_callback_hsw+0x3f9> + .byte 196,98,125,24,21,76,6,0,0 // vbroadcastss 0x64c(%rip),%ymm10 # 405c <_sk_callback_hsw+0x411> + .byte 196,98,125,24,29,71,6,0,0 // vbroadcastss 0x647(%rip),%ymm11 # 4060 <_sk_callback_hsw+0x415> .byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11 .byte 196,65,36,89,193 // vmulps %ymm9,%ymm11,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -11070,16 +11073,16 @@ HIDDEN _sk_bicubic_n1x_hsw FUNCTION(_sk_bicubic_n1x_hsw) _sk_bicubic_n1x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,14,6,0,0 // vbroadcastss 0x60e(%rip),%ymm0 # 4018 <_sk_callback_hsw+0x3fd> + .byte 196,226,125,24,5,42,6,0,0 // vbroadcastss 0x62a(%rip),%ymm0 # 4064 <_sk_callback_hsw+0x419> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,5,6,0,0 // vbroadcastss 0x605(%rip),%ymm8 # 401c <_sk_callback_hsw+0x401> + .byte 196,98,125,24,5,33,6,0,0 // vbroadcastss 0x621(%rip),%ymm8 # 4068 <_sk_callback_hsw+0x41d> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,251,5,0,0 // vbroadcastss 0x5fb(%rip),%ymm9 # 4020 <_sk_callback_hsw+0x405> - .byte 196,98,125,24,21,246,5,0,0 // vbroadcastss 0x5f6(%rip),%ymm10 # 4024 <_sk_callback_hsw+0x409> + .byte 196,98,125,24,13,23,6,0,0 // vbroadcastss 0x617(%rip),%ymm9 # 406c <_sk_callback_hsw+0x421> + .byte 196,98,125,24,21,18,6,0,0 // vbroadcastss 0x612(%rip),%ymm10 # 4070 <_sk_callback_hsw+0x425> .byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,13,236,5,0,0 // vbroadcastss 0x5ec(%rip),%ymm9 # 4028 <_sk_callback_hsw+0x40d> + .byte 196,98,125,24,13,8,6,0,0 // vbroadcastss 0x608(%rip),%ymm9 # 4074 <_sk_callback_hsw+0x429> .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9 - .byte 196,98,125,24,21,226,5,0,0 // vbroadcastss 0x5e2(%rip),%ymm10 # 402c <_sk_callback_hsw+0x411> + .byte 196,98,125,24,21,254,5,0,0 // vbroadcastss 0x5fe(%rip),%ymm10 # 4078 <_sk_callback_hsw+0x42d> .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10 .byte 197,124,17,144,128,0,0,0 // vmovups %ymm10,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -11090,14 +11093,14 @@ HIDDEN _sk_bicubic_p1x_hsw FUNCTION(_sk_bicubic_p1x_hsw) _sk_bicubic_p1x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,202,5,0,0 // vbroadcastss 0x5ca(%rip),%ymm8 # 4030 <_sk_callback_hsw+0x415> + .byte 196,98,125,24,5,230,5,0,0 // vbroadcastss 0x5e6(%rip),%ymm8 # 407c <_sk_callback_hsw+0x431> .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 - .byte 196,98,125,24,21,188,5,0,0 // vbroadcastss 0x5bc(%rip),%ymm10 # 4034 <_sk_callback_hsw+0x419> - .byte 196,98,125,24,29,183,5,0,0 // vbroadcastss 0x5b7(%rip),%ymm11 # 4038 <_sk_callback_hsw+0x41d> + .byte 196,98,125,24,21,216,5,0,0 // vbroadcastss 0x5d8(%rip),%ymm10 # 4080 <_sk_callback_hsw+0x435> + .byte 196,98,125,24,29,211,5,0,0 // vbroadcastss 0x5d3(%rip),%ymm11 # 4084 <_sk_callback_hsw+0x439> .byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11 .byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11 - .byte 196,98,125,24,5,168,5,0,0 // vbroadcastss 0x5a8(%rip),%ymm8 # 403c <_sk_callback_hsw+0x421> + .byte 196,98,125,24,5,196,5,0,0 // vbroadcastss 0x5c4(%rip),%ymm8 # 4088 <_sk_callback_hsw+0x43d> .byte 196,66,53,184,195 // vfmadd231ps %ymm11,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -11108,12 +11111,12 @@ HIDDEN _sk_bicubic_p3x_hsw FUNCTION(_sk_bicubic_p3x_hsw) _sk_bicubic_p3x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,144,5,0,0 // vbroadcastss 0x590(%rip),%ymm0 # 4040 <_sk_callback_hsw+0x425> + .byte 196,226,125,24,5,172,5,0,0 // vbroadcastss 0x5ac(%rip),%ymm0 # 408c <_sk_callback_hsw+0x441> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,125,5,0,0 // vbroadcastss 0x57d(%rip),%ymm10 # 4044 <_sk_callback_hsw+0x429> - .byte 196,98,125,24,29,120,5,0,0 // vbroadcastss 0x578(%rip),%ymm11 # 4048 <_sk_callback_hsw+0x42d> + .byte 196,98,125,24,21,153,5,0,0 // vbroadcastss 0x599(%rip),%ymm10 # 4090 <_sk_callback_hsw+0x445> + .byte 196,98,125,24,29,148,5,0,0 // vbroadcastss 0x594(%rip),%ymm11 # 4094 <_sk_callback_hsw+0x449> .byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11 .byte 196,65,52,89,195 // vmulps %ymm11,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -11125,13 +11128,13 @@ HIDDEN _sk_bicubic_n3y_hsw FUNCTION(_sk_bicubic_n3y_hsw) _sk_bicubic_n3y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,91,5,0,0 // vbroadcastss 0x55b(%rip),%ymm1 # 404c <_sk_callback_hsw+0x431> + .byte 196,226,125,24,13,119,5,0,0 // vbroadcastss 0x577(%rip),%ymm1 # 4098 <_sk_callback_hsw+0x44d> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,81,5,0,0 // vbroadcastss 0x551(%rip),%ymm8 # 4050 <_sk_callback_hsw+0x435> + .byte 196,98,125,24,5,109,5,0,0 // vbroadcastss 0x56d(%rip),%ymm8 # 409c <_sk_callback_hsw+0x451> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,66,5,0,0 // vbroadcastss 0x542(%rip),%ymm10 # 4054 <_sk_callback_hsw+0x439> - .byte 196,98,125,24,29,61,5,0,0 // vbroadcastss 0x53d(%rip),%ymm11 # 4058 <_sk_callback_hsw+0x43d> + .byte 196,98,125,24,21,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm10 # 40a0 <_sk_callback_hsw+0x455> + .byte 196,98,125,24,29,89,5,0,0 // vbroadcastss 0x559(%rip),%ymm11 # 40a4 <_sk_callback_hsw+0x459> .byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11 .byte 196,65,36,89,193 // vmulps %ymm9,%ymm11,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -11143,16 +11146,16 @@ HIDDEN _sk_bicubic_n1y_hsw FUNCTION(_sk_bicubic_n1y_hsw) _sk_bicubic_n1y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,32,5,0,0 // vbroadcastss 0x520(%rip),%ymm1 # 405c <_sk_callback_hsw+0x441> + .byte 196,226,125,24,13,60,5,0,0 // vbroadcastss 0x53c(%rip),%ymm1 # 40a8 <_sk_callback_hsw+0x45d> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,22,5,0,0 // vbroadcastss 0x516(%rip),%ymm8 # 4060 <_sk_callback_hsw+0x445> + .byte 196,98,125,24,5,50,5,0,0 // vbroadcastss 0x532(%rip),%ymm8 # 40ac <_sk_callback_hsw+0x461> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,12,5,0,0 // vbroadcastss 0x50c(%rip),%ymm9 # 4064 <_sk_callback_hsw+0x449> - .byte 196,98,125,24,21,7,5,0,0 // vbroadcastss 0x507(%rip),%ymm10 # 4068 <_sk_callback_hsw+0x44d> + .byte 196,98,125,24,13,40,5,0,0 // vbroadcastss 0x528(%rip),%ymm9 # 40b0 <_sk_callback_hsw+0x465> + .byte 196,98,125,24,21,35,5,0,0 // vbroadcastss 0x523(%rip),%ymm10 # 40b4 <_sk_callback_hsw+0x469> .byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,13,253,4,0,0 // vbroadcastss 0x4fd(%rip),%ymm9 # 406c <_sk_callback_hsw+0x451> + .byte 196,98,125,24,13,25,5,0,0 // vbroadcastss 0x519(%rip),%ymm9 # 40b8 <_sk_callback_hsw+0x46d> .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9 - .byte 196,98,125,24,21,243,4,0,0 // vbroadcastss 0x4f3(%rip),%ymm10 # 4070 <_sk_callback_hsw+0x455> + .byte 196,98,125,24,21,15,5,0,0 // vbroadcastss 0x50f(%rip),%ymm10 # 40bc <_sk_callback_hsw+0x471> .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10 .byte 197,124,17,144,160,0,0,0 // vmovups %ymm10,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -11163,14 +11166,14 @@ HIDDEN _sk_bicubic_p1y_hsw FUNCTION(_sk_bicubic_p1y_hsw) _sk_bicubic_p1y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,219,4,0,0 // vbroadcastss 0x4db(%rip),%ymm8 # 4074 <_sk_callback_hsw+0x459> + .byte 196,98,125,24,5,247,4,0,0 // vbroadcastss 0x4f7(%rip),%ymm8 # 40c0 <_sk_callback_hsw+0x475> .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 - .byte 196,98,125,24,21,204,4,0,0 // vbroadcastss 0x4cc(%rip),%ymm10 # 4078 <_sk_callback_hsw+0x45d> - .byte 196,98,125,24,29,199,4,0,0 // vbroadcastss 0x4c7(%rip),%ymm11 # 407c <_sk_callback_hsw+0x461> + .byte 196,98,125,24,21,232,4,0,0 // vbroadcastss 0x4e8(%rip),%ymm10 # 40c4 <_sk_callback_hsw+0x479> + .byte 196,98,125,24,29,227,4,0,0 // vbroadcastss 0x4e3(%rip),%ymm11 # 40c8 <_sk_callback_hsw+0x47d> .byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11 .byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11 - .byte 196,98,125,24,5,184,4,0,0 // vbroadcastss 0x4b8(%rip),%ymm8 # 4080 <_sk_callback_hsw+0x465> + .byte 196,98,125,24,5,212,4,0,0 // vbroadcastss 0x4d4(%rip),%ymm8 # 40cc <_sk_callback_hsw+0x481> .byte 196,66,53,184,195 // vfmadd231ps %ymm11,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -11181,12 +11184,12 @@ HIDDEN _sk_bicubic_p3y_hsw FUNCTION(_sk_bicubic_p3y_hsw) _sk_bicubic_p3y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,160,4,0,0 // vbroadcastss 0x4a0(%rip),%ymm1 # 4084 <_sk_callback_hsw+0x469> + .byte 196,226,125,24,13,188,4,0,0 // vbroadcastss 0x4bc(%rip),%ymm1 # 40d0 <_sk_callback_hsw+0x485> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,140,4,0,0 // vbroadcastss 0x48c(%rip),%ymm10 # 4088 <_sk_callback_hsw+0x46d> - .byte 196,98,125,24,29,135,4,0,0 // vbroadcastss 0x487(%rip),%ymm11 # 408c <_sk_callback_hsw+0x471> + .byte 196,98,125,24,21,168,4,0,0 // vbroadcastss 0x4a8(%rip),%ymm10 # 40d4 <_sk_callback_hsw+0x489> + .byte 196,98,125,24,29,163,4,0,0 // vbroadcastss 0x4a3(%rip),%ymm11 # 40d8 <_sk_callback_hsw+0x48d> .byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11 .byte 196,65,52,89,195 // vmulps %ymm11,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -11326,11 +11329,23 @@ BALIGN4 .byte 0,128,64,171,170,42 // add %al,0x2aaaab40(%rax) .byte 62,0,0 // add %al,%ds:(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 129,128,128,59,129,128,128,59,0,248 // addl $0xf8003b80,-0x7f7ec480(%rax) .byte 0,0 // add %al,(%rax) - .byte 8,33 // or %ah,(%rcx) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 64,171 // rex stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,0,0 // add %al,%ds:(%rax) + .byte 128,191,0,0,192,64,171 // cmpb $0xab,0x40c00000(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 190,129,128,128,59 // mov $0x3b808081,%esi + .byte 129,128,128,59,0,248,0,0,8,33 // addl $0x21080000,-0x7ffc480(%rax) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 3dfd <.literal4+0xbd> + .byte 224,7 // loopne 3e49 <.literal4+0xd9> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -11344,10 +11359,10 @@ BALIGN4 .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) .byte 0,52,255 // add %dh,(%rdi,%rdi,8) .byte 255 // (bad) - .byte 127,0 // jg 3e28 <.literal4+0xe8> + .byte 127,0 // jg 3e74 <.literal4+0x104> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 3ea1 <.literal4+0x161> + .byte 119,115 // ja 3eed <.literal4+0x17d> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -11361,10 +11376,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 3e5c <.literal4+0x11c> + .byte 127,0 // jg 3ea8 <.literal4+0x138> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 3ed5 <.literal4+0x195> + .byte 119,115 // ja 3f21 <.literal4+0x1b1> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -11378,10 +11393,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 3e90 <.literal4+0x150> + .byte 127,0 // jg 3edc <.literal4+0x16c> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 3f09 <.literal4+0x1c9> + .byte 119,115 // ja 3f55 <.literal4+0x1e5> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -11395,10 +11410,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 3ec4 <.literal4+0x184> + .byte 127,0 // jg 3f10 <.literal4+0x1a0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 3f3d <.literal4+0x1fd> + .byte 119,115 // ja 3f89 <.literal4+0x219> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -11411,7 +11426,7 @@ BALIGN4 .byte 0,75,0 // add %cl,0x0(%rbx) .byte 0,128,63,0,0,200 // add %al,-0x37ffffc1(%rax) .byte 66,0,0 // rex.X add %al,(%rax) - .byte 127,67 // jg 3f3b <.literal4+0x1fb> + .byte 127,67 // jg 3f87 <.literal4+0x217> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -11423,10 +11438,10 @@ BALIGN4 .byte 190,80,128,3,62 // mov $0x3e038050,%esi .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 3f5b <.literal4+0x21b> + .byte 118,63 // jbe 3fa7 <.literal4+0x237> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 3f6f <.literal4+0x22f> + .byte 127,67 // jg 3fbb <.literal4+0x24b> .byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax) .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax) @@ -11435,7 +11450,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 3f51 <.literal4+0x211> + .byte 224,7 // loopne 3f9d <.literal4+0x22d> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -11447,7 +11462,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 3f6d <.literal4+0x22d> + .byte 224,7 // loopne 3fb9 <.literal4+0x249> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -11458,7 +11473,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 3fc2 <.literal4+0x282> + .byte 124,66 // jl 400e <.literal4+0x29e> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax) @@ -11476,9 +11491,9 @@ BALIGN4 .byte 137,136,136,59,15,0 // mov %ecx,0xf3b88(%rax) .byte 0,0 // add %al,(%rax) .byte 137,136,136,61,0,0 // mov %ecx,0x3d88(%rax) - .byte 112,65 // jo 4005 <.literal4+0x2c5> + .byte 112,65 // jo 4051 <.literal4+0x2e1> .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 4013 <.literal4+0x2d3> + .byte 127,67 // jg 405f <.literal4+0x2ef> .byte 128,0,128 // addb $0x80,(%rax) .byte 55 // (bad) .byte 128,0,128 // addb $0x80,(%rax) @@ -11486,7 +11501,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 255 // (bad) - .byte 127,71 // jg 4027 <.literal4+0x2e7> + .byte 127,71 // jg 4073 <.literal4+0x303> .byte 208 // (bad) .byte 179,89 // mov $0x59,%bl .byte 62,89 // ds pop %rcx @@ -11572,16 +11587,16 @@ BALIGN32 .byte 0,0 // add %al,(%rax) .byte 1,255 // add %edi,%edi .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0040c8 <_sk_callback_hsw+0xa0004ad> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004108 <_sk_callback_hsw+0xa0004bd> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120040d0 <_sk_callback_hsw+0x120004b5> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004110 <_sk_callback_hsw+0x120004c5> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0040d8 <_sk_callback_hsw+0x1a0004bd> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004118 <_sk_callback_hsw+0x1a0004cd> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30040e0 <_sk_callback_hsw+0x30004c5> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004120 <_sk_callback_hsw+0x30004d5> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -11624,16 +11639,16 @@ BALIGN32 .byte 0,0 // add %al,(%rax) .byte 1,255 // add %edi,%edi .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004128 <_sk_callback_hsw+0xa00050d> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004168 <_sk_callback_hsw+0xa00051d> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004130 <_sk_callback_hsw+0x12000515> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004170 <_sk_callback_hsw+0x12000525> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004138 <_sk_callback_hsw+0x1a00051d> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004178 <_sk_callback_hsw+0x1a00052d> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004140 <_sk_callback_hsw+0x3000525> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004180 <_sk_callback_hsw+0x3000535> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -11676,16 +11691,16 @@ BALIGN32 .byte 0,0 // add %al,(%rax) .byte 1,255 // add %edi,%edi .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004188 <_sk_callback_hsw+0xa00056d> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0041c8 <_sk_callback_hsw+0xa00057d> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004190 <_sk_callback_hsw+0x12000575> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120041d0 <_sk_callback_hsw+0x12000585> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004198 <_sk_callback_hsw+0x1a00057d> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0041d8 <_sk_callback_hsw+0x1a00058d> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30041a0 <_sk_callback_hsw+0x3000585> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30041e0 <_sk_callback_hsw+0x3000595> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -11728,16 +11743,16 @@ BALIGN32 .byte 0,0 // add %al,(%rax) .byte 1,255 // add %edi,%edi .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0041e8 <_sk_callback_hsw+0xa0005cd> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004228 <_sk_callback_hsw+0xa0005dd> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120041f0 <_sk_callback_hsw+0x120005d5> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004230 <_sk_callback_hsw+0x120005e5> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0041f8 <_sk_callback_hsw+0x1a0005dd> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004238 <_sk_callback_hsw+0x1a0005ed> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004200 <_sk_callback_hsw+0x30005e5> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004240 <_sk_callback_hsw+0x30005f5> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -11858,14 +11873,14 @@ _sk_seed_shader_avx: .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,107,83,0,0 // vbroadcastss 0x536b(%rip),%ymm1 # 5434 <_sk_callback_avx+0x126> + .byte 196,226,125,24,13,111,83,0,0 // vbroadcastss 0x536f(%rip),%ymm1 # 5438 <_sk_callback_avx+0x126> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,21,79,83,0,0 // vbroadcastss 0x534f(%rip),%ymm2 # 5438 <_sk_callback_avx+0x12a> + .byte 196,226,125,24,21,83,83,0,0 // vbroadcastss 0x5353(%rip),%ymm2 # 543c <_sk_callback_avx+0x12a> .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 @@ -11901,7 +11916,7 @@ HIDDEN _sk_srcatop_avx FUNCTION(_sk_srcatop_avx) _sk_srcatop_avx: .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 - .byte 196,98,125,24,5,255,82,0,0 // vbroadcastss 0x52ff(%rip),%ymm8 # 543c <_sk_callback_avx+0x12e> + .byte 196,98,125,24,5,3,83,0,0 // vbroadcastss 0x5303(%rip),%ymm8 # 5440 <_sk_callback_avx+0x12e> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -11922,7 +11937,7 @@ HIDDEN _sk_dstatop_avx FUNCTION(_sk_dstatop_avx) _sk_dstatop_avx: .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 - .byte 196,98,125,24,13,193,82,0,0 // vbroadcastss 0x52c1(%rip),%ymm9 # 5440 <_sk_callback_avx+0x132> + .byte 196,98,125,24,13,197,82,0,0 // vbroadcastss 0x52c5(%rip),%ymm9 # 5444 <_sk_callback_avx+0x132> .byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 @@ -11964,7 +11979,7 @@ HIDDEN _sk_srcout_avx .globl _sk_srcout_avx FUNCTION(_sk_srcout_avx) _sk_srcout_avx: - .byte 196,98,125,24,5,96,82,0,0 // vbroadcastss 0x5260(%rip),%ymm8 # 5444 <_sk_callback_avx+0x136> + .byte 196,98,125,24,5,100,82,0,0 // vbroadcastss 0x5264(%rip),%ymm8 # 5448 <_sk_callback_avx+0x136> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -11977,7 +11992,7 @@ HIDDEN _sk_dstout_avx .globl _sk_dstout_avx FUNCTION(_sk_dstout_avx) _sk_dstout_avx: - .byte 196,226,125,24,5,67,82,0,0 // vbroadcastss 0x5243(%rip),%ymm0 # 5448 <_sk_callback_avx+0x13a> + .byte 196,226,125,24,5,71,82,0,0 // vbroadcastss 0x5247(%rip),%ymm0 # 544c <_sk_callback_avx+0x13a> .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 @@ -11990,7 +12005,7 @@ HIDDEN _sk_srcover_avx .globl _sk_srcover_avx FUNCTION(_sk_srcover_avx) _sk_srcover_avx: - .byte 196,98,125,24,5,38,82,0,0 // vbroadcastss 0x5226(%rip),%ymm8 # 544c <_sk_callback_avx+0x13e> + .byte 196,98,125,24,5,42,82,0,0 // vbroadcastss 0x522a(%rip),%ymm8 # 5450 <_sk_callback_avx+0x13e> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -12007,7 +12022,7 @@ HIDDEN _sk_dstover_avx .globl _sk_dstover_avx FUNCTION(_sk_dstover_avx) _sk_dstover_avx: - .byte 196,98,125,24,5,249,81,0,0 // vbroadcastss 0x51f9(%rip),%ymm8 # 5450 <_sk_callback_avx+0x142> + .byte 196,98,125,24,5,253,81,0,0 // vbroadcastss 0x51fd(%rip),%ymm8 # 5454 <_sk_callback_avx+0x142> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 @@ -12035,7 +12050,7 @@ HIDDEN _sk_multiply_avx .globl _sk_multiply_avx FUNCTION(_sk_multiply_avx) _sk_multiply_avx: - .byte 196,98,125,24,5,184,81,0,0 // vbroadcastss 0x51b8(%rip),%ymm8 # 5454 <_sk_callback_avx+0x146> + .byte 196,98,125,24,5,188,81,0,0 // vbroadcastss 0x51bc(%rip),%ymm8 # 5458 <_sk_callback_avx+0x146> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -12095,7 +12110,7 @@ HIDDEN _sk_xor__avx .globl _sk_xor__avx FUNCTION(_sk_xor__avx) _sk_xor__avx: - .byte 196,98,125,24,5,7,81,0,0 // vbroadcastss 0x5107(%rip),%ymm8 # 5458 <_sk_callback_avx+0x14a> + .byte 196,98,125,24,5,11,81,0,0 // vbroadcastss 0x510b(%rip),%ymm8 # 545c <_sk_callback_avx+0x14a> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -12132,7 +12147,7 @@ _sk_darken_avx: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,135,80,0,0 // vbroadcastss 0x5087(%rip),%ymm8 # 545c <_sk_callback_avx+0x14e> + .byte 196,98,125,24,5,139,80,0,0 // vbroadcastss 0x508b(%rip),%ymm8 # 5460 <_sk_callback_avx+0x14e> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -12158,7 +12173,7 @@ _sk_lighten_avx: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,51,80,0,0 // vbroadcastss 0x5033(%rip),%ymm8 # 5460 <_sk_callback_avx+0x152> + .byte 196,98,125,24,5,55,80,0,0 // vbroadcastss 0x5037(%rip),%ymm8 # 5464 <_sk_callback_avx+0x152> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -12187,7 +12202,7 @@ _sk_difference_avx: .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,211,79,0,0 // vbroadcastss 0x4fd3(%rip),%ymm8 # 5464 <_sk_callback_avx+0x156> + .byte 196,98,125,24,5,215,79,0,0 // vbroadcastss 0x4fd7(%rip),%ymm8 # 5468 <_sk_callback_avx+0x156> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -12210,7 +12225,7 @@ _sk_exclusion_avx: .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,142,79,0,0 // vbroadcastss 0x4f8e(%rip),%ymm8 # 5468 <_sk_callback_avx+0x15a> + .byte 196,98,125,24,5,146,79,0,0 // vbroadcastss 0x4f92(%rip),%ymm8 # 546c <_sk_callback_avx+0x15a> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -12221,7 +12236,7 @@ HIDDEN _sk_colorburn_avx .globl _sk_colorburn_avx FUNCTION(_sk_colorburn_avx) _sk_colorburn_avx: - .byte 196,98,125,24,5,121,79,0,0 // vbroadcastss 0x4f79(%rip),%ymm8 # 546c <_sk_callback_avx+0x15e> + .byte 196,98,125,24,5,125,79,0,0 // vbroadcastss 0x4f7d(%rip),%ymm8 # 5470 <_sk_callback_avx+0x15e> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 @@ -12283,7 +12298,7 @@ HIDDEN _sk_colordodge_avx FUNCTION(_sk_colordodge_avx) _sk_colordodge_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 196,98,125,24,13,117,78,0,0 // vbroadcastss 0x4e75(%rip),%ymm9 # 5470 <_sk_callback_avx+0x162> + .byte 196,98,125,24,13,121,78,0,0 // vbroadcastss 0x4e79(%rip),%ymm9 # 5474 <_sk_callback_avx+0x162> .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 @@ -12340,7 +12355,7 @@ HIDDEN _sk_hardlight_avx .globl _sk_hardlight_avx FUNCTION(_sk_hardlight_avx) _sk_hardlight_avx: - .byte 196,98,125,24,5,135,77,0,0 // vbroadcastss 0x4d87(%rip),%ymm8 # 5474 <_sk_callback_avx+0x166> + .byte 196,98,125,24,5,139,77,0,0 // vbroadcastss 0x4d8b(%rip),%ymm8 # 5478 <_sk_callback_avx+0x166> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -12395,7 +12410,7 @@ HIDDEN _sk_overlay_avx .globl _sk_overlay_avx FUNCTION(_sk_overlay_avx) _sk_overlay_avx: - .byte 196,98,125,24,5,176,76,0,0 // vbroadcastss 0x4cb0(%rip),%ymm8 # 5478 <_sk_callback_avx+0x16a> + .byte 196,98,125,24,5,180,76,0,0 // vbroadcastss 0x4cb4(%rip),%ymm8 # 547c <_sk_callback_avx+0x16a> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -12461,10 +12476,10 @@ _sk_softlight_avx: .byte 196,65,60,88,192 // vaddps %ymm8,%ymm8,%ymm8 .byte 196,65,60,89,216 // vmulps %ymm8,%ymm8,%ymm11 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 - .byte 196,98,125,24,29,167,75,0,0 // vbroadcastss 0x4ba7(%rip),%ymm11 # 5480 <_sk_callback_avx+0x172> + .byte 196,98,125,24,29,171,75,0,0 // vbroadcastss 0x4bab(%rip),%ymm11 # 5484 <_sk_callback_avx+0x172> .byte 196,65,28,88,235 // vaddps %ymm11,%ymm12,%ymm13 .byte 196,65,20,89,192 // vmulps %ymm8,%ymm13,%ymm8 - .byte 196,98,125,24,45,152,75,0,0 // vbroadcastss 0x4b98(%rip),%ymm13 # 5484 <_sk_callback_avx+0x176> + .byte 196,98,125,24,45,156,75,0,0 // vbroadcastss 0x4b9c(%rip),%ymm13 # 5488 <_sk_callback_avx+0x176> .byte 196,65,28,89,245 // vmulps %ymm13,%ymm12,%ymm14 .byte 196,65,12,88,192 // vaddps %ymm8,%ymm14,%ymm8 .byte 196,65,124,82,244 // vrsqrtps %ymm12,%ymm14 @@ -12475,7 +12490,7 @@ _sk_softlight_avx: .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15 .byte 196,67,13,74,240,240 // vblendvps %ymm15,%ymm8,%ymm14,%ymm14 .byte 197,116,88,249 // vaddps %ymm1,%ymm1,%ymm15 - .byte 196,98,125,24,5,86,75,0,0 // vbroadcastss 0x4b56(%rip),%ymm8 # 547c <_sk_callback_avx+0x16e> + .byte 196,98,125,24,5,90,75,0,0 // vbroadcastss 0x4b5a(%rip),%ymm8 # 5480 <_sk_callback_avx+0x16e> .byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12 .byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0 .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 @@ -12582,7 +12597,7 @@ HIDDEN _sk_clamp_1_avx .globl _sk_clamp_1_avx FUNCTION(_sk_clamp_1_avx) _sk_clamp_1_avx: - .byte 196,98,125,24,5,166,73,0,0 // vbroadcastss 0x49a6(%rip),%ymm8 # 5488 <_sk_callback_avx+0x17a> + .byte 196,98,125,24,5,170,73,0,0 // vbroadcastss 0x49aa(%rip),%ymm8 # 548c <_sk_callback_avx+0x17a> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 @@ -12594,7 +12609,7 @@ HIDDEN _sk_clamp_a_avx .globl _sk_clamp_a_avx FUNCTION(_sk_clamp_a_avx) _sk_clamp_a_avx: - .byte 196,98,125,24,5,137,73,0,0 // vbroadcastss 0x4989(%rip),%ymm8 # 548c <_sk_callback_avx+0x17e> + .byte 196,98,125,24,5,141,73,0,0 // vbroadcastss 0x498d(%rip),%ymm8 # 5490 <_sk_callback_avx+0x17e> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 @@ -12680,7 +12695,7 @@ FUNCTION(_sk_unpremul_avx) _sk_unpremul_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 - .byte 196,98,125,24,21,209,72,0,0 // vbroadcastss 0x48d1(%rip),%ymm10 # 5490 <_sk_callback_avx+0x182> + .byte 196,98,125,24,21,213,72,0,0 // vbroadcastss 0x48d5(%rip),%ymm10 # 5494 <_sk_callback_avx+0x182> .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 @@ -12693,17 +12708,17 @@ HIDDEN _sk_from_srgb_avx .globl _sk_from_srgb_avx FUNCTION(_sk_from_srgb_avx) _sk_from_srgb_avx: - .byte 196,98,125,24,5,178,72,0,0 // vbroadcastss 0x48b2(%rip),%ymm8 # 5494 <_sk_callback_avx+0x186> + .byte 196,98,125,24,5,182,72,0,0 // vbroadcastss 0x48b6(%rip),%ymm8 # 5498 <_sk_callback_avx+0x186> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 - .byte 196,98,125,24,29,164,72,0,0 // vbroadcastss 0x48a4(%rip),%ymm11 # 5498 <_sk_callback_avx+0x18a> + .byte 196,98,125,24,29,168,72,0,0 // vbroadcastss 0x48a8(%rip),%ymm11 # 549c <_sk_callback_avx+0x18a> .byte 196,65,124,89,227 // vmulps %ymm11,%ymm0,%ymm12 - .byte 196,98,125,24,45,154,72,0,0 // vbroadcastss 0x489a(%rip),%ymm13 # 549c <_sk_callback_avx+0x18e> + .byte 196,98,125,24,45,158,72,0,0 // vbroadcastss 0x489e(%rip),%ymm13 # 54a0 <_sk_callback_avx+0x18e> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,139,72,0,0 // vbroadcastss 0x488b(%rip),%ymm12 # 54a0 <_sk_callback_avx+0x192> + .byte 196,98,125,24,37,143,72,0,0 // vbroadcastss 0x488f(%rip),%ymm12 # 54a4 <_sk_callback_avx+0x192> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,53,129,72,0,0 // vbroadcastss 0x4881(%rip),%ymm14 # 54a4 <_sk_callback_avx+0x196> + .byte 196,98,125,24,53,133,72,0,0 // vbroadcastss 0x4885(%rip),%ymm14 # 54a8 <_sk_callback_avx+0x196> .byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0 .byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0 .byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9 @@ -12732,18 +12747,18 @@ _sk_to_srgb_avx: .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 .byte 196,65,124,83,200 // vrcpps %ymm8,%ymm9 .byte 196,65,124,82,208 // vrsqrtps %ymm8,%ymm10 - .byte 196,98,125,24,5,12,72,0,0 // vbroadcastss 0x480c(%rip),%ymm8 # 54a8 <_sk_callback_avx+0x19a> + .byte 196,98,125,24,5,16,72,0,0 // vbroadcastss 0x4810(%rip),%ymm8 # 54ac <_sk_callback_avx+0x19a> .byte 196,65,124,89,216 // vmulps %ymm8,%ymm0,%ymm11 - .byte 196,98,125,24,37,2,72,0,0 // vbroadcastss 0x4802(%rip),%ymm12 # 54ac <_sk_callback_avx+0x19e> + .byte 196,98,125,24,37,6,72,0,0 // vbroadcastss 0x4806(%rip),%ymm12 # 54b0 <_sk_callback_avx+0x19e> .byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9 - .byte 196,98,125,24,45,248,71,0,0 // vbroadcastss 0x47f8(%rip),%ymm13 # 54b0 <_sk_callback_avx+0x1a2> + .byte 196,98,125,24,45,252,71,0,0 // vbroadcastss 0x47fc(%rip),%ymm13 # 54b4 <_sk_callback_avx+0x1a2> .byte 196,65,52,88,205 // vaddps %ymm13,%ymm9,%ymm9 - .byte 196,98,125,24,53,238,71,0,0 // vbroadcastss 0x47ee(%rip),%ymm14 # 54b4 <_sk_callback_avx+0x1a6> + .byte 196,98,125,24,53,242,71,0,0 // vbroadcastss 0x47f2(%rip),%ymm14 # 54b8 <_sk_callback_avx+0x1a6> .byte 196,65,44,89,214 // vmulps %ymm14,%ymm10,%ymm10 .byte 196,65,44,88,201 // vaddps %ymm9,%ymm10,%ymm9 - .byte 196,98,125,24,21,223,71,0,0 // vbroadcastss 0x47df(%rip),%ymm10 # 54b8 <_sk_callback_avx+0x1aa> + .byte 196,98,125,24,21,227,71,0,0 // vbroadcastss 0x47e3(%rip),%ymm10 # 54bc <_sk_callback_avx+0x1aa> .byte 196,65,44,93,201 // vminps %ymm9,%ymm10,%ymm9 - .byte 196,98,125,24,61,213,71,0,0 // vbroadcastss 0x47d5(%rip),%ymm15 # 54bc <_sk_callback_avx+0x1ae> + .byte 196,98,125,24,61,217,71,0,0 // vbroadcastss 0x47d9(%rip),%ymm15 # 54c0 <_sk_callback_avx+0x1ae> .byte 196,193,124,194,199,1 // vcmpltps %ymm15,%ymm0,%ymm0 .byte 196,195,53,74,195,0 // vblendvps %ymm0,%ymm11,%ymm9,%ymm0 .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 @@ -12780,7 +12795,7 @@ _sk_rgb_to_hsl_avx: .byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9 .byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9 .byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,29,59,71,0,0 // vbroadcastss 0x473b(%rip),%ymm11 # 54c0 <_sk_callback_avx+0x1b2> + .byte 196,98,125,24,29,63,71,0,0 // vbroadcastss 0x473f(%rip),%ymm11 # 54c4 <_sk_callback_avx+0x1b2> .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11 .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12 .byte 196,65,28,89,227 // vmulps %ymm11,%ymm12,%ymm12 @@ -12790,19 +12805,19 @@ _sk_rgb_to_hsl_avx: .byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2 .byte 197,252,92,201 // vsubps %ymm1,%ymm0,%ymm1 .byte 196,193,116,89,203 // vmulps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,20,71,0,0 // vbroadcastss 0x4714(%rip),%ymm11 # 54cc <_sk_callback_avx+0x1be> + .byte 196,98,125,24,29,24,71,0,0 // vbroadcastss 0x4718(%rip),%ymm11 # 54d0 <_sk_callback_avx+0x1be> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,2,71,0,0 // vbroadcastss 0x4702(%rip),%ymm11 # 54c8 <_sk_callback_avx+0x1ba> + .byte 196,98,125,24,29,6,71,0,0 // vbroadcastss 0x4706(%rip),%ymm11 # 54cc <_sk_callback_avx+0x1ba> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,227,117,74,202,224 // vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,234,70,0,0 // vbroadcastss 0x46ea(%rip),%ymm2 # 54c4 <_sk_callback_avx+0x1b6> + .byte 196,226,125,24,21,238,70,0,0 // vbroadcastss 0x46ee(%rip),%ymm2 # 54c8 <_sk_callback_avx+0x1b6> .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14 .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2 .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0 .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 .byte 196,227,117,74,194,0 // vblendvps %ymm0,%ymm2,%ymm1,%ymm0 .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1 - .byte 196,98,125,24,37,209,70,0,0 // vbroadcastss 0x46d1(%rip),%ymm12 # 54d4 <_sk_callback_avx+0x1c6> + .byte 196,98,125,24,37,213,70,0,0 // vbroadcastss 0x46d5(%rip),%ymm12 # 54d8 <_sk_callback_avx+0x1c6> .byte 196,193,116,89,212 // vmulps %ymm12,%ymm1,%ymm2 .byte 197,28,194,226,1 // vcmpltps %ymm2,%ymm12,%ymm12 .byte 196,65,36,92,216 // vsubps %ymm8,%ymm11,%ymm11 @@ -12812,7 +12827,7 @@ _sk_rgb_to_hsl_avx: .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 .byte 196,195,125,74,198,128 // vblendvps %ymm8,%ymm14,%ymm0,%ymm0 .byte 196,195,117,74,206,128 // vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - .byte 196,98,125,24,5,148,70,0,0 // vbroadcastss 0x4694(%rip),%ymm8 # 54d0 <_sk_callback_avx+0x1c2> + .byte 196,98,125,24,5,152,70,0,0 // vbroadcastss 0x4698(%rip),%ymm8 # 54d4 <_sk_callback_avx+0x1c2> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -12821,123 +12836,119 @@ HIDDEN _sk_hsl_to_rgb_avx .globl _sk_hsl_to_rgb_avx FUNCTION(_sk_hsl_to_rgb_avx) _sk_hsl_to_rgb_avx: - .byte 72,131,236,56 // sub $0x38,%rsp - .byte 197,252,17,60,36 // vmovups %ymm7,(%rsp) - .byte 197,252,17,116,36,224 // vmovups %ymm6,-0x20(%rsp) - .byte 197,252,17,108,36,192 // vmovups %ymm5,-0x40(%rsp) - .byte 197,252,17,100,36,160 // vmovups %ymm4,-0x60(%rsp) - .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) - .byte 197,252,40,226 // vmovaps %ymm2,%ymm4 - .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 - .byte 197,252,40,216 // vmovaps %ymm0,%ymm3 + .byte 72,131,236,120 // sub $0x78,%rsp + .byte 197,252,17,124,36,64 // vmovups %ymm7,0x40(%rsp) + .byte 197,252,17,116,36,32 // vmovups %ymm6,0x20(%rsp) + .byte 197,252,17,44,36 // vmovups %ymm5,(%rsp) + .byte 197,252,17,100,36,224 // vmovups %ymm4,-0x20(%rsp) + .byte 197,252,17,92,36,192 // vmovups %ymm3,-0x40(%rsp) + .byte 197,252,40,242 // vmovaps %ymm2,%ymm6 + .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp) + .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 184,0,0,0,63 // mov $0x3f000000,%eax .byte 197,249,110,192 // vmovd %eax,%xmm0 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 .byte 196,99,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm8 - .byte 196,193,92,194,192,1 // vcmpltps %ymm8,%ymm4,%ymm0 + .byte 196,193,76,194,192,1 // vcmpltps %ymm8,%ymm6,%ymm0 + .byte 196,98,125,24,21,68,70,0,0 // vbroadcastss 0x4644(%rip),%ymm10 # 54dc <_sk_callback_avx+0x1ca> + .byte 196,193,116,88,218 // vaddps %ymm10,%ymm1,%ymm3 + .byte 197,228,89,222 // vmulps %ymm6,%ymm3,%ymm3 + .byte 197,244,88,230 // vaddps %ymm6,%ymm1,%ymm4 + .byte 197,244,89,238 // vmulps %ymm6,%ymm1,%ymm5 + .byte 197,220,92,229 // vsubps %ymm5,%ymm4,%ymm4 + .byte 196,99,93,74,203,0 // vblendvps %ymm0,%ymm3,%ymm4,%ymm9 + .byte 196,226,125,24,5,36,70,0,0 // vbroadcastss 0x4624(%rip),%ymm0 # 54e0 <_sk_callback_avx+0x1ce> + .byte 197,236,88,200 // vaddps %ymm0,%ymm2,%ymm1 + .byte 65,184,0,0,0,0 // mov $0x0,%r8d .byte 184,0,0,128,63 // mov $0x3f800000,%eax - .byte 197,249,110,208 // vmovd %eax,%xmm2 - .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,99,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm10 - .byte 197,172,88,213 // vaddps %ymm5,%ymm10,%ymm2 - .byte 197,236,89,212 // vmulps %ymm4,%ymm2,%ymm2 - .byte 197,84,88,204 // vaddps %ymm4,%ymm5,%ymm9 - .byte 197,84,89,220 // vmulps %ymm4,%ymm5,%ymm11 - .byte 196,65,52,92,203 // vsubps %ymm11,%ymm9,%ymm9 - .byte 196,99,53,74,202,0 // vblendvps %ymm0,%ymm2,%ymm9,%ymm9 - .byte 65,184,0,0,0,64 // mov $0x40000000,%r8d - .byte 184,171,170,170,62 // mov $0x3eaaaaab,%eax - .byte 197,249,110,192 // vmovd %eax,%xmm0 - .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 - .byte 197,28,88,251 // vaddps %ymm3,%ymm12,%ymm15 - .byte 184,0,0,0,0 // mov $0x0,%eax - .byte 197,249,110,192 // vmovd %eax,%xmm0 - .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,99,125,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 - .byte 196,193,44,194,199,1 // vcmpltps %ymm15,%ymm10,%ymm0 - .byte 196,193,4,92,210 // vsubps %ymm10,%ymm15,%ymm2 - .byte 196,227,5,74,194,0 // vblendvps %ymm0,%ymm2,%ymm15,%ymm0 - .byte 196,193,4,194,213,1 // vcmpltps %ymm13,%ymm15,%ymm2 - .byte 196,65,44,88,223 // vaddps %ymm15,%ymm10,%ymm11 - .byte 196,195,125,74,203,32 // vblendvps %ymm2,%ymm11,%ymm0,%ymm1 - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 - .byte 196,65,124,92,217 // vsubps %ymm9,%ymm0,%ymm11 + .byte 197,249,110,216 // vmovd %eax,%xmm3 + .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 + .byte 196,99,101,24,227,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm12 + .byte 197,156,194,217,1 // vcmpltps %ymm1,%ymm12,%ymm3 + .byte 196,98,125,24,53,251,69,0,0 // vbroadcastss 0x45fb(%rip),%ymm14 # 54e4 <_sk_callback_avx+0x1d2> + .byte 196,193,116,88,230 // vaddps %ymm14,%ymm1,%ymm4 + .byte 196,227,117,74,220,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm3 + .byte 196,193,121,110,224 // vmovd %r8d,%xmm4 + .byte 196,227,121,4,228,0 // vpermilps $0x0,%xmm4,%xmm4 + .byte 196,99,93,24,252,1 // vinsertf128 $0x1,%xmm4,%ymm4,%ymm15 + .byte 196,193,116,194,231,1 // vcmpltps %ymm15,%ymm1,%ymm4 + .byte 196,193,116,88,234 // vaddps %ymm10,%ymm1,%ymm5 + .byte 196,227,101,74,197,64 // vblendvps %ymm4,%ymm5,%ymm3,%ymm0 + .byte 197,204,88,222 // vaddps %ymm6,%ymm6,%ymm3 + .byte 196,65,100,92,217 // vsubps %ymm9,%ymm3,%ymm11 + .byte 196,193,52,92,219 // vsubps %ymm11,%ymm9,%ymm3 + .byte 196,226,125,24,37,187,69,0,0 // vbroadcastss 0x45bb(%rip),%ymm4 # 54e8 <_sk_callback_avx+0x1d6> + .byte 197,100,89,236 // vmulps %ymm4,%ymm3,%ymm13 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d - .byte 184,0,0,192,64 // mov $0x40c00000,%eax - .byte 197,249,110,192 // vmovd %eax,%xmm0 - .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 196,193,52,92,211 // vsubps %ymm11,%ymm9,%ymm2 - .byte 197,108,89,240 // vmulps %ymm0,%ymm2,%ymm14 .byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax - .byte 197,249,110,192 // vmovd %eax,%xmm0 - .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2 - .byte 197,140,89,210 // vmulps %ymm2,%ymm14,%ymm2 - .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 - .byte 197,244,194,248,1 // vcmpltps %ymm0,%ymm1,%ymm7 - .byte 196,227,37,74,210,112 // vblendvps %ymm7,%ymm2,%ymm11,%ymm2 - .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 - .byte 196,195,109,74,249,112 // vblendvps %ymm7,%ymm9,%ymm2,%ymm7 - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1 - .byte 196,65,4,89,254 // vmulps %ymm14,%ymm15,%ymm15 - .byte 196,65,36,88,255 // vaddps %ymm15,%ymm11,%ymm15 - .byte 196,67,69,74,255,16 // vblendvps %ymm1,%ymm15,%ymm7,%ymm15 - .byte 197,172,194,203,1 // vcmpltps %ymm3,%ymm10,%ymm1 - .byte 196,193,100,92,250 // vsubps %ymm10,%ymm3,%ymm7 - .byte 196,227,101,74,207,16 // vblendvps %ymm1,%ymm7,%ymm3,%ymm1 - .byte 196,193,100,194,253,1 // vcmpltps %ymm13,%ymm3,%ymm7 - .byte 197,172,88,243 // vaddps %ymm3,%ymm10,%ymm6 - .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1 - .byte 197,252,92,241 // vsubps %ymm1,%ymm0,%ymm6 - .byte 197,140,89,246 // vmulps %ymm6,%ymm14,%ymm6 - .byte 197,164,88,246 // vaddps %ymm6,%ymm11,%ymm6 - .byte 197,244,194,248,1 // vcmpltps %ymm0,%ymm1,%ymm7 - .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6 - .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 - .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6 - .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1 - .byte 197,140,89,251 // vmulps %ymm3,%ymm14,%ymm7 + .byte 197,249,110,216 // vmovd %eax,%xmm3 + .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 + .byte 196,227,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm5 + .byte 196,226,125,24,37,151,69,0,0 // vbroadcastss 0x4597(%rip),%ymm4 # 54ec <_sk_callback_avx+0x1da> + .byte 197,220,92,216 // vsubps %ymm0,%ymm4,%ymm3 + .byte 197,148,89,219 // vmulps %ymm3,%ymm13,%ymm3 + .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 + .byte 197,252,194,253,1 // vcmpltps %ymm5,%ymm0,%ymm7 + .byte 196,227,37,74,219,112 // vblendvps %ymm7,%ymm3,%ymm11,%ymm3 + .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7 + .byte 196,195,101,74,249,112 // vblendvps %ymm7,%ymm9,%ymm3,%ymm7 + .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 + .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 + .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 + .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0 + .byte 196,193,116,89,205 // vmulps %ymm13,%ymm1,%ymm1 + .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 + .byte 196,227,69,74,193,0 // vblendvps %ymm0,%ymm1,%ymm7,%ymm0 + .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp) + .byte 197,156,194,202,1 // vcmpltps %ymm2,%ymm12,%ymm1 + .byte 196,193,108,88,254 // vaddps %ymm14,%ymm2,%ymm7 + .byte 196,227,109,74,207,16 // vblendvps %ymm1,%ymm7,%ymm2,%ymm1 + .byte 196,193,108,194,255,1 // vcmpltps %ymm15,%ymm2,%ymm7 + .byte 196,193,108,88,194 // vaddps %ymm10,%ymm2,%ymm0 + .byte 196,227,117,74,192,112 // vblendvps %ymm7,%ymm0,%ymm1,%ymm0 + .byte 197,220,92,200 // vsubps %ymm0,%ymm4,%ymm1 + .byte 197,148,89,201 // vmulps %ymm1,%ymm13,%ymm1 + .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 + .byte 197,252,194,253,1 // vcmpltps %ymm5,%ymm0,%ymm7 + .byte 196,227,37,74,201,112 // vblendvps %ymm7,%ymm1,%ymm11,%ymm1 + .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7 + .byte 196,195,117,74,201,112 // vblendvps %ymm7,%ymm9,%ymm1,%ymm1 + .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0 + .byte 197,148,89,250 // vmulps %ymm2,%ymm13,%ymm7 .byte 197,164,88,255 // vaddps %ymm7,%ymm11,%ymm7 - .byte 196,227,77,74,207,16 // vblendvps %ymm1,%ymm7,%ymm6,%ymm1 - .byte 196,193,100,92,220 // vsubps %ymm12,%ymm3,%ymm3 - .byte 197,172,194,243,1 // vcmpltps %ymm3,%ymm10,%ymm6 - .byte 196,193,100,92,250 // vsubps %ymm10,%ymm3,%ymm7 - .byte 196,227,101,74,247,96 // vblendvps %ymm6,%ymm7,%ymm3,%ymm6 - .byte 196,193,100,194,253,1 // vcmpltps %ymm13,%ymm3,%ymm7 - .byte 197,44,88,211 // vaddps %ymm3,%ymm10,%ymm10 - .byte 196,195,77,74,242,112 // vblendvps %ymm7,%ymm10,%ymm6,%ymm6 - .byte 197,204,194,248,1 // vcmpltps %ymm0,%ymm6,%ymm7 - .byte 197,252,92,198 // vsubps %ymm6,%ymm0,%ymm0 - .byte 197,140,89,192 // vmulps %ymm0,%ymm14,%ymm0 + .byte 196,227,117,74,207,0 // vblendvps %ymm0,%ymm7,%ymm1,%ymm1 + .byte 196,226,125,24,5,237,68,0,0 // vbroadcastss 0x44ed(%rip),%ymm0 # 54f0 <_sk_callback_avx+0x1de> + .byte 197,236,88,192 // vaddps %ymm0,%ymm2,%ymm0 + .byte 197,156,194,208,1 // vcmpltps %ymm0,%ymm12,%ymm2 + .byte 196,193,124,88,254 // vaddps %ymm14,%ymm0,%ymm7 + .byte 196,227,125,74,215,32 // vblendvps %ymm2,%ymm7,%ymm0,%ymm2 + .byte 196,193,124,194,255,1 // vcmpltps %ymm15,%ymm0,%ymm7 + .byte 196,65,124,88,210 // vaddps %ymm10,%ymm0,%ymm10 + .byte 196,195,109,74,210,112 // vblendvps %ymm7,%ymm10,%ymm2,%ymm2 + .byte 197,236,194,237,1 // vcmpltps %ymm5,%ymm2,%ymm5 + .byte 197,220,92,226 // vsubps %ymm2,%ymm4,%ymm4 + .byte 197,148,89,228 // vmulps %ymm4,%ymm13,%ymm4 + .byte 197,164,88,228 // vaddps %ymm4,%ymm11,%ymm4 + .byte 196,227,37,74,228,80 // vblendvps %ymm5,%ymm4,%ymm11,%ymm4 + .byte 196,193,108,194,232,1 // vcmpltps %ymm8,%ymm2,%ymm5 + .byte 196,195,93,74,225,80 // vblendvps %ymm5,%ymm9,%ymm4,%ymm4 + .byte 197,236,194,211,1 // vcmpltps %ymm3,%ymm2,%ymm2 + .byte 196,193,124,89,197 // vmulps %ymm13,%ymm0,%ymm0 .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 - .byte 196,227,37,74,192,112 // vblendvps %ymm7,%ymm0,%ymm11,%ymm0 - .byte 196,193,76,194,248,1 // vcmpltps %ymm8,%ymm6,%ymm7 - .byte 196,195,125,74,193,112 // vblendvps %ymm7,%ymm9,%ymm0,%ymm0 - .byte 197,204,194,210,1 // vcmpltps %ymm2,%ymm6,%ymm2 - .byte 196,193,100,89,222 // vmulps %ymm14,%ymm3,%ymm3 - .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 - .byte 196,227,125,74,211,32 // vblendvps %ymm2,%ymm3,%ymm0,%ymm2 + .byte 196,227,93,74,208,32 // vblendvps %ymm2,%ymm0,%ymm4,%ymm2 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 - .byte 197,212,194,216,0 // vcmpeqps %ymm0,%ymm5,%ymm3 - .byte 196,227,5,74,196,48 // vblendvps %ymm3,%ymm4,%ymm15,%ymm0 - .byte 196,227,117,74,204,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm1 - .byte 196,227,109,74,212,48 // vblendvps %ymm3,%ymm4,%ymm2,%ymm2 + .byte 197,252,194,92,36,160,0 // vcmpeqps -0x60(%rsp),%ymm0,%ymm3 + .byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0 + .byte 196,227,125,74,198,48 // vblendvps %ymm3,%ymm6,%ymm0,%ymm0 + .byte 196,227,117,74,206,48 // vblendvps %ymm3,%ymm6,%ymm1,%ymm1 + .byte 196,227,109,74,214,48 // vblendvps %ymm3,%ymm6,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 197,252,16,92,36,128 // vmovups -0x80(%rsp),%ymm3 - .byte 197,252,16,100,36,160 // vmovups -0x60(%rsp),%ymm4 - .byte 197,252,16,108,36,192 // vmovups -0x40(%rsp),%ymm5 - .byte 197,252,16,116,36,224 // vmovups -0x20(%rsp),%ymm6 - .byte 197,252,16,60,36 // vmovups (%rsp),%ymm7 - .byte 72,131,196,56 // add $0x38,%rsp + .byte 197,252,16,92,36,192 // vmovups -0x40(%rsp),%ymm3 + .byte 197,252,16,100,36,224 // vmovups -0x20(%rsp),%ymm4 + .byte 197,252,16,44,36 // vmovups (%rsp),%ymm5 + .byte 197,252,16,116,36,32 // vmovups 0x20(%rsp),%ymm6 + .byte 197,252,16,124,36,64 // vmovups 0x40(%rsp),%ymm7 + .byte 72,131,196,120 // add $0x78,%rsp .byte 255,224 // jmpq *%rax HIDDEN _sk_scale_1_float_avx @@ -12962,14 +12973,14 @@ _sk_scale_u8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,68 // jne 1110 <_sk_scale_u8_avx+0x54> + .byte 117,68 // jne 1116 <_sk_scale_u8_avx+0x54> .byte 197,122,126,0 // vmovq (%rax),%xmm8 .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,228,67,0,0 // vbroadcastss 0x43e4(%rip),%ymm9 # 54d8 <_sk_callback_avx+0x1ca> + .byte 196,98,125,24,13,250,67,0,0 // vbroadcastss 0x43fa(%rip),%ymm9 # 54f4 <_sk_callback_avx+0x1e2> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -12987,9 +12998,9 @@ _sk_scale_u8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 1118 <_sk_scale_u8_avx+0x5c> + .byte 117,234 // jne 111e <_sk_scale_u8_avx+0x5c> .byte 196,65,249,110,193 // vmovq %r9,%xmm8 - .byte 235,155 // jmp 10d0 <_sk_scale_u8_avx+0x14> + .byte 235,155 // jmp 10d6 <_sk_scale_u8_avx+0x14> HIDDEN _sk_lerp_1_float_avx .globl _sk_lerp_1_float_avx @@ -13021,14 +13032,14 @@ _sk_lerp_u8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,104 // jne 11ec <_sk_lerp_u8_avx+0x78> + .byte 117,104 // jne 11f2 <_sk_lerp_u8_avx+0x78> .byte 197,122,126,0 // vmovq (%rax),%xmm8 .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,48,67,0,0 // vbroadcastss 0x4330(%rip),%ymm9 # 54dc <_sk_callback_avx+0x1ce> + .byte 196,98,125,24,13,70,67,0,0 // vbroadcastss 0x4346(%rip),%ymm9 # 54f8 <_sk_callback_avx+0x1e6> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 @@ -13054,9 +13065,9 @@ _sk_lerp_u8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 11f4 <_sk_lerp_u8_avx+0x80> + .byte 117,234 // jne 11fa <_sk_lerp_u8_avx+0x80> .byte 196,65,249,110,193 // vmovq %r9,%xmm8 - .byte 233,116,255,255,255 // jmpq 1188 <_sk_lerp_u8_avx+0x14> + .byte 233,116,255,255,255 // jmpq 118e <_sk_lerp_u8_avx+0x14> HIDDEN _sk_lerp_565_avx .globl _sk_lerp_565_avx @@ -13065,26 +13076,26 @@ _sk_lerp_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,174,0,0,0 // jne 12d0 <_sk_lerp_565_avx+0xbc> + .byte 15,133,174,0,0,0 // jne 12d6 <_sk_lerp_565_avx+0xbc> .byte 196,65,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm8 .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 .byte 197,185,105,219 // vpunpckhwd %xmm3,%xmm8,%xmm3 .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 - .byte 196,98,125,24,5,156,66,0,0 // vbroadcastss 0x429c(%rip),%ymm8 # 54e0 <_sk_callback_avx+0x1d2> + .byte 196,98,125,24,5,178,66,0,0 // vbroadcastss 0x42b2(%rip),%ymm8 # 54fc <_sk_callback_avx+0x1ea> .byte 196,65,100,84,192 // vandps %ymm8,%ymm3,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,141,66,0,0 // vbroadcastss 0x428d(%rip),%ymm9 # 54e4 <_sk_callback_avx+0x1d6> + .byte 196,98,125,24,13,163,66,0,0 // vbroadcastss 0x42a3(%rip),%ymm9 # 5500 <_sk_callback_avx+0x1ee> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,131,66,0,0 // vbroadcastss 0x4283(%rip),%ymm9 # 54e8 <_sk_callback_avx+0x1da> + .byte 196,98,125,24,13,153,66,0,0 // vbroadcastss 0x4299(%rip),%ymm9 # 5504 <_sk_callback_avx+0x1f2> .byte 196,65,100,84,201 // vandps %ymm9,%ymm3,%ymm9 .byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9 - .byte 196,98,125,24,21,116,66,0,0 // vbroadcastss 0x4274(%rip),%ymm10 # 54ec <_sk_callback_avx+0x1de> + .byte 196,98,125,24,21,138,66,0,0 // vbroadcastss 0x428a(%rip),%ymm10 # 5508 <_sk_callback_avx+0x1f6> .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,21,106,66,0,0 // vbroadcastss 0x426a(%rip),%ymm10 # 54f0 <_sk_callback_avx+0x1e2> + .byte 196,98,125,24,21,128,66,0,0 // vbroadcastss 0x4280(%rip),%ymm10 # 550c <_sk_callback_avx+0x1fa> .byte 196,193,100,84,218 // vandps %ymm10,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,21,92,66,0,0 // vbroadcastss 0x425c(%rip),%ymm10 # 54f4 <_sk_callback_avx+0x1e6> + .byte 196,98,125,24,21,114,66,0,0 // vbroadcastss 0x4272(%rip),%ymm10 # 5510 <_sk_callback_avx+0x1fe> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 @@ -13096,16 +13107,16 @@ _sk_lerp_565_avx: .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,42,66,0,0 // vbroadcastss 0x422a(%rip),%ymm3 # 54f8 <_sk_callback_avx+0x1ea> + .byte 196,226,125,24,29,64,66,0,0 // vbroadcastss 0x4240(%rip),%ymm3 # 5514 <_sk_callback_avx+0x202> .byte 255,224 // jmpq *%rax .byte 65,137,200 // mov %ecx,%r8d .byte 65,128,224,7 // and $0x7,%r8b .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,63,255,255,255 // ja 1228 <_sk_lerp_565_avx+0x14> + .byte 15,135,63,255,255,255 // ja 122e <_sk_lerp_565_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 1340 <_sk_lerp_565_avx+0x12c> + .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 1344 <_sk_lerp_565_avx+0x12a> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -13117,26 +13128,27 @@ _sk_lerp_565_avx: .byte 196,65,57,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 .byte 196,65,57,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 .byte 196,65,57,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 - .byte 233,235,254,255,255 // jmpq 1228 <_sk_lerp_565_avx+0x14> - .byte 15,31,0 // nopl (%rax) - .byte 241 // icebp + .byte 233,235,254,255,255 // jmpq 122e <_sk_lerp_565_avx+0x14> + .byte 144 // nop + .byte 243,255 // repz (bad) .byte 255 // (bad) .byte 255 // (bad) + .byte 235,255 // jmp 1349 <_sk_lerp_565_avx+0x12f> .byte 255 // (bad) - .byte 233,255,255,255,225 // jmpq ffffffffe2001348 <_sk_callback_avx+0xffffffffe1ffc03a> + .byte 255,227 // jmpq *%rbx .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 217,255 // fcos + .byte 219,255 // (bad) .byte 255 // (bad) - .byte 255,209 // callq *%rcx + .byte 255,211 // callq *%rbx .byte 255 // (bad) .byte 255 // (bad) - .byte 255,201 // dec %ecx + .byte 255,203 // dec %ebx .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 189 // .byte 0xbd + .byte 191 // .byte 0xbf .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -13148,7 +13160,7 @@ _sk_load_tables_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,0 // mov (%rax),%r8 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,26,2,0,0 // jne 1584 <_sk_load_tables_avx+0x228> + .byte 15,133,26,2,0,0 // jne 1588 <_sk_load_tables_avx+0x228> .byte 196,65,124,16,4,184 // vmovups (%r8,%rdi,4),%ymm8 .byte 85 // push %rbp .byte 65,87 // push %r15 @@ -13156,7 +13168,7 @@ _sk_load_tables_avx: .byte 65,85 // push %r13 .byte 65,84 // push %r12 .byte 83 // push %rbx - .byte 197,124,40,13,62,68,0,0 // vmovaps 0x443e(%rip),%ymm9 # 57c0 <_sk_callback_avx+0x4b2> + .byte 197,124,40,13,90,68,0,0 // vmovaps 0x445a(%rip),%ymm9 # 57e0 <_sk_callback_avx+0x4ce> .byte 196,193,60,84,193 // vandps %ymm9,%ymm8,%ymm0 .byte 196,193,249,126,193 // vmovq %xmm0,%r9 .byte 69,137,203 // mov %r9d,%r11d @@ -13248,7 +13260,7 @@ _sk_load_tables_avx: .byte 196,193,97,114,210,24 // vpsrld $0x18,%xmm10,%xmm3 .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,139,63,0,0 // vbroadcastss 0x3f8b(%rip),%ymm8 # 54fc <_sk_callback_avx+0x1ee> + .byte 196,98,125,24,5,163,63,0,0 // vbroadcastss 0x3fa3(%rip),%ymm8 # 5518 <_sk_callback_avx+0x206> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx @@ -13263,9 +13275,9 @@ _sk_load_tables_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 15,135,211,253,255,255 // ja 1370 <_sk_load_tables_avx+0x14> + .byte 15,135,211,253,255,255 // ja 1374 <_sk_load_tables_avx+0x14> .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,140,0,0,0 // lea 0x8c(%rip),%r10 # 1634 <_sk_load_tables_avx+0x2d8> + .byte 76,141,21,140,0,0,0 // lea 0x8c(%rip),%r10 # 1638 <_sk_load_tables_avx+0x2d8> .byte 79,99,12,138 // movslq (%r10,%r9,4),%r9 .byte 77,1,209 // add %r10,%r9 .byte 65,255,225 // jmpq *%r9 @@ -13288,7 +13300,7 @@ _sk_load_tables_avx: .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 .byte 196,195,57,34,4,184,0 // vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0 .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 - .byte 233,62,253,255,255 // jmpq 1370 <_sk_load_tables_avx+0x14> + .byte 233,62,253,255,255 // jmpq 1374 <_sk_load_tables_avx+0x14> .byte 102,144 // xchg %ax,%ax .byte 236 // in (%dx),%al .byte 255 // (bad) @@ -13306,7 +13318,7 @@ _sk_load_tables_avx: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 126,255 // jle 164d <_sk_load_tables_avx+0x2f1> + .byte 126,255 // jle 1651 <_sk_load_tables_avx+0x2f1> .byte 255 // (bad) .byte 255 // .byte 0xff @@ -13318,7 +13330,7 @@ _sk_load_tables_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,113,2,0,0 // jne 18d7 <_sk_load_tables_u16_be_avx+0x287> + .byte 15,133,113,2,0,0 // jne 18db <_sk_load_tables_u16_be_avx+0x287> .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8 .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2 .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3 @@ -13340,7 +13352,7 @@ _sk_load_tables_u16_be_avx: .byte 197,177,108,208 // vpunpcklqdq %xmm0,%xmm9,%xmm2 .byte 197,177,109,200 // vpunpckhqdq %xmm0,%xmm9,%xmm1 .byte 196,65,57,108,212 // vpunpcklqdq %xmm12,%xmm8,%xmm10 - .byte 197,121,111,29,126,65,0,0 // vmovdqa 0x417e(%rip),%xmm11 # 5840 <_sk_callback_avx+0x532> + .byte 197,121,111,29,154,65,0,0 // vmovdqa 0x419a(%rip),%xmm11 # 5860 <_sk_callback_avx+0x54e> .byte 196,193,105,219,195 // vpand %xmm11,%xmm2,%xmm0 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 196,193,121,105,209 // vpunpckhwd %xmm9,%xmm0,%xmm2 @@ -13439,7 +13451,7 @@ _sk_load_tables_u16_be_avx: .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3 .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,60,60,0,0 // vbroadcastss 0x3c3c(%rip),%ymm8 # 5500 <_sk_callback_avx+0x1f2> + .byte 196,98,125,24,5,84,60,0,0 // vbroadcastss 0x3c54(%rip),%ymm8 # 551c <_sk_callback_avx+0x20a> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx @@ -13452,29 +13464,29 @@ _sk_load_tables_u16_be_avx: .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 193d <_sk_load_tables_u16_be_avx+0x2ed> + .byte 116,85 // je 1941 <_sk_load_tables_u16_be_avx+0x2ed> .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 193d <_sk_load_tables_u16_be_avx+0x2ed> + .byte 114,72 // jb 1941 <_sk_load_tables_u16_be_avx+0x2ed> .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 194a <_sk_load_tables_u16_be_avx+0x2fa> + .byte 116,72 // je 194e <_sk_load_tables_u16_be_avx+0x2fa> .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 194a <_sk_load_tables_u16_be_avx+0x2fa> + .byte 114,59 // jb 194e <_sk_load_tables_u16_be_avx+0x2fa> .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,97,253,255,255 // je 1681 <_sk_load_tables_u16_be_avx+0x31> + .byte 15,132,97,253,255,255 // je 1685 <_sk_load_tables_u16_be_avx+0x31> .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,80,253,255,255 // jb 1681 <_sk_load_tables_u16_be_avx+0x31> + .byte 15,130,80,253,255,255 // jb 1685 <_sk_load_tables_u16_be_avx+0x31> .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9 - .byte 233,68,253,255,255 // jmpq 1681 <_sk_load_tables_u16_be_avx+0x31> + .byte 233,68,253,255,255 // jmpq 1685 <_sk_load_tables_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,55,253,255,255 // jmpq 1681 <_sk_load_tables_u16_be_avx+0x31> + .byte 233,55,253,255,255 // jmpq 1685 <_sk_load_tables_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,46,253,255,255 // jmpq 1681 <_sk_load_tables_u16_be_avx+0x31> + .byte 233,46,253,255,255 // jmpq 1685 <_sk_load_tables_u16_be_avx+0x31> HIDDEN _sk_load_tables_rgb_u16_be_avx .globl _sk_load_tables_rgb_u16_be_avx @@ -13484,7 +13496,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,93,2,0,0 // jne 1bc2 <_sk_load_tables_rgb_u16_be_avx+0x26f> + .byte 15,133,93,2,0,0 // jne 1bc6 <_sk_load_tables_rgb_u16_be_avx+0x26f> .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0 .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2 .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1 @@ -13511,7 +13523,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 197,185,108,202 // vpunpcklqdq %xmm2,%xmm8,%xmm1 .byte 197,185,109,210 // vpunpckhqdq %xmm2,%xmm8,%xmm2 .byte 197,121,108,195 // vpunpcklqdq %xmm3,%xmm0,%xmm8 - .byte 197,121,111,13,119,62,0,0 // vmovdqa 0x3e77(%rip),%xmm9 # 5850 <_sk_callback_avx+0x542> + .byte 197,121,111,13,147,62,0,0 // vmovdqa 0x3e93(%rip),%xmm9 # 5870 <_sk_callback_avx+0x55e> .byte 196,193,113,219,193 // vpand %xmm9,%xmm1,%xmm0 .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10 .byte 196,193,121,105,202 // vpunpckhwd %xmm10,%xmm0,%xmm1 @@ -13603,7 +13615,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 .byte 196,195,109,24,208,1 // vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,78,57,0,0 // vbroadcastss 0x394e(%rip),%ymm3 # 5504 <_sk_callback_avx+0x1f6> + .byte 196,226,125,24,29,102,57,0,0 // vbroadcastss 0x3966(%rip),%ymm3 # 5520 <_sk_callback_avx+0x20e> .byte 91 // pop %rbx .byte 65,92 // pop %r12 .byte 65,93 // pop %r13 @@ -13614,36 +13626,36 @@ _sk_load_tables_rgb_u16_be_avx: .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0 .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 1bdb <_sk_load_tables_rgb_u16_be_avx+0x288> - .byte 233,190,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 117,5 // jne 1bdf <_sk_load_tables_rgb_u16_be_avx+0x288> + .byte 233,190,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1 .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 1c0a <_sk_load_tables_rgb_u16_be_avx+0x2b7> + .byte 114,26 // jb 1c0e <_sk_load_tables_rgb_u16_be_avx+0x2b7> .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1 .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 1c0f <_sk_load_tables_rgb_u16_be_avx+0x2bc> - .byte 233,143,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> - .byte 233,138,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 1c13 <_sk_load_tables_rgb_u16_be_avx+0x2bc> + .byte 233,143,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,138,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1 .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 1c3e <_sk_load_tables_rgb_u16_be_avx+0x2eb> + .byte 114,26 // jb 1c42 <_sk_load_tables_rgb_u16_be_avx+0x2eb> .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1 .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 1c43 <_sk_load_tables_rgb_u16_be_avx+0x2f0> - .byte 233,91,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> - .byte 233,86,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 1c47 <_sk_load_tables_rgb_u16_be_avx+0x2f0> + .byte 233,91,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,86,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3 .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 1c6c <_sk_load_tables_rgb_u16_be_avx+0x319> + .byte 114,20 // jb 1c70 <_sk_load_tables_rgb_u16_be_avx+0x319> .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3 .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 - .byte 233,45,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> - .byte 233,40,253,255,255 // jmpq 1999 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,45,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,40,253,255,255 // jmpq 199d <_sk_load_tables_rgb_u16_be_avx+0x46> HIDDEN _sk_byte_tables_avx .globl _sk_byte_tables_avx @@ -13656,7 +13668,7 @@ _sk_byte_tables_avx: .byte 65,84 // push %r12 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,130,56,0,0 // vbroadcastss 0x3882(%rip),%ymm8 # 5508 <_sk_callback_avx+0x1fa> + .byte 196,98,125,24,5,154,56,0,0 // vbroadcastss 0x389a(%rip),%ymm8 # 5524 <_sk_callback_avx+0x212> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 @@ -13693,7 +13705,7 @@ _sk_byte_tables_avx: .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,208,55,0,0 // vbroadcastss 0x37d0(%rip),%ymm9 # 550c <_sk_callback_avx+0x1fe> + .byte 196,98,125,24,13,232,55,0,0 // vbroadcastss 0x37e8(%rip),%ymm9 # 5528 <_sk_callback_avx+0x216> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 @@ -13855,7 +13867,7 @@ _sk_byte_tables_rgb_avx: .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,246,52,0,0 // vbroadcastss 0x34f6(%rip),%ymm9 # 5510 <_sk_callback_avx+0x202> + .byte 196,98,125,24,13,14,53,0,0 // vbroadcastss 0x350e(%rip),%ymm9 # 552c <_sk_callback_avx+0x21a> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 @@ -14152,36 +14164,36 @@ _sk_parametric_r_avx: .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,216 // vcvtdq2ps %ymm0,%ymm11 - .byte 196,98,125,24,37,84,48,0,0 // vbroadcastss 0x3054(%rip),%ymm12 # 5514 <_sk_callback_avx+0x206> + .byte 196,98,125,24,37,108,48,0,0 // vbroadcastss 0x306c(%rip),%ymm12 # 5530 <_sk_callback_avx+0x21e> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,74,48,0,0 // vbroadcastss 0x304a(%rip),%ymm12 # 5518 <_sk_callback_avx+0x20a> + .byte 196,98,125,24,37,98,48,0,0 // vbroadcastss 0x3062(%rip),%ymm12 # 5534 <_sk_callback_avx+0x222> .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,64,48,0,0 // vbroadcastss 0x3040(%rip),%ymm12 # 551c <_sk_callback_avx+0x20e> + .byte 196,98,125,24,37,88,48,0,0 // vbroadcastss 0x3058(%rip),%ymm12 # 5538 <_sk_callback_avx+0x226> .byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,54,48,0,0 // vbroadcastss 0x3036(%rip),%ymm12 # 5520 <_sk_callback_avx+0x212> + .byte 196,98,125,24,37,78,48,0,0 // vbroadcastss 0x304e(%rip),%ymm12 # 553c <_sk_callback_avx+0x22a> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,44,48,0,0 // vbroadcastss 0x302c(%rip),%ymm12 # 5524 <_sk_callback_avx+0x216> + .byte 196,98,125,24,37,68,48,0,0 // vbroadcastss 0x3044(%rip),%ymm12 # 5540 <_sk_callback_avx+0x22e> .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,29,48,0,0 // vbroadcastss 0x301d(%rip),%ymm12 # 5528 <_sk_callback_avx+0x21a> + .byte 196,98,125,24,37,53,48,0,0 // vbroadcastss 0x3035(%rip),%ymm12 # 5544 <_sk_callback_avx+0x232> .byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,19,48,0,0 // vbroadcastss 0x3013(%rip),%ymm12 # 552c <_sk_callback_avx+0x21e> + .byte 196,98,125,24,37,43,48,0,0 // vbroadcastss 0x302b(%rip),%ymm12 # 5548 <_sk_callback_avx+0x236> .byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0 .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10 - .byte 196,98,125,24,29,247,47,0,0 // vbroadcastss 0x2ff7(%rip),%ymm11 # 5530 <_sk_callback_avx+0x222> + .byte 196,98,125,24,29,15,48,0,0 // vbroadcastss 0x300f(%rip),%ymm11 # 554c <_sk_callback_avx+0x23a> .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,237,47,0,0 // vbroadcastss 0x2fed(%rip),%ymm11 # 5534 <_sk_callback_avx+0x226> + .byte 196,98,125,24,29,5,48,0,0 // vbroadcastss 0x3005(%rip),%ymm11 # 5550 <_sk_callback_avx+0x23e> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,222,47,0,0 // vbroadcastss 0x2fde(%rip),%ymm11 # 5538 <_sk_callback_avx+0x22a> + .byte 196,98,125,24,29,246,47,0,0 // vbroadcastss 0x2ff6(%rip),%ymm11 # 5554 <_sk_callback_avx+0x242> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,212,47,0,0 // vbroadcastss 0x2fd4(%rip),%ymm11 # 553c <_sk_callback_avx+0x22e> + .byte 196,98,125,24,29,236,47,0,0 // vbroadcastss 0x2fec(%rip),%ymm11 # 5558 <_sk_callback_avx+0x246> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 - .byte 196,98,125,24,21,197,47,0,0 // vbroadcastss 0x2fc5(%rip),%ymm10 # 5540 <_sk_callback_avx+0x232> + .byte 196,98,125,24,21,221,47,0,0 // vbroadcastss 0x2fdd(%rip),%ymm10 # 555c <_sk_callback_avx+0x24a> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -14189,7 +14201,7 @@ _sk_parametric_r_avx: .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,156,47,0,0 // vbroadcastss 0x2f9c(%rip),%ymm8 # 5544 <_sk_callback_avx+0x236> + .byte 196,98,125,24,5,180,47,0,0 // vbroadcastss 0x2fb4(%rip),%ymm8 # 5560 <_sk_callback_avx+0x24e> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14211,36 +14223,36 @@ _sk_parametric_g_avx: .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,217 // vcvtdq2ps %ymm1,%ymm11 - .byte 196,98,125,24,37,77,47,0,0 // vbroadcastss 0x2f4d(%rip),%ymm12 # 5548 <_sk_callback_avx+0x23a> + .byte 196,98,125,24,37,101,47,0,0 // vbroadcastss 0x2f65(%rip),%ymm12 # 5564 <_sk_callback_avx+0x252> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,67,47,0,0 // vbroadcastss 0x2f43(%rip),%ymm12 # 554c <_sk_callback_avx+0x23e> + .byte 196,98,125,24,37,91,47,0,0 // vbroadcastss 0x2f5b(%rip),%ymm12 # 5568 <_sk_callback_avx+0x256> .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,57,47,0,0 // vbroadcastss 0x2f39(%rip),%ymm12 # 5550 <_sk_callback_avx+0x242> + .byte 196,98,125,24,37,81,47,0,0 // vbroadcastss 0x2f51(%rip),%ymm12 # 556c <_sk_callback_avx+0x25a> .byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,47,47,0,0 // vbroadcastss 0x2f2f(%rip),%ymm12 # 5554 <_sk_callback_avx+0x246> + .byte 196,98,125,24,37,71,47,0,0 // vbroadcastss 0x2f47(%rip),%ymm12 # 5570 <_sk_callback_avx+0x25e> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,37,47,0,0 // vbroadcastss 0x2f25(%rip),%ymm12 # 5558 <_sk_callback_avx+0x24a> + .byte 196,98,125,24,37,61,47,0,0 // vbroadcastss 0x2f3d(%rip),%ymm12 # 5574 <_sk_callback_avx+0x262> .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,22,47,0,0 // vbroadcastss 0x2f16(%rip),%ymm12 # 555c <_sk_callback_avx+0x24e> + .byte 196,98,125,24,37,46,47,0,0 // vbroadcastss 0x2f2e(%rip),%ymm12 # 5578 <_sk_callback_avx+0x266> .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,12,47,0,0 // vbroadcastss 0x2f0c(%rip),%ymm12 # 5560 <_sk_callback_avx+0x252> + .byte 196,98,125,24,37,36,47,0,0 // vbroadcastss 0x2f24(%rip),%ymm12 # 557c <_sk_callback_avx+0x26a> .byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1 .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10 - .byte 196,98,125,24,29,240,46,0,0 // vbroadcastss 0x2ef0(%rip),%ymm11 # 5564 <_sk_callback_avx+0x256> + .byte 196,98,125,24,29,8,47,0,0 // vbroadcastss 0x2f08(%rip),%ymm11 # 5580 <_sk_callback_avx+0x26e> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,230,46,0,0 // vbroadcastss 0x2ee6(%rip),%ymm11 # 5568 <_sk_callback_avx+0x25a> + .byte 196,98,125,24,29,254,46,0,0 // vbroadcastss 0x2efe(%rip),%ymm11 # 5584 <_sk_callback_avx+0x272> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,215,46,0,0 // vbroadcastss 0x2ed7(%rip),%ymm11 # 556c <_sk_callback_avx+0x25e> + .byte 196,98,125,24,29,239,46,0,0 // vbroadcastss 0x2eef(%rip),%ymm11 # 5588 <_sk_callback_avx+0x276> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,205,46,0,0 // vbroadcastss 0x2ecd(%rip),%ymm11 # 5570 <_sk_callback_avx+0x262> + .byte 196,98,125,24,29,229,46,0,0 // vbroadcastss 0x2ee5(%rip),%ymm11 # 558c <_sk_callback_avx+0x27a> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 - .byte 196,98,125,24,21,190,46,0,0 // vbroadcastss 0x2ebe(%rip),%ymm10 # 5574 <_sk_callback_avx+0x266> + .byte 196,98,125,24,21,214,46,0,0 // vbroadcastss 0x2ed6(%rip),%ymm10 # 5590 <_sk_callback_avx+0x27e> .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -14248,7 +14260,7 @@ _sk_parametric_g_avx: .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,5,149,46,0,0 // vbroadcastss 0x2e95(%rip),%ymm8 # 5578 <_sk_callback_avx+0x26a> + .byte 196,98,125,24,5,173,46,0,0 // vbroadcastss 0x2ead(%rip),%ymm8 # 5594 <_sk_callback_avx+0x282> .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14270,36 +14282,36 @@ _sk_parametric_b_avx: .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,218 // vcvtdq2ps %ymm2,%ymm11 - .byte 196,98,125,24,37,70,46,0,0 // vbroadcastss 0x2e46(%rip),%ymm12 # 557c <_sk_callback_avx+0x26e> + .byte 196,98,125,24,37,94,46,0,0 // vbroadcastss 0x2e5e(%rip),%ymm12 # 5598 <_sk_callback_avx+0x286> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,60,46,0,0 // vbroadcastss 0x2e3c(%rip),%ymm12 # 5580 <_sk_callback_avx+0x272> + .byte 196,98,125,24,37,84,46,0,0 // vbroadcastss 0x2e54(%rip),%ymm12 # 559c <_sk_callback_avx+0x28a> .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,50,46,0,0 // vbroadcastss 0x2e32(%rip),%ymm12 # 5584 <_sk_callback_avx+0x276> + .byte 196,98,125,24,37,74,46,0,0 // vbroadcastss 0x2e4a(%rip),%ymm12 # 55a0 <_sk_callback_avx+0x28e> .byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,40,46,0,0 // vbroadcastss 0x2e28(%rip),%ymm12 # 5588 <_sk_callback_avx+0x27a> + .byte 196,98,125,24,37,64,46,0,0 // vbroadcastss 0x2e40(%rip),%ymm12 # 55a4 <_sk_callback_avx+0x292> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,30,46,0,0 // vbroadcastss 0x2e1e(%rip),%ymm12 # 558c <_sk_callback_avx+0x27e> + .byte 196,98,125,24,37,54,46,0,0 // vbroadcastss 0x2e36(%rip),%ymm12 # 55a8 <_sk_callback_avx+0x296> .byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,15,46,0,0 // vbroadcastss 0x2e0f(%rip),%ymm12 # 5590 <_sk_callback_avx+0x282> + .byte 196,98,125,24,37,39,46,0,0 // vbroadcastss 0x2e27(%rip),%ymm12 # 55ac <_sk_callback_avx+0x29a> .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,5,46,0,0 // vbroadcastss 0x2e05(%rip),%ymm12 # 5594 <_sk_callback_avx+0x286> + .byte 196,98,125,24,37,29,46,0,0 // vbroadcastss 0x2e1d(%rip),%ymm12 # 55b0 <_sk_callback_avx+0x29e> .byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2 .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10 - .byte 196,98,125,24,29,233,45,0,0 // vbroadcastss 0x2de9(%rip),%ymm11 # 5598 <_sk_callback_avx+0x28a> + .byte 196,98,125,24,29,1,46,0,0 // vbroadcastss 0x2e01(%rip),%ymm11 # 55b4 <_sk_callback_avx+0x2a2> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,223,45,0,0 // vbroadcastss 0x2ddf(%rip),%ymm11 # 559c <_sk_callback_avx+0x28e> + .byte 196,98,125,24,29,247,45,0,0 // vbroadcastss 0x2df7(%rip),%ymm11 # 55b8 <_sk_callback_avx+0x2a6> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,208,45,0,0 // vbroadcastss 0x2dd0(%rip),%ymm11 # 55a0 <_sk_callback_avx+0x292> + .byte 196,98,125,24,29,232,45,0,0 // vbroadcastss 0x2de8(%rip),%ymm11 # 55bc <_sk_callback_avx+0x2aa> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,198,45,0,0 // vbroadcastss 0x2dc6(%rip),%ymm11 # 55a4 <_sk_callback_avx+0x296> + .byte 196,98,125,24,29,222,45,0,0 // vbroadcastss 0x2dde(%rip),%ymm11 # 55c0 <_sk_callback_avx+0x2ae> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 - .byte 196,98,125,24,21,183,45,0,0 // vbroadcastss 0x2db7(%rip),%ymm10 # 55a8 <_sk_callback_avx+0x29a> + .byte 196,98,125,24,21,207,45,0,0 // vbroadcastss 0x2dcf(%rip),%ymm10 # 55c4 <_sk_callback_avx+0x2b2> .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -14307,7 +14319,7 @@ _sk_parametric_b_avx: .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,142,45,0,0 // vbroadcastss 0x2d8e(%rip),%ymm8 # 55ac <_sk_callback_avx+0x29e> + .byte 196,98,125,24,5,166,45,0,0 // vbroadcastss 0x2da6(%rip),%ymm8 # 55c8 <_sk_callback_avx+0x2b6> .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14329,36 +14341,36 @@ _sk_parametric_a_avx: .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,219 // vcvtdq2ps %ymm3,%ymm11 - .byte 196,98,125,24,37,63,45,0,0 // vbroadcastss 0x2d3f(%rip),%ymm12 # 55b0 <_sk_callback_avx+0x2a2> + .byte 196,98,125,24,37,87,45,0,0 // vbroadcastss 0x2d57(%rip),%ymm12 # 55cc <_sk_callback_avx+0x2ba> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,53,45,0,0 // vbroadcastss 0x2d35(%rip),%ymm12 # 55b4 <_sk_callback_avx+0x2a6> + .byte 196,98,125,24,37,77,45,0,0 // vbroadcastss 0x2d4d(%rip),%ymm12 # 55d0 <_sk_callback_avx+0x2be> .byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,43,45,0,0 // vbroadcastss 0x2d2b(%rip),%ymm12 # 55b8 <_sk_callback_avx+0x2aa> + .byte 196,98,125,24,37,67,45,0,0 // vbroadcastss 0x2d43(%rip),%ymm12 # 55d4 <_sk_callback_avx+0x2c2> .byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,33,45,0,0 // vbroadcastss 0x2d21(%rip),%ymm12 # 55bc <_sk_callback_avx+0x2ae> + .byte 196,98,125,24,37,57,45,0,0 // vbroadcastss 0x2d39(%rip),%ymm12 # 55d8 <_sk_callback_avx+0x2c6> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,23,45,0,0 // vbroadcastss 0x2d17(%rip),%ymm12 # 55c0 <_sk_callback_avx+0x2b2> + .byte 196,98,125,24,37,47,45,0,0 // vbroadcastss 0x2d2f(%rip),%ymm12 # 55dc <_sk_callback_avx+0x2ca> .byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,8,45,0,0 // vbroadcastss 0x2d08(%rip),%ymm12 # 55c4 <_sk_callback_avx+0x2b6> + .byte 196,98,125,24,37,32,45,0,0 // vbroadcastss 0x2d20(%rip),%ymm12 # 55e0 <_sk_callback_avx+0x2ce> .byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,254,44,0,0 // vbroadcastss 0x2cfe(%rip),%ymm12 # 55c8 <_sk_callback_avx+0x2ba> + .byte 196,98,125,24,37,22,45,0,0 // vbroadcastss 0x2d16(%rip),%ymm12 # 55e4 <_sk_callback_avx+0x2d2> .byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3 .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10 - .byte 196,98,125,24,29,226,44,0,0 // vbroadcastss 0x2ce2(%rip),%ymm11 # 55cc <_sk_callback_avx+0x2be> + .byte 196,98,125,24,29,250,44,0,0 // vbroadcastss 0x2cfa(%rip),%ymm11 # 55e8 <_sk_callback_avx+0x2d6> .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,216,44,0,0 // vbroadcastss 0x2cd8(%rip),%ymm11 # 55d0 <_sk_callback_avx+0x2c2> + .byte 196,98,125,24,29,240,44,0,0 // vbroadcastss 0x2cf0(%rip),%ymm11 # 55ec <_sk_callback_avx+0x2da> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,201,44,0,0 // vbroadcastss 0x2cc9(%rip),%ymm11 # 55d4 <_sk_callback_avx+0x2c6> + .byte 196,98,125,24,29,225,44,0,0 // vbroadcastss 0x2ce1(%rip),%ymm11 # 55f0 <_sk_callback_avx+0x2de> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,191,44,0,0 // vbroadcastss 0x2cbf(%rip),%ymm11 # 55d8 <_sk_callback_avx+0x2ca> + .byte 196,98,125,24,29,215,44,0,0 // vbroadcastss 0x2cd7(%rip),%ymm11 # 55f4 <_sk_callback_avx+0x2e2> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 - .byte 196,98,125,24,21,176,44,0,0 // vbroadcastss 0x2cb0(%rip),%ymm10 # 55dc <_sk_callback_avx+0x2ce> + .byte 196,98,125,24,21,200,44,0,0 // vbroadcastss 0x2cc8(%rip),%ymm10 # 55f8 <_sk_callback_avx+0x2e6> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -14366,7 +14378,7 @@ _sk_parametric_a_avx: .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 - .byte 196,98,125,24,5,135,44,0,0 // vbroadcastss 0x2c87(%rip),%ymm8 # 55e0 <_sk_callback_avx+0x2d2> + .byte 196,98,125,24,5,159,44,0,0 // vbroadcastss 0x2c9f(%rip),%ymm8 # 55fc <_sk_callback_avx+0x2ea> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14375,31 +14387,31 @@ HIDDEN _sk_lab_to_xyz_avx .globl _sk_lab_to_xyz_avx FUNCTION(_sk_lab_to_xyz_avx) _sk_lab_to_xyz_avx: - .byte 196,98,125,24,5,121,44,0,0 // vbroadcastss 0x2c79(%rip),%ymm8 # 55e4 <_sk_callback_avx+0x2d6> + .byte 196,98,125,24,5,145,44,0,0 // vbroadcastss 0x2c91(%rip),%ymm8 # 5600 <_sk_callback_avx+0x2ee> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,111,44,0,0 // vbroadcastss 0x2c6f(%rip),%ymm8 # 55e8 <_sk_callback_avx+0x2da> + .byte 196,98,125,24,5,135,44,0,0 // vbroadcastss 0x2c87(%rip),%ymm8 # 5604 <_sk_callback_avx+0x2f2> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,13,101,44,0,0 // vbroadcastss 0x2c65(%rip),%ymm9 # 55ec <_sk_callback_avx+0x2de> + .byte 196,98,125,24,13,125,44,0,0 // vbroadcastss 0x2c7d(%rip),%ymm9 # 5608 <_sk_callback_avx+0x2f6> .byte 196,193,116,88,201 // vaddps %ymm9,%ymm1,%ymm1 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 - .byte 196,98,125,24,5,81,44,0,0 // vbroadcastss 0x2c51(%rip),%ymm8 # 55f0 <_sk_callback_avx+0x2e2> + .byte 196,98,125,24,5,105,44,0,0 // vbroadcastss 0x2c69(%rip),%ymm8 # 560c <_sk_callback_avx+0x2fa> .byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,71,44,0,0 // vbroadcastss 0x2c47(%rip),%ymm8 # 55f4 <_sk_callback_avx+0x2e6> + .byte 196,98,125,24,5,95,44,0,0 // vbroadcastss 0x2c5f(%rip),%ymm8 # 5610 <_sk_callback_avx+0x2fe> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,61,44,0,0 // vbroadcastss 0x2c3d(%rip),%ymm8 # 55f8 <_sk_callback_avx+0x2ea> + .byte 196,98,125,24,5,85,44,0,0 // vbroadcastss 0x2c55(%rip),%ymm8 # 5614 <_sk_callback_avx+0x302> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1 - .byte 196,98,125,24,5,47,44,0,0 // vbroadcastss 0x2c2f(%rip),%ymm8 # 55fc <_sk_callback_avx+0x2ee> + .byte 196,98,125,24,5,71,44,0,0 // vbroadcastss 0x2c47(%rip),%ymm8 # 5618 <_sk_callback_avx+0x306> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 197,252,92,210 // vsubps %ymm2,%ymm0,%ymm2 .byte 197,116,89,193 // vmulps %ymm1,%ymm1,%ymm8 .byte 196,65,116,89,192 // vmulps %ymm8,%ymm1,%ymm8 - .byte 196,98,125,24,13,24,44,0,0 // vbroadcastss 0x2c18(%rip),%ymm9 # 5600 <_sk_callback_avx+0x2f2> + .byte 196,98,125,24,13,48,44,0,0 // vbroadcastss 0x2c30(%rip),%ymm9 # 561c <_sk_callback_avx+0x30a> .byte 196,65,52,194,208,1 // vcmpltps %ymm8,%ymm9,%ymm10 - .byte 196,98,125,24,29,13,44,0,0 // vbroadcastss 0x2c0d(%rip),%ymm11 # 5604 <_sk_callback_avx+0x2f6> + .byte 196,98,125,24,29,37,44,0,0 // vbroadcastss 0x2c25(%rip),%ymm11 # 5620 <_sk_callback_avx+0x30e> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,37,3,44,0,0 // vbroadcastss 0x2c03(%rip),%ymm12 # 5608 <_sk_callback_avx+0x2fa> + .byte 196,98,125,24,37,27,44,0,0 // vbroadcastss 0x2c1b(%rip),%ymm12 # 5624 <_sk_callback_avx+0x312> .byte 196,193,116,89,204 // vmulps %ymm12,%ymm1,%ymm1 .byte 196,67,117,74,192,160 // vblendvps %ymm10,%ymm8,%ymm1,%ymm8 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1 @@ -14414,9 +14426,9 @@ _sk_lab_to_xyz_avx: .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,193,108,89,212 // vmulps %ymm12,%ymm2,%ymm2 .byte 196,227,109,74,208,144 // vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - .byte 196,226,125,24,5,185,43,0,0 // vbroadcastss 0x2bb9(%rip),%ymm0 # 560c <_sk_callback_avx+0x2fe> + .byte 196,226,125,24,5,209,43,0,0 // vbroadcastss 0x2bd1(%rip),%ymm0 # 5628 <_sk_callback_avx+0x316> .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,176,43,0,0 // vbroadcastss 0x2bb0(%rip),%ymm8 # 5610 <_sk_callback_avx+0x302> + .byte 196,98,125,24,5,200,43,0,0 // vbroadcastss 0x2bc8(%rip),%ymm8 # 562c <_sk_callback_avx+0x31a> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14430,14 +14442,14 @@ _sk_load_a8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,62 // jne 2ab7 <_sk_load_a8_avx+0x4e> + .byte 117,62 // jne 2abb <_sk_load_a8_avx+0x4e> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,116,43,0,0 // vbroadcastss 0x2b74(%rip),%ymm1 # 5614 <_sk_callback_avx+0x306> + .byte 196,226,125,24,13,140,43,0,0 // vbroadcastss 0x2b8c(%rip),%ymm1 # 5630 <_sk_callback_avx+0x31e> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -14454,9 +14466,9 @@ _sk_load_a8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 2abf <_sk_load_a8_avx+0x56> + .byte 117,234 // jne 2ac3 <_sk_load_a8_avx+0x56> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,161 // jmp 2a7d <_sk_load_a8_avx+0x14> + .byte 235,161 // jmp 2a81 <_sk_load_a8_avx+0x14> HIDDEN _sk_gather_a8_avx .globl _sk_gather_a8_avx @@ -14506,7 +14518,7 @@ _sk_gather_a8_avx: .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,105,42,0,0 // vbroadcastss 0x2a69(%rip),%ymm1 # 5618 <_sk_callback_avx+0x30a> + .byte 196,226,125,24,13,129,42,0,0 // vbroadcastss 0x2a81(%rip),%ymm1 # 5634 <_sk_callback_avx+0x322> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -14524,14 +14536,14 @@ FUNCTION(_sk_store_a8_avx) _sk_store_a8_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,68,42,0,0 // vbroadcastss 0x2a44(%rip),%ymm8 # 561c <_sk_callback_avx+0x30e> + .byte 196,98,125,24,5,92,42,0,0 // vbroadcastss 0x2a5c(%rip),%ymm8 # 5638 <_sk_callback_avx+0x326> .byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 2c01 <_sk_store_a8_avx+0x37> + .byte 117,10 // jne 2c05 <_sk_store_a8_avx+0x37> .byte 196,65,123,17,4,58 // vmovsd %xmm8,(%r10,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14539,10 +14551,10 @@ _sk_store_a8_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 2bfd <_sk_store_a8_avx+0x33> + .byte 119,236 // ja 2c01 <_sk_store_a8_avx+0x33> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 2c64 <_sk_store_a8_avx+0x9a> + .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 2c68 <_sk_store_a8_avx+0x9a> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -14553,7 +14565,7 @@ _sk_store_a8_avx: .byte 196,67,121,20,68,58,2,4 // vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1) .byte 196,67,121,20,68,58,1,2 // vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1) .byte 196,67,121,20,4,58,0 // vpextrb $0x0,%xmm8,(%r10,%rdi,1) - .byte 235,154 // jmp 2bfd <_sk_store_a8_avx+0x33> + .byte 235,154 // jmp 2c01 <_sk_store_a8_avx+0x33> .byte 144 // nop .byte 246,255 // idiv %bh .byte 255 // (bad) @@ -14587,17 +14599,17 @@ _sk_load_g8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,67 // jne 2cd3 <_sk_load_g8_avx+0x53> + .byte 117,67 // jne 2cd7 <_sk_load_g8_avx+0x53> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,105,41,0,0 // vbroadcastss 0x2969(%rip),%ymm1 # 5620 <_sk_callback_avx+0x312> + .byte 196,226,125,24,13,129,41,0,0 // vbroadcastss 0x2981(%rip),%ymm1 # 563c <_sk_callback_avx+0x32a> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,94,41,0,0 // vbroadcastss 0x295e(%rip),%ymm3 # 5624 <_sk_callback_avx+0x316> + .byte 196,226,125,24,29,118,41,0,0 // vbroadcastss 0x2976(%rip),%ymm3 # 5640 <_sk_callback_avx+0x32e> .byte 76,137,193 // mov %r8,%rcx .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 @@ -14611,9 +14623,9 @@ _sk_load_g8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 2cdb <_sk_load_g8_avx+0x5b> + .byte 117,234 // jne 2cdf <_sk_load_g8_avx+0x5b> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,156 // jmp 2c94 <_sk_load_g8_avx+0x14> + .byte 235,156 // jmp 2c98 <_sk_load_g8_avx+0x14> HIDDEN _sk_gather_g8_avx .globl _sk_gather_g8_avx @@ -14663,10 +14675,10 @@ _sk_gather_g8_avx: .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,93,40,0,0 // vbroadcastss 0x285d(%rip),%ymm1 # 5628 <_sk_callback_avx+0x31a> + .byte 196,226,125,24,13,117,40,0,0 // vbroadcastss 0x2875(%rip),%ymm1 # 5644 <_sk_callback_avx+0x332> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,82,40,0,0 // vbroadcastss 0x2852(%rip),%ymm3 # 562c <_sk_callback_avx+0x31e> + .byte 196,226,125,24,29,106,40,0,0 // vbroadcastss 0x286a(%rip),%ymm3 # 5648 <_sk_callback_avx+0x336> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 91 // pop %rbx @@ -14682,9 +14694,9 @@ _sk_gather_i8_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 2dfa <_sk_gather_i8_avx+0xf> + .byte 116,5 // je 2dfe <_sk_gather_i8_avx+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 2dfc <_sk_gather_i8_avx+0x11> + .byte 235,2 // jmp 2e00 <_sk_gather_i8_avx+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -14746,10 +14758,10 @@ _sk_gather_i8_avx: .byte 196,163,121,34,4,163,2 // vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0 .byte 196,163,121,34,28,19,3 // vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - .byte 197,124,40,21,186,40,0,0 // vmovaps 0x28ba(%rip),%ymm10 # 57e0 <_sk_callback_avx+0x4d2> + .byte 197,124,40,21,214,40,0,0 // vmovaps 0x28d6(%rip),%ymm10 # 5800 <_sk_callback_avx+0x4ee> .byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,248,38,0,0 // vbroadcastss 0x26f8(%rip),%ymm9 # 5630 <_sk_callback_avx+0x322> + .byte 196,98,125,24,13,16,39,0,0 // vbroadcastss 0x2710(%rip),%ymm9 # 564c <_sk_callback_avx+0x33a> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 @@ -14783,38 +14795,38 @@ _sk_load_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,128,0,0,0 // jne 3030 <_sk_load_565_avx+0x8e> + .byte 15,133,128,0,0,0 // jne 3034 <_sk_load_565_avx+0x8e> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - .byte 196,226,125,24,5,98,38,0,0 // vbroadcastss 0x2662(%rip),%ymm0 # 5634 <_sk_callback_avx+0x326> + .byte 196,226,125,24,5,122,38,0,0 // vbroadcastss 0x267a(%rip),%ymm0 # 5650 <_sk_callback_avx+0x33e> .byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,85,38,0,0 // vbroadcastss 0x2655(%rip),%ymm1 # 5638 <_sk_callback_avx+0x32a> + .byte 196,226,125,24,13,109,38,0,0 // vbroadcastss 0x266d(%rip),%ymm1 # 5654 <_sk_callback_avx+0x342> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,76,38,0,0 // vbroadcastss 0x264c(%rip),%ymm1 # 563c <_sk_callback_avx+0x32e> + .byte 196,226,125,24,13,100,38,0,0 // vbroadcastss 0x2664(%rip),%ymm1 # 5658 <_sk_callback_avx+0x346> .byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,63,38,0,0 // vbroadcastss 0x263f(%rip),%ymm3 # 5640 <_sk_callback_avx+0x332> + .byte 196,226,125,24,29,87,38,0,0 // vbroadcastss 0x2657(%rip),%ymm3 # 565c <_sk_callback_avx+0x34a> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,24,29,54,38,0,0 // vbroadcastss 0x2636(%rip),%ymm3 # 5644 <_sk_callback_avx+0x336> + .byte 196,226,125,24,29,78,38,0,0 // vbroadcastss 0x264e(%rip),%ymm3 # 5660 <_sk_callback_avx+0x34e> .byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,41,38,0,0 // vbroadcastss 0x2629(%rip),%ymm3 # 5648 <_sk_callback_avx+0x33a> + .byte 196,226,125,24,29,65,38,0,0 // vbroadcastss 0x2641(%rip),%ymm3 # 5664 <_sk_callback_avx+0x352> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,30,38,0,0 // vbroadcastss 0x261e(%rip),%ymm3 # 564c <_sk_callback_avx+0x33e> + .byte 196,226,125,24,29,54,38,0,0 // vbroadcastss 0x2636(%rip),%ymm3 # 5668 <_sk_callback_avx+0x356> .byte 255,224 // jmpq *%rax .byte 65,137,200 // mov %ecx,%r8d .byte 65,128,224,7 // and $0x7,%r8b .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,110,255,255,255 // ja 2fb6 <_sk_load_565_avx+0x14> + .byte 15,135,110,255,255,255 // ja 2fba <_sk_load_565_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 309c <_sk_load_565_avx+0xfa> + .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 30a0 <_sk_load_565_avx+0xfa> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -14826,7 +14838,7 @@ _sk_load_565_avx: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,26,255,255,255 // jmpq 2fb6 <_sk_load_565_avx+0x14> + .byte 233,26,255,255,255 // jmpq 2fba <_sk_load_565_avx+0x14> .byte 244 // hlt .byte 255 // (bad) .byte 255 // (bad) @@ -14904,23 +14916,23 @@ _sk_gather_565_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - .byte 196,226,125,24,5,190,36,0,0 // vbroadcastss 0x24be(%rip),%ymm0 # 5650 <_sk_callback_avx+0x342> + .byte 196,226,125,24,5,214,36,0,0 // vbroadcastss 0x24d6(%rip),%ymm0 # 566c <_sk_callback_avx+0x35a> .byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,177,36,0,0 // vbroadcastss 0x24b1(%rip),%ymm1 # 5654 <_sk_callback_avx+0x346> + .byte 196,226,125,24,13,201,36,0,0 // vbroadcastss 0x24c9(%rip),%ymm1 # 5670 <_sk_callback_avx+0x35e> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,168,36,0,0 // vbroadcastss 0x24a8(%rip),%ymm1 # 5658 <_sk_callback_avx+0x34a> + .byte 196,226,125,24,13,192,36,0,0 // vbroadcastss 0x24c0(%rip),%ymm1 # 5674 <_sk_callback_avx+0x362> .byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,155,36,0,0 // vbroadcastss 0x249b(%rip),%ymm3 # 565c <_sk_callback_avx+0x34e> + .byte 196,226,125,24,29,179,36,0,0 // vbroadcastss 0x24b3(%rip),%ymm3 # 5678 <_sk_callback_avx+0x366> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,24,29,146,36,0,0 // vbroadcastss 0x2492(%rip),%ymm3 # 5660 <_sk_callback_avx+0x352> + .byte 196,226,125,24,29,170,36,0,0 // vbroadcastss 0x24aa(%rip),%ymm3 # 567c <_sk_callback_avx+0x36a> .byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,133,36,0,0 // vbroadcastss 0x2485(%rip),%ymm3 # 5664 <_sk_callback_avx+0x356> + .byte 196,226,125,24,29,157,36,0,0 // vbroadcastss 0x249d(%rip),%ymm3 # 5680 <_sk_callback_avx+0x36e> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,122,36,0,0 // vbroadcastss 0x247a(%rip),%ymm3 # 5668 <_sk_callback_avx+0x35a> + .byte 196,226,125,24,29,146,36,0,0 // vbroadcastss 0x2492(%rip),%ymm3 # 5684 <_sk_callback_avx+0x372> .byte 91 // pop %rbx .byte 65,92 // pop %r12 .byte 65,94 // pop %r14 @@ -14934,14 +14946,14 @@ FUNCTION(_sk_store_565_avx) _sk_store_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,102,36,0,0 // vbroadcastss 0x2466(%rip),%ymm8 # 566c <_sk_callback_avx+0x35e> + .byte 196,98,125,24,5,126,36,0,0 // vbroadcastss 0x247e(%rip),%ymm8 # 5688 <_sk_callback_avx+0x376> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 .byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - .byte 196,98,125,24,21,63,36,0,0 // vbroadcastss 0x243f(%rip),%ymm10 # 5670 <_sk_callback_avx+0x362> + .byte 196,98,125,24,21,87,36,0,0 // vbroadcastss 0x2457(%rip),%ymm10 # 568c <_sk_callback_avx+0x37a> .byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 .byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11 @@ -14955,7 +14967,7 @@ _sk_store_565_avx: .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 3281 <_sk_store_565_avx+0x89> + .byte 117,10 // jne 3285 <_sk_store_565_avx+0x89> .byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14963,9 +14975,9 @@ _sk_store_565_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 327d <_sk_store_565_avx+0x85> + .byte 119,236 // ja 3281 <_sk_store_565_avx+0x85> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 32e0 <_sk_store_565_avx+0xe8> + .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 32e4 <_sk_store_565_avx+0xe8> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -14976,7 +14988,7 @@ _sk_store_565_avx: .byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) .byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) .byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2) - .byte 235,159 // jmp 327d <_sk_store_565_avx+0x85> + .byte 235,159 // jmp 3281 <_sk_store_565_avx+0x85> .byte 102,144 // xchg %ax,%ax .byte 245 // cmc .byte 255 // (bad) @@ -15009,31 +15021,31 @@ _sk_load_4444_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,152,0,0,0 // jne 33a2 <_sk_load_4444_avx+0xa6> + .byte 15,133,152,0,0,0 // jne 33a6 <_sk_load_4444_avx+0xa6> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - .byte 196,226,125,24,5,72,35,0,0 // vbroadcastss 0x2348(%rip),%ymm0 # 5674 <_sk_callback_avx+0x366> + .byte 196,226,125,24,5,96,35,0,0 // vbroadcastss 0x2360(%rip),%ymm0 # 5690 <_sk_callback_avx+0x37e> .byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,59,35,0,0 // vbroadcastss 0x233b(%rip),%ymm1 # 5678 <_sk_callback_avx+0x36a> + .byte 196,226,125,24,13,83,35,0,0 // vbroadcastss 0x2353(%rip),%ymm1 # 5694 <_sk_callback_avx+0x382> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,50,35,0,0 // vbroadcastss 0x2332(%rip),%ymm1 # 567c <_sk_callback_avx+0x36e> + .byte 196,226,125,24,13,74,35,0,0 // vbroadcastss 0x234a(%rip),%ymm1 # 5698 <_sk_callback_avx+0x386> .byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,37,35,0,0 // vbroadcastss 0x2325(%rip),%ymm2 # 5680 <_sk_callback_avx+0x372> + .byte 196,226,125,24,21,61,35,0,0 // vbroadcastss 0x233d(%rip),%ymm2 # 569c <_sk_callback_avx+0x38a> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,28,35,0,0 // vbroadcastss 0x231c(%rip),%ymm2 # 5684 <_sk_callback_avx+0x376> + .byte 196,226,125,24,21,52,35,0,0 // vbroadcastss 0x2334(%rip),%ymm2 # 56a0 <_sk_callback_avx+0x38e> .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,15,35,0,0 // vbroadcastss 0x230f(%rip),%ymm8 # 5688 <_sk_callback_avx+0x37a> + .byte 196,98,125,24,5,39,35,0,0 // vbroadcastss 0x2327(%rip),%ymm8 # 56a4 <_sk_callback_avx+0x392> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,5,35,0,0 // vbroadcastss 0x2305(%rip),%ymm8 # 568c <_sk_callback_avx+0x37e> + .byte 196,98,125,24,5,29,35,0,0 // vbroadcastss 0x231d(%rip),%ymm8 # 56a8 <_sk_callback_avx+0x396> .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,247,34,0,0 // vbroadcastss 0x22f7(%rip),%ymm8 # 5690 <_sk_callback_avx+0x382> + .byte 196,98,125,24,5,15,35,0,0 // vbroadcastss 0x230f(%rip),%ymm8 # 56ac <_sk_callback_avx+0x39a> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15042,9 +15054,9 @@ _sk_load_4444_avx: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,86,255,255,255 // ja 3310 <_sk_load_4444_avx+0x14> + .byte 15,135,86,255,255,255 // ja 3314 <_sk_load_4444_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3410 <_sk_load_4444_avx+0x114> + .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3414 <_sk_load_4444_avx+0x114> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15056,7 +15068,7 @@ _sk_load_4444_avx: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,2,255,255,255 // jmpq 3310 <_sk_load_4444_avx+0x14> + .byte 233,2,255,255,255 // jmpq 3314 <_sk_load_4444_avx+0x14> .byte 102,144 // xchg %ax,%ax .byte 242,255 // repnz (bad) .byte 255 // (bad) @@ -15135,25 +15147,25 @@ _sk_gather_4444_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - .byte 196,226,125,24,5,142,33,0,0 // vbroadcastss 0x218e(%rip),%ymm0 # 5694 <_sk_callback_avx+0x386> + .byte 196,226,125,24,5,166,33,0,0 // vbroadcastss 0x21a6(%rip),%ymm0 # 56b0 <_sk_callback_avx+0x39e> .byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,129,33,0,0 // vbroadcastss 0x2181(%rip),%ymm1 # 5698 <_sk_callback_avx+0x38a> + .byte 196,226,125,24,13,153,33,0,0 // vbroadcastss 0x2199(%rip),%ymm1 # 56b4 <_sk_callback_avx+0x3a2> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,120,33,0,0 // vbroadcastss 0x2178(%rip),%ymm1 # 569c <_sk_callback_avx+0x38e> + .byte 196,226,125,24,13,144,33,0,0 // vbroadcastss 0x2190(%rip),%ymm1 # 56b8 <_sk_callback_avx+0x3a6> .byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,107,33,0,0 // vbroadcastss 0x216b(%rip),%ymm2 # 56a0 <_sk_callback_avx+0x392> + .byte 196,226,125,24,21,131,33,0,0 // vbroadcastss 0x2183(%rip),%ymm2 # 56bc <_sk_callback_avx+0x3aa> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,98,33,0,0 // vbroadcastss 0x2162(%rip),%ymm2 # 56a4 <_sk_callback_avx+0x396> + .byte 196,226,125,24,21,122,33,0,0 // vbroadcastss 0x217a(%rip),%ymm2 # 56c0 <_sk_callback_avx+0x3ae> .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,85,33,0,0 // vbroadcastss 0x2155(%rip),%ymm8 # 56a8 <_sk_callback_avx+0x39a> + .byte 196,98,125,24,5,109,33,0,0 // vbroadcastss 0x216d(%rip),%ymm8 # 56c4 <_sk_callback_avx+0x3b2> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,75,33,0,0 // vbroadcastss 0x214b(%rip),%ymm8 # 56ac <_sk_callback_avx+0x39e> + .byte 196,98,125,24,5,99,33,0,0 // vbroadcastss 0x2163(%rip),%ymm8 # 56c8 <_sk_callback_avx+0x3b6> .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,61,33,0,0 // vbroadcastss 0x213d(%rip),%ymm8 # 56b0 <_sk_callback_avx+0x3a2> + .byte 196,98,125,24,5,85,33,0,0 // vbroadcastss 0x2155(%rip),%ymm8 # 56cc <_sk_callback_avx+0x3ba> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx @@ -15169,7 +15181,7 @@ FUNCTION(_sk_store_4444_avx) _sk_store_4444_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,34,33,0,0 // vbroadcastss 0x2122(%rip),%ymm8 # 56b4 <_sk_callback_avx+0x3a6> + .byte 196,98,125,24,5,58,33,0,0 // vbroadcastss 0x213a(%rip),%ymm8 # 56d0 <_sk_callback_avx+0x3be> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10 @@ -15196,7 +15208,7 @@ _sk_store_4444_avx: .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 362b <_sk_store_4444_avx+0xa7> + .byte 117,10 // jne 362f <_sk_store_4444_avx+0xa7> .byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15204,9 +15216,9 @@ _sk_store_4444_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 3627 <_sk_store_4444_avx+0xa3> + .byte 119,236 // ja 362b <_sk_store_4444_avx+0xa3> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,66,0,0,0 // lea 0x42(%rip),%r9 # 3688 <_sk_store_4444_avx+0x104> + .byte 76,141,13,66,0,0,0 // lea 0x42(%rip),%r9 # 368c <_sk_store_4444_avx+0x104> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15217,7 +15229,7 @@ _sk_store_4444_avx: .byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) .byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) .byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2) - .byte 235,159 // jmp 3627 <_sk_store_4444_avx+0xa3> + .byte 235,159 // jmp 362b <_sk_store_4444_avx+0xa3> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -15248,12 +15260,12 @@ _sk_load_8888_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,135,0,0,0 // jne 3739 <_sk_load_8888_avx+0x95> + .byte 15,133,135,0,0,0 // jne 373d <_sk_load_8888_avx+0x95> .byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9 - .byte 197,124,40,21,64,33,0,0 // vmovaps 0x2140(%rip),%ymm10 # 5800 <_sk_callback_avx+0x4f2> + .byte 197,124,40,21,92,33,0,0 // vmovaps 0x215c(%rip),%ymm10 # 5820 <_sk_callback_avx+0x50e> .byte 196,193,52,84,194 // vandps %ymm10,%ymm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,230,31,0,0 // vbroadcastss 0x1fe6(%rip),%ymm8 # 56b8 <_sk_callback_avx+0x3aa> + .byte 196,98,125,24,5,254,31,0,0 // vbroadcastss 0x1ffe(%rip),%ymm8 # 56d4 <_sk_callback_avx+0x3c2> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 196,193,113,114,209,8 // vpsrld $0x8,%xmm9,%xmm1 .byte 196,99,125,25,203,1 // vextractf128 $0x1,%ymm9,%xmm3 @@ -15280,9 +15292,9 @@ _sk_load_8888_avx: .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,102,255,255,255 // ja 36b8 <_sk_load_8888_avx+0x14> + .byte 15,135,102,255,255,255 // ja 36bc <_sk_load_8888_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,139,0,0,0 // lea 0x8b(%rip),%r9 # 37e8 <_sk_load_8888_avx+0x144> + .byte 76,141,13,139,0,0,0 // lea 0x8b(%rip),%r9 # 37ec <_sk_load_8888_avx+0x144> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15305,7 +15317,7 @@ _sk_load_8888_avx: .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 .byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 - .byte 233,210,254,255,255 // jmpq 36b8 <_sk_load_8888_avx+0x14> + .byte 233,210,254,255,255 // jmpq 36bc <_sk_load_8888_avx+0x14> .byte 102,144 // xchg %ax,%ax .byte 236 // in (%dx),%al .byte 255 // (bad) @@ -15323,7 +15335,7 @@ _sk_load_8888_avx: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 126,255 // jle 3801 <_sk_load_8888_avx+0x15d> + .byte 126,255 // jle 3805 <_sk_load_8888_avx+0x15d> .byte 255 // (bad) .byte 255 // .byte 0xff @@ -15368,10 +15380,10 @@ _sk_gather_8888_avx: .byte 196,131,121,34,4,152,2 // vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 .byte 196,131,121,34,28,144,3 // vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - .byte 197,124,40,21,106,31,0,0 // vmovaps 0x1f6a(%rip),%ymm10 # 5820 <_sk_callback_avx+0x512> + .byte 197,124,40,21,134,31,0,0 // vmovaps 0x1f86(%rip),%ymm10 # 5840 <_sk_callback_avx+0x52e> .byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,244,29,0,0 // vbroadcastss 0x1df4(%rip),%ymm9 # 56bc <_sk_callback_avx+0x3ae> + .byte 196,98,125,24,13,12,30,0,0 // vbroadcastss 0x1e0c(%rip),%ymm9 # 56d8 <_sk_callback_avx+0x3c6> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 @@ -15403,7 +15415,7 @@ FUNCTION(_sk_store_8888_avx) _sk_store_8888_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,130,29,0,0 // vbroadcastss 0x1d82(%rip),%ymm8 # 56c0 <_sk_callback_avx+0x3b2> + .byte 196,98,125,24,5,154,29,0,0 // vbroadcastss 0x1d9a(%rip),%ymm8 # 56dc <_sk_callback_avx+0x3ca> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,65,116,89,208 // vmulps %ymm8,%ymm1,%ymm10 @@ -15428,7 +15440,7 @@ _sk_store_8888_avx: .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8 .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 39cc <_sk_store_8888_avx+0x9c> + .byte 117,10 // jne 39d0 <_sk_store_8888_avx+0x9c> .byte 196,65,124,17,4,186 // vmovups %ymm8,(%r10,%rdi,4) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15436,9 +15448,9 @@ _sk_store_8888_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 39c8 <_sk_store_8888_avx+0x98> + .byte 119,236 // ja 39cc <_sk_store_8888_avx+0x98> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,85,0,0,0 // lea 0x55(%rip),%r9 # 3a3c <_sk_store_8888_avx+0x10c> + .byte 76,141,13,85,0,0,0 // lea 0x55(%rip),%r9 # 3a40 <_sk_store_8888_avx+0x10c> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15452,7 +15464,7 @@ _sk_store_8888_avx: .byte 196,67,121,22,68,186,8,2 // vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4) .byte 196,67,121,22,68,186,4,1 // vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4) .byte 196,65,121,126,4,186 // vmovd %xmm8,(%r10,%rdi,4) - .byte 235,143 // jmp 39c8 <_sk_store_8888_avx+0x98> + .byte 235,143 // jmp 39cc <_sk_store_8888_avx+0x98> .byte 15,31,0 // nopl (%rax) .byte 245 // cmc .byte 255 // (bad) @@ -15490,7 +15502,7 @@ _sk_load_f16_avx: .byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp) .byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp) .byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp) - .byte 15,133,141,2,0,0 // jne 3d0f <_sk_load_f16_avx+0x2b7> + .byte 15,133,141,2,0,0 // jne 3d13 <_sk_load_f16_avx+0x2b7> .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 .byte 197,249,16,76,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm1 @@ -15508,13 +15520,13 @@ _sk_load_f16_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - .byte 196,98,125,24,37,233,27,0,0 // vbroadcastss 0x1be9(%rip),%ymm12 # 56c4 <_sk_callback_avx+0x3b6> + .byte 196,98,125,24,37,1,28,0,0 // vbroadcastss 0x1c01(%rip),%ymm12 # 56e0 <_sk_callback_avx+0x3ce> .byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1 .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 - .byte 196,98,121,24,29,213,27,0,0 // vbroadcastss 0x1bd5(%rip),%xmm11 # 56c8 <_sk_callback_avx+0x3ba> + .byte 196,98,121,24,29,237,27,0,0 // vbroadcastss 0x1bed(%rip),%xmm11 # 56e4 <_sk_callback_avx+0x3d2> .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 - .byte 196,98,121,24,45,203,27,0,0 // vbroadcastss 0x1bcb(%rip),%xmm13 # 56cc <_sk_callback_avx+0x3be> + .byte 196,98,121,24,45,227,27,0,0 // vbroadcastss 0x1be3(%rip),%xmm13 # 56e8 <_sk_callback_avx+0x3d6> .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -15528,7 +15540,7 @@ _sk_load_f16_avx: .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 .byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0 .byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1 - .byte 196,226,121,24,29,129,27,0,0 // vbroadcastss 0x1b81(%rip),%xmm3 # 56d0 <_sk_callback_avx+0x3c2> + .byte 196,226,121,24,29,153,27,0,0 // vbroadcastss 0x1b99(%rip),%xmm3 # 56ec <_sk_callback_avx+0x3da> .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -15621,29 +15633,29 @@ _sk_load_f16_avx: .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,79 // je 3d6e <_sk_load_f16_avx+0x316> + .byte 116,79 // je 3d72 <_sk_load_f16_avx+0x316> .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,67 // jb 3d6e <_sk_load_f16_avx+0x316> + .byte 114,67 // jb 3d72 <_sk_load_f16_avx+0x316> .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,68 // je 3d7b <_sk_load_f16_avx+0x323> + .byte 116,68 // je 3d7f <_sk_load_f16_avx+0x323> .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,56 // jb 3d7b <_sk_load_f16_avx+0x323> + .byte 114,56 // jb 3d7f <_sk_load_f16_avx+0x323> .byte 197,251,16,76,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,70,253,255,255 // je 3a99 <_sk_load_f16_avx+0x41> + .byte 15,132,70,253,255,255 // je 3a9d <_sk_load_f16_avx+0x41> .byte 197,241,22,76,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,54,253,255,255 // jb 3a99 <_sk_load_f16_avx+0x41> + .byte 15,130,54,253,255,255 // jb 3a9d <_sk_load_f16_avx+0x41> .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 - .byte 233,43,253,255,255 // jmpq 3a99 <_sk_load_f16_avx+0x41> + .byte 233,43,253,255,255 // jmpq 3a9d <_sk_load_f16_avx+0x41> .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,30,253,255,255 // jmpq 3a99 <_sk_load_f16_avx+0x41> + .byte 233,30,253,255,255 // jmpq 3a9d <_sk_load_f16_avx+0x41> .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 - .byte 233,21,253,255,255 // jmpq 3a99 <_sk_load_f16_avx+0x41> + .byte 233,21,253,255,255 // jmpq 3a9d <_sk_load_f16_avx+0x41> HIDDEN _sk_gather_f16_avx .globl _sk_gather_f16_avx @@ -15707,13 +15719,13 @@ _sk_gather_f16_avx: .byte 197,249,105,210 // vpunpckhwd %xmm2,%xmm0,%xmm2 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - .byte 196,98,125,24,37,69,24,0,0 // vbroadcastss 0x1845(%rip),%ymm12 # 56d4 <_sk_callback_avx+0x3c6> + .byte 196,98,125,24,37,93,24,0,0 // vbroadcastss 0x185d(%rip),%ymm12 # 56f0 <_sk_callback_avx+0x3de> .byte 196,193,124,84,212 // vandps %ymm12,%ymm0,%ymm2 .byte 197,252,87,194 // vxorps %ymm2,%ymm0,%ymm0 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 - .byte 196,98,121,24,29,49,24,0,0 // vbroadcastss 0x1831(%rip),%xmm11 # 56d8 <_sk_callback_avx+0x3ca> + .byte 196,98,121,24,29,73,24,0,0 // vbroadcastss 0x1849(%rip),%xmm11 # 56f4 <_sk_callback_avx+0x3e2> .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 - .byte 196,98,121,24,45,39,24,0,0 // vbroadcastss 0x1827(%rip),%xmm13 # 56dc <_sk_callback_avx+0x3ce> + .byte 196,98,121,24,45,63,24,0,0 // vbroadcastss 0x183f(%rip),%xmm13 # 56f8 <_sk_callback_avx+0x3e6> .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -15727,7 +15739,7 @@ _sk_gather_f16_avx: .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 .byte 197,252,86,194 // vorps %ymm2,%ymm0,%ymm0 .byte 196,227,125,25,194,1 // vextractf128 $0x1,%ymm0,%xmm2 - .byte 196,226,121,24,29,221,23,0,0 // vbroadcastss 0x17dd(%rip),%xmm3 # 56e0 <_sk_callback_avx+0x3d2> + .byte 196,226,121,24,29,245,23,0,0 // vbroadcastss 0x17f5(%rip),%xmm3 # 56fc <_sk_callback_avx+0x3ea> .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 @@ -15831,12 +15843,12 @@ _sk_store_f16_avx: .byte 197,252,17,52,36 // vmovups %ymm6,(%rsp) .byte 197,252,17,108,36,224 // vmovups %ymm5,-0x20(%rsp) .byte 197,252,17,100,36,192 // vmovups %ymm4,-0x40(%rsp) - .byte 196,98,125,24,13,246,21,0,0 // vbroadcastss 0x15f6(%rip),%ymm9 # 56e4 <_sk_callback_avx+0x3d6> + .byte 196,98,125,24,13,14,22,0,0 // vbroadcastss 0x160e(%rip),%ymm9 # 5700 <_sk_callback_avx+0x3ee> .byte 196,65,124,84,209 // vandps %ymm9,%ymm0,%ymm10 .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp) .byte 196,65,124,87,218 // vxorps %ymm10,%ymm0,%ymm11 .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 - .byte 196,98,121,24,5,219,21,0,0 // vbroadcastss 0x15db(%rip),%xmm8 # 56e8 <_sk_callback_avx+0x3da> + .byte 196,98,121,24,5,243,21,0,0 // vbroadcastss 0x15f3(%rip),%xmm8 # 5704 <_sk_callback_avx+0x3f2> .byte 196,65,57,102,236 // vpcmpgtd %xmm12,%xmm8,%xmm13 .byte 196,65,57,102,243 // vpcmpgtd %xmm11,%xmm8,%xmm14 .byte 196,67,13,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 @@ -15846,7 +15858,7 @@ _sk_store_f16_avx: .byte 196,67,13,24,242,1 // vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11 .byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12 - .byte 196,98,125,24,21,162,21,0,0 // vbroadcastss 0x15a2(%rip),%ymm10 # 56ec <_sk_callback_avx+0x3de> + .byte 196,98,125,24,21,186,21,0,0 // vbroadcastss 0x15ba(%rip),%ymm10 # 5708 <_sk_callback_avx+0x3f6> .byte 196,65,12,86,242 // vorps %ymm10,%ymm14,%ymm14 .byte 196,67,125,25,247,1 // vextractf128 $0x1,%ymm14,%xmm15 .byte 196,65,1,254,228 // vpaddd %xmm12,%xmm15,%xmm12 @@ -15928,7 +15940,7 @@ _sk_store_f16_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 117,66 // jne 4328 <_sk_store_f16_avx+0x25e> + .byte 117,66 // jne 432c <_sk_store_f16_avx+0x25e> .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8) .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8) .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8) @@ -15944,22 +15956,22 @@ _sk_store_f16_avx: .byte 255,224 // jmpq *%rax .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,202 // je 42fd <_sk_store_f16_avx+0x233> + .byte 116,202 // je 4301 <_sk_store_f16_avx+0x233> .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,190 // jb 42fd <_sk_store_f16_avx+0x233> + .byte 114,190 // jb 4301 <_sk_store_f16_avx+0x233> .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) - .byte 116,182 // je 42fd <_sk_store_f16_avx+0x233> + .byte 116,182 // je 4301 <_sk_store_f16_avx+0x233> .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,170 // jb 42fd <_sk_store_f16_avx+0x233> + .byte 114,170 // jb 4301 <_sk_store_f16_avx+0x233> .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) - .byte 116,162 // je 42fd <_sk_store_f16_avx+0x233> + .byte 116,162 // je 4301 <_sk_store_f16_avx+0x233> .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,150 // jb 42fd <_sk_store_f16_avx+0x233> + .byte 114,150 // jb 4301 <_sk_store_f16_avx+0x233> .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) - .byte 235,142 // jmp 42fd <_sk_store_f16_avx+0x233> + .byte 235,142 // jmp 4301 <_sk_store_f16_avx+0x233> HIDDEN _sk_load_u16_be_avx .globl _sk_load_u16_be_avx @@ -15969,7 +15981,7 @@ _sk_load_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,253,0,0,0 // jne 4482 <_sk_load_u16_be_avx+0x113> + .byte 15,133,253,0,0,0 // jne 4486 <_sk_load_u16_be_avx+0x113> .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -15991,7 +16003,7 @@ _sk_load_u16_be_avx: .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,29,250,18,0,0 // vbroadcastss 0x12fa(%rip),%ymm11 # 56f0 <_sk_callback_avx+0x3e2> + .byte 196,98,125,24,29,18,19,0,0 // vbroadcastss 0x1312(%rip),%ymm11 # 570c <_sk_callback_avx+0x3fa> .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,177,109,202 // vpunpckhqdq %xmm2,%xmm9,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -16025,29 +16037,29 @@ _sk_load_u16_be_avx: .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 44e8 <_sk_load_u16_be_avx+0x179> + .byte 116,85 // je 44ec <_sk_load_u16_be_avx+0x179> .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 44e8 <_sk_load_u16_be_avx+0x179> + .byte 114,72 // jb 44ec <_sk_load_u16_be_avx+0x179> .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 44f5 <_sk_load_u16_be_avx+0x186> + .byte 116,72 // je 44f9 <_sk_load_u16_be_avx+0x186> .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 44f5 <_sk_load_u16_be_avx+0x186> + .byte 114,59 // jb 44f9 <_sk_load_u16_be_avx+0x186> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,213,254,255,255 // je 43a0 <_sk_load_u16_be_avx+0x31> + .byte 15,132,213,254,255,255 // je 43a4 <_sk_load_u16_be_avx+0x31> .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,196,254,255,255 // jb 43a0 <_sk_load_u16_be_avx+0x31> + .byte 15,130,196,254,255,255 // jb 43a4 <_sk_load_u16_be_avx+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,184,254,255,255 // jmpq 43a0 <_sk_load_u16_be_avx+0x31> + .byte 233,184,254,255,255 // jmpq 43a4 <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,171,254,255,255 // jmpq 43a0 <_sk_load_u16_be_avx+0x31> + .byte 233,171,254,255,255 // jmpq 43a4 <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,162,254,255,255 // jmpq 43a0 <_sk_load_u16_be_avx+0x31> + .byte 233,162,254,255,255 // jmpq 43a4 <_sk_load_u16_be_avx+0x31> HIDDEN _sk_load_rgb_u16_be_avx .globl _sk_load_rgb_u16_be_avx @@ -16057,7 +16069,7 @@ _sk_load_rgb_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,243,0,0,0 // jne 4603 <_sk_load_rgb_u16_be_avx+0x105> + .byte 15,133,243,0,0,0 // jne 4607 <_sk_load_rgb_u16_be_avx+0x105> .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -16084,7 +16096,7 @@ _sk_load_rgb_u16_be_avx: .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,29,90,17,0,0 // vbroadcastss 0x115a(%rip),%ymm11 # 56f4 <_sk_callback_avx+0x3e6> + .byte 196,98,125,24,29,114,17,0,0 // vbroadcastss 0x1172(%rip),%ymm11 # 5710 <_sk_callback_avx+0x3fe> .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -16105,41 +16117,41 @@ _sk_load_rgb_u16_be_avx: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,247,16,0,0 // vbroadcastss 0x10f7(%rip),%ymm3 # 56f8 <_sk_callback_avx+0x3ea> + .byte 196,226,125,24,29,15,17,0,0 // vbroadcastss 0x110f(%rip),%ymm3 # 5714 <_sk_callback_avx+0x402> .byte 255,224 // jmpq *%rax .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 461c <_sk_load_rgb_u16_be_avx+0x11e> - .byte 233,40,255,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,5 // jne 4620 <_sk_load_rgb_u16_be_avx+0x11e> + .byte 233,40,255,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 464b <_sk_load_rgb_u16_be_avx+0x14d> + .byte 114,26 // jb 464f <_sk_load_rgb_u16_be_avx+0x14d> .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 4650 <_sk_load_rgb_u16_be_avx+0x152> - .byte 233,249,254,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,244,254,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 4654 <_sk_load_rgb_u16_be_avx+0x152> + .byte 233,249,254,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,244,254,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 467f <_sk_load_rgb_u16_be_avx+0x181> + .byte 114,26 // jb 4683 <_sk_load_rgb_u16_be_avx+0x181> .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 4684 <_sk_load_rgb_u16_be_avx+0x186> - .byte 233,197,254,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,192,254,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 4688 <_sk_load_rgb_u16_be_avx+0x186> + .byte 233,197,254,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,192,254,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 46ad <_sk_load_rgb_u16_be_avx+0x1af> + .byte 114,20 // jb 46b1 <_sk_load_rgb_u16_be_avx+0x1af> .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - .byte 233,151,254,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,146,254,255,255 // jmpq 4544 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,151,254,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,146,254,255,255 // jmpq 4548 <_sk_load_rgb_u16_be_avx+0x46> HIDDEN _sk_store_u16_be_avx .globl _sk_store_u16_be_avx @@ -16148,7 +16160,7 @@ _sk_store_u16_be_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax - .byte 196,98,125,24,5,52,16,0,0 // vbroadcastss 0x1034(%rip),%ymm8 # 56fc <_sk_callback_avx+0x3ee> + .byte 196,98,125,24,5,76,16,0,0 // vbroadcastss 0x104c(%rip),%ymm8 # 5718 <_sk_callback_avx+0x406> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 @@ -16186,7 +16198,7 @@ _sk_store_u16_be_avx: .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,31 // jne 47ac <_sk_store_u16_be_avx+0xfa> + .byte 117,31 // jne 47b0 <_sk_store_u16_be_avx+0xfa> .byte 196,65,120,17,28,64 // vmovups %xmm11,(%r8,%rax,2) .byte 196,65,120,17,84,64,16 // vmovups %xmm10,0x10(%r8,%rax,2) .byte 196,65,120,17,76,64,32 // vmovups %xmm9,0x20(%r8,%rax,2) @@ -16195,22 +16207,22 @@ _sk_store_u16_be_avx: .byte 255,224 // jmpq *%rax .byte 196,65,121,214,28,64 // vmovq %xmm11,(%r8,%rax,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 116,240 // je 47ac <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,23,92,64,8 // vmovhpd %xmm11,0x8(%r8,%rax,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 114,227 // jb 47ac <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,214,84,64,16 // vmovq %xmm10,0x10(%r8,%rax,2) - .byte 116,218 // je 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 116,218 // je 47ac <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,23,84,64,24 // vmovhpd %xmm10,0x18(%r8,%rax,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 114,205 // jb 47ac <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,214,76,64,32 // vmovq %xmm9,0x20(%r8,%rax,2) - .byte 116,196 // je 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 116,196 // je 47ac <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,23,76,64,40 // vmovhpd %xmm9,0x28(%r8,%rax,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 114,183 // jb 47ac <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,214,68,64,48 // vmovq %xmm8,0x30(%r8,%rax,2) - .byte 235,174 // jmp 47a8 <_sk_store_u16_be_avx+0xf6> + .byte 235,174 // jmp 47ac <_sk_store_u16_be_avx+0xf6> HIDDEN _sk_load_f32_avx .globl _sk_load_f32_avx @@ -16218,10 +16230,10 @@ FUNCTION(_sk_load_f32_avx) _sk_load_f32_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 119,110 // ja 4870 <_sk_load_f32_avx+0x76> + .byte 119,110 // ja 4874 <_sk_load_f32_avx+0x76> .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 - .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 4898 <_sk_load_f32_avx+0x9e> + .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 489c <_sk_load_f32_avx+0x9e> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -16280,7 +16292,7 @@ _sk_store_f32_avx: .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 .byte 72,133,201 // test %rcx,%rcx - .byte 117,55 // jne 4925 <_sk_store_f32_avx+0x6d> + .byte 117,55 // jne 4929 <_sk_store_f32_avx+0x6d> .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -16293,22 +16305,22 @@ _sk_store_f32_avx: .byte 255,224 // jmpq *%rax .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 4921 <_sk_store_f32_avx+0x69> + .byte 116,240 // je 4925 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 4921 <_sk_store_f32_avx+0x69> + .byte 114,227 // jb 4925 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 4921 <_sk_store_f32_avx+0x69> + .byte 116,218 // je 4925 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 4921 <_sk_store_f32_avx+0x69> + .byte 114,205 // jb 4925 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - .byte 116,195 // je 4921 <_sk_store_f32_avx+0x69> + .byte 116,195 // je 4925 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,181 // jb 4921 <_sk_store_f32_avx+0x69> + .byte 114,181 // jb 4925 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - .byte 235,171 // jmp 4921 <_sk_store_f32_avx+0x69> + .byte 235,171 // jmp 4925 <_sk_store_f32_avx+0x69> HIDDEN _sk_clamp_x_avx .globl _sk_clamp_x_avx @@ -16444,12 +16456,12 @@ HIDDEN _sk_luminance_to_alpha_avx .globl _sk_luminance_to_alpha_avx FUNCTION(_sk_luminance_to_alpha_avx) _sk_luminance_to_alpha_avx: - .byte 196,226,125,24,29,191,11,0,0 // vbroadcastss 0xbbf(%rip),%ymm3 # 5700 <_sk_callback_avx+0x3f2> + .byte 196,226,125,24,29,215,11,0,0 // vbroadcastss 0xbd7(%rip),%ymm3 # 571c <_sk_callback_avx+0x40a> .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,29,182,11,0,0 // vbroadcastss 0xbb6(%rip),%ymm3 # 5704 <_sk_callback_avx+0x3f6> + .byte 196,226,125,24,29,206,11,0,0 // vbroadcastss 0xbce(%rip),%ymm3 # 5720 <_sk_callback_avx+0x40e> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,169,11,0,0 // vbroadcastss 0xba9(%rip),%ymm1 # 5708 <_sk_callback_avx+0x3fa> + .byte 196,226,125,24,13,193,11,0,0 // vbroadcastss 0xbc1(%rip),%ymm1 # 5724 <_sk_callback_avx+0x412> .byte 197,236,89,201 // vmulps %ymm1,%ymm2,%ymm1 .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -16627,7 +16639,7 @@ _sk_linear_gradient_avx: .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3 .byte 76,139,0 // mov (%rax),%r8 .byte 77,133,192 // test %r8,%r8 - .byte 15,132,146,0,0,0 // je 4eb5 <_sk_linear_gradient_avx+0xb8> + .byte 15,132,146,0,0,0 // je 4eb9 <_sk_linear_gradient_avx+0xb8> .byte 72,139,64,8 // mov 0x8(%rax),%rax .byte 72,131,192,32 // add $0x20,%rax .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 @@ -16654,8 +16666,8 @@ _sk_linear_gradient_avx: .byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3 .byte 72,131,192,36 // add $0x24,%rax .byte 73,255,200 // dec %r8 - .byte 117,140 // jne 4e3f <_sk_linear_gradient_avx+0x42> - .byte 235,20 // jmp 4ec9 <_sk_linear_gradient_avx+0xcc> + .byte 117,140 // jne 4e43 <_sk_linear_gradient_avx+0x42> + .byte 235,20 // jmp 4ecd <_sk_linear_gradient_avx+0xcc> .byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 @@ -16702,7 +16714,7 @@ HIDDEN _sk_save_xy_avx FUNCTION(_sk_save_xy_avx) _sk_save_xy_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,181,7,0,0 // vbroadcastss 0x7b5(%rip),%ymm8 # 570c <_sk_callback_avx+0x3fe> + .byte 196,98,125,24,5,205,7,0,0 // vbroadcastss 0x7cd(%rip),%ymm8 # 5728 <_sk_callback_avx+0x416> .byte 196,65,124,88,200 // vaddps %ymm8,%ymm0,%ymm9 .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 @@ -16739,9 +16751,9 @@ HIDDEN _sk_bilinear_nx_avx FUNCTION(_sk_bilinear_nx_avx) _sk_bilinear_nx_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,65,7,0,0 // vbroadcastss 0x741(%rip),%ymm0 # 5710 <_sk_callback_avx+0x402> + .byte 196,226,125,24,5,89,7,0,0 // vbroadcastss 0x759(%rip),%ymm0 # 572c <_sk_callback_avx+0x41a> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,56,7,0,0 // vbroadcastss 0x738(%rip),%ymm8 # 5714 <_sk_callback_avx+0x406> + .byte 196,98,125,24,5,80,7,0,0 // vbroadcastss 0x750(%rip),%ymm8 # 5730 <_sk_callback_avx+0x41e> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -16752,7 +16764,7 @@ HIDDEN _sk_bilinear_px_avx FUNCTION(_sk_bilinear_px_avx) _sk_bilinear_px_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,32,7,0,0 // vbroadcastss 0x720(%rip),%ymm0 # 5718 <_sk_callback_avx+0x40a> + .byte 196,226,125,24,5,56,7,0,0 // vbroadcastss 0x738(%rip),%ymm0 # 5734 <_sk_callback_avx+0x422> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -16764,9 +16776,9 @@ HIDDEN _sk_bilinear_ny_avx FUNCTION(_sk_bilinear_ny_avx) _sk_bilinear_ny_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,4,7,0,0 // vbroadcastss 0x704(%rip),%ymm1 # 571c <_sk_callback_avx+0x40e> + .byte 196,226,125,24,13,28,7,0,0 // vbroadcastss 0x71c(%rip),%ymm1 # 5738 <_sk_callback_avx+0x426> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,250,6,0,0 // vbroadcastss 0x6fa(%rip),%ymm8 # 5720 <_sk_callback_avx+0x412> + .byte 196,98,125,24,5,18,7,0,0 // vbroadcastss 0x712(%rip),%ymm8 # 573c <_sk_callback_avx+0x42a> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -16777,7 +16789,7 @@ HIDDEN _sk_bilinear_py_avx FUNCTION(_sk_bilinear_py_avx) _sk_bilinear_py_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,226,6,0,0 // vbroadcastss 0x6e2(%rip),%ymm1 # 5724 <_sk_callback_avx+0x416> + .byte 196,226,125,24,13,250,6,0,0 // vbroadcastss 0x6fa(%rip),%ymm1 # 5740 <_sk_callback_avx+0x42e> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -16789,14 +16801,14 @@ HIDDEN _sk_bicubic_n3x_avx FUNCTION(_sk_bicubic_n3x_avx) _sk_bicubic_n3x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,197,6,0,0 // vbroadcastss 0x6c5(%rip),%ymm0 # 5728 <_sk_callback_avx+0x41a> + .byte 196,226,125,24,5,221,6,0,0 // vbroadcastss 0x6dd(%rip),%ymm0 # 5744 <_sk_callback_avx+0x432> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,188,6,0,0 // vbroadcastss 0x6bc(%rip),%ymm8 # 572c <_sk_callback_avx+0x41e> + .byte 196,98,125,24,5,212,6,0,0 // vbroadcastss 0x6d4(%rip),%ymm8 # 5748 <_sk_callback_avx+0x436> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,173,6,0,0 // vbroadcastss 0x6ad(%rip),%ymm10 # 5730 <_sk_callback_avx+0x422> + .byte 196,98,125,24,21,197,6,0,0 // vbroadcastss 0x6c5(%rip),%ymm10 # 574c <_sk_callback_avx+0x43a> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,163,6,0,0 // vbroadcastss 0x6a3(%rip),%ymm10 # 5734 <_sk_callback_avx+0x426> + .byte 196,98,125,24,21,187,6,0,0 // vbroadcastss 0x6bb(%rip),%ymm10 # 5750 <_sk_callback_avx+0x43e> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -16808,19 +16820,19 @@ HIDDEN _sk_bicubic_n1x_avx FUNCTION(_sk_bicubic_n1x_avx) _sk_bicubic_n1x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,134,6,0,0 // vbroadcastss 0x686(%rip),%ymm0 # 5738 <_sk_callback_avx+0x42a> + .byte 196,226,125,24,5,158,6,0,0 // vbroadcastss 0x69e(%rip),%ymm0 # 5754 <_sk_callback_avx+0x442> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,125,6,0,0 // vbroadcastss 0x67d(%rip),%ymm8 # 573c <_sk_callback_avx+0x42e> + .byte 196,98,125,24,5,149,6,0,0 // vbroadcastss 0x695(%rip),%ymm8 # 5758 <_sk_callback_avx+0x446> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,115,6,0,0 // vbroadcastss 0x673(%rip),%ymm9 # 5740 <_sk_callback_avx+0x432> + .byte 196,98,125,24,13,139,6,0,0 // vbroadcastss 0x68b(%rip),%ymm9 # 575c <_sk_callback_avx+0x44a> .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,105,6,0,0 // vbroadcastss 0x669(%rip),%ymm10 # 5744 <_sk_callback_avx+0x436> + .byte 196,98,125,24,21,129,6,0,0 // vbroadcastss 0x681(%rip),%ymm10 # 5760 <_sk_callback_avx+0x44e> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,90,6,0,0 // vbroadcastss 0x65a(%rip),%ymm10 # 5748 <_sk_callback_avx+0x43a> + .byte 196,98,125,24,21,114,6,0,0 // vbroadcastss 0x672(%rip),%ymm10 # 5764 <_sk_callback_avx+0x452> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,75,6,0,0 // vbroadcastss 0x64b(%rip),%ymm9 # 574c <_sk_callback_avx+0x43e> + .byte 196,98,125,24,13,99,6,0,0 // vbroadcastss 0x663(%rip),%ymm9 # 5768 <_sk_callback_avx+0x456> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -16831,17 +16843,17 @@ HIDDEN _sk_bicubic_p1x_avx FUNCTION(_sk_bicubic_p1x_avx) _sk_bicubic_p1x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,51,6,0,0 // vbroadcastss 0x633(%rip),%ymm8 # 5750 <_sk_callback_avx+0x442> + .byte 196,98,125,24,5,75,6,0,0 // vbroadcastss 0x64b(%rip),%ymm8 # 576c <_sk_callback_avx+0x45a> .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 - .byte 196,98,125,24,21,37,6,0,0 // vbroadcastss 0x625(%rip),%ymm10 # 5754 <_sk_callback_avx+0x446> + .byte 196,98,125,24,21,61,6,0,0 // vbroadcastss 0x63d(%rip),%ymm10 # 5770 <_sk_callback_avx+0x45e> .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 - .byte 196,98,125,24,29,27,6,0,0 // vbroadcastss 0x61b(%rip),%ymm11 # 5758 <_sk_callback_avx+0x44a> + .byte 196,98,125,24,29,51,6,0,0 // vbroadcastss 0x633(%rip),%ymm11 # 5774 <_sk_callback_avx+0x462> .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 - .byte 196,98,125,24,13,2,6,0,0 // vbroadcastss 0x602(%rip),%ymm9 # 575c <_sk_callback_avx+0x44e> + .byte 196,98,125,24,13,26,6,0,0 // vbroadcastss 0x61a(%rip),%ymm9 # 5778 <_sk_callback_avx+0x466> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -16852,13 +16864,13 @@ HIDDEN _sk_bicubic_p3x_avx FUNCTION(_sk_bicubic_p3x_avx) _sk_bicubic_p3x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,234,5,0,0 // vbroadcastss 0x5ea(%rip),%ymm0 # 5760 <_sk_callback_avx+0x452> + .byte 196,226,125,24,5,2,6,0,0 // vbroadcastss 0x602(%rip),%ymm0 # 577c <_sk_callback_avx+0x46a> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,215,5,0,0 // vbroadcastss 0x5d7(%rip),%ymm10 # 5764 <_sk_callback_avx+0x456> + .byte 196,98,125,24,21,239,5,0,0 // vbroadcastss 0x5ef(%rip),%ymm10 # 5780 <_sk_callback_avx+0x46e> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,205,5,0,0 // vbroadcastss 0x5cd(%rip),%ymm10 # 5768 <_sk_callback_avx+0x45a> + .byte 196,98,125,24,21,229,5,0,0 // vbroadcastss 0x5e5(%rip),%ymm10 # 5784 <_sk_callback_avx+0x472> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -16870,14 +16882,14 @@ HIDDEN _sk_bicubic_n3y_avx FUNCTION(_sk_bicubic_n3y_avx) _sk_bicubic_n3y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,176,5,0,0 // vbroadcastss 0x5b0(%rip),%ymm1 # 576c <_sk_callback_avx+0x45e> + .byte 196,226,125,24,13,200,5,0,0 // vbroadcastss 0x5c8(%rip),%ymm1 # 5788 <_sk_callback_avx+0x476> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,166,5,0,0 // vbroadcastss 0x5a6(%rip),%ymm8 # 5770 <_sk_callback_avx+0x462> + .byte 196,98,125,24,5,190,5,0,0 // vbroadcastss 0x5be(%rip),%ymm8 # 578c <_sk_callback_avx+0x47a> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,151,5,0,0 // vbroadcastss 0x597(%rip),%ymm10 # 5774 <_sk_callback_avx+0x466> + .byte 196,98,125,24,21,175,5,0,0 // vbroadcastss 0x5af(%rip),%ymm10 # 5790 <_sk_callback_avx+0x47e> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,141,5,0,0 // vbroadcastss 0x58d(%rip),%ymm10 # 5778 <_sk_callback_avx+0x46a> + .byte 196,98,125,24,21,165,5,0,0 // vbroadcastss 0x5a5(%rip),%ymm10 # 5794 <_sk_callback_avx+0x482> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -16889,19 +16901,19 @@ HIDDEN _sk_bicubic_n1y_avx FUNCTION(_sk_bicubic_n1y_avx) _sk_bicubic_n1y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,112,5,0,0 // vbroadcastss 0x570(%rip),%ymm1 # 577c <_sk_callback_avx+0x46e> + .byte 196,226,125,24,13,136,5,0,0 // vbroadcastss 0x588(%rip),%ymm1 # 5798 <_sk_callback_avx+0x486> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,102,5,0,0 // vbroadcastss 0x566(%rip),%ymm8 # 5780 <_sk_callback_avx+0x472> + .byte 196,98,125,24,5,126,5,0,0 // vbroadcastss 0x57e(%rip),%ymm8 # 579c <_sk_callback_avx+0x48a> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,92,5,0,0 // vbroadcastss 0x55c(%rip),%ymm9 # 5784 <_sk_callback_avx+0x476> + .byte 196,98,125,24,13,116,5,0,0 // vbroadcastss 0x574(%rip),%ymm9 # 57a0 <_sk_callback_avx+0x48e> .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,82,5,0,0 // vbroadcastss 0x552(%rip),%ymm10 # 5788 <_sk_callback_avx+0x47a> + .byte 196,98,125,24,21,106,5,0,0 // vbroadcastss 0x56a(%rip),%ymm10 # 57a4 <_sk_callback_avx+0x492> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,67,5,0,0 // vbroadcastss 0x543(%rip),%ymm10 # 578c <_sk_callback_avx+0x47e> + .byte 196,98,125,24,21,91,5,0,0 // vbroadcastss 0x55b(%rip),%ymm10 # 57a8 <_sk_callback_avx+0x496> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,52,5,0,0 // vbroadcastss 0x534(%rip),%ymm9 # 5790 <_sk_callback_avx+0x482> + .byte 196,98,125,24,13,76,5,0,0 // vbroadcastss 0x54c(%rip),%ymm9 # 57ac <_sk_callback_avx+0x49a> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -16912,17 +16924,17 @@ HIDDEN _sk_bicubic_p1y_avx FUNCTION(_sk_bicubic_p1y_avx) _sk_bicubic_p1y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,28,5,0,0 // vbroadcastss 0x51c(%rip),%ymm8 # 5794 <_sk_callback_avx+0x486> + .byte 196,98,125,24,5,52,5,0,0 // vbroadcastss 0x534(%rip),%ymm8 # 57b0 <_sk_callback_avx+0x49e> .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 - .byte 196,98,125,24,21,13,5,0,0 // vbroadcastss 0x50d(%rip),%ymm10 # 5798 <_sk_callback_avx+0x48a> + .byte 196,98,125,24,21,37,5,0,0 // vbroadcastss 0x525(%rip),%ymm10 # 57b4 <_sk_callback_avx+0x4a2> .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 - .byte 196,98,125,24,29,3,5,0,0 // vbroadcastss 0x503(%rip),%ymm11 # 579c <_sk_callback_avx+0x48e> + .byte 196,98,125,24,29,27,5,0,0 // vbroadcastss 0x51b(%rip),%ymm11 # 57b8 <_sk_callback_avx+0x4a6> .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 - .byte 196,98,125,24,13,234,4,0,0 // vbroadcastss 0x4ea(%rip),%ymm9 # 57a0 <_sk_callback_avx+0x492> + .byte 196,98,125,24,13,2,5,0,0 // vbroadcastss 0x502(%rip),%ymm9 # 57bc <_sk_callback_avx+0x4aa> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -16933,13 +16945,13 @@ HIDDEN _sk_bicubic_p3y_avx FUNCTION(_sk_bicubic_p3y_avx) _sk_bicubic_p3y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,210,4,0,0 // vbroadcastss 0x4d2(%rip),%ymm1 # 57a4 <_sk_callback_avx+0x496> + .byte 196,226,125,24,13,234,4,0,0 // vbroadcastss 0x4ea(%rip),%ymm1 # 57c0 <_sk_callback_avx+0x4ae> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,190,4,0,0 // vbroadcastss 0x4be(%rip),%ymm10 # 57a8 <_sk_callback_avx+0x49a> + .byte 196,98,125,24,21,214,4,0,0 // vbroadcastss 0x4d6(%rip),%ymm10 # 57c4 <_sk_callback_avx+0x4b2> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,180,4,0,0 // vbroadcastss 0x4b4(%rip),%ymm10 # 57ac <_sk_callback_avx+0x49e> + .byte 196,98,125,24,21,204,4,0,0 // vbroadcastss 0x4cc(%rip),%ymm10 # 57c8 <_sk_callback_avx+0x4b6> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -17080,11 +17092,21 @@ BALIGN4 .byte 0,128,64,171,170,42 // add %al,0x2aaaab40(%rax) .byte 62,0,0 // add %al,%ds:(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 129,128,128,59,129,128,128,59,0,248 // addl $0xf8003b80,-0x7f7ec480(%rax) .byte 0,0 // add %al,(%rax) - .byte 8,33 // or %ah,(%rcx) + .byte 128,63,171 // cmpb $0xab,(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,0,0 // add %al,%ds:(%rax) + .byte 128,191,0,0,192,64,171 // cmpb $0xab,0x40c00000(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 190,129,128,128,59 // mov $0x3b808081,%esi + .byte 129,128,128,59,0,248,0,0,8,33 // addl $0x21080000,-0x7ffc480(%rax) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 54f1 <.literal4+0xbd> + .byte 224,7 // loopne 550d <.literal4+0xd5> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -17098,10 +17120,10 @@ BALIGN4 .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) .byte 0,52,255 // add %dh,(%rdi,%rdi,8) .byte 255 // (bad) - .byte 127,0 // jg 551c <.literal4+0xe8> + .byte 127,0 // jg 5538 <.literal4+0x100> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5595 <.literal4+0x161> + .byte 119,115 // ja 55b1 <.literal4+0x179> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -17115,10 +17137,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5550 <.literal4+0x11c> + .byte 127,0 // jg 556c <.literal4+0x134> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 55c9 <.literal4+0x195> + .byte 119,115 // ja 55e5 <.literal4+0x1ad> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -17132,10 +17154,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5584 <.literal4+0x150> + .byte 127,0 // jg 55a0 <.literal4+0x168> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 55fd <.literal4+0x1c9> + .byte 119,115 // ja 5619 <.literal4+0x1e1> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -17149,10 +17171,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 55b8 <.literal4+0x184> + .byte 127,0 // jg 55d4 <.literal4+0x19c> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5631 <.literal4+0x1fd> + .byte 119,115 // ja 564d <.literal4+0x215> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -17165,7 +17187,7 @@ BALIGN4 .byte 0,75,0 // add %cl,0x0(%rbx) .byte 0,128,63,0,0,200 // add %al,-0x37ffffc1(%rax) .byte 66,0,0 // rex.X add %al,(%rax) - .byte 127,67 // jg 562f <.literal4+0x1fb> + .byte 127,67 // jg 564b <.literal4+0x213> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -17177,10 +17199,10 @@ BALIGN4 .byte 190,80,128,3,62 // mov $0x3e038050,%esi .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 564f <.literal4+0x21b> + .byte 118,63 // jbe 566b <.literal4+0x233> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 5663 <.literal4+0x22f> + .byte 127,67 // jg 567f <.literal4+0x247> .byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax) .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax) @@ -17189,7 +17211,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 5645 <.literal4+0x211> + .byte 224,7 // loopne 5661 <.literal4+0x229> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -17201,7 +17223,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 5661 <.literal4+0x22d> + .byte 224,7 // loopne 567d <.literal4+0x245> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -17212,7 +17234,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 56b6 <.literal4+0x282> + .byte 124,66 // jl 56d2 <.literal4+0x29a> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax) @@ -17230,9 +17252,9 @@ BALIGN4 .byte 137,136,136,59,15,0 // mov %ecx,0xf3b88(%rax) .byte 0,0 // add %al,(%rax) .byte 137,136,136,61,0,0 // mov %ecx,0x3d88(%rax) - .byte 112,65 // jo 56f9 <.literal4+0x2c5> + .byte 112,65 // jo 5715 <.literal4+0x2dd> .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 5707 <.literal4+0x2d3> + .byte 127,67 // jg 5723 <.literal4+0x2eb> .byte 0,128,0,0,0,0 // add %al,0x0(%rax) .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax) .byte 0,0 // add %al,(%rax) @@ -17248,7 +17270,7 @@ BALIGN4 .byte 0,128,55,0,0,128 // add %al,-0x7fffffc9(%rax) .byte 63 // (bad) .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 5747 <.literal4+0x313> + .byte 127,71 // jg 5763 <.literal4+0x32b> .byte 208 // (bad) .byte 179,89 // mov $0x59,%bl .byte 62,89 // ds pop %rcx @@ -17464,7 +17486,7 @@ _sk_seed_shader_sse41: .byte 102,15,110,199 // movd %edi,%xmm0 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 - .byte 15,40,21,100,57,0,0 // movaps 0x3964(%rip),%xmm2 # 39e0 <_sk_callback_sse41+0xd5> + .byte 15,40,21,132,57,0,0 // movaps 0x3984(%rip),%xmm2 # 3a00 <_sk_callback_sse41+0xdd> .byte 15,88,202 // addps %xmm2,%xmm1 .byte 15,16,2 // movups (%rdx),%xmm0 .byte 15,88,193 // addps %xmm1,%xmm0 @@ -17473,7 +17495,7 @@ _sk_seed_shader_sse41: .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 15,88,202 // addps %xmm2,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,21,83,57,0,0 // movaps 0x3953(%rip),%xmm2 # 39f0 <_sk_callback_sse41+0xe5> + .byte 15,40,21,115,57,0,0 // movaps 0x3973(%rip),%xmm2 # 3a10 <_sk_callback_sse41+0xed> .byte 15,87,219 // xorps %xmm3,%xmm3 .byte 15,87,228 // xorps %xmm4,%xmm4 .byte 15,87,237 // xorps %xmm5,%xmm5 @@ -17513,7 +17535,7 @@ HIDDEN _sk_srcatop_sse41 FUNCTION(_sk_srcatop_sse41) _sk_srcatop_sse41: .byte 15,89,199 // mulps %xmm7,%xmm0 - .byte 68,15,40,5,14,57,0,0 // movaps 0x390e(%rip),%xmm8 # 3a00 <_sk_callback_sse41+0xf5> + .byte 68,15,40,5,46,57,0,0 // movaps 0x392e(%rip),%xmm8 # 3a20 <_sk_callback_sse41+0xfd> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -17538,7 +17560,7 @@ FUNCTION(_sk_dstatop_sse41) _sk_dstatop_sse41: .byte 68,15,40,195 // movaps %xmm3,%xmm8 .byte 68,15,89,196 // mulps %xmm4,%xmm8 - .byte 68,15,40,13,209,56,0,0 // movaps 0x38d1(%rip),%xmm9 # 3a10 <_sk_callback_sse41+0x105> + .byte 68,15,40,13,241,56,0,0 // movaps 0x38f1(%rip),%xmm9 # 3a30 <_sk_callback_sse41+0x10d> .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,88,192 // addps %xmm8,%xmm0 @@ -17585,7 +17607,7 @@ HIDDEN _sk_srcout_sse41 .globl _sk_srcout_sse41 FUNCTION(_sk_srcout_sse41) _sk_srcout_sse41: - .byte 68,15,40,5,117,56,0,0 // movaps 0x3875(%rip),%xmm8 # 3a20 <_sk_callback_sse41+0x115> + .byte 68,15,40,5,149,56,0,0 // movaps 0x3895(%rip),%xmm8 # 3a40 <_sk_callback_sse41+0x11d> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 @@ -17598,7 +17620,7 @@ HIDDEN _sk_dstout_sse41 .globl _sk_dstout_sse41 FUNCTION(_sk_dstout_sse41) _sk_dstout_sse41: - .byte 68,15,40,5,101,56,0,0 // movaps 0x3865(%rip),%xmm8 # 3a30 <_sk_callback_sse41+0x125> + .byte 68,15,40,5,133,56,0,0 // movaps 0x3885(%rip),%xmm8 # 3a50 <_sk_callback_sse41+0x12d> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 15,89,196 // mulps %xmm4,%xmm0 @@ -17615,7 +17637,7 @@ HIDDEN _sk_srcover_sse41 .globl _sk_srcover_sse41 FUNCTION(_sk_srcover_sse41) _sk_srcover_sse41: - .byte 68,15,40,5,72,56,0,0 // movaps 0x3848(%rip),%xmm8 # 3a40 <_sk_callback_sse41+0x135> + .byte 68,15,40,5,104,56,0,0 // movaps 0x3868(%rip),%xmm8 # 3a60 <_sk_callback_sse41+0x13d> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -17635,7 +17657,7 @@ HIDDEN _sk_dstover_sse41 .globl _sk_dstover_sse41 FUNCTION(_sk_dstover_sse41) _sk_dstover_sse41: - .byte 68,15,40,5,28,56,0,0 // movaps 0x381c(%rip),%xmm8 # 3a50 <_sk_callback_sse41+0x145> + .byte 68,15,40,5,60,56,0,0 // movaps 0x383c(%rip),%xmm8 # 3a70 <_sk_callback_sse41+0x14d> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -17663,7 +17685,7 @@ HIDDEN _sk_multiply_sse41 .globl _sk_multiply_sse41 FUNCTION(_sk_multiply_sse41) _sk_multiply_sse41: - .byte 68,15,40,5,240,55,0,0 // movaps 0x37f0(%rip),%xmm8 # 3a60 <_sk_callback_sse41+0x155> + .byte 68,15,40,5,16,56,0,0 // movaps 0x3810(%rip),%xmm8 # 3a80 <_sk_callback_sse41+0x15d> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 69,15,40,209 // movaps %xmm9,%xmm10 @@ -17739,7 +17761,7 @@ HIDDEN _sk_xor__sse41 FUNCTION(_sk_xor__sse41) _sk_xor__sse41: .byte 68,15,40,195 // movaps %xmm3,%xmm8 - .byte 15,40,29,33,55,0,0 // movaps 0x3721(%rip),%xmm3 # 3a70 <_sk_callback_sse41+0x165> + .byte 15,40,29,65,55,0,0 // movaps 0x3741(%rip),%xmm3 # 3a90 <_sk_callback_sse41+0x16d> .byte 68,15,40,203 // movaps %xmm3,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 @@ -17787,7 +17809,7 @@ _sk_darken_sse41: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,95,209 // maxps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,140,54,0,0 // movaps 0x368c(%rip),%xmm2 # 3a80 <_sk_callback_sse41+0x175> + .byte 15,40,21,172,54,0,0 // movaps 0x36ac(%rip),%xmm2 # 3aa0 <_sk_callback_sse41+0x17d> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -17821,7 +17843,7 @@ _sk_lighten_sse41: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,49,54,0,0 // movaps 0x3631(%rip),%xmm2 # 3a90 <_sk_callback_sse41+0x185> + .byte 15,40,21,81,54,0,0 // movaps 0x3651(%rip),%xmm2 # 3ab0 <_sk_callback_sse41+0x18d> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -17858,7 +17880,7 @@ _sk_difference_sse41: .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,203,53,0,0 // movaps 0x35cb(%rip),%xmm2 # 3aa0 <_sk_callback_sse41+0x195> + .byte 15,40,21,235,53,0,0 // movaps 0x35eb(%rip),%xmm2 # 3ac0 <_sk_callback_sse41+0x19d> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -17885,7 +17907,7 @@ _sk_exclusion_sse41: .byte 15,89,214 // mulps %xmm6,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,202 // subps %xmm2,%xmm9 - .byte 15,40,13,140,53,0,0 // movaps 0x358c(%rip),%xmm1 # 3ab0 <_sk_callback_sse41+0x1a5> + .byte 15,40,13,172,53,0,0 // movaps 0x35ac(%rip),%xmm1 # 3ad0 <_sk_callback_sse41+0x1ad> .byte 15,92,203 // subps %xmm3,%xmm1 .byte 15,89,207 // mulps %xmm7,%xmm1 .byte 15,88,217 // addps %xmm1,%xmm3 @@ -17899,7 +17921,7 @@ HIDDEN _sk_colorburn_sse41 FUNCTION(_sk_colorburn_sse41) _sk_colorburn_sse41: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,40,21,123,53,0,0 // movaps 0x357b(%rip),%xmm10 # 3ac0 <_sk_callback_sse41+0x1b5> + .byte 68,15,40,21,155,53,0,0 // movaps 0x359b(%rip),%xmm10 # 3ae0 <_sk_callback_sse41+0x1bd> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 69,15,40,203 // movaps %xmm11,%xmm9 @@ -17981,7 +18003,7 @@ HIDDEN _sk_colordodge_sse41 FUNCTION(_sk_colordodge_sse41) _sk_colordodge_sse41: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,40,21,89,52,0,0 // movaps 0x3459(%rip),%xmm10 # 3ad0 <_sk_callback_sse41+0x1c5> + .byte 68,15,40,21,121,52,0,0 // movaps 0x3479(%rip),%xmm10 # 3af0 <_sk_callback_sse41+0x1cd> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 69,15,40,227 // movaps %xmm11,%xmm12 @@ -18063,7 +18085,7 @@ _sk_hardlight_sse41: .byte 15,40,244 // movaps %xmm4,%xmm6 .byte 15,40,227 // movaps %xmm3,%xmm4 .byte 68,15,40,200 // movaps %xmm0,%xmm9 - .byte 68,15,40,21,50,51,0,0 // movaps 0x3332(%rip),%xmm10 # 3ae0 <_sk_callback_sse41+0x1d5> + .byte 68,15,40,21,82,51,0,0 // movaps 0x3352(%rip),%xmm10 # 3b00 <_sk_callback_sse41+0x1dd> .byte 65,15,40,234 // movaps %xmm10,%xmm5 .byte 15,92,239 // subps %xmm7,%xmm5 .byte 15,40,197 // movaps %xmm5,%xmm0 @@ -18146,7 +18168,7 @@ FUNCTION(_sk_overlay_sse41) _sk_overlay_sse41: .byte 68,15,40,201 // movaps %xmm1,%xmm9 .byte 68,15,40,240 // movaps %xmm0,%xmm14 - .byte 68,15,40,21,23,50,0,0 // movaps 0x3217(%rip),%xmm10 # 3af0 <_sk_callback_sse41+0x1e5> + .byte 68,15,40,21,55,50,0,0 // movaps 0x3237(%rip),%xmm10 # 3b10 <_sk_callback_sse41+0x1ed> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 65,15,40,195 // movaps %xmm11,%xmm0 @@ -18231,7 +18253,7 @@ _sk_softlight_sse41: .byte 15,40,198 // movaps %xmm6,%xmm0 .byte 15,94,199 // divps %xmm7,%xmm0 .byte 65,15,84,193 // andps %xmm9,%xmm0 - .byte 15,40,13,238,48,0,0 // movaps 0x30ee(%rip),%xmm1 # 3b00 <_sk_callback_sse41+0x1f5> + .byte 15,40,13,14,49,0,0 // movaps 0x310e(%rip),%xmm1 # 3b20 <_sk_callback_sse41+0x1fd> .byte 68,15,40,209 // movaps %xmm1,%xmm10 .byte 68,15,92,208 // subps %xmm0,%xmm10 .byte 68,15,40,240 // movaps %xmm0,%xmm14 @@ -18244,10 +18266,10 @@ _sk_softlight_sse41: .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 15,89,210 // mulps %xmm2,%xmm2 .byte 15,88,208 // addps %xmm0,%xmm2 - .byte 68,15,40,45,204,48,0,0 // movaps 0x30cc(%rip),%xmm13 # 3b10 <_sk_callback_sse41+0x205> + .byte 68,15,40,45,236,48,0,0 // movaps 0x30ec(%rip),%xmm13 # 3b30 <_sk_callback_sse41+0x20d> .byte 69,15,88,245 // addps %xmm13,%xmm14 .byte 68,15,89,242 // mulps %xmm2,%xmm14 - .byte 68,15,40,37,204,48,0,0 // movaps 0x30cc(%rip),%xmm12 # 3b20 <_sk_callback_sse41+0x215> + .byte 68,15,40,37,236,48,0,0 // movaps 0x30ec(%rip),%xmm12 # 3b40 <_sk_callback_sse41+0x21d> .byte 69,15,89,252 // mulps %xmm12,%xmm15 .byte 69,15,88,254 // addps %xmm14,%xmm15 .byte 15,40,198 // movaps %xmm6,%xmm0 @@ -18395,7 +18417,7 @@ HIDDEN _sk_clamp_1_sse41 .globl _sk_clamp_1_sse41 FUNCTION(_sk_clamp_1_sse41) _sk_clamp_1_sse41: - .byte 68,15,40,5,222,46,0,0 // movaps 0x2ede(%rip),%xmm8 # 3b30 <_sk_callback_sse41+0x225> + .byte 68,15,40,5,254,46,0,0 // movaps 0x2efe(%rip),%xmm8 # 3b50 <_sk_callback_sse41+0x22d> .byte 65,15,93,192 // minps %xmm8,%xmm0 .byte 65,15,93,200 // minps %xmm8,%xmm1 .byte 65,15,93,208 // minps %xmm8,%xmm2 @@ -18407,7 +18429,7 @@ HIDDEN _sk_clamp_a_sse41 .globl _sk_clamp_a_sse41 FUNCTION(_sk_clamp_a_sse41) _sk_clamp_a_sse41: - .byte 15,93,29,211,46,0,0 // minps 0x2ed3(%rip),%xmm3 # 3b40 <_sk_callback_sse41+0x235> + .byte 15,93,29,243,46,0,0 // minps 0x2ef3(%rip),%xmm3 # 3b60 <_sk_callback_sse41+0x23d> .byte 15,93,195 // minps %xmm3,%xmm0 .byte 15,93,203 // minps %xmm3,%xmm1 .byte 15,93,211 // minps %xmm3,%xmm2 @@ -18494,7 +18516,7 @@ HIDDEN _sk_unpremul_sse41 FUNCTION(_sk_unpremul_sse41) _sk_unpremul_sse41: .byte 69,15,87,192 // xorps %xmm8,%xmm8 - .byte 68,15,40,13,62,46,0,0 // movaps 0x2e3e(%rip),%xmm9 # 3b50 <_sk_callback_sse41+0x245> + .byte 68,15,40,13,94,46,0,0 // movaps 0x2e5e(%rip),%xmm9 # 3b70 <_sk_callback_sse41+0x24d> .byte 68,15,94,203 // divps %xmm3,%xmm9 .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8 .byte 69,15,84,193 // andps %xmm9,%xmm8 @@ -18508,20 +18530,20 @@ HIDDEN _sk_from_srgb_sse41 .globl _sk_from_srgb_sse41 FUNCTION(_sk_from_srgb_sse41) _sk_from_srgb_sse41: - .byte 68,15,40,29,41,46,0,0 // movaps 0x2e29(%rip),%xmm11 # 3b60 <_sk_callback_sse41+0x255> + .byte 68,15,40,29,73,46,0,0 // movaps 0x2e49(%rip),%xmm11 # 3b80 <_sk_callback_sse41+0x25d> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 .byte 68,15,40,208 // movaps %xmm0,%xmm10 .byte 69,15,89,210 // mulps %xmm10,%xmm10 - .byte 68,15,40,37,33,46,0,0 // movaps 0x2e21(%rip),%xmm12 # 3b70 <_sk_callback_sse41+0x265> + .byte 68,15,40,37,65,46,0,0 // movaps 0x2e41(%rip),%xmm12 # 3b90 <_sk_callback_sse41+0x26d> .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 69,15,89,196 // mulps %xmm12,%xmm8 - .byte 68,15,40,45,33,46,0,0 // movaps 0x2e21(%rip),%xmm13 # 3b80 <_sk_callback_sse41+0x275> + .byte 68,15,40,45,65,46,0,0 // movaps 0x2e41(%rip),%xmm13 # 3ba0 <_sk_callback_sse41+0x27d> .byte 69,15,88,197 // addps %xmm13,%xmm8 .byte 69,15,89,194 // mulps %xmm10,%xmm8 - .byte 68,15,40,53,33,46,0,0 // movaps 0x2e21(%rip),%xmm14 # 3b90 <_sk_callback_sse41+0x285> + .byte 68,15,40,53,65,46,0,0 // movaps 0x2e41(%rip),%xmm14 # 3bb0 <_sk_callback_sse41+0x28d> .byte 69,15,88,198 // addps %xmm14,%xmm8 - .byte 68,15,40,61,37,46,0,0 // movaps 0x2e25(%rip),%xmm15 # 3ba0 <_sk_callback_sse41+0x295> + .byte 68,15,40,61,69,46,0,0 // movaps 0x2e45(%rip),%xmm15 # 3bc0 <_sk_callback_sse41+0x29d> .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 .byte 102,69,15,56,20,193 // blendvps %xmm0,%xmm9,%xmm8 .byte 68,15,40,209 // movaps %xmm1,%xmm10 @@ -18566,20 +18588,20 @@ _sk_to_srgb_sse41: .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 .byte 69,15,83,200 // rcpps %xmm8,%xmm9 .byte 69,15,82,208 // rsqrtps %xmm8,%xmm10 - .byte 68,15,40,29,149,45,0,0 // movaps 0x2d95(%rip),%xmm11 # 3bb0 <_sk_callback_sse41+0x2a5> + .byte 68,15,40,29,181,45,0,0 // movaps 0x2db5(%rip),%xmm11 # 3bd0 <_sk_callback_sse41+0x2ad> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 65,15,89,203 // mulps %xmm11,%xmm1 - .byte 68,15,40,37,150,45,0,0 // movaps 0x2d96(%rip),%xmm12 # 3bc0 <_sk_callback_sse41+0x2b5> + .byte 68,15,40,37,182,45,0,0 // movaps 0x2db6(%rip),%xmm12 # 3be0 <_sk_callback_sse41+0x2bd> .byte 69,15,89,204 // mulps %xmm12,%xmm9 - .byte 68,15,40,45,154,45,0,0 // movaps 0x2d9a(%rip),%xmm13 # 3bd0 <_sk_callback_sse41+0x2c5> + .byte 68,15,40,45,186,45,0,0 // movaps 0x2dba(%rip),%xmm13 # 3bf0 <_sk_callback_sse41+0x2cd> .byte 69,15,88,205 // addps %xmm13,%xmm9 - .byte 68,15,40,53,158,45,0,0 // movaps 0x2d9e(%rip),%xmm14 # 3be0 <_sk_callback_sse41+0x2d5> + .byte 68,15,40,53,190,45,0,0 // movaps 0x2dbe(%rip),%xmm14 # 3c00 <_sk_callback_sse41+0x2dd> .byte 69,15,89,214 // mulps %xmm14,%xmm10 .byte 69,15,88,209 // addps %xmm9,%xmm10 - .byte 68,15,40,5,158,45,0,0 // movaps 0x2d9e(%rip),%xmm8 # 3bf0 <_sk_callback_sse41+0x2e5> + .byte 68,15,40,5,190,45,0,0 // movaps 0x2dbe(%rip),%xmm8 # 3c10 <_sk_callback_sse41+0x2ed> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,93,202 // minps %xmm10,%xmm9 - .byte 68,15,40,61,158,45,0,0 // movaps 0x2d9e(%rip),%xmm15 # 3c00 <_sk_callback_sse41+0x2f5> + .byte 68,15,40,61,190,45,0,0 // movaps 0x2dbe(%rip),%xmm15 # 3c20 <_sk_callback_sse41+0x2fd> .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 .byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9 .byte 15,82,194 // rsqrtps %xmm2,%xmm0 @@ -18633,7 +18655,7 @@ _sk_rgb_to_hsl_sse41: .byte 68,15,93,226 // minps %xmm2,%xmm12 .byte 65,15,40,203 // movaps %xmm11,%xmm1 .byte 65,15,92,204 // subps %xmm12,%xmm1 - .byte 68,15,40,53,239,44,0,0 // movaps 0x2cef(%rip),%xmm14 # 3c10 <_sk_callback_sse41+0x305> + .byte 68,15,40,53,15,45,0,0 // movaps 0x2d0f(%rip),%xmm14 # 3c30 <_sk_callback_sse41+0x30d> .byte 68,15,94,241 // divps %xmm1,%xmm14 .byte 69,15,40,211 // movaps %xmm11,%xmm10 .byte 69,15,194,208,0 // cmpeqps %xmm8,%xmm10 @@ -18642,27 +18664,27 @@ _sk_rgb_to_hsl_sse41: .byte 65,15,89,198 // mulps %xmm14,%xmm0 .byte 69,15,40,249 // movaps %xmm9,%xmm15 .byte 68,15,194,250,1 // cmpltps %xmm2,%xmm15 - .byte 68,15,84,61,214,44,0,0 // andps 0x2cd6(%rip),%xmm15 # 3c20 <_sk_callback_sse41+0x315> + .byte 68,15,84,61,246,44,0,0 // andps 0x2cf6(%rip),%xmm15 # 3c40 <_sk_callback_sse41+0x31d> .byte 68,15,88,248 // addps %xmm0,%xmm15 .byte 65,15,40,195 // movaps %xmm11,%xmm0 .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0 .byte 65,15,92,208 // subps %xmm8,%xmm2 .byte 65,15,89,214 // mulps %xmm14,%xmm2 - .byte 68,15,40,45,201,44,0,0 // movaps 0x2cc9(%rip),%xmm13 # 3c30 <_sk_callback_sse41+0x325> + .byte 68,15,40,45,233,44,0,0 // movaps 0x2ce9(%rip),%xmm13 # 3c50 <_sk_callback_sse41+0x32d> .byte 65,15,88,213 // addps %xmm13,%xmm2 .byte 69,15,92,193 // subps %xmm9,%xmm8 .byte 69,15,89,198 // mulps %xmm14,%xmm8 - .byte 68,15,88,5,197,44,0,0 // addps 0x2cc5(%rip),%xmm8 # 3c40 <_sk_callback_sse41+0x335> + .byte 68,15,88,5,229,44,0,0 // addps 0x2ce5(%rip),%xmm8 # 3c60 <_sk_callback_sse41+0x33d> .byte 102,68,15,56,20,194 // blendvps %xmm0,%xmm2,%xmm8 .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 102,69,15,56,20,199 // blendvps %xmm0,%xmm15,%xmm8 - .byte 68,15,89,5,189,44,0,0 // mulps 0x2cbd(%rip),%xmm8 # 3c50 <_sk_callback_sse41+0x345> + .byte 68,15,89,5,221,44,0,0 // mulps 0x2cdd(%rip),%xmm8 # 3c70 <_sk_callback_sse41+0x34d> .byte 69,15,40,203 // movaps %xmm11,%xmm9 .byte 69,15,194,204,4 // cmpneqps %xmm12,%xmm9 .byte 69,15,84,193 // andps %xmm9,%xmm8 .byte 69,15,92,235 // subps %xmm11,%xmm13 .byte 69,15,88,220 // addps %xmm12,%xmm11 - .byte 15,40,5,177,44,0,0 // movaps 0x2cb1(%rip),%xmm0 # 3c60 <_sk_callback_sse41+0x355> + .byte 15,40,5,209,44,0,0 // movaps 0x2cd1(%rip),%xmm0 # 3c80 <_sk_callback_sse41+0x35d> .byte 65,15,40,211 // movaps %xmm11,%xmm2 .byte 15,89,208 // mulps %xmm0,%xmm2 .byte 15,194,194,1 // cmpltps %xmm2,%xmm0 @@ -18684,152 +18706,150 @@ _sk_hsl_to_rgb_sse41: .byte 15,41,108,36,224 // movaps %xmm5,-0x20(%rsp) .byte 15,41,100,36,208 // movaps %xmm4,-0x30(%rsp) .byte 15,41,92,36,192 // movaps %xmm3,-0x40(%rsp) - .byte 68,15,40,208 // movaps %xmm0,%xmm10 + .byte 68,15,40,216 // movaps %xmm0,%xmm11 .byte 184,0,0,0,63 // mov $0x3f000000,%eax .byte 102,15,110,216 // movd %eax,%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 15,41,92,36,128 // movaps %xmm3,-0x80(%rsp) .byte 15,40,194 // movaps %xmm2,%xmm0 .byte 15,194,195,1 // cmpltps %xmm3,%xmm0 - .byte 184,0,0,128,63 // mov $0x3f800000,%eax - .byte 102,68,15,110,192 // movd %eax,%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 65,15,40,224 // movaps %xmm8,%xmm4 - .byte 15,88,225 // addps %xmm1,%xmm4 - .byte 15,89,226 // mulps %xmm2,%xmm4 + .byte 15,40,45,124,44,0,0 // movaps 0x2c7c(%rip),%xmm5 # 3c90 <_sk_callback_sse41+0x36d> + .byte 15,40,241 // movaps %xmm1,%xmm6 + .byte 15,40,225 // movaps %xmm1,%xmm4 .byte 15,40,217 // movaps %xmm1,%xmm3 - .byte 15,40,249 // movaps %xmm1,%xmm7 - .byte 15,88,250 // addps %xmm2,%xmm7 + .byte 15,88,221 // addps %xmm5,%xmm3 + .byte 15,40,253 // movaps %xmm5,%xmm7 .byte 15,89,218 // mulps %xmm2,%xmm3 + .byte 15,88,242 // addps %xmm2,%xmm6 + .byte 15,89,226 // mulps %xmm2,%xmm4 .byte 15,40,234 // movaps %xmm2,%xmm5 - .byte 15,92,251 // subps %xmm3,%xmm7 - .byte 102,15,56,20,252 // blendvps %xmm0,%xmm4,%xmm7 - .byte 184,0,0,0,64 // mov $0x40000000,%eax - .byte 185,171,170,170,62 // mov $0x3eaaaaab,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 + .byte 15,92,244 // subps %xmm4,%xmm6 + .byte 102,15,56,20,243 // blendvps %xmm0,%xmm3,%xmm6 + .byte 68,15,40,61,97,44,0,0 // movaps 0x2c61(%rip),%xmm15 # 3ca0 <_sk_callback_sse41+0x37d> + .byte 69,15,88,251 // addps %xmm11,%xmm15 + .byte 184,0,0,0,0 // mov $0x0,%eax + .byte 185,0,0,128,63 // mov $0x3f800000,%ecx + .byte 102,68,15,110,201 // movd %ecx,%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 65,15,40,193 // movaps %xmm9,%xmm0 + .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 + .byte 65,15,40,215 // movaps %xmm15,%xmm2 + .byte 15,88,21,69,44,0,0 // addps 0x2c45(%rip),%xmm2 # 3cb0 <_sk_callback_sse41+0x38d> + .byte 69,15,40,231 // movaps %xmm15,%xmm12 + .byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12 + .byte 102,15,110,208 // movd %eax,%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 15,41,84,36,176 // movaps %xmm2,-0x50(%rsp) - .byte 65,15,88,210 // addps %xmm10,%xmm2 - .byte 185,0,0,0,0 // mov $0x0,%ecx - .byte 65,15,40,192 // movaps %xmm8,%xmm0 + .byte 15,41,84,36,160 // movaps %xmm2,-0x60(%rsp) + .byte 65,15,40,199 // movaps %xmm15,%xmm0 .byte 15,194,194,1 // cmpltps %xmm2,%xmm0 - .byte 15,40,218 // movaps %xmm2,%xmm3 - .byte 65,15,92,216 // subps %xmm8,%xmm3 - .byte 68,15,40,226 // movaps %xmm2,%xmm12 - .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12 - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 15,40,194 // movaps %xmm2,%xmm0 - .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 - .byte 68,15,41,116,36,160 // movaps %xmm14,-0x60(%rsp) - .byte 65,15,40,216 // movaps %xmm8,%xmm3 - .byte 15,88,218 // addps %xmm2,%xmm3 - .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12 - .byte 102,68,15,110,200 // movd %eax,%xmm9 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 15,40,229 // movaps %xmm5,%xmm4 - .byte 15,41,100,36,144 // movaps %xmm4,-0x70(%rsp) - .byte 68,15,89,204 // mulps %xmm4,%xmm9 - .byte 68,15,92,207 // subps %xmm7,%xmm9 + .byte 65,15,40,215 // movaps %xmm15,%xmm2 + .byte 15,88,215 // addps %xmm7,%xmm2 + .byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12 + .byte 15,40,221 // movaps %xmm5,%xmm3 + .byte 15,41,92,36,176 // movaps %xmm3,-0x50(%rsp) + .byte 68,15,40,195 // movaps %xmm3,%xmm8 + .byte 69,15,88,192 // addps %xmm8,%xmm8 + .byte 68,15,92,198 // subps %xmm6,%xmm8 .byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax - .byte 15,40,199 // movaps %xmm7,%xmm0 - .byte 65,15,92,193 // subps %xmm9,%xmm0 - .byte 185,0,0,192,64 // mov $0x40c00000,%ecx - .byte 102,15,110,241 // movd %ecx,%xmm6 - .byte 15,198,246,0 // shufps $0x0,%xmm6,%xmm6 - .byte 15,89,240 // mulps %xmm0,%xmm6 + .byte 15,40,214 // movaps %xmm6,%xmm2 + .byte 65,15,92,208 // subps %xmm8,%xmm2 + .byte 15,89,21,2,44,0,0 // mulps 0x2c02(%rip),%xmm2 # 3cc0 <_sk_callback_sse41+0x39d> .byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 15,40,235 // movaps %xmm3,%xmm5 + .byte 102,15,110,193 // movd %ecx,%xmm0 + .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 + .byte 15,41,68,36,144 // movaps %xmm0,-0x70(%rsp) + .byte 15,40,37,249,43,0,0 // movaps 0x2bf9(%rip),%xmm4 # 3cd0 <_sk_callback_sse41+0x3ad> + .byte 15,40,236 // movaps %xmm4,%xmm5 .byte 65,15,92,236 // subps %xmm12,%xmm5 .byte 69,15,40,236 // movaps %xmm12,%xmm13 - .byte 69,15,40,252 // movaps %xmm12,%xmm15 - .byte 68,15,194,227,1 // cmpltps %xmm3,%xmm12 - .byte 15,89,238 // mulps %xmm6,%xmm5 - .byte 65,15,88,233 // addps %xmm9,%xmm5 - .byte 69,15,40,217 // movaps %xmm9,%xmm11 + .byte 69,15,40,244 // movaps %xmm12,%xmm14 + .byte 68,15,194,224,1 // cmpltps %xmm0,%xmm12 + .byte 15,89,234 // mulps %xmm2,%xmm5 + .byte 65,15,88,232 // addps %xmm8,%xmm5 + .byte 69,15,40,208 // movaps %xmm8,%xmm10 .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 102,68,15,56,20,221 // blendvps %xmm0,%xmm5,%xmm11 - .byte 68,15,194,124,36,128,1 // cmpltps -0x80(%rsp),%xmm15 - .byte 65,15,40,199 // movaps %xmm15,%xmm0 - .byte 102,68,15,56,20,223 // blendvps %xmm0,%xmm7,%xmm11 - .byte 102,15,110,232 // movd %eax,%xmm5 - .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5 - .byte 68,15,194,237,1 // cmpltps %xmm5,%xmm13 - .byte 15,89,214 // mulps %xmm6,%xmm2 - .byte 65,15,88,209 // addps %xmm9,%xmm2 + .byte 102,68,15,56,20,213 // blendvps %xmm0,%xmm5,%xmm10 + .byte 15,40,108,36,128 // movaps -0x80(%rsp),%xmm5 + .byte 68,15,194,245,1 // cmpltps %xmm5,%xmm14 + .byte 65,15,40,198 // movaps %xmm14,%xmm0 + .byte 102,68,15,56,20,214 // blendvps %xmm0,%xmm6,%xmm10 + .byte 102,15,110,248 // movd %eax,%xmm7 + .byte 15,198,255,0 // shufps $0x0,%xmm7,%xmm7 + .byte 68,15,194,239,1 // cmpltps %xmm7,%xmm13 + .byte 68,15,89,250 // mulps %xmm2,%xmm15 + .byte 69,15,88,248 // addps %xmm8,%xmm15 .byte 65,15,40,197 // movaps %xmm13,%xmm0 - .byte 102,68,15,56,20,218 // blendvps %xmm0,%xmm2,%xmm11 + .byte 102,69,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm10 .byte 69,15,87,228 // xorps %xmm12,%xmm12 .byte 68,15,194,225,0 // cmpeqps %xmm1,%xmm12 .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 102,68,15,56,20,220 // blendvps %xmm0,%xmm4,%xmm11 - .byte 65,15,40,192 // movaps %xmm8,%xmm0 - .byte 65,15,194,194,1 // cmpltps %xmm10,%xmm0 - .byte 65,15,40,202 // movaps %xmm10,%xmm1 - .byte 65,15,92,200 // subps %xmm8,%xmm1 - .byte 69,15,40,234 // movaps %xmm10,%xmm13 + .byte 102,68,15,56,20,211 // blendvps %xmm0,%xmm3,%xmm10 + .byte 65,15,40,193 // movaps %xmm9,%xmm0 + .byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0 + .byte 65,15,40,203 // movaps %xmm11,%xmm1 + .byte 15,88,13,86,43,0,0 // addps 0x2b56(%rip),%xmm1 # 3cb0 <_sk_callback_sse41+0x38d> + .byte 69,15,40,235 // movaps %xmm11,%xmm13 .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13 - .byte 65,15,40,194 // movaps %xmm10,%xmm0 - .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 - .byte 65,15,40,200 // movaps %xmm8,%xmm1 - .byte 65,15,88,202 // addps %xmm10,%xmm1 + .byte 65,15,40,195 // movaps %xmm11,%xmm0 + .byte 15,194,68,36,160,1 // cmpltps -0x60(%rsp),%xmm0 + .byte 65,15,40,203 // movaps %xmm11,%xmm1 + .byte 15,88,13,23,43,0,0 // addps 0x2b17(%rip),%xmm1 # 3c90 <_sk_callback_sse41+0x36d> .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13 - .byte 68,15,40,243 // movaps %xmm3,%xmm14 - .byte 69,15,92,245 // subps %xmm13,%xmm14 - .byte 65,15,40,229 // movaps %xmm13,%xmm4 + .byte 15,40,220 // movaps %xmm4,%xmm3 + .byte 65,15,92,221 // subps %xmm13,%xmm3 + .byte 69,15,40,245 // movaps %xmm13,%xmm14 .byte 69,15,40,253 // movaps %xmm13,%xmm15 - .byte 68,15,194,235,1 // cmpltps %xmm3,%xmm13 - .byte 68,15,89,246 // mulps %xmm6,%xmm14 - .byte 69,15,88,241 // addps %xmm9,%xmm14 - .byte 65,15,40,201 // movaps %xmm9,%xmm1 + .byte 68,15,194,108,36,144,1 // cmpltps -0x70(%rsp),%xmm13 + .byte 15,89,218 // mulps %xmm2,%xmm3 + .byte 65,15,88,216 // addps %xmm8,%xmm3 + .byte 65,15,40,200 // movaps %xmm8,%xmm1 .byte 65,15,40,197 // movaps %xmm13,%xmm0 - .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1 - .byte 68,15,40,116,36,128 // movaps -0x80(%rsp),%xmm14 - .byte 69,15,194,254,1 // cmpltps %xmm14,%xmm15 + .byte 102,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm1 + .byte 68,15,194,253,1 // cmpltps %xmm5,%xmm15 .byte 65,15,40,199 // movaps %xmm15,%xmm0 - .byte 102,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm1 - .byte 15,194,229,1 // cmpltps %xmm5,%xmm4 - .byte 15,40,214 // movaps %xmm6,%xmm2 - .byte 65,15,89,210 // mulps %xmm10,%xmm2 - .byte 65,15,88,209 // addps %xmm9,%xmm2 - .byte 15,40,196 // movaps %xmm4,%xmm0 - .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1 + .byte 102,15,56,20,206 // blendvps %xmm0,%xmm6,%xmm1 + .byte 68,15,194,247,1 // cmpltps %xmm7,%xmm14 + .byte 15,40,218 // movaps %xmm2,%xmm3 + .byte 65,15,89,219 // mulps %xmm11,%xmm3 + .byte 65,15,88,216 // addps %xmm8,%xmm3 + .byte 65,15,40,198 // movaps %xmm14,%xmm0 + .byte 102,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm1 .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 68,15,40,124,36,144 // movaps -0x70(%rsp),%xmm15 - .byte 102,65,15,56,20,207 // blendvps %xmm0,%xmm15,%xmm1 - .byte 68,15,92,84,36,176 // subps -0x50(%rsp),%xmm10 - .byte 65,15,40,210 // movaps %xmm10,%xmm2 - .byte 65,15,92,208 // subps %xmm8,%xmm2 - .byte 69,15,40,232 // movaps %xmm8,%xmm13 - .byte 69,15,194,194,1 // cmpltps %xmm10,%xmm8 - .byte 65,15,40,226 // movaps %xmm10,%xmm4 - .byte 65,15,40,192 // movaps %xmm8,%xmm0 - .byte 102,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm4 - .byte 65,15,40,194 // movaps %xmm10,%xmm0 + .byte 15,40,108,36,176 // movaps -0x50(%rsp),%xmm5 + .byte 102,15,56,20,205 // blendvps %xmm0,%xmm5,%xmm1 + .byte 68,15,88,29,250,42,0,0 // addps 0x2afa(%rip),%xmm11 # 3ce0 <_sk_callback_sse41+0x3bd> + .byte 69,15,194,203,1 // cmpltps %xmm11,%xmm9 + .byte 15,40,29,190,42,0,0 // movaps 0x2abe(%rip),%xmm3 # 3cb0 <_sk_callback_sse41+0x38d> + .byte 65,15,88,219 // addps %xmm11,%xmm3 + .byte 69,15,40,235 // movaps %xmm11,%xmm13 + .byte 65,15,40,193 // movaps %xmm9,%xmm0 + .byte 102,68,15,56,20,235 // blendvps %xmm0,%xmm3,%xmm13 + .byte 65,15,40,195 // movaps %xmm11,%xmm0 .byte 15,194,68,36,160,1 // cmpltps -0x60(%rsp),%xmm0 - .byte 69,15,88,234 // addps %xmm10,%xmm13 - .byte 102,65,15,56,20,229 // blendvps %xmm0,%xmm13,%xmm4 - .byte 68,15,89,214 // mulps %xmm6,%xmm10 - .byte 15,40,196 // movaps %xmm4,%xmm0 - .byte 15,194,195,1 // cmpltps %xmm3,%xmm0 - .byte 15,92,220 // subps %xmm4,%xmm3 - .byte 15,89,222 // mulps %xmm6,%xmm3 - .byte 69,15,88,209 // addps %xmm9,%xmm10 - .byte 65,15,88,217 // addps %xmm9,%xmm3 - .byte 102,68,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm9 - .byte 15,40,196 // movaps %xmm4,%xmm0 - .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 - .byte 102,68,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm9 - .byte 15,40,196 // movaps %xmm4,%xmm0 - .byte 15,194,197,1 // cmpltps %xmm5,%xmm0 - .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9 + .byte 15,40,29,123,42,0,0 // movaps 0x2a7b(%rip),%xmm3 # 3c90 <_sk_callback_sse41+0x36d> + .byte 65,15,88,219 // addps %xmm11,%xmm3 + .byte 102,68,15,56,20,235 // blendvps %xmm0,%xmm3,%xmm13 + .byte 65,15,92,229 // subps %xmm13,%xmm4 + .byte 69,15,40,205 // movaps %xmm13,%xmm9 + .byte 69,15,40,245 // movaps %xmm13,%xmm14 + .byte 68,15,194,108,36,144,1 // cmpltps -0x70(%rsp),%xmm13 + .byte 68,15,89,218 // mulps %xmm2,%xmm11 + .byte 15,89,226 // mulps %xmm2,%xmm4 + .byte 69,15,88,216 // addps %xmm8,%xmm11 + .byte 65,15,88,224 // addps %xmm8,%xmm4 + .byte 65,15,40,197 // movaps %xmm13,%xmm0 + .byte 102,68,15,56,20,196 // blendvps %xmm0,%xmm4,%xmm8 + .byte 68,15,194,116,36,128,1 // cmpltps -0x80(%rsp),%xmm14 + .byte 65,15,40,198 // movaps %xmm14,%xmm0 + .byte 102,68,15,56,20,198 // blendvps %xmm0,%xmm6,%xmm8 + .byte 68,15,194,207,1 // cmpltps %xmm7,%xmm9 + .byte 65,15,40,193 // movaps %xmm9,%xmm0 + .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8 .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 102,69,15,56,20,207 // blendvps %xmm0,%xmm15,%xmm9 + .byte 102,68,15,56,20,197 // blendvps %xmm0,%xmm5,%xmm8 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 65,15,40,195 // movaps %xmm11,%xmm0 - .byte 65,15,40,209 // movaps %xmm9,%xmm2 + .byte 65,15,40,194 // movaps %xmm10,%xmm0 + .byte 65,15,40,208 // movaps %xmm8,%xmm2 .byte 15,40,92,36,192 // movaps -0x40(%rsp),%xmm3 .byte 15,40,100,36,208 // movaps -0x30(%rsp),%xmm4 .byte 15,40,108,36,224 // movaps -0x20(%rsp),%xmm5 @@ -18860,7 +18880,7 @@ _sk_scale_u8_sse41: .byte 72,139,0 // mov (%rax),%rax .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,179,41,0,0 // mulps 0x29b3(%rip),%xmm8 # 3c70 <_sk_callback_sse41+0x365> + .byte 68,15,89,5,27,42,0,0 // mulps 0x2a1b(%rip),%xmm8 # 3cf0 <_sk_callback_sse41+0x3cd> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 65,15,89,208 // mulps %xmm8,%xmm2 @@ -18898,7 +18918,7 @@ _sk_lerp_u8_sse41: .byte 72,139,0 // mov (%rax),%rax .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,95,41,0,0 // mulps 0x295f(%rip),%xmm8 # 3c80 <_sk_callback_sse41+0x375> + .byte 68,15,89,5,199,41,0,0 // mulps 0x29c7(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x3dd> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -18921,17 +18941,17 @@ _sk_lerp_565_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 102,68,15,56,51,4,120 // pmovzxwd (%rax,%rdi,2),%xmm8 - .byte 102,15,111,29,47,41,0,0 // movdqa 0x292f(%rip),%xmm3 # 3c90 <_sk_callback_sse41+0x385> + .byte 102,15,111,29,151,41,0,0 // movdqa 0x2997(%rip),%xmm3 # 3d10 <_sk_callback_sse41+0x3ed> .byte 102,65,15,219,216 // pand %xmm8,%xmm3 .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 - .byte 68,15,89,13,46,41,0,0 // mulps 0x292e(%rip),%xmm9 # 3ca0 <_sk_callback_sse41+0x395> - .byte 102,15,111,29,54,41,0,0 // movdqa 0x2936(%rip),%xmm3 # 3cb0 <_sk_callback_sse41+0x3a5> + .byte 68,15,89,13,150,41,0,0 // mulps 0x2996(%rip),%xmm9 # 3d20 <_sk_callback_sse41+0x3fd> + .byte 102,15,111,29,158,41,0,0 // movdqa 0x299e(%rip),%xmm3 # 3d30 <_sk_callback_sse41+0x40d> .byte 102,65,15,219,216 // pand %xmm8,%xmm3 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,55,41,0,0 // mulps 0x2937(%rip),%xmm3 # 3cc0 <_sk_callback_sse41+0x3b5> - .byte 102,68,15,219,5,62,41,0,0 // pand 0x293e(%rip),%xmm8 # 3cd0 <_sk_callback_sse41+0x3c5> + .byte 15,89,29,159,41,0,0 // mulps 0x299f(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x41d> + .byte 102,68,15,219,5,166,41,0,0 // pand 0x29a6(%rip),%xmm8 # 3d50 <_sk_callback_sse41+0x42d> .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,66,41,0,0 // mulps 0x2942(%rip),%xmm8 # 3ce0 <_sk_callback_sse41+0x3d5> + .byte 68,15,89,5,170,41,0,0 // mulps 0x29aa(%rip),%xmm8 # 3d60 <_sk_callback_sse41+0x43d> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -18942,7 +18962,7 @@ _sk_lerp_565_sse41: .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 15,88,214 // addps %xmm6,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,44,41,0,0 // movaps 0x292c(%rip),%xmm3 # 3cf0 <_sk_callback_sse41+0x3e5> + .byte 15,40,29,148,41,0,0 // movaps 0x2994(%rip),%xmm3 # 3d70 <_sk_callback_sse41+0x44d> .byte 255,224 // jmpq *%rax HIDDEN _sk_load_tables_sse41 @@ -18953,7 +18973,7 @@ _sk_load_tables_sse41: .byte 76,139,0 // mov (%rax),%r8 .byte 76,139,72,8 // mov 0x8(%rax),%r9 .byte 243,69,15,111,4,184 // movdqu (%r8,%rdi,4),%xmm8 - .byte 102,15,111,5,35,41,0,0 // movdqa 0x2923(%rip),%xmm0 # 3d00 <_sk_callback_sse41+0x3f5> + .byte 102,15,111,5,139,41,0,0 // movdqa 0x298b(%rip),%xmm0 # 3d80 <_sk_callback_sse41+0x45d> .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,73,15,58,22,192,1 // pextrq $0x1,%xmm0,%r8 .byte 102,72,15,126,193 // movq %xmm0,%rcx @@ -18968,7 +18988,7 @@ _sk_load_tables_sse41: .byte 102,15,58,33,193,48 // insertps $0x30,%xmm1,%xmm0 .byte 76,139,64,16 // mov 0x10(%rax),%r8 .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1 - .byte 102,15,56,0,13,222,40,0,0 // pshufb 0x28de(%rip),%xmm1 # 3d10 <_sk_callback_sse41+0x405> + .byte 102,15,56,0,13,70,41,0,0 // pshufb 0x2946(%rip),%xmm1 # 3d90 <_sk_callback_sse41+0x46d> .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 102,72,15,126,201 // movq %xmm1,%rcx .byte 68,15,182,209 // movzbl %cl,%r10d @@ -18983,7 +19003,7 @@ _sk_load_tables_sse41: .byte 102,15,58,33,202,48 // insertps $0x30,%xmm2,%xmm1 .byte 76,139,64,24 // mov 0x18(%rax),%r8 .byte 102,65,15,111,208 // movdqa %xmm8,%xmm2 - .byte 102,15,56,0,21,154,40,0,0 // pshufb 0x289a(%rip),%xmm2 # 3d20 <_sk_callback_sse41+0x415> + .byte 102,15,56,0,21,2,41,0,0 // pshufb 0x2902(%rip),%xmm2 # 3da0 <_sk_callback_sse41+0x47d> .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx .byte 102,72,15,126,208 // movq %xmm2,%rax .byte 68,15,182,200 // movzbl %al,%r9d @@ -18998,7 +19018,7 @@ _sk_load_tables_sse41: .byte 102,15,58,33,211,48 // insertps $0x30,%xmm3,%xmm2 .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8 .byte 65,15,91,216 // cvtdq2ps %xmm8,%xmm3 - .byte 15,89,29,87,40,0,0 // mulps 0x2857(%rip),%xmm3 # 3d30 <_sk_callback_sse41+0x425> + .byte 15,89,29,191,40,0,0 // mulps 0x28bf(%rip),%xmm3 # 3db0 <_sk_callback_sse41+0x48d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -19017,7 +19037,7 @@ _sk_load_tables_u16_be_sse41: .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9 - .byte 102,68,15,111,5,42,40,0,0 // movdqa 0x282a(%rip),%xmm8 # 3d40 <_sk_callback_sse41+0x435> + .byte 102,68,15,111,5,146,40,0,0 // movdqa 0x2892(%rip),%xmm8 # 3dc0 <_sk_callback_sse41+0x49d> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0 @@ -19034,7 +19054,7 @@ _sk_load_tables_u16_be_sse41: .byte 243,67,15,16,20,8 // movss (%r8,%r9,1),%xmm2 .byte 102,15,58,33,194,48 // insertps $0x30,%xmm2,%xmm0 .byte 76,139,64,16 // mov 0x10(%rax),%r8 - .byte 102,15,56,0,13,221,39,0,0 // pshufb 0x27dd(%rip),%xmm1 # 3d50 <_sk_callback_sse41+0x445> + .byte 102,15,56,0,13,69,40,0,0 // pshufb 0x2845(%rip),%xmm1 # 3dd0 <_sk_callback_sse41+0x4ad> .byte 102,15,56,51,201 // pmovzxwd %xmm1,%xmm1 .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 102,72,15,126,201 // movq %xmm1,%rcx @@ -19070,7 +19090,7 @@ _sk_load_tables_u16_be_sse41: .byte 102,65,15,235,216 // por %xmm8,%xmm3 .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,43,39,0,0 // mulps 0x272b(%rip),%xmm3 # 3d60 <_sk_callback_sse41+0x455> + .byte 15,89,29,147,39,0,0 // mulps 0x2793(%rip),%xmm3 # 3de0 <_sk_callback_sse41+0x4bd> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -19092,7 +19112,7 @@ _sk_load_tables_rgb_u16_be_sse41: .byte 102,68,15,97,200 // punpcklwd %xmm0,%xmm9 .byte 102,15,111,202 // movdqa %xmm2,%xmm1 .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 - .byte 102,68,15,111,5,237,38,0,0 // movdqa 0x26ed(%rip),%xmm8 # 3d70 <_sk_callback_sse41+0x465> + .byte 102,68,15,111,5,85,39,0,0 // movdqa 0x2755(%rip),%xmm8 # 3df0 <_sk_callback_sse41+0x4cd> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0 @@ -19109,7 +19129,7 @@ _sk_load_tables_rgb_u16_be_sse41: .byte 243,67,15,16,28,8 // movss (%r8,%r9,1),%xmm3 .byte 102,15,58,33,195,48 // insertps $0x30,%xmm3,%xmm0 .byte 76,139,64,16 // mov 0x10(%rax),%r8 - .byte 102,15,56,0,13,160,38,0,0 // pshufb 0x26a0(%rip),%xmm1 # 3d80 <_sk_callback_sse41+0x475> + .byte 102,15,56,0,13,8,39,0,0 // pshufb 0x2708(%rip),%xmm1 # 3e00 <_sk_callback_sse41+0x4dd> .byte 102,15,56,51,201 // pmovzxwd %xmm1,%xmm1 .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 102,72,15,126,201 // movq %xmm1,%rcx @@ -19140,7 +19160,7 @@ _sk_load_tables_rgb_u16_be_sse41: .byte 243,65,15,16,28,8 // movss (%r8,%rcx,1),%xmm3 .byte 102,15,58,33,211,48 // insertps $0x30,%xmm3,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,11,38,0,0 // movaps 0x260b(%rip),%xmm3 # 3d90 <_sk_callback_sse41+0x485> + .byte 15,40,29,115,38,0,0 // movaps 0x2673(%rip),%xmm3 # 3e10 <_sk_callback_sse41+0x4ed> .byte 255,224 // jmpq *%rax HIDDEN _sk_byte_tables_sse41 @@ -19150,7 +19170,7 @@ _sk_byte_tables_sse41: .byte 65,86 // push %r14 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,12,38,0,0 // movaps 0x260c(%rip),%xmm8 # 3da0 <_sk_callback_sse41+0x495> + .byte 68,15,40,5,116,38,0,0 // movaps 0x2674(%rip),%xmm8 # 3e20 <_sk_callback_sse41+0x4fd> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx @@ -19169,7 +19189,7 @@ _sk_byte_tables_sse41: .byte 102,15,58,32,193,3 // pinsrb $0x3,%ecx,%xmm0 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,13,189,37,0,0 // movaps 0x25bd(%rip),%xmm9 # 3db0 <_sk_callback_sse41+0x4a5> + .byte 68,15,40,13,37,38,0,0 // movaps 0x2625(%rip),%xmm9 # 3e30 <_sk_callback_sse41+0x50d> .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -19260,7 +19280,7 @@ _sk_byte_tables_rgb_sse41: .byte 102,15,58,32,193,3 // pinsrb $0x3,%ecx,%xmm0 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,13,69,36,0,0 // movaps 0x2445(%rip),%xmm9 # 3dc0 <_sk_callback_sse41+0x4b5> + .byte 68,15,40,13,173,36,0,0 // movaps 0x24ad(%rip),%xmm9 # 3e40 <_sk_callback_sse41+0x51d> .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -19437,31 +19457,31 @@ _sk_parametric_r_sse41: .byte 69,15,88,208 // addps %xmm8,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,194 // cvtdq2ps %xmm10,%xmm8 - .byte 68,15,89,5,156,33,0,0 // mulps 0x219c(%rip),%xmm8 # 3dd0 <_sk_callback_sse41+0x4c5> - .byte 68,15,84,21,164,33,0,0 // andps 0x21a4(%rip),%xmm10 # 3de0 <_sk_callback_sse41+0x4d5> - .byte 68,15,86,21,172,33,0,0 // orps 0x21ac(%rip),%xmm10 # 3df0 <_sk_callback_sse41+0x4e5> - .byte 68,15,88,5,180,33,0,0 // addps 0x21b4(%rip),%xmm8 # 3e00 <_sk_callback_sse41+0x4f5> - .byte 68,15,40,37,188,33,0,0 // movaps 0x21bc(%rip),%xmm12 # 3e10 <_sk_callback_sse41+0x505> + .byte 68,15,89,5,4,34,0,0 // mulps 0x2204(%rip),%xmm8 # 3e50 <_sk_callback_sse41+0x52d> + .byte 68,15,84,21,12,34,0,0 // andps 0x220c(%rip),%xmm10 # 3e60 <_sk_callback_sse41+0x53d> + .byte 68,15,86,21,20,34,0,0 // orps 0x2214(%rip),%xmm10 # 3e70 <_sk_callback_sse41+0x54d> + .byte 68,15,88,5,28,34,0,0 // addps 0x221c(%rip),%xmm8 # 3e80 <_sk_callback_sse41+0x55d> + .byte 68,15,40,37,36,34,0,0 // movaps 0x2224(%rip),%xmm12 # 3e90 <_sk_callback_sse41+0x56d> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,196 // subps %xmm12,%xmm8 - .byte 68,15,88,21,188,33,0,0 // addps 0x21bc(%rip),%xmm10 # 3e20 <_sk_callback_sse41+0x515> - .byte 68,15,40,37,196,33,0,0 // movaps 0x21c4(%rip),%xmm12 # 3e30 <_sk_callback_sse41+0x525> + .byte 68,15,88,21,36,34,0,0 // addps 0x2224(%rip),%xmm10 # 3ea0 <_sk_callback_sse41+0x57d> + .byte 68,15,40,37,44,34,0,0 // movaps 0x222c(%rip),%xmm12 # 3eb0 <_sk_callback_sse41+0x58d> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,196 // subps %xmm12,%xmm8 .byte 69,15,89,195 // mulps %xmm11,%xmm8 .byte 102,69,15,58,8,208,1 // roundps $0x1,%xmm8,%xmm10 .byte 69,15,40,216 // movaps %xmm8,%xmm11 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 68,15,88,5,177,33,0,0 // addps 0x21b1(%rip),%xmm8 # 3e40 <_sk_callback_sse41+0x535> - .byte 68,15,40,21,185,33,0,0 // movaps 0x21b9(%rip),%xmm10 # 3e50 <_sk_callback_sse41+0x545> + .byte 68,15,88,5,25,34,0,0 // addps 0x2219(%rip),%xmm8 # 3ec0 <_sk_callback_sse41+0x59d> + .byte 68,15,40,21,33,34,0,0 // movaps 0x2221(%rip),%xmm10 # 3ed0 <_sk_callback_sse41+0x5ad> .byte 69,15,89,211 // mulps %xmm11,%xmm10 .byte 69,15,92,194 // subps %xmm10,%xmm8 - .byte 68,15,40,21,185,33,0,0 // movaps 0x21b9(%rip),%xmm10 # 3e60 <_sk_callback_sse41+0x555> + .byte 68,15,40,21,33,34,0,0 // movaps 0x2221(%rip),%xmm10 # 3ee0 <_sk_callback_sse41+0x5bd> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 68,15,40,29,189,33,0,0 // movaps 0x21bd(%rip),%xmm11 # 3e70 <_sk_callback_sse41+0x565> + .byte 68,15,40,29,37,34,0,0 // movaps 0x2225(%rip),%xmm11 # 3ef0 <_sk_callback_sse41+0x5cd> .byte 69,15,94,218 // divps %xmm10,%xmm11 .byte 69,15,88,216 // addps %xmm8,%xmm11 - .byte 68,15,89,29,189,33,0,0 // mulps 0x21bd(%rip),%xmm11 # 3e80 <_sk_callback_sse41+0x575> + .byte 68,15,89,29,37,34,0,0 // mulps 0x2225(%rip),%xmm11 # 3f00 <_sk_callback_sse41+0x5dd> .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 .byte 243,68,15,16,64,20 // movss 0x14(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 @@ -19469,7 +19489,7 @@ _sk_parametric_r_sse41: .byte 102,69,15,56,20,193 // blendvps %xmm0,%xmm9,%xmm8 .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 68,15,95,192 // maxps %xmm0,%xmm8 - .byte 68,15,93,5,164,33,0,0 // minps 0x21a4(%rip),%xmm8 # 3e90 <_sk_callback_sse41+0x585> + .byte 68,15,93,5,12,34,0,0 // minps 0x220c(%rip),%xmm8 # 3f10 <_sk_callback_sse41+0x5ed> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -19499,31 +19519,31 @@ _sk_parametric_g_sse41: .byte 68,15,88,217 // addps %xmm1,%xmm11 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 68,15,89,37,69,33,0,0 // mulps 0x2145(%rip),%xmm12 # 3ea0 <_sk_callback_sse41+0x595> - .byte 68,15,84,29,77,33,0,0 // andps 0x214d(%rip),%xmm11 # 3eb0 <_sk_callback_sse41+0x5a5> - .byte 68,15,86,29,85,33,0,0 // orps 0x2155(%rip),%xmm11 # 3ec0 <_sk_callback_sse41+0x5b5> - .byte 68,15,88,37,93,33,0,0 // addps 0x215d(%rip),%xmm12 # 3ed0 <_sk_callback_sse41+0x5c5> - .byte 15,40,13,102,33,0,0 // movaps 0x2166(%rip),%xmm1 # 3ee0 <_sk_callback_sse41+0x5d5> + .byte 68,15,89,37,173,33,0,0 // mulps 0x21ad(%rip),%xmm12 # 3f20 <_sk_callback_sse41+0x5fd> + .byte 68,15,84,29,181,33,0,0 // andps 0x21b5(%rip),%xmm11 # 3f30 <_sk_callback_sse41+0x60d> + .byte 68,15,86,29,189,33,0,0 // orps 0x21bd(%rip),%xmm11 # 3f40 <_sk_callback_sse41+0x61d> + .byte 68,15,88,37,197,33,0,0 // addps 0x21c5(%rip),%xmm12 # 3f50 <_sk_callback_sse41+0x62d> + .byte 15,40,13,206,33,0,0 // movaps 0x21ce(%rip),%xmm1 # 3f60 <_sk_callback_sse41+0x63d> .byte 65,15,89,203 // mulps %xmm11,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 - .byte 68,15,88,29,102,33,0,0 // addps 0x2166(%rip),%xmm11 # 3ef0 <_sk_callback_sse41+0x5e5> - .byte 15,40,13,111,33,0,0 // movaps 0x216f(%rip),%xmm1 # 3f00 <_sk_callback_sse41+0x5f5> + .byte 68,15,88,29,206,33,0,0 // addps 0x21ce(%rip),%xmm11 # 3f70 <_sk_callback_sse41+0x64d> + .byte 15,40,13,215,33,0,0 // movaps 0x21d7(%rip),%xmm1 # 3f80 <_sk_callback_sse41+0x65d> .byte 65,15,94,203 // divps %xmm11,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 102,69,15,58,8,212,1 // roundps $0x1,%xmm12,%xmm10 .byte 69,15,40,220 // movaps %xmm12,%xmm11 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 68,15,88,37,92,33,0,0 // addps 0x215c(%rip),%xmm12 # 3f10 <_sk_callback_sse41+0x605> - .byte 15,40,13,101,33,0,0 // movaps 0x2165(%rip),%xmm1 # 3f20 <_sk_callback_sse41+0x615> + .byte 68,15,88,37,196,33,0,0 // addps 0x21c4(%rip),%xmm12 # 3f90 <_sk_callback_sse41+0x66d> + .byte 15,40,13,205,33,0,0 // movaps 0x21cd(%rip),%xmm1 # 3fa0 <_sk_callback_sse41+0x67d> .byte 65,15,89,203 // mulps %xmm11,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 - .byte 68,15,40,21,101,33,0,0 // movaps 0x2165(%rip),%xmm10 # 3f30 <_sk_callback_sse41+0x625> + .byte 68,15,40,21,205,33,0,0 // movaps 0x21cd(%rip),%xmm10 # 3fb0 <_sk_callback_sse41+0x68d> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 15,40,13,106,33,0,0 // movaps 0x216a(%rip),%xmm1 # 3f40 <_sk_callback_sse41+0x635> + .byte 15,40,13,210,33,0,0 // movaps 0x21d2(%rip),%xmm1 # 3fc0 <_sk_callback_sse41+0x69d> .byte 65,15,94,202 // divps %xmm10,%xmm1 .byte 65,15,88,204 // addps %xmm12,%xmm1 - .byte 15,89,13,107,33,0,0 // mulps 0x216b(%rip),%xmm1 # 3f50 <_sk_callback_sse41+0x645> + .byte 15,89,13,211,33,0,0 // mulps 0x21d3(%rip),%xmm1 # 3fd0 <_sk_callback_sse41+0x6ad> .byte 102,68,15,91,209 // cvtps2dq %xmm1,%xmm10 .byte 243,15,16,72,20 // movss 0x14(%rax),%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 @@ -19531,7 +19551,7 @@ _sk_parametric_g_sse41: .byte 102,65,15,56,20,201 // blendvps %xmm0,%xmm9,%xmm1 .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 15,95,200 // maxps %xmm0,%xmm1 - .byte 15,93,13,86,33,0,0 // minps 0x2156(%rip),%xmm1 # 3f60 <_sk_callback_sse41+0x655> + .byte 15,93,13,190,33,0,0 // minps 0x21be(%rip),%xmm1 # 3fe0 <_sk_callback_sse41+0x6bd> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -19561,31 +19581,31 @@ _sk_parametric_b_sse41: .byte 68,15,88,218 // addps %xmm2,%xmm11 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 68,15,89,37,247,32,0,0 // mulps 0x20f7(%rip),%xmm12 # 3f70 <_sk_callback_sse41+0x665> - .byte 68,15,84,29,255,32,0,0 // andps 0x20ff(%rip),%xmm11 # 3f80 <_sk_callback_sse41+0x675> - .byte 68,15,86,29,7,33,0,0 // orps 0x2107(%rip),%xmm11 # 3f90 <_sk_callback_sse41+0x685> - .byte 68,15,88,37,15,33,0,0 // addps 0x210f(%rip),%xmm12 # 3fa0 <_sk_callback_sse41+0x695> - .byte 15,40,21,24,33,0,0 // movaps 0x2118(%rip),%xmm2 # 3fb0 <_sk_callback_sse41+0x6a5> + .byte 68,15,89,37,95,33,0,0 // mulps 0x215f(%rip),%xmm12 # 3ff0 <_sk_callback_sse41+0x6cd> + .byte 68,15,84,29,103,33,0,0 // andps 0x2167(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x6dd> + .byte 68,15,86,29,111,33,0,0 // orps 0x216f(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x6ed> + .byte 68,15,88,37,119,33,0,0 // addps 0x2177(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x6fd> + .byte 15,40,21,128,33,0,0 // movaps 0x2180(%rip),%xmm2 # 4030 <_sk_callback_sse41+0x70d> .byte 65,15,89,211 // mulps %xmm11,%xmm2 .byte 68,15,92,226 // subps %xmm2,%xmm12 - .byte 68,15,88,29,24,33,0,0 // addps 0x2118(%rip),%xmm11 # 3fc0 <_sk_callback_sse41+0x6b5> - .byte 15,40,21,33,33,0,0 // movaps 0x2121(%rip),%xmm2 # 3fd0 <_sk_callback_sse41+0x6c5> + .byte 68,15,88,29,128,33,0,0 // addps 0x2180(%rip),%xmm11 # 4040 <_sk_callback_sse41+0x71d> + .byte 15,40,21,137,33,0,0 // movaps 0x2189(%rip),%xmm2 # 4050 <_sk_callback_sse41+0x72d> .byte 65,15,94,211 // divps %xmm11,%xmm2 .byte 68,15,92,226 // subps %xmm2,%xmm12 .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 102,69,15,58,8,212,1 // roundps $0x1,%xmm12,%xmm10 .byte 69,15,40,220 // movaps %xmm12,%xmm11 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 68,15,88,37,14,33,0,0 // addps 0x210e(%rip),%xmm12 # 3fe0 <_sk_callback_sse41+0x6d5> - .byte 15,40,21,23,33,0,0 // movaps 0x2117(%rip),%xmm2 # 3ff0 <_sk_callback_sse41+0x6e5> + .byte 68,15,88,37,118,33,0,0 // addps 0x2176(%rip),%xmm12 # 4060 <_sk_callback_sse41+0x73d> + .byte 15,40,21,127,33,0,0 // movaps 0x217f(%rip),%xmm2 # 4070 <_sk_callback_sse41+0x74d> .byte 65,15,89,211 // mulps %xmm11,%xmm2 .byte 68,15,92,226 // subps %xmm2,%xmm12 - .byte 68,15,40,21,23,33,0,0 // movaps 0x2117(%rip),%xmm10 # 4000 <_sk_callback_sse41+0x6f5> + .byte 68,15,40,21,127,33,0,0 // movaps 0x217f(%rip),%xmm10 # 4080 <_sk_callback_sse41+0x75d> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 15,40,21,28,33,0,0 // movaps 0x211c(%rip),%xmm2 # 4010 <_sk_callback_sse41+0x705> + .byte 15,40,21,132,33,0,0 // movaps 0x2184(%rip),%xmm2 # 4090 <_sk_callback_sse41+0x76d> .byte 65,15,94,210 // divps %xmm10,%xmm2 .byte 65,15,88,212 // addps %xmm12,%xmm2 - .byte 15,89,21,29,33,0,0 // mulps 0x211d(%rip),%xmm2 # 4020 <_sk_callback_sse41+0x715> + .byte 15,89,21,133,33,0,0 // mulps 0x2185(%rip),%xmm2 # 40a0 <_sk_callback_sse41+0x77d> .byte 102,68,15,91,210 // cvtps2dq %xmm2,%xmm10 .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 @@ -19593,7 +19613,7 @@ _sk_parametric_b_sse41: .byte 102,65,15,56,20,209 // blendvps %xmm0,%xmm9,%xmm2 .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 15,95,208 // maxps %xmm0,%xmm2 - .byte 15,93,21,8,33,0,0 // minps 0x2108(%rip),%xmm2 # 4030 <_sk_callback_sse41+0x725> + .byte 15,93,21,112,33,0,0 // minps 0x2170(%rip),%xmm2 # 40b0 <_sk_callback_sse41+0x78d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -19623,31 +19643,31 @@ _sk_parametric_a_sse41: .byte 68,15,88,219 // addps %xmm3,%xmm11 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 68,15,89,37,169,32,0,0 // mulps 0x20a9(%rip),%xmm12 # 4040 <_sk_callback_sse41+0x735> - .byte 68,15,84,29,177,32,0,0 // andps 0x20b1(%rip),%xmm11 # 4050 <_sk_callback_sse41+0x745> - .byte 68,15,86,29,185,32,0,0 // orps 0x20b9(%rip),%xmm11 # 4060 <_sk_callback_sse41+0x755> - .byte 68,15,88,37,193,32,0,0 // addps 0x20c1(%rip),%xmm12 # 4070 <_sk_callback_sse41+0x765> - .byte 15,40,29,202,32,0,0 // movaps 0x20ca(%rip),%xmm3 # 4080 <_sk_callback_sse41+0x775> + .byte 68,15,89,37,17,33,0,0 // mulps 0x2111(%rip),%xmm12 # 40c0 <_sk_callback_sse41+0x79d> + .byte 68,15,84,29,25,33,0,0 // andps 0x2119(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x7ad> + .byte 68,15,86,29,33,33,0,0 // orps 0x2121(%rip),%xmm11 # 40e0 <_sk_callback_sse41+0x7bd> + .byte 68,15,88,37,41,33,0,0 // addps 0x2129(%rip),%xmm12 # 40f0 <_sk_callback_sse41+0x7cd> + .byte 15,40,29,50,33,0,0 // movaps 0x2132(%rip),%xmm3 # 4100 <_sk_callback_sse41+0x7dd> .byte 65,15,89,219 // mulps %xmm11,%xmm3 .byte 68,15,92,227 // subps %xmm3,%xmm12 - .byte 68,15,88,29,202,32,0,0 // addps 0x20ca(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x785> - .byte 15,40,29,211,32,0,0 // movaps 0x20d3(%rip),%xmm3 # 40a0 <_sk_callback_sse41+0x795> + .byte 68,15,88,29,50,33,0,0 // addps 0x2132(%rip),%xmm11 # 4110 <_sk_callback_sse41+0x7ed> + .byte 15,40,29,59,33,0,0 // movaps 0x213b(%rip),%xmm3 # 4120 <_sk_callback_sse41+0x7fd> .byte 65,15,94,219 // divps %xmm11,%xmm3 .byte 68,15,92,227 // subps %xmm3,%xmm12 .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 102,69,15,58,8,212,1 // roundps $0x1,%xmm12,%xmm10 .byte 69,15,40,220 // movaps %xmm12,%xmm11 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 68,15,88,37,192,32,0,0 // addps 0x20c0(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x7a5> - .byte 15,40,29,201,32,0,0 // movaps 0x20c9(%rip),%xmm3 # 40c0 <_sk_callback_sse41+0x7b5> + .byte 68,15,88,37,40,33,0,0 // addps 0x2128(%rip),%xmm12 # 4130 <_sk_callback_sse41+0x80d> + .byte 15,40,29,49,33,0,0 // movaps 0x2131(%rip),%xmm3 # 4140 <_sk_callback_sse41+0x81d> .byte 65,15,89,219 // mulps %xmm11,%xmm3 .byte 68,15,92,227 // subps %xmm3,%xmm12 - .byte 68,15,40,21,201,32,0,0 // movaps 0x20c9(%rip),%xmm10 # 40d0 <_sk_callback_sse41+0x7c5> + .byte 68,15,40,21,49,33,0,0 // movaps 0x2131(%rip),%xmm10 # 4150 <_sk_callback_sse41+0x82d> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 15,40,29,206,32,0,0 // movaps 0x20ce(%rip),%xmm3 # 40e0 <_sk_callback_sse41+0x7d5> + .byte 15,40,29,54,33,0,0 // movaps 0x2136(%rip),%xmm3 # 4160 <_sk_callback_sse41+0x83d> .byte 65,15,94,218 // divps %xmm10,%xmm3 .byte 65,15,88,220 // addps %xmm12,%xmm3 - .byte 15,89,29,207,32,0,0 // mulps 0x20cf(%rip),%xmm3 # 40f0 <_sk_callback_sse41+0x7e5> + .byte 15,89,29,55,33,0,0 // mulps 0x2137(%rip),%xmm3 # 4170 <_sk_callback_sse41+0x84d> .byte 102,68,15,91,211 // cvtps2dq %xmm3,%xmm10 .byte 243,15,16,88,20 // movss 0x14(%rax),%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 @@ -19655,7 +19675,7 @@ _sk_parametric_a_sse41: .byte 102,65,15,56,20,217 // blendvps %xmm0,%xmm9,%xmm3 .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 15,95,216 // maxps %xmm0,%xmm3 - .byte 15,93,29,186,32,0,0 // minps 0x20ba(%rip),%xmm3 # 4100 <_sk_callback_sse41+0x7f5> + .byte 15,93,29,34,33,0,0 // minps 0x2122(%rip),%xmm3 # 4180 <_sk_callback_sse41+0x85d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -19665,29 +19685,29 @@ HIDDEN _sk_lab_to_xyz_sse41 FUNCTION(_sk_lab_to_xyz_sse41) _sk_lab_to_xyz_sse41: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,89,5,182,32,0,0 // mulps 0x20b6(%rip),%xmm8 # 4110 <_sk_callback_sse41+0x805> - .byte 68,15,40,13,190,32,0,0 // movaps 0x20be(%rip),%xmm9 # 4120 <_sk_callback_sse41+0x815> + .byte 68,15,89,5,30,33,0,0 // mulps 0x211e(%rip),%xmm8 # 4190 <_sk_callback_sse41+0x86d> + .byte 68,15,40,13,38,33,0,0 // movaps 0x2126(%rip),%xmm9 # 41a0 <_sk_callback_sse41+0x87d> .byte 65,15,89,201 // mulps %xmm9,%xmm1 - .byte 15,40,5,195,32,0,0 // movaps 0x20c3(%rip),%xmm0 # 4130 <_sk_callback_sse41+0x825> + .byte 15,40,5,43,33,0,0 // movaps 0x212b(%rip),%xmm0 # 41b0 <_sk_callback_sse41+0x88d> .byte 15,88,200 // addps %xmm0,%xmm1 .byte 65,15,89,209 // mulps %xmm9,%xmm2 .byte 15,88,208 // addps %xmm0,%xmm2 - .byte 68,15,88,5,193,32,0,0 // addps 0x20c1(%rip),%xmm8 # 4140 <_sk_callback_sse41+0x835> - .byte 68,15,89,5,201,32,0,0 // mulps 0x20c9(%rip),%xmm8 # 4150 <_sk_callback_sse41+0x845> - .byte 15,89,13,210,32,0,0 // mulps 0x20d2(%rip),%xmm1 # 4160 <_sk_callback_sse41+0x855> + .byte 68,15,88,5,41,33,0,0 // addps 0x2129(%rip),%xmm8 # 41c0 <_sk_callback_sse41+0x89d> + .byte 68,15,89,5,49,33,0,0 // mulps 0x2131(%rip),%xmm8 # 41d0 <_sk_callback_sse41+0x8ad> + .byte 15,89,13,58,33,0,0 // mulps 0x213a(%rip),%xmm1 # 41e0 <_sk_callback_sse41+0x8bd> .byte 65,15,88,200 // addps %xmm8,%xmm1 - .byte 15,89,21,215,32,0,0 // mulps 0x20d7(%rip),%xmm2 # 4170 <_sk_callback_sse41+0x865> + .byte 15,89,21,63,33,0,0 // mulps 0x213f(%rip),%xmm2 # 41f0 <_sk_callback_sse41+0x8cd> .byte 69,15,40,208 // movaps %xmm8,%xmm10 .byte 68,15,92,210 // subps %xmm2,%xmm10 .byte 68,15,40,217 // movaps %xmm1,%xmm11 .byte 69,15,89,219 // mulps %xmm11,%xmm11 .byte 68,15,89,217 // mulps %xmm1,%xmm11 - .byte 68,15,40,13,203,32,0,0 // movaps 0x20cb(%rip),%xmm9 # 4180 <_sk_callback_sse41+0x875> + .byte 68,15,40,13,51,33,0,0 // movaps 0x2133(%rip),%xmm9 # 4200 <_sk_callback_sse41+0x8dd> .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0 - .byte 15,40,21,203,32,0,0 // movaps 0x20cb(%rip),%xmm2 # 4190 <_sk_callback_sse41+0x885> + .byte 15,40,21,51,33,0,0 // movaps 0x2133(%rip),%xmm2 # 4210 <_sk_callback_sse41+0x8ed> .byte 15,88,202 // addps %xmm2,%xmm1 - .byte 68,15,40,37,208,32,0,0 // movaps 0x20d0(%rip),%xmm12 # 41a0 <_sk_callback_sse41+0x895> + .byte 68,15,40,37,56,33,0,0 // movaps 0x2138(%rip),%xmm12 # 4220 <_sk_callback_sse41+0x8fd> .byte 65,15,89,204 // mulps %xmm12,%xmm1 .byte 102,65,15,56,20,203 // blendvps %xmm0,%xmm11,%xmm1 .byte 69,15,40,216 // movaps %xmm8,%xmm11 @@ -19706,8 +19726,8 @@ _sk_lab_to_xyz_sse41: .byte 65,15,89,212 // mulps %xmm12,%xmm2 .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2 - .byte 15,89,13,137,32,0,0 // mulps 0x2089(%rip),%xmm1 # 41b0 <_sk_callback_sse41+0x8a5> - .byte 15,89,21,146,32,0,0 // mulps 0x2092(%rip),%xmm2 # 41c0 <_sk_callback_sse41+0x8b5> + .byte 15,89,13,241,32,0,0 // mulps 0x20f1(%rip),%xmm1 # 4230 <_sk_callback_sse41+0x90d> + .byte 15,89,21,250,32,0,0 // mulps 0x20fa(%rip),%xmm2 # 4240 <_sk_callback_sse41+0x91d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,40,193 // movaps %xmm1,%xmm0 .byte 65,15,40,200 // movaps %xmm8,%xmm1 @@ -19721,7 +19741,7 @@ _sk_load_a8_sse41: .byte 72,139,0 // mov (%rax),%rax .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0 .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,130,32,0,0 // mulps 0x2082(%rip),%xmm3 # 41d0 <_sk_callback_sse41+0x8c5> + .byte 15,89,29,234,32,0,0 // mulps 0x20ea(%rip),%xmm3 # 4250 <_sk_callback_sse41+0x92d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 15,87,201 // xorps %xmm1,%xmm1 @@ -19754,7 +19774,7 @@ _sk_gather_a8_sse41: .byte 102,15,58,32,192,3 // pinsrb $0x3,%eax,%xmm0 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,22,32,0,0 // mulps 0x2016(%rip),%xmm3 # 41e0 <_sk_callback_sse41+0x8d5> + .byte 15,89,29,126,32,0,0 // mulps 0x207e(%rip),%xmm3 # 4260 <_sk_callback_sse41+0x93d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 102,15,239,201 // pxor %xmm1,%xmm1 @@ -19767,7 +19787,7 @@ FUNCTION(_sk_store_a8_sse41) _sk_store_a8_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,10,32,0,0 // movaps 0x200a(%rip),%xmm8 # 41f0 <_sk_callback_sse41+0x8e5> + .byte 68,15,40,5,114,32,0,0 // movaps 0x2072(%rip),%xmm8 # 4270 <_sk_callback_sse41+0x94d> .byte 68,15,89,195 // mulps %xmm3,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 @@ -19784,9 +19804,9 @@ _sk_load_g8_sse41: .byte 72,139,0 // mov (%rax),%rax .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,231,31,0,0 // mulps 0x1fe7(%rip),%xmm0 # 4200 <_sk_callback_sse41+0x8f5> + .byte 15,89,5,79,32,0,0 // mulps 0x204f(%rip),%xmm0 # 4280 <_sk_callback_sse41+0x95d> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,238,31,0,0 // movaps 0x1fee(%rip),%xmm3 # 4210 <_sk_callback_sse41+0x905> + .byte 15,40,29,86,32,0,0 // movaps 0x2056(%rip),%xmm3 # 4290 <_sk_callback_sse41+0x96d> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 255,224 // jmpq *%rax @@ -19817,9 +19837,9 @@ _sk_gather_g8_sse41: .byte 102,15,58,32,192,3 // pinsrb $0x3,%eax,%xmm0 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,135,31,0,0 // mulps 0x1f87(%rip),%xmm0 # 4220 <_sk_callback_sse41+0x915> + .byte 15,89,5,239,31,0,0 // mulps 0x1fef(%rip),%xmm0 # 42a0 <_sk_callback_sse41+0x97d> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,142,31,0,0 // movaps 0x1f8e(%rip),%xmm3 # 4230 <_sk_callback_sse41+0x925> + .byte 15,40,29,246,31,0,0 // movaps 0x1ff6(%rip),%xmm3 # 42b0 <_sk_callback_sse41+0x98d> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 255,224 // jmpq *%rax @@ -19831,9 +19851,9 @@ _sk_gather_i8_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 22b9 <_sk_gather_i8_sse41+0xf> + .byte 116,5 // je 22d1 <_sk_gather_i8_sse41+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 22bb <_sk_gather_i8_sse41+0x11> + .byte 235,2 // jmp 22d3 <_sk_gather_i8_sse41+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 @@ -19864,17 +19884,17 @@ _sk_gather_i8_sse41: .byte 102,15,58,34,28,8,1 // pinsrd $0x1,(%rax,%rcx,1),%xmm3 .byte 102,66,15,58,34,28,144,2 // pinsrd $0x2,(%rax,%r10,4),%xmm3 .byte 102,66,15,58,34,28,8,3 // pinsrd $0x3,(%rax,%r9,1),%xmm3 - .byte 102,15,111,5,229,30,0,0 // movdqa 0x1ee5(%rip),%xmm0 # 4240 <_sk_callback_sse41+0x935> + .byte 102,15,111,5,77,31,0,0 // movdqa 0x1f4d(%rip),%xmm0 # 42c0 <_sk_callback_sse41+0x99d> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,230,30,0,0 // movaps 0x1ee6(%rip),%xmm8 # 4250 <_sk_callback_sse41+0x945> + .byte 68,15,40,5,78,31,0,0 // movaps 0x1f4e(%rip),%xmm8 # 42d0 <_sk_callback_sse41+0x9ad> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,15,56,0,13,229,30,0,0 // pshufb 0x1ee5(%rip),%xmm1 # 4260 <_sk_callback_sse41+0x955> + .byte 102,15,56,0,13,77,31,0,0 // pshufb 0x1f4d(%rip),%xmm1 # 42e0 <_sk_callback_sse41+0x9bd> .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,15,56,0,21,225,30,0,0 // pshufb 0x1ee1(%rip),%xmm2 # 4270 <_sk_callback_sse41+0x965> + .byte 102,15,56,0,21,73,31,0,0 // pshufb 0x1f49(%rip),%xmm2 # 42f0 <_sk_callback_sse41+0x9cd> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 102,15,114,211,24 // psrld $0x18,%xmm3 @@ -19890,19 +19910,19 @@ _sk_load_565_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 102,15,56,51,20,120 // pmovzxwd (%rax,%rdi,2),%xmm2 - .byte 102,15,111,5,199,30,0,0 // movdqa 0x1ec7(%rip),%xmm0 # 4280 <_sk_callback_sse41+0x975> + .byte 102,15,111,5,47,31,0,0 // movdqa 0x1f2f(%rip),%xmm0 # 4300 <_sk_callback_sse41+0x9dd> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,201,30,0,0 // mulps 0x1ec9(%rip),%xmm0 # 4290 <_sk_callback_sse41+0x985> - .byte 102,15,111,13,209,30,0,0 // movdqa 0x1ed1(%rip),%xmm1 # 42a0 <_sk_callback_sse41+0x995> + .byte 15,89,5,49,31,0,0 // mulps 0x1f31(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x9ed> + .byte 102,15,111,13,57,31,0,0 // movdqa 0x1f39(%rip),%xmm1 # 4320 <_sk_callback_sse41+0x9fd> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,211,30,0,0 // mulps 0x1ed3(%rip),%xmm1 # 42b0 <_sk_callback_sse41+0x9a5> - .byte 102,15,219,21,219,30,0,0 // pand 0x1edb(%rip),%xmm2 # 42c0 <_sk_callback_sse41+0x9b5> + .byte 15,89,13,59,31,0,0 // mulps 0x1f3b(%rip),%xmm1 # 4330 <_sk_callback_sse41+0xa0d> + .byte 102,15,219,21,67,31,0,0 // pand 0x1f43(%rip),%xmm2 # 4340 <_sk_callback_sse41+0xa1d> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,225,30,0,0 // mulps 0x1ee1(%rip),%xmm2 # 42d0 <_sk_callback_sse41+0x9c5> + .byte 15,89,21,73,31,0,0 // mulps 0x1f49(%rip),%xmm2 # 4350 <_sk_callback_sse41+0xa2d> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,232,30,0,0 // movaps 0x1ee8(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x9d5> + .byte 15,40,29,80,31,0,0 // movaps 0x1f50(%rip),%xmm3 # 4360 <_sk_callback_sse41+0xa3d> .byte 255,224 // jmpq *%rax HIDDEN _sk_gather_565_sse41 @@ -19930,19 +19950,19 @@ _sk_gather_565_sse41: .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0 .byte 102,15,56,51,208 // pmovzxwd %xmm0,%xmm2 - .byte 102,15,111,5,141,30,0,0 // movdqa 0x1e8d(%rip),%xmm0 # 42f0 <_sk_callback_sse41+0x9e5> + .byte 102,15,111,5,245,30,0,0 // movdqa 0x1ef5(%rip),%xmm0 # 4370 <_sk_callback_sse41+0xa4d> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,143,30,0,0 // mulps 0x1e8f(%rip),%xmm0 # 4300 <_sk_callback_sse41+0x9f5> - .byte 102,15,111,13,151,30,0,0 // movdqa 0x1e97(%rip),%xmm1 # 4310 <_sk_callback_sse41+0xa05> + .byte 15,89,5,247,30,0,0 // mulps 0x1ef7(%rip),%xmm0 # 4380 <_sk_callback_sse41+0xa5d> + .byte 102,15,111,13,255,30,0,0 // movdqa 0x1eff(%rip),%xmm1 # 4390 <_sk_callback_sse41+0xa6d> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,153,30,0,0 // mulps 0x1e99(%rip),%xmm1 # 4320 <_sk_callback_sse41+0xa15> - .byte 102,15,219,21,161,30,0,0 // pand 0x1ea1(%rip),%xmm2 # 4330 <_sk_callback_sse41+0xa25> + .byte 15,89,13,1,31,0,0 // mulps 0x1f01(%rip),%xmm1 # 43a0 <_sk_callback_sse41+0xa7d> + .byte 102,15,219,21,9,31,0,0 // pand 0x1f09(%rip),%xmm2 # 43b0 <_sk_callback_sse41+0xa8d> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,167,30,0,0 // mulps 0x1ea7(%rip),%xmm2 # 4340 <_sk_callback_sse41+0xa35> + .byte 15,89,21,15,31,0,0 // mulps 0x1f0f(%rip),%xmm2 # 43c0 <_sk_callback_sse41+0xa9d> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,174,30,0,0 // movaps 0x1eae(%rip),%xmm3 # 4350 <_sk_callback_sse41+0xa45> + .byte 15,40,29,22,31,0,0 // movaps 0x1f16(%rip),%xmm3 # 43d0 <_sk_callback_sse41+0xaad> .byte 255,224 // jmpq *%rax HIDDEN _sk_store_565_sse41 @@ -19951,12 +19971,12 @@ FUNCTION(_sk_store_565_sse41) _sk_store_565_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,175,30,0,0 // movaps 0x1eaf(%rip),%xmm8 # 4360 <_sk_callback_sse41+0xa55> + .byte 68,15,40,5,23,31,0,0 // movaps 0x1f17(%rip),%xmm8 # 43e0 <_sk_callback_sse41+0xabd> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 .byte 102,65,15,114,241,11 // pslld $0xb,%xmm9 - .byte 68,15,40,21,164,30,0,0 // movaps 0x1ea4(%rip),%xmm10 # 4370 <_sk_callback_sse41+0xa65> + .byte 68,15,40,21,12,31,0,0 // movaps 0x1f0c(%rip),%xmm10 # 43f0 <_sk_callback_sse41+0xacd> .byte 68,15,89,209 // mulps %xmm1,%xmm10 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10 @@ -19976,21 +19996,21 @@ _sk_load_4444_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 102,15,56,51,28,120 // pmovzxwd (%rax,%rdi,2),%xmm3 - .byte 102,15,111,5,111,30,0,0 // movdqa 0x1e6f(%rip),%xmm0 # 4380 <_sk_callback_sse41+0xa75> + .byte 102,15,111,5,215,30,0,0 // movdqa 0x1ed7(%rip),%xmm0 # 4400 <_sk_callback_sse41+0xadd> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,113,30,0,0 // mulps 0x1e71(%rip),%xmm0 # 4390 <_sk_callback_sse41+0xa85> - .byte 102,15,111,13,121,30,0,0 // movdqa 0x1e79(%rip),%xmm1 # 43a0 <_sk_callback_sse41+0xa95> + .byte 15,89,5,217,30,0,0 // mulps 0x1ed9(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xaed> + .byte 102,15,111,13,225,30,0,0 // movdqa 0x1ee1(%rip),%xmm1 # 4420 <_sk_callback_sse41+0xafd> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,123,30,0,0 // mulps 0x1e7b(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0xaa5> - .byte 102,15,111,21,131,30,0,0 // movdqa 0x1e83(%rip),%xmm2 # 43c0 <_sk_callback_sse41+0xab5> + .byte 15,89,13,227,30,0,0 // mulps 0x1ee3(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xb0d> + .byte 102,15,111,21,235,30,0,0 // movdqa 0x1eeb(%rip),%xmm2 # 4440 <_sk_callback_sse41+0xb1d> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,133,30,0,0 // mulps 0x1e85(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0xac5> - .byte 102,15,219,29,141,30,0,0 // pand 0x1e8d(%rip),%xmm3 # 43e0 <_sk_callback_sse41+0xad5> + .byte 15,89,21,237,30,0,0 // mulps 0x1eed(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xb2d> + .byte 102,15,219,29,245,30,0,0 // pand 0x1ef5(%rip),%xmm3 # 4460 <_sk_callback_sse41+0xb3d> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,147,30,0,0 // mulps 0x1e93(%rip),%xmm3 # 43f0 <_sk_callback_sse41+0xae5> + .byte 15,89,29,251,30,0,0 // mulps 0x1efb(%rip),%xmm3 # 4470 <_sk_callback_sse41+0xb4d> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20019,21 +20039,21 @@ _sk_gather_4444_sse41: .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0 .byte 102,15,56,51,216 // pmovzxwd %xmm0,%xmm3 - .byte 102,15,111,5,54,30,0,0 // movdqa 0x1e36(%rip),%xmm0 # 4400 <_sk_callback_sse41+0xaf5> + .byte 102,15,111,5,158,30,0,0 // movdqa 0x1e9e(%rip),%xmm0 # 4480 <_sk_callback_sse41+0xb5d> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,56,30,0,0 // mulps 0x1e38(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xb05> - .byte 102,15,111,13,64,30,0,0 // movdqa 0x1e40(%rip),%xmm1 # 4420 <_sk_callback_sse41+0xb15> + .byte 15,89,5,160,30,0,0 // mulps 0x1ea0(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xb6d> + .byte 102,15,111,13,168,30,0,0 // movdqa 0x1ea8(%rip),%xmm1 # 44a0 <_sk_callback_sse41+0xb7d> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,66,30,0,0 // mulps 0x1e42(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xb25> - .byte 102,15,111,21,74,30,0,0 // movdqa 0x1e4a(%rip),%xmm2 # 4440 <_sk_callback_sse41+0xb35> + .byte 15,89,13,170,30,0,0 // mulps 0x1eaa(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xb8d> + .byte 102,15,111,21,178,30,0,0 // movdqa 0x1eb2(%rip),%xmm2 # 44c0 <_sk_callback_sse41+0xb9d> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,76,30,0,0 // mulps 0x1e4c(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xb45> - .byte 102,15,219,29,84,30,0,0 // pand 0x1e54(%rip),%xmm3 # 4460 <_sk_callback_sse41+0xb55> + .byte 15,89,21,180,30,0,0 // mulps 0x1eb4(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xbad> + .byte 102,15,219,29,188,30,0,0 // pand 0x1ebc(%rip),%xmm3 # 44e0 <_sk_callback_sse41+0xbbd> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,90,30,0,0 // mulps 0x1e5a(%rip),%xmm3 # 4470 <_sk_callback_sse41+0xb65> + .byte 15,89,29,194,30,0,0 // mulps 0x1ec2(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xbcd> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20043,7 +20063,7 @@ FUNCTION(_sk_store_4444_sse41) _sk_store_4444_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,89,30,0,0 // movaps 0x1e59(%rip),%xmm8 # 4480 <_sk_callback_sse41+0xb75> + .byte 68,15,40,5,193,30,0,0 // movaps 0x1ec1(%rip),%xmm8 # 4500 <_sk_callback_sse41+0xbdd> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 @@ -20073,17 +20093,17 @@ _sk_load_8888_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 15,16,28,184 // movups (%rax,%rdi,4),%xmm3 - .byte 15,40,5,248,29,0,0 // movaps 0x1df8(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xb85> + .byte 15,40,5,96,30,0,0 // movaps 0x1e60(%rip),%xmm0 # 4510 <_sk_callback_sse41+0xbed> .byte 15,84,195 // andps %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,250,29,0,0 // movaps 0x1dfa(%rip),%xmm8 # 44a0 <_sk_callback_sse41+0xb95> + .byte 68,15,40,5,98,30,0,0 // movaps 0x1e62(%rip),%xmm8 # 4520 <_sk_callback_sse41+0xbfd> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,40,203 // movaps %xmm3,%xmm1 - .byte 102,15,56,0,13,250,29,0,0 // pshufb 0x1dfa(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xba5> + .byte 102,15,56,0,13,98,30,0,0 // pshufb 0x1e62(%rip),%xmm1 # 4530 <_sk_callback_sse41+0xc0d> .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 15,40,211 // movaps %xmm3,%xmm2 - .byte 102,15,56,0,21,247,29,0,0 // pshufb 0x1df7(%rip),%xmm2 # 44c0 <_sk_callback_sse41+0xbb5> + .byte 102,15,56,0,21,95,30,0,0 // pshufb 0x1e5f(%rip),%xmm2 # 4540 <_sk_callback_sse41+0xc1d> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 102,15,114,211,24 // psrld $0x18,%xmm3 @@ -20114,17 +20134,17 @@ _sk_gather_8888_sse41: .byte 102,65,15,58,34,28,129,1 // pinsrd $0x1,(%r9,%rax,4),%xmm3 .byte 102,67,15,58,34,28,145,2 // pinsrd $0x2,(%r9,%r10,4),%xmm3 .byte 102,65,15,58,34,28,137,3 // pinsrd $0x3,(%r9,%rcx,4),%xmm3 - .byte 102,15,111,5,144,29,0,0 // movdqa 0x1d90(%rip),%xmm0 # 44d0 <_sk_callback_sse41+0xbc5> + .byte 102,15,111,5,248,29,0,0 // movdqa 0x1df8(%rip),%xmm0 # 4550 <_sk_callback_sse41+0xc2d> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,145,29,0,0 // movaps 0x1d91(%rip),%xmm8 # 44e0 <_sk_callback_sse41+0xbd5> + .byte 68,15,40,5,249,29,0,0 // movaps 0x1df9(%rip),%xmm8 # 4560 <_sk_callback_sse41+0xc3d> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,15,56,0,13,144,29,0,0 // pshufb 0x1d90(%rip),%xmm1 # 44f0 <_sk_callback_sse41+0xbe5> + .byte 102,15,56,0,13,248,29,0,0 // pshufb 0x1df8(%rip),%xmm1 # 4570 <_sk_callback_sse41+0xc4d> .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,15,56,0,21,140,29,0,0 // pshufb 0x1d8c(%rip),%xmm2 # 4500 <_sk_callback_sse41+0xbf5> + .byte 102,15,56,0,21,244,29,0,0 // pshufb 0x1df4(%rip),%xmm2 # 4580 <_sk_callback_sse41+0xc5d> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 102,15,114,211,24 // psrld $0x18,%xmm3 @@ -20139,7 +20159,7 @@ FUNCTION(_sk_store_8888_sse41) _sk_store_8888_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,120,29,0,0 // movaps 0x1d78(%rip),%xmm8 # 4510 <_sk_callback_sse41+0xc05> + .byte 68,15,40,5,224,29,0,0 // movaps 0x1de0(%rip),%xmm8 # 4590 <_sk_callback_sse41+0xc6d> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 @@ -20176,18 +20196,18 @@ _sk_load_f16_sse41: .byte 102,68,15,97,216 // punpcklwd %xmm0,%xmm11 .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9 .byte 102,65,15,56,51,203 // pmovzxwd %xmm11,%xmm1 - .byte 102,68,15,111,5,241,28,0,0 // movdqa 0x1cf1(%rip),%xmm8 # 4520 <_sk_callback_sse41+0xc15> + .byte 102,68,15,111,5,89,29,0,0 // movdqa 0x1d59(%rip),%xmm8 # 45a0 <_sk_callback_sse41+0xc7d> .byte 102,15,111,209 // movdqa %xmm1,%xmm2 .byte 102,65,15,219,208 // pand %xmm8,%xmm2 .byte 102,15,239,202 // pxor %xmm2,%xmm1 - .byte 102,15,111,29,236,28,0,0 // movdqa 0x1cec(%rip),%xmm3 # 4530 <_sk_callback_sse41+0xc25> + .byte 102,15,111,29,84,29,0,0 // movdqa 0x1d54(%rip),%xmm3 # 45b0 <_sk_callback_sse41+0xc8d> .byte 102,15,114,242,16 // pslld $0x10,%xmm2 .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,15,56,63,195 // pmaxud %xmm3,%xmm0 .byte 102,15,118,193 // pcmpeqd %xmm1,%xmm0 .byte 102,15,114,241,13 // pslld $0xd,%xmm1 .byte 102,15,235,202 // por %xmm2,%xmm1 - .byte 102,68,15,111,21,216,28,0,0 // movdqa 0x1cd8(%rip),%xmm10 # 4540 <_sk_callback_sse41+0xc35> + .byte 102,68,15,111,21,64,29,0,0 // movdqa 0x1d40(%rip),%xmm10 # 45c0 <_sk_callback_sse41+0xc9d> .byte 102,65,15,254,202 // paddd %xmm10,%xmm1 .byte 102,15,219,193 // pand %xmm1,%xmm0 .byte 102,65,15,115,219,8 // psrldq $0x8,%xmm11 @@ -20260,18 +20280,18 @@ _sk_gather_f16_sse41: .byte 102,68,15,97,218 // punpcklwd %xmm2,%xmm11 .byte 102,68,15,105,202 // punpckhwd %xmm2,%xmm9 .byte 102,65,15,56,51,203 // pmovzxwd %xmm11,%xmm1 - .byte 102,68,15,111,5,150,27,0,0 // movdqa 0x1b96(%rip),%xmm8 # 4550 <_sk_callback_sse41+0xc45> + .byte 102,68,15,111,5,254,27,0,0 // movdqa 0x1bfe(%rip),%xmm8 # 45d0 <_sk_callback_sse41+0xcad> .byte 102,15,111,209 // movdqa %xmm1,%xmm2 .byte 102,65,15,219,208 // pand %xmm8,%xmm2 .byte 102,15,239,202 // pxor %xmm2,%xmm1 - .byte 102,15,111,29,145,27,0,0 // movdqa 0x1b91(%rip),%xmm3 # 4560 <_sk_callback_sse41+0xc55> + .byte 102,15,111,29,249,27,0,0 // movdqa 0x1bf9(%rip),%xmm3 # 45e0 <_sk_callback_sse41+0xcbd> .byte 102,15,114,242,16 // pslld $0x10,%xmm2 .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,15,56,63,195 // pmaxud %xmm3,%xmm0 .byte 102,15,118,193 // pcmpeqd %xmm1,%xmm0 .byte 102,15,114,241,13 // pslld $0xd,%xmm1 .byte 102,15,235,202 // por %xmm2,%xmm1 - .byte 102,68,15,111,21,125,27,0,0 // movdqa 0x1b7d(%rip),%xmm10 # 4570 <_sk_callback_sse41+0xc65> + .byte 102,68,15,111,21,229,27,0,0 // movdqa 0x1be5(%rip),%xmm10 # 45f0 <_sk_callback_sse41+0xccd> .byte 102,65,15,254,202 // paddd %xmm10,%xmm1 .byte 102,15,219,193 // pand %xmm1,%xmm0 .byte 102,65,15,115,219,8 // psrldq $0x8,%xmm11 @@ -20319,17 +20339,17 @@ FUNCTION(_sk_store_f16_sse41) _sk_store_f16_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 102,68,15,111,21,179,26,0,0 // movdqa 0x1ab3(%rip),%xmm10 # 4580 <_sk_callback_sse41+0xc75> + .byte 102,68,15,111,21,27,27,0,0 // movdqa 0x1b1b(%rip),%xmm10 # 4600 <_sk_callback_sse41+0xcdd> .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13 .byte 102,69,15,219,234 // pand %xmm10,%xmm13 .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 - .byte 102,68,15,111,13,166,26,0,0 // movdqa 0x1aa6(%rip),%xmm9 # 4590 <_sk_callback_sse41+0xc85> + .byte 102,68,15,111,13,14,27,0,0 // movdqa 0x1b0e(%rip),%xmm9 # 4610 <_sk_callback_sse41+0xced> .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13 .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8 .byte 102,69,15,102,196 // pcmpgtd %xmm12,%xmm8 .byte 102,65,15,114,212,13 // psrld $0xd,%xmm12 - .byte 102,68,15,111,29,151,26,0,0 // movdqa 0x1a97(%rip),%xmm11 # 45a0 <_sk_callback_sse41+0xc95> + .byte 102,68,15,111,29,255,26,0,0 // movdqa 0x1aff(%rip),%xmm11 # 4620 <_sk_callback_sse41+0xcfd> .byte 102,69,15,235,235 // por %xmm11,%xmm13 .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 .byte 102,69,15,223,197 // pandn %xmm13,%xmm8 @@ -20399,7 +20419,7 @@ _sk_load_u16_be_sse41: .byte 102,15,235,200 // por %xmm0,%xmm1 .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,102,25,0,0 // movaps 0x1966(%rip),%xmm8 # 45b0 <_sk_callback_sse41+0xca5> + .byte 68,15,40,5,206,25,0,0 // movaps 0x19ce(%rip),%xmm8 # 4630 <_sk_callback_sse41+0xd0d> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 .byte 102,15,113,241,8 // psllw $0x8,%xmm1 @@ -20451,7 +20471,7 @@ _sk_load_rgb_u16_be_sse41: .byte 102,15,235,193 // por %xmm1,%xmm0 .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,167,24,0,0 // movaps 0x18a7(%rip),%xmm8 # 45c0 <_sk_callback_sse41+0xcb5> + .byte 68,15,40,5,15,25,0,0 // movaps 0x190f(%rip),%xmm8 # 4640 <_sk_callback_sse41+0xd1d> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 .byte 102,15,113,241,8 // psllw $0x8,%xmm1 @@ -20468,7 +20488,7 @@ _sk_load_rgb_u16_be_sse41: .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,110,24,0,0 // movaps 0x186e(%rip),%xmm3 # 45d0 <_sk_callback_sse41+0xcc5> + .byte 15,40,29,214,24,0,0 // movaps 0x18d6(%rip),%xmm3 # 4650 <_sk_callback_sse41+0xd2d> .byte 255,224 // jmpq *%rax HIDDEN _sk_store_u16_be_sse41 @@ -20477,7 +20497,7 @@ FUNCTION(_sk_store_u16_be_sse41) _sk_store_u16_be_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,13,111,24,0,0 // movaps 0x186f(%rip),%xmm9 # 45e0 <_sk_callback_sse41+0xcd5> + .byte 68,15,40,13,215,24,0,0 // movaps 0x18d7(%rip),%xmm9 # 4660 <_sk_callback_sse41+0xd3d> .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 69,15,89,193 // mulps %xmm9,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 @@ -20700,10 +20720,10 @@ HIDDEN _sk_luminance_to_alpha_sse41 FUNCTION(_sk_luminance_to_alpha_sse41) _sk_luminance_to_alpha_sse41: .byte 15,40,218 // movaps %xmm2,%xmm3 - .byte 15,89,5,141,21,0,0 // mulps 0x158d(%rip),%xmm0 # 45f0 <_sk_callback_sse41+0xce5> - .byte 15,89,13,150,21,0,0 // mulps 0x1596(%rip),%xmm1 # 4600 <_sk_callback_sse41+0xcf5> + .byte 15,89,5,245,21,0,0 // mulps 0x15f5(%rip),%xmm0 # 4670 <_sk_callback_sse41+0xd4d> + .byte 15,89,13,254,21,0,0 // mulps 0x15fe(%rip),%xmm1 # 4680 <_sk_callback_sse41+0xd5d> .byte 15,88,200 // addps %xmm0,%xmm1 - .byte 15,89,29,156,21,0,0 // mulps 0x159c(%rip),%xmm3 # 4610 <_sk_callback_sse41+0xd05> + .byte 15,89,29,4,22,0,0 // mulps 0x1604(%rip),%xmm3 # 4690 <_sk_callback_sse41+0xd6d> .byte 15,88,217 // addps %xmm1,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 @@ -20936,7 +20956,7 @@ _sk_linear_gradient_sse41: .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 .byte 72,139,8 // mov (%rax),%rcx .byte 72,133,201 // test %rcx,%rcx - .byte 15,132,254,0,0,0 // je 3516 <_sk_linear_gradient_sse41+0x138> + .byte 15,132,254,0,0,0 // je 352e <_sk_linear_gradient_sse41+0x138> .byte 15,41,100,36,168 // movaps %xmm4,-0x58(%rsp) .byte 15,41,108,36,184 // movaps %xmm5,-0x48(%rsp) .byte 15,41,116,36,200 // movaps %xmm6,-0x38(%rsp) @@ -20986,12 +21006,12 @@ _sk_linear_gradient_sse41: .byte 15,40,196 // movaps %xmm4,%xmm0 .byte 72,131,192,36 // add $0x24,%rax .byte 72,255,201 // dec %rcx - .byte 15,133,65,255,255,255 // jne 3441 <_sk_linear_gradient_sse41+0x63> + .byte 15,133,65,255,255,255 // jne 3459 <_sk_linear_gradient_sse41+0x63> .byte 15,40,124,36,216 // movaps -0x28(%rsp),%xmm7 .byte 15,40,116,36,200 // movaps -0x38(%rsp),%xmm6 .byte 15,40,108,36,184 // movaps -0x48(%rsp),%xmm5 .byte 15,40,100,36,168 // movaps -0x58(%rsp),%xmm4 - .byte 235,13 // jmp 3523 <_sk_linear_gradient_sse41+0x145> + .byte 235,13 // jmp 353b <_sk_linear_gradient_sse41+0x145> .byte 15,87,201 // xorps %xmm1,%xmm1 .byte 15,87,210 // xorps %xmm2,%xmm2 .byte 15,87,219 // xorps %xmm3,%xmm3 @@ -21046,7 +21066,7 @@ HIDDEN _sk_save_xy_sse41 FUNCTION(_sk_save_xy_sse41) _sk_save_xy_sse41: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,94,16,0,0 // movaps 0x105e(%rip),%xmm8 # 4620 <_sk_callback_sse41+0xd15> + .byte 68,15,40,5,198,16,0,0 // movaps 0x10c6(%rip),%xmm8 # 46a0 <_sk_callback_sse41+0xd7d> .byte 15,17,0 // movups %xmm0,(%rax) .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,88,200 // addps %xmm8,%xmm9 @@ -21090,8 +21110,8 @@ _sk_bilinear_nx_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,224,15,0,0 // addps 0xfe0(%rip),%xmm0 # 4630 <_sk_callback_sse41+0xd25> - .byte 68,15,40,13,232,15,0,0 // movaps 0xfe8(%rip),%xmm9 # 4640 <_sk_callback_sse41+0xd35> + .byte 15,88,5,72,16,0,0 // addps 0x1048(%rip),%xmm0 # 46b0 <_sk_callback_sse41+0xd8d> + .byte 68,15,40,13,80,16,0,0 // movaps 0x1050(%rip),%xmm9 # 46c0 <_sk_callback_sse41+0xd9d> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21104,7 +21124,7 @@ _sk_bilinear_px_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,215,15,0,0 // addps 0xfd7(%rip),%xmm0 # 4650 <_sk_callback_sse41+0xd45> + .byte 15,88,5,63,16,0,0 // addps 0x103f(%rip),%xmm0 # 46d0 <_sk_callback_sse41+0xdad> .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -21116,8 +21136,8 @@ _sk_bilinear_ny_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,201,15,0,0 // addps 0xfc9(%rip),%xmm1 # 4660 <_sk_callback_sse41+0xd55> - .byte 68,15,40,13,209,15,0,0 // movaps 0xfd1(%rip),%xmm9 # 4670 <_sk_callback_sse41+0xd65> + .byte 15,88,13,49,16,0,0 // addps 0x1031(%rip),%xmm1 # 46e0 <_sk_callback_sse41+0xdbd> + .byte 68,15,40,13,57,16,0,0 // movaps 0x1039(%rip),%xmm9 # 46f0 <_sk_callback_sse41+0xdcd> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21130,7 +21150,7 @@ _sk_bilinear_py_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,191,15,0,0 // addps 0xfbf(%rip),%xmm1 # 4680 <_sk_callback_sse41+0xd75> + .byte 15,88,13,39,16,0,0 // addps 0x1027(%rip),%xmm1 # 4700 <_sk_callback_sse41+0xddd> .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -21142,13 +21162,13 @@ _sk_bicubic_n3x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,178,15,0,0 // addps 0xfb2(%rip),%xmm0 # 4690 <_sk_callback_sse41+0xd85> - .byte 68,15,40,13,186,15,0,0 // movaps 0xfba(%rip),%xmm9 # 46a0 <_sk_callback_sse41+0xd95> + .byte 15,88,5,26,16,0,0 // addps 0x101a(%rip),%xmm0 # 4710 <_sk_callback_sse41+0xded> + .byte 68,15,40,13,34,16,0,0 // movaps 0x1022(%rip),%xmm9 # 4720 <_sk_callback_sse41+0xdfd> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,89,192 // mulps %xmm8,%xmm8 - .byte 68,15,89,13,182,15,0,0 // mulps 0xfb6(%rip),%xmm9 # 46b0 <_sk_callback_sse41+0xda5> - .byte 68,15,88,13,190,15,0,0 // addps 0xfbe(%rip),%xmm9 # 46c0 <_sk_callback_sse41+0xdb5> + .byte 68,15,89,13,30,16,0,0 // mulps 0x101e(%rip),%xmm9 # 4730 <_sk_callback_sse41+0xe0d> + .byte 68,15,88,13,38,16,0,0 // addps 0x1026(%rip),%xmm9 # 4740 <_sk_callback_sse41+0xe1d> .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21161,16 +21181,16 @@ _sk_bicubic_n1x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,173,15,0,0 // addps 0xfad(%rip),%xmm0 # 46d0 <_sk_callback_sse41+0xdc5> - .byte 68,15,40,13,181,15,0,0 // movaps 0xfb5(%rip),%xmm9 # 46e0 <_sk_callback_sse41+0xdd5> + .byte 15,88,5,21,16,0,0 // addps 0x1015(%rip),%xmm0 # 4750 <_sk_callback_sse41+0xe2d> + .byte 68,15,40,13,29,16,0,0 // movaps 0x101d(%rip),%xmm9 # 4760 <_sk_callback_sse41+0xe3d> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,185,15,0,0 // movaps 0xfb9(%rip),%xmm8 # 46f0 <_sk_callback_sse41+0xde5> + .byte 68,15,40,5,33,16,0,0 // movaps 0x1021(%rip),%xmm8 # 4770 <_sk_callback_sse41+0xe4d> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,189,15,0,0 // addps 0xfbd(%rip),%xmm8 # 4700 <_sk_callback_sse41+0xdf5> + .byte 68,15,88,5,37,16,0,0 // addps 0x1025(%rip),%xmm8 # 4780 <_sk_callback_sse41+0xe5d> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,193,15,0,0 // addps 0xfc1(%rip),%xmm8 # 4710 <_sk_callback_sse41+0xe05> + .byte 68,15,88,5,41,16,0,0 // addps 0x1029(%rip),%xmm8 # 4790 <_sk_callback_sse41+0xe6d> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,197,15,0,0 // addps 0xfc5(%rip),%xmm8 # 4720 <_sk_callback_sse41+0xe15> + .byte 68,15,88,5,45,16,0,0 // addps 0x102d(%rip),%xmm8 # 47a0 <_sk_callback_sse41+0xe7d> .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -21180,17 +21200,17 @@ HIDDEN _sk_bicubic_p1x_sse41 FUNCTION(_sk_bicubic_p1x_sse41) _sk_bicubic_p1x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,191,15,0,0 // movaps 0xfbf(%rip),%xmm8 # 4730 <_sk_callback_sse41+0xe25> + .byte 68,15,40,5,39,16,0,0 // movaps 0x1027(%rip),%xmm8 # 47b0 <_sk_callback_sse41+0xe8d> .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 .byte 65,15,88,192 // addps %xmm8,%xmm0 - .byte 68,15,40,21,187,15,0,0 // movaps 0xfbb(%rip),%xmm10 # 4740 <_sk_callback_sse41+0xe35> + .byte 68,15,40,21,35,16,0,0 // movaps 0x1023(%rip),%xmm10 # 47c0 <_sk_callback_sse41+0xe9d> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,191,15,0,0 // addps 0xfbf(%rip),%xmm10 # 4750 <_sk_callback_sse41+0xe45> + .byte 68,15,88,21,39,16,0,0 // addps 0x1027(%rip),%xmm10 # 47d0 <_sk_callback_sse41+0xead> .byte 69,15,89,209 // mulps %xmm9,%xmm10 .byte 69,15,88,208 // addps %xmm8,%xmm10 .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,187,15,0,0 // addps 0xfbb(%rip),%xmm10 # 4760 <_sk_callback_sse41+0xe55> + .byte 68,15,88,21,35,16,0,0 // addps 0x1023(%rip),%xmm10 # 47e0 <_sk_callback_sse41+0xebd> .byte 68,15,17,144,128,0,0,0 // movups %xmm10,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -21202,11 +21222,11 @@ _sk_bicubic_p3x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,174,15,0,0 // addps 0xfae(%rip),%xmm0 # 4770 <_sk_callback_sse41+0xe65> + .byte 15,88,5,22,16,0,0 // addps 0x1016(%rip),%xmm0 # 47f0 <_sk_callback_sse41+0xecd> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,174,15,0,0 // mulps 0xfae(%rip),%xmm8 # 4780 <_sk_callback_sse41+0xe75> - .byte 68,15,88,5,182,15,0,0 // addps 0xfb6(%rip),%xmm8 # 4790 <_sk_callback_sse41+0xe85> + .byte 68,15,89,5,22,16,0,0 // mulps 0x1016(%rip),%xmm8 # 4800 <_sk_callback_sse41+0xedd> + .byte 68,15,88,5,30,16,0,0 // addps 0x101e(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xeed> .byte 69,15,89,193 // mulps %xmm9,%xmm8 .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21219,13 +21239,13 @@ _sk_bicubic_n3y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,164,15,0,0 // addps 0xfa4(%rip),%xmm1 # 47a0 <_sk_callback_sse41+0xe95> - .byte 68,15,40,13,172,15,0,0 // movaps 0xfac(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0xea5> + .byte 15,88,13,12,16,0,0 // addps 0x100c(%rip),%xmm1 # 4820 <_sk_callback_sse41+0xefd> + .byte 68,15,40,13,20,16,0,0 // movaps 0x1014(%rip),%xmm9 # 4830 <_sk_callback_sse41+0xf0d> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,89,192 // mulps %xmm8,%xmm8 - .byte 68,15,89,13,168,15,0,0 // mulps 0xfa8(%rip),%xmm9 # 47c0 <_sk_callback_sse41+0xeb5> - .byte 68,15,88,13,176,15,0,0 // addps 0xfb0(%rip),%xmm9 # 47d0 <_sk_callback_sse41+0xec5> + .byte 68,15,89,13,16,16,0,0 // mulps 0x1010(%rip),%xmm9 # 4840 <_sk_callback_sse41+0xf1d> + .byte 68,15,88,13,24,16,0,0 // addps 0x1018(%rip),%xmm9 # 4850 <_sk_callback_sse41+0xf2d> .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21238,16 +21258,16 @@ _sk_bicubic_n1y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,158,15,0,0 // addps 0xf9e(%rip),%xmm1 # 47e0 <_sk_callback_sse41+0xed5> - .byte 68,15,40,13,166,15,0,0 // movaps 0xfa6(%rip),%xmm9 # 47f0 <_sk_callback_sse41+0xee5> + .byte 15,88,13,6,16,0,0 // addps 0x1006(%rip),%xmm1 # 4860 <_sk_callback_sse41+0xf3d> + .byte 68,15,40,13,14,16,0,0 // movaps 0x100e(%rip),%xmm9 # 4870 <_sk_callback_sse41+0xf4d> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,170,15,0,0 // movaps 0xfaa(%rip),%xmm8 # 4800 <_sk_callback_sse41+0xef5> + .byte 68,15,40,5,18,16,0,0 // movaps 0x1012(%rip),%xmm8 # 4880 <_sk_callback_sse41+0xf5d> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,174,15,0,0 // addps 0xfae(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xf05> + .byte 68,15,88,5,22,16,0,0 // addps 0x1016(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xf6d> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,178,15,0,0 // addps 0xfb2(%rip),%xmm8 # 4820 <_sk_callback_sse41+0xf15> + .byte 68,15,88,5,26,16,0,0 // addps 0x101a(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xf7d> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,182,15,0,0 // addps 0xfb6(%rip),%xmm8 # 4830 <_sk_callback_sse41+0xf25> + .byte 68,15,88,5,30,16,0,0 // addps 0x101e(%rip),%xmm8 # 48b0 <_sk_callback_sse41+0xf8d> .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -21257,17 +21277,17 @@ HIDDEN _sk_bicubic_p1y_sse41 FUNCTION(_sk_bicubic_p1y_sse41) _sk_bicubic_p1y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,176,15,0,0 // movaps 0xfb0(%rip),%xmm8 # 4840 <_sk_callback_sse41+0xf35> + .byte 68,15,40,5,24,16,0,0 // movaps 0x1018(%rip),%xmm8 # 48c0 <_sk_callback_sse41+0xf9d> .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 .byte 65,15,88,200 // addps %xmm8,%xmm1 - .byte 68,15,40,21,171,15,0,0 // movaps 0xfab(%rip),%xmm10 # 4850 <_sk_callback_sse41+0xf45> + .byte 68,15,40,21,19,16,0,0 // movaps 0x1013(%rip),%xmm10 # 48d0 <_sk_callback_sse41+0xfad> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,175,15,0,0 // addps 0xfaf(%rip),%xmm10 # 4860 <_sk_callback_sse41+0xf55> + .byte 68,15,88,21,23,16,0,0 // addps 0x1017(%rip),%xmm10 # 48e0 <_sk_callback_sse41+0xfbd> .byte 69,15,89,209 // mulps %xmm9,%xmm10 .byte 69,15,88,208 // addps %xmm8,%xmm10 .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,171,15,0,0 // addps 0xfab(%rip),%xmm10 # 4870 <_sk_callback_sse41+0xf65> + .byte 68,15,88,21,19,16,0,0 // addps 0x1013(%rip),%xmm10 # 48f0 <_sk_callback_sse41+0xfcd> .byte 68,15,17,144,160,0,0,0 // movups %xmm10,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -21279,11 +21299,11 @@ _sk_bicubic_p3y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,157,15,0,0 // addps 0xf9d(%rip),%xmm1 # 4880 <_sk_callback_sse41+0xf75> + .byte 15,88,13,5,16,0,0 // addps 0x1005(%rip),%xmm1 # 4900 <_sk_callback_sse41+0xfdd> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,157,15,0,0 // mulps 0xf9d(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xf85> - .byte 68,15,88,5,165,15,0,0 // addps 0xfa5(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xf95> + .byte 68,15,89,5,5,16,0,0 // mulps 0x1005(%rip),%xmm8 # 4910 <_sk_callback_sse41+0xfed> + .byte 68,15,88,5,13,16,0,0 // addps 0x100d(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xffd> .byte 69,15,89,193 // mulps %xmm9,%xmm8 .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21468,11 +21488,11 @@ BALIGN16 .byte 0,128,191,0,0,128 // add %al,-0x7fffff41(%rax) .byte 191,0,0,224,64 // mov $0x40e00000,%edi .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 3b68 <.literal16+0x188> + .byte 224,64 // loopne 3b88 <.literal16+0x188> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 3b6c <.literal16+0x18c> + .byte 224,64 // loopne 3b8c <.literal16+0x18c> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 3b70 <.literal16+0x190> + .byte 224,64 // loopne 3b90 <.literal16+0x190> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -21611,6 +21631,49 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .byte 0,0 // add %al,(%rax) + .byte 128,63,171 // cmpb $0xab,(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,171 // ds stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,171 // ds stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,171 // ds stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,0,0 // add %al,%ds:(%rax) + .byte 128,191,0,0,128,191,0 // cmpb $0x0,-0x40800000(%rdi) + .byte 0,128,191,0,0,128 // add %al,-0x7fffff41(%rax) + .byte 191,0,0,192,64 // mov $0x40c00000,%edi + .byte 0,0 // add %al,(%rax) + .byte 192,64,0,0 // rolb $0x0,0x0(%rax) + .byte 192,64,0,0 // rolb $0x0,0x0(%rax) + .byte 192,64,171,170 // rolb $0xaa,-0x55(%rax) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 190,171,170,170,190 // mov $0xbeaaaaab,%esi + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 190,171,170,170,190 // mov $0xbeaaaaab,%esi .byte 129,128,128,59,129,128,128,59,129,128// addl $0x80813b80,-0x7f7ec480(%rax) .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,129,128,128,59 // addb $0x3b,-0x7f7f7ec5(%rax) @@ -21632,13 +21695,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 3cb9 <.literal16+0x2d9> + .byte 224,7 // loopne 3d39 <.literal16+0x339> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 3cbd <.literal16+0x2dd> + .byte 224,7 // loopne 3d3d <.literal16+0x33d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 3cc1 <.literal16+0x2e1> + .byte 224,7 // loopne 3d41 <.literal16+0x341> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 3cc5 <.literal16+0x2e5> + .byte 224,7 // loopne 3d45 <.literal16+0x345> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21678,10 +21741,10 @@ BALIGN16 .byte 0,1 // add %al,(%rcx) .byte 255 // (bad) .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a003d18 <_sk_callback_sse41+0xa00040d> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a003d98 <_sk_callback_sse41+0xa000475> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3003d20 <_sk_callback_sse41+0x3000415> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3003da0 <_sk_callback_sse41+0x300047d> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -21736,11 +21799,11 @@ BALIGN16 .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,127,67 // add %bh,0x43(%rdi) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 3deb <.literal16+0x40b> + .byte 127,67 // jg 3e6b <.literal16+0x46b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 3def <.literal16+0x40f> + .byte 127,67 // jg 3e6f <.literal16+0x46f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 3df3 <.literal16+0x413> + .byte 127,67 // jg 3e73 <.literal16+0x473> .byte 129,128,128,59,129,128,128,59,129,128// addl $0x80813b80,-0x7f7ec480(%rax) .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,129,128,128,59 // addb $0x3b,-0x7f7f7ec5(%rax) @@ -21755,16 +21818,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 3de4 <.literal16+0x404> + .byte 127,0 // jg 3e64 <.literal16+0x464> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3de8 <.literal16+0x408> + .byte 127,0 // jg 3e68 <.literal16+0x468> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3dec <.literal16+0x40c> + .byte 127,0 // jg 3e6c <.literal16+0x46c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3df0 <.literal16+0x410> + .byte 127,0 // jg 3e70 <.literal16+0x470> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -21773,7 +21836,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 3e75 <.literal16+0x495> + .byte 119,115 // ja 3ef5 <.literal16+0x4f5> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -21784,7 +21847,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 3dd9 <.literal16+0x3f9> + .byte 117,191 // jne 3e59 <.literal16+0x459> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -21796,7 +21859,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a37e1a <_sk_callback_sse41+0xffffffffe9a3450f> + .byte 233,220,63,163,233 // jmpq ffffffffe9a37e9a <_sk_callback_sse41+0xffffffffe9a34577> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -21851,16 +21914,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 3eb4 <.literal16+0x4d4> + .byte 127,0 // jg 3f34 <.literal16+0x534> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3eb8 <.literal16+0x4d8> + .byte 127,0 // jg 3f38 <.literal16+0x538> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3ebc <.literal16+0x4dc> + .byte 127,0 // jg 3f3c <.literal16+0x53c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3ec0 <.literal16+0x4e0> + .byte 127,0 // jg 3f40 <.literal16+0x540> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -21869,7 +21932,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 3f45 <.literal16+0x565> + .byte 119,115 // ja 3fc5 <.literal16+0x5c5> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -21880,7 +21943,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 3ea9 <.literal16+0x4c9> + .byte 117,191 // jne 3f29 <.literal16+0x529> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -21892,7 +21955,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a37eea <_sk_callback_sse41+0xffffffffe9a345df> + .byte 233,220,63,163,233 // jmpq ffffffffe9a37f6a <_sk_callback_sse41+0xffffffffe9a34647> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -21947,16 +22010,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 3f84 <.literal16+0x5a4> + .byte 127,0 // jg 4004 <.literal16+0x604> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3f88 <.literal16+0x5a8> + .byte 127,0 // jg 4008 <.literal16+0x608> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3f8c <.literal16+0x5ac> + .byte 127,0 // jg 400c <.literal16+0x60c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 3f90 <.literal16+0x5b0> + .byte 127,0 // jg 4010 <.literal16+0x610> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -21965,7 +22028,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4015 <.literal16+0x635> + .byte 119,115 // ja 4095 <.literal16+0x695> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -21976,7 +22039,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 3f79 <.literal16+0x599> + .byte 117,191 // jne 3ff9 <.literal16+0x5f9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -21988,7 +22051,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a37fba <_sk_callback_sse41+0xffffffffe9a346af> + .byte 233,220,63,163,233 // jmpq ffffffffe9a3803a <_sk_callback_sse41+0xffffffffe9a34717> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -22043,16 +22106,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 4054 <.literal16+0x674> + .byte 127,0 // jg 40d4 <.literal16+0x6d4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4058 <.literal16+0x678> + .byte 127,0 // jg 40d8 <.literal16+0x6d8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 405c <.literal16+0x67c> + .byte 127,0 // jg 40dc <.literal16+0x6dc> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4060 <.literal16+0x680> + .byte 127,0 // jg 40e0 <.literal16+0x6e0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -22061,7 +22124,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 40e5 <.literal16+0x705> + .byte 119,115 // ja 4165 <.literal16+0x765> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -22072,7 +22135,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 4049 <.literal16+0x669> + .byte 117,191 // jne 40c9 <.literal16+0x6c9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -22084,7 +22147,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a3808a <_sk_callback_sse41+0xffffffffe9a3477f> + .byte 233,220,63,163,233 // jmpq ffffffffe9a3810a <_sk_callback_sse41+0xffffffffe9a347e7> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -22135,13 +22198,13 @@ BALIGN16 .byte 200,66,0,0 // enterq $0x42,$0x0 .byte 200,66,0,0 // enterq $0x42,$0x0 .byte 200,66,0,0 // enterq $0x42,$0x0 - .byte 127,67 // jg 4167 <.literal16+0x787> + .byte 127,67 // jg 41e7 <.literal16+0x7e7> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 416b <.literal16+0x78b> + .byte 127,67 // jg 41eb <.literal16+0x7eb> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 416f <.literal16+0x78f> + .byte 127,67 // jg 41ef <.literal16+0x7ef> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 4173 <.literal16+0x793> + .byte 127,67 // jg 41f3 <.literal16+0x7f3> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -22188,16 +22251,16 @@ BALIGN16 .byte 128,3,62 // addb $0x3e,(%rbx) .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 41f3 <.literal16+0x813> + .byte 118,63 // jbe 4273 <.literal16+0x873> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 41f7 <.literal16+0x817> + .byte 118,63 // jbe 4277 <.literal16+0x877> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 41fb <.literal16+0x81b> + .byte 118,63 // jbe 427b <.literal16+0x87b> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 41ff <.literal16+0x81f> + .byte 118,63 // jbe 427f <.literal16+0x87f> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 246,64,83,63 // testb $0x3f,0x53(%rax) @@ -22209,11 +22272,11 @@ BALIGN16 .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,127,67 // add %bh,0x43(%rdi) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 423b <.literal16+0x85b> + .byte 127,67 // jg 42bb <.literal16+0x8bb> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 423f <.literal16+0x85f> + .byte 127,67 // jg 42bf <.literal16+0x8bf> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 4243 <.literal16+0x863> + .byte 127,67 // jg 42c3 <.literal16+0x8c3> .byte 129,128,128,59,129,128,128,59,129,128// addl $0x80813b80,-0x7f7ec480(%rax) .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,0,0,128,63 // addb $0x3f,-0x7fffffc5(%rax) @@ -22242,7 +22305,7 @@ BALIGN16 .byte 5,255,255,255,9 // add $0x9ffffff,%eax .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004270 <_sk_callback_sse41+0x3000965> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 30042f0 <_sk_callback_sse41+0x30009cd> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -22271,13 +22334,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 42a9 <.literal16+0x8c9> + .byte 224,7 // loopne 4329 <.literal16+0x929> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 42ad <.literal16+0x8cd> + .byte 224,7 // loopne 432d <.literal16+0x92d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 42b1 <.literal16+0x8d1> + .byte 224,7 // loopne 4331 <.literal16+0x931> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 42b5 <.literal16+0x8d5> + .byte 224,7 // loopne 4335 <.literal16+0x935> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -22323,13 +22386,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 4319 <.literal16+0x939> + .byte 224,7 // loopne 4399 <.literal16+0x999> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 431d <.literal16+0x93d> + .byte 224,7 // loopne 439d <.literal16+0x99d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 4321 <.literal16+0x941> + .byte 224,7 // loopne 43a1 <.literal16+0x9a1> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 4325 <.literal16+0x945> + .byte 224,7 // loopne 43a5 <.literal16+0x9a5> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -22367,13 +22430,13 @@ BALIGN16 .byte 65,0,0 // add %al,(%r8) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 43b6 <.literal16+0x9d6> + .byte 124,66 // jl 4436 <.literal16+0xa36> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 43ba <.literal16+0x9da> + .byte 124,66 // jl 443a <.literal16+0xa3a> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 43be <.literal16+0x9de> + .byte 124,66 // jl 443e <.literal16+0xa3e> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 43c2 <.literal16+0x9e2> + .byte 124,66 // jl 4442 <.literal16+0xa42> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 0,240 // add %dh,%al @@ -22463,13 +22526,13 @@ BALIGN16 .byte 136,136,61,137,136,136 // mov %cl,-0x777776c3(%rax) .byte 61,137,136,136,61 // cmp $0x3d888889,%eax .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 44c5 <.literal16+0xae5> + .byte 112,65 // jo 4545 <.literal16+0xb45> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 44c9 <.literal16+0xae9> + .byte 112,65 // jo 4549 <.literal16+0xb49> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 44cd <.literal16+0xaed> + .byte 112,65 // jo 454d <.literal16+0xb4d> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 44d1 <.literal16+0xaf1> + .byte 112,65 // jo 4551 <.literal16+0xb51> .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -22484,7 +22547,7 @@ BALIGN16 .byte 5,255,255,255,9 // add $0x9ffffff,%eax .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 30044c0 <_sk_callback_sse41+0x3000bb5> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004540 <_sk_callback_sse41+0x3000c1d> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -22511,7 +22574,7 @@ BALIGN16 .byte 5,255,255,255,9 // add $0x9ffffff,%eax .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004500 <_sk_callback_sse41+0x3000bf5> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004580 <_sk_callback_sse41+0x3000c5d> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -22526,11 +22589,11 @@ BALIGN16 .byte 255,0 // incl (%rax) .byte 0,127,67 // add %bh,0x43(%rdi) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 455b <.literal16+0xb7b> + .byte 127,67 // jg 45db <.literal16+0xbdb> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 455f <.literal16+0xb7f> + .byte 127,67 // jg 45df <.literal16+0xbdf> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 4563 <.literal16+0xb83> + .byte 127,67 // jg 45e3 <.literal16+0xbe3> .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) .byte 0,0 // add %al,(%rax) .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) @@ -22606,13 +22669,13 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 255 // (bad) - .byte 127,71 // jg 462b <.literal16+0xc4b> + .byte 127,71 // jg 46ab <.literal16+0xcab> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 462f <.literal16+0xc4f> + .byte 127,71 // jg 46af <.literal16+0xcaf> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 4633 <.literal16+0xc53> + .byte 127,71 // jg 46b3 <.literal16+0xcb3> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 4637 <.literal16+0xc57> + .byte 127,71 // jg 46b7 <.literal16+0xcb7> .byte 208 // (bad) .byte 179,89 // mov $0x59,%bl .byte 62,208 // ds (bad) @@ -22696,11 +22759,11 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,114 // cmpb $0x72,(%rdi) .byte 28,199 // sbb $0xc7,%al - .byte 62,114,28 // jb,pt 46d2 <.literal16+0xcf2> + .byte 62,114,28 // jb,pt 4752 <.literal16+0xd52> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 46d6 <.literal16+0xcf6> + .byte 62,114,28 // jb,pt 4756 <.literal16+0xd56> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 46da <.literal16+0xcfa> + .byte 62,114,28 // jb,pt 475a <.literal16+0xd5a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -22744,7 +22807,7 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d565 <_sk_callback_sse41+0x3d639c5a> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d5e5 <_sk_callback_sse41+0x3d639cc2> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -22770,7 +22833,7 @@ BALIGN16 .byte 0,192 // add %al,%al .byte 63 // (bad) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d5a5 <_sk_callback_sse41+0x3d639c9a> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d625 <_sk_callback_sse41+0x3d639d02> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -22779,13 +22842,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 479e <.literal16+0xdbe> + .byte 114,28 // jb 481e <.literal16+0xe1e> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 47a2 <.literal16+0xdc2> + .byte 62,114,28 // jb,pt 4822 <.literal16+0xe22> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 47a6 <.literal16+0xdc6> + .byte 62,114,28 // jb,pt 4826 <.literal16+0xe26> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 47aa <.literal16+0xdca> + .byte 62,114,28 // jb,pt 482a <.literal16+0xe2a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -22806,11 +22869,11 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,114 // cmpb $0x72,(%rdi) .byte 28,199 // sbb $0xc7,%al - .byte 62,114,28 // jb,pt 47e2 <.literal16+0xe02> + .byte 62,114,28 // jb,pt 4862 <.literal16+0xe62> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 47e6 <.literal16+0xe06> + .byte 62,114,28 // jb,pt 4866 <.literal16+0xe66> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 47ea <.literal16+0xe0a> + .byte 62,114,28 // jb,pt 486a <.literal16+0xe6a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -22854,7 +22917,7 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d675 <_sk_callback_sse41+0x3d639d6a> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d6f5 <_sk_callback_sse41+0x3d639dd2> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -22880,7 +22943,7 @@ BALIGN16 .byte 0,192 // add %al,%al .byte 63 // (bad) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d6b5 <_sk_callback_sse41+0x3d639daa> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d735 <_sk_callback_sse41+0x3d639e12> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -22889,13 +22952,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 48ae <.literal16+0xece> + .byte 114,28 // jb 492e <.literal16+0xf2e> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 48b2 <_sk_callback_sse41+0xfa7> + .byte 62,114,28 // jb,pt 4932 <_sk_callback_sse41+0x100f> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 48b6 <_sk_callback_sse41+0xfab> + .byte 62,114,28 // jb,pt 4936 <_sk_callback_sse41+0x1013> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 48ba <_sk_callback_sse41+0xfaf> + .byte 62,114,28 // jb,pt 493a <_sk_callback_sse41+0x1017> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -22965,7 +23028,7 @@ _sk_seed_shader_sse2: .byte 102,15,110,199 // movd %edi,%xmm0 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 - .byte 15,40,21,100,61,0,0 // movaps 0x3d64(%rip),%xmm2 # 3de0 <_sk_callback_sse2+0xd8> + .byte 15,40,21,132,61,0,0 // movaps 0x3d84(%rip),%xmm2 # 3e00 <_sk_callback_sse2+0xd9> .byte 15,88,202 // addps %xmm2,%xmm1 .byte 15,16,2 // movups (%rdx),%xmm0 .byte 15,88,193 // addps %xmm1,%xmm0 @@ -22974,7 +23037,7 @@ _sk_seed_shader_sse2: .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 15,88,202 // addps %xmm2,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,21,83,61,0,0 // movaps 0x3d53(%rip),%xmm2 # 3df0 <_sk_callback_sse2+0xe8> + .byte 15,40,21,115,61,0,0 // movaps 0x3d73(%rip),%xmm2 # 3e10 <_sk_callback_sse2+0xe9> .byte 15,87,219 // xorps %xmm3,%xmm3 .byte 15,87,228 // xorps %xmm4,%xmm4 .byte 15,87,237 // xorps %xmm5,%xmm5 @@ -23014,7 +23077,7 @@ HIDDEN _sk_srcatop_sse2 FUNCTION(_sk_srcatop_sse2) _sk_srcatop_sse2: .byte 15,89,199 // mulps %xmm7,%xmm0 - .byte 68,15,40,5,14,61,0,0 // movaps 0x3d0e(%rip),%xmm8 # 3e00 <_sk_callback_sse2+0xf8> + .byte 68,15,40,5,46,61,0,0 // movaps 0x3d2e(%rip),%xmm8 # 3e20 <_sk_callback_sse2+0xf9> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -23039,7 +23102,7 @@ FUNCTION(_sk_dstatop_sse2) _sk_dstatop_sse2: .byte 68,15,40,195 // movaps %xmm3,%xmm8 .byte 68,15,89,196 // mulps %xmm4,%xmm8 - .byte 68,15,40,13,209,60,0,0 // movaps 0x3cd1(%rip),%xmm9 # 3e10 <_sk_callback_sse2+0x108> + .byte 68,15,40,13,241,60,0,0 // movaps 0x3cf1(%rip),%xmm9 # 3e30 <_sk_callback_sse2+0x109> .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,88,192 // addps %xmm8,%xmm0 @@ -23086,7 +23149,7 @@ HIDDEN _sk_srcout_sse2 .globl _sk_srcout_sse2 FUNCTION(_sk_srcout_sse2) _sk_srcout_sse2: - .byte 68,15,40,5,117,60,0,0 // movaps 0x3c75(%rip),%xmm8 # 3e20 <_sk_callback_sse2+0x118> + .byte 68,15,40,5,149,60,0,0 // movaps 0x3c95(%rip),%xmm8 # 3e40 <_sk_callback_sse2+0x119> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 @@ -23099,7 +23162,7 @@ HIDDEN _sk_dstout_sse2 .globl _sk_dstout_sse2 FUNCTION(_sk_dstout_sse2) _sk_dstout_sse2: - .byte 68,15,40,5,101,60,0,0 // movaps 0x3c65(%rip),%xmm8 # 3e30 <_sk_callback_sse2+0x128> + .byte 68,15,40,5,133,60,0,0 // movaps 0x3c85(%rip),%xmm8 # 3e50 <_sk_callback_sse2+0x129> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 15,89,196 // mulps %xmm4,%xmm0 @@ -23116,7 +23179,7 @@ HIDDEN _sk_srcover_sse2 .globl _sk_srcover_sse2 FUNCTION(_sk_srcover_sse2) _sk_srcover_sse2: - .byte 68,15,40,5,72,60,0,0 // movaps 0x3c48(%rip),%xmm8 # 3e40 <_sk_callback_sse2+0x138> + .byte 68,15,40,5,104,60,0,0 // movaps 0x3c68(%rip),%xmm8 # 3e60 <_sk_callback_sse2+0x139> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -23136,7 +23199,7 @@ HIDDEN _sk_dstover_sse2 .globl _sk_dstover_sse2 FUNCTION(_sk_dstover_sse2) _sk_dstover_sse2: - .byte 68,15,40,5,28,60,0,0 // movaps 0x3c1c(%rip),%xmm8 # 3e50 <_sk_callback_sse2+0x148> + .byte 68,15,40,5,60,60,0,0 // movaps 0x3c3c(%rip),%xmm8 # 3e70 <_sk_callback_sse2+0x149> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -23164,7 +23227,7 @@ HIDDEN _sk_multiply_sse2 .globl _sk_multiply_sse2 FUNCTION(_sk_multiply_sse2) _sk_multiply_sse2: - .byte 68,15,40,5,240,59,0,0 // movaps 0x3bf0(%rip),%xmm8 # 3e60 <_sk_callback_sse2+0x158> + .byte 68,15,40,5,16,60,0,0 // movaps 0x3c10(%rip),%xmm8 # 3e80 <_sk_callback_sse2+0x159> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 69,15,40,209 // movaps %xmm9,%xmm10 @@ -23240,7 +23303,7 @@ HIDDEN _sk_xor__sse2 FUNCTION(_sk_xor__sse2) _sk_xor__sse2: .byte 68,15,40,195 // movaps %xmm3,%xmm8 - .byte 15,40,29,33,59,0,0 // movaps 0x3b21(%rip),%xmm3 # 3e70 <_sk_callback_sse2+0x168> + .byte 15,40,29,65,59,0,0 // movaps 0x3b41(%rip),%xmm3 # 3e90 <_sk_callback_sse2+0x169> .byte 68,15,40,203 // movaps %xmm3,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 @@ -23288,7 +23351,7 @@ _sk_darken_sse2: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,95,209 // maxps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,140,58,0,0 // movaps 0x3a8c(%rip),%xmm2 # 3e80 <_sk_callback_sse2+0x178> + .byte 15,40,21,172,58,0,0 // movaps 0x3aac(%rip),%xmm2 # 3ea0 <_sk_callback_sse2+0x179> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -23322,7 +23385,7 @@ _sk_lighten_sse2: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,49,58,0,0 // movaps 0x3a31(%rip),%xmm2 # 3e90 <_sk_callback_sse2+0x188> + .byte 15,40,21,81,58,0,0 // movaps 0x3a51(%rip),%xmm2 # 3eb0 <_sk_callback_sse2+0x189> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -23359,7 +23422,7 @@ _sk_difference_sse2: .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,203,57,0,0 // movaps 0x39cb(%rip),%xmm2 # 3ea0 <_sk_callback_sse2+0x198> + .byte 15,40,21,235,57,0,0 // movaps 0x39eb(%rip),%xmm2 # 3ec0 <_sk_callback_sse2+0x199> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -23386,7 +23449,7 @@ _sk_exclusion_sse2: .byte 15,89,214 // mulps %xmm6,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,202 // subps %xmm2,%xmm9 - .byte 15,40,13,140,57,0,0 // movaps 0x398c(%rip),%xmm1 # 3eb0 <_sk_callback_sse2+0x1a8> + .byte 15,40,13,172,57,0,0 // movaps 0x39ac(%rip),%xmm1 # 3ed0 <_sk_callback_sse2+0x1a9> .byte 15,92,203 // subps %xmm3,%xmm1 .byte 15,89,207 // mulps %xmm7,%xmm1 .byte 15,88,217 // addps %xmm1,%xmm3 @@ -23400,7 +23463,7 @@ HIDDEN _sk_colorburn_sse2 FUNCTION(_sk_colorburn_sse2) _sk_colorburn_sse2: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,40,21,123,57,0,0 // movaps 0x397b(%rip),%xmm10 # 3ec0 <_sk_callback_sse2+0x1b8> + .byte 68,15,40,21,155,57,0,0 // movaps 0x399b(%rip),%xmm10 # 3ee0 <_sk_callback_sse2+0x1b9> .byte 69,15,40,202 // movaps %xmm10,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 69,15,40,217 // movaps %xmm9,%xmm11 @@ -23494,7 +23557,7 @@ HIDDEN _sk_colordodge_sse2 FUNCTION(_sk_colordodge_sse2) _sk_colordodge_sse2: .byte 68,15,40,200 // movaps %xmm0,%xmm9 - .byte 68,15,40,21,49,56,0,0 // movaps 0x3831(%rip),%xmm10 # 3ed0 <_sk_callback_sse2+0x1c8> + .byte 68,15,40,21,81,56,0,0 // movaps 0x3851(%rip),%xmm10 # 3ef0 <_sk_callback_sse2+0x1c9> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 69,15,40,227 // movaps %xmm11,%xmm12 @@ -23588,7 +23651,7 @@ _sk_hardlight_sse2: .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp) .byte 15,40,245 // movaps %xmm5,%xmm6 .byte 15,40,236 // movaps %xmm4,%xmm5 - .byte 68,15,40,29,230,54,0,0 // movaps 0x36e6(%rip),%xmm11 # 3ee0 <_sk_callback_sse2+0x1d8> + .byte 68,15,40,29,6,55,0,0 // movaps 0x3706(%rip),%xmm11 # 3f00 <_sk_callback_sse2+0x1d9> .byte 69,15,40,211 // movaps %xmm11,%xmm10 .byte 68,15,92,215 // subps %xmm7,%xmm10 .byte 69,15,40,194 // movaps %xmm10,%xmm8 @@ -23676,7 +23739,7 @@ FUNCTION(_sk_overlay_sse2) _sk_overlay_sse2: .byte 68,15,40,193 // movaps %xmm1,%xmm8 .byte 68,15,40,232 // movaps %xmm0,%xmm13 - .byte 68,15,40,13,180,53,0,0 // movaps 0x35b4(%rip),%xmm9 # 3ef0 <_sk_callback_sse2+0x1e8> + .byte 68,15,40,13,212,53,0,0 // movaps 0x35d4(%rip),%xmm9 # 3f10 <_sk_callback_sse2+0x1e9> .byte 69,15,40,209 // movaps %xmm9,%xmm10 .byte 68,15,92,215 // subps %xmm7,%xmm10 .byte 69,15,40,218 // movaps %xmm10,%xmm11 @@ -23767,7 +23830,7 @@ _sk_softlight_sse2: .byte 68,15,40,213 // movaps %xmm5,%xmm10 .byte 68,15,94,215 // divps %xmm7,%xmm10 .byte 69,15,84,212 // andps %xmm12,%xmm10 - .byte 68,15,40,13,113,52,0,0 // movaps 0x3471(%rip),%xmm9 # 3f00 <_sk_callback_sse2+0x1f8> + .byte 68,15,40,13,145,52,0,0 // movaps 0x3491(%rip),%xmm9 # 3f20 <_sk_callback_sse2+0x1f9> .byte 69,15,40,249 // movaps %xmm9,%xmm15 .byte 69,15,92,250 // subps %xmm10,%xmm15 .byte 69,15,40,218 // movaps %xmm10,%xmm11 @@ -23780,10 +23843,10 @@ _sk_softlight_sse2: .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 15,89,192 // mulps %xmm0,%xmm0 .byte 65,15,88,194 // addps %xmm10,%xmm0 - .byte 68,15,40,53,75,52,0,0 // movaps 0x344b(%rip),%xmm14 # 3f10 <_sk_callback_sse2+0x208> + .byte 68,15,40,53,107,52,0,0 // movaps 0x346b(%rip),%xmm14 # 3f30 <_sk_callback_sse2+0x209> .byte 69,15,88,222 // addps %xmm14,%xmm11 .byte 68,15,89,216 // mulps %xmm0,%xmm11 - .byte 68,15,40,21,75,52,0,0 // movaps 0x344b(%rip),%xmm10 # 3f20 <_sk_callback_sse2+0x218> + .byte 68,15,40,21,107,52,0,0 // movaps 0x346b(%rip),%xmm10 # 3f40 <_sk_callback_sse2+0x219> .byte 69,15,89,234 // mulps %xmm10,%xmm13 .byte 69,15,88,235 // addps %xmm11,%xmm13 .byte 15,88,228 // addps %xmm4,%xmm4 @@ -23935,7 +23998,7 @@ HIDDEN _sk_clamp_1_sse2 .globl _sk_clamp_1_sse2 FUNCTION(_sk_clamp_1_sse2) _sk_clamp_1_sse2: - .byte 68,15,40,5,90,50,0,0 // movaps 0x325a(%rip),%xmm8 # 3f30 <_sk_callback_sse2+0x228> + .byte 68,15,40,5,122,50,0,0 // movaps 0x327a(%rip),%xmm8 # 3f50 <_sk_callback_sse2+0x229> .byte 65,15,93,192 // minps %xmm8,%xmm0 .byte 65,15,93,200 // minps %xmm8,%xmm1 .byte 65,15,93,208 // minps %xmm8,%xmm2 @@ -23947,7 +24010,7 @@ HIDDEN _sk_clamp_a_sse2 .globl _sk_clamp_a_sse2 FUNCTION(_sk_clamp_a_sse2) _sk_clamp_a_sse2: - .byte 15,93,29,79,50,0,0 // minps 0x324f(%rip),%xmm3 # 3f40 <_sk_callback_sse2+0x238> + .byte 15,93,29,111,50,0,0 // minps 0x326f(%rip),%xmm3 # 3f60 <_sk_callback_sse2+0x239> .byte 15,93,195 // minps %xmm3,%xmm0 .byte 15,93,203 // minps %xmm3,%xmm1 .byte 15,93,211 // minps %xmm3,%xmm2 @@ -24034,7 +24097,7 @@ HIDDEN _sk_unpremul_sse2 FUNCTION(_sk_unpremul_sse2) _sk_unpremul_sse2: .byte 69,15,87,192 // xorps %xmm8,%xmm8 - .byte 68,15,40,13,186,49,0,0 // movaps 0x31ba(%rip),%xmm9 # 3f50 <_sk_callback_sse2+0x248> + .byte 68,15,40,13,218,49,0,0 // movaps 0x31da(%rip),%xmm9 # 3f70 <_sk_callback_sse2+0x249> .byte 68,15,94,203 // divps %xmm3,%xmm9 .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8 .byte 69,15,84,193 // andps %xmm9,%xmm8 @@ -24048,20 +24111,20 @@ HIDDEN _sk_from_srgb_sse2 .globl _sk_from_srgb_sse2 FUNCTION(_sk_from_srgb_sse2) _sk_from_srgb_sse2: - .byte 68,15,40,5,165,49,0,0 // movaps 0x31a5(%rip),%xmm8 # 3f60 <_sk_callback_sse2+0x258> + .byte 68,15,40,5,197,49,0,0 // movaps 0x31c5(%rip),%xmm8 # 3f80 <_sk_callback_sse2+0x259> .byte 68,15,40,232 // movaps %xmm0,%xmm13 .byte 69,15,89,232 // mulps %xmm8,%xmm13 .byte 68,15,40,216 // movaps %xmm0,%xmm11 .byte 69,15,89,219 // mulps %xmm11,%xmm11 - .byte 68,15,40,13,157,49,0,0 // movaps 0x319d(%rip),%xmm9 # 3f70 <_sk_callback_sse2+0x268> + .byte 68,15,40,13,189,49,0,0 // movaps 0x31bd(%rip),%xmm9 # 3f90 <_sk_callback_sse2+0x269> .byte 68,15,40,240 // movaps %xmm0,%xmm14 .byte 69,15,89,241 // mulps %xmm9,%xmm14 - .byte 68,15,40,21,157,49,0,0 // movaps 0x319d(%rip),%xmm10 # 3f80 <_sk_callback_sse2+0x278> + .byte 68,15,40,21,189,49,0,0 // movaps 0x31bd(%rip),%xmm10 # 3fa0 <_sk_callback_sse2+0x279> .byte 69,15,88,242 // addps %xmm10,%xmm14 .byte 69,15,89,243 // mulps %xmm11,%xmm14 - .byte 68,15,40,29,157,49,0,0 // movaps 0x319d(%rip),%xmm11 # 3f90 <_sk_callback_sse2+0x288> + .byte 68,15,40,29,189,49,0,0 // movaps 0x31bd(%rip),%xmm11 # 3fb0 <_sk_callback_sse2+0x289> .byte 69,15,88,243 // addps %xmm11,%xmm14 - .byte 68,15,40,37,161,49,0,0 // movaps 0x31a1(%rip),%xmm12 # 3fa0 <_sk_callback_sse2+0x298> + .byte 68,15,40,37,193,49,0,0 // movaps 0x31c1(%rip),%xmm12 # 3fc0 <_sk_callback_sse2+0x299> .byte 65,15,194,196,1 // cmpltps %xmm12,%xmm0 .byte 68,15,84,232 // andps %xmm0,%xmm13 .byte 65,15,85,198 // andnps %xmm14,%xmm0 @@ -24100,20 +24163,20 @@ _sk_to_srgb_sse2: .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 .byte 69,15,83,200 // rcpps %xmm8,%xmm9 .byte 69,15,82,232 // rsqrtps %xmm8,%xmm13 - .byte 68,15,40,5,38,49,0,0 // movaps 0x3126(%rip),%xmm8 # 3fb0 <_sk_callback_sse2+0x2a8> + .byte 68,15,40,5,70,49,0,0 // movaps 0x3146(%rip),%xmm8 # 3fd0 <_sk_callback_sse2+0x2a9> .byte 68,15,40,240 // movaps %xmm0,%xmm14 .byte 69,15,89,240 // mulps %xmm8,%xmm14 - .byte 68,15,40,21,38,49,0,0 // movaps 0x3126(%rip),%xmm10 # 3fc0 <_sk_callback_sse2+0x2b8> + .byte 68,15,40,21,70,49,0,0 // movaps 0x3146(%rip),%xmm10 # 3fe0 <_sk_callback_sse2+0x2b9> .byte 69,15,89,202 // mulps %xmm10,%xmm9 - .byte 68,15,40,29,42,49,0,0 // movaps 0x312a(%rip),%xmm11 # 3fd0 <_sk_callback_sse2+0x2c8> + .byte 68,15,40,29,74,49,0,0 // movaps 0x314a(%rip),%xmm11 # 3ff0 <_sk_callback_sse2+0x2c9> .byte 69,15,88,203 // addps %xmm11,%xmm9 - .byte 68,15,40,37,46,49,0,0 // movaps 0x312e(%rip),%xmm12 # 3fe0 <_sk_callback_sse2+0x2d8> + .byte 68,15,40,37,78,49,0,0 // movaps 0x314e(%rip),%xmm12 # 4000 <_sk_callback_sse2+0x2d9> .byte 69,15,89,236 // mulps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,40,13,46,49,0,0 // movaps 0x312e(%rip),%xmm9 # 3ff0 <_sk_callback_sse2+0x2e8> + .byte 68,15,40,13,78,49,0,0 // movaps 0x314e(%rip),%xmm9 # 4010 <_sk_callback_sse2+0x2e9> .byte 69,15,40,249 // movaps %xmm9,%xmm15 .byte 69,15,93,253 // minps %xmm13,%xmm15 - .byte 68,15,40,45,46,49,0,0 // movaps 0x312e(%rip),%xmm13 # 4000 <_sk_callback_sse2+0x2f8> + .byte 68,15,40,45,78,49,0,0 // movaps 0x314e(%rip),%xmm13 # 4020 <_sk_callback_sse2+0x2f9> .byte 65,15,194,197,1 // cmpltps %xmm13,%xmm0 .byte 68,15,84,240 // andps %xmm0,%xmm14 .byte 65,15,85,199 // andnps %xmm15,%xmm0 @@ -24163,7 +24226,7 @@ _sk_rgb_to_hsl_sse2: .byte 68,15,93,218 // minps %xmm2,%xmm11 .byte 65,15,40,202 // movaps %xmm10,%xmm1 .byte 65,15,92,203 // subps %xmm11,%xmm1 - .byte 68,15,40,45,135,48,0,0 // movaps 0x3087(%rip),%xmm13 # 4010 <_sk_callback_sse2+0x308> + .byte 68,15,40,45,167,48,0,0 // movaps 0x30a7(%rip),%xmm13 # 4030 <_sk_callback_sse2+0x309> .byte 68,15,94,233 // divps %xmm1,%xmm13 .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 @@ -24172,30 +24235,30 @@ _sk_rgb_to_hsl_sse2: .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,40,241 // movaps %xmm9,%xmm14 .byte 68,15,194,242,1 // cmpltps %xmm2,%xmm14 - .byte 68,15,84,53,109,48,0,0 // andps 0x306d(%rip),%xmm14 # 4020 <_sk_callback_sse2+0x318> + .byte 68,15,84,53,141,48,0,0 // andps 0x308d(%rip),%xmm14 # 4040 <_sk_callback_sse2+0x319> .byte 69,15,88,244 // addps %xmm12,%xmm14 .byte 69,15,40,250 // movaps %xmm10,%xmm15 .byte 69,15,194,249,0 // cmpeqps %xmm9,%xmm15 .byte 65,15,92,208 // subps %xmm8,%xmm2 .byte 65,15,89,213 // mulps %xmm13,%xmm2 - .byte 68,15,40,37,96,48,0,0 // movaps 0x3060(%rip),%xmm12 # 4030 <_sk_callback_sse2+0x328> + .byte 68,15,40,37,128,48,0,0 // movaps 0x3080(%rip),%xmm12 # 4050 <_sk_callback_sse2+0x329> .byte 65,15,88,212 // addps %xmm12,%xmm2 .byte 69,15,92,193 // subps %xmm9,%xmm8 .byte 69,15,89,197 // mulps %xmm13,%xmm8 - .byte 68,15,88,5,92,48,0,0 // addps 0x305c(%rip),%xmm8 # 4040 <_sk_callback_sse2+0x338> + .byte 68,15,88,5,124,48,0,0 // addps 0x307c(%rip),%xmm8 # 4060 <_sk_callback_sse2+0x339> .byte 65,15,84,215 // andps %xmm15,%xmm2 .byte 69,15,85,248 // andnps %xmm8,%xmm15 .byte 68,15,86,250 // orps %xmm2,%xmm15 .byte 68,15,84,240 // andps %xmm0,%xmm14 .byte 65,15,85,199 // andnps %xmm15,%xmm0 .byte 65,15,86,198 // orps %xmm14,%xmm0 - .byte 15,89,5,77,48,0,0 // mulps 0x304d(%rip),%xmm0 # 4050 <_sk_callback_sse2+0x348> + .byte 15,89,5,109,48,0,0 // mulps 0x306d(%rip),%xmm0 # 4070 <_sk_callback_sse2+0x349> .byte 69,15,40,194 // movaps %xmm10,%xmm8 .byte 69,15,194,195,4 // cmpneqps %xmm11,%xmm8 .byte 65,15,84,192 // andps %xmm8,%xmm0 .byte 69,15,92,226 // subps %xmm10,%xmm12 .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 68,15,40,13,64,48,0,0 // movaps 0x3040(%rip),%xmm9 # 4060 <_sk_callback_sse2+0x358> + .byte 68,15,40,13,96,48,0,0 // movaps 0x3060(%rip),%xmm9 # 4080 <_sk_callback_sse2+0x359> .byte 65,15,40,210 // movaps %xmm10,%xmm2 .byte 65,15,89,209 // mulps %xmm9,%xmm2 .byte 68,15,194,202,1 // cmpltps %xmm2,%xmm9 @@ -24212,180 +24275,185 @@ HIDDEN _sk_hsl_to_rgb_sse2 .globl _sk_hsl_to_rgb_sse2 FUNCTION(_sk_hsl_to_rgb_sse2) _sk_hsl_to_rgb_sse2: - .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp) - .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp) - .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp) - .byte 15,41,100,36,184 // movaps %xmm4,-0x48(%rsp) - .byte 15,41,92,36,168 // movaps %xmm3,-0x58(%rsp) + .byte 72,131,236,40 // sub $0x28,%rsp + .byte 15,41,124,36,16 // movaps %xmm7,0x10(%rsp) + .byte 15,41,52,36 // movaps %xmm6,(%rsp) + .byte 15,41,108,36,240 // movaps %xmm5,-0x10(%rsp) + .byte 15,41,100,36,224 // movaps %xmm4,-0x20(%rsp) + .byte 15,41,92,36,208 // movaps %xmm3,-0x30(%rsp) .byte 68,15,40,210 // movaps %xmm2,%xmm10 - .byte 15,40,209 // movaps %xmm1,%xmm2 - .byte 15,40,240 // movaps %xmm0,%xmm6 + .byte 15,40,217 // movaps %xmm1,%xmm3 + .byte 15,40,232 // movaps %xmm0,%xmm5 .byte 184,0,0,0,63 // mov $0x3f000000,%eax - .byte 102,68,15,110,240 // movd %eax,%xmm14 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,40,202 // movaps %xmm10,%xmm9 - .byte 69,15,194,206,1 // cmpltps %xmm14,%xmm9 - .byte 184,0,0,128,63 // mov $0x3f800000,%eax - .byte 102,68,15,110,224 // movd %eax,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 15,88,194 // addps %xmm2,%xmm0 - .byte 65,15,89,194 // mulps %xmm10,%xmm0 - .byte 15,40,218 // movaps %xmm2,%xmm3 - .byte 69,15,87,219 // xorps %xmm11,%xmm11 - .byte 68,15,194,218,0 // cmpeqps %xmm2,%xmm11 - .byte 65,15,88,210 // addps %xmm10,%xmm2 + .byte 102,15,110,192 // movd %eax,%xmm0 + .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 + .byte 69,15,40,194 // movaps %xmm10,%xmm8 + .byte 68,15,194,192,1 // cmpltps %xmm0,%xmm8 + .byte 68,15,40,216 // movaps %xmm0,%xmm11 + .byte 68,15,40,37,255,47,0,0 // movaps 0x2fff(%rip),%xmm12 # 4090 <_sk_callback_sse2+0x369> + .byte 15,40,195 // movaps %xmm3,%xmm0 + .byte 15,40,211 // movaps %xmm3,%xmm2 + .byte 15,87,201 // xorps %xmm1,%xmm1 + .byte 15,194,203,0 // cmpeqps %xmm3,%xmm1 + .byte 15,41,76,36,176 // movaps %xmm1,-0x50(%rsp) + .byte 65,15,88,220 // addps %xmm12,%xmm3 .byte 65,15,89,218 // mulps %xmm10,%xmm3 - .byte 15,92,211 // subps %xmm3,%xmm2 - .byte 65,15,84,193 // andps %xmm9,%xmm0 - .byte 68,15,85,202 // andnps %xmm2,%xmm9 - .byte 68,15,86,200 // orps %xmm0,%xmm9 - .byte 184,0,0,0,64 // mov $0x40000000,%eax - .byte 185,171,170,170,62 // mov $0x3eaaaaab,%ecx - .byte 102,68,15,110,193 // movd %ecx,%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,41,68,36,152 // movaps %xmm8,-0x68(%rsp) - .byte 68,15,88,198 // addps %xmm6,%xmm8 - .byte 185,0,0,0,0 // mov $0x0,%ecx - .byte 102,15,110,233 // movd %ecx,%xmm5 - .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5 - .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 65,15,194,192,1 // cmpltps %xmm8,%xmm0 - .byte 65,15,40,216 // movaps %xmm8,%xmm3 - .byte 65,15,92,220 // subps %xmm12,%xmm3 - .byte 15,84,216 // andps %xmm0,%xmm3 - .byte 65,15,85,192 // andnps %xmm8,%xmm0 - .byte 15,86,195 // orps %xmm3,%xmm0 - .byte 65,15,40,216 // movaps %xmm8,%xmm3 - .byte 15,194,221,1 // cmpltps %xmm5,%xmm3 - .byte 65,15,40,212 // movaps %xmm12,%xmm2 - .byte 65,15,88,208 // addps %xmm8,%xmm2 - .byte 15,84,211 // andps %xmm3,%xmm2 - .byte 15,85,216 // andnps %xmm0,%xmm3 - .byte 15,86,218 // orps %xmm2,%xmm3 - .byte 102,68,15,110,232 // movd %eax,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 69,15,92,233 // subps %xmm9,%xmm13 + .byte 65,15,88,194 // addps %xmm10,%xmm0 + .byte 65,15,89,210 // mulps %xmm10,%xmm2 + .byte 15,92,194 // subps %xmm2,%xmm0 + .byte 65,15,84,216 // andps %xmm8,%xmm3 + .byte 68,15,85,192 // andnps %xmm0,%xmm8 + .byte 68,15,86,195 // orps %xmm3,%xmm8 + .byte 15,40,29,215,47,0,0 // movaps 0x2fd7(%rip),%xmm3 # 40a0 <_sk_callback_sse2+0x379> + .byte 15,88,221 // addps %xmm5,%xmm3 + .byte 184,0,0,0,0 // mov $0x0,%eax + .byte 185,0,0,128,63 // mov $0x3f800000,%ecx + .byte 102,68,15,110,241 // movd %ecx,%xmm14 + .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 + .byte 65,15,40,214 // movaps %xmm14,%xmm2 + .byte 15,194,211,1 // cmpltps %xmm3,%xmm2 + .byte 68,15,40,61,192,47,0,0 // movaps 0x2fc0(%rip),%xmm15 # 40b0 <_sk_callback_sse2+0x389> + .byte 15,40,195 // movaps %xmm3,%xmm0 + .byte 65,15,88,199 // addps %xmm15,%xmm0 + .byte 15,84,194 // andps %xmm2,%xmm0 + .byte 15,85,211 // andnps %xmm3,%xmm2 + .byte 15,86,208 // orps %xmm0,%xmm2 + .byte 102,15,110,200 // movd %eax,%xmm1 + .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 + .byte 15,41,76,36,128 // movaps %xmm1,-0x80(%rsp) + .byte 15,40,195 // movaps %xmm3,%xmm0 + .byte 15,194,193,1 // cmpltps %xmm1,%xmm0 + .byte 15,40,227 // movaps %xmm3,%xmm4 + .byte 65,15,88,228 // addps %xmm12,%xmm4 + .byte 15,84,224 // andps %xmm0,%xmm4 + .byte 15,85,194 // andnps %xmm2,%xmm0 + .byte 15,86,196 // orps %xmm4,%xmm0 + .byte 69,15,40,234 // movaps %xmm10,%xmm13 + .byte 69,15,88,237 // addps %xmm13,%xmm13 + .byte 69,15,92,232 // subps %xmm8,%xmm13 .byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax - .byte 65,15,40,193 // movaps %xmm9,%xmm0 - .byte 65,15,92,197 // subps %xmm13,%xmm0 - .byte 185,0,0,192,64 // mov $0x40c00000,%ecx - .byte 102,68,15,110,249 // movd %ecx,%xmm15 - .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 68,15,89,248 // mulps %xmm0,%xmm15 + .byte 69,15,40,200 // movaps %xmm8,%xmm9 + .byte 69,15,92,205 // subps %xmm13,%xmm9 + .byte 68,15,89,13,123,47,0,0 // mulps 0x2f7b(%rip),%xmm9 # 40c0 <_sk_callback_sse2+0x399> .byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx - .byte 102,15,110,225 // movd %ecx,%xmm4 - .byte 15,198,228,0 // shufps $0x0,%xmm4,%xmm4 - .byte 15,40,212 // movaps %xmm4,%xmm2 - .byte 15,92,211 // subps %xmm3,%xmm2 - .byte 15,40,203 // movaps %xmm3,%xmm1 - .byte 15,40,195 // movaps %xmm3,%xmm0 - .byte 15,194,220,1 // cmpltps %xmm4,%xmm3 - .byte 65,15,89,215 // mulps %xmm15,%xmm2 - .byte 65,15,88,213 // addps %xmm13,%xmm2 - .byte 15,84,211 // andps %xmm3,%xmm2 - .byte 65,15,85,221 // andnps %xmm13,%xmm3 - .byte 15,86,218 // orps %xmm2,%xmm3 - .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 - .byte 65,15,40,209 // movaps %xmm9,%xmm2 - .byte 15,84,208 // andps %xmm0,%xmm2 - .byte 15,85,195 // andnps %xmm3,%xmm0 - .byte 15,86,194 // orps %xmm2,%xmm0 - .byte 102,15,110,248 // movd %eax,%xmm7 + .byte 102,15,110,249 // movd %ecx,%xmm7 .byte 15,198,255,0 // shufps $0x0,%xmm7,%xmm7 + .byte 15,41,124,36,144 // movaps %xmm7,-0x70(%rsp) + .byte 15,40,53,114,47,0,0 // movaps 0x2f72(%rip),%xmm6 # 40d0 <_sk_callback_sse2+0x3a9> + .byte 15,40,230 // movaps %xmm6,%xmm4 + .byte 15,92,224 // subps %xmm0,%xmm4 + .byte 15,40,208 // movaps %xmm0,%xmm2 + .byte 15,40,200 // movaps %xmm0,%xmm1 + .byte 15,194,199,1 // cmpltps %xmm7,%xmm0 + .byte 65,15,89,225 // mulps %xmm9,%xmm4 + .byte 65,15,88,229 // addps %xmm13,%xmm4 + .byte 15,84,224 // andps %xmm0,%xmm4 + .byte 65,15,85,197 // andnps %xmm13,%xmm0 + .byte 15,86,196 // orps %xmm4,%xmm0 + .byte 65,15,40,251 // movaps %xmm11,%xmm7 + .byte 15,41,124,36,160 // movaps %xmm7,-0x60(%rsp) .byte 15,194,207,1 // cmpltps %xmm7,%xmm1 - .byte 69,15,89,199 // mulps %xmm15,%xmm8 - .byte 69,15,88,197 // addps %xmm13,%xmm8 - .byte 68,15,84,193 // andps %xmm1,%xmm8 + .byte 65,15,40,224 // movaps %xmm8,%xmm4 + .byte 15,84,225 // andps %xmm1,%xmm4 .byte 15,85,200 // andnps %xmm0,%xmm1 - .byte 65,15,86,200 // orps %xmm8,%xmm1 - .byte 69,15,40,195 // movaps %xmm11,%xmm8 - .byte 68,15,85,193 // andnps %xmm1,%xmm8 - .byte 65,15,40,196 // movaps %xmm12,%xmm0 - .byte 15,194,198,1 // cmpltps %xmm6,%xmm0 - .byte 15,40,206 // movaps %xmm6,%xmm1 - .byte 65,15,92,204 // subps %xmm12,%xmm1 + .byte 15,86,204 // orps %xmm4,%xmm1 + .byte 102,15,110,224 // movd %eax,%xmm4 + .byte 15,198,228,0 // shufps $0x0,%xmm4,%xmm4 + .byte 15,194,212,1 // cmpltps %xmm4,%xmm2 + .byte 65,15,89,217 // mulps %xmm9,%xmm3 + .byte 65,15,88,221 // addps %xmm13,%xmm3 + .byte 15,84,218 // andps %xmm2,%xmm3 + .byte 15,85,209 // andnps %xmm1,%xmm2 + .byte 15,86,211 // orps %xmm3,%xmm2 + .byte 68,15,40,92,36,176 // movaps -0x50(%rsp),%xmm11 + .byte 65,15,40,195 // movaps %xmm11,%xmm0 + .byte 15,85,194 // andnps %xmm2,%xmm0 + .byte 15,41,68,36,192 // movaps %xmm0,-0x40(%rsp) + .byte 65,15,40,198 // movaps %xmm14,%xmm0 + .byte 15,194,197,1 // cmpltps %xmm5,%xmm0 + .byte 15,40,205 // movaps %xmm5,%xmm1 + .byte 65,15,88,207 // addps %xmm15,%xmm1 .byte 15,84,200 // andps %xmm0,%xmm1 - .byte 15,85,198 // andnps %xmm6,%xmm0 + .byte 15,85,197 // andnps %xmm5,%xmm0 .byte 15,86,193 // orps %xmm1,%xmm0 - .byte 15,40,206 // movaps %xmm6,%xmm1 - .byte 15,194,205,1 // cmpltps %xmm5,%xmm1 - .byte 65,15,40,212 // movaps %xmm12,%xmm2 - .byte 15,88,214 // addps %xmm6,%xmm2 + .byte 15,40,205 // movaps %xmm5,%xmm1 + .byte 15,194,76,36,128,1 // cmpltps -0x80(%rsp),%xmm1 + .byte 15,40,213 // movaps %xmm5,%xmm2 + .byte 65,15,88,212 // addps %xmm12,%xmm2 .byte 15,84,209 // andps %xmm1,%xmm2 .byte 15,85,200 // andnps %xmm0,%xmm1 .byte 15,86,202 // orps %xmm2,%xmm1 - .byte 15,40,196 // movaps %xmm4,%xmm0 + .byte 15,40,198 // movaps %xmm6,%xmm0 .byte 15,92,193 // subps %xmm1,%xmm0 - .byte 15,40,217 // movaps %xmm1,%xmm3 .byte 15,40,209 // movaps %xmm1,%xmm2 - .byte 15,194,204,1 // cmpltps %xmm4,%xmm1 - .byte 65,15,89,199 // mulps %xmm15,%xmm0 + .byte 15,40,217 // movaps %xmm1,%xmm3 + .byte 15,194,76,36,144,1 // cmpltps -0x70(%rsp),%xmm1 + .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,88,197 // addps %xmm13,%xmm0 .byte 15,84,193 // andps %xmm1,%xmm0 .byte 65,15,85,205 // andnps %xmm13,%xmm1 .byte 15,86,200 // orps %xmm0,%xmm1 - .byte 65,15,194,214,1 // cmpltps %xmm14,%xmm2 - .byte 65,15,40,193 // movaps %xmm9,%xmm0 - .byte 15,84,194 // andps %xmm2,%xmm0 - .byte 15,85,209 // andnps %xmm1,%xmm2 - .byte 15,86,208 // orps %xmm0,%xmm2 .byte 15,194,223,1 // cmpltps %xmm7,%xmm3 - .byte 65,15,40,199 // movaps %xmm15,%xmm0 - .byte 15,89,198 // mulps %xmm6,%xmm0 - .byte 65,15,88,197 // addps %xmm13,%xmm0 + .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 15,84,195 // andps %xmm3,%xmm0 - .byte 15,85,218 // andnps %xmm2,%xmm3 + .byte 15,85,217 // andnps %xmm1,%xmm3 .byte 15,86,216 // orps %xmm0,%xmm3 - .byte 65,15,40,203 // movaps %xmm11,%xmm1 - .byte 15,85,203 // andnps %xmm3,%xmm1 - .byte 15,92,116,36,152 // subps -0x68(%rsp),%xmm6 - .byte 15,40,198 // movaps %xmm6,%xmm0 - .byte 15,194,197,1 // cmpltps %xmm5,%xmm0 - .byte 15,40,214 // movaps %xmm6,%xmm2 - .byte 65,15,92,212 // subps %xmm12,%xmm2 - .byte 65,15,40,220 // movaps %xmm12,%xmm3 - .byte 68,15,194,230,1 // cmpltps %xmm6,%xmm12 - .byte 65,15,84,212 // andps %xmm12,%xmm2 - .byte 68,15,85,230 // andnps %xmm6,%xmm12 - .byte 68,15,86,226 // orps %xmm2,%xmm12 - .byte 15,88,222 // addps %xmm6,%xmm3 - .byte 15,84,216 // andps %xmm0,%xmm3 - .byte 65,15,85,196 // andnps %xmm12,%xmm0 - .byte 15,86,195 // orps %xmm3,%xmm0 - .byte 15,40,232 // movaps %xmm0,%xmm5 - .byte 15,194,239,1 // cmpltps %xmm7,%xmm5 - .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 15,194,212,1 // cmpltps %xmm4,%xmm2 - .byte 15,92,224 // subps %xmm0,%xmm4 - .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 - .byte 65,15,89,247 // mulps %xmm15,%xmm6 - .byte 65,15,89,231 // mulps %xmm15,%xmm4 + .byte 65,15,40,193 // movaps %xmm9,%xmm0 + .byte 15,89,197 // mulps %xmm5,%xmm0 + .byte 65,15,88,197 // addps %xmm13,%xmm0 + .byte 15,84,194 // andps %xmm2,%xmm0 + .byte 15,85,211 // andnps %xmm3,%xmm2 + .byte 15,86,208 // orps %xmm0,%xmm2 + .byte 65,15,40,219 // movaps %xmm11,%xmm3 + .byte 15,85,218 // andnps %xmm2,%xmm3 + .byte 15,88,45,139,46,0,0 // addps 0x2e8b(%rip),%xmm5 # 40e0 <_sk_callback_sse2+0x3b9> + .byte 15,40,197 // movaps %xmm5,%xmm0 + .byte 15,194,68,36,128,1 // cmpltps -0x80(%rsp),%xmm0 + .byte 68,15,194,245,1 // cmpltps %xmm5,%xmm14 + .byte 68,15,88,253 // addps %xmm5,%xmm15 + .byte 69,15,84,254 // andps %xmm14,%xmm15 + .byte 68,15,85,245 // andnps %xmm5,%xmm14 + .byte 69,15,86,247 // orps %xmm15,%xmm14 + .byte 68,15,88,229 // addps %xmm5,%xmm12 + .byte 68,15,84,224 // andps %xmm0,%xmm12 + .byte 65,15,85,198 // andnps %xmm14,%xmm0 + .byte 65,15,86,196 // orps %xmm12,%xmm0 + .byte 15,40,248 // movaps %xmm0,%xmm7 + .byte 15,194,252,1 // cmpltps %xmm4,%xmm7 + .byte 15,40,200 // movaps %xmm0,%xmm1 + .byte 15,194,76,36,160,1 // cmpltps -0x60(%rsp),%xmm1 + .byte 15,92,240 // subps %xmm0,%xmm6 + .byte 15,194,68,36,144,1 // cmpltps -0x70(%rsp),%xmm0 + .byte 65,15,89,233 // mulps %xmm9,%xmm5 + .byte 65,15,89,241 // mulps %xmm9,%xmm6 + .byte 65,15,88,237 // addps %xmm13,%xmm5 .byte 65,15,88,245 // addps %xmm13,%xmm6 - .byte 65,15,88,229 // addps %xmm13,%xmm4 - .byte 15,84,226 // andps %xmm2,%xmm4 - .byte 65,15,85,213 // andnps %xmm13,%xmm2 - .byte 15,86,212 // orps %xmm4,%xmm2 - .byte 68,15,84,200 // andps %xmm0,%xmm9 - .byte 15,85,194 // andnps %xmm2,%xmm0 - .byte 65,15,86,193 // orps %xmm9,%xmm0 - .byte 15,84,245 // andps %xmm5,%xmm6 - .byte 15,85,232 // andnps %xmm0,%xmm5 - .byte 15,86,238 // orps %xmm6,%xmm5 + .byte 15,84,240 // andps %xmm0,%xmm6 + .byte 65,15,85,197 // andnps %xmm13,%xmm0 + .byte 15,86,198 // orps %xmm6,%xmm0 + .byte 68,15,84,193 // andps %xmm1,%xmm8 + .byte 15,85,200 // andnps %xmm0,%xmm1 + .byte 65,15,86,200 // orps %xmm8,%xmm1 + .byte 15,84,239 // andps %xmm7,%xmm5 + .byte 15,85,249 // andnps %xmm1,%xmm7 + .byte 15,86,253 // orps %xmm5,%xmm7 .byte 69,15,84,211 // andps %xmm11,%xmm10 - .byte 68,15,85,221 // andnps %xmm5,%xmm11 - .byte 69,15,86,194 // orps %xmm10,%xmm8 + .byte 68,15,85,223 // andnps %xmm7,%xmm11 + .byte 15,40,76,36,192 // movaps -0x40(%rsp),%xmm1 .byte 65,15,86,202 // orps %xmm10,%xmm1 + .byte 65,15,86,218 // orps %xmm10,%xmm3 .byte 69,15,86,211 // orps %xmm11,%xmm10 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 65,15,40,192 // movaps %xmm8,%xmm0 + .byte 15,40,193 // movaps %xmm1,%xmm0 + .byte 15,40,203 // movaps %xmm3,%xmm1 .byte 65,15,40,210 // movaps %xmm10,%xmm2 - .byte 15,40,92,36,168 // movaps -0x58(%rsp),%xmm3 - .byte 15,40,100,36,184 // movaps -0x48(%rsp),%xmm4 - .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5 - .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6 - .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7 + .byte 15,40,92,36,208 // movaps -0x30(%rsp),%xmm3 + .byte 15,40,100,36,224 // movaps -0x20(%rsp),%xmm4 + .byte 15,40,108,36,240 // movaps -0x10(%rsp),%xmm5 + .byte 15,40,52,36 // movaps (%rsp),%xmm6 + .byte 15,40,124,36,16 // movaps 0x10(%rsp),%xmm7 + .byte 72,131,196,40 // add $0x28,%rsp .byte 255,224 // jmpq *%rax HIDDEN _sk_scale_1_float_sse2 @@ -24413,7 +24481,7 @@ _sk_scale_u8_sse2: .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8 .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,60,45,0,0 // mulps 0x2d3c(%rip),%xmm8 # 4070 <_sk_callback_sse2+0x368> + .byte 68,15,89,5,157,45,0,0 // mulps 0x2d9d(%rip),%xmm8 # 40f0 <_sk_callback_sse2+0x3c9> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 65,15,89,208 // mulps %xmm8,%xmm2 @@ -24454,7 +24522,7 @@ _sk_lerp_u8_sse2: .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8 .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,218,44,0,0 // mulps 0x2cda(%rip),%xmm8 # 4080 <_sk_callback_sse2+0x378> + .byte 68,15,89,5,59,45,0,0 // mulps 0x2d3b(%rip),%xmm8 # 4100 <_sk_callback_sse2+0x3d9> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -24479,17 +24547,17 @@ _sk_lerp_565_sse2: .byte 243,68,15,126,4,120 // movq (%rax,%rdi,2),%xmm8 .byte 102,15,239,219 // pxor %xmm3,%xmm3 .byte 102,68,15,97,195 // punpcklwd %xmm3,%xmm8 - .byte 102,15,111,29,162,44,0,0 // movdqa 0x2ca2(%rip),%xmm3 # 4090 <_sk_callback_sse2+0x388> + .byte 102,15,111,29,3,45,0,0 // movdqa 0x2d03(%rip),%xmm3 # 4110 <_sk_callback_sse2+0x3e9> .byte 102,65,15,219,216 // pand %xmm8,%xmm3 .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 - .byte 68,15,89,13,161,44,0,0 // mulps 0x2ca1(%rip),%xmm9 # 40a0 <_sk_callback_sse2+0x398> - .byte 102,15,111,29,169,44,0,0 // movdqa 0x2ca9(%rip),%xmm3 # 40b0 <_sk_callback_sse2+0x3a8> + .byte 68,15,89,13,2,45,0,0 // mulps 0x2d02(%rip),%xmm9 # 4120 <_sk_callback_sse2+0x3f9> + .byte 102,15,111,29,10,45,0,0 // movdqa 0x2d0a(%rip),%xmm3 # 4130 <_sk_callback_sse2+0x409> .byte 102,65,15,219,216 // pand %xmm8,%xmm3 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,170,44,0,0 // mulps 0x2caa(%rip),%xmm3 # 40c0 <_sk_callback_sse2+0x3b8> - .byte 102,68,15,219,5,177,44,0,0 // pand 0x2cb1(%rip),%xmm8 # 40d0 <_sk_callback_sse2+0x3c8> + .byte 15,89,29,11,45,0,0 // mulps 0x2d0b(%rip),%xmm3 # 4140 <_sk_callback_sse2+0x419> + .byte 102,68,15,219,5,18,45,0,0 // pand 0x2d12(%rip),%xmm8 # 4150 <_sk_callback_sse2+0x429> .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,181,44,0,0 // mulps 0x2cb5(%rip),%xmm8 # 40e0 <_sk_callback_sse2+0x3d8> + .byte 68,15,89,5,22,45,0,0 // mulps 0x2d16(%rip),%xmm8 # 4160 <_sk_callback_sse2+0x439> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -24500,7 +24568,7 @@ _sk_lerp_565_sse2: .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 15,88,214 // addps %xmm6,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,159,44,0,0 // movaps 0x2c9f(%rip),%xmm3 # 40f0 <_sk_callback_sse2+0x3e8> + .byte 15,40,29,0,45,0,0 // movaps 0x2d00(%rip),%xmm3 # 4170 <_sk_callback_sse2+0x449> .byte 255,224 // jmpq *%rax HIDDEN _sk_load_tables_sse2 @@ -24511,7 +24579,7 @@ _sk_load_tables_sse2: .byte 76,139,0 // mov (%rax),%r8 .byte 76,139,72,8 // mov 0x8(%rax),%r9 .byte 243,69,15,111,12,184 // movdqu (%r8,%rdi,4),%xmm9 - .byte 102,68,15,111,5,149,44,0,0 // movdqa 0x2c95(%rip),%xmm8 # 4100 <_sk_callback_sse2+0x3f8> + .byte 102,68,15,111,5,246,44,0,0 // movdqa 0x2cf6(%rip),%xmm8 # 4180 <_sk_callback_sse2+0x459> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 @@ -24566,7 +24634,7 @@ _sk_load_tables_sse2: .byte 65,15,20,208 // unpcklps %xmm8,%xmm2 .byte 102,65,15,114,209,24 // psrld $0x18,%xmm9 .byte 65,15,91,217 // cvtdq2ps %xmm9,%xmm3 - .byte 15,89,29,162,43,0,0 // mulps 0x2ba2(%rip),%xmm3 # 4110 <_sk_callback_sse2+0x408> + .byte 15,89,29,3,44,0,0 // mulps 0x2c03(%rip),%xmm3 # 4190 <_sk_callback_sse2+0x469> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24585,7 +24653,7 @@ _sk_load_tables_u16_be_sse2: .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9 - .byte 102,68,15,111,21,117,43,0,0 // movdqa 0x2b75(%rip),%xmm10 # 4120 <_sk_callback_sse2+0x418> + .byte 102,68,15,111,21,214,43,0,0 // movdqa 0x2bd6(%rip),%xmm10 # 41a0 <_sk_callback_sse2+0x479> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,194 // pand %xmm10,%xmm0 .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 @@ -24646,7 +24714,7 @@ _sk_load_tables_u16_be_sse2: .byte 102,65,15,235,217 // por %xmm9,%xmm3 .byte 102,65,15,97,216 // punpcklwd %xmm8,%xmm3 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,100,42,0,0 // mulps 0x2a64(%rip),%xmm3 # 4130 <_sk_callback_sse2+0x428> + .byte 15,89,29,197,42,0,0 // mulps 0x2ac5(%rip),%xmm3 # 41b0 <_sk_callback_sse2+0x489> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24668,7 +24736,7 @@ _sk_load_tables_rgb_u16_be_sse2: .byte 102,68,15,97,208 // punpcklwd %xmm0,%xmm10 .byte 102,65,15,111,195 // movdqa %xmm11,%xmm0 .byte 102,65,15,97,194 // punpcklwd %xmm10,%xmm0 - .byte 102,68,15,111,5,36,42,0,0 // movdqa 0x2a24(%rip),%xmm8 # 4140 <_sk_callback_sse2+0x438> + .byte 102,68,15,111,5,133,42,0,0 // movdqa 0x2a85(%rip),%xmm8 # 41c0 <_sk_callback_sse2+0x499> .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 @@ -24723,7 +24791,7 @@ _sk_load_tables_rgb_u16_be_sse2: .byte 15,20,211 // unpcklps %xmm3,%xmm2 .byte 65,15,20,208 // unpcklps %xmm8,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,51,41,0,0 // movaps 0x2933(%rip),%xmm3 # 4150 <_sk_callback_sse2+0x448> + .byte 15,40,29,148,41,0,0 // movaps 0x2994(%rip),%xmm3 # 41d0 <_sk_callback_sse2+0x4a9> .byte 255,224 // jmpq *%rax HIDDEN _sk_byte_tables_sse2 @@ -24733,7 +24801,7 @@ _sk_byte_tables_sse2: .byte 65,86 // push %r14 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,52,41,0,0 // movaps 0x2934(%rip),%xmm8 # 4160 <_sk_callback_sse2+0x458> + .byte 68,15,40,5,149,41,0,0 // movaps 0x2995(%rip),%xmm8 # 41e0 <_sk_callback_sse2+0x4b9> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 .byte 102,72,15,126,193 // movq %xmm0,%rcx @@ -24760,7 +24828,7 @@ _sk_byte_tables_sse2: .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,21,209,40,0,0 // movaps 0x28d1(%rip),%xmm10 # 4170 <_sk_callback_sse2+0x468> + .byte 68,15,40,21,50,41,0,0 // movaps 0x2932(%rip),%xmm10 # 41f0 <_sk_callback_sse2+0x4c9> .byte 65,15,89,194 // mulps %xmm10,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -24876,7 +24944,7 @@ _sk_byte_tables_rgb_sse2: .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,21,36,39,0,0 // movaps 0x2724(%rip),%xmm10 # 4180 <_sk_callback_sse2+0x478> + .byte 68,15,40,21,133,39,0,0 // movaps 0x2785(%rip),%xmm10 # 4200 <_sk_callback_sse2+0x4d9> .byte 65,15,89,194 // mulps %xmm10,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -25073,15 +25141,15 @@ _sk_parametric_r_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,99,36,0,0 // mulps 0x2463(%rip),%xmm9 # 4190 <_sk_callback_sse2+0x488> - .byte 68,15,84,21,107,36,0,0 // andps 0x246b(%rip),%xmm10 # 41a0 <_sk_callback_sse2+0x498> - .byte 68,15,86,21,115,36,0,0 // orps 0x2473(%rip),%xmm10 # 41b0 <_sk_callback_sse2+0x4a8> - .byte 68,15,88,13,123,36,0,0 // addps 0x247b(%rip),%xmm9 # 41c0 <_sk_callback_sse2+0x4b8> - .byte 68,15,40,37,131,36,0,0 // movaps 0x2483(%rip),%xmm12 # 41d0 <_sk_callback_sse2+0x4c8> + .byte 68,15,89,13,196,36,0,0 // mulps 0x24c4(%rip),%xmm9 # 4210 <_sk_callback_sse2+0x4e9> + .byte 68,15,84,21,204,36,0,0 // andps 0x24cc(%rip),%xmm10 # 4220 <_sk_callback_sse2+0x4f9> + .byte 68,15,86,21,212,36,0,0 // orps 0x24d4(%rip),%xmm10 # 4230 <_sk_callback_sse2+0x509> + .byte 68,15,88,13,220,36,0,0 // addps 0x24dc(%rip),%xmm9 # 4240 <_sk_callback_sse2+0x519> + .byte 68,15,40,37,228,36,0,0 // movaps 0x24e4(%rip),%xmm12 # 4250 <_sk_callback_sse2+0x529> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,131,36,0,0 // addps 0x2483(%rip),%xmm10 # 41e0 <_sk_callback_sse2+0x4d8> - .byte 68,15,40,37,139,36,0,0 // movaps 0x248b(%rip),%xmm12 # 41f0 <_sk_callback_sse2+0x4e8> + .byte 68,15,88,21,228,36,0,0 // addps 0x24e4(%rip),%xmm10 # 4260 <_sk_callback_sse2+0x539> + .byte 68,15,40,37,236,36,0,0 // movaps 0x24ec(%rip),%xmm12 # 4270 <_sk_callback_sse2+0x549> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -25089,22 +25157,22 @@ _sk_parametric_r_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,117,36,0,0 // movaps 0x2475(%rip),%xmm10 # 4200 <_sk_callback_sse2+0x4f8> + .byte 68,15,40,21,214,36,0,0 // movaps 0x24d6(%rip),%xmm10 # 4280 <_sk_callback_sse2+0x559> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,105,36,0,0 // addps 0x2469(%rip),%xmm9 # 4210 <_sk_callback_sse2+0x508> - .byte 68,15,40,37,113,36,0,0 // movaps 0x2471(%rip),%xmm12 # 4220 <_sk_callback_sse2+0x518> + .byte 68,15,88,13,202,36,0,0 // addps 0x24ca(%rip),%xmm9 # 4290 <_sk_callback_sse2+0x569> + .byte 68,15,40,37,210,36,0,0 // movaps 0x24d2(%rip),%xmm12 # 42a0 <_sk_callback_sse2+0x579> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,113,36,0,0 // movaps 0x2471(%rip),%xmm12 # 4230 <_sk_callback_sse2+0x528> + .byte 68,15,40,37,210,36,0,0 // movaps 0x24d2(%rip),%xmm12 # 42b0 <_sk_callback_sse2+0x589> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,117,36,0,0 // movaps 0x2475(%rip),%xmm13 # 4240 <_sk_callback_sse2+0x538> + .byte 68,15,40,45,214,36,0,0 // movaps 0x24d6(%rip),%xmm13 # 42c0 <_sk_callback_sse2+0x599> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,117,36,0,0 // mulps 0x2475(%rip),%xmm13 # 4250 <_sk_callback_sse2+0x548> + .byte 68,15,89,45,214,36,0,0 // mulps 0x24d6(%rip),%xmm13 # 42d0 <_sk_callback_sse2+0x5a9> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -25140,15 +25208,15 @@ _sk_parametric_g_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,245,35,0,0 // mulps 0x23f5(%rip),%xmm9 # 4260 <_sk_callback_sse2+0x558> - .byte 68,15,84,21,253,35,0,0 // andps 0x23fd(%rip),%xmm10 # 4270 <_sk_callback_sse2+0x568> - .byte 68,15,86,21,5,36,0,0 // orps 0x2405(%rip),%xmm10 # 4280 <_sk_callback_sse2+0x578> - .byte 68,15,88,13,13,36,0,0 // addps 0x240d(%rip),%xmm9 # 4290 <_sk_callback_sse2+0x588> - .byte 68,15,40,37,21,36,0,0 // movaps 0x2415(%rip),%xmm12 # 42a0 <_sk_callback_sse2+0x598> + .byte 68,15,89,13,86,36,0,0 // mulps 0x2456(%rip),%xmm9 # 42e0 <_sk_callback_sse2+0x5b9> + .byte 68,15,84,21,94,36,0,0 // andps 0x245e(%rip),%xmm10 # 42f0 <_sk_callback_sse2+0x5c9> + .byte 68,15,86,21,102,36,0,0 // orps 0x2466(%rip),%xmm10 # 4300 <_sk_callback_sse2+0x5d9> + .byte 68,15,88,13,110,36,0,0 // addps 0x246e(%rip),%xmm9 # 4310 <_sk_callback_sse2+0x5e9> + .byte 68,15,40,37,118,36,0,0 // movaps 0x2476(%rip),%xmm12 # 4320 <_sk_callback_sse2+0x5f9> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,21,36,0,0 // addps 0x2415(%rip),%xmm10 # 42b0 <_sk_callback_sse2+0x5a8> - .byte 68,15,40,37,29,36,0,0 // movaps 0x241d(%rip),%xmm12 # 42c0 <_sk_callback_sse2+0x5b8> + .byte 68,15,88,21,118,36,0,0 // addps 0x2476(%rip),%xmm10 # 4330 <_sk_callback_sse2+0x609> + .byte 68,15,40,37,126,36,0,0 // movaps 0x247e(%rip),%xmm12 # 4340 <_sk_callback_sse2+0x619> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -25156,22 +25224,22 @@ _sk_parametric_g_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,7,36,0,0 // movaps 0x2407(%rip),%xmm10 # 42d0 <_sk_callback_sse2+0x5c8> + .byte 68,15,40,21,104,36,0,0 // movaps 0x2468(%rip),%xmm10 # 4350 <_sk_callback_sse2+0x629> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,251,35,0,0 // addps 0x23fb(%rip),%xmm9 # 42e0 <_sk_callback_sse2+0x5d8> - .byte 68,15,40,37,3,36,0,0 // movaps 0x2403(%rip),%xmm12 # 42f0 <_sk_callback_sse2+0x5e8> + .byte 68,15,88,13,92,36,0,0 // addps 0x245c(%rip),%xmm9 # 4360 <_sk_callback_sse2+0x639> + .byte 68,15,40,37,100,36,0,0 // movaps 0x2464(%rip),%xmm12 # 4370 <_sk_callback_sse2+0x649> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,3,36,0,0 // movaps 0x2403(%rip),%xmm12 # 4300 <_sk_callback_sse2+0x5f8> + .byte 68,15,40,37,100,36,0,0 // movaps 0x2464(%rip),%xmm12 # 4380 <_sk_callback_sse2+0x659> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,7,36,0,0 // movaps 0x2407(%rip),%xmm13 # 4310 <_sk_callback_sse2+0x608> + .byte 68,15,40,45,104,36,0,0 // movaps 0x2468(%rip),%xmm13 # 4390 <_sk_callback_sse2+0x669> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,7,36,0,0 // mulps 0x2407(%rip),%xmm13 # 4320 <_sk_callback_sse2+0x618> + .byte 68,15,89,45,104,36,0,0 // mulps 0x2468(%rip),%xmm13 # 43a0 <_sk_callback_sse2+0x679> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -25207,15 +25275,15 @@ _sk_parametric_b_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,135,35,0,0 // mulps 0x2387(%rip),%xmm9 # 4330 <_sk_callback_sse2+0x628> - .byte 68,15,84,21,143,35,0,0 // andps 0x238f(%rip),%xmm10 # 4340 <_sk_callback_sse2+0x638> - .byte 68,15,86,21,151,35,0,0 // orps 0x2397(%rip),%xmm10 # 4350 <_sk_callback_sse2+0x648> - .byte 68,15,88,13,159,35,0,0 // addps 0x239f(%rip),%xmm9 # 4360 <_sk_callback_sse2+0x658> - .byte 68,15,40,37,167,35,0,0 // movaps 0x23a7(%rip),%xmm12 # 4370 <_sk_callback_sse2+0x668> + .byte 68,15,89,13,232,35,0,0 // mulps 0x23e8(%rip),%xmm9 # 43b0 <_sk_callback_sse2+0x689> + .byte 68,15,84,21,240,35,0,0 // andps 0x23f0(%rip),%xmm10 # 43c0 <_sk_callback_sse2+0x699> + .byte 68,15,86,21,248,35,0,0 // orps 0x23f8(%rip),%xmm10 # 43d0 <_sk_callback_sse2+0x6a9> + .byte 68,15,88,13,0,36,0,0 // addps 0x2400(%rip),%xmm9 # 43e0 <_sk_callback_sse2+0x6b9> + .byte 68,15,40,37,8,36,0,0 // movaps 0x2408(%rip),%xmm12 # 43f0 <_sk_callback_sse2+0x6c9> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,167,35,0,0 // addps 0x23a7(%rip),%xmm10 # 4380 <_sk_callback_sse2+0x678> - .byte 68,15,40,37,175,35,0,0 // movaps 0x23af(%rip),%xmm12 # 4390 <_sk_callback_sse2+0x688> + .byte 68,15,88,21,8,36,0,0 // addps 0x2408(%rip),%xmm10 # 4400 <_sk_callback_sse2+0x6d9> + .byte 68,15,40,37,16,36,0,0 // movaps 0x2410(%rip),%xmm12 # 4410 <_sk_callback_sse2+0x6e9> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -25223,22 +25291,22 @@ _sk_parametric_b_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,153,35,0,0 // movaps 0x2399(%rip),%xmm10 # 43a0 <_sk_callback_sse2+0x698> + .byte 68,15,40,21,250,35,0,0 // movaps 0x23fa(%rip),%xmm10 # 4420 <_sk_callback_sse2+0x6f9> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,141,35,0,0 // addps 0x238d(%rip),%xmm9 # 43b0 <_sk_callback_sse2+0x6a8> - .byte 68,15,40,37,149,35,0,0 // movaps 0x2395(%rip),%xmm12 # 43c0 <_sk_callback_sse2+0x6b8> + .byte 68,15,88,13,238,35,0,0 // addps 0x23ee(%rip),%xmm9 # 4430 <_sk_callback_sse2+0x709> + .byte 68,15,40,37,246,35,0,0 // movaps 0x23f6(%rip),%xmm12 # 4440 <_sk_callback_sse2+0x719> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,149,35,0,0 // movaps 0x2395(%rip),%xmm12 # 43d0 <_sk_callback_sse2+0x6c8> + .byte 68,15,40,37,246,35,0,0 // movaps 0x23f6(%rip),%xmm12 # 4450 <_sk_callback_sse2+0x729> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,153,35,0,0 // movaps 0x2399(%rip),%xmm13 # 43e0 <_sk_callback_sse2+0x6d8> + .byte 68,15,40,45,250,35,0,0 // movaps 0x23fa(%rip),%xmm13 # 4460 <_sk_callback_sse2+0x739> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,153,35,0,0 // mulps 0x2399(%rip),%xmm13 # 43f0 <_sk_callback_sse2+0x6e8> + .byte 68,15,89,45,250,35,0,0 // mulps 0x23fa(%rip),%xmm13 # 4470 <_sk_callback_sse2+0x749> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -25274,15 +25342,15 @@ _sk_parametric_a_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,25,35,0,0 // mulps 0x2319(%rip),%xmm9 # 4400 <_sk_callback_sse2+0x6f8> - .byte 68,15,84,21,33,35,0,0 // andps 0x2321(%rip),%xmm10 # 4410 <_sk_callback_sse2+0x708> - .byte 68,15,86,21,41,35,0,0 // orps 0x2329(%rip),%xmm10 # 4420 <_sk_callback_sse2+0x718> - .byte 68,15,88,13,49,35,0,0 // addps 0x2331(%rip),%xmm9 # 4430 <_sk_callback_sse2+0x728> - .byte 68,15,40,37,57,35,0,0 // movaps 0x2339(%rip),%xmm12 # 4440 <_sk_callback_sse2+0x738> + .byte 68,15,89,13,122,35,0,0 // mulps 0x237a(%rip),%xmm9 # 4480 <_sk_callback_sse2+0x759> + .byte 68,15,84,21,130,35,0,0 // andps 0x2382(%rip),%xmm10 # 4490 <_sk_callback_sse2+0x769> + .byte 68,15,86,21,138,35,0,0 // orps 0x238a(%rip),%xmm10 # 44a0 <_sk_callback_sse2+0x779> + .byte 68,15,88,13,146,35,0,0 // addps 0x2392(%rip),%xmm9 # 44b0 <_sk_callback_sse2+0x789> + .byte 68,15,40,37,154,35,0,0 // movaps 0x239a(%rip),%xmm12 # 44c0 <_sk_callback_sse2+0x799> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,57,35,0,0 // addps 0x2339(%rip),%xmm10 # 4450 <_sk_callback_sse2+0x748> - .byte 68,15,40,37,65,35,0,0 // movaps 0x2341(%rip),%xmm12 # 4460 <_sk_callback_sse2+0x758> + .byte 68,15,88,21,154,35,0,0 // addps 0x239a(%rip),%xmm10 # 44d0 <_sk_callback_sse2+0x7a9> + .byte 68,15,40,37,162,35,0,0 // movaps 0x23a2(%rip),%xmm12 # 44e0 <_sk_callback_sse2+0x7b9> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -25290,22 +25358,22 @@ _sk_parametric_a_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,43,35,0,0 // movaps 0x232b(%rip),%xmm10 # 4470 <_sk_callback_sse2+0x768> + .byte 68,15,40,21,140,35,0,0 // movaps 0x238c(%rip),%xmm10 # 44f0 <_sk_callback_sse2+0x7c9> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,31,35,0,0 // addps 0x231f(%rip),%xmm9 # 4480 <_sk_callback_sse2+0x778> - .byte 68,15,40,37,39,35,0,0 // movaps 0x2327(%rip),%xmm12 # 4490 <_sk_callback_sse2+0x788> + .byte 68,15,88,13,128,35,0,0 // addps 0x2380(%rip),%xmm9 # 4500 <_sk_callback_sse2+0x7d9> + .byte 68,15,40,37,136,35,0,0 // movaps 0x2388(%rip),%xmm12 # 4510 <_sk_callback_sse2+0x7e9> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,39,35,0,0 // movaps 0x2327(%rip),%xmm12 # 44a0 <_sk_callback_sse2+0x798> + .byte 68,15,40,37,136,35,0,0 // movaps 0x2388(%rip),%xmm12 # 4520 <_sk_callback_sse2+0x7f9> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,43,35,0,0 // movaps 0x232b(%rip),%xmm13 # 44b0 <_sk_callback_sse2+0x7a8> + .byte 68,15,40,45,140,35,0,0 // movaps 0x238c(%rip),%xmm13 # 4530 <_sk_callback_sse2+0x809> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,43,35,0,0 // mulps 0x232b(%rip),%xmm13 # 44c0 <_sk_callback_sse2+0x7b8> + .byte 68,15,89,45,140,35,0,0 // mulps 0x238c(%rip),%xmm13 # 4540 <_sk_callback_sse2+0x819> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -25322,29 +25390,29 @@ HIDDEN _sk_lab_to_xyz_sse2 .globl _sk_lab_to_xyz_sse2 FUNCTION(_sk_lab_to_xyz_sse2) _sk_lab_to_xyz_sse2: - .byte 15,89,5,8,35,0,0 // mulps 0x2308(%rip),%xmm0 # 44d0 <_sk_callback_sse2+0x7c8> - .byte 68,15,40,5,16,35,0,0 // movaps 0x2310(%rip),%xmm8 # 44e0 <_sk_callback_sse2+0x7d8> + .byte 15,89,5,105,35,0,0 // mulps 0x2369(%rip),%xmm0 # 4550 <_sk_callback_sse2+0x829> + .byte 68,15,40,5,113,35,0,0 // movaps 0x2371(%rip),%xmm8 # 4560 <_sk_callback_sse2+0x839> .byte 65,15,89,200 // mulps %xmm8,%xmm1 - .byte 68,15,40,13,20,35,0,0 // movaps 0x2314(%rip),%xmm9 # 44f0 <_sk_callback_sse2+0x7e8> + .byte 68,15,40,13,117,35,0,0 // movaps 0x2375(%rip),%xmm9 # 4570 <_sk_callback_sse2+0x849> .byte 65,15,88,201 // addps %xmm9,%xmm1 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 65,15,88,209 // addps %xmm9,%xmm2 - .byte 15,88,5,17,35,0,0 // addps 0x2311(%rip),%xmm0 # 4500 <_sk_callback_sse2+0x7f8> - .byte 15,89,5,26,35,0,0 // mulps 0x231a(%rip),%xmm0 # 4510 <_sk_callback_sse2+0x808> - .byte 15,89,13,35,35,0,0 // mulps 0x2323(%rip),%xmm1 # 4520 <_sk_callback_sse2+0x818> + .byte 15,88,5,114,35,0,0 // addps 0x2372(%rip),%xmm0 # 4580 <_sk_callback_sse2+0x859> + .byte 15,89,5,123,35,0,0 // mulps 0x237b(%rip),%xmm0 # 4590 <_sk_callback_sse2+0x869> + .byte 15,89,13,132,35,0,0 // mulps 0x2384(%rip),%xmm1 # 45a0 <_sk_callback_sse2+0x879> .byte 15,88,200 // addps %xmm0,%xmm1 - .byte 15,89,21,41,35,0,0 // mulps 0x2329(%rip),%xmm2 # 4530 <_sk_callback_sse2+0x828> + .byte 15,89,21,138,35,0,0 // mulps 0x238a(%rip),%xmm2 # 45b0 <_sk_callback_sse2+0x889> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 68,15,92,202 // subps %xmm2,%xmm9 .byte 68,15,40,225 // movaps %xmm1,%xmm12 .byte 69,15,89,228 // mulps %xmm12,%xmm12 .byte 68,15,89,225 // mulps %xmm1,%xmm12 - .byte 15,40,21,30,35,0,0 // movaps 0x231e(%rip),%xmm2 # 4540 <_sk_callback_sse2+0x838> + .byte 15,40,21,127,35,0,0 // movaps 0x237f(%rip),%xmm2 # 45c0 <_sk_callback_sse2+0x899> .byte 68,15,40,194 // movaps %xmm2,%xmm8 .byte 69,15,194,196,1 // cmpltps %xmm12,%xmm8 - .byte 68,15,40,21,29,35,0,0 // movaps 0x231d(%rip),%xmm10 # 4550 <_sk_callback_sse2+0x848> + .byte 68,15,40,21,126,35,0,0 // movaps 0x237e(%rip),%xmm10 # 45d0 <_sk_callback_sse2+0x8a9> .byte 65,15,88,202 // addps %xmm10,%xmm1 - .byte 68,15,40,29,33,35,0,0 // movaps 0x2321(%rip),%xmm11 # 4560 <_sk_callback_sse2+0x858> + .byte 68,15,40,29,130,35,0,0 // movaps 0x2382(%rip),%xmm11 # 45e0 <_sk_callback_sse2+0x8b9> .byte 65,15,89,203 // mulps %xmm11,%xmm1 .byte 69,15,84,224 // andps %xmm8,%xmm12 .byte 68,15,85,193 // andnps %xmm1,%xmm8 @@ -25368,8 +25436,8 @@ _sk_lab_to_xyz_sse2: .byte 15,84,194 // andps %xmm2,%xmm0 .byte 65,15,85,209 // andnps %xmm9,%xmm2 .byte 15,86,208 // orps %xmm0,%xmm2 - .byte 68,15,89,5,209,34,0,0 // mulps 0x22d1(%rip),%xmm8 # 4570 <_sk_callback_sse2+0x868> - .byte 15,89,21,218,34,0,0 // mulps 0x22da(%rip),%xmm2 # 4580 <_sk_callback_sse2+0x878> + .byte 68,15,89,5,50,35,0,0 // mulps 0x2332(%rip),%xmm8 # 45f0 <_sk_callback_sse2+0x8c9> + .byte 15,89,21,59,35,0,0 // mulps 0x233b(%rip),%xmm2 # 4600 <_sk_callback_sse2+0x8d9> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -25385,7 +25453,7 @@ _sk_load_a8_sse2: .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,194,34,0,0 // mulps 0x22c2(%rip),%xmm3 # 4590 <_sk_callback_sse2+0x888> + .byte 15,89,29,35,35,0,0 // mulps 0x2323(%rip),%xmm3 # 4610 <_sk_callback_sse2+0x8e9> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 102,15,239,201 // pxor %xmm1,%xmm1 @@ -25430,7 +25498,7 @@ _sk_gather_a8_sse2: .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,49,34,0,0 // mulps 0x2231(%rip),%xmm3 # 45a0 <_sk_callback_sse2+0x898> + .byte 15,89,29,146,34,0,0 // mulps 0x2292(%rip),%xmm3 # 4620 <_sk_callback_sse2+0x8f9> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 102,15,239,201 // pxor %xmm1,%xmm1 @@ -25443,7 +25511,7 @@ FUNCTION(_sk_store_a8_sse2) _sk_store_a8_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,37,34,0,0 // movaps 0x2225(%rip),%xmm8 # 45b0 <_sk_callback_sse2+0x8a8> + .byte 68,15,40,5,134,34,0,0 // movaps 0x2286(%rip),%xmm8 # 4630 <_sk_callback_sse2+0x909> .byte 68,15,89,195 // mulps %xmm3,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 @@ -25465,9 +25533,9 @@ _sk_load_g8_sse2: .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,236,33,0,0 // mulps 0x21ec(%rip),%xmm0 # 45c0 <_sk_callback_sse2+0x8b8> + .byte 15,89,5,77,34,0,0 // mulps 0x224d(%rip),%xmm0 # 4640 <_sk_callback_sse2+0x919> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,243,33,0,0 // movaps 0x21f3(%rip),%xmm3 # 45d0 <_sk_callback_sse2+0x8c8> + .byte 15,40,29,84,34,0,0 // movaps 0x2254(%rip),%xmm3 # 4650 <_sk_callback_sse2+0x929> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 255,224 // jmpq *%rax @@ -25510,9 +25578,9 @@ _sk_gather_g8_sse2: .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,104,33,0,0 // mulps 0x2168(%rip),%xmm0 # 45e0 <_sk_callback_sse2+0x8d8> + .byte 15,89,5,201,33,0,0 // mulps 0x21c9(%rip),%xmm0 # 4660 <_sk_callback_sse2+0x939> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,111,33,0,0 // movaps 0x216f(%rip),%xmm3 # 45f0 <_sk_callback_sse2+0x8e8> + .byte 15,40,29,208,33,0,0 // movaps 0x21d0(%rip),%xmm3 # 4670 <_sk_callback_sse2+0x949> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 255,224 // jmpq *%rax @@ -25524,9 +25592,9 @@ _sk_gather_i8_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 2498 <_sk_gather_i8_sse2+0xf> + .byte 116,5 // je 24b7 <_sk_gather_i8_sse2+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 249a <_sk_gather_i8_sse2+0x11> + .byte 235,2 // jmp 24b9 <_sk_gather_i8_sse2+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 @@ -25575,11 +25643,11 @@ _sk_gather_i8_sse2: .byte 102,67,15,110,12,136 // movd (%r8,%r9,4),%xmm1 .byte 102,68,15,98,201 // punpckldq %xmm1,%xmm9 .byte 102,68,15,98,200 // punpckldq %xmm0,%xmm9 - .byte 102,15,111,21,142,32,0,0 // movdqa 0x208e(%rip),%xmm2 # 4600 <_sk_callback_sse2+0x8f8> + .byte 102,15,111,21,239,32,0,0 // movdqa 0x20ef(%rip),%xmm2 # 4680 <_sk_callback_sse2+0x959> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,138,32,0,0 // movaps 0x208a(%rip),%xmm8 # 4610 <_sk_callback_sse2+0x908> + .byte 68,15,40,5,235,32,0,0 // movaps 0x20eb(%rip),%xmm8 # 4690 <_sk_callback_sse2+0x969> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,114,209,8 // psrld $0x8,%xmm1 @@ -25606,19 +25674,19 @@ _sk_load_565_sse2: .byte 243,15,126,20,120 // movq (%rax,%rdi,2),%xmm2 .byte 102,15,239,192 // pxor %xmm0,%xmm0 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 - .byte 102,15,111,5,64,32,0,0 // movdqa 0x2040(%rip),%xmm0 # 4620 <_sk_callback_sse2+0x918> + .byte 102,15,111,5,161,32,0,0 // movdqa 0x20a1(%rip),%xmm0 # 46a0 <_sk_callback_sse2+0x979> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,66,32,0,0 // mulps 0x2042(%rip),%xmm0 # 4630 <_sk_callback_sse2+0x928> - .byte 102,15,111,13,74,32,0,0 // movdqa 0x204a(%rip),%xmm1 # 4640 <_sk_callback_sse2+0x938> + .byte 15,89,5,163,32,0,0 // mulps 0x20a3(%rip),%xmm0 # 46b0 <_sk_callback_sse2+0x989> + .byte 102,15,111,13,171,32,0,0 // movdqa 0x20ab(%rip),%xmm1 # 46c0 <_sk_callback_sse2+0x999> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,76,32,0,0 // mulps 0x204c(%rip),%xmm1 # 4650 <_sk_callback_sse2+0x948> - .byte 102,15,219,21,84,32,0,0 // pand 0x2054(%rip),%xmm2 # 4660 <_sk_callback_sse2+0x958> + .byte 15,89,13,173,32,0,0 // mulps 0x20ad(%rip),%xmm1 # 46d0 <_sk_callback_sse2+0x9a9> + .byte 102,15,219,21,181,32,0,0 // pand 0x20b5(%rip),%xmm2 # 46e0 <_sk_callback_sse2+0x9b9> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,90,32,0,0 // mulps 0x205a(%rip),%xmm2 # 4670 <_sk_callback_sse2+0x968> + .byte 15,89,21,187,32,0,0 // mulps 0x20bb(%rip),%xmm2 # 46f0 <_sk_callback_sse2+0x9c9> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,97,32,0,0 // movaps 0x2061(%rip),%xmm3 # 4680 <_sk_callback_sse2+0x978> + .byte 15,40,29,194,32,0,0 // movaps 0x20c2(%rip),%xmm3 # 4700 <_sk_callback_sse2+0x9d9> .byte 255,224 // jmpq *%rax HIDDEN _sk_gather_565_sse2 @@ -25653,19 +25721,19 @@ _sk_gather_565_sse2: .byte 102,15,196,208,3 // pinsrw $0x3,%eax,%xmm2 .byte 102,15,239,192 // pxor %xmm0,%xmm0 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 - .byte 102,15,111,5,234,31,0,0 // movdqa 0x1fea(%rip),%xmm0 # 4690 <_sk_callback_sse2+0x988> + .byte 102,15,111,5,75,32,0,0 // movdqa 0x204b(%rip),%xmm0 # 4710 <_sk_callback_sse2+0x9e9> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,236,31,0,0 // mulps 0x1fec(%rip),%xmm0 # 46a0 <_sk_callback_sse2+0x998> - .byte 102,15,111,13,244,31,0,0 // movdqa 0x1ff4(%rip),%xmm1 # 46b0 <_sk_callback_sse2+0x9a8> + .byte 15,89,5,77,32,0,0 // mulps 0x204d(%rip),%xmm0 # 4720 <_sk_callback_sse2+0x9f9> + .byte 102,15,111,13,85,32,0,0 // movdqa 0x2055(%rip),%xmm1 # 4730 <_sk_callback_sse2+0xa09> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,246,31,0,0 // mulps 0x1ff6(%rip),%xmm1 # 46c0 <_sk_callback_sse2+0x9b8> - .byte 102,15,219,21,254,31,0,0 // pand 0x1ffe(%rip),%xmm2 # 46d0 <_sk_callback_sse2+0x9c8> + .byte 15,89,13,87,32,0,0 // mulps 0x2057(%rip),%xmm1 # 4740 <_sk_callback_sse2+0xa19> + .byte 102,15,219,21,95,32,0,0 // pand 0x205f(%rip),%xmm2 # 4750 <_sk_callback_sse2+0xa29> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,4,32,0,0 // mulps 0x2004(%rip),%xmm2 # 46e0 <_sk_callback_sse2+0x9d8> + .byte 15,89,21,101,32,0,0 // mulps 0x2065(%rip),%xmm2 # 4760 <_sk_callback_sse2+0xa39> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,11,32,0,0 // movaps 0x200b(%rip),%xmm3 # 46f0 <_sk_callback_sse2+0x9e8> + .byte 15,40,29,108,32,0,0 // movaps 0x206c(%rip),%xmm3 # 4770 <_sk_callback_sse2+0xa49> .byte 255,224 // jmpq *%rax HIDDEN _sk_store_565_sse2 @@ -25674,12 +25742,12 @@ FUNCTION(_sk_store_565_sse2) _sk_store_565_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,12,32,0,0 // movaps 0x200c(%rip),%xmm8 # 4700 <_sk_callback_sse2+0x9f8> + .byte 68,15,40,5,109,32,0,0 // movaps 0x206d(%rip),%xmm8 # 4780 <_sk_callback_sse2+0xa59> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 .byte 102,65,15,114,241,11 // pslld $0xb,%xmm9 - .byte 68,15,40,21,1,32,0,0 // movaps 0x2001(%rip),%xmm10 # 4710 <_sk_callback_sse2+0xa08> + .byte 68,15,40,21,98,32,0,0 // movaps 0x2062(%rip),%xmm10 # 4790 <_sk_callback_sse2+0xa69> .byte 68,15,89,209 // mulps %xmm1,%xmm10 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10 @@ -25703,21 +25771,21 @@ _sk_load_4444_sse2: .byte 243,15,126,28,120 // movq (%rax,%rdi,2),%xmm3 .byte 102,15,239,192 // pxor %xmm0,%xmm0 .byte 102,15,97,216 // punpcklwd %xmm0,%xmm3 - .byte 102,15,111,5,186,31,0,0 // movdqa 0x1fba(%rip),%xmm0 # 4720 <_sk_callback_sse2+0xa18> + .byte 102,15,111,5,27,32,0,0 // movdqa 0x201b(%rip),%xmm0 # 47a0 <_sk_callback_sse2+0xa79> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,188,31,0,0 // mulps 0x1fbc(%rip),%xmm0 # 4730 <_sk_callback_sse2+0xa28> - .byte 102,15,111,13,196,31,0,0 // movdqa 0x1fc4(%rip),%xmm1 # 4740 <_sk_callback_sse2+0xa38> + .byte 15,89,5,29,32,0,0 // mulps 0x201d(%rip),%xmm0 # 47b0 <_sk_callback_sse2+0xa89> + .byte 102,15,111,13,37,32,0,0 // movdqa 0x2025(%rip),%xmm1 # 47c0 <_sk_callback_sse2+0xa99> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,198,31,0,0 // mulps 0x1fc6(%rip),%xmm1 # 4750 <_sk_callback_sse2+0xa48> - .byte 102,15,111,21,206,31,0,0 // movdqa 0x1fce(%rip),%xmm2 # 4760 <_sk_callback_sse2+0xa58> + .byte 15,89,13,39,32,0,0 // mulps 0x2027(%rip),%xmm1 # 47d0 <_sk_callback_sse2+0xaa9> + .byte 102,15,111,21,47,32,0,0 // movdqa 0x202f(%rip),%xmm2 # 47e0 <_sk_callback_sse2+0xab9> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,208,31,0,0 // mulps 0x1fd0(%rip),%xmm2 # 4770 <_sk_callback_sse2+0xa68> - .byte 102,15,219,29,216,31,0,0 // pand 0x1fd8(%rip),%xmm3 # 4780 <_sk_callback_sse2+0xa78> + .byte 15,89,21,49,32,0,0 // mulps 0x2031(%rip),%xmm2 # 47f0 <_sk_callback_sse2+0xac9> + .byte 102,15,219,29,57,32,0,0 // pand 0x2039(%rip),%xmm3 # 4800 <_sk_callback_sse2+0xad9> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,222,31,0,0 // mulps 0x1fde(%rip),%xmm3 # 4790 <_sk_callback_sse2+0xa88> + .byte 15,89,29,63,32,0,0 // mulps 0x203f(%rip),%xmm3 # 4810 <_sk_callback_sse2+0xae9> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -25753,21 +25821,21 @@ _sk_gather_4444_sse2: .byte 102,15,196,216,3 // pinsrw $0x3,%eax,%xmm3 .byte 102,15,239,192 // pxor %xmm0,%xmm0 .byte 102,15,97,216 // punpcklwd %xmm0,%xmm3 - .byte 102,15,111,5,101,31,0,0 // movdqa 0x1f65(%rip),%xmm0 # 47a0 <_sk_callback_sse2+0xa98> + .byte 102,15,111,5,198,31,0,0 // movdqa 0x1fc6(%rip),%xmm0 # 4820 <_sk_callback_sse2+0xaf9> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,103,31,0,0 // mulps 0x1f67(%rip),%xmm0 # 47b0 <_sk_callback_sse2+0xaa8> - .byte 102,15,111,13,111,31,0,0 // movdqa 0x1f6f(%rip),%xmm1 # 47c0 <_sk_callback_sse2+0xab8> + .byte 15,89,5,200,31,0,0 // mulps 0x1fc8(%rip),%xmm0 # 4830 <_sk_callback_sse2+0xb09> + .byte 102,15,111,13,208,31,0,0 // movdqa 0x1fd0(%rip),%xmm1 # 4840 <_sk_callback_sse2+0xb19> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,113,31,0,0 // mulps 0x1f71(%rip),%xmm1 # 47d0 <_sk_callback_sse2+0xac8> - .byte 102,15,111,21,121,31,0,0 // movdqa 0x1f79(%rip),%xmm2 # 47e0 <_sk_callback_sse2+0xad8> + .byte 15,89,13,210,31,0,0 // mulps 0x1fd2(%rip),%xmm1 # 4850 <_sk_callback_sse2+0xb29> + .byte 102,15,111,21,218,31,0,0 // movdqa 0x1fda(%rip),%xmm2 # 4860 <_sk_callback_sse2+0xb39> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,123,31,0,0 // mulps 0x1f7b(%rip),%xmm2 # 47f0 <_sk_callback_sse2+0xae8> - .byte 102,15,219,29,131,31,0,0 // pand 0x1f83(%rip),%xmm3 # 4800 <_sk_callback_sse2+0xaf8> + .byte 15,89,21,220,31,0,0 // mulps 0x1fdc(%rip),%xmm2 # 4870 <_sk_callback_sse2+0xb49> + .byte 102,15,219,29,228,31,0,0 // pand 0x1fe4(%rip),%xmm3 # 4880 <_sk_callback_sse2+0xb59> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,137,31,0,0 // mulps 0x1f89(%rip),%xmm3 # 4810 <_sk_callback_sse2+0xb08> + .byte 15,89,29,234,31,0,0 // mulps 0x1fea(%rip),%xmm3 # 4890 <_sk_callback_sse2+0xb69> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -25777,7 +25845,7 @@ FUNCTION(_sk_store_4444_sse2) _sk_store_4444_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,136,31,0,0 // movaps 0x1f88(%rip),%xmm8 # 4820 <_sk_callback_sse2+0xb18> + .byte 68,15,40,5,233,31,0,0 // movaps 0x1fe9(%rip),%xmm8 # 48a0 <_sk_callback_sse2+0xb79> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 @@ -25809,11 +25877,11 @@ _sk_load_8888_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 68,15,16,12,184 // movups (%rax,%rdi,4),%xmm9 - .byte 15,40,21,27,31,0,0 // movaps 0x1f1b(%rip),%xmm2 # 4830 <_sk_callback_sse2+0xb28> + .byte 15,40,21,124,31,0,0 // movaps 0x1f7c(%rip),%xmm2 # 48b0 <_sk_callback_sse2+0xb89> .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 15,84,194 // andps %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,25,31,0,0 // movaps 0x1f19(%rip),%xmm8 # 4840 <_sk_callback_sse2+0xb38> + .byte 68,15,40,5,122,31,0,0 // movaps 0x1f7a(%rip),%xmm8 # 48c0 <_sk_callback_sse2+0xb99> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,40,201 // movaps %xmm9,%xmm1 .byte 102,15,114,209,8 // psrld $0x8,%xmm1 @@ -25862,11 +25930,11 @@ _sk_gather_8888_sse2: .byte 102,67,15,110,12,129 // movd (%r9,%r8,4),%xmm1 .byte 102,68,15,98,201 // punpckldq %xmm1,%xmm9 .byte 102,68,15,98,200 // punpckldq %xmm0,%xmm9 - .byte 102,15,111,21,106,30,0,0 // movdqa 0x1e6a(%rip),%xmm2 # 4850 <_sk_callback_sse2+0xb48> + .byte 102,15,111,21,203,30,0,0 // movdqa 0x1ecb(%rip),%xmm2 # 48d0 <_sk_callback_sse2+0xba9> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,102,30,0,0 // movaps 0x1e66(%rip),%xmm8 # 4860 <_sk_callback_sse2+0xb58> + .byte 68,15,40,5,199,30,0,0 // movaps 0x1ec7(%rip),%xmm8 # 48e0 <_sk_callback_sse2+0xbb9> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,114,209,8 // psrld $0x8,%xmm1 @@ -25890,7 +25958,7 @@ FUNCTION(_sk_store_8888_sse2) _sk_store_8888_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,41,30,0,0 // movaps 0x1e29(%rip),%xmm8 # 4870 <_sk_callback_sse2+0xb68> + .byte 68,15,40,5,138,30,0,0 // movaps 0x1e8a(%rip),%xmm8 # 48f0 <_sk_callback_sse2+0xbc9> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 @@ -25929,7 +25997,7 @@ _sk_load_f16_sse2: .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1 - .byte 102,68,15,111,13,153,29,0,0 // movdqa 0x1d99(%rip),%xmm9 # 4880 <_sk_callback_sse2+0xb78> + .byte 102,68,15,111,13,250,29,0,0 // movdqa 0x1dfa(%rip),%xmm9 # 4900 <_sk_callback_sse2+0xbd9> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,193 // pand %xmm9,%xmm0 .byte 102,15,239,200 // pxor %xmm0,%xmm1 @@ -25937,11 +26005,11 @@ _sk_load_f16_sse2: .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 .byte 102,68,15,235,232 // por %xmm0,%xmm13 - .byte 102,68,15,111,29,126,29,0,0 // movdqa 0x1d7e(%rip),%xmm11 # 4890 <_sk_callback_sse2+0xb88> + .byte 102,68,15,111,29,223,29,0,0 // movdqa 0x1ddf(%rip),%xmm11 # 4910 <_sk_callback_sse2+0xbe9> .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 - .byte 102,68,15,111,37,128,29,0,0 // movdqa 0x1d80(%rip),%xmm12 # 48a0 <_sk_callback_sse2+0xb98> + .byte 102,68,15,111,37,225,29,0,0 // movdqa 0x1de1(%rip),%xmm12 # 4920 <_sk_callback_sse2+0xbf9> .byte 102,65,15,239,204 // pxor %xmm12,%xmm1 - .byte 102,15,111,29,131,29,0,0 // movdqa 0x1d83(%rip),%xmm3 # 48b0 <_sk_callback_sse2+0xba8> + .byte 102,15,111,29,228,29,0,0 // movdqa 0x1de4(%rip),%xmm3 # 4930 <_sk_callback_sse2+0xc09> .byte 102,15,111,195 // movdqa %xmm3,%xmm0 .byte 102,15,102,193 // pcmpgtd %xmm1,%xmm0 .byte 102,65,15,223,197 // pandn %xmm13,%xmm0 @@ -26027,7 +26095,7 @@ _sk_gather_f16_sse2: .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1 - .byte 102,68,15,111,13,17,28,0,0 // movdqa 0x1c11(%rip),%xmm9 # 48c0 <_sk_callback_sse2+0xbb8> + .byte 102,68,15,111,13,114,28,0,0 // movdqa 0x1c72(%rip),%xmm9 # 4940 <_sk_callback_sse2+0xc19> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,193 // pand %xmm9,%xmm0 .byte 102,15,239,200 // pxor %xmm0,%xmm1 @@ -26035,11 +26103,11 @@ _sk_gather_f16_sse2: .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 .byte 102,68,15,235,232 // por %xmm0,%xmm13 - .byte 102,68,15,111,29,246,27,0,0 // movdqa 0x1bf6(%rip),%xmm11 # 48d0 <_sk_callback_sse2+0xbc8> + .byte 102,68,15,111,29,87,28,0,0 // movdqa 0x1c57(%rip),%xmm11 # 4950 <_sk_callback_sse2+0xc29> .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 - .byte 102,68,15,111,37,248,27,0,0 // movdqa 0x1bf8(%rip),%xmm12 # 48e0 <_sk_callback_sse2+0xbd8> + .byte 102,68,15,111,37,89,28,0,0 // movdqa 0x1c59(%rip),%xmm12 # 4960 <_sk_callback_sse2+0xc39> .byte 102,65,15,239,204 // pxor %xmm12,%xmm1 - .byte 102,15,111,29,251,27,0,0 // movdqa 0x1bfb(%rip),%xmm3 # 48f0 <_sk_callback_sse2+0xbe8> + .byte 102,15,111,29,92,28,0,0 // movdqa 0x1c5c(%rip),%xmm3 # 4970 <_sk_callback_sse2+0xc49> .byte 102,15,111,195 // movdqa %xmm3,%xmm0 .byte 102,15,102,193 // pcmpgtd %xmm1,%xmm0 .byte 102,65,15,223,197 // pandn %xmm13,%xmm0 @@ -26092,17 +26160,17 @@ FUNCTION(_sk_store_f16_sse2) _sk_store_f16_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 102,68,15,111,21,35,27,0,0 // movdqa 0x1b23(%rip),%xmm10 # 4900 <_sk_callback_sse2+0xbf8> + .byte 102,68,15,111,21,132,27,0,0 // movdqa 0x1b84(%rip),%xmm10 # 4980 <_sk_callback_sse2+0xc59> .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13 .byte 102,69,15,219,234 // pand %xmm10,%xmm13 .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 - .byte 102,68,15,111,13,22,27,0,0 // movdqa 0x1b16(%rip),%xmm9 # 4910 <_sk_callback_sse2+0xc08> + .byte 102,68,15,111,13,119,27,0,0 // movdqa 0x1b77(%rip),%xmm9 # 4990 <_sk_callback_sse2+0xc69> .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13 .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8 .byte 102,69,15,102,196 // pcmpgtd %xmm12,%xmm8 .byte 102,65,15,114,212,13 // psrld $0xd,%xmm12 - .byte 102,68,15,111,29,7,27,0,0 // movdqa 0x1b07(%rip),%xmm11 # 4920 <_sk_callback_sse2+0xc18> + .byte 102,68,15,111,29,104,27,0,0 // movdqa 0x1b68(%rip),%xmm11 # 49a0 <_sk_callback_sse2+0xc79> .byte 102,69,15,235,235 // por %xmm11,%xmm13 .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 @@ -26181,7 +26249,7 @@ _sk_load_u16_be_sse2: .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 .byte 15,91,193 // cvtdq2ps %xmm1,%xmm0 - .byte 68,15,40,5,165,25,0,0 // movaps 0x19a5(%rip),%xmm8 # 4930 <_sk_callback_sse2+0xc28> + .byte 68,15,40,5,6,26,0,0 // movaps 0x1a06(%rip),%xmm8 # 49b0 <_sk_callback_sse2+0xc89> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 .byte 102,15,113,241,8 // psllw $0x8,%xmm1 @@ -26234,7 +26302,7 @@ _sk_load_rgb_u16_be_sse2: .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 102,65,15,97,192 // punpcklwd %xmm8,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,13,225,24,0,0 // movaps 0x18e1(%rip),%xmm9 # 4940 <_sk_callback_sse2+0xc38> + .byte 68,15,40,13,66,25,0,0 // movaps 0x1942(%rip),%xmm9 # 49c0 <_sk_callback_sse2+0xc99> .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 .byte 102,15,113,241,8 // psllw $0x8,%xmm1 @@ -26251,7 +26319,7 @@ _sk_load_rgb_u16_be_sse2: .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,209 // mulps %xmm9,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,168,24,0,0 // movaps 0x18a8(%rip),%xmm3 # 4950 <_sk_callback_sse2+0xc48> + .byte 15,40,29,9,25,0,0 // movaps 0x1909(%rip),%xmm3 # 49d0 <_sk_callback_sse2+0xca9> .byte 255,224 // jmpq *%rax HIDDEN _sk_store_u16_be_sse2 @@ -26260,7 +26328,7 @@ FUNCTION(_sk_store_u16_be_sse2) _sk_store_u16_be_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,13,169,24,0,0 // movaps 0x18a9(%rip),%xmm9 # 4960 <_sk_callback_sse2+0xc58> + .byte 68,15,40,13,10,25,0,0 // movaps 0x190a(%rip),%xmm9 # 49e0 <_sk_callback_sse2+0xcb9> .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 69,15,89,193 // mulps %xmm9,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 @@ -26410,7 +26478,7 @@ _sk_repeat_x_sse2: .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 - .byte 68,15,84,13,147,22,0,0 // andps 0x1693(%rip),%xmm9 # 4970 <_sk_callback_sse2+0xc68> + .byte 68,15,84,13,244,22,0,0 // andps 0x16f4(%rip),%xmm9 # 49f0 <_sk_callback_sse2+0xcc9> .byte 69,15,92,209 // subps %xmm9,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,92,194 // subps %xmm10,%xmm0 @@ -26432,7 +26500,7 @@ _sk_repeat_y_sse2: .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 - .byte 68,15,84,13,91,22,0,0 // andps 0x165b(%rip),%xmm9 # 4980 <_sk_callback_sse2+0xc78> + .byte 68,15,84,13,188,22,0,0 // andps 0x16bc(%rip),%xmm9 # 4a00 <_sk_callback_sse2+0xcd9> .byte 69,15,92,209 // subps %xmm9,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,92,202 // subps %xmm10,%xmm1 @@ -26458,7 +26526,7 @@ _sk_mirror_x_sse2: .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 - .byte 68,15,84,21,17,22,0,0 // andps 0x1611(%rip),%xmm10 # 4990 <_sk_callback_sse2+0xc88> + .byte 68,15,84,21,114,22,0,0 // andps 0x1672(%rip),%xmm10 # 4a10 <_sk_callback_sse2+0xce9> .byte 69,15,87,228 // xorps %xmm12,%xmm12 .byte 69,15,92,218 // subps %xmm10,%xmm11 .byte 69,15,89,216 // mulps %xmm8,%xmm11 @@ -26488,7 +26556,7 @@ _sk_mirror_y_sse2: .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 - .byte 68,15,84,21,183,21,0,0 // andps 0x15b7(%rip),%xmm10 # 49a0 <_sk_callback_sse2+0xc98> + .byte 68,15,84,21,24,22,0,0 // andps 0x1618(%rip),%xmm10 # 4a20 <_sk_callback_sse2+0xcf9> .byte 69,15,87,228 // xorps %xmm12,%xmm12 .byte 69,15,92,218 // subps %xmm10,%xmm11 .byte 69,15,89,216 // mulps %xmm8,%xmm11 @@ -26507,10 +26575,10 @@ HIDDEN _sk_luminance_to_alpha_sse2 FUNCTION(_sk_luminance_to_alpha_sse2) _sk_luminance_to_alpha_sse2: .byte 15,40,218 // movaps %xmm2,%xmm3 - .byte 15,89,5,143,21,0,0 // mulps 0x158f(%rip),%xmm0 # 49b0 <_sk_callback_sse2+0xca8> - .byte 15,89,13,152,21,0,0 // mulps 0x1598(%rip),%xmm1 # 49c0 <_sk_callback_sse2+0xcb8> + .byte 15,89,5,240,21,0,0 // mulps 0x15f0(%rip),%xmm0 # 4a30 <_sk_callback_sse2+0xd09> + .byte 15,89,13,249,21,0,0 // mulps 0x15f9(%rip),%xmm1 # 4a40 <_sk_callback_sse2+0xd19> .byte 15,88,200 // addps %xmm0,%xmm1 - .byte 15,89,29,158,21,0,0 // mulps 0x159e(%rip),%xmm3 # 49d0 <_sk_callback_sse2+0xcc8> + .byte 15,89,29,255,21,0,0 // mulps 0x15ff(%rip),%xmm3 # 4a50 <_sk_callback_sse2+0xd29> .byte 15,88,217 // addps %xmm1,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 @@ -26743,7 +26811,7 @@ _sk_linear_gradient_sse2: .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 72,139,8 // mov (%rax),%rcx .byte 72,133,201 // test %rcx,%rcx - .byte 15,132,15,1,0,0 // je 38e5 <_sk_linear_gradient_sse2+0x149> + .byte 15,132,15,1,0,0 // je 3904 <_sk_linear_gradient_sse2+0x149> .byte 72,139,64,8 // mov 0x8(%rax),%rax .byte 72,131,192,32 // add $0x20,%rax .byte 69,15,87,192 // xorps %xmm8,%xmm8 @@ -26804,8 +26872,8 @@ _sk_linear_gradient_sse2: .byte 69,15,86,231 // orps %xmm15,%xmm12 .byte 72,131,192,36 // add $0x24,%rax .byte 72,255,201 // dec %rcx - .byte 15,133,8,255,255,255 // jne 37eb <_sk_linear_gradient_sse2+0x4f> - .byte 235,13 // jmp 38f2 <_sk_linear_gradient_sse2+0x156> + .byte 15,133,8,255,255,255 // jne 380a <_sk_linear_gradient_sse2+0x4f> + .byte 235,13 // jmp 3911 <_sk_linear_gradient_sse2+0x156> .byte 15,87,201 // xorps %xmm1,%xmm1 .byte 15,87,210 // xorps %xmm2,%xmm2 .byte 15,87,219 // xorps %xmm3,%xmm3 @@ -26860,7 +26928,7 @@ HIDDEN _sk_save_xy_sse2 FUNCTION(_sk_save_xy_sse2) _sk_save_xy_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,79,16,0,0 // movaps 0x104f(%rip),%xmm8 # 49e0 <_sk_callback_sse2+0xcd8> + .byte 68,15,40,5,176,16,0,0 // movaps 0x10b0(%rip),%xmm8 # 4a60 <_sk_callback_sse2+0xd39> .byte 15,17,0 // movups %xmm0,(%rax) .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,88,200 // addps %xmm8,%xmm9 @@ -26868,7 +26936,7 @@ _sk_save_xy_sse2: .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,40,217 // movaps %xmm9,%xmm11 .byte 69,15,194,218,1 // cmpltps %xmm10,%xmm11 - .byte 68,15,40,37,58,16,0,0 // movaps 0x103a(%rip),%xmm12 # 49f0 <_sk_callback_sse2+0xce8> + .byte 68,15,40,37,155,16,0,0 // movaps 0x109b(%rip),%xmm12 # 4a70 <_sk_callback_sse2+0xd49> .byte 69,15,84,220 // andps %xmm12,%xmm11 .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 69,15,92,202 // subps %xmm10,%xmm9 @@ -26915,8 +26983,8 @@ _sk_bilinear_nx_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,179,15,0,0 // addps 0xfb3(%rip),%xmm0 # 4a00 <_sk_callback_sse2+0xcf8> - .byte 68,15,40,13,187,15,0,0 // movaps 0xfbb(%rip),%xmm9 # 4a10 <_sk_callback_sse2+0xd08> + .byte 15,88,5,20,16,0,0 // addps 0x1014(%rip),%xmm0 # 4a80 <_sk_callback_sse2+0xd59> + .byte 68,15,40,13,28,16,0,0 // movaps 0x101c(%rip),%xmm9 # 4a90 <_sk_callback_sse2+0xd69> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -26929,7 +26997,7 @@ _sk_bilinear_px_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,170,15,0,0 // addps 0xfaa(%rip),%xmm0 # 4a20 <_sk_callback_sse2+0xd18> + .byte 15,88,5,11,16,0,0 // addps 0x100b(%rip),%xmm0 # 4aa0 <_sk_callback_sse2+0xd79> .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -26941,8 +27009,8 @@ _sk_bilinear_ny_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,156,15,0,0 // addps 0xf9c(%rip),%xmm1 # 4a30 <_sk_callback_sse2+0xd28> - .byte 68,15,40,13,164,15,0,0 // movaps 0xfa4(%rip),%xmm9 # 4a40 <_sk_callback_sse2+0xd38> + .byte 15,88,13,253,15,0,0 // addps 0xffd(%rip),%xmm1 # 4ab0 <_sk_callback_sse2+0xd89> + .byte 68,15,40,13,5,16,0,0 // movaps 0x1005(%rip),%xmm9 # 4ac0 <_sk_callback_sse2+0xd99> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -26955,7 +27023,7 @@ _sk_bilinear_py_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,146,15,0,0 // addps 0xf92(%rip),%xmm1 # 4a50 <_sk_callback_sse2+0xd48> + .byte 15,88,13,243,15,0,0 // addps 0xff3(%rip),%xmm1 # 4ad0 <_sk_callback_sse2+0xda9> .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -26967,13 +27035,13 @@ _sk_bicubic_n3x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,133,15,0,0 // addps 0xf85(%rip),%xmm0 # 4a60 <_sk_callback_sse2+0xd58> - .byte 68,15,40,13,141,15,0,0 // movaps 0xf8d(%rip),%xmm9 # 4a70 <_sk_callback_sse2+0xd68> + .byte 15,88,5,230,15,0,0 // addps 0xfe6(%rip),%xmm0 # 4ae0 <_sk_callback_sse2+0xdb9> + .byte 68,15,40,13,238,15,0,0 // movaps 0xfee(%rip),%xmm9 # 4af0 <_sk_callback_sse2+0xdc9> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,89,192 // mulps %xmm8,%xmm8 - .byte 68,15,89,13,137,15,0,0 // mulps 0xf89(%rip),%xmm9 # 4a80 <_sk_callback_sse2+0xd78> - .byte 68,15,88,13,145,15,0,0 // addps 0xf91(%rip),%xmm9 # 4a90 <_sk_callback_sse2+0xd88> + .byte 68,15,89,13,234,15,0,0 // mulps 0xfea(%rip),%xmm9 # 4b00 <_sk_callback_sse2+0xdd9> + .byte 68,15,88,13,242,15,0,0 // addps 0xff2(%rip),%xmm9 # 4b10 <_sk_callback_sse2+0xde9> .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -26986,16 +27054,16 @@ _sk_bicubic_n1x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,128,15,0,0 // addps 0xf80(%rip),%xmm0 # 4aa0 <_sk_callback_sse2+0xd98> - .byte 68,15,40,13,136,15,0,0 // movaps 0xf88(%rip),%xmm9 # 4ab0 <_sk_callback_sse2+0xda8> + .byte 15,88,5,225,15,0,0 // addps 0xfe1(%rip),%xmm0 # 4b20 <_sk_callback_sse2+0xdf9> + .byte 68,15,40,13,233,15,0,0 // movaps 0xfe9(%rip),%xmm9 # 4b30 <_sk_callback_sse2+0xe09> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,140,15,0,0 // movaps 0xf8c(%rip),%xmm8 # 4ac0 <_sk_callback_sse2+0xdb8> + .byte 68,15,40,5,237,15,0,0 // movaps 0xfed(%rip),%xmm8 # 4b40 <_sk_callback_sse2+0xe19> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,144,15,0,0 // addps 0xf90(%rip),%xmm8 # 4ad0 <_sk_callback_sse2+0xdc8> + .byte 68,15,88,5,241,15,0,0 // addps 0xff1(%rip),%xmm8 # 4b50 <_sk_callback_sse2+0xe29> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,148,15,0,0 // addps 0xf94(%rip),%xmm8 # 4ae0 <_sk_callback_sse2+0xdd8> + .byte 68,15,88,5,245,15,0,0 // addps 0xff5(%rip),%xmm8 # 4b60 <_sk_callback_sse2+0xe39> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,152,15,0,0 // addps 0xf98(%rip),%xmm8 # 4af0 <_sk_callback_sse2+0xde8> + .byte 68,15,88,5,249,15,0,0 // addps 0xff9(%rip),%xmm8 # 4b70 <_sk_callback_sse2+0xe49> .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -27005,17 +27073,17 @@ HIDDEN _sk_bicubic_p1x_sse2 FUNCTION(_sk_bicubic_p1x_sse2) _sk_bicubic_p1x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,146,15,0,0 // movaps 0xf92(%rip),%xmm8 # 4b00 <_sk_callback_sse2+0xdf8> + .byte 68,15,40,5,243,15,0,0 // movaps 0xff3(%rip),%xmm8 # 4b80 <_sk_callback_sse2+0xe59> .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 .byte 65,15,88,192 // addps %xmm8,%xmm0 - .byte 68,15,40,21,142,15,0,0 // movaps 0xf8e(%rip),%xmm10 # 4b10 <_sk_callback_sse2+0xe08> + .byte 68,15,40,21,239,15,0,0 // movaps 0xfef(%rip),%xmm10 # 4b90 <_sk_callback_sse2+0xe69> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,146,15,0,0 // addps 0xf92(%rip),%xmm10 # 4b20 <_sk_callback_sse2+0xe18> + .byte 68,15,88,21,243,15,0,0 // addps 0xff3(%rip),%xmm10 # 4ba0 <_sk_callback_sse2+0xe79> .byte 69,15,89,209 // mulps %xmm9,%xmm10 .byte 69,15,88,208 // addps %xmm8,%xmm10 .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,142,15,0,0 // addps 0xf8e(%rip),%xmm10 # 4b30 <_sk_callback_sse2+0xe28> + .byte 68,15,88,21,239,15,0,0 // addps 0xfef(%rip),%xmm10 # 4bb0 <_sk_callback_sse2+0xe89> .byte 68,15,17,144,128,0,0,0 // movups %xmm10,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -27027,11 +27095,11 @@ _sk_bicubic_p3x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,129,15,0,0 // addps 0xf81(%rip),%xmm0 # 4b40 <_sk_callback_sse2+0xe38> + .byte 15,88,5,226,15,0,0 // addps 0xfe2(%rip),%xmm0 # 4bc0 <_sk_callback_sse2+0xe99> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,129,15,0,0 // mulps 0xf81(%rip),%xmm8 # 4b50 <_sk_callback_sse2+0xe48> - .byte 68,15,88,5,137,15,0,0 // addps 0xf89(%rip),%xmm8 # 4b60 <_sk_callback_sse2+0xe58> + .byte 68,15,89,5,226,15,0,0 // mulps 0xfe2(%rip),%xmm8 # 4bd0 <_sk_callback_sse2+0xea9> + .byte 68,15,88,5,234,15,0,0 // addps 0xfea(%rip),%xmm8 # 4be0 <_sk_callback_sse2+0xeb9> .byte 69,15,89,193 // mulps %xmm9,%xmm8 .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -27044,13 +27112,13 @@ _sk_bicubic_n3y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,119,15,0,0 // addps 0xf77(%rip),%xmm1 # 4b70 <_sk_callback_sse2+0xe68> - .byte 68,15,40,13,127,15,0,0 // movaps 0xf7f(%rip),%xmm9 # 4b80 <_sk_callback_sse2+0xe78> + .byte 15,88,13,216,15,0,0 // addps 0xfd8(%rip),%xmm1 # 4bf0 <_sk_callback_sse2+0xec9> + .byte 68,15,40,13,224,15,0,0 // movaps 0xfe0(%rip),%xmm9 # 4c00 <_sk_callback_sse2+0xed9> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,89,192 // mulps %xmm8,%xmm8 - .byte 68,15,89,13,123,15,0,0 // mulps 0xf7b(%rip),%xmm9 # 4b90 <_sk_callback_sse2+0xe88> - .byte 68,15,88,13,131,15,0,0 // addps 0xf83(%rip),%xmm9 # 4ba0 <_sk_callback_sse2+0xe98> + .byte 68,15,89,13,220,15,0,0 // mulps 0xfdc(%rip),%xmm9 # 4c10 <_sk_callback_sse2+0xee9> + .byte 68,15,88,13,228,15,0,0 // addps 0xfe4(%rip),%xmm9 # 4c20 <_sk_callback_sse2+0xef9> .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -27063,16 +27131,16 @@ _sk_bicubic_n1y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,113,15,0,0 // addps 0xf71(%rip),%xmm1 # 4bb0 <_sk_callback_sse2+0xea8> - .byte 68,15,40,13,121,15,0,0 // movaps 0xf79(%rip),%xmm9 # 4bc0 <_sk_callback_sse2+0xeb8> + .byte 15,88,13,210,15,0,0 // addps 0xfd2(%rip),%xmm1 # 4c30 <_sk_callback_sse2+0xf09> + .byte 68,15,40,13,218,15,0,0 // movaps 0xfda(%rip),%xmm9 # 4c40 <_sk_callback_sse2+0xf19> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,125,15,0,0 // movaps 0xf7d(%rip),%xmm8 # 4bd0 <_sk_callback_sse2+0xec8> + .byte 68,15,40,5,222,15,0,0 // movaps 0xfde(%rip),%xmm8 # 4c50 <_sk_callback_sse2+0xf29> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,129,15,0,0 // addps 0xf81(%rip),%xmm8 # 4be0 <_sk_callback_sse2+0xed8> + .byte 68,15,88,5,226,15,0,0 // addps 0xfe2(%rip),%xmm8 # 4c60 <_sk_callback_sse2+0xf39> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,133,15,0,0 // addps 0xf85(%rip),%xmm8 # 4bf0 <_sk_callback_sse2+0xee8> + .byte 68,15,88,5,230,15,0,0 // addps 0xfe6(%rip),%xmm8 # 4c70 <_sk_callback_sse2+0xf49> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,137,15,0,0 // addps 0xf89(%rip),%xmm8 # 4c00 <_sk_callback_sse2+0xef8> + .byte 68,15,88,5,234,15,0,0 // addps 0xfea(%rip),%xmm8 # 4c80 <_sk_callback_sse2+0xf59> .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -27082,17 +27150,17 @@ HIDDEN _sk_bicubic_p1y_sse2 FUNCTION(_sk_bicubic_p1y_sse2) _sk_bicubic_p1y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,131,15,0,0 // movaps 0xf83(%rip),%xmm8 # 4c10 <_sk_callback_sse2+0xf08> + .byte 68,15,40,5,228,15,0,0 // movaps 0xfe4(%rip),%xmm8 # 4c90 <_sk_callback_sse2+0xf69> .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 .byte 65,15,88,200 // addps %xmm8,%xmm1 - .byte 68,15,40,21,126,15,0,0 // movaps 0xf7e(%rip),%xmm10 # 4c20 <_sk_callback_sse2+0xf18> + .byte 68,15,40,21,223,15,0,0 // movaps 0xfdf(%rip),%xmm10 # 4ca0 <_sk_callback_sse2+0xf79> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,130,15,0,0 // addps 0xf82(%rip),%xmm10 # 4c30 <_sk_callback_sse2+0xf28> + .byte 68,15,88,21,227,15,0,0 // addps 0xfe3(%rip),%xmm10 # 4cb0 <_sk_callback_sse2+0xf89> .byte 69,15,89,209 // mulps %xmm9,%xmm10 .byte 69,15,88,208 // addps %xmm8,%xmm10 .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,126,15,0,0 // addps 0xf7e(%rip),%xmm10 # 4c40 <_sk_callback_sse2+0xf38> + .byte 68,15,88,21,223,15,0,0 // addps 0xfdf(%rip),%xmm10 # 4cc0 <_sk_callback_sse2+0xf99> .byte 68,15,17,144,160,0,0,0 // movups %xmm10,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -27104,11 +27172,11 @@ _sk_bicubic_p3y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,112,15,0,0 // addps 0xf70(%rip),%xmm1 # 4c50 <_sk_callback_sse2+0xf48> + .byte 15,88,13,209,15,0,0 // addps 0xfd1(%rip),%xmm1 # 4cd0 <_sk_callback_sse2+0xfa9> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,112,15,0,0 // mulps 0xf70(%rip),%xmm8 # 4c60 <_sk_callback_sse2+0xf58> - .byte 68,15,88,5,120,15,0,0 // addps 0xf78(%rip),%xmm8 # 4c70 <_sk_callback_sse2+0xf68> + .byte 68,15,89,5,209,15,0,0 // mulps 0xfd1(%rip),%xmm8 # 4ce0 <_sk_callback_sse2+0xfb9> + .byte 68,15,88,5,217,15,0,0 // addps 0xfd9(%rip),%xmm8 # 4cf0 <_sk_callback_sse2+0xfc9> .byte 69,15,89,193 // mulps %xmm9,%xmm8 .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -27293,11 +27361,11 @@ BALIGN16 .byte 0,128,191,0,0,128 // add %al,-0x7fffff41(%rax) .byte 191,0,0,224,64 // mov $0x40e00000,%edi .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 3f68 <.literal16+0x188> + .byte 224,64 // loopne 3f88 <.literal16+0x188> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 3f6c <.literal16+0x18c> + .byte 224,64 // loopne 3f8c <.literal16+0x18c> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 3f70 <.literal16+0x190> + .byte 224,64 // loopne 3f90 <.literal16+0x190> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -27436,6 +27504,49 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .byte 0,0 // add %al,(%rax) + .byte 128,63,171 // cmpb $0xab,(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,171 // ds stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,171 // ds stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,171 // ds stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 62,0,0 // add %al,%ds:(%rax) + .byte 128,191,0,0,128,191,0 // cmpb $0x0,-0x40800000(%rdi) + .byte 0,128,191,0,0,128 // add %al,-0x7fffff41(%rax) + .byte 191,0,0,192,64 // mov $0x40c00000,%edi + .byte 0,0 // add %al,(%rax) + .byte 192,64,0,0 // rolb $0x0,0x0(%rax) + .byte 192,64,0,0 // rolb $0x0,0x0(%rax) + .byte 192,64,171,170 // rolb $0xaa,-0x55(%rax) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 42,63 // sub (%rdi),%bh + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 190,171,170,170,190 // mov $0xbeaaaaab,%esi + .byte 171 // stos %eax,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 170 // stos %al,%es:(%rdi) + .byte 190,171,170,170,190 // mov $0xbeaaaaab,%esi .byte 129,128,128,59,129,128,128,59,129,128// addl $0x80813b80,-0x7f7ec480(%rax) .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,129,128,128,59 // addb $0x3b,-0x7f7f7ec5(%rax) @@ -27457,13 +27568,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 40b9 <.literal16+0x2d9> + .byte 224,7 // loopne 4139 <.literal16+0x339> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 40bd <.literal16+0x2dd> + .byte 224,7 // loopne 413d <.literal16+0x33d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 40c1 <.literal16+0x2e1> + .byte 224,7 // loopne 4141 <.literal16+0x341> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 40c5 <.literal16+0x2e5> + .byte 224,7 // loopne 4145 <.literal16+0x345> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -27532,11 +27643,11 @@ BALIGN16 .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,127,67 // add %bh,0x43(%rdi) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 41ab <.literal16+0x3cb> + .byte 127,67 // jg 422b <.literal16+0x42b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 41af <.literal16+0x3cf> + .byte 127,67 // jg 422f <.literal16+0x42f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 41b3 <.literal16+0x3d3> + .byte 127,67 // jg 4233 <.literal16+0x433> .byte 129,128,128,59,129,128,128,59,129,128// addl $0x80813b80,-0x7f7ec480(%rax) .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,129,128,128,59 // addb $0x3b,-0x7f7f7ec5(%rax) @@ -27551,16 +27662,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 41a4 <.literal16+0x3c4> + .byte 127,0 // jg 4224 <.literal16+0x424> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 41a8 <.literal16+0x3c8> + .byte 127,0 // jg 4228 <.literal16+0x428> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 41ac <.literal16+0x3cc> + .byte 127,0 // jg 422c <.literal16+0x42c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 41b0 <.literal16+0x3d0> + .byte 127,0 // jg 4230 <.literal16+0x430> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -27569,7 +27680,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4235 <.literal16+0x455> + .byte 119,115 // ja 42b5 <.literal16+0x4b5> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -27580,7 +27691,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 4199 <.literal16+0x3b9> + .byte 117,191 // jne 4219 <.literal16+0x419> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -27592,7 +27703,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a381da <_sk_callback_sse2+0xffffffffe9a344d2> + .byte 233,220,63,163,233 // jmpq ffffffffe9a3825a <_sk_callback_sse2+0xffffffffe9a34533> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -27646,16 +27757,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 4274 <.literal16+0x494> + .byte 127,0 // jg 42f4 <.literal16+0x4f4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4278 <.literal16+0x498> + .byte 127,0 // jg 42f8 <.literal16+0x4f8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 427c <.literal16+0x49c> + .byte 127,0 // jg 42fc <.literal16+0x4fc> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4280 <.literal16+0x4a0> + .byte 127,0 // jg 4300 <.literal16+0x500> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -27664,7 +27775,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4305 <.literal16+0x525> + .byte 119,115 // ja 4385 <.literal16+0x585> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -27675,7 +27786,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 4269 <.literal16+0x489> + .byte 117,191 // jne 42e9 <.literal16+0x4e9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -27687,7 +27798,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a382aa <_sk_callback_sse2+0xffffffffe9a345a2> + .byte 233,220,63,163,233 // jmpq ffffffffe9a3832a <_sk_callback_sse2+0xffffffffe9a34603> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -27741,16 +27852,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 4344 <.literal16+0x564> + .byte 127,0 // jg 43c4 <.literal16+0x5c4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4348 <.literal16+0x568> + .byte 127,0 // jg 43c8 <.literal16+0x5c8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 434c <.literal16+0x56c> + .byte 127,0 // jg 43cc <.literal16+0x5cc> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4350 <.literal16+0x570> + .byte 127,0 // jg 43d0 <.literal16+0x5d0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -27759,7 +27870,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 43d5 <.literal16+0x5f5> + .byte 119,115 // ja 4455 <.literal16+0x655> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -27770,7 +27881,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 4339 <.literal16+0x559> + .byte 117,191 // jne 43b9 <.literal16+0x5b9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -27782,7 +27893,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a3837a <_sk_callback_sse2+0xffffffffe9a34672> + .byte 233,220,63,163,233 // jmpq ffffffffe9a383fa <_sk_callback_sse2+0xffffffffe9a346d3> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -27836,16 +27947,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 4414 <.literal16+0x634> + .byte 127,0 // jg 4494 <.literal16+0x694> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4418 <.literal16+0x638> + .byte 127,0 // jg 4498 <.literal16+0x698> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 441c <.literal16+0x63c> + .byte 127,0 // jg 449c <.literal16+0x69c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 4420 <.literal16+0x640> + .byte 127,0 // jg 44a0 <.literal16+0x6a0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -27854,7 +27965,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 44a5 <.literal16+0x6c5> + .byte 119,115 // ja 4525 <.literal16+0x725> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -27865,7 +27976,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 4409 <.literal16+0x629> + .byte 117,191 // jne 4489 <.literal16+0x689> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -27877,7 +27988,7 @@ BALIGN16 .byte 249 // stc .byte 68,180,62 // rex.R mov $0x3e,%spl .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9 - .byte 233,220,63,163,233 // jmpq ffffffffe9a3844a <_sk_callback_sse2+0xffffffffe9a34742> + .byte 233,220,63,163,233 // jmpq ffffffffe9a384ca <_sk_callback_sse2+0xffffffffe9a347a3> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -27927,13 +28038,13 @@ BALIGN16 .byte 200,66,0,0 // enterq $0x42,$0x0 .byte 200,66,0,0 // enterq $0x42,$0x0 .byte 200,66,0,0 // enterq $0x42,$0x0 - .byte 127,67 // jg 4527 <.literal16+0x747> + .byte 127,67 // jg 45a7 <.literal16+0x7a7> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 452b <.literal16+0x74b> + .byte 127,67 // jg 45ab <.literal16+0x7ab> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 452f <.literal16+0x74f> + .byte 127,67 // jg 45af <.literal16+0x7af> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 4533 <.literal16+0x753> + .byte 127,67 // jg 45b3 <.literal16+0x7b3> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -27980,16 +28091,16 @@ BALIGN16 .byte 128,3,62 // addb $0x3e,(%rbx) .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 45b3 <.literal16+0x7d3> + .byte 118,63 // jbe 4633 <.literal16+0x833> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 45b7 <.literal16+0x7d7> + .byte 118,63 // jbe 4637 <.literal16+0x837> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 45bb <.literal16+0x7db> + .byte 118,63 // jbe 463b <.literal16+0x83b> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 45bf <.literal16+0x7df> + .byte 118,63 // jbe 463f <.literal16+0x83f> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 246,64,83,63 // testb $0x3f,0x53(%rax) @@ -28001,11 +28112,11 @@ BALIGN16 .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,127,67 // add %bh,0x43(%rdi) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 45fb <.literal16+0x81b> + .byte 127,67 // jg 467b <.literal16+0x87b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 45ff <.literal16+0x81f> + .byte 127,67 // jg 467f <.literal16+0x87f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 4603 <.literal16+0x823> + .byte 127,67 // jg 4683 <.literal16+0x883> .byte 129,128,128,59,129,128,128,59,129,128// addl $0x80813b80,-0x7f7ec480(%rax) .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,0,0,128,63 // addb $0x3f,-0x7fffffc5(%rax) @@ -28045,13 +28156,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 4649 <.literal16+0x869> + .byte 224,7 // loopne 46c9 <.literal16+0x8c9> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 464d <.literal16+0x86d> + .byte 224,7 // loopne 46cd <.literal16+0x8cd> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 4651 <.literal16+0x871> + .byte 224,7 // loopne 46d1 <.literal16+0x8d1> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 4655 <.literal16+0x875> + .byte 224,7 // loopne 46d5 <.literal16+0x8d5> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -28097,13 +28208,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 46b9 <.literal16+0x8d9> + .byte 224,7 // loopne 4739 <.literal16+0x939> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 46bd <.literal16+0x8dd> + .byte 224,7 // loopne 473d <.literal16+0x93d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 46c1 <.literal16+0x8e1> + .byte 224,7 // loopne 4741 <.literal16+0x941> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 46c5 <.literal16+0x8e5> + .byte 224,7 // loopne 4745 <.literal16+0x945> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -28141,13 +28252,13 @@ BALIGN16 .byte 65,0,0 // add %al,(%r8) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 4756 <.literal16+0x976> + .byte 124,66 // jl 47d6 <.literal16+0x9d6> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 475a <.literal16+0x97a> + .byte 124,66 // jl 47da <.literal16+0x9da> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 475e <.literal16+0x97e> + .byte 124,66 // jl 47de <.literal16+0x9de> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 4762 <.literal16+0x982> + .byte 124,66 // jl 47e2 <.literal16+0x9e2> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 0,240 // add %dh,%al @@ -28237,13 +28348,13 @@ BALIGN16 .byte 136,136,61,137,136,136 // mov %cl,-0x777776c3(%rax) .byte 61,137,136,136,61 // cmp $0x3d888889,%eax .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 4865 <.literal16+0xa85> + .byte 112,65 // jo 48e5 <.literal16+0xae5> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 4869 <.literal16+0xa89> + .byte 112,65 // jo 48e9 <.literal16+0xae9> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 486d <.literal16+0xa8d> + .byte 112,65 // jo 48ed <.literal16+0xaed> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 4871 <.literal16+0xa91> + .byte 112,65 // jo 48f1 <.literal16+0xaf1> .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -28265,11 +28376,11 @@ BALIGN16 .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,0,0,127,67 // addb $0x43,0x7f00003b(%rax) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 48bb <.literal16+0xadb> + .byte 127,67 // jg 493b <.literal16+0xb3b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 48bf <.literal16+0xadf> + .byte 127,67 // jg 493f <.literal16+0xb3f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 48c3 <.literal16+0xae3> + .byte 127,67 // jg 4943 <.literal16+0xb43> .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) .byte 0,0 // add %al,(%rax) .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) @@ -28345,13 +28456,13 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 255 // (bad) - .byte 127,71 // jg 49ab <.literal16+0xbcb> + .byte 127,71 // jg 4a2b <.literal16+0xc2b> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 49af <.literal16+0xbcf> + .byte 127,71 // jg 4a2f <.literal16+0xc2f> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 49b3 <.literal16+0xbd3> + .byte 127,71 // jg 4a33 <.literal16+0xc33> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 49b7 <.literal16+0xbd7> + .byte 127,71 // jg 4a37 <.literal16+0xc37> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -28462,11 +28573,11 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,114 // cmpb $0x72,(%rdi) .byte 28,199 // sbb $0xc7,%al - .byte 62,114,28 // jb,pt 4aa2 <.literal16+0xcc2> + .byte 62,114,28 // jb,pt 4b22 <.literal16+0xd22> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4aa6 <.literal16+0xcc6> + .byte 62,114,28 // jb,pt 4b26 <.literal16+0xd26> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4aaa <.literal16+0xcca> + .byte 62,114,28 // jb,pt 4b2a <.literal16+0xd2a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -28510,7 +28621,7 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d935 <_sk_callback_sse2+0x3d639c2d> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d9b5 <_sk_callback_sse2+0x3d639c8e> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -28536,7 +28647,7 @@ BALIGN16 .byte 0,192 // add %al,%al .byte 63 // (bad) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d975 <_sk_callback_sse2+0x3d639c6d> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d9f5 <_sk_callback_sse2+0x3d639cce> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -28545,13 +28656,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 4b6e <.literal16+0xd8e> + .byte 114,28 // jb 4bee <.literal16+0xdee> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4b72 <.literal16+0xd92> + .byte 62,114,28 // jb,pt 4bf2 <.literal16+0xdf2> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4b76 <.literal16+0xd96> + .byte 62,114,28 // jb,pt 4bf6 <.literal16+0xdf6> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4b7a <.literal16+0xd9a> + .byte 62,114,28 // jb,pt 4bfa <.literal16+0xdfa> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -28572,11 +28683,11 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,114 // cmpb $0x72,(%rdi) .byte 28,199 // sbb $0xc7,%al - .byte 62,114,28 // jb,pt 4bb2 <.literal16+0xdd2> + .byte 62,114,28 // jb,pt 4c32 <.literal16+0xe32> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4bb6 <.literal16+0xdd6> + .byte 62,114,28 // jb,pt 4c36 <.literal16+0xe36> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4bba <.literal16+0xdda> + .byte 62,114,28 // jb,pt 4c3a <.literal16+0xe3a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -28620,7 +28731,7 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63da45 <_sk_callback_sse2+0x3d639d3d> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63dac5 <_sk_callback_sse2+0x3d639d9e> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -28646,7 +28757,7 @@ BALIGN16 .byte 0,192 // add %al,%al .byte 63 // (bad) .byte 57,142,99,61,57,142 // cmp %ecx,-0x71c6c29d(%rsi) - .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63da85 <_sk_callback_sse2+0x3d639d7d> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63db05 <_sk_callback_sse2+0x3d639dde> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -28655,13 +28766,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 4c7e <.literal16+0xe9e> + .byte 114,28 // jb 4cfe <.literal16+0xefe> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4c82 <_sk_callback_sse2+0xf7a> + .byte 62,114,28 // jb,pt 4d02 <_sk_callback_sse2+0xfdb> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4c86 <_sk_callback_sse2+0xf7e> + .byte 62,114,28 // jb,pt 4d06 <_sk_callback_sse2+0xfdf> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 4c8a <_sk_callback_sse2+0xf82> + .byte 62,114,28 // jb,pt 4d0a <_sk_callback_sse2+0xfe3> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 822c90685f..488227b4c3 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -106,14 +106,14 @@ _sk_seed_shader_hsw LABEL PROC DB 197,249,110,199 ; vmovd %edi,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,130,60,0,0 ; vbroadcastss 0x3c82(%rip),%ymm1 # 3ddc <_sk_callback_hsw+0x119> + DB 196,226,125,24,13,194,60,0,0 ; vbroadcastss 0x3cc2(%rip),%ymm1 # 3e1c <_sk_callback_hsw+0x119> 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 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,102,60,0,0 ; vbroadcastss 0x3c66(%rip),%ymm2 # 3de0 <_sk_callback_hsw+0x11d> + DB 196,226,125,24,21,166,60,0,0 ; vbroadcastss 0x3ca6(%rip),%ymm2 # 3e20 <_sk_callback_hsw+0x11d> 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 @@ -143,7 +143,7 @@ _sk_clear_hsw LABEL PROC PUBLIC _sk_srcatop_hsw _sk_srcatop_hsw LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,22,60,0,0 ; vbroadcastss 0x3c16(%rip),%ymm8 # 3de4 <_sk_callback_hsw+0x121> + DB 196,98,125,24,5,86,60,0,0 ; vbroadcastss 0x3c56(%rip),%ymm8 # 3e24 <_sk_callback_hsw+0x121> 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 @@ -157,7 +157,7 @@ _sk_srcatop_hsw LABEL PROC PUBLIC _sk_dstatop_hsw _sk_dstatop_hsw LABEL PROC - DB 196,98,125,24,5,233,59,0,0 ; vbroadcastss 0x3be9(%rip),%ymm8 # 3de8 <_sk_callback_hsw+0x125> + DB 196,98,125,24,5,41,60,0,0 ; vbroadcastss 0x3c29(%rip),%ymm8 # 3e28 <_sk_callback_hsw+0x125> 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 @@ -190,7 +190,7 @@ _sk_dstin_hsw LABEL PROC PUBLIC _sk_srcout_hsw _sk_srcout_hsw LABEL PROC - DB 196,98,125,24,5,144,59,0,0 ; vbroadcastss 0x3b90(%rip),%ymm8 # 3dec <_sk_callback_hsw+0x129> + DB 196,98,125,24,5,208,59,0,0 ; vbroadcastss 0x3bd0(%rip),%ymm8 # 3e2c <_sk_callback_hsw+0x129> 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 @@ -201,7 +201,7 @@ _sk_srcout_hsw LABEL PROC PUBLIC _sk_dstout_hsw _sk_dstout_hsw LABEL PROC - DB 196,226,125,24,5,115,59,0,0 ; vbroadcastss 0x3b73(%rip),%ymm0 # 3df0 <_sk_callback_hsw+0x12d> + DB 196,226,125,24,5,179,59,0,0 ; vbroadcastss 0x3bb3(%rip),%ymm0 # 3e30 <_sk_callback_hsw+0x12d> 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 @@ -212,7 +212,7 @@ _sk_dstout_hsw LABEL PROC PUBLIC _sk_srcover_hsw _sk_srcover_hsw LABEL PROC - DB 196,98,125,24,5,86,59,0,0 ; vbroadcastss 0x3b56(%rip),%ymm8 # 3df4 <_sk_callback_hsw+0x131> + DB 196,98,125,24,5,150,59,0,0 ; vbroadcastss 0x3b96(%rip),%ymm8 # 3e34 <_sk_callback_hsw+0x131> 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 @@ -223,7 +223,7 @@ _sk_srcover_hsw LABEL PROC PUBLIC _sk_dstover_hsw _sk_dstover_hsw LABEL PROC - DB 196,98,125,24,5,53,59,0,0 ; vbroadcastss 0x3b35(%rip),%ymm8 # 3df8 <_sk_callback_hsw+0x135> + DB 196,98,125,24,5,117,59,0,0 ; vbroadcastss 0x3b75(%rip),%ymm8 # 3e38 <_sk_callback_hsw+0x135> 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 @@ -243,7 +243,7 @@ _sk_modulate_hsw LABEL PROC PUBLIC _sk_multiply_hsw _sk_multiply_hsw LABEL PROC - DB 196,98,125,24,5,0,59,0,0 ; vbroadcastss 0x3b00(%rip),%ymm8 # 3dfc <_sk_callback_hsw+0x139> + DB 196,98,125,24,5,64,59,0,0 ; vbroadcastss 0x3b40(%rip),%ymm8 # 3e3c <_sk_callback_hsw+0x139> 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 @@ -285,7 +285,7 @@ _sk_screen_hsw LABEL PROC PUBLIC _sk_xor__hsw _sk_xor__hsw LABEL PROC - DB 196,98,125,24,5,123,58,0,0 ; vbroadcastss 0x3a7b(%rip),%ymm8 # 3e00 <_sk_callback_hsw+0x13d> + DB 196,98,125,24,5,187,58,0,0 ; vbroadcastss 0x3abb(%rip),%ymm8 # 3e40 <_sk_callback_hsw+0x13d> 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 @@ -317,7 +317,7 @@ _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 196,98,125,24,5,3,58,0,0 ; vbroadcastss 0x3a03(%rip),%ymm8 # 3e04 <_sk_callback_hsw+0x141> + DB 196,98,125,24,5,67,58,0,0 ; vbroadcastss 0x3a43(%rip),%ymm8 # 3e44 <_sk_callback_hsw+0x141> 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 @@ -340,7 +340,7 @@ _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 196,98,125,24,5,178,57,0,0 ; vbroadcastss 0x39b2(%rip),%ymm8 # 3e08 <_sk_callback_hsw+0x145> + DB 196,98,125,24,5,242,57,0,0 ; vbroadcastss 0x39f2(%rip),%ymm8 # 3e48 <_sk_callback_hsw+0x145> 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 @@ -366,7 +366,7 @@ _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 196,98,125,24,5,85,57,0,0 ; vbroadcastss 0x3955(%rip),%ymm8 # 3e0c <_sk_callback_hsw+0x149> + DB 196,98,125,24,5,149,57,0,0 ; vbroadcastss 0x3995(%rip),%ymm8 # 3e4c <_sk_callback_hsw+0x149> 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 @@ -386,7 +386,7 @@ _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 196,98,125,24,5,19,57,0,0 ; vbroadcastss 0x3913(%rip),%ymm8 # 3e10 <_sk_callback_hsw+0x14d> + DB 196,98,125,24,5,83,57,0,0 ; vbroadcastss 0x3953(%rip),%ymm8 # 3e50 <_sk_callback_hsw+0x14d> 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 @@ -394,7 +394,7 @@ _sk_exclusion_hsw LABEL PROC PUBLIC _sk_colorburn_hsw _sk_colorburn_hsw LABEL PROC - DB 196,98,125,24,5,1,57,0,0 ; vbroadcastss 0x3901(%rip),%ymm8 # 3e14 <_sk_callback_hsw+0x151> + DB 196,98,125,24,5,65,57,0,0 ; vbroadcastss 0x3941(%rip),%ymm8 # 3e54 <_sk_callback_hsw+0x151> 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 @@ -450,7 +450,7 @@ _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 196,98,125,24,13,12,56,0,0 ; vbroadcastss 0x380c(%rip),%ymm9 # 3e18 <_sk_callback_hsw+0x155> + DB 196,98,125,24,13,76,56,0,0 ; vbroadcastss 0x384c(%rip),%ymm9 # 3e58 <_sk_callback_hsw+0x155> 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 @@ -501,7 +501,7 @@ _sk_colordodge_hsw LABEL PROC PUBLIC _sk_hardlight_hsw _sk_hardlight_hsw LABEL PROC - DB 196,98,125,24,5,45,55,0,0 ; vbroadcastss 0x372d(%rip),%ymm8 # 3e1c <_sk_callback_hsw+0x159> + DB 196,98,125,24,5,109,55,0,0 ; vbroadcastss 0x376d(%rip),%ymm8 # 3e5c <_sk_callback_hsw+0x159> 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 @@ -550,7 +550,7 @@ _sk_hardlight_hsw LABEL PROC PUBLIC _sk_overlay_hsw _sk_overlay_hsw LABEL PROC - DB 196,98,125,24,5,101,54,0,0 ; vbroadcastss 0x3665(%rip),%ymm8 # 3e20 <_sk_callback_hsw+0x15d> + DB 196,98,125,24,5,165,54,0,0 ; vbroadcastss 0x36a5(%rip),%ymm8 # 3e60 <_sk_callback_hsw+0x15d> 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 @@ -610,10 +610,10 @@ _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 196,98,125,24,29,108,53,0,0 ; vbroadcastss 0x356c(%rip),%ymm11 # 3e28 <_sk_callback_hsw+0x165> + DB 196,98,125,24,29,172,53,0,0 ; vbroadcastss 0x35ac(%rip),%ymm11 # 3e68 <_sk_callback_hsw+0x165> DB 196,65,20,88,227 ; vaddps %ymm11,%ymm13,%ymm12 DB 196,65,28,89,192 ; vmulps %ymm8,%ymm12,%ymm8 - DB 196,98,125,24,37,93,53,0,0 ; vbroadcastss 0x355d(%rip),%ymm12 # 3e2c <_sk_callback_hsw+0x169> + DB 196,98,125,24,37,157,53,0,0 ; vbroadcastss 0x359d(%rip),%ymm12 # 3e6c <_sk_callback_hsw+0x169> 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 @@ -623,7 +623,7 @@ _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 196,98,125,24,5,32,53,0,0 ; vbroadcastss 0x3520(%rip),%ymm8 # 3e24 <_sk_callback_hsw+0x161> + DB 196,98,125,24,5,96,53,0,0 ; vbroadcastss 0x3560(%rip),%ymm8 # 3e64 <_sk_callback_hsw+0x161> 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 @@ -713,7 +713,7 @@ _sk_clamp_0_hsw LABEL PROC PUBLIC _sk_clamp_1_hsw _sk_clamp_1_hsw LABEL PROC - DB 196,98,125,24,5,163,51,0,0 ; vbroadcastss 0x33a3(%rip),%ymm8 # 3e30 <_sk_callback_hsw+0x16d> + DB 196,98,125,24,5,227,51,0,0 ; vbroadcastss 0x33e3(%rip),%ymm8 # 3e70 <_sk_callback_hsw+0x16d> 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 @@ -723,7 +723,7 @@ _sk_clamp_1_hsw LABEL PROC PUBLIC _sk_clamp_a_hsw _sk_clamp_a_hsw LABEL PROC - DB 196,98,125,24,5,134,51,0,0 ; vbroadcastss 0x3386(%rip),%ymm8 # 3e34 <_sk_callback_hsw+0x171> + DB 196,98,125,24,5,198,51,0,0 ; vbroadcastss 0x33c6(%rip),%ymm8 # 3e74 <_sk_callback_hsw+0x171> 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 @@ -795,7 +795,7 @@ 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 196,98,125,24,21,206,50,0,0 ; vbroadcastss 0x32ce(%rip),%ymm10 # 3e38 <_sk_callback_hsw+0x175> + DB 196,98,125,24,21,14,51,0,0 ; vbroadcastss 0x330e(%rip),%ymm10 # 3e78 <_sk_callback_hsw+0x175> 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 @@ -806,16 +806,16 @@ _sk_unpremul_hsw LABEL PROC PUBLIC _sk_from_srgb_hsw _sk_from_srgb_hsw LABEL PROC - DB 196,98,125,24,5,175,50,0,0 ; vbroadcastss 0x32af(%rip),%ymm8 # 3e3c <_sk_callback_hsw+0x179> + DB 196,98,125,24,5,239,50,0,0 ; vbroadcastss 0x32ef(%rip),%ymm8 # 3e7c <_sk_callback_hsw+0x179> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,161,50,0,0 ; vbroadcastss 0x32a1(%rip),%ymm11 # 3e40 <_sk_callback_hsw+0x17d> - DB 196,98,125,24,37,156,50,0,0 ; vbroadcastss 0x329c(%rip),%ymm12 # 3e44 <_sk_callback_hsw+0x181> + DB 196,98,125,24,29,225,50,0,0 ; vbroadcastss 0x32e1(%rip),%ymm11 # 3e80 <_sk_callback_hsw+0x17d> + DB 196,98,125,24,37,220,50,0,0 ; vbroadcastss 0x32dc(%rip),%ymm12 # 3e84 <_sk_callback_hsw+0x181> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13 - DB 196,98,125,24,53,141,50,0,0 ; vbroadcastss 0x328d(%rip),%ymm14 # 3e48 <_sk_callback_hsw+0x185> + DB 196,98,125,24,53,205,50,0,0 ; vbroadcastss 0x32cd(%rip),%ymm14 # 3e88 <_sk_callback_hsw+0x185> DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13 - DB 196,98,125,24,21,131,50,0,0 ; vbroadcastss 0x3283(%rip),%ymm10 # 3e4c <_sk_callback_hsw+0x189> + DB 196,98,125,24,21,195,50,0,0 ; vbroadcastss 0x32c3(%rip),%ymm10 # 3e8c <_sk_callback_hsw+0x189> 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 @@ -839,16 +839,16 @@ _sk_to_srgb_hsw LABEL PROC DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8 DB 196,65,124,83,200 ; vrcpps %ymm8,%ymm9 DB 196,65,124,82,208 ; vrsqrtps %ymm8,%ymm10 - DB 196,98,125,24,5,29,50,0,0 ; vbroadcastss 0x321d(%rip),%ymm8 # 3e50 <_sk_callback_hsw+0x18d> + DB 196,98,125,24,5,93,50,0,0 ; vbroadcastss 0x325d(%rip),%ymm8 # 3e90 <_sk_callback_hsw+0x18d> DB 196,65,124,89,216 ; vmulps %ymm8,%ymm0,%ymm11 - DB 196,98,125,24,37,19,50,0,0 ; vbroadcastss 0x3213(%rip),%ymm12 # 3e54 <_sk_callback_hsw+0x191> - DB 196,98,125,24,45,14,50,0,0 ; vbroadcastss 0x320e(%rip),%ymm13 # 3e58 <_sk_callback_hsw+0x195> + DB 196,98,125,24,37,83,50,0,0 ; vbroadcastss 0x3253(%rip),%ymm12 # 3e94 <_sk_callback_hsw+0x191> + DB 196,98,125,24,45,78,50,0,0 ; vbroadcastss 0x324e(%rip),%ymm13 # 3e98 <_sk_callback_hsw+0x195> DB 196,66,21,168,204 ; vfmadd213ps %ymm12,%ymm13,%ymm9 - DB 196,98,125,24,53,4,50,0,0 ; vbroadcastss 0x3204(%rip),%ymm14 # 3e5c <_sk_callback_hsw+0x199> + DB 196,98,125,24,53,68,50,0,0 ; vbroadcastss 0x3244(%rip),%ymm14 # 3e9c <_sk_callback_hsw+0x199> DB 196,66,13,184,202 ; vfmadd231ps %ymm10,%ymm14,%ymm9 - DB 196,98,125,24,21,250,49,0,0 ; vbroadcastss 0x31fa(%rip),%ymm10 # 3e60 <_sk_callback_hsw+0x19d> + DB 196,98,125,24,21,58,50,0,0 ; vbroadcastss 0x323a(%rip),%ymm10 # 3ea0 <_sk_callback_hsw+0x19d> DB 196,65,44,93,201 ; vminps %ymm9,%ymm10,%ymm9 - DB 196,98,125,24,61,240,49,0,0 ; vbroadcastss 0x31f0(%rip),%ymm15 # 3e64 <_sk_callback_hsw+0x1a1> + DB 196,98,125,24,61,48,50,0,0 ; vbroadcastss 0x3230(%rip),%ymm15 # 3ea4 <_sk_callback_hsw+0x1a1> DB 196,193,124,194,199,1 ; vcmpltps %ymm15,%ymm0,%ymm0 DB 196,195,53,74,195,0 ; vblendvps %ymm0,%ymm11,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -879,26 +879,26 @@ _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 196,98,125,24,29,106,49,0,0 ; vbroadcastss 0x316a(%rip),%ymm11 # 3e68 <_sk_callback_hsw+0x1a5> + DB 196,98,125,24,29,170,49,0,0 ; vbroadcastss 0x31aa(%rip),%ymm11 # 3ea8 <_sk_callback_hsw+0x1a5> 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 196,98,125,24,53,87,49,0,0 ; vbroadcastss 0x3157(%rip),%ymm14 # 3e6c <_sk_callback_hsw+0x1a9> + DB 196,98,125,24,53,151,49,0,0 ; vbroadcastss 0x3197(%rip),%ymm14 # 3eac <_sk_callback_hsw+0x1a9> 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 196,98,125,24,53,62,49,0,0 ; vbroadcastss 0x313e(%rip),%ymm14 # 3e74 <_sk_callback_hsw+0x1b1> + DB 196,98,125,24,53,126,49,0,0 ; vbroadcastss 0x317e(%rip),%ymm14 # 3eb4 <_sk_callback_hsw+0x1b1> DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13 - DB 196,98,125,24,53,44,49,0,0 ; vbroadcastss 0x312c(%rip),%ymm14 # 3e70 <_sk_callback_hsw+0x1ad> + DB 196,98,125,24,53,108,49,0,0 ; vbroadcastss 0x316c(%rip),%ymm14 # 3eb0 <_sk_callback_hsw+0x1ad> 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 196,98,125,24,29,15,49,0,0 ; vbroadcastss 0x310f(%rip),%ymm11 # 3e7c <_sk_callback_hsw+0x1b9> + DB 196,98,125,24,29,79,49,0,0 ; vbroadcastss 0x314f(%rip),%ymm11 # 3ebc <_sk_callback_hsw+0x1b9> 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 @@ -908,112 +908,115 @@ _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 196,98,125,24,5,210,48,0,0 ; vbroadcastss 0x30d2(%rip),%ymm8 # 3e78 <_sk_callback_hsw+0x1b5> + DB 196,98,125,24,5,18,49,0,0 ; vbroadcastss 0x3112(%rip),%ymm8 # 3eb8 <_sk_callback_hsw+0x1b5> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_hsl_to_rgb_hsw _sk_hsl_to_rgb_hsw LABEL PROC - 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 197,252,17,28,36 ; vmovups %ymm3,(%rsp) - DB 197,252,40,233 ; vmovaps %ymm1,%ymm5 + DB 72,129,236,248,0,0,0 ; sub $0xf8,%rsp + DB 197,252,17,188,36,192,0,0,0 ; vmovups %ymm7,0xc0(%rsp) + DB 197,252,17,180,36,160,0,0,0 ; vmovups %ymm6,0xa0(%rsp) + DB 197,252,17,172,36,128,0,0,0 ; vmovups %ymm5,0x80(%rsp) + DB 197,252,17,100,36,96 ; vmovups %ymm4,0x60(%rsp) + DB 197,252,17,92,36,64 ; vmovups %ymm3,0x40(%rsp) + DB 197,252,40,234 ; vmovaps %ymm2,%ymm5 + DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 184,0,0,0,63 ; mov $0x3f000000,%eax - DB 197,249,110,200 ; vmovd %eax,%xmm1 - DB 196,98,125,88,193 ; vpbroadcastd %xmm1,%ymm8 - DB 196,193,108,194,200,1 ; vcmpltps %ymm8,%ymm2,%ymm1 + DB 197,249,110,192 ; vmovd %eax,%xmm0 + DB 196,98,125,88,192 ; vpbroadcastd %xmm0,%ymm8 + DB 196,193,84,194,192,1 ; vcmpltps %ymm8,%ymm5,%ymm0 + DB 196,98,125,24,21,190,48,0,0 ; vbroadcastss 0x30be(%rip),%ymm10 # 3ec0 <_sk_callback_hsw+0x1bd> + DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp) + DB 196,193,116,88,218 ; vaddps %ymm10,%ymm1,%ymm3 + DB 197,228,89,221 ; vmulps %ymm5,%ymm3,%ymm3 + DB 197,244,88,229 ; vaddps %ymm5,%ymm1,%ymm4 + DB 196,226,117,188,229 ; vfnmadd231ps %ymm5,%ymm1,%ymm4 + DB 196,99,93,74,203,0 ; vblendvps %ymm0,%ymm3,%ymm4,%ymm9 + DB 196,226,125,24,13,159,48,0,0 ; vbroadcastss 0x309f(%rip),%ymm1 # 3ec8 <_sk_callback_hsw+0x1c5> + DB 197,236,88,241 ; vaddps %ymm1,%ymm2,%ymm6 + DB 65,184,0,0,0,0 ; mov $0x0,%r8d DB 184,0,0,128,63 ; mov $0x3f800000,%eax - DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 196,98,125,88,212 ; vpbroadcastd %xmm4,%ymm10 - DB 197,172,88,229 ; vaddps %ymm5,%ymm10,%ymm4 - DB 197,220,89,226 ; vmulps %ymm2,%ymm4,%ymm4 - DB 197,84,88,202 ; vaddps %ymm2,%ymm5,%ymm9 - DB 196,98,85,188,202 ; vfnmadd231ps %ymm2,%ymm5,%ymm9 - DB 196,99,53,74,204,16 ; vblendvps %ymm1,%ymm4,%ymm9,%ymm9 - DB 65,184,0,0,0,64 ; mov $0x40000000,%r8d - DB 184,171,170,170,62 ; mov $0x3eaaaaab,%eax - DB 197,249,110,200 ; vmovd %eax,%xmm1 - DB 196,98,125,88,233 ; vpbroadcastd %xmm1,%ymm13 - DB 197,148,88,224 ; vaddps %ymm0,%ymm13,%ymm4 - DB 184,0,0,0,0 ; mov $0x0,%eax DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,98,125,88,225 ; vpbroadcastd %xmm1,%ymm12 - DB 197,172,194,204,1 ; vcmpltps %ymm4,%ymm10,%ymm1 - DB 196,65,92,92,218 ; vsubps %ymm10,%ymm4,%ymm11 - DB 196,195,93,74,203,16 ; vblendvps %ymm1,%ymm11,%ymm4,%ymm1 - DB 196,65,92,194,220,1 ; vcmpltps %ymm12,%ymm4,%ymm11 - DB 197,44,88,244 ; vaddps %ymm4,%ymm10,%ymm14 - DB 196,195,117,74,206,176 ; vblendvps %ymm11,%ymm14,%ymm1,%ymm1 + DB 197,156,194,206,1 ; vcmpltps %ymm6,%ymm12,%ymm1 + DB 196,98,125,24,45,125,48,0,0 ; vbroadcastss 0x307d(%rip),%ymm13 # 3ecc <_sk_callback_hsw+0x1c9> + DB 196,193,76,88,221 ; vaddps %ymm13,%ymm6,%ymm3 + DB 196,227,77,74,203,16 ; vblendvps %ymm1,%ymm3,%ymm6,%ymm1 DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,219 ; vpbroadcastd %xmm3,%ymm11 - DB 196,66,109,170,217 ; vfmsub213ps %ymm9,%ymm2,%ymm11 + DB 196,98,125,88,251 ; vpbroadcastd %xmm3,%ymm15 + DB 196,193,76,194,223,1 ; vcmpltps %ymm15,%ymm6,%ymm3 + DB 196,193,76,88,226 ; vaddps %ymm10,%ymm6,%ymm4 + DB 196,227,117,74,196,48 ; vblendvps %ymm3,%ymm4,%ymm1,%ymm0 + DB 196,98,125,24,29,70,48,0,0 ; vbroadcastss 0x3046(%rip),%ymm11 # 3ec4 <_sk_callback_hsw+0x1c1> + DB 196,66,85,170,217 ; vfmsub213ps %ymm9,%ymm5,%ymm11 + DB 196,193,52,92,203 ; vsubps %ymm11,%ymm9,%ymm1 + DB 196,226,125,24,29,63,48,0,0 ; vbroadcastss 0x303f(%rip),%ymm3 # 3ed0 <_sk_callback_hsw+0x1cd> + DB 197,116,89,243 ; vmulps %ymm3,%ymm1,%ymm14 DB 65,184,171,170,42,62 ; mov $0x3e2aaaab,%r8d - DB 184,0,0,192,64 ; mov $0x40c00000,%eax - DB 197,249,110,216 ; vmovd %eax,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 196,65,52,92,243 ; vsubps %ymm11,%ymm9,%ymm14 - DB 197,12,89,243 ; vmulps %ymm3,%ymm14,%ymm14 DB 184,171,170,42,63 ; mov $0x3f2aaaab,%eax - DB 197,249,110,216 ; vmovd %eax,%xmm3 - DB 196,98,125,88,251 ; vpbroadcastd %xmm3,%ymm15 - DB 197,132,92,217 ; vsubps %ymm1,%ymm15,%ymm3 - DB 196,194,13,168,219 ; vfmadd213ps %ymm11,%ymm14,%ymm3 - DB 196,193,116,194,255,1 ; vcmpltps %ymm15,%ymm1,%ymm7 - DB 196,227,37,74,219,112 ; vblendvps %ymm7,%ymm3,%ymm11,%ymm3 - DB 196,193,116,194,248,1 ; vcmpltps %ymm8,%ymm1,%ymm7 - DB 196,195,101,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm3,%ymm7 - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 197,244,194,203,1 ; vcmpltps %ymm3,%ymm1,%ymm1 - DB 196,194,13,168,227 ; vfmadd213ps %ymm11,%ymm14,%ymm4 - DB 196,227,69,74,228,16 ; vblendvps %ymm1,%ymm4,%ymm7,%ymm4 - DB 197,172,194,200,1 ; vcmpltps %ymm0,%ymm10,%ymm1 - DB 196,193,124,92,250 ; vsubps %ymm10,%ymm0,%ymm7 - DB 196,227,125,74,207,16 ; vblendvps %ymm1,%ymm7,%ymm0,%ymm1 - DB 196,193,124,194,252,1 ; vcmpltps %ymm12,%ymm0,%ymm7 - DB 197,172,88,240 ; vaddps %ymm0,%ymm10,%ymm6 - DB 196,227,117,74,206,112 ; vblendvps %ymm7,%ymm6,%ymm1,%ymm1 - DB 197,132,92,241 ; vsubps %ymm1,%ymm15,%ymm6 + DB 197,249,110,200 ; vmovd %eax,%xmm1 + DB 196,226,125,88,225 ; vpbroadcastd %xmm1,%ymm4 + DB 196,226,125,24,29,34,48,0,0 ; vbroadcastss 0x3022(%rip),%ymm3 # 3ed4 <_sk_callback_hsw+0x1d1> + DB 197,228,92,200 ; vsubps %ymm0,%ymm3,%ymm1 + DB 196,194,13,168,203 ; vfmadd213ps %ymm11,%ymm14,%ymm1 + DB 197,252,194,252,1 ; vcmpltps %ymm4,%ymm0,%ymm7 + DB 196,227,37,74,201,112 ; vblendvps %ymm7,%ymm1,%ymm11,%ymm1 + DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7 + DB 196,195,117,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm1,%ymm7 + DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 + DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 + DB 197,252,194,193,1 ; vcmpltps %ymm1,%ymm0,%ymm0 + DB 196,194,13,168,243 ; vfmadd213ps %ymm11,%ymm14,%ymm6 + DB 196,227,69,74,198,0 ; vblendvps %ymm0,%ymm6,%ymm7,%ymm0 + DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) + DB 197,156,194,194,1 ; vcmpltps %ymm2,%ymm12,%ymm0 + DB 196,193,108,88,253 ; vaddps %ymm13,%ymm2,%ymm7 + DB 196,227,109,74,199,0 ; vblendvps %ymm0,%ymm7,%ymm2,%ymm0 + DB 196,193,108,194,255,1 ; vcmpltps %ymm15,%ymm2,%ymm7 + DB 196,193,108,88,242 ; vaddps %ymm10,%ymm2,%ymm6 + DB 196,227,125,74,198,112 ; vblendvps %ymm7,%ymm6,%ymm0,%ymm0 + DB 197,228,92,240 ; vsubps %ymm0,%ymm3,%ymm6 DB 196,194,13,168,243 ; vfmadd213ps %ymm11,%ymm14,%ymm6 - DB 196,193,116,194,255,1 ; vcmpltps %ymm15,%ymm1,%ymm7 + DB 197,252,194,252,1 ; vcmpltps %ymm4,%ymm0,%ymm7 DB 196,227,37,74,246,112 ; vblendvps %ymm7,%ymm6,%ymm11,%ymm6 - DB 196,193,116,194,248,1 ; vcmpltps %ymm8,%ymm1,%ymm7 + DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7 DB 196,195,77,74,241,112 ; vblendvps %ymm7,%ymm9,%ymm6,%ymm6 - DB 197,244,194,203,1 ; vcmpltps %ymm3,%ymm1,%ymm1 - DB 196,193,124,92,253 ; vsubps %ymm13,%ymm0,%ymm7 - DB 196,194,13,168,195 ; vfmadd213ps %ymm11,%ymm14,%ymm0 - DB 196,227,77,74,200,16 ; vblendvps %ymm1,%ymm0,%ymm6,%ymm1 - DB 197,172,194,199,1 ; vcmpltps %ymm7,%ymm10,%ymm0 - DB 196,193,68,92,242 ; vsubps %ymm10,%ymm7,%ymm6 - DB 196,227,69,74,198,0 ; vblendvps %ymm0,%ymm6,%ymm7,%ymm0 - DB 196,193,68,194,244,1 ; vcmpltps %ymm12,%ymm7,%ymm6 - DB 197,44,88,215 ; vaddps %ymm7,%ymm10,%ymm10 - DB 196,195,125,74,194,96 ; vblendvps %ymm6,%ymm10,%ymm0,%ymm0 + DB 197,252,194,193,1 ; vcmpltps %ymm1,%ymm0,%ymm0 + DB 197,252,40,250 ; vmovaps %ymm2,%ymm7 DB 196,194,13,168,251 ; vfmadd213ps %ymm11,%ymm14,%ymm7 - DB 197,132,92,240 ; vsubps %ymm0,%ymm15,%ymm6 - DB 196,194,13,168,243 ; vfmadd213ps %ymm11,%ymm14,%ymm6 - DB 196,65,124,194,215,1 ; vcmpltps %ymm15,%ymm0,%ymm10 - DB 196,227,37,74,246,160 ; vblendvps %ymm10,%ymm6,%ymm11,%ymm6 - DB 196,65,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm8 - DB 196,195,77,74,241,128 ; vblendvps %ymm8,%ymm9,%ymm6,%ymm6 - DB 197,252,194,195,1 ; vcmpltps %ymm3,%ymm0,%ymm0 - DB 196,227,77,74,223,0 ; vblendvps %ymm0,%ymm7,%ymm6,%ymm3 + DB 196,227,77,74,247,0 ; vblendvps %ymm0,%ymm7,%ymm6,%ymm6 + DB 196,226,125,24,5,137,47,0,0 ; vbroadcastss 0x2f89(%rip),%ymm0 # 3ed8 <_sk_callback_hsw+0x1d5> + DB 197,236,88,192 ; vaddps %ymm0,%ymm2,%ymm0 + DB 197,156,194,208,1 ; vcmpltps %ymm0,%ymm12,%ymm2 + DB 196,193,124,88,253 ; vaddps %ymm13,%ymm0,%ymm7 + DB 196,227,125,74,215,32 ; vblendvps %ymm2,%ymm7,%ymm0,%ymm2 + DB 196,193,124,194,255,1 ; vcmpltps %ymm15,%ymm0,%ymm7 + DB 196,65,124,88,210 ; vaddps %ymm10,%ymm0,%ymm10 + DB 196,195,109,74,210,112 ; vblendvps %ymm7,%ymm10,%ymm2,%ymm2 + DB 196,194,13,168,195 ; vfmadd213ps %ymm11,%ymm14,%ymm0 + DB 197,228,92,218 ; vsubps %ymm2,%ymm3,%ymm3 + DB 196,194,13,168,219 ; vfmadd213ps %ymm11,%ymm14,%ymm3 + DB 197,236,194,228,1 ; vcmpltps %ymm4,%ymm2,%ymm4 + DB 196,227,37,74,219,64 ; vblendvps %ymm4,%ymm3,%ymm11,%ymm3 + DB 196,193,108,194,224,1 ; vcmpltps %ymm8,%ymm2,%ymm4 + DB 196,195,101,74,217,64 ; vblendvps %ymm4,%ymm9,%ymm3,%ymm3 + DB 197,236,194,201,1 ; vcmpltps %ymm1,%ymm2,%ymm1 + DB 196,227,101,74,208,16 ; vblendvps %ymm1,%ymm0,%ymm3,%ymm2 DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 - DB 197,212,194,232,0 ; vcmpeqps %ymm0,%ymm5,%ymm5 - DB 196,227,93,74,194,80 ; vblendvps %ymm5,%ymm2,%ymm4,%ymm0 - DB 196,227,117,74,202,80 ; vblendvps %ymm5,%ymm2,%ymm1,%ymm1 - DB 196,227,101,74,210,80 ; vblendvps %ymm5,%ymm2,%ymm3,%ymm2 + DB 197,252,194,92,36,32,0 ; vcmpeqps 0x20(%rsp),%ymm0,%ymm3 + DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 + DB 196,227,125,74,197,48 ; vblendvps %ymm3,%ymm5,%ymm0,%ymm0 + DB 196,227,77,74,205,48 ; vblendvps %ymm3,%ymm5,%ymm6,%ymm1 + DB 196,227,109,74,213,48 ; vblendvps %ymm3,%ymm5,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,28,36 ; vmovups (%rsp),%ymm3 - 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 72,129,196,184,0,0,0 ; add $0xb8,%rsp + DB 197,252,16,92,36,64 ; vmovups 0x40(%rsp),%ymm3 + DB 197,252,16,100,36,96 ; vmovups 0x60(%rsp),%ymm4 + DB 197,252,16,172,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm5 + DB 197,252,16,180,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm6 + DB 197,252,16,188,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm7 + DB 72,129,196,248,0,0,0 ; add $0xf8,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_scale_1_float_hsw @@ -1034,11 +1037,11 @@ _sk_scale_u8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,51 ; jne 1019 <_sk_scale_u8_hsw+0x43> + DB 117,51 ; jne 1056 <_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 196,98,125,24,13,131,46,0,0 ; vbroadcastss 0x2e83(%rip),%ymm9 # 3e80 <_sk_callback_hsw+0x1bd> + DB 196,98,125,24,13,162,46,0,0 ; vbroadcastss 0x2ea2(%rip),%ymm9 # 3edc <_sk_callback_hsw+0x1d9> 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 @@ -1056,9 +1059,9 @@ _sk_scale_u8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 1021 <_sk_scale_u8_hsw+0x4b> + DB 117,234 ; jne 105e <_sk_scale_u8_hsw+0x4b> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,172 ; jmp fea <_sk_scale_u8_hsw+0x14> + DB 235,172 ; jmp 1027 <_sk_scale_u8_hsw+0x14> PUBLIC _sk_lerp_1_float_hsw _sk_lerp_1_float_hsw LABEL PROC @@ -1082,11 +1085,11 @@ _sk_lerp_u8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,71 ; jne 10c4 <_sk_lerp_u8_hsw+0x57> + DB 117,71 ; jne 1101 <_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 196,98,125,24,13,240,45,0,0 ; vbroadcastss 0x2df0(%rip),%ymm9 # 3e84 <_sk_callback_hsw+0x1c1> + DB 196,98,125,24,13,15,46,0,0 ; vbroadcastss 0x2e0f(%rip),%ymm9 # 3ee0 <_sk_callback_hsw+0x1dd> 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 @@ -1108,32 +1111,32 @@ _sk_lerp_u8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 10cc <_sk_lerp_u8_hsw+0x5f> + DB 117,234 ; jne 1109 <_sk_lerp_u8_hsw+0x5f> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,152 ; jmp 1081 <_sk_lerp_u8_hsw+0x14> + DB 235,152 ; jmp 10be <_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 15,133,149,0,0,0 ; jne 118c <_sk_lerp_565_hsw+0xa3> + DB 15,133,149,0,0,0 ; jne 11c9 <_sk_lerp_565_hsw+0xa3> DB 196,193,122,111,28,122 ; vmovdqu (%r10,%rdi,2),%xmm3 DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 - DB 196,98,125,88,5,125,45,0,0 ; vpbroadcastd 0x2d7d(%rip),%ymm8 # 3e88 <_sk_callback_hsw+0x1c5> + DB 196,98,125,88,5,156,45,0,0 ; vpbroadcastd 0x2d9c(%rip),%ymm8 # 3ee4 <_sk_callback_hsw+0x1e1> DB 196,65,101,219,192 ; vpand %ymm8,%ymm3,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,110,45,0,0 ; vbroadcastss 0x2d6e(%rip),%ymm9 # 3e8c <_sk_callback_hsw+0x1c9> + DB 196,98,125,24,13,141,45,0,0 ; vbroadcastss 0x2d8d(%rip),%ymm9 # 3ee8 <_sk_callback_hsw+0x1e5> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,88,13,100,45,0,0 ; vpbroadcastd 0x2d64(%rip),%ymm9 # 3e90 <_sk_callback_hsw+0x1cd> + DB 196,98,125,88,13,131,45,0,0 ; vpbroadcastd 0x2d83(%rip),%ymm9 # 3eec <_sk_callback_hsw+0x1e9> DB 196,65,101,219,201 ; vpand %ymm9,%ymm3,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,85,45,0,0 ; vbroadcastss 0x2d55(%rip),%ymm10 # 3e94 <_sk_callback_hsw+0x1d1> + DB 196,98,125,24,21,116,45,0,0 ; vbroadcastss 0x2d74(%rip),%ymm10 # 3ef0 <_sk_callback_hsw+0x1ed> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,88,21,75,45,0,0 ; vpbroadcastd 0x2d4b(%rip),%ymm10 # 3e98 <_sk_callback_hsw+0x1d5> + DB 196,98,125,88,21,106,45,0,0 ; vpbroadcastd 0x2d6a(%rip),%ymm10 # 3ef4 <_sk_callback_hsw+0x1f1> DB 196,193,101,219,218 ; vpand %ymm10,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,21,61,45,0,0 ; vbroadcastss 0x2d3d(%rip),%ymm10 # 3e9c <_sk_callback_hsw+0x1d9> + DB 196,98,125,24,21,92,45,0,0 ; vbroadcastss 0x2d5c(%rip),%ymm10 # 3ef8 <_sk_callback_hsw+0x1f5> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -1142,16 +1145,16 @@ _sk_lerp_565_hsw LABEL PROC DB 197,236,92,214 ; vsubps %ymm6,%ymm2,%ymm2 DB 196,226,101,168,214 ; vfmadd213ps %ymm6,%ymm3,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,22,45,0,0 ; vbroadcastss 0x2d16(%rip),%ymm3 # 3ea0 <_sk_callback_hsw+0x1dd> + DB 196,226,125,24,29,53,45,0,0 ; vbroadcastss 0x2d35(%rip),%ymm3 # 3efc <_sk_callback_hsw+0x1f9> DB 255,224 ; jmpq *%rax DB 65,137,200 ; mov %ecx,%r8d DB 65,128,224,7 ; and $0x7,%r8b DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,89,255,255,255 ; ja 10fd <_sk_lerp_565_hsw+0x14> + DB 15,135,89,255,255,255 ; ja 113a <_sk_lerp_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 11f8 <_sk_lerp_565_hsw+0x10f> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1238 <_sk_lerp_565_hsw+0x112> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -1163,27 +1166,26 @@ _sk_lerp_565_hsw LABEL PROC DB 196,193,97,196,92,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3 DB 196,193,97,196,92,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3 DB 196,193,97,196,28,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3 - DB 233,5,255,255,255 ; jmpq 10fd <_sk_lerp_565_hsw+0x14> - DB 244 ; hlt + DB 233,5,255,255,255 ; jmpq 113a <_sk_lerp_565_hsw+0x14> + DB 15,31,0 ; nopl (%rax) + DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 236 ; in (%dx),%al + DB 233,255,255,255,225 ; jmpq ffffffffe2001240 <_sk_callback_hsw+0xffffffffe1ffd53d> DB 255 ; (bad) DB 255 ; (bad) - DB 255,228 ; jmpq *%rsp DB 255 ; (bad) + DB 217,255 ; fcos DB 255 ; (bad) + DB 255,209 ; callq *%rcx DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) - DB 255 ; (bad) - DB 255,212 ; callq *%rsp DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) - DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax + DB 189 ; .byte 0xbd DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -1195,23 +1197,23 @@ _sk_load_tables_hsw LABEL PROC DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 - DB 117,105 ; jne 1292 <_sk_load_tables_hsw+0x7e> + DB 117,105 ; jne 12d2 <_sk_load_tables_hsw+0x7e> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 - DB 197,229,219,13,10,47,0,0 ; vpand 0x2f0a(%rip),%ymm3,%ymm1 # 4140 <_sk_callback_hsw+0x47d> + DB 197,229,219,13,42,47,0,0 ; vpand 0x2f2a(%rip),%ymm3,%ymm1 # 41a0 <_sk_callback_hsw+0x49d> 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 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 DB 196,226,109,146,4,137 ; vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0 - DB 196,226,101,0,21,10,47,0,0 ; vpshufb 0x2f0a(%rip),%ymm3,%ymm2 # 4160 <_sk_callback_hsw+0x49d> + DB 196,226,101,0,21,42,47,0,0 ; vpshufb 0x2f2a(%rip),%ymm3,%ymm2 # 41c0 <_sk_callback_hsw+0x4bd> DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 DB 196,194,53,146,12,145 ; vgatherdps %ymm9,(%r9,%ymm2,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,98,101,0,13,18,47,0,0 ; vpshufb 0x2f12(%rip),%ymm3,%ymm9 # 4180 <_sk_callback_hsw+0x4bd> + DB 196,98,101,0,13,50,47,0,0 ; vpshufb 0x2f32(%rip),%ymm3,%ymm9 # 41e0 <_sk_callback_hsw+0x4dd> 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 196,98,125,24,5,30,44,0,0 ; vbroadcastss 0x2c1e(%rip),%ymm8 # 3ea4 <_sk_callback_hsw+0x1e1> + DB 196,98,125,24,5,58,44,0,0 ; vbroadcastss 0x2c3a(%rip),%ymm8 # 3f00 <_sk_callback_hsw+0x1fd> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx @@ -1224,7 +1226,7 @@ _sk_load_tables_hsw LABEL PROC DB 196,193,249,110,194 ; vmovq %r10,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3 - DB 233,115,255,255,255 ; jmpq 122e <_sk_load_tables_hsw+0x1a> + DB 233,115,255,255,255 ; jmpq 126e <_sk_load_tables_hsw+0x1a> PUBLIC _sk_load_tables_u16_be_hsw _sk_load_tables_u16_be_hsw LABEL PROC @@ -1232,7 +1234,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC 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 139a <_sk_load_tables_u16_be_hsw+0xdf> + DB 15,133,201,0,0,0 ; jne 13da <_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 @@ -1248,7 +1250,7 @@ _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 197,121,111,21,158,47,0,0 ; vmovdqa 0x2f9e(%rip),%xmm10 # 42c0 <_sk_callback_hsw+0x5fd> + DB 197,121,111,21,190,47,0,0 ; vmovdqa 0x2fbe(%rip),%xmm10 # 4320 <_sk_callback_hsw+0x61d> 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 @@ -1270,36 +1272,36 @@ _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 196,98,125,24,5,23,43,0,0 ; vbroadcastss 0x2b17(%rip),%ymm8 # 3ea8 <_sk_callback_hsw+0x1e5> + DB 196,98,125,24,5,51,43,0,0 ; vbroadcastss 0x2b33(%rip),%ymm8 # 3f04 <_sk_callback_hsw+0x201> 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 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 1400 <_sk_load_tables_u16_be_hsw+0x145> + DB 116,85 ; je 1440 <_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 1400 <_sk_load_tables_u16_be_hsw+0x145> + DB 114,72 ; jb 1440 <_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 140d <_sk_load_tables_u16_be_hsw+0x152> + DB 116,72 ; je 144d <_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 140d <_sk_load_tables_u16_be_hsw+0x152> + DB 114,59 ; jb 144d <_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 12ec <_sk_load_tables_u16_be_hsw+0x31> + DB 15,132,9,255,255,255 ; je 132c <_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 12ec <_sk_load_tables_u16_be_hsw+0x31> + DB 15,130,248,254,255,255 ; jb 132c <_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 12ec <_sk_load_tables_u16_be_hsw+0x31> + DB 233,236,254,255,255 ; jmpq 132c <_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 12ec <_sk_load_tables_u16_be_hsw+0x31> + DB 233,223,254,255,255 ; jmpq 132c <_sk_load_tables_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,214,254,255,255 ; jmpq 12ec <_sk_load_tables_u16_be_hsw+0x31> + DB 233,214,254,255,255 ; jmpq 132c <_sk_load_tables_u16_be_hsw+0x31> PUBLIC _sk_load_tables_rgb_u16_be_hsw _sk_load_tables_rgb_u16_be_hsw LABEL PROC @@ -1307,7 +1309,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC 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 14e9 <_sk_load_tables_rgb_u16_be_hsw+0xd3> + DB 15,133,193,0,0,0 ; jne 1529 <_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 @@ -1328,7 +1330,7 @@ _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 197,121,111,13,62,46,0,0 ; vmovdqa 0x2e3e(%rip),%xmm9 # 42d0 <_sk_callback_hsw+0x60d> + DB 197,121,111,13,94,46,0,0 ; vmovdqa 0x2e5e(%rip),%xmm9 # 4330 <_sk_callback_hsw+0x62d> 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 @@ -1345,41 +1347,41 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC 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 196,226,125,24,29,197,41,0,0 ; vbroadcastss 0x29c5(%rip),%ymm3 # 3eac <_sk_callback_hsw+0x1e9> + DB 196,226,125,24,29,225,41,0,0 ; vbroadcastss 0x29e1(%rip),%ymm3 # 3f08 <_sk_callback_hsw+0x205> 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 1502 <_sk_load_tables_rgb_u16_be_hsw+0xec> - DB 233,90,255,255,255 ; jmpq 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 117,5 ; jne 1542 <_sk_load_tables_rgb_u16_be_hsw+0xec> + DB 233,90,255,255,255 ; jmpq 149c <_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 1531 <_sk_load_tables_rgb_u16_be_hsw+0x11b> + DB 114,26 ; jb 1571 <_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 1536 <_sk_load_tables_rgb_u16_be_hsw+0x120> - DB 233,43,255,255,255 ; jmpq 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 233,38,255,255,255 ; jmpq 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 1576 <_sk_load_tables_rgb_u16_be_hsw+0x120> + DB 233,43,255,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,38,255,255,255 ; jmpq 149c <_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 1565 <_sk_load_tables_rgb_u16_be_hsw+0x14f> + DB 114,26 ; jb 15a5 <_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 156a <_sk_load_tables_rgb_u16_be_hsw+0x154> - DB 233,247,254,255,255 ; jmpq 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 233,242,254,255,255 ; jmpq 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 15aa <_sk_load_tables_rgb_u16_be_hsw+0x154> + DB 233,247,254,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,242,254,255,255 ; jmpq 149c <_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 1593 <_sk_load_tables_rgb_u16_be_hsw+0x17d> + DB 114,20 ; jb 15d3 <_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 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> - DB 233,196,254,255,255 ; jmpq 145c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,201,254,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46> + DB 233,196,254,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46> PUBLIC _sk_byte_tables_hsw _sk_byte_tables_hsw LABEL PROC @@ -1390,7 +1392,7 @@ _sk_byte_tables_hsw LABEL PROC DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,3,41,0,0 ; vbroadcastss 0x2903(%rip),%ymm8 # 3eb0 <_sk_callback_hsw+0x1ed> + DB 196,98,125,24,5,31,41,0,0 ; vbroadcastss 0x291f(%rip),%ymm8 # 3f0c <_sk_callback_hsw+0x209> 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 @@ -1427,7 +1429,7 @@ _sk_byte_tables_hsw LABEL PROC 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 196,98,125,24,13,84,40,0,0 ; vbroadcastss 0x2854(%rip),%ymm9 # 3eb4 <_sk_callback_hsw+0x1f1> + DB 196,98,125,24,13,112,40,0,0 ; vbroadcastss 0x2870(%rip),%ymm9 # 3f10 <_sk_callback_hsw+0x20d> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 @@ -1586,7 +1588,7 @@ _sk_byte_tables_rgb_hsw LABEL PROC 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 196,98,125,24,13,141,37,0,0 ; vbroadcastss 0x258d(%rip),%ymm9 # 3eb8 <_sk_callback_hsw+0x1f5> + DB 196,98,125,24,13,169,37,0,0 ; vbroadcastss 0x25a9(%rip),%ymm9 # 3f14 <_sk_callback_hsw+0x211> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 @@ -1739,33 +1741,33 @@ _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 196,98,125,24,37,64,35,0,0 ; vbroadcastss 0x2340(%rip),%ymm12 # 3ebc <_sk_callback_hsw+0x1f9> - DB 196,98,125,24,45,59,35,0,0 ; vbroadcastss 0x233b(%rip),%ymm13 # 3ec0 <_sk_callback_hsw+0x1fd> + DB 196,98,125,24,37,92,35,0,0 ; vbroadcastss 0x235c(%rip),%ymm12 # 3f18 <_sk_callback_hsw+0x215> + DB 196,98,125,24,45,87,35,0,0 ; vbroadcastss 0x2357(%rip),%ymm13 # 3f1c <_sk_callback_hsw+0x219> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,49,35,0,0 ; vbroadcastss 0x2331(%rip),%ymm13 # 3ec4 <_sk_callback_hsw+0x201> + DB 196,98,125,24,45,77,35,0,0 ; vbroadcastss 0x234d(%rip),%ymm13 # 3f20 <_sk_callback_hsw+0x21d> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,39,35,0,0 ; vbroadcastss 0x2327(%rip),%ymm13 # 3ec8 <_sk_callback_hsw+0x205> + DB 196,98,125,24,45,67,35,0,0 ; vbroadcastss 0x2343(%rip),%ymm13 # 3f24 <_sk_callback_hsw+0x221> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,29,35,0,0 ; vbroadcastss 0x231d(%rip),%ymm11 # 3ecc <_sk_callback_hsw+0x209> + DB 196,98,125,24,29,57,35,0,0 ; vbroadcastss 0x2339(%rip),%ymm11 # 3f28 <_sk_callback_hsw+0x225> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,19,35,0,0 ; vbroadcastss 0x2313(%rip),%ymm12 # 3ed0 <_sk_callback_hsw+0x20d> + DB 196,98,125,24,37,47,35,0,0 ; vbroadcastss 0x232f(%rip),%ymm12 # 3f2c <_sk_callback_hsw+0x229> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,9,35,0,0 ; vbroadcastss 0x2309(%rip),%ymm12 # 3ed4 <_sk_callback_hsw+0x211> + DB 196,98,125,24,37,37,35,0,0 ; vbroadcastss 0x2325(%rip),%ymm12 # 3f30 <_sk_callback_hsw+0x22d> 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 196,98,125,24,29,234,34,0,0 ; vbroadcastss 0x22ea(%rip),%ymm11 # 3ed8 <_sk_callback_hsw+0x215> + DB 196,98,125,24,29,6,35,0,0 ; vbroadcastss 0x2306(%rip),%ymm11 # 3f34 <_sk_callback_hsw+0x231> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,224,34,0,0 ; vbroadcastss 0x22e0(%rip),%ymm11 # 3edc <_sk_callback_hsw+0x219> + DB 196,98,125,24,29,252,34,0,0 ; vbroadcastss 0x22fc(%rip),%ymm11 # 3f38 <_sk_callback_hsw+0x235> DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11 - DB 196,226,125,24,5,214,34,0,0 ; vbroadcastss 0x22d6(%rip),%ymm0 # 3ee0 <_sk_callback_hsw+0x21d> + DB 196,226,125,24,5,242,34,0,0 ; vbroadcastss 0x22f2(%rip),%ymm0 # 3f3c <_sk_callback_hsw+0x239> DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,204,34,0,0 ; vbroadcastss 0x22cc(%rip),%ymm10 # 3ee4 <_sk_callback_hsw+0x221> + DB 196,98,125,24,21,232,34,0,0 ; vbroadcastss 0x22e8(%rip),%ymm10 # 3f40 <_sk_callback_hsw+0x23d> DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0 DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,21,191,34,0,0 ; vbroadcastss 0x22bf(%rip),%ymm10 # 3ee8 <_sk_callback_hsw+0x225> + DB 196,98,125,24,21,219,34,0,0 ; vbroadcastss 0x22db(%rip),%ymm10 # 3f44 <_sk_callback_hsw+0x241> 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 @@ -1773,7 +1775,7 @@ _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 196,98,125,24,5,150,34,0,0 ; vbroadcastss 0x2296(%rip),%ymm8 # 3eec <_sk_callback_hsw+0x229> + DB 196,98,125,24,5,178,34,0,0 ; vbroadcastss 0x22b2(%rip),%ymm8 # 3f48 <_sk_callback_hsw+0x245> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1791,33 +1793,33 @@ _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 196,98,125,24,37,78,34,0,0 ; vbroadcastss 0x224e(%rip),%ymm12 # 3ef0 <_sk_callback_hsw+0x22d> - DB 196,98,125,24,45,73,34,0,0 ; vbroadcastss 0x2249(%rip),%ymm13 # 3ef4 <_sk_callback_hsw+0x231> + DB 196,98,125,24,37,106,34,0,0 ; vbroadcastss 0x226a(%rip),%ymm12 # 3f4c <_sk_callback_hsw+0x249> + DB 196,98,125,24,45,101,34,0,0 ; vbroadcastss 0x2265(%rip),%ymm13 # 3f50 <_sk_callback_hsw+0x24d> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,63,34,0,0 ; vbroadcastss 0x223f(%rip),%ymm13 # 3ef8 <_sk_callback_hsw+0x235> + DB 196,98,125,24,45,91,34,0,0 ; vbroadcastss 0x225b(%rip),%ymm13 # 3f54 <_sk_callback_hsw+0x251> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,53,34,0,0 ; vbroadcastss 0x2235(%rip),%ymm13 # 3efc <_sk_callback_hsw+0x239> + DB 196,98,125,24,45,81,34,0,0 ; vbroadcastss 0x2251(%rip),%ymm13 # 3f58 <_sk_callback_hsw+0x255> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,43,34,0,0 ; vbroadcastss 0x222b(%rip),%ymm11 # 3f00 <_sk_callback_hsw+0x23d> + DB 196,98,125,24,29,71,34,0,0 ; vbroadcastss 0x2247(%rip),%ymm11 # 3f5c <_sk_callback_hsw+0x259> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,33,34,0,0 ; vbroadcastss 0x2221(%rip),%ymm12 # 3f04 <_sk_callback_hsw+0x241> + DB 196,98,125,24,37,61,34,0,0 ; vbroadcastss 0x223d(%rip),%ymm12 # 3f60 <_sk_callback_hsw+0x25d> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,23,34,0,0 ; vbroadcastss 0x2217(%rip),%ymm12 # 3f08 <_sk_callback_hsw+0x245> + DB 196,98,125,24,37,51,34,0,0 ; vbroadcastss 0x2233(%rip),%ymm12 # 3f64 <_sk_callback_hsw+0x261> 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 196,98,125,24,29,248,33,0,0 ; vbroadcastss 0x21f8(%rip),%ymm11 # 3f0c <_sk_callback_hsw+0x249> + DB 196,98,125,24,29,20,34,0,0 ; vbroadcastss 0x2214(%rip),%ymm11 # 3f68 <_sk_callback_hsw+0x265> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,238,33,0,0 ; vbroadcastss 0x21ee(%rip),%ymm11 # 3f10 <_sk_callback_hsw+0x24d> + DB 196,98,125,24,29,10,34,0,0 ; vbroadcastss 0x220a(%rip),%ymm11 # 3f6c <_sk_callback_hsw+0x269> DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11 - DB 196,226,125,24,13,228,33,0,0 ; vbroadcastss 0x21e4(%rip),%ymm1 # 3f14 <_sk_callback_hsw+0x251> + DB 196,226,125,24,13,0,34,0,0 ; vbroadcastss 0x2200(%rip),%ymm1 # 3f70 <_sk_callback_hsw+0x26d> DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,218,33,0,0 ; vbroadcastss 0x21da(%rip),%ymm10 # 3f18 <_sk_callback_hsw+0x255> + DB 196,98,125,24,21,246,33,0,0 ; vbroadcastss 0x21f6(%rip),%ymm10 # 3f74 <_sk_callback_hsw+0x271> DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 196,98,125,24,21,205,33,0,0 ; vbroadcastss 0x21cd(%rip),%ymm10 # 3f1c <_sk_callback_hsw+0x259> + DB 196,98,125,24,21,233,33,0,0 ; vbroadcastss 0x21e9(%rip),%ymm10 # 3f78 <_sk_callback_hsw+0x275> 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 @@ -1825,7 +1827,7 @@ _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 196,98,125,24,5,164,33,0,0 ; vbroadcastss 0x21a4(%rip),%ymm8 # 3f20 <_sk_callback_hsw+0x25d> + DB 196,98,125,24,5,192,33,0,0 ; vbroadcastss 0x21c0(%rip),%ymm8 # 3f7c <_sk_callback_hsw+0x279> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1843,33 +1845,33 @@ _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 196,98,125,24,37,92,33,0,0 ; vbroadcastss 0x215c(%rip),%ymm12 # 3f24 <_sk_callback_hsw+0x261> - DB 196,98,125,24,45,87,33,0,0 ; vbroadcastss 0x2157(%rip),%ymm13 # 3f28 <_sk_callback_hsw+0x265> + DB 196,98,125,24,37,120,33,0,0 ; vbroadcastss 0x2178(%rip),%ymm12 # 3f80 <_sk_callback_hsw+0x27d> + DB 196,98,125,24,45,115,33,0,0 ; vbroadcastss 0x2173(%rip),%ymm13 # 3f84 <_sk_callback_hsw+0x281> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,77,33,0,0 ; vbroadcastss 0x214d(%rip),%ymm13 # 3f2c <_sk_callback_hsw+0x269> + DB 196,98,125,24,45,105,33,0,0 ; vbroadcastss 0x2169(%rip),%ymm13 # 3f88 <_sk_callback_hsw+0x285> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,67,33,0,0 ; vbroadcastss 0x2143(%rip),%ymm13 # 3f30 <_sk_callback_hsw+0x26d> + DB 196,98,125,24,45,95,33,0,0 ; vbroadcastss 0x215f(%rip),%ymm13 # 3f8c <_sk_callback_hsw+0x289> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,57,33,0,0 ; vbroadcastss 0x2139(%rip),%ymm11 # 3f34 <_sk_callback_hsw+0x271> + DB 196,98,125,24,29,85,33,0,0 ; vbroadcastss 0x2155(%rip),%ymm11 # 3f90 <_sk_callback_hsw+0x28d> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,47,33,0,0 ; vbroadcastss 0x212f(%rip),%ymm12 # 3f38 <_sk_callback_hsw+0x275> + DB 196,98,125,24,37,75,33,0,0 ; vbroadcastss 0x214b(%rip),%ymm12 # 3f94 <_sk_callback_hsw+0x291> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,37,33,0,0 ; vbroadcastss 0x2125(%rip),%ymm12 # 3f3c <_sk_callback_hsw+0x279> + DB 196,98,125,24,37,65,33,0,0 ; vbroadcastss 0x2141(%rip),%ymm12 # 3f98 <_sk_callback_hsw+0x295> 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 196,98,125,24,29,6,33,0,0 ; vbroadcastss 0x2106(%rip),%ymm11 # 3f40 <_sk_callback_hsw+0x27d> + DB 196,98,125,24,29,34,33,0,0 ; vbroadcastss 0x2122(%rip),%ymm11 # 3f9c <_sk_callback_hsw+0x299> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,252,32,0,0 ; vbroadcastss 0x20fc(%rip),%ymm11 # 3f44 <_sk_callback_hsw+0x281> + DB 196,98,125,24,29,24,33,0,0 ; vbroadcastss 0x2118(%rip),%ymm11 # 3fa0 <_sk_callback_hsw+0x29d> DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11 - DB 196,226,125,24,21,242,32,0,0 ; vbroadcastss 0x20f2(%rip),%ymm2 # 3f48 <_sk_callback_hsw+0x285> + DB 196,226,125,24,21,14,33,0,0 ; vbroadcastss 0x210e(%rip),%ymm2 # 3fa4 <_sk_callback_hsw+0x2a1> DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,232,32,0,0 ; vbroadcastss 0x20e8(%rip),%ymm10 # 3f4c <_sk_callback_hsw+0x289> + DB 196,98,125,24,21,4,33,0,0 ; vbroadcastss 0x2104(%rip),%ymm10 # 3fa8 <_sk_callback_hsw+0x2a5> DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2 DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 196,98,125,24,21,219,32,0,0 ; vbroadcastss 0x20db(%rip),%ymm10 # 3f50 <_sk_callback_hsw+0x28d> + DB 196,98,125,24,21,247,32,0,0 ; vbroadcastss 0x20f7(%rip),%ymm10 # 3fac <_sk_callback_hsw+0x2a9> 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 @@ -1877,7 +1879,7 @@ _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 196,98,125,24,5,178,32,0,0 ; vbroadcastss 0x20b2(%rip),%ymm8 # 3f54 <_sk_callback_hsw+0x291> + DB 196,98,125,24,5,206,32,0,0 ; vbroadcastss 0x20ce(%rip),%ymm8 # 3fb0 <_sk_callback_hsw+0x2ad> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1895,33 +1897,33 @@ _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 196,98,125,24,37,106,32,0,0 ; vbroadcastss 0x206a(%rip),%ymm12 # 3f58 <_sk_callback_hsw+0x295> - DB 196,98,125,24,45,101,32,0,0 ; vbroadcastss 0x2065(%rip),%ymm13 # 3f5c <_sk_callback_hsw+0x299> + DB 196,98,125,24,37,134,32,0,0 ; vbroadcastss 0x2086(%rip),%ymm12 # 3fb4 <_sk_callback_hsw+0x2b1> + DB 196,98,125,24,45,129,32,0,0 ; vbroadcastss 0x2081(%rip),%ymm13 # 3fb8 <_sk_callback_hsw+0x2b5> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,91,32,0,0 ; vbroadcastss 0x205b(%rip),%ymm13 # 3f60 <_sk_callback_hsw+0x29d> + DB 196,98,125,24,45,119,32,0,0 ; vbroadcastss 0x2077(%rip),%ymm13 # 3fbc <_sk_callback_hsw+0x2b9> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,81,32,0,0 ; vbroadcastss 0x2051(%rip),%ymm13 # 3f64 <_sk_callback_hsw+0x2a1> + DB 196,98,125,24,45,109,32,0,0 ; vbroadcastss 0x206d(%rip),%ymm13 # 3fc0 <_sk_callback_hsw+0x2bd> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,71,32,0,0 ; vbroadcastss 0x2047(%rip),%ymm11 # 3f68 <_sk_callback_hsw+0x2a5> + DB 196,98,125,24,29,99,32,0,0 ; vbroadcastss 0x2063(%rip),%ymm11 # 3fc4 <_sk_callback_hsw+0x2c1> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,61,32,0,0 ; vbroadcastss 0x203d(%rip),%ymm12 # 3f6c <_sk_callback_hsw+0x2a9> + DB 196,98,125,24,37,89,32,0,0 ; vbroadcastss 0x2059(%rip),%ymm12 # 3fc8 <_sk_callback_hsw+0x2c5> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,51,32,0,0 ; vbroadcastss 0x2033(%rip),%ymm12 # 3f70 <_sk_callback_hsw+0x2ad> + DB 196,98,125,24,37,79,32,0,0 ; vbroadcastss 0x204f(%rip),%ymm12 # 3fcc <_sk_callback_hsw+0x2c9> 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 196,98,125,24,29,20,32,0,0 ; vbroadcastss 0x2014(%rip),%ymm11 # 3f74 <_sk_callback_hsw+0x2b1> + DB 196,98,125,24,29,48,32,0,0 ; vbroadcastss 0x2030(%rip),%ymm11 # 3fd0 <_sk_callback_hsw+0x2cd> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,10,32,0,0 ; vbroadcastss 0x200a(%rip),%ymm11 # 3f78 <_sk_callback_hsw+0x2b5> + DB 196,98,125,24,29,38,32,0,0 ; vbroadcastss 0x2026(%rip),%ymm11 # 3fd4 <_sk_callback_hsw+0x2d1> DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11 - DB 196,226,125,24,29,0,32,0,0 ; vbroadcastss 0x2000(%rip),%ymm3 # 3f7c <_sk_callback_hsw+0x2b9> + DB 196,226,125,24,29,28,32,0,0 ; vbroadcastss 0x201c(%rip),%ymm3 # 3fd8 <_sk_callback_hsw+0x2d5> DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,246,31,0,0 ; vbroadcastss 0x1ff6(%rip),%ymm10 # 3f80 <_sk_callback_hsw+0x2bd> + DB 196,98,125,24,21,18,32,0,0 ; vbroadcastss 0x2012(%rip),%ymm10 # 3fdc <_sk_callback_hsw+0x2d9> DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3 DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 196,98,125,24,21,233,31,0,0 ; vbroadcastss 0x1fe9(%rip),%ymm10 # 3f84 <_sk_callback_hsw+0x2c1> + DB 196,98,125,24,21,5,32,0,0 ; vbroadcastss 0x2005(%rip),%ymm10 # 3fe0 <_sk_callback_hsw+0x2dd> 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 @@ -1929,33 +1931,33 @@ _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 196,98,125,24,5,192,31,0,0 ; vbroadcastss 0x1fc0(%rip),%ymm8 # 3f88 <_sk_callback_hsw+0x2c5> + DB 196,98,125,24,5,220,31,0,0 ; vbroadcastss 0x1fdc(%rip),%ymm8 # 3fe4 <_sk_callback_hsw+0x2e1> 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 196,98,125,24,5,178,31,0,0 ; vbroadcastss 0x1fb2(%rip),%ymm8 # 3f8c <_sk_callback_hsw+0x2c9> - DB 196,98,125,24,13,173,31,0,0 ; vbroadcastss 0x1fad(%rip),%ymm9 # 3f90 <_sk_callback_hsw+0x2cd> - DB 196,98,125,24,21,168,31,0,0 ; vbroadcastss 0x1fa8(%rip),%ymm10 # 3f94 <_sk_callback_hsw+0x2d1> + DB 196,98,125,24,5,206,31,0,0 ; vbroadcastss 0x1fce(%rip),%ymm8 # 3fe8 <_sk_callback_hsw+0x2e5> + DB 196,98,125,24,13,201,31,0,0 ; vbroadcastss 0x1fc9(%rip),%ymm9 # 3fec <_sk_callback_hsw+0x2e9> + DB 196,98,125,24,21,196,31,0,0 ; vbroadcastss 0x1fc4(%rip),%ymm10 # 3ff0 <_sk_callback_hsw+0x2ed> DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1 DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2 - DB 196,98,125,24,13,153,31,0,0 ; vbroadcastss 0x1f99(%rip),%ymm9 # 3f98 <_sk_callback_hsw+0x2d5> + DB 196,98,125,24,13,181,31,0,0 ; vbroadcastss 0x1fb5(%rip),%ymm9 # 3ff4 <_sk_callback_hsw+0x2f1> DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9 - DB 196,226,125,24,5,143,31,0,0 ; vbroadcastss 0x1f8f(%rip),%ymm0 # 3f9c <_sk_callback_hsw+0x2d9> + DB 196,226,125,24,5,171,31,0,0 ; vbroadcastss 0x1fab(%rip),%ymm0 # 3ff8 <_sk_callback_hsw+0x2f5> DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 - DB 196,98,125,24,5,134,31,0,0 ; vbroadcastss 0x1f86(%rip),%ymm8 # 3fa0 <_sk_callback_hsw+0x2dd> + DB 196,98,125,24,5,162,31,0,0 ; vbroadcastss 0x1fa2(%rip),%ymm8 # 3ffc <_sk_callback_hsw+0x2f9> DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8 - DB 196,98,125,24,13,124,31,0,0 ; vbroadcastss 0x1f7c(%rip),%ymm9 # 3fa4 <_sk_callback_hsw+0x2e1> + DB 196,98,125,24,13,152,31,0,0 ; vbroadcastss 0x1f98(%rip),%ymm9 # 4000 <_sk_callback_hsw+0x2fd> 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 196,226,125,24,21,105,31,0,0 ; vbroadcastss 0x1f69(%rip),%ymm2 # 3fa8 <_sk_callback_hsw+0x2e5> + DB 196,226,125,24,21,133,31,0,0 ; vbroadcastss 0x1f85(%rip),%ymm2 # 4004 <_sk_callback_hsw+0x301> DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10 - DB 196,98,125,24,29,95,31,0,0 ; vbroadcastss 0x1f5f(%rip),%ymm11 # 3fac <_sk_callback_hsw+0x2e9> + DB 196,98,125,24,29,123,31,0,0 ; vbroadcastss 0x1f7b(%rip),%ymm11 # 4008 <_sk_callback_hsw+0x305> DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,37,85,31,0,0 ; vbroadcastss 0x1f55(%rip),%ymm12 # 3fb0 <_sk_callback_hsw+0x2ed> + DB 196,98,125,24,37,113,31,0,0 ; vbroadcastss 0x1f71(%rip),%ymm12 # 400c <_sk_callback_hsw+0x309> 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 @@ -1970,9 +1972,9 @@ _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 196,226,125,24,5,10,31,0,0 ; vbroadcastss 0x1f0a(%rip),%ymm0 # 3fb4 <_sk_callback_hsw+0x2f1> + DB 196,226,125,24,5,38,31,0,0 ; vbroadcastss 0x1f26(%rip),%ymm0 # 4010 <_sk_callback_hsw+0x30d> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,1,31,0,0 ; vbroadcastss 0x1f01(%rip),%ymm8 # 3fb8 <_sk_callback_hsw+0x2f5> + DB 196,98,125,24,5,29,31,0,0 ; vbroadcastss 0x1f1d(%rip),%ymm8 # 4014 <_sk_callback_hsw+0x311> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1984,11 +1986,11 @@ _sk_load_a8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,45 ; jne 20fd <_sk_load_a8_hsw+0x3d> + DB 117,45 ; jne 213d <_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 196,226,125,24,13,214,30,0,0 ; vbroadcastss 0x1ed6(%rip),%ymm1 # 3fbc <_sk_callback_hsw+0x2f9> + DB 196,226,125,24,13,242,30,0,0 ; vbroadcastss 0x1ef2(%rip),%ymm1 # 4018 <_sk_callback_hsw+0x315> 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 @@ -2005,9 +2007,9 @@ _sk_load_a8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2105 <_sk_load_a8_hsw+0x45> + DB 117,234 ; jne 2145 <_sk_load_a8_hsw+0x45> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,178 ; jmp 20d4 <_sk_load_a8_hsw+0x14> + DB 235,178 ; jmp 2114 <_sk_load_a8_hsw+0x14> PUBLIC _sk_gather_a8_hsw _sk_gather_a8_hsw LABEL PROC @@ -2051,7 +2053,7 @@ _sk_gather_a8_hsw LABEL PROC 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 196,226,125,24,13,225,29,0,0 ; vbroadcastss 0x1de1(%rip),%ymm1 # 3fc0 <_sk_callback_hsw+0x2fd> + DB 196,226,125,24,13,253,29,0,0 ; vbroadcastss 0x1dfd(%rip),%ymm1 # 401c <_sk_callback_hsw+0x319> 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 @@ -2067,14 +2069,14 @@ 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 196,98,125,24,5,188,29,0,0 ; vbroadcastss 0x1dbc(%rip),%ymm8 # 3fc4 <_sk_callback_hsw+0x301> + DB 196,98,125,24,5,216,29,0,0 ; vbroadcastss 0x1dd8(%rip),%ymm8 # 4020 <_sk_callback_hsw+0x31d> 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 2231 <_sk_store_a8_hsw+0x37> + DB 117,10 ; jne 2271 <_sk_store_a8_hsw+0x37> DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2082,10 +2084,10 @@ _sk_store_a8_hsw LABEL PROC 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 222d <_sk_store_a8_hsw+0x33> + DB 119,236 ; ja 226d <_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 # 2294 <_sk_store_a8_hsw+0x9a> + DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 22d4 <_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 @@ -2096,7 +2098,7 @@ _sk_store_a8_hsw LABEL PROC 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 222d <_sk_store_a8_hsw+0x33> + DB 235,154 ; jmp 226d <_sk_store_a8_hsw+0x33> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -2128,14 +2130,14 @@ _sk_load_g8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,50 ; jne 22f2 <_sk_load_g8_hsw+0x42> + DB 117,50 ; jne 2332 <_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 196,226,125,24,13,242,28,0,0 ; vbroadcastss 0x1cf2(%rip),%ymm1 # 3fc8 <_sk_callback_hsw+0x305> + DB 196,226,125,24,13,14,29,0,0 ; vbroadcastss 0x1d0e(%rip),%ymm1 # 4024 <_sk_callback_hsw+0x321> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,231,28,0,0 ; vbroadcastss 0x1ce7(%rip),%ymm3 # 3fcc <_sk_callback_hsw+0x309> + DB 196,226,125,24,29,3,29,0,0 ; vbroadcastss 0x1d03(%rip),%ymm3 # 4028 <_sk_callback_hsw+0x325> DB 76,137,193 ; mov %r8,%rcx DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 @@ -2149,9 +2151,9 @@ _sk_load_g8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 22fa <_sk_load_g8_hsw+0x4a> + DB 117,234 ; jne 233a <_sk_load_g8_hsw+0x4a> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,173 ; jmp 22c4 <_sk_load_g8_hsw+0x14> + DB 235,173 ; jmp 2304 <_sk_load_g8_hsw+0x14> PUBLIC _sk_gather_g8_hsw _sk_gather_g8_hsw LABEL PROC @@ -2195,10 +2197,10 @@ _sk_gather_g8_hsw LABEL PROC 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 196,226,125,24,13,252,27,0,0 ; vbroadcastss 0x1bfc(%rip),%ymm1 # 3fd0 <_sk_callback_hsw+0x30d> + DB 196,226,125,24,13,24,28,0,0 ; vbroadcastss 0x1c18(%rip),%ymm1 # 402c <_sk_callback_hsw+0x329> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,241,27,0,0 ; vbroadcastss 0x1bf1(%rip),%ymm3 # 3fd4 <_sk_callback_hsw+0x311> + DB 196,226,125,24,29,13,28,0,0 ; vbroadcastss 0x1c0d(%rip),%ymm3 # 4030 <_sk_callback_hsw+0x32d> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx @@ -2212,9 +2214,9 @@ _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 2403 <_sk_gather_i8_hsw+0xf> + DB 116,5 ; je 2443 <_sk_gather_i8_hsw+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 2405 <_sk_gather_i8_hsw+0x11> + DB 235,2 ; jmp 2445 <_sk_gather_i8_hsw+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -2252,14 +2254,14 @@ _sk_gather_i8_hsw LABEL PROC DB 73,139,64,8 ; mov 0x8(%r8),%rax DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,226,117,144,28,128 ; vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 - DB 197,229,219,5,237,28,0,0 ; vpand 0x1ced(%rip),%ymm3,%ymm0 # 41a0 <_sk_callback_hsw+0x4dd> + DB 197,229,219,5,13,29,0,0 ; vpand 0x1d0d(%rip),%ymm3,%ymm0 # 4200 <_sk_callback_hsw+0x4fd> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,24,27,0,0 ; vbroadcastss 0x1b18(%rip),%ymm8 # 3fd8 <_sk_callback_hsw+0x315> + DB 196,98,125,24,5,52,27,0,0 ; vbroadcastss 0x1b34(%rip),%ymm8 # 4034 <_sk_callback_hsw+0x331> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,242,28,0,0 ; vpshufb 0x1cf2(%rip),%ymm3,%ymm1 # 41c0 <_sk_callback_hsw+0x4fd> + DB 196,226,101,0,13,18,29,0,0 ; vpshufb 0x1d12(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x51d> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,0,29,0,0 ; vpshufb 0x1d00(%rip),%ymm3,%ymm2 # 41e0 <_sk_callback_hsw+0x51d> + DB 196,226,101,0,21,32,29,0,0 ; vpshufb 0x1d20(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x53d> 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 @@ -2278,35 +2280,35 @@ _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 2580 <_sk_load_565_hsw+0x7c> + DB 117,114 ; jne 25c0 <_sk_load_565_hsw+0x7c> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,186,26,0,0 ; vpbroadcastd 0x1aba(%rip),%ymm0 # 3fdc <_sk_callback_hsw+0x319> + DB 196,226,125,88,5,214,26,0,0 ; vpbroadcastd 0x1ad6(%rip),%ymm0 # 4038 <_sk_callback_hsw+0x335> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,173,26,0,0 ; vbroadcastss 0x1aad(%rip),%ymm1 # 3fe0 <_sk_callback_hsw+0x31d> + DB 196,226,125,24,13,201,26,0,0 ; vbroadcastss 0x1ac9(%rip),%ymm1 # 403c <_sk_callback_hsw+0x339> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,164,26,0,0 ; vpbroadcastd 0x1aa4(%rip),%ymm1 # 3fe4 <_sk_callback_hsw+0x321> + DB 196,226,125,88,13,192,26,0,0 ; vpbroadcastd 0x1ac0(%rip),%ymm1 # 4040 <_sk_callback_hsw+0x33d> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,151,26,0,0 ; vbroadcastss 0x1a97(%rip),%ymm3 # 3fe8 <_sk_callback_hsw+0x325> + DB 196,226,125,24,29,179,26,0,0 ; vbroadcastss 0x1ab3(%rip),%ymm3 # 4044 <_sk_callback_hsw+0x341> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,142,26,0,0 ; vpbroadcastd 0x1a8e(%rip),%ymm3 # 3fec <_sk_callback_hsw+0x329> + DB 196,226,125,88,29,170,26,0,0 ; vpbroadcastd 0x1aaa(%rip),%ymm3 # 4048 <_sk_callback_hsw+0x345> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,129,26,0,0 ; vbroadcastss 0x1a81(%rip),%ymm3 # 3ff0 <_sk_callback_hsw+0x32d> + DB 196,226,125,24,29,157,26,0,0 ; vbroadcastss 0x1a9d(%rip),%ymm3 # 404c <_sk_callback_hsw+0x349> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,118,26,0,0 ; vbroadcastss 0x1a76(%rip),%ymm3 # 3ff4 <_sk_callback_hsw+0x331> + DB 196,226,125,24,29,146,26,0,0 ; vbroadcastss 0x1a92(%rip),%ymm3 # 4050 <_sk_callback_hsw+0x34d> DB 255,224 ; jmpq *%rax DB 65,137,200 ; mov %ecx,%r8d DB 65,128,224,7 ; and $0x7,%r8b 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 2514 <_sk_load_565_hsw+0x10> + DB 119,128 ; ja 2554 <_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 # 25e8 <_sk_load_565_hsw+0xe4> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2628 <_sk_load_565_hsw+0xe4> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2318,7 +2320,7 @@ _sk_load_565_hsw LABEL PROC 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 2514 <_sk_load_565_hsw+0x10> + DB 233,44,255,255,255 ; jmpq 2554 <_sk_load_565_hsw+0x10> DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) @@ -2386,23 +2388,23 @@ _sk_gather_565_hsw LABEL PROC DB 65,15,183,4,88 ; movzwl (%r8,%rbx,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,57,25,0,0 ; vpbroadcastd 0x1939(%rip),%ymm0 # 3ff8 <_sk_callback_hsw+0x335> + DB 196,226,125,88,5,85,25,0,0 ; vpbroadcastd 0x1955(%rip),%ymm0 # 4054 <_sk_callback_hsw+0x351> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,44,25,0,0 ; vbroadcastss 0x192c(%rip),%ymm1 # 3ffc <_sk_callback_hsw+0x339> + DB 196,226,125,24,13,72,25,0,0 ; vbroadcastss 0x1948(%rip),%ymm1 # 4058 <_sk_callback_hsw+0x355> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,35,25,0,0 ; vpbroadcastd 0x1923(%rip),%ymm1 # 4000 <_sk_callback_hsw+0x33d> + DB 196,226,125,88,13,63,25,0,0 ; vpbroadcastd 0x193f(%rip),%ymm1 # 405c <_sk_callback_hsw+0x359> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,22,25,0,0 ; vbroadcastss 0x1916(%rip),%ymm3 # 4004 <_sk_callback_hsw+0x341> + DB 196,226,125,24,29,50,25,0,0 ; vbroadcastss 0x1932(%rip),%ymm3 # 4060 <_sk_callback_hsw+0x35d> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,13,25,0,0 ; vpbroadcastd 0x190d(%rip),%ymm3 # 4008 <_sk_callback_hsw+0x345> + DB 196,226,125,88,29,41,25,0,0 ; vpbroadcastd 0x1929(%rip),%ymm3 # 4064 <_sk_callback_hsw+0x361> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,0,25,0,0 ; vbroadcastss 0x1900(%rip),%ymm3 # 400c <_sk_callback_hsw+0x349> + DB 196,226,125,24,29,28,25,0,0 ; vbroadcastss 0x191c(%rip),%ymm3 # 4068 <_sk_callback_hsw+0x365> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,245,24,0,0 ; vbroadcastss 0x18f5(%rip),%ymm3 # 4010 <_sk_callback_hsw+0x34d> + DB 196,226,125,24,29,17,25,0,0 ; vbroadcastss 0x1911(%rip),%ymm3 # 406c <_sk_callback_hsw+0x369> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 @@ -2413,11 +2415,11 @@ 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 196,98,125,24,5,226,24,0,0 ; vbroadcastss 0x18e2(%rip),%ymm8 # 4014 <_sk_callback_hsw+0x351> + DB 196,98,125,24,5,254,24,0,0 ; vbroadcastss 0x18fe(%rip),%ymm8 # 4070 <_sk_callback_hsw+0x36d> 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 196,98,125,24,21,205,24,0,0 ; vbroadcastss 0x18cd(%rip),%ymm10 # 4018 <_sk_callback_hsw+0x355> + DB 196,98,125,24,21,233,24,0,0 ; vbroadcastss 0x18e9(%rip),%ymm10 # 4074 <_sk_callback_hsw+0x371> 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 @@ -2428,7 +2430,7 @@ _sk_store_565_hsw LABEL PROC 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 2789 <_sk_store_565_hsw+0x65> + DB 117,10 ; jne 27c9 <_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 @@ -2436,9 +2438,9 @@ _sk_store_565_hsw LABEL PROC 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 2785 <_sk_store_565_hsw+0x61> + DB 119,236 ; ja 27c5 <_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 # 27e8 <_sk_store_565_hsw+0xc4> + DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2828 <_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 @@ -2449,7 +2451,7 @@ _sk_store_565_hsw LABEL PROC 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 2785 <_sk_store_565_hsw+0x61> + DB 235,159 ; jmp 27c5 <_sk_store_565_hsw+0x61> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -2480,28 +2482,28 @@ _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 289c <_sk_load_4444_hsw+0x98> + DB 15,133,138,0,0,0 ; jne 28dc <_sk_load_4444_hsw+0x98> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,246,23,0,0 ; vpbroadcastd 0x17f6(%rip),%ymm0 # 401c <_sk_callback_hsw+0x359> + DB 196,226,125,88,5,18,24,0,0 ; vpbroadcastd 0x1812(%rip),%ymm0 # 4078 <_sk_callback_hsw+0x375> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,233,23,0,0 ; vbroadcastss 0x17e9(%rip),%ymm1 # 4020 <_sk_callback_hsw+0x35d> + DB 196,226,125,24,13,5,24,0,0 ; vbroadcastss 0x1805(%rip),%ymm1 # 407c <_sk_callback_hsw+0x379> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,224,23,0,0 ; vpbroadcastd 0x17e0(%rip),%ymm1 # 4024 <_sk_callback_hsw+0x361> + DB 196,226,125,88,13,252,23,0,0 ; vpbroadcastd 0x17fc(%rip),%ymm1 # 4080 <_sk_callback_hsw+0x37d> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,211,23,0,0 ; vbroadcastss 0x17d3(%rip),%ymm2 # 4028 <_sk_callback_hsw+0x365> + DB 196,226,125,24,21,239,23,0,0 ; vbroadcastss 0x17ef(%rip),%ymm2 # 4084 <_sk_callback_hsw+0x381> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,202,23,0,0 ; vpbroadcastd 0x17ca(%rip),%ymm2 # 402c <_sk_callback_hsw+0x369> + DB 196,226,125,88,21,230,23,0,0 ; vpbroadcastd 0x17e6(%rip),%ymm2 # 4088 <_sk_callback_hsw+0x385> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,189,23,0,0 ; vbroadcastss 0x17bd(%rip),%ymm8 # 4030 <_sk_callback_hsw+0x36d> + DB 196,98,125,24,5,217,23,0,0 ; vbroadcastss 0x17d9(%rip),%ymm8 # 408c <_sk_callback_hsw+0x389> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,179,23,0,0 ; vpbroadcastd 0x17b3(%rip),%ymm8 # 4034 <_sk_callback_hsw+0x371> + DB 196,98,125,88,5,207,23,0,0 ; vpbroadcastd 0x17cf(%rip),%ymm8 # 4090 <_sk_callback_hsw+0x38d> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,165,23,0,0 ; vbroadcastss 0x17a5(%rip),%ymm8 # 4038 <_sk_callback_hsw+0x375> + DB 196,98,125,24,5,193,23,0,0 ; vbroadcastss 0x17c1(%rip),%ymm8 # 4094 <_sk_callback_hsw+0x391> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2510,9 +2512,9 @@ _sk_load_4444_hsw LABEL PROC 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 2818 <_sk_load_4444_hsw+0x14> + DB 15,135,100,255,255,255 ; ja 2858 <_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 # 2908 <_sk_load_4444_hsw+0x104> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2948 <_sk_load_4444_hsw+0x104> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2524,7 +2526,7 @@ _sk_load_4444_hsw LABEL PROC 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 2818 <_sk_load_4444_hsw+0x14> + DB 233,16,255,255,255 ; jmpq 2858 <_sk_load_4444_hsw+0x14> DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) @@ -2592,25 +2594,25 @@ _sk_gather_4444_hsw LABEL PROC DB 65,15,183,4,88 ; movzwl (%r8,%rbx,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,93,22,0,0 ; vpbroadcastd 0x165d(%rip),%ymm0 # 403c <_sk_callback_hsw+0x379> + DB 196,226,125,88,5,121,22,0,0 ; vpbroadcastd 0x1679(%rip),%ymm0 # 4098 <_sk_callback_hsw+0x395> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,80,22,0,0 ; vbroadcastss 0x1650(%rip),%ymm1 # 4040 <_sk_callback_hsw+0x37d> + DB 196,226,125,24,13,108,22,0,0 ; vbroadcastss 0x166c(%rip),%ymm1 # 409c <_sk_callback_hsw+0x399> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,71,22,0,0 ; vpbroadcastd 0x1647(%rip),%ymm1 # 4044 <_sk_callback_hsw+0x381> + DB 196,226,125,88,13,99,22,0,0 ; vpbroadcastd 0x1663(%rip),%ymm1 # 40a0 <_sk_callback_hsw+0x39d> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,58,22,0,0 ; vbroadcastss 0x163a(%rip),%ymm2 # 4048 <_sk_callback_hsw+0x385> + DB 196,226,125,24,21,86,22,0,0 ; vbroadcastss 0x1656(%rip),%ymm2 # 40a4 <_sk_callback_hsw+0x3a1> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,49,22,0,0 ; vpbroadcastd 0x1631(%rip),%ymm2 # 404c <_sk_callback_hsw+0x389> + DB 196,226,125,88,21,77,22,0,0 ; vpbroadcastd 0x164d(%rip),%ymm2 # 40a8 <_sk_callback_hsw+0x3a5> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,36,22,0,0 ; vbroadcastss 0x1624(%rip),%ymm8 # 4050 <_sk_callback_hsw+0x38d> + DB 196,98,125,24,5,64,22,0,0 ; vbroadcastss 0x1640(%rip),%ymm8 # 40ac <_sk_callback_hsw+0x3a9> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,26,22,0,0 ; vpbroadcastd 0x161a(%rip),%ymm8 # 4054 <_sk_callback_hsw+0x391> + DB 196,98,125,88,5,54,22,0,0 ; vpbroadcastd 0x1636(%rip),%ymm8 # 40b0 <_sk_callback_hsw+0x3ad> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,12,22,0,0 ; vbroadcastss 0x160c(%rip),%ymm8 # 4058 <_sk_callback_hsw+0x395> + DB 196,98,125,24,5,40,22,0,0 ; vbroadcastss 0x1628(%rip),%ymm8 # 40b4 <_sk_callback_hsw+0x3b1> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -2623,7 +2625,7 @@ 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 196,98,125,24,5,242,21,0,0 ; vbroadcastss 0x15f2(%rip),%ymm8 # 405c <_sk_callback_hsw+0x399> + DB 196,98,125,24,5,14,22,0,0 ; vbroadcastss 0x160e(%rip),%ymm8 # 40b8 <_sk_callback_hsw+0x3b5> 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 @@ -2641,7 +2643,7 @@ _sk_store_4444_hsw LABEL PROC 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 2acd <_sk_store_4444_hsw+0x71> + DB 117,10 ; jne 2b0d <_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 @@ -2649,9 +2651,9 @@ _sk_store_4444_hsw LABEL PROC 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 2ac9 <_sk_store_4444_hsw+0x6d> + DB 119,236 ; ja 2b09 <_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 # 2b2c <_sk_store_4444_hsw+0xd0> + DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2b6c <_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 @@ -2662,7 +2664,7 @@ _sk_store_4444_hsw LABEL PROC 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 2ac9 <_sk_store_4444_hsw+0x6d> + DB 235,159 ; jmp 2b09 <_sk_store_4444_hsw+0x6d> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -2695,16 +2697,16 @@ _sk_load_8888_hsw LABEL PROC DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 - DB 117,88 ; jne 2bb5 <_sk_load_8888_hsw+0x6d> + DB 117,88 ; jne 2bf5 <_sk_load_8888_hsw+0x6d> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 - DB 197,229,219,5,150,22,0,0 ; vpand 0x1696(%rip),%ymm3,%ymm0 # 4200 <_sk_callback_hsw+0x53d> + DB 197,229,219,5,182,22,0,0 ; vpand 0x16b6(%rip),%ymm3,%ymm0 # 4260 <_sk_callback_hsw+0x55d> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,233,20,0,0 ; vbroadcastss 0x14e9(%rip),%ymm8 # 4060 <_sk_callback_hsw+0x39d> + DB 196,98,125,24,5,5,21,0,0 ; vbroadcastss 0x1505(%rip),%ymm8 # 40bc <_sk_callback_hsw+0x3b9> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,155,22,0,0 ; vpshufb 0x169b(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x55d> + DB 196,226,101,0,13,187,22,0,0 ; vpshufb 0x16bb(%rip),%ymm3,%ymm1 # 4280 <_sk_callback_hsw+0x57d> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,169,22,0,0 ; vpshufb 0x16a9(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x57d> + DB 196,226,101,0,21,201,22,0,0 ; vpshufb 0x16c9(%rip),%ymm3,%ymm2 # 42a0 <_sk_callback_hsw+0x59d> 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 @@ -2721,7 +2723,7 @@ _sk_load_8888_hsw LABEL PROC DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3 - DB 235,135 ; jmp 2b62 <_sk_load_8888_hsw+0x1a> + DB 235,135 ; jmp 2ba2 <_sk_load_8888_hsw+0x1a> PUBLIC _sk_gather_8888_hsw _sk_gather_8888_hsw LABEL PROC @@ -2734,14 +2736,14 @@ _sk_gather_8888_hsw LABEL PROC 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 197,229,219,5,87,22,0,0 ; vpand 0x1657(%rip),%ymm3,%ymm0 # 4260 <_sk_callback_hsw+0x59d> + DB 197,229,219,5,119,22,0,0 ; vpand 0x1677(%rip),%ymm3,%ymm0 # 42c0 <_sk_callback_hsw+0x5bd> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,78,20,0,0 ; vbroadcastss 0x144e(%rip),%ymm8 # 4064 <_sk_callback_hsw+0x3a1> + DB 196,98,125,24,5,106,20,0,0 ; vbroadcastss 0x146a(%rip),%ymm8 # 40c0 <_sk_callback_hsw+0x3bd> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,92,22,0,0 ; vpshufb 0x165c(%rip),%ymm3,%ymm1 # 4280 <_sk_callback_hsw+0x5bd> + DB 196,226,101,0,13,124,22,0,0 ; vpshufb 0x167c(%rip),%ymm3,%ymm1 # 42e0 <_sk_callback_hsw+0x5dd> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,106,22,0,0 ; vpshufb 0x166a(%rip),%ymm3,%ymm2 # 42a0 <_sk_callback_hsw+0x5dd> + DB 196,226,101,0,21,138,22,0,0 ; vpshufb 0x168a(%rip),%ymm3,%ymm2 # 4300 <_sk_callback_hsw+0x5fd> 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 @@ -2756,7 +2758,7 @@ _sk_store_8888_hsw LABEL PROC 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 196,98,125,24,5,254,19,0,0 ; vbroadcastss 0x13fe(%rip),%ymm8 # 4068 <_sk_callback_hsw+0x3a5> + DB 196,98,125,24,5,26,20,0,0 ; vbroadcastss 0x141a(%rip),%ymm8 # 40c4 <_sk_callback_hsw+0x3c1> 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 @@ -2772,7 +2774,7 @@ _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 2cc4 <_sk_store_8888_hsw+0x73> + DB 117,12 ; jne 2d04 <_sk_store_8888_hsw+0x73> DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx @@ -2785,14 +2787,14 @@ _sk_store_8888_hsw LABEL PROC DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9 DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9) - DB 235,211 ; jmp 2cbd <_sk_store_8888_hsw+0x6c> + DB 235,211 ; jmp 2cfd <_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 2d55 <_sk_load_f16_hsw+0x6b> + DB 117,97 ; jne 2d95 <_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 @@ -2818,29 +2820,29 @@ _sk_load_f16_hsw LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 2db4 <_sk_load_f16_hsw+0xca> + DB 116,79 ; je 2df4 <_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 2db4 <_sk_load_f16_hsw+0xca> + DB 114,67 ; jb 2df4 <_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 2dc1 <_sk_load_f16_hsw+0xd7> + DB 116,68 ; je 2e01 <_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 2dc1 <_sk_load_f16_hsw+0xd7> + DB 114,56 ; jb 2e01 <_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 2d0b <_sk_load_f16_hsw+0x21> + DB 15,132,114,255,255,255 ; je 2d4b <_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 2d0b <_sk_load_f16_hsw+0x21> + DB 15,130,98,255,255,255 ; jb 2d4b <_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 2d0b <_sk_load_f16_hsw+0x21> + DB 233,87,255,255,255 ; jmpq 2d4b <_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 2d0b <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq 2d4b <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq 2d0b <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 2d4b <_sk_load_f16_hsw+0x21> PUBLIC _sk_gather_f16_hsw _sk_gather_f16_hsw LABEL PROC @@ -2894,7 +2896,7 @@ _sk_store_f16_hsw LABEL PROC 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 2eb9 <_sk_store_f16_hsw+0x65> + DB 117,27 ; jne 2ef9 <_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) @@ -2903,22 +2905,22 @@ _sk_store_f16_hsw LABEL PROC 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 2eb5 <_sk_store_f16_hsw+0x61> + DB 116,241 ; je 2ef5 <_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 2eb5 <_sk_store_f16_hsw+0x61> + DB 114,229 ; jb 2ef5 <_sk_store_f16_hsw+0x61> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je 2eb5 <_sk_store_f16_hsw+0x61> + DB 116,221 ; je 2ef5 <_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 2eb5 <_sk_store_f16_hsw+0x61> + DB 114,209 ; jb 2ef5 <_sk_store_f16_hsw+0x61> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je 2eb5 <_sk_store_f16_hsw+0x61> + DB 116,201 ; je 2ef5 <_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 2eb5 <_sk_store_f16_hsw+0x61> + DB 114,189 ; jb 2ef5 <_sk_store_f16_hsw+0x61> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp 2eb5 <_sk_store_f16_hsw+0x61> + DB 235,181 ; jmp 2ef5 <_sk_store_f16_hsw+0x61> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_hsw LABEL PROC @@ -2926,7 +2928,7 @@ _sk_load_u16_be_hsw LABEL PROC 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 2fe2 <_sk_load_u16_be_hsw+0xe2> + DB 15,133,204,0,0,0 ; jne 3022 <_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 @@ -2945,7 +2947,7 @@ _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 196,98,125,24,21,245,16,0,0 ; vbroadcastss 0x10f5(%rip),%ymm10 # 406c <_sk_callback_hsw+0x3a9> + DB 196,98,125,24,21,17,17,0,0 ; vbroadcastss 0x1111(%rip),%ymm10 # 40c8 <_sk_callback_hsw+0x3c5> 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 @@ -2973,29 +2975,29 @@ _sk_load_u16_be_hsw LABEL PROC DB 196,65,123,16,4,64 ; vmovsd (%r8,%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 3048 <_sk_load_u16_be_hsw+0x148> + DB 116,85 ; je 3088 <_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 3048 <_sk_load_u16_be_hsw+0x148> + DB 114,72 ; jb 3088 <_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 3055 <_sk_load_u16_be_hsw+0x155> + DB 116,72 ; je 3095 <_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 3055 <_sk_load_u16_be_hsw+0x155> + DB 114,59 ; jb 3095 <_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 2f31 <_sk_load_u16_be_hsw+0x31> + DB 15,132,6,255,255,255 ; je 2f71 <_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 2f31 <_sk_load_u16_be_hsw+0x31> + DB 15,130,245,254,255,255 ; jb 2f71 <_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 2f31 <_sk_load_u16_be_hsw+0x31> + DB 233,233,254,255,255 ; jmpq 2f71 <_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 2f31 <_sk_load_u16_be_hsw+0x31> + DB 233,220,254,255,255 ; jmpq 2f71 <_sk_load_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,211,254,255,255 ; jmpq 2f31 <_sk_load_u16_be_hsw+0x31> + DB 233,211,254,255,255 ; jmpq 2f71 <_sk_load_u16_be_hsw+0x31> PUBLIC _sk_load_rgb_u16_be_hsw _sk_load_rgb_u16_be_hsw LABEL PROC @@ -3003,7 +3005,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC 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 313c <_sk_load_rgb_u16_be_hsw+0xde> + DB 15,133,204,0,0,0 ; jne 317c <_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 @@ -3027,7 +3029,7 @@ _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 196,98,125,24,21,134,15,0,0 ; vbroadcastss 0xf86(%rip),%ymm10 # 4070 <_sk_callback_hsw+0x3ad> + DB 196,98,125,24,21,162,15,0,0 ; vbroadcastss 0xfa2(%rip),%ymm10 # 40cc <_sk_callback_hsw+0x3c9> 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 @@ -3044,48 +3046,48 @@ _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 196,226,125,24,29,58,15,0,0 ; vbroadcastss 0xf3a(%rip),%ymm3 # 4074 <_sk_callback_hsw+0x3b1> + DB 196,226,125,24,29,86,15,0,0 ; vbroadcastss 0xf56(%rip),%ymm3 # 40d0 <_sk_callback_hsw+0x3cd> 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 3155 <_sk_load_rgb_u16_be_hsw+0xf7> - DB 233,79,255,255,255 ; jmpq 30a4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,5 ; jne 3195 <_sk_load_rgb_u16_be_hsw+0xf7> + DB 233,79,255,255,255 ; jmpq 30e4 <_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 3184 <_sk_load_rgb_u16_be_hsw+0x126> + DB 114,26 ; jb 31c4 <_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 3189 <_sk_load_rgb_u16_be_hsw+0x12b> - DB 233,32,255,255,255 ; jmpq 30a4 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,27,255,255,255 ; jmpq 30a4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 31c9 <_sk_load_rgb_u16_be_hsw+0x12b> + DB 233,32,255,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,27,255,255,255 ; jmpq 30e4 <_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 31b8 <_sk_load_rgb_u16_be_hsw+0x15a> + DB 114,26 ; jb 31f8 <_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 31bd <_sk_load_rgb_u16_be_hsw+0x15f> - DB 233,236,254,255,255 ; jmpq 30a4 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,231,254,255,255 ; jmpq 30a4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 31fd <_sk_load_rgb_u16_be_hsw+0x15f> + DB 233,236,254,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,231,254,255,255 ; jmpq 30e4 <_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 31e6 <_sk_load_rgb_u16_be_hsw+0x188> + DB 114,20 ; jb 3226 <_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 30a4 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,185,254,255,255 ; jmpq 30a4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,190,254,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,185,254,255,255 ; jmpq 30e4 <_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 196,98,125,24,5,119,14,0,0 ; vbroadcastss 0xe77(%rip),%ymm8 # 4078 <_sk_callback_hsw+0x3b5> + DB 196,98,125,24,5,147,14,0,0 ; vbroadcastss 0xe93(%rip),%ymm8 # 40d4 <_sk_callback_hsw+0x3d1> 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 @@ -3123,7 +3125,7 @@ _sk_store_u16_be_hsw LABEL PROC 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 32e5 <_sk_store_u16_be_hsw+0xfa> + DB 117,31 ; jne 3325 <_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) @@ -3132,31 +3134,31 @@ _sk_store_u16_be_hsw LABEL PROC 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 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 116,240 ; je 3321 <_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 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 114,227 ; jb 3321 <_sk_store_u16_be_hsw+0xf6> DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 116,218 ; je 3321 <_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 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 114,205 ; jb 3321 <_sk_store_u16_be_hsw+0xf6> DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 116,196 ; je 3321 <_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 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 114,183 ; jb 3321 <_sk_store_u16_be_hsw+0xf6> DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp 32e1 <_sk_store_u16_be_hsw+0xf6> + DB 235,174 ; jmp 3321 <_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 33a9 <_sk_load_f32_hsw+0x76> + DB 119,110 ; ja 33e9 <_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 # 33d4 <_sk_load_f32_hsw+0xa1> + DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 3414 <_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 @@ -3215,7 +3217,7 @@ _sk_store_f32_hsw LABEL PROC 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 3461 <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 34a1 <_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 @@ -3228,22 +3230,22 @@ _sk_store_f32_hsw LABEL PROC 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 345d <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 349d <_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 345d <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 349d <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 345d <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 349d <_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 345d <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 349d <_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 345d <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 349d <_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 345d <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 349d <_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 345d <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 349d <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -3339,11 +3341,11 @@ _sk_mirror_y_hsw LABEL PROC PUBLIC _sk_luminance_to_alpha_hsw _sk_luminance_to_alpha_hsw LABEL PROC - DB 196,226,125,24,29,145,10,0,0 ; vbroadcastss 0xa91(%rip),%ymm3 # 407c <_sk_callback_hsw+0x3b9> - DB 196,98,125,24,5,140,10,0,0 ; vbroadcastss 0xa8c(%rip),%ymm8 # 4080 <_sk_callback_hsw+0x3bd> + DB 196,226,125,24,29,173,10,0,0 ; vbroadcastss 0xaad(%rip),%ymm3 # 40d8 <_sk_callback_hsw+0x3d5> + DB 196,98,125,24,5,168,10,0,0 ; vbroadcastss 0xaa8(%rip),%ymm8 # 40dc <_sk_callback_hsw+0x3d9> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1 - DB 196,226,125,24,29,125,10,0,0 ; vbroadcastss 0xa7d(%rip),%ymm3 # 4084 <_sk_callback_hsw+0x3c1> + DB 196,226,125,24,29,153,10,0,0 ; vbroadcastss 0xa99(%rip),%ymm3 # 40e0 <_sk_callback_hsw+0x3dd> 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 @@ -3478,7 +3480,7 @@ _sk_linear_gradient_hsw LABEL PROC DB 196,98,125,24,72,28 ; vbroadcastss 0x1c(%rax),%ymm9 DB 76,139,0 ; mov (%rax),%r8 DB 77,133,192 ; test %r8,%r8 - DB 15,132,143,0,0,0 ; je 38df <_sk_linear_gradient_hsw+0xb5> + DB 15,132,143,0,0,0 ; je 391f <_sk_linear_gradient_hsw+0xb5> DB 72,139,64,8 ; mov 0x8(%rax),%rax DB 72,131,192,32 ; add $0x20,%rax DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12 @@ -3505,8 +3507,8 @@ _sk_linear_gradient_hsw LABEL PROC DB 196,67,13,74,201,208 ; vblendvps %ymm13,%ymm9,%ymm14,%ymm9 DB 72,131,192,36 ; add $0x24,%rax DB 73,255,200 ; dec %r8 - DB 117,140 ; jne 3869 <_sk_linear_gradient_hsw+0x3f> - DB 235,17 ; jmp 38f0 <_sk_linear_gradient_hsw+0xc6> + DB 117,140 ; jne 38a9 <_sk_linear_gradient_hsw+0x3f> + DB 235,17 ; jmp 3930 <_sk_linear_gradient_hsw+0xc6> DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 @@ -3541,7 +3543,7 @@ _sk_linear_gradient_2stops_hsw LABEL PROC PUBLIC _sk_save_xy_hsw _sk_save_xy_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,36,7,0,0 ; vbroadcastss 0x724(%rip),%ymm8 # 4088 <_sk_callback_hsw+0x3c5> + DB 196,98,125,24,5,64,7,0,0 ; vbroadcastss 0x740(%rip),%ymm8 # 40e4 <_sk_callback_hsw+0x3e1> 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 @@ -3571,9 +3573,9 @@ _sk_accumulate_hsw LABEL PROC PUBLIC _sk_bilinear_nx_hsw _sk_bilinear_nx_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,184,6,0,0 ; vbroadcastss 0x6b8(%rip),%ymm0 # 408c <_sk_callback_hsw+0x3c9> + DB 196,226,125,24,5,212,6,0,0 ; vbroadcastss 0x6d4(%rip),%ymm0 # 40e8 <_sk_callback_hsw+0x3e5> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,175,6,0,0 ; vbroadcastss 0x6af(%rip),%ymm8 # 4090 <_sk_callback_hsw+0x3cd> + DB 196,98,125,24,5,203,6,0,0 ; vbroadcastss 0x6cb(%rip),%ymm8 # 40ec <_sk_callback_hsw+0x3e9> 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 @@ -3582,7 +3584,7 @@ _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 196,226,125,24,5,151,6,0,0 ; vbroadcastss 0x697(%rip),%ymm0 # 4094 <_sk_callback_hsw+0x3d1> + DB 196,226,125,24,5,179,6,0,0 ; vbroadcastss 0x6b3(%rip),%ymm0 # 40f0 <_sk_callback_hsw+0x3ed> 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) @@ -3592,9 +3594,9 @@ _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 196,226,125,24,13,123,6,0,0 ; vbroadcastss 0x67b(%rip),%ymm1 # 4098 <_sk_callback_hsw+0x3d5> + DB 196,226,125,24,13,151,6,0,0 ; vbroadcastss 0x697(%rip),%ymm1 # 40f4 <_sk_callback_hsw+0x3f1> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,113,6,0,0 ; vbroadcastss 0x671(%rip),%ymm8 # 409c <_sk_callback_hsw+0x3d9> + DB 196,98,125,24,5,141,6,0,0 ; vbroadcastss 0x68d(%rip),%ymm8 # 40f8 <_sk_callback_hsw+0x3f5> 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 @@ -3603,7 +3605,7 @@ _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 196,226,125,24,13,89,6,0,0 ; vbroadcastss 0x659(%rip),%ymm1 # 40a0 <_sk_callback_hsw+0x3dd> + DB 196,226,125,24,13,117,6,0,0 ; vbroadcastss 0x675(%rip),%ymm1 # 40fc <_sk_callback_hsw+0x3f9> 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) @@ -3613,13 +3615,13 @@ _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 196,226,125,24,5,60,6,0,0 ; vbroadcastss 0x63c(%rip),%ymm0 # 40a4 <_sk_callback_hsw+0x3e1> + DB 196,226,125,24,5,88,6,0,0 ; vbroadcastss 0x658(%rip),%ymm0 # 4100 <_sk_callback_hsw+0x3fd> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,51,6,0,0 ; vbroadcastss 0x633(%rip),%ymm8 # 40a8 <_sk_callback_hsw+0x3e5> + DB 196,98,125,24,5,79,6,0,0 ; vbroadcastss 0x64f(%rip),%ymm8 # 4104 <_sk_callback_hsw+0x401> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,36,6,0,0 ; vbroadcastss 0x624(%rip),%ymm10 # 40ac <_sk_callback_hsw+0x3e9> - DB 196,98,125,24,29,31,6,0,0 ; vbroadcastss 0x61f(%rip),%ymm11 # 40b0 <_sk_callback_hsw+0x3ed> + DB 196,98,125,24,21,64,6,0,0 ; vbroadcastss 0x640(%rip),%ymm10 # 4108 <_sk_callback_hsw+0x405> + DB 196,98,125,24,29,59,6,0,0 ; vbroadcastss 0x63b(%rip),%ymm11 # 410c <_sk_callback_hsw+0x409> 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) @@ -3629,16 +3631,16 @@ _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 196,226,125,24,5,2,6,0,0 ; vbroadcastss 0x602(%rip),%ymm0 # 40b4 <_sk_callback_hsw+0x3f1> + DB 196,226,125,24,5,30,6,0,0 ; vbroadcastss 0x61e(%rip),%ymm0 # 4110 <_sk_callback_hsw+0x40d> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,249,5,0,0 ; vbroadcastss 0x5f9(%rip),%ymm8 # 40b8 <_sk_callback_hsw+0x3f5> + DB 196,98,125,24,5,21,6,0,0 ; vbroadcastss 0x615(%rip),%ymm8 # 4114 <_sk_callback_hsw+0x411> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,239,5,0,0 ; vbroadcastss 0x5ef(%rip),%ymm9 # 40bc <_sk_callback_hsw+0x3f9> - DB 196,98,125,24,21,234,5,0,0 ; vbroadcastss 0x5ea(%rip),%ymm10 # 40c0 <_sk_callback_hsw+0x3fd> + DB 196,98,125,24,13,11,6,0,0 ; vbroadcastss 0x60b(%rip),%ymm9 # 4118 <_sk_callback_hsw+0x415> + DB 196,98,125,24,21,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm10 # 411c <_sk_callback_hsw+0x419> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,224,5,0,0 ; vbroadcastss 0x5e0(%rip),%ymm9 # 40c4 <_sk_callback_hsw+0x401> + DB 196,98,125,24,13,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm9 # 4120 <_sk_callback_hsw+0x41d> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,214,5,0,0 ; vbroadcastss 0x5d6(%rip),%ymm10 # 40c8 <_sk_callback_hsw+0x405> + DB 196,98,125,24,21,242,5,0,0 ; vbroadcastss 0x5f2(%rip),%ymm10 # 4124 <_sk_callback_hsw+0x421> 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 @@ -3647,14 +3649,14 @@ _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 196,98,125,24,5,190,5,0,0 ; vbroadcastss 0x5be(%rip),%ymm8 # 40cc <_sk_callback_hsw+0x409> + DB 196,98,125,24,5,218,5,0,0 ; vbroadcastss 0x5da(%rip),%ymm8 # 4128 <_sk_callback_hsw+0x425> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,176,5,0,0 ; vbroadcastss 0x5b0(%rip),%ymm10 # 40d0 <_sk_callback_hsw+0x40d> - DB 196,98,125,24,29,171,5,0,0 ; vbroadcastss 0x5ab(%rip),%ymm11 # 40d4 <_sk_callback_hsw+0x411> + DB 196,98,125,24,21,204,5,0,0 ; vbroadcastss 0x5cc(%rip),%ymm10 # 412c <_sk_callback_hsw+0x429> + DB 196,98,125,24,29,199,5,0,0 ; vbroadcastss 0x5c7(%rip),%ymm11 # 4130 <_sk_callback_hsw+0x42d> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,156,5,0,0 ; vbroadcastss 0x59c(%rip),%ymm8 # 40d8 <_sk_callback_hsw+0x415> + DB 196,98,125,24,5,184,5,0,0 ; vbroadcastss 0x5b8(%rip),%ymm8 # 4134 <_sk_callback_hsw+0x431> 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 @@ -3663,12 +3665,12 @@ _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 196,226,125,24,5,132,5,0,0 ; vbroadcastss 0x584(%rip),%ymm0 # 40dc <_sk_callback_hsw+0x419> + DB 196,226,125,24,5,160,5,0,0 ; vbroadcastss 0x5a0(%rip),%ymm0 # 4138 <_sk_callback_hsw+0x435> 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 196,98,125,24,21,113,5,0,0 ; vbroadcastss 0x571(%rip),%ymm10 # 40e0 <_sk_callback_hsw+0x41d> - DB 196,98,125,24,29,108,5,0,0 ; vbroadcastss 0x56c(%rip),%ymm11 # 40e4 <_sk_callback_hsw+0x421> + DB 196,98,125,24,21,141,5,0,0 ; vbroadcastss 0x58d(%rip),%ymm10 # 413c <_sk_callback_hsw+0x439> + DB 196,98,125,24,29,136,5,0,0 ; vbroadcastss 0x588(%rip),%ymm11 # 4140 <_sk_callback_hsw+0x43d> 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) @@ -3678,13 +3680,13 @@ _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 196,226,125,24,13,79,5,0,0 ; vbroadcastss 0x54f(%rip),%ymm1 # 40e8 <_sk_callback_hsw+0x425> + DB 196,226,125,24,13,107,5,0,0 ; vbroadcastss 0x56b(%rip),%ymm1 # 4144 <_sk_callback_hsw+0x441> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,69,5,0,0 ; vbroadcastss 0x545(%rip),%ymm8 # 40ec <_sk_callback_hsw+0x429> + DB 196,98,125,24,5,97,5,0,0 ; vbroadcastss 0x561(%rip),%ymm8 # 4148 <_sk_callback_hsw+0x445> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,54,5,0,0 ; vbroadcastss 0x536(%rip),%ymm10 # 40f0 <_sk_callback_hsw+0x42d> - DB 196,98,125,24,29,49,5,0,0 ; vbroadcastss 0x531(%rip),%ymm11 # 40f4 <_sk_callback_hsw+0x431> + DB 196,98,125,24,21,82,5,0,0 ; vbroadcastss 0x552(%rip),%ymm10 # 414c <_sk_callback_hsw+0x449> + DB 196,98,125,24,29,77,5,0,0 ; vbroadcastss 0x54d(%rip),%ymm11 # 4150 <_sk_callback_hsw+0x44d> 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) @@ -3694,16 +3696,16 @@ _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 196,226,125,24,13,20,5,0,0 ; vbroadcastss 0x514(%rip),%ymm1 # 40f8 <_sk_callback_hsw+0x435> + DB 196,226,125,24,13,48,5,0,0 ; vbroadcastss 0x530(%rip),%ymm1 # 4154 <_sk_callback_hsw+0x451> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,10,5,0,0 ; vbroadcastss 0x50a(%rip),%ymm8 # 40fc <_sk_callback_hsw+0x439> + DB 196,98,125,24,5,38,5,0,0 ; vbroadcastss 0x526(%rip),%ymm8 # 4158 <_sk_callback_hsw+0x455> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,0,5,0,0 ; vbroadcastss 0x500(%rip),%ymm9 # 4100 <_sk_callback_hsw+0x43d> - DB 196,98,125,24,21,251,4,0,0 ; vbroadcastss 0x4fb(%rip),%ymm10 # 4104 <_sk_callback_hsw+0x441> + DB 196,98,125,24,13,28,5,0,0 ; vbroadcastss 0x51c(%rip),%ymm9 # 415c <_sk_callback_hsw+0x459> + DB 196,98,125,24,21,23,5,0,0 ; vbroadcastss 0x517(%rip),%ymm10 # 4160 <_sk_callback_hsw+0x45d> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,241,4,0,0 ; vbroadcastss 0x4f1(%rip),%ymm9 # 4108 <_sk_callback_hsw+0x445> + DB 196,98,125,24,13,13,5,0,0 ; vbroadcastss 0x50d(%rip),%ymm9 # 4164 <_sk_callback_hsw+0x461> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,231,4,0,0 ; vbroadcastss 0x4e7(%rip),%ymm10 # 410c <_sk_callback_hsw+0x449> + DB 196,98,125,24,21,3,5,0,0 ; vbroadcastss 0x503(%rip),%ymm10 # 4168 <_sk_callback_hsw+0x465> 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 @@ -3712,14 +3714,14 @@ _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 196,98,125,24,5,207,4,0,0 ; vbroadcastss 0x4cf(%rip),%ymm8 # 4110 <_sk_callback_hsw+0x44d> + DB 196,98,125,24,5,235,4,0,0 ; vbroadcastss 0x4eb(%rip),%ymm8 # 416c <_sk_callback_hsw+0x469> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,192,4,0,0 ; vbroadcastss 0x4c0(%rip),%ymm10 # 4114 <_sk_callback_hsw+0x451> - DB 196,98,125,24,29,187,4,0,0 ; vbroadcastss 0x4bb(%rip),%ymm11 # 4118 <_sk_callback_hsw+0x455> + DB 196,98,125,24,21,220,4,0,0 ; vbroadcastss 0x4dc(%rip),%ymm10 # 4170 <_sk_callback_hsw+0x46d> + DB 196,98,125,24,29,215,4,0,0 ; vbroadcastss 0x4d7(%rip),%ymm11 # 4174 <_sk_callback_hsw+0x471> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,172,4,0,0 ; vbroadcastss 0x4ac(%rip),%ymm8 # 411c <_sk_callback_hsw+0x459> + DB 196,98,125,24,5,200,4,0,0 ; vbroadcastss 0x4c8(%rip),%ymm8 # 4178 <_sk_callback_hsw+0x475> 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 @@ -3728,12 +3730,12 @@ _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 196,226,125,24,13,148,4,0,0 ; vbroadcastss 0x494(%rip),%ymm1 # 4120 <_sk_callback_hsw+0x45d> + DB 196,226,125,24,13,176,4,0,0 ; vbroadcastss 0x4b0(%rip),%ymm1 # 417c <_sk_callback_hsw+0x479> 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 196,98,125,24,21,128,4,0,0 ; vbroadcastss 0x480(%rip),%ymm10 # 4124 <_sk_callback_hsw+0x461> - DB 196,98,125,24,29,123,4,0,0 ; vbroadcastss 0x47b(%rip),%ymm11 # 4128 <_sk_callback_hsw+0x465> + DB 196,98,125,24,21,156,4,0,0 ; vbroadcastss 0x49c(%rip),%ymm10 # 4180 <_sk_callback_hsw+0x47d> + DB 196,98,125,24,29,151,4,0,0 ; vbroadcastss 0x497(%rip),%ymm11 # 4184 <_sk_callback_hsw+0x481> 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) @@ -3863,11 +3865,23 @@ ALIGN 4 DB 0,128,64,171,170,42 ; add %al,0x2aaaab40(%rax) DB 62,0,0 ; add %al,%ds:(%rax) DB 0,63 ; add %bh,(%rdi) - DB 129,128,128,59,129,128,128,59,0,248 ; addl $0xf8003b80,-0x7f7ec480(%rax) DB 0,0 ; add %al,(%rax) - DB 8,33 ; or %ah,(%rcx) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 64,171 ; rex stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,0,0 ; add %al,%ds:(%rax) + DB 128,191,0,0,192,64,171 ; cmpb $0xab,0x40c00000(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + 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 3e99 <.literal4+0xbd> + DB 224,7 ; loopne 3ef5 <.literal4+0xd9> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -3881,10 +3895,10 @@ 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 3ec4 <.literal4+0xe8> + DB 127,0 ; jg 3f20 <.literal4+0x104> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 3f3d <.literal4+0x161> + DB 119,115 ; ja 3f99 <.literal4+0x17d> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -3898,10 +3912,10 @@ 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 3ef8 <.literal4+0x11c> + DB 127,0 ; jg 3f54 <.literal4+0x138> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 3f71 <.literal4+0x195> + DB 119,115 ; ja 3fcd <.literal4+0x1b1> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -3915,10 +3929,10 @@ 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 3f2c <.literal4+0x150> + DB 127,0 ; jg 3f88 <.literal4+0x16c> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 3fa5 <.literal4+0x1c9> + DB 119,115 ; ja 4001 <.literal4+0x1e5> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -3932,10 +3946,10 @@ 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 3f60 <.literal4+0x184> + DB 127,0 ; jg 3fbc <.literal4+0x1a0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 3fd9 <.literal4+0x1fd> + DB 119,115 ; ja 4035 <.literal4+0x219> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -3948,7 +3962,7 @@ 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 3fd7 <.literal4+0x1fb> + DB 127,67 ; jg 4033 <.literal4+0x217> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -3960,10 +3974,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 3ff7 <.literal4+0x21b> + DB 118,63 ; jbe 4053 <.literal4+0x237> 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 400b <.literal4+0x22f> + DB 127,67 ; jg 4067 <.literal4+0x24b> 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) @@ -3972,7 +3986,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 3fed <.literal4+0x211> + DB 224,7 ; loopne 4049 <.literal4+0x22d> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -3984,7 +3998,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4009 <.literal4+0x22d> + DB 224,7 ; loopne 4065 <.literal4+0x249> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -3995,7 +4009,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 405e <.literal4+0x282> + DB 124,66 ; jl 40ba <.literal4+0x29e> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -4013,9 +4027,9 @@ 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 40a1 <.literal4+0x2c5> + DB 112,65 ; jo 40fd <.literal4+0x2e1> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 40af <.literal4+0x2d3> + DB 127,67 ; jg 410b <.literal4+0x2ef> DB 128,0,128 ; addb $0x80,(%rax) DB 55 ; (bad) DB 128,0,128 ; addb $0x80,(%rax) @@ -4023,7 +4037,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 40c3 <.literal4+0x2e7> + DB 127,71 ; jg 411f <.literal4+0x303> DB 208 ; (bad) DB 179,89 ; mov $0x59,%bl DB 62,89 ; ds pop %rcx @@ -4109,16 +4123,16 @@ 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) # a004168 <_sk_callback_hsw+0xa0004a5> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0041c8 <_sk_callback_hsw+0xa0004c5> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004170 <_sk_callback_hsw+0x120004ad> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120041d0 <_sk_callback_hsw+0x120004cd> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004178 <_sk_callback_hsw+0x1a0004b5> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0041d8 <_sk_callback_hsw+0x1a0004d5> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004180 <_sk_callback_hsw+0x30004bd> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30041e0 <_sk_callback_hsw+0x30004dd> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -4161,16 +4175,16 @@ 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) # a0041c8 <_sk_callback_hsw+0xa000505> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004228 <_sk_callback_hsw+0xa000525> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120041d0 <_sk_callback_hsw+0x1200050d> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004230 <_sk_callback_hsw+0x1200052d> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0041d8 <_sk_callback_hsw+0x1a000515> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004238 <_sk_callback_hsw+0x1a000535> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30041e0 <_sk_callback_hsw+0x300051d> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004240 <_sk_callback_hsw+0x300053d> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -4213,16 +4227,16 @@ 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) # a004228 <_sk_callback_hsw+0xa000565> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004288 <_sk_callback_hsw+0xa000585> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004230 <_sk_callback_hsw+0x1200056d> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004290 <_sk_callback_hsw+0x1200058d> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004238 <_sk_callback_hsw+0x1a000575> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004298 <_sk_callback_hsw+0x1a000595> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004240 <_sk_callback_hsw+0x300057d> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30042a0 <_sk_callback_hsw+0x300059d> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -4265,16 +4279,16 @@ 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) # a004288 <_sk_callback_hsw+0xa0005c5> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0042e8 <_sk_callback_hsw+0xa0005e5> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004290 <_sk_callback_hsw+0x120005cd> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120042f0 <_sk_callback_hsw+0x120005ed> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004298 <_sk_callback_hsw+0x1a0005d5> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0042f8 <_sk_callback_hsw+0x1a0005f5> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30042a0 <_sk_callback_hsw+0x30005dd> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004300 <_sk_callback_hsw+0x30005fd> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -4416,14 +4430,14 @@ _sk_seed_shader_avx LABEL PROC 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 196,226,125,24,13,140,83,0,0 ; vbroadcastss 0x538c(%rip),%ymm1 # 54ec <_sk_callback_avx+0x11a> + DB 196,226,125,24,13,156,83,0,0 ; vbroadcastss 0x539c(%rip),%ymm1 # 54fc <_sk_callback_avx+0x11a> 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 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,112,83,0,0 ; vbroadcastss 0x5370(%rip),%ymm2 # 54f0 <_sk_callback_avx+0x11e> + DB 196,226,125,24,21,128,83,0,0 ; vbroadcastss 0x5380(%rip),%ymm2 # 5500 <_sk_callback_avx+0x11e> 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 @@ -4453,7 +4467,7 @@ _sk_clear_avx LABEL PROC PUBLIC _sk_srcatop_avx _sk_srcatop_avx LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,32,83,0,0 ; vbroadcastss 0x5320(%rip),%ymm8 # 54f4 <_sk_callback_avx+0x122> + DB 196,98,125,24,5,48,83,0,0 ; vbroadcastss 0x5330(%rip),%ymm8 # 5504 <_sk_callback_avx+0x122> 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 @@ -4472,7 +4486,7 @@ _sk_srcatop_avx LABEL PROC PUBLIC _sk_dstatop_avx _sk_dstatop_avx LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 - DB 196,98,125,24,13,226,82,0,0 ; vbroadcastss 0x52e2(%rip),%ymm9 # 54f8 <_sk_callback_avx+0x126> + DB 196,98,125,24,13,242,82,0,0 ; vbroadcastss 0x52f2(%rip),%ymm9 # 5508 <_sk_callback_avx+0x126> 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 @@ -4508,7 +4522,7 @@ _sk_dstin_avx LABEL PROC PUBLIC _sk_srcout_avx _sk_srcout_avx LABEL PROC - DB 196,98,125,24,5,129,82,0,0 ; vbroadcastss 0x5281(%rip),%ymm8 # 54fc <_sk_callback_avx+0x12a> + DB 196,98,125,24,5,145,82,0,0 ; vbroadcastss 0x5291(%rip),%ymm8 # 550c <_sk_callback_avx+0x12a> 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 @@ -4519,7 +4533,7 @@ _sk_srcout_avx LABEL PROC PUBLIC _sk_dstout_avx _sk_dstout_avx LABEL PROC - DB 196,226,125,24,5,100,82,0,0 ; vbroadcastss 0x5264(%rip),%ymm0 # 5500 <_sk_callback_avx+0x12e> + DB 196,226,125,24,5,116,82,0,0 ; vbroadcastss 0x5274(%rip),%ymm0 # 5510 <_sk_callback_avx+0x12e> 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 @@ -4530,7 +4544,7 @@ _sk_dstout_avx LABEL PROC PUBLIC _sk_srcover_avx _sk_srcover_avx LABEL PROC - DB 196,98,125,24,5,71,82,0,0 ; vbroadcastss 0x5247(%rip),%ymm8 # 5504 <_sk_callback_avx+0x132> + DB 196,98,125,24,5,87,82,0,0 ; vbroadcastss 0x5257(%rip),%ymm8 # 5514 <_sk_callback_avx+0x132> 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 @@ -4545,7 +4559,7 @@ _sk_srcover_avx LABEL PROC PUBLIC _sk_dstover_avx _sk_dstover_avx LABEL PROC - DB 196,98,125,24,5,26,82,0,0 ; vbroadcastss 0x521a(%rip),%ymm8 # 5508 <_sk_callback_avx+0x136> + DB 196,98,125,24,5,42,82,0,0 ; vbroadcastss 0x522a(%rip),%ymm8 # 5518 <_sk_callback_avx+0x136> 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 @@ -4569,7 +4583,7 @@ _sk_modulate_avx LABEL PROC PUBLIC _sk_multiply_avx _sk_multiply_avx LABEL PROC - DB 196,98,125,24,5,217,81,0,0 ; vbroadcastss 0x51d9(%rip),%ymm8 # 550c <_sk_callback_avx+0x13a> + DB 196,98,125,24,5,233,81,0,0 ; vbroadcastss 0x51e9(%rip),%ymm8 # 551c <_sk_callback_avx+0x13a> 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 @@ -4623,7 +4637,7 @@ _sk_screen_avx LABEL PROC PUBLIC _sk_xor__avx _sk_xor__avx LABEL PROC - DB 196,98,125,24,5,40,81,0,0 ; vbroadcastss 0x5128(%rip),%ymm8 # 5510 <_sk_callback_avx+0x13e> + DB 196,98,125,24,5,56,81,0,0 ; vbroadcastss 0x5138(%rip),%ymm8 # 5520 <_sk_callback_avx+0x13e> 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 @@ -4658,7 +4672,7 @@ _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 196,98,125,24,5,168,80,0,0 ; vbroadcastss 0x50a8(%rip),%ymm8 # 5514 <_sk_callback_avx+0x142> + DB 196,98,125,24,5,184,80,0,0 ; vbroadcastss 0x50b8(%rip),%ymm8 # 5524 <_sk_callback_avx+0x142> 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 @@ -4682,7 +4696,7 @@ _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 196,98,125,24,5,84,80,0,0 ; vbroadcastss 0x5054(%rip),%ymm8 # 5518 <_sk_callback_avx+0x146> + DB 196,98,125,24,5,100,80,0,0 ; vbroadcastss 0x5064(%rip),%ymm8 # 5528 <_sk_callback_avx+0x146> 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 @@ -4709,7 +4723,7 @@ _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 196,98,125,24,5,244,79,0,0 ; vbroadcastss 0x4ff4(%rip),%ymm8 # 551c <_sk_callback_avx+0x14a> + DB 196,98,125,24,5,4,80,0,0 ; vbroadcastss 0x5004(%rip),%ymm8 # 552c <_sk_callback_avx+0x14a> 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 @@ -4730,7 +4744,7 @@ _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 196,98,125,24,5,175,79,0,0 ; vbroadcastss 0x4faf(%rip),%ymm8 # 5520 <_sk_callback_avx+0x14e> + DB 196,98,125,24,5,191,79,0,0 ; vbroadcastss 0x4fbf(%rip),%ymm8 # 5530 <_sk_callback_avx+0x14e> 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 @@ -4739,7 +4753,7 @@ _sk_exclusion_avx LABEL PROC PUBLIC _sk_colorburn_avx _sk_colorburn_avx LABEL PROC - DB 196,98,125,24,5,154,79,0,0 ; vbroadcastss 0x4f9a(%rip),%ymm8 # 5524 <_sk_callback_avx+0x152> + DB 196,98,125,24,5,170,79,0,0 ; vbroadcastss 0x4faa(%rip),%ymm8 # 5534 <_sk_callback_avx+0x152> 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 @@ -4799,7 +4813,7 @@ _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 196,98,125,24,13,150,78,0,0 ; vbroadcastss 0x4e96(%rip),%ymm9 # 5528 <_sk_callback_avx+0x156> + DB 196,98,125,24,13,166,78,0,0 ; vbroadcastss 0x4ea6(%rip),%ymm9 # 5538 <_sk_callback_avx+0x156> 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 @@ -4854,7 +4868,7 @@ _sk_colordodge_avx LABEL PROC PUBLIC _sk_hardlight_avx _sk_hardlight_avx LABEL PROC - DB 196,98,125,24,5,168,77,0,0 ; vbroadcastss 0x4da8(%rip),%ymm8 # 552c <_sk_callback_avx+0x15a> + DB 196,98,125,24,5,184,77,0,0 ; vbroadcastss 0x4db8(%rip),%ymm8 # 553c <_sk_callback_avx+0x15a> 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 @@ -4907,7 +4921,7 @@ _sk_hardlight_avx LABEL PROC PUBLIC _sk_overlay_avx _sk_overlay_avx LABEL PROC - DB 196,98,125,24,5,209,76,0,0 ; vbroadcastss 0x4cd1(%rip),%ymm8 # 5530 <_sk_callback_avx+0x15e> + DB 196,98,125,24,5,225,76,0,0 ; vbroadcastss 0x4ce1(%rip),%ymm8 # 5540 <_sk_callback_avx+0x15e> 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 @@ -4972,10 +4986,10 @@ _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 196,98,125,24,29,196,75,0,0 ; vbroadcastss 0x4bc4(%rip),%ymm11 # 5538 <_sk_callback_avx+0x166> + DB 196,98,125,24,29,212,75,0,0 ; vbroadcastss 0x4bd4(%rip),%ymm11 # 5548 <_sk_callback_avx+0x166> DB 196,65,28,88,235 ; vaddps %ymm11,%ymm12,%ymm13 DB 196,65,20,89,192 ; vmulps %ymm8,%ymm13,%ymm8 - DB 196,98,125,24,45,181,75,0,0 ; vbroadcastss 0x4bb5(%rip),%ymm13 # 553c <_sk_callback_avx+0x16a> + DB 196,98,125,24,45,197,75,0,0 ; vbroadcastss 0x4bc5(%rip),%ymm13 # 554c <_sk_callback_avx+0x16a> 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 @@ -4986,7 +5000,7 @@ _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 196,98,125,24,5,115,75,0,0 ; vbroadcastss 0x4b73(%rip),%ymm8 # 5534 <_sk_callback_avx+0x162> + DB 196,98,125,24,5,131,75,0,0 ; vbroadcastss 0x4b83(%rip),%ymm8 # 5544 <_sk_callback_avx+0x162> 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 @@ -5090,7 +5104,7 @@ _sk_clamp_0_avx LABEL PROC PUBLIC _sk_clamp_1_avx _sk_clamp_1_avx LABEL PROC - DB 196,98,125,24,5,193,73,0,0 ; vbroadcastss 0x49c1(%rip),%ymm8 # 5540 <_sk_callback_avx+0x16e> + DB 196,98,125,24,5,209,73,0,0 ; vbroadcastss 0x49d1(%rip),%ymm8 # 5550 <_sk_callback_avx+0x16e> 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 @@ -5100,7 +5114,7 @@ _sk_clamp_1_avx LABEL PROC PUBLIC _sk_clamp_a_avx _sk_clamp_a_avx LABEL PROC - DB 196,98,125,24,5,164,73,0,0 ; vbroadcastss 0x49a4(%rip),%ymm8 # 5544 <_sk_callback_avx+0x172> + DB 196,98,125,24,5,180,73,0,0 ; vbroadcastss 0x49b4(%rip),%ymm8 # 5554 <_sk_callback_avx+0x172> 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 @@ -5172,7 +5186,7 @@ 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 196,98,125,24,21,236,72,0,0 ; vbroadcastss 0x48ec(%rip),%ymm10 # 5548 <_sk_callback_avx+0x176> + DB 196,98,125,24,21,252,72,0,0 ; vbroadcastss 0x48fc(%rip),%ymm10 # 5558 <_sk_callback_avx+0x176> 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 @@ -5183,17 +5197,17 @@ _sk_unpremul_avx LABEL PROC PUBLIC _sk_from_srgb_avx _sk_from_srgb_avx LABEL PROC - DB 196,98,125,24,5,205,72,0,0 ; vbroadcastss 0x48cd(%rip),%ymm8 # 554c <_sk_callback_avx+0x17a> + DB 196,98,125,24,5,221,72,0,0 ; vbroadcastss 0x48dd(%rip),%ymm8 # 555c <_sk_callback_avx+0x17a> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,191,72,0,0 ; vbroadcastss 0x48bf(%rip),%ymm11 # 5550 <_sk_callback_avx+0x17e> + DB 196,98,125,24,29,207,72,0,0 ; vbroadcastss 0x48cf(%rip),%ymm11 # 5560 <_sk_callback_avx+0x17e> DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12 - DB 196,98,125,24,45,181,72,0,0 ; vbroadcastss 0x48b5(%rip),%ymm13 # 5554 <_sk_callback_avx+0x182> + DB 196,98,125,24,45,197,72,0,0 ; vbroadcastss 0x48c5(%rip),%ymm13 # 5564 <_sk_callback_avx+0x182> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,166,72,0,0 ; vbroadcastss 0x48a6(%rip),%ymm12 # 5558 <_sk_callback_avx+0x186> + DB 196,98,125,24,37,182,72,0,0 ; vbroadcastss 0x48b6(%rip),%ymm12 # 5568 <_sk_callback_avx+0x186> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,53,156,72,0,0 ; vbroadcastss 0x489c(%rip),%ymm14 # 555c <_sk_callback_avx+0x18a> + DB 196,98,125,24,53,172,72,0,0 ; vbroadcastss 0x48ac(%rip),%ymm14 # 556c <_sk_callback_avx+0x18a> 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 @@ -5220,18 +5234,18 @@ _sk_to_srgb_avx LABEL PROC DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8 DB 196,65,124,83,200 ; vrcpps %ymm8,%ymm9 DB 196,65,124,82,208 ; vrsqrtps %ymm8,%ymm10 - DB 196,98,125,24,5,39,72,0,0 ; vbroadcastss 0x4827(%rip),%ymm8 # 5560 <_sk_callback_avx+0x18e> + DB 196,98,125,24,5,55,72,0,0 ; vbroadcastss 0x4837(%rip),%ymm8 # 5570 <_sk_callback_avx+0x18e> DB 196,65,124,89,216 ; vmulps %ymm8,%ymm0,%ymm11 - DB 196,98,125,24,37,29,72,0,0 ; vbroadcastss 0x481d(%rip),%ymm12 # 5564 <_sk_callback_avx+0x192> + DB 196,98,125,24,37,45,72,0,0 ; vbroadcastss 0x482d(%rip),%ymm12 # 5574 <_sk_callback_avx+0x192> DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 - DB 196,98,125,24,45,19,72,0,0 ; vbroadcastss 0x4813(%rip),%ymm13 # 5568 <_sk_callback_avx+0x196> + DB 196,98,125,24,45,35,72,0,0 ; vbroadcastss 0x4823(%rip),%ymm13 # 5578 <_sk_callback_avx+0x196> DB 196,65,52,88,205 ; vaddps %ymm13,%ymm9,%ymm9 - DB 196,98,125,24,53,9,72,0,0 ; vbroadcastss 0x4809(%rip),%ymm14 # 556c <_sk_callback_avx+0x19a> + DB 196,98,125,24,53,25,72,0,0 ; vbroadcastss 0x4819(%rip),%ymm14 # 557c <_sk_callback_avx+0x19a> DB 196,65,44,89,214 ; vmulps %ymm14,%ymm10,%ymm10 DB 196,65,44,88,201 ; vaddps %ymm9,%ymm10,%ymm9 - DB 196,98,125,24,21,250,71,0,0 ; vbroadcastss 0x47fa(%rip),%ymm10 # 5570 <_sk_callback_avx+0x19e> + DB 196,98,125,24,21,10,72,0,0 ; vbroadcastss 0x480a(%rip),%ymm10 # 5580 <_sk_callback_avx+0x19e> DB 196,65,44,93,201 ; vminps %ymm9,%ymm10,%ymm9 - DB 196,98,125,24,61,240,71,0,0 ; vbroadcastss 0x47f0(%rip),%ymm15 # 5574 <_sk_callback_avx+0x1a2> + DB 196,98,125,24,61,0,72,0,0 ; vbroadcastss 0x4800(%rip),%ymm15 # 5584 <_sk_callback_avx+0x1a2> DB 196,193,124,194,199,1 ; vcmpltps %ymm15,%ymm0,%ymm0 DB 196,195,53,74,195,0 ; vblendvps %ymm0,%ymm11,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -5266,7 +5280,7 @@ _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 196,98,125,24,29,86,71,0,0 ; vbroadcastss 0x4756(%rip),%ymm11 # 5578 <_sk_callback_avx+0x1a6> + DB 196,98,125,24,29,102,71,0,0 ; vbroadcastss 0x4766(%rip),%ymm11 # 5588 <_sk_callback_avx+0x1a6> 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 @@ -5276,19 +5290,19 @@ _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 196,98,125,24,29,47,71,0,0 ; vbroadcastss 0x472f(%rip),%ymm11 # 5584 <_sk_callback_avx+0x1b2> + DB 196,98,125,24,29,63,71,0,0 ; vbroadcastss 0x473f(%rip),%ymm11 # 5594 <_sk_callback_avx+0x1b2> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,29,71,0,0 ; vbroadcastss 0x471d(%rip),%ymm11 # 5580 <_sk_callback_avx+0x1ae> + DB 196,98,125,24,29,45,71,0,0 ; vbroadcastss 0x472d(%rip),%ymm11 # 5590 <_sk_callback_avx+0x1ae> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,227,117,74,202,224 ; vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,5,71,0,0 ; vbroadcastss 0x4705(%rip),%ymm2 # 557c <_sk_callback_avx+0x1aa> + DB 196,226,125,24,21,21,71,0,0 ; vbroadcastss 0x4715(%rip),%ymm2 # 558c <_sk_callback_avx+0x1aa> 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 196,98,125,24,37,236,70,0,0 ; vbroadcastss 0x46ec(%rip),%ymm12 # 558c <_sk_callback_avx+0x1ba> + DB 196,98,125,24,37,252,70,0,0 ; vbroadcastss 0x46fc(%rip),%ymm12 # 559c <_sk_callback_avx+0x1ba> 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 @@ -5298,130 +5312,126 @@ _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 196,98,125,24,5,175,70,0,0 ; vbroadcastss 0x46af(%rip),%ymm8 # 5588 <_sk_callback_avx+0x1b6> + DB 196,98,125,24,5,191,70,0,0 ; vbroadcastss 0x46bf(%rip),%ymm8 # 5598 <_sk_callback_avx+0x1b6> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_hsl_to_rgb_avx _sk_hsl_to_rgb_avx LABEL PROC - 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 197,252,17,28,36 ; vmovups %ymm3,(%rsp) - DB 197,252,40,226 ; vmovaps %ymm2,%ymm4 - DB 197,252,40,233 ; vmovaps %ymm1,%ymm5 - DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 + DB 72,129,236,248,0,0,0 ; sub $0xf8,%rsp + DB 197,252,17,188,36,192,0,0,0 ; vmovups %ymm7,0xc0(%rsp) + DB 197,252,17,180,36,160,0,0,0 ; vmovups %ymm6,0xa0(%rsp) + DB 197,252,17,172,36,128,0,0,0 ; vmovups %ymm5,0x80(%rsp) + DB 197,252,17,100,36,96 ; vmovups %ymm4,0x60(%rsp) + DB 197,252,17,92,36,64 ; vmovups %ymm3,0x40(%rsp) + DB 197,252,40,242 ; vmovaps %ymm2,%ymm6 + DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp) + DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 184,0,0,0,63 ; mov $0x3f000000,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 DB 196,99,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm8 - DB 196,193,92,194,192,1 ; vcmpltps %ymm8,%ymm4,%ymm0 + DB 196,193,76,194,192,1 ; vcmpltps %ymm8,%ymm6,%ymm0 + DB 196,98,125,24,21,94,70,0,0 ; vbroadcastss 0x465e(%rip),%ymm10 # 55a0 <_sk_callback_avx+0x1be> + DB 196,193,116,88,218 ; vaddps %ymm10,%ymm1,%ymm3 + DB 197,228,89,222 ; vmulps %ymm6,%ymm3,%ymm3 + DB 197,244,88,230 ; vaddps %ymm6,%ymm1,%ymm4 + DB 197,244,89,238 ; vmulps %ymm6,%ymm1,%ymm5 + DB 197,220,92,229 ; vsubps %ymm5,%ymm4,%ymm4 + DB 196,99,93,74,203,0 ; vblendvps %ymm0,%ymm3,%ymm4,%ymm9 + DB 196,226,125,24,5,62,70,0,0 ; vbroadcastss 0x463e(%rip),%ymm0 # 55a4 <_sk_callback_avx+0x1c2> + DB 197,236,88,200 ; vaddps %ymm0,%ymm2,%ymm1 + DB 65,184,0,0,0,0 ; mov $0x0,%r8d DB 184,0,0,128,63 ; mov $0x3f800000,%eax - DB 197,249,110,208 ; vmovd %eax,%xmm2 - DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,99,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm10 - DB 197,172,88,213 ; vaddps %ymm5,%ymm10,%ymm2 - DB 197,236,89,212 ; vmulps %ymm4,%ymm2,%ymm2 - DB 197,84,88,204 ; vaddps %ymm4,%ymm5,%ymm9 - DB 197,84,89,220 ; vmulps %ymm4,%ymm5,%ymm11 - DB 196,65,52,92,203 ; vsubps %ymm11,%ymm9,%ymm9 - DB 196,99,53,74,202,0 ; vblendvps %ymm0,%ymm2,%ymm9,%ymm9 - DB 65,184,0,0,0,64 ; mov $0x40000000,%r8d - DB 184,171,170,170,62 ; mov $0x3eaaaaab,%eax - DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,99,125,24,224,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 - DB 197,28,88,251 ; vaddps %ymm3,%ymm12,%ymm15 - DB 184,0,0,0,0 ; mov $0x0,%eax - DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,99,125,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 - DB 196,193,44,194,199,1 ; vcmpltps %ymm15,%ymm10,%ymm0 - DB 196,193,4,92,210 ; vsubps %ymm10,%ymm15,%ymm2 - DB 196,227,5,74,194,0 ; vblendvps %ymm0,%ymm2,%ymm15,%ymm0 - DB 196,193,4,194,213,1 ; vcmpltps %ymm13,%ymm15,%ymm2 - DB 196,65,44,88,223 ; vaddps %ymm15,%ymm10,%ymm11 - DB 196,195,125,74,203,32 ; vblendvps %ymm2,%ymm11,%ymm0,%ymm1 - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,252,89,196 ; vmulps %ymm4,%ymm0,%ymm0 - DB 196,65,124,92,217 ; vsubps %ymm9,%ymm0,%ymm11 + DB 197,249,110,216 ; vmovd %eax,%xmm3 + DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 + DB 196,99,101,24,227,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm12 + DB 197,156,194,217,1 ; vcmpltps %ymm1,%ymm12,%ymm3 + DB 196,98,125,24,53,21,70,0,0 ; vbroadcastss 0x4615(%rip),%ymm14 # 55a8 <_sk_callback_avx+0x1c6> + DB 196,193,116,88,230 ; vaddps %ymm14,%ymm1,%ymm4 + DB 196,227,117,74,220,48 ; vblendvps %ymm3,%ymm4,%ymm1,%ymm3 + DB 196,193,121,110,224 ; vmovd %r8d,%xmm4 + DB 196,227,121,4,228,0 ; vpermilps $0x0,%xmm4,%xmm4 + DB 196,99,93,24,252,1 ; vinsertf128 $0x1,%xmm4,%ymm4,%ymm15 + DB 196,193,116,194,231,1 ; vcmpltps %ymm15,%ymm1,%ymm4 + DB 196,193,116,88,234 ; vaddps %ymm10,%ymm1,%ymm5 + DB 196,227,101,74,197,64 ; vblendvps %ymm4,%ymm5,%ymm3,%ymm0 + DB 197,204,88,222 ; vaddps %ymm6,%ymm6,%ymm3 + DB 196,65,100,92,217 ; vsubps %ymm9,%ymm3,%ymm11 + DB 196,193,52,92,219 ; vsubps %ymm11,%ymm9,%ymm3 + DB 196,226,125,24,37,213,69,0,0 ; vbroadcastss 0x45d5(%rip),%ymm4 # 55ac <_sk_callback_avx+0x1ca> + DB 197,100,89,236 ; vmulps %ymm4,%ymm3,%ymm13 DB 65,184,171,170,42,62 ; mov $0x3e2aaaab,%r8d - DB 184,0,0,192,64 ; mov $0x40c00000,%eax - DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 196,193,52,92,211 ; vsubps %ymm11,%ymm9,%ymm2 - DB 197,108,89,240 ; vmulps %ymm0,%ymm2,%ymm14 DB 184,171,170,42,63 ; mov $0x3f2aaaab,%eax - DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2 - DB 197,140,89,210 ; vmulps %ymm2,%ymm14,%ymm2 - DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 197,244,194,248,1 ; vcmpltps %ymm0,%ymm1,%ymm7 - DB 196,227,37,74,210,112 ; vblendvps %ymm7,%ymm2,%ymm11,%ymm2 - DB 196,193,116,194,248,1 ; vcmpltps %ymm8,%ymm1,%ymm7 - DB 196,195,109,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm2,%ymm7 - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,244,194,202,1 ; vcmpltps %ymm2,%ymm1,%ymm1 - DB 196,65,4,89,254 ; vmulps %ymm14,%ymm15,%ymm15 - DB 196,65,36,88,255 ; vaddps %ymm15,%ymm11,%ymm15 - DB 196,67,69,74,255,16 ; vblendvps %ymm1,%ymm15,%ymm7,%ymm15 - DB 197,172,194,203,1 ; vcmpltps %ymm3,%ymm10,%ymm1 - DB 196,193,100,92,250 ; vsubps %ymm10,%ymm3,%ymm7 - DB 196,227,101,74,207,16 ; vblendvps %ymm1,%ymm7,%ymm3,%ymm1 - DB 196,193,100,194,253,1 ; vcmpltps %ymm13,%ymm3,%ymm7 - DB 197,172,88,243 ; vaddps %ymm3,%ymm10,%ymm6 - DB 196,227,117,74,206,112 ; vblendvps %ymm7,%ymm6,%ymm1,%ymm1 - DB 197,252,92,241 ; vsubps %ymm1,%ymm0,%ymm6 - DB 197,140,89,246 ; vmulps %ymm6,%ymm14,%ymm6 - DB 197,164,88,246 ; vaddps %ymm6,%ymm11,%ymm6 - DB 197,244,194,248,1 ; vcmpltps %ymm0,%ymm1,%ymm7 - DB 196,227,37,74,246,112 ; vblendvps %ymm7,%ymm6,%ymm11,%ymm6 - DB 196,193,116,194,248,1 ; vcmpltps %ymm8,%ymm1,%ymm7 - DB 196,195,77,74,241,112 ; vblendvps %ymm7,%ymm9,%ymm6,%ymm6 - DB 197,244,194,202,1 ; vcmpltps %ymm2,%ymm1,%ymm1 - DB 197,140,89,251 ; vmulps %ymm3,%ymm14,%ymm7 + DB 197,249,110,216 ; vmovd %eax,%xmm3 + DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 + DB 196,227,101,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm5 + DB 196,226,125,24,37,177,69,0,0 ; vbroadcastss 0x45b1(%rip),%ymm4 # 55b0 <_sk_callback_avx+0x1ce> + DB 197,220,92,216 ; vsubps %ymm0,%ymm4,%ymm3 + DB 197,148,89,219 ; vmulps %ymm3,%ymm13,%ymm3 + DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 + DB 197,252,194,253,1 ; vcmpltps %ymm5,%ymm0,%ymm7 + DB 196,227,37,74,219,112 ; vblendvps %ymm7,%ymm3,%ymm11,%ymm3 + DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7 + DB 196,195,101,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm3,%ymm7 + DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 + DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 + DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 + DB 197,252,194,195,1 ; vcmpltps %ymm3,%ymm0,%ymm0 + DB 196,193,116,89,205 ; vmulps %ymm13,%ymm1,%ymm1 + DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 + DB 196,227,69,74,193,0 ; vblendvps %ymm0,%ymm1,%ymm7,%ymm0 + DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) + DB 197,156,194,202,1 ; vcmpltps %ymm2,%ymm12,%ymm1 + DB 196,193,108,88,254 ; vaddps %ymm14,%ymm2,%ymm7 + DB 196,227,109,74,207,16 ; vblendvps %ymm1,%ymm7,%ymm2,%ymm1 + DB 196,193,108,194,255,1 ; vcmpltps %ymm15,%ymm2,%ymm7 + DB 196,193,108,88,194 ; vaddps %ymm10,%ymm2,%ymm0 + DB 196,227,117,74,192,112 ; vblendvps %ymm7,%ymm0,%ymm1,%ymm0 + DB 197,220,92,200 ; vsubps %ymm0,%ymm4,%ymm1 + DB 197,148,89,201 ; vmulps %ymm1,%ymm13,%ymm1 + DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 + DB 197,252,194,253,1 ; vcmpltps %ymm5,%ymm0,%ymm7 + DB 196,227,37,74,201,112 ; vblendvps %ymm7,%ymm1,%ymm11,%ymm1 + DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7 + DB 196,195,117,74,201,112 ; vblendvps %ymm7,%ymm9,%ymm1,%ymm1 + DB 197,252,194,195,1 ; vcmpltps %ymm3,%ymm0,%ymm0 + DB 197,148,89,250 ; vmulps %ymm2,%ymm13,%ymm7 DB 197,164,88,255 ; vaddps %ymm7,%ymm11,%ymm7 - DB 196,227,77,74,207,16 ; vblendvps %ymm1,%ymm7,%ymm6,%ymm1 - DB 196,193,100,92,220 ; vsubps %ymm12,%ymm3,%ymm3 - DB 197,172,194,243,1 ; vcmpltps %ymm3,%ymm10,%ymm6 - DB 196,193,100,92,250 ; vsubps %ymm10,%ymm3,%ymm7 - DB 196,227,101,74,247,96 ; vblendvps %ymm6,%ymm7,%ymm3,%ymm6 - DB 196,193,100,194,253,1 ; vcmpltps %ymm13,%ymm3,%ymm7 - DB 197,44,88,211 ; vaddps %ymm3,%ymm10,%ymm10 - DB 196,195,77,74,242,112 ; vblendvps %ymm7,%ymm10,%ymm6,%ymm6 - DB 197,204,194,248,1 ; vcmpltps %ymm0,%ymm6,%ymm7 - DB 197,252,92,198 ; vsubps %ymm6,%ymm0,%ymm0 - DB 197,140,89,192 ; vmulps %ymm0,%ymm14,%ymm0 + DB 196,227,117,74,207,0 ; vblendvps %ymm0,%ymm7,%ymm1,%ymm1 + DB 196,226,125,24,5,8,69,0,0 ; vbroadcastss 0x4508(%rip),%ymm0 # 55b4 <_sk_callback_avx+0x1d2> + DB 197,236,88,192 ; vaddps %ymm0,%ymm2,%ymm0 + DB 197,156,194,208,1 ; vcmpltps %ymm0,%ymm12,%ymm2 + DB 196,193,124,88,254 ; vaddps %ymm14,%ymm0,%ymm7 + DB 196,227,125,74,215,32 ; vblendvps %ymm2,%ymm7,%ymm0,%ymm2 + DB 196,193,124,194,255,1 ; vcmpltps %ymm15,%ymm0,%ymm7 + DB 196,65,124,88,210 ; vaddps %ymm10,%ymm0,%ymm10 + DB 196,195,109,74,210,112 ; vblendvps %ymm7,%ymm10,%ymm2,%ymm2 + DB 197,236,194,237,1 ; vcmpltps %ymm5,%ymm2,%ymm5 + DB 197,220,92,226 ; vsubps %ymm2,%ymm4,%ymm4 + DB 197,148,89,228 ; vmulps %ymm4,%ymm13,%ymm4 + DB 197,164,88,228 ; vaddps %ymm4,%ymm11,%ymm4 + DB 196,227,37,74,228,80 ; vblendvps %ymm5,%ymm4,%ymm11,%ymm4 + DB 196,193,108,194,232,1 ; vcmpltps %ymm8,%ymm2,%ymm5 + DB 196,195,93,74,225,80 ; vblendvps %ymm5,%ymm9,%ymm4,%ymm4 + DB 197,236,194,211,1 ; vcmpltps %ymm3,%ymm2,%ymm2 + DB 196,193,124,89,197 ; vmulps %ymm13,%ymm0,%ymm0 DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 196,227,37,74,192,112 ; vblendvps %ymm7,%ymm0,%ymm11,%ymm0 - DB 196,193,76,194,248,1 ; vcmpltps %ymm8,%ymm6,%ymm7 - DB 196,195,125,74,193,112 ; vblendvps %ymm7,%ymm9,%ymm0,%ymm0 - DB 197,204,194,210,1 ; vcmpltps %ymm2,%ymm6,%ymm2 - DB 196,193,100,89,222 ; vmulps %ymm14,%ymm3,%ymm3 - DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 196,227,125,74,211,32 ; vblendvps %ymm2,%ymm3,%ymm0,%ymm2 + DB 196,227,93,74,208,32 ; vblendvps %ymm2,%ymm0,%ymm4,%ymm2 DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 - DB 197,212,194,216,0 ; vcmpeqps %ymm0,%ymm5,%ymm3 - DB 196,227,5,74,196,48 ; vblendvps %ymm3,%ymm4,%ymm15,%ymm0 - DB 196,227,117,74,204,48 ; vblendvps %ymm3,%ymm4,%ymm1,%ymm1 - DB 196,227,109,74,212,48 ; vblendvps %ymm3,%ymm4,%ymm2,%ymm2 + DB 197,252,194,92,36,32,0 ; vcmpeqps 0x20(%rsp),%ymm0,%ymm3 + DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 + DB 196,227,125,74,198,48 ; vblendvps %ymm3,%ymm6,%ymm0,%ymm0 + DB 196,227,117,74,206,48 ; vblendvps %ymm3,%ymm6,%ymm1,%ymm1 + DB 196,227,109,74,214,48 ; vblendvps %ymm3,%ymm6,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,28,36 ; vmovups (%rsp),%ymm3 - 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 72,129,196,184,0,0,0 ; add $0xb8,%rsp + DB 197,252,16,92,36,64 ; vmovups 0x40(%rsp),%ymm3 + DB 197,252,16,100,36,96 ; vmovups 0x60(%rsp),%ymm4 + DB 197,252,16,172,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm5 + DB 197,252,16,180,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm6 + DB 197,252,16,188,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm7 + DB 72,129,196,248,0,0,0 ; add $0xf8,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_scale_1_float_avx @@ -5442,14 +5452,14 @@ _sk_scale_u8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,68 ; jne 11b9 <_sk_scale_u8_avx+0x54> + DB 117,68 ; jne 11cb <_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 196,98,125,24,13,243,67,0,0 ; vbroadcastss 0x43f3(%rip),%ymm9 # 5590 <_sk_callback_avx+0x1be> + DB 196,98,125,24,13,9,68,0,0 ; vbroadcastss 0x4409(%rip),%ymm9 # 55b8 <_sk_callback_avx+0x1d6> 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 @@ -5467,9 +5477,9 @@ _sk_scale_u8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 11c1 <_sk_scale_u8_avx+0x5c> + DB 117,234 ; jne 11d3 <_sk_scale_u8_avx+0x5c> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,155 ; jmp 1179 <_sk_scale_u8_avx+0x14> + DB 235,155 ; jmp 118b <_sk_scale_u8_avx+0x14> PUBLIC _sk_lerp_1_float_avx _sk_lerp_1_float_avx LABEL PROC @@ -5497,14 +5507,14 @@ _sk_lerp_u8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,104 ; jne 1295 <_sk_lerp_u8_avx+0x78> + DB 117,104 ; jne 12a7 <_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 196,98,125,24,13,63,67,0,0 ; vbroadcastss 0x433f(%rip),%ymm9 # 5594 <_sk_callback_avx+0x1c2> + DB 196,98,125,24,13,85,67,0,0 ; vbroadcastss 0x4355(%rip),%ymm9 # 55bc <_sk_callback_avx+0x1da> 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 @@ -5530,35 +5540,35 @@ _sk_lerp_u8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 129d <_sk_lerp_u8_avx+0x80> + DB 117,234 ; jne 12af <_sk_lerp_u8_avx+0x80> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 233,116,255,255,255 ; jmpq 1231 <_sk_lerp_u8_avx+0x14> + DB 233,116,255,255,255 ; jmpq 1243 <_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 15,133,174,0,0,0 ; jne 1379 <_sk_lerp_565_avx+0xbc> + DB 15,133,174,0,0,0 ; jne 138b <_sk_lerp_565_avx+0xbc> DB 196,65,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm8 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 197,185,105,219 ; vpunpckhwd %xmm3,%xmm8,%xmm3 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 - DB 196,98,125,24,5,171,66,0,0 ; vbroadcastss 0x42ab(%rip),%ymm8 # 5598 <_sk_callback_avx+0x1c6> + DB 196,98,125,24,5,193,66,0,0 ; vbroadcastss 0x42c1(%rip),%ymm8 # 55c0 <_sk_callback_avx+0x1de> DB 196,65,100,84,192 ; vandps %ymm8,%ymm3,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,156,66,0,0 ; vbroadcastss 0x429c(%rip),%ymm9 # 559c <_sk_callback_avx+0x1ca> + DB 196,98,125,24,13,178,66,0,0 ; vbroadcastss 0x42b2(%rip),%ymm9 # 55c4 <_sk_callback_avx+0x1e2> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,146,66,0,0 ; vbroadcastss 0x4292(%rip),%ymm9 # 55a0 <_sk_callback_avx+0x1ce> + DB 196,98,125,24,13,168,66,0,0 ; vbroadcastss 0x42a8(%rip),%ymm9 # 55c8 <_sk_callback_avx+0x1e6> DB 196,65,100,84,201 ; vandps %ymm9,%ymm3,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,131,66,0,0 ; vbroadcastss 0x4283(%rip),%ymm10 # 55a4 <_sk_callback_avx+0x1d2> + DB 196,98,125,24,21,153,66,0,0 ; vbroadcastss 0x4299(%rip),%ymm10 # 55cc <_sk_callback_avx+0x1ea> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,121,66,0,0 ; vbroadcastss 0x4279(%rip),%ymm10 # 55a8 <_sk_callback_avx+0x1d6> + DB 196,98,125,24,21,143,66,0,0 ; vbroadcastss 0x428f(%rip),%ymm10 # 55d0 <_sk_callback_avx+0x1ee> DB 196,193,100,84,218 ; vandps %ymm10,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,21,107,66,0,0 ; vbroadcastss 0x426b(%rip),%ymm10 # 55ac <_sk_callback_avx+0x1da> + DB 196,98,125,24,21,129,66,0,0 ; vbroadcastss 0x4281(%rip),%ymm10 # 55d4 <_sk_callback_avx+0x1f2> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 @@ -5570,16 +5580,16 @@ _sk_lerp_565_avx LABEL PROC DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 197,236,88,214 ; vaddps %ymm6,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,57,66,0,0 ; vbroadcastss 0x4239(%rip),%ymm3 # 55b0 <_sk_callback_avx+0x1de> + DB 196,226,125,24,29,79,66,0,0 ; vbroadcastss 0x424f(%rip),%ymm3 # 55d8 <_sk_callback_avx+0x1f6> DB 255,224 ; jmpq *%rax DB 65,137,200 ; mov %ecx,%r8d DB 65,128,224,7 ; and $0x7,%r8b 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 15,135,63,255,255,255 ; ja 12d1 <_sk_lerp_565_avx+0x14> + DB 15,135,63,255,255,255 ; ja 12e3 <_sk_lerp_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 13e8 <_sk_lerp_565_avx+0x12b> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 13f8 <_sk_lerp_565_avx+0x129> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -5591,28 +5601,27 @@ _sk_lerp_565_avx LABEL PROC 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 233,235,254,255,255 ; jmpq 12d1 <_sk_lerp_565_avx+0x14> - DB 102,144 ; xchg %ax,%ax - DB 242,255 ; repnz (bad) + DB 233,235,254,255,255 ; jmpq 12e3 <_sk_lerp_565_avx+0x14> + DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) + DB 236 ; in (%dx),%al DB 255 ; (bad) - DB 255,226 ; jmpq *%rdx DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) - DB 218,255 ; (bad) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) + DB 255,212 ; callq *%rsp DB 255 ; (bad) - DB 255,202 ; dec %edx DB 255 ; (bad) + DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 190 ; .byte 0xbe + DB 255,192 ; inc %eax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -5622,7 +5631,7 @@ _sk_load_tables_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,26,2,0,0 ; jne 162c <_sk_load_tables_avx+0x228> + DB 15,133,26,2,0,0 ; jne 163c <_sk_load_tables_avx+0x228> DB 196,65,124,16,4,184 ; vmovups (%r8,%rdi,4),%ymm8 DB 85 ; push %rbp DB 65,87 ; push %r15 @@ -5630,7 +5639,7 @@ _sk_load_tables_avx LABEL PROC DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 197,124,40,13,86,68,0,0 ; vmovaps 0x4456(%rip),%ymm9 # 5880 <_sk_callback_avx+0x4ae> + DB 197,124,40,13,102,68,0,0 ; vmovaps 0x4466(%rip),%ymm9 # 58a0 <_sk_callback_avx+0x4be> DB 196,193,60,84,193 ; vandps %ymm9,%ymm8,%ymm0 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 DB 69,137,203 ; mov %r9d,%r11d @@ -5722,7 +5731,7 @@ _sk_load_tables_avx LABEL PROC DB 196,193,97,114,210,24 ; vpsrld $0x18,%xmm10,%xmm3 DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,155,63,0,0 ; vbroadcastss 0x3f9b(%rip),%ymm8 # 55b4 <_sk_callback_avx+0x1e2> + DB 196,98,125,24,5,179,63,0,0 ; vbroadcastss 0x3fb3(%rip),%ymm8 # 55dc <_sk_callback_avx+0x1fa> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -5737,9 +5746,9 @@ _sk_load_tables_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,211,253,255,255 ; ja 1418 <_sk_load_tables_avx+0x14> + DB 15,135,211,253,255,255 ; ja 1428 <_sk_load_tables_avx+0x14> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,140,0,0,0 ; lea 0x8c(%rip),%r10 # 16dc <_sk_load_tables_avx+0x2d8> + DB 76,141,21,140,0,0,0 ; lea 0x8c(%rip),%r10 # 16ec <_sk_load_tables_avx+0x2d8> DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9 DB 77,1,209 ; add %r10,%r9 DB 65,255,225 ; jmpq *%r9 @@ -5762,7 +5771,7 @@ _sk_load_tables_avx LABEL PROC DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8 DB 196,195,57,34,4,184,0 ; vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0 DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8 - DB 233,62,253,255,255 ; jmpq 1418 <_sk_load_tables_avx+0x14> + DB 233,62,253,255,255 ; jmpq 1428 <_sk_load_tables_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 236 ; in (%dx),%al DB 255 ; (bad) @@ -5780,7 +5789,7 @@ _sk_load_tables_avx LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 126,255 ; jle 16f5 <_sk_load_tables_avx+0x2f1> + DB 126,255 ; jle 1705 <_sk_load_tables_avx+0x2f1> DB 255 ; (bad) DB 255 ; .byte 0xff @@ -5790,7 +5799,7 @@ _sk_load_tables_u16_be_avx LABEL PROC 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 197f <_sk_load_tables_u16_be_avx+0x287> + DB 15,133,113,2,0,0 ; jne 198f <_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 @@ -5812,7 +5821,7 @@ _sk_load_tables_u16_be_avx LABEL PROC 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 197,121,111,29,150,65,0,0 ; vmovdqa 0x4196(%rip),%xmm11 # 5900 <_sk_callback_avx+0x52e> + DB 197,121,111,29,166,65,0,0 ; vmovdqa 0x41a6(%rip),%xmm11 # 5920 <_sk_callback_avx+0x53e> DB 196,193,105,219,195 ; vpand %xmm11,%xmm2,%xmm0 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,193,121,105,209 ; vpunpckhwd %xmm9,%xmm0,%xmm2 @@ -5911,7 +5920,7 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,76,60,0,0 ; vbroadcastss 0x3c4c(%rip),%ymm8 # 55b8 <_sk_callback_avx+0x1e6> + DB 196,98,125,24,5,100,60,0,0 ; vbroadcastss 0x3c64(%rip),%ymm8 # 55e0 <_sk_callback_avx+0x1fe> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -5924,29 +5933,29 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 19e5 <_sk_load_tables_u16_be_avx+0x2ed> + DB 116,85 ; je 19f5 <_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 19e5 <_sk_load_tables_u16_be_avx+0x2ed> + DB 114,72 ; jb 19f5 <_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 19f2 <_sk_load_tables_u16_be_avx+0x2fa> + DB 116,72 ; je 1a02 <_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 19f2 <_sk_load_tables_u16_be_avx+0x2fa> + DB 114,59 ; jb 1a02 <_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 1729 <_sk_load_tables_u16_be_avx+0x31> + DB 15,132,97,253,255,255 ; je 1739 <_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 1729 <_sk_load_tables_u16_be_avx+0x31> + DB 15,130,80,253,255,255 ; jb 1739 <_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 1729 <_sk_load_tables_u16_be_avx+0x31> + DB 233,68,253,255,255 ; jmpq 1739 <_sk_load_tables_u16_be_avx+0x31> 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 1729 <_sk_load_tables_u16_be_avx+0x31> + DB 233,55,253,255,255 ; jmpq 1739 <_sk_load_tables_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,46,253,255,255 ; jmpq 1729 <_sk_load_tables_u16_be_avx+0x31> + DB 233,46,253,255,255 ; jmpq 1739 <_sk_load_tables_u16_be_avx+0x31> PUBLIC _sk_load_tables_rgb_u16_be_avx _sk_load_tables_rgb_u16_be_avx LABEL PROC @@ -5954,7 +5963,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC 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 1c6a <_sk_load_tables_rgb_u16_be_avx+0x26f> + DB 15,133,93,2,0,0 ; jne 1c7a <_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 @@ -5981,7 +5990,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC 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 197,121,111,13,143,62,0,0 ; vmovdqa 0x3e8f(%rip),%xmm9 # 5910 <_sk_callback_avx+0x53e> + DB 197,121,111,13,159,62,0,0 ; vmovdqa 0x3e9f(%rip),%xmm9 # 5930 <_sk_callback_avx+0x54e> DB 196,193,113,219,193 ; vpand %xmm9,%xmm1,%xmm0 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 @@ -6073,7 +6082,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC 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 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,94,57,0,0 ; vbroadcastss 0x395e(%rip),%ymm3 # 55bc <_sk_callback_avx+0x1ea> + DB 196,226,125,24,29,118,57,0,0 ; vbroadcastss 0x3976(%rip),%ymm3 # 55e4 <_sk_callback_avx+0x202> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,93 ; pop %r13 @@ -6084,36 +6093,36 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC 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 1c83 <_sk_load_tables_rgb_u16_be_avx+0x288> - DB 233,190,253,255,255 ; jmpq 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 117,5 ; jne 1c93 <_sk_load_tables_rgb_u16_be_avx+0x288> + DB 233,190,253,255,255 ; jmpq 1a51 <_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 1cb2 <_sk_load_tables_rgb_u16_be_avx+0x2b7> + DB 114,26 ; jb 1cc2 <_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 1cb7 <_sk_load_tables_rgb_u16_be_avx+0x2bc> - DB 233,143,253,255,255 ; jmpq 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,138,253,255,255 ; jmpq 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 1cc7 <_sk_load_tables_rgb_u16_be_avx+0x2bc> + DB 233,143,253,255,255 ; jmpq 1a51 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,138,253,255,255 ; jmpq 1a51 <_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 1ce6 <_sk_load_tables_rgb_u16_be_avx+0x2eb> + DB 114,26 ; jb 1cf6 <_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 1ceb <_sk_load_tables_rgb_u16_be_avx+0x2f0> - DB 233,91,253,255,255 ; jmpq 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,86,253,255,255 ; jmpq 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 1cfb <_sk_load_tables_rgb_u16_be_avx+0x2f0> + DB 233,91,253,255,255 ; jmpq 1a51 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,86,253,255,255 ; jmpq 1a51 <_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 1d14 <_sk_load_tables_rgb_u16_be_avx+0x319> + DB 114,20 ; jb 1d24 <_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 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,40,253,255,255 ; jmpq 1a41 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,45,253,255,255 ; jmpq 1a51 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,40,253,255,255 ; jmpq 1a51 <_sk_load_tables_rgb_u16_be_avx+0x46> PUBLIC _sk_byte_tables_avx _sk_byte_tables_avx LABEL PROC @@ -6124,7 +6133,7 @@ _sk_byte_tables_avx LABEL PROC DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,146,56,0,0 ; vbroadcastss 0x3892(%rip),%ymm8 # 55c0 <_sk_callback_avx+0x1ee> + DB 196,98,125,24,5,170,56,0,0 ; vbroadcastss 0x38aa(%rip),%ymm8 # 55e8 <_sk_callback_avx+0x206> 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 @@ -6161,7 +6170,7 @@ _sk_byte_tables_avx LABEL PROC 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 196,98,125,24,13,224,55,0,0 ; vbroadcastss 0x37e0(%rip),%ymm9 # 55c4 <_sk_callback_avx+0x1f2> + DB 196,98,125,24,13,248,55,0,0 ; vbroadcastss 0x37f8(%rip),%ymm9 # 55ec <_sk_callback_avx+0x20a> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 @@ -6321,7 +6330,7 @@ _sk_byte_tables_rgb_avx LABEL PROC 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 196,98,125,24,13,6,53,0,0 ; vbroadcastss 0x3506(%rip),%ymm9 # 55c8 <_sk_callback_avx+0x1f6> + DB 196,98,125,24,13,30,53,0,0 ; vbroadcastss 0x351e(%rip),%ymm9 # 55f0 <_sk_callback_avx+0x20e> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 @@ -6608,36 +6617,36 @@ _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 196,98,125,24,37,100,48,0,0 ; vbroadcastss 0x3064(%rip),%ymm12 # 55cc <_sk_callback_avx+0x1fa> + DB 196,98,125,24,37,124,48,0,0 ; vbroadcastss 0x307c(%rip),%ymm12 # 55f4 <_sk_callback_avx+0x212> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,90,48,0,0 ; vbroadcastss 0x305a(%rip),%ymm12 # 55d0 <_sk_callback_avx+0x1fe> + DB 196,98,125,24,37,114,48,0,0 ; vbroadcastss 0x3072(%rip),%ymm12 # 55f8 <_sk_callback_avx+0x216> DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,80,48,0,0 ; vbroadcastss 0x3050(%rip),%ymm12 # 55d4 <_sk_callback_avx+0x202> + DB 196,98,125,24,37,104,48,0,0 ; vbroadcastss 0x3068(%rip),%ymm12 # 55fc <_sk_callback_avx+0x21a> DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,70,48,0,0 ; vbroadcastss 0x3046(%rip),%ymm12 # 55d8 <_sk_callback_avx+0x206> + DB 196,98,125,24,37,94,48,0,0 ; vbroadcastss 0x305e(%rip),%ymm12 # 5600 <_sk_callback_avx+0x21e> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,60,48,0,0 ; vbroadcastss 0x303c(%rip),%ymm12 # 55dc <_sk_callback_avx+0x20a> + DB 196,98,125,24,37,84,48,0,0 ; vbroadcastss 0x3054(%rip),%ymm12 # 5604 <_sk_callback_avx+0x222> DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,45,48,0,0 ; vbroadcastss 0x302d(%rip),%ymm12 # 55e0 <_sk_callback_avx+0x20e> + DB 196,98,125,24,37,69,48,0,0 ; vbroadcastss 0x3045(%rip),%ymm12 # 5608 <_sk_callback_avx+0x226> DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,35,48,0,0 ; vbroadcastss 0x3023(%rip),%ymm12 # 55e4 <_sk_callback_avx+0x212> + DB 196,98,125,24,37,59,48,0,0 ; vbroadcastss 0x303b(%rip),%ymm12 # 560c <_sk_callback_avx+0x22a> 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 196,98,125,24,29,7,48,0,0 ; vbroadcastss 0x3007(%rip),%ymm11 # 55e8 <_sk_callback_avx+0x216> + DB 196,98,125,24,29,31,48,0,0 ; vbroadcastss 0x301f(%rip),%ymm11 # 5610 <_sk_callback_avx+0x22e> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,253,47,0,0 ; vbroadcastss 0x2ffd(%rip),%ymm11 # 55ec <_sk_callback_avx+0x21a> + DB 196,98,125,24,29,21,48,0,0 ; vbroadcastss 0x3015(%rip),%ymm11 # 5614 <_sk_callback_avx+0x232> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,238,47,0,0 ; vbroadcastss 0x2fee(%rip),%ymm11 # 55f0 <_sk_callback_avx+0x21e> + DB 196,98,125,24,29,6,48,0,0 ; vbroadcastss 0x3006(%rip),%ymm11 # 5618 <_sk_callback_avx+0x236> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,228,47,0,0 ; vbroadcastss 0x2fe4(%rip),%ymm11 # 55f4 <_sk_callback_avx+0x222> + DB 196,98,125,24,29,252,47,0,0 ; vbroadcastss 0x2ffc(%rip),%ymm11 # 561c <_sk_callback_avx+0x23a> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,213,47,0,0 ; vbroadcastss 0x2fd5(%rip),%ymm10 # 55f8 <_sk_callback_avx+0x226> + DB 196,98,125,24,21,237,47,0,0 ; vbroadcastss 0x2fed(%rip),%ymm10 # 5620 <_sk_callback_avx+0x23e> 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 @@ -6645,7 +6654,7 @@ _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 196,98,125,24,5,172,47,0,0 ; vbroadcastss 0x2fac(%rip),%ymm8 # 55fc <_sk_callback_avx+0x22a> + DB 196,98,125,24,5,196,47,0,0 ; vbroadcastss 0x2fc4(%rip),%ymm8 # 5624 <_sk_callback_avx+0x242> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6665,36 +6674,36 @@ _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 196,98,125,24,37,93,47,0,0 ; vbroadcastss 0x2f5d(%rip),%ymm12 # 5600 <_sk_callback_avx+0x22e> + DB 196,98,125,24,37,117,47,0,0 ; vbroadcastss 0x2f75(%rip),%ymm12 # 5628 <_sk_callback_avx+0x246> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,83,47,0,0 ; vbroadcastss 0x2f53(%rip),%ymm12 # 5604 <_sk_callback_avx+0x232> + DB 196,98,125,24,37,107,47,0,0 ; vbroadcastss 0x2f6b(%rip),%ymm12 # 562c <_sk_callback_avx+0x24a> DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,73,47,0,0 ; vbroadcastss 0x2f49(%rip),%ymm12 # 5608 <_sk_callback_avx+0x236> + DB 196,98,125,24,37,97,47,0,0 ; vbroadcastss 0x2f61(%rip),%ymm12 # 5630 <_sk_callback_avx+0x24e> DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,63,47,0,0 ; vbroadcastss 0x2f3f(%rip),%ymm12 # 560c <_sk_callback_avx+0x23a> + DB 196,98,125,24,37,87,47,0,0 ; vbroadcastss 0x2f57(%rip),%ymm12 # 5634 <_sk_callback_avx+0x252> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,53,47,0,0 ; vbroadcastss 0x2f35(%rip),%ymm12 # 5610 <_sk_callback_avx+0x23e> + DB 196,98,125,24,37,77,47,0,0 ; vbroadcastss 0x2f4d(%rip),%ymm12 # 5638 <_sk_callback_avx+0x256> DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,38,47,0,0 ; vbroadcastss 0x2f26(%rip),%ymm12 # 5614 <_sk_callback_avx+0x242> + DB 196,98,125,24,37,62,47,0,0 ; vbroadcastss 0x2f3e(%rip),%ymm12 # 563c <_sk_callback_avx+0x25a> DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,28,47,0,0 ; vbroadcastss 0x2f1c(%rip),%ymm12 # 5618 <_sk_callback_avx+0x246> + DB 196,98,125,24,37,52,47,0,0 ; vbroadcastss 0x2f34(%rip),%ymm12 # 5640 <_sk_callback_avx+0x25e> 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 196,98,125,24,29,0,47,0,0 ; vbroadcastss 0x2f00(%rip),%ymm11 # 561c <_sk_callback_avx+0x24a> + DB 196,98,125,24,29,24,47,0,0 ; vbroadcastss 0x2f18(%rip),%ymm11 # 5644 <_sk_callback_avx+0x262> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,246,46,0,0 ; vbroadcastss 0x2ef6(%rip),%ymm11 # 5620 <_sk_callback_avx+0x24e> + DB 196,98,125,24,29,14,47,0,0 ; vbroadcastss 0x2f0e(%rip),%ymm11 # 5648 <_sk_callback_avx+0x266> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,231,46,0,0 ; vbroadcastss 0x2ee7(%rip),%ymm11 # 5624 <_sk_callback_avx+0x252> + DB 196,98,125,24,29,255,46,0,0 ; vbroadcastss 0x2eff(%rip),%ymm11 # 564c <_sk_callback_avx+0x26a> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,221,46,0,0 ; vbroadcastss 0x2edd(%rip),%ymm11 # 5628 <_sk_callback_avx+0x256> + DB 196,98,125,24,29,245,46,0,0 ; vbroadcastss 0x2ef5(%rip),%ymm11 # 5650 <_sk_callback_avx+0x26e> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,206,46,0,0 ; vbroadcastss 0x2ece(%rip),%ymm10 # 562c <_sk_callback_avx+0x25a> + DB 196,98,125,24,21,230,46,0,0 ; vbroadcastss 0x2ee6(%rip),%ymm10 # 5654 <_sk_callback_avx+0x272> 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 @@ -6702,7 +6711,7 @@ _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 196,98,125,24,5,165,46,0,0 ; vbroadcastss 0x2ea5(%rip),%ymm8 # 5630 <_sk_callback_avx+0x25e> + DB 196,98,125,24,5,189,46,0,0 ; vbroadcastss 0x2ebd(%rip),%ymm8 # 5658 <_sk_callback_avx+0x276> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6722,36 +6731,36 @@ _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 196,98,125,24,37,86,46,0,0 ; vbroadcastss 0x2e56(%rip),%ymm12 # 5634 <_sk_callback_avx+0x262> + DB 196,98,125,24,37,110,46,0,0 ; vbroadcastss 0x2e6e(%rip),%ymm12 # 565c <_sk_callback_avx+0x27a> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,76,46,0,0 ; vbroadcastss 0x2e4c(%rip),%ymm12 # 5638 <_sk_callback_avx+0x266> + DB 196,98,125,24,37,100,46,0,0 ; vbroadcastss 0x2e64(%rip),%ymm12 # 5660 <_sk_callback_avx+0x27e> DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,66,46,0,0 ; vbroadcastss 0x2e42(%rip),%ymm12 # 563c <_sk_callback_avx+0x26a> + DB 196,98,125,24,37,90,46,0,0 ; vbroadcastss 0x2e5a(%rip),%ymm12 # 5664 <_sk_callback_avx+0x282> DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,56,46,0,0 ; vbroadcastss 0x2e38(%rip),%ymm12 # 5640 <_sk_callback_avx+0x26e> + DB 196,98,125,24,37,80,46,0,0 ; vbroadcastss 0x2e50(%rip),%ymm12 # 5668 <_sk_callback_avx+0x286> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,46,46,0,0 ; vbroadcastss 0x2e2e(%rip),%ymm12 # 5644 <_sk_callback_avx+0x272> + DB 196,98,125,24,37,70,46,0,0 ; vbroadcastss 0x2e46(%rip),%ymm12 # 566c <_sk_callback_avx+0x28a> DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,31,46,0,0 ; vbroadcastss 0x2e1f(%rip),%ymm12 # 5648 <_sk_callback_avx+0x276> + DB 196,98,125,24,37,55,46,0,0 ; vbroadcastss 0x2e37(%rip),%ymm12 # 5670 <_sk_callback_avx+0x28e> DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,21,46,0,0 ; vbroadcastss 0x2e15(%rip),%ymm12 # 564c <_sk_callback_avx+0x27a> + DB 196,98,125,24,37,45,46,0,0 ; vbroadcastss 0x2e2d(%rip),%ymm12 # 5674 <_sk_callback_avx+0x292> 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 196,98,125,24,29,249,45,0,0 ; vbroadcastss 0x2df9(%rip),%ymm11 # 5650 <_sk_callback_avx+0x27e> + DB 196,98,125,24,29,17,46,0,0 ; vbroadcastss 0x2e11(%rip),%ymm11 # 5678 <_sk_callback_avx+0x296> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,239,45,0,0 ; vbroadcastss 0x2def(%rip),%ymm11 # 5654 <_sk_callback_avx+0x282> + DB 196,98,125,24,29,7,46,0,0 ; vbroadcastss 0x2e07(%rip),%ymm11 # 567c <_sk_callback_avx+0x29a> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,224,45,0,0 ; vbroadcastss 0x2de0(%rip),%ymm11 # 5658 <_sk_callback_avx+0x286> + DB 196,98,125,24,29,248,45,0,0 ; vbroadcastss 0x2df8(%rip),%ymm11 # 5680 <_sk_callback_avx+0x29e> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,214,45,0,0 ; vbroadcastss 0x2dd6(%rip),%ymm11 # 565c <_sk_callback_avx+0x28a> + DB 196,98,125,24,29,238,45,0,0 ; vbroadcastss 0x2dee(%rip),%ymm11 # 5684 <_sk_callback_avx+0x2a2> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,199,45,0,0 ; vbroadcastss 0x2dc7(%rip),%ymm10 # 5660 <_sk_callback_avx+0x28e> + DB 196,98,125,24,21,223,45,0,0 ; vbroadcastss 0x2ddf(%rip),%ymm10 # 5688 <_sk_callback_avx+0x2a6> 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 @@ -6759,7 +6768,7 @@ _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 196,98,125,24,5,158,45,0,0 ; vbroadcastss 0x2d9e(%rip),%ymm8 # 5664 <_sk_callback_avx+0x292> + DB 196,98,125,24,5,182,45,0,0 ; vbroadcastss 0x2db6(%rip),%ymm8 # 568c <_sk_callback_avx+0x2aa> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6779,36 +6788,36 @@ _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 196,98,125,24,37,79,45,0,0 ; vbroadcastss 0x2d4f(%rip),%ymm12 # 5668 <_sk_callback_avx+0x296> + DB 196,98,125,24,37,103,45,0,0 ; vbroadcastss 0x2d67(%rip),%ymm12 # 5690 <_sk_callback_avx+0x2ae> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,69,45,0,0 ; vbroadcastss 0x2d45(%rip),%ymm12 # 566c <_sk_callback_avx+0x29a> + DB 196,98,125,24,37,93,45,0,0 ; vbroadcastss 0x2d5d(%rip),%ymm12 # 5694 <_sk_callback_avx+0x2b2> DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,59,45,0,0 ; vbroadcastss 0x2d3b(%rip),%ymm12 # 5670 <_sk_callback_avx+0x29e> + DB 196,98,125,24,37,83,45,0,0 ; vbroadcastss 0x2d53(%rip),%ymm12 # 5698 <_sk_callback_avx+0x2b6> DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,49,45,0,0 ; vbroadcastss 0x2d31(%rip),%ymm12 # 5674 <_sk_callback_avx+0x2a2> + DB 196,98,125,24,37,73,45,0,0 ; vbroadcastss 0x2d49(%rip),%ymm12 # 569c <_sk_callback_avx+0x2ba> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,39,45,0,0 ; vbroadcastss 0x2d27(%rip),%ymm12 # 5678 <_sk_callback_avx+0x2a6> + DB 196,98,125,24,37,63,45,0,0 ; vbroadcastss 0x2d3f(%rip),%ymm12 # 56a0 <_sk_callback_avx+0x2be> DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,24,45,0,0 ; vbroadcastss 0x2d18(%rip),%ymm12 # 567c <_sk_callback_avx+0x2aa> + DB 196,98,125,24,37,48,45,0,0 ; vbroadcastss 0x2d30(%rip),%ymm12 # 56a4 <_sk_callback_avx+0x2c2> DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,14,45,0,0 ; vbroadcastss 0x2d0e(%rip),%ymm12 # 5680 <_sk_callback_avx+0x2ae> + DB 196,98,125,24,37,38,45,0,0 ; vbroadcastss 0x2d26(%rip),%ymm12 # 56a8 <_sk_callback_avx+0x2c6> 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 196,98,125,24,29,242,44,0,0 ; vbroadcastss 0x2cf2(%rip),%ymm11 # 5684 <_sk_callback_avx+0x2b2> + DB 196,98,125,24,29,10,45,0,0 ; vbroadcastss 0x2d0a(%rip),%ymm11 # 56ac <_sk_callback_avx+0x2ca> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,232,44,0,0 ; vbroadcastss 0x2ce8(%rip),%ymm11 # 5688 <_sk_callback_avx+0x2b6> + DB 196,98,125,24,29,0,45,0,0 ; vbroadcastss 0x2d00(%rip),%ymm11 # 56b0 <_sk_callback_avx+0x2ce> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,217,44,0,0 ; vbroadcastss 0x2cd9(%rip),%ymm11 # 568c <_sk_callback_avx+0x2ba> + DB 196,98,125,24,29,241,44,0,0 ; vbroadcastss 0x2cf1(%rip),%ymm11 # 56b4 <_sk_callback_avx+0x2d2> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,207,44,0,0 ; vbroadcastss 0x2ccf(%rip),%ymm11 # 5690 <_sk_callback_avx+0x2be> + DB 196,98,125,24,29,231,44,0,0 ; vbroadcastss 0x2ce7(%rip),%ymm11 # 56b8 <_sk_callback_avx+0x2d6> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,192,44,0,0 ; vbroadcastss 0x2cc0(%rip),%ymm10 # 5694 <_sk_callback_avx+0x2c2> + DB 196,98,125,24,21,216,44,0,0 ; vbroadcastss 0x2cd8(%rip),%ymm10 # 56bc <_sk_callback_avx+0x2da> 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 @@ -6816,38 +6825,38 @@ _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 196,98,125,24,5,151,44,0,0 ; vbroadcastss 0x2c97(%rip),%ymm8 # 5698 <_sk_callback_avx+0x2c6> + DB 196,98,125,24,5,175,44,0,0 ; vbroadcastss 0x2caf(%rip),%ymm8 # 56c0 <_sk_callback_avx+0x2de> 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 196,98,125,24,5,137,44,0,0 ; vbroadcastss 0x2c89(%rip),%ymm8 # 569c <_sk_callback_avx+0x2ca> + DB 196,98,125,24,5,161,44,0,0 ; vbroadcastss 0x2ca1(%rip),%ymm8 # 56c4 <_sk_callback_avx+0x2e2> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,127,44,0,0 ; vbroadcastss 0x2c7f(%rip),%ymm8 # 56a0 <_sk_callback_avx+0x2ce> + DB 196,98,125,24,5,151,44,0,0 ; vbroadcastss 0x2c97(%rip),%ymm8 # 56c8 <_sk_callback_avx+0x2e6> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,13,117,44,0,0 ; vbroadcastss 0x2c75(%rip),%ymm9 # 56a4 <_sk_callback_avx+0x2d2> + DB 196,98,125,24,13,141,44,0,0 ; vbroadcastss 0x2c8d(%rip),%ymm9 # 56cc <_sk_callback_avx+0x2ea> 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 196,98,125,24,5,97,44,0,0 ; vbroadcastss 0x2c61(%rip),%ymm8 # 56a8 <_sk_callback_avx+0x2d6> + DB 196,98,125,24,5,121,44,0,0 ; vbroadcastss 0x2c79(%rip),%ymm8 # 56d0 <_sk_callback_avx+0x2ee> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,87,44,0,0 ; vbroadcastss 0x2c57(%rip),%ymm8 # 56ac <_sk_callback_avx+0x2da> + DB 196,98,125,24,5,111,44,0,0 ; vbroadcastss 0x2c6f(%rip),%ymm8 # 56d4 <_sk_callback_avx+0x2f2> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,77,44,0,0 ; vbroadcastss 0x2c4d(%rip),%ymm8 # 56b0 <_sk_callback_avx+0x2de> + DB 196,98,125,24,5,101,44,0,0 ; vbroadcastss 0x2c65(%rip),%ymm8 # 56d8 <_sk_callback_avx+0x2f6> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1 - DB 196,98,125,24,5,63,44,0,0 ; vbroadcastss 0x2c3f(%rip),%ymm8 # 56b4 <_sk_callback_avx+0x2e2> + DB 196,98,125,24,5,87,44,0,0 ; vbroadcastss 0x2c57(%rip),%ymm8 # 56dc <_sk_callback_avx+0x2fa> 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 196,98,125,24,13,40,44,0,0 ; vbroadcastss 0x2c28(%rip),%ymm9 # 56b8 <_sk_callback_avx+0x2e6> + DB 196,98,125,24,13,64,44,0,0 ; vbroadcastss 0x2c40(%rip),%ymm9 # 56e0 <_sk_callback_avx+0x2fe> DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,29,44,0,0 ; vbroadcastss 0x2c1d(%rip),%ymm11 # 56bc <_sk_callback_avx+0x2ea> + DB 196,98,125,24,29,53,44,0,0 ; vbroadcastss 0x2c35(%rip),%ymm11 # 56e4 <_sk_callback_avx+0x302> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,37,19,44,0,0 ; vbroadcastss 0x2c13(%rip),%ymm12 # 56c0 <_sk_callback_avx+0x2ee> + DB 196,98,125,24,37,43,44,0,0 ; vbroadcastss 0x2c2b(%rip),%ymm12 # 56e8 <_sk_callback_avx+0x306> 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 @@ -6862,9 +6871,9 @@ _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 196,226,125,24,5,201,43,0,0 ; vbroadcastss 0x2bc9(%rip),%ymm0 # 56c4 <_sk_callback_avx+0x2f2> + DB 196,226,125,24,5,225,43,0,0 ; vbroadcastss 0x2be1(%rip),%ymm0 # 56ec <_sk_callback_avx+0x30a> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,192,43,0,0 ; vbroadcastss 0x2bc0(%rip),%ymm8 # 56c8 <_sk_callback_avx+0x2f6> + DB 196,98,125,24,5,216,43,0,0 ; vbroadcastss 0x2bd8(%rip),%ymm8 # 56f0 <_sk_callback_avx+0x30e> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6876,14 +6885,14 @@ _sk_load_a8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,62 ; jne 2b5f <_sk_load_a8_avx+0x4e> + DB 117,62 ; jne 2b6f <_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 196,226,125,24,13,132,43,0,0 ; vbroadcastss 0x2b84(%rip),%ymm1 # 56cc <_sk_callback_avx+0x2fa> + DB 196,226,125,24,13,156,43,0,0 ; vbroadcastss 0x2b9c(%rip),%ymm1 # 56f4 <_sk_callback_avx+0x312> 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 @@ -6900,9 +6909,9 @@ _sk_load_a8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2b67 <_sk_load_a8_avx+0x56> + DB 117,234 ; jne 2b77 <_sk_load_a8_avx+0x56> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,161 ; jmp 2b25 <_sk_load_a8_avx+0x14> + DB 235,161 ; jmp 2b35 <_sk_load_a8_avx+0x14> PUBLIC _sk_gather_a8_avx _sk_gather_a8_avx LABEL PROC @@ -6950,7 +6959,7 @@ _sk_gather_a8_avx LABEL PROC 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 196,226,125,24,13,121,42,0,0 ; vbroadcastss 0x2a79(%rip),%ymm1 # 56d0 <_sk_callback_avx+0x2fe> + DB 196,226,125,24,13,145,42,0,0 ; vbroadcastss 0x2a91(%rip),%ymm1 # 56f8 <_sk_callback_avx+0x316> 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 @@ -6966,14 +6975,14 @@ 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 196,98,125,24,5,84,42,0,0 ; vbroadcastss 0x2a54(%rip),%ymm8 # 56d4 <_sk_callback_avx+0x302> + DB 196,98,125,24,5,108,42,0,0 ; vbroadcastss 0x2a6c(%rip),%ymm8 # 56fc <_sk_callback_avx+0x31a> 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 2ca9 <_sk_store_a8_avx+0x37> + DB 117,10 ; jne 2cb9 <_sk_store_a8_avx+0x37> DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6981,10 +6990,10 @@ _sk_store_a8_avx LABEL PROC 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 2ca5 <_sk_store_a8_avx+0x33> + DB 119,236 ; ja 2cb5 <_sk_store_a8_avx+0x33> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 2d0c <_sk_store_a8_avx+0x9a> + DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 2d1c <_sk_store_a8_avx+0x9a> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -6995,7 +7004,7 @@ _sk_store_a8_avx LABEL PROC 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 2ca5 <_sk_store_a8_avx+0x33> + DB 235,154 ; jmp 2cb5 <_sk_store_a8_avx+0x33> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -7027,17 +7036,17 @@ _sk_load_g8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,67 ; jne 2d7b <_sk_load_g8_avx+0x53> + DB 117,67 ; jne 2d8b <_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 196,226,125,24,13,121,41,0,0 ; vbroadcastss 0x2979(%rip),%ymm1 # 56d8 <_sk_callback_avx+0x306> + DB 196,226,125,24,13,145,41,0,0 ; vbroadcastss 0x2991(%rip),%ymm1 # 5700 <_sk_callback_avx+0x31e> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,110,41,0,0 ; vbroadcastss 0x296e(%rip),%ymm3 # 56dc <_sk_callback_avx+0x30a> + DB 196,226,125,24,29,134,41,0,0 ; vbroadcastss 0x2986(%rip),%ymm3 # 5704 <_sk_callback_avx+0x322> DB 76,137,193 ; mov %r8,%rcx DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 @@ -7051,9 +7060,9 @@ _sk_load_g8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2d83 <_sk_load_g8_avx+0x5b> + DB 117,234 ; jne 2d93 <_sk_load_g8_avx+0x5b> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,156 ; jmp 2d3c <_sk_load_g8_avx+0x14> + DB 235,156 ; jmp 2d4c <_sk_load_g8_avx+0x14> PUBLIC _sk_gather_g8_avx _sk_gather_g8_avx LABEL PROC @@ -7101,10 +7110,10 @@ _sk_gather_g8_avx LABEL PROC 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 196,226,125,24,13,109,40,0,0 ; vbroadcastss 0x286d(%rip),%ymm1 # 56e0 <_sk_callback_avx+0x30e> + DB 196,226,125,24,13,133,40,0,0 ; vbroadcastss 0x2885(%rip),%ymm1 # 5708 <_sk_callback_avx+0x326> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,98,40,0,0 ; vbroadcastss 0x2862(%rip),%ymm3 # 56e4 <_sk_callback_avx+0x312> + DB 196,226,125,24,29,122,40,0,0 ; vbroadcastss 0x287a(%rip),%ymm3 # 570c <_sk_callback_avx+0x32a> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx @@ -7118,9 +7127,9 @@ _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 2ea2 <_sk_gather_i8_avx+0xf> + DB 116,5 ; je 2eb2 <_sk_gather_i8_avx+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 2ea4 <_sk_gather_i8_avx+0x11> + DB 235,2 ; jmp 2eb4 <_sk_gather_i8_avx+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -7182,10 +7191,10 @@ _sk_gather_i8_avx LABEL PROC 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 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,210,40,0,0 ; vmovaps 0x28d2(%rip),%ymm10 # 58a0 <_sk_callback_avx+0x4ce> + DB 197,124,40,21,226,40,0,0 ; vmovaps 0x28e2(%rip),%ymm10 # 58c0 <_sk_callback_avx+0x4de> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,8,39,0,0 ; vbroadcastss 0x2708(%rip),%ymm9 # 56e8 <_sk_callback_avx+0x316> + DB 196,98,125,24,13,32,39,0,0 ; vbroadcastss 0x2720(%rip),%ymm9 # 5710 <_sk_callback_avx+0x32e> 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 @@ -7217,38 +7226,38 @@ _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 30d8 <_sk_load_565_avx+0x8e> + DB 15,133,128,0,0,0 ; jne 30e8 <_sk_load_565_avx+0x8e> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,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 196,226,125,24,5,114,38,0,0 ; vbroadcastss 0x2672(%rip),%ymm0 # 56ec <_sk_callback_avx+0x31a> + DB 196,226,125,24,5,138,38,0,0 ; vbroadcastss 0x268a(%rip),%ymm0 # 5714 <_sk_callback_avx+0x332> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,101,38,0,0 ; vbroadcastss 0x2665(%rip),%ymm1 # 56f0 <_sk_callback_avx+0x31e> + DB 196,226,125,24,13,125,38,0,0 ; vbroadcastss 0x267d(%rip),%ymm1 # 5718 <_sk_callback_avx+0x336> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,92,38,0,0 ; vbroadcastss 0x265c(%rip),%ymm1 # 56f4 <_sk_callback_avx+0x322> + DB 196,226,125,24,13,116,38,0,0 ; vbroadcastss 0x2674(%rip),%ymm1 # 571c <_sk_callback_avx+0x33a> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,79,38,0,0 ; vbroadcastss 0x264f(%rip),%ymm3 # 56f8 <_sk_callback_avx+0x326> + DB 196,226,125,24,29,103,38,0,0 ; vbroadcastss 0x2667(%rip),%ymm3 # 5720 <_sk_callback_avx+0x33e> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,70,38,0,0 ; vbroadcastss 0x2646(%rip),%ymm3 # 56fc <_sk_callback_avx+0x32a> + DB 196,226,125,24,29,94,38,0,0 ; vbroadcastss 0x265e(%rip),%ymm3 # 5724 <_sk_callback_avx+0x342> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,57,38,0,0 ; vbroadcastss 0x2639(%rip),%ymm3 # 5700 <_sk_callback_avx+0x32e> + DB 196,226,125,24,29,81,38,0,0 ; vbroadcastss 0x2651(%rip),%ymm3 # 5728 <_sk_callback_avx+0x346> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,46,38,0,0 ; vbroadcastss 0x262e(%rip),%ymm3 # 5704 <_sk_callback_avx+0x332> + DB 196,226,125,24,29,70,38,0,0 ; vbroadcastss 0x2646(%rip),%ymm3 # 572c <_sk_callback_avx+0x34a> DB 255,224 ; jmpq *%rax DB 65,137,200 ; mov %ecx,%r8d DB 65,128,224,7 ; and $0x7,%r8b 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 305e <_sk_load_565_avx+0x14> + DB 15,135,110,255,255,255 ; ja 306e <_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 # 3144 <_sk_load_565_avx+0xfa> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 3154 <_sk_load_565_avx+0xfa> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7260,7 +7269,7 @@ _sk_load_565_avx LABEL PROC 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 305e <_sk_load_565_avx+0x14> + DB 233,26,255,255,255 ; jmpq 306e <_sk_load_565_avx+0x14> DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) @@ -7336,23 +7345,23 @@ _sk_gather_565_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,206,36,0,0 ; vbroadcastss 0x24ce(%rip),%ymm0 # 5708 <_sk_callback_avx+0x336> + DB 196,226,125,24,5,230,36,0,0 ; vbroadcastss 0x24e6(%rip),%ymm0 # 5730 <_sk_callback_avx+0x34e> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,193,36,0,0 ; vbroadcastss 0x24c1(%rip),%ymm1 # 570c <_sk_callback_avx+0x33a> + DB 196,226,125,24,13,217,36,0,0 ; vbroadcastss 0x24d9(%rip),%ymm1 # 5734 <_sk_callback_avx+0x352> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,184,36,0,0 ; vbroadcastss 0x24b8(%rip),%ymm1 # 5710 <_sk_callback_avx+0x33e> + DB 196,226,125,24,13,208,36,0,0 ; vbroadcastss 0x24d0(%rip),%ymm1 # 5738 <_sk_callback_avx+0x356> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,171,36,0,0 ; vbroadcastss 0x24ab(%rip),%ymm3 # 5714 <_sk_callback_avx+0x342> + DB 196,226,125,24,29,195,36,0,0 ; vbroadcastss 0x24c3(%rip),%ymm3 # 573c <_sk_callback_avx+0x35a> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,162,36,0,0 ; vbroadcastss 0x24a2(%rip),%ymm3 # 5718 <_sk_callback_avx+0x346> + DB 196,226,125,24,29,186,36,0,0 ; vbroadcastss 0x24ba(%rip),%ymm3 # 5740 <_sk_callback_avx+0x35e> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,149,36,0,0 ; vbroadcastss 0x2495(%rip),%ymm3 # 571c <_sk_callback_avx+0x34a> + DB 196,226,125,24,29,173,36,0,0 ; vbroadcastss 0x24ad(%rip),%ymm3 # 5744 <_sk_callback_avx+0x362> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,138,36,0,0 ; vbroadcastss 0x248a(%rip),%ymm3 # 5720 <_sk_callback_avx+0x34e> + DB 196,226,125,24,29,162,36,0,0 ; vbroadcastss 0x24a2(%rip),%ymm3 # 5748 <_sk_callback_avx+0x366> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 @@ -7364,14 +7373,14 @@ 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 196,98,125,24,5,118,36,0,0 ; vbroadcastss 0x2476(%rip),%ymm8 # 5724 <_sk_callback_avx+0x352> + DB 196,98,125,24,5,142,36,0,0 ; vbroadcastss 0x248e(%rip),%ymm8 # 574c <_sk_callback_avx+0x36a> 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 196,98,125,24,21,79,36,0,0 ; vbroadcastss 0x244f(%rip),%ymm10 # 5728 <_sk_callback_avx+0x356> + DB 196,98,125,24,21,103,36,0,0 ; vbroadcastss 0x2467(%rip),%ymm10 # 5750 <_sk_callback_avx+0x36e> 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 @@ -7385,7 +7394,7 @@ _sk_store_565_avx LABEL PROC 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 3329 <_sk_store_565_avx+0x89> + DB 117,10 ; jne 3339 <_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 @@ -7393,9 +7402,9 @@ _sk_store_565_avx LABEL PROC 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 3325 <_sk_store_565_avx+0x85> + DB 119,236 ; ja 3335 <_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 # 3388 <_sk_store_565_avx+0xe8> + DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3398 <_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 @@ -7406,7 +7415,7 @@ _sk_store_565_avx LABEL PROC 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 3325 <_sk_store_565_avx+0x85> + DB 235,159 ; jmp 3335 <_sk_store_565_avx+0x85> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -7437,31 +7446,31 @@ _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 344a <_sk_load_4444_avx+0xa6> + DB 15,133,152,0,0,0 ; jne 345a <_sk_load_4444_avx+0xa6> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,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 196,226,125,24,5,88,35,0,0 ; vbroadcastss 0x2358(%rip),%ymm0 # 572c <_sk_callback_avx+0x35a> + DB 196,226,125,24,5,112,35,0,0 ; vbroadcastss 0x2370(%rip),%ymm0 # 5754 <_sk_callback_avx+0x372> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,75,35,0,0 ; vbroadcastss 0x234b(%rip),%ymm1 # 5730 <_sk_callback_avx+0x35e> + DB 196,226,125,24,13,99,35,0,0 ; vbroadcastss 0x2363(%rip),%ymm1 # 5758 <_sk_callback_avx+0x376> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,66,35,0,0 ; vbroadcastss 0x2342(%rip),%ymm1 # 5734 <_sk_callback_avx+0x362> + DB 196,226,125,24,13,90,35,0,0 ; vbroadcastss 0x235a(%rip),%ymm1 # 575c <_sk_callback_avx+0x37a> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,53,35,0,0 ; vbroadcastss 0x2335(%rip),%ymm2 # 5738 <_sk_callback_avx+0x366> + DB 196,226,125,24,21,77,35,0,0 ; vbroadcastss 0x234d(%rip),%ymm2 # 5760 <_sk_callback_avx+0x37e> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,44,35,0,0 ; vbroadcastss 0x232c(%rip),%ymm2 # 573c <_sk_callback_avx+0x36a> + DB 196,226,125,24,21,68,35,0,0 ; vbroadcastss 0x2344(%rip),%ymm2 # 5764 <_sk_callback_avx+0x382> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,31,35,0,0 ; vbroadcastss 0x231f(%rip),%ymm8 # 5740 <_sk_callback_avx+0x36e> + DB 196,98,125,24,5,55,35,0,0 ; vbroadcastss 0x2337(%rip),%ymm8 # 5768 <_sk_callback_avx+0x386> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,21,35,0,0 ; vbroadcastss 0x2315(%rip),%ymm8 # 5744 <_sk_callback_avx+0x372> + DB 196,98,125,24,5,45,35,0,0 ; vbroadcastss 0x232d(%rip),%ymm8 # 576c <_sk_callback_avx+0x38a> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,7,35,0,0 ; vbroadcastss 0x2307(%rip),%ymm8 # 5748 <_sk_callback_avx+0x376> + DB 196,98,125,24,5,31,35,0,0 ; vbroadcastss 0x231f(%rip),%ymm8 # 5770 <_sk_callback_avx+0x38e> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -7470,9 +7479,9 @@ _sk_load_4444_avx LABEL PROC 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 33b8 <_sk_load_4444_avx+0x14> + DB 15,135,86,255,255,255 ; ja 33c8 <_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 # 34b8 <_sk_load_4444_avx+0x114> + DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 34c8 <_sk_load_4444_avx+0x114> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7484,7 +7493,7 @@ _sk_load_4444_avx LABEL PROC 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 33b8 <_sk_load_4444_avx+0x14> + DB 233,2,255,255,255 ; jmpq 33c8 <_sk_load_4444_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 242,255 ; repnz (bad) DB 255 ; (bad) @@ -7561,25 +7570,25 @@ _sk_gather_4444_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,158,33,0,0 ; vbroadcastss 0x219e(%rip),%ymm0 # 574c <_sk_callback_avx+0x37a> + DB 196,226,125,24,5,182,33,0,0 ; vbroadcastss 0x21b6(%rip),%ymm0 # 5774 <_sk_callback_avx+0x392> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,145,33,0,0 ; vbroadcastss 0x2191(%rip),%ymm1 # 5750 <_sk_callback_avx+0x37e> + DB 196,226,125,24,13,169,33,0,0 ; vbroadcastss 0x21a9(%rip),%ymm1 # 5778 <_sk_callback_avx+0x396> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,136,33,0,0 ; vbroadcastss 0x2188(%rip),%ymm1 # 5754 <_sk_callback_avx+0x382> + DB 196,226,125,24,13,160,33,0,0 ; vbroadcastss 0x21a0(%rip),%ymm1 # 577c <_sk_callback_avx+0x39a> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,123,33,0,0 ; vbroadcastss 0x217b(%rip),%ymm2 # 5758 <_sk_callback_avx+0x386> + DB 196,226,125,24,21,147,33,0,0 ; vbroadcastss 0x2193(%rip),%ymm2 # 5780 <_sk_callback_avx+0x39e> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,114,33,0,0 ; vbroadcastss 0x2172(%rip),%ymm2 # 575c <_sk_callback_avx+0x38a> + DB 196,226,125,24,21,138,33,0,0 ; vbroadcastss 0x218a(%rip),%ymm2 # 5784 <_sk_callback_avx+0x3a2> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,101,33,0,0 ; vbroadcastss 0x2165(%rip),%ymm8 # 5760 <_sk_callback_avx+0x38e> + DB 196,98,125,24,5,125,33,0,0 ; vbroadcastss 0x217d(%rip),%ymm8 # 5788 <_sk_callback_avx+0x3a6> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,91,33,0,0 ; vbroadcastss 0x215b(%rip),%ymm8 # 5764 <_sk_callback_avx+0x392> + DB 196,98,125,24,5,115,33,0,0 ; vbroadcastss 0x2173(%rip),%ymm8 # 578c <_sk_callback_avx+0x3aa> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,77,33,0,0 ; vbroadcastss 0x214d(%rip),%ymm8 # 5768 <_sk_callback_avx+0x396> + DB 196,98,125,24,5,101,33,0,0 ; vbroadcastss 0x2165(%rip),%ymm8 # 5790 <_sk_callback_avx+0x3ae> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -7593,7 +7602,7 @@ 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 196,98,125,24,5,50,33,0,0 ; vbroadcastss 0x2132(%rip),%ymm8 # 576c <_sk_callback_avx+0x39a> + DB 196,98,125,24,5,74,33,0,0 ; vbroadcastss 0x214a(%rip),%ymm8 # 5794 <_sk_callback_avx+0x3b2> 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 @@ -7620,7 +7629,7 @@ _sk_store_4444_avx LABEL PROC 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 36d3 <_sk_store_4444_avx+0xa7> + DB 117,10 ; jne 36e3 <_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 @@ -7628,9 +7637,9 @@ _sk_store_4444_avx LABEL PROC 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 36cf <_sk_store_4444_avx+0xa3> + DB 119,236 ; ja 36df <_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 # 3730 <_sk_store_4444_avx+0x104> + DB 76,141,13,66,0,0,0 ; lea 0x42(%rip),%r9 # 3740 <_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 @@ -7641,7 +7650,7 @@ _sk_store_4444_avx LABEL PROC 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 36cf <_sk_store_4444_avx+0xa3> + DB 235,159 ; jmp 36df <_sk_store_4444_avx+0xa3> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -7670,12 +7679,12 @@ _sk_load_8888_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 15,133,135,0,0,0 ; jne 37e1 <_sk_load_8888_avx+0x95> + DB 15,133,135,0,0,0 ; jne 37f1 <_sk_load_8888_avx+0x95> DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9 - DB 197,124,40,21,88,33,0,0 ; vmovaps 0x2158(%rip),%ymm10 # 58c0 <_sk_callback_avx+0x4ee> + DB 197,124,40,21,104,33,0,0 ; vmovaps 0x2168(%rip),%ymm10 # 58e0 <_sk_callback_avx+0x4fe> DB 196,193,52,84,194 ; vandps %ymm10,%ymm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,246,31,0,0 ; vbroadcastss 0x1ff6(%rip),%ymm8 # 5770 <_sk_callback_avx+0x39e> + DB 196,98,125,24,5,14,32,0,0 ; vbroadcastss 0x200e(%rip),%ymm8 # 5798 <_sk_callback_avx+0x3b6> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,193,113,114,209,8 ; vpsrld $0x8,%xmm9,%xmm1 DB 196,99,125,25,203,1 ; vextractf128 $0x1,%ymm9,%xmm3 @@ -7702,9 +7711,9 @@ _sk_load_8888_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,102,255,255,255 ; ja 3760 <_sk_load_8888_avx+0x14> + DB 15,135,102,255,255,255 ; ja 3770 <_sk_load_8888_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,139,0,0,0 ; lea 0x8b(%rip),%r9 # 3890 <_sk_load_8888_avx+0x144> + DB 76,141,13,139,0,0,0 ; lea 0x8b(%rip),%r9 # 38a0 <_sk_load_8888_avx+0x144> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7727,7 +7736,7 @@ _sk_load_8888_avx LABEL PROC DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 DB 196,195,49,34,4,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 - DB 233,210,254,255,255 ; jmpq 3760 <_sk_load_8888_avx+0x14> + DB 233,210,254,255,255 ; jmpq 3770 <_sk_load_8888_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 236 ; in (%dx),%al DB 255 ; (bad) @@ -7745,7 +7754,7 @@ _sk_load_8888_avx LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 126,255 ; jle 38a9 <_sk_load_8888_avx+0x15d> + DB 126,255 ; jle 38b9 <_sk_load_8888_avx+0x15d> DB 255 ; (bad) DB 255 ; .byte 0xff @@ -7788,10 +7797,10 @@ _sk_gather_8888_avx LABEL PROC 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 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,130,31,0,0 ; vmovaps 0x1f82(%rip),%ymm10 # 58e0 <_sk_callback_avx+0x50e> + DB 197,124,40,21,146,31,0,0 ; vmovaps 0x1f92(%rip),%ymm10 # 5900 <_sk_callback_avx+0x51e> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,4,30,0,0 ; vbroadcastss 0x1e04(%rip),%ymm9 # 5774 <_sk_callback_avx+0x3a2> + DB 196,98,125,24,13,28,30,0,0 ; vbroadcastss 0x1e1c(%rip),%ymm9 # 579c <_sk_callback_avx+0x3ba> 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 @@ -7821,7 +7830,7 @@ PUBLIC _sk_store_8888_avx _sk_store_8888_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,146,29,0,0 ; vbroadcastss 0x1d92(%rip),%ymm8 # 5778 <_sk_callback_avx+0x3a6> + DB 196,98,125,24,5,170,29,0,0 ; vbroadcastss 0x1daa(%rip),%ymm8 # 57a0 <_sk_callback_avx+0x3be> 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 @@ -7846,7 +7855,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3a74 <_sk_store_8888_avx+0x9c> + DB 117,10 ; jne 3a84 <_sk_store_8888_avx+0x9c> DB 196,65,124,17,4,186 ; vmovups %ymm8,(%r10,%rdi,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -7854,9 +7863,9 @@ _sk_store_8888_avx LABEL PROC 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 3a70 <_sk_store_8888_avx+0x98> + DB 119,236 ; ja 3a80 <_sk_store_8888_avx+0x98> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,85,0,0,0 ; lea 0x55(%rip),%r9 # 3ae4 <_sk_store_8888_avx+0x10c> + DB 76,141,13,85,0,0,0 ; lea 0x55(%rip),%r9 # 3af4 <_sk_store_8888_avx+0x10c> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7870,7 +7879,7 @@ _sk_store_8888_avx LABEL PROC DB 196,67,121,22,68,186,8,2 ; vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4) DB 196,67,121,22,68,186,4,1 ; vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4) DB 196,65,121,126,4,186 ; vmovd %xmm8,(%r10,%rdi,4) - DB 235,143 ; jmp 3a70 <_sk_store_8888_avx+0x98> + DB 235,143 ; jmp 3a80 <_sk_store_8888_avx+0x98> DB 15,31,0 ; nopl (%rax) DB 245 ; cmc DB 255 ; (bad) @@ -7906,7 +7915,7 @@ _sk_load_f16_avx LABEL PROC 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 3dbb <_sk_load_f16_avx+0x2bb> + DB 15,133,143,2,0,0 ; jne 3dcb <_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 @@ -7924,13 +7933,13 @@ _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 196,98,125,24,37,247,27,0,0 ; vbroadcastss 0x1bf7(%rip),%ymm12 # 577c <_sk_callback_avx+0x3aa> + DB 196,98,125,24,37,15,28,0,0 ; vbroadcastss 0x1c0f(%rip),%ymm12 # 57a4 <_sk_callback_avx+0x3c2> 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 196,98,121,24,29,227,27,0,0 ; vbroadcastss 0x1be3(%rip),%xmm11 # 5780 <_sk_callback_avx+0x3ae> + DB 196,98,121,24,29,251,27,0,0 ; vbroadcastss 0x1bfb(%rip),%xmm11 # 57a8 <_sk_callback_avx+0x3c6> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,217,27,0,0 ; vbroadcastss 0x1bd9(%rip),%xmm13 # 5784 <_sk_callback_avx+0x3b2> + DB 196,98,121,24,45,241,27,0,0 ; vbroadcastss 0x1bf1(%rip),%xmm13 # 57ac <_sk_callback_avx+0x3ca> 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 @@ -7944,7 +7953,7 @@ _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 196,226,121,24,29,143,27,0,0 ; vbroadcastss 0x1b8f(%rip),%xmm3 # 5788 <_sk_callback_avx+0x3b6> + DB 196,226,121,24,29,167,27,0,0 ; vbroadcastss 0x1ba7(%rip),%xmm3 # 57b0 <_sk_callback_avx+0x3ce> 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 @@ -8037,29 +8046,29 @@ _sk_load_f16_avx LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 3e1a <_sk_load_f16_avx+0x31a> + DB 116,79 ; je 3e2a <_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 3e1a <_sk_load_f16_avx+0x31a> + DB 114,67 ; jb 3e2a <_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 3e27 <_sk_load_f16_avx+0x327> + DB 116,68 ; je 3e37 <_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 3e27 <_sk_load_f16_avx+0x327> + DB 114,56 ; jb 3e37 <_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 3b43 <_sk_load_f16_avx+0x43> + DB 15,132,68,253,255,255 ; je 3b53 <_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 3b43 <_sk_load_f16_avx+0x43> + DB 15,130,52,253,255,255 ; jb 3b53 <_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 3b43 <_sk_load_f16_avx+0x43> + DB 233,41,253,255,255 ; jmpq 3b53 <_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 3b43 <_sk_load_f16_avx+0x43> + DB 233,28,253,255,255 ; jmpq 3b53 <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 - DB 233,19,253,255,255 ; jmpq 3b43 <_sk_load_f16_avx+0x43> + DB 233,19,253,255,255 ; jmpq 3b53 <_sk_load_f16_avx+0x43> PUBLIC _sk_gather_f16_avx _sk_gather_f16_avx LABEL PROC @@ -8121,13 +8130,13 @@ _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 196,98,125,24,37,79,24,0,0 ; vbroadcastss 0x184f(%rip),%ymm12 # 578c <_sk_callback_avx+0x3ba> + DB 196,98,125,24,37,103,24,0,0 ; vbroadcastss 0x1867(%rip),%ymm12 # 57b4 <_sk_callback_avx+0x3d2> 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 196,98,121,24,29,59,24,0,0 ; vbroadcastss 0x183b(%rip),%xmm11 # 5790 <_sk_callback_avx+0x3be> + DB 196,98,121,24,29,83,24,0,0 ; vbroadcastss 0x1853(%rip),%xmm11 # 57b8 <_sk_callback_avx+0x3d6> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,49,24,0,0 ; vbroadcastss 0x1831(%rip),%xmm13 # 5794 <_sk_callback_avx+0x3c2> + DB 196,98,121,24,45,73,24,0,0 ; vbroadcastss 0x1849(%rip),%xmm13 # 57bc <_sk_callback_avx+0x3da> 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 @@ -8141,7 +8150,7 @@ _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 196,226,121,24,29,231,23,0,0 ; vbroadcastss 0x17e7(%rip),%xmm3 # 5798 <_sk_callback_avx+0x3c6> + DB 196,226,121,24,29,255,23,0,0 ; vbroadcastss 0x17ff(%rip),%xmm3 # 57c0 <_sk_callback_avx+0x3de> 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 @@ -8243,12 +8252,12 @@ _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 196,98,125,24,13,244,21,0,0 ; vbroadcastss 0x15f4(%rip),%ymm9 # 579c <_sk_callback_avx+0x3ca> + DB 196,98,125,24,13,12,22,0,0 ; vbroadcastss 0x160c(%rip),%ymm9 # 57c4 <_sk_callback_avx+0x3e2> 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 196,98,121,24,5,218,21,0,0 ; vbroadcastss 0x15da(%rip),%xmm8 # 57a0 <_sk_callback_avx+0x3ce> + DB 196,98,121,24,5,242,21,0,0 ; vbroadcastss 0x15f2(%rip),%xmm8 # 57c8 <_sk_callback_avx+0x3e6> 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 @@ -8258,7 +8267,7 @@ _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 196,98,125,24,21,161,21,0,0 ; vbroadcastss 0x15a1(%rip),%ymm10 # 57a4 <_sk_callback_avx+0x3d2> + DB 196,98,125,24,21,185,21,0,0 ; vbroadcastss 0x15b9(%rip),%ymm10 # 57cc <_sk_callback_avx+0x3ea> 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 @@ -8340,7 +8349,7 @@ _sk_store_f16_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 117,75 ; jne 43ea <_sk_store_f16_avx+0x270> + DB 117,75 ; jne 43fa <_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) @@ -8356,22 +8365,22 @@ _sk_store_f16_avx LABEL PROC 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 43b6 <_sk_store_f16_avx+0x23c> + DB 116,193 ; je 43c6 <_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 43b6 <_sk_store_f16_avx+0x23c> + DB 114,181 ; jb 43c6 <_sk_store_f16_avx+0x23c> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,173 ; je 43b6 <_sk_store_f16_avx+0x23c> + DB 116,173 ; je 43c6 <_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 43b6 <_sk_store_f16_avx+0x23c> + DB 114,161 ; jb 43c6 <_sk_store_f16_avx+0x23c> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,153 ; je 43b6 <_sk_store_f16_avx+0x23c> + DB 116,153 ; je 43c6 <_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 43b6 <_sk_store_f16_avx+0x23c> + DB 114,141 ; jb 43c6 <_sk_store_f16_avx+0x23c> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,133 ; jmp 43b6 <_sk_store_f16_avx+0x23c> + DB 235,133 ; jmp 43c6 <_sk_store_f16_avx+0x23c> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC @@ -8379,7 +8388,7 @@ _sk_load_u16_be_avx LABEL PROC 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 4544 <_sk_load_u16_be_avx+0x113> + DB 15,133,253,0,0,0 ; jne 4554 <_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 @@ -8401,7 +8410,7 @@ _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 196,98,125,24,29,240,18,0,0 ; vbroadcastss 0x12f0(%rip),%ymm11 # 57a8 <_sk_callback_avx+0x3d6> + DB 196,98,125,24,29,8,19,0,0 ; vbroadcastss 0x1308(%rip),%ymm11 # 57d0 <_sk_callback_avx+0x3ee> 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 @@ -8435,29 +8444,29 @@ _sk_load_u16_be_avx LABEL PROC DB 196,65,123,16,4,64 ; vmovsd (%r8,%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 45aa <_sk_load_u16_be_avx+0x179> + DB 116,85 ; je 45ba <_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 45aa <_sk_load_u16_be_avx+0x179> + DB 114,72 ; jb 45ba <_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 45b7 <_sk_load_u16_be_avx+0x186> + DB 116,72 ; je 45c7 <_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 45b7 <_sk_load_u16_be_avx+0x186> + DB 114,59 ; jb 45c7 <_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 4462 <_sk_load_u16_be_avx+0x31> + DB 15,132,213,254,255,255 ; je 4472 <_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 4462 <_sk_load_u16_be_avx+0x31> + DB 15,130,196,254,255,255 ; jb 4472 <_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 4462 <_sk_load_u16_be_avx+0x31> + DB 233,184,254,255,255 ; jmpq 4472 <_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 4462 <_sk_load_u16_be_avx+0x31> + DB 233,171,254,255,255 ; jmpq 4472 <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,162,254,255,255 ; jmpq 4462 <_sk_load_u16_be_avx+0x31> + DB 233,162,254,255,255 ; jmpq 4472 <_sk_load_u16_be_avx+0x31> PUBLIC _sk_load_rgb_u16_be_avx _sk_load_rgb_u16_be_avx LABEL PROC @@ -8465,7 +8474,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC 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 46c5 <_sk_load_rgb_u16_be_avx+0x105> + DB 15,133,243,0,0,0 ; jne 46d5 <_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 @@ -8492,7 +8501,7 @@ _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 196,98,125,24,29,80,17,0,0 ; vbroadcastss 0x1150(%rip),%ymm11 # 57ac <_sk_callback_avx+0x3da> + DB 196,98,125,24,29,104,17,0,0 ; vbroadcastss 0x1168(%rip),%ymm11 # 57d4 <_sk_callback_avx+0x3f2> 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 @@ -8513,48 +8522,48 @@ _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 196,226,125,24,29,237,16,0,0 ; vbroadcastss 0x10ed(%rip),%ymm3 # 57b0 <_sk_callback_avx+0x3de> + DB 196,226,125,24,29,5,17,0,0 ; vbroadcastss 0x1105(%rip),%ymm3 # 57d8 <_sk_callback_avx+0x3f6> 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 46de <_sk_load_rgb_u16_be_avx+0x11e> - DB 233,40,255,255,255 ; jmpq 4606 <_sk_load_rgb_u16_be_avx+0x46> + DB 117,5 ; jne 46ee <_sk_load_rgb_u16_be_avx+0x11e> + DB 233,40,255,255,255 ; jmpq 4616 <_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 470d <_sk_load_rgb_u16_be_avx+0x14d> + DB 114,26 ; jb 471d <_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 4712 <_sk_load_rgb_u16_be_avx+0x152> - DB 233,249,254,255,255 ; jmpq 4606 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,244,254,255,255 ; jmpq 4606 <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 4722 <_sk_load_rgb_u16_be_avx+0x152> + DB 233,249,254,255,255 ; jmpq 4616 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,244,254,255,255 ; jmpq 4616 <_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 4741 <_sk_load_rgb_u16_be_avx+0x181> + DB 114,26 ; jb 4751 <_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 4746 <_sk_load_rgb_u16_be_avx+0x186> - DB 233,197,254,255,255 ; jmpq 4606 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,192,254,255,255 ; jmpq 4606 <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 4756 <_sk_load_rgb_u16_be_avx+0x186> + DB 233,197,254,255,255 ; jmpq 4616 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,192,254,255,255 ; jmpq 4616 <_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 476f <_sk_load_rgb_u16_be_avx+0x1af> + DB 114,20 ; jb 477f <_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 4606 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,146,254,255,255 ; jmpq 4606 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,151,254,255,255 ; jmpq 4616 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,146,254,255,255 ; jmpq 4616 <_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 196,98,125,24,5,42,16,0,0 ; vbroadcastss 0x102a(%rip),%ymm8 # 57b4 <_sk_callback_avx+0x3e2> + DB 196,98,125,24,5,66,16,0,0 ; vbroadcastss 0x1042(%rip),%ymm8 # 57dc <_sk_callback_avx+0x3fa> 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 @@ -8592,7 +8601,7 @@ _sk_store_u16_be_avx LABEL PROC 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 486e <_sk_store_u16_be_avx+0xfa> + DB 117,31 ; jne 487e <_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) @@ -8601,31 +8610,31 @@ _sk_store_u16_be_avx LABEL PROC 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 486a <_sk_store_u16_be_avx+0xf6> + DB 116,240 ; je 487a <_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 486a <_sk_store_u16_be_avx+0xf6> + DB 114,227 ; jb 487a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je 486a <_sk_store_u16_be_avx+0xf6> + DB 116,218 ; je 487a <_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 486a <_sk_store_u16_be_avx+0xf6> + DB 114,205 ; jb 487a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je 486a <_sk_store_u16_be_avx+0xf6> + DB 116,196 ; je 487a <_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 486a <_sk_store_u16_be_avx+0xf6> + DB 114,183 ; jb 487a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp 486a <_sk_store_u16_be_avx+0xf6> + DB 235,174 ; jmp 487a <_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 4932 <_sk_load_f32_avx+0x76> + DB 119,110 ; ja 4942 <_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 # 495c <_sk_load_f32_avx+0xa0> + DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 496c <_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 @@ -8682,7 +8691,7 @@ _sk_store_f32_avx LABEL PROC 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 49e9 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 49f9 <_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 @@ -8695,22 +8704,22 @@ _sk_store_f32_avx LABEL PROC 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 49e5 <_sk_store_f32_avx+0x69> + DB 116,240 ; je 49f5 <_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 49e5 <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 49f5 <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 49e5 <_sk_store_f32_avx+0x69> + DB 116,218 ; je 49f5 <_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 49e5 <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 49f5 <_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 49e5 <_sk_store_f32_avx+0x69> + DB 116,195 ; je 49f5 <_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 49e5 <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 49f5 <_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 49e5 <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 49f5 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -8832,12 +8841,12 @@ _sk_mirror_y_avx LABEL PROC PUBLIC _sk_luminance_to_alpha_avx _sk_luminance_to_alpha_avx LABEL PROC - DB 196,226,125,24,29,179,11,0,0 ; vbroadcastss 0xbb3(%rip),%ymm3 # 57b8 <_sk_callback_avx+0x3e6> + DB 196,226,125,24,29,203,11,0,0 ; vbroadcastss 0xbcb(%rip),%ymm3 # 57e0 <_sk_callback_avx+0x3fe> DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,29,170,11,0,0 ; vbroadcastss 0xbaa(%rip),%ymm3 # 57bc <_sk_callback_avx+0x3ea> + DB 196,226,125,24,29,194,11,0,0 ; vbroadcastss 0xbc2(%rip),%ymm3 # 57e4 <_sk_callback_avx+0x402> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,157,11,0,0 ; vbroadcastss 0xb9d(%rip),%ymm1 # 57c0 <_sk_callback_avx+0x3ee> + DB 196,226,125,24,13,181,11,0,0 ; vbroadcastss 0xbb5(%rip),%ymm1 # 57e8 <_sk_callback_avx+0x406> 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 @@ -9005,7 +9014,7 @@ _sk_linear_gradient_avx LABEL PROC DB 196,226,125,24,88,28 ; vbroadcastss 0x1c(%rax),%ymm3 DB 76,139,0 ; mov (%rax),%r8 DB 77,133,192 ; test %r8,%r8 - DB 15,132,146,0,0,0 ; je 4f79 <_sk_linear_gradient_avx+0xb8> + DB 15,132,146,0,0,0 ; je 4f89 <_sk_linear_gradient_avx+0xb8> DB 72,139,64,8 ; mov 0x8(%rax),%rax DB 72,131,192,32 ; add $0x20,%rax DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12 @@ -9032,8 +9041,8 @@ _sk_linear_gradient_avx LABEL PROC DB 196,227,13,74,219,208 ; vblendvps %ymm13,%ymm3,%ymm14,%ymm3 DB 72,131,192,36 ; add $0x24,%rax DB 73,255,200 ; dec %r8 - DB 117,140 ; jne 4f03 <_sk_linear_gradient_avx+0x42> - DB 235,20 ; jmp 4f8d <_sk_linear_gradient_avx+0xcc> + DB 117,140 ; jne 4f13 <_sk_linear_gradient_avx+0x42> + DB 235,20 ; jmp 4f9d <_sk_linear_gradient_avx+0xcc> DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 @@ -9076,7 +9085,7 @@ _sk_linear_gradient_2stops_avx LABEL PROC PUBLIC _sk_save_xy_avx _sk_save_xy_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,169,7,0,0 ; vbroadcastss 0x7a9(%rip),%ymm8 # 57c4 <_sk_callback_avx+0x3f2> + DB 196,98,125,24,5,193,7,0,0 ; vbroadcastss 0x7c1(%rip),%ymm8 # 57ec <_sk_callback_avx+0x40a> 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 @@ -9109,9 +9118,9 @@ _sk_accumulate_avx LABEL PROC PUBLIC _sk_bilinear_nx_avx _sk_bilinear_nx_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,53,7,0,0 ; vbroadcastss 0x735(%rip),%ymm0 # 57c8 <_sk_callback_avx+0x3f6> + DB 196,226,125,24,5,77,7,0,0 ; vbroadcastss 0x74d(%rip),%ymm0 # 57f0 <_sk_callback_avx+0x40e> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,44,7,0,0 ; vbroadcastss 0x72c(%rip),%ymm8 # 57cc <_sk_callback_avx+0x3fa> + DB 196,98,125,24,5,68,7,0,0 ; vbroadcastss 0x744(%rip),%ymm8 # 57f4 <_sk_callback_avx+0x412> 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 @@ -9120,7 +9129,7 @@ _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 196,226,125,24,5,20,7,0,0 ; vbroadcastss 0x714(%rip),%ymm0 # 57d0 <_sk_callback_avx+0x3fe> + DB 196,226,125,24,5,44,7,0,0 ; vbroadcastss 0x72c(%rip),%ymm0 # 57f8 <_sk_callback_avx+0x416> 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) @@ -9130,9 +9139,9 @@ _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 196,226,125,24,13,248,6,0,0 ; vbroadcastss 0x6f8(%rip),%ymm1 # 57d4 <_sk_callback_avx+0x402> + DB 196,226,125,24,13,16,7,0,0 ; vbroadcastss 0x710(%rip),%ymm1 # 57fc <_sk_callback_avx+0x41a> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,238,6,0,0 ; vbroadcastss 0x6ee(%rip),%ymm8 # 57d8 <_sk_callback_avx+0x406> + DB 196,98,125,24,5,6,7,0,0 ; vbroadcastss 0x706(%rip),%ymm8 # 5800 <_sk_callback_avx+0x41e> 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 @@ -9141,7 +9150,7 @@ _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 196,226,125,24,13,214,6,0,0 ; vbroadcastss 0x6d6(%rip),%ymm1 # 57dc <_sk_callback_avx+0x40a> + DB 196,226,125,24,13,238,6,0,0 ; vbroadcastss 0x6ee(%rip),%ymm1 # 5804 <_sk_callback_avx+0x422> 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) @@ -9151,14 +9160,14 @@ _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 196,226,125,24,5,185,6,0,0 ; vbroadcastss 0x6b9(%rip),%ymm0 # 57e0 <_sk_callback_avx+0x40e> + DB 196,226,125,24,5,209,6,0,0 ; vbroadcastss 0x6d1(%rip),%ymm0 # 5808 <_sk_callback_avx+0x426> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,176,6,0,0 ; vbroadcastss 0x6b0(%rip),%ymm8 # 57e4 <_sk_callback_avx+0x412> + DB 196,98,125,24,5,200,6,0,0 ; vbroadcastss 0x6c8(%rip),%ymm8 # 580c <_sk_callback_avx+0x42a> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,161,6,0,0 ; vbroadcastss 0x6a1(%rip),%ymm10 # 57e8 <_sk_callback_avx+0x416> + DB 196,98,125,24,21,185,6,0,0 ; vbroadcastss 0x6b9(%rip),%ymm10 # 5810 <_sk_callback_avx+0x42e> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,151,6,0,0 ; vbroadcastss 0x697(%rip),%ymm10 # 57ec <_sk_callback_avx+0x41a> + DB 196,98,125,24,21,175,6,0,0 ; vbroadcastss 0x6af(%rip),%ymm10 # 5814 <_sk_callback_avx+0x432> 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) @@ -9168,19 +9177,19 @@ _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 196,226,125,24,5,122,6,0,0 ; vbroadcastss 0x67a(%rip),%ymm0 # 57f0 <_sk_callback_avx+0x41e> + DB 196,226,125,24,5,146,6,0,0 ; vbroadcastss 0x692(%rip),%ymm0 # 5818 <_sk_callback_avx+0x436> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,113,6,0,0 ; vbroadcastss 0x671(%rip),%ymm8 # 57f4 <_sk_callback_avx+0x422> + DB 196,98,125,24,5,137,6,0,0 ; vbroadcastss 0x689(%rip),%ymm8 # 581c <_sk_callback_avx+0x43a> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,103,6,0,0 ; vbroadcastss 0x667(%rip),%ymm9 # 57f8 <_sk_callback_avx+0x426> + DB 196,98,125,24,13,127,6,0,0 ; vbroadcastss 0x67f(%rip),%ymm9 # 5820 <_sk_callback_avx+0x43e> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,93,6,0,0 ; vbroadcastss 0x65d(%rip),%ymm10 # 57fc <_sk_callback_avx+0x42a> + DB 196,98,125,24,21,117,6,0,0 ; vbroadcastss 0x675(%rip),%ymm10 # 5824 <_sk_callback_avx+0x442> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,78,6,0,0 ; vbroadcastss 0x64e(%rip),%ymm10 # 5800 <_sk_callback_avx+0x42e> + DB 196,98,125,24,21,102,6,0,0 ; vbroadcastss 0x666(%rip),%ymm10 # 5828 <_sk_callback_avx+0x446> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,63,6,0,0 ; vbroadcastss 0x63f(%rip),%ymm9 # 5804 <_sk_callback_avx+0x432> + DB 196,98,125,24,13,87,6,0,0 ; vbroadcastss 0x657(%rip),%ymm9 # 582c <_sk_callback_avx+0x44a> 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 @@ -9189,17 +9198,17 @@ _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 196,98,125,24,5,39,6,0,0 ; vbroadcastss 0x627(%rip),%ymm8 # 5808 <_sk_callback_avx+0x436> + DB 196,98,125,24,5,63,6,0,0 ; vbroadcastss 0x63f(%rip),%ymm8 # 5830 <_sk_callback_avx+0x44e> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,25,6,0,0 ; vbroadcastss 0x619(%rip),%ymm10 # 580c <_sk_callback_avx+0x43a> + DB 196,98,125,24,21,49,6,0,0 ; vbroadcastss 0x631(%rip),%ymm10 # 5834 <_sk_callback_avx+0x452> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,15,6,0,0 ; vbroadcastss 0x60f(%rip),%ymm11 # 5810 <_sk_callback_avx+0x43e> + DB 196,98,125,24,29,39,6,0,0 ; vbroadcastss 0x627(%rip),%ymm11 # 5838 <_sk_callback_avx+0x456> 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 196,98,125,24,13,246,5,0,0 ; vbroadcastss 0x5f6(%rip),%ymm9 # 5814 <_sk_callback_avx+0x442> + DB 196,98,125,24,13,14,6,0,0 ; vbroadcastss 0x60e(%rip),%ymm9 # 583c <_sk_callback_avx+0x45a> 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 @@ -9208,13 +9217,13 @@ _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 196,226,125,24,5,222,5,0,0 ; vbroadcastss 0x5de(%rip),%ymm0 # 5818 <_sk_callback_avx+0x446> + DB 196,226,125,24,5,246,5,0,0 ; vbroadcastss 0x5f6(%rip),%ymm0 # 5840 <_sk_callback_avx+0x45e> 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 196,98,125,24,21,203,5,0,0 ; vbroadcastss 0x5cb(%rip),%ymm10 # 581c <_sk_callback_avx+0x44a> + DB 196,98,125,24,21,227,5,0,0 ; vbroadcastss 0x5e3(%rip),%ymm10 # 5844 <_sk_callback_avx+0x462> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,193,5,0,0 ; vbroadcastss 0x5c1(%rip),%ymm10 # 5820 <_sk_callback_avx+0x44e> + DB 196,98,125,24,21,217,5,0,0 ; vbroadcastss 0x5d9(%rip),%ymm10 # 5848 <_sk_callback_avx+0x466> 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) @@ -9224,14 +9233,14 @@ _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 196,226,125,24,13,164,5,0,0 ; vbroadcastss 0x5a4(%rip),%ymm1 # 5824 <_sk_callback_avx+0x452> + DB 196,226,125,24,13,188,5,0,0 ; vbroadcastss 0x5bc(%rip),%ymm1 # 584c <_sk_callback_avx+0x46a> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,154,5,0,0 ; vbroadcastss 0x59a(%rip),%ymm8 # 5828 <_sk_callback_avx+0x456> + DB 196,98,125,24,5,178,5,0,0 ; vbroadcastss 0x5b2(%rip),%ymm8 # 5850 <_sk_callback_avx+0x46e> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,139,5,0,0 ; vbroadcastss 0x58b(%rip),%ymm10 # 582c <_sk_callback_avx+0x45a> + DB 196,98,125,24,21,163,5,0,0 ; vbroadcastss 0x5a3(%rip),%ymm10 # 5854 <_sk_callback_avx+0x472> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,129,5,0,0 ; vbroadcastss 0x581(%rip),%ymm10 # 5830 <_sk_callback_avx+0x45e> + DB 196,98,125,24,21,153,5,0,0 ; vbroadcastss 0x599(%rip),%ymm10 # 5858 <_sk_callback_avx+0x476> 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) @@ -9241,19 +9250,19 @@ _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 196,226,125,24,13,100,5,0,0 ; vbroadcastss 0x564(%rip),%ymm1 # 5834 <_sk_callback_avx+0x462> + DB 196,226,125,24,13,124,5,0,0 ; vbroadcastss 0x57c(%rip),%ymm1 # 585c <_sk_callback_avx+0x47a> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,90,5,0,0 ; vbroadcastss 0x55a(%rip),%ymm8 # 5838 <_sk_callback_avx+0x466> + DB 196,98,125,24,5,114,5,0,0 ; vbroadcastss 0x572(%rip),%ymm8 # 5860 <_sk_callback_avx+0x47e> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,80,5,0,0 ; vbroadcastss 0x550(%rip),%ymm9 # 583c <_sk_callback_avx+0x46a> + DB 196,98,125,24,13,104,5,0,0 ; vbroadcastss 0x568(%rip),%ymm9 # 5864 <_sk_callback_avx+0x482> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,70,5,0,0 ; vbroadcastss 0x546(%rip),%ymm10 # 5840 <_sk_callback_avx+0x46e> + DB 196,98,125,24,21,94,5,0,0 ; vbroadcastss 0x55e(%rip),%ymm10 # 5868 <_sk_callback_avx+0x486> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,55,5,0,0 ; vbroadcastss 0x537(%rip),%ymm10 # 5844 <_sk_callback_avx+0x472> + DB 196,98,125,24,21,79,5,0,0 ; vbroadcastss 0x54f(%rip),%ymm10 # 586c <_sk_callback_avx+0x48a> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,40,5,0,0 ; vbroadcastss 0x528(%rip),%ymm9 # 5848 <_sk_callback_avx+0x476> + DB 196,98,125,24,13,64,5,0,0 ; vbroadcastss 0x540(%rip),%ymm9 # 5870 <_sk_callback_avx+0x48e> 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 @@ -9262,17 +9271,17 @@ _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 196,98,125,24,5,16,5,0,0 ; vbroadcastss 0x510(%rip),%ymm8 # 584c <_sk_callback_avx+0x47a> + DB 196,98,125,24,5,40,5,0,0 ; vbroadcastss 0x528(%rip),%ymm8 # 5874 <_sk_callback_avx+0x492> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,1,5,0,0 ; vbroadcastss 0x501(%rip),%ymm10 # 5850 <_sk_callback_avx+0x47e> + DB 196,98,125,24,21,25,5,0,0 ; vbroadcastss 0x519(%rip),%ymm10 # 5878 <_sk_callback_avx+0x496> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,247,4,0,0 ; vbroadcastss 0x4f7(%rip),%ymm11 # 5854 <_sk_callback_avx+0x482> + DB 196,98,125,24,29,15,5,0,0 ; vbroadcastss 0x50f(%rip),%ymm11 # 587c <_sk_callback_avx+0x49a> 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 196,98,125,24,13,222,4,0,0 ; vbroadcastss 0x4de(%rip),%ymm9 # 5858 <_sk_callback_avx+0x486> + DB 196,98,125,24,13,246,4,0,0 ; vbroadcastss 0x4f6(%rip),%ymm9 # 5880 <_sk_callback_avx+0x49e> 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 @@ -9281,13 +9290,13 @@ _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 196,226,125,24,13,198,4,0,0 ; vbroadcastss 0x4c6(%rip),%ymm1 # 585c <_sk_callback_avx+0x48a> + DB 196,226,125,24,13,222,4,0,0 ; vbroadcastss 0x4de(%rip),%ymm1 # 5884 <_sk_callback_avx+0x4a2> 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 196,98,125,24,21,178,4,0,0 ; vbroadcastss 0x4b2(%rip),%ymm10 # 5860 <_sk_callback_avx+0x48e> + DB 196,98,125,24,21,202,4,0,0 ; vbroadcastss 0x4ca(%rip),%ymm10 # 5888 <_sk_callback_avx+0x4a6> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,168,4,0,0 ; vbroadcastss 0x4a8(%rip),%ymm10 # 5864 <_sk_callback_avx+0x492> + DB 196,98,125,24,21,192,4,0,0 ; vbroadcastss 0x4c0(%rip),%ymm10 # 588c <_sk_callback_avx+0x4aa> 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) @@ -9418,11 +9427,21 @@ ALIGN 4 DB 0,128,64,171,170,42 ; add %al,0x2aaaab40(%rax) DB 62,0,0 ; add %al,%ds:(%rax) DB 0,63 ; add %bh,(%rdi) - DB 129,128,128,59,129,128,128,59,0,248 ; addl $0xf8003b80,-0x7f7ec480(%rax) DB 0,0 ; add %al,(%rax) - DB 8,33 ; or %ah,(%rcx) + DB 128,63,171 ; cmpb $0xab,(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,0,0 ; add %al,%ds:(%rax) + DB 128,191,0,0,192,64,171 ; cmpb $0xab,0x40c00000(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + 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 55a9 <.literal4+0xbd> + DB 224,7 ; loopne 55d1 <.literal4+0xd5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -9436,10 +9455,10 @@ 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 55d4 <.literal4+0xe8> + DB 127,0 ; jg 55fc <.literal4+0x100> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 564d <.literal4+0x161> + DB 119,115 ; ja 5675 <.literal4+0x179> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -9453,10 +9472,10 @@ 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 5608 <.literal4+0x11c> + DB 127,0 ; jg 5630 <.literal4+0x134> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5681 <.literal4+0x195> + DB 119,115 ; ja 56a9 <.literal4+0x1ad> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -9470,10 +9489,10 @@ 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 563c <.literal4+0x150> + DB 127,0 ; jg 5664 <.literal4+0x168> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 56b5 <.literal4+0x1c9> + DB 119,115 ; ja 56dd <.literal4+0x1e1> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -9487,10 +9506,10 @@ 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 5670 <.literal4+0x184> + DB 127,0 ; jg 5698 <.literal4+0x19c> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 56e9 <.literal4+0x1fd> + DB 119,115 ; ja 5711 <.literal4+0x215> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -9503,7 +9522,7 @@ 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 56e7 <.literal4+0x1fb> + DB 127,67 ; jg 570f <.literal4+0x213> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -9515,10 +9534,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5707 <.literal4+0x21b> + DB 118,63 ; jbe 572f <.literal4+0x233> 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 571b <.literal4+0x22f> + DB 127,67 ; jg 5743 <.literal4+0x247> 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) @@ -9527,7 +9546,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 56fd <.literal4+0x211> + DB 224,7 ; loopne 5725 <.literal4+0x229> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -9539,7 +9558,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5719 <.literal4+0x22d> + DB 224,7 ; loopne 5741 <.literal4+0x245> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -9550,7 +9569,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 576e <.literal4+0x282> + DB 124,66 ; jl 5796 <.literal4+0x29a> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -9568,9 +9587,9 @@ 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 57b1 <.literal4+0x2c5> + DB 112,65 ; jo 57d9 <.literal4+0x2dd> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 57bf <.literal4+0x2d3> + DB 127,67 ; jg 57e7 <.literal4+0x2eb> 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) @@ -9586,7 +9605,7 @@ 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 57ff <.literal4+0x313> + DB 127,71 ; jg 5827 <.literal4+0x32b> DB 208 ; (bad) DB 179,89 ; mov $0x59,%bl DB 62,89 ; ds pop %rcx @@ -9823,7 +9842,7 @@ _sk_seed_shader_sse41 LABEL PROC DB 102,15,110,199 ; movd %edi,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 - DB 15,40,21,97,57,0,0 ; movaps 0x3961(%rip),%xmm2 # 3a70 <_sk_callback_sse41+0xaf> + DB 15,40,21,129,57,0,0 ; movaps 0x3981(%rip),%xmm2 # 3a90 <_sk_callback_sse41+0xb7> DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,16,2 ; movups (%rdx),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 @@ -9832,7 +9851,7 @@ _sk_seed_shader_sse41 LABEL PROC DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,80,57,0,0 ; movaps 0x3950(%rip),%xmm2 # 3a80 <_sk_callback_sse41+0xbf> + DB 15,40,21,112,57,0,0 ; movaps 0x3970(%rip),%xmm2 # 3aa0 <_sk_callback_sse41+0xc7> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -9866,7 +9885,7 @@ _sk_clear_sse41 LABEL PROC PUBLIC _sk_srcatop_sse41 _sk_srcatop_sse41 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,11,57,0,0 ; movaps 0x390b(%rip),%xmm8 # 3a90 <_sk_callback_sse41+0xcf> + DB 68,15,40,5,43,57,0,0 ; movaps 0x392b(%rip),%xmm8 # 3ab0 <_sk_callback_sse41+0xd7> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -9889,7 +9908,7 @@ 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 68,15,40,13,206,56,0,0 ; movaps 0x38ce(%rip),%xmm9 # 3aa0 <_sk_callback_sse41+0xdf> + DB 68,15,40,13,238,56,0,0 ; movaps 0x38ee(%rip),%xmm9 # 3ac0 <_sk_callback_sse41+0xe7> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -9930,7 +9949,7 @@ _sk_dstin_sse41 LABEL PROC PUBLIC _sk_srcout_sse41 _sk_srcout_sse41 LABEL PROC - DB 68,15,40,5,114,56,0,0 ; movaps 0x3872(%rip),%xmm8 # 3ab0 <_sk_callback_sse41+0xef> + DB 68,15,40,5,146,56,0,0 ; movaps 0x3892(%rip),%xmm8 # 3ad0 <_sk_callback_sse41+0xf7> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -9941,7 +9960,7 @@ _sk_srcout_sse41 LABEL PROC PUBLIC _sk_dstout_sse41 _sk_dstout_sse41 LABEL PROC - DB 68,15,40,5,98,56,0,0 ; movaps 0x3862(%rip),%xmm8 # 3ac0 <_sk_callback_sse41+0xff> + DB 68,15,40,5,130,56,0,0 ; movaps 0x3882(%rip),%xmm8 # 3ae0 <_sk_callback_sse41+0x107> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -9956,7 +9975,7 @@ _sk_dstout_sse41 LABEL PROC PUBLIC _sk_srcover_sse41 _sk_srcover_sse41 LABEL PROC - DB 68,15,40,5,69,56,0,0 ; movaps 0x3845(%rip),%xmm8 # 3ad0 <_sk_callback_sse41+0x10f> + DB 68,15,40,5,101,56,0,0 ; movaps 0x3865(%rip),%xmm8 # 3af0 <_sk_callback_sse41+0x117> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -9974,7 +9993,7 @@ _sk_srcover_sse41 LABEL PROC PUBLIC _sk_dstover_sse41 _sk_dstover_sse41 LABEL PROC - DB 68,15,40,5,25,56,0,0 ; movaps 0x3819(%rip),%xmm8 # 3ae0 <_sk_callback_sse41+0x11f> + DB 68,15,40,5,57,56,0,0 ; movaps 0x3839(%rip),%xmm8 # 3b00 <_sk_callback_sse41+0x127> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -9998,7 +10017,7 @@ _sk_modulate_sse41 LABEL PROC PUBLIC _sk_multiply_sse41 _sk_multiply_sse41 LABEL PROC - DB 68,15,40,5,237,55,0,0 ; movaps 0x37ed(%rip),%xmm8 # 3af0 <_sk_callback_sse41+0x12f> + DB 68,15,40,5,13,56,0,0 ; movaps 0x380d(%rip),%xmm8 # 3b10 <_sk_callback_sse41+0x137> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -10068,7 +10087,7 @@ _sk_screen_sse41 LABEL PROC PUBLIC _sk_xor__sse41 _sk_xor__sse41 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,30,55,0,0 ; movaps 0x371e(%rip),%xmm3 # 3b00 <_sk_callback_sse41+0x13f> + DB 15,40,29,62,55,0,0 ; movaps 0x373e(%rip),%xmm3 # 3b20 <_sk_callback_sse41+0x147> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -10114,7 +10133,7 @@ _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 15,40,21,137,54,0,0 ; movaps 0x3689(%rip),%xmm2 # 3b10 <_sk_callback_sse41+0x14f> + DB 15,40,21,169,54,0,0 ; movaps 0x36a9(%rip),%xmm2 # 3b30 <_sk_callback_sse41+0x157> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -10146,7 +10165,7 @@ _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 15,40,21,46,54,0,0 ; movaps 0x362e(%rip),%xmm2 # 3b20 <_sk_callback_sse41+0x15f> + DB 15,40,21,78,54,0,0 ; movaps 0x364e(%rip),%xmm2 # 3b40 <_sk_callback_sse41+0x167> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -10181,7 +10200,7 @@ _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 15,40,21,200,53,0,0 ; movaps 0x35c8(%rip),%xmm2 # 3b30 <_sk_callback_sse41+0x16f> + DB 15,40,21,232,53,0,0 ; movaps 0x35e8(%rip),%xmm2 # 3b50 <_sk_callback_sse41+0x177> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -10206,7 +10225,7 @@ _sk_exclusion_sse41 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,202 ; subps %xmm2,%xmm9 - DB 15,40,13,137,53,0,0 ; movaps 0x3589(%rip),%xmm1 # 3b40 <_sk_callback_sse41+0x17f> + DB 15,40,13,169,53,0,0 ; movaps 0x35a9(%rip),%xmm1 # 3b60 <_sk_callback_sse41+0x187> DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,207 ; mulps %xmm7,%xmm1 DB 15,88,217 ; addps %xmm1,%xmm3 @@ -10218,7 +10237,7 @@ _sk_exclusion_sse41 LABEL PROC PUBLIC _sk_colorburn_sse41 _sk_colorburn_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,120,53,0,0 ; movaps 0x3578(%rip),%xmm10 # 3b50 <_sk_callback_sse41+0x18f> + DB 68,15,40,21,152,53,0,0 ; movaps 0x3598(%rip),%xmm10 # 3b70 <_sk_callback_sse41+0x197> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,203 ; movaps %xmm11,%xmm9 @@ -10298,7 +10317,7 @@ _sk_colorburn_sse41 LABEL PROC PUBLIC _sk_colordodge_sse41 _sk_colordodge_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,86,52,0,0 ; movaps 0x3456(%rip),%xmm10 # 3b60 <_sk_callback_sse41+0x19f> + DB 68,15,40,21,118,52,0,0 ; movaps 0x3476(%rip),%xmm10 # 3b80 <_sk_callback_sse41+0x1a7> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -10379,7 +10398,7 @@ _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 68,15,40,21,44,51,0,0 ; movaps 0x332c(%rip),%xmm10 # 3b70 <_sk_callback_sse41+0x1af> + DB 68,15,40,21,76,51,0,0 ; movaps 0x334c(%rip),%xmm10 # 3b90 <_sk_callback_sse41+0x1b7> DB 65,15,40,234 ; movaps %xmm10,%xmm5 DB 15,92,239 ; subps %xmm7,%xmm5 DB 15,40,197 ; movaps %xmm5,%xmm0 @@ -10461,7 +10480,7 @@ 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 68,15,40,21,14,50,0,0 ; movaps 0x320e(%rip),%xmm10 # 3b80 <_sk_callback_sse41+0x1bf> + DB 68,15,40,21,46,50,0,0 ; movaps 0x322e(%rip),%xmm10 # 3ba0 <_sk_callback_sse41+0x1c7> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 65,15,40,195 ; movaps %xmm11,%xmm0 @@ -10545,7 +10564,7 @@ _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 15,40,13,225,48,0,0 ; movaps 0x30e1(%rip),%xmm1 # 3b90 <_sk_callback_sse41+0x1cf> + DB 15,40,13,1,49,0,0 ; movaps 0x3101(%rip),%xmm1 # 3bb0 <_sk_callback_sse41+0x1d7> DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 68,15,40,240 ; movaps %xmm0,%xmm14 @@ -10558,10 +10577,10 @@ _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 68,15,40,45,191,48,0,0 ; movaps 0x30bf(%rip),%xmm13 # 3ba0 <_sk_callback_sse41+0x1df> + DB 68,15,40,45,223,48,0,0 ; movaps 0x30df(%rip),%xmm13 # 3bc0 <_sk_callback_sse41+0x1e7> DB 69,15,88,245 ; addps %xmm13,%xmm14 DB 68,15,89,242 ; mulps %xmm2,%xmm14 - DB 68,15,40,37,191,48,0,0 ; movaps 0x30bf(%rip),%xmm12 # 3bb0 <_sk_callback_sse41+0x1ef> + DB 68,15,40,37,223,48,0,0 ; movaps 0x30df(%rip),%xmm12 # 3bd0 <_sk_callback_sse41+0x1f7> DB 69,15,89,252 ; mulps %xmm12,%xmm15 DB 69,15,88,254 ; addps %xmm14,%xmm15 DB 15,40,198 ; movaps %xmm6,%xmm0 @@ -10706,7 +10725,7 @@ _sk_clamp_0_sse41 LABEL PROC PUBLIC _sk_clamp_1_sse41 _sk_clamp_1_sse41 LABEL PROC - DB 68,15,40,5,207,46,0,0 ; movaps 0x2ecf(%rip),%xmm8 # 3bc0 <_sk_callback_sse41+0x1ff> + DB 68,15,40,5,239,46,0,0 ; movaps 0x2eef(%rip),%xmm8 # 3be0 <_sk_callback_sse41+0x207> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -10716,7 +10735,7 @@ _sk_clamp_1_sse41 LABEL PROC PUBLIC _sk_clamp_a_sse41 _sk_clamp_a_sse41 LABEL PROC - DB 15,93,29,196,46,0,0 ; minps 0x2ec4(%rip),%xmm3 # 3bd0 <_sk_callback_sse41+0x20f> + DB 15,93,29,228,46,0,0 ; minps 0x2ee4(%rip),%xmm3 # 3bf0 <_sk_callback_sse41+0x217> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -10789,7 +10808,7 @@ _sk_premul_sse41 LABEL PROC PUBLIC _sk_unpremul_sse41 _sk_unpremul_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,47,46,0,0 ; movaps 0x2e2f(%rip),%xmm9 # 3be0 <_sk_callback_sse41+0x21f> + DB 68,15,40,13,79,46,0,0 ; movaps 0x2e4f(%rip),%xmm9 # 3c00 <_sk_callback_sse41+0x227> 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 @@ -10801,20 +10820,20 @@ _sk_unpremul_sse41 LABEL PROC PUBLIC _sk_from_srgb_sse41 _sk_from_srgb_sse41 LABEL PROC - DB 68,15,40,29,26,46,0,0 ; movaps 0x2e1a(%rip),%xmm11 # 3bf0 <_sk_callback_sse41+0x22f> + DB 68,15,40,29,58,46,0,0 ; movaps 0x2e3a(%rip),%xmm11 # 3c10 <_sk_callback_sse41+0x237> 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 68,15,40,37,18,46,0,0 ; movaps 0x2e12(%rip),%xmm12 # 3c00 <_sk_callback_sse41+0x23f> + DB 68,15,40,37,50,46,0,0 ; movaps 0x2e32(%rip),%xmm12 # 3c20 <_sk_callback_sse41+0x247> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,18,46,0,0 ; movaps 0x2e12(%rip),%xmm13 # 3c10 <_sk_callback_sse41+0x24f> + DB 68,15,40,45,50,46,0,0 ; movaps 0x2e32(%rip),%xmm13 # 3c30 <_sk_callback_sse41+0x257> DB 69,15,88,197 ; addps %xmm13,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 - DB 68,15,40,53,18,46,0,0 ; movaps 0x2e12(%rip),%xmm14 # 3c20 <_sk_callback_sse41+0x25f> + DB 68,15,40,53,50,46,0,0 ; movaps 0x2e32(%rip),%xmm14 # 3c40 <_sk_callback_sse41+0x267> DB 69,15,88,198 ; addps %xmm14,%xmm8 - DB 68,15,40,61,22,46,0,0 ; movaps 0x2e16(%rip),%xmm15 # 3c30 <_sk_callback_sse41+0x26f> + DB 68,15,40,61,54,46,0,0 ; movaps 0x2e36(%rip),%xmm15 # 3c50 <_sk_callback_sse41+0x277> 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 @@ -10858,20 +10877,20 @@ _sk_to_srgb_sse41 LABEL PROC DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 DB 69,15,83,200 ; rcpps %xmm8,%xmm9 DB 69,15,82,208 ; rsqrtps %xmm8,%xmm10 - DB 68,15,40,29,131,45,0,0 ; movaps 0x2d83(%rip),%xmm11 # 3c40 <_sk_callback_sse41+0x27f> + DB 68,15,40,29,163,45,0,0 ; movaps 0x2da3(%rip),%xmm11 # 3c60 <_sk_callback_sse41+0x287> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 65,15,89,203 ; mulps %xmm11,%xmm1 - DB 68,15,40,37,132,45,0,0 ; movaps 0x2d84(%rip),%xmm12 # 3c50 <_sk_callback_sse41+0x28f> + DB 68,15,40,37,164,45,0,0 ; movaps 0x2da4(%rip),%xmm12 # 3c70 <_sk_callback_sse41+0x297> DB 69,15,89,204 ; mulps %xmm12,%xmm9 - DB 68,15,40,45,136,45,0,0 ; movaps 0x2d88(%rip),%xmm13 # 3c60 <_sk_callback_sse41+0x29f> + DB 68,15,40,45,168,45,0,0 ; movaps 0x2da8(%rip),%xmm13 # 3c80 <_sk_callback_sse41+0x2a7> DB 69,15,88,205 ; addps %xmm13,%xmm9 - DB 68,15,40,53,140,45,0,0 ; movaps 0x2d8c(%rip),%xmm14 # 3c70 <_sk_callback_sse41+0x2af> + DB 68,15,40,53,172,45,0,0 ; movaps 0x2dac(%rip),%xmm14 # 3c90 <_sk_callback_sse41+0x2b7> DB 69,15,89,214 ; mulps %xmm14,%xmm10 DB 69,15,88,209 ; addps %xmm9,%xmm10 - DB 68,15,40,5,140,45,0,0 ; movaps 0x2d8c(%rip),%xmm8 # 3c80 <_sk_callback_sse41+0x2bf> + DB 68,15,40,5,172,45,0,0 ; movaps 0x2dac(%rip),%xmm8 # 3ca0 <_sk_callback_sse41+0x2c7> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,93,202 ; minps %xmm10,%xmm9 - DB 68,15,40,61,140,45,0,0 ; movaps 0x2d8c(%rip),%xmm15 # 3c90 <_sk_callback_sse41+0x2cf> + DB 68,15,40,61,172,45,0,0 ; movaps 0x2dac(%rip),%xmm15 # 3cb0 <_sk_callback_sse41+0x2d7> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 102,68,15,56,20,201 ; blendvps %xmm0,%xmm1,%xmm9 DB 15,82,194 ; rsqrtps %xmm2,%xmm0 @@ -10924,7 +10943,7 @@ _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 68,15,40,53,218,44,0,0 ; movaps 0x2cda(%rip),%xmm14 # 3ca0 <_sk_callback_sse41+0x2df> + DB 68,15,40,53,250,44,0,0 ; movaps 0x2cfa(%rip),%xmm14 # 3cc0 <_sk_callback_sse41+0x2e7> 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 @@ -10933,27 +10952,27 @@ _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 68,15,84,61,193,44,0,0 ; andps 0x2cc1(%rip),%xmm15 # 3cb0 <_sk_callback_sse41+0x2ef> + DB 68,15,84,61,225,44,0,0 ; andps 0x2ce1(%rip),%xmm15 # 3cd0 <_sk_callback_sse41+0x2f7> 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 68,15,40,45,180,44,0,0 ; movaps 0x2cb4(%rip),%xmm13 # 3cc0 <_sk_callback_sse41+0x2ff> + DB 68,15,40,45,212,44,0,0 ; movaps 0x2cd4(%rip),%xmm13 # 3ce0 <_sk_callback_sse41+0x307> 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 68,15,88,5,176,44,0,0 ; addps 0x2cb0(%rip),%xmm8 # 3cd0 <_sk_callback_sse41+0x30f> + DB 68,15,88,5,208,44,0,0 ; addps 0x2cd0(%rip),%xmm8 # 3cf0 <_sk_callback_sse41+0x317> 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 68,15,89,5,168,44,0,0 ; mulps 0x2ca8(%rip),%xmm8 # 3ce0 <_sk_callback_sse41+0x31f> + DB 68,15,89,5,200,44,0,0 ; mulps 0x2cc8(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x327> 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 15,40,5,156,44,0,0 ; movaps 0x2c9c(%rip),%xmm0 # 3cf0 <_sk_callback_sse41+0x32f> + DB 15,40,5,188,44,0,0 ; movaps 0x2cbc(%rip),%xmm0 # 3d10 <_sk_callback_sse41+0x337> DB 65,15,40,211 ; movaps %xmm11,%xmm2 DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,194,194,1 ; cmpltps %xmm2,%xmm0 @@ -10973,152 +10992,150 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,108,36,96 ; movaps %xmm5,0x60(%rsp) DB 15,41,100,36,80 ; movaps %xmm4,0x50(%rsp) DB 15,41,92,36,64 ; movaps %xmm3,0x40(%rsp) - DB 68,15,40,208 ; movaps %xmm0,%xmm10 + DB 68,15,40,216 ; movaps %xmm0,%xmm11 DB 184,0,0,0,63 ; mov $0x3f000000,%eax DB 102,15,110,216 ; movd %eax,%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 15,41,28,36 ; movaps %xmm3,(%rsp) DB 15,40,194 ; movaps %xmm2,%xmm0 DB 15,194,195,1 ; cmpltps %xmm3,%xmm0 - DB 184,0,0,128,63 ; mov $0x3f800000,%eax - DB 102,68,15,110,192 ; movd %eax,%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 65,15,40,224 ; movaps %xmm8,%xmm4 - DB 15,88,225 ; addps %xmm1,%xmm4 - DB 15,89,226 ; mulps %xmm2,%xmm4 + DB 15,40,45,97,44,0,0 ; movaps 0x2c61(%rip),%xmm5 # 3d20 <_sk_callback_sse41+0x347> + DB 15,40,241 ; movaps %xmm1,%xmm6 + DB 15,40,225 ; movaps %xmm1,%xmm4 DB 15,40,217 ; movaps %xmm1,%xmm3 - DB 15,40,249 ; movaps %xmm1,%xmm7 - DB 15,88,250 ; addps %xmm2,%xmm7 + DB 15,88,221 ; addps %xmm5,%xmm3 + DB 15,40,253 ; movaps %xmm5,%xmm7 DB 15,89,218 ; mulps %xmm2,%xmm3 + DB 15,88,242 ; addps %xmm2,%xmm6 + DB 15,89,226 ; mulps %xmm2,%xmm4 DB 15,40,234 ; movaps %xmm2,%xmm5 - DB 15,92,251 ; subps %xmm3,%xmm7 - DB 102,15,56,20,252 ; blendvps %xmm0,%xmm4,%xmm7 - DB 184,0,0,0,64 ; mov $0x40000000,%eax - DB 185,171,170,170,62 ; mov $0x3eaaaaab,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 + DB 15,92,244 ; subps %xmm4,%xmm6 + DB 102,15,56,20,243 ; blendvps %xmm0,%xmm3,%xmm6 + DB 68,15,40,61,70,44,0,0 ; movaps 0x2c46(%rip),%xmm15 # 3d30 <_sk_callback_sse41+0x357> + DB 69,15,88,251 ; addps %xmm11,%xmm15 + DB 184,0,0,0,0 ; mov $0x0,%eax + DB 185,0,0,128,63 ; mov $0x3f800000,%ecx + DB 102,68,15,110,201 ; movd %ecx,%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 + DB 65,15,40,215 ; movaps %xmm15,%xmm2 + DB 15,88,21,42,44,0,0 ; addps 0x2c2a(%rip),%xmm2 # 3d40 <_sk_callback_sse41+0x367> + DB 69,15,40,231 ; movaps %xmm15,%xmm12 + DB 102,68,15,56,20,226 ; blendvps %xmm0,%xmm2,%xmm12 + DB 102,15,110,208 ; movd %eax,%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 15,41,84,36,48 ; movaps %xmm2,0x30(%rsp) - DB 65,15,88,210 ; addps %xmm10,%xmm2 - DB 185,0,0,0,0 ; mov $0x0,%ecx - DB 65,15,40,192 ; movaps %xmm8,%xmm0 + DB 15,41,84,36,32 ; movaps %xmm2,0x20(%rsp) + DB 65,15,40,199 ; movaps %xmm15,%xmm0 DB 15,194,194,1 ; cmpltps %xmm2,%xmm0 - DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 65,15,92,216 ; subps %xmm8,%xmm3 - DB 68,15,40,226 ; movaps %xmm2,%xmm12 - DB 102,68,15,56,20,227 ; blendvps %xmm0,%xmm3,%xmm12 - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 15,40,194 ; movaps %xmm2,%xmm0 - DB 65,15,194,198,1 ; cmpltps %xmm14,%xmm0 - DB 68,15,41,116,36,32 ; movaps %xmm14,0x20(%rsp) - DB 65,15,40,216 ; movaps %xmm8,%xmm3 - DB 15,88,218 ; addps %xmm2,%xmm3 - DB 102,68,15,56,20,227 ; blendvps %xmm0,%xmm3,%xmm12 - DB 102,68,15,110,200 ; movd %eax,%xmm9 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 15,40,229 ; movaps %xmm5,%xmm4 - DB 15,41,100,36,16 ; movaps %xmm4,0x10(%rsp) - DB 68,15,89,204 ; mulps %xmm4,%xmm9 - DB 68,15,92,207 ; subps %xmm7,%xmm9 + DB 65,15,40,215 ; movaps %xmm15,%xmm2 + DB 15,88,215 ; addps %xmm7,%xmm2 + DB 102,68,15,56,20,226 ; blendvps %xmm0,%xmm2,%xmm12 + DB 15,40,221 ; movaps %xmm5,%xmm3 + DB 15,41,92,36,48 ; movaps %xmm3,0x30(%rsp) + DB 68,15,40,195 ; movaps %xmm3,%xmm8 + DB 69,15,88,192 ; addps %xmm8,%xmm8 + DB 68,15,92,198 ; subps %xmm6,%xmm8 DB 184,171,170,42,62 ; mov $0x3e2aaaab,%eax - DB 15,40,199 ; movaps %xmm7,%xmm0 - DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 185,0,0,192,64 ; mov $0x40c00000,%ecx - DB 102,15,110,241 ; movd %ecx,%xmm6 - DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 - DB 15,89,240 ; mulps %xmm0,%xmm6 + DB 15,40,214 ; movaps %xmm6,%xmm2 + DB 65,15,92,208 ; subps %xmm8,%xmm2 + DB 15,89,21,231,43,0,0 ; mulps 0x2be7(%rip),%xmm2 # 3d50 <_sk_callback_sse41+0x377> DB 185,171,170,42,63 ; mov $0x3f2aaaab,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 15,40,235 ; movaps %xmm3,%xmm5 + DB 102,15,110,193 ; movd %ecx,%xmm0 + DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0 + DB 15,41,68,36,16 ; movaps %xmm0,0x10(%rsp) + DB 15,40,37,222,43,0,0 ; movaps 0x2bde(%rip),%xmm4 # 3d60 <_sk_callback_sse41+0x387> + DB 15,40,236 ; movaps %xmm4,%xmm5 DB 65,15,92,236 ; subps %xmm12,%xmm5 DB 69,15,40,236 ; movaps %xmm12,%xmm13 - DB 69,15,40,252 ; movaps %xmm12,%xmm15 - DB 68,15,194,227,1 ; cmpltps %xmm3,%xmm12 - DB 15,89,238 ; mulps %xmm6,%xmm5 - DB 65,15,88,233 ; addps %xmm9,%xmm5 - DB 69,15,40,217 ; movaps %xmm9,%xmm11 + DB 69,15,40,244 ; movaps %xmm12,%xmm14 + DB 68,15,194,224,1 ; cmpltps %xmm0,%xmm12 + DB 15,89,234 ; mulps %xmm2,%xmm5 + DB 65,15,88,232 ; addps %xmm8,%xmm5 + DB 69,15,40,208 ; movaps %xmm8,%xmm10 DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 102,68,15,56,20,221 ; blendvps %xmm0,%xmm5,%xmm11 - DB 68,15,194,60,36,1 ; cmpltps (%rsp),%xmm15 - DB 65,15,40,199 ; movaps %xmm15,%xmm0 - DB 102,68,15,56,20,223 ; blendvps %xmm0,%xmm7,%xmm11 - DB 102,15,110,232 ; movd %eax,%xmm5 - DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 - DB 68,15,194,237,1 ; cmpltps %xmm5,%xmm13 - DB 15,89,214 ; mulps %xmm6,%xmm2 - DB 65,15,88,209 ; addps %xmm9,%xmm2 + DB 102,68,15,56,20,213 ; blendvps %xmm0,%xmm5,%xmm10 + DB 15,40,44,36 ; movaps (%rsp),%xmm5 + DB 68,15,194,245,1 ; cmpltps %xmm5,%xmm14 + DB 65,15,40,198 ; movaps %xmm14,%xmm0 + DB 102,68,15,56,20,214 ; blendvps %xmm0,%xmm6,%xmm10 + DB 102,15,110,248 ; movd %eax,%xmm7 + DB 15,198,255,0 ; shufps $0x0,%xmm7,%xmm7 + DB 68,15,194,239,1 ; cmpltps %xmm7,%xmm13 + DB 68,15,89,250 ; mulps %xmm2,%xmm15 + DB 69,15,88,248 ; addps %xmm8,%xmm15 DB 65,15,40,197 ; movaps %xmm13,%xmm0 - DB 102,68,15,56,20,218 ; blendvps %xmm0,%xmm2,%xmm11 + DB 102,69,15,56,20,215 ; blendvps %xmm0,%xmm15,%xmm10 DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 68,15,194,225,0 ; cmpeqps %xmm1,%xmm12 DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 102,68,15,56,20,220 ; blendvps %xmm0,%xmm4,%xmm11 - DB 65,15,40,192 ; movaps %xmm8,%xmm0 - DB 65,15,194,194,1 ; cmpltps %xmm10,%xmm0 - DB 65,15,40,202 ; movaps %xmm10,%xmm1 - DB 65,15,92,200 ; subps %xmm8,%xmm1 - DB 69,15,40,234 ; movaps %xmm10,%xmm13 + DB 102,68,15,56,20,211 ; blendvps %xmm0,%xmm3,%xmm10 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 + DB 65,15,40,203 ; movaps %xmm11,%xmm1 + DB 15,88,13,60,43,0,0 ; addps 0x2b3c(%rip),%xmm1 # 3d40 <_sk_callback_sse41+0x367> + DB 69,15,40,235 ; movaps %xmm11,%xmm13 DB 102,68,15,56,20,233 ; blendvps %xmm0,%xmm1,%xmm13 - DB 65,15,40,194 ; movaps %xmm10,%xmm0 - DB 65,15,194,198,1 ; cmpltps %xmm14,%xmm0 - DB 65,15,40,200 ; movaps %xmm8,%xmm1 - DB 65,15,88,202 ; addps %xmm10,%xmm1 + DB 65,15,40,195 ; movaps %xmm11,%xmm0 + DB 15,194,68,36,32,1 ; cmpltps 0x20(%rsp),%xmm0 + DB 65,15,40,203 ; movaps %xmm11,%xmm1 + DB 15,88,13,253,42,0,0 ; addps 0x2afd(%rip),%xmm1 # 3d20 <_sk_callback_sse41+0x347> DB 102,68,15,56,20,233 ; blendvps %xmm0,%xmm1,%xmm13 - DB 68,15,40,243 ; movaps %xmm3,%xmm14 - DB 69,15,92,245 ; subps %xmm13,%xmm14 - DB 65,15,40,229 ; movaps %xmm13,%xmm4 + DB 15,40,220 ; movaps %xmm4,%xmm3 + DB 65,15,92,221 ; subps %xmm13,%xmm3 + DB 69,15,40,245 ; movaps %xmm13,%xmm14 DB 69,15,40,253 ; movaps %xmm13,%xmm15 - DB 68,15,194,235,1 ; cmpltps %xmm3,%xmm13 - DB 68,15,89,246 ; mulps %xmm6,%xmm14 - DB 69,15,88,241 ; addps %xmm9,%xmm14 - DB 65,15,40,201 ; movaps %xmm9,%xmm1 + DB 68,15,194,108,36,16,1 ; cmpltps 0x10(%rsp),%xmm13 + DB 15,89,218 ; mulps %xmm2,%xmm3 + DB 65,15,88,216 ; addps %xmm8,%xmm3 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 DB 65,15,40,197 ; movaps %xmm13,%xmm0 - DB 102,65,15,56,20,206 ; blendvps %xmm0,%xmm14,%xmm1 - DB 68,15,40,52,36 ; movaps (%rsp),%xmm14 - DB 69,15,194,254,1 ; cmpltps %xmm14,%xmm15 + DB 102,15,56,20,203 ; blendvps %xmm0,%xmm3,%xmm1 + DB 68,15,194,253,1 ; cmpltps %xmm5,%xmm15 DB 65,15,40,199 ; movaps %xmm15,%xmm0 - DB 102,15,56,20,207 ; blendvps %xmm0,%xmm7,%xmm1 - DB 15,194,229,1 ; cmpltps %xmm5,%xmm4 - DB 15,40,214 ; movaps %xmm6,%xmm2 - DB 65,15,89,210 ; mulps %xmm10,%xmm2 - DB 65,15,88,209 ; addps %xmm9,%xmm2 - DB 15,40,196 ; movaps %xmm4,%xmm0 - DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 + DB 102,15,56,20,206 ; blendvps %xmm0,%xmm6,%xmm1 + DB 68,15,194,247,1 ; cmpltps %xmm7,%xmm14 + DB 15,40,218 ; movaps %xmm2,%xmm3 + DB 65,15,89,219 ; mulps %xmm11,%xmm3 + DB 65,15,88,216 ; addps %xmm8,%xmm3 + DB 65,15,40,198 ; movaps %xmm14,%xmm0 + DB 102,15,56,20,203 ; blendvps %xmm0,%xmm3,%xmm1 DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 68,15,40,124,36,16 ; movaps 0x10(%rsp),%xmm15 - DB 102,65,15,56,20,207 ; blendvps %xmm0,%xmm15,%xmm1 - DB 68,15,92,84,36,48 ; subps 0x30(%rsp),%xmm10 - DB 65,15,40,210 ; movaps %xmm10,%xmm2 - DB 65,15,92,208 ; subps %xmm8,%xmm2 - DB 69,15,40,232 ; movaps %xmm8,%xmm13 - DB 69,15,194,194,1 ; cmpltps %xmm10,%xmm8 - DB 65,15,40,226 ; movaps %xmm10,%xmm4 - DB 65,15,40,192 ; movaps %xmm8,%xmm0 - DB 102,15,56,20,226 ; blendvps %xmm0,%xmm2,%xmm4 - DB 65,15,40,194 ; movaps %xmm10,%xmm0 + DB 15,40,108,36,48 ; movaps 0x30(%rsp),%xmm5 + DB 102,15,56,20,205 ; blendvps %xmm0,%xmm5,%xmm1 + DB 68,15,88,29,224,42,0,0 ; addps 0x2ae0(%rip),%xmm11 # 3d70 <_sk_callback_sse41+0x397> + DB 69,15,194,203,1 ; cmpltps %xmm11,%xmm9 + DB 15,40,29,164,42,0,0 ; movaps 0x2aa4(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x367> + DB 65,15,88,219 ; addps %xmm11,%xmm3 + DB 69,15,40,235 ; movaps %xmm11,%xmm13 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 102,68,15,56,20,235 ; blendvps %xmm0,%xmm3,%xmm13 + DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 15,194,68,36,32,1 ; cmpltps 0x20(%rsp),%xmm0 - DB 69,15,88,234 ; addps %xmm10,%xmm13 - DB 102,65,15,56,20,229 ; blendvps %xmm0,%xmm13,%xmm4 - DB 68,15,89,214 ; mulps %xmm6,%xmm10 - DB 15,40,196 ; movaps %xmm4,%xmm0 - DB 15,194,195,1 ; cmpltps %xmm3,%xmm0 - DB 15,92,220 ; subps %xmm4,%xmm3 - DB 15,89,222 ; mulps %xmm6,%xmm3 - DB 69,15,88,209 ; addps %xmm9,%xmm10 - DB 65,15,88,217 ; addps %xmm9,%xmm3 - DB 102,68,15,56,20,203 ; blendvps %xmm0,%xmm3,%xmm9 - DB 15,40,196 ; movaps %xmm4,%xmm0 - DB 65,15,194,198,1 ; cmpltps %xmm14,%xmm0 - DB 102,68,15,56,20,207 ; blendvps %xmm0,%xmm7,%xmm9 - DB 15,40,196 ; movaps %xmm4,%xmm0 - DB 15,194,197,1 ; cmpltps %xmm5,%xmm0 - DB 102,69,15,56,20,202 ; blendvps %xmm0,%xmm10,%xmm9 + DB 15,40,29,97,42,0,0 ; movaps 0x2a61(%rip),%xmm3 # 3d20 <_sk_callback_sse41+0x347> + DB 65,15,88,219 ; addps %xmm11,%xmm3 + DB 102,68,15,56,20,235 ; blendvps %xmm0,%xmm3,%xmm13 + DB 65,15,92,229 ; subps %xmm13,%xmm4 + DB 69,15,40,205 ; movaps %xmm13,%xmm9 + DB 69,15,40,245 ; movaps %xmm13,%xmm14 + DB 68,15,194,108,36,16,1 ; cmpltps 0x10(%rsp),%xmm13 + DB 68,15,89,218 ; mulps %xmm2,%xmm11 + DB 15,89,226 ; mulps %xmm2,%xmm4 + DB 69,15,88,216 ; addps %xmm8,%xmm11 + DB 65,15,88,224 ; addps %xmm8,%xmm4 + DB 65,15,40,197 ; movaps %xmm13,%xmm0 + DB 102,68,15,56,20,196 ; blendvps %xmm0,%xmm4,%xmm8 + DB 68,15,194,52,36,1 ; cmpltps (%rsp),%xmm14 + DB 65,15,40,198 ; movaps %xmm14,%xmm0 + DB 102,68,15,56,20,198 ; blendvps %xmm0,%xmm6,%xmm8 + DB 68,15,194,207,1 ; cmpltps %xmm7,%xmm9 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 102,69,15,56,20,195 ; blendvps %xmm0,%xmm11,%xmm8 DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 102,69,15,56,20,207 ; blendvps %xmm0,%xmm15,%xmm9 + DB 102,68,15,56,20,197 ; blendvps %xmm0,%xmm5,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 65,15,40,195 ; movaps %xmm11,%xmm0 - DB 65,15,40,209 ; movaps %xmm9,%xmm2 + DB 65,15,40,194 ; movaps %xmm10,%xmm0 + DB 65,15,40,208 ; movaps %xmm8,%xmm2 DB 15,40,92,36,64 ; movaps 0x40(%rsp),%xmm3 DB 15,40,100,36,80 ; movaps 0x50(%rsp),%xmm4 DB 15,40,108,36,96 ; movaps 0x60(%rsp),%xmm5 @@ -11145,7 +11162,7 @@ _sk_scale_u8_sse41 LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 102,68,15,56,49,4,56 ; pmovzxbd (%rax,%rdi,1),%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,147,41,0,0 ; mulps 0x2993(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x33f> + DB 68,15,89,5,251,41,0,0 ; mulps 0x29fb(%rip),%xmm8 # 3d80 <_sk_callback_sse41+0x3a7> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -11179,7 +11196,7 @@ _sk_lerp_u8_sse41 LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 102,68,15,56,49,4,56 ; pmovzxbd (%rax,%rdi,1),%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,63,41,0,0 ; mulps 0x293f(%rip),%xmm8 # 3d10 <_sk_callback_sse41+0x34f> + DB 68,15,89,5,167,41,0,0 ; mulps 0x29a7(%rip),%xmm8 # 3d90 <_sk_callback_sse41+0x3b7> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -11200,17 +11217,17 @@ _sk_lerp_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 102,68,15,56,51,4,120 ; pmovzxwd (%rax,%rdi,2),%xmm8 - DB 102,15,111,29,15,41,0,0 ; movdqa 0x290f(%rip),%xmm3 # 3d20 <_sk_callback_sse41+0x35f> + DB 102,15,111,29,119,41,0,0 ; movdqa 0x2977(%rip),%xmm3 # 3da0 <_sk_callback_sse41+0x3c7> DB 102,65,15,219,216 ; pand %xmm8,%xmm3 DB 68,15,91,203 ; cvtdq2ps %xmm3,%xmm9 - DB 68,15,89,13,14,41,0,0 ; mulps 0x290e(%rip),%xmm9 # 3d30 <_sk_callback_sse41+0x36f> - DB 102,15,111,29,22,41,0,0 ; movdqa 0x2916(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x37f> + DB 68,15,89,13,118,41,0,0 ; mulps 0x2976(%rip),%xmm9 # 3db0 <_sk_callback_sse41+0x3d7> + DB 102,15,111,29,126,41,0,0 ; movdqa 0x297e(%rip),%xmm3 # 3dc0 <_sk_callback_sse41+0x3e7> DB 102,65,15,219,216 ; pand %xmm8,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,23,41,0,0 ; mulps 0x2917(%rip),%xmm3 # 3d50 <_sk_callback_sse41+0x38f> - DB 102,68,15,219,5,30,41,0,0 ; pand 0x291e(%rip),%xmm8 # 3d60 <_sk_callback_sse41+0x39f> + DB 15,89,29,127,41,0,0 ; mulps 0x297f(%rip),%xmm3 # 3dd0 <_sk_callback_sse41+0x3f7> + DB 102,68,15,219,5,134,41,0,0 ; pand 0x2986(%rip),%xmm8 # 3de0 <_sk_callback_sse41+0x407> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,34,41,0,0 ; mulps 0x2922(%rip),%xmm8 # 3d70 <_sk_callback_sse41+0x3af> + DB 68,15,89,5,138,41,0,0 ; mulps 0x298a(%rip),%xmm8 # 3df0 <_sk_callback_sse41+0x417> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -11221,7 +11238,7 @@ _sk_lerp_565_sse41 LABEL PROC DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 15,88,214 ; addps %xmm6,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,12,41,0,0 ; movaps 0x290c(%rip),%xmm3 # 3d80 <_sk_callback_sse41+0x3bf> + DB 15,40,29,116,41,0,0 ; movaps 0x2974(%rip),%xmm3 # 3e00 <_sk_callback_sse41+0x427> DB 255,224 ; jmpq *%rax PUBLIC _sk_load_tables_sse41 @@ -11230,7 +11247,7 @@ _sk_load_tables_sse41 LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 76,139,72,8 ; mov 0x8(%rax),%r9 DB 243,69,15,111,4,184 ; movdqu (%r8,%rdi,4),%xmm8 - DB 102,15,111,5,3,41,0,0 ; movdqa 0x2903(%rip),%xmm0 # 3d90 <_sk_callback_sse41+0x3cf> + DB 102,15,111,5,107,41,0,0 ; movdqa 0x296b(%rip),%xmm0 # 3e10 <_sk_callback_sse41+0x437> DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 DB 102,72,15,126,193 ; movq %xmm0,%rcx @@ -11245,7 +11262,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,15,58,33,193,48 ; insertps $0x30,%xmm1,%xmm0 DB 76,139,64,16 ; mov 0x10(%rax),%r8 DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 - DB 102,15,56,0,13,190,40,0,0 ; pshufb 0x28be(%rip),%xmm1 # 3da0 <_sk_callback_sse41+0x3df> + DB 102,15,56,0,13,38,41,0,0 ; pshufb 0x2926(%rip),%xmm1 # 3e20 <_sk_callback_sse41+0x447> DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,201 ; movq %xmm1,%rcx DB 68,15,182,209 ; movzbl %cl,%r10d @@ -11260,7 +11277,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 DB 76,139,64,24 ; mov 0x18(%rax),%r8 DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 - DB 102,15,56,0,21,122,40,0,0 ; pshufb 0x287a(%rip),%xmm2 # 3db0 <_sk_callback_sse41+0x3ef> + DB 102,15,56,0,21,226,40,0,0 ; pshufb 0x28e2(%rip),%xmm2 # 3e30 <_sk_callback_sse41+0x457> DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx DB 102,72,15,126,208 ; movq %xmm2,%rax DB 68,15,182,200 ; movzbl %al,%r9d @@ -11275,7 +11292,7 @@ _sk_load_tables_sse41 LABEL PROC 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 15,89,29,55,40,0,0 ; mulps 0x2837(%rip),%xmm3 # 3dc0 <_sk_callback_sse41+0x3ff> + DB 15,89,29,159,40,0,0 ; mulps 0x289f(%rip),%xmm3 # 3e40 <_sk_callback_sse41+0x467> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11292,7 +11309,7 @@ _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 102,68,15,111,5,10,40,0,0 ; movdqa 0x280a(%rip),%xmm8 # 3dd0 <_sk_callback_sse41+0x40f> + DB 102,68,15,111,5,114,40,0,0 ; movdqa 0x2872(%rip),%xmm8 # 3e50 <_sk_callback_sse41+0x477> 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 @@ -11309,7 +11326,7 @@ _sk_load_tables_u16_be_sse41 LABEL PROC DB 243,67,15,16,20,8 ; movss (%r8,%r9,1),%xmm2 DB 102,15,58,33,194,48 ; insertps $0x30,%xmm2,%xmm0 DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 102,15,56,0,13,189,39,0,0 ; pshufb 0x27bd(%rip),%xmm1 # 3de0 <_sk_callback_sse41+0x41f> + DB 102,15,56,0,13,37,40,0,0 ; pshufb 0x2825(%rip),%xmm1 # 3e60 <_sk_callback_sse41+0x487> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,201 ; movq %xmm1,%rcx @@ -11345,7 +11362,7 @@ _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 15,89,29,11,39,0,0 ; mulps 0x270b(%rip),%xmm3 # 3df0 <_sk_callback_sse41+0x42f> + DB 15,89,29,115,39,0,0 ; mulps 0x2773(%rip),%xmm3 # 3e70 <_sk_callback_sse41+0x497> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11365,7 +11382,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC 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 102,68,15,111,5,205,38,0,0 ; movdqa 0x26cd(%rip),%xmm8 # 3e00 <_sk_callback_sse41+0x43f> + DB 102,68,15,111,5,53,39,0,0 ; movdqa 0x2735(%rip),%xmm8 # 3e80 <_sk_callback_sse41+0x4a7> 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 @@ -11382,7 +11399,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 243,67,15,16,28,8 ; movss (%r8,%r9,1),%xmm3 DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0 DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 102,15,56,0,13,128,38,0,0 ; pshufb 0x2680(%rip),%xmm1 # 3e10 <_sk_callback_sse41+0x44f> + DB 102,15,56,0,13,232,38,0,0 ; pshufb 0x26e8(%rip),%xmm1 # 3e90 <_sk_callback_sse41+0x4b7> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,201 ; movq %xmm1,%rcx @@ -11413,7 +11430,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 243,65,15,16,28,8 ; movss (%r8,%rcx,1),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,235,37,0,0 ; movaps 0x25eb(%rip),%xmm3 # 3e20 <_sk_callback_sse41+0x45f> + DB 15,40,29,83,38,0,0 ; movaps 0x2653(%rip),%xmm3 # 3ea0 <_sk_callback_sse41+0x4c7> DB 255,224 ; jmpq *%rax PUBLIC _sk_byte_tables_sse41 @@ -11421,7 +11438,7 @@ _sk_byte_tables_sse41 LABEL PROC DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,236,37,0,0 ; movaps 0x25ec(%rip),%xmm8 # 3e30 <_sk_callback_sse41+0x46f> + DB 68,15,40,5,84,38,0,0 ; movaps 0x2654(%rip),%xmm8 # 3eb0 <_sk_callback_sse41+0x4d7> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx @@ -11440,7 +11457,7 @@ _sk_byte_tables_sse41 LABEL PROC DB 102,15,58,32,193,3 ; pinsrb $0x3,%ecx,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,157,37,0,0 ; movaps 0x259d(%rip),%xmm9 # 3e40 <_sk_callback_sse41+0x47f> + DB 68,15,40,13,5,38,0,0 ; movaps 0x2605(%rip),%xmm9 # 3ec0 <_sk_callback_sse41+0x4e7> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -11529,7 +11546,7 @@ _sk_byte_tables_rgb_sse41 LABEL PROC DB 102,15,58,32,193,3 ; pinsrb $0x3,%ecx,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,37,36,0,0 ; movaps 0x2425(%rip),%xmm9 # 3e50 <_sk_callback_sse41+0x48f> + DB 68,15,40,13,141,36,0,0 ; movaps 0x248d(%rip),%xmm9 # 3ed0 <_sk_callback_sse41+0x4f7> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -11696,31 +11713,31 @@ _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 68,15,89,5,124,33,0,0 ; mulps 0x217c(%rip),%xmm8 # 3e60 <_sk_callback_sse41+0x49f> - DB 68,15,84,21,132,33,0,0 ; andps 0x2184(%rip),%xmm10 # 3e70 <_sk_callback_sse41+0x4af> - DB 68,15,86,21,140,33,0,0 ; orps 0x218c(%rip),%xmm10 # 3e80 <_sk_callback_sse41+0x4bf> - DB 68,15,88,5,148,33,0,0 ; addps 0x2194(%rip),%xmm8 # 3e90 <_sk_callback_sse41+0x4cf> - DB 68,15,40,37,156,33,0,0 ; movaps 0x219c(%rip),%xmm12 # 3ea0 <_sk_callback_sse41+0x4df> + DB 68,15,89,5,228,33,0,0 ; mulps 0x21e4(%rip),%xmm8 # 3ee0 <_sk_callback_sse41+0x507> + DB 68,15,84,21,236,33,0,0 ; andps 0x21ec(%rip),%xmm10 # 3ef0 <_sk_callback_sse41+0x517> + DB 68,15,86,21,244,33,0,0 ; orps 0x21f4(%rip),%xmm10 # 3f00 <_sk_callback_sse41+0x527> + DB 68,15,88,5,252,33,0,0 ; addps 0x21fc(%rip),%xmm8 # 3f10 <_sk_callback_sse41+0x537> + DB 68,15,40,37,4,34,0,0 ; movaps 0x2204(%rip),%xmm12 # 3f20 <_sk_callback_sse41+0x547> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 68,15,88,21,156,33,0,0 ; addps 0x219c(%rip),%xmm10 # 3eb0 <_sk_callback_sse41+0x4ef> - DB 68,15,40,37,164,33,0,0 ; movaps 0x21a4(%rip),%xmm12 # 3ec0 <_sk_callback_sse41+0x4ff> + DB 68,15,88,21,4,34,0,0 ; addps 0x2204(%rip),%xmm10 # 3f30 <_sk_callback_sse41+0x557> + DB 68,15,40,37,12,34,0,0 ; movaps 0x220c(%rip),%xmm12 # 3f40 <_sk_callback_sse41+0x567> 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 68,15,88,5,145,33,0,0 ; addps 0x2191(%rip),%xmm8 # 3ed0 <_sk_callback_sse41+0x50f> - DB 68,15,40,21,153,33,0,0 ; movaps 0x2199(%rip),%xmm10 # 3ee0 <_sk_callback_sse41+0x51f> + DB 68,15,88,5,249,33,0,0 ; addps 0x21f9(%rip),%xmm8 # 3f50 <_sk_callback_sse41+0x577> + DB 68,15,40,21,1,34,0,0 ; movaps 0x2201(%rip),%xmm10 # 3f60 <_sk_callback_sse41+0x587> DB 69,15,89,211 ; mulps %xmm11,%xmm10 DB 69,15,92,194 ; subps %xmm10,%xmm8 - DB 68,15,40,21,153,33,0,0 ; movaps 0x2199(%rip),%xmm10 # 3ef0 <_sk_callback_sse41+0x52f> + DB 68,15,40,21,1,34,0,0 ; movaps 0x2201(%rip),%xmm10 # 3f70 <_sk_callback_sse41+0x597> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 68,15,40,29,157,33,0,0 ; movaps 0x219d(%rip),%xmm11 # 3f00 <_sk_callback_sse41+0x53f> + DB 68,15,40,29,5,34,0,0 ; movaps 0x2205(%rip),%xmm11 # 3f80 <_sk_callback_sse41+0x5a7> DB 69,15,94,218 ; divps %xmm10,%xmm11 DB 69,15,88,216 ; addps %xmm8,%xmm11 - DB 68,15,89,29,157,33,0,0 ; mulps 0x219d(%rip),%xmm11 # 3f10 <_sk_callback_sse41+0x54f> + DB 68,15,89,29,5,34,0,0 ; mulps 0x2205(%rip),%xmm11 # 3f90 <_sk_callback_sse41+0x5b7> 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 @@ -11728,7 +11745,7 @@ _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 68,15,93,5,132,33,0,0 ; minps 0x2184(%rip),%xmm8 # 3f20 <_sk_callback_sse41+0x55f> + DB 68,15,93,5,236,33,0,0 ; minps 0x21ec(%rip),%xmm8 # 3fa0 <_sk_callback_sse41+0x5c7> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -11756,31 +11773,31 @@ _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 68,15,89,37,37,33,0,0 ; mulps 0x2125(%rip),%xmm12 # 3f30 <_sk_callback_sse41+0x56f> - DB 68,15,84,29,45,33,0,0 ; andps 0x212d(%rip),%xmm11 # 3f40 <_sk_callback_sse41+0x57f> - DB 68,15,86,29,53,33,0,0 ; orps 0x2135(%rip),%xmm11 # 3f50 <_sk_callback_sse41+0x58f> - DB 68,15,88,37,61,33,0,0 ; addps 0x213d(%rip),%xmm12 # 3f60 <_sk_callback_sse41+0x59f> - DB 15,40,13,70,33,0,0 ; movaps 0x2146(%rip),%xmm1 # 3f70 <_sk_callback_sse41+0x5af> + DB 68,15,89,37,141,33,0,0 ; mulps 0x218d(%rip),%xmm12 # 3fb0 <_sk_callback_sse41+0x5d7> + DB 68,15,84,29,149,33,0,0 ; andps 0x2195(%rip),%xmm11 # 3fc0 <_sk_callback_sse41+0x5e7> + DB 68,15,86,29,157,33,0,0 ; orps 0x219d(%rip),%xmm11 # 3fd0 <_sk_callback_sse41+0x5f7> + DB 68,15,88,37,165,33,0,0 ; addps 0x21a5(%rip),%xmm12 # 3fe0 <_sk_callback_sse41+0x607> + DB 15,40,13,174,33,0,0 ; movaps 0x21ae(%rip),%xmm1 # 3ff0 <_sk_callback_sse41+0x617> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,88,29,70,33,0,0 ; addps 0x2146(%rip),%xmm11 # 3f80 <_sk_callback_sse41+0x5bf> - DB 15,40,13,79,33,0,0 ; movaps 0x214f(%rip),%xmm1 # 3f90 <_sk_callback_sse41+0x5cf> + DB 68,15,88,29,174,33,0,0 ; addps 0x21ae(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x627> + DB 15,40,13,183,33,0,0 ; movaps 0x21b7(%rip),%xmm1 # 4010 <_sk_callback_sse41+0x637> 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 68,15,88,37,60,33,0,0 ; addps 0x213c(%rip),%xmm12 # 3fa0 <_sk_callback_sse41+0x5df> - DB 15,40,13,69,33,0,0 ; movaps 0x2145(%rip),%xmm1 # 3fb0 <_sk_callback_sse41+0x5ef> + DB 68,15,88,37,164,33,0,0 ; addps 0x21a4(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x647> + DB 15,40,13,173,33,0,0 ; movaps 0x21ad(%rip),%xmm1 # 4030 <_sk_callback_sse41+0x657> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,40,21,69,33,0,0 ; movaps 0x2145(%rip),%xmm10 # 3fc0 <_sk_callback_sse41+0x5ff> + DB 68,15,40,21,173,33,0,0 ; movaps 0x21ad(%rip),%xmm10 # 4040 <_sk_callback_sse41+0x667> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,13,74,33,0,0 ; movaps 0x214a(%rip),%xmm1 # 3fd0 <_sk_callback_sse41+0x60f> + DB 15,40,13,178,33,0,0 ; movaps 0x21b2(%rip),%xmm1 # 4050 <_sk_callback_sse41+0x677> DB 65,15,94,202 ; divps %xmm10,%xmm1 DB 65,15,88,204 ; addps %xmm12,%xmm1 - DB 15,89,13,75,33,0,0 ; mulps 0x214b(%rip),%xmm1 # 3fe0 <_sk_callback_sse41+0x61f> + DB 15,89,13,179,33,0,0 ; mulps 0x21b3(%rip),%xmm1 # 4060 <_sk_callback_sse41+0x687> 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 @@ -11788,7 +11805,7 @@ _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 15,93,13,54,33,0,0 ; minps 0x2136(%rip),%xmm1 # 3ff0 <_sk_callback_sse41+0x62f> + DB 15,93,13,158,33,0,0 ; minps 0x219e(%rip),%xmm1 # 4070 <_sk_callback_sse41+0x697> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -11816,31 +11833,31 @@ _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 68,15,89,37,215,32,0,0 ; mulps 0x20d7(%rip),%xmm12 # 4000 <_sk_callback_sse41+0x63f> - DB 68,15,84,29,223,32,0,0 ; andps 0x20df(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x64f> - DB 68,15,86,29,231,32,0,0 ; orps 0x20e7(%rip),%xmm11 # 4020 <_sk_callback_sse41+0x65f> - DB 68,15,88,37,239,32,0,0 ; addps 0x20ef(%rip),%xmm12 # 4030 <_sk_callback_sse41+0x66f> - DB 15,40,21,248,32,0,0 ; movaps 0x20f8(%rip),%xmm2 # 4040 <_sk_callback_sse41+0x67f> + DB 68,15,89,37,63,33,0,0 ; mulps 0x213f(%rip),%xmm12 # 4080 <_sk_callback_sse41+0x6a7> + DB 68,15,84,29,71,33,0,0 ; andps 0x2147(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x6b7> + DB 68,15,86,29,79,33,0,0 ; orps 0x214f(%rip),%xmm11 # 40a0 <_sk_callback_sse41+0x6c7> + DB 68,15,88,37,87,33,0,0 ; addps 0x2157(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x6d7> + DB 15,40,21,96,33,0,0 ; movaps 0x2160(%rip),%xmm2 # 40c0 <_sk_callback_sse41+0x6e7> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,88,29,248,32,0,0 ; addps 0x20f8(%rip),%xmm11 # 4050 <_sk_callback_sse41+0x68f> - DB 15,40,21,1,33,0,0 ; movaps 0x2101(%rip),%xmm2 # 4060 <_sk_callback_sse41+0x69f> + DB 68,15,88,29,96,33,0,0 ; addps 0x2160(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x6f7> + DB 15,40,21,105,33,0,0 ; movaps 0x2169(%rip),%xmm2 # 40e0 <_sk_callback_sse41+0x707> 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 68,15,88,37,238,32,0,0 ; addps 0x20ee(%rip),%xmm12 # 4070 <_sk_callback_sse41+0x6af> - DB 15,40,21,247,32,0,0 ; movaps 0x20f7(%rip),%xmm2 # 4080 <_sk_callback_sse41+0x6bf> + DB 68,15,88,37,86,33,0,0 ; addps 0x2156(%rip),%xmm12 # 40f0 <_sk_callback_sse41+0x717> + DB 15,40,21,95,33,0,0 ; movaps 0x215f(%rip),%xmm2 # 4100 <_sk_callback_sse41+0x727> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,40,21,247,32,0,0 ; movaps 0x20f7(%rip),%xmm10 # 4090 <_sk_callback_sse41+0x6cf> + DB 68,15,40,21,95,33,0,0 ; movaps 0x215f(%rip),%xmm10 # 4110 <_sk_callback_sse41+0x737> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,21,252,32,0,0 ; movaps 0x20fc(%rip),%xmm2 # 40a0 <_sk_callback_sse41+0x6df> + DB 15,40,21,100,33,0,0 ; movaps 0x2164(%rip),%xmm2 # 4120 <_sk_callback_sse41+0x747> DB 65,15,94,210 ; divps %xmm10,%xmm2 DB 65,15,88,212 ; addps %xmm12,%xmm2 - DB 15,89,21,253,32,0,0 ; mulps 0x20fd(%rip),%xmm2 # 40b0 <_sk_callback_sse41+0x6ef> + DB 15,89,21,101,33,0,0 ; mulps 0x2165(%rip),%xmm2 # 4130 <_sk_callback_sse41+0x757> 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 @@ -11848,7 +11865,7 @@ _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 15,93,21,232,32,0,0 ; minps 0x20e8(%rip),%xmm2 # 40c0 <_sk_callback_sse41+0x6ff> + DB 15,93,21,80,33,0,0 ; minps 0x2150(%rip),%xmm2 # 4140 <_sk_callback_sse41+0x767> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -11876,31 +11893,31 @@ _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 68,15,89,37,137,32,0,0 ; mulps 0x2089(%rip),%xmm12 # 40d0 <_sk_callback_sse41+0x70f> - DB 68,15,84,29,145,32,0,0 ; andps 0x2091(%rip),%xmm11 # 40e0 <_sk_callback_sse41+0x71f> - DB 68,15,86,29,153,32,0,0 ; orps 0x2099(%rip),%xmm11 # 40f0 <_sk_callback_sse41+0x72f> - DB 68,15,88,37,161,32,0,0 ; addps 0x20a1(%rip),%xmm12 # 4100 <_sk_callback_sse41+0x73f> - DB 15,40,29,170,32,0,0 ; movaps 0x20aa(%rip),%xmm3 # 4110 <_sk_callback_sse41+0x74f> + DB 68,15,89,37,241,32,0,0 ; mulps 0x20f1(%rip),%xmm12 # 4150 <_sk_callback_sse41+0x777> + DB 68,15,84,29,249,32,0,0 ; andps 0x20f9(%rip),%xmm11 # 4160 <_sk_callback_sse41+0x787> + DB 68,15,86,29,1,33,0,0 ; orps 0x2101(%rip),%xmm11 # 4170 <_sk_callback_sse41+0x797> + DB 68,15,88,37,9,33,0,0 ; addps 0x2109(%rip),%xmm12 # 4180 <_sk_callback_sse41+0x7a7> + DB 15,40,29,18,33,0,0 ; movaps 0x2112(%rip),%xmm3 # 4190 <_sk_callback_sse41+0x7b7> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,88,29,170,32,0,0 ; addps 0x20aa(%rip),%xmm11 # 4120 <_sk_callback_sse41+0x75f> - DB 15,40,29,179,32,0,0 ; movaps 0x20b3(%rip),%xmm3 # 4130 <_sk_callback_sse41+0x76f> + DB 68,15,88,29,18,33,0,0 ; addps 0x2112(%rip),%xmm11 # 41a0 <_sk_callback_sse41+0x7c7> + DB 15,40,29,27,33,0,0 ; movaps 0x211b(%rip),%xmm3 # 41b0 <_sk_callback_sse41+0x7d7> 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 68,15,88,37,160,32,0,0 ; addps 0x20a0(%rip),%xmm12 # 4140 <_sk_callback_sse41+0x77f> - DB 15,40,29,169,32,0,0 ; movaps 0x20a9(%rip),%xmm3 # 4150 <_sk_callback_sse41+0x78f> + DB 68,15,88,37,8,33,0,0 ; addps 0x2108(%rip),%xmm12 # 41c0 <_sk_callback_sse41+0x7e7> + DB 15,40,29,17,33,0,0 ; movaps 0x2111(%rip),%xmm3 # 41d0 <_sk_callback_sse41+0x7f7> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,40,21,169,32,0,0 ; movaps 0x20a9(%rip),%xmm10 # 4160 <_sk_callback_sse41+0x79f> + DB 68,15,40,21,17,33,0,0 ; movaps 0x2111(%rip),%xmm10 # 41e0 <_sk_callback_sse41+0x807> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,29,174,32,0,0 ; movaps 0x20ae(%rip),%xmm3 # 4170 <_sk_callback_sse41+0x7af> + DB 15,40,29,22,33,0,0 ; movaps 0x2116(%rip),%xmm3 # 41f0 <_sk_callback_sse41+0x817> DB 65,15,94,218 ; divps %xmm10,%xmm3 DB 65,15,88,220 ; addps %xmm12,%xmm3 - DB 15,89,29,175,32,0,0 ; mulps 0x20af(%rip),%xmm3 # 4180 <_sk_callback_sse41+0x7bf> + DB 15,89,29,23,33,0,0 ; mulps 0x2117(%rip),%xmm3 # 4200 <_sk_callback_sse41+0x827> 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 @@ -11908,7 +11925,7 @@ _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 15,93,29,154,32,0,0 ; minps 0x209a(%rip),%xmm3 # 4190 <_sk_callback_sse41+0x7cf> + DB 15,93,29,2,33,0,0 ; minps 0x2102(%rip),%xmm3 # 4210 <_sk_callback_sse41+0x837> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -11916,29 +11933,29 @@ _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 68,15,89,5,150,32,0,0 ; mulps 0x2096(%rip),%xmm8 # 41a0 <_sk_callback_sse41+0x7df> - DB 68,15,40,13,158,32,0,0 ; movaps 0x209e(%rip),%xmm9 # 41b0 <_sk_callback_sse41+0x7ef> + DB 68,15,89,5,254,32,0,0 ; mulps 0x20fe(%rip),%xmm8 # 4220 <_sk_callback_sse41+0x847> + DB 68,15,40,13,6,33,0,0 ; movaps 0x2106(%rip),%xmm9 # 4230 <_sk_callback_sse41+0x857> DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 15,40,5,163,32,0,0 ; movaps 0x20a3(%rip),%xmm0 # 41c0 <_sk_callback_sse41+0x7ff> + DB 15,40,5,11,33,0,0 ; movaps 0x210b(%rip),%xmm0 # 4240 <_sk_callback_sse41+0x867> DB 15,88,200 ; addps %xmm0,%xmm1 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 68,15,88,5,161,32,0,0 ; addps 0x20a1(%rip),%xmm8 # 41d0 <_sk_callback_sse41+0x80f> - DB 68,15,89,5,169,32,0,0 ; mulps 0x20a9(%rip),%xmm8 # 41e0 <_sk_callback_sse41+0x81f> - DB 15,89,13,178,32,0,0 ; mulps 0x20b2(%rip),%xmm1 # 41f0 <_sk_callback_sse41+0x82f> + DB 68,15,88,5,9,33,0,0 ; addps 0x2109(%rip),%xmm8 # 4250 <_sk_callback_sse41+0x877> + DB 68,15,89,5,17,33,0,0 ; mulps 0x2111(%rip),%xmm8 # 4260 <_sk_callback_sse41+0x887> + DB 15,89,13,26,33,0,0 ; mulps 0x211a(%rip),%xmm1 # 4270 <_sk_callback_sse41+0x897> DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 15,89,21,183,32,0,0 ; mulps 0x20b7(%rip),%xmm2 # 4200 <_sk_callback_sse41+0x83f> + DB 15,89,21,31,33,0,0 ; mulps 0x211f(%rip),%xmm2 # 4280 <_sk_callback_sse41+0x8a7> 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 68,15,40,13,171,32,0,0 ; movaps 0x20ab(%rip),%xmm9 # 4210 <_sk_callback_sse41+0x84f> + DB 68,15,40,13,19,33,0,0 ; movaps 0x2113(%rip),%xmm9 # 4290 <_sk_callback_sse41+0x8b7> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 15,40,21,171,32,0,0 ; movaps 0x20ab(%rip),%xmm2 # 4220 <_sk_callback_sse41+0x85f> + DB 15,40,21,19,33,0,0 ; movaps 0x2113(%rip),%xmm2 # 42a0 <_sk_callback_sse41+0x8c7> DB 15,88,202 ; addps %xmm2,%xmm1 - DB 68,15,40,37,176,32,0,0 ; movaps 0x20b0(%rip),%xmm12 # 4230 <_sk_callback_sse41+0x86f> + DB 68,15,40,37,24,33,0,0 ; movaps 0x2118(%rip),%xmm12 # 42b0 <_sk_callback_sse41+0x8d7> 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 @@ -11957,8 +11974,8 @@ _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 15,89,13,105,32,0,0 ; mulps 0x2069(%rip),%xmm1 # 4240 <_sk_callback_sse41+0x87f> - DB 15,89,21,114,32,0,0 ; mulps 0x2072(%rip),%xmm2 # 4250 <_sk_callback_sse41+0x88f> + DB 15,89,13,209,32,0,0 ; mulps 0x20d1(%rip),%xmm1 # 42c0 <_sk_callback_sse41+0x8e7> + DB 15,89,21,218,32,0,0 ; mulps 0x20da(%rip),%xmm2 # 42d0 <_sk_callback_sse41+0x8f7> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,40,200 ; movaps %xmm8,%xmm1 @@ -11970,7 +11987,7 @@ _sk_load_a8_sse41 LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 102,15,56,49,4,56 ; pmovzxbd (%rax,%rdi,1),%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,98,32,0,0 ; mulps 0x2062(%rip),%xmm3 # 4260 <_sk_callback_sse41+0x89f> + DB 15,89,29,202,32,0,0 ; mulps 0x20ca(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x907> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 @@ -12001,7 +12018,7 @@ _sk_gather_a8_sse41 LABEL PROC 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 15,89,29,246,31,0,0 ; mulps 0x1ff6(%rip),%xmm3 # 4270 <_sk_callback_sse41+0x8af> + DB 15,89,29,94,32,0,0 ; mulps 0x205e(%rip),%xmm3 # 42f0 <_sk_callback_sse41+0x917> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -12012,7 +12029,7 @@ PUBLIC _sk_store_a8_sse41 _sk_store_a8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,234,31,0,0 ; movaps 0x1fea(%rip),%xmm8 # 4280 <_sk_callback_sse41+0x8bf> + DB 68,15,40,5,82,32,0,0 ; movaps 0x2052(%rip),%xmm8 # 4300 <_sk_callback_sse41+0x927> 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 @@ -12027,9 +12044,9 @@ _sk_load_g8_sse41 LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 102,15,56,49,4,56 ; pmovzxbd (%rax,%rdi,1),%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,199,31,0,0 ; mulps 0x1fc7(%rip),%xmm0 # 4290 <_sk_callback_sse41+0x8cf> + DB 15,89,5,47,32,0,0 ; mulps 0x202f(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x937> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,206,31,0,0 ; movaps 0x1fce(%rip),%xmm3 # 42a0 <_sk_callback_sse41+0x8df> + DB 15,40,29,54,32,0,0 ; movaps 0x2036(%rip),%xmm3 # 4320 <_sk_callback_sse41+0x947> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -12058,9 +12075,9 @@ _sk_gather_g8_sse41 LABEL PROC 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 15,89,5,103,31,0,0 ; mulps 0x1f67(%rip),%xmm0 # 42b0 <_sk_callback_sse41+0x8ef> + DB 15,89,5,207,31,0,0 ; mulps 0x1fcf(%rip),%xmm0 # 4330 <_sk_callback_sse41+0x957> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,110,31,0,0 ; movaps 0x1f6e(%rip),%xmm3 # 42c0 <_sk_callback_sse41+0x8ff> + DB 15,40,29,214,31,0,0 ; movaps 0x1fd6(%rip),%xmm3 # 4340 <_sk_callback_sse41+0x967> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -12070,9 +12087,9 @@ _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 2369 <_sk_gather_i8_sse41+0xf> + DB 116,5 ; je 2381 <_sk_gather_i8_sse41+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 236b <_sk_gather_i8_sse41+0x11> + DB 235,2 ; jmp 2383 <_sk_gather_i8_sse41+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 @@ -12103,17 +12120,17 @@ _sk_gather_i8_sse41 LABEL PROC DB 102,15,58,34,28,8,1 ; pinsrd $0x1,(%rax,%rcx,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 102,15,111,5,197,30,0,0 ; movdqa 0x1ec5(%rip),%xmm0 # 42d0 <_sk_callback_sse41+0x90f> + DB 102,15,111,5,45,31,0,0 ; movdqa 0x1f2d(%rip),%xmm0 # 4350 <_sk_callback_sse41+0x977> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,198,30,0,0 ; movaps 0x1ec6(%rip),%xmm8 # 42e0 <_sk_callback_sse41+0x91f> + DB 68,15,40,5,46,31,0,0 ; movaps 0x1f2e(%rip),%xmm8 # 4360 <_sk_callback_sse41+0x987> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,197,30,0,0 ; pshufb 0x1ec5(%rip),%xmm1 # 42f0 <_sk_callback_sse41+0x92f> + DB 102,15,56,0,13,45,31,0,0 ; pshufb 0x1f2d(%rip),%xmm1 # 4370 <_sk_callback_sse41+0x997> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,193,30,0,0 ; pshufb 0x1ec1(%rip),%xmm2 # 4300 <_sk_callback_sse41+0x93f> + DB 102,15,56,0,21,41,31,0,0 ; pshufb 0x1f29(%rip),%xmm2 # 4380 <_sk_callback_sse41+0x9a7> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -12127,19 +12144,19 @@ _sk_load_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 102,15,56,51,20,120 ; pmovzxwd (%rax,%rdi,2),%xmm2 - DB 102,15,111,5,167,30,0,0 ; movdqa 0x1ea7(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x94f> + DB 102,15,111,5,15,31,0,0 ; movdqa 0x1f0f(%rip),%xmm0 # 4390 <_sk_callback_sse41+0x9b7> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,169,30,0,0 ; mulps 0x1ea9(%rip),%xmm0 # 4320 <_sk_callback_sse41+0x95f> - DB 102,15,111,13,177,30,0,0 ; movdqa 0x1eb1(%rip),%xmm1 # 4330 <_sk_callback_sse41+0x96f> + DB 15,89,5,17,31,0,0 ; mulps 0x1f11(%rip),%xmm0 # 43a0 <_sk_callback_sse41+0x9c7> + DB 102,15,111,13,25,31,0,0 ; movdqa 0x1f19(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0x9d7> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,179,30,0,0 ; mulps 0x1eb3(%rip),%xmm1 # 4340 <_sk_callback_sse41+0x97f> - DB 102,15,219,21,187,30,0,0 ; pand 0x1ebb(%rip),%xmm2 # 4350 <_sk_callback_sse41+0x98f> + DB 15,89,13,27,31,0,0 ; mulps 0x1f1b(%rip),%xmm1 # 43c0 <_sk_callback_sse41+0x9e7> + DB 102,15,219,21,35,31,0,0 ; pand 0x1f23(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0x9f7> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,193,30,0,0 ; mulps 0x1ec1(%rip),%xmm2 # 4360 <_sk_callback_sse41+0x99f> + DB 15,89,21,41,31,0,0 ; mulps 0x1f29(%rip),%xmm2 # 43e0 <_sk_callback_sse41+0xa07> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,200,30,0,0 ; movaps 0x1ec8(%rip),%xmm3 # 4370 <_sk_callback_sse41+0x9af> + DB 15,40,29,48,31,0,0 ; movaps 0x1f30(%rip),%xmm3 # 43f0 <_sk_callback_sse41+0xa17> DB 255,224 ; jmpq *%rax PUBLIC _sk_gather_565_sse41 @@ -12165,31 +12182,31 @@ _sk_gather_565_sse41 LABEL PROC 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 102,15,111,5,109,30,0,0 ; movdqa 0x1e6d(%rip),%xmm0 # 4380 <_sk_callback_sse41+0x9bf> + DB 102,15,111,5,213,30,0,0 ; movdqa 0x1ed5(%rip),%xmm0 # 4400 <_sk_callback_sse41+0xa27> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,111,30,0,0 ; mulps 0x1e6f(%rip),%xmm0 # 4390 <_sk_callback_sse41+0x9cf> - DB 102,15,111,13,119,30,0,0 ; movdqa 0x1e77(%rip),%xmm1 # 43a0 <_sk_callback_sse41+0x9df> + DB 15,89,5,215,30,0,0 ; mulps 0x1ed7(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xa37> + DB 102,15,111,13,223,30,0,0 ; movdqa 0x1edf(%rip),%xmm1 # 4420 <_sk_callback_sse41+0xa47> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,121,30,0,0 ; mulps 0x1e79(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0x9ef> - DB 102,15,219,21,129,30,0,0 ; pand 0x1e81(%rip),%xmm2 # 43c0 <_sk_callback_sse41+0x9ff> + DB 15,89,13,225,30,0,0 ; mulps 0x1ee1(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xa57> + DB 102,15,219,21,233,30,0,0 ; pand 0x1ee9(%rip),%xmm2 # 4440 <_sk_callback_sse41+0xa67> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,135,30,0,0 ; mulps 0x1e87(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0xa0f> + DB 15,89,21,239,30,0,0 ; mulps 0x1eef(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xa77> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,142,30,0,0 ; movaps 0x1e8e(%rip),%xmm3 # 43e0 <_sk_callback_sse41+0xa1f> + DB 15,40,29,246,30,0,0 ; movaps 0x1ef6(%rip),%xmm3 # 4460 <_sk_callback_sse41+0xa87> 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 68,15,40,5,143,30,0,0 ; movaps 0x1e8f(%rip),%xmm8 # 43f0 <_sk_callback_sse41+0xa2f> + DB 68,15,40,5,247,30,0,0 ; movaps 0x1ef7(%rip),%xmm8 # 4470 <_sk_callback_sse41+0xa97> 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 68,15,40,21,132,30,0,0 ; movaps 0x1e84(%rip),%xmm10 # 4400 <_sk_callback_sse41+0xa3f> + DB 68,15,40,21,236,30,0,0 ; movaps 0x1eec(%rip),%xmm10 # 4480 <_sk_callback_sse41+0xaa7> 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 @@ -12207,21 +12224,21 @@ _sk_load_4444_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 102,15,56,51,28,120 ; pmovzxwd (%rax,%rdi,2),%xmm3 - DB 102,15,111,5,79,30,0,0 ; movdqa 0x1e4f(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xa4f> + DB 102,15,111,5,183,30,0,0 ; movdqa 0x1eb7(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xab7> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,81,30,0,0 ; mulps 0x1e51(%rip),%xmm0 # 4420 <_sk_callback_sse41+0xa5f> - DB 102,15,111,13,89,30,0,0 ; movdqa 0x1e59(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xa6f> + DB 15,89,5,185,30,0,0 ; mulps 0x1eb9(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xac7> + DB 102,15,111,13,193,30,0,0 ; movdqa 0x1ec1(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xad7> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,91,30,0,0 ; mulps 0x1e5b(%rip),%xmm1 # 4440 <_sk_callback_sse41+0xa7f> - DB 102,15,111,21,99,30,0,0 ; movdqa 0x1e63(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xa8f> + DB 15,89,13,195,30,0,0 ; mulps 0x1ec3(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xae7> + DB 102,15,111,21,203,30,0,0 ; movdqa 0x1ecb(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xaf7> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,101,30,0,0 ; mulps 0x1e65(%rip),%xmm2 # 4460 <_sk_callback_sse41+0xa9f> - DB 102,15,219,29,109,30,0,0 ; pand 0x1e6d(%rip),%xmm3 # 4470 <_sk_callback_sse41+0xaaf> + DB 15,89,21,205,30,0,0 ; mulps 0x1ecd(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xb07> + DB 102,15,219,29,213,30,0,0 ; pand 0x1ed5(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xb17> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,115,30,0,0 ; mulps 0x1e73(%rip),%xmm3 # 4480 <_sk_callback_sse41+0xabf> + DB 15,89,29,219,30,0,0 ; mulps 0x1edb(%rip),%xmm3 # 4500 <_sk_callback_sse41+0xb27> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -12248,21 +12265,21 @@ _sk_gather_4444_sse41 LABEL PROC 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 102,15,111,5,22,30,0,0 ; movdqa 0x1e16(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xacf> + DB 102,15,111,5,126,30,0,0 ; movdqa 0x1e7e(%rip),%xmm0 # 4510 <_sk_callback_sse41+0xb37> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,24,30,0,0 ; mulps 0x1e18(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xadf> - DB 102,15,111,13,32,30,0,0 ; movdqa 0x1e20(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xaef> + DB 15,89,5,128,30,0,0 ; mulps 0x1e80(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xb47> + DB 102,15,111,13,136,30,0,0 ; movdqa 0x1e88(%rip),%xmm1 # 4530 <_sk_callback_sse41+0xb57> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,34,30,0,0 ; mulps 0x1e22(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xaff> - DB 102,15,111,21,42,30,0,0 ; movdqa 0x1e2a(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xb0f> + DB 15,89,13,138,30,0,0 ; mulps 0x1e8a(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xb67> + DB 102,15,111,21,146,30,0,0 ; movdqa 0x1e92(%rip),%xmm2 # 4550 <_sk_callback_sse41+0xb77> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,44,30,0,0 ; mulps 0x1e2c(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xb1f> - DB 102,15,219,29,52,30,0,0 ; pand 0x1e34(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xb2f> + DB 15,89,21,148,30,0,0 ; mulps 0x1e94(%rip),%xmm2 # 4560 <_sk_callback_sse41+0xb87> + DB 102,15,219,29,156,30,0,0 ; pand 0x1e9c(%rip),%xmm3 # 4570 <_sk_callback_sse41+0xb97> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,58,30,0,0 ; mulps 0x1e3a(%rip),%xmm3 # 4500 <_sk_callback_sse41+0xb3f> + DB 15,89,29,162,30,0,0 ; mulps 0x1ea2(%rip),%xmm3 # 4580 <_sk_callback_sse41+0xba7> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -12270,7 +12287,7 @@ 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 68,15,40,5,57,30,0,0 ; movaps 0x1e39(%rip),%xmm8 # 4510 <_sk_callback_sse41+0xb4f> + DB 68,15,40,5,161,30,0,0 ; movaps 0x1ea1(%rip),%xmm8 # 4590 <_sk_callback_sse41+0xbb7> 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 @@ -12298,17 +12315,17 @@ _sk_load_8888_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 15,16,28,184 ; movups (%rax,%rdi,4),%xmm3 - DB 15,40,5,216,29,0,0 ; movaps 0x1dd8(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xb5f> + DB 15,40,5,64,30,0,0 ; movaps 0x1e40(%rip),%xmm0 # 45a0 <_sk_callback_sse41+0xbc7> DB 15,84,195 ; andps %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,218,29,0,0 ; movaps 0x1dda(%rip),%xmm8 # 4530 <_sk_callback_sse41+0xb6f> + DB 68,15,40,5,66,30,0,0 ; movaps 0x1e42(%rip),%xmm8 # 45b0 <_sk_callback_sse41+0xbd7> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,40,203 ; movaps %xmm3,%xmm1 - DB 102,15,56,0,13,218,29,0,0 ; pshufb 0x1dda(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xb7f> + DB 102,15,56,0,13,66,30,0,0 ; pshufb 0x1e42(%rip),%xmm1 # 45c0 <_sk_callback_sse41+0xbe7> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 15,40,211 ; movaps %xmm3,%xmm2 - DB 102,15,56,0,21,215,29,0,0 ; pshufb 0x1dd7(%rip),%xmm2 # 4550 <_sk_callback_sse41+0xb8f> + DB 102,15,56,0,21,63,30,0,0 ; pshufb 0x1e3f(%rip),%xmm2 # 45d0 <_sk_callback_sse41+0xbf7> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -12337,17 +12354,17 @@ _sk_gather_8888_sse41 LABEL PROC DB 102,65,15,58,34,28,129,1 ; pinsrd $0x1,(%r9,%rax,4),%xmm3 DB 102,67,15,58,34,28,145,2 ; pinsrd $0x2,(%r9,%r10,4),%xmm3 DB 102,65,15,58,34,28,137,3 ; pinsrd $0x3,(%r9,%rcx,4),%xmm3 - DB 102,15,111,5,112,29,0,0 ; movdqa 0x1d70(%rip),%xmm0 # 4560 <_sk_callback_sse41+0xb9f> + DB 102,15,111,5,216,29,0,0 ; movdqa 0x1dd8(%rip),%xmm0 # 45e0 <_sk_callback_sse41+0xc07> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,113,29,0,0 ; movaps 0x1d71(%rip),%xmm8 # 4570 <_sk_callback_sse41+0xbaf> + DB 68,15,40,5,217,29,0,0 ; movaps 0x1dd9(%rip),%xmm8 # 45f0 <_sk_callback_sse41+0xc17> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,112,29,0,0 ; pshufb 0x1d70(%rip),%xmm1 # 4580 <_sk_callback_sse41+0xbbf> + DB 102,15,56,0,13,216,29,0,0 ; pshufb 0x1dd8(%rip),%xmm1 # 4600 <_sk_callback_sse41+0xc27> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,108,29,0,0 ; pshufb 0x1d6c(%rip),%xmm2 # 4590 <_sk_callback_sse41+0xbcf> + DB 102,15,56,0,21,212,29,0,0 ; pshufb 0x1dd4(%rip),%xmm2 # 4610 <_sk_callback_sse41+0xc37> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -12360,7 +12377,7 @@ 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 68,15,40,5,88,29,0,0 ; movaps 0x1d58(%rip),%xmm8 # 45a0 <_sk_callback_sse41+0xbdf> + DB 68,15,40,5,192,29,0,0 ; movaps 0x1dc0(%rip),%xmm8 # 4620 <_sk_callback_sse41+0xc47> 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 @@ -12395,18 +12412,18 @@ _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 102,68,15,111,5,209,28,0,0 ; movdqa 0x1cd1(%rip),%xmm8 # 45b0 <_sk_callback_sse41+0xbef> + DB 102,68,15,111,5,57,29,0,0 ; movdqa 0x1d39(%rip),%xmm8 # 4630 <_sk_callback_sse41+0xc57> 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 102,15,111,29,204,28,0,0 ; movdqa 0x1ccc(%rip),%xmm3 # 45c0 <_sk_callback_sse41+0xbff> + DB 102,15,111,29,52,29,0,0 ; movdqa 0x1d34(%rip),%xmm3 # 4640 <_sk_callback_sse41+0xc67> 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 102,68,15,111,21,184,28,0,0 ; movdqa 0x1cb8(%rip),%xmm10 # 45d0 <_sk_callback_sse41+0xc0f> + DB 102,68,15,111,21,32,29,0,0 ; movdqa 0x1d20(%rip),%xmm10 # 4650 <_sk_callback_sse41+0xc77> 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 @@ -12477,18 +12494,18 @@ _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 102,68,15,111,5,118,27,0,0 ; movdqa 0x1b76(%rip),%xmm8 # 45e0 <_sk_callback_sse41+0xc1f> + DB 102,68,15,111,5,222,27,0,0 ; movdqa 0x1bde(%rip),%xmm8 # 4660 <_sk_callback_sse41+0xc87> 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 102,15,111,29,113,27,0,0 ; movdqa 0x1b71(%rip),%xmm3 # 45f0 <_sk_callback_sse41+0xc2f> + DB 102,15,111,29,217,27,0,0 ; movdqa 0x1bd9(%rip),%xmm3 # 4670 <_sk_callback_sse41+0xc97> 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 102,68,15,111,21,93,27,0,0 ; movdqa 0x1b5d(%rip),%xmm10 # 4600 <_sk_callback_sse41+0xc3f> + DB 102,68,15,111,21,197,27,0,0 ; movdqa 0x1bc5(%rip),%xmm10 # 4680 <_sk_callback_sse41+0xca7> 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 @@ -12534,17 +12551,17 @@ 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 102,68,15,111,21,147,26,0,0 ; movdqa 0x1a93(%rip),%xmm10 # 4610 <_sk_callback_sse41+0xc4f> + DB 102,68,15,111,21,251,26,0,0 ; movdqa 0x1afb(%rip),%xmm10 # 4690 <_sk_callback_sse41+0xcb7> DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 DB 102,69,15,219,234 ; pand %xmm10,%xmm13 DB 102,69,15,239,229 ; pxor %xmm13,%xmm12 - DB 102,68,15,111,13,134,26,0,0 ; movdqa 0x1a86(%rip),%xmm9 # 4620 <_sk_callback_sse41+0xc5f> + DB 102,68,15,111,13,238,26,0,0 ; movdqa 0x1aee(%rip),%xmm9 # 46a0 <_sk_callback_sse41+0xcc7> DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13 DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 DB 102,69,15,102,196 ; pcmpgtd %xmm12,%xmm8 DB 102,65,15,114,212,13 ; psrld $0xd,%xmm12 - DB 102,68,15,111,29,119,26,0,0 ; movdqa 0x1a77(%rip),%xmm11 # 4630 <_sk_callback_sse41+0xc6f> + DB 102,68,15,111,29,223,26,0,0 ; movdqa 0x1adf(%rip),%xmm11 # 46b0 <_sk_callback_sse41+0xcd7> DB 102,69,15,235,235 ; por %xmm11,%xmm13 DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 DB 102,69,15,223,197 ; pandn %xmm13,%xmm8 @@ -12612,7 +12629,7 @@ _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 68,15,40,5,70,25,0,0 ; movaps 0x1946(%rip),%xmm8 # 4640 <_sk_callback_sse41+0xc7f> + DB 68,15,40,5,174,25,0,0 ; movaps 0x19ae(%rip),%xmm8 # 46c0 <_sk_callback_sse41+0xce7> 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 @@ -12662,7 +12679,7 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 102,15,235,193 ; por %xmm1,%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,135,24,0,0 ; movaps 0x1887(%rip),%xmm8 # 4650 <_sk_callback_sse41+0xc8f> + DB 68,15,40,5,239,24,0,0 ; movaps 0x18ef(%rip),%xmm8 # 46d0 <_sk_callback_sse41+0xcf7> 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 @@ -12679,14 +12696,14 @@ _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 15,40,29,78,24,0,0 ; movaps 0x184e(%rip),%xmm3 # 4660 <_sk_callback_sse41+0xc9f> + DB 15,40,29,182,24,0,0 ; movaps 0x18b6(%rip),%xmm3 # 46e0 <_sk_callback_sse41+0xd07> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_u16_be_sse41 _sk_store_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,13,79,24,0,0 ; movaps 0x184f(%rip),%xmm9 # 4670 <_sk_callback_sse41+0xcaf> + DB 68,15,40,13,183,24,0,0 ; movaps 0x18b7(%rip),%xmm9 # 46f0 <_sk_callback_sse41+0xd17> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 @@ -12891,10 +12908,10 @@ _sk_mirror_y_sse41 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse41 _sk_luminance_to_alpha_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,109,21,0,0 ; mulps 0x156d(%rip),%xmm0 # 4680 <_sk_callback_sse41+0xcbf> - DB 15,89,13,118,21,0,0 ; mulps 0x1576(%rip),%xmm1 # 4690 <_sk_callback_sse41+0xccf> + DB 15,89,5,213,21,0,0 ; mulps 0x15d5(%rip),%xmm0 # 4700 <_sk_callback_sse41+0xd27> + DB 15,89,13,222,21,0,0 ; mulps 0x15de(%rip),%xmm1 # 4710 <_sk_callback_sse41+0xd37> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,124,21,0,0 ; mulps 0x157c(%rip),%xmm3 # 46a0 <_sk_callback_sse41+0xcdf> + DB 15,89,29,228,21,0,0 ; mulps 0x15e4(%rip),%xmm3 # 4720 <_sk_callback_sse41+0xd47> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -13117,7 +13134,7 @@ _sk_linear_gradient_sse41 LABEL PROC DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 DB 72,139,8 ; mov (%rax),%rcx DB 72,133,201 ; test %rcx,%rcx - DB 15,132,4,1,0,0 ; je 35cc <_sk_linear_gradient_sse41+0x13e> + DB 15,132,4,1,0,0 ; je 35e4 <_sk_linear_gradient_sse41+0x13e> DB 72,131,236,88 ; sub $0x58,%rsp DB 15,41,36,36 ; movaps %xmm4,(%rsp) DB 15,41,108,36,16 ; movaps %xmm5,0x10(%rsp) @@ -13168,13 +13185,13 @@ _sk_linear_gradient_sse41 LABEL PROC DB 15,40,196 ; movaps %xmm4,%xmm0 DB 72,131,192,36 ; add $0x24,%rax DB 72,255,201 ; dec %rcx - DB 15,133,65,255,255,255 ; jne 34f4 <_sk_linear_gradient_sse41+0x66> + DB 15,133,65,255,255,255 ; jne 350c <_sk_linear_gradient_sse41+0x66> DB 15,40,124,36,48 ; movaps 0x30(%rsp),%xmm7 DB 15,40,116,36,32 ; movaps 0x20(%rsp),%xmm6 DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5 DB 15,40,36,36 ; movaps (%rsp),%xmm4 DB 72,131,196,88 ; add $0x58,%rsp - DB 235,13 ; jmp 35d9 <_sk_linear_gradient_sse41+0x14b> + DB 235,13 ; jmp 35f1 <_sk_linear_gradient_sse41+0x14b> DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 DB 15,87,219 ; xorps %xmm3,%xmm3 @@ -13225,7 +13242,7 @@ _sk_linear_gradient_2stops_sse41 LABEL PROC PUBLIC _sk_save_xy_sse41 _sk_save_xy_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,56,16,0,0 ; movaps 0x1038(%rip),%xmm8 # 46b0 <_sk_callback_sse41+0xcef> + DB 68,15,40,5,160,16,0,0 ; movaps 0x10a0(%rip),%xmm8 # 4730 <_sk_callback_sse41+0xd57> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -13265,8 +13282,8 @@ _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 15,88,5,186,15,0,0 ; addps 0xfba(%rip),%xmm0 # 46c0 <_sk_callback_sse41+0xcff> - DB 68,15,40,13,194,15,0,0 ; movaps 0xfc2(%rip),%xmm9 # 46d0 <_sk_callback_sse41+0xd0f> + DB 15,88,5,34,16,0,0 ; addps 0x1022(%rip),%xmm0 # 4740 <_sk_callback_sse41+0xd67> + DB 68,15,40,13,42,16,0,0 ; movaps 0x102a(%rip),%xmm9 # 4750 <_sk_callback_sse41+0xd77> 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 @@ -13277,7 +13294,7 @@ _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 15,88,5,177,15,0,0 ; addps 0xfb1(%rip),%xmm0 # 46e0 <_sk_callback_sse41+0xd1f> + DB 15,88,5,25,16,0,0 ; addps 0x1019(%rip),%xmm0 # 4760 <_sk_callback_sse41+0xd87> 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 @@ -13287,8 +13304,8 @@ _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 15,88,13,163,15,0,0 ; addps 0xfa3(%rip),%xmm1 # 46f0 <_sk_callback_sse41+0xd2f> - DB 68,15,40,13,171,15,0,0 ; movaps 0xfab(%rip),%xmm9 # 4700 <_sk_callback_sse41+0xd3f> + DB 15,88,13,11,16,0,0 ; addps 0x100b(%rip),%xmm1 # 4770 <_sk_callback_sse41+0xd97> + DB 68,15,40,13,19,16,0,0 ; movaps 0x1013(%rip),%xmm9 # 4780 <_sk_callback_sse41+0xda7> 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 @@ -13299,7 +13316,7 @@ _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 15,88,13,153,15,0,0 ; addps 0xf99(%rip),%xmm1 # 4710 <_sk_callback_sse41+0xd4f> + DB 15,88,13,1,16,0,0 ; addps 0x1001(%rip),%xmm1 # 4790 <_sk_callback_sse41+0xdb7> 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 @@ -13309,13 +13326,13 @@ _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 15,88,5,140,15,0,0 ; addps 0xf8c(%rip),%xmm0 # 4720 <_sk_callback_sse41+0xd5f> - DB 68,15,40,13,148,15,0,0 ; movaps 0xf94(%rip),%xmm9 # 4730 <_sk_callback_sse41+0xd6f> + DB 15,88,5,244,15,0,0 ; addps 0xff4(%rip),%xmm0 # 47a0 <_sk_callback_sse41+0xdc7> + DB 68,15,40,13,252,15,0,0 ; movaps 0xffc(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0xdd7> 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 68,15,89,13,144,15,0,0 ; mulps 0xf90(%rip),%xmm9 # 4740 <_sk_callback_sse41+0xd7f> - DB 68,15,88,13,152,15,0,0 ; addps 0xf98(%rip),%xmm9 # 4750 <_sk_callback_sse41+0xd8f> + DB 68,15,89,13,248,15,0,0 ; mulps 0xff8(%rip),%xmm9 # 47c0 <_sk_callback_sse41+0xde7> + DB 68,15,88,13,0,16,0,0 ; addps 0x1000(%rip),%xmm9 # 47d0 <_sk_callback_sse41+0xdf7> 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 @@ -13326,16 +13343,16 @@ _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 15,88,5,135,15,0,0 ; addps 0xf87(%rip),%xmm0 # 4760 <_sk_callback_sse41+0xd9f> - DB 68,15,40,13,143,15,0,0 ; movaps 0xf8f(%rip),%xmm9 # 4770 <_sk_callback_sse41+0xdaf> + DB 15,88,5,239,15,0,0 ; addps 0xfef(%rip),%xmm0 # 47e0 <_sk_callback_sse41+0xe07> + DB 68,15,40,13,247,15,0,0 ; movaps 0xff7(%rip),%xmm9 # 47f0 <_sk_callback_sse41+0xe17> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,147,15,0,0 ; movaps 0xf93(%rip),%xmm8 # 4780 <_sk_callback_sse41+0xdbf> + DB 68,15,40,5,251,15,0,0 ; movaps 0xffb(%rip),%xmm8 # 4800 <_sk_callback_sse41+0xe27> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,151,15,0,0 ; addps 0xf97(%rip),%xmm8 # 4790 <_sk_callback_sse41+0xdcf> + DB 68,15,88,5,255,15,0,0 ; addps 0xfff(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xe37> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,155,15,0,0 ; addps 0xf9b(%rip),%xmm8 # 47a0 <_sk_callback_sse41+0xddf> + DB 68,15,88,5,3,16,0,0 ; addps 0x1003(%rip),%xmm8 # 4820 <_sk_callback_sse41+0xe47> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,159,15,0,0 ; addps 0xf9f(%rip),%xmm8 # 47b0 <_sk_callback_sse41+0xdef> + DB 68,15,88,5,7,16,0,0 ; addps 0x1007(%rip),%xmm8 # 4830 <_sk_callback_sse41+0xe57> 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 @@ -13343,17 +13360,17 @@ _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 68,15,40,5,153,15,0,0 ; movaps 0xf99(%rip),%xmm8 # 47c0 <_sk_callback_sse41+0xdff> + DB 68,15,40,5,1,16,0,0 ; movaps 0x1001(%rip),%xmm8 # 4840 <_sk_callback_sse41+0xe67> 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 68,15,40,21,149,15,0,0 ; movaps 0xf95(%rip),%xmm10 # 47d0 <_sk_callback_sse41+0xe0f> + DB 68,15,40,21,253,15,0,0 ; movaps 0xffd(%rip),%xmm10 # 4850 <_sk_callback_sse41+0xe77> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,153,15,0,0 ; addps 0xf99(%rip),%xmm10 # 47e0 <_sk_callback_sse41+0xe1f> + DB 68,15,88,21,1,16,0,0 ; addps 0x1001(%rip),%xmm10 # 4860 <_sk_callback_sse41+0xe87> 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 68,15,88,21,149,15,0,0 ; addps 0xf95(%rip),%xmm10 # 47f0 <_sk_callback_sse41+0xe2f> + DB 68,15,88,21,253,15,0,0 ; addps 0xffd(%rip),%xmm10 # 4870 <_sk_callback_sse41+0xe97> 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 @@ -13363,11 +13380,11 @@ _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 15,88,5,136,15,0,0 ; addps 0xf88(%rip),%xmm0 # 4800 <_sk_callback_sse41+0xe3f> + DB 15,88,5,240,15,0,0 ; addps 0xff0(%rip),%xmm0 # 4880 <_sk_callback_sse41+0xea7> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,136,15,0,0 ; mulps 0xf88(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xe4f> - DB 68,15,88,5,144,15,0,0 ; addps 0xf90(%rip),%xmm8 # 4820 <_sk_callback_sse41+0xe5f> + DB 68,15,89,5,240,15,0,0 ; mulps 0xff0(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xeb7> + DB 68,15,88,5,248,15,0,0 ; addps 0xff8(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xec7> 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 @@ -13378,13 +13395,13 @@ _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 15,88,13,126,15,0,0 ; addps 0xf7e(%rip),%xmm1 # 4830 <_sk_callback_sse41+0xe6f> - DB 68,15,40,13,134,15,0,0 ; movaps 0xf86(%rip),%xmm9 # 4840 <_sk_callback_sse41+0xe7f> + DB 15,88,13,230,15,0,0 ; addps 0xfe6(%rip),%xmm1 # 48b0 <_sk_callback_sse41+0xed7> + DB 68,15,40,13,238,15,0,0 ; movaps 0xfee(%rip),%xmm9 # 48c0 <_sk_callback_sse41+0xee7> 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 68,15,89,13,130,15,0,0 ; mulps 0xf82(%rip),%xmm9 # 4850 <_sk_callback_sse41+0xe8f> - DB 68,15,88,13,138,15,0,0 ; addps 0xf8a(%rip),%xmm9 # 4860 <_sk_callback_sse41+0xe9f> + DB 68,15,89,13,234,15,0,0 ; mulps 0xfea(%rip),%xmm9 # 48d0 <_sk_callback_sse41+0xef7> + DB 68,15,88,13,242,15,0,0 ; addps 0xff2(%rip),%xmm9 # 48e0 <_sk_callback_sse41+0xf07> 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 @@ -13395,16 +13412,16 @@ _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 15,88,13,120,15,0,0 ; addps 0xf78(%rip),%xmm1 # 4870 <_sk_callback_sse41+0xeaf> - DB 68,15,40,13,128,15,0,0 ; movaps 0xf80(%rip),%xmm9 # 4880 <_sk_callback_sse41+0xebf> + DB 15,88,13,224,15,0,0 ; addps 0xfe0(%rip),%xmm1 # 48f0 <_sk_callback_sse41+0xf17> + DB 68,15,40,13,232,15,0,0 ; movaps 0xfe8(%rip),%xmm9 # 4900 <_sk_callback_sse41+0xf27> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,132,15,0,0 ; movaps 0xf84(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xecf> + DB 68,15,40,5,236,15,0,0 ; movaps 0xfec(%rip),%xmm8 # 4910 <_sk_callback_sse41+0xf37> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,136,15,0,0 ; addps 0xf88(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xedf> + DB 68,15,88,5,240,15,0,0 ; addps 0xff0(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xf47> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,140,15,0,0 ; addps 0xf8c(%rip),%xmm8 # 48b0 <_sk_callback_sse41+0xeef> + DB 68,15,88,5,244,15,0,0 ; addps 0xff4(%rip),%xmm8 # 4930 <_sk_callback_sse41+0xf57> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,144,15,0,0 ; addps 0xf90(%rip),%xmm8 # 48c0 <_sk_callback_sse41+0xeff> + DB 68,15,88,5,248,15,0,0 ; addps 0xff8(%rip),%xmm8 # 4940 <_sk_callback_sse41+0xf67> 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 @@ -13412,17 +13429,17 @@ _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 68,15,40,5,138,15,0,0 ; movaps 0xf8a(%rip),%xmm8 # 48d0 <_sk_callback_sse41+0xf0f> + DB 68,15,40,5,242,15,0,0 ; movaps 0xff2(%rip),%xmm8 # 4950 <_sk_callback_sse41+0xf77> 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 68,15,40,21,133,15,0,0 ; movaps 0xf85(%rip),%xmm10 # 48e0 <_sk_callback_sse41+0xf1f> + DB 68,15,40,21,237,15,0,0 ; movaps 0xfed(%rip),%xmm10 # 4960 <_sk_callback_sse41+0xf87> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,137,15,0,0 ; addps 0xf89(%rip),%xmm10 # 48f0 <_sk_callback_sse41+0xf2f> + DB 68,15,88,21,241,15,0,0 ; addps 0xff1(%rip),%xmm10 # 4970 <_sk_callback_sse41+0xf97> 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 68,15,88,21,133,15,0,0 ; addps 0xf85(%rip),%xmm10 # 4900 <_sk_callback_sse41+0xf3f> + DB 68,15,88,21,237,15,0,0 ; addps 0xfed(%rip),%xmm10 # 4980 <_sk_callback_sse41+0xfa7> 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 @@ -13432,11 +13449,11 @@ _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 15,88,13,119,15,0,0 ; addps 0xf77(%rip),%xmm1 # 4910 <_sk_callback_sse41+0xf4f> + DB 15,88,13,223,15,0,0 ; addps 0xfdf(%rip),%xmm1 # 4990 <_sk_callback_sse41+0xfb7> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,119,15,0,0 ; mulps 0xf77(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xf5f> - DB 68,15,88,5,127,15,0,0 ; addps 0xf7f(%rip),%xmm8 # 4930 <_sk_callback_sse41+0xf6f> + DB 68,15,89,5,223,15,0,0 ; mulps 0xfdf(%rip),%xmm8 # 49a0 <_sk_callback_sse41+0xfc7> + DB 68,15,88,5,231,15,0,0 ; addps 0xfe7(%rip),%xmm8 # 49b0 <_sk_callback_sse41+0xfd7> 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 @@ -13607,11 +13624,11 @@ ALIGN 16 DB 0,128,191,0,0,128 ; add %al,-0x7fffff41(%rax) DB 191,0,0,224,64 ; mov $0x40e00000,%edi DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 3bf8 <.literal16+0x188> + DB 224,64 ; loopne 3c18 <.literal16+0x188> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 3bfc <.literal16+0x18c> + DB 224,64 ; loopne 3c1c <.literal16+0x18c> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 3c00 <.literal16+0x190> + DB 224,64 ; loopne 3c20 <.literal16+0x190> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -13750,6 +13767,49 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + DB 0,0 ; add %al,(%rax) + DB 128,63,171 ; cmpb $0xab,(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,171 ; ds stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,171 ; ds stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,171 ; ds stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,0,0 ; add %al,%ds:(%rax) + DB 128,191,0,0,128,191,0 ; cmpb $0x0,-0x40800000(%rdi) + DB 0,128,191,0,0,128 ; add %al,-0x7fffff41(%rax) + DB 191,0,0,192,64 ; mov $0x40c00000,%edi + DB 0,0 ; add %al,(%rax) + DB 192,64,0,0 ; rolb $0x0,0x0(%rax) + DB 192,64,0,0 ; rolb $0x0,0x0(%rax) + DB 192,64,171,170 ; rolb $0xaa,-0x55(%rax) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 190,171,170,170,190 ; mov $0xbeaaaaab,%esi + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 190,171,170,170,190 ; mov $0xbeaaaaab,%esi 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) @@ -13771,13 +13831,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 3d49 <.literal16+0x2d9> + DB 224,7 ; loopne 3dc9 <.literal16+0x339> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 3d4d <.literal16+0x2dd> + DB 224,7 ; loopne 3dcd <.literal16+0x33d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 3d51 <.literal16+0x2e1> + DB 224,7 ; loopne 3dd1 <.literal16+0x341> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 3d55 <.literal16+0x2e5> + DB 224,7 ; loopne 3dd5 <.literal16+0x345> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -13817,10 +13877,10 @@ ALIGN 16 DB 0,1 ; add %al,(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a003da8 <_sk_callback_sse41+0xa0003e7> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a003e28 <_sk_callback_sse41+0xa00044f> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3003db0 <_sk_callback_sse41+0x30003ef> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3003e30 <_sk_callback_sse41+0x3000457> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -13875,11 +13935,11 @@ 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 3e7b <.literal16+0x40b> + DB 127,67 ; jg 3efb <.literal16+0x46b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 3e7f <.literal16+0x40f> + DB 127,67 ; jg 3eff <.literal16+0x46f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 3e83 <.literal16+0x413> + DB 127,67 ; jg 3f03 <.literal16+0x473> 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) @@ -13894,16 +13954,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 3e74 <.literal16+0x404> + DB 127,0 ; jg 3ef4 <.literal16+0x464> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 3e78 <.literal16+0x408> + DB 127,0 ; jg 3ef8 <.literal16+0x468> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 3e7c <.literal16+0x40c> + DB 127,0 ; jg 3efc <.literal16+0x46c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 3e80 <.literal16+0x410> + DB 127,0 ; jg 3f00 <.literal16+0x470> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -13912,7 +13972,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 3f05 <.literal16+0x495> + DB 119,115 ; ja 3f85 <.literal16+0x4f5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -13923,7 +13983,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 3e69 <.literal16+0x3f9> + DB 117,191 ; jne 3ee9 <.literal16+0x459> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -13935,7 +13995,7 @@ 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 ffffffffe9a37eaa <_sk_callback_sse41+0xffffffffe9a344e9> + DB 233,220,63,163,233 ; jmpq ffffffffe9a37f2a <_sk_callback_sse41+0xffffffffe9a34551> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -13990,16 +14050,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 3f44 <.literal16+0x4d4> + DB 127,0 ; jg 3fc4 <.literal16+0x534> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 3f48 <.literal16+0x4d8> + DB 127,0 ; jg 3fc8 <.literal16+0x538> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 3f4c <.literal16+0x4dc> + DB 127,0 ; jg 3fcc <.literal16+0x53c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 3f50 <.literal16+0x4e0> + DB 127,0 ; jg 3fd0 <.literal16+0x540> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -14008,7 +14068,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 3fd5 <.literal16+0x565> + DB 119,115 ; ja 4055 <.literal16+0x5c5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -14019,7 +14079,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 3f39 <.literal16+0x4c9> + DB 117,191 ; jne 3fb9 <.literal16+0x529> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -14031,7 +14091,7 @@ 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 ffffffffe9a37f7a <_sk_callback_sse41+0xffffffffe9a345b9> + DB 233,220,63,163,233 ; jmpq ffffffffe9a37ffa <_sk_callback_sse41+0xffffffffe9a34621> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -14086,16 +14146,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 4014 <.literal16+0x5a4> + DB 127,0 ; jg 4094 <.literal16+0x604> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 4018 <.literal16+0x5a8> + DB 127,0 ; jg 4098 <.literal16+0x608> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 401c <.literal16+0x5ac> + DB 127,0 ; jg 409c <.literal16+0x60c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 4020 <.literal16+0x5b0> + DB 127,0 ; jg 40a0 <.literal16+0x610> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -14104,7 +14164,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 40a5 <.literal16+0x635> + DB 119,115 ; ja 4125 <.literal16+0x695> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -14115,7 +14175,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 4009 <.literal16+0x599> + DB 117,191 ; jne 4089 <.literal16+0x5f9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -14127,7 +14187,7 @@ 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 ffffffffe9a3804a <_sk_callback_sse41+0xffffffffe9a34689> + DB 233,220,63,163,233 ; jmpq ffffffffe9a380ca <_sk_callback_sse41+0xffffffffe9a346f1> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -14182,16 +14242,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 40e4 <.literal16+0x674> + DB 127,0 ; jg 4164 <.literal16+0x6d4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 40e8 <.literal16+0x678> + DB 127,0 ; jg 4168 <.literal16+0x6d8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 40ec <.literal16+0x67c> + DB 127,0 ; jg 416c <.literal16+0x6dc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 40f0 <.literal16+0x680> + DB 127,0 ; jg 4170 <.literal16+0x6e0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -14200,7 +14260,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4175 <.literal16+0x705> + DB 119,115 ; ja 41f5 <.literal16+0x765> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -14211,7 +14271,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 40d9 <.literal16+0x669> + DB 117,191 ; jne 4159 <.literal16+0x6c9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -14223,7 +14283,7 @@ 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 ffffffffe9a3811a <_sk_callback_sse41+0xffffffffe9a34759> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3819a <_sk_callback_sse41+0xffffffffe9a347c1> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -14274,13 +14334,13 @@ 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 41f7 <.literal16+0x787> + DB 127,67 ; jg 4277 <.literal16+0x7e7> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 41fb <.literal16+0x78b> + DB 127,67 ; jg 427b <.literal16+0x7eb> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 41ff <.literal16+0x78f> + DB 127,67 ; jg 427f <.literal16+0x7ef> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 4203 <.literal16+0x793> + DB 127,67 ; jg 4283 <.literal16+0x7f3> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -14327,16 +14387,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 4283 <.literal16+0x813> + DB 118,63 ; jbe 4303 <.literal16+0x873> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 4287 <.literal16+0x817> + DB 118,63 ; jbe 4307 <.literal16+0x877> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 428b <.literal16+0x81b> + DB 118,63 ; jbe 430b <.literal16+0x87b> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 428f <.literal16+0x81f> + DB 118,63 ; jbe 430f <.literal16+0x87f> 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) @@ -14348,11 +14408,11 @@ 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 42cb <.literal16+0x85b> + DB 127,67 ; jg 434b <.literal16+0x8bb> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 42cf <.literal16+0x85f> + DB 127,67 ; jg 434f <.literal16+0x8bf> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 42d3 <.literal16+0x863> + DB 127,67 ; jg 4353 <.literal16+0x8c3> DB 129,128,128,59,129,128,128,59,129,128; addl $0x80813b80,-0x7f7ec480(%rax) DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,0,0,128,63 ; addb $0x3f,-0x7fffffc5(%rax) @@ -14381,7 +14441,7 @@ 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) # 3004300 <_sk_callback_sse41+0x300093f> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004380 <_sk_callback_sse41+0x30009a7> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -14410,13 +14470,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4339 <.literal16+0x8c9> + DB 224,7 ; loopne 43b9 <.literal16+0x929> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 433d <.literal16+0x8cd> + DB 224,7 ; loopne 43bd <.literal16+0x92d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 4341 <.literal16+0x8d1> + DB 224,7 ; loopne 43c1 <.literal16+0x931> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 4345 <.literal16+0x8d5> + DB 224,7 ; loopne 43c5 <.literal16+0x935> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -14462,13 +14522,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 43a9 <.literal16+0x939> + DB 224,7 ; loopne 4429 <.literal16+0x999> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 43ad <.literal16+0x93d> + DB 224,7 ; loopne 442d <.literal16+0x99d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 43b1 <.literal16+0x941> + DB 224,7 ; loopne 4431 <.literal16+0x9a1> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 43b5 <.literal16+0x945> + DB 224,7 ; loopne 4435 <.literal16+0x9a5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -14506,13 +14566,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 4446 <.literal16+0x9d6> + DB 124,66 ; jl 44c6 <.literal16+0xa36> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 444a <.literal16+0x9da> + DB 124,66 ; jl 44ca <.literal16+0xa3a> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 444e <.literal16+0x9de> + DB 124,66 ; jl 44ce <.literal16+0xa3e> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 4452 <.literal16+0x9e2> + DB 124,66 ; jl 44d2 <.literal16+0xa42> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -14602,13 +14662,13 @@ 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 4555 <.literal16+0xae5> + DB 112,65 ; jo 45d5 <.literal16+0xb45> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 4559 <.literal16+0xae9> + DB 112,65 ; jo 45d9 <.literal16+0xb49> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 455d <.literal16+0xaed> + DB 112,65 ; jo 45dd <.literal16+0xb4d> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 4561 <.literal16+0xaf1> + DB 112,65 ; jo 45e1 <.literal16+0xb51> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -14623,7 +14683,7 @@ 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) # 3004550 <_sk_callback_sse41+0x3000b8f> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30045d0 <_sk_callback_sse41+0x3000bf7> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -14650,7 +14710,7 @@ 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) # 3004590 <_sk_callback_sse41+0x3000bcf> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004610 <_sk_callback_sse41+0x3000c37> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -14665,11 +14725,11 @@ 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 45eb <.literal16+0xb7b> + DB 127,67 ; jg 466b <.literal16+0xbdb> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 45ef <.literal16+0xb7f> + DB 127,67 ; jg 466f <.literal16+0xbdf> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 45f3 <.literal16+0xb83> + DB 127,67 ; jg 4673 <.literal16+0xbe3> 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) @@ -14745,13 +14805,13 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 46bb <.literal16+0xc4b> + DB 127,71 ; jg 473b <.literal16+0xcab> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 46bf <.literal16+0xc4f> + DB 127,71 ; jg 473f <.literal16+0xcaf> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 46c3 <.literal16+0xc53> + DB 127,71 ; jg 4743 <.literal16+0xcb3> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 46c7 <.literal16+0xc57> + DB 127,71 ; jg 4747 <.literal16+0xcb7> DB 208 ; (bad) DB 179,89 ; mov $0x59,%bl DB 62,208 ; ds (bad) @@ -14835,11 +14895,11 @@ 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 4762 <.literal16+0xcf2> + DB 62,114,28 ; jb,pt 47e2 <.literal16+0xd52> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4766 <.literal16+0xcf6> + DB 62,114,28 ; jb,pt 47e6 <.literal16+0xd56> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 476a <.literal16+0xcfa> + DB 62,114,28 ; jb,pt 47ea <.literal16+0xd5a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -14883,7 +14943,7 @@ 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 # 3d63d5f5 <_sk_callback_sse41+0x3d639c34> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d675 <_sk_callback_sse41+0x3d639c9c> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -14909,7 +14969,7 @@ 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 # 3d63d635 <_sk_callback_sse41+0x3d639c74> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d6b5 <_sk_callback_sse41+0x3d639cdc> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -14918,13 +14978,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 482e <.literal16+0xdbe> + DB 114,28 ; jb 48ae <.literal16+0xe1e> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4832 <.literal16+0xdc2> + DB 62,114,28 ; jb,pt 48b2 <.literal16+0xe22> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4836 <.literal16+0xdc6> + DB 62,114,28 ; jb,pt 48b6 <.literal16+0xe26> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 483a <.literal16+0xdca> + DB 62,114,28 ; jb,pt 48ba <.literal16+0xe2a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -14945,11 +15005,11 @@ 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 4872 <.literal16+0xe02> + DB 62,114,28 ; jb,pt 48f2 <.literal16+0xe62> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4876 <.literal16+0xe06> + DB 62,114,28 ; jb,pt 48f6 <.literal16+0xe66> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 487a <.literal16+0xe0a> + DB 62,114,28 ; jb,pt 48fa <.literal16+0xe6a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -14993,7 +15053,7 @@ 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 # 3d63d705 <_sk_callback_sse41+0x3d639d44> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d785 <_sk_callback_sse41+0x3d639dac> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -15019,7 +15079,7 @@ 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 # 3d63d745 <_sk_callback_sse41+0x3d639d84> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d7c5 <_sk_callback_sse41+0x3d639dec> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -15028,13 +15088,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 493e <.literal16+0xece> + DB 114,28 ; jb 49be <.literal16+0xf2e> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4942 <_sk_callback_sse41+0xf81> + DB 62,114,28 ; jb,pt 49c2 <_sk_callback_sse41+0xfe9> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4946 <_sk_callback_sse41+0xf85> + DB 62,114,28 ; jb,pt 49c6 <_sk_callback_sse41+0xfed> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 494a <_sk_callback_sse41+0xf89> + DB 62,114,28 ; jb,pt 49ca <_sk_callback_sse41+0xff1> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -15125,7 +15185,7 @@ _sk_seed_shader_sse2 LABEL PROC DB 102,15,110,199 ; movd %edi,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 - DB 15,40,21,81,61,0,0 ; movaps 0x3d51(%rip),%xmm2 # 3e60 <_sk_callback_sse2+0xb3> + DB 15,40,21,129,61,0,0 ; movaps 0x3d81(%rip),%xmm2 # 3e90 <_sk_callback_sse2+0xb9> DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,16,2 ; movups (%rdx),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 @@ -15134,7 +15194,7 @@ _sk_seed_shader_sse2 LABEL PROC DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,64,61,0,0 ; movaps 0x3d40(%rip),%xmm2 # 3e70 <_sk_callback_sse2+0xc3> + DB 15,40,21,112,61,0,0 ; movaps 0x3d70(%rip),%xmm2 # 3ea0 <_sk_callback_sse2+0xc9> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -15168,7 +15228,7 @@ _sk_clear_sse2 LABEL PROC PUBLIC _sk_srcatop_sse2 _sk_srcatop_sse2 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,251,60,0,0 ; movaps 0x3cfb(%rip),%xmm8 # 3e80 <_sk_callback_sse2+0xd3> + DB 68,15,40,5,43,61,0,0 ; movaps 0x3d2b(%rip),%xmm8 # 3eb0 <_sk_callback_sse2+0xd9> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -15191,7 +15251,7 @@ 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 68,15,40,13,190,60,0,0 ; movaps 0x3cbe(%rip),%xmm9 # 3e90 <_sk_callback_sse2+0xe3> + DB 68,15,40,13,238,60,0,0 ; movaps 0x3cee(%rip),%xmm9 # 3ec0 <_sk_callback_sse2+0xe9> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -15232,7 +15292,7 @@ _sk_dstin_sse2 LABEL PROC PUBLIC _sk_srcout_sse2 _sk_srcout_sse2 LABEL PROC - DB 68,15,40,5,98,60,0,0 ; movaps 0x3c62(%rip),%xmm8 # 3ea0 <_sk_callback_sse2+0xf3> + DB 68,15,40,5,146,60,0,0 ; movaps 0x3c92(%rip),%xmm8 # 3ed0 <_sk_callback_sse2+0xf9> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -15243,7 +15303,7 @@ _sk_srcout_sse2 LABEL PROC PUBLIC _sk_dstout_sse2 _sk_dstout_sse2 LABEL PROC - DB 68,15,40,5,82,60,0,0 ; movaps 0x3c52(%rip),%xmm8 # 3eb0 <_sk_callback_sse2+0x103> + DB 68,15,40,5,130,60,0,0 ; movaps 0x3c82(%rip),%xmm8 # 3ee0 <_sk_callback_sse2+0x109> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -15258,7 +15318,7 @@ _sk_dstout_sse2 LABEL PROC PUBLIC _sk_srcover_sse2 _sk_srcover_sse2 LABEL PROC - DB 68,15,40,5,53,60,0,0 ; movaps 0x3c35(%rip),%xmm8 # 3ec0 <_sk_callback_sse2+0x113> + DB 68,15,40,5,101,60,0,0 ; movaps 0x3c65(%rip),%xmm8 # 3ef0 <_sk_callback_sse2+0x119> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -15276,7 +15336,7 @@ _sk_srcover_sse2 LABEL PROC PUBLIC _sk_dstover_sse2 _sk_dstover_sse2 LABEL PROC - DB 68,15,40,5,9,60,0,0 ; movaps 0x3c09(%rip),%xmm8 # 3ed0 <_sk_callback_sse2+0x123> + DB 68,15,40,5,57,60,0,0 ; movaps 0x3c39(%rip),%xmm8 # 3f00 <_sk_callback_sse2+0x129> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -15300,7 +15360,7 @@ _sk_modulate_sse2 LABEL PROC PUBLIC _sk_multiply_sse2 _sk_multiply_sse2 LABEL PROC - DB 68,15,40,5,221,59,0,0 ; movaps 0x3bdd(%rip),%xmm8 # 3ee0 <_sk_callback_sse2+0x133> + DB 68,15,40,5,13,60,0,0 ; movaps 0x3c0d(%rip),%xmm8 # 3f10 <_sk_callback_sse2+0x139> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -15370,7 +15430,7 @@ _sk_screen_sse2 LABEL PROC PUBLIC _sk_xor__sse2 _sk_xor__sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,14,59,0,0 ; movaps 0x3b0e(%rip),%xmm3 # 3ef0 <_sk_callback_sse2+0x143> + DB 15,40,29,62,59,0,0 ; movaps 0x3b3e(%rip),%xmm3 # 3f20 <_sk_callback_sse2+0x149> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -15416,7 +15476,7 @@ _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 15,40,21,121,58,0,0 ; movaps 0x3a79(%rip),%xmm2 # 3f00 <_sk_callback_sse2+0x153> + DB 15,40,21,169,58,0,0 ; movaps 0x3aa9(%rip),%xmm2 # 3f30 <_sk_callback_sse2+0x159> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -15448,7 +15508,7 @@ _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 15,40,21,30,58,0,0 ; movaps 0x3a1e(%rip),%xmm2 # 3f10 <_sk_callback_sse2+0x163> + DB 15,40,21,78,58,0,0 ; movaps 0x3a4e(%rip),%xmm2 # 3f40 <_sk_callback_sse2+0x169> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -15483,7 +15543,7 @@ _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 15,40,21,184,57,0,0 ; movaps 0x39b8(%rip),%xmm2 # 3f20 <_sk_callback_sse2+0x173> + DB 15,40,21,232,57,0,0 ; movaps 0x39e8(%rip),%xmm2 # 3f50 <_sk_callback_sse2+0x179> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -15508,7 +15568,7 @@ _sk_exclusion_sse2 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,202 ; subps %xmm2,%xmm9 - DB 15,40,13,121,57,0,0 ; movaps 0x3979(%rip),%xmm1 # 3f30 <_sk_callback_sse2+0x183> + DB 15,40,13,169,57,0,0 ; movaps 0x39a9(%rip),%xmm1 # 3f60 <_sk_callback_sse2+0x189> DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,207 ; mulps %xmm7,%xmm1 DB 15,88,217 ; addps %xmm1,%xmm3 @@ -15520,7 +15580,7 @@ _sk_exclusion_sse2 LABEL PROC PUBLIC _sk_colorburn_sse2 _sk_colorburn_sse2 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,104,57,0,0 ; movaps 0x3968(%rip),%xmm10 # 3f40 <_sk_callback_sse2+0x193> + DB 68,15,40,21,152,57,0,0 ; movaps 0x3998(%rip),%xmm10 # 3f70 <_sk_callback_sse2+0x199> DB 69,15,40,202 ; movaps %xmm10,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,217 ; movaps %xmm9,%xmm11 @@ -15612,7 +15672,7 @@ _sk_colorburn_sse2 LABEL PROC PUBLIC _sk_colordodge_sse2 _sk_colordodge_sse2 LABEL PROC DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 68,15,40,21,30,56,0,0 ; movaps 0x381e(%rip),%xmm10 # 3f50 <_sk_callback_sse2+0x1a3> + DB 68,15,40,21,78,56,0,0 ; movaps 0x384e(%rip),%xmm10 # 3f80 <_sk_callback_sse2+0x1a9> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -15705,7 +15765,7 @@ _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 68,15,40,29,208,54,0,0 ; movaps 0x36d0(%rip),%xmm11 # 3f60 <_sk_callback_sse2+0x1b3> + DB 68,15,40,29,0,55,0,0 ; movaps 0x3700(%rip),%xmm11 # 3f90 <_sk_callback_sse2+0x1b9> DB 69,15,40,211 ; movaps %xmm11,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,194 ; movaps %xmm10,%xmm8 @@ -15792,7 +15852,7 @@ 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 68,15,40,13,155,53,0,0 ; movaps 0x359b(%rip),%xmm9 # 3f70 <_sk_callback_sse2+0x1c3> + DB 68,15,40,13,203,53,0,0 ; movaps 0x35cb(%rip),%xmm9 # 3fa0 <_sk_callback_sse2+0x1c9> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -15882,7 +15942,7 @@ _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 68,15,40,13,85,52,0,0 ; movaps 0x3455(%rip),%xmm9 # 3f80 <_sk_callback_sse2+0x1d3> + DB 68,15,40,13,133,52,0,0 ; movaps 0x3485(%rip),%xmm9 # 3fb0 <_sk_callback_sse2+0x1d9> DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 69,15,92,250 ; subps %xmm10,%xmm15 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -15895,10 +15955,10 @@ _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 68,15,40,53,47,52,0,0 ; movaps 0x342f(%rip),%xmm14 # 3f90 <_sk_callback_sse2+0x1e3> + DB 68,15,40,53,95,52,0,0 ; movaps 0x345f(%rip),%xmm14 # 3fc0 <_sk_callback_sse2+0x1e9> DB 69,15,88,222 ; addps %xmm14,%xmm11 DB 68,15,89,216 ; mulps %xmm0,%xmm11 - DB 68,15,40,21,47,52,0,0 ; movaps 0x342f(%rip),%xmm10 # 3fa0 <_sk_callback_sse2+0x1f3> + DB 68,15,40,21,95,52,0,0 ; movaps 0x345f(%rip),%xmm10 # 3fd0 <_sk_callback_sse2+0x1f9> DB 69,15,89,234 ; mulps %xmm10,%xmm13 DB 69,15,88,235 ; addps %xmm11,%xmm13 DB 15,88,228 ; addps %xmm4,%xmm4 @@ -16047,7 +16107,7 @@ _sk_clamp_0_sse2 LABEL PROC PUBLIC _sk_clamp_1_sse2 _sk_clamp_1_sse2 LABEL PROC - DB 68,15,40,5,59,50,0,0 ; movaps 0x323b(%rip),%xmm8 # 3fb0 <_sk_callback_sse2+0x203> + DB 68,15,40,5,107,50,0,0 ; movaps 0x326b(%rip),%xmm8 # 3fe0 <_sk_callback_sse2+0x209> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -16057,7 +16117,7 @@ _sk_clamp_1_sse2 LABEL PROC PUBLIC _sk_clamp_a_sse2 _sk_clamp_a_sse2 LABEL PROC - DB 15,93,29,48,50,0,0 ; minps 0x3230(%rip),%xmm3 # 3fc0 <_sk_callback_sse2+0x213> + DB 15,93,29,96,50,0,0 ; minps 0x3260(%rip),%xmm3 # 3ff0 <_sk_callback_sse2+0x219> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -16130,7 +16190,7 @@ _sk_premul_sse2 LABEL PROC PUBLIC _sk_unpremul_sse2 _sk_unpremul_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,155,49,0,0 ; movaps 0x319b(%rip),%xmm9 # 3fd0 <_sk_callback_sse2+0x223> + DB 68,15,40,13,203,49,0,0 ; movaps 0x31cb(%rip),%xmm9 # 4000 <_sk_callback_sse2+0x229> 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 @@ -16142,20 +16202,20 @@ _sk_unpremul_sse2 LABEL PROC PUBLIC _sk_from_srgb_sse2 _sk_from_srgb_sse2 LABEL PROC - DB 68,15,40,5,134,49,0,0 ; movaps 0x3186(%rip),%xmm8 # 3fe0 <_sk_callback_sse2+0x233> + DB 68,15,40,5,182,49,0,0 ; movaps 0x31b6(%rip),%xmm8 # 4010 <_sk_callback_sse2+0x239> 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 68,15,40,13,126,49,0,0 ; movaps 0x317e(%rip),%xmm9 # 3ff0 <_sk_callback_sse2+0x243> + DB 68,15,40,13,174,49,0,0 ; movaps 0x31ae(%rip),%xmm9 # 4020 <_sk_callback_sse2+0x249> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,241 ; mulps %xmm9,%xmm14 - DB 68,15,40,21,126,49,0,0 ; movaps 0x317e(%rip),%xmm10 # 4000 <_sk_callback_sse2+0x253> + DB 68,15,40,21,174,49,0,0 ; movaps 0x31ae(%rip),%xmm10 # 4030 <_sk_callback_sse2+0x259> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 69,15,89,243 ; mulps %xmm11,%xmm14 - DB 68,15,40,29,126,49,0,0 ; movaps 0x317e(%rip),%xmm11 # 4010 <_sk_callback_sse2+0x263> + DB 68,15,40,29,174,49,0,0 ; movaps 0x31ae(%rip),%xmm11 # 4040 <_sk_callback_sse2+0x269> DB 69,15,88,243 ; addps %xmm11,%xmm14 - DB 68,15,40,37,130,49,0,0 ; movaps 0x3182(%rip),%xmm12 # 4020 <_sk_callback_sse2+0x273> + DB 68,15,40,37,178,49,0,0 ; movaps 0x31b2(%rip),%xmm12 # 4050 <_sk_callback_sse2+0x279> 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 @@ -16192,20 +16252,20 @@ _sk_to_srgb_sse2 LABEL PROC DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 DB 69,15,83,200 ; rcpps %xmm8,%xmm9 DB 69,15,82,232 ; rsqrtps %xmm8,%xmm13 - DB 68,15,40,5,7,49,0,0 ; movaps 0x3107(%rip),%xmm8 # 4030 <_sk_callback_sse2+0x283> + DB 68,15,40,5,55,49,0,0 ; movaps 0x3137(%rip),%xmm8 # 4060 <_sk_callback_sse2+0x289> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,240 ; mulps %xmm8,%xmm14 - DB 68,15,40,21,7,49,0,0 ; movaps 0x3107(%rip),%xmm10 # 4040 <_sk_callback_sse2+0x293> + DB 68,15,40,21,55,49,0,0 ; movaps 0x3137(%rip),%xmm10 # 4070 <_sk_callback_sse2+0x299> DB 69,15,89,202 ; mulps %xmm10,%xmm9 - DB 68,15,40,29,11,49,0,0 ; movaps 0x310b(%rip),%xmm11 # 4050 <_sk_callback_sse2+0x2a3> + DB 68,15,40,29,59,49,0,0 ; movaps 0x313b(%rip),%xmm11 # 4080 <_sk_callback_sse2+0x2a9> DB 69,15,88,203 ; addps %xmm11,%xmm9 - DB 68,15,40,37,15,49,0,0 ; movaps 0x310f(%rip),%xmm12 # 4060 <_sk_callback_sse2+0x2b3> + DB 68,15,40,37,63,49,0,0 ; movaps 0x313f(%rip),%xmm12 # 4090 <_sk_callback_sse2+0x2b9> DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,40,13,15,49,0,0 ; movaps 0x310f(%rip),%xmm9 # 4070 <_sk_callback_sse2+0x2c3> + DB 68,15,40,13,63,49,0,0 ; movaps 0x313f(%rip),%xmm9 # 40a0 <_sk_callback_sse2+0x2c9> DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 69,15,93,253 ; minps %xmm13,%xmm15 - DB 68,15,40,45,15,49,0,0 ; movaps 0x310f(%rip),%xmm13 # 4080 <_sk_callback_sse2+0x2d3> + DB 68,15,40,45,63,49,0,0 ; movaps 0x313f(%rip),%xmm13 # 40b0 <_sk_callback_sse2+0x2d9> DB 65,15,194,197,1 ; cmpltps %xmm13,%xmm0 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,199 ; andnps %xmm15,%xmm0 @@ -16253,7 +16313,7 @@ _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 68,15,40,45,104,48,0,0 ; movaps 0x3068(%rip),%xmm13 # 4090 <_sk_callback_sse2+0x2e3> + DB 68,15,40,45,152,48,0,0 ; movaps 0x3098(%rip),%xmm13 # 40c0 <_sk_callback_sse2+0x2e9> 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 @@ -16262,30 +16322,30 @@ _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 68,15,84,53,78,48,0,0 ; andps 0x304e(%rip),%xmm14 # 40a0 <_sk_callback_sse2+0x2f3> + DB 68,15,84,53,126,48,0,0 ; andps 0x307e(%rip),%xmm14 # 40d0 <_sk_callback_sse2+0x2f9> 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 68,15,40,37,65,48,0,0 ; movaps 0x3041(%rip),%xmm12 # 40b0 <_sk_callback_sse2+0x303> + DB 68,15,40,37,113,48,0,0 ; movaps 0x3071(%rip),%xmm12 # 40e0 <_sk_callback_sse2+0x309> 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 68,15,88,5,61,48,0,0 ; addps 0x303d(%rip),%xmm8 # 40c0 <_sk_callback_sse2+0x313> + DB 68,15,88,5,109,48,0,0 ; addps 0x306d(%rip),%xmm8 # 40f0 <_sk_callback_sse2+0x319> 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 15,89,5,46,48,0,0 ; mulps 0x302e(%rip),%xmm0 # 40d0 <_sk_callback_sse2+0x323> + DB 15,89,5,94,48,0,0 ; mulps 0x305e(%rip),%xmm0 # 4100 <_sk_callback_sse2+0x329> 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 68,15,40,13,33,48,0,0 ; movaps 0x3021(%rip),%xmm9 # 40e0 <_sk_callback_sse2+0x333> + DB 68,15,40,13,81,48,0,0 ; movaps 0x3051(%rip),%xmm9 # 4110 <_sk_callback_sse2+0x339> 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 @@ -16300,182 +16360,185 @@ _sk_rgb_to_hsl_sse2 LABEL PROC PUBLIC _sk_hsl_to_rgb_sse2 _sk_hsl_to_rgb_sse2 LABEL PROC - DB 72,131,236,104 ; sub $0x68,%rsp - DB 15,41,124,36,80 ; movaps %xmm7,0x50(%rsp) - DB 15,41,116,36,64 ; movaps %xmm6,0x40(%rsp) - DB 15,41,108,36,48 ; movaps %xmm5,0x30(%rsp) - DB 15,41,100,36,32 ; movaps %xmm4,0x20(%rsp) - DB 15,41,92,36,16 ; movaps %xmm3,0x10(%rsp) + DB 72,129,236,168,0,0,0 ; sub $0xa8,%rsp + DB 15,41,188,36,144,0,0,0 ; movaps %xmm7,0x90(%rsp) + DB 15,41,180,36,128,0,0,0 ; movaps %xmm6,0x80(%rsp) + DB 15,41,108,36,112 ; movaps %xmm5,0x70(%rsp) + DB 15,41,100,36,96 ; movaps %xmm4,0x60(%rsp) + DB 15,41,92,36,80 ; movaps %xmm3,0x50(%rsp) DB 68,15,40,210 ; movaps %xmm2,%xmm10 - DB 15,40,209 ; movaps %xmm1,%xmm2 - DB 15,40,240 ; movaps %xmm0,%xmm6 + DB 15,40,217 ; movaps %xmm1,%xmm3 + DB 15,40,232 ; movaps %xmm0,%xmm5 DB 184,0,0,0,63 ; mov $0x3f000000,%eax - DB 102,68,15,110,240 ; movd %eax,%xmm14 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,40,202 ; movaps %xmm10,%xmm9 - DB 69,15,194,206,1 ; cmpltps %xmm14,%xmm9 - DB 184,0,0,128,63 ; mov $0x3f800000,%eax - DB 102,68,15,110,224 ; movd %eax,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 15,88,194 ; addps %xmm2,%xmm0 - DB 65,15,89,194 ; mulps %xmm10,%xmm0 - DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 69,15,87,219 ; xorps %xmm11,%xmm11 - DB 68,15,194,218,0 ; cmpeqps %xmm2,%xmm11 - DB 65,15,88,210 ; addps %xmm10,%xmm2 + DB 102,15,110,192 ; movd %eax,%xmm0 + DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0 + DB 69,15,40,194 ; movaps %xmm10,%xmm8 + DB 68,15,194,192,1 ; cmpltps %xmm0,%xmm8 + DB 68,15,40,216 ; movaps %xmm0,%xmm11 + DB 68,15,40,37,230,47,0,0 ; movaps 0x2fe6(%rip),%xmm12 # 4120 <_sk_callback_sse2+0x349> + DB 15,40,195 ; movaps %xmm3,%xmm0 + DB 15,40,211 ; movaps %xmm3,%xmm2 + DB 15,87,201 ; xorps %xmm1,%xmm1 + DB 15,194,203,0 ; cmpeqps %xmm3,%xmm1 + DB 15,41,76,36,48 ; movaps %xmm1,0x30(%rsp) + DB 65,15,88,220 ; addps %xmm12,%xmm3 DB 65,15,89,218 ; mulps %xmm10,%xmm3 - DB 15,92,211 ; subps %xmm3,%xmm2 - DB 65,15,84,193 ; andps %xmm9,%xmm0 - DB 68,15,85,202 ; andnps %xmm2,%xmm9 - DB 68,15,86,200 ; orps %xmm0,%xmm9 - DB 184,0,0,0,64 ; mov $0x40000000,%eax - DB 185,171,170,170,62 ; mov $0x3eaaaaab,%ecx - DB 102,68,15,110,193 ; movd %ecx,%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,41,4,36 ; movaps %xmm8,(%rsp) - DB 68,15,88,198 ; addps %xmm6,%xmm8 - DB 185,0,0,0,0 ; mov $0x0,%ecx - DB 102,15,110,233 ; movd %ecx,%xmm5 - DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 - DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 65,15,194,192,1 ; cmpltps %xmm8,%xmm0 - DB 65,15,40,216 ; movaps %xmm8,%xmm3 - DB 65,15,92,220 ; subps %xmm12,%xmm3 - DB 15,84,216 ; andps %xmm0,%xmm3 - DB 65,15,85,192 ; andnps %xmm8,%xmm0 - DB 15,86,195 ; orps %xmm3,%xmm0 - DB 65,15,40,216 ; movaps %xmm8,%xmm3 - DB 15,194,221,1 ; cmpltps %xmm5,%xmm3 - DB 65,15,40,212 ; movaps %xmm12,%xmm2 - DB 65,15,88,208 ; addps %xmm8,%xmm2 - DB 15,84,211 ; andps %xmm3,%xmm2 - DB 15,85,216 ; andnps %xmm0,%xmm3 - DB 15,86,218 ; orps %xmm2,%xmm3 - DB 102,68,15,110,232 ; movd %eax,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 69,15,92,233 ; subps %xmm9,%xmm13 + DB 65,15,88,194 ; addps %xmm10,%xmm0 + DB 65,15,89,210 ; mulps %xmm10,%xmm2 + DB 15,92,194 ; subps %xmm2,%xmm0 + DB 65,15,84,216 ; andps %xmm8,%xmm3 + DB 68,15,85,192 ; andnps %xmm0,%xmm8 + DB 68,15,86,195 ; orps %xmm3,%xmm8 + DB 15,40,29,190,47,0,0 ; movaps 0x2fbe(%rip),%xmm3 # 4130 <_sk_callback_sse2+0x359> + DB 15,88,221 ; addps %xmm5,%xmm3 + DB 184,0,0,0,0 ; mov $0x0,%eax + DB 185,0,0,128,63 ; mov $0x3f800000,%ecx + DB 102,68,15,110,241 ; movd %ecx,%xmm14 + DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 + DB 65,15,40,214 ; movaps %xmm14,%xmm2 + DB 15,194,211,1 ; cmpltps %xmm3,%xmm2 + DB 68,15,40,61,167,47,0,0 ; movaps 0x2fa7(%rip),%xmm15 # 4140 <_sk_callback_sse2+0x369> + DB 15,40,195 ; movaps %xmm3,%xmm0 + DB 65,15,88,199 ; addps %xmm15,%xmm0 + DB 15,84,194 ; andps %xmm2,%xmm0 + DB 15,85,211 ; andnps %xmm3,%xmm2 + DB 15,86,208 ; orps %xmm0,%xmm2 + DB 102,15,110,200 ; movd %eax,%xmm1 + DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 + DB 15,41,12,36 ; movaps %xmm1,(%rsp) + DB 15,40,195 ; movaps %xmm3,%xmm0 + DB 15,194,193,1 ; cmpltps %xmm1,%xmm0 + DB 15,40,227 ; movaps %xmm3,%xmm4 + DB 65,15,88,228 ; addps %xmm12,%xmm4 + DB 15,84,224 ; andps %xmm0,%xmm4 + DB 15,85,194 ; andnps %xmm2,%xmm0 + DB 15,86,196 ; orps %xmm4,%xmm0 + DB 69,15,40,234 ; movaps %xmm10,%xmm13 + DB 69,15,88,237 ; addps %xmm13,%xmm13 + DB 69,15,92,232 ; subps %xmm8,%xmm13 DB 184,171,170,42,62 ; mov $0x3e2aaaab,%eax - DB 65,15,40,193 ; movaps %xmm9,%xmm0 - DB 65,15,92,197 ; subps %xmm13,%xmm0 - DB 185,0,0,192,64 ; mov $0x40c00000,%ecx - DB 102,68,15,110,249 ; movd %ecx,%xmm15 - DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 68,15,89,248 ; mulps %xmm0,%xmm15 + DB 69,15,40,200 ; movaps %xmm8,%xmm9 + DB 69,15,92,205 ; subps %xmm13,%xmm9 + DB 68,15,89,13,99,47,0,0 ; mulps 0x2f63(%rip),%xmm9 # 4150 <_sk_callback_sse2+0x379> DB 185,171,170,42,63 ; mov $0x3f2aaaab,%ecx - DB 102,15,110,225 ; movd %ecx,%xmm4 - DB 15,198,228,0 ; shufps $0x0,%xmm4,%xmm4 - DB 15,40,212 ; movaps %xmm4,%xmm2 - DB 15,92,211 ; subps %xmm3,%xmm2 - DB 15,40,203 ; movaps %xmm3,%xmm1 - DB 15,40,195 ; movaps %xmm3,%xmm0 - DB 15,194,220,1 ; cmpltps %xmm4,%xmm3 - DB 65,15,89,215 ; mulps %xmm15,%xmm2 - DB 65,15,88,213 ; addps %xmm13,%xmm2 - DB 15,84,211 ; andps %xmm3,%xmm2 - DB 65,15,85,221 ; andnps %xmm13,%xmm3 - DB 15,86,218 ; orps %xmm2,%xmm3 - DB 65,15,194,198,1 ; cmpltps %xmm14,%xmm0 - DB 65,15,40,209 ; movaps %xmm9,%xmm2 - DB 15,84,208 ; andps %xmm0,%xmm2 - DB 15,85,195 ; andnps %xmm3,%xmm0 - DB 15,86,194 ; orps %xmm2,%xmm0 - DB 102,15,110,248 ; movd %eax,%xmm7 + DB 102,15,110,249 ; movd %ecx,%xmm7 DB 15,198,255,0 ; shufps $0x0,%xmm7,%xmm7 + DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp) + DB 15,40,53,90,47,0,0 ; movaps 0x2f5a(%rip),%xmm6 # 4160 <_sk_callback_sse2+0x389> + DB 15,40,230 ; movaps %xmm6,%xmm4 + DB 15,92,224 ; subps %xmm0,%xmm4 + DB 15,40,208 ; movaps %xmm0,%xmm2 + DB 15,40,200 ; movaps %xmm0,%xmm1 + DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 + DB 65,15,89,225 ; mulps %xmm9,%xmm4 + DB 65,15,88,229 ; addps %xmm13,%xmm4 + DB 15,84,224 ; andps %xmm0,%xmm4 + DB 65,15,85,197 ; andnps %xmm13,%xmm0 + DB 15,86,196 ; orps %xmm4,%xmm0 + DB 65,15,40,251 ; movaps %xmm11,%xmm7 + DB 15,41,124,36,32 ; movaps %xmm7,0x20(%rsp) DB 15,194,207,1 ; cmpltps %xmm7,%xmm1 - DB 69,15,89,199 ; mulps %xmm15,%xmm8 - DB 69,15,88,197 ; addps %xmm13,%xmm8 - DB 68,15,84,193 ; andps %xmm1,%xmm8 + DB 65,15,40,224 ; movaps %xmm8,%xmm4 + DB 15,84,225 ; andps %xmm1,%xmm4 DB 15,85,200 ; andnps %xmm0,%xmm1 - DB 65,15,86,200 ; orps %xmm8,%xmm1 - DB 69,15,40,195 ; movaps %xmm11,%xmm8 - DB 68,15,85,193 ; andnps %xmm1,%xmm8 - DB 65,15,40,196 ; movaps %xmm12,%xmm0 - DB 15,194,198,1 ; cmpltps %xmm6,%xmm0 - DB 15,40,206 ; movaps %xmm6,%xmm1 - DB 65,15,92,204 ; subps %xmm12,%xmm1 + DB 15,86,204 ; orps %xmm4,%xmm1 + DB 102,15,110,224 ; movd %eax,%xmm4 + DB 15,198,228,0 ; shufps $0x0,%xmm4,%xmm4 + DB 15,194,212,1 ; cmpltps %xmm4,%xmm2 + DB 65,15,89,217 ; mulps %xmm9,%xmm3 + DB 65,15,88,221 ; addps %xmm13,%xmm3 + DB 15,84,218 ; andps %xmm2,%xmm3 + DB 15,85,209 ; andnps %xmm1,%xmm2 + DB 15,86,211 ; orps %xmm3,%xmm2 + DB 68,15,40,92,36,48 ; movaps 0x30(%rsp),%xmm11 + DB 65,15,40,195 ; movaps %xmm11,%xmm0 + DB 15,85,194 ; andnps %xmm2,%xmm0 + DB 15,41,68,36,64 ; movaps %xmm0,0x40(%rsp) + DB 65,15,40,198 ; movaps %xmm14,%xmm0 + DB 15,194,197,1 ; cmpltps %xmm5,%xmm0 + DB 15,40,205 ; movaps %xmm5,%xmm1 + DB 65,15,88,207 ; addps %xmm15,%xmm1 DB 15,84,200 ; andps %xmm0,%xmm1 - DB 15,85,198 ; andnps %xmm6,%xmm0 + DB 15,85,197 ; andnps %xmm5,%xmm0 DB 15,86,193 ; orps %xmm1,%xmm0 - DB 15,40,206 ; movaps %xmm6,%xmm1 - DB 15,194,205,1 ; cmpltps %xmm5,%xmm1 - DB 65,15,40,212 ; movaps %xmm12,%xmm2 - DB 15,88,214 ; addps %xmm6,%xmm2 + DB 15,40,205 ; movaps %xmm5,%xmm1 + DB 15,194,12,36,1 ; cmpltps (%rsp),%xmm1 + DB 15,40,213 ; movaps %xmm5,%xmm2 + DB 65,15,88,212 ; addps %xmm12,%xmm2 DB 15,84,209 ; andps %xmm1,%xmm2 DB 15,85,200 ; andnps %xmm0,%xmm1 DB 15,86,202 ; orps %xmm2,%xmm1 - DB 15,40,196 ; movaps %xmm4,%xmm0 + DB 15,40,198 ; movaps %xmm6,%xmm0 DB 15,92,193 ; subps %xmm1,%xmm0 - DB 15,40,217 ; movaps %xmm1,%xmm3 DB 15,40,209 ; movaps %xmm1,%xmm2 - DB 15,194,204,1 ; cmpltps %xmm4,%xmm1 - DB 65,15,89,199 ; mulps %xmm15,%xmm0 + DB 15,40,217 ; movaps %xmm1,%xmm3 + DB 15,194,76,36,16,1 ; cmpltps 0x10(%rsp),%xmm1 + DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,197 ; addps %xmm13,%xmm0 DB 15,84,193 ; andps %xmm1,%xmm0 DB 65,15,85,205 ; andnps %xmm13,%xmm1 DB 15,86,200 ; orps %xmm0,%xmm1 - DB 65,15,194,214,1 ; cmpltps %xmm14,%xmm2 - DB 65,15,40,193 ; movaps %xmm9,%xmm0 - DB 15,84,194 ; andps %xmm2,%xmm0 - DB 15,85,209 ; andnps %xmm1,%xmm2 - DB 15,86,208 ; orps %xmm0,%xmm2 DB 15,194,223,1 ; cmpltps %xmm7,%xmm3 - DB 65,15,40,199 ; movaps %xmm15,%xmm0 - DB 15,89,198 ; mulps %xmm6,%xmm0 - DB 65,15,88,197 ; addps %xmm13,%xmm0 + DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,84,195 ; andps %xmm3,%xmm0 - DB 15,85,218 ; andnps %xmm2,%xmm3 + DB 15,85,217 ; andnps %xmm1,%xmm3 DB 15,86,216 ; orps %xmm0,%xmm3 - DB 65,15,40,203 ; movaps %xmm11,%xmm1 - DB 15,85,203 ; andnps %xmm3,%xmm1 - DB 15,92,52,36 ; subps (%rsp),%xmm6 - DB 15,40,198 ; movaps %xmm6,%xmm0 - DB 15,194,197,1 ; cmpltps %xmm5,%xmm0 - DB 15,40,214 ; movaps %xmm6,%xmm2 - DB 65,15,92,212 ; subps %xmm12,%xmm2 - DB 65,15,40,220 ; movaps %xmm12,%xmm3 - DB 68,15,194,230,1 ; cmpltps %xmm6,%xmm12 - DB 65,15,84,212 ; andps %xmm12,%xmm2 - DB 68,15,85,230 ; andnps %xmm6,%xmm12 - DB 68,15,86,226 ; orps %xmm2,%xmm12 - DB 15,88,222 ; addps %xmm6,%xmm3 - DB 15,84,216 ; andps %xmm0,%xmm3 - DB 65,15,85,196 ; andnps %xmm12,%xmm0 - DB 15,86,195 ; orps %xmm3,%xmm0 - DB 15,40,232 ; movaps %xmm0,%xmm5 - DB 15,194,239,1 ; cmpltps %xmm7,%xmm5 - DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,194,212,1 ; cmpltps %xmm4,%xmm2 - DB 15,92,224 ; subps %xmm0,%xmm4 - DB 65,15,194,198,1 ; cmpltps %xmm14,%xmm0 - DB 65,15,89,247 ; mulps %xmm15,%xmm6 - DB 65,15,89,231 ; mulps %xmm15,%xmm4 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 15,89,197 ; mulps %xmm5,%xmm0 + DB 65,15,88,197 ; addps %xmm13,%xmm0 + DB 15,84,194 ; andps %xmm2,%xmm0 + DB 15,85,211 ; andnps %xmm3,%xmm2 + DB 15,86,208 ; orps %xmm0,%xmm2 + DB 65,15,40,219 ; movaps %xmm11,%xmm3 + DB 15,85,218 ; andnps %xmm2,%xmm3 + DB 15,88,45,116,46,0,0 ; addps 0x2e74(%rip),%xmm5 # 4170 <_sk_callback_sse2+0x399> + DB 15,40,197 ; movaps %xmm5,%xmm0 + DB 15,194,4,36,1 ; cmpltps (%rsp),%xmm0 + DB 68,15,194,245,1 ; cmpltps %xmm5,%xmm14 + DB 68,15,88,253 ; addps %xmm5,%xmm15 + DB 69,15,84,254 ; andps %xmm14,%xmm15 + DB 68,15,85,245 ; andnps %xmm5,%xmm14 + DB 69,15,86,247 ; orps %xmm15,%xmm14 + DB 68,15,88,229 ; addps %xmm5,%xmm12 + DB 68,15,84,224 ; andps %xmm0,%xmm12 + DB 65,15,85,198 ; andnps %xmm14,%xmm0 + DB 65,15,86,196 ; orps %xmm12,%xmm0 + DB 15,40,248 ; movaps %xmm0,%xmm7 + DB 15,194,252,1 ; cmpltps %xmm4,%xmm7 + DB 15,40,200 ; movaps %xmm0,%xmm1 + DB 15,194,76,36,32,1 ; cmpltps 0x20(%rsp),%xmm1 + DB 15,92,240 ; subps %xmm0,%xmm6 + DB 15,194,68,36,16,1 ; cmpltps 0x10(%rsp),%xmm0 + DB 65,15,89,233 ; mulps %xmm9,%xmm5 + DB 65,15,89,241 ; mulps %xmm9,%xmm6 + DB 65,15,88,237 ; addps %xmm13,%xmm5 DB 65,15,88,245 ; addps %xmm13,%xmm6 - DB 65,15,88,229 ; addps %xmm13,%xmm4 - DB 15,84,226 ; andps %xmm2,%xmm4 - DB 65,15,85,213 ; andnps %xmm13,%xmm2 - DB 15,86,212 ; orps %xmm4,%xmm2 - DB 68,15,84,200 ; andps %xmm0,%xmm9 - DB 15,85,194 ; andnps %xmm2,%xmm0 - DB 65,15,86,193 ; orps %xmm9,%xmm0 - DB 15,84,245 ; andps %xmm5,%xmm6 - DB 15,85,232 ; andnps %xmm0,%xmm5 - DB 15,86,238 ; orps %xmm6,%xmm5 + DB 15,84,240 ; andps %xmm0,%xmm6 + DB 65,15,85,197 ; andnps %xmm13,%xmm0 + DB 15,86,198 ; orps %xmm6,%xmm0 + DB 68,15,84,193 ; andps %xmm1,%xmm8 + DB 15,85,200 ; andnps %xmm0,%xmm1 + DB 65,15,86,200 ; orps %xmm8,%xmm1 + DB 15,84,239 ; andps %xmm7,%xmm5 + DB 15,85,249 ; andnps %xmm1,%xmm7 + DB 15,86,253 ; orps %xmm5,%xmm7 DB 69,15,84,211 ; andps %xmm11,%xmm10 - DB 68,15,85,221 ; andnps %xmm5,%xmm11 - DB 69,15,86,194 ; orps %xmm10,%xmm8 + DB 68,15,85,223 ; andnps %xmm7,%xmm11 + DB 15,40,76,36,64 ; movaps 0x40(%rsp),%xmm1 DB 65,15,86,202 ; orps %xmm10,%xmm1 + DB 65,15,86,218 ; orps %xmm10,%xmm3 DB 69,15,86,211 ; orps %xmm11,%xmm10 DB 72,173 ; lods %ds:(%rsi),%rax - DB 65,15,40,192 ; movaps %xmm8,%xmm0 + DB 15,40,193 ; movaps %xmm1,%xmm0 + DB 15,40,203 ; movaps %xmm3,%xmm1 DB 65,15,40,210 ; movaps %xmm10,%xmm2 - DB 15,40,92,36,16 ; movaps 0x10(%rsp),%xmm3 - DB 15,40,100,36,32 ; movaps 0x20(%rsp),%xmm4 - DB 15,40,108,36,48 ; movaps 0x30(%rsp),%xmm5 - DB 15,40,116,36,64 ; movaps 0x40(%rsp),%xmm6 - DB 15,40,124,36,80 ; movaps 0x50(%rsp),%xmm7 - DB 72,131,196,104 ; add $0x68,%rsp + DB 15,40,92,36,80 ; movaps 0x50(%rsp),%xmm3 + DB 15,40,100,36,96 ; movaps 0x60(%rsp),%xmm4 + DB 15,40,108,36,112 ; movaps 0x70(%rsp),%xmm5 + DB 15,40,180,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm6 + DB 15,40,188,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm7 + DB 72,129,196,168,0,0,0 ; add $0xa8,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_scale_1_float_sse2 @@ -16499,7 +16562,7 @@ _sk_scale_u8_sse2 LABEL PROC DB 102,69,15,96,193 ; punpcklbw %xmm9,%xmm8 DB 102,69,15,97,193 ; punpcklwd %xmm9,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,23,45,0,0 ; mulps 0x2d17(%rip),%xmm8 # 40f0 <_sk_callback_sse2+0x343> + DB 68,15,89,5,125,45,0,0 ; mulps 0x2d7d(%rip),%xmm8 # 4180 <_sk_callback_sse2+0x3a9> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -16536,7 +16599,7 @@ _sk_lerp_u8_sse2 LABEL PROC DB 102,69,15,96,193 ; punpcklbw %xmm9,%xmm8 DB 102,69,15,97,193 ; punpcklwd %xmm9,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,181,44,0,0 ; mulps 0x2cb5(%rip),%xmm8 # 4100 <_sk_callback_sse2+0x353> + DB 68,15,89,5,27,45,0,0 ; mulps 0x2d1b(%rip),%xmm8 # 4190 <_sk_callback_sse2+0x3b9> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -16559,17 +16622,17 @@ _sk_lerp_565_sse2 LABEL PROC DB 243,68,15,126,4,120 ; movq (%rax,%rdi,2),%xmm8 DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,68,15,97,195 ; punpcklwd %xmm3,%xmm8 - DB 102,15,111,29,125,44,0,0 ; movdqa 0x2c7d(%rip),%xmm3 # 4110 <_sk_callback_sse2+0x363> + DB 102,15,111,29,227,44,0,0 ; movdqa 0x2ce3(%rip),%xmm3 # 41a0 <_sk_callback_sse2+0x3c9> DB 102,65,15,219,216 ; pand %xmm8,%xmm3 DB 68,15,91,203 ; cvtdq2ps %xmm3,%xmm9 - DB 68,15,89,13,124,44,0,0 ; mulps 0x2c7c(%rip),%xmm9 # 4120 <_sk_callback_sse2+0x373> - DB 102,15,111,29,132,44,0,0 ; movdqa 0x2c84(%rip),%xmm3 # 4130 <_sk_callback_sse2+0x383> + DB 68,15,89,13,226,44,0,0 ; mulps 0x2ce2(%rip),%xmm9 # 41b0 <_sk_callback_sse2+0x3d9> + DB 102,15,111,29,234,44,0,0 ; movdqa 0x2cea(%rip),%xmm3 # 41c0 <_sk_callback_sse2+0x3e9> DB 102,65,15,219,216 ; pand %xmm8,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,133,44,0,0 ; mulps 0x2c85(%rip),%xmm3 # 4140 <_sk_callback_sse2+0x393> - DB 102,68,15,219,5,140,44,0,0 ; pand 0x2c8c(%rip),%xmm8 # 4150 <_sk_callback_sse2+0x3a3> + DB 15,89,29,235,44,0,0 ; mulps 0x2ceb(%rip),%xmm3 # 41d0 <_sk_callback_sse2+0x3f9> + DB 102,68,15,219,5,242,44,0,0 ; pand 0x2cf2(%rip),%xmm8 # 41e0 <_sk_callback_sse2+0x409> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,144,44,0,0 ; mulps 0x2c90(%rip),%xmm8 # 4160 <_sk_callback_sse2+0x3b3> + DB 68,15,89,5,246,44,0,0 ; mulps 0x2cf6(%rip),%xmm8 # 41f0 <_sk_callback_sse2+0x419> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -16580,7 +16643,7 @@ _sk_lerp_565_sse2 LABEL PROC DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 15,88,214 ; addps %xmm6,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,122,44,0,0 ; movaps 0x2c7a(%rip),%xmm3 # 4170 <_sk_callback_sse2+0x3c3> + DB 15,40,29,224,44,0,0 ; movaps 0x2ce0(%rip),%xmm3 # 4200 <_sk_callback_sse2+0x429> DB 255,224 ; jmpq *%rax PUBLIC _sk_load_tables_sse2 @@ -16589,7 +16652,7 @@ _sk_load_tables_sse2 LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 76,139,72,8 ; mov 0x8(%rax),%r9 DB 243,69,15,111,12,184 ; movdqu (%r8,%rdi,4),%xmm9 - DB 102,68,15,111,5,112,44,0,0 ; movdqa 0x2c70(%rip),%xmm8 # 4180 <_sk_callback_sse2+0x3d3> + DB 102,68,15,111,5,214,44,0,0 ; movdqa 0x2cd6(%rip),%xmm8 # 4210 <_sk_callback_sse2+0x439> 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 @@ -16644,7 +16707,7 @@ _sk_load_tables_sse2 LABEL PROC 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 15,89,29,125,43,0,0 ; mulps 0x2b7d(%rip),%xmm3 # 4190 <_sk_callback_sse2+0x3e3> + DB 15,89,29,227,43,0,0 ; mulps 0x2be3(%rip),%xmm3 # 4220 <_sk_callback_sse2+0x449> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16661,7 +16724,7 @@ _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 102,68,15,111,21,80,43,0,0 ; movdqa 0x2b50(%rip),%xmm10 # 41a0 <_sk_callback_sse2+0x3f3> + DB 102,68,15,111,21,182,43,0,0 ; movdqa 0x2bb6(%rip),%xmm10 # 4230 <_sk_callback_sse2+0x459> 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 @@ -16722,7 +16785,7 @@ _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 15,89,29,63,42,0,0 ; mulps 0x2a3f(%rip),%xmm3 # 41b0 <_sk_callback_sse2+0x403> + DB 15,89,29,165,42,0,0 ; mulps 0x2aa5(%rip),%xmm3 # 4240 <_sk_callback_sse2+0x469> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16742,7 +16805,7 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC 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 102,68,15,111,5,255,41,0,0 ; movdqa 0x29ff(%rip),%xmm8 # 41c0 <_sk_callback_sse2+0x413> + DB 102,68,15,111,5,101,42,0,0 ; movdqa 0x2a65(%rip),%xmm8 # 4250 <_sk_callback_sse2+0x479> 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 @@ -16797,7 +16860,7 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,14,41,0,0 ; movaps 0x290e(%rip),%xmm3 # 41d0 <_sk_callback_sse2+0x423> + DB 15,40,29,116,41,0,0 ; movaps 0x2974(%rip),%xmm3 # 4260 <_sk_callback_sse2+0x489> DB 255,224 ; jmpq *%rax PUBLIC _sk_byte_tables_sse2 @@ -16805,7 +16868,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,15,41,0,0 ; movaps 0x290f(%rip),%xmm8 # 41e0 <_sk_callback_sse2+0x433> + DB 68,15,40,5,117,41,0,0 ; movaps 0x2975(%rip),%xmm8 # 4270 <_sk_callback_sse2+0x499> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,72,15,126,193 ; movq %xmm0,%rcx @@ -16832,7 +16895,7 @@ _sk_byte_tables_sse2 LABEL PROC 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 68,15,40,21,172,40,0,0 ; movaps 0x28ac(%rip),%xmm10 # 41f0 <_sk_callback_sse2+0x443> + DB 68,15,40,21,18,41,0,0 ; movaps 0x2912(%rip),%xmm10 # 4280 <_sk_callback_sse2+0x4a9> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -16946,7 +17009,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC 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 68,15,40,21,255,38,0,0 ; movaps 0x26ff(%rip),%xmm10 # 4200 <_sk_callback_sse2+0x453> + DB 68,15,40,21,101,39,0,0 ; movaps 0x2765(%rip),%xmm10 # 4290 <_sk_callback_sse2+0x4b9> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -17133,15 +17196,15 @@ _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 68,15,89,13,62,36,0,0 ; mulps 0x243e(%rip),%xmm9 # 4210 <_sk_callback_sse2+0x463> - DB 68,15,84,21,70,36,0,0 ; andps 0x2446(%rip),%xmm10 # 4220 <_sk_callback_sse2+0x473> - DB 68,15,86,21,78,36,0,0 ; orps 0x244e(%rip),%xmm10 # 4230 <_sk_callback_sse2+0x483> - DB 68,15,88,13,86,36,0,0 ; addps 0x2456(%rip),%xmm9 # 4240 <_sk_callback_sse2+0x493> - DB 68,15,40,37,94,36,0,0 ; movaps 0x245e(%rip),%xmm12 # 4250 <_sk_callback_sse2+0x4a3> + DB 68,15,89,13,164,36,0,0 ; mulps 0x24a4(%rip),%xmm9 # 42a0 <_sk_callback_sse2+0x4c9> + DB 68,15,84,21,172,36,0,0 ; andps 0x24ac(%rip),%xmm10 # 42b0 <_sk_callback_sse2+0x4d9> + DB 68,15,86,21,180,36,0,0 ; orps 0x24b4(%rip),%xmm10 # 42c0 <_sk_callback_sse2+0x4e9> + DB 68,15,88,13,188,36,0,0 ; addps 0x24bc(%rip),%xmm9 # 42d0 <_sk_callback_sse2+0x4f9> + DB 68,15,40,37,196,36,0,0 ; movaps 0x24c4(%rip),%xmm12 # 42e0 <_sk_callback_sse2+0x509> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,94,36,0,0 ; addps 0x245e(%rip),%xmm10 # 4260 <_sk_callback_sse2+0x4b3> - DB 68,15,40,37,102,36,0,0 ; movaps 0x2466(%rip),%xmm12 # 4270 <_sk_callback_sse2+0x4c3> + DB 68,15,88,21,196,36,0,0 ; addps 0x24c4(%rip),%xmm10 # 42f0 <_sk_callback_sse2+0x519> + DB 68,15,40,37,204,36,0,0 ; movaps 0x24cc(%rip),%xmm12 # 4300 <_sk_callback_sse2+0x529> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -17149,22 +17212,22 @@ _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 68,15,40,21,80,36,0,0 ; movaps 0x2450(%rip),%xmm10 # 4280 <_sk_callback_sse2+0x4d3> + DB 68,15,40,21,182,36,0,0 ; movaps 0x24b6(%rip),%xmm10 # 4310 <_sk_callback_sse2+0x539> 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 68,15,88,13,68,36,0,0 ; addps 0x2444(%rip),%xmm9 # 4290 <_sk_callback_sse2+0x4e3> - DB 68,15,40,37,76,36,0,0 ; movaps 0x244c(%rip),%xmm12 # 42a0 <_sk_callback_sse2+0x4f3> + DB 68,15,88,13,170,36,0,0 ; addps 0x24aa(%rip),%xmm9 # 4320 <_sk_callback_sse2+0x549> + DB 68,15,40,37,178,36,0,0 ; movaps 0x24b2(%rip),%xmm12 # 4330 <_sk_callback_sse2+0x559> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,76,36,0,0 ; movaps 0x244c(%rip),%xmm12 # 42b0 <_sk_callback_sse2+0x503> + DB 68,15,40,37,178,36,0,0 ; movaps 0x24b2(%rip),%xmm12 # 4340 <_sk_callback_sse2+0x569> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,80,36,0,0 ; movaps 0x2450(%rip),%xmm13 # 42c0 <_sk_callback_sse2+0x513> + DB 68,15,40,45,182,36,0,0 ; movaps 0x24b6(%rip),%xmm13 # 4350 <_sk_callback_sse2+0x579> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,80,36,0,0 ; mulps 0x2450(%rip),%xmm13 # 42d0 <_sk_callback_sse2+0x523> + DB 68,15,89,45,182,36,0,0 ; mulps 0x24b6(%rip),%xmm13 # 4360 <_sk_callback_sse2+0x589> 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 @@ -17198,15 +17261,15 @@ _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 68,15,89,13,208,35,0,0 ; mulps 0x23d0(%rip),%xmm9 # 42e0 <_sk_callback_sse2+0x533> - DB 68,15,84,21,216,35,0,0 ; andps 0x23d8(%rip),%xmm10 # 42f0 <_sk_callback_sse2+0x543> - DB 68,15,86,21,224,35,0,0 ; orps 0x23e0(%rip),%xmm10 # 4300 <_sk_callback_sse2+0x553> - DB 68,15,88,13,232,35,0,0 ; addps 0x23e8(%rip),%xmm9 # 4310 <_sk_callback_sse2+0x563> - DB 68,15,40,37,240,35,0,0 ; movaps 0x23f0(%rip),%xmm12 # 4320 <_sk_callback_sse2+0x573> + DB 68,15,89,13,54,36,0,0 ; mulps 0x2436(%rip),%xmm9 # 4370 <_sk_callback_sse2+0x599> + DB 68,15,84,21,62,36,0,0 ; andps 0x243e(%rip),%xmm10 # 4380 <_sk_callback_sse2+0x5a9> + DB 68,15,86,21,70,36,0,0 ; orps 0x2446(%rip),%xmm10 # 4390 <_sk_callback_sse2+0x5b9> + DB 68,15,88,13,78,36,0,0 ; addps 0x244e(%rip),%xmm9 # 43a0 <_sk_callback_sse2+0x5c9> + DB 68,15,40,37,86,36,0,0 ; movaps 0x2456(%rip),%xmm12 # 43b0 <_sk_callback_sse2+0x5d9> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,240,35,0,0 ; addps 0x23f0(%rip),%xmm10 # 4330 <_sk_callback_sse2+0x583> - DB 68,15,40,37,248,35,0,0 ; movaps 0x23f8(%rip),%xmm12 # 4340 <_sk_callback_sse2+0x593> + DB 68,15,88,21,86,36,0,0 ; addps 0x2456(%rip),%xmm10 # 43c0 <_sk_callback_sse2+0x5e9> + DB 68,15,40,37,94,36,0,0 ; movaps 0x245e(%rip),%xmm12 # 43d0 <_sk_callback_sse2+0x5f9> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -17214,22 +17277,22 @@ _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 68,15,40,21,226,35,0,0 ; movaps 0x23e2(%rip),%xmm10 # 4350 <_sk_callback_sse2+0x5a3> + DB 68,15,40,21,72,36,0,0 ; movaps 0x2448(%rip),%xmm10 # 43e0 <_sk_callback_sse2+0x609> 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 68,15,88,13,214,35,0,0 ; addps 0x23d6(%rip),%xmm9 # 4360 <_sk_callback_sse2+0x5b3> - DB 68,15,40,37,222,35,0,0 ; movaps 0x23de(%rip),%xmm12 # 4370 <_sk_callback_sse2+0x5c3> + DB 68,15,88,13,60,36,0,0 ; addps 0x243c(%rip),%xmm9 # 43f0 <_sk_callback_sse2+0x619> + DB 68,15,40,37,68,36,0,0 ; movaps 0x2444(%rip),%xmm12 # 4400 <_sk_callback_sse2+0x629> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,222,35,0,0 ; movaps 0x23de(%rip),%xmm12 # 4380 <_sk_callback_sse2+0x5d3> + DB 68,15,40,37,68,36,0,0 ; movaps 0x2444(%rip),%xmm12 # 4410 <_sk_callback_sse2+0x639> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,226,35,0,0 ; movaps 0x23e2(%rip),%xmm13 # 4390 <_sk_callback_sse2+0x5e3> + DB 68,15,40,45,72,36,0,0 ; movaps 0x2448(%rip),%xmm13 # 4420 <_sk_callback_sse2+0x649> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,226,35,0,0 ; mulps 0x23e2(%rip),%xmm13 # 43a0 <_sk_callback_sse2+0x5f3> + DB 68,15,89,45,72,36,0,0 ; mulps 0x2448(%rip),%xmm13 # 4430 <_sk_callback_sse2+0x659> 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 @@ -17263,15 +17326,15 @@ _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 68,15,89,13,98,35,0,0 ; mulps 0x2362(%rip),%xmm9 # 43b0 <_sk_callback_sse2+0x603> - DB 68,15,84,21,106,35,0,0 ; andps 0x236a(%rip),%xmm10 # 43c0 <_sk_callback_sse2+0x613> - DB 68,15,86,21,114,35,0,0 ; orps 0x2372(%rip),%xmm10 # 43d0 <_sk_callback_sse2+0x623> - DB 68,15,88,13,122,35,0,0 ; addps 0x237a(%rip),%xmm9 # 43e0 <_sk_callback_sse2+0x633> - DB 68,15,40,37,130,35,0,0 ; movaps 0x2382(%rip),%xmm12 # 43f0 <_sk_callback_sse2+0x643> + DB 68,15,89,13,200,35,0,0 ; mulps 0x23c8(%rip),%xmm9 # 4440 <_sk_callback_sse2+0x669> + DB 68,15,84,21,208,35,0,0 ; andps 0x23d0(%rip),%xmm10 # 4450 <_sk_callback_sse2+0x679> + DB 68,15,86,21,216,35,0,0 ; orps 0x23d8(%rip),%xmm10 # 4460 <_sk_callback_sse2+0x689> + DB 68,15,88,13,224,35,0,0 ; addps 0x23e0(%rip),%xmm9 # 4470 <_sk_callback_sse2+0x699> + DB 68,15,40,37,232,35,0,0 ; movaps 0x23e8(%rip),%xmm12 # 4480 <_sk_callback_sse2+0x6a9> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,130,35,0,0 ; addps 0x2382(%rip),%xmm10 # 4400 <_sk_callback_sse2+0x653> - DB 68,15,40,37,138,35,0,0 ; movaps 0x238a(%rip),%xmm12 # 4410 <_sk_callback_sse2+0x663> + DB 68,15,88,21,232,35,0,0 ; addps 0x23e8(%rip),%xmm10 # 4490 <_sk_callback_sse2+0x6b9> + DB 68,15,40,37,240,35,0,0 ; movaps 0x23f0(%rip),%xmm12 # 44a0 <_sk_callback_sse2+0x6c9> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -17279,22 +17342,22 @@ _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 68,15,40,21,116,35,0,0 ; movaps 0x2374(%rip),%xmm10 # 4420 <_sk_callback_sse2+0x673> + DB 68,15,40,21,218,35,0,0 ; movaps 0x23da(%rip),%xmm10 # 44b0 <_sk_callback_sse2+0x6d9> 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 68,15,88,13,104,35,0,0 ; addps 0x2368(%rip),%xmm9 # 4430 <_sk_callback_sse2+0x683> - DB 68,15,40,37,112,35,0,0 ; movaps 0x2370(%rip),%xmm12 # 4440 <_sk_callback_sse2+0x693> + DB 68,15,88,13,206,35,0,0 ; addps 0x23ce(%rip),%xmm9 # 44c0 <_sk_callback_sse2+0x6e9> + DB 68,15,40,37,214,35,0,0 ; movaps 0x23d6(%rip),%xmm12 # 44d0 <_sk_callback_sse2+0x6f9> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,112,35,0,0 ; movaps 0x2370(%rip),%xmm12 # 4450 <_sk_callback_sse2+0x6a3> + DB 68,15,40,37,214,35,0,0 ; movaps 0x23d6(%rip),%xmm12 # 44e0 <_sk_callback_sse2+0x709> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,116,35,0,0 ; movaps 0x2374(%rip),%xmm13 # 4460 <_sk_callback_sse2+0x6b3> + DB 68,15,40,45,218,35,0,0 ; movaps 0x23da(%rip),%xmm13 # 44f0 <_sk_callback_sse2+0x719> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,116,35,0,0 ; mulps 0x2374(%rip),%xmm13 # 4470 <_sk_callback_sse2+0x6c3> + DB 68,15,89,45,218,35,0,0 ; mulps 0x23da(%rip),%xmm13 # 4500 <_sk_callback_sse2+0x729> 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 @@ -17328,15 +17391,15 @@ _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 68,15,89,13,244,34,0,0 ; mulps 0x22f4(%rip),%xmm9 # 4480 <_sk_callback_sse2+0x6d3> - DB 68,15,84,21,252,34,0,0 ; andps 0x22fc(%rip),%xmm10 # 4490 <_sk_callback_sse2+0x6e3> - DB 68,15,86,21,4,35,0,0 ; orps 0x2304(%rip),%xmm10 # 44a0 <_sk_callback_sse2+0x6f3> - DB 68,15,88,13,12,35,0,0 ; addps 0x230c(%rip),%xmm9 # 44b0 <_sk_callback_sse2+0x703> - DB 68,15,40,37,20,35,0,0 ; movaps 0x2314(%rip),%xmm12 # 44c0 <_sk_callback_sse2+0x713> + DB 68,15,89,13,90,35,0,0 ; mulps 0x235a(%rip),%xmm9 # 4510 <_sk_callback_sse2+0x739> + DB 68,15,84,21,98,35,0,0 ; andps 0x2362(%rip),%xmm10 # 4520 <_sk_callback_sse2+0x749> + DB 68,15,86,21,106,35,0,0 ; orps 0x236a(%rip),%xmm10 # 4530 <_sk_callback_sse2+0x759> + DB 68,15,88,13,114,35,0,0 ; addps 0x2372(%rip),%xmm9 # 4540 <_sk_callback_sse2+0x769> + DB 68,15,40,37,122,35,0,0 ; movaps 0x237a(%rip),%xmm12 # 4550 <_sk_callback_sse2+0x779> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,20,35,0,0 ; addps 0x2314(%rip),%xmm10 # 44d0 <_sk_callback_sse2+0x723> - DB 68,15,40,37,28,35,0,0 ; movaps 0x231c(%rip),%xmm12 # 44e0 <_sk_callback_sse2+0x733> + DB 68,15,88,21,122,35,0,0 ; addps 0x237a(%rip),%xmm10 # 4560 <_sk_callback_sse2+0x789> + DB 68,15,40,37,130,35,0,0 ; movaps 0x2382(%rip),%xmm12 # 4570 <_sk_callback_sse2+0x799> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -17344,22 +17407,22 @@ _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 68,15,40,21,6,35,0,0 ; movaps 0x2306(%rip),%xmm10 # 44f0 <_sk_callback_sse2+0x743> + DB 68,15,40,21,108,35,0,0 ; movaps 0x236c(%rip),%xmm10 # 4580 <_sk_callback_sse2+0x7a9> 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 68,15,88,13,250,34,0,0 ; addps 0x22fa(%rip),%xmm9 # 4500 <_sk_callback_sse2+0x753> - DB 68,15,40,37,2,35,0,0 ; movaps 0x2302(%rip),%xmm12 # 4510 <_sk_callback_sse2+0x763> + DB 68,15,88,13,96,35,0,0 ; addps 0x2360(%rip),%xmm9 # 4590 <_sk_callback_sse2+0x7b9> + DB 68,15,40,37,104,35,0,0 ; movaps 0x2368(%rip),%xmm12 # 45a0 <_sk_callback_sse2+0x7c9> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,2,35,0,0 ; movaps 0x2302(%rip),%xmm12 # 4520 <_sk_callback_sse2+0x773> + DB 68,15,40,37,104,35,0,0 ; movaps 0x2368(%rip),%xmm12 # 45b0 <_sk_callback_sse2+0x7d9> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,6,35,0,0 ; movaps 0x2306(%rip),%xmm13 # 4530 <_sk_callback_sse2+0x783> + DB 68,15,40,45,108,35,0,0 ; movaps 0x236c(%rip),%xmm13 # 45c0 <_sk_callback_sse2+0x7e9> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,6,35,0,0 ; mulps 0x2306(%rip),%xmm13 # 4540 <_sk_callback_sse2+0x793> + DB 68,15,89,45,108,35,0,0 ; mulps 0x236c(%rip),%xmm13 # 45d0 <_sk_callback_sse2+0x7f9> 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 @@ -17374,29 +17437,29 @@ _sk_parametric_a_sse2 LABEL PROC PUBLIC _sk_lab_to_xyz_sse2 _sk_lab_to_xyz_sse2 LABEL PROC - DB 15,89,5,227,34,0,0 ; mulps 0x22e3(%rip),%xmm0 # 4550 <_sk_callback_sse2+0x7a3> - DB 68,15,40,5,235,34,0,0 ; movaps 0x22eb(%rip),%xmm8 # 4560 <_sk_callback_sse2+0x7b3> + DB 15,89,5,73,35,0,0 ; mulps 0x2349(%rip),%xmm0 # 45e0 <_sk_callback_sse2+0x809> + DB 68,15,40,5,81,35,0,0 ; movaps 0x2351(%rip),%xmm8 # 45f0 <_sk_callback_sse2+0x819> DB 65,15,89,200 ; mulps %xmm8,%xmm1 - DB 68,15,40,13,239,34,0,0 ; movaps 0x22ef(%rip),%xmm9 # 4570 <_sk_callback_sse2+0x7c3> + DB 68,15,40,13,85,35,0,0 ; movaps 0x2355(%rip),%xmm9 # 4600 <_sk_callback_sse2+0x829> 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 15,88,5,236,34,0,0 ; addps 0x22ec(%rip),%xmm0 # 4580 <_sk_callback_sse2+0x7d3> - DB 15,89,5,245,34,0,0 ; mulps 0x22f5(%rip),%xmm0 # 4590 <_sk_callback_sse2+0x7e3> - DB 15,89,13,254,34,0,0 ; mulps 0x22fe(%rip),%xmm1 # 45a0 <_sk_callback_sse2+0x7f3> + DB 15,88,5,82,35,0,0 ; addps 0x2352(%rip),%xmm0 # 4610 <_sk_callback_sse2+0x839> + DB 15,89,5,91,35,0,0 ; mulps 0x235b(%rip),%xmm0 # 4620 <_sk_callback_sse2+0x849> + DB 15,89,13,100,35,0,0 ; mulps 0x2364(%rip),%xmm1 # 4630 <_sk_callback_sse2+0x859> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,21,4,35,0,0 ; mulps 0x2304(%rip),%xmm2 # 45b0 <_sk_callback_sse2+0x803> + DB 15,89,21,106,35,0,0 ; mulps 0x236a(%rip),%xmm2 # 4640 <_sk_callback_sse2+0x869> 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 15,40,21,249,34,0,0 ; movaps 0x22f9(%rip),%xmm2 # 45c0 <_sk_callback_sse2+0x813> + DB 15,40,21,95,35,0,0 ; movaps 0x235f(%rip),%xmm2 # 4650 <_sk_callback_sse2+0x879> DB 68,15,40,194 ; movaps %xmm2,%xmm8 DB 69,15,194,196,1 ; cmpltps %xmm12,%xmm8 - DB 68,15,40,21,248,34,0,0 ; movaps 0x22f8(%rip),%xmm10 # 45d0 <_sk_callback_sse2+0x823> + DB 68,15,40,21,94,35,0,0 ; movaps 0x235e(%rip),%xmm10 # 4660 <_sk_callback_sse2+0x889> DB 65,15,88,202 ; addps %xmm10,%xmm1 - DB 68,15,40,29,252,34,0,0 ; movaps 0x22fc(%rip),%xmm11 # 45e0 <_sk_callback_sse2+0x833> + DB 68,15,40,29,98,35,0,0 ; movaps 0x2362(%rip),%xmm11 # 4670 <_sk_callback_sse2+0x899> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 69,15,84,224 ; andps %xmm8,%xmm12 DB 68,15,85,193 ; andnps %xmm1,%xmm8 @@ -17420,8 +17483,8 @@ _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 68,15,89,5,172,34,0,0 ; mulps 0x22ac(%rip),%xmm8 # 45f0 <_sk_callback_sse2+0x843> - DB 15,89,21,181,34,0,0 ; mulps 0x22b5(%rip),%xmm2 # 4600 <_sk_callback_sse2+0x853> + DB 68,15,89,5,18,35,0,0 ; mulps 0x2312(%rip),%xmm8 # 4680 <_sk_callback_sse2+0x8a9> + DB 15,89,21,27,35,0,0 ; mulps 0x231b(%rip),%xmm2 # 4690 <_sk_callback_sse2+0x8b9> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -17435,7 +17498,7 @@ _sk_load_a8_sse2 LABEL PROC DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,157,34,0,0 ; mulps 0x229d(%rip),%xmm3 # 4610 <_sk_callback_sse2+0x863> + DB 15,89,29,3,35,0,0 ; mulps 0x2303(%rip),%xmm3 # 46a0 <_sk_callback_sse2+0x8c9> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -17478,7 +17541,7 @@ _sk_gather_a8_sse2 LABEL PROC DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,12,34,0,0 ; mulps 0x220c(%rip),%xmm3 # 4620 <_sk_callback_sse2+0x873> + DB 15,89,29,114,34,0,0 ; mulps 0x2272(%rip),%xmm3 # 46b0 <_sk_callback_sse2+0x8d9> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -17489,7 +17552,7 @@ PUBLIC _sk_store_a8_sse2 _sk_store_a8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,0,34,0,0 ; movaps 0x2200(%rip),%xmm8 # 4630 <_sk_callback_sse2+0x883> + DB 68,15,40,5,102,34,0,0 ; movaps 0x2266(%rip),%xmm8 # 46c0 <_sk_callback_sse2+0x8e9> 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 @@ -17509,9 +17572,9 @@ _sk_load_g8_sse2 LABEL PROC DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,199,33,0,0 ; mulps 0x21c7(%rip),%xmm0 # 4640 <_sk_callback_sse2+0x893> + DB 15,89,5,45,34,0,0 ; mulps 0x222d(%rip),%xmm0 # 46d0 <_sk_callback_sse2+0x8f9> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,206,33,0,0 ; movaps 0x21ce(%rip),%xmm3 # 4650 <_sk_callback_sse2+0x8a3> + DB 15,40,29,52,34,0,0 ; movaps 0x2234(%rip),%xmm3 # 46e0 <_sk_callback_sse2+0x909> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -17552,9 +17615,9 @@ _sk_gather_g8_sse2 LABEL PROC DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,67,33,0,0 ; mulps 0x2143(%rip),%xmm0 # 4660 <_sk_callback_sse2+0x8b3> + DB 15,89,5,169,33,0,0 ; mulps 0x21a9(%rip),%xmm0 # 46f0 <_sk_callback_sse2+0x919> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,74,33,0,0 ; movaps 0x214a(%rip),%xmm3 # 4670 <_sk_callback_sse2+0x8c3> + DB 15,40,29,176,33,0,0 ; movaps 0x21b0(%rip),%xmm3 # 4700 <_sk_callback_sse2+0x929> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -17564,9 +17627,9 @@ _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 253d <_sk_gather_i8_sse2+0xf> + DB 116,5 ; je 2567 <_sk_gather_i8_sse2+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 253f <_sk_gather_i8_sse2+0x11> + DB 235,2 ; jmp 2569 <_sk_gather_i8_sse2+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 @@ -17615,11 +17678,11 @@ _sk_gather_i8_sse2 LABEL PROC DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,105,32,0,0 ; movdqa 0x2069(%rip),%xmm2 # 4680 <_sk_callback_sse2+0x8d3> + DB 102,15,111,21,207,32,0,0 ; movdqa 0x20cf(%rip),%xmm2 # 4710 <_sk_callback_sse2+0x939> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,101,32,0,0 ; movaps 0x2065(%rip),%xmm8 # 4690 <_sk_callback_sse2+0x8e3> + DB 68,15,40,5,203,32,0,0 ; movaps 0x20cb(%rip),%xmm8 # 4720 <_sk_callback_sse2+0x949> 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 @@ -17644,19 +17707,19 @@ _sk_load_565_sse2 LABEL PROC DB 243,15,126,20,120 ; movq (%rax,%rdi,2),%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,27,32,0,0 ; movdqa 0x201b(%rip),%xmm0 # 46a0 <_sk_callback_sse2+0x8f3> + DB 102,15,111,5,129,32,0,0 ; movdqa 0x2081(%rip),%xmm0 # 4730 <_sk_callback_sse2+0x959> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,29,32,0,0 ; mulps 0x201d(%rip),%xmm0 # 46b0 <_sk_callback_sse2+0x903> - DB 102,15,111,13,37,32,0,0 ; movdqa 0x2025(%rip),%xmm1 # 46c0 <_sk_callback_sse2+0x913> + DB 15,89,5,131,32,0,0 ; mulps 0x2083(%rip),%xmm0 # 4740 <_sk_callback_sse2+0x969> + DB 102,15,111,13,139,32,0,0 ; movdqa 0x208b(%rip),%xmm1 # 4750 <_sk_callback_sse2+0x979> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,39,32,0,0 ; mulps 0x2027(%rip),%xmm1 # 46d0 <_sk_callback_sse2+0x923> - DB 102,15,219,21,47,32,0,0 ; pand 0x202f(%rip),%xmm2 # 46e0 <_sk_callback_sse2+0x933> + DB 15,89,13,141,32,0,0 ; mulps 0x208d(%rip),%xmm1 # 4760 <_sk_callback_sse2+0x989> + DB 102,15,219,21,149,32,0,0 ; pand 0x2095(%rip),%xmm2 # 4770 <_sk_callback_sse2+0x999> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,53,32,0,0 ; mulps 0x2035(%rip),%xmm2 # 46f0 <_sk_callback_sse2+0x943> + DB 15,89,21,155,32,0,0 ; mulps 0x209b(%rip),%xmm2 # 4780 <_sk_callback_sse2+0x9a9> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,60,32,0,0 ; movaps 0x203c(%rip),%xmm3 # 4700 <_sk_callback_sse2+0x953> + DB 15,40,29,162,32,0,0 ; movaps 0x20a2(%rip),%xmm3 # 4790 <_sk_callback_sse2+0x9b9> DB 255,224 ; jmpq *%rax PUBLIC _sk_gather_565_sse2 @@ -17689,31 +17752,31 @@ _sk_gather_565_sse2 LABEL PROC 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 102,15,111,5,197,31,0,0 ; movdqa 0x1fc5(%rip),%xmm0 # 4710 <_sk_callback_sse2+0x963> + DB 102,15,111,5,43,32,0,0 ; movdqa 0x202b(%rip),%xmm0 # 47a0 <_sk_callback_sse2+0x9c9> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,199,31,0,0 ; mulps 0x1fc7(%rip),%xmm0 # 4720 <_sk_callback_sse2+0x973> - DB 102,15,111,13,207,31,0,0 ; movdqa 0x1fcf(%rip),%xmm1 # 4730 <_sk_callback_sse2+0x983> + DB 15,89,5,45,32,0,0 ; mulps 0x202d(%rip),%xmm0 # 47b0 <_sk_callback_sse2+0x9d9> + DB 102,15,111,13,53,32,0,0 ; movdqa 0x2035(%rip),%xmm1 # 47c0 <_sk_callback_sse2+0x9e9> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,209,31,0,0 ; mulps 0x1fd1(%rip),%xmm1 # 4740 <_sk_callback_sse2+0x993> - DB 102,15,219,21,217,31,0,0 ; pand 0x1fd9(%rip),%xmm2 # 4750 <_sk_callback_sse2+0x9a3> + DB 15,89,13,55,32,0,0 ; mulps 0x2037(%rip),%xmm1 # 47d0 <_sk_callback_sse2+0x9f9> + DB 102,15,219,21,63,32,0,0 ; pand 0x203f(%rip),%xmm2 # 47e0 <_sk_callback_sse2+0xa09> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,223,31,0,0 ; mulps 0x1fdf(%rip),%xmm2 # 4760 <_sk_callback_sse2+0x9b3> + DB 15,89,21,69,32,0,0 ; mulps 0x2045(%rip),%xmm2 # 47f0 <_sk_callback_sse2+0xa19> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,230,31,0,0 ; movaps 0x1fe6(%rip),%xmm3 # 4770 <_sk_callback_sse2+0x9c3> + DB 15,40,29,76,32,0,0 ; movaps 0x204c(%rip),%xmm3 # 4800 <_sk_callback_sse2+0xa29> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_sse2 _sk_store_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,231,31,0,0 ; movaps 0x1fe7(%rip),%xmm8 # 4780 <_sk_callback_sse2+0x9d3> + DB 68,15,40,5,77,32,0,0 ; movaps 0x204d(%rip),%xmm8 # 4810 <_sk_callback_sse2+0xa39> 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 68,15,40,21,220,31,0,0 ; movaps 0x1fdc(%rip),%xmm10 # 4790 <_sk_callback_sse2+0x9e3> + DB 68,15,40,21,66,32,0,0 ; movaps 0x2042(%rip),%xmm10 # 4820 <_sk_callback_sse2+0xa49> 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 @@ -17735,21 +17798,21 @@ _sk_load_4444_sse2 LABEL PROC DB 243,15,126,28,120 ; movq (%rax,%rdi,2),%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,149,31,0,0 ; movdqa 0x1f95(%rip),%xmm0 # 47a0 <_sk_callback_sse2+0x9f3> + DB 102,15,111,5,251,31,0,0 ; movdqa 0x1ffb(%rip),%xmm0 # 4830 <_sk_callback_sse2+0xa59> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,151,31,0,0 ; mulps 0x1f97(%rip),%xmm0 # 47b0 <_sk_callback_sse2+0xa03> - DB 102,15,111,13,159,31,0,0 ; movdqa 0x1f9f(%rip),%xmm1 # 47c0 <_sk_callback_sse2+0xa13> + DB 15,89,5,253,31,0,0 ; mulps 0x1ffd(%rip),%xmm0 # 4840 <_sk_callback_sse2+0xa69> + DB 102,15,111,13,5,32,0,0 ; movdqa 0x2005(%rip),%xmm1 # 4850 <_sk_callback_sse2+0xa79> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,161,31,0,0 ; mulps 0x1fa1(%rip),%xmm1 # 47d0 <_sk_callback_sse2+0xa23> - DB 102,15,111,21,169,31,0,0 ; movdqa 0x1fa9(%rip),%xmm2 # 47e0 <_sk_callback_sse2+0xa33> + DB 15,89,13,7,32,0,0 ; mulps 0x2007(%rip),%xmm1 # 4860 <_sk_callback_sse2+0xa89> + DB 102,15,111,21,15,32,0,0 ; movdqa 0x200f(%rip),%xmm2 # 4870 <_sk_callback_sse2+0xa99> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,171,31,0,0 ; mulps 0x1fab(%rip),%xmm2 # 47f0 <_sk_callback_sse2+0xa43> - DB 102,15,219,29,179,31,0,0 ; pand 0x1fb3(%rip),%xmm3 # 4800 <_sk_callback_sse2+0xa53> + DB 15,89,21,17,32,0,0 ; mulps 0x2011(%rip),%xmm2 # 4880 <_sk_callback_sse2+0xaa9> + DB 102,15,219,29,25,32,0,0 ; pand 0x2019(%rip),%xmm3 # 4890 <_sk_callback_sse2+0xab9> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,185,31,0,0 ; mulps 0x1fb9(%rip),%xmm3 # 4810 <_sk_callback_sse2+0xa63> + DB 15,89,29,31,32,0,0 ; mulps 0x201f(%rip),%xmm3 # 48a0 <_sk_callback_sse2+0xac9> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -17783,21 +17846,21 @@ _sk_gather_4444_sse2 LABEL PROC 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 102,15,111,5,64,31,0,0 ; movdqa 0x1f40(%rip),%xmm0 # 4820 <_sk_callback_sse2+0xa73> + DB 102,15,111,5,166,31,0,0 ; movdqa 0x1fa6(%rip),%xmm0 # 48b0 <_sk_callback_sse2+0xad9> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,66,31,0,0 ; mulps 0x1f42(%rip),%xmm0 # 4830 <_sk_callback_sse2+0xa83> - DB 102,15,111,13,74,31,0,0 ; movdqa 0x1f4a(%rip),%xmm1 # 4840 <_sk_callback_sse2+0xa93> + DB 15,89,5,168,31,0,0 ; mulps 0x1fa8(%rip),%xmm0 # 48c0 <_sk_callback_sse2+0xae9> + DB 102,15,111,13,176,31,0,0 ; movdqa 0x1fb0(%rip),%xmm1 # 48d0 <_sk_callback_sse2+0xaf9> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,76,31,0,0 ; mulps 0x1f4c(%rip),%xmm1 # 4850 <_sk_callback_sse2+0xaa3> - DB 102,15,111,21,84,31,0,0 ; movdqa 0x1f54(%rip),%xmm2 # 4860 <_sk_callback_sse2+0xab3> + DB 15,89,13,178,31,0,0 ; mulps 0x1fb2(%rip),%xmm1 # 48e0 <_sk_callback_sse2+0xb09> + DB 102,15,111,21,186,31,0,0 ; movdqa 0x1fba(%rip),%xmm2 # 48f0 <_sk_callback_sse2+0xb19> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,86,31,0,0 ; mulps 0x1f56(%rip),%xmm2 # 4870 <_sk_callback_sse2+0xac3> - DB 102,15,219,29,94,31,0,0 ; pand 0x1f5e(%rip),%xmm3 # 4880 <_sk_callback_sse2+0xad3> + DB 15,89,21,188,31,0,0 ; mulps 0x1fbc(%rip),%xmm2 # 4900 <_sk_callback_sse2+0xb29> + DB 102,15,219,29,196,31,0,0 ; pand 0x1fc4(%rip),%xmm3 # 4910 <_sk_callback_sse2+0xb39> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,100,31,0,0 ; mulps 0x1f64(%rip),%xmm3 # 4890 <_sk_callback_sse2+0xae3> + DB 15,89,29,202,31,0,0 ; mulps 0x1fca(%rip),%xmm3 # 4920 <_sk_callback_sse2+0xb49> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -17805,7 +17868,7 @@ PUBLIC _sk_store_4444_sse2 _sk_store_4444_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,99,31,0,0 ; movaps 0x1f63(%rip),%xmm8 # 48a0 <_sk_callback_sse2+0xaf3> + DB 68,15,40,5,201,31,0,0 ; movaps 0x1fc9(%rip),%xmm8 # 4930 <_sk_callback_sse2+0xb59> 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 @@ -17835,11 +17898,11 @@ _sk_load_8888_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 68,15,16,12,184 ; movups (%rax,%rdi,4),%xmm9 - DB 15,40,21,246,30,0,0 ; movaps 0x1ef6(%rip),%xmm2 # 48b0 <_sk_callback_sse2+0xb03> + DB 15,40,21,92,31,0,0 ; movaps 0x1f5c(%rip),%xmm2 # 4940 <_sk_callback_sse2+0xb69> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,84,194 ; andps %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,244,30,0,0 ; movaps 0x1ef4(%rip),%xmm8 # 48c0 <_sk_callback_sse2+0xb13> + DB 68,15,40,5,90,31,0,0 ; movaps 0x1f5a(%rip),%xmm8 # 4950 <_sk_callback_sse2+0xb79> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,40,201 ; movaps %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -17886,11 +17949,11 @@ _sk_gather_8888_sse2 LABEL PROC DB 102,67,15,110,12,129 ; movd (%r9,%r8,4),%xmm1 DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,69,30,0,0 ; movdqa 0x1e45(%rip),%xmm2 # 48d0 <_sk_callback_sse2+0xb23> + DB 102,15,111,21,171,30,0,0 ; movdqa 0x1eab(%rip),%xmm2 # 4960 <_sk_callback_sse2+0xb89> 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 68,15,40,5,65,30,0,0 ; movaps 0x1e41(%rip),%xmm8 # 48e0 <_sk_callback_sse2+0xb33> + DB 68,15,40,5,167,30,0,0 ; movaps 0x1ea7(%rip),%xmm8 # 4970 <_sk_callback_sse2+0xb99> 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 @@ -17912,7 +17975,7 @@ 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 68,15,40,5,4,30,0,0 ; movaps 0x1e04(%rip),%xmm8 # 48f0 <_sk_callback_sse2+0xb43> + DB 68,15,40,5,106,30,0,0 ; movaps 0x1e6a(%rip),%xmm8 # 4980 <_sk_callback_sse2+0xba9> 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 @@ -17949,7 +18012,7 @@ _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 102,68,15,111,13,116,29,0,0 ; movdqa 0x1d74(%rip),%xmm9 # 4900 <_sk_callback_sse2+0xb53> + DB 102,68,15,111,13,218,29,0,0 ; movdqa 0x1dda(%rip),%xmm9 # 4990 <_sk_callback_sse2+0xbb9> 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 @@ -17957,11 +18020,11 @@ _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 102,68,15,111,29,89,29,0,0 ; movdqa 0x1d59(%rip),%xmm11 # 4910 <_sk_callback_sse2+0xb63> + DB 102,68,15,111,29,191,29,0,0 ; movdqa 0x1dbf(%rip),%xmm11 # 49a0 <_sk_callback_sse2+0xbc9> DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,91,29,0,0 ; movdqa 0x1d5b(%rip),%xmm12 # 4920 <_sk_callback_sse2+0xb73> + DB 102,68,15,111,37,193,29,0,0 ; movdqa 0x1dc1(%rip),%xmm12 # 49b0 <_sk_callback_sse2+0xbd9> DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,94,29,0,0 ; movdqa 0x1d5e(%rip),%xmm3 # 4930 <_sk_callback_sse2+0xb83> + DB 102,15,111,29,196,29,0,0 ; movdqa 0x1dc4(%rip),%xmm3 # 49c0 <_sk_callback_sse2+0xbe9> 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 @@ -18045,7 +18108,7 @@ _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 102,68,15,111,13,236,27,0,0 ; movdqa 0x1bec(%rip),%xmm9 # 4940 <_sk_callback_sse2+0xb93> + DB 102,68,15,111,13,82,28,0,0 ; movdqa 0x1c52(%rip),%xmm9 # 49d0 <_sk_callback_sse2+0xbf9> 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 @@ -18053,11 +18116,11 @@ _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 102,68,15,111,29,209,27,0,0 ; movdqa 0x1bd1(%rip),%xmm11 # 4950 <_sk_callback_sse2+0xba3> + DB 102,68,15,111,29,55,28,0,0 ; movdqa 0x1c37(%rip),%xmm11 # 49e0 <_sk_callback_sse2+0xc09> DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,211,27,0,0 ; movdqa 0x1bd3(%rip),%xmm12 # 4960 <_sk_callback_sse2+0xbb3> + DB 102,68,15,111,37,57,28,0,0 ; movdqa 0x1c39(%rip),%xmm12 # 49f0 <_sk_callback_sse2+0xc19> DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,214,27,0,0 ; movdqa 0x1bd6(%rip),%xmm3 # 4970 <_sk_callback_sse2+0xbc3> + DB 102,15,111,29,60,28,0,0 ; movdqa 0x1c3c(%rip),%xmm3 # 4a00 <_sk_callback_sse2+0xc29> 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 @@ -18108,17 +18171,17 @@ 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 102,68,15,111,21,254,26,0,0 ; movdqa 0x1afe(%rip),%xmm10 # 4980 <_sk_callback_sse2+0xbd3> + DB 102,68,15,111,21,100,27,0,0 ; movdqa 0x1b64(%rip),%xmm10 # 4a10 <_sk_callback_sse2+0xc39> DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 DB 102,69,15,219,234 ; pand %xmm10,%xmm13 DB 102,69,15,239,229 ; pxor %xmm13,%xmm12 - DB 102,68,15,111,13,241,26,0,0 ; movdqa 0x1af1(%rip),%xmm9 # 4990 <_sk_callback_sse2+0xbe3> + DB 102,68,15,111,13,87,27,0,0 ; movdqa 0x1b57(%rip),%xmm9 # 4a20 <_sk_callback_sse2+0xc49> DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13 DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 DB 102,69,15,102,196 ; pcmpgtd %xmm12,%xmm8 DB 102,65,15,114,212,13 ; psrld $0xd,%xmm12 - DB 102,68,15,111,29,226,26,0,0 ; movdqa 0x1ae2(%rip),%xmm11 # 49a0 <_sk_callback_sse2+0xbf3> + DB 102,68,15,111,29,72,27,0,0 ; movdqa 0x1b48(%rip),%xmm11 # 4a30 <_sk_callback_sse2+0xc59> DB 102,69,15,235,235 ; por %xmm11,%xmm13 DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 @@ -18195,7 +18258,7 @@ _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 68,15,40,5,128,25,0,0 ; movaps 0x1980(%rip),%xmm8 # 49b0 <_sk_callback_sse2+0xc03> + DB 68,15,40,5,230,25,0,0 ; movaps 0x19e6(%rip),%xmm8 # 4a40 <_sk_callback_sse2+0xc69> 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 @@ -18246,7 +18309,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,65,15,97,192 ; punpcklwd %xmm8,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,188,24,0,0 ; movaps 0x18bc(%rip),%xmm9 # 49c0 <_sk_callback_sse2+0xc13> + DB 68,15,40,13,34,25,0,0 ; movaps 0x1922(%rip),%xmm9 # 4a50 <_sk_callback_sse2+0xc79> 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 @@ -18263,14 +18326,14 @@ _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 15,40,29,131,24,0,0 ; movaps 0x1883(%rip),%xmm3 # 49d0 <_sk_callback_sse2+0xc23> + DB 15,40,29,233,24,0,0 ; movaps 0x18e9(%rip),%xmm3 # 4a60 <_sk_callback_sse2+0xc89> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_u16_be_sse2 _sk_store_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,13,132,24,0,0 ; movaps 0x1884(%rip),%xmm9 # 49e0 <_sk_callback_sse2+0xc33> + DB 68,15,40,13,234,24,0,0 ; movaps 0x18ea(%rip),%xmm9 # 4a70 <_sk_callback_sse2+0xc99> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 @@ -18410,7 +18473,7 @@ _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 68,15,84,13,110,22,0,0 ; andps 0x166e(%rip),%xmm9 # 49f0 <_sk_callback_sse2+0xc43> + DB 68,15,84,13,212,22,0,0 ; andps 0x16d4(%rip),%xmm9 # 4a80 <_sk_callback_sse2+0xca9> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,194 ; subps %xmm10,%xmm0 @@ -18430,7 +18493,7 @@ _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 68,15,84,13,54,22,0,0 ; andps 0x1636(%rip),%xmm9 # 4a00 <_sk_callback_sse2+0xc53> + DB 68,15,84,13,156,22,0,0 ; andps 0x169c(%rip),%xmm9 # 4a90 <_sk_callback_sse2+0xcb9> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,202 ; subps %xmm10,%xmm1 @@ -18454,7 +18517,7 @@ _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 68,15,84,21,236,21,0,0 ; andps 0x15ec(%rip),%xmm10 # 4a10 <_sk_callback_sse2+0xc63> + DB 68,15,84,21,82,22,0,0 ; andps 0x1652(%rip),%xmm10 # 4aa0 <_sk_callback_sse2+0xcc9> DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 69,15,89,216 ; mulps %xmm8,%xmm11 @@ -18482,7 +18545,7 @@ _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 68,15,84,21,146,21,0,0 ; andps 0x1592(%rip),%xmm10 # 4a20 <_sk_callback_sse2+0xc73> + DB 68,15,84,21,248,21,0,0 ; andps 0x15f8(%rip),%xmm10 # 4ab0 <_sk_callback_sse2+0xcd9> DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 69,15,89,216 ; mulps %xmm8,%xmm11 @@ -18499,10 +18562,10 @@ _sk_mirror_y_sse2 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse2 _sk_luminance_to_alpha_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,106,21,0,0 ; mulps 0x156a(%rip),%xmm0 # 4a30 <_sk_callback_sse2+0xc83> - DB 15,89,13,115,21,0,0 ; mulps 0x1573(%rip),%xmm1 # 4a40 <_sk_callback_sse2+0xc93> + DB 15,89,5,208,21,0,0 ; mulps 0x15d0(%rip),%xmm0 # 4ac0 <_sk_callback_sse2+0xce9> + DB 15,89,13,217,21,0,0 ; mulps 0x15d9(%rip),%xmm1 # 4ad0 <_sk_callback_sse2+0xcf9> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,121,21,0,0 ; mulps 0x1579(%rip),%xmm3 # 4a50 <_sk_callback_sse2+0xca3> + DB 15,89,29,223,21,0,0 ; mulps 0x15df(%rip),%xmm3 # 4ae0 <_sk_callback_sse2+0xd09> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -18725,7 +18788,7 @@ _sk_linear_gradient_sse2 LABEL PROC DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 72,139,8 ; mov (%rax),%rcx DB 72,133,201 ; test %rcx,%rcx - DB 15,132,15,1,0,0 ; je 398a <_sk_linear_gradient_sse2+0x149> + DB 15,132,15,1,0,0 ; je 39b4 <_sk_linear_gradient_sse2+0x149> DB 72,139,64,8 ; mov 0x8(%rax),%rax DB 72,131,192,32 ; add $0x20,%rax DB 69,15,87,192 ; xorps %xmm8,%xmm8 @@ -18786,8 +18849,8 @@ _sk_linear_gradient_sse2 LABEL PROC DB 69,15,86,231 ; orps %xmm15,%xmm12 DB 72,131,192,36 ; add $0x24,%rax DB 72,255,201 ; dec %rcx - DB 15,133,8,255,255,255 ; jne 3890 <_sk_linear_gradient_sse2+0x4f> - DB 235,13 ; jmp 3997 <_sk_linear_gradient_sse2+0x156> + DB 15,133,8,255,255,255 ; jne 38ba <_sk_linear_gradient_sse2+0x4f> + DB 235,13 ; jmp 39c1 <_sk_linear_gradient_sse2+0x156> DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 DB 15,87,219 ; xorps %xmm3,%xmm3 @@ -18838,7 +18901,7 @@ _sk_linear_gradient_2stops_sse2 LABEL PROC PUBLIC _sk_save_xy_sse2 _sk_save_xy_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,42,16,0,0 ; movaps 0x102a(%rip),%xmm8 # 4a60 <_sk_callback_sse2+0xcb3> + DB 68,15,40,5,144,16,0,0 ; movaps 0x1090(%rip),%xmm8 # 4af0 <_sk_callback_sse2+0xd19> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -18846,7 +18909,7 @@ _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 68,15,40,37,21,16,0,0 ; movaps 0x1015(%rip),%xmm12 # 4a70 <_sk_callback_sse2+0xcc3> + DB 68,15,40,37,123,16,0,0 ; movaps 0x107b(%rip),%xmm12 # 4b00 <_sk_callback_sse2+0xd29> DB 69,15,84,220 ; andps %xmm12,%xmm11 DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 69,15,92,202 ; subps %xmm10,%xmm9 @@ -18889,8 +18952,8 @@ _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 15,88,5,142,15,0,0 ; addps 0xf8e(%rip),%xmm0 # 4a80 <_sk_callback_sse2+0xcd3> - DB 68,15,40,13,150,15,0,0 ; movaps 0xf96(%rip),%xmm9 # 4a90 <_sk_callback_sse2+0xce3> + DB 15,88,5,244,15,0,0 ; addps 0xff4(%rip),%xmm0 # 4b10 <_sk_callback_sse2+0xd39> + DB 68,15,40,13,252,15,0,0 ; movaps 0xffc(%rip),%xmm9 # 4b20 <_sk_callback_sse2+0xd49> 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 @@ -18901,7 +18964,7 @@ _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 15,88,5,133,15,0,0 ; addps 0xf85(%rip),%xmm0 # 4aa0 <_sk_callback_sse2+0xcf3> + DB 15,88,5,235,15,0,0 ; addps 0xfeb(%rip),%xmm0 # 4b30 <_sk_callback_sse2+0xd59> 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 @@ -18911,8 +18974,8 @@ _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 15,88,13,119,15,0,0 ; addps 0xf77(%rip),%xmm1 # 4ab0 <_sk_callback_sse2+0xd03> - DB 68,15,40,13,127,15,0,0 ; movaps 0xf7f(%rip),%xmm9 # 4ac0 <_sk_callback_sse2+0xd13> + DB 15,88,13,221,15,0,0 ; addps 0xfdd(%rip),%xmm1 # 4b40 <_sk_callback_sse2+0xd69> + DB 68,15,40,13,229,15,0,0 ; movaps 0xfe5(%rip),%xmm9 # 4b50 <_sk_callback_sse2+0xd79> 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 @@ -18923,7 +18986,7 @@ _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 15,88,13,109,15,0,0 ; addps 0xf6d(%rip),%xmm1 # 4ad0 <_sk_callback_sse2+0xd23> + DB 15,88,13,211,15,0,0 ; addps 0xfd3(%rip),%xmm1 # 4b60 <_sk_callback_sse2+0xd89> 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 @@ -18933,13 +18996,13 @@ _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 15,88,5,96,15,0,0 ; addps 0xf60(%rip),%xmm0 # 4ae0 <_sk_callback_sse2+0xd33> - DB 68,15,40,13,104,15,0,0 ; movaps 0xf68(%rip),%xmm9 # 4af0 <_sk_callback_sse2+0xd43> + DB 15,88,5,198,15,0,0 ; addps 0xfc6(%rip),%xmm0 # 4b70 <_sk_callback_sse2+0xd99> + DB 68,15,40,13,206,15,0,0 ; movaps 0xfce(%rip),%xmm9 # 4b80 <_sk_callback_sse2+0xda9> 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 68,15,89,13,100,15,0,0 ; mulps 0xf64(%rip),%xmm9 # 4b00 <_sk_callback_sse2+0xd53> - DB 68,15,88,13,108,15,0,0 ; addps 0xf6c(%rip),%xmm9 # 4b10 <_sk_callback_sse2+0xd63> + DB 68,15,89,13,202,15,0,0 ; mulps 0xfca(%rip),%xmm9 # 4b90 <_sk_callback_sse2+0xdb9> + DB 68,15,88,13,210,15,0,0 ; addps 0xfd2(%rip),%xmm9 # 4ba0 <_sk_callback_sse2+0xdc9> 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 @@ -18950,16 +19013,16 @@ _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 15,88,5,91,15,0,0 ; addps 0xf5b(%rip),%xmm0 # 4b20 <_sk_callback_sse2+0xd73> - DB 68,15,40,13,99,15,0,0 ; movaps 0xf63(%rip),%xmm9 # 4b30 <_sk_callback_sse2+0xd83> + DB 15,88,5,193,15,0,0 ; addps 0xfc1(%rip),%xmm0 # 4bb0 <_sk_callback_sse2+0xdd9> + DB 68,15,40,13,201,15,0,0 ; movaps 0xfc9(%rip),%xmm9 # 4bc0 <_sk_callback_sse2+0xde9> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,103,15,0,0 ; movaps 0xf67(%rip),%xmm8 # 4b40 <_sk_callback_sse2+0xd93> + DB 68,15,40,5,205,15,0,0 ; movaps 0xfcd(%rip),%xmm8 # 4bd0 <_sk_callback_sse2+0xdf9> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,107,15,0,0 ; addps 0xf6b(%rip),%xmm8 # 4b50 <_sk_callback_sse2+0xda3> + DB 68,15,88,5,209,15,0,0 ; addps 0xfd1(%rip),%xmm8 # 4be0 <_sk_callback_sse2+0xe09> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,111,15,0,0 ; addps 0xf6f(%rip),%xmm8 # 4b60 <_sk_callback_sse2+0xdb3> + DB 68,15,88,5,213,15,0,0 ; addps 0xfd5(%rip),%xmm8 # 4bf0 <_sk_callback_sse2+0xe19> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,115,15,0,0 ; addps 0xf73(%rip),%xmm8 # 4b70 <_sk_callback_sse2+0xdc3> + DB 68,15,88,5,217,15,0,0 ; addps 0xfd9(%rip),%xmm8 # 4c00 <_sk_callback_sse2+0xe29> 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 @@ -18967,17 +19030,17 @@ _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 68,15,40,5,109,15,0,0 ; movaps 0xf6d(%rip),%xmm8 # 4b80 <_sk_callback_sse2+0xdd3> + DB 68,15,40,5,211,15,0,0 ; movaps 0xfd3(%rip),%xmm8 # 4c10 <_sk_callback_sse2+0xe39> 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 68,15,40,21,105,15,0,0 ; movaps 0xf69(%rip),%xmm10 # 4b90 <_sk_callback_sse2+0xde3> + DB 68,15,40,21,207,15,0,0 ; movaps 0xfcf(%rip),%xmm10 # 4c20 <_sk_callback_sse2+0xe49> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,109,15,0,0 ; addps 0xf6d(%rip),%xmm10 # 4ba0 <_sk_callback_sse2+0xdf3> + DB 68,15,88,21,211,15,0,0 ; addps 0xfd3(%rip),%xmm10 # 4c30 <_sk_callback_sse2+0xe59> 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 68,15,88,21,105,15,0,0 ; addps 0xf69(%rip),%xmm10 # 4bb0 <_sk_callback_sse2+0xe03> + DB 68,15,88,21,207,15,0,0 ; addps 0xfcf(%rip),%xmm10 # 4c40 <_sk_callback_sse2+0xe69> 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 @@ -18987,11 +19050,11 @@ _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 15,88,5,92,15,0,0 ; addps 0xf5c(%rip),%xmm0 # 4bc0 <_sk_callback_sse2+0xe13> + DB 15,88,5,194,15,0,0 ; addps 0xfc2(%rip),%xmm0 # 4c50 <_sk_callback_sse2+0xe79> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,92,15,0,0 ; mulps 0xf5c(%rip),%xmm8 # 4bd0 <_sk_callback_sse2+0xe23> - DB 68,15,88,5,100,15,0,0 ; addps 0xf64(%rip),%xmm8 # 4be0 <_sk_callback_sse2+0xe33> + DB 68,15,89,5,194,15,0,0 ; mulps 0xfc2(%rip),%xmm8 # 4c60 <_sk_callback_sse2+0xe89> + DB 68,15,88,5,202,15,0,0 ; addps 0xfca(%rip),%xmm8 # 4c70 <_sk_callback_sse2+0xe99> 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 @@ -19002,13 +19065,13 @@ _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 15,88,13,82,15,0,0 ; addps 0xf52(%rip),%xmm1 # 4bf0 <_sk_callback_sse2+0xe43> - DB 68,15,40,13,90,15,0,0 ; movaps 0xf5a(%rip),%xmm9 # 4c00 <_sk_callback_sse2+0xe53> + DB 15,88,13,184,15,0,0 ; addps 0xfb8(%rip),%xmm1 # 4c80 <_sk_callback_sse2+0xea9> + DB 68,15,40,13,192,15,0,0 ; movaps 0xfc0(%rip),%xmm9 # 4c90 <_sk_callback_sse2+0xeb9> 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 68,15,89,13,86,15,0,0 ; mulps 0xf56(%rip),%xmm9 # 4c10 <_sk_callback_sse2+0xe63> - DB 68,15,88,13,94,15,0,0 ; addps 0xf5e(%rip),%xmm9 # 4c20 <_sk_callback_sse2+0xe73> + DB 68,15,89,13,188,15,0,0 ; mulps 0xfbc(%rip),%xmm9 # 4ca0 <_sk_callback_sse2+0xec9> + DB 68,15,88,13,196,15,0,0 ; addps 0xfc4(%rip),%xmm9 # 4cb0 <_sk_callback_sse2+0xed9> 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 @@ -19019,16 +19082,16 @@ _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 15,88,13,76,15,0,0 ; addps 0xf4c(%rip),%xmm1 # 4c30 <_sk_callback_sse2+0xe83> - DB 68,15,40,13,84,15,0,0 ; movaps 0xf54(%rip),%xmm9 # 4c40 <_sk_callback_sse2+0xe93> + DB 15,88,13,178,15,0,0 ; addps 0xfb2(%rip),%xmm1 # 4cc0 <_sk_callback_sse2+0xee9> + DB 68,15,40,13,186,15,0,0 ; movaps 0xfba(%rip),%xmm9 # 4cd0 <_sk_callback_sse2+0xef9> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,88,15,0,0 ; movaps 0xf58(%rip),%xmm8 # 4c50 <_sk_callback_sse2+0xea3> + DB 68,15,40,5,190,15,0,0 ; movaps 0xfbe(%rip),%xmm8 # 4ce0 <_sk_callback_sse2+0xf09> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,92,15,0,0 ; addps 0xf5c(%rip),%xmm8 # 4c60 <_sk_callback_sse2+0xeb3> + DB 68,15,88,5,194,15,0,0 ; addps 0xfc2(%rip),%xmm8 # 4cf0 <_sk_callback_sse2+0xf19> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,96,15,0,0 ; addps 0xf60(%rip),%xmm8 # 4c70 <_sk_callback_sse2+0xec3> + DB 68,15,88,5,198,15,0,0 ; addps 0xfc6(%rip),%xmm8 # 4d00 <_sk_callback_sse2+0xf29> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,100,15,0,0 ; addps 0xf64(%rip),%xmm8 # 4c80 <_sk_callback_sse2+0xed3> + DB 68,15,88,5,202,15,0,0 ; addps 0xfca(%rip),%xmm8 # 4d10 <_sk_callback_sse2+0xf39> 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 @@ -19036,17 +19099,17 @@ _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 68,15,40,5,94,15,0,0 ; movaps 0xf5e(%rip),%xmm8 # 4c90 <_sk_callback_sse2+0xee3> + DB 68,15,40,5,196,15,0,0 ; movaps 0xfc4(%rip),%xmm8 # 4d20 <_sk_callback_sse2+0xf49> 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 68,15,40,21,89,15,0,0 ; movaps 0xf59(%rip),%xmm10 # 4ca0 <_sk_callback_sse2+0xef3> + DB 68,15,40,21,191,15,0,0 ; movaps 0xfbf(%rip),%xmm10 # 4d30 <_sk_callback_sse2+0xf59> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,93,15,0,0 ; addps 0xf5d(%rip),%xmm10 # 4cb0 <_sk_callback_sse2+0xf03> + DB 68,15,88,21,195,15,0,0 ; addps 0xfc3(%rip),%xmm10 # 4d40 <_sk_callback_sse2+0xf69> 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 68,15,88,21,89,15,0,0 ; addps 0xf59(%rip),%xmm10 # 4cc0 <_sk_callback_sse2+0xf13> + DB 68,15,88,21,191,15,0,0 ; addps 0xfbf(%rip),%xmm10 # 4d50 <_sk_callback_sse2+0xf79> 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 @@ -19056,11 +19119,11 @@ _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 15,88,13,75,15,0,0 ; addps 0xf4b(%rip),%xmm1 # 4cd0 <_sk_callback_sse2+0xf23> + DB 15,88,13,177,15,0,0 ; addps 0xfb1(%rip),%xmm1 # 4d60 <_sk_callback_sse2+0xf89> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,75,15,0,0 ; mulps 0xf4b(%rip),%xmm8 # 4ce0 <_sk_callback_sse2+0xf33> - DB 68,15,88,5,83,15,0,0 ; addps 0xf53(%rip),%xmm8 # 4cf0 <_sk_callback_sse2+0xf43> + DB 68,15,89,5,177,15,0,0 ; mulps 0xfb1(%rip),%xmm8 # 4d70 <_sk_callback_sse2+0xf99> + DB 68,15,88,5,185,15,0,0 ; addps 0xfb9(%rip),%xmm8 # 4d80 <_sk_callback_sse2+0xfa9> 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 @@ -19231,11 +19294,11 @@ ALIGN 16 DB 0,128,191,0,0,128 ; add %al,-0x7fffff41(%rax) DB 191,0,0,224,64 ; mov $0x40e00000,%edi DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 3fe8 <.literal16+0x188> + DB 224,64 ; loopne 4018 <.literal16+0x188> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 3fec <.literal16+0x18c> + DB 224,64 ; loopne 401c <.literal16+0x18c> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 3ff0 <.literal16+0x190> + DB 224,64 ; loopne 4020 <.literal16+0x190> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -19374,6 +19437,49 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + DB 0,0 ; add %al,(%rax) + DB 128,63,171 ; cmpb $0xab,(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,171 ; ds stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,171 ; ds stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,171 ; ds stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 62,0,0 ; add %al,%ds:(%rax) + DB 128,191,0,0,128,191,0 ; cmpb $0x0,-0x40800000(%rdi) + DB 0,128,191,0,0,128 ; add %al,-0x7fffff41(%rax) + DB 191,0,0,192,64 ; mov $0x40c00000,%edi + DB 0,0 ; add %al,(%rax) + DB 192,64,0,0 ; rolb $0x0,0x0(%rax) + DB 192,64,0,0 ; rolb $0x0,0x0(%rax) + DB 192,64,171,170 ; rolb $0xaa,-0x55(%rax) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 42,63 ; sub (%rdi),%bh + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 190,171,170,170,190 ; mov $0xbeaaaaab,%esi + DB 171 ; stos %eax,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 170 ; stos %al,%es:(%rdi) + DB 190,171,170,170,190 ; mov $0xbeaaaaab,%esi 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) @@ -19395,13 +19501,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4139 <.literal16+0x2d9> + DB 224,7 ; loopne 41c9 <.literal16+0x339> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 413d <.literal16+0x2dd> + DB 224,7 ; loopne 41cd <.literal16+0x33d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 4141 <.literal16+0x2e1> + DB 224,7 ; loopne 41d1 <.literal16+0x341> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 4145 <.literal16+0x2e5> + DB 224,7 ; loopne 41d5 <.literal16+0x345> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -19470,11 +19576,11 @@ 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 422b <.literal16+0x3cb> + DB 127,67 ; jg 42bb <.literal16+0x42b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 422f <.literal16+0x3cf> + DB 127,67 ; jg 42bf <.literal16+0x42f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 4233 <.literal16+0x3d3> + DB 127,67 ; jg 42c3 <.literal16+0x433> 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) @@ -19489,16 +19595,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 4224 <.literal16+0x3c4> + DB 127,0 ; jg 42b4 <.literal16+0x424> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 4228 <.literal16+0x3c8> + DB 127,0 ; jg 42b8 <.literal16+0x428> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 422c <.literal16+0x3cc> + DB 127,0 ; jg 42bc <.literal16+0x42c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 4230 <.literal16+0x3d0> + DB 127,0 ; jg 42c0 <.literal16+0x430> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -19507,7 +19613,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 42b5 <.literal16+0x455> + DB 119,115 ; ja 4345 <.literal16+0x4b5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -19518,7 +19624,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 4219 <.literal16+0x3b9> + DB 117,191 ; jne 42a9 <.literal16+0x419> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -19530,7 +19636,7 @@ 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 ffffffffe9a3825a <_sk_callback_sse2+0xffffffffe9a344ad> + DB 233,220,63,163,233 ; jmpq ffffffffe9a382ea <_sk_callback_sse2+0xffffffffe9a34513> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -19584,16 +19690,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 42f4 <.literal16+0x494> + DB 127,0 ; jg 4384 <.literal16+0x4f4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 42f8 <.literal16+0x498> + DB 127,0 ; jg 4388 <.literal16+0x4f8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 42fc <.literal16+0x49c> + DB 127,0 ; jg 438c <.literal16+0x4fc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 4300 <.literal16+0x4a0> + DB 127,0 ; jg 4390 <.literal16+0x500> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -19602,7 +19708,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4385 <.literal16+0x525> + DB 119,115 ; ja 4415 <.literal16+0x585> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -19613,7 +19719,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 42e9 <.literal16+0x489> + DB 117,191 ; jne 4379 <.literal16+0x4e9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -19625,7 +19731,7 @@ 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 ffffffffe9a3832a <_sk_callback_sse2+0xffffffffe9a3457d> + DB 233,220,63,163,233 ; jmpq ffffffffe9a383ba <_sk_callback_sse2+0xffffffffe9a345e3> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -19679,16 +19785,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 43c4 <.literal16+0x564> + DB 127,0 ; jg 4454 <.literal16+0x5c4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 43c8 <.literal16+0x568> + DB 127,0 ; jg 4458 <.literal16+0x5c8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 43cc <.literal16+0x56c> + DB 127,0 ; jg 445c <.literal16+0x5cc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 43d0 <.literal16+0x570> + DB 127,0 ; jg 4460 <.literal16+0x5d0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -19697,7 +19803,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4455 <.literal16+0x5f5> + DB 119,115 ; ja 44e5 <.literal16+0x655> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -19708,7 +19814,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 43b9 <.literal16+0x559> + DB 117,191 ; jne 4449 <.literal16+0x5b9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -19720,7 +19826,7 @@ 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 ffffffffe9a383fa <_sk_callback_sse2+0xffffffffe9a3464d> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3848a <_sk_callback_sse2+0xffffffffe9a346b3> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -19774,16 +19880,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 4494 <.literal16+0x634> + DB 127,0 ; jg 4524 <.literal16+0x694> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 4498 <.literal16+0x638> + DB 127,0 ; jg 4528 <.literal16+0x698> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 449c <.literal16+0x63c> + DB 127,0 ; jg 452c <.literal16+0x69c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 44a0 <.literal16+0x640> + DB 127,0 ; jg 4530 <.literal16+0x6a0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -19792,7 +19898,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4525 <.literal16+0x6c5> + DB 119,115 ; ja 45b5 <.literal16+0x725> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -19803,7 +19909,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 4489 <.literal16+0x629> + DB 117,191 ; jne 4519 <.literal16+0x689> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -19815,7 +19921,7 @@ 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 ffffffffe9a384ca <_sk_callback_sse2+0xffffffffe9a3471d> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3855a <_sk_callback_sse2+0xffffffffe9a34783> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -19865,13 +19971,13 @@ 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 45a7 <.literal16+0x747> + DB 127,67 ; jg 4637 <.literal16+0x7a7> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 45ab <.literal16+0x74b> + DB 127,67 ; jg 463b <.literal16+0x7ab> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 45af <.literal16+0x74f> + DB 127,67 ; jg 463f <.literal16+0x7af> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 45b3 <.literal16+0x753> + DB 127,67 ; jg 4643 <.literal16+0x7b3> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -19918,16 +20024,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 4633 <.literal16+0x7d3> + DB 118,63 ; jbe 46c3 <.literal16+0x833> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 4637 <.literal16+0x7d7> + DB 118,63 ; jbe 46c7 <.literal16+0x837> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 463b <.literal16+0x7db> + DB 118,63 ; jbe 46cb <.literal16+0x83b> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 463f <.literal16+0x7df> + DB 118,63 ; jbe 46cf <.literal16+0x83f> 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) @@ -19939,11 +20045,11 @@ 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 467b <.literal16+0x81b> + DB 127,67 ; jg 470b <.literal16+0x87b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 467f <.literal16+0x81f> + DB 127,67 ; jg 470f <.literal16+0x87f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 4683 <.literal16+0x823> + DB 127,67 ; jg 4713 <.literal16+0x883> DB 129,128,128,59,129,128,128,59,129,128; addl $0x80813b80,-0x7f7ec480(%rax) DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,0,0,128,63 ; addb $0x3f,-0x7fffffc5(%rax) @@ -19983,13 +20089,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 46c9 <.literal16+0x869> + DB 224,7 ; loopne 4759 <.literal16+0x8c9> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 46cd <.literal16+0x86d> + DB 224,7 ; loopne 475d <.literal16+0x8cd> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 46d1 <.literal16+0x871> + DB 224,7 ; loopne 4761 <.literal16+0x8d1> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 46d5 <.literal16+0x875> + DB 224,7 ; loopne 4765 <.literal16+0x8d5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -20035,13 +20141,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4739 <.literal16+0x8d9> + DB 224,7 ; loopne 47c9 <.literal16+0x939> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 473d <.literal16+0x8dd> + DB 224,7 ; loopne 47cd <.literal16+0x93d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 4741 <.literal16+0x8e1> + DB 224,7 ; loopne 47d1 <.literal16+0x941> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 4745 <.literal16+0x8e5> + DB 224,7 ; loopne 47d5 <.literal16+0x945> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -20079,13 +20185,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 47d6 <.literal16+0x976> + DB 124,66 ; jl 4866 <.literal16+0x9d6> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 47da <.literal16+0x97a> + DB 124,66 ; jl 486a <.literal16+0x9da> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 47de <.literal16+0x97e> + DB 124,66 ; jl 486e <.literal16+0x9de> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 47e2 <.literal16+0x982> + DB 124,66 ; jl 4872 <.literal16+0x9e2> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -20175,13 +20281,13 @@ 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 48e5 <.literal16+0xa85> + DB 112,65 ; jo 4975 <.literal16+0xae5> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 48e9 <.literal16+0xa89> + DB 112,65 ; jo 4979 <.literal16+0xae9> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 48ed <.literal16+0xa8d> + DB 112,65 ; jo 497d <.literal16+0xaed> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 48f1 <.literal16+0xa91> + DB 112,65 ; jo 4981 <.literal16+0xaf1> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -20203,11 +20309,11 @@ 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 493b <.literal16+0xadb> + DB 127,67 ; jg 49cb <.literal16+0xb3b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 493f <.literal16+0xadf> + DB 127,67 ; jg 49cf <.literal16+0xb3f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 4943 <.literal16+0xae3> + DB 127,67 ; jg 49d3 <.literal16+0xb43> 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) @@ -20283,13 +20389,13 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 4a2b <.literal16+0xbcb> + DB 127,71 ; jg 4abb <.literal16+0xc2b> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 4a2f <.literal16+0xbcf> + DB 127,71 ; jg 4abf <.literal16+0xc2f> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 4a33 <.literal16+0xbd3> + DB 127,71 ; jg 4ac3 <.literal16+0xc33> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 4a37 <.literal16+0xbd7> + DB 127,71 ; jg 4ac7 <.literal16+0xc37> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -20400,11 +20506,11 @@ 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 4b22 <.literal16+0xcc2> + DB 62,114,28 ; jb,pt 4bb2 <.literal16+0xd22> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4b26 <.literal16+0xcc6> + DB 62,114,28 ; jb,pt 4bb6 <.literal16+0xd26> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4b2a <.literal16+0xcca> + DB 62,114,28 ; jb,pt 4bba <.literal16+0xd2a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -20448,7 +20554,7 @@ 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 # 3d63d9b5 <_sk_callback_sse2+0x3d639c08> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63da45 <_sk_callback_sse2+0x3d639c6e> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -20474,7 +20580,7 @@ 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 # 3d63d9f5 <_sk_callback_sse2+0x3d639c48> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63da85 <_sk_callback_sse2+0x3d639cae> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -20483,13 +20589,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 4bee <.literal16+0xd8e> + DB 114,28 ; jb 4c7e <.literal16+0xdee> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4bf2 <.literal16+0xd92> + DB 62,114,28 ; jb,pt 4c82 <.literal16+0xdf2> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4bf6 <.literal16+0xd96> + DB 62,114,28 ; jb,pt 4c86 <.literal16+0xdf6> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4bfa <.literal16+0xd9a> + DB 62,114,28 ; jb,pt 4c8a <.literal16+0xdfa> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -20510,11 +20616,11 @@ 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 4c32 <.literal16+0xdd2> + DB 62,114,28 ; jb,pt 4cc2 <.literal16+0xe32> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4c36 <.literal16+0xdd6> + DB 62,114,28 ; jb,pt 4cc6 <.literal16+0xe36> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4c3a <.literal16+0xdda> + DB 62,114,28 ; jb,pt 4cca <.literal16+0xe3a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -20558,7 +20664,7 @@ 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 # 3d63dac5 <_sk_callback_sse2+0x3d639d18> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63db55 <_sk_callback_sse2+0x3d639d7e> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -20584,7 +20690,7 @@ 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 # 3d63db05 <_sk_callback_sse2+0x3d639d58> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63db95 <_sk_callback_sse2+0x3d639dbe> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -20593,13 +20699,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 4cfe <.literal16+0xe9e> + DB 114,28 ; jb 4d8e <.literal16+0xefe> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4d02 <_sk_callback_sse2+0xf55> + DB 62,114,28 ; jb,pt 4d92 <_sk_callback_sse2+0xfbb> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4d06 <_sk_callback_sse2+0xf59> + DB 62,114,28 ; jb,pt 4d96 <_sk_callback_sse2+0xfbf> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 4d0a <_sk_callback_sse2+0xf5d> + DB 62,114,28 ; jb,pt 4d9a <_sk_callback_sse2+0xfc3> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp index 8d45dd3196..09ebebf8d4 100644 --- a/src/jumper/SkJumper_stages.cpp +++ b/src/jumper/SkJumper_stages.cpp @@ -504,23 +504,23 @@ STAGE(hsl_to_rgb) { s = g, l = b; - F q = if_then_else(l < 0.5_f, l*(1.0_f + s), l + s - l*s), - p = 2.0_f*l - q; + F q = if_then_else(l < 0.5_f, l*(1.0f + s), l + s - l*s), + p = 2.0f*l - q; auto hue_to_rgb = [&](F t) { - F t2 = if_then_else(t < 0.0_f, t + 1.0_f, - if_then_else(t > 1.0_f, t - 1.0_f, + F t2 = if_then_else(t < 0.0_f, t + 1.0f, + if_then_else(t > 1.0_f, t - 1.0f, t)); - return if_then_else(t2 < C(1/6.0f), p + (q-p)*6.0_f*t, + return if_then_else(t2 < C(1/6.0f), p + (q-p)*6.0f*t, if_then_else(t2 < C(3/6.0f), q, - if_then_else(t2 < C(4/6.0f), p + (q-p)*6.0_f*(C(4/6.0f) - t2), + if_then_else(t2 < C(4/6.0f), p + (q-p)*6.0f*((4/6.0f) - t2), p))); }; - r = if_then_else(s == 0, l, hue_to_rgb(h + C(1/3.0f))); - g = if_then_else(s == 0, l, hue_to_rgb(h )); - b = if_then_else(s == 0, l, hue_to_rgb(h - C(1/3.0f))); + r = if_then_else(s == 0, l, hue_to_rgb(h + (1/3.0f))); + g = if_then_else(s == 0, l, hue_to_rgb(h )); + b = if_then_else(s == 0, l, hue_to_rgb(h - (1/3.0f))); } STAGE(scale_1_float) { |