diff options
| author |  Mike Klein <mtklein@chromium.org> | 2017-04-27 08:59:55 -0400 |
|---|---|---|
| committer |  Skia Commit-Bot <skia-commit-bot@chromium.org> | 2017-04-27 15:30:07 +0000 |
| commit | b4bbc64adee74f99ddfc30ad4e5043f98380bdbb (patch) | |
| tree | cf5ffc26961271454629a7cc078e9cf0b6ca31c2 | |
| parent | e51125557f28c2fc3dfbf4a8553a7c06bc1a0286 (diff) | |
clear out C(), _i, and _f constants from SkJumper_vectors.h
I think this is just gonna be a bunch of baby steps for now.
This gets everything in SkJumper_vectors.h.
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: Ic87faa9bf6380a4fc9a577936dad8c3a9c48472e
Reviewed-on: https://skia-review.googlesource.com/14441
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
| -rw-r--r-- | src/jumper/SkJumper_generated.S | 4158 | ||||
| -rw-r--r-- | src/jumper/SkJumper_generated_win.S | 4161 | ||||
| -rw-r--r-- | src/jumper/SkJumper_vectors.h | 34 |
3 files changed, 4719 insertions, 3634 deletions
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index ed23be731f..a3ea251a7c 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 196,226,125,24,13,117,65,0,0 // vbroadcastss 0x4175(%rip),%ymm1 # 4238 <_sk_callback_hsw+0x126> + .byte 196,226,125,24,13,61,64,0,0 // vbroadcastss 0x403d(%rip),%ymm1 # 4100 <_sk_callback_hsw+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 196,226,125,24,21,89,65,0,0 // vbroadcastss 0x4159(%rip),%ymm2 # 423c <_sk_callback_hsw+0x12a> + .byte 196,226,125,24,21,33,64,0,0 // vbroadcastss 0x4021(%rip),%ymm2 # 4104 <_sk_callback_hsw+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 @@ -9205,60 +9205,36 @@ _sk_parametric_r_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,125,168,211 // vfmadd213ps %ymm11,%ymm0,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 - .byte 196,65,125,219,210 // vpand %ymm10,%ymm0,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 - .byte 197,45,235,208 // vpor %ymm0,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,240 // vpbroadcastd %xmm0,%ymm14 - .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,216 // vpbroadcastd %xmm0,%ymm11 - .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 - .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0 - .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 - .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 - .byte 197,28,89,216 // vmulps %ymm0,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .byte 196,98,125,24,37,243,36,0,0 // vbroadcastss 0x24f3(%rip),%ymm12 # 4108 <_sk_callback_hsw+0x12e> + .byte 196,98,125,24,45,238,36,0,0 // vbroadcastss 0x24ee(%rip),%ymm13 # 410c <_sk_callback_hsw+0x132> + .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,228,36,0,0 // vbroadcastss 0x24e4(%rip),%ymm13 # 4110 <_sk_callback_hsw+0x136> + .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,218,36,0,0 // vbroadcastss 0x24da(%rip),%ymm13 # 4114 <_sk_callback_hsw+0x13a> + .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 + .byte 196,98,125,24,29,208,36,0,0 // vbroadcastss 0x24d0(%rip),%ymm11 # 4118 <_sk_callback_hsw+0x13e> + .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 + .byte 196,98,125,24,37,198,36,0,0 // vbroadcastss 0x24c6(%rip),%ymm12 # 411c <_sk_callback_hsw+0x142> + .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 + .byte 196,98,125,24,37,188,36,0,0 // vbroadcastss 0x24bc(%rip),%ymm12 # 4120 <_sk_callback_hsw+0x146> + .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,224 // vpbroadcastd %xmm0,%ymm12 - .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 - .byte 196,65,124,88,219 // vaddps %ymm11,%ymm0,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 - .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,216 // vpbroadcastd %xmm0,%ymm11 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 + .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 196,98,125,24,29,157,36,0,0 // vbroadcastss 0x249d(%rip),%ymm11 # 4124 <_sk_callback_hsw+0x14a> + .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 + .byte 196,98,125,24,29,147,36,0,0 // vbroadcastss 0x2493(%rip),%ymm11 # 4128 <_sk_callback_hsw+0x14e> + .byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11 + .byte 196,226,125,24,5,137,36,0,0 // vbroadcastss 0x2489(%rip),%ymm0 # 412c <_sk_callback_hsw+0x152> .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 - .byte 197,164,94,192 // vdivps %ymm0,%ymm11,%ymm0 - .byte 197,148,88,192 // vaddps %ymm0,%ymm13,%ymm0 - .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 + .byte 196,98,125,24,21,127,36,0,0 // vbroadcastss 0x247f(%rip),%ymm10 # 4130 <_sk_callback_hsw+0x156> + .byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0 + .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 + .byte 196,98,125,24,21,114,36,0,0 // vbroadcastss 0x2472(%rip),%ymm10 # 4134 <_sk_callback_hsw+0x15a> + .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 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 @@ -9285,60 +9261,36 @@ _sk_parametric_g_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,117,168,211 // vfmadd213ps %ymm11,%ymm1,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 - .byte 196,65,117,219,210 // vpand %ymm10,%ymm1,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 - .byte 197,45,235,209 // vpor %ymm1,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,241 // vpbroadcastd %xmm1,%ymm14 - .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,217 // vpbroadcastd %xmm1,%ymm11 - .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 - .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 - .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 - .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 - .byte 197,28,89,217 // vmulps %ymm1,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .byte 196,98,125,24,37,249,35,0,0 // vbroadcastss 0x23f9(%rip),%ymm12 # 4138 <_sk_callback_hsw+0x15e> + .byte 196,98,125,24,45,244,35,0,0 // vbroadcastss 0x23f4(%rip),%ymm13 # 413c <_sk_callback_hsw+0x162> + .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,234,35,0,0 // vbroadcastss 0x23ea(%rip),%ymm13 # 4140 <_sk_callback_hsw+0x166> + .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,224,35,0,0 // vbroadcastss 0x23e0(%rip),%ymm13 # 4144 <_sk_callback_hsw+0x16a> + .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 + .byte 196,98,125,24,29,214,35,0,0 // vbroadcastss 0x23d6(%rip),%ymm11 # 4148 <_sk_callback_hsw+0x16e> + .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 + .byte 196,98,125,24,37,204,35,0,0 // vbroadcastss 0x23cc(%rip),%ymm12 # 414c <_sk_callback_hsw+0x172> + .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 + .byte 196,98,125,24,37,194,35,0,0 // vbroadcastss 0x23c2(%rip),%ymm12 # 4150 <_sk_callback_hsw+0x176> + .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12 - .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 - .byte 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 - .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,217 // vpbroadcastd %xmm1,%ymm11 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 + .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 196,98,125,24,29,163,35,0,0 // vbroadcastss 0x23a3(%rip),%ymm11 # 4154 <_sk_callback_hsw+0x17a> + .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 + .byte 196,98,125,24,29,153,35,0,0 // vbroadcastss 0x2399(%rip),%ymm11 # 4158 <_sk_callback_hsw+0x17e> + .byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11 + .byte 196,226,125,24,13,143,35,0,0 // vbroadcastss 0x238f(%rip),%ymm1 # 415c <_sk_callback_hsw+0x182> .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 - .byte 197,164,94,201 // vdivps %ymm1,%ymm11,%ymm1 - .byte 197,148,88,201 // vaddps %ymm1,%ymm13,%ymm1 - .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 + .byte 196,98,125,24,21,133,35,0,0 // vbroadcastss 0x2385(%rip),%ymm10 # 4160 <_sk_callback_hsw+0x186> + .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 + .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 + .byte 196,98,125,24,21,120,35,0,0 // vbroadcastss 0x2378(%rip),%ymm10 # 4164 <_sk_callback_hsw+0x18a> + .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 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 @@ -9365,60 +9317,36 @@ _sk_parametric_b_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,109,168,211 // vfmadd213ps %ymm11,%ymm2,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 - .byte 196,65,109,219,210 // vpand %ymm10,%ymm2,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 - .byte 197,45,235,210 // vpor %ymm2,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,242 // vpbroadcastd %xmm2,%ymm14 - .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,218 // vpbroadcastd %xmm2,%ymm11 - .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 - .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 - .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 - .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 - .byte 197,28,89,218 // vmulps %ymm2,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .byte 196,98,125,24,37,255,34,0,0 // vbroadcastss 0x22ff(%rip),%ymm12 # 4168 <_sk_callback_hsw+0x18e> + .byte 196,98,125,24,45,250,34,0,0 // vbroadcastss 0x22fa(%rip),%ymm13 # 416c <_sk_callback_hsw+0x192> + .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,240,34,0,0 // vbroadcastss 0x22f0(%rip),%ymm13 # 4170 <_sk_callback_hsw+0x196> + .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,230,34,0,0 // vbroadcastss 0x22e6(%rip),%ymm13 # 4174 <_sk_callback_hsw+0x19a> + .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 + .byte 196,98,125,24,29,220,34,0,0 // vbroadcastss 0x22dc(%rip),%ymm11 # 4178 <_sk_callback_hsw+0x19e> + .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 + .byte 196,98,125,24,37,210,34,0,0 // vbroadcastss 0x22d2(%rip),%ymm12 # 417c <_sk_callback_hsw+0x1a2> + .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 + .byte 196,98,125,24,37,200,34,0,0 // vbroadcastss 0x22c8(%rip),%ymm12 # 4180 <_sk_callback_hsw+0x1a6> + .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,226 // vpbroadcastd %xmm2,%ymm12 - .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 - .byte 196,65,108,88,219 // vaddps %ymm11,%ymm2,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 - .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,218 // vpbroadcastd %xmm2,%ymm11 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 + .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 196,98,125,24,29,169,34,0,0 // vbroadcastss 0x22a9(%rip),%ymm11 # 4184 <_sk_callback_hsw+0x1aa> + .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 + .byte 196,98,125,24,29,159,34,0,0 // vbroadcastss 0x229f(%rip),%ymm11 # 4188 <_sk_callback_hsw+0x1ae> + .byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11 + .byte 196,226,125,24,21,149,34,0,0 // vbroadcastss 0x2295(%rip),%ymm2 # 418c <_sk_callback_hsw+0x1b2> .byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2 - .byte 197,164,94,210 // vdivps %ymm2,%ymm11,%ymm2 - .byte 197,148,88,210 // vaddps %ymm2,%ymm13,%ymm2 - .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 + .byte 196,98,125,24,21,139,34,0,0 // vbroadcastss 0x228b(%rip),%ymm10 # 4190 <_sk_callback_hsw+0x1b6> + .byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2 + .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 + .byte 196,98,125,24,21,126,34,0,0 // vbroadcastss 0x227e(%rip),%ymm10 # 4194 <_sk_callback_hsw+0x1ba> + .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 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 @@ -9445,60 +9373,36 @@ _sk_parametric_a_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,101,168,211 // vfmadd213ps %ymm11,%ymm3,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,226,125,24,24 // vbroadcastss (%rax),%ymm3 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 196,65,101,219,210 // vpand %ymm10,%ymm3,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 197,45,235,211 // vpor %ymm3,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,243 // vpbroadcastd %xmm3,%ymm14 - .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 - .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 197,172,88,219 // vaddps %ymm3,%ymm10,%ymm3 - .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 - .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 - .byte 197,28,89,219 // vmulps %ymm3,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .byte 196,98,125,24,37,5,34,0,0 // vbroadcastss 0x2205(%rip),%ymm12 # 4198 <_sk_callback_hsw+0x1be> + .byte 196,98,125,24,45,0,34,0,0 // vbroadcastss 0x2200(%rip),%ymm13 # 419c <_sk_callback_hsw+0x1c2> + .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,246,33,0,0 // vbroadcastss 0x21f6(%rip),%ymm13 # 41a0 <_sk_callback_hsw+0x1c6> + .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 + .byte 196,98,125,24,45,236,33,0,0 // vbroadcastss 0x21ec(%rip),%ymm13 # 41a4 <_sk_callback_hsw+0x1ca> + .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 + .byte 196,98,125,24,29,226,33,0,0 // vbroadcastss 0x21e2(%rip),%ymm11 # 41a8 <_sk_callback_hsw+0x1ce> + .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 + .byte 196,98,125,24,37,216,33,0,0 // vbroadcastss 0x21d8(%rip),%ymm12 # 41ac <_sk_callback_hsw+0x1d2> + .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 + .byte 196,98,125,24,37,206,33,0,0 // vbroadcastss 0x21ce(%rip),%ymm12 # 41b0 <_sk_callback_hsw+0x1d6> + .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,227 // vpbroadcastd %xmm3,%ymm12 - .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 - .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 + .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 196,98,125,24,29,175,33,0,0 // vbroadcastss 0x21af(%rip),%ymm11 # 41b4 <_sk_callback_hsw+0x1da> + .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 + .byte 196,98,125,24,29,165,33,0,0 // vbroadcastss 0x21a5(%rip),%ymm11 # 41b8 <_sk_callback_hsw+0x1de> + .byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11 + .byte 196,226,125,24,29,155,33,0,0 // vbroadcastss 0x219b(%rip),%ymm3 # 41bc <_sk_callback_hsw+0x1e2> .byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3 - .byte 197,164,94,219 // vdivps %ymm3,%ymm11,%ymm3 - .byte 197,148,88,219 // vaddps %ymm3,%ymm13,%ymm3 - .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 + .byte 196,98,125,24,21,145,33,0,0 // vbroadcastss 0x2191(%rip),%ymm10 # 41c0 <_sk_callback_hsw+0x1e6> + .byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3 + .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 + .byte 196,98,125,24,21,132,33,0,0 // vbroadcastss 0x2184(%rip),%ymm10 # 41c4 <_sk_callback_hsw+0x1ea> + .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 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 @@ -9590,7 +9494,7 @@ _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 117,50 // jne 2321 <_sk_load_a8_hsw+0x42> + .byte 117,50 // jne 21e9 <_sk_load_a8_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 @@ -9613,9 +9517,9 @@ _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 2329 <_sk_load_a8_hsw+0x4a> + .byte 117,234 // jne 21f1 <_sk_load_a8_hsw+0x4a> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,173 // jmp 22f3 <_sk_load_a8_hsw+0x14> + .byte 235,173 // jmp 21bb <_sk_load_a8_hsw+0x14> HIDDEN _sk_gather_a8_hsw .globl _sk_gather_a8_hsw @@ -9690,7 +9594,7 @@ _sk_store_a8_hsw: .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 245e <_sk_store_a8_hsw+0x3b> + .byte 117,10 // jne 2326 <_sk_store_a8_hsw+0x3b> .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -9698,10 +9602,10 @@ _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 245a <_sk_store_a8_hsw+0x37> + .byte 119,236 // ja 2322 <_sk_store_a8_hsw+0x37> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 24c0 <_sk_store_a8_hsw+0x9d> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2388 <_sk_store_a8_hsw+0x9d> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -9712,7 +9616,7 @@ _sk_store_a8_hsw: .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) - .byte 235,154 // jmp 245a <_sk_store_a8_hsw+0x37> + .byte 235,154 // jmp 2322 <_sk_store_a8_hsw+0x37> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -9745,7 +9649,7 @@ _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 117,60 // jne 2528 <_sk_load_g8_hsw+0x4c> + .byte 117,60 // jne 23f0 <_sk_load_g8_hsw+0x4c> .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 @@ -9770,9 +9674,9 @@ _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 2530 <_sk_load_g8_hsw+0x54> + .byte 117,234 // jne 23f8 <_sk_load_g8_hsw+0x54> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,163 // jmp 24f0 <_sk_load_g8_hsw+0x14> + .byte 235,163 // jmp 23b8 <_sk_load_g8_hsw+0x14> HIDDEN _sk_gather_g8_hsw .globl _sk_gather_g8_hsw @@ -9841,9 +9745,9 @@ _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 2643 <_sk_gather_i8_hsw+0xf> + .byte 116,5 // je 250b <_sk_gather_i8_hsw+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 2645 <_sk_gather_i8_hsw+0x11> + .byte 235,2 // jmp 250d <_sk_gather_i8_hsw+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -9916,7 +9820,7 @@ _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 15,133,149,0,0,0 // jne 27f7 <_sk_load_565_hsw+0xa3> + .byte 15,133,149,0,0,0 // jne 26bf <_sk_load_565_hsw+0xa3> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 .byte 184,0,248,0,0 // mov $0xf800,%eax @@ -9956,9 +9860,9 @@ _sk_load_565_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 15,135,89,255,255,255 // ja 2768 <_sk_load_565_hsw+0x14> + .byte 15,135,89,255,255,255 // ja 2630 <_sk_load_565_hsw+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 2864 <_sk_load_565_hsw+0x110> + .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 272c <_sk_load_565_hsw+0x110> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -9970,12 +9874,12 @@ _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 233,5,255,255,255 // jmpq 2768 <_sk_load_565_hsw+0x14> + .byte 233,5,255,255,255 // jmpq 2630 <_sk_load_565_hsw+0x14> .byte 144 // nop .byte 243,255 // repz (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 235,255 // jmp 2869 <_sk_load_565_hsw+0x115> + .byte 235,255 // jmp 2731 <_sk_load_565_hsw+0x115> .byte 255 // (bad) .byte 255,227 // jmpq *%rbx .byte 255 // (bad) @@ -10102,7 +10006,7 @@ _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 2a2f <_sk_store_565_hsw+0x6c> + .byte 117,10 // jne 28f7 <_sk_store_565_hsw+0x6c> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -10110,9 +10014,9 @@ _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 2a2b <_sk_store_565_hsw+0x68> + .byte 119,236 // ja 28f3 <_sk_store_565_hsw+0x68> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2a8c <_sk_store_565_hsw+0xc9> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2954 <_sk_store_565_hsw+0xc9> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -10123,7 +10027,7 @@ _sk_store_565_hsw: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 2a2b <_sk_store_565_hsw+0x68> + .byte 235,159 // jmp 28f3 <_sk_store_565_hsw+0x68> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -10154,7 +10058,7 @@ _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 15,133,179,0,0,0 // jne 2b69 <_sk_load_4444_hsw+0xc1> + .byte 15,133,179,0,0,0 // jne 2a31 <_sk_load_4444_hsw+0xc1> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 196,98,125,51,200 // vpmovzxwd %xmm0,%ymm9 .byte 184,0,240,0,0 // mov $0xf000,%eax @@ -10200,9 +10104,9 @@ _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 15,135,59,255,255,255 // ja 2abc <_sk_load_4444_hsw+0x14> + .byte 15,135,59,255,255,255 // ja 2984 <_sk_load_4444_hsw+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 2bd8 <_sk_load_4444_hsw+0x130> + .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 2aa0 <_sk_load_4444_hsw+0x130> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -10214,13 +10118,13 @@ _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 233,231,254,255,255 // jmpq 2abc <_sk_load_4444_hsw+0x14> + .byte 233,231,254,255,255 // jmpq 2984 <_sk_load_4444_hsw+0x14> .byte 15,31,0 // nopl (%rax) .byte 241 // icebp .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,225 // jmpq ffffffffe2002be0 <_sk_callback_hsw+0xffffffffe1ffeace> + .byte 233,255,255,255,225 // jmpq ffffffffe2002aa8 <_sk_callback_hsw+0xffffffffe1ffeace> .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) @@ -10352,7 +10256,7 @@ _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 2dc7 <_sk_store_4444_hsw+0x72> + .byte 117,10 // jne 2c8f <_sk_store_4444_hsw+0x72> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -10360,9 +10264,9 @@ _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 2dc3 <_sk_store_4444_hsw+0x6e> + .byte 119,236 // ja 2c8b <_sk_store_4444_hsw+0x6e> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2e24 <_sk_store_4444_hsw+0xcf> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2cec <_sk_store_4444_hsw+0xcf> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -10373,7 +10277,7 @@ _sk_store_4444_hsw: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 2dc3 <_sk_store_4444_hsw+0x6e> + .byte 235,159 // jmp 2c8b <_sk_store_4444_hsw+0x6e> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -10406,7 +10310,7 @@ _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 117,104 // jne 2ebd <_sk_load_8888_hsw+0x7d> + .byte 117,104 // jne 2d85 <_sk_load_8888_hsw+0x7d> .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 .byte 184,255,0,0,0 // mov $0xff,%eax .byte 197,249,110,192 // vmovd %eax,%xmm0 @@ -10439,7 +10343,7 @@ _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 233,116,255,255,255 // jmpq 2e5a <_sk_load_8888_hsw+0x1a> + .byte 233,116,255,255,255 // jmpq 2d22 <_sk_load_8888_hsw+0x1a> HIDDEN _sk_gather_8888_hsw .globl _sk_gather_8888_hsw @@ -10503,7 +10407,7 @@ _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 2fe0 <_sk_store_8888_hsw+0x74> + .byte 117,12 // jne 2ea8 <_sk_store_8888_hsw+0x74> .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,193 // mov %r8,%rcx @@ -10516,7 +10420,7 @@ _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 2fd9 <_sk_store_8888_hsw+0x6d> + .byte 235,211 // jmp 2ea1 <_sk_store_8888_hsw+0x6d> HIDDEN _sk_load_f16_hsw .globl _sk_load_f16_hsw @@ -10525,7 +10429,7 @@ _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 3071 <_sk_load_f16_hsw+0x6b> + .byte 117,97 // jne 2f39 <_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 @@ -10551,29 +10455,29 @@ _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 30d0 <_sk_load_f16_hsw+0xca> + .byte 116,79 // je 2f98 <_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 30d0 <_sk_load_f16_hsw+0xca> + .byte 114,67 // jb 2f98 <_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 30dd <_sk_load_f16_hsw+0xd7> + .byte 116,68 // je 2fa5 <_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 30dd <_sk_load_f16_hsw+0xd7> + .byte 114,56 // jb 2fa5 <_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 3027 <_sk_load_f16_hsw+0x21> + .byte 15,132,114,255,255,255 // je 2eef <_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 3027 <_sk_load_f16_hsw+0x21> + .byte 15,130,98,255,255,255 // jb 2eef <_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 3027 <_sk_load_f16_hsw+0x21> + .byte 233,87,255,255,255 // jmpq 2eef <_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 3027 <_sk_load_f16_hsw+0x21> + .byte 233,74,255,255,255 // jmpq 2eef <_sk_load_f16_hsw+0x21> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,65,255,255,255 // jmpq 3027 <_sk_load_f16_hsw+0x21> + .byte 233,65,255,255,255 // jmpq 2eef <_sk_load_f16_hsw+0x21> HIDDEN _sk_gather_f16_hsw .globl _sk_gather_f16_hsw @@ -10631,7 +10535,7 @@ _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 31d5 <_sk_store_f16_hsw+0x65> + .byte 117,27 // jne 309d <_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) @@ -10640,22 +10544,22 @@ _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 31d1 <_sk_store_f16_hsw+0x61> + .byte 116,241 // je 3099 <_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 31d1 <_sk_store_f16_hsw+0x61> + .byte 114,229 // jb 3099 <_sk_store_f16_hsw+0x61> .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) - .byte 116,221 // je 31d1 <_sk_store_f16_hsw+0x61> + .byte 116,221 // je 3099 <_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 31d1 <_sk_store_f16_hsw+0x61> + .byte 114,209 // jb 3099 <_sk_store_f16_hsw+0x61> .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) - .byte 116,201 // je 31d1 <_sk_store_f16_hsw+0x61> + .byte 116,201 // je 3099 <_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 31d1 <_sk_store_f16_hsw+0x61> + .byte 114,189 // jb 3099 <_sk_store_f16_hsw+0x61> .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) - .byte 235,181 // jmp 31d1 <_sk_store_f16_hsw+0x61> + .byte 235,181 // jmp 3099 <_sk_store_f16_hsw+0x61> HIDDEN _sk_load_u16_be_hsw .globl _sk_load_u16_be_hsw @@ -10665,7 +10569,7 @@ _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 15,133,205,0,0,0 // jne 32ff <_sk_load_u16_be_hsw+0xe3> + .byte 15,133,205,0,0,0 // jne 31c7 <_sk_load_u16_be_hsw+0xe3> .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 @@ -10714,29 +10618,29 @@ _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 3365 <_sk_load_u16_be_hsw+0x149> + .byte 116,85 // je 322d <_sk_load_u16_be_hsw+0x149> .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 3365 <_sk_load_u16_be_hsw+0x149> + .byte 114,72 // jb 322d <_sk_load_u16_be_hsw+0x149> .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 3372 <_sk_load_u16_be_hsw+0x156> + .byte 116,72 // je 323a <_sk_load_u16_be_hsw+0x156> .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 3372 <_sk_load_u16_be_hsw+0x156> + .byte 114,59 // jb 323a <_sk_load_u16_be_hsw+0x156> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,5,255,255,255 // je 324d <_sk_load_u16_be_hsw+0x31> + .byte 15,132,5,255,255,255 // je 3115 <_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 15,130,244,254,255,255 // jb 324d <_sk_load_u16_be_hsw+0x31> + .byte 15,130,244,254,255,255 // jb 3115 <_sk_load_u16_be_hsw+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,232,254,255,255 // jmpq 324d <_sk_load_u16_be_hsw+0x31> + .byte 233,232,254,255,255 // jmpq 3115 <_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 233,219,254,255,255 // jmpq 324d <_sk_load_u16_be_hsw+0x31> + .byte 233,219,254,255,255 // jmpq 3115 <_sk_load_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,210,254,255,255 // jmpq 324d <_sk_load_u16_be_hsw+0x31> + .byte 233,210,254,255,255 // jmpq 3115 <_sk_load_u16_be_hsw+0x31> HIDDEN _sk_load_rgb_u16_be_hsw .globl _sk_load_rgb_u16_be_hsw @@ -10746,7 +10650,7 @@ _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 15,133,211,0,0,0 // jne 3460 <_sk_load_rgb_u16_be_hsw+0xe5> + .byte 15,133,211,0,0,0 // jne 3328 <_sk_load_rgb_u16_be_hsw+0xe5> .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 @@ -10796,36 +10700,36 @@ _sk_load_rgb_u16_be_hsw: .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 3479 <_sk_load_rgb_u16_be_hsw+0xfe> - .byte 233,72,255,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,5 // jne 3341 <_sk_load_rgb_u16_be_hsw+0xfe> + .byte 233,72,255,255,255 // jmpq 3289 <_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 34a8 <_sk_load_rgb_u16_be_hsw+0x12d> + .byte 114,26 // jb 3370 <_sk_load_rgb_u16_be_hsw+0x12d> .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 34ad <_sk_load_rgb_u16_be_hsw+0x132> - .byte 233,25,255,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,20,255,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 3375 <_sk_load_rgb_u16_be_hsw+0x132> + .byte 233,25,255,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,20,255,255,255 // jmpq 3289 <_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 34dc <_sk_load_rgb_u16_be_hsw+0x161> + .byte 114,26 // jb 33a4 <_sk_load_rgb_u16_be_hsw+0x161> .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 34e1 <_sk_load_rgb_u16_be_hsw+0x166> - .byte 233,229,254,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,224,254,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 33a9 <_sk_load_rgb_u16_be_hsw+0x166> + .byte 233,229,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,224,254,255,255 // jmpq 3289 <_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 350a <_sk_load_rgb_u16_be_hsw+0x18f> + .byte 114,20 // jb 33d2 <_sk_load_rgb_u16_be_hsw+0x18f> .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 233,183,254,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,178,254,255,255 // jmpq 33c1 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,183,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,178,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46> HIDDEN _sk_store_u16_be_hsw .globl _sk_store_u16_be_hsw @@ -10874,7 +10778,7 @@ _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 360a <_sk_store_u16_be_hsw+0xfb> + .byte 117,31 // jne 34d2 <_sk_store_u16_be_hsw+0xfb> .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2) .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2) .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2) @@ -10883,22 +10787,22 @@ _sk_store_u16_be_hsw: .byte 255,224 // jmpq *%rax .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 116,240 // je 34ce <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 114,227 // jb 34ce <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2) - .byte 116,218 // je 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 116,218 // je 34ce <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 114,205 // jb 34ce <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2) - .byte 116,196 // je 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 116,196 // je 34ce <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 114,183 // jb 34ce <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2) - .byte 235,174 // jmp 3606 <_sk_store_u16_be_hsw+0xf7> + .byte 235,174 // jmp 34ce <_sk_store_u16_be_hsw+0xf7> HIDDEN _sk_load_f32_hsw .globl _sk_load_f32_hsw @@ -10906,10 +10810,10 @@ 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 36ce <_sk_load_f32_hsw+0x76> + .byte 119,110 // ja 3596 <_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 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 36f8 <_sk_load_f32_hsw+0xa0> + .byte 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 35c0 <_sk_load_f32_hsw+0xa0> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -10968,7 +10872,7 @@ _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 3785 <_sk_store_f32_hsw+0x6d> + .byte 117,55 // jne 364d <_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 @@ -10981,22 +10885,22 @@ _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 3781 <_sk_store_f32_hsw+0x69> + .byte 116,240 // je 3649 <_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 3781 <_sk_store_f32_hsw+0x69> + .byte 114,227 // jb 3649 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 3781 <_sk_store_f32_hsw+0x69> + .byte 116,218 // je 3649 <_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 3781 <_sk_store_f32_hsw+0x69> + .byte 114,205 // jb 3649 <_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 3781 <_sk_store_f32_hsw+0x69> + .byte 116,195 // je 3649 <_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 3781 <_sk_store_f32_hsw+0x69> + .byte 114,181 // jb 3649 <_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 3781 <_sk_store_f32_hsw+0x69> + .byte 235,171 // jmp 3649 <_sk_store_f32_hsw+0x69> HIDDEN _sk_clamp_x_hsw .globl _sk_clamp_x_hsw @@ -11261,7 +11165,7 @@ _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 3c11 <_sk_linear_gradient_hsw+0xb5> + .byte 15,132,143,0,0,0 // je 3ad9 <_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 @@ -11288,8 +11192,8 @@ _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 3b9b <_sk_linear_gradient_hsw+0x3f> - .byte 235,17 // jmp 3c22 <_sk_linear_gradient_hsw+0xc6> + .byte 117,140 // jne 3a63 <_sk_linear_gradient_hsw+0x3f> + .byte 235,17 // jmp 3aea <_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 @@ -11715,8 +11619,76 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) - .byte 128 // .byte 0x80 - .byte 63 // (bad) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4110 <.literal4+0x10> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4189 <.literal4+0x89> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4140 <.literal4+0x40> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 41b9 <.literal4+0xb9> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4170 <.literal4+0x70> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 41e9 <_sk_callback_hsw+0x20f> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 41a0 <.literal4+0xa0> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4219 <_sk_callback_hsw+0x23f> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0 // .byte 0x0 + .byte 75 // rex.WXB BALIGN32 HIDDEN _sk_start_pipeline_avx @@ -11795,14 +11767,14 @@ _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 196,226,125,24,13,63,93,0,0 // vbroadcastss 0x5d3f(%rip),%ymm1 # 5e08 <_sk_callback_avx+0x126> + .byte 196,226,125,24,13,135,91,0,0 // vbroadcastss 0x5b87(%rip),%ymm1 # 5c50 <_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 196,226,125,24,21,35,93,0,0 // vbroadcastss 0x5d23(%rip),%ymm2 # 5e0c <_sk_callback_avx+0x12a> + .byte 196,226,125,24,21,107,91,0,0 // vbroadcastss 0x5b6b(%rip),%ymm2 # 5c54 <_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 @@ -14263,76 +14235,40 @@ _sk_parametric_r_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 - .byte 196,65,124,88,211 // vaddps %ymm11,%ymm0,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 - .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .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 197,36,89,216 // vmulps %ymm0,%ymm11,%ymm11 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 196,65,124,84,210 // vandps %ymm10,%ymm0,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,44,86,208 // vorps %ymm0,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .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 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .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 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 - .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%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 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .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 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0 - .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 + .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 196,98,125,24,37,198,52,0,0 // vbroadcastss 0x34c6(%rip),%ymm12 # 5c58 <_sk_callback_avx+0x12e> + .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,188,52,0,0 // vbroadcastss 0x34bc(%rip),%ymm12 # 5c5c <_sk_callback_avx+0x132> + .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0 + .byte 196,98,125,24,37,178,52,0,0 // vbroadcastss 0x34b2(%rip),%ymm12 # 5c60 <_sk_callback_avx+0x136> + .byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0 + .byte 196,98,125,24,37,168,52,0,0 // vbroadcastss 0x34a8(%rip),%ymm12 # 5c64 <_sk_callback_avx+0x13a> + .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,158,52,0,0 // vbroadcastss 0x349e(%rip),%ymm12 # 5c68 <_sk_callback_avx+0x13e> + .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 + .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,143,52,0,0 // vbroadcastss 0x348f(%rip),%ymm12 # 5c6c <_sk_callback_avx+0x142> + .byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0 + .byte 196,98,125,24,37,133,52,0,0 // vbroadcastss 0x3485(%rip),%ymm12 # 5c70 <_sk_callback_avx+0x146> + .byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0 .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 - .byte 197,28,89,216 // vmulps %ymm0,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .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 196,98,125,24,29,105,52,0,0 // vbroadcastss 0x3469(%rip),%ymm11 # 5c74 <_sk_callback_avx+0x14a> + .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 + .byte 196,98,125,24,29,95,52,0,0 // vbroadcastss 0x345f(%rip),%ymm11 # 5c78 <_sk_callback_avx+0x14e> + .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 + .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0 + .byte 196,98,125,24,29,80,52,0,0 // vbroadcastss 0x3450(%rip),%ymm11 # 5c7c <_sk_callback_avx+0x152> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%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 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .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 196,65,124,88,219 // vaddps %ymm11,%ymm0,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .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 196,98,125,24,29,70,52,0,0 // vbroadcastss 0x3446(%rip),%ymm11 # 5c80 <_sk_callback_avx+0x156> + .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 + .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 + .byte 196,98,125,24,21,55,52,0,0 // vbroadcastss 0x3437(%rip),%ymm10 # 5c84 <_sk_callback_avx+0x15a> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 - .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%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 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .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 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 - .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 - .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 - .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 @@ -14361,76 +14297,40 @@ _sk_parametric_g_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 - .byte 196,65,116,88,211 // vaddps %ymm11,%ymm1,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 - .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,36,89,217 // vmulps %ymm1,%ymm11,%ymm11 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 196,65,116,84,210 // vandps %ymm10,%ymm1,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,44,86,209 // vorps %ymm1,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 - .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,99,117,24,233,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 - .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 + .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 196,98,125,24,37,176,51,0,0 // vbroadcastss 0x33b0(%rip),%ymm12 # 5c88 <_sk_callback_avx+0x15e> + .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,166,51,0,0 // vbroadcastss 0x33a6(%rip),%ymm12 # 5c8c <_sk_callback_avx+0x162> + .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1 + .byte 196,98,125,24,37,156,51,0,0 // vbroadcastss 0x339c(%rip),%ymm12 # 5c90 <_sk_callback_avx+0x166> + .byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1 + .byte 196,98,125,24,37,146,51,0,0 // vbroadcastss 0x3392(%rip),%ymm12 # 5c94 <_sk_callback_avx+0x16a> + .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,136,51,0,0 // vbroadcastss 0x3388(%rip),%ymm12 # 5c98 <_sk_callback_avx+0x16e> + .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12 + .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,121,51,0,0 // vbroadcastss 0x3379(%rip),%ymm12 # 5c9c <_sk_callback_avx+0x172> + .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 + .byte 196,98,125,24,37,111,51,0,0 // vbroadcastss 0x336f(%rip),%ymm12 # 5ca0 <_sk_callback_avx+0x176> + .byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1 .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 - .byte 197,28,89,217 // vmulps %ymm1,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .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 196,98,125,24,29,83,51,0,0 // vbroadcastss 0x3353(%rip),%ymm11 # 5ca4 <_sk_callback_avx+0x17a> + .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 + .byte 196,98,125,24,29,73,51,0,0 // vbroadcastss 0x3349(%rip),%ymm11 # 5ca8 <_sk_callback_avx+0x17e> + .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 + .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1 + .byte 196,98,125,24,29,58,51,0,0 // vbroadcastss 0x333a(%rip),%ymm11 # 5cac <_sk_callback_avx+0x182> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 - .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 + .byte 196,98,125,24,29,48,51,0,0 // vbroadcastss 0x3330(%rip),%ymm11 # 5cb0 <_sk_callback_avx+0x186> + .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 + .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 + .byte 196,98,125,24,21,33,51,0,0 // vbroadcastss 0x3321(%rip),%ymm10 # 5cb4 <_sk_callback_avx+0x18a> .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 - .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,99,117,24,233,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 - .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 - .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 - .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 @@ -14459,76 +14359,40 @@ _sk_parametric_b_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 - .byte 196,65,108,88,211 // vaddps %ymm11,%ymm2,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 - .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .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 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 - .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 196,65,108,84,210 // vandps %ymm10,%ymm2,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 - .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,44,86,210 // vorps %ymm2,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .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 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .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 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 - .byte 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,99,109,24,234,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .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 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 - .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 + .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 196,98,125,24,37,154,50,0,0 // vbroadcastss 0x329a(%rip),%ymm12 # 5cb8 <_sk_callback_avx+0x18e> + .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,144,50,0,0 // vbroadcastss 0x3290(%rip),%ymm12 # 5cbc <_sk_callback_avx+0x192> + .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2 + .byte 196,98,125,24,37,134,50,0,0 // vbroadcastss 0x3286(%rip),%ymm12 # 5cc0 <_sk_callback_avx+0x196> + .byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2 + .byte 196,98,125,24,37,124,50,0,0 // vbroadcastss 0x327c(%rip),%ymm12 # 5cc4 <_sk_callback_avx+0x19a> + .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,114,50,0,0 // vbroadcastss 0x3272(%rip),%ymm12 # 5cc8 <_sk_callback_avx+0x19e> + .byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12 + .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,99,50,0,0 // vbroadcastss 0x3263(%rip),%ymm12 # 5ccc <_sk_callback_avx+0x1a2> + .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 + .byte 196,98,125,24,37,89,50,0,0 // vbroadcastss 0x3259(%rip),%ymm12 # 5cd0 <_sk_callback_avx+0x1a6> + .byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2 .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 - .byte 197,28,89,218 // vmulps %ymm2,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .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 196,98,125,24,29,61,50,0,0 // vbroadcastss 0x323d(%rip),%ymm11 # 5cd4 <_sk_callback_avx+0x1aa> + .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 + .byte 196,98,125,24,29,51,50,0,0 // vbroadcastss 0x3233(%rip),%ymm11 # 5cd8 <_sk_callback_avx+0x1ae> + .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 + .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2 + .byte 196,98,125,24,29,36,50,0,0 // vbroadcastss 0x3224(%rip),%ymm11 # 5cdc <_sk_callback_avx+0x1b2> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,99,109,24,226,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm12 - .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .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 196,65,108,88,219 // vaddps %ymm11,%ymm2,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .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 196,98,125,24,29,26,50,0,0 // vbroadcastss 0x321a(%rip),%ymm11 # 5ce0 <_sk_callback_avx+0x1b6> + .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 + .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 + .byte 196,98,125,24,21,11,50,0,0 // vbroadcastss 0x320b(%rip),%ymm10 # 5ce4 <_sk_callback_avx+0x1ba> .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 - .byte 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,99,109,24,234,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .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 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2 - .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 - .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 - .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 @@ -14557,76 +14421,40 @@ _sk_parametric_a_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 - .byte 196,65,100,88,211 // vaddps %ymm11,%ymm3,%ymm10 - .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 - .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 - .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d - .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 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 - .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 - .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 196,65,100,84,210 // vandps %ymm10,%ymm3,%ymm10 - .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 - .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,44,86,211 // vorps %ymm3,%ymm10,%ymm10 - .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d - .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 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 - .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d - .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 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 - .byte 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 - .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 - .byte 196,99,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 - .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d - .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 197,172,88,219 // vaddps %ymm3,%ymm10,%ymm3 - .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 + .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 196,98,125,24,37,132,49,0,0 // vbroadcastss 0x3184(%rip),%ymm12 # 5ce8 <_sk_callback_avx+0x1be> + .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,122,49,0,0 // vbroadcastss 0x317a(%rip),%ymm12 # 5cec <_sk_callback_avx+0x1c2> + .byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3 + .byte 196,98,125,24,37,112,49,0,0 // vbroadcastss 0x3170(%rip),%ymm12 # 5cf0 <_sk_callback_avx+0x1c6> + .byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3 + .byte 196,98,125,24,37,102,49,0,0 // vbroadcastss 0x3166(%rip),%ymm12 # 5cf4 <_sk_callback_avx+0x1ca> + .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,92,49,0,0 // vbroadcastss 0x315c(%rip),%ymm12 # 5cf8 <_sk_callback_avx+0x1ce> + .byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12 + .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 + .byte 196,98,125,24,37,77,49,0,0 // vbroadcastss 0x314d(%rip),%ymm12 # 5cfc <_sk_callback_avx+0x1d2> + .byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3 + .byte 196,98,125,24,37,67,49,0,0 // vbroadcastss 0x3143(%rip),%ymm12 # 5d00 <_sk_callback_avx+0x1d6> + .byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3 .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 - .byte 197,28,89,219 // vmulps %ymm3,%ymm12,%ymm11 - .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 + .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 196,98,125,24,29,39,49,0,0 // vbroadcastss 0x3127(%rip),%ymm11 # 5d04 <_sk_callback_avx+0x1da> + .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 + .byte 196,98,125,24,29,29,49,0,0 // vbroadcastss 0x311d(%rip),%ymm11 # 5d08 <_sk_callback_avx+0x1de> + .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 + .byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3 + .byte 196,98,125,24,29,14,49,0,0 // vbroadcastss 0x310e(%rip),%ymm11 # 5d0c <_sk_callback_avx+0x1e2> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%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 65,184,81,140,242,66 // mov $0x42f28c51,%r8d - .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 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 - .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d - .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 196,98,125,24,29,4,49,0,0 // vbroadcastss 0x3104(%rip),%ymm11 # 5d10 <_sk_callback_avx+0x1e6> + .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 + .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 + .byte 196,98,125,24,21,245,48,0,0 // vbroadcastss 0x30f5(%rip),%ymm10 # 5d14 <_sk_callback_avx+0x1ea> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 - .byte 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 - .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 - .byte 196,99,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 - .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d - .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 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3 - .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 - .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 - .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 @@ -14736,7 +14564,7 @@ _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 117,74 // jne 30b8 <_sk_load_a8_avx+0x5a> + .byte 117,74 // jne 2e44 <_sk_load_a8_avx+0x5a> .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 @@ -14763,9 +14591,9 @@ _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 30c0 <_sk_load_a8_avx+0x62> + .byte 117,234 // jne 2e4c <_sk_load_a8_avx+0x62> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,149 // jmp 3072 <_sk_load_a8_avx+0x14> + .byte 235,149 // jmp 2dfe <_sk_load_a8_avx+0x14> HIDDEN _sk_gather_a8_avx .globl _sk_gather_a8_avx @@ -14846,7 +14674,7 @@ _sk_store_a8_avx: .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 3219 <_sk_store_a8_avx+0x42> + .byte 117,10 // jne 2fa5 <_sk_store_a8_avx+0x42> .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -14854,10 +14682,10 @@ _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 3215 <_sk_store_a8_avx+0x3e> + .byte 119,236 // ja 2fa1 <_sk_store_a8_avx+0x3e> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 327c <_sk_store_a8_avx+0xa5> + .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 3008 <_sk_store_a8_avx+0xa5> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -14868,7 +14696,7 @@ _sk_store_a8_avx: .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) - .byte 235,154 // jmp 3215 <_sk_store_a8_avx+0x3e> + .byte 235,154 // jmp 2fa1 <_sk_store_a8_avx+0x3e> .byte 144 // nop .byte 246,255 // idiv %bh .byte 255 // (bad) @@ -14902,7 +14730,7 @@ _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 117,91 // jne 3303 <_sk_load_g8_avx+0x6b> + .byte 117,91 // jne 308f <_sk_load_g8_avx+0x6b> .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 @@ -14932,9 +14760,9 @@ _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 330b <_sk_load_g8_avx+0x73> + .byte 117,234 // jne 3097 <_sk_load_g8_avx+0x73> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,132 // jmp 32ac <_sk_load_g8_avx+0x14> + .byte 235,132 // jmp 3038 <_sk_load_g8_avx+0x14> HIDDEN _sk_gather_g8_avx .globl _sk_gather_g8_avx @@ -15009,9 +14837,9 @@ _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 3442 <_sk_gather_i8_avx+0xf> + .byte 116,5 // je 31ce <_sk_gather_i8_avx+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 3444 <_sk_gather_i8_avx+0x11> + .byte 235,2 // jmp 31d0 <_sk_gather_i8_avx+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -15116,7 +14944,7 @@ _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 15,133,209,0,0,0 // jne 36de <_sk_load_565_avx+0xdf> + .byte 15,133,209,0,0,0 // jne 346a <_sk_load_565_avx+0xdf> .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 @@ -15166,9 +14994,9 @@ _sk_load_565_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 15,135,29,255,255,255 // ja 3613 <_sk_load_565_avx+0x14> + .byte 15,135,29,255,255,255 // ja 339f <_sk_load_565_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 374c <_sk_load_565_avx+0x14d> + .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 34d8 <_sk_load_565_avx+0x14d> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15180,7 +15008,7 @@ _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 233,201,254,255,255 // jmpq 3613 <_sk_load_565_avx+0x14> + .byte 233,201,254,255,255 // jmpq 339f <_sk_load_565_avx+0x14> .byte 102,144 // xchg %ax,%ax .byte 242,255 // repnz (bad) .byte 255 // (bad) @@ -15337,7 +15165,7 @@ _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 3997 <_sk_store_565_avx+0x9e> + .byte 117,10 // jne 3723 <_sk_store_565_avx+0x9e> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15345,9 +15173,9 @@ _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 3993 <_sk_store_565_avx+0x9a> + .byte 119,236 // ja 371f <_sk_store_565_avx+0x9a> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 39f4 <_sk_store_565_avx+0xfb> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 3780 <_sk_store_565_avx+0xfb> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -15358,7 +15186,7 @@ _sk_store_565_avx: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 3993 <_sk_store_565_avx+0x9a> + .byte 235,159 // jmp 371f <_sk_store_565_avx+0x9a> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -15389,7 +15217,7 @@ _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 15,133,245,0,0,0 // jne 3b13 <_sk_load_4444_avx+0x103> + .byte 15,133,245,0,0,0 // jne 389f <_sk_load_4444_avx+0x103> .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 @@ -15446,9 +15274,9 @@ _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 15,135,249,254,255,255 // ja 3a24 <_sk_load_4444_avx+0x14> + .byte 15,135,249,254,255,255 // ja 37b0 <_sk_load_4444_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 3b80 <_sk_load_4444_avx+0x170> + .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 390c <_sk_load_4444_avx+0x170> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15460,12 +15288,12 @@ _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 233,165,254,255,255 // jmpq 3a24 <_sk_load_4444_avx+0x14> + .byte 233,165,254,255,255 // jmpq 37b0 <_sk_load_4444_avx+0x14> .byte 144 // nop .byte 243,255 // repz (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 235,255 // jmp 3b85 <_sk_load_4444_avx+0x175> + .byte 235,255 // jmp 3911 <_sk_load_4444_avx+0x175> .byte 255 // (bad) .byte 255,227 // jmpq *%rbx .byte 255 // (bad) @@ -15626,7 +15454,7 @@ _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 3e00 <_sk_store_4444_avx+0xaf> + .byte 117,10 // jne 3b8c <_sk_store_4444_avx+0xaf> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15634,9 +15462,9 @@ _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 3dfc <_sk_store_4444_avx+0xab> + .byte 119,236 // ja 3b88 <_sk_store_4444_avx+0xab> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3e60 <_sk_store_4444_avx+0x10f> + .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3bec <_sk_store_4444_avx+0x10f> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -15647,7 +15475,7 @@ _sk_store_4444_avx: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 3dfc <_sk_store_4444_avx+0xab> + .byte 235,159 // jmp 3b88 <_sk_store_4444_avx+0xab> .byte 15,31,0 // nopl (%rax) .byte 244 // hlt .byte 255 // (bad) @@ -15680,7 +15508,7 @@ _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 15,133,157,0,0,0 // jne 3f27 <_sk_load_8888_avx+0xab> + .byte 15,133,157,0,0,0 // jne 3cb3 <_sk_load_8888_avx+0xab> .byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9 .byte 184,255,0,0,0 // mov $0xff,%eax .byte 197,249,110,192 // vmovd %eax,%xmm0 @@ -15718,9 +15546,9 @@ _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 15,135,80,255,255,255 // ja 3e90 <_sk_load_8888_avx+0x14> + .byte 15,135,80,255,255,255 // ja 3c1c <_sk_load_8888_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 3fd4 <_sk_load_8888_avx+0x158> + .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 3d60 <_sk_load_8888_avx+0x158> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15743,7 +15571,7 @@ _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 233,188,254,255,255 // jmpq 3e90 <_sk_load_8888_avx+0x14> + .byte 233,188,254,255,255 // jmpq 3c1c <_sk_load_8888_avx+0x14> .byte 238 // out %al,(%dx) .byte 255 // (bad) .byte 255 // (bad) @@ -15873,7 +15701,7 @@ _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 41d5 <_sk_store_8888_avx+0xa4> + .byte 117,10 // jne 3f61 <_sk_store_8888_avx+0xa4> .byte 196,65,124,17,4,185 // vmovups %ymm8,(%r9,%rdi,4) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15881,9 +15709,9 @@ _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 41d1 <_sk_store_8888_avx+0xa0> + .byte 119,236 // ja 3f5d <_sk_store_8888_avx+0xa0> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,84,0,0,0 // lea 0x54(%rip),%r8 # 4244 <_sk_store_8888_avx+0x113> + .byte 76,141,5,84,0,0,0 // lea 0x54(%rip),%r8 # 3fd0 <_sk_store_8888_avx+0x113> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -15897,7 +15725,7 @@ _sk_store_8888_avx: .byte 196,67,121,22,68,185,8,2 // vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) .byte 196,67,121,22,68,185,4,1 // vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) .byte 196,65,121,126,4,185 // vmovd %xmm8,(%r9,%rdi,4) - .byte 235,143 // jmp 41d1 <_sk_store_8888_avx+0xa0> + .byte 235,143 // jmp 3f5d <_sk_store_8888_avx+0xa0> .byte 102,144 // xchg %ax,%ax .byte 246,255 // idiv %bh .byte 255 // (bad) @@ -15926,13 +15754,15 @@ HIDDEN _sk_load_f16_avx .globl _sk_load_f16_avx FUNCTION(_sk_load_f16_avx) _sk_load_f16_avx: + .byte 72,131,236,24 // sub $0x18,%rsp .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 197,252,17,124,36,200 // vmovups %ymm7,-0x38(%rsp) - .byte 197,252,17,116,36,168 // vmovups %ymm6,-0x58(%rsp) - .byte 197,252,17,108,36,136 // vmovups %ymm5,-0x78(%rsp) - .byte 15,133,46,2,0,0 // jne 44ae <_sk_load_f16_avx+0x24e> + .byte 197,252,17,124,36,224 // vmovups %ymm7,-0x20(%rsp) + .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 42a3 <_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 @@ -15941,137 +15771,151 @@ _sk_load_f16_avx: .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 .byte 196,193,113,97,217 // vpunpcklwd %xmm9,%xmm1,%xmm3 .byte 196,193,113,105,201 // vpunpckhwd %xmm9,%xmm1,%xmm1 - .byte 197,121,97,242 // vpunpcklwd %xmm2,%xmm0,%xmm14 + .byte 197,121,97,250 // vpunpcklwd %xmm2,%xmm0,%xmm15 .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8 - .byte 197,97,97,249 // vpunpcklwd %xmm1,%xmm3,%xmm15 - .byte 197,97,105,217 // vpunpckhwd %xmm1,%xmm3,%xmm11 - .byte 196,193,9,108,199 // vpunpcklqdq %xmm15,%xmm14,%xmm0 - .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 - .byte 196,193,121,105,201 // vpunpckhwd %xmm9,%xmm0,%xmm1 + .byte 197,225,97,209 // vpunpcklwd %xmm1,%xmm3,%xmm2 + .byte 197,97,105,201 // vpunpckhwd %xmm1,%xmm3,%xmm9 + .byte 197,129,108,194 // vpunpcklqdq %xmm2,%xmm15,%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 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - .byte 184,0,128,0,0 // mov $0x8000,%eax - .byte 197,249,110,200 // vmovd %eax,%xmm1 - .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 - .byte 196,99,117,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm10 - .byte 196,193,124,84,202 // vandps %ymm10,%ymm0,%ymm1 + .byte 196,98,125,24,37,169,28,0,0 // vbroadcastss 0x1ca9(%rip),%ymm12 # 5d18 <_sk_callback_avx+0x1ee> + .byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1 .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0 - .byte 184,0,4,0,0 // mov $0x400,%eax - .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 - .byte 197,249,110,208 // vmovd %eax,%xmm2 - .byte 197,121,112,226,0 // vpshufd $0x0,%xmm2,%xmm12 - .byte 197,153,102,211 // vpcmpgtd %xmm3,%xmm12,%xmm2 - .byte 197,25,102,232 // vpcmpgtd %xmm0,%xmm12,%xmm13 - .byte 196,227,21,24,242,1 // vinsertf128 $0x1,%xmm2,%ymm13,%ymm6 - .byte 196,227,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm2 - .byte 197,145,114,242,16 // vpslld $0x10,%xmm2,%xmm13 - .byte 197,233,114,243,13 // vpslld $0xd,%xmm3,%xmm2 - .byte 184,0,0,0,56 // mov $0x38000000,%eax - .byte 197,249,110,216 // vmovd %eax,%xmm3 - .byte 197,249,112,235,0 // vpshufd $0x0,%xmm3,%xmm5 - .byte 197,145,254,253 // vpaddd %xmm5,%xmm13,%xmm7 - .byte 197,193,254,210 // vpaddd %xmm2,%xmm7,%xmm2 + .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 + .byte 196,98,121,24,29,149,28,0,0 // vbroadcastss 0x1c95(%rip),%xmm11 # 5d1c <_sk_callback_avx+0x1f2> + .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 + .byte 196,98,121,24,45,139,28,0,0 // vbroadcastss 0x1c8b(%rip),%xmm13 # 5d20 <_sk_callback_avx+0x1f6> + .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 + .byte 196,99,45,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm10,%ymm10 + .byte 197,225,114,241,16 // vpslld $0x10,%xmm1,%xmm3 + .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 .byte 197,241,114,241,16 // vpslld $0x10,%xmm1,%xmm1 + .byte 196,227,101,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0 - .byte 197,241,254,205 // vpaddd %xmm5,%xmm1,%xmm1 - .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 - .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - .byte 196,65,20,87,237 // vxorps %ymm13,%ymm13,%ymm13 - .byte 196,195,125,74,197,96 // vblendvps %ymm6,%ymm13,%ymm0,%ymm0 - .byte 196,193,9,109,207 // vpunpckhqdq %xmm15,%xmm14,%xmm1 - .byte 196,193,113,105,209 // vpunpckhwd %xmm9,%xmm1,%xmm2 + .byte 196,193,97,114,246,13 // vpslld $0xd,%xmm14,%xmm3 + .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 196,226,121,24,29,65,28,0,0 // vbroadcastss 0x1c41(%rip),%xmm3 # 5d24 <_sk_callback_avx+0x1fa> + .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 + .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14 + .byte 196,195,125,74,198,160 // vblendvps %ymm10,%ymm14,%ymm0,%ymm0 + .byte 197,129,109,202 // vpunpckhqdq %xmm2,%xmm15,%xmm1 + .byte 197,217,239,228 // vpxor %xmm4,%xmm4,%xmm4 + .byte 197,241,105,212 // vpunpckhwd %xmm4,%xmm1,%xmm2 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - .byte 196,193,116,84,210 // vandps %ymm10,%ymm1,%ymm2 + .byte 196,193,116,84,212 // vandps %ymm12,%ymm1,%ymm2 .byte 197,244,87,202 // vxorps %ymm2,%ymm1,%ymm1 - .byte 196,227,125,25,206,1 // vextractf128 $0x1,%ymm1,%xmm6 - .byte 197,153,102,254 // vpcmpgtd %xmm6,%xmm12,%xmm7 - .byte 197,25,102,241 // vpcmpgtd %xmm1,%xmm12,%xmm14 - .byte 196,99,13,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm14,%ymm14 - .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7 - .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 - .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 - .byte 197,193,254,253 // vpaddd %xmm5,%xmm7,%xmm7 - .byte 197,193,254,246 // vpaddd %xmm6,%xmm7,%xmm6 + .byte 196,195,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm10 + .byte 196,193,40,87,251 // vxorps %xmm11,%xmm10,%xmm7 + .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7 + .byte 196,193,112,87,243 // vxorps %xmm11,%xmm1,%xmm6 + .byte 197,145,102,246 // vpcmpgtd %xmm6,%xmm13,%xmm6 + .byte 196,227,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 + .byte 197,193,114,242,16 // vpslld $0x10,%xmm2,%xmm7 + .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2 + .byte 196,227,69,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm7,%ymm2 .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1 - .byte 197,233,254,213 // vpaddd %xmm5,%xmm2,%xmm2 - .byte 197,233,254,201 // vpaddd %xmm1,%xmm2,%xmm1 - .byte 196,227,117,24,206,1 // vinsertf128 $0x1,%xmm6,%ymm1,%ymm1 - .byte 196,195,117,74,205,224 // vblendvps %ymm14,%ymm13,%ymm1,%ymm1 - .byte 196,193,57,108,211 // vpunpcklqdq %xmm11,%xmm8,%xmm2 - .byte 196,193,105,105,241 // vpunpckhwd %xmm9,%xmm2,%xmm6 + .byte 196,193,65,114,242,13 // vpslld $0xd,%xmm10,%xmm7 + .byte 196,227,117,24,207,1 // vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 + .byte 197,244,86,202 // vorps %ymm2,%ymm1,%ymm1 + .byte 196,227,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm2 + .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 + .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 + .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 + .byte 196,195,117,74,206,96 // vblendvps %ymm6,%ymm14,%ymm1,%ymm1 + .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2 + .byte 197,233,105,244 // vpunpckhwd %xmm4,%xmm2,%xmm6 + .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - .byte 196,193,108,84,242 // vandps %ymm10,%ymm2,%ymm6 + .byte 196,193,108,84,244 // vandps %ymm12,%ymm2,%ymm6 .byte 197,236,87,214 // vxorps %ymm6,%ymm2,%ymm2 - .byte 196,195,125,25,214,1 // vextractf128 $0x1,%ymm2,%xmm14 - .byte 196,193,25,102,254 // vpcmpgtd %xmm14,%xmm12,%xmm7 - .byte 197,25,102,250 // vpcmpgtd %xmm2,%xmm12,%xmm15 - .byte 196,99,5,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm15,%ymm15 - .byte 196,227,125,25,247,1 // vextractf128 $0x1,%ymm6,%xmm7 - .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 - .byte 196,193,9,114,246,13 // vpslld $0xd,%xmm14,%xmm14 - .byte 197,193,254,253 // vpaddd %xmm5,%xmm7,%xmm7 - .byte 196,193,65,254,254 // vpaddd %xmm14,%xmm7,%xmm7 + .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7 + .byte 196,193,64,87,235 // vxorps %xmm11,%xmm7,%xmm5 + .byte 197,145,102,237 // vpcmpgtd %xmm5,%xmm13,%xmm5 + .byte 196,193,104,87,227 // vxorps %xmm11,%xmm2,%xmm4 + .byte 197,145,102,228 // vpcmpgtd %xmm4,%xmm13,%xmm4 + .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + .byte 197,209,114,246,16 // vpslld $0x10,%xmm6,%xmm5 + .byte 196,227,125,25,246,1 // vextractf128 $0x1,%ymm6,%xmm6 .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6 + .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2 - .byte 197,201,254,245 // vpaddd %xmm5,%xmm6,%xmm6 - .byte 197,201,254,210 // vpaddd %xmm2,%xmm6,%xmm2 - .byte 196,227,109,24,215,1 // vinsertf128 $0x1,%xmm7,%ymm2,%ymm2 - .byte 196,195,109,74,213,240 // vblendvps %ymm15,%ymm13,%ymm2,%ymm2 - .byte 196,193,57,109,243 // vpunpckhqdq %xmm11,%xmm8,%xmm6 - .byte 196,193,73,105,249 // vpunpckhwd %xmm9,%xmm6,%xmm7 - .byte 196,226,121,51,246 // vpmovzxwd %xmm6,%xmm6 - .byte 196,227,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 - .byte 196,193,76,84,250 // vandps %ymm10,%ymm6,%ymm7 - .byte 197,204,87,247 // vxorps %ymm7,%ymm6,%ymm6 - .byte 196,195,125,25,240,1 // vextractf128 $0x1,%ymm6,%xmm8 - .byte 196,65,25,102,200 // vpcmpgtd %xmm8,%xmm12,%xmm9 - .byte 197,25,102,214 // vpcmpgtd %xmm6,%xmm12,%xmm10 - .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - .byte 196,227,125,25,251,1 // vextractf128 $0x1,%ymm7,%xmm3 - .byte 197,225,114,243,16 // vpslld $0x10,%xmm3,%xmm3 - .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 - .byte 197,193,254,253 // vpaddd %xmm5,%xmm7,%xmm7 - .byte 197,225,254,221 // vpaddd %xmm5,%xmm3,%xmm3 - .byte 196,193,81,114,240,13 // vpslld $0xd,%xmm8,%xmm5 - .byte 197,225,254,221 // vpaddd %xmm5,%xmm3,%xmm3 - .byte 197,209,114,246,13 // vpslld $0xd,%xmm6,%xmm5 - .byte 197,193,254,237 // vpaddd %xmm5,%xmm7,%xmm5 - .byte 196,227,85,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm5,%ymm3 - .byte 196,195,101,74,221,144 // vblendvps %ymm9,%ymm13,%ymm3,%ymm3 - .byte 72,173 // lods %ds:(%rsi),%rax - .byte 197,252,16,108,36,136 // vmovups -0x78(%rsp),%ymm5 - .byte 197,252,16,116,36,168 // vmovups -0x58(%rsp),%ymm6 - .byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7 + .byte 197,201,114,247,13 // vpslld $0xd,%xmm7,%xmm6 + .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 + .byte 197,236,86,213 // vorps %ymm5,%ymm2,%ymm2 + .byte 196,227,125,25,213,1 // vextractf128 $0x1,%ymm2,%xmm5 + .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5 + .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 + .byte 196,227,109,24,213,1 // vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 + .byte 196,195,109,74,214,64 // vblendvps %ymm4,%ymm14,%ymm2,%ymm2 + .byte 196,193,57,109,225 // vpunpckhqdq %xmm9,%xmm8,%xmm4 + .byte 196,193,89,105,234 // vpunpckhwd %xmm10,%xmm4,%xmm5 + .byte 196,226,121,51,228 // vpmovzxwd %xmm4,%xmm4 + .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + .byte 196,193,92,84,236 // vandps %ymm12,%ymm4,%ymm5 + .byte 197,220,87,229 // vxorps %ymm5,%ymm4,%ymm4 + .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6 + .byte 196,193,72,87,251 // vxorps %xmm11,%xmm6,%xmm7 + .byte 197,17,102,199 // vpcmpgtd %xmm7,%xmm13,%xmm8 + .byte 196,193,88,87,251 // vxorps %xmm11,%xmm4,%xmm7 + .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7 + .byte 196,195,69,24,248,1 // vinsertf128 $0x1,%xmm8,%ymm7,%ymm7 + .byte 197,185,114,245,16 // vpslld $0x10,%xmm5,%xmm8 + .byte 196,227,125,25,237,1 // vextractf128 $0x1,%ymm5,%xmm5 + .byte 197,209,114,245,16 // vpslld $0x10,%xmm5,%xmm5 + .byte 196,227,61,24,237,1 // vinsertf128 $0x1,%xmm5,%ymm8,%ymm5 + .byte 197,217,114,244,13 // vpslld $0xd,%xmm4,%xmm4 + .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 + .byte 196,227,93,24,230,1 // vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 + .byte 197,220,86,229 // vorps %ymm5,%ymm4,%ymm4 + .byte 196,227,125,25,229,1 // vextractf128 $0x1,%ymm4,%xmm5 + .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5 + .byte 197,217,254,219 // vpaddd %xmm3,%xmm4,%xmm3 + .byte 196,227,101,24,221,1 // vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 + .byte 196,195,101,74,222,112 // vblendvps %ymm7,%ymm14,%ymm3,%ymm3 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 197,252,16,100,36,128 // vmovups -0x80(%rsp),%ymm4 + .byte 197,252,16,108,36,160 // vmovups -0x60(%rsp),%ymm5 + .byte 197,252,16,116,36,192 // vmovups -0x40(%rsp),%ymm6 + .byte 197,252,16,124,36,224 // vmovups -0x20(%rsp),%ymm7 + .byte 72,131,196,24 // add $0x18,%rsp .byte 255,224 // jmpq *%rax .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 450d <_sk_load_f16_avx+0x2ad> + .byte 116,79 // je 4302 <_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 450d <_sk_load_f16_avx+0x2ad> + .byte 114,67 // jb 4302 <_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 451a <_sk_load_f16_avx+0x2ba> + .byte 116,68 // je 430f <_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 451a <_sk_load_f16_avx+0x2ba> + .byte 114,56 // jb 430f <_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 15,132,165,253,255,255 // je 4297 <_sk_load_f16_avx+0x37> + .byte 15,132,70,253,255,255 // je 402d <_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 15,130,149,253,255,255 // jb 4297 <_sk_load_f16_avx+0x37> + .byte 15,130,54,253,255,255 // jb 402d <_sk_load_f16_avx+0x41> .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 - .byte 233,138,253,255,255 // jmpq 4297 <_sk_load_f16_avx+0x37> + .byte 233,43,253,255,255 // jmpq 402d <_sk_load_f16_avx+0x41> .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,125,253,255,255 // jmpq 4297 <_sk_load_f16_avx+0x37> + .byte 233,30,253,255,255 // jmpq 402d <_sk_load_f16_avx+0x41> .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 - .byte 233,116,253,255,255 // jmpq 4297 <_sk_load_f16_avx+0x37> + .byte 233,21,253,255,255 // jmpq 402d <_sk_load_f16_avx+0x41> HIDDEN _sk_gather_f16_avx .globl _sk_gather_f16_avx @@ -16085,7 +15929,7 @@ _sk_gather_f16_avx: .byte 197,252,17,124,36,224 // vmovups %ymm7,-0x20(%rsp) .byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp) .byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp) - .byte 197,252,17,100,36,128 // vmovups %ymm4,-0x80(%rsp) + .byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp) .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,0 // mov (%rax),%r8 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 @@ -16126,106 +15970,118 @@ _sk_gather_f16_avx: .byte 197,177,105,201 // vpunpckhwd %xmm1,%xmm9,%xmm1 .byte 197,169,97,211 // vpunpcklwd %xmm3,%xmm10,%xmm2 .byte 197,169,105,219 // vpunpckhwd %xmm3,%xmm10,%xmm3 - .byte 197,121,97,241 // vpunpcklwd %xmm1,%xmm0,%xmm14 + .byte 197,121,97,249 // vpunpcklwd %xmm1,%xmm0,%xmm15 .byte 197,121,105,193 // vpunpckhwd %xmm1,%xmm0,%xmm8 - .byte 197,105,97,251 // vpunpcklwd %xmm3,%xmm2,%xmm15 - .byte 197,105,105,219 // vpunpckhwd %xmm3,%xmm2,%xmm11 - .byte 196,193,9,108,199 // vpunpcklqdq %xmm15,%xmm14,%xmm0 - .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 - .byte 196,193,121,105,209 // vpunpckhwd %xmm9,%xmm0,%xmm2 + .byte 197,233,97,203 // vpunpcklwd %xmm3,%xmm2,%xmm1 + .byte 197,105,105,203 // vpunpckhwd %xmm3,%xmm2,%xmm9 + .byte 197,129,108,193 // vpunpcklqdq %xmm1,%xmm15,%xmm0 + .byte 197,233,239,210 // vpxor %xmm2,%xmm2,%xmm2 + .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 184,0,128,0,0 // mov $0x8000,%eax - .byte 197,249,110,208 // vmovd %eax,%xmm2 - .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 - .byte 196,99,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm10 - .byte 196,193,124,84,210 // vandps %ymm10,%ymm0,%ymm2 + .byte 196,98,125,24,37,5,25,0,0 // vbroadcastss 0x1905(%rip),%ymm12 # 5d28 <_sk_callback_avx+0x1fe> + .byte 196,193,124,84,212 // vandps %ymm12,%ymm0,%ymm2 .byte 197,252,87,194 // vxorps %ymm2,%ymm0,%ymm0 - .byte 184,0,4,0,0 // mov $0x400,%eax - .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 - .byte 197,249,110,200 // vmovd %eax,%xmm1 - .byte 197,121,112,225,0 // vpshufd $0x0,%xmm1,%xmm12 - .byte 197,153,102,203 // vpcmpgtd %xmm3,%xmm12,%xmm1 - .byte 197,25,102,232 // vpcmpgtd %xmm0,%xmm12,%xmm13 - .byte 196,227,21,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm13,%ymm4 - .byte 196,227,125,25,209,1 // vextractf128 $0x1,%ymm2,%xmm1 - .byte 197,145,114,241,16 // vpslld $0x10,%xmm1,%xmm13 - .byte 197,241,114,243,13 // vpslld $0xd,%xmm3,%xmm1 - .byte 184,0,0,0,56 // mov $0x38000000,%eax - .byte 197,249,110,216 // vmovd %eax,%xmm3 - .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 - .byte 197,145,254,251 // vpaddd %xmm3,%xmm13,%xmm7 - .byte 197,193,254,201 // vpaddd %xmm1,%xmm7,%xmm1 + .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 + .byte 196,98,121,24,29,241,24,0,0 // vbroadcastss 0x18f1(%rip),%xmm11 # 5d2c <_sk_callback_avx+0x202> + .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 + .byte 196,98,121,24,45,231,24,0,0 // vbroadcastss 0x18e7(%rip),%xmm13 # 5d30 <_sk_callback_avx+0x206> + .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 + .byte 196,99,45,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm10,%ymm10 + .byte 197,225,114,242,16 // vpslld $0x10,%xmm2,%xmm3 + .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2 + .byte 196,227,101,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0 + .byte 196,193,97,114,246,13 // vpslld $0xd,%xmm14,%xmm3 + .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 196,226,121,24,29,157,24,0,0 // vbroadcastss 0x189d(%rip),%xmm3 # 5d34 <_sk_callback_avx+0x20a> .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 - .byte 197,233,254,192 // vpaddd %xmm0,%xmm2,%xmm0 - .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - .byte 196,65,20,87,237 // vxorps %ymm13,%ymm13,%ymm13 - .byte 196,195,125,74,197,64 // vblendvps %ymm4,%ymm13,%ymm0,%ymm0 - .byte 196,193,9,109,207 // vpunpckhqdq %xmm15,%xmm14,%xmm1 - .byte 196,193,113,105,209 // vpunpckhwd %xmm9,%xmm1,%xmm2 + .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 + .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 + .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14 + .byte 196,195,125,74,198,160 // vblendvps %ymm10,%ymm14,%ymm0,%ymm0 + .byte 197,129,109,201 // vpunpckhqdq %xmm1,%xmm15,%xmm1 + .byte 197,217,239,228 // vpxor %xmm4,%xmm4,%xmm4 + .byte 197,241,105,212 // vpunpckhwd %xmm4,%xmm1,%xmm2 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - .byte 196,193,116,84,210 // vandps %ymm10,%ymm1,%ymm2 + .byte 196,193,116,84,212 // vandps %ymm12,%ymm1,%ymm2 .byte 197,244,87,202 // vxorps %ymm2,%ymm1,%ymm1 - .byte 196,227,125,25,204,1 // vextractf128 $0x1,%ymm1,%xmm4 - .byte 197,153,102,252 // vpcmpgtd %xmm4,%xmm12,%xmm7 - .byte 197,25,102,241 // vpcmpgtd %xmm1,%xmm12,%xmm14 - .byte 196,227,13,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm14,%ymm7 - .byte 196,227,125,25,214,1 // vextractf128 $0x1,%ymm2,%xmm6 - .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6 - .byte 197,217,114,244,13 // vpslld $0xd,%xmm4,%xmm4 - .byte 197,201,254,243 // vpaddd %xmm3,%xmm6,%xmm6 - .byte 197,201,254,228 // vpaddd %xmm4,%xmm6,%xmm4 + .byte 196,195,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm10 + .byte 196,193,40,87,251 // vxorps %xmm11,%xmm10,%xmm7 + .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7 + .byte 196,193,112,87,243 // vxorps %xmm11,%xmm1,%xmm6 + .byte 197,145,102,246 // vpcmpgtd %xmm6,%xmm13,%xmm6 + .byte 196,227,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 + .byte 197,193,114,242,16 // vpslld $0x10,%xmm2,%xmm7 + .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2 + .byte 196,227,69,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm7,%ymm2 .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1 + .byte 196,193,65,114,242,13 // vpslld $0xd,%xmm10,%xmm7 + .byte 196,227,117,24,207,1 // vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 + .byte 197,244,86,202 // vorps %ymm2,%ymm1,%ymm1 + .byte 196,227,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm2 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 - .byte 197,233,254,201 // vpaddd %xmm1,%xmm2,%xmm1 - .byte 196,227,117,24,204,1 // vinsertf128 $0x1,%xmm4,%ymm1,%ymm1 - .byte 196,195,117,74,205,112 // vblendvps %ymm7,%ymm13,%ymm1,%ymm1 - .byte 196,193,57,108,211 // vpunpcklqdq %xmm11,%xmm8,%xmm2 - .byte 196,193,105,105,225 // vpunpckhwd %xmm9,%xmm2,%xmm4 + .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 + .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 + .byte 196,195,117,74,206,96 // vblendvps %ymm6,%ymm14,%ymm1,%ymm1 + .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2 + .byte 197,233,105,244 // vpunpckhwd %xmm4,%xmm2,%xmm6 + .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 - .byte 196,227,109,24,212,1 // vinsertf128 $0x1,%xmm4,%ymm2,%ymm2 - .byte 196,193,108,84,226 // vandps %ymm10,%ymm2,%ymm4 - .byte 197,236,87,212 // vxorps %ymm4,%ymm2,%ymm2 - .byte 196,227,125,25,214,1 // vextractf128 $0x1,%ymm2,%xmm6 - .byte 197,153,102,254 // vpcmpgtd %xmm6,%xmm12,%xmm7 - .byte 197,25,102,242 // vpcmpgtd %xmm2,%xmm12,%xmm14 - .byte 196,227,13,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm14,%ymm7 - .byte 196,227,125,25,229,1 // vextractf128 $0x1,%ymm4,%xmm5 - .byte 197,209,114,245,16 // vpslld $0x10,%xmm5,%xmm5 - .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 - .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5 - .byte 197,209,254,238 // vpaddd %xmm6,%xmm5,%xmm5 - .byte 197,217,114,244,16 // vpslld $0x10,%xmm4,%xmm4 + .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 + .byte 196,193,108,84,244 // vandps %ymm12,%ymm2,%ymm6 + .byte 197,236,87,214 // vxorps %ymm6,%ymm2,%ymm2 + .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7 + .byte 196,193,64,87,235 // vxorps %xmm11,%xmm7,%xmm5 + .byte 197,145,102,237 // vpcmpgtd %xmm5,%xmm13,%xmm5 + .byte 196,193,104,87,227 // vxorps %xmm11,%xmm2,%xmm4 + .byte 197,145,102,228 // vpcmpgtd %xmm4,%xmm13,%xmm4 + .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + .byte 197,209,114,246,16 // vpslld $0x10,%xmm6,%xmm5 + .byte 196,227,125,25,246,1 // vextractf128 $0x1,%ymm6,%xmm6 + .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6 + .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2 - .byte 197,217,254,227 // vpaddd %xmm3,%xmm4,%xmm4 - .byte 197,217,254,210 // vpaddd %xmm2,%xmm4,%xmm2 + .byte 197,201,114,247,13 // vpslld $0xd,%xmm7,%xmm6 + .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 + .byte 197,236,86,213 // vorps %ymm5,%ymm2,%ymm2 + .byte 196,227,125,25,213,1 // vextractf128 $0x1,%ymm2,%xmm5 + .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5 + .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 .byte 196,227,109,24,213,1 // vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 - .byte 196,195,109,74,213,112 // vblendvps %ymm7,%ymm13,%ymm2,%ymm2 - .byte 196,193,57,109,227 // vpunpckhqdq %xmm11,%xmm8,%xmm4 - .byte 196,193,89,105,233 // vpunpckhwd %xmm9,%xmm4,%xmm5 + .byte 196,195,109,74,214,64 // vblendvps %ymm4,%ymm14,%ymm2,%ymm2 + .byte 196,193,57,109,225 // vpunpckhqdq %xmm9,%xmm8,%xmm4 + .byte 196,193,89,105,234 // vpunpckhwd %xmm10,%xmm4,%xmm5 .byte 196,226,121,51,228 // vpmovzxwd %xmm4,%xmm4 .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - .byte 196,193,92,84,234 // vandps %ymm10,%ymm4,%ymm5 + .byte 196,193,92,84,236 // vandps %ymm12,%ymm4,%ymm5 .byte 197,220,87,229 // vxorps %ymm5,%ymm4,%ymm4 .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6 - .byte 197,153,102,254 // vpcmpgtd %xmm6,%xmm12,%xmm7 - .byte 197,25,102,196 // vpcmpgtd %xmm4,%xmm12,%xmm8 - .byte 196,99,61,24,199,1 // vinsertf128 $0x1,%xmm7,%ymm8,%ymm8 - .byte 196,227,125,25,239,1 // vextractf128 $0x1,%ymm5,%xmm7 - .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 + .byte 196,193,72,87,251 // vxorps %xmm11,%xmm6,%xmm7 + .byte 197,17,102,199 // vpcmpgtd %xmm7,%xmm13,%xmm8 + .byte 196,193,88,87,251 // vxorps %xmm11,%xmm4,%xmm7 + .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7 + .byte 196,195,69,24,248,1 // vinsertf128 $0x1,%xmm8,%ymm7,%ymm7 + .byte 197,185,114,245,16 // vpslld $0x10,%xmm5,%xmm8 + .byte 196,227,125,25,237,1 // vextractf128 $0x1,%ymm5,%xmm5 .byte 197,209,114,245,16 // vpslld $0x10,%xmm5,%xmm5 - .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5 - .byte 197,193,254,219 // vpaddd %xmm3,%xmm7,%xmm3 - .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 - .byte 197,225,254,222 // vpaddd %xmm6,%xmm3,%xmm3 + .byte 196,227,61,24,237,1 // vinsertf128 $0x1,%xmm5,%ymm8,%ymm5 .byte 197,217,114,244,13 // vpslld $0xd,%xmm4,%xmm4 - .byte 197,209,254,228 // vpaddd %xmm4,%xmm5,%xmm4 - .byte 196,227,93,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm4,%ymm3 - .byte 196,195,101,74,221,128 // vblendvps %ymm8,%ymm13,%ymm3,%ymm3 + .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 + .byte 196,227,93,24,230,1 // vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 + .byte 197,220,86,229 // vorps %ymm5,%ymm4,%ymm4 + .byte 196,227,125,25,229,1 // vextractf128 $0x1,%ymm4,%xmm5 + .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5 + .byte 197,217,254,219 // vpaddd %xmm3,%xmm4,%xmm3 + .byte 196,227,101,24,221,1 // vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 + .byte 196,195,101,74,222,112 // vblendvps %ymm7,%ymm14,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,16,100,36,128 // vmovups -0x80(%rsp),%ymm4 .byte 197,252,16,108,36,160 // vmovups -0x60(%rsp),%ymm5 @@ -16247,139 +16103,135 @@ _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 197,252,40,225 // vmovaps %ymm1,%ymm4 - .byte 72,173 // lods %ds:(%rsi),%rax - .byte 76,139,0 // mov (%rax),%r8 - .byte 184,0,0,0,128 // mov $0x80000000,%eax - .byte 197,121,110,192 // vmovd %eax,%xmm8 - .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 - .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 - .byte 197,60,84,208 // vandps %ymm0,%ymm8,%ymm10 + .byte 196,98,125,24,13,182,22,0,0 // vbroadcastss 0x16b6(%rip),%ymm9 # 5d38 <_sk_callback_avx+0x20e> + .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 184,0,0,128,56 // mov $0x38800000,%eax .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 - .byte 197,121,110,200 // vmovd %eax,%xmm9 - .byte 196,65,121,112,201,0 // vpshufd $0x0,%xmm9,%xmm9 - .byte 196,65,49,102,236 // vpcmpgtd %xmm12,%xmm9,%xmm13 - .byte 196,65,49,102,243 // vpcmpgtd %xmm11,%xmm9,%xmm14 + .byte 196,98,121,24,5,155,22,0,0 // vbroadcastss 0x169b(%rip),%xmm8 # 5d3c <_sk_callback_avx+0x212> + .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 - .byte 196,67,125,25,214,1 // vextractf128 $0x1,%ymm10,%xmm14 - .byte 196,193,9,114,214,16 // vpsrld $0x10,%xmm14,%xmm14 - .byte 196,193,1,114,210,16 // vpsrld $0x10,%xmm10,%xmm15 + .byte 196,193,9,114,210,16 // vpsrld $0x10,%xmm10,%xmm14 + .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 + .byte 196,193,41,114,210,16 // vpsrld $0x10,%xmm10,%xmm10 + .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 184,0,192,1,0 // mov $0x1c000,%eax - .byte 197,121,110,208 // vmovd %eax,%xmm10 - .byte 196,65,121,112,210,0 // vpshufd $0x0,%xmm10,%xmm10 - .byte 196,65,1,250,250 // vpsubd %xmm10,%xmm15,%xmm15 - .byte 196,65,9,250,242 // vpsubd %xmm10,%xmm14,%xmm14 - .byte 196,65,9,254,228 // vpaddd %xmm12,%xmm14,%xmm12 - .byte 196,65,1,254,219 // vpaddd %xmm11,%xmm15,%xmm11 + .byte 196,98,125,24,21,98,22,0,0 // vbroadcastss 0x1662(%rip),%ymm10 # 5d40 <_sk_callback_avx+0x216> + .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 + .byte 196,65,9,254,219 // vpaddd %xmm11,%xmm14,%xmm11 .byte 196,67,37,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm11,%ymm12 - .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 - .byte 196,99,29,74,225,208 // vblendvps %ymm13,%ymm1,%ymm12,%ymm12 - .byte 197,60,84,236 // vandps %ymm4,%ymm8,%ymm13 - .byte 197,252,17,100,36,128 // vmovups %ymm4,-0x80(%rsp) - .byte 196,65,92,87,245 // vxorps %ymm13,%ymm4,%ymm14 + .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 + .byte 196,99,29,74,224,208 // vblendvps %ymm13,%ymm0,%ymm12,%ymm12 + .byte 196,65,116,84,233 // vandps %ymm9,%ymm1,%ymm13 + .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp) + .byte 196,65,116,87,245 // vxorps %ymm13,%ymm1,%ymm14 .byte 196,67,125,25,247,1 // vextractf128 $0x1,%ymm14,%xmm15 - .byte 196,193,49,102,255 // vpcmpgtd %xmm15,%xmm9,%xmm7 - .byte 196,65,49,102,222 // vpcmpgtd %xmm14,%xmm9,%xmm11 - .byte 196,99,37,24,223,1 // vinsertf128 $0x1,%xmm7,%ymm11,%ymm11 + .byte 196,193,57,102,255 // vpcmpgtd %xmm15,%xmm8,%xmm7 + .byte 196,65,57,102,222 // vpcmpgtd %xmm14,%xmm8,%xmm11 + .byte 196,227,37,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm11,%ymm7 + .byte 196,193,33,114,213,16 // vpsrld $0x10,%xmm13,%xmm11 .byte 196,99,125,25,238,1 // vextractf128 $0x1,%ymm13,%xmm6 .byte 197,201,114,214,16 // vpsrld $0x10,%xmm6,%xmm6 - .byte 196,193,65,114,215,13 // vpsrld $0xd,%xmm15,%xmm7 - .byte 196,193,73,250,242 // vpsubd %xmm10,%xmm6,%xmm6 - .byte 197,73,254,255 // vpaddd %xmm7,%xmm6,%xmm15 - .byte 196,193,65,114,213,16 // vpsrld $0x10,%xmm13,%xmm7 - .byte 196,193,73,114,214,13 // vpsrld $0xd,%xmm14,%xmm6 - .byte 196,193,65,250,250 // vpsubd %xmm10,%xmm7,%xmm7 - .byte 197,193,254,246 // vpaddd %xmm6,%xmm7,%xmm6 - .byte 196,195,77,24,247,1 // vinsertf128 $0x1,%xmm15,%ymm6,%ymm6 - .byte 196,99,77,74,233,176 // vblendvps %ymm11,%ymm1,%ymm6,%ymm13 - .byte 197,188,84,242 // vandps %ymm2,%ymm8,%ymm6 - .byte 197,252,17,84,36,160 // vmovups %ymm2,-0x60(%rsp) - .byte 197,236,87,254 // vxorps %ymm6,%ymm2,%ymm7 - .byte 196,195,125,25,251,1 // vextractf128 $0x1,%ymm7,%xmm11 - .byte 196,65,49,102,243 // vpcmpgtd %xmm11,%xmm9,%xmm14 - .byte 197,49,102,255 // vpcmpgtd %xmm7,%xmm9,%xmm15 - .byte 196,67,5,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm15,%ymm14 + .byte 196,227,37,24,246,1 // vinsertf128 $0x1,%xmm6,%ymm11,%ymm6 + .byte 196,193,33,114,215,13 // vpsrld $0xd,%xmm15,%xmm11 + .byte 196,193,76,86,242 // vorps %ymm10,%ymm6,%ymm6 .byte 196,227,125,25,245,1 // vextractf128 $0x1,%ymm6,%xmm5 - .byte 197,129,114,213,16 // vpsrld $0x10,%xmm5,%xmm15 - .byte 196,193,81,114,211,13 // vpsrld $0xd,%xmm11,%xmm5 - .byte 196,193,1,250,226 // vpsubd %xmm10,%xmm15,%xmm4 + .byte 196,193,81,254,235 // vpaddd %xmm11,%xmm5,%xmm5 + .byte 196,193,89,114,214,13 // vpsrld $0xd,%xmm14,%xmm4 + .byte 197,201,254,228 // vpaddd %xmm4,%xmm6,%xmm4 + .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + .byte 196,99,93,74,232,112 // vblendvps %ymm7,%ymm0,%ymm4,%ymm13 + .byte 196,193,108,84,225 // vandps %ymm9,%ymm2,%ymm4 + .byte 197,236,87,236 // vxorps %ymm4,%ymm2,%ymm5 + .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6 + .byte 197,185,102,254 // vpcmpgtd %xmm6,%xmm8,%xmm7 + .byte 197,57,102,221 // vpcmpgtd %xmm5,%xmm8,%xmm11 + .byte 196,227,37,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm11,%ymm7 + .byte 197,161,114,212,16 // vpsrld $0x10,%xmm4,%xmm11 + .byte 196,227,125,25,228,1 // vextractf128 $0x1,%ymm4,%xmm4 + .byte 197,217,114,212,16 // vpsrld $0x10,%xmm4,%xmm4 + .byte 196,227,37,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm11,%ymm4 + .byte 197,201,114,214,13 // vpsrld $0xd,%xmm6,%xmm6 + .byte 196,193,92,86,226 // vorps %ymm10,%ymm4,%ymm4 + .byte 196,227,125,25,225,1 // vextractf128 $0x1,%ymm4,%xmm1 + .byte 197,241,254,206 // vpaddd %xmm6,%xmm1,%xmm1 + .byte 197,209,114,213,13 // vpsrld $0xd,%xmm5,%xmm5 .byte 197,217,254,229 // vpaddd %xmm5,%xmm4,%xmm4 - .byte 197,209,114,214,16 // vpsrld $0x10,%xmm6,%xmm5 - .byte 197,201,114,215,13 // vpsrld $0xd,%xmm7,%xmm6 - .byte 196,193,81,250,234 // vpsubd %xmm10,%xmm5,%xmm5 - .byte 197,209,254,238 // vpaddd %xmm6,%xmm5,%xmm5 - .byte 196,227,85,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 - .byte 196,99,93,74,217,224 // vblendvps %ymm14,%ymm1,%ymm4,%ymm11 - .byte 197,188,84,235 // vandps %ymm3,%ymm8,%ymm5 - .byte 197,228,87,245 // vxorps %ymm5,%ymm3,%ymm6 - .byte 196,227,125,25,247,1 // vextractf128 $0x1,%ymm6,%xmm7 - .byte 197,177,102,231 // vpcmpgtd %xmm7,%xmm9,%xmm4 - .byte 197,49,102,198 // vpcmpgtd %xmm6,%xmm9,%xmm8 + .byte 196,227,93,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 + .byte 196,99,117,74,216,112 // vblendvps %ymm7,%ymm0,%ymm1,%ymm11 + .byte 196,193,100,84,225 // vandps %ymm9,%ymm3,%ymm4 + .byte 197,228,87,236 // vxorps %ymm4,%ymm3,%ymm5 + .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6 + .byte 197,185,102,254 // vpcmpgtd %xmm6,%xmm8,%xmm7 + .byte 197,57,102,197 // vpcmpgtd %xmm5,%xmm8,%xmm8 + .byte 196,227,61,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm8,%ymm7 + .byte 197,185,114,212,16 // vpsrld $0x10,%xmm4,%xmm8 + .byte 196,227,125,25,228,1 // vextractf128 $0x1,%ymm4,%xmm4 + .byte 197,217,114,212,16 // vpsrld $0x10,%xmm4,%xmm4 .byte 196,227,61,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm8,%ymm4 - .byte 196,227,125,25,234,1 // vextractf128 $0x1,%ymm5,%xmm2 - .byte 197,233,114,210,16 // vpsrld $0x10,%xmm2,%xmm2 - .byte 197,209,114,213,16 // vpsrld $0x10,%xmm5,%xmm5 - .byte 196,193,81,250,234 // vpsubd %xmm10,%xmm5,%xmm5 - .byte 196,193,105,250,210 // vpsubd %xmm10,%xmm2,%xmm2 - .byte 197,193,114,215,13 // vpsrld $0xd,%xmm7,%xmm7 - .byte 197,233,254,215 // vpaddd %xmm7,%xmm2,%xmm2 + .byte 196,193,92,86,226 // vorps %ymm10,%ymm4,%ymm4 .byte 197,201,114,214,13 // vpsrld $0xd,%xmm6,%xmm6 - .byte 197,209,254,238 // vpaddd %xmm6,%xmm5,%xmm5 - .byte 196,227,85,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm5,%ymm2 - .byte 196,227,109,74,209,64 // vblendvps %ymm4,%ymm1,%ymm2,%ymm2 - .byte 196,99,125,25,225,1 // vextractf128 $0x1,%ymm12,%xmm1 - .byte 196,226,25,43,201 // vpackusdw %xmm1,%xmm12,%xmm1 + .byte 196,227,125,25,225,1 // vextractf128 $0x1,%ymm4,%xmm1 + .byte 197,241,254,206 // vpaddd %xmm6,%xmm1,%xmm1 + .byte 197,209,114,213,13 // vpsrld $0xd,%xmm5,%xmm5 + .byte 197,217,254,229 // vpaddd %xmm5,%xmm4,%xmm4 + .byte 196,227,93,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 + .byte 196,227,117,74,200,112 // vblendvps %ymm7,%ymm0,%ymm1,%ymm1 + .byte 196,99,125,25,224,1 // vextractf128 $0x1,%ymm12,%xmm0 + .byte 196,226,25,43,192 // vpackusdw %xmm0,%xmm12,%xmm0 .byte 196,99,125,25,236,1 // vextractf128 $0x1,%ymm13,%xmm4 .byte 196,226,17,43,228 // vpackusdw %xmm4,%xmm13,%xmm4 .byte 196,99,125,25,221,1 // vextractf128 $0x1,%ymm11,%xmm5 - .byte 196,226,33,43,237 // vpackusdw %xmm5,%xmm11,%xmm5 - .byte 196,227,125,25,214,1 // vextractf128 $0x1,%ymm2,%xmm6 - .byte 196,226,105,43,214 // vpackusdw %xmm6,%xmm2,%xmm2 - .byte 197,241,97,244 // vpunpcklwd %xmm4,%xmm1,%xmm6 - .byte 197,241,105,204 // vpunpckhwd %xmm4,%xmm1,%xmm1 - .byte 197,209,97,226 // vpunpcklwd %xmm2,%xmm5,%xmm4 - .byte 197,209,105,210 // vpunpckhwd %xmm2,%xmm5,%xmm2 - .byte 197,73,98,220 // vpunpckldq %xmm4,%xmm6,%xmm11 - .byte 197,73,106,212 // vpunpckhdq %xmm4,%xmm6,%xmm10 - .byte 197,113,98,202 // vpunpckldq %xmm2,%xmm1,%xmm9 - .byte 197,113,106,194 // vpunpckhdq %xmm2,%xmm1,%xmm8 + .byte 196,226,33,43,245 // vpackusdw %xmm5,%xmm11,%xmm6 + .byte 196,227,125,25,205,1 // vextractf128 $0x1,%ymm1,%xmm5 + .byte 196,226,113,43,205 // vpackusdw %xmm5,%xmm1,%xmm1 + .byte 197,249,97,236 // vpunpcklwd %xmm4,%xmm0,%xmm5 + .byte 197,249,105,196 // vpunpckhwd %xmm4,%xmm0,%xmm0 + .byte 197,201,97,225 // vpunpcklwd %xmm1,%xmm6,%xmm4 + .byte 197,201,105,201 // vpunpckhwd %xmm1,%xmm6,%xmm1 + .byte 197,81,98,220 // vpunpckldq %xmm4,%xmm5,%xmm11 + .byte 197,81,106,212 // vpunpckhdq %xmm4,%xmm5,%xmm10 + .byte 197,121,98,201 // vpunpckldq %xmm1,%xmm0,%xmm9 + .byte 197,121,106,193 // vpunpckhdq %xmm1,%xmm0,%xmm8 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 117,70 // jne 4a6e <_sk_store_f16_avx+0x25f> - .byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8) - .byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8) - .byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8) - .byte 196,65,122,127,68,248,48 // vmovdqu %xmm8,0x30(%r8,%rdi,8) - .byte 72,173 // lods %ds:(%rsi),%rax - .byte 197,252,16,76,36,128 // vmovups -0x80(%rsp),%ymm1 - .byte 197,252,16,84,36,160 // vmovups -0x60(%rsp),%ymm2 + .byte 117,66 // jne 48bc <_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) + .byte 197,122,127,68,248,48 // vmovdqu %xmm8,0x30(%rax,%rdi,8) + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0 + .byte 197,252,16,76,36,160 // vmovups -0x60(%rsp),%ymm1 .byte 197,252,16,100,36,192 // vmovups -0x40(%rsp),%ymm4 .byte 197,252,16,108,36,224 // vmovups -0x20(%rsp),%ymm5 .byte 197,252,16,52,36 // vmovups (%rsp),%ymm6 .byte 197,252,16,124,36,32 // vmovups 0x20(%rsp),%ymm7 .byte 72,131,196,88 // add $0x58,%rsp .byte 255,224 // jmpq *%rax - .byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8) + .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,201 // je 4a43 <_sk_store_f16_avx+0x234> - .byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8) + .byte 116,202 // je 4891 <_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 114,188 // jb 4a43 <_sk_store_f16_avx+0x234> - .byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8) - .byte 116,179 // je 4a43 <_sk_store_f16_avx+0x234> - .byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8) + .byte 114,190 // jb 4891 <_sk_store_f16_avx+0x233> + .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) + .byte 116,182 // je 4891 <_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 114,166 // jb 4a43 <_sk_store_f16_avx+0x234> - .byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8) - .byte 116,157 // je 4a43 <_sk_store_f16_avx+0x234> - .byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8) + .byte 114,170 // jb 4891 <_sk_store_f16_avx+0x233> + .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) + .byte 116,162 // je 4891 <_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 114,144 // jb 4a43 <_sk_store_f16_avx+0x234> - .byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8) - .byte 235,135 // jmp 4a43 <_sk_store_f16_avx+0x234> + .byte 114,150 // jb 4891 <_sk_store_f16_avx+0x233> + .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) + .byte 235,142 // jmp 4891 <_sk_store_f16_avx+0x233> HIDDEN _sk_load_u16_be_avx .globl _sk_load_u16_be_avx @@ -16389,7 +16241,7 @@ _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 15,133,5,1,0,0 // jne 4bd7 <_sk_load_u16_be_avx+0x11b> + .byte 15,133,5,1,0,0 // jne 4a1e <_sk_load_u16_be_avx+0x11b> .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 @@ -16448,29 +16300,29 @@ _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 4c3d <_sk_load_u16_be_avx+0x181> + .byte 116,85 // je 4a84 <_sk_load_u16_be_avx+0x181> .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 4c3d <_sk_load_u16_be_avx+0x181> + .byte 114,72 // jb 4a84 <_sk_load_u16_be_avx+0x181> .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 4c4a <_sk_load_u16_be_avx+0x18e> + .byte 116,72 // je 4a91 <_sk_load_u16_be_avx+0x18e> .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 4c4a <_sk_load_u16_be_avx+0x18e> + .byte 114,59 // jb 4a91 <_sk_load_u16_be_avx+0x18e> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,205,254,255,255 // je 4aed <_sk_load_u16_be_avx+0x31> + .byte 15,132,205,254,255,255 // je 4934 <_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 15,130,188,254,255,255 // jb 4aed <_sk_load_u16_be_avx+0x31> + .byte 15,130,188,254,255,255 // jb 4934 <_sk_load_u16_be_avx+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,176,254,255,255 // jmpq 4aed <_sk_load_u16_be_avx+0x31> + .byte 233,176,254,255,255 // jmpq 4934 <_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 233,163,254,255,255 // jmpq 4aed <_sk_load_u16_be_avx+0x31> + .byte 233,163,254,255,255 // jmpq 4934 <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,154,254,255,255 // jmpq 4aed <_sk_load_u16_be_avx+0x31> + .byte 233,154,254,255,255 // jmpq 4934 <_sk_load_u16_be_avx+0x31> HIDDEN _sk_load_rgb_u16_be_avx .globl _sk_load_rgb_u16_be_avx @@ -16480,7 +16332,7 @@ _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 15,133,8,1,0,0 // jne 4d6d <_sk_load_rgb_u16_be_avx+0x11a> + .byte 15,133,8,1,0,0 // jne 4bb4 <_sk_load_rgb_u16_be_avx+0x11a> .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 @@ -16539,36 +16391,36 @@ _sk_load_rgb_u16_be_avx: .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 4d86 <_sk_load_rgb_u16_be_avx+0x133> - .byte 233,19,255,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,5 // jne 4bcd <_sk_load_rgb_u16_be_avx+0x133> + .byte 233,19,255,255,255 // jmpq 4ae0 <_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 4db5 <_sk_load_rgb_u16_be_avx+0x162> + .byte 114,26 // jb 4bfc <_sk_load_rgb_u16_be_avx+0x162> .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 4dba <_sk_load_rgb_u16_be_avx+0x167> - .byte 233,228,254,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,223,254,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 4c01 <_sk_load_rgb_u16_be_avx+0x167> + .byte 233,228,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,223,254,255,255 // jmpq 4ae0 <_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 4de9 <_sk_load_rgb_u16_be_avx+0x196> + .byte 114,26 // jb 4c30 <_sk_load_rgb_u16_be_avx+0x196> .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 4dee <_sk_load_rgb_u16_be_avx+0x19b> - .byte 233,176,254,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,171,254,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 4c35 <_sk_load_rgb_u16_be_avx+0x19b> + .byte 233,176,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,171,254,255,255 // jmpq 4ae0 <_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 4e17 <_sk_load_rgb_u16_be_avx+0x1c4> + .byte 114,20 // jb 4c5e <_sk_load_rgb_u16_be_avx+0x1c4> .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 233,130,254,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,125,254,255,255 // jmpq 4c99 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,130,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,125,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46> HIDDEN _sk_store_u16_be_avx .globl _sk_store_u16_be_avx @@ -16618,7 +16470,7 @@ _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 4f1e <_sk_store_u16_be_avx+0x102> + .byte 117,31 // jne 4d65 <_sk_store_u16_be_avx+0x102> .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2) .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2) .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2) @@ -16627,22 +16479,22 @@ _sk_store_u16_be_avx: .byte 255,224 // jmpq *%rax .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 116,240 // je 4d61 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 114,227 // jb 4d61 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2) - .byte 116,218 // je 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 116,218 // je 4d61 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 114,205 // jb 4d61 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2) - .byte 116,196 // je 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 116,196 // je 4d61 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 114,183 // jb 4d61 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2) - .byte 235,174 // jmp 4f1a <_sk_store_u16_be_avx+0xfe> + .byte 235,174 // jmp 4d61 <_sk_store_u16_be_avx+0xfe> HIDDEN _sk_load_f32_avx .globl _sk_load_f32_avx @@ -16650,10 +16502,10 @@ 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 4fe2 <_sk_load_f32_avx+0x76> + .byte 119,110 // ja 4e29 <_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 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 500c <_sk_load_f32_avx+0xa0> + .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 4e54 <_sk_load_f32_avx+0xa1> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -16679,19 +16531,21 @@ _sk_load_f32_avx: .byte 196,193,101,21,216 // vunpckhpd %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax - .byte 102,144 // xchg %ax,%ax - .byte 131,255,255 // cmp $0xffffffff,%edi - .byte 255,202 // dec %edx + .byte 15,31,0 // nopl (%rax) + .byte 130 // (bad) .byte 255 // (bad) .byte 255 // (bad) + .byte 255,201 // dec %ecx .byte 255 // (bad) - .byte 189,255,255,255,176 // mov $0xb0ffffff,%ebp .byte 255 // (bad) .byte 255 // (bad) - .byte 255,163,255,255,255,155 // jmpq *-0x64000001(%rbx) + .byte 188,255,255,255,175 // mov $0xafffffff,%esp .byte 255 // (bad) .byte 255 // (bad) - .byte 255,147,255,255,255,139 // callq *-0x74000001(%rbx) + .byte 255,162,255,255,255,154 // jmpq *-0x65000001(%rdx) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,146,255,255,255,138 // callq *-0x75000001(%rdx) .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -16712,7 +16566,7 @@ _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 5099 <_sk_store_f32_avx+0x6d> + .byte 117,55 // jne 4ee1 <_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 @@ -16725,22 +16579,22 @@ _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 5095 <_sk_store_f32_avx+0x69> + .byte 116,240 // je 4edd <_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 5095 <_sk_store_f32_avx+0x69> + .byte 114,227 // jb 4edd <_sk_store_f32_avx+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 5095 <_sk_store_f32_avx+0x69> + .byte 116,218 // je 4edd <_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 5095 <_sk_store_f32_avx+0x69> + .byte 114,205 // jb 4edd <_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 5095 <_sk_store_f32_avx+0x69> + .byte 116,195 // je 4edd <_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 5095 <_sk_store_f32_avx+0x69> + .byte 114,181 // jb 4edd <_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 5095 <_sk_store_f32_avx+0x69> + .byte 235,171 // jmp 4edd <_sk_store_f32_avx+0x69> HIDDEN _sk_clamp_x_avx .globl _sk_clamp_x_avx @@ -17068,7 +16922,7 @@ _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 564d <_sk_linear_gradient_avx+0xb8> + .byte 15,132,146,0,0,0 // je 5495 <_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 @@ -17095,8 +16949,8 @@ _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 55d7 <_sk_linear_gradient_avx+0x42> - .byte 235,20 // jmp 5661 <_sk_linear_gradient_avx+0xcc> + .byte 117,140 // jne 541f <_sk_linear_gradient_avx+0x42> + .byte 235,20 // jmp 54a9 <_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 @@ -17590,8 +17444,88 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) - .byte 128 // .byte 0x80 - .byte 63 // (bad) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 5c60 <.literal4+0x10> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 5cd9 <.literal4+0x89> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 5c90 <.literal4+0x40> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 5d09 <.literal4+0xb9> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 5cc0 <.literal4+0x70> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 5d39 <.literal4+0xe9> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 5cf0 <.literal4+0xa0> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 5d69 <_sk_callback_avx+0x23f> + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,249,68,180 // mov $0xb444f93f,%edi + .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9 + .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 128,0,0 // addb $0x0,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .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) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,128,56 // add %al,0x38800000(%rax) + .byte 0,64,254 // add %al,-0x2(%rax) + .byte 255 // .byte 0xff BALIGN32 HIDDEN _sk_start_pipeline_sse41 @@ -17652,7 +17586,7 @@ _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 15,40,21,132,64,0,0 // movaps 0x4084(%rip),%xmm2 # 4100 <_sk_callback_sse41+0xe2> + .byte 15,40,21,196,62,0,0 // movaps 0x3ec4(%rip),%xmm2 # 3f40 <_sk_callback_sse41+0xe0> .byte 15,88,202 // addps %xmm2,%xmm1 .byte 15,16,2 // movups (%rdx),%xmm0 .byte 15,88,193 // addps %xmm1,%xmm0 @@ -17661,7 +17595,7 @@ _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 15,40,21,115,64,0,0 // movaps 0x4073(%rip),%xmm2 # 4110 <_sk_callback_sse41+0xf2> + .byte 15,40,21,179,62,0,0 // movaps 0x3eb3(%rip),%xmm2 # 3f50 <_sk_callback_sse41+0xf0> .byte 15,87,219 // xorps %xmm3,%xmm3 .byte 15,87,228 // xorps %xmm4,%xmm4 .byte 15,87,237 // xorps %xmm5,%xmm5 @@ -19743,74 +19677,44 @@ _sk_parametric_r_sse41: .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 68,15,89,200 // mulps %xmm0,%xmm9 - .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 68,15,89,216 // mulps %xmm0,%xmm11 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,208 // mulps %xmm0,%xmm10 .byte 65,15,194,192,2 // cmpleps %xmm8,%xmm0 .byte 243,68,15,16,64,24 // movss 0x18(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 69,15,88,200 // addps %xmm8,%xmm9 - .byte 243,68,15,16,16 // movss (%rax),%xmm10 + .byte 243,68,15,16,24 // movss (%rax),%xmm11 .byte 243,68,15,16,64,8 // movss 0x8(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 69,15,88,216 // addps %xmm8,%xmm11 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,193 // movd %ecx,%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 69,15,89,196 // mulps %xmm12,%xmm8 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 102,69,15,112,228,0 // pshufd $0x0,%xmm12,%xmm12 - .byte 102,69,15,219,227 // pand %xmm11,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,92,196 // subps %xmm12,%xmm8 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,227 // mulps %xmm11,%xmm12 + .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 68,15,89,5,149,33,0,0 // mulps 0x2195(%rip),%xmm8 # 3f60 <_sk_callback_sse41+0x100> + .byte 68,15,84,21,157,33,0,0 // andps 0x219d(%rip),%xmm10 # 3f70 <_sk_callback_sse41+0x110> + .byte 68,15,86,21,165,33,0,0 // orps 0x21a5(%rip),%xmm10 # 3f80 <_sk_callback_sse41+0x120> + .byte 68,15,88,5,173,33,0,0 // addps 0x21ad(%rip),%xmm8 # 3f90 <_sk_callback_sse41+0x130> + .byte 68,15,40,37,181,33,0,0 // movaps 0x21b5(%rip),%xmm12 # 3fa0 <_sk_callback_sse41+0x140> + .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,196 // subps %xmm12,%xmm8 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,88,235 // addps %xmm11,%xmm13 - .byte 69,15,94,229 // divps %xmm13,%xmm12 + .byte 68,15,88,21,181,33,0,0 // addps 0x21b5(%rip),%xmm10 # 3fb0 <_sk_callback_sse41+0x150> + .byte 68,15,40,37,189,33,0,0 // movaps 0x21bd(%rip),%xmm12 # 3fc0 <_sk_callback_sse41+0x160> + .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,196 // subps %xmm12,%xmm8 - .byte 69,15,89,194 // mulps %xmm10,%xmm8 - .byte 102,69,15,58,8,216,1 // roundps $0x1,%xmm8,%xmm11 - .byte 185,0,0,0,75 // mov $0x4b000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,224 // addps %xmm8,%xmm12 - .byte 69,15,92,195 // subps %xmm11,%xmm8 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,89,216 // mulps %xmm8,%xmm11 - .byte 69,15,92,227 // subps %xmm11,%xmm12 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,92,232 // subps %xmm8,%xmm13 - .byte 69,15,94,221 // divps %xmm13,%xmm11 - .byte 69,15,88,220 // addps %xmm12,%xmm11 - .byte 69,15,89,218 // mulps %xmm10,%xmm11 + .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 68,15,88,5,170,33,0,0 // addps 0x21aa(%rip),%xmm8 # 3fd0 <_sk_callback_sse41+0x170> + .byte 68,15,40,21,178,33,0,0 // movaps 0x21b2(%rip),%xmm10 # 3fe0 <_sk_callback_sse41+0x180> + .byte 69,15,89,211 // mulps %xmm11,%xmm10 + .byte 69,15,92,194 // subps %xmm10,%xmm8 + .byte 68,15,40,21,178,33,0,0 // movaps 0x21b2(%rip),%xmm10 # 3ff0 <_sk_callback_sse41+0x190> + .byte 69,15,92,211 // subps %xmm11,%xmm10 + .byte 68,15,40,29,182,33,0,0 // movaps 0x21b6(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x1a0> + .byte 69,15,94,218 // divps %xmm10,%xmm11 + .byte 69,15,88,216 // addps %xmm8,%xmm11 + .byte 68,15,89,29,182,33,0,0 // mulps 0x21b6(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x1b0> .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 @@ -19851,62 +19755,32 @@ _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 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 102,68,15,112,225,0 // pshufd $0x0,%xmm1,%xmm12 - .byte 102,69,15,219,227 // pand %xmm11,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 102,68,15,112,217,0 // pshufd $0x0,%xmm1,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 68,15,92,233 // subps %xmm1,%xmm13 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 + .byte 68,15,89,37,37,33,0,0 // mulps 0x2125(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x1c0> + .byte 68,15,84,29,45,33,0,0 // andps 0x212d(%rip),%xmm11 # 4030 <_sk_callback_sse41+0x1d0> + .byte 68,15,86,29,53,33,0,0 // orps 0x2135(%rip),%xmm11 # 4040 <_sk_callback_sse41+0x1e0> + .byte 68,15,88,37,61,33,0,0 // addps 0x213d(%rip),%xmm12 # 4050 <_sk_callback_sse41+0x1f0> + .byte 15,40,13,70,33,0,0 // movaps 0x2146(%rip),%xmm1 # 4060 <_sk_callback_sse41+0x200> .byte 65,15,89,203 // mulps %xmm11,%xmm1 - .byte 68,15,92,233 // subps %xmm1,%xmm13 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 65,15,88,203 // addps %xmm11,%xmm1 - .byte 68,15,94,225 // divps %xmm1,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 - .byte 185,0,0,0,75 // mov $0x4b000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 - .byte 69,15,92,235 // subps %xmm11,%xmm13 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 65,15,89,205 // mulps %xmm13,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 65,15,92,205 // subps %xmm13,%xmm1 - .byte 68,15,94,217 // divps %xmm1,%xmm11 - .byte 69,15,88,220 // addps %xmm12,%xmm11 - .byte 69,15,89,218 // mulps %xmm10,%xmm11 - .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 + .byte 68,15,88,29,70,33,0,0 // addps 0x2146(%rip),%xmm11 # 4070 <_sk_callback_sse41+0x210> + .byte 15,40,13,79,33,0,0 // movaps 0x214f(%rip),%xmm1 # 4080 <_sk_callback_sse41+0x220> + .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 68,15,88,37,60,33,0,0 // addps 0x213c(%rip),%xmm12 # 4090 <_sk_callback_sse41+0x230> + .byte 15,40,13,69,33,0,0 // movaps 0x2145(%rip),%xmm1 # 40a0 <_sk_callback_sse41+0x240> + .byte 65,15,89,203 // mulps %xmm11,%xmm1 + .byte 68,15,92,225 // subps %xmm1,%xmm12 + .byte 68,15,40,21,69,33,0,0 // movaps 0x2145(%rip),%xmm10 # 40b0 <_sk_callback_sse41+0x250> + .byte 69,15,92,211 // subps %xmm11,%xmm10 + .byte 15,40,13,74,33,0,0 // movaps 0x214a(%rip),%xmm1 # 40c0 <_sk_callback_sse41+0x260> + .byte 65,15,94,202 // divps %xmm10,%xmm1 + .byte 65,15,88,204 // addps %xmm12,%xmm1 + .byte 15,89,13,75,33,0,0 // mulps 0x214b(%rip),%xmm1 # 40d0 <_sk_callback_sse41+0x270> + .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 .byte 65,15,88,202 // addps %xmm10,%xmm1 @@ -19946,62 +19820,32 @@ _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 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 102,68,15,112,226,0 // pshufd $0x0,%xmm2,%xmm12 - .byte 102,69,15,219,227 // pand %xmm11,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 68,15,92,234 // subps %xmm2,%xmm13 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 + .byte 68,15,89,37,190,32,0,0 // mulps 0x20be(%rip),%xmm12 # 40e0 <_sk_callback_sse41+0x280> + .byte 68,15,84,29,198,32,0,0 // andps 0x20c6(%rip),%xmm11 # 40f0 <_sk_callback_sse41+0x290> + .byte 68,15,86,29,206,32,0,0 // orps 0x20ce(%rip),%xmm11 # 4100 <_sk_callback_sse41+0x2a0> + .byte 68,15,88,37,214,32,0,0 // addps 0x20d6(%rip),%xmm12 # 4110 <_sk_callback_sse41+0x2b0> + .byte 15,40,21,223,32,0,0 // movaps 0x20df(%rip),%xmm2 # 4120 <_sk_callback_sse41+0x2c0> .byte 65,15,89,211 // mulps %xmm11,%xmm2 - .byte 68,15,92,234 // subps %xmm2,%xmm13 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 65,15,88,211 // addps %xmm11,%xmm2 - .byte 68,15,94,226 // divps %xmm2,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 - .byte 185,0,0,0,75 // mov $0x4b000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 - .byte 69,15,92,235 // subps %xmm11,%xmm13 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 65,15,89,213 // mulps %xmm13,%xmm2 .byte 68,15,92,226 // subps %xmm2,%xmm12 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 65,15,92,213 // subps %xmm13,%xmm2 - .byte 68,15,94,218 // divps %xmm2,%xmm11 - .byte 69,15,88,220 // addps %xmm12,%xmm11 - .byte 69,15,89,218 // mulps %xmm10,%xmm11 - .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 + .byte 68,15,88,29,223,32,0,0 // addps 0x20df(%rip),%xmm11 # 4130 <_sk_callback_sse41+0x2d0> + .byte 15,40,21,232,32,0,0 // movaps 0x20e8(%rip),%xmm2 # 4140 <_sk_callback_sse41+0x2e0> + .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 68,15,88,37,213,32,0,0 // addps 0x20d5(%rip),%xmm12 # 4150 <_sk_callback_sse41+0x2f0> + .byte 15,40,21,222,32,0,0 // movaps 0x20de(%rip),%xmm2 # 4160 <_sk_callback_sse41+0x300> + .byte 65,15,89,211 // mulps %xmm11,%xmm2 + .byte 68,15,92,226 // subps %xmm2,%xmm12 + .byte 68,15,40,21,222,32,0,0 // movaps 0x20de(%rip),%xmm10 # 4170 <_sk_callback_sse41+0x310> + .byte 69,15,92,211 // subps %xmm11,%xmm10 + .byte 15,40,21,227,32,0,0 // movaps 0x20e3(%rip),%xmm2 # 4180 <_sk_callback_sse41+0x320> + .byte 65,15,94,210 // divps %xmm10,%xmm2 + .byte 65,15,88,212 // addps %xmm12,%xmm2 + .byte 15,89,21,228,32,0,0 // mulps 0x20e4(%rip),%xmm2 # 4190 <_sk_callback_sse41+0x330> + .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 .byte 65,15,88,210 // addps %xmm10,%xmm2 @@ -20041,62 +19885,32 @@ _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 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 102,68,15,112,227,0 // pshufd $0x0,%xmm3,%xmm12 - .byte 102,69,15,219,227 // pand %xmm11,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 102,68,15,112,219,0 // pshufd $0x0,%xmm3,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 68,15,92,235 // subps %xmm3,%xmm13 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 + .byte 68,15,89,37,87,32,0,0 // mulps 0x2057(%rip),%xmm12 # 41a0 <_sk_callback_sse41+0x340> + .byte 68,15,84,29,95,32,0,0 // andps 0x205f(%rip),%xmm11 # 41b0 <_sk_callback_sse41+0x350> + .byte 68,15,86,29,103,32,0,0 // orps 0x2067(%rip),%xmm11 # 41c0 <_sk_callback_sse41+0x360> + .byte 68,15,88,37,111,32,0,0 // addps 0x206f(%rip),%xmm12 # 41d0 <_sk_callback_sse41+0x370> + .byte 15,40,29,120,32,0,0 // movaps 0x2078(%rip),%xmm3 # 41e0 <_sk_callback_sse41+0x380> .byte 65,15,89,219 // mulps %xmm11,%xmm3 - .byte 68,15,92,235 // subps %xmm3,%xmm13 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 65,15,88,219 // addps %xmm11,%xmm3 - .byte 68,15,94,227 // divps %xmm3,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 - .byte 185,0,0,0,75 // mov $0x4b000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 - .byte 69,15,92,235 // subps %xmm11,%xmm13 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 65,15,89,221 // mulps %xmm13,%xmm3 .byte 68,15,92,227 // subps %xmm3,%xmm12 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 65,15,92,221 // subps %xmm13,%xmm3 - .byte 68,15,94,219 // divps %xmm3,%xmm11 - .byte 69,15,88,220 // addps %xmm12,%xmm11 - .byte 69,15,89,218 // mulps %xmm10,%xmm11 - .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 + .byte 68,15,88,29,120,32,0,0 // addps 0x2078(%rip),%xmm11 # 41f0 <_sk_callback_sse41+0x390> + .byte 15,40,29,129,32,0,0 // movaps 0x2081(%rip),%xmm3 # 4200 <_sk_callback_sse41+0x3a0> + .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 68,15,88,37,110,32,0,0 // addps 0x206e(%rip),%xmm12 # 4210 <_sk_callback_sse41+0x3b0> + .byte 15,40,29,119,32,0,0 // movaps 0x2077(%rip),%xmm3 # 4220 <_sk_callback_sse41+0x3c0> + .byte 65,15,89,219 // mulps %xmm11,%xmm3 + .byte 68,15,92,227 // subps %xmm3,%xmm12 + .byte 68,15,40,21,119,32,0,0 // movaps 0x2077(%rip),%xmm10 # 4230 <_sk_callback_sse41+0x3d0> + .byte 69,15,92,211 // subps %xmm11,%xmm10 + .byte 15,40,29,124,32,0,0 // movaps 0x207c(%rip),%xmm3 # 4240 <_sk_callback_sse41+0x3e0> + .byte 65,15,94,218 // divps %xmm10,%xmm3 + .byte 65,15,88,220 // addps %xmm12,%xmm3 + .byte 15,89,29,125,32,0,0 // mulps 0x207d(%rip),%xmm3 # 4250 <_sk_callback_sse41+0x3f0> + .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 .byte 65,15,88,218 // addps %xmm10,%xmm3 @@ -20329,9 +20143,9 @@ _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 26b5 <_sk_gather_i8_sse41+0xf> + .byte 116,5 // je 250a <_sk_gather_i8_sse41+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 26b7 <_sk_gather_i8_sse41+0x11> + .byte 235,2 // jmp 250c <_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 @@ -20777,59 +20591,57 @@ _sk_load_f16_sse41: .byte 102,69,15,111,217 // movdqa %xmm9,%xmm11 .byte 102,68,15,97,216 // punpcklwd %xmm0,%xmm11 .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9 - .byte 102,69,15,56,51,227 // pmovzxwd %xmm11,%xmm12 - .byte 184,0,128,0,0 // mov $0x8000,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,68,15,112,192,0 // pshufd $0x0,%xmm0,%xmm8 - .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 + .byte 102,65,15,56,51,203 // pmovzxwd %xmm11,%xmm1 + .byte 102,68,15,111,5,152,22,0,0 // movdqa 0x1698(%rip),%xmm8 # 4260 <_sk_callback_sse41+0x400> + .byte 102,15,111,209 // movdqa %xmm1,%xmm2 .byte 102,65,15,219,208 // pand %xmm8,%xmm2 - .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 - .byte 184,0,4,0,0 // mov $0x400,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 + .byte 102,15,239,202 // pxor %xmm2,%xmm1 + .byte 102,15,111,29,147,22,0,0 // movdqa 0x1693(%rip),%xmm3 # 4270 <_sk_callback_sse41+0x410> .byte 102,15,114,242,16 // pslld $0x10,%xmm2 - .byte 102,15,111,195 // movdqa %xmm3,%xmm0 - .byte 102,65,15,102,196 // pcmpgtd %xmm12,%xmm0 - .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 - .byte 184,0,0,0,56 // mov $0x38000000,%eax - .byte 102,15,110,200 // movd %eax,%xmm1 - .byte 102,68,15,112,209,0 // pshufd $0x0,%xmm1,%xmm10 - .byte 102,65,15,254,210 // paddd %xmm10,%xmm2 - .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 - .byte 102,15,223,194 // pandn %xmm2,%xmm0 + .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 102,68,15,111,21,127,22,0,0 // movdqa 0x167f(%rip),%xmm10 # 4280 <_sk_callback_sse41+0x420> + .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 .byte 102,69,15,56,51,219 // pmovzxwd %xmm11,%xmm11 .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2 .byte 102,65,15,219,208 // pand %xmm8,%xmm2 .byte 102,68,15,239,218 // pxor %xmm2,%xmm11 .byte 102,15,114,242,16 // pslld $0x10,%xmm2 - .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,65,15,102,203 // pcmpgtd %xmm11,%xmm1 + .byte 102,65,15,111,203 // movdqa %xmm11,%xmm1 + .byte 102,15,56,63,203 // pmaxud %xmm3,%xmm1 + .byte 102,65,15,118,203 // pcmpeqd %xmm11,%xmm1 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11 - .byte 102,65,15,254,210 // paddd %xmm10,%xmm2 - .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 - .byte 102,15,223,202 // pandn %xmm2,%xmm1 + .byte 102,68,15,235,218 // por %xmm2,%xmm11 + .byte 102,69,15,254,218 // paddd %xmm10,%xmm11 + .byte 102,65,15,219,203 // pand %xmm11,%xmm1 .byte 102,69,15,56,51,217 // pmovzxwd %xmm9,%xmm11 .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12 .byte 102,69,15,219,224 // pand %xmm8,%xmm12 .byte 102,69,15,239,220 // pxor %xmm12,%xmm11 .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12 - .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,65,15,102,211 // pcmpgtd %xmm11,%xmm2 + .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2 + .byte 102,15,56,63,211 // pmaxud %xmm3,%xmm2 + .byte 102,65,15,118,211 // pcmpeqd %xmm11,%xmm2 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11 - .byte 102,69,15,254,226 // paddd %xmm10,%xmm12 - .byte 102,69,15,254,227 // paddd %xmm11,%xmm12 - .byte 102,65,15,223,212 // pandn %xmm12,%xmm2 + .byte 102,69,15,235,220 // por %xmm12,%xmm11 + .byte 102,69,15,254,218 // paddd %xmm10,%xmm11 + .byte 102,65,15,219,211 // pand %xmm11,%xmm2 .byte 102,65,15,115,217,8 // psrldq $0x8,%xmm9 .byte 102,69,15,56,51,201 // pmovzxwd %xmm9,%xmm9 .byte 102,69,15,219,193 // pand %xmm9,%xmm8 .byte 102,69,15,239,200 // pxor %xmm8,%xmm9 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 - .byte 102,65,15,102,217 // pcmpgtd %xmm9,%xmm3 + .byte 102,65,15,56,63,217 // pmaxud %xmm9,%xmm3 + .byte 102,65,15,118,217 // pcmpeqd %xmm9,%xmm3 .byte 102,65,15,114,241,13 // pslld $0xd,%xmm9 - .byte 102,69,15,254,194 // paddd %xmm10,%xmm8 - .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 - .byte 102,65,15,223,216 // pandn %xmm8,%xmm3 + .byte 102,69,15,235,200 // por %xmm8,%xmm9 + .byte 102,69,15,254,202 // paddd %xmm10,%xmm9 + .byte 102,65,15,219,217 // pand %xmm9,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20863,59 +20675,57 @@ _sk_gather_f16_sse41: .byte 102,69,15,111,217 // movdqa %xmm9,%xmm11 .byte 102,68,15,97,218 // punpcklwd %xmm2,%xmm11 .byte 102,68,15,105,202 // punpckhwd %xmm2,%xmm9 - .byte 102,69,15,56,51,227 // pmovzxwd %xmm11,%xmm12 - .byte 184,0,128,0,0 // mov $0x8000,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,68,15,112,192,0 // pshufd $0x0,%xmm0,%xmm8 - .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 + .byte 102,65,15,56,51,203 // pmovzxwd %xmm11,%xmm1 + .byte 102,68,15,111,5,61,21,0,0 // movdqa 0x153d(%rip),%xmm8 # 4290 <_sk_callback_sse41+0x430> + .byte 102,15,111,209 // movdqa %xmm1,%xmm2 .byte 102,65,15,219,208 // pand %xmm8,%xmm2 - .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 - .byte 184,0,4,0,0 // mov $0x400,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 + .byte 102,15,239,202 // pxor %xmm2,%xmm1 + .byte 102,15,111,29,56,21,0,0 // movdqa 0x1538(%rip),%xmm3 # 42a0 <_sk_callback_sse41+0x440> .byte 102,15,114,242,16 // pslld $0x10,%xmm2 - .byte 102,15,111,195 // movdqa %xmm3,%xmm0 - .byte 102,65,15,102,196 // pcmpgtd %xmm12,%xmm0 - .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 - .byte 184,0,0,0,56 // mov $0x38000000,%eax - .byte 102,15,110,200 // movd %eax,%xmm1 - .byte 102,68,15,112,209,0 // pshufd $0x0,%xmm1,%xmm10 - .byte 102,65,15,254,210 // paddd %xmm10,%xmm2 - .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 - .byte 102,15,223,194 // pandn %xmm2,%xmm0 + .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 102,68,15,111,21,36,21,0,0 // movdqa 0x1524(%rip),%xmm10 # 42b0 <_sk_callback_sse41+0x450> + .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 .byte 102,69,15,56,51,219 // pmovzxwd %xmm11,%xmm11 .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2 .byte 102,65,15,219,208 // pand %xmm8,%xmm2 .byte 102,68,15,239,218 // pxor %xmm2,%xmm11 .byte 102,15,114,242,16 // pslld $0x10,%xmm2 - .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,65,15,102,203 // pcmpgtd %xmm11,%xmm1 + .byte 102,65,15,111,203 // movdqa %xmm11,%xmm1 + .byte 102,15,56,63,203 // pmaxud %xmm3,%xmm1 + .byte 102,65,15,118,203 // pcmpeqd %xmm11,%xmm1 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11 - .byte 102,65,15,254,210 // paddd %xmm10,%xmm2 - .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 - .byte 102,15,223,202 // pandn %xmm2,%xmm1 + .byte 102,68,15,235,218 // por %xmm2,%xmm11 + .byte 102,69,15,254,218 // paddd %xmm10,%xmm11 + .byte 102,65,15,219,203 // pand %xmm11,%xmm1 .byte 102,69,15,56,51,217 // pmovzxwd %xmm9,%xmm11 .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12 .byte 102,69,15,219,224 // pand %xmm8,%xmm12 .byte 102,69,15,239,220 // pxor %xmm12,%xmm11 .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12 - .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,65,15,102,211 // pcmpgtd %xmm11,%xmm2 + .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2 + .byte 102,15,56,63,211 // pmaxud %xmm3,%xmm2 + .byte 102,65,15,118,211 // pcmpeqd %xmm11,%xmm2 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11 - .byte 102,69,15,254,226 // paddd %xmm10,%xmm12 - .byte 102,69,15,254,227 // paddd %xmm11,%xmm12 - .byte 102,65,15,223,212 // pandn %xmm12,%xmm2 + .byte 102,69,15,235,220 // por %xmm12,%xmm11 + .byte 102,69,15,254,218 // paddd %xmm10,%xmm11 + .byte 102,65,15,219,211 // pand %xmm11,%xmm2 .byte 102,65,15,115,217,8 // psrldq $0x8,%xmm9 .byte 102,69,15,56,51,201 // pmovzxwd %xmm9,%xmm9 .byte 102,69,15,219,193 // pand %xmm9,%xmm8 .byte 102,69,15,239,200 // pxor %xmm8,%xmm9 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 - .byte 102,65,15,102,217 // pcmpgtd %xmm9,%xmm3 + .byte 102,65,15,56,63,217 // pmaxud %xmm9,%xmm3 + .byte 102,65,15,118,217 // pcmpeqd %xmm9,%xmm3 .byte 102,65,15,114,241,13 // pslld $0xd,%xmm9 - .byte 102,69,15,254,194 // paddd %xmm10,%xmm8 - .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 - .byte 102,65,15,223,216 // pandn %xmm8,%xmm3 + .byte 102,69,15,235,200 // por %xmm8,%xmm9 + .byte 102,69,15,254,202 // paddd %xmm10,%xmm9 + .byte 102,65,15,219,217 // pand %xmm9,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20925,63 +20735,57 @@ FUNCTION(_sk_store_f16_sse41) _sk_store_f16_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 185,0,0,0,128 // mov $0x80000000,%ecx - .byte 102,68,15,110,193 // movd %ecx,%xmm8 - .byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9 - .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 - .byte 102,68,15,219,224 // pand %xmm0,%xmm12 + .byte 102,68,15,111,21,90,20,0,0 // movdqa 0x145a(%rip),%xmm10 # 42c0 <_sk_callback_sse41+0x460> + .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13 - .byte 102,69,15,239,236 // pxor %xmm12,%xmm13 - .byte 185,0,0,128,56 // mov $0x38800000,%ecx - .byte 102,68,15,110,193 // movd %ecx,%xmm8 - .byte 102,69,15,112,208,0 // pshufd $0x0,%xmm8,%xmm10 - .byte 102,65,15,114,212,16 // psrld $0x10,%xmm12 - .byte 102,69,15,111,194 // movdqa %xmm10,%xmm8 - .byte 102,69,15,102,197 // pcmpgtd %xmm13,%xmm8 - .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 - .byte 185,0,192,1,0 // mov $0x1c000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 - .byte 102,69,15,250,227 // psubd %xmm11,%xmm12 - .byte 102,69,15,254,229 // paddd %xmm13,%xmm12 - .byte 102,69,15,223,196 // pandn %xmm12,%xmm8 + .byte 102,69,15,219,234 // pand %xmm10,%xmm13 + .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 + .byte 102,68,15,111,13,77,20,0,0 // movdqa 0x144d(%rip),%xmm9 # 42d0 <_sk_callback_sse41+0x470> + .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 102,68,15,111,29,62,20,0,0 // movdqa 0x143e(%rip),%xmm11 # 42e0 <_sk_callback_sse41+0x480> + .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 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 - .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13 - .byte 102,68,15,219,233 // pand %xmm1,%xmm13 + .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 .byte 102,68,15,111,241 // movdqa %xmm1,%xmm14 - .byte 102,69,15,239,245 // pxor %xmm13,%xmm14 - .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13 - .byte 102,69,15,111,226 // movdqa %xmm10,%xmm12 - .byte 102,69,15,102,230 // pcmpgtd %xmm14,%xmm12 - .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 - .byte 102,69,15,250,235 // psubd %xmm11,%xmm13 - .byte 102,69,15,254,238 // paddd %xmm14,%xmm13 - .byte 102,69,15,223,229 // pandn %xmm13,%xmm12 + .byte 102,69,15,219,242 // pand %xmm10,%xmm14 + .byte 102,69,15,239,238 // pxor %xmm14,%xmm13 + .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14 + .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 + .byte 102,69,15,102,229 // pcmpgtd %xmm13,%xmm12 + .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 + .byte 102,69,15,235,243 // por %xmm11,%xmm14 + .byte 102,69,15,254,245 // paddd %xmm13,%xmm14 + .byte 102,69,15,223,230 // pandn %xmm14,%xmm12 .byte 102,69,15,56,43,228 // packusdw %xmm12,%xmm12 - .byte 102,69,15,111,241 // movdqa %xmm9,%xmm14 - .byte 102,68,15,219,242 // pand %xmm2,%xmm14 + .byte 102,68,15,111,242 // movdqa %xmm2,%xmm14 .byte 102,68,15,111,250 // movdqa %xmm2,%xmm15 - .byte 102,69,15,239,254 // pxor %xmm14,%xmm15 - .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14 - .byte 102,69,15,111,234 // movdqa %xmm10,%xmm13 - .byte 102,69,15,102,239 // pcmpgtd %xmm15,%xmm13 - .byte 102,65,15,114,215,13 // psrld $0xd,%xmm15 - .byte 102,69,15,250,243 // psubd %xmm11,%xmm14 - .byte 102,69,15,254,247 // paddd %xmm15,%xmm14 - .byte 102,69,15,223,238 // pandn %xmm14,%xmm13 + .byte 102,69,15,219,250 // pand %xmm10,%xmm15 + .byte 102,69,15,239,247 // pxor %xmm15,%xmm14 + .byte 102,65,15,114,215,16 // psrld $0x10,%xmm15 + .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13 + .byte 102,69,15,102,238 // pcmpgtd %xmm14,%xmm13 + .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 + .byte 102,69,15,235,251 // por %xmm11,%xmm15 + .byte 102,69,15,254,254 // paddd %xmm14,%xmm15 + .byte 102,69,15,223,239 // pandn %xmm15,%xmm13 .byte 102,69,15,56,43,237 // packusdw %xmm13,%xmm13 - .byte 102,68,15,219,203 // pand %xmm3,%xmm9 + .byte 102,68,15,219,211 // pand %xmm3,%xmm10 .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14 - .byte 102,69,15,239,241 // pxor %xmm9,%xmm14 - .byte 102,65,15,114,209,16 // psrld $0x10,%xmm9 - .byte 102,69,15,102,214 // pcmpgtd %xmm14,%xmm10 + .byte 102,69,15,239,242 // pxor %xmm10,%xmm14 + .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10 + .byte 102,69,15,102,206 // pcmpgtd %xmm14,%xmm9 .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 - .byte 102,69,15,250,203 // psubd %xmm11,%xmm9 - .byte 102,69,15,254,206 // paddd %xmm14,%xmm9 - .byte 102,69,15,223,209 // pandn %xmm9,%xmm10 - .byte 102,69,15,56,43,210 // packusdw %xmm10,%xmm10 + .byte 102,69,15,235,211 // por %xmm11,%xmm10 + .byte 102,69,15,254,214 // paddd %xmm14,%xmm10 + .byte 102,69,15,223,202 // pandn %xmm10,%xmm9 + .byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9 .byte 102,69,15,97,196 // punpcklwd %xmm12,%xmm8 - .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13 + .byte 102,69,15,97,233 // punpcklwd %xmm9,%xmm13 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 .byte 102,69,15,98,205 // punpckldq %xmm13,%xmm9 .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) @@ -21564,7 +21368,7 @@ _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 3aa0 <_sk_linear_gradient_sse41+0x138> + .byte 15,132,254,0,0,0 // je 38e2 <_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) @@ -21614,12 +21418,12 @@ _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 39cb <_sk_linear_gradient_sse41+0x63> + .byte 15,133,65,255,255,255 // jne 380d <_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 3aad <_sk_linear_gradient_sse41+0x145> + .byte 235,13 // jmp 38ef <_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 @@ -22106,8 +21910,423 @@ BALIGN16 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) .byte 63 // (bad) .byte 0,0 // add %al,(%rax) - .byte 128 // .byte 0x80 + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 3f74 <.literal16+0x34> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 3f78 <.literal16+0x38> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 3f7c <.literal16+0x3c> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 3f80 <.literal16+0x40> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4005 <.literal16+0xc5> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 3f69 <.literal16+0x29> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a37faa <_sk_callback_sse41+0xffffffffe9a3414a> + .byte 220,63 // fdivrl (%rdi) + .byte 81 // push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,52,0 // add %dh,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4034 <.literal16+0xf4> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4038 <.literal16+0xf8> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 403c <.literal16+0xfc> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4040 <.literal16+0x100> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 40c5 <.literal16+0x185> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 4029 <.literal16+0xe9> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a3806a <_sk_callback_sse41+0xffffffffe9a3420a> + .byte 220,63 // fdivrl (%rdi) + .byte 81 // push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,52,0 // add %dh,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 40f4 <.literal16+0x1b4> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 40f8 <.literal16+0x1b8> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 40fc <.literal16+0x1bc> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4100 <.literal16+0x1c0> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4185 <.literal16+0x245> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 40e9 <.literal16+0x1a9> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a3812a <_sk_callback_sse41+0xffffffffe9a342ca> + .byte 220,63 // fdivrl (%rdi) + .byte 81 // push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,52,0 // add %dh,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 41b4 <.literal16+0x274> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 41b8 <.literal16+0x278> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 41bc <.literal16+0x27c> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 41c0 <.literal16+0x280> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4245 <.literal16+0x305> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) + .byte 117,191 // jne 41a9 <.literal16+0x269> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a381ea <_sk_callback_sse41+0xffffffffe9a3438a> + .byte 220,63 // fdivrl (%rdi) + .byte 81 // push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,128,0,0,0,128 // rex.WXB add %al,-0x80000000(%r8) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,4,0 // add %al,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 4,0 // add $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 4,0 // add $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 4,0 // add $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 128,0,0 // addb $0x0,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,4 // add %al,0x4000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,4,0 // add %al,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 4,0 // add $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 4,0 // add $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 128,0,0 // addb $0x0,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,128,56 // add %al,0x38800000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 128,56,0 // cmpb $0x0,(%rax) + .byte 0,128,56,0,0,128 // add %al,-0x7fffffc8(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 64,254 // rex (bad) + .byte 255,0 // incl (%rax) + .byte 64,254 // rex (bad) + .byte 255,0 // incl (%rax) + .byte 64,254 // rex (bad) + .byte 255,0 // incl (%rax) + .byte 64,254 // rex (bad) + .byte 255 // .byte 0xff BALIGN32 HIDDEN _sk_start_pipeline_sse2 @@ -22168,7 +22387,7 @@ _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 15,40,21,148,68,0,0 // movaps 0x4494(%rip),%xmm2 # 4510 <_sk_callback_sse2+0xd8> + .byte 15,40,21,212,66,0,0 // movaps 0x42d4(%rip),%xmm2 # 4350 <_sk_callback_sse2+0xe0> .byte 15,88,202 // addps %xmm2,%xmm1 .byte 15,16,2 // movups (%rdx),%xmm0 .byte 15,88,193 // addps %xmm1,%xmm0 @@ -22177,7 +22396,7 @@ _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 15,40,21,131,68,0,0 // movaps 0x4483(%rip),%xmm2 # 4520 <_sk_callback_sse2+0xe8> + .byte 15,40,21,195,66,0,0 // movaps 0x42c3(%rip),%xmm2 # 4360 <_sk_callback_sse2+0xf0> .byte 15,87,219 // xorps %xmm3,%xmm3 .byte 15,87,228 // xorps %xmm4,%xmm4 .byte 15,87,237 // xorps %xmm5,%xmm5 @@ -24384,92 +24603,62 @@ _sk_parametric_r_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,192 // mulps %xmm0,%xmm8 - .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 68,15,89,216 // mulps %xmm0,%xmm11 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,208 // mulps %xmm0,%xmm10 .byte 65,15,194,193,2 // cmpleps %xmm9,%xmm0 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 - .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,218 // addps %xmm10,%xmm11 - .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,212 // mulps %xmm12,%xmm10 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 - .byte 102,69,15,219,235 // pand %xmm11,%xmm13 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,235,229 // por %xmm13,%xmm12 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,89,220 // mulps %xmm12,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 243,68,15,16,24 // movss (%rax),%xmm11 + .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,244 // addps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 - .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 - .byte 69,15,40,234 // movaps %xmm10,%xmm13 - .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,250 // addps %xmm10,%xmm15 - .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 + .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,84,211 // andps %xmm11,%xmm10 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,226 // subps %xmm10,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 102,69,15,110,208 // movd %r8d,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,229 // mulps %xmm13,%xmm12 - .byte 69,15,92,252 // subps %xmm12,%xmm15 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,245 // subps %xmm13,%xmm14 - .byte 69,15,94,230 // divps %xmm14,%xmm12 - .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 + .byte 68,15,89,13,131,36,0,0 // mulps 0x2483(%rip),%xmm9 # 4370 <_sk_callback_sse2+0x100> + .byte 68,15,84,21,139,36,0,0 // andps 0x248b(%rip),%xmm10 # 4380 <_sk_callback_sse2+0x110> + .byte 68,15,86,21,147,36,0,0 // orps 0x2493(%rip),%xmm10 # 4390 <_sk_callback_sse2+0x120> + .byte 68,15,88,13,155,36,0,0 // addps 0x249b(%rip),%xmm9 # 43a0 <_sk_callback_sse2+0x130> + .byte 68,15,40,37,163,36,0,0 // movaps 0x24a3(%rip),%xmm12 # 43b0 <_sk_callback_sse2+0x140> .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 - .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,226 // addps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 68,15,88,21,163,36,0,0 // addps 0x24a3(%rip),%xmm10 # 43c0 <_sk_callback_sse2+0x150> + .byte 68,15,40,37,171,36,0,0 // movaps 0x24ab(%rip),%xmm12 # 43d0 <_sk_callback_sse2+0x160> + .byte 69,15,94,226 // divps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 69,15,89,203 // mulps %xmm11,%xmm9 + .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 + .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12 + .byte 68,15,84,37,149,36,0,0 // andps 0x2495(%rip),%xmm12 # 43e0 <_sk_callback_sse2+0x170> + .byte 69,15,87,210 // xorps %xmm10,%xmm10 + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 68,15,88,13,141,36,0,0 // addps 0x248d(%rip),%xmm9 # 43f0 <_sk_callback_sse2+0x180> + .byte 68,15,40,29,149,36,0,0 // movaps 0x2495(%rip),%xmm11 # 4400 <_sk_callback_sse2+0x190> + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,203 // subps %xmm11,%xmm9 + .byte 68,15,40,29,149,36,0,0 // movaps 0x2495(%rip),%xmm11 # 4410 <_sk_callback_sse2+0x1a0> + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 68,15,40,37,153,36,0,0 // movaps 0x2499(%rip),%xmm12 # 4420 <_sk_callback_sse2+0x1b0> + .byte 69,15,94,227 // divps %xmm11,%xmm12 + .byte 69,15,88,225 // addps %xmm9,%xmm12 + .byte 68,15,89,37,153,36,0,0 // mulps 0x2499(%rip),%xmm12 # 4430 <_sk_callback_sse2+0x1c0> + .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9 + .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,88,217 // addps %xmm9,%xmm11 .byte 68,15,84,192 // andps %xmm0,%xmm8 - .byte 65,15,85,196 // andnps %xmm12,%xmm0 + .byte 65,15,85,195 // andnps %xmm11,%xmm0 .byte 65,15,86,192 // orps %xmm8,%xmm0 - .byte 65,15,95,193 // maxps %xmm9,%xmm0 - .byte 65,15,93,195 // minps %xmm11,%xmm0 + .byte 65,15,95,194 // maxps %xmm10,%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 65,15,93,192 // minps %xmm8,%xmm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24483,92 +24672,62 @@ _sk_parametric_g_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,193 // mulps %xmm1,%xmm8 - .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 68,15,89,217 // mulps %xmm1,%xmm11 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,209 // mulps %xmm1,%xmm10 .byte 65,15,194,201,2 // cmpleps %xmm9,%xmm1 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 - .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,218 // addps %xmm10,%xmm11 - .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,212 // mulps %xmm12,%xmm10 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 - .byte 102,69,15,219,235 // pand %xmm11,%xmm13 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,235,229 // por %xmm13,%xmm12 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,89,220 // mulps %xmm12,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 243,68,15,16,24 // movss (%rax),%xmm11 + .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,244 // addps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 - .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 - .byte 69,15,40,234 // movaps %xmm10,%xmm13 - .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,250 // addps %xmm10,%xmm15 - .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 + .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,84,211 // andps %xmm11,%xmm10 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,226 // subps %xmm10,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 102,69,15,110,208 // movd %r8d,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,229 // mulps %xmm13,%xmm12 - .byte 69,15,92,252 // subps %xmm12,%xmm15 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,245 // subps %xmm13,%xmm14 - .byte 69,15,94,230 // divps %xmm14,%xmm12 - .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 + .byte 68,15,89,13,10,36,0,0 // mulps 0x240a(%rip),%xmm9 # 4440 <_sk_callback_sse2+0x1d0> + .byte 68,15,84,21,18,36,0,0 // andps 0x2412(%rip),%xmm10 # 4450 <_sk_callback_sse2+0x1e0> + .byte 68,15,86,21,26,36,0,0 // orps 0x241a(%rip),%xmm10 # 4460 <_sk_callback_sse2+0x1f0> + .byte 68,15,88,13,34,36,0,0 // addps 0x2422(%rip),%xmm9 # 4470 <_sk_callback_sse2+0x200> + .byte 68,15,40,37,42,36,0,0 // movaps 0x242a(%rip),%xmm12 # 4480 <_sk_callback_sse2+0x210> .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 - .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,226 // addps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 68,15,88,21,42,36,0,0 // addps 0x242a(%rip),%xmm10 # 4490 <_sk_callback_sse2+0x220> + .byte 68,15,40,37,50,36,0,0 // movaps 0x2432(%rip),%xmm12 # 44a0 <_sk_callback_sse2+0x230> + .byte 69,15,94,226 // divps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 69,15,89,203 // mulps %xmm11,%xmm9 + .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 + .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12 + .byte 68,15,84,37,28,36,0,0 // andps 0x241c(%rip),%xmm12 # 44b0 <_sk_callback_sse2+0x240> + .byte 69,15,87,210 // xorps %xmm10,%xmm10 + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 68,15,88,13,20,36,0,0 // addps 0x2414(%rip),%xmm9 # 44c0 <_sk_callback_sse2+0x250> + .byte 68,15,40,29,28,36,0,0 // movaps 0x241c(%rip),%xmm11 # 44d0 <_sk_callback_sse2+0x260> + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,203 // subps %xmm11,%xmm9 + .byte 68,15,40,29,28,36,0,0 // movaps 0x241c(%rip),%xmm11 # 44e0 <_sk_callback_sse2+0x270> + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 68,15,40,37,32,36,0,0 // movaps 0x2420(%rip),%xmm12 # 44f0 <_sk_callback_sse2+0x280> + .byte 69,15,94,227 // divps %xmm11,%xmm12 + .byte 69,15,88,225 // addps %xmm9,%xmm12 + .byte 68,15,89,37,32,36,0,0 // mulps 0x2420(%rip),%xmm12 # 4500 <_sk_callback_sse2+0x290> + .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9 + .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,88,217 // addps %xmm9,%xmm11 .byte 68,15,84,193 // andps %xmm1,%xmm8 - .byte 65,15,85,204 // andnps %xmm12,%xmm1 + .byte 65,15,85,203 // andnps %xmm11,%xmm1 .byte 65,15,86,200 // orps %xmm8,%xmm1 - .byte 65,15,95,201 // maxps %xmm9,%xmm1 - .byte 65,15,93,203 // minps %xmm11,%xmm1 + .byte 65,15,95,202 // maxps %xmm10,%xmm1 + .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 65,15,93,200 // minps %xmm8,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24582,92 +24741,62 @@ _sk_parametric_b_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,194 // mulps %xmm2,%xmm8 - .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 68,15,89,218 // mulps %xmm2,%xmm11 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,210 // mulps %xmm2,%xmm10 .byte 65,15,194,209,2 // cmpleps %xmm9,%xmm2 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 - .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,218 // addps %xmm10,%xmm11 - .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,212 // mulps %xmm12,%xmm10 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 - .byte 102,69,15,219,235 // pand %xmm11,%xmm13 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,235,229 // por %xmm13,%xmm12 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,89,220 // mulps %xmm12,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 243,68,15,16,24 // movss (%rax),%xmm11 + .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,244 // addps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 - .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 - .byte 69,15,40,234 // movaps %xmm10,%xmm13 - .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,250 // addps %xmm10,%xmm15 - .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 + .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,84,211 // andps %xmm11,%xmm10 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,226 // subps %xmm10,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 102,69,15,110,208 // movd %r8d,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,229 // mulps %xmm13,%xmm12 - .byte 69,15,92,252 // subps %xmm12,%xmm15 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,245 // subps %xmm13,%xmm14 - .byte 69,15,94,230 // divps %xmm14,%xmm12 - .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 + .byte 68,15,89,13,145,35,0,0 // mulps 0x2391(%rip),%xmm9 # 4510 <_sk_callback_sse2+0x2a0> + .byte 68,15,84,21,153,35,0,0 // andps 0x2399(%rip),%xmm10 # 4520 <_sk_callback_sse2+0x2b0> + .byte 68,15,86,21,161,35,0,0 // orps 0x23a1(%rip),%xmm10 # 4530 <_sk_callback_sse2+0x2c0> + .byte 68,15,88,13,169,35,0,0 // addps 0x23a9(%rip),%xmm9 # 4540 <_sk_callback_sse2+0x2d0> + .byte 68,15,40,37,177,35,0,0 // movaps 0x23b1(%rip),%xmm12 # 4550 <_sk_callback_sse2+0x2e0> .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 - .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,226 // addps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 68,15,88,21,177,35,0,0 // addps 0x23b1(%rip),%xmm10 # 4560 <_sk_callback_sse2+0x2f0> + .byte 68,15,40,37,185,35,0,0 // movaps 0x23b9(%rip),%xmm12 # 4570 <_sk_callback_sse2+0x300> + .byte 69,15,94,226 // divps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 69,15,89,203 // mulps %xmm11,%xmm9 + .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 + .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12 + .byte 68,15,84,37,163,35,0,0 // andps 0x23a3(%rip),%xmm12 # 4580 <_sk_callback_sse2+0x310> + .byte 69,15,87,210 // xorps %xmm10,%xmm10 + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 68,15,88,13,155,35,0,0 // addps 0x239b(%rip),%xmm9 # 4590 <_sk_callback_sse2+0x320> + .byte 68,15,40,29,163,35,0,0 // movaps 0x23a3(%rip),%xmm11 # 45a0 <_sk_callback_sse2+0x330> + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,203 // subps %xmm11,%xmm9 + .byte 68,15,40,29,163,35,0,0 // movaps 0x23a3(%rip),%xmm11 # 45b0 <_sk_callback_sse2+0x340> + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 68,15,40,37,167,35,0,0 // movaps 0x23a7(%rip),%xmm12 # 45c0 <_sk_callback_sse2+0x350> + .byte 69,15,94,227 // divps %xmm11,%xmm12 + .byte 69,15,88,225 // addps %xmm9,%xmm12 + .byte 68,15,89,37,167,35,0,0 // mulps 0x23a7(%rip),%xmm12 # 45d0 <_sk_callback_sse2+0x360> + .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9 + .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,88,217 // addps %xmm9,%xmm11 .byte 68,15,84,194 // andps %xmm2,%xmm8 - .byte 65,15,85,212 // andnps %xmm12,%xmm2 + .byte 65,15,85,211 // andnps %xmm11,%xmm2 .byte 65,15,86,208 // orps %xmm8,%xmm2 - .byte 65,15,95,209 // maxps %xmm9,%xmm2 - .byte 65,15,93,211 // minps %xmm11,%xmm2 + .byte 65,15,95,210 // maxps %xmm10,%xmm2 + .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 65,15,93,208 // minps %xmm8,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24681,92 +24810,62 @@ _sk_parametric_a_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,195 // mulps %xmm3,%xmm8 - .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 68,15,89,219 // mulps %xmm3,%xmm11 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,211 // mulps %xmm3,%xmm10 .byte 65,15,194,217,2 // cmpleps %xmm9,%xmm3 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 - .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,218 // addps %xmm10,%xmm11 - .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 - .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,212 // mulps %xmm12,%xmm10 - .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 - .byte 102,69,15,219,235 // pand %xmm11,%xmm13 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,235,229 // por %xmm13,%xmm12 - .byte 185,119,115,248,66 // mov $0x42f87377,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,89,220 // mulps %xmm12,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d - .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 243,68,15,16,24 // movss (%rax),%xmm11 + .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,244 // addps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 - .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 - .byte 69,15,40,234 // movaps %xmm10,%xmm13 - .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,250 // addps %xmm10,%xmm15 - .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 + .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,84,211 // andps %xmm11,%xmm10 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,226 // subps %xmm10,%xmm12 - .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 102,69,15,110,208 // movd %r8d,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,229 // mulps %xmm13,%xmm12 - .byte 69,15,92,252 // subps %xmm12,%xmm15 - .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx - .byte 102,68,15,110,241 // movd %ecx,%xmm14 - .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,245 // subps %xmm13,%xmm14 - .byte 69,15,94,230 // divps %xmm14,%xmm12 - .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 + .byte 68,15,89,13,24,35,0,0 // mulps 0x2318(%rip),%xmm9 # 45e0 <_sk_callback_sse2+0x370> + .byte 68,15,84,21,32,35,0,0 // andps 0x2320(%rip),%xmm10 # 45f0 <_sk_callback_sse2+0x380> + .byte 68,15,86,21,40,35,0,0 // orps 0x2328(%rip),%xmm10 # 4600 <_sk_callback_sse2+0x390> + .byte 68,15,88,13,48,35,0,0 // addps 0x2330(%rip),%xmm9 # 4610 <_sk_callback_sse2+0x3a0> + .byte 68,15,40,37,56,35,0,0 // movaps 0x2338(%rip),%xmm12 # 4620 <_sk_callback_sse2+0x3b0> .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 - .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,226 // addps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 68,15,88,21,56,35,0,0 // addps 0x2338(%rip),%xmm10 # 4630 <_sk_callback_sse2+0x3c0> + .byte 68,15,40,37,64,35,0,0 // movaps 0x2340(%rip),%xmm12 # 4640 <_sk_callback_sse2+0x3d0> + .byte 69,15,94,226 // divps %xmm10,%xmm12 + .byte 69,15,92,204 // subps %xmm12,%xmm9 + .byte 69,15,89,203 // mulps %xmm11,%xmm9 + .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 + .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12 + .byte 68,15,84,37,42,35,0,0 // andps 0x232a(%rip),%xmm12 # 4650 <_sk_callback_sse2+0x3e0> + .byte 69,15,87,210 // xorps %xmm10,%xmm10 + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 68,15,88,13,34,35,0,0 // addps 0x2322(%rip),%xmm9 # 4660 <_sk_callback_sse2+0x3f0> + .byte 68,15,40,29,42,35,0,0 // movaps 0x232a(%rip),%xmm11 # 4670 <_sk_callback_sse2+0x400> + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,203 // subps %xmm11,%xmm9 + .byte 68,15,40,29,42,35,0,0 // movaps 0x232a(%rip),%xmm11 # 4680 <_sk_callback_sse2+0x410> + .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 68,15,40,37,46,35,0,0 // movaps 0x232e(%rip),%xmm12 # 4690 <_sk_callback_sse2+0x420> + .byte 69,15,94,227 // divps %xmm11,%xmm12 + .byte 69,15,88,225 // addps %xmm9,%xmm12 + .byte 68,15,89,37,46,35,0,0 // mulps 0x232e(%rip),%xmm12 # 46a0 <_sk_callback_sse2+0x430> + .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9 + .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,88,217 // addps %xmm9,%xmm11 .byte 68,15,84,195 // andps %xmm3,%xmm8 - .byte 65,15,85,220 // andnps %xmm12,%xmm3 + .byte 65,15,85,219 // andnps %xmm11,%xmm3 .byte 65,15,86,216 // orps %xmm8,%xmm3 - .byte 65,15,95,217 // maxps %xmm9,%xmm3 - .byte 65,15,93,219 // minps %xmm11,%xmm3 + .byte 65,15,95,218 // maxps %xmm10,%xmm3 + .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 65,15,93,216 // minps %xmm8,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -25024,9 +25123,9 @@ _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 28cc <_sk_gather_i8_sse2+0xf> + .byte 116,5 // je 2718 <_sk_gather_i8_sse2+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 28ce <_sk_gather_i8_sse2+0x11> + .byte 235,2 // jmp 271a <_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 @@ -25518,65 +25617,68 @@ _sk_load_f16_sse2: .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 .byte 102,68,15,97,193 // punpcklwd %xmm1,%xmm8 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 - .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 - .byte 102,68,15,97,224 // punpcklwd %xmm0,%xmm12 + .byte 102,69,15,111,240 // movdqa %xmm8,%xmm14 + .byte 102,68,15,97,240 // punpcklwd %xmm0,%xmm14 .byte 102,68,15,105,192 // punpckhwd %xmm0,%xmm8 .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 - .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 - .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13 - .byte 184,0,128,0,0 // mov $0x8000,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,68,15,112,200,0 // pshufd $0x0,%xmm0,%xmm9 - .byte 102,65,15,111,205 // movdqa %xmm13,%xmm1 + .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 + .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1 + .byte 102,68,15,111,13,29,24,0,0 // movdqa 0x181d(%rip),%xmm9 # 46b0 <_sk_callback_sse2+0x440> + .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 + .byte 102,15,114,240,16 // pslld $0x10,%xmm0 + .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 102,68,15,111,29,2,24,0,0 // movdqa 0x1802(%rip),%xmm11 # 46c0 <_sk_callback_sse2+0x450> + .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 + .byte 102,68,15,111,37,4,24,0,0 // movdqa 0x1804(%rip),%xmm12 # 46d0 <_sk_callback_sse2+0x460> + .byte 102,65,15,239,204 // pxor %xmm12,%xmm1 + .byte 102,15,111,29,7,24,0,0 // movdqa 0x1807(%rip),%xmm3 # 46e0 <_sk_callback_sse2+0x470> + .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 + .byte 102,65,15,115,222,8 // psrldq $0x8,%xmm14 + .byte 102,69,15,97,242 // punpcklwd %xmm10,%xmm14 + .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 .byte 102,65,15,219,201 // pand %xmm9,%xmm1 - .byte 102,68,15,239,233 // pxor %xmm1,%xmm13 - .byte 184,0,4,0,0 // mov $0x400,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 + .byte 102,68,15,239,241 // pxor %xmm1,%xmm14 .byte 102,15,114,241,16 // pslld $0x10,%xmm1 - .byte 102,15,111,195 // movdqa %xmm3,%xmm0 - .byte 102,65,15,102,197 // pcmpgtd %xmm13,%xmm0 - .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 - .byte 184,0,0,0,56 // mov $0x38000000,%eax - .byte 102,15,110,208 // movd %eax,%xmm2 - .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 - .byte 102,65,15,254,203 // paddd %xmm11,%xmm1 - .byte 102,65,15,254,205 // paddd %xmm13,%xmm1 - .byte 102,15,223,193 // pandn %xmm1,%xmm0 - .byte 102,65,15,115,220,8 // psrldq $0x8,%xmm12 - .byte 102,69,15,97,226 // punpcklwd %xmm10,%xmm12 - .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 - .byte 102,65,15,219,209 // pand %xmm9,%xmm2 - .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 - .byte 102,15,114,242,16 // pslld $0x10,%xmm2 - .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,65,15,102,204 // pcmpgtd %xmm12,%xmm1 - .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 + .byte 102,65,15,111,214 // movdqa %xmm14,%xmm2 + .byte 102,15,114,242,13 // pslld $0xd,%xmm2 + .byte 102,15,235,209 // por %xmm1,%xmm2 .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 - .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 + .byte 102,69,15,239,244 // pxor %xmm12,%xmm14 + .byte 102,15,111,203 // movdqa %xmm3,%xmm1 + .byte 102,65,15,102,206 // pcmpgtd %xmm14,%xmm1 .byte 102,15,223,202 // pandn %xmm2,%xmm1 - .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 - .byte 102,69,15,97,226 // punpcklwd %xmm10,%xmm12 - .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 - .byte 102,69,15,219,233 // pand %xmm9,%xmm13 - .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 - .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 + .byte 102,69,15,111,232 // movdqa %xmm8,%xmm13 + .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13 + .byte 102,65,15,111,213 // movdqa %xmm13,%xmm2 + .byte 102,65,15,219,209 // pand %xmm9,%xmm2 + .byte 102,68,15,239,234 // pxor %xmm2,%xmm13 + .byte 102,15,114,242,16 // pslld $0x10,%xmm2 + .byte 102,69,15,111,245 // movdqa %xmm13,%xmm14 + .byte 102,65,15,114,246,13 // pslld $0xd,%xmm14 + .byte 102,68,15,235,242 // por %xmm2,%xmm14 + .byte 102,69,15,254,243 // paddd %xmm11,%xmm14 + .byte 102,69,15,239,236 // pxor %xmm12,%xmm13 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,65,15,102,212 // pcmpgtd %xmm12,%xmm2 - .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 - .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 - .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 - .byte 102,65,15,223,213 // pandn %xmm13,%xmm2 + .byte 102,65,15,102,213 // pcmpgtd %xmm13,%xmm2 + .byte 102,65,15,223,214 // pandn %xmm14,%xmm2 .byte 102,65,15,115,216,8 // psrldq $0x8,%xmm8 .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 .byte 102,69,15,219,200 // pand %xmm8,%xmm9 .byte 102,69,15,239,193 // pxor %xmm9,%xmm8 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 + .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 + .byte 102,65,15,114,242,13 // pslld $0xd,%xmm10 + .byte 102,69,15,235,209 // por %xmm9,%xmm10 + .byte 102,69,15,254,211 // paddd %xmm11,%xmm10 + .byte 102,69,15,239,196 // pxor %xmm12,%xmm8 .byte 102,65,15,102,216 // pcmpgtd %xmm8,%xmm3 - .byte 102,65,15,114,240,13 // pslld $0xd,%xmm8 - .byte 102,69,15,254,203 // paddd %xmm11,%xmm9 - .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 - .byte 102,65,15,223,217 // pandn %xmm9,%xmm3 + .byte 102,65,15,223,218 // pandn %xmm10,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -25613,65 +25715,68 @@ _sk_gather_f16_sse2: .byte 102,68,15,111,193 // movdqa %xmm1,%xmm8 .byte 102,68,15,97,194 // punpcklwd %xmm2,%xmm8 .byte 102,15,105,202 // punpckhwd %xmm2,%xmm1 - .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 - .byte 102,68,15,97,225 // punpcklwd %xmm1,%xmm12 + .byte 102,69,15,111,240 // movdqa %xmm8,%xmm14 + .byte 102,68,15,97,241 // punpcklwd %xmm1,%xmm14 .byte 102,68,15,105,193 // punpckhwd %xmm1,%xmm8 .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 - .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 - .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13 - .byte 184,0,128,0,0 // mov $0x8000,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,68,15,112,200,0 // pshufd $0x0,%xmm0,%xmm9 - .byte 102,65,15,111,205 // movdqa %xmm13,%xmm1 + .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 + .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1 + .byte 102,68,15,111,13,149,22,0,0 // movdqa 0x1695(%rip),%xmm9 # 46f0 <_sk_callback_sse2+0x480> + .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 + .byte 102,15,114,240,16 // pslld $0x10,%xmm0 + .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 102,68,15,111,29,122,22,0,0 // movdqa 0x167a(%rip),%xmm11 # 4700 <_sk_callback_sse2+0x490> + .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 + .byte 102,68,15,111,37,124,22,0,0 // movdqa 0x167c(%rip),%xmm12 # 4710 <_sk_callback_sse2+0x4a0> + .byte 102,65,15,239,204 // pxor %xmm12,%xmm1 + .byte 102,15,111,29,127,22,0,0 // movdqa 0x167f(%rip),%xmm3 # 4720 <_sk_callback_sse2+0x4b0> + .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 + .byte 102,65,15,115,222,8 // psrldq $0x8,%xmm14 + .byte 102,69,15,97,242 // punpcklwd %xmm10,%xmm14 + .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 .byte 102,65,15,219,201 // pand %xmm9,%xmm1 - .byte 102,68,15,239,233 // pxor %xmm1,%xmm13 - .byte 184,0,4,0,0 // mov $0x400,%eax - .byte 102,15,110,192 // movd %eax,%xmm0 - .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 + .byte 102,68,15,239,241 // pxor %xmm1,%xmm14 .byte 102,15,114,241,16 // pslld $0x10,%xmm1 - .byte 102,15,111,195 // movdqa %xmm3,%xmm0 - .byte 102,65,15,102,197 // pcmpgtd %xmm13,%xmm0 - .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 - .byte 184,0,0,0,56 // mov $0x38000000,%eax - .byte 102,15,110,208 // movd %eax,%xmm2 - .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 - .byte 102,65,15,254,203 // paddd %xmm11,%xmm1 - .byte 102,65,15,254,205 // paddd %xmm13,%xmm1 - .byte 102,15,223,193 // pandn %xmm1,%xmm0 - .byte 102,65,15,115,220,8 // psrldq $0x8,%xmm12 - .byte 102,69,15,97,226 // punpcklwd %xmm10,%xmm12 - .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 - .byte 102,65,15,219,209 // pand %xmm9,%xmm2 - .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 - .byte 102,15,114,242,16 // pslld $0x10,%xmm2 - .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,65,15,102,204 // pcmpgtd %xmm12,%xmm1 - .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 + .byte 102,65,15,111,214 // movdqa %xmm14,%xmm2 + .byte 102,15,114,242,13 // pslld $0xd,%xmm2 + .byte 102,15,235,209 // por %xmm1,%xmm2 .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 - .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 + .byte 102,69,15,239,244 // pxor %xmm12,%xmm14 + .byte 102,15,111,203 // movdqa %xmm3,%xmm1 + .byte 102,65,15,102,206 // pcmpgtd %xmm14,%xmm1 .byte 102,15,223,202 // pandn %xmm2,%xmm1 - .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 - .byte 102,69,15,97,226 // punpcklwd %xmm10,%xmm12 - .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 - .byte 102,69,15,219,233 // pand %xmm9,%xmm13 - .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 - .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 + .byte 102,69,15,111,232 // movdqa %xmm8,%xmm13 + .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13 + .byte 102,65,15,111,213 // movdqa %xmm13,%xmm2 + .byte 102,65,15,219,209 // pand %xmm9,%xmm2 + .byte 102,68,15,239,234 // pxor %xmm2,%xmm13 + .byte 102,15,114,242,16 // pslld $0x10,%xmm2 + .byte 102,69,15,111,245 // movdqa %xmm13,%xmm14 + .byte 102,65,15,114,246,13 // pslld $0xd,%xmm14 + .byte 102,68,15,235,242 // por %xmm2,%xmm14 + .byte 102,69,15,254,243 // paddd %xmm11,%xmm14 + .byte 102,69,15,239,236 // pxor %xmm12,%xmm13 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,65,15,102,212 // pcmpgtd %xmm12,%xmm2 - .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 - .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 - .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 - .byte 102,65,15,223,213 // pandn %xmm13,%xmm2 + .byte 102,65,15,102,213 // pcmpgtd %xmm13,%xmm2 + .byte 102,65,15,223,214 // pandn %xmm14,%xmm2 .byte 102,65,15,115,216,8 // psrldq $0x8,%xmm8 .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 .byte 102,69,15,219,200 // pand %xmm8,%xmm9 .byte 102,69,15,239,193 // pxor %xmm9,%xmm8 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 + .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 + .byte 102,65,15,114,242,13 // pslld $0xd,%xmm10 + .byte 102,69,15,235,209 // por %xmm9,%xmm10 + .byte 102,69,15,254,211 // paddd %xmm11,%xmm10 + .byte 102,69,15,239,196 // pxor %xmm12,%xmm8 .byte 102,65,15,102,216 // pcmpgtd %xmm8,%xmm3 - .byte 102,65,15,114,240,13 // pslld $0xd,%xmm8 - .byte 102,69,15,254,203 // paddd %xmm11,%xmm9 - .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 - .byte 102,65,15,223,217 // pandn %xmm9,%xmm3 + .byte 102,65,15,223,218 // pandn %xmm10,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -25681,71 +25786,65 @@ FUNCTION(_sk_store_f16_sse2) _sk_store_f16_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 185,0,0,0,128 // mov $0x80000000,%ecx - .byte 102,68,15,110,193 // movd %ecx,%xmm8 - .byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9 - .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 - .byte 102,68,15,219,224 // pand %xmm0,%xmm12 + .byte 102,68,15,111,21,167,21,0,0 // movdqa 0x15a7(%rip),%xmm10 # 4730 <_sk_callback_sse2+0x4c0> + .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13 - .byte 102,69,15,239,236 // pxor %xmm12,%xmm13 - .byte 185,0,0,128,56 // mov $0x38800000,%ecx - .byte 102,68,15,110,193 // movd %ecx,%xmm8 - .byte 102,69,15,112,208,0 // pshufd $0x0,%xmm8,%xmm10 - .byte 102,65,15,114,212,16 // psrld $0x10,%xmm12 - .byte 102,69,15,111,194 // movdqa %xmm10,%xmm8 - .byte 102,69,15,102,197 // pcmpgtd %xmm13,%xmm8 - .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 - .byte 185,0,192,1,0 // mov $0x1c000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 - .byte 102,69,15,250,227 // psubd %xmm11,%xmm12 - .byte 102,69,15,254,229 // paddd %xmm13,%xmm12 - .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12 - .byte 102,65,15,114,228,16 // psrad $0x10,%xmm12 - .byte 102,69,15,223,196 // pandn %xmm12,%xmm8 - .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 - .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13 - .byte 102,68,15,219,233 // pand %xmm1,%xmm13 - .byte 102,68,15,111,241 // movdqa %xmm1,%xmm14 - .byte 102,69,15,239,245 // pxor %xmm13,%xmm14 + .byte 102,69,15,219,234 // pand %xmm10,%xmm13 + .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 + .byte 102,68,15,111,13,154,21,0,0 // movdqa 0x159a(%rip),%xmm9 # 4740 <_sk_callback_sse2+0x4d0> .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13 - .byte 102,69,15,111,226 // movdqa %xmm10,%xmm12 - .byte 102,69,15,102,230 // pcmpgtd %xmm14,%xmm12 - .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 - .byte 102,69,15,250,235 // psubd %xmm11,%xmm13 - .byte 102,69,15,254,238 // paddd %xmm14,%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 102,68,15,111,29,139,21,0,0 // movdqa 0x158b(%rip),%xmm11 # 4750 <_sk_callback_sse2+0x4e0> + .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 .byte 102,65,15,114,229,16 // psrad $0x10,%xmm13 - .byte 102,69,15,223,229 // pandn %xmm13,%xmm12 - .byte 102,69,15,107,228 // packssdw %xmm12,%xmm12 - .byte 102,69,15,111,241 // movdqa %xmm9,%xmm14 - .byte 102,68,15,219,242 // pand %xmm2,%xmm14 - .byte 102,68,15,111,250 // movdqa %xmm2,%xmm15 - .byte 102,69,15,239,254 // pxor %xmm14,%xmm15 + .byte 102,69,15,223,197 // pandn %xmm13,%xmm8 + .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 + .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 + .byte 102,68,15,111,241 // movdqa %xmm1,%xmm14 + .byte 102,69,15,219,242 // pand %xmm10,%xmm14 + .byte 102,69,15,239,238 // pxor %xmm14,%xmm13 .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14 - .byte 102,69,15,111,234 // movdqa %xmm10,%xmm13 - .byte 102,69,15,102,239 // pcmpgtd %xmm15,%xmm13 - .byte 102,65,15,114,215,13 // psrld $0xd,%xmm15 - .byte 102,69,15,250,243 // psubd %xmm11,%xmm14 - .byte 102,69,15,254,247 // paddd %xmm15,%xmm14 + .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 + .byte 102,69,15,102,229 // pcmpgtd %xmm13,%xmm12 + .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 + .byte 102,69,15,235,243 // por %xmm11,%xmm14 + .byte 102,69,15,254,245 // paddd %xmm13,%xmm14 .byte 102,65,15,114,246,16 // pslld $0x10,%xmm14 .byte 102,65,15,114,230,16 // psrad $0x10,%xmm14 - .byte 102,69,15,223,238 // pandn %xmm14,%xmm13 + .byte 102,69,15,223,230 // pandn %xmm14,%xmm12 + .byte 102,69,15,107,228 // packssdw %xmm12,%xmm12 + .byte 102,68,15,111,242 // movdqa %xmm2,%xmm14 + .byte 102,68,15,111,250 // movdqa %xmm2,%xmm15 + .byte 102,69,15,219,250 // pand %xmm10,%xmm15 + .byte 102,69,15,239,247 // pxor %xmm15,%xmm14 + .byte 102,65,15,114,215,16 // psrld $0x10,%xmm15 + .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13 + .byte 102,69,15,102,238 // pcmpgtd %xmm14,%xmm13 + .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 + .byte 102,69,15,235,251 // por %xmm11,%xmm15 + .byte 102,69,15,254,254 // paddd %xmm14,%xmm15 + .byte 102,65,15,114,247,16 // pslld $0x10,%xmm15 + .byte 102,65,15,114,231,16 // psrad $0x10,%xmm15 + .byte 102,69,15,223,239 // pandn %xmm15,%xmm13 .byte 102,69,15,107,237 // packssdw %xmm13,%xmm13 - .byte 102,68,15,219,203 // pand %xmm3,%xmm9 + .byte 102,68,15,219,211 // pand %xmm3,%xmm10 .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14 - .byte 102,69,15,239,241 // pxor %xmm9,%xmm14 - .byte 102,65,15,114,209,16 // psrld $0x10,%xmm9 - .byte 102,69,15,102,214 // pcmpgtd %xmm14,%xmm10 + .byte 102,69,15,239,242 // pxor %xmm10,%xmm14 + .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10 + .byte 102,69,15,102,206 // pcmpgtd %xmm14,%xmm9 .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 - .byte 102,69,15,250,203 // psubd %xmm11,%xmm9 - .byte 102,69,15,254,206 // paddd %xmm14,%xmm9 - .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 - .byte 102,65,15,114,225,16 // psrad $0x10,%xmm9 - .byte 102,69,15,223,209 // pandn %xmm9,%xmm10 - .byte 102,69,15,107,210 // packssdw %xmm10,%xmm10 + .byte 102,69,15,235,211 // por %xmm11,%xmm10 + .byte 102,69,15,254,214 // paddd %xmm14,%xmm10 + .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10 + .byte 102,65,15,114,226,16 // psrad $0x10,%xmm10 + .byte 102,69,15,223,202 // pandn %xmm10,%xmm9 + .byte 102,69,15,107,201 // packssdw %xmm9,%xmm9 .byte 102,69,15,97,196 // punpcklwd %xmm12,%xmm8 - .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13 + .byte 102,69,15,97,233 // punpcklwd %xmm9,%xmm13 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 .byte 102,69,15,98,205 // punpckldq %xmm13,%xmm9 .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) @@ -26013,11 +26112,8 @@ _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 184,0,0,128,63 // mov $0x3f800000,%eax - .byte 102,68,15,110,216 // movd %eax,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,84,217 // andps %xmm9,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 + .byte 68,15,84,13,188,16,0,0 // andps 0x10bc(%rip),%xmm9 # 4760 <_sk_callback_sse2+0x4f0> + .byte 69,15,92,209 // subps %xmm9,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,92,194 // subps %xmm10,%xmm0 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 @@ -26038,11 +26134,8 @@ _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 184,0,0,128,63 // mov $0x3f800000,%eax - .byte 102,68,15,110,216 // movd %eax,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,84,217 // andps %xmm9,%xmm11 - .byte 69,15,92,211 // subps %xmm11,%xmm10 + .byte 68,15,84,13,132,16,0,0 // andps 0x1084(%rip),%xmm9 # 4770 <_sk_callback_sse2+0x500> + .byte 69,15,92,209 // subps %xmm9,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,92,202 // subps %xmm10,%xmm1 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 @@ -26056,31 +26149,28 @@ HIDDEN _sk_mirror_x_sse2 FUNCTION(_sk_mirror_x_sse2) _sk_mirror_x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 69,15,40,193 // movaps %xmm9,%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 65,15,92,192 // subps %xmm8,%xmm0 - .byte 243,69,15,88,201 // addss %xmm9,%xmm9 + .byte 243,68,15,16,0 // movss (%rax),%xmm8 + .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 65,15,92,193 // subps %xmm9,%xmm0 + .byte 243,69,15,88,192 // addss %xmm8,%xmm8 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,40,208 // movaps %xmm0,%xmm10 - .byte 69,15,94,209 // divps %xmm9,%xmm10 + .byte 69,15,94,208 // divps %xmm8,%xmm10 .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 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 69,15,84,226 // andps %xmm10,%xmm12 - .byte 69,15,87,210 // xorps %xmm10,%xmm10 - .byte 69,15,92,220 // subps %xmm12,%xmm11 - .byte 69,15,89,217 // mulps %xmm9,%xmm11 + .byte 68,15,84,21,58,16,0,0 // andps 0x103a(%rip),%xmm10 # 4780 <_sk_callback_sse2+0x510> + .byte 69,15,87,228 // xorps %xmm12,%xmm12 + .byte 69,15,92,218 // subps %xmm10,%xmm11 + .byte 69,15,89,216 // mulps %xmm8,%xmm11 .byte 65,15,92,195 // subps %xmm11,%xmm0 - .byte 65,15,92,192 // subps %xmm8,%xmm0 - .byte 68,15,92,208 // subps %xmm0,%xmm10 - .byte 65,15,84,194 // andps %xmm10,%xmm0 - .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 - .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 - .byte 65,15,93,193 // minps %xmm9,%xmm0 + .byte 65,15,92,193 // subps %xmm9,%xmm0 + .byte 68,15,92,224 // subps %xmm0,%xmm12 + .byte 65,15,84,196 // andps %xmm12,%xmm0 + .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 + .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 + .byte 65,15,93,192 // minps %xmm8,%xmm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -26089,31 +26179,28 @@ HIDDEN _sk_mirror_y_sse2 FUNCTION(_sk_mirror_y_sse2) _sk_mirror_y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 69,15,40,193 // movaps %xmm9,%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 65,15,92,200 // subps %xmm8,%xmm1 - .byte 243,69,15,88,201 // addss %xmm9,%xmm9 + .byte 243,68,15,16,0 // movss (%rax),%xmm8 + .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 65,15,92,201 // subps %xmm9,%xmm1 + .byte 243,69,15,88,192 // addss %xmm8,%xmm8 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,40,209 // movaps %xmm1,%xmm10 - .byte 69,15,94,209 // divps %xmm9,%xmm10 + .byte 69,15,94,208 // divps %xmm8,%xmm10 .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 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 69,15,84,226 // andps %xmm10,%xmm12 - .byte 69,15,87,210 // xorps %xmm10,%xmm10 - .byte 69,15,92,220 // subps %xmm12,%xmm11 - .byte 69,15,89,217 // mulps %xmm9,%xmm11 + .byte 68,15,84,21,224,15,0,0 // andps 0xfe0(%rip),%xmm10 # 4790 <_sk_callback_sse2+0x520> + .byte 69,15,87,228 // xorps %xmm12,%xmm12 + .byte 69,15,92,218 // subps %xmm10,%xmm11 + .byte 69,15,89,216 // mulps %xmm8,%xmm11 .byte 65,15,92,203 // subps %xmm11,%xmm1 - .byte 65,15,92,200 // subps %xmm8,%xmm1 - .byte 68,15,92,209 // subps %xmm1,%xmm10 - .byte 65,15,84,202 // andps %xmm10,%xmm1 - .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 - .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 - .byte 65,15,93,201 // minps %xmm9,%xmm1 + .byte 65,15,92,201 // subps %xmm9,%xmm1 + .byte 68,15,92,225 // subps %xmm1,%xmm12 + .byte 65,15,84,204 // andps %xmm12,%xmm1 + .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 + .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 + .byte 65,15,93,200 // minps %xmm8,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -26366,7 +26453,7 @@ _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 3e85 <_sk_linear_gradient_sse2+0x149> + .byte 15,132,15,1,0,0 // je 3cc4 <_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 @@ -26427,8 +26514,8 @@ _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 3d8b <_sk_linear_gradient_sse2+0x4f> - .byte 235,13 // jmp 3e92 <_sk_linear_gradient_sse2+0x156> + .byte 15,133,8,255,255,255 // jne 3bca <_sk_linear_gradient_sse2+0x4f> + .byte 235,13 // jmp 3cd1 <_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 @@ -26492,9 +26579,7 @@ _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 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 68,15,40,37,7,10,0,0 // movaps 0xa07(%rip),%xmm12 # 47a0 <_sk_callback_sse2+0x530> .byte 69,15,84,220 // andps %xmm12,%xmm11 .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 69,15,92,202 // subps %xmm10,%xmm9 @@ -26928,6 +27013,467 @@ BALIGN16 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) .byte 63 // (bad) .byte 0,0 // add %al,(%rax) - .byte 128 // .byte 0x80 + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4384 <.literal16+0x34> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4388 <.literal16+0x38> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 438c <.literal16+0x3c> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4390 <.literal16+0x40> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4415 <.literal16+0xc5> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 4379 <.literal16+0x29> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a383ba <_sk_callback_sse2+0xffffffffe9a3414a> + .byte 220,63 // fdivrl (%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 128,63,81 // cmpb $0x51,(%rdi) + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,52,0 // add %dh,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4454 <.literal16+0x104> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4458 <.literal16+0x108> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 445c <.literal16+0x10c> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4460 <.literal16+0x110> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 44e5 <.literal16+0x195> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 4449 <.literal16+0xf9> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a3848a <_sk_callback_sse2+0xffffffffe9a3421a> + .byte 220,63 // fdivrl (%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 128,63,81 // cmpb $0x51,(%rdi) + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,52,0 // add %dh,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 4524 <.literal16+0x1d4> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4528 <.literal16+0x1d8> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 452c <.literal16+0x1dc> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4530 <.literal16+0x1e0> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 45b5 <.literal16+0x265> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 4519 <.literal16+0x1c9> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a3855a <_sk_callback_sse2+0xffffffffe9a342ea> + .byte 220,63 // fdivrl (%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 128,63,81 // cmpb $0x51,(%rdi) + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,52,0 // add %dh,(%rax,%rax,1) + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,0 // xor $0x0,%al + .byte 0,0 // add %al,(%rax) + .byte 52,255 // xor $0xff,%al + .byte 255 // (bad) + .byte 127,0 // jg 45f4 <.literal16+0x2a4> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 45f8 <.literal16+0x2a8> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 45fc <.literal16+0x2ac> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 127,0 // jg 4600 <.literal16+0x2b0> + .byte 0,0 // add %al,(%rax) + .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 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 119,115 // ja 4685 <.literal16+0x335> + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,119,115 // retq $0x7377 + .byte 248 // clc + .byte 194,117,191 // retq $0xbf75 + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 117,191 // jne 45e9 <.literal16+0x299> + .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi + .byte 63 // (bad) + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .byte 249 // stc + .byte 68,180,62 // rex.R mov $0x3e,%spl + .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 ffffffffe9a3862a <_sk_callback_sse2+0xffffffffe9a343ba> + .byte 220,63 // fdivrl (%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 128,63,81 // cmpb $0x51,(%rdi) + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,81 // rex.X push %rcx + .byte 140,242 // mov %?,%edx + .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi + .byte 63 // (bad) + .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi + .byte 63 // (bad) + .byte 248 // clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,248 // rex clc + .byte 245 // cmc + .byte 154 // (bad) + .byte 64,254 // rex (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,254 // rex.B (bad) + .byte 210,221 // rcr %cl,%ch + .byte 65,0,0 // add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,0 // rex.WXB add %al,(%r8) + .byte 0,75,0 // add %cl,0x0(%rbx) + .byte 0,0 // add %al,(%rax) + .byte 75,0,128,0,0,0,128 // rex.WXB add %al,-0x80000000(%r8) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,4,0 // add %al,(%rax,%rax,1) + .byte 128,0,4 // addb $0x4,(%rax) + .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax) + .byte 0,4,0 // add %al,(%rax,%rax,1) + .byte 128,0,128 // addb $0x80,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,0 // add %al,0x0(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,56 // add %bh,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,4,0 // add %al,(%rax,%rax,1) + .byte 128,0,4 // addb $0x4,(%rax) + .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax) + .byte 0,4,0 // add %al,(%rax,%rax,1) + .byte 128,0,0 // addb $0x0,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax) + .byte 0,0 // add %al,(%rax) + .byte 128,56,0 // cmpb $0x0,(%rax) + .byte 0,128,56,0,0,128 // add %al,-0x7fffffc8(%rax) + .byte 56,0 // cmp %al,(%rax) + .byte 0,128,56,0,64,254 // add %al,-0x1bfffc8(%rax) + .byte 255,0 // incl (%rax) + .byte 64,254 // rex (bad) + .byte 255,0 // incl (%rax) + .byte 64,254 // rex (bad) + .byte 255,0 // incl (%rax) + .byte 64,254 // rex (bad) + .byte 255,0 // incl (%rax) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .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 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .byte 0,0 // add %al,(%rax) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .byte 0,0 // add %al,(%rax) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .byte 0,0 // add %al,(%rax) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) + .byte 63 // (bad) + .byte 0,0 // add %al,(%rax) + .byte 128,63,0 // cmpb $0x0,(%rdi) + .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) .byte 63 // (bad) #endif diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 12fc90ee38..2258440ca5 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 196,226,125,24,13,130,65,0,0 ; vbroadcastss 0x4182(%rip),%ymm1 # 42dc <_sk_callback_hsw+0x11a> + DB 196,226,125,24,13,74,64,0,0 ; vbroadcastss 0x404a(%rip),%ymm1 # 41a4 <_sk_callback_hsw+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 196,226,125,24,21,102,65,0,0 ; vbroadcastss 0x4166(%rip),%ymm2 # 42e0 <_sk_callback_hsw+0x11e> + DB 196,226,125,24,21,46,64,0,0 ; vbroadcastss 0x402e(%rip),%ymm2 # 41a8 <_sk_callback_hsw+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 @@ -1862,60 +1862,36 @@ _sk_parametric_r_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,226,125,24,0 ; vbroadcastss (%rax),%ymm0 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,232 ; vpbroadcastd %xmm0,%ymm13 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 - DB 196,65,125,219,210 ; vpand %ymm10,%ymm0,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 - DB 197,45,235,208 ; vpor %ymm0,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,240 ; vpbroadcastd %xmm0,%ymm14 - DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,216 ; vpbroadcastd %xmm0,%ymm11 - DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,232 ; vpbroadcastd %xmm0,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 - DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 - DB 197,148,94,192 ; vdivps %ymm0,%ymm13,%ymm0 - DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 - DB 197,28,89,216 ; vmulps %ymm0,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + DB 196,98,125,24,37,231,36,0,0 ; vbroadcastss 0x24e7(%rip),%ymm12 # 41ac <_sk_callback_hsw+0x122> + DB 196,98,125,24,45,226,36,0,0 ; vbroadcastss 0x24e2(%rip),%ymm13 # 41b0 <_sk_callback_hsw+0x126> + DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,216,36,0,0 ; vbroadcastss 0x24d8(%rip),%ymm13 # 41b4 <_sk_callback_hsw+0x12a> + DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,206,36,0,0 ; vbroadcastss 0x24ce(%rip),%ymm13 # 41b8 <_sk_callback_hsw+0x12e> + DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 + DB 196,98,125,24,29,196,36,0,0 ; vbroadcastss 0x24c4(%rip),%ymm11 # 41bc <_sk_callback_hsw+0x132> + DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 + DB 196,98,125,24,37,186,36,0,0 ; vbroadcastss 0x24ba(%rip),%ymm12 # 41c0 <_sk_callback_hsw+0x136> + DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 + DB 196,98,125,24,37,176,36,0,0 ; vbroadcastss 0x24b0(%rip),%ymm12 # 41c4 <_sk_callback_hsw+0x13a> + DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,224 ; vpbroadcastd %xmm0,%ymm12 - DB 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 - DB 196,65,124,88,219 ; vaddps %ymm11,%ymm0,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,232 ; vpbroadcastd %xmm0,%ymm13 - DB 196,66,45,172,235 ; vfnmadd213ps %ymm11,%ymm10,%ymm13 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,216 ; vpbroadcastd %xmm0,%ymm11 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 + 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 196,98,125,24,29,145,36,0,0 ; vbroadcastss 0x2491(%rip),%ymm11 # 41c8 <_sk_callback_hsw+0x13e> + DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 + DB 196,98,125,24,29,135,36,0,0 ; vbroadcastss 0x2487(%rip),%ymm11 # 41cc <_sk_callback_hsw+0x142> + DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11 + DB 196,226,125,24,5,125,36,0,0 ; vbroadcastss 0x247d(%rip),%ymm0 # 41d0 <_sk_callback_hsw+0x146> DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 197,164,94,192 ; vdivps %ymm0,%ymm11,%ymm0 - DB 197,148,88,192 ; vaddps %ymm0,%ymm13,%ymm0 - DB 197,156,89,192 ; vmulps %ymm0,%ymm12,%ymm0 + DB 196,98,125,24,21,115,36,0,0 ; vbroadcastss 0x2473(%rip),%ymm10 # 41d4 <_sk_callback_hsw+0x14a> + DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0 + DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 + DB 196,98,125,24,21,102,36,0,0 ; vbroadcastss 0x2466(%rip),%ymm10 # 41d8 <_sk_callback_hsw+0x14e> + 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 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 @@ -1940,60 +1916,36 @@ _sk_parametric_g_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,226,125,24,8 ; vbroadcastss (%rax),%ymm1 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,233 ; vpbroadcastd %xmm1,%ymm13 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 - DB 196,65,117,219,210 ; vpand %ymm10,%ymm1,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 - DB 197,45,235,209 ; vpor %ymm1,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,241 ; vpbroadcastd %xmm1,%ymm14 - DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,217 ; vpbroadcastd %xmm1,%ymm11 - DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,233 ; vpbroadcastd %xmm1,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 - DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1 - DB 197,148,94,201 ; vdivps %ymm1,%ymm13,%ymm1 - DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 - DB 197,28,89,217 ; vmulps %ymm1,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + DB 196,98,125,24,37,237,35,0,0 ; vbroadcastss 0x23ed(%rip),%ymm12 # 41dc <_sk_callback_hsw+0x152> + DB 196,98,125,24,45,232,35,0,0 ; vbroadcastss 0x23e8(%rip),%ymm13 # 41e0 <_sk_callback_hsw+0x156> + DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,222,35,0,0 ; vbroadcastss 0x23de(%rip),%ymm13 # 41e4 <_sk_callback_hsw+0x15a> + DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,212,35,0,0 ; vbroadcastss 0x23d4(%rip),%ymm13 # 41e8 <_sk_callback_hsw+0x15e> + DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 + DB 196,98,125,24,29,202,35,0,0 ; vbroadcastss 0x23ca(%rip),%ymm11 # 41ec <_sk_callback_hsw+0x162> + DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 + DB 196,98,125,24,37,192,35,0,0 ; vbroadcastss 0x23c0(%rip),%ymm12 # 41f0 <_sk_callback_hsw+0x166> + DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 + DB 196,98,125,24,37,182,35,0,0 ; vbroadcastss 0x23b6(%rip),%ymm12 # 41f4 <_sk_callback_hsw+0x16a> + DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,225 ; vpbroadcastd %xmm1,%ymm12 - DB 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 - DB 196,65,116,88,219 ; vaddps %ymm11,%ymm1,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,233 ; vpbroadcastd %xmm1,%ymm13 - DB 196,66,45,172,235 ; vfnmadd213ps %ymm11,%ymm10,%ymm13 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,217 ; vpbroadcastd %xmm1,%ymm11 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 + 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 196,98,125,24,29,151,35,0,0 ; vbroadcastss 0x2397(%rip),%ymm11 # 41f8 <_sk_callback_hsw+0x16e> + DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 + DB 196,98,125,24,29,141,35,0,0 ; vbroadcastss 0x238d(%rip),%ymm11 # 41fc <_sk_callback_hsw+0x172> + DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11 + DB 196,226,125,24,13,131,35,0,0 ; vbroadcastss 0x2383(%rip),%ymm1 # 4200 <_sk_callback_hsw+0x176> DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 - DB 197,164,94,201 ; vdivps %ymm1,%ymm11,%ymm1 - DB 197,148,88,201 ; vaddps %ymm1,%ymm13,%ymm1 - DB 197,156,89,201 ; vmulps %ymm1,%ymm12,%ymm1 + DB 196,98,125,24,21,121,35,0,0 ; vbroadcastss 0x2379(%rip),%ymm10 # 4204 <_sk_callback_hsw+0x17a> + DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 + DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 + DB 196,98,125,24,21,108,35,0,0 ; vbroadcastss 0x236c(%rip),%ymm10 # 4208 <_sk_callback_hsw+0x17e> + 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 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 @@ -2018,60 +1970,36 @@ _sk_parametric_b_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,234 ; vpbroadcastd %xmm2,%ymm13 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 - DB 196,65,109,219,210 ; vpand %ymm10,%ymm2,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 - DB 197,45,235,210 ; vpor %ymm2,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,242 ; vpbroadcastd %xmm2,%ymm14 - DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,218 ; vpbroadcastd %xmm2,%ymm11 - DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,234 ; vpbroadcastd %xmm2,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 - DB 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 - DB 197,148,94,210 ; vdivps %ymm2,%ymm13,%ymm2 - DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 - DB 197,28,89,218 ; vmulps %ymm2,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + DB 196,98,125,24,37,243,34,0,0 ; vbroadcastss 0x22f3(%rip),%ymm12 # 420c <_sk_callback_hsw+0x182> + DB 196,98,125,24,45,238,34,0,0 ; vbroadcastss 0x22ee(%rip),%ymm13 # 4210 <_sk_callback_hsw+0x186> + DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,228,34,0,0 ; vbroadcastss 0x22e4(%rip),%ymm13 # 4214 <_sk_callback_hsw+0x18a> + DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,218,34,0,0 ; vbroadcastss 0x22da(%rip),%ymm13 # 4218 <_sk_callback_hsw+0x18e> + DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 + DB 196,98,125,24,29,208,34,0,0 ; vbroadcastss 0x22d0(%rip),%ymm11 # 421c <_sk_callback_hsw+0x192> + DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 + DB 196,98,125,24,37,198,34,0,0 ; vbroadcastss 0x22c6(%rip),%ymm12 # 4220 <_sk_callback_hsw+0x196> + DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 + DB 196,98,125,24,37,188,34,0,0 ; vbroadcastss 0x22bc(%rip),%ymm12 # 4224 <_sk_callback_hsw+0x19a> + DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,226 ; vpbroadcastd %xmm2,%ymm12 - DB 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 - DB 196,65,108,88,219 ; vaddps %ymm11,%ymm2,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,234 ; vpbroadcastd %xmm2,%ymm13 - DB 196,66,45,172,235 ; vfnmadd213ps %ymm11,%ymm10,%ymm13 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,218 ; vpbroadcastd %xmm2,%ymm11 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 + 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 196,98,125,24,29,157,34,0,0 ; vbroadcastss 0x229d(%rip),%ymm11 # 4228 <_sk_callback_hsw+0x19e> + DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 + DB 196,98,125,24,29,147,34,0,0 ; vbroadcastss 0x2293(%rip),%ymm11 # 422c <_sk_callback_hsw+0x1a2> + DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11 + DB 196,226,125,24,21,137,34,0,0 ; vbroadcastss 0x2289(%rip),%ymm2 # 4230 <_sk_callback_hsw+0x1a6> DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2 - DB 197,164,94,210 ; vdivps %ymm2,%ymm11,%ymm2 - DB 197,148,88,210 ; vaddps %ymm2,%ymm13,%ymm2 - DB 197,156,89,210 ; vmulps %ymm2,%ymm12,%ymm2 + DB 196,98,125,24,21,127,34,0,0 ; vbroadcastss 0x227f(%rip),%ymm10 # 4234 <_sk_callback_hsw+0x1aa> + DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2 + DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 + DB 196,98,125,24,21,114,34,0,0 ; vbroadcastss 0x2272(%rip),%ymm10 # 4238 <_sk_callback_hsw+0x1ae> + 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 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 @@ -2096,60 +2024,36 @@ _sk_parametric_a_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,226,125,24,24 ; vbroadcastss (%rax),%ymm3 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,235 ; vpbroadcastd %xmm3,%ymm13 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 196,65,101,219,210 ; vpand %ymm10,%ymm3,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 197,45,235,211 ; vpor %ymm3,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,243 ; vpbroadcastd %xmm3,%ymm14 - DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,219 ; vpbroadcastd %xmm3,%ymm11 - DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,235 ; vpbroadcastd %xmm3,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 197,172,88,219 ; vaddps %ymm3,%ymm10,%ymm3 - DB 197,148,94,219 ; vdivps %ymm3,%ymm13,%ymm3 - DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 - DB 197,28,89,219 ; vmulps %ymm3,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + DB 196,98,125,24,37,249,33,0,0 ; vbroadcastss 0x21f9(%rip),%ymm12 # 423c <_sk_callback_hsw+0x1b2> + DB 196,98,125,24,45,244,33,0,0 ; vbroadcastss 0x21f4(%rip),%ymm13 # 4240 <_sk_callback_hsw+0x1b6> + DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,234,33,0,0 ; vbroadcastss 0x21ea(%rip),%ymm13 # 4244 <_sk_callback_hsw+0x1ba> + DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 + DB 196,98,125,24,45,224,33,0,0 ; vbroadcastss 0x21e0(%rip),%ymm13 # 4248 <_sk_callback_hsw+0x1be> + DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 + DB 196,98,125,24,29,214,33,0,0 ; vbroadcastss 0x21d6(%rip),%ymm11 # 424c <_sk_callback_hsw+0x1c2> + DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 + DB 196,98,125,24,37,204,33,0,0 ; vbroadcastss 0x21cc(%rip),%ymm12 # 4250 <_sk_callback_hsw+0x1c6> + DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 + DB 196,98,125,24,37,194,33,0,0 ; vbroadcastss 0x21c2(%rip),%ymm12 # 4254 <_sk_callback_hsw+0x1ca> + DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,227 ; vpbroadcastd %xmm3,%ymm12 - DB 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,235 ; vpbroadcastd %xmm3,%ymm13 - DB 196,66,45,172,235 ; vfnmadd213ps %ymm11,%ymm10,%ymm13 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,219 ; vpbroadcastd %xmm3,%ymm11 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 + 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 196,98,125,24,29,163,33,0,0 ; vbroadcastss 0x21a3(%rip),%ymm11 # 4258 <_sk_callback_hsw+0x1ce> + DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 + DB 196,98,125,24,29,153,33,0,0 ; vbroadcastss 0x2199(%rip),%ymm11 # 425c <_sk_callback_hsw+0x1d2> + DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11 + DB 196,226,125,24,29,143,33,0,0 ; vbroadcastss 0x218f(%rip),%ymm3 # 4260 <_sk_callback_hsw+0x1d6> DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3 - DB 197,164,94,219 ; vdivps %ymm3,%ymm11,%ymm3 - DB 197,148,88,219 ; vaddps %ymm3,%ymm13,%ymm3 - DB 197,156,89,219 ; vmulps %ymm3,%ymm12,%ymm3 + DB 196,98,125,24,21,133,33,0,0 ; vbroadcastss 0x2185(%rip),%ymm10 # 4264 <_sk_callback_hsw+0x1da> + DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3 + DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 + DB 196,98,125,24,21,120,33,0,0 ; vbroadcastss 0x2178(%rip),%ymm10 # 4268 <_sk_callback_hsw+0x1de> + 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 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 @@ -2237,7 +2141,7 @@ _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 117,50 ; jne 23d1 <_sk_load_a8_hsw+0x42> + DB 117,50 ; jne 2299 <_sk_load_a8_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 @@ -2260,9 +2164,9 @@ _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 23d9 <_sk_load_a8_hsw+0x4a> + DB 117,234 ; jne 22a1 <_sk_load_a8_hsw+0x4a> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,173 ; jmp 23a3 <_sk_load_a8_hsw+0x14> + DB 235,173 ; jmp 226b <_sk_load_a8_hsw+0x14> PUBLIC _sk_gather_a8_hsw _sk_gather_a8_hsw LABEL PROC @@ -2333,7 +2237,7 @@ _sk_store_a8_hsw LABEL PROC 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 250e <_sk_store_a8_hsw+0x3b> + DB 117,10 ; jne 23d6 <_sk_store_a8_hsw+0x3b> DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2341,10 +2245,10 @@ _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 250a <_sk_store_a8_hsw+0x37> + DB 119,236 ; ja 23d2 <_sk_store_a8_hsw+0x37> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2570 <_sk_store_a8_hsw+0x9d> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2438 <_sk_store_a8_hsw+0x9d> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2355,7 +2259,7 @@ _sk_store_a8_hsw LABEL PROC DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,154 ; jmp 250a <_sk_store_a8_hsw+0x37> + DB 235,154 ; jmp 23d2 <_sk_store_a8_hsw+0x37> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -2386,7 +2290,7 @@ _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 117,60 ; jne 25d8 <_sk_load_g8_hsw+0x4c> + DB 117,60 ; jne 24a0 <_sk_load_g8_hsw+0x4c> 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 @@ -2411,9 +2315,9 @@ _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 25e0 <_sk_load_g8_hsw+0x54> + DB 117,234 ; jne 24a8 <_sk_load_g8_hsw+0x54> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,163 ; jmp 25a0 <_sk_load_g8_hsw+0x14> + DB 235,163 ; jmp 2468 <_sk_load_g8_hsw+0x14> PUBLIC _sk_gather_g8_hsw _sk_gather_g8_hsw LABEL PROC @@ -2478,9 +2382,9 @@ _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 26f3 <_sk_gather_i8_hsw+0xf> + DB 116,5 ; je 25bb <_sk_gather_i8_hsw+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 26f5 <_sk_gather_i8_hsw+0x11> + DB 235,2 ; jmp 25bd <_sk_gather_i8_hsw+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -2551,7 +2455,7 @@ _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 15,133,149,0,0,0 ; jne 28a7 <_sk_load_565_hsw+0xa3> + DB 15,133,149,0,0,0 ; jne 276f <_sk_load_565_hsw+0xa3> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 DB 184,0,248,0,0 ; mov $0xf800,%eax @@ -2591,9 +2495,9 @@ _sk_load_565_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 15,135,89,255,255,255 ; ja 2818 <_sk_load_565_hsw+0x14> + DB 15,135,89,255,255,255 ; ja 26e0 <_sk_load_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 2914 <_sk_load_565_hsw+0x110> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 27dc <_sk_load_565_hsw+0x110> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2605,12 +2509,12 @@ _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 233,5,255,255,255 ; jmpq 2818 <_sk_load_565_hsw+0x14> + DB 233,5,255,255,255 ; jmpq 26e0 <_sk_load_565_hsw+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 2919 <_sk_load_565_hsw+0x115> + DB 235,255 ; jmp 27e1 <_sk_load_565_hsw+0x115> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -2733,7 +2637,7 @@ _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 2adf <_sk_store_565_hsw+0x6c> + DB 117,10 ; jne 29a7 <_sk_store_565_hsw+0x6c> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2741,9 +2645,9 @@ _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 2adb <_sk_store_565_hsw+0x68> + DB 119,236 ; ja 29a3 <_sk_store_565_hsw+0x68> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2b3c <_sk_store_565_hsw+0xc9> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2a04 <_sk_store_565_hsw+0xc9> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2754,7 +2658,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 2adb <_sk_store_565_hsw+0x68> + DB 235,159 ; jmp 29a3 <_sk_store_565_hsw+0x68> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -2783,7 +2687,7 @@ _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 15,133,179,0,0,0 ; jne 2c19 <_sk_load_4444_hsw+0xc1> + DB 15,133,179,0,0,0 ; jne 2ae1 <_sk_load_4444_hsw+0xc1> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,98,125,51,200 ; vpmovzxwd %xmm0,%ymm9 DB 184,0,240,0,0 ; mov $0xf000,%eax @@ -2829,9 +2733,9 @@ _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 15,135,59,255,255,255 ; ja 2b6c <_sk_load_4444_hsw+0x14> + DB 15,135,59,255,255,255 ; ja 2a34 <_sk_load_4444_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 2c88 <_sk_load_4444_hsw+0x130> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 2b50 <_sk_load_4444_hsw+0x130> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2843,13 +2747,13 @@ _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 233,231,254,255,255 ; jmpq 2b6c <_sk_load_4444_hsw+0x14> + DB 233,231,254,255,255 ; jmpq 2a34 <_sk_load_4444_hsw+0x14> DB 15,31,0 ; nopl (%rax) DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2002c90 <_sk_callback_hsw+0xffffffffe1ffeace> + DB 233,255,255,255,225 ; jmpq ffffffffe2002b58 <_sk_callback_hsw+0xffffffffe1ffeace> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -2977,7 +2881,7 @@ _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 2e77 <_sk_store_4444_hsw+0x72> + DB 117,10 ; jne 2d3f <_sk_store_4444_hsw+0x72> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2985,9 +2889,9 @@ _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 2e73 <_sk_store_4444_hsw+0x6e> + DB 119,236 ; ja 2d3b <_sk_store_4444_hsw+0x6e> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2ed4 <_sk_store_4444_hsw+0xcf> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2d9c <_sk_store_4444_hsw+0xcf> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2998,7 +2902,7 @@ _sk_store_4444_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 2e73 <_sk_store_4444_hsw+0x6e> + DB 235,159 ; jmp 2d3b <_sk_store_4444_hsw+0x6e> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -3029,7 +2933,7 @@ _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 117,104 ; jne 2f6d <_sk_load_8888_hsw+0x7d> + DB 117,104 ; jne 2e35 <_sk_load_8888_hsw+0x7d> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -3062,7 +2966,7 @@ _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 233,116,255,255,255 ; jmpq 2f0a <_sk_load_8888_hsw+0x1a> + DB 233,116,255,255,255 ; jmpq 2dd2 <_sk_load_8888_hsw+0x1a> PUBLIC _sk_gather_8888_hsw _sk_gather_8888_hsw LABEL PROC @@ -3122,7 +3026,7 @@ _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 3090 <_sk_store_8888_hsw+0x74> + DB 117,12 ; jne 2f58 <_sk_store_8888_hsw+0x74> DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx @@ -3135,14 +3039,14 @@ _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 3089 <_sk_store_8888_hsw+0x6d> + DB 235,211 ; jmp 2f51 <_sk_store_8888_hsw+0x6d> 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 3121 <_sk_load_f16_hsw+0x6b> + DB 117,97 ; jne 2fe9 <_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 @@ -3168,29 +3072,29 @@ _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 3180 <_sk_load_f16_hsw+0xca> + DB 116,79 ; je 3048 <_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 3180 <_sk_load_f16_hsw+0xca> + DB 114,67 ; jb 3048 <_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 318d <_sk_load_f16_hsw+0xd7> + DB 116,68 ; je 3055 <_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 318d <_sk_load_f16_hsw+0xd7> + DB 114,56 ; jb 3055 <_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 30d7 <_sk_load_f16_hsw+0x21> + DB 15,132,114,255,255,255 ; je 2f9f <_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 30d7 <_sk_load_f16_hsw+0x21> + DB 15,130,98,255,255,255 ; jb 2f9f <_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 30d7 <_sk_load_f16_hsw+0x21> + DB 233,87,255,255,255 ; jmpq 2f9f <_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 30d7 <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq 2f9f <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq 30d7 <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 2f9f <_sk_load_f16_hsw+0x21> PUBLIC _sk_gather_f16_hsw _sk_gather_f16_hsw LABEL PROC @@ -3244,7 +3148,7 @@ _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 3285 <_sk_store_f16_hsw+0x65> + DB 117,27 ; jne 314d <_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) @@ -3253,22 +3157,22 @@ _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 3281 <_sk_store_f16_hsw+0x61> + DB 116,241 ; je 3149 <_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 3281 <_sk_store_f16_hsw+0x61> + DB 114,229 ; jb 3149 <_sk_store_f16_hsw+0x61> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je 3281 <_sk_store_f16_hsw+0x61> + DB 116,221 ; je 3149 <_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 3281 <_sk_store_f16_hsw+0x61> + DB 114,209 ; jb 3149 <_sk_store_f16_hsw+0x61> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je 3281 <_sk_store_f16_hsw+0x61> + DB 116,201 ; je 3149 <_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 3281 <_sk_store_f16_hsw+0x61> + DB 114,189 ; jb 3149 <_sk_store_f16_hsw+0x61> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp 3281 <_sk_store_f16_hsw+0x61> + DB 235,181 ; jmp 3149 <_sk_store_f16_hsw+0x61> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_hsw LABEL PROC @@ -3276,7 +3180,7 @@ _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 15,133,205,0,0,0 ; jne 33af <_sk_load_u16_be_hsw+0xe3> + DB 15,133,205,0,0,0 ; jne 3277 <_sk_load_u16_be_hsw+0xe3> 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 @@ -3325,29 +3229,29 @@ _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 3415 <_sk_load_u16_be_hsw+0x149> + DB 116,85 ; je 32dd <_sk_load_u16_be_hsw+0x149> 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 3415 <_sk_load_u16_be_hsw+0x149> + DB 114,72 ; jb 32dd <_sk_load_u16_be_hsw+0x149> 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 3422 <_sk_load_u16_be_hsw+0x156> + DB 116,72 ; je 32ea <_sk_load_u16_be_hsw+0x156> 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 3422 <_sk_load_u16_be_hsw+0x156> + DB 114,59 ; jb 32ea <_sk_load_u16_be_hsw+0x156> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,5,255,255,255 ; je 32fd <_sk_load_u16_be_hsw+0x31> + DB 15,132,5,255,255,255 ; je 31c5 <_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 15,130,244,254,255,255 ; jb 32fd <_sk_load_u16_be_hsw+0x31> + DB 15,130,244,254,255,255 ; jb 31c5 <_sk_load_u16_be_hsw+0x31> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,232,254,255,255 ; jmpq 32fd <_sk_load_u16_be_hsw+0x31> + DB 233,232,254,255,255 ; jmpq 31c5 <_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 233,219,254,255,255 ; jmpq 32fd <_sk_load_u16_be_hsw+0x31> + DB 233,219,254,255,255 ; jmpq 31c5 <_sk_load_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,210,254,255,255 ; jmpq 32fd <_sk_load_u16_be_hsw+0x31> + DB 233,210,254,255,255 ; jmpq 31c5 <_sk_load_u16_be_hsw+0x31> PUBLIC _sk_load_rgb_u16_be_hsw _sk_load_rgb_u16_be_hsw LABEL PROC @@ -3355,7 +3259,7 @@ _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 15,133,211,0,0,0 ; jne 3510 <_sk_load_rgb_u16_be_hsw+0xe5> + DB 15,133,211,0,0,0 ; jne 33d8 <_sk_load_rgb_u16_be_hsw+0xe5> 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 @@ -3405,36 +3309,36 @@ _sk_load_rgb_u16_be_hsw LABEL PROC 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 3529 <_sk_load_rgb_u16_be_hsw+0xfe> - DB 233,72,255,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,5 ; jne 33f1 <_sk_load_rgb_u16_be_hsw+0xfe> + DB 233,72,255,255,255 ; jmpq 3339 <_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 3558 <_sk_load_rgb_u16_be_hsw+0x12d> + DB 114,26 ; jb 3420 <_sk_load_rgb_u16_be_hsw+0x12d> 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 355d <_sk_load_rgb_u16_be_hsw+0x132> - DB 233,25,255,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,20,255,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 3425 <_sk_load_rgb_u16_be_hsw+0x132> + DB 233,25,255,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,20,255,255,255 ; jmpq 3339 <_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 358c <_sk_load_rgb_u16_be_hsw+0x161> + DB 114,26 ; jb 3454 <_sk_load_rgb_u16_be_hsw+0x161> 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 3591 <_sk_load_rgb_u16_be_hsw+0x166> - DB 233,229,254,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,224,254,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 3459 <_sk_load_rgb_u16_be_hsw+0x166> + DB 233,229,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,224,254,255,255 ; jmpq 3339 <_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 35ba <_sk_load_rgb_u16_be_hsw+0x18f> + DB 114,20 ; jb 3482 <_sk_load_rgb_u16_be_hsw+0x18f> 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 233,183,254,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,178,254,255,255 ; jmpq 3471 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,183,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,178,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46> PUBLIC _sk_store_u16_be_hsw _sk_store_u16_be_hsw LABEL PROC @@ -3481,7 +3385,7 @@ _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 36ba <_sk_store_u16_be_hsw+0xfb> + DB 117,31 ; jne 3582 <_sk_store_u16_be_hsw+0xfb> DB 196,1,120,17,28,72 ; vmovups %xmm11,(%r8,%r9,2) DB 196,1,120,17,84,72,16 ; vmovups %xmm10,0x10(%r8,%r9,2) DB 196,1,120,17,76,72,32 ; vmovups %xmm9,0x20(%r8,%r9,2) @@ -3490,31 +3394,31 @@ _sk_store_u16_be_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,1,121,214,28,72 ; vmovq %xmm11,(%r8,%r9,2) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 116,240 ; je 357e <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,23,92,72,8 ; vmovhpd %xmm11,0x8(%r8,%r9,2) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 114,227 ; jb 357e <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,214,84,72,16 ; vmovq %xmm10,0x10(%r8,%r9,2) - DB 116,218 ; je 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 116,218 ; je 357e <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,23,84,72,24 ; vmovhpd %xmm10,0x18(%r8,%r9,2) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 114,205 ; jb 357e <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,214,76,72,32 ; vmovq %xmm9,0x20(%r8,%r9,2) - DB 116,196 ; je 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 116,196 ; je 357e <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,23,76,72,40 ; vmovhpd %xmm9,0x28(%r8,%r9,2) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 114,183 ; jb 357e <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,214,68,72,48 ; vmovq %xmm8,0x30(%r8,%r9,2) - DB 235,174 ; jmp 36b6 <_sk_store_u16_be_hsw+0xf7> + DB 235,174 ; jmp 357e <_sk_store_u16_be_hsw+0xf7> 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 377e <_sk_load_f32_hsw+0x76> + DB 119,110 ; ja 3646 <_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 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 37a8 <_sk_load_f32_hsw+0xa0> + DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 3670 <_sk_load_f32_hsw+0xa0> DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -3571,7 +3475,7 @@ _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 3835 <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 36fd <_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 @@ -3584,22 +3488,22 @@ _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 3831 <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 36f9 <_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 3831 <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 36f9 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 3831 <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 36f9 <_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 3831 <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 36f9 <_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 3831 <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 36f9 <_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 3831 <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 36f9 <_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 3831 <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 36f9 <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -3840,7 +3744,7 @@ _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 3cc1 <_sk_linear_gradient_hsw+0xb5> + DB 15,132,143,0,0,0 ; je 3b89 <_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 @@ -3867,8 +3771,8 @@ _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 3c4b <_sk_linear_gradient_hsw+0x3f> - DB 235,17 ; jmp 3cd2 <_sk_linear_gradient_hsw+0xc6> + DB 117,140 ; jne 3b13 <_sk_linear_gradient_hsw+0x3f> + DB 235,17 ; jmp 3b9a <_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 @@ -4254,8 +4158,76 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) - DB 128 ; .byte 0x80 - DB 63 ; (bad) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 41b4 <.literal4+0x10> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 422d <.literal4+0x89> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 41e4 <.literal4+0x40> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 425d <.literal4+0xb9> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4214 <.literal4+0x70> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 428d <_sk_callback_hsw+0x203> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4244 <.literal4+0xa0> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 42bd <_sk_callback_hsw+0x233> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0 ; .byte 0x0 + DB 75 ; rex.WXB ALIGN 32 PUBLIC _sk_start_pipeline_avx @@ -4355,14 +4327,14 @@ _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 196,226,125,24,13,100,93,0,0 ; vbroadcastss 0x5d64(%rip),%ymm1 # 5ec4 <_sk_callback_avx+0x11a> + DB 196,226,125,24,13,164,91,0,0 ; vbroadcastss 0x5ba4(%rip),%ymm1 # 5d04 <_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 196,226,125,24,21,72,93,0,0 ; vbroadcastss 0x5d48(%rip),%ymm2 # 5ec8 <_sk_callback_avx+0x11e> + DB 196,226,125,24,21,136,91,0,0 ; vbroadcastss 0x5b88(%rip),%ymm2 # 5d08 <_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 @@ -6719,76 +6691,40 @@ _sk_parametric_r_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 - DB 196,65,124,88,211 ; vaddps %ymm11,%ymm0,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 - DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - 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 197,36,89,216 ; vmulps %ymm0,%ymm11,%ymm11 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 196,65,124,84,210 ; vandps %ymm10,%ymm0,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,44,86,208 ; vorps %ymm0,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - 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 197,36,92,216 ; vsubps %ymm0,%ymm11,%ymm11 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - 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 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 - DB 197,36,92,216 ; vsubps %ymm0,%ymm11,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%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 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - 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 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 - DB 197,148,94,192 ; vdivps %ymm0,%ymm13,%ymm0 + 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 196,98,125,24,37,210,52,0,0 ; vbroadcastss 0x34d2(%rip),%ymm12 # 5d0c <_sk_callback_avx+0x122> + DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,200,52,0,0 ; vbroadcastss 0x34c8(%rip),%ymm12 # 5d10 <_sk_callback_avx+0x126> + DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0 + DB 196,98,125,24,37,190,52,0,0 ; vbroadcastss 0x34be(%rip),%ymm12 # 5d14 <_sk_callback_avx+0x12a> + DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0 + DB 196,98,125,24,37,180,52,0,0 ; vbroadcastss 0x34b4(%rip),%ymm12 # 5d18 <_sk_callback_avx+0x12e> + DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,170,52,0,0 ; vbroadcastss 0x34aa(%rip),%ymm12 # 5d1c <_sk_callback_avx+0x132> + DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 + DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,155,52,0,0 ; vbroadcastss 0x349b(%rip),%ymm12 # 5d20 <_sk_callback_avx+0x136> + DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 + DB 196,98,125,24,37,145,52,0,0 ; vbroadcastss 0x3491(%rip),%ymm12 # 5d24 <_sk_callback_avx+0x13a> + DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 - DB 197,28,89,216 ; vmulps %ymm0,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + 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 196,98,125,24,29,117,52,0,0 ; vbroadcastss 0x3475(%rip),%ymm11 # 5d28 <_sk_callback_avx+0x13e> + DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 + DB 196,98,125,24,29,107,52,0,0 ; vbroadcastss 0x346b(%rip),%ymm11 # 5d2c <_sk_callback_avx+0x142> + DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 + DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 + DB 196,98,125,24,29,92,52,0,0 ; vbroadcastss 0x345c(%rip),%ymm11 # 5d30 <_sk_callback_avx+0x146> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%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 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - 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 196,65,124,88,219 ; vaddps %ymm11,%ymm0,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - 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 196,98,125,24,29,82,52,0,0 ; vbroadcastss 0x3452(%rip),%ymm11 # 5d34 <_sk_callback_avx+0x14a> + DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 + DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 + DB 196,98,125,24,21,67,52,0,0 ; vbroadcastss 0x3443(%rip),%ymm10 # 5d38 <_sk_callback_avx+0x14e> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 - DB 197,36,92,216 ; vsubps %ymm0,%ymm11,%ymm11 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%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 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - 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 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 197,148,94,192 ; vdivps %ymm0,%ymm13,%ymm0 - DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 197,156,89,192 ; vmulps %ymm0,%ymm12,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 @@ -6815,76 +6751,40 @@ _sk_parametric_g_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 - DB 196,65,116,88,211 ; vaddps %ymm11,%ymm1,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 - DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,36,89,217 ; vmulps %ymm1,%ymm11,%ymm11 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 196,65,116,84,210 ; vandps %ymm10,%ymm1,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,44,86,209 ; vorps %ymm1,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 - DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,99,117,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1 - DB 197,148,94,201 ; vdivps %ymm1,%ymm13,%ymm1 + 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 196,98,125,24,37,188,51,0,0 ; vbroadcastss 0x33bc(%rip),%ymm12 # 5d3c <_sk_callback_avx+0x152> + DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,178,51,0,0 ; vbroadcastss 0x33b2(%rip),%ymm12 # 5d40 <_sk_callback_avx+0x156> + DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1 + DB 196,98,125,24,37,168,51,0,0 ; vbroadcastss 0x33a8(%rip),%ymm12 # 5d44 <_sk_callback_avx+0x15a> + DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1 + DB 196,98,125,24,37,158,51,0,0 ; vbroadcastss 0x339e(%rip),%ymm12 # 5d48 <_sk_callback_avx+0x15e> + DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,148,51,0,0 ; vbroadcastss 0x3394(%rip),%ymm12 # 5d4c <_sk_callback_avx+0x162> + DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12 + DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,133,51,0,0 ; vbroadcastss 0x3385(%rip),%ymm12 # 5d50 <_sk_callback_avx+0x166> + DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1 + DB 196,98,125,24,37,123,51,0,0 ; vbroadcastss 0x337b(%rip),%ymm12 # 5d54 <_sk_callback_avx+0x16a> + DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1 DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 - DB 197,28,89,217 ; vmulps %ymm1,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + 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 196,98,125,24,29,95,51,0,0 ; vbroadcastss 0x335f(%rip),%ymm11 # 5d58 <_sk_callback_avx+0x16e> + DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 + DB 196,98,125,24,29,85,51,0,0 ; vbroadcastss 0x3355(%rip),%ymm11 # 5d5c <_sk_callback_avx+0x172> + DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 + DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 + DB 196,98,125,24,29,70,51,0,0 ; vbroadcastss 0x3346(%rip),%ymm11 # 5d60 <_sk_callback_avx+0x176> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,99,117,24,225,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 - DB 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 196,65,116,88,219 ; vaddps %ymm11,%ymm1,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 + DB 196,98,125,24,29,60,51,0,0 ; vbroadcastss 0x333c(%rip),%ymm11 # 5d64 <_sk_callback_avx+0x17a> + DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 + DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 + DB 196,98,125,24,21,45,51,0,0 ; vbroadcastss 0x332d(%rip),%ymm10 # 5d68 <_sk_callback_avx+0x17e> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 - DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,99,117,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 - DB 197,148,94,201 ; vdivps %ymm1,%ymm13,%ymm1 - DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 197,156,89,201 ; vmulps %ymm1,%ymm12,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 @@ -6911,76 +6811,40 @@ _sk_parametric_b_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 - DB 196,65,108,88,211 ; vaddps %ymm11,%ymm2,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 - DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - 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 197,36,89,218 ; vmulps %ymm2,%ymm11,%ymm11 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 - DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 196,65,108,84,210 ; vandps %ymm10,%ymm2,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 - DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,44,86,210 ; vorps %ymm2,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - 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 197,36,92,218 ; vsubps %ymm2,%ymm11,%ymm11 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - 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 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 - DB 197,36,92,218 ; vsubps %ymm2,%ymm11,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,99,109,24,234,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - 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 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 - DB 197,148,94,210 ; vdivps %ymm2,%ymm13,%ymm2 + 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 196,98,125,24,37,166,50,0,0 ; vbroadcastss 0x32a6(%rip),%ymm12 # 5d6c <_sk_callback_avx+0x182> + DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,156,50,0,0 ; vbroadcastss 0x329c(%rip),%ymm12 # 5d70 <_sk_callback_avx+0x186> + DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2 + DB 196,98,125,24,37,146,50,0,0 ; vbroadcastss 0x3292(%rip),%ymm12 # 5d74 <_sk_callback_avx+0x18a> + DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2 + DB 196,98,125,24,37,136,50,0,0 ; vbroadcastss 0x3288(%rip),%ymm12 # 5d78 <_sk_callback_avx+0x18e> + DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,126,50,0,0 ; vbroadcastss 0x327e(%rip),%ymm12 # 5d7c <_sk_callback_avx+0x192> + DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12 + DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,111,50,0,0 ; vbroadcastss 0x326f(%rip),%ymm12 # 5d80 <_sk_callback_avx+0x196> + DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 + DB 196,98,125,24,37,101,50,0,0 ; vbroadcastss 0x3265(%rip),%ymm12 # 5d84 <_sk_callback_avx+0x19a> + DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2 DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 - DB 197,28,89,218 ; vmulps %ymm2,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + 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 196,98,125,24,29,73,50,0,0 ; vbroadcastss 0x3249(%rip),%ymm11 # 5d88 <_sk_callback_avx+0x19e> + DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 + DB 196,98,125,24,29,63,50,0,0 ; vbroadcastss 0x323f(%rip),%ymm11 # 5d8c <_sk_callback_avx+0x1a2> + DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 + DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 + DB 196,98,125,24,29,48,50,0,0 ; vbroadcastss 0x3230(%rip),%ymm11 # 5d90 <_sk_callback_avx+0x1a6> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,99,109,24,226,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm12 - DB 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - 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 196,65,108,88,219 ; vaddps %ymm11,%ymm2,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - 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 196,98,125,24,29,38,50,0,0 ; vbroadcastss 0x3226(%rip),%ymm11 # 5d94 <_sk_callback_avx+0x1aa> + DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 + DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 + DB 196,98,125,24,21,23,50,0,0 ; vbroadcastss 0x3217(%rip),%ymm10 # 5d98 <_sk_callback_avx+0x1ae> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 - DB 197,36,92,218 ; vsubps %ymm2,%ymm11,%ymm11 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,99,109,24,234,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - 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 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2 - DB 197,148,94,210 ; vdivps %ymm2,%ymm13,%ymm2 - DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 197,156,89,210 ; vmulps %ymm2,%ymm12,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 @@ -7007,76 +6871,40 @@ _sk_parametric_a_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 - DB 196,65,100,88,211 ; vaddps %ymm11,%ymm3,%ymm10 - DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 - DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d - 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 197,36,89,219 ; vmulps %ymm3,%ymm11,%ymm11 - DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3 - DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 196,65,100,84,210 ; vandps %ymm10,%ymm3,%ymm10 - DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3 - DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,44,86,211 ; vorps %ymm3,%ymm10,%ymm10 - DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d - 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 197,36,92,219 ; vsubps %ymm3,%ymm11,%ymm11 - DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d - 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 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 - DB 197,36,92,219 ; vsubps %ymm3,%ymm11,%ymm11 - DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 - DB 196,99,101,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 - DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d - 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 197,172,88,219 ; vaddps %ymm3,%ymm10,%ymm3 - DB 197,148,94,219 ; vdivps %ymm3,%ymm13,%ymm3 + 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 196,98,125,24,37,144,49,0,0 ; vbroadcastss 0x3190(%rip),%ymm12 # 5d9c <_sk_callback_avx+0x1b2> + DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,134,49,0,0 ; vbroadcastss 0x3186(%rip),%ymm12 # 5da0 <_sk_callback_avx+0x1b6> + DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3 + DB 196,98,125,24,37,124,49,0,0 ; vbroadcastss 0x317c(%rip),%ymm12 # 5da4 <_sk_callback_avx+0x1ba> + DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3 + DB 196,98,125,24,37,114,49,0,0 ; vbroadcastss 0x3172(%rip),%ymm12 # 5da8 <_sk_callback_avx+0x1be> + DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,104,49,0,0 ; vbroadcastss 0x3168(%rip),%ymm12 # 5dac <_sk_callback_avx+0x1c2> + DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12 + DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 + DB 196,98,125,24,37,89,49,0,0 ; vbroadcastss 0x3159(%rip),%ymm12 # 5db0 <_sk_callback_avx+0x1c6> + DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3 + DB 196,98,125,24,37,79,49,0,0 ; vbroadcastss 0x314f(%rip),%ymm12 # 5db4 <_sk_callback_avx+0x1ca> + DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3 DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 - DB 197,28,89,219 ; vmulps %ymm3,%ymm12,%ymm11 - DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 + 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 196,98,125,24,29,51,49,0,0 ; vbroadcastss 0x3133(%rip),%ymm11 # 5db8 <_sk_callback_avx+0x1ce> + DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 + DB 196,98,125,24,29,41,49,0,0 ; vbroadcastss 0x3129(%rip),%ymm11 # 5dbc <_sk_callback_avx+0x1d2> + DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 + DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3 + DB 196,98,125,24,29,26,49,0,0 ; vbroadcastss 0x311a(%rip),%ymm11 # 5dc0 <_sk_callback_avx+0x1d6> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%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 65,184,81,140,242,66 ; mov $0x42f28c51,%r8d - 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 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 - DB 65,184,141,188,190,63 ; mov $0x3fbebc8d,%r8d - 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 196,98,125,24,29,16,49,0,0 ; vbroadcastss 0x3110(%rip),%ymm11 # 5dc4 <_sk_callback_avx+0x1da> + DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 + DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 + DB 196,98,125,24,21,1,49,0,0 ; vbroadcastss 0x3101(%rip),%ymm10 # 5dc8 <_sk_callback_avx+0x1de> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 - DB 197,36,92,219 ; vsubps %ymm3,%ymm11,%ymm11 - DB 65,184,254,210,221,65 ; mov $0x41ddd2fe,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 - DB 196,99,101,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 - DB 65,184,248,245,154,64 ; mov $0x409af5f8,%r8d - 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 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3 - DB 197,148,94,219 ; vdivps %ymm3,%ymm13,%ymm3 - DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 197,156,89,219 ; vmulps %ymm3,%ymm12,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 @@ -7182,7 +7010,7 @@ _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 117,74 ; jne 3160 <_sk_load_a8_avx+0x5a> + DB 117,74 ; jne 2eec <_sk_load_a8_avx+0x5a> 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 @@ -7209,9 +7037,9 @@ _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 3168 <_sk_load_a8_avx+0x62> + DB 117,234 ; jne 2ef4 <_sk_load_a8_avx+0x62> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,149 ; jmp 311a <_sk_load_a8_avx+0x14> + DB 235,149 ; jmp 2ea6 <_sk_load_a8_avx+0x14> PUBLIC _sk_gather_a8_avx _sk_gather_a8_avx LABEL PROC @@ -7288,7 +7116,7 @@ _sk_store_a8_avx LABEL PROC 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 32c1 <_sk_store_a8_avx+0x42> + DB 117,10 ; jne 304d <_sk_store_a8_avx+0x42> DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -7296,10 +7124,10 @@ _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 32bd <_sk_store_a8_avx+0x3e> + DB 119,236 ; ja 3049 <_sk_store_a8_avx+0x3e> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 3324 <_sk_store_a8_avx+0xa5> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 30b0 <_sk_store_a8_avx+0xa5> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -7310,7 +7138,7 @@ _sk_store_a8_avx LABEL PROC DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,154 ; jmp 32bd <_sk_store_a8_avx+0x3e> + DB 235,154 ; jmp 3049 <_sk_store_a8_avx+0x3e> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -7342,7 +7170,7 @@ _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 117,91 ; jne 33ab <_sk_load_g8_avx+0x6b> + DB 117,91 ; jne 3137 <_sk_load_g8_avx+0x6b> 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 @@ -7372,9 +7200,9 @@ _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 33b3 <_sk_load_g8_avx+0x73> + DB 117,234 ; jne 313f <_sk_load_g8_avx+0x73> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,132 ; jmp 3354 <_sk_load_g8_avx+0x14> + DB 235,132 ; jmp 30e0 <_sk_load_g8_avx+0x14> PUBLIC _sk_gather_g8_avx _sk_gather_g8_avx LABEL PROC @@ -7445,9 +7273,9 @@ _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 34ea <_sk_gather_i8_avx+0xf> + DB 116,5 ; je 3276 <_sk_gather_i8_avx+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 34ec <_sk_gather_i8_avx+0x11> + DB 235,2 ; jmp 3278 <_sk_gather_i8_avx+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -7550,7 +7378,7 @@ _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 15,133,209,0,0,0 ; jne 3786 <_sk_load_565_avx+0xdf> + DB 15,133,209,0,0,0 ; jne 3512 <_sk_load_565_avx+0xdf> 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 @@ -7600,9 +7428,9 @@ _sk_load_565_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 15,135,29,255,255,255 ; ja 36bb <_sk_load_565_avx+0x14> + DB 15,135,29,255,255,255 ; ja 3447 <_sk_load_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 37f4 <_sk_load_565_avx+0x14d> + DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 3580 <_sk_load_565_avx+0x14d> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7614,7 +7442,7 @@ _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 233,201,254,255,255 ; jmpq 36bb <_sk_load_565_avx+0x14> + DB 233,201,254,255,255 ; jmpq 3447 <_sk_load_565_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 242,255 ; repnz (bad) DB 255 ; (bad) @@ -7767,7 +7595,7 @@ _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 3a3f <_sk_store_565_avx+0x9e> + DB 117,10 ; jne 37cb <_sk_store_565_avx+0x9e> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -7775,9 +7603,9 @@ _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 3a3b <_sk_store_565_avx+0x9a> + DB 119,236 ; ja 37c7 <_sk_store_565_avx+0x9a> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 3a9c <_sk_store_565_avx+0xfb> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 3828 <_sk_store_565_avx+0xfb> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -7788,7 +7616,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 3a3b <_sk_store_565_avx+0x9a> + DB 235,159 ; jmp 37c7 <_sk_store_565_avx+0x9a> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -7817,7 +7645,7 @@ _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 15,133,245,0,0,0 ; jne 3bbb <_sk_load_4444_avx+0x103> + DB 15,133,245,0,0,0 ; jne 3947 <_sk_load_4444_avx+0x103> 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 @@ -7874,9 +7702,9 @@ _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 15,135,249,254,255,255 ; ja 3acc <_sk_load_4444_avx+0x14> + DB 15,135,249,254,255,255 ; ja 3858 <_sk_load_4444_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 3c28 <_sk_load_4444_avx+0x170> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 39b4 <_sk_load_4444_avx+0x170> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7888,12 +7716,12 @@ _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 233,165,254,255,255 ; jmpq 3acc <_sk_load_4444_avx+0x14> + DB 233,165,254,255,255 ; jmpq 3858 <_sk_load_4444_avx+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 3c2d <_sk_load_4444_avx+0x175> + DB 235,255 ; jmp 39b9 <_sk_load_4444_avx+0x175> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -8050,7 +7878,7 @@ _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 3ea8 <_sk_store_4444_avx+0xaf> + DB 117,10 ; jne 3c34 <_sk_store_4444_avx+0xaf> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8058,9 +7886,9 @@ _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 3ea4 <_sk_store_4444_avx+0xab> + DB 119,236 ; ja 3c30 <_sk_store_4444_avx+0xab> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 3f08 <_sk_store_4444_avx+0x10f> + DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 3c94 <_sk_store_4444_avx+0x10f> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -8071,7 +7899,7 @@ _sk_store_4444_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 3ea4 <_sk_store_4444_avx+0xab> + DB 235,159 ; jmp 3c30 <_sk_store_4444_avx+0xab> DB 15,31,0 ; nopl (%rax) DB 244 ; hlt DB 255 ; (bad) @@ -8102,7 +7930,7 @@ _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 15,133,157,0,0,0 ; jne 3fcf <_sk_load_8888_avx+0xab> + DB 15,133,157,0,0,0 ; jne 3d5b <_sk_load_8888_avx+0xab> DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -8140,9 +7968,9 @@ _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 15,135,80,255,255,255 ; ja 3f38 <_sk_load_8888_avx+0x14> + DB 15,135,80,255,255,255 ; ja 3cc4 <_sk_load_8888_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 407c <_sk_load_8888_avx+0x158> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 3e08 <_sk_load_8888_avx+0x158> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -8165,7 +7993,7 @@ _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 233,188,254,255,255 ; jmpq 3f38 <_sk_load_8888_avx+0x14> + DB 233,188,254,255,255 ; jmpq 3cc4 <_sk_load_8888_avx+0x14> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -8291,7 +8119,7 @@ _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 427d <_sk_store_8888_avx+0xa4> + DB 117,10 ; jne 4009 <_sk_store_8888_avx+0xa4> DB 196,65,124,17,4,185 ; vmovups %ymm8,(%r9,%rdi,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8299,9 +8127,9 @@ _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 4279 <_sk_store_8888_avx+0xa0> + DB 119,236 ; ja 4005 <_sk_store_8888_avx+0xa0> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 42ec <_sk_store_8888_avx+0x113> + DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 4078 <_sk_store_8888_avx+0x113> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -8315,7 +8143,7 @@ _sk_store_8888_avx LABEL PROC DB 196,67,121,22,68,185,8,2 ; vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) DB 196,67,121,22,68,185,4,1 ; vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) DB 196,65,121,126,4,185 ; vmovd %xmm8,(%r9,%rdi,4) - DB 235,143 ; jmp 4279 <_sk_store_8888_avx+0xa0> + DB 235,143 ; jmp 4005 <_sk_store_8888_avx+0xa0> DB 102,144 ; xchg %ax,%ax DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -8342,14 +8170,15 @@ _sk_store_8888_avx LABEL PROC PUBLIC _sk_load_f16_avx _sk_load_f16_avx LABEL PROC - DB 72,131,236,120 ; sub $0x78,%rsp + DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 197,252,17,124,36,64 ; vmovups %ymm7,0x40(%rsp) - DB 197,252,17,116,36,32 ; vmovups %ymm6,0x20(%rsp) - DB 197,252,17,44,36 ; vmovups %ymm5,(%rsp) - DB 15,133,49,2,0,0 ; jne 455c <_sk_load_f16_avx+0x254> + DB 197,252,17,124,36,96 ; vmovups %ymm7,0x60(%rsp) + DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) + DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) + DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) + DB 15,133,143,2,0,0 ; jne 434f <_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 @@ -8358,138 +8187,151 @@ _sk_load_f16_avx LABEL PROC DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,113,97,217 ; vpunpcklwd %xmm9,%xmm1,%xmm3 DB 196,193,113,105,201 ; vpunpckhwd %xmm9,%xmm1,%xmm1 - DB 197,121,97,242 ; vpunpcklwd %xmm2,%xmm0,%xmm14 + DB 197,121,97,250 ; vpunpcklwd %xmm2,%xmm0,%xmm15 DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8 - DB 197,97,97,249 ; vpunpcklwd %xmm1,%xmm3,%xmm15 - DB 197,97,105,217 ; vpunpckhwd %xmm1,%xmm3,%xmm11 - DB 196,193,9,108,199 ; vpunpcklqdq %xmm15,%xmm14,%xmm0 - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,193,121,105,201 ; vpunpckhwd %xmm9,%xmm0,%xmm1 + DB 197,225,97,209 ; vpunpcklwd %xmm1,%xmm3,%xmm2 + DB 197,97,105,201 ; vpunpckhwd %xmm1,%xmm3,%xmm9 + DB 197,129,108,194 ; vpunpcklqdq %xmm2,%xmm15,%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 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 184,0,128,0,0 ; mov $0x8000,%eax - DB 197,249,110,200 ; vmovd %eax,%xmm1 - DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 - DB 196,99,117,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm10 - DB 196,193,124,84,202 ; vandps %ymm10,%ymm0,%ymm1 + DB 196,98,125,24,37,179,28,0,0 ; vbroadcastss 0x1cb3(%rip),%ymm12 # 5dcc <_sk_callback_avx+0x1e2> + DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 - DB 184,0,4,0,0 ; mov $0x400,%eax - DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 - DB 197,249,110,208 ; vmovd %eax,%xmm2 - DB 197,121,112,226,0 ; vpshufd $0x0,%xmm2,%xmm12 - DB 197,153,102,211 ; vpcmpgtd %xmm3,%xmm12,%xmm2 - DB 197,25,102,232 ; vpcmpgtd %xmm0,%xmm12,%xmm13 - DB 196,227,21,24,242,1 ; vinsertf128 $0x1,%xmm2,%ymm13,%ymm6 - DB 196,227,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm2 - DB 197,145,114,242,16 ; vpslld $0x10,%xmm2,%xmm13 - DB 197,233,114,243,13 ; vpslld $0xd,%xmm3,%xmm2 - DB 184,0,0,0,56 ; mov $0x38000000,%eax - DB 197,249,110,216 ; vmovd %eax,%xmm3 - DB 197,249,112,235,0 ; vpshufd $0x0,%xmm3,%xmm5 - DB 197,145,254,253 ; vpaddd %xmm5,%xmm13,%xmm7 - DB 197,193,254,210 ; vpaddd %xmm2,%xmm7,%xmm2 + DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 + DB 196,98,121,24,29,159,28,0,0 ; vbroadcastss 0x1c9f(%rip),%xmm11 # 5dd0 <_sk_callback_avx+0x1e6> + DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 + DB 196,98,121,24,45,149,28,0,0 ; vbroadcastss 0x1c95(%rip),%xmm13 # 5dd4 <_sk_callback_avx+0x1ea> + 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 + DB 196,99,45,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm10,%ymm10 + DB 197,225,114,241,16 ; vpslld $0x10,%xmm1,%xmm3 + DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 DB 197,241,114,241,16 ; vpslld $0x10,%xmm1,%xmm1 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 DB 197,249,114,240,13 ; vpslld $0xd,%xmm0,%xmm0 - DB 197,241,254,205 ; vpaddd %xmm5,%xmm1,%xmm1 - DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - DB 196,65,20,87,237 ; vxorps %ymm13,%ymm13,%ymm13 - DB 196,195,125,74,197,96 ; vblendvps %ymm6,%ymm13,%ymm0,%ymm0 - DB 196,193,9,109,207 ; vpunpckhqdq %xmm15,%xmm14,%xmm1 - DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2 + DB 196,193,97,114,246,13 ; vpslld $0xd,%xmm14,%xmm3 + 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 196,226,121,24,29,75,28,0,0 ; vbroadcastss 0x1c4b(%rip),%xmm3 # 5dd8 <_sk_callback_avx+0x1ee> + 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 + DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 + DB 196,195,125,74,198,160 ; vblendvps %ymm10,%ymm14,%ymm0,%ymm0 + DB 197,129,109,202 ; vpunpckhqdq %xmm2,%xmm15,%xmm1 + DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 + DB 197,241,105,212 ; vpunpckhwd %xmm4,%xmm1,%xmm2 DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,193,116,84,210 ; vandps %ymm10,%ymm1,%ymm2 + DB 196,193,116,84,212 ; vandps %ymm12,%ymm1,%ymm2 DB 197,244,87,202 ; vxorps %ymm2,%ymm1,%ymm1 - DB 196,227,125,25,206,1 ; vextractf128 $0x1,%ymm1,%xmm6 - DB 197,153,102,254 ; vpcmpgtd %xmm6,%xmm12,%xmm7 - DB 197,25,102,241 ; vpcmpgtd %xmm1,%xmm12,%xmm14 - DB 196,99,13,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm14,%ymm14 - DB 196,227,125,25,215,1 ; vextractf128 $0x1,%ymm2,%xmm7 - DB 197,193,114,247,16 ; vpslld $0x10,%xmm7,%xmm7 - DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 - DB 197,193,254,253 ; vpaddd %xmm5,%xmm7,%xmm7 - DB 197,193,254,246 ; vpaddd %xmm6,%xmm7,%xmm6 + DB 196,195,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm10 + DB 196,193,40,87,251 ; vxorps %xmm11,%xmm10,%xmm7 + DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 + DB 196,193,112,87,243 ; vxorps %xmm11,%xmm1,%xmm6 + DB 197,145,102,246 ; vpcmpgtd %xmm6,%xmm13,%xmm6 + DB 196,227,77,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 + DB 197,193,114,242,16 ; vpslld $0x10,%xmm2,%xmm7 + DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 + DB 196,227,69,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm7,%ymm2 DB 197,241,114,241,13 ; vpslld $0xd,%xmm1,%xmm1 - DB 197,233,254,213 ; vpaddd %xmm5,%xmm2,%xmm2 - DB 197,233,254,201 ; vpaddd %xmm1,%xmm2,%xmm1 - DB 196,227,117,24,206,1 ; vinsertf128 $0x1,%xmm6,%ymm1,%ymm1 - DB 196,195,117,74,205,224 ; vblendvps %ymm14,%ymm13,%ymm1,%ymm1 - DB 196,193,57,108,211 ; vpunpcklqdq %xmm11,%xmm8,%xmm2 - DB 196,193,105,105,241 ; vpunpckhwd %xmm9,%xmm2,%xmm6 + DB 196,193,65,114,242,13 ; vpslld $0xd,%xmm10,%xmm7 + DB 196,227,117,24,207,1 ; vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 + DB 197,244,86,202 ; vorps %ymm2,%ymm1,%ymm1 + DB 196,227,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm2 + DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 + DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 + DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 + DB 196,195,117,74,206,96 ; vblendvps %ymm6,%ymm14,%ymm1,%ymm1 + DB 196,193,57,108,209 ; vpunpcklqdq %xmm9,%xmm8,%xmm2 + DB 197,233,105,244 ; vpunpckhwd %xmm4,%xmm2,%xmm6 + DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - DB 196,193,108,84,242 ; vandps %ymm10,%ymm2,%ymm6 + DB 196,193,108,84,244 ; vandps %ymm12,%ymm2,%ymm6 DB 197,236,87,214 ; vxorps %ymm6,%ymm2,%ymm2 - DB 196,195,125,25,214,1 ; vextractf128 $0x1,%ymm2,%xmm14 - DB 196,193,25,102,254 ; vpcmpgtd %xmm14,%xmm12,%xmm7 - DB 197,25,102,250 ; vpcmpgtd %xmm2,%xmm12,%xmm15 - DB 196,99,5,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm15,%ymm15 - DB 196,227,125,25,247,1 ; vextractf128 $0x1,%ymm6,%xmm7 - DB 197,193,114,247,16 ; vpslld $0x10,%xmm7,%xmm7 - DB 196,193,9,114,246,13 ; vpslld $0xd,%xmm14,%xmm14 - DB 197,193,254,253 ; vpaddd %xmm5,%xmm7,%xmm7 - DB 196,193,65,254,254 ; vpaddd %xmm14,%xmm7,%xmm7 + DB 196,227,125,25,215,1 ; vextractf128 $0x1,%ymm2,%xmm7 + DB 196,193,64,87,235 ; vxorps %xmm11,%xmm7,%xmm5 + DB 197,145,102,237 ; vpcmpgtd %xmm5,%xmm13,%xmm5 + DB 196,193,104,87,227 ; vxorps %xmm11,%xmm2,%xmm4 + DB 197,145,102,228 ; vpcmpgtd %xmm4,%xmm13,%xmm4 + DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + DB 197,209,114,246,16 ; vpslld $0x10,%xmm6,%xmm5 + DB 196,227,125,25,246,1 ; vextractf128 $0x1,%ymm6,%xmm6 DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 + DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 - DB 197,201,254,245 ; vpaddd %xmm5,%xmm6,%xmm6 - DB 197,201,254,210 ; vpaddd %xmm2,%xmm6,%xmm2 - DB 196,227,109,24,215,1 ; vinsertf128 $0x1,%xmm7,%ymm2,%ymm2 - DB 196,195,109,74,213,240 ; vblendvps %ymm15,%ymm13,%ymm2,%ymm2 - DB 196,193,57,109,243 ; vpunpckhqdq %xmm11,%xmm8,%xmm6 - DB 196,193,73,105,249 ; vpunpckhwd %xmm9,%xmm6,%xmm7 - DB 196,226,121,51,246 ; vpmovzxwd %xmm6,%xmm6 - DB 196,227,77,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 - DB 196,193,76,84,250 ; vandps %ymm10,%ymm6,%ymm7 - DB 197,204,87,247 ; vxorps %ymm7,%ymm6,%ymm6 - DB 196,195,125,25,240,1 ; vextractf128 $0x1,%ymm6,%xmm8 - DB 196,65,25,102,200 ; vpcmpgtd %xmm8,%xmm12,%xmm9 - DB 197,25,102,214 ; vpcmpgtd %xmm6,%xmm12,%xmm10 - DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,227,125,25,251,1 ; vextractf128 $0x1,%ymm7,%xmm3 - DB 197,225,114,243,16 ; vpslld $0x10,%xmm3,%xmm3 - DB 197,193,114,247,16 ; vpslld $0x10,%xmm7,%xmm7 - DB 197,193,254,253 ; vpaddd %xmm5,%xmm7,%xmm7 - DB 197,225,254,221 ; vpaddd %xmm5,%xmm3,%xmm3 - DB 196,193,81,114,240,13 ; vpslld $0xd,%xmm8,%xmm5 - DB 197,225,254,221 ; vpaddd %xmm5,%xmm3,%xmm3 - DB 197,209,114,246,13 ; vpslld $0xd,%xmm6,%xmm5 - DB 197,193,254,237 ; vpaddd %xmm5,%xmm7,%xmm5 - DB 196,227,85,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm5,%ymm3 - DB 196,195,101,74,221,144 ; vblendvps %ymm9,%ymm13,%ymm3,%ymm3 - DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,44,36 ; vmovups (%rsp),%ymm5 - DB 197,252,16,116,36,32 ; vmovups 0x20(%rsp),%ymm6 - DB 197,252,16,124,36,64 ; vmovups 0x40(%rsp),%ymm7 - DB 72,131,196,120 ; add $0x78,%rsp + DB 197,201,114,247,13 ; vpslld $0xd,%xmm7,%xmm6 + DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 + DB 197,236,86,213 ; vorps %ymm5,%ymm2,%ymm2 + DB 196,227,125,25,213,1 ; vextractf128 $0x1,%ymm2,%xmm5 + DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 + DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 + DB 196,227,109,24,213,1 ; vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 + DB 196,195,109,74,214,64 ; vblendvps %ymm4,%ymm14,%ymm2,%ymm2 + DB 196,193,57,109,225 ; vpunpckhqdq %xmm9,%xmm8,%xmm4 + DB 196,193,89,105,234 ; vpunpckhwd %xmm10,%xmm4,%xmm5 + DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 + DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + DB 196,193,92,84,236 ; vandps %ymm12,%ymm4,%ymm5 + DB 197,220,87,229 ; vxorps %ymm5,%ymm4,%ymm4 + DB 196,227,125,25,230,1 ; vextractf128 $0x1,%ymm4,%xmm6 + DB 196,193,72,87,251 ; vxorps %xmm11,%xmm6,%xmm7 + DB 197,17,102,199 ; vpcmpgtd %xmm7,%xmm13,%xmm8 + DB 196,193,88,87,251 ; vxorps %xmm11,%xmm4,%xmm7 + DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 + DB 196,195,69,24,248,1 ; vinsertf128 $0x1,%xmm8,%ymm7,%ymm7 + DB 197,185,114,245,16 ; vpslld $0x10,%xmm5,%xmm8 + DB 196,227,125,25,237,1 ; vextractf128 $0x1,%ymm5,%xmm5 + DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 + DB 196,227,61,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm8,%ymm5 + DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 + DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 + DB 196,227,93,24,230,1 ; vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 + DB 197,220,86,229 ; vorps %ymm5,%ymm4,%ymm4 + DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 + DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 + DB 197,217,254,219 ; vpaddd %xmm3,%xmm4,%xmm3 + DB 196,227,101,24,221,1 ; vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 + DB 196,195,101,74,222,112 ; vblendvps %ymm7,%ymm14,%ymm3,%ymm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 + DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 + DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 + DB 197,252,16,124,36,96 ; vmovups 0x60(%rsp),%ymm7 + DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax 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 45bb <_sk_load_f16_avx+0x2b3> + DB 116,79 ; je 43ae <_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 45bb <_sk_load_f16_avx+0x2b3> + DB 114,67 ; jb 43ae <_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 45c8 <_sk_load_f16_avx+0x2c0> + DB 116,68 ; je 43bb <_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 45c8 <_sk_load_f16_avx+0x2c0> + DB 114,56 ; jb 43bb <_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 15,132,162,253,255,255 ; je 4342 <_sk_load_f16_avx+0x3a> + DB 15,132,68,253,255,255 ; je 40d7 <_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 15,130,146,253,255,255 ; jb 4342 <_sk_load_f16_avx+0x3a> + DB 15,130,52,253,255,255 ; jb 40d7 <_sk_load_f16_avx+0x43> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,135,253,255,255 ; jmpq 4342 <_sk_load_f16_avx+0x3a> + DB 233,41,253,255,255 ; jmpq 40d7 <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,122,253,255,255 ; jmpq 4342 <_sk_load_f16_avx+0x3a> + DB 233,28,253,255,255 ; jmpq 40d7 <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 - DB 233,113,253,255,255 ; jmpq 4342 <_sk_load_f16_avx+0x3a> + DB 233,19,253,255,255 ; jmpq 40d7 <_sk_load_f16_avx+0x43> PUBLIC _sk_gather_f16_avx _sk_gather_f16_avx LABEL PROC @@ -8501,7 +8343,7 @@ _sk_gather_f16_avx LABEL PROC DB 197,252,17,124,36,96 ; vmovups %ymm7,0x60(%rsp) DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) - DB 197,252,17,36,36 ; vmovups %ymm4,(%rsp) + DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 @@ -8542,106 +8384,118 @@ _sk_gather_f16_avx LABEL PROC DB 197,177,105,201 ; vpunpckhwd %xmm1,%xmm9,%xmm1 DB 197,169,97,211 ; vpunpcklwd %xmm3,%xmm10,%xmm2 DB 197,169,105,219 ; vpunpckhwd %xmm3,%xmm10,%xmm3 - DB 197,121,97,241 ; vpunpcklwd %xmm1,%xmm0,%xmm14 + DB 197,121,97,249 ; vpunpcklwd %xmm1,%xmm0,%xmm15 DB 197,121,105,193 ; vpunpckhwd %xmm1,%xmm0,%xmm8 - DB 197,105,97,251 ; vpunpcklwd %xmm3,%xmm2,%xmm15 - DB 197,105,105,219 ; vpunpckhwd %xmm3,%xmm2,%xmm11 - DB 196,193,9,108,199 ; vpunpcklqdq %xmm15,%xmm14,%xmm0 - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,193,121,105,209 ; vpunpckhwd %xmm9,%xmm0,%xmm2 + DB 197,233,97,203 ; vpunpcklwd %xmm3,%xmm2,%xmm1 + DB 197,105,105,203 ; vpunpckhwd %xmm3,%xmm2,%xmm9 + DB 197,129,108,193 ; vpunpcklqdq %xmm1,%xmm15,%xmm0 + DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 + 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 184,0,128,0,0 ; mov $0x8000,%eax - DB 197,249,110,208 ; vmovd %eax,%xmm2 - DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 - DB 196,99,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm10 - DB 196,193,124,84,210 ; vandps %ymm10,%ymm0,%ymm2 + DB 196,98,125,24,37,11,25,0,0 ; vbroadcastss 0x190b(%rip),%ymm12 # 5ddc <_sk_callback_avx+0x1f2> + DB 196,193,124,84,212 ; vandps %ymm12,%ymm0,%ymm2 DB 197,252,87,194 ; vxorps %ymm2,%ymm0,%ymm0 - DB 184,0,4,0,0 ; mov $0x400,%eax - DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 - DB 197,249,110,200 ; vmovd %eax,%xmm1 - DB 197,121,112,225,0 ; vpshufd $0x0,%xmm1,%xmm12 - DB 197,153,102,203 ; vpcmpgtd %xmm3,%xmm12,%xmm1 - DB 197,25,102,232 ; vpcmpgtd %xmm0,%xmm12,%xmm13 - DB 196,227,21,24,225,1 ; vinsertf128 $0x1,%xmm1,%ymm13,%ymm4 - DB 196,227,125,25,209,1 ; vextractf128 $0x1,%ymm2,%xmm1 - DB 197,145,114,241,16 ; vpslld $0x10,%xmm1,%xmm13 - DB 197,241,114,243,13 ; vpslld $0xd,%xmm3,%xmm1 - DB 184,0,0,0,56 ; mov $0x38000000,%eax - DB 197,249,110,216 ; vmovd %eax,%xmm3 - DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3 - DB 197,145,254,251 ; vpaddd %xmm3,%xmm13,%xmm7 - DB 197,193,254,201 ; vpaddd %xmm1,%xmm7,%xmm1 + DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 + DB 196,98,121,24,29,247,24,0,0 ; vbroadcastss 0x18f7(%rip),%xmm11 # 5de0 <_sk_callback_avx+0x1f6> + DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 + DB 196,98,121,24,45,237,24,0,0 ; vbroadcastss 0x18ed(%rip),%xmm13 # 5de4 <_sk_callback_avx+0x1fa> + 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 + DB 196,99,45,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm10,%ymm10 + DB 197,225,114,242,16 ; vpslld $0x10,%xmm2,%xmm3 + DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 DB 197,249,114,240,13 ; vpslld $0xd,%xmm0,%xmm0 + DB 196,193,97,114,246,13 ; vpslld $0xd,%xmm14,%xmm3 + 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 196,226,121,24,29,163,24,0,0 ; vbroadcastss 0x18a3(%rip),%xmm3 # 5de8 <_sk_callback_avx+0x1fe> DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 - DB 197,233,254,192 ; vpaddd %xmm0,%xmm2,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,65,20,87,237 ; vxorps %ymm13,%ymm13,%ymm13 - DB 196,195,125,74,197,64 ; vblendvps %ymm4,%ymm13,%ymm0,%ymm0 - DB 196,193,9,109,207 ; vpunpckhqdq %xmm15,%xmm14,%xmm1 - DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2 + DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 + DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 + DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 + DB 196,195,125,74,198,160 ; vblendvps %ymm10,%ymm14,%ymm0,%ymm0 + DB 197,129,109,201 ; vpunpckhqdq %xmm1,%xmm15,%xmm1 + DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 + DB 197,241,105,212 ; vpunpckhwd %xmm4,%xmm1,%xmm2 DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,193,116,84,210 ; vandps %ymm10,%ymm1,%ymm2 + DB 196,193,116,84,212 ; vandps %ymm12,%ymm1,%ymm2 DB 197,244,87,202 ; vxorps %ymm2,%ymm1,%ymm1 - DB 196,227,125,25,204,1 ; vextractf128 $0x1,%ymm1,%xmm4 - DB 197,153,102,252 ; vpcmpgtd %xmm4,%xmm12,%xmm7 - DB 197,25,102,241 ; vpcmpgtd %xmm1,%xmm12,%xmm14 - DB 196,227,13,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm14,%ymm7 - DB 196,227,125,25,214,1 ; vextractf128 $0x1,%ymm2,%xmm6 - DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 - DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 - DB 197,201,254,243 ; vpaddd %xmm3,%xmm6,%xmm6 - DB 197,201,254,228 ; vpaddd %xmm4,%xmm6,%xmm4 + DB 196,195,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm10 + DB 196,193,40,87,251 ; vxorps %xmm11,%xmm10,%xmm7 + DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 + DB 196,193,112,87,243 ; vxorps %xmm11,%xmm1,%xmm6 + DB 197,145,102,246 ; vpcmpgtd %xmm6,%xmm13,%xmm6 + DB 196,227,77,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 + DB 197,193,114,242,16 ; vpslld $0x10,%xmm2,%xmm7 + DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 + DB 196,227,69,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm7,%ymm2 DB 197,241,114,241,13 ; vpslld $0xd,%xmm1,%xmm1 + DB 196,193,65,114,242,13 ; vpslld $0xd,%xmm10,%xmm7 + DB 196,227,117,24,207,1 ; vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 + DB 197,244,86,202 ; vorps %ymm2,%ymm1,%ymm1 + DB 196,227,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm2 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 - DB 197,233,254,201 ; vpaddd %xmm1,%xmm2,%xmm1 - DB 196,227,117,24,204,1 ; vinsertf128 $0x1,%xmm4,%ymm1,%ymm1 - DB 196,195,117,74,205,112 ; vblendvps %ymm7,%ymm13,%ymm1,%ymm1 - DB 196,193,57,108,211 ; vpunpcklqdq %xmm11,%xmm8,%xmm2 - DB 196,193,105,105,225 ; vpunpckhwd %xmm9,%xmm2,%xmm4 + DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 + DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 + DB 196,195,117,74,206,96 ; vblendvps %ymm6,%ymm14,%ymm1,%ymm1 + DB 196,193,57,108,209 ; vpunpcklqdq %xmm9,%xmm8,%xmm2 + DB 197,233,105,244 ; vpunpckhwd %xmm4,%xmm2,%xmm6 + DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,227,109,24,212,1 ; vinsertf128 $0x1,%xmm4,%ymm2,%ymm2 - DB 196,193,108,84,226 ; vandps %ymm10,%ymm2,%ymm4 - DB 197,236,87,212 ; vxorps %ymm4,%ymm2,%ymm2 - DB 196,227,125,25,214,1 ; vextractf128 $0x1,%ymm2,%xmm6 - DB 197,153,102,254 ; vpcmpgtd %xmm6,%xmm12,%xmm7 - DB 197,25,102,242 ; vpcmpgtd %xmm2,%xmm12,%xmm14 - DB 196,227,13,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm14,%ymm7 - DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 - DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 - DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 - DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 - DB 197,209,254,238 ; vpaddd %xmm6,%xmm5,%xmm5 - DB 197,217,114,244,16 ; vpslld $0x10,%xmm4,%xmm4 + DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 + DB 196,193,108,84,244 ; vandps %ymm12,%ymm2,%ymm6 + DB 197,236,87,214 ; vxorps %ymm6,%ymm2,%ymm2 + DB 196,227,125,25,215,1 ; vextractf128 $0x1,%ymm2,%xmm7 + DB 196,193,64,87,235 ; vxorps %xmm11,%xmm7,%xmm5 + DB 197,145,102,237 ; vpcmpgtd %xmm5,%xmm13,%xmm5 + DB 196,193,104,87,227 ; vxorps %xmm11,%xmm2,%xmm4 + DB 197,145,102,228 ; vpcmpgtd %xmm4,%xmm13,%xmm4 + DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + DB 197,209,114,246,16 ; vpslld $0x10,%xmm6,%xmm5 + DB 196,227,125,25,246,1 ; vextractf128 $0x1,%ymm6,%xmm6 + DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 + DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 - DB 197,217,254,227 ; vpaddd %xmm3,%xmm4,%xmm4 - DB 197,217,254,210 ; vpaddd %xmm2,%xmm4,%xmm2 + DB 197,201,114,247,13 ; vpslld $0xd,%xmm7,%xmm6 + DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 + DB 197,236,86,213 ; vorps %ymm5,%ymm2,%ymm2 + DB 196,227,125,25,213,1 ; vextractf128 $0x1,%ymm2,%xmm5 + DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 + DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,109,24,213,1 ; vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 - DB 196,195,109,74,213,112 ; vblendvps %ymm7,%ymm13,%ymm2,%ymm2 - DB 196,193,57,109,227 ; vpunpckhqdq %xmm11,%xmm8,%xmm4 - DB 196,193,89,105,233 ; vpunpckhwd %xmm9,%xmm4,%xmm5 + DB 196,195,109,74,214,64 ; vblendvps %ymm4,%ymm14,%ymm2,%ymm2 + DB 196,193,57,109,225 ; vpunpckhqdq %xmm9,%xmm8,%xmm4 + DB 196,193,89,105,234 ; vpunpckhwd %xmm10,%xmm4,%xmm5 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 196,193,92,84,234 ; vandps %ymm10,%ymm4,%ymm5 + DB 196,193,92,84,236 ; vandps %ymm12,%ymm4,%ymm5 DB 197,220,87,229 ; vxorps %ymm5,%ymm4,%ymm4 DB 196,227,125,25,230,1 ; vextractf128 $0x1,%ymm4,%xmm6 - DB 197,153,102,254 ; vpcmpgtd %xmm6,%xmm12,%xmm7 - DB 197,25,102,196 ; vpcmpgtd %xmm4,%xmm12,%xmm8 - DB 196,99,61,24,199,1 ; vinsertf128 $0x1,%xmm7,%ymm8,%ymm8 - DB 196,227,125,25,239,1 ; vextractf128 $0x1,%ymm5,%xmm7 - DB 197,193,114,247,16 ; vpslld $0x10,%xmm7,%xmm7 + DB 196,193,72,87,251 ; vxorps %xmm11,%xmm6,%xmm7 + DB 197,17,102,199 ; vpcmpgtd %xmm7,%xmm13,%xmm8 + DB 196,193,88,87,251 ; vxorps %xmm11,%xmm4,%xmm7 + DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 + DB 196,195,69,24,248,1 ; vinsertf128 $0x1,%xmm8,%ymm7,%ymm7 + DB 197,185,114,245,16 ; vpslld $0x10,%xmm5,%xmm8 + DB 196,227,125,25,237,1 ; vextractf128 $0x1,%ymm5,%xmm5 DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 - DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 - DB 197,193,254,219 ; vpaddd %xmm3,%xmm7,%xmm3 - DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 - DB 197,225,254,222 ; vpaddd %xmm6,%xmm3,%xmm3 + DB 196,227,61,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm8,%ymm5 DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 - DB 197,209,254,228 ; vpaddd %xmm4,%xmm5,%xmm4 - DB 196,227,93,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm3 - DB 196,195,101,74,221,128 ; vblendvps %ymm8,%ymm13,%ymm3,%ymm3 + DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 + DB 196,227,93,24,230,1 ; vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 + DB 197,220,86,229 ; vorps %ymm5,%ymm4,%ymm4 + DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 + DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 + DB 197,217,254,219 ; vpaddd %xmm3,%xmm4,%xmm3 + DB 196,227,101,24,221,1 ; vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 + DB 196,195,101,74,222,112 ; vblendvps %ymm7,%ymm14,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 @@ -8661,139 +8515,135 @@ _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 197,252,40,225 ; vmovaps %ymm1,%ymm4 - DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 184,0,0,0,128 ; mov $0x80000000,%eax - DB 197,121,110,192 ; vmovd %eax,%xmm8 - DB 196,65,121,112,192,0 ; vpshufd $0x0,%xmm8,%xmm8 - DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 - DB 197,60,84,208 ; vandps %ymm0,%ymm8,%ymm10 + DB 196,98,125,24,13,176,22,0,0 ; vbroadcastss 0x16b0(%rip),%ymm9 # 5dec <_sk_callback_avx+0x202> + 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 184,0,0,128,56 ; mov $0x38800000,%eax DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12 - DB 197,121,110,200 ; vmovd %eax,%xmm9 - DB 196,65,121,112,201,0 ; vpshufd $0x0,%xmm9,%xmm9 - DB 196,65,49,102,236 ; vpcmpgtd %xmm12,%xmm9,%xmm13 - DB 196,65,49,102,243 ; vpcmpgtd %xmm11,%xmm9,%xmm14 + DB 196,98,121,24,5,150,22,0,0 ; vbroadcastss 0x1696(%rip),%xmm8 # 5df0 <_sk_callback_avx+0x206> + 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 - DB 196,67,125,25,214,1 ; vextractf128 $0x1,%ymm10,%xmm14 - DB 196,193,9,114,214,16 ; vpsrld $0x10,%xmm14,%xmm14 - DB 196,193,1,114,210,16 ; vpsrld $0x10,%xmm10,%xmm15 + DB 196,193,9,114,210,16 ; vpsrld $0x10,%xmm10,%xmm14 + DB 196,67,125,25,210,1 ; vextractf128 $0x1,%ymm10,%xmm10 + DB 196,193,41,114,210,16 ; vpsrld $0x10,%xmm10,%xmm10 + 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 184,0,192,1,0 ; mov $0x1c000,%eax - DB 197,121,110,208 ; vmovd %eax,%xmm10 - DB 196,65,121,112,210,0 ; vpshufd $0x0,%xmm10,%xmm10 - DB 196,65,1,250,250 ; vpsubd %xmm10,%xmm15,%xmm15 - DB 196,65,9,250,242 ; vpsubd %xmm10,%xmm14,%xmm14 - DB 196,65,9,254,228 ; vpaddd %xmm12,%xmm14,%xmm12 - DB 196,65,1,254,219 ; vpaddd %xmm11,%xmm15,%xmm11 + DB 196,98,125,24,21,93,22,0,0 ; vbroadcastss 0x165d(%rip),%ymm10 # 5df4 <_sk_callback_avx+0x20a> + 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 + DB 196,65,9,254,219 ; vpaddd %xmm11,%xmm14,%xmm11 DB 196,67,37,24,228,1 ; vinsertf128 $0x1,%xmm12,%ymm11,%ymm12 - DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 - DB 196,99,29,74,225,208 ; vblendvps %ymm13,%ymm1,%ymm12,%ymm12 - DB 197,60,84,236 ; vandps %ymm4,%ymm8,%ymm13 - DB 197,252,17,36,36 ; vmovups %ymm4,(%rsp) - DB 196,65,92,87,245 ; vxorps %ymm13,%ymm4,%ymm14 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,99,29,74,224,208 ; vblendvps %ymm13,%ymm0,%ymm12,%ymm12 + DB 196,65,116,84,233 ; vandps %ymm9,%ymm1,%ymm13 + DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp) + DB 196,65,116,87,245 ; vxorps %ymm13,%ymm1,%ymm14 DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15 - DB 196,193,49,102,255 ; vpcmpgtd %xmm15,%xmm9,%xmm7 - DB 196,65,49,102,222 ; vpcmpgtd %xmm14,%xmm9,%xmm11 - DB 196,99,37,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm11,%ymm11 + DB 196,193,57,102,255 ; vpcmpgtd %xmm15,%xmm8,%xmm7 + DB 196,65,57,102,222 ; vpcmpgtd %xmm14,%xmm8,%xmm11 + DB 196,227,37,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm11,%ymm7 + DB 196,193,33,114,213,16 ; vpsrld $0x10,%xmm13,%xmm11 DB 196,99,125,25,238,1 ; vextractf128 $0x1,%ymm13,%xmm6 DB 197,201,114,214,16 ; vpsrld $0x10,%xmm6,%xmm6 - DB 196,193,65,114,215,13 ; vpsrld $0xd,%xmm15,%xmm7 - DB 196,193,73,250,242 ; vpsubd %xmm10,%xmm6,%xmm6 - DB 197,73,254,255 ; vpaddd %xmm7,%xmm6,%xmm15 - DB 196,193,65,114,213,16 ; vpsrld $0x10,%xmm13,%xmm7 - DB 196,193,73,114,214,13 ; vpsrld $0xd,%xmm14,%xmm6 - DB 196,193,65,250,250 ; vpsubd %xmm10,%xmm7,%xmm7 - DB 197,193,254,246 ; vpaddd %xmm6,%xmm7,%xmm6 - DB 196,195,77,24,247,1 ; vinsertf128 $0x1,%xmm15,%ymm6,%ymm6 - DB 196,99,77,74,233,176 ; vblendvps %ymm11,%ymm1,%ymm6,%ymm13 - DB 197,188,84,242 ; vandps %ymm2,%ymm8,%ymm6 - DB 197,252,17,84,36,32 ; vmovups %ymm2,0x20(%rsp) - DB 197,236,87,254 ; vxorps %ymm6,%ymm2,%ymm7 - DB 196,195,125,25,251,1 ; vextractf128 $0x1,%ymm7,%xmm11 - DB 196,65,49,102,243 ; vpcmpgtd %xmm11,%xmm9,%xmm14 - DB 197,49,102,255 ; vpcmpgtd %xmm7,%xmm9,%xmm15 - DB 196,67,5,24,246,1 ; vinsertf128 $0x1,%xmm14,%ymm15,%ymm14 + DB 196,227,37,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm11,%ymm6 + DB 196,193,33,114,215,13 ; vpsrld $0xd,%xmm15,%xmm11 + DB 196,193,76,86,242 ; vorps %ymm10,%ymm6,%ymm6 DB 196,227,125,25,245,1 ; vextractf128 $0x1,%ymm6,%xmm5 - DB 197,129,114,213,16 ; vpsrld $0x10,%xmm5,%xmm15 - DB 196,193,81,114,211,13 ; vpsrld $0xd,%xmm11,%xmm5 - DB 196,193,1,250,226 ; vpsubd %xmm10,%xmm15,%xmm4 + DB 196,193,81,254,235 ; vpaddd %xmm11,%xmm5,%xmm5 + DB 196,193,89,114,214,13 ; vpsrld $0xd,%xmm14,%xmm4 + DB 197,201,254,228 ; vpaddd %xmm4,%xmm6,%xmm4 + DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + DB 196,99,93,74,232,112 ; vblendvps %ymm7,%ymm0,%ymm4,%ymm13 + DB 196,193,108,84,225 ; vandps %ymm9,%ymm2,%ymm4 + DB 197,236,87,236 ; vxorps %ymm4,%ymm2,%ymm5 + DB 196,227,125,25,238,1 ; vextractf128 $0x1,%ymm5,%xmm6 + DB 197,185,102,254 ; vpcmpgtd %xmm6,%xmm8,%xmm7 + DB 197,57,102,221 ; vpcmpgtd %xmm5,%xmm8,%xmm11 + DB 196,227,37,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm11,%ymm7 + DB 197,161,114,212,16 ; vpsrld $0x10,%xmm4,%xmm11 + DB 196,227,125,25,228,1 ; vextractf128 $0x1,%ymm4,%xmm4 + DB 197,217,114,212,16 ; vpsrld $0x10,%xmm4,%xmm4 + DB 196,227,37,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm11,%ymm4 + DB 197,201,114,214,13 ; vpsrld $0xd,%xmm6,%xmm6 + DB 196,193,92,86,226 ; vorps %ymm10,%ymm4,%ymm4 + DB 196,227,125,25,225,1 ; vextractf128 $0x1,%ymm4,%xmm1 + DB 197,241,254,206 ; vpaddd %xmm6,%xmm1,%xmm1 + DB 197,209,114,213,13 ; vpsrld $0xd,%xmm5,%xmm5 DB 197,217,254,229 ; vpaddd %xmm5,%xmm4,%xmm4 - DB 197,209,114,214,16 ; vpsrld $0x10,%xmm6,%xmm5 - DB 197,201,114,215,13 ; vpsrld $0xd,%xmm7,%xmm6 - DB 196,193,81,250,234 ; vpsubd %xmm10,%xmm5,%xmm5 - DB 197,209,254,238 ; vpaddd %xmm6,%xmm5,%xmm5 - DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 - DB 196,99,93,74,217,224 ; vblendvps %ymm14,%ymm1,%ymm4,%ymm11 - DB 197,188,84,235 ; vandps %ymm3,%ymm8,%ymm5 - DB 197,228,87,245 ; vxorps %ymm5,%ymm3,%ymm6 - DB 196,227,125,25,247,1 ; vextractf128 $0x1,%ymm6,%xmm7 - DB 197,177,102,231 ; vpcmpgtd %xmm7,%xmm9,%xmm4 - DB 197,49,102,198 ; vpcmpgtd %xmm6,%xmm9,%xmm8 + DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 + DB 196,99,117,74,216,112 ; vblendvps %ymm7,%ymm0,%ymm1,%ymm11 + DB 196,193,100,84,225 ; vandps %ymm9,%ymm3,%ymm4 + DB 197,228,87,236 ; vxorps %ymm4,%ymm3,%ymm5 + DB 196,227,125,25,238,1 ; vextractf128 $0x1,%ymm5,%xmm6 + DB 197,185,102,254 ; vpcmpgtd %xmm6,%xmm8,%xmm7 + DB 197,57,102,197 ; vpcmpgtd %xmm5,%xmm8,%xmm8 + DB 196,227,61,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm8,%ymm7 + DB 197,185,114,212,16 ; vpsrld $0x10,%xmm4,%xmm8 + DB 196,227,125,25,228,1 ; vextractf128 $0x1,%ymm4,%xmm4 + DB 197,217,114,212,16 ; vpsrld $0x10,%xmm4,%xmm4 DB 196,227,61,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm8,%ymm4 - DB 196,227,125,25,234,1 ; vextractf128 $0x1,%ymm5,%xmm2 - DB 197,233,114,210,16 ; vpsrld $0x10,%xmm2,%xmm2 - DB 197,209,114,213,16 ; vpsrld $0x10,%xmm5,%xmm5 - DB 196,193,81,250,234 ; vpsubd %xmm10,%xmm5,%xmm5 - DB 196,193,105,250,210 ; vpsubd %xmm10,%xmm2,%xmm2 - DB 197,193,114,215,13 ; vpsrld $0xd,%xmm7,%xmm7 - DB 197,233,254,215 ; vpaddd %xmm7,%xmm2,%xmm2 + DB 196,193,92,86,226 ; vorps %ymm10,%ymm4,%ymm4 DB 197,201,114,214,13 ; vpsrld $0xd,%xmm6,%xmm6 - DB 197,209,254,238 ; vpaddd %xmm6,%xmm5,%xmm5 - DB 196,227,85,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm5,%ymm2 - DB 196,227,109,74,209,64 ; vblendvps %ymm4,%ymm1,%ymm2,%ymm2 - DB 196,99,125,25,225,1 ; vextractf128 $0x1,%ymm12,%xmm1 - DB 196,226,25,43,201 ; vpackusdw %xmm1,%xmm12,%xmm1 + DB 196,227,125,25,225,1 ; vextractf128 $0x1,%ymm4,%xmm1 + DB 197,241,254,206 ; vpaddd %xmm6,%xmm1,%xmm1 + DB 197,209,114,213,13 ; vpsrld $0xd,%xmm5,%xmm5 + DB 197,217,254,229 ; vpaddd %xmm5,%xmm4,%xmm4 + DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 + DB 196,227,117,74,200,112 ; vblendvps %ymm7,%ymm0,%ymm1,%ymm1 + DB 196,99,125,25,224,1 ; vextractf128 $0x1,%ymm12,%xmm0 + DB 196,226,25,43,192 ; vpackusdw %xmm0,%xmm12,%xmm0 DB 196,99,125,25,236,1 ; vextractf128 $0x1,%ymm13,%xmm4 DB 196,226,17,43,228 ; vpackusdw %xmm4,%xmm13,%xmm4 DB 196,99,125,25,221,1 ; vextractf128 $0x1,%ymm11,%xmm5 - DB 196,226,33,43,237 ; vpackusdw %xmm5,%xmm11,%xmm5 - DB 196,227,125,25,214,1 ; vextractf128 $0x1,%ymm2,%xmm6 - DB 196,226,105,43,214 ; vpackusdw %xmm6,%xmm2,%xmm2 - DB 197,241,97,244 ; vpunpcklwd %xmm4,%xmm1,%xmm6 - DB 197,241,105,204 ; vpunpckhwd %xmm4,%xmm1,%xmm1 - DB 197,209,97,226 ; vpunpcklwd %xmm2,%xmm5,%xmm4 - DB 197,209,105,210 ; vpunpckhwd %xmm2,%xmm5,%xmm2 - DB 197,73,98,220 ; vpunpckldq %xmm4,%xmm6,%xmm11 - DB 197,73,106,212 ; vpunpckhdq %xmm4,%xmm6,%xmm10 - DB 197,113,98,202 ; vpunpckldq %xmm2,%xmm1,%xmm9 - DB 197,113,106,194 ; vpunpckhdq %xmm2,%xmm1,%xmm8 + DB 196,226,33,43,245 ; vpackusdw %xmm5,%xmm11,%xmm6 + DB 196,227,125,25,205,1 ; vextractf128 $0x1,%ymm1,%xmm5 + DB 196,226,113,43,205 ; vpackusdw %xmm5,%xmm1,%xmm1 + DB 197,249,97,236 ; vpunpcklwd %xmm4,%xmm0,%xmm5 + DB 197,249,105,196 ; vpunpckhwd %xmm4,%xmm0,%xmm0 + DB 197,201,97,225 ; vpunpcklwd %xmm1,%xmm6,%xmm4 + DB 197,201,105,201 ; vpunpckhwd %xmm1,%xmm6,%xmm1 + DB 197,81,98,220 ; vpunpckldq %xmm4,%xmm5,%xmm11 + DB 197,81,106,212 ; vpunpckhdq %xmm4,%xmm5,%xmm10 + DB 197,121,98,201 ; vpunpckldq %xmm1,%xmm0,%xmm9 + DB 197,121,106,193 ; vpunpckhdq %xmm1,%xmm0,%xmm8 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 117,79 ; jne 4b32 <_sk_store_f16_avx+0x271> - DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8) - DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8) - DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8) - DB 196,65,122,127,68,248,48 ; vmovdqu %xmm8,0x30(%r8,%rdi,8) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,12,36 ; vmovups (%rsp),%ymm1 - DB 197,252,16,84,36,32 ; vmovups 0x20(%rsp),%ymm2 + DB 117,75 ; jne 497e <_sk_store_f16_avx+0x270> + DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8) + DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8) + DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8) + DB 197,122,127,68,248,48 ; vmovdqu %xmm8,0x30(%rax,%rdi,8) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 + DB 197,252,16,76,36,32 ; vmovups 0x20(%rsp),%ymm1 DB 197,252,16,100,36,64 ; vmovups 0x40(%rsp),%ymm4 DB 197,252,16,108,36,96 ; vmovups 0x60(%rsp),%ymm5 DB 197,252,16,180,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm6 DB 197,252,16,188,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm7 DB 72,129,196,216,0,0,0 ; add $0xd8,%rsp DB 255,224 ; jmpq *%rax - DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8) + DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,192 ; je 4afe <_sk_store_f16_avx+0x23d> - DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8) + DB 116,193 ; je 494a <_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 114,179 ; jb 4afe <_sk_store_f16_avx+0x23d> - DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,170 ; je 4afe <_sk_store_f16_avx+0x23d> - DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8) + DB 114,181 ; jb 494a <_sk_store_f16_avx+0x23c> + DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) + DB 116,173 ; je 494a <_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 114,157 ; jb 4afe <_sk_store_f16_avx+0x23d> - DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,148 ; je 4afe <_sk_store_f16_avx+0x23d> - DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8) + DB 114,161 ; jb 494a <_sk_store_f16_avx+0x23c> + DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) + DB 116,153 ; je 494a <_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 114,135 ; jb 4afe <_sk_store_f16_avx+0x23d> - DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 233,123,255,255,255 ; jmpq 4afe <_sk_store_f16_avx+0x23d> + DB 114,141 ; jb 494a <_sk_store_f16_avx+0x23c> + DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) + DB 235,133 ; jmp 494a <_sk_store_f16_avx+0x23c> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC @@ -8801,7 +8651,7 @@ _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 15,133,5,1,0,0 ; jne 4c9e <_sk_load_u16_be_avx+0x11b> + DB 15,133,5,1,0,0 ; jne 4ae0 <_sk_load_u16_be_avx+0x11b> 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 @@ -8860,29 +8710,29 @@ _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 4d04 <_sk_load_u16_be_avx+0x181> + DB 116,85 ; je 4b46 <_sk_load_u16_be_avx+0x181> 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 4d04 <_sk_load_u16_be_avx+0x181> + DB 114,72 ; jb 4b46 <_sk_load_u16_be_avx+0x181> 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 4d11 <_sk_load_u16_be_avx+0x18e> + DB 116,72 ; je 4b53 <_sk_load_u16_be_avx+0x18e> 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 4d11 <_sk_load_u16_be_avx+0x18e> + DB 114,59 ; jb 4b53 <_sk_load_u16_be_avx+0x18e> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,205,254,255,255 ; je 4bb4 <_sk_load_u16_be_avx+0x31> + DB 15,132,205,254,255,255 ; je 49f6 <_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 15,130,188,254,255,255 ; jb 4bb4 <_sk_load_u16_be_avx+0x31> + DB 15,130,188,254,255,255 ; jb 49f6 <_sk_load_u16_be_avx+0x31> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,176,254,255,255 ; jmpq 4bb4 <_sk_load_u16_be_avx+0x31> + DB 233,176,254,255,255 ; jmpq 49f6 <_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 233,163,254,255,255 ; jmpq 4bb4 <_sk_load_u16_be_avx+0x31> + DB 233,163,254,255,255 ; jmpq 49f6 <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,154,254,255,255 ; jmpq 4bb4 <_sk_load_u16_be_avx+0x31> + DB 233,154,254,255,255 ; jmpq 49f6 <_sk_load_u16_be_avx+0x31> PUBLIC _sk_load_rgb_u16_be_avx _sk_load_rgb_u16_be_avx LABEL PROC @@ -8890,7 +8740,7 @@ _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 15,133,8,1,0,0 ; jne 4e34 <_sk_load_rgb_u16_be_avx+0x11a> + DB 15,133,8,1,0,0 ; jne 4c76 <_sk_load_rgb_u16_be_avx+0x11a> 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 @@ -8949,36 +8799,36 @@ _sk_load_rgb_u16_be_avx LABEL PROC 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 4e4d <_sk_load_rgb_u16_be_avx+0x133> - DB 233,19,255,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> + DB 117,5 ; jne 4c8f <_sk_load_rgb_u16_be_avx+0x133> + DB 233,19,255,255,255 ; jmpq 4ba2 <_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 4e7c <_sk_load_rgb_u16_be_avx+0x162> + DB 114,26 ; jb 4cbe <_sk_load_rgb_u16_be_avx+0x162> 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 4e81 <_sk_load_rgb_u16_be_avx+0x167> - DB 233,228,254,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,223,254,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 4cc3 <_sk_load_rgb_u16_be_avx+0x167> + DB 233,228,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,223,254,255,255 ; jmpq 4ba2 <_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 4eb0 <_sk_load_rgb_u16_be_avx+0x196> + DB 114,26 ; jb 4cf2 <_sk_load_rgb_u16_be_avx+0x196> 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 4eb5 <_sk_load_rgb_u16_be_avx+0x19b> - DB 233,176,254,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,171,254,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 4cf7 <_sk_load_rgb_u16_be_avx+0x19b> + DB 233,176,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,171,254,255,255 ; jmpq 4ba2 <_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 4ede <_sk_load_rgb_u16_be_avx+0x1c4> + DB 114,20 ; jb 4d20 <_sk_load_rgb_u16_be_avx+0x1c4> 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 233,130,254,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> - DB 233,125,254,255,255 ; jmpq 4d60 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,130,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,125,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC @@ -9026,7 +8876,7 @@ _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 4fe5 <_sk_store_u16_be_avx+0x102> + DB 117,31 ; jne 4e27 <_sk_store_u16_be_avx+0x102> DB 196,1,120,17,28,72 ; vmovups %xmm11,(%r8,%r9,2) DB 196,1,120,17,84,72,16 ; vmovups %xmm10,0x10(%r8,%r9,2) DB 196,1,120,17,76,72,32 ; vmovups %xmm9,0x20(%r8,%r9,2) @@ -9035,31 +8885,31 @@ _sk_store_u16_be_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,1,121,214,28,72 ; vmovq %xmm11,(%r8,%r9,2) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 116,240 ; je 4e23 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,23,92,72,8 ; vmovhpd %xmm11,0x8(%r8,%r9,2) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 114,227 ; jb 4e23 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,214,84,72,16 ; vmovq %xmm10,0x10(%r8,%r9,2) - DB 116,218 ; je 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 116,218 ; je 4e23 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,23,84,72,24 ; vmovhpd %xmm10,0x18(%r8,%r9,2) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 114,205 ; jb 4e23 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,214,76,72,32 ; vmovq %xmm9,0x20(%r8,%r9,2) - DB 116,196 ; je 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 116,196 ; je 4e23 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,23,76,72,40 ; vmovhpd %xmm9,0x28(%r8,%r9,2) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 114,183 ; jb 4e23 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,214,68,72,48 ; vmovq %xmm8,0x30(%r8,%r9,2) - DB 235,174 ; jmp 4fe1 <_sk_store_u16_be_avx+0xfe> + DB 235,174 ; jmp 4e23 <_sk_store_u16_be_avx+0xfe> 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 50a9 <_sk_load_f32_avx+0x76> + DB 119,110 ; ja 4eeb <_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 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 50d4 <_sk_load_f32_avx+0xa1> + DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 4f14 <_sk_load_f32_avx+0x9f> DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -9085,22 +8935,19 @@ _sk_load_f32_avx LABEL PROC DB 196,193,101,21,216 ; vunpckhpd %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax - DB 15,31,0 ; nopl (%rax) - DB 130 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,201 ; dec %ecx - DB 255 ; (bad) + DB 144 ; nop + DB 132,255 ; test %bh,%bh DB 255 ; (bad) + DB 255,203 ; dec %ebx DB 255 ; (bad) - DB 188,255,255,255,175 ; mov $0xafffffff,%esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,162,255,255,255,154 ; jmpq *-0x65000001(%rdx) + DB 190,255,255,255,177 ; mov $0xb1ffffff,%esi DB 255 ; (bad) DB 255 ; (bad) - DB 255,146,255,255,255,138 ; callq *-0x75000001(%rdx) + DB 255,164,255,255,255,156,255 ; jmpq *-0x630001(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,148,255,255,255,140,255 ; callq *-0x730001(%rdi,%rdi,8) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -9118,7 +8965,7 @@ _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 5161 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 4fa1 <_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 @@ -9131,22 +8978,22 @@ _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 515d <_sk_store_f32_avx+0x69> + DB 116,240 ; je 4f9d <_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 515d <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 4f9d <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 515d <_sk_store_f32_avx+0x69> + DB 116,218 ; je 4f9d <_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 515d <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 4f9d <_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 515d <_sk_store_f32_avx+0x69> + DB 116,195 ; je 4f9d <_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 515d <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 4f9d <_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 515d <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 4f9d <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -9450,7 +9297,7 @@ _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 5715 <_sk_linear_gradient_avx+0xb8> + DB 15,132,146,0,0,0 ; je 5555 <_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 @@ -9477,8 +9324,8 @@ _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 569f <_sk_linear_gradient_avx+0x42> - DB 235,20 ; jmp 5729 <_sk_linear_gradient_avx+0xcc> + DB 117,140 ; jne 54df <_sk_linear_gradient_avx+0x42> + DB 235,20 ; jmp 5569 <_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 @@ -9932,8 +9779,88 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) - DB 128 ; .byte 0x80 - DB 63 ; (bad) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 5d14 <.literal4+0x10> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 5d8d <.literal4+0x89> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 5d44 <.literal4+0x40> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 5dbd <.literal4+0xb9> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 5d74 <.literal4+0x70> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 5ded <.literal4+0xe9> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 5da4 <.literal4+0xa0> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 5e1d <_sk_callback_avx+0x233> + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,249,68,180 ; mov $0xb444f93f,%edi + DB 62,163,233,220,63,81,140,242,66,141 ; movabs %eax,%ds:0x8d42f28c513fdce9 + DB 188,190,63,248,245 ; mov $0xf5f83fbe,%esp + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 128,0,0 ; addb $0x0,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + 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) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,128,56 ; add %al,0x38800000(%rax) + DB 0,64,254 ; add %al,-0x2(%rax) + DB 255 ; .byte 0xff ALIGN 32 PUBLIC _sk_start_pipeline_sse41 @@ -10015,7 +9942,7 @@ _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 15,40,21,129,64,0,0 ; movaps 0x4081(%rip),%xmm2 # 4190 <_sk_callback_sse41+0xb6> + DB 15,40,21,193,62,0,0 ; movaps 0x3ec1(%rip),%xmm2 # 3fd0 <_sk_callback_sse41+0xb4> DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,16,2 ; movups (%rdx),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 @@ -10024,7 +9951,7 @@ _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 15,40,21,112,64,0,0 ; movaps 0x4070(%rip),%xmm2 # 41a0 <_sk_callback_sse41+0xc6> + DB 15,40,21,176,62,0,0 ; movaps 0x3eb0(%rip),%xmm2 # 3fe0 <_sk_callback_sse41+0xc4> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -12008,74 +11935,44 @@ _sk_parametric_r_sse41 LABEL PROC DB 243,68,15,16,72,12 ; movss 0xc(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,89,200 ; mulps %xmm0,%xmm9 - DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 68,15,89,216 ; mulps %xmm0,%xmm11 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,208 ; mulps %xmm0,%xmm10 DB 65,15,194,192,2 ; cmpleps %xmm8,%xmm0 DB 243,68,15,16,64,24 ; movss 0x18(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 69,15,88,200 ; addps %xmm8,%xmm9 - DB 243,68,15,16,16 ; movss (%rax),%xmm10 + DB 243,68,15,16,24 ; movss (%rax),%xmm11 DB 243,68,15,16,64,8 ; movss 0x8(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 69,15,88,216 ; addps %xmm8,%xmm11 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,193 ; movd %ecx,%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 102,69,15,112,228,0 ; pshufd $0x0,%xmm12,%xmm12 - DB 102,69,15,219,227 ; pand %xmm11,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,219,0 ; pshufd $0x0,%xmm11,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,227 ; mulps %xmm11,%xmm12 + 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 68,15,89,5,111,33,0,0 ; mulps 0x216f(%rip),%xmm8 # 3ff0 <_sk_callback_sse41+0xd4> + DB 68,15,84,21,119,33,0,0 ; andps 0x2177(%rip),%xmm10 # 4000 <_sk_callback_sse41+0xe4> + DB 68,15,86,21,127,33,0,0 ; orps 0x217f(%rip),%xmm10 # 4010 <_sk_callback_sse41+0xf4> + DB 68,15,88,5,135,33,0,0 ; addps 0x2187(%rip),%xmm8 # 4020 <_sk_callback_sse41+0x104> + DB 68,15,40,37,143,33,0,0 ; movaps 0x218f(%rip),%xmm12 # 4030 <_sk_callback_sse41+0x114> + DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,88,235 ; addps %xmm11,%xmm13 - DB 69,15,94,229 ; divps %xmm13,%xmm12 + DB 68,15,88,21,143,33,0,0 ; addps 0x218f(%rip),%xmm10 # 4040 <_sk_callback_sse41+0x124> + DB 68,15,40,37,151,33,0,0 ; movaps 0x2197(%rip),%xmm12 # 4050 <_sk_callback_sse41+0x134> + DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 69,15,89,194 ; mulps %xmm10,%xmm8 - DB 102,69,15,58,8,216,1 ; roundps $0x1,%xmm8,%xmm11 - DB 185,0,0,0,75 ; mov $0x4b000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,224 ; addps %xmm8,%xmm12 - DB 69,15,92,195 ; subps %xmm11,%xmm8 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,89,216 ; mulps %xmm8,%xmm11 - DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,92,232 ; subps %xmm8,%xmm13 - DB 69,15,94,221 ; divps %xmm13,%xmm11 - DB 69,15,88,220 ; addps %xmm12,%xmm11 - DB 69,15,89,218 ; mulps %xmm10,%xmm11 + 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 68,15,88,5,132,33,0,0 ; addps 0x2184(%rip),%xmm8 # 4060 <_sk_callback_sse41+0x144> + DB 68,15,40,21,140,33,0,0 ; movaps 0x218c(%rip),%xmm10 # 4070 <_sk_callback_sse41+0x154> + DB 69,15,89,211 ; mulps %xmm11,%xmm10 + DB 69,15,92,194 ; subps %xmm10,%xmm8 + DB 68,15,40,21,140,33,0,0 ; movaps 0x218c(%rip),%xmm10 # 4080 <_sk_callback_sse41+0x164> + DB 69,15,92,211 ; subps %xmm11,%xmm10 + DB 68,15,40,29,144,33,0,0 ; movaps 0x2190(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x174> + DB 69,15,94,218 ; divps %xmm10,%xmm11 + DB 69,15,88,216 ; addps %xmm8,%xmm11 + DB 68,15,89,29,144,33,0,0 ; mulps 0x2190(%rip),%xmm11 # 40a0 <_sk_callback_sse41+0x184> 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 @@ -12114,62 +12011,32 @@ _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 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 102,68,15,112,225,0 ; pshufd $0x0,%xmm1,%xmm12 - DB 102,69,15,219,227 ; pand %xmm11,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 102,68,15,112,217,0 ; pshufd $0x0,%xmm1,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 68,15,92,233 ; subps %xmm1,%xmm13 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 + DB 68,15,89,37,255,32,0,0 ; mulps 0x20ff(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x194> + DB 68,15,84,29,7,33,0,0 ; andps 0x2107(%rip),%xmm11 # 40c0 <_sk_callback_sse41+0x1a4> + DB 68,15,86,29,15,33,0,0 ; orps 0x210f(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x1b4> + DB 68,15,88,37,23,33,0,0 ; addps 0x2117(%rip),%xmm12 # 40e0 <_sk_callback_sse41+0x1c4> + DB 15,40,13,32,33,0,0 ; movaps 0x2120(%rip),%xmm1 # 40f0 <_sk_callback_sse41+0x1d4> DB 65,15,89,203 ; mulps %xmm11,%xmm1 - DB 68,15,92,233 ; subps %xmm1,%xmm13 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 65,15,88,203 ; addps %xmm11,%xmm1 - DB 68,15,94,225 ; divps %xmm1,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 102,69,15,58,8,221,1 ; roundps $0x1,%xmm13,%xmm11 - DB 185,0,0,0,75 ; mov $0x4b000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 - DB 69,15,92,235 ; subps %xmm11,%xmm13 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 65,15,89,205 ; mulps %xmm13,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 65,15,92,205 ; subps %xmm13,%xmm1 - DB 68,15,94,217 ; divps %xmm1,%xmm11 - DB 69,15,88,220 ; addps %xmm12,%xmm11 - DB 69,15,89,218 ; mulps %xmm10,%xmm11 - DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 + DB 68,15,88,29,32,33,0,0 ; addps 0x2120(%rip),%xmm11 # 4100 <_sk_callback_sse41+0x1e4> + DB 15,40,13,41,33,0,0 ; movaps 0x2129(%rip),%xmm1 # 4110 <_sk_callback_sse41+0x1f4> + 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 68,15,88,37,22,33,0,0 ; addps 0x2116(%rip),%xmm12 # 4120 <_sk_callback_sse41+0x204> + DB 15,40,13,31,33,0,0 ; movaps 0x211f(%rip),%xmm1 # 4130 <_sk_callback_sse41+0x214> + DB 65,15,89,203 ; mulps %xmm11,%xmm1 + DB 68,15,92,225 ; subps %xmm1,%xmm12 + DB 68,15,40,21,31,33,0,0 ; movaps 0x211f(%rip),%xmm10 # 4140 <_sk_callback_sse41+0x224> + DB 69,15,92,211 ; subps %xmm11,%xmm10 + DB 15,40,13,36,33,0,0 ; movaps 0x2124(%rip),%xmm1 # 4150 <_sk_callback_sse41+0x234> + DB 65,15,94,202 ; divps %xmm10,%xmm1 + DB 65,15,88,204 ; addps %xmm12,%xmm1 + DB 15,89,13,37,33,0,0 ; mulps 0x2125(%rip),%xmm1 # 4160 <_sk_callback_sse41+0x244> + 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 DB 65,15,88,202 ; addps %xmm10,%xmm1 @@ -12207,62 +12074,32 @@ _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 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 102,68,15,112,226,0 ; pshufd $0x0,%xmm2,%xmm12 - DB 102,69,15,219,227 ; pand %xmm11,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 102,68,15,112,218,0 ; pshufd $0x0,%xmm2,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 68,15,92,234 ; subps %xmm2,%xmm13 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 68,15,89,37,152,32,0,0 ; mulps 0x2098(%rip),%xmm12 # 4170 <_sk_callback_sse41+0x254> + DB 68,15,84,29,160,32,0,0 ; andps 0x20a0(%rip),%xmm11 # 4180 <_sk_callback_sse41+0x264> + DB 68,15,86,29,168,32,0,0 ; orps 0x20a8(%rip),%xmm11 # 4190 <_sk_callback_sse41+0x274> + DB 68,15,88,37,176,32,0,0 ; addps 0x20b0(%rip),%xmm12 # 41a0 <_sk_callback_sse41+0x284> + DB 15,40,21,185,32,0,0 ; movaps 0x20b9(%rip),%xmm2 # 41b0 <_sk_callback_sse41+0x294> DB 65,15,89,211 ; mulps %xmm11,%xmm2 - DB 68,15,92,234 ; subps %xmm2,%xmm13 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 65,15,88,211 ; addps %xmm11,%xmm2 - DB 68,15,94,226 ; divps %xmm2,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 102,69,15,58,8,221,1 ; roundps $0x1,%xmm13,%xmm11 - DB 185,0,0,0,75 ; mov $0x4b000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 - DB 69,15,92,235 ; subps %xmm11,%xmm13 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 65,15,89,213 ; mulps %xmm13,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 65,15,92,213 ; subps %xmm13,%xmm2 - DB 68,15,94,218 ; divps %xmm2,%xmm11 - DB 69,15,88,220 ; addps %xmm12,%xmm11 - DB 69,15,89,218 ; mulps %xmm10,%xmm11 - DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 + DB 68,15,88,29,185,32,0,0 ; addps 0x20b9(%rip),%xmm11 # 41c0 <_sk_callback_sse41+0x2a4> + DB 15,40,21,194,32,0,0 ; movaps 0x20c2(%rip),%xmm2 # 41d0 <_sk_callback_sse41+0x2b4> + 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 68,15,88,37,175,32,0,0 ; addps 0x20af(%rip),%xmm12 # 41e0 <_sk_callback_sse41+0x2c4> + DB 15,40,21,184,32,0,0 ; movaps 0x20b8(%rip),%xmm2 # 41f0 <_sk_callback_sse41+0x2d4> + DB 65,15,89,211 ; mulps %xmm11,%xmm2 + DB 68,15,92,226 ; subps %xmm2,%xmm12 + DB 68,15,40,21,184,32,0,0 ; movaps 0x20b8(%rip),%xmm10 # 4200 <_sk_callback_sse41+0x2e4> + DB 69,15,92,211 ; subps %xmm11,%xmm10 + DB 15,40,21,189,32,0,0 ; movaps 0x20bd(%rip),%xmm2 # 4210 <_sk_callback_sse41+0x2f4> + DB 65,15,94,210 ; divps %xmm10,%xmm2 + DB 65,15,88,212 ; addps %xmm12,%xmm2 + DB 15,89,21,190,32,0,0 ; mulps 0x20be(%rip),%xmm2 # 4220 <_sk_callback_sse41+0x304> + 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 DB 65,15,88,210 ; addps %xmm10,%xmm2 @@ -12300,62 +12137,32 @@ _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 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 102,68,15,112,227,0 ; pshufd $0x0,%xmm3,%xmm12 - DB 102,69,15,219,227 ; pand %xmm11,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 102,68,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 68,15,92,235 ; subps %xmm3,%xmm13 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 + DB 68,15,89,37,49,32,0,0 ; mulps 0x2031(%rip),%xmm12 # 4230 <_sk_callback_sse41+0x314> + DB 68,15,84,29,57,32,0,0 ; andps 0x2039(%rip),%xmm11 # 4240 <_sk_callback_sse41+0x324> + DB 68,15,86,29,65,32,0,0 ; orps 0x2041(%rip),%xmm11 # 4250 <_sk_callback_sse41+0x334> + DB 68,15,88,37,73,32,0,0 ; addps 0x2049(%rip),%xmm12 # 4260 <_sk_callback_sse41+0x344> + DB 15,40,29,82,32,0,0 ; movaps 0x2052(%rip),%xmm3 # 4270 <_sk_callback_sse41+0x354> DB 65,15,89,219 ; mulps %xmm11,%xmm3 - DB 68,15,92,235 ; subps %xmm3,%xmm13 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 65,15,88,219 ; addps %xmm11,%xmm3 - DB 68,15,94,227 ; divps %xmm3,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 102,69,15,58,8,221,1 ; roundps $0x1,%xmm13,%xmm11 - DB 185,0,0,0,75 ; mov $0x4b000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 - DB 69,15,92,235 ; subps %xmm11,%xmm13 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 65,15,89,221 ; mulps %xmm13,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 65,15,92,221 ; subps %xmm13,%xmm3 - DB 68,15,94,219 ; divps %xmm3,%xmm11 - DB 69,15,88,220 ; addps %xmm12,%xmm11 - DB 69,15,89,218 ; mulps %xmm10,%xmm11 - DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 + DB 68,15,88,29,82,32,0,0 ; addps 0x2052(%rip),%xmm11 # 4280 <_sk_callback_sse41+0x364> + DB 15,40,29,91,32,0,0 ; movaps 0x205b(%rip),%xmm3 # 4290 <_sk_callback_sse41+0x374> + 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 68,15,88,37,72,32,0,0 ; addps 0x2048(%rip),%xmm12 # 42a0 <_sk_callback_sse41+0x384> + DB 15,40,29,81,32,0,0 ; movaps 0x2051(%rip),%xmm3 # 42b0 <_sk_callback_sse41+0x394> + DB 65,15,89,219 ; mulps %xmm11,%xmm3 + DB 68,15,92,227 ; subps %xmm3,%xmm12 + DB 68,15,40,21,81,32,0,0 ; movaps 0x2051(%rip),%xmm10 # 42c0 <_sk_callback_sse41+0x3a4> + DB 69,15,92,211 ; subps %xmm11,%xmm10 + DB 15,40,29,86,32,0,0 ; movaps 0x2056(%rip),%xmm3 # 42d0 <_sk_callback_sse41+0x3b4> + DB 65,15,94,218 ; divps %xmm10,%xmm3 + DB 65,15,88,220 ; addps %xmm12,%xmm3 + DB 15,89,29,87,32,0,0 ; mulps 0x2057(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x3c4> + 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 DB 65,15,88,218 ; addps %xmm10,%xmm3 @@ -12574,9 +12381,9 @@ _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 276b <_sk_gather_i8_sse41+0xf> + DB 116,5 ; je 25c0 <_sk_gather_i8_sse41+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 276d <_sk_gather_i8_sse41+0x11> + DB 235,2 ; jmp 25c2 <_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 @@ -13002,59 +12809,57 @@ _sk_load_f16_sse41 LABEL PROC DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 DB 102,68,15,97,216 ; punpcklwd %xmm0,%xmm11 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,69,15,56,51,227 ; pmovzxwd %xmm11,%xmm12 - DB 184,0,128,0,0 ; mov $0x8000,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,68,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm8 - DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 + DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1 + DB 102,68,15,111,5,114,22,0,0 ; movdqa 0x1672(%rip),%xmm8 # 42f0 <_sk_callback_sse41+0x3d4> + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 - DB 102,68,15,239,226 ; pxor %xmm2,%xmm12 - DB 184,0,4,0,0 ; mov $0x400,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,15,112,216,0 ; pshufd $0x0,%xmm0,%xmm3 + DB 102,15,239,202 ; pxor %xmm2,%xmm1 + DB 102,15,111,29,109,22,0,0 ; movdqa 0x166d(%rip),%xmm3 # 4300 <_sk_callback_sse41+0x3e4> DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,65,15,102,196 ; pcmpgtd %xmm12,%xmm0 - DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 - DB 184,0,0,0,56 ; mov $0x38000000,%eax - DB 102,15,110,200 ; movd %eax,%xmm1 - DB 102,68,15,112,209,0 ; pshufd $0x0,%xmm1,%xmm10 - DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 - DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 - DB 102,15,223,194 ; pandn %xmm2,%xmm0 + 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 102,68,15,111,21,89,22,0,0 ; movdqa 0x1659(%rip),%xmm10 # 4310 <_sk_callback_sse41+0x3f4> + 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 DB 102,69,15,56,51,219 ; pmovzxwd %xmm11,%xmm11 DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 DB 102,68,15,239,218 ; pxor %xmm2,%xmm11 DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,102,203 ; pcmpgtd %xmm11,%xmm1 + DB 102,65,15,111,203 ; movdqa %xmm11,%xmm1 + DB 102,15,56,63,203 ; pmaxud %xmm3,%xmm1 + DB 102,65,15,118,203 ; pcmpeqd %xmm11,%xmm1 DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 - DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,15,223,202 ; pandn %xmm2,%xmm1 + DB 102,68,15,235,218 ; por %xmm2,%xmm11 + DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 + DB 102,65,15,219,203 ; pand %xmm11,%xmm1 DB 102,69,15,56,51,217 ; pmovzxwd %xmm9,%xmm11 DB 102,69,15,111,227 ; movdqa %xmm11,%xmm12 DB 102,69,15,219,224 ; pand %xmm8,%xmm12 DB 102,69,15,239,220 ; pxor %xmm12,%xmm11 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,65,15,102,211 ; pcmpgtd %xmm11,%xmm2 + DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 + DB 102,15,56,63,211 ; pmaxud %xmm3,%xmm2 + DB 102,65,15,118,211 ; pcmpeqd %xmm11,%xmm2 DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,69,15,254,226 ; paddd %xmm10,%xmm12 - DB 102,69,15,254,227 ; paddd %xmm11,%xmm12 - DB 102,65,15,223,212 ; pandn %xmm12,%xmm2 + DB 102,69,15,235,220 ; por %xmm12,%xmm11 + DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 + DB 102,65,15,219,211 ; pand %xmm11,%xmm2 DB 102,65,15,115,217,8 ; psrldq $0x8,%xmm9 DB 102,69,15,56,51,201 ; pmovzxwd %xmm9,%xmm9 DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 - DB 102,65,15,102,217 ; pcmpgtd %xmm9,%xmm3 + DB 102,65,15,56,63,217 ; pmaxud %xmm9,%xmm3 + DB 102,65,15,118,217 ; pcmpeqd %xmm9,%xmm3 DB 102,65,15,114,241,13 ; pslld $0xd,%xmm9 - DB 102,69,15,254,194 ; paddd %xmm10,%xmm8 - DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 - DB 102,65,15,223,216 ; pandn %xmm8,%xmm3 + DB 102,69,15,235,200 ; por %xmm8,%xmm9 + DB 102,69,15,254,202 ; paddd %xmm10,%xmm9 + DB 102,65,15,219,217 ; pand %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -13086,59 +12891,57 @@ _sk_gather_f16_sse41 LABEL PROC DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 DB 102,68,15,97,218 ; punpcklwd %xmm2,%xmm11 DB 102,68,15,105,202 ; punpckhwd %xmm2,%xmm9 - DB 102,69,15,56,51,227 ; pmovzxwd %xmm11,%xmm12 - DB 184,0,128,0,0 ; mov $0x8000,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,68,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm8 - DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 + DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1 + DB 102,68,15,111,5,23,21,0,0 ; movdqa 0x1517(%rip),%xmm8 # 4320 <_sk_callback_sse41+0x404> + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 - DB 102,68,15,239,226 ; pxor %xmm2,%xmm12 - DB 184,0,4,0,0 ; mov $0x400,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,15,112,216,0 ; pshufd $0x0,%xmm0,%xmm3 + DB 102,15,239,202 ; pxor %xmm2,%xmm1 + DB 102,15,111,29,18,21,0,0 ; movdqa 0x1512(%rip),%xmm3 # 4330 <_sk_callback_sse41+0x414> DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,65,15,102,196 ; pcmpgtd %xmm12,%xmm0 - DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 - DB 184,0,0,0,56 ; mov $0x38000000,%eax - DB 102,15,110,200 ; movd %eax,%xmm1 - DB 102,68,15,112,209,0 ; pshufd $0x0,%xmm1,%xmm10 - DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 - DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 - DB 102,15,223,194 ; pandn %xmm2,%xmm0 + 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 102,68,15,111,21,254,20,0,0 ; movdqa 0x14fe(%rip),%xmm10 # 4340 <_sk_callback_sse41+0x424> + 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 DB 102,69,15,56,51,219 ; pmovzxwd %xmm11,%xmm11 DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 DB 102,65,15,219,208 ; pand %xmm8,%xmm2 DB 102,68,15,239,218 ; pxor %xmm2,%xmm11 DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,102,203 ; pcmpgtd %xmm11,%xmm1 + DB 102,65,15,111,203 ; movdqa %xmm11,%xmm1 + DB 102,15,56,63,203 ; pmaxud %xmm3,%xmm1 + DB 102,65,15,118,203 ; pcmpeqd %xmm11,%xmm1 DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 - DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,15,223,202 ; pandn %xmm2,%xmm1 + DB 102,68,15,235,218 ; por %xmm2,%xmm11 + DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 + DB 102,65,15,219,203 ; pand %xmm11,%xmm1 DB 102,69,15,56,51,217 ; pmovzxwd %xmm9,%xmm11 DB 102,69,15,111,227 ; movdqa %xmm11,%xmm12 DB 102,69,15,219,224 ; pand %xmm8,%xmm12 DB 102,69,15,239,220 ; pxor %xmm12,%xmm11 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,65,15,102,211 ; pcmpgtd %xmm11,%xmm2 + DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 + DB 102,15,56,63,211 ; pmaxud %xmm3,%xmm2 + DB 102,65,15,118,211 ; pcmpeqd %xmm11,%xmm2 DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,69,15,254,226 ; paddd %xmm10,%xmm12 - DB 102,69,15,254,227 ; paddd %xmm11,%xmm12 - DB 102,65,15,223,212 ; pandn %xmm12,%xmm2 + DB 102,69,15,235,220 ; por %xmm12,%xmm11 + DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 + DB 102,65,15,219,211 ; pand %xmm11,%xmm2 DB 102,65,15,115,217,8 ; psrldq $0x8,%xmm9 DB 102,69,15,56,51,201 ; pmovzxwd %xmm9,%xmm9 DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 - DB 102,65,15,102,217 ; pcmpgtd %xmm9,%xmm3 + DB 102,65,15,56,63,217 ; pmaxud %xmm9,%xmm3 + DB 102,65,15,118,217 ; pcmpeqd %xmm9,%xmm3 DB 102,65,15,114,241,13 ; pslld $0xd,%xmm9 - DB 102,69,15,254,194 ; paddd %xmm10,%xmm8 - DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 - DB 102,65,15,223,216 ; pandn %xmm8,%xmm3 + DB 102,69,15,235,200 ; por %xmm8,%xmm9 + DB 102,69,15,254,202 ; paddd %xmm10,%xmm9 + DB 102,65,15,219,217 ; pand %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -13146,63 +12949,57 @@ 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 185,0,0,0,128 ; mov $0x80000000,%ecx - DB 102,68,15,110,193 ; movd %ecx,%xmm8 - DB 102,69,15,112,200,0 ; pshufd $0x0,%xmm8,%xmm9 - DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 - DB 102,68,15,219,224 ; pand %xmm0,%xmm12 + DB 102,68,15,111,21,52,20,0,0 ; movdqa 0x1434(%rip),%xmm10 # 4350 <_sk_callback_sse41+0x434> + DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 - DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 - DB 185,0,0,128,56 ; mov $0x38800000,%ecx - DB 102,68,15,110,193 ; movd %ecx,%xmm8 - DB 102,69,15,112,208,0 ; pshufd $0x0,%xmm8,%xmm10 - DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 - DB 102,69,15,111,194 ; movdqa %xmm10,%xmm8 - DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 - DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 185,0,192,1,0 ; mov $0x1c000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,219,0 ; pshufd $0x0,%xmm11,%xmm11 - DB 102,69,15,250,227 ; psubd %xmm11,%xmm12 - DB 102,69,15,254,229 ; paddd %xmm13,%xmm12 - DB 102,69,15,223,196 ; pandn %xmm12,%xmm8 + DB 102,69,15,219,234 ; pand %xmm10,%xmm13 + DB 102,69,15,239,229 ; pxor %xmm13,%xmm12 + DB 102,68,15,111,13,39,20,0,0 ; movdqa 0x1427(%rip),%xmm9 # 4360 <_sk_callback_sse41+0x444> + 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 102,68,15,111,29,24,20,0,0 ; movdqa 0x1418(%rip),%xmm11 # 4370 <_sk_callback_sse41+0x454> + 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 DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 - DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 - DB 102,68,15,219,233 ; pand %xmm1,%xmm13 + DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,69,15,239,245 ; pxor %xmm13,%xmm14 - DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13 - DB 102,69,15,111,226 ; movdqa %xmm10,%xmm12 - DB 102,69,15,102,230 ; pcmpgtd %xmm14,%xmm12 - DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,250,235 ; psubd %xmm11,%xmm13 - DB 102,69,15,254,238 ; paddd %xmm14,%xmm13 - DB 102,69,15,223,229 ; pandn %xmm13,%xmm12 + DB 102,69,15,219,242 ; pand %xmm10,%xmm14 + DB 102,69,15,239,238 ; pxor %xmm14,%xmm13 + DB 102,65,15,114,214,16 ; psrld $0x10,%xmm14 + DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 + DB 102,69,15,102,229 ; pcmpgtd %xmm13,%xmm12 + DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 + DB 102,69,15,235,243 ; por %xmm11,%xmm14 + DB 102,69,15,254,245 ; paddd %xmm13,%xmm14 + DB 102,69,15,223,230 ; pandn %xmm14,%xmm12 DB 102,69,15,56,43,228 ; packusdw %xmm12,%xmm12 - DB 102,69,15,111,241 ; movdqa %xmm9,%xmm14 - DB 102,68,15,219,242 ; pand %xmm2,%xmm14 + DB 102,68,15,111,242 ; movdqa %xmm2,%xmm14 DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 - DB 102,69,15,239,254 ; pxor %xmm14,%xmm15 - DB 102,65,15,114,214,16 ; psrld $0x10,%xmm14 - DB 102,69,15,111,234 ; movdqa %xmm10,%xmm13 - DB 102,69,15,102,239 ; pcmpgtd %xmm15,%xmm13 - DB 102,65,15,114,215,13 ; psrld $0xd,%xmm15 - DB 102,69,15,250,243 ; psubd %xmm11,%xmm14 - DB 102,69,15,254,247 ; paddd %xmm15,%xmm14 - DB 102,69,15,223,238 ; pandn %xmm14,%xmm13 + DB 102,69,15,219,250 ; pand %xmm10,%xmm15 + DB 102,69,15,239,247 ; pxor %xmm15,%xmm14 + DB 102,65,15,114,215,16 ; psrld $0x10,%xmm15 + DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 + DB 102,69,15,102,238 ; pcmpgtd %xmm14,%xmm13 + DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 + DB 102,69,15,235,251 ; por %xmm11,%xmm15 + DB 102,69,15,254,254 ; paddd %xmm14,%xmm15 + DB 102,69,15,223,239 ; pandn %xmm15,%xmm13 DB 102,69,15,56,43,237 ; packusdw %xmm13,%xmm13 - DB 102,68,15,219,203 ; pand %xmm3,%xmm9 + DB 102,68,15,219,211 ; pand %xmm3,%xmm10 DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,69,15,239,241 ; pxor %xmm9,%xmm14 - DB 102,65,15,114,209,16 ; psrld $0x10,%xmm9 - DB 102,69,15,102,214 ; pcmpgtd %xmm14,%xmm10 + DB 102,69,15,239,242 ; pxor %xmm10,%xmm14 + DB 102,65,15,114,210,16 ; psrld $0x10,%xmm10 + DB 102,69,15,102,206 ; pcmpgtd %xmm14,%xmm9 DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,250,203 ; psubd %xmm11,%xmm9 - DB 102,69,15,254,206 ; paddd %xmm14,%xmm9 - DB 102,69,15,223,209 ; pandn %xmm9,%xmm10 - DB 102,69,15,56,43,210 ; packusdw %xmm10,%xmm10 + DB 102,69,15,235,211 ; por %xmm11,%xmm10 + DB 102,69,15,254,214 ; paddd %xmm14,%xmm10 + DB 102,69,15,223,202 ; pandn %xmm10,%xmm9 + DB 102,69,15,56,43,201 ; packusdw %xmm9,%xmm9 DB 102,69,15,97,196 ; punpcklwd %xmm12,%xmm8 - DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 + DB 102,69,15,97,233 ; punpcklwd %xmm9,%xmm13 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,69,15,98,205 ; punpckldq %xmm13,%xmm9 DB 243,68,15,127,12,248 ; movdqu %xmm9,(%rax,%rdi,8) @@ -13751,7 +13548,7 @@ _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 3b5c <_sk_linear_gradient_sse41+0x13e> + DB 15,132,4,1,0,0 ; je 399e <_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) @@ -13802,13 +13599,13 @@ _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 3a84 <_sk_linear_gradient_sse41+0x66> + DB 15,133,65,255,255,255 ; jne 38c6 <_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 3b69 <_sk_linear_gradient_sse41+0x14b> + DB 235,13 ; jmp 39ab <_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 @@ -14251,8 +14048,423 @@ ALIGN 16 DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) DB 63 ; (bad) DB 0,0 ; add %al,(%rax) - DB 128 ; .byte 0x80 + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4004 <.literal16+0x34> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4008 <.literal16+0x38> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 400c <.literal16+0x3c> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4010 <.literal16+0x40> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 4095 <.literal16+0xc5> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 3ff9 <.literal16+0x29> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a3803a <_sk_callback_sse41+0xffffffffe9a3411e> + DB 220,63 ; fdivrl (%rdi) + DB 81 ; push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,52,0 ; add %dh,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 40c4 <.literal16+0xf4> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 40c8 <.literal16+0xf8> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 40cc <.literal16+0xfc> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 40d0 <.literal16+0x100> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 4155 <.literal16+0x185> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 40b9 <.literal16+0xe9> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a380fa <_sk_callback_sse41+0xffffffffe9a341de> + DB 220,63 ; fdivrl (%rdi) + DB 81 ; push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,52,0 ; add %dh,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4184 <.literal16+0x1b4> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4188 <.literal16+0x1b8> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 418c <.literal16+0x1bc> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4190 <.literal16+0x1c0> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 4215 <.literal16+0x245> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 4179 <.literal16+0x1a9> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a381ba <_sk_callback_sse41+0xffffffffe9a3429e> + DB 220,63 ; fdivrl (%rdi) + DB 81 ; push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,52,0 ; add %dh,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4244 <.literal16+0x274> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4248 <.literal16+0x278> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 424c <.literal16+0x27c> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4250 <.literal16+0x280> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 42d5 <.literal16+0x305> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 4239 <.literal16+0x269> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a3827a <_sk_callback_sse41+0xffffffffe9a3435e> + DB 220,63 ; fdivrl (%rdi) + DB 81 ; push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,128,0,0,0,128 ; rex.WXB add %al,-0x80000000(%r8) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 128,0,0 ; addb $0x0,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,4 ; add %al,0x4000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 128,0,0 ; addb $0x0,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,128,56 ; add %al,0x38800000(%rax) + DB 0,0 ; add %al,(%rax) + DB 128,56,0 ; cmpb $0x0,(%rax) + DB 0,128,56,0,0,128 ; add %al,-0x7fffffc8(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 64,254 ; rex (bad) + DB 255,0 ; incl (%rax) + DB 64,254 ; rex (bad) + DB 255,0 ; incl (%rax) + DB 64,254 ; rex (bad) + DB 255,0 ; incl (%rax) + DB 64,254 ; rex (bad) + DB 255 ; .byte 0xff ALIGN 32 PUBLIC _sk_start_pipeline_sse2 @@ -14334,7 +14546,7 @@ _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 15,40,21,129,68,0,0 ; movaps 0x4481(%rip),%xmm2 # 4590 <_sk_callback_sse2+0xb3> + DB 15,40,21,177,66,0,0 ; movaps 0x42b1(%rip),%xmm2 # 43c0 <_sk_callback_sse2+0xab> DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,16,2 ; movups (%rdx),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 @@ -14343,7 +14555,7 @@ _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 15,40,21,112,68,0,0 ; movaps 0x4470(%rip),%xmm2 # 45a0 <_sk_callback_sse2+0xc3> + DB 15,40,21,160,66,0,0 ; movaps 0x42a0(%rip),%xmm2 # 43d0 <_sk_callback_sse2+0xbb> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -16450,92 +16662,62 @@ _sk_parametric_r_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,192 ; mulps %xmm0,%xmm8 - DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 68,15,89,216 ; mulps %xmm0,%xmm11 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,208 ; mulps %xmm0,%xmm10 DB 65,15,194,193,2 ; cmpleps %xmm9,%xmm0 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 - DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,218 ; addps %xmm10,%xmm11 - DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,212 ; mulps %xmm12,%xmm10 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 - DB 102,69,15,219,235 ; pand %xmm11,%xmm13 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,235,229 ; por %xmm13,%xmm12 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,89,220 ; mulps %xmm12,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 243,68,15,16,24 ; movss (%rax),%xmm11 + DB 243,68,15,16,72,8 ; movss 0x8(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,244 ; addps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 - DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 - DB 69,15,40,234 ; movaps %xmm10,%xmm13 - DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,250 ; addps %xmm10,%xmm15 - DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 + DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,84,211 ; andps %xmm11,%xmm10 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,226 ; subps %xmm10,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 102,69,15,110,208 ; movd %r8d,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,229 ; mulps %xmm13,%xmm12 - DB 69,15,92,252 ; subps %xmm12,%xmm15 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,245 ; subps %xmm13,%xmm14 - DB 69,15,94,230 ; divps %xmm14,%xmm12 - DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 + DB 68,15,89,13,78,36,0,0 ; mulps 0x244e(%rip),%xmm9 # 43e0 <_sk_callback_sse2+0xcb> + DB 68,15,84,21,86,36,0,0 ; andps 0x2456(%rip),%xmm10 # 43f0 <_sk_callback_sse2+0xdb> + DB 68,15,86,21,94,36,0,0 ; orps 0x245e(%rip),%xmm10 # 4400 <_sk_callback_sse2+0xeb> + DB 68,15,88,13,102,36,0,0 ; addps 0x2466(%rip),%xmm9 # 4410 <_sk_callback_sse2+0xfb> + DB 68,15,40,37,110,36,0,0 ; movaps 0x246e(%rip),%xmm12 # 4420 <_sk_callback_sse2+0x10b> DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 - DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,226 ; addps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 68,15,88,21,110,36,0,0 ; addps 0x246e(%rip),%xmm10 # 4430 <_sk_callback_sse2+0x11b> + DB 68,15,40,37,118,36,0,0 ; movaps 0x2476(%rip),%xmm12 # 4440 <_sk_callback_sse2+0x12b> + DB 69,15,94,226 ; divps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 69,15,89,203 ; mulps %xmm11,%xmm9 + DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 + DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12 + DB 68,15,84,37,96,36,0,0 ; andps 0x2460(%rip),%xmm12 # 4450 <_sk_callback_sse2+0x13b> + DB 69,15,87,210 ; xorps %xmm10,%xmm10 + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 68,15,88,13,88,36,0,0 ; addps 0x2458(%rip),%xmm9 # 4460 <_sk_callback_sse2+0x14b> + DB 68,15,40,29,96,36,0,0 ; movaps 0x2460(%rip),%xmm11 # 4470 <_sk_callback_sse2+0x15b> + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,203 ; subps %xmm11,%xmm9 + DB 68,15,40,29,96,36,0,0 ; movaps 0x2460(%rip),%xmm11 # 4480 <_sk_callback_sse2+0x16b> + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 68,15,40,37,100,36,0,0 ; movaps 0x2464(%rip),%xmm12 # 4490 <_sk_callback_sse2+0x17b> + DB 69,15,94,227 ; divps %xmm11,%xmm12 + DB 69,15,88,225 ; addps %xmm9,%xmm12 + DB 68,15,89,37,100,36,0,0 ; mulps 0x2464(%rip),%xmm12 # 44a0 <_sk_callback_sse2+0x18b> + DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9 + DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,88,217 ; addps %xmm9,%xmm11 DB 68,15,84,192 ; andps %xmm0,%xmm8 - DB 65,15,85,196 ; andnps %xmm12,%xmm0 + DB 65,15,85,195 ; andnps %xmm11,%xmm0 DB 65,15,86,192 ; orps %xmm8,%xmm0 - DB 65,15,95,193 ; maxps %xmm9,%xmm0 - DB 65,15,93,195 ; minps %xmm11,%xmm0 + DB 65,15,95,194 ; maxps %xmm10,%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 65,15,93,192 ; minps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16547,92 +16729,62 @@ _sk_parametric_g_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,193 ; mulps %xmm1,%xmm8 - DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 68,15,89,217 ; mulps %xmm1,%xmm11 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,209 ; mulps %xmm1,%xmm10 DB 65,15,194,201,2 ; cmpleps %xmm9,%xmm1 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 - DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,218 ; addps %xmm10,%xmm11 - DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,212 ; mulps %xmm12,%xmm10 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 - DB 102,69,15,219,235 ; pand %xmm11,%xmm13 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,235,229 ; por %xmm13,%xmm12 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,89,220 ; mulps %xmm12,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 243,68,15,16,24 ; movss (%rax),%xmm11 + DB 243,68,15,16,72,8 ; movss 0x8(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,244 ; addps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 - DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 - DB 69,15,40,234 ; movaps %xmm10,%xmm13 - DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,250 ; addps %xmm10,%xmm15 - DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 + DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,84,211 ; andps %xmm11,%xmm10 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,226 ; subps %xmm10,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 102,69,15,110,208 ; movd %r8d,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,229 ; mulps %xmm13,%xmm12 - DB 69,15,92,252 ; subps %xmm12,%xmm15 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,245 ; subps %xmm13,%xmm14 - DB 69,15,94,230 ; divps %xmm14,%xmm12 - DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 + DB 68,15,89,13,213,35,0,0 ; mulps 0x23d5(%rip),%xmm9 # 44b0 <_sk_callback_sse2+0x19b> + DB 68,15,84,21,221,35,0,0 ; andps 0x23dd(%rip),%xmm10 # 44c0 <_sk_callback_sse2+0x1ab> + DB 68,15,86,21,229,35,0,0 ; orps 0x23e5(%rip),%xmm10 # 44d0 <_sk_callback_sse2+0x1bb> + DB 68,15,88,13,237,35,0,0 ; addps 0x23ed(%rip),%xmm9 # 44e0 <_sk_callback_sse2+0x1cb> + DB 68,15,40,37,245,35,0,0 ; movaps 0x23f5(%rip),%xmm12 # 44f0 <_sk_callback_sse2+0x1db> DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 - DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,226 ; addps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 68,15,88,21,245,35,0,0 ; addps 0x23f5(%rip),%xmm10 # 4500 <_sk_callback_sse2+0x1eb> + DB 68,15,40,37,253,35,0,0 ; movaps 0x23fd(%rip),%xmm12 # 4510 <_sk_callback_sse2+0x1fb> + DB 69,15,94,226 ; divps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 69,15,89,203 ; mulps %xmm11,%xmm9 + DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 + DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12 + DB 68,15,84,37,231,35,0,0 ; andps 0x23e7(%rip),%xmm12 # 4520 <_sk_callback_sse2+0x20b> + DB 69,15,87,210 ; xorps %xmm10,%xmm10 + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 68,15,88,13,223,35,0,0 ; addps 0x23df(%rip),%xmm9 # 4530 <_sk_callback_sse2+0x21b> + DB 68,15,40,29,231,35,0,0 ; movaps 0x23e7(%rip),%xmm11 # 4540 <_sk_callback_sse2+0x22b> + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,203 ; subps %xmm11,%xmm9 + DB 68,15,40,29,231,35,0,0 ; movaps 0x23e7(%rip),%xmm11 # 4550 <_sk_callback_sse2+0x23b> + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 68,15,40,37,235,35,0,0 ; movaps 0x23eb(%rip),%xmm12 # 4560 <_sk_callback_sse2+0x24b> + DB 69,15,94,227 ; divps %xmm11,%xmm12 + DB 69,15,88,225 ; addps %xmm9,%xmm12 + DB 68,15,89,37,235,35,0,0 ; mulps 0x23eb(%rip),%xmm12 # 4570 <_sk_callback_sse2+0x25b> + DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9 + DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,88,217 ; addps %xmm9,%xmm11 DB 68,15,84,193 ; andps %xmm1,%xmm8 - DB 65,15,85,204 ; andnps %xmm12,%xmm1 + DB 65,15,85,203 ; andnps %xmm11,%xmm1 DB 65,15,86,200 ; orps %xmm8,%xmm1 - DB 65,15,95,201 ; maxps %xmm9,%xmm1 - DB 65,15,93,203 ; minps %xmm11,%xmm1 + DB 65,15,95,202 ; maxps %xmm10,%xmm1 + 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 65,15,93,200 ; minps %xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16644,92 +16796,62 @@ _sk_parametric_b_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,194 ; mulps %xmm2,%xmm8 - DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 68,15,89,218 ; mulps %xmm2,%xmm11 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,210 ; mulps %xmm2,%xmm10 DB 65,15,194,209,2 ; cmpleps %xmm9,%xmm2 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 - DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,218 ; addps %xmm10,%xmm11 - DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,212 ; mulps %xmm12,%xmm10 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 - DB 102,69,15,219,235 ; pand %xmm11,%xmm13 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,235,229 ; por %xmm13,%xmm12 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,89,220 ; mulps %xmm12,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 243,68,15,16,24 ; movss (%rax),%xmm11 + DB 243,68,15,16,72,8 ; movss 0x8(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,244 ; addps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 - DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 - DB 69,15,40,234 ; movaps %xmm10,%xmm13 - DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,250 ; addps %xmm10,%xmm15 - DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 + DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,84,211 ; andps %xmm11,%xmm10 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,226 ; subps %xmm10,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 102,69,15,110,208 ; movd %r8d,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,229 ; mulps %xmm13,%xmm12 - DB 69,15,92,252 ; subps %xmm12,%xmm15 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,245 ; subps %xmm13,%xmm14 - DB 69,15,94,230 ; divps %xmm14,%xmm12 - DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 + DB 68,15,89,13,92,35,0,0 ; mulps 0x235c(%rip),%xmm9 # 4580 <_sk_callback_sse2+0x26b> + DB 68,15,84,21,100,35,0,0 ; andps 0x2364(%rip),%xmm10 # 4590 <_sk_callback_sse2+0x27b> + DB 68,15,86,21,108,35,0,0 ; orps 0x236c(%rip),%xmm10 # 45a0 <_sk_callback_sse2+0x28b> + DB 68,15,88,13,116,35,0,0 ; addps 0x2374(%rip),%xmm9 # 45b0 <_sk_callback_sse2+0x29b> + DB 68,15,40,37,124,35,0,0 ; movaps 0x237c(%rip),%xmm12 # 45c0 <_sk_callback_sse2+0x2ab> DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 - DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,226 ; addps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 68,15,88,21,124,35,0,0 ; addps 0x237c(%rip),%xmm10 # 45d0 <_sk_callback_sse2+0x2bb> + DB 68,15,40,37,132,35,0,0 ; movaps 0x2384(%rip),%xmm12 # 45e0 <_sk_callback_sse2+0x2cb> + DB 69,15,94,226 ; divps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 69,15,89,203 ; mulps %xmm11,%xmm9 + DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 + DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12 + DB 68,15,84,37,110,35,0,0 ; andps 0x236e(%rip),%xmm12 # 45f0 <_sk_callback_sse2+0x2db> + DB 69,15,87,210 ; xorps %xmm10,%xmm10 + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 68,15,88,13,102,35,0,0 ; addps 0x2366(%rip),%xmm9 # 4600 <_sk_callback_sse2+0x2eb> + DB 68,15,40,29,110,35,0,0 ; movaps 0x236e(%rip),%xmm11 # 4610 <_sk_callback_sse2+0x2fb> + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,203 ; subps %xmm11,%xmm9 + DB 68,15,40,29,110,35,0,0 ; movaps 0x236e(%rip),%xmm11 # 4620 <_sk_callback_sse2+0x30b> + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 68,15,40,37,114,35,0,0 ; movaps 0x2372(%rip),%xmm12 # 4630 <_sk_callback_sse2+0x31b> + DB 69,15,94,227 ; divps %xmm11,%xmm12 + DB 69,15,88,225 ; addps %xmm9,%xmm12 + DB 68,15,89,37,114,35,0,0 ; mulps 0x2372(%rip),%xmm12 # 4640 <_sk_callback_sse2+0x32b> + DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9 + DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,88,217 ; addps %xmm9,%xmm11 DB 68,15,84,194 ; andps %xmm2,%xmm8 - DB 65,15,85,212 ; andnps %xmm12,%xmm2 + DB 65,15,85,211 ; andnps %xmm11,%xmm2 DB 65,15,86,208 ; orps %xmm8,%xmm2 - DB 65,15,95,209 ; maxps %xmm9,%xmm2 - DB 65,15,93,211 ; minps %xmm11,%xmm2 + DB 65,15,95,210 ; maxps %xmm10,%xmm2 + 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 65,15,93,208 ; minps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16741,92 +16863,62 @@ _sk_parametric_a_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,195 ; mulps %xmm3,%xmm8 - DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 68,15,89,219 ; mulps %xmm3,%xmm11 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,211 ; mulps %xmm3,%xmm10 DB 65,15,194,217,2 ; cmpleps %xmm9,%xmm3 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 - DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,218 ; addps %xmm10,%xmm11 - DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,212 ; mulps %xmm12,%xmm10 - DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 - DB 102,69,15,219,235 ; pand %xmm11,%xmm13 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,235,229 ; por %xmm13,%xmm12 - DB 185,119,115,248,66 ; mov $0x42f87377,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,89,220 ; mulps %xmm12,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d - DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 243,68,15,16,24 ; movss (%rax),%xmm11 + DB 243,68,15,16,72,8 ; movss 0x8(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,244 ; addps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 - DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 - DB 69,15,40,234 ; movaps %xmm10,%xmm13 - DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,250 ; addps %xmm10,%xmm15 - DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 + DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,84,211 ; andps %xmm11,%xmm10 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,226 ; subps %xmm10,%xmm12 - DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 102,69,15,110,208 ; movd %r8d,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,229 ; mulps %xmm13,%xmm12 - DB 69,15,92,252 ; subps %xmm12,%xmm15 - DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx - DB 102,68,15,110,241 ; movd %ecx,%xmm14 - DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,245 ; subps %xmm13,%xmm14 - DB 69,15,94,230 ; divps %xmm14,%xmm12 - DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 + DB 68,15,89,13,227,34,0,0 ; mulps 0x22e3(%rip),%xmm9 # 4650 <_sk_callback_sse2+0x33b> + DB 68,15,84,21,235,34,0,0 ; andps 0x22eb(%rip),%xmm10 # 4660 <_sk_callback_sse2+0x34b> + DB 68,15,86,21,243,34,0,0 ; orps 0x22f3(%rip),%xmm10 # 4670 <_sk_callback_sse2+0x35b> + DB 68,15,88,13,251,34,0,0 ; addps 0x22fb(%rip),%xmm9 # 4680 <_sk_callback_sse2+0x36b> + DB 68,15,40,37,3,35,0,0 ; movaps 0x2303(%rip),%xmm12 # 4690 <_sk_callback_sse2+0x37b> DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 - DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,226 ; addps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 68,15,88,21,3,35,0,0 ; addps 0x2303(%rip),%xmm10 # 46a0 <_sk_callback_sse2+0x38b> + DB 68,15,40,37,11,35,0,0 ; movaps 0x230b(%rip),%xmm12 # 46b0 <_sk_callback_sse2+0x39b> + DB 69,15,94,226 ; divps %xmm10,%xmm12 + DB 69,15,92,204 ; subps %xmm12,%xmm9 + DB 69,15,89,203 ; mulps %xmm11,%xmm9 + DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 + DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12 + DB 68,15,84,37,245,34,0,0 ; andps 0x22f5(%rip),%xmm12 # 46c0 <_sk_callback_sse2+0x3ab> + DB 69,15,87,210 ; xorps %xmm10,%xmm10 + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 68,15,88,13,237,34,0,0 ; addps 0x22ed(%rip),%xmm9 # 46d0 <_sk_callback_sse2+0x3bb> + DB 68,15,40,29,245,34,0,0 ; movaps 0x22f5(%rip),%xmm11 # 46e0 <_sk_callback_sse2+0x3cb> + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,203 ; subps %xmm11,%xmm9 + DB 68,15,40,29,245,34,0,0 ; movaps 0x22f5(%rip),%xmm11 # 46f0 <_sk_callback_sse2+0x3db> + DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 68,15,40,37,249,34,0,0 ; movaps 0x22f9(%rip),%xmm12 # 4700 <_sk_callback_sse2+0x3eb> + DB 69,15,94,227 ; divps %xmm11,%xmm12 + DB 69,15,88,225 ; addps %xmm9,%xmm12 + DB 68,15,89,37,249,34,0,0 ; mulps 0x22f9(%rip),%xmm12 # 4710 <_sk_callback_sse2+0x3fb> + DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9 + DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,88,217 ; addps %xmm9,%xmm11 DB 68,15,84,195 ; andps %xmm3,%xmm8 - DB 65,15,85,220 ; andnps %xmm12,%xmm3 + DB 65,15,85,219 ; andnps %xmm11,%xmm3 DB 65,15,86,216 ; orps %xmm8,%xmm3 - DB 65,15,95,217 ; maxps %xmm9,%xmm3 - DB 65,15,93,219 ; minps %xmm11,%xmm3 + DB 65,15,95,218 ; maxps %xmm10,%xmm3 + 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 65,15,93,216 ; minps %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -17070,9 +17162,9 @@ _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 2971 <_sk_gather_i8_sse2+0xf> + DB 116,5 ; je 27bd <_sk_gather_i8_sse2+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 2973 <_sk_gather_i8_sse2+0x11> + DB 235,2 ; jmp 27bf <_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 @@ -17544,65 +17636,68 @@ _sk_load_f16_sse2 LABEL PROC DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 DB 102,68,15,97,193 ; punpcklwd %xmm1,%xmm8 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,69,15,111,224 ; movdqa %xmm8,%xmm12 - DB 102,68,15,97,224 ; punpcklwd %xmm0,%xmm12 + DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 + DB 102,68,15,97,240 ; punpcklwd %xmm0,%xmm14 DB 102,68,15,105,192 ; punpckhwd %xmm0,%xmm8 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 - DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 - DB 184,0,128,0,0 ; mov $0x8000,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,68,15,112,200,0 ; pshufd $0x0,%xmm0,%xmm9 - DB 102,65,15,111,205 ; movdqa %xmm13,%xmm1 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 + DB 102,68,15,111,13,232,23,0,0 ; movdqa 0x17e8(%rip),%xmm9 # 4720 <_sk_callback_sse2+0x40b> + 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 + DB 102,15,114,240,16 ; pslld $0x10,%xmm0 + 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 102,68,15,111,29,205,23,0,0 ; movdqa 0x17cd(%rip),%xmm11 # 4730 <_sk_callback_sse2+0x41b> + DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 + DB 102,68,15,111,37,207,23,0,0 ; movdqa 0x17cf(%rip),%xmm12 # 4740 <_sk_callback_sse2+0x42b> + DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 + DB 102,15,111,29,210,23,0,0 ; movdqa 0x17d2(%rip),%xmm3 # 4750 <_sk_callback_sse2+0x43b> + 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 + DB 102,65,15,115,222,8 ; psrldq $0x8,%xmm14 + DB 102,69,15,97,242 ; punpcklwd %xmm10,%xmm14 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 DB 102,65,15,219,201 ; pand %xmm9,%xmm1 - DB 102,68,15,239,233 ; pxor %xmm1,%xmm13 - DB 184,0,4,0,0 ; mov $0x400,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,15,112,216,0 ; pshufd $0x0,%xmm0,%xmm3 + DB 102,68,15,239,241 ; pxor %xmm1,%xmm14 DB 102,15,114,241,16 ; pslld $0x10,%xmm1 - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,65,15,102,197 ; pcmpgtd %xmm13,%xmm0 - DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 - DB 184,0,0,0,56 ; mov $0x38000000,%eax - DB 102,15,110,208 ; movd %eax,%xmm2 - DB 102,68,15,112,218,0 ; pshufd $0x0,%xmm2,%xmm11 - DB 102,65,15,254,203 ; paddd %xmm11,%xmm1 - DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,65,15,115,220,8 ; psrldq $0x8,%xmm12 - DB 102,69,15,97,226 ; punpcklwd %xmm10,%xmm12 - DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 - DB 102,65,15,219,209 ; pand %xmm9,%xmm2 - DB 102,68,15,239,226 ; pxor %xmm2,%xmm12 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,102,204 ; pcmpgtd %xmm12,%xmm1 - DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 + DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,15,235,209 ; por %xmm1,%xmm2 DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 + DB 102,69,15,239,244 ; pxor %xmm12,%xmm14 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,65,15,102,206 ; pcmpgtd %xmm14,%xmm1 DB 102,15,223,202 ; pandn %xmm2,%xmm1 - DB 102,69,15,111,224 ; movdqa %xmm8,%xmm12 - DB 102,69,15,97,226 ; punpcklwd %xmm10,%xmm12 - DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 - DB 102,69,15,219,233 ; pand %xmm9,%xmm13 - DB 102,69,15,239,229 ; pxor %xmm13,%xmm12 - DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 + DB 102,69,15,111,232 ; movdqa %xmm8,%xmm13 + DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 + DB 102,65,15,111,213 ; movdqa %xmm13,%xmm2 + DB 102,65,15,219,209 ; pand %xmm9,%xmm2 + DB 102,68,15,239,234 ; pxor %xmm2,%xmm13 + DB 102,15,114,242,16 ; pslld $0x10,%xmm2 + DB 102,69,15,111,245 ; movdqa %xmm13,%xmm14 + DB 102,65,15,114,246,13 ; pslld $0xd,%xmm14 + DB 102,68,15,235,242 ; por %xmm2,%xmm14 + DB 102,69,15,254,243 ; paddd %xmm11,%xmm14 + DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,65,15,102,212 ; pcmpgtd %xmm12,%xmm2 - DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 - DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 - DB 102,65,15,223,213 ; pandn %xmm13,%xmm2 + DB 102,65,15,102,213 ; pcmpgtd %xmm13,%xmm2 + DB 102,65,15,223,214 ; pandn %xmm14,%xmm2 DB 102,65,15,115,216,8 ; psrldq $0x8,%xmm8 DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8 DB 102,69,15,219,200 ; pand %xmm8,%xmm9 DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 + DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 + DB 102,65,15,114,242,13 ; pslld $0xd,%xmm10 + DB 102,69,15,235,209 ; por %xmm9,%xmm10 + DB 102,69,15,254,211 ; paddd %xmm11,%xmm10 + DB 102,69,15,239,196 ; pxor %xmm12,%xmm8 DB 102,65,15,102,216 ; pcmpgtd %xmm8,%xmm3 - DB 102,65,15,114,240,13 ; pslld $0xd,%xmm8 - DB 102,69,15,254,203 ; paddd %xmm11,%xmm9 - DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 - DB 102,65,15,223,217 ; pandn %xmm9,%xmm3 + DB 102,65,15,223,218 ; pandn %xmm10,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -17637,65 +17732,68 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 DB 102,68,15,97,194 ; punpcklwd %xmm2,%xmm8 DB 102,15,105,202 ; punpckhwd %xmm2,%xmm1 - DB 102,69,15,111,224 ; movdqa %xmm8,%xmm12 - DB 102,68,15,97,225 ; punpcklwd %xmm1,%xmm12 + DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 + DB 102,68,15,97,241 ; punpcklwd %xmm1,%xmm14 DB 102,68,15,105,193 ; punpckhwd %xmm1,%xmm8 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 - DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 - DB 184,0,128,0,0 ; mov $0x8000,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,68,15,112,200,0 ; pshufd $0x0,%xmm0,%xmm9 - DB 102,65,15,111,205 ; movdqa %xmm13,%xmm1 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 + DB 102,68,15,111,13,96,22,0,0 ; movdqa 0x1660(%rip),%xmm9 # 4760 <_sk_callback_sse2+0x44b> + 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 + DB 102,15,114,240,16 ; pslld $0x10,%xmm0 + 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 102,68,15,111,29,69,22,0,0 ; movdqa 0x1645(%rip),%xmm11 # 4770 <_sk_callback_sse2+0x45b> + DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 + DB 102,68,15,111,37,71,22,0,0 ; movdqa 0x1647(%rip),%xmm12 # 4780 <_sk_callback_sse2+0x46b> + DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 + DB 102,15,111,29,74,22,0,0 ; movdqa 0x164a(%rip),%xmm3 # 4790 <_sk_callback_sse2+0x47b> + 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 + DB 102,65,15,115,222,8 ; psrldq $0x8,%xmm14 + DB 102,69,15,97,242 ; punpcklwd %xmm10,%xmm14 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 DB 102,65,15,219,201 ; pand %xmm9,%xmm1 - DB 102,68,15,239,233 ; pxor %xmm1,%xmm13 - DB 184,0,4,0,0 ; mov $0x400,%eax - DB 102,15,110,192 ; movd %eax,%xmm0 - DB 102,15,112,216,0 ; pshufd $0x0,%xmm0,%xmm3 + DB 102,68,15,239,241 ; pxor %xmm1,%xmm14 DB 102,15,114,241,16 ; pslld $0x10,%xmm1 - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,65,15,102,197 ; pcmpgtd %xmm13,%xmm0 - DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 - DB 184,0,0,0,56 ; mov $0x38000000,%eax - DB 102,15,110,208 ; movd %eax,%xmm2 - DB 102,68,15,112,218,0 ; pshufd $0x0,%xmm2,%xmm11 - DB 102,65,15,254,203 ; paddd %xmm11,%xmm1 - DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,65,15,115,220,8 ; psrldq $0x8,%xmm12 - DB 102,69,15,97,226 ; punpcklwd %xmm10,%xmm12 - DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 - DB 102,65,15,219,209 ; pand %xmm9,%xmm2 - DB 102,68,15,239,226 ; pxor %xmm2,%xmm12 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,102,204 ; pcmpgtd %xmm12,%xmm1 - DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 + DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,15,235,209 ; por %xmm1,%xmm2 DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 + DB 102,69,15,239,244 ; pxor %xmm12,%xmm14 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,65,15,102,206 ; pcmpgtd %xmm14,%xmm1 DB 102,15,223,202 ; pandn %xmm2,%xmm1 - DB 102,69,15,111,224 ; movdqa %xmm8,%xmm12 - DB 102,69,15,97,226 ; punpcklwd %xmm10,%xmm12 - DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 - DB 102,69,15,219,233 ; pand %xmm9,%xmm13 - DB 102,69,15,239,229 ; pxor %xmm13,%xmm12 - DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 + DB 102,69,15,111,232 ; movdqa %xmm8,%xmm13 + DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 + DB 102,65,15,111,213 ; movdqa %xmm13,%xmm2 + DB 102,65,15,219,209 ; pand %xmm9,%xmm2 + DB 102,68,15,239,234 ; pxor %xmm2,%xmm13 + DB 102,15,114,242,16 ; pslld $0x10,%xmm2 + DB 102,69,15,111,245 ; movdqa %xmm13,%xmm14 + DB 102,65,15,114,246,13 ; pslld $0xd,%xmm14 + DB 102,68,15,235,242 ; por %xmm2,%xmm14 + DB 102,69,15,254,243 ; paddd %xmm11,%xmm14 + DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,65,15,102,212 ; pcmpgtd %xmm12,%xmm2 - DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 - DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 - DB 102,65,15,223,213 ; pandn %xmm13,%xmm2 + DB 102,65,15,102,213 ; pcmpgtd %xmm13,%xmm2 + DB 102,65,15,223,214 ; pandn %xmm14,%xmm2 DB 102,65,15,115,216,8 ; psrldq $0x8,%xmm8 DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8 DB 102,69,15,219,200 ; pand %xmm8,%xmm9 DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 + DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 + DB 102,65,15,114,242,13 ; pslld $0xd,%xmm10 + DB 102,69,15,235,209 ; por %xmm9,%xmm10 + DB 102,69,15,254,211 ; paddd %xmm11,%xmm10 + DB 102,69,15,239,196 ; pxor %xmm12,%xmm8 DB 102,65,15,102,216 ; pcmpgtd %xmm8,%xmm3 - DB 102,65,15,114,240,13 ; pslld $0xd,%xmm8 - DB 102,69,15,254,203 ; paddd %xmm11,%xmm9 - DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 - DB 102,65,15,223,217 ; pandn %xmm9,%xmm3 + DB 102,65,15,223,218 ; pandn %xmm10,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -17703,71 +17801,65 @@ 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 185,0,0,0,128 ; mov $0x80000000,%ecx - DB 102,68,15,110,193 ; movd %ecx,%xmm8 - DB 102,69,15,112,200,0 ; pshufd $0x0,%xmm8,%xmm9 - DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 - DB 102,68,15,219,224 ; pand %xmm0,%xmm12 + DB 102,68,15,111,21,114,21,0,0 ; movdqa 0x1572(%rip),%xmm10 # 47a0 <_sk_callback_sse2+0x48b> + DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 - DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 - DB 185,0,0,128,56 ; mov $0x38800000,%ecx - DB 102,68,15,110,193 ; movd %ecx,%xmm8 - DB 102,69,15,112,208,0 ; pshufd $0x0,%xmm8,%xmm10 - DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 - DB 102,69,15,111,194 ; movdqa %xmm10,%xmm8 - DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 - DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 185,0,192,1,0 ; mov $0x1c000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,219,0 ; pshufd $0x0,%xmm11,%xmm11 - DB 102,69,15,250,227 ; psubd %xmm11,%xmm12 - DB 102,69,15,254,229 ; paddd %xmm13,%xmm12 - DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,65,15,114,228,16 ; psrad $0x10,%xmm12 - DB 102,69,15,223,196 ; pandn %xmm12,%xmm8 - DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 - DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 - DB 102,68,15,219,233 ; pand %xmm1,%xmm13 - DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,69,15,239,245 ; pxor %xmm13,%xmm14 + DB 102,69,15,219,234 ; pand %xmm10,%xmm13 + DB 102,69,15,239,229 ; pxor %xmm13,%xmm12 + DB 102,68,15,111,13,101,21,0,0 ; movdqa 0x1565(%rip),%xmm9 # 47b0 <_sk_callback_sse2+0x49b> DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13 - DB 102,69,15,111,226 ; movdqa %xmm10,%xmm12 - DB 102,69,15,102,230 ; pcmpgtd %xmm14,%xmm12 - DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,250,235 ; psubd %xmm11,%xmm13 - DB 102,69,15,254,238 ; paddd %xmm14,%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 102,68,15,111,29,86,21,0,0 ; movdqa 0x1556(%rip),%xmm11 # 47c0 <_sk_callback_sse2+0x4ab> + 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 DB 102,65,15,114,229,16 ; psrad $0x10,%xmm13 - DB 102,69,15,223,229 ; pandn %xmm13,%xmm12 - DB 102,69,15,107,228 ; packssdw %xmm12,%xmm12 - DB 102,69,15,111,241 ; movdqa %xmm9,%xmm14 - DB 102,68,15,219,242 ; pand %xmm2,%xmm14 - DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 - DB 102,69,15,239,254 ; pxor %xmm14,%xmm15 + DB 102,69,15,223,197 ; pandn %xmm13,%xmm8 + DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 + DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 + DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 + DB 102,69,15,219,242 ; pand %xmm10,%xmm14 + DB 102,69,15,239,238 ; pxor %xmm14,%xmm13 DB 102,65,15,114,214,16 ; psrld $0x10,%xmm14 - DB 102,69,15,111,234 ; movdqa %xmm10,%xmm13 - DB 102,69,15,102,239 ; pcmpgtd %xmm15,%xmm13 - DB 102,65,15,114,215,13 ; psrld $0xd,%xmm15 - DB 102,69,15,250,243 ; psubd %xmm11,%xmm14 - DB 102,69,15,254,247 ; paddd %xmm15,%xmm14 + DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 + DB 102,69,15,102,229 ; pcmpgtd %xmm13,%xmm12 + DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 + DB 102,69,15,235,243 ; por %xmm11,%xmm14 + DB 102,69,15,254,245 ; paddd %xmm13,%xmm14 DB 102,65,15,114,246,16 ; pslld $0x10,%xmm14 DB 102,65,15,114,230,16 ; psrad $0x10,%xmm14 - DB 102,69,15,223,238 ; pandn %xmm14,%xmm13 + DB 102,69,15,223,230 ; pandn %xmm14,%xmm12 + DB 102,69,15,107,228 ; packssdw %xmm12,%xmm12 + DB 102,68,15,111,242 ; movdqa %xmm2,%xmm14 + DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 + DB 102,69,15,219,250 ; pand %xmm10,%xmm15 + DB 102,69,15,239,247 ; pxor %xmm15,%xmm14 + DB 102,65,15,114,215,16 ; psrld $0x10,%xmm15 + DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 + DB 102,69,15,102,238 ; pcmpgtd %xmm14,%xmm13 + DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 + DB 102,69,15,235,251 ; por %xmm11,%xmm15 + DB 102,69,15,254,254 ; paddd %xmm14,%xmm15 + DB 102,65,15,114,247,16 ; pslld $0x10,%xmm15 + DB 102,65,15,114,231,16 ; psrad $0x10,%xmm15 + DB 102,69,15,223,239 ; pandn %xmm15,%xmm13 DB 102,69,15,107,237 ; packssdw %xmm13,%xmm13 - DB 102,68,15,219,203 ; pand %xmm3,%xmm9 + DB 102,68,15,219,211 ; pand %xmm3,%xmm10 DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,69,15,239,241 ; pxor %xmm9,%xmm14 - DB 102,65,15,114,209,16 ; psrld $0x10,%xmm9 - DB 102,69,15,102,214 ; pcmpgtd %xmm14,%xmm10 + DB 102,69,15,239,242 ; pxor %xmm10,%xmm14 + DB 102,65,15,114,210,16 ; psrld $0x10,%xmm10 + DB 102,69,15,102,206 ; pcmpgtd %xmm14,%xmm9 DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,250,203 ; psubd %xmm11,%xmm9 - DB 102,69,15,254,206 ; paddd %xmm14,%xmm9 - DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 - DB 102,65,15,114,225,16 ; psrad $0x10,%xmm9 - DB 102,69,15,223,209 ; pandn %xmm9,%xmm10 - DB 102,69,15,107,210 ; packssdw %xmm10,%xmm10 + DB 102,69,15,235,211 ; por %xmm11,%xmm10 + DB 102,69,15,254,214 ; paddd %xmm14,%xmm10 + DB 102,65,15,114,242,16 ; pslld $0x10,%xmm10 + DB 102,65,15,114,226,16 ; psrad $0x10,%xmm10 + DB 102,69,15,223,202 ; pandn %xmm10,%xmm9 + DB 102,69,15,107,201 ; packssdw %xmm9,%xmm9 DB 102,69,15,97,196 ; punpcklwd %xmm12,%xmm8 - DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 + DB 102,69,15,97,233 ; punpcklwd %xmm9,%xmm13 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,69,15,98,205 ; punpckldq %xmm13,%xmm9 DB 243,68,15,127,12,248 ; movdqu %xmm9,(%rax,%rdi,8) @@ -18019,11 +18111,8 @@ _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 184,0,0,128,63 ; mov $0x3f800000,%eax - DB 102,68,15,110,216 ; movd %eax,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,84,217 ; andps %xmm9,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 + DB 68,15,84,13,135,16,0,0 ; andps 0x1087(%rip),%xmm9 # 47d0 <_sk_callback_sse2+0x4bb> + DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,194 ; subps %xmm10,%xmm0 DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 @@ -18042,11 +18131,8 @@ _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 184,0,0,128,63 ; mov $0x3f800000,%eax - DB 102,68,15,110,216 ; movd %eax,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,84,217 ; andps %xmm9,%xmm11 - DB 69,15,92,211 ; subps %xmm11,%xmm10 + DB 68,15,84,13,79,16,0,0 ; andps 0x104f(%rip),%xmm9 # 47e0 <_sk_callback_sse2+0x4cb> + DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,202 ; subps %xmm10,%xmm1 DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 @@ -18058,62 +18144,56 @@ _sk_repeat_y_sse2 LABEL PROC PUBLIC _sk_mirror_x_sse2 _sk_mirror_x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 69,15,40,193 ; movaps %xmm9,%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 65,15,92,192 ; subps %xmm8,%xmm0 - DB 243,69,15,88,201 ; addss %xmm9,%xmm9 + DB 243,68,15,16,0 ; movss (%rax),%xmm8 + DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 243,69,15,88,192 ; addss %xmm8,%xmm8 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,40,208 ; movaps %xmm0,%xmm10 - DB 69,15,94,209 ; divps %xmm9,%xmm10 + DB 69,15,94,208 ; divps %xmm8,%xmm10 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 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 69,15,84,226 ; andps %xmm10,%xmm12 - DB 69,15,87,210 ; xorps %xmm10,%xmm10 - DB 69,15,92,220 ; subps %xmm12,%xmm11 - DB 69,15,89,217 ; mulps %xmm9,%xmm11 + DB 68,15,84,21,5,16,0,0 ; andps 0x1005(%rip),%xmm10 # 47f0 <_sk_callback_sse2+0x4db> + DB 69,15,87,228 ; xorps %xmm12,%xmm12 + DB 69,15,92,218 ; subps %xmm10,%xmm11 + DB 69,15,89,216 ; mulps %xmm8,%xmm11 DB 65,15,92,195 ; subps %xmm11,%xmm0 - DB 65,15,92,192 ; subps %xmm8,%xmm0 - DB 68,15,92,208 ; subps %xmm0,%xmm10 - DB 65,15,84,194 ; andps %xmm10,%xmm0 - DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 - DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 - DB 65,15,93,193 ; minps %xmm9,%xmm0 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 68,15,92,224 ; subps %xmm0,%xmm12 + DB 65,15,84,196 ; andps %xmm12,%xmm0 + DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 + DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 + DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_mirror_y_sse2 _sk_mirror_y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 69,15,40,193 ; movaps %xmm9,%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 65,15,92,200 ; subps %xmm8,%xmm1 - DB 243,69,15,88,201 ; addss %xmm9,%xmm9 + DB 243,68,15,16,0 ; movss (%rax),%xmm8 + DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 65,15,92,201 ; subps %xmm9,%xmm1 + DB 243,69,15,88,192 ; addss %xmm8,%xmm8 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,40,209 ; movaps %xmm1,%xmm10 - DB 69,15,94,209 ; divps %xmm9,%xmm10 + DB 69,15,94,208 ; divps %xmm8,%xmm10 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 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 69,15,84,226 ; andps %xmm10,%xmm12 - DB 69,15,87,210 ; xorps %xmm10,%xmm10 - DB 69,15,92,220 ; subps %xmm12,%xmm11 - DB 69,15,89,217 ; mulps %xmm9,%xmm11 + DB 68,15,84,21,171,15,0,0 ; andps 0xfab(%rip),%xmm10 # 4800 <_sk_callback_sse2+0x4eb> + DB 69,15,87,228 ; xorps %xmm12,%xmm12 + DB 69,15,92,218 ; subps %xmm10,%xmm11 + DB 69,15,89,216 ; mulps %xmm8,%xmm11 DB 65,15,92,203 ; subps %xmm11,%xmm1 - DB 65,15,92,200 ; subps %xmm8,%xmm1 - DB 68,15,92,209 ; subps %xmm1,%xmm10 - DB 65,15,84,202 ; andps %xmm10,%xmm1 - DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 - DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 - DB 65,15,93,201 ; minps %xmm9,%xmm1 + DB 65,15,92,201 ; subps %xmm9,%xmm1 + DB 68,15,92,225 ; subps %xmm1,%xmm12 + DB 65,15,84,204 ; andps %xmm12,%xmm1 + DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 + DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 + DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -18354,7 +18434,7 @@ _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 3f2a <_sk_linear_gradient_sse2+0x149> + DB 15,132,15,1,0,0 ; je 3d69 <_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 @@ -18415,8 +18495,8 @@ _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 3e30 <_sk_linear_gradient_sse2+0x4f> - DB 235,13 ; jmp 3f37 <_sk_linear_gradient_sse2+0x156> + DB 15,133,8,255,255,255 ; jne 3c6f <_sk_linear_gradient_sse2+0x4f> + DB 235,13 ; jmp 3d76 <_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 @@ -18476,9 +18556,7 @@ _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 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 68,15,40,37,210,9,0,0 ; movaps 0x9d2(%rip),%xmm12 # 4810 <_sk_callback_sse2+0x4fb> DB 69,15,84,220 ; andps %xmm12,%xmm11 DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 69,15,92,202 ; subps %xmm10,%xmm9 @@ -18872,7 +18950,468 @@ ALIGN 16 DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) DB 63 ; (bad) DB 0,0 ; add %al,(%rax) - DB 128 ; .byte 0x80 + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 43f4 <.literal16+0x34> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 43f8 <.literal16+0x38> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 43fc <.literal16+0x3c> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4400 <.literal16+0x40> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 4485 <.literal16+0xc5> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 43e9 <.literal16+0x29> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a3842a <_sk_callback_sse2+0xffffffffe9a34115> + DB 220,63 ; fdivrl (%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 128,63,81 ; cmpb $0x51,(%rdi) + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,52,0 ; add %dh,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 44c4 <.literal16+0x104> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 44c8 <.literal16+0x108> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 44cc <.literal16+0x10c> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 44d0 <.literal16+0x110> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 4555 <.literal16+0x195> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 44b9 <.literal16+0xf9> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a384fa <_sk_callback_sse2+0xffffffffe9a341e5> + DB 220,63 ; fdivrl (%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 128,63,81 ; cmpb $0x51,(%rdi) + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,52,0 ; add %dh,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4594 <.literal16+0x1d4> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4598 <.literal16+0x1d8> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 459c <.literal16+0x1dc> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 45a0 <.literal16+0x1e0> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 4625 <.literal16+0x265> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 4589 <.literal16+0x1c9> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a385ca <_sk_callback_sse2+0xffffffffe9a342b5> + DB 220,63 ; fdivrl (%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 128,63,81 ; cmpb $0x51,(%rdi) + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,52,0 ; add %dh,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,0 ; xor $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 52,255 ; xor $0xff,%al + DB 255 ; (bad) + DB 127,0 ; jg 4664 <.literal16+0x2a4> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4668 <.literal16+0x2a8> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 466c <.literal16+0x2ac> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 4670 <.literal16+0x2b0> + DB 0,0 ; add %al,(%rax) + 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 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 119,115 ; ja 46f5 <.literal16+0x335> + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,119,115 ; retq $0x7377 + DB 248 ; clc + DB 194,117,191 ; retq $0xbf75 + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 117,191 ; jne 4659 <.literal16+0x299> + DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi + DB 63 ; (bad) + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + DB 249 ; stc + DB 68,180,62 ; rex.R mov $0x3e,%spl + 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 ffffffffe9a3869a <_sk_callback_sse2+0xffffffffe9a34385> + DB 220,63 ; fdivrl (%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 128,63,81 ; cmpb $0x51,(%rdi) + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,81 ; rex.X push %rcx + DB 140,242 ; mov %?,%edx + DB 66,141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%r15,4),%edi + DB 63 ; (bad) + DB 141,188,190,63,141,188,190 ; lea -0x414372c1(%rsi,%rdi,4),%edi + DB 63 ; (bad) + DB 248 ; clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,248 ; rex clc + DB 245 ; cmc + DB 154 ; (bad) + DB 64,254 ; rex (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,254 ; rex.B (bad) + DB 210,221 ; rcr %cl,%ch + DB 65,0,0 ; add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,0 ; rex.WXB add %al,(%r8) + DB 0,75,0 ; add %cl,0x0(%rbx) + DB 0,0 ; add %al,(%rax) + DB 75,0,128,0,0,0,128 ; rex.WXB add %al,-0x80000000(%r8) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 128,0,4 ; addb $0x4,(%rax) + DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 128,0,128 ; addb $0x80,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,0 ; add %al,0x0(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,56 ; add %bh,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 128,0,4 ; addb $0x4,(%rax) + DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 128,0,0 ; addb $0x0,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) + DB 0,0 ; add %al,(%rax) + DB 128,56,0 ; cmpb $0x0,(%rax) + DB 0,128,56,0,0,128 ; add %al,-0x7fffffc8(%rax) + DB 56,0 ; cmp %al,(%rax) + DB 0,128,56,0,64,254 ; add %al,-0x1bfffc8(%rax) + DB 255,0 ; incl (%rax) + DB 64,254 ; rex (bad) + DB 255,0 ; incl (%rax) + DB 64,254 ; rex (bad) + DB 255,0 ; incl (%rax) + DB 64,254 ; rex (bad) + DB 255,0 ; incl (%rax) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + 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 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + DB 0,0 ; add %al,(%rax) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + DB 0,0 ; add %al,(%rax) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + DB 0,0 ; add %al,(%rax) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) + DB 63 ; (bad) + DB 0,0 ; add %al,(%rax) + DB 128,63,0 ; cmpb $0x0,(%rdi) + DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) DB 63 ; (bad) ENDIF END diff --git a/src/jumper/SkJumper_vectors.h b/src/jumper/SkJumper_vectors.h index 57f36be0dd..504aca0a82 100644 --- a/src/jumper/SkJumper_vectors.h +++ b/src/jumper/SkJumper_vectors.h @@ -162,7 +162,7 @@ SI F floor_(F v) { F roundtrip = vcvt_f32_s32(vcvt_s32_f32(v)); - return roundtrip - if_then_else(roundtrip > v, 1.0_f, 0); + return roundtrip - if_then_else(roundtrip > v, 1, 0); } template <typename T> @@ -467,7 +467,7 @@ return _mm_floor_ps(v); #else F roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v)); - return roundtrip - if_then_else(roundtrip > v, 1.0_f, 0); + return roundtrip - if_then_else(roundtrip > v, 1, 0); #endif } @@ -576,21 +576,21 @@ SI F fract(F v) { return v - floor_(v); } // See http://www.machinedlearnings.com/2011/06/fast-approximate-logarithm-exponential.html. SI F approx_log2(F x) { // e - 127 is a fair approximation of log2(x) in its own right... - F e = cast(bit_cast<U32>(x)) * C(1.0f / (1<<23)); + F e = cast(bit_cast<U32>(x)) * (1.0f / (1<<23)); // ... but using the mantissa to refine its error is _much_ better. - F m = bit_cast<F>((bit_cast<U32>(x) & 0x007fffff_i) | 0x3f000000_i); + F m = bit_cast<F>((bit_cast<U32>(x) & 0x007fffff) | 0x3f000000); return e - - 124.225514990_f - - 1.498030302_f * m - - 1.725879990_f / (0.3520887068_f + m); + - 124.225514990f + - 1.498030302f * m + - 1.725879990f / (0.3520887068f + m); } SI F approx_pow2(F x) { F f = fract(x); - return bit_cast<F>(round(C(1.0f * (1<<23)), - x + 121.274057500_f - - 1.490129070_f * f - + 27.728023300_f / (4.84252568_f - f))); + return bit_cast<F>(round(1.0f * (1<<23), + x + 121.274057500f + - 1.490129070f * f + + 27.728023300f / (4.84252568f - f))); } SI F approx_powf(F x, F y) { @@ -611,13 +611,13 @@ SI F from_half(U16 h) { #else // Remember, a half is 1-5-10 (sign-exponent-mantissa) with 15 exponent bias. U32 sem = expand(h), - s = sem & 0x8000_i, + s = sem & 0x8000, em = sem ^ s; // Convert to 1-8-23 float with 127 bias, flushing denorm halfs (including zero) to zero. - auto denorm = (I32)em < 0x0400_i; // I32 comparison is often quicker, and always safe here. + auto denorm = (I32)em < 0x0400; // I32 comparison is often quicker, and always safe here. return if_then_else(denorm, F(0) - , bit_cast<F>( (s<<16) + (em<<13) + C((127-15)<<23) )); + , bit_cast<F>( (s<<16) + (em<<13) + ((127-15)<<23) )); #endif } @@ -636,13 +636,13 @@ SI U16 to_half(F f) { #else // Remember, a float is 1-8-23 (sign-exponent-mantissa) with 127 exponent bias. U32 sem = bit_cast<U32>(f), - s = sem & 0x80000000_i, + s = sem & 0x80000000, em = sem ^ s; // Convert to 1-5-10 half with 15 bias, flushing denorm halfs (including zero) to zero. - auto denorm = (I32)em < 0x38800000_i; // I32 comparison is often quicker, and always safe here. + auto denorm = (I32)em < 0x38800000; // I32 comparison is often quicker, and always safe here. return pack(if_then_else(denorm, U32(0) - , (s>>16) + (em>>13) - C((127-15)<<10))); + , (s>>16) + (em>>13) - ((127-15)<<10))); #endif } |