aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/jumper
diff options
context:
space:
mode:
authorGravatar Mike Klein <mtklein@chromium.org>2017-06-13 13:56:09 -0400
committerGravatar Skia Commit-Bot <skia-commit-bot@chromium.org>2017-06-13 18:58:43 +0000
commitf046594d82b84bf96c18f6fd0cb14c16bd3b8708 (patch)
tree85de9973282ca723235ba6fa2d459348282138b5 /src/jumper
parentc0f8e426c59eec6c720b8e1329dcb966cf1b6800 (diff)
Remove AVX+ special case for load<U8>().
We haven't really needed this since we got constants working. Change-Id: Ie9de8df861959696ed44fc2a64e259cca786bfe5 Reviewed-on: https://skia-review.googlesource.com/19670 Reviewed-by: Herb Derby <herb@google.com> Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/jumper')
-rw-r--r--src/jumper/SkJumper_generated.S2494
-rw-r--r--src/jumper/SkJumper_generated_win.S2477
-rw-r--r--src/jumper/SkJumper_stages.cpp19
3 files changed, 2802 insertions, 2188 deletions
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S
index e568ed9db9..34d01ce9af 100644
--- a/src/jumper/SkJumper_generated.S
+++ b/src/jumper/SkJumper_generated.S
@@ -9794,7 +9794,7 @@ _sk_seed_shader_hsw:
.byte 197,249,110,194 // vmovd %edx,%xmm0
.byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,194,72,0,0 // vbroadcastss 0x48c2(%rip),%ymm1 # 499c <_sk_callback_hsw+0x144>
+ .byte 196,226,125,24,13,158,74,0,0 // vbroadcastss 0x4a9e(%rip),%ymm1 # 4b78 <_sk_callback_hsw+0x144>
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
.byte 197,252,88,7 // vaddps (%rdi),%ymm0,%ymm0
.byte 197,249,110,209 // vmovd %ecx,%xmm2
@@ -9802,7 +9802,7 @@ _sk_seed_shader_hsw:
.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 196,226,125,24,21,162,72,0,0 // vbroadcastss 0x48a2(%rip),%ymm2 # 49a0 <_sk_callback_hsw+0x148>
+ .byte 196,226,125,24,21,126,74,0,0 // vbroadcastss 0x4a7e(%rip),%ymm2 # 4b7c <_sk_callback_hsw+0x148>
.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
@@ -9821,13 +9821,13 @@ _sk_dither_hsw:
.byte 197,121,110,201 // vmovd %ecx,%xmm9
.byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
.byte 196,65,53,239,200 // vpxor %ymm8,%ymm9,%ymm9
- .byte 196,98,125,88,21,105,72,0,0 // vpbroadcastd 0x4869(%rip),%ymm10 # 49a4 <_sk_callback_hsw+0x14c>
+ .byte 196,98,125,88,21,69,74,0,0 // vpbroadcastd 0x4a45(%rip),%ymm10 # 4b80 <_sk_callback_hsw+0x14c>
.byte 196,65,53,219,218 // vpand %ymm10,%ymm9,%ymm11
.byte 196,193,37,114,243,5 // vpslld $0x5,%ymm11,%ymm11
.byte 196,65,61,219,210 // vpand %ymm10,%ymm8,%ymm10
.byte 196,193,45,114,242,4 // vpslld $0x4,%ymm10,%ymm10
- .byte 196,98,125,88,37,78,72,0,0 // vpbroadcastd 0x484e(%rip),%ymm12 # 49a8 <_sk_callback_hsw+0x150>
- .byte 196,98,125,88,45,73,72,0,0 // vpbroadcastd 0x4849(%rip),%ymm13 # 49ac <_sk_callback_hsw+0x154>
+ .byte 196,98,125,88,37,42,74,0,0 // vpbroadcastd 0x4a2a(%rip),%ymm12 # 4b84 <_sk_callback_hsw+0x150>
+ .byte 196,98,125,88,45,37,74,0,0 // vpbroadcastd 0x4a25(%rip),%ymm13 # 4b88 <_sk_callback_hsw+0x154>
.byte 196,65,53,219,245 // vpand %ymm13,%ymm9,%ymm14
.byte 196,193,13,114,246,2 // vpslld $0x2,%ymm14,%ymm14
.byte 196,65,61,219,237 // vpand %ymm13,%ymm8,%ymm13
@@ -9842,8 +9842,8 @@ _sk_dither_hsw:
.byte 196,65,61,235,194 // vpor %ymm10,%ymm8,%ymm8
.byte 196,65,61,235,193 // vpor %ymm9,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,251,71,0,0 // vbroadcastss 0x47fb(%rip),%ymm9 # 49b0 <_sk_callback_hsw+0x158>
- .byte 196,98,125,24,21,246,71,0,0 // vbroadcastss 0x47f6(%rip),%ymm10 # 49b4 <_sk_callback_hsw+0x15c>
+ .byte 196,98,125,24,13,215,73,0,0 // vbroadcastss 0x49d7(%rip),%ymm9 # 4b8c <_sk_callback_hsw+0x158>
+ .byte 196,98,125,24,21,210,73,0,0 // vbroadcastss 0x49d2(%rip),%ymm10 # 4b90 <_sk_callback_hsw+0x15c>
.byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10
.byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
@@ -9912,7 +9912,7 @@ HIDDEN _sk_srcatop_hsw
FUNCTION(_sk_srcatop_hsw)
_sk_srcatop_hsw:
.byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
- .byte 196,98,125,24,5,78,71,0,0 // vbroadcastss 0x474e(%rip),%ymm8 # 49b8 <_sk_callback_hsw+0x160>
+ .byte 196,98,125,24,5,42,73,0,0 // vbroadcastss 0x492a(%rip),%ymm8 # 4b94 <_sk_callback_hsw+0x160>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 196,226,61,184,196 // vfmadd231ps %ymm4,%ymm8,%ymm0
.byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
@@ -9928,7 +9928,7 @@ HIDDEN _sk_dstatop_hsw
.globl _sk_dstatop_hsw
FUNCTION(_sk_dstatop_hsw)
_sk_dstatop_hsw:
- .byte 196,98,125,24,5,33,71,0,0 // vbroadcastss 0x4721(%rip),%ymm8 # 49bc <_sk_callback_hsw+0x164>
+ .byte 196,98,125,24,5,253,72,0,0 // vbroadcastss 0x48fd(%rip),%ymm8 # 4b98 <_sk_callback_hsw+0x164>
.byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
.byte 196,226,101,184,196 // vfmadd231ps %ymm4,%ymm3,%ymm0
@@ -9967,7 +9967,7 @@ HIDDEN _sk_srcout_hsw
.globl _sk_srcout_hsw
FUNCTION(_sk_srcout_hsw)
_sk_srcout_hsw:
- .byte 196,98,125,24,5,200,70,0,0 // vbroadcastss 0x46c8(%rip),%ymm8 # 49c0 <_sk_callback_hsw+0x168>
+ .byte 196,98,125,24,5,164,72,0,0 // vbroadcastss 0x48a4(%rip),%ymm8 # 4b9c <_sk_callback_hsw+0x168>
.byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
.byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
@@ -9980,7 +9980,7 @@ HIDDEN _sk_dstout_hsw
.globl _sk_dstout_hsw
FUNCTION(_sk_dstout_hsw)
_sk_dstout_hsw:
- .byte 196,226,125,24,5,171,70,0,0 // vbroadcastss 0x46ab(%rip),%ymm0 # 49c4 <_sk_callback_hsw+0x16c>
+ .byte 196,226,125,24,5,135,72,0,0 // vbroadcastss 0x4887(%rip),%ymm0 # 4ba0 <_sk_callback_hsw+0x16c>
.byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3
.byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0
.byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1
@@ -9993,7 +9993,7 @@ HIDDEN _sk_srcover_hsw
.globl _sk_srcover_hsw
FUNCTION(_sk_srcover_hsw)
_sk_srcover_hsw:
- .byte 196,98,125,24,5,142,70,0,0 // vbroadcastss 0x468e(%rip),%ymm8 # 49c8 <_sk_callback_hsw+0x170>
+ .byte 196,98,125,24,5,106,72,0,0 // vbroadcastss 0x486a(%rip),%ymm8 # 4ba4 <_sk_callback_hsw+0x170>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0
.byte 196,194,85,184,200 // vfmadd231ps %ymm8,%ymm5,%ymm1
@@ -10006,7 +10006,7 @@ HIDDEN _sk_dstover_hsw
.globl _sk_dstover_hsw
FUNCTION(_sk_dstover_hsw)
_sk_dstover_hsw:
- .byte 196,98,125,24,5,109,70,0,0 // vbroadcastss 0x466d(%rip),%ymm8 # 49cc <_sk_callback_hsw+0x174>
+ .byte 196,98,125,24,5,73,72,0,0 // vbroadcastss 0x4849(%rip),%ymm8 # 4ba8 <_sk_callback_hsw+0x174>
.byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
.byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0
.byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1
@@ -10030,7 +10030,7 @@ HIDDEN _sk_multiply_hsw
.globl _sk_multiply_hsw
FUNCTION(_sk_multiply_hsw)
_sk_multiply_hsw:
- .byte 196,98,125,24,5,56,70,0,0 // vbroadcastss 0x4638(%rip),%ymm8 # 49d0 <_sk_callback_hsw+0x178>
+ .byte 196,98,125,24,5,20,72,0,0 // vbroadcastss 0x4814(%rip),%ymm8 # 4bac <_sk_callback_hsw+0x178>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -10078,7 +10078,7 @@ HIDDEN _sk_xor__hsw
.globl _sk_xor__hsw
FUNCTION(_sk_xor__hsw)
_sk_xor__hsw:
- .byte 196,98,125,24,5,179,69,0,0 // vbroadcastss 0x45b3(%rip),%ymm8 # 49d4 <_sk_callback_hsw+0x17c>
+ .byte 196,98,125,24,5,143,71,0,0 // vbroadcastss 0x478f(%rip),%ymm8 # 4bb0 <_sk_callback_hsw+0x17c>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -10112,7 +10112,7 @@ _sk_darken_hsw:
.byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
.byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,59,69,0,0 // vbroadcastss 0x453b(%rip),%ymm8 # 49d8 <_sk_callback_hsw+0x180>
+ .byte 196,98,125,24,5,23,71,0,0 // vbroadcastss 0x4717(%rip),%ymm8 # 4bb4 <_sk_callback_hsw+0x180>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -10137,7 +10137,7 @@ _sk_lighten_hsw:
.byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
.byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,234,68,0,0 // vbroadcastss 0x44ea(%rip),%ymm8 # 49dc <_sk_callback_hsw+0x184>
+ .byte 196,98,125,24,5,198,70,0,0 // vbroadcastss 0x46c6(%rip),%ymm8 # 4bb8 <_sk_callback_hsw+0x184>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -10165,7 +10165,7 @@ _sk_difference_hsw:
.byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
.byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,141,68,0,0 // vbroadcastss 0x448d(%rip),%ymm8 # 49e0 <_sk_callback_hsw+0x188>
+ .byte 196,98,125,24,5,105,70,0,0 // vbroadcastss 0x4669(%rip),%ymm8 # 4bbc <_sk_callback_hsw+0x188>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -10187,7 +10187,7 @@ _sk_exclusion_hsw:
.byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
.byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,75,68,0,0 // vbroadcastss 0x444b(%rip),%ymm8 # 49e4 <_sk_callback_hsw+0x18c>
+ .byte 196,98,125,24,5,39,70,0,0 // vbroadcastss 0x4627(%rip),%ymm8 # 4bc0 <_sk_callback_hsw+0x18c>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -10197,7 +10197,7 @@ HIDDEN _sk_colorburn_hsw
.globl _sk_colorburn_hsw
FUNCTION(_sk_colorburn_hsw)
_sk_colorburn_hsw:
- .byte 196,98,125,24,5,57,68,0,0 // vbroadcastss 0x4439(%rip),%ymm8 # 49e8 <_sk_callback_hsw+0x190>
+ .byte 196,98,125,24,5,21,70,0,0 // vbroadcastss 0x4615(%rip),%ymm8 # 4bc4 <_sk_callback_hsw+0x190>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11
.byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
@@ -10255,7 +10255,7 @@ HIDDEN _sk_colordodge_hsw
FUNCTION(_sk_colordodge_hsw)
_sk_colordodge_hsw:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
- .byte 196,98,125,24,13,68,67,0,0 // vbroadcastss 0x4344(%rip),%ymm9 # 49ec <_sk_callback_hsw+0x194>
+ .byte 196,98,125,24,13,32,69,0,0 // vbroadcastss 0x4520(%rip),%ymm9 # 4bc8 <_sk_callback_hsw+0x194>
.byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10
.byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11
.byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9
@@ -10308,7 +10308,7 @@ HIDDEN _sk_hardlight_hsw
.globl _sk_hardlight_hsw
FUNCTION(_sk_hardlight_hsw)
_sk_hardlight_hsw:
- .byte 196,98,125,24,5,101,66,0,0 // vbroadcastss 0x4265(%rip),%ymm8 # 49f0 <_sk_callback_hsw+0x198>
+ .byte 196,98,125,24,5,65,68,0,0 // vbroadcastss 0x4441(%rip),%ymm8 # 4bcc <_sk_callback_hsw+0x198>
.byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
.byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -10359,7 +10359,7 @@ HIDDEN _sk_overlay_hsw
.globl _sk_overlay_hsw
FUNCTION(_sk_overlay_hsw)
_sk_overlay_hsw:
- .byte 196,98,125,24,5,157,65,0,0 // vbroadcastss 0x419d(%rip),%ymm8 # 49f4 <_sk_callback_hsw+0x19c>
+ .byte 196,98,125,24,5,121,67,0,0 // vbroadcastss 0x4379(%rip),%ymm8 # 4bd0 <_sk_callback_hsw+0x19c>
.byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
.byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -10420,10 +10420,10 @@ _sk_softlight_hsw:
.byte 196,65,20,88,197 // vaddps %ymm13,%ymm13,%ymm8
.byte 196,65,60,88,192 // vaddps %ymm8,%ymm8,%ymm8
.byte 196,66,61,168,192 // vfmadd213ps %ymm8,%ymm8,%ymm8
- .byte 196,98,125,24,29,168,64,0,0 // vbroadcastss 0x40a8(%rip),%ymm11 # 49fc <_sk_callback_hsw+0x1a4>
+ .byte 196,98,125,24,29,132,66,0,0 // vbroadcastss 0x4284(%rip),%ymm11 # 4bd8 <_sk_callback_hsw+0x1a4>
.byte 196,65,20,88,227 // vaddps %ymm11,%ymm13,%ymm12
.byte 196,65,28,89,192 // vmulps %ymm8,%ymm12,%ymm8
- .byte 196,98,125,24,37,153,64,0,0 // vbroadcastss 0x4099(%rip),%ymm12 # 4a00 <_sk_callback_hsw+0x1a8>
+ .byte 196,98,125,24,37,117,66,0,0 // vbroadcastss 0x4275(%rip),%ymm12 # 4bdc <_sk_callback_hsw+0x1a8>
.byte 196,66,21,184,196 // vfmadd231ps %ymm12,%ymm13,%ymm8
.byte 196,65,124,82,245 // vrsqrtps %ymm13,%ymm14
.byte 196,65,124,83,246 // vrcpps %ymm14,%ymm14
@@ -10433,7 +10433,7 @@ _sk_softlight_hsw:
.byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15
.byte 196,67,13,74,240,240 // vblendvps %ymm15,%ymm8,%ymm14,%ymm14
.byte 197,116,88,249 // vaddps %ymm1,%ymm1,%ymm15
- .byte 196,98,125,24,5,92,64,0,0 // vbroadcastss 0x405c(%rip),%ymm8 # 49f8 <_sk_callback_hsw+0x1a0>
+ .byte 196,98,125,24,5,56,66,0,0 // vbroadcastss 0x4238(%rip),%ymm8 # 4bd4 <_sk_callback_hsw+0x1a0>
.byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13
.byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0
.byte 196,98,125,168,235 // vfmadd213ps %ymm3,%ymm0,%ymm13
@@ -10546,11 +10546,11 @@ _sk_hue_hsw:
.byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10
.byte 196,65,44,94,214 // vdivps %ymm14,%ymm10,%ymm10
.byte 196,67,45,74,224,240 // vblendvps %ymm15,%ymm8,%ymm10,%ymm12
- .byte 196,98,125,24,53,96,62,0,0 // vbroadcastss 0x3e60(%rip),%ymm14 # 4a04 <_sk_callback_hsw+0x1ac>
- .byte 196,98,125,24,61,91,62,0,0 // vbroadcastss 0x3e5b(%rip),%ymm15 # 4a08 <_sk_callback_hsw+0x1b0>
+ .byte 196,98,125,24,53,60,64,0,0 // vbroadcastss 0x403c(%rip),%ymm14 # 4be0 <_sk_callback_hsw+0x1ac>
+ .byte 196,98,125,24,61,55,64,0,0 // vbroadcastss 0x4037(%rip),%ymm15 # 4be4 <_sk_callback_hsw+0x1b0>
.byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13
.byte 196,66,93,184,238 // vfmadd231ps %ymm14,%ymm4,%ymm13
- .byte 196,226,125,24,5,76,62,0,0 // vbroadcastss 0x3e4c(%rip),%ymm0 # 4a0c <_sk_callback_hsw+0x1b4>
+ .byte 196,226,125,24,5,40,64,0,0 // vbroadcastss 0x4028(%rip),%ymm0 # 4be8 <_sk_callback_hsw+0x1b4>
.byte 196,98,77,184,232 // vfmadd231ps %ymm0,%ymm6,%ymm13
.byte 196,65,116,89,215 // vmulps %ymm15,%ymm1,%ymm10
.byte 196,66,53,184,214 // vfmadd231ps %ymm14,%ymm9,%ymm10
@@ -10605,7 +10605,7 @@ _sk_hue_hsw:
.byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0
.byte 196,65,36,95,200 // vmaxps %ymm8,%ymm11,%ymm9
.byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8
- .byte 196,226,125,24,13,57,61,0,0 // vbroadcastss 0x3d39(%rip),%ymm1 # 4a10 <_sk_callback_hsw+0x1b8>
+ .byte 196,226,125,24,13,21,63,0,0 // vbroadcastss 0x3f15(%rip),%ymm1 # 4bec <_sk_callback_hsw+0x1b8>
.byte 197,116,92,215 // vsubps %ymm7,%ymm1,%ymm10
.byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
.byte 197,116,92,219 // vsubps %ymm3,%ymm1,%ymm11
@@ -10659,11 +10659,11 @@ _sk_saturation_hsw:
.byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10
.byte 196,65,44,94,214 // vdivps %ymm14,%ymm10,%ymm10
.byte 196,67,45,74,224,240 // vblendvps %ymm15,%ymm8,%ymm10,%ymm12
- .byte 196,98,125,24,53,80,60,0,0 // vbroadcastss 0x3c50(%rip),%ymm14 # 4a14 <_sk_callback_hsw+0x1bc>
- .byte 196,98,125,24,61,75,60,0,0 // vbroadcastss 0x3c4b(%rip),%ymm15 # 4a18 <_sk_callback_hsw+0x1c0>
+ .byte 196,98,125,24,53,44,62,0,0 // vbroadcastss 0x3e2c(%rip),%ymm14 # 4bf0 <_sk_callback_hsw+0x1bc>
+ .byte 196,98,125,24,61,39,62,0,0 // vbroadcastss 0x3e27(%rip),%ymm15 # 4bf4 <_sk_callback_hsw+0x1c0>
.byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13
.byte 196,66,93,184,238 // vfmadd231ps %ymm14,%ymm4,%ymm13
- .byte 196,226,125,24,5,60,60,0,0 // vbroadcastss 0x3c3c(%rip),%ymm0 # 4a1c <_sk_callback_hsw+0x1c4>
+ .byte 196,226,125,24,5,24,62,0,0 // vbroadcastss 0x3e18(%rip),%ymm0 # 4bf8 <_sk_callback_hsw+0x1c4>
.byte 196,98,77,184,232 // vfmadd231ps %ymm0,%ymm6,%ymm13
.byte 196,65,116,89,215 // vmulps %ymm15,%ymm1,%ymm10
.byte 196,66,53,184,214 // vfmadd231ps %ymm14,%ymm9,%ymm10
@@ -10718,7 +10718,7 @@ _sk_saturation_hsw:
.byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0
.byte 196,65,36,95,200 // vmaxps %ymm8,%ymm11,%ymm9
.byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8
- .byte 196,226,125,24,13,41,59,0,0 // vbroadcastss 0x3b29(%rip),%ymm1 # 4a20 <_sk_callback_hsw+0x1c8>
+ .byte 196,226,125,24,13,5,61,0,0 // vbroadcastss 0x3d05(%rip),%ymm1 # 4bfc <_sk_callback_hsw+0x1c8>
.byte 197,116,92,215 // vsubps %ymm7,%ymm1,%ymm10
.byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
.byte 197,116,92,219 // vsubps %ymm3,%ymm1,%ymm11
@@ -10746,11 +10746,11 @@ _sk_color_hsw:
.byte 197,108,89,199 // vmulps %ymm7,%ymm2,%ymm8
.byte 197,116,89,215 // vmulps %ymm7,%ymm1,%ymm10
.byte 197,52,89,223 // vmulps %ymm7,%ymm9,%ymm11
- .byte 196,98,125,24,45,194,58,0,0 // vbroadcastss 0x3ac2(%rip),%ymm13 # 4a24 <_sk_callback_hsw+0x1cc>
- .byte 196,98,125,24,53,189,58,0,0 // vbroadcastss 0x3abd(%rip),%ymm14 # 4a28 <_sk_callback_hsw+0x1d0>
+ .byte 196,98,125,24,45,158,60,0,0 // vbroadcastss 0x3c9e(%rip),%ymm13 # 4c00 <_sk_callback_hsw+0x1cc>
+ .byte 196,98,125,24,53,153,60,0,0 // vbroadcastss 0x3c99(%rip),%ymm14 # 4c04 <_sk_callback_hsw+0x1d0>
.byte 196,65,84,89,230 // vmulps %ymm14,%ymm5,%ymm12
.byte 196,66,93,184,229 // vfmadd231ps %ymm13,%ymm4,%ymm12
- .byte 196,98,125,24,61,174,58,0,0 // vbroadcastss 0x3aae(%rip),%ymm15 # 4a2c <_sk_callback_hsw+0x1d4>
+ .byte 196,98,125,24,61,138,60,0,0 // vbroadcastss 0x3c8a(%rip),%ymm15 # 4c08 <_sk_callback_hsw+0x1d4>
.byte 196,66,77,184,231 // vfmadd231ps %ymm15,%ymm6,%ymm12
.byte 196,65,44,89,206 // vmulps %ymm14,%ymm10,%ymm9
.byte 196,66,61,184,205 // vfmadd231ps %ymm13,%ymm8,%ymm9
@@ -10806,7 +10806,7 @@ _sk_color_hsw:
.byte 196,193,116,95,206 // vmaxps %ymm14,%ymm1,%ymm1
.byte 196,65,44,95,198 // vmaxps %ymm14,%ymm10,%ymm8
.byte 196,65,124,95,206 // vmaxps %ymm14,%ymm0,%ymm9
- .byte 196,226,125,24,5,144,57,0,0 // vbroadcastss 0x3990(%rip),%ymm0 # 4a30 <_sk_callback_hsw+0x1d8>
+ .byte 196,226,125,24,5,108,59,0,0 // vbroadcastss 0x3b6c(%rip),%ymm0 # 4c0c <_sk_callback_hsw+0x1d8>
.byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10
.byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
.byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11
@@ -10834,11 +10834,11 @@ _sk_luminosity_hsw:
.byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8
.byte 197,100,89,213 // vmulps %ymm5,%ymm3,%ymm10
.byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11
- .byte 196,98,125,24,45,41,57,0,0 // vbroadcastss 0x3929(%rip),%ymm13 # 4a34 <_sk_callback_hsw+0x1dc>
- .byte 196,98,125,24,53,36,57,0,0 // vbroadcastss 0x3924(%rip),%ymm14 # 4a38 <_sk_callback_hsw+0x1e0>
+ .byte 196,98,125,24,45,5,59,0,0 // vbroadcastss 0x3b05(%rip),%ymm13 # 4c10 <_sk_callback_hsw+0x1dc>
+ .byte 196,98,125,24,53,0,59,0,0 // vbroadcastss 0x3b00(%rip),%ymm14 # 4c14 <_sk_callback_hsw+0x1e0>
.byte 196,65,116,89,230 // vmulps %ymm14,%ymm1,%ymm12
.byte 196,66,109,184,229 // vfmadd231ps %ymm13,%ymm2,%ymm12
- .byte 196,98,125,24,61,21,57,0,0 // vbroadcastss 0x3915(%rip),%ymm15 # 4a3c <_sk_callback_hsw+0x1e4>
+ .byte 196,98,125,24,61,241,58,0,0 // vbroadcastss 0x3af1(%rip),%ymm15 # 4c18 <_sk_callback_hsw+0x1e4>
.byte 196,66,53,184,231 // vfmadd231ps %ymm15,%ymm9,%ymm12
.byte 196,65,44,89,206 // vmulps %ymm14,%ymm10,%ymm9
.byte 196,66,61,184,205 // vfmadd231ps %ymm13,%ymm8,%ymm9
@@ -10894,7 +10894,7 @@ _sk_luminosity_hsw:
.byte 196,193,116,95,206 // vmaxps %ymm14,%ymm1,%ymm1
.byte 196,65,44,95,198 // vmaxps %ymm14,%ymm10,%ymm8
.byte 196,65,124,95,206 // vmaxps %ymm14,%ymm0,%ymm9
- .byte 196,226,125,24,5,247,55,0,0 // vbroadcastss 0x37f7(%rip),%ymm0 # 4a40 <_sk_callback_hsw+0x1e8>
+ .byte 196,226,125,24,5,211,57,0,0 // vbroadcastss 0x39d3(%rip),%ymm0 # 4c1c <_sk_callback_hsw+0x1e8>
.byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10
.byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
.byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11
@@ -10922,17 +10922,17 @@ _sk_srcover_rgba_8888_hsw:
.byte 77,133,192 // test %r8,%r8
.byte 15,133,180,0,0,0 // jne 1358 <_sk_srcover_rgba_8888_hsw+0xcd>
.byte 196,193,124,16,58 // vmovups (%r10),%ymm7
- .byte 197,196,84,37,15,59,0,0 // vandps 0x3b0f(%rip),%ymm7,%ymm4 # 4dc0 <_sk_callback_hsw+0x568>
+ .byte 197,196,84,37,207,60,0,0 // vandps 0x3ccf(%rip),%ymm7,%ymm4 # 4f80 <_sk_callback_hsw+0x54c>
.byte 197,252,91,228 // vcvtdq2ps %ymm4,%ymm4
- .byte 196,226,69,0,45,34,59,0,0 // vpshufb 0x3b22(%rip),%ymm7,%ymm5 # 4de0 <_sk_callback_hsw+0x588>
+ .byte 196,226,69,0,45,226,60,0,0 // vpshufb 0x3ce2(%rip),%ymm7,%ymm5 # 4fa0 <_sk_callback_hsw+0x56c>
.byte 197,252,91,237 // vcvtdq2ps %ymm5,%ymm5
- .byte 196,226,69,0,53,53,59,0,0 // vpshufb 0x3b35(%rip),%ymm7,%ymm6 # 4e00 <_sk_callback_hsw+0x5a8>
+ .byte 196,226,69,0,53,245,60,0,0 // vpshufb 0x3cf5(%rip),%ymm7,%ymm6 # 4fc0 <_sk_callback_hsw+0x58c>
.byte 197,252,91,246 // vcvtdq2ps %ymm6,%ymm6
.byte 197,197,114,215,24 // vpsrld $0x18,%ymm7,%ymm7
.byte 197,252,91,255 // vcvtdq2ps %ymm7,%ymm7
- .byte 196,98,125,24,5,99,55,0,0 // vbroadcastss 0x3763(%rip),%ymm8 # 4a44 <_sk_callback_hsw+0x1ec>
+ .byte 196,98,125,24,5,63,57,0,0 // vbroadcastss 0x393f(%rip),%ymm8 # 4c20 <_sk_callback_hsw+0x1ec>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
- .byte 196,98,125,24,13,90,55,0,0 // vbroadcastss 0x375a(%rip),%ymm9 # 4a48 <_sk_callback_hsw+0x1f0>
+ .byte 196,98,125,24,13,54,57,0,0 // vbroadcastss 0x3936(%rip),%ymm9 # 4c24 <_sk_callback_hsw+0x1f0>
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
.byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0
.byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1
@@ -10992,7 +10992,7 @@ HIDDEN _sk_clamp_1_hsw
.globl _sk_clamp_1_hsw
FUNCTION(_sk_clamp_1_hsw)
_sk_clamp_1_hsw:
- .byte 196,98,125,24,5,127,54,0,0 // vbroadcastss 0x367f(%rip),%ymm8 # 4a4c <_sk_callback_hsw+0x1f4>
+ .byte 196,98,125,24,5,91,56,0,0 // vbroadcastss 0x385b(%rip),%ymm8 # 4c28 <_sk_callback_hsw+0x1f4>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
@@ -11004,7 +11004,7 @@ HIDDEN _sk_clamp_a_hsw
.globl _sk_clamp_a_hsw
FUNCTION(_sk_clamp_a_hsw)
_sk_clamp_a_hsw:
- .byte 196,98,125,24,5,98,54,0,0 // vbroadcastss 0x3662(%rip),%ymm8 # 4a50 <_sk_callback_hsw+0x1f8>
+ .byte 196,98,125,24,5,62,56,0,0 // vbroadcastss 0x383e(%rip),%ymm8 # 4c2c <_sk_callback_hsw+0x1f8>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0
.byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1
@@ -11090,7 +11090,7 @@ FUNCTION(_sk_unpremul_hsw)
_sk_unpremul_hsw:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9
- .byte 196,98,125,24,21,170,53,0,0 // vbroadcastss 0x35aa(%rip),%ymm10 # 4a54 <_sk_callback_hsw+0x1fc>
+ .byte 196,98,125,24,21,134,55,0,0 // vbroadcastss 0x3786(%rip),%ymm10 # 4c30 <_sk_callback_hsw+0x1fc>
.byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10
.byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
@@ -11103,16 +11103,16 @@ HIDDEN _sk_from_srgb_hsw
.globl _sk_from_srgb_hsw
FUNCTION(_sk_from_srgb_hsw)
_sk_from_srgb_hsw:
- .byte 196,98,125,24,5,139,53,0,0 // vbroadcastss 0x358b(%rip),%ymm8 # 4a58 <_sk_callback_hsw+0x200>
+ .byte 196,98,125,24,5,103,55,0,0 // vbroadcastss 0x3767(%rip),%ymm8 # 4c34 <_sk_callback_hsw+0x200>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
- .byte 196,98,125,24,29,125,53,0,0 // vbroadcastss 0x357d(%rip),%ymm11 # 4a5c <_sk_callback_hsw+0x204>
- .byte 196,98,125,24,37,120,53,0,0 // vbroadcastss 0x3578(%rip),%ymm12 # 4a60 <_sk_callback_hsw+0x208>
+ .byte 196,98,125,24,29,89,55,0,0 // vbroadcastss 0x3759(%rip),%ymm11 # 4c38 <_sk_callback_hsw+0x204>
+ .byte 196,98,125,24,37,84,55,0,0 // vbroadcastss 0x3754(%rip),%ymm12 # 4c3c <_sk_callback_hsw+0x208>
.byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13
.byte 196,66,125,168,235 // vfmadd213ps %ymm11,%ymm0,%ymm13
- .byte 196,98,125,24,53,105,53,0,0 // vbroadcastss 0x3569(%rip),%ymm14 # 4a64 <_sk_callback_hsw+0x20c>
+ .byte 196,98,125,24,53,69,55,0,0 // vbroadcastss 0x3745(%rip),%ymm14 # 4c40 <_sk_callback_hsw+0x20c>
.byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13
- .byte 196,98,125,24,21,95,53,0,0 // vbroadcastss 0x355f(%rip),%ymm10 # 4a68 <_sk_callback_hsw+0x210>
+ .byte 196,98,125,24,21,59,55,0,0 // vbroadcastss 0x373b(%rip),%ymm10 # 4c44 <_sk_callback_hsw+0x210>
.byte 196,193,124,194,194,1 // vcmpltps %ymm10,%ymm0,%ymm0
.byte 196,195,21,74,193,0 // vblendvps %ymm0,%ymm9,%ymm13,%ymm0
.byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9
@@ -11136,19 +11136,19 @@ HIDDEN _sk_to_srgb_hsw
FUNCTION(_sk_to_srgb_hsw)
_sk_to_srgb_hsw:
.byte 197,124,82,200 // vrsqrtps %ymm0,%ymm9
- .byte 196,98,125,24,5,3,53,0,0 // vbroadcastss 0x3503(%rip),%ymm8 # 4a6c <_sk_callback_hsw+0x214>
+ .byte 196,98,125,24,5,223,54,0,0 // vbroadcastss 0x36df(%rip),%ymm8 # 4c48 <_sk_callback_hsw+0x214>
.byte 196,65,124,89,208 // vmulps %ymm8,%ymm0,%ymm10
- .byte 196,98,125,24,29,249,52,0,0 // vbroadcastss 0x34f9(%rip),%ymm11 # 4a70 <_sk_callback_hsw+0x218>
- .byte 196,98,125,24,37,244,52,0,0 // vbroadcastss 0x34f4(%rip),%ymm12 # 4a74 <_sk_callback_hsw+0x21c>
+ .byte 196,98,125,24,29,213,54,0,0 // vbroadcastss 0x36d5(%rip),%ymm11 # 4c4c <_sk_callback_hsw+0x218>
+ .byte 196,98,125,24,37,208,54,0,0 // vbroadcastss 0x36d0(%rip),%ymm12 # 4c50 <_sk_callback_hsw+0x21c>
.byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13
.byte 196,66,53,168,235 // vfmadd213ps %ymm11,%ymm9,%ymm13
- .byte 196,98,125,24,53,229,52,0,0 // vbroadcastss 0x34e5(%rip),%ymm14 # 4a78 <_sk_callback_hsw+0x220>
+ .byte 196,98,125,24,53,193,54,0,0 // vbroadcastss 0x36c1(%rip),%ymm14 # 4c54 <_sk_callback_hsw+0x220>
.byte 196,66,53,168,238 // vfmadd213ps %ymm14,%ymm9,%ymm13
- .byte 196,98,125,24,61,219,52,0,0 // vbroadcastss 0x34db(%rip),%ymm15 # 4a7c <_sk_callback_hsw+0x224>
+ .byte 196,98,125,24,61,183,54,0,0 // vbroadcastss 0x36b7(%rip),%ymm15 # 4c58 <_sk_callback_hsw+0x224>
.byte 196,65,52,88,207 // vaddps %ymm15,%ymm9,%ymm9
.byte 196,65,124,83,201 // vrcpps %ymm9,%ymm9
.byte 196,65,20,89,201 // vmulps %ymm9,%ymm13,%ymm9
- .byte 196,98,125,24,45,199,52,0,0 // vbroadcastss 0x34c7(%rip),%ymm13 # 4a80 <_sk_callback_hsw+0x228>
+ .byte 196,98,125,24,45,163,54,0,0 // vbroadcastss 0x36a3(%rip),%ymm13 # 4c5c <_sk_callback_hsw+0x228>
.byte 196,193,124,194,197,1 // vcmpltps %ymm13,%ymm0,%ymm0
.byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0
.byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9
@@ -11182,26 +11182,26 @@ _sk_rgb_to_hsl_hsw:
.byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9
.byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9
.byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,29,60,52,0,0 // vbroadcastss 0x343c(%rip),%ymm11 # 4a84 <_sk_callback_hsw+0x22c>
+ .byte 196,98,125,24,29,24,54,0,0 // vbroadcastss 0x3618(%rip),%ymm11 # 4c60 <_sk_callback_hsw+0x22c>
.byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11
.byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12
.byte 197,116,194,234,1 // vcmpltps %ymm2,%ymm1,%ymm13
- .byte 196,98,125,24,53,41,52,0,0 // vbroadcastss 0x3429(%rip),%ymm14 # 4a88 <_sk_callback_hsw+0x230>
+ .byte 196,98,125,24,53,5,54,0,0 // vbroadcastss 0x3605(%rip),%ymm14 # 4c64 <_sk_callback_hsw+0x230>
.byte 196,65,4,87,255 // vxorps %ymm15,%ymm15,%ymm15
.byte 196,67,5,74,238,208 // vblendvps %ymm13,%ymm14,%ymm15,%ymm13
.byte 196,66,37,168,229 // vfmadd213ps %ymm13,%ymm11,%ymm12
.byte 197,236,92,208 // vsubps %ymm0,%ymm2,%ymm2
.byte 197,124,92,233 // vsubps %ymm1,%ymm0,%ymm13
- .byte 196,98,125,24,53,16,52,0,0 // vbroadcastss 0x3410(%rip),%ymm14 # 4a90 <_sk_callback_hsw+0x238>
+ .byte 196,98,125,24,53,236,53,0,0 // vbroadcastss 0x35ec(%rip),%ymm14 # 4c6c <_sk_callback_hsw+0x238>
.byte 196,66,37,168,238 // vfmadd213ps %ymm14,%ymm11,%ymm13
- .byte 196,98,125,24,53,254,51,0,0 // vbroadcastss 0x33fe(%rip),%ymm14 # 4a8c <_sk_callback_hsw+0x234>
+ .byte 196,98,125,24,53,218,53,0,0 // vbroadcastss 0x35da(%rip),%ymm14 # 4c68 <_sk_callback_hsw+0x234>
.byte 196,194,37,168,214 // vfmadd213ps %ymm14,%ymm11,%ymm2
.byte 197,188,194,201,0 // vcmpeqps %ymm1,%ymm8,%ymm1
.byte 196,227,21,74,202,16 // vblendvps %ymm1,%ymm2,%ymm13,%ymm1
.byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0
.byte 196,195,117,74,196,0 // vblendvps %ymm0,%ymm12,%ymm1,%ymm0
.byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1
- .byte 196,98,125,24,29,225,51,0,0 // vbroadcastss 0x33e1(%rip),%ymm11 # 4a98 <_sk_callback_hsw+0x240>
+ .byte 196,98,125,24,29,189,53,0,0 // vbroadcastss 0x35bd(%rip),%ymm11 # 4c74 <_sk_callback_hsw+0x240>
.byte 196,193,116,89,211 // vmulps %ymm11,%ymm1,%ymm2
.byte 197,36,194,218,1 // vcmpltps %ymm2,%ymm11,%ymm11
.byte 196,65,12,92,224 // vsubps %ymm8,%ymm14,%ymm12
@@ -11211,7 +11211,7 @@ _sk_rgb_to_hsl_hsw:
.byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
.byte 196,195,125,74,199,128 // vblendvps %ymm8,%ymm15,%ymm0,%ymm0
.byte 196,195,117,74,207,128 // vblendvps %ymm8,%ymm15,%ymm1,%ymm1
- .byte 196,98,125,24,5,164,51,0,0 // vbroadcastss 0x33a4(%rip),%ymm8 # 4a94 <_sk_callback_hsw+0x23c>
+ .byte 196,98,125,24,5,128,53,0,0 // vbroadcastss 0x3580(%rip),%ymm8 # 4c70 <_sk_callback_hsw+0x23c>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -11228,30 +11228,30 @@ _sk_hsl_to_rgb_hsw:
.byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp)
.byte 197,252,40,233 // vmovaps %ymm1,%ymm5
.byte 197,252,40,224 // vmovaps %ymm0,%ymm4
- .byte 196,98,125,24,5,113,51,0,0 // vbroadcastss 0x3371(%rip),%ymm8 # 4a9c <_sk_callback_hsw+0x244>
+ .byte 196,98,125,24,5,77,53,0,0 // vbroadcastss 0x354d(%rip),%ymm8 # 4c78 <_sk_callback_hsw+0x244>
.byte 197,60,194,202,2 // vcmpleps %ymm2,%ymm8,%ymm9
.byte 197,84,89,210 // vmulps %ymm2,%ymm5,%ymm10
.byte 196,65,84,92,218 // vsubps %ymm10,%ymm5,%ymm11
.byte 196,67,45,74,203,144 // vblendvps %ymm9,%ymm11,%ymm10,%ymm9
.byte 197,52,88,210 // vaddps %ymm2,%ymm9,%ymm10
- .byte 196,98,125,24,13,84,51,0,0 // vbroadcastss 0x3354(%rip),%ymm9 # 4aa0 <_sk_callback_hsw+0x248>
+ .byte 196,98,125,24,13,48,53,0,0 // vbroadcastss 0x3530(%rip),%ymm9 # 4c7c <_sk_callback_hsw+0x248>
.byte 196,66,109,170,202 // vfmsub213ps %ymm10,%ymm2,%ymm9
- .byte 196,98,125,24,29,74,51,0,0 // vbroadcastss 0x334a(%rip),%ymm11 # 4aa4 <_sk_callback_hsw+0x24c>
+ .byte 196,98,125,24,29,38,53,0,0 // vbroadcastss 0x3526(%rip),%ymm11 # 4c80 <_sk_callback_hsw+0x24c>
.byte 196,65,92,88,219 // vaddps %ymm11,%ymm4,%ymm11
.byte 196,67,125,8,227,1 // vroundps $0x1,%ymm11,%ymm12
.byte 196,65,36,92,252 // vsubps %ymm12,%ymm11,%ymm15
.byte 196,65,44,92,217 // vsubps %ymm9,%ymm10,%ymm11
- .byte 196,98,125,24,45,52,51,0,0 // vbroadcastss 0x3334(%rip),%ymm13 # 4aac <_sk_callback_hsw+0x254>
+ .byte 196,98,125,24,45,16,53,0,0 // vbroadcastss 0x3510(%rip),%ymm13 # 4c88 <_sk_callback_hsw+0x254>
.byte 196,193,4,89,197 // vmulps %ymm13,%ymm15,%ymm0
- .byte 196,98,125,24,53,42,51,0,0 // vbroadcastss 0x332a(%rip),%ymm14 # 4ab0 <_sk_callback_hsw+0x258>
+ .byte 196,98,125,24,53,6,53,0,0 // vbroadcastss 0x3506(%rip),%ymm14 # 4c8c <_sk_callback_hsw+0x258>
.byte 197,12,92,224 // vsubps %ymm0,%ymm14,%ymm12
.byte 196,66,37,168,225 // vfmadd213ps %ymm9,%ymm11,%ymm12
- .byte 196,226,125,24,29,16,51,0,0 // vbroadcastss 0x3310(%rip),%ymm3 # 4aa8 <_sk_callback_hsw+0x250>
+ .byte 196,226,125,24,29,236,52,0,0 // vbroadcastss 0x34ec(%rip),%ymm3 # 4c84 <_sk_callback_hsw+0x250>
.byte 196,193,100,194,255,2 // vcmpleps %ymm15,%ymm3,%ymm7
.byte 196,195,29,74,249,112 // vblendvps %ymm7,%ymm9,%ymm12,%ymm7
.byte 196,65,60,194,231,2 // vcmpleps %ymm15,%ymm8,%ymm12
.byte 196,227,45,74,255,192 // vblendvps %ymm12,%ymm7,%ymm10,%ymm7
- .byte 196,98,125,24,37,251,50,0,0 // vbroadcastss 0x32fb(%rip),%ymm12 # 4ab4 <_sk_callback_hsw+0x25c>
+ .byte 196,98,125,24,37,215,52,0,0 // vbroadcastss 0x34d7(%rip),%ymm12 # 4c90 <_sk_callback_hsw+0x25c>
.byte 196,65,28,194,255,2 // vcmpleps %ymm15,%ymm12,%ymm15
.byte 196,194,37,168,193 // vfmadd213ps %ymm9,%ymm11,%ymm0
.byte 196,99,125,74,255,240 // vblendvps %ymm15,%ymm7,%ymm0,%ymm15
@@ -11267,7 +11267,7 @@ _sk_hsl_to_rgb_hsw:
.byte 197,156,194,192,2 // vcmpleps %ymm0,%ymm12,%ymm0
.byte 196,194,37,168,249 // vfmadd213ps %ymm9,%ymm11,%ymm7
.byte 196,227,69,74,201,0 // vblendvps %ymm0,%ymm1,%ymm7,%ymm1
- .byte 196,226,125,24,5,167,50,0,0 // vbroadcastss 0x32a7(%rip),%ymm0 # 4ab8 <_sk_callback_hsw+0x260>
+ .byte 196,226,125,24,5,131,52,0,0 // vbroadcastss 0x3483(%rip),%ymm0 # 4c94 <_sk_callback_hsw+0x260>
.byte 197,220,88,192 // vaddps %ymm0,%ymm4,%ymm0
.byte 196,227,125,8,224,1 // vroundps $0x1,%ymm0,%ymm4
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
@@ -11312,38 +11312,69 @@ HIDDEN _sk_scale_u8_hsw
.globl _sk_scale_u8_hsw
FUNCTION(_sk_scale_u8_hsw)
_sk_scale_u8_hsw:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,51 // jne 18f1 <_sk_scale_u8_hsw+0x43>
- .byte 197,122,126,0 // vmovq (%rax),%xmm8
- .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8
+ .byte 117,58 // jne 18f2 <_sk_scale_u8_hsw+0x44>
+ .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8
+ .byte 197,57,219,5,154,56,0,0 // vpand 0x389a(%rip),%xmm8,%xmm8 # 5160 <_sk_callback_hsw+0x72c>
+ .byte 196,66,125,51,192 // vpmovzxwd %xmm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,231,49,0,0 // vbroadcastss 0x31e7(%rip),%ymm9 # 4abc <_sk_callback_hsw+0x264>
+ .byte 196,98,125,24,13,191,51,0,0 // vbroadcastss 0x33bf(%rip),%ymm9 # 4c98 <_sk_callback_hsw+0x264>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
.byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
.byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
.byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 76,137,201 // mov %r9,%rcx
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 18fa <_sk_scale_u8_hsw+0x4c>
- .byte 196,65,249,110,194 // vmovq %r10,%xmm8
- .byte 91 // pop %rbx
- .byte 235,171 // jmp 18c2 <_sk_scale_u8_hsw+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,183 // ja 18be <_sk_scale_u8_hsw+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,98,0,0,0 // lea 0x62(%rip),%r10 # 1974 <_sk_scale_u8_hsw+0xc6>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 197,57,196,192,6 // vpinsrw $0x6,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,5 // vpinsrw $0x5,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,4 // vpinsrw $0x4,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,3 // vpinsrw $0x3,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,2 // vpinsrw $0x2,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,1 // vpinsrw $0x1,%eax,%xmm8,%xmm8
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,57,196,192,0 // vpinsrw $0x0,%eax,%xmm8,%xmm8
+ .byte 233,77,255,255,255 // jmpq 18be <_sk_scale_u8_hsw+0x10>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 238 // out %al,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,227 // jmpq *%rbx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 216,255 // fdivr %st(7),%st
+ .byte 255 // (bad)
+ .byte 255,205 // dec %ebp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,194 // inc %edx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,183,255,255,255,167 // pushq -0x58000001(%rdi)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_lerp_1_float_hsw
.globl _sk_lerp_1_float_hsw
@@ -11366,16 +11397,15 @@ HIDDEN _sk_lerp_u8_hsw
.globl _sk_lerp_u8_hsw
FUNCTION(_sk_lerp_u8_hsw)
_sk_lerp_u8_hsw:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,71 // jne 199d <_sk_lerp_u8_hsw+0x57>
- .byte 197,122,126,0 // vmovq (%rax),%xmm8
- .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8
+ .byte 117,78 // jne 1a17 <_sk_lerp_u8_hsw+0x58>
+ .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8
+ .byte 197,57,219,5,153,55,0,0 // vpand 0x3799(%rip),%xmm8,%xmm8 # 5170 <_sk_callback_hsw+0x73c>
+ .byte 196,66,125,51,192 // vpmovzxwd %xmm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,83,49,0,0 // vbroadcastss 0x3153(%rip),%ymm9 # 4ac0 <_sk_callback_hsw+0x268>
+ .byte 196,98,125,24,13,178,50,0,0 // vbroadcastss 0x32b2(%rip),%ymm9 # 4c9c <_sk_callback_hsw+0x268>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
.byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0
@@ -11386,22 +11416,55 @@ _sk_lerp_u8_hsw:
.byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3
.byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 76,137,201 // mov %r9,%rcx
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 19a6 <_sk_lerp_u8_hsw+0x60>
- .byte 196,65,249,110,194 // vmovq %r10,%xmm8
- .byte 91 // pop %rbx
- .byte 235,151 // jmp 195a <_sk_lerp_u8_hsw+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,163 // ja 19cf <_sk_lerp_u8_hsw+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,97,0,0,0 // lea 0x61(%rip),%r10 # 1a98 <_sk_lerp_u8_hsw+0xd9>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 197,57,196,192,6 // vpinsrw $0x6,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,5 // vpinsrw $0x5,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,4 // vpinsrw $0x4,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,3 // vpinsrw $0x3,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,2 // vpinsrw $0x2,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,1 // vpinsrw $0x1,%eax,%xmm8,%xmm8
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,57,196,192,0 // vpinsrw $0x0,%eax,%xmm8,%xmm8
+ .byte 233,57,255,255,255 // jmpq 19cf <_sk_lerp_u8_hsw+0x10>
+ .byte 102,144 // xchg %ax,%ax
+ .byte 239 // out %eax,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,228 // jmpq *%rsp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 217,255 // fcos
+ .byte 255 // (bad)
+ .byte 255,206 // dec %esi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,195 // inc %ebx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 184,255,255,255,168 // mov $0xa8ffffff,%eax
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_lerp_565_hsw
.globl _sk_lerp_565_hsw
@@ -11410,23 +11473,23 @@ _sk_lerp_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,169,0,0,0 // jne 1a7a <_sk_lerp_565_hsw+0xb7>
+ .byte 15,133,169,0,0,0 // jne 1b6b <_sk_lerp_565_hsw+0xb7>
.byte 196,65,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm8
.byte 196,66,125,51,192 // vpmovzxwd %xmm8,%ymm8
- .byte 196,98,125,88,13,223,48,0,0 // vpbroadcastd 0x30df(%rip),%ymm9 # 4ac4 <_sk_callback_hsw+0x26c>
+ .byte 196,98,125,88,13,202,49,0,0 // vpbroadcastd 0x31ca(%rip),%ymm9 # 4ca0 <_sk_callback_hsw+0x26c>
.byte 196,65,61,219,201 // vpand %ymm9,%ymm8,%ymm9
.byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9
- .byte 196,98,125,24,21,208,48,0,0 // vbroadcastss 0x30d0(%rip),%ymm10 # 4ac8 <_sk_callback_hsw+0x270>
+ .byte 196,98,125,24,21,187,49,0,0 // vbroadcastss 0x31bb(%rip),%ymm10 # 4ca4 <_sk_callback_hsw+0x270>
.byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,88,21,198,48,0,0 // vpbroadcastd 0x30c6(%rip),%ymm10 # 4acc <_sk_callback_hsw+0x274>
+ .byte 196,98,125,88,21,177,49,0,0 // vpbroadcastd 0x31b1(%rip),%ymm10 # 4ca8 <_sk_callback_hsw+0x274>
.byte 196,65,61,219,210 // vpand %ymm10,%ymm8,%ymm10
.byte 196,65,124,91,210 // vcvtdq2ps %ymm10,%ymm10
- .byte 196,98,125,24,29,183,48,0,0 // vbroadcastss 0x30b7(%rip),%ymm11 # 4ad0 <_sk_callback_hsw+0x278>
+ .byte 196,98,125,24,29,162,49,0,0 // vbroadcastss 0x31a2(%rip),%ymm11 # 4cac <_sk_callback_hsw+0x278>
.byte 196,65,44,89,211 // vmulps %ymm11,%ymm10,%ymm10
- .byte 196,98,125,88,29,173,48,0,0 // vpbroadcastd 0x30ad(%rip),%ymm11 # 4ad4 <_sk_callback_hsw+0x27c>
+ .byte 196,98,125,88,29,152,49,0,0 // vpbroadcastd 0x3198(%rip),%ymm11 # 4cb0 <_sk_callback_hsw+0x27c>
.byte 196,65,61,219,195 // vpand %ymm11,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,29,158,48,0,0 // vbroadcastss 0x309e(%rip),%ymm11 # 4ad8 <_sk_callback_hsw+0x280>
+ .byte 196,98,125,24,29,137,49,0,0 // vbroadcastss 0x3189(%rip),%ymm11 # 4cb4 <_sk_callback_hsw+0x280>
.byte 196,65,60,89,195 // vmulps %ymm11,%ymm8,%ymm8
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
.byte 196,226,53,168,196 // vfmadd213ps %ymm4,%ymm9,%ymm0
@@ -11447,9 +11510,9 @@ _sk_lerp_565_hsw:
.byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 15,135,68,255,255,255 // ja 19d7 <_sk_lerp_565_hsw+0x14>
+ .byte 15,135,68,255,255,255 // ja 1ac8 <_sk_lerp_565_hsw+0x14>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,74,0,0,0 // lea 0x4a(%rip),%r10 # 1ae8 <_sk_lerp_565_hsw+0x125>
+ .byte 76,141,21,77,0,0,0 // lea 0x4d(%rip),%r10 # 1bdc <_sk_lerp_565_hsw+0x128>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -11461,28 +11524,26 @@ _sk_lerp_565_hsw:
.byte 196,65,57,196,68,83,4,2 // vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm8,%xmm8
.byte 196,65,57,196,68,83,2,1 // vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm8,%xmm8
.byte 196,65,57,196,4,83,0 // vpinsrw $0x0,(%r11,%rdx,2),%xmm8,%xmm8
- .byte 233,239,254,255,255 // jmpq 19d7 <_sk_lerp_565_hsw+0x14>
- .byte 244 // hlt
- .byte 255 // (bad)
- .byte 255 // (bad)
+ .byte 233,239,254,255,255 // jmpq 1ac8 <_sk_lerp_565_hsw+0x14>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 241 // icebp
.byte 255 // (bad)
- .byte 236 // in (%dx),%al
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,228 // jmpq *%rsp
+ .byte 233,255,255,255,225 // jmpq ffffffffe2001be4 <_sk_callback_hsw+0xffffffffe1ffd1b0>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 220,255 // fdivr %st,%st(7)
+ .byte 217,255 // fcos
.byte 255 // (bad)
- .byte 255,212 // callq *%rsp
+ .byte 255,209 // callq *%rcx
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,204 // dec %esp
+ .byte 255,201 // dec %ecx
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 191 // .byte 0xbf
+ .byte 188 // .byte 0xbc
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -11496,23 +11557,23 @@ _sk_load_tables_hsw:
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 76,3,16 // add (%rax),%r10
.byte 77,133,192 // test %r8,%r8
- .byte 117,105 // jne 1b82 <_sk_load_tables_hsw+0x7e>
+ .byte 117,105 // jne 1c76 <_sk_load_tables_hsw+0x7e>
.byte 196,193,124,16,26 // vmovups (%r10),%ymm3
- .byte 197,228,84,13,250,50,0,0 // vandps 0x32fa(%rip),%ymm3,%ymm1 # 4e20 <_sk_callback_hsw+0x5c8>
+ .byte 197,228,84,13,198,51,0,0 // vandps 0x33c6(%rip),%ymm3,%ymm1 # 4fe0 <_sk_callback_hsw+0x5ac>
.byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
.byte 72,139,72,8 // mov 0x8(%rax),%rcx
.byte 76,139,80,16 // mov 0x10(%rax),%r10
.byte 197,237,118,210 // vpcmpeqd %ymm2,%ymm2,%ymm2
.byte 196,226,109,146,4,137 // vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0
- .byte 196,226,101,0,21,250,50,0,0 // vpshufb 0x32fa(%rip),%ymm3,%ymm2 # 4e40 <_sk_callback_hsw+0x5e8>
+ .byte 196,226,101,0,21,198,51,0,0 // vpshufb 0x33c6(%rip),%ymm3,%ymm2 # 5000 <_sk_callback_hsw+0x5cc>
.byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
.byte 196,194,53,146,12,146 // vgatherdps %ymm9,(%r10,%ymm2,4),%ymm1
.byte 72,139,64,24 // mov 0x18(%rax),%rax
- .byte 196,98,101,0,13,2,51,0,0 // vpshufb 0x3302(%rip),%ymm3,%ymm9 # 4e60 <_sk_callback_hsw+0x608>
+ .byte 196,98,101,0,13,206,51,0,0 // vpshufb 0x33ce(%rip),%ymm3,%ymm9 # 5020 <_sk_callback_hsw+0x5ec>
.byte 196,162,61,146,20,136 // vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2
.byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,102,47,0,0 // vbroadcastss 0x2f66(%rip),%ymm8 # 4adc <_sk_callback_hsw+0x284>
+ .byte 196,98,125,24,5,78,48,0,0 // vbroadcastss 0x304e(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x284>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,201 // mov %r9,%rcx
@@ -11525,7 +11586,7 @@ _sk_load_tables_hsw:
.byte 196,193,249,110,195 // vmovq %r11,%xmm0
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
.byte 196,194,125,44,26 // vmaskmovps (%r10),%ymm0,%ymm3
- .byte 233,115,255,255,255 // jmpq 1b1e <_sk_load_tables_hsw+0x1a>
+ .byte 233,115,255,255,255 // jmpq 1c12 <_sk_load_tables_hsw+0x1a>
HIDDEN _sk_load_tables_u16_be_hsw
.globl _sk_load_tables_u16_be_hsw
@@ -11535,7 +11596,7 @@ _sk_load_tables_u16_be_hsw:
.byte 76,139,8 // mov (%rax),%r9
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,201,0,0,0 // jne 1c8a <_sk_load_tables_u16_be_hsw+0xdf>
+ .byte 15,133,201,0,0,0 // jne 1d7e <_sk_load_tables_u16_be_hsw+0xdf>
.byte 196,1,121,16,4,81 // vmovupd (%r9,%r10,2),%xmm8
.byte 196,129,121,16,84,81,16 // vmovupd 0x10(%r9,%r10,2),%xmm2
.byte 196,129,121,16,92,81,32 // vmovupd 0x20(%r9,%r10,2),%xmm3
@@ -11551,7 +11612,7 @@ _sk_load_tables_u16_be_hsw:
.byte 197,185,108,200 // vpunpcklqdq %xmm0,%xmm8,%xmm1
.byte 197,185,109,208 // vpunpckhqdq %xmm0,%xmm8,%xmm2
.byte 197,49,108,195 // vpunpcklqdq %xmm3,%xmm9,%xmm8
- .byte 197,121,111,21,142,51,0,0 // vmovdqa 0x338e(%rip),%xmm10 # 4fa0 <_sk_callback_hsw+0x748>
+ .byte 197,121,111,21,122,52,0,0 // vmovdqa 0x347a(%rip),%xmm10 # 5180 <_sk_callback_hsw+0x74c>
.byte 196,193,113,219,194 // vpand %xmm10,%xmm1,%xmm0
.byte 196,226,125,51,200 // vpmovzxwd %xmm0,%ymm1
.byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11
@@ -11573,36 +11634,36 @@ _sk_load_tables_u16_be_hsw:
.byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3
.byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,95,46,0,0 // vbroadcastss 0x2e5f(%rip),%ymm8 # 4ae0 <_sk_callback_hsw+0x288>
+ .byte 196,98,125,24,5,71,47,0,0 // vbroadcastss 0x2f47(%rip),%ymm8 # 4cbc <_sk_callback_hsw+0x288>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
.byte 196,1,123,16,4,81 // vmovsd (%r9,%r10,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,85 // je 1cf0 <_sk_load_tables_u16_be_hsw+0x145>
+ .byte 116,85 // je 1de4 <_sk_load_tables_u16_be_hsw+0x145>
.byte 196,1,57,22,68,81,8 // vmovhpd 0x8(%r9,%r10,2),%xmm8,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,72 // jb 1cf0 <_sk_load_tables_u16_be_hsw+0x145>
+ .byte 114,72 // jb 1de4 <_sk_load_tables_u16_be_hsw+0x145>
.byte 196,129,123,16,84,81,16 // vmovsd 0x10(%r9,%r10,2),%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 116,72 // je 1cfd <_sk_load_tables_u16_be_hsw+0x152>
+ .byte 116,72 // je 1df1 <_sk_load_tables_u16_be_hsw+0x152>
.byte 196,129,105,22,84,81,24 // vmovhpd 0x18(%r9,%r10,2),%xmm2,%xmm2
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,59 // jb 1cfd <_sk_load_tables_u16_be_hsw+0x152>
+ .byte 114,59 // jb 1df1 <_sk_load_tables_u16_be_hsw+0x152>
.byte 196,129,123,16,92,81,32 // vmovsd 0x20(%r9,%r10,2),%xmm3
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 15,132,9,255,255,255 // je 1bdc <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 15,132,9,255,255,255 // je 1cd0 <_sk_load_tables_u16_be_hsw+0x31>
.byte 196,129,97,22,92,81,40 // vmovhpd 0x28(%r9,%r10,2),%xmm3,%xmm3
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 15,130,248,254,255,255 // jb 1bdc <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 15,130,248,254,255,255 // jb 1cd0 <_sk_load_tables_u16_be_hsw+0x31>
.byte 196,1,122,126,76,81,48 // vmovq 0x30(%r9,%r10,2),%xmm9
- .byte 233,236,254,255,255 // jmpq 1bdc <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,236,254,255,255 // jmpq 1cd0 <_sk_load_tables_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,223,254,255,255 // jmpq 1bdc <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,223,254,255,255 // jmpq 1cd0 <_sk_load_tables_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,214,254,255,255 // jmpq 1bdc <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,214,254,255,255 // jmpq 1cd0 <_sk_load_tables_u16_be_hsw+0x31>
HIDDEN _sk_load_tables_rgb_u16_be_hsw
.globl _sk_load_tables_rgb_u16_be_hsw
@@ -11612,7 +11673,7 @@ _sk_load_tables_rgb_u16_be_hsw:
.byte 76,139,8 // mov (%rax),%r9
.byte 76,141,20,82 // lea (%rdx,%rdx,2),%r10
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,193,0,0,0 // jne 1dd9 <_sk_load_tables_rgb_u16_be_hsw+0xd3>
+ .byte 15,133,193,0,0,0 // jne 1ecd <_sk_load_tables_rgb_u16_be_hsw+0xd3>
.byte 196,129,122,111,4,81 // vmovdqu (%r9,%r10,2),%xmm0
.byte 196,129,122,111,84,81,12 // vmovdqu 0xc(%r9,%r10,2),%xmm2
.byte 196,129,122,111,76,81,24 // vmovdqu 0x18(%r9,%r10,2),%xmm1
@@ -11633,7 +11694,7 @@ _sk_load_tables_rgb_u16_be_hsw:
.byte 197,185,108,218 // vpunpcklqdq %xmm2,%xmm8,%xmm3
.byte 197,185,109,210 // vpunpckhqdq %xmm2,%xmm8,%xmm2
.byte 197,121,108,193 // vpunpcklqdq %xmm1,%xmm0,%xmm8
- .byte 197,121,111,13,46,50,0,0 // vmovdqa 0x322e(%rip),%xmm9 # 4fb0 <_sk_callback_hsw+0x758>
+ .byte 197,121,111,13,26,51,0,0 // vmovdqa 0x331a(%rip),%xmm9 # 5190 <_sk_callback_hsw+0x75c>
.byte 196,193,97,219,193 // vpand %xmm9,%xmm3,%xmm0
.byte 196,226,125,51,200 // vpmovzxwd %xmm0,%ymm1
.byte 197,229,118,219 // vpcmpeqd %ymm3,%ymm3,%ymm3
@@ -11650,48 +11711,48 @@ _sk_load_tables_rgb_u16_be_hsw:
.byte 196,98,125,51,194 // vpmovzxwd %xmm2,%ymm8
.byte 196,162,101,146,20,128 // vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,13,45,0,0 // vbroadcastss 0x2d0d(%rip),%ymm3 # 4ae4 <_sk_callback_hsw+0x28c>
+ .byte 196,226,125,24,29,245,45,0,0 // vbroadcastss 0x2df5(%rip),%ymm3 # 4cc0 <_sk_callback_hsw+0x28c>
.byte 255,224 // jmpq *%rax
.byte 196,129,121,110,4,81 // vmovd (%r9,%r10,2),%xmm0
.byte 196,129,121,196,68,81,4,2 // vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 117,5 // jne 1df2 <_sk_load_tables_rgb_u16_be_hsw+0xec>
- .byte 233,90,255,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,5 // jne 1ee6 <_sk_load_tables_rgb_u16_be_hsw+0xec>
+ .byte 233,90,255,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
.byte 196,129,121,110,76,81,6 // vmovd 0x6(%r9,%r10,2),%xmm1
.byte 196,1,113,196,68,81,10,2 // vpinsrw $0x2,0xa(%r9,%r10,2),%xmm1,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,26 // jb 1e21 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
+ .byte 114,26 // jb 1f15 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
.byte 196,129,121,110,76,81,12 // vmovd 0xc(%r9,%r10,2),%xmm1
.byte 196,129,113,196,84,81,16,2 // vpinsrw $0x2,0x10(%r9,%r10,2),%xmm1,%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 117,10 // jne 1e26 <_sk_load_tables_rgb_u16_be_hsw+0x120>
- .byte 233,43,255,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,38,255,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 1f1a <_sk_load_tables_rgb_u16_be_hsw+0x120>
+ .byte 233,43,255,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,38,255,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
.byte 196,129,121,110,76,81,18 // vmovd 0x12(%r9,%r10,2),%xmm1
.byte 196,1,113,196,76,81,22,2 // vpinsrw $0x2,0x16(%r9,%r10,2),%xmm1,%xmm9
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,26 // jb 1e55 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
+ .byte 114,26 // jb 1f49 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
.byte 196,129,121,110,76,81,24 // vmovd 0x18(%r9,%r10,2),%xmm1
.byte 196,129,113,196,76,81,28,2 // vpinsrw $0x2,0x1c(%r9,%r10,2),%xmm1,%xmm1
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 117,10 // jne 1e5a <_sk_load_tables_rgb_u16_be_hsw+0x154>
- .byte 233,247,254,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,242,254,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 1f4e <_sk_load_tables_rgb_u16_be_hsw+0x154>
+ .byte 233,247,254,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,242,254,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
.byte 196,129,121,110,92,81,30 // vmovd 0x1e(%r9,%r10,2),%xmm3
.byte 196,1,97,196,92,81,34,2 // vpinsrw $0x2,0x22(%r9,%r10,2),%xmm3,%xmm11
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,20 // jb 1e83 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
+ .byte 114,20 // jb 1f77 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
.byte 196,129,121,110,92,81,36 // vmovd 0x24(%r9,%r10,2),%xmm3
.byte 196,129,97,196,92,81,40,2 // vpinsrw $0x2,0x28(%r9,%r10,2),%xmm3,%xmm3
- .byte 233,201,254,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,196,254,255,255 // jmpq 1d4c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,201,254,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,196,254,255,255 // jmpq 1e40 <_sk_load_tables_rgb_u16_be_hsw+0x46>
HIDDEN _sk_byte_tables_hsw
.globl _sk_byte_tables_hsw
FUNCTION(_sk_byte_tables_hsw)
_sk_byte_tables_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,85,44,0,0 // vbroadcastss 0x2c55(%rip),%ymm8 # 4ae8 <_sk_callback_hsw+0x290>
+ .byte 196,98,125,24,5,61,45,0,0 // vbroadcastss 0x2d3d(%rip),%ymm8 # 4cc4 <_sk_callback_hsw+0x290>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 197,125,91,200 // vcvtps2dq %ymm0,%ymm9
.byte 196,65,249,126,201 // vmovq %xmm9,%r9
@@ -11813,7 +11874,7 @@ _sk_byte_tables_hsw:
.byte 67,15,182,4,26 // movzbl (%r10,%r11,1),%eax
.byte 196,194,125,49,193 // vpmovzxbd %xmm9,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,250,41,0,0 // vbroadcastss 0x29fa(%rip),%ymm8 # 4aec <_sk_callback_hsw+0x294>
+ .byte 196,98,125,24,5,226,42,0,0 // vbroadcastss 0x2ae2(%rip),%ymm8 # 4cc8 <_sk_callback_hsw+0x294>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
@@ -11929,7 +11990,7 @@ _sk_byte_tables_rgb_hsw:
.byte 67,15,182,4,26 // movzbl (%r10,%r11,1),%eax
.byte 196,194,125,49,193 // vpmovzxbd %xmm9,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,230,39,0,0 // vbroadcastss 0x27e6(%rip),%ymm8 # 4af0 <_sk_callback_hsw+0x298>
+ .byte 196,98,125,24,5,206,40,0,0 // vbroadcastss 0x28ce(%rip),%ymm8 # 4ccc <_sk_callback_hsw+0x298>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
@@ -12028,33 +12089,33 @@ _sk_parametric_r_hsw:
.byte 196,66,125,168,211 // vfmadd213ps %ymm11,%ymm0,%ymm10
.byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0
.byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
- .byte 196,98,125,24,37,192,38,0,0 // vbroadcastss 0x26c0(%rip),%ymm12 # 4af4 <_sk_callback_hsw+0x29c>
- .byte 196,98,125,24,45,187,38,0,0 // vbroadcastss 0x26bb(%rip),%ymm13 # 4af8 <_sk_callback_hsw+0x2a0>
+ .byte 196,98,125,24,37,168,39,0,0 // vbroadcastss 0x27a8(%rip),%ymm12 # 4cd0 <_sk_callback_hsw+0x29c>
+ .byte 196,98,125,24,45,163,39,0,0 // vbroadcastss 0x27a3(%rip),%ymm13 # 4cd4 <_sk_callback_hsw+0x2a0>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,177,38,0,0 // vbroadcastss 0x26b1(%rip),%ymm13 # 4afc <_sk_callback_hsw+0x2a4>
+ .byte 196,98,125,24,45,153,39,0,0 // vbroadcastss 0x2799(%rip),%ymm13 # 4cd8 <_sk_callback_hsw+0x2a4>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,167,38,0,0 // vbroadcastss 0x26a7(%rip),%ymm13 # 4b00 <_sk_callback_hsw+0x2a8>
+ .byte 196,98,125,24,45,143,39,0,0 // vbroadcastss 0x278f(%rip),%ymm13 # 4cdc <_sk_callback_hsw+0x2a8>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,157,38,0,0 // vbroadcastss 0x269d(%rip),%ymm11 # 4b04 <_sk_callback_hsw+0x2ac>
+ .byte 196,98,125,24,29,133,39,0,0 // vbroadcastss 0x2785(%rip),%ymm11 # 4ce0 <_sk_callback_hsw+0x2ac>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,147,38,0,0 // vbroadcastss 0x2693(%rip),%ymm12 # 4b08 <_sk_callback_hsw+0x2b0>
+ .byte 196,98,125,24,37,123,39,0,0 // vbroadcastss 0x277b(%rip),%ymm12 # 4ce4 <_sk_callback_hsw+0x2b0>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,137,38,0,0 // vbroadcastss 0x2689(%rip),%ymm12 # 4b0c <_sk_callback_hsw+0x2b4>
+ .byte 196,98,125,24,37,113,39,0,0 // vbroadcastss 0x2771(%rip),%ymm12 # 4ce8 <_sk_callback_hsw+0x2b4>
.byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
.byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0
.byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10
.byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10
- .byte 196,98,125,24,29,106,38,0,0 // vbroadcastss 0x266a(%rip),%ymm11 # 4b10 <_sk_callback_hsw+0x2b8>
+ .byte 196,98,125,24,29,82,39,0,0 // vbroadcastss 0x2752(%rip),%ymm11 # 4cec <_sk_callback_hsw+0x2b8>
.byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,96,38,0,0 // vbroadcastss 0x2660(%rip),%ymm11 # 4b14 <_sk_callback_hsw+0x2bc>
+ .byte 196,98,125,24,29,72,39,0,0 // vbroadcastss 0x2748(%rip),%ymm11 # 4cf0 <_sk_callback_hsw+0x2bc>
.byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11
- .byte 196,226,125,24,5,86,38,0,0 // vbroadcastss 0x2656(%rip),%ymm0 # 4b18 <_sk_callback_hsw+0x2c0>
+ .byte 196,226,125,24,5,62,39,0,0 // vbroadcastss 0x273e(%rip),%ymm0 # 4cf4 <_sk_callback_hsw+0x2c0>
.byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0
- .byte 196,98,125,24,21,76,38,0,0 // vbroadcastss 0x264c(%rip),%ymm10 # 4b1c <_sk_callback_hsw+0x2c4>
+ .byte 196,98,125,24,21,52,39,0,0 // vbroadcastss 0x2734(%rip),%ymm10 # 4cf8 <_sk_callback_hsw+0x2c4>
.byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0
.byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
- .byte 196,98,125,24,21,63,38,0,0 // vbroadcastss 0x263f(%rip),%ymm10 # 4b20 <_sk_callback_hsw+0x2c8>
+ .byte 196,98,125,24,21,39,39,0,0 // vbroadcastss 0x2727(%rip),%ymm10 # 4cfc <_sk_callback_hsw+0x2c8>
.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
@@ -12062,7 +12123,7 @@ _sk_parametric_r_hsw:
.byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,22,38,0,0 // vbroadcastss 0x2616(%rip),%ymm8 # 4b24 <_sk_callback_hsw+0x2cc>
+ .byte 196,98,125,24,5,254,38,0,0 // vbroadcastss 0x26fe(%rip),%ymm8 # 4d00 <_sk_callback_hsw+0x2cc>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12082,33 +12143,33 @@ _sk_parametric_g_hsw:
.byte 196,66,117,168,211 // vfmadd213ps %ymm11,%ymm1,%ymm10
.byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1
.byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
- .byte 196,98,125,24,37,206,37,0,0 // vbroadcastss 0x25ce(%rip),%ymm12 # 4b28 <_sk_callback_hsw+0x2d0>
- .byte 196,98,125,24,45,201,37,0,0 // vbroadcastss 0x25c9(%rip),%ymm13 # 4b2c <_sk_callback_hsw+0x2d4>
+ .byte 196,98,125,24,37,182,38,0,0 // vbroadcastss 0x26b6(%rip),%ymm12 # 4d04 <_sk_callback_hsw+0x2d0>
+ .byte 196,98,125,24,45,177,38,0,0 // vbroadcastss 0x26b1(%rip),%ymm13 # 4d08 <_sk_callback_hsw+0x2d4>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,191,37,0,0 // vbroadcastss 0x25bf(%rip),%ymm13 # 4b30 <_sk_callback_hsw+0x2d8>
+ .byte 196,98,125,24,45,167,38,0,0 // vbroadcastss 0x26a7(%rip),%ymm13 # 4d0c <_sk_callback_hsw+0x2d8>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,181,37,0,0 // vbroadcastss 0x25b5(%rip),%ymm13 # 4b34 <_sk_callback_hsw+0x2dc>
+ .byte 196,98,125,24,45,157,38,0,0 // vbroadcastss 0x269d(%rip),%ymm13 # 4d10 <_sk_callback_hsw+0x2dc>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,171,37,0,0 // vbroadcastss 0x25ab(%rip),%ymm11 # 4b38 <_sk_callback_hsw+0x2e0>
+ .byte 196,98,125,24,29,147,38,0,0 // vbroadcastss 0x2693(%rip),%ymm11 # 4d14 <_sk_callback_hsw+0x2e0>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,161,37,0,0 // vbroadcastss 0x25a1(%rip),%ymm12 # 4b3c <_sk_callback_hsw+0x2e4>
+ .byte 196,98,125,24,37,137,38,0,0 // vbroadcastss 0x2689(%rip),%ymm12 # 4d18 <_sk_callback_hsw+0x2e4>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,151,37,0,0 // vbroadcastss 0x2597(%rip),%ymm12 # 4b40 <_sk_callback_hsw+0x2e8>
+ .byte 196,98,125,24,37,127,38,0,0 // vbroadcastss 0x267f(%rip),%ymm12 # 4d1c <_sk_callback_hsw+0x2e8>
.byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
.byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1
.byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10
.byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
- .byte 196,98,125,24,29,120,37,0,0 // vbroadcastss 0x2578(%rip),%ymm11 # 4b44 <_sk_callback_hsw+0x2ec>
+ .byte 196,98,125,24,29,96,38,0,0 // vbroadcastss 0x2660(%rip),%ymm11 # 4d20 <_sk_callback_hsw+0x2ec>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,110,37,0,0 // vbroadcastss 0x256e(%rip),%ymm11 # 4b48 <_sk_callback_hsw+0x2f0>
+ .byte 196,98,125,24,29,86,38,0,0 // vbroadcastss 0x2656(%rip),%ymm11 # 4d24 <_sk_callback_hsw+0x2f0>
.byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11
- .byte 196,226,125,24,13,100,37,0,0 // vbroadcastss 0x2564(%rip),%ymm1 # 4b4c <_sk_callback_hsw+0x2f4>
+ .byte 196,226,125,24,13,76,38,0,0 // vbroadcastss 0x264c(%rip),%ymm1 # 4d28 <_sk_callback_hsw+0x2f4>
.byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1
- .byte 196,98,125,24,21,90,37,0,0 // vbroadcastss 0x255a(%rip),%ymm10 # 4b50 <_sk_callback_hsw+0x2f8>
+ .byte 196,98,125,24,21,66,38,0,0 // vbroadcastss 0x2642(%rip),%ymm10 # 4d2c <_sk_callback_hsw+0x2f8>
.byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
.byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
- .byte 196,98,125,24,21,77,37,0,0 // vbroadcastss 0x254d(%rip),%ymm10 # 4b54 <_sk_callback_hsw+0x2fc>
+ .byte 196,98,125,24,21,53,38,0,0 // vbroadcastss 0x2635(%rip),%ymm10 # 4d30 <_sk_callback_hsw+0x2fc>
.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
@@ -12116,7 +12177,7 @@ _sk_parametric_g_hsw:
.byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1
- .byte 196,98,125,24,5,36,37,0,0 // vbroadcastss 0x2524(%rip),%ymm8 # 4b58 <_sk_callback_hsw+0x300>
+ .byte 196,98,125,24,5,12,38,0,0 // vbroadcastss 0x260c(%rip),%ymm8 # 4d34 <_sk_callback_hsw+0x300>
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12136,33 +12197,33 @@ _sk_parametric_b_hsw:
.byte 196,66,109,168,211 // vfmadd213ps %ymm11,%ymm2,%ymm10
.byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2
.byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
- .byte 196,98,125,24,37,220,36,0,0 // vbroadcastss 0x24dc(%rip),%ymm12 # 4b5c <_sk_callback_hsw+0x304>
- .byte 196,98,125,24,45,215,36,0,0 // vbroadcastss 0x24d7(%rip),%ymm13 # 4b60 <_sk_callback_hsw+0x308>
+ .byte 196,98,125,24,37,196,37,0,0 // vbroadcastss 0x25c4(%rip),%ymm12 # 4d38 <_sk_callback_hsw+0x304>
+ .byte 196,98,125,24,45,191,37,0,0 // vbroadcastss 0x25bf(%rip),%ymm13 # 4d3c <_sk_callback_hsw+0x308>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,205,36,0,0 // vbroadcastss 0x24cd(%rip),%ymm13 # 4b64 <_sk_callback_hsw+0x30c>
+ .byte 196,98,125,24,45,181,37,0,0 // vbroadcastss 0x25b5(%rip),%ymm13 # 4d40 <_sk_callback_hsw+0x30c>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,195,36,0,0 // vbroadcastss 0x24c3(%rip),%ymm13 # 4b68 <_sk_callback_hsw+0x310>
+ .byte 196,98,125,24,45,171,37,0,0 // vbroadcastss 0x25ab(%rip),%ymm13 # 4d44 <_sk_callback_hsw+0x310>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,185,36,0,0 // vbroadcastss 0x24b9(%rip),%ymm11 # 4b6c <_sk_callback_hsw+0x314>
+ .byte 196,98,125,24,29,161,37,0,0 // vbroadcastss 0x25a1(%rip),%ymm11 # 4d48 <_sk_callback_hsw+0x314>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,175,36,0,0 // vbroadcastss 0x24af(%rip),%ymm12 # 4b70 <_sk_callback_hsw+0x318>
+ .byte 196,98,125,24,37,151,37,0,0 // vbroadcastss 0x2597(%rip),%ymm12 # 4d4c <_sk_callback_hsw+0x318>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,165,36,0,0 // vbroadcastss 0x24a5(%rip),%ymm12 # 4b74 <_sk_callback_hsw+0x31c>
+ .byte 196,98,125,24,37,141,37,0,0 // vbroadcastss 0x258d(%rip),%ymm12 # 4d50 <_sk_callback_hsw+0x31c>
.byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
.byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2
.byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10
.byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10
- .byte 196,98,125,24,29,134,36,0,0 // vbroadcastss 0x2486(%rip),%ymm11 # 4b78 <_sk_callback_hsw+0x320>
+ .byte 196,98,125,24,29,110,37,0,0 // vbroadcastss 0x256e(%rip),%ymm11 # 4d54 <_sk_callback_hsw+0x320>
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,124,36,0,0 // vbroadcastss 0x247c(%rip),%ymm11 # 4b7c <_sk_callback_hsw+0x324>
+ .byte 196,98,125,24,29,100,37,0,0 // vbroadcastss 0x2564(%rip),%ymm11 # 4d58 <_sk_callback_hsw+0x324>
.byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11
- .byte 196,226,125,24,21,114,36,0,0 // vbroadcastss 0x2472(%rip),%ymm2 # 4b80 <_sk_callback_hsw+0x328>
+ .byte 196,226,125,24,21,90,37,0,0 // vbroadcastss 0x255a(%rip),%ymm2 # 4d5c <_sk_callback_hsw+0x328>
.byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2
- .byte 196,98,125,24,21,104,36,0,0 // vbroadcastss 0x2468(%rip),%ymm10 # 4b84 <_sk_callback_hsw+0x32c>
+ .byte 196,98,125,24,21,80,37,0,0 // vbroadcastss 0x2550(%rip),%ymm10 # 4d60 <_sk_callback_hsw+0x32c>
.byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2
.byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
- .byte 196,98,125,24,21,91,36,0,0 // vbroadcastss 0x245b(%rip),%ymm10 # 4b88 <_sk_callback_hsw+0x330>
+ .byte 196,98,125,24,21,67,37,0,0 // vbroadcastss 0x2543(%rip),%ymm10 # 4d64 <_sk_callback_hsw+0x330>
.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
@@ -12170,7 +12231,7 @@ _sk_parametric_b_hsw:
.byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,50,36,0,0 // vbroadcastss 0x2432(%rip),%ymm8 # 4b8c <_sk_callback_hsw+0x334>
+ .byte 196,98,125,24,5,26,37,0,0 // vbroadcastss 0x251a(%rip),%ymm8 # 4d68 <_sk_callback_hsw+0x334>
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12190,33 +12251,33 @@ _sk_parametric_a_hsw:
.byte 196,66,101,168,211 // vfmadd213ps %ymm11,%ymm3,%ymm10
.byte 196,226,125,24,24 // vbroadcastss (%rax),%ymm3
.byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
- .byte 196,98,125,24,37,234,35,0,0 // vbroadcastss 0x23ea(%rip),%ymm12 # 4b90 <_sk_callback_hsw+0x338>
- .byte 196,98,125,24,45,229,35,0,0 // vbroadcastss 0x23e5(%rip),%ymm13 # 4b94 <_sk_callback_hsw+0x33c>
+ .byte 196,98,125,24,37,210,36,0,0 // vbroadcastss 0x24d2(%rip),%ymm12 # 4d6c <_sk_callback_hsw+0x338>
+ .byte 196,98,125,24,45,205,36,0,0 // vbroadcastss 0x24cd(%rip),%ymm13 # 4d70 <_sk_callback_hsw+0x33c>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,219,35,0,0 // vbroadcastss 0x23db(%rip),%ymm13 # 4b98 <_sk_callback_hsw+0x340>
+ .byte 196,98,125,24,45,195,36,0,0 // vbroadcastss 0x24c3(%rip),%ymm13 # 4d74 <_sk_callback_hsw+0x340>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,209,35,0,0 // vbroadcastss 0x23d1(%rip),%ymm13 # 4b9c <_sk_callback_hsw+0x344>
+ .byte 196,98,125,24,45,185,36,0,0 // vbroadcastss 0x24b9(%rip),%ymm13 # 4d78 <_sk_callback_hsw+0x344>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,199,35,0,0 // vbroadcastss 0x23c7(%rip),%ymm11 # 4ba0 <_sk_callback_hsw+0x348>
+ .byte 196,98,125,24,29,175,36,0,0 // vbroadcastss 0x24af(%rip),%ymm11 # 4d7c <_sk_callback_hsw+0x348>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,189,35,0,0 // vbroadcastss 0x23bd(%rip),%ymm12 # 4ba4 <_sk_callback_hsw+0x34c>
+ .byte 196,98,125,24,37,165,36,0,0 // vbroadcastss 0x24a5(%rip),%ymm12 # 4d80 <_sk_callback_hsw+0x34c>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,179,35,0,0 // vbroadcastss 0x23b3(%rip),%ymm12 # 4ba8 <_sk_callback_hsw+0x350>
+ .byte 196,98,125,24,37,155,36,0,0 // vbroadcastss 0x249b(%rip),%ymm12 # 4d84 <_sk_callback_hsw+0x350>
.byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
.byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3
.byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10
.byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10
- .byte 196,98,125,24,29,148,35,0,0 // vbroadcastss 0x2394(%rip),%ymm11 # 4bac <_sk_callback_hsw+0x354>
+ .byte 196,98,125,24,29,124,36,0,0 // vbroadcastss 0x247c(%rip),%ymm11 # 4d88 <_sk_callback_hsw+0x354>
.byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,138,35,0,0 // vbroadcastss 0x238a(%rip),%ymm11 # 4bb0 <_sk_callback_hsw+0x358>
+ .byte 196,98,125,24,29,114,36,0,0 // vbroadcastss 0x2472(%rip),%ymm11 # 4d8c <_sk_callback_hsw+0x358>
.byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11
- .byte 196,226,125,24,29,128,35,0,0 // vbroadcastss 0x2380(%rip),%ymm3 # 4bb4 <_sk_callback_hsw+0x35c>
+ .byte 196,226,125,24,29,104,36,0,0 // vbroadcastss 0x2468(%rip),%ymm3 # 4d90 <_sk_callback_hsw+0x35c>
.byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3
- .byte 196,98,125,24,21,118,35,0,0 // vbroadcastss 0x2376(%rip),%ymm10 # 4bb8 <_sk_callback_hsw+0x360>
+ .byte 196,98,125,24,21,94,36,0,0 // vbroadcastss 0x245e(%rip),%ymm10 # 4d94 <_sk_callback_hsw+0x360>
.byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3
.byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
- .byte 196,98,125,24,21,105,35,0,0 // vbroadcastss 0x2369(%rip),%ymm10 # 4bbc <_sk_callback_hsw+0x364>
+ .byte 196,98,125,24,21,81,36,0,0 // vbroadcastss 0x2451(%rip),%ymm10 # 4d98 <_sk_callback_hsw+0x364>
.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
@@ -12224,7 +12285,7 @@ _sk_parametric_a_hsw:
.byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3
- .byte 196,98,125,24,5,64,35,0,0 // vbroadcastss 0x2340(%rip),%ymm8 # 4bc0 <_sk_callback_hsw+0x368>
+ .byte 196,98,125,24,5,40,36,0,0 // vbroadcastss 0x2428(%rip),%ymm8 # 4d9c <_sk_callback_hsw+0x368>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12233,26 +12294,26 @@ HIDDEN _sk_lab_to_xyz_hsw
.globl _sk_lab_to_xyz_hsw
FUNCTION(_sk_lab_to_xyz_hsw)
_sk_lab_to_xyz_hsw:
- .byte 196,98,125,24,5,50,35,0,0 // vbroadcastss 0x2332(%rip),%ymm8 # 4bc4 <_sk_callback_hsw+0x36c>
- .byte 196,98,125,24,13,45,35,0,0 // vbroadcastss 0x232d(%rip),%ymm9 # 4bc8 <_sk_callback_hsw+0x370>
- .byte 196,98,125,24,21,40,35,0,0 // vbroadcastss 0x2328(%rip),%ymm10 # 4bcc <_sk_callback_hsw+0x374>
+ .byte 196,98,125,24,5,26,36,0,0 // vbroadcastss 0x241a(%rip),%ymm8 # 4da0 <_sk_callback_hsw+0x36c>
+ .byte 196,98,125,24,13,21,36,0,0 // vbroadcastss 0x2415(%rip),%ymm9 # 4da4 <_sk_callback_hsw+0x370>
+ .byte 196,98,125,24,21,16,36,0,0 // vbroadcastss 0x2410(%rip),%ymm10 # 4da8 <_sk_callback_hsw+0x374>
.byte 196,194,53,168,202 // vfmadd213ps %ymm10,%ymm9,%ymm1
.byte 196,194,53,168,210 // vfmadd213ps %ymm10,%ymm9,%ymm2
- .byte 196,98,125,24,13,25,35,0,0 // vbroadcastss 0x2319(%rip),%ymm9 # 4bd0 <_sk_callback_hsw+0x378>
+ .byte 196,98,125,24,13,1,36,0,0 // vbroadcastss 0x2401(%rip),%ymm9 # 4dac <_sk_callback_hsw+0x378>
.byte 196,66,125,184,200 // vfmadd231ps %ymm8,%ymm0,%ymm9
- .byte 196,226,125,24,5,15,35,0,0 // vbroadcastss 0x230f(%rip),%ymm0 # 4bd4 <_sk_callback_hsw+0x37c>
+ .byte 196,226,125,24,5,247,35,0,0 // vbroadcastss 0x23f7(%rip),%ymm0 # 4db0 <_sk_callback_hsw+0x37c>
.byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
- .byte 196,98,125,24,5,6,35,0,0 // vbroadcastss 0x2306(%rip),%ymm8 # 4bd8 <_sk_callback_hsw+0x380>
+ .byte 196,98,125,24,5,238,35,0,0 // vbroadcastss 0x23ee(%rip),%ymm8 # 4db4 <_sk_callback_hsw+0x380>
.byte 196,98,117,168,192 // vfmadd213ps %ymm0,%ymm1,%ymm8
- .byte 196,98,125,24,13,252,34,0,0 // vbroadcastss 0x22fc(%rip),%ymm9 # 4bdc <_sk_callback_hsw+0x384>
+ .byte 196,98,125,24,13,228,35,0,0 // vbroadcastss 0x23e4(%rip),%ymm9 # 4db8 <_sk_callback_hsw+0x384>
.byte 196,98,109,172,200 // vfnmadd213ps %ymm0,%ymm2,%ymm9
.byte 196,193,60,89,200 // vmulps %ymm8,%ymm8,%ymm1
.byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
- .byte 196,226,125,24,21,233,34,0,0 // vbroadcastss 0x22e9(%rip),%ymm2 # 4be0 <_sk_callback_hsw+0x388>
+ .byte 196,226,125,24,21,209,35,0,0 // vbroadcastss 0x23d1(%rip),%ymm2 # 4dbc <_sk_callback_hsw+0x388>
.byte 197,108,194,209,1 // vcmpltps %ymm1,%ymm2,%ymm10
- .byte 196,98,125,24,29,223,34,0,0 // vbroadcastss 0x22df(%rip),%ymm11 # 4be4 <_sk_callback_hsw+0x38c>
+ .byte 196,98,125,24,29,199,35,0,0 // vbroadcastss 0x23c7(%rip),%ymm11 # 4dc0 <_sk_callback_hsw+0x38c>
.byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
- .byte 196,98,125,24,37,213,34,0,0 // vbroadcastss 0x22d5(%rip),%ymm12 # 4be8 <_sk_callback_hsw+0x390>
+ .byte 196,98,125,24,37,189,35,0,0 // vbroadcastss 0x23bd(%rip),%ymm12 # 4dc4 <_sk_callback_hsw+0x390>
.byte 196,65,60,89,196 // vmulps %ymm12,%ymm8,%ymm8
.byte 196,99,61,74,193,160 // vblendvps %ymm10,%ymm1,%ymm8,%ymm8
.byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1
@@ -12267,9 +12328,9 @@ _sk_lab_to_xyz_hsw:
.byte 196,65,52,88,203 // vaddps %ymm11,%ymm9,%ymm9
.byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9
.byte 196,227,53,74,208,32 // vblendvps %ymm2,%ymm0,%ymm9,%ymm2
- .byte 196,226,125,24,5,138,34,0,0 // vbroadcastss 0x228a(%rip),%ymm0 # 4bec <_sk_callback_hsw+0x394>
+ .byte 196,226,125,24,5,114,35,0,0 // vbroadcastss 0x2372(%rip),%ymm0 # 4dc8 <_sk_callback_hsw+0x394>
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,129,34,0,0 // vbroadcastss 0x2281(%rip),%ymm8 # 4bf0 <_sk_callback_hsw+0x398>
+ .byte 196,98,125,24,5,105,35,0,0 // vbroadcastss 0x2369(%rip),%ymm8 # 4dcc <_sk_callback_hsw+0x398>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12278,37 +12339,68 @@ HIDDEN _sk_load_a8_hsw
.globl _sk_load_a8_hsw
FUNCTION(_sk_load_a8_hsw)
_sk_load_a8_hsw:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,45 // jne 29b5 <_sk_load_a8_hsw+0x3d>
- .byte 197,250,126,0 // vmovq (%rax),%xmm0
- .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
+ .byte 117,52 // jne 2aaa <_sk_load_a8_hsw+0x3e>
+ .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0
+ .byte 197,249,219,5,28,39,0,0 // vpand 0x271c(%rip),%xmm0,%xmm0 # 51a0 <_sk_callback_hsw+0x76c>
+ .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,86,34,0,0 // vbroadcastss 0x2256(%rip),%ymm1 # 4bf4 <_sk_callback_hsw+0x39c>
+ .byte 196,226,125,24,13,58,35,0,0 // vbroadcastss 0x233a(%rip),%ymm1 # 4dd0 <_sk_callback_hsw+0x39c>
.byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
.byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
.byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
- .byte 76,137,201 // mov %r9,%rcx
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 29be <_sk_load_a8_hsw+0x46>
- .byte 196,193,249,110,194 // vmovq %r10,%xmm0
- .byte 91 // pop %rbx
- .byte 235,177 // jmp 298c <_sk_load_a8_hsw+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,190 // ja 2a7c <_sk_load_a8_hsw+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,95,0,0,0 // lea 0x5f(%rip),%r10 # 2b28 <_sk_load_a8_hsw+0xbc>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,2 // vpinsrw $0x2,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,1 // vpinsrw $0x1,%eax,%xmm0,%xmm0
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,249,196,192,0 // vpinsrw $0x0,%eax,%xmm0,%xmm0
+ .byte 233,85,255,255,255 // jmpq 2a7c <_sk_load_a8_hsw+0x10>
+ .byte 144 // nop
+ .byte 240,255 // lock (bad)
+ .byte 255 // (bad)
+ .byte 255,229 // jmpq *%rbp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 218,255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,207 // dec %edi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,196 // inc %esp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 185,255,255,255,170 // mov $0xaaffffff,%ecx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_gather_a8_hsw
.globl _sk_gather_a8_hsw
@@ -12350,7 +12442,7 @@ _sk_gather_a8_hsw:
.byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%xmm0
.byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,103,33,0,0 // vbroadcastss 0x2167(%rip),%ymm1 # 4bf8 <_sk_callback_hsw+0x3a0>
+ .byte 196,226,125,24,13,218,33,0,0 // vbroadcastss 0x21da(%rip),%ymm1 # 4dd4 <_sk_callback_hsw+0x3a0>
.byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
@@ -12364,14 +12456,14 @@ FUNCTION(_sk_store_a8_hsw)
_sk_store_a8_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
- .byte 196,98,125,24,5,73,33,0,0 // vbroadcastss 0x2149(%rip),%ymm8 # 4bfc <_sk_callback_hsw+0x3a4>
+ .byte 196,98,125,24,5,188,33,0,0 // vbroadcastss 0x21bc(%rip),%ymm8 # 4dd8 <_sk_callback_hsw+0x3a4>
.byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8
.byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
.byte 77,133,192 // test %r8,%r8
- .byte 117,10 // jne 2adc <_sk_store_a8_hsw+0x37>
+ .byte 117,10 // jne 2c45 <_sk_store_a8_hsw+0x37>
.byte 196,65,123,17,4,19 // vmovsd %xmm8,(%r11,%rdx,1)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12379,10 +12471,10 @@ _sk_store_a8_hsw:
.byte 65,128,225,7 // and $0x7,%r9b
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,236 // ja 2ad8 <_sk_store_a8_hsw+0x33>
+ .byte 119,236 // ja 2c41 <_sk_store_a8_hsw+0x33>
.byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,68,0,0,0 // lea 0x44(%rip),%r10 # 2b40 <_sk_store_a8_hsw+0x9b>
+ .byte 76,141,21,67,0,0,0 // lea 0x43(%rip),%r10 # 2ca8 <_sk_store_a8_hsw+0x9a>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -12393,28 +12485,27 @@ _sk_store_a8_hsw:
.byte 196,67,121,20,68,19,2,4 // vpextrb $0x4,%xmm8,0x2(%r11,%rdx,1)
.byte 196,67,121,20,68,19,1,2 // vpextrb $0x2,%xmm8,0x1(%r11,%rdx,1)
.byte 196,67,121,20,4,19,0 // vpextrb $0x0,%xmm8,(%r11,%rdx,1)
- .byte 235,154 // jmp 2ad8 <_sk_store_a8_hsw+0x33>
- .byte 102,144 // xchg %ax,%ax
- .byte 245 // cmc
- .byte 255 // (bad)
+ .byte 235,154 // jmp 2c41 <_sk_store_a8_hsw+0x33>
+ .byte 144 // nop
+ .byte 246,255 // idiv %bh
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 237 // in (%dx),%eax
+ .byte 238 // out %al,(%dx)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,229 // jmpq *%rbp
+ .byte 255,230 // jmpq *%rsi
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 221,255 // (bad)
+ .byte 222,255 // fdivrp %st,%st(7)
.byte 255 // (bad)
- .byte 255,213 // callq *%rbp
+ .byte 255,214 // callq *%rsi
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,205 // dec %ebp
+ .byte 255,206 // dec %esi
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,197 // inc %ebp
+ .byte 255,198 // inc %esi
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -12423,37 +12514,68 @@ HIDDEN _sk_load_g8_hsw
.globl _sk_load_g8_hsw
FUNCTION(_sk_load_g8_hsw)
_sk_load_g8_hsw:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,50 // jne 2b9e <_sk_load_g8_hsw+0x42>
- .byte 197,250,126,0 // vmovq (%rax),%xmm0
- .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
+ .byte 117,57 // jne 2d07 <_sk_load_g8_hsw+0x43>
+ .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0
+ .byte 197,249,219,5,212,36,0,0 // vpand 0x24d4(%rip),%xmm0,%xmm0 # 51b0 <_sk_callback_hsw+0x77c>
+ .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,126,32,0,0 // vbroadcastss 0x207e(%rip),%ymm1 # 4c00 <_sk_callback_hsw+0x3a8>
+ .byte 196,226,125,24,13,238,32,0,0 // vbroadcastss 0x20ee(%rip),%ymm1 # 4ddc <_sk_callback_hsw+0x3a8>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,115,32,0,0 // vbroadcastss 0x2073(%rip),%ymm3 # 4c04 <_sk_callback_hsw+0x3ac>
- .byte 76,137,201 // mov %r9,%rcx
+ .byte 196,226,125,24,29,227,32,0,0 // vbroadcastss 0x20e3(%rip),%ymm3 # 4de0 <_sk_callback_hsw+0x3ac>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 2ba7 <_sk_load_g8_hsw+0x4b>
- .byte 196,193,249,110,194 // vmovq %r10,%xmm0
- .byte 91 // pop %rbx
- .byte 235,172 // jmp 2b70 <_sk_load_g8_hsw+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,185 // ja 2cd4 <_sk_load_g8_hsw+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,94,0,0,0 // lea 0x5e(%rip),%r10 # 2d84 <_sk_load_g8_hsw+0xc0>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,2 // vpinsrw $0x2,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,1 // vpinsrw $0x1,%eax,%xmm0,%xmm0
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,249,196,192,0 // vpinsrw $0x0,%eax,%xmm0,%xmm0
+ .byte 233,80,255,255,255 // jmpq 2cd4 <_sk_load_g8_hsw+0x10>
+ .byte 241 // icebp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,230 // jmpq *%rsi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 219,255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,208 // callq *%rax
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,197 // inc %ebp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 186,255,255,255,171 // mov $0xabffffff,%edx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_gather_g8_hsw
.globl _sk_gather_g8_hsw
@@ -12495,10 +12617,10 @@ _sk_gather_g8_hsw:
.byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%xmm0
.byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,142,31,0,0 // vbroadcastss 0x1f8e(%rip),%ymm1 # 4c08 <_sk_callback_hsw+0x3b0>
+ .byte 196,226,125,24,13,142,31,0,0 // vbroadcastss 0x1f8e(%rip),%ymm1 # 4de4 <_sk_callback_hsw+0x3b0>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,131,31,0,0 // vbroadcastss 0x1f83(%rip),%ymm3 # 4c0c <_sk_callback_hsw+0x3b4>
+ .byte 196,226,125,24,29,131,31,0,0 // vbroadcastss 0x1f83(%rip),%ymm3 # 4de8 <_sk_callback_hsw+0x3b4>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 255,224 // jmpq *%rax
@@ -12510,9 +12632,9 @@ _sk_gather_i8_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,137,193 // mov %rax,%r9
.byte 77,133,201 // test %r9,%r9
- .byte 116,5 // je 2ca2 <_sk_gather_i8_hsw+0xf>
+ .byte 116,5 // je 2e7e <_sk_gather_i8_hsw+0xf>
.byte 76,137,200 // mov %r9,%rax
- .byte 235,2 // jmp 2ca4 <_sk_gather_i8_hsw+0x11>
+ .byte 235,2 // jmp 2e80 <_sk_gather_i8_hsw+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 83 // push %rbx
.byte 76,139,16 // mov (%rax),%r10
@@ -12546,14 +12668,14 @@ _sk_gather_i8_hsw:
.byte 73,139,65,8 // mov 0x8(%r9),%rax
.byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
.byte 196,226,117,144,28,128 // vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3
- .byte 197,229,219,5,55,33,0,0 // vpand 0x2137(%rip),%ymm3,%ymm0 # 4e80 <_sk_callback_hsw+0x628>
+ .byte 197,229,219,5,27,33,0,0 // vpand 0x211b(%rip),%ymm3,%ymm0 # 5040 <_sk_callback_hsw+0x60c>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,186,30,0,0 // vbroadcastss 0x1eba(%rip),%ymm8 # 4c10 <_sk_callback_hsw+0x3b8>
+ .byte 196,98,125,24,5,186,30,0,0 // vbroadcastss 0x1eba(%rip),%ymm8 # 4dec <_sk_callback_hsw+0x3b8>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,60,33,0,0 // vpshufb 0x213c(%rip),%ymm3,%ymm1 # 4ea0 <_sk_callback_hsw+0x648>
+ .byte 196,226,101,0,13,32,33,0,0 // vpshufb 0x2120(%rip),%ymm3,%ymm1 # 5060 <_sk_callback_hsw+0x62c>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,74,33,0,0 // vpshufb 0x214a(%rip),%ymm3,%ymm2 # 4ec0 <_sk_callback_hsw+0x668>
+ .byte 196,226,101,0,21,46,33,0,0 // vpshufb 0x212e(%rip),%ymm3,%ymm2 # 5080 <_sk_callback_hsw+0x64c>
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
@@ -12570,35 +12692,35 @@ _sk_load_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,114 // jne 2e0e <_sk_load_565_hsw+0x7c>
+ .byte 117,114 // jne 2fea <_sk_load_565_hsw+0x7c>
.byte 196,193,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm0
.byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2
- .byte 196,226,125,88,5,100,30,0,0 // vpbroadcastd 0x1e64(%rip),%ymm0 # 4c14 <_sk_callback_hsw+0x3bc>
+ .byte 196,226,125,88,5,100,30,0,0 // vpbroadcastd 0x1e64(%rip),%ymm0 # 4df0 <_sk_callback_hsw+0x3bc>
.byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,87,30,0,0 // vbroadcastss 0x1e57(%rip),%ymm1 # 4c18 <_sk_callback_hsw+0x3c0>
+ .byte 196,226,125,24,13,87,30,0,0 // vbroadcastss 0x1e57(%rip),%ymm1 # 4df4 <_sk_callback_hsw+0x3c0>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,78,30,0,0 // vpbroadcastd 0x1e4e(%rip),%ymm1 # 4c1c <_sk_callback_hsw+0x3c4>
+ .byte 196,226,125,88,13,78,30,0,0 // vpbroadcastd 0x1e4e(%rip),%ymm1 # 4df8 <_sk_callback_hsw+0x3c4>
.byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,65,30,0,0 // vbroadcastss 0x1e41(%rip),%ymm3 # 4c20 <_sk_callback_hsw+0x3c8>
+ .byte 196,226,125,24,29,65,30,0,0 // vbroadcastss 0x1e41(%rip),%ymm3 # 4dfc <_sk_callback_hsw+0x3c8>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,88,29,56,30,0,0 // vpbroadcastd 0x1e38(%rip),%ymm3 # 4c24 <_sk_callback_hsw+0x3cc>
+ .byte 196,226,125,88,29,56,30,0,0 // vpbroadcastd 0x1e38(%rip),%ymm3 # 4e00 <_sk_callback_hsw+0x3cc>
.byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,43,30,0,0 // vbroadcastss 0x1e2b(%rip),%ymm3 # 4c28 <_sk_callback_hsw+0x3d0>
+ .byte 196,226,125,24,29,43,30,0,0 // vbroadcastss 0x1e2b(%rip),%ymm3 # 4e04 <_sk_callback_hsw+0x3d0>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,32,30,0,0 // vbroadcastss 0x1e20(%rip),%ymm3 # 4c2c <_sk_callback_hsw+0x3d4>
+ .byte 196,226,125,24,29,32,30,0,0 // vbroadcastss 0x1e20(%rip),%ymm3 # 4e08 <_sk_callback_hsw+0x3d4>
.byte 255,224 // jmpq *%rax
.byte 69,137,193 // mov %r8d,%r9d
.byte 65,128,225,7 // and $0x7,%r9b
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,128 // ja 2da2 <_sk_load_565_hsw+0x10>
+ .byte 119,128 // ja 2f7e <_sk_load_565_hsw+0x10>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,75,0,0,0 // lea 0x4b(%rip),%r10 # 2e78 <_sk_load_565_hsw+0xe6>
+ .byte 76,141,21,75,0,0,0 // lea 0x4b(%rip),%r10 # 3054 <_sk_load_565_hsw+0xe6>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -12610,7 +12732,7 @@ _sk_load_565_hsw:
.byte 196,193,121,196,68,83,4,2 // vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,68,83,2,1 // vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,4,83,0 // vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- .byte 233,44,255,255,255 // jmpq 2da2 <_sk_load_565_hsw+0x10>
+ .byte 233,44,255,255,255 // jmpq 2f7e <_sk_load_565_hsw+0x10>
.byte 102,144 // xchg %ax,%ax
.byte 242,255 // repnz (bad)
.byte 255 // (bad)
@@ -12677,23 +12799,23 @@ _sk_gather_565_hsw:
.byte 67,15,183,4,89 // movzwl (%r9,%r11,2),%eax
.byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0
.byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2
- .byte 196,226,125,88,5,235,28,0,0 // vpbroadcastd 0x1ceb(%rip),%ymm0 # 4c30 <_sk_callback_hsw+0x3d8>
+ .byte 196,226,125,88,5,235,28,0,0 // vpbroadcastd 0x1ceb(%rip),%ymm0 # 4e0c <_sk_callback_hsw+0x3d8>
.byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,222,28,0,0 // vbroadcastss 0x1cde(%rip),%ymm1 # 4c34 <_sk_callback_hsw+0x3dc>
+ .byte 196,226,125,24,13,222,28,0,0 // vbroadcastss 0x1cde(%rip),%ymm1 # 4e10 <_sk_callback_hsw+0x3dc>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,213,28,0,0 // vpbroadcastd 0x1cd5(%rip),%ymm1 # 4c38 <_sk_callback_hsw+0x3e0>
+ .byte 196,226,125,88,13,213,28,0,0 // vpbroadcastd 0x1cd5(%rip),%ymm1 # 4e14 <_sk_callback_hsw+0x3e0>
.byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,200,28,0,0 // vbroadcastss 0x1cc8(%rip),%ymm3 # 4c3c <_sk_callback_hsw+0x3e4>
+ .byte 196,226,125,24,29,200,28,0,0 // vbroadcastss 0x1cc8(%rip),%ymm3 # 4e18 <_sk_callback_hsw+0x3e4>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,88,29,191,28,0,0 // vpbroadcastd 0x1cbf(%rip),%ymm3 # 4c40 <_sk_callback_hsw+0x3e8>
+ .byte 196,226,125,88,29,191,28,0,0 // vpbroadcastd 0x1cbf(%rip),%ymm3 # 4e1c <_sk_callback_hsw+0x3e8>
.byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,178,28,0,0 // vbroadcastss 0x1cb2(%rip),%ymm3 # 4c44 <_sk_callback_hsw+0x3ec>
+ .byte 196,226,125,24,29,178,28,0,0 // vbroadcastss 0x1cb2(%rip),%ymm3 # 4e20 <_sk_callback_hsw+0x3ec>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,167,28,0,0 // vbroadcastss 0x1ca7(%rip),%ymm3 # 4c48 <_sk_callback_hsw+0x3f0>
+ .byte 196,226,125,24,29,167,28,0,0 // vbroadcastss 0x1ca7(%rip),%ymm3 # 4e24 <_sk_callback_hsw+0x3f0>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_565_hsw
@@ -12702,11 +12824,11 @@ FUNCTION(_sk_store_565_hsw)
_sk_store_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
- .byte 196,98,125,24,5,155,28,0,0 // vbroadcastss 0x1c9b(%rip),%ymm8 # 4c4c <_sk_callback_hsw+0x3f4>
+ .byte 196,98,125,24,5,155,28,0,0 // vbroadcastss 0x1c9b(%rip),%ymm8 # 4e28 <_sk_callback_hsw+0x3f4>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,193,53,114,241,11 // vpslld $0xb,%ymm9,%ymm9
- .byte 196,98,125,24,21,134,28,0,0 // vbroadcastss 0x1c86(%rip),%ymm10 # 4c50 <_sk_callback_hsw+0x3f8>
+ .byte 196,98,125,24,21,134,28,0,0 // vbroadcastss 0x1c86(%rip),%ymm10 # 4e2c <_sk_callback_hsw+0x3f8>
.byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10
.byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
.byte 196,193,45,114,242,5 // vpslld $0x5,%ymm10,%ymm10
@@ -12717,7 +12839,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,10 // jne 3008 <_sk_store_565_hsw+0x65>
+ .byte 117,10 // jne 31e4 <_sk_store_565_hsw+0x65>
.byte 196,65,122,127,4,83 // vmovdqu %xmm8,(%r11,%rdx,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12725,9 +12847,9 @@ _sk_store_565_hsw:
.byte 65,128,225,7 // and $0x7,%r9b
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,236 // ja 3004 <_sk_store_565_hsw+0x61>
+ .byte 119,236 // ja 31e0 <_sk_store_565_hsw+0x61>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,69,0,0,0 // lea 0x45(%rip),%r10 # 3068 <_sk_store_565_hsw+0xc5>
+ .byte 76,141,21,69,0,0,0 // lea 0x45(%rip),%r10 # 3244 <_sk_store_565_hsw+0xc5>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -12738,7 +12860,7 @@ _sk_store_565_hsw:
.byte 196,67,121,21,68,83,4,2 // vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
.byte 196,67,121,21,68,83,2,1 // vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
.byte 196,67,121,21,4,83,0 // vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- .byte 235,159 // jmp 3004 <_sk_store_565_hsw+0x61>
+ .byte 235,159 // jmp 31e0 <_sk_store_565_hsw+0x61>
.byte 15,31,0 // nopl (%rax)
.byte 244 // hlt
.byte 255 // (bad)
@@ -12771,28 +12893,28 @@ _sk_load_4444_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,138,0,0,0 // jne 311c <_sk_load_4444_hsw+0x98>
+ .byte 15,133,138,0,0,0 // jne 32f8 <_sk_load_4444_hsw+0x98>
.byte 196,193,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm0
.byte 196,226,125,51,216 // vpmovzxwd %xmm0,%ymm3
- .byte 196,226,125,88,5,174,27,0,0 // vpbroadcastd 0x1bae(%rip),%ymm0 # 4c54 <_sk_callback_hsw+0x3fc>
+ .byte 196,226,125,88,5,174,27,0,0 // vpbroadcastd 0x1bae(%rip),%ymm0 # 4e30 <_sk_callback_hsw+0x3fc>
.byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,161,27,0,0 // vbroadcastss 0x1ba1(%rip),%ymm1 # 4c58 <_sk_callback_hsw+0x400>
+ .byte 196,226,125,24,13,161,27,0,0 // vbroadcastss 0x1ba1(%rip),%ymm1 # 4e34 <_sk_callback_hsw+0x400>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,152,27,0,0 // vpbroadcastd 0x1b98(%rip),%ymm1 # 4c5c <_sk_callback_hsw+0x404>
+ .byte 196,226,125,88,13,152,27,0,0 // vpbroadcastd 0x1b98(%rip),%ymm1 # 4e38 <_sk_callback_hsw+0x404>
.byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,139,27,0,0 // vbroadcastss 0x1b8b(%rip),%ymm2 # 4c60 <_sk_callback_hsw+0x408>
+ .byte 196,226,125,24,21,139,27,0,0 // vbroadcastss 0x1b8b(%rip),%ymm2 # 4e3c <_sk_callback_hsw+0x408>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,88,21,130,27,0,0 // vpbroadcastd 0x1b82(%rip),%ymm2 # 4c64 <_sk_callback_hsw+0x40c>
+ .byte 196,226,125,88,21,130,27,0,0 // vpbroadcastd 0x1b82(%rip),%ymm2 # 4e40 <_sk_callback_hsw+0x40c>
.byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,117,27,0,0 // vbroadcastss 0x1b75(%rip),%ymm8 # 4c68 <_sk_callback_hsw+0x410>
+ .byte 196,98,125,24,5,117,27,0,0 // vbroadcastss 0x1b75(%rip),%ymm8 # 4e44 <_sk_callback_hsw+0x410>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,88,5,107,27,0,0 // vpbroadcastd 0x1b6b(%rip),%ymm8 # 4c6c <_sk_callback_hsw+0x414>
+ .byte 196,98,125,88,5,107,27,0,0 // vpbroadcastd 0x1b6b(%rip),%ymm8 # 4e48 <_sk_callback_hsw+0x414>
.byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,93,27,0,0 // vbroadcastss 0x1b5d(%rip),%ymm8 # 4c70 <_sk_callback_hsw+0x418>
+ .byte 196,98,125,24,5,93,27,0,0 // vbroadcastss 0x1b5d(%rip),%ymm8 # 4e4c <_sk_callback_hsw+0x418>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12801,9 +12923,9 @@ _sk_load_4444_hsw:
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 15,135,100,255,255,255 // ja 3098 <_sk_load_4444_hsw+0x14>
+ .byte 15,135,100,255,255,255 // ja 3274 <_sk_load_4444_hsw+0x14>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,73,0,0,0 // lea 0x49(%rip),%r10 # 3188 <_sk_load_4444_hsw+0x104>
+ .byte 76,141,21,73,0,0,0 // lea 0x49(%rip),%r10 # 3364 <_sk_load_4444_hsw+0x104>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -12815,7 +12937,7 @@ _sk_load_4444_hsw:
.byte 196,193,121,196,68,83,4,2 // vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,68,83,2,1 // vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,4,83,0 // vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- .byte 233,16,255,255,255 // jmpq 3098 <_sk_load_4444_hsw+0x14>
+ .byte 233,16,255,255,255 // jmpq 3274 <_sk_load_4444_hsw+0x14>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -12881,25 +13003,25 @@ _sk_gather_4444_hsw:
.byte 67,15,183,4,89 // movzwl (%r9,%r11,2),%eax
.byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0
.byte 196,226,125,51,216 // vpmovzxwd %xmm0,%ymm3
- .byte 196,226,125,88,5,31,26,0,0 // vpbroadcastd 0x1a1f(%rip),%ymm0 # 4c74 <_sk_callback_hsw+0x41c>
+ .byte 196,226,125,88,5,31,26,0,0 // vpbroadcastd 0x1a1f(%rip),%ymm0 # 4e50 <_sk_callback_hsw+0x41c>
.byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,18,26,0,0 // vbroadcastss 0x1a12(%rip),%ymm1 # 4c78 <_sk_callback_hsw+0x420>
+ .byte 196,226,125,24,13,18,26,0,0 // vbroadcastss 0x1a12(%rip),%ymm1 # 4e54 <_sk_callback_hsw+0x420>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,9,26,0,0 // vpbroadcastd 0x1a09(%rip),%ymm1 # 4c7c <_sk_callback_hsw+0x424>
+ .byte 196,226,125,88,13,9,26,0,0 // vpbroadcastd 0x1a09(%rip),%ymm1 # 4e58 <_sk_callback_hsw+0x424>
.byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,252,25,0,0 // vbroadcastss 0x19fc(%rip),%ymm2 # 4c80 <_sk_callback_hsw+0x428>
+ .byte 196,226,125,24,21,252,25,0,0 // vbroadcastss 0x19fc(%rip),%ymm2 # 4e5c <_sk_callback_hsw+0x428>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,88,21,243,25,0,0 // vpbroadcastd 0x19f3(%rip),%ymm2 # 4c84 <_sk_callback_hsw+0x42c>
+ .byte 196,226,125,88,21,243,25,0,0 // vpbroadcastd 0x19f3(%rip),%ymm2 # 4e60 <_sk_callback_hsw+0x42c>
.byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,230,25,0,0 // vbroadcastss 0x19e6(%rip),%ymm8 # 4c88 <_sk_callback_hsw+0x430>
+ .byte 196,98,125,24,5,230,25,0,0 // vbroadcastss 0x19e6(%rip),%ymm8 # 4e64 <_sk_callback_hsw+0x430>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,88,5,220,25,0,0 // vpbroadcastd 0x19dc(%rip),%ymm8 # 4c8c <_sk_callback_hsw+0x434>
+ .byte 196,98,125,88,5,220,25,0,0 // vpbroadcastd 0x19dc(%rip),%ymm8 # 4e68 <_sk_callback_hsw+0x434>
.byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,206,25,0,0 // vbroadcastss 0x19ce(%rip),%ymm8 # 4c90 <_sk_callback_hsw+0x438>
+ .byte 196,98,125,24,5,206,25,0,0 // vbroadcastss 0x19ce(%rip),%ymm8 # 4e6c <_sk_callback_hsw+0x438>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12910,7 +13032,7 @@ FUNCTION(_sk_store_4444_hsw)
_sk_store_4444_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
- .byte 196,98,125,24,5,187,25,0,0 // vbroadcastss 0x19bb(%rip),%ymm8 # 4c94 <_sk_callback_hsw+0x43c>
+ .byte 196,98,125,24,5,187,25,0,0 // vbroadcastss 0x19bb(%rip),%ymm8 # 4e70 <_sk_callback_hsw+0x43c>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,193,53,114,241,12 // vpslld $0xc,%ymm9,%ymm9
@@ -12928,7 +13050,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,10 // jne 333c <_sk_store_4444_hsw+0x71>
+ .byte 117,10 // jne 3518 <_sk_store_4444_hsw+0x71>
.byte 196,65,122,127,4,83 // vmovdqu %xmm8,(%r11,%rdx,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -12936,9 +13058,9 @@ _sk_store_4444_hsw:
.byte 65,128,225,7 // and $0x7,%r9b
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,236 // ja 3338 <_sk_store_4444_hsw+0x6d>
+ .byte 119,236 // ja 3514 <_sk_store_4444_hsw+0x6d>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,69,0,0,0 // lea 0x45(%rip),%r10 # 339c <_sk_store_4444_hsw+0xd1>
+ .byte 76,141,21,69,0,0,0 // lea 0x45(%rip),%r10 # 3578 <_sk_store_4444_hsw+0xd1>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -12949,7 +13071,7 @@ _sk_store_4444_hsw:
.byte 196,67,121,21,68,83,4,2 // vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
.byte 196,67,121,21,68,83,2,1 // vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
.byte 196,67,121,21,4,83,0 // vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- .byte 235,159 // jmp 3338 <_sk_store_4444_hsw+0x6d>
+ .byte 235,159 // jmp 3514 <_sk_store_4444_hsw+0x6d>
.byte 15,31,0 // nopl (%rax)
.byte 244 // hlt
.byte 255 // (bad)
@@ -12984,16 +13106,16 @@ _sk_load_8888_hsw:
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 76,3,16 // add (%rax),%r10
.byte 77,133,192 // test %r8,%r8
- .byte 117,88 // jne 3425 <_sk_load_8888_hsw+0x6d>
+ .byte 117,88 // jne 3601 <_sk_load_8888_hsw+0x6d>
.byte 196,193,124,16,26 // vmovups (%r10),%ymm3
- .byte 197,228,84,5,6,27,0,0 // vandps 0x1b06(%rip),%ymm3,%ymm0 # 4ee0 <_sk_callback_hsw+0x688>
+ .byte 197,228,84,5,234,26,0,0 // vandps 0x1aea(%rip),%ymm3,%ymm0 # 50a0 <_sk_callback_hsw+0x66c>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,177,24,0,0 // vbroadcastss 0x18b1(%rip),%ymm8 # 4c98 <_sk_callback_hsw+0x440>
+ .byte 196,98,125,24,5,177,24,0,0 // vbroadcastss 0x18b1(%rip),%ymm8 # 4e74 <_sk_callback_hsw+0x440>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,11,27,0,0 // vpshufb 0x1b0b(%rip),%ymm3,%ymm1 # 4f00 <_sk_callback_hsw+0x6a8>
+ .byte 196,226,101,0,13,239,26,0,0 // vpshufb 0x1aef(%rip),%ymm3,%ymm1 # 50c0 <_sk_callback_hsw+0x68c>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,25,27,0,0 // vpshufb 0x1b19(%rip),%ymm3,%ymm2 # 4f20 <_sk_callback_hsw+0x6c8>
+ .byte 196,226,101,0,21,253,26,0,0 // vpshufb 0x1afd(%rip),%ymm3,%ymm2 # 50e0 <_sk_callback_hsw+0x6ac>
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
@@ -13010,7 +13132,7 @@ _sk_load_8888_hsw:
.byte 196,225,249,110,192 // vmovq %rax,%xmm0
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
.byte 196,194,125,44,26 // vmaskmovps (%r10),%ymm0,%ymm3
- .byte 235,135 // jmp 33d2 <_sk_load_8888_hsw+0x1a>
+ .byte 235,135 // jmp 35ae <_sk_load_8888_hsw+0x1a>
HIDDEN _sk_gather_8888_hsw
.globl _sk_gather_8888_hsw
@@ -13025,14 +13147,14 @@ _sk_gather_8888_hsw:
.byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
.byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
.byte 196,194,117,144,28,129 // vpgatherdd %ymm1,(%r9,%ymm0,4),%ymm3
- .byte 197,229,219,5,199,26,0,0 // vpand 0x1ac7(%rip),%ymm3,%ymm0 # 4f40 <_sk_callback_hsw+0x6e8>
+ .byte 197,229,219,5,171,26,0,0 // vpand 0x1aab(%rip),%ymm3,%ymm0 # 5100 <_sk_callback_hsw+0x6cc>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,22,24,0,0 // vbroadcastss 0x1816(%rip),%ymm8 # 4c9c <_sk_callback_hsw+0x444>
+ .byte 196,98,125,24,5,22,24,0,0 // vbroadcastss 0x1816(%rip),%ymm8 # 4e78 <_sk_callback_hsw+0x444>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,204,26,0,0 // vpshufb 0x1acc(%rip),%ymm3,%ymm1 # 4f60 <_sk_callback_hsw+0x708>
+ .byte 196,226,101,0,13,176,26,0,0 // vpshufb 0x1ab0(%rip),%ymm3,%ymm1 # 5120 <_sk_callback_hsw+0x6ec>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,218,26,0,0 // vpshufb 0x1ada(%rip),%ymm3,%ymm2 # 4f80 <_sk_callback_hsw+0x728>
+ .byte 196,226,101,0,21,190,26,0,0 // vpshufb 0x1abe(%rip),%ymm3,%ymm2 # 5140 <_sk_callback_hsw+0x70c>
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
@@ -13049,7 +13171,7 @@ _sk_store_8888_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 76,3,16 // add (%rax),%r10
- .byte 196,98,125,24,5,198,23,0,0 // vbroadcastss 0x17c6(%rip),%ymm8 # 4ca0 <_sk_callback_hsw+0x448>
+ .byte 196,98,125,24,5,198,23,0,0 // vbroadcastss 0x17c6(%rip),%ymm8 # 4e7c <_sk_callback_hsw+0x448>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,65,116,89,208 // vmulps %ymm8,%ymm1,%ymm10
@@ -13065,7 +13187,7 @@ _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 3534 <_sk_store_8888_hsw+0x73>
+ .byte 117,12 // jne 3710 <_sk_store_8888_hsw+0x73>
.byte 196,65,124,17,2 // vmovups %ymm8,(%r10)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,201 // mov %r9,%rcx
@@ -13078,7 +13200,7 @@ _sk_store_8888_hsw:
.byte 196,97,249,110,200 // vmovq %rax,%xmm9
.byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9
.byte 196,66,53,46,2 // vmaskmovps %ymm8,%ymm9,(%r10)
- .byte 235,211 // jmp 352d <_sk_store_8888_hsw+0x6c>
+ .byte 235,211 // jmp 3709 <_sk_store_8888_hsw+0x6c>
HIDDEN _sk_load_f16_hsw
.globl _sk_load_f16_hsw
@@ -13087,7 +13209,7 @@ _sk_load_f16_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,97 // jne 35c5 <_sk_load_f16_hsw+0x6b>
+ .byte 117,97 // jne 37a1 <_sk_load_f16_hsw+0x6b>
.byte 197,121,16,4,208 // vmovupd (%rax,%rdx,8),%xmm8
.byte 197,249,16,84,208,16 // vmovupd 0x10(%rax,%rdx,8),%xmm2
.byte 197,249,16,92,208,32 // vmovupd 0x20(%rax,%rdx,8),%xmm3
@@ -13113,29 +13235,29 @@ _sk_load_f16_hsw:
.byte 197,123,16,4,208 // vmovsd (%rax,%rdx,8),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,79 // je 3624 <_sk_load_f16_hsw+0xca>
+ .byte 116,79 // je 3800 <_sk_load_f16_hsw+0xca>
.byte 197,57,22,68,208,8 // vmovhpd 0x8(%rax,%rdx,8),%xmm8,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,67 // jb 3624 <_sk_load_f16_hsw+0xca>
+ .byte 114,67 // jb 3800 <_sk_load_f16_hsw+0xca>
.byte 197,251,16,84,208,16 // vmovsd 0x10(%rax,%rdx,8),%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 116,68 // je 3631 <_sk_load_f16_hsw+0xd7>
+ .byte 116,68 // je 380d <_sk_load_f16_hsw+0xd7>
.byte 197,233,22,84,208,24 // vmovhpd 0x18(%rax,%rdx,8),%xmm2,%xmm2
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,56 // jb 3631 <_sk_load_f16_hsw+0xd7>
+ .byte 114,56 // jb 380d <_sk_load_f16_hsw+0xd7>
.byte 197,251,16,92,208,32 // vmovsd 0x20(%rax,%rdx,8),%xmm3
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 15,132,114,255,255,255 // je 357b <_sk_load_f16_hsw+0x21>
+ .byte 15,132,114,255,255,255 // je 3757 <_sk_load_f16_hsw+0x21>
.byte 197,225,22,92,208,40 // vmovhpd 0x28(%rax,%rdx,8),%xmm3,%xmm3
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 15,130,98,255,255,255 // jb 357b <_sk_load_f16_hsw+0x21>
+ .byte 15,130,98,255,255,255 // jb 3757 <_sk_load_f16_hsw+0x21>
.byte 197,122,126,76,208,48 // vmovq 0x30(%rax,%rdx,8),%xmm9
- .byte 233,87,255,255,255 // jmpq 357b <_sk_load_f16_hsw+0x21>
+ .byte 233,87,255,255,255 // jmpq 3757 <_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 357b <_sk_load_f16_hsw+0x21>
+ .byte 233,74,255,255,255 // jmpq 3757 <_sk_load_f16_hsw+0x21>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,65,255,255,255 // jmpq 357b <_sk_load_f16_hsw+0x21>
+ .byte 233,65,255,255,255 // jmpq 3757 <_sk_load_f16_hsw+0x21>
HIDDEN _sk_gather_f16_hsw
.globl _sk_gather_f16_hsw
@@ -13193,7 +13315,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,27 // jne 3729 <_sk_store_f16_hsw+0x65>
+ .byte 117,27 // jne 3905 <_sk_store_f16_hsw+0x65>
.byte 197,120,17,28,208 // vmovups %xmm11,(%rax,%rdx,8)
.byte 197,120,17,84,208,16 // vmovups %xmm10,0x10(%rax,%rdx,8)
.byte 197,120,17,76,208,32 // vmovups %xmm9,0x20(%rax,%rdx,8)
@@ -13202,22 +13324,22 @@ _sk_store_f16_hsw:
.byte 255,224 // jmpq *%rax
.byte 197,121,214,28,208 // vmovq %xmm11,(%rax,%rdx,8)
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,241 // je 3725 <_sk_store_f16_hsw+0x61>
+ .byte 116,241 // je 3901 <_sk_store_f16_hsw+0x61>
.byte 197,121,23,92,208,8 // vmovhpd %xmm11,0x8(%rax,%rdx,8)
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,229 // jb 3725 <_sk_store_f16_hsw+0x61>
+ .byte 114,229 // jb 3901 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,84,208,16 // vmovq %xmm10,0x10(%rax,%rdx,8)
- .byte 116,221 // je 3725 <_sk_store_f16_hsw+0x61>
+ .byte 116,221 // je 3901 <_sk_store_f16_hsw+0x61>
.byte 197,121,23,84,208,24 // vmovhpd %xmm10,0x18(%rax,%rdx,8)
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,209 // jb 3725 <_sk_store_f16_hsw+0x61>
+ .byte 114,209 // jb 3901 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,76,208,32 // vmovq %xmm9,0x20(%rax,%rdx,8)
- .byte 116,201 // je 3725 <_sk_store_f16_hsw+0x61>
+ .byte 116,201 // je 3901 <_sk_store_f16_hsw+0x61>
.byte 197,121,23,76,208,40 // vmovhpd %xmm9,0x28(%rax,%rdx,8)
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,189 // jb 3725 <_sk_store_f16_hsw+0x61>
+ .byte 114,189 // jb 3901 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,68,208,48 // vmovq %xmm8,0x30(%rax,%rdx,8)
- .byte 235,181 // jmp 3725 <_sk_store_f16_hsw+0x61>
+ .byte 235,181 // jmp 3901 <_sk_store_f16_hsw+0x61>
HIDDEN _sk_load_u16_be_hsw
.globl _sk_load_u16_be_hsw
@@ -13227,7 +13349,7 @@ _sk_load_u16_be_hsw:
.byte 76,139,8 // mov (%rax),%r9
.byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,204,0,0,0 // jne 3852 <_sk_load_u16_be_hsw+0xe2>
+ .byte 15,133,204,0,0,0 // jne 3a2e <_sk_load_u16_be_hsw+0xe2>
.byte 196,65,121,16,4,65 // vmovupd (%r9,%rax,2),%xmm8
.byte 196,193,121,16,84,65,16 // vmovupd 0x10(%r9,%rax,2),%xmm2
.byte 196,193,121,16,92,65,32 // vmovupd 0x20(%r9,%rax,2),%xmm3
@@ -13246,7 +13368,7 @@ _sk_load_u16_be_hsw:
.byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0
.byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,21,189,20,0,0 // vbroadcastss 0x14bd(%rip),%ymm10 # 4ca4 <_sk_callback_hsw+0x44c>
+ .byte 196,98,125,24,21,189,20,0,0 // vbroadcastss 0x14bd(%rip),%ymm10 # 4e80 <_sk_callback_hsw+0x44c>
.byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0
.byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
.byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
@@ -13274,29 +13396,29 @@ _sk_load_u16_be_hsw:
.byte 196,65,123,16,4,65 // vmovsd (%r9,%rax,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,85 // je 38b8 <_sk_load_u16_be_hsw+0x148>
+ .byte 116,85 // je 3a94 <_sk_load_u16_be_hsw+0x148>
.byte 196,65,57,22,68,65,8 // vmovhpd 0x8(%r9,%rax,2),%xmm8,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,72 // jb 38b8 <_sk_load_u16_be_hsw+0x148>
+ .byte 114,72 // jb 3a94 <_sk_load_u16_be_hsw+0x148>
.byte 196,193,123,16,84,65,16 // vmovsd 0x10(%r9,%rax,2),%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 116,72 // je 38c5 <_sk_load_u16_be_hsw+0x155>
+ .byte 116,72 // je 3aa1 <_sk_load_u16_be_hsw+0x155>
.byte 196,193,105,22,84,65,24 // vmovhpd 0x18(%r9,%rax,2),%xmm2,%xmm2
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,59 // jb 38c5 <_sk_load_u16_be_hsw+0x155>
+ .byte 114,59 // jb 3aa1 <_sk_load_u16_be_hsw+0x155>
.byte 196,193,123,16,92,65,32 // vmovsd 0x20(%r9,%rax,2),%xmm3
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 15,132,6,255,255,255 // je 37a1 <_sk_load_u16_be_hsw+0x31>
+ .byte 15,132,6,255,255,255 // je 397d <_sk_load_u16_be_hsw+0x31>
.byte 196,193,97,22,92,65,40 // vmovhpd 0x28(%r9,%rax,2),%xmm3,%xmm3
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 15,130,245,254,255,255 // jb 37a1 <_sk_load_u16_be_hsw+0x31>
+ .byte 15,130,245,254,255,255 // jb 397d <_sk_load_u16_be_hsw+0x31>
.byte 196,65,122,126,76,65,48 // vmovq 0x30(%r9,%rax,2),%xmm9
- .byte 233,233,254,255,255 // jmpq 37a1 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,233,254,255,255 // jmpq 397d <_sk_load_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,220,254,255,255 // jmpq 37a1 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,220,254,255,255 // jmpq 397d <_sk_load_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,211,254,255,255 // jmpq 37a1 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,211,254,255,255 // jmpq 397d <_sk_load_u16_be_hsw+0x31>
HIDDEN _sk_load_rgb_u16_be_hsw
.globl _sk_load_rgb_u16_be_hsw
@@ -13306,7 +13428,7 @@ _sk_load_rgb_u16_be_hsw:
.byte 76,139,8 // mov (%rax),%r9
.byte 72,141,4,82 // lea (%rdx,%rdx,2),%rax
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,204,0,0,0 // jne 39ac <_sk_load_rgb_u16_be_hsw+0xde>
+ .byte 15,133,204,0,0,0 // jne 3b88 <_sk_load_rgb_u16_be_hsw+0xde>
.byte 196,193,122,111,4,65 // vmovdqu (%r9,%rax,2),%xmm0
.byte 196,193,122,111,84,65,12 // vmovdqu 0xc(%r9,%rax,2),%xmm2
.byte 196,193,122,111,76,65,24 // vmovdqu 0x18(%r9,%rax,2),%xmm1
@@ -13330,7 +13452,7 @@ _sk_load_rgb_u16_be_hsw:
.byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0
.byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,21,78,19,0,0 // vbroadcastss 0x134e(%rip),%ymm10 # 4ca8 <_sk_callback_hsw+0x450>
+ .byte 196,98,125,24,21,78,19,0,0 // vbroadcastss 0x134e(%rip),%ymm10 # 4e84 <_sk_callback_hsw+0x450>
.byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0
.byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
.byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
@@ -13347,41 +13469,41 @@ _sk_load_rgb_u16_be_hsw:
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,2,19,0,0 // vbroadcastss 0x1302(%rip),%ymm3 # 4cac <_sk_callback_hsw+0x454>
+ .byte 196,226,125,24,29,2,19,0,0 // vbroadcastss 0x1302(%rip),%ymm3 # 4e88 <_sk_callback_hsw+0x454>
.byte 255,224 // jmpq *%rax
.byte 196,193,121,110,4,65 // vmovd (%r9,%rax,2),%xmm0
.byte 196,193,121,196,68,65,4,2 // vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 117,5 // jne 39c5 <_sk_load_rgb_u16_be_hsw+0xf7>
- .byte 233,79,255,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,5 // jne 3ba1 <_sk_load_rgb_u16_be_hsw+0xf7>
+ .byte 233,79,255,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
.byte 196,193,121,110,76,65,6 // vmovd 0x6(%r9,%rax,2),%xmm1
.byte 196,65,113,196,68,65,10,2 // vpinsrw $0x2,0xa(%r9,%rax,2),%xmm1,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,26 // jb 39f4 <_sk_load_rgb_u16_be_hsw+0x126>
+ .byte 114,26 // jb 3bd0 <_sk_load_rgb_u16_be_hsw+0x126>
.byte 196,193,121,110,76,65,12 // vmovd 0xc(%r9,%rax,2),%xmm1
.byte 196,193,113,196,84,65,16,2 // vpinsrw $0x2,0x10(%r9,%rax,2),%xmm1,%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 117,10 // jne 39f9 <_sk_load_rgb_u16_be_hsw+0x12b>
- .byte 233,32,255,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,27,255,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 3bd5 <_sk_load_rgb_u16_be_hsw+0x12b>
+ .byte 233,32,255,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,27,255,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
.byte 196,193,121,110,76,65,18 // vmovd 0x12(%r9,%rax,2),%xmm1
.byte 196,65,113,196,76,65,22,2 // vpinsrw $0x2,0x16(%r9,%rax,2),%xmm1,%xmm9
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,26 // jb 3a28 <_sk_load_rgb_u16_be_hsw+0x15a>
+ .byte 114,26 // jb 3c04 <_sk_load_rgb_u16_be_hsw+0x15a>
.byte 196,193,121,110,76,65,24 // vmovd 0x18(%r9,%rax,2),%xmm1
.byte 196,193,113,196,76,65,28,2 // vpinsrw $0x2,0x1c(%r9,%rax,2),%xmm1,%xmm1
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 117,10 // jne 3a2d <_sk_load_rgb_u16_be_hsw+0x15f>
- .byte 233,236,254,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,231,254,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 3c09 <_sk_load_rgb_u16_be_hsw+0x15f>
+ .byte 233,236,254,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,231,254,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
.byte 196,193,121,110,92,65,30 // vmovd 0x1e(%r9,%rax,2),%xmm3
.byte 196,65,97,196,92,65,34,2 // vpinsrw $0x2,0x22(%r9,%rax,2),%xmm3,%xmm11
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,20 // jb 3a56 <_sk_load_rgb_u16_be_hsw+0x188>
+ .byte 114,20 // jb 3c32 <_sk_load_rgb_u16_be_hsw+0x188>
.byte 196,193,121,110,92,65,36 // vmovd 0x24(%r9,%rax,2),%xmm3
.byte 196,193,97,196,92,65,40,2 // vpinsrw $0x2,0x28(%r9,%rax,2),%xmm3,%xmm3
- .byte 233,190,254,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,185,254,255,255 // jmpq 3914 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,190,254,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,185,254,255,255 // jmpq 3af0 <_sk_load_rgb_u16_be_hsw+0x46>
HIDDEN _sk_store_u16_be_hsw
.globl _sk_store_u16_be_hsw
@@ -13390,7 +13512,7 @@ _sk_store_u16_be_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,8 // mov (%rax),%r9
.byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax
- .byte 196,98,125,24,5,63,18,0,0 // vbroadcastss 0x123f(%rip),%ymm8 # 4cb0 <_sk_callback_hsw+0x458>
+ .byte 196,98,125,24,5,63,18,0,0 // vbroadcastss 0x123f(%rip),%ymm8 # 4e8c <_sk_callback_hsw+0x458>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10
@@ -13428,7 +13550,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,31 // jne 3b55 <_sk_store_u16_be_hsw+0xfa>
+ .byte 117,31 // jne 3d31 <_sk_store_u16_be_hsw+0xfa>
.byte 196,65,120,17,28,65 // vmovups %xmm11,(%r9,%rax,2)
.byte 196,65,120,17,84,65,16 // vmovups %xmm10,0x10(%r9,%rax,2)
.byte 196,65,120,17,76,65,32 // vmovups %xmm9,0x20(%r9,%rax,2)
@@ -13437,22 +13559,22 @@ _sk_store_u16_be_hsw:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,214,28,65 // vmovq %xmm11,(%r9,%rax,2)
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,240 // je 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,240 // je 3d2d <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,23,92,65,8 // vmovhpd %xmm11,0x8(%r9,%rax,2)
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,227 // jb 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,227 // jb 3d2d <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,84,65,16 // vmovq %xmm10,0x10(%r9,%rax,2)
- .byte 116,218 // je 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,218 // je 3d2d <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,23,84,65,24 // vmovhpd %xmm10,0x18(%r9,%rax,2)
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,205 // jb 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,205 // jb 3d2d <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,76,65,32 // vmovq %xmm9,0x20(%r9,%rax,2)
- .byte 116,196 // je 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,196 // je 3d2d <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,23,76,65,40 // vmovhpd %xmm9,0x28(%r9,%rax,2)
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,183 // jb 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,183 // jb 3d2d <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,68,65,48 // vmovq %xmm8,0x30(%r9,%rax,2)
- .byte 235,174 // jmp 3b51 <_sk_store_u16_be_hsw+0xf6>
+ .byte 235,174 // jmp 3d2d <_sk_store_u16_be_hsw+0xf6>
HIDDEN _sk_load_f32_hsw
.globl _sk_load_f32_hsw
@@ -13460,10 +13582,10 @@ FUNCTION(_sk_load_f32_hsw)
_sk_load_f32_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 119,110 // ja 3c19 <_sk_load_f32_hsw+0x76>
+ .byte 119,110 // ja 3df5 <_sk_load_f32_hsw+0x76>
.byte 76,139,8 // mov (%rax),%r9
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
- .byte 76,141,29,135,0,0,0 // lea 0x87(%rip),%r11 # 3c44 <_sk_load_f32_hsw+0xa1>
+ .byte 76,141,29,135,0,0,0 // lea 0x87(%rip),%r11 # 3e20 <_sk_load_f32_hsw+0xa1>
.byte 75,99,4,131 // movslq (%r11,%r8,4),%rax
.byte 76,1,216 // add %r11,%rax
.byte 255,224 // jmpq *%rax
@@ -13524,7 +13646,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,55 // jne 3cd1 <_sk_store_f32_hsw+0x6d>
+ .byte 117,55 // jne 3ead <_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
@@ -13537,22 +13659,22 @@ _sk_store_f32_hsw:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,17,20,129 // vmovupd %xmm10,(%r9,%rax,4)
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,240 // je 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 116,240 // je 3ea9 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,76,129,16 // vmovupd %xmm9,0x10(%r9,%rax,4)
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,227 // jb 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 114,227 // jb 3ea9 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,68,129,32 // vmovupd %xmm8,0x20(%r9,%rax,4)
- .byte 116,218 // je 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 116,218 // je 3ea9 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,92,129,48 // vmovupd %xmm11,0x30(%r9,%rax,4)
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,205 // jb 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 114,205 // jb 3ea9 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,84,129,64,1 // vextractf128 $0x1,%ymm10,0x40(%r9,%rax,4)
- .byte 116,195 // je 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 116,195 // je 3ea9 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,76,129,80,1 // vextractf128 $0x1,%ymm9,0x50(%r9,%rax,4)
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,181 // jb 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 114,181 // jb 3ea9 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,68,129,96,1 // vextractf128 $0x1,%ymm8,0x60(%r9,%rax,4)
- .byte 235,171 // jmp 3ccd <_sk_store_f32_hsw+0x69>
+ .byte 235,171 // jmp 3ea9 <_sk_store_f32_hsw+0x69>
HIDDEN _sk_clamp_x_hsw
.globl _sk_clamp_x_hsw
@@ -13650,7 +13772,7 @@ FUNCTION(_sk_clamp_x_1_hsw)
_sk_clamp_x_1_hsw:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,146,14,0,0 // vbroadcastss 0xe92(%rip),%ymm8 # 4cb4 <_sk_callback_hsw+0x45c>
+ .byte 196,98,125,24,5,146,14,0,0 // vbroadcastss 0xe92(%rip),%ymm8 # 4e90 <_sk_callback_hsw+0x45c>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -13668,9 +13790,9 @@ HIDDEN _sk_mirror_x_1_hsw
.globl _sk_mirror_x_1_hsw
FUNCTION(_sk_mirror_x_1_hsw)
_sk_mirror_x_1_hsw:
- .byte 196,98,125,24,5,117,14,0,0 // vbroadcastss 0xe75(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x460>
+ .byte 196,98,125,24,5,117,14,0,0 // vbroadcastss 0xe75(%rip),%ymm8 # 4e94 <_sk_callback_hsw+0x460>
.byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,13,107,14,0,0 // vbroadcastss 0xe6b(%rip),%ymm9 # 4cbc <_sk_callback_hsw+0x464>
+ .byte 196,98,125,24,13,107,14,0,0 // vbroadcastss 0xe6b(%rip),%ymm9 # 4e98 <_sk_callback_hsw+0x464>
.byte 196,65,124,89,201 // vmulps %ymm9,%ymm0,%ymm9
.byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9
.byte 196,65,52,88,201 // vaddps %ymm9,%ymm9,%ymm9
@@ -13686,11 +13808,11 @@ HIDDEN _sk_luminance_to_alpha_hsw
.globl _sk_luminance_to_alpha_hsw
FUNCTION(_sk_luminance_to_alpha_hsw)
_sk_luminance_to_alpha_hsw:
- .byte 196,226,125,24,29,59,14,0,0 // vbroadcastss 0xe3b(%rip),%ymm3 # 4cc0 <_sk_callback_hsw+0x468>
- .byte 196,98,125,24,5,54,14,0,0 // vbroadcastss 0xe36(%rip),%ymm8 # 4cc4 <_sk_callback_hsw+0x46c>
+ .byte 196,226,125,24,29,59,14,0,0 // vbroadcastss 0xe3b(%rip),%ymm3 # 4e9c <_sk_callback_hsw+0x468>
+ .byte 196,98,125,24,5,54,14,0,0 // vbroadcastss 0xe36(%rip),%ymm8 # 4ea0 <_sk_callback_hsw+0x46c>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
.byte 196,226,125,184,203 // vfmadd231ps %ymm3,%ymm0,%ymm1
- .byte 196,226,125,24,29,39,14,0,0 // vbroadcastss 0xe27(%rip),%ymm3 # 4cc8 <_sk_callback_hsw+0x470>
+ .byte 196,226,125,24,29,39,14,0,0 // vbroadcastss 0xe27(%rip),%ymm3 # 4ea4 <_sk_callback_hsw+0x470>
.byte 196,226,109,168,217 // vfmadd213ps %ymm1,%ymm2,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
@@ -13864,9 +13986,9 @@ _sk_evenly_spaced_gradient_hsw:
.byte 76,139,72,8 // mov 0x8(%rax),%r9
.byte 77,137,211 // mov %r10,%r11
.byte 73,255,203 // dec %r11
- .byte 120,7 // js 415a <_sk_evenly_spaced_gradient_hsw+0x19>
+ .byte 120,7 // js 4336 <_sk_evenly_spaced_gradient_hsw+0x19>
.byte 196,193,242,42,203 // vcvtsi2ss %r11,%xmm1,%xmm1
- .byte 235,22 // jmp 4170 <_sk_evenly_spaced_gradient_hsw+0x2f>
+ .byte 235,22 // jmp 434c <_sk_evenly_spaced_gradient_hsw+0x2f>
.byte 76,137,219 // mov %r11,%rbx
.byte 72,209,235 // shr %rbx
.byte 65,131,227,1 // and $0x1,%r11d
@@ -13877,7 +13999,7 @@ _sk_evenly_spaced_gradient_hsw:
.byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1
.byte 197,126,91,217 // vcvttps2dq %ymm1,%ymm11
.byte 73,131,250,8 // cmp $0x8,%r10
- .byte 119,70 // ja 41c9 <_sk_evenly_spaced_gradient_hsw+0x88>
+ .byte 119,70 // ja 43a5 <_sk_evenly_spaced_gradient_hsw+0x88>
.byte 196,66,37,22,1 // vpermps (%r9),%ymm11,%ymm8
.byte 72,139,88,40 // mov 0x28(%rax),%rbx
.byte 196,98,37,22,11 // vpermps (%rbx),%ymm11,%ymm9
@@ -13893,7 +14015,7 @@ _sk_evenly_spaced_gradient_hsw:
.byte 196,226,37,22,27 // vpermps (%rbx),%ymm11,%ymm3
.byte 72,139,64,64 // mov 0x40(%rax),%rax
.byte 196,98,37,22,40 // vpermps (%rax),%ymm11,%ymm13
- .byte 235,110 // jmp 4237 <_sk_evenly_spaced_gradient_hsw+0xf6>
+ .byte 235,110 // jmp 4413 <_sk_evenly_spaced_gradient_hsw+0xf6>
.byte 196,65,13,118,246 // vpcmpeqd %ymm14,%ymm14,%ymm14
.byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
.byte 196,2,117,146,4,153 // vgatherdps %ymm1,(%r9,%ymm11,4),%ymm8
@@ -13930,14 +14052,14 @@ HIDDEN _sk_gauss_a_to_rgba_hsw
.globl _sk_gauss_a_to_rgba_hsw
FUNCTION(_sk_gauss_a_to_rgba_hsw)
_sk_gauss_a_to_rgba_hsw:
- .byte 196,226,125,24,5,111,10,0,0 // vbroadcastss 0xa6f(%rip),%ymm0 # 4ccc <_sk_callback_hsw+0x474>
- .byte 196,226,125,24,13,106,10,0,0 // vbroadcastss 0xa6a(%rip),%ymm1 # 4cd0 <_sk_callback_hsw+0x478>
+ .byte 196,226,125,24,5,111,10,0,0 // vbroadcastss 0xa6f(%rip),%ymm0 # 4ea8 <_sk_callback_hsw+0x474>
+ .byte 196,226,125,24,13,106,10,0,0 // vbroadcastss 0xa6a(%rip),%ymm1 # 4eac <_sk_callback_hsw+0x478>
.byte 196,226,101,168,200 // vfmadd213ps %ymm0,%ymm3,%ymm1
- .byte 196,226,125,24,5,96,10,0,0 // vbroadcastss 0xa60(%rip),%ymm0 # 4cd4 <_sk_callback_hsw+0x47c>
+ .byte 196,226,125,24,5,96,10,0,0 // vbroadcastss 0xa60(%rip),%ymm0 # 4eb0 <_sk_callback_hsw+0x47c>
.byte 196,226,101,184,193 // vfmadd231ps %ymm1,%ymm3,%ymm0
- .byte 196,226,125,24,13,86,10,0,0 // vbroadcastss 0xa56(%rip),%ymm1 # 4cd8 <_sk_callback_hsw+0x480>
+ .byte 196,226,125,24,13,86,10,0,0 // vbroadcastss 0xa56(%rip),%ymm1 # 4eb4 <_sk_callback_hsw+0x480>
.byte 196,226,101,184,200 // vfmadd231ps %ymm0,%ymm3,%ymm1
- .byte 196,226,125,24,5,76,10,0,0 // vbroadcastss 0xa4c(%rip),%ymm0 # 4cdc <_sk_callback_hsw+0x484>
+ .byte 196,226,125,24,5,76,10,0,0 // vbroadcastss 0xa4c(%rip),%ymm0 # 4eb8 <_sk_callback_hsw+0x484>
.byte 196,226,101,184,193 // vfmadd231ps %ymm1,%ymm3,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
@@ -13952,11 +14074,11 @@ _sk_gradient_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,8 // mov (%rax),%r9
.byte 73,131,249,1 // cmp $0x1,%r9
- .byte 15,134,180,0,0,0 // jbe 4368 <_sk_gradient_hsw+0xc3>
+ .byte 15,134,180,0,0,0 // jbe 4544 <_sk_gradient_hsw+0xc3>
.byte 76,139,80,72 // mov 0x48(%rax),%r10
.byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
.byte 65,187,1,0,0,0 // mov $0x1,%r11d
- .byte 196,226,125,24,21,21,10,0,0 // vbroadcastss 0xa15(%rip),%ymm2 # 4ce0 <_sk_callback_hsw+0x488>
+ .byte 196,226,125,24,21,21,10,0,0 // vbroadcastss 0xa15(%rip),%ymm2 # 4ebc <_sk_callback_hsw+0x488>
.byte 196,65,53,239,201 // vpxor %ymm9,%ymm9,%ymm9
.byte 196,130,125,24,28,154 // vbroadcastss (%r10,%r11,4),%ymm3
.byte 197,228,194,216,2 // vcmpleps %ymm0,%ymm3,%ymm3
@@ -13964,10 +14086,10 @@ _sk_gradient_hsw:
.byte 196,65,101,254,201 // vpaddd %ymm9,%ymm3,%ymm9
.byte 73,255,195 // inc %r11
.byte 77,57,217 // cmp %r11,%r9
- .byte 117,226 // jne 42d0 <_sk_gradient_hsw+0x2b>
+ .byte 117,226 // jne 44ac <_sk_gradient_hsw+0x2b>
.byte 76,139,80,8 // mov 0x8(%rax),%r10
.byte 73,131,249,8 // cmp $0x8,%r9
- .byte 118,121 // jbe 4371 <_sk_gradient_hsw+0xcc>
+ .byte 118,121 // jbe 454d <_sk_gradient_hsw+0xcc>
.byte 196,65,13,118,246 // vpcmpeqd %ymm14,%ymm14,%ymm14
.byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
.byte 196,2,117,146,4,138 // vgatherdps %ymm1,(%r10,%ymm9,4),%ymm8
@@ -13991,7 +14113,7 @@ _sk_gradient_hsw:
.byte 196,130,21,146,28,137 // vgatherdps %ymm13,(%r9,%ymm9,4),%ymm3
.byte 72,139,64,64 // mov 0x40(%rax),%rax
.byte 196,34,13,146,44,136 // vgatherdps %ymm14,(%rax,%ymm9,4),%ymm13
- .byte 235,77 // jmp 43b5 <_sk_gradient_hsw+0x110>
+ .byte 235,77 // jmp 4591 <_sk_gradient_hsw+0x110>
.byte 76,139,80,8 // mov 0x8(%rax),%r10
.byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
.byte 196,66,53,22,2 // vpermps (%r10),%ymm9,%ymm8
@@ -14051,24 +14173,24 @@ _sk_xy_to_unit_angle_hsw:
.byte 196,65,52,95,226 // vmaxps %ymm10,%ymm9,%ymm12
.byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11
.byte 196,65,36,89,227 // vmulps %ymm11,%ymm11,%ymm12
- .byte 196,98,125,24,45,148,8,0,0 // vbroadcastss 0x894(%rip),%ymm13 # 4ce4 <_sk_callback_hsw+0x48c>
- .byte 196,98,125,24,53,143,8,0,0 // vbroadcastss 0x88f(%rip),%ymm14 # 4ce8 <_sk_callback_hsw+0x490>
+ .byte 196,98,125,24,45,148,8,0,0 // vbroadcastss 0x894(%rip),%ymm13 # 4ec0 <_sk_callback_hsw+0x48c>
+ .byte 196,98,125,24,53,143,8,0,0 // vbroadcastss 0x88f(%rip),%ymm14 # 4ec4 <_sk_callback_hsw+0x490>
.byte 196,66,29,184,245 // vfmadd231ps %ymm13,%ymm12,%ymm14
- .byte 196,98,125,24,45,133,8,0,0 // vbroadcastss 0x885(%rip),%ymm13 # 4cec <_sk_callback_hsw+0x494>
+ .byte 196,98,125,24,45,133,8,0,0 // vbroadcastss 0x885(%rip),%ymm13 # 4ec8 <_sk_callback_hsw+0x494>
.byte 196,66,29,184,238 // vfmadd231ps %ymm14,%ymm12,%ymm13
- .byte 196,98,125,24,53,123,8,0,0 // vbroadcastss 0x87b(%rip),%ymm14 # 4cf0 <_sk_callback_hsw+0x498>
+ .byte 196,98,125,24,53,123,8,0,0 // vbroadcastss 0x87b(%rip),%ymm14 # 4ecc <_sk_callback_hsw+0x498>
.byte 196,66,29,184,245 // vfmadd231ps %ymm13,%ymm12,%ymm14
.byte 196,65,36,89,222 // vmulps %ymm14,%ymm11,%ymm11
.byte 196,65,52,194,202,1 // vcmpltps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,24,21,102,8,0,0 // vbroadcastss 0x866(%rip),%ymm10 # 4cf4 <_sk_callback_hsw+0x49c>
+ .byte 196,98,125,24,21,102,8,0,0 // vbroadcastss 0x866(%rip),%ymm10 # 4ed0 <_sk_callback_hsw+0x49c>
.byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10
.byte 196,67,37,74,202,144 // vblendvps %ymm9,%ymm10,%ymm11,%ymm9
.byte 196,193,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,21,80,8,0,0 // vbroadcastss 0x850(%rip),%ymm10 # 4cf8 <_sk_callback_hsw+0x4a0>
+ .byte 196,98,125,24,21,80,8,0,0 // vbroadcastss 0x850(%rip),%ymm10 # 4ed4 <_sk_callback_hsw+0x4a0>
.byte 196,65,44,92,209 // vsubps %ymm9,%ymm10,%ymm10
.byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0
.byte 196,65,116,194,200,1 // vcmpltps %ymm8,%ymm1,%ymm9
- .byte 196,98,125,24,21,58,8,0,0 // vbroadcastss 0x83a(%rip),%ymm10 # 4cfc <_sk_callback_hsw+0x4a4>
+ .byte 196,98,125,24,21,58,8,0,0 // vbroadcastss 0x83a(%rip),%ymm10 # 4ed8 <_sk_callback_hsw+0x4a4>
.byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10
.byte 196,195,125,74,194,144 // vblendvps %ymm9,%ymm10,%ymm0,%ymm0
.byte 196,65,124,194,200,3 // vcmpunordps %ymm8,%ymm0,%ymm9
@@ -14091,7 +14213,7 @@ HIDDEN _sk_save_xy_hsw
FUNCTION(_sk_save_xy_hsw)
_sk_save_xy_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,7,8,0,0 // vbroadcastss 0x807(%rip),%ymm8 # 4d00 <_sk_callback_hsw+0x4a8>
+ .byte 196,98,125,24,5,7,8,0,0 // vbroadcastss 0x807(%rip),%ymm8 # 4edc <_sk_callback_hsw+0x4a8>
.byte 196,65,124,88,200 // vaddps %ymm8,%ymm0,%ymm9
.byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10
.byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9
@@ -14125,9 +14247,9 @@ HIDDEN _sk_bilinear_nx_hsw
FUNCTION(_sk_bilinear_nx_hsw)
_sk_bilinear_nx_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,155,7,0,0 // vbroadcastss 0x79b(%rip),%ymm0 # 4d04 <_sk_callback_hsw+0x4ac>
+ .byte 196,226,125,24,5,155,7,0,0 // vbroadcastss 0x79b(%rip),%ymm0 # 4ee0 <_sk_callback_hsw+0x4ac>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,146,7,0,0 // vbroadcastss 0x792(%rip),%ymm8 # 4d08 <_sk_callback_hsw+0x4b0>
+ .byte 196,98,125,24,5,146,7,0,0 // vbroadcastss 0x792(%rip),%ymm8 # 4ee4 <_sk_callback_hsw+0x4b0>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -14138,7 +14260,7 @@ HIDDEN _sk_bilinear_px_hsw
FUNCTION(_sk_bilinear_px_hsw)
_sk_bilinear_px_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,122,7,0,0 // vbroadcastss 0x77a(%rip),%ymm0 # 4d0c <_sk_callback_hsw+0x4b4>
+ .byte 196,226,125,24,5,122,7,0,0 // vbroadcastss 0x77a(%rip),%ymm0 # 4ee8 <_sk_callback_hsw+0x4b4>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
.byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
@@ -14150,9 +14272,9 @@ HIDDEN _sk_bilinear_ny_hsw
FUNCTION(_sk_bilinear_ny_hsw)
_sk_bilinear_ny_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,94,7,0,0 // vbroadcastss 0x75e(%rip),%ymm1 # 4d10 <_sk_callback_hsw+0x4b8>
+ .byte 196,226,125,24,13,94,7,0,0 // vbroadcastss 0x75e(%rip),%ymm1 # 4eec <_sk_callback_hsw+0x4b8>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,84,7,0,0 // vbroadcastss 0x754(%rip),%ymm8 # 4d14 <_sk_callback_hsw+0x4bc>
+ .byte 196,98,125,24,5,84,7,0,0 // vbroadcastss 0x754(%rip),%ymm8 # 4ef0 <_sk_callback_hsw+0x4bc>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -14163,7 +14285,7 @@ HIDDEN _sk_bilinear_py_hsw
FUNCTION(_sk_bilinear_py_hsw)
_sk_bilinear_py_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,60,7,0,0 // vbroadcastss 0x73c(%rip),%ymm1 # 4d18 <_sk_callback_hsw+0x4c0>
+ .byte 196,226,125,24,13,60,7,0,0 // vbroadcastss 0x73c(%rip),%ymm1 # 4ef4 <_sk_callback_hsw+0x4c0>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
.byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
@@ -14175,13 +14297,13 @@ HIDDEN _sk_bicubic_n3x_hsw
FUNCTION(_sk_bicubic_n3x_hsw)
_sk_bicubic_n3x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,31,7,0,0 // vbroadcastss 0x71f(%rip),%ymm0 # 4d1c <_sk_callback_hsw+0x4c4>
+ .byte 196,226,125,24,5,31,7,0,0 // vbroadcastss 0x71f(%rip),%ymm0 # 4ef8 <_sk_callback_hsw+0x4c4>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,22,7,0,0 // vbroadcastss 0x716(%rip),%ymm8 # 4d20 <_sk_callback_hsw+0x4c8>
+ .byte 196,98,125,24,5,22,7,0,0 // vbroadcastss 0x716(%rip),%ymm8 # 4efc <_sk_callback_hsw+0x4c8>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,7,7,0,0 // vbroadcastss 0x707(%rip),%ymm10 # 4d24 <_sk_callback_hsw+0x4cc>
- .byte 196,98,125,24,29,2,7,0,0 // vbroadcastss 0x702(%rip),%ymm11 # 4d28 <_sk_callback_hsw+0x4d0>
+ .byte 196,98,125,24,21,7,7,0,0 // vbroadcastss 0x707(%rip),%ymm10 # 4f00 <_sk_callback_hsw+0x4cc>
+ .byte 196,98,125,24,29,2,7,0,0 // vbroadcastss 0x702(%rip),%ymm11 # 4f04 <_sk_callback_hsw+0x4d0>
.byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11
.byte 196,65,36,89,193 // vmulps %ymm9,%ymm11,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
@@ -14193,16 +14315,16 @@ HIDDEN _sk_bicubic_n1x_hsw
FUNCTION(_sk_bicubic_n1x_hsw)
_sk_bicubic_n1x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,229,6,0,0 // vbroadcastss 0x6e5(%rip),%ymm0 # 4d2c <_sk_callback_hsw+0x4d4>
+ .byte 196,226,125,24,5,229,6,0,0 // vbroadcastss 0x6e5(%rip),%ymm0 # 4f08 <_sk_callback_hsw+0x4d4>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,220,6,0,0 // vbroadcastss 0x6dc(%rip),%ymm8 # 4d30 <_sk_callback_hsw+0x4d8>
+ .byte 196,98,125,24,5,220,6,0,0 // vbroadcastss 0x6dc(%rip),%ymm8 # 4f0c <_sk_callback_hsw+0x4d8>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,210,6,0,0 // vbroadcastss 0x6d2(%rip),%ymm9 # 4d34 <_sk_callback_hsw+0x4dc>
- .byte 196,98,125,24,21,205,6,0,0 // vbroadcastss 0x6cd(%rip),%ymm10 # 4d38 <_sk_callback_hsw+0x4e0>
+ .byte 196,98,125,24,13,210,6,0,0 // vbroadcastss 0x6d2(%rip),%ymm9 # 4f10 <_sk_callback_hsw+0x4dc>
+ .byte 196,98,125,24,21,205,6,0,0 // vbroadcastss 0x6cd(%rip),%ymm10 # 4f14 <_sk_callback_hsw+0x4e0>
.byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,13,195,6,0,0 // vbroadcastss 0x6c3(%rip),%ymm9 # 4d3c <_sk_callback_hsw+0x4e4>
+ .byte 196,98,125,24,13,195,6,0,0 // vbroadcastss 0x6c3(%rip),%ymm9 # 4f18 <_sk_callback_hsw+0x4e4>
.byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
- .byte 196,98,125,24,21,185,6,0,0 // vbroadcastss 0x6b9(%rip),%ymm10 # 4d40 <_sk_callback_hsw+0x4e8>
+ .byte 196,98,125,24,21,185,6,0,0 // vbroadcastss 0x6b9(%rip),%ymm10 # 4f1c <_sk_callback_hsw+0x4e8>
.byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10
.byte 197,124,17,144,128,0,0,0 // vmovups %ymm10,0x80(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -14213,14 +14335,14 @@ HIDDEN _sk_bicubic_p1x_hsw
FUNCTION(_sk_bicubic_p1x_hsw)
_sk_bicubic_p1x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,161,6,0,0 // vbroadcastss 0x6a1(%rip),%ymm8 # 4d44 <_sk_callback_hsw+0x4ec>
+ .byte 196,98,125,24,5,161,6,0,0 // vbroadcastss 0x6a1(%rip),%ymm8 # 4f20 <_sk_callback_hsw+0x4ec>
.byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
.byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
- .byte 196,98,125,24,21,147,6,0,0 // vbroadcastss 0x693(%rip),%ymm10 # 4d48 <_sk_callback_hsw+0x4f0>
- .byte 196,98,125,24,29,142,6,0,0 // vbroadcastss 0x68e(%rip),%ymm11 # 4d4c <_sk_callback_hsw+0x4f4>
+ .byte 196,98,125,24,21,147,6,0,0 // vbroadcastss 0x693(%rip),%ymm10 # 4f24 <_sk_callback_hsw+0x4f0>
+ .byte 196,98,125,24,29,142,6,0,0 // vbroadcastss 0x68e(%rip),%ymm11 # 4f28 <_sk_callback_hsw+0x4f4>
.byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11
.byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11
- .byte 196,98,125,24,5,127,6,0,0 // vbroadcastss 0x67f(%rip),%ymm8 # 4d50 <_sk_callback_hsw+0x4f8>
+ .byte 196,98,125,24,5,127,6,0,0 // vbroadcastss 0x67f(%rip),%ymm8 # 4f2c <_sk_callback_hsw+0x4f8>
.byte 196,66,53,184,195 // vfmadd231ps %ymm11,%ymm9,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -14231,12 +14353,12 @@ HIDDEN _sk_bicubic_p3x_hsw
FUNCTION(_sk_bicubic_p3x_hsw)
_sk_bicubic_p3x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,103,6,0,0 // vbroadcastss 0x667(%rip),%ymm0 # 4d54 <_sk_callback_hsw+0x4fc>
+ .byte 196,226,125,24,5,103,6,0,0 // vbroadcastss 0x667(%rip),%ymm0 # 4f30 <_sk_callback_hsw+0x4fc>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
.byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,84,6,0,0 // vbroadcastss 0x654(%rip),%ymm10 # 4d58 <_sk_callback_hsw+0x500>
- .byte 196,98,125,24,29,79,6,0,0 // vbroadcastss 0x64f(%rip),%ymm11 # 4d5c <_sk_callback_hsw+0x504>
+ .byte 196,98,125,24,21,84,6,0,0 // vbroadcastss 0x654(%rip),%ymm10 # 4f34 <_sk_callback_hsw+0x500>
+ .byte 196,98,125,24,29,79,6,0,0 // vbroadcastss 0x64f(%rip),%ymm11 # 4f38 <_sk_callback_hsw+0x504>
.byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11
.byte 196,65,52,89,195 // vmulps %ymm11,%ymm9,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
@@ -14248,13 +14370,13 @@ HIDDEN _sk_bicubic_n3y_hsw
FUNCTION(_sk_bicubic_n3y_hsw)
_sk_bicubic_n3y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,50,6,0,0 // vbroadcastss 0x632(%rip),%ymm1 # 4d60 <_sk_callback_hsw+0x508>
+ .byte 196,226,125,24,13,50,6,0,0 // vbroadcastss 0x632(%rip),%ymm1 # 4f3c <_sk_callback_hsw+0x508>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,40,6,0,0 // vbroadcastss 0x628(%rip),%ymm8 # 4d64 <_sk_callback_hsw+0x50c>
+ .byte 196,98,125,24,5,40,6,0,0 // vbroadcastss 0x628(%rip),%ymm8 # 4f40 <_sk_callback_hsw+0x50c>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,25,6,0,0 // vbroadcastss 0x619(%rip),%ymm10 # 4d68 <_sk_callback_hsw+0x510>
- .byte 196,98,125,24,29,20,6,0,0 // vbroadcastss 0x614(%rip),%ymm11 # 4d6c <_sk_callback_hsw+0x514>
+ .byte 196,98,125,24,21,25,6,0,0 // vbroadcastss 0x619(%rip),%ymm10 # 4f44 <_sk_callback_hsw+0x510>
+ .byte 196,98,125,24,29,20,6,0,0 // vbroadcastss 0x614(%rip),%ymm11 # 4f48 <_sk_callback_hsw+0x514>
.byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11
.byte 196,65,36,89,193 // vmulps %ymm9,%ymm11,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
@@ -14266,16 +14388,16 @@ HIDDEN _sk_bicubic_n1y_hsw
FUNCTION(_sk_bicubic_n1y_hsw)
_sk_bicubic_n1y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,247,5,0,0 // vbroadcastss 0x5f7(%rip),%ymm1 # 4d70 <_sk_callback_hsw+0x518>
+ .byte 196,226,125,24,13,247,5,0,0 // vbroadcastss 0x5f7(%rip),%ymm1 # 4f4c <_sk_callback_hsw+0x518>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,237,5,0,0 // vbroadcastss 0x5ed(%rip),%ymm8 # 4d74 <_sk_callback_hsw+0x51c>
+ .byte 196,98,125,24,5,237,5,0,0 // vbroadcastss 0x5ed(%rip),%ymm8 # 4f50 <_sk_callback_hsw+0x51c>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,227,5,0,0 // vbroadcastss 0x5e3(%rip),%ymm9 # 4d78 <_sk_callback_hsw+0x520>
- .byte 196,98,125,24,21,222,5,0,0 // vbroadcastss 0x5de(%rip),%ymm10 # 4d7c <_sk_callback_hsw+0x524>
+ .byte 196,98,125,24,13,227,5,0,0 // vbroadcastss 0x5e3(%rip),%ymm9 # 4f54 <_sk_callback_hsw+0x520>
+ .byte 196,98,125,24,21,222,5,0,0 // vbroadcastss 0x5de(%rip),%ymm10 # 4f58 <_sk_callback_hsw+0x524>
.byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,13,212,5,0,0 // vbroadcastss 0x5d4(%rip),%ymm9 # 4d80 <_sk_callback_hsw+0x528>
+ .byte 196,98,125,24,13,212,5,0,0 // vbroadcastss 0x5d4(%rip),%ymm9 # 4f5c <_sk_callback_hsw+0x528>
.byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
- .byte 196,98,125,24,21,202,5,0,0 // vbroadcastss 0x5ca(%rip),%ymm10 # 4d84 <_sk_callback_hsw+0x52c>
+ .byte 196,98,125,24,21,202,5,0,0 // vbroadcastss 0x5ca(%rip),%ymm10 # 4f60 <_sk_callback_hsw+0x52c>
.byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10
.byte 197,124,17,144,160,0,0,0 // vmovups %ymm10,0xa0(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -14286,14 +14408,14 @@ HIDDEN _sk_bicubic_p1y_hsw
FUNCTION(_sk_bicubic_p1y_hsw)
_sk_bicubic_p1y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,178,5,0,0 // vbroadcastss 0x5b2(%rip),%ymm8 # 4d88 <_sk_callback_hsw+0x530>
+ .byte 196,98,125,24,5,178,5,0,0 // vbroadcastss 0x5b2(%rip),%ymm8 # 4f64 <_sk_callback_hsw+0x530>
.byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
.byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
- .byte 196,98,125,24,21,163,5,0,0 // vbroadcastss 0x5a3(%rip),%ymm10 # 4d8c <_sk_callback_hsw+0x534>
- .byte 196,98,125,24,29,158,5,0,0 // vbroadcastss 0x59e(%rip),%ymm11 # 4d90 <_sk_callback_hsw+0x538>
+ .byte 196,98,125,24,21,163,5,0,0 // vbroadcastss 0x5a3(%rip),%ymm10 # 4f68 <_sk_callback_hsw+0x534>
+ .byte 196,98,125,24,29,158,5,0,0 // vbroadcastss 0x59e(%rip),%ymm11 # 4f6c <_sk_callback_hsw+0x538>
.byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11
.byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11
- .byte 196,98,125,24,5,143,5,0,0 // vbroadcastss 0x58f(%rip),%ymm8 # 4d94 <_sk_callback_hsw+0x53c>
+ .byte 196,98,125,24,5,143,5,0,0 // vbroadcastss 0x58f(%rip),%ymm8 # 4f70 <_sk_callback_hsw+0x53c>
.byte 196,66,53,184,195 // vfmadd231ps %ymm11,%ymm9,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -14304,12 +14426,12 @@ HIDDEN _sk_bicubic_p3y_hsw
FUNCTION(_sk_bicubic_p3y_hsw)
_sk_bicubic_p3y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,119,5,0,0 // vbroadcastss 0x577(%rip),%ymm1 # 4d98 <_sk_callback_hsw+0x540>
+ .byte 196,226,125,24,13,119,5,0,0 // vbroadcastss 0x577(%rip),%ymm1 # 4f74 <_sk_callback_hsw+0x540>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
.byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,99,5,0,0 // vbroadcastss 0x563(%rip),%ymm10 # 4d9c <_sk_callback_hsw+0x544>
- .byte 196,98,125,24,29,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm11 # 4da0 <_sk_callback_hsw+0x548>
+ .byte 196,98,125,24,21,99,5,0,0 // vbroadcastss 0x563(%rip),%ymm10 # 4f78 <_sk_callback_hsw+0x544>
+ .byte 196,98,125,24,29,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm11 # 4f7c <_sk_callback_hsw+0x548>
.byte 196,66,61,168,218 // vfmadd213ps %ymm10,%ymm8,%ymm11
.byte 196,65,52,89,195 // vmulps %ymm11,%ymm9,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
@@ -14439,25 +14561,25 @@ BALIGN4
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 4a4d <.literal4+0xb1>
+ .byte 71,225,61 // rex.RXB loope 4c29 <.literal4+0xb1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,154 // cmpb $0x9a,(%rdi)
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 4a5d <.literal4+0xc1>
+ .byte 71,225,61 // rex.RXB loope 4c39 <.literal4+0xc1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,154 // cmpb $0x9a,(%rdi)
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 4a6d <.literal4+0xd1>
+ .byte 71,225,61 // rex.RXB loope 4c49 <.literal4+0xd1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,154 // cmpb $0x9a,(%rdi)
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 4a7d <.literal4+0xe1>
+ .byte 71,225,61 // rex.RXB loope 4c59 <.literal4+0xe1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax)
@@ -14507,7 +14629,7 @@ BALIGN4
.byte 190,129,128,128,59 // mov $0x3b808081,%esi
.byte 129,128,128,59,0,248,0,0,8,33 // addl $0x21080000,-0x7ffc480(%rax)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4ad5 <.literal4+0x139>
+ .byte 224,7 // loopne 4cb1 <.literal4+0x139>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -14523,10 +14645,10 @@ BALIGN4
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
.byte 0,52,255 // add %dh,(%rdi,%rdi,8)
.byte 255 // (bad)
- .byte 127,0 // jg 4afc <.literal4+0x160>
+ .byte 127,0 // jg 4cd8 <.literal4+0x160>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4b75 <.literal4+0x1d9>
+ .byte 119,115 // ja 4d51 <.literal4+0x1d9>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -14540,10 +14662,10 @@ BALIGN4
.byte 0,128,63,0,0,0 // add %al,0x3f(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4b30 <.literal4+0x194>
+ .byte 127,0 // jg 4d0c <.literal4+0x194>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4ba9 <.literal4+0x20d>
+ .byte 119,115 // ja 4d85 <.literal4+0x20d>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -14557,10 +14679,10 @@ BALIGN4
.byte 0,128,63,0,0,0 // add %al,0x3f(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4b64 <.literal4+0x1c8>
+ .byte 127,0 // jg 4d40 <.literal4+0x1c8>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4bdd <.literal4+0x241>
+ .byte 119,115 // ja 4db9 <.literal4+0x241>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -14574,10 +14696,10 @@ BALIGN4
.byte 0,128,63,0,0,0 // add %al,0x3f(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4b98 <.literal4+0x1fc>
+ .byte 127,0 // jg 4d74 <.literal4+0x1fc>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4c11 <.literal4+0x275>
+ .byte 119,115 // ja 4ded <.literal4+0x275>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -14590,7 +14712,7 @@ BALIGN4
.byte 0,75,0 // add %cl,0x0(%rbx)
.byte 0,128,63,0,0,200 // add %al,-0x37ffffc1(%rax)
.byte 66,0,0 // rex.X add %al,(%rax)
- .byte 127,67 // jg 4c0f <.literal4+0x273>
+ .byte 127,67 // jg 4deb <.literal4+0x273>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -14602,10 +14724,10 @@ BALIGN4
.byte 190,80,128,3,62 // mov $0x3e038050,%esi
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 4c2f <.literal4+0x293>
+ .byte 118,63 // jbe 4e0b <.literal4+0x293>
.byte 246,64,83,63 // testb $0x3f,0x53(%rax)
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 4c43 <.literal4+0x2a7>
+ .byte 127,67 // jg 4e1f <.literal4+0x2a7>
.byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax)
.byte 128,59,0 // cmpb $0x0,(%rbx)
.byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax)
@@ -14614,7 +14736,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4c25 <.literal4+0x289>
+ .byte 224,7 // loopne 4e01 <.literal4+0x289>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -14626,7 +14748,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4c41 <.literal4+0x2a5>
+ .byte 224,7 // loopne 4e1d <.literal4+0x2a5>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -14637,7 +14759,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 4c96 <.literal4+0x2fa>
+ .byte 124,66 // jl 4e72 <.literal4+0x2fa>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax)
@@ -14655,9 +14777,9 @@ BALIGN4
.byte 137,136,136,59,15,0 // mov %ecx,0xf3b88(%rax)
.byte 0,0 // add %al,(%rax)
.byte 137,136,136,61,0,0 // mov %ecx,0x3d88(%rax)
- .byte 112,65 // jo 4cd9 <.literal4+0x33d>
+ .byte 112,65 // jo 4eb5 <.literal4+0x33d>
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 4ce7 <.literal4+0x34b>
+ .byte 127,67 // jg 4ec3 <.literal4+0x34b>
.byte 128,0,128 // addb $0x80,(%rax)
.byte 55 // (bad)
.byte 128,0,128 // addb $0x80,(%rax)
@@ -14665,7 +14787,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 255 // (bad)
- .byte 127,71 // jg 4cfb <.literal4+0x35f>
+ .byte 127,71 // jg 4ed7 <.literal4+0x35f>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,191,0,0,0 // add %al,0xbf(%rax)
@@ -14773,16 +14895,16 @@ BALIGN32
.byte 0,0 // add %al,(%rax)
.byte 1,255 // add %edi,%edi
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004de8 <_sk_callback_hsw+0xa000590>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004fa8 <_sk_callback_hsw+0xa000574>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004df0 <_sk_callback_hsw+0x12000598>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004fb0 <_sk_callback_hsw+0x1200057c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004df8 <_sk_callback_hsw+0x1a0005a0>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004fb8 <_sk_callback_hsw+0x1a000584>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004e00 <_sk_callback_hsw+0x30005a8>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004fc0 <_sk_callback_hsw+0x300058c>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -14825,16 +14947,16 @@ BALIGN32
.byte 0,0 // add %al,(%rax)
.byte 1,255 // add %edi,%edi
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004e48 <_sk_callback_hsw+0xa0005f0>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005008 <_sk_callback_hsw+0xa0005d4>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004e50 <_sk_callback_hsw+0x120005f8>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005010 <_sk_callback_hsw+0x120005dc>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004e58 <_sk_callback_hsw+0x1a000600>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005018 <_sk_callback_hsw+0x1a0005e4>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004e60 <_sk_callback_hsw+0x3000608>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005020 <_sk_callback_hsw+0x30005ec>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -14877,16 +14999,16 @@ BALIGN32
.byte 0,0 // add %al,(%rax)
.byte 1,255 // add %edi,%edi
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004ea8 <_sk_callback_hsw+0xa000650>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005068 <_sk_callback_hsw+0xa000634>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004eb0 <_sk_callback_hsw+0x12000658>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005070 <_sk_callback_hsw+0x1200063c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004eb8 <_sk_callback_hsw+0x1a000660>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005078 <_sk_callback_hsw+0x1a000644>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004ec0 <_sk_callback_hsw+0x3000668>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005080 <_sk_callback_hsw+0x300064c>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -14929,16 +15051,16 @@ BALIGN32
.byte 0,0 // add %al,(%rax)
.byte 1,255 // add %edi,%edi
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004f08 <_sk_callback_hsw+0xa0006b0>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0050c8 <_sk_callback_hsw+0xa000694>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004f10 <_sk_callback_hsw+0x120006b8>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120050d0 <_sk_callback_hsw+0x1200069c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004f18 <_sk_callback_hsw+0x1a0006c0>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0050d8 <_sk_callback_hsw+0x1a0006a4>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004f20 <_sk_callback_hsw+0x30006c8>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30050e0 <_sk_callback_hsw+0x30006ac>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -14981,16 +15103,16 @@ BALIGN32
.byte 0,0 // add %al,(%rax)
.byte 1,255 // add %edi,%edi
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004f68 <_sk_callback_hsw+0xa000710>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005128 <_sk_callback_hsw+0xa0006f4>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004f70 <_sk_callback_hsw+0x12000718>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005130 <_sk_callback_hsw+0x120006fc>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004f78 <_sk_callback_hsw+0x1a000720>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005138 <_sk_callback_hsw+0x1a000704>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004f80 <_sk_callback_hsw+0x3000728>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005140 <_sk_callback_hsw+0x300070c>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -15033,6 +15155,38 @@ BALIGN16
.byte 255,0 // incl (%rax)
.byte 255,0 // incl (%rax)
.byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
BALIGN32
HIDDEN _sk_start_pipeline_avx
@@ -15119,7 +15273,7 @@ _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 196,226,125,24,13,96,100,0,0 // vbroadcastss 0x6460(%rip),%ymm1 # 6540 <_sk_callback_avx+0x144>
+ .byte 196,226,125,24,13,48,102,0,0 // vbroadcastss 0x6630(%rip),%ymm1 # 6710 <_sk_callback_avx+0x144>
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
.byte 197,252,88,7 // vaddps (%rdi),%ymm0,%ymm0
.byte 197,249,110,209 // vmovd %ecx,%xmm2
@@ -15128,7 +15282,7 @@ _sk_seed_shader_avx:
.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 196,226,125,24,21,58,100,0,0 // vbroadcastss 0x643a(%rip),%ymm2 # 6544 <_sk_callback_avx+0x148>
+ .byte 196,226,125,24,21,10,102,0,0 // vbroadcastss 0x660a(%rip),%ymm2 # 6714 <_sk_callback_avx+0x148>
.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
@@ -15152,7 +15306,7 @@ _sk_dither_avx:
.byte 196,65,121,112,201,0 // vpshufd $0x0,%xmm9,%xmm9
.byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
.byte 196,65,52,87,208 // vxorps %ymm8,%ymm9,%ymm10
- .byte 196,98,125,24,29,227,99,0,0 // vbroadcastss 0x63e3(%rip),%ymm11 # 6548 <_sk_callback_avx+0x14c>
+ .byte 196,98,125,24,29,179,101,0,0 // vbroadcastss 0x65b3(%rip),%ymm11 # 6718 <_sk_callback_avx+0x14c>
.byte 196,65,44,84,203 // vandps %ymm11,%ymm10,%ymm9
.byte 196,193,25,114,241,5 // vpslld $0x5,%xmm9,%xmm12
.byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9
@@ -15163,8 +15317,8 @@ _sk_dither_avx:
.byte 196,67,125,25,219,1 // vextractf128 $0x1,%ymm11,%xmm11
.byte 196,193,33,114,243,4 // vpslld $0x4,%xmm11,%xmm11
.byte 196,67,29,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm12,%ymm11
- .byte 196,98,125,24,37,164,99,0,0 // vbroadcastss 0x63a4(%rip),%ymm12 # 654c <_sk_callback_avx+0x150>
- .byte 196,98,125,24,45,159,99,0,0 // vbroadcastss 0x639f(%rip),%ymm13 # 6550 <_sk_callback_avx+0x154>
+ .byte 196,98,125,24,37,116,101,0,0 // vbroadcastss 0x6574(%rip),%ymm12 # 671c <_sk_callback_avx+0x150>
+ .byte 196,98,125,24,45,111,101,0,0 // vbroadcastss 0x656f(%rip),%ymm13 # 6720 <_sk_callback_avx+0x154>
.byte 196,65,44,84,245 // vandps %ymm13,%ymm10,%ymm14
.byte 196,193,1,114,246,2 // vpslld $0x2,%xmm14,%xmm15
.byte 196,67,125,25,246,1 // vextractf128 $0x1,%ymm14,%xmm14
@@ -15191,9 +15345,9 @@ _sk_dither_avx:
.byte 196,65,60,86,193 // vorps %ymm9,%ymm8,%ymm8
.byte 196,65,60,86,194 // vorps %ymm10,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,10,99,0,0 // vbroadcastss 0x630a(%rip),%ymm9 # 6554 <_sk_callback_avx+0x158>
+ .byte 196,98,125,24,13,218,100,0,0 // vbroadcastss 0x64da(%rip),%ymm9 # 6724 <_sk_callback_avx+0x158>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,0,99,0,0 // vbroadcastss 0x6300(%rip),%ymm9 # 6558 <_sk_callback_avx+0x15c>
+ .byte 196,98,125,24,13,208,100,0,0 // vbroadcastss 0x64d0(%rip),%ymm9 # 6728 <_sk_callback_avx+0x15c>
.byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
.byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
@@ -15262,7 +15416,7 @@ HIDDEN _sk_srcatop_avx
FUNCTION(_sk_srcatop_avx)
_sk_srcatop_avx:
.byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
- .byte 196,98,125,24,5,88,98,0,0 // vbroadcastss 0x6258(%rip),%ymm8 # 655c <_sk_callback_avx+0x160>
+ .byte 196,98,125,24,5,40,100,0,0 // vbroadcastss 0x6428(%rip),%ymm8 # 672c <_sk_callback_avx+0x160>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9
.byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0
@@ -15283,7 +15437,7 @@ HIDDEN _sk_dstatop_avx
FUNCTION(_sk_dstatop_avx)
_sk_dstatop_avx:
.byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8
- .byte 196,98,125,24,13,26,98,0,0 // vbroadcastss 0x621a(%rip),%ymm9 # 6560 <_sk_callback_avx+0x164>
+ .byte 196,98,125,24,13,234,99,0,0 // vbroadcastss 0x63ea(%rip),%ymm9 # 6730 <_sk_callback_avx+0x164>
.byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9
.byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
.byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0
@@ -15325,7 +15479,7 @@ HIDDEN _sk_srcout_avx
.globl _sk_srcout_avx
FUNCTION(_sk_srcout_avx)
_sk_srcout_avx:
- .byte 196,98,125,24,5,185,97,0,0 // vbroadcastss 0x61b9(%rip),%ymm8 # 6564 <_sk_callback_avx+0x168>
+ .byte 196,98,125,24,5,137,99,0,0 // vbroadcastss 0x6389(%rip),%ymm8 # 6734 <_sk_callback_avx+0x168>
.byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
.byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
@@ -15338,7 +15492,7 @@ HIDDEN _sk_dstout_avx
.globl _sk_dstout_avx
FUNCTION(_sk_dstout_avx)
_sk_dstout_avx:
- .byte 196,226,125,24,5,156,97,0,0 // vbroadcastss 0x619c(%rip),%ymm0 # 6568 <_sk_callback_avx+0x16c>
+ .byte 196,226,125,24,5,108,99,0,0 // vbroadcastss 0x636c(%rip),%ymm0 # 6738 <_sk_callback_avx+0x16c>
.byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3
.byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0
.byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1
@@ -15351,7 +15505,7 @@ HIDDEN _sk_srcover_avx
.globl _sk_srcover_avx
FUNCTION(_sk_srcover_avx)
_sk_srcover_avx:
- .byte 196,98,125,24,5,127,97,0,0 // vbroadcastss 0x617f(%rip),%ymm8 # 656c <_sk_callback_avx+0x170>
+ .byte 196,98,125,24,5,79,99,0,0 // vbroadcastss 0x634f(%rip),%ymm8 # 673c <_sk_callback_avx+0x170>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9
.byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0
@@ -15368,7 +15522,7 @@ HIDDEN _sk_dstover_avx
.globl _sk_dstover_avx
FUNCTION(_sk_dstover_avx)
_sk_dstover_avx:
- .byte 196,98,125,24,5,82,97,0,0 // vbroadcastss 0x6152(%rip),%ymm8 # 6570 <_sk_callback_avx+0x174>
+ .byte 196,98,125,24,5,34,99,0,0 // vbroadcastss 0x6322(%rip),%ymm8 # 6740 <_sk_callback_avx+0x174>
.byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
.byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
@@ -15396,7 +15550,7 @@ HIDDEN _sk_multiply_avx
.globl _sk_multiply_avx
FUNCTION(_sk_multiply_avx)
_sk_multiply_avx:
- .byte 196,98,125,24,5,17,97,0,0 // vbroadcastss 0x6111(%rip),%ymm8 # 6574 <_sk_callback_avx+0x178>
+ .byte 196,98,125,24,5,225,98,0,0 // vbroadcastss 0x62e1(%rip),%ymm8 # 6744 <_sk_callback_avx+0x178>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -15456,7 +15610,7 @@ HIDDEN _sk_xor__avx
.globl _sk_xor__avx
FUNCTION(_sk_xor__avx)
_sk_xor__avx:
- .byte 196,98,125,24,5,96,96,0,0 // vbroadcastss 0x6060(%rip),%ymm8 # 6578 <_sk_callback_avx+0x17c>
+ .byte 196,98,125,24,5,48,98,0,0 // vbroadcastss 0x6230(%rip),%ymm8 # 6748 <_sk_callback_avx+0x17c>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -15493,7 +15647,7 @@ _sk_darken_avx:
.byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
.byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,224,95,0,0 // vbroadcastss 0x5fe0(%rip),%ymm8 # 657c <_sk_callback_avx+0x180>
+ .byte 196,98,125,24,5,176,97,0,0 // vbroadcastss 0x61b0(%rip),%ymm8 # 674c <_sk_callback_avx+0x180>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
.byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
@@ -15519,7 +15673,7 @@ _sk_lighten_avx:
.byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
.byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,140,95,0,0 // vbroadcastss 0x5f8c(%rip),%ymm8 # 6580 <_sk_callback_avx+0x184>
+ .byte 196,98,125,24,5,92,97,0,0 // vbroadcastss 0x615c(%rip),%ymm8 # 6750 <_sk_callback_avx+0x184>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
.byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
@@ -15548,7 +15702,7 @@ _sk_difference_avx:
.byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
.byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,44,95,0,0 // vbroadcastss 0x5f2c(%rip),%ymm8 # 6584 <_sk_callback_avx+0x188>
+ .byte 196,98,125,24,5,252,96,0,0 // vbroadcastss 0x60fc(%rip),%ymm8 # 6754 <_sk_callback_avx+0x188>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
.byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
@@ -15571,7 +15725,7 @@ _sk_exclusion_avx:
.byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
.byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
.byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
- .byte 196,98,125,24,5,231,94,0,0 // vbroadcastss 0x5ee7(%rip),%ymm8 # 6588 <_sk_callback_avx+0x18c>
+ .byte 196,98,125,24,5,183,96,0,0 // vbroadcastss 0x60b7(%rip),%ymm8 # 6758 <_sk_callback_avx+0x18c>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
.byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
.byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
@@ -15582,7 +15736,7 @@ HIDDEN _sk_colorburn_avx
.globl _sk_colorburn_avx
FUNCTION(_sk_colorburn_avx)
_sk_colorburn_avx:
- .byte 196,98,125,24,5,210,94,0,0 // vbroadcastss 0x5ed2(%rip),%ymm8 # 658c <_sk_callback_avx+0x190>
+ .byte 196,98,125,24,5,162,96,0,0 // vbroadcastss 0x60a2(%rip),%ymm8 # 675c <_sk_callback_avx+0x190>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11
.byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
@@ -15644,7 +15798,7 @@ HIDDEN _sk_colordodge_avx
FUNCTION(_sk_colordodge_avx)
_sk_colordodge_avx:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
- .byte 196,98,125,24,13,206,93,0,0 // vbroadcastss 0x5dce(%rip),%ymm9 # 6590 <_sk_callback_avx+0x194>
+ .byte 196,98,125,24,13,158,95,0,0 // vbroadcastss 0x5f9e(%rip),%ymm9 # 6760 <_sk_callback_avx+0x194>
.byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10
.byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11
.byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9
@@ -15701,7 +15855,7 @@ HIDDEN _sk_hardlight_avx
.globl _sk_hardlight_avx
FUNCTION(_sk_hardlight_avx)
_sk_hardlight_avx:
- .byte 196,98,125,24,5,224,92,0,0 // vbroadcastss 0x5ce0(%rip),%ymm8 # 6594 <_sk_callback_avx+0x198>
+ .byte 196,98,125,24,5,176,94,0,0 // vbroadcastss 0x5eb0(%rip),%ymm8 # 6764 <_sk_callback_avx+0x198>
.byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
.byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -15756,7 +15910,7 @@ HIDDEN _sk_overlay_avx
.globl _sk_overlay_avx
FUNCTION(_sk_overlay_avx)
_sk_overlay_avx:
- .byte 196,98,125,24,5,9,92,0,0 // vbroadcastss 0x5c09(%rip),%ymm8 # 6598 <_sk_callback_avx+0x19c>
+ .byte 196,98,125,24,5,217,93,0,0 // vbroadcastss 0x5dd9(%rip),%ymm8 # 6768 <_sk_callback_avx+0x19c>
.byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
.byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -15822,10 +15976,10 @@ _sk_softlight_avx:
.byte 196,65,60,88,192 // vaddps %ymm8,%ymm8,%ymm8
.byte 196,65,60,89,216 // vmulps %ymm8,%ymm8,%ymm11
.byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
- .byte 196,98,125,24,29,0,91,0,0 // vbroadcastss 0x5b00(%rip),%ymm11 # 65a0 <_sk_callback_avx+0x1a4>
+ .byte 196,98,125,24,29,208,92,0,0 // vbroadcastss 0x5cd0(%rip),%ymm11 # 6770 <_sk_callback_avx+0x1a4>
.byte 196,65,28,88,235 // vaddps %ymm11,%ymm12,%ymm13
.byte 196,65,20,89,192 // vmulps %ymm8,%ymm13,%ymm8
- .byte 196,98,125,24,45,241,90,0,0 // vbroadcastss 0x5af1(%rip),%ymm13 # 65a4 <_sk_callback_avx+0x1a8>
+ .byte 196,98,125,24,45,193,92,0,0 // vbroadcastss 0x5cc1(%rip),%ymm13 # 6774 <_sk_callback_avx+0x1a8>
.byte 196,65,28,89,245 // vmulps %ymm13,%ymm12,%ymm14
.byte 196,65,12,88,192 // vaddps %ymm8,%ymm14,%ymm8
.byte 196,65,124,82,244 // vrsqrtps %ymm12,%ymm14
@@ -15836,7 +15990,7 @@ _sk_softlight_avx:
.byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15
.byte 196,67,13,74,240,240 // vblendvps %ymm15,%ymm8,%ymm14,%ymm14
.byte 197,116,88,249 // vaddps %ymm1,%ymm1,%ymm15
- .byte 196,98,125,24,5,175,90,0,0 // vbroadcastss 0x5aaf(%rip),%ymm8 # 659c <_sk_callback_avx+0x1a0>
+ .byte 196,98,125,24,5,127,92,0,0 // vbroadcastss 0x5c7f(%rip),%ymm8 # 676c <_sk_callback_avx+0x1a0>
.byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12
.byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0
.byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12
@@ -15963,12 +16117,12 @@ _sk_hue_avx:
.byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11
.byte 196,65,36,94,222 // vdivps %ymm14,%ymm11,%ymm11
.byte 196,67,37,74,224,240 // vblendvps %ymm15,%ymm8,%ymm11,%ymm12
- .byte 196,98,125,24,53,126,88,0,0 // vbroadcastss 0x587e(%rip),%ymm14 # 65a8 <_sk_callback_avx+0x1ac>
+ .byte 196,98,125,24,53,78,90,0,0 // vbroadcastss 0x5a4e(%rip),%ymm14 # 6778 <_sk_callback_avx+0x1ac>
.byte 196,65,92,89,222 // vmulps %ymm14,%ymm4,%ymm11
- .byte 196,98,125,24,61,116,88,0,0 // vbroadcastss 0x5874(%rip),%ymm15 # 65ac <_sk_callback_avx+0x1b0>
+ .byte 196,98,125,24,61,68,90,0,0 // vbroadcastss 0x5a44(%rip),%ymm15 # 677c <_sk_callback_avx+0x1b0>
.byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13
.byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11
- .byte 196,226,125,24,5,101,88,0,0 // vbroadcastss 0x5865(%rip),%ymm0 # 65b0 <_sk_callback_avx+0x1b4>
+ .byte 196,226,125,24,5,53,90,0,0 // vbroadcastss 0x5a35(%rip),%ymm0 # 6780 <_sk_callback_avx+0x1b4>
.byte 197,76,89,232 // vmulps %ymm0,%ymm6,%ymm13
.byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11
.byte 196,65,52,89,238 // vmulps %ymm14,%ymm9,%ymm13
@@ -16029,7 +16183,7 @@ _sk_hue_avx:
.byte 196,65,36,95,208 // vmaxps %ymm8,%ymm11,%ymm10
.byte 196,195,109,74,209,240 // vblendvps %ymm15,%ymm9,%ymm2,%ymm2
.byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,62,87,0,0 // vbroadcastss 0x573e(%rip),%ymm8 # 65b4 <_sk_callback_avx+0x1b8>
+ .byte 196,98,125,24,5,14,89,0,0 // vbroadcastss 0x590e(%rip),%ymm8 # 6784 <_sk_callback_avx+0x1b8>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -16086,12 +16240,12 @@ _sk_saturation_avx:
.byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11
.byte 196,65,36,94,222 // vdivps %ymm14,%ymm11,%ymm11
.byte 196,67,37,74,224,240 // vblendvps %ymm15,%ymm8,%ymm11,%ymm12
- .byte 196,98,125,24,53,76,86,0,0 // vbroadcastss 0x564c(%rip),%ymm14 # 65b8 <_sk_callback_avx+0x1bc>
+ .byte 196,98,125,24,53,28,88,0,0 // vbroadcastss 0x581c(%rip),%ymm14 # 6788 <_sk_callback_avx+0x1bc>
.byte 196,65,92,89,222 // vmulps %ymm14,%ymm4,%ymm11
- .byte 196,98,125,24,61,66,86,0,0 // vbroadcastss 0x5642(%rip),%ymm15 # 65bc <_sk_callback_avx+0x1c0>
+ .byte 196,98,125,24,61,18,88,0,0 // vbroadcastss 0x5812(%rip),%ymm15 # 678c <_sk_callback_avx+0x1c0>
.byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13
.byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11
- .byte 196,226,125,24,5,51,86,0,0 // vbroadcastss 0x5633(%rip),%ymm0 # 65c0 <_sk_callback_avx+0x1c4>
+ .byte 196,226,125,24,5,3,88,0,0 // vbroadcastss 0x5803(%rip),%ymm0 # 6790 <_sk_callback_avx+0x1c4>
.byte 197,76,89,232 // vmulps %ymm0,%ymm6,%ymm13
.byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11
.byte 196,65,52,89,238 // vmulps %ymm14,%ymm9,%ymm13
@@ -16152,7 +16306,7 @@ _sk_saturation_avx:
.byte 196,65,36,95,208 // vmaxps %ymm8,%ymm11,%ymm10
.byte 196,195,109,74,209,240 // vblendvps %ymm15,%ymm9,%ymm2,%ymm2
.byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,12,85,0,0 // vbroadcastss 0x550c(%rip),%ymm8 # 65c4 <_sk_callback_avx+0x1c8>
+ .byte 196,98,125,24,5,220,86,0,0 // vbroadcastss 0x56dc(%rip),%ymm8 # 6794 <_sk_callback_avx+0x1c8>
.byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
.byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
@@ -16181,12 +16335,12 @@ _sk_color_avx:
.byte 197,252,17,68,36,168 // vmovups %ymm0,-0x58(%rsp)
.byte 197,124,89,199 // vmulps %ymm7,%ymm0,%ymm8
.byte 197,116,89,207 // vmulps %ymm7,%ymm1,%ymm9
- .byte 196,98,125,24,45,162,84,0,0 // vbroadcastss 0x54a2(%rip),%ymm13 # 65c8 <_sk_callback_avx+0x1cc>
+ .byte 196,98,125,24,45,114,86,0,0 // vbroadcastss 0x5672(%rip),%ymm13 # 6798 <_sk_callback_avx+0x1cc>
.byte 196,65,92,89,213 // vmulps %ymm13,%ymm4,%ymm10
- .byte 196,98,125,24,53,152,84,0,0 // vbroadcastss 0x5498(%rip),%ymm14 # 65cc <_sk_callback_avx+0x1d0>
+ .byte 196,98,125,24,53,104,86,0,0 // vbroadcastss 0x5668(%rip),%ymm14 # 679c <_sk_callback_avx+0x1d0>
.byte 196,65,84,89,222 // vmulps %ymm14,%ymm5,%ymm11
.byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
- .byte 196,98,125,24,61,137,84,0,0 // vbroadcastss 0x5489(%rip),%ymm15 # 65d0 <_sk_callback_avx+0x1d4>
+ .byte 196,98,125,24,61,89,86,0,0 // vbroadcastss 0x5659(%rip),%ymm15 # 67a0 <_sk_callback_avx+0x1d4>
.byte 196,65,76,89,223 // vmulps %ymm15,%ymm6,%ymm11
.byte 196,193,44,88,195 // vaddps %ymm11,%ymm10,%ymm0
.byte 196,65,60,89,221 // vmulps %ymm13,%ymm8,%ymm11
@@ -16249,7 +16403,7 @@ _sk_color_avx:
.byte 196,65,44,95,207 // vmaxps %ymm15,%ymm10,%ymm9
.byte 196,195,37,74,192,0 // vblendvps %ymm0,%ymm8,%ymm11,%ymm0
.byte 196,65,124,95,199 // vmaxps %ymm15,%ymm0,%ymm8
- .byte 196,226,125,24,5,80,83,0,0 // vbroadcastss 0x5350(%rip),%ymm0 # 65d4 <_sk_callback_avx+0x1d8>
+ .byte 196,226,125,24,5,32,85,0,0 // vbroadcastss 0x5520(%rip),%ymm0 # 67a4 <_sk_callback_avx+0x1d8>
.byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10
.byte 197,172,89,84,36,168 // vmulps -0x58(%rsp),%ymm10,%ymm2
.byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11
@@ -16279,12 +16433,12 @@ _sk_luminosity_avx:
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8
.byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
- .byte 196,98,125,24,45,226,82,0,0 // vbroadcastss 0x52e2(%rip),%ymm13 # 65d8 <_sk_callback_avx+0x1dc>
+ .byte 196,98,125,24,45,178,84,0,0 // vbroadcastss 0x54b2(%rip),%ymm13 # 67a8 <_sk_callback_avx+0x1dc>
.byte 196,65,108,89,213 // vmulps %ymm13,%ymm2,%ymm10
- .byte 196,98,125,24,53,216,82,0,0 // vbroadcastss 0x52d8(%rip),%ymm14 # 65dc <_sk_callback_avx+0x1e0>
+ .byte 196,98,125,24,53,168,84,0,0 // vbroadcastss 0x54a8(%rip),%ymm14 # 67ac <_sk_callback_avx+0x1e0>
.byte 196,65,116,89,222 // vmulps %ymm14,%ymm1,%ymm11
.byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
- .byte 196,98,125,24,61,201,82,0,0 // vbroadcastss 0x52c9(%rip),%ymm15 # 65e0 <_sk_callback_avx+0x1e4>
+ .byte 196,98,125,24,61,153,84,0,0 // vbroadcastss 0x5499(%rip),%ymm15 # 67b0 <_sk_callback_avx+0x1e4>
.byte 196,65,28,89,223 // vmulps %ymm15,%ymm12,%ymm11
.byte 196,193,44,88,195 // vaddps %ymm11,%ymm10,%ymm0
.byte 196,65,60,89,221 // vmulps %ymm13,%ymm8,%ymm11
@@ -16347,7 +16501,7 @@ _sk_luminosity_avx:
.byte 196,65,44,95,207 // vmaxps %ymm15,%ymm10,%ymm9
.byte 196,195,37,74,192,0 // vblendvps %ymm0,%ymm8,%ymm11,%ymm0
.byte 196,65,124,95,199 // vmaxps %ymm15,%ymm0,%ymm8
- .byte 196,226,125,24,5,144,81,0,0 // vbroadcastss 0x5190(%rip),%ymm0 # 65e4 <_sk_callback_avx+0x1e8>
+ .byte 196,226,125,24,5,96,83,0,0 // vbroadcastss 0x5360(%rip),%ymm0 # 67b4 <_sk_callback_avx+0x1e8>
.byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10
.byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
.byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11
@@ -16378,7 +16532,7 @@ _sk_srcover_rgba_8888_avx:
.byte 77,133,192 // test %r8,%r8
.byte 15,133,43,1,0,0 // jne 15e3 <_sk_srcover_rgba_8888_avx+0x144>
.byte 196,193,124,16,58 // vmovups (%r10),%ymm7
- .byte 197,124,40,13,91,85,0,0 // vmovaps 0x555b(%rip),%ymm9 # 6a20 <_sk_callback_avx+0x624>
+ .byte 197,124,40,13,91,87,0,0 // vmovaps 0x575b(%rip),%ymm9 # 6c20 <_sk_callback_avx+0x654>
.byte 196,193,68,84,225 // vandps %ymm9,%ymm7,%ymm4
.byte 197,252,91,228 // vcvtdq2ps %ymm4,%ymm4
.byte 197,209,114,215,8 // vpsrld $0x8,%xmm7,%xmm5
@@ -16396,9 +16550,9 @@ _sk_srcover_rgba_8888_avx:
.byte 196,193,65,114,208,24 // vpsrld $0x18,%xmm8,%xmm7
.byte 196,227,53,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm9,%ymm7
.byte 197,252,91,255 // vcvtdq2ps %ymm7,%ymm7
- .byte 196,98,125,24,5,194,80,0,0 // vbroadcastss 0x50c2(%rip),%ymm8 # 65e8 <_sk_callback_avx+0x1ec>
+ .byte 196,98,125,24,5,146,82,0,0 // vbroadcastss 0x5292(%rip),%ymm8 # 67b8 <_sk_callback_avx+0x1ec>
.byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
- .byte 196,98,125,24,13,185,80,0,0 // vbroadcastss 0x50b9(%rip),%ymm9 # 65ec <_sk_callback_avx+0x1f0>
+ .byte 196,98,125,24,13,137,82,0,0 // vbroadcastss 0x5289(%rip),%ymm9 # 67bc <_sk_callback_avx+0x1f0>
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
.byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10
.byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
@@ -16443,9 +16597,9 @@ _sk_srcover_rgba_8888_avx:
.byte 72,211,232 // shr %cl,%rax
.byte 196,225,249,110,224 // vmovq %rax,%xmm4
.byte 196,226,121,48,228 // vpmovzxbw %xmm4,%xmm4
- .byte 196,226,89,0,45,101,83,0,0 // vpshufb 0x5365(%rip),%xmm4,%xmm5 # 6970 <_sk_callback_avx+0x574>
+ .byte 196,226,89,0,45,53,85,0,0 // vpshufb 0x5535(%rip),%xmm4,%xmm5 # 6b40 <_sk_callback_avx+0x574>
.byte 196,226,121,33,237 // vpmovsxbd %xmm5,%xmm5
- .byte 196,226,89,0,37,103,83,0,0 // vpshufb 0x5367(%rip),%xmm4,%xmm4 # 6980 <_sk_callback_avx+0x584>
+ .byte 196,226,89,0,37,55,85,0,0 // vpshufb 0x5537(%rip),%xmm4,%xmm4 # 6b50 <_sk_callback_avx+0x584>
.byte 196,226,121,33,228 // vpmovsxbd %xmm4,%xmm4
.byte 196,227,85,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm5,%ymm4
.byte 196,194,93,44,58 // vmaskmovps (%r10),%ymm4,%ymm7
@@ -16457,9 +16611,9 @@ _sk_srcover_rgba_8888_avx:
.byte 72,211,232 // shr %cl,%rax
.byte 196,97,249,110,200 // vmovq %rax,%xmm9
.byte 196,66,121,48,201 // vpmovzxbw %xmm9,%xmm9
- .byte 196,98,49,0,21,26,83,0,0 // vpshufb 0x531a(%rip),%xmm9,%xmm10 # 6970 <_sk_callback_avx+0x574>
+ .byte 196,98,49,0,21,234,84,0,0 // vpshufb 0x54ea(%rip),%xmm9,%xmm10 # 6b40 <_sk_callback_avx+0x574>
.byte 196,66,121,33,210 // vpmovsxbd %xmm10,%xmm10
- .byte 196,98,49,0,13,28,83,0,0 // vpshufb 0x531c(%rip),%xmm9,%xmm9 # 6980 <_sk_callback_avx+0x584>
+ .byte 196,98,49,0,13,236,84,0,0 // vpshufb 0x54ec(%rip),%xmm9,%xmm9 # 6b50 <_sk_callback_avx+0x584>
.byte 196,66,121,33,201 // vpmovsxbd %xmm9,%xmm9
.byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
.byte 196,66,53,46,2 // vmaskmovps %ymm8,%ymm9,(%r10)
@@ -16481,7 +16635,7 @@ HIDDEN _sk_clamp_1_avx
.globl _sk_clamp_1_avx
FUNCTION(_sk_clamp_1_avx)
_sk_clamp_1_avx:
- .byte 196,98,125,24,5,81,79,0,0 // vbroadcastss 0x4f51(%rip),%ymm8 # 65f0 <_sk_callback_avx+0x1f4>
+ .byte 196,98,125,24,5,33,81,0,0 // vbroadcastss 0x5121(%rip),%ymm8 # 67c0 <_sk_callback_avx+0x1f4>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
@@ -16493,7 +16647,7 @@ HIDDEN _sk_clamp_a_avx
.globl _sk_clamp_a_avx
FUNCTION(_sk_clamp_a_avx)
_sk_clamp_a_avx:
- .byte 196,98,125,24,5,52,79,0,0 // vbroadcastss 0x4f34(%rip),%ymm8 # 65f4 <_sk_callback_avx+0x1f8>
+ .byte 196,98,125,24,5,4,81,0,0 // vbroadcastss 0x5104(%rip),%ymm8 # 67c4 <_sk_callback_avx+0x1f8>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0
.byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1
@@ -16579,7 +16733,7 @@ FUNCTION(_sk_unpremul_avx)
_sk_unpremul_avx:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9
- .byte 196,98,125,24,21,124,78,0,0 // vbroadcastss 0x4e7c(%rip),%ymm10 # 65f8 <_sk_callback_avx+0x1fc>
+ .byte 196,98,125,24,21,76,80,0,0 // vbroadcastss 0x504c(%rip),%ymm10 # 67c8 <_sk_callback_avx+0x1fc>
.byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10
.byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
@@ -16592,17 +16746,17 @@ HIDDEN _sk_from_srgb_avx
.globl _sk_from_srgb_avx
FUNCTION(_sk_from_srgb_avx)
_sk_from_srgb_avx:
- .byte 196,98,125,24,5,93,78,0,0 // vbroadcastss 0x4e5d(%rip),%ymm8 # 65fc <_sk_callback_avx+0x200>
+ .byte 196,98,125,24,5,45,80,0,0 // vbroadcastss 0x502d(%rip),%ymm8 # 67cc <_sk_callback_avx+0x200>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
- .byte 196,98,125,24,29,79,78,0,0 // vbroadcastss 0x4e4f(%rip),%ymm11 # 6600 <_sk_callback_avx+0x204>
+ .byte 196,98,125,24,29,31,80,0,0 // vbroadcastss 0x501f(%rip),%ymm11 # 67d0 <_sk_callback_avx+0x204>
.byte 196,65,124,89,227 // vmulps %ymm11,%ymm0,%ymm12
- .byte 196,98,125,24,45,69,78,0,0 // vbroadcastss 0x4e45(%rip),%ymm13 # 6604 <_sk_callback_avx+0x208>
+ .byte 196,98,125,24,45,21,80,0,0 // vbroadcastss 0x5015(%rip),%ymm13 # 67d4 <_sk_callback_avx+0x208>
.byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
.byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,54,78,0,0 // vbroadcastss 0x4e36(%rip),%ymm12 # 6608 <_sk_callback_avx+0x20c>
+ .byte 196,98,125,24,37,6,80,0,0 // vbroadcastss 0x5006(%rip),%ymm12 # 67d8 <_sk_callback_avx+0x20c>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,53,44,78,0,0 // vbroadcastss 0x4e2c(%rip),%ymm14 # 660c <_sk_callback_avx+0x210>
+ .byte 196,98,125,24,53,252,79,0,0 // vbroadcastss 0x4ffc(%rip),%ymm14 # 67dc <_sk_callback_avx+0x210>
.byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0
.byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0
.byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9
@@ -16629,20 +16783,20 @@ HIDDEN _sk_to_srgb_avx
FUNCTION(_sk_to_srgb_avx)
_sk_to_srgb_avx:
.byte 197,124,82,200 // vrsqrtps %ymm0,%ymm9
- .byte 196,98,125,24,5,193,77,0,0 // vbroadcastss 0x4dc1(%rip),%ymm8 # 6610 <_sk_callback_avx+0x214>
+ .byte 196,98,125,24,5,145,79,0,0 // vbroadcastss 0x4f91(%rip),%ymm8 # 67e0 <_sk_callback_avx+0x214>
.byte 196,65,124,89,208 // vmulps %ymm8,%ymm0,%ymm10
- .byte 196,98,125,24,29,183,77,0,0 // vbroadcastss 0x4db7(%rip),%ymm11 # 6614 <_sk_callback_avx+0x218>
+ .byte 196,98,125,24,29,135,79,0,0 // vbroadcastss 0x4f87(%rip),%ymm11 # 67e4 <_sk_callback_avx+0x218>
.byte 196,65,52,89,227 // vmulps %ymm11,%ymm9,%ymm12
- .byte 196,98,125,24,45,173,77,0,0 // vbroadcastss 0x4dad(%rip),%ymm13 # 6618 <_sk_callback_avx+0x21c>
+ .byte 196,98,125,24,45,125,79,0,0 // vbroadcastss 0x4f7d(%rip),%ymm13 # 67e8 <_sk_callback_avx+0x21c>
.byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
.byte 196,65,52,89,228 // vmulps %ymm12,%ymm9,%ymm12
- .byte 196,98,125,24,53,158,77,0,0 // vbroadcastss 0x4d9e(%rip),%ymm14 # 661c <_sk_callback_avx+0x220>
+ .byte 196,98,125,24,53,110,79,0,0 // vbroadcastss 0x4f6e(%rip),%ymm14 # 67ec <_sk_callback_avx+0x220>
.byte 196,65,28,88,230 // vaddps %ymm14,%ymm12,%ymm12
- .byte 196,98,125,24,61,148,77,0,0 // vbroadcastss 0x4d94(%rip),%ymm15 # 6620 <_sk_callback_avx+0x224>
+ .byte 196,98,125,24,61,100,79,0,0 // vbroadcastss 0x4f64(%rip),%ymm15 # 67f0 <_sk_callback_avx+0x224>
.byte 196,65,52,88,207 // vaddps %ymm15,%ymm9,%ymm9
.byte 196,65,124,83,201 // vrcpps %ymm9,%ymm9
.byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9
- .byte 196,98,125,24,37,128,77,0,0 // vbroadcastss 0x4d80(%rip),%ymm12 # 6624 <_sk_callback_avx+0x228>
+ .byte 196,98,125,24,37,80,79,0,0 // vbroadcastss 0x4f50(%rip),%ymm12 # 67f4 <_sk_callback_avx+0x228>
.byte 196,193,124,194,196,1 // vcmpltps %ymm12,%ymm0,%ymm0
.byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0
.byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9
@@ -16679,7 +16833,7 @@ _sk_rgb_to_hsl_avx:
.byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9
.byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9
.byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,29,230,76,0,0 // vbroadcastss 0x4ce6(%rip),%ymm11 # 6628 <_sk_callback_avx+0x22c>
+ .byte 196,98,125,24,29,182,78,0,0 // vbroadcastss 0x4eb6(%rip),%ymm11 # 67f8 <_sk_callback_avx+0x22c>
.byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11
.byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12
.byte 196,65,28,89,227 // vmulps %ymm11,%ymm12,%ymm12
@@ -16689,19 +16843,19 @@ _sk_rgb_to_hsl_avx:
.byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2
.byte 197,252,92,201 // vsubps %ymm1,%ymm0,%ymm1
.byte 196,193,116,89,203 // vmulps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,191,76,0,0 // vbroadcastss 0x4cbf(%rip),%ymm11 # 6634 <_sk_callback_avx+0x238>
+ .byte 196,98,125,24,29,143,78,0,0 // vbroadcastss 0x4e8f(%rip),%ymm11 # 6804 <_sk_callback_avx+0x238>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,173,76,0,0 // vbroadcastss 0x4cad(%rip),%ymm11 # 6630 <_sk_callback_avx+0x234>
+ .byte 196,98,125,24,29,125,78,0,0 // vbroadcastss 0x4e7d(%rip),%ymm11 # 6800 <_sk_callback_avx+0x234>
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
.byte 196,227,117,74,202,224 // vblendvps %ymm14,%ymm2,%ymm1,%ymm1
- .byte 196,226,125,24,21,149,76,0,0 // vbroadcastss 0x4c95(%rip),%ymm2 # 662c <_sk_callback_avx+0x230>
+ .byte 196,226,125,24,21,101,78,0,0 // vbroadcastss 0x4e65(%rip),%ymm2 # 67fc <_sk_callback_avx+0x230>
.byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14
.byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2
.byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0
.byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2
.byte 196,227,117,74,194,0 // vblendvps %ymm0,%ymm2,%ymm1,%ymm0
.byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1
- .byte 196,98,125,24,37,124,76,0,0 // vbroadcastss 0x4c7c(%rip),%ymm12 # 663c <_sk_callback_avx+0x240>
+ .byte 196,98,125,24,37,76,78,0,0 // vbroadcastss 0x4e4c(%rip),%ymm12 # 680c <_sk_callback_avx+0x240>
.byte 196,193,116,89,212 // vmulps %ymm12,%ymm1,%ymm2
.byte 197,28,194,226,1 // vcmpltps %ymm2,%ymm12,%ymm12
.byte 196,65,36,92,216 // vsubps %ymm8,%ymm11,%ymm11
@@ -16711,7 +16865,7 @@ _sk_rgb_to_hsl_avx:
.byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
.byte 196,195,125,74,198,128 // vblendvps %ymm8,%ymm14,%ymm0,%ymm0
.byte 196,195,117,74,206,128 // vblendvps %ymm8,%ymm14,%ymm1,%ymm1
- .byte 196,98,125,24,5,63,76,0,0 // vbroadcastss 0x4c3f(%rip),%ymm8 # 6638 <_sk_callback_avx+0x23c>
+ .byte 196,98,125,24,5,15,78,0,0 // vbroadcastss 0x4e0f(%rip),%ymm8 # 6808 <_sk_callback_avx+0x23c>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -16728,7 +16882,7 @@ _sk_hsl_to_rgb_avx:
.byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp)
.byte 197,252,40,225 // vmovaps %ymm1,%ymm4
.byte 197,252,40,216 // vmovaps %ymm0,%ymm3
- .byte 196,98,125,24,5,12,76,0,0 // vbroadcastss 0x4c0c(%rip),%ymm8 # 6640 <_sk_callback_avx+0x244>
+ .byte 196,98,125,24,5,220,77,0,0 // vbroadcastss 0x4ddc(%rip),%ymm8 # 6810 <_sk_callback_avx+0x244>
.byte 197,60,194,202,2 // vcmpleps %ymm2,%ymm8,%ymm9
.byte 197,92,89,210 // vmulps %ymm2,%ymm4,%ymm10
.byte 196,65,92,92,218 // vsubps %ymm10,%ymm4,%ymm11
@@ -16736,23 +16890,23 @@ _sk_hsl_to_rgb_avx:
.byte 197,52,88,210 // vaddps %ymm2,%ymm9,%ymm10
.byte 197,108,88,202 // vaddps %ymm2,%ymm2,%ymm9
.byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,24,29,230,75,0,0 // vbroadcastss 0x4be6(%rip),%ymm11 # 6644 <_sk_callback_avx+0x248>
+ .byte 196,98,125,24,29,182,77,0,0 // vbroadcastss 0x4db6(%rip),%ymm11 # 6814 <_sk_callback_avx+0x248>
.byte 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11
.byte 196,67,125,8,227,1 // vroundps $0x1,%ymm11,%ymm12
.byte 196,65,36,92,252 // vsubps %ymm12,%ymm11,%ymm15
.byte 196,65,44,92,217 // vsubps %ymm9,%ymm10,%ymm11
- .byte 196,98,125,24,37,208,75,0,0 // vbroadcastss 0x4bd0(%rip),%ymm12 # 664c <_sk_callback_avx+0x250>
+ .byte 196,98,125,24,37,160,77,0,0 // vbroadcastss 0x4da0(%rip),%ymm12 # 681c <_sk_callback_avx+0x250>
.byte 196,193,4,89,196 // vmulps %ymm12,%ymm15,%ymm0
- .byte 196,98,125,24,45,198,75,0,0 // vbroadcastss 0x4bc6(%rip),%ymm13 # 6650 <_sk_callback_avx+0x254>
+ .byte 196,98,125,24,45,150,77,0,0 // vbroadcastss 0x4d96(%rip),%ymm13 # 6820 <_sk_callback_avx+0x254>
.byte 197,20,92,240 // vsubps %ymm0,%ymm13,%ymm14
.byte 196,65,36,89,246 // vmulps %ymm14,%ymm11,%ymm14
.byte 196,65,52,88,246 // vaddps %ymm14,%ymm9,%ymm14
- .byte 196,226,125,24,13,167,75,0,0 // vbroadcastss 0x4ba7(%rip),%ymm1 # 6648 <_sk_callback_avx+0x24c>
+ .byte 196,226,125,24,13,119,77,0,0 // vbroadcastss 0x4d77(%rip),%ymm1 # 6818 <_sk_callback_avx+0x24c>
.byte 196,193,116,194,255,2 // vcmpleps %ymm15,%ymm1,%ymm7
.byte 196,195,13,74,249,112 // vblendvps %ymm7,%ymm9,%ymm14,%ymm7
.byte 196,65,60,194,247,2 // vcmpleps %ymm15,%ymm8,%ymm14
.byte 196,227,45,74,255,224 // vblendvps %ymm14,%ymm7,%ymm10,%ymm7
- .byte 196,98,125,24,53,146,75,0,0 // vbroadcastss 0x4b92(%rip),%ymm14 # 6654 <_sk_callback_avx+0x258>
+ .byte 196,98,125,24,53,98,77,0,0 // vbroadcastss 0x4d62(%rip),%ymm14 # 6824 <_sk_callback_avx+0x258>
.byte 196,65,12,194,255,2 // vcmpleps %ymm15,%ymm14,%ymm15
.byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0
.byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0
@@ -16771,7 +16925,7 @@ _sk_hsl_to_rgb_avx:
.byte 197,164,89,247 // vmulps %ymm7,%ymm11,%ymm6
.byte 197,180,88,246 // vaddps %ymm6,%ymm9,%ymm6
.byte 196,227,77,74,237,0 // vblendvps %ymm0,%ymm5,%ymm6,%ymm5
- .byte 196,226,125,24,5,52,75,0,0 // vbroadcastss 0x4b34(%rip),%ymm0 # 6658 <_sk_callback_avx+0x25c>
+ .byte 196,226,125,24,5,4,77,0,0 // vbroadcastss 0x4d04(%rip),%ymm0 # 6828 <_sk_callback_avx+0x25c>
.byte 197,228,88,192 // vaddps %ymm0,%ymm3,%ymm0
.byte 196,227,125,8,216,1 // vroundps $0x1,%ymm0,%ymm3
.byte 197,252,92,195 // vsubps %ymm3,%ymm0,%ymm0
@@ -16818,41 +16972,73 @@ HIDDEN _sk_scale_u8_avx
.globl _sk_scale_u8_avx
FUNCTION(_sk_scale_u8_avx)
_sk_scale_u8_avx:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,68 // jne 1c1b <_sk_scale_u8_avx+0x54>
- .byte 197,122,126,0 // vmovq (%rax),%xmm8
- .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9
- .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8
- .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8
- .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
+ .byte 117,74 // jne 1c1b <_sk_scale_u8_avx+0x54>
+ .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8
+ .byte 197,57,219,5,129,79,0,0 // vpand 0x4f81(%rip),%xmm8,%xmm8 # 6b60 <_sk_callback_avx+0x594>
+ .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
+ .byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9
+ .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8
+ .byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,93,74,0,0 // vbroadcastss 0x4a5d(%rip),%ymm9 # 665c <_sk_callback_avx+0x260>
+ .byte 196,98,125,24,13,42,76,0,0 // vbroadcastss 0x4c2a(%rip),%ymm9 # 682c <_sk_callback_avx+0x260>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
.byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
.byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
.byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 76,137,201 // mov %r9,%rcx
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 1c24 <_sk_scale_u8_avx+0x5d>
- .byte 196,65,249,110,194 // vmovq %r10,%xmm8
- .byte 91 // pop %rbx
- .byte 235,154 // jmp 1bdb <_sk_scale_u8_avx+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,167 // ja 1bd7 <_sk_scale_u8_avx+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,97,0,0,0 // lea 0x61(%rip),%r10 # 1c9c <_sk_scale_u8_avx+0xd5>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 197,57,196,192,6 // vpinsrw $0x6,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,5 // vpinsrw $0x5,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,4 // vpinsrw $0x4,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,3 // vpinsrw $0x3,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,2 // vpinsrw $0x2,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,1 // vpinsrw $0x1,%eax,%xmm8,%xmm8
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,57,196,192,0 // vpinsrw $0x0,%eax,%xmm8,%xmm8
+ .byte 233,61,255,255,255 // jmpq 1bd7 <_sk_scale_u8_avx+0x10>
+ .byte 102,144 // xchg %ax,%ax
+ .byte 239 // out %eax,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,228 // jmpq *%rsp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 217,255 // fcos
+ .byte 255 // (bad)
+ .byte 255,206 // dec %esi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,195 // inc %ebx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 184,255,255,255,168 // mov $0xa8ffffff,%eax
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_lerp_1_float_avx
.globl _sk_lerp_1_float_avx
@@ -16879,19 +17065,18 @@ HIDDEN _sk_lerp_u8_avx
.globl _sk_lerp_u8_avx
FUNCTION(_sk_lerp_u8_avx)
_sk_lerp_u8_avx:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,104 // jne 1cf8 <_sk_lerp_u8_avx+0x78>
- .byte 197,122,126,0 // vmovq (%rax),%xmm8
- .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9
- .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8
- .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8
- .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
+ .byte 117,110 // jne 1d6f <_sk_lerp_u8_avx+0x78>
+ .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8
+ .byte 197,57,219,5,97,78,0,0 // vpand 0x4e61(%rip),%xmm8,%xmm8 # 6b70 <_sk_callback_avx+0x5a4>
+ .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
+ .byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9
+ .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8
+ .byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,168,73,0,0 // vbroadcastss 0x49a8(%rip),%ymm9 # 6660 <_sk_callback_avx+0x264>
+ .byte 196,98,125,24,13,254,74,0,0 // vbroadcastss 0x4afe(%rip),%ymm9 # 6830 <_sk_callback_avx+0x264>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
@@ -16906,22 +17091,55 @@ _sk_lerp_u8_avx:
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 76,137,201 // mov %r9,%rcx
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 1d01 <_sk_lerp_u8_avx+0x81>
- .byte 196,65,249,110,194 // vmovq %r10,%xmm8
- .byte 91 // pop %rbx
- .byte 233,115,255,255,255 // jmpq 1c94 <_sk_lerp_u8_avx+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,131 // ja 1d07 <_sk_lerp_u8_avx+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,97,0,0,0 // lea 0x61(%rip),%r10 # 1df0 <_sk_lerp_u8_avx+0xf9>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
+ .byte 197,57,196,192,6 // vpinsrw $0x6,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,5 // vpinsrw $0x5,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,4 // vpinsrw $0x4,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,3 // vpinsrw $0x3,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,2 // vpinsrw $0x2,%eax,%xmm8,%xmm8
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,57,196,192,1 // vpinsrw $0x1,%eax,%xmm8,%xmm8
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,57,196,192,0 // vpinsrw $0x0,%eax,%xmm8,%xmm8
+ .byte 233,25,255,255,255 // jmpq 1d07 <_sk_lerp_u8_avx+0x10>
+ .byte 102,144 // xchg %ax,%ax
+ .byte 239 // out %eax,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,228 // jmpq *%rsp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 217,255 // fcos
+ .byte 255 // (bad)
+ .byte 255,206 // dec %esi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,195 // inc %ebx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 184,255,255,255,168 // mov $0xa8ffffff,%eax
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_lerp_565_avx
.globl _sk_lerp_565_avx
@@ -16930,26 +17148,26 @@ _sk_lerp_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,208,0,0,0 // jne 1dff <_sk_lerp_565_avx+0xde>
+ .byte 15,133,208,0,0,0 // jne 1eea <_sk_lerp_565_avx+0xde>
.byte 196,65,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9
.byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8
.byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,17,73,0,0 // vbroadcastss 0x4911(%rip),%ymm9 # 6664 <_sk_callback_avx+0x268>
+ .byte 196,98,125,24,13,246,73,0,0 // vbroadcastss 0x49f6(%rip),%ymm9 # 6834 <_sk_callback_avx+0x268>
.byte 196,65,60,84,201 // vandps %ymm9,%ymm8,%ymm9
.byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9
- .byte 196,98,125,24,21,2,73,0,0 // vbroadcastss 0x4902(%rip),%ymm10 # 6668 <_sk_callback_avx+0x26c>
+ .byte 196,98,125,24,21,231,73,0,0 // vbroadcastss 0x49e7(%rip),%ymm10 # 6838 <_sk_callback_avx+0x26c>
.byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,24,21,248,72,0,0 // vbroadcastss 0x48f8(%rip),%ymm10 # 666c <_sk_callback_avx+0x270>
+ .byte 196,98,125,24,21,221,73,0,0 // vbroadcastss 0x49dd(%rip),%ymm10 # 683c <_sk_callback_avx+0x270>
.byte 196,65,60,84,210 // vandps %ymm10,%ymm8,%ymm10
.byte 196,65,124,91,210 // vcvtdq2ps %ymm10,%ymm10
- .byte 196,98,125,24,29,233,72,0,0 // vbroadcastss 0x48e9(%rip),%ymm11 # 6670 <_sk_callback_avx+0x274>
+ .byte 196,98,125,24,29,206,73,0,0 // vbroadcastss 0x49ce(%rip),%ymm11 # 6840 <_sk_callback_avx+0x274>
.byte 196,65,44,89,211 // vmulps %ymm11,%ymm10,%ymm10
- .byte 196,98,125,24,29,223,72,0,0 // vbroadcastss 0x48df(%rip),%ymm11 # 6674 <_sk_callback_avx+0x278>
+ .byte 196,98,125,24,29,196,73,0,0 // vbroadcastss 0x49c4(%rip),%ymm11 # 6844 <_sk_callback_avx+0x278>
.byte 196,65,60,84,195 // vandps %ymm11,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,29,208,72,0,0 // vbroadcastss 0x48d0(%rip),%ymm11 # 6678 <_sk_callback_avx+0x27c>
+ .byte 196,98,125,24,29,181,73,0,0 // vbroadcastss 0x49b5(%rip),%ymm11 # 6848 <_sk_callback_avx+0x27c>
.byte 196,65,60,89,195 // vmulps %ymm11,%ymm8,%ymm8
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
@@ -16976,9 +17194,9 @@ _sk_lerp_565_avx:
.byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 15,135,29,255,255,255 // ja 1d35 <_sk_lerp_565_avx+0x14>
+ .byte 15,135,29,255,255,255 // ja 1e20 <_sk_lerp_565_avx+0x14>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,77,0,0,0 // lea 0x4d(%rip),%r10 # 1e70 <_sk_lerp_565_avx+0x14f>
+ .byte 76,141,21,74,0,0,0 // lea 0x4a(%rip),%r10 # 1f58 <_sk_lerp_565_avx+0x14c>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -16990,26 +17208,28 @@ _sk_lerp_565_avx:
.byte 196,65,57,196,68,83,4,2 // vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm8,%xmm8
.byte 196,65,57,196,68,83,2,1 // vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm8,%xmm8
.byte 196,65,57,196,4,83,0 // vpinsrw $0x0,(%r11,%rdx,2),%xmm8,%xmm8
- .byte 233,200,254,255,255 // jmpq 1d35 <_sk_lerp_565_avx+0x14>
- .byte 15,31,0 // nopl (%rax)
- .byte 241 // icebp
+ .byte 233,200,254,255,255 // jmpq 1e20 <_sk_lerp_565_avx+0x14>
+ .byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 233,255,255,255,225 // jmpq ffffffffe2001e78 <_sk_callback_avx+0xffffffffe1ffba7c>
+ .byte 236 // in (%dx),%al
.byte 255 // (bad)
.byte 255 // (bad)
+ .byte 255,228 // jmpq *%rsp
.byte 255 // (bad)
- .byte 217,255 // fcos
.byte 255 // (bad)
- .byte 255,209 // callq *%rcx
.byte 255 // (bad)
+ .byte 220,255 // fdivr %st,%st(7)
+ .byte 255 // (bad)
+ .byte 255,212 // callq *%rsp
.byte 255 // (bad)
- .byte 255,201 // dec %ecx
.byte 255 // (bad)
+ .byte 255,204 // dec %esp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 188 // .byte 0xbc
+ .byte 255 // (bad)
+ .byte 191 // .byte 0xbf
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -17025,9 +17245,9 @@ _sk_load_tables_avx:
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 76,3,16 // add (%rax),%r10
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,248,1,0,0 // jne 20a4 <_sk_load_tables_avx+0x218>
+ .byte 15,133,248,1,0,0 // jne 218c <_sk_load_tables_avx+0x218>
.byte 196,65,124,16,18 // vmovups (%r10),%ymm10
- .byte 197,124,40,13,135,75,0,0 // vmovaps 0x4b87(%rip),%ymm9 # 6a40 <_sk_callback_avx+0x644>
+ .byte 197,124,40,13,159,76,0,0 // vmovaps 0x4c9f(%rip),%ymm9 # 6c40 <_sk_callback_avx+0x674>
.byte 196,193,44,84,201 // vandps %ymm9,%ymm10,%ymm1
.byte 196,227,125,25,200,1 // vextractf128 $0x1,%ymm1,%xmm0
.byte 196,193,249,126,195 // vmovq %xmm0,%r11
@@ -17119,7 +17339,7 @@ _sk_load_tables_avx:
.byte 196,193,65,114,208,24 // vpsrld $0x18,%xmm8,%xmm7
.byte 196,227,101,24,223,1 // vinsertf128 $0x1,%xmm7,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,226,125,24,61,234,69,0,0 // vbroadcastss 0x45ea(%rip),%ymm7 # 667c <_sk_callback_avx+0x280>
+ .byte 196,226,125,24,61,210,70,0,0 // vbroadcastss 0x46d2(%rip),%ymm7 # 684c <_sk_callback_avx+0x280>
.byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,201 // mov %r9,%rcx
@@ -17133,13 +17353,13 @@ _sk_load_tables_avx:
.byte 73,211,235 // shr %cl,%r11
.byte 196,193,249,110,195 // vmovq %r11,%xmm0
.byte 196,226,121,48,192 // vpmovzxbw %xmm0,%xmm0
- .byte 196,226,121,0,13,196,72,0,0 // vpshufb 0x48c4(%rip),%xmm0,%xmm1 # 6990 <_sk_callback_avx+0x594>
+ .byte 196,226,121,0,13,204,73,0,0 // vpshufb 0x49cc(%rip),%xmm0,%xmm1 # 6b80 <_sk_callback_avx+0x5b4>
.byte 196,226,121,33,201 // vpmovsxbd %xmm1,%xmm1
- .byte 196,226,121,0,5,198,72,0,0 // vpshufb 0x48c6(%rip),%xmm0,%xmm0 # 69a0 <_sk_callback_avx+0x5a4>
+ .byte 196,226,121,0,5,206,73,0,0 // vpshufb 0x49ce(%rip),%xmm0,%xmm0 # 6b90 <_sk_callback_avx+0x5c4>
.byte 196,226,121,33,192 // vpmovsxbd %xmm0,%xmm0
.byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
.byte 196,66,125,44,18 // vmaskmovps (%r10),%ymm0,%ymm10
- .byte 233,194,253,255,255 // jmpq 1eb1 <_sk_load_tables_avx+0x25>
+ .byte 233,194,253,255,255 // jmpq 1f99 <_sk_load_tables_avx+0x25>
HIDDEN _sk_load_tables_u16_be_avx
.globl _sk_load_tables_u16_be_avx
@@ -17150,7 +17370,7 @@ _sk_load_tables_u16_be_avx:
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 77,133,192 // test %r8,%r8
.byte 197,252,17,124,36,200 // vmovups %ymm7,-0x38(%rsp)
- .byte 15,133,84,2,0,0 // jne 235f <_sk_load_tables_u16_be_avx+0x270>
+ .byte 15,133,84,2,0,0 // jne 2447 <_sk_load_tables_u16_be_avx+0x270>
.byte 196,1,121,16,4,81 // vmovupd (%r9,%r10,2),%xmm8
.byte 196,129,121,16,84,81,16 // vmovupd 0x10(%r9,%r10,2),%xmm2
.byte 196,129,121,16,92,81,32 // vmovupd 0x20(%r9,%r10,2),%xmm3
@@ -17165,7 +17385,7 @@ _sk_load_tables_u16_be_avx:
.byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11
.byte 197,177,108,200 // vpunpcklqdq %xmm0,%xmm9,%xmm1
.byte 197,49,109,224 // vpunpckhqdq %xmm0,%xmm9,%xmm12
- .byte 197,121,111,21,88,72,0,0 // vmovdqa 0x4858(%rip),%xmm10 # 69b0 <_sk_callback_avx+0x5b4>
+ .byte 197,121,111,21,96,73,0,0 // vmovdqa 0x4960(%rip),%xmm10 # 6ba0 <_sk_callback_avx+0x5d4>
.byte 196,193,113,219,202 // vpand %xmm10,%xmm1,%xmm1
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 196,193,113,105,209 // vpunpckhwd %xmm9,%xmm1,%xmm2
@@ -17263,7 +17483,7 @@ _sk_load_tables_u16_be_avx:
.byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3
.byte 196,227,101,24,223,1 // vinsertf128 $0x1,%xmm7,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,226,125,24,61,47,67,0,0 // vbroadcastss 0x432f(%rip),%ymm7 # 6680 <_sk_callback_avx+0x284>
+ .byte 196,226,125,24,61,23,68,0,0 // vbroadcastss 0x4417(%rip),%ymm7 # 6850 <_sk_callback_avx+0x284>
.byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7
@@ -17271,29 +17491,29 @@ _sk_load_tables_u16_be_avx:
.byte 196,1,123,16,4,81 // vmovsd (%r9,%r10,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,85 // je 23c5 <_sk_load_tables_u16_be_avx+0x2d6>
+ .byte 116,85 // je 24ad <_sk_load_tables_u16_be_avx+0x2d6>
.byte 196,1,57,22,68,81,8 // vmovhpd 0x8(%r9,%r10,2),%xmm8,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,72 // jb 23c5 <_sk_load_tables_u16_be_avx+0x2d6>
+ .byte 114,72 // jb 24ad <_sk_load_tables_u16_be_avx+0x2d6>
.byte 196,129,123,16,84,81,16 // vmovsd 0x10(%r9,%r10,2),%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 116,72 // je 23d2 <_sk_load_tables_u16_be_avx+0x2e3>
+ .byte 116,72 // je 24ba <_sk_load_tables_u16_be_avx+0x2e3>
.byte 196,129,105,22,84,81,24 // vmovhpd 0x18(%r9,%r10,2),%xmm2,%xmm2
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,59 // jb 23d2 <_sk_load_tables_u16_be_avx+0x2e3>
+ .byte 114,59 // jb 24ba <_sk_load_tables_u16_be_avx+0x2e3>
.byte 196,129,123,16,92,81,32 // vmovsd 0x20(%r9,%r10,2),%xmm3
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 15,132,126,253,255,255 // je 2126 <_sk_load_tables_u16_be_avx+0x37>
+ .byte 15,132,126,253,255,255 // je 220e <_sk_load_tables_u16_be_avx+0x37>
.byte 196,129,97,22,92,81,40 // vmovhpd 0x28(%r9,%r10,2),%xmm3,%xmm3
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 15,130,109,253,255,255 // jb 2126 <_sk_load_tables_u16_be_avx+0x37>
+ .byte 15,130,109,253,255,255 // jb 220e <_sk_load_tables_u16_be_avx+0x37>
.byte 196,1,122,126,76,81,48 // vmovq 0x30(%r9,%r10,2),%xmm9
- .byte 233,97,253,255,255 // jmpq 2126 <_sk_load_tables_u16_be_avx+0x37>
+ .byte 233,97,253,255,255 // jmpq 220e <_sk_load_tables_u16_be_avx+0x37>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,84,253,255,255 // jmpq 2126 <_sk_load_tables_u16_be_avx+0x37>
+ .byte 233,84,253,255,255 // jmpq 220e <_sk_load_tables_u16_be_avx+0x37>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,75,253,255,255 // jmpq 2126 <_sk_load_tables_u16_be_avx+0x37>
+ .byte 233,75,253,255,255 // jmpq 220e <_sk_load_tables_u16_be_avx+0x37>
HIDDEN _sk_load_tables_rgb_u16_be_avx
.globl _sk_load_tables_rgb_u16_be_avx
@@ -17305,7 +17525,7 @@ _sk_load_tables_rgb_u16_be_avx:
.byte 77,133,192 // test %r8,%r8
.byte 197,252,17,124,36,200 // vmovups %ymm7,-0x38(%rsp)
.byte 197,252,17,116,36,168 // vmovups %ymm6,-0x58(%rsp)
- .byte 15,133,71,2,0,0 // jne 2640 <_sk_load_tables_rgb_u16_be_avx+0x265>
+ .byte 15,133,71,2,0,0 // jne 2728 <_sk_load_tables_rgb_u16_be_avx+0x265>
.byte 196,129,122,111,4,81 // vmovdqu (%r9,%r10,2),%xmm0
.byte 196,129,122,111,84,81,12 // vmovdqu 0xc(%r9,%r10,2),%xmm2
.byte 196,129,122,111,76,81,24 // vmovdqu 0x18(%r9,%r10,2),%xmm1
@@ -17326,7 +17546,7 @@ _sk_load_tables_rgb_u16_be_avx:
.byte 197,185,108,218 // vpunpcklqdq %xmm2,%xmm8,%xmm3
.byte 197,57,109,218 // vpunpckhqdq %xmm2,%xmm8,%xmm11
.byte 197,121,108,193 // vpunpcklqdq %xmm1,%xmm0,%xmm8
- .byte 197,121,111,13,93,69,0,0 // vmovdqa 0x455d(%rip),%xmm9 # 69c0 <_sk_callback_avx+0x5c4>
+ .byte 197,121,111,13,101,70,0,0 // vmovdqa 0x4665(%rip),%xmm9 # 6bb0 <_sk_callback_avx+0x5e4>
.byte 196,193,97,219,193 // vpand %xmm9,%xmm3,%xmm0
.byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10
.byte 196,193,121,105,202 // vpunpckhwd %xmm10,%xmm0,%xmm1
@@ -17416,50 +17636,50 @@ _sk_load_tables_rgb_u16_be_avx:
.byte 196,195,105,33,211,48 // vinsertps $0x30,%xmm11,%xmm2,%xmm2
.byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,82,64,0,0 // vbroadcastss 0x4052(%rip),%ymm3 # 6684 <_sk_callback_avx+0x288>
+ .byte 196,226,125,24,29,58,65,0,0 // vbroadcastss 0x413a(%rip),%ymm3 # 6854 <_sk_callback_avx+0x288>
.byte 197,252,16,116,36,168 // vmovups -0x58(%rsp),%ymm6
.byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7
.byte 255,224 // jmpq *%rax
.byte 196,129,121,110,4,81 // vmovd (%r9,%r10,2),%xmm0
.byte 196,129,121,196,68,81,4,2 // vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 117,5 // jne 2659 <_sk_load_tables_rgb_u16_be_avx+0x27e>
- .byte 233,212,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 117,5 // jne 2741 <_sk_load_tables_rgb_u16_be_avx+0x27e>
+ .byte 233,212,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
.byte 196,129,121,110,76,81,6 // vmovd 0x6(%r9,%r10,2),%xmm1
.byte 196,1,113,196,68,81,10,2 // vpinsrw $0x2,0xa(%r9,%r10,2),%xmm1,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,26 // jb 2688 <_sk_load_tables_rgb_u16_be_avx+0x2ad>
+ .byte 114,26 // jb 2770 <_sk_load_tables_rgb_u16_be_avx+0x2ad>
.byte 196,129,121,110,76,81,12 // vmovd 0xc(%r9,%r10,2),%xmm1
.byte 196,129,113,196,84,81,16,2 // vpinsrw $0x2,0x10(%r9,%r10,2),%xmm1,%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 117,10 // jne 268d <_sk_load_tables_rgb_u16_be_avx+0x2b2>
- .byte 233,165,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
- .byte 233,160,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 117,10 // jne 2775 <_sk_load_tables_rgb_u16_be_avx+0x2b2>
+ .byte 233,165,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 233,160,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
.byte 196,129,121,110,76,81,18 // vmovd 0x12(%r9,%r10,2),%xmm1
.byte 196,1,113,196,76,81,22,2 // vpinsrw $0x2,0x16(%r9,%r10,2),%xmm1,%xmm9
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,26 // jb 26bc <_sk_load_tables_rgb_u16_be_avx+0x2e1>
+ .byte 114,26 // jb 27a4 <_sk_load_tables_rgb_u16_be_avx+0x2e1>
.byte 196,129,121,110,76,81,24 // vmovd 0x18(%r9,%r10,2),%xmm1
.byte 196,129,113,196,76,81,28,2 // vpinsrw $0x2,0x1c(%r9,%r10,2),%xmm1,%xmm1
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 117,10 // jne 26c1 <_sk_load_tables_rgb_u16_be_avx+0x2e6>
- .byte 233,113,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
- .byte 233,108,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 117,10 // jne 27a9 <_sk_load_tables_rgb_u16_be_avx+0x2e6>
+ .byte 233,113,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 233,108,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
.byte 196,129,121,110,92,81,30 // vmovd 0x1e(%r9,%r10,2),%xmm3
.byte 196,1,97,196,92,81,34,2 // vpinsrw $0x2,0x22(%r9,%r10,2),%xmm3,%xmm11
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,20 // jb 26ea <_sk_load_tables_rgb_u16_be_avx+0x30f>
+ .byte 114,20 // jb 27d2 <_sk_load_tables_rgb_u16_be_avx+0x30f>
.byte 196,129,121,110,92,81,36 // vmovd 0x24(%r9,%r10,2),%xmm3
.byte 196,129,97,196,92,81,40,2 // vpinsrw $0x2,0x28(%r9,%r10,2),%xmm3,%xmm3
- .byte 233,67,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
- .byte 233,62,253,255,255 // jmpq 242d <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 233,67,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
+ .byte 233,62,253,255,255 // jmpq 2515 <_sk_load_tables_rgb_u16_be_avx+0x52>
HIDDEN _sk_byte_tables_avx
.globl _sk_byte_tables_avx
FUNCTION(_sk_byte_tables_avx)
_sk_byte_tables_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,142,63,0,0 // vbroadcastss 0x3f8e(%rip),%ymm8 # 6688 <_sk_callback_avx+0x28c>
+ .byte 196,98,125,24,5,118,64,0,0 // vbroadcastss 0x4076(%rip),%ymm8 # 6858 <_sk_callback_avx+0x28c>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 197,125,91,200 // vcvtps2dq %ymm0,%ymm9
.byte 196,65,249,126,201 // vmovq %xmm9,%r9
@@ -17578,7 +17798,7 @@ _sk_byte_tables_avx:
.byte 196,194,121,49,204 // vpmovzxbd %xmm12,%xmm1
.byte 196,194,121,49,213 // vpmovzxbd %xmm13,%xmm2
.byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
- .byte 196,98,125,24,13,56,61,0,0 // vbroadcastss 0x3d38(%rip),%ymm9 # 668c <_sk_callback_avx+0x290>
+ .byte 196,98,125,24,13,32,62,0,0 // vbroadcastss 0x3e20(%rip),%ymm9 # 685c <_sk_callback_avx+0x290>
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1
@@ -17694,7 +17914,7 @@ _sk_byte_tables_rgb_avx:
.byte 196,194,121,49,203 // vpmovzxbd %xmm11,%xmm1
.byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,13,25,59,0,0 // vbroadcastss 0x3b19(%rip),%ymm9 # 6690 <_sk_callback_avx+0x294>
+ .byte 196,98,125,24,13,1,60,0,0 // vbroadcastss 0x3c01(%rip),%ymm9 # 6860 <_sk_callback_avx+0x294>
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
.byte 196,194,121,49,202 // vpmovzxbd %xmm10,%xmm1
.byte 196,194,121,49,212 // vpmovzxbd %xmm12,%xmm2
@@ -17899,36 +18119,36 @@ _sk_parametric_r_avx:
.byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
.byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
.byte 197,124,91,216 // vcvtdq2ps %ymm0,%ymm11
- .byte 196,98,125,24,37,198,55,0,0 // vbroadcastss 0x37c6(%rip),%ymm12 # 6694 <_sk_callback_avx+0x298>
+ .byte 196,98,125,24,37,174,56,0,0 // vbroadcastss 0x38ae(%rip),%ymm12 # 6864 <_sk_callback_avx+0x298>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,188,55,0,0 // vbroadcastss 0x37bc(%rip),%ymm12 # 6698 <_sk_callback_avx+0x29c>
+ .byte 196,98,125,24,37,164,56,0,0 // vbroadcastss 0x38a4(%rip),%ymm12 # 6868 <_sk_callback_avx+0x29c>
.byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0
- .byte 196,98,125,24,37,178,55,0,0 // vbroadcastss 0x37b2(%rip),%ymm12 # 669c <_sk_callback_avx+0x2a0>
+ .byte 196,98,125,24,37,154,56,0,0 // vbroadcastss 0x389a(%rip),%ymm12 # 686c <_sk_callback_avx+0x2a0>
.byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0
- .byte 196,98,125,24,37,168,55,0,0 // vbroadcastss 0x37a8(%rip),%ymm12 # 66a0 <_sk_callback_avx+0x2a4>
+ .byte 196,98,125,24,37,144,56,0,0 // vbroadcastss 0x3890(%rip),%ymm12 # 6870 <_sk_callback_avx+0x2a4>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,158,55,0,0 // vbroadcastss 0x379e(%rip),%ymm12 # 66a4 <_sk_callback_avx+0x2a8>
+ .byte 196,98,125,24,37,134,56,0,0 // vbroadcastss 0x3886(%rip),%ymm12 # 6874 <_sk_callback_avx+0x2a8>
.byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,143,55,0,0 // vbroadcastss 0x378f(%rip),%ymm12 # 66a8 <_sk_callback_avx+0x2ac>
+ .byte 196,98,125,24,37,119,56,0,0 // vbroadcastss 0x3877(%rip),%ymm12 # 6878 <_sk_callback_avx+0x2ac>
.byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0
- .byte 196,98,125,24,37,133,55,0,0 // vbroadcastss 0x3785(%rip),%ymm12 # 66ac <_sk_callback_avx+0x2b0>
+ .byte 196,98,125,24,37,109,56,0,0 // vbroadcastss 0x386d(%rip),%ymm12 # 687c <_sk_callback_avx+0x2b0>
.byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0
.byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0
.byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
.byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10
.byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10
- .byte 196,98,125,24,29,105,55,0,0 // vbroadcastss 0x3769(%rip),%ymm11 # 66b0 <_sk_callback_avx+0x2b4>
+ .byte 196,98,125,24,29,81,56,0,0 // vbroadcastss 0x3851(%rip),%ymm11 # 6880 <_sk_callback_avx+0x2b4>
.byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,95,55,0,0 // vbroadcastss 0x375f(%rip),%ymm11 # 66b4 <_sk_callback_avx+0x2b8>
+ .byte 196,98,125,24,29,71,56,0,0 // vbroadcastss 0x3847(%rip),%ymm11 # 6884 <_sk_callback_avx+0x2b8>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,80,55,0,0 // vbroadcastss 0x3750(%rip),%ymm11 # 66b8 <_sk_callback_avx+0x2bc>
+ .byte 196,98,125,24,29,56,56,0,0 // vbroadcastss 0x3838(%rip),%ymm11 # 6888 <_sk_callback_avx+0x2bc>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,70,55,0,0 // vbroadcastss 0x3746(%rip),%ymm11 # 66bc <_sk_callback_avx+0x2c0>
+ .byte 196,98,125,24,29,46,56,0,0 // vbroadcastss 0x382e(%rip),%ymm11 # 688c <_sk_callback_avx+0x2c0>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
- .byte 196,98,125,24,21,55,55,0,0 // vbroadcastss 0x3737(%rip),%ymm10 # 66c0 <_sk_callback_avx+0x2c4>
+ .byte 196,98,125,24,21,31,56,0,0 // vbroadcastss 0x381f(%rip),%ymm10 # 6890 <_sk_callback_avx+0x2c4>
.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
@@ -17936,7 +18156,7 @@ _sk_parametric_r_avx:
.byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,14,55,0,0 // vbroadcastss 0x370e(%rip),%ymm8 # 66c4 <_sk_callback_avx+0x2c8>
+ .byte 196,98,125,24,5,246,55,0,0 // vbroadcastss 0x37f6(%rip),%ymm8 # 6894 <_sk_callback_avx+0x2c8>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -17958,36 +18178,36 @@ _sk_parametric_g_avx:
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
.byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
.byte 197,124,91,217 // vcvtdq2ps %ymm1,%ymm11
- .byte 196,98,125,24,37,191,54,0,0 // vbroadcastss 0x36bf(%rip),%ymm12 # 66c8 <_sk_callback_avx+0x2cc>
+ .byte 196,98,125,24,37,167,55,0,0 // vbroadcastss 0x37a7(%rip),%ymm12 # 6898 <_sk_callback_avx+0x2cc>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,181,54,0,0 // vbroadcastss 0x36b5(%rip),%ymm12 # 66cc <_sk_callback_avx+0x2d0>
+ .byte 196,98,125,24,37,157,55,0,0 // vbroadcastss 0x379d(%rip),%ymm12 # 689c <_sk_callback_avx+0x2d0>
.byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1
- .byte 196,98,125,24,37,171,54,0,0 // vbroadcastss 0x36ab(%rip),%ymm12 # 66d0 <_sk_callback_avx+0x2d4>
+ .byte 196,98,125,24,37,147,55,0,0 // vbroadcastss 0x3793(%rip),%ymm12 # 68a0 <_sk_callback_avx+0x2d4>
.byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1
- .byte 196,98,125,24,37,161,54,0,0 // vbroadcastss 0x36a1(%rip),%ymm12 # 66d4 <_sk_callback_avx+0x2d8>
+ .byte 196,98,125,24,37,137,55,0,0 // vbroadcastss 0x3789(%rip),%ymm12 # 68a4 <_sk_callback_avx+0x2d8>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,151,54,0,0 // vbroadcastss 0x3697(%rip),%ymm12 # 66d8 <_sk_callback_avx+0x2dc>
+ .byte 196,98,125,24,37,127,55,0,0 // vbroadcastss 0x377f(%rip),%ymm12 # 68a8 <_sk_callback_avx+0x2dc>
.byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,136,54,0,0 // vbroadcastss 0x3688(%rip),%ymm12 # 66dc <_sk_callback_avx+0x2e0>
+ .byte 196,98,125,24,37,112,55,0,0 // vbroadcastss 0x3770(%rip),%ymm12 # 68ac <_sk_callback_avx+0x2e0>
.byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1
- .byte 196,98,125,24,37,126,54,0,0 // vbroadcastss 0x367e(%rip),%ymm12 # 66e0 <_sk_callback_avx+0x2e4>
+ .byte 196,98,125,24,37,102,55,0,0 // vbroadcastss 0x3766(%rip),%ymm12 # 68b0 <_sk_callback_avx+0x2e4>
.byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1
.byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1
.byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
.byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10
.byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
- .byte 196,98,125,24,29,98,54,0,0 // vbroadcastss 0x3662(%rip),%ymm11 # 66e4 <_sk_callback_avx+0x2e8>
+ .byte 196,98,125,24,29,74,55,0,0 // vbroadcastss 0x374a(%rip),%ymm11 # 68b4 <_sk_callback_avx+0x2e8>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,88,54,0,0 // vbroadcastss 0x3658(%rip),%ymm11 # 66e8 <_sk_callback_avx+0x2ec>
+ .byte 196,98,125,24,29,64,55,0,0 // vbroadcastss 0x3740(%rip),%ymm11 # 68b8 <_sk_callback_avx+0x2ec>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,73,54,0,0 // vbroadcastss 0x3649(%rip),%ymm11 # 66ec <_sk_callback_avx+0x2f0>
+ .byte 196,98,125,24,29,49,55,0,0 // vbroadcastss 0x3731(%rip),%ymm11 # 68bc <_sk_callback_avx+0x2f0>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,63,54,0,0 // vbroadcastss 0x363f(%rip),%ymm11 # 66f0 <_sk_callback_avx+0x2f4>
+ .byte 196,98,125,24,29,39,55,0,0 // vbroadcastss 0x3727(%rip),%ymm11 # 68c0 <_sk_callback_avx+0x2f4>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
- .byte 196,98,125,24,21,48,54,0,0 // vbroadcastss 0x3630(%rip),%ymm10 # 66f4 <_sk_callback_avx+0x2f8>
+ .byte 196,98,125,24,21,24,55,0,0 // vbroadcastss 0x3718(%rip),%ymm10 # 68c4 <_sk_callback_avx+0x2f8>
.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
@@ -17995,7 +18215,7 @@ _sk_parametric_g_avx:
.byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1
- .byte 196,98,125,24,5,7,54,0,0 // vbroadcastss 0x3607(%rip),%ymm8 # 66f8 <_sk_callback_avx+0x2fc>
+ .byte 196,98,125,24,5,239,54,0,0 // vbroadcastss 0x36ef(%rip),%ymm8 # 68c8 <_sk_callback_avx+0x2fc>
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18017,36 +18237,36 @@ _sk_parametric_b_avx:
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
.byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
.byte 197,124,91,218 // vcvtdq2ps %ymm2,%ymm11
- .byte 196,98,125,24,37,184,53,0,0 // vbroadcastss 0x35b8(%rip),%ymm12 # 66fc <_sk_callback_avx+0x300>
+ .byte 196,98,125,24,37,160,54,0,0 // vbroadcastss 0x36a0(%rip),%ymm12 # 68cc <_sk_callback_avx+0x300>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,174,53,0,0 // vbroadcastss 0x35ae(%rip),%ymm12 # 6700 <_sk_callback_avx+0x304>
+ .byte 196,98,125,24,37,150,54,0,0 // vbroadcastss 0x3696(%rip),%ymm12 # 68d0 <_sk_callback_avx+0x304>
.byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2
- .byte 196,98,125,24,37,164,53,0,0 // vbroadcastss 0x35a4(%rip),%ymm12 # 6704 <_sk_callback_avx+0x308>
+ .byte 196,98,125,24,37,140,54,0,0 // vbroadcastss 0x368c(%rip),%ymm12 # 68d4 <_sk_callback_avx+0x308>
.byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2
- .byte 196,98,125,24,37,154,53,0,0 // vbroadcastss 0x359a(%rip),%ymm12 # 6708 <_sk_callback_avx+0x30c>
+ .byte 196,98,125,24,37,130,54,0,0 // vbroadcastss 0x3682(%rip),%ymm12 # 68d8 <_sk_callback_avx+0x30c>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,144,53,0,0 // vbroadcastss 0x3590(%rip),%ymm12 # 670c <_sk_callback_avx+0x310>
+ .byte 196,98,125,24,37,120,54,0,0 // vbroadcastss 0x3678(%rip),%ymm12 # 68dc <_sk_callback_avx+0x310>
.byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,129,53,0,0 // vbroadcastss 0x3581(%rip),%ymm12 # 6710 <_sk_callback_avx+0x314>
+ .byte 196,98,125,24,37,105,54,0,0 // vbroadcastss 0x3669(%rip),%ymm12 # 68e0 <_sk_callback_avx+0x314>
.byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2
- .byte 196,98,125,24,37,119,53,0,0 // vbroadcastss 0x3577(%rip),%ymm12 # 6714 <_sk_callback_avx+0x318>
+ .byte 196,98,125,24,37,95,54,0,0 // vbroadcastss 0x365f(%rip),%ymm12 # 68e4 <_sk_callback_avx+0x318>
.byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2
.byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2
.byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
.byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10
.byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10
- .byte 196,98,125,24,29,91,53,0,0 // vbroadcastss 0x355b(%rip),%ymm11 # 6718 <_sk_callback_avx+0x31c>
+ .byte 196,98,125,24,29,67,54,0,0 // vbroadcastss 0x3643(%rip),%ymm11 # 68e8 <_sk_callback_avx+0x31c>
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,81,53,0,0 // vbroadcastss 0x3551(%rip),%ymm11 # 671c <_sk_callback_avx+0x320>
+ .byte 196,98,125,24,29,57,54,0,0 // vbroadcastss 0x3639(%rip),%ymm11 # 68ec <_sk_callback_avx+0x320>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,66,53,0,0 // vbroadcastss 0x3542(%rip),%ymm11 # 6720 <_sk_callback_avx+0x324>
+ .byte 196,98,125,24,29,42,54,0,0 // vbroadcastss 0x362a(%rip),%ymm11 # 68f0 <_sk_callback_avx+0x324>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,56,53,0,0 // vbroadcastss 0x3538(%rip),%ymm11 # 6724 <_sk_callback_avx+0x328>
+ .byte 196,98,125,24,29,32,54,0,0 // vbroadcastss 0x3620(%rip),%ymm11 # 68f4 <_sk_callback_avx+0x328>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
- .byte 196,98,125,24,21,41,53,0,0 // vbroadcastss 0x3529(%rip),%ymm10 # 6728 <_sk_callback_avx+0x32c>
+ .byte 196,98,125,24,21,17,54,0,0 // vbroadcastss 0x3611(%rip),%ymm10 # 68f8 <_sk_callback_avx+0x32c>
.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
@@ -18054,7 +18274,7 @@ _sk_parametric_b_avx:
.byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,0,53,0,0 // vbroadcastss 0x3500(%rip),%ymm8 # 672c <_sk_callback_avx+0x330>
+ .byte 196,98,125,24,5,232,53,0,0 // vbroadcastss 0x35e8(%rip),%ymm8 # 68fc <_sk_callback_avx+0x330>
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18076,36 +18296,36 @@ _sk_parametric_a_avx:
.byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
.byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
.byte 197,124,91,219 // vcvtdq2ps %ymm3,%ymm11
- .byte 196,98,125,24,37,177,52,0,0 // vbroadcastss 0x34b1(%rip),%ymm12 # 6730 <_sk_callback_avx+0x334>
+ .byte 196,98,125,24,37,153,53,0,0 // vbroadcastss 0x3599(%rip),%ymm12 # 6900 <_sk_callback_avx+0x334>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,167,52,0,0 // vbroadcastss 0x34a7(%rip),%ymm12 # 6734 <_sk_callback_avx+0x338>
+ .byte 196,98,125,24,37,143,53,0,0 // vbroadcastss 0x358f(%rip),%ymm12 # 6904 <_sk_callback_avx+0x338>
.byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3
- .byte 196,98,125,24,37,157,52,0,0 // vbroadcastss 0x349d(%rip),%ymm12 # 6738 <_sk_callback_avx+0x33c>
+ .byte 196,98,125,24,37,133,53,0,0 // vbroadcastss 0x3585(%rip),%ymm12 # 6908 <_sk_callback_avx+0x33c>
.byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3
- .byte 196,98,125,24,37,147,52,0,0 // vbroadcastss 0x3493(%rip),%ymm12 # 673c <_sk_callback_avx+0x340>
+ .byte 196,98,125,24,37,123,53,0,0 // vbroadcastss 0x357b(%rip),%ymm12 # 690c <_sk_callback_avx+0x340>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,137,52,0,0 // vbroadcastss 0x3489(%rip),%ymm12 # 6740 <_sk_callback_avx+0x344>
+ .byte 196,98,125,24,37,113,53,0,0 // vbroadcastss 0x3571(%rip),%ymm12 # 6910 <_sk_callback_avx+0x344>
.byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,122,52,0,0 // vbroadcastss 0x347a(%rip),%ymm12 # 6744 <_sk_callback_avx+0x348>
+ .byte 196,98,125,24,37,98,53,0,0 // vbroadcastss 0x3562(%rip),%ymm12 # 6914 <_sk_callback_avx+0x348>
.byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3
- .byte 196,98,125,24,37,112,52,0,0 // vbroadcastss 0x3470(%rip),%ymm12 # 6748 <_sk_callback_avx+0x34c>
+ .byte 196,98,125,24,37,88,53,0,0 // vbroadcastss 0x3558(%rip),%ymm12 # 6918 <_sk_callback_avx+0x34c>
.byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3
.byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3
.byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3
.byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10
.byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10
- .byte 196,98,125,24,29,84,52,0,0 // vbroadcastss 0x3454(%rip),%ymm11 # 674c <_sk_callback_avx+0x350>
+ .byte 196,98,125,24,29,60,53,0,0 // vbroadcastss 0x353c(%rip),%ymm11 # 691c <_sk_callback_avx+0x350>
.byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,74,52,0,0 // vbroadcastss 0x344a(%rip),%ymm11 # 6750 <_sk_callback_avx+0x354>
+ .byte 196,98,125,24,29,50,53,0,0 // vbroadcastss 0x3532(%rip),%ymm11 # 6920 <_sk_callback_avx+0x354>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,59,52,0,0 // vbroadcastss 0x343b(%rip),%ymm11 # 6754 <_sk_callback_avx+0x358>
+ .byte 196,98,125,24,29,35,53,0,0 // vbroadcastss 0x3523(%rip),%ymm11 # 6924 <_sk_callback_avx+0x358>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,49,52,0,0 // vbroadcastss 0x3431(%rip),%ymm11 # 6758 <_sk_callback_avx+0x35c>
+ .byte 196,98,125,24,29,25,53,0,0 // vbroadcastss 0x3519(%rip),%ymm11 # 6928 <_sk_callback_avx+0x35c>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3
- .byte 196,98,125,24,21,34,52,0,0 // vbroadcastss 0x3422(%rip),%ymm10 # 675c <_sk_callback_avx+0x360>
+ .byte 196,98,125,24,21,10,53,0,0 // vbroadcastss 0x350a(%rip),%ymm10 # 692c <_sk_callback_avx+0x360>
.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
@@ -18113,7 +18333,7 @@ _sk_parametric_a_avx:
.byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3
- .byte 196,98,125,24,5,249,51,0,0 // vbroadcastss 0x33f9(%rip),%ymm8 # 6760 <_sk_callback_avx+0x364>
+ .byte 196,98,125,24,5,225,52,0,0 // vbroadcastss 0x34e1(%rip),%ymm8 # 6930 <_sk_callback_avx+0x364>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18122,31 +18342,31 @@ HIDDEN _sk_lab_to_xyz_avx
.globl _sk_lab_to_xyz_avx
FUNCTION(_sk_lab_to_xyz_avx)
_sk_lab_to_xyz_avx:
- .byte 196,98,125,24,5,235,51,0,0 // vbroadcastss 0x33eb(%rip),%ymm8 # 6764 <_sk_callback_avx+0x368>
+ .byte 196,98,125,24,5,211,52,0,0 // vbroadcastss 0x34d3(%rip),%ymm8 # 6934 <_sk_callback_avx+0x368>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,225,51,0,0 // vbroadcastss 0x33e1(%rip),%ymm8 # 6768 <_sk_callback_avx+0x36c>
+ .byte 196,98,125,24,5,201,52,0,0 // vbroadcastss 0x34c9(%rip),%ymm8 # 6938 <_sk_callback_avx+0x36c>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,98,125,24,13,215,51,0,0 // vbroadcastss 0x33d7(%rip),%ymm9 # 676c <_sk_callback_avx+0x370>
+ .byte 196,98,125,24,13,191,52,0,0 // vbroadcastss 0x34bf(%rip),%ymm9 # 693c <_sk_callback_avx+0x370>
.byte 196,193,116,88,201 // vaddps %ymm9,%ymm1,%ymm1
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2
- .byte 196,98,125,24,5,195,51,0,0 // vbroadcastss 0x33c3(%rip),%ymm8 # 6770 <_sk_callback_avx+0x374>
+ .byte 196,98,125,24,5,171,52,0,0 // vbroadcastss 0x34ab(%rip),%ymm8 # 6940 <_sk_callback_avx+0x374>
.byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,185,51,0,0 // vbroadcastss 0x33b9(%rip),%ymm8 # 6774 <_sk_callback_avx+0x378>
+ .byte 196,98,125,24,5,161,52,0,0 // vbroadcastss 0x34a1(%rip),%ymm8 # 6944 <_sk_callback_avx+0x378>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,175,51,0,0 // vbroadcastss 0x33af(%rip),%ymm8 # 6778 <_sk_callback_avx+0x37c>
+ .byte 196,98,125,24,5,151,52,0,0 // vbroadcastss 0x3497(%rip),%ymm8 # 6948 <_sk_callback_avx+0x37c>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
.byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1
- .byte 196,98,125,24,5,161,51,0,0 // vbroadcastss 0x33a1(%rip),%ymm8 # 677c <_sk_callback_avx+0x380>
+ .byte 196,98,125,24,5,137,52,0,0 // vbroadcastss 0x3489(%rip),%ymm8 # 694c <_sk_callback_avx+0x380>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 197,252,92,210 // vsubps %ymm2,%ymm0,%ymm2
.byte 197,116,89,193 // vmulps %ymm1,%ymm1,%ymm8
.byte 196,65,116,89,192 // vmulps %ymm8,%ymm1,%ymm8
- .byte 196,98,125,24,13,138,51,0,0 // vbroadcastss 0x338a(%rip),%ymm9 # 6780 <_sk_callback_avx+0x384>
+ .byte 196,98,125,24,13,114,52,0,0 // vbroadcastss 0x3472(%rip),%ymm9 # 6950 <_sk_callback_avx+0x384>
.byte 196,65,52,194,208,1 // vcmpltps %ymm8,%ymm9,%ymm10
- .byte 196,98,125,24,29,127,51,0,0 // vbroadcastss 0x337f(%rip),%ymm11 # 6784 <_sk_callback_avx+0x388>
+ .byte 196,98,125,24,29,103,52,0,0 // vbroadcastss 0x3467(%rip),%ymm11 # 6954 <_sk_callback_avx+0x388>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,37,117,51,0,0 // vbroadcastss 0x3375(%rip),%ymm12 # 6788 <_sk_callback_avx+0x38c>
+ .byte 196,98,125,24,37,93,52,0,0 // vbroadcastss 0x345d(%rip),%ymm12 # 6958 <_sk_callback_avx+0x38c>
.byte 196,193,116,89,204 // vmulps %ymm12,%ymm1,%ymm1
.byte 196,67,117,74,192,160 // vblendvps %ymm10,%ymm8,%ymm1,%ymm8
.byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1
@@ -18161,9 +18381,9 @@ _sk_lab_to_xyz_avx:
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
.byte 196,193,108,89,212 // vmulps %ymm12,%ymm2,%ymm2
.byte 196,227,109,74,208,144 // vblendvps %ymm9,%ymm0,%ymm2,%ymm2
- .byte 196,226,125,24,5,43,51,0,0 // vbroadcastss 0x332b(%rip),%ymm0 # 678c <_sk_callback_avx+0x390>
+ .byte 196,226,125,24,5,19,52,0,0 // vbroadcastss 0x3413(%rip),%ymm0 # 695c <_sk_callback_avx+0x390>
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,34,51,0,0 // vbroadcastss 0x3322(%rip),%ymm8 # 6790 <_sk_callback_avx+0x394>
+ .byte 196,98,125,24,5,10,52,0,0 // vbroadcastss 0x340a(%rip),%ymm8 # 6960 <_sk_callback_avx+0x394>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18172,40 +18392,71 @@ HIDDEN _sk_load_a8_avx
.globl _sk_load_a8_avx
FUNCTION(_sk_load_a8_avx)
_sk_load_a8_avx:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,62 // jne 34c5 <_sk_load_a8_avx+0x4e>
- .byte 197,250,126,0 // vmovq (%rax),%xmm0
- .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
- .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
- .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
- .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
+ .byte 117,66 // jne 35ab <_sk_load_a8_avx+0x4c>
+ .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0
+ .byte 197,249,219,5,73,54,0,0 // vpand 0x3649(%rip),%xmm0,%xmm0 # 6bc0 <_sk_callback_avx+0x5f4>
+ .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 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,230,50,0,0 // vbroadcastss 0x32e6(%rip),%ymm1 # 6794 <_sk_callback_avx+0x398>
+ .byte 196,226,125,24,13,205,51,0,0 // vbroadcastss 0x33cd(%rip),%ymm1 # 6964 <_sk_callback_avx+0x398>
.byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
.byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
.byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
- .byte 76,137,201 // mov %r9,%rcx
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 34ce <_sk_load_a8_avx+0x57>
- .byte 196,193,249,110,194 // vmovq %r10,%xmm0
- .byte 91 // pop %rbx
- .byte 235,160 // jmp 348b <_sk_load_a8_avx+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,176 // ja 356f <_sk_load_a8_avx+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,94,0,0,0 // lea 0x5e(%rip),%r10 # 3628 <_sk_load_a8_avx+0xc9>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,2 // vpinsrw $0x2,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,1 // vpinsrw $0x1,%eax,%xmm0,%xmm0
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,249,196,192,0 // vpinsrw $0x0,%eax,%xmm0,%xmm0
+ .byte 233,71,255,255,255 // jmpq 356f <_sk_load_a8_avx+0x10>
+ .byte 241 // icebp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,230 // jmpq *%rsi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 219,255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,208 // callq *%rax
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,197 // inc %ebp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 186,255,255,255,171 // mov $0xabffffff,%edx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_gather_a8_avx
.globl _sk_gather_a8_avx
@@ -18251,7 +18502,7 @@ _sk_gather_a8_avx:
.byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
.byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,225,49,0,0 // vbroadcastss 0x31e1(%rip),%ymm1 # 6798 <_sk_callback_avx+0x39c>
+ .byte 196,226,125,24,13,88,50,0,0 // vbroadcastss 0x3258(%rip),%ymm1 # 6968 <_sk_callback_avx+0x39c>
.byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
@@ -18265,14 +18516,14 @@ FUNCTION(_sk_store_a8_avx)
_sk_store_a8_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
- .byte 196,98,125,24,5,195,49,0,0 // vbroadcastss 0x31c3(%rip),%ymm8 # 679c <_sk_callback_avx+0x3a0>
+ .byte 196,98,125,24,5,58,50,0,0 // vbroadcastss 0x323a(%rip),%ymm8 # 696c <_sk_callback_avx+0x3a0>
.byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8
.byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
.byte 77,133,192 // test %r8,%r8
- .byte 117,10 // jne 3602 <_sk_store_a8_avx+0x37>
+ .byte 117,10 // jne 375b <_sk_store_a8_avx+0x37>
.byte 196,65,123,17,4,19 // vmovsd %xmm8,(%r11,%rdx,1)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18280,10 +18531,10 @@ _sk_store_a8_avx:
.byte 65,128,225,7 // and $0x7,%r9b
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,236 // ja 35fe <_sk_store_a8_avx+0x33>
+ .byte 119,236 // ja 3757 <_sk_store_a8_avx+0x33>
.byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,66,0,0,0 // lea 0x42(%rip),%r10 # 3664 <_sk_store_a8_avx+0x99>
+ .byte 76,141,21,69,0,0,0 // lea 0x45(%rip),%r10 # 37c0 <_sk_store_a8_avx+0x9c>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -18294,26 +18545,28 @@ _sk_store_a8_avx:
.byte 196,67,121,20,68,19,2,4 // vpextrb $0x4,%xmm8,0x2(%r11,%rdx,1)
.byte 196,67,121,20,68,19,1,2 // vpextrb $0x2,%xmm8,0x1(%r11,%rdx,1)
.byte 196,67,121,20,4,19,0 // vpextrb $0x0,%xmm8,(%r11,%rdx,1)
- .byte 235,154 // jmp 35fe <_sk_store_a8_avx+0x33>
- .byte 247,255 // idiv %edi
+ .byte 235,154 // jmp 3757 <_sk_store_a8_avx+0x33>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 239 // out %eax,(%dx)
+ .byte 255 // (bad)
+ .byte 236 // in (%dx),%al
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,231 // jmpq *%rdi
+ .byte 255,228 // jmpq *%rsp
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 223,255 // (bad)
+ .byte 220,255 // fdivr %st,%st(7)
.byte 255 // (bad)
- .byte 255,215 // callq *%rdi
+ .byte 255,212 // callq *%rsp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,207 // dec %edi
+ .byte 255,204 // dec %esp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,199 // inc %edi
+ .byte 255,196 // inc %esp
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -18322,40 +18575,72 @@ HIDDEN _sk_load_g8_avx
.globl _sk_load_g8_avx
FUNCTION(_sk_load_g8_avx)
_sk_load_g8_avx:
- .byte 73,137,201 // mov %rcx,%r9
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 72,139,0 // mov (%rax),%rax
- .byte 72,1,208 // add %rdx,%rax
+ .byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 117,67 // jne 36d3 <_sk_load_g8_avx+0x53>
- .byte 197,250,126,0 // vmovq (%rax),%xmm0
- .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
- .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
- .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
- .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
+ .byte 117,71 // jne 382d <_sk_load_g8_avx+0x51>
+ .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0
+ .byte 197,249,219,5,220,51,0,0 // vpand 0x33dc(%rip),%xmm0,%xmm0 # 6bd0 <_sk_callback_avx+0x604>
+ .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 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,233,48,0,0 // vbroadcastss 0x30e9(%rip),%ymm1 # 67a0 <_sk_callback_avx+0x3a4>
+ .byte 196,226,125,24,13,92,49,0,0 // vbroadcastss 0x315c(%rip),%ymm1 # 6970 <_sk_callback_avx+0x3a4>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,222,48,0,0 // vbroadcastss 0x30de(%rip),%ymm3 # 67a4 <_sk_callback_avx+0x3a8>
- .byte 76,137,201 // mov %r9,%rcx
+ .byte 196,226,125,24,29,81,49,0,0 // vbroadcastss 0x3151(%rip),%ymm3 # 6974 <_sk_callback_avx+0x3a8>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 255,224 // jmpq *%rax
- .byte 83 // push %rbx
- .byte 49,201 // xor %ecx,%ecx
- .byte 77,137,195 // mov %r8,%r11
- .byte 69,49,210 // xor %r10d,%r10d
- .byte 15,182,24 // movzbl (%rax),%ebx
- .byte 72,255,192 // inc %rax
- .byte 72,211,227 // shl %cl,%rbx
- .byte 73,9,218 // or %rbx,%r10
- .byte 72,131,193,8 // add $0x8,%rcx
- .byte 73,255,203 // dec %r11
- .byte 117,235 // jne 36dc <_sk_load_g8_avx+0x5c>
- .byte 196,193,249,110,194 // vmovq %r10,%xmm0
- .byte 91 // pop %rbx
- .byte 235,155 // jmp 3694 <_sk_load_g8_avx+0x14>
+ .byte 69,137,193 // mov %r8d,%r9d
+ .byte 65,128,225,7 // and $0x7,%r9b
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 65,254,201 // dec %r9b
+ .byte 65,128,249,6 // cmp $0x6,%r9b
+ .byte 119,171 // ja 37ec <_sk_load_g8_avx+0x10>
+ .byte 69,15,182,201 // movzbl %r9b,%r9d
+ .byte 76,141,21,96,0,0,0 // lea 0x60(%rip),%r10 # 38ac <_sk_load_g8_avx+0xd0>
+ .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
+ .byte 76,1,208 // add %r10,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,2 // vpinsrw $0x2,%eax,%xmm0,%xmm0
+ .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax
+ .byte 197,249,196,192,1 // vpinsrw $0x1,%eax,%xmm0,%xmm0
+ .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax
+ .byte 197,249,196,192,0 // vpinsrw $0x0,%eax,%xmm0,%xmm0
+ .byte 233,66,255,255,255 // jmpq 37ec <_sk_load_g8_avx+0x10>
+ .byte 102,144 // xchg %ax,%ax
+ .byte 239 // out %eax,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,228 // jmpq *%rsp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 217,255 // fcos
+ .byte 255 // (bad)
+ .byte 255,206 // dec %esi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,195 // inc %ebx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 184,255,255,255,169 // mov $0xa9ffffff,%eax
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
HIDDEN _sk_gather_g8_avx
.globl _sk_gather_g8_avx
@@ -18401,10 +18686,10 @@ _sk_gather_g8_avx:
.byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
.byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,227,47,0,0 // vbroadcastss 0x2fe3(%rip),%ymm1 # 67a8 <_sk_callback_avx+0x3ac>
+ .byte 196,226,125,24,13,228,47,0,0 // vbroadcastss 0x2fe4(%rip),%ymm1 # 6978 <_sk_callback_avx+0x3ac>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,216,47,0,0 // vbroadcastss 0x2fd8(%rip),%ymm3 # 67ac <_sk_callback_avx+0x3b0>
+ .byte 196,226,125,24,29,217,47,0,0 // vbroadcastss 0x2fd9(%rip),%ymm3 # 697c <_sk_callback_avx+0x3b0>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 255,224 // jmpq *%rax
@@ -18416,9 +18701,9 @@ _sk_gather_i8_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,137,193 // mov %rax,%r9
.byte 77,133,201 // test %r9,%r9
- .byte 116,5 // je 37ed <_sk_gather_i8_avx+0xf>
+ .byte 116,5 // je 39bc <_sk_gather_i8_avx+0xf>
.byte 76,137,200 // mov %r9,%rax
- .byte 235,2 // jmp 37ef <_sk_gather_i8_avx+0x11>
+ .byte 235,2 // jmp 39be <_sk_gather_i8_avx+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 83 // push %rbx
.byte 76,139,16 // mov (%rax),%r10
@@ -18476,10 +18761,10 @@ _sk_gather_i8_avx:
.byte 72,193,235,30 // shr $0x1e,%rbx
.byte 196,195,121,34,28,27,3 // vpinsrd $0x3,(%r11,%rbx,1),%xmm0,%xmm3
.byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
- .byte 197,124,40,21,83,49,0,0 // vmovaps 0x3153(%rip),%ymm10 # 6a60 <_sk_callback_avx+0x664>
+ .byte 197,124,40,21,132,49,0,0 // vmovaps 0x3184(%rip),%ymm10 # 6c60 <_sk_callback_avx+0x694>
.byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,13,145,46,0,0 // vbroadcastss 0x2e91(%rip),%ymm9 # 67b0 <_sk_callback_avx+0x3b4>
+ .byte 196,98,125,24,13,146,46,0,0 // vbroadcastss 0x2e92(%rip),%ymm9 # 6980 <_sk_callback_avx+0x3b4>
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
.byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1
.byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2
@@ -18509,38 +18794,38 @@ _sk_load_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,128,0,0,0 // jne 3a0f <_sk_load_565_avx+0x8e>
+ .byte 15,133,128,0,0,0 // jne 3bde <_sk_load_565_avx+0x8e>
.byte 196,193,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm0
.byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
- .byte 196,226,125,24,5,3,46,0,0 // vbroadcastss 0x2e03(%rip),%ymm0 # 67b4 <_sk_callback_avx+0x3b8>
+ .byte 196,226,125,24,5,4,46,0,0 // vbroadcastss 0x2e04(%rip),%ymm0 # 6984 <_sk_callback_avx+0x3b8>
.byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,246,45,0,0 // vbroadcastss 0x2df6(%rip),%ymm1 # 67b8 <_sk_callback_avx+0x3bc>
+ .byte 196,226,125,24,13,247,45,0,0 // vbroadcastss 0x2df7(%rip),%ymm1 # 6988 <_sk_callback_avx+0x3bc>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,237,45,0,0 // vbroadcastss 0x2ded(%rip),%ymm1 # 67bc <_sk_callback_avx+0x3c0>
+ .byte 196,226,125,24,13,238,45,0,0 // vbroadcastss 0x2dee(%rip),%ymm1 # 698c <_sk_callback_avx+0x3c0>
.byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,224,45,0,0 // vbroadcastss 0x2de0(%rip),%ymm3 # 67c0 <_sk_callback_avx+0x3c4>
+ .byte 196,226,125,24,29,225,45,0,0 // vbroadcastss 0x2de1(%rip),%ymm3 # 6990 <_sk_callback_avx+0x3c4>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,24,29,215,45,0,0 // vbroadcastss 0x2dd7(%rip),%ymm3 # 67c4 <_sk_callback_avx+0x3c8>
+ .byte 196,226,125,24,29,216,45,0,0 // vbroadcastss 0x2dd8(%rip),%ymm3 # 6994 <_sk_callback_avx+0x3c8>
.byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,202,45,0,0 // vbroadcastss 0x2dca(%rip),%ymm3 # 67c8 <_sk_callback_avx+0x3cc>
+ .byte 196,226,125,24,29,203,45,0,0 // vbroadcastss 0x2dcb(%rip),%ymm3 # 6998 <_sk_callback_avx+0x3cc>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,191,45,0,0 // vbroadcastss 0x2dbf(%rip),%ymm3 # 67cc <_sk_callback_avx+0x3d0>
+ .byte 196,226,125,24,29,192,45,0,0 // vbroadcastss 0x2dc0(%rip),%ymm3 # 699c <_sk_callback_avx+0x3d0>
.byte 255,224 // jmpq *%rax
.byte 69,137,193 // mov %r8d,%r9d
.byte 65,128,225,7 // and $0x7,%r9b
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 15,135,110,255,255,255 // ja 3995 <_sk_load_565_avx+0x14>
+ .byte 15,135,110,255,255,255 // ja 3b64 <_sk_load_565_avx+0x14>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,74,0,0,0 // lea 0x4a(%rip),%r10 # 3a7c <_sk_load_565_avx+0xfb>
+ .byte 76,141,21,75,0,0,0 // lea 0x4b(%rip),%r10 # 3c4c <_sk_load_565_avx+0xfc>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -18552,27 +18837,28 @@ _sk_load_565_avx:
.byte 196,193,121,196,68,83,4,2 // vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,68,83,2,1 // vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,4,83,0 // vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- .byte 233,26,255,255,255 // jmpq 3995 <_sk_load_565_avx+0x14>
- .byte 144 // nop
- .byte 243,255 // repz (bad)
+ .byte 233,26,255,255,255 // jmpq 3b64 <_sk_load_565_avx+0x14>
+ .byte 102,144 // xchg %ax,%ax
+ .byte 242,255 // repnz (bad)
+ .byte 255 // (bad)
.byte 255 // (bad)
+ .byte 234 // (bad)
.byte 255 // (bad)
- .byte 235,255 // jmp 3a81 <_sk_load_565_avx+0x100>
.byte 255 // (bad)
- .byte 255,227 // jmpq *%rbx
+ .byte 255,226 // jmpq *%rdx
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 219,255 // (bad)
+ .byte 218,255 // (bad)
.byte 255 // (bad)
- .byte 255,211 // callq *%rbx
+ .byte 255,210 // callq *%rdx
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,203 // dec %ebx
+ .byte 255,202 // dec %edx
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 191 // .byte 0xbf
+ .byte 190 // .byte 0xbe
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -18625,23 +18911,23 @@ _sk_gather_565_avx:
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
- .byte 196,226,125,24,5,102,44,0,0 // vbroadcastss 0x2c66(%rip),%ymm0 # 67d0 <_sk_callback_avx+0x3d4>
+ .byte 196,226,125,24,5,102,44,0,0 // vbroadcastss 0x2c66(%rip),%ymm0 # 69a0 <_sk_callback_avx+0x3d4>
.byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,89,44,0,0 // vbroadcastss 0x2c59(%rip),%ymm1 # 67d4 <_sk_callback_avx+0x3d8>
+ .byte 196,226,125,24,13,89,44,0,0 // vbroadcastss 0x2c59(%rip),%ymm1 # 69a4 <_sk_callback_avx+0x3d8>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,80,44,0,0 // vbroadcastss 0x2c50(%rip),%ymm1 # 67d8 <_sk_callback_avx+0x3dc>
+ .byte 196,226,125,24,13,80,44,0,0 // vbroadcastss 0x2c50(%rip),%ymm1 # 69a8 <_sk_callback_avx+0x3dc>
.byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,67,44,0,0 // vbroadcastss 0x2c43(%rip),%ymm3 # 67dc <_sk_callback_avx+0x3e0>
+ .byte 196,226,125,24,29,67,44,0,0 // vbroadcastss 0x2c43(%rip),%ymm3 # 69ac <_sk_callback_avx+0x3e0>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,24,29,58,44,0,0 // vbroadcastss 0x2c3a(%rip),%ymm3 # 67e0 <_sk_callback_avx+0x3e4>
+ .byte 196,226,125,24,29,58,44,0,0 // vbroadcastss 0x2c3a(%rip),%ymm3 # 69b0 <_sk_callback_avx+0x3e4>
.byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,45,44,0,0 // vbroadcastss 0x2c2d(%rip),%ymm3 # 67e4 <_sk_callback_avx+0x3e8>
+ .byte 196,226,125,24,29,45,44,0,0 // vbroadcastss 0x2c2d(%rip),%ymm3 # 69b4 <_sk_callback_avx+0x3e8>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,34,44,0,0 // vbroadcastss 0x2c22(%rip),%ymm3 # 67e8 <_sk_callback_avx+0x3ec>
+ .byte 196,226,125,24,29,34,44,0,0 // vbroadcastss 0x2c22(%rip),%ymm3 # 69b8 <_sk_callback_avx+0x3ec>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_565_avx
@@ -18650,14 +18936,14 @@ FUNCTION(_sk_store_565_avx)
_sk_store_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
- .byte 196,98,125,24,5,22,44,0,0 // vbroadcastss 0x2c16(%rip),%ymm8 # 67ec <_sk_callback_avx+0x3f0>
+ .byte 196,98,125,24,5,22,44,0,0 // vbroadcastss 0x2c16(%rip),%ymm8 # 69bc <_sk_callback_avx+0x3f0>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10
.byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9
.byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9
.byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
- .byte 196,98,125,24,21,239,43,0,0 // vbroadcastss 0x2bef(%rip),%ymm10 # 67f0 <_sk_callback_avx+0x3f4>
+ .byte 196,98,125,24,21,239,43,0,0 // vbroadcastss 0x2bef(%rip),%ymm10 # 69c0 <_sk_callback_avx+0x3f4>
.byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10
.byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
.byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11
@@ -18671,7 +18957,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,10 // jne 3c51 <_sk_store_565_avx+0x89>
+ .byte 117,10 // jne 3e21 <_sk_store_565_avx+0x89>
.byte 196,65,122,127,4,83 // vmovdqu %xmm8,(%r11,%rdx,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18679,9 +18965,9 @@ _sk_store_565_avx:
.byte 65,128,225,7 // and $0x7,%r9b
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,236 // ja 3c4d <_sk_store_565_avx+0x85>
+ .byte 119,236 // ja 3e1d <_sk_store_565_avx+0x85>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,68,0,0,0 // lea 0x44(%rip),%r10 # 3cb0 <_sk_store_565_avx+0xe8>
+ .byte 76,141,21,68,0,0,0 // lea 0x44(%rip),%r10 # 3e80 <_sk_store_565_avx+0xe8>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -18692,7 +18978,7 @@ _sk_store_565_avx:
.byte 196,67,121,21,68,83,4,2 // vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
.byte 196,67,121,21,68,83,2,1 // vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
.byte 196,67,121,21,4,83,0 // vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- .byte 235,159 // jmp 3c4d <_sk_store_565_avx+0x85>
+ .byte 235,159 // jmp 3e1d <_sk_store_565_avx+0x85>
.byte 102,144 // xchg %ax,%ax
.byte 245 // cmc
.byte 255 // (bad)
@@ -18725,31 +19011,31 @@ _sk_load_4444_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,152,0,0,0 // jne 3d72 <_sk_load_4444_avx+0xa6>
+ .byte 15,133,152,0,0,0 // jne 3f42 <_sk_load_4444_avx+0xa6>
.byte 196,193,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm0
.byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3
- .byte 196,226,125,24,5,248,42,0,0 // vbroadcastss 0x2af8(%rip),%ymm0 # 67f4 <_sk_callback_avx+0x3f8>
+ .byte 196,226,125,24,5,248,42,0,0 // vbroadcastss 0x2af8(%rip),%ymm0 # 69c4 <_sk_callback_avx+0x3f8>
.byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,235,42,0,0 // vbroadcastss 0x2aeb(%rip),%ymm1 # 67f8 <_sk_callback_avx+0x3fc>
+ .byte 196,226,125,24,13,235,42,0,0 // vbroadcastss 0x2aeb(%rip),%ymm1 # 69c8 <_sk_callback_avx+0x3fc>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,226,42,0,0 // vbroadcastss 0x2ae2(%rip),%ymm1 # 67fc <_sk_callback_avx+0x400>
+ .byte 196,226,125,24,13,226,42,0,0 // vbroadcastss 0x2ae2(%rip),%ymm1 # 69cc <_sk_callback_avx+0x400>
.byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,213,42,0,0 // vbroadcastss 0x2ad5(%rip),%ymm2 # 6800 <_sk_callback_avx+0x404>
+ .byte 196,226,125,24,21,213,42,0,0 // vbroadcastss 0x2ad5(%rip),%ymm2 # 69d0 <_sk_callback_avx+0x404>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,24,21,204,42,0,0 // vbroadcastss 0x2acc(%rip),%ymm2 # 6804 <_sk_callback_avx+0x408>
+ .byte 196,226,125,24,21,204,42,0,0 // vbroadcastss 0x2acc(%rip),%ymm2 # 69d4 <_sk_callback_avx+0x408>
.byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,191,42,0,0 // vbroadcastss 0x2abf(%rip),%ymm8 # 6808 <_sk_callback_avx+0x40c>
+ .byte 196,98,125,24,5,191,42,0,0 // vbroadcastss 0x2abf(%rip),%ymm8 # 69d8 <_sk_callback_avx+0x40c>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,181,42,0,0 // vbroadcastss 0x2ab5(%rip),%ymm8 # 680c <_sk_callback_avx+0x410>
+ .byte 196,98,125,24,5,181,42,0,0 // vbroadcastss 0x2ab5(%rip),%ymm8 # 69dc <_sk_callback_avx+0x410>
.byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,167,42,0,0 // vbroadcastss 0x2aa7(%rip),%ymm8 # 6810 <_sk_callback_avx+0x414>
+ .byte 196,98,125,24,5,167,42,0,0 // vbroadcastss 0x2aa7(%rip),%ymm8 # 69e0 <_sk_callback_avx+0x414>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18758,9 +19044,9 @@ _sk_load_4444_avx:
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 15,135,86,255,255,255 // ja 3ce0 <_sk_load_4444_avx+0x14>
+ .byte 15,135,86,255,255,255 // ja 3eb0 <_sk_load_4444_avx+0x14>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,75,0,0,0 // lea 0x4b(%rip),%r10 # 3de0 <_sk_load_4444_avx+0x114>
+ .byte 76,141,21,75,0,0,0 // lea 0x4b(%rip),%r10 # 3fb0 <_sk_load_4444_avx+0x114>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -18772,7 +19058,7 @@ _sk_load_4444_avx:
.byte 196,193,121,196,68,83,4,2 // vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,68,83,2,1 // vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
.byte 196,193,121,196,4,83,0 // vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- .byte 233,2,255,255,255 // jmpq 3ce0 <_sk_load_4444_avx+0x14>
+ .byte 233,2,255,255,255 // jmpq 3eb0 <_sk_load_4444_avx+0x14>
.byte 102,144 // xchg %ax,%ax
.byte 242,255 // repnz (bad)
.byte 255 // (bad)
@@ -18846,25 +19132,25 @@ _sk_gather_4444_avx:
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3
- .byte 196,226,125,24,5,70,41,0,0 // vbroadcastss 0x2946(%rip),%ymm0 # 6814 <_sk_callback_avx+0x418>
+ .byte 196,226,125,24,5,70,41,0,0 // vbroadcastss 0x2946(%rip),%ymm0 # 69e4 <_sk_callback_avx+0x418>
.byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,57,41,0,0 // vbroadcastss 0x2939(%rip),%ymm1 # 6818 <_sk_callback_avx+0x41c>
+ .byte 196,226,125,24,13,57,41,0,0 // vbroadcastss 0x2939(%rip),%ymm1 # 69e8 <_sk_callback_avx+0x41c>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,48,41,0,0 // vbroadcastss 0x2930(%rip),%ymm1 # 681c <_sk_callback_avx+0x420>
+ .byte 196,226,125,24,13,48,41,0,0 // vbroadcastss 0x2930(%rip),%ymm1 # 69ec <_sk_callback_avx+0x420>
.byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,35,41,0,0 // vbroadcastss 0x2923(%rip),%ymm2 # 6820 <_sk_callback_avx+0x424>
+ .byte 196,226,125,24,21,35,41,0,0 // vbroadcastss 0x2923(%rip),%ymm2 # 69f0 <_sk_callback_avx+0x424>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,24,21,26,41,0,0 // vbroadcastss 0x291a(%rip),%ymm2 # 6824 <_sk_callback_avx+0x428>
+ .byte 196,226,125,24,21,26,41,0,0 // vbroadcastss 0x291a(%rip),%ymm2 # 69f4 <_sk_callback_avx+0x428>
.byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,13,41,0,0 // vbroadcastss 0x290d(%rip),%ymm8 # 6828 <_sk_callback_avx+0x42c>
+ .byte 196,98,125,24,5,13,41,0,0 // vbroadcastss 0x290d(%rip),%ymm8 # 69f8 <_sk_callback_avx+0x42c>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,3,41,0,0 // vbroadcastss 0x2903(%rip),%ymm8 # 682c <_sk_callback_avx+0x430>
+ .byte 196,98,125,24,5,3,41,0,0 // vbroadcastss 0x2903(%rip),%ymm8 # 69fc <_sk_callback_avx+0x430>
.byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,245,40,0,0 // vbroadcastss 0x28f5(%rip),%ymm8 # 6830 <_sk_callback_avx+0x434>
+ .byte 196,98,125,24,5,245,40,0,0 // vbroadcastss 0x28f5(%rip),%ymm8 # 6a00 <_sk_callback_avx+0x434>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18875,7 +19161,7 @@ FUNCTION(_sk_store_4444_avx)
_sk_store_4444_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,24 // mov (%rax),%r11
- .byte 196,98,125,24,5,226,40,0,0 // vbroadcastss 0x28e2(%rip),%ymm8 # 6834 <_sk_callback_avx+0x438>
+ .byte 196,98,125,24,5,226,40,0,0 // vbroadcastss 0x28e2(%rip),%ymm8 # 6a04 <_sk_callback_avx+0x438>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10
@@ -18902,7 +19188,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,10 // jne 3feb <_sk_store_4444_avx+0xa7>
+ .byte 117,10 // jne 41bb <_sk_store_4444_avx+0xa7>
.byte 196,65,122,127,4,83 // vmovdqu %xmm8,(%r11,%rdx,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -18910,9 +19196,9 @@ _sk_store_4444_avx:
.byte 65,128,225,7 // and $0x7,%r9b
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 119,236 // ja 3fe7 <_sk_store_4444_avx+0xa3>
+ .byte 119,236 // ja 41b7 <_sk_store_4444_avx+0xa3>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,66,0,0,0 // lea 0x42(%rip),%r10 # 4048 <_sk_store_4444_avx+0x104>
+ .byte 76,141,21,66,0,0,0 // lea 0x42(%rip),%r10 # 4218 <_sk_store_4444_avx+0x104>
.byte 75,99,4,138 // movslq (%r10,%r9,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -18923,7 +19209,7 @@ _sk_store_4444_avx:
.byte 196,67,121,21,68,83,4,2 // vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
.byte 196,67,121,21,68,83,2,1 // vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
.byte 196,67,121,21,4,83,0 // vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- .byte 235,159 // jmp 3fe7 <_sk_store_4444_avx+0xa3>
+ .byte 235,159 // jmp 41b7 <_sk_store_4444_avx+0xa3>
.byte 247,255 // idiv %edi
.byte 255 // (bad)
.byte 255 // (bad)
@@ -18956,12 +19242,12 @@ _sk_load_8888_avx:
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 76,3,16 // add (%rax),%r10
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,137,0,0,0 // jne 4106 <_sk_load_8888_avx+0xa2>
+ .byte 15,133,137,0,0,0 // jne 42d6 <_sk_load_8888_avx+0xa2>
.byte 196,193,124,16,26 // vmovups (%r10),%ymm3
- .byte 197,124,40,21,246,41,0,0 // vmovaps 0x29f6(%rip),%ymm10 # 6a80 <_sk_callback_avx+0x684>
+ .byte 197,124,40,21,38,42,0,0 // vmovaps 0x2a26(%rip),%ymm10 # 6c80 <_sk_callback_avx+0x6b4>
.byte 196,193,100,84,194 // vandps %ymm10,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,156,39,0,0 // vbroadcastss 0x279c(%rip),%ymm8 # 6838 <_sk_callback_avx+0x43c>
+ .byte 196,98,125,24,5,156,39,0,0 // vbroadcastss 0x279c(%rip),%ymm8 # 6a08 <_sk_callback_avx+0x43c>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1
.byte 196,195,125,25,217,1 // vextractf128 $0x1,%ymm3,%xmm9
@@ -18991,13 +19277,13 @@ _sk_load_8888_avx:
.byte 72,211,232 // shr %cl,%rax
.byte 196,225,249,110,192 // vmovq %rax,%xmm0
.byte 196,226,121,48,192 // vpmovzxbw %xmm0,%xmm0
- .byte 196,226,121,0,13,162,40,0,0 // vpshufb 0x28a2(%rip),%xmm0,%xmm1 # 69d0 <_sk_callback_avx+0x5d4>
+ .byte 196,226,121,0,13,226,40,0,0 // vpshufb 0x28e2(%rip),%xmm0,%xmm1 # 6be0 <_sk_callback_avx+0x614>
.byte 196,226,121,33,201 // vpmovsxbd %xmm1,%xmm1
- .byte 196,226,121,0,5,164,40,0,0 // vpshufb 0x28a4(%rip),%xmm0,%xmm0 # 69e0 <_sk_callback_avx+0x5e4>
+ .byte 196,226,121,0,5,228,40,0,0 // vpshufb 0x28e4(%rip),%xmm0,%xmm0 # 6bf0 <_sk_callback_avx+0x624>
.byte 196,226,121,33,192 // vpmovsxbd %xmm0,%xmm0
.byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
.byte 196,194,125,44,26 // vmaskmovps (%r10),%ymm0,%ymm3
- .byte 233,49,255,255,255 // jmpq 4082 <_sk_load_8888_avx+0x1e>
+ .byte 233,49,255,255,255 // jmpq 4252 <_sk_load_8888_avx+0x1e>
HIDDEN _sk_gather_8888_avx
.globl _sk_gather_8888_avx
@@ -19036,10 +19322,10 @@ _sk_gather_8888_avx:
.byte 73,193,234,32 // shr $0x20,%r10
.byte 196,131,121,34,28,145,3 // vpinsrd $0x3,(%r9,%r10,4),%xmm0,%xmm3
.byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
- .byte 197,124,40,21,164,40,0,0 // vmovaps 0x28a4(%rip),%ymm10 # 6aa0 <_sk_callback_avx+0x6a4>
+ .byte 197,124,40,21,212,40,0,0 // vmovaps 0x28d4(%rip),%ymm10 # 6ca0 <_sk_callback_avx+0x6d4>
.byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,13,46,38,0,0 // vbroadcastss 0x262e(%rip),%ymm9 # 683c <_sk_callback_avx+0x440>
+ .byte 196,98,125,24,13,46,38,0,0 // vbroadcastss 0x262e(%rip),%ymm9 # 6a0c <_sk_callback_avx+0x440>
.byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
.byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1
.byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2
@@ -19069,7 +19355,7 @@ _sk_store_8888_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
.byte 76,3,16 // add (%rax),%r10
- .byte 196,98,125,24,5,184,37,0,0 // vbroadcastss 0x25b8(%rip),%ymm8 # 6840 <_sk_callback_avx+0x444>
+ .byte 196,98,125,24,5,184,37,0,0 // vbroadcastss 0x25b8(%rip),%ymm8 # 6a10 <_sk_callback_avx+0x444>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,65,116,89,208 // vmulps %ymm8,%ymm1,%ymm10
@@ -19094,7 +19380,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,12 // jne 4318 <_sk_store_8888_avx+0xa9>
+ .byte 117,12 // jne 44e8 <_sk_store_8888_avx+0xa9>
.byte 196,65,124,17,2 // vmovups %ymm8,(%r10)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,201 // mov %r9,%rcx
@@ -19106,13 +19392,13 @@ _sk_store_8888_avx:
.byte 72,211,232 // shr %cl,%rax
.byte 196,97,249,110,200 // vmovq %rax,%xmm9
.byte 196,66,121,48,201 // vpmovzxbw %xmm9,%xmm9
- .byte 196,98,49,0,21,176,38,0,0 // vpshufb 0x26b0(%rip),%xmm9,%xmm10 # 69f0 <_sk_callback_avx+0x5f4>
+ .byte 196,98,49,0,21,240,38,0,0 // vpshufb 0x26f0(%rip),%xmm9,%xmm10 # 6c00 <_sk_callback_avx+0x634>
.byte 196,66,121,33,210 // vpmovsxbd %xmm10,%xmm10
- .byte 196,98,49,0,13,178,38,0,0 // vpshufb 0x26b2(%rip),%xmm9,%xmm9 # 6a00 <_sk_callback_avx+0x604>
+ .byte 196,98,49,0,13,242,38,0,0 // vpshufb 0x26f2(%rip),%xmm9,%xmm9 # 6c10 <_sk_callback_avx+0x644>
.byte 196,66,121,33,201 // vpmovsxbd %xmm9,%xmm9
.byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
.byte 196,66,53,46,2 // vmaskmovps %ymm8,%ymm9,(%r10)
- .byte 235,177 // jmp 4311 <_sk_store_8888_avx+0xa2>
+ .byte 235,177 // jmp 44e1 <_sk_store_8888_avx+0xa2>
HIDDEN _sk_load_f16_avx
.globl _sk_load_f16_avx
@@ -19126,7 +19412,7 @@ _sk_load_f16_avx:
.byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp)
.byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp)
.byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp)
- .byte 15,133,141,2,0,0 // jne 4617 <_sk_load_f16_avx+0x2b7>
+ .byte 15,133,141,2,0,0 // jne 47e7 <_sk_load_f16_avx+0x2b7>
.byte 197,121,16,4,208 // vmovupd (%rax,%rdx,8),%xmm8
.byte 197,249,16,84,208,16 // vmovupd 0x10(%rax,%rdx,8),%xmm2
.byte 197,249,16,76,208,32 // vmovupd 0x20(%rax,%rdx,8),%xmm1
@@ -19144,13 +19430,13 @@ _sk_load_f16_avx:
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- .byte 196,98,125,24,37,97,36,0,0 // vbroadcastss 0x2461(%rip),%ymm12 # 6844 <_sk_callback_avx+0x448>
+ .byte 196,98,125,24,37,97,36,0,0 // vbroadcastss 0x2461(%rip),%ymm12 # 6a14 <_sk_callback_avx+0x448>
.byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1
.byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0
.byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14
- .byte 196,98,121,24,29,77,36,0,0 // vbroadcastss 0x244d(%rip),%xmm11 # 6848 <_sk_callback_avx+0x44c>
+ .byte 196,98,121,24,29,77,36,0,0 // vbroadcastss 0x244d(%rip),%xmm11 # 6a18 <_sk_callback_avx+0x44c>
.byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3
- .byte 196,98,121,24,45,67,36,0,0 // vbroadcastss 0x2443(%rip),%xmm13 # 684c <_sk_callback_avx+0x450>
+ .byte 196,98,121,24,45,67,36,0,0 // vbroadcastss 0x2443(%rip),%xmm13 # 6a1c <_sk_callback_avx+0x450>
.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
@@ -19164,7 +19450,7 @@ _sk_load_f16_avx:
.byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
.byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0
.byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1
- .byte 196,226,121,24,29,249,35,0,0 // vbroadcastss 0x23f9(%rip),%xmm3 # 6850 <_sk_callback_avx+0x454>
+ .byte 196,226,121,24,29,249,35,0,0 // vbroadcastss 0x23f9(%rip),%xmm3 # 6a20 <_sk_callback_avx+0x454>
.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
@@ -19257,29 +19543,29 @@ _sk_load_f16_avx:
.byte 197,123,16,4,208 // vmovsd (%rax,%rdx,8),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,79 // je 4676 <_sk_load_f16_avx+0x316>
+ .byte 116,79 // je 4846 <_sk_load_f16_avx+0x316>
.byte 197,57,22,68,208,8 // vmovhpd 0x8(%rax,%rdx,8),%xmm8,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,67 // jb 4676 <_sk_load_f16_avx+0x316>
+ .byte 114,67 // jb 4846 <_sk_load_f16_avx+0x316>
.byte 197,251,16,84,208,16 // vmovsd 0x10(%rax,%rdx,8),%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 116,68 // je 4683 <_sk_load_f16_avx+0x323>
+ .byte 116,68 // je 4853 <_sk_load_f16_avx+0x323>
.byte 197,233,22,84,208,24 // vmovhpd 0x18(%rax,%rdx,8),%xmm2,%xmm2
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,56 // jb 4683 <_sk_load_f16_avx+0x323>
+ .byte 114,56 // jb 4853 <_sk_load_f16_avx+0x323>
.byte 197,251,16,76,208,32 // vmovsd 0x20(%rax,%rdx,8),%xmm1
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 15,132,70,253,255,255 // je 43a1 <_sk_load_f16_avx+0x41>
+ .byte 15,132,70,253,255,255 // je 4571 <_sk_load_f16_avx+0x41>
.byte 197,241,22,76,208,40 // vmovhpd 0x28(%rax,%rdx,8),%xmm1,%xmm1
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 15,130,54,253,255,255 // jb 43a1 <_sk_load_f16_avx+0x41>
+ .byte 15,130,54,253,255,255 // jb 4571 <_sk_load_f16_avx+0x41>
.byte 197,122,126,76,208,48 // vmovq 0x30(%rax,%rdx,8),%xmm9
- .byte 233,43,253,255,255 // jmpq 43a1 <_sk_load_f16_avx+0x41>
+ .byte 233,43,253,255,255 // jmpq 4571 <_sk_load_f16_avx+0x41>
.byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,30,253,255,255 // jmpq 43a1 <_sk_load_f16_avx+0x41>
+ .byte 233,30,253,255,255 // jmpq 4571 <_sk_load_f16_avx+0x41>
.byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1
- .byte 233,21,253,255,255 // jmpq 43a1 <_sk_load_f16_avx+0x41>
+ .byte 233,21,253,255,255 // jmpq 4571 <_sk_load_f16_avx+0x41>
HIDDEN _sk_gather_f16_avx
.globl _sk_gather_f16_avx
@@ -19340,13 +19626,13 @@ _sk_gather_f16_avx:
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- .byte 196,98,125,24,37,190,32,0,0 // vbroadcastss 0x20be(%rip),%ymm12 # 6854 <_sk_callback_avx+0x458>
+ .byte 196,98,125,24,37,190,32,0,0 // vbroadcastss 0x20be(%rip),%ymm12 # 6a24 <_sk_callback_avx+0x458>
.byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1
.byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0
.byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14
- .byte 196,98,121,24,29,170,32,0,0 // vbroadcastss 0x20aa(%rip),%xmm11 # 6858 <_sk_callback_avx+0x45c>
+ .byte 196,98,121,24,29,170,32,0,0 // vbroadcastss 0x20aa(%rip),%xmm11 # 6a28 <_sk_callback_avx+0x45c>
.byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3
- .byte 196,98,121,24,45,160,32,0,0 // vbroadcastss 0x20a0(%rip),%xmm13 # 685c <_sk_callback_avx+0x460>
+ .byte 196,98,121,24,45,160,32,0,0 // vbroadcastss 0x20a0(%rip),%xmm13 # 6a2c <_sk_callback_avx+0x460>
.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
@@ -19360,7 +19646,7 @@ _sk_gather_f16_avx:
.byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
.byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0
.byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1
- .byte 196,226,121,24,29,86,32,0,0 // vbroadcastss 0x2056(%rip),%xmm3 # 6860 <_sk_callback_avx+0x464>
+ .byte 196,226,121,24,29,86,32,0,0 // vbroadcastss 0x2056(%rip),%xmm3 # 6a30 <_sk_callback_avx+0x464>
.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
@@ -19460,12 +19746,12 @@ _sk_store_f16_avx:
.byte 197,252,17,52,36 // vmovups %ymm6,(%rsp)
.byte 197,252,17,108,36,224 // vmovups %ymm5,-0x20(%rsp)
.byte 197,252,17,100,36,192 // vmovups %ymm4,-0x40(%rsp)
- .byte 196,98,125,24,13,118,30,0,0 // vbroadcastss 0x1e76(%rip),%ymm9 # 6864 <_sk_callback_avx+0x468>
+ .byte 196,98,125,24,13,118,30,0,0 // vbroadcastss 0x1e76(%rip),%ymm9 # 6a34 <_sk_callback_avx+0x468>
.byte 196,65,124,84,209 // vandps %ymm9,%ymm0,%ymm10
.byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp)
.byte 196,65,124,87,218 // vxorps %ymm10,%ymm0,%ymm11
.byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12
- .byte 196,98,121,24,5,91,30,0,0 // vbroadcastss 0x1e5b(%rip),%xmm8 # 6868 <_sk_callback_avx+0x46c>
+ .byte 196,98,121,24,5,91,30,0,0 // vbroadcastss 0x1e5b(%rip),%xmm8 # 6a38 <_sk_callback_avx+0x46c>
.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
@@ -19475,7 +19761,7 @@ _sk_store_f16_avx:
.byte 196,67,13,24,242,1 // vinsertf128 $0x1,%xmm10,%ymm14,%ymm14
.byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11
.byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12
- .byte 196,98,125,24,21,34,30,0,0 // vbroadcastss 0x1e22(%rip),%ymm10 # 686c <_sk_callback_avx+0x470>
+ .byte 196,98,125,24,21,34,30,0,0 // vbroadcastss 0x1e22(%rip),%ymm10 # 6a3c <_sk_callback_avx+0x470>
.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
@@ -19557,7 +19843,7 @@ _sk_store_f16_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,66 // jne 4c28 <_sk_store_f16_avx+0x25e>
+ .byte 117,66 // jne 4df8 <_sk_store_f16_avx+0x25e>
.byte 197,120,17,28,208 // vmovups %xmm11,(%rax,%rdx,8)
.byte 197,120,17,84,208,16 // vmovups %xmm10,0x10(%rax,%rdx,8)
.byte 197,120,17,76,208,32 // vmovups %xmm9,0x20(%rax,%rdx,8)
@@ -19573,22 +19859,22 @@ _sk_store_f16_avx:
.byte 255,224 // jmpq *%rax
.byte 197,121,214,28,208 // vmovq %xmm11,(%rax,%rdx,8)
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,202 // je 4bfd <_sk_store_f16_avx+0x233>
+ .byte 116,202 // je 4dcd <_sk_store_f16_avx+0x233>
.byte 197,121,23,92,208,8 // vmovhpd %xmm11,0x8(%rax,%rdx,8)
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,190 // jb 4bfd <_sk_store_f16_avx+0x233>
+ .byte 114,190 // jb 4dcd <_sk_store_f16_avx+0x233>
.byte 197,121,214,84,208,16 // vmovq %xmm10,0x10(%rax,%rdx,8)
- .byte 116,182 // je 4bfd <_sk_store_f16_avx+0x233>
+ .byte 116,182 // je 4dcd <_sk_store_f16_avx+0x233>
.byte 197,121,23,84,208,24 // vmovhpd %xmm10,0x18(%rax,%rdx,8)
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,170 // jb 4bfd <_sk_store_f16_avx+0x233>
+ .byte 114,170 // jb 4dcd <_sk_store_f16_avx+0x233>
.byte 197,121,214,76,208,32 // vmovq %xmm9,0x20(%rax,%rdx,8)
- .byte 116,162 // je 4bfd <_sk_store_f16_avx+0x233>
+ .byte 116,162 // je 4dcd <_sk_store_f16_avx+0x233>
.byte 197,121,23,76,208,40 // vmovhpd %xmm9,0x28(%rax,%rdx,8)
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,150 // jb 4bfd <_sk_store_f16_avx+0x233>
+ .byte 114,150 // jb 4dcd <_sk_store_f16_avx+0x233>
.byte 197,121,214,68,208,48 // vmovq %xmm8,0x30(%rax,%rdx,8)
- .byte 235,142 // jmp 4bfd <_sk_store_f16_avx+0x233>
+ .byte 235,142 // jmp 4dcd <_sk_store_f16_avx+0x233>
HIDDEN _sk_load_u16_be_avx
.globl _sk_load_u16_be_avx
@@ -19598,7 +19884,7 @@ _sk_load_u16_be_avx:
.byte 76,139,8 // mov (%rax),%r9
.byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,253,0,0,0 // jne 4d82 <_sk_load_u16_be_avx+0x113>
+ .byte 15,133,253,0,0,0 // jne 4f52 <_sk_load_u16_be_avx+0x113>
.byte 196,65,121,16,4,65 // vmovupd (%r9,%rax,2),%xmm8
.byte 196,193,121,16,84,65,16 // vmovupd 0x10(%r9,%rax,2),%xmm2
.byte 196,193,121,16,92,65,32 // vmovupd 0x20(%r9,%rax,2),%xmm3
@@ -19620,7 +19906,7 @@ _sk_load_u16_be_avx:
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,29,122,27,0,0 // vbroadcastss 0x1b7a(%rip),%ymm11 # 6870 <_sk_callback_avx+0x474>
+ .byte 196,98,125,24,29,122,27,0,0 // vbroadcastss 0x1b7a(%rip),%ymm11 # 6a40 <_sk_callback_avx+0x474>
.byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0
.byte 197,177,109,202 // vpunpckhqdq %xmm2,%xmm9,%xmm1
.byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
@@ -19654,29 +19940,29 @@ _sk_load_u16_be_avx:
.byte 196,65,123,16,4,65 // vmovsd (%r9,%rax,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,85 // je 4de8 <_sk_load_u16_be_avx+0x179>
+ .byte 116,85 // je 4fb8 <_sk_load_u16_be_avx+0x179>
.byte 196,65,57,22,68,65,8 // vmovhpd 0x8(%r9,%rax,2),%xmm8,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,72 // jb 4de8 <_sk_load_u16_be_avx+0x179>
+ .byte 114,72 // jb 4fb8 <_sk_load_u16_be_avx+0x179>
.byte 196,193,123,16,84,65,16 // vmovsd 0x10(%r9,%rax,2),%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 116,72 // je 4df5 <_sk_load_u16_be_avx+0x186>
+ .byte 116,72 // je 4fc5 <_sk_load_u16_be_avx+0x186>
.byte 196,193,105,22,84,65,24 // vmovhpd 0x18(%r9,%rax,2),%xmm2,%xmm2
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,59 // jb 4df5 <_sk_load_u16_be_avx+0x186>
+ .byte 114,59 // jb 4fc5 <_sk_load_u16_be_avx+0x186>
.byte 196,193,123,16,92,65,32 // vmovsd 0x20(%r9,%rax,2),%xmm3
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 15,132,213,254,255,255 // je 4ca0 <_sk_load_u16_be_avx+0x31>
+ .byte 15,132,213,254,255,255 // je 4e70 <_sk_load_u16_be_avx+0x31>
.byte 196,193,97,22,92,65,40 // vmovhpd 0x28(%r9,%rax,2),%xmm3,%xmm3
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 15,130,196,254,255,255 // jb 4ca0 <_sk_load_u16_be_avx+0x31>
+ .byte 15,130,196,254,255,255 // jb 4e70 <_sk_load_u16_be_avx+0x31>
.byte 196,65,122,126,76,65,48 // vmovq 0x30(%r9,%rax,2),%xmm9
- .byte 233,184,254,255,255 // jmpq 4ca0 <_sk_load_u16_be_avx+0x31>
+ .byte 233,184,254,255,255 // jmpq 4e70 <_sk_load_u16_be_avx+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,171,254,255,255 // jmpq 4ca0 <_sk_load_u16_be_avx+0x31>
+ .byte 233,171,254,255,255 // jmpq 4e70 <_sk_load_u16_be_avx+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,162,254,255,255 // jmpq 4ca0 <_sk_load_u16_be_avx+0x31>
+ .byte 233,162,254,255,255 // jmpq 4e70 <_sk_load_u16_be_avx+0x31>
HIDDEN _sk_load_rgb_u16_be_avx
.globl _sk_load_rgb_u16_be_avx
@@ -19686,7 +19972,7 @@ _sk_load_rgb_u16_be_avx:
.byte 76,139,8 // mov (%rax),%r9
.byte 72,141,4,82 // lea (%rdx,%rdx,2),%rax
.byte 77,133,192 // test %r8,%r8
- .byte 15,133,243,0,0,0 // jne 4f03 <_sk_load_rgb_u16_be_avx+0x105>
+ .byte 15,133,243,0,0,0 // jne 50d3 <_sk_load_rgb_u16_be_avx+0x105>
.byte 196,193,122,111,4,65 // vmovdqu (%r9,%rax,2),%xmm0
.byte 196,193,122,111,84,65,12 // vmovdqu 0xc(%r9,%rax,2),%xmm2
.byte 196,193,122,111,76,65,24 // vmovdqu 0x18(%r9,%rax,2),%xmm1
@@ -19713,7 +19999,7 @@ _sk_load_rgb_u16_be_avx:
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,29,218,25,0,0 // vbroadcastss 0x19da(%rip),%ymm11 # 6874 <_sk_callback_avx+0x478>
+ .byte 196,98,125,24,29,218,25,0,0 // vbroadcastss 0x19da(%rip),%ymm11 # 6a44 <_sk_callback_avx+0x478>
.byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0
.byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
.byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
@@ -19734,41 +20020,41 @@ _sk_load_rgb_u16_be_avx:
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,119,25,0,0 // vbroadcastss 0x1977(%rip),%ymm3 # 6878 <_sk_callback_avx+0x47c>
+ .byte 196,226,125,24,29,119,25,0,0 // vbroadcastss 0x1977(%rip),%ymm3 # 6a48 <_sk_callback_avx+0x47c>
.byte 255,224 // jmpq *%rax
.byte 196,193,121,110,4,65 // vmovd (%r9,%rax,2),%xmm0
.byte 196,193,121,196,68,65,4,2 // vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 117,5 // jne 4f1c <_sk_load_rgb_u16_be_avx+0x11e>
- .byte 233,40,255,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 117,5 // jne 50ec <_sk_load_rgb_u16_be_avx+0x11e>
+ .byte 233,40,255,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
.byte 196,193,121,110,76,65,6 // vmovd 0x6(%r9,%rax,2),%xmm1
.byte 196,65,113,196,68,65,10,2 // vpinsrw $0x2,0xa(%r9,%rax,2),%xmm1,%xmm8
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,26 // jb 4f4b <_sk_load_rgb_u16_be_avx+0x14d>
+ .byte 114,26 // jb 511b <_sk_load_rgb_u16_be_avx+0x14d>
.byte 196,193,121,110,76,65,12 // vmovd 0xc(%r9,%rax,2),%xmm1
.byte 196,193,113,196,84,65,16,2 // vpinsrw $0x2,0x10(%r9,%rax,2),%xmm1,%xmm2
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 117,10 // jne 4f50 <_sk_load_rgb_u16_be_avx+0x152>
- .byte 233,249,254,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
- .byte 233,244,254,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 117,10 // jne 5120 <_sk_load_rgb_u16_be_avx+0x152>
+ .byte 233,249,254,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,244,254,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
.byte 196,193,121,110,76,65,18 // vmovd 0x12(%r9,%rax,2),%xmm1
.byte 196,65,113,196,76,65,22,2 // vpinsrw $0x2,0x16(%r9,%rax,2),%xmm1,%xmm9
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,26 // jb 4f7f <_sk_load_rgb_u16_be_avx+0x181>
+ .byte 114,26 // jb 514f <_sk_load_rgb_u16_be_avx+0x181>
.byte 196,193,121,110,76,65,24 // vmovd 0x18(%r9,%rax,2),%xmm1
.byte 196,193,113,196,76,65,28,2 // vpinsrw $0x2,0x1c(%r9,%rax,2),%xmm1,%xmm1
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 117,10 // jne 4f84 <_sk_load_rgb_u16_be_avx+0x186>
- .byte 233,197,254,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
- .byte 233,192,254,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 117,10 // jne 5154 <_sk_load_rgb_u16_be_avx+0x186>
+ .byte 233,197,254,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,192,254,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
.byte 196,193,121,110,92,65,30 // vmovd 0x1e(%r9,%rax,2),%xmm3
.byte 196,65,97,196,92,65,34,2 // vpinsrw $0x2,0x22(%r9,%rax,2),%xmm3,%xmm11
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,20 // jb 4fad <_sk_load_rgb_u16_be_avx+0x1af>
+ .byte 114,20 // jb 517d <_sk_load_rgb_u16_be_avx+0x1af>
.byte 196,193,121,110,92,65,36 // vmovd 0x24(%r9,%rax,2),%xmm3
.byte 196,193,97,196,92,65,40,2 // vpinsrw $0x2,0x28(%r9,%rax,2),%xmm3,%xmm3
- .byte 233,151,254,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
- .byte 233,146,254,255,255 // jmpq 4e44 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,151,254,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,146,254,255,255 // jmpq 5014 <_sk_load_rgb_u16_be_avx+0x46>
HIDDEN _sk_store_u16_be_avx
.globl _sk_store_u16_be_avx
@@ -19777,7 +20063,7 @@ _sk_store_u16_be_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,8 // mov (%rax),%r9
.byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax
- .byte 196,98,125,24,5,180,24,0,0 // vbroadcastss 0x18b4(%rip),%ymm8 # 687c <_sk_callback_avx+0x480>
+ .byte 196,98,125,24,5,180,24,0,0 // vbroadcastss 0x18b4(%rip),%ymm8 # 6a4c <_sk_callback_avx+0x480>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
.byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10
@@ -19815,7 +20101,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,31 // jne 50ac <_sk_store_u16_be_avx+0xfa>
+ .byte 117,31 // jne 527c <_sk_store_u16_be_avx+0xfa>
.byte 196,65,120,17,28,65 // vmovups %xmm11,(%r9,%rax,2)
.byte 196,65,120,17,84,65,16 // vmovups %xmm10,0x10(%r9,%rax,2)
.byte 196,65,120,17,76,65,32 // vmovups %xmm9,0x20(%r9,%rax,2)
@@ -19824,22 +20110,22 @@ _sk_store_u16_be_avx:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,214,28,65 // vmovq %xmm11,(%r9,%rax,2)
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,240 // je 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 116,240 // je 5278 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,23,92,65,8 // vmovhpd %xmm11,0x8(%r9,%rax,2)
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,227 // jb 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 114,227 // jb 5278 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,214,84,65,16 // vmovq %xmm10,0x10(%r9,%rax,2)
- .byte 116,218 // je 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 116,218 // je 5278 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,23,84,65,24 // vmovhpd %xmm10,0x18(%r9,%rax,2)
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,205 // jb 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 114,205 // jb 5278 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,214,76,65,32 // vmovq %xmm9,0x20(%r9,%rax,2)
- .byte 116,196 // je 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 116,196 // je 5278 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,23,76,65,40 // vmovhpd %xmm9,0x28(%r9,%rax,2)
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,183 // jb 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 114,183 // jb 5278 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,214,68,65,48 // vmovq %xmm8,0x30(%r9,%rax,2)
- .byte 235,174 // jmp 50a8 <_sk_store_u16_be_avx+0xf6>
+ .byte 235,174 // jmp 5278 <_sk_store_u16_be_avx+0xf6>
HIDDEN _sk_load_f32_avx
.globl _sk_load_f32_avx
@@ -19847,10 +20133,10 @@ FUNCTION(_sk_load_f32_avx)
_sk_load_f32_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 119,110 // ja 5170 <_sk_load_f32_avx+0x76>
+ .byte 119,110 // ja 5340 <_sk_load_f32_avx+0x76>
.byte 76,139,8 // mov (%rax),%r9
.byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10
- .byte 76,141,29,132,0,0,0 // lea 0x84(%rip),%r11 # 5198 <_sk_load_f32_avx+0x9e>
+ .byte 76,141,29,132,0,0,0 // lea 0x84(%rip),%r11 # 5368 <_sk_load_f32_avx+0x9e>
.byte 75,99,4,131 // movslq (%r11,%r8,4),%rax
.byte 76,1,216 // add %r11,%rax
.byte 255,224 // jmpq *%rax
@@ -19909,7 +20195,7 @@ _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 77,133,192 // test %r8,%r8
- .byte 117,55 // jne 5225 <_sk_store_f32_avx+0x6d>
+ .byte 117,55 // jne 53f5 <_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
@@ -19922,22 +20208,22 @@ _sk_store_f32_avx:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,17,20,129 // vmovupd %xmm10,(%r9,%rax,4)
.byte 73,131,248,1 // cmp $0x1,%r8
- .byte 116,240 // je 5221 <_sk_store_f32_avx+0x69>
+ .byte 116,240 // je 53f1 <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,76,129,16 // vmovupd %xmm9,0x10(%r9,%rax,4)
.byte 73,131,248,3 // cmp $0x3,%r8
- .byte 114,227 // jb 5221 <_sk_store_f32_avx+0x69>
+ .byte 114,227 // jb 53f1 <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,68,129,32 // vmovupd %xmm8,0x20(%r9,%rax,4)
- .byte 116,218 // je 5221 <_sk_store_f32_avx+0x69>
+ .byte 116,218 // je 53f1 <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,92,129,48 // vmovupd %xmm11,0x30(%r9,%rax,4)
.byte 73,131,248,5 // cmp $0x5,%r8
- .byte 114,205 // jb 5221 <_sk_store_f32_avx+0x69>
+ .byte 114,205 // jb 53f1 <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,84,129,64,1 // vextractf128 $0x1,%ymm10,0x40(%r9,%rax,4)
- .byte 116,195 // je 5221 <_sk_store_f32_avx+0x69>
+ .byte 116,195 // je 53f1 <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,76,129,80,1 // vextractf128 $0x1,%ymm9,0x50(%r9,%rax,4)
.byte 73,131,248,7 // cmp $0x7,%r8
- .byte 114,181 // jb 5221 <_sk_store_f32_avx+0x69>
+ .byte 114,181 // jb 53f1 <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,68,129,96,1 // vextractf128 $0x1,%ymm8,0x60(%r9,%rax,4)
- .byte 235,171 // jmp 5221 <_sk_store_f32_avx+0x69>
+ .byte 235,171 // jmp 53f1 <_sk_store_f32_avx+0x69>
HIDDEN _sk_clamp_x_avx
.globl _sk_clamp_x_avx
@@ -20041,7 +20327,7 @@ FUNCTION(_sk_clamp_x_1_avx)
_sk_clamp_x_1_avx:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,230,20,0,0 // vbroadcastss 0x14e6(%rip),%ymm8 # 6880 <_sk_callback_avx+0x484>
+ .byte 196,98,125,24,5,230,20,0,0 // vbroadcastss 0x14e6(%rip),%ymm8 # 6a50 <_sk_callback_avx+0x484>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -20059,9 +20345,9 @@ HIDDEN _sk_mirror_x_1_avx
.globl _sk_mirror_x_1_avx
FUNCTION(_sk_mirror_x_1_avx)
_sk_mirror_x_1_avx:
- .byte 196,98,125,24,5,201,20,0,0 // vbroadcastss 0x14c9(%rip),%ymm8 # 6884 <_sk_callback_avx+0x488>
+ .byte 196,98,125,24,5,201,20,0,0 // vbroadcastss 0x14c9(%rip),%ymm8 # 6a54 <_sk_callback_avx+0x488>
.byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,13,191,20,0,0 // vbroadcastss 0x14bf(%rip),%ymm9 # 6888 <_sk_callback_avx+0x48c>
+ .byte 196,98,125,24,13,191,20,0,0 // vbroadcastss 0x14bf(%rip),%ymm9 # 6a58 <_sk_callback_avx+0x48c>
.byte 196,65,124,89,201 // vmulps %ymm9,%ymm0,%ymm9
.byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9
.byte 196,65,52,88,201 // vaddps %ymm9,%ymm9,%ymm9
@@ -20077,12 +20363,12 @@ HIDDEN _sk_luminance_to_alpha_avx
.globl _sk_luminance_to_alpha_avx
FUNCTION(_sk_luminance_to_alpha_avx)
_sk_luminance_to_alpha_avx:
- .byte 196,226,125,24,29,143,20,0,0 // vbroadcastss 0x148f(%rip),%ymm3 # 688c <_sk_callback_avx+0x490>
+ .byte 196,226,125,24,29,143,20,0,0 // vbroadcastss 0x148f(%rip),%ymm3 # 6a5c <_sk_callback_avx+0x490>
.byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
- .byte 196,226,125,24,29,134,20,0,0 // vbroadcastss 0x1486(%rip),%ymm3 # 6890 <_sk_callback_avx+0x494>
+ .byte 196,226,125,24,29,134,20,0,0 // vbroadcastss 0x1486(%rip),%ymm3 # 6a60 <_sk_callback_avx+0x494>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,121,20,0,0 // vbroadcastss 0x1479(%rip),%ymm1 # 6894 <_sk_callback_avx+0x498>
+ .byte 196,226,125,24,13,121,20,0,0 // vbroadcastss 0x1479(%rip),%ymm1 # 6a64 <_sk_callback_avx+0x498>
.byte 197,236,89,201 // vmulps %ymm1,%ymm2,%ymm1
.byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20302,9 +20588,9 @@ _sk_evenly_spaced_gradient_avx:
.byte 72,139,24 // mov (%rax),%rbx
.byte 72,139,104,8 // mov 0x8(%rax),%rbp
.byte 72,255,203 // dec %rbx
- .byte 120,7 // js 5777 <_sk_evenly_spaced_gradient_avx+0x25>
+ .byte 120,7 // js 5947 <_sk_evenly_spaced_gradient_avx+0x25>
.byte 196,225,242,42,203 // vcvtsi2ss %rbx,%xmm1,%xmm1
- .byte 235,21 // jmp 578c <_sk_evenly_spaced_gradient_avx+0x3a>
+ .byte 235,21 // jmp 595c <_sk_evenly_spaced_gradient_avx+0x3a>
.byte 73,137,217 // mov %rbx,%r9
.byte 73,209,233 // shr %r9
.byte 131,227,1 // and $0x1,%ebx
@@ -20462,18 +20748,18 @@ HIDDEN _sk_gauss_a_to_rgba_avx
.globl _sk_gauss_a_to_rgba_avx
FUNCTION(_sk_gauss_a_to_rgba_avx)
_sk_gauss_a_to_rgba_avx:
- .byte 196,226,125,24,5,168,13,0,0 // vbroadcastss 0xda8(%rip),%ymm0 # 6898 <_sk_callback_avx+0x49c>
+ .byte 196,226,125,24,5,168,13,0,0 // vbroadcastss 0xda8(%rip),%ymm0 # 6a68 <_sk_callback_avx+0x49c>
.byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0
- .byte 196,226,125,24,13,159,13,0,0 // vbroadcastss 0xd9f(%rip),%ymm1 # 689c <_sk_callback_avx+0x4a0>
+ .byte 196,226,125,24,13,159,13,0,0 // vbroadcastss 0xd9f(%rip),%ymm1 # 6a6c <_sk_callback_avx+0x4a0>
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
.byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
- .byte 196,226,125,24,13,146,13,0,0 // vbroadcastss 0xd92(%rip),%ymm1 # 68a0 <_sk_callback_avx+0x4a4>
+ .byte 196,226,125,24,13,146,13,0,0 // vbroadcastss 0xd92(%rip),%ymm1 # 6a70 <_sk_callback_avx+0x4a4>
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
.byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
- .byte 196,226,125,24,13,133,13,0,0 // vbroadcastss 0xd85(%rip),%ymm1 # 68a4 <_sk_callback_avx+0x4a8>
+ .byte 196,226,125,24,13,133,13,0,0 // vbroadcastss 0xd85(%rip),%ymm1 # 6a74 <_sk_callback_avx+0x4a8>
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
.byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
- .byte 196,226,125,24,13,120,13,0,0 // vbroadcastss 0xd78(%rip),%ymm1 # 68a8 <_sk_callback_avx+0x4ac>
+ .byte 196,226,125,24,13,120,13,0,0 // vbroadcastss 0xd78(%rip),%ymm1 # 6a78 <_sk_callback_avx+0x4ac>
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
@@ -20496,12 +20782,12 @@ _sk_gradient_avx:
.byte 76,139,8 // mov (%rax),%r9
.byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
.byte 73,131,249,2 // cmp $0x2,%r9
- .byte 114,80 // jb 5bb3 <_sk_gradient_avx+0x6f>
+ .byte 114,80 // jb 5d83 <_sk_gradient_avx+0x6f>
.byte 72,139,88,72 // mov 0x48(%rax),%rbx
.byte 73,255,201 // dec %r9
.byte 72,131,195,4 // add $0x4,%rbx
.byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
- .byte 196,98,125,24,21,48,13,0,0 // vbroadcastss 0xd30(%rip),%ymm10 # 68ac <_sk_callback_avx+0x4b0>
+ .byte 196,98,125,24,21,48,13,0,0 // vbroadcastss 0xd30(%rip),%ymm10 # 6a7c <_sk_callback_avx+0x4b0>
.byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
.byte 196,98,125,24,3 // vbroadcastss (%rbx),%ymm8
.byte 197,60,194,192,2 // vcmpleps %ymm0,%ymm8,%ymm8
@@ -20513,7 +20799,7 @@ _sk_gradient_avx:
.byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
.byte 72,131,195,4 // add $0x4,%rbx
.byte 73,255,201 // dec %r9
- .byte 117,205 // jne 5b80 <_sk_gradient_avx+0x3c>
+ .byte 117,205 // jne 5d50 <_sk_gradient_avx+0x3c>
.byte 196,195,249,22,201,1 // vpextrq $0x1,%xmm1,%r9
.byte 69,137,202 // mov %r9d,%r10d
.byte 73,193,233,32 // shr $0x20,%r9
@@ -20696,27 +20982,27 @@ _sk_xy_to_unit_angle_avx:
.byte 196,65,52,95,226 // vmaxps %ymm10,%ymm9,%ymm12
.byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11
.byte 196,65,36,89,227 // vmulps %ymm11,%ymm11,%ymm12
- .byte 196,98,125,24,45,38,9,0,0 // vbroadcastss 0x926(%rip),%ymm13 # 68b0 <_sk_callback_avx+0x4b4>
+ .byte 196,98,125,24,45,38,9,0,0 // vbroadcastss 0x926(%rip),%ymm13 # 6a80 <_sk_callback_avx+0x4b4>
.byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13
- .byte 196,98,125,24,53,28,9,0,0 // vbroadcastss 0x91c(%rip),%ymm14 # 68b4 <_sk_callback_avx+0x4b8>
+ .byte 196,98,125,24,53,28,9,0,0 // vbroadcastss 0x91c(%rip),%ymm14 # 6a84 <_sk_callback_avx+0x4b8>
.byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
.byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13
- .byte 196,98,125,24,53,13,9,0,0 // vbroadcastss 0x90d(%rip),%ymm14 # 68b8 <_sk_callback_avx+0x4bc>
+ .byte 196,98,125,24,53,13,9,0,0 // vbroadcastss 0x90d(%rip),%ymm14 # 6a88 <_sk_callback_avx+0x4bc>
.byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
.byte 196,65,28,89,229 // vmulps %ymm13,%ymm12,%ymm12
- .byte 196,98,125,24,45,254,8,0,0 // vbroadcastss 0x8fe(%rip),%ymm13 # 68bc <_sk_callback_avx+0x4c0>
+ .byte 196,98,125,24,45,254,8,0,0 // vbroadcastss 0x8fe(%rip),%ymm13 # 6a8c <_sk_callback_avx+0x4c0>
.byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
.byte 196,65,52,194,202,1 // vcmpltps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,24,21,233,8,0,0 // vbroadcastss 0x8e9(%rip),%ymm10 # 68c0 <_sk_callback_avx+0x4c4>
+ .byte 196,98,125,24,21,233,8,0,0 // vbroadcastss 0x8e9(%rip),%ymm10 # 6a90 <_sk_callback_avx+0x4c4>
.byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10
.byte 196,67,37,74,202,144 // vblendvps %ymm9,%ymm10,%ymm11,%ymm9
.byte 196,193,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,21,211,8,0,0 // vbroadcastss 0x8d3(%rip),%ymm10 # 68c4 <_sk_callback_avx+0x4c8>
+ .byte 196,98,125,24,21,211,8,0,0 // vbroadcastss 0x8d3(%rip),%ymm10 # 6a94 <_sk_callback_avx+0x4c8>
.byte 196,65,44,92,209 // vsubps %ymm9,%ymm10,%ymm10
.byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0
.byte 196,65,116,194,200,1 // vcmpltps %ymm8,%ymm1,%ymm9
- .byte 196,98,125,24,21,189,8,0,0 // vbroadcastss 0x8bd(%rip),%ymm10 # 68c8 <_sk_callback_avx+0x4cc>
+ .byte 196,98,125,24,21,189,8,0,0 // vbroadcastss 0x8bd(%rip),%ymm10 # 6a98 <_sk_callback_avx+0x4cc>
.byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10
.byte 196,195,125,74,194,144 // vblendvps %ymm9,%ymm10,%ymm0,%ymm0
.byte 196,65,124,194,200,3 // vcmpunordps %ymm8,%ymm0,%ymm9
@@ -20740,7 +21026,7 @@ HIDDEN _sk_save_xy_avx
FUNCTION(_sk_save_xy_avx)
_sk_save_xy_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,135,8,0,0 // vbroadcastss 0x887(%rip),%ymm8 # 68cc <_sk_callback_avx+0x4d0>
+ .byte 196,98,125,24,5,135,8,0,0 // vbroadcastss 0x887(%rip),%ymm8 # 6a9c <_sk_callback_avx+0x4d0>
.byte 196,65,124,88,200 // vaddps %ymm8,%ymm0,%ymm9
.byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10
.byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9
@@ -20777,9 +21063,9 @@ HIDDEN _sk_bilinear_nx_avx
FUNCTION(_sk_bilinear_nx_avx)
_sk_bilinear_nx_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,19,8,0,0 // vbroadcastss 0x813(%rip),%ymm0 # 68d0 <_sk_callback_avx+0x4d4>
+ .byte 196,226,125,24,5,19,8,0,0 // vbroadcastss 0x813(%rip),%ymm0 # 6aa0 <_sk_callback_avx+0x4d4>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,10,8,0,0 // vbroadcastss 0x80a(%rip),%ymm8 # 68d4 <_sk_callback_avx+0x4d8>
+ .byte 196,98,125,24,5,10,8,0,0 // vbroadcastss 0x80a(%rip),%ymm8 # 6aa4 <_sk_callback_avx+0x4d8>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20790,7 +21076,7 @@ HIDDEN _sk_bilinear_px_avx
FUNCTION(_sk_bilinear_px_avx)
_sk_bilinear_px_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,242,7,0,0 // vbroadcastss 0x7f2(%rip),%ymm0 # 68d8 <_sk_callback_avx+0x4dc>
+ .byte 196,226,125,24,5,242,7,0,0 // vbroadcastss 0x7f2(%rip),%ymm0 # 6aa8 <_sk_callback_avx+0x4dc>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
.byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
@@ -20802,9 +21088,9 @@ HIDDEN _sk_bilinear_ny_avx
FUNCTION(_sk_bilinear_ny_avx)
_sk_bilinear_ny_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,214,7,0,0 // vbroadcastss 0x7d6(%rip),%ymm1 # 68dc <_sk_callback_avx+0x4e0>
+ .byte 196,226,125,24,13,214,7,0,0 // vbroadcastss 0x7d6(%rip),%ymm1 # 6aac <_sk_callback_avx+0x4e0>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,204,7,0,0 // vbroadcastss 0x7cc(%rip),%ymm8 # 68e0 <_sk_callback_avx+0x4e4>
+ .byte 196,98,125,24,5,204,7,0,0 // vbroadcastss 0x7cc(%rip),%ymm8 # 6ab0 <_sk_callback_avx+0x4e4>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20815,7 +21101,7 @@ HIDDEN _sk_bilinear_py_avx
FUNCTION(_sk_bilinear_py_avx)
_sk_bilinear_py_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,180,7,0,0 // vbroadcastss 0x7b4(%rip),%ymm1 # 68e4 <_sk_callback_avx+0x4e8>
+ .byte 196,226,125,24,13,180,7,0,0 // vbroadcastss 0x7b4(%rip),%ymm1 # 6ab4 <_sk_callback_avx+0x4e8>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
.byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
@@ -20827,14 +21113,14 @@ HIDDEN _sk_bicubic_n3x_avx
FUNCTION(_sk_bicubic_n3x_avx)
_sk_bicubic_n3x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,151,7,0,0 // vbroadcastss 0x797(%rip),%ymm0 # 68e8 <_sk_callback_avx+0x4ec>
+ .byte 196,226,125,24,5,151,7,0,0 // vbroadcastss 0x797(%rip),%ymm0 # 6ab8 <_sk_callback_avx+0x4ec>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,142,7,0,0 // vbroadcastss 0x78e(%rip),%ymm8 # 68ec <_sk_callback_avx+0x4f0>
+ .byte 196,98,125,24,5,142,7,0,0 // vbroadcastss 0x78e(%rip),%ymm8 # 6abc <_sk_callback_avx+0x4f0>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,127,7,0,0 // vbroadcastss 0x77f(%rip),%ymm10 # 68f0 <_sk_callback_avx+0x4f4>
+ .byte 196,98,125,24,21,127,7,0,0 // vbroadcastss 0x77f(%rip),%ymm10 # 6ac0 <_sk_callback_avx+0x4f4>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,117,7,0,0 // vbroadcastss 0x775(%rip),%ymm10 # 68f4 <_sk_callback_avx+0x4f8>
+ .byte 196,98,125,24,21,117,7,0,0 // vbroadcastss 0x775(%rip),%ymm10 # 6ac4 <_sk_callback_avx+0x4f8>
.byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
@@ -20846,19 +21132,19 @@ HIDDEN _sk_bicubic_n1x_avx
FUNCTION(_sk_bicubic_n1x_avx)
_sk_bicubic_n1x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,88,7,0,0 // vbroadcastss 0x758(%rip),%ymm0 # 68f8 <_sk_callback_avx+0x4fc>
+ .byte 196,226,125,24,5,88,7,0,0 // vbroadcastss 0x758(%rip),%ymm0 # 6ac8 <_sk_callback_avx+0x4fc>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,79,7,0,0 // vbroadcastss 0x74f(%rip),%ymm8 # 68fc <_sk_callback_avx+0x500>
+ .byte 196,98,125,24,5,79,7,0,0 // vbroadcastss 0x74f(%rip),%ymm8 # 6acc <_sk_callback_avx+0x500>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,69,7,0,0 // vbroadcastss 0x745(%rip),%ymm9 # 6900 <_sk_callback_avx+0x504>
+ .byte 196,98,125,24,13,69,7,0,0 // vbroadcastss 0x745(%rip),%ymm9 # 6ad0 <_sk_callback_avx+0x504>
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,59,7,0,0 // vbroadcastss 0x73b(%rip),%ymm10 # 6904 <_sk_callback_avx+0x508>
+ .byte 196,98,125,24,21,59,7,0,0 // vbroadcastss 0x73b(%rip),%ymm10 # 6ad4 <_sk_callback_avx+0x508>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,44,7,0,0 // vbroadcastss 0x72c(%rip),%ymm10 # 6908 <_sk_callback_avx+0x50c>
+ .byte 196,98,125,24,21,44,7,0,0 // vbroadcastss 0x72c(%rip),%ymm10 # 6ad8 <_sk_callback_avx+0x50c>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,29,7,0,0 // vbroadcastss 0x71d(%rip),%ymm9 # 690c <_sk_callback_avx+0x510>
+ .byte 196,98,125,24,13,29,7,0,0 // vbroadcastss 0x71d(%rip),%ymm9 # 6adc <_sk_callback_avx+0x510>
.byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20869,17 +21155,17 @@ HIDDEN _sk_bicubic_p1x_avx
FUNCTION(_sk_bicubic_p1x_avx)
_sk_bicubic_p1x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,5,7,0,0 // vbroadcastss 0x705(%rip),%ymm8 # 6910 <_sk_callback_avx+0x514>
+ .byte 196,98,125,24,5,5,7,0,0 // vbroadcastss 0x705(%rip),%ymm8 # 6ae0 <_sk_callback_avx+0x514>
.byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
.byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
- .byte 196,98,125,24,21,247,6,0,0 // vbroadcastss 0x6f7(%rip),%ymm10 # 6914 <_sk_callback_avx+0x518>
+ .byte 196,98,125,24,21,247,6,0,0 // vbroadcastss 0x6f7(%rip),%ymm10 # 6ae4 <_sk_callback_avx+0x518>
.byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
- .byte 196,98,125,24,29,237,6,0,0 // vbroadcastss 0x6ed(%rip),%ymm11 # 6918 <_sk_callback_avx+0x51c>
+ .byte 196,98,125,24,29,237,6,0,0 // vbroadcastss 0x6ed(%rip),%ymm11 # 6ae8 <_sk_callback_avx+0x51c>
.byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
.byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
.byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
- .byte 196,98,125,24,13,212,6,0,0 // vbroadcastss 0x6d4(%rip),%ymm9 # 691c <_sk_callback_avx+0x520>
+ .byte 196,98,125,24,13,212,6,0,0 // vbroadcastss 0x6d4(%rip),%ymm9 # 6aec <_sk_callback_avx+0x520>
.byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20890,13 +21176,13 @@ HIDDEN _sk_bicubic_p3x_avx
FUNCTION(_sk_bicubic_p3x_avx)
_sk_bicubic_p3x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,188,6,0,0 // vbroadcastss 0x6bc(%rip),%ymm0 # 6920 <_sk_callback_avx+0x524>
+ .byte 196,226,125,24,5,188,6,0,0 // vbroadcastss 0x6bc(%rip),%ymm0 # 6af0 <_sk_callback_avx+0x524>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
.byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,169,6,0,0 // vbroadcastss 0x6a9(%rip),%ymm10 # 6924 <_sk_callback_avx+0x528>
+ .byte 196,98,125,24,21,169,6,0,0 // vbroadcastss 0x6a9(%rip),%ymm10 # 6af4 <_sk_callback_avx+0x528>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,159,6,0,0 // vbroadcastss 0x69f(%rip),%ymm10 # 6928 <_sk_callback_avx+0x52c>
+ .byte 196,98,125,24,21,159,6,0,0 // vbroadcastss 0x69f(%rip),%ymm10 # 6af8 <_sk_callback_avx+0x52c>
.byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
.byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
@@ -20908,14 +21194,14 @@ HIDDEN _sk_bicubic_n3y_avx
FUNCTION(_sk_bicubic_n3y_avx)
_sk_bicubic_n3y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,130,6,0,0 // vbroadcastss 0x682(%rip),%ymm1 # 692c <_sk_callback_avx+0x530>
+ .byte 196,226,125,24,13,130,6,0,0 // vbroadcastss 0x682(%rip),%ymm1 # 6afc <_sk_callback_avx+0x530>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,120,6,0,0 // vbroadcastss 0x678(%rip),%ymm8 # 6930 <_sk_callback_avx+0x534>
+ .byte 196,98,125,24,5,120,6,0,0 // vbroadcastss 0x678(%rip),%ymm8 # 6b00 <_sk_callback_avx+0x534>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,105,6,0,0 // vbroadcastss 0x669(%rip),%ymm10 # 6934 <_sk_callback_avx+0x538>
+ .byte 196,98,125,24,21,105,6,0,0 // vbroadcastss 0x669(%rip),%ymm10 # 6b04 <_sk_callback_avx+0x538>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,95,6,0,0 // vbroadcastss 0x65f(%rip),%ymm10 # 6938 <_sk_callback_avx+0x53c>
+ .byte 196,98,125,24,21,95,6,0,0 // vbroadcastss 0x65f(%rip),%ymm10 # 6b08 <_sk_callback_avx+0x53c>
.byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
@@ -20927,19 +21213,19 @@ HIDDEN _sk_bicubic_n1y_avx
FUNCTION(_sk_bicubic_n1y_avx)
_sk_bicubic_n1y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,66,6,0,0 // vbroadcastss 0x642(%rip),%ymm1 # 693c <_sk_callback_avx+0x540>
+ .byte 196,226,125,24,13,66,6,0,0 // vbroadcastss 0x642(%rip),%ymm1 # 6b0c <_sk_callback_avx+0x540>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,56,6,0,0 // vbroadcastss 0x638(%rip),%ymm8 # 6940 <_sk_callback_avx+0x544>
+ .byte 196,98,125,24,5,56,6,0,0 // vbroadcastss 0x638(%rip),%ymm8 # 6b10 <_sk_callback_avx+0x544>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,46,6,0,0 // vbroadcastss 0x62e(%rip),%ymm9 # 6944 <_sk_callback_avx+0x548>
+ .byte 196,98,125,24,13,46,6,0,0 // vbroadcastss 0x62e(%rip),%ymm9 # 6b14 <_sk_callback_avx+0x548>
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,36,6,0,0 // vbroadcastss 0x624(%rip),%ymm10 # 6948 <_sk_callback_avx+0x54c>
+ .byte 196,98,125,24,21,36,6,0,0 // vbroadcastss 0x624(%rip),%ymm10 # 6b18 <_sk_callback_avx+0x54c>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,21,6,0,0 // vbroadcastss 0x615(%rip),%ymm10 # 694c <_sk_callback_avx+0x550>
+ .byte 196,98,125,24,21,21,6,0,0 // vbroadcastss 0x615(%rip),%ymm10 # 6b1c <_sk_callback_avx+0x550>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,6,6,0,0 // vbroadcastss 0x606(%rip),%ymm9 # 6950 <_sk_callback_avx+0x554>
+ .byte 196,98,125,24,13,6,6,0,0 // vbroadcastss 0x606(%rip),%ymm9 # 6b20 <_sk_callback_avx+0x554>
.byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20950,17 +21236,17 @@ HIDDEN _sk_bicubic_p1y_avx
FUNCTION(_sk_bicubic_p1y_avx)
_sk_bicubic_p1y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,238,5,0,0 // vbroadcastss 0x5ee(%rip),%ymm8 # 6954 <_sk_callback_avx+0x558>
+ .byte 196,98,125,24,5,238,5,0,0 // vbroadcastss 0x5ee(%rip),%ymm8 # 6b24 <_sk_callback_avx+0x558>
.byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
.byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
- .byte 196,98,125,24,21,223,5,0,0 // vbroadcastss 0x5df(%rip),%ymm10 # 6958 <_sk_callback_avx+0x55c>
+ .byte 196,98,125,24,21,223,5,0,0 // vbroadcastss 0x5df(%rip),%ymm10 # 6b28 <_sk_callback_avx+0x55c>
.byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
- .byte 196,98,125,24,29,213,5,0,0 // vbroadcastss 0x5d5(%rip),%ymm11 # 695c <_sk_callback_avx+0x560>
+ .byte 196,98,125,24,29,213,5,0,0 // vbroadcastss 0x5d5(%rip),%ymm11 # 6b2c <_sk_callback_avx+0x560>
.byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
.byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
.byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
- .byte 196,98,125,24,13,188,5,0,0 // vbroadcastss 0x5bc(%rip),%ymm9 # 6960 <_sk_callback_avx+0x564>
+ .byte 196,98,125,24,13,188,5,0,0 // vbroadcastss 0x5bc(%rip),%ymm9 # 6b30 <_sk_callback_avx+0x564>
.byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
.byte 72,173 // lods %ds:(%rsi),%rax
@@ -20971,13 +21257,13 @@ HIDDEN _sk_bicubic_p3y_avx
FUNCTION(_sk_bicubic_p3y_avx)
_sk_bicubic_p3y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,164,5,0,0 // vbroadcastss 0x5a4(%rip),%ymm1 # 6964 <_sk_callback_avx+0x568>
+ .byte 196,226,125,24,13,164,5,0,0 // vbroadcastss 0x5a4(%rip),%ymm1 # 6b34 <_sk_callback_avx+0x568>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
.byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,144,5,0,0 // vbroadcastss 0x590(%rip),%ymm10 # 6968 <_sk_callback_avx+0x56c>
+ .byte 196,98,125,24,21,144,5,0,0 // vbroadcastss 0x590(%rip),%ymm10 # 6b38 <_sk_callback_avx+0x56c>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,134,5,0,0 // vbroadcastss 0x586(%rip),%ymm10 # 696c <_sk_callback_avx+0x570>
+ .byte 196,98,125,24,21,134,5,0,0 // vbroadcastss 0x586(%rip),%ymm10 # 6b3c <_sk_callback_avx+0x570>
.byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8
.byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
.byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
@@ -21107,25 +21393,25 @@ BALIGN4
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 65f1 <.literal4+0xb1>
+ .byte 71,225,61 // rex.RXB loope 67c1 <.literal4+0xb1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,154 // cmpb $0x9a,(%rdi)
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 6601 <.literal4+0xc1>
+ .byte 71,225,61 // rex.RXB loope 67d1 <.literal4+0xc1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,154 // cmpb $0x9a,(%rdi)
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 6611 <.literal4+0xd1>
+ .byte 71,225,61 // rex.RXB loope 67e1 <.literal4+0xd1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,154 // cmpb $0x9a,(%rdi)
.byte 153 // cltd
.byte 153 // cltd
.byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax
- .byte 71,225,61 // rex.RXB loope 6621 <.literal4+0xe1>
+ .byte 71,225,61 // rex.RXB loope 67f1 <.literal4+0xe1>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax)
@@ -21177,7 +21463,7 @@ BALIGN4
.byte 190,129,128,128,59 // mov $0x3b808081,%esi
.byte 129,128,128,59,0,248,0,0,8,33 // addl $0x21080000,-0x7ffc480(%rax)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 6675 <.literal4+0x135>
+ .byte 224,7 // loopne 6845 <.literal4+0x135>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -21193,10 +21479,10 @@ BALIGN4
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
.byte 0,52,255 // add %dh,(%rdi,%rdi,8)
.byte 255 // (bad)
- .byte 127,0 // jg 669c <.literal4+0x15c>
+ .byte 127,0 // jg 686c <.literal4+0x15c>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 6715 <.literal4+0x1d5>
+ .byte 119,115 // ja 68e5 <.literal4+0x1d5>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -21210,10 +21496,10 @@ BALIGN4
.byte 0,128,63,0,0,0 // add %al,0x3f(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 66d0 <.literal4+0x190>
+ .byte 127,0 // jg 68a0 <.literal4+0x190>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 6749 <.literal4+0x209>
+ .byte 119,115 // ja 6919 <.literal4+0x209>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -21227,10 +21513,10 @@ BALIGN4
.byte 0,128,63,0,0,0 // add %al,0x3f(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 6704 <.literal4+0x1c4>
+ .byte 127,0 // jg 68d4 <.literal4+0x1c4>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 677d <.literal4+0x23d>
+ .byte 119,115 // ja 694d <.literal4+0x23d>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -21244,10 +21530,10 @@ BALIGN4
.byte 0,128,63,0,0,0 // add %al,0x3f(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 6738 <.literal4+0x1f8>
+ .byte 127,0 // jg 6908 <.literal4+0x1f8>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 67b1 <.literal4+0x271>
+ .byte 119,115 // ja 6981 <.literal4+0x271>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -21260,7 +21546,7 @@ BALIGN4
.byte 0,75,0 // add %cl,0x0(%rbx)
.byte 0,128,63,0,0,200 // add %al,-0x37ffffc1(%rax)
.byte 66,0,0 // rex.X add %al,(%rax)
- .byte 127,67 // jg 67af <.literal4+0x26f>
+ .byte 127,67 // jg 697f <.literal4+0x26f>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -21272,10 +21558,10 @@ BALIGN4
.byte 190,80,128,3,62 // mov $0x3e038050,%esi
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 67cf <.literal4+0x28f>
+ .byte 118,63 // jbe 699f <.literal4+0x28f>
.byte 246,64,83,63 // testb $0x3f,0x53(%rax)
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 67e3 <.literal4+0x2a3>
+ .byte 127,67 // jg 69b3 <.literal4+0x2a3>
.byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax)
.byte 128,59,0 // cmpb $0x0,(%rbx)
.byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax)
@@ -21284,7 +21570,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 67c5 <.literal4+0x285>
+ .byte 224,7 // loopne 6995 <.literal4+0x285>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -21296,7 +21582,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 67e1 <.literal4+0x2a1>
+ .byte 224,7 // loopne 69b1 <.literal4+0x2a1>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -21307,7 +21593,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 6836 <.literal4+0x2f6>
+ .byte 124,66 // jl 6a06 <.literal4+0x2f6>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax)
@@ -21325,9 +21611,9 @@ BALIGN4
.byte 137,136,136,59,15,0 // mov %ecx,0xf3b88(%rax)
.byte 0,0 // add %al,(%rax)
.byte 137,136,136,61,0,0 // mov %ecx,0x3d88(%rax)
- .byte 112,65 // jo 6879 <.literal4+0x339>
+ .byte 112,65 // jo 6a49 <.literal4+0x339>
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 6887 <.literal4+0x347>
+ .byte 127,67 // jg 6a57 <.literal4+0x347>
.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)
@@ -21343,7 +21629,7 @@ BALIGN4
.byte 0,128,55,0,0,128 // add %al,-0x7fffffc9(%rax)
.byte 63 // (bad)
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 68c7 <.literal4+0x387>
+ .byte 127,71 // jg 6a97 <.literal4+0x387>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,191,0,0,0 // add %al,0xbf(%rax)
@@ -21456,6 +21742,22 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 0,0 // add %al,(%rax)
.byte 0,0 // add %al,(%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
.byte 0,2 // add %al,(%rdx)
.byte 4,6 // add $0x6,%al
.byte 0,0 // add %al,(%rax)
@@ -21488,6 +21790,22 @@ BALIGN16
.byte 255,0 // incl (%rax)
.byte 255,0 // incl (%rax)
.byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
+ .byte 255,0 // incl (%rax)
.byte 0,2 // add %al,(%rdx)
.byte 4,6 // add $0x6,%al
.byte 0,0 // add %al,(%rax)
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index 99b5d7a374..d7cdd3bc0a 100644
--- a/src/jumper/SkJumper_generated_win.S
+++ b/src/jumper/SkJumper_generated_win.S
@@ -110,7 +110,7 @@ _sk_seed_shader_hsw LABEL PROC
DB 197,249,110,194 ; vmovd %edx,%xmm0
DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,213,72,0,0 ; vbroadcastss 0x48d5(%rip),%ymm1 # 4a48 <_sk_callback_hsw+0x12c>
+ DB 196,226,125,24,13,173,74,0,0 ; vbroadcastss 0x4aad(%rip),%ymm1 # 4c20 <_sk_callback_hsw+0x12c>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,88,7 ; vaddps (%rdi),%ymm0,%ymm0
DB 197,249,110,209 ; vmovd %ecx,%xmm2
@@ -118,7 +118,7 @@ _sk_seed_shader_hsw LABEL PROC
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 196,226,125,24,21,181,72,0,0 ; vbroadcastss 0x48b5(%rip),%ymm2 # 4a4c <_sk_callback_hsw+0x130>
+ DB 196,226,125,24,21,141,74,0,0 ; vbroadcastss 0x4a8d(%rip),%ymm2 # 4c24 <_sk_callback_hsw+0x130>
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
@@ -135,13 +135,13 @@ _sk_dither_hsw LABEL PROC
DB 197,121,110,201 ; vmovd %ecx,%xmm9
DB 196,66,125,88,201 ; vpbroadcastd %xmm9,%ymm9
DB 196,65,53,239,200 ; vpxor %ymm8,%ymm9,%ymm9
- DB 196,98,125,88,21,124,72,0,0 ; vpbroadcastd 0x487c(%rip),%ymm10 # 4a50 <_sk_callback_hsw+0x134>
+ DB 196,98,125,88,21,84,74,0,0 ; vpbroadcastd 0x4a54(%rip),%ymm10 # 4c28 <_sk_callback_hsw+0x134>
DB 196,65,53,219,218 ; vpand %ymm10,%ymm9,%ymm11
DB 196,193,37,114,243,5 ; vpslld $0x5,%ymm11,%ymm11
DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10
DB 196,193,45,114,242,4 ; vpslld $0x4,%ymm10,%ymm10
- DB 196,98,125,88,37,97,72,0,0 ; vpbroadcastd 0x4861(%rip),%ymm12 # 4a54 <_sk_callback_hsw+0x138>
- DB 196,98,125,88,45,92,72,0,0 ; vpbroadcastd 0x485c(%rip),%ymm13 # 4a58 <_sk_callback_hsw+0x13c>
+ DB 196,98,125,88,37,57,74,0,0 ; vpbroadcastd 0x4a39(%rip),%ymm12 # 4c2c <_sk_callback_hsw+0x138>
+ DB 196,98,125,88,45,52,74,0,0 ; vpbroadcastd 0x4a34(%rip),%ymm13 # 4c30 <_sk_callback_hsw+0x13c>
DB 196,65,53,219,245 ; vpand %ymm13,%ymm9,%ymm14
DB 196,193,13,114,246,2 ; vpslld $0x2,%ymm14,%ymm14
DB 196,65,61,219,237 ; vpand %ymm13,%ymm8,%ymm13
@@ -156,8 +156,8 @@ _sk_dither_hsw LABEL PROC
DB 196,65,61,235,194 ; vpor %ymm10,%ymm8,%ymm8
DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,14,72,0,0 ; vbroadcastss 0x480e(%rip),%ymm9 # 4a5c <_sk_callback_hsw+0x140>
- DB 196,98,125,24,21,9,72,0,0 ; vbroadcastss 0x4809(%rip),%ymm10 # 4a60 <_sk_callback_hsw+0x144>
+ DB 196,98,125,24,13,230,73,0,0 ; vbroadcastss 0x49e6(%rip),%ymm9 # 4c34 <_sk_callback_hsw+0x140>
+ DB 196,98,125,24,21,225,73,0,0 ; vbroadcastss 0x49e1(%rip),%ymm10 # 4c38 <_sk_callback_hsw+0x144>
DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10
DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
@@ -216,7 +216,7 @@ _sk_clear_hsw LABEL PROC
PUBLIC _sk_srcatop_hsw
_sk_srcatop_hsw LABEL PROC
DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0
- DB 196,98,125,24,5,97,71,0,0 ; vbroadcastss 0x4761(%rip),%ymm8 # 4a64 <_sk_callback_hsw+0x148>
+ DB 196,98,125,24,5,57,73,0,0 ; vbroadcastss 0x4939(%rip),%ymm8 # 4c3c <_sk_callback_hsw+0x148>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 196,226,61,184,196 ; vfmadd231ps %ymm4,%ymm8,%ymm0
DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1
@@ -230,7 +230,7 @@ _sk_srcatop_hsw LABEL PROC
PUBLIC _sk_dstatop_hsw
_sk_dstatop_hsw LABEL PROC
- DB 196,98,125,24,5,52,71,0,0 ; vbroadcastss 0x4734(%rip),%ymm8 # 4a68 <_sk_callback_hsw+0x14c>
+ DB 196,98,125,24,5,12,73,0,0 ; vbroadcastss 0x490c(%rip),%ymm8 # 4c40 <_sk_callback_hsw+0x14c>
DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
DB 196,226,101,184,196 ; vfmadd231ps %ymm4,%ymm3,%ymm0
@@ -263,7 +263,7 @@ _sk_dstin_hsw LABEL PROC
PUBLIC _sk_srcout_hsw
_sk_srcout_hsw LABEL PROC
- DB 196,98,125,24,5,219,70,0,0 ; vbroadcastss 0x46db(%rip),%ymm8 # 4a6c <_sk_callback_hsw+0x150>
+ DB 196,98,125,24,5,179,72,0,0 ; vbroadcastss 0x48b3(%rip),%ymm8 # 4c44 <_sk_callback_hsw+0x150>
DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
@@ -274,7 +274,7 @@ _sk_srcout_hsw LABEL PROC
PUBLIC _sk_dstout_hsw
_sk_dstout_hsw LABEL PROC
- DB 196,226,125,24,5,190,70,0,0 ; vbroadcastss 0x46be(%rip),%ymm0 # 4a70 <_sk_callback_hsw+0x154>
+ DB 196,226,125,24,5,150,72,0,0 ; vbroadcastss 0x4896(%rip),%ymm0 # 4c48 <_sk_callback_hsw+0x154>
DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3
DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0
DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1
@@ -285,7 +285,7 @@ _sk_dstout_hsw LABEL PROC
PUBLIC _sk_srcover_hsw
_sk_srcover_hsw LABEL PROC
- DB 196,98,125,24,5,161,70,0,0 ; vbroadcastss 0x46a1(%rip),%ymm8 # 4a74 <_sk_callback_hsw+0x158>
+ DB 196,98,125,24,5,121,72,0,0 ; vbroadcastss 0x4879(%rip),%ymm8 # 4c4c <_sk_callback_hsw+0x158>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0
DB 196,194,85,184,200 ; vfmadd231ps %ymm8,%ymm5,%ymm1
@@ -296,7 +296,7 @@ _sk_srcover_hsw LABEL PROC
PUBLIC _sk_dstover_hsw
_sk_dstover_hsw LABEL PROC
- DB 196,98,125,24,5,128,70,0,0 ; vbroadcastss 0x4680(%rip),%ymm8 # 4a78 <_sk_callback_hsw+0x15c>
+ DB 196,98,125,24,5,88,72,0,0 ; vbroadcastss 0x4858(%rip),%ymm8 # 4c50 <_sk_callback_hsw+0x15c>
DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8
DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0
DB 196,226,61,168,205 ; vfmadd213ps %ymm5,%ymm8,%ymm1
@@ -316,7 +316,7 @@ _sk_modulate_hsw LABEL PROC
PUBLIC _sk_multiply_hsw
_sk_multiply_hsw LABEL PROC
- DB 196,98,125,24,5,75,70,0,0 ; vbroadcastss 0x464b(%rip),%ymm8 # 4a7c <_sk_callback_hsw+0x160>
+ DB 196,98,125,24,5,35,72,0,0 ; vbroadcastss 0x4823(%rip),%ymm8 # 4c54 <_sk_callback_hsw+0x160>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -358,7 +358,7 @@ _sk_screen_hsw LABEL PROC
PUBLIC _sk_xor__hsw
_sk_xor__hsw LABEL PROC
- DB 196,98,125,24,5,198,69,0,0 ; vbroadcastss 0x45c6(%rip),%ymm8 # 4a80 <_sk_callback_hsw+0x164>
+ DB 196,98,125,24,5,158,71,0,0 ; vbroadcastss 0x479e(%rip),%ymm8 # 4c58 <_sk_callback_hsw+0x164>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -390,7 +390,7 @@ _sk_darken_hsw LABEL PROC
DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9
DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,78,69,0,0 ; vbroadcastss 0x454e(%rip),%ymm8 # 4a84 <_sk_callback_hsw+0x168>
+ DB 196,98,125,24,5,38,71,0,0 ; vbroadcastss 0x4726(%rip),%ymm8 # 4c5c <_sk_callback_hsw+0x168>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -413,7 +413,7 @@ _sk_lighten_hsw LABEL PROC
DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9
DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,253,68,0,0 ; vbroadcastss 0x44fd(%rip),%ymm8 # 4a88 <_sk_callback_hsw+0x16c>
+ DB 196,98,125,24,5,213,70,0,0 ; vbroadcastss 0x46d5(%rip),%ymm8 # 4c60 <_sk_callback_hsw+0x16c>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -439,7 +439,7 @@ _sk_difference_hsw LABEL PROC
DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2
DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,160,68,0,0 ; vbroadcastss 0x44a0(%rip),%ymm8 # 4a8c <_sk_callback_hsw+0x170>
+ DB 196,98,125,24,5,120,70,0,0 ; vbroadcastss 0x4678(%rip),%ymm8 # 4c64 <_sk_callback_hsw+0x170>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -459,7 +459,7 @@ _sk_exclusion_hsw LABEL PROC
DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2
DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,94,68,0,0 ; vbroadcastss 0x445e(%rip),%ymm8 # 4a90 <_sk_callback_hsw+0x174>
+ DB 196,98,125,24,5,54,70,0,0 ; vbroadcastss 0x4636(%rip),%ymm8 # 4c68 <_sk_callback_hsw+0x174>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -467,7 +467,7 @@ _sk_exclusion_hsw LABEL PROC
PUBLIC _sk_colorburn_hsw
_sk_colorburn_hsw LABEL PROC
- DB 196,98,125,24,5,76,68,0,0 ; vbroadcastss 0x444c(%rip),%ymm8 # 4a94 <_sk_callback_hsw+0x178>
+ DB 196,98,125,24,5,36,70,0,0 ; vbroadcastss 0x4624(%rip),%ymm8 # 4c6c <_sk_callback_hsw+0x178>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11
DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10
@@ -523,7 +523,7 @@ _sk_colorburn_hsw LABEL PROC
PUBLIC _sk_colordodge_hsw
_sk_colordodge_hsw LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
- DB 196,98,125,24,13,87,67,0,0 ; vbroadcastss 0x4357(%rip),%ymm9 # 4a98 <_sk_callback_hsw+0x17c>
+ DB 196,98,125,24,13,47,69,0,0 ; vbroadcastss 0x452f(%rip),%ymm9 # 4c70 <_sk_callback_hsw+0x17c>
DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10
DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11
DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9
@@ -574,7 +574,7 @@ _sk_colordodge_hsw LABEL PROC
PUBLIC _sk_hardlight_hsw
_sk_hardlight_hsw LABEL PROC
- DB 196,98,125,24,5,120,66,0,0 ; vbroadcastss 0x4278(%rip),%ymm8 # 4a9c <_sk_callback_hsw+0x180>
+ DB 196,98,125,24,5,80,68,0,0 ; vbroadcastss 0x4450(%rip),%ymm8 # 4c74 <_sk_callback_hsw+0x180>
DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10
DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -623,7 +623,7 @@ _sk_hardlight_hsw LABEL PROC
PUBLIC _sk_overlay_hsw
_sk_overlay_hsw LABEL PROC
- DB 196,98,125,24,5,176,65,0,0 ; vbroadcastss 0x41b0(%rip),%ymm8 # 4aa0 <_sk_callback_hsw+0x184>
+ DB 196,98,125,24,5,136,67,0,0 ; vbroadcastss 0x4388(%rip),%ymm8 # 4c78 <_sk_callback_hsw+0x184>
DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10
DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -683,10 +683,10 @@ _sk_softlight_hsw LABEL PROC
DB 196,65,20,88,197 ; vaddps %ymm13,%ymm13,%ymm8
DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8
DB 196,66,61,168,192 ; vfmadd213ps %ymm8,%ymm8,%ymm8
- DB 196,98,125,24,29,183,64,0,0 ; vbroadcastss 0x40b7(%rip),%ymm11 # 4aa8 <_sk_callback_hsw+0x18c>
+ DB 196,98,125,24,29,143,66,0,0 ; vbroadcastss 0x428f(%rip),%ymm11 # 4c80 <_sk_callback_hsw+0x18c>
DB 196,65,20,88,227 ; vaddps %ymm11,%ymm13,%ymm12
DB 196,65,28,89,192 ; vmulps %ymm8,%ymm12,%ymm8
- DB 196,98,125,24,37,168,64,0,0 ; vbroadcastss 0x40a8(%rip),%ymm12 # 4aac <_sk_callback_hsw+0x190>
+ DB 196,98,125,24,37,128,66,0,0 ; vbroadcastss 0x4280(%rip),%ymm12 # 4c84 <_sk_callback_hsw+0x190>
DB 196,66,21,184,196 ; vfmadd231ps %ymm12,%ymm13,%ymm8
DB 196,65,124,82,245 ; vrsqrtps %ymm13,%ymm14
DB 196,65,124,83,246 ; vrcpps %ymm14,%ymm14
@@ -696,7 +696,7 @@ _sk_softlight_hsw LABEL PROC
DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15
DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14
DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15
- DB 196,98,125,24,5,107,64,0,0 ; vbroadcastss 0x406b(%rip),%ymm8 # 4aa4 <_sk_callback_hsw+0x188>
+ DB 196,98,125,24,5,67,66,0,0 ; vbroadcastss 0x4243(%rip),%ymm8 # 4c7c <_sk_callback_hsw+0x188>
DB 196,65,60,92,237 ; vsubps %ymm13,%ymm8,%ymm13
DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0
DB 196,98,125,168,235 ; vfmadd213ps %ymm3,%ymm0,%ymm13
@@ -809,11 +809,11 @@ _sk_hue_hsw LABEL PROC
DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10
DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10
DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12
- DB 196,98,125,24,53,106,62,0,0 ; vbroadcastss 0x3e6a(%rip),%ymm14 # 4ab0 <_sk_callback_hsw+0x194>
- DB 196,98,125,24,61,101,62,0,0 ; vbroadcastss 0x3e65(%rip),%ymm15 # 4ab4 <_sk_callback_hsw+0x198>
+ DB 196,98,125,24,53,66,64,0,0 ; vbroadcastss 0x4042(%rip),%ymm14 # 4c88 <_sk_callback_hsw+0x194>
+ DB 196,98,125,24,61,61,64,0,0 ; vbroadcastss 0x403d(%rip),%ymm15 # 4c8c <_sk_callback_hsw+0x198>
DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13
DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13
- DB 196,226,125,24,5,86,62,0,0 ; vbroadcastss 0x3e56(%rip),%ymm0 # 4ab8 <_sk_callback_hsw+0x19c>
+ DB 196,226,125,24,5,46,64,0,0 ; vbroadcastss 0x402e(%rip),%ymm0 # 4c90 <_sk_callback_hsw+0x19c>
DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13
DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10
DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10
@@ -868,7 +868,7 @@ _sk_hue_hsw LABEL PROC
DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0
DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9
DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8
- DB 196,226,125,24,13,67,61,0,0 ; vbroadcastss 0x3d43(%rip),%ymm1 # 4abc <_sk_callback_hsw+0x1a0>
+ DB 196,226,125,24,13,27,63,0,0 ; vbroadcastss 0x3f1b(%rip),%ymm1 # 4c94 <_sk_callback_hsw+0x1a0>
DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11
@@ -922,11 +922,11 @@ _sk_saturation_hsw LABEL PROC
DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10
DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10
DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12
- DB 196,98,125,24,53,84,60,0,0 ; vbroadcastss 0x3c54(%rip),%ymm14 # 4ac0 <_sk_callback_hsw+0x1a4>
- DB 196,98,125,24,61,79,60,0,0 ; vbroadcastss 0x3c4f(%rip),%ymm15 # 4ac4 <_sk_callback_hsw+0x1a8>
+ DB 196,98,125,24,53,44,62,0,0 ; vbroadcastss 0x3e2c(%rip),%ymm14 # 4c98 <_sk_callback_hsw+0x1a4>
+ DB 196,98,125,24,61,39,62,0,0 ; vbroadcastss 0x3e27(%rip),%ymm15 # 4c9c <_sk_callback_hsw+0x1a8>
DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13
DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13
- DB 196,226,125,24,5,64,60,0,0 ; vbroadcastss 0x3c40(%rip),%ymm0 # 4ac8 <_sk_callback_hsw+0x1ac>
+ DB 196,226,125,24,5,24,62,0,0 ; vbroadcastss 0x3e18(%rip),%ymm0 # 4ca0 <_sk_callback_hsw+0x1ac>
DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13
DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10
DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10
@@ -981,7 +981,7 @@ _sk_saturation_hsw LABEL PROC
DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0
DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9
DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8
- DB 196,226,125,24,13,45,59,0,0 ; vbroadcastss 0x3b2d(%rip),%ymm1 # 4acc <_sk_callback_hsw+0x1b0>
+ DB 196,226,125,24,13,5,61,0,0 ; vbroadcastss 0x3d05(%rip),%ymm1 # 4ca4 <_sk_callback_hsw+0x1b0>
DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11
@@ -1009,11 +1009,11 @@ _sk_color_hsw LABEL PROC
DB 197,108,89,199 ; vmulps %ymm7,%ymm2,%ymm8
DB 197,116,89,215 ; vmulps %ymm7,%ymm1,%ymm10
DB 197,52,89,223 ; vmulps %ymm7,%ymm9,%ymm11
- DB 196,98,125,24,45,192,58,0,0 ; vbroadcastss 0x3ac0(%rip),%ymm13 # 4ad0 <_sk_callback_hsw+0x1b4>
- DB 196,98,125,24,53,187,58,0,0 ; vbroadcastss 0x3abb(%rip),%ymm14 # 4ad4 <_sk_callback_hsw+0x1b8>
+ DB 196,98,125,24,45,152,60,0,0 ; vbroadcastss 0x3c98(%rip),%ymm13 # 4ca8 <_sk_callback_hsw+0x1b4>
+ DB 196,98,125,24,53,147,60,0,0 ; vbroadcastss 0x3c93(%rip),%ymm14 # 4cac <_sk_callback_hsw+0x1b8>
DB 196,65,84,89,230 ; vmulps %ymm14,%ymm5,%ymm12
DB 196,66,93,184,229 ; vfmadd231ps %ymm13,%ymm4,%ymm12
- DB 196,98,125,24,61,172,58,0,0 ; vbroadcastss 0x3aac(%rip),%ymm15 # 4ad8 <_sk_callback_hsw+0x1bc>
+ DB 196,98,125,24,61,132,60,0,0 ; vbroadcastss 0x3c84(%rip),%ymm15 # 4cb0 <_sk_callback_hsw+0x1bc>
DB 196,66,77,184,231 ; vfmadd231ps %ymm15,%ymm6,%ymm12
DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9
DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9
@@ -1069,7 +1069,7 @@ _sk_color_hsw LABEL PROC
DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1
DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8
DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9
- DB 196,226,125,24,5,142,57,0,0 ; vbroadcastss 0x398e(%rip),%ymm0 # 4adc <_sk_callback_hsw+0x1c0>
+ DB 196,226,125,24,5,102,59,0,0 ; vbroadcastss 0x3b66(%rip),%ymm0 # 4cb4 <_sk_callback_hsw+0x1c0>
DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11
@@ -1097,11 +1097,11 @@ _sk_luminosity_hsw LABEL PROC
DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8
DB 197,100,89,213 ; vmulps %ymm5,%ymm3,%ymm10
DB 197,100,89,222 ; vmulps %ymm6,%ymm3,%ymm11
- DB 196,98,125,24,45,33,57,0,0 ; vbroadcastss 0x3921(%rip),%ymm13 # 4ae0 <_sk_callback_hsw+0x1c4>
- DB 196,98,125,24,53,28,57,0,0 ; vbroadcastss 0x391c(%rip),%ymm14 # 4ae4 <_sk_callback_hsw+0x1c8>
+ DB 196,98,125,24,45,249,58,0,0 ; vbroadcastss 0x3af9(%rip),%ymm13 # 4cb8 <_sk_callback_hsw+0x1c4>
+ DB 196,98,125,24,53,244,58,0,0 ; vbroadcastss 0x3af4(%rip),%ymm14 # 4cbc <_sk_callback_hsw+0x1c8>
DB 196,65,116,89,230 ; vmulps %ymm14,%ymm1,%ymm12
DB 196,66,109,184,229 ; vfmadd231ps %ymm13,%ymm2,%ymm12
- DB 196,98,125,24,61,13,57,0,0 ; vbroadcastss 0x390d(%rip),%ymm15 # 4ae8 <_sk_callback_hsw+0x1cc>
+ DB 196,98,125,24,61,229,58,0,0 ; vbroadcastss 0x3ae5(%rip),%ymm15 # 4cc0 <_sk_callback_hsw+0x1cc>
DB 196,66,53,184,231 ; vfmadd231ps %ymm15,%ymm9,%ymm12
DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9
DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9
@@ -1157,7 +1157,7 @@ _sk_luminosity_hsw LABEL PROC
DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1
DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8
DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9
- DB 196,226,125,24,5,239,55,0,0 ; vbroadcastss 0x37ef(%rip),%ymm0 # 4aec <_sk_callback_hsw+0x1d0>
+ DB 196,226,125,24,5,199,57,0,0 ; vbroadcastss 0x39c7(%rip),%ymm0 # 4cc4 <_sk_callback_hsw+0x1d0>
DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11
@@ -1184,17 +1184,17 @@ _sk_srcover_rgba_8888_hsw LABEL PROC
DB 77,133,192 ; test %r8,%r8
DB 15,133,180,0,0,0 ; jne 140f <_sk_srcover_rgba_8888_hsw+0xcd>
DB 196,193,124,16,58 ; vmovups (%r10),%ymm7
- DB 197,196,84,37,248,58,0,0 ; vandps 0x3af8(%rip),%ymm7,%ymm4 # 4e60 <_sk_callback_hsw+0x544>
+ DB 197,196,84,37,216,60,0,0 ; vandps 0x3cd8(%rip),%ymm7,%ymm4 # 5040 <_sk_callback_hsw+0x54c>
DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4
- DB 196,226,69,0,45,11,59,0,0 ; vpshufb 0x3b0b(%rip),%ymm7,%ymm5 # 4e80 <_sk_callback_hsw+0x564>
+ DB 196,226,69,0,45,235,60,0,0 ; vpshufb 0x3ceb(%rip),%ymm7,%ymm5 # 5060 <_sk_callback_hsw+0x56c>
DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5
- DB 196,226,69,0,53,30,59,0,0 ; vpshufb 0x3b1e(%rip),%ymm7,%ymm6 # 4ea0 <_sk_callback_hsw+0x584>
+ DB 196,226,69,0,53,254,60,0,0 ; vpshufb 0x3cfe(%rip),%ymm7,%ymm6 # 5080 <_sk_callback_hsw+0x58c>
DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6
DB 197,197,114,215,24 ; vpsrld $0x18,%ymm7,%ymm7
DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7
- DB 196,98,125,24,5,88,55,0,0 ; vbroadcastss 0x3758(%rip),%ymm8 # 4af0 <_sk_callback_hsw+0x1d4>
+ DB 196,98,125,24,5,48,57,0,0 ; vbroadcastss 0x3930(%rip),%ymm8 # 4cc8 <_sk_callback_hsw+0x1d4>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
- DB 196,98,125,24,13,79,55,0,0 ; vbroadcastss 0x374f(%rip),%ymm9 # 4af4 <_sk_callback_hsw+0x1d8>
+ DB 196,98,125,24,13,39,57,0,0 ; vbroadcastss 0x3927(%rip),%ymm9 # 4ccc <_sk_callback_hsw+0x1d8>
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0
DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1
@@ -1250,7 +1250,7 @@ _sk_clamp_0_hsw LABEL PROC
PUBLIC _sk_clamp_1_hsw
_sk_clamp_1_hsw LABEL PROC
- DB 196,98,125,24,5,116,54,0,0 ; vbroadcastss 0x3674(%rip),%ymm8 # 4af8 <_sk_callback_hsw+0x1dc>
+ DB 196,98,125,24,5,76,56,0,0 ; vbroadcastss 0x384c(%rip),%ymm8 # 4cd0 <_sk_callback_hsw+0x1dc>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
@@ -1260,7 +1260,7 @@ _sk_clamp_1_hsw LABEL PROC
PUBLIC _sk_clamp_a_hsw
_sk_clamp_a_hsw LABEL PROC
- DB 196,98,125,24,5,87,54,0,0 ; vbroadcastss 0x3657(%rip),%ymm8 # 4afc <_sk_callback_hsw+0x1e0>
+ DB 196,98,125,24,5,47,56,0,0 ; vbroadcastss 0x382f(%rip),%ymm8 # 4cd4 <_sk_callback_hsw+0x1e0>
DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3
DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0
DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1
@@ -1332,7 +1332,7 @@ PUBLIC _sk_unpremul_hsw
_sk_unpremul_hsw LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9
- DB 196,98,125,24,21,159,53,0,0 ; vbroadcastss 0x359f(%rip),%ymm10 # 4b00 <_sk_callback_hsw+0x1e4>
+ DB 196,98,125,24,21,119,55,0,0 ; vbroadcastss 0x3777(%rip),%ymm10 # 4cd8 <_sk_callback_hsw+0x1e4>
DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10
DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
@@ -1343,16 +1343,16 @@ _sk_unpremul_hsw LABEL PROC
PUBLIC _sk_from_srgb_hsw
_sk_from_srgb_hsw LABEL PROC
- DB 196,98,125,24,5,128,53,0,0 ; vbroadcastss 0x3580(%rip),%ymm8 # 4b04 <_sk_callback_hsw+0x1e8>
+ DB 196,98,125,24,5,88,55,0,0 ; vbroadcastss 0x3758(%rip),%ymm8 # 4cdc <_sk_callback_hsw+0x1e8>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10
- DB 196,98,125,24,29,114,53,0,0 ; vbroadcastss 0x3572(%rip),%ymm11 # 4b08 <_sk_callback_hsw+0x1ec>
- DB 196,98,125,24,37,109,53,0,0 ; vbroadcastss 0x356d(%rip),%ymm12 # 4b0c <_sk_callback_hsw+0x1f0>
+ DB 196,98,125,24,29,74,55,0,0 ; vbroadcastss 0x374a(%rip),%ymm11 # 4ce0 <_sk_callback_hsw+0x1ec>
+ DB 196,98,125,24,37,69,55,0,0 ; vbroadcastss 0x3745(%rip),%ymm12 # 4ce4 <_sk_callback_hsw+0x1f0>
DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13
DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13
- DB 196,98,125,24,53,94,53,0,0 ; vbroadcastss 0x355e(%rip),%ymm14 # 4b10 <_sk_callback_hsw+0x1f4>
+ DB 196,98,125,24,53,54,55,0,0 ; vbroadcastss 0x3736(%rip),%ymm14 # 4ce8 <_sk_callback_hsw+0x1f4>
DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13
- DB 196,98,125,24,21,84,53,0,0 ; vbroadcastss 0x3554(%rip),%ymm10 # 4b14 <_sk_callback_hsw+0x1f8>
+ DB 196,98,125,24,21,44,55,0,0 ; vbroadcastss 0x372c(%rip),%ymm10 # 4cec <_sk_callback_hsw+0x1f8>
DB 196,193,124,194,194,1 ; vcmpltps %ymm10,%ymm0,%ymm0
DB 196,195,21,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm13,%ymm0
DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9
@@ -1374,19 +1374,19 @@ _sk_from_srgb_hsw LABEL PROC
PUBLIC _sk_to_srgb_hsw
_sk_to_srgb_hsw LABEL PROC
DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9
- DB 196,98,125,24,5,248,52,0,0 ; vbroadcastss 0x34f8(%rip),%ymm8 # 4b18 <_sk_callback_hsw+0x1fc>
+ DB 196,98,125,24,5,208,54,0,0 ; vbroadcastss 0x36d0(%rip),%ymm8 # 4cf0 <_sk_callback_hsw+0x1fc>
DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10
- DB 196,98,125,24,29,238,52,0,0 ; vbroadcastss 0x34ee(%rip),%ymm11 # 4b1c <_sk_callback_hsw+0x200>
- DB 196,98,125,24,37,233,52,0,0 ; vbroadcastss 0x34e9(%rip),%ymm12 # 4b20 <_sk_callback_hsw+0x204>
+ DB 196,98,125,24,29,198,54,0,0 ; vbroadcastss 0x36c6(%rip),%ymm11 # 4cf4 <_sk_callback_hsw+0x200>
+ DB 196,98,125,24,37,193,54,0,0 ; vbroadcastss 0x36c1(%rip),%ymm12 # 4cf8 <_sk_callback_hsw+0x204>
DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13
DB 196,66,53,168,235 ; vfmadd213ps %ymm11,%ymm9,%ymm13
- DB 196,98,125,24,53,218,52,0,0 ; vbroadcastss 0x34da(%rip),%ymm14 # 4b24 <_sk_callback_hsw+0x208>
+ DB 196,98,125,24,53,178,54,0,0 ; vbroadcastss 0x36b2(%rip),%ymm14 # 4cfc <_sk_callback_hsw+0x208>
DB 196,66,53,168,238 ; vfmadd213ps %ymm14,%ymm9,%ymm13
- DB 196,98,125,24,61,208,52,0,0 ; vbroadcastss 0x34d0(%rip),%ymm15 # 4b28 <_sk_callback_hsw+0x20c>
+ DB 196,98,125,24,61,168,54,0,0 ; vbroadcastss 0x36a8(%rip),%ymm15 # 4d00 <_sk_callback_hsw+0x20c>
DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9
DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9
DB 196,65,20,89,201 ; vmulps %ymm9,%ymm13,%ymm9
- DB 196,98,125,24,45,188,52,0,0 ; vbroadcastss 0x34bc(%rip),%ymm13 # 4b2c <_sk_callback_hsw+0x210>
+ DB 196,98,125,24,45,148,54,0,0 ; vbroadcastss 0x3694(%rip),%ymm13 # 4d04 <_sk_callback_hsw+0x210>
DB 196,193,124,194,197,1 ; vcmpltps %ymm13,%ymm0,%ymm0
DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0
DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9
@@ -1418,26 +1418,26 @@ _sk_rgb_to_hsl_hsw LABEL PROC
DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9
DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9
DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,29,49,52,0,0 ; vbroadcastss 0x3431(%rip),%ymm11 # 4b30 <_sk_callback_hsw+0x214>
+ DB 196,98,125,24,29,9,54,0,0 ; vbroadcastss 0x3609(%rip),%ymm11 # 4d08 <_sk_callback_hsw+0x214>
DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11
DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12
DB 197,116,194,234,1 ; vcmpltps %ymm2,%ymm1,%ymm13
- DB 196,98,125,24,53,30,52,0,0 ; vbroadcastss 0x341e(%rip),%ymm14 # 4b34 <_sk_callback_hsw+0x218>
+ DB 196,98,125,24,53,246,53,0,0 ; vbroadcastss 0x35f6(%rip),%ymm14 # 4d0c <_sk_callback_hsw+0x218>
DB 196,65,4,87,255 ; vxorps %ymm15,%ymm15,%ymm15
DB 196,67,5,74,238,208 ; vblendvps %ymm13,%ymm14,%ymm15,%ymm13
DB 196,66,37,168,229 ; vfmadd213ps %ymm13,%ymm11,%ymm12
DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2
DB 197,124,92,233 ; vsubps %ymm1,%ymm0,%ymm13
- DB 196,98,125,24,53,5,52,0,0 ; vbroadcastss 0x3405(%rip),%ymm14 # 4b3c <_sk_callback_hsw+0x220>
+ DB 196,98,125,24,53,221,53,0,0 ; vbroadcastss 0x35dd(%rip),%ymm14 # 4d14 <_sk_callback_hsw+0x220>
DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13
- DB 196,98,125,24,53,243,51,0,0 ; vbroadcastss 0x33f3(%rip),%ymm14 # 4b38 <_sk_callback_hsw+0x21c>
+ DB 196,98,125,24,53,203,53,0,0 ; vbroadcastss 0x35cb(%rip),%ymm14 # 4d10 <_sk_callback_hsw+0x21c>
DB 196,194,37,168,214 ; vfmadd213ps %ymm14,%ymm11,%ymm2
DB 197,188,194,201,0 ; vcmpeqps %ymm1,%ymm8,%ymm1
DB 196,227,21,74,202,16 ; vblendvps %ymm1,%ymm2,%ymm13,%ymm1
DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0
DB 196,195,117,74,196,0 ; vblendvps %ymm0,%ymm12,%ymm1,%ymm0
DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1
- DB 196,98,125,24,29,214,51,0,0 ; vbroadcastss 0x33d6(%rip),%ymm11 # 4b44 <_sk_callback_hsw+0x228>
+ DB 196,98,125,24,29,174,53,0,0 ; vbroadcastss 0x35ae(%rip),%ymm11 # 4d1c <_sk_callback_hsw+0x228>
DB 196,193,116,89,211 ; vmulps %ymm11,%ymm1,%ymm2
DB 197,36,194,218,1 ; vcmpltps %ymm2,%ymm11,%ymm11
DB 196,65,12,92,224 ; vsubps %ymm8,%ymm14,%ymm12
@@ -1447,7 +1447,7 @@ _sk_rgb_to_hsl_hsw LABEL PROC
DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1
DB 196,195,125,74,199,128 ; vblendvps %ymm8,%ymm15,%ymm0,%ymm0
DB 196,195,117,74,207,128 ; vblendvps %ymm8,%ymm15,%ymm1,%ymm1
- DB 196,98,125,24,5,153,51,0,0 ; vbroadcastss 0x3399(%rip),%ymm8 # 4b40 <_sk_callback_hsw+0x224>
+ DB 196,98,125,24,5,113,53,0,0 ; vbroadcastss 0x3571(%rip),%ymm8 # 4d18 <_sk_callback_hsw+0x224>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1462,30 +1462,30 @@ _sk_hsl_to_rgb_hsw LABEL PROC
DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp)
DB 197,252,40,233 ; vmovaps %ymm1,%ymm5
DB 197,252,40,224 ; vmovaps %ymm0,%ymm4
- DB 196,98,125,24,5,96,51,0,0 ; vbroadcastss 0x3360(%rip),%ymm8 # 4b48 <_sk_callback_hsw+0x22c>
+ DB 196,98,125,24,5,56,53,0,0 ; vbroadcastss 0x3538(%rip),%ymm8 # 4d20 <_sk_callback_hsw+0x22c>
DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9
DB 197,84,89,210 ; vmulps %ymm2,%ymm5,%ymm10
DB 196,65,84,92,218 ; vsubps %ymm10,%ymm5,%ymm11
DB 196,67,45,74,203,144 ; vblendvps %ymm9,%ymm11,%ymm10,%ymm9
DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10
- DB 196,98,125,24,13,67,51,0,0 ; vbroadcastss 0x3343(%rip),%ymm9 # 4b4c <_sk_callback_hsw+0x230>
+ DB 196,98,125,24,13,27,53,0,0 ; vbroadcastss 0x351b(%rip),%ymm9 # 4d24 <_sk_callback_hsw+0x230>
DB 196,66,109,170,202 ; vfmsub213ps %ymm10,%ymm2,%ymm9
- DB 196,98,125,24,29,57,51,0,0 ; vbroadcastss 0x3339(%rip),%ymm11 # 4b50 <_sk_callback_hsw+0x234>
+ DB 196,98,125,24,29,17,53,0,0 ; vbroadcastss 0x3511(%rip),%ymm11 # 4d28 <_sk_callback_hsw+0x234>
DB 196,65,92,88,219 ; vaddps %ymm11,%ymm4,%ymm11
DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12
DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15
DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11
- DB 196,98,125,24,45,35,51,0,0 ; vbroadcastss 0x3323(%rip),%ymm13 # 4b58 <_sk_callback_hsw+0x23c>
+ DB 196,98,125,24,45,251,52,0,0 ; vbroadcastss 0x34fb(%rip),%ymm13 # 4d30 <_sk_callback_hsw+0x23c>
DB 196,193,4,89,197 ; vmulps %ymm13,%ymm15,%ymm0
- DB 196,98,125,24,53,25,51,0,0 ; vbroadcastss 0x3319(%rip),%ymm14 # 4b5c <_sk_callback_hsw+0x240>
+ DB 196,98,125,24,53,241,52,0,0 ; vbroadcastss 0x34f1(%rip),%ymm14 # 4d34 <_sk_callback_hsw+0x240>
DB 197,12,92,224 ; vsubps %ymm0,%ymm14,%ymm12
DB 196,66,37,168,225 ; vfmadd213ps %ymm9,%ymm11,%ymm12
- DB 196,226,125,24,29,255,50,0,0 ; vbroadcastss 0x32ff(%rip),%ymm3 # 4b54 <_sk_callback_hsw+0x238>
+ DB 196,226,125,24,29,215,52,0,0 ; vbroadcastss 0x34d7(%rip),%ymm3 # 4d2c <_sk_callback_hsw+0x238>
DB 196,193,100,194,255,2 ; vcmpleps %ymm15,%ymm3,%ymm7
DB 196,195,29,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm12,%ymm7
DB 196,65,60,194,231,2 ; vcmpleps %ymm15,%ymm8,%ymm12
DB 196,227,45,74,255,192 ; vblendvps %ymm12,%ymm7,%ymm10,%ymm7
- DB 196,98,125,24,37,234,50,0,0 ; vbroadcastss 0x32ea(%rip),%ymm12 # 4b60 <_sk_callback_hsw+0x244>
+ DB 196,98,125,24,37,194,52,0,0 ; vbroadcastss 0x34c2(%rip),%ymm12 # 4d38 <_sk_callback_hsw+0x244>
DB 196,65,28,194,255,2 ; vcmpleps %ymm15,%ymm12,%ymm15
DB 196,194,37,168,193 ; vfmadd213ps %ymm9,%ymm11,%ymm0
DB 196,99,125,74,255,240 ; vblendvps %ymm15,%ymm7,%ymm0,%ymm15
@@ -1501,7 +1501,7 @@ _sk_hsl_to_rgb_hsw LABEL PROC
DB 197,156,194,192,2 ; vcmpleps %ymm0,%ymm12,%ymm0
DB 196,194,37,168,249 ; vfmadd213ps %ymm9,%ymm11,%ymm7
DB 196,227,69,74,201,0 ; vblendvps %ymm0,%ymm1,%ymm7,%ymm1
- DB 196,226,125,24,5,150,50,0,0 ; vbroadcastss 0x3296(%rip),%ymm0 # 4b64 <_sk_callback_hsw+0x248>
+ DB 196,226,125,24,5,110,52,0,0 ; vbroadcastss 0x346e(%rip),%ymm0 # 4d3c <_sk_callback_hsw+0x248>
DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0
DB 196,227,125,8,224,1 ; vroundps $0x1,%ymm0,%ymm4
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
@@ -1542,38 +1542,69 @@ _sk_scale_1_float_hsw LABEL PROC
PUBLIC _sk_scale_u8_hsw
_sk_scale_u8_hsw LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,51 ; jne 19b4 <_sk_scale_u8_hsw+0x43>
- DB 197,122,126,0 ; vmovq (%rax),%xmm8
- DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8
+ DB 117,58 ; jne 19b5 <_sk_scale_u8_hsw+0x44>
+ DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8
+ DB 197,57,219,5,151,56,0,0 ; vpand 0x3897(%rip),%xmm8,%xmm8 # 5220 <_sk_callback_hsw+0x72c>
+ DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,208,49,0,0 ; vbroadcastss 0x31d0(%rip),%ymm9 # 4b68 <_sk_callback_hsw+0x24c>
+ DB 196,98,125,24,13,164,51,0,0 ; vbroadcastss 0x33a4(%rip),%ymm9 # 4d40 <_sk_callback_hsw+0x24c>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 76,137,201 ; mov %r9,%rcx
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 19bd <_sk_scale_u8_hsw+0x4c>
- DB 196,65,249,110,194 ; vmovq %r10,%xmm8
- DB 91 ; pop %rbx
- DB 235,171 ; jmp 1985 <_sk_scale_u8_hsw+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,183 ; ja 1981 <_sk_scale_u8_hsw+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,95,0,0,0 ; lea 0x5f(%rip),%r10 # 1a34 <_sk_scale_u8_hsw+0xc3>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 197,57,196,192,6 ; vpinsrw $0x6,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,57,196,192,5 ; vpinsrw $0x5,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,57,196,192,4 ; vpinsrw $0x4,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,57,196,192,3 ; vpinsrw $0x3,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,57,196,192,2 ; vpinsrw $0x2,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,57,196,192,1 ; vpinsrw $0x1,%eax,%xmm8,%xmm8
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,57,196,192,0 ; vpinsrw $0x0,%eax,%xmm8,%xmm8
+ DB 233,77,255,255,255 ; jmpq 1981 <_sk_scale_u8_hsw+0x10>
+ DB 241 ; icebp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,230 ; jmpq *%rsi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 219,255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,208 ; callq *%rax
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,197 ; inc %ebp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 186,255,255,255,170 ; mov $0xaaffffff,%edx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_lerp_1_float_hsw
_sk_lerp_1_float_hsw LABEL PROC
@@ -1592,16 +1623,15 @@ _sk_lerp_1_float_hsw LABEL PROC
PUBLIC _sk_lerp_u8_hsw
_sk_lerp_u8_hsw LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,71 ; jne 1a60 <_sk_lerp_u8_hsw+0x57>
- DB 197,122,126,0 ; vmovq (%rax),%xmm8
- DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8
+ DB 117,78 ; jne 1ad7 <_sk_lerp_u8_hsw+0x58>
+ DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8
+ DB 197,57,219,5,153,55,0,0 ; vpand 0x3799(%rip),%xmm8,%xmm8 # 5230 <_sk_callback_hsw+0x73c>
+ DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,60,49,0,0 ; vbroadcastss 0x313c(%rip),%ymm9 # 4b6c <_sk_callback_hsw+0x250>
+ DB 196,98,125,24,13,154,50,0,0 ; vbroadcastss 0x329a(%rip),%ymm9 # 4d44 <_sk_callback_hsw+0x250>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0
@@ -1612,45 +1642,78 @@ _sk_lerp_u8_hsw LABEL PROC
DB 197,228,92,223 ; vsubps %ymm7,%ymm3,%ymm3
DB 196,226,61,168,223 ; vfmadd213ps %ymm7,%ymm8,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 76,137,201 ; mov %r9,%rcx
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 1a69 <_sk_lerp_u8_hsw+0x60>
- DB 196,65,249,110,194 ; vmovq %r10,%xmm8
- DB 91 ; pop %rbx
- DB 235,151 ; jmp 1a1d <_sk_lerp_u8_hsw+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,163 ; ja 1a8f <_sk_lerp_u8_hsw+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,97,0,0,0 ; lea 0x61(%rip),%r10 # 1b58 <_sk_lerp_u8_hsw+0xd9>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 197,57,196,192,6 ; vpinsrw $0x6,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,57,196,192,5 ; vpinsrw $0x5,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,57,196,192,4 ; vpinsrw $0x4,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,57,196,192,3 ; vpinsrw $0x3,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,57,196,192,2 ; vpinsrw $0x2,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,57,196,192,1 ; vpinsrw $0x1,%eax,%xmm8,%xmm8
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,57,196,192,0 ; vpinsrw $0x0,%eax,%xmm8,%xmm8
+ DB 233,57,255,255,255 ; jmpq 1a8f <_sk_lerp_u8_hsw+0x10>
+ DB 102,144 ; xchg %ax,%ax
+ DB 239 ; out %eax,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 217,255 ; fcos
+ DB 255 ; (bad)
+ DB 255,206 ; dec %esi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,195 ; inc %ebx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 184,255,255,255,168 ; mov $0xa8ffffff,%eax
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_lerp_565_hsw
_sk_lerp_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 15,133,169,0,0,0 ; jne 1b3d <_sk_lerp_565_hsw+0xb7>
+ DB 15,133,169,0,0,0 ; jne 1c2b <_sk_lerp_565_hsw+0xb7>
DB 196,65,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm8
DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8
- DB 196,98,125,88,13,200,48,0,0 ; vpbroadcastd 0x30c8(%rip),%ymm9 # 4b70 <_sk_callback_hsw+0x254>
+ DB 196,98,125,88,13,178,49,0,0 ; vpbroadcastd 0x31b2(%rip),%ymm9 # 4d48 <_sk_callback_hsw+0x254>
DB 196,65,61,219,201 ; vpand %ymm9,%ymm8,%ymm9
DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9
- DB 196,98,125,24,21,185,48,0,0 ; vbroadcastss 0x30b9(%rip),%ymm10 # 4b74 <_sk_callback_hsw+0x258>
+ DB 196,98,125,24,21,163,49,0,0 ; vbroadcastss 0x31a3(%rip),%ymm10 # 4d4c <_sk_callback_hsw+0x258>
DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
- DB 196,98,125,88,21,175,48,0,0 ; vpbroadcastd 0x30af(%rip),%ymm10 # 4b78 <_sk_callback_hsw+0x25c>
+ DB 196,98,125,88,21,153,49,0,0 ; vpbroadcastd 0x3199(%rip),%ymm10 # 4d50 <_sk_callback_hsw+0x25c>
DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10
DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10
- DB 196,98,125,24,29,160,48,0,0 ; vbroadcastss 0x30a0(%rip),%ymm11 # 4b7c <_sk_callback_hsw+0x260>
+ DB 196,98,125,24,29,138,49,0,0 ; vbroadcastss 0x318a(%rip),%ymm11 # 4d54 <_sk_callback_hsw+0x260>
DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10
- DB 196,98,125,88,29,150,48,0,0 ; vpbroadcastd 0x3096(%rip),%ymm11 # 4b80 <_sk_callback_hsw+0x264>
+ DB 196,98,125,88,29,128,49,0,0 ; vpbroadcastd 0x3180(%rip),%ymm11 # 4d58 <_sk_callback_hsw+0x264>
DB 196,65,61,219,195 ; vpand %ymm11,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,29,135,48,0,0 ; vbroadcastss 0x3087(%rip),%ymm11 # 4b84 <_sk_callback_hsw+0x268>
+ DB 196,98,125,24,29,113,49,0,0 ; vbroadcastss 0x3171(%rip),%ymm11 # 4d5c <_sk_callback_hsw+0x268>
DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
DB 196,226,53,168,196 ; vfmadd213ps %ymm4,%ymm9,%ymm0
@@ -1671,9 +1734,9 @@ _sk_lerp_565_hsw LABEL PROC
DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 15,135,68,255,255,255 ; ja 1a9a <_sk_lerp_565_hsw+0x14>
+ DB 15,135,68,255,255,255 ; ja 1b88 <_sk_lerp_565_hsw+0x14>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 1bac <_sk_lerp_565_hsw+0x126>
+ DB 76,141,21,77,0,0,0 ; lea 0x4d(%rip),%r10 # 1c9c <_sk_lerp_565_hsw+0x128>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -1685,27 +1748,26 @@ _sk_lerp_565_hsw LABEL PROC
DB 196,65,57,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm8,%xmm8
DB 196,65,57,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm8,%xmm8
DB 196,65,57,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm8,%xmm8
- DB 233,239,254,255,255 ; jmpq 1a9a <_sk_lerp_565_hsw+0x14>
- DB 144 ; nop
- DB 243,255 ; repz (bad)
+ DB 233,239,254,255,255 ; jmpq 1b88 <_sk_lerp_565_hsw+0x14>
+ DB 15,31,0 ; nopl (%rax)
+ DB 241 ; icebp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 1bb1 <_sk_lerp_565_hsw+0x12b>
DB 255 ; (bad)
- DB 255,227 ; jmpq *%rbx
+ DB 233,255,255,255,225 ; jmpq ffffffffe2001ca4 <_sk_callback_hsw+0xffffffffe1ffd1b0>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 219,255 ; (bad)
+ DB 217,255 ; fcos
DB 255 ; (bad)
- DB 255,211 ; callq *%rbx
+ DB 255,209 ; callq *%rcx
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,203 ; dec %ebx
+ DB 255,201 ; dec %ecx
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 190 ; .byte 0xbe
+ DB 188 ; .byte 0xbc
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -1717,23 +1779,23 @@ _sk_load_tables_hsw LABEL PROC
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 76,3,16 ; add (%rax),%r10
DB 77,133,192 ; test %r8,%r8
- DB 117,105 ; jne 1c46 <_sk_load_tables_hsw+0x7e>
+ DB 117,105 ; jne 1d36 <_sk_load_tables_hsw+0x7e>
DB 196,193,124,16,26 ; vmovups (%r10),%ymm3
- DB 197,228,84,13,214,50,0,0 ; vandps 0x32d6(%rip),%ymm3,%ymm1 # 4ec0 <_sk_callback_hsw+0x5a4>
+ DB 197,228,84,13,198,51,0,0 ; vandps 0x33c6(%rip),%ymm3,%ymm1 # 50a0 <_sk_callback_hsw+0x5ac>
DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8
DB 72,139,72,8 ; mov 0x8(%rax),%rcx
DB 76,139,80,16 ; mov 0x10(%rax),%r10
DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2
DB 196,226,109,146,4,137 ; vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0
- DB 196,226,101,0,21,214,50,0,0 ; vpshufb 0x32d6(%rip),%ymm3,%ymm2 # 4ee0 <_sk_callback_hsw+0x5c4>
+ DB 196,226,101,0,21,198,51,0,0 ; vpshufb 0x33c6(%rip),%ymm3,%ymm2 # 50c0 <_sk_callback_hsw+0x5cc>
DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9
DB 196,194,53,146,12,146 ; vgatherdps %ymm9,(%r10,%ymm2,4),%ymm1
DB 72,139,64,24 ; mov 0x18(%rax),%rax
- DB 196,98,101,0,13,222,50,0,0 ; vpshufb 0x32de(%rip),%ymm3,%ymm9 # 4f00 <_sk_callback_hsw+0x5e4>
+ DB 196,98,101,0,13,206,51,0,0 ; vpshufb 0x33ce(%rip),%ymm3,%ymm9 # 50e0 <_sk_callback_hsw+0x5ec>
DB 196,162,61,146,20,136 ; vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2
DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,78,47,0,0 ; vbroadcastss 0x2f4e(%rip),%ymm8 # 4b88 <_sk_callback_hsw+0x26c>
+ DB 196,98,125,24,5,54,48,0,0 ; vbroadcastss 0x3036(%rip),%ymm8 # 4d60 <_sk_callback_hsw+0x26c>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,201 ; mov %r9,%rcx
@@ -1746,7 +1808,7 @@ _sk_load_tables_hsw LABEL PROC
DB 196,193,249,110,195 ; vmovq %r11,%xmm0
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
DB 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3
- DB 233,115,255,255,255 ; jmpq 1be2 <_sk_load_tables_hsw+0x1a>
+ DB 233,115,255,255,255 ; jmpq 1cd2 <_sk_load_tables_hsw+0x1a>
PUBLIC _sk_load_tables_u16_be_hsw
_sk_load_tables_u16_be_hsw LABEL PROC
@@ -1754,7 +1816,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 77,133,192 ; test %r8,%r8
- DB 15,133,201,0,0,0 ; jne 1d4e <_sk_load_tables_u16_be_hsw+0xdf>
+ DB 15,133,201,0,0,0 ; jne 1e3e <_sk_load_tables_u16_be_hsw+0xdf>
DB 196,1,121,16,4,81 ; vmovupd (%r9,%r10,2),%xmm8
DB 196,129,121,16,84,81,16 ; vmovupd 0x10(%r9,%r10,2),%xmm2
DB 196,129,121,16,92,81,32 ; vmovupd 0x20(%r9,%r10,2),%xmm3
@@ -1770,7 +1832,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 197,185,108,200 ; vpunpcklqdq %xmm0,%xmm8,%xmm1
DB 197,185,109,208 ; vpunpckhqdq %xmm0,%xmm8,%xmm2
DB 197,49,108,195 ; vpunpcklqdq %xmm3,%xmm9,%xmm8
- DB 197,121,111,21,106,51,0,0 ; vmovdqa 0x336a(%rip),%xmm10 # 5040 <_sk_callback_hsw+0x724>
+ DB 197,121,111,21,122,52,0,0 ; vmovdqa 0x347a(%rip),%xmm10 # 5240 <_sk_callback_hsw+0x74c>
DB 196,193,113,219,194 ; vpand %xmm10,%xmm1,%xmm0
DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1
DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11
@@ -1792,36 +1854,36 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3
DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,71,46,0,0 ; vbroadcastss 0x2e47(%rip),%ymm8 # 4b8c <_sk_callback_hsw+0x270>
+ DB 196,98,125,24,5,47,47,0,0 ; vbroadcastss 0x2f2f(%rip),%ymm8 # 4d64 <_sk_callback_hsw+0x270>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
DB 196,1,123,16,4,81 ; vmovsd (%r9,%r10,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,85 ; je 1db4 <_sk_load_tables_u16_be_hsw+0x145>
+ DB 116,85 ; je 1ea4 <_sk_load_tables_u16_be_hsw+0x145>
DB 196,1,57,22,68,81,8 ; vmovhpd 0x8(%r9,%r10,2),%xmm8,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,72 ; jb 1db4 <_sk_load_tables_u16_be_hsw+0x145>
+ DB 114,72 ; jb 1ea4 <_sk_load_tables_u16_be_hsw+0x145>
DB 196,129,123,16,84,81,16 ; vmovsd 0x10(%r9,%r10,2),%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 116,72 ; je 1dc1 <_sk_load_tables_u16_be_hsw+0x152>
+ DB 116,72 ; je 1eb1 <_sk_load_tables_u16_be_hsw+0x152>
DB 196,129,105,22,84,81,24 ; vmovhpd 0x18(%r9,%r10,2),%xmm2,%xmm2
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,59 ; jb 1dc1 <_sk_load_tables_u16_be_hsw+0x152>
+ DB 114,59 ; jb 1eb1 <_sk_load_tables_u16_be_hsw+0x152>
DB 196,129,123,16,92,81,32 ; vmovsd 0x20(%r9,%r10,2),%xmm3
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 15,132,9,255,255,255 ; je 1ca0 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,132,9,255,255,255 ; je 1d90 <_sk_load_tables_u16_be_hsw+0x31>
DB 196,129,97,22,92,81,40 ; vmovhpd 0x28(%r9,%r10,2),%xmm3,%xmm3
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 15,130,248,254,255,255 ; jb 1ca0 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,130,248,254,255,255 ; jb 1d90 <_sk_load_tables_u16_be_hsw+0x31>
DB 196,1,122,126,76,81,48 ; vmovq 0x30(%r9,%r10,2),%xmm9
- DB 233,236,254,255,255 ; jmpq 1ca0 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,236,254,255,255 ; jmpq 1d90 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,223,254,255,255 ; jmpq 1ca0 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,223,254,255,255 ; jmpq 1d90 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,214,254,255,255 ; jmpq 1ca0 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,214,254,255,255 ; jmpq 1d90 <_sk_load_tables_u16_be_hsw+0x31>
PUBLIC _sk_load_tables_rgb_u16_be_hsw
_sk_load_tables_rgb_u16_be_hsw LABEL PROC
@@ -1829,7 +1891,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 76,141,20,82 ; lea (%rdx,%rdx,2),%r10
DB 77,133,192 ; test %r8,%r8
- DB 15,133,193,0,0,0 ; jne 1e9d <_sk_load_tables_rgb_u16_be_hsw+0xd3>
+ DB 15,133,193,0,0,0 ; jne 1f8d <_sk_load_tables_rgb_u16_be_hsw+0xd3>
DB 196,129,122,111,4,81 ; vmovdqu (%r9,%r10,2),%xmm0
DB 196,129,122,111,84,81,12 ; vmovdqu 0xc(%r9,%r10,2),%xmm2
DB 196,129,122,111,76,81,24 ; vmovdqu 0x18(%r9,%r10,2),%xmm1
@@ -1850,7 +1912,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3
DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2
DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8
- DB 197,121,111,13,10,50,0,0 ; vmovdqa 0x320a(%rip),%xmm9 # 5050 <_sk_callback_hsw+0x734>
+ DB 197,121,111,13,26,51,0,0 ; vmovdqa 0x331a(%rip),%xmm9 # 5250 <_sk_callback_hsw+0x75c>
DB 196,193,97,219,193 ; vpand %xmm9,%xmm3,%xmm0
DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1
DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3
@@ -1867,46 +1929,46 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8
DB 196,162,101,146,20,128 ; vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,245,44,0,0 ; vbroadcastss 0x2cf5(%rip),%ymm3 # 4b90 <_sk_callback_hsw+0x274>
+ DB 196,226,125,24,29,221,45,0,0 ; vbroadcastss 0x2ddd(%rip),%ymm3 # 4d68 <_sk_callback_hsw+0x274>
DB 255,224 ; jmpq *%rax
DB 196,129,121,110,4,81 ; vmovd (%r9,%r10,2),%xmm0
DB 196,129,121,196,68,81,4,2 ; vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 117,5 ; jne 1eb6 <_sk_load_tables_rgb_u16_be_hsw+0xec>
- DB 233,90,255,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 1fa6 <_sk_load_tables_rgb_u16_be_hsw+0xec>
+ DB 233,90,255,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,76,81,6 ; vmovd 0x6(%r9,%r10,2),%xmm1
DB 196,1,113,196,68,81,10,2 ; vpinsrw $0x2,0xa(%r9,%r10,2),%xmm1,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,26 ; jb 1ee5 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
+ DB 114,26 ; jb 1fd5 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
DB 196,129,121,110,76,81,12 ; vmovd 0xc(%r9,%r10,2),%xmm1
DB 196,129,113,196,84,81,16,2 ; vpinsrw $0x2,0x10(%r9,%r10,2),%xmm1,%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 117,10 ; jne 1eea <_sk_load_tables_rgb_u16_be_hsw+0x120>
- DB 233,43,255,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,38,255,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 1fda <_sk_load_tables_rgb_u16_be_hsw+0x120>
+ DB 233,43,255,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,38,255,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,76,81,18 ; vmovd 0x12(%r9,%r10,2),%xmm1
DB 196,1,113,196,76,81,22,2 ; vpinsrw $0x2,0x16(%r9,%r10,2),%xmm1,%xmm9
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,26 ; jb 1f19 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
+ DB 114,26 ; jb 2009 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
DB 196,129,121,110,76,81,24 ; vmovd 0x18(%r9,%r10,2),%xmm1
DB 196,129,113,196,76,81,28,2 ; vpinsrw $0x2,0x1c(%r9,%r10,2),%xmm1,%xmm1
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 117,10 ; jne 1f1e <_sk_load_tables_rgb_u16_be_hsw+0x154>
- DB 233,247,254,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,242,254,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 200e <_sk_load_tables_rgb_u16_be_hsw+0x154>
+ DB 233,247,254,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,242,254,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,92,81,30 ; vmovd 0x1e(%r9,%r10,2),%xmm3
DB 196,1,97,196,92,81,34,2 ; vpinsrw $0x2,0x22(%r9,%r10,2),%xmm3,%xmm11
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,20 ; jb 1f47 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
+ DB 114,20 ; jb 2037 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
DB 196,129,121,110,92,81,36 ; vmovd 0x24(%r9,%r10,2),%xmm3
DB 196,129,97,196,92,81,40,2 ; vpinsrw $0x2,0x28(%r9,%r10,2),%xmm3,%xmm3
- DB 233,201,254,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,196,254,255,255 ; jmpq 1e10 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,201,254,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,196,254,255,255 ; jmpq 1f00 <_sk_load_tables_rgb_u16_be_hsw+0x46>
PUBLIC _sk_byte_tables_hsw
_sk_byte_tables_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,61,44,0,0 ; vbroadcastss 0x2c3d(%rip),%ymm8 # 4b94 <_sk_callback_hsw+0x278>
+ DB 196,98,125,24,5,37,45,0,0 ; vbroadcastss 0x2d25(%rip),%ymm8 # 4d6c <_sk_callback_hsw+0x278>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9
DB 196,65,249,126,201 ; vmovq %xmm9,%r9
@@ -2028,7 +2090,7 @@ _sk_byte_tables_hsw LABEL PROC
DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax
DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,226,41,0,0 ; vbroadcastss 0x29e2(%rip),%ymm8 # 4b98 <_sk_callback_hsw+0x27c>
+ DB 196,98,125,24,5,202,42,0,0 ; vbroadcastss 0x2aca(%rip),%ymm8 # 4d70 <_sk_callback_hsw+0x27c>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
@@ -2142,7 +2204,7 @@ _sk_byte_tables_rgb_hsw LABEL PROC
DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax
DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,206,39,0,0 ; vbroadcastss 0x27ce(%rip),%ymm8 # 4b9c <_sk_callback_hsw+0x280>
+ DB 196,98,125,24,5,182,40,0,0 ; vbroadcastss 0x28b6(%rip),%ymm8 # 4d74 <_sk_callback_hsw+0x280>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
@@ -2231,33 +2293,33 @@ _sk_parametric_r_hsw LABEL PROC
DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10
DB 196,226,125,24,0 ; vbroadcastss (%rax),%ymm0
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,168,38,0,0 ; vbroadcastss 0x26a8(%rip),%ymm12 # 4ba0 <_sk_callback_hsw+0x284>
- DB 196,98,125,24,45,163,38,0,0 ; vbroadcastss 0x26a3(%rip),%ymm13 # 4ba4 <_sk_callback_hsw+0x288>
+ DB 196,98,125,24,37,144,39,0,0 ; vbroadcastss 0x2790(%rip),%ymm12 # 4d78 <_sk_callback_hsw+0x284>
+ DB 196,98,125,24,45,139,39,0,0 ; vbroadcastss 0x278b(%rip),%ymm13 # 4d7c <_sk_callback_hsw+0x288>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,153,38,0,0 ; vbroadcastss 0x2699(%rip),%ymm13 # 4ba8 <_sk_callback_hsw+0x28c>
+ DB 196,98,125,24,45,129,39,0,0 ; vbroadcastss 0x2781(%rip),%ymm13 # 4d80 <_sk_callback_hsw+0x28c>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,143,38,0,0 ; vbroadcastss 0x268f(%rip),%ymm13 # 4bac <_sk_callback_hsw+0x290>
+ DB 196,98,125,24,45,119,39,0,0 ; vbroadcastss 0x2777(%rip),%ymm13 # 4d84 <_sk_callback_hsw+0x290>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,133,38,0,0 ; vbroadcastss 0x2685(%rip),%ymm11 # 4bb0 <_sk_callback_hsw+0x294>
+ DB 196,98,125,24,29,109,39,0,0 ; vbroadcastss 0x276d(%rip),%ymm11 # 4d88 <_sk_callback_hsw+0x294>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,123,38,0,0 ; vbroadcastss 0x267b(%rip),%ymm12 # 4bb4 <_sk_callback_hsw+0x298>
+ DB 196,98,125,24,37,99,39,0,0 ; vbroadcastss 0x2763(%rip),%ymm12 # 4d8c <_sk_callback_hsw+0x298>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,113,38,0,0 ; vbroadcastss 0x2671(%rip),%ymm12 # 4bb8 <_sk_callback_hsw+0x29c>
+ DB 196,98,125,24,37,89,39,0,0 ; vbroadcastss 0x2759(%rip),%ymm12 # 4d90 <_sk_callback_hsw+0x29c>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0
DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10
DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10
- DB 196,98,125,24,29,82,38,0,0 ; vbroadcastss 0x2652(%rip),%ymm11 # 4bbc <_sk_callback_hsw+0x2a0>
+ DB 196,98,125,24,29,58,39,0,0 ; vbroadcastss 0x273a(%rip),%ymm11 # 4d94 <_sk_callback_hsw+0x2a0>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,72,38,0,0 ; vbroadcastss 0x2648(%rip),%ymm11 # 4bc0 <_sk_callback_hsw+0x2a4>
+ DB 196,98,125,24,29,48,39,0,0 ; vbroadcastss 0x2730(%rip),%ymm11 # 4d98 <_sk_callback_hsw+0x2a4>
DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11
- DB 196,226,125,24,5,62,38,0,0 ; vbroadcastss 0x263e(%rip),%ymm0 # 4bc4 <_sk_callback_hsw+0x2a8>
+ DB 196,226,125,24,5,38,39,0,0 ; vbroadcastss 0x2726(%rip),%ymm0 # 4d9c <_sk_callback_hsw+0x2a8>
DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,52,38,0,0 ; vbroadcastss 0x2634(%rip),%ymm10 # 4bc8 <_sk_callback_hsw+0x2ac>
+ DB 196,98,125,24,21,28,39,0,0 ; vbroadcastss 0x271c(%rip),%ymm10 # 4da0 <_sk_callback_hsw+0x2ac>
DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0
DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0
- DB 196,98,125,24,21,39,38,0,0 ; vbroadcastss 0x2627(%rip),%ymm10 # 4bcc <_sk_callback_hsw+0x2b0>
+ DB 196,98,125,24,21,15,39,0,0 ; vbroadcastss 0x270f(%rip),%ymm10 # 4da4 <_sk_callback_hsw+0x2b0>
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
@@ -2265,7 +2327,7 @@ _sk_parametric_r_hsw LABEL PROC
DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,254,37,0,0 ; vbroadcastss 0x25fe(%rip),%ymm8 # 4bd0 <_sk_callback_hsw+0x2b4>
+ DB 196,98,125,24,5,230,38,0,0 ; vbroadcastss 0x26e6(%rip),%ymm8 # 4da8 <_sk_callback_hsw+0x2b4>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2283,33 +2345,33 @@ _sk_parametric_g_hsw LABEL PROC
DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10
DB 196,226,125,24,8 ; vbroadcastss (%rax),%ymm1
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,182,37,0,0 ; vbroadcastss 0x25b6(%rip),%ymm12 # 4bd4 <_sk_callback_hsw+0x2b8>
- DB 196,98,125,24,45,177,37,0,0 ; vbroadcastss 0x25b1(%rip),%ymm13 # 4bd8 <_sk_callback_hsw+0x2bc>
+ DB 196,98,125,24,37,158,38,0,0 ; vbroadcastss 0x269e(%rip),%ymm12 # 4dac <_sk_callback_hsw+0x2b8>
+ DB 196,98,125,24,45,153,38,0,0 ; vbroadcastss 0x2699(%rip),%ymm13 # 4db0 <_sk_callback_hsw+0x2bc>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,167,37,0,0 ; vbroadcastss 0x25a7(%rip),%ymm13 # 4bdc <_sk_callback_hsw+0x2c0>
+ DB 196,98,125,24,45,143,38,0,0 ; vbroadcastss 0x268f(%rip),%ymm13 # 4db4 <_sk_callback_hsw+0x2c0>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,157,37,0,0 ; vbroadcastss 0x259d(%rip),%ymm13 # 4be0 <_sk_callback_hsw+0x2c4>
+ DB 196,98,125,24,45,133,38,0,0 ; vbroadcastss 0x2685(%rip),%ymm13 # 4db8 <_sk_callback_hsw+0x2c4>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,147,37,0,0 ; vbroadcastss 0x2593(%rip),%ymm11 # 4be4 <_sk_callback_hsw+0x2c8>
+ DB 196,98,125,24,29,123,38,0,0 ; vbroadcastss 0x267b(%rip),%ymm11 # 4dbc <_sk_callback_hsw+0x2c8>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,137,37,0,0 ; vbroadcastss 0x2589(%rip),%ymm12 # 4be8 <_sk_callback_hsw+0x2cc>
+ DB 196,98,125,24,37,113,38,0,0 ; vbroadcastss 0x2671(%rip),%ymm12 # 4dc0 <_sk_callback_hsw+0x2cc>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,127,37,0,0 ; vbroadcastss 0x257f(%rip),%ymm12 # 4bec <_sk_callback_hsw+0x2d0>
+ DB 196,98,125,24,37,103,38,0,0 ; vbroadcastss 0x2667(%rip),%ymm12 # 4dc4 <_sk_callback_hsw+0x2d0>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1
DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10
DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
- DB 196,98,125,24,29,96,37,0,0 ; vbroadcastss 0x2560(%rip),%ymm11 # 4bf0 <_sk_callback_hsw+0x2d4>
+ DB 196,98,125,24,29,72,38,0,0 ; vbroadcastss 0x2648(%rip),%ymm11 # 4dc8 <_sk_callback_hsw+0x2d4>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,86,37,0,0 ; vbroadcastss 0x2556(%rip),%ymm11 # 4bf4 <_sk_callback_hsw+0x2d8>
+ DB 196,98,125,24,29,62,38,0,0 ; vbroadcastss 0x263e(%rip),%ymm11 # 4dcc <_sk_callback_hsw+0x2d8>
DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11
- DB 196,226,125,24,13,76,37,0,0 ; vbroadcastss 0x254c(%rip),%ymm1 # 4bf8 <_sk_callback_hsw+0x2dc>
+ DB 196,226,125,24,13,52,38,0,0 ; vbroadcastss 0x2634(%rip),%ymm1 # 4dd0 <_sk_callback_hsw+0x2dc>
DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,66,37,0,0 ; vbroadcastss 0x2542(%rip),%ymm10 # 4bfc <_sk_callback_hsw+0x2e0>
+ DB 196,98,125,24,21,42,38,0,0 ; vbroadcastss 0x262a(%rip),%ymm10 # 4dd4 <_sk_callback_hsw+0x2e0>
DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1
DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1
- DB 196,98,125,24,21,53,37,0,0 ; vbroadcastss 0x2535(%rip),%ymm10 # 4c00 <_sk_callback_hsw+0x2e4>
+ DB 196,98,125,24,21,29,38,0,0 ; vbroadcastss 0x261d(%rip),%ymm10 # 4dd8 <_sk_callback_hsw+0x2e4>
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
@@ -2317,7 +2379,7 @@ _sk_parametric_g_hsw LABEL PROC
DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1
- DB 196,98,125,24,5,12,37,0,0 ; vbroadcastss 0x250c(%rip),%ymm8 # 4c04 <_sk_callback_hsw+0x2e8>
+ DB 196,98,125,24,5,244,37,0,0 ; vbroadcastss 0x25f4(%rip),%ymm8 # 4ddc <_sk_callback_hsw+0x2e8>
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2335,33 +2397,33 @@ _sk_parametric_b_hsw LABEL PROC
DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10
DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,196,36,0,0 ; vbroadcastss 0x24c4(%rip),%ymm12 # 4c08 <_sk_callback_hsw+0x2ec>
- DB 196,98,125,24,45,191,36,0,0 ; vbroadcastss 0x24bf(%rip),%ymm13 # 4c0c <_sk_callback_hsw+0x2f0>
+ DB 196,98,125,24,37,172,37,0,0 ; vbroadcastss 0x25ac(%rip),%ymm12 # 4de0 <_sk_callback_hsw+0x2ec>
+ DB 196,98,125,24,45,167,37,0,0 ; vbroadcastss 0x25a7(%rip),%ymm13 # 4de4 <_sk_callback_hsw+0x2f0>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,181,36,0,0 ; vbroadcastss 0x24b5(%rip),%ymm13 # 4c10 <_sk_callback_hsw+0x2f4>
+ DB 196,98,125,24,45,157,37,0,0 ; vbroadcastss 0x259d(%rip),%ymm13 # 4de8 <_sk_callback_hsw+0x2f4>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,171,36,0,0 ; vbroadcastss 0x24ab(%rip),%ymm13 # 4c14 <_sk_callback_hsw+0x2f8>
+ DB 196,98,125,24,45,147,37,0,0 ; vbroadcastss 0x2593(%rip),%ymm13 # 4dec <_sk_callback_hsw+0x2f8>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,161,36,0,0 ; vbroadcastss 0x24a1(%rip),%ymm11 # 4c18 <_sk_callback_hsw+0x2fc>
+ DB 196,98,125,24,29,137,37,0,0 ; vbroadcastss 0x2589(%rip),%ymm11 # 4df0 <_sk_callback_hsw+0x2fc>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,151,36,0,0 ; vbroadcastss 0x2497(%rip),%ymm12 # 4c1c <_sk_callback_hsw+0x300>
+ DB 196,98,125,24,37,127,37,0,0 ; vbroadcastss 0x257f(%rip),%ymm12 # 4df4 <_sk_callback_hsw+0x300>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,141,36,0,0 ; vbroadcastss 0x248d(%rip),%ymm12 # 4c20 <_sk_callback_hsw+0x304>
+ DB 196,98,125,24,37,117,37,0,0 ; vbroadcastss 0x2575(%rip),%ymm12 # 4df8 <_sk_callback_hsw+0x304>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2
DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10
DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10
- DB 196,98,125,24,29,110,36,0,0 ; vbroadcastss 0x246e(%rip),%ymm11 # 4c24 <_sk_callback_hsw+0x308>
+ DB 196,98,125,24,29,86,37,0,0 ; vbroadcastss 0x2556(%rip),%ymm11 # 4dfc <_sk_callback_hsw+0x308>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,100,36,0,0 ; vbroadcastss 0x2464(%rip),%ymm11 # 4c28 <_sk_callback_hsw+0x30c>
+ DB 196,98,125,24,29,76,37,0,0 ; vbroadcastss 0x254c(%rip),%ymm11 # 4e00 <_sk_callback_hsw+0x30c>
DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11
- DB 196,226,125,24,21,90,36,0,0 ; vbroadcastss 0x245a(%rip),%ymm2 # 4c2c <_sk_callback_hsw+0x310>
+ DB 196,226,125,24,21,66,37,0,0 ; vbroadcastss 0x2542(%rip),%ymm2 # 4e04 <_sk_callback_hsw+0x310>
DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,80,36,0,0 ; vbroadcastss 0x2450(%rip),%ymm10 # 4c30 <_sk_callback_hsw+0x314>
+ DB 196,98,125,24,21,56,37,0,0 ; vbroadcastss 0x2538(%rip),%ymm10 # 4e08 <_sk_callback_hsw+0x314>
DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2
DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2
- DB 196,98,125,24,21,67,36,0,0 ; vbroadcastss 0x2443(%rip),%ymm10 # 4c34 <_sk_callback_hsw+0x318>
+ DB 196,98,125,24,21,43,37,0,0 ; vbroadcastss 0x252b(%rip),%ymm10 # 4e0c <_sk_callback_hsw+0x318>
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
@@ -2369,7 +2431,7 @@ _sk_parametric_b_hsw LABEL PROC
DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,26,36,0,0 ; vbroadcastss 0x241a(%rip),%ymm8 # 4c38 <_sk_callback_hsw+0x31c>
+ DB 196,98,125,24,5,2,37,0,0 ; vbroadcastss 0x2502(%rip),%ymm8 # 4e10 <_sk_callback_hsw+0x31c>
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2387,33 +2449,33 @@ _sk_parametric_a_hsw LABEL PROC
DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10
DB 196,226,125,24,24 ; vbroadcastss (%rax),%ymm3
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,210,35,0,0 ; vbroadcastss 0x23d2(%rip),%ymm12 # 4c3c <_sk_callback_hsw+0x320>
- DB 196,98,125,24,45,205,35,0,0 ; vbroadcastss 0x23cd(%rip),%ymm13 # 4c40 <_sk_callback_hsw+0x324>
+ DB 196,98,125,24,37,186,36,0,0 ; vbroadcastss 0x24ba(%rip),%ymm12 # 4e14 <_sk_callback_hsw+0x320>
+ DB 196,98,125,24,45,181,36,0,0 ; vbroadcastss 0x24b5(%rip),%ymm13 # 4e18 <_sk_callback_hsw+0x324>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,195,35,0,0 ; vbroadcastss 0x23c3(%rip),%ymm13 # 4c44 <_sk_callback_hsw+0x328>
+ DB 196,98,125,24,45,171,36,0,0 ; vbroadcastss 0x24ab(%rip),%ymm13 # 4e1c <_sk_callback_hsw+0x328>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,185,35,0,0 ; vbroadcastss 0x23b9(%rip),%ymm13 # 4c48 <_sk_callback_hsw+0x32c>
+ DB 196,98,125,24,45,161,36,0,0 ; vbroadcastss 0x24a1(%rip),%ymm13 # 4e20 <_sk_callback_hsw+0x32c>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,175,35,0,0 ; vbroadcastss 0x23af(%rip),%ymm11 # 4c4c <_sk_callback_hsw+0x330>
+ DB 196,98,125,24,29,151,36,0,0 ; vbroadcastss 0x2497(%rip),%ymm11 # 4e24 <_sk_callback_hsw+0x330>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,165,35,0,0 ; vbroadcastss 0x23a5(%rip),%ymm12 # 4c50 <_sk_callback_hsw+0x334>
+ DB 196,98,125,24,37,141,36,0,0 ; vbroadcastss 0x248d(%rip),%ymm12 # 4e28 <_sk_callback_hsw+0x334>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,155,35,0,0 ; vbroadcastss 0x239b(%rip),%ymm12 # 4c54 <_sk_callback_hsw+0x338>
+ DB 196,98,125,24,37,131,36,0,0 ; vbroadcastss 0x2483(%rip),%ymm12 # 4e2c <_sk_callback_hsw+0x338>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10
DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10
- DB 196,98,125,24,29,124,35,0,0 ; vbroadcastss 0x237c(%rip),%ymm11 # 4c58 <_sk_callback_hsw+0x33c>
+ DB 196,98,125,24,29,100,36,0,0 ; vbroadcastss 0x2464(%rip),%ymm11 # 4e30 <_sk_callback_hsw+0x33c>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,114,35,0,0 ; vbroadcastss 0x2372(%rip),%ymm11 # 4c5c <_sk_callback_hsw+0x340>
+ DB 196,98,125,24,29,90,36,0,0 ; vbroadcastss 0x245a(%rip),%ymm11 # 4e34 <_sk_callback_hsw+0x340>
DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11
- DB 196,226,125,24,29,104,35,0,0 ; vbroadcastss 0x2368(%rip),%ymm3 # 4c60 <_sk_callback_hsw+0x344>
+ DB 196,226,125,24,29,80,36,0,0 ; vbroadcastss 0x2450(%rip),%ymm3 # 4e38 <_sk_callback_hsw+0x344>
DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,94,35,0,0 ; vbroadcastss 0x235e(%rip),%ymm10 # 4c64 <_sk_callback_hsw+0x348>
+ DB 196,98,125,24,21,70,36,0,0 ; vbroadcastss 0x2446(%rip),%ymm10 # 4e3c <_sk_callback_hsw+0x348>
DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3
DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3
- DB 196,98,125,24,21,81,35,0,0 ; vbroadcastss 0x2351(%rip),%ymm10 # 4c68 <_sk_callback_hsw+0x34c>
+ DB 196,98,125,24,21,57,36,0,0 ; vbroadcastss 0x2439(%rip),%ymm10 # 4e40 <_sk_callback_hsw+0x34c>
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
@@ -2421,33 +2483,33 @@ _sk_parametric_a_hsw LABEL PROC
DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3
- DB 196,98,125,24,5,40,35,0,0 ; vbroadcastss 0x2328(%rip),%ymm8 # 4c6c <_sk_callback_hsw+0x350>
+ DB 196,98,125,24,5,16,36,0,0 ; vbroadcastss 0x2410(%rip),%ymm8 # 4e44 <_sk_callback_hsw+0x350>
DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
PUBLIC _sk_lab_to_xyz_hsw
_sk_lab_to_xyz_hsw LABEL PROC
- DB 196,98,125,24,5,26,35,0,0 ; vbroadcastss 0x231a(%rip),%ymm8 # 4c70 <_sk_callback_hsw+0x354>
- DB 196,98,125,24,13,21,35,0,0 ; vbroadcastss 0x2315(%rip),%ymm9 # 4c74 <_sk_callback_hsw+0x358>
- DB 196,98,125,24,21,16,35,0,0 ; vbroadcastss 0x2310(%rip),%ymm10 # 4c78 <_sk_callback_hsw+0x35c>
+ DB 196,98,125,24,5,2,36,0,0 ; vbroadcastss 0x2402(%rip),%ymm8 # 4e48 <_sk_callback_hsw+0x354>
+ DB 196,98,125,24,13,253,35,0,0 ; vbroadcastss 0x23fd(%rip),%ymm9 # 4e4c <_sk_callback_hsw+0x358>
+ DB 196,98,125,24,21,248,35,0,0 ; vbroadcastss 0x23f8(%rip),%ymm10 # 4e50 <_sk_callback_hsw+0x35c>
DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1
DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2
- DB 196,98,125,24,13,1,35,0,0 ; vbroadcastss 0x2301(%rip),%ymm9 # 4c7c <_sk_callback_hsw+0x360>
+ DB 196,98,125,24,13,233,35,0,0 ; vbroadcastss 0x23e9(%rip),%ymm9 # 4e54 <_sk_callback_hsw+0x360>
DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9
- DB 196,226,125,24,5,247,34,0,0 ; vbroadcastss 0x22f7(%rip),%ymm0 # 4c80 <_sk_callback_hsw+0x364>
+ DB 196,226,125,24,5,223,35,0,0 ; vbroadcastss 0x23df(%rip),%ymm0 # 4e58 <_sk_callback_hsw+0x364>
DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0
- DB 196,98,125,24,5,238,34,0,0 ; vbroadcastss 0x22ee(%rip),%ymm8 # 4c84 <_sk_callback_hsw+0x368>
+ DB 196,98,125,24,5,214,35,0,0 ; vbroadcastss 0x23d6(%rip),%ymm8 # 4e5c <_sk_callback_hsw+0x368>
DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8
- DB 196,98,125,24,13,228,34,0,0 ; vbroadcastss 0x22e4(%rip),%ymm9 # 4c88 <_sk_callback_hsw+0x36c>
+ DB 196,98,125,24,13,204,35,0,0 ; vbroadcastss 0x23cc(%rip),%ymm9 # 4e60 <_sk_callback_hsw+0x36c>
DB 196,98,109,172,200 ; vfnmadd213ps %ymm0,%ymm2,%ymm9
DB 196,193,60,89,200 ; vmulps %ymm8,%ymm8,%ymm1
DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
- DB 196,226,125,24,21,209,34,0,0 ; vbroadcastss 0x22d1(%rip),%ymm2 # 4c8c <_sk_callback_hsw+0x370>
+ DB 196,226,125,24,21,185,35,0,0 ; vbroadcastss 0x23b9(%rip),%ymm2 # 4e64 <_sk_callback_hsw+0x370>
DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10
- DB 196,98,125,24,29,199,34,0,0 ; vbroadcastss 0x22c7(%rip),%ymm11 # 4c90 <_sk_callback_hsw+0x374>
+ DB 196,98,125,24,29,175,35,0,0 ; vbroadcastss 0x23af(%rip),%ymm11 # 4e68 <_sk_callback_hsw+0x374>
DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8
- DB 196,98,125,24,37,189,34,0,0 ; vbroadcastss 0x22bd(%rip),%ymm12 # 4c94 <_sk_callback_hsw+0x378>
+ DB 196,98,125,24,37,165,35,0,0 ; vbroadcastss 0x23a5(%rip),%ymm12 # 4e6c <_sk_callback_hsw+0x378>
DB 196,65,60,89,196 ; vmulps %ymm12,%ymm8,%ymm8
DB 196,99,61,74,193,160 ; vblendvps %ymm10,%ymm1,%ymm8,%ymm8
DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1
@@ -2462,46 +2524,77 @@ _sk_lab_to_xyz_hsw LABEL PROC
DB 196,65,52,88,203 ; vaddps %ymm11,%ymm9,%ymm9
DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9
DB 196,227,53,74,208,32 ; vblendvps %ymm2,%ymm0,%ymm9,%ymm2
- DB 196,226,125,24,5,114,34,0,0 ; vbroadcastss 0x2272(%rip),%ymm0 # 4c98 <_sk_callback_hsw+0x37c>
+ DB 196,226,125,24,5,90,35,0,0 ; vbroadcastss 0x235a(%rip),%ymm0 # 4e70 <_sk_callback_hsw+0x37c>
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,105,34,0,0 ; vbroadcastss 0x2269(%rip),%ymm8 # 4c9c <_sk_callback_hsw+0x380>
+ DB 196,98,125,24,5,81,35,0,0 ; vbroadcastss 0x2351(%rip),%ymm8 # 4e74 <_sk_callback_hsw+0x380>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
PUBLIC _sk_load_a8_hsw
_sk_load_a8_hsw LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,45 ; jne 2a79 <_sk_load_a8_hsw+0x3d>
- DB 197,250,126,0 ; vmovq (%rax),%xmm0
- DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
+ DB 117,52 ; jne 2b6a <_sk_load_a8_hsw+0x3e>
+ DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0
+ DB 197,249,219,5,28,39,0,0 ; vpand 0x271c(%rip),%xmm0,%xmm0 # 5260 <_sk_callback_hsw+0x76c>
+ DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,62,34,0,0 ; vbroadcastss 0x223e(%rip),%ymm1 # 4ca0 <_sk_callback_hsw+0x384>
+ DB 196,226,125,24,13,34,35,0,0 ; vbroadcastss 0x2322(%rip),%ymm1 # 4e78 <_sk_callback_hsw+0x384>
DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2
- DB 76,137,201 ; mov %r9,%rcx
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 2a82 <_sk_load_a8_hsw+0x46>
- DB 196,193,249,110,194 ; vmovq %r10,%xmm0
- DB 91 ; pop %rbx
- DB 235,177 ; jmp 2a50 <_sk_load_a8_hsw+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,190 ; ja 2b3c <_sk_load_a8_hsw+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,95,0,0,0 ; lea 0x5f(%rip),%r10 # 2be8 <_sk_load_a8_hsw+0xbc>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,249,196,192,5 ; vpinsrw $0x5,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,249,196,192,4 ; vpinsrw $0x4,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,249,196,192,3 ; vpinsrw $0x3,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,249,196,192,1 ; vpinsrw $0x1,%eax,%xmm0,%xmm0
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,249,196,192,0 ; vpinsrw $0x0,%eax,%xmm0,%xmm0
+ DB 233,85,255,255,255 ; jmpq 2b3c <_sk_load_a8_hsw+0x10>
+ DB 144 ; nop
+ DB 240,255 ; lock (bad)
+ DB 255 ; (bad)
+ DB 255,229 ; jmpq *%rbp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 218,255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,207 ; dec %edi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,196 ; inc %esp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 185,255,255,255,170 ; mov $0xaaffffff,%ecx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_gather_a8_hsw
_sk_gather_a8_hsw LABEL PROC
@@ -2541,7 +2634,7 @@ _sk_gather_a8_hsw LABEL PROC
DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,79,33,0,0 ; vbroadcastss 0x214f(%rip),%ymm1 # 4ca4 <_sk_callback_hsw+0x388>
+ DB 196,226,125,24,13,194,33,0,0 ; vbroadcastss 0x21c2(%rip),%ymm1 # 4e7c <_sk_callback_hsw+0x388>
DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
@@ -2553,14 +2646,14 @@ PUBLIC _sk_store_a8_hsw
_sk_store_a8_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
- DB 196,98,125,24,5,49,33,0,0 ; vbroadcastss 0x2131(%rip),%ymm8 # 4ca8 <_sk_callback_hsw+0x38c>
+ DB 196,98,125,24,5,164,33,0,0 ; vbroadcastss 0x21a4(%rip),%ymm8 # 4e80 <_sk_callback_hsw+0x38c>
DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8
DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
DB 77,133,192 ; test %r8,%r8
- DB 117,10 ; jne 2ba0 <_sk_store_a8_hsw+0x37>
+ DB 117,10 ; jne 2d05 <_sk_store_a8_hsw+0x37>
DB 196,65,123,17,4,19 ; vmovsd %xmm8,(%r11,%rdx,1)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2568,10 +2661,10 @@ _sk_store_a8_hsw LABEL PROC
DB 65,128,225,7 ; and $0x7,%r9b
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,236 ; ja 2b9c <_sk_store_a8_hsw+0x33>
+ DB 119,236 ; ja 2d01 <_sk_store_a8_hsw+0x33>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 2c04 <_sk_store_a8_hsw+0x9b>
+ DB 76,141,21,67,0,0,0 ; lea 0x43(%rip),%r10 # 2d68 <_sk_store_a8_hsw+0x9a>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -2582,65 +2675,95 @@ _sk_store_a8_hsw LABEL PROC
DB 196,67,121,20,68,19,2,4 ; vpextrb $0x4,%xmm8,0x2(%r11,%rdx,1)
DB 196,67,121,20,68,19,1,2 ; vpextrb $0x2,%xmm8,0x1(%r11,%rdx,1)
DB 196,67,121,20,4,19,0 ; vpextrb $0x0,%xmm8,(%r11,%rdx,1)
- DB 235,154 ; jmp 2b9c <_sk_store_a8_hsw+0x33>
- DB 102,144 ; xchg %ax,%ax
- DB 245 ; cmc
+ DB 235,154 ; jmp 2d01 <_sk_store_a8_hsw+0x33>
+ DB 144 ; nop
+ DB 246,255 ; idiv %bh
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 238 ; out %al,(%dx)
DB 255 ; (bad)
- DB 237 ; in (%dx),%eax
DB 255 ; (bad)
+ DB 255,230 ; jmpq *%rsi
DB 255 ; (bad)
- DB 255,229 ; jmpq *%rbp
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 222,255 ; fdivrp %st,%st(7)
DB 255 ; (bad)
- DB 221,255 ; (bad)
+ DB 255,214 ; callq *%rsi
DB 255 ; (bad)
- DB 255,213 ; callq *%rbp
DB 255 ; (bad)
+ DB 255,206 ; dec %esi
DB 255 ; (bad)
- DB 255,205 ; dec %ebp
DB 255 ; (bad)
- DB 255 ; (bad)
- DB 255,197 ; inc %ebp
+ DB 255,198 ; inc %esi
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
PUBLIC _sk_load_g8_hsw
_sk_load_g8_hsw LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,50 ; jne 2c62 <_sk_load_g8_hsw+0x42>
- DB 197,250,126,0 ; vmovq (%rax),%xmm0
- DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
+ DB 117,57 ; jne 2dc7 <_sk_load_g8_hsw+0x43>
+ DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0
+ DB 197,249,219,5,212,36,0,0 ; vpand 0x24d4(%rip),%xmm0,%xmm0 # 5270 <_sk_callback_hsw+0x77c>
+ DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,102,32,0,0 ; vbroadcastss 0x2066(%rip),%ymm1 # 4cac <_sk_callback_hsw+0x390>
+ DB 196,226,125,24,13,214,32,0,0 ; vbroadcastss 0x20d6(%rip),%ymm1 # 4e84 <_sk_callback_hsw+0x390>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,91,32,0,0 ; vbroadcastss 0x205b(%rip),%ymm3 # 4cb0 <_sk_callback_hsw+0x394>
- DB 76,137,201 ; mov %r9,%rcx
+ DB 196,226,125,24,29,203,32,0,0 ; vbroadcastss 0x20cb(%rip),%ymm3 # 4e88 <_sk_callback_hsw+0x394>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 2c6b <_sk_load_g8_hsw+0x4b>
- DB 196,193,249,110,194 ; vmovq %r10,%xmm0
- DB 91 ; pop %rbx
- DB 235,172 ; jmp 2c34 <_sk_load_g8_hsw+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,185 ; ja 2d94 <_sk_load_g8_hsw+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,94,0,0,0 ; lea 0x5e(%rip),%r10 # 2e44 <_sk_load_g8_hsw+0xc0>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,249,196,192,5 ; vpinsrw $0x5,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,249,196,192,4 ; vpinsrw $0x4,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,249,196,192,3 ; vpinsrw $0x3,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,249,196,192,1 ; vpinsrw $0x1,%eax,%xmm0,%xmm0
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,249,196,192,0 ; vpinsrw $0x0,%eax,%xmm0,%xmm0
+ DB 233,80,255,255,255 ; jmpq 2d94 <_sk_load_g8_hsw+0x10>
+ DB 241 ; icebp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,230 ; jmpq *%rsi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 219,255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,208 ; callq *%rax
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,197 ; inc %ebp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 186,255,255,255,171 ; mov $0xabffffff,%edx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_gather_g8_hsw
_sk_gather_g8_hsw LABEL PROC
@@ -2680,10 +2803,10 @@ _sk_gather_g8_hsw LABEL PROC
DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,118,31,0,0 ; vbroadcastss 0x1f76(%rip),%ymm1 # 4cb4 <_sk_callback_hsw+0x398>
+ DB 196,226,125,24,13,118,31,0,0 ; vbroadcastss 0x1f76(%rip),%ymm1 # 4e8c <_sk_callback_hsw+0x398>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,107,31,0,0 ; vbroadcastss 0x1f6b(%rip),%ymm3 # 4cb8 <_sk_callback_hsw+0x39c>
+ DB 196,226,125,24,29,107,31,0,0 ; vbroadcastss 0x1f6b(%rip),%ymm3 # 4e90 <_sk_callback_hsw+0x39c>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 255,224 ; jmpq *%rax
@@ -2693,9 +2816,9 @@ _sk_gather_i8_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,137,193 ; mov %rax,%r9
DB 77,133,201 ; test %r9,%r9
- DB 116,5 ; je 2d66 <_sk_gather_i8_hsw+0xf>
+ DB 116,5 ; je 2f3e <_sk_gather_i8_hsw+0xf>
DB 76,137,200 ; mov %r9,%rax
- DB 235,2 ; jmp 2d68 <_sk_gather_i8_hsw+0x11>
+ DB 235,2 ; jmp 2f40 <_sk_gather_i8_hsw+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 83 ; push %rbx
DB 76,139,16 ; mov (%rax),%r10
@@ -2729,14 +2852,14 @@ _sk_gather_i8_hsw LABEL PROC
DB 73,139,65,8 ; mov 0x8(%r9),%rax
DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1
DB 196,226,117,144,28,128 ; vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3
- DB 197,229,219,5,19,33,0,0 ; vpand 0x2113(%rip),%ymm3,%ymm0 # 4f20 <_sk_callback_hsw+0x604>
+ DB 197,229,219,5,27,33,0,0 ; vpand 0x211b(%rip),%ymm3,%ymm0 # 5100 <_sk_callback_hsw+0x60c>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,162,30,0,0 ; vbroadcastss 0x1ea2(%rip),%ymm8 # 4cbc <_sk_callback_hsw+0x3a0>
+ DB 196,98,125,24,5,162,30,0,0 ; vbroadcastss 0x1ea2(%rip),%ymm8 # 4e94 <_sk_callback_hsw+0x3a0>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,24,33,0,0 ; vpshufb 0x2118(%rip),%ymm3,%ymm1 # 4f40 <_sk_callback_hsw+0x624>
+ DB 196,226,101,0,13,32,33,0,0 ; vpshufb 0x2120(%rip),%ymm3,%ymm1 # 5120 <_sk_callback_hsw+0x62c>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,38,33,0,0 ; vpshufb 0x2126(%rip),%ymm3,%ymm2 # 4f60 <_sk_callback_hsw+0x644>
+ DB 196,226,101,0,21,46,33,0,0 ; vpshufb 0x212e(%rip),%ymm3,%ymm2 # 5140 <_sk_callback_hsw+0x64c>
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3
@@ -2751,35 +2874,35 @@ _sk_load_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,114 ; jne 2ed2 <_sk_load_565_hsw+0x7c>
+ DB 117,114 ; jne 30aa <_sk_load_565_hsw+0x7c>
DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0
DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2
- DB 196,226,125,88,5,76,30,0,0 ; vpbroadcastd 0x1e4c(%rip),%ymm0 # 4cc0 <_sk_callback_hsw+0x3a4>
+ DB 196,226,125,88,5,76,30,0,0 ; vpbroadcastd 0x1e4c(%rip),%ymm0 # 4e98 <_sk_callback_hsw+0x3a4>
DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,63,30,0,0 ; vbroadcastss 0x1e3f(%rip),%ymm1 # 4cc4 <_sk_callback_hsw+0x3a8>
+ DB 196,226,125,24,13,63,30,0,0 ; vbroadcastss 0x1e3f(%rip),%ymm1 # 4e9c <_sk_callback_hsw+0x3a8>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,54,30,0,0 ; vpbroadcastd 0x1e36(%rip),%ymm1 # 4cc8 <_sk_callback_hsw+0x3ac>
+ DB 196,226,125,88,13,54,30,0,0 ; vpbroadcastd 0x1e36(%rip),%ymm1 # 4ea0 <_sk_callback_hsw+0x3ac>
DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,41,30,0,0 ; vbroadcastss 0x1e29(%rip),%ymm3 # 4ccc <_sk_callback_hsw+0x3b0>
+ DB 196,226,125,24,29,41,30,0,0 ; vbroadcastss 0x1e29(%rip),%ymm3 # 4ea4 <_sk_callback_hsw+0x3b0>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,88,29,32,30,0,0 ; vpbroadcastd 0x1e20(%rip),%ymm3 # 4cd0 <_sk_callback_hsw+0x3b4>
+ DB 196,226,125,88,29,32,30,0,0 ; vpbroadcastd 0x1e20(%rip),%ymm3 # 4ea8 <_sk_callback_hsw+0x3b4>
DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,19,30,0,0 ; vbroadcastss 0x1e13(%rip),%ymm3 # 4cd4 <_sk_callback_hsw+0x3b8>
+ DB 196,226,125,24,29,19,30,0,0 ; vbroadcastss 0x1e13(%rip),%ymm3 # 4eac <_sk_callback_hsw+0x3b8>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,8,30,0,0 ; vbroadcastss 0x1e08(%rip),%ymm3 # 4cd8 <_sk_callback_hsw+0x3bc>
+ DB 196,226,125,24,29,8,30,0,0 ; vbroadcastss 0x1e08(%rip),%ymm3 # 4eb0 <_sk_callback_hsw+0x3bc>
DB 255,224 ; jmpq *%rax
DB 69,137,193 ; mov %r8d,%r9d
DB 65,128,225,7 ; and $0x7,%r9b
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,128 ; ja 2e66 <_sk_load_565_hsw+0x10>
+ DB 119,128 ; ja 303e <_sk_load_565_hsw+0x10>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 2f3c <_sk_load_565_hsw+0xe6>
+ DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 3114 <_sk_load_565_hsw+0xe6>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -2791,7 +2914,7 @@ _sk_load_565_hsw LABEL PROC
DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- DB 233,44,255,255,255 ; jmpq 2e66 <_sk_load_565_hsw+0x10>
+ DB 233,44,255,255,255 ; jmpq 303e <_sk_load_565_hsw+0x10>
DB 102,144 ; xchg %ax,%ax
DB 242,255 ; repnz (bad)
DB 255 ; (bad)
@@ -2856,34 +2979,34 @@ _sk_gather_565_hsw LABEL PROC
DB 67,15,183,4,89 ; movzwl (%r9,%r11,2),%eax
DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0
DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2
- DB 196,226,125,88,5,211,28,0,0 ; vpbroadcastd 0x1cd3(%rip),%ymm0 # 4cdc <_sk_callback_hsw+0x3c0>
+ DB 196,226,125,88,5,211,28,0,0 ; vpbroadcastd 0x1cd3(%rip),%ymm0 # 4eb4 <_sk_callback_hsw+0x3c0>
DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,198,28,0,0 ; vbroadcastss 0x1cc6(%rip),%ymm1 # 4ce0 <_sk_callback_hsw+0x3c4>
+ DB 196,226,125,24,13,198,28,0,0 ; vbroadcastss 0x1cc6(%rip),%ymm1 # 4eb8 <_sk_callback_hsw+0x3c4>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,189,28,0,0 ; vpbroadcastd 0x1cbd(%rip),%ymm1 # 4ce4 <_sk_callback_hsw+0x3c8>
+ DB 196,226,125,88,13,189,28,0,0 ; vpbroadcastd 0x1cbd(%rip),%ymm1 # 4ebc <_sk_callback_hsw+0x3c8>
DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,176,28,0,0 ; vbroadcastss 0x1cb0(%rip),%ymm3 # 4ce8 <_sk_callback_hsw+0x3cc>
+ DB 196,226,125,24,29,176,28,0,0 ; vbroadcastss 0x1cb0(%rip),%ymm3 # 4ec0 <_sk_callback_hsw+0x3cc>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,88,29,167,28,0,0 ; vpbroadcastd 0x1ca7(%rip),%ymm3 # 4cec <_sk_callback_hsw+0x3d0>
+ DB 196,226,125,88,29,167,28,0,0 ; vpbroadcastd 0x1ca7(%rip),%ymm3 # 4ec4 <_sk_callback_hsw+0x3d0>
DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,154,28,0,0 ; vbroadcastss 0x1c9a(%rip),%ymm3 # 4cf0 <_sk_callback_hsw+0x3d4>
+ DB 196,226,125,24,29,154,28,0,0 ; vbroadcastss 0x1c9a(%rip),%ymm3 # 4ec8 <_sk_callback_hsw+0x3d4>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,143,28,0,0 ; vbroadcastss 0x1c8f(%rip),%ymm3 # 4cf4 <_sk_callback_hsw+0x3d8>
+ DB 196,226,125,24,29,143,28,0,0 ; vbroadcastss 0x1c8f(%rip),%ymm3 # 4ecc <_sk_callback_hsw+0x3d8>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_store_565_hsw
_sk_store_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
- DB 196,98,125,24,5,131,28,0,0 ; vbroadcastss 0x1c83(%rip),%ymm8 # 4cf8 <_sk_callback_hsw+0x3dc>
+ DB 196,98,125,24,5,131,28,0,0 ; vbroadcastss 0x1c83(%rip),%ymm8 # 4ed0 <_sk_callback_hsw+0x3dc>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,193,53,114,241,11 ; vpslld $0xb,%ymm9,%ymm9
- DB 196,98,125,24,21,110,28,0,0 ; vbroadcastss 0x1c6e(%rip),%ymm10 # 4cfc <_sk_callback_hsw+0x3e0>
+ DB 196,98,125,24,21,110,28,0,0 ; vbroadcastss 0x1c6e(%rip),%ymm10 # 4ed4 <_sk_callback_hsw+0x3e0>
DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10
DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10
DB 196,193,45,114,242,5 ; vpslld $0x5,%ymm10,%ymm10
@@ -2894,7 +3017,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,10 ; jne 30cc <_sk_store_565_hsw+0x65>
+ DB 117,10 ; jne 32a4 <_sk_store_565_hsw+0x65>
DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2902,9 +3025,9 @@ _sk_store_565_hsw LABEL PROC
DB 65,128,225,7 ; and $0x7,%r9b
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,236 ; ja 30c8 <_sk_store_565_hsw+0x61>
+ DB 119,236 ; ja 32a0 <_sk_store_565_hsw+0x61>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,69,0,0,0 ; lea 0x45(%rip),%r10 # 312c <_sk_store_565_hsw+0xc5>
+ DB 76,141,21,69,0,0,0 ; lea 0x45(%rip),%r10 # 3304 <_sk_store_565_hsw+0xc5>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -2915,7 +3038,7 @@ _sk_store_565_hsw LABEL PROC
DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- DB 235,159 ; jmp 30c8 <_sk_store_565_hsw+0x61>
+ DB 235,159 ; jmp 32a0 <_sk_store_565_hsw+0x61>
DB 15,31,0 ; nopl (%rax)
DB 244 ; hlt
DB 255 ; (bad)
@@ -2946,28 +3069,28 @@ _sk_load_4444_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 15,133,138,0,0,0 ; jne 31e0 <_sk_load_4444_hsw+0x98>
+ DB 15,133,138,0,0,0 ; jne 33b8 <_sk_load_4444_hsw+0x98>
DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0
DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3
- DB 196,226,125,88,5,150,27,0,0 ; vpbroadcastd 0x1b96(%rip),%ymm0 # 4d00 <_sk_callback_hsw+0x3e4>
+ DB 196,226,125,88,5,150,27,0,0 ; vpbroadcastd 0x1b96(%rip),%ymm0 # 4ed8 <_sk_callback_hsw+0x3e4>
DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,137,27,0,0 ; vbroadcastss 0x1b89(%rip),%ymm1 # 4d04 <_sk_callback_hsw+0x3e8>
+ DB 196,226,125,24,13,137,27,0,0 ; vbroadcastss 0x1b89(%rip),%ymm1 # 4edc <_sk_callback_hsw+0x3e8>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,128,27,0,0 ; vpbroadcastd 0x1b80(%rip),%ymm1 # 4d08 <_sk_callback_hsw+0x3ec>
+ DB 196,226,125,88,13,128,27,0,0 ; vpbroadcastd 0x1b80(%rip),%ymm1 # 4ee0 <_sk_callback_hsw+0x3ec>
DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,115,27,0,0 ; vbroadcastss 0x1b73(%rip),%ymm2 # 4d0c <_sk_callback_hsw+0x3f0>
+ DB 196,226,125,24,21,115,27,0,0 ; vbroadcastss 0x1b73(%rip),%ymm2 # 4ee4 <_sk_callback_hsw+0x3f0>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,88,21,106,27,0,0 ; vpbroadcastd 0x1b6a(%rip),%ymm2 # 4d10 <_sk_callback_hsw+0x3f4>
+ DB 196,226,125,88,21,106,27,0,0 ; vpbroadcastd 0x1b6a(%rip),%ymm2 # 4ee8 <_sk_callback_hsw+0x3f4>
DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,93,27,0,0 ; vbroadcastss 0x1b5d(%rip),%ymm8 # 4d14 <_sk_callback_hsw+0x3f8>
+ DB 196,98,125,24,5,93,27,0,0 ; vbroadcastss 0x1b5d(%rip),%ymm8 # 4eec <_sk_callback_hsw+0x3f8>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,88,5,83,27,0,0 ; vpbroadcastd 0x1b53(%rip),%ymm8 # 4d18 <_sk_callback_hsw+0x3fc>
+ DB 196,98,125,88,5,83,27,0,0 ; vpbroadcastd 0x1b53(%rip),%ymm8 # 4ef0 <_sk_callback_hsw+0x3fc>
DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,69,27,0,0 ; vbroadcastss 0x1b45(%rip),%ymm8 # 4d1c <_sk_callback_hsw+0x400>
+ DB 196,98,125,24,5,69,27,0,0 ; vbroadcastss 0x1b45(%rip),%ymm8 # 4ef4 <_sk_callback_hsw+0x400>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2976,9 +3099,9 @@ _sk_load_4444_hsw LABEL PROC
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 15,135,100,255,255,255 ; ja 315c <_sk_load_4444_hsw+0x14>
+ DB 15,135,100,255,255,255 ; ja 3334 <_sk_load_4444_hsw+0x14>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,73,0,0,0 ; lea 0x49(%rip),%r10 # 324c <_sk_load_4444_hsw+0x104>
+ DB 76,141,21,73,0,0,0 ; lea 0x49(%rip),%r10 # 3424 <_sk_load_4444_hsw+0x104>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -2990,7 +3113,7 @@ _sk_load_4444_hsw LABEL PROC
DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- DB 233,16,255,255,255 ; jmpq 315c <_sk_load_4444_hsw+0x14>
+ DB 233,16,255,255,255 ; jmpq 3334 <_sk_load_4444_hsw+0x14>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -3054,25 +3177,25 @@ _sk_gather_4444_hsw LABEL PROC
DB 67,15,183,4,89 ; movzwl (%r9,%r11,2),%eax
DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0
DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3
- DB 196,226,125,88,5,7,26,0,0 ; vpbroadcastd 0x1a07(%rip),%ymm0 # 4d20 <_sk_callback_hsw+0x404>
+ DB 196,226,125,88,5,7,26,0,0 ; vpbroadcastd 0x1a07(%rip),%ymm0 # 4ef8 <_sk_callback_hsw+0x404>
DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,250,25,0,0 ; vbroadcastss 0x19fa(%rip),%ymm1 # 4d24 <_sk_callback_hsw+0x408>
+ DB 196,226,125,24,13,250,25,0,0 ; vbroadcastss 0x19fa(%rip),%ymm1 # 4efc <_sk_callback_hsw+0x408>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,241,25,0,0 ; vpbroadcastd 0x19f1(%rip),%ymm1 # 4d28 <_sk_callback_hsw+0x40c>
+ DB 196,226,125,88,13,241,25,0,0 ; vpbroadcastd 0x19f1(%rip),%ymm1 # 4f00 <_sk_callback_hsw+0x40c>
DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,228,25,0,0 ; vbroadcastss 0x19e4(%rip),%ymm2 # 4d2c <_sk_callback_hsw+0x410>
+ DB 196,226,125,24,21,228,25,0,0 ; vbroadcastss 0x19e4(%rip),%ymm2 # 4f04 <_sk_callback_hsw+0x410>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,88,21,219,25,0,0 ; vpbroadcastd 0x19db(%rip),%ymm2 # 4d30 <_sk_callback_hsw+0x414>
+ DB 196,226,125,88,21,219,25,0,0 ; vpbroadcastd 0x19db(%rip),%ymm2 # 4f08 <_sk_callback_hsw+0x414>
DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,206,25,0,0 ; vbroadcastss 0x19ce(%rip),%ymm8 # 4d34 <_sk_callback_hsw+0x418>
+ DB 196,98,125,24,5,206,25,0,0 ; vbroadcastss 0x19ce(%rip),%ymm8 # 4f0c <_sk_callback_hsw+0x418>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,88,5,196,25,0,0 ; vpbroadcastd 0x19c4(%rip),%ymm8 # 4d38 <_sk_callback_hsw+0x41c>
+ DB 196,98,125,88,5,196,25,0,0 ; vpbroadcastd 0x19c4(%rip),%ymm8 # 4f10 <_sk_callback_hsw+0x41c>
DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,182,25,0,0 ; vbroadcastss 0x19b6(%rip),%ymm8 # 4d3c <_sk_callback_hsw+0x420>
+ DB 196,98,125,24,5,182,25,0,0 ; vbroadcastss 0x19b6(%rip),%ymm8 # 4f14 <_sk_callback_hsw+0x420>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -3081,7 +3204,7 @@ PUBLIC _sk_store_4444_hsw
_sk_store_4444_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
- DB 196,98,125,24,5,163,25,0,0 ; vbroadcastss 0x19a3(%rip),%ymm8 # 4d40 <_sk_callback_hsw+0x424>
+ DB 196,98,125,24,5,163,25,0,0 ; vbroadcastss 0x19a3(%rip),%ymm8 # 4f18 <_sk_callback_hsw+0x424>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,193,53,114,241,12 ; vpslld $0xc,%ymm9,%ymm9
@@ -3099,7 +3222,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,10 ; jne 3400 <_sk_store_4444_hsw+0x71>
+ DB 117,10 ; jne 35d8 <_sk_store_4444_hsw+0x71>
DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -3107,9 +3230,9 @@ _sk_store_4444_hsw LABEL PROC
DB 65,128,225,7 ; and $0x7,%r9b
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,236 ; ja 33fc <_sk_store_4444_hsw+0x6d>
+ DB 119,236 ; ja 35d4 <_sk_store_4444_hsw+0x6d>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,69,0,0,0 ; lea 0x45(%rip),%r10 # 3460 <_sk_store_4444_hsw+0xd1>
+ DB 76,141,21,69,0,0,0 ; lea 0x45(%rip),%r10 # 3638 <_sk_store_4444_hsw+0xd1>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -3120,7 +3243,7 @@ _sk_store_4444_hsw LABEL PROC
DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- DB 235,159 ; jmp 33fc <_sk_store_4444_hsw+0x6d>
+ DB 235,159 ; jmp 35d4 <_sk_store_4444_hsw+0x6d>
DB 15,31,0 ; nopl (%rax)
DB 244 ; hlt
DB 255 ; (bad)
@@ -3153,16 +3276,16 @@ _sk_load_8888_hsw LABEL PROC
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 76,3,16 ; add (%rax),%r10
DB 77,133,192 ; test %r8,%r8
- DB 117,88 ; jne 34e9 <_sk_load_8888_hsw+0x6d>
+ DB 117,88 ; jne 36c1 <_sk_load_8888_hsw+0x6d>
DB 196,193,124,16,26 ; vmovups (%r10),%ymm3
- DB 197,228,84,5,226,26,0,0 ; vandps 0x1ae2(%rip),%ymm3,%ymm0 # 4f80 <_sk_callback_hsw+0x664>
+ DB 197,228,84,5,234,26,0,0 ; vandps 0x1aea(%rip),%ymm3,%ymm0 # 5160 <_sk_callback_hsw+0x66c>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,153,24,0,0 ; vbroadcastss 0x1899(%rip),%ymm8 # 4d44 <_sk_callback_hsw+0x428>
+ DB 196,98,125,24,5,153,24,0,0 ; vbroadcastss 0x1899(%rip),%ymm8 # 4f1c <_sk_callback_hsw+0x428>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,231,26,0,0 ; vpshufb 0x1ae7(%rip),%ymm3,%ymm1 # 4fa0 <_sk_callback_hsw+0x684>
+ DB 196,226,101,0,13,239,26,0,0 ; vpshufb 0x1aef(%rip),%ymm3,%ymm1 # 5180 <_sk_callback_hsw+0x68c>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,245,26,0,0 ; vpshufb 0x1af5(%rip),%ymm3,%ymm2 # 4fc0 <_sk_callback_hsw+0x6a4>
+ DB 196,226,101,0,21,253,26,0,0 ; vpshufb 0x1afd(%rip),%ymm3,%ymm2 # 51a0 <_sk_callback_hsw+0x6ac>
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3
@@ -3179,7 +3302,7 @@ _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 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3
- DB 235,135 ; jmp 3496 <_sk_load_8888_hsw+0x1a>
+ DB 235,135 ; jmp 366e <_sk_load_8888_hsw+0x1a>
PUBLIC _sk_gather_8888_hsw
_sk_gather_8888_hsw LABEL PROC
@@ -3192,14 +3315,14 @@ _sk_gather_8888_hsw LABEL PROC
DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0
DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1
DB 196,194,117,144,28,129 ; vpgatherdd %ymm1,(%r9,%ymm0,4),%ymm3
- DB 197,229,219,5,163,26,0,0 ; vpand 0x1aa3(%rip),%ymm3,%ymm0 # 4fe0 <_sk_callback_hsw+0x6c4>
+ DB 197,229,219,5,171,26,0,0 ; vpand 0x1aab(%rip),%ymm3,%ymm0 # 51c0 <_sk_callback_hsw+0x6cc>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,254,23,0,0 ; vbroadcastss 0x17fe(%rip),%ymm8 # 4d48 <_sk_callback_hsw+0x42c>
+ DB 196,98,125,24,5,254,23,0,0 ; vbroadcastss 0x17fe(%rip),%ymm8 # 4f20 <_sk_callback_hsw+0x42c>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,168,26,0,0 ; vpshufb 0x1aa8(%rip),%ymm3,%ymm1 # 5000 <_sk_callback_hsw+0x6e4>
+ DB 196,226,101,0,13,176,26,0,0 ; vpshufb 0x1ab0(%rip),%ymm3,%ymm1 # 51e0 <_sk_callback_hsw+0x6ec>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,182,26,0,0 ; vpshufb 0x1ab6(%rip),%ymm3,%ymm2 # 5020 <_sk_callback_hsw+0x704>
+ DB 196,226,101,0,21,190,26,0,0 ; vpshufb 0x1abe(%rip),%ymm3,%ymm2 # 5200 <_sk_callback_hsw+0x70c>
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3
@@ -3214,7 +3337,7 @@ _sk_store_8888_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 76,3,16 ; add (%rax),%r10
- DB 196,98,125,24,5,174,23,0,0 ; vbroadcastss 0x17ae(%rip),%ymm8 # 4d4c <_sk_callback_hsw+0x430>
+ DB 196,98,125,24,5,174,23,0,0 ; vbroadcastss 0x17ae(%rip),%ymm8 # 4f24 <_sk_callback_hsw+0x430>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10
@@ -3230,7 +3353,7 @@ _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 35f8 <_sk_store_8888_hsw+0x73>
+ DB 117,12 ; jne 37d0 <_sk_store_8888_hsw+0x73>
DB 196,65,124,17,2 ; vmovups %ymm8,(%r10)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,201 ; mov %r9,%rcx
@@ -3243,14 +3366,14 @@ _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 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10)
- DB 235,211 ; jmp 35f1 <_sk_store_8888_hsw+0x6c>
+ DB 235,211 ; jmp 37c9 <_sk_store_8888_hsw+0x6c>
PUBLIC _sk_load_f16_hsw
_sk_load_f16_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,97 ; jne 3689 <_sk_load_f16_hsw+0x6b>
+ DB 117,97 ; jne 3861 <_sk_load_f16_hsw+0x6b>
DB 197,121,16,4,208 ; vmovupd (%rax,%rdx,8),%xmm8
DB 197,249,16,84,208,16 ; vmovupd 0x10(%rax,%rdx,8),%xmm2
DB 197,249,16,92,208,32 ; vmovupd 0x20(%rax,%rdx,8),%xmm3
@@ -3276,29 +3399,29 @@ _sk_load_f16_hsw LABEL PROC
DB 197,123,16,4,208 ; vmovsd (%rax,%rdx,8),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,79 ; je 36e8 <_sk_load_f16_hsw+0xca>
+ DB 116,79 ; je 38c0 <_sk_load_f16_hsw+0xca>
DB 197,57,22,68,208,8 ; vmovhpd 0x8(%rax,%rdx,8),%xmm8,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,67 ; jb 36e8 <_sk_load_f16_hsw+0xca>
+ DB 114,67 ; jb 38c0 <_sk_load_f16_hsw+0xca>
DB 197,251,16,84,208,16 ; vmovsd 0x10(%rax,%rdx,8),%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 116,68 ; je 36f5 <_sk_load_f16_hsw+0xd7>
+ DB 116,68 ; je 38cd <_sk_load_f16_hsw+0xd7>
DB 197,233,22,84,208,24 ; vmovhpd 0x18(%rax,%rdx,8),%xmm2,%xmm2
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,56 ; jb 36f5 <_sk_load_f16_hsw+0xd7>
+ DB 114,56 ; jb 38cd <_sk_load_f16_hsw+0xd7>
DB 197,251,16,92,208,32 ; vmovsd 0x20(%rax,%rdx,8),%xmm3
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 15,132,114,255,255,255 ; je 363f <_sk_load_f16_hsw+0x21>
+ DB 15,132,114,255,255,255 ; je 3817 <_sk_load_f16_hsw+0x21>
DB 197,225,22,92,208,40 ; vmovhpd 0x28(%rax,%rdx,8),%xmm3,%xmm3
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 15,130,98,255,255,255 ; jb 363f <_sk_load_f16_hsw+0x21>
+ DB 15,130,98,255,255,255 ; jb 3817 <_sk_load_f16_hsw+0x21>
DB 197,122,126,76,208,48 ; vmovq 0x30(%rax,%rdx,8),%xmm9
- DB 233,87,255,255,255 ; jmpq 363f <_sk_load_f16_hsw+0x21>
+ DB 233,87,255,255,255 ; jmpq 3817 <_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 363f <_sk_load_f16_hsw+0x21>
+ DB 233,74,255,255,255 ; jmpq 3817 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,65,255,255,255 ; jmpq 363f <_sk_load_f16_hsw+0x21>
+ DB 233,65,255,255,255 ; jmpq 3817 <_sk_load_f16_hsw+0x21>
PUBLIC _sk_gather_f16_hsw
_sk_gather_f16_hsw LABEL PROC
@@ -3352,7 +3475,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,27 ; jne 37ed <_sk_store_f16_hsw+0x65>
+ DB 117,27 ; jne 39c5 <_sk_store_f16_hsw+0x65>
DB 197,120,17,28,208 ; vmovups %xmm11,(%rax,%rdx,8)
DB 197,120,17,84,208,16 ; vmovups %xmm10,0x10(%rax,%rdx,8)
DB 197,120,17,76,208,32 ; vmovups %xmm9,0x20(%rax,%rdx,8)
@@ -3361,22 +3484,22 @@ _sk_store_f16_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,208 ; vmovq %xmm11,(%rax,%rdx,8)
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,241 ; je 37e9 <_sk_store_f16_hsw+0x61>
+ DB 116,241 ; je 39c1 <_sk_store_f16_hsw+0x61>
DB 197,121,23,92,208,8 ; vmovhpd %xmm11,0x8(%rax,%rdx,8)
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,229 ; jb 37e9 <_sk_store_f16_hsw+0x61>
+ DB 114,229 ; jb 39c1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,84,208,16 ; vmovq %xmm10,0x10(%rax,%rdx,8)
- DB 116,221 ; je 37e9 <_sk_store_f16_hsw+0x61>
+ DB 116,221 ; je 39c1 <_sk_store_f16_hsw+0x61>
DB 197,121,23,84,208,24 ; vmovhpd %xmm10,0x18(%rax,%rdx,8)
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,209 ; jb 37e9 <_sk_store_f16_hsw+0x61>
+ DB 114,209 ; jb 39c1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,76,208,32 ; vmovq %xmm9,0x20(%rax,%rdx,8)
- DB 116,201 ; je 37e9 <_sk_store_f16_hsw+0x61>
+ DB 116,201 ; je 39c1 <_sk_store_f16_hsw+0x61>
DB 197,121,23,76,208,40 ; vmovhpd %xmm9,0x28(%rax,%rdx,8)
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,189 ; jb 37e9 <_sk_store_f16_hsw+0x61>
+ DB 114,189 ; jb 39c1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,68,208,48 ; vmovq %xmm8,0x30(%rax,%rdx,8)
- DB 235,181 ; jmp 37e9 <_sk_store_f16_hsw+0x61>
+ DB 235,181 ; jmp 39c1 <_sk_store_f16_hsw+0x61>
PUBLIC _sk_load_u16_be_hsw
_sk_load_u16_be_hsw LABEL PROC
@@ -3384,7 +3507,7 @@ _sk_load_u16_be_hsw LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax
DB 77,133,192 ; test %r8,%r8
- DB 15,133,204,0,0,0 ; jne 3916 <_sk_load_u16_be_hsw+0xe2>
+ DB 15,133,204,0,0,0 ; jne 3aee <_sk_load_u16_be_hsw+0xe2>
DB 196,65,121,16,4,65 ; vmovupd (%r9,%rax,2),%xmm8
DB 196,193,121,16,84,65,16 ; vmovupd 0x10(%r9,%rax,2),%xmm2
DB 196,193,121,16,92,65,32 ; vmovupd 0x20(%r9,%rax,2),%xmm3
@@ -3403,7 +3526,7 @@ _sk_load_u16_be_hsw LABEL PROC
DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0
DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,21,165,20,0,0 ; vbroadcastss 0x14a5(%rip),%ymm10 # 4d50 <_sk_callback_hsw+0x434>
+ DB 196,98,125,24,21,165,20,0,0 ; vbroadcastss 0x14a5(%rip),%ymm10 # 4f28 <_sk_callback_hsw+0x434>
DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0
DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1
DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2
@@ -3431,29 +3554,29 @@ _sk_load_u16_be_hsw LABEL PROC
DB 196,65,123,16,4,65 ; vmovsd (%r9,%rax,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,85 ; je 397c <_sk_load_u16_be_hsw+0x148>
+ DB 116,85 ; je 3b54 <_sk_load_u16_be_hsw+0x148>
DB 196,65,57,22,68,65,8 ; vmovhpd 0x8(%r9,%rax,2),%xmm8,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,72 ; jb 397c <_sk_load_u16_be_hsw+0x148>
+ DB 114,72 ; jb 3b54 <_sk_load_u16_be_hsw+0x148>
DB 196,193,123,16,84,65,16 ; vmovsd 0x10(%r9,%rax,2),%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 116,72 ; je 3989 <_sk_load_u16_be_hsw+0x155>
+ DB 116,72 ; je 3b61 <_sk_load_u16_be_hsw+0x155>
DB 196,193,105,22,84,65,24 ; vmovhpd 0x18(%r9,%rax,2),%xmm2,%xmm2
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,59 ; jb 3989 <_sk_load_u16_be_hsw+0x155>
+ DB 114,59 ; jb 3b61 <_sk_load_u16_be_hsw+0x155>
DB 196,193,123,16,92,65,32 ; vmovsd 0x20(%r9,%rax,2),%xmm3
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 15,132,6,255,255,255 ; je 3865 <_sk_load_u16_be_hsw+0x31>
+ DB 15,132,6,255,255,255 ; je 3a3d <_sk_load_u16_be_hsw+0x31>
DB 196,193,97,22,92,65,40 ; vmovhpd 0x28(%r9,%rax,2),%xmm3,%xmm3
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 15,130,245,254,255,255 ; jb 3865 <_sk_load_u16_be_hsw+0x31>
+ DB 15,130,245,254,255,255 ; jb 3a3d <_sk_load_u16_be_hsw+0x31>
DB 196,65,122,126,76,65,48 ; vmovq 0x30(%r9,%rax,2),%xmm9
- DB 233,233,254,255,255 ; jmpq 3865 <_sk_load_u16_be_hsw+0x31>
+ DB 233,233,254,255,255 ; jmpq 3a3d <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,220,254,255,255 ; jmpq 3865 <_sk_load_u16_be_hsw+0x31>
+ DB 233,220,254,255,255 ; jmpq 3a3d <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,211,254,255,255 ; jmpq 3865 <_sk_load_u16_be_hsw+0x31>
+ DB 233,211,254,255,255 ; jmpq 3a3d <_sk_load_u16_be_hsw+0x31>
PUBLIC _sk_load_rgb_u16_be_hsw
_sk_load_rgb_u16_be_hsw LABEL PROC
@@ -3461,7 +3584,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 72,141,4,82 ; lea (%rdx,%rdx,2),%rax
DB 77,133,192 ; test %r8,%r8
- DB 15,133,204,0,0,0 ; jne 3a70 <_sk_load_rgb_u16_be_hsw+0xde>
+ DB 15,133,204,0,0,0 ; jne 3c48 <_sk_load_rgb_u16_be_hsw+0xde>
DB 196,193,122,111,4,65 ; vmovdqu (%r9,%rax,2),%xmm0
DB 196,193,122,111,84,65,12 ; vmovdqu 0xc(%r9,%rax,2),%xmm2
DB 196,193,122,111,76,65,24 ; vmovdqu 0x18(%r9,%rax,2),%xmm1
@@ -3485,7 +3608,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0
DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,21,54,19,0,0 ; vbroadcastss 0x1336(%rip),%ymm10 # 4d54 <_sk_callback_hsw+0x438>
+ DB 196,98,125,24,21,54,19,0,0 ; vbroadcastss 0x1336(%rip),%ymm10 # 4f2c <_sk_callback_hsw+0x438>
DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0
DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1
DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2
@@ -3502,48 +3625,48 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,234,18,0,0 ; vbroadcastss 0x12ea(%rip),%ymm3 # 4d58 <_sk_callback_hsw+0x43c>
+ DB 196,226,125,24,29,234,18,0,0 ; vbroadcastss 0x12ea(%rip),%ymm3 # 4f30 <_sk_callback_hsw+0x43c>
DB 255,224 ; jmpq *%rax
DB 196,193,121,110,4,65 ; vmovd (%r9,%rax,2),%xmm0
DB 196,193,121,196,68,65,4,2 ; vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 117,5 ; jne 3a89 <_sk_load_rgb_u16_be_hsw+0xf7>
- DB 233,79,255,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 3c61 <_sk_load_rgb_u16_be_hsw+0xf7>
+ DB 233,79,255,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,76,65,6 ; vmovd 0x6(%r9,%rax,2),%xmm1
DB 196,65,113,196,68,65,10,2 ; vpinsrw $0x2,0xa(%r9,%rax,2),%xmm1,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,26 ; jb 3ab8 <_sk_load_rgb_u16_be_hsw+0x126>
+ DB 114,26 ; jb 3c90 <_sk_load_rgb_u16_be_hsw+0x126>
DB 196,193,121,110,76,65,12 ; vmovd 0xc(%r9,%rax,2),%xmm1
DB 196,193,113,196,84,65,16,2 ; vpinsrw $0x2,0x10(%r9,%rax,2),%xmm1,%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 117,10 ; jne 3abd <_sk_load_rgb_u16_be_hsw+0x12b>
- DB 233,32,255,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,27,255,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 3c95 <_sk_load_rgb_u16_be_hsw+0x12b>
+ DB 233,32,255,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,27,255,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,76,65,18 ; vmovd 0x12(%r9,%rax,2),%xmm1
DB 196,65,113,196,76,65,22,2 ; vpinsrw $0x2,0x16(%r9,%rax,2),%xmm1,%xmm9
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,26 ; jb 3aec <_sk_load_rgb_u16_be_hsw+0x15a>
+ DB 114,26 ; jb 3cc4 <_sk_load_rgb_u16_be_hsw+0x15a>
DB 196,193,121,110,76,65,24 ; vmovd 0x18(%r9,%rax,2),%xmm1
DB 196,193,113,196,76,65,28,2 ; vpinsrw $0x2,0x1c(%r9,%rax,2),%xmm1,%xmm1
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 117,10 ; jne 3af1 <_sk_load_rgb_u16_be_hsw+0x15f>
- DB 233,236,254,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,231,254,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 3cc9 <_sk_load_rgb_u16_be_hsw+0x15f>
+ DB 233,236,254,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,231,254,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,92,65,30 ; vmovd 0x1e(%r9,%rax,2),%xmm3
DB 196,65,97,196,92,65,34,2 ; vpinsrw $0x2,0x22(%r9,%rax,2),%xmm3,%xmm11
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,20 ; jb 3b1a <_sk_load_rgb_u16_be_hsw+0x188>
+ DB 114,20 ; jb 3cf2 <_sk_load_rgb_u16_be_hsw+0x188>
DB 196,193,121,110,92,65,36 ; vmovd 0x24(%r9,%rax,2),%xmm3
DB 196,193,97,196,92,65,40,2 ; vpinsrw $0x2,0x28(%r9,%rax,2),%xmm3,%xmm3
- DB 233,190,254,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,185,254,255,255 ; jmpq 39d8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,190,254,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,185,254,255,255 ; jmpq 3bb0 <_sk_load_rgb_u16_be_hsw+0x46>
PUBLIC _sk_store_u16_be_hsw
_sk_store_u16_be_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,8 ; mov (%rax),%r9
DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax
- DB 196,98,125,24,5,39,18,0,0 ; vbroadcastss 0x1227(%rip),%ymm8 # 4d5c <_sk_callback_hsw+0x440>
+ DB 196,98,125,24,5,39,18,0,0 ; vbroadcastss 0x1227(%rip),%ymm8 # 4f34 <_sk_callback_hsw+0x440>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10
@@ -3581,7 +3704,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,31 ; jne 3c19 <_sk_store_u16_be_hsw+0xfa>
+ DB 117,31 ; jne 3df1 <_sk_store_u16_be_hsw+0xfa>
DB 196,65,120,17,28,65 ; vmovups %xmm11,(%r9,%rax,2)
DB 196,65,120,17,84,65,16 ; vmovups %xmm10,0x10(%r9,%rax,2)
DB 196,65,120,17,76,65,32 ; vmovups %xmm9,0x20(%r9,%rax,2)
@@ -3590,31 +3713,31 @@ _sk_store_u16_be_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,214,28,65 ; vmovq %xmm11,(%r9,%rax,2)
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,240 ; je 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 116,240 ; je 3ded <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,23,92,65,8 ; vmovhpd %xmm11,0x8(%r9,%rax,2)
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,227 ; jb 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 114,227 ; jb 3ded <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,84,65,16 ; vmovq %xmm10,0x10(%r9,%rax,2)
- DB 116,218 ; je 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 116,218 ; je 3ded <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,23,84,65,24 ; vmovhpd %xmm10,0x18(%r9,%rax,2)
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,205 ; jb 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 114,205 ; jb 3ded <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,76,65,32 ; vmovq %xmm9,0x20(%r9,%rax,2)
- DB 116,196 ; je 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 116,196 ; je 3ded <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,23,76,65,40 ; vmovhpd %xmm9,0x28(%r9,%rax,2)
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,183 ; jb 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 114,183 ; jb 3ded <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,68,65,48 ; vmovq %xmm8,0x30(%r9,%rax,2)
- DB 235,174 ; jmp 3c15 <_sk_store_u16_be_hsw+0xf6>
+ DB 235,174 ; jmp 3ded <_sk_store_u16_be_hsw+0xf6>
PUBLIC _sk_load_f32_hsw
_sk_load_f32_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 119,110 ; ja 3cdd <_sk_load_f32_hsw+0x76>
+ DB 119,110 ; ja 3eb5 <_sk_load_f32_hsw+0x76>
DB 76,139,8 ; mov (%rax),%r9
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
- DB 76,141,29,135,0,0,0 ; lea 0x87(%rip),%r11 # 3d08 <_sk_load_f32_hsw+0xa1>
+ DB 76,141,29,135,0,0,0 ; lea 0x87(%rip),%r11 # 3ee0 <_sk_load_f32_hsw+0xa1>
DB 75,99,4,131 ; movslq (%r11,%r8,4),%rax
DB 76,1,216 ; add %r11,%rax
DB 255,224 ; jmpq *%rax
@@ -3673,7 +3796,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,55 ; jne 3d95 <_sk_store_f32_hsw+0x6d>
+ DB 117,55 ; jne 3f6d <_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
@@ -3686,22 +3809,22 @@ _sk_store_f32_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,17,20,129 ; vmovupd %xmm10,(%r9,%rax,4)
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,240 ; je 3d91 <_sk_store_f32_hsw+0x69>
+ DB 116,240 ; je 3f69 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,76,129,16 ; vmovupd %xmm9,0x10(%r9,%rax,4)
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,227 ; jb 3d91 <_sk_store_f32_hsw+0x69>
+ DB 114,227 ; jb 3f69 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,68,129,32 ; vmovupd %xmm8,0x20(%r9,%rax,4)
- DB 116,218 ; je 3d91 <_sk_store_f32_hsw+0x69>
+ DB 116,218 ; je 3f69 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,92,129,48 ; vmovupd %xmm11,0x30(%r9,%rax,4)
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,205 ; jb 3d91 <_sk_store_f32_hsw+0x69>
+ DB 114,205 ; jb 3f69 <_sk_store_f32_hsw+0x69>
DB 196,67,125,25,84,129,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r9,%rax,4)
- DB 116,195 ; je 3d91 <_sk_store_f32_hsw+0x69>
+ DB 116,195 ; je 3f69 <_sk_store_f32_hsw+0x69>
DB 196,67,125,25,76,129,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r9,%rax,4)
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,181 ; jb 3d91 <_sk_store_f32_hsw+0x69>
+ DB 114,181 ; jb 3f69 <_sk_store_f32_hsw+0x69>
DB 196,67,125,25,68,129,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r9,%rax,4)
- DB 235,171 ; jmp 3d91 <_sk_store_f32_hsw+0x69>
+ DB 235,171 ; jmp 3f69 <_sk_store_f32_hsw+0x69>
PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
@@ -3785,7 +3908,7 @@ PUBLIC _sk_clamp_x_1_hsw
_sk_clamp_x_1_hsw LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,122,14,0,0 ; vbroadcastss 0xe7a(%rip),%ymm8 # 4d60 <_sk_callback_hsw+0x444>
+ DB 196,98,125,24,5,122,14,0,0 ; vbroadcastss 0xe7a(%rip),%ymm8 # 4f38 <_sk_callback_hsw+0x444>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -3799,9 +3922,9 @@ _sk_repeat_x_1_hsw LABEL PROC
PUBLIC _sk_mirror_x_1_hsw
_sk_mirror_x_1_hsw LABEL PROC
- DB 196,98,125,24,5,93,14,0,0 ; vbroadcastss 0xe5d(%rip),%ymm8 # 4d64 <_sk_callback_hsw+0x448>
+ DB 196,98,125,24,5,93,14,0,0 ; vbroadcastss 0xe5d(%rip),%ymm8 # 4f3c <_sk_callback_hsw+0x448>
DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,13,83,14,0,0 ; vbroadcastss 0xe53(%rip),%ymm9 # 4d68 <_sk_callback_hsw+0x44c>
+ DB 196,98,125,24,13,83,14,0,0 ; vbroadcastss 0xe53(%rip),%ymm9 # 4f40 <_sk_callback_hsw+0x44c>
DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9
DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9
DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9
@@ -3815,11 +3938,11 @@ _sk_mirror_x_1_hsw LABEL PROC
PUBLIC _sk_luminance_to_alpha_hsw
_sk_luminance_to_alpha_hsw LABEL PROC
- DB 196,226,125,24,29,35,14,0,0 ; vbroadcastss 0xe23(%rip),%ymm3 # 4d6c <_sk_callback_hsw+0x450>
- DB 196,98,125,24,5,30,14,0,0 ; vbroadcastss 0xe1e(%rip),%ymm8 # 4d70 <_sk_callback_hsw+0x454>
+ DB 196,226,125,24,29,35,14,0,0 ; vbroadcastss 0xe23(%rip),%ymm3 # 4f44 <_sk_callback_hsw+0x450>
+ DB 196,98,125,24,5,30,14,0,0 ; vbroadcastss 0xe1e(%rip),%ymm8 # 4f48 <_sk_callback_hsw+0x454>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1
- DB 196,226,125,24,29,15,14,0,0 ; vbroadcastss 0xe0f(%rip),%ymm3 # 4d74 <_sk_callback_hsw+0x458>
+ DB 196,226,125,24,29,15,14,0,0 ; vbroadcastss 0xe0f(%rip),%ymm3 # 4f4c <_sk_callback_hsw+0x458>
DB 196,226,109,168,217 ; vfmadd213ps %ymm1,%ymm2,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
@@ -3981,9 +4104,9 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
DB 76,139,72,8 ; mov 0x8(%rax),%r9
DB 77,137,211 ; mov %r10,%r11
DB 73,255,203 ; dec %r11
- DB 120,7 ; js 421e <_sk_evenly_spaced_gradient_hsw+0x19>
+ DB 120,7 ; js 43f6 <_sk_evenly_spaced_gradient_hsw+0x19>
DB 196,193,242,42,203 ; vcvtsi2ss %r11,%xmm1,%xmm1
- DB 235,22 ; jmp 4234 <_sk_evenly_spaced_gradient_hsw+0x2f>
+ DB 235,22 ; jmp 440c <_sk_evenly_spaced_gradient_hsw+0x2f>
DB 76,137,219 ; mov %r11,%rbx
DB 72,209,235 ; shr %rbx
DB 65,131,227,1 ; and $0x1,%r11d
@@ -3994,7 +4117,7 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
DB 197,244,89,200 ; vmulps %ymm0,%ymm1,%ymm1
DB 197,126,91,217 ; vcvttps2dq %ymm1,%ymm11
DB 73,131,250,8 ; cmp $0x8,%r10
- DB 119,70 ; ja 428d <_sk_evenly_spaced_gradient_hsw+0x88>
+ DB 119,70 ; ja 4465 <_sk_evenly_spaced_gradient_hsw+0x88>
DB 196,66,37,22,1 ; vpermps (%r9),%ymm11,%ymm8
DB 72,139,88,40 ; mov 0x28(%rax),%rbx
DB 196,98,37,22,11 ; vpermps (%rbx),%ymm11,%ymm9
@@ -4010,7 +4133,7 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
DB 196,226,37,22,27 ; vpermps (%rbx),%ymm11,%ymm3
DB 72,139,64,64 ; mov 0x40(%rax),%rax
DB 196,98,37,22,40 ; vpermps (%rax),%ymm11,%ymm13
- DB 235,110 ; jmp 42fb <_sk_evenly_spaced_gradient_hsw+0xf6>
+ DB 235,110 ; jmp 44d3 <_sk_evenly_spaced_gradient_hsw+0xf6>
DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14
DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1
DB 196,2,117,146,4,153 ; vgatherdps %ymm1,(%r9,%ymm11,4),%ymm8
@@ -4045,14 +4168,14 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
PUBLIC _sk_gauss_a_to_rgba_hsw
_sk_gauss_a_to_rgba_hsw LABEL PROC
- DB 196,226,125,24,5,87,10,0,0 ; vbroadcastss 0xa57(%rip),%ymm0 # 4d78 <_sk_callback_hsw+0x45c>
- DB 196,226,125,24,13,82,10,0,0 ; vbroadcastss 0xa52(%rip),%ymm1 # 4d7c <_sk_callback_hsw+0x460>
+ DB 196,226,125,24,5,87,10,0,0 ; vbroadcastss 0xa57(%rip),%ymm0 # 4f50 <_sk_callback_hsw+0x45c>
+ DB 196,226,125,24,13,82,10,0,0 ; vbroadcastss 0xa52(%rip),%ymm1 # 4f54 <_sk_callback_hsw+0x460>
DB 196,226,101,168,200 ; vfmadd213ps %ymm0,%ymm3,%ymm1
- DB 196,226,125,24,5,72,10,0,0 ; vbroadcastss 0xa48(%rip),%ymm0 # 4d80 <_sk_callback_hsw+0x464>
+ DB 196,226,125,24,5,72,10,0,0 ; vbroadcastss 0xa48(%rip),%ymm0 # 4f58 <_sk_callback_hsw+0x464>
DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0
- DB 196,226,125,24,13,62,10,0,0 ; vbroadcastss 0xa3e(%rip),%ymm1 # 4d84 <_sk_callback_hsw+0x468>
+ DB 196,226,125,24,13,62,10,0,0 ; vbroadcastss 0xa3e(%rip),%ymm1 # 4f5c <_sk_callback_hsw+0x468>
DB 196,226,101,184,200 ; vfmadd231ps %ymm0,%ymm3,%ymm1
- DB 196,226,125,24,5,52,10,0,0 ; vbroadcastss 0xa34(%rip),%ymm0 # 4d88 <_sk_callback_hsw+0x46c>
+ DB 196,226,125,24,5,52,10,0,0 ; vbroadcastss 0xa34(%rip),%ymm0 # 4f60 <_sk_callback_hsw+0x46c>
DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
@@ -4065,11 +4188,11 @@ _sk_gradient_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,8 ; mov (%rax),%r9
DB 73,131,249,1 ; cmp $0x1,%r9
- DB 15,134,180,0,0,0 ; jbe 442c <_sk_gradient_hsw+0xc3>
+ DB 15,134,180,0,0,0 ; jbe 4604 <_sk_gradient_hsw+0xc3>
DB 76,139,80,72 ; mov 0x48(%rax),%r10
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 65,187,1,0,0,0 ; mov $0x1,%r11d
- DB 196,226,125,24,21,253,9,0,0 ; vbroadcastss 0x9fd(%rip),%ymm2 # 4d8c <_sk_callback_hsw+0x470>
+ DB 196,226,125,24,21,253,9,0,0 ; vbroadcastss 0x9fd(%rip),%ymm2 # 4f64 <_sk_callback_hsw+0x470>
DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9
DB 196,130,125,24,28,154 ; vbroadcastss (%r10,%r11,4),%ymm3
DB 197,228,194,216,2 ; vcmpleps %ymm0,%ymm3,%ymm3
@@ -4077,10 +4200,10 @@ _sk_gradient_hsw LABEL PROC
DB 196,65,101,254,201 ; vpaddd %ymm9,%ymm3,%ymm9
DB 73,255,195 ; inc %r11
DB 77,57,217 ; cmp %r11,%r9
- DB 117,226 ; jne 4394 <_sk_gradient_hsw+0x2b>
+ DB 117,226 ; jne 456c <_sk_gradient_hsw+0x2b>
DB 76,139,80,8 ; mov 0x8(%rax),%r10
DB 73,131,249,8 ; cmp $0x8,%r9
- DB 118,121 ; jbe 4435 <_sk_gradient_hsw+0xcc>
+ DB 118,121 ; jbe 460d <_sk_gradient_hsw+0xcc>
DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14
DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1
DB 196,2,117,146,4,138 ; vgatherdps %ymm1,(%r10,%ymm9,4),%ymm8
@@ -4104,7 +4227,7 @@ _sk_gradient_hsw LABEL PROC
DB 196,130,21,146,28,137 ; vgatherdps %ymm13,(%r9,%ymm9,4),%ymm3
DB 72,139,64,64 ; mov 0x40(%rax),%rax
DB 196,34,13,146,44,136 ; vgatherdps %ymm14,(%rax,%ymm9,4),%ymm13
- DB 235,77 ; jmp 4479 <_sk_gradient_hsw+0x110>
+ DB 235,77 ; jmp 4651 <_sk_gradient_hsw+0x110>
DB 76,139,80,8 ; mov 0x8(%rax),%r10
DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9
DB 196,66,53,22,2 ; vpermps (%r10),%ymm9,%ymm8
@@ -4160,24 +4283,24 @@ _sk_xy_to_unit_angle_hsw LABEL PROC
DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12
DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11
DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12
- DB 196,98,125,24,45,124,8,0,0 ; vbroadcastss 0x87c(%rip),%ymm13 # 4d90 <_sk_callback_hsw+0x474>
- DB 196,98,125,24,53,119,8,0,0 ; vbroadcastss 0x877(%rip),%ymm14 # 4d94 <_sk_callback_hsw+0x478>
+ DB 196,98,125,24,45,124,8,0,0 ; vbroadcastss 0x87c(%rip),%ymm13 # 4f68 <_sk_callback_hsw+0x474>
+ DB 196,98,125,24,53,119,8,0,0 ; vbroadcastss 0x877(%rip),%ymm14 # 4f6c <_sk_callback_hsw+0x478>
DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14
- DB 196,98,125,24,45,109,8,0,0 ; vbroadcastss 0x86d(%rip),%ymm13 # 4d98 <_sk_callback_hsw+0x47c>
+ DB 196,98,125,24,45,109,8,0,0 ; vbroadcastss 0x86d(%rip),%ymm13 # 4f70 <_sk_callback_hsw+0x47c>
DB 196,66,29,184,238 ; vfmadd231ps %ymm14,%ymm12,%ymm13
- DB 196,98,125,24,53,99,8,0,0 ; vbroadcastss 0x863(%rip),%ymm14 # 4d9c <_sk_callback_hsw+0x480>
+ DB 196,98,125,24,53,99,8,0,0 ; vbroadcastss 0x863(%rip),%ymm14 # 4f74 <_sk_callback_hsw+0x480>
DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14
DB 196,65,36,89,222 ; vmulps %ymm14,%ymm11,%ymm11
DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9
- DB 196,98,125,24,21,78,8,0,0 ; vbroadcastss 0x84e(%rip),%ymm10 # 4da0 <_sk_callback_hsw+0x484>
+ DB 196,98,125,24,21,78,8,0,0 ; vbroadcastss 0x84e(%rip),%ymm10 # 4f78 <_sk_callback_hsw+0x484>
DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10
DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9
DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,21,56,8,0,0 ; vbroadcastss 0x838(%rip),%ymm10 # 4da4 <_sk_callback_hsw+0x488>
+ DB 196,98,125,24,21,56,8,0,0 ; vbroadcastss 0x838(%rip),%ymm10 # 4f7c <_sk_callback_hsw+0x488>
DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10
DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0
DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9
- DB 196,98,125,24,21,34,8,0,0 ; vbroadcastss 0x822(%rip),%ymm10 # 4da8 <_sk_callback_hsw+0x48c>
+ DB 196,98,125,24,21,34,8,0,0 ; vbroadcastss 0x822(%rip),%ymm10 # 4f80 <_sk_callback_hsw+0x48c>
DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10
DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0
DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9
@@ -4196,7 +4319,7 @@ _sk_xy_to_radius_hsw LABEL PROC
PUBLIC _sk_save_xy_hsw
_sk_save_xy_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,239,7,0,0 ; vbroadcastss 0x7ef(%rip),%ymm8 # 4dac <_sk_callback_hsw+0x490>
+ DB 196,98,125,24,5,239,7,0,0 ; vbroadcastss 0x7ef(%rip),%ymm8 # 4f84 <_sk_callback_hsw+0x490>
DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9
DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10
DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9
@@ -4226,9 +4349,9 @@ _sk_accumulate_hsw LABEL PROC
PUBLIC _sk_bilinear_nx_hsw
_sk_bilinear_nx_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,131,7,0,0 ; vbroadcastss 0x783(%rip),%ymm0 # 4db0 <_sk_callback_hsw+0x494>
+ DB 196,226,125,24,5,131,7,0,0 ; vbroadcastss 0x783(%rip),%ymm0 # 4f88 <_sk_callback_hsw+0x494>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,122,7,0,0 ; vbroadcastss 0x77a(%rip),%ymm8 # 4db4 <_sk_callback_hsw+0x498>
+ DB 196,98,125,24,5,122,7,0,0 ; vbroadcastss 0x77a(%rip),%ymm8 # 4f8c <_sk_callback_hsw+0x498>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -4237,7 +4360,7 @@ _sk_bilinear_nx_hsw LABEL PROC
PUBLIC _sk_bilinear_px_hsw
_sk_bilinear_px_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,98,7,0,0 ; vbroadcastss 0x762(%rip),%ymm0 # 4db8 <_sk_callback_hsw+0x49c>
+ DB 196,226,125,24,5,98,7,0,0 ; vbroadcastss 0x762(%rip),%ymm0 # 4f90 <_sk_callback_hsw+0x49c>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
@@ -4247,9 +4370,9 @@ _sk_bilinear_px_hsw LABEL PROC
PUBLIC _sk_bilinear_ny_hsw
_sk_bilinear_ny_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,70,7,0,0 ; vbroadcastss 0x746(%rip),%ymm1 # 4dbc <_sk_callback_hsw+0x4a0>
+ DB 196,226,125,24,13,70,7,0,0 ; vbroadcastss 0x746(%rip),%ymm1 # 4f94 <_sk_callback_hsw+0x4a0>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,60,7,0,0 ; vbroadcastss 0x73c(%rip),%ymm8 # 4dc0 <_sk_callback_hsw+0x4a4>
+ DB 196,98,125,24,5,60,7,0,0 ; vbroadcastss 0x73c(%rip),%ymm8 # 4f98 <_sk_callback_hsw+0x4a4>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -4258,7 +4381,7 @@ _sk_bilinear_ny_hsw LABEL PROC
PUBLIC _sk_bilinear_py_hsw
_sk_bilinear_py_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,36,7,0,0 ; vbroadcastss 0x724(%rip),%ymm1 # 4dc4 <_sk_callback_hsw+0x4a8>
+ DB 196,226,125,24,13,36,7,0,0 ; vbroadcastss 0x724(%rip),%ymm1 # 4f9c <_sk_callback_hsw+0x4a8>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
@@ -4268,13 +4391,13 @@ _sk_bilinear_py_hsw LABEL PROC
PUBLIC _sk_bicubic_n3x_hsw
_sk_bicubic_n3x_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,7,7,0,0 ; vbroadcastss 0x707(%rip),%ymm0 # 4dc8 <_sk_callback_hsw+0x4ac>
+ DB 196,226,125,24,5,7,7,0,0 ; vbroadcastss 0x707(%rip),%ymm0 # 4fa0 <_sk_callback_hsw+0x4ac>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,254,6,0,0 ; vbroadcastss 0x6fe(%rip),%ymm8 # 4dcc <_sk_callback_hsw+0x4b0>
+ DB 196,98,125,24,5,254,6,0,0 ; vbroadcastss 0x6fe(%rip),%ymm8 # 4fa4 <_sk_callback_hsw+0x4b0>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,239,6,0,0 ; vbroadcastss 0x6ef(%rip),%ymm10 # 4dd0 <_sk_callback_hsw+0x4b4>
- DB 196,98,125,24,29,234,6,0,0 ; vbroadcastss 0x6ea(%rip),%ymm11 # 4dd4 <_sk_callback_hsw+0x4b8>
+ DB 196,98,125,24,21,239,6,0,0 ; vbroadcastss 0x6ef(%rip),%ymm10 # 4fa8 <_sk_callback_hsw+0x4b4>
+ DB 196,98,125,24,29,234,6,0,0 ; vbroadcastss 0x6ea(%rip),%ymm11 # 4fac <_sk_callback_hsw+0x4b8>
DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11
DB 196,65,36,89,193 ; vmulps %ymm9,%ymm11,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
@@ -4284,16 +4407,16 @@ _sk_bicubic_n3x_hsw LABEL PROC
PUBLIC _sk_bicubic_n1x_hsw
_sk_bicubic_n1x_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,205,6,0,0 ; vbroadcastss 0x6cd(%rip),%ymm0 # 4dd8 <_sk_callback_hsw+0x4bc>
+ DB 196,226,125,24,5,205,6,0,0 ; vbroadcastss 0x6cd(%rip),%ymm0 # 4fb0 <_sk_callback_hsw+0x4bc>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,196,6,0,0 ; vbroadcastss 0x6c4(%rip),%ymm8 # 4ddc <_sk_callback_hsw+0x4c0>
+ DB 196,98,125,24,5,196,6,0,0 ; vbroadcastss 0x6c4(%rip),%ymm8 # 4fb4 <_sk_callback_hsw+0x4c0>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,186,6,0,0 ; vbroadcastss 0x6ba(%rip),%ymm9 # 4de0 <_sk_callback_hsw+0x4c4>
- DB 196,98,125,24,21,181,6,0,0 ; vbroadcastss 0x6b5(%rip),%ymm10 # 4de4 <_sk_callback_hsw+0x4c8>
+ DB 196,98,125,24,13,186,6,0,0 ; vbroadcastss 0x6ba(%rip),%ymm9 # 4fb8 <_sk_callback_hsw+0x4c4>
+ DB 196,98,125,24,21,181,6,0,0 ; vbroadcastss 0x6b5(%rip),%ymm10 # 4fbc <_sk_callback_hsw+0x4c8>
DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,13,171,6,0,0 ; vbroadcastss 0x6ab(%rip),%ymm9 # 4de8 <_sk_callback_hsw+0x4cc>
+ DB 196,98,125,24,13,171,6,0,0 ; vbroadcastss 0x6ab(%rip),%ymm9 # 4fc0 <_sk_callback_hsw+0x4cc>
DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9
- DB 196,98,125,24,21,161,6,0,0 ; vbroadcastss 0x6a1(%rip),%ymm10 # 4dec <_sk_callback_hsw+0x4d0>
+ DB 196,98,125,24,21,161,6,0,0 ; vbroadcastss 0x6a1(%rip),%ymm10 # 4fc4 <_sk_callback_hsw+0x4d0>
DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10
DB 197,124,17,144,128,0,0,0 ; vmovups %ymm10,0x80(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -4302,14 +4425,14 @@ _sk_bicubic_n1x_hsw LABEL PROC
PUBLIC _sk_bicubic_p1x_hsw
_sk_bicubic_p1x_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,137,6,0,0 ; vbroadcastss 0x689(%rip),%ymm8 # 4df0 <_sk_callback_hsw+0x4d4>
+ DB 196,98,125,24,5,137,6,0,0 ; vbroadcastss 0x689(%rip),%ymm8 # 4fc8 <_sk_callback_hsw+0x4d4>
DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0
DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9
- DB 196,98,125,24,21,123,6,0,0 ; vbroadcastss 0x67b(%rip),%ymm10 # 4df4 <_sk_callback_hsw+0x4d8>
- DB 196,98,125,24,29,118,6,0,0 ; vbroadcastss 0x676(%rip),%ymm11 # 4df8 <_sk_callback_hsw+0x4dc>
+ DB 196,98,125,24,21,123,6,0,0 ; vbroadcastss 0x67b(%rip),%ymm10 # 4fcc <_sk_callback_hsw+0x4d8>
+ DB 196,98,125,24,29,118,6,0,0 ; vbroadcastss 0x676(%rip),%ymm11 # 4fd0 <_sk_callback_hsw+0x4dc>
DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11
DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11
- DB 196,98,125,24,5,103,6,0,0 ; vbroadcastss 0x667(%rip),%ymm8 # 4dfc <_sk_callback_hsw+0x4e0>
+ DB 196,98,125,24,5,103,6,0,0 ; vbroadcastss 0x667(%rip),%ymm8 # 4fd4 <_sk_callback_hsw+0x4e0>
DB 196,66,53,184,195 ; vfmadd231ps %ymm11,%ymm9,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -4318,12 +4441,12 @@ _sk_bicubic_p1x_hsw LABEL PROC
PUBLIC _sk_bicubic_p3x_hsw
_sk_bicubic_p3x_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,79,6,0,0 ; vbroadcastss 0x64f(%rip),%ymm0 # 4e00 <_sk_callback_hsw+0x4e4>
+ DB 196,226,125,24,5,79,6,0,0 ; vbroadcastss 0x64f(%rip),%ymm0 # 4fd8 <_sk_callback_hsw+0x4e4>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,60,6,0,0 ; vbroadcastss 0x63c(%rip),%ymm10 # 4e04 <_sk_callback_hsw+0x4e8>
- DB 196,98,125,24,29,55,6,0,0 ; vbroadcastss 0x637(%rip),%ymm11 # 4e08 <_sk_callback_hsw+0x4ec>
+ DB 196,98,125,24,21,60,6,0,0 ; vbroadcastss 0x63c(%rip),%ymm10 # 4fdc <_sk_callback_hsw+0x4e8>
+ DB 196,98,125,24,29,55,6,0,0 ; vbroadcastss 0x637(%rip),%ymm11 # 4fe0 <_sk_callback_hsw+0x4ec>
DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11
DB 196,65,52,89,195 ; vmulps %ymm11,%ymm9,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
@@ -4333,13 +4456,13 @@ _sk_bicubic_p3x_hsw LABEL PROC
PUBLIC _sk_bicubic_n3y_hsw
_sk_bicubic_n3y_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,26,6,0,0 ; vbroadcastss 0x61a(%rip),%ymm1 # 4e0c <_sk_callback_hsw+0x4f0>
+ DB 196,226,125,24,13,26,6,0,0 ; vbroadcastss 0x61a(%rip),%ymm1 # 4fe4 <_sk_callback_hsw+0x4f0>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm8 # 4e10 <_sk_callback_hsw+0x4f4>
+ DB 196,98,125,24,5,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm8 # 4fe8 <_sk_callback_hsw+0x4f4>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,1,6,0,0 ; vbroadcastss 0x601(%rip),%ymm10 # 4e14 <_sk_callback_hsw+0x4f8>
- DB 196,98,125,24,29,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm11 # 4e18 <_sk_callback_hsw+0x4fc>
+ DB 196,98,125,24,21,1,6,0,0 ; vbroadcastss 0x601(%rip),%ymm10 # 4fec <_sk_callback_hsw+0x4f8>
+ DB 196,98,125,24,29,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm11 # 4ff0 <_sk_callback_hsw+0x4fc>
DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11
DB 196,65,36,89,193 ; vmulps %ymm9,%ymm11,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
@@ -4349,16 +4472,16 @@ _sk_bicubic_n3y_hsw LABEL PROC
PUBLIC _sk_bicubic_n1y_hsw
_sk_bicubic_n1y_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,223,5,0,0 ; vbroadcastss 0x5df(%rip),%ymm1 # 4e1c <_sk_callback_hsw+0x500>
+ DB 196,226,125,24,13,223,5,0,0 ; vbroadcastss 0x5df(%rip),%ymm1 # 4ff4 <_sk_callback_hsw+0x500>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,213,5,0,0 ; vbroadcastss 0x5d5(%rip),%ymm8 # 4e20 <_sk_callback_hsw+0x504>
+ DB 196,98,125,24,5,213,5,0,0 ; vbroadcastss 0x5d5(%rip),%ymm8 # 4ff8 <_sk_callback_hsw+0x504>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,203,5,0,0 ; vbroadcastss 0x5cb(%rip),%ymm9 # 4e24 <_sk_callback_hsw+0x508>
- DB 196,98,125,24,21,198,5,0,0 ; vbroadcastss 0x5c6(%rip),%ymm10 # 4e28 <_sk_callback_hsw+0x50c>
+ DB 196,98,125,24,13,203,5,0,0 ; vbroadcastss 0x5cb(%rip),%ymm9 # 4ffc <_sk_callback_hsw+0x508>
+ DB 196,98,125,24,21,198,5,0,0 ; vbroadcastss 0x5c6(%rip),%ymm10 # 5000 <_sk_callback_hsw+0x50c>
DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,13,188,5,0,0 ; vbroadcastss 0x5bc(%rip),%ymm9 # 4e2c <_sk_callback_hsw+0x510>
+ DB 196,98,125,24,13,188,5,0,0 ; vbroadcastss 0x5bc(%rip),%ymm9 # 5004 <_sk_callback_hsw+0x510>
DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9
- DB 196,98,125,24,21,178,5,0,0 ; vbroadcastss 0x5b2(%rip),%ymm10 # 4e30 <_sk_callback_hsw+0x514>
+ DB 196,98,125,24,21,178,5,0,0 ; vbroadcastss 0x5b2(%rip),%ymm10 # 5008 <_sk_callback_hsw+0x514>
DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10
DB 197,124,17,144,160,0,0,0 ; vmovups %ymm10,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -4367,14 +4490,14 @@ _sk_bicubic_n1y_hsw LABEL PROC
PUBLIC _sk_bicubic_p1y_hsw
_sk_bicubic_p1y_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,154,5,0,0 ; vbroadcastss 0x59a(%rip),%ymm8 # 4e34 <_sk_callback_hsw+0x518>
+ DB 196,98,125,24,5,154,5,0,0 ; vbroadcastss 0x59a(%rip),%ymm8 # 500c <_sk_callback_hsw+0x518>
DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1
DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9
- DB 196,98,125,24,21,139,5,0,0 ; vbroadcastss 0x58b(%rip),%ymm10 # 4e38 <_sk_callback_hsw+0x51c>
- DB 196,98,125,24,29,134,5,0,0 ; vbroadcastss 0x586(%rip),%ymm11 # 4e3c <_sk_callback_hsw+0x520>
+ DB 196,98,125,24,21,139,5,0,0 ; vbroadcastss 0x58b(%rip),%ymm10 # 5010 <_sk_callback_hsw+0x51c>
+ DB 196,98,125,24,29,134,5,0,0 ; vbroadcastss 0x586(%rip),%ymm11 # 5014 <_sk_callback_hsw+0x520>
DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11
DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11
- DB 196,98,125,24,5,119,5,0,0 ; vbroadcastss 0x577(%rip),%ymm8 # 4e40 <_sk_callback_hsw+0x524>
+ DB 196,98,125,24,5,119,5,0,0 ; vbroadcastss 0x577(%rip),%ymm8 # 5018 <_sk_callback_hsw+0x524>
DB 196,66,53,184,195 ; vfmadd231ps %ymm11,%ymm9,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -4383,12 +4506,12 @@ _sk_bicubic_p1y_hsw LABEL PROC
PUBLIC _sk_bicubic_p3y_hsw
_sk_bicubic_p3y_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,95,5,0,0 ; vbroadcastss 0x55f(%rip),%ymm1 # 4e44 <_sk_callback_hsw+0x528>
+ DB 196,226,125,24,13,95,5,0,0 ; vbroadcastss 0x55f(%rip),%ymm1 # 501c <_sk_callback_hsw+0x528>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,75,5,0,0 ; vbroadcastss 0x54b(%rip),%ymm10 # 4e48 <_sk_callback_hsw+0x52c>
- DB 196,98,125,24,29,70,5,0,0 ; vbroadcastss 0x546(%rip),%ymm11 # 4e4c <_sk_callback_hsw+0x530>
+ DB 196,98,125,24,21,75,5,0,0 ; vbroadcastss 0x54b(%rip),%ymm10 # 5020 <_sk_callback_hsw+0x52c>
+ DB 196,98,125,24,29,70,5,0,0 ; vbroadcastss 0x546(%rip),%ymm11 # 5024 <_sk_callback_hsw+0x530>
DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11
DB 196,65,52,89,195 ; vmulps %ymm11,%ymm9,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
@@ -4510,25 +4633,25 @@ ALIGN 4
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 4af9 <.literal4+0xb1>
+ DB 71,225,61 ; rex.RXB loope 4cd1 <.literal4+0xb1>
DB 0,0 ; add %al,(%rax)
DB 128,63,154 ; cmpb $0x9a,(%rdi)
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 4b09 <.literal4+0xc1>
+ DB 71,225,61 ; rex.RXB loope 4ce1 <.literal4+0xc1>
DB 0,0 ; add %al,(%rax)
DB 128,63,154 ; cmpb $0x9a,(%rdi)
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 4b19 <.literal4+0xd1>
+ DB 71,225,61 ; rex.RXB loope 4cf1 <.literal4+0xd1>
DB 0,0 ; add %al,(%rax)
DB 128,63,154 ; cmpb $0x9a,(%rdi)
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 4b29 <.literal4+0xe1>
+ DB 71,225,61 ; rex.RXB loope 4d01 <.literal4+0xe1>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax)
@@ -4578,7 +4701,7 @@ ALIGN 4
DB 190,129,128,128,59 ; mov $0x3b808081,%esi
DB 129,128,128,59,0,248,0,0,8,33 ; addl $0x21080000,-0x7ffc480(%rax)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4b81 <.literal4+0x139>
+ DB 224,7 ; loopne 4d59 <.literal4+0x139>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -4594,10 +4717,10 @@ ALIGN 4
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
DB 0,52,255 ; add %dh,(%rdi,%rdi,8)
DB 255 ; (bad)
- DB 127,0 ; jg 4ba8 <.literal4+0x160>
+ DB 127,0 ; jg 4d80 <.literal4+0x160>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4c21 <.literal4+0x1d9>
+ DB 119,115 ; ja 4df9 <.literal4+0x1d9>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4611,10 +4734,10 @@ ALIGN 4
DB 0,128,63,0,0,0 ; add %al,0x3f(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4bdc <.literal4+0x194>
+ DB 127,0 ; jg 4db4 <.literal4+0x194>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4c55 <.literal4+0x20d>
+ DB 119,115 ; ja 4e2d <.literal4+0x20d>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4628,10 +4751,10 @@ ALIGN 4
DB 0,128,63,0,0,0 ; add %al,0x3f(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4c10 <.literal4+0x1c8>
+ DB 127,0 ; jg 4de8 <.literal4+0x1c8>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4c89 <.literal4+0x241>
+ DB 119,115 ; ja 4e61 <.literal4+0x241>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4645,10 +4768,10 @@ ALIGN 4
DB 0,128,63,0,0,0 ; add %al,0x3f(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4c44 <.literal4+0x1fc>
+ DB 127,0 ; jg 4e1c <.literal4+0x1fc>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4cbd <.literal4+0x275>
+ DB 119,115 ; ja 4e95 <.literal4+0x275>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4661,7 +4784,7 @@ ALIGN 4
DB 0,75,0 ; add %cl,0x0(%rbx)
DB 0,128,63,0,0,200 ; add %al,-0x37ffffc1(%rax)
DB 66,0,0 ; rex.X add %al,(%rax)
- DB 127,67 ; jg 4cbb <.literal4+0x273>
+ DB 127,67 ; jg 4e93 <.literal4+0x273>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -4673,10 +4796,10 @@ ALIGN 4
DB 190,80,128,3,62 ; mov $0x3e038050,%esi
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4cdb <.literal4+0x293>
+ DB 118,63 ; jbe 4eb3 <.literal4+0x293>
DB 246,64,83,63 ; testb $0x3f,0x53(%rax)
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 4cef <.literal4+0x2a7>
+ DB 127,67 ; jg 4ec7 <.literal4+0x2a7>
DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax)
DB 128,59,0 ; cmpb $0x0,(%rbx)
DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax)
@@ -4685,7 +4808,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4cd1 <.literal4+0x289>
+ DB 224,7 ; loopne 4ea9 <.literal4+0x289>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -4697,7 +4820,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4ced <.literal4+0x2a5>
+ DB 224,7 ; loopne 4ec5 <.literal4+0x2a5>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -4708,7 +4831,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 4d42 <.literal4+0x2fa>
+ DB 124,66 ; jl 4f1a <.literal4+0x2fa>
DB 0,240 ; add %dh,%al
DB 0,0 ; add %al,(%rax)
DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax)
@@ -4726,9 +4849,9 @@ ALIGN 4
DB 137,136,136,59,15,0 ; mov %ecx,0xf3b88(%rax)
DB 0,0 ; add %al,(%rax)
DB 137,136,136,61,0,0 ; mov %ecx,0x3d88(%rax)
- DB 112,65 ; jo 4d85 <.literal4+0x33d>
+ DB 112,65 ; jo 4f5d <.literal4+0x33d>
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 4d93 <.literal4+0x34b>
+ DB 127,67 ; jg 4f6b <.literal4+0x34b>
DB 128,0,128 ; addb $0x80,(%rax)
DB 55 ; (bad)
DB 128,0,128 ; addb $0x80,(%rax)
@@ -4736,7 +4859,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 255 ; (bad)
- DB 127,71 ; jg 4da7 <.literal4+0x35f>
+ DB 127,71 ; jg 4f7f <.literal4+0x35f>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,191,0,0,0 ; add %al,0xbf(%rax)
@@ -4844,16 +4967,16 @@ ALIGN 32
DB 0,0 ; add %al,(%rax)
DB 1,255 ; add %edi,%edi
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004e88 <_sk_callback_hsw+0xa00056c>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005068 <_sk_callback_hsw+0xa000574>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004e90 <_sk_callback_hsw+0x12000574>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005070 <_sk_callback_hsw+0x1200057c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004e98 <_sk_callback_hsw+0x1a00057c>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005078 <_sk_callback_hsw+0x1a000584>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004ea0 <_sk_callback_hsw+0x3000584>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005080 <_sk_callback_hsw+0x300058c>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4896,16 +5019,16 @@ ALIGN 32
DB 0,0 ; add %al,(%rax)
DB 1,255 ; add %edi,%edi
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004ee8 <_sk_callback_hsw+0xa0005cc>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0050c8 <_sk_callback_hsw+0xa0005d4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004ef0 <_sk_callback_hsw+0x120005d4>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120050d0 <_sk_callback_hsw+0x120005dc>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004ef8 <_sk_callback_hsw+0x1a0005dc>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0050d8 <_sk_callback_hsw+0x1a0005e4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004f00 <_sk_callback_hsw+0x30005e4>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30050e0 <_sk_callback_hsw+0x30005ec>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4948,16 +5071,16 @@ ALIGN 32
DB 0,0 ; add %al,(%rax)
DB 1,255 ; add %edi,%edi
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004f48 <_sk_callback_hsw+0xa00062c>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005128 <_sk_callback_hsw+0xa000634>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004f50 <_sk_callback_hsw+0x12000634>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005130 <_sk_callback_hsw+0x1200063c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004f58 <_sk_callback_hsw+0x1a00063c>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005138 <_sk_callback_hsw+0x1a000644>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004f60 <_sk_callback_hsw+0x3000644>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005140 <_sk_callback_hsw+0x300064c>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -5000,16 +5123,16 @@ ALIGN 32
DB 0,0 ; add %al,(%rax)
DB 1,255 ; add %edi,%edi
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004fa8 <_sk_callback_hsw+0xa00068c>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005188 <_sk_callback_hsw+0xa000694>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004fb0 <_sk_callback_hsw+0x12000694>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005190 <_sk_callback_hsw+0x1200069c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004fb8 <_sk_callback_hsw+0x1a00069c>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005198 <_sk_callback_hsw+0x1a0006a4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004fc0 <_sk_callback_hsw+0x30006a4>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30051a0 <_sk_callback_hsw+0x30006ac>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -5052,16 +5175,16 @@ ALIGN 32
DB 0,0 ; add %al,(%rax)
DB 1,255 ; add %edi,%edi
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005008 <_sk_callback_hsw+0xa0006ec>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0051e8 <_sk_callback_hsw+0xa0006f4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005010 <_sk_callback_hsw+0x120006f4>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120051f0 <_sk_callback_hsw+0x120006fc>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005018 <_sk_callback_hsw+0x1a0006fc>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0051f8 <_sk_callback_hsw+0x1a000704>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005020 <_sk_callback_hsw+0x3000704>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005200 <_sk_callback_hsw+0x300070c>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -5104,6 +5227,38 @@ ALIGN 16
DB 255,0 ; incl (%rax)
DB 255,0 ; incl (%rax)
DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
ALIGN 32
PUBLIC _sk_start_pipeline_avx
@@ -5207,7 +5362,7 @@ _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 196,226,125,24,13,183,100,0,0 ; vbroadcastss 0x64b7(%rip),%ymm1 # 6630 <_sk_callback_avx+0x12c>
+ DB 196,226,125,24,13,139,102,0,0 ; vbroadcastss 0x668b(%rip),%ymm1 # 6804 <_sk_callback_avx+0x12c>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,88,7 ; vaddps (%rdi),%ymm0,%ymm0
DB 197,249,110,209 ; vmovd %ecx,%xmm2
@@ -5216,7 +5371,7 @@ _sk_seed_shader_avx LABEL PROC
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 196,226,125,24,21,145,100,0,0 ; vbroadcastss 0x6491(%rip),%ymm2 # 6634 <_sk_callback_avx+0x130>
+ DB 196,226,125,24,21,101,102,0,0 ; vbroadcastss 0x6665(%rip),%ymm2 # 6808 <_sk_callback_avx+0x130>
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
@@ -5238,7 +5393,7 @@ _sk_dither_avx LABEL PROC
DB 196,65,121,112,201,0 ; vpshufd $0x0,%xmm9,%xmm9
DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
DB 196,65,52,87,208 ; vxorps %ymm8,%ymm9,%ymm10
- DB 196,98,125,24,29,58,100,0,0 ; vbroadcastss 0x643a(%rip),%ymm11 # 6638 <_sk_callback_avx+0x134>
+ DB 196,98,125,24,29,14,102,0,0 ; vbroadcastss 0x660e(%rip),%ymm11 # 680c <_sk_callback_avx+0x134>
DB 196,65,44,84,203 ; vandps %ymm11,%ymm10,%ymm9
DB 196,193,25,114,241,5 ; vpslld $0x5,%xmm9,%xmm12
DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9
@@ -5249,8 +5404,8 @@ _sk_dither_avx LABEL PROC
DB 196,67,125,25,219,1 ; vextractf128 $0x1,%ymm11,%xmm11
DB 196,193,33,114,243,4 ; vpslld $0x4,%xmm11,%xmm11
DB 196,67,29,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm12,%ymm11
- DB 196,98,125,24,37,251,99,0,0 ; vbroadcastss 0x63fb(%rip),%ymm12 # 663c <_sk_callback_avx+0x138>
- DB 196,98,125,24,45,246,99,0,0 ; vbroadcastss 0x63f6(%rip),%ymm13 # 6640 <_sk_callback_avx+0x13c>
+ DB 196,98,125,24,37,207,101,0,0 ; vbroadcastss 0x65cf(%rip),%ymm12 # 6810 <_sk_callback_avx+0x138>
+ DB 196,98,125,24,45,202,101,0,0 ; vbroadcastss 0x65ca(%rip),%ymm13 # 6814 <_sk_callback_avx+0x13c>
DB 196,65,44,84,245 ; vandps %ymm13,%ymm10,%ymm14
DB 196,193,1,114,246,2 ; vpslld $0x2,%xmm14,%xmm15
DB 196,67,125,25,246,1 ; vextractf128 $0x1,%ymm14,%xmm14
@@ -5277,9 +5432,9 @@ _sk_dither_avx LABEL PROC
DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8
DB 196,65,60,86,194 ; vorps %ymm10,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,97,99,0,0 ; vbroadcastss 0x6361(%rip),%ymm9 # 6644 <_sk_callback_avx+0x140>
+ DB 196,98,125,24,13,53,101,0,0 ; vbroadcastss 0x6535(%rip),%ymm9 # 6818 <_sk_callback_avx+0x140>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,87,99,0,0 ; vbroadcastss 0x6357(%rip),%ymm9 # 6648 <_sk_callback_avx+0x144>
+ DB 196,98,125,24,13,43,101,0,0 ; vbroadcastss 0x652b(%rip),%ymm9 # 681c <_sk_callback_avx+0x144>
DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8
DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
@@ -5338,7 +5493,7 @@ _sk_clear_avx LABEL PROC
PUBLIC _sk_srcatop_avx
_sk_srcatop_avx LABEL PROC
DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0
- DB 196,98,125,24,5,175,98,0,0 ; vbroadcastss 0x62af(%rip),%ymm8 # 664c <_sk_callback_avx+0x148>
+ DB 196,98,125,24,5,131,100,0,0 ; vbroadcastss 0x6483(%rip),%ymm8 # 6820 <_sk_callback_avx+0x148>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9
DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0
@@ -5357,7 +5512,7 @@ _sk_srcatop_avx LABEL PROC
PUBLIC _sk_dstatop_avx
_sk_dstatop_avx LABEL PROC
DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8
- DB 196,98,125,24,13,113,98,0,0 ; vbroadcastss 0x6271(%rip),%ymm9 # 6650 <_sk_callback_avx+0x14c>
+ DB 196,98,125,24,13,69,100,0,0 ; vbroadcastss 0x6445(%rip),%ymm9 # 6824 <_sk_callback_avx+0x14c>
DB 197,52,92,207 ; vsubps %ymm7,%ymm9,%ymm9
DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0
DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0
@@ -5393,7 +5548,7 @@ _sk_dstin_avx LABEL PROC
PUBLIC _sk_srcout_avx
_sk_srcout_avx LABEL PROC
- DB 196,98,125,24,5,16,98,0,0 ; vbroadcastss 0x6210(%rip),%ymm8 # 6654 <_sk_callback_avx+0x150>
+ DB 196,98,125,24,5,228,99,0,0 ; vbroadcastss 0x63e4(%rip),%ymm8 # 6828 <_sk_callback_avx+0x150>
DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
@@ -5404,7 +5559,7 @@ _sk_srcout_avx LABEL PROC
PUBLIC _sk_dstout_avx
_sk_dstout_avx LABEL PROC
- DB 196,226,125,24,5,243,97,0,0 ; vbroadcastss 0x61f3(%rip),%ymm0 # 6658 <_sk_callback_avx+0x154>
+ DB 196,226,125,24,5,199,99,0,0 ; vbroadcastss 0x63c7(%rip),%ymm0 # 682c <_sk_callback_avx+0x154>
DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3
DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0
DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1
@@ -5415,7 +5570,7 @@ _sk_dstout_avx LABEL PROC
PUBLIC _sk_srcover_avx
_sk_srcover_avx LABEL PROC
- DB 196,98,125,24,5,214,97,0,0 ; vbroadcastss 0x61d6(%rip),%ymm8 # 665c <_sk_callback_avx+0x158>
+ DB 196,98,125,24,5,170,99,0,0 ; vbroadcastss 0x63aa(%rip),%ymm8 # 6830 <_sk_callback_avx+0x158>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9
DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0
@@ -5430,7 +5585,7 @@ _sk_srcover_avx LABEL PROC
PUBLIC _sk_dstover_avx
_sk_dstover_avx LABEL PROC
- DB 196,98,125,24,5,169,97,0,0 ; vbroadcastss 0x61a9(%rip),%ymm8 # 6660 <_sk_callback_avx+0x15c>
+ DB 196,98,125,24,5,125,99,0,0 ; vbroadcastss 0x637d(%rip),%ymm8 # 6834 <_sk_callback_avx+0x15c>
DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0
@@ -5454,7 +5609,7 @@ _sk_modulate_avx LABEL PROC
PUBLIC _sk_multiply_avx
_sk_multiply_avx LABEL PROC
- DB 196,98,125,24,5,104,97,0,0 ; vbroadcastss 0x6168(%rip),%ymm8 # 6664 <_sk_callback_avx+0x160>
+ DB 196,98,125,24,5,60,99,0,0 ; vbroadcastss 0x633c(%rip),%ymm8 # 6838 <_sk_callback_avx+0x160>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -5508,7 +5663,7 @@ _sk_screen_avx LABEL PROC
PUBLIC _sk_xor__avx
_sk_xor__avx LABEL PROC
- DB 196,98,125,24,5,183,96,0,0 ; vbroadcastss 0x60b7(%rip),%ymm8 # 6668 <_sk_callback_avx+0x164>
+ DB 196,98,125,24,5,139,98,0,0 ; vbroadcastss 0x628b(%rip),%ymm8 # 683c <_sk_callback_avx+0x164>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -5543,7 +5698,7 @@ _sk_darken_avx LABEL PROC
DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9
DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,55,96,0,0 ; vbroadcastss 0x6037(%rip),%ymm8 # 666c <_sk_callback_avx+0x168>
+ DB 196,98,125,24,5,11,98,0,0 ; vbroadcastss 0x620b(%rip),%ymm8 # 6840 <_sk_callback_avx+0x168>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8
DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3
@@ -5567,7 +5722,7 @@ _sk_lighten_avx LABEL PROC
DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9
DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,227,95,0,0 ; vbroadcastss 0x5fe3(%rip),%ymm8 # 6670 <_sk_callback_avx+0x16c>
+ DB 196,98,125,24,5,183,97,0,0 ; vbroadcastss 0x61b7(%rip),%ymm8 # 6844 <_sk_callback_avx+0x16c>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8
DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3
@@ -5594,7 +5749,7 @@ _sk_difference_avx LABEL PROC
DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2
DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,131,95,0,0 ; vbroadcastss 0x5f83(%rip),%ymm8 # 6674 <_sk_callback_avx+0x170>
+ DB 196,98,125,24,5,87,97,0,0 ; vbroadcastss 0x6157(%rip),%ymm8 # 6848 <_sk_callback_avx+0x170>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8
DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3
@@ -5615,7 +5770,7 @@ _sk_exclusion_avx LABEL PROC
DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2
DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2
DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2
- DB 196,98,125,24,5,62,95,0,0 ; vbroadcastss 0x5f3e(%rip),%ymm8 # 6678 <_sk_callback_avx+0x174>
+ DB 196,98,125,24,5,18,97,0,0 ; vbroadcastss 0x6112(%rip),%ymm8 # 684c <_sk_callback_avx+0x174>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8
DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3
@@ -5624,7 +5779,7 @@ _sk_exclusion_avx LABEL PROC
PUBLIC _sk_colorburn_avx
_sk_colorburn_avx LABEL PROC
- DB 196,98,125,24,5,41,95,0,0 ; vbroadcastss 0x5f29(%rip),%ymm8 # 667c <_sk_callback_avx+0x178>
+ DB 196,98,125,24,5,253,96,0,0 ; vbroadcastss 0x60fd(%rip),%ymm8 # 6850 <_sk_callback_avx+0x178>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11
DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10
@@ -5684,7 +5839,7 @@ _sk_colorburn_avx LABEL PROC
PUBLIC _sk_colordodge_avx
_sk_colordodge_avx LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
- DB 196,98,125,24,13,37,94,0,0 ; vbroadcastss 0x5e25(%rip),%ymm9 # 6680 <_sk_callback_avx+0x17c>
+ DB 196,98,125,24,13,249,95,0,0 ; vbroadcastss 0x5ff9(%rip),%ymm9 # 6854 <_sk_callback_avx+0x17c>
DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10
DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11
DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9
@@ -5739,7 +5894,7 @@ _sk_colordodge_avx LABEL PROC
PUBLIC _sk_hardlight_avx
_sk_hardlight_avx LABEL PROC
- DB 196,98,125,24,5,55,93,0,0 ; vbroadcastss 0x5d37(%rip),%ymm8 # 6684 <_sk_callback_avx+0x180>
+ DB 196,98,125,24,5,11,95,0,0 ; vbroadcastss 0x5f0b(%rip),%ymm8 # 6858 <_sk_callback_avx+0x180>
DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10
DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -5792,7 +5947,7 @@ _sk_hardlight_avx LABEL PROC
PUBLIC _sk_overlay_avx
_sk_overlay_avx LABEL PROC
- DB 196,98,125,24,5,96,92,0,0 ; vbroadcastss 0x5c60(%rip),%ymm8 # 6688 <_sk_callback_avx+0x184>
+ DB 196,98,125,24,5,52,94,0,0 ; vbroadcastss 0x5e34(%rip),%ymm8 # 685c <_sk_callback_avx+0x184>
DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10
DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -5857,10 +6012,10 @@ _sk_softlight_avx LABEL PROC
DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8
DB 196,65,60,89,216 ; vmulps %ymm8,%ymm8,%ymm11
DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8
- DB 196,98,125,24,29,83,91,0,0 ; vbroadcastss 0x5b53(%rip),%ymm11 # 6690 <_sk_callback_avx+0x18c>
+ DB 196,98,125,24,29,39,93,0,0 ; vbroadcastss 0x5d27(%rip),%ymm11 # 6864 <_sk_callback_avx+0x18c>
DB 196,65,28,88,235 ; vaddps %ymm11,%ymm12,%ymm13
DB 196,65,20,89,192 ; vmulps %ymm8,%ymm13,%ymm8
- DB 196,98,125,24,45,68,91,0,0 ; vbroadcastss 0x5b44(%rip),%ymm13 # 6694 <_sk_callback_avx+0x190>
+ DB 196,98,125,24,45,24,93,0,0 ; vbroadcastss 0x5d18(%rip),%ymm13 # 6868 <_sk_callback_avx+0x190>
DB 196,65,28,89,245 ; vmulps %ymm13,%ymm12,%ymm14
DB 196,65,12,88,192 ; vaddps %ymm8,%ymm14,%ymm8
DB 196,65,124,82,244 ; vrsqrtps %ymm12,%ymm14
@@ -5871,7 +6026,7 @@ _sk_softlight_avx LABEL PROC
DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15
DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14
DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15
- DB 196,98,125,24,5,2,91,0,0 ; vbroadcastss 0x5b02(%rip),%ymm8 # 668c <_sk_callback_avx+0x188>
+ DB 196,98,125,24,5,214,92,0,0 ; vbroadcastss 0x5cd6(%rip),%ymm8 # 6860 <_sk_callback_avx+0x188>
DB 196,65,60,92,228 ; vsubps %ymm12,%ymm8,%ymm12
DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0
DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12
@@ -5998,12 +6153,12 @@ _sk_hue_avx LABEL PROC
DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11
DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11
DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12
- DB 196,98,125,24,53,204,88,0,0 ; vbroadcastss 0x58cc(%rip),%ymm14 # 6698 <_sk_callback_avx+0x194>
+ DB 196,98,125,24,53,160,90,0,0 ; vbroadcastss 0x5aa0(%rip),%ymm14 # 686c <_sk_callback_avx+0x194>
DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11
- DB 196,98,125,24,61,194,88,0,0 ; vbroadcastss 0x58c2(%rip),%ymm15 # 669c <_sk_callback_avx+0x198>
+ DB 196,98,125,24,61,150,90,0,0 ; vbroadcastss 0x5a96(%rip),%ymm15 # 6870 <_sk_callback_avx+0x198>
DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13
DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11
- DB 196,226,125,24,5,179,88,0,0 ; vbroadcastss 0x58b3(%rip),%ymm0 # 66a0 <_sk_callback_avx+0x19c>
+ DB 196,226,125,24,5,135,90,0,0 ; vbroadcastss 0x5a87(%rip),%ymm0 # 6874 <_sk_callback_avx+0x19c>
DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13
DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11
DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13
@@ -6064,7 +6219,7 @@ _sk_hue_avx LABEL PROC
DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10
DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2
DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,140,87,0,0 ; vbroadcastss 0x578c(%rip),%ymm8 # 66a4 <_sk_callback_avx+0x1a0>
+ DB 196,98,125,24,5,96,89,0,0 ; vbroadcastss 0x5960(%rip),%ymm8 # 6878 <_sk_callback_avx+0x1a0>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -6121,12 +6276,12 @@ _sk_saturation_avx LABEL PROC
DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11
DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11
DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12
- DB 196,98,125,24,53,148,86,0,0 ; vbroadcastss 0x5694(%rip),%ymm14 # 66a8 <_sk_callback_avx+0x1a4>
+ DB 196,98,125,24,53,104,88,0,0 ; vbroadcastss 0x5868(%rip),%ymm14 # 687c <_sk_callback_avx+0x1a4>
DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11
- DB 196,98,125,24,61,138,86,0,0 ; vbroadcastss 0x568a(%rip),%ymm15 # 66ac <_sk_callback_avx+0x1a8>
+ DB 196,98,125,24,61,94,88,0,0 ; vbroadcastss 0x585e(%rip),%ymm15 # 6880 <_sk_callback_avx+0x1a8>
DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13
DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11
- DB 196,226,125,24,5,123,86,0,0 ; vbroadcastss 0x567b(%rip),%ymm0 # 66b0 <_sk_callback_avx+0x1ac>
+ DB 196,226,125,24,5,79,88,0,0 ; vbroadcastss 0x584f(%rip),%ymm0 # 6884 <_sk_callback_avx+0x1ac>
DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13
DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11
DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13
@@ -6187,7 +6342,7 @@ _sk_saturation_avx LABEL PROC
DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10
DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2
DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,84,85,0,0 ; vbroadcastss 0x5554(%rip),%ymm8 # 66b4 <_sk_callback_avx+0x1b0>
+ DB 196,98,125,24,5,40,87,0,0 ; vbroadcastss 0x5728(%rip),%ymm8 # 6888 <_sk_callback_avx+0x1b0>
DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9
DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
@@ -6216,12 +6371,12 @@ _sk_color_avx LABEL PROC
DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp)
DB 197,124,89,199 ; vmulps %ymm7,%ymm0,%ymm8
DB 197,116,89,207 ; vmulps %ymm7,%ymm1,%ymm9
- DB 196,98,125,24,45,228,84,0,0 ; vbroadcastss 0x54e4(%rip),%ymm13 # 66b8 <_sk_callback_avx+0x1b4>
+ DB 196,98,125,24,45,184,86,0,0 ; vbroadcastss 0x56b8(%rip),%ymm13 # 688c <_sk_callback_avx+0x1b4>
DB 196,65,92,89,213 ; vmulps %ymm13,%ymm4,%ymm10
- DB 196,98,125,24,53,218,84,0,0 ; vbroadcastss 0x54da(%rip),%ymm14 # 66bc <_sk_callback_avx+0x1b8>
+ DB 196,98,125,24,53,174,86,0,0 ; vbroadcastss 0x56ae(%rip),%ymm14 # 6890 <_sk_callback_avx+0x1b8>
DB 196,65,84,89,222 ; vmulps %ymm14,%ymm5,%ymm11
DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10
- DB 196,98,125,24,61,203,84,0,0 ; vbroadcastss 0x54cb(%rip),%ymm15 # 66c0 <_sk_callback_avx+0x1bc>
+ DB 196,98,125,24,61,159,86,0,0 ; vbroadcastss 0x569f(%rip),%ymm15 # 6894 <_sk_callback_avx+0x1bc>
DB 196,65,76,89,223 ; vmulps %ymm15,%ymm6,%ymm11
DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0
DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11
@@ -6284,7 +6439,7 @@ _sk_color_avx LABEL PROC
DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9
DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0
DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8
- DB 196,226,125,24,5,146,83,0,0 ; vbroadcastss 0x5392(%rip),%ymm0 # 66c4 <_sk_callback_avx+0x1c0>
+ DB 196,226,125,24,5,102,85,0,0 ; vbroadcastss 0x5566(%rip),%ymm0 # 6898 <_sk_callback_avx+0x1c0>
DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10
DB 197,172,89,84,36,32 ; vmulps 0x20(%rsp),%ymm10,%ymm2
DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11
@@ -6314,12 +6469,12 @@ _sk_luminosity_avx LABEL PROC
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8
DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9
- DB 196,98,125,24,45,30,83,0,0 ; vbroadcastss 0x531e(%rip),%ymm13 # 66c8 <_sk_callback_avx+0x1c4>
+ DB 196,98,125,24,45,242,84,0,0 ; vbroadcastss 0x54f2(%rip),%ymm13 # 689c <_sk_callback_avx+0x1c4>
DB 196,65,108,89,213 ; vmulps %ymm13,%ymm2,%ymm10
- DB 196,98,125,24,53,20,83,0,0 ; vbroadcastss 0x5314(%rip),%ymm14 # 66cc <_sk_callback_avx+0x1c8>
+ DB 196,98,125,24,53,232,84,0,0 ; vbroadcastss 0x54e8(%rip),%ymm14 # 68a0 <_sk_callback_avx+0x1c8>
DB 196,65,116,89,222 ; vmulps %ymm14,%ymm1,%ymm11
DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10
- DB 196,98,125,24,61,5,83,0,0 ; vbroadcastss 0x5305(%rip),%ymm15 # 66d0 <_sk_callback_avx+0x1cc>
+ DB 196,98,125,24,61,217,84,0,0 ; vbroadcastss 0x54d9(%rip),%ymm15 # 68a4 <_sk_callback_avx+0x1cc>
DB 196,65,28,89,223 ; vmulps %ymm15,%ymm12,%ymm11
DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0
DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11
@@ -6382,7 +6537,7 @@ _sk_luminosity_avx LABEL PROC
DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9
DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0
DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8
- DB 196,226,125,24,5,204,81,0,0 ; vbroadcastss 0x51cc(%rip),%ymm0 # 66d4 <_sk_callback_avx+0x1d0>
+ DB 196,226,125,24,5,160,83,0,0 ; vbroadcastss 0x53a0(%rip),%ymm0 # 68a8 <_sk_callback_avx+0x1d0>
DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11
@@ -6413,7 +6568,7 @@ _sk_srcover_rgba_8888_avx LABEL PROC
DB 77,133,192 ; test %r8,%r8
DB 15,133,47,1,0,0 ; jne 16a2 <_sk_srcover_rgba_8888_avx+0x14c>
DB 196,193,124,16,58 ; vmovups (%r10),%ymm7
- DB 197,124,40,13,128,85,0,0 ; vmovaps 0x5580(%rip),%ymm9 # 6b00 <_sk_callback_avx+0x5fc>
+ DB 197,124,40,13,160,87,0,0 ; vmovaps 0x57a0(%rip),%ymm9 # 6d20 <_sk_callback_avx+0x648>
DB 196,193,68,84,225 ; vandps %ymm9,%ymm7,%ymm4
DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4
DB 197,209,114,215,8 ; vpsrld $0x8,%xmm7,%xmm5
@@ -6431,9 +6586,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC
DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7
DB 196,227,53,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm9,%ymm7
DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7
- DB 196,98,125,24,5,247,80,0,0 ; vbroadcastss 0x50f7(%rip),%ymm8 # 66d8 <_sk_callback_avx+0x1d4>
+ DB 196,98,125,24,5,203,82,0,0 ; vbroadcastss 0x52cb(%rip),%ymm8 # 68ac <_sk_callback_avx+0x1d4>
DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8
- DB 196,98,125,24,13,238,80,0,0 ; vbroadcastss 0x50ee(%rip),%ymm9 # 66dc <_sk_callback_avx+0x1d8>
+ DB 196,98,125,24,13,194,82,0,0 ; vbroadcastss 0x52c2(%rip),%ymm9 # 68b0 <_sk_callback_avx+0x1d8>
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
DB 197,60,89,212 ; vmulps %ymm4,%ymm8,%ymm10
DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
@@ -6479,9 +6634,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC
DB 72,211,232 ; shr %cl,%rax
DB 196,225,249,110,224 ; vmovq %rax,%xmm4
DB 196,226,121,48,228 ; vpmovzxbw %xmm4,%xmm4
- DB 196,226,89,0,45,150,83,0,0 ; vpshufb 0x5396(%rip),%xmm4,%xmm5 # 6a60 <_sk_callback_avx+0x55c>
+ DB 196,226,89,0,45,118,85,0,0 ; vpshufb 0x5576(%rip),%xmm4,%xmm5 # 6c40 <_sk_callback_avx+0x568>
DB 196,226,121,33,237 ; vpmovsxbd %xmm5,%xmm5
- DB 196,226,89,0,37,152,83,0,0 ; vpshufb 0x5398(%rip),%xmm4,%xmm4 # 6a70 <_sk_callback_avx+0x56c>
+ DB 196,226,89,0,37,120,85,0,0 ; vpshufb 0x5578(%rip),%xmm4,%xmm4 # 6c50 <_sk_callback_avx+0x578>
DB 196,226,121,33,228 ; vpmovsxbd %xmm4,%xmm4
DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4
DB 196,194,93,44,58 ; vmaskmovps (%r10),%ymm4,%ymm7
@@ -6493,9 +6648,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC
DB 72,211,232 ; shr %cl,%rax
DB 196,97,249,110,200 ; vmovq %rax,%xmm9
DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9
- DB 196,98,49,0,21,75,83,0,0 ; vpshufb 0x534b(%rip),%xmm9,%xmm10 # 6a60 <_sk_callback_avx+0x55c>
+ DB 196,98,49,0,21,43,85,0,0 ; vpshufb 0x552b(%rip),%xmm9,%xmm10 # 6c40 <_sk_callback_avx+0x568>
DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10
- DB 196,98,49,0,13,77,83,0,0 ; vpshufb 0x534d(%rip),%xmm9,%xmm9 # 6a70 <_sk_callback_avx+0x56c>
+ DB 196,98,49,0,13,45,85,0,0 ; vpshufb 0x552d(%rip),%xmm9,%xmm9 # 6c50 <_sk_callback_avx+0x578>
DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9
DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10)
@@ -6513,7 +6668,7 @@ _sk_clamp_0_avx LABEL PROC
PUBLIC _sk_clamp_1_avx
_sk_clamp_1_avx LABEL PROC
- DB 196,98,125,24,5,130,79,0,0 ; vbroadcastss 0x4f82(%rip),%ymm8 # 66e0 <_sk_callback_avx+0x1dc>
+ DB 196,98,125,24,5,86,81,0,0 ; vbroadcastss 0x5156(%rip),%ymm8 # 68b4 <_sk_callback_avx+0x1dc>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
@@ -6523,7 +6678,7 @@ _sk_clamp_1_avx LABEL PROC
PUBLIC _sk_clamp_a_avx
_sk_clamp_a_avx LABEL PROC
- DB 196,98,125,24,5,101,79,0,0 ; vbroadcastss 0x4f65(%rip),%ymm8 # 66e4 <_sk_callback_avx+0x1e0>
+ DB 196,98,125,24,5,57,81,0,0 ; vbroadcastss 0x5139(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x1e0>
DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3
DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0
DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1
@@ -6595,7 +6750,7 @@ PUBLIC _sk_unpremul_avx
_sk_unpremul_avx LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9
- DB 196,98,125,24,21,173,78,0,0 ; vbroadcastss 0x4ead(%rip),%ymm10 # 66e8 <_sk_callback_avx+0x1e4>
+ DB 196,98,125,24,21,129,80,0,0 ; vbroadcastss 0x5081(%rip),%ymm10 # 68bc <_sk_callback_avx+0x1e4>
DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10
DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
@@ -6606,17 +6761,17 @@ _sk_unpremul_avx LABEL PROC
PUBLIC _sk_from_srgb_avx
_sk_from_srgb_avx LABEL PROC
- DB 196,98,125,24,5,142,78,0,0 ; vbroadcastss 0x4e8e(%rip),%ymm8 # 66ec <_sk_callback_avx+0x1e8>
+ DB 196,98,125,24,5,98,80,0,0 ; vbroadcastss 0x5062(%rip),%ymm8 # 68c0 <_sk_callback_avx+0x1e8>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10
- DB 196,98,125,24,29,128,78,0,0 ; vbroadcastss 0x4e80(%rip),%ymm11 # 66f0 <_sk_callback_avx+0x1ec>
+ DB 196,98,125,24,29,84,80,0,0 ; vbroadcastss 0x5054(%rip),%ymm11 # 68c4 <_sk_callback_avx+0x1ec>
DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12
- DB 196,98,125,24,45,118,78,0,0 ; vbroadcastss 0x4e76(%rip),%ymm13 # 66f4 <_sk_callback_avx+0x1f0>
+ DB 196,98,125,24,45,74,80,0,0 ; vbroadcastss 0x504a(%rip),%ymm13 # 68c8 <_sk_callback_avx+0x1f0>
DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12
DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,103,78,0,0 ; vbroadcastss 0x4e67(%rip),%ymm12 # 66f8 <_sk_callback_avx+0x1f4>
+ DB 196,98,125,24,37,59,80,0,0 ; vbroadcastss 0x503b(%rip),%ymm12 # 68cc <_sk_callback_avx+0x1f4>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,53,93,78,0,0 ; vbroadcastss 0x4e5d(%rip),%ymm14 # 66fc <_sk_callback_avx+0x1f8>
+ DB 196,98,125,24,53,49,80,0,0 ; vbroadcastss 0x5031(%rip),%ymm14 # 68d0 <_sk_callback_avx+0x1f8>
DB 196,193,124,194,198,1 ; vcmpltps %ymm14,%ymm0,%ymm0
DB 196,195,45,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm10,%ymm0
DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9
@@ -6641,20 +6796,20 @@ _sk_from_srgb_avx LABEL PROC
PUBLIC _sk_to_srgb_avx
_sk_to_srgb_avx LABEL PROC
DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9
- DB 196,98,125,24,5,242,77,0,0 ; vbroadcastss 0x4df2(%rip),%ymm8 # 6700 <_sk_callback_avx+0x1fc>
+ DB 196,98,125,24,5,198,79,0,0 ; vbroadcastss 0x4fc6(%rip),%ymm8 # 68d4 <_sk_callback_avx+0x1fc>
DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10
- DB 196,98,125,24,29,232,77,0,0 ; vbroadcastss 0x4de8(%rip),%ymm11 # 6704 <_sk_callback_avx+0x200>
+ DB 196,98,125,24,29,188,79,0,0 ; vbroadcastss 0x4fbc(%rip),%ymm11 # 68d8 <_sk_callback_avx+0x200>
DB 196,65,52,89,227 ; vmulps %ymm11,%ymm9,%ymm12
- DB 196,98,125,24,45,222,77,0,0 ; vbroadcastss 0x4dde(%rip),%ymm13 # 6708 <_sk_callback_avx+0x204>
+ DB 196,98,125,24,45,178,79,0,0 ; vbroadcastss 0x4fb2(%rip),%ymm13 # 68dc <_sk_callback_avx+0x204>
DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12
DB 196,65,52,89,228 ; vmulps %ymm12,%ymm9,%ymm12
- DB 196,98,125,24,53,207,77,0,0 ; vbroadcastss 0x4dcf(%rip),%ymm14 # 670c <_sk_callback_avx+0x208>
+ DB 196,98,125,24,53,163,79,0,0 ; vbroadcastss 0x4fa3(%rip),%ymm14 # 68e0 <_sk_callback_avx+0x208>
DB 196,65,28,88,230 ; vaddps %ymm14,%ymm12,%ymm12
- DB 196,98,125,24,61,197,77,0,0 ; vbroadcastss 0x4dc5(%rip),%ymm15 # 6710 <_sk_callback_avx+0x20c>
+ DB 196,98,125,24,61,153,79,0,0 ; vbroadcastss 0x4f99(%rip),%ymm15 # 68e4 <_sk_callback_avx+0x20c>
DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9
DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9
DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9
- DB 196,98,125,24,37,177,77,0,0 ; vbroadcastss 0x4db1(%rip),%ymm12 # 6714 <_sk_callback_avx+0x210>
+ DB 196,98,125,24,37,133,79,0,0 ; vbroadcastss 0x4f85(%rip),%ymm12 # 68e8 <_sk_callback_avx+0x210>
DB 196,193,124,194,196,1 ; vcmpltps %ymm12,%ymm0,%ymm0
DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0
DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9
@@ -6689,7 +6844,7 @@ _sk_rgb_to_hsl_avx LABEL PROC
DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9
DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9
DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,29,23,77,0,0 ; vbroadcastss 0x4d17(%rip),%ymm11 # 6718 <_sk_callback_avx+0x214>
+ DB 196,98,125,24,29,235,78,0,0 ; vbroadcastss 0x4eeb(%rip),%ymm11 # 68ec <_sk_callback_avx+0x214>
DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11
DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12
DB 196,65,28,89,227 ; vmulps %ymm11,%ymm12,%ymm12
@@ -6699,19 +6854,19 @@ _sk_rgb_to_hsl_avx LABEL PROC
DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2
DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1
DB 196,193,116,89,203 ; vmulps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,240,76,0,0 ; vbroadcastss 0x4cf0(%rip),%ymm11 # 6724 <_sk_callback_avx+0x220>
+ DB 196,98,125,24,29,196,78,0,0 ; vbroadcastss 0x4ec4(%rip),%ymm11 # 68f8 <_sk_callback_avx+0x220>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,222,76,0,0 ; vbroadcastss 0x4cde(%rip),%ymm11 # 6720 <_sk_callback_avx+0x21c>
+ DB 196,98,125,24,29,178,78,0,0 ; vbroadcastss 0x4eb2(%rip),%ymm11 # 68f4 <_sk_callback_avx+0x21c>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
DB 196,227,117,74,202,224 ; vblendvps %ymm14,%ymm2,%ymm1,%ymm1
- DB 196,226,125,24,21,198,76,0,0 ; vbroadcastss 0x4cc6(%rip),%ymm2 # 671c <_sk_callback_avx+0x218>
+ DB 196,226,125,24,21,154,78,0,0 ; vbroadcastss 0x4e9a(%rip),%ymm2 # 68f0 <_sk_callback_avx+0x218>
DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14
DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2
DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0
DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2
DB 196,227,117,74,194,0 ; vblendvps %ymm0,%ymm2,%ymm1,%ymm0
DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1
- DB 196,98,125,24,37,173,76,0,0 ; vbroadcastss 0x4cad(%rip),%ymm12 # 672c <_sk_callback_avx+0x228>
+ DB 196,98,125,24,37,129,78,0,0 ; vbroadcastss 0x4e81(%rip),%ymm12 # 6900 <_sk_callback_avx+0x228>
DB 196,193,116,89,212 ; vmulps %ymm12,%ymm1,%ymm2
DB 197,28,194,226,1 ; vcmpltps %ymm2,%ymm12,%ymm12
DB 196,65,36,92,216 ; vsubps %ymm8,%ymm11,%ymm11
@@ -6721,7 +6876,7 @@ _sk_rgb_to_hsl_avx LABEL PROC
DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1
DB 196,195,125,74,198,128 ; vblendvps %ymm8,%ymm14,%ymm0,%ymm0
DB 196,195,117,74,206,128 ; vblendvps %ymm8,%ymm14,%ymm1,%ymm1
- DB 196,98,125,24,5,112,76,0,0 ; vbroadcastss 0x4c70(%rip),%ymm8 # 6728 <_sk_callback_avx+0x224>
+ DB 196,98,125,24,5,68,78,0,0 ; vbroadcastss 0x4e44(%rip),%ymm8 # 68fc <_sk_callback_avx+0x224>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -6736,7 +6891,7 @@ _sk_hsl_to_rgb_avx LABEL PROC
DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp)
DB 197,252,40,225 ; vmovaps %ymm1,%ymm4
DB 197,252,40,216 ; vmovaps %ymm0,%ymm3
- DB 196,98,125,24,5,55,76,0,0 ; vbroadcastss 0x4c37(%rip),%ymm8 # 6730 <_sk_callback_avx+0x22c>
+ DB 196,98,125,24,5,11,78,0,0 ; vbroadcastss 0x4e0b(%rip),%ymm8 # 6904 <_sk_callback_avx+0x22c>
DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9
DB 197,92,89,210 ; vmulps %ymm2,%ymm4,%ymm10
DB 196,65,92,92,218 ; vsubps %ymm10,%ymm4,%ymm11
@@ -6744,23 +6899,23 @@ _sk_hsl_to_rgb_avx LABEL PROC
DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10
DB 197,108,88,202 ; vaddps %ymm2,%ymm2,%ymm9
DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9
- DB 196,98,125,24,29,17,76,0,0 ; vbroadcastss 0x4c11(%rip),%ymm11 # 6734 <_sk_callback_avx+0x230>
+ DB 196,98,125,24,29,229,77,0,0 ; vbroadcastss 0x4de5(%rip),%ymm11 # 6908 <_sk_callback_avx+0x230>
DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11
DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12
DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15
DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11
- DB 196,98,125,24,37,251,75,0,0 ; vbroadcastss 0x4bfb(%rip),%ymm12 # 673c <_sk_callback_avx+0x238>
+ DB 196,98,125,24,37,207,77,0,0 ; vbroadcastss 0x4dcf(%rip),%ymm12 # 6910 <_sk_callback_avx+0x238>
DB 196,193,4,89,196 ; vmulps %ymm12,%ymm15,%ymm0
- DB 196,98,125,24,45,241,75,0,0 ; vbroadcastss 0x4bf1(%rip),%ymm13 # 6740 <_sk_callback_avx+0x23c>
+ DB 196,98,125,24,45,197,77,0,0 ; vbroadcastss 0x4dc5(%rip),%ymm13 # 6914 <_sk_callback_avx+0x23c>
DB 197,20,92,240 ; vsubps %ymm0,%ymm13,%ymm14
DB 196,65,36,89,246 ; vmulps %ymm14,%ymm11,%ymm14
DB 196,65,52,88,246 ; vaddps %ymm14,%ymm9,%ymm14
- DB 196,226,125,24,13,210,75,0,0 ; vbroadcastss 0x4bd2(%rip),%ymm1 # 6738 <_sk_callback_avx+0x234>
+ DB 196,226,125,24,13,166,77,0,0 ; vbroadcastss 0x4da6(%rip),%ymm1 # 690c <_sk_callback_avx+0x234>
DB 196,193,116,194,255,2 ; vcmpleps %ymm15,%ymm1,%ymm7
DB 196,195,13,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm14,%ymm7
DB 196,65,60,194,247,2 ; vcmpleps %ymm15,%ymm8,%ymm14
DB 196,227,45,74,255,224 ; vblendvps %ymm14,%ymm7,%ymm10,%ymm7
- DB 196,98,125,24,53,189,75,0,0 ; vbroadcastss 0x4bbd(%rip),%ymm14 # 6744 <_sk_callback_avx+0x240>
+ DB 196,98,125,24,53,145,77,0,0 ; vbroadcastss 0x4d91(%rip),%ymm14 # 6918 <_sk_callback_avx+0x240>
DB 196,65,12,194,255,2 ; vcmpleps %ymm15,%ymm14,%ymm15
DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0
DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0
@@ -6779,7 +6934,7 @@ _sk_hsl_to_rgb_avx LABEL PROC
DB 197,164,89,247 ; vmulps %ymm7,%ymm11,%ymm6
DB 197,180,88,246 ; vaddps %ymm6,%ymm9,%ymm6
DB 196,227,77,74,237,0 ; vblendvps %ymm0,%ymm5,%ymm6,%ymm5
- DB 196,226,125,24,5,95,75,0,0 ; vbroadcastss 0x4b5f(%rip),%ymm0 # 6748 <_sk_callback_avx+0x244>
+ DB 196,226,125,24,5,51,77,0,0 ; vbroadcastss 0x4d33(%rip),%ymm0 # 691c <_sk_callback_avx+0x244>
DB 197,228,88,192 ; vaddps %ymm0,%ymm3,%ymm0
DB 196,227,125,8,216,1 ; vroundps $0x1,%ymm0,%ymm3
DB 197,252,92,195 ; vsubps %ymm3,%ymm0,%ymm0
@@ -6822,41 +6977,72 @@ _sk_scale_1_float_avx LABEL PROC
PUBLIC _sk_scale_u8_avx
_sk_scale_u8_avx LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,68 ; jne 1ce6 <_sk_scale_u8_avx+0x54>
- DB 197,122,126,0 ; vmovq (%rax),%xmm8
- DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9
- DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8
- DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8
- DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
+ DB 117,74 ; jne 1ce6 <_sk_scale_u8_avx+0x54>
+ DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8
+ DB 197,57,219,5,182,79,0,0 ; vpand 0x4fb6(%rip),%xmm8,%xmm8 # 6c60 <_sk_callback_avx+0x588>
+ DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
+ DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9
+ DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8
+ DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,130,74,0,0 ; vbroadcastss 0x4a82(%rip),%ymm9 # 674c <_sk_callback_avx+0x248>
+ DB 196,98,125,24,13,83,76,0,0 ; vbroadcastss 0x4c53(%rip),%ymm9 # 6920 <_sk_callback_avx+0x248>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 76,137,201 ; mov %r9,%rcx
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 1cef <_sk_scale_u8_avx+0x5d>
- DB 196,65,249,110,194 ; vmovq %r10,%xmm8
- DB 91 ; pop %rbx
- DB 235,154 ; jmp 1ca6 <_sk_scale_u8_avx+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,167 ; ja 1ca2 <_sk_scale_u8_avx+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,98,0,0,0 ; lea 0x62(%rip),%r10 # 1d68 <_sk_scale_u8_avx+0xd6>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 197,57,196,192,6 ; vpinsrw $0x6,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,57,196,192,5 ; vpinsrw $0x5,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,57,196,192,4 ; vpinsrw $0x4,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,57,196,192,3 ; vpinsrw $0x3,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,57,196,192,2 ; vpinsrw $0x2,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,57,196,192,1 ; vpinsrw $0x1,%eax,%xmm8,%xmm8
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,57,196,192,0 ; vpinsrw $0x0,%eax,%xmm8,%xmm8
+ DB 233,61,255,255,255 ; jmpq 1ca2 <_sk_scale_u8_avx+0x10>
+ DB 15,31,0 ; nopl (%rax)
+ DB 238 ; out %al,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,227 ; jmpq *%rbx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 216,255 ; fdivr %st(7),%st
+ DB 255 ; (bad)
+ DB 255,205 ; dec %ebp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,194 ; inc %edx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,183,255,255,255,167 ; pushq -0x58000001(%rdi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_lerp_1_float_avx
_sk_lerp_1_float_avx LABEL PROC
@@ -6879,19 +7065,18 @@ _sk_lerp_1_float_avx LABEL PROC
PUBLIC _sk_lerp_u8_avx
_sk_lerp_u8_avx LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,104 ; jne 1dc3 <_sk_lerp_u8_avx+0x78>
- DB 197,122,126,0 ; vmovq (%rax),%xmm8
- DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9
- DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8
- DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8
- DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
+ DB 117,110 ; jne 1e3b <_sk_lerp_u8_avx+0x78>
+ DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8
+ DB 197,57,219,5,149,78,0,0 ; vpand 0x4e95(%rip),%xmm8,%xmm8 # 6c70 <_sk_callback_avx+0x598>
+ DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
+ DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9
+ DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8
+ DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,205,73,0,0 ; vbroadcastss 0x49cd(%rip),%ymm9 # 6750 <_sk_callback_avx+0x24c>
+ DB 196,98,125,24,13,38,75,0,0 ; vbroadcastss 0x4b26(%rip),%ymm9 # 6924 <_sk_callback_avx+0x24c>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
@@ -6906,48 +7091,81 @@ _sk_lerp_u8_avx LABEL PROC
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 197,228,88,223 ; vaddps %ymm7,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 76,137,201 ; mov %r9,%rcx
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 1dcc <_sk_lerp_u8_avx+0x81>
- DB 196,65,249,110,194 ; vmovq %r10,%xmm8
- DB 91 ; pop %rbx
- DB 233,115,255,255,255 ; jmpq 1d5f <_sk_lerp_u8_avx+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,131 ; ja 1dd3 <_sk_lerp_u8_avx+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,97,0,0,0 ; lea 0x61(%rip),%r10 # 1ebc <_sk_lerp_u8_avx+0xf9>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
+ DB 197,57,196,192,6 ; vpinsrw $0x6,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,57,196,192,5 ; vpinsrw $0x5,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,57,196,192,4 ; vpinsrw $0x4,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,57,196,192,3 ; vpinsrw $0x3,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,57,196,192,2 ; vpinsrw $0x2,%eax,%xmm8,%xmm8
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,57,196,192,1 ; vpinsrw $0x1,%eax,%xmm8,%xmm8
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,57,196,192,0 ; vpinsrw $0x0,%eax,%xmm8,%xmm8
+ DB 233,25,255,255,255 ; jmpq 1dd3 <_sk_lerp_u8_avx+0x10>
+ DB 102,144 ; xchg %ax,%ax
+ DB 239 ; out %eax,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 217,255 ; fcos
+ DB 255 ; (bad)
+ DB 255,206 ; dec %esi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,195 ; inc %ebx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 184,255,255,255,168 ; mov $0xa8ffffff,%eax
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_lerp_565_avx
_sk_lerp_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 15,133,208,0,0,0 ; jne 1eca <_sk_lerp_565_avx+0xde>
+ DB 15,133,208,0,0,0 ; jne 1fb6 <_sk_lerp_565_avx+0xde>
DB 196,65,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9
DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8
DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8
- DB 196,98,125,24,13,54,73,0,0 ; vbroadcastss 0x4936(%rip),%ymm9 # 6754 <_sk_callback_avx+0x250>
+ DB 196,98,125,24,13,30,74,0,0 ; vbroadcastss 0x4a1e(%rip),%ymm9 # 6928 <_sk_callback_avx+0x250>
DB 196,65,60,84,201 ; vandps %ymm9,%ymm8,%ymm9
DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9
- DB 196,98,125,24,21,39,73,0,0 ; vbroadcastss 0x4927(%rip),%ymm10 # 6758 <_sk_callback_avx+0x254>
+ DB 196,98,125,24,21,15,74,0,0 ; vbroadcastss 0x4a0f(%rip),%ymm10 # 692c <_sk_callback_avx+0x254>
DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
- DB 196,98,125,24,21,29,73,0,0 ; vbroadcastss 0x491d(%rip),%ymm10 # 675c <_sk_callback_avx+0x258>
+ DB 196,98,125,24,21,5,74,0,0 ; vbroadcastss 0x4a05(%rip),%ymm10 # 6930 <_sk_callback_avx+0x258>
DB 196,65,60,84,210 ; vandps %ymm10,%ymm8,%ymm10
DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10
- DB 196,98,125,24,29,14,73,0,0 ; vbroadcastss 0x490e(%rip),%ymm11 # 6760 <_sk_callback_avx+0x25c>
+ DB 196,98,125,24,29,246,73,0,0 ; vbroadcastss 0x49f6(%rip),%ymm11 # 6934 <_sk_callback_avx+0x25c>
DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10
- DB 196,98,125,24,29,4,73,0,0 ; vbroadcastss 0x4904(%rip),%ymm11 # 6764 <_sk_callback_avx+0x260>
+ DB 196,98,125,24,29,236,73,0,0 ; vbroadcastss 0x49ec(%rip),%ymm11 # 6938 <_sk_callback_avx+0x260>
DB 196,65,60,84,195 ; vandps %ymm11,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,29,245,72,0,0 ; vbroadcastss 0x48f5(%rip),%ymm11 # 6768 <_sk_callback_avx+0x264>
+ DB 196,98,125,24,29,221,73,0,0 ; vbroadcastss 0x49dd(%rip),%ymm11 # 693c <_sk_callback_avx+0x264>
DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
@@ -6974,9 +7192,9 @@ _sk_lerp_565_avx LABEL PROC
DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 15,135,29,255,255,255 ; ja 1e00 <_sk_lerp_565_avx+0x14>
+ DB 15,135,29,255,255,255 ; ja 1eec <_sk_lerp_565_avx+0x14>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,74,0,0,0 ; lea 0x4a(%rip),%r10 # 1f38 <_sk_lerp_565_avx+0x14c>
+ DB 76,141,21,74,0,0,0 ; lea 0x4a(%rip),%r10 # 2024 <_sk_lerp_565_avx+0x14c>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -6988,7 +7206,7 @@ _sk_lerp_565_avx LABEL PROC
DB 196,65,57,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm8,%xmm8
DB 196,65,57,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm8,%xmm8
DB 196,65,57,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm8,%xmm8
- DB 233,200,254,255,255 ; jmpq 1e00 <_sk_lerp_565_avx+0x14>
+ DB 233,200,254,255,255 ; jmpq 1eec <_sk_lerp_565_avx+0x14>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -7024,9 +7242,9 @@ _sk_load_tables_avx LABEL PROC
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 76,3,16 ; add (%rax),%r10
DB 77,133,192 ; test %r8,%r8
- DB 15,133,251,1,0,0 ; jne 2172 <_sk_load_tables_avx+0x21e>
+ DB 15,133,251,1,0,0 ; jne 225e <_sk_load_tables_avx+0x21e>
DB 196,65,124,16,18 ; vmovups (%r10),%ymm10
- DB 197,124,40,13,156,75,0,0 ; vmovaps 0x4b9c(%rip),%ymm9 # 6b20 <_sk_callback_avx+0x61c>
+ DB 197,124,40,13,208,76,0,0 ; vmovaps 0x4cd0(%rip),%ymm9 # 6d40 <_sk_callback_avx+0x668>
DB 196,193,44,84,201 ; vandps %ymm9,%ymm10,%ymm1
DB 196,227,125,25,200,1 ; vextractf128 $0x1,%ymm1,%xmm0
DB 196,193,249,126,195 ; vmovq %xmm0,%r11
@@ -7118,7 +7336,7 @@ _sk_load_tables_avx LABEL PROC
DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7
DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,226,125,24,61,15,70,0,0 ; vbroadcastss 0x460f(%rip),%ymm7 # 676c <_sk_callback_avx+0x268>
+ DB 196,226,125,24,61,247,70,0,0 ; vbroadcastss 0x46f7(%rip),%ymm7 # 6940 <_sk_callback_avx+0x268>
DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,201 ; mov %r9,%rcx
@@ -7133,13 +7351,13 @@ _sk_load_tables_avx LABEL PROC
DB 73,211,235 ; shr %cl,%r11
DB 196,193,249,110,195 ; vmovq %r11,%xmm0
DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0
- DB 196,226,121,0,13,230,72,0,0 ; vpshufb 0x48e6(%rip),%xmm0,%xmm1 # 6a80 <_sk_callback_avx+0x57c>
+ DB 196,226,121,0,13,250,73,0,0 ; vpshufb 0x49fa(%rip),%xmm0,%xmm1 # 6c80 <_sk_callback_avx+0x5a8>
DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1
- DB 196,226,121,0,5,232,72,0,0 ; vpshufb 0x48e8(%rip),%xmm0,%xmm0 # 6a90 <_sk_callback_avx+0x58c>
+ DB 196,226,121,0,5,252,73,0,0 ; vpshufb 0x49fc(%rip),%xmm0,%xmm0 # 6c90 <_sk_callback_avx+0x5b8>
DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 196,66,125,44,18 ; vmaskmovps (%r10),%ymm0,%ymm10
- DB 233,191,253,255,255 ; jmpq 1f7c <_sk_load_tables_avx+0x28>
+ DB 233,191,253,255,255 ; jmpq 2068 <_sk_load_tables_avx+0x28>
PUBLIC _sk_load_tables_u16_be_avx
_sk_load_tables_u16_be_avx LABEL PROC
@@ -7149,7 +7367,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 77,133,192 ; test %r8,%r8
DB 197,252,17,60,36 ; vmovups %ymm7,(%rsp)
- DB 15,133,87,2,0,0 ; jne 2433 <_sk_load_tables_u16_be_avx+0x276>
+ DB 15,133,87,2,0,0 ; jne 251f <_sk_load_tables_u16_be_avx+0x276>
DB 196,1,121,16,4,81 ; vmovupd (%r9,%r10,2),%xmm8
DB 196,129,121,16,84,81,16 ; vmovupd 0x10(%r9,%r10,2),%xmm2
DB 196,129,121,16,92,81,32 ; vmovupd 0x20(%r9,%r10,2),%xmm3
@@ -7164,7 +7382,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 197,113,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm11
DB 197,177,108,200 ; vpunpcklqdq %xmm0,%xmm9,%xmm1
DB 197,49,109,224 ; vpunpckhqdq %xmm0,%xmm9,%xmm12
- DB 197,121,111,21,119,72,0,0 ; vmovdqa 0x4877(%rip),%xmm10 # 6aa0 <_sk_callback_avx+0x59c>
+ DB 197,121,111,21,139,73,0,0 ; vmovdqa 0x498b(%rip),%xmm10 # 6ca0 <_sk_callback_avx+0x5c8>
DB 196,193,113,219,202 ; vpand %xmm10,%xmm1,%xmm1
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2
@@ -7262,7 +7480,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3
DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,226,125,24,61,78,67,0,0 ; vbroadcastss 0x434e(%rip),%ymm7 # 6770 <_sk_callback_avx+0x26c>
+ DB 196,226,125,24,61,54,68,0,0 ; vbroadcastss 0x4436(%rip),%ymm7 # 6944 <_sk_callback_avx+0x26c>
DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7
@@ -7271,29 +7489,29 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 196,1,123,16,4,81 ; vmovsd (%r9,%r10,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,85 ; je 2499 <_sk_load_tables_u16_be_avx+0x2dc>
+ DB 116,85 ; je 2585 <_sk_load_tables_u16_be_avx+0x2dc>
DB 196,1,57,22,68,81,8 ; vmovhpd 0x8(%r9,%r10,2),%xmm8,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,72 ; jb 2499 <_sk_load_tables_u16_be_avx+0x2dc>
+ DB 114,72 ; jb 2585 <_sk_load_tables_u16_be_avx+0x2dc>
DB 196,129,123,16,84,81,16 ; vmovsd 0x10(%r9,%r10,2),%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 116,72 ; je 24a6 <_sk_load_tables_u16_be_avx+0x2e9>
+ DB 116,72 ; je 2592 <_sk_load_tables_u16_be_avx+0x2e9>
DB 196,129,105,22,84,81,24 ; vmovhpd 0x18(%r9,%r10,2),%xmm2,%xmm2
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,59 ; jb 24a6 <_sk_load_tables_u16_be_avx+0x2e9>
+ DB 114,59 ; jb 2592 <_sk_load_tables_u16_be_avx+0x2e9>
DB 196,129,123,16,92,81,32 ; vmovsd 0x20(%r9,%r10,2),%xmm3
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 15,132,123,253,255,255 ; je 21f7 <_sk_load_tables_u16_be_avx+0x3a>
+ DB 15,132,123,253,255,255 ; je 22e3 <_sk_load_tables_u16_be_avx+0x3a>
DB 196,129,97,22,92,81,40 ; vmovhpd 0x28(%r9,%r10,2),%xmm3,%xmm3
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 15,130,106,253,255,255 ; jb 21f7 <_sk_load_tables_u16_be_avx+0x3a>
+ DB 15,130,106,253,255,255 ; jb 22e3 <_sk_load_tables_u16_be_avx+0x3a>
DB 196,1,122,126,76,81,48 ; vmovq 0x30(%r9,%r10,2),%xmm9
- DB 233,94,253,255,255 ; jmpq 21f7 <_sk_load_tables_u16_be_avx+0x3a>
+ DB 233,94,253,255,255 ; jmpq 22e3 <_sk_load_tables_u16_be_avx+0x3a>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,81,253,255,255 ; jmpq 21f7 <_sk_load_tables_u16_be_avx+0x3a>
+ DB 233,81,253,255,255 ; jmpq 22e3 <_sk_load_tables_u16_be_avx+0x3a>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,72,253,255,255 ; jmpq 21f7 <_sk_load_tables_u16_be_avx+0x3a>
+ DB 233,72,253,255,255 ; jmpq 22e3 <_sk_load_tables_u16_be_avx+0x3a>
PUBLIC _sk_load_tables_rgb_u16_be_avx
_sk_load_tables_rgb_u16_be_avx LABEL PROC
@@ -7304,7 +7522,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 77,133,192 ; test %r8,%r8
DB 197,252,17,124,36,32 ; vmovups %ymm7,0x20(%rsp)
DB 197,252,17,52,36 ; vmovups %ymm6,(%rsp)
- DB 15,133,74,2,0,0 ; jne 271a <_sk_load_tables_rgb_u16_be_avx+0x26b>
+ DB 15,133,74,2,0,0 ; jne 2806 <_sk_load_tables_rgb_u16_be_avx+0x26b>
DB 196,129,122,111,4,81 ; vmovdqu (%r9,%r10,2),%xmm0
DB 196,129,122,111,84,81,12 ; vmovdqu 0xc(%r9,%r10,2),%xmm2
DB 196,129,122,111,76,81,24 ; vmovdqu 0x18(%r9,%r10,2),%xmm1
@@ -7325,7 +7543,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3
DB 197,57,109,218 ; vpunpckhqdq %xmm2,%xmm8,%xmm11
DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8
- DB 197,121,111,13,118,69,0,0 ; vmovdqa 0x4576(%rip),%xmm9 # 6ab0 <_sk_callback_avx+0x5ac>
+ DB 197,121,111,13,138,70,0,0 ; vmovdqa 0x468a(%rip),%xmm9 # 6cb0 <_sk_callback_avx+0x5d8>
DB 196,193,97,219,193 ; vpand %xmm9,%xmm3,%xmm0
DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10
DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1
@@ -7415,7 +7633,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 196,195,105,33,211,48 ; vinsertps $0x30,%xmm11,%xmm2,%xmm2
DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,107,64,0,0 ; vbroadcastss 0x406b(%rip),%ymm3 # 6774 <_sk_callback_avx+0x270>
+ DB 196,226,125,24,29,83,65,0,0 ; vbroadcastss 0x4153(%rip),%ymm3 # 6948 <_sk_callback_avx+0x270>
DB 197,252,16,52,36 ; vmovups (%rsp),%ymm6
DB 197,252,16,124,36,32 ; vmovups 0x20(%rsp),%ymm7
DB 72,131,196,88 ; add $0x58,%rsp
@@ -7423,41 +7641,41 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 196,129,121,110,4,81 ; vmovd (%r9,%r10,2),%xmm0
DB 196,129,121,196,68,81,4,2 ; vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 117,5 ; jne 2733 <_sk_load_tables_rgb_u16_be_avx+0x284>
- DB 233,209,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 117,5 ; jne 281f <_sk_load_tables_rgb_u16_be_avx+0x284>
+ DB 233,209,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
DB 196,129,121,110,76,81,6 ; vmovd 0x6(%r9,%r10,2),%xmm1
DB 196,1,113,196,68,81,10,2 ; vpinsrw $0x2,0xa(%r9,%r10,2),%xmm1,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,26 ; jb 2762 <_sk_load_tables_rgb_u16_be_avx+0x2b3>
+ DB 114,26 ; jb 284e <_sk_load_tables_rgb_u16_be_avx+0x2b3>
DB 196,129,121,110,76,81,12 ; vmovd 0xc(%r9,%r10,2),%xmm1
DB 196,129,113,196,84,81,16,2 ; vpinsrw $0x2,0x10(%r9,%r10,2),%xmm1,%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 117,10 ; jne 2767 <_sk_load_tables_rgb_u16_be_avx+0x2b8>
- DB 233,162,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
- DB 233,157,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 117,10 ; jne 2853 <_sk_load_tables_rgb_u16_be_avx+0x2b8>
+ DB 233,162,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 233,157,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
DB 196,129,121,110,76,81,18 ; vmovd 0x12(%r9,%r10,2),%xmm1
DB 196,1,113,196,76,81,22,2 ; vpinsrw $0x2,0x16(%r9,%r10,2),%xmm1,%xmm9
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,26 ; jb 2796 <_sk_load_tables_rgb_u16_be_avx+0x2e7>
+ DB 114,26 ; jb 2882 <_sk_load_tables_rgb_u16_be_avx+0x2e7>
DB 196,129,121,110,76,81,24 ; vmovd 0x18(%r9,%r10,2),%xmm1
DB 196,129,113,196,76,81,28,2 ; vpinsrw $0x2,0x1c(%r9,%r10,2),%xmm1,%xmm1
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 117,10 ; jne 279b <_sk_load_tables_rgb_u16_be_avx+0x2ec>
- DB 233,110,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
- DB 233,105,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 117,10 ; jne 2887 <_sk_load_tables_rgb_u16_be_avx+0x2ec>
+ DB 233,110,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 233,105,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
DB 196,129,121,110,92,81,30 ; vmovd 0x1e(%r9,%r10,2),%xmm3
DB 196,1,97,196,92,81,34,2 ; vpinsrw $0x2,0x22(%r9,%r10,2),%xmm3,%xmm11
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,20 ; jb 27c4 <_sk_load_tables_rgb_u16_be_avx+0x315>
+ DB 114,20 ; jb 28b0 <_sk_load_tables_rgb_u16_be_avx+0x315>
DB 196,129,121,110,92,81,36 ; vmovd 0x24(%r9,%r10,2),%xmm3
DB 196,129,97,196,92,81,40,2 ; vpinsrw $0x2,0x28(%r9,%r10,2),%xmm3,%xmm3
- DB 233,64,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
- DB 233,59,253,255,255 ; jmpq 2504 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 233,64,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
+ DB 233,59,253,255,255 ; jmpq 25f0 <_sk_load_tables_rgb_u16_be_avx+0x55>
PUBLIC _sk_byte_tables_avx
_sk_byte_tables_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,164,63,0,0 ; vbroadcastss 0x3fa4(%rip),%ymm8 # 6778 <_sk_callback_avx+0x274>
+ DB 196,98,125,24,5,140,64,0,0 ; vbroadcastss 0x408c(%rip),%ymm8 # 694c <_sk_callback_avx+0x274>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9
DB 196,65,249,126,201 ; vmovq %xmm9,%r9
@@ -7576,7 +7794,7 @@ _sk_byte_tables_avx LABEL PROC
DB 196,194,121,49,204 ; vpmovzxbd %xmm12,%xmm1
DB 196,194,121,49,213 ; vpmovzxbd %xmm13,%xmm2
DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
- DB 196,98,125,24,13,78,61,0,0 ; vbroadcastss 0x3d4e(%rip),%ymm9 # 677c <_sk_callback_avx+0x278>
+ DB 196,98,125,24,13,54,62,0,0 ; vbroadcastss 0x3e36(%rip),%ymm9 # 6950 <_sk_callback_avx+0x278>
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1
@@ -7690,7 +7908,7 @@ _sk_byte_tables_rgb_avx LABEL PROC
DB 196,194,121,49,203 ; vpmovzxbd %xmm11,%xmm1
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,13,47,59,0,0 ; vbroadcastss 0x3b2f(%rip),%ymm9 # 6780 <_sk_callback_avx+0x27c>
+ DB 196,98,125,24,13,23,60,0,0 ; vbroadcastss 0x3c17(%rip),%ymm9 # 6954 <_sk_callback_avx+0x27c>
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
DB 196,194,121,49,202 ; vpmovzxbd %xmm10,%xmm1
DB 196,194,121,49,212 ; vpmovzxbd %xmm12,%xmm2
@@ -7885,36 +8103,36 @@ _sk_parametric_r_avx LABEL PROC
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,216 ; vcvtdq2ps %ymm0,%ymm11
- DB 196,98,125,24,37,220,55,0,0 ; vbroadcastss 0x37dc(%rip),%ymm12 # 6784 <_sk_callback_avx+0x280>
+ DB 196,98,125,24,37,196,56,0,0 ; vbroadcastss 0x38c4(%rip),%ymm12 # 6958 <_sk_callback_avx+0x280>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,210,55,0,0 ; vbroadcastss 0x37d2(%rip),%ymm12 # 6788 <_sk_callback_avx+0x284>
+ DB 196,98,125,24,37,186,56,0,0 ; vbroadcastss 0x38ba(%rip),%ymm12 # 695c <_sk_callback_avx+0x284>
DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,200,55,0,0 ; vbroadcastss 0x37c8(%rip),%ymm12 # 678c <_sk_callback_avx+0x288>
+ DB 196,98,125,24,37,176,56,0,0 ; vbroadcastss 0x38b0(%rip),%ymm12 # 6960 <_sk_callback_avx+0x288>
DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,190,55,0,0 ; vbroadcastss 0x37be(%rip),%ymm12 # 6790 <_sk_callback_avx+0x28c>
+ DB 196,98,125,24,37,166,56,0,0 ; vbroadcastss 0x38a6(%rip),%ymm12 # 6964 <_sk_callback_avx+0x28c>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,180,55,0,0 ; vbroadcastss 0x37b4(%rip),%ymm12 # 6794 <_sk_callback_avx+0x290>
+ DB 196,98,125,24,37,156,56,0,0 ; vbroadcastss 0x389c(%rip),%ymm12 # 6968 <_sk_callback_avx+0x290>
DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,165,55,0,0 ; vbroadcastss 0x37a5(%rip),%ymm12 # 6798 <_sk_callback_avx+0x294>
+ DB 196,98,125,24,37,141,56,0,0 ; vbroadcastss 0x388d(%rip),%ymm12 # 696c <_sk_callback_avx+0x294>
DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,155,55,0,0 ; vbroadcastss 0x379b(%rip),%ymm12 # 679c <_sk_callback_avx+0x298>
+ DB 196,98,125,24,37,131,56,0,0 ; vbroadcastss 0x3883(%rip),%ymm12 # 6970 <_sk_callback_avx+0x298>
DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0
DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0
DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10
DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10
- DB 196,98,125,24,29,127,55,0,0 ; vbroadcastss 0x377f(%rip),%ymm11 # 67a0 <_sk_callback_avx+0x29c>
+ DB 196,98,125,24,29,103,56,0,0 ; vbroadcastss 0x3867(%rip),%ymm11 # 6974 <_sk_callback_avx+0x29c>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,117,55,0,0 ; vbroadcastss 0x3775(%rip),%ymm11 # 67a4 <_sk_callback_avx+0x2a0>
+ DB 196,98,125,24,29,93,56,0,0 ; vbroadcastss 0x385d(%rip),%ymm11 # 6978 <_sk_callback_avx+0x2a0>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,102,55,0,0 ; vbroadcastss 0x3766(%rip),%ymm11 # 67a8 <_sk_callback_avx+0x2a4>
+ DB 196,98,125,24,29,78,56,0,0 ; vbroadcastss 0x384e(%rip),%ymm11 # 697c <_sk_callback_avx+0x2a4>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,92,55,0,0 ; vbroadcastss 0x375c(%rip),%ymm11 # 67ac <_sk_callback_avx+0x2a8>
+ DB 196,98,125,24,29,68,56,0,0 ; vbroadcastss 0x3844(%rip),%ymm11 # 6980 <_sk_callback_avx+0x2a8>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,77,55,0,0 ; vbroadcastss 0x374d(%rip),%ymm10 # 67b0 <_sk_callback_avx+0x2ac>
+ DB 196,98,125,24,21,53,56,0,0 ; vbroadcastss 0x3835(%rip),%ymm10 # 6984 <_sk_callback_avx+0x2ac>
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
@@ -7922,7 +8140,7 @@ _sk_parametric_r_avx LABEL PROC
DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,36,55,0,0 ; vbroadcastss 0x3724(%rip),%ymm8 # 67b4 <_sk_callback_avx+0x2b0>
+ DB 196,98,125,24,5,12,56,0,0 ; vbroadcastss 0x380c(%rip),%ymm8 # 6988 <_sk_callback_avx+0x2b0>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7942,36 +8160,36 @@ _sk_parametric_g_avx LABEL PROC
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,217 ; vcvtdq2ps %ymm1,%ymm11
- DB 196,98,125,24,37,213,54,0,0 ; vbroadcastss 0x36d5(%rip),%ymm12 # 67b8 <_sk_callback_avx+0x2b4>
+ DB 196,98,125,24,37,189,55,0,0 ; vbroadcastss 0x37bd(%rip),%ymm12 # 698c <_sk_callback_avx+0x2b4>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,203,54,0,0 ; vbroadcastss 0x36cb(%rip),%ymm12 # 67bc <_sk_callback_avx+0x2b8>
+ DB 196,98,125,24,37,179,55,0,0 ; vbroadcastss 0x37b3(%rip),%ymm12 # 6990 <_sk_callback_avx+0x2b8>
DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,193,54,0,0 ; vbroadcastss 0x36c1(%rip),%ymm12 # 67c0 <_sk_callback_avx+0x2bc>
+ DB 196,98,125,24,37,169,55,0,0 ; vbroadcastss 0x37a9(%rip),%ymm12 # 6994 <_sk_callback_avx+0x2bc>
DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,183,54,0,0 ; vbroadcastss 0x36b7(%rip),%ymm12 # 67c4 <_sk_callback_avx+0x2c0>
+ DB 196,98,125,24,37,159,55,0,0 ; vbroadcastss 0x379f(%rip),%ymm12 # 6998 <_sk_callback_avx+0x2c0>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,173,54,0,0 ; vbroadcastss 0x36ad(%rip),%ymm12 # 67c8 <_sk_callback_avx+0x2c4>
+ DB 196,98,125,24,37,149,55,0,0 ; vbroadcastss 0x3795(%rip),%ymm12 # 699c <_sk_callback_avx+0x2c4>
DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,158,54,0,0 ; vbroadcastss 0x369e(%rip),%ymm12 # 67cc <_sk_callback_avx+0x2c8>
+ DB 196,98,125,24,37,134,55,0,0 ; vbroadcastss 0x3786(%rip),%ymm12 # 69a0 <_sk_callback_avx+0x2c8>
DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,148,54,0,0 ; vbroadcastss 0x3694(%rip),%ymm12 # 67d0 <_sk_callback_avx+0x2cc>
+ DB 196,98,125,24,37,124,55,0,0 ; vbroadcastss 0x377c(%rip),%ymm12 # 69a4 <_sk_callback_avx+0x2cc>
DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1
DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1
DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10
DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
- DB 196,98,125,24,29,120,54,0,0 ; vbroadcastss 0x3678(%rip),%ymm11 # 67d4 <_sk_callback_avx+0x2d0>
+ DB 196,98,125,24,29,96,55,0,0 ; vbroadcastss 0x3760(%rip),%ymm11 # 69a8 <_sk_callback_avx+0x2d0>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,110,54,0,0 ; vbroadcastss 0x366e(%rip),%ymm11 # 67d8 <_sk_callback_avx+0x2d4>
+ DB 196,98,125,24,29,86,55,0,0 ; vbroadcastss 0x3756(%rip),%ymm11 # 69ac <_sk_callback_avx+0x2d4>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,95,54,0,0 ; vbroadcastss 0x365f(%rip),%ymm11 # 67dc <_sk_callback_avx+0x2d8>
+ DB 196,98,125,24,29,71,55,0,0 ; vbroadcastss 0x3747(%rip),%ymm11 # 69b0 <_sk_callback_avx+0x2d8>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,85,54,0,0 ; vbroadcastss 0x3655(%rip),%ymm11 # 67e0 <_sk_callback_avx+0x2dc>
+ DB 196,98,125,24,29,61,55,0,0 ; vbroadcastss 0x373d(%rip),%ymm11 # 69b4 <_sk_callback_avx+0x2dc>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,70,54,0,0 ; vbroadcastss 0x3646(%rip),%ymm10 # 67e4 <_sk_callback_avx+0x2e0>
+ DB 196,98,125,24,21,46,55,0,0 ; vbroadcastss 0x372e(%rip),%ymm10 # 69b8 <_sk_callback_avx+0x2e0>
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
@@ -7979,7 +8197,7 @@ _sk_parametric_g_avx LABEL PROC
DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1
- DB 196,98,125,24,5,29,54,0,0 ; vbroadcastss 0x361d(%rip),%ymm8 # 67e8 <_sk_callback_avx+0x2e4>
+ DB 196,98,125,24,5,5,55,0,0 ; vbroadcastss 0x3705(%rip),%ymm8 # 69bc <_sk_callback_avx+0x2e4>
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7999,36 +8217,36 @@ _sk_parametric_b_avx LABEL PROC
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,218 ; vcvtdq2ps %ymm2,%ymm11
- DB 196,98,125,24,37,206,53,0,0 ; vbroadcastss 0x35ce(%rip),%ymm12 # 67ec <_sk_callback_avx+0x2e8>
+ DB 196,98,125,24,37,182,54,0,0 ; vbroadcastss 0x36b6(%rip),%ymm12 # 69c0 <_sk_callback_avx+0x2e8>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,196,53,0,0 ; vbroadcastss 0x35c4(%rip),%ymm12 # 67f0 <_sk_callback_avx+0x2ec>
+ DB 196,98,125,24,37,172,54,0,0 ; vbroadcastss 0x36ac(%rip),%ymm12 # 69c4 <_sk_callback_avx+0x2ec>
DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,186,53,0,0 ; vbroadcastss 0x35ba(%rip),%ymm12 # 67f4 <_sk_callback_avx+0x2f0>
+ DB 196,98,125,24,37,162,54,0,0 ; vbroadcastss 0x36a2(%rip),%ymm12 # 69c8 <_sk_callback_avx+0x2f0>
DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,176,53,0,0 ; vbroadcastss 0x35b0(%rip),%ymm12 # 67f8 <_sk_callback_avx+0x2f4>
+ DB 196,98,125,24,37,152,54,0,0 ; vbroadcastss 0x3698(%rip),%ymm12 # 69cc <_sk_callback_avx+0x2f4>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,166,53,0,0 ; vbroadcastss 0x35a6(%rip),%ymm12 # 67fc <_sk_callback_avx+0x2f8>
+ DB 196,98,125,24,37,142,54,0,0 ; vbroadcastss 0x368e(%rip),%ymm12 # 69d0 <_sk_callback_avx+0x2f8>
DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,151,53,0,0 ; vbroadcastss 0x3597(%rip),%ymm12 # 6800 <_sk_callback_avx+0x2fc>
+ DB 196,98,125,24,37,127,54,0,0 ; vbroadcastss 0x367f(%rip),%ymm12 # 69d4 <_sk_callback_avx+0x2fc>
DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,141,53,0,0 ; vbroadcastss 0x358d(%rip),%ymm12 # 6804 <_sk_callback_avx+0x300>
+ DB 196,98,125,24,37,117,54,0,0 ; vbroadcastss 0x3675(%rip),%ymm12 # 69d8 <_sk_callback_avx+0x300>
DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2
DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10
DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10
- DB 196,98,125,24,29,113,53,0,0 ; vbroadcastss 0x3571(%rip),%ymm11 # 6808 <_sk_callback_avx+0x304>
+ DB 196,98,125,24,29,89,54,0,0 ; vbroadcastss 0x3659(%rip),%ymm11 # 69dc <_sk_callback_avx+0x304>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,103,53,0,0 ; vbroadcastss 0x3567(%rip),%ymm11 # 680c <_sk_callback_avx+0x308>
+ DB 196,98,125,24,29,79,54,0,0 ; vbroadcastss 0x364f(%rip),%ymm11 # 69e0 <_sk_callback_avx+0x308>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,88,53,0,0 ; vbroadcastss 0x3558(%rip),%ymm11 # 6810 <_sk_callback_avx+0x30c>
+ DB 196,98,125,24,29,64,54,0,0 ; vbroadcastss 0x3640(%rip),%ymm11 # 69e4 <_sk_callback_avx+0x30c>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,78,53,0,0 ; vbroadcastss 0x354e(%rip),%ymm11 # 6814 <_sk_callback_avx+0x310>
+ DB 196,98,125,24,29,54,54,0,0 ; vbroadcastss 0x3636(%rip),%ymm11 # 69e8 <_sk_callback_avx+0x310>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,63,53,0,0 ; vbroadcastss 0x353f(%rip),%ymm10 # 6818 <_sk_callback_avx+0x314>
+ DB 196,98,125,24,21,39,54,0,0 ; vbroadcastss 0x3627(%rip),%ymm10 # 69ec <_sk_callback_avx+0x314>
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
@@ -8036,7 +8254,7 @@ _sk_parametric_b_avx LABEL PROC
DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,22,53,0,0 ; vbroadcastss 0x3516(%rip),%ymm8 # 681c <_sk_callback_avx+0x318>
+ DB 196,98,125,24,5,254,53,0,0 ; vbroadcastss 0x35fe(%rip),%ymm8 # 69f0 <_sk_callback_avx+0x318>
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -8056,36 +8274,36 @@ _sk_parametric_a_avx LABEL PROC
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,219 ; vcvtdq2ps %ymm3,%ymm11
- DB 196,98,125,24,37,199,52,0,0 ; vbroadcastss 0x34c7(%rip),%ymm12 # 6820 <_sk_callback_avx+0x31c>
+ DB 196,98,125,24,37,175,53,0,0 ; vbroadcastss 0x35af(%rip),%ymm12 # 69f4 <_sk_callback_avx+0x31c>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,189,52,0,0 ; vbroadcastss 0x34bd(%rip),%ymm12 # 6824 <_sk_callback_avx+0x320>
+ DB 196,98,125,24,37,165,53,0,0 ; vbroadcastss 0x35a5(%rip),%ymm12 # 69f8 <_sk_callback_avx+0x320>
DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,179,52,0,0 ; vbroadcastss 0x34b3(%rip),%ymm12 # 6828 <_sk_callback_avx+0x324>
+ DB 196,98,125,24,37,155,53,0,0 ; vbroadcastss 0x359b(%rip),%ymm12 # 69fc <_sk_callback_avx+0x324>
DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,169,52,0,0 ; vbroadcastss 0x34a9(%rip),%ymm12 # 682c <_sk_callback_avx+0x328>
+ DB 196,98,125,24,37,145,53,0,0 ; vbroadcastss 0x3591(%rip),%ymm12 # 6a00 <_sk_callback_avx+0x328>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,159,52,0,0 ; vbroadcastss 0x349f(%rip),%ymm12 # 6830 <_sk_callback_avx+0x32c>
+ DB 196,98,125,24,37,135,53,0,0 ; vbroadcastss 0x3587(%rip),%ymm12 # 6a04 <_sk_callback_avx+0x32c>
DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,144,52,0,0 ; vbroadcastss 0x3490(%rip),%ymm12 # 6834 <_sk_callback_avx+0x330>
+ DB 196,98,125,24,37,120,53,0,0 ; vbroadcastss 0x3578(%rip),%ymm12 # 6a08 <_sk_callback_avx+0x330>
DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,134,52,0,0 ; vbroadcastss 0x3486(%rip),%ymm12 # 6838 <_sk_callback_avx+0x334>
+ DB 196,98,125,24,37,110,53,0,0 ; vbroadcastss 0x356e(%rip),%ymm12 # 6a0c <_sk_callback_avx+0x334>
DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3
DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3
DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3
DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10
DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10
- DB 196,98,125,24,29,106,52,0,0 ; vbroadcastss 0x346a(%rip),%ymm11 # 683c <_sk_callback_avx+0x338>
+ DB 196,98,125,24,29,82,53,0,0 ; vbroadcastss 0x3552(%rip),%ymm11 # 6a10 <_sk_callback_avx+0x338>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,96,52,0,0 ; vbroadcastss 0x3460(%rip),%ymm11 # 6840 <_sk_callback_avx+0x33c>
+ DB 196,98,125,24,29,72,53,0,0 ; vbroadcastss 0x3548(%rip),%ymm11 # 6a14 <_sk_callback_avx+0x33c>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,81,52,0,0 ; vbroadcastss 0x3451(%rip),%ymm11 # 6844 <_sk_callback_avx+0x340>
+ DB 196,98,125,24,29,57,53,0,0 ; vbroadcastss 0x3539(%rip),%ymm11 # 6a18 <_sk_callback_avx+0x340>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,71,52,0,0 ; vbroadcastss 0x3447(%rip),%ymm11 # 6848 <_sk_callback_avx+0x344>
+ DB 196,98,125,24,29,47,53,0,0 ; vbroadcastss 0x352f(%rip),%ymm11 # 6a1c <_sk_callback_avx+0x344>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,56,52,0,0 ; vbroadcastss 0x3438(%rip),%ymm10 # 684c <_sk_callback_avx+0x348>
+ DB 196,98,125,24,21,32,53,0,0 ; vbroadcastss 0x3520(%rip),%ymm10 # 6a20 <_sk_callback_avx+0x348>
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
@@ -8093,38 +8311,38 @@ _sk_parametric_a_avx LABEL PROC
DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3
- DB 196,98,125,24,5,15,52,0,0 ; vbroadcastss 0x340f(%rip),%ymm8 # 6850 <_sk_callback_avx+0x34c>
+ DB 196,98,125,24,5,247,52,0,0 ; vbroadcastss 0x34f7(%rip),%ymm8 # 6a24 <_sk_callback_avx+0x34c>
DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
PUBLIC _sk_lab_to_xyz_avx
_sk_lab_to_xyz_avx LABEL PROC
- DB 196,98,125,24,5,1,52,0,0 ; vbroadcastss 0x3401(%rip),%ymm8 # 6854 <_sk_callback_avx+0x350>
+ DB 196,98,125,24,5,233,52,0,0 ; vbroadcastss 0x34e9(%rip),%ymm8 # 6a28 <_sk_callback_avx+0x350>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,247,51,0,0 ; vbroadcastss 0x33f7(%rip),%ymm8 # 6858 <_sk_callback_avx+0x354>
+ DB 196,98,125,24,5,223,52,0,0 ; vbroadcastss 0x34df(%rip),%ymm8 # 6a2c <_sk_callback_avx+0x354>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,98,125,24,13,237,51,0,0 ; vbroadcastss 0x33ed(%rip),%ymm9 # 685c <_sk_callback_avx+0x358>
+ DB 196,98,125,24,13,213,52,0,0 ; vbroadcastss 0x34d5(%rip),%ymm9 # 6a30 <_sk_callback_avx+0x358>
DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 196,193,108,88,209 ; vaddps %ymm9,%ymm2,%ymm2
- DB 196,98,125,24,5,217,51,0,0 ; vbroadcastss 0x33d9(%rip),%ymm8 # 6860 <_sk_callback_avx+0x35c>
+ DB 196,98,125,24,5,193,52,0,0 ; vbroadcastss 0x34c1(%rip),%ymm8 # 6a34 <_sk_callback_avx+0x35c>
DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,207,51,0,0 ; vbroadcastss 0x33cf(%rip),%ymm8 # 6864 <_sk_callback_avx+0x360>
+ DB 196,98,125,24,5,183,52,0,0 ; vbroadcastss 0x34b7(%rip),%ymm8 # 6a38 <_sk_callback_avx+0x360>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,197,51,0,0 ; vbroadcastss 0x33c5(%rip),%ymm8 # 6868 <_sk_callback_avx+0x364>
+ DB 196,98,125,24,5,173,52,0,0 ; vbroadcastss 0x34ad(%rip),%ymm8 # 6a3c <_sk_callback_avx+0x364>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1
- DB 196,98,125,24,5,183,51,0,0 ; vbroadcastss 0x33b7(%rip),%ymm8 # 686c <_sk_callback_avx+0x368>
+ DB 196,98,125,24,5,159,52,0,0 ; vbroadcastss 0x349f(%rip),%ymm8 # 6a40 <_sk_callback_avx+0x368>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 197,252,92,210 ; vsubps %ymm2,%ymm0,%ymm2
DB 197,116,89,193 ; vmulps %ymm1,%ymm1,%ymm8
DB 196,65,116,89,192 ; vmulps %ymm8,%ymm1,%ymm8
- DB 196,98,125,24,13,160,51,0,0 ; vbroadcastss 0x33a0(%rip),%ymm9 # 6870 <_sk_callback_avx+0x36c>
+ DB 196,98,125,24,13,136,52,0,0 ; vbroadcastss 0x3488(%rip),%ymm9 # 6a44 <_sk_callback_avx+0x36c>
DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10
- DB 196,98,125,24,29,149,51,0,0 ; vbroadcastss 0x3395(%rip),%ymm11 # 6874 <_sk_callback_avx+0x370>
+ DB 196,98,125,24,29,125,52,0,0 ; vbroadcastss 0x347d(%rip),%ymm11 # 6a48 <_sk_callback_avx+0x370>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,37,139,51,0,0 ; vbroadcastss 0x338b(%rip),%ymm12 # 6878 <_sk_callback_avx+0x374>
+ DB 196,98,125,24,37,115,52,0,0 ; vbroadcastss 0x3473(%rip),%ymm12 # 6a4c <_sk_callback_avx+0x374>
DB 196,193,116,89,204 ; vmulps %ymm12,%ymm1,%ymm1
DB 196,67,117,74,192,160 ; vblendvps %ymm10,%ymm8,%ymm1,%ymm8
DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1
@@ -8139,49 +8357,81 @@ _sk_lab_to_xyz_avx LABEL PROC
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
DB 196,193,108,89,212 ; vmulps %ymm12,%ymm2,%ymm2
DB 196,227,109,74,208,144 ; vblendvps %ymm9,%ymm0,%ymm2,%ymm2
- DB 196,226,125,24,5,65,51,0,0 ; vbroadcastss 0x3341(%rip),%ymm0 # 687c <_sk_callback_avx+0x378>
+ DB 196,226,125,24,5,41,52,0,0 ; vbroadcastss 0x3429(%rip),%ymm0 # 6a50 <_sk_callback_avx+0x378>
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,56,51,0,0 ; vbroadcastss 0x3338(%rip),%ymm8 # 6880 <_sk_callback_avx+0x37c>
+ DB 196,98,125,24,5,32,52,0,0 ; vbroadcastss 0x3420(%rip),%ymm8 # 6a54 <_sk_callback_avx+0x37c>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
PUBLIC _sk_load_a8_avx
_sk_load_a8_avx LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,62 ; jne 359f <_sk_load_a8_avx+0x4e>
- DB 197,250,126,0 ; vmovq (%rax),%xmm0
- DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
- DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
- DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
- DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
+ DB 117,66 ; jne 3689 <_sk_load_a8_avx+0x4c>
+ DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0
+ DB 197,249,219,5,107,54,0,0 ; vpand 0x366b(%rip),%xmm0,%xmm0 # 6cc0 <_sk_callback_avx+0x5e8>
+ 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 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,252,50,0,0 ; vbroadcastss 0x32fc(%rip),%ymm1 # 6884 <_sk_callback_avx+0x380>
+ DB 196,226,125,24,13,227,51,0,0 ; vbroadcastss 0x33e3(%rip),%ymm1 # 6a58 <_sk_callback_avx+0x380>
DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2
- DB 76,137,201 ; mov %r9,%rcx
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 35a8 <_sk_load_a8_avx+0x57>
- DB 196,193,249,110,194 ; vmovq %r10,%xmm0
- DB 91 ; pop %rbx
- DB 235,160 ; jmp 3565 <_sk_load_a8_avx+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,176 ; ja 364d <_sk_load_a8_avx+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,96,0,0,0 ; lea 0x60(%rip),%r10 # 3708 <_sk_load_a8_avx+0xcb>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,249,196,192,5 ; vpinsrw $0x5,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,249,196,192,4 ; vpinsrw $0x4,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,249,196,192,3 ; vpinsrw $0x3,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,249,196,192,1 ; vpinsrw $0x1,%eax,%xmm0,%xmm0
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,249,196,192,0 ; vpinsrw $0x0,%eax,%xmm0,%xmm0
+ DB 233,71,255,255,255 ; jmpq 364d <_sk_load_a8_avx+0x10>
+ DB 102,144 ; xchg %ax,%ax
+ DB 239 ; out %eax,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 217,255 ; fcos
+ DB 255 ; (bad)
+ DB 255,206 ; dec %esi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,195 ; inc %ebx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 184,255,255,255,169 ; mov $0xa9ffffff,%eax
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_gather_a8_avx
_sk_gather_a8_avx LABEL PROC
@@ -8225,7 +8475,7 @@ _sk_gather_a8_avx LABEL PROC
DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,247,49,0,0 ; vbroadcastss 0x31f7(%rip),%ymm1 # 6888 <_sk_callback_avx+0x384>
+ DB 196,226,125,24,13,108,50,0,0 ; vbroadcastss 0x326c(%rip),%ymm1 # 6a5c <_sk_callback_avx+0x384>
DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
@@ -8237,14 +8487,14 @@ PUBLIC _sk_store_a8_avx
_sk_store_a8_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
- DB 196,98,125,24,5,217,49,0,0 ; vbroadcastss 0x31d9(%rip),%ymm8 # 688c <_sk_callback_avx+0x388>
+ DB 196,98,125,24,5,78,50,0,0 ; vbroadcastss 0x324e(%rip),%ymm8 # 6a60 <_sk_callback_avx+0x388>
DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8
DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
DB 77,133,192 ; test %r8,%r8
- DB 117,10 ; jne 36dc <_sk_store_a8_avx+0x37>
+ DB 117,10 ; jne 383b <_sk_store_a8_avx+0x37>
DB 196,65,123,17,4,19 ; vmovsd %xmm8,(%r11,%rdx,1)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -8252,10 +8502,10 @@ _sk_store_a8_avx LABEL PROC
DB 65,128,225,7 ; and $0x7,%r9b
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,236 ; ja 36d8 <_sk_store_a8_avx+0x33>
+ DB 119,236 ; ja 3837 <_sk_store_a8_avx+0x33>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 3740 <_sk_store_a8_avx+0x9b>
+ DB 76,141,21,69,0,0,0 ; lea 0x45(%rip),%r10 # 38a0 <_sk_store_a8_avx+0x9c>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8266,68 +8516,100 @@ _sk_store_a8_avx LABEL PROC
DB 196,67,121,20,68,19,2,4 ; vpextrb $0x4,%xmm8,0x2(%r11,%rdx,1)
DB 196,67,121,20,68,19,1,2 ; vpextrb $0x2,%xmm8,0x1(%r11,%rdx,1)
DB 196,67,121,20,4,19,0 ; vpextrb $0x0,%xmm8,(%r11,%rdx,1)
- DB 235,154 ; jmp 36d8 <_sk_store_a8_avx+0x33>
- DB 102,144 ; xchg %ax,%ax
- DB 245 ; cmc
+ DB 235,154 ; jmp 3837 <_sk_store_a8_avx+0x33>
+ DB 15,31,0 ; nopl (%rax)
+ DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 237 ; in (%dx),%eax
+ DB 236 ; in (%dx),%al
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,229 ; jmpq *%rbp
+ DB 255,228 ; jmpq *%rsp
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 221,255 ; (bad)
+ DB 220,255 ; fdivr %st,%st(7)
DB 255 ; (bad)
- DB 255,213 ; callq *%rbp
+ DB 255,212 ; callq *%rsp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,205 ; dec %ebp
+ DB 255,204 ; dec %esp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,197 ; inc %ebp
+ DB 255,196 ; inc %esp
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
PUBLIC _sk_load_g8_avx
_sk_load_g8_avx LABEL PROC
- DB 73,137,201 ; mov %rcx,%r9
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,0 ; mov (%rax),%rax
- DB 72,1,208 ; add %rdx,%rax
+ DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 117,67 ; jne 37af <_sk_load_g8_avx+0x53>
- DB 197,250,126,0 ; vmovq (%rax),%xmm0
- DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
- DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
- DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
- DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
+ DB 117,71 ; jne 390d <_sk_load_g8_avx+0x51>
+ DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0
+ DB 197,249,219,5,252,51,0,0 ; vpand 0x33fc(%rip),%xmm0,%xmm0 # 6cd0 <_sk_callback_avx+0x5f8>
+ 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 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,253,48,0,0 ; vbroadcastss 0x30fd(%rip),%ymm1 # 6890 <_sk_callback_avx+0x38c>
+ DB 196,226,125,24,13,112,49,0,0 ; vbroadcastss 0x3170(%rip),%ymm1 # 6a64 <_sk_callback_avx+0x38c>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,242,48,0,0 ; vbroadcastss 0x30f2(%rip),%ymm3 # 6894 <_sk_callback_avx+0x390>
- DB 76,137,201 ; mov %r9,%rcx
+ DB 196,226,125,24,29,101,49,0,0 ; vbroadcastss 0x3165(%rip),%ymm3 # 6a68 <_sk_callback_avx+0x390>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 255,224 ; jmpq *%rax
- DB 83 ; push %rbx
- DB 49,201 ; xor %ecx,%ecx
- DB 77,137,195 ; mov %r8,%r11
- DB 69,49,210 ; xor %r10d,%r10d
- DB 15,182,24 ; movzbl (%rax),%ebx
- DB 72,255,192 ; inc %rax
- DB 72,211,227 ; shl %cl,%rbx
- DB 73,9,218 ; or %rbx,%r10
- DB 72,131,193,8 ; add $0x8,%rcx
- DB 73,255,203 ; dec %r11
- DB 117,235 ; jne 37b8 <_sk_load_g8_avx+0x5c>
- DB 196,193,249,110,194 ; vmovq %r10,%xmm0
- DB 91 ; pop %rbx
- DB 235,155 ; jmp 3770 <_sk_load_g8_avx+0x14>
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 119,171 ; ja 38cc <_sk_load_g8_avx+0x10>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,96,0,0,0 ; lea 0x60(%rip),%r10 # 398c <_sk_load_g8_avx+0xd0>
+ DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
+ DB 76,1,208 ; add %r10,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax
+ DB 197,249,196,192,5 ; vpinsrw $0x5,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax
+ DB 197,249,196,192,4 ; vpinsrw $0x4,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax
+ DB 197,249,196,192,3 ; vpinsrw $0x3,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax
+ DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0
+ DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax
+ DB 197,249,196,192,1 ; vpinsrw $0x1,%eax,%xmm0,%xmm0
+ DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax
+ DB 197,249,196,192,0 ; vpinsrw $0x0,%eax,%xmm0,%xmm0
+ DB 233,66,255,255,255 ; jmpq 38cc <_sk_load_g8_avx+0x10>
+ DB 102,144 ; xchg %ax,%ax
+ DB 239 ; out %eax,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 217,255 ; fcos
+ DB 255 ; (bad)
+ DB 255,206 ; dec %esi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,195 ; inc %ebx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 184,255,255,255,169 ; mov $0xa9ffffff,%eax
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
PUBLIC _sk_gather_g8_avx
_sk_gather_g8_avx LABEL PROC
@@ -8371,10 +8653,10 @@ _sk_gather_g8_avx LABEL PROC
DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,247,47,0,0 ; vbroadcastss 0x2ff7(%rip),%ymm1 # 6898 <_sk_callback_avx+0x394>
+ DB 196,226,125,24,13,248,47,0,0 ; vbroadcastss 0x2ff8(%rip),%ymm1 # 6a6c <_sk_callback_avx+0x394>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,236,47,0,0 ; vbroadcastss 0x2fec(%rip),%ymm3 # 689c <_sk_callback_avx+0x398>
+ DB 196,226,125,24,29,237,47,0,0 ; vbroadcastss 0x2fed(%rip),%ymm3 # 6a70 <_sk_callback_avx+0x398>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 255,224 ; jmpq *%rax
@@ -8384,9 +8666,9 @@ _sk_gather_i8_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,137,193 ; mov %rax,%r9
DB 77,133,201 ; test %r9,%r9
- DB 116,5 ; je 38c9 <_sk_gather_i8_avx+0xf>
+ DB 116,5 ; je 3a9c <_sk_gather_i8_avx+0xf>
DB 76,137,200 ; mov %r9,%rax
- DB 235,2 ; jmp 38cb <_sk_gather_i8_avx+0x11>
+ DB 235,2 ; jmp 3a9e <_sk_gather_i8_avx+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 83 ; push %rbx
DB 76,139,16 ; mov (%rax),%r10
@@ -8444,10 +8726,10 @@ _sk_gather_i8_avx LABEL PROC
DB 72,193,235,30 ; shr $0x1e,%rbx
DB 196,195,121,34,28,27,3 ; vpinsrd $0x3,(%r11,%rbx,1),%xmm0,%xmm3
DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
- DB 197,124,40,21,87,49,0,0 ; vmovaps 0x3157(%rip),%ymm10 # 6b40 <_sk_callback_avx+0x63c>
+ DB 197,124,40,21,164,49,0,0 ; vmovaps 0x31a4(%rip),%ymm10 # 6d60 <_sk_callback_avx+0x688>
DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,13,165,46,0,0 ; vbroadcastss 0x2ea5(%rip),%ymm9 # 68a0 <_sk_callback_avx+0x39c>
+ DB 196,98,125,24,13,166,46,0,0 ; vbroadcastss 0x2ea6(%rip),%ymm9 # 6a74 <_sk_callback_avx+0x39c>
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1
DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2
@@ -8475,38 +8757,38 @@ _sk_load_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 15,133,128,0,0,0 ; jne 3aeb <_sk_load_565_avx+0x8e>
+ DB 15,133,128,0,0,0 ; jne 3cbe <_sk_load_565_avx+0x8e>
DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0
DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
- DB 196,226,125,24,5,23,46,0,0 ; vbroadcastss 0x2e17(%rip),%ymm0 # 68a4 <_sk_callback_avx+0x3a0>
+ DB 196,226,125,24,5,24,46,0,0 ; vbroadcastss 0x2e18(%rip),%ymm0 # 6a78 <_sk_callback_avx+0x3a0>
DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,10,46,0,0 ; vbroadcastss 0x2e0a(%rip),%ymm1 # 68a8 <_sk_callback_avx+0x3a4>
+ DB 196,226,125,24,13,11,46,0,0 ; vbroadcastss 0x2e0b(%rip),%ymm1 # 6a7c <_sk_callback_avx+0x3a4>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,1,46,0,0 ; vbroadcastss 0x2e01(%rip),%ymm1 # 68ac <_sk_callback_avx+0x3a8>
+ DB 196,226,125,24,13,2,46,0,0 ; vbroadcastss 0x2e02(%rip),%ymm1 # 6a80 <_sk_callback_avx+0x3a8>
DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,244,45,0,0 ; vbroadcastss 0x2df4(%rip),%ymm3 # 68b0 <_sk_callback_avx+0x3ac>
+ DB 196,226,125,24,29,245,45,0,0 ; vbroadcastss 0x2df5(%rip),%ymm3 # 6a84 <_sk_callback_avx+0x3ac>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,24,29,235,45,0,0 ; vbroadcastss 0x2deb(%rip),%ymm3 # 68b4 <_sk_callback_avx+0x3b0>
+ DB 196,226,125,24,29,236,45,0,0 ; vbroadcastss 0x2dec(%rip),%ymm3 # 6a88 <_sk_callback_avx+0x3b0>
DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,222,45,0,0 ; vbroadcastss 0x2dde(%rip),%ymm3 # 68b8 <_sk_callback_avx+0x3b4>
+ DB 196,226,125,24,29,223,45,0,0 ; vbroadcastss 0x2ddf(%rip),%ymm3 # 6a8c <_sk_callback_avx+0x3b4>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,211,45,0,0 ; vbroadcastss 0x2dd3(%rip),%ymm3 # 68bc <_sk_callback_avx+0x3b8>
+ DB 196,226,125,24,29,212,45,0,0 ; vbroadcastss 0x2dd4(%rip),%ymm3 # 6a90 <_sk_callback_avx+0x3b8>
DB 255,224 ; jmpq *%rax
DB 69,137,193 ; mov %r8d,%r9d
DB 65,128,225,7 ; and $0x7,%r9b
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 15,135,110,255,255,255 ; ja 3a71 <_sk_load_565_avx+0x14>
+ DB 15,135,110,255,255,255 ; ja 3c44 <_sk_load_565_avx+0x14>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,74,0,0,0 ; lea 0x4a(%rip),%r10 # 3b58 <_sk_load_565_avx+0xfb>
+ DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 3d2c <_sk_load_565_avx+0xfc>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8518,27 +8800,28 @@ _sk_load_565_avx LABEL PROC
DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- DB 233,26,255,255,255 ; jmpq 3a71 <_sk_load_565_avx+0x14>
- DB 144 ; nop
- DB 243,255 ; repz (bad)
+ DB 233,26,255,255,255 ; jmpq 3c44 <_sk_load_565_avx+0x14>
+ DB 102,144 ; xchg %ax,%ax
+ DB 242,255 ; repnz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 3b5d <_sk_load_565_avx+0x100>
+ DB 234 ; (bad)
DB 255 ; (bad)
- DB 255,227 ; jmpq *%rbx
DB 255 ; (bad)
+ DB 255,226 ; jmpq *%rdx
DB 255 ; (bad)
DB 255 ; (bad)
- DB 219,255 ; (bad)
DB 255 ; (bad)
- DB 255,211 ; callq *%rbx
+ DB 218,255 ; (bad)
DB 255 ; (bad)
+ DB 255,210 ; callq *%rdx
DB 255 ; (bad)
- DB 255,203 ; dec %ebx
DB 255 ; (bad)
+ DB 255,202 ; dec %edx
DB 255 ; (bad)
DB 255 ; (bad)
- DB 191 ; .byte 0xbf
+ DB 255 ; (bad)
+ DB 190 ; .byte 0xbe
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -8589,37 +8872,37 @@ _sk_gather_565_avx LABEL PROC
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
- DB 196,226,125,24,5,122,44,0,0 ; vbroadcastss 0x2c7a(%rip),%ymm0 # 68c0 <_sk_callback_avx+0x3bc>
+ DB 196,226,125,24,5,122,44,0,0 ; vbroadcastss 0x2c7a(%rip),%ymm0 # 6a94 <_sk_callback_avx+0x3bc>
DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,109,44,0,0 ; vbroadcastss 0x2c6d(%rip),%ymm1 # 68c4 <_sk_callback_avx+0x3c0>
+ DB 196,226,125,24,13,109,44,0,0 ; vbroadcastss 0x2c6d(%rip),%ymm1 # 6a98 <_sk_callback_avx+0x3c0>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,100,44,0,0 ; vbroadcastss 0x2c64(%rip),%ymm1 # 68c8 <_sk_callback_avx+0x3c4>
+ DB 196,226,125,24,13,100,44,0,0 ; vbroadcastss 0x2c64(%rip),%ymm1 # 6a9c <_sk_callback_avx+0x3c4>
DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,87,44,0,0 ; vbroadcastss 0x2c57(%rip),%ymm3 # 68cc <_sk_callback_avx+0x3c8>
+ DB 196,226,125,24,29,87,44,0,0 ; vbroadcastss 0x2c57(%rip),%ymm3 # 6aa0 <_sk_callback_avx+0x3c8>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,24,29,78,44,0,0 ; vbroadcastss 0x2c4e(%rip),%ymm3 # 68d0 <_sk_callback_avx+0x3cc>
+ DB 196,226,125,24,29,78,44,0,0 ; vbroadcastss 0x2c4e(%rip),%ymm3 # 6aa4 <_sk_callback_avx+0x3cc>
DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,65,44,0,0 ; vbroadcastss 0x2c41(%rip),%ymm3 # 68d4 <_sk_callback_avx+0x3d0>
+ DB 196,226,125,24,29,65,44,0,0 ; vbroadcastss 0x2c41(%rip),%ymm3 # 6aa8 <_sk_callback_avx+0x3d0>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,54,44,0,0 ; vbroadcastss 0x2c36(%rip),%ymm3 # 68d8 <_sk_callback_avx+0x3d4>
+ DB 196,226,125,24,29,54,44,0,0 ; vbroadcastss 0x2c36(%rip),%ymm3 # 6aac <_sk_callback_avx+0x3d4>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_store_565_avx
_sk_store_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
- DB 196,98,125,24,5,42,44,0,0 ; vbroadcastss 0x2c2a(%rip),%ymm8 # 68dc <_sk_callback_avx+0x3d8>
+ DB 196,98,125,24,5,42,44,0,0 ; vbroadcastss 0x2c2a(%rip),%ymm8 # 6ab0 <_sk_callback_avx+0x3d8>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,193,41,114,241,11 ; vpslld $0xb,%xmm9,%xmm10
DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9
DB 196,193,49,114,241,11 ; vpslld $0xb,%xmm9,%xmm9
DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
- DB 196,98,125,24,21,3,44,0,0 ; vbroadcastss 0x2c03(%rip),%ymm10 # 68e0 <_sk_callback_avx+0x3dc>
+ DB 196,98,125,24,21,3,44,0,0 ; vbroadcastss 0x2c03(%rip),%ymm10 # 6ab4 <_sk_callback_avx+0x3dc>
DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10
DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10
DB 196,193,33,114,242,5 ; vpslld $0x5,%xmm10,%xmm11
@@ -8633,7 +8916,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,10 ; jne 3d2d <_sk_store_565_avx+0x89>
+ DB 117,10 ; jne 3f01 <_sk_store_565_avx+0x89>
DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -8641,9 +8924,9 @@ _sk_store_565_avx LABEL PROC
DB 65,128,225,7 ; and $0x7,%r9b
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,236 ; ja 3d29 <_sk_store_565_avx+0x85>
+ DB 119,236 ; ja 3efd <_sk_store_565_avx+0x85>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 3d8c <_sk_store_565_avx+0xe8>
+ DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 3f60 <_sk_store_565_avx+0xe8>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8654,7 +8937,7 @@ _sk_store_565_avx LABEL PROC
DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- DB 235,159 ; jmp 3d29 <_sk_store_565_avx+0x85>
+ DB 235,159 ; jmp 3efd <_sk_store_565_avx+0x85>
DB 102,144 ; xchg %ax,%ax
DB 245 ; cmc
DB 255 ; (bad)
@@ -8685,31 +8968,31 @@ _sk_load_4444_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
DB 77,133,192 ; test %r8,%r8
- DB 15,133,152,0,0,0 ; jne 3e4e <_sk_load_4444_avx+0xa6>
+ DB 15,133,152,0,0,0 ; jne 4022 <_sk_load_4444_avx+0xa6>
DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0
DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3
- DB 196,226,125,24,5,12,43,0,0 ; vbroadcastss 0x2b0c(%rip),%ymm0 # 68e4 <_sk_callback_avx+0x3e0>
+ DB 196,226,125,24,5,12,43,0,0 ; vbroadcastss 0x2b0c(%rip),%ymm0 # 6ab8 <_sk_callback_avx+0x3e0>
DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,255,42,0,0 ; vbroadcastss 0x2aff(%rip),%ymm1 # 68e8 <_sk_callback_avx+0x3e4>
+ DB 196,226,125,24,13,255,42,0,0 ; vbroadcastss 0x2aff(%rip),%ymm1 # 6abc <_sk_callback_avx+0x3e4>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,246,42,0,0 ; vbroadcastss 0x2af6(%rip),%ymm1 # 68ec <_sk_callback_avx+0x3e8>
+ DB 196,226,125,24,13,246,42,0,0 ; vbroadcastss 0x2af6(%rip),%ymm1 # 6ac0 <_sk_callback_avx+0x3e8>
DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,233,42,0,0 ; vbroadcastss 0x2ae9(%rip),%ymm2 # 68f0 <_sk_callback_avx+0x3ec>
+ DB 196,226,125,24,21,233,42,0,0 ; vbroadcastss 0x2ae9(%rip),%ymm2 # 6ac4 <_sk_callback_avx+0x3ec>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,24,21,224,42,0,0 ; vbroadcastss 0x2ae0(%rip),%ymm2 # 68f4 <_sk_callback_avx+0x3f0>
+ DB 196,226,125,24,21,224,42,0,0 ; vbroadcastss 0x2ae0(%rip),%ymm2 # 6ac8 <_sk_callback_avx+0x3f0>
DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,211,42,0,0 ; vbroadcastss 0x2ad3(%rip),%ymm8 # 68f8 <_sk_callback_avx+0x3f4>
+ DB 196,98,125,24,5,211,42,0,0 ; vbroadcastss 0x2ad3(%rip),%ymm8 # 6acc <_sk_callback_avx+0x3f4>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,201,42,0,0 ; vbroadcastss 0x2ac9(%rip),%ymm8 # 68fc <_sk_callback_avx+0x3f8>
+ DB 196,98,125,24,5,201,42,0,0 ; vbroadcastss 0x2ac9(%rip),%ymm8 # 6ad0 <_sk_callback_avx+0x3f8>
DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,187,42,0,0 ; vbroadcastss 0x2abb(%rip),%ymm8 # 6900 <_sk_callback_avx+0x3fc>
+ DB 196,98,125,24,5,187,42,0,0 ; vbroadcastss 0x2abb(%rip),%ymm8 # 6ad4 <_sk_callback_avx+0x3fc>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -8718,9 +9001,9 @@ _sk_load_4444_avx LABEL PROC
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 15,135,86,255,255,255 ; ja 3dbc <_sk_load_4444_avx+0x14>
+ DB 15,135,86,255,255,255 ; ja 3f90 <_sk_load_4444_avx+0x14>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 3ebc <_sk_load_4444_avx+0x114>
+ DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 4090 <_sk_load_4444_avx+0x114>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8732,7 +9015,7 @@ _sk_load_4444_avx LABEL PROC
DB 196,193,121,196,68,83,4,2 ; vpinsrw $0x2,0x4(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,68,83,2,1 ; vpinsrw $0x1,0x2(%r11,%rdx,2),%xmm0,%xmm0
DB 196,193,121,196,4,83,0 ; vpinsrw $0x0,(%r11,%rdx,2),%xmm0,%xmm0
- DB 233,2,255,255,255 ; jmpq 3dbc <_sk_load_4444_avx+0x14>
+ DB 233,2,255,255,255 ; jmpq 3f90 <_sk_load_4444_avx+0x14>
DB 102,144 ; xchg %ax,%ax
DB 242,255 ; repnz (bad)
DB 255 ; (bad)
@@ -8804,25 +9087,25 @@ _sk_gather_4444_avx LABEL PROC
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3
- DB 196,226,125,24,5,90,41,0,0 ; vbroadcastss 0x295a(%rip),%ymm0 # 6904 <_sk_callback_avx+0x400>
+ DB 196,226,125,24,5,90,41,0,0 ; vbroadcastss 0x295a(%rip),%ymm0 # 6ad8 <_sk_callback_avx+0x400>
DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,77,41,0,0 ; vbroadcastss 0x294d(%rip),%ymm1 # 6908 <_sk_callback_avx+0x404>
+ DB 196,226,125,24,13,77,41,0,0 ; vbroadcastss 0x294d(%rip),%ymm1 # 6adc <_sk_callback_avx+0x404>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,68,41,0,0 ; vbroadcastss 0x2944(%rip),%ymm1 # 690c <_sk_callback_avx+0x408>
+ DB 196,226,125,24,13,68,41,0,0 ; vbroadcastss 0x2944(%rip),%ymm1 # 6ae0 <_sk_callback_avx+0x408>
DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,55,41,0,0 ; vbroadcastss 0x2937(%rip),%ymm2 # 6910 <_sk_callback_avx+0x40c>
+ DB 196,226,125,24,21,55,41,0,0 ; vbroadcastss 0x2937(%rip),%ymm2 # 6ae4 <_sk_callback_avx+0x40c>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,24,21,46,41,0,0 ; vbroadcastss 0x292e(%rip),%ymm2 # 6914 <_sk_callback_avx+0x410>
+ DB 196,226,125,24,21,46,41,0,0 ; vbroadcastss 0x292e(%rip),%ymm2 # 6ae8 <_sk_callback_avx+0x410>
DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,33,41,0,0 ; vbroadcastss 0x2921(%rip),%ymm8 # 6918 <_sk_callback_avx+0x414>
+ DB 196,98,125,24,5,33,41,0,0 ; vbroadcastss 0x2921(%rip),%ymm8 # 6aec <_sk_callback_avx+0x414>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,23,41,0,0 ; vbroadcastss 0x2917(%rip),%ymm8 # 691c <_sk_callback_avx+0x418>
+ DB 196,98,125,24,5,23,41,0,0 ; vbroadcastss 0x2917(%rip),%ymm8 # 6af0 <_sk_callback_avx+0x418>
DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,9,41,0,0 ; vbroadcastss 0x2909(%rip),%ymm8 # 6920 <_sk_callback_avx+0x41c>
+ DB 196,98,125,24,5,9,41,0,0 ; vbroadcastss 0x2909(%rip),%ymm8 # 6af4 <_sk_callback_avx+0x41c>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -8831,7 +9114,7 @@ PUBLIC _sk_store_4444_avx
_sk_store_4444_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,24 ; mov (%rax),%r11
- DB 196,98,125,24,5,246,40,0,0 ; vbroadcastss 0x28f6(%rip),%ymm8 # 6924 <_sk_callback_avx+0x420>
+ DB 196,98,125,24,5,246,40,0,0 ; vbroadcastss 0x28f6(%rip),%ymm8 # 6af8 <_sk_callback_avx+0x420>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,193,41,114,241,12 ; vpslld $0xc,%xmm9,%xmm10
@@ -8858,7 +9141,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,10 ; jne 40c7 <_sk_store_4444_avx+0xa7>
+ DB 117,10 ; jne 429b <_sk_store_4444_avx+0xa7>
DB 196,65,122,127,4,83 ; vmovdqu %xmm8,(%r11,%rdx,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -8866,9 +9149,9 @@ _sk_store_4444_avx LABEL PROC
DB 65,128,225,7 ; and $0x7,%r9b
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 119,236 ; ja 40c3 <_sk_store_4444_avx+0xa3>
+ DB 119,236 ; ja 4297 <_sk_store_4444_avx+0xa3>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 4124 <_sk_store_4444_avx+0x104>
+ DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 42f8 <_sk_store_4444_avx+0x104>
DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8879,7 +9162,7 @@ _sk_store_4444_avx LABEL PROC
DB 196,67,121,21,68,83,4,2 ; vpextrw $0x2,%xmm8,0x4(%r11,%rdx,2)
DB 196,67,121,21,68,83,2,1 ; vpextrw $0x1,%xmm8,0x2(%r11,%rdx,2)
DB 196,67,121,21,4,83,0 ; vpextrw $0x0,%xmm8,(%r11,%rdx,2)
- DB 235,159 ; jmp 40c3 <_sk_store_4444_avx+0xa3>
+ DB 235,159 ; jmp 4297 <_sk_store_4444_avx+0xa3>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -8911,12 +9194,12 @@ _sk_load_8888_avx LABEL PROC
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 76,3,16 ; add (%rax),%r10
DB 77,133,192 ; test %r8,%r8
- DB 15,133,139,0,0,0 ; jne 41e5 <_sk_load_8888_avx+0xa5>
+ DB 15,133,139,0,0,0 ; jne 43b9 <_sk_load_8888_avx+0xa5>
DB 196,193,124,16,26 ; vmovups (%r10),%ymm3
- DB 197,124,40,21,249,41,0,0 ; vmovaps 0x29f9(%rip),%ymm10 # 6b60 <_sk_callback_avx+0x65c>
+ DB 197,124,40,21,69,42,0,0 ; vmovaps 0x2a45(%rip),%ymm10 # 6d80 <_sk_callback_avx+0x6a8>
DB 196,193,100,84,194 ; vandps %ymm10,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,175,39,0,0 ; vbroadcastss 0x27af(%rip),%ymm8 # 6928 <_sk_callback_avx+0x424>
+ DB 196,98,125,24,5,175,39,0,0 ; vbroadcastss 0x27af(%rip),%ymm8 # 6afc <_sk_callback_avx+0x424>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 197,241,114,211,8 ; vpsrld $0x8,%xmm3,%xmm1
DB 196,195,125,25,217,1 ; vextractf128 $0x1,%ymm3,%xmm9
@@ -8947,13 +9230,13 @@ _sk_load_8888_avx LABEL PROC
DB 72,211,232 ; shr %cl,%rax
DB 196,225,249,110,192 ; vmovq %rax,%xmm0
DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0
- DB 196,226,121,0,13,179,40,0,0 ; vpshufb 0x28b3(%rip),%xmm0,%xmm1 # 6ac0 <_sk_callback_avx+0x5bc>
+ DB 196,226,121,0,13,255,40,0,0 ; vpshufb 0x28ff(%rip),%xmm0,%xmm1 # 6ce0 <_sk_callback_avx+0x608>
DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1
- DB 196,226,121,0,5,181,40,0,0 ; vpshufb 0x28b5(%rip),%xmm0,%xmm0 # 6ad0 <_sk_callback_avx+0x5cc>
+ DB 196,226,121,0,5,1,41,0,0 ; vpshufb 0x2901(%rip),%xmm0,%xmm0 # 6cf0 <_sk_callback_avx+0x618>
DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3
- DB 233,47,255,255,255 ; jmpq 415f <_sk_load_8888_avx+0x1f>
+ DB 233,47,255,255,255 ; jmpq 4333 <_sk_load_8888_avx+0x1f>
PUBLIC _sk_gather_8888_avx
_sk_gather_8888_avx LABEL PROC
@@ -8990,10 +9273,10 @@ _sk_gather_8888_avx LABEL PROC
DB 73,193,234,32 ; shr $0x20,%r10
DB 196,131,121,34,28,145,3 ; vpinsrd $0x3,(%r9,%r10,4),%xmm0,%xmm3
DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
- DB 197,124,40,21,165,40,0,0 ; vmovaps 0x28a5(%rip),%ymm10 # 6b80 <_sk_callback_avx+0x67c>
+ DB 197,124,40,21,241,40,0,0 ; vmovaps 0x28f1(%rip),%ymm10 # 6da0 <_sk_callback_avx+0x6c8>
DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,13,63,38,0,0 ; vbroadcastss 0x263f(%rip),%ymm9 # 692c <_sk_callback_avx+0x428>
+ DB 196,98,125,24,13,63,38,0,0 ; vbroadcastss 0x263f(%rip),%ymm9 # 6b00 <_sk_callback_avx+0x428>
DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0
DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1
DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2
@@ -9022,7 +9305,7 @@ _sk_store_8888_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
DB 76,3,16 ; add (%rax),%r10
- DB 196,98,125,24,5,200,37,0,0 ; vbroadcastss 0x25c8(%rip),%ymm8 # 6930 <_sk_callback_avx+0x42c>
+ DB 196,98,125,24,5,200,37,0,0 ; vbroadcastss 0x25c8(%rip),%ymm8 # 6b04 <_sk_callback_avx+0x42c>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10
@@ -9047,7 +9330,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,14 ; jne 43fa <_sk_store_8888_avx+0xac>
+ DB 117,14 ; jne 45ce <_sk_store_8888_avx+0xac>
DB 196,65,124,17,2 ; vmovups %ymm8,(%r10)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,201 ; mov %r9,%rcx
@@ -9060,13 +9343,13 @@ _sk_store_8888_avx LABEL PROC
DB 72,211,232 ; shr %cl,%rax
DB 196,97,249,110,200 ; vmovq %rax,%xmm9
DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9
- DB 196,98,49,0,21,190,38,0,0 ; vpshufb 0x26be(%rip),%xmm9,%xmm10 # 6ae0 <_sk_callback_avx+0x5dc>
+ DB 196,98,49,0,21,10,39,0,0 ; vpshufb 0x270a(%rip),%xmm9,%xmm10 # 6d00 <_sk_callback_avx+0x628>
DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10
- DB 196,98,49,0,13,192,38,0,0 ; vpshufb 0x26c0(%rip),%xmm9,%xmm9 # 6af0 <_sk_callback_avx+0x5ec>
+ DB 196,98,49,0,13,12,39,0,0 ; vpshufb 0x270c(%rip),%xmm9,%xmm9 # 6d10 <_sk_callback_avx+0x638>
DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9
DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10)
- DB 235,175 ; jmp 43f1 <_sk_store_8888_avx+0xa3>
+ DB 235,175 ; jmp 45c5 <_sk_store_8888_avx+0xa3>
PUBLIC _sk_load_f16_avx
_sk_load_f16_avx LABEL PROC
@@ -9078,7 +9361,7 @@ _sk_load_f16_avx LABEL PROC
DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp)
DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp)
DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp)
- DB 15,133,143,2,0,0 ; jne 46fd <_sk_load_f16_avx+0x2bb>
+ DB 15,133,143,2,0,0 ; jne 48d1 <_sk_load_f16_avx+0x2bb>
DB 197,121,16,4,208 ; vmovupd (%rax,%rdx,8),%xmm8
DB 197,249,16,84,208,16 ; vmovupd 0x10(%rax,%rdx,8),%xmm2
DB 197,249,16,76,208,32 ; vmovupd 0x20(%rax,%rdx,8),%xmm1
@@ -9096,13 +9379,13 @@ _sk_load_f16_avx LABEL PROC
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- DB 196,98,125,24,37,109,36,0,0 ; vbroadcastss 0x246d(%rip),%ymm12 # 6934 <_sk_callback_avx+0x430>
+ DB 196,98,125,24,37,109,36,0,0 ; vbroadcastss 0x246d(%rip),%ymm12 # 6b08 <_sk_callback_avx+0x430>
DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1
DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0
DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14
- DB 196,98,121,24,29,89,36,0,0 ; vbroadcastss 0x2459(%rip),%xmm11 # 6938 <_sk_callback_avx+0x434>
+ DB 196,98,121,24,29,89,36,0,0 ; vbroadcastss 0x2459(%rip),%xmm11 # 6b0c <_sk_callback_avx+0x434>
DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3
- DB 196,98,121,24,45,79,36,0,0 ; vbroadcastss 0x244f(%rip),%xmm13 # 693c <_sk_callback_avx+0x438>
+ DB 196,98,121,24,45,79,36,0,0 ; vbroadcastss 0x244f(%rip),%xmm13 # 6b10 <_sk_callback_avx+0x438>
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
@@ -9116,7 +9399,7 @@ _sk_load_f16_avx LABEL PROC
DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0
DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1
- DB 196,226,121,24,29,5,36,0,0 ; vbroadcastss 0x2405(%rip),%xmm3 # 6940 <_sk_callback_avx+0x43c>
+ DB 196,226,121,24,29,5,36,0,0 ; vbroadcastss 0x2405(%rip),%xmm3 # 6b14 <_sk_callback_avx+0x43c>
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
@@ -9209,29 +9492,29 @@ _sk_load_f16_avx LABEL PROC
DB 197,123,16,4,208 ; vmovsd (%rax,%rdx,8),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,79 ; je 475c <_sk_load_f16_avx+0x31a>
+ DB 116,79 ; je 4930 <_sk_load_f16_avx+0x31a>
DB 197,57,22,68,208,8 ; vmovhpd 0x8(%rax,%rdx,8),%xmm8,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,67 ; jb 475c <_sk_load_f16_avx+0x31a>
+ DB 114,67 ; jb 4930 <_sk_load_f16_avx+0x31a>
DB 197,251,16,84,208,16 ; vmovsd 0x10(%rax,%rdx,8),%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 116,68 ; je 4769 <_sk_load_f16_avx+0x327>
+ DB 116,68 ; je 493d <_sk_load_f16_avx+0x327>
DB 197,233,22,84,208,24 ; vmovhpd 0x18(%rax,%rdx,8),%xmm2,%xmm2
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,56 ; jb 4769 <_sk_load_f16_avx+0x327>
+ DB 114,56 ; jb 493d <_sk_load_f16_avx+0x327>
DB 197,251,16,76,208,32 ; vmovsd 0x20(%rax,%rdx,8),%xmm1
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 15,132,68,253,255,255 ; je 4485 <_sk_load_f16_avx+0x43>
+ DB 15,132,68,253,255,255 ; je 4659 <_sk_load_f16_avx+0x43>
DB 197,241,22,76,208,40 ; vmovhpd 0x28(%rax,%rdx,8),%xmm1,%xmm1
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 15,130,52,253,255,255 ; jb 4485 <_sk_load_f16_avx+0x43>
+ DB 15,130,52,253,255,255 ; jb 4659 <_sk_load_f16_avx+0x43>
DB 197,122,126,76,208,48 ; vmovq 0x30(%rax,%rdx,8),%xmm9
- DB 233,41,253,255,255 ; jmpq 4485 <_sk_load_f16_avx+0x43>
+ DB 233,41,253,255,255 ; jmpq 4659 <_sk_load_f16_avx+0x43>
DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,28,253,255,255 ; jmpq 4485 <_sk_load_f16_avx+0x43>
+ DB 233,28,253,255,255 ; jmpq 4659 <_sk_load_f16_avx+0x43>
DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1
- DB 233,19,253,255,255 ; jmpq 4485 <_sk_load_f16_avx+0x43>
+ DB 233,19,253,255,255 ; jmpq 4659 <_sk_load_f16_avx+0x43>
PUBLIC _sk_gather_f16_avx
_sk_gather_f16_avx LABEL PROC
@@ -9290,13 +9573,13 @@ _sk_gather_f16_avx LABEL PROC
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- DB 196,98,125,24,37,198,32,0,0 ; vbroadcastss 0x20c6(%rip),%ymm12 # 6944 <_sk_callback_avx+0x440>
+ DB 196,98,125,24,37,198,32,0,0 ; vbroadcastss 0x20c6(%rip),%ymm12 # 6b18 <_sk_callback_avx+0x440>
DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1
DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0
DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14
- DB 196,98,121,24,29,178,32,0,0 ; vbroadcastss 0x20b2(%rip),%xmm11 # 6948 <_sk_callback_avx+0x444>
+ DB 196,98,121,24,29,178,32,0,0 ; vbroadcastss 0x20b2(%rip),%xmm11 # 6b1c <_sk_callback_avx+0x444>
DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3
- DB 196,98,121,24,45,168,32,0,0 ; vbroadcastss 0x20a8(%rip),%xmm13 # 694c <_sk_callback_avx+0x448>
+ DB 196,98,121,24,45,168,32,0,0 ; vbroadcastss 0x20a8(%rip),%xmm13 # 6b20 <_sk_callback_avx+0x448>
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
@@ -9310,7 +9593,7 @@ _sk_gather_f16_avx LABEL PROC
DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0
DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1
- DB 196,226,121,24,29,94,32,0,0 ; vbroadcastss 0x205e(%rip),%xmm3 # 6950 <_sk_callback_avx+0x44c>
+ DB 196,226,121,24,29,94,32,0,0 ; vbroadcastss 0x205e(%rip),%xmm3 # 6b24 <_sk_callback_avx+0x44c>
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
@@ -9408,12 +9691,12 @@ _sk_store_f16_avx LABEL PROC
DB 197,252,17,180,36,128,0,0,0 ; vmovups %ymm6,0x80(%rsp)
DB 197,252,17,108,36,96 ; vmovups %ymm5,0x60(%rsp)
DB 197,252,17,100,36,64 ; vmovups %ymm4,0x40(%rsp)
- DB 196,98,125,24,13,114,30,0,0 ; vbroadcastss 0x1e72(%rip),%ymm9 # 6954 <_sk_callback_avx+0x450>
+ DB 196,98,125,24,13,114,30,0,0 ; vbroadcastss 0x1e72(%rip),%ymm9 # 6b28 <_sk_callback_avx+0x450>
DB 196,65,124,84,209 ; vandps %ymm9,%ymm0,%ymm10
DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp)
DB 196,65,124,87,218 ; vxorps %ymm10,%ymm0,%ymm11
DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12
- DB 196,98,121,24,5,88,30,0,0 ; vbroadcastss 0x1e58(%rip),%xmm8 # 6958 <_sk_callback_avx+0x454>
+ DB 196,98,121,24,5,88,30,0,0 ; vbroadcastss 0x1e58(%rip),%xmm8 # 6b2c <_sk_callback_avx+0x454>
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
@@ -9423,7 +9706,7 @@ _sk_store_f16_avx LABEL PROC
DB 196,67,13,24,242,1 ; vinsertf128 $0x1,%xmm10,%ymm14,%ymm14
DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11
DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12
- DB 196,98,125,24,21,31,30,0,0 ; vbroadcastss 0x1e1f(%rip),%ymm10 # 695c <_sk_callback_avx+0x458>
+ DB 196,98,125,24,21,31,30,0,0 ; vbroadcastss 0x1e1f(%rip),%ymm10 # 6b30 <_sk_callback_avx+0x458>
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
@@ -9505,7 +9788,7 @@ _sk_store_f16_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,75 ; jne 4d24 <_sk_store_f16_avx+0x270>
+ DB 117,75 ; jne 4ef8 <_sk_store_f16_avx+0x270>
DB 197,120,17,28,208 ; vmovups %xmm11,(%rax,%rdx,8)
DB 197,120,17,84,208,16 ; vmovups %xmm10,0x10(%rax,%rdx,8)
DB 197,120,17,76,208,32 ; vmovups %xmm9,0x20(%rax,%rdx,8)
@@ -9521,22 +9804,22 @@ _sk_store_f16_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,208 ; vmovq %xmm11,(%rax,%rdx,8)
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,193 ; je 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 116,193 ; je 4ec4 <_sk_store_f16_avx+0x23c>
DB 197,121,23,92,208,8 ; vmovhpd %xmm11,0x8(%rax,%rdx,8)
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,181 ; jb 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 114,181 ; jb 4ec4 <_sk_store_f16_avx+0x23c>
DB 197,121,214,84,208,16 ; vmovq %xmm10,0x10(%rax,%rdx,8)
- DB 116,173 ; je 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 116,173 ; je 4ec4 <_sk_store_f16_avx+0x23c>
DB 197,121,23,84,208,24 ; vmovhpd %xmm10,0x18(%rax,%rdx,8)
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,161 ; jb 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 114,161 ; jb 4ec4 <_sk_store_f16_avx+0x23c>
DB 197,121,214,76,208,32 ; vmovq %xmm9,0x20(%rax,%rdx,8)
- DB 116,153 ; je 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 116,153 ; je 4ec4 <_sk_store_f16_avx+0x23c>
DB 197,121,23,76,208,40 ; vmovhpd %xmm9,0x28(%rax,%rdx,8)
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,141 ; jb 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 114,141 ; jb 4ec4 <_sk_store_f16_avx+0x23c>
DB 197,121,214,68,208,48 ; vmovq %xmm8,0x30(%rax,%rdx,8)
- DB 235,133 ; jmp 4cf0 <_sk_store_f16_avx+0x23c>
+ DB 235,133 ; jmp 4ec4 <_sk_store_f16_avx+0x23c>
PUBLIC _sk_load_u16_be_avx
_sk_load_u16_be_avx LABEL PROC
@@ -9544,7 +9827,7 @@ _sk_load_u16_be_avx LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax
DB 77,133,192 ; test %r8,%r8
- DB 15,133,253,0,0,0 ; jne 4e7e <_sk_load_u16_be_avx+0x113>
+ DB 15,133,253,0,0,0 ; jne 5052 <_sk_load_u16_be_avx+0x113>
DB 196,65,121,16,4,65 ; vmovupd (%r9,%rax,2),%xmm8
DB 196,193,121,16,84,65,16 ; vmovupd 0x10(%r9,%rax,2),%xmm2
DB 196,193,121,16,92,65,32 ; vmovupd 0x20(%r9,%rax,2),%xmm3
@@ -9566,7 +9849,7 @@ _sk_load_u16_be_avx LABEL PROC
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,29,110,27,0,0 ; vbroadcastss 0x1b6e(%rip),%ymm11 # 6960 <_sk_callback_avx+0x45c>
+ DB 196,98,125,24,29,110,27,0,0 ; vbroadcastss 0x1b6e(%rip),%ymm11 # 6b34 <_sk_callback_avx+0x45c>
DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0
DB 197,177,109,202 ; vpunpckhqdq %xmm2,%xmm9,%xmm1
DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2
@@ -9600,29 +9883,29 @@ _sk_load_u16_be_avx LABEL PROC
DB 196,65,123,16,4,65 ; vmovsd (%r9,%rax,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,85 ; je 4ee4 <_sk_load_u16_be_avx+0x179>
+ DB 116,85 ; je 50b8 <_sk_load_u16_be_avx+0x179>
DB 196,65,57,22,68,65,8 ; vmovhpd 0x8(%r9,%rax,2),%xmm8,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,72 ; jb 4ee4 <_sk_load_u16_be_avx+0x179>
+ DB 114,72 ; jb 50b8 <_sk_load_u16_be_avx+0x179>
DB 196,193,123,16,84,65,16 ; vmovsd 0x10(%r9,%rax,2),%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 116,72 ; je 4ef1 <_sk_load_u16_be_avx+0x186>
+ DB 116,72 ; je 50c5 <_sk_load_u16_be_avx+0x186>
DB 196,193,105,22,84,65,24 ; vmovhpd 0x18(%r9,%rax,2),%xmm2,%xmm2
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,59 ; jb 4ef1 <_sk_load_u16_be_avx+0x186>
+ DB 114,59 ; jb 50c5 <_sk_load_u16_be_avx+0x186>
DB 196,193,123,16,92,65,32 ; vmovsd 0x20(%r9,%rax,2),%xmm3
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 15,132,213,254,255,255 ; je 4d9c <_sk_load_u16_be_avx+0x31>
+ DB 15,132,213,254,255,255 ; je 4f70 <_sk_load_u16_be_avx+0x31>
DB 196,193,97,22,92,65,40 ; vmovhpd 0x28(%r9,%rax,2),%xmm3,%xmm3
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 15,130,196,254,255,255 ; jb 4d9c <_sk_load_u16_be_avx+0x31>
+ DB 15,130,196,254,255,255 ; jb 4f70 <_sk_load_u16_be_avx+0x31>
DB 196,65,122,126,76,65,48 ; vmovq 0x30(%r9,%rax,2),%xmm9
- DB 233,184,254,255,255 ; jmpq 4d9c <_sk_load_u16_be_avx+0x31>
+ DB 233,184,254,255,255 ; jmpq 4f70 <_sk_load_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,171,254,255,255 ; jmpq 4d9c <_sk_load_u16_be_avx+0x31>
+ DB 233,171,254,255,255 ; jmpq 4f70 <_sk_load_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,162,254,255,255 ; jmpq 4d9c <_sk_load_u16_be_avx+0x31>
+ DB 233,162,254,255,255 ; jmpq 4f70 <_sk_load_u16_be_avx+0x31>
PUBLIC _sk_load_rgb_u16_be_avx
_sk_load_rgb_u16_be_avx LABEL PROC
@@ -9630,7 +9913,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 72,141,4,82 ; lea (%rdx,%rdx,2),%rax
DB 77,133,192 ; test %r8,%r8
- DB 15,133,243,0,0,0 ; jne 4fff <_sk_load_rgb_u16_be_avx+0x105>
+ DB 15,133,243,0,0,0 ; jne 51d3 <_sk_load_rgb_u16_be_avx+0x105>
DB 196,193,122,111,4,65 ; vmovdqu (%r9,%rax,2),%xmm0
DB 196,193,122,111,84,65,12 ; vmovdqu 0xc(%r9,%rax,2),%xmm2
DB 196,193,122,111,76,65,24 ; vmovdqu 0x18(%r9,%rax,2),%xmm1
@@ -9657,7 +9940,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,29,206,25,0,0 ; vbroadcastss 0x19ce(%rip),%ymm11 # 6964 <_sk_callback_avx+0x460>
+ DB 196,98,125,24,29,206,25,0,0 ; vbroadcastss 0x19ce(%rip),%ymm11 # 6b38 <_sk_callback_avx+0x460>
DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0
DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1
DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2
@@ -9678,48 +9961,48 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,107,25,0,0 ; vbroadcastss 0x196b(%rip),%ymm3 # 6968 <_sk_callback_avx+0x464>
+ DB 196,226,125,24,29,107,25,0,0 ; vbroadcastss 0x196b(%rip),%ymm3 # 6b3c <_sk_callback_avx+0x464>
DB 255,224 ; jmpq *%rax
DB 196,193,121,110,4,65 ; vmovd (%r9,%rax,2),%xmm0
DB 196,193,121,196,68,65,4,2 ; vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 117,5 ; jne 5018 <_sk_load_rgb_u16_be_avx+0x11e>
- DB 233,40,255,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,5 ; jne 51ec <_sk_load_rgb_u16_be_avx+0x11e>
+ DB 233,40,255,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,76,65,6 ; vmovd 0x6(%r9,%rax,2),%xmm1
DB 196,65,113,196,68,65,10,2 ; vpinsrw $0x2,0xa(%r9,%rax,2),%xmm1,%xmm8
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,26 ; jb 5047 <_sk_load_rgb_u16_be_avx+0x14d>
+ DB 114,26 ; jb 521b <_sk_load_rgb_u16_be_avx+0x14d>
DB 196,193,121,110,76,65,12 ; vmovd 0xc(%r9,%rax,2),%xmm1
DB 196,193,113,196,84,65,16,2 ; vpinsrw $0x2,0x10(%r9,%rax,2),%xmm1,%xmm2
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 117,10 ; jne 504c <_sk_load_rgb_u16_be_avx+0x152>
- DB 233,249,254,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,244,254,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 5220 <_sk_load_rgb_u16_be_avx+0x152>
+ DB 233,249,254,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,244,254,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,76,65,18 ; vmovd 0x12(%r9,%rax,2),%xmm1
DB 196,65,113,196,76,65,22,2 ; vpinsrw $0x2,0x16(%r9,%rax,2),%xmm1,%xmm9
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,26 ; jb 507b <_sk_load_rgb_u16_be_avx+0x181>
+ DB 114,26 ; jb 524f <_sk_load_rgb_u16_be_avx+0x181>
DB 196,193,121,110,76,65,24 ; vmovd 0x18(%r9,%rax,2),%xmm1
DB 196,193,113,196,76,65,28,2 ; vpinsrw $0x2,0x1c(%r9,%rax,2),%xmm1,%xmm1
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 117,10 ; jne 5080 <_sk_load_rgb_u16_be_avx+0x186>
- DB 233,197,254,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,192,254,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 5254 <_sk_load_rgb_u16_be_avx+0x186>
+ DB 233,197,254,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,192,254,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,92,65,30 ; vmovd 0x1e(%r9,%rax,2),%xmm3
DB 196,65,97,196,92,65,34,2 ; vpinsrw $0x2,0x22(%r9,%rax,2),%xmm3,%xmm11
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,20 ; jb 50a9 <_sk_load_rgb_u16_be_avx+0x1af>
+ DB 114,20 ; jb 527d <_sk_load_rgb_u16_be_avx+0x1af>
DB 196,193,121,110,92,65,36 ; vmovd 0x24(%r9,%rax,2),%xmm3
DB 196,193,97,196,92,65,40,2 ; vpinsrw $0x2,0x28(%r9,%rax,2),%xmm3,%xmm3
- DB 233,151,254,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,146,254,255,255 ; jmpq 4f40 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,151,254,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,146,254,255,255 ; jmpq 5114 <_sk_load_rgb_u16_be_avx+0x46>
PUBLIC _sk_store_u16_be_avx
_sk_store_u16_be_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,8 ; mov (%rax),%r9
DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax
- DB 196,98,125,24,5,168,24,0,0 ; vbroadcastss 0x18a8(%rip),%ymm8 # 696c <_sk_callback_avx+0x468>
+ DB 196,98,125,24,5,168,24,0,0 ; vbroadcastss 0x18a8(%rip),%ymm8 # 6b40 <_sk_callback_avx+0x468>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10
@@ -9757,7 +10040,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,31 ; jne 51a8 <_sk_store_u16_be_avx+0xfa>
+ DB 117,31 ; jne 537c <_sk_store_u16_be_avx+0xfa>
DB 196,65,120,17,28,65 ; vmovups %xmm11,(%r9,%rax,2)
DB 196,65,120,17,84,65,16 ; vmovups %xmm10,0x10(%r9,%rax,2)
DB 196,65,120,17,76,65,32 ; vmovups %xmm9,0x20(%r9,%rax,2)
@@ -9766,31 +10049,31 @@ _sk_store_u16_be_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,214,28,65 ; vmovq %xmm11,(%r9,%rax,2)
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,240 ; je 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 116,240 ; je 5378 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,23,92,65,8 ; vmovhpd %xmm11,0x8(%r9,%rax,2)
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,227 ; jb 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 114,227 ; jb 5378 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,84,65,16 ; vmovq %xmm10,0x10(%r9,%rax,2)
- DB 116,218 ; je 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 116,218 ; je 5378 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,23,84,65,24 ; vmovhpd %xmm10,0x18(%r9,%rax,2)
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,205 ; jb 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 114,205 ; jb 5378 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,76,65,32 ; vmovq %xmm9,0x20(%r9,%rax,2)
- DB 116,196 ; je 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 116,196 ; je 5378 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,23,76,65,40 ; vmovhpd %xmm9,0x28(%r9,%rax,2)
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,183 ; jb 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 114,183 ; jb 5378 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,68,65,48 ; vmovq %xmm8,0x30(%r9,%rax,2)
- DB 235,174 ; jmp 51a4 <_sk_store_u16_be_avx+0xf6>
+ DB 235,174 ; jmp 5378 <_sk_store_u16_be_avx+0xf6>
PUBLIC _sk_load_f32_avx
_sk_load_f32_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 119,110 ; ja 526c <_sk_load_f32_avx+0x76>
+ DB 119,110 ; ja 5440 <_sk_load_f32_avx+0x76>
DB 76,139,8 ; mov (%rax),%r9
DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10
- DB 76,141,29,132,0,0,0 ; lea 0x84(%rip),%r11 # 5294 <_sk_load_f32_avx+0x9e>
+ DB 76,141,29,132,0,0,0 ; lea 0x84(%rip),%r11 # 5468 <_sk_load_f32_avx+0x9e>
DB 75,99,4,131 ; movslq (%r11,%r8,4),%rax
DB 76,1,216 ; add %r11,%rax
DB 255,224 ; jmpq *%rax
@@ -9847,7 +10130,7 @@ _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 77,133,192 ; test %r8,%r8
- DB 117,55 ; jne 5321 <_sk_store_f32_avx+0x6d>
+ DB 117,55 ; jne 54f5 <_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
@@ -9860,22 +10143,22 @@ _sk_store_f32_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,17,20,129 ; vmovupd %xmm10,(%r9,%rax,4)
DB 73,131,248,1 ; cmp $0x1,%r8
- DB 116,240 ; je 531d <_sk_store_f32_avx+0x69>
+ DB 116,240 ; je 54f1 <_sk_store_f32_avx+0x69>
DB 196,65,121,17,76,129,16 ; vmovupd %xmm9,0x10(%r9,%rax,4)
DB 73,131,248,3 ; cmp $0x3,%r8
- DB 114,227 ; jb 531d <_sk_store_f32_avx+0x69>
+ DB 114,227 ; jb 54f1 <_sk_store_f32_avx+0x69>
DB 196,65,121,17,68,129,32 ; vmovupd %xmm8,0x20(%r9,%rax,4)
- DB 116,218 ; je 531d <_sk_store_f32_avx+0x69>
+ DB 116,218 ; je 54f1 <_sk_store_f32_avx+0x69>
DB 196,65,121,17,92,129,48 ; vmovupd %xmm11,0x30(%r9,%rax,4)
DB 73,131,248,5 ; cmp $0x5,%r8
- DB 114,205 ; jb 531d <_sk_store_f32_avx+0x69>
+ DB 114,205 ; jb 54f1 <_sk_store_f32_avx+0x69>
DB 196,67,125,25,84,129,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r9,%rax,4)
- DB 116,195 ; je 531d <_sk_store_f32_avx+0x69>
+ DB 116,195 ; je 54f1 <_sk_store_f32_avx+0x69>
DB 196,67,125,25,76,129,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r9,%rax,4)
DB 73,131,248,7 ; cmp $0x7,%r8
- DB 114,181 ; jb 531d <_sk_store_f32_avx+0x69>
+ DB 114,181 ; jb 54f1 <_sk_store_f32_avx+0x69>
DB 196,67,125,25,68,129,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r9,%rax,4)
- DB 235,171 ; jmp 531d <_sk_store_f32_avx+0x69>
+ DB 235,171 ; jmp 54f1 <_sk_store_f32_avx+0x69>
PUBLIC _sk_clamp_x_avx
_sk_clamp_x_avx LABEL PROC
@@ -9965,7 +10248,7 @@ PUBLIC _sk_clamp_x_1_avx
_sk_clamp_x_1_avx LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,218,20,0,0 ; vbroadcastss 0x14da(%rip),%ymm8 # 6970 <_sk_callback_avx+0x46c>
+ DB 196,98,125,24,5,218,20,0,0 ; vbroadcastss 0x14da(%rip),%ymm8 # 6b44 <_sk_callback_avx+0x46c>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -9979,9 +10262,9 @@ _sk_repeat_x_1_avx LABEL PROC
PUBLIC _sk_mirror_x_1_avx
_sk_mirror_x_1_avx LABEL PROC
- DB 196,98,125,24,5,189,20,0,0 ; vbroadcastss 0x14bd(%rip),%ymm8 # 6974 <_sk_callback_avx+0x470>
+ DB 196,98,125,24,5,189,20,0,0 ; vbroadcastss 0x14bd(%rip),%ymm8 # 6b48 <_sk_callback_avx+0x470>
DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,13,179,20,0,0 ; vbroadcastss 0x14b3(%rip),%ymm9 # 6978 <_sk_callback_avx+0x474>
+ DB 196,98,125,24,13,179,20,0,0 ; vbroadcastss 0x14b3(%rip),%ymm9 # 6b4c <_sk_callback_avx+0x474>
DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9
DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9
DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9
@@ -9995,12 +10278,12 @@ _sk_mirror_x_1_avx LABEL PROC
PUBLIC _sk_luminance_to_alpha_avx
_sk_luminance_to_alpha_avx LABEL PROC
- DB 196,226,125,24,29,131,20,0,0 ; vbroadcastss 0x1483(%rip),%ymm3 # 697c <_sk_callback_avx+0x478>
+ DB 196,226,125,24,29,131,20,0,0 ; vbroadcastss 0x1483(%rip),%ymm3 # 6b50 <_sk_callback_avx+0x478>
DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0
- DB 196,226,125,24,29,122,20,0,0 ; vbroadcastss 0x147a(%rip),%ymm3 # 6980 <_sk_callback_avx+0x47c>
+ DB 196,226,125,24,29,122,20,0,0 ; vbroadcastss 0x147a(%rip),%ymm3 # 6b54 <_sk_callback_avx+0x47c>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,109,20,0,0 ; vbroadcastss 0x146d(%rip),%ymm1 # 6984 <_sk_callback_avx+0x480>
+ DB 196,226,125,24,13,109,20,0,0 ; vbroadcastss 0x146d(%rip),%ymm1 # 6b58 <_sk_callback_avx+0x480>
DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1
DB 197,252,88,217 ; vaddps %ymm1,%ymm0,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10209,9 +10492,9 @@ _sk_evenly_spaced_gradient_avx LABEL PROC
DB 72,139,24 ; mov (%rax),%rbx
DB 72,139,104,8 ; mov 0x8(%rax),%rbp
DB 72,255,203 ; dec %rbx
- DB 120,7 ; js 5876 <_sk_evenly_spaced_gradient_avx+0x28>
+ DB 120,7 ; js 5a4a <_sk_evenly_spaced_gradient_avx+0x28>
DB 196,225,242,42,203 ; vcvtsi2ss %rbx,%xmm1,%xmm1
- DB 235,21 ; jmp 588b <_sk_evenly_spaced_gradient_avx+0x3d>
+ DB 235,21 ; jmp 5a5f <_sk_evenly_spaced_gradient_avx+0x3d>
DB 73,137,217 ; mov %rbx,%r9
DB 73,209,233 ; shr %r9
DB 131,227,1 ; and $0x1,%ebx
@@ -10368,18 +10651,18 @@ _sk_evenly_spaced_gradient_avx LABEL PROC
PUBLIC _sk_gauss_a_to_rgba_avx
_sk_gauss_a_to_rgba_avx LABEL PROC
- DB 196,226,125,24,5,150,13,0,0 ; vbroadcastss 0xd96(%rip),%ymm0 # 6988 <_sk_callback_avx+0x484>
+ DB 196,226,125,24,5,150,13,0,0 ; vbroadcastss 0xd96(%rip),%ymm0 # 6b5c <_sk_callback_avx+0x484>
DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0
- DB 196,226,125,24,13,141,13,0,0 ; vbroadcastss 0xd8d(%rip),%ymm1 # 698c <_sk_callback_avx+0x488>
+ DB 196,226,125,24,13,141,13,0,0 ; vbroadcastss 0xd8d(%rip),%ymm1 # 6b60 <_sk_callback_avx+0x488>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0
- DB 196,226,125,24,13,128,13,0,0 ; vbroadcastss 0xd80(%rip),%ymm1 # 6990 <_sk_callback_avx+0x48c>
+ DB 196,226,125,24,13,128,13,0,0 ; vbroadcastss 0xd80(%rip),%ymm1 # 6b64 <_sk_callback_avx+0x48c>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0
- DB 196,226,125,24,13,115,13,0,0 ; vbroadcastss 0xd73(%rip),%ymm1 # 6994 <_sk_callback_avx+0x490>
+ DB 196,226,125,24,13,115,13,0,0 ; vbroadcastss 0xd73(%rip),%ymm1 # 6b68 <_sk_callback_avx+0x490>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0
- DB 196,226,125,24,13,102,13,0,0 ; vbroadcastss 0xd66(%rip),%ymm1 # 6998 <_sk_callback_avx+0x494>
+ DB 196,226,125,24,13,102,13,0,0 ; vbroadcastss 0xd66(%rip),%ymm1 # 6b6c <_sk_callback_avx+0x494>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
@@ -10401,12 +10684,12 @@ _sk_gradient_avx LABEL PROC
DB 76,139,8 ; mov (%rax),%r9
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 73,131,249,2 ; cmp $0x2,%r9
- DB 114,80 ; jb 5cb8 <_sk_gradient_avx+0x72>
+ DB 114,80 ; jb 5e8c <_sk_gradient_avx+0x72>
DB 72,139,88,72 ; mov 0x48(%rax),%rbx
DB 73,255,201 ; dec %r9
DB 72,131,195,4 ; add $0x4,%rbx
DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9
- DB 196,98,125,24,21,27,13,0,0 ; vbroadcastss 0xd1b(%rip),%ymm10 # 699c <_sk_callback_avx+0x498>
+ DB 196,98,125,24,21,27,13,0,0 ; vbroadcastss 0xd1b(%rip),%ymm10 # 6b70 <_sk_callback_avx+0x498>
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 196,98,125,24,3 ; vbroadcastss (%rbx),%ymm8
DB 197,60,194,192,2 ; vcmpleps %ymm0,%ymm8,%ymm8
@@ -10418,7 +10701,7 @@ _sk_gradient_avx LABEL PROC
DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
DB 72,131,195,4 ; add $0x4,%rbx
DB 73,255,201 ; dec %r9
- DB 117,205 ; jne 5c85 <_sk_gradient_avx+0x3f>
+ DB 117,205 ; jne 5e59 <_sk_gradient_avx+0x3f>
DB 196,195,249,22,201,1 ; vpextrq $0x1,%xmm1,%r9
DB 69,137,202 ; mov %r9d,%r10d
DB 73,193,233,32 ; shr $0x20,%r9
@@ -10598,27 +10881,27 @@ _sk_xy_to_unit_angle_avx LABEL PROC
DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12
DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11
DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12
- DB 196,98,125,24,45,14,9,0,0 ; vbroadcastss 0x90e(%rip),%ymm13 # 69a0 <_sk_callback_avx+0x49c>
+ DB 196,98,125,24,45,14,9,0,0 ; vbroadcastss 0x90e(%rip),%ymm13 # 6b74 <_sk_callback_avx+0x49c>
DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13
- DB 196,98,125,24,53,4,9,0,0 ; vbroadcastss 0x904(%rip),%ymm14 # 69a4 <_sk_callback_avx+0x4a0>
+ DB 196,98,125,24,53,4,9,0,0 ; vbroadcastss 0x904(%rip),%ymm14 # 6b78 <_sk_callback_avx+0x4a0>
DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13
DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13
- DB 196,98,125,24,53,245,8,0,0 ; vbroadcastss 0x8f5(%rip),%ymm14 # 69a8 <_sk_callback_avx+0x4a4>
+ DB 196,98,125,24,53,245,8,0,0 ; vbroadcastss 0x8f5(%rip),%ymm14 # 6b7c <_sk_callback_avx+0x4a4>
DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13
DB 196,65,28,89,229 ; vmulps %ymm13,%ymm12,%ymm12
- DB 196,98,125,24,45,230,8,0,0 ; vbroadcastss 0x8e6(%rip),%ymm13 # 69ac <_sk_callback_avx+0x4a8>
+ DB 196,98,125,24,45,230,8,0,0 ; vbroadcastss 0x8e6(%rip),%ymm13 # 6b80 <_sk_callback_avx+0x4a8>
DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9
- DB 196,98,125,24,21,209,8,0,0 ; vbroadcastss 0x8d1(%rip),%ymm10 # 69b0 <_sk_callback_avx+0x4ac>
+ DB 196,98,125,24,21,209,8,0,0 ; vbroadcastss 0x8d1(%rip),%ymm10 # 6b84 <_sk_callback_avx+0x4ac>
DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10
DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9
DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,21,187,8,0,0 ; vbroadcastss 0x8bb(%rip),%ymm10 # 69b4 <_sk_callback_avx+0x4b0>
+ DB 196,98,125,24,21,187,8,0,0 ; vbroadcastss 0x8bb(%rip),%ymm10 # 6b88 <_sk_callback_avx+0x4b0>
DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10
DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0
DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9
- DB 196,98,125,24,21,165,8,0,0 ; vbroadcastss 0x8a5(%rip),%ymm10 # 69b8 <_sk_callback_avx+0x4b4>
+ DB 196,98,125,24,21,165,8,0,0 ; vbroadcastss 0x8a5(%rip),%ymm10 # 6b8c <_sk_callback_avx+0x4b4>
DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10
DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0
DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9
@@ -10638,7 +10921,7 @@ _sk_xy_to_radius_avx LABEL PROC
PUBLIC _sk_save_xy_avx
_sk_save_xy_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,111,8,0,0 ; vbroadcastss 0x86f(%rip),%ymm8 # 69bc <_sk_callback_avx+0x4b8>
+ DB 196,98,125,24,5,111,8,0,0 ; vbroadcastss 0x86f(%rip),%ymm8 # 6b90 <_sk_callback_avx+0x4b8>
DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9
DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10
DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9
@@ -10671,9 +10954,9 @@ _sk_accumulate_avx LABEL PROC
PUBLIC _sk_bilinear_nx_avx
_sk_bilinear_nx_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,251,7,0,0 ; vbroadcastss 0x7fb(%rip),%ymm0 # 69c0 <_sk_callback_avx+0x4bc>
+ DB 196,226,125,24,5,251,7,0,0 ; vbroadcastss 0x7fb(%rip),%ymm0 # 6b94 <_sk_callback_avx+0x4bc>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,242,7,0,0 ; vbroadcastss 0x7f2(%rip),%ymm8 # 69c4 <_sk_callback_avx+0x4c0>
+ DB 196,98,125,24,5,242,7,0,0 ; vbroadcastss 0x7f2(%rip),%ymm8 # 6b98 <_sk_callback_avx+0x4c0>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10682,7 +10965,7 @@ _sk_bilinear_nx_avx LABEL PROC
PUBLIC _sk_bilinear_px_avx
_sk_bilinear_px_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,218,7,0,0 ; vbroadcastss 0x7da(%rip),%ymm0 # 69c8 <_sk_callback_avx+0x4c4>
+ DB 196,226,125,24,5,218,7,0,0 ; vbroadcastss 0x7da(%rip),%ymm0 # 6b9c <_sk_callback_avx+0x4c4>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
@@ -10692,9 +10975,9 @@ _sk_bilinear_px_avx LABEL PROC
PUBLIC _sk_bilinear_ny_avx
_sk_bilinear_ny_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,190,7,0,0 ; vbroadcastss 0x7be(%rip),%ymm1 # 69cc <_sk_callback_avx+0x4c8>
+ DB 196,226,125,24,13,190,7,0,0 ; vbroadcastss 0x7be(%rip),%ymm1 # 6ba0 <_sk_callback_avx+0x4c8>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,180,7,0,0 ; vbroadcastss 0x7b4(%rip),%ymm8 # 69d0 <_sk_callback_avx+0x4cc>
+ DB 196,98,125,24,5,180,7,0,0 ; vbroadcastss 0x7b4(%rip),%ymm8 # 6ba4 <_sk_callback_avx+0x4cc>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10703,7 +10986,7 @@ _sk_bilinear_ny_avx LABEL PROC
PUBLIC _sk_bilinear_py_avx
_sk_bilinear_py_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,156,7,0,0 ; vbroadcastss 0x79c(%rip),%ymm1 # 69d4 <_sk_callback_avx+0x4d0>
+ DB 196,226,125,24,13,156,7,0,0 ; vbroadcastss 0x79c(%rip),%ymm1 # 6ba8 <_sk_callback_avx+0x4d0>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
@@ -10713,14 +10996,14 @@ _sk_bilinear_py_avx LABEL PROC
PUBLIC _sk_bicubic_n3x_avx
_sk_bicubic_n3x_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,127,7,0,0 ; vbroadcastss 0x77f(%rip),%ymm0 # 69d8 <_sk_callback_avx+0x4d4>
+ DB 196,226,125,24,5,127,7,0,0 ; vbroadcastss 0x77f(%rip),%ymm0 # 6bac <_sk_callback_avx+0x4d4>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,118,7,0,0 ; vbroadcastss 0x776(%rip),%ymm8 # 69dc <_sk_callback_avx+0x4d8>
+ DB 196,98,125,24,5,118,7,0,0 ; vbroadcastss 0x776(%rip),%ymm8 # 6bb0 <_sk_callback_avx+0x4d8>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,103,7,0,0 ; vbroadcastss 0x767(%rip),%ymm10 # 69e0 <_sk_callback_avx+0x4dc>
+ DB 196,98,125,24,21,103,7,0,0 ; vbroadcastss 0x767(%rip),%ymm10 # 6bb4 <_sk_callback_avx+0x4dc>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,93,7,0,0 ; vbroadcastss 0x75d(%rip),%ymm10 # 69e4 <_sk_callback_avx+0x4e0>
+ DB 196,98,125,24,21,93,7,0,0 ; vbroadcastss 0x75d(%rip),%ymm10 # 6bb8 <_sk_callback_avx+0x4e0>
DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
@@ -10730,19 +11013,19 @@ _sk_bicubic_n3x_avx LABEL PROC
PUBLIC _sk_bicubic_n1x_avx
_sk_bicubic_n1x_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,64,7,0,0 ; vbroadcastss 0x740(%rip),%ymm0 # 69e8 <_sk_callback_avx+0x4e4>
+ DB 196,226,125,24,5,64,7,0,0 ; vbroadcastss 0x740(%rip),%ymm0 # 6bbc <_sk_callback_avx+0x4e4>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,55,7,0,0 ; vbroadcastss 0x737(%rip),%ymm8 # 69ec <_sk_callback_avx+0x4e8>
+ DB 196,98,125,24,5,55,7,0,0 ; vbroadcastss 0x737(%rip),%ymm8 # 6bc0 <_sk_callback_avx+0x4e8>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,45,7,0,0 ; vbroadcastss 0x72d(%rip),%ymm9 # 69f0 <_sk_callback_avx+0x4ec>
+ DB 196,98,125,24,13,45,7,0,0 ; vbroadcastss 0x72d(%rip),%ymm9 # 6bc4 <_sk_callback_avx+0x4ec>
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,35,7,0,0 ; vbroadcastss 0x723(%rip),%ymm10 # 69f4 <_sk_callback_avx+0x4f0>
+ DB 196,98,125,24,21,35,7,0,0 ; vbroadcastss 0x723(%rip),%ymm10 # 6bc8 <_sk_callback_avx+0x4f0>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,20,7,0,0 ; vbroadcastss 0x714(%rip),%ymm10 # 69f8 <_sk_callback_avx+0x4f4>
+ DB 196,98,125,24,21,20,7,0,0 ; vbroadcastss 0x714(%rip),%ymm10 # 6bcc <_sk_callback_avx+0x4f4>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,5,7,0,0 ; vbroadcastss 0x705(%rip),%ymm9 # 69fc <_sk_callback_avx+0x4f8>
+ DB 196,98,125,24,13,5,7,0,0 ; vbroadcastss 0x705(%rip),%ymm9 # 6bd0 <_sk_callback_avx+0x4f8>
DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10751,17 +11034,17 @@ _sk_bicubic_n1x_avx LABEL PROC
PUBLIC _sk_bicubic_p1x_avx
_sk_bicubic_p1x_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,237,6,0,0 ; vbroadcastss 0x6ed(%rip),%ymm8 # 6a00 <_sk_callback_avx+0x4fc>
+ DB 196,98,125,24,5,237,6,0,0 ; vbroadcastss 0x6ed(%rip),%ymm8 # 6bd4 <_sk_callback_avx+0x4fc>
DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0
DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9
- DB 196,98,125,24,21,223,6,0,0 ; vbroadcastss 0x6df(%rip),%ymm10 # 6a04 <_sk_callback_avx+0x500>
+ DB 196,98,125,24,21,223,6,0,0 ; vbroadcastss 0x6df(%rip),%ymm10 # 6bd8 <_sk_callback_avx+0x500>
DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10
- DB 196,98,125,24,29,213,6,0,0 ; vbroadcastss 0x6d5(%rip),%ymm11 # 6a08 <_sk_callback_avx+0x504>
+ DB 196,98,125,24,29,213,6,0,0 ; vbroadcastss 0x6d5(%rip),%ymm11 # 6bdc <_sk_callback_avx+0x504>
DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10
DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10
DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
- DB 196,98,125,24,13,188,6,0,0 ; vbroadcastss 0x6bc(%rip),%ymm9 # 6a0c <_sk_callback_avx+0x508>
+ DB 196,98,125,24,13,188,6,0,0 ; vbroadcastss 0x6bc(%rip),%ymm9 # 6be0 <_sk_callback_avx+0x508>
DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10770,13 +11053,13 @@ _sk_bicubic_p1x_avx LABEL PROC
PUBLIC _sk_bicubic_p3x_avx
_sk_bicubic_p3x_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,164,6,0,0 ; vbroadcastss 0x6a4(%rip),%ymm0 # 6a10 <_sk_callback_avx+0x50c>
+ DB 196,226,125,24,5,164,6,0,0 ; vbroadcastss 0x6a4(%rip),%ymm0 # 6be4 <_sk_callback_avx+0x50c>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,145,6,0,0 ; vbroadcastss 0x691(%rip),%ymm10 # 6a14 <_sk_callback_avx+0x510>
+ DB 196,98,125,24,21,145,6,0,0 ; vbroadcastss 0x691(%rip),%ymm10 # 6be8 <_sk_callback_avx+0x510>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,135,6,0,0 ; vbroadcastss 0x687(%rip),%ymm10 # 6a18 <_sk_callback_avx+0x514>
+ DB 196,98,125,24,21,135,6,0,0 ; vbroadcastss 0x687(%rip),%ymm10 # 6bec <_sk_callback_avx+0x514>
DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax)
@@ -10786,14 +11069,14 @@ _sk_bicubic_p3x_avx LABEL PROC
PUBLIC _sk_bicubic_n3y_avx
_sk_bicubic_n3y_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,106,6,0,0 ; vbroadcastss 0x66a(%rip),%ymm1 # 6a1c <_sk_callback_avx+0x518>
+ DB 196,226,125,24,13,106,6,0,0 ; vbroadcastss 0x66a(%rip),%ymm1 # 6bf0 <_sk_callback_avx+0x518>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,96,6,0,0 ; vbroadcastss 0x660(%rip),%ymm8 # 6a20 <_sk_callback_avx+0x51c>
+ DB 196,98,125,24,5,96,6,0,0 ; vbroadcastss 0x660(%rip),%ymm8 # 6bf4 <_sk_callback_avx+0x51c>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,81,6,0,0 ; vbroadcastss 0x651(%rip),%ymm10 # 6a24 <_sk_callback_avx+0x520>
+ DB 196,98,125,24,21,81,6,0,0 ; vbroadcastss 0x651(%rip),%ymm10 # 6bf8 <_sk_callback_avx+0x520>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,71,6,0,0 ; vbroadcastss 0x647(%rip),%ymm10 # 6a28 <_sk_callback_avx+0x524>
+ DB 196,98,125,24,21,71,6,0,0 ; vbroadcastss 0x647(%rip),%ymm10 # 6bfc <_sk_callback_avx+0x524>
DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
@@ -10803,19 +11086,19 @@ _sk_bicubic_n3y_avx LABEL PROC
PUBLIC _sk_bicubic_n1y_avx
_sk_bicubic_n1y_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,42,6,0,0 ; vbroadcastss 0x62a(%rip),%ymm1 # 6a2c <_sk_callback_avx+0x528>
+ DB 196,226,125,24,13,42,6,0,0 ; vbroadcastss 0x62a(%rip),%ymm1 # 6c00 <_sk_callback_avx+0x528>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,32,6,0,0 ; vbroadcastss 0x620(%rip),%ymm8 # 6a30 <_sk_callback_avx+0x52c>
+ DB 196,98,125,24,5,32,6,0,0 ; vbroadcastss 0x620(%rip),%ymm8 # 6c04 <_sk_callback_avx+0x52c>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,22,6,0,0 ; vbroadcastss 0x616(%rip),%ymm9 # 6a34 <_sk_callback_avx+0x530>
+ DB 196,98,125,24,13,22,6,0,0 ; vbroadcastss 0x616(%rip),%ymm9 # 6c08 <_sk_callback_avx+0x530>
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,12,6,0,0 ; vbroadcastss 0x60c(%rip),%ymm10 # 6a38 <_sk_callback_avx+0x534>
+ DB 196,98,125,24,21,12,6,0,0 ; vbroadcastss 0x60c(%rip),%ymm10 # 6c0c <_sk_callback_avx+0x534>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,253,5,0,0 ; vbroadcastss 0x5fd(%rip),%ymm10 # 6a3c <_sk_callback_avx+0x538>
+ DB 196,98,125,24,21,253,5,0,0 ; vbroadcastss 0x5fd(%rip),%ymm10 # 6c10 <_sk_callback_avx+0x538>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,238,5,0,0 ; vbroadcastss 0x5ee(%rip),%ymm9 # 6a40 <_sk_callback_avx+0x53c>
+ DB 196,98,125,24,13,238,5,0,0 ; vbroadcastss 0x5ee(%rip),%ymm9 # 6c14 <_sk_callback_avx+0x53c>
DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10824,17 +11107,17 @@ _sk_bicubic_n1y_avx LABEL PROC
PUBLIC _sk_bicubic_p1y_avx
_sk_bicubic_p1y_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,214,5,0,0 ; vbroadcastss 0x5d6(%rip),%ymm8 # 6a44 <_sk_callback_avx+0x540>
+ DB 196,98,125,24,5,214,5,0,0 ; vbroadcastss 0x5d6(%rip),%ymm8 # 6c18 <_sk_callback_avx+0x540>
DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1
DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9
- DB 196,98,125,24,21,199,5,0,0 ; vbroadcastss 0x5c7(%rip),%ymm10 # 6a48 <_sk_callback_avx+0x544>
+ DB 196,98,125,24,21,199,5,0,0 ; vbroadcastss 0x5c7(%rip),%ymm10 # 6c1c <_sk_callback_avx+0x544>
DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10
- DB 196,98,125,24,29,189,5,0,0 ; vbroadcastss 0x5bd(%rip),%ymm11 # 6a4c <_sk_callback_avx+0x548>
+ DB 196,98,125,24,29,189,5,0,0 ; vbroadcastss 0x5bd(%rip),%ymm11 # 6c20 <_sk_callback_avx+0x548>
DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10
DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10
DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
- DB 196,98,125,24,13,164,5,0,0 ; vbroadcastss 0x5a4(%rip),%ymm9 # 6a50 <_sk_callback_avx+0x54c>
+ DB 196,98,125,24,13,164,5,0,0 ; vbroadcastss 0x5a4(%rip),%ymm9 # 6c24 <_sk_callback_avx+0x54c>
DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10843,13 +11126,13 @@ _sk_bicubic_p1y_avx LABEL PROC
PUBLIC _sk_bicubic_p3y_avx
_sk_bicubic_p3y_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,13,140,5,0,0 ; vbroadcastss 0x58c(%rip),%ymm1 # 6a54 <_sk_callback_avx+0x550>
+ DB 196,226,125,24,13,140,5,0,0 ; vbroadcastss 0x58c(%rip),%ymm1 # 6c28 <_sk_callback_avx+0x550>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,120,5,0,0 ; vbroadcastss 0x578(%rip),%ymm10 # 6a58 <_sk_callback_avx+0x554>
+ DB 196,98,125,24,21,120,5,0,0 ; vbroadcastss 0x578(%rip),%ymm10 # 6c2c <_sk_callback_avx+0x554>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,110,5,0,0 ; vbroadcastss 0x56e(%rip),%ymm10 # 6a5c <_sk_callback_avx+0x558>
+ DB 196,98,125,24,21,110,5,0,0 ; vbroadcastss 0x56e(%rip),%ymm10 # 6c30 <_sk_callback_avx+0x558>
DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8
DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8
DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax)
@@ -10971,25 +11254,25 @@ ALIGN 4
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 66e1 <.literal4+0xb1>
+ DB 71,225,61 ; rex.RXB loope 68b5 <.literal4+0xb1>
DB 0,0 ; add %al,(%rax)
DB 128,63,154 ; cmpb $0x9a,(%rdi)
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 66f1 <.literal4+0xc1>
+ DB 71,225,61 ; rex.RXB loope 68c5 <.literal4+0xc1>
DB 0,0 ; add %al,(%rax)
DB 128,63,154 ; cmpb $0x9a,(%rdi)
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 6701 <.literal4+0xd1>
+ DB 71,225,61 ; rex.RXB loope 68d5 <.literal4+0xd1>
DB 0,0 ; add %al,(%rax)
DB 128,63,154 ; cmpb $0x9a,(%rdi)
DB 153 ; cltd
DB 153 ; cltd
DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax
- DB 71,225,61 ; rex.RXB loope 6711 <.literal4+0xe1>
+ DB 71,225,61 ; rex.RXB loope 68e5 <.literal4+0xe1>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax)
@@ -11041,7 +11324,7 @@ ALIGN 4
DB 190,129,128,128,59 ; mov $0x3b808081,%esi
DB 129,128,128,59,0,248,0,0,8,33 ; addl $0x21080000,-0x7ffc480(%rax)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 6765 <.literal4+0x135>
+ DB 224,7 ; loopne 6939 <.literal4+0x135>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -11057,10 +11340,10 @@ ALIGN 4
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
DB 0,52,255 ; add %dh,(%rdi,%rdi,8)
DB 255 ; (bad)
- DB 127,0 ; jg 678c <.literal4+0x15c>
+ DB 127,0 ; jg 6960 <.literal4+0x15c>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 6805 <.literal4+0x1d5>
+ DB 119,115 ; ja 69d9 <.literal4+0x1d5>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -11074,10 +11357,10 @@ ALIGN 4
DB 0,128,63,0,0,0 ; add %al,0x3f(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 67c0 <.literal4+0x190>
+ DB 127,0 ; jg 6994 <.literal4+0x190>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 6839 <.literal4+0x209>
+ DB 119,115 ; ja 6a0d <.literal4+0x209>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -11091,10 +11374,10 @@ ALIGN 4
DB 0,128,63,0,0,0 ; add %al,0x3f(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 67f4 <.literal4+0x1c4>
+ DB 127,0 ; jg 69c8 <.literal4+0x1c4>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 686d <.literal4+0x23d>
+ DB 119,115 ; ja 6a41 <.literal4+0x23d>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -11108,10 +11391,10 @@ ALIGN 4
DB 0,128,63,0,0,0 ; add %al,0x3f(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 6828 <.literal4+0x1f8>
+ DB 127,0 ; jg 69fc <.literal4+0x1f8>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 68a1 <.literal4+0x271>
+ DB 119,115 ; ja 6a75 <.literal4+0x271>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -11124,7 +11407,7 @@ ALIGN 4
DB 0,75,0 ; add %cl,0x0(%rbx)
DB 0,128,63,0,0,200 ; add %al,-0x37ffffc1(%rax)
DB 66,0,0 ; rex.X add %al,(%rax)
- DB 127,67 ; jg 689f <.literal4+0x26f>
+ DB 127,67 ; jg 6a73 <.literal4+0x26f>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -11136,10 +11419,10 @@ ALIGN 4
DB 190,80,128,3,62 ; mov $0x3e038050,%esi
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 68bf <.literal4+0x28f>
+ DB 118,63 ; jbe 6a93 <.literal4+0x28f>
DB 246,64,83,63 ; testb $0x3f,0x53(%rax)
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 68d3 <.literal4+0x2a3>
+ DB 127,67 ; jg 6aa7 <.literal4+0x2a3>
DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax)
DB 128,59,0 ; cmpb $0x0,(%rbx)
DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax)
@@ -11148,7 +11431,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 68b5 <.literal4+0x285>
+ DB 224,7 ; loopne 6a89 <.literal4+0x285>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -11160,7 +11443,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 68d1 <.literal4+0x2a1>
+ DB 224,7 ; loopne 6aa5 <.literal4+0x2a1>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -11171,7 +11454,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 6926 <.literal4+0x2f6>
+ DB 124,66 ; jl 6afa <.literal4+0x2f6>
DB 0,240 ; add %dh,%al
DB 0,0 ; add %al,(%rax)
DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax)
@@ -11189,9 +11472,9 @@ ALIGN 4
DB 137,136,136,59,15,0 ; mov %ecx,0xf3b88(%rax)
DB 0,0 ; add %al,(%rax)
DB 137,136,136,61,0,0 ; mov %ecx,0x3d88(%rax)
- DB 112,65 ; jo 6969 <.literal4+0x339>
+ DB 112,65 ; jo 6b3d <.literal4+0x339>
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 6977 <.literal4+0x347>
+ DB 127,67 ; jg 6b4b <.literal4+0x347>
DB 0,128,0,0,0,0 ; add %al,0x0(%rax)
DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax)
DB 0,0 ; add %al,(%rax)
@@ -11207,7 +11490,7 @@ ALIGN 4
DB 0,128,55,0,0,128 ; add %al,-0x7fffffc9(%rax)
DB 63 ; (bad)
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 69b7 <.literal4+0x387>
+ DB 127,71 ; jg 6b8b <.literal4+0x387>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,191,0,0,0 ; add %al,0xbf(%rax)
@@ -11320,6 +11603,22 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 0,0 ; add %al,(%rax)
DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
DB 0,2 ; add %al,(%rdx)
DB 4,6 ; add $0x6,%al
DB 0,0 ; add %al,(%rax)
@@ -11352,6 +11651,22 @@ ALIGN 16
DB 255,0 ; incl (%rax)
DB 255,0 ; incl (%rax)
DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
DB 0,2 ; add %al,(%rdx)
DB 4,6 ; add $0x6,%al
DB 0,0 ; add %al,(%rax)
diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp
index 42f4b934f2..eae3cd84f1 100644
--- a/src/jumper/SkJumper_stages.cpp
+++ b/src/jumper/SkJumper_stages.cpp
@@ -129,25 +129,6 @@ SI void store(T* dst, V v, size_t tail) {
unaligned_store(dst, v);
}
-// This doesn't look strictly necessary, but without it Clang would generate load() using
-// compiler-generated constants that we can't support. This version doesn't need constants.
-#if defined(JUMPER) && defined(__AVX__)
- template <>
- inline U8 load(const uint8_t* src, size_t tail) {
- if (__builtin_expect(tail, 0)) {
- uint64_t v = 0;
- size_t shift = 0;
- #pragma nounroll
- while (tail --> 0) {
- v |= (uint64_t)*src++ << shift;
- shift += 8;
- }
- return unaligned_load<U8>(&v);
- }
- return unaligned_load<U8>(src);
- }
-#endif
-
// AVX adds some mask loads and stores that make for shorter, faster code.
#if defined(JUMPER) && defined(__AVX__)
SI U32 mask(size_t tail) {