fix SkJumper radial gradient precision
authorMike Klein <mtklein@chromium.org>
Mon, 15 May 2017 19:55:54 +0000 (15:55 -0400)
committerSkia Commit-Bot <skia-commit-bot@chromium.org>
Mon, 15 May 2017 20:37:16 +0000 (20:37 +0000)
rcp(rsqrt(x)) doesn't have enough precision when x is a coordinate.
(It's fine when x is a color, like in the softlight blend mode.)

Adds a GM to test this.  It used to look quite ugly.

Change-Id: Icec295c2e2f50ae7a5e3e33c62270f632a58f65c
Reviewed-on: https://skia-review.googlesource.com/16914
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>

gm/radial_gradient_precision.cpp [new file with mode: 0644]
gn/gm.gni
src/jumper/SkJumper_generated.S
src/jumper/SkJumper_generated_win.S
src/jumper/SkJumper_stages.cpp
src/jumper/SkJumper_vectors.h

diff --git a/gm/radial_gradient_precision.cpp b/gm/radial_gradient_precision.cpp
new file mode 100644 (file)
index 0000000..b4e9a11
--- /dev/null
@@ -0,0 +1,22 @@
+/*
+ * Copyright 2017 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "gm.h"
+#include "SkGradientShader.h"
+
+// All we're looking for here is that we see a smooth gradient.
+DEF_SIMPLE_GM(radial_gradient_precision, canvas, 200, 200) {
+    SkPoint  center   = {100000, 100000};
+    SkScalar radius   = 40;
+    SkColor  colors[] = {SK_ColorBLACK, SK_ColorWHITE};
+
+    SkPaint p;
+    p.setShader(SkGradientShader::MakeRadial(center, radius,
+                                             colors, nullptr, SK_ARRAY_COUNT(colors),
+                                             SkShader::kRepeat_TileMode));
+    canvas->drawPaint(p);
+}
index 4bcd103..2e1c8af 100644 (file)
--- a/gn/gm.gni
+++ b/gn/gm.gni
@@ -235,6 +235,7 @@ gm_sources = [
   "$_gm/poly2poly.cpp",
   "$_gm/polygons.cpp",
   "$_gm/quadpaths.cpp",
+  "$_gm/radial_gradient_precision.cpp",
   "$_gm/readpixels.cpp",
   "$_gm/recordopts.cpp",
   "$_gm/rectangletexture.cpp",
index 82dd6b3..a5a5add 100644 (file)
@@ -3809,16 +3809,10 @@ HIDDEN _sk_xy_to_radius_aarch64
 .globl _sk_xy_to_radius_aarch64
 FUNCTION(_sk_xy_to_radius_aarch64)
 _sk_xy_to_radius_aarch64:
+  .long  0xf8408423                          // ldr           x3, [x1], #8
   .long  0x6e21dc30                          // fmul          v16.4s, v1.4s, v1.4s
   .long  0x4e20cc10                          // fmla          v16.4s, v0.4s, v0.4s
-  .long  0x6ea1da00                          // frsqrte       v0.4s, v16.4s
-  .long  0x6e20dc11                          // fmul          v17.4s, v0.4s, v0.4s
-  .long  0x4eb1fe10                          // frsqrts       v16.4s, v16.4s, v17.4s
-  .long  0x6e30dc00                          // fmul          v0.4s, v0.4s, v16.4s
-  .long  0xf8408423                          // ldr           x3, [x1], #8
-  .long  0x4ea1d810                          // frecpe        v16.4s, v0.4s
-  .long  0x4e30fc00                          // frecps        v0.4s, v0.4s, v16.4s
-  .long  0x6e20de00                          // fmul          v0.4s, v16.4s, v0.4s
+  .long  0x6ea1fa00                          // fsqrt         v0.4s, v16.4s
   .long  0xd61f0060                          // br            x3
 
 HIDDEN _sk_save_xy_aarch64
@@ -8284,12 +8278,14 @@ _sk_xy_to_radius_vfp4:
   .long  0xf2410da0                          // vadd.f32      d16, d17, d16
   .long  0xf3fb15a0                          // vrsqrte.f32   d17, d16
   .long  0xf3412db1                          // vmul.f32      d18, d17, d17
-  .long  0xf2600fb2                          // vrsqrts.f32   d16, d16, d18
-  .long  0xf3410db0                          // vmul.f32      d16, d17, d16
-  .long  0xf3fb1520                          // vrecpe.f32    d17, d16
-  .long  0xf2400fb1                          // vrecps.f32    d16, d16, d17
-  .long  0xf3010db0                          // vmul.f32      d0, d17, d16
+  .long  0xf2602fb2                          // vrsqrts.f32   d18, d16, d18
+  .long  0xf3411db2                          // vmul.f32      d17, d17, d18
+  .long  0xf3412db1                          // vmul.f32      d18, d17, d17
+  .long  0xf2602fb2                          // vrsqrts.f32   d18, d16, d18
+  .long  0xf3411db2                          // vmul.f32      d17, d17, d18
+  .long  0xf3000db1                          // vmul.f32      d0, d16, d17
   .long  0xe12fff13                          // bx            r3
+  .long  0xe320f000                          // nop           {0}
 
 HIDDEN _sk_save_xy_vfp4
 .globl _sk_save_xy_vfp4
@@ -8726,14 +8722,14 @@ _sk_seed_shader_hsw:
   .byte  197,249,110,199                     // vmovd         %edi,%xmm0
   .byte  196,226,125,88,192                  // vpbroadcastd  %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,189,70,0,0        // vbroadcastss  0x46bd(%rip),%ymm1        # 4780 <_sk_callback_hsw+0x128>
+  .byte  196,226,125,24,13,185,70,0,0        // vbroadcastss  0x46b9(%rip),%ymm1        # 477c <_sk_callback_hsw+0x128>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  197,252,88,2                        // vaddps        (%rdx),%ymm0,%ymm0
   .byte  196,226,125,24,16                   // vbroadcastss  (%rax),%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  197,236,88,201                      // vaddps        %ymm1,%ymm2,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,21,161,70,0,0        // vbroadcastss  0x46a1(%rip),%ymm2        # 4784 <_sk_callback_hsw+0x12c>
+  .byte  196,226,125,24,21,157,70,0,0        // vbroadcastss  0x469d(%rip),%ymm2        # 4780 <_sk_callback_hsw+0x12c>
   .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
@@ -8754,13 +8750,13 @@ _sk_dither_hsw:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  196,66,125,88,8                     // vpbroadcastd  (%r8),%ymm9
   .byte  196,65,61,239,201                   // vpxor         %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,88,21,96,70,0,0          // vpbroadcastd  0x4660(%rip),%ymm10        # 4788 <_sk_callback_hsw+0x130>
+  .byte  196,98,125,88,21,92,70,0,0          // vpbroadcastd  0x465c(%rip),%ymm10        # 4784 <_sk_callback_hsw+0x130>
   .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  196,98,125,88,37,69,70,0,0          // vpbroadcastd  0x4645(%rip),%ymm12        # 478c <_sk_callback_hsw+0x134>
-  .byte  196,98,125,88,45,64,70,0,0          // vpbroadcastd  0x4640(%rip),%ymm13        # 4790 <_sk_callback_hsw+0x138>
+  .byte  196,98,125,88,37,65,70,0,0          // vpbroadcastd  0x4641(%rip),%ymm12        # 4788 <_sk_callback_hsw+0x134>
+  .byte  196,98,125,88,45,60,70,0,0          // vpbroadcastd  0x463c(%rip),%ymm13        # 478c <_sk_callback_hsw+0x138>
   .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
@@ -8775,8 +8771,8 @@ _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  196,98,125,24,13,242,69,0,0         // vbroadcastss  0x45f2(%rip),%ymm9        # 4794 <_sk_callback_hsw+0x13c>
-  .byte  196,98,125,24,21,237,69,0,0         // vbroadcastss  0x45ed(%rip),%ymm10        # 4798 <_sk_callback_hsw+0x140>
+  .byte  196,98,125,24,13,238,69,0,0         // vbroadcastss  0x45ee(%rip),%ymm9        # 4790 <_sk_callback_hsw+0x13c>
+  .byte  196,98,125,24,21,233,69,0,0         // vbroadcastss  0x45e9(%rip),%ymm10        # 4794 <_sk_callback_hsw+0x140>
   .byte  196,66,61,184,209                   // vfmadd231ps   %ymm9,%ymm8,%ymm10
   .byte  196,98,125,24,64,8                  // vbroadcastss  0x8(%rax),%ymm8
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
@@ -8838,7 +8834,7 @@ HIDDEN _sk_srcatop_hsw
 FUNCTION(_sk_srcatop_hsw)
 _sk_srcatop_hsw:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,97,69,0,0           // vbroadcastss  0x4561(%rip),%ymm8        # 479c <_sk_callback_hsw+0x144>
+  .byte  196,98,125,24,5,93,69,0,0           // vbroadcastss  0x455d(%rip),%ymm8        # 4798 <_sk_callback_hsw+0x144>
   .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
@@ -8854,7 +8850,7 @@ HIDDEN _sk_dstatop_hsw
 .globl _sk_dstatop_hsw
 FUNCTION(_sk_dstatop_hsw)
 _sk_dstatop_hsw:
-  .byte  196,98,125,24,5,52,69,0,0           // vbroadcastss  0x4534(%rip),%ymm8        # 47a0 <_sk_callback_hsw+0x148>
+  .byte  196,98,125,24,5,48,69,0,0           // vbroadcastss  0x4530(%rip),%ymm8        # 479c <_sk_callback_hsw+0x148>
   .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
@@ -8893,7 +8889,7 @@ HIDDEN _sk_srcout_hsw
 .globl _sk_srcout_hsw
 FUNCTION(_sk_srcout_hsw)
 _sk_srcout_hsw:
-  .byte  196,98,125,24,5,219,68,0,0          // vbroadcastss  0x44db(%rip),%ymm8        # 47a4 <_sk_callback_hsw+0x14c>
+  .byte  196,98,125,24,5,215,68,0,0          // vbroadcastss  0x44d7(%rip),%ymm8        # 47a0 <_sk_callback_hsw+0x14c>
   .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
@@ -8906,7 +8902,7 @@ HIDDEN _sk_dstout_hsw
 .globl _sk_dstout_hsw
 FUNCTION(_sk_dstout_hsw)
 _sk_dstout_hsw:
-  .byte  196,226,125,24,5,190,68,0,0         // vbroadcastss  0x44be(%rip),%ymm0        # 47a8 <_sk_callback_hsw+0x150>
+  .byte  196,226,125,24,5,186,68,0,0         // vbroadcastss  0x44ba(%rip),%ymm0        # 47a4 <_sk_callback_hsw+0x150>
   .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
@@ -8919,7 +8915,7 @@ HIDDEN _sk_srcover_hsw
 .globl _sk_srcover_hsw
 FUNCTION(_sk_srcover_hsw)
 _sk_srcover_hsw:
-  .byte  196,98,125,24,5,161,68,0,0          // vbroadcastss  0x44a1(%rip),%ymm8        # 47ac <_sk_callback_hsw+0x154>
+  .byte  196,98,125,24,5,157,68,0,0          // vbroadcastss  0x449d(%rip),%ymm8        # 47a8 <_sk_callback_hsw+0x154>
   .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
@@ -8932,7 +8928,7 @@ HIDDEN _sk_dstover_hsw
 .globl _sk_dstover_hsw
 FUNCTION(_sk_dstover_hsw)
 _sk_dstover_hsw:
-  .byte  196,98,125,24,5,128,68,0,0          // vbroadcastss  0x4480(%rip),%ymm8        # 47b0 <_sk_callback_hsw+0x158>
+  .byte  196,98,125,24,5,124,68,0,0          // vbroadcastss  0x447c(%rip),%ymm8        # 47ac <_sk_callback_hsw+0x158>
   .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
@@ -8956,7 +8952,7 @@ HIDDEN _sk_multiply_hsw
 .globl _sk_multiply_hsw
 FUNCTION(_sk_multiply_hsw)
 _sk_multiply_hsw:
-  .byte  196,98,125,24,5,75,68,0,0           // vbroadcastss  0x444b(%rip),%ymm8        # 47b4 <_sk_callback_hsw+0x15c>
+  .byte  196,98,125,24,5,71,68,0,0           // vbroadcastss  0x4447(%rip),%ymm8        # 47b0 <_sk_callback_hsw+0x15c>
   .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
@@ -9004,7 +9000,7 @@ HIDDEN _sk_xor__hsw
 .globl _sk_xor__hsw
 FUNCTION(_sk_xor__hsw)
 _sk_xor__hsw:
-  .byte  196,98,125,24,5,198,67,0,0          // vbroadcastss  0x43c6(%rip),%ymm8        # 47b8 <_sk_callback_hsw+0x160>
+  .byte  196,98,125,24,5,194,67,0,0          // vbroadcastss  0x43c2(%rip),%ymm8        # 47b4 <_sk_callback_hsw+0x160>
   .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
@@ -9038,7 +9034,7 @@ _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  196,98,125,24,5,78,67,0,0           // vbroadcastss  0x434e(%rip),%ymm8        # 47bc <_sk_callback_hsw+0x164>
+  .byte  196,98,125,24,5,74,67,0,0           // vbroadcastss  0x434a(%rip),%ymm8        # 47b8 <_sk_callback_hsw+0x164>
   .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
@@ -9063,7 +9059,7 @@ _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  196,98,125,24,5,253,66,0,0          // vbroadcastss  0x42fd(%rip),%ymm8        # 47c0 <_sk_callback_hsw+0x168>
+  .byte  196,98,125,24,5,249,66,0,0          // vbroadcastss  0x42f9(%rip),%ymm8        # 47bc <_sk_callback_hsw+0x168>
   .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
@@ -9091,7 +9087,7 @@ _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  196,98,125,24,5,160,66,0,0          // vbroadcastss  0x42a0(%rip),%ymm8        # 47c4 <_sk_callback_hsw+0x16c>
+  .byte  196,98,125,24,5,156,66,0,0          // vbroadcastss  0x429c(%rip),%ymm8        # 47c0 <_sk_callback_hsw+0x16c>
   .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
@@ -9113,7 +9109,7 @@ _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  196,98,125,24,5,94,66,0,0           // vbroadcastss  0x425e(%rip),%ymm8        # 47c8 <_sk_callback_hsw+0x170>
+  .byte  196,98,125,24,5,90,66,0,0           // vbroadcastss  0x425a(%rip),%ymm8        # 47c4 <_sk_callback_hsw+0x170>
   .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
@@ -9123,7 +9119,7 @@ HIDDEN _sk_colorburn_hsw
 .globl _sk_colorburn_hsw
 FUNCTION(_sk_colorburn_hsw)
 _sk_colorburn_hsw:
-  .byte  196,98,125,24,5,76,66,0,0           // vbroadcastss  0x424c(%rip),%ymm8        # 47cc <_sk_callback_hsw+0x174>
+  .byte  196,98,125,24,5,72,66,0,0           // vbroadcastss  0x4248(%rip),%ymm8        # 47c8 <_sk_callback_hsw+0x174>
   .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
@@ -9181,7 +9177,7 @@ HIDDEN _sk_colordodge_hsw
 FUNCTION(_sk_colordodge_hsw)
 _sk_colordodge_hsw:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,13,87,65,0,0          // vbroadcastss  0x4157(%rip),%ymm9        # 47d0 <_sk_callback_hsw+0x178>
+  .byte  196,98,125,24,13,83,65,0,0          // vbroadcastss  0x4153(%rip),%ymm9        # 47cc <_sk_callback_hsw+0x178>
   .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
@@ -9234,7 +9230,7 @@ HIDDEN _sk_hardlight_hsw
 .globl _sk_hardlight_hsw
 FUNCTION(_sk_hardlight_hsw)
 _sk_hardlight_hsw:
-  .byte  196,98,125,24,5,120,64,0,0          // vbroadcastss  0x4078(%rip),%ymm8        # 47d4 <_sk_callback_hsw+0x17c>
+  .byte  196,98,125,24,5,116,64,0,0          // vbroadcastss  0x4074(%rip),%ymm8        # 47d0 <_sk_callback_hsw+0x17c>
   .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
@@ -9285,7 +9281,7 @@ HIDDEN _sk_overlay_hsw
 .globl _sk_overlay_hsw
 FUNCTION(_sk_overlay_hsw)
 _sk_overlay_hsw:
-  .byte  196,98,125,24,5,176,63,0,0          // vbroadcastss  0x3fb0(%rip),%ymm8        # 47d8 <_sk_callback_hsw+0x180>
+  .byte  196,98,125,24,5,172,63,0,0          // vbroadcastss  0x3fac(%rip),%ymm8        # 47d4 <_sk_callback_hsw+0x180>
   .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
@@ -9346,10 +9342,10 @@ _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  196,98,125,24,29,187,62,0,0         // vbroadcastss  0x3ebb(%rip),%ymm11        # 47e0 <_sk_callback_hsw+0x188>
+  .byte  196,98,125,24,29,183,62,0,0         // vbroadcastss  0x3eb7(%rip),%ymm11        # 47dc <_sk_callback_hsw+0x188>
   .byte  196,65,20,88,227                    // vaddps        %ymm11,%ymm13,%ymm12
   .byte  196,65,28,89,192                    // vmulps        %ymm8,%ymm12,%ymm8
-  .byte  196,98,125,24,37,172,62,0,0         // vbroadcastss  0x3eac(%rip),%ymm12        # 47e4 <_sk_callback_hsw+0x18c>
+  .byte  196,98,125,24,37,168,62,0,0         // vbroadcastss  0x3ea8(%rip),%ymm12        # 47e0 <_sk_callback_hsw+0x18c>
   .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
@@ -9359,7 +9355,7 @@ _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  196,98,125,24,5,111,62,0,0          // vbroadcastss  0x3e6f(%rip),%ymm8        # 47dc <_sk_callback_hsw+0x184>
+  .byte  196,98,125,24,5,107,62,0,0          // vbroadcastss  0x3e6b(%rip),%ymm8        # 47d8 <_sk_callback_hsw+0x184>
   .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
@@ -9472,11 +9468,11 @@ _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  196,98,125,24,53,115,60,0,0         // vbroadcastss  0x3c73(%rip),%ymm14        # 47e8 <_sk_callback_hsw+0x190>
-  .byte  196,98,125,24,61,110,60,0,0         // vbroadcastss  0x3c6e(%rip),%ymm15        # 47ec <_sk_callback_hsw+0x194>
+  .byte  196,98,125,24,53,111,60,0,0         // vbroadcastss  0x3c6f(%rip),%ymm14        # 47e4 <_sk_callback_hsw+0x190>
+  .byte  196,98,125,24,61,106,60,0,0         // vbroadcastss  0x3c6a(%rip),%ymm15        # 47e8 <_sk_callback_hsw+0x194>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,66,93,184,238                   // vfmadd231ps   %ymm14,%ymm4,%ymm13
-  .byte  196,226,125,24,5,95,60,0,0          // vbroadcastss  0x3c5f(%rip),%ymm0        # 47f0 <_sk_callback_hsw+0x198>
+  .byte  196,226,125,24,5,91,60,0,0          // vbroadcastss  0x3c5b(%rip),%ymm0        # 47ec <_sk_callback_hsw+0x198>
   .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
@@ -9531,7 +9527,7 @@ _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  196,226,125,24,13,76,59,0,0         // vbroadcastss  0x3b4c(%rip),%ymm1        # 47f4 <_sk_callback_hsw+0x19c>
+  .byte  196,226,125,24,13,72,59,0,0         // vbroadcastss  0x3b48(%rip),%ymm1        # 47f0 <_sk_callback_hsw+0x19c>
   .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
@@ -9585,11 +9581,11 @@ _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  196,98,125,24,53,99,58,0,0          // vbroadcastss  0x3a63(%rip),%ymm14        # 47f8 <_sk_callback_hsw+0x1a0>
-  .byte  196,98,125,24,61,94,58,0,0          // vbroadcastss  0x3a5e(%rip),%ymm15        # 47fc <_sk_callback_hsw+0x1a4>
+  .byte  196,98,125,24,53,95,58,0,0          // vbroadcastss  0x3a5f(%rip),%ymm14        # 47f4 <_sk_callback_hsw+0x1a0>
+  .byte  196,98,125,24,61,90,58,0,0          // vbroadcastss  0x3a5a(%rip),%ymm15        # 47f8 <_sk_callback_hsw+0x1a4>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,66,93,184,238                   // vfmadd231ps   %ymm14,%ymm4,%ymm13
-  .byte  196,226,125,24,5,79,58,0,0          // vbroadcastss  0x3a4f(%rip),%ymm0        # 4800 <_sk_callback_hsw+0x1a8>
+  .byte  196,226,125,24,5,75,58,0,0          // vbroadcastss  0x3a4b(%rip),%ymm0        # 47fc <_sk_callback_hsw+0x1a8>
   .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
@@ -9644,7 +9640,7 @@ _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  196,226,125,24,13,60,57,0,0         // vbroadcastss  0x393c(%rip),%ymm1        # 4804 <_sk_callback_hsw+0x1ac>
+  .byte  196,226,125,24,13,56,57,0,0         // vbroadcastss  0x3938(%rip),%ymm1        # 4800 <_sk_callback_hsw+0x1ac>
   .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
@@ -9672,11 +9668,11 @@ _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  196,98,125,24,45,213,56,0,0         // vbroadcastss  0x38d5(%rip),%ymm13        # 4808 <_sk_callback_hsw+0x1b0>
-  .byte  196,98,125,24,53,208,56,0,0         // vbroadcastss  0x38d0(%rip),%ymm14        # 480c <_sk_callback_hsw+0x1b4>
+  .byte  196,98,125,24,45,209,56,0,0         // vbroadcastss  0x38d1(%rip),%ymm13        # 4804 <_sk_callback_hsw+0x1b0>
+  .byte  196,98,125,24,53,204,56,0,0         // vbroadcastss  0x38cc(%rip),%ymm14        # 4808 <_sk_callback_hsw+0x1b4>
   .byte  196,65,84,89,230                    // vmulps        %ymm14,%ymm5,%ymm12
   .byte  196,66,93,184,229                   // vfmadd231ps   %ymm13,%ymm4,%ymm12
-  .byte  196,98,125,24,61,193,56,0,0         // vbroadcastss  0x38c1(%rip),%ymm15        # 4810 <_sk_callback_hsw+0x1b8>
+  .byte  196,98,125,24,61,189,56,0,0         // vbroadcastss  0x38bd(%rip),%ymm15        # 480c <_sk_callback_hsw+0x1b8>
   .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
@@ -9732,7 +9728,7 @@ _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  196,226,125,24,5,163,55,0,0         // vbroadcastss  0x37a3(%rip),%ymm0        # 4814 <_sk_callback_hsw+0x1bc>
+  .byte  196,226,125,24,5,159,55,0,0         // vbroadcastss  0x379f(%rip),%ymm0        # 4810 <_sk_callback_hsw+0x1bc>
   .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
@@ -9760,11 +9756,11 @@ _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  196,98,125,24,45,60,55,0,0          // vbroadcastss  0x373c(%rip),%ymm13        # 4818 <_sk_callback_hsw+0x1c0>
-  .byte  196,98,125,24,53,55,55,0,0          // vbroadcastss  0x3737(%rip),%ymm14        # 481c <_sk_callback_hsw+0x1c4>
+  .byte  196,98,125,24,45,56,55,0,0          // vbroadcastss  0x3738(%rip),%ymm13        # 4814 <_sk_callback_hsw+0x1c0>
+  .byte  196,98,125,24,53,51,55,0,0          // vbroadcastss  0x3733(%rip),%ymm14        # 4818 <_sk_callback_hsw+0x1c4>
   .byte  196,65,116,89,230                   // vmulps        %ymm14,%ymm1,%ymm12
   .byte  196,66,109,184,229                  // vfmadd231ps   %ymm13,%ymm2,%ymm12
-  .byte  196,98,125,24,61,40,55,0,0          // vbroadcastss  0x3728(%rip),%ymm15        # 4820 <_sk_callback_hsw+0x1c8>
+  .byte  196,98,125,24,61,36,55,0,0          // vbroadcastss  0x3724(%rip),%ymm15        # 481c <_sk_callback_hsw+0x1c8>
   .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
@@ -9820,7 +9816,7 @@ _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  196,226,125,24,5,10,54,0,0          // vbroadcastss  0x360a(%rip),%ymm0        # 4824 <_sk_callback_hsw+0x1cc>
+  .byte  196,226,125,24,5,6,54,0,0           // vbroadcastss  0x3606(%rip),%ymm0        # 4820 <_sk_callback_hsw+0x1cc>
   .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
@@ -9853,7 +9849,7 @@ HIDDEN _sk_clamp_1_hsw
 .globl _sk_clamp_1_hsw
 FUNCTION(_sk_clamp_1_hsw)
 _sk_clamp_1_hsw:
-  .byte  196,98,125,24,5,166,53,0,0          // vbroadcastss  0x35a6(%rip),%ymm8        # 4828 <_sk_callback_hsw+0x1d0>
+  .byte  196,98,125,24,5,162,53,0,0          // vbroadcastss  0x35a2(%rip),%ymm8        # 4824 <_sk_callback_hsw+0x1d0>
   .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
@@ -9865,7 +9861,7 @@ HIDDEN _sk_clamp_a_hsw
 .globl _sk_clamp_a_hsw
 FUNCTION(_sk_clamp_a_hsw)
 _sk_clamp_a_hsw:
-  .byte  196,98,125,24,5,137,53,0,0          // vbroadcastss  0x3589(%rip),%ymm8        # 482c <_sk_callback_hsw+0x1d4>
+  .byte  196,98,125,24,5,133,53,0,0          // vbroadcastss  0x3585(%rip),%ymm8        # 4828 <_sk_callback_hsw+0x1d4>
   .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
@@ -9951,7 +9947,7 @@ 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  196,98,125,24,21,209,52,0,0         // vbroadcastss  0x34d1(%rip),%ymm10        # 4830 <_sk_callback_hsw+0x1d8>
+  .byte  196,98,125,24,21,205,52,0,0         // vbroadcastss  0x34cd(%rip),%ymm10        # 482c <_sk_callback_hsw+0x1d8>
   .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
@@ -9964,16 +9960,16 @@ HIDDEN _sk_from_srgb_hsw
 .globl _sk_from_srgb_hsw
 FUNCTION(_sk_from_srgb_hsw)
 _sk_from_srgb_hsw:
-  .byte  196,98,125,24,5,178,52,0,0          // vbroadcastss  0x34b2(%rip),%ymm8        # 4834 <_sk_callback_hsw+0x1dc>
+  .byte  196,98,125,24,5,174,52,0,0          // vbroadcastss  0x34ae(%rip),%ymm8        # 4830 <_sk_callback_hsw+0x1dc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,164,52,0,0         // vbroadcastss  0x34a4(%rip),%ymm11        # 4838 <_sk_callback_hsw+0x1e0>
-  .byte  196,98,125,24,37,159,52,0,0         // vbroadcastss  0x349f(%rip),%ymm12        # 483c <_sk_callback_hsw+0x1e4>
+  .byte  196,98,125,24,29,160,52,0,0         // vbroadcastss  0x34a0(%rip),%ymm11        # 4834 <_sk_callback_hsw+0x1e0>
+  .byte  196,98,125,24,37,155,52,0,0         // vbroadcastss  0x349b(%rip),%ymm12        # 4838 <_sk_callback_hsw+0x1e4>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,125,168,235                  // vfmadd213ps   %ymm11,%ymm0,%ymm13
-  .byte  196,98,125,24,53,144,52,0,0         // vbroadcastss  0x3490(%rip),%ymm14        # 4840 <_sk_callback_hsw+0x1e8>
+  .byte  196,98,125,24,53,140,52,0,0         // vbroadcastss  0x348c(%rip),%ymm14        # 483c <_sk_callback_hsw+0x1e8>
   .byte  196,66,45,168,238                   // vfmadd213ps   %ymm14,%ymm10,%ymm13
-  .byte  196,98,125,24,21,134,52,0,0         // vbroadcastss  0x3486(%rip),%ymm10        # 4844 <_sk_callback_hsw+0x1ec>
+  .byte  196,98,125,24,21,130,52,0,0         // vbroadcastss  0x3482(%rip),%ymm10        # 4840 <_sk_callback_hsw+0x1ec>
   .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
@@ -9999,16 +9995,16 @@ _sk_to_srgb_hsw:
   .byte  197,124,82,192                      // vrsqrtps      %ymm0,%ymm8
   .byte  196,65,124,83,200                   // vrcpps        %ymm8,%ymm9
   .byte  196,65,124,82,208                   // vrsqrtps      %ymm8,%ymm10
-  .byte  196,98,125,24,5,32,52,0,0           // vbroadcastss  0x3420(%rip),%ymm8        # 4848 <_sk_callback_hsw+0x1f0>
+  .byte  196,98,125,24,5,28,52,0,0           // vbroadcastss  0x341c(%rip),%ymm8        # 4844 <_sk_callback_hsw+0x1f0>
   .byte  196,65,124,89,216                   // vmulps        %ymm8,%ymm0,%ymm11
-  .byte  196,98,125,24,37,22,52,0,0          // vbroadcastss  0x3416(%rip),%ymm12        # 484c <_sk_callback_hsw+0x1f4>
-  .byte  196,98,125,24,45,17,52,0,0          // vbroadcastss  0x3411(%rip),%ymm13        # 4850 <_sk_callback_hsw+0x1f8>
+  .byte  196,98,125,24,37,18,52,0,0          // vbroadcastss  0x3412(%rip),%ymm12        # 4848 <_sk_callback_hsw+0x1f4>
+  .byte  196,98,125,24,45,13,52,0,0          // vbroadcastss  0x340d(%rip),%ymm13        # 484c <_sk_callback_hsw+0x1f8>
   .byte  196,66,21,168,204                   // vfmadd213ps   %ymm12,%ymm13,%ymm9
-  .byte  196,98,125,24,53,7,52,0,0           // vbroadcastss  0x3407(%rip),%ymm14        # 4854 <_sk_callback_hsw+0x1fc>
+  .byte  196,98,125,24,53,3,52,0,0           // vbroadcastss  0x3403(%rip),%ymm14        # 4850 <_sk_callback_hsw+0x1fc>
   .byte  196,66,13,184,202                   // vfmadd231ps   %ymm10,%ymm14,%ymm9
-  .byte  196,98,125,24,21,253,51,0,0         // vbroadcastss  0x33fd(%rip),%ymm10        # 4858 <_sk_callback_hsw+0x200>
+  .byte  196,98,125,24,21,249,51,0,0         // vbroadcastss  0x33f9(%rip),%ymm10        # 4854 <_sk_callback_hsw+0x200>
   .byte  196,65,44,93,201                    // vminps        %ymm9,%ymm10,%ymm9
-  .byte  196,98,125,24,61,243,51,0,0         // vbroadcastss  0x33f3(%rip),%ymm15        # 485c <_sk_callback_hsw+0x204>
+  .byte  196,98,125,24,61,239,51,0,0         // vbroadcastss  0x33ef(%rip),%ymm15        # 4858 <_sk_callback_hsw+0x204>
   .byte  196,193,124,194,199,1               // vcmpltps      %ymm15,%ymm0,%ymm0
   .byte  196,195,53,74,195,0                 // vblendvps     %ymm0,%ymm11,%ymm9,%ymm0
   .byte  197,124,82,201                      // vrsqrtps      %ymm1,%ymm9
@@ -10041,26 +10037,26 @@ _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  196,98,125,24,29,109,51,0,0         // vbroadcastss  0x336d(%rip),%ymm11        # 4860 <_sk_callback_hsw+0x208>
+  .byte  196,98,125,24,29,105,51,0,0         // vbroadcastss  0x3369(%rip),%ymm11        # 485c <_sk_callback_hsw+0x208>
   .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  196,98,125,24,53,90,51,0,0          // vbroadcastss  0x335a(%rip),%ymm14        # 4864 <_sk_callback_hsw+0x20c>
+  .byte  196,98,125,24,53,86,51,0,0          // vbroadcastss  0x3356(%rip),%ymm14        # 4860 <_sk_callback_hsw+0x20c>
   .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  196,98,125,24,53,65,51,0,0          // vbroadcastss  0x3341(%rip),%ymm14        # 486c <_sk_callback_hsw+0x214>
+  .byte  196,98,125,24,53,61,51,0,0          // vbroadcastss  0x333d(%rip),%ymm14        # 4868 <_sk_callback_hsw+0x214>
   .byte  196,66,37,168,238                   // vfmadd213ps   %ymm14,%ymm11,%ymm13
-  .byte  196,98,125,24,53,47,51,0,0          // vbroadcastss  0x332f(%rip),%ymm14        # 4868 <_sk_callback_hsw+0x210>
+  .byte  196,98,125,24,53,43,51,0,0          // vbroadcastss  0x332b(%rip),%ymm14        # 4864 <_sk_callback_hsw+0x210>
   .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  196,98,125,24,29,18,51,0,0          // vbroadcastss  0x3312(%rip),%ymm11        # 4874 <_sk_callback_hsw+0x21c>
+  .byte  196,98,125,24,29,14,51,0,0          // vbroadcastss  0x330e(%rip),%ymm11        # 4870 <_sk_callback_hsw+0x21c>
   .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
@@ -10070,7 +10066,7 @@ _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  196,98,125,24,5,213,50,0,0          // vbroadcastss  0x32d5(%rip),%ymm8        # 4870 <_sk_callback_hsw+0x218>
+  .byte  196,98,125,24,5,209,50,0,0          // vbroadcastss  0x32d1(%rip),%ymm8        # 486c <_sk_callback_hsw+0x218>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -10087,30 +10083,30 @@ _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  196,98,125,24,5,162,50,0,0          // vbroadcastss  0x32a2(%rip),%ymm8        # 4878 <_sk_callback_hsw+0x220>
+  .byte  196,98,125,24,5,158,50,0,0          // vbroadcastss  0x329e(%rip),%ymm8        # 4874 <_sk_callback_hsw+0x220>
   .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  196,98,125,24,13,133,50,0,0         // vbroadcastss  0x3285(%rip),%ymm9        # 487c <_sk_callback_hsw+0x224>
+  .byte  196,98,125,24,13,129,50,0,0         // vbroadcastss  0x3281(%rip),%ymm9        # 4878 <_sk_callback_hsw+0x224>
   .byte  196,66,109,170,202                  // vfmsub213ps   %ymm10,%ymm2,%ymm9
-  .byte  196,98,125,24,29,123,50,0,0         // vbroadcastss  0x327b(%rip),%ymm11        # 4880 <_sk_callback_hsw+0x228>
+  .byte  196,98,125,24,29,119,50,0,0         // vbroadcastss  0x3277(%rip),%ymm11        # 487c <_sk_callback_hsw+0x228>
   .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  196,98,125,24,45,101,50,0,0         // vbroadcastss  0x3265(%rip),%ymm13        # 4888 <_sk_callback_hsw+0x230>
+  .byte  196,98,125,24,45,97,50,0,0          // vbroadcastss  0x3261(%rip),%ymm13        # 4884 <_sk_callback_hsw+0x230>
   .byte  196,193,4,89,197                    // vmulps        %ymm13,%ymm15,%ymm0
-  .byte  196,98,125,24,53,91,50,0,0          // vbroadcastss  0x325b(%rip),%ymm14        # 488c <_sk_callback_hsw+0x234>
+  .byte  196,98,125,24,53,87,50,0,0          // vbroadcastss  0x3257(%rip),%ymm14        # 4888 <_sk_callback_hsw+0x234>
   .byte  197,12,92,224                       // vsubps        %ymm0,%ymm14,%ymm12
   .byte  196,66,37,168,225                   // vfmadd213ps   %ymm9,%ymm11,%ymm12
-  .byte  196,226,125,24,29,65,50,0,0         // vbroadcastss  0x3241(%rip),%ymm3        # 4884 <_sk_callback_hsw+0x22c>
+  .byte  196,226,125,24,29,61,50,0,0         // vbroadcastss  0x323d(%rip),%ymm3        # 4880 <_sk_callback_hsw+0x22c>
   .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  196,98,125,24,37,44,50,0,0          // vbroadcastss  0x322c(%rip),%ymm12        # 4890 <_sk_callback_hsw+0x238>
+  .byte  196,98,125,24,37,40,50,0,0          // vbroadcastss  0x3228(%rip),%ymm12        # 488c <_sk_callback_hsw+0x238>
   .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
@@ -10126,7 +10122,7 @@ _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  196,226,125,24,5,216,49,0,0         // vbroadcastss  0x31d8(%rip),%ymm0        # 4894 <_sk_callback_hsw+0x23c>
+  .byte  196,226,125,24,5,212,49,0,0         // vbroadcastss  0x31d4(%rip),%ymm0        # 4890 <_sk_callback_hsw+0x23c>
   .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
@@ -10180,7 +10176,7 @@ _sk_scale_u8_hsw:
   .byte  197,122,126,0                       // vmovq         (%rax),%xmm8
   .byte  196,66,125,49,192                   // vpmovzxbd     %xmm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,24,49,0,0          // vbroadcastss  0x3118(%rip),%ymm9        # 4898 <_sk_callback_hsw+0x240>
+  .byte  196,98,125,24,13,20,49,0,0          // vbroadcastss  0x3114(%rip),%ymm9        # 4894 <_sk_callback_hsw+0x240>
   .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
@@ -10232,7 +10228,7 @@ _sk_lerp_u8_hsw:
   .byte  197,122,126,0                       // vmovq         (%rax),%xmm8
   .byte  196,66,125,49,192                   // vpmovzxbd     %xmm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,133,48,0,0         // vbroadcastss  0x3085(%rip),%ymm9        # 489c <_sk_callback_hsw+0x244>
+  .byte  196,98,125,24,13,129,48,0,0         // vbroadcastss  0x3081(%rip),%ymm9        # 4898 <_sk_callback_hsw+0x244>
   .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
@@ -10268,20 +10264,20 @@ _sk_lerp_565_hsw:
   .byte  15,133,169,0,0,0                    // jne           1923 <_sk_lerp_565_hsw+0xb7>
   .byte  196,65,122,111,4,122                // vmovdqu       (%r10,%rdi,2),%xmm8
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
-  .byte  196,98,125,88,13,18,48,0,0          // vpbroadcastd  0x3012(%rip),%ymm9        # 48a0 <_sk_callback_hsw+0x248>
+  .byte  196,98,125,88,13,14,48,0,0          // vpbroadcastd  0x300e(%rip),%ymm9        # 489c <_sk_callback_hsw+0x248>
   .byte  196,65,61,219,201                   // vpand         %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,3,48,0,0           // vbroadcastss  0x3003(%rip),%ymm10        # 48a4 <_sk_callback_hsw+0x24c>
+  .byte  196,98,125,24,21,255,47,0,0         // vbroadcastss  0x2fff(%rip),%ymm10        # 48a0 <_sk_callback_hsw+0x24c>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,88,21,249,47,0,0         // vpbroadcastd  0x2ff9(%rip),%ymm10        # 48a8 <_sk_callback_hsw+0x250>
+  .byte  196,98,125,88,21,245,47,0,0         // vpbroadcastd  0x2ff5(%rip),%ymm10        # 48a4 <_sk_callback_hsw+0x250>
   .byte  196,65,61,219,210                   // vpand         %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,234,47,0,0         // vbroadcastss  0x2fea(%rip),%ymm11        # 48ac <_sk_callback_hsw+0x254>
+  .byte  196,98,125,24,29,230,47,0,0         // vbroadcastss  0x2fe6(%rip),%ymm11        # 48a8 <_sk_callback_hsw+0x254>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,88,29,224,47,0,0         // vpbroadcastd  0x2fe0(%rip),%ymm11        # 48b0 <_sk_callback_hsw+0x258>
+  .byte  196,98,125,88,29,220,47,0,0         // vpbroadcastd  0x2fdc(%rip),%ymm11        # 48ac <_sk_callback_hsw+0x258>
   .byte  196,65,61,219,195                   // vpand         %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,209,47,0,0         // vbroadcastss  0x2fd1(%rip),%ymm11        # 48b4 <_sk_callback_hsw+0x25c>
+  .byte  196,98,125,24,29,205,47,0,0         // vbroadcastss  0x2fcd(%rip),%ymm11        # 48b0 <_sk_callback_hsw+0x25c>
   .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
@@ -10322,7 +10318,7 @@ _sk_lerp_565_hsw:
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,225                 // jmpq          ffffffffe200199c <_sk_callback_hsw+0xffffffffe1ffd344>
+  .byte  233,255,255,255,225                 // jmpq          ffffffffe200199c <_sk_callback_hsw+0xffffffffe1ffd348>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
@@ -10351,21 +10347,21 @@ _sk_load_tables_hsw:
   .byte  77,133,192                          // test          %r8,%r8
   .byte  117,105                             // jne           1a2e <_sk_load_tables_hsw+0x7e>
   .byte  196,193,126,111,25                  // vmovdqu       (%r9),%ymm3
-  .byte  197,229,219,13,142,49,0,0           // vpand         0x318e(%rip),%ymm3,%ymm1        # 4b60 <_sk_callback_hsw+0x508>
+  .byte  197,229,219,13,142,49,0,0           // vpand         0x318e(%rip),%ymm3,%ymm1        # 4b60 <_sk_callback_hsw+0x50c>
   .byte  196,65,61,118,192                   // vpcmpeqd      %ymm8,%ymm8,%ymm8
   .byte  72,139,72,8                         // mov           0x8(%rax),%rcx
   .byte  76,139,72,16                        // mov           0x10(%rax),%r9
   .byte  197,237,118,210                     // vpcmpeqd      %ymm2,%ymm2,%ymm2
   .byte  196,226,109,146,4,137               // vgatherdps    %ymm2,(%rcx,%ymm1,4),%ymm0
-  .byte  196,226,101,0,21,142,49,0,0         // vpshufb       0x318e(%rip),%ymm3,%ymm2        # 4b80 <_sk_callback_hsw+0x528>
+  .byte  196,226,101,0,21,142,49,0,0         // vpshufb       0x318e(%rip),%ymm3,%ymm2        # 4b80 <_sk_callback_hsw+0x52c>
   .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
   .byte  196,194,53,146,12,145               // vgatherdps    %ymm9,(%r9,%ymm2,4),%ymm1
   .byte  72,139,64,24                        // mov           0x18(%rax),%rax
-  .byte  196,98,101,0,13,150,49,0,0          // vpshufb       0x3196(%rip),%ymm3,%ymm9        # 4ba0 <_sk_callback_hsw+0x548>
+  .byte  196,98,101,0,13,150,49,0,0          // vpshufb       0x3196(%rip),%ymm3,%ymm9        # 4ba0 <_sk_callback_hsw+0x54c>
   .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  196,98,125,24,5,150,46,0,0          // vbroadcastss  0x2e96(%rip),%ymm8        # 48b8 <_sk_callback_hsw+0x260>
+  .byte  196,98,125,24,5,146,46,0,0          // vbroadcastss  0x2e92(%rip),%ymm8        # 48b4 <_sk_callback_hsw+0x260>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,137,193                          // mov           %r8,%rcx
@@ -10404,7 +10400,7 @@ _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  197,121,111,21,34,50,0,0            // vmovdqa       0x3222(%rip),%xmm10        # 4ce0 <_sk_callback_hsw+0x688>
+  .byte  197,121,111,21,34,50,0,0            // vmovdqa       0x3222(%rip),%xmm10        # 4ce0 <_sk_callback_hsw+0x68c>
   .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
@@ -10426,7 +10422,7 @@ _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  196,98,125,24,5,143,45,0,0          // vbroadcastss  0x2d8f(%rip),%ymm8        # 48bc <_sk_callback_hsw+0x264>
+  .byte  196,98,125,24,5,139,45,0,0          // vbroadcastss  0x2d8b(%rip),%ymm8        # 48b8 <_sk_callback_hsw+0x264>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -10486,7 +10482,7 @@ _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  197,121,111,13,194,48,0,0           // vmovdqa       0x30c2(%rip),%xmm9        # 4cf0 <_sk_callback_hsw+0x698>
+  .byte  197,121,111,13,194,48,0,0           // vmovdqa       0x30c2(%rip),%xmm9        # 4cf0 <_sk_callback_hsw+0x69c>
   .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
@@ -10503,7 +10499,7 @@ _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  196,226,125,24,29,61,44,0,0         // vbroadcastss  0x2c3d(%rip),%ymm3        # 48c0 <_sk_callback_hsw+0x268>
+  .byte  196,226,125,24,29,57,44,0,0         // vbroadcastss  0x2c39(%rip),%ymm3        # 48bc <_sk_callback_hsw+0x268>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,129,121,110,4,72                // vmovd         (%r8,%r9,2),%xmm0
   .byte  196,129,121,196,68,72,4,2           // vpinsrw       $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
@@ -10550,7 +10546,7 @@ _sk_byte_tables_hsw:
   .byte  65,84                               // push          %r12
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,123,43,0,0          // vbroadcastss  0x2b7b(%rip),%ymm8        # 48c4 <_sk_callback_hsw+0x26c>
+  .byte  196,98,125,24,5,119,43,0,0          // vbroadcastss  0x2b77(%rip),%ymm8        # 48c0 <_sk_callback_hsw+0x26c>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,253,91,192                      // vcvtps2dq     %ymm0,%ymm0
   .byte  196,195,249,22,192,1                // vpextrq       $0x1,%xmm0,%r8
@@ -10587,7 +10583,7 @@ _sk_byte_tables_hsw:
   .byte  196,227,121,32,197,7                // vpinsrb       $0x7,%ebp,%xmm0,%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,204,42,0,0         // vbroadcastss  0x2acc(%rip),%ymm9        # 48c8 <_sk_callback_hsw+0x270>
+  .byte  196,98,125,24,13,200,42,0,0         // vbroadcastss  0x2ac8(%rip),%ymm9        # 48c4 <_sk_callback_hsw+0x270>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
@@ -10748,7 +10744,7 @@ _sk_byte_tables_rgb_hsw:
   .byte  196,227,121,32,197,7                // vpinsrb       $0x7,%ebp,%xmm0,%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,5,40,0,0           // vbroadcastss  0x2805(%rip),%ymm9        # 48cc <_sk_callback_hsw+0x274>
+  .byte  196,98,125,24,13,1,40,0,0           // vbroadcastss  0x2801(%rip),%ymm9        # 48c8 <_sk_callback_hsw+0x274>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
@@ -10911,33 +10907,33 @@ _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  196,98,125,24,37,184,37,0,0         // vbroadcastss  0x25b8(%rip),%ymm12        # 48d0 <_sk_callback_hsw+0x278>
-  .byte  196,98,125,24,45,179,37,0,0         // vbroadcastss  0x25b3(%rip),%ymm13        # 48d4 <_sk_callback_hsw+0x27c>
+  .byte  196,98,125,24,37,180,37,0,0         // vbroadcastss  0x25b4(%rip),%ymm12        # 48cc <_sk_callback_hsw+0x278>
+  .byte  196,98,125,24,45,175,37,0,0         // vbroadcastss  0x25af(%rip),%ymm13        # 48d0 <_sk_callback_hsw+0x27c>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,169,37,0,0         // vbroadcastss  0x25a9(%rip),%ymm13        # 48d8 <_sk_callback_hsw+0x280>
+  .byte  196,98,125,24,45,165,37,0,0         // vbroadcastss  0x25a5(%rip),%ymm13        # 48d4 <_sk_callback_hsw+0x280>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,159,37,0,0         // vbroadcastss  0x259f(%rip),%ymm13        # 48dc <_sk_callback_hsw+0x284>
+  .byte  196,98,125,24,45,155,37,0,0         // vbroadcastss  0x259b(%rip),%ymm13        # 48d8 <_sk_callback_hsw+0x284>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,149,37,0,0         // vbroadcastss  0x2595(%rip),%ymm11        # 48e0 <_sk_callback_hsw+0x288>
+  .byte  196,98,125,24,29,145,37,0,0         // vbroadcastss  0x2591(%rip),%ymm11        # 48dc <_sk_callback_hsw+0x288>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,139,37,0,0         // vbroadcastss  0x258b(%rip),%ymm12        # 48e4 <_sk_callback_hsw+0x28c>
+  .byte  196,98,125,24,37,135,37,0,0         // vbroadcastss  0x2587(%rip),%ymm12        # 48e0 <_sk_callback_hsw+0x28c>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,129,37,0,0         // vbroadcastss  0x2581(%rip),%ymm12        # 48e8 <_sk_callback_hsw+0x290>
+  .byte  196,98,125,24,37,125,37,0,0         // vbroadcastss  0x257d(%rip),%ymm12        # 48e4 <_sk_callback_hsw+0x290>
   .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  196,98,125,24,29,98,37,0,0          // vbroadcastss  0x2562(%rip),%ymm11        # 48ec <_sk_callback_hsw+0x294>
+  .byte  196,98,125,24,29,94,37,0,0          // vbroadcastss  0x255e(%rip),%ymm11        # 48e8 <_sk_callback_hsw+0x294>
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,88,37,0,0          // vbroadcastss  0x2558(%rip),%ymm11        # 48f0 <_sk_callback_hsw+0x298>
+  .byte  196,98,125,24,29,84,37,0,0          // vbroadcastss  0x2554(%rip),%ymm11        # 48ec <_sk_callback_hsw+0x298>
   .byte  196,98,45,172,216                   // vfnmadd213ps  %ymm0,%ymm10,%ymm11
-  .byte  196,226,125,24,5,78,37,0,0          // vbroadcastss  0x254e(%rip),%ymm0        # 48f4 <_sk_callback_hsw+0x29c>
+  .byte  196,226,125,24,5,74,37,0,0          // vbroadcastss  0x254a(%rip),%ymm0        # 48f0 <_sk_callback_hsw+0x29c>
   .byte  196,193,124,92,194                  // vsubps        %ymm10,%ymm0,%ymm0
-  .byte  196,98,125,24,21,68,37,0,0          // vbroadcastss  0x2544(%rip),%ymm10        # 48f8 <_sk_callback_hsw+0x2a0>
+  .byte  196,98,125,24,21,64,37,0,0          // vbroadcastss  0x2540(%rip),%ymm10        # 48f4 <_sk_callback_hsw+0x2a0>
   .byte  197,172,94,192                      // vdivps        %ymm0,%ymm10,%ymm0
   .byte  197,164,88,192                      // vaddps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,21,55,37,0,0          // vbroadcastss  0x2537(%rip),%ymm10        # 48fc <_sk_callback_hsw+0x2a4>
+  .byte  196,98,125,24,21,51,37,0,0          // vbroadcastss  0x2533(%rip),%ymm10        # 48f8 <_sk_callback_hsw+0x2a4>
   .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
@@ -10945,7 +10941,7 @@ _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  196,98,125,24,5,14,37,0,0           // vbroadcastss  0x250e(%rip),%ymm8        # 4900 <_sk_callback_hsw+0x2a8>
+  .byte  196,98,125,24,5,10,37,0,0           // vbroadcastss  0x250a(%rip),%ymm8        # 48fc <_sk_callback_hsw+0x2a8>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -10965,33 +10961,33 @@ _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  196,98,125,24,37,198,36,0,0         // vbroadcastss  0x24c6(%rip),%ymm12        # 4904 <_sk_callback_hsw+0x2ac>
-  .byte  196,98,125,24,45,193,36,0,0         // vbroadcastss  0x24c1(%rip),%ymm13        # 4908 <_sk_callback_hsw+0x2b0>
+  .byte  196,98,125,24,37,194,36,0,0         // vbroadcastss  0x24c2(%rip),%ymm12        # 4900 <_sk_callback_hsw+0x2ac>
+  .byte  196,98,125,24,45,189,36,0,0         // vbroadcastss  0x24bd(%rip),%ymm13        # 4904 <_sk_callback_hsw+0x2b0>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,183,36,0,0         // vbroadcastss  0x24b7(%rip),%ymm13        # 490c <_sk_callback_hsw+0x2b4>
+  .byte  196,98,125,24,45,179,36,0,0         // vbroadcastss  0x24b3(%rip),%ymm13        # 4908 <_sk_callback_hsw+0x2b4>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,173,36,0,0         // vbroadcastss  0x24ad(%rip),%ymm13        # 4910 <_sk_callback_hsw+0x2b8>
+  .byte  196,98,125,24,45,169,36,0,0         // vbroadcastss  0x24a9(%rip),%ymm13        # 490c <_sk_callback_hsw+0x2b8>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,163,36,0,0         // vbroadcastss  0x24a3(%rip),%ymm11        # 4914 <_sk_callback_hsw+0x2bc>
+  .byte  196,98,125,24,29,159,36,0,0         // vbroadcastss  0x249f(%rip),%ymm11        # 4910 <_sk_callback_hsw+0x2bc>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,153,36,0,0         // vbroadcastss  0x2499(%rip),%ymm12        # 4918 <_sk_callback_hsw+0x2c0>
+  .byte  196,98,125,24,37,149,36,0,0         // vbroadcastss  0x2495(%rip),%ymm12        # 4914 <_sk_callback_hsw+0x2c0>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,143,36,0,0         // vbroadcastss  0x248f(%rip),%ymm12        # 491c <_sk_callback_hsw+0x2c4>
+  .byte  196,98,125,24,37,139,36,0,0         // vbroadcastss  0x248b(%rip),%ymm12        # 4918 <_sk_callback_hsw+0x2c4>
   .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  196,98,125,24,29,112,36,0,0         // vbroadcastss  0x2470(%rip),%ymm11        # 4920 <_sk_callback_hsw+0x2c8>
+  .byte  196,98,125,24,29,108,36,0,0         // vbroadcastss  0x246c(%rip),%ymm11        # 491c <_sk_callback_hsw+0x2c8>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,102,36,0,0         // vbroadcastss  0x2466(%rip),%ymm11        # 4924 <_sk_callback_hsw+0x2cc>
+  .byte  196,98,125,24,29,98,36,0,0          // vbroadcastss  0x2462(%rip),%ymm11        # 4920 <_sk_callback_hsw+0x2cc>
   .byte  196,98,45,172,217                   // vfnmadd213ps  %ymm1,%ymm10,%ymm11
-  .byte  196,226,125,24,13,92,36,0,0         // vbroadcastss  0x245c(%rip),%ymm1        # 4928 <_sk_callback_hsw+0x2d0>
+  .byte  196,226,125,24,13,88,36,0,0         // vbroadcastss  0x2458(%rip),%ymm1        # 4924 <_sk_callback_hsw+0x2d0>
   .byte  196,193,116,92,202                  // vsubps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,21,82,36,0,0          // vbroadcastss  0x2452(%rip),%ymm10        # 492c <_sk_callback_hsw+0x2d4>
+  .byte  196,98,125,24,21,78,36,0,0          // vbroadcastss  0x244e(%rip),%ymm10        # 4928 <_sk_callback_hsw+0x2d4>
   .byte  197,172,94,201                      // vdivps        %ymm1,%ymm10,%ymm1
   .byte  197,164,88,201                      // vaddps        %ymm1,%ymm11,%ymm1
-  .byte  196,98,125,24,21,69,36,0,0          // vbroadcastss  0x2445(%rip),%ymm10        # 4930 <_sk_callback_hsw+0x2d8>
+  .byte  196,98,125,24,21,65,36,0,0          // vbroadcastss  0x2441(%rip),%ymm10        # 492c <_sk_callback_hsw+0x2d8>
   .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
@@ -10999,7 +10995,7 @@ _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  196,98,125,24,5,28,36,0,0           // vbroadcastss  0x241c(%rip),%ymm8        # 4934 <_sk_callback_hsw+0x2dc>
+  .byte  196,98,125,24,5,24,36,0,0           // vbroadcastss  0x2418(%rip),%ymm8        # 4930 <_sk_callback_hsw+0x2dc>
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11019,33 +11015,33 @@ _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  196,98,125,24,37,212,35,0,0         // vbroadcastss  0x23d4(%rip),%ymm12        # 4938 <_sk_callback_hsw+0x2e0>
-  .byte  196,98,125,24,45,207,35,0,0         // vbroadcastss  0x23cf(%rip),%ymm13        # 493c <_sk_callback_hsw+0x2e4>
+  .byte  196,98,125,24,37,208,35,0,0         // vbroadcastss  0x23d0(%rip),%ymm12        # 4934 <_sk_callback_hsw+0x2e0>
+  .byte  196,98,125,24,45,203,35,0,0         // vbroadcastss  0x23cb(%rip),%ymm13        # 4938 <_sk_callback_hsw+0x2e4>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,197,35,0,0         // vbroadcastss  0x23c5(%rip),%ymm13        # 4940 <_sk_callback_hsw+0x2e8>
+  .byte  196,98,125,24,45,193,35,0,0         // vbroadcastss  0x23c1(%rip),%ymm13        # 493c <_sk_callback_hsw+0x2e8>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,187,35,0,0         // vbroadcastss  0x23bb(%rip),%ymm13        # 4944 <_sk_callback_hsw+0x2ec>
+  .byte  196,98,125,24,45,183,35,0,0         // vbroadcastss  0x23b7(%rip),%ymm13        # 4940 <_sk_callback_hsw+0x2ec>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,177,35,0,0         // vbroadcastss  0x23b1(%rip),%ymm11        # 4948 <_sk_callback_hsw+0x2f0>
+  .byte  196,98,125,24,29,173,35,0,0         // vbroadcastss  0x23ad(%rip),%ymm11        # 4944 <_sk_callback_hsw+0x2f0>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,167,35,0,0         // vbroadcastss  0x23a7(%rip),%ymm12        # 494c <_sk_callback_hsw+0x2f4>
+  .byte  196,98,125,24,37,163,35,0,0         // vbroadcastss  0x23a3(%rip),%ymm12        # 4948 <_sk_callback_hsw+0x2f4>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,157,35,0,0         // vbroadcastss  0x239d(%rip),%ymm12        # 4950 <_sk_callback_hsw+0x2f8>
+  .byte  196,98,125,24,37,153,35,0,0         // vbroadcastss  0x2399(%rip),%ymm12        # 494c <_sk_callback_hsw+0x2f8>
   .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  196,98,125,24,29,126,35,0,0         // vbroadcastss  0x237e(%rip),%ymm11        # 4954 <_sk_callback_hsw+0x2fc>
+  .byte  196,98,125,24,29,122,35,0,0         // vbroadcastss  0x237a(%rip),%ymm11        # 4950 <_sk_callback_hsw+0x2fc>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,116,35,0,0         // vbroadcastss  0x2374(%rip),%ymm11        # 4958 <_sk_callback_hsw+0x300>
+  .byte  196,98,125,24,29,112,35,0,0         // vbroadcastss  0x2370(%rip),%ymm11        # 4954 <_sk_callback_hsw+0x300>
   .byte  196,98,45,172,218                   // vfnmadd213ps  %ymm2,%ymm10,%ymm11
-  .byte  196,226,125,24,21,106,35,0,0        // vbroadcastss  0x236a(%rip),%ymm2        # 495c <_sk_callback_hsw+0x304>
+  .byte  196,226,125,24,21,102,35,0,0        // vbroadcastss  0x2366(%rip),%ymm2        # 4958 <_sk_callback_hsw+0x304>
   .byte  196,193,108,92,210                  // vsubps        %ymm10,%ymm2,%ymm2
-  .byte  196,98,125,24,21,96,35,0,0          // vbroadcastss  0x2360(%rip),%ymm10        # 4960 <_sk_callback_hsw+0x308>
+  .byte  196,98,125,24,21,92,35,0,0          // vbroadcastss  0x235c(%rip),%ymm10        # 495c <_sk_callback_hsw+0x308>
   .byte  197,172,94,210                      // vdivps        %ymm2,%ymm10,%ymm2
   .byte  197,164,88,210                      // vaddps        %ymm2,%ymm11,%ymm2
-  .byte  196,98,125,24,21,83,35,0,0          // vbroadcastss  0x2353(%rip),%ymm10        # 4964 <_sk_callback_hsw+0x30c>
+  .byte  196,98,125,24,21,79,35,0,0          // vbroadcastss  0x234f(%rip),%ymm10        # 4960 <_sk_callback_hsw+0x30c>
   .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
@@ -11053,7 +11049,7 @@ _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  196,98,125,24,5,42,35,0,0           // vbroadcastss  0x232a(%rip),%ymm8        # 4968 <_sk_callback_hsw+0x310>
+  .byte  196,98,125,24,5,38,35,0,0           // vbroadcastss  0x2326(%rip),%ymm8        # 4964 <_sk_callback_hsw+0x310>
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11073,33 +11069,33 @@ _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  196,98,125,24,37,226,34,0,0         // vbroadcastss  0x22e2(%rip),%ymm12        # 496c <_sk_callback_hsw+0x314>
-  .byte  196,98,125,24,45,221,34,0,0         // vbroadcastss  0x22dd(%rip),%ymm13        # 4970 <_sk_callback_hsw+0x318>
+  .byte  196,98,125,24,37,222,34,0,0         // vbroadcastss  0x22de(%rip),%ymm12        # 4968 <_sk_callback_hsw+0x314>
+  .byte  196,98,125,24,45,217,34,0,0         // vbroadcastss  0x22d9(%rip),%ymm13        # 496c <_sk_callback_hsw+0x318>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,211,34,0,0         // vbroadcastss  0x22d3(%rip),%ymm13        # 4974 <_sk_callback_hsw+0x31c>
+  .byte  196,98,125,24,45,207,34,0,0         // vbroadcastss  0x22cf(%rip),%ymm13        # 4970 <_sk_callback_hsw+0x31c>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,201,34,0,0         // vbroadcastss  0x22c9(%rip),%ymm13        # 4978 <_sk_callback_hsw+0x320>
+  .byte  196,98,125,24,45,197,34,0,0         // vbroadcastss  0x22c5(%rip),%ymm13        # 4974 <_sk_callback_hsw+0x320>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,191,34,0,0         // vbroadcastss  0x22bf(%rip),%ymm11        # 497c <_sk_callback_hsw+0x324>
+  .byte  196,98,125,24,29,187,34,0,0         // vbroadcastss  0x22bb(%rip),%ymm11        # 4978 <_sk_callback_hsw+0x324>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,181,34,0,0         // vbroadcastss  0x22b5(%rip),%ymm12        # 4980 <_sk_callback_hsw+0x328>
+  .byte  196,98,125,24,37,177,34,0,0         // vbroadcastss  0x22b1(%rip),%ymm12        # 497c <_sk_callback_hsw+0x328>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,171,34,0,0         // vbroadcastss  0x22ab(%rip),%ymm12        # 4984 <_sk_callback_hsw+0x32c>
+  .byte  196,98,125,24,37,167,34,0,0         // vbroadcastss  0x22a7(%rip),%ymm12        # 4980 <_sk_callback_hsw+0x32c>
   .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  196,98,125,24,29,140,34,0,0         // vbroadcastss  0x228c(%rip),%ymm11        # 4988 <_sk_callback_hsw+0x330>
+  .byte  196,98,125,24,29,136,34,0,0         // vbroadcastss  0x2288(%rip),%ymm11        # 4984 <_sk_callback_hsw+0x330>
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,130,34,0,0         // vbroadcastss  0x2282(%rip),%ymm11        # 498c <_sk_callback_hsw+0x334>
+  .byte  196,98,125,24,29,126,34,0,0         // vbroadcastss  0x227e(%rip),%ymm11        # 4988 <_sk_callback_hsw+0x334>
   .byte  196,98,45,172,219                   // vfnmadd213ps  %ymm3,%ymm10,%ymm11
-  .byte  196,226,125,24,29,120,34,0,0        // vbroadcastss  0x2278(%rip),%ymm3        # 4990 <_sk_callback_hsw+0x338>
+  .byte  196,226,125,24,29,116,34,0,0        // vbroadcastss  0x2274(%rip),%ymm3        # 498c <_sk_callback_hsw+0x338>
   .byte  196,193,100,92,218                  // vsubps        %ymm10,%ymm3,%ymm3
-  .byte  196,98,125,24,21,110,34,0,0         // vbroadcastss  0x226e(%rip),%ymm10        # 4994 <_sk_callback_hsw+0x33c>
+  .byte  196,98,125,24,21,106,34,0,0         // vbroadcastss  0x226a(%rip),%ymm10        # 4990 <_sk_callback_hsw+0x33c>
   .byte  197,172,94,219                      // vdivps        %ymm3,%ymm10,%ymm3
   .byte  197,164,88,219                      // vaddps        %ymm3,%ymm11,%ymm3
-  .byte  196,98,125,24,21,97,34,0,0          // vbroadcastss  0x2261(%rip),%ymm10        # 4998 <_sk_callback_hsw+0x340>
+  .byte  196,98,125,24,21,93,34,0,0          // vbroadcastss  0x225d(%rip),%ymm10        # 4994 <_sk_callback_hsw+0x340>
   .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
@@ -11107,7 +11103,7 @@ _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  196,98,125,24,5,56,34,0,0           // vbroadcastss  0x2238(%rip),%ymm8        # 499c <_sk_callback_hsw+0x344>
+  .byte  196,98,125,24,5,52,34,0,0           // vbroadcastss  0x2234(%rip),%ymm8        # 4998 <_sk_callback_hsw+0x344>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11116,26 +11112,26 @@ HIDDEN _sk_lab_to_xyz_hsw
 .globl _sk_lab_to_xyz_hsw
 FUNCTION(_sk_lab_to_xyz_hsw)
 _sk_lab_to_xyz_hsw:
-  .byte  196,98,125,24,5,42,34,0,0           // vbroadcastss  0x222a(%rip),%ymm8        # 49a0 <_sk_callback_hsw+0x348>
-  .byte  196,98,125,24,13,37,34,0,0          // vbroadcastss  0x2225(%rip),%ymm9        # 49a4 <_sk_callback_hsw+0x34c>
-  .byte  196,98,125,24,21,32,34,0,0          // vbroadcastss  0x2220(%rip),%ymm10        # 49a8 <_sk_callback_hsw+0x350>
+  .byte  196,98,125,24,5,38,34,0,0           // vbroadcastss  0x2226(%rip),%ymm8        # 499c <_sk_callback_hsw+0x348>
+  .byte  196,98,125,24,13,33,34,0,0          // vbroadcastss  0x2221(%rip),%ymm9        # 49a0 <_sk_callback_hsw+0x34c>
+  .byte  196,98,125,24,21,28,34,0,0          // vbroadcastss  0x221c(%rip),%ymm10        # 49a4 <_sk_callback_hsw+0x350>
   .byte  196,194,53,168,202                  // vfmadd213ps   %ymm10,%ymm9,%ymm1
   .byte  196,194,53,168,210                  // vfmadd213ps   %ymm10,%ymm9,%ymm2
-  .byte  196,98,125,24,13,17,34,0,0          // vbroadcastss  0x2211(%rip),%ymm9        # 49ac <_sk_callback_hsw+0x354>
+  .byte  196,98,125,24,13,13,34,0,0          // vbroadcastss  0x220d(%rip),%ymm9        # 49a8 <_sk_callback_hsw+0x354>
   .byte  196,66,125,184,200                  // vfmadd231ps   %ymm8,%ymm0,%ymm9
-  .byte  196,226,125,24,5,7,34,0,0           // vbroadcastss  0x2207(%rip),%ymm0        # 49b0 <_sk_callback_hsw+0x358>
+  .byte  196,226,125,24,5,3,34,0,0           // vbroadcastss  0x2203(%rip),%ymm0        # 49ac <_sk_callback_hsw+0x358>
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
-  .byte  196,98,125,24,5,254,33,0,0          // vbroadcastss  0x21fe(%rip),%ymm8        # 49b4 <_sk_callback_hsw+0x35c>
+  .byte  196,98,125,24,5,250,33,0,0          // vbroadcastss  0x21fa(%rip),%ymm8        # 49b0 <_sk_callback_hsw+0x35c>
   .byte  196,98,117,168,192                  // vfmadd213ps   %ymm0,%ymm1,%ymm8
-  .byte  196,98,125,24,13,244,33,0,0         // vbroadcastss  0x21f4(%rip),%ymm9        # 49b8 <_sk_callback_hsw+0x360>
+  .byte  196,98,125,24,13,240,33,0,0         // vbroadcastss  0x21f0(%rip),%ymm9        # 49b4 <_sk_callback_hsw+0x360>
   .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  196,226,125,24,21,225,33,0,0        // vbroadcastss  0x21e1(%rip),%ymm2        # 49bc <_sk_callback_hsw+0x364>
+  .byte  196,226,125,24,21,221,33,0,0        // vbroadcastss  0x21dd(%rip),%ymm2        # 49b8 <_sk_callback_hsw+0x364>
   .byte  197,108,194,209,1                   // vcmpltps      %ymm1,%ymm2,%ymm10
-  .byte  196,98,125,24,29,215,33,0,0         // vbroadcastss  0x21d7(%rip),%ymm11        # 49c0 <_sk_callback_hsw+0x368>
+  .byte  196,98,125,24,29,211,33,0,0         // vbroadcastss  0x21d3(%rip),%ymm11        # 49bc <_sk_callback_hsw+0x368>
   .byte  196,65,60,88,195                    // vaddps        %ymm11,%ymm8,%ymm8
-  .byte  196,98,125,24,37,205,33,0,0         // vbroadcastss  0x21cd(%rip),%ymm12        # 49c4 <_sk_callback_hsw+0x36c>
+  .byte  196,98,125,24,37,201,33,0,0         // vbroadcastss  0x21c9(%rip),%ymm12        # 49c0 <_sk_callback_hsw+0x36c>
   .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
@@ -11150,9 +11146,9 @@ _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  196,226,125,24,5,130,33,0,0         // vbroadcastss  0x2182(%rip),%ymm0        # 49c8 <_sk_callback_hsw+0x370>
+  .byte  196,226,125,24,5,126,33,0,0         // vbroadcastss  0x217e(%rip),%ymm0        # 49c4 <_sk_callback_hsw+0x370>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,121,33,0,0          // vbroadcastss  0x2179(%rip),%ymm8        # 49cc <_sk_callback_hsw+0x374>
+  .byte  196,98,125,24,5,117,33,0,0          // vbroadcastss  0x2175(%rip),%ymm8        # 49c8 <_sk_callback_hsw+0x374>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11170,7 +11166,7 @@ _sk_load_a8_hsw:
   .byte  197,250,126,0                       // vmovq         (%rax),%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,78,33,0,0         // vbroadcastss  0x214e(%rip),%ymm1        # 49d0 <_sk_callback_hsw+0x378>
+  .byte  196,226,125,24,13,74,33,0,0         // vbroadcastss  0x214a(%rip),%ymm1        # 49cc <_sk_callback_hsw+0x378>
   .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
@@ -11235,7 +11231,7 @@ _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  196,226,125,24,13,89,32,0,0         // vbroadcastss  0x2059(%rip),%ymm1        # 49d4 <_sk_callback_hsw+0x37c>
+  .byte  196,226,125,24,13,85,32,0,0         // vbroadcastss  0x2055(%rip),%ymm1        # 49d0 <_sk_callback_hsw+0x37c>
   .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
@@ -11253,7 +11249,7 @@ FUNCTION(_sk_store_a8_hsw)
 _sk_store_a8_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,52,32,0,0           // vbroadcastss  0x2034(%rip),%ymm8        # 49d8 <_sk_callback_hsw+0x380>
+  .byte  196,98,125,24,5,48,32,0,0           // vbroadcastss  0x2030(%rip),%ymm8        # 49d4 <_sk_callback_hsw+0x380>
   .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
@@ -11320,10 +11316,10 @@ _sk_load_g8_hsw:
   .byte  197,250,126,0                       // vmovq         (%rax),%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,106,31,0,0        // vbroadcastss  0x1f6a(%rip),%ymm1        # 49dc <_sk_callback_hsw+0x384>
+  .byte  196,226,125,24,13,102,31,0,0        // vbroadcastss  0x1f66(%rip),%ymm1        # 49d8 <_sk_callback_hsw+0x384>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,95,31,0,0         // vbroadcastss  0x1f5f(%rip),%ymm3        # 49e0 <_sk_callback_hsw+0x388>
+  .byte  196,226,125,24,29,91,31,0,0         // vbroadcastss  0x1f5b(%rip),%ymm3        # 49dc <_sk_callback_hsw+0x388>
   .byte  76,137,193                          // mov           %r8,%rcx
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
@@ -11385,10 +11381,10 @@ _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  196,226,125,24,13,116,30,0,0        // vbroadcastss  0x1e74(%rip),%ymm1        # 49e4 <_sk_callback_hsw+0x38c>
+  .byte  196,226,125,24,13,112,30,0,0        // vbroadcastss  0x1e70(%rip),%ymm1        # 49e0 <_sk_callback_hsw+0x38c>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,105,30,0,0        // vbroadcastss  0x1e69(%rip),%ymm3        # 49e8 <_sk_callback_hsw+0x390>
+  .byte  196,226,125,24,29,101,30,0,0        // vbroadcastss  0x1e65(%rip),%ymm3        # 49e4 <_sk_callback_hsw+0x390>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -11444,14 +11440,14 @@ _sk_gather_i8_hsw:
   .byte  73,139,64,8                         // mov           0x8(%r8),%rax
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
   .byte  196,226,117,144,28,128              // vpgatherdd    %ymm1,(%rax,%ymm0,4),%ymm3
-  .byte  197,229,219,5,113,31,0,0            // vpand         0x1f71(%rip),%ymm3,%ymm0        # 4bc0 <_sk_callback_hsw+0x568>
+  .byte  197,229,219,5,113,31,0,0            // vpand         0x1f71(%rip),%ymm3,%ymm0        # 4bc0 <_sk_callback_hsw+0x56c>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,144,29,0,0          // vbroadcastss  0x1d90(%rip),%ymm8        # 49ec <_sk_callback_hsw+0x394>
+  .byte  196,98,125,24,5,140,29,0,0          // vbroadcastss  0x1d8c(%rip),%ymm8        # 49e8 <_sk_callback_hsw+0x394>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,118,31,0,0         // vpshufb       0x1f76(%rip),%ymm3,%ymm1        # 4be0 <_sk_callback_hsw+0x588>
+  .byte  196,226,101,0,13,118,31,0,0         // vpshufb       0x1f76(%rip),%ymm3,%ymm1        # 4be0 <_sk_callback_hsw+0x58c>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,132,31,0,0         // vpshufb       0x1f84(%rip),%ymm3,%ymm2        # 4c00 <_sk_callback_hsw+0x5a8>
+  .byte  196,226,101,0,21,132,31,0,0         // vpshufb       0x1f84(%rip),%ymm3,%ymm2        # 4c00 <_sk_callback_hsw+0x5ac>
   .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
@@ -11475,23 +11471,23 @@ _sk_load_565_hsw:
   .byte  117,114                             // jne           2d1c <_sk_load_565_hsw+0x7c>
   .byte  196,193,122,111,4,122               // vmovdqu       (%r10,%rdi,2),%xmm0
   .byte  196,226,125,51,208                  // vpmovzxwd     %xmm0,%ymm2
-  .byte  196,226,125,88,5,50,29,0,0          // vpbroadcastd  0x1d32(%rip),%ymm0        # 49f0 <_sk_callback_hsw+0x398>
+  .byte  196,226,125,88,5,46,29,0,0          // vpbroadcastd  0x1d2e(%rip),%ymm0        # 49ec <_sk_callback_hsw+0x398>
   .byte  197,237,219,192                     // vpand         %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,37,29,0,0         // vbroadcastss  0x1d25(%rip),%ymm1        # 49f4 <_sk_callback_hsw+0x39c>
+  .byte  196,226,125,24,13,33,29,0,0         // vbroadcastss  0x1d21(%rip),%ymm1        # 49f0 <_sk_callback_hsw+0x39c>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,28,29,0,0         // vpbroadcastd  0x1d1c(%rip),%ymm1        # 49f8 <_sk_callback_hsw+0x3a0>
+  .byte  196,226,125,88,13,24,29,0,0         // vpbroadcastd  0x1d18(%rip),%ymm1        # 49f4 <_sk_callback_hsw+0x3a0>
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,15,29,0,0         // vbroadcastss  0x1d0f(%rip),%ymm3        # 49fc <_sk_callback_hsw+0x3a4>
+  .byte  196,226,125,24,29,11,29,0,0         // vbroadcastss  0x1d0b(%rip),%ymm3        # 49f8 <_sk_callback_hsw+0x3a4>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,88,29,6,29,0,0          // vpbroadcastd  0x1d06(%rip),%ymm3        # 4a00 <_sk_callback_hsw+0x3a8>
+  .byte  196,226,125,88,29,2,29,0,0          // vpbroadcastd  0x1d02(%rip),%ymm3        # 49fc <_sk_callback_hsw+0x3a8>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,249,28,0,0        // vbroadcastss  0x1cf9(%rip),%ymm3        # 4a04 <_sk_callback_hsw+0x3ac>
+  .byte  196,226,125,24,29,245,28,0,0        // vbroadcastss  0x1cf5(%rip),%ymm3        # 4a00 <_sk_callback_hsw+0x3ac>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,238,28,0,0        // vbroadcastss  0x1cee(%rip),%ymm3        # 4a08 <_sk_callback_hsw+0x3b0>
+  .byte  196,226,125,24,29,234,28,0,0        // vbroadcastss  0x1cea(%rip),%ymm3        # 4a04 <_sk_callback_hsw+0x3b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,200                          // mov           %ecx,%r8d
   .byte  65,128,224,7                        // and           $0x7,%r8b
@@ -11582,23 +11578,23 @@ _sk_gather_565_hsw:
   .byte  65,15,183,4,88                      // movzwl        (%r8,%rbx,2),%eax
   .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,51,208                  // vpmovzxwd     %xmm0,%ymm2
-  .byte  196,226,125,88,5,177,27,0,0         // vpbroadcastd  0x1bb1(%rip),%ymm0        # 4a0c <_sk_callback_hsw+0x3b4>
+  .byte  196,226,125,88,5,173,27,0,0         // vpbroadcastd  0x1bad(%rip),%ymm0        # 4a08 <_sk_callback_hsw+0x3b4>
   .byte  197,237,219,192                     // vpand         %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,164,27,0,0        // vbroadcastss  0x1ba4(%rip),%ymm1        # 4a10 <_sk_callback_hsw+0x3b8>
+  .byte  196,226,125,24,13,160,27,0,0        // vbroadcastss  0x1ba0(%rip),%ymm1        # 4a0c <_sk_callback_hsw+0x3b8>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,155,27,0,0        // vpbroadcastd  0x1b9b(%rip),%ymm1        # 4a14 <_sk_callback_hsw+0x3bc>
+  .byte  196,226,125,88,13,151,27,0,0        // vpbroadcastd  0x1b97(%rip),%ymm1        # 4a10 <_sk_callback_hsw+0x3bc>
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,142,27,0,0        // vbroadcastss  0x1b8e(%rip),%ymm3        # 4a18 <_sk_callback_hsw+0x3c0>
+  .byte  196,226,125,24,29,138,27,0,0        // vbroadcastss  0x1b8a(%rip),%ymm3        # 4a14 <_sk_callback_hsw+0x3c0>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,88,29,133,27,0,0        // vpbroadcastd  0x1b85(%rip),%ymm3        # 4a1c <_sk_callback_hsw+0x3c4>
+  .byte  196,226,125,88,29,129,27,0,0        // vpbroadcastd  0x1b81(%rip),%ymm3        # 4a18 <_sk_callback_hsw+0x3c4>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,120,27,0,0        // vbroadcastss  0x1b78(%rip),%ymm3        # 4a20 <_sk_callback_hsw+0x3c8>
+  .byte  196,226,125,24,29,116,27,0,0        // vbroadcastss  0x1b74(%rip),%ymm3        # 4a1c <_sk_callback_hsw+0x3c8>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,109,27,0,0        // vbroadcastss  0x1b6d(%rip),%ymm3        # 4a24 <_sk_callback_hsw+0x3cc>
+  .byte  196,226,125,24,29,105,27,0,0        // vbroadcastss  0x1b69(%rip),%ymm3        # 4a20 <_sk_callback_hsw+0x3cc>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -11611,11 +11607,11 @@ FUNCTION(_sk_store_565_hsw)
 _sk_store_565_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,90,27,0,0           // vbroadcastss  0x1b5a(%rip),%ymm8        # 4a28 <_sk_callback_hsw+0x3d0>
+  .byte  196,98,125,24,5,86,27,0,0           // vbroadcastss  0x1b56(%rip),%ymm8        # 4a24 <_sk_callback_hsw+0x3d0>
   .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  196,98,125,24,21,69,27,0,0          // vbroadcastss  0x1b45(%rip),%ymm10        # 4a2c <_sk_callback_hsw+0x3d4>
+  .byte  196,98,125,24,21,65,27,0,0          // vbroadcastss  0x1b41(%rip),%ymm10        # 4a28 <_sk_callback_hsw+0x3d4>
   .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
@@ -11683,25 +11679,25 @@ _sk_load_4444_hsw:
   .byte  15,133,138,0,0,0                    // jne           3038 <_sk_load_4444_hsw+0x98>
   .byte  196,193,122,111,4,122               // vmovdqu       (%r10,%rdi,2),%xmm0
   .byte  196,226,125,51,216                  // vpmovzxwd     %xmm0,%ymm3
-  .byte  196,226,125,88,5,110,26,0,0         // vpbroadcastd  0x1a6e(%rip),%ymm0        # 4a30 <_sk_callback_hsw+0x3d8>
+  .byte  196,226,125,88,5,106,26,0,0         // vpbroadcastd  0x1a6a(%rip),%ymm0        # 4a2c <_sk_callback_hsw+0x3d8>
   .byte  197,229,219,192                     // vpand         %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,97,26,0,0         // vbroadcastss  0x1a61(%rip),%ymm1        # 4a34 <_sk_callback_hsw+0x3dc>
+  .byte  196,226,125,24,13,93,26,0,0         // vbroadcastss  0x1a5d(%rip),%ymm1        # 4a30 <_sk_callback_hsw+0x3dc>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,88,26,0,0         // vpbroadcastd  0x1a58(%rip),%ymm1        # 4a38 <_sk_callback_hsw+0x3e0>
+  .byte  196,226,125,88,13,84,26,0,0         // vpbroadcastd  0x1a54(%rip),%ymm1        # 4a34 <_sk_callback_hsw+0x3e0>
   .byte  197,229,219,201                     // vpand         %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,75,26,0,0         // vbroadcastss  0x1a4b(%rip),%ymm2        # 4a3c <_sk_callback_hsw+0x3e4>
+  .byte  196,226,125,24,21,71,26,0,0         // vbroadcastss  0x1a47(%rip),%ymm2        # 4a38 <_sk_callback_hsw+0x3e4>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,88,21,66,26,0,0         // vpbroadcastd  0x1a42(%rip),%ymm2        # 4a40 <_sk_callback_hsw+0x3e8>
+  .byte  196,226,125,88,21,62,26,0,0         // vpbroadcastd  0x1a3e(%rip),%ymm2        # 4a3c <_sk_callback_hsw+0x3e8>
   .byte  197,229,219,210                     // vpand         %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,53,26,0,0           // vbroadcastss  0x1a35(%rip),%ymm8        # 4a44 <_sk_callback_hsw+0x3ec>
+  .byte  196,98,125,24,5,49,26,0,0           // vbroadcastss  0x1a31(%rip),%ymm8        # 4a40 <_sk_callback_hsw+0x3ec>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,88,5,43,26,0,0           // vpbroadcastd  0x1a2b(%rip),%ymm8        # 4a48 <_sk_callback_hsw+0x3f0>
+  .byte  196,98,125,88,5,39,26,0,0           // vpbroadcastd  0x1a27(%rip),%ymm8        # 4a44 <_sk_callback_hsw+0x3f0>
   .byte  196,193,101,219,216                 // vpand         %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,29,26,0,0           // vbroadcastss  0x1a1d(%rip),%ymm8        # 4a4c <_sk_callback_hsw+0x3f4>
+  .byte  196,98,125,24,5,25,26,0,0           // vbroadcastss  0x1a19(%rip),%ymm8        # 4a48 <_sk_callback_hsw+0x3f4>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11794,25 +11790,25 @@ _sk_gather_4444_hsw:
   .byte  65,15,183,4,88                      // movzwl        (%r8,%rbx,2),%eax
   .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,51,216                  // vpmovzxwd     %xmm0,%ymm3
-  .byte  196,226,125,88,5,213,24,0,0         // vpbroadcastd  0x18d5(%rip),%ymm0        # 4a50 <_sk_callback_hsw+0x3f8>
+  .byte  196,226,125,88,5,209,24,0,0         // vpbroadcastd  0x18d1(%rip),%ymm0        # 4a4c <_sk_callback_hsw+0x3f8>
   .byte  197,229,219,192                     // vpand         %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,200,24,0,0        // vbroadcastss  0x18c8(%rip),%ymm1        # 4a54 <_sk_callback_hsw+0x3fc>
+  .byte  196,226,125,24,13,196,24,0,0        // vbroadcastss  0x18c4(%rip),%ymm1        # 4a50 <_sk_callback_hsw+0x3fc>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,191,24,0,0        // vpbroadcastd  0x18bf(%rip),%ymm1        # 4a58 <_sk_callback_hsw+0x400>
+  .byte  196,226,125,88,13,187,24,0,0        // vpbroadcastd  0x18bb(%rip),%ymm1        # 4a54 <_sk_callback_hsw+0x400>
   .byte  197,229,219,201                     // vpand         %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,178,24,0,0        // vbroadcastss  0x18b2(%rip),%ymm2        # 4a5c <_sk_callback_hsw+0x404>
+  .byte  196,226,125,24,21,174,24,0,0        // vbroadcastss  0x18ae(%rip),%ymm2        # 4a58 <_sk_callback_hsw+0x404>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,88,21,169,24,0,0        // vpbroadcastd  0x18a9(%rip),%ymm2        # 4a60 <_sk_callback_hsw+0x408>
+  .byte  196,226,125,88,21,165,24,0,0        // vpbroadcastd  0x18a5(%rip),%ymm2        # 4a5c <_sk_callback_hsw+0x408>
   .byte  197,229,219,210                     // vpand         %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,156,24,0,0          // vbroadcastss  0x189c(%rip),%ymm8        # 4a64 <_sk_callback_hsw+0x40c>
+  .byte  196,98,125,24,5,152,24,0,0          // vbroadcastss  0x1898(%rip),%ymm8        # 4a60 <_sk_callback_hsw+0x40c>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,88,5,146,24,0,0          // vpbroadcastd  0x1892(%rip),%ymm8        # 4a68 <_sk_callback_hsw+0x410>
+  .byte  196,98,125,88,5,142,24,0,0          // vpbroadcastd  0x188e(%rip),%ymm8        # 4a64 <_sk_callback_hsw+0x410>
   .byte  196,193,101,219,216                 // vpand         %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,132,24,0,0          // vbroadcastss  0x1884(%rip),%ymm8        # 4a6c <_sk_callback_hsw+0x414>
+  .byte  196,98,125,24,5,128,24,0,0          // vbroadcastss  0x1880(%rip),%ymm8        # 4a68 <_sk_callback_hsw+0x414>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -11827,7 +11823,7 @@ FUNCTION(_sk_store_4444_hsw)
 _sk_store_4444_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,106,24,0,0          // vbroadcastss  0x186a(%rip),%ymm8        # 4a70 <_sk_callback_hsw+0x418>
+  .byte  196,98,125,24,5,102,24,0,0          // vbroadcastss  0x1866(%rip),%ymm8        # 4a6c <_sk_callback_hsw+0x418>
   .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
@@ -11903,14 +11899,14 @@ _sk_load_8888_hsw:
   .byte  77,133,192                          // test          %r8,%r8
   .byte  117,88                              // jne           3351 <_sk_load_8888_hsw+0x6d>
   .byte  196,193,126,111,25                  // vmovdqu       (%r9),%ymm3
-  .byte  197,229,219,5,26,25,0,0             // vpand         0x191a(%rip),%ymm3,%ymm0        # 4c20 <_sk_callback_hsw+0x5c8>
+  .byte  197,229,219,5,26,25,0,0             // vpand         0x191a(%rip),%ymm3,%ymm0        # 4c20 <_sk_callback_hsw+0x5cc>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,97,23,0,0           // vbroadcastss  0x1761(%rip),%ymm8        # 4a74 <_sk_callback_hsw+0x41c>
+  .byte  196,98,125,24,5,93,23,0,0           // vbroadcastss  0x175d(%rip),%ymm8        # 4a70 <_sk_callback_hsw+0x41c>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,31,25,0,0          // vpshufb       0x191f(%rip),%ymm3,%ymm1        # 4c40 <_sk_callback_hsw+0x5e8>
+  .byte  196,226,101,0,13,31,25,0,0          // vpshufb       0x191f(%rip),%ymm3,%ymm1        # 4c40 <_sk_callback_hsw+0x5ec>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,45,25,0,0          // vpshufb       0x192d(%rip),%ymm3,%ymm2        # 4c60 <_sk_callback_hsw+0x608>
+  .byte  196,226,101,0,21,45,25,0,0          // vpshufb       0x192d(%rip),%ymm3,%ymm2        # 4c60 <_sk_callback_hsw+0x60c>
   .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
@@ -11942,14 +11938,14 @@ _sk_gather_8888_hsw:
   .byte  197,245,254,192                     // vpaddd        %ymm0,%ymm1,%ymm0
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
   .byte  196,194,117,144,28,128              // vpgatherdd    %ymm1,(%r8,%ymm0,4),%ymm3
-  .byte  197,229,219,5,219,24,0,0            // vpand         0x18db(%rip),%ymm3,%ymm0        # 4c80 <_sk_callback_hsw+0x628>
+  .byte  197,229,219,5,219,24,0,0            // vpand         0x18db(%rip),%ymm3,%ymm0        # 4c80 <_sk_callback_hsw+0x62c>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,198,22,0,0          // vbroadcastss  0x16c6(%rip),%ymm8        # 4a78 <_sk_callback_hsw+0x420>
+  .byte  196,98,125,24,5,194,22,0,0          // vbroadcastss  0x16c2(%rip),%ymm8        # 4a74 <_sk_callback_hsw+0x420>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,224,24,0,0         // vpshufb       0x18e0(%rip),%ymm3,%ymm1        # 4ca0 <_sk_callback_hsw+0x648>
+  .byte  196,226,101,0,13,224,24,0,0         // vpshufb       0x18e0(%rip),%ymm3,%ymm1        # 4ca0 <_sk_callback_hsw+0x64c>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,238,24,0,0         // vpshufb       0x18ee(%rip),%ymm3,%ymm2        # 4cc0 <_sk_callback_hsw+0x668>
+  .byte  196,226,101,0,21,238,24,0,0         // vpshufb       0x18ee(%rip),%ymm3,%ymm2        # 4cc0 <_sk_callback_hsw+0x66c>
   .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
@@ -11966,7 +11962,7 @@ _sk_store_8888_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,141,12,189,0,0,0,0               // lea           0x0(,%rdi,4),%r9
   .byte  76,3,8                              // add           (%rax),%r9
-  .byte  196,98,125,24,5,118,22,0,0          // vbroadcastss  0x1676(%rip),%ymm8        # 4a7c <_sk_callback_hsw+0x424>
+  .byte  196,98,125,24,5,114,22,0,0          // vbroadcastss  0x1672(%rip),%ymm8        # 4a78 <_sk_callback_hsw+0x424>
   .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
@@ -12163,7 +12159,7 @@ _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  196,98,125,24,21,109,19,0,0         // vbroadcastss  0x136d(%rip),%ymm10        # 4a80 <_sk_callback_hsw+0x428>
+  .byte  196,98,125,24,21,105,19,0,0         // vbroadcastss  0x1369(%rip),%ymm10        # 4a7c <_sk_callback_hsw+0x428>
   .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
@@ -12247,7 +12243,7 @@ _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  196,98,125,24,21,254,17,0,0         // vbroadcastss  0x11fe(%rip),%ymm10        # 4a84 <_sk_callback_hsw+0x42c>
+  .byte  196,98,125,24,21,250,17,0,0         // vbroadcastss  0x11fa(%rip),%ymm10        # 4a80 <_sk_callback_hsw+0x42c>
   .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
@@ -12264,7 +12260,7 @@ _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  196,226,125,24,29,178,17,0,0        // vbroadcastss  0x11b2(%rip),%ymm3        # 4a88 <_sk_callback_hsw+0x430>
+  .byte  196,226,125,24,29,174,17,0,0        // vbroadcastss  0x11ae(%rip),%ymm3        # 4a84 <_sk_callback_hsw+0x430>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,4,64                // vmovd         (%r8,%rax,2),%xmm0
   .byte  196,193,121,196,68,64,4,2           // vpinsrw       $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
@@ -12307,7 +12303,7 @@ _sk_store_u16_be_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  72,141,4,189,0,0,0,0                // lea           0x0(,%rdi,4),%rax
-  .byte  196,98,125,24,5,239,16,0,0          // vbroadcastss  0x10ef(%rip),%ymm8        # 4a8c <_sk_callback_hsw+0x434>
+  .byte  196,98,125,24,5,235,16,0,0          // vbroadcastss  0x10eb(%rip),%ymm8        # 4a88 <_sk_callback_hsw+0x434>
   .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
@@ -12567,11 +12563,11 @@ HIDDEN _sk_luminance_to_alpha_hsw
 .globl _sk_luminance_to_alpha_hsw
 FUNCTION(_sk_luminance_to_alpha_hsw)
 _sk_luminance_to_alpha_hsw:
-  .byte  196,226,125,24,29,63,13,0,0         // vbroadcastss  0xd3f(%rip),%ymm3        # 4a90 <_sk_callback_hsw+0x438>
-  .byte  196,98,125,24,5,58,13,0,0           // vbroadcastss  0xd3a(%rip),%ymm8        # 4a94 <_sk_callback_hsw+0x43c>
+  .byte  196,226,125,24,29,59,13,0,0         // vbroadcastss  0xd3b(%rip),%ymm3        # 4a8c <_sk_callback_hsw+0x438>
+  .byte  196,98,125,24,5,54,13,0,0           // vbroadcastss  0xd36(%rip),%ymm8        # 4a90 <_sk_callback_hsw+0x43c>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  196,226,125,184,203                 // vfmadd231ps   %ymm3,%ymm0,%ymm1
-  .byte  196,226,125,24,29,43,13,0,0         // vbroadcastss  0xd2b(%rip),%ymm3        # 4a98 <_sk_callback_hsw+0x440>
+  .byte  196,226,125,24,29,39,13,0,0         // vbroadcastss  0xd27(%rip),%ymm3        # 4a94 <_sk_callback_hsw+0x440>
   .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
@@ -12786,7 +12782,7 @@ _sk_gradient_hsw:
   .byte  76,139,72,72                        // mov           0x48(%rax),%r9
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  65,186,1,0,0,0                      // mov           $0x1,%r10d
-  .byte  196,226,125,24,21,213,9,0,0         // vbroadcastss  0x9d5(%rip),%ymm2        # 4a9c <_sk_callback_hsw+0x444>
+  .byte  196,226,125,24,21,209,9,0,0         // vbroadcastss  0x9d1(%rip),%ymm2        # 4a98 <_sk_callback_hsw+0x444>
   .byte  196,65,53,239,201                   // vpxor         %ymm9,%ymm9,%ymm9
   .byte  196,130,125,24,28,145               // vbroadcastss  (%r9,%r10,4),%ymm3
   .byte  197,228,194,216,2                   // vcmpleps      %ymm0,%ymm3,%ymm3
@@ -12881,24 +12877,24 @@ _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  196,98,125,24,45,84,8,0,0           // vbroadcastss  0x854(%rip),%ymm13        # 4aa0 <_sk_callback_hsw+0x448>
-  .byte  196,98,125,24,53,79,8,0,0           // vbroadcastss  0x84f(%rip),%ymm14        # 4aa4 <_sk_callback_hsw+0x44c>
+  .byte  196,98,125,24,45,80,8,0,0           // vbroadcastss  0x850(%rip),%ymm13        # 4a9c <_sk_callback_hsw+0x448>
+  .byte  196,98,125,24,53,75,8,0,0           // vbroadcastss  0x84b(%rip),%ymm14        # 4aa0 <_sk_callback_hsw+0x44c>
   .byte  196,66,29,184,245                   // vfmadd231ps   %ymm13,%ymm12,%ymm14
-  .byte  196,98,125,24,45,69,8,0,0           // vbroadcastss  0x845(%rip),%ymm13        # 4aa8 <_sk_callback_hsw+0x450>
+  .byte  196,98,125,24,45,65,8,0,0           // vbroadcastss  0x841(%rip),%ymm13        # 4aa4 <_sk_callback_hsw+0x450>
   .byte  196,66,29,184,238                   // vfmadd231ps   %ymm14,%ymm12,%ymm13
-  .byte  196,98,125,24,53,59,8,0,0           // vbroadcastss  0x83b(%rip),%ymm14        # 4aac <_sk_callback_hsw+0x454>
+  .byte  196,98,125,24,53,55,8,0,0           // vbroadcastss  0x837(%rip),%ymm14        # 4aa8 <_sk_callback_hsw+0x454>
   .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  196,98,125,24,21,38,8,0,0           // vbroadcastss  0x826(%rip),%ymm10        # 4ab0 <_sk_callback_hsw+0x458>
+  .byte  196,98,125,24,21,34,8,0,0           // vbroadcastss  0x822(%rip),%ymm10        # 4aac <_sk_callback_hsw+0x458>
   .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  196,98,125,24,21,16,8,0,0           // vbroadcastss  0x810(%rip),%ymm10        # 4ab4 <_sk_callback_hsw+0x45c>
+  .byte  196,98,125,24,21,12,8,0,0           // vbroadcastss  0x80c(%rip),%ymm10        # 4ab0 <_sk_callback_hsw+0x45c>
   .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  196,98,125,24,21,250,7,0,0          // vbroadcastss  0x7fa(%rip),%ymm10        # 4ab8 <_sk_callback_hsw+0x460>
+  .byte  196,98,125,24,21,246,7,0,0          // vbroadcastss  0x7f6(%rip),%ymm10        # 4ab4 <_sk_callback_hsw+0x460>
   .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
@@ -12912,8 +12908,7 @@ FUNCTION(_sk_xy_to_radius_hsw)
 _sk_xy_to_radius_hsw:
   .byte  197,116,89,193                      // vmulps        %ymm1,%ymm1,%ymm8
   .byte  196,98,125,184,192                  // vfmadd231ps   %ymm0,%ymm0,%ymm8
-  .byte  196,193,124,82,192                  // vrsqrtps      %ymm8,%ymm0
-  .byte  197,252,83,192                      // vrcpps        %ymm0,%ymm0
+  .byte  196,193,124,81,192                  // vsqrtps       %ymm8,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -12922,7 +12917,7 @@ HIDDEN _sk_save_xy_hsw
 FUNCTION(_sk_save_xy_hsw)
 _sk_save_xy_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,195,7,0,0           // vbroadcastss  0x7c3(%rip),%ymm8        # 4abc <_sk_callback_hsw+0x464>
+  .byte  196,98,125,24,5,195,7,0,0           // vbroadcastss  0x7c3(%rip),%ymm8        # 4ab8 <_sk_callback_hsw+0x464>
   .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
@@ -12956,9 +12951,9 @@ HIDDEN _sk_bilinear_nx_hsw
 FUNCTION(_sk_bilinear_nx_hsw)
 _sk_bilinear_nx_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,87,7,0,0           // vbroadcastss  0x757(%rip),%ymm0        # 4ac0 <_sk_callback_hsw+0x468>
+  .byte  196,226,125,24,5,87,7,0,0           // vbroadcastss  0x757(%rip),%ymm0        # 4abc <_sk_callback_hsw+0x468>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,78,7,0,0            // vbroadcastss  0x74e(%rip),%ymm8        # 4ac4 <_sk_callback_hsw+0x46c>
+  .byte  196,98,125,24,5,78,7,0,0            // vbroadcastss  0x74e(%rip),%ymm8        # 4ac0 <_sk_callback_hsw+0x46c>
   .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
@@ -12969,7 +12964,7 @@ HIDDEN _sk_bilinear_px_hsw
 FUNCTION(_sk_bilinear_px_hsw)
 _sk_bilinear_px_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,54,7,0,0           // vbroadcastss  0x736(%rip),%ymm0        # 4ac8 <_sk_callback_hsw+0x470>
+  .byte  196,226,125,24,5,54,7,0,0           // vbroadcastss  0x736(%rip),%ymm0        # 4ac4 <_sk_callback_hsw+0x470>
   .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)
@@ -12981,9 +12976,9 @@ HIDDEN _sk_bilinear_ny_hsw
 FUNCTION(_sk_bilinear_ny_hsw)
 _sk_bilinear_ny_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,26,7,0,0          // vbroadcastss  0x71a(%rip),%ymm1        # 4acc <_sk_callback_hsw+0x474>
+  .byte  196,226,125,24,13,26,7,0,0          // vbroadcastss  0x71a(%rip),%ymm1        # 4ac8 <_sk_callback_hsw+0x474>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,16,7,0,0            // vbroadcastss  0x710(%rip),%ymm8        # 4ad0 <_sk_callback_hsw+0x478>
+  .byte  196,98,125,24,5,16,7,0,0            // vbroadcastss  0x710(%rip),%ymm8        # 4acc <_sk_callback_hsw+0x478>
   .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
@@ -12994,7 +12989,7 @@ HIDDEN _sk_bilinear_py_hsw
 FUNCTION(_sk_bilinear_py_hsw)
 _sk_bilinear_py_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,248,6,0,0         // vbroadcastss  0x6f8(%rip),%ymm1        # 4ad4 <_sk_callback_hsw+0x47c>
+  .byte  196,226,125,24,13,248,6,0,0         // vbroadcastss  0x6f8(%rip),%ymm1        # 4ad0 <_sk_callback_hsw+0x47c>
   .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)
@@ -13006,13 +13001,13 @@ HIDDEN _sk_bicubic_n3x_hsw
 FUNCTION(_sk_bicubic_n3x_hsw)
 _sk_bicubic_n3x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,219,6,0,0          // vbroadcastss  0x6db(%rip),%ymm0        # 4ad8 <_sk_callback_hsw+0x480>
+  .byte  196,226,125,24,5,219,6,0,0          // vbroadcastss  0x6db(%rip),%ymm0        # 4ad4 <_sk_callback_hsw+0x480>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,210,6,0,0           // vbroadcastss  0x6d2(%rip),%ymm8        # 4adc <_sk_callback_hsw+0x484>
+  .byte  196,98,125,24,5,210,6,0,0           // vbroadcastss  0x6d2(%rip),%ymm8        # 4ad8 <_sk_callback_hsw+0x484>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,195,6,0,0          // vbroadcastss  0x6c3(%rip),%ymm10        # 4ae0 <_sk_callback_hsw+0x488>
-  .byte  196,98,125,24,29,190,6,0,0          // vbroadcastss  0x6be(%rip),%ymm11        # 4ae4 <_sk_callback_hsw+0x48c>
+  .byte  196,98,125,24,21,195,6,0,0          // vbroadcastss  0x6c3(%rip),%ymm10        # 4adc <_sk_callback_hsw+0x488>
+  .byte  196,98,125,24,29,190,6,0,0          // vbroadcastss  0x6be(%rip),%ymm11        # 4ae0 <_sk_callback_hsw+0x48c>
   .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)
@@ -13024,16 +13019,16 @@ HIDDEN _sk_bicubic_n1x_hsw
 FUNCTION(_sk_bicubic_n1x_hsw)
 _sk_bicubic_n1x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,161,6,0,0          // vbroadcastss  0x6a1(%rip),%ymm0        # 4ae8 <_sk_callback_hsw+0x490>
+  .byte  196,226,125,24,5,161,6,0,0          // vbroadcastss  0x6a1(%rip),%ymm0        # 4ae4 <_sk_callback_hsw+0x490>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,152,6,0,0           // vbroadcastss  0x698(%rip),%ymm8        # 4aec <_sk_callback_hsw+0x494>
+  .byte  196,98,125,24,5,152,6,0,0           // vbroadcastss  0x698(%rip),%ymm8        # 4ae8 <_sk_callback_hsw+0x494>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,142,6,0,0          // vbroadcastss  0x68e(%rip),%ymm9        # 4af0 <_sk_callback_hsw+0x498>
-  .byte  196,98,125,24,21,137,6,0,0          // vbroadcastss  0x689(%rip),%ymm10        # 4af4 <_sk_callback_hsw+0x49c>
+  .byte  196,98,125,24,13,142,6,0,0          // vbroadcastss  0x68e(%rip),%ymm9        # 4aec <_sk_callback_hsw+0x498>
+  .byte  196,98,125,24,21,137,6,0,0          // vbroadcastss  0x689(%rip),%ymm10        # 4af0 <_sk_callback_hsw+0x49c>
   .byte  196,66,61,168,209                   // vfmadd213ps   %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,13,127,6,0,0          // vbroadcastss  0x67f(%rip),%ymm9        # 4af8 <_sk_callback_hsw+0x4a0>
+  .byte  196,98,125,24,13,127,6,0,0          // vbroadcastss  0x67f(%rip),%ymm9        # 4af4 <_sk_callback_hsw+0x4a0>
   .byte  196,66,61,184,202                   // vfmadd231ps   %ymm10,%ymm8,%ymm9
-  .byte  196,98,125,24,21,117,6,0,0          // vbroadcastss  0x675(%rip),%ymm10        # 4afc <_sk_callback_hsw+0x4a4>
+  .byte  196,98,125,24,21,117,6,0,0          // vbroadcastss  0x675(%rip),%ymm10        # 4af8 <_sk_callback_hsw+0x4a4>
   .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
@@ -13044,14 +13039,14 @@ HIDDEN _sk_bicubic_p1x_hsw
 FUNCTION(_sk_bicubic_p1x_hsw)
 _sk_bicubic_p1x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,93,6,0,0            // vbroadcastss  0x65d(%rip),%ymm8        # 4b00 <_sk_callback_hsw+0x4a8>
+  .byte  196,98,125,24,5,93,6,0,0            // vbroadcastss  0x65d(%rip),%ymm8        # 4afc <_sk_callback_hsw+0x4a8>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,72,64                    // vmovups       0x40(%rax),%ymm9
-  .byte  196,98,125,24,21,79,6,0,0           // vbroadcastss  0x64f(%rip),%ymm10        # 4b04 <_sk_callback_hsw+0x4ac>
-  .byte  196,98,125,24,29,74,6,0,0           // vbroadcastss  0x64a(%rip),%ymm11        # 4b08 <_sk_callback_hsw+0x4b0>
+  .byte  196,98,125,24,21,79,6,0,0           // vbroadcastss  0x64f(%rip),%ymm10        # 4b00 <_sk_callback_hsw+0x4ac>
+  .byte  196,98,125,24,29,74,6,0,0           // vbroadcastss  0x64a(%rip),%ymm11        # 4b04 <_sk_callback_hsw+0x4b0>
   .byte  196,66,53,168,218                   // vfmadd213ps   %ymm10,%ymm9,%ymm11
   .byte  196,66,53,168,216                   // vfmadd213ps   %ymm8,%ymm9,%ymm11
-  .byte  196,98,125,24,5,59,6,0,0            // vbroadcastss  0x63b(%rip),%ymm8        # 4b0c <_sk_callback_hsw+0x4b4>
+  .byte  196,98,125,24,5,59,6,0,0            // vbroadcastss  0x63b(%rip),%ymm8        # 4b08 <_sk_callback_hsw+0x4b4>
   .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
@@ -13062,12 +13057,12 @@ HIDDEN _sk_bicubic_p3x_hsw
 FUNCTION(_sk_bicubic_p3x_hsw)
 _sk_bicubic_p3x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,35,6,0,0           // vbroadcastss  0x623(%rip),%ymm0        # 4b10 <_sk_callback_hsw+0x4b8>
+  .byte  196,226,125,24,5,35,6,0,0           // vbroadcastss  0x623(%rip),%ymm0        # 4b0c <_sk_callback_hsw+0x4b8>
   .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  196,98,125,24,21,16,6,0,0           // vbroadcastss  0x610(%rip),%ymm10        # 4b14 <_sk_callback_hsw+0x4bc>
-  .byte  196,98,125,24,29,11,6,0,0           // vbroadcastss  0x60b(%rip),%ymm11        # 4b18 <_sk_callback_hsw+0x4c0>
+  .byte  196,98,125,24,21,16,6,0,0           // vbroadcastss  0x610(%rip),%ymm10        # 4b10 <_sk_callback_hsw+0x4bc>
+  .byte  196,98,125,24,29,11,6,0,0           // vbroadcastss  0x60b(%rip),%ymm11        # 4b14 <_sk_callback_hsw+0x4c0>
   .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)
@@ -13079,13 +13074,13 @@ HIDDEN _sk_bicubic_n3y_hsw
 FUNCTION(_sk_bicubic_n3y_hsw)
 _sk_bicubic_n3y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,238,5,0,0         // vbroadcastss  0x5ee(%rip),%ymm1        # 4b1c <_sk_callback_hsw+0x4c4>
+  .byte  196,226,125,24,13,238,5,0,0         // vbroadcastss  0x5ee(%rip),%ymm1        # 4b18 <_sk_callback_hsw+0x4c4>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,228,5,0,0           // vbroadcastss  0x5e4(%rip),%ymm8        # 4b20 <_sk_callback_hsw+0x4c8>
+  .byte  196,98,125,24,5,228,5,0,0           // vbroadcastss  0x5e4(%rip),%ymm8        # 4b1c <_sk_callback_hsw+0x4c8>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,213,5,0,0          // vbroadcastss  0x5d5(%rip),%ymm10        # 4b24 <_sk_callback_hsw+0x4cc>
-  .byte  196,98,125,24,29,208,5,0,0          // vbroadcastss  0x5d0(%rip),%ymm11        # 4b28 <_sk_callback_hsw+0x4d0>
+  .byte  196,98,125,24,21,213,5,0,0          // vbroadcastss  0x5d5(%rip),%ymm10        # 4b20 <_sk_callback_hsw+0x4cc>
+  .byte  196,98,125,24,29,208,5,0,0          // vbroadcastss  0x5d0(%rip),%ymm11        # 4b24 <_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  197,124,17,128,160,0,0,0            // vmovups       %ymm8,0xa0(%rax)
@@ -13097,16 +13092,16 @@ HIDDEN _sk_bicubic_n1y_hsw
 FUNCTION(_sk_bicubic_n1y_hsw)
 _sk_bicubic_n1y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,179,5,0,0         // vbroadcastss  0x5b3(%rip),%ymm1        # 4b2c <_sk_callback_hsw+0x4d4>
+  .byte  196,226,125,24,13,179,5,0,0         // vbroadcastss  0x5b3(%rip),%ymm1        # 4b28 <_sk_callback_hsw+0x4d4>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,169,5,0,0           // vbroadcastss  0x5a9(%rip),%ymm8        # 4b30 <_sk_callback_hsw+0x4d8>
+  .byte  196,98,125,24,5,169,5,0,0           // vbroadcastss  0x5a9(%rip),%ymm8        # 4b2c <_sk_callback_hsw+0x4d8>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,159,5,0,0          // vbroadcastss  0x59f(%rip),%ymm9        # 4b34 <_sk_callback_hsw+0x4dc>
-  .byte  196,98,125,24,21,154,5,0,0          // vbroadcastss  0x59a(%rip),%ymm10        # 4b38 <_sk_callback_hsw+0x4e0>
+  .byte  196,98,125,24,13,159,5,0,0          // vbroadcastss  0x59f(%rip),%ymm9        # 4b30 <_sk_callback_hsw+0x4dc>
+  .byte  196,98,125,24,21,154,5,0,0          // vbroadcastss  0x59a(%rip),%ymm10        # 4b34 <_sk_callback_hsw+0x4e0>
   .byte  196,66,61,168,209                   // vfmadd213ps   %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,13,144,5,0,0          // vbroadcastss  0x590(%rip),%ymm9        # 4b3c <_sk_callback_hsw+0x4e4>
+  .byte  196,98,125,24,13,144,5,0,0          // vbroadcastss  0x590(%rip),%ymm9        # 4b38 <_sk_callback_hsw+0x4e4>
   .byte  196,66,61,184,202                   // vfmadd231ps   %ymm10,%ymm8,%ymm9
-  .byte  196,98,125,24,21,134,5,0,0          // vbroadcastss  0x586(%rip),%ymm10        # 4b40 <_sk_callback_hsw+0x4e8>
+  .byte  196,98,125,24,21,134,5,0,0          // vbroadcastss  0x586(%rip),%ymm10        # 4b3c <_sk_callback_hsw+0x4e8>
   .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
@@ -13117,14 +13112,14 @@ HIDDEN _sk_bicubic_p1y_hsw
 FUNCTION(_sk_bicubic_p1y_hsw)
 _sk_bicubic_p1y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,110,5,0,0           // vbroadcastss  0x56e(%rip),%ymm8        # 4b44 <_sk_callback_hsw+0x4ec>
+  .byte  196,98,125,24,5,110,5,0,0           // vbroadcastss  0x56e(%rip),%ymm8        # 4b40 <_sk_callback_hsw+0x4ec>
   .byte  197,188,88,72,32                    // vaddps        0x20(%rax),%ymm8,%ymm1
   .byte  197,124,16,72,96                    // vmovups       0x60(%rax),%ymm9
-  .byte  196,98,125,24,21,95,5,0,0           // vbroadcastss  0x55f(%rip),%ymm10        # 4b48 <_sk_callback_hsw+0x4f0>
-  .byte  196,98,125,24,29,90,5,0,0           // vbroadcastss  0x55a(%rip),%ymm11        # 4b4c <_sk_callback_hsw+0x4f4>
+  .byte  196,98,125,24,21,95,5,0,0           // vbroadcastss  0x55f(%rip),%ymm10        # 4b44 <_sk_callback_hsw+0x4f0>
+  .byte  196,98,125,24,29,90,5,0,0           // vbroadcastss  0x55a(%rip),%ymm11        # 4b48 <_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  196,98,125,24,5,75,5,0,0            // vbroadcastss  0x54b(%rip),%ymm8        # 4b50 <_sk_callback_hsw+0x4f8>
+  .byte  196,98,125,24,5,75,5,0,0            // vbroadcastss  0x54b(%rip),%ymm8        # 4b4c <_sk_callback_hsw+0x4f8>
   .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
@@ -13135,12 +13130,12 @@ HIDDEN _sk_bicubic_p3y_hsw
 FUNCTION(_sk_bicubic_p3y_hsw)
 _sk_bicubic_p3y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,51,5,0,0          // vbroadcastss  0x533(%rip),%ymm1        # 4b54 <_sk_callback_hsw+0x4fc>
+  .byte  196,226,125,24,13,51,5,0,0          // vbroadcastss  0x533(%rip),%ymm1        # 4b50 <_sk_callback_hsw+0x4fc>
   .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  196,98,125,24,21,31,5,0,0           // vbroadcastss  0x51f(%rip),%ymm10        # 4b58 <_sk_callback_hsw+0x500>
-  .byte  196,98,125,24,29,26,5,0,0           // vbroadcastss  0x51a(%rip),%ymm11        # 4b5c <_sk_callback_hsw+0x504>
+  .byte  196,98,125,24,21,31,5,0,0           // vbroadcastss  0x51f(%rip),%ymm10        # 4b54 <_sk_callback_hsw+0x500>
+  .byte  196,98,125,24,29,26,5,0,0           // vbroadcastss  0x51a(%rip),%ymm11        # 4b58 <_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  197,124,17,128,160,0,0,0            // vmovups       %ymm8,0xa0(%rax)
@@ -13264,25 +13259,25 @@ 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 4831 <.literal4+0xb1>
+  .byte  71,225,61                           // rex.RXB       loope 482d <.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 4841 <.literal4+0xc1>
+  .byte  71,225,61                           // rex.RXB       loope 483d <.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 4851 <.literal4+0xd1>
+  .byte  71,225,61                           // rex.RXB       loope 484d <.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 4861 <.literal4+0xe1>
+  .byte  71,225,61                           // rex.RXB       loope 485d <.literal4+0xe1>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -13332,7 +13327,7 @@ 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        48b1 <.literal4+0x131>
+  .byte  224,7                               // loopne        48ad <.literal4+0x131>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -13348,10 +13343,10 @@ 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            48d8 <.literal4+0x158>
+  .byte  127,0                               // jg            48d4 <.literal4+0x158>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4951 <.literal4+0x1d1>
+  .byte  119,115                             // ja            494d <.literal4+0x1d1>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13365,10 +13360,10 @@ 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            490c <.literal4+0x18c>
+  .byte  127,0                               // jg            4908 <.literal4+0x18c>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4985 <.literal4+0x205>
+  .byte  119,115                             // ja            4981 <.literal4+0x205>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13382,10 +13377,10 @@ 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            4940 <.literal4+0x1c0>
+  .byte  127,0                               // jg            493c <.literal4+0x1c0>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            49b9 <.literal4+0x239>
+  .byte  119,115                             // ja            49b5 <.literal4+0x239>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13399,10 +13394,10 @@ 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            4974 <.literal4+0x1f4>
+  .byte  127,0                               // jg            4970 <.literal4+0x1f4>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            49ed <.literal4+0x26d>
+  .byte  119,115                             // ja            49e9 <.literal4+0x26d>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13415,7 +13410,7 @@ 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            49eb <.literal4+0x26b>
+  .byte  127,67                              // jg            49e7 <.literal4+0x26b>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,195                               // add           %al,%bl
   .byte  0,0                                 // add           %al,(%rax)
@@ -13427,10 +13422,10 @@ BALIGN4
   .byte  190,80,128,3,62                     // mov           $0x3e038050,%esi
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           4a0b <.literal4+0x28b>
+  .byte  118,63                              // jbe           4a07 <.literal4+0x28b>
   .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            4a1f <.literal4+0x29f>
+  .byte  127,67                              // jg            4a1b <.literal4+0x29f>
   .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)
@@ -13439,7 +13434,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        4a01 <.literal4+0x281>
+  .byte  224,7                               // loopne        49fd <.literal4+0x281>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -13451,7 +13446,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        4a1d <.literal4+0x29d>
+  .byte  224,7                               // loopne        4a19 <.literal4+0x29d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -13462,7 +13457,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            4a72 <.literal4+0x2f2>
+  .byte  124,66                              // jl            4a6e <.literal4+0x2f2>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  137,136,136,55,0,15                 // mov           %ecx,0xf003788(%rax)
@@ -13480,9 +13475,9 @@ 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            4ab5 <.literal4+0x335>
+  .byte  112,65                              // jo            4ab1 <.literal4+0x335>
   .byte  129,128,128,59,129,128,128,59,0,0   // addl          $0x3b80,-0x7f7ec480(%rax)
-  .byte  127,67                              // jg            4ac3 <.literal4+0x343>
+  .byte  127,67                              // jg            4abf <.literal4+0x343>
   .byte  128,0,128                           // addb          $0x80,(%rax)
   .byte  55                                  // (bad)
   .byte  128,0,128                           // addb          $0x80,(%rax)
@@ -13490,7 +13485,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  255                                 // (bad)
-  .byte  127,71                              // jg            4ad7 <.literal4+0x357>
+  .byte  127,71                              // jg            4ad3 <.literal4+0x357>
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,89                               // ds            pop %rcx
@@ -13590,16 +13585,16 @@ 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)        # a004b88 <_sk_callback_hsw+0xa000530>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004b88 <_sk_callback_hsw+0xa000534>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004b90 <_sk_callback_hsw+0x12000538>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004b90 <_sk_callback_hsw+0x1200053c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004b98 <_sk_callback_hsw+0x1a000540>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004b98 <_sk_callback_hsw+0x1a000544>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004ba0 <_sk_callback_hsw+0x3000548>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004ba0 <_sk_callback_hsw+0x300054c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13642,16 +13637,16 @@ 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)        # a004be8 <_sk_callback_hsw+0xa000590>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004be8 <_sk_callback_hsw+0xa000594>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004bf0 <_sk_callback_hsw+0x12000598>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004bf0 <_sk_callback_hsw+0x1200059c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004bf8 <_sk_callback_hsw+0x1a0005a0>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004bf8 <_sk_callback_hsw+0x1a0005a4>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004c00 <_sk_callback_hsw+0x30005a8>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004c00 <_sk_callback_hsw+0x30005ac>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13694,16 +13689,16 @@ 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)        # a004c48 <_sk_callback_hsw+0xa0005f0>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004c48 <_sk_callback_hsw+0xa0005f4>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004c50 <_sk_callback_hsw+0x120005f8>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004c50 <_sk_callback_hsw+0x120005fc>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004c58 <_sk_callback_hsw+0x1a000600>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004c58 <_sk_callback_hsw+0x1a000604>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004c60 <_sk_callback_hsw+0x3000608>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004c60 <_sk_callback_hsw+0x300060c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13746,16 +13741,16 @@ 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)        # a004ca8 <_sk_callback_hsw+0xa000650>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004ca8 <_sk_callback_hsw+0xa000654>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004cb0 <_sk_callback_hsw+0x12000658>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004cb0 <_sk_callback_hsw+0x1200065c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004cb8 <_sk_callback_hsw+0x1a000660>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004cb8 <_sk_callback_hsw+0x1a000664>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004cc0 <_sk_callback_hsw+0x3000668>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004cc0 <_sk_callback_hsw+0x300066c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13876,14 +13871,14 @@ _sk_seed_shader_avx:
   .byte  197,249,112,192,0                   // vpshufd       $0x0,%xmm0,%xmm0
   .byte  196,227,125,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,167,98,0,0        // vbroadcastss  0x62a7(%rip),%ymm1        # 6370 <_sk_callback_avx+0x126>
+  .byte  196,226,125,24,13,163,98,0,0        // vbroadcastss  0x62a3(%rip),%ymm1        # 636c <_sk_callback_avx+0x126>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  197,252,88,2                        // vaddps        (%rdx),%ymm0,%ymm0
   .byte  196,226,125,24,16                   // vbroadcastss  (%rax),%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  197,236,88,201                      // vaddps        %ymm1,%ymm2,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,21,139,98,0,0        // vbroadcastss  0x628b(%rip),%ymm2        # 6374 <_sk_callback_avx+0x12a>
+  .byte  196,226,125,24,21,135,98,0,0        // vbroadcastss  0x6287(%rip),%ymm2        # 6370 <_sk_callback_avx+0x12a>
   .byte  197,228,87,219                      // vxorps        %ymm3,%ymm3,%ymm3
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
   .byte  197,212,87,237                      // vxorps        %ymm5,%ymm5,%ymm5
@@ -13905,7 +13900,7 @@ _sk_dither_avx:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  196,66,125,24,8                     // vbroadcastss  (%r8),%ymm9
   .byte  196,65,60,87,209                    // vxorps        %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,29,67,98,0,0          // vbroadcastss  0x6243(%rip),%ymm11        # 6378 <_sk_callback_avx+0x12e>
+  .byte  196,98,125,24,29,63,98,0,0          // vbroadcastss  0x623f(%rip),%ymm11        # 6374 <_sk_callback_avx+0x12e>
   .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
@@ -13916,8 +13911,8 @@ _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  196,98,125,24,37,4,98,0,0           // vbroadcastss  0x6204(%rip),%ymm12        # 637c <_sk_callback_avx+0x132>
-  .byte  196,98,125,24,45,255,97,0,0         // vbroadcastss  0x61ff(%rip),%ymm13        # 6380 <_sk_callback_avx+0x136>
+  .byte  196,98,125,24,37,0,98,0,0           // vbroadcastss  0x6200(%rip),%ymm12        # 6378 <_sk_callback_avx+0x132>
+  .byte  196,98,125,24,45,251,97,0,0         // vbroadcastss  0x61fb(%rip),%ymm13        # 637c <_sk_callback_avx+0x136>
   .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
@@ -13944,9 +13939,9 @@ _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  196,98,125,24,13,106,97,0,0         // vbroadcastss  0x616a(%rip),%ymm9        # 6384 <_sk_callback_avx+0x13a>
+  .byte  196,98,125,24,13,102,97,0,0         // vbroadcastss  0x6166(%rip),%ymm9        # 6380 <_sk_callback_avx+0x13a>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,96,97,0,0          // vbroadcastss  0x6160(%rip),%ymm9        # 6388 <_sk_callback_avx+0x13e>
+  .byte  196,98,125,24,13,92,97,0,0          // vbroadcastss  0x615c(%rip),%ymm9        # 6384 <_sk_callback_avx+0x13e>
   .byte  196,65,60,88,193                    // vaddps        %ymm9,%ymm8,%ymm8
   .byte  196,98,125,24,72,8                  // vbroadcastss  0x8(%rax),%ymm9
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
@@ -14008,7 +14003,7 @@ HIDDEN _sk_srcatop_avx
 FUNCTION(_sk_srcatop_avx)
 _sk_srcatop_avx:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,212,96,0,0          // vbroadcastss  0x60d4(%rip),%ymm8        # 638c <_sk_callback_avx+0x142>
+  .byte  196,98,125,24,5,208,96,0,0          // vbroadcastss  0x60d0(%rip),%ymm8        # 6388 <_sk_callback_avx+0x142>
   .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
@@ -14029,7 +14024,7 @@ HIDDEN _sk_dstatop_avx
 FUNCTION(_sk_dstatop_avx)
 _sk_dstatop_avx:
   .byte  197,100,89,196                      // vmulps        %ymm4,%ymm3,%ymm8
-  .byte  196,98,125,24,13,150,96,0,0         // vbroadcastss  0x6096(%rip),%ymm9        # 6390 <_sk_callback_avx+0x146>
+  .byte  196,98,125,24,13,146,96,0,0         // vbroadcastss  0x6092(%rip),%ymm9        # 638c <_sk_callback_avx+0x146>
   .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
@@ -14071,7 +14066,7 @@ HIDDEN _sk_srcout_avx
 .globl _sk_srcout_avx
 FUNCTION(_sk_srcout_avx)
 _sk_srcout_avx:
-  .byte  196,98,125,24,5,53,96,0,0           // vbroadcastss  0x6035(%rip),%ymm8        # 6394 <_sk_callback_avx+0x14a>
+  .byte  196,98,125,24,5,49,96,0,0           // vbroadcastss  0x6031(%rip),%ymm8        # 6390 <_sk_callback_avx+0x14a>
   .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
@@ -14084,7 +14079,7 @@ HIDDEN _sk_dstout_avx
 .globl _sk_dstout_avx
 FUNCTION(_sk_dstout_avx)
 _sk_dstout_avx:
-  .byte  196,226,125,24,5,24,96,0,0          // vbroadcastss  0x6018(%rip),%ymm0        # 6398 <_sk_callback_avx+0x14e>
+  .byte  196,226,125,24,5,20,96,0,0          // vbroadcastss  0x6014(%rip),%ymm0        # 6394 <_sk_callback_avx+0x14e>
   .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
@@ -14097,7 +14092,7 @@ HIDDEN _sk_srcover_avx
 .globl _sk_srcover_avx
 FUNCTION(_sk_srcover_avx)
 _sk_srcover_avx:
-  .byte  196,98,125,24,5,251,95,0,0          // vbroadcastss  0x5ffb(%rip),%ymm8        # 639c <_sk_callback_avx+0x152>
+  .byte  196,98,125,24,5,247,95,0,0          // vbroadcastss  0x5ff7(%rip),%ymm8        # 6398 <_sk_callback_avx+0x152>
   .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
@@ -14114,7 +14109,7 @@ HIDDEN _sk_dstover_avx
 .globl _sk_dstover_avx
 FUNCTION(_sk_dstover_avx)
 _sk_dstover_avx:
-  .byte  196,98,125,24,5,206,95,0,0          // vbroadcastss  0x5fce(%rip),%ymm8        # 63a0 <_sk_callback_avx+0x156>
+  .byte  196,98,125,24,5,202,95,0,0          // vbroadcastss  0x5fca(%rip),%ymm8        # 639c <_sk_callback_avx+0x156>
   .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
@@ -14142,7 +14137,7 @@ HIDDEN _sk_multiply_avx
 .globl _sk_multiply_avx
 FUNCTION(_sk_multiply_avx)
 _sk_multiply_avx:
-  .byte  196,98,125,24,5,141,95,0,0          // vbroadcastss  0x5f8d(%rip),%ymm8        # 63a4 <_sk_callback_avx+0x15a>
+  .byte  196,98,125,24,5,137,95,0,0          // vbroadcastss  0x5f89(%rip),%ymm8        # 63a0 <_sk_callback_avx+0x15a>
   .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
@@ -14202,7 +14197,7 @@ HIDDEN _sk_xor__avx
 .globl _sk_xor__avx
 FUNCTION(_sk_xor__avx)
 _sk_xor__avx:
-  .byte  196,98,125,24,5,220,94,0,0          // vbroadcastss  0x5edc(%rip),%ymm8        # 63a8 <_sk_callback_avx+0x15e>
+  .byte  196,98,125,24,5,216,94,0,0          // vbroadcastss  0x5ed8(%rip),%ymm8        # 63a4 <_sk_callback_avx+0x15e>
   .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
@@ -14239,7 +14234,7 @@ _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  196,98,125,24,5,92,94,0,0           // vbroadcastss  0x5e5c(%rip),%ymm8        # 63ac <_sk_callback_avx+0x162>
+  .byte  196,98,125,24,5,88,94,0,0           // vbroadcastss  0x5e58(%rip),%ymm8        # 63a8 <_sk_callback_avx+0x162>
   .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
@@ -14265,7 +14260,7 @@ _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  196,98,125,24,5,8,94,0,0            // vbroadcastss  0x5e08(%rip),%ymm8        # 63b0 <_sk_callback_avx+0x166>
+  .byte  196,98,125,24,5,4,94,0,0            // vbroadcastss  0x5e04(%rip),%ymm8        # 63ac <_sk_callback_avx+0x166>
   .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
@@ -14294,7 +14289,7 @@ _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  196,98,125,24,5,168,93,0,0          // vbroadcastss  0x5da8(%rip),%ymm8        # 63b4 <_sk_callback_avx+0x16a>
+  .byte  196,98,125,24,5,164,93,0,0          // vbroadcastss  0x5da4(%rip),%ymm8        # 63b0 <_sk_callback_avx+0x16a>
   .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
@@ -14317,7 +14312,7 @@ _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  196,98,125,24,5,99,93,0,0           // vbroadcastss  0x5d63(%rip),%ymm8        # 63b8 <_sk_callback_avx+0x16e>
+  .byte  196,98,125,24,5,95,93,0,0           // vbroadcastss  0x5d5f(%rip),%ymm8        # 63b4 <_sk_callback_avx+0x16e>
   .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
@@ -14328,7 +14323,7 @@ HIDDEN _sk_colorburn_avx
 .globl _sk_colorburn_avx
 FUNCTION(_sk_colorburn_avx)
 _sk_colorburn_avx:
-  .byte  196,98,125,24,5,78,93,0,0           // vbroadcastss  0x5d4e(%rip),%ymm8        # 63bc <_sk_callback_avx+0x172>
+  .byte  196,98,125,24,5,74,93,0,0           // vbroadcastss  0x5d4a(%rip),%ymm8        # 63b8 <_sk_callback_avx+0x172>
   .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
@@ -14390,7 +14385,7 @@ HIDDEN _sk_colordodge_avx
 FUNCTION(_sk_colordodge_avx)
 _sk_colordodge_avx:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,13,74,92,0,0          // vbroadcastss  0x5c4a(%rip),%ymm9        # 63c0 <_sk_callback_avx+0x176>
+  .byte  196,98,125,24,13,70,92,0,0          // vbroadcastss  0x5c46(%rip),%ymm9        # 63bc <_sk_callback_avx+0x176>
   .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
@@ -14447,7 +14442,7 @@ HIDDEN _sk_hardlight_avx
 .globl _sk_hardlight_avx
 FUNCTION(_sk_hardlight_avx)
 _sk_hardlight_avx:
-  .byte  196,98,125,24,5,92,91,0,0           // vbroadcastss  0x5b5c(%rip),%ymm8        # 63c4 <_sk_callback_avx+0x17a>
+  .byte  196,98,125,24,5,88,91,0,0           // vbroadcastss  0x5b58(%rip),%ymm8        # 63c0 <_sk_callback_avx+0x17a>
   .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
@@ -14502,7 +14497,7 @@ HIDDEN _sk_overlay_avx
 .globl _sk_overlay_avx
 FUNCTION(_sk_overlay_avx)
 _sk_overlay_avx:
-  .byte  196,98,125,24,5,133,90,0,0          // vbroadcastss  0x5a85(%rip),%ymm8        # 63c8 <_sk_callback_avx+0x17e>
+  .byte  196,98,125,24,5,129,90,0,0          // vbroadcastss  0x5a81(%rip),%ymm8        # 63c4 <_sk_callback_avx+0x17e>
   .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
@@ -14568,10 +14563,10 @@ _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  196,98,125,24,29,124,89,0,0         // vbroadcastss  0x597c(%rip),%ymm11        # 63d0 <_sk_callback_avx+0x186>
+  .byte  196,98,125,24,29,120,89,0,0         // vbroadcastss  0x5978(%rip),%ymm11        # 63cc <_sk_callback_avx+0x186>
   .byte  196,65,28,88,235                    // vaddps        %ymm11,%ymm12,%ymm13
   .byte  196,65,20,89,192                    // vmulps        %ymm8,%ymm13,%ymm8
-  .byte  196,98,125,24,45,109,89,0,0         // vbroadcastss  0x596d(%rip),%ymm13        # 63d4 <_sk_callback_avx+0x18a>
+  .byte  196,98,125,24,45,105,89,0,0         // vbroadcastss  0x5969(%rip),%ymm13        # 63d0 <_sk_callback_avx+0x18a>
   .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
@@ -14582,7 +14577,7 @@ _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  196,98,125,24,5,43,89,0,0           // vbroadcastss  0x592b(%rip),%ymm8        # 63cc <_sk_callback_avx+0x182>
+  .byte  196,98,125,24,5,39,89,0,0           // vbroadcastss  0x5927(%rip),%ymm8        # 63c8 <_sk_callback_avx+0x182>
   .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
@@ -14709,12 +14704,12 @@ _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  196,98,125,24,53,250,86,0,0         // vbroadcastss  0x56fa(%rip),%ymm14        # 63d8 <_sk_callback_avx+0x18e>
+  .byte  196,98,125,24,53,246,86,0,0         // vbroadcastss  0x56f6(%rip),%ymm14        # 63d4 <_sk_callback_avx+0x18e>
   .byte  196,65,92,89,222                    // vmulps        %ymm14,%ymm4,%ymm11
-  .byte  196,98,125,24,61,240,86,0,0         // vbroadcastss  0x56f0(%rip),%ymm15        # 63dc <_sk_callback_avx+0x192>
+  .byte  196,98,125,24,61,236,86,0,0         // vbroadcastss  0x56ec(%rip),%ymm15        # 63d8 <_sk_callback_avx+0x192>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
-  .byte  196,226,125,24,5,225,86,0,0         // vbroadcastss  0x56e1(%rip),%ymm0        # 63e0 <_sk_callback_avx+0x196>
+  .byte  196,226,125,24,5,221,86,0,0         // vbroadcastss  0x56dd(%rip),%ymm0        # 63dc <_sk_callback_avx+0x196>
   .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
@@ -14775,7 +14770,7 @@ _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  196,98,125,24,5,186,85,0,0          // vbroadcastss  0x55ba(%rip),%ymm8        # 63e4 <_sk_callback_avx+0x19a>
+  .byte  196,98,125,24,5,182,85,0,0          // vbroadcastss  0x55b6(%rip),%ymm8        # 63e0 <_sk_callback_avx+0x19a>
   .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
@@ -14832,12 +14827,12 @@ _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  196,98,125,24,53,200,84,0,0         // vbroadcastss  0x54c8(%rip),%ymm14        # 63e8 <_sk_callback_avx+0x19e>
+  .byte  196,98,125,24,53,196,84,0,0         // vbroadcastss  0x54c4(%rip),%ymm14        # 63e4 <_sk_callback_avx+0x19e>
   .byte  196,65,92,89,222                    // vmulps        %ymm14,%ymm4,%ymm11
-  .byte  196,98,125,24,61,190,84,0,0         // vbroadcastss  0x54be(%rip),%ymm15        # 63ec <_sk_callback_avx+0x1a2>
+  .byte  196,98,125,24,61,186,84,0,0         // vbroadcastss  0x54ba(%rip),%ymm15        # 63e8 <_sk_callback_avx+0x1a2>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
-  .byte  196,226,125,24,5,175,84,0,0         // vbroadcastss  0x54af(%rip),%ymm0        # 63f0 <_sk_callback_avx+0x1a6>
+  .byte  196,226,125,24,5,171,84,0,0         // vbroadcastss  0x54ab(%rip),%ymm0        # 63ec <_sk_callback_avx+0x1a6>
   .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
@@ -14898,7 +14893,7 @@ _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  196,98,125,24,5,136,83,0,0          // vbroadcastss  0x5388(%rip),%ymm8        # 63f4 <_sk_callback_avx+0x1aa>
+  .byte  196,98,125,24,5,132,83,0,0          // vbroadcastss  0x5384(%rip),%ymm8        # 63f0 <_sk_callback_avx+0x1aa>
   .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
@@ -14927,12 +14922,12 @@ _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  196,98,125,24,45,30,83,0,0          // vbroadcastss  0x531e(%rip),%ymm13        # 63f8 <_sk_callback_avx+0x1ae>
+  .byte  196,98,125,24,45,26,83,0,0          // vbroadcastss  0x531a(%rip),%ymm13        # 63f4 <_sk_callback_avx+0x1ae>
   .byte  196,65,92,89,213                    // vmulps        %ymm13,%ymm4,%ymm10
-  .byte  196,98,125,24,53,20,83,0,0          // vbroadcastss  0x5314(%rip),%ymm14        # 63fc <_sk_callback_avx+0x1b2>
+  .byte  196,98,125,24,53,16,83,0,0          // vbroadcastss  0x5310(%rip),%ymm14        # 63f8 <_sk_callback_avx+0x1b2>
   .byte  196,65,84,89,222                    // vmulps        %ymm14,%ymm5,%ymm11
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,61,5,83,0,0           // vbroadcastss  0x5305(%rip),%ymm15        # 6400 <_sk_callback_avx+0x1b6>
+  .byte  196,98,125,24,61,1,83,0,0           // vbroadcastss  0x5301(%rip),%ymm15        # 63fc <_sk_callback_avx+0x1b6>
   .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
@@ -14995,7 +14990,7 @@ _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  196,226,125,24,5,204,81,0,0         // vbroadcastss  0x51cc(%rip),%ymm0        # 6404 <_sk_callback_avx+0x1ba>
+  .byte  196,226,125,24,5,200,81,0,0         // vbroadcastss  0x51c8(%rip),%ymm0        # 6400 <_sk_callback_avx+0x1ba>
   .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
@@ -15025,12 +15020,12 @@ _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  196,98,125,24,45,94,81,0,0          // vbroadcastss  0x515e(%rip),%ymm13        # 6408 <_sk_callback_avx+0x1be>
+  .byte  196,98,125,24,45,90,81,0,0          // vbroadcastss  0x515a(%rip),%ymm13        # 6404 <_sk_callback_avx+0x1be>
   .byte  196,65,108,89,213                   // vmulps        %ymm13,%ymm2,%ymm10
-  .byte  196,98,125,24,53,84,81,0,0          // vbroadcastss  0x5154(%rip),%ymm14        # 640c <_sk_callback_avx+0x1c2>
+  .byte  196,98,125,24,53,80,81,0,0          // vbroadcastss  0x5150(%rip),%ymm14        # 6408 <_sk_callback_avx+0x1c2>
   .byte  196,65,116,89,222                   // vmulps        %ymm14,%ymm1,%ymm11
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,61,69,81,0,0          // vbroadcastss  0x5145(%rip),%ymm15        # 6410 <_sk_callback_avx+0x1c6>
+  .byte  196,98,125,24,61,65,81,0,0          // vbroadcastss  0x5141(%rip),%ymm15        # 640c <_sk_callback_avx+0x1c6>
   .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
@@ -15093,7 +15088,7 @@ _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  196,226,125,24,5,12,80,0,0          // vbroadcastss  0x500c(%rip),%ymm0        # 6414 <_sk_callback_avx+0x1ca>
+  .byte  196,226,125,24,5,8,80,0,0           // vbroadcastss  0x5008(%rip),%ymm0        # 6410 <_sk_callback_avx+0x1ca>
   .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
@@ -15129,7 +15124,7 @@ HIDDEN _sk_clamp_1_avx
 .globl _sk_clamp_1_avx
 FUNCTION(_sk_clamp_1_avx)
 _sk_clamp_1_avx:
-  .byte  196,98,125,24,5,159,79,0,0          // vbroadcastss  0x4f9f(%rip),%ymm8        # 6418 <_sk_callback_avx+0x1ce>
+  .byte  196,98,125,24,5,155,79,0,0          // vbroadcastss  0x4f9b(%rip),%ymm8        # 6414 <_sk_callback_avx+0x1ce>
   .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
@@ -15141,7 +15136,7 @@ HIDDEN _sk_clamp_a_avx
 .globl _sk_clamp_a_avx
 FUNCTION(_sk_clamp_a_avx)
 _sk_clamp_a_avx:
-  .byte  196,98,125,24,5,130,79,0,0          // vbroadcastss  0x4f82(%rip),%ymm8        # 641c <_sk_callback_avx+0x1d2>
+  .byte  196,98,125,24,5,126,79,0,0          // vbroadcastss  0x4f7e(%rip),%ymm8        # 6418 <_sk_callback_avx+0x1d2>
   .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
@@ -15227,7 +15222,7 @@ 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  196,98,125,24,21,202,78,0,0         // vbroadcastss  0x4eca(%rip),%ymm10        # 6420 <_sk_callback_avx+0x1d6>
+  .byte  196,98,125,24,21,198,78,0,0         // vbroadcastss  0x4ec6(%rip),%ymm10        # 641c <_sk_callback_avx+0x1d6>
   .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
@@ -15240,17 +15235,17 @@ HIDDEN _sk_from_srgb_avx
 .globl _sk_from_srgb_avx
 FUNCTION(_sk_from_srgb_avx)
 _sk_from_srgb_avx:
-  .byte  196,98,125,24,5,171,78,0,0          // vbroadcastss  0x4eab(%rip),%ymm8        # 6424 <_sk_callback_avx+0x1da>
+  .byte  196,98,125,24,5,167,78,0,0          // vbroadcastss  0x4ea7(%rip),%ymm8        # 6420 <_sk_callback_avx+0x1da>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,157,78,0,0         // vbroadcastss  0x4e9d(%rip),%ymm11        # 6428 <_sk_callback_avx+0x1de>
+  .byte  196,98,125,24,29,153,78,0,0         // vbroadcastss  0x4e99(%rip),%ymm11        # 6424 <_sk_callback_avx+0x1de>
   .byte  196,65,124,89,227                   // vmulps        %ymm11,%ymm0,%ymm12
-  .byte  196,98,125,24,45,147,78,0,0         // vbroadcastss  0x4e93(%rip),%ymm13        # 642c <_sk_callback_avx+0x1e2>
+  .byte  196,98,125,24,45,143,78,0,0         // vbroadcastss  0x4e8f(%rip),%ymm13        # 6428 <_sk_callback_avx+0x1e2>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,44,89,212                    // vmulps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,132,78,0,0         // vbroadcastss  0x4e84(%rip),%ymm12        # 6430 <_sk_callback_avx+0x1e6>
+  .byte  196,98,125,24,37,128,78,0,0         // vbroadcastss  0x4e80(%rip),%ymm12        # 642c <_sk_callback_avx+0x1e6>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,53,122,78,0,0         // vbroadcastss  0x4e7a(%rip),%ymm14        # 6434 <_sk_callback_avx+0x1ea>
+  .byte  196,98,125,24,53,118,78,0,0         // vbroadcastss  0x4e76(%rip),%ymm14        # 6430 <_sk_callback_avx+0x1ea>
   .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
@@ -15279,18 +15274,18 @@ _sk_to_srgb_avx:
   .byte  197,124,82,192                      // vrsqrtps      %ymm0,%ymm8
   .byte  196,65,124,83,200                   // vrcpps        %ymm8,%ymm9
   .byte  196,65,124,82,208                   // vrsqrtps      %ymm8,%ymm10
-  .byte  196,98,125,24,5,5,78,0,0            // vbroadcastss  0x4e05(%rip),%ymm8        # 6438 <_sk_callback_avx+0x1ee>
+  .byte  196,98,125,24,5,1,78,0,0            // vbroadcastss  0x4e01(%rip),%ymm8        # 6434 <_sk_callback_avx+0x1ee>
   .byte  196,65,124,89,216                   // vmulps        %ymm8,%ymm0,%ymm11
-  .byte  196,98,125,24,37,251,77,0,0         // vbroadcastss  0x4dfb(%rip),%ymm12        # 643c <_sk_callback_avx+0x1f2>
+  .byte  196,98,125,24,37,247,77,0,0         // vbroadcastss  0x4df7(%rip),%ymm12        # 6438 <_sk_callback_avx+0x1f2>
   .byte  196,65,52,89,204                    // vmulps        %ymm12,%ymm9,%ymm9
-  .byte  196,98,125,24,45,241,77,0,0         // vbroadcastss  0x4df1(%rip),%ymm13        # 6440 <_sk_callback_avx+0x1f6>
+  .byte  196,98,125,24,45,237,77,0,0         // vbroadcastss  0x4ded(%rip),%ymm13        # 643c <_sk_callback_avx+0x1f6>
   .byte  196,65,52,88,205                    // vaddps        %ymm13,%ymm9,%ymm9
-  .byte  196,98,125,24,53,231,77,0,0         // vbroadcastss  0x4de7(%rip),%ymm14        # 6444 <_sk_callback_avx+0x1fa>
+  .byte  196,98,125,24,53,227,77,0,0         // vbroadcastss  0x4de3(%rip),%ymm14        # 6440 <_sk_callback_avx+0x1fa>
   .byte  196,65,44,89,214                    // vmulps        %ymm14,%ymm10,%ymm10
   .byte  196,65,44,88,201                    // vaddps        %ymm9,%ymm10,%ymm9
-  .byte  196,98,125,24,21,216,77,0,0         // vbroadcastss  0x4dd8(%rip),%ymm10        # 6448 <_sk_callback_avx+0x1fe>
+  .byte  196,98,125,24,21,212,77,0,0         // vbroadcastss  0x4dd4(%rip),%ymm10        # 6444 <_sk_callback_avx+0x1fe>
   .byte  196,65,44,93,201                    // vminps        %ymm9,%ymm10,%ymm9
-  .byte  196,98,125,24,61,206,77,0,0         // vbroadcastss  0x4dce(%rip),%ymm15        # 644c <_sk_callback_avx+0x202>
+  .byte  196,98,125,24,61,202,77,0,0         // vbroadcastss  0x4dca(%rip),%ymm15        # 6448 <_sk_callback_avx+0x202>
   .byte  196,193,124,194,199,1               // vcmpltps      %ymm15,%ymm0,%ymm0
   .byte  196,195,53,74,195,0                 // vblendvps     %ymm0,%ymm11,%ymm9,%ymm0
   .byte  197,124,82,201                      // vrsqrtps      %ymm1,%ymm9
@@ -15327,7 +15322,7 @@ _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  196,98,125,24,29,52,77,0,0          // vbroadcastss  0x4d34(%rip),%ymm11        # 6450 <_sk_callback_avx+0x206>
+  .byte  196,98,125,24,29,48,77,0,0          // vbroadcastss  0x4d30(%rip),%ymm11        # 644c <_sk_callback_avx+0x206>
   .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
@@ -15337,19 +15332,19 @@ _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  196,98,125,24,29,13,77,0,0          // vbroadcastss  0x4d0d(%rip),%ymm11        # 645c <_sk_callback_avx+0x212>
+  .byte  196,98,125,24,29,9,77,0,0           // vbroadcastss  0x4d09(%rip),%ymm11        # 6458 <_sk_callback_avx+0x212>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,251,76,0,0         // vbroadcastss  0x4cfb(%rip),%ymm11        # 6458 <_sk_callback_avx+0x20e>
+  .byte  196,98,125,24,29,247,76,0,0         // vbroadcastss  0x4cf7(%rip),%ymm11        # 6454 <_sk_callback_avx+0x20e>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
   .byte  196,227,117,74,202,224              // vblendvps     %ymm14,%ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,227,76,0,0        // vbroadcastss  0x4ce3(%rip),%ymm2        # 6454 <_sk_callback_avx+0x20a>
+  .byte  196,226,125,24,21,223,76,0,0        // vbroadcastss  0x4cdf(%rip),%ymm2        # 6450 <_sk_callback_avx+0x20a>
   .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  196,98,125,24,37,202,76,0,0         // vbroadcastss  0x4cca(%rip),%ymm12        # 6464 <_sk_callback_avx+0x21a>
+  .byte  196,98,125,24,37,198,76,0,0         // vbroadcastss  0x4cc6(%rip),%ymm12        # 6460 <_sk_callback_avx+0x21a>
   .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
@@ -15359,7 +15354,7 @@ _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  196,98,125,24,5,141,76,0,0          // vbroadcastss  0x4c8d(%rip),%ymm8        # 6460 <_sk_callback_avx+0x216>
+  .byte  196,98,125,24,5,137,76,0,0          // vbroadcastss  0x4c89(%rip),%ymm8        # 645c <_sk_callback_avx+0x216>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15376,7 +15371,7 @@ _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  196,98,125,24,5,90,76,0,0           // vbroadcastss  0x4c5a(%rip),%ymm8        # 6468 <_sk_callback_avx+0x21e>
+  .byte  196,98,125,24,5,86,76,0,0           // vbroadcastss  0x4c56(%rip),%ymm8        # 6464 <_sk_callback_avx+0x21e>
   .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
@@ -15384,23 +15379,23 @@ _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  196,98,125,24,29,52,76,0,0          // vbroadcastss  0x4c34(%rip),%ymm11        # 646c <_sk_callback_avx+0x222>
+  .byte  196,98,125,24,29,48,76,0,0          // vbroadcastss  0x4c30(%rip),%ymm11        # 6468 <_sk_callback_avx+0x222>
   .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  196,98,125,24,37,30,76,0,0          // vbroadcastss  0x4c1e(%rip),%ymm12        # 6474 <_sk_callback_avx+0x22a>
+  .byte  196,98,125,24,37,26,76,0,0          // vbroadcastss  0x4c1a(%rip),%ymm12        # 6470 <_sk_callback_avx+0x22a>
   .byte  196,193,4,89,196                    // vmulps        %ymm12,%ymm15,%ymm0
-  .byte  196,98,125,24,45,20,76,0,0          // vbroadcastss  0x4c14(%rip),%ymm13        # 6478 <_sk_callback_avx+0x22e>
+  .byte  196,98,125,24,45,16,76,0,0          // vbroadcastss  0x4c10(%rip),%ymm13        # 6474 <_sk_callback_avx+0x22e>
   .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  196,226,125,24,13,245,75,0,0        // vbroadcastss  0x4bf5(%rip),%ymm1        # 6470 <_sk_callback_avx+0x226>
+  .byte  196,226,125,24,13,241,75,0,0        // vbroadcastss  0x4bf1(%rip),%ymm1        # 646c <_sk_callback_avx+0x226>
   .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  196,98,125,24,53,224,75,0,0         // vbroadcastss  0x4be0(%rip),%ymm14        # 647c <_sk_callback_avx+0x232>
+  .byte  196,98,125,24,53,220,75,0,0         // vbroadcastss  0x4bdc(%rip),%ymm14        # 6478 <_sk_callback_avx+0x232>
   .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
@@ -15419,7 +15414,7 @@ _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  196,226,125,24,5,130,75,0,0         // vbroadcastss  0x4b82(%rip),%ymm0        # 6480 <_sk_callback_avx+0x236>
+  .byte  196,226,125,24,5,126,75,0,0         // vbroadcastss  0x4b7e(%rip),%ymm0        # 647c <_sk_callback_avx+0x236>
   .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
@@ -15478,7 +15473,7 @@ _sk_scale_u8_avx:
   .byte  196,66,121,49,192                   // vpmovzxbd     %xmm8,%xmm8
   .byte  196,67,53,24,192,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,171,74,0,0         // vbroadcastss  0x4aab(%rip),%ymm9        # 6484 <_sk_callback_avx+0x23a>
+  .byte  196,98,125,24,13,167,74,0,0         // vbroadcastss  0x4aa7(%rip),%ymm9        # 6480 <_sk_callback_avx+0x23a>
   .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
@@ -15537,7 +15532,7 @@ _sk_lerp_u8_avx:
   .byte  196,66,121,49,192                   // vpmovzxbd     %xmm8,%xmm8
   .byte  196,67,53,24,192,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,247,73,0,0         // vbroadcastss  0x49f7(%rip),%ymm9        # 6488 <_sk_callback_avx+0x23e>
+  .byte  196,98,125,24,13,243,73,0,0         // vbroadcastss  0x49f3(%rip),%ymm9        # 6484 <_sk_callback_avx+0x23e>
   .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
@@ -15580,20 +15575,20 @@ _sk_lerp_565_avx:
   .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  196,98,125,24,13,97,73,0,0          // vbroadcastss  0x4961(%rip),%ymm9        # 648c <_sk_callback_avx+0x242>
+  .byte  196,98,125,24,13,93,73,0,0          // vbroadcastss  0x495d(%rip),%ymm9        # 6488 <_sk_callback_avx+0x242>
   .byte  196,65,60,84,201                    // vandps        %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,82,73,0,0          // vbroadcastss  0x4952(%rip),%ymm10        # 6490 <_sk_callback_avx+0x246>
+  .byte  196,98,125,24,21,78,73,0,0          // vbroadcastss  0x494e(%rip),%ymm10        # 648c <_sk_callback_avx+0x246>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,21,72,73,0,0          // vbroadcastss  0x4948(%rip),%ymm10        # 6494 <_sk_callback_avx+0x24a>
+  .byte  196,98,125,24,21,68,73,0,0          // vbroadcastss  0x4944(%rip),%ymm10        # 6490 <_sk_callback_avx+0x24a>
   .byte  196,65,60,84,210                    // vandps        %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,57,73,0,0          // vbroadcastss  0x4939(%rip),%ymm11        # 6498 <_sk_callback_avx+0x24e>
+  .byte  196,98,125,24,29,53,73,0,0          // vbroadcastss  0x4935(%rip),%ymm11        # 6494 <_sk_callback_avx+0x24e>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,29,47,73,0,0          // vbroadcastss  0x492f(%rip),%ymm11        # 649c <_sk_callback_avx+0x252>
+  .byte  196,98,125,24,29,43,73,0,0          // vbroadcastss  0x492b(%rip),%ymm11        # 6498 <_sk_callback_avx+0x252>
   .byte  196,65,60,84,195                    // vandps        %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,32,73,0,0          // vbroadcastss  0x4920(%rip),%ymm11        # 64a0 <_sk_callback_avx+0x256>
+  .byte  196,98,125,24,29,28,73,0,0          // vbroadcastss  0x491c(%rip),%ymm11        # 649c <_sk_callback_avx+0x256>
   .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
@@ -15640,7 +15635,7 @@ _sk_lerp_565_avx:
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,225                 // jmpq          ffffffffe2001c50 <_sk_callback_avx+0xffffffffe1ffba06>
+  .byte  233,255,255,255,225                 // jmpq          ffffffffe2001c50 <_sk_callback_avx+0xffffffffe1ffba0a>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
@@ -15673,7 +15668,7 @@ _sk_load_tables_avx:
   .byte  65,85                               // push          %r13
   .byte  65,84                               // push          %r12
   .byte  83                                  // push          %rbx
-  .byte  197,124,40,13,246,74,0,0            // vmovaps       0x4af6(%rip),%ymm9        # 6780 <_sk_callback_avx+0x536>
+  .byte  197,124,40,13,246,74,0,0            // vmovaps       0x4af6(%rip),%ymm9        # 6780 <_sk_callback_avx+0x53a>
   .byte  196,193,60,84,193                   // vandps        %ymm9,%ymm8,%ymm0
   .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
   .byte  69,137,203                          // mov           %r9d,%r11d
@@ -15765,7 +15760,7 @@ _sk_load_tables_avx:
   .byte  196,193,97,114,210,24               // vpsrld        $0x18,%xmm10,%xmm3
   .byte  196,227,61,24,219,1                 // vinsertf128   $0x1,%xmm3,%ymm8,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,43,70,0,0           // vbroadcastss  0x462b(%rip),%ymm8        # 64a4 <_sk_callback_avx+0x25a>
+  .byte  196,98,125,24,5,39,70,0,0           // vbroadcastss  0x4627(%rip),%ymm8        # 64a0 <_sk_callback_avx+0x25a>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -15857,7 +15852,7 @@ _sk_load_tables_u16_be_avx:
   .byte  197,177,108,208                     // vpunpcklqdq   %xmm0,%xmm9,%xmm2
   .byte  197,177,109,200                     // vpunpckhqdq   %xmm0,%xmm9,%xmm1
   .byte  196,65,57,108,212                   // vpunpcklqdq   %xmm12,%xmm8,%xmm10
-  .byte  197,121,111,29,54,72,0,0            // vmovdqa       0x4836(%rip),%xmm11        # 6800 <_sk_callback_avx+0x5b6>
+  .byte  197,121,111,29,54,72,0,0            // vmovdqa       0x4836(%rip),%xmm11        # 6800 <_sk_callback_avx+0x5ba>
   .byte  196,193,105,219,195                 // vpand         %xmm11,%xmm2,%xmm0
   .byte  196,65,49,239,201                   // vpxor         %xmm9,%xmm9,%xmm9
   .byte  196,193,121,105,209                 // vpunpckhwd    %xmm9,%xmm0,%xmm2
@@ -15956,7 +15951,7 @@ _sk_load_tables_u16_be_avx:
   .byte  196,226,121,51,219                  // vpmovzxwd     %xmm3,%xmm3
   .byte  196,195,101,24,216,1                // vinsertf128   $0x1,%xmm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,220,66,0,0          // vbroadcastss  0x42dc(%rip),%ymm8        # 64a8 <_sk_callback_avx+0x25e>
+  .byte  196,98,125,24,5,216,66,0,0          // vbroadcastss  0x42d8(%rip),%ymm8        # 64a4 <_sk_callback_avx+0x25e>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -16028,7 +16023,7 @@ _sk_load_tables_rgb_u16_be_avx:
   .byte  197,185,108,202                     // vpunpcklqdq   %xmm2,%xmm8,%xmm1
   .byte  197,185,109,210                     // vpunpckhqdq   %xmm2,%xmm8,%xmm2
   .byte  197,121,108,195                     // vpunpcklqdq   %xmm3,%xmm0,%xmm8
-  .byte  197,121,111,13,47,69,0,0            // vmovdqa       0x452f(%rip),%xmm9        # 6810 <_sk_callback_avx+0x5c6>
+  .byte  197,121,111,13,47,69,0,0            // vmovdqa       0x452f(%rip),%xmm9        # 6810 <_sk_callback_avx+0x5ca>
   .byte  196,193,113,219,193                 // vpand         %xmm9,%xmm1,%xmm0
   .byte  196,65,41,239,210                   // vpxor         %xmm10,%xmm10,%xmm10
   .byte  196,193,121,105,202                 // vpunpckhwd    %xmm10,%xmm0,%xmm1
@@ -16120,7 +16115,7 @@ _sk_load_tables_rgb_u16_be_avx:
   .byte  196,227,105,33,211,48               // vinsertps     $0x30,%xmm3,%xmm2,%xmm2
   .byte  196,195,109,24,208,1                // vinsertf128   $0x1,%xmm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,238,63,0,0        // vbroadcastss  0x3fee(%rip),%ymm3        # 64ac <_sk_callback_avx+0x262>
+  .byte  196,226,125,24,29,234,63,0,0        // vbroadcastss  0x3fea(%rip),%ymm3        # 64a8 <_sk_callback_avx+0x262>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,93                               // pop           %r13
@@ -16173,7 +16168,7 @@ _sk_byte_tables_avx:
   .byte  65,84                               // push          %r12
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,34,63,0,0           // vbroadcastss  0x3f22(%rip),%ymm8        # 64b0 <_sk_callback_avx+0x266>
+  .byte  196,98,125,24,5,30,63,0,0           // vbroadcastss  0x3f1e(%rip),%ymm8        # 64ac <_sk_callback_avx+0x266>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,253,91,192                      // vcvtps2dq     %ymm0,%ymm0
   .byte  196,195,249,22,192,1                // vpextrq       $0x1,%xmm0,%r8
@@ -16210,7 +16205,7 @@ _sk_byte_tables_avx:
   .byte  196,226,121,49,192                  // vpmovzxbd     %xmm0,%xmm0
   .byte  196,227,53,24,192,1                 // vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,112,62,0,0         // vbroadcastss  0x3e70(%rip),%ymm9        # 64b4 <_sk_callback_avx+0x26a>
+  .byte  196,98,125,24,13,108,62,0,0         // vbroadcastss  0x3e6c(%rip),%ymm9        # 64b0 <_sk_callback_avx+0x26a>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
@@ -16372,7 +16367,7 @@ _sk_byte_tables_rgb_avx:
   .byte  196,226,121,49,192                  // vpmovzxbd     %xmm0,%xmm0
   .byte  196,227,53,24,192,1                 // vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,150,59,0,0         // vbroadcastss  0x3b96(%rip),%ymm9        # 64b8 <_sk_callback_avx+0x26e>
+  .byte  196,98,125,24,13,146,59,0,0         // vbroadcastss  0x3b92(%rip),%ymm9        # 64b4 <_sk_callback_avx+0x26e>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
@@ -16669,36 +16664,36 @@ _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  196,98,125,24,37,244,54,0,0         // vbroadcastss  0x36f4(%rip),%ymm12        # 64bc <_sk_callback_avx+0x272>
+  .byte  196,98,125,24,37,240,54,0,0         // vbroadcastss  0x36f0(%rip),%ymm12        # 64b8 <_sk_callback_avx+0x272>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,234,54,0,0         // vbroadcastss  0x36ea(%rip),%ymm12        # 64c0 <_sk_callback_avx+0x276>
+  .byte  196,98,125,24,37,230,54,0,0         // vbroadcastss  0x36e6(%rip),%ymm12        # 64bc <_sk_callback_avx+0x276>
   .byte  196,193,124,84,196                  // vandps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,224,54,0,0         // vbroadcastss  0x36e0(%rip),%ymm12        # 64c4 <_sk_callback_avx+0x27a>
+  .byte  196,98,125,24,37,220,54,0,0         // vbroadcastss  0x36dc(%rip),%ymm12        # 64c0 <_sk_callback_avx+0x27a>
   .byte  196,193,124,86,196                  // vorps         %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,214,54,0,0         // vbroadcastss  0x36d6(%rip),%ymm12        # 64c8 <_sk_callback_avx+0x27e>
+  .byte  196,98,125,24,37,210,54,0,0         // vbroadcastss  0x36d2(%rip),%ymm12        # 64c4 <_sk_callback_avx+0x27e>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,204,54,0,0         // vbroadcastss  0x36cc(%rip),%ymm12        # 64cc <_sk_callback_avx+0x282>
+  .byte  196,98,125,24,37,200,54,0,0         // vbroadcastss  0x36c8(%rip),%ymm12        # 64c8 <_sk_callback_avx+0x282>
   .byte  196,65,124,89,228                   // vmulps        %ymm12,%ymm0,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,189,54,0,0         // vbroadcastss  0x36bd(%rip),%ymm12        # 64d0 <_sk_callback_avx+0x286>
+  .byte  196,98,125,24,37,185,54,0,0         // vbroadcastss  0x36b9(%rip),%ymm12        # 64cc <_sk_callback_avx+0x286>
   .byte  196,193,124,88,196                  // vaddps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,179,54,0,0         // vbroadcastss  0x36b3(%rip),%ymm12        # 64d4 <_sk_callback_avx+0x28a>
+  .byte  196,98,125,24,37,175,54,0,0         // vbroadcastss  0x36af(%rip),%ymm12        # 64d0 <_sk_callback_avx+0x28a>
   .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  196,98,125,24,29,151,54,0,0         // vbroadcastss  0x3697(%rip),%ymm11        # 64d8 <_sk_callback_avx+0x28e>
+  .byte  196,98,125,24,29,147,54,0,0         // vbroadcastss  0x3693(%rip),%ymm11        # 64d4 <_sk_callback_avx+0x28e>
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,141,54,0,0         // vbroadcastss  0x368d(%rip),%ymm11        # 64dc <_sk_callback_avx+0x292>
+  .byte  196,98,125,24,29,137,54,0,0         // vbroadcastss  0x3689(%rip),%ymm11        # 64d8 <_sk_callback_avx+0x292>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,124,92,195                  // vsubps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,126,54,0,0         // vbroadcastss  0x367e(%rip),%ymm11        # 64e0 <_sk_callback_avx+0x296>
+  .byte  196,98,125,24,29,122,54,0,0         // vbroadcastss  0x367a(%rip),%ymm11        # 64dc <_sk_callback_avx+0x296>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,116,54,0,0         // vbroadcastss  0x3674(%rip),%ymm11        # 64e4 <_sk_callback_avx+0x29a>
+  .byte  196,98,125,24,29,112,54,0,0         // vbroadcastss  0x3670(%rip),%ymm11        # 64e0 <_sk_callback_avx+0x29a>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,124,88,194                  // vaddps        %ymm10,%ymm0,%ymm0
-  .byte  196,98,125,24,21,101,54,0,0         // vbroadcastss  0x3665(%rip),%ymm10        # 64e8 <_sk_callback_avx+0x29e>
+  .byte  196,98,125,24,21,97,54,0,0          // vbroadcastss  0x3661(%rip),%ymm10        # 64e4 <_sk_callback_avx+0x29e>
   .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
@@ -16706,7 +16701,7 @@ _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  196,98,125,24,5,60,54,0,0           // vbroadcastss  0x363c(%rip),%ymm8        # 64ec <_sk_callback_avx+0x2a2>
+  .byte  196,98,125,24,5,56,54,0,0           // vbroadcastss  0x3638(%rip),%ymm8        # 64e8 <_sk_callback_avx+0x2a2>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16728,36 +16723,36 @@ _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  196,98,125,24,37,237,53,0,0         // vbroadcastss  0x35ed(%rip),%ymm12        # 64f0 <_sk_callback_avx+0x2a6>
+  .byte  196,98,125,24,37,233,53,0,0         // vbroadcastss  0x35e9(%rip),%ymm12        # 64ec <_sk_callback_avx+0x2a6>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,227,53,0,0         // vbroadcastss  0x35e3(%rip),%ymm12        # 64f4 <_sk_callback_avx+0x2aa>
+  .byte  196,98,125,24,37,223,53,0,0         // vbroadcastss  0x35df(%rip),%ymm12        # 64f0 <_sk_callback_avx+0x2aa>
   .byte  196,193,116,84,204                  // vandps        %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,217,53,0,0         // vbroadcastss  0x35d9(%rip),%ymm12        # 64f8 <_sk_callback_avx+0x2ae>
+  .byte  196,98,125,24,37,213,53,0,0         // vbroadcastss  0x35d5(%rip),%ymm12        # 64f4 <_sk_callback_avx+0x2ae>
   .byte  196,193,116,86,204                  // vorps         %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,207,53,0,0         // vbroadcastss  0x35cf(%rip),%ymm12        # 64fc <_sk_callback_avx+0x2b2>
+  .byte  196,98,125,24,37,203,53,0,0         // vbroadcastss  0x35cb(%rip),%ymm12        # 64f8 <_sk_callback_avx+0x2b2>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,197,53,0,0         // vbroadcastss  0x35c5(%rip),%ymm12        # 6500 <_sk_callback_avx+0x2b6>
+  .byte  196,98,125,24,37,193,53,0,0         // vbroadcastss  0x35c1(%rip),%ymm12        # 64fc <_sk_callback_avx+0x2b6>
   .byte  196,65,116,89,228                   // vmulps        %ymm12,%ymm1,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,182,53,0,0         // vbroadcastss  0x35b6(%rip),%ymm12        # 6504 <_sk_callback_avx+0x2ba>
+  .byte  196,98,125,24,37,178,53,0,0         // vbroadcastss  0x35b2(%rip),%ymm12        # 6500 <_sk_callback_avx+0x2ba>
   .byte  196,193,116,88,204                  // vaddps        %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,172,53,0,0         // vbroadcastss  0x35ac(%rip),%ymm12        # 6508 <_sk_callback_avx+0x2be>
+  .byte  196,98,125,24,37,168,53,0,0         // vbroadcastss  0x35a8(%rip),%ymm12        # 6504 <_sk_callback_avx+0x2be>
   .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  196,98,125,24,29,144,53,0,0         // vbroadcastss  0x3590(%rip),%ymm11        # 650c <_sk_callback_avx+0x2c2>
+  .byte  196,98,125,24,29,140,53,0,0         // vbroadcastss  0x358c(%rip),%ymm11        # 6508 <_sk_callback_avx+0x2c2>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,134,53,0,0         // vbroadcastss  0x3586(%rip),%ymm11        # 6510 <_sk_callback_avx+0x2c6>
+  .byte  196,98,125,24,29,130,53,0,0         // vbroadcastss  0x3582(%rip),%ymm11        # 650c <_sk_callback_avx+0x2c6>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,116,92,203                  // vsubps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,119,53,0,0         // vbroadcastss  0x3577(%rip),%ymm11        # 6514 <_sk_callback_avx+0x2ca>
+  .byte  196,98,125,24,29,115,53,0,0         // vbroadcastss  0x3573(%rip),%ymm11        # 6510 <_sk_callback_avx+0x2ca>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,109,53,0,0         // vbroadcastss  0x356d(%rip),%ymm11        # 6518 <_sk_callback_avx+0x2ce>
+  .byte  196,98,125,24,29,105,53,0,0         // vbroadcastss  0x3569(%rip),%ymm11        # 6514 <_sk_callback_avx+0x2ce>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,116,88,202                  // vaddps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,21,94,53,0,0          // vbroadcastss  0x355e(%rip),%ymm10        # 651c <_sk_callback_avx+0x2d2>
+  .byte  196,98,125,24,21,90,53,0,0          // vbroadcastss  0x355a(%rip),%ymm10        # 6518 <_sk_callback_avx+0x2d2>
   .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
@@ -16765,7 +16760,7 @@ _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  196,98,125,24,5,53,53,0,0           // vbroadcastss  0x3535(%rip),%ymm8        # 6520 <_sk_callback_avx+0x2d6>
+  .byte  196,98,125,24,5,49,53,0,0           // vbroadcastss  0x3531(%rip),%ymm8        # 651c <_sk_callback_avx+0x2d6>
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16787,36 +16782,36 @@ _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  196,98,125,24,37,230,52,0,0         // vbroadcastss  0x34e6(%rip),%ymm12        # 6524 <_sk_callback_avx+0x2da>
+  .byte  196,98,125,24,37,226,52,0,0         // vbroadcastss  0x34e2(%rip),%ymm12        # 6520 <_sk_callback_avx+0x2da>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,220,52,0,0         // vbroadcastss  0x34dc(%rip),%ymm12        # 6528 <_sk_callback_avx+0x2de>
+  .byte  196,98,125,24,37,216,52,0,0         // vbroadcastss  0x34d8(%rip),%ymm12        # 6524 <_sk_callback_avx+0x2de>
   .byte  196,193,108,84,212                  // vandps        %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,210,52,0,0         // vbroadcastss  0x34d2(%rip),%ymm12        # 652c <_sk_callback_avx+0x2e2>
+  .byte  196,98,125,24,37,206,52,0,0         // vbroadcastss  0x34ce(%rip),%ymm12        # 6528 <_sk_callback_avx+0x2e2>
   .byte  196,193,108,86,212                  // vorps         %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,200,52,0,0         // vbroadcastss  0x34c8(%rip),%ymm12        # 6530 <_sk_callback_avx+0x2e6>
+  .byte  196,98,125,24,37,196,52,0,0         // vbroadcastss  0x34c4(%rip),%ymm12        # 652c <_sk_callback_avx+0x2e6>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,190,52,0,0         // vbroadcastss  0x34be(%rip),%ymm12        # 6534 <_sk_callback_avx+0x2ea>
+  .byte  196,98,125,24,37,186,52,0,0         // vbroadcastss  0x34ba(%rip),%ymm12        # 6530 <_sk_callback_avx+0x2ea>
   .byte  196,65,108,89,228                   // vmulps        %ymm12,%ymm2,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,175,52,0,0         // vbroadcastss  0x34af(%rip),%ymm12        # 6538 <_sk_callback_avx+0x2ee>
+  .byte  196,98,125,24,37,171,52,0,0         // vbroadcastss  0x34ab(%rip),%ymm12        # 6534 <_sk_callback_avx+0x2ee>
   .byte  196,193,108,88,212                  // vaddps        %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,165,52,0,0         // vbroadcastss  0x34a5(%rip),%ymm12        # 653c <_sk_callback_avx+0x2f2>
+  .byte  196,98,125,24,37,161,52,0,0         // vbroadcastss  0x34a1(%rip),%ymm12        # 6538 <_sk_callback_avx+0x2f2>
   .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  196,98,125,24,29,137,52,0,0         // vbroadcastss  0x3489(%rip),%ymm11        # 6540 <_sk_callback_avx+0x2f6>
+  .byte  196,98,125,24,29,133,52,0,0         // vbroadcastss  0x3485(%rip),%ymm11        # 653c <_sk_callback_avx+0x2f6>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,127,52,0,0         // vbroadcastss  0x347f(%rip),%ymm11        # 6544 <_sk_callback_avx+0x2fa>
+  .byte  196,98,125,24,29,123,52,0,0         // vbroadcastss  0x347b(%rip),%ymm11        # 6540 <_sk_callback_avx+0x2fa>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,108,92,211                  // vsubps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,112,52,0,0         // vbroadcastss  0x3470(%rip),%ymm11        # 6548 <_sk_callback_avx+0x2fe>
+  .byte  196,98,125,24,29,108,52,0,0         // vbroadcastss  0x346c(%rip),%ymm11        # 6544 <_sk_callback_avx+0x2fe>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,102,52,0,0         // vbroadcastss  0x3466(%rip),%ymm11        # 654c <_sk_callback_avx+0x302>
+  .byte  196,98,125,24,29,98,52,0,0          // vbroadcastss  0x3462(%rip),%ymm11        # 6548 <_sk_callback_avx+0x302>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,108,88,210                  // vaddps        %ymm10,%ymm2,%ymm2
-  .byte  196,98,125,24,21,87,52,0,0          // vbroadcastss  0x3457(%rip),%ymm10        # 6550 <_sk_callback_avx+0x306>
+  .byte  196,98,125,24,21,83,52,0,0          // vbroadcastss  0x3453(%rip),%ymm10        # 654c <_sk_callback_avx+0x306>
   .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
@@ -16824,7 +16819,7 @@ _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  196,98,125,24,5,46,52,0,0           // vbroadcastss  0x342e(%rip),%ymm8        # 6554 <_sk_callback_avx+0x30a>
+  .byte  196,98,125,24,5,42,52,0,0           // vbroadcastss  0x342a(%rip),%ymm8        # 6550 <_sk_callback_avx+0x30a>
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16846,36 +16841,36 @@ _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  196,98,125,24,37,223,51,0,0         // vbroadcastss  0x33df(%rip),%ymm12        # 6558 <_sk_callback_avx+0x30e>
+  .byte  196,98,125,24,37,219,51,0,0         // vbroadcastss  0x33db(%rip),%ymm12        # 6554 <_sk_callback_avx+0x30e>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,213,51,0,0         // vbroadcastss  0x33d5(%rip),%ymm12        # 655c <_sk_callback_avx+0x312>
+  .byte  196,98,125,24,37,209,51,0,0         // vbroadcastss  0x33d1(%rip),%ymm12        # 6558 <_sk_callback_avx+0x312>
   .byte  196,193,100,84,220                  // vandps        %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,203,51,0,0         // vbroadcastss  0x33cb(%rip),%ymm12        # 6560 <_sk_callback_avx+0x316>
+  .byte  196,98,125,24,37,199,51,0,0         // vbroadcastss  0x33c7(%rip),%ymm12        # 655c <_sk_callback_avx+0x316>
   .byte  196,193,100,86,220                  // vorps         %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,193,51,0,0         // vbroadcastss  0x33c1(%rip),%ymm12        # 6564 <_sk_callback_avx+0x31a>
+  .byte  196,98,125,24,37,189,51,0,0         // vbroadcastss  0x33bd(%rip),%ymm12        # 6560 <_sk_callback_avx+0x31a>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,183,51,0,0         // vbroadcastss  0x33b7(%rip),%ymm12        # 6568 <_sk_callback_avx+0x31e>
+  .byte  196,98,125,24,37,179,51,0,0         // vbroadcastss  0x33b3(%rip),%ymm12        # 6564 <_sk_callback_avx+0x31e>
   .byte  196,65,100,89,228                   // vmulps        %ymm12,%ymm3,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,168,51,0,0         // vbroadcastss  0x33a8(%rip),%ymm12        # 656c <_sk_callback_avx+0x322>
+  .byte  196,98,125,24,37,164,51,0,0         // vbroadcastss  0x33a4(%rip),%ymm12        # 6568 <_sk_callback_avx+0x322>
   .byte  196,193,100,88,220                  // vaddps        %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,158,51,0,0         // vbroadcastss  0x339e(%rip),%ymm12        # 6570 <_sk_callback_avx+0x326>
+  .byte  196,98,125,24,37,154,51,0,0         // vbroadcastss  0x339a(%rip),%ymm12        # 656c <_sk_callback_avx+0x326>
   .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  196,98,125,24,29,130,51,0,0         // vbroadcastss  0x3382(%rip),%ymm11        # 6574 <_sk_callback_avx+0x32a>
+  .byte  196,98,125,24,29,126,51,0,0         // vbroadcastss  0x337e(%rip),%ymm11        # 6570 <_sk_callback_avx+0x32a>
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,120,51,0,0         // vbroadcastss  0x3378(%rip),%ymm11        # 6578 <_sk_callback_avx+0x32e>
+  .byte  196,98,125,24,29,116,51,0,0         // vbroadcastss  0x3374(%rip),%ymm11        # 6574 <_sk_callback_avx+0x32e>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,100,92,219                  // vsubps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,105,51,0,0         // vbroadcastss  0x3369(%rip),%ymm11        # 657c <_sk_callback_avx+0x332>
+  .byte  196,98,125,24,29,101,51,0,0         // vbroadcastss  0x3365(%rip),%ymm11        # 6578 <_sk_callback_avx+0x332>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,95,51,0,0          // vbroadcastss  0x335f(%rip),%ymm11        # 6580 <_sk_callback_avx+0x336>
+  .byte  196,98,125,24,29,91,51,0,0          // vbroadcastss  0x335b(%rip),%ymm11        # 657c <_sk_callback_avx+0x336>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,100,88,218                  // vaddps        %ymm10,%ymm3,%ymm3
-  .byte  196,98,125,24,21,80,51,0,0          // vbroadcastss  0x3350(%rip),%ymm10        # 6584 <_sk_callback_avx+0x33a>
+  .byte  196,98,125,24,21,76,51,0,0          // vbroadcastss  0x334c(%rip),%ymm10        # 6580 <_sk_callback_avx+0x33a>
   .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
@@ -16883,7 +16878,7 @@ _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  196,98,125,24,5,39,51,0,0           // vbroadcastss  0x3327(%rip),%ymm8        # 6588 <_sk_callback_avx+0x33e>
+  .byte  196,98,125,24,5,35,51,0,0           // vbroadcastss  0x3323(%rip),%ymm8        # 6584 <_sk_callback_avx+0x33e>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16892,31 +16887,31 @@ HIDDEN _sk_lab_to_xyz_avx
 .globl _sk_lab_to_xyz_avx
 FUNCTION(_sk_lab_to_xyz_avx)
 _sk_lab_to_xyz_avx:
-  .byte  196,98,125,24,5,25,51,0,0           // vbroadcastss  0x3319(%rip),%ymm8        # 658c <_sk_callback_avx+0x342>
+  .byte  196,98,125,24,5,21,51,0,0           // vbroadcastss  0x3315(%rip),%ymm8        # 6588 <_sk_callback_avx+0x342>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,15,51,0,0           // vbroadcastss  0x330f(%rip),%ymm8        # 6590 <_sk_callback_avx+0x346>
+  .byte  196,98,125,24,5,11,51,0,0           // vbroadcastss  0x330b(%rip),%ymm8        # 658c <_sk_callback_avx+0x346>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,98,125,24,13,5,51,0,0           // vbroadcastss  0x3305(%rip),%ymm9        # 6594 <_sk_callback_avx+0x34a>
+  .byte  196,98,125,24,13,1,51,0,0           // vbroadcastss  0x3301(%rip),%ymm9        # 6590 <_sk_callback_avx+0x34a>
   .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  196,98,125,24,5,241,50,0,0          // vbroadcastss  0x32f1(%rip),%ymm8        # 6598 <_sk_callback_avx+0x34e>
+  .byte  196,98,125,24,5,237,50,0,0          // vbroadcastss  0x32ed(%rip),%ymm8        # 6594 <_sk_callback_avx+0x34e>
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,231,50,0,0          // vbroadcastss  0x32e7(%rip),%ymm8        # 659c <_sk_callback_avx+0x352>
+  .byte  196,98,125,24,5,227,50,0,0          // vbroadcastss  0x32e3(%rip),%ymm8        # 6598 <_sk_callback_avx+0x352>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,221,50,0,0          // vbroadcastss  0x32dd(%rip),%ymm8        # 65a0 <_sk_callback_avx+0x356>
+  .byte  196,98,125,24,5,217,50,0,0          // vbroadcastss  0x32d9(%rip),%ymm8        # 659c <_sk_callback_avx+0x356>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  197,252,88,201                      // vaddps        %ymm1,%ymm0,%ymm1
-  .byte  196,98,125,24,5,207,50,0,0          // vbroadcastss  0x32cf(%rip),%ymm8        # 65a4 <_sk_callback_avx+0x35a>
+  .byte  196,98,125,24,5,203,50,0,0          // vbroadcastss  0x32cb(%rip),%ymm8        # 65a0 <_sk_callback_avx+0x35a>
   .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  196,98,125,24,13,184,50,0,0         // vbroadcastss  0x32b8(%rip),%ymm9        # 65a8 <_sk_callback_avx+0x35e>
+  .byte  196,98,125,24,13,180,50,0,0         // vbroadcastss  0x32b4(%rip),%ymm9        # 65a4 <_sk_callback_avx+0x35e>
   .byte  196,65,52,194,208,1                 // vcmpltps      %ymm8,%ymm9,%ymm10
-  .byte  196,98,125,24,29,173,50,0,0         // vbroadcastss  0x32ad(%rip),%ymm11        # 65ac <_sk_callback_avx+0x362>
+  .byte  196,98,125,24,29,169,50,0,0         // vbroadcastss  0x32a9(%rip),%ymm11        # 65a8 <_sk_callback_avx+0x362>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,37,163,50,0,0         // vbroadcastss  0x32a3(%rip),%ymm12        # 65b0 <_sk_callback_avx+0x366>
+  .byte  196,98,125,24,37,159,50,0,0         // vbroadcastss  0x329f(%rip),%ymm12        # 65ac <_sk_callback_avx+0x366>
   .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
@@ -16931,9 +16926,9 @@ _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  196,226,125,24,5,89,50,0,0          // vbroadcastss  0x3259(%rip),%ymm0        # 65b4 <_sk_callback_avx+0x36a>
+  .byte  196,226,125,24,5,85,50,0,0          // vbroadcastss  0x3255(%rip),%ymm0        # 65b0 <_sk_callback_avx+0x36a>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,80,50,0,0           // vbroadcastss  0x3250(%rip),%ymm8        # 65b8 <_sk_callback_avx+0x36e>
+  .byte  196,98,125,24,5,76,50,0,0           // vbroadcastss  0x324c(%rip),%ymm8        # 65b4 <_sk_callback_avx+0x36e>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16954,7 +16949,7 @@ _sk_load_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  196,226,125,24,13,20,50,0,0         // vbroadcastss  0x3214(%rip),%ymm1        # 65bc <_sk_callback_avx+0x372>
+  .byte  196,226,125,24,13,16,50,0,0         // vbroadcastss  0x3210(%rip),%ymm1        # 65b8 <_sk_callback_avx+0x372>
   .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
@@ -17023,7 +17018,7 @@ _sk_gather_a8_avx:
   .byte  196,226,121,49,201                  // vpmovzxbd     %xmm1,%xmm1
   .byte  196,227,125,24,193,1                // vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,9,49,0,0          // vbroadcastss  0x3109(%rip),%ymm1        # 65c0 <_sk_callback_avx+0x376>
+  .byte  196,226,125,24,13,5,49,0,0          // vbroadcastss  0x3105(%rip),%ymm1        # 65bc <_sk_callback_avx+0x376>
   .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
@@ -17041,7 +17036,7 @@ FUNCTION(_sk_store_a8_avx)
 _sk_store_a8_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,228,48,0,0          // vbroadcastss  0x30e4(%rip),%ymm8        # 65c4 <_sk_callback_avx+0x37a>
+  .byte  196,98,125,24,5,224,48,0,0          // vbroadcastss  0x30e0(%rip),%ymm8        # 65c0 <_sk_callback_avx+0x37a>
   .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
@@ -17111,10 +17106,10 @@ _sk_load_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  196,226,125,24,13,9,48,0,0          // vbroadcastss  0x3009(%rip),%ymm1        # 65c8 <_sk_callback_avx+0x37e>
+  .byte  196,226,125,24,13,5,48,0,0          // vbroadcastss  0x3005(%rip),%ymm1        # 65c4 <_sk_callback_avx+0x37e>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,254,47,0,0        // vbroadcastss  0x2ffe(%rip),%ymm3        # 65cc <_sk_callback_avx+0x382>
+  .byte  196,226,125,24,29,250,47,0,0        // vbroadcastss  0x2ffa(%rip),%ymm3        # 65c8 <_sk_callback_avx+0x382>
   .byte  76,137,193                          // mov           %r8,%rcx
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
@@ -17180,10 +17175,10 @@ _sk_gather_g8_avx:
   .byte  196,226,121,49,201                  // vpmovzxbd     %xmm1,%xmm1
   .byte  196,227,125,24,193,1                // vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,253,46,0,0        // vbroadcastss  0x2efd(%rip),%ymm1        # 65d0 <_sk_callback_avx+0x386>
+  .byte  196,226,125,24,13,249,46,0,0        // vbroadcastss  0x2ef9(%rip),%ymm1        # 65cc <_sk_callback_avx+0x386>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,242,46,0,0        // vbroadcastss  0x2ef2(%rip),%ymm3        # 65d4 <_sk_callback_avx+0x38a>
+  .byte  196,226,125,24,29,238,46,0,0        // vbroadcastss  0x2eee(%rip),%ymm3        # 65d0 <_sk_callback_avx+0x38a>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -17263,10 +17258,10 @@ _sk_gather_i8_avx:
   .byte  196,163,121,34,4,163,2              // vpinsrd       $0x2,(%rbx,%r12,4),%xmm0,%xmm0
   .byte  196,163,121,34,28,19,3              // vpinsrd       $0x3,(%rbx,%r10,1),%xmm0,%xmm3
   .byte  196,227,61,24,195,1                 // vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  .byte  197,124,40,21,114,47,0,0            // vmovaps       0x2f72(%rip),%ymm10        # 67a0 <_sk_callback_avx+0x556>
+  .byte  197,124,40,21,114,47,0,0            // vmovaps       0x2f72(%rip),%ymm10        # 67a0 <_sk_callback_avx+0x55a>
   .byte  196,193,124,84,194                  // vandps        %ymm10,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,152,45,0,0         // vbroadcastss  0x2d98(%rip),%ymm9        # 65d8 <_sk_callback_avx+0x38e>
+  .byte  196,98,125,24,13,148,45,0,0         // vbroadcastss  0x2d94(%rip),%ymm9        # 65d4 <_sk_callback_avx+0x38e>
   .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
@@ -17306,23 +17301,23 @@ _sk_load_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  196,226,125,24,5,2,45,0,0           // vbroadcastss  0x2d02(%rip),%ymm0        # 65dc <_sk_callback_avx+0x392>
+  .byte  196,226,125,24,5,254,44,0,0         // vbroadcastss  0x2cfe(%rip),%ymm0        # 65d8 <_sk_callback_avx+0x392>
   .byte  197,236,84,192                      // vandps        %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,245,44,0,0        // vbroadcastss  0x2cf5(%rip),%ymm1        # 65e0 <_sk_callback_avx+0x396>
+  .byte  196,226,125,24,13,241,44,0,0        // vbroadcastss  0x2cf1(%rip),%ymm1        # 65dc <_sk_callback_avx+0x396>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,236,44,0,0        // vbroadcastss  0x2cec(%rip),%ymm1        # 65e4 <_sk_callback_avx+0x39a>
+  .byte  196,226,125,24,13,232,44,0,0        // vbroadcastss  0x2ce8(%rip),%ymm1        # 65e0 <_sk_callback_avx+0x39a>
   .byte  197,236,84,201                      // vandps        %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,223,44,0,0        // vbroadcastss  0x2cdf(%rip),%ymm3        # 65e8 <_sk_callback_avx+0x39e>
+  .byte  196,226,125,24,29,219,44,0,0        // vbroadcastss  0x2cdb(%rip),%ymm3        # 65e4 <_sk_callback_avx+0x39e>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,24,29,214,44,0,0        // vbroadcastss  0x2cd6(%rip),%ymm3        # 65ec <_sk_callback_avx+0x3a2>
+  .byte  196,226,125,24,29,210,44,0,0        // vbroadcastss  0x2cd2(%rip),%ymm3        # 65e8 <_sk_callback_avx+0x3a2>
   .byte  197,236,84,211                      // vandps        %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,201,44,0,0        // vbroadcastss  0x2cc9(%rip),%ymm3        # 65f0 <_sk_callback_avx+0x3a6>
+  .byte  196,226,125,24,29,197,44,0,0        // vbroadcastss  0x2cc5(%rip),%ymm3        # 65ec <_sk_callback_avx+0x3a6>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,190,44,0,0        // vbroadcastss  0x2cbe(%rip),%ymm3        # 65f4 <_sk_callback_avx+0x3aa>
+  .byte  196,226,125,24,29,186,44,0,0        // vbroadcastss  0x2cba(%rip),%ymm3        # 65f0 <_sk_callback_avx+0x3aa>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,200                          // mov           %ecx,%r8d
   .byte  65,128,224,7                        // and           $0x7,%r8b
@@ -17421,23 +17416,23 @@ _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  196,226,125,24,5,94,43,0,0          // vbroadcastss  0x2b5e(%rip),%ymm0        # 65f8 <_sk_callback_avx+0x3ae>
+  .byte  196,226,125,24,5,90,43,0,0          // vbroadcastss  0x2b5a(%rip),%ymm0        # 65f4 <_sk_callback_avx+0x3ae>
   .byte  197,236,84,192                      // vandps        %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,81,43,0,0         // vbroadcastss  0x2b51(%rip),%ymm1        # 65fc <_sk_callback_avx+0x3b2>
+  .byte  196,226,125,24,13,77,43,0,0         // vbroadcastss  0x2b4d(%rip),%ymm1        # 65f8 <_sk_callback_avx+0x3b2>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,72,43,0,0         // vbroadcastss  0x2b48(%rip),%ymm1        # 6600 <_sk_callback_avx+0x3b6>
+  .byte  196,226,125,24,13,68,43,0,0         // vbroadcastss  0x2b44(%rip),%ymm1        # 65fc <_sk_callback_avx+0x3b6>
   .byte  197,236,84,201                      // vandps        %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,59,43,0,0         // vbroadcastss  0x2b3b(%rip),%ymm3        # 6604 <_sk_callback_avx+0x3ba>
+  .byte  196,226,125,24,29,55,43,0,0         // vbroadcastss  0x2b37(%rip),%ymm3        # 6600 <_sk_callback_avx+0x3ba>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,24,29,50,43,0,0         // vbroadcastss  0x2b32(%rip),%ymm3        # 6608 <_sk_callback_avx+0x3be>
+  .byte  196,226,125,24,29,46,43,0,0         // vbroadcastss  0x2b2e(%rip),%ymm3        # 6604 <_sk_callback_avx+0x3be>
   .byte  197,236,84,211                      // vandps        %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,37,43,0,0         // vbroadcastss  0x2b25(%rip),%ymm3        # 660c <_sk_callback_avx+0x3c2>
+  .byte  196,226,125,24,29,33,43,0,0         // vbroadcastss  0x2b21(%rip),%ymm3        # 6608 <_sk_callback_avx+0x3c2>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,26,43,0,0         // vbroadcastss  0x2b1a(%rip),%ymm3        # 6610 <_sk_callback_avx+0x3c6>
+  .byte  196,226,125,24,29,22,43,0,0         // vbroadcastss  0x2b16(%rip),%ymm3        # 660c <_sk_callback_avx+0x3c6>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -17451,14 +17446,14 @@ FUNCTION(_sk_store_565_avx)
 _sk_store_565_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,6,43,0,0            // vbroadcastss  0x2b06(%rip),%ymm8        # 6614 <_sk_callback_avx+0x3ca>
+  .byte  196,98,125,24,5,2,43,0,0            // vbroadcastss  0x2b02(%rip),%ymm8        # 6610 <_sk_callback_avx+0x3ca>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,41,114,241,11               // 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  196,98,125,24,21,223,42,0,0         // vbroadcastss  0x2adf(%rip),%ymm10        # 6618 <_sk_callback_avx+0x3ce>
+  .byte  196,98,125,24,21,219,42,0,0         // vbroadcastss  0x2adb(%rip),%ymm10        # 6614 <_sk_callback_avx+0x3ce>
   .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
@@ -17532,25 +17527,25 @@ _sk_load_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  196,226,125,24,5,232,41,0,0         // vbroadcastss  0x29e8(%rip),%ymm0        # 661c <_sk_callback_avx+0x3d2>
+  .byte  196,226,125,24,5,228,41,0,0         // vbroadcastss  0x29e4(%rip),%ymm0        # 6618 <_sk_callback_avx+0x3d2>
   .byte  197,228,84,192                      // vandps        %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,219,41,0,0        // vbroadcastss  0x29db(%rip),%ymm1        # 6620 <_sk_callback_avx+0x3d6>
+  .byte  196,226,125,24,13,215,41,0,0        // vbroadcastss  0x29d7(%rip),%ymm1        # 661c <_sk_callback_avx+0x3d6>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,210,41,0,0        // vbroadcastss  0x29d2(%rip),%ymm1        # 6624 <_sk_callback_avx+0x3da>
+  .byte  196,226,125,24,13,206,41,0,0        // vbroadcastss  0x29ce(%rip),%ymm1        # 6620 <_sk_callback_avx+0x3da>
   .byte  197,228,84,201                      // vandps        %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,197,41,0,0        // vbroadcastss  0x29c5(%rip),%ymm2        # 6628 <_sk_callback_avx+0x3de>
+  .byte  196,226,125,24,21,193,41,0,0        // vbroadcastss  0x29c1(%rip),%ymm2        # 6624 <_sk_callback_avx+0x3de>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,188,41,0,0        // vbroadcastss  0x29bc(%rip),%ymm2        # 662c <_sk_callback_avx+0x3e2>
+  .byte  196,226,125,24,21,184,41,0,0        // vbroadcastss  0x29b8(%rip),%ymm2        # 6628 <_sk_callback_avx+0x3e2>
   .byte  197,228,84,210                      // vandps        %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,175,41,0,0          // vbroadcastss  0x29af(%rip),%ymm8        # 6630 <_sk_callback_avx+0x3e6>
+  .byte  196,98,125,24,5,171,41,0,0          // vbroadcastss  0x29ab(%rip),%ymm8        # 662c <_sk_callback_avx+0x3e6>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,165,41,0,0          // vbroadcastss  0x29a5(%rip),%ymm8        # 6634 <_sk_callback_avx+0x3ea>
+  .byte  196,98,125,24,5,161,41,0,0          // vbroadcastss  0x29a1(%rip),%ymm8        # 6630 <_sk_callback_avx+0x3ea>
   .byte  196,193,100,84,216                  // vandps        %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,151,41,0,0          // vbroadcastss  0x2997(%rip),%ymm8        # 6638 <_sk_callback_avx+0x3ee>
+  .byte  196,98,125,24,5,147,41,0,0          // vbroadcastss  0x2993(%rip),%ymm8        # 6634 <_sk_callback_avx+0x3ee>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -17652,25 +17647,25 @@ _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  196,226,125,24,5,46,40,0,0          // vbroadcastss  0x282e(%rip),%ymm0        # 663c <_sk_callback_avx+0x3f2>
+  .byte  196,226,125,24,5,42,40,0,0          // vbroadcastss  0x282a(%rip),%ymm0        # 6638 <_sk_callback_avx+0x3f2>
   .byte  197,228,84,192                      // vandps        %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,33,40,0,0         // vbroadcastss  0x2821(%rip),%ymm1        # 6640 <_sk_callback_avx+0x3f6>
+  .byte  196,226,125,24,13,29,40,0,0         // vbroadcastss  0x281d(%rip),%ymm1        # 663c <_sk_callback_avx+0x3f6>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,24,40,0,0         // vbroadcastss  0x2818(%rip),%ymm1        # 6644 <_sk_callback_avx+0x3fa>
+  .byte  196,226,125,24,13,20,40,0,0         // vbroadcastss  0x2814(%rip),%ymm1        # 6640 <_sk_callback_avx+0x3fa>
   .byte  197,228,84,201                      // vandps        %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,11,40,0,0         // vbroadcastss  0x280b(%rip),%ymm2        # 6648 <_sk_callback_avx+0x3fe>
+  .byte  196,226,125,24,21,7,40,0,0          // vbroadcastss  0x2807(%rip),%ymm2        # 6644 <_sk_callback_avx+0x3fe>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,2,40,0,0          // vbroadcastss  0x2802(%rip),%ymm2        # 664c <_sk_callback_avx+0x402>
+  .byte  196,226,125,24,21,254,39,0,0        // vbroadcastss  0x27fe(%rip),%ymm2        # 6648 <_sk_callback_avx+0x402>
   .byte  197,228,84,210                      // vandps        %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,245,39,0,0          // vbroadcastss  0x27f5(%rip),%ymm8        # 6650 <_sk_callback_avx+0x406>
+  .byte  196,98,125,24,5,241,39,0,0          // vbroadcastss  0x27f1(%rip),%ymm8        # 664c <_sk_callback_avx+0x406>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,235,39,0,0          // vbroadcastss  0x27eb(%rip),%ymm8        # 6654 <_sk_callback_avx+0x40a>
+  .byte  196,98,125,24,5,231,39,0,0          // vbroadcastss  0x27e7(%rip),%ymm8        # 6650 <_sk_callback_avx+0x40a>
   .byte  196,193,100,84,216                  // vandps        %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,221,39,0,0          // vbroadcastss  0x27dd(%rip),%ymm8        # 6658 <_sk_callback_avx+0x40e>
+  .byte  196,98,125,24,5,217,39,0,0          // vbroadcastss  0x27d9(%rip),%ymm8        # 6654 <_sk_callback_avx+0x40e>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -17686,7 +17681,7 @@ FUNCTION(_sk_store_4444_avx)
 _sk_store_4444_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,194,39,0,0          // vbroadcastss  0x27c2(%rip),%ymm8        # 665c <_sk_callback_avx+0x412>
+  .byte  196,98,125,24,5,190,39,0,0          // vbroadcastss  0x27be(%rip),%ymm8        # 6658 <_sk_callback_avx+0x412>
   .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
@@ -17767,10 +17762,10 @@ _sk_load_8888_avx:
   .byte  72,133,201                          // test          %rcx,%rcx
   .byte  15,133,135,0,0,0                    // jne           4041 <_sk_load_8888_avx+0x95>
   .byte  196,65,124,16,12,186                // vmovups       (%r10,%rdi,4),%ymm9
-  .byte  197,124,40,21,248,39,0,0            // vmovaps       0x27f8(%rip),%ymm10        # 67c0 <_sk_callback_avx+0x576>
+  .byte  197,124,40,21,248,39,0,0            // vmovaps       0x27f8(%rip),%ymm10        # 67c0 <_sk_callback_avx+0x57a>
   .byte  196,193,52,84,194                   // vandps        %ymm10,%ymm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,134,38,0,0          // vbroadcastss  0x2686(%rip),%ymm8        # 6660 <_sk_callback_avx+0x416>
+  .byte  196,98,125,24,5,130,38,0,0          // vbroadcastss  0x2682(%rip),%ymm8        # 665c <_sk_callback_avx+0x416>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  196,193,113,114,209,8               // vpsrld        $0x8,%xmm9,%xmm1
   .byte  196,99,125,25,203,1                 // vextractf128  $0x1,%ymm9,%xmm3
@@ -17885,10 +17880,10 @@ _sk_gather_8888_avx:
   .byte  196,131,121,34,4,152,2              // vpinsrd       $0x2,(%r8,%r11,4),%xmm0,%xmm0
   .byte  196,131,121,34,28,144,3             // vpinsrd       $0x3,(%r8,%r10,4),%xmm0,%xmm3
   .byte  196,227,61,24,195,1                 // vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  .byte  197,124,40,21,34,38,0,0             // vmovaps       0x2622(%rip),%ymm10        # 67e0 <_sk_callback_avx+0x596>
+  .byte  197,124,40,21,34,38,0,0             // vmovaps       0x2622(%rip),%ymm10        # 67e0 <_sk_callback_avx+0x59a>
   .byte  196,193,124,84,194                  // vandps        %ymm10,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,148,36,0,0         // vbroadcastss  0x2494(%rip),%ymm9        # 6664 <_sk_callback_avx+0x41a>
+  .byte  196,98,125,24,13,144,36,0,0         // vbroadcastss  0x2490(%rip),%ymm9        # 6660 <_sk_callback_avx+0x41a>
   .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
@@ -17920,7 +17915,7 @@ FUNCTION(_sk_store_8888_avx)
 _sk_store_8888_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,34,36,0,0           // vbroadcastss  0x2422(%rip),%ymm8        # 6668 <_sk_callback_avx+0x41e>
+  .byte  196,98,125,24,5,30,36,0,0           // vbroadcastss  0x241e(%rip),%ymm8        # 6664 <_sk_callback_avx+0x41e>
   .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
@@ -18025,13 +18020,13 @@ _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  196,98,125,24,37,137,34,0,0         // vbroadcastss  0x2289(%rip),%ymm12        # 666c <_sk_callback_avx+0x422>
+  .byte  196,98,125,24,37,133,34,0,0         // vbroadcastss  0x2285(%rip),%ymm12        # 6668 <_sk_callback_avx+0x422>
   .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  196,98,121,24,29,117,34,0,0         // vbroadcastss  0x2275(%rip),%xmm11        # 6670 <_sk_callback_avx+0x426>
+  .byte  196,98,121,24,29,113,34,0,0         // vbroadcastss  0x2271(%rip),%xmm11        # 666c <_sk_callback_avx+0x426>
   .byte  196,193,8,87,219                    // vxorps        %xmm11,%xmm14,%xmm3
-  .byte  196,98,121,24,45,107,34,0,0         // vbroadcastss  0x226b(%rip),%xmm13        # 6674 <_sk_callback_avx+0x42a>
+  .byte  196,98,121,24,45,103,34,0,0         // vbroadcastss  0x2267(%rip),%xmm13        # 6670 <_sk_callback_avx+0x42a>
   .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
@@ -18045,7 +18040,7 @@ _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  196,226,121,24,29,33,34,0,0         // vbroadcastss  0x2221(%rip),%xmm3        # 6678 <_sk_callback_avx+0x42e>
+  .byte  196,226,121,24,29,29,34,0,0         // vbroadcastss  0x221d(%rip),%xmm3        # 6674 <_sk_callback_avx+0x42e>
   .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
@@ -18224,13 +18219,13 @@ _sk_gather_f16_avx:
   .byte  197,249,105,210                     // vpunpckhwd    %xmm2,%xmm0,%xmm2
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,125,24,194,1                // vinsertf128   $0x1,%xmm2,%ymm0,%ymm0
-  .byte  196,98,125,24,37,229,30,0,0         // vbroadcastss  0x1ee5(%rip),%ymm12        # 667c <_sk_callback_avx+0x432>
+  .byte  196,98,125,24,37,225,30,0,0         // vbroadcastss  0x1ee1(%rip),%ymm12        # 6678 <_sk_callback_avx+0x432>
   .byte  196,193,124,84,212                  // vandps        %ymm12,%ymm0,%ymm2
   .byte  197,252,87,194                      // vxorps        %ymm2,%ymm0,%ymm0
   .byte  196,195,125,25,198,1                // vextractf128  $0x1,%ymm0,%xmm14
-  .byte  196,98,121,24,29,209,30,0,0         // vbroadcastss  0x1ed1(%rip),%xmm11        # 6680 <_sk_callback_avx+0x436>
+  .byte  196,98,121,24,29,205,30,0,0         // vbroadcastss  0x1ecd(%rip),%xmm11        # 667c <_sk_callback_avx+0x436>
   .byte  196,193,8,87,219                    // vxorps        %xmm11,%xmm14,%xmm3
-  .byte  196,98,121,24,45,199,30,0,0         // vbroadcastss  0x1ec7(%rip),%xmm13        # 6684 <_sk_callback_avx+0x43a>
+  .byte  196,98,121,24,45,195,30,0,0         // vbroadcastss  0x1ec3(%rip),%xmm13        # 6680 <_sk_callback_avx+0x43a>
   .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
@@ -18244,7 +18239,7 @@ _sk_gather_f16_avx:
   .byte  196,227,125,24,195,1                // vinsertf128   $0x1,%xmm3,%ymm0,%ymm0
   .byte  197,252,86,194                      // vorps         %ymm2,%ymm0,%ymm0
   .byte  196,227,125,25,194,1                // vextractf128  $0x1,%ymm0,%xmm2
-  .byte  196,226,121,24,29,125,30,0,0        // vbroadcastss  0x1e7d(%rip),%xmm3        # 6688 <_sk_callback_avx+0x43e>
+  .byte  196,226,121,24,29,121,30,0,0        // vbroadcastss  0x1e79(%rip),%xmm3        # 6684 <_sk_callback_avx+0x43e>
   .byte  197,233,254,211                     // vpaddd        %xmm3,%xmm2,%xmm2
   .byte  197,249,254,195                     // vpaddd        %xmm3,%xmm0,%xmm0
   .byte  196,227,125,24,194,1                // vinsertf128   $0x1,%xmm2,%ymm0,%ymm0
@@ -18348,12 +18343,12 @@ _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  196,98,125,24,13,150,28,0,0         // vbroadcastss  0x1c96(%rip),%ymm9        # 668c <_sk_callback_avx+0x442>
+  .byte  196,98,125,24,13,146,28,0,0         // vbroadcastss  0x1c92(%rip),%ymm9        # 6688 <_sk_callback_avx+0x442>
   .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  196,98,121,24,5,123,28,0,0          // vbroadcastss  0x1c7b(%rip),%xmm8        # 6690 <_sk_callback_avx+0x446>
+  .byte  196,98,121,24,5,119,28,0,0          // vbroadcastss  0x1c77(%rip),%xmm8        # 668c <_sk_callback_avx+0x446>
   .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
@@ -18363,7 +18358,7 @@ _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  196,98,125,24,21,66,28,0,0          // vbroadcastss  0x1c42(%rip),%ymm10        # 6694 <_sk_callback_avx+0x44a>
+  .byte  196,98,125,24,21,62,28,0,0          // vbroadcastss  0x1c3e(%rip),%ymm10        # 6690 <_sk_callback_avx+0x44a>
   .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
@@ -18508,7 +18503,7 @@ _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  196,98,125,24,29,154,25,0,0         // vbroadcastss  0x199a(%rip),%ymm11        # 6698 <_sk_callback_avx+0x44e>
+  .byte  196,98,125,24,29,150,25,0,0         // vbroadcastss  0x1996(%rip),%ymm11        # 6694 <_sk_callback_avx+0x44e>
   .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
@@ -18601,7 +18596,7 @@ _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  196,98,125,24,29,250,23,0,0         // vbroadcastss  0x17fa(%rip),%ymm11        # 669c <_sk_callback_avx+0x452>
+  .byte  196,98,125,24,29,246,23,0,0         // vbroadcastss  0x17f6(%rip),%ymm11        # 6698 <_sk_callback_avx+0x452>
   .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
@@ -18622,7 +18617,7 @@ _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  196,226,125,24,29,151,23,0,0        // vbroadcastss  0x1797(%rip),%ymm3        # 66a0 <_sk_callback_avx+0x456>
+  .byte  196,226,125,24,29,147,23,0,0        // vbroadcastss  0x1793(%rip),%ymm3        # 669c <_sk_callback_avx+0x456>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,4,64                // vmovd         (%r8,%rax,2),%xmm0
   .byte  196,193,121,196,68,64,4,2           // vpinsrw       $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
@@ -18665,7 +18660,7 @@ _sk_store_u16_be_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  72,141,4,189,0,0,0,0                // lea           0x0(,%rdi,4),%rax
-  .byte  196,98,125,24,5,212,22,0,0          // vbroadcastss  0x16d4(%rip),%ymm8        # 66a4 <_sk_callback_avx+0x45a>
+  .byte  196,98,125,24,5,208,22,0,0          // vbroadcastss  0x16d0(%rip),%ymm8        # 66a0 <_sk_callback_avx+0x45a>
   .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
@@ -18931,12 +18926,12 @@ HIDDEN _sk_luminance_to_alpha_avx
 .globl _sk_luminance_to_alpha_avx
 FUNCTION(_sk_luminance_to_alpha_avx)
 _sk_luminance_to_alpha_avx:
-  .byte  196,226,125,24,29,251,18,0,0        // vbroadcastss  0x12fb(%rip),%ymm3        # 66a8 <_sk_callback_avx+0x45e>
+  .byte  196,226,125,24,29,247,18,0,0        // vbroadcastss  0x12f7(%rip),%ymm3        # 66a4 <_sk_callback_avx+0x45e>
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,24,29,242,18,0,0        // vbroadcastss  0x12f2(%rip),%ymm3        # 66ac <_sk_callback_avx+0x462>
+  .byte  196,226,125,24,29,238,18,0,0        // vbroadcastss  0x12ee(%rip),%ymm3        # 66a8 <_sk_callback_avx+0x462>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,229,18,0,0        // vbroadcastss  0x12e5(%rip),%ymm1        # 66b0 <_sk_callback_avx+0x466>
+  .byte  196,226,125,24,13,225,18,0,0        // vbroadcastss  0x12e1(%rip),%ymm1        # 66ac <_sk_callback_avx+0x466>
   .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
@@ -19291,7 +19286,7 @@ _sk_gradient_avx:
   .byte  73,255,200                          // dec           %r8
   .byte  72,131,195,4                        // add           $0x4,%rbx
   .byte  196,65,52,87,201                    // vxorps        %ymm9,%ymm9,%ymm9
-  .byte  196,98,125,24,21,192,12,0,0         // vbroadcastss  0xcc0(%rip),%ymm10        # 66b4 <_sk_callback_avx+0x46a>
+  .byte  196,98,125,24,21,188,12,0,0         // vbroadcastss  0xcbc(%rip),%ymm10        # 66b0 <_sk_callback_avx+0x46a>
   .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
@@ -19485,27 +19480,27 @@ _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  196,98,125,24,45,228,8,0,0          // vbroadcastss  0x8e4(%rip),%ymm13        # 66b8 <_sk_callback_avx+0x46e>
+  .byte  196,98,125,24,45,224,8,0,0          // vbroadcastss  0x8e0(%rip),%ymm13        # 66b4 <_sk_callback_avx+0x46e>
   .byte  196,65,28,89,237                    // vmulps        %ymm13,%ymm12,%ymm13
-  .byte  196,98,125,24,53,218,8,0,0          // vbroadcastss  0x8da(%rip),%ymm14        # 66bc <_sk_callback_avx+0x472>
+  .byte  196,98,125,24,53,214,8,0,0          // vbroadcastss  0x8d6(%rip),%ymm14        # 66b8 <_sk_callback_avx+0x472>
   .byte  196,65,20,88,238                    // vaddps        %ymm14,%ymm13,%ymm13
   .byte  196,65,28,89,237                    // vmulps        %ymm13,%ymm12,%ymm13
-  .byte  196,98,125,24,53,203,8,0,0          // vbroadcastss  0x8cb(%rip),%ymm14        # 66c0 <_sk_callback_avx+0x476>
+  .byte  196,98,125,24,53,199,8,0,0          // vbroadcastss  0x8c7(%rip),%ymm14        # 66bc <_sk_callback_avx+0x476>
   .byte  196,65,20,88,238                    // vaddps        %ymm14,%ymm13,%ymm13
   .byte  196,65,28,89,229                    // vmulps        %ymm13,%ymm12,%ymm12
-  .byte  196,98,125,24,45,188,8,0,0          // vbroadcastss  0x8bc(%rip),%ymm13        # 66c4 <_sk_callback_avx+0x47a>
+  .byte  196,98,125,24,45,184,8,0,0          // vbroadcastss  0x8b8(%rip),%ymm13        # 66c0 <_sk_callback_avx+0x47a>
   .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  196,98,125,24,21,167,8,0,0          // vbroadcastss  0x8a7(%rip),%ymm10        # 66c8 <_sk_callback_avx+0x47e>
+  .byte  196,98,125,24,21,163,8,0,0          // vbroadcastss  0x8a3(%rip),%ymm10        # 66c4 <_sk_callback_avx+0x47e>
   .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  196,98,125,24,21,145,8,0,0          // vbroadcastss  0x891(%rip),%ymm10        # 66cc <_sk_callback_avx+0x482>
+  .byte  196,98,125,24,21,141,8,0,0          // vbroadcastss  0x88d(%rip),%ymm10        # 66c8 <_sk_callback_avx+0x482>
   .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  196,98,125,24,21,123,8,0,0          // vbroadcastss  0x87b(%rip),%ymm10        # 66d0 <_sk_callback_avx+0x486>
+  .byte  196,98,125,24,21,119,8,0,0          // vbroadcastss  0x877(%rip),%ymm10        # 66cc <_sk_callback_avx+0x486>
   .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
@@ -19520,8 +19515,7 @@ _sk_xy_to_radius_avx:
   .byte  197,252,89,192                      // vmulps        %ymm0,%ymm0,%ymm0
   .byte  197,116,89,193                      // vmulps        %ymm1,%ymm1,%ymm8
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  197,252,82,192                      // vrsqrtps      %ymm0,%ymm0
-  .byte  197,252,83,192                      // vrcpps        %ymm0,%ymm0
+  .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -19530,7 +19524,7 @@ HIDDEN _sk_save_xy_avx
 FUNCTION(_sk_save_xy_avx)
 _sk_save_xy_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,65,8,0,0            // vbroadcastss  0x841(%rip),%ymm8        # 66d4 <_sk_callback_avx+0x48a>
+  .byte  196,98,125,24,5,65,8,0,0            // vbroadcastss  0x841(%rip),%ymm8        # 66d0 <_sk_callback_avx+0x48a>
   .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
@@ -19567,9 +19561,9 @@ HIDDEN _sk_bilinear_nx_avx
 FUNCTION(_sk_bilinear_nx_avx)
 _sk_bilinear_nx_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,205,7,0,0          // vbroadcastss  0x7cd(%rip),%ymm0        # 66d8 <_sk_callback_avx+0x48e>
+  .byte  196,226,125,24,5,205,7,0,0          // vbroadcastss  0x7cd(%rip),%ymm0        # 66d4 <_sk_callback_avx+0x48e>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,196,7,0,0           // vbroadcastss  0x7c4(%rip),%ymm8        # 66dc <_sk_callback_avx+0x492>
+  .byte  196,98,125,24,5,196,7,0,0           // vbroadcastss  0x7c4(%rip),%ymm8        # 66d8 <_sk_callback_avx+0x492>
   .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
@@ -19580,7 +19574,7 @@ HIDDEN _sk_bilinear_px_avx
 FUNCTION(_sk_bilinear_px_avx)
 _sk_bilinear_px_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,172,7,0,0          // vbroadcastss  0x7ac(%rip),%ymm0        # 66e0 <_sk_callback_avx+0x496>
+  .byte  196,226,125,24,5,172,7,0,0          // vbroadcastss  0x7ac(%rip),%ymm0        # 66dc <_sk_callback_avx+0x496>
   .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)
@@ -19592,9 +19586,9 @@ HIDDEN _sk_bilinear_ny_avx
 FUNCTION(_sk_bilinear_ny_avx)
 _sk_bilinear_ny_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,144,7,0,0         // vbroadcastss  0x790(%rip),%ymm1        # 66e4 <_sk_callback_avx+0x49a>
+  .byte  196,226,125,24,13,144,7,0,0         // vbroadcastss  0x790(%rip),%ymm1        # 66e0 <_sk_callback_avx+0x49a>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,134,7,0,0           // vbroadcastss  0x786(%rip),%ymm8        # 66e8 <_sk_callback_avx+0x49e>
+  .byte  196,98,125,24,5,134,7,0,0           // vbroadcastss  0x786(%rip),%ymm8        # 66e4 <_sk_callback_avx+0x49e>
   .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
@@ -19605,7 +19599,7 @@ HIDDEN _sk_bilinear_py_avx
 FUNCTION(_sk_bilinear_py_avx)
 _sk_bilinear_py_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,110,7,0,0         // vbroadcastss  0x76e(%rip),%ymm1        # 66ec <_sk_callback_avx+0x4a2>
+  .byte  196,226,125,24,13,110,7,0,0         // vbroadcastss  0x76e(%rip),%ymm1        # 66e8 <_sk_callback_avx+0x4a2>
   .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)
@@ -19617,14 +19611,14 @@ HIDDEN _sk_bicubic_n3x_avx
 FUNCTION(_sk_bicubic_n3x_avx)
 _sk_bicubic_n3x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,81,7,0,0           // vbroadcastss  0x751(%rip),%ymm0        # 66f0 <_sk_callback_avx+0x4a6>
+  .byte  196,226,125,24,5,81,7,0,0           // vbroadcastss  0x751(%rip),%ymm0        # 66ec <_sk_callback_avx+0x4a6>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,72,7,0,0            // vbroadcastss  0x748(%rip),%ymm8        # 66f4 <_sk_callback_avx+0x4aa>
+  .byte  196,98,125,24,5,72,7,0,0            // vbroadcastss  0x748(%rip),%ymm8        # 66f0 <_sk_callback_avx+0x4aa>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,57,7,0,0           // vbroadcastss  0x739(%rip),%ymm10        # 66f8 <_sk_callback_avx+0x4ae>
+  .byte  196,98,125,24,21,57,7,0,0           // vbroadcastss  0x739(%rip),%ymm10        # 66f4 <_sk_callback_avx+0x4ae>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,47,7,0,0           // vbroadcastss  0x72f(%rip),%ymm10        # 66fc <_sk_callback_avx+0x4b2>
+  .byte  196,98,125,24,21,47,7,0,0           // vbroadcastss  0x72f(%rip),%ymm10        # 66f8 <_sk_callback_avx+0x4b2>
   .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)
@@ -19636,19 +19630,19 @@ HIDDEN _sk_bicubic_n1x_avx
 FUNCTION(_sk_bicubic_n1x_avx)
 _sk_bicubic_n1x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,18,7,0,0           // vbroadcastss  0x712(%rip),%ymm0        # 6700 <_sk_callback_avx+0x4b6>
+  .byte  196,226,125,24,5,18,7,0,0           // vbroadcastss  0x712(%rip),%ymm0        # 66fc <_sk_callback_avx+0x4b6>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,9,7,0,0             // vbroadcastss  0x709(%rip),%ymm8        # 6704 <_sk_callback_avx+0x4ba>
+  .byte  196,98,125,24,5,9,7,0,0             // vbroadcastss  0x709(%rip),%ymm8        # 6700 <_sk_callback_avx+0x4ba>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,255,6,0,0          // vbroadcastss  0x6ff(%rip),%ymm9        # 6708 <_sk_callback_avx+0x4be>
+  .byte  196,98,125,24,13,255,6,0,0          // vbroadcastss  0x6ff(%rip),%ymm9        # 6704 <_sk_callback_avx+0x4be>
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,245,6,0,0          // vbroadcastss  0x6f5(%rip),%ymm10        # 670c <_sk_callback_avx+0x4c2>
+  .byte  196,98,125,24,21,245,6,0,0          // vbroadcastss  0x6f5(%rip),%ymm10        # 6708 <_sk_callback_avx+0x4c2>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,230,6,0,0          // vbroadcastss  0x6e6(%rip),%ymm10        # 6710 <_sk_callback_avx+0x4c6>
+  .byte  196,98,125,24,21,230,6,0,0          // vbroadcastss  0x6e6(%rip),%ymm10        # 670c <_sk_callback_avx+0x4c6>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,215,6,0,0          // vbroadcastss  0x6d7(%rip),%ymm9        # 6714 <_sk_callback_avx+0x4ca>
+  .byte  196,98,125,24,13,215,6,0,0          // vbroadcastss  0x6d7(%rip),%ymm9        # 6710 <_sk_callback_avx+0x4ca>
   .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
@@ -19659,17 +19653,17 @@ HIDDEN _sk_bicubic_p1x_avx
 FUNCTION(_sk_bicubic_p1x_avx)
 _sk_bicubic_p1x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,191,6,0,0           // vbroadcastss  0x6bf(%rip),%ymm8        # 6718 <_sk_callback_avx+0x4ce>
+  .byte  196,98,125,24,5,191,6,0,0           // vbroadcastss  0x6bf(%rip),%ymm8        # 6714 <_sk_callback_avx+0x4ce>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,72,64                    // vmovups       0x40(%rax),%ymm9
-  .byte  196,98,125,24,21,177,6,0,0          // vbroadcastss  0x6b1(%rip),%ymm10        # 671c <_sk_callback_avx+0x4d2>
+  .byte  196,98,125,24,21,177,6,0,0          // vbroadcastss  0x6b1(%rip),%ymm10        # 6718 <_sk_callback_avx+0x4d2>
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
-  .byte  196,98,125,24,29,167,6,0,0          // vbroadcastss  0x6a7(%rip),%ymm11        # 6720 <_sk_callback_avx+0x4d6>
+  .byte  196,98,125,24,29,167,6,0,0          // vbroadcastss  0x6a7(%rip),%ymm11        # 671c <_sk_callback_avx+0x4d6>
   .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  196,98,125,24,13,142,6,0,0          // vbroadcastss  0x68e(%rip),%ymm9        # 6724 <_sk_callback_avx+0x4da>
+  .byte  196,98,125,24,13,142,6,0,0          // vbroadcastss  0x68e(%rip),%ymm9        # 6720 <_sk_callback_avx+0x4da>
   .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
@@ -19680,13 +19674,13 @@ HIDDEN _sk_bicubic_p3x_avx
 FUNCTION(_sk_bicubic_p3x_avx)
 _sk_bicubic_p3x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,118,6,0,0          // vbroadcastss  0x676(%rip),%ymm0        # 6728 <_sk_callback_avx+0x4de>
+  .byte  196,226,125,24,5,118,6,0,0          // vbroadcastss  0x676(%rip),%ymm0        # 6724 <_sk_callback_avx+0x4de>
   .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  196,98,125,24,21,99,6,0,0           // vbroadcastss  0x663(%rip),%ymm10        # 672c <_sk_callback_avx+0x4e2>
+  .byte  196,98,125,24,21,99,6,0,0           // vbroadcastss  0x663(%rip),%ymm10        # 6728 <_sk_callback_avx+0x4e2>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,89,6,0,0           // vbroadcastss  0x659(%rip),%ymm10        # 6730 <_sk_callback_avx+0x4e6>
+  .byte  196,98,125,24,21,89,6,0,0           // vbroadcastss  0x659(%rip),%ymm10        # 672c <_sk_callback_avx+0x4e6>
   .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)
@@ -19698,14 +19692,14 @@ HIDDEN _sk_bicubic_n3y_avx
 FUNCTION(_sk_bicubic_n3y_avx)
 _sk_bicubic_n3y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,60,6,0,0          // vbroadcastss  0x63c(%rip),%ymm1        # 6734 <_sk_callback_avx+0x4ea>
+  .byte  196,226,125,24,13,60,6,0,0          // vbroadcastss  0x63c(%rip),%ymm1        # 6730 <_sk_callback_avx+0x4ea>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,50,6,0,0            // vbroadcastss  0x632(%rip),%ymm8        # 6738 <_sk_callback_avx+0x4ee>
+  .byte  196,98,125,24,5,50,6,0,0            // vbroadcastss  0x632(%rip),%ymm8        # 6734 <_sk_callback_avx+0x4ee>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,35,6,0,0           // vbroadcastss  0x623(%rip),%ymm10        # 673c <_sk_callback_avx+0x4f2>
+  .byte  196,98,125,24,21,35,6,0,0           // vbroadcastss  0x623(%rip),%ymm10        # 6738 <_sk_callback_avx+0x4f2>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,25,6,0,0           // vbroadcastss  0x619(%rip),%ymm10        # 6740 <_sk_callback_avx+0x4f6>
+  .byte  196,98,125,24,21,25,6,0,0           // vbroadcastss  0x619(%rip),%ymm10        # 673c <_sk_callback_avx+0x4f6>
   .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)
@@ -19717,19 +19711,19 @@ HIDDEN _sk_bicubic_n1y_avx
 FUNCTION(_sk_bicubic_n1y_avx)
 _sk_bicubic_n1y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,252,5,0,0         // vbroadcastss  0x5fc(%rip),%ymm1        # 6744 <_sk_callback_avx+0x4fa>
+  .byte  196,226,125,24,13,252,5,0,0         // vbroadcastss  0x5fc(%rip),%ymm1        # 6740 <_sk_callback_avx+0x4fa>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,242,5,0,0           // vbroadcastss  0x5f2(%rip),%ymm8        # 6748 <_sk_callback_avx+0x4fe>
+  .byte  196,98,125,24,5,242,5,0,0           // vbroadcastss  0x5f2(%rip),%ymm8        # 6744 <_sk_callback_avx+0x4fe>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,232,5,0,0          // vbroadcastss  0x5e8(%rip),%ymm9        # 674c <_sk_callback_avx+0x502>
+  .byte  196,98,125,24,13,232,5,0,0          // vbroadcastss  0x5e8(%rip),%ymm9        # 6748 <_sk_callback_avx+0x502>
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,222,5,0,0          // vbroadcastss  0x5de(%rip),%ymm10        # 6750 <_sk_callback_avx+0x506>
+  .byte  196,98,125,24,21,222,5,0,0          // vbroadcastss  0x5de(%rip),%ymm10        # 674c <_sk_callback_avx+0x506>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,207,5,0,0          // vbroadcastss  0x5cf(%rip),%ymm10        # 6754 <_sk_callback_avx+0x50a>
+  .byte  196,98,125,24,21,207,5,0,0          // vbroadcastss  0x5cf(%rip),%ymm10        # 6750 <_sk_callback_avx+0x50a>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,192,5,0,0          // vbroadcastss  0x5c0(%rip),%ymm9        # 6758 <_sk_callback_avx+0x50e>
+  .byte  196,98,125,24,13,192,5,0,0          // vbroadcastss  0x5c0(%rip),%ymm9        # 6754 <_sk_callback_avx+0x50e>
   .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
@@ -19740,17 +19734,17 @@ HIDDEN _sk_bicubic_p1y_avx
 FUNCTION(_sk_bicubic_p1y_avx)
 _sk_bicubic_p1y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,168,5,0,0           // vbroadcastss  0x5a8(%rip),%ymm8        # 675c <_sk_callback_avx+0x512>
+  .byte  196,98,125,24,5,168,5,0,0           // vbroadcastss  0x5a8(%rip),%ymm8        # 6758 <_sk_callback_avx+0x512>
   .byte  197,188,88,72,32                    // vaddps        0x20(%rax),%ymm8,%ymm1
   .byte  197,124,16,72,96                    // vmovups       0x60(%rax),%ymm9
-  .byte  196,98,125,24,21,153,5,0,0          // vbroadcastss  0x599(%rip),%ymm10        # 6760 <_sk_callback_avx+0x516>
+  .byte  196,98,125,24,21,153,5,0,0          // vbroadcastss  0x599(%rip),%ymm10        # 675c <_sk_callback_avx+0x516>
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
-  .byte  196,98,125,24,29,143,5,0,0          // vbroadcastss  0x58f(%rip),%ymm11        # 6764 <_sk_callback_avx+0x51a>
+  .byte  196,98,125,24,29,143,5,0,0          // vbroadcastss  0x58f(%rip),%ymm11        # 6760 <_sk_callback_avx+0x51a>
   .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  196,98,125,24,13,118,5,0,0          // vbroadcastss  0x576(%rip),%ymm9        # 6768 <_sk_callback_avx+0x51e>
+  .byte  196,98,125,24,13,118,5,0,0          // vbroadcastss  0x576(%rip),%ymm9        # 6764 <_sk_callback_avx+0x51e>
   .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
@@ -19761,13 +19755,13 @@ HIDDEN _sk_bicubic_p3y_avx
 FUNCTION(_sk_bicubic_p3y_avx)
 _sk_bicubic_p3y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,94,5,0,0          // vbroadcastss  0x55e(%rip),%ymm1        # 676c <_sk_callback_avx+0x522>
+  .byte  196,226,125,24,13,94,5,0,0          // vbroadcastss  0x55e(%rip),%ymm1        # 6768 <_sk_callback_avx+0x522>
   .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  196,98,125,24,21,74,5,0,0           // vbroadcastss  0x54a(%rip),%ymm10        # 6770 <_sk_callback_avx+0x526>
+  .byte  196,98,125,24,21,74,5,0,0           // vbroadcastss  0x54a(%rip),%ymm10        # 676c <_sk_callback_avx+0x526>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,64,5,0,0           // vbroadcastss  0x540(%rip),%ymm10        # 6774 <_sk_callback_avx+0x52a>
+  .byte  196,98,125,24,21,64,5,0,0           // vbroadcastss  0x540(%rip),%ymm10        # 6770 <_sk_callback_avx+0x52a>
   .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)
@@ -19891,25 +19885,25 @@ 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 6421 <.literal4+0xb1>
+  .byte  71,225,61                           // rex.RXB       loope 641d <.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 6431 <.literal4+0xc1>
+  .byte  71,225,61                           // rex.RXB       loope 642d <.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 6441 <.literal4+0xd1>
+  .byte  71,225,61                           // rex.RXB       loope 643d <.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 6451 <.literal4+0xe1>
+  .byte  71,225,61                           // rex.RXB       loope 644d <.literal4+0xe1>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -19959,7 +19953,7 @@ 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        649d <.literal4+0x12d>
+  .byte  224,7                               // loopne        6499 <.literal4+0x12d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -19975,10 +19969,10 @@ 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            64c4 <.literal4+0x154>
+  .byte  127,0                               // jg            64c0 <.literal4+0x154>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            653d <.literal4+0x1cd>
+  .byte  119,115                             // ja            6539 <.literal4+0x1cd>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -19992,10 +19986,10 @@ 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            64f8 <.literal4+0x188>
+  .byte  127,0                               // jg            64f4 <.literal4+0x188>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            6571 <.literal4+0x201>
+  .byte  119,115                             // ja            656d <.literal4+0x201>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20009,10 +20003,10 @@ 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            652c <.literal4+0x1bc>
+  .byte  127,0                               // jg            6528 <.literal4+0x1bc>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            65a5 <.literal4+0x235>
+  .byte  119,115                             // ja            65a1 <.literal4+0x235>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20026,10 +20020,10 @@ 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            6560 <.literal4+0x1f0>
+  .byte  127,0                               // jg            655c <.literal4+0x1f0>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            65d9 <.literal4+0x269>
+  .byte  119,115                             // ja            65d5 <.literal4+0x269>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20042,7 +20036,7 @@ 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            65d7 <.literal4+0x267>
+  .byte  127,67                              // jg            65d3 <.literal4+0x267>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,195                               // add           %al,%bl
   .byte  0,0                                 // add           %al,(%rax)
@@ -20054,10 +20048,10 @@ BALIGN4
   .byte  190,80,128,3,62                     // mov           $0x3e038050,%esi
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           65f7 <.literal4+0x287>
+  .byte  118,63                              // jbe           65f3 <.literal4+0x287>
   .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            660b <.literal4+0x29b>
+  .byte  127,67                              // jg            6607 <.literal4+0x29b>
   .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)
@@ -20066,7 +20060,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        65ed <.literal4+0x27d>
+  .byte  224,7                               // loopne        65e9 <.literal4+0x27d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -20078,7 +20072,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        6609 <.literal4+0x299>
+  .byte  224,7                               // loopne        6605 <.literal4+0x299>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -20089,7 +20083,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            665e <.literal4+0x2ee>
+  .byte  124,66                              // jl            665a <.literal4+0x2ee>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  137,136,136,55,0,15                 // mov           %ecx,0xf003788(%rax)
@@ -20107,9 +20101,9 @@ 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            66a1 <.literal4+0x331>
+  .byte  112,65                              // jo            669d <.literal4+0x331>
   .byte  129,128,128,59,129,128,128,59,0,0   // addl          $0x3b80,-0x7f7ec480(%rax)
-  .byte  127,67                              // jg            66af <.literal4+0x33f>
+  .byte  127,67                              // jg            66ab <.literal4+0x33f>
   .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)
@@ -20125,7 +20119,7 @@ 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            66ef <.literal4+0x37f>
+  .byte  127,71                              // jg            66eb <.literal4+0x37f>
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,89                               // ds            pop %rcx
@@ -20355,7 +20349,7 @@ _sk_seed_shader_sse41:
   .byte  102,15,110,199                      // movd          %edi,%xmm0
   .byte  102,15,112,192,0                    // pshufd        $0x0,%xmm0,%xmm0
   .byte  15,91,200                           // cvtdq2ps      %xmm0,%xmm1
-  .byte  15,40,21,116,70,0,0                 // movaps        0x4674(%rip),%xmm2        # 46f0 <_sk_callback_sse41+0xe0>
+  .byte  15,40,21,116,70,0,0                 // movaps        0x4674(%rip),%xmm2        # 46f0 <_sk_callback_sse41+0xe3>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  15,16,2                             // movups        (%rdx),%xmm0
   .byte  15,88,193                           // addps         %xmm1,%xmm0
@@ -20364,7 +20358,7 @@ _sk_seed_shader_sse41:
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,21,99,70,0,0                  // movaps        0x4663(%rip),%xmm2        # 4700 <_sk_callback_sse41+0xf0>
+  .byte  15,40,21,99,70,0,0                  // movaps        0x4663(%rip),%xmm2        # 4700 <_sk_callback_sse41+0xf3>
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
@@ -20387,14 +20381,14 @@ _sk_dither_sse41:
   .byte  102,68,15,110,1                     // movd          (%rcx),%xmm8
   .byte  102,69,15,112,192,0                 // pshufd        $0x0,%xmm8,%xmm8
   .byte  102,69,15,239,193                   // pxor          %xmm9,%xmm8
-  .byte  102,68,15,111,21,40,70,0,0          // movdqa        0x4628(%rip),%xmm10        # 4710 <_sk_callback_sse41+0x100>
+  .byte  102,68,15,111,21,40,70,0,0          // movdqa        0x4628(%rip),%xmm10        # 4710 <_sk_callback_sse41+0x103>
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,219,218                   // pand          %xmm10,%xmm11
   .byte  102,65,15,114,243,5                 // pslld         $0x5,%xmm11
   .byte  102,69,15,219,209                   // pand          %xmm9,%xmm10
   .byte  102,65,15,114,242,4                 // pslld         $0x4,%xmm10
-  .byte  102,68,15,111,37,20,70,0,0          // movdqa        0x4614(%rip),%xmm12        # 4720 <_sk_callback_sse41+0x110>
-  .byte  102,68,15,111,45,27,70,0,0          // movdqa        0x461b(%rip),%xmm13        # 4730 <_sk_callback_sse41+0x120>
+  .byte  102,68,15,111,37,20,70,0,0          // movdqa        0x4614(%rip),%xmm12        # 4720 <_sk_callback_sse41+0x113>
+  .byte  102,68,15,111,45,27,70,0,0          // movdqa        0x461b(%rip),%xmm13        # 4730 <_sk_callback_sse41+0x123>
   .byte  102,69,15,111,240                   // movdqa        %xmm8,%xmm14
   .byte  102,69,15,219,245                   // pand          %xmm13,%xmm14
   .byte  102,65,15,114,246,2                 // pslld         $0x2,%xmm14
@@ -20410,8 +20404,8 @@ _sk_dither_sse41:
   .byte  102,69,15,235,245                   // por           %xmm13,%xmm14
   .byte  102,69,15,235,240                   // por           %xmm8,%xmm14
   .byte  69,15,91,198                        // cvtdq2ps      %xmm14,%xmm8
-  .byte  68,15,89,5,214,69,0,0               // mulps         0x45d6(%rip),%xmm8        # 4740 <_sk_callback_sse41+0x130>
-  .byte  68,15,88,5,222,69,0,0               // addps         0x45de(%rip),%xmm8        # 4750 <_sk_callback_sse41+0x140>
+  .byte  68,15,89,5,214,69,0,0               // mulps         0x45d6(%rip),%xmm8        # 4740 <_sk_callback_sse41+0x133>
+  .byte  68,15,88,5,222,69,0,0               // addps         0x45de(%rip),%xmm8        # 4750 <_sk_callback_sse41+0x143>
   .byte  243,68,15,16,72,8                   // movss         0x8(%rax),%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
@@ -20477,7 +20471,7 @@ HIDDEN _sk_srcatop_sse41
 FUNCTION(_sk_srcatop_sse41)
 _sk_srcatop_sse41:
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  68,15,40,5,97,69,0,0                // movaps        0x4561(%rip),%xmm8        # 4760 <_sk_callback_sse41+0x150>
+  .byte  68,15,40,5,97,69,0,0                // movaps        0x4561(%rip),%xmm8        # 4760 <_sk_callback_sse41+0x153>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -20502,7 +20496,7 @@ FUNCTION(_sk_dstatop_sse41)
 _sk_dstatop_sse41:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
   .byte  68,15,89,196                        // mulps         %xmm4,%xmm8
-  .byte  68,15,40,13,36,69,0,0               // movaps        0x4524(%rip),%xmm9        # 4770 <_sk_callback_sse41+0x160>
+  .byte  68,15,40,13,36,69,0,0               // movaps        0x4524(%rip),%xmm9        # 4770 <_sk_callback_sse41+0x163>
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
@@ -20549,7 +20543,7 @@ HIDDEN _sk_srcout_sse41
 .globl _sk_srcout_sse41
 FUNCTION(_sk_srcout_sse41)
 _sk_srcout_sse41:
-  .byte  68,15,40,5,200,68,0,0               // movaps        0x44c8(%rip),%xmm8        # 4780 <_sk_callback_sse41+0x170>
+  .byte  68,15,40,5,200,68,0,0               // movaps        0x44c8(%rip),%xmm8        # 4780 <_sk_callback_sse41+0x173>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
@@ -20562,7 +20556,7 @@ HIDDEN _sk_dstout_sse41
 .globl _sk_dstout_sse41
 FUNCTION(_sk_dstout_sse41)
 _sk_dstout_sse41:
-  .byte  68,15,40,5,184,68,0,0               // movaps        0x44b8(%rip),%xmm8        # 4790 <_sk_callback_sse41+0x180>
+  .byte  68,15,40,5,184,68,0,0               // movaps        0x44b8(%rip),%xmm8        # 4790 <_sk_callback_sse41+0x183>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -20579,7 +20573,7 @@ HIDDEN _sk_srcover_sse41
 .globl _sk_srcover_sse41
 FUNCTION(_sk_srcover_sse41)
 _sk_srcover_sse41:
-  .byte  68,15,40,5,155,68,0,0               // movaps        0x449b(%rip),%xmm8        # 47a0 <_sk_callback_sse41+0x190>
+  .byte  68,15,40,5,155,68,0,0               // movaps        0x449b(%rip),%xmm8        # 47a0 <_sk_callback_sse41+0x193>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -20599,7 +20593,7 @@ HIDDEN _sk_dstover_sse41
 .globl _sk_dstover_sse41
 FUNCTION(_sk_dstover_sse41)
 _sk_dstover_sse41:
-  .byte  68,15,40,5,111,68,0,0               // movaps        0x446f(%rip),%xmm8        # 47b0 <_sk_callback_sse41+0x1a0>
+  .byte  68,15,40,5,111,68,0,0               // movaps        0x446f(%rip),%xmm8        # 47b0 <_sk_callback_sse41+0x1a3>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -20627,7 +20621,7 @@ HIDDEN _sk_multiply_sse41
 .globl _sk_multiply_sse41
 FUNCTION(_sk_multiply_sse41)
 _sk_multiply_sse41:
-  .byte  68,15,40,5,67,68,0,0                // movaps        0x4443(%rip),%xmm8        # 47c0 <_sk_callback_sse41+0x1b0>
+  .byte  68,15,40,5,67,68,0,0                // movaps        0x4443(%rip),%xmm8        # 47c0 <_sk_callback_sse41+0x1b3>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
@@ -20703,7 +20697,7 @@ HIDDEN _sk_xor__sse41
 FUNCTION(_sk_xor__sse41)
 _sk_xor__sse41:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
-  .byte  15,40,29,116,67,0,0                 // movaps        0x4374(%rip),%xmm3        # 47d0 <_sk_callback_sse41+0x1c0>
+  .byte  15,40,29,116,67,0,0                 // movaps        0x4374(%rip),%xmm3        # 47d0 <_sk_callback_sse41+0x1c3>
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
@@ -20751,7 +20745,7 @@ _sk_darken_sse41:
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,95,209                        // maxps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,223,66,0,0                 // movaps        0x42df(%rip),%xmm2        # 47e0 <_sk_callback_sse41+0x1d0>
+  .byte  15,40,21,223,66,0,0                 // movaps        0x42df(%rip),%xmm2        # 47e0 <_sk_callback_sse41+0x1d3>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -20785,7 +20779,7 @@ _sk_lighten_sse41:
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,132,66,0,0                 // movaps        0x4284(%rip),%xmm2        # 47f0 <_sk_callback_sse41+0x1e0>
+  .byte  15,40,21,132,66,0,0                 // movaps        0x4284(%rip),%xmm2        # 47f0 <_sk_callback_sse41+0x1e3>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -20822,7 +20816,7 @@ _sk_difference_sse41:
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,30,66,0,0                  // movaps        0x421e(%rip),%xmm2        # 4800 <_sk_callback_sse41+0x1f0>
+  .byte  15,40,21,30,66,0,0                  // movaps        0x421e(%rip),%xmm2        # 4800 <_sk_callback_sse41+0x1f3>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -20849,7 +20843,7 @@ _sk_exclusion_sse41:
   .byte  15,89,214                           // mulps         %xmm6,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,202                        // subps         %xmm2,%xmm9
-  .byte  15,40,13,223,65,0,0                 // movaps        0x41df(%rip),%xmm1        # 4810 <_sk_callback_sse41+0x200>
+  .byte  15,40,13,223,65,0,0                 // movaps        0x41df(%rip),%xmm1        # 4810 <_sk_callback_sse41+0x203>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,207                           // mulps         %xmm7,%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -20863,7 +20857,7 @@ HIDDEN _sk_colorburn_sse41
 FUNCTION(_sk_colorburn_sse41)
 _sk_colorburn_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,206,65,0,0              // movaps        0x41ce(%rip),%xmm10        # 4820 <_sk_callback_sse41+0x210>
+  .byte  68,15,40,21,206,65,0,0              // movaps        0x41ce(%rip),%xmm10        # 4820 <_sk_callback_sse41+0x213>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,203                        // movaps        %xmm11,%xmm9
@@ -20945,7 +20939,7 @@ HIDDEN _sk_colordodge_sse41
 FUNCTION(_sk_colordodge_sse41)
 _sk_colordodge_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,172,64,0,0              // movaps        0x40ac(%rip),%xmm10        # 4830 <_sk_callback_sse41+0x220>
+  .byte  68,15,40,21,172,64,0,0              // movaps        0x40ac(%rip),%xmm10        # 4830 <_sk_callback_sse41+0x223>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -21027,7 +21021,7 @@ _sk_hardlight_sse41:
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
-  .byte  68,15,40,21,133,63,0,0              // movaps        0x3f85(%rip),%xmm10        # 4840 <_sk_callback_sse41+0x230>
+  .byte  68,15,40,21,133,63,0,0              // movaps        0x3f85(%rip),%xmm10        # 4840 <_sk_callback_sse41+0x233>
   .byte  65,15,40,234                        // movaps        %xmm10,%xmm5
   .byte  15,92,239                           // subps         %xmm7,%xmm5
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
@@ -21110,7 +21104,7 @@ FUNCTION(_sk_overlay_sse41)
 _sk_overlay_sse41:
   .byte  68,15,40,201                        // movaps        %xmm1,%xmm9
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
-  .byte  68,15,40,21,106,62,0,0              // movaps        0x3e6a(%rip),%xmm10        # 4850 <_sk_callback_sse41+0x240>
+  .byte  68,15,40,21,106,62,0,0              // movaps        0x3e6a(%rip),%xmm10        # 4850 <_sk_callback_sse41+0x243>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
@@ -21195,7 +21189,7 @@ _sk_softlight_sse41:
   .byte  15,40,198                           // movaps        %xmm6,%xmm0
   .byte  15,94,199                           // divps         %xmm7,%xmm0
   .byte  65,15,84,193                        // andps         %xmm9,%xmm0
-  .byte  15,40,13,65,61,0,0                  // movaps        0x3d41(%rip),%xmm1        # 4860 <_sk_callback_sse41+0x250>
+  .byte  15,40,13,65,61,0,0                  // movaps        0x3d41(%rip),%xmm1        # 4860 <_sk_callback_sse41+0x253>
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  68,15,92,208                        // subps         %xmm0,%xmm10
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
@@ -21208,10 +21202,10 @@ _sk_softlight_sse41:
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,89,210                           // mulps         %xmm2,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
-  .byte  68,15,40,45,31,61,0,0               // movaps        0x3d1f(%rip),%xmm13        # 4870 <_sk_callback_sse41+0x260>
+  .byte  68,15,40,45,31,61,0,0               // movaps        0x3d1f(%rip),%xmm13        # 4870 <_sk_callback_sse41+0x263>
   .byte  69,15,88,245                        // addps         %xmm13,%xmm14
   .byte  68,15,89,242                        // mulps         %xmm2,%xmm14
-  .byte  68,15,40,37,31,61,0,0               // movaps        0x3d1f(%rip),%xmm12        # 4880 <_sk_callback_sse41+0x270>
+  .byte  68,15,40,37,31,61,0,0               // movaps        0x3d1f(%rip),%xmm12        # 4880 <_sk_callback_sse41+0x273>
   .byte  69,15,89,252                        // mulps         %xmm12,%xmm15
   .byte  69,15,88,254                        // addps         %xmm14,%xmm15
   .byte  15,40,198                           // movaps        %xmm6,%xmm0
@@ -21397,12 +21391,12 @@ _sk_hue_sse41:
   .byte  68,15,84,208                        // andps         %xmm0,%xmm10
   .byte  15,84,200                           // andps         %xmm0,%xmm1
   .byte  68,15,84,232                        // andps         %xmm0,%xmm13
-  .byte  15,40,5,138,58,0,0                  // movaps        0x3a8a(%rip),%xmm0        # 4890 <_sk_callback_sse41+0x280>
+  .byte  15,40,5,138,58,0,0                  // movaps        0x3a8a(%rip),%xmm0        # 4890 <_sk_callback_sse41+0x283>
   .byte  68,15,89,224                        // mulps         %xmm0,%xmm12
-  .byte  15,40,21,143,58,0,0                 // movaps        0x3a8f(%rip),%xmm2        # 48a0 <_sk_callback_sse41+0x290>
+  .byte  15,40,21,143,58,0,0                 // movaps        0x3a8f(%rip),%xmm2        # 48a0 <_sk_callback_sse41+0x293>
   .byte  15,89,250                           // mulps         %xmm2,%xmm7
   .byte  65,15,88,252                        // addps         %xmm12,%xmm7
-  .byte  68,15,40,53,144,58,0,0              // movaps        0x3a90(%rip),%xmm14        # 48b0 <_sk_callback_sse41+0x2a0>
+  .byte  68,15,40,53,144,58,0,0              // movaps        0x3a90(%rip),%xmm14        # 48b0 <_sk_callback_sse41+0x2a3>
   .byte  68,15,40,252                        // movaps        %xmm4,%xmm15
   .byte  69,15,89,254                        // mulps         %xmm14,%xmm15
   .byte  68,15,88,255                        // addps         %xmm7,%xmm15
@@ -21485,7 +21479,7 @@ _sk_hue_sse41:
   .byte  65,15,88,214                        // addps         %xmm14,%xmm2
   .byte  15,40,196                           // movaps        %xmm4,%xmm0
   .byte  102,15,56,20,202                    // blendvps      %xmm0,%xmm2,%xmm1
-  .byte  68,15,40,13,84,57,0,0               // movaps        0x3954(%rip),%xmm9        # 48c0 <_sk_callback_sse41+0x2b0>
+  .byte  68,15,40,13,84,57,0,0               // movaps        0x3954(%rip),%xmm9        # 48c0 <_sk_callback_sse41+0x2b3>
   .byte  65,15,40,225                        // movaps        %xmm9,%xmm4
   .byte  15,92,229                           // subps         %xmm5,%xmm4
   .byte  15,40,68,36,200                     // movaps        -0x38(%rsp),%xmm0
@@ -21579,14 +21573,14 @@ _sk_saturation_sse41:
   .byte  68,15,84,215                        // andps         %xmm7,%xmm10
   .byte  68,15,84,223                        // andps         %xmm7,%xmm11
   .byte  68,15,84,199                        // andps         %xmm7,%xmm8
-  .byte  15,40,21,14,56,0,0                  // movaps        0x380e(%rip),%xmm2        # 48d0 <_sk_callback_sse41+0x2c0>
+  .byte  15,40,21,14,56,0,0                  // movaps        0x380e(%rip),%xmm2        # 48d0 <_sk_callback_sse41+0x2c3>
   .byte  15,40,221                           // movaps        %xmm5,%xmm3
   .byte  15,89,218                           // mulps         %xmm2,%xmm3
-  .byte  15,40,13,17,56,0,0                  // movaps        0x3811(%rip),%xmm1        # 48e0 <_sk_callback_sse41+0x2d0>
+  .byte  15,40,13,17,56,0,0                  // movaps        0x3811(%rip),%xmm1        # 48e0 <_sk_callback_sse41+0x2d3>
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
   .byte  15,89,249                           // mulps         %xmm1,%xmm7
   .byte  15,88,251                           // addps         %xmm3,%xmm7
-  .byte  68,15,40,45,16,56,0,0               // movaps        0x3810(%rip),%xmm13        # 48f0 <_sk_callback_sse41+0x2e0>
+  .byte  68,15,40,45,16,56,0,0               // movaps        0x3810(%rip),%xmm13        # 48f0 <_sk_callback_sse41+0x2e3>
   .byte  69,15,89,245                        // mulps         %xmm13,%xmm14
   .byte  68,15,88,247                        // addps         %xmm7,%xmm14
   .byte  65,15,40,218                        // movaps        %xmm10,%xmm3
@@ -21667,7 +21661,7 @@ _sk_saturation_sse41:
   .byte  65,15,88,253                        // addps         %xmm13,%xmm7
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  102,68,15,56,20,223                 // blendvps      %xmm0,%xmm7,%xmm11
-  .byte  68,15,40,13,214,54,0,0              // movaps        0x36d6(%rip),%xmm9        # 4900 <_sk_callback_sse41+0x2f0>
+  .byte  68,15,40,13,214,54,0,0              // movaps        0x36d6(%rip),%xmm9        # 4900 <_sk_callback_sse41+0x2f3>
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  68,15,92,204                        // subps         %xmm4,%xmm9
   .byte  15,40,124,36,168                    // movaps        -0x58(%rsp),%xmm7
@@ -21722,14 +21716,14 @@ _sk_color_sse41:
   .byte  15,40,231                           // movaps        %xmm7,%xmm4
   .byte  68,15,89,244                        // mulps         %xmm4,%xmm14
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
-  .byte  68,15,40,13,33,54,0,0               // movaps        0x3621(%rip),%xmm9        # 4910 <_sk_callback_sse41+0x300>
+  .byte  68,15,40,13,33,54,0,0               // movaps        0x3621(%rip),%xmm9        # 4910 <_sk_callback_sse41+0x303>
   .byte  65,15,40,250                        // movaps        %xmm10,%xmm7
   .byte  65,15,89,249                        // mulps         %xmm9,%xmm7
-  .byte  68,15,40,21,33,54,0,0               // movaps        0x3621(%rip),%xmm10        # 4920 <_sk_callback_sse41+0x310>
+  .byte  68,15,40,21,33,54,0,0               // movaps        0x3621(%rip),%xmm10        # 4920 <_sk_callback_sse41+0x313>
   .byte  65,15,40,219                        // movaps        %xmm11,%xmm3
   .byte  65,15,89,218                        // mulps         %xmm10,%xmm3
   .byte  15,88,223                           // addps         %xmm7,%xmm3
-  .byte  68,15,40,29,30,54,0,0               // movaps        0x361e(%rip),%xmm11        # 4930 <_sk_callback_sse41+0x320>
+  .byte  68,15,40,29,30,54,0,0               // movaps        0x361e(%rip),%xmm11        # 4930 <_sk_callback_sse41+0x323>
   .byte  69,15,40,236                        // movaps        %xmm12,%xmm13
   .byte  69,15,89,235                        // mulps         %xmm11,%xmm13
   .byte  68,15,88,235                        // addps         %xmm3,%xmm13
@@ -21814,7 +21808,7 @@ _sk_color_sse41:
   .byte  65,15,88,251                        // addps         %xmm11,%xmm7
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  102,15,56,20,207                    // blendvps      %xmm0,%xmm7,%xmm1
-  .byte  68,15,40,13,218,52,0,0              // movaps        0x34da(%rip),%xmm9        # 4940 <_sk_callback_sse41+0x330>
+  .byte  68,15,40,13,218,52,0,0              // movaps        0x34da(%rip),%xmm9        # 4940 <_sk_callback_sse41+0x333>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  68,15,89,192                        // mulps         %xmm0,%xmm8
@@ -21866,13 +21860,13 @@ _sk_luminosity_sse41:
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,89,205                        // mulps         %xmm5,%xmm9
-  .byte  68,15,40,5,50,52,0,0                // movaps        0x3432(%rip),%xmm8        # 4950 <_sk_callback_sse41+0x340>
+  .byte  68,15,40,5,50,52,0,0                // movaps        0x3432(%rip),%xmm8        # 4950 <_sk_callback_sse41+0x343>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
-  .byte  68,15,40,21,54,52,0,0               // movaps        0x3436(%rip),%xmm10        # 4960 <_sk_callback_sse41+0x350>
+  .byte  68,15,40,21,54,52,0,0               // movaps        0x3436(%rip),%xmm10        # 4960 <_sk_callback_sse41+0x353>
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
   .byte  65,15,89,234                        // mulps         %xmm10,%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
-  .byte  68,15,40,37,52,52,0,0               // movaps        0x3434(%rip),%xmm12        # 4970 <_sk_callback_sse41+0x360>
+  .byte  68,15,40,37,52,52,0,0               // movaps        0x3434(%rip),%xmm12        # 4970 <_sk_callback_sse41+0x363>
   .byte  68,15,40,242                        // movaps        %xmm2,%xmm14
   .byte  69,15,89,244                        // mulps         %xmm12,%xmm14
   .byte  68,15,88,245                        // addps         %xmm5,%xmm14
@@ -21957,7 +21951,7 @@ _sk_luminosity_sse41:
   .byte  65,15,88,244                        // addps         %xmm12,%xmm6
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  102,68,15,56,20,206                 // blendvps      %xmm0,%xmm6,%xmm9
-  .byte  15,40,5,234,50,0,0                  // movaps        0x32ea(%rip),%xmm0        # 4980 <_sk_callback_sse41+0x370>
+  .byte  15,40,5,234,50,0,0                  // movaps        0x32ea(%rip),%xmm0        # 4980 <_sk_callback_sse41+0x373>
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,92,215                           // subps         %xmm7,%xmm2
   .byte  15,89,226                           // mulps         %xmm2,%xmm4
@@ -22006,7 +22000,7 @@ HIDDEN _sk_clamp_1_sse41
 .globl _sk_clamp_1_sse41
 FUNCTION(_sk_clamp_1_sse41)
 _sk_clamp_1_sse41:
-  .byte  68,15,40,5,109,50,0,0               // movaps        0x326d(%rip),%xmm8        # 4990 <_sk_callback_sse41+0x380>
+  .byte  68,15,40,5,109,50,0,0               // movaps        0x326d(%rip),%xmm8        # 4990 <_sk_callback_sse41+0x383>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
   .byte  65,15,93,208                        // minps         %xmm8,%xmm2
@@ -22018,7 +22012,7 @@ HIDDEN _sk_clamp_a_sse41
 .globl _sk_clamp_a_sse41
 FUNCTION(_sk_clamp_a_sse41)
 _sk_clamp_a_sse41:
-  .byte  15,93,29,98,50,0,0                  // minps         0x3262(%rip),%xmm3        # 49a0 <_sk_callback_sse41+0x390>
+  .byte  15,93,29,98,50,0,0                  // minps         0x3262(%rip),%xmm3        # 49a0 <_sk_callback_sse41+0x393>
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -22105,7 +22099,7 @@ HIDDEN _sk_unpremul_sse41
 FUNCTION(_sk_unpremul_sse41)
 _sk_unpremul_sse41:
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
-  .byte  68,15,40,13,205,49,0,0              // movaps        0x31cd(%rip),%xmm9        # 49b0 <_sk_callback_sse41+0x3a0>
+  .byte  68,15,40,13,205,49,0,0              // movaps        0x31cd(%rip),%xmm9        # 49b0 <_sk_callback_sse41+0x3a3>
   .byte  68,15,94,203                        // divps         %xmm3,%xmm9
   .byte  68,15,194,195,4                     // cmpneqps      %xmm3,%xmm8
   .byte  69,15,84,193                        // andps         %xmm9,%xmm8
@@ -22119,20 +22113,20 @@ HIDDEN _sk_from_srgb_sse41
 .globl _sk_from_srgb_sse41
 FUNCTION(_sk_from_srgb_sse41)
 _sk_from_srgb_sse41:
-  .byte  68,15,40,29,184,49,0,0              // movaps        0x31b8(%rip),%xmm11        # 49c0 <_sk_callback_sse41+0x3b0>
+  .byte  68,15,40,29,184,49,0,0              // movaps        0x31b8(%rip),%xmm11        # 49c0 <_sk_callback_sse41+0x3b3>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
   .byte  68,15,40,208                        // movaps        %xmm0,%xmm10
   .byte  69,15,89,210                        // mulps         %xmm10,%xmm10
-  .byte  68,15,40,37,176,49,0,0              // movaps        0x31b0(%rip),%xmm12        # 49d0 <_sk_callback_sse41+0x3c0>
+  .byte  68,15,40,37,176,49,0,0              // movaps        0x31b0(%rip),%xmm12        # 49d0 <_sk_callback_sse41+0x3c3>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,196                        // mulps         %xmm12,%xmm8
-  .byte  68,15,40,45,176,49,0,0              // movaps        0x31b0(%rip),%xmm13        # 49e0 <_sk_callback_sse41+0x3d0>
+  .byte  68,15,40,45,176,49,0,0              // movaps        0x31b0(%rip),%xmm13        # 49e0 <_sk_callback_sse41+0x3d3>
   .byte  69,15,88,197                        // addps         %xmm13,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
-  .byte  68,15,40,53,176,49,0,0              // movaps        0x31b0(%rip),%xmm14        # 49f0 <_sk_callback_sse41+0x3e0>
+  .byte  68,15,40,53,176,49,0,0              // movaps        0x31b0(%rip),%xmm14        # 49f0 <_sk_callback_sse41+0x3e3>
   .byte  69,15,88,198                        // addps         %xmm14,%xmm8
-  .byte  68,15,40,61,180,49,0,0              // movaps        0x31b4(%rip),%xmm15        # 4a00 <_sk_callback_sse41+0x3f0>
+  .byte  68,15,40,61,180,49,0,0              // movaps        0x31b4(%rip),%xmm15        # 4a00 <_sk_callback_sse41+0x3f3>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  102,69,15,56,20,193                 // blendvps      %xmm0,%xmm9,%xmm8
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
@@ -22177,20 +22171,20 @@ _sk_to_srgb_sse41:
   .byte  68,15,82,192                        // rsqrtps       %xmm0,%xmm8
   .byte  69,15,83,200                        // rcpps         %xmm8,%xmm9
   .byte  69,15,82,208                        // rsqrtps       %xmm8,%xmm10
-  .byte  68,15,40,29,36,49,0,0               // movaps        0x3124(%rip),%xmm11        # 4a10 <_sk_callback_sse41+0x400>
+  .byte  68,15,40,29,36,49,0,0               // movaps        0x3124(%rip),%xmm11        # 4a10 <_sk_callback_sse41+0x403>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
-  .byte  68,15,40,37,37,49,0,0               // movaps        0x3125(%rip),%xmm12        # 4a20 <_sk_callback_sse41+0x410>
+  .byte  68,15,40,37,37,49,0,0               // movaps        0x3125(%rip),%xmm12        # 4a20 <_sk_callback_sse41+0x413>
   .byte  69,15,89,204                        // mulps         %xmm12,%xmm9
-  .byte  68,15,40,45,41,49,0,0               // movaps        0x3129(%rip),%xmm13        # 4a30 <_sk_callback_sse41+0x420>
+  .byte  68,15,40,45,41,49,0,0               // movaps        0x3129(%rip),%xmm13        # 4a30 <_sk_callback_sse41+0x423>
   .byte  69,15,88,205                        // addps         %xmm13,%xmm9
-  .byte  68,15,40,53,45,49,0,0               // movaps        0x312d(%rip),%xmm14        # 4a40 <_sk_callback_sse41+0x430>
+  .byte  68,15,40,53,45,49,0,0               // movaps        0x312d(%rip),%xmm14        # 4a40 <_sk_callback_sse41+0x433>
   .byte  69,15,89,214                        // mulps         %xmm14,%xmm10
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
-  .byte  68,15,40,5,45,49,0,0                // movaps        0x312d(%rip),%xmm8        # 4a50 <_sk_callback_sse41+0x440>
+  .byte  68,15,40,5,45,49,0,0                // movaps        0x312d(%rip),%xmm8        # 4a50 <_sk_callback_sse41+0x443>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,93,202                        // minps         %xmm10,%xmm9
-  .byte  68,15,40,61,45,49,0,0               // movaps        0x312d(%rip),%xmm15        # 4a60 <_sk_callback_sse41+0x450>
+  .byte  68,15,40,61,45,49,0,0               // movaps        0x312d(%rip),%xmm15        # 4a60 <_sk_callback_sse41+0x453>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  102,68,15,56,20,201                 // blendvps      %xmm0,%xmm1,%xmm9
   .byte  15,82,194                           // rsqrtps       %xmm2,%xmm0
@@ -22244,7 +22238,7 @@ _sk_rgb_to_hsl_sse41:
   .byte  68,15,93,226                        // minps         %xmm2,%xmm12
   .byte  65,15,40,203                        // movaps        %xmm11,%xmm1
   .byte  65,15,92,204                        // subps         %xmm12,%xmm1
-  .byte  68,15,40,53,126,48,0,0              // movaps        0x307e(%rip),%xmm14        # 4a70 <_sk_callback_sse41+0x460>
+  .byte  68,15,40,53,126,48,0,0              // movaps        0x307e(%rip),%xmm14        # 4a70 <_sk_callback_sse41+0x463>
   .byte  68,15,94,241                        // divps         %xmm1,%xmm14
   .byte  69,15,40,211                        // movaps        %xmm11,%xmm10
   .byte  69,15,194,208,0                     // cmpeqps       %xmm8,%xmm10
@@ -22253,27 +22247,27 @@ _sk_rgb_to_hsl_sse41:
   .byte  65,15,89,198                        // mulps         %xmm14,%xmm0
   .byte  69,15,40,249                        // movaps        %xmm9,%xmm15
   .byte  68,15,194,250,1                     // cmpltps       %xmm2,%xmm15
-  .byte  68,15,84,61,101,48,0,0              // andps         0x3065(%rip),%xmm15        # 4a80 <_sk_callback_sse41+0x470>
+  .byte  68,15,84,61,101,48,0,0              // andps         0x3065(%rip),%xmm15        # 4a80 <_sk_callback_sse41+0x473>
   .byte  68,15,88,248                        // addps         %xmm0,%xmm15
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  65,15,194,193,0                     // cmpeqps       %xmm9,%xmm0
   .byte  65,15,92,208                        // subps         %xmm8,%xmm2
   .byte  65,15,89,214                        // mulps         %xmm14,%xmm2
-  .byte  68,15,40,45,88,48,0,0               // movaps        0x3058(%rip),%xmm13        # 4a90 <_sk_callback_sse41+0x480>
+  .byte  68,15,40,45,88,48,0,0               // movaps        0x3058(%rip),%xmm13        # 4a90 <_sk_callback_sse41+0x483>
   .byte  65,15,88,213                        // addps         %xmm13,%xmm2
   .byte  69,15,92,193                        // subps         %xmm9,%xmm8
   .byte  69,15,89,198                        // mulps         %xmm14,%xmm8
-  .byte  68,15,88,5,84,48,0,0                // addps         0x3054(%rip),%xmm8        # 4aa0 <_sk_callback_sse41+0x490>
+  .byte  68,15,88,5,84,48,0,0                // addps         0x3054(%rip),%xmm8        # 4aa0 <_sk_callback_sse41+0x493>
   .byte  102,68,15,56,20,194                 // blendvps      %xmm0,%xmm2,%xmm8
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  102,69,15,56,20,199                 // blendvps      %xmm0,%xmm15,%xmm8
-  .byte  68,15,89,5,76,48,0,0                // mulps         0x304c(%rip),%xmm8        # 4ab0 <_sk_callback_sse41+0x4a0>
+  .byte  68,15,89,5,76,48,0,0                // mulps         0x304c(%rip),%xmm8        # 4ab0 <_sk_callback_sse41+0x4a3>
   .byte  69,15,40,203                        // movaps        %xmm11,%xmm9
   .byte  69,15,194,204,4                     // cmpneqps      %xmm12,%xmm9
   .byte  69,15,84,193                        // andps         %xmm9,%xmm8
   .byte  69,15,92,235                        // subps         %xmm11,%xmm13
   .byte  69,15,88,220                        // addps         %xmm12,%xmm11
-  .byte  15,40,5,64,48,0,0                   // movaps        0x3040(%rip),%xmm0        # 4ac0 <_sk_callback_sse41+0x4b0>
+  .byte  15,40,5,64,48,0,0                   // movaps        0x3040(%rip),%xmm0        # 4ac0 <_sk_callback_sse41+0x4b3>
   .byte  65,15,40,211                        // movaps        %xmm11,%xmm2
   .byte  15,89,208                           // mulps         %xmm0,%xmm2
   .byte  15,194,194,1                        // cmpltps       %xmm2,%xmm0
@@ -22295,7 +22289,7 @@ _sk_hsl_to_rgb_sse41:
   .byte  15,41,100,36,184                    // movaps        %xmm4,-0x48(%rsp)
   .byte  15,41,92,36,168                     // movaps        %xmm3,-0x58(%rsp)
   .byte  68,15,40,208                        // movaps        %xmm0,%xmm10
-  .byte  68,15,40,13,6,48,0,0                // movaps        0x3006(%rip),%xmm9        # 4ad0 <_sk_callback_sse41+0x4c0>
+  .byte  68,15,40,13,6,48,0,0                // movaps        0x3006(%rip),%xmm9        # 4ad0 <_sk_callback_sse41+0x4c3>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,194,194,2                        // cmpleps       %xmm2,%xmm0
   .byte  15,40,217                           // movaps        %xmm1,%xmm3
@@ -22308,19 +22302,19 @@ _sk_hsl_to_rgb_sse41:
   .byte  15,41,84,36,152                     // movaps        %xmm2,-0x68(%rsp)
   .byte  69,15,88,192                        // addps         %xmm8,%xmm8
   .byte  68,15,92,197                        // subps         %xmm5,%xmm8
-  .byte  68,15,40,53,225,47,0,0              // movaps        0x2fe1(%rip),%xmm14        # 4ae0 <_sk_callback_sse41+0x4d0>
+  .byte  68,15,40,53,225,47,0,0              // movaps        0x2fe1(%rip),%xmm14        # 4ae0 <_sk_callback_sse41+0x4d3>
   .byte  69,15,88,242                        // addps         %xmm10,%xmm14
   .byte  102,65,15,58,8,198,1                // roundps       $0x1,%xmm14,%xmm0
   .byte  68,15,92,240                        // subps         %xmm0,%xmm14
-  .byte  68,15,40,29,218,47,0,0              // movaps        0x2fda(%rip),%xmm11        # 4af0 <_sk_callback_sse41+0x4e0>
+  .byte  68,15,40,29,218,47,0,0              // movaps        0x2fda(%rip),%xmm11        # 4af0 <_sk_callback_sse41+0x4e3>
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  65,15,194,198,2                     // cmpleps       %xmm14,%xmm0
   .byte  15,40,245                           // movaps        %xmm5,%xmm6
   .byte  65,15,92,240                        // subps         %xmm8,%xmm6
-  .byte  15,40,61,211,47,0,0                 // movaps        0x2fd3(%rip),%xmm7        # 4b00 <_sk_callback_sse41+0x4f0>
+  .byte  15,40,61,211,47,0,0                 // movaps        0x2fd3(%rip),%xmm7        # 4b00 <_sk_callback_sse41+0x4f3>
   .byte  69,15,40,238                        // movaps        %xmm14,%xmm13
   .byte  68,15,89,239                        // mulps         %xmm7,%xmm13
-  .byte  15,40,29,212,47,0,0                 // movaps        0x2fd4(%rip),%xmm3        # 4b10 <_sk_callback_sse41+0x500>
+  .byte  15,40,29,212,47,0,0                 // movaps        0x2fd4(%rip),%xmm3        # 4b10 <_sk_callback_sse41+0x503>
   .byte  68,15,40,227                        // movaps        %xmm3,%xmm12
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  68,15,89,230                        // mulps         %xmm6,%xmm12
@@ -22330,7 +22324,7 @@ _sk_hsl_to_rgb_sse41:
   .byte  65,15,194,198,2                     // cmpleps       %xmm14,%xmm0
   .byte  68,15,40,253                        // movaps        %xmm5,%xmm15
   .byte  102,69,15,56,20,252                 // blendvps      %xmm0,%xmm12,%xmm15
-  .byte  68,15,40,37,179,47,0,0              // movaps        0x2fb3(%rip),%xmm12        # 4b20 <_sk_callback_sse41+0x510>
+  .byte  68,15,40,37,179,47,0,0              // movaps        0x2fb3(%rip),%xmm12        # 4b20 <_sk_callback_sse41+0x513>
   .byte  65,15,40,196                        // movaps        %xmm12,%xmm0
   .byte  65,15,194,198,2                     // cmpleps       %xmm14,%xmm0
   .byte  68,15,89,238                        // mulps         %xmm6,%xmm13
@@ -22364,7 +22358,7 @@ _sk_hsl_to_rgb_sse41:
   .byte  65,15,40,198                        // movaps        %xmm14,%xmm0
   .byte  15,40,84,36,152                     // movaps        -0x68(%rsp),%xmm2
   .byte  102,15,56,20,202                    // blendvps      %xmm0,%xmm2,%xmm1
-  .byte  68,15,88,21,43,47,0,0               // addps         0x2f2b(%rip),%xmm10        # 4b30 <_sk_callback_sse41+0x520>
+  .byte  68,15,88,21,43,47,0,0               // addps         0x2f2b(%rip),%xmm10        # 4b30 <_sk_callback_sse41+0x523>
   .byte  102,65,15,58,8,194,1                // roundps       $0x1,%xmm10,%xmm0
   .byte  68,15,92,208                        // subps         %xmm0,%xmm10
   .byte  69,15,194,218,2                     // cmpleps       %xmm10,%xmm11
@@ -22416,7 +22410,7 @@ _sk_scale_u8_sse41:
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,68,15,56,49,4,56                // pmovzxbd      (%rax,%rdi,1),%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,136,46,0,0               // mulps         0x2e88(%rip),%xmm8        # 4b40 <_sk_callback_sse41+0x530>
+  .byte  68,15,89,5,136,46,0,0               // mulps         0x2e88(%rip),%xmm8        # 4b40 <_sk_callback_sse41+0x533>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
@@ -22454,7 +22448,7 @@ _sk_lerp_u8_sse41:
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,68,15,56,49,4,56                // pmovzxbd      (%rax,%rdi,1),%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,52,46,0,0                // mulps         0x2e34(%rip),%xmm8        # 4b50 <_sk_callback_sse41+0x540>
+  .byte  68,15,89,5,52,46,0,0                // mulps         0x2e34(%rip),%xmm8        # 4b50 <_sk_callback_sse41+0x543>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -22477,17 +22471,17 @@ _sk_lerp_565_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,68,15,56,51,20,120              // pmovzxwd      (%rax,%rdi,2),%xmm10
-  .byte  102,68,15,111,5,3,46,0,0            // movdqa        0x2e03(%rip),%xmm8        # 4b60 <_sk_callback_sse41+0x550>
+  .byte  102,68,15,111,5,3,46,0,0            // movdqa        0x2e03(%rip),%xmm8        # 4b60 <_sk_callback_sse41+0x553>
   .byte  102,69,15,219,194                   // pand          %xmm10,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,2,46,0,0                 // mulps         0x2e02(%rip),%xmm8        # 4b70 <_sk_callback_sse41+0x560>
-  .byte  102,68,15,111,13,9,46,0,0           // movdqa        0x2e09(%rip),%xmm9        # 4b80 <_sk_callback_sse41+0x570>
+  .byte  68,15,89,5,2,46,0,0                 // mulps         0x2e02(%rip),%xmm8        # 4b70 <_sk_callback_sse41+0x563>
+  .byte  102,68,15,111,13,9,46,0,0           // movdqa        0x2e09(%rip),%xmm9        # 4b80 <_sk_callback_sse41+0x573>
   .byte  102,69,15,219,202                   // pand          %xmm10,%xmm9
   .byte  69,15,91,201                        // cvtdq2ps      %xmm9,%xmm9
-  .byte  68,15,89,13,8,46,0,0                // mulps         0x2e08(%rip),%xmm9        # 4b90 <_sk_callback_sse41+0x580>
-  .byte  102,68,15,219,21,15,46,0,0          // pand          0x2e0f(%rip),%xmm10        # 4ba0 <_sk_callback_sse41+0x590>
+  .byte  68,15,89,13,8,46,0,0                // mulps         0x2e08(%rip),%xmm9        # 4b90 <_sk_callback_sse41+0x583>
+  .byte  102,68,15,219,21,15,46,0,0          // pand          0x2e0f(%rip),%xmm10        # 4ba0 <_sk_callback_sse41+0x593>
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
-  .byte  68,15,89,21,19,46,0,0               // mulps         0x2e13(%rip),%xmm10        # 4bb0 <_sk_callback_sse41+0x5a0>
+  .byte  68,15,89,21,19,46,0,0               // mulps         0x2e13(%rip),%xmm10        # 4bb0 <_sk_callback_sse41+0x5a3>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -22518,7 +22512,7 @@ _sk_load_tables_sse41:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  76,139,72,8                         // mov           0x8(%rax),%r9
   .byte  243,69,15,111,4,184                 // movdqu        (%r8,%rdi,4),%xmm8
-  .byte  102,15,111,5,196,45,0,0             // movdqa        0x2dc4(%rip),%xmm0        # 4bc0 <_sk_callback_sse41+0x5b0>
+  .byte  102,15,111,5,196,45,0,0             // movdqa        0x2dc4(%rip),%xmm0        # 4bc0 <_sk_callback_sse41+0x5b3>
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,73,15,58,22,192,1               // pextrq        $0x1,%xmm0,%r8
   .byte  102,72,15,126,193                   // movq          %xmm0,%rcx
@@ -22533,7 +22527,7 @@ _sk_load_tables_sse41:
   .byte  102,15,58,33,193,48                 // insertps      $0x30,%xmm1,%xmm0
   .byte  76,139,64,16                        // mov           0x10(%rax),%r8
   .byte  102,65,15,111,200                   // movdqa        %xmm8,%xmm1
-  .byte  102,15,56,0,13,127,45,0,0           // pshufb        0x2d7f(%rip),%xmm1        # 4bd0 <_sk_callback_sse41+0x5c0>
+  .byte  102,15,56,0,13,127,45,0,0           // pshufb        0x2d7f(%rip),%xmm1        # 4bd0 <_sk_callback_sse41+0x5c3>
   .byte  102,73,15,58,22,201,1               // pextrq        $0x1,%xmm1,%r9
   .byte  102,72,15,126,201                   // movq          %xmm1,%rcx
   .byte  68,15,182,209                       // movzbl        %cl,%r10d
@@ -22548,7 +22542,7 @@ _sk_load_tables_sse41:
   .byte  102,15,58,33,202,48                 // insertps      $0x30,%xmm2,%xmm1
   .byte  76,139,64,24                        // mov           0x18(%rax),%r8
   .byte  102,65,15,111,208                   // movdqa        %xmm8,%xmm2
-  .byte  102,15,56,0,21,59,45,0,0            // pshufb        0x2d3b(%rip),%xmm2        # 4be0 <_sk_callback_sse41+0x5d0>
+  .byte  102,15,56,0,21,59,45,0,0            // pshufb        0x2d3b(%rip),%xmm2        # 4be0 <_sk_callback_sse41+0x5d3>
   .byte  102,72,15,58,22,209,1               // pextrq        $0x1,%xmm2,%rcx
   .byte  102,72,15,126,208                   // movq          %xmm2,%rax
   .byte  68,15,182,200                       // movzbl        %al,%r9d
@@ -22563,7 +22557,7 @@ _sk_load_tables_sse41:
   .byte  102,15,58,33,211,48                 // insertps      $0x30,%xmm3,%xmm2
   .byte  102,65,15,114,208,24                // psrld         $0x18,%xmm8
   .byte  65,15,91,216                        // cvtdq2ps      %xmm8,%xmm3
-  .byte  15,89,29,248,44,0,0                 // mulps         0x2cf8(%rip),%xmm3        # 4bf0 <_sk_callback_sse41+0x5e0>
+  .byte  15,89,29,248,44,0,0                 // mulps         0x2cf8(%rip),%xmm3        # 4bf0 <_sk_callback_sse41+0x5e3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -22582,7 +22576,7 @@ _sk_load_tables_u16_be_sse41:
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,97,200                       // punpcklwd     %xmm0,%xmm1
   .byte  102,68,15,105,200                   // punpckhwd     %xmm0,%xmm9
-  .byte  102,68,15,111,5,203,44,0,0          // movdqa        0x2ccb(%rip),%xmm8        # 4c00 <_sk_callback_sse41+0x5f0>
+  .byte  102,68,15,111,5,203,44,0,0          // movdqa        0x2ccb(%rip),%xmm8        # 4c00 <_sk_callback_sse41+0x5f3>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,15,56,51,192                    // pmovzxwd      %xmm0,%xmm0
@@ -22599,7 +22593,7 @@ _sk_load_tables_u16_be_sse41:
   .byte  243,67,15,16,20,8                   // movss         (%r8,%r9,1),%xmm2
   .byte  102,15,58,33,194,48                 // insertps      $0x30,%xmm2,%xmm0
   .byte  76,139,64,16                        // mov           0x10(%rax),%r8
-  .byte  102,15,56,0,13,126,44,0,0           // pshufb        0x2c7e(%rip),%xmm1        # 4c10 <_sk_callback_sse41+0x600>
+  .byte  102,15,56,0,13,126,44,0,0           // pshufb        0x2c7e(%rip),%xmm1        # 4c10 <_sk_callback_sse41+0x603>
   .byte  102,15,56,51,201                    // pmovzxwd      %xmm1,%xmm1
   .byte  102,73,15,58,22,201,1               // pextrq        $0x1,%xmm1,%r9
   .byte  102,72,15,126,201                   // movq          %xmm1,%rcx
@@ -22635,7 +22629,7 @@ _sk_load_tables_u16_be_sse41:
   .byte  102,65,15,235,216                   // por           %xmm8,%xmm3
   .byte  102,15,56,51,219                    // pmovzxwd      %xmm3,%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,204,43,0,0                 // mulps         0x2bcc(%rip),%xmm3        # 4c20 <_sk_callback_sse41+0x610>
+  .byte  15,89,29,204,43,0,0                 // mulps         0x2bcc(%rip),%xmm3        # 4c20 <_sk_callback_sse41+0x613>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -22657,7 +22651,7 @@ _sk_load_tables_rgb_u16_be_sse41:
   .byte  102,68,15,97,200                    // punpcklwd     %xmm0,%xmm9
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,65,15,97,201                    // punpcklwd     %xmm9,%xmm1
-  .byte  102,68,15,111,5,142,43,0,0          // movdqa        0x2b8e(%rip),%xmm8        # 4c30 <_sk_callback_sse41+0x620>
+  .byte  102,68,15,111,5,142,43,0,0          // movdqa        0x2b8e(%rip),%xmm8        # 4c30 <_sk_callback_sse41+0x623>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,15,56,51,192                    // pmovzxwd      %xmm0,%xmm0
@@ -22674,7 +22668,7 @@ _sk_load_tables_rgb_u16_be_sse41:
   .byte  243,67,15,16,28,8                   // movss         (%r8,%r9,1),%xmm3
   .byte  102,15,58,33,195,48                 // insertps      $0x30,%xmm3,%xmm0
   .byte  76,139,64,16                        // mov           0x10(%rax),%r8
-  .byte  102,15,56,0,13,65,43,0,0            // pshufb        0x2b41(%rip),%xmm1        # 4c40 <_sk_callback_sse41+0x630>
+  .byte  102,15,56,0,13,65,43,0,0            // pshufb        0x2b41(%rip),%xmm1        # 4c40 <_sk_callback_sse41+0x633>
   .byte  102,15,56,51,201                    // pmovzxwd      %xmm1,%xmm1
   .byte  102,73,15,58,22,201,1               // pextrq        $0x1,%xmm1,%r9
   .byte  102,72,15,126,201                   // movq          %xmm1,%rcx
@@ -22705,7 +22699,7 @@ _sk_load_tables_rgb_u16_be_sse41:
   .byte  243,65,15,16,28,8                   // movss         (%r8,%rcx,1),%xmm3
   .byte  102,15,58,33,211,48                 // insertps      $0x30,%xmm3,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,172,42,0,0                 // movaps        0x2aac(%rip),%xmm3        # 4c50 <_sk_callback_sse41+0x640>
+  .byte  15,40,29,172,42,0,0                 // movaps        0x2aac(%rip),%xmm3        # 4c50 <_sk_callback_sse41+0x643>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_byte_tables_sse41
@@ -22715,7 +22709,7 @@ _sk_byte_tables_sse41:
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,173,42,0,0               // movaps        0x2aad(%rip),%xmm8        # 4c60 <_sk_callback_sse41+0x650>
+  .byte  68,15,40,5,173,42,0,0               // movaps        0x2aad(%rip),%xmm8        # 4c60 <_sk_callback_sse41+0x653>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  102,72,15,58,22,193,1               // pextrq        $0x1,%xmm0,%rcx
@@ -22734,7 +22728,7 @@ _sk_byte_tables_sse41:
   .byte  102,15,58,32,193,3                  // pinsrb        $0x3,%ecx,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,13,94,42,0,0               // movaps        0x2a5e(%rip),%xmm9        # 4c70 <_sk_callback_sse41+0x660>
+  .byte  68,15,40,13,94,42,0,0               // movaps        0x2a5e(%rip),%xmm9        # 4c70 <_sk_callback_sse41+0x663>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -22825,7 +22819,7 @@ _sk_byte_tables_rgb_sse41:
   .byte  102,15,58,32,193,3                  // pinsrb        $0x3,%ecx,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,13,230,40,0,0              // movaps        0x28e6(%rip),%xmm9        # 4c80 <_sk_callback_sse41+0x670>
+  .byte  68,15,40,13,230,40,0,0              // movaps        0x28e6(%rip),%xmm9        # 4c80 <_sk_callback_sse41+0x673>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -23002,31 +22996,31 @@ _sk_parametric_r_sse41:
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,194                        // cvtdq2ps      %xmm10,%xmm8
-  .byte  68,15,89,5,61,38,0,0                // mulps         0x263d(%rip),%xmm8        # 4c90 <_sk_callback_sse41+0x680>
-  .byte  68,15,84,21,69,38,0,0               // andps         0x2645(%rip),%xmm10        # 4ca0 <_sk_callback_sse41+0x690>
-  .byte  68,15,86,21,77,38,0,0               // orps          0x264d(%rip),%xmm10        # 4cb0 <_sk_callback_sse41+0x6a0>
-  .byte  68,15,88,5,85,38,0,0                // addps         0x2655(%rip),%xmm8        # 4cc0 <_sk_callback_sse41+0x6b0>
-  .byte  68,15,40,37,93,38,0,0               // movaps        0x265d(%rip),%xmm12        # 4cd0 <_sk_callback_sse41+0x6c0>
+  .byte  68,15,89,5,61,38,0,0                // mulps         0x263d(%rip),%xmm8        # 4c90 <_sk_callback_sse41+0x683>
+  .byte  68,15,84,21,69,38,0,0               // andps         0x2645(%rip),%xmm10        # 4ca0 <_sk_callback_sse41+0x693>
+  .byte  68,15,86,21,77,38,0,0               // orps          0x264d(%rip),%xmm10        # 4cb0 <_sk_callback_sse41+0x6a3>
+  .byte  68,15,88,5,85,38,0,0                // addps         0x2655(%rip),%xmm8        # 4cc0 <_sk_callback_sse41+0x6b3>
+  .byte  68,15,40,37,93,38,0,0               // movaps        0x265d(%rip),%xmm12        # 4cd0 <_sk_callback_sse41+0x6c3>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,196                        // subps         %xmm12,%xmm8
-  .byte  68,15,88,21,93,38,0,0               // addps         0x265d(%rip),%xmm10        # 4ce0 <_sk_callback_sse41+0x6d0>
-  .byte  68,15,40,37,101,38,0,0              // movaps        0x2665(%rip),%xmm12        # 4cf0 <_sk_callback_sse41+0x6e0>
+  .byte  68,15,88,21,93,38,0,0               // addps         0x265d(%rip),%xmm10        # 4ce0 <_sk_callback_sse41+0x6d3>
+  .byte  68,15,40,37,101,38,0,0              // movaps        0x2665(%rip),%xmm12        # 4cf0 <_sk_callback_sse41+0x6e3>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,196                        // subps         %xmm12,%xmm8
   .byte  69,15,89,195                        // mulps         %xmm11,%xmm8
   .byte  102,69,15,58,8,208,1                // roundps       $0x1,%xmm8,%xmm10
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,5,82,38,0,0                // addps         0x2652(%rip),%xmm8        # 4d00 <_sk_callback_sse41+0x6f0>
-  .byte  68,15,40,21,90,38,0,0               // movaps        0x265a(%rip),%xmm10        # 4d10 <_sk_callback_sse41+0x700>
+  .byte  68,15,88,5,82,38,0,0                // addps         0x2652(%rip),%xmm8        # 4d00 <_sk_callback_sse41+0x6f3>
+  .byte  68,15,40,21,90,38,0,0               // movaps        0x265a(%rip),%xmm10        # 4d10 <_sk_callback_sse41+0x703>
   .byte  69,15,89,211                        // mulps         %xmm11,%xmm10
   .byte  69,15,92,194                        // subps         %xmm10,%xmm8
-  .byte  68,15,40,21,90,38,0,0               // movaps        0x265a(%rip),%xmm10        # 4d20 <_sk_callback_sse41+0x710>
+  .byte  68,15,40,21,90,38,0,0               // movaps        0x265a(%rip),%xmm10        # 4d20 <_sk_callback_sse41+0x713>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  68,15,40,29,94,38,0,0               // movaps        0x265e(%rip),%xmm11        # 4d30 <_sk_callback_sse41+0x720>
+  .byte  68,15,40,29,94,38,0,0               // movaps        0x265e(%rip),%xmm11        # 4d30 <_sk_callback_sse41+0x723>
   .byte  69,15,94,218                        // divps         %xmm10,%xmm11
   .byte  69,15,88,216                        // addps         %xmm8,%xmm11
-  .byte  68,15,89,29,94,38,0,0               // mulps         0x265e(%rip),%xmm11        # 4d40 <_sk_callback_sse41+0x730>
+  .byte  68,15,89,29,94,38,0,0               // mulps         0x265e(%rip),%xmm11        # 4d40 <_sk_callback_sse41+0x733>
   .byte  102,69,15,91,211                    // cvtps2dq      %xmm11,%xmm10
   .byte  243,68,15,16,64,20                  // movss         0x14(%rax),%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
@@ -23034,7 +23028,7 @@ _sk_parametric_r_sse41:
   .byte  102,69,15,56,20,193                 // blendvps      %xmm0,%xmm9,%xmm8
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
-  .byte  68,15,93,5,69,38,0,0                // minps         0x2645(%rip),%xmm8        # 4d50 <_sk_callback_sse41+0x740>
+  .byte  68,15,93,5,69,38,0,0                // minps         0x2645(%rip),%xmm8        # 4d50 <_sk_callback_sse41+0x743>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23064,31 +23058,31 @@ _sk_parametric_g_sse41:
   .byte  68,15,88,217                        // addps         %xmm1,%xmm11
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,91,227                        // cvtdq2ps      %xmm11,%xmm12
-  .byte  68,15,89,37,230,37,0,0              // mulps         0x25e6(%rip),%xmm12        # 4d60 <_sk_callback_sse41+0x750>
-  .byte  68,15,84,29,238,37,0,0              // andps         0x25ee(%rip),%xmm11        # 4d70 <_sk_callback_sse41+0x760>
-  .byte  68,15,86,29,246,37,0,0              // orps          0x25f6(%rip),%xmm11        # 4d80 <_sk_callback_sse41+0x770>
-  .byte  68,15,88,37,254,37,0,0              // addps         0x25fe(%rip),%xmm12        # 4d90 <_sk_callback_sse41+0x780>
-  .byte  15,40,13,7,38,0,0                   // movaps        0x2607(%rip),%xmm1        # 4da0 <_sk_callback_sse41+0x790>
+  .byte  68,15,89,37,230,37,0,0              // mulps         0x25e6(%rip),%xmm12        # 4d60 <_sk_callback_sse41+0x753>
+  .byte  68,15,84,29,238,37,0,0              // andps         0x25ee(%rip),%xmm11        # 4d70 <_sk_callback_sse41+0x763>
+  .byte  68,15,86,29,246,37,0,0              // orps          0x25f6(%rip),%xmm11        # 4d80 <_sk_callback_sse41+0x773>
+  .byte  68,15,88,37,254,37,0,0              // addps         0x25fe(%rip),%xmm12        # 4d90 <_sk_callback_sse41+0x783>
+  .byte  15,40,13,7,38,0,0                   // movaps        0x2607(%rip),%xmm1        # 4da0 <_sk_callback_sse41+0x793>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
-  .byte  68,15,88,29,7,38,0,0                // addps         0x2607(%rip),%xmm11        # 4db0 <_sk_callback_sse41+0x7a0>
-  .byte  15,40,13,16,38,0,0                  // movaps        0x2610(%rip),%xmm1        # 4dc0 <_sk_callback_sse41+0x7b0>
+  .byte  68,15,88,29,7,38,0,0                // addps         0x2607(%rip),%xmm11        # 4db0 <_sk_callback_sse41+0x7a3>
+  .byte  15,40,13,16,38,0,0                  // movaps        0x2610(%rip),%xmm1        # 4dc0 <_sk_callback_sse41+0x7b3>
   .byte  65,15,94,203                        // divps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  102,69,15,58,8,212,1                // roundps       $0x1,%xmm12,%xmm10
   .byte  69,15,40,220                        // movaps        %xmm12,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,37,253,37,0,0              // addps         0x25fd(%rip),%xmm12        # 4dd0 <_sk_callback_sse41+0x7c0>
-  .byte  15,40,13,6,38,0,0                   // movaps        0x2606(%rip),%xmm1        # 4de0 <_sk_callback_sse41+0x7d0>
+  .byte  68,15,88,37,253,37,0,0              // addps         0x25fd(%rip),%xmm12        # 4dd0 <_sk_callback_sse41+0x7c3>
+  .byte  15,40,13,6,38,0,0                   // movaps        0x2606(%rip),%xmm1        # 4de0 <_sk_callback_sse41+0x7d3>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
-  .byte  68,15,40,21,6,38,0,0                // movaps        0x2606(%rip),%xmm10        # 4df0 <_sk_callback_sse41+0x7e0>
+  .byte  68,15,40,21,6,38,0,0                // movaps        0x2606(%rip),%xmm10        # 4df0 <_sk_callback_sse41+0x7e3>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,13,11,38,0,0                  // movaps        0x260b(%rip),%xmm1        # 4e00 <_sk_callback_sse41+0x7f0>
+  .byte  15,40,13,11,38,0,0                  // movaps        0x260b(%rip),%xmm1        # 4e00 <_sk_callback_sse41+0x7f3>
   .byte  65,15,94,202                        // divps         %xmm10,%xmm1
   .byte  65,15,88,204                        // addps         %xmm12,%xmm1
-  .byte  15,89,13,12,38,0,0                  // mulps         0x260c(%rip),%xmm1        # 4e10 <_sk_callback_sse41+0x800>
+  .byte  15,89,13,12,38,0,0                  // mulps         0x260c(%rip),%xmm1        # 4e10 <_sk_callback_sse41+0x803>
   .byte  102,68,15,91,209                    // cvtps2dq      %xmm1,%xmm10
   .byte  243,15,16,72,20                     // movss         0x14(%rax),%xmm1
   .byte  15,198,201,0                        // shufps        $0x0,%xmm1,%xmm1
@@ -23096,7 +23090,7 @@ _sk_parametric_g_sse41:
   .byte  102,65,15,56,20,201                 // blendvps      %xmm0,%xmm9,%xmm1
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,95,200                           // maxps         %xmm0,%xmm1
-  .byte  15,93,13,247,37,0,0                 // minps         0x25f7(%rip),%xmm1        # 4e20 <_sk_callback_sse41+0x810>
+  .byte  15,93,13,247,37,0,0                 // minps         0x25f7(%rip),%xmm1        # 4e20 <_sk_callback_sse41+0x813>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23126,31 +23120,31 @@ _sk_parametric_b_sse41:
   .byte  68,15,88,218                        // addps         %xmm2,%xmm11
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,91,227                        // cvtdq2ps      %xmm11,%xmm12
-  .byte  68,15,89,37,152,37,0,0              // mulps         0x2598(%rip),%xmm12        # 4e30 <_sk_callback_sse41+0x820>
-  .byte  68,15,84,29,160,37,0,0              // andps         0x25a0(%rip),%xmm11        # 4e40 <_sk_callback_sse41+0x830>
-  .byte  68,15,86,29,168,37,0,0              // orps          0x25a8(%rip),%xmm11        # 4e50 <_sk_callback_sse41+0x840>
-  .byte  68,15,88,37,176,37,0,0              // addps         0x25b0(%rip),%xmm12        # 4e60 <_sk_callback_sse41+0x850>
-  .byte  15,40,21,185,37,0,0                 // movaps        0x25b9(%rip),%xmm2        # 4e70 <_sk_callback_sse41+0x860>
+  .byte  68,15,89,37,152,37,0,0              // mulps         0x2598(%rip),%xmm12        # 4e30 <_sk_callback_sse41+0x823>
+  .byte  68,15,84,29,160,37,0,0              // andps         0x25a0(%rip),%xmm11        # 4e40 <_sk_callback_sse41+0x833>
+  .byte  68,15,86,29,168,37,0,0              // orps          0x25a8(%rip),%xmm11        # 4e50 <_sk_callback_sse41+0x843>
+  .byte  68,15,88,37,176,37,0,0              // addps         0x25b0(%rip),%xmm12        # 4e60 <_sk_callback_sse41+0x853>
+  .byte  15,40,21,185,37,0,0                 // movaps        0x25b9(%rip),%xmm2        # 4e70 <_sk_callback_sse41+0x863>
   .byte  65,15,89,211                        // mulps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
-  .byte  68,15,88,29,185,37,0,0              // addps         0x25b9(%rip),%xmm11        # 4e80 <_sk_callback_sse41+0x870>
-  .byte  15,40,21,194,37,0,0                 // movaps        0x25c2(%rip),%xmm2        # 4e90 <_sk_callback_sse41+0x880>
+  .byte  68,15,88,29,185,37,0,0              // addps         0x25b9(%rip),%xmm11        # 4e80 <_sk_callback_sse41+0x873>
+  .byte  15,40,21,194,37,0,0                 // movaps        0x25c2(%rip),%xmm2        # 4e90 <_sk_callback_sse41+0x883>
   .byte  65,15,94,211                        // divps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  102,69,15,58,8,212,1                // roundps       $0x1,%xmm12,%xmm10
   .byte  69,15,40,220                        // movaps        %xmm12,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,37,175,37,0,0              // addps         0x25af(%rip),%xmm12        # 4ea0 <_sk_callback_sse41+0x890>
-  .byte  15,40,21,184,37,0,0                 // movaps        0x25b8(%rip),%xmm2        # 4eb0 <_sk_callback_sse41+0x8a0>
+  .byte  68,15,88,37,175,37,0,0              // addps         0x25af(%rip),%xmm12        # 4ea0 <_sk_callback_sse41+0x893>
+  .byte  15,40,21,184,37,0,0                 // movaps        0x25b8(%rip),%xmm2        # 4eb0 <_sk_callback_sse41+0x8a3>
   .byte  65,15,89,211                        // mulps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
-  .byte  68,15,40,21,184,37,0,0              // movaps        0x25b8(%rip),%xmm10        # 4ec0 <_sk_callback_sse41+0x8b0>
+  .byte  68,15,40,21,184,37,0,0              // movaps        0x25b8(%rip),%xmm10        # 4ec0 <_sk_callback_sse41+0x8b3>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,21,189,37,0,0                 // movaps        0x25bd(%rip),%xmm2        # 4ed0 <_sk_callback_sse41+0x8c0>
+  .byte  15,40,21,189,37,0,0                 // movaps        0x25bd(%rip),%xmm2        # 4ed0 <_sk_callback_sse41+0x8c3>
   .byte  65,15,94,210                        // divps         %xmm10,%xmm2
   .byte  65,15,88,212                        // addps         %xmm12,%xmm2
-  .byte  15,89,21,190,37,0,0                 // mulps         0x25be(%rip),%xmm2        # 4ee0 <_sk_callback_sse41+0x8d0>
+  .byte  15,89,21,190,37,0,0                 // mulps         0x25be(%rip),%xmm2        # 4ee0 <_sk_callback_sse41+0x8d3>
   .byte  102,68,15,91,210                    // cvtps2dq      %xmm2,%xmm10
   .byte  243,15,16,80,20                     // movss         0x14(%rax),%xmm2
   .byte  15,198,210,0                        // shufps        $0x0,%xmm2,%xmm2
@@ -23158,7 +23152,7 @@ _sk_parametric_b_sse41:
   .byte  102,65,15,56,20,209                 // blendvps      %xmm0,%xmm9,%xmm2
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,95,208                           // maxps         %xmm0,%xmm2
-  .byte  15,93,21,169,37,0,0                 // minps         0x25a9(%rip),%xmm2        # 4ef0 <_sk_callback_sse41+0x8e0>
+  .byte  15,93,21,169,37,0,0                 // minps         0x25a9(%rip),%xmm2        # 4ef0 <_sk_callback_sse41+0x8e3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23188,31 +23182,31 @@ _sk_parametric_a_sse41:
   .byte  68,15,88,219                        // addps         %xmm3,%xmm11
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,91,227                        // cvtdq2ps      %xmm11,%xmm12
-  .byte  68,15,89,37,74,37,0,0               // mulps         0x254a(%rip),%xmm12        # 4f00 <_sk_callback_sse41+0x8f0>
-  .byte  68,15,84,29,82,37,0,0               // andps         0x2552(%rip),%xmm11        # 4f10 <_sk_callback_sse41+0x900>
-  .byte  68,15,86,29,90,37,0,0               // orps          0x255a(%rip),%xmm11        # 4f20 <_sk_callback_sse41+0x910>
-  .byte  68,15,88,37,98,37,0,0               // addps         0x2562(%rip),%xmm12        # 4f30 <_sk_callback_sse41+0x920>
-  .byte  15,40,29,107,37,0,0                 // movaps        0x256b(%rip),%xmm3        # 4f40 <_sk_callback_sse41+0x930>
+  .byte  68,15,89,37,74,37,0,0               // mulps         0x254a(%rip),%xmm12        # 4f00 <_sk_callback_sse41+0x8f3>
+  .byte  68,15,84,29,82,37,0,0               // andps         0x2552(%rip),%xmm11        # 4f10 <_sk_callback_sse41+0x903>
+  .byte  68,15,86,29,90,37,0,0               // orps          0x255a(%rip),%xmm11        # 4f20 <_sk_callback_sse41+0x913>
+  .byte  68,15,88,37,98,37,0,0               // addps         0x2562(%rip),%xmm12        # 4f30 <_sk_callback_sse41+0x923>
+  .byte  15,40,29,107,37,0,0                 // movaps        0x256b(%rip),%xmm3        # 4f40 <_sk_callback_sse41+0x933>
   .byte  65,15,89,219                        // mulps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
-  .byte  68,15,88,29,107,37,0,0              // addps         0x256b(%rip),%xmm11        # 4f50 <_sk_callback_sse41+0x940>
-  .byte  15,40,29,116,37,0,0                 // movaps        0x2574(%rip),%xmm3        # 4f60 <_sk_callback_sse41+0x950>
+  .byte  68,15,88,29,107,37,0,0              // addps         0x256b(%rip),%xmm11        # 4f50 <_sk_callback_sse41+0x943>
+  .byte  15,40,29,116,37,0,0                 // movaps        0x2574(%rip),%xmm3        # 4f60 <_sk_callback_sse41+0x953>
   .byte  65,15,94,219                        // divps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  102,69,15,58,8,212,1                // roundps       $0x1,%xmm12,%xmm10
   .byte  69,15,40,220                        // movaps        %xmm12,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,37,97,37,0,0               // addps         0x2561(%rip),%xmm12        # 4f70 <_sk_callback_sse41+0x960>
-  .byte  15,40,29,106,37,0,0                 // movaps        0x256a(%rip),%xmm3        # 4f80 <_sk_callback_sse41+0x970>
+  .byte  68,15,88,37,97,37,0,0               // addps         0x2561(%rip),%xmm12        # 4f70 <_sk_callback_sse41+0x963>
+  .byte  15,40,29,106,37,0,0                 // movaps        0x256a(%rip),%xmm3        # 4f80 <_sk_callback_sse41+0x973>
   .byte  65,15,89,219                        // mulps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
-  .byte  68,15,40,21,106,37,0,0              // movaps        0x256a(%rip),%xmm10        # 4f90 <_sk_callback_sse41+0x980>
+  .byte  68,15,40,21,106,37,0,0              // movaps        0x256a(%rip),%xmm10        # 4f90 <_sk_callback_sse41+0x983>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,29,111,37,0,0                 // movaps        0x256f(%rip),%xmm3        # 4fa0 <_sk_callback_sse41+0x990>
+  .byte  15,40,29,111,37,0,0                 // movaps        0x256f(%rip),%xmm3        # 4fa0 <_sk_callback_sse41+0x993>
   .byte  65,15,94,218                        // divps         %xmm10,%xmm3
   .byte  65,15,88,220                        // addps         %xmm12,%xmm3
-  .byte  15,89,29,112,37,0,0                 // mulps         0x2570(%rip),%xmm3        # 4fb0 <_sk_callback_sse41+0x9a0>
+  .byte  15,89,29,112,37,0,0                 // mulps         0x2570(%rip),%xmm3        # 4fb0 <_sk_callback_sse41+0x9a3>
   .byte  102,68,15,91,211                    // cvtps2dq      %xmm3,%xmm10
   .byte  243,15,16,88,20                     // movss         0x14(%rax),%xmm3
   .byte  15,198,219,0                        // shufps        $0x0,%xmm3,%xmm3
@@ -23220,7 +23214,7 @@ _sk_parametric_a_sse41:
   .byte  102,65,15,56,20,217                 // blendvps      %xmm0,%xmm9,%xmm3
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,95,216                           // maxps         %xmm0,%xmm3
-  .byte  15,93,29,91,37,0,0                  // minps         0x255b(%rip),%xmm3        # 4fc0 <_sk_callback_sse41+0x9b0>
+  .byte  15,93,29,91,37,0,0                  // minps         0x255b(%rip),%xmm3        # 4fc0 <_sk_callback_sse41+0x9b3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23230,29 +23224,29 @@ HIDDEN _sk_lab_to_xyz_sse41
 FUNCTION(_sk_lab_to_xyz_sse41)
 _sk_lab_to_xyz_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,89,5,87,37,0,0                // mulps         0x2557(%rip),%xmm8        # 4fd0 <_sk_callback_sse41+0x9c0>
-  .byte  68,15,40,13,95,37,0,0               // movaps        0x255f(%rip),%xmm9        # 4fe0 <_sk_callback_sse41+0x9d0>
+  .byte  68,15,89,5,87,37,0,0                // mulps         0x2557(%rip),%xmm8        # 4fd0 <_sk_callback_sse41+0x9c3>
+  .byte  68,15,40,13,95,37,0,0               // movaps        0x255f(%rip),%xmm9        # 4fe0 <_sk_callback_sse41+0x9d3>
   .byte  65,15,89,201                        // mulps         %xmm9,%xmm1
-  .byte  15,40,5,100,37,0,0                  // movaps        0x2564(%rip),%xmm0        # 4ff0 <_sk_callback_sse41+0x9e0>
+  .byte  15,40,5,100,37,0,0                  // movaps        0x2564(%rip),%xmm0        # 4ff0 <_sk_callback_sse41+0x9e3>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
-  .byte  68,15,88,5,98,37,0,0                // addps         0x2562(%rip),%xmm8        # 5000 <_sk_callback_sse41+0x9f0>
-  .byte  68,15,89,5,106,37,0,0               // mulps         0x256a(%rip),%xmm8        # 5010 <_sk_callback_sse41+0xa00>
-  .byte  15,89,13,115,37,0,0                 // mulps         0x2573(%rip),%xmm1        # 5020 <_sk_callback_sse41+0xa10>
+  .byte  68,15,88,5,98,37,0,0                // addps         0x2562(%rip),%xmm8        # 5000 <_sk_callback_sse41+0x9f3>
+  .byte  68,15,89,5,106,37,0,0               // mulps         0x256a(%rip),%xmm8        # 5010 <_sk_callback_sse41+0xa03>
+  .byte  15,89,13,115,37,0,0                 // mulps         0x2573(%rip),%xmm1        # 5020 <_sk_callback_sse41+0xa13>
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  15,89,21,120,37,0,0                 // mulps         0x2578(%rip),%xmm2        # 5030 <_sk_callback_sse41+0xa20>
+  .byte  15,89,21,120,37,0,0                 // mulps         0x2578(%rip),%xmm2        # 5030 <_sk_callback_sse41+0xa23>
   .byte  69,15,40,208                        // movaps        %xmm8,%xmm10
   .byte  68,15,92,210                        // subps         %xmm2,%xmm10
   .byte  68,15,40,217                        // movaps        %xmm1,%xmm11
   .byte  69,15,89,219                        // mulps         %xmm11,%xmm11
   .byte  68,15,89,217                        // mulps         %xmm1,%xmm11
-  .byte  68,15,40,13,108,37,0,0              // movaps        0x256c(%rip),%xmm9        # 5040 <_sk_callback_sse41+0xa30>
+  .byte  68,15,40,13,108,37,0,0              // movaps        0x256c(%rip),%xmm9        # 5040 <_sk_callback_sse41+0xa33>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  65,15,194,195,1                     // cmpltps       %xmm11,%xmm0
-  .byte  15,40,21,108,37,0,0                 // movaps        0x256c(%rip),%xmm2        # 5050 <_sk_callback_sse41+0xa40>
+  .byte  15,40,21,108,37,0,0                 // movaps        0x256c(%rip),%xmm2        # 5050 <_sk_callback_sse41+0xa43>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
-  .byte  68,15,40,37,113,37,0,0              // movaps        0x2571(%rip),%xmm12        # 5060 <_sk_callback_sse41+0xa50>
+  .byte  68,15,40,37,113,37,0,0              // movaps        0x2571(%rip),%xmm12        # 5060 <_sk_callback_sse41+0xa53>
   .byte  65,15,89,204                        // mulps         %xmm12,%xmm1
   .byte  102,65,15,56,20,203                 // blendvps      %xmm0,%xmm11,%xmm1
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
@@ -23271,8 +23265,8 @@ _sk_lab_to_xyz_sse41:
   .byte  65,15,89,212                        // mulps         %xmm12,%xmm2
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  102,65,15,56,20,211                 // blendvps      %xmm0,%xmm11,%xmm2
-  .byte  15,89,13,42,37,0,0                  // mulps         0x252a(%rip),%xmm1        # 5070 <_sk_callback_sse41+0xa60>
-  .byte  15,89,21,51,37,0,0                  // mulps         0x2533(%rip),%xmm2        # 5080 <_sk_callback_sse41+0xa70>
+  .byte  15,89,13,42,37,0,0                  // mulps         0x252a(%rip),%xmm1        # 5070 <_sk_callback_sse41+0xa63>
+  .byte  15,89,21,51,37,0,0                  // mulps         0x2533(%rip),%xmm2        # 5080 <_sk_callback_sse41+0xa73>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  65,15,40,200                        // movaps        %xmm8,%xmm1
@@ -23286,7 +23280,7 @@ _sk_load_a8_sse41:
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,15,56,49,4,56                   // pmovzxbd      (%rax,%rdi,1),%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,35,37,0,0                  // mulps         0x2523(%rip),%xmm3        # 5090 <_sk_callback_sse41+0xa80>
+  .byte  15,89,29,35,37,0,0                  // mulps         0x2523(%rip),%xmm3        # 5090 <_sk_callback_sse41+0xa83>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
@@ -23319,7 +23313,7 @@ _sk_gather_a8_sse41:
   .byte  102,15,58,32,192,3                  // pinsrb        $0x3,%eax,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,183,36,0,0                 // mulps         0x24b7(%rip),%xmm3        # 50a0 <_sk_callback_sse41+0xa90>
+  .byte  15,89,29,183,36,0,0                 // mulps         0x24b7(%rip),%xmm3        # 50a0 <_sk_callback_sse41+0xa93>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -23332,7 +23326,7 @@ FUNCTION(_sk_store_a8_sse41)
 _sk_store_a8_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,171,36,0,0               // movaps        0x24ab(%rip),%xmm8        # 50b0 <_sk_callback_sse41+0xaa0>
+  .byte  68,15,40,5,171,36,0,0               // movaps        0x24ab(%rip),%xmm8        # 50b0 <_sk_callback_sse41+0xaa3>
   .byte  68,15,89,195                        // mulps         %xmm3,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
   .byte  102,69,15,56,43,192                 // packusdw      %xmm8,%xmm8
@@ -23349,9 +23343,9 @@ _sk_load_g8_sse41:
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,15,56,49,4,56                   // pmovzxbd      (%rax,%rdi,1),%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,136,36,0,0                  // mulps         0x2488(%rip),%xmm0        # 50c0 <_sk_callback_sse41+0xab0>
+  .byte  15,89,5,136,36,0,0                  // mulps         0x2488(%rip),%xmm0        # 50c0 <_sk_callback_sse41+0xab3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,143,36,0,0                 // movaps        0x248f(%rip),%xmm3        # 50d0 <_sk_callback_sse41+0xac0>
+  .byte  15,40,29,143,36,0,0                 // movaps        0x248f(%rip),%xmm3        # 50d0 <_sk_callback_sse41+0xac3>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -23382,9 +23376,9 @@ _sk_gather_g8_sse41:
   .byte  102,15,58,32,192,3                  // pinsrb        $0x3,%eax,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,40,36,0,0                   // mulps         0x2428(%rip),%xmm0        # 50e0 <_sk_callback_sse41+0xad0>
+  .byte  15,89,5,40,36,0,0                   // mulps         0x2428(%rip),%xmm0        # 50e0 <_sk_callback_sse41+0xad3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,47,36,0,0                  // movaps        0x242f(%rip),%xmm3        # 50f0 <_sk_callback_sse41+0xae0>
+  .byte  15,40,29,47,36,0,0                  // movaps        0x242f(%rip),%xmm3        # 50f0 <_sk_callback_sse41+0xae3>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -23429,17 +23423,17 @@ _sk_gather_i8_sse41:
   .byte  102,15,58,34,28,8,1                 // pinsrd        $0x1,(%rax,%rcx,1),%xmm3
   .byte  102,66,15,58,34,28,144,2            // pinsrd        $0x2,(%rax,%r10,4),%xmm3
   .byte  102,66,15,58,34,28,8,3              // pinsrd        $0x3,(%rax,%r9,1),%xmm3
-  .byte  102,15,111,5,134,35,0,0             // movdqa        0x2386(%rip),%xmm0        # 5100 <_sk_callback_sse41+0xaf0>
+  .byte  102,15,111,5,134,35,0,0             // movdqa        0x2386(%rip),%xmm0        # 5100 <_sk_callback_sse41+0xaf3>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,135,35,0,0               // movaps        0x2387(%rip),%xmm8        # 5110 <_sk_callback_sse41+0xb00>
+  .byte  68,15,40,5,135,35,0,0               // movaps        0x2387(%rip),%xmm8        # 5110 <_sk_callback_sse41+0xb03>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
-  .byte  102,15,56,0,13,134,35,0,0           // pshufb        0x2386(%rip),%xmm1        # 5120 <_sk_callback_sse41+0xb10>
+  .byte  102,15,56,0,13,134,35,0,0           // pshufb        0x2386(%rip),%xmm1        # 5120 <_sk_callback_sse41+0xb13>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,211                      // movdqa        %xmm3,%xmm2
-  .byte  102,15,56,0,21,130,35,0,0           // pshufb        0x2382(%rip),%xmm2        # 5130 <_sk_callback_sse41+0xb20>
+  .byte  102,15,56,0,21,130,35,0,0           // pshufb        0x2382(%rip),%xmm2        # 5130 <_sk_callback_sse41+0xb23>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -23455,19 +23449,19 @@ _sk_load_565_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,15,56,51,20,120                 // pmovzxwd      (%rax,%rdi,2),%xmm2
-  .byte  102,15,111,5,104,35,0,0             // movdqa        0x2368(%rip),%xmm0        # 5140 <_sk_callback_sse41+0xb30>
+  .byte  102,15,111,5,104,35,0,0             // movdqa        0x2368(%rip),%xmm0        # 5140 <_sk_callback_sse41+0xb33>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,106,35,0,0                  // mulps         0x236a(%rip),%xmm0        # 5150 <_sk_callback_sse41+0xb40>
-  .byte  102,15,111,13,114,35,0,0            // movdqa        0x2372(%rip),%xmm1        # 5160 <_sk_callback_sse41+0xb50>
+  .byte  15,89,5,106,35,0,0                  // mulps         0x236a(%rip),%xmm0        # 5150 <_sk_callback_sse41+0xb43>
+  .byte  102,15,111,13,114,35,0,0            // movdqa        0x2372(%rip),%xmm1        # 5160 <_sk_callback_sse41+0xb53>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,116,35,0,0                 // mulps         0x2374(%rip),%xmm1        # 5170 <_sk_callback_sse41+0xb60>
-  .byte  102,15,219,21,124,35,0,0            // pand          0x237c(%rip),%xmm2        # 5180 <_sk_callback_sse41+0xb70>
+  .byte  15,89,13,116,35,0,0                 // mulps         0x2374(%rip),%xmm1        # 5170 <_sk_callback_sse41+0xb63>
+  .byte  102,15,219,21,124,35,0,0            // pand          0x237c(%rip),%xmm2        # 5180 <_sk_callback_sse41+0xb73>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,130,35,0,0                 // mulps         0x2382(%rip),%xmm2        # 5190 <_sk_callback_sse41+0xb80>
+  .byte  15,89,21,130,35,0,0                 // mulps         0x2382(%rip),%xmm2        # 5190 <_sk_callback_sse41+0xb83>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,137,35,0,0                 // movaps        0x2389(%rip),%xmm3        # 51a0 <_sk_callback_sse41+0xb90>
+  .byte  15,40,29,137,35,0,0                 // movaps        0x2389(%rip),%xmm3        # 51a0 <_sk_callback_sse41+0xb93>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_gather_565_sse41
@@ -23495,19 +23489,19 @@ _sk_gather_565_sse41:
   .byte  65,15,183,4,65                      // movzwl        (%r9,%rax,2),%eax
   .byte  102,15,196,192,3                    // pinsrw        $0x3,%eax,%xmm0
   .byte  102,15,56,51,208                    // pmovzxwd      %xmm0,%xmm2
-  .byte  102,15,111,5,46,35,0,0              // movdqa        0x232e(%rip),%xmm0        # 51b0 <_sk_callback_sse41+0xba0>
+  .byte  102,15,111,5,46,35,0,0              // movdqa        0x232e(%rip),%xmm0        # 51b0 <_sk_callback_sse41+0xba3>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,48,35,0,0                   // mulps         0x2330(%rip),%xmm0        # 51c0 <_sk_callback_sse41+0xbb0>
-  .byte  102,15,111,13,56,35,0,0             // movdqa        0x2338(%rip),%xmm1        # 51d0 <_sk_callback_sse41+0xbc0>
+  .byte  15,89,5,48,35,0,0                   // mulps         0x2330(%rip),%xmm0        # 51c0 <_sk_callback_sse41+0xbb3>
+  .byte  102,15,111,13,56,35,0,0             // movdqa        0x2338(%rip),%xmm1        # 51d0 <_sk_callback_sse41+0xbc3>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,58,35,0,0                  // mulps         0x233a(%rip),%xmm1        # 51e0 <_sk_callback_sse41+0xbd0>
-  .byte  102,15,219,21,66,35,0,0             // pand          0x2342(%rip),%xmm2        # 51f0 <_sk_callback_sse41+0xbe0>
+  .byte  15,89,13,58,35,0,0                  // mulps         0x233a(%rip),%xmm1        # 51e0 <_sk_callback_sse41+0xbd3>
+  .byte  102,15,219,21,66,35,0,0             // pand          0x2342(%rip),%xmm2        # 51f0 <_sk_callback_sse41+0xbe3>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,72,35,0,0                  // mulps         0x2348(%rip),%xmm2        # 5200 <_sk_callback_sse41+0xbf0>
+  .byte  15,89,21,72,35,0,0                  // mulps         0x2348(%rip),%xmm2        # 5200 <_sk_callback_sse41+0xbf3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,79,35,0,0                  // movaps        0x234f(%rip),%xmm3        # 5210 <_sk_callback_sse41+0xc00>
+  .byte  15,40,29,79,35,0,0                  // movaps        0x234f(%rip),%xmm3        # 5210 <_sk_callback_sse41+0xc03>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_565_sse41
@@ -23516,12 +23510,12 @@ FUNCTION(_sk_store_565_sse41)
 _sk_store_565_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,80,35,0,0                // movaps        0x2350(%rip),%xmm8        # 5220 <_sk_callback_sse41+0xc10>
+  .byte  68,15,40,5,80,35,0,0                // movaps        0x2350(%rip),%xmm8        # 5220 <_sk_callback_sse41+0xc13>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
   .byte  102,65,15,114,241,11                // pslld         $0xb,%xmm9
-  .byte  68,15,40,21,69,35,0,0               // movaps        0x2345(%rip),%xmm10        # 5230 <_sk_callback_sse41+0xc20>
+  .byte  68,15,40,21,69,35,0,0               // movaps        0x2345(%rip),%xmm10        # 5230 <_sk_callback_sse41+0xc23>
   .byte  68,15,89,209                        // mulps         %xmm1,%xmm10
   .byte  102,69,15,91,210                    // cvtps2dq      %xmm10,%xmm10
   .byte  102,65,15,114,242,5                 // pslld         $0x5,%xmm10
@@ -23541,21 +23535,21 @@ _sk_load_4444_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  102,15,56,51,28,120                 // pmovzxwd      (%rax,%rdi,2),%xmm3
-  .byte  102,15,111,5,16,35,0,0              // movdqa        0x2310(%rip),%xmm0        # 5240 <_sk_callback_sse41+0xc30>
+  .byte  102,15,111,5,16,35,0,0              // movdqa        0x2310(%rip),%xmm0        # 5240 <_sk_callback_sse41+0xc33>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,18,35,0,0                   // mulps         0x2312(%rip),%xmm0        # 5250 <_sk_callback_sse41+0xc40>
-  .byte  102,15,111,13,26,35,0,0             // movdqa        0x231a(%rip),%xmm1        # 5260 <_sk_callback_sse41+0xc50>
+  .byte  15,89,5,18,35,0,0                   // mulps         0x2312(%rip),%xmm0        # 5250 <_sk_callback_sse41+0xc43>
+  .byte  102,15,111,13,26,35,0,0             // movdqa        0x231a(%rip),%xmm1        # 5260 <_sk_callback_sse41+0xc53>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,28,35,0,0                  // mulps         0x231c(%rip),%xmm1        # 5270 <_sk_callback_sse41+0xc60>
-  .byte  102,15,111,21,36,35,0,0             // movdqa        0x2324(%rip),%xmm2        # 5280 <_sk_callback_sse41+0xc70>
+  .byte  15,89,13,28,35,0,0                  // mulps         0x231c(%rip),%xmm1        # 5270 <_sk_callback_sse41+0xc63>
+  .byte  102,15,111,21,36,35,0,0             // movdqa        0x2324(%rip),%xmm2        # 5280 <_sk_callback_sse41+0xc73>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,38,35,0,0                  // mulps         0x2326(%rip),%xmm2        # 5290 <_sk_callback_sse41+0xc80>
-  .byte  102,15,219,29,46,35,0,0             // pand          0x232e(%rip),%xmm3        # 52a0 <_sk_callback_sse41+0xc90>
+  .byte  15,89,21,38,35,0,0                  // mulps         0x2326(%rip),%xmm2        # 5290 <_sk_callback_sse41+0xc83>
+  .byte  102,15,219,29,46,35,0,0             // pand          0x232e(%rip),%xmm3        # 52a0 <_sk_callback_sse41+0xc93>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,52,35,0,0                  // mulps         0x2334(%rip),%xmm3        # 52b0 <_sk_callback_sse41+0xca0>
+  .byte  15,89,29,52,35,0,0                  // mulps         0x2334(%rip),%xmm3        # 52b0 <_sk_callback_sse41+0xca3>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -23584,21 +23578,21 @@ _sk_gather_4444_sse41:
   .byte  65,15,183,4,65                      // movzwl        (%r9,%rax,2),%eax
   .byte  102,15,196,192,3                    // pinsrw        $0x3,%eax,%xmm0
   .byte  102,15,56,51,216                    // pmovzxwd      %xmm0,%xmm3
-  .byte  102,15,111,5,215,34,0,0             // movdqa        0x22d7(%rip),%xmm0        # 52c0 <_sk_callback_sse41+0xcb0>
+  .byte  102,15,111,5,215,34,0,0             // movdqa        0x22d7(%rip),%xmm0        # 52c0 <_sk_callback_sse41+0xcb3>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,217,34,0,0                  // mulps         0x22d9(%rip),%xmm0        # 52d0 <_sk_callback_sse41+0xcc0>
-  .byte  102,15,111,13,225,34,0,0            // movdqa        0x22e1(%rip),%xmm1        # 52e0 <_sk_callback_sse41+0xcd0>
+  .byte  15,89,5,217,34,0,0                  // mulps         0x22d9(%rip),%xmm0        # 52d0 <_sk_callback_sse41+0xcc3>
+  .byte  102,15,111,13,225,34,0,0            // movdqa        0x22e1(%rip),%xmm1        # 52e0 <_sk_callback_sse41+0xcd3>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,227,34,0,0                 // mulps         0x22e3(%rip),%xmm1        # 52f0 <_sk_callback_sse41+0xce0>
-  .byte  102,15,111,21,235,34,0,0            // movdqa        0x22eb(%rip),%xmm2        # 5300 <_sk_callback_sse41+0xcf0>
+  .byte  15,89,13,227,34,0,0                 // mulps         0x22e3(%rip),%xmm1        # 52f0 <_sk_callback_sse41+0xce3>
+  .byte  102,15,111,21,235,34,0,0            // movdqa        0x22eb(%rip),%xmm2        # 5300 <_sk_callback_sse41+0xcf3>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,237,34,0,0                 // mulps         0x22ed(%rip),%xmm2        # 5310 <_sk_callback_sse41+0xd00>
-  .byte  102,15,219,29,245,34,0,0            // pand          0x22f5(%rip),%xmm3        # 5320 <_sk_callback_sse41+0xd10>
+  .byte  15,89,21,237,34,0,0                 // mulps         0x22ed(%rip),%xmm2        # 5310 <_sk_callback_sse41+0xd03>
+  .byte  102,15,219,29,245,34,0,0            // pand          0x22f5(%rip),%xmm3        # 5320 <_sk_callback_sse41+0xd13>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,251,34,0,0                 // mulps         0x22fb(%rip),%xmm3        # 5330 <_sk_callback_sse41+0xd20>
+  .byte  15,89,29,251,34,0,0                 // mulps         0x22fb(%rip),%xmm3        # 5330 <_sk_callback_sse41+0xd23>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -23608,7 +23602,7 @@ FUNCTION(_sk_store_4444_sse41)
 _sk_store_4444_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,250,34,0,0               // movaps        0x22fa(%rip),%xmm8        # 5340 <_sk_callback_sse41+0xd30>
+  .byte  68,15,40,5,250,34,0,0               // movaps        0x22fa(%rip),%xmm8        # 5340 <_sk_callback_sse41+0xd33>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -23638,17 +23632,17 @@ _sk_load_8888_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  15,16,28,184                        // movups        (%rax,%rdi,4),%xmm3
-  .byte  15,40,5,153,34,0,0                  // movaps        0x2299(%rip),%xmm0        # 5350 <_sk_callback_sse41+0xd40>
+  .byte  15,40,5,153,34,0,0                  // movaps        0x2299(%rip),%xmm0        # 5350 <_sk_callback_sse41+0xd43>
   .byte  15,84,195                           // andps         %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,155,34,0,0               // movaps        0x229b(%rip),%xmm8        # 5360 <_sk_callback_sse41+0xd50>
+  .byte  68,15,40,5,155,34,0,0               // movaps        0x229b(%rip),%xmm8        # 5360 <_sk_callback_sse41+0xd53>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,40,203                           // movaps        %xmm3,%xmm1
-  .byte  102,15,56,0,13,155,34,0,0           // pshufb        0x229b(%rip),%xmm1        # 5370 <_sk_callback_sse41+0xd60>
+  .byte  102,15,56,0,13,155,34,0,0           // pshufb        0x229b(%rip),%xmm1        # 5370 <_sk_callback_sse41+0xd63>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  15,40,211                           // movaps        %xmm3,%xmm2
-  .byte  102,15,56,0,21,152,34,0,0           // pshufb        0x2298(%rip),%xmm2        # 5380 <_sk_callback_sse41+0xd70>
+  .byte  102,15,56,0,21,152,34,0,0           // pshufb        0x2298(%rip),%xmm2        # 5380 <_sk_callback_sse41+0xd73>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -23679,17 +23673,17 @@ _sk_gather_8888_sse41:
   .byte  102,65,15,58,34,28,129,1            // pinsrd        $0x1,(%r9,%rax,4),%xmm3
   .byte  102,67,15,58,34,28,145,2            // pinsrd        $0x2,(%r9,%r10,4),%xmm3
   .byte  102,65,15,58,34,28,137,3            // pinsrd        $0x3,(%r9,%rcx,4),%xmm3
-  .byte  102,15,111,5,49,34,0,0              // movdqa        0x2231(%rip),%xmm0        # 5390 <_sk_callback_sse41+0xd80>
+  .byte  102,15,111,5,49,34,0,0              // movdqa        0x2231(%rip),%xmm0        # 5390 <_sk_callback_sse41+0xd83>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,50,34,0,0                // movaps        0x2232(%rip),%xmm8        # 53a0 <_sk_callback_sse41+0xd90>
+  .byte  68,15,40,5,50,34,0,0                // movaps        0x2232(%rip),%xmm8        # 53a0 <_sk_callback_sse41+0xd93>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
-  .byte  102,15,56,0,13,49,34,0,0            // pshufb        0x2231(%rip),%xmm1        # 53b0 <_sk_callback_sse41+0xda0>
+  .byte  102,15,56,0,13,49,34,0,0            // pshufb        0x2231(%rip),%xmm1        # 53b0 <_sk_callback_sse41+0xda3>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,211                      // movdqa        %xmm3,%xmm2
-  .byte  102,15,56,0,21,45,34,0,0            // pshufb        0x222d(%rip),%xmm2        # 53c0 <_sk_callback_sse41+0xdb0>
+  .byte  102,15,56,0,21,45,34,0,0            // pshufb        0x222d(%rip),%xmm2        # 53c0 <_sk_callback_sse41+0xdb3>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -23704,7 +23698,7 @@ FUNCTION(_sk_store_8888_sse41)
 _sk_store_8888_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,25,34,0,0                // movaps        0x2219(%rip),%xmm8        # 53d0 <_sk_callback_sse41+0xdc0>
+  .byte  68,15,40,5,25,34,0,0                // movaps        0x2219(%rip),%xmm8        # 53d0 <_sk_callback_sse41+0xdc3>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -23741,18 +23735,18 @@ _sk_load_f16_sse41:
   .byte  102,68,15,97,216                    // punpcklwd     %xmm0,%xmm11
   .byte  102,68,15,105,200                   // punpckhwd     %xmm0,%xmm9
   .byte  102,65,15,56,51,203                 // pmovzxwd      %xmm11,%xmm1
-  .byte  102,68,15,111,5,146,33,0,0          // movdqa        0x2192(%rip),%xmm8        # 53e0 <_sk_callback_sse41+0xdd0>
+  .byte  102,68,15,111,5,146,33,0,0          // movdqa        0x2192(%rip),%xmm8        # 53e0 <_sk_callback_sse41+0xdd3>
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,65,15,219,208                   // pand          %xmm8,%xmm2
   .byte  102,15,239,202                      // pxor          %xmm2,%xmm1
-  .byte  102,15,111,29,141,33,0,0            // movdqa        0x218d(%rip),%xmm3        # 53f0 <_sk_callback_sse41+0xde0>
+  .byte  102,15,111,29,141,33,0,0            // movdqa        0x218d(%rip),%xmm3        # 53f0 <_sk_callback_sse41+0xde3>
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,56,63,195                    // pmaxud        %xmm3,%xmm0
   .byte  102,15,118,193                      // pcmpeqd       %xmm1,%xmm0
   .byte  102,15,114,241,13                   // pslld         $0xd,%xmm1
   .byte  102,15,235,202                      // por           %xmm2,%xmm1
-  .byte  102,68,15,111,21,121,33,0,0         // movdqa        0x2179(%rip),%xmm10        # 5400 <_sk_callback_sse41+0xdf0>
+  .byte  102,68,15,111,21,121,33,0,0         // movdqa        0x2179(%rip),%xmm10        # 5400 <_sk_callback_sse41+0xdf3>
   .byte  102,65,15,254,202                   // paddd         %xmm10,%xmm1
   .byte  102,15,219,193                      // pand          %xmm1,%xmm0
   .byte  102,65,15,115,219,8                 // psrldq        $0x8,%xmm11
@@ -23825,18 +23819,18 @@ _sk_gather_f16_sse41:
   .byte  102,68,15,97,218                    // punpcklwd     %xmm2,%xmm11
   .byte  102,68,15,105,202                   // punpckhwd     %xmm2,%xmm9
   .byte  102,65,15,56,51,203                 // pmovzxwd      %xmm11,%xmm1
-  .byte  102,68,15,111,5,55,32,0,0           // movdqa        0x2037(%rip),%xmm8        # 5410 <_sk_callback_sse41+0xe00>
+  .byte  102,68,15,111,5,55,32,0,0           // movdqa        0x2037(%rip),%xmm8        # 5410 <_sk_callback_sse41+0xe03>
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,65,15,219,208                   // pand          %xmm8,%xmm2
   .byte  102,15,239,202                      // pxor          %xmm2,%xmm1
-  .byte  102,15,111,29,50,32,0,0             // movdqa        0x2032(%rip),%xmm3        # 5420 <_sk_callback_sse41+0xe10>
+  .byte  102,15,111,29,50,32,0,0             // movdqa        0x2032(%rip),%xmm3        # 5420 <_sk_callback_sse41+0xe13>
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,56,63,195                    // pmaxud        %xmm3,%xmm0
   .byte  102,15,118,193                      // pcmpeqd       %xmm1,%xmm0
   .byte  102,15,114,241,13                   // pslld         $0xd,%xmm1
   .byte  102,15,235,202                      // por           %xmm2,%xmm1
-  .byte  102,68,15,111,21,30,32,0,0          // movdqa        0x201e(%rip),%xmm10        # 5430 <_sk_callback_sse41+0xe20>
+  .byte  102,68,15,111,21,30,32,0,0          // movdqa        0x201e(%rip),%xmm10        # 5430 <_sk_callback_sse41+0xe23>
   .byte  102,65,15,254,202                   // paddd         %xmm10,%xmm1
   .byte  102,15,219,193                      // pand          %xmm1,%xmm0
   .byte  102,65,15,115,219,8                 // psrldq        $0x8,%xmm11
@@ -23884,17 +23878,17 @@ FUNCTION(_sk_store_f16_sse41)
 _sk_store_f16_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  102,68,15,111,21,84,31,0,0          // movdqa        0x1f54(%rip),%xmm10        # 5440 <_sk_callback_sse41+0xe30>
+  .byte  102,68,15,111,21,84,31,0,0          // movdqa        0x1f54(%rip),%xmm10        # 5440 <_sk_callback_sse41+0xe33>
   .byte  102,68,15,111,224                   // movdqa        %xmm0,%xmm12
   .byte  102,68,15,111,232                   // movdqa        %xmm0,%xmm13
   .byte  102,69,15,219,234                   // pand          %xmm10,%xmm13
   .byte  102,69,15,239,229                   // pxor          %xmm13,%xmm12
-  .byte  102,68,15,111,13,71,31,0,0          // movdqa        0x1f47(%rip),%xmm9        # 5450 <_sk_callback_sse41+0xe40>
+  .byte  102,68,15,111,13,71,31,0,0          // movdqa        0x1f47(%rip),%xmm9        # 5450 <_sk_callback_sse41+0xe43>
   .byte  102,65,15,114,213,16                // psrld         $0x10,%xmm13
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
   .byte  102,69,15,102,196                   // pcmpgtd       %xmm12,%xmm8
   .byte  102,65,15,114,212,13                // psrld         $0xd,%xmm12
-  .byte  102,68,15,111,29,56,31,0,0          // movdqa        0x1f38(%rip),%xmm11        # 5460 <_sk_callback_sse41+0xe50>
+  .byte  102,68,15,111,29,56,31,0,0          // movdqa        0x1f38(%rip),%xmm11        # 5460 <_sk_callback_sse41+0xe53>
   .byte  102,69,15,235,235                   // por           %xmm11,%xmm13
   .byte  102,69,15,254,236                   // paddd         %xmm12,%xmm13
   .byte  102,69,15,223,197                   // pandn         %xmm13,%xmm8
@@ -23964,7 +23958,7 @@ _sk_load_u16_be_sse41:
   .byte  102,15,235,200                      // por           %xmm0,%xmm1
   .byte  102,15,56,51,193                    // pmovzxwd      %xmm1,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,7,30,0,0                 // movaps        0x1e07(%rip),%xmm8        # 5470 <_sk_callback_sse41+0xe60>
+  .byte  68,15,40,5,7,30,0,0                 // movaps        0x1e07(%rip),%xmm8        # 5470 <_sk_callback_sse41+0xe63>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -24016,7 +24010,7 @@ _sk_load_rgb_u16_be_sse41:
   .byte  102,15,235,193                      // por           %xmm1,%xmm0
   .byte  102,15,56,51,192                    // pmovzxwd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,72,29,0,0                // movaps        0x1d48(%rip),%xmm8        # 5480 <_sk_callback_sse41+0xe70>
+  .byte  68,15,40,5,72,29,0,0                // movaps        0x1d48(%rip),%xmm8        # 5480 <_sk_callback_sse41+0xe73>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -24033,7 +24027,7 @@ _sk_load_rgb_u16_be_sse41:
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,15,29,0,0                  // movaps        0x1d0f(%rip),%xmm3        # 5490 <_sk_callback_sse41+0xe80>
+  .byte  15,40,29,15,29,0,0                  // movaps        0x1d0f(%rip),%xmm3        # 5490 <_sk_callback_sse41+0xe83>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_u16_be_sse41
@@ -24042,7 +24036,7 @@ FUNCTION(_sk_store_u16_be_sse41)
 _sk_store_u16_be_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,13,16,29,0,0               // movaps        0x1d10(%rip),%xmm9        # 54a0 <_sk_callback_sse41+0xe90>
+  .byte  68,15,40,13,16,29,0,0               // movaps        0x1d10(%rip),%xmm9        # 54a0 <_sk_callback_sse41+0xe93>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
@@ -24253,10 +24247,10 @@ HIDDEN _sk_luminance_to_alpha_sse41
 FUNCTION(_sk_luminance_to_alpha_sse41)
 _sk_luminance_to_alpha_sse41:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
-  .byte  15,89,5,108,26,0,0                  // mulps         0x1a6c(%rip),%xmm0        # 54b0 <_sk_callback_sse41+0xea0>
-  .byte  15,89,13,117,26,0,0                 // mulps         0x1a75(%rip),%xmm1        # 54c0 <_sk_callback_sse41+0xeb0>
+  .byte  15,89,5,108,26,0,0                  // mulps         0x1a6c(%rip),%xmm0        # 54b0 <_sk_callback_sse41+0xea3>
+  .byte  15,89,13,117,26,0,0                 // mulps         0x1a75(%rip),%xmm1        # 54c0 <_sk_callback_sse41+0xeb3>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,29,123,26,0,0                 // mulps         0x1a7b(%rip),%xmm3        # 54d0 <_sk_callback_sse41+0xec0>
+  .byte  15,89,29,123,26,0,0                 // mulps         0x1a7b(%rip),%xmm3        # 54d0 <_sk_callback_sse41+0xec3>
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -24580,7 +24574,7 @@ _sk_gradient_sse41:
   .byte  73,255,200                          // dec           %r8
   .byte  72,131,193,4                        // add           $0x4,%rcx
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
-  .byte  15,40,21,48,21,0,0                  // movaps        0x1530(%rip),%xmm2        # 54e0 <_sk_callback_sse41+0xed0>
+  .byte  15,40,21,48,21,0,0                  // movaps        0x1530(%rip),%xmm2        # 54e0 <_sk_callback_sse41+0xed3>
   .byte  243,15,16,25                        // movss         (%rcx),%xmm3
   .byte  15,198,219,0                        // shufps        $0x0,%xmm3,%xmm3
   .byte  15,194,216,2                        // cmpleps       %xmm0,%xmm3
@@ -24719,26 +24713,26 @@ _sk_xy_to_unit_angle_sse41:
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,40,236                        // movaps        %xmm12,%xmm13
   .byte  69,15,89,237                        // mulps         %xmm13,%xmm13
-  .byte  68,15,40,21,210,18,0,0              // movaps        0x12d2(%rip),%xmm10        # 54f0 <_sk_callback_sse41+0xee0>
+  .byte  68,15,40,21,210,18,0,0              // movaps        0x12d2(%rip),%xmm10        # 54f0 <_sk_callback_sse41+0xee3>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,214,18,0,0              // addps         0x12d6(%rip),%xmm10        # 5500 <_sk_callback_sse41+0xef0>
+  .byte  68,15,88,21,214,18,0,0              // addps         0x12d6(%rip),%xmm10        # 5500 <_sk_callback_sse41+0xef3>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,218,18,0,0              // addps         0x12da(%rip),%xmm10        # 5510 <_sk_callback_sse41+0xf00>
+  .byte  68,15,88,21,218,18,0,0              // addps         0x12da(%rip),%xmm10        # 5510 <_sk_callback_sse41+0xf03>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,222,18,0,0              // addps         0x12de(%rip),%xmm10        # 5520 <_sk_callback_sse41+0xf10>
+  .byte  68,15,88,21,222,18,0,0              // addps         0x12de(%rip),%xmm10        # 5520 <_sk_callback_sse41+0xf13>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,194,195,1                     // cmpltps       %xmm11,%xmm0
-  .byte  68,15,40,29,221,18,0,0              // movaps        0x12dd(%rip),%xmm11        # 5530 <_sk_callback_sse41+0xf20>
+  .byte  68,15,40,29,221,18,0,0              // movaps        0x12dd(%rip),%xmm11        # 5530 <_sk_callback_sse41+0xf23>
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  102,69,15,56,20,211                 // blendvps      %xmm0,%xmm11,%xmm10
   .byte  69,15,194,200,1                     // cmpltps       %xmm8,%xmm9
-  .byte  68,15,40,29,214,18,0,0              // movaps        0x12d6(%rip),%xmm11        # 5540 <_sk_callback_sse41+0xf30>
+  .byte  68,15,40,29,214,18,0,0              // movaps        0x12d6(%rip),%xmm11        # 5540 <_sk_callback_sse41+0xf33>
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  102,69,15,56,20,211                 // blendvps      %xmm0,%xmm11,%xmm10
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  65,15,194,192,1                     // cmpltps       %xmm8,%xmm0
-  .byte  68,15,40,13,200,18,0,0              // movaps        0x12c8(%rip),%xmm9        # 5550 <_sk_callback_sse41+0xf40>
+  .byte  68,15,40,13,200,18,0,0              // movaps        0x12c8(%rip),%xmm9        # 5550 <_sk_callback_sse41+0xf43>
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
   .byte  102,69,15,56,20,209                 // blendvps      %xmm0,%xmm9,%xmm10
   .byte  69,15,194,194,7                     // cmpordps      %xmm10,%xmm8
@@ -24755,8 +24749,7 @@ _sk_xy_to_radius_sse41:
   .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
   .byte  68,15,88,192                        // addps         %xmm0,%xmm8
-  .byte  65,15,82,192                        // rsqrtps       %xmm8,%xmm0
-  .byte  15,83,192                           // rcpps         %xmm0,%xmm0
+  .byte  65,15,81,192                        // sqrtps        %xmm8,%xmm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -24765,7 +24758,7 @@ HIDDEN _sk_save_xy_sse41
 FUNCTION(_sk_save_xy_sse41)
 _sk_save_xy_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,153,18,0,0               // movaps        0x1299(%rip),%xmm8        # 5560 <_sk_callback_sse41+0xf50>
+  .byte  68,15,40,5,156,18,0,0               // movaps        0x129c(%rip),%xmm8        # 5560 <_sk_callback_sse41+0xf53>
   .byte  15,17,0                             // movups        %xmm0,(%rax)
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,88,200                        // addps         %xmm8,%xmm9
@@ -24809,8 +24802,8 @@ _sk_bilinear_nx_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,27,18,0,0                   // addps         0x121b(%rip),%xmm0        # 5570 <_sk_callback_sse41+0xf60>
-  .byte  68,15,40,13,35,18,0,0               // movaps        0x1223(%rip),%xmm9        # 5580 <_sk_callback_sse41+0xf70>
+  .byte  15,88,5,30,18,0,0                   // addps         0x121e(%rip),%xmm0        # 5570 <_sk_callback_sse41+0xf63>
+  .byte  68,15,40,13,38,18,0,0               // movaps        0x1226(%rip),%xmm9        # 5580 <_sk_callback_sse41+0xf73>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,128,0,0,0              // movups        %xmm9,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -24823,7 +24816,7 @@ _sk_bilinear_px_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,18,18,0,0                   // addps         0x1212(%rip),%xmm0        # 5590 <_sk_callback_sse41+0xf80>
+  .byte  15,88,5,21,18,0,0                   // addps         0x1215(%rip),%xmm0        # 5590 <_sk_callback_sse41+0xf83>
   .byte  68,15,17,128,128,0,0,0              // movups        %xmm8,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -24835,8 +24828,8 @@ _sk_bilinear_ny_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,4,18,0,0                   // addps         0x1204(%rip),%xmm1        # 55a0 <_sk_callback_sse41+0xf90>
-  .byte  68,15,40,13,12,18,0,0               // movaps        0x120c(%rip),%xmm9        # 55b0 <_sk_callback_sse41+0xfa0>
+  .byte  15,88,13,7,18,0,0                   // addps         0x1207(%rip),%xmm1        # 55a0 <_sk_callback_sse41+0xf93>
+  .byte  68,15,40,13,15,18,0,0               // movaps        0x120f(%rip),%xmm9        # 55b0 <_sk_callback_sse41+0xfa3>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,160,0,0,0              // movups        %xmm9,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -24849,7 +24842,7 @@ _sk_bilinear_py_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,250,17,0,0                 // addps         0x11fa(%rip),%xmm1        # 55c0 <_sk_callback_sse41+0xfb0>
+  .byte  15,88,13,253,17,0,0                 // addps         0x11fd(%rip),%xmm1        # 55c0 <_sk_callback_sse41+0xfb3>
   .byte  68,15,17,128,160,0,0,0              // movups        %xmm8,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -24861,13 +24854,13 @@ _sk_bicubic_n3x_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,237,17,0,0                  // addps         0x11ed(%rip),%xmm0        # 55d0 <_sk_callback_sse41+0xfc0>
-  .byte  68,15,40,13,245,17,0,0              // movaps        0x11f5(%rip),%xmm9        # 55e0 <_sk_callback_sse41+0xfd0>
+  .byte  15,88,5,240,17,0,0                  // addps         0x11f0(%rip),%xmm0        # 55d0 <_sk_callback_sse41+0xfc3>
+  .byte  68,15,40,13,248,17,0,0              // movaps        0x11f8(%rip),%xmm9        # 55e0 <_sk_callback_sse41+0xfd3>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,241,17,0,0              // mulps         0x11f1(%rip),%xmm9        # 55f0 <_sk_callback_sse41+0xfe0>
-  .byte  68,15,88,13,249,17,0,0              // addps         0x11f9(%rip),%xmm9        # 5600 <_sk_callback_sse41+0xff0>
+  .byte  68,15,89,13,244,17,0,0              // mulps         0x11f4(%rip),%xmm9        # 55f0 <_sk_callback_sse41+0xfe3>
+  .byte  68,15,88,13,252,17,0,0              // addps         0x11fc(%rip),%xmm9        # 5600 <_sk_callback_sse41+0xff3>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,128,0,0,0              // movups        %xmm9,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -24880,16 +24873,16 @@ _sk_bicubic_n1x_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,232,17,0,0                  // addps         0x11e8(%rip),%xmm0        # 5610 <_sk_callback_sse41+0x1000>
-  .byte  68,15,40,13,240,17,0,0              // movaps        0x11f0(%rip),%xmm9        # 5620 <_sk_callback_sse41+0x1010>
+  .byte  15,88,5,235,17,0,0                  // addps         0x11eb(%rip),%xmm0        # 5610 <_sk_callback_sse41+0x1003>
+  .byte  68,15,40,13,243,17,0,0              // movaps        0x11f3(%rip),%xmm9        # 5620 <_sk_callback_sse41+0x1013>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,244,17,0,0               // movaps        0x11f4(%rip),%xmm8        # 5630 <_sk_callback_sse41+0x1020>
+  .byte  68,15,40,5,247,17,0,0               // movaps        0x11f7(%rip),%xmm8        # 5630 <_sk_callback_sse41+0x1023>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,248,17,0,0               // addps         0x11f8(%rip),%xmm8        # 5640 <_sk_callback_sse41+0x1030>
+  .byte  68,15,88,5,251,17,0,0               // addps         0x11fb(%rip),%xmm8        # 5640 <_sk_callback_sse41+0x1033>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,252,17,0,0               // addps         0x11fc(%rip),%xmm8        # 5650 <_sk_callback_sse41+0x1040>
+  .byte  68,15,88,5,255,17,0,0               // addps         0x11ff(%rip),%xmm8        # 5650 <_sk_callback_sse41+0x1043>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,0,18,0,0                 // addps         0x1200(%rip),%xmm8        # 5660 <_sk_callback_sse41+0x1050>
+  .byte  68,15,88,5,3,18,0,0                 // addps         0x1203(%rip),%xmm8        # 5660 <_sk_callback_sse41+0x1053>
   .byte  68,15,17,128,128,0,0,0              // movups        %xmm8,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -24899,17 +24892,17 @@ HIDDEN _sk_bicubic_p1x_sse41
 FUNCTION(_sk_bicubic_p1x_sse41)
 _sk_bicubic_p1x_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,250,17,0,0               // movaps        0x11fa(%rip),%xmm8        # 5670 <_sk_callback_sse41+0x1060>
+  .byte  68,15,40,5,253,17,0,0               // movaps        0x11fd(%rip),%xmm8        # 5670 <_sk_callback_sse41+0x1063>
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,72,64                      // movups        0x40(%rax),%xmm9
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,21,246,17,0,0              // movaps        0x11f6(%rip),%xmm10        # 5680 <_sk_callback_sse41+0x1070>
+  .byte  68,15,40,21,249,17,0,0              // movaps        0x11f9(%rip),%xmm10        # 5680 <_sk_callback_sse41+0x1073>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,250,17,0,0              // addps         0x11fa(%rip),%xmm10        # 5690 <_sk_callback_sse41+0x1080>
+  .byte  68,15,88,21,253,17,0,0              // addps         0x11fd(%rip),%xmm10        # 5690 <_sk_callback_sse41+0x1083>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,246,17,0,0              // addps         0x11f6(%rip),%xmm10        # 56a0 <_sk_callback_sse41+0x1090>
+  .byte  68,15,88,21,249,17,0,0              // addps         0x11f9(%rip),%xmm10        # 56a0 <_sk_callback_sse41+0x1093>
   .byte  68,15,17,144,128,0,0,0              // movups        %xmm10,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -24921,11 +24914,11 @@ _sk_bicubic_p3x_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,233,17,0,0                  // addps         0x11e9(%rip),%xmm0        # 56b0 <_sk_callback_sse41+0x10a0>
+  .byte  15,88,5,236,17,0,0                  // addps         0x11ec(%rip),%xmm0        # 56b0 <_sk_callback_sse41+0x10a3>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,233,17,0,0               // mulps         0x11e9(%rip),%xmm8        # 56c0 <_sk_callback_sse41+0x10b0>
-  .byte  68,15,88,5,241,17,0,0               // addps         0x11f1(%rip),%xmm8        # 56d0 <_sk_callback_sse41+0x10c0>
+  .byte  68,15,89,5,236,17,0,0               // mulps         0x11ec(%rip),%xmm8        # 56c0 <_sk_callback_sse41+0x10b3>
+  .byte  68,15,88,5,244,17,0,0               // addps         0x11f4(%rip),%xmm8        # 56d0 <_sk_callback_sse41+0x10c3>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,128,0,0,0              // movups        %xmm8,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -24938,13 +24931,13 @@ _sk_bicubic_n3y_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,223,17,0,0                 // addps         0x11df(%rip),%xmm1        # 56e0 <_sk_callback_sse41+0x10d0>
-  .byte  68,15,40,13,231,17,0,0              // movaps        0x11e7(%rip),%xmm9        # 56f0 <_sk_callback_sse41+0x10e0>
+  .byte  15,88,13,226,17,0,0                 // addps         0x11e2(%rip),%xmm1        # 56e0 <_sk_callback_sse41+0x10d3>
+  .byte  68,15,40,13,234,17,0,0              // movaps        0x11ea(%rip),%xmm9        # 56f0 <_sk_callback_sse41+0x10e3>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,227,17,0,0              // mulps         0x11e3(%rip),%xmm9        # 5700 <_sk_callback_sse41+0x10f0>
-  .byte  68,15,88,13,235,17,0,0              // addps         0x11eb(%rip),%xmm9        # 5710 <_sk_callback_sse41+0x1100>
+  .byte  68,15,89,13,230,17,0,0              // mulps         0x11e6(%rip),%xmm9        # 5700 <_sk_callback_sse41+0x10f3>
+  .byte  68,15,88,13,238,17,0,0              // addps         0x11ee(%rip),%xmm9        # 5710 <_sk_callback_sse41+0x1103>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,160,0,0,0              // movups        %xmm9,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -24957,16 +24950,16 @@ _sk_bicubic_n1y_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,217,17,0,0                 // addps         0x11d9(%rip),%xmm1        # 5720 <_sk_callback_sse41+0x1110>
-  .byte  68,15,40,13,225,17,0,0              // movaps        0x11e1(%rip),%xmm9        # 5730 <_sk_callback_sse41+0x1120>
+  .byte  15,88,13,220,17,0,0                 // addps         0x11dc(%rip),%xmm1        # 5720 <_sk_callback_sse41+0x1113>
+  .byte  68,15,40,13,228,17,0,0              // movaps        0x11e4(%rip),%xmm9        # 5730 <_sk_callback_sse41+0x1123>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,229,17,0,0               // movaps        0x11e5(%rip),%xmm8        # 5740 <_sk_callback_sse41+0x1130>
+  .byte  68,15,40,5,232,17,0,0               // movaps        0x11e8(%rip),%xmm8        # 5740 <_sk_callback_sse41+0x1133>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,233,17,0,0               // addps         0x11e9(%rip),%xmm8        # 5750 <_sk_callback_sse41+0x1140>
+  .byte  68,15,88,5,236,17,0,0               // addps         0x11ec(%rip),%xmm8        # 5750 <_sk_callback_sse41+0x1143>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,237,17,0,0               // addps         0x11ed(%rip),%xmm8        # 5760 <_sk_callback_sse41+0x1150>
+  .byte  68,15,88,5,240,17,0,0               // addps         0x11f0(%rip),%xmm8        # 5760 <_sk_callback_sse41+0x1153>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,241,17,0,0               // addps         0x11f1(%rip),%xmm8        # 5770 <_sk_callback_sse41+0x1160>
+  .byte  68,15,88,5,244,17,0,0               // addps         0x11f4(%rip),%xmm8        # 5770 <_sk_callback_sse41+0x1163>
   .byte  68,15,17,128,160,0,0,0              // movups        %xmm8,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -24976,17 +24969,17 @@ HIDDEN _sk_bicubic_p1y_sse41
 FUNCTION(_sk_bicubic_p1y_sse41)
 _sk_bicubic_p1y_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,235,17,0,0               // movaps        0x11eb(%rip),%xmm8        # 5780 <_sk_callback_sse41+0x1170>
+  .byte  68,15,40,5,238,17,0,0               // movaps        0x11ee(%rip),%xmm8        # 5780 <_sk_callback_sse41+0x1173>
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,72,96                      // movups        0x60(%rax),%xmm9
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  68,15,40,21,230,17,0,0              // movaps        0x11e6(%rip),%xmm10        # 5790 <_sk_callback_sse41+0x1180>
+  .byte  68,15,40,21,233,17,0,0              // movaps        0x11e9(%rip),%xmm10        # 5790 <_sk_callback_sse41+0x1183>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,234,17,0,0              // addps         0x11ea(%rip),%xmm10        # 57a0 <_sk_callback_sse41+0x1190>
+  .byte  68,15,88,21,237,17,0,0              // addps         0x11ed(%rip),%xmm10        # 57a0 <_sk_callback_sse41+0x1193>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,230,17,0,0              // addps         0x11e6(%rip),%xmm10        # 57b0 <_sk_callback_sse41+0x11a0>
+  .byte  68,15,88,21,233,17,0,0              // addps         0x11e9(%rip),%xmm10        # 57b0 <_sk_callback_sse41+0x11a3>
   .byte  68,15,17,144,160,0,0,0              // movups        %xmm10,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -24998,11 +24991,11 @@ _sk_bicubic_p3y_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,216,17,0,0                 // addps         0x11d8(%rip),%xmm1        # 57c0 <_sk_callback_sse41+0x11b0>
+  .byte  15,88,13,219,17,0,0                 // addps         0x11db(%rip),%xmm1        # 57c0 <_sk_callback_sse41+0x11b3>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,216,17,0,0               // mulps         0x11d8(%rip),%xmm8        # 57d0 <_sk_callback_sse41+0x11c0>
-  .byte  68,15,88,5,224,17,0,0               // addps         0x11e0(%rip),%xmm8        # 57e0 <_sk_callback_sse41+0x11d0>
+  .byte  68,15,89,5,219,17,0,0               // mulps         0x11db(%rip),%xmm8        # 57d0 <_sk_callback_sse41+0x11c3>
+  .byte  68,15,88,5,227,17,0,0               // addps         0x11e3(%rip),%xmm8        # 57e0 <_sk_callback_sse41+0x11d3>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,160,0,0,0              // movups        %xmm8,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -25608,10 +25601,10 @@ BALIGN16
   .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)        # a004bd8 <_sk_callback_sse41+0xa0005c8>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004bd8 <_sk_callback_sse41+0xa0005cb>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3004be0 <_sk_callback_sse41+0x30005d0>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3004be0 <_sk_callback_sse41+0x30005d3>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -25726,7 +25719,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38cda <_sk_callback_sse41+0xffffffffe9a346ca>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38cda <_sk_callback_sse41+0xffffffffe9a346cd>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -25822,7 +25815,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38daa <_sk_callback_sse41+0xffffffffe9a3479a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38daa <_sk_callback_sse41+0xffffffffe9a3479d>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -25918,7 +25911,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38e7a <_sk_callback_sse41+0xffffffffe9a3486a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38e7a <_sk_callback_sse41+0xffffffffe9a3486d>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -26014,7 +26007,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38f4a <_sk_callback_sse41+0xffffffffe9a3493a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38f4a <_sk_callback_sse41+0xffffffffe9a3493d>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -26172,7 +26165,7 @@ BALIGN16
   .byte  5,255,255,255,9                     // add           $0x9ffffff,%eax
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3005130 <_sk_callback_sse41+0x3000b20>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3005130 <_sk_callback_sse41+0x3000b23>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -26414,7 +26407,7 @@ BALIGN16
   .byte  5,255,255,255,9                     // add           $0x9ffffff,%eax
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3005380 <_sk_callback_sse41+0x3000d70>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3005380 <_sk_callback_sse41+0x3000d73>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -26441,7 +26434,7 @@ BALIGN16
   .byte  5,255,255,255,9                     // add           $0x9ffffff,%eax
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 30053c0 <_sk_callback_sse41+0x3000db0>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 30053c0 <_sk_callback_sse41+0x3000db3>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -26724,7 +26717,7 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e4a5 <_sk_callback_sse41+0x3d639e95>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e4a5 <_sk_callback_sse41+0x3d639e98>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -26750,7 +26743,7 @@ BALIGN16
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e4e5 <_sk_callback_sse41+0x3d639ed5>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e4e5 <_sk_callback_sse41+0x3d639ed8>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -26834,7 +26827,7 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e5b5 <_sk_callback_sse41+0x3d639fa5>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e5b5 <_sk_callback_sse41+0x3d639fa8>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -26860,7 +26853,7 @@ BALIGN16
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e5f5 <_sk_callback_sse41+0x3d639fe5>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e5f5 <_sk_callback_sse41+0x3d639fe8>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -26871,11 +26864,11 @@ BALIGN16
   .byte  63                                  // (bad)
   .byte  114,28                              // jb            57ee <.literal16+0x10fe>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         57f2 <_sk_callback_sse41+0x11e2>
+  .byte  62,114,28                           // jb,pt         57f2 <_sk_callback_sse41+0x11e5>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         57f6 <_sk_callback_sse41+0x11e6>
+  .byte  62,114,28                           // jb,pt         57f6 <_sk_callback_sse41+0x11e9>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         57fa <_sk_callback_sse41+0x11ea>
+  .byte  62,114,28                           // jb,pt         57fa <_sk_callback_sse41+0x11ed>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -26945,7 +26938,7 @@ _sk_seed_shader_sse2:
   .byte  102,15,110,199                      // movd          %edi,%xmm0
   .byte  102,15,112,192,0                    // pshufd        $0x0,%xmm0,%xmm0
   .byte  15,91,200                           // cvtdq2ps      %xmm0,%xmm1
-  .byte  15,40,21,228,74,0,0                 // movaps        0x4ae4(%rip),%xmm2        # 4b60 <_sk_callback_sse2+0xdc>
+  .byte  15,40,21,228,74,0,0                 // movaps        0x4ae4(%rip),%xmm2        # 4b60 <_sk_callback_sse2+0xdf>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  15,16,2                             // movups        (%rdx),%xmm0
   .byte  15,88,193                           // addps         %xmm1,%xmm0
@@ -26954,7 +26947,7 @@ _sk_seed_shader_sse2:
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,21,211,74,0,0                 // movaps        0x4ad3(%rip),%xmm2        # 4b70 <_sk_callback_sse2+0xec>
+  .byte  15,40,21,211,74,0,0                 // movaps        0x4ad3(%rip),%xmm2        # 4b70 <_sk_callback_sse2+0xef>
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
@@ -26977,14 +26970,14 @@ _sk_dither_sse2:
   .byte  102,68,15,110,1                     // movd          (%rcx),%xmm8
   .byte  102,69,15,112,192,0                 // pshufd        $0x0,%xmm8,%xmm8
   .byte  102,69,15,239,193                   // pxor          %xmm9,%xmm8
-  .byte  102,68,15,111,21,152,74,0,0         // movdqa        0x4a98(%rip),%xmm10        # 4b80 <_sk_callback_sse2+0xfc>
+  .byte  102,68,15,111,21,152,74,0,0         // movdqa        0x4a98(%rip),%xmm10        # 4b80 <_sk_callback_sse2+0xff>
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,219,218                   // pand          %xmm10,%xmm11
   .byte  102,65,15,114,243,5                 // pslld         $0x5,%xmm11
   .byte  102,69,15,219,209                   // pand          %xmm9,%xmm10
   .byte  102,65,15,114,242,4                 // pslld         $0x4,%xmm10
-  .byte  102,68,15,111,37,132,74,0,0         // movdqa        0x4a84(%rip),%xmm12        # 4b90 <_sk_callback_sse2+0x10c>
-  .byte  102,68,15,111,45,139,74,0,0         // movdqa        0x4a8b(%rip),%xmm13        # 4ba0 <_sk_callback_sse2+0x11c>
+  .byte  102,68,15,111,37,132,74,0,0         // movdqa        0x4a84(%rip),%xmm12        # 4b90 <_sk_callback_sse2+0x10f>
+  .byte  102,68,15,111,45,139,74,0,0         // movdqa        0x4a8b(%rip),%xmm13        # 4ba0 <_sk_callback_sse2+0x11f>
   .byte  102,69,15,111,240                   // movdqa        %xmm8,%xmm14
   .byte  102,69,15,219,245                   // pand          %xmm13,%xmm14
   .byte  102,65,15,114,246,2                 // pslld         $0x2,%xmm14
@@ -27000,8 +26993,8 @@ _sk_dither_sse2:
   .byte  102,69,15,235,245                   // por           %xmm13,%xmm14
   .byte  102,69,15,235,240                   // por           %xmm8,%xmm14
   .byte  69,15,91,198                        // cvtdq2ps      %xmm14,%xmm8
-  .byte  68,15,89,5,70,74,0,0                // mulps         0x4a46(%rip),%xmm8        # 4bb0 <_sk_callback_sse2+0x12c>
-  .byte  68,15,88,5,78,74,0,0                // addps         0x4a4e(%rip),%xmm8        # 4bc0 <_sk_callback_sse2+0x13c>
+  .byte  68,15,89,5,70,74,0,0                // mulps         0x4a46(%rip),%xmm8        # 4bb0 <_sk_callback_sse2+0x12f>
+  .byte  68,15,88,5,78,74,0,0                // addps         0x4a4e(%rip),%xmm8        # 4bc0 <_sk_callback_sse2+0x13f>
   .byte  243,68,15,16,72,8                   // movss         0x8(%rax),%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
@@ -27067,7 +27060,7 @@ HIDDEN _sk_srcatop_sse2
 FUNCTION(_sk_srcatop_sse2)
 _sk_srcatop_sse2:
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  68,15,40,5,209,73,0,0               // movaps        0x49d1(%rip),%xmm8        # 4bd0 <_sk_callback_sse2+0x14c>
+  .byte  68,15,40,5,209,73,0,0               // movaps        0x49d1(%rip),%xmm8        # 4bd0 <_sk_callback_sse2+0x14f>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -27092,7 +27085,7 @@ FUNCTION(_sk_dstatop_sse2)
 _sk_dstatop_sse2:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
   .byte  68,15,89,196                        // mulps         %xmm4,%xmm8
-  .byte  68,15,40,13,148,73,0,0              // movaps        0x4994(%rip),%xmm9        # 4be0 <_sk_callback_sse2+0x15c>
+  .byte  68,15,40,13,148,73,0,0              // movaps        0x4994(%rip),%xmm9        # 4be0 <_sk_callback_sse2+0x15f>
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
@@ -27139,7 +27132,7 @@ HIDDEN _sk_srcout_sse2
 .globl _sk_srcout_sse2
 FUNCTION(_sk_srcout_sse2)
 _sk_srcout_sse2:
-  .byte  68,15,40,5,56,73,0,0                // movaps        0x4938(%rip),%xmm8        # 4bf0 <_sk_callback_sse2+0x16c>
+  .byte  68,15,40,5,56,73,0,0                // movaps        0x4938(%rip),%xmm8        # 4bf0 <_sk_callback_sse2+0x16f>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
@@ -27152,7 +27145,7 @@ HIDDEN _sk_dstout_sse2
 .globl _sk_dstout_sse2
 FUNCTION(_sk_dstout_sse2)
 _sk_dstout_sse2:
-  .byte  68,15,40,5,40,73,0,0                // movaps        0x4928(%rip),%xmm8        # 4c00 <_sk_callback_sse2+0x17c>
+  .byte  68,15,40,5,40,73,0,0                // movaps        0x4928(%rip),%xmm8        # 4c00 <_sk_callback_sse2+0x17f>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -27169,7 +27162,7 @@ HIDDEN _sk_srcover_sse2
 .globl _sk_srcover_sse2
 FUNCTION(_sk_srcover_sse2)
 _sk_srcover_sse2:
-  .byte  68,15,40,5,11,73,0,0                // movaps        0x490b(%rip),%xmm8        # 4c10 <_sk_callback_sse2+0x18c>
+  .byte  68,15,40,5,11,73,0,0                // movaps        0x490b(%rip),%xmm8        # 4c10 <_sk_callback_sse2+0x18f>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -27189,7 +27182,7 @@ HIDDEN _sk_dstover_sse2
 .globl _sk_dstover_sse2
 FUNCTION(_sk_dstover_sse2)
 _sk_dstover_sse2:
-  .byte  68,15,40,5,223,72,0,0               // movaps        0x48df(%rip),%xmm8        # 4c20 <_sk_callback_sse2+0x19c>
+  .byte  68,15,40,5,223,72,0,0               // movaps        0x48df(%rip),%xmm8        # 4c20 <_sk_callback_sse2+0x19f>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -27217,7 +27210,7 @@ HIDDEN _sk_multiply_sse2
 .globl _sk_multiply_sse2
 FUNCTION(_sk_multiply_sse2)
 _sk_multiply_sse2:
-  .byte  68,15,40,5,179,72,0,0               // movaps        0x48b3(%rip),%xmm8        # 4c30 <_sk_callback_sse2+0x1ac>
+  .byte  68,15,40,5,179,72,0,0               // movaps        0x48b3(%rip),%xmm8        # 4c30 <_sk_callback_sse2+0x1af>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
@@ -27293,7 +27286,7 @@ HIDDEN _sk_xor__sse2
 FUNCTION(_sk_xor__sse2)
 _sk_xor__sse2:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
-  .byte  15,40,29,228,71,0,0                 // movaps        0x47e4(%rip),%xmm3        # 4c40 <_sk_callback_sse2+0x1bc>
+  .byte  15,40,29,228,71,0,0                 // movaps        0x47e4(%rip),%xmm3        # 4c40 <_sk_callback_sse2+0x1bf>
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
@@ -27341,7 +27334,7 @@ _sk_darken_sse2:
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,95,209                        // maxps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,79,71,0,0                  // movaps        0x474f(%rip),%xmm2        # 4c50 <_sk_callback_sse2+0x1cc>
+  .byte  15,40,21,79,71,0,0                  // movaps        0x474f(%rip),%xmm2        # 4c50 <_sk_callback_sse2+0x1cf>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -27375,7 +27368,7 @@ _sk_lighten_sse2:
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,244,70,0,0                 // movaps        0x46f4(%rip),%xmm2        # 4c60 <_sk_callback_sse2+0x1dc>
+  .byte  15,40,21,244,70,0,0                 // movaps        0x46f4(%rip),%xmm2        # 4c60 <_sk_callback_sse2+0x1df>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -27412,7 +27405,7 @@ _sk_difference_sse2:
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,142,70,0,0                 // movaps        0x468e(%rip),%xmm2        # 4c70 <_sk_callback_sse2+0x1ec>
+  .byte  15,40,21,142,70,0,0                 // movaps        0x468e(%rip),%xmm2        # 4c70 <_sk_callback_sse2+0x1ef>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -27439,7 +27432,7 @@ _sk_exclusion_sse2:
   .byte  15,89,214                           // mulps         %xmm6,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,202                        // subps         %xmm2,%xmm9
-  .byte  15,40,13,79,70,0,0                  // movaps        0x464f(%rip),%xmm1        # 4c80 <_sk_callback_sse2+0x1fc>
+  .byte  15,40,13,79,70,0,0                  // movaps        0x464f(%rip),%xmm1        # 4c80 <_sk_callback_sse2+0x1ff>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,207                           // mulps         %xmm7,%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -27453,7 +27446,7 @@ HIDDEN _sk_colorburn_sse2
 FUNCTION(_sk_colorburn_sse2)
 _sk_colorburn_sse2:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,62,70,0,0               // movaps        0x463e(%rip),%xmm10        # 4c90 <_sk_callback_sse2+0x20c>
+  .byte  68,15,40,21,62,70,0,0               // movaps        0x463e(%rip),%xmm10        # 4c90 <_sk_callback_sse2+0x20f>
   .byte  69,15,40,202                        // movaps        %xmm10,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,217                        // movaps        %xmm9,%xmm11
@@ -27547,7 +27540,7 @@ HIDDEN _sk_colordodge_sse2
 FUNCTION(_sk_colordodge_sse2)
 _sk_colordodge_sse2:
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
-  .byte  68,15,40,21,244,68,0,0              // movaps        0x44f4(%rip),%xmm10        # 4ca0 <_sk_callback_sse2+0x21c>
+  .byte  68,15,40,21,244,68,0,0              // movaps        0x44f4(%rip),%xmm10        # 4ca0 <_sk_callback_sse2+0x21f>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -27641,7 +27634,7 @@ _sk_hardlight_sse2:
   .byte  15,41,116,36,232                    // movaps        %xmm6,-0x18(%rsp)
   .byte  15,40,245                           // movaps        %xmm5,%xmm6
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
-  .byte  68,15,40,29,169,67,0,0              // movaps        0x43a9(%rip),%xmm11        # 4cb0 <_sk_callback_sse2+0x22c>
+  .byte  68,15,40,29,169,67,0,0              // movaps        0x43a9(%rip),%xmm11        # 4cb0 <_sk_callback_sse2+0x22f>
   .byte  69,15,40,211                        // movaps        %xmm11,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,194                        // movaps        %xmm10,%xmm8
@@ -27729,7 +27722,7 @@ FUNCTION(_sk_overlay_sse2)
 _sk_overlay_sse2:
   .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
   .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
-  .byte  68,15,40,13,119,66,0,0              // movaps        0x4277(%rip),%xmm9        # 4cc0 <_sk_callback_sse2+0x23c>
+  .byte  68,15,40,13,119,66,0,0              // movaps        0x4277(%rip),%xmm9        # 4cc0 <_sk_callback_sse2+0x23f>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -27820,7 +27813,7 @@ _sk_softlight_sse2:
   .byte  68,15,40,213                        // movaps        %xmm5,%xmm10
   .byte  68,15,94,215                        // divps         %xmm7,%xmm10
   .byte  69,15,84,212                        // andps         %xmm12,%xmm10
-  .byte  68,15,40,13,52,65,0,0               // movaps        0x4134(%rip),%xmm9        # 4cd0 <_sk_callback_sse2+0x24c>
+  .byte  68,15,40,13,52,65,0,0               // movaps        0x4134(%rip),%xmm9        # 4cd0 <_sk_callback_sse2+0x24f>
   .byte  69,15,40,249                        // movaps        %xmm9,%xmm15
   .byte  69,15,92,250                        // subps         %xmm10,%xmm15
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -27833,10 +27826,10 @@ _sk_softlight_sse2:
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  65,15,88,194                        // addps         %xmm10,%xmm0
-  .byte  68,15,40,53,14,65,0,0               // movaps        0x410e(%rip),%xmm14        # 4ce0 <_sk_callback_sse2+0x25c>
+  .byte  68,15,40,53,14,65,0,0               // movaps        0x410e(%rip),%xmm14        # 4ce0 <_sk_callback_sse2+0x25f>
   .byte  69,15,88,222                        // addps         %xmm14,%xmm11
   .byte  68,15,89,216                        // mulps         %xmm0,%xmm11
-  .byte  68,15,40,21,14,65,0,0               // movaps        0x410e(%rip),%xmm10        # 4cf0 <_sk_callback_sse2+0x26c>
+  .byte  68,15,40,21,14,65,0,0               // movaps        0x410e(%rip),%xmm10        # 4cf0 <_sk_callback_sse2+0x26f>
   .byte  69,15,89,234                        // mulps         %xmm10,%xmm13
   .byte  69,15,88,235                        // addps         %xmm11,%xmm13
   .byte  15,88,228                           // addps         %xmm4,%xmm4
@@ -27981,7 +27974,7 @@ _sk_hue_sse2:
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  68,15,89,211                        // mulps         %xmm3,%xmm10
-  .byte  68,15,40,5,81,63,0,0                // movaps        0x3f51(%rip),%xmm8        # 4d30 <_sk_callback_sse2+0x2ac>
+  .byte  68,15,40,5,81,63,0,0                // movaps        0x3f51(%rip),%xmm8        # 4d30 <_sk_callback_sse2+0x2af>
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
   .byte  15,40,207                           // movaps        %xmm7,%xmm1
   .byte  68,15,92,217                        // subps         %xmm1,%xmm11
@@ -28027,12 +28020,12 @@ _sk_hue_sse2:
   .byte  69,15,84,206                        // andps         %xmm14,%xmm9
   .byte  69,15,84,214                        // andps         %xmm14,%xmm10
   .byte  65,15,84,214                        // andps         %xmm14,%xmm2
-  .byte  68,15,40,61,101,62,0,0              // movaps        0x3e65(%rip),%xmm15        # 4d00 <_sk_callback_sse2+0x27c>
+  .byte  68,15,40,61,101,62,0,0              // movaps        0x3e65(%rip),%xmm15        # 4d00 <_sk_callback_sse2+0x27f>
   .byte  65,15,89,231                        // mulps         %xmm15,%xmm4
-  .byte  15,40,5,106,62,0,0                  // movaps        0x3e6a(%rip),%xmm0        # 4d10 <_sk_callback_sse2+0x28c>
+  .byte  15,40,5,106,62,0,0                  // movaps        0x3e6a(%rip),%xmm0        # 4d10 <_sk_callback_sse2+0x28f>
   .byte  15,89,240                           // mulps         %xmm0,%xmm6
   .byte  15,88,244                           // addps         %xmm4,%xmm6
-  .byte  68,15,40,53,108,62,0,0              // movaps        0x3e6c(%rip),%xmm14        # 4d20 <_sk_callback_sse2+0x29c>
+  .byte  68,15,40,53,108,62,0,0              // movaps        0x3e6c(%rip),%xmm14        # 4d20 <_sk_callback_sse2+0x29f>
   .byte  68,15,40,239                        // movaps        %xmm7,%xmm13
   .byte  69,15,89,238                        // mulps         %xmm14,%xmm13
   .byte  68,15,88,238                        // addps         %xmm6,%xmm13
@@ -28209,14 +28202,14 @@ _sk_saturation_sse2:
   .byte  68,15,84,211                        // andps         %xmm3,%xmm10
   .byte  68,15,84,203                        // andps         %xmm3,%xmm9
   .byte  15,84,195                           // andps         %xmm3,%xmm0
-  .byte  68,15,40,5,1,60,0,0                 // movaps        0x3c01(%rip),%xmm8        # 4d40 <_sk_callback_sse2+0x2bc>
+  .byte  68,15,40,5,1,60,0,0                 // movaps        0x3c01(%rip),%xmm8        # 4d40 <_sk_callback_sse2+0x2bf>
   .byte  15,40,214                           // movaps        %xmm6,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
-  .byte  15,40,13,3,60,0,0                   // movaps        0x3c03(%rip),%xmm1        # 4d50 <_sk_callback_sse2+0x2cc>
+  .byte  15,40,13,3,60,0,0                   // movaps        0x3c03(%rip),%xmm1        # 4d50 <_sk_callback_sse2+0x2cf>
   .byte  15,40,221                           // movaps        %xmm5,%xmm3
   .byte  15,89,217                           // mulps         %xmm1,%xmm3
   .byte  15,88,218                           // addps         %xmm2,%xmm3
-  .byte  68,15,40,37,2,60,0,0                // movaps        0x3c02(%rip),%xmm12        # 4d60 <_sk_callback_sse2+0x2dc>
+  .byte  68,15,40,37,2,60,0,0                // movaps        0x3c02(%rip),%xmm12        # 4d60 <_sk_callback_sse2+0x2df>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
   .byte  68,15,88,235                        // addps         %xmm3,%xmm13
   .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
@@ -28261,7 +28254,7 @@ _sk_saturation_sse2:
   .byte  15,40,223                           // movaps        %xmm7,%xmm3
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,89,221                           // mulps         %xmm5,%xmm3
-  .byte  68,15,40,5,103,59,0,0               // movaps        0x3b67(%rip),%xmm8        # 4d70 <_sk_callback_sse2+0x2ec>
+  .byte  68,15,40,5,103,59,0,0               // movaps        0x3b67(%rip),%xmm8        # 4d70 <_sk_callback_sse2+0x2ef>
   .byte  65,15,40,224                        // movaps        %xmm8,%xmm4
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  15,88,253                           // addps         %xmm5,%xmm7
@@ -28362,14 +28355,14 @@ _sk_color_sse2:
   .byte  68,15,40,213                        // movaps        %xmm5,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
   .byte  65,15,40,208                        // movaps        %xmm8,%xmm2
-  .byte  68,15,40,45,5,58,0,0                // movaps        0x3a05(%rip),%xmm13        # 4d80 <_sk_callback_sse2+0x2fc>
+  .byte  68,15,40,45,5,58,0,0                // movaps        0x3a05(%rip),%xmm13        # 4d80 <_sk_callback_sse2+0x2ff>
   .byte  68,15,40,198                        // movaps        %xmm6,%xmm8
   .byte  69,15,89,197                        // mulps         %xmm13,%xmm8
-  .byte  68,15,40,53,5,58,0,0                // movaps        0x3a05(%rip),%xmm14        # 4d90 <_sk_callback_sse2+0x30c>
+  .byte  68,15,40,53,5,58,0,0                // movaps        0x3a05(%rip),%xmm14        # 4d90 <_sk_callback_sse2+0x30f>
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  65,15,89,198                        // mulps         %xmm14,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,29,1,58,0,0                // movaps        0x3a01(%rip),%xmm11        # 4da0 <_sk_callback_sse2+0x31c>
+  .byte  68,15,40,29,1,58,0,0                // movaps        0x3a01(%rip),%xmm11        # 4da0 <_sk_callback_sse2+0x31f>
   .byte  69,15,89,227                        // mulps         %xmm11,%xmm12
   .byte  68,15,88,224                        // addps         %xmm0,%xmm12
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
@@ -28377,7 +28370,7 @@ _sk_color_sse2:
   .byte  69,15,40,250                        // movaps        %xmm10,%xmm15
   .byte  69,15,89,254                        // mulps         %xmm14,%xmm15
   .byte  68,15,88,248                        // addps         %xmm0,%xmm15
-  .byte  68,15,40,5,237,57,0,0               // movaps        0x39ed(%rip),%xmm8        # 4db0 <_sk_callback_sse2+0x32c>
+  .byte  68,15,40,5,237,57,0,0               // movaps        0x39ed(%rip),%xmm8        # 4db0 <_sk_callback_sse2+0x32f>
   .byte  65,15,40,224                        // movaps        %xmm8,%xmm4
   .byte  15,92,226                           // subps         %xmm2,%xmm4
   .byte  15,89,252                           // mulps         %xmm4,%xmm7
@@ -28513,15 +28506,15 @@ _sk_luminosity_sse2:
   .byte  68,15,40,205                        // movaps        %xmm5,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
   .byte  15,89,222                           // mulps         %xmm6,%xmm3
-  .byte  68,15,40,37,4,56,0,0                // movaps        0x3804(%rip),%xmm12        # 4dc0 <_sk_callback_sse2+0x33c>
+  .byte  68,15,40,37,4,56,0,0                // movaps        0x3804(%rip),%xmm12        # 4dc0 <_sk_callback_sse2+0x33f>
   .byte  68,15,40,199                        // movaps        %xmm7,%xmm8
   .byte  69,15,89,196                        // mulps         %xmm12,%xmm8
-  .byte  68,15,40,45,4,56,0,0                // movaps        0x3804(%rip),%xmm13        # 4dd0 <_sk_callback_sse2+0x34c>
+  .byte  68,15,40,45,4,56,0,0                // movaps        0x3804(%rip),%xmm13        # 4dd0 <_sk_callback_sse2+0x34f>
   .byte  68,15,40,241                        // movaps        %xmm1,%xmm14
   .byte  69,15,89,245                        // mulps         %xmm13,%xmm14
   .byte  69,15,88,240                        // addps         %xmm8,%xmm14
-  .byte  68,15,40,29,0,56,0,0                // movaps        0x3800(%rip),%xmm11        # 4de0 <_sk_callback_sse2+0x35c>
-  .byte  68,15,40,5,8,56,0,0                 // movaps        0x3808(%rip),%xmm8        # 4df0 <_sk_callback_sse2+0x36c>
+  .byte  68,15,40,29,0,56,0,0                // movaps        0x3800(%rip),%xmm11        # 4de0 <_sk_callback_sse2+0x35f>
+  .byte  68,15,40,5,8,56,0,0                 // movaps        0x3808(%rip),%xmm8        # 4df0 <_sk_callback_sse2+0x36f>
   .byte  69,15,40,248                        // movaps        %xmm8,%xmm15
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  68,15,92,248                        // subps         %xmm0,%xmm15
@@ -28666,7 +28659,7 @@ HIDDEN _sk_clamp_1_sse2
 .globl _sk_clamp_1_sse2
 FUNCTION(_sk_clamp_1_sse2)
 _sk_clamp_1_sse2:
-  .byte  68,15,40,5,17,54,0,0                // movaps        0x3611(%rip),%xmm8        # 4e00 <_sk_callback_sse2+0x37c>
+  .byte  68,15,40,5,17,54,0,0                // movaps        0x3611(%rip),%xmm8        # 4e00 <_sk_callback_sse2+0x37f>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
   .byte  65,15,93,208                        // minps         %xmm8,%xmm2
@@ -28678,7 +28671,7 @@ HIDDEN _sk_clamp_a_sse2
 .globl _sk_clamp_a_sse2
 FUNCTION(_sk_clamp_a_sse2)
 _sk_clamp_a_sse2:
-  .byte  15,93,29,6,54,0,0                   // minps         0x3606(%rip),%xmm3        # 4e10 <_sk_callback_sse2+0x38c>
+  .byte  15,93,29,6,54,0,0                   // minps         0x3606(%rip),%xmm3        # 4e10 <_sk_callback_sse2+0x38f>
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -28765,7 +28758,7 @@ HIDDEN _sk_unpremul_sse2
 FUNCTION(_sk_unpremul_sse2)
 _sk_unpremul_sse2:
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
-  .byte  68,15,40,13,113,53,0,0              // movaps        0x3571(%rip),%xmm9        # 4e20 <_sk_callback_sse2+0x39c>
+  .byte  68,15,40,13,113,53,0,0              // movaps        0x3571(%rip),%xmm9        # 4e20 <_sk_callback_sse2+0x39f>
   .byte  68,15,94,203                        // divps         %xmm3,%xmm9
   .byte  68,15,194,195,4                     // cmpneqps      %xmm3,%xmm8
   .byte  69,15,84,193                        // andps         %xmm9,%xmm8
@@ -28779,20 +28772,20 @@ HIDDEN _sk_from_srgb_sse2
 .globl _sk_from_srgb_sse2
 FUNCTION(_sk_from_srgb_sse2)
 _sk_from_srgb_sse2:
-  .byte  68,15,40,5,92,53,0,0                // movaps        0x355c(%rip),%xmm8        # 4e30 <_sk_callback_sse2+0x3ac>
+  .byte  68,15,40,5,92,53,0,0                // movaps        0x355c(%rip),%xmm8        # 4e30 <_sk_callback_sse2+0x3af>
   .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
   .byte  69,15,89,232                        // mulps         %xmm8,%xmm13
   .byte  68,15,40,216                        // movaps        %xmm0,%xmm11
   .byte  69,15,89,219                        // mulps         %xmm11,%xmm11
-  .byte  68,15,40,13,84,53,0,0               // movaps        0x3554(%rip),%xmm9        # 4e40 <_sk_callback_sse2+0x3bc>
+  .byte  68,15,40,13,84,53,0,0               // movaps        0x3554(%rip),%xmm9        # 4e40 <_sk_callback_sse2+0x3bf>
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
   .byte  69,15,89,241                        // mulps         %xmm9,%xmm14
-  .byte  68,15,40,21,84,53,0,0               // movaps        0x3554(%rip),%xmm10        # 4e50 <_sk_callback_sse2+0x3cc>
+  .byte  68,15,40,21,84,53,0,0               // movaps        0x3554(%rip),%xmm10        # 4e50 <_sk_callback_sse2+0x3cf>
   .byte  69,15,88,242                        // addps         %xmm10,%xmm14
   .byte  69,15,89,243                        // mulps         %xmm11,%xmm14
-  .byte  68,15,40,29,84,53,0,0               // movaps        0x3554(%rip),%xmm11        # 4e60 <_sk_callback_sse2+0x3dc>
+  .byte  68,15,40,29,84,53,0,0               // movaps        0x3554(%rip),%xmm11        # 4e60 <_sk_callback_sse2+0x3df>
   .byte  69,15,88,243                        // addps         %xmm11,%xmm14
-  .byte  68,15,40,37,88,53,0,0               // movaps        0x3558(%rip),%xmm12        # 4e70 <_sk_callback_sse2+0x3ec>
+  .byte  68,15,40,37,88,53,0,0               // movaps        0x3558(%rip),%xmm12        # 4e70 <_sk_callback_sse2+0x3ef>
   .byte  65,15,194,196,1                     // cmpltps       %xmm12,%xmm0
   .byte  68,15,84,232                        // andps         %xmm0,%xmm13
   .byte  65,15,85,198                        // andnps        %xmm14,%xmm0
@@ -28831,20 +28824,20 @@ _sk_to_srgb_sse2:
   .byte  68,15,82,192                        // rsqrtps       %xmm0,%xmm8
   .byte  69,15,83,200                        // rcpps         %xmm8,%xmm9
   .byte  69,15,82,232                        // rsqrtps       %xmm8,%xmm13
-  .byte  68,15,40,5,221,52,0,0               // movaps        0x34dd(%rip),%xmm8        # 4e80 <_sk_callback_sse2+0x3fc>
+  .byte  68,15,40,5,221,52,0,0               // movaps        0x34dd(%rip),%xmm8        # 4e80 <_sk_callback_sse2+0x3ff>
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
   .byte  69,15,89,240                        // mulps         %xmm8,%xmm14
-  .byte  68,15,40,21,221,52,0,0              // movaps        0x34dd(%rip),%xmm10        # 4e90 <_sk_callback_sse2+0x40c>
+  .byte  68,15,40,21,221,52,0,0              // movaps        0x34dd(%rip),%xmm10        # 4e90 <_sk_callback_sse2+0x40f>
   .byte  69,15,89,202                        // mulps         %xmm10,%xmm9
-  .byte  68,15,40,29,225,52,0,0              // movaps        0x34e1(%rip),%xmm11        # 4ea0 <_sk_callback_sse2+0x41c>
+  .byte  68,15,40,29,225,52,0,0              // movaps        0x34e1(%rip),%xmm11        # 4ea0 <_sk_callback_sse2+0x41f>
   .byte  69,15,88,203                        // addps         %xmm11,%xmm9
-  .byte  68,15,40,37,229,52,0,0              // movaps        0x34e5(%rip),%xmm12        # 4eb0 <_sk_callback_sse2+0x42c>
+  .byte  68,15,40,37,229,52,0,0              // movaps        0x34e5(%rip),%xmm12        # 4eb0 <_sk_callback_sse2+0x42f>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,40,13,229,52,0,0              // movaps        0x34e5(%rip),%xmm9        # 4ec0 <_sk_callback_sse2+0x43c>
+  .byte  68,15,40,13,229,52,0,0              // movaps        0x34e5(%rip),%xmm9        # 4ec0 <_sk_callback_sse2+0x43f>
   .byte  69,15,40,249                        // movaps        %xmm9,%xmm15
   .byte  69,15,93,253                        // minps         %xmm13,%xmm15
-  .byte  68,15,40,45,229,52,0,0              // movaps        0x34e5(%rip),%xmm13        # 4ed0 <_sk_callback_sse2+0x44c>
+  .byte  68,15,40,45,229,52,0,0              // movaps        0x34e5(%rip),%xmm13        # 4ed0 <_sk_callback_sse2+0x44f>
   .byte  65,15,194,197,1                     // cmpltps       %xmm13,%xmm0
   .byte  68,15,84,240                        // andps         %xmm0,%xmm14
   .byte  65,15,85,199                        // andnps        %xmm15,%xmm0
@@ -28894,7 +28887,7 @@ _sk_rgb_to_hsl_sse2:
   .byte  68,15,93,218                        // minps         %xmm2,%xmm11
   .byte  65,15,40,202                        // movaps        %xmm10,%xmm1
   .byte  65,15,92,203                        // subps         %xmm11,%xmm1
-  .byte  68,15,40,45,62,52,0,0               // movaps        0x343e(%rip),%xmm13        # 4ee0 <_sk_callback_sse2+0x45c>
+  .byte  68,15,40,45,62,52,0,0               // movaps        0x343e(%rip),%xmm13        # 4ee0 <_sk_callback_sse2+0x45f>
   .byte  68,15,94,233                        // divps         %xmm1,%xmm13
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  65,15,194,192,0                     // cmpeqps       %xmm8,%xmm0
@@ -28903,30 +28896,30 @@ _sk_rgb_to_hsl_sse2:
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,40,241                        // movaps        %xmm9,%xmm14
   .byte  68,15,194,242,1                     // cmpltps       %xmm2,%xmm14
-  .byte  68,15,84,53,36,52,0,0               // andps         0x3424(%rip),%xmm14        # 4ef0 <_sk_callback_sse2+0x46c>
+  .byte  68,15,84,53,36,52,0,0               // andps         0x3424(%rip),%xmm14        # 4ef0 <_sk_callback_sse2+0x46f>
   .byte  69,15,88,244                        // addps         %xmm12,%xmm14
   .byte  69,15,40,250                        // movaps        %xmm10,%xmm15
   .byte  69,15,194,249,0                     // cmpeqps       %xmm9,%xmm15
   .byte  65,15,92,208                        // subps         %xmm8,%xmm2
   .byte  65,15,89,213                        // mulps         %xmm13,%xmm2
-  .byte  68,15,40,37,23,52,0,0               // movaps        0x3417(%rip),%xmm12        # 4f00 <_sk_callback_sse2+0x47c>
+  .byte  68,15,40,37,23,52,0,0               // movaps        0x3417(%rip),%xmm12        # 4f00 <_sk_callback_sse2+0x47f>
   .byte  65,15,88,212                        // addps         %xmm12,%xmm2
   .byte  69,15,92,193                        // subps         %xmm9,%xmm8
   .byte  69,15,89,197                        // mulps         %xmm13,%xmm8
-  .byte  68,15,88,5,19,52,0,0                // addps         0x3413(%rip),%xmm8        # 4f10 <_sk_callback_sse2+0x48c>
+  .byte  68,15,88,5,19,52,0,0                // addps         0x3413(%rip),%xmm8        # 4f10 <_sk_callback_sse2+0x48f>
   .byte  65,15,84,215                        // andps         %xmm15,%xmm2
   .byte  69,15,85,248                        // andnps        %xmm8,%xmm15
   .byte  68,15,86,250                        // orps          %xmm2,%xmm15
   .byte  68,15,84,240                        // andps         %xmm0,%xmm14
   .byte  65,15,85,199                        // andnps        %xmm15,%xmm0
   .byte  65,15,86,198                        // orps          %xmm14,%xmm0
-  .byte  15,89,5,4,52,0,0                    // mulps         0x3404(%rip),%xmm0        # 4f20 <_sk_callback_sse2+0x49c>
+  .byte  15,89,5,4,52,0,0                    // mulps         0x3404(%rip),%xmm0        # 4f20 <_sk_callback_sse2+0x49f>
   .byte  69,15,40,194                        // movaps        %xmm10,%xmm8
   .byte  69,15,194,195,4                     // cmpneqps      %xmm11,%xmm8
   .byte  65,15,84,192                        // andps         %xmm8,%xmm0
   .byte  69,15,92,226                        // subps         %xmm10,%xmm12
   .byte  69,15,88,211                        // addps         %xmm11,%xmm10
-  .byte  68,15,40,13,247,51,0,0              // movaps        0x33f7(%rip),%xmm9        # 4f30 <_sk_callback_sse2+0x4ac>
+  .byte  68,15,40,13,247,51,0,0              // movaps        0x33f7(%rip),%xmm9        # 4f30 <_sk_callback_sse2+0x4af>
   .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  68,15,194,202,1                     // cmpltps       %xmm2,%xmm9
@@ -28950,7 +28943,7 @@ _sk_hsl_to_rgb_sse2:
   .byte  15,41,92,36,168                     // movaps        %xmm3,-0x58(%rsp)
   .byte  68,15,40,218                        // movaps        %xmm2,%xmm11
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
-  .byte  68,15,40,13,182,51,0,0              // movaps        0x33b6(%rip),%xmm9        # 4f40 <_sk_callback_sse2+0x4bc>
+  .byte  68,15,40,13,182,51,0,0              // movaps        0x33b6(%rip),%xmm9        # 4f40 <_sk_callback_sse2+0x4bf>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  69,15,194,211,2                     // cmpleps       %xmm11,%xmm10
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
@@ -28967,28 +28960,28 @@ _sk_hsl_to_rgb_sse2:
   .byte  69,15,88,211                        // addps         %xmm11,%xmm10
   .byte  69,15,88,219                        // addps         %xmm11,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  15,40,5,127,51,0,0                  // movaps        0x337f(%rip),%xmm0        # 4f50 <_sk_callback_sse2+0x4cc>
+  .byte  15,40,5,127,51,0,0                  // movaps        0x337f(%rip),%xmm0        # 4f50 <_sk_callback_sse2+0x4cf>
   .byte  15,88,198                           // addps         %xmm6,%xmm0
   .byte  243,15,91,200                       // cvttps2dq     %xmm0,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
   .byte  15,194,217,1                        // cmpltps       %xmm1,%xmm3
-  .byte  15,84,29,119,51,0,0                 // andps         0x3377(%rip),%xmm3        # 4f60 <_sk_callback_sse2+0x4dc>
+  .byte  15,84,29,119,51,0,0                 // andps         0x3377(%rip),%xmm3        # 4f60 <_sk_callback_sse2+0x4df>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,92,193                           // subps         %xmm1,%xmm0
-  .byte  68,15,40,45,121,51,0,0              // movaps        0x3379(%rip),%xmm13        # 4f70 <_sk_callback_sse2+0x4ec>
+  .byte  68,15,40,45,121,51,0,0              // movaps        0x3379(%rip),%xmm13        # 4f70 <_sk_callback_sse2+0x4ef>
   .byte  69,15,40,197                        // movaps        %xmm13,%xmm8
   .byte  68,15,194,192,2                     // cmpleps       %xmm0,%xmm8
   .byte  69,15,40,242                        // movaps        %xmm10,%xmm14
   .byte  69,15,92,243                        // subps         %xmm11,%xmm14
   .byte  65,15,40,217                        // movaps        %xmm9,%xmm3
   .byte  15,194,216,2                        // cmpleps       %xmm0,%xmm3
-  .byte  15,40,21,137,51,0,0                 // movaps        0x3389(%rip),%xmm2        # 4fa0 <_sk_callback_sse2+0x51c>
+  .byte  15,40,21,137,51,0,0                 // movaps        0x3389(%rip),%xmm2        # 4fa0 <_sk_callback_sse2+0x51f>
   .byte  68,15,40,250                        // movaps        %xmm2,%xmm15
   .byte  68,15,194,248,2                     // cmpleps       %xmm0,%xmm15
-  .byte  15,40,13,89,51,0,0                  // movaps        0x3359(%rip),%xmm1        # 4f80 <_sk_callback_sse2+0x4fc>
+  .byte  15,40,13,89,51,0,0                  // movaps        0x3359(%rip),%xmm1        # 4f80 <_sk_callback_sse2+0x4ff>
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
-  .byte  15,40,45,95,51,0,0                  // movaps        0x335f(%rip),%xmm5        # 4f90 <_sk_callback_sse2+0x50c>
+  .byte  15,40,45,95,51,0,0                  // movaps        0x335f(%rip),%xmm5        # 4f90 <_sk_callback_sse2+0x50f>
   .byte  15,40,229                           // movaps        %xmm5,%xmm4
   .byte  15,92,224                           // subps         %xmm0,%xmm4
   .byte  65,15,89,230                        // mulps         %xmm14,%xmm4
@@ -29011,7 +29004,7 @@ _sk_hsl_to_rgb_sse2:
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  15,40,222                           // movaps        %xmm6,%xmm3
   .byte  15,194,216,1                        // cmpltps       %xmm0,%xmm3
-  .byte  15,84,29,212,50,0,0                 // andps         0x32d4(%rip),%xmm3        # 4f60 <_sk_callback_sse2+0x4dc>
+  .byte  15,84,29,212,50,0,0                 // andps         0x32d4(%rip),%xmm3        # 4f60 <_sk_callback_sse2+0x4df>
   .byte  15,92,195                           // subps         %xmm3,%xmm0
   .byte  68,15,40,230                        // movaps        %xmm6,%xmm12
   .byte  68,15,92,224                        // subps         %xmm0,%xmm12
@@ -29041,12 +29034,12 @@ _sk_hsl_to_rgb_sse2:
   .byte  15,40,124,36,136                    // movaps        -0x78(%rsp),%xmm7
   .byte  15,40,231                           // movaps        %xmm7,%xmm4
   .byte  15,85,227                           // andnps        %xmm3,%xmm4
-  .byte  15,88,53,172,50,0,0                 // addps         0x32ac(%rip),%xmm6        # 4fb0 <_sk_callback_sse2+0x52c>
+  .byte  15,88,53,172,50,0,0                 // addps         0x32ac(%rip),%xmm6        # 4fb0 <_sk_callback_sse2+0x52f>
   .byte  243,15,91,198                       // cvttps2dq     %xmm6,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  15,40,222                           // movaps        %xmm6,%xmm3
   .byte  15,194,216,1                        // cmpltps       %xmm0,%xmm3
-  .byte  15,84,29,71,50,0,0                  // andps         0x3247(%rip),%xmm3        # 4f60 <_sk_callback_sse2+0x4dc>
+  .byte  15,84,29,71,50,0,0                  // andps         0x3247(%rip),%xmm3        # 4f60 <_sk_callback_sse2+0x4df>
   .byte  15,92,195                           // subps         %xmm3,%xmm0
   .byte  15,92,240                           // subps         %xmm0,%xmm6
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
@@ -29110,7 +29103,7 @@ _sk_scale_u8_sse2:
   .byte  102,69,15,96,193                    // punpcklbw     %xmm9,%xmm8
   .byte  102,69,15,97,193                    // punpcklwd     %xmm9,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,213,49,0,0               // mulps         0x31d5(%rip),%xmm8        # 4fc0 <_sk_callback_sse2+0x53c>
+  .byte  68,15,89,5,213,49,0,0               // mulps         0x31d5(%rip),%xmm8        # 4fc0 <_sk_callback_sse2+0x53f>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
@@ -29151,7 +29144,7 @@ _sk_lerp_u8_sse2:
   .byte  102,69,15,96,193                    // punpcklbw     %xmm9,%xmm8
   .byte  102,69,15,97,193                    // punpcklwd     %xmm9,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,115,49,0,0               // mulps         0x3173(%rip),%xmm8        # 4fd0 <_sk_callback_sse2+0x54c>
+  .byte  68,15,89,5,115,49,0,0               // mulps         0x3173(%rip),%xmm8        # 4fd0 <_sk_callback_sse2+0x54f>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -29176,17 +29169,17 @@ _sk_lerp_565_sse2:
   .byte  243,68,15,126,20,120                // movq          (%rax,%rdi,2),%xmm10
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,69,15,97,208                    // punpcklwd     %xmm8,%xmm10
-  .byte  102,68,15,111,5,57,49,0,0           // movdqa        0x3139(%rip),%xmm8        # 4fe0 <_sk_callback_sse2+0x55c>
+  .byte  102,68,15,111,5,57,49,0,0           // movdqa        0x3139(%rip),%xmm8        # 4fe0 <_sk_callback_sse2+0x55f>
   .byte  102,69,15,219,194                   // pand          %xmm10,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,56,49,0,0                // mulps         0x3138(%rip),%xmm8        # 4ff0 <_sk_callback_sse2+0x56c>
-  .byte  102,68,15,111,13,63,49,0,0          // movdqa        0x313f(%rip),%xmm9        # 5000 <_sk_callback_sse2+0x57c>
+  .byte  68,15,89,5,56,49,0,0                // mulps         0x3138(%rip),%xmm8        # 4ff0 <_sk_callback_sse2+0x56f>
+  .byte  102,68,15,111,13,63,49,0,0          // movdqa        0x313f(%rip),%xmm9        # 5000 <_sk_callback_sse2+0x57f>
   .byte  102,69,15,219,202                   // pand          %xmm10,%xmm9
   .byte  69,15,91,201                        // cvtdq2ps      %xmm9,%xmm9
-  .byte  68,15,89,13,62,49,0,0               // mulps         0x313e(%rip),%xmm9        # 5010 <_sk_callback_sse2+0x58c>
-  .byte  102,68,15,219,21,69,49,0,0          // pand          0x3145(%rip),%xmm10        # 5020 <_sk_callback_sse2+0x59c>
+  .byte  68,15,89,13,62,49,0,0               // mulps         0x313e(%rip),%xmm9        # 5010 <_sk_callback_sse2+0x58f>
+  .byte  102,68,15,219,21,69,49,0,0          // pand          0x3145(%rip),%xmm10        # 5020 <_sk_callback_sse2+0x59f>
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
-  .byte  68,15,89,21,73,49,0,0               // mulps         0x3149(%rip),%xmm10        # 5030 <_sk_callback_sse2+0x5ac>
+  .byte  68,15,89,21,73,49,0,0               // mulps         0x3149(%rip),%xmm10        # 5030 <_sk_callback_sse2+0x5af>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -29217,7 +29210,7 @@ _sk_load_tables_sse2:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  76,139,72,8                         // mov           0x8(%rax),%r9
   .byte  243,69,15,111,12,184                // movdqu        (%r8,%rdi,4),%xmm9
-  .byte  102,68,15,111,5,249,48,0,0          // movdqa        0x30f9(%rip),%xmm8        # 5040 <_sk_callback_sse2+0x5bc>
+  .byte  102,68,15,111,5,249,48,0,0          // movdqa        0x30f9(%rip),%xmm8        # 5040 <_sk_callback_sse2+0x5bf>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
@@ -29272,7 +29265,7 @@ _sk_load_tables_sse2:
   .byte  65,15,20,208                        // unpcklps      %xmm8,%xmm2
   .byte  102,65,15,114,209,24                // psrld         $0x18,%xmm9
   .byte  65,15,91,217                        // cvtdq2ps      %xmm9,%xmm3
-  .byte  15,89,29,6,48,0,0                   // mulps         0x3006(%rip),%xmm3        # 5050 <_sk_callback_sse2+0x5cc>
+  .byte  15,89,29,6,48,0,0                   // mulps         0x3006(%rip),%xmm3        # 5050 <_sk_callback_sse2+0x5cf>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -29291,7 +29284,7 @@ _sk_load_tables_u16_be_sse2:
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,97,200                       // punpcklwd     %xmm0,%xmm1
   .byte  102,68,15,105,200                   // punpckhwd     %xmm0,%xmm9
-  .byte  102,68,15,111,21,217,47,0,0         // movdqa        0x2fd9(%rip),%xmm10        # 5060 <_sk_callback_sse2+0x5dc>
+  .byte  102,68,15,111,21,217,47,0,0         // movdqa        0x2fd9(%rip),%xmm10        # 5060 <_sk_callback_sse2+0x5df>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,194                   // pand          %xmm10,%xmm0
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
@@ -29352,7 +29345,7 @@ _sk_load_tables_u16_be_sse2:
   .byte  102,65,15,235,217                   // por           %xmm9,%xmm3
   .byte  102,65,15,97,216                    // punpcklwd     %xmm8,%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,200,46,0,0                 // mulps         0x2ec8(%rip),%xmm3        # 5070 <_sk_callback_sse2+0x5ec>
+  .byte  15,89,29,200,46,0,0                 // mulps         0x2ec8(%rip),%xmm3        # 5070 <_sk_callback_sse2+0x5ef>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -29374,7 +29367,7 @@ _sk_load_tables_rgb_u16_be_sse2:
   .byte  102,68,15,97,208                    // punpcklwd     %xmm0,%xmm10
   .byte  102,65,15,111,195                   // movdqa        %xmm11,%xmm0
   .byte  102,65,15,97,194                    // punpcklwd     %xmm10,%xmm0
-  .byte  102,68,15,111,5,136,46,0,0          // movdqa        0x2e88(%rip),%xmm8        # 5080 <_sk_callback_sse2+0x5fc>
+  .byte  102,68,15,111,5,136,46,0,0          // movdqa        0x2e88(%rip),%xmm8        # 5080 <_sk_callback_sse2+0x5ff>
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
@@ -29429,7 +29422,7 @@ _sk_load_tables_rgb_u16_be_sse2:
   .byte  15,20,211                           // unpcklps      %xmm3,%xmm2
   .byte  65,15,20,208                        // unpcklps      %xmm8,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,151,45,0,0                 // movaps        0x2d97(%rip),%xmm3        # 5090 <_sk_callback_sse2+0x60c>
+  .byte  15,40,29,151,45,0,0                 // movaps        0x2d97(%rip),%xmm3        # 5090 <_sk_callback_sse2+0x60f>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_byte_tables_sse2
@@ -29439,7 +29432,7 @@ _sk_byte_tables_sse2:
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,152,45,0,0               // movaps        0x2d98(%rip),%xmm8        # 50a0 <_sk_callback_sse2+0x61c>
+  .byte  68,15,40,5,152,45,0,0               // movaps        0x2d98(%rip),%xmm8        # 50a0 <_sk_callback_sse2+0x61f>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  102,72,15,126,193                   // movq          %xmm0,%rcx
@@ -29466,7 +29459,7 @@ _sk_byte_tables_sse2:
   .byte  102,65,15,96,193                    // punpcklbw     %xmm9,%xmm0
   .byte  102,65,15,97,193                    // punpcklwd     %xmm9,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,21,53,45,0,0               // movaps        0x2d35(%rip),%xmm10        # 50b0 <_sk_callback_sse2+0x62c>
+  .byte  68,15,40,21,53,45,0,0               // movaps        0x2d35(%rip),%xmm10        # 50b0 <_sk_callback_sse2+0x62f>
   .byte  65,15,89,194                        // mulps         %xmm10,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -29582,7 +29575,7 @@ _sk_byte_tables_rgb_sse2:
   .byte  102,65,15,96,193                    // punpcklbw     %xmm9,%xmm0
   .byte  102,65,15,97,193                    // punpcklwd     %xmm9,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,21,136,43,0,0              // movaps        0x2b88(%rip),%xmm10        # 50c0 <_sk_callback_sse2+0x63c>
+  .byte  68,15,40,21,136,43,0,0              // movaps        0x2b88(%rip),%xmm10        # 50c0 <_sk_callback_sse2+0x63f>
   .byte  65,15,89,194                        // mulps         %xmm10,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -29779,15 +29772,15 @@ _sk_parametric_r_sse2:
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,199,40,0,0              // mulps         0x28c7(%rip),%xmm9        # 50d0 <_sk_callback_sse2+0x64c>
-  .byte  68,15,84,21,207,40,0,0              // andps         0x28cf(%rip),%xmm10        # 50e0 <_sk_callback_sse2+0x65c>
-  .byte  68,15,86,21,215,40,0,0              // orps          0x28d7(%rip),%xmm10        # 50f0 <_sk_callback_sse2+0x66c>
-  .byte  68,15,88,13,223,40,0,0              // addps         0x28df(%rip),%xmm9        # 5100 <_sk_callback_sse2+0x67c>
-  .byte  68,15,40,37,231,40,0,0              // movaps        0x28e7(%rip),%xmm12        # 5110 <_sk_callback_sse2+0x68c>
+  .byte  68,15,89,13,199,40,0,0              // mulps         0x28c7(%rip),%xmm9        # 50d0 <_sk_callback_sse2+0x64f>
+  .byte  68,15,84,21,207,40,0,0              // andps         0x28cf(%rip),%xmm10        # 50e0 <_sk_callback_sse2+0x65f>
+  .byte  68,15,86,21,215,40,0,0              // orps          0x28d7(%rip),%xmm10        # 50f0 <_sk_callback_sse2+0x66f>
+  .byte  68,15,88,13,223,40,0,0              // addps         0x28df(%rip),%xmm9        # 5100 <_sk_callback_sse2+0x67f>
+  .byte  68,15,40,37,231,40,0,0              // movaps        0x28e7(%rip),%xmm12        # 5110 <_sk_callback_sse2+0x68f>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,231,40,0,0              // addps         0x28e7(%rip),%xmm10        # 5120 <_sk_callback_sse2+0x69c>
-  .byte  68,15,40,37,239,40,0,0              // movaps        0x28ef(%rip),%xmm12        # 5130 <_sk_callback_sse2+0x6ac>
+  .byte  68,15,88,21,231,40,0,0              // addps         0x28e7(%rip),%xmm10        # 5120 <_sk_callback_sse2+0x69f>
+  .byte  68,15,40,37,239,40,0,0              // movaps        0x28ef(%rip),%xmm12        # 5130 <_sk_callback_sse2+0x6af>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -29795,22 +29788,22 @@ _sk_parametric_r_sse2:
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,217,40,0,0              // movaps        0x28d9(%rip),%xmm10        # 5140 <_sk_callback_sse2+0x6bc>
+  .byte  68,15,40,21,217,40,0,0              // movaps        0x28d9(%rip),%xmm10        # 5140 <_sk_callback_sse2+0x6bf>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,205,40,0,0              // addps         0x28cd(%rip),%xmm9        # 5150 <_sk_callback_sse2+0x6cc>
-  .byte  68,15,40,37,213,40,0,0              // movaps        0x28d5(%rip),%xmm12        # 5160 <_sk_callback_sse2+0x6dc>
+  .byte  68,15,88,13,205,40,0,0              // addps         0x28cd(%rip),%xmm9        # 5150 <_sk_callback_sse2+0x6cf>
+  .byte  68,15,40,37,213,40,0,0              // movaps        0x28d5(%rip),%xmm12        # 5160 <_sk_callback_sse2+0x6df>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,213,40,0,0              // movaps        0x28d5(%rip),%xmm12        # 5170 <_sk_callback_sse2+0x6ec>
+  .byte  68,15,40,37,213,40,0,0              // movaps        0x28d5(%rip),%xmm12        # 5170 <_sk_callback_sse2+0x6ef>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,217,40,0,0              // movaps        0x28d9(%rip),%xmm13        # 5180 <_sk_callback_sse2+0x6fc>
+  .byte  68,15,40,45,217,40,0,0              // movaps        0x28d9(%rip),%xmm13        # 5180 <_sk_callback_sse2+0x6ff>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,217,40,0,0              // mulps         0x28d9(%rip),%xmm13        # 5190 <_sk_callback_sse2+0x70c>
+  .byte  68,15,89,45,217,40,0,0              // mulps         0x28d9(%rip),%xmm13        # 5190 <_sk_callback_sse2+0x70f>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -29846,15 +29839,15 @@ _sk_parametric_g_sse2:
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,89,40,0,0               // mulps         0x2859(%rip),%xmm9        # 51a0 <_sk_callback_sse2+0x71c>
-  .byte  68,15,84,21,97,40,0,0               // andps         0x2861(%rip),%xmm10        # 51b0 <_sk_callback_sse2+0x72c>
-  .byte  68,15,86,21,105,40,0,0              // orps          0x2869(%rip),%xmm10        # 51c0 <_sk_callback_sse2+0x73c>
-  .byte  68,15,88,13,113,40,0,0              // addps         0x2871(%rip),%xmm9        # 51d0 <_sk_callback_sse2+0x74c>
-  .byte  68,15,40,37,121,40,0,0              // movaps        0x2879(%rip),%xmm12        # 51e0 <_sk_callback_sse2+0x75c>
+  .byte  68,15,89,13,89,40,0,0               // mulps         0x2859(%rip),%xmm9        # 51a0 <_sk_callback_sse2+0x71f>
+  .byte  68,15,84,21,97,40,0,0               // andps         0x2861(%rip),%xmm10        # 51b0 <_sk_callback_sse2+0x72f>
+  .byte  68,15,86,21,105,40,0,0              // orps          0x2869(%rip),%xmm10        # 51c0 <_sk_callback_sse2+0x73f>
+  .byte  68,15,88,13,113,40,0,0              // addps         0x2871(%rip),%xmm9        # 51d0 <_sk_callback_sse2+0x74f>
+  .byte  68,15,40,37,121,40,0,0              // movaps        0x2879(%rip),%xmm12        # 51e0 <_sk_callback_sse2+0x75f>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,121,40,0,0              // addps         0x2879(%rip),%xmm10        # 51f0 <_sk_callback_sse2+0x76c>
-  .byte  68,15,40,37,129,40,0,0              // movaps        0x2881(%rip),%xmm12        # 5200 <_sk_callback_sse2+0x77c>
+  .byte  68,15,88,21,121,40,0,0              // addps         0x2879(%rip),%xmm10        # 51f0 <_sk_callback_sse2+0x76f>
+  .byte  68,15,40,37,129,40,0,0              // movaps        0x2881(%rip),%xmm12        # 5200 <_sk_callback_sse2+0x77f>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -29862,22 +29855,22 @@ _sk_parametric_g_sse2:
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,107,40,0,0              // movaps        0x286b(%rip),%xmm10        # 5210 <_sk_callback_sse2+0x78c>
+  .byte  68,15,40,21,107,40,0,0              // movaps        0x286b(%rip),%xmm10        # 5210 <_sk_callback_sse2+0x78f>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,95,40,0,0               // addps         0x285f(%rip),%xmm9        # 5220 <_sk_callback_sse2+0x79c>
-  .byte  68,15,40,37,103,40,0,0              // movaps        0x2867(%rip),%xmm12        # 5230 <_sk_callback_sse2+0x7ac>
+  .byte  68,15,88,13,95,40,0,0               // addps         0x285f(%rip),%xmm9        # 5220 <_sk_callback_sse2+0x79f>
+  .byte  68,15,40,37,103,40,0,0              // movaps        0x2867(%rip),%xmm12        # 5230 <_sk_callback_sse2+0x7af>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,103,40,0,0              // movaps        0x2867(%rip),%xmm12        # 5240 <_sk_callback_sse2+0x7bc>
+  .byte  68,15,40,37,103,40,0,0              // movaps        0x2867(%rip),%xmm12        # 5240 <_sk_callback_sse2+0x7bf>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,107,40,0,0              // movaps        0x286b(%rip),%xmm13        # 5250 <_sk_callback_sse2+0x7cc>
+  .byte  68,15,40,45,107,40,0,0              // movaps        0x286b(%rip),%xmm13        # 5250 <_sk_callback_sse2+0x7cf>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,107,40,0,0              // mulps         0x286b(%rip),%xmm13        # 5260 <_sk_callback_sse2+0x7dc>
+  .byte  68,15,89,45,107,40,0,0              // mulps         0x286b(%rip),%xmm13        # 5260 <_sk_callback_sse2+0x7df>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -29913,15 +29906,15 @@ _sk_parametric_b_sse2:
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,235,39,0,0              // mulps         0x27eb(%rip),%xmm9        # 5270 <_sk_callback_sse2+0x7ec>
-  .byte  68,15,84,21,243,39,0,0              // andps         0x27f3(%rip),%xmm10        # 5280 <_sk_callback_sse2+0x7fc>
-  .byte  68,15,86,21,251,39,0,0              // orps          0x27fb(%rip),%xmm10        # 5290 <_sk_callback_sse2+0x80c>
-  .byte  68,15,88,13,3,40,0,0                // addps         0x2803(%rip),%xmm9        # 52a0 <_sk_callback_sse2+0x81c>
-  .byte  68,15,40,37,11,40,0,0               // movaps        0x280b(%rip),%xmm12        # 52b0 <_sk_callback_sse2+0x82c>
+  .byte  68,15,89,13,235,39,0,0              // mulps         0x27eb(%rip),%xmm9        # 5270 <_sk_callback_sse2+0x7ef>
+  .byte  68,15,84,21,243,39,0,0              // andps         0x27f3(%rip),%xmm10        # 5280 <_sk_callback_sse2+0x7ff>
+  .byte  68,15,86,21,251,39,0,0              // orps          0x27fb(%rip),%xmm10        # 5290 <_sk_callback_sse2+0x80f>
+  .byte  68,15,88,13,3,40,0,0                // addps         0x2803(%rip),%xmm9        # 52a0 <_sk_callback_sse2+0x81f>
+  .byte  68,15,40,37,11,40,0,0               // movaps        0x280b(%rip),%xmm12        # 52b0 <_sk_callback_sse2+0x82f>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,11,40,0,0               // addps         0x280b(%rip),%xmm10        # 52c0 <_sk_callback_sse2+0x83c>
-  .byte  68,15,40,37,19,40,0,0               // movaps        0x2813(%rip),%xmm12        # 52d0 <_sk_callback_sse2+0x84c>
+  .byte  68,15,88,21,11,40,0,0               // addps         0x280b(%rip),%xmm10        # 52c0 <_sk_callback_sse2+0x83f>
+  .byte  68,15,40,37,19,40,0,0               // movaps        0x2813(%rip),%xmm12        # 52d0 <_sk_callback_sse2+0x84f>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -29929,22 +29922,22 @@ _sk_parametric_b_sse2:
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,253,39,0,0              // movaps        0x27fd(%rip),%xmm10        # 52e0 <_sk_callback_sse2+0x85c>
+  .byte  68,15,40,21,253,39,0,0              // movaps        0x27fd(%rip),%xmm10        # 52e0 <_sk_callback_sse2+0x85f>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,241,39,0,0              // addps         0x27f1(%rip),%xmm9        # 52f0 <_sk_callback_sse2+0x86c>
-  .byte  68,15,40,37,249,39,0,0              // movaps        0x27f9(%rip),%xmm12        # 5300 <_sk_callback_sse2+0x87c>
+  .byte  68,15,88,13,241,39,0,0              // addps         0x27f1(%rip),%xmm9        # 52f0 <_sk_callback_sse2+0x86f>
+  .byte  68,15,40,37,249,39,0,0              // movaps        0x27f9(%rip),%xmm12        # 5300 <_sk_callback_sse2+0x87f>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,249,39,0,0              // movaps        0x27f9(%rip),%xmm12        # 5310 <_sk_callback_sse2+0x88c>
+  .byte  68,15,40,37,249,39,0,0              // movaps        0x27f9(%rip),%xmm12        # 5310 <_sk_callback_sse2+0x88f>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,253,39,0,0              // movaps        0x27fd(%rip),%xmm13        # 5320 <_sk_callback_sse2+0x89c>
+  .byte  68,15,40,45,253,39,0,0              // movaps        0x27fd(%rip),%xmm13        # 5320 <_sk_callback_sse2+0x89f>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,253,39,0,0              // mulps         0x27fd(%rip),%xmm13        # 5330 <_sk_callback_sse2+0x8ac>
+  .byte  68,15,89,45,253,39,0,0              // mulps         0x27fd(%rip),%xmm13        # 5330 <_sk_callback_sse2+0x8af>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -29980,15 +29973,15 @@ _sk_parametric_a_sse2:
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,125,39,0,0              // mulps         0x277d(%rip),%xmm9        # 5340 <_sk_callback_sse2+0x8bc>
-  .byte  68,15,84,21,133,39,0,0              // andps         0x2785(%rip),%xmm10        # 5350 <_sk_callback_sse2+0x8cc>
-  .byte  68,15,86,21,141,39,0,0              // orps          0x278d(%rip),%xmm10        # 5360 <_sk_callback_sse2+0x8dc>
-  .byte  68,15,88,13,149,39,0,0              // addps         0x2795(%rip),%xmm9        # 5370 <_sk_callback_sse2+0x8ec>
-  .byte  68,15,40,37,157,39,0,0              // movaps        0x279d(%rip),%xmm12        # 5380 <_sk_callback_sse2+0x8fc>
+  .byte  68,15,89,13,125,39,0,0              // mulps         0x277d(%rip),%xmm9        # 5340 <_sk_callback_sse2+0x8bf>
+  .byte  68,15,84,21,133,39,0,0              // andps         0x2785(%rip),%xmm10        # 5350 <_sk_callback_sse2+0x8cf>
+  .byte  68,15,86,21,141,39,0,0              // orps          0x278d(%rip),%xmm10        # 5360 <_sk_callback_sse2+0x8df>
+  .byte  68,15,88,13,149,39,0,0              // addps         0x2795(%rip),%xmm9        # 5370 <_sk_callback_sse2+0x8ef>
+  .byte  68,15,40,37,157,39,0,0              // movaps        0x279d(%rip),%xmm12        # 5380 <_sk_callback_sse2+0x8ff>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,157,39,0,0              // addps         0x279d(%rip),%xmm10        # 5390 <_sk_callback_sse2+0x90c>
-  .byte  68,15,40,37,165,39,0,0              // movaps        0x27a5(%rip),%xmm12        # 53a0 <_sk_callback_sse2+0x91c>
+  .byte  68,15,88,21,157,39,0,0              // addps         0x279d(%rip),%xmm10        # 5390 <_sk_callback_sse2+0x90f>
+  .byte  68,15,40,37,165,39,0,0              // movaps        0x27a5(%rip),%xmm12        # 53a0 <_sk_callback_sse2+0x91f>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -29996,22 +29989,22 @@ _sk_parametric_a_sse2:
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,143,39,0,0              // movaps        0x278f(%rip),%xmm10        # 53b0 <_sk_callback_sse2+0x92c>
+  .byte  68,15,40,21,143,39,0,0              // movaps        0x278f(%rip),%xmm10        # 53b0 <_sk_callback_sse2+0x92f>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,131,39,0,0              // addps         0x2783(%rip),%xmm9        # 53c0 <_sk_callback_sse2+0x93c>
-  .byte  68,15,40,37,139,39,0,0              // movaps        0x278b(%rip),%xmm12        # 53d0 <_sk_callback_sse2+0x94c>
+  .byte  68,15,88,13,131,39,0,0              // addps         0x2783(%rip),%xmm9        # 53c0 <_sk_callback_sse2+0x93f>
+  .byte  68,15,40,37,139,39,0,0              // movaps        0x278b(%rip),%xmm12        # 53d0 <_sk_callback_sse2+0x94f>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,139,39,0,0              // movaps        0x278b(%rip),%xmm12        # 53e0 <_sk_callback_sse2+0x95c>
+  .byte  68,15,40,37,139,39,0,0              // movaps        0x278b(%rip),%xmm12        # 53e0 <_sk_callback_sse2+0x95f>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,143,39,0,0              // movaps        0x278f(%rip),%xmm13        # 53f0 <_sk_callback_sse2+0x96c>
+  .byte  68,15,40,45,143,39,0,0              // movaps        0x278f(%rip),%xmm13        # 53f0 <_sk_callback_sse2+0x96f>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,143,39,0,0              // mulps         0x278f(%rip),%xmm13        # 5400 <_sk_callback_sse2+0x97c>
+  .byte  68,15,89,45,143,39,0,0              // mulps         0x278f(%rip),%xmm13        # 5400 <_sk_callback_sse2+0x97f>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -30028,29 +30021,29 @@ HIDDEN _sk_lab_to_xyz_sse2
 .globl _sk_lab_to_xyz_sse2
 FUNCTION(_sk_lab_to_xyz_sse2)
 _sk_lab_to_xyz_sse2:
-  .byte  15,89,5,108,39,0,0                  // mulps         0x276c(%rip),%xmm0        # 5410 <_sk_callback_sse2+0x98c>
-  .byte  68,15,40,5,116,39,0,0               // movaps        0x2774(%rip),%xmm8        # 5420 <_sk_callback_sse2+0x99c>
+  .byte  15,89,5,108,39,0,0                  // mulps         0x276c(%rip),%xmm0        # 5410 <_sk_callback_sse2+0x98f>
+  .byte  68,15,40,5,116,39,0,0               // movaps        0x2774(%rip),%xmm8        # 5420 <_sk_callback_sse2+0x99f>
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
-  .byte  68,15,40,13,120,39,0,0              // movaps        0x2778(%rip),%xmm9        # 5430 <_sk_callback_sse2+0x9ac>
+  .byte  68,15,40,13,120,39,0,0              // movaps        0x2778(%rip),%xmm9        # 5430 <_sk_callback_sse2+0x9af>
   .byte  65,15,88,201                        // addps         %xmm9,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  65,15,88,209                        // addps         %xmm9,%xmm2
-  .byte  15,88,5,117,39,0,0                  // addps         0x2775(%rip),%xmm0        # 5440 <_sk_callback_sse2+0x9bc>
-  .byte  15,89,5,126,39,0,0                  // mulps         0x277e(%rip),%xmm0        # 5450 <_sk_callback_sse2+0x9cc>
-  .byte  15,89,13,135,39,0,0                 // mulps         0x2787(%rip),%xmm1        # 5460 <_sk_callback_sse2+0x9dc>
+  .byte  15,88,5,117,39,0,0                  // addps         0x2775(%rip),%xmm0        # 5440 <_sk_callback_sse2+0x9bf>
+  .byte  15,89,5,126,39,0,0                  // mulps         0x277e(%rip),%xmm0        # 5450 <_sk_callback_sse2+0x9cf>
+  .byte  15,89,13,135,39,0,0                 // mulps         0x2787(%rip),%xmm1        # 5460 <_sk_callback_sse2+0x9df>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,21,141,39,0,0                 // mulps         0x278d(%rip),%xmm2        # 5470 <_sk_callback_sse2+0x9ec>
+  .byte  15,89,21,141,39,0,0                 // mulps         0x278d(%rip),%xmm2        # 5470 <_sk_callback_sse2+0x9ef>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  68,15,92,202                        // subps         %xmm2,%xmm9
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
   .byte  68,15,89,225                        // mulps         %xmm1,%xmm12
-  .byte  15,40,21,130,39,0,0                 // movaps        0x2782(%rip),%xmm2        # 5480 <_sk_callback_sse2+0x9fc>
+  .byte  15,40,21,130,39,0,0                 // movaps        0x2782(%rip),%xmm2        # 5480 <_sk_callback_sse2+0x9ff>
   .byte  68,15,40,194                        // movaps        %xmm2,%xmm8
   .byte  69,15,194,196,1                     // cmpltps       %xmm12,%xmm8
-  .byte  68,15,40,21,129,39,0,0              // movaps        0x2781(%rip),%xmm10        # 5490 <_sk_callback_sse2+0xa0c>
+  .byte  68,15,40,21,129,39,0,0              // movaps        0x2781(%rip),%xmm10        # 5490 <_sk_callback_sse2+0xa0f>
   .byte  65,15,88,202                        // addps         %xmm10,%xmm1
-  .byte  68,15,40,29,133,39,0,0              // movaps        0x2785(%rip),%xmm11        # 54a0 <_sk_callback_sse2+0xa1c>
+  .byte  68,15,40,29,133,39,0,0              // movaps        0x2785(%rip),%xmm11        # 54a0 <_sk_callback_sse2+0xa1f>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  69,15,84,224                        // andps         %xmm8,%xmm12
   .byte  68,15,85,193                        // andnps        %xmm1,%xmm8
@@ -30074,8 +30067,8 @@ _sk_lab_to_xyz_sse2:
   .byte  15,84,194                           // andps         %xmm2,%xmm0
   .byte  65,15,85,209                        // andnps        %xmm9,%xmm2
   .byte  15,86,208                           // orps          %xmm0,%xmm2
-  .byte  68,15,89,5,53,39,0,0                // mulps         0x2735(%rip),%xmm8        # 54b0 <_sk_callback_sse2+0xa2c>
-  .byte  15,89,21,62,39,0,0                  // mulps         0x273e(%rip),%xmm2        # 54c0 <_sk_callback_sse2+0xa3c>
+  .byte  68,15,89,5,53,39,0,0                // mulps         0x2735(%rip),%xmm8        # 54b0 <_sk_callback_sse2+0xa2f>
+  .byte  15,89,21,62,39,0,0                  // mulps         0x273e(%rip),%xmm2        # 54c0 <_sk_callback_sse2+0xa3f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -30091,7 +30084,7 @@ _sk_load_a8_sse2:
   .byte  102,15,96,193                       // punpcklbw     %xmm1,%xmm0
   .byte  102,15,97,193                       // punpcklwd     %xmm1,%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,38,39,0,0                  // mulps         0x2726(%rip),%xmm3        # 54d0 <_sk_callback_sse2+0xa4c>
+  .byte  15,89,29,38,39,0,0                  // mulps         0x2726(%rip),%xmm3        # 54d0 <_sk_callback_sse2+0xa4f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -30136,7 +30129,7 @@ _sk_gather_a8_sse2:
   .byte  102,15,96,193                       // punpcklbw     %xmm1,%xmm0
   .byte  102,15,97,193                       // punpcklwd     %xmm1,%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,149,38,0,0                 // mulps         0x2695(%rip),%xmm3        # 54e0 <_sk_callback_sse2+0xa5c>
+  .byte  15,89,29,149,38,0,0                 // mulps         0x2695(%rip),%xmm3        # 54e0 <_sk_callback_sse2+0xa5f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -30149,7 +30142,7 @@ FUNCTION(_sk_store_a8_sse2)
 _sk_store_a8_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,137,38,0,0               // movaps        0x2689(%rip),%xmm8        # 54f0 <_sk_callback_sse2+0xa6c>
+  .byte  68,15,40,5,137,38,0,0               // movaps        0x2689(%rip),%xmm8        # 54f0 <_sk_callback_sse2+0xa6f>
   .byte  68,15,89,195                        // mulps         %xmm3,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
   .byte  102,65,15,114,240,16                // pslld         $0x10,%xmm8
@@ -30171,9 +30164,9 @@ _sk_load_g8_sse2:
   .byte  102,15,96,193                       // punpcklbw     %xmm1,%xmm0
   .byte  102,15,97,193                       // punpcklwd     %xmm1,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,80,38,0,0                   // mulps         0x2650(%rip),%xmm0        # 5500 <_sk_callback_sse2+0xa7c>
+  .byte  15,89,5,80,38,0,0                   // mulps         0x2650(%rip),%xmm0        # 5500 <_sk_callback_sse2+0xa7f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,87,38,0,0                  // movaps        0x2657(%rip),%xmm3        # 5510 <_sk_callback_sse2+0xa8c>
+  .byte  15,40,29,87,38,0,0                  // movaps        0x2657(%rip),%xmm3        # 5510 <_sk_callback_sse2+0xa8f>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -30216,9 +30209,9 @@ _sk_gather_g8_sse2:
   .byte  102,15,96,193                       // punpcklbw     %xmm1,%xmm0
   .byte  102,15,97,193                       // punpcklwd     %xmm1,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,204,37,0,0                  // mulps         0x25cc(%rip),%xmm0        # 5520 <_sk_callback_sse2+0xa9c>
+  .byte  15,89,5,204,37,0,0                  // mulps         0x25cc(%rip),%xmm0        # 5520 <_sk_callback_sse2+0xa9f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,211,37,0,0                 // movaps        0x25d3(%rip),%xmm3        # 5530 <_sk_callback_sse2+0xaac>
+  .byte  15,40,29,211,37,0,0                 // movaps        0x25d3(%rip),%xmm3        # 5530 <_sk_callback_sse2+0xaaf>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -30281,11 +30274,11 @@ _sk_gather_i8_sse2:
   .byte  102,67,15,110,12,136                // movd          (%r8,%r9,4),%xmm1
   .byte  102,68,15,98,201                    // punpckldq     %xmm1,%xmm9
   .byte  102,68,15,98,200                    // punpckldq     %xmm0,%xmm9
-  .byte  102,15,111,21,242,36,0,0            // movdqa        0x24f2(%rip),%xmm2        # 5540 <_sk_callback_sse2+0xabc>
+  .byte  102,15,111,21,242,36,0,0            // movdqa        0x24f2(%rip),%xmm2        # 5540 <_sk_callback_sse2+0xabf>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,238,36,0,0               // movaps        0x24ee(%rip),%xmm8        # 5550 <_sk_callback_sse2+0xacc>
+  .byte  68,15,40,5,238,36,0,0               // movaps        0x24ee(%rip),%xmm8        # 5550 <_sk_callback_sse2+0xacf>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -30312,19 +30305,19 @@ _sk_load_565_sse2:
   .byte  243,15,126,20,120                   // movq          (%rax,%rdi,2),%xmm2
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,97,208                       // punpcklwd     %xmm0,%xmm2
-  .byte  102,15,111,5,164,36,0,0             // movdqa        0x24a4(%rip),%xmm0        # 5560 <_sk_callback_sse2+0xadc>
+  .byte  102,15,111,5,164,36,0,0             // movdqa        0x24a4(%rip),%xmm0        # 5560 <_sk_callback_sse2+0xadf>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,166,36,0,0                  // mulps         0x24a6(%rip),%xmm0        # 5570 <_sk_callback_sse2+0xaec>
-  .byte  102,15,111,13,174,36,0,0            // movdqa        0x24ae(%rip),%xmm1        # 5580 <_sk_callback_sse2+0xafc>
+  .byte  15,89,5,166,36,0,0                  // mulps         0x24a6(%rip),%xmm0        # 5570 <_sk_callback_sse2+0xaef>
+  .byte  102,15,111,13,174,36,0,0            // movdqa        0x24ae(%rip),%xmm1        # 5580 <_sk_callback_sse2+0xaff>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,176,36,0,0                 // mulps         0x24b0(%rip),%xmm1        # 5590 <_sk_callback_sse2+0xb0c>
-  .byte  102,15,219,21,184,36,0,0            // pand          0x24b8(%rip),%xmm2        # 55a0 <_sk_callback_sse2+0xb1c>
+  .byte  15,89,13,176,36,0,0                 // mulps         0x24b0(%rip),%xmm1        # 5590 <_sk_callback_sse2+0xb0f>
+  .byte  102,15,219,21,184,36,0,0            // pand          0x24b8(%rip),%xmm2        # 55a0 <_sk_callback_sse2+0xb1f>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,190,36,0,0                 // mulps         0x24be(%rip),%xmm2        # 55b0 <_sk_callback_sse2+0xb2c>
+  .byte  15,89,21,190,36,0,0                 // mulps         0x24be(%rip),%xmm2        # 55b0 <_sk_callback_sse2+0xb2f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,197,36,0,0                 // movaps        0x24c5(%rip),%xmm3        # 55c0 <_sk_callback_sse2+0xb3c>
+  .byte  15,40,29,197,36,0,0                 // movaps        0x24c5(%rip),%xmm3        # 55c0 <_sk_callback_sse2+0xb3f>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_gather_565_sse2
@@ -30359,19 +30352,19 @@ _sk_gather_565_sse2:
   .byte  102,15,196,208,3                    // pinsrw        $0x3,%eax,%xmm2
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,97,208                       // punpcklwd     %xmm0,%xmm2
-  .byte  102,15,111,5,78,36,0,0              // movdqa        0x244e(%rip),%xmm0        # 55d0 <_sk_callback_sse2+0xb4c>
+  .byte  102,15,111,5,78,36,0,0              // movdqa        0x244e(%rip),%xmm0        # 55d0 <_sk_callback_sse2+0xb4f>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,80,36,0,0                   // mulps         0x2450(%rip),%xmm0        # 55e0 <_sk_callback_sse2+0xb5c>
-  .byte  102,15,111,13,88,36,0,0             // movdqa        0x2458(%rip),%xmm1        # 55f0 <_sk_callback_sse2+0xb6c>
+  .byte  15,89,5,80,36,0,0                   // mulps         0x2450(%rip),%xmm0        # 55e0 <_sk_callback_sse2+0xb5f>
+  .byte  102,15,111,13,88,36,0,0             // movdqa        0x2458(%rip),%xmm1        # 55f0 <_sk_callback_sse2+0xb6f>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,90,36,0,0                  // mulps         0x245a(%rip),%xmm1        # 5600 <_sk_callback_sse2+0xb7c>
-  .byte  102,15,219,21,98,36,0,0             // pand          0x2462(%rip),%xmm2        # 5610 <_sk_callback_sse2+0xb8c>
+  .byte  15,89,13,90,36,0,0                  // mulps         0x245a(%rip),%xmm1        # 5600 <_sk_callback_sse2+0xb7f>
+  .byte  102,15,219,21,98,36,0,0             // pand          0x2462(%rip),%xmm2        # 5610 <_sk_callback_sse2+0xb8f>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,104,36,0,0                 // mulps         0x2468(%rip),%xmm2        # 5620 <_sk_callback_sse2+0xb9c>
+  .byte  15,89,21,104,36,0,0                 // mulps         0x2468(%rip),%xmm2        # 5620 <_sk_callback_sse2+0xb9f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,111,36,0,0                 // movaps        0x246f(%rip),%xmm3        # 5630 <_sk_callback_sse2+0xbac>
+  .byte  15,40,29,111,36,0,0                 // movaps        0x246f(%rip),%xmm3        # 5630 <_sk_callback_sse2+0xbaf>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_565_sse2
@@ -30380,12 +30373,12 @@ FUNCTION(_sk_store_565_sse2)
 _sk_store_565_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,112,36,0,0               // movaps        0x2470(%rip),%xmm8        # 5640 <_sk_callback_sse2+0xbbc>
+  .byte  68,15,40,5,112,36,0,0               // movaps        0x2470(%rip),%xmm8        # 5640 <_sk_callback_sse2+0xbbf>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
   .byte  102,65,15,114,241,11                // pslld         $0xb,%xmm9
-  .byte  68,15,40,21,101,36,0,0              // movaps        0x2465(%rip),%xmm10        # 5650 <_sk_callback_sse2+0xbcc>
+  .byte  68,15,40,21,101,36,0,0              // movaps        0x2465(%rip),%xmm10        # 5650 <_sk_callback_sse2+0xbcf>
   .byte  68,15,89,209                        // mulps         %xmm1,%xmm10
   .byte  102,69,15,91,210                    // cvtps2dq      %xmm10,%xmm10
   .byte  102,65,15,114,242,5                 // pslld         $0x5,%xmm10
@@ -30409,21 +30402,21 @@ _sk_load_4444_sse2:
   .byte  243,15,126,28,120                   // movq          (%rax,%rdi,2),%xmm3
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,97,216                       // punpcklwd     %xmm0,%xmm3
-  .byte  102,15,111,5,30,36,0,0              // movdqa        0x241e(%rip),%xmm0        # 5660 <_sk_callback_sse2+0xbdc>
+  .byte  102,15,111,5,30,36,0,0              // movdqa        0x241e(%rip),%xmm0        # 5660 <_sk_callback_sse2+0xbdf>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,32,36,0,0                   // mulps         0x2420(%rip),%xmm0        # 5670 <_sk_callback_sse2+0xbec>
-  .byte  102,15,111,13,40,36,0,0             // movdqa        0x2428(%rip),%xmm1        # 5680 <_sk_callback_sse2+0xbfc>
+  .byte  15,89,5,32,36,0,0                   // mulps         0x2420(%rip),%xmm0        # 5670 <_sk_callback_sse2+0xbef>
+  .byte  102,15,111,13,40,36,0,0             // movdqa        0x2428(%rip),%xmm1        # 5680 <_sk_callback_sse2+0xbff>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,42,36,0,0                  // mulps         0x242a(%rip),%xmm1        # 5690 <_sk_callback_sse2+0xc0c>
-  .byte  102,15,111,21,50,36,0,0             // movdqa        0x2432(%rip),%xmm2        # 56a0 <_sk_callback_sse2+0xc1c>
+  .byte  15,89,13,42,36,0,0                  // mulps         0x242a(%rip),%xmm1        # 5690 <_sk_callback_sse2+0xc0f>
+  .byte  102,15,111,21,50,36,0,0             // movdqa        0x2432(%rip),%xmm2        # 56a0 <_sk_callback_sse2+0xc1f>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,52,36,0,0                  // mulps         0x2434(%rip),%xmm2        # 56b0 <_sk_callback_sse2+0xc2c>
-  .byte  102,15,219,29,60,36,0,0             // pand          0x243c(%rip),%xmm3        # 56c0 <_sk_callback_sse2+0xc3c>
+  .byte  15,89,21,52,36,0,0                  // mulps         0x2434(%rip),%xmm2        # 56b0 <_sk_callback_sse2+0xc2f>
+  .byte  102,15,219,29,60,36,0,0             // pand          0x243c(%rip),%xmm3        # 56c0 <_sk_callback_sse2+0xc3f>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,66,36,0,0                  // mulps         0x2442(%rip),%xmm3        # 56d0 <_sk_callback_sse2+0xc4c>
+  .byte  15,89,29,66,36,0,0                  // mulps         0x2442(%rip),%xmm3        # 56d0 <_sk_callback_sse2+0xc4f>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -30459,21 +30452,21 @@ _sk_gather_4444_sse2:
   .byte  102,15,196,216,3                    // pinsrw        $0x3,%eax,%xmm3
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,97,216                       // punpcklwd     %xmm0,%xmm3
-  .byte  102,15,111,5,201,35,0,0             // movdqa        0x23c9(%rip),%xmm0        # 56e0 <_sk_callback_sse2+0xc5c>
+  .byte  102,15,111,5,201,35,0,0             // movdqa        0x23c9(%rip),%xmm0        # 56e0 <_sk_callback_sse2+0xc5f>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,203,35,0,0                  // mulps         0x23cb(%rip),%xmm0        # 56f0 <_sk_callback_sse2+0xc6c>
-  .byte  102,15,111,13,211,35,0,0            // movdqa        0x23d3(%rip),%xmm1        # 5700 <_sk_callback_sse2+0xc7c>
+  .byte  15,89,5,203,35,0,0                  // mulps         0x23cb(%rip),%xmm0        # 56f0 <_sk_callback_sse2+0xc6f>
+  .byte  102,15,111,13,211,35,0,0            // movdqa        0x23d3(%rip),%xmm1        # 5700 <_sk_callback_sse2+0xc7f>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,213,35,0,0                 // mulps         0x23d5(%rip),%xmm1        # 5710 <_sk_callback_sse2+0xc8c>
-  .byte  102,15,111,21,221,35,0,0            // movdqa        0x23dd(%rip),%xmm2        # 5720 <_sk_callback_sse2+0xc9c>
+  .byte  15,89,13,213,35,0,0                 // mulps         0x23d5(%rip),%xmm1        # 5710 <_sk_callback_sse2+0xc8f>
+  .byte  102,15,111,21,221,35,0,0            // movdqa        0x23dd(%rip),%xmm2        # 5720 <_sk_callback_sse2+0xc9f>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,223,35,0,0                 // mulps         0x23df(%rip),%xmm2        # 5730 <_sk_callback_sse2+0xcac>
-  .byte  102,15,219,29,231,35,0,0            // pand          0x23e7(%rip),%xmm3        # 5740 <_sk_callback_sse2+0xcbc>
+  .byte  15,89,21,223,35,0,0                 // mulps         0x23df(%rip),%xmm2        # 5730 <_sk_callback_sse2+0xcaf>
+  .byte  102,15,219,29,231,35,0,0            // pand          0x23e7(%rip),%xmm3        # 5740 <_sk_callback_sse2+0xcbf>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,237,35,0,0                 // mulps         0x23ed(%rip),%xmm3        # 5750 <_sk_callback_sse2+0xccc>
+  .byte  15,89,29,237,35,0,0                 // mulps         0x23ed(%rip),%xmm3        # 5750 <_sk_callback_sse2+0xccf>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -30483,7 +30476,7 @@ FUNCTION(_sk_store_4444_sse2)
 _sk_store_4444_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,236,35,0,0               // movaps        0x23ec(%rip),%xmm8        # 5760 <_sk_callback_sse2+0xcdc>
+  .byte  68,15,40,5,236,35,0,0               // movaps        0x23ec(%rip),%xmm8        # 5760 <_sk_callback_sse2+0xcdf>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -30515,11 +30508,11 @@ _sk_load_8888_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  68,15,16,12,184                     // movups        (%rax,%rdi,4),%xmm9
-  .byte  15,40,21,127,35,0,0                 // movaps        0x237f(%rip),%xmm2        # 5770 <_sk_callback_sse2+0xcec>
+  .byte  15,40,21,127,35,0,0                 // movaps        0x237f(%rip),%xmm2        # 5770 <_sk_callback_sse2+0xcef>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,84,194                           // andps         %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,125,35,0,0               // movaps        0x237d(%rip),%xmm8        # 5780 <_sk_callback_sse2+0xcfc>
+  .byte  68,15,40,5,125,35,0,0               // movaps        0x237d(%rip),%xmm8        # 5780 <_sk_callback_sse2+0xcff>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,40,201                        // movaps        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -30568,11 +30561,11 @@ _sk_gather_8888_sse2:
   .byte  102,67,15,110,12,129                // movd          (%r9,%r8,4),%xmm1
   .byte  102,68,15,98,201                    // punpckldq     %xmm1,%xmm9
   .byte  102,68,15,98,200                    // punpckldq     %xmm0,%xmm9
-  .byte  102,15,111,21,206,34,0,0            // movdqa        0x22ce(%rip),%xmm2        # 5790 <_sk_callback_sse2+0xd0c>
+  .byte  102,15,111,21,206,34,0,0            // movdqa        0x22ce(%rip),%xmm2        # 5790 <_sk_callback_sse2+0xd0f>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,202,34,0,0               // movaps        0x22ca(%rip),%xmm8        # 57a0 <_sk_callback_sse2+0xd1c>
+  .byte  68,15,40,5,202,34,0,0               // movaps        0x22ca(%rip),%xmm8        # 57a0 <_sk_callback_sse2+0xd1f>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -30596,7 +30589,7 @@ FUNCTION(_sk_store_8888_sse2)
 _sk_store_8888_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,141,34,0,0               // movaps        0x228d(%rip),%xmm8        # 57b0 <_sk_callback_sse2+0xd2c>
+  .byte  68,15,40,5,141,34,0,0               // movaps        0x228d(%rip),%xmm8        # 57b0 <_sk_callback_sse2+0xd2f>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -30635,7 +30628,7 @@ _sk_load_f16_sse2:
   .byte  102,69,15,239,210                   // pxor          %xmm10,%xmm10
   .byte  102,65,15,111,206                   // movdqa        %xmm14,%xmm1
   .byte  102,65,15,97,202                    // punpcklwd     %xmm10,%xmm1
-  .byte  102,68,15,111,13,253,33,0,0         // movdqa        0x21fd(%rip),%xmm9        # 57c0 <_sk_callback_sse2+0xd3c>
+  .byte  102,68,15,111,13,253,33,0,0         // movdqa        0x21fd(%rip),%xmm9        # 57c0 <_sk_callback_sse2+0xd3f>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
   .byte  102,15,239,200                      // pxor          %xmm0,%xmm1
@@ -30643,11 +30636,11 @@ _sk_load_f16_sse2:
   .byte  102,68,15,111,233                   // movdqa        %xmm1,%xmm13
   .byte  102,65,15,114,245,13                // pslld         $0xd,%xmm13
   .byte  102,68,15,235,232                   // por           %xmm0,%xmm13
-  .byte  102,68,15,111,29,226,33,0,0         // movdqa        0x21e2(%rip),%xmm11        # 57d0 <_sk_callback_sse2+0xd4c>
+  .byte  102,68,15,111,29,226,33,0,0         // movdqa        0x21e2(%rip),%xmm11        # 57d0 <_sk_callback_sse2+0xd4f>
   .byte  102,69,15,254,235                   // paddd         %xmm11,%xmm13
-  .byte  102,68,15,111,37,228,33,0,0         // movdqa        0x21e4(%rip),%xmm12        # 57e0 <_sk_callback_sse2+0xd5c>
+  .byte  102,68,15,111,37,228,33,0,0         // movdqa        0x21e4(%rip),%xmm12        # 57e0 <_sk_callback_sse2+0xd5f>
   .byte  102,65,15,239,204                   // pxor          %xmm12,%xmm1
-  .byte  102,15,111,29,231,33,0,0            // movdqa        0x21e7(%rip),%xmm3        # 57f0 <_sk_callback_sse2+0xd6c>
+  .byte  102,15,111,29,231,33,0,0            // movdqa        0x21e7(%rip),%xmm3        # 57f0 <_sk_callback_sse2+0xd6f>
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,102,193                      // pcmpgtd       %xmm1,%xmm0
   .byte  102,65,15,223,197                   // pandn         %xmm13,%xmm0
@@ -30733,7 +30726,7 @@ _sk_gather_f16_sse2:
   .byte  102,69,15,239,210                   // pxor          %xmm10,%xmm10
   .byte  102,65,15,111,206                   // movdqa        %xmm14,%xmm1
   .byte  102,65,15,97,202                    // punpcklwd     %xmm10,%xmm1
-  .byte  102,68,15,111,13,117,32,0,0         // movdqa        0x2075(%rip),%xmm9        # 5800 <_sk_callback_sse2+0xd7c>
+  .byte  102,68,15,111,13,117,32,0,0         // movdqa        0x2075(%rip),%xmm9        # 5800 <_sk_callback_sse2+0xd7f>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
   .byte  102,15,239,200                      // pxor          %xmm0,%xmm1
@@ -30741,11 +30734,11 @@ _sk_gather_f16_sse2:
   .byte  102,68,15,111,233                   // movdqa        %xmm1,%xmm13
   .byte  102,65,15,114,245,13                // pslld         $0xd,%xmm13
   .byte  102,68,15,235,232                   // por           %xmm0,%xmm13
-  .byte  102,68,15,111,29,90,32,0,0          // movdqa        0x205a(%rip),%xmm11        # 5810 <_sk_callback_sse2+0xd8c>
+  .byte  102,68,15,111,29,90,32,0,0          // movdqa        0x205a(%rip),%xmm11        # 5810 <_sk_callback_sse2+0xd8f>
   .byte  102,69,15,254,235                   // paddd         %xmm11,%xmm13
-  .byte  102,68,15,111,37,92,32,0,0          // movdqa        0x205c(%rip),%xmm12        # 5820 <_sk_callback_sse2+0xd9c>
+  .byte  102,68,15,111,37,92,32,0,0          // movdqa        0x205c(%rip),%xmm12        # 5820 <_sk_callback_sse2+0xd9f>
   .byte  102,65,15,239,204                   // pxor          %xmm12,%xmm1
-  .byte  102,15,111,29,95,32,0,0             // movdqa        0x205f(%rip),%xmm3        # 5830 <_sk_callback_sse2+0xdac>
+  .byte  102,15,111,29,95,32,0,0             // movdqa        0x205f(%rip),%xmm3        # 5830 <_sk_callback_sse2+0xdaf>
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,102,193                      // pcmpgtd       %xmm1,%xmm0
   .byte  102,65,15,223,197                   // pandn         %xmm13,%xmm0
@@ -30798,17 +30791,17 @@ FUNCTION(_sk_store_f16_sse2)
 _sk_store_f16_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  102,68,15,111,21,135,31,0,0         // movdqa        0x1f87(%rip),%xmm10        # 5840 <_sk_callback_sse2+0xdbc>
+  .byte  102,68,15,111,21,135,31,0,0         // movdqa        0x1f87(%rip),%xmm10        # 5840 <_sk_callback_sse2+0xdbf>
   .byte  102,68,15,111,224                   // movdqa        %xmm0,%xmm12
   .byte  102,68,15,111,232                   // movdqa        %xmm0,%xmm13
   .byte  102,69,15,219,234                   // pand          %xmm10,%xmm13
   .byte  102,69,15,239,229                   // pxor          %xmm13,%xmm12
-  .byte  102,68,15,111,13,122,31,0,0         // movdqa        0x1f7a(%rip),%xmm9        # 5850 <_sk_callback_sse2+0xdcc>
+  .byte  102,68,15,111,13,122,31,0,0         // movdqa        0x1f7a(%rip),%xmm9        # 5850 <_sk_callback_sse2+0xdcf>
   .byte  102,65,15,114,213,16                // psrld         $0x10,%xmm13
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
   .byte  102,69,15,102,196                   // pcmpgtd       %xmm12,%xmm8
   .byte  102,65,15,114,212,13                // psrld         $0xd,%xmm12
-  .byte  102,68,15,111,29,107,31,0,0         // movdqa        0x1f6b(%rip),%xmm11        # 5860 <_sk_callback_sse2+0xddc>
+  .byte  102,68,15,111,29,107,31,0,0         // movdqa        0x1f6b(%rip),%xmm11        # 5860 <_sk_callback_sse2+0xddf>
   .byte  102,69,15,235,235                   // por           %xmm11,%xmm13
   .byte  102,69,15,254,236                   // paddd         %xmm12,%xmm13
   .byte  102,65,15,114,245,16                // pslld         $0x10,%xmm13
@@ -30887,7 +30880,7 @@ _sk_load_u16_be_sse2:
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,65,15,97,201                    // punpcklwd     %xmm9,%xmm1
   .byte  15,91,193                           // cvtdq2ps      %xmm1,%xmm0
-  .byte  68,15,40,5,9,30,0,0                 // movaps        0x1e09(%rip),%xmm8        # 5870 <_sk_callback_sse2+0xdec>
+  .byte  68,15,40,5,9,30,0,0                 // movaps        0x1e09(%rip),%xmm8        # 5870 <_sk_callback_sse2+0xdef>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -30940,7 +30933,7 @@ _sk_load_rgb_u16_be_sse2:
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,65,15,97,192                    // punpcklwd     %xmm8,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,13,69,29,0,0               // movaps        0x1d45(%rip),%xmm9        # 5880 <_sk_callback_sse2+0xdfc>
+  .byte  68,15,40,13,69,29,0,0               // movaps        0x1d45(%rip),%xmm9        # 5880 <_sk_callback_sse2+0xdff>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -30957,7 +30950,7 @@ _sk_load_rgb_u16_be_sse2:
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,12,29,0,0                  // movaps        0x1d0c(%rip),%xmm3        # 5890 <_sk_callback_sse2+0xe0c>
+  .byte  15,40,29,12,29,0,0                  // movaps        0x1d0c(%rip),%xmm3        # 5890 <_sk_callback_sse2+0xe0f>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_u16_be_sse2
@@ -30966,7 +30959,7 @@ FUNCTION(_sk_store_u16_be_sse2)
 _sk_store_u16_be_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,13,13,29,0,0               // movaps        0x1d0d(%rip),%xmm9        # 58a0 <_sk_callback_sse2+0xe1c>
+  .byte  68,15,40,13,13,29,0,0               // movaps        0x1d0d(%rip),%xmm9        # 58a0 <_sk_callback_sse2+0xe1f>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
@@ -31112,7 +31105,7 @@ _sk_repeat_x_sse2:
   .byte  243,69,15,91,209                    // cvttps2dq     %xmm9,%xmm10
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,84,13,13,27,0,0               // andps         0x1b0d(%rip),%xmm9        # 58b0 <_sk_callback_sse2+0xe2c>
+  .byte  68,15,84,13,13,27,0,0               // andps         0x1b0d(%rip),%xmm9        # 58b0 <_sk_callback_sse2+0xe2f>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
   .byte  65,15,92,194                        // subps         %xmm10,%xmm0
@@ -31132,7 +31125,7 @@ _sk_repeat_y_sse2:
   .byte  243,69,15,91,209                    // cvttps2dq     %xmm9,%xmm10
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,84,13,223,26,0,0              // andps         0x1adf(%rip),%xmm9        # 58c0 <_sk_callback_sse2+0xe3c>
+  .byte  68,15,84,13,223,26,0,0              // andps         0x1adf(%rip),%xmm9        # 58c0 <_sk_callback_sse2+0xe3f>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
   .byte  65,15,92,202                        // subps         %xmm10,%xmm1
@@ -31156,7 +31149,7 @@ _sk_mirror_x_sse2:
   .byte  243,69,15,91,218                    // cvttps2dq     %xmm10,%xmm11
   .byte  69,15,91,219                        // cvtdq2ps      %xmm11,%xmm11
   .byte  69,15,194,211,1                     // cmpltps       %xmm11,%xmm10
-  .byte  68,15,84,21,159,26,0,0              // andps         0x1a9f(%rip),%xmm10        # 58d0 <_sk_callback_sse2+0xe4c>
+  .byte  68,15,84,21,159,26,0,0              // andps         0x1a9f(%rip),%xmm10        # 58d0 <_sk_callback_sse2+0xe4f>
   .byte  69,15,87,228                        // xorps         %xmm12,%xmm12
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
@@ -31184,7 +31177,7 @@ _sk_mirror_y_sse2:
   .byte  243,69,15,91,218                    // cvttps2dq     %xmm10,%xmm11
   .byte  69,15,91,219                        // cvtdq2ps      %xmm11,%xmm11
   .byte  69,15,194,211,1                     // cmpltps       %xmm11,%xmm10
-  .byte  68,15,84,21,79,26,0,0               // andps         0x1a4f(%rip),%xmm10        # 58e0 <_sk_callback_sse2+0xe5c>
+  .byte  68,15,84,21,79,26,0,0               // andps         0x1a4f(%rip),%xmm10        # 58e0 <_sk_callback_sse2+0xe5f>
   .byte  69,15,87,228                        // xorps         %xmm12,%xmm12
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
@@ -31201,10 +31194,10 @@ HIDDEN _sk_luminance_to_alpha_sse2
 FUNCTION(_sk_luminance_to_alpha_sse2)
 _sk_luminance_to_alpha_sse2:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
-  .byte  15,89,5,49,26,0,0                   // mulps         0x1a31(%rip),%xmm0        # 58f0 <_sk_callback_sse2+0xe6c>
-  .byte  15,89,13,58,26,0,0                  // mulps         0x1a3a(%rip),%xmm1        # 5900 <_sk_callback_sse2+0xe7c>
+  .byte  15,89,5,49,26,0,0                   // mulps         0x1a31(%rip),%xmm0        # 58f0 <_sk_callback_sse2+0xe6f>
+  .byte  15,89,13,58,26,0,0                  // mulps         0x1a3a(%rip),%xmm1        # 5900 <_sk_callback_sse2+0xe7f>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,29,64,26,0,0                  // mulps         0x1a40(%rip),%xmm3        # 5910 <_sk_callback_sse2+0xe8c>
+  .byte  15,89,29,64,26,0,0                  // mulps         0x1a40(%rip),%xmm3        # 5910 <_sk_callback_sse2+0xe8f>
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -31537,7 +31530,7 @@ _sk_gradient_sse2:
   .byte  73,255,200                          // dec           %r8
   .byte  72,131,193,4                        // add           $0x4,%rcx
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
-  .byte  15,40,21,21,21,0,0                  // movaps        0x1515(%rip),%xmm2        # 5920 <_sk_callback_sse2+0xe9c>
+  .byte  15,40,21,21,21,0,0                  // movaps        0x1515(%rip),%xmm2        # 5920 <_sk_callback_sse2+0xe9f>
   .byte  243,15,16,25                        // movss         (%rcx),%xmm3
   .byte  15,198,219,0                        // shufps        $0x0,%xmm3,%xmm3
   .byte  15,194,216,2                        // cmpleps       %xmm0,%xmm3
@@ -31685,29 +31678,29 @@ _sk_xy_to_unit_angle_sse2:
   .byte  69,15,94,220                        // divps         %xmm12,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
-  .byte  68,15,40,45,215,18,0,0              // movaps        0x12d7(%rip),%xmm13        # 5930 <_sk_callback_sse2+0xeac>
+  .byte  68,15,40,45,215,18,0,0              // movaps        0x12d7(%rip),%xmm13        # 5930 <_sk_callback_sse2+0xeaf>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,219,18,0,0              // addps         0x12db(%rip),%xmm13        # 5940 <_sk_callback_sse2+0xebc>
+  .byte  68,15,88,45,219,18,0,0              // addps         0x12db(%rip),%xmm13        # 5940 <_sk_callback_sse2+0xebf>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,223,18,0,0              // addps         0x12df(%rip),%xmm13        # 5950 <_sk_callback_sse2+0xecc>
+  .byte  68,15,88,45,223,18,0,0              // addps         0x12df(%rip),%xmm13        # 5950 <_sk_callback_sse2+0xecf>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,227,18,0,0              // addps         0x12e3(%rip),%xmm13        # 5960 <_sk_callback_sse2+0xedc>
+  .byte  68,15,88,45,227,18,0,0              // addps         0x12e3(%rip),%xmm13        # 5960 <_sk_callback_sse2+0xedf>
   .byte  69,15,89,235                        // mulps         %xmm11,%xmm13
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,40,21,226,18,0,0              // movaps        0x12e2(%rip),%xmm10        # 5970 <_sk_callback_sse2+0xeec>
+  .byte  68,15,40,21,226,18,0,0              // movaps        0x12e2(%rip),%xmm10        # 5970 <_sk_callback_sse2+0xeef>
   .byte  69,15,92,213                        // subps         %xmm13,%xmm10
   .byte  69,15,84,209                        // andps         %xmm9,%xmm10
   .byte  69,15,85,205                        // andnps        %xmm13,%xmm9
   .byte  69,15,86,202                        // orps          %xmm10,%xmm9
   .byte  68,15,194,192,1                     // cmpltps       %xmm0,%xmm8
-  .byte  68,15,40,21,213,18,0,0              // movaps        0x12d5(%rip),%xmm10        # 5980 <_sk_callback_sse2+0xefc>
+  .byte  68,15,40,21,213,18,0,0              // movaps        0x12d5(%rip),%xmm10        # 5980 <_sk_callback_sse2+0xeff>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,84,208                        // andps         %xmm8,%xmm10
   .byte  69,15,85,193                        // andnps        %xmm9,%xmm8
   .byte  69,15,86,194                        // orps          %xmm10,%xmm8
   .byte  68,15,40,201                        // movaps        %xmm1,%xmm9
   .byte  68,15,194,200,1                     // cmpltps       %xmm0,%xmm9
-  .byte  68,15,40,21,196,18,0,0              // movaps        0x12c4(%rip),%xmm10        # 5990 <_sk_callback_sse2+0xf0c>
+  .byte  68,15,40,21,196,18,0,0              // movaps        0x12c4(%rip),%xmm10        # 5990 <_sk_callback_sse2+0xf0f>
   .byte  69,15,92,208                        // subps         %xmm8,%xmm10
   .byte  69,15,84,209                        // andps         %xmm9,%xmm10
   .byte  69,15,85,200                        // andnps        %xmm8,%xmm9
@@ -31725,8 +31718,7 @@ _sk_xy_to_radius_sse2:
   .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
   .byte  68,15,88,192                        // addps         %xmm0,%xmm8
-  .byte  65,15,82,192                        // rsqrtps       %xmm8,%xmm0
-  .byte  15,83,192                           // rcpps         %xmm0,%xmm0
+  .byte  65,15,81,192                        // sqrtps        %xmm8,%xmm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -31735,7 +31727,7 @@ HIDDEN _sk_save_xy_sse2
 FUNCTION(_sk_save_xy_sse2)
 _sk_save_xy_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,147,18,0,0               // movaps        0x1293(%rip),%xmm8        # 59a0 <_sk_callback_sse2+0xf1c>
+  .byte  68,15,40,5,150,18,0,0               // movaps        0x1296(%rip),%xmm8        # 59a0 <_sk_callback_sse2+0xf1f>
   .byte  15,17,0                             // movups        %xmm0,(%rax)
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,88,200                        // addps         %xmm8,%xmm9
@@ -31743,7 +31735,7 @@ _sk_save_xy_sse2:
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
   .byte  69,15,40,217                        // movaps        %xmm9,%xmm11
   .byte  69,15,194,218,1                     // cmpltps       %xmm10,%xmm11
-  .byte  68,15,40,37,126,18,0,0              // movaps        0x127e(%rip),%xmm12        # 59b0 <_sk_callback_sse2+0xf2c>
+  .byte  68,15,40,37,129,18,0,0              // movaps        0x1281(%rip),%xmm12        # 59b0 <_sk_callback_sse2+0xf2f>
   .byte  69,15,84,220                        // andps         %xmm12,%xmm11
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
@@ -31790,8 +31782,8 @@ _sk_bilinear_nx_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,247,17,0,0                  // addps         0x11f7(%rip),%xmm0        # 59c0 <_sk_callback_sse2+0xf3c>
-  .byte  68,15,40,13,255,17,0,0              // movaps        0x11ff(%rip),%xmm9        # 59d0 <_sk_callback_sse2+0xf4c>
+  .byte  15,88,5,250,17,0,0                  // addps         0x11fa(%rip),%xmm0        # 59c0 <_sk_callback_sse2+0xf3f>
+  .byte  68,15,40,13,2,18,0,0                // movaps        0x1202(%rip),%xmm9        # 59d0 <_sk_callback_sse2+0xf4f>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,128,0,0,0              // movups        %xmm9,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31804,7 +31796,7 @@ _sk_bilinear_px_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,238,17,0,0                  // addps         0x11ee(%rip),%xmm0        # 59e0 <_sk_callback_sse2+0xf5c>
+  .byte  15,88,5,241,17,0,0                  // addps         0x11f1(%rip),%xmm0        # 59e0 <_sk_callback_sse2+0xf5f>
   .byte  68,15,17,128,128,0,0,0              // movups        %xmm8,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -31816,8 +31808,8 @@ _sk_bilinear_ny_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,224,17,0,0                 // addps         0x11e0(%rip),%xmm1        # 59f0 <_sk_callback_sse2+0xf6c>
-  .byte  68,15,40,13,232,17,0,0              // movaps        0x11e8(%rip),%xmm9        # 5a00 <_sk_callback_sse2+0xf7c>
+  .byte  15,88,13,227,17,0,0                 // addps         0x11e3(%rip),%xmm1        # 59f0 <_sk_callback_sse2+0xf6f>
+  .byte  68,15,40,13,235,17,0,0              // movaps        0x11eb(%rip),%xmm9        # 5a00 <_sk_callback_sse2+0xf7f>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,160,0,0,0              // movups        %xmm9,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31830,7 +31822,7 @@ _sk_bilinear_py_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,214,17,0,0                 // addps         0x11d6(%rip),%xmm1        # 5a10 <_sk_callback_sse2+0xf8c>
+  .byte  15,88,13,217,17,0,0                 // addps         0x11d9(%rip),%xmm1        # 5a10 <_sk_callback_sse2+0xf8f>
   .byte  68,15,17,128,160,0,0,0              // movups        %xmm8,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -31842,13 +31834,13 @@ _sk_bicubic_n3x_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,201,17,0,0                  // addps         0x11c9(%rip),%xmm0        # 5a20 <_sk_callback_sse2+0xf9c>
-  .byte  68,15,40,13,209,17,0,0              // movaps        0x11d1(%rip),%xmm9        # 5a30 <_sk_callback_sse2+0xfac>
+  .byte  15,88,5,204,17,0,0                  // addps         0x11cc(%rip),%xmm0        # 5a20 <_sk_callback_sse2+0xf9f>
+  .byte  68,15,40,13,212,17,0,0              // movaps        0x11d4(%rip),%xmm9        # 5a30 <_sk_callback_sse2+0xfaf>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,205,17,0,0              // mulps         0x11cd(%rip),%xmm9        # 5a40 <_sk_callback_sse2+0xfbc>
-  .byte  68,15,88,13,213,17,0,0              // addps         0x11d5(%rip),%xmm9        # 5a50 <_sk_callback_sse2+0xfcc>
+  .byte  68,15,89,13,208,17,0,0              // mulps         0x11d0(%rip),%xmm9        # 5a40 <_sk_callback_sse2+0xfbf>
+  .byte  68,15,88,13,216,17,0,0              // addps         0x11d8(%rip),%xmm9        # 5a50 <_sk_callback_sse2+0xfcf>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,128,0,0,0              // movups        %xmm9,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31861,16 +31853,16 @@ _sk_bicubic_n1x_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,196,17,0,0                  // addps         0x11c4(%rip),%xmm0        # 5a60 <_sk_callback_sse2+0xfdc>
-  .byte  68,15,40,13,204,17,0,0              // movaps        0x11cc(%rip),%xmm9        # 5a70 <_sk_callback_sse2+0xfec>
+  .byte  15,88,5,199,17,0,0                  // addps         0x11c7(%rip),%xmm0        # 5a60 <_sk_callback_sse2+0xfdf>
+  .byte  68,15,40,13,207,17,0,0              // movaps        0x11cf(%rip),%xmm9        # 5a70 <_sk_callback_sse2+0xfef>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,208,17,0,0               // movaps        0x11d0(%rip),%xmm8        # 5a80 <_sk_callback_sse2+0xffc>
+  .byte  68,15,40,5,211,17,0,0               // movaps        0x11d3(%rip),%xmm8        # 5a80 <_sk_callback_sse2+0xfff>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,212,17,0,0               // addps         0x11d4(%rip),%xmm8        # 5a90 <_sk_callback_sse2+0x100c>
+  .byte  68,15,88,5,215,17,0,0               // addps         0x11d7(%rip),%xmm8        # 5a90 <_sk_callback_sse2+0x100f>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,216,17,0,0               // addps         0x11d8(%rip),%xmm8        # 5aa0 <_sk_callback_sse2+0x101c>
+  .byte  68,15,88,5,219,17,0,0               // addps         0x11db(%rip),%xmm8        # 5aa0 <_sk_callback_sse2+0x101f>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,220,17,0,0               // addps         0x11dc(%rip),%xmm8        # 5ab0 <_sk_callback_sse2+0x102c>
+  .byte  68,15,88,5,223,17,0,0               // addps         0x11df(%rip),%xmm8        # 5ab0 <_sk_callback_sse2+0x102f>
   .byte  68,15,17,128,128,0,0,0              // movups        %xmm8,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -31880,17 +31872,17 @@ HIDDEN _sk_bicubic_p1x_sse2
 FUNCTION(_sk_bicubic_p1x_sse2)
 _sk_bicubic_p1x_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,214,17,0,0               // movaps        0x11d6(%rip),%xmm8        # 5ac0 <_sk_callback_sse2+0x103c>
+  .byte  68,15,40,5,217,17,0,0               // movaps        0x11d9(%rip),%xmm8        # 5ac0 <_sk_callback_sse2+0x103f>
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,72,64                      // movups        0x40(%rax),%xmm9
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,21,210,17,0,0              // movaps        0x11d2(%rip),%xmm10        # 5ad0 <_sk_callback_sse2+0x104c>
+  .byte  68,15,40,21,213,17,0,0              // movaps        0x11d5(%rip),%xmm10        # 5ad0 <_sk_callback_sse2+0x104f>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,214,17,0,0              // addps         0x11d6(%rip),%xmm10        # 5ae0 <_sk_callback_sse2+0x105c>
+  .byte  68,15,88,21,217,17,0,0              // addps         0x11d9(%rip),%xmm10        # 5ae0 <_sk_callback_sse2+0x105f>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,210,17,0,0              // addps         0x11d2(%rip),%xmm10        # 5af0 <_sk_callback_sse2+0x106c>
+  .byte  68,15,88,21,213,17,0,0              // addps         0x11d5(%rip),%xmm10        # 5af0 <_sk_callback_sse2+0x106f>
   .byte  68,15,17,144,128,0,0,0              // movups        %xmm10,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -31902,11 +31894,11 @@ _sk_bicubic_p3x_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,64,64                      // movups        0x40(%rax),%xmm8
-  .byte  15,88,5,197,17,0,0                  // addps         0x11c5(%rip),%xmm0        # 5b00 <_sk_callback_sse2+0x107c>
+  .byte  15,88,5,200,17,0,0                  // addps         0x11c8(%rip),%xmm0        # 5b00 <_sk_callback_sse2+0x107f>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,197,17,0,0               // mulps         0x11c5(%rip),%xmm8        # 5b10 <_sk_callback_sse2+0x108c>
-  .byte  68,15,88,5,205,17,0,0               // addps         0x11cd(%rip),%xmm8        # 5b20 <_sk_callback_sse2+0x109c>
+  .byte  68,15,89,5,200,17,0,0               // mulps         0x11c8(%rip),%xmm8        # 5b10 <_sk_callback_sse2+0x108f>
+  .byte  68,15,88,5,208,17,0,0               // addps         0x11d0(%rip),%xmm8        # 5b20 <_sk_callback_sse2+0x109f>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,128,0,0,0              // movups        %xmm8,0x80(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31919,13 +31911,13 @@ _sk_bicubic_n3y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,187,17,0,0                 // addps         0x11bb(%rip),%xmm1        # 5b30 <_sk_callback_sse2+0x10ac>
-  .byte  68,15,40,13,195,17,0,0              // movaps        0x11c3(%rip),%xmm9        # 5b40 <_sk_callback_sse2+0x10bc>
+  .byte  15,88,13,190,17,0,0                 // addps         0x11be(%rip),%xmm1        # 5b30 <_sk_callback_sse2+0x10af>
+  .byte  68,15,40,13,198,17,0,0              // movaps        0x11c6(%rip),%xmm9        # 5b40 <_sk_callback_sse2+0x10bf>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,191,17,0,0              // mulps         0x11bf(%rip),%xmm9        # 5b50 <_sk_callback_sse2+0x10cc>
-  .byte  68,15,88,13,199,17,0,0              // addps         0x11c7(%rip),%xmm9        # 5b60 <_sk_callback_sse2+0x10dc>
+  .byte  68,15,89,13,194,17,0,0              // mulps         0x11c2(%rip),%xmm9        # 5b50 <_sk_callback_sse2+0x10cf>
+  .byte  68,15,88,13,202,17,0,0              // addps         0x11ca(%rip),%xmm9        # 5b60 <_sk_callback_sse2+0x10df>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,160,0,0,0              // movups        %xmm9,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31938,16 +31930,16 @@ _sk_bicubic_n1y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,181,17,0,0                 // addps         0x11b5(%rip),%xmm1        # 5b70 <_sk_callback_sse2+0x10ec>
-  .byte  68,15,40,13,189,17,0,0              // movaps        0x11bd(%rip),%xmm9        # 5b80 <_sk_callback_sse2+0x10fc>
+  .byte  15,88,13,184,17,0,0                 // addps         0x11b8(%rip),%xmm1        # 5b70 <_sk_callback_sse2+0x10ef>
+  .byte  68,15,40,13,192,17,0,0              // movaps        0x11c0(%rip),%xmm9        # 5b80 <_sk_callback_sse2+0x10ff>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,193,17,0,0               // movaps        0x11c1(%rip),%xmm8        # 5b90 <_sk_callback_sse2+0x110c>
+  .byte  68,15,40,5,196,17,0,0               // movaps        0x11c4(%rip),%xmm8        # 5b90 <_sk_callback_sse2+0x110f>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,197,17,0,0               // addps         0x11c5(%rip),%xmm8        # 5ba0 <_sk_callback_sse2+0x111c>
+  .byte  68,15,88,5,200,17,0,0               // addps         0x11c8(%rip),%xmm8        # 5ba0 <_sk_callback_sse2+0x111f>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,201,17,0,0               // addps         0x11c9(%rip),%xmm8        # 5bb0 <_sk_callback_sse2+0x112c>
+  .byte  68,15,88,5,204,17,0,0               // addps         0x11cc(%rip),%xmm8        # 5bb0 <_sk_callback_sse2+0x112f>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,205,17,0,0               // addps         0x11cd(%rip),%xmm8        # 5bc0 <_sk_callback_sse2+0x113c>
+  .byte  68,15,88,5,208,17,0,0               // addps         0x11d0(%rip),%xmm8        # 5bc0 <_sk_callback_sse2+0x113f>
   .byte  68,15,17,128,160,0,0,0              // movups        %xmm8,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -31957,17 +31949,17 @@ HIDDEN _sk_bicubic_p1y_sse2
 FUNCTION(_sk_bicubic_p1y_sse2)
 _sk_bicubic_p1y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,199,17,0,0               // movaps        0x11c7(%rip),%xmm8        # 5bd0 <_sk_callback_sse2+0x114c>
+  .byte  68,15,40,5,202,17,0,0               // movaps        0x11ca(%rip),%xmm8        # 5bd0 <_sk_callback_sse2+0x114f>
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,72,96                      // movups        0x60(%rax),%xmm9
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  68,15,40,21,194,17,0,0              // movaps        0x11c2(%rip),%xmm10        # 5be0 <_sk_callback_sse2+0x115c>
+  .byte  68,15,40,21,197,17,0,0              // movaps        0x11c5(%rip),%xmm10        # 5be0 <_sk_callback_sse2+0x115f>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,198,17,0,0              // addps         0x11c6(%rip),%xmm10        # 5bf0 <_sk_callback_sse2+0x116c>
+  .byte  68,15,88,21,201,17,0,0              // addps         0x11c9(%rip),%xmm10        # 5bf0 <_sk_callback_sse2+0x116f>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,194,17,0,0              // addps         0x11c2(%rip),%xmm10        # 5c00 <_sk_callback_sse2+0x117c>
+  .byte  68,15,88,21,197,17,0,0              // addps         0x11c5(%rip),%xmm10        # 5c00 <_sk_callback_sse2+0x117f>
   .byte  68,15,17,144,160,0,0,0              // movups        %xmm10,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -31979,11 +31971,11 @@ _sk_bicubic_p3y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,32                         // movups        0x20(%rax),%xmm1
   .byte  68,15,16,64,96                      // movups        0x60(%rax),%xmm8
-  .byte  15,88,13,180,17,0,0                 // addps         0x11b4(%rip),%xmm1        # 5c10 <_sk_callback_sse2+0x118c>
+  .byte  15,88,13,183,17,0,0                 // addps         0x11b7(%rip),%xmm1        # 5c10 <_sk_callback_sse2+0x118f>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,180,17,0,0               // mulps         0x11b4(%rip),%xmm8        # 5c20 <_sk_callback_sse2+0x119c>
-  .byte  68,15,88,5,188,17,0,0               // addps         0x11bc(%rip),%xmm8        # 5c30 <_sk_callback_sse2+0x11ac>
+  .byte  68,15,89,5,183,17,0,0               // mulps         0x11b7(%rip),%xmm8        # 5c20 <_sk_callback_sse2+0x119f>
+  .byte  68,15,88,5,191,17,0,0               // addps         0x11bf(%rip),%xmm8        # 5c30 <_sk_callback_sse2+0x11af>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,160,0,0,0              // movups        %xmm8,0xa0(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -32685,7 +32677,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3911a <_sk_callback_sse2+0xffffffffe9a34696>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3911a <_sk_callback_sse2+0xffffffffe9a34699>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -32780,7 +32772,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a391ea <_sk_callback_sse2+0xffffffffe9a34766>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a391ea <_sk_callback_sse2+0xffffffffe9a34769>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -32875,7 +32867,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a392ba <_sk_callback_sse2+0xffffffffe9a34836>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a392ba <_sk_callback_sse2+0xffffffffe9a34839>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -32970,7 +32962,7 @@ BALIGN16
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3938a <_sk_callback_sse2+0xffffffffe9a34906>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3938a <_sk_callback_sse2+0xffffffffe9a34909>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -33653,7 +33645,7 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e8f5 <_sk_callback_sse2+0x3d639e71>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e8f5 <_sk_callback_sse2+0x3d639e74>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -33679,7 +33671,7 @@ BALIGN16
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e935 <_sk_callback_sse2+0x3d639eb1>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e935 <_sk_callback_sse2+0x3d639eb4>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -33763,7 +33755,7 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63ea05 <_sk_callback_sse2+0x3d639f81>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63ea05 <_sk_callback_sse2+0x3d639f84>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -33789,7 +33781,7 @@ BALIGN16
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63ea45 <_sk_callback_sse2+0x3d639fc1>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63ea45 <_sk_callback_sse2+0x3d639fc4>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -33800,11 +33792,11 @@ BALIGN16
   .byte  63                                  // (bad)
   .byte  114,28                              // jb            5c3e <.literal16+0x10de>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c42 <_sk_callback_sse2+0x11be>
+  .byte  62,114,28                           // jb,pt         5c42 <_sk_callback_sse2+0x11c1>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c46 <_sk_callback_sse2+0x11c2>
+  .byte  62,114,28                           // jb,pt         5c46 <_sk_callback_sse2+0x11c5>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c4a <_sk_callback_sse2+0x11c6>
+  .byte  62,114,28                           // jb,pt         5c4a <_sk_callback_sse2+0x11c9>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
index d55a0a4..f759020 100644 (file)
@@ -106,14 +106,14 @@ _sk_seed_shader_hsw LABEL PROC
   DB  197,249,110,199                     ; vmovd         %edi,%xmm0
   DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,218,70,0,0        ; vbroadcastss  0x46da(%rip),%ymm1        # 4834 <_sk_callback_hsw+0x11c>
+  DB  196,226,125,24,13,214,70,0,0        ; vbroadcastss  0x46d6(%rip),%ymm1        # 4830 <_sk_callback_hsw+0x11c>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  197,252,88,2                        ; vaddps        (%rdx),%ymm0,%ymm0
   DB  196,226,125,24,16                   ; vbroadcastss  (%rax),%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  197,236,88,201                      ; vaddps        %ymm1,%ymm2,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,21,190,70,0,0        ; vbroadcastss  0x46be(%rip),%ymm2        # 4838 <_sk_callback_hsw+0x120>
+  DB  196,226,125,24,21,186,70,0,0        ; vbroadcastss  0x46ba(%rip),%ymm2        # 4834 <_sk_callback_hsw+0x120>
   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
@@ -132,13 +132,13 @@ _sk_dither_hsw LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  196,66,125,88,8                     ; vpbroadcastd  (%r8),%ymm9
   DB  196,65,61,239,201                   ; vpxor         %ymm9,%ymm8,%ymm9
-  DB  196,98,125,88,21,125,70,0,0         ; vpbroadcastd  0x467d(%rip),%ymm10        # 483c <_sk_callback_hsw+0x124>
+  DB  196,98,125,88,21,121,70,0,0         ; vpbroadcastd  0x4679(%rip),%ymm10        # 4838 <_sk_callback_hsw+0x124>
   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  196,98,125,88,37,98,70,0,0          ; vpbroadcastd  0x4662(%rip),%ymm12        # 4840 <_sk_callback_hsw+0x128>
-  DB  196,98,125,88,45,93,70,0,0          ; vpbroadcastd  0x465d(%rip),%ymm13        # 4844 <_sk_callback_hsw+0x12c>
+  DB  196,98,125,88,37,94,70,0,0          ; vpbroadcastd  0x465e(%rip),%ymm12        # 483c <_sk_callback_hsw+0x128>
+  DB  196,98,125,88,45,89,70,0,0          ; vpbroadcastd  0x4659(%rip),%ymm13        # 4840 <_sk_callback_hsw+0x12c>
   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
@@ -153,8 +153,8 @@ _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  196,98,125,24,13,15,70,0,0          ; vbroadcastss  0x460f(%rip),%ymm9        # 4848 <_sk_callback_hsw+0x130>
-  DB  196,98,125,24,21,10,70,0,0          ; vbroadcastss  0x460a(%rip),%ymm10        # 484c <_sk_callback_hsw+0x134>
+  DB  196,98,125,24,13,11,70,0,0          ; vbroadcastss  0x460b(%rip),%ymm9        # 4844 <_sk_callback_hsw+0x130>
+  DB  196,98,125,24,21,6,70,0,0           ; vbroadcastss  0x4606(%rip),%ymm10        # 4848 <_sk_callback_hsw+0x134>
   DB  196,66,61,184,209                   ; vfmadd231ps   %ymm9,%ymm8,%ymm10
   DB  196,98,125,24,64,8                  ; vbroadcastss  0x8(%rax),%ymm8
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
@@ -206,7 +206,7 @@ _sk_clear_hsw LABEL PROC
 PUBLIC _sk_srcatop_hsw
 _sk_srcatop_hsw LABEL PROC
   DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
-  DB  196,98,125,24,5,126,69,0,0          ; vbroadcastss  0x457e(%rip),%ymm8        # 4850 <_sk_callback_hsw+0x138>
+  DB  196,98,125,24,5,122,69,0,0          ; vbroadcastss  0x457a(%rip),%ymm8        # 484c <_sk_callback_hsw+0x138>
   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
@@ -220,7 +220,7 @@ _sk_srcatop_hsw LABEL PROC
 
 PUBLIC _sk_dstatop_hsw
 _sk_dstatop_hsw LABEL PROC
-  DB  196,98,125,24,5,81,69,0,0           ; vbroadcastss  0x4551(%rip),%ymm8        # 4854 <_sk_callback_hsw+0x13c>
+  DB  196,98,125,24,5,77,69,0,0           ; vbroadcastss  0x454d(%rip),%ymm8        # 4850 <_sk_callback_hsw+0x13c>
   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
@@ -253,7 +253,7 @@ _sk_dstin_hsw LABEL PROC
 
 PUBLIC _sk_srcout_hsw
 _sk_srcout_hsw LABEL PROC
-  DB  196,98,125,24,5,248,68,0,0          ; vbroadcastss  0x44f8(%rip),%ymm8        # 4858 <_sk_callback_hsw+0x140>
+  DB  196,98,125,24,5,244,68,0,0          ; vbroadcastss  0x44f4(%rip),%ymm8        # 4854 <_sk_callback_hsw+0x140>
   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
@@ -264,7 +264,7 @@ _sk_srcout_hsw LABEL PROC
 
 PUBLIC _sk_dstout_hsw
 _sk_dstout_hsw LABEL PROC
-  DB  196,226,125,24,5,219,68,0,0         ; vbroadcastss  0x44db(%rip),%ymm0        # 485c <_sk_callback_hsw+0x144>
+  DB  196,226,125,24,5,215,68,0,0         ; vbroadcastss  0x44d7(%rip),%ymm0        # 4858 <_sk_callback_hsw+0x144>
   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
@@ -275,7 +275,7 @@ _sk_dstout_hsw LABEL PROC
 
 PUBLIC _sk_srcover_hsw
 _sk_srcover_hsw LABEL PROC
-  DB  196,98,125,24,5,190,68,0,0          ; vbroadcastss  0x44be(%rip),%ymm8        # 4860 <_sk_callback_hsw+0x148>
+  DB  196,98,125,24,5,186,68,0,0          ; vbroadcastss  0x44ba(%rip),%ymm8        # 485c <_sk_callback_hsw+0x148>
   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
@@ -286,7 +286,7 @@ _sk_srcover_hsw LABEL PROC
 
 PUBLIC _sk_dstover_hsw
 _sk_dstover_hsw LABEL PROC
-  DB  196,98,125,24,5,157,68,0,0          ; vbroadcastss  0x449d(%rip),%ymm8        # 4864 <_sk_callback_hsw+0x14c>
+  DB  196,98,125,24,5,153,68,0,0          ; vbroadcastss  0x4499(%rip),%ymm8        # 4860 <_sk_callback_hsw+0x14c>
   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
@@ -306,7 +306,7 @@ _sk_modulate_hsw LABEL PROC
 
 PUBLIC _sk_multiply_hsw
 _sk_multiply_hsw LABEL PROC
-  DB  196,98,125,24,5,104,68,0,0          ; vbroadcastss  0x4468(%rip),%ymm8        # 4868 <_sk_callback_hsw+0x150>
+  DB  196,98,125,24,5,100,68,0,0          ; vbroadcastss  0x4464(%rip),%ymm8        # 4864 <_sk_callback_hsw+0x150>
   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
@@ -348,7 +348,7 @@ _sk_screen_hsw LABEL PROC
 
 PUBLIC _sk_xor__hsw
 _sk_xor__hsw LABEL PROC
-  DB  196,98,125,24,5,227,67,0,0          ; vbroadcastss  0x43e3(%rip),%ymm8        # 486c <_sk_callback_hsw+0x154>
+  DB  196,98,125,24,5,223,67,0,0          ; vbroadcastss  0x43df(%rip),%ymm8        # 4868 <_sk_callback_hsw+0x154>
   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
@@ -380,7 +380,7 @@ _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  196,98,125,24,5,107,67,0,0          ; vbroadcastss  0x436b(%rip),%ymm8        # 4870 <_sk_callback_hsw+0x158>
+  DB  196,98,125,24,5,103,67,0,0          ; vbroadcastss  0x4367(%rip),%ymm8        # 486c <_sk_callback_hsw+0x158>
   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
@@ -403,7 +403,7 @@ _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  196,98,125,24,5,26,67,0,0           ; vbroadcastss  0x431a(%rip),%ymm8        # 4874 <_sk_callback_hsw+0x15c>
+  DB  196,98,125,24,5,22,67,0,0           ; vbroadcastss  0x4316(%rip),%ymm8        # 4870 <_sk_callback_hsw+0x15c>
   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
@@ -429,7 +429,7 @@ _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  196,98,125,24,5,189,66,0,0          ; vbroadcastss  0x42bd(%rip),%ymm8        # 4878 <_sk_callback_hsw+0x160>
+  DB  196,98,125,24,5,185,66,0,0          ; vbroadcastss  0x42b9(%rip),%ymm8        # 4874 <_sk_callback_hsw+0x160>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -449,7 +449,7 @@ _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  196,98,125,24,5,123,66,0,0          ; vbroadcastss  0x427b(%rip),%ymm8        # 487c <_sk_callback_hsw+0x164>
+  DB  196,98,125,24,5,119,66,0,0          ; vbroadcastss  0x4277(%rip),%ymm8        # 4878 <_sk_callback_hsw+0x164>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -457,7 +457,7 @@ _sk_exclusion_hsw LABEL PROC
 
 PUBLIC _sk_colorburn_hsw
 _sk_colorburn_hsw LABEL PROC
-  DB  196,98,125,24,5,105,66,0,0          ; vbroadcastss  0x4269(%rip),%ymm8        # 4880 <_sk_callback_hsw+0x168>
+  DB  196,98,125,24,5,101,66,0,0          ; vbroadcastss  0x4265(%rip),%ymm8        # 487c <_sk_callback_hsw+0x168>
   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
@@ -513,7 +513,7 @@ _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  196,98,125,24,13,116,65,0,0         ; vbroadcastss  0x4174(%rip),%ymm9        # 4884 <_sk_callback_hsw+0x16c>
+  DB  196,98,125,24,13,112,65,0,0         ; vbroadcastss  0x4170(%rip),%ymm9        # 4880 <_sk_callback_hsw+0x16c>
   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
@@ -564,7 +564,7 @@ _sk_colordodge_hsw LABEL PROC
 
 PUBLIC _sk_hardlight_hsw
 _sk_hardlight_hsw LABEL PROC
-  DB  196,98,125,24,5,149,64,0,0          ; vbroadcastss  0x4095(%rip),%ymm8        # 4888 <_sk_callback_hsw+0x170>
+  DB  196,98,125,24,5,145,64,0,0          ; vbroadcastss  0x4091(%rip),%ymm8        # 4884 <_sk_callback_hsw+0x170>
   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
@@ -613,7 +613,7 @@ _sk_hardlight_hsw LABEL PROC
 
 PUBLIC _sk_overlay_hsw
 _sk_overlay_hsw LABEL PROC
-  DB  196,98,125,24,5,205,63,0,0          ; vbroadcastss  0x3fcd(%rip),%ymm8        # 488c <_sk_callback_hsw+0x174>
+  DB  196,98,125,24,5,201,63,0,0          ; vbroadcastss  0x3fc9(%rip),%ymm8        # 4888 <_sk_callback_hsw+0x174>
   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
@@ -673,10 +673,10 @@ _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  196,98,125,24,29,212,62,0,0         ; vbroadcastss  0x3ed4(%rip),%ymm11        # 4894 <_sk_callback_hsw+0x17c>
+  DB  196,98,125,24,29,208,62,0,0         ; vbroadcastss  0x3ed0(%rip),%ymm11        # 4890 <_sk_callback_hsw+0x17c>
   DB  196,65,20,88,227                    ; vaddps        %ymm11,%ymm13,%ymm12
   DB  196,65,28,89,192                    ; vmulps        %ymm8,%ymm12,%ymm8
-  DB  196,98,125,24,37,197,62,0,0         ; vbroadcastss  0x3ec5(%rip),%ymm12        # 4898 <_sk_callback_hsw+0x180>
+  DB  196,98,125,24,37,193,62,0,0         ; vbroadcastss  0x3ec1(%rip),%ymm12        # 4894 <_sk_callback_hsw+0x180>
   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
@@ -686,7 +686,7 @@ _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  196,98,125,24,5,136,62,0,0          ; vbroadcastss  0x3e88(%rip),%ymm8        # 4890 <_sk_callback_hsw+0x178>
+  DB  196,98,125,24,5,132,62,0,0          ; vbroadcastss  0x3e84(%rip),%ymm8        # 488c <_sk_callback_hsw+0x178>
   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
@@ -799,11 +799,11 @@ _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  196,98,125,24,53,135,60,0,0         ; vbroadcastss  0x3c87(%rip),%ymm14        # 489c <_sk_callback_hsw+0x184>
-  DB  196,98,125,24,61,130,60,0,0         ; vbroadcastss  0x3c82(%rip),%ymm15        # 48a0 <_sk_callback_hsw+0x188>
+  DB  196,98,125,24,53,131,60,0,0         ; vbroadcastss  0x3c83(%rip),%ymm14        # 4898 <_sk_callback_hsw+0x184>
+  DB  196,98,125,24,61,126,60,0,0         ; vbroadcastss  0x3c7e(%rip),%ymm15        # 489c <_sk_callback_hsw+0x188>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,66,93,184,238                   ; vfmadd231ps   %ymm14,%ymm4,%ymm13
-  DB  196,226,125,24,5,115,60,0,0         ; vbroadcastss  0x3c73(%rip),%ymm0        # 48a4 <_sk_callback_hsw+0x18c>
+  DB  196,226,125,24,5,111,60,0,0         ; vbroadcastss  0x3c6f(%rip),%ymm0        # 48a0 <_sk_callback_hsw+0x18c>
   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
@@ -858,7 +858,7 @@ _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  196,226,125,24,13,96,59,0,0         ; vbroadcastss  0x3b60(%rip),%ymm1        # 48a8 <_sk_callback_hsw+0x190>
+  DB  196,226,125,24,13,92,59,0,0         ; vbroadcastss  0x3b5c(%rip),%ymm1        # 48a4 <_sk_callback_hsw+0x190>
   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
@@ -912,11 +912,11 @@ _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  196,98,125,24,53,113,58,0,0         ; vbroadcastss  0x3a71(%rip),%ymm14        # 48ac <_sk_callback_hsw+0x194>
-  DB  196,98,125,24,61,108,58,0,0         ; vbroadcastss  0x3a6c(%rip),%ymm15        # 48b0 <_sk_callback_hsw+0x198>
+  DB  196,98,125,24,53,109,58,0,0         ; vbroadcastss  0x3a6d(%rip),%ymm14        # 48a8 <_sk_callback_hsw+0x194>
+  DB  196,98,125,24,61,104,58,0,0         ; vbroadcastss  0x3a68(%rip),%ymm15        # 48ac <_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  196,226,125,24,5,93,58,0,0          ; vbroadcastss  0x3a5d(%rip),%ymm0        # 48b4 <_sk_callback_hsw+0x19c>
+  DB  196,226,125,24,5,89,58,0,0          ; vbroadcastss  0x3a59(%rip),%ymm0        # 48b0 <_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
@@ -971,7 +971,7 @@ _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  196,226,125,24,13,74,57,0,0         ; vbroadcastss  0x394a(%rip),%ymm1        # 48b8 <_sk_callback_hsw+0x1a0>
+  DB  196,226,125,24,13,70,57,0,0         ; vbroadcastss  0x3946(%rip),%ymm1        # 48b4 <_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
@@ -999,11 +999,11 @@ _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  196,98,125,24,45,221,56,0,0         ; vbroadcastss  0x38dd(%rip),%ymm13        # 48bc <_sk_callback_hsw+0x1a4>
-  DB  196,98,125,24,53,216,56,0,0         ; vbroadcastss  0x38d8(%rip),%ymm14        # 48c0 <_sk_callback_hsw+0x1a8>
+  DB  196,98,125,24,45,217,56,0,0         ; vbroadcastss  0x38d9(%rip),%ymm13        # 48b8 <_sk_callback_hsw+0x1a4>
+  DB  196,98,125,24,53,212,56,0,0         ; vbroadcastss  0x38d4(%rip),%ymm14        # 48bc <_sk_callback_hsw+0x1a8>
   DB  196,65,84,89,230                    ; vmulps        %ymm14,%ymm5,%ymm12
   DB  196,66,93,184,229                   ; vfmadd231ps   %ymm13,%ymm4,%ymm12
-  DB  196,98,125,24,61,201,56,0,0         ; vbroadcastss  0x38c9(%rip),%ymm15        # 48c4 <_sk_callback_hsw+0x1ac>
+  DB  196,98,125,24,61,197,56,0,0         ; vbroadcastss  0x38c5(%rip),%ymm15        # 48c0 <_sk_callback_hsw+0x1ac>
   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
@@ -1059,7 +1059,7 @@ _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  196,226,125,24,5,171,55,0,0         ; vbroadcastss  0x37ab(%rip),%ymm0        # 48c8 <_sk_callback_hsw+0x1b0>
+  DB  196,226,125,24,5,167,55,0,0         ; vbroadcastss  0x37a7(%rip),%ymm0        # 48c4 <_sk_callback_hsw+0x1b0>
   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
@@ -1087,11 +1087,11 @@ _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  196,98,125,24,45,62,55,0,0          ; vbroadcastss  0x373e(%rip),%ymm13        # 48cc <_sk_callback_hsw+0x1b4>
-  DB  196,98,125,24,53,57,55,0,0          ; vbroadcastss  0x3739(%rip),%ymm14        # 48d0 <_sk_callback_hsw+0x1b8>
+  DB  196,98,125,24,45,58,55,0,0          ; vbroadcastss  0x373a(%rip),%ymm13        # 48c8 <_sk_callback_hsw+0x1b4>
+  DB  196,98,125,24,53,53,55,0,0          ; vbroadcastss  0x3735(%rip),%ymm14        # 48cc <_sk_callback_hsw+0x1b8>
   DB  196,65,116,89,230                   ; vmulps        %ymm14,%ymm1,%ymm12
   DB  196,66,109,184,229                  ; vfmadd231ps   %ymm13,%ymm2,%ymm12
-  DB  196,98,125,24,61,42,55,0,0          ; vbroadcastss  0x372a(%rip),%ymm15        # 48d4 <_sk_callback_hsw+0x1bc>
+  DB  196,98,125,24,61,38,55,0,0          ; vbroadcastss  0x3726(%rip),%ymm15        # 48d0 <_sk_callback_hsw+0x1bc>
   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
@@ -1147,7 +1147,7 @@ _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  196,226,125,24,5,12,54,0,0          ; vbroadcastss  0x360c(%rip),%ymm0        # 48d8 <_sk_callback_hsw+0x1c0>
+  DB  196,226,125,24,5,8,54,0,0           ; vbroadcastss  0x3608(%rip),%ymm0        # 48d4 <_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
@@ -1177,7 +1177,7 @@ _sk_clamp_0_hsw LABEL PROC
 
 PUBLIC _sk_clamp_1_hsw
 _sk_clamp_1_hsw LABEL PROC
-  DB  196,98,125,24,5,165,53,0,0          ; vbroadcastss  0x35a5(%rip),%ymm8        # 48dc <_sk_callback_hsw+0x1c4>
+  DB  196,98,125,24,5,161,53,0,0          ; vbroadcastss  0x35a1(%rip),%ymm8        # 48d8 <_sk_callback_hsw+0x1c4>
   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
@@ -1187,7 +1187,7 @@ _sk_clamp_1_hsw LABEL PROC
 
 PUBLIC _sk_clamp_a_hsw
 _sk_clamp_a_hsw LABEL PROC
-  DB  196,98,125,24,5,136,53,0,0          ; vbroadcastss  0x3588(%rip),%ymm8        # 48e0 <_sk_callback_hsw+0x1c8>
+  DB  196,98,125,24,5,132,53,0,0          ; vbroadcastss  0x3584(%rip),%ymm8        # 48dc <_sk_callback_hsw+0x1c8>
   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
@@ -1259,7 +1259,7 @@ 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  196,98,125,24,21,208,52,0,0         ; vbroadcastss  0x34d0(%rip),%ymm10        # 48e4 <_sk_callback_hsw+0x1cc>
+  DB  196,98,125,24,21,204,52,0,0         ; vbroadcastss  0x34cc(%rip),%ymm10        # 48e0 <_sk_callback_hsw+0x1cc>
   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
@@ -1270,16 +1270,16 @@ _sk_unpremul_hsw LABEL PROC
 
 PUBLIC _sk_from_srgb_hsw
 _sk_from_srgb_hsw LABEL PROC
-  DB  196,98,125,24,5,177,52,0,0          ; vbroadcastss  0x34b1(%rip),%ymm8        # 48e8 <_sk_callback_hsw+0x1d0>
+  DB  196,98,125,24,5,173,52,0,0          ; vbroadcastss  0x34ad(%rip),%ymm8        # 48e4 <_sk_callback_hsw+0x1d0>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
-  DB  196,98,125,24,29,163,52,0,0         ; vbroadcastss  0x34a3(%rip),%ymm11        # 48ec <_sk_callback_hsw+0x1d4>
-  DB  196,98,125,24,37,158,52,0,0         ; vbroadcastss  0x349e(%rip),%ymm12        # 48f0 <_sk_callback_hsw+0x1d8>
+  DB  196,98,125,24,29,159,52,0,0         ; vbroadcastss  0x349f(%rip),%ymm11        # 48e8 <_sk_callback_hsw+0x1d4>
+  DB  196,98,125,24,37,154,52,0,0         ; vbroadcastss  0x349a(%rip),%ymm12        # 48ec <_sk_callback_hsw+0x1d8>
   DB  196,65,124,40,236                   ; vmovaps       %ymm12,%ymm13
   DB  196,66,125,168,235                  ; vfmadd213ps   %ymm11,%ymm0,%ymm13
-  DB  196,98,125,24,53,143,52,0,0         ; vbroadcastss  0x348f(%rip),%ymm14        # 48f4 <_sk_callback_hsw+0x1dc>
+  DB  196,98,125,24,53,139,52,0,0         ; vbroadcastss  0x348b(%rip),%ymm14        # 48f0 <_sk_callback_hsw+0x1dc>
   DB  196,66,45,168,238                   ; vfmadd213ps   %ymm14,%ymm10,%ymm13
-  DB  196,98,125,24,21,133,52,0,0         ; vbroadcastss  0x3485(%rip),%ymm10        # 48f8 <_sk_callback_hsw+0x1e0>
+  DB  196,98,125,24,21,129,52,0,0         ; vbroadcastss  0x3481(%rip),%ymm10        # 48f4 <_sk_callback_hsw+0x1e0>
   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
@@ -1303,16 +1303,16 @@ _sk_to_srgb_hsw LABEL PROC
   DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
   DB  196,65,124,83,200                   ; vrcpps        %ymm8,%ymm9
   DB  196,65,124,82,208                   ; vrsqrtps      %ymm8,%ymm10
-  DB  196,98,125,24,5,31,52,0,0           ; vbroadcastss  0x341f(%rip),%ymm8        # 48fc <_sk_callback_hsw+0x1e4>
+  DB  196,98,125,24,5,27,52,0,0           ; vbroadcastss  0x341b(%rip),%ymm8        # 48f8 <_sk_callback_hsw+0x1e4>
   DB  196,65,124,89,216                   ; vmulps        %ymm8,%ymm0,%ymm11
-  DB  196,98,125,24,37,21,52,0,0          ; vbroadcastss  0x3415(%rip),%ymm12        # 4900 <_sk_callback_hsw+0x1e8>
-  DB  196,98,125,24,45,16,52,0,0          ; vbroadcastss  0x3410(%rip),%ymm13        # 4904 <_sk_callback_hsw+0x1ec>
+  DB  196,98,125,24,37,17,52,0,0          ; vbroadcastss  0x3411(%rip),%ymm12        # 48fc <_sk_callback_hsw+0x1e8>
+  DB  196,98,125,24,45,12,52,0,0          ; vbroadcastss  0x340c(%rip),%ymm13        # 4900 <_sk_callback_hsw+0x1ec>
   DB  196,66,21,168,204                   ; vfmadd213ps   %ymm12,%ymm13,%ymm9
-  DB  196,98,125,24,53,6,52,0,0           ; vbroadcastss  0x3406(%rip),%ymm14        # 4908 <_sk_callback_hsw+0x1f0>
+  DB  196,98,125,24,53,2,52,0,0           ; vbroadcastss  0x3402(%rip),%ymm14        # 4904 <_sk_callback_hsw+0x1f0>
   DB  196,66,13,184,202                   ; vfmadd231ps   %ymm10,%ymm14,%ymm9
-  DB  196,98,125,24,21,252,51,0,0         ; vbroadcastss  0x33fc(%rip),%ymm10        # 490c <_sk_callback_hsw+0x1f4>
+  DB  196,98,125,24,21,248,51,0,0         ; vbroadcastss  0x33f8(%rip),%ymm10        # 4908 <_sk_callback_hsw+0x1f4>
   DB  196,65,44,93,201                    ; vminps        %ymm9,%ymm10,%ymm9
-  DB  196,98,125,24,61,242,51,0,0         ; vbroadcastss  0x33f2(%rip),%ymm15        # 4910 <_sk_callback_hsw+0x1f8>
+  DB  196,98,125,24,61,238,51,0,0         ; vbroadcastss  0x33ee(%rip),%ymm15        # 490c <_sk_callback_hsw+0x1f8>
   DB  196,193,124,194,199,1               ; vcmpltps      %ymm15,%ymm0,%ymm0
   DB  196,195,53,74,195,0                 ; vblendvps     %ymm0,%ymm11,%ymm9,%ymm0
   DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
@@ -1343,26 +1343,26 @@ _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  196,98,125,24,29,108,51,0,0         ; vbroadcastss  0x336c(%rip),%ymm11        # 4914 <_sk_callback_hsw+0x1fc>
+  DB  196,98,125,24,29,104,51,0,0         ; vbroadcastss  0x3368(%rip),%ymm11        # 4910 <_sk_callback_hsw+0x1fc>
   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  196,98,125,24,53,89,51,0,0          ; vbroadcastss  0x3359(%rip),%ymm14        # 4918 <_sk_callback_hsw+0x200>
+  DB  196,98,125,24,53,85,51,0,0          ; vbroadcastss  0x3355(%rip),%ymm14        # 4914 <_sk_callback_hsw+0x200>
   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  196,98,125,24,53,64,51,0,0          ; vbroadcastss  0x3340(%rip),%ymm14        # 4920 <_sk_callback_hsw+0x208>
+  DB  196,98,125,24,53,60,51,0,0          ; vbroadcastss  0x333c(%rip),%ymm14        # 491c <_sk_callback_hsw+0x208>
   DB  196,66,37,168,238                   ; vfmadd213ps   %ymm14,%ymm11,%ymm13
-  DB  196,98,125,24,53,46,51,0,0          ; vbroadcastss  0x332e(%rip),%ymm14        # 491c <_sk_callback_hsw+0x204>
+  DB  196,98,125,24,53,42,51,0,0          ; vbroadcastss  0x332a(%rip),%ymm14        # 4918 <_sk_callback_hsw+0x204>
   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  196,98,125,24,29,17,51,0,0          ; vbroadcastss  0x3311(%rip),%ymm11        # 4928 <_sk_callback_hsw+0x210>
+  DB  196,98,125,24,29,13,51,0,0          ; vbroadcastss  0x330d(%rip),%ymm11        # 4924 <_sk_callback_hsw+0x210>
   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
@@ -1372,7 +1372,7 @@ _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  196,98,125,24,5,212,50,0,0          ; vbroadcastss  0x32d4(%rip),%ymm8        # 4924 <_sk_callback_hsw+0x20c>
+  DB  196,98,125,24,5,208,50,0,0          ; vbroadcastss  0x32d0(%rip),%ymm8        # 4920 <_sk_callback_hsw+0x20c>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -1387,30 +1387,30 @@ _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  196,98,125,24,5,155,50,0,0          ; vbroadcastss  0x329b(%rip),%ymm8        # 492c <_sk_callback_hsw+0x214>
+  DB  196,98,125,24,5,151,50,0,0          ; vbroadcastss  0x3297(%rip),%ymm8        # 4928 <_sk_callback_hsw+0x214>
   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  196,98,125,24,13,126,50,0,0         ; vbroadcastss  0x327e(%rip),%ymm9        # 4930 <_sk_callback_hsw+0x218>
+  DB  196,98,125,24,13,122,50,0,0         ; vbroadcastss  0x327a(%rip),%ymm9        # 492c <_sk_callback_hsw+0x218>
   DB  196,66,109,170,202                  ; vfmsub213ps   %ymm10,%ymm2,%ymm9
-  DB  196,98,125,24,29,116,50,0,0         ; vbroadcastss  0x3274(%rip),%ymm11        # 4934 <_sk_callback_hsw+0x21c>
+  DB  196,98,125,24,29,112,50,0,0         ; vbroadcastss  0x3270(%rip),%ymm11        # 4930 <_sk_callback_hsw+0x21c>
   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  196,98,125,24,45,94,50,0,0          ; vbroadcastss  0x325e(%rip),%ymm13        # 493c <_sk_callback_hsw+0x224>
+  DB  196,98,125,24,45,90,50,0,0          ; vbroadcastss  0x325a(%rip),%ymm13        # 4938 <_sk_callback_hsw+0x224>
   DB  196,193,4,89,197                    ; vmulps        %ymm13,%ymm15,%ymm0
-  DB  196,98,125,24,53,84,50,0,0          ; vbroadcastss  0x3254(%rip),%ymm14        # 4940 <_sk_callback_hsw+0x228>
+  DB  196,98,125,24,53,80,50,0,0          ; vbroadcastss  0x3250(%rip),%ymm14        # 493c <_sk_callback_hsw+0x228>
   DB  197,12,92,224                       ; vsubps        %ymm0,%ymm14,%ymm12
   DB  196,66,37,168,225                   ; vfmadd213ps   %ymm9,%ymm11,%ymm12
-  DB  196,226,125,24,29,58,50,0,0         ; vbroadcastss  0x323a(%rip),%ymm3        # 4938 <_sk_callback_hsw+0x220>
+  DB  196,226,125,24,29,54,50,0,0         ; vbroadcastss  0x3236(%rip),%ymm3        # 4934 <_sk_callback_hsw+0x220>
   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  196,98,125,24,37,37,50,0,0          ; vbroadcastss  0x3225(%rip),%ymm12        # 4944 <_sk_callback_hsw+0x22c>
+  DB  196,98,125,24,37,33,50,0,0          ; vbroadcastss  0x3221(%rip),%ymm12        # 4940 <_sk_callback_hsw+0x22c>
   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
@@ -1426,7 +1426,7 @@ _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  196,226,125,24,5,209,49,0,0         ; vbroadcastss  0x31d1(%rip),%ymm0        # 4948 <_sk_callback_hsw+0x230>
+  DB  196,226,125,24,5,205,49,0,0         ; vbroadcastss  0x31cd(%rip),%ymm0        # 4944 <_sk_callback_hsw+0x230>
   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
@@ -1476,7 +1476,7 @@ _sk_scale_u8_hsw LABEL PROC
   DB  197,122,126,0                       ; vmovq         (%rax),%xmm8
   DB  196,66,125,49,192                   ; vpmovzxbd     %xmm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,11,49,0,0          ; vbroadcastss  0x310b(%rip),%ymm9        # 494c <_sk_callback_hsw+0x234>
+  DB  196,98,125,24,13,7,49,0,0           ; vbroadcastss  0x3107(%rip),%ymm9        # 4948 <_sk_callback_hsw+0x234>
   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
@@ -1524,7 +1524,7 @@ _sk_lerp_u8_hsw LABEL PROC
   DB  197,122,126,0                       ; vmovq         (%rax),%xmm8
   DB  196,66,125,49,192                   ; vpmovzxbd     %xmm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,120,48,0,0         ; vbroadcastss  0x3078(%rip),%ymm9        # 4950 <_sk_callback_hsw+0x238>
+  DB  196,98,125,24,13,116,48,0,0         ; vbroadcastss  0x3074(%rip),%ymm9        # 494c <_sk_callback_hsw+0x238>
   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
@@ -1558,20 +1558,20 @@ _sk_lerp_565_hsw LABEL PROC
   DB  15,133,169,0,0,0                    ; jne           19e4 <_sk_lerp_565_hsw+0xb7>
   DB  196,65,122,111,4,122                ; vmovdqu       (%r10,%rdi,2),%xmm8
   DB  196,66,125,51,192                   ; vpmovzxwd     %xmm8,%ymm8
-  DB  196,98,125,88,13,5,48,0,0           ; vpbroadcastd  0x3005(%rip),%ymm9        # 4954 <_sk_callback_hsw+0x23c>
+  DB  196,98,125,88,13,1,48,0,0           ; vpbroadcastd  0x3001(%rip),%ymm9        # 4950 <_sk_callback_hsw+0x23c>
   DB  196,65,61,219,201                   ; vpand         %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,246,47,0,0         ; vbroadcastss  0x2ff6(%rip),%ymm10        # 4958 <_sk_callback_hsw+0x240>
+  DB  196,98,125,24,21,242,47,0,0         ; vbroadcastss  0x2ff2(%rip),%ymm10        # 4954 <_sk_callback_hsw+0x240>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,88,21,236,47,0,0         ; vpbroadcastd  0x2fec(%rip),%ymm10        # 495c <_sk_callback_hsw+0x244>
+  DB  196,98,125,88,21,232,47,0,0         ; vpbroadcastd  0x2fe8(%rip),%ymm10        # 4958 <_sk_callback_hsw+0x244>
   DB  196,65,61,219,210                   ; vpand         %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,221,47,0,0         ; vbroadcastss  0x2fdd(%rip),%ymm11        # 4960 <_sk_callback_hsw+0x248>
+  DB  196,98,125,24,29,217,47,0,0         ; vbroadcastss  0x2fd9(%rip),%ymm11        # 495c <_sk_callback_hsw+0x248>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,88,29,211,47,0,0         ; vpbroadcastd  0x2fd3(%rip),%ymm11        # 4964 <_sk_callback_hsw+0x24c>
+  DB  196,98,125,88,29,207,47,0,0         ; vpbroadcastd  0x2fcf(%rip),%ymm11        # 4960 <_sk_callback_hsw+0x24c>
   DB  196,65,61,219,195                   ; vpand         %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,196,47,0,0         ; vbroadcastss  0x2fc4(%rip),%ymm11        # 4968 <_sk_callback_hsw+0x250>
+  DB  196,98,125,24,29,192,47,0,0         ; vbroadcastss  0x2fc0(%rip),%ymm11        # 4964 <_sk_callback_hsw+0x250>
   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
@@ -1641,21 +1641,21 @@ _sk_load_tables_hsw LABEL PROC
   DB  77,133,192                          ; test          %r8,%r8
   DB  117,105                             ; jne           1aee <_sk_load_tables_hsw+0x7e>
   DB  196,193,126,111,25                  ; vmovdqu       (%r9),%ymm3
-  DB  197,229,219,13,142,49,0,0           ; vpand         0x318e(%rip),%ymm3,%ymm1        # 4c20 <_sk_callback_hsw+0x508>
+  DB  197,229,219,13,142,49,0,0           ; vpand         0x318e(%rip),%ymm3,%ymm1        # 4c20 <_sk_callback_hsw+0x50c>
   DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
   DB  72,139,72,8                         ; mov           0x8(%rax),%rcx
   DB  76,139,72,16                        ; mov           0x10(%rax),%r9
   DB  197,237,118,210                     ; vpcmpeqd      %ymm2,%ymm2,%ymm2
   DB  196,226,109,146,4,137               ; vgatherdps    %ymm2,(%rcx,%ymm1,4),%ymm0
-  DB  196,226,101,0,21,142,49,0,0         ; vpshufb       0x318e(%rip),%ymm3,%ymm2        # 4c40 <_sk_callback_hsw+0x528>
+  DB  196,226,101,0,21,142,49,0,0         ; vpshufb       0x318e(%rip),%ymm3,%ymm2        # 4c40 <_sk_callback_hsw+0x52c>
   DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
   DB  196,194,53,146,12,145               ; vgatherdps    %ymm9,(%r9,%ymm2,4),%ymm1
   DB  72,139,64,24                        ; mov           0x18(%rax),%rax
-  DB  196,98,101,0,13,150,49,0,0          ; vpshufb       0x3196(%rip),%ymm3,%ymm9        # 4c60 <_sk_callback_hsw+0x548>
+  DB  196,98,101,0,13,150,49,0,0          ; vpshufb       0x3196(%rip),%ymm3,%ymm9        # 4c60 <_sk_callback_hsw+0x54c>
   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  196,98,125,24,5,138,46,0,0          ; vbroadcastss  0x2e8a(%rip),%ymm8        # 496c <_sk_callback_hsw+0x254>
+  DB  196,98,125,24,5,134,46,0,0          ; vbroadcastss  0x2e86(%rip),%ymm8        # 4968 <_sk_callback_hsw+0x254>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,137,193                          ; mov           %r8,%rcx
@@ -1692,7 +1692,7 @@ _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  197,121,111,21,34,50,0,0            ; vmovdqa       0x3222(%rip),%xmm10        # 4da0 <_sk_callback_hsw+0x688>
+  DB  197,121,111,21,34,50,0,0            ; vmovdqa       0x3222(%rip),%xmm10        # 4da0 <_sk_callback_hsw+0x68c>
   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
@@ -1714,7 +1714,7 @@ _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  196,98,125,24,5,131,45,0,0          ; vbroadcastss  0x2d83(%rip),%ymm8        # 4970 <_sk_callback_hsw+0x258>
+  DB  196,98,125,24,5,127,45,0,0          ; vbroadcastss  0x2d7f(%rip),%ymm8        # 496c <_sk_callback_hsw+0x258>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -1772,7 +1772,7 @@ _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  197,121,111,13,194,48,0,0           ; vmovdqa       0x30c2(%rip),%xmm9        # 4db0 <_sk_callback_hsw+0x698>
+  DB  197,121,111,13,194,48,0,0           ; vmovdqa       0x30c2(%rip),%xmm9        # 4db0 <_sk_callback_hsw+0x69c>
   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
@@ -1789,7 +1789,7 @@ _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  196,226,125,24,29,49,44,0,0         ; vbroadcastss  0x2c31(%rip),%ymm3        # 4974 <_sk_callback_hsw+0x25c>
+  DB  196,226,125,24,29,45,44,0,0         ; vbroadcastss  0x2c2d(%rip),%ymm3        # 4970 <_sk_callback_hsw+0x25c>
   DB  255,224                             ; jmpq          *%rax
   DB  196,129,121,110,4,72                ; vmovd         (%r8,%r9,2),%xmm0
   DB  196,129,121,196,68,72,4,2           ; vpinsrw       $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
@@ -1834,7 +1834,7 @@ _sk_byte_tables_hsw LABEL PROC
   DB  65,84                               ; push          %r12
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,111,43,0,0          ; vbroadcastss  0x2b6f(%rip),%ymm8        # 4978 <_sk_callback_hsw+0x260>
+  DB  196,98,125,24,5,107,43,0,0          ; vbroadcastss  0x2b6b(%rip),%ymm8        # 4974 <_sk_callback_hsw+0x260>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
   DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
@@ -1871,7 +1871,7 @@ _sk_byte_tables_hsw LABEL PROC
   DB  196,227,121,32,197,7                ; vpinsrb       $0x7,%ebp,%xmm0,%xmm0
   DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,192,42,0,0         ; vbroadcastss  0x2ac0(%rip),%ymm9        # 497c <_sk_callback_hsw+0x264>
+  DB  196,98,125,24,13,188,42,0,0         ; vbroadcastss  0x2abc(%rip),%ymm9        # 4978 <_sk_callback_hsw+0x264>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
@@ -2030,7 +2030,7 @@ _sk_byte_tables_rgb_hsw LABEL PROC
   DB  196,227,121,32,197,7                ; vpinsrb       $0x7,%ebp,%xmm0,%xmm0
   DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,249,39,0,0         ; vbroadcastss  0x27f9(%rip),%ymm9        # 4980 <_sk_callback_hsw+0x268>
+  DB  196,98,125,24,13,245,39,0,0         ; vbroadcastss  0x27f5(%rip),%ymm9        # 497c <_sk_callback_hsw+0x268>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
   DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
@@ -2183,33 +2183,33 @@ _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  196,98,125,24,37,172,37,0,0         ; vbroadcastss  0x25ac(%rip),%ymm12        # 4984 <_sk_callback_hsw+0x26c>
-  DB  196,98,125,24,45,167,37,0,0         ; vbroadcastss  0x25a7(%rip),%ymm13        # 4988 <_sk_callback_hsw+0x270>
+  DB  196,98,125,24,37,168,37,0,0         ; vbroadcastss  0x25a8(%rip),%ymm12        # 4980 <_sk_callback_hsw+0x26c>
+  DB  196,98,125,24,45,163,37,0,0         ; vbroadcastss  0x25a3(%rip),%ymm13        # 4984 <_sk_callback_hsw+0x270>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,157,37,0,0         ; vbroadcastss  0x259d(%rip),%ymm13        # 498c <_sk_callback_hsw+0x274>
+  DB  196,98,125,24,45,153,37,0,0         ; vbroadcastss  0x2599(%rip),%ymm13        # 4988 <_sk_callback_hsw+0x274>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,147,37,0,0         ; vbroadcastss  0x2593(%rip),%ymm13        # 4990 <_sk_callback_hsw+0x278>
+  DB  196,98,125,24,45,143,37,0,0         ; vbroadcastss  0x258f(%rip),%ymm13        # 498c <_sk_callback_hsw+0x278>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,137,37,0,0         ; vbroadcastss  0x2589(%rip),%ymm11        # 4994 <_sk_callback_hsw+0x27c>
+  DB  196,98,125,24,29,133,37,0,0         ; vbroadcastss  0x2585(%rip),%ymm11        # 4990 <_sk_callback_hsw+0x27c>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,127,37,0,0         ; vbroadcastss  0x257f(%rip),%ymm12        # 4998 <_sk_callback_hsw+0x280>
+  DB  196,98,125,24,37,123,37,0,0         ; vbroadcastss  0x257b(%rip),%ymm12        # 4994 <_sk_callback_hsw+0x280>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,117,37,0,0         ; vbroadcastss  0x2575(%rip),%ymm12        # 499c <_sk_callback_hsw+0x284>
+  DB  196,98,125,24,37,113,37,0,0         ; vbroadcastss  0x2571(%rip),%ymm12        # 4998 <_sk_callback_hsw+0x284>
   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  196,98,125,24,29,86,37,0,0          ; vbroadcastss  0x2556(%rip),%ymm11        # 49a0 <_sk_callback_hsw+0x288>
+  DB  196,98,125,24,29,82,37,0,0          ; vbroadcastss  0x2552(%rip),%ymm11        # 499c <_sk_callback_hsw+0x288>
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,76,37,0,0          ; vbroadcastss  0x254c(%rip),%ymm11        # 49a4 <_sk_callback_hsw+0x28c>
+  DB  196,98,125,24,29,72,37,0,0          ; vbroadcastss  0x2548(%rip),%ymm11        # 49a0 <_sk_callback_hsw+0x28c>
   DB  196,98,45,172,216                   ; vfnmadd213ps  %ymm0,%ymm10,%ymm11
-  DB  196,226,125,24,5,66,37,0,0          ; vbroadcastss  0x2542(%rip),%ymm0        # 49a8 <_sk_callback_hsw+0x290>
+  DB  196,226,125,24,5,62,37,0,0          ; vbroadcastss  0x253e(%rip),%ymm0        # 49a4 <_sk_callback_hsw+0x290>
   DB  196,193,124,92,194                  ; vsubps        %ymm10,%ymm0,%ymm0
-  DB  196,98,125,24,21,56,37,0,0          ; vbroadcastss  0x2538(%rip),%ymm10        # 49ac <_sk_callback_hsw+0x294>
+  DB  196,98,125,24,21,52,37,0,0          ; vbroadcastss  0x2534(%rip),%ymm10        # 49a8 <_sk_callback_hsw+0x294>
   DB  197,172,94,192                      ; vdivps        %ymm0,%ymm10,%ymm0
   DB  197,164,88,192                      ; vaddps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,21,43,37,0,0          ; vbroadcastss  0x252b(%rip),%ymm10        # 49b0 <_sk_callback_hsw+0x298>
+  DB  196,98,125,24,21,39,37,0,0          ; vbroadcastss  0x2527(%rip),%ymm10        # 49ac <_sk_callback_hsw+0x298>
   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
@@ -2217,7 +2217,7 @@ _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  196,98,125,24,5,2,37,0,0            ; vbroadcastss  0x2502(%rip),%ymm8        # 49b4 <_sk_callback_hsw+0x29c>
+  DB  196,98,125,24,5,254,36,0,0          ; vbroadcastss  0x24fe(%rip),%ymm8        # 49b0 <_sk_callback_hsw+0x29c>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2235,33 +2235,33 @@ _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  196,98,125,24,37,186,36,0,0         ; vbroadcastss  0x24ba(%rip),%ymm12        # 49b8 <_sk_callback_hsw+0x2a0>
-  DB  196,98,125,24,45,181,36,0,0         ; vbroadcastss  0x24b5(%rip),%ymm13        # 49bc <_sk_callback_hsw+0x2a4>
+  DB  196,98,125,24,37,182,36,0,0         ; vbroadcastss  0x24b6(%rip),%ymm12        # 49b4 <_sk_callback_hsw+0x2a0>
+  DB  196,98,125,24,45,177,36,0,0         ; vbroadcastss  0x24b1(%rip),%ymm13        # 49b8 <_sk_callback_hsw+0x2a4>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,171,36,0,0         ; vbroadcastss  0x24ab(%rip),%ymm13        # 49c0 <_sk_callback_hsw+0x2a8>
+  DB  196,98,125,24,45,167,36,0,0         ; vbroadcastss  0x24a7(%rip),%ymm13        # 49bc <_sk_callback_hsw+0x2a8>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,161,36,0,0         ; vbroadcastss  0x24a1(%rip),%ymm13        # 49c4 <_sk_callback_hsw+0x2ac>
+  DB  196,98,125,24,45,157,36,0,0         ; vbroadcastss  0x249d(%rip),%ymm13        # 49c0 <_sk_callback_hsw+0x2ac>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,151,36,0,0         ; vbroadcastss  0x2497(%rip),%ymm11        # 49c8 <_sk_callback_hsw+0x2b0>
+  DB  196,98,125,24,29,147,36,0,0         ; vbroadcastss  0x2493(%rip),%ymm11        # 49c4 <_sk_callback_hsw+0x2b0>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,141,36,0,0         ; vbroadcastss  0x248d(%rip),%ymm12        # 49cc <_sk_callback_hsw+0x2b4>
+  DB  196,98,125,24,37,137,36,0,0         ; vbroadcastss  0x2489(%rip),%ymm12        # 49c8 <_sk_callback_hsw+0x2b4>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,131,36,0,0         ; vbroadcastss  0x2483(%rip),%ymm12        # 49d0 <_sk_callback_hsw+0x2b8>
+  DB  196,98,125,24,37,127,36,0,0         ; vbroadcastss  0x247f(%rip),%ymm12        # 49cc <_sk_callback_hsw+0x2b8>
   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  196,98,125,24,29,100,36,0,0         ; vbroadcastss  0x2464(%rip),%ymm11        # 49d4 <_sk_callback_hsw+0x2bc>
+  DB  196,98,125,24,29,96,36,0,0          ; vbroadcastss  0x2460(%rip),%ymm11        # 49d0 <_sk_callback_hsw+0x2bc>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,90,36,0,0          ; vbroadcastss  0x245a(%rip),%ymm11        # 49d8 <_sk_callback_hsw+0x2c0>
+  DB  196,98,125,24,29,86,36,0,0          ; vbroadcastss  0x2456(%rip),%ymm11        # 49d4 <_sk_callback_hsw+0x2c0>
   DB  196,98,45,172,217                   ; vfnmadd213ps  %ymm1,%ymm10,%ymm11
-  DB  196,226,125,24,13,80,36,0,0         ; vbroadcastss  0x2450(%rip),%ymm1        # 49dc <_sk_callback_hsw+0x2c4>
+  DB  196,226,125,24,13,76,36,0,0         ; vbroadcastss  0x244c(%rip),%ymm1        # 49d8 <_sk_callback_hsw+0x2c4>
   DB  196,193,116,92,202                  ; vsubps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,21,70,36,0,0          ; vbroadcastss  0x2446(%rip),%ymm10        # 49e0 <_sk_callback_hsw+0x2c8>
+  DB  196,98,125,24,21,66,36,0,0          ; vbroadcastss  0x2442(%rip),%ymm10        # 49dc <_sk_callback_hsw+0x2c8>
   DB  197,172,94,201                      ; vdivps        %ymm1,%ymm10,%ymm1
   DB  197,164,88,201                      ; vaddps        %ymm1,%ymm11,%ymm1
-  DB  196,98,125,24,21,57,36,0,0          ; vbroadcastss  0x2439(%rip),%ymm10        # 49e4 <_sk_callback_hsw+0x2cc>
+  DB  196,98,125,24,21,53,36,0,0          ; vbroadcastss  0x2435(%rip),%ymm10        # 49e0 <_sk_callback_hsw+0x2cc>
   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
@@ -2269,7 +2269,7 @@ _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  196,98,125,24,5,16,36,0,0           ; vbroadcastss  0x2410(%rip),%ymm8        # 49e8 <_sk_callback_hsw+0x2d0>
+  DB  196,98,125,24,5,12,36,0,0           ; vbroadcastss  0x240c(%rip),%ymm8        # 49e4 <_sk_callback_hsw+0x2d0>
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2287,33 +2287,33 @@ _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  196,98,125,24,37,200,35,0,0         ; vbroadcastss  0x23c8(%rip),%ymm12        # 49ec <_sk_callback_hsw+0x2d4>
-  DB  196,98,125,24,45,195,35,0,0         ; vbroadcastss  0x23c3(%rip),%ymm13        # 49f0 <_sk_callback_hsw+0x2d8>
+  DB  196,98,125,24,37,196,35,0,0         ; vbroadcastss  0x23c4(%rip),%ymm12        # 49e8 <_sk_callback_hsw+0x2d4>
+  DB  196,98,125,24,45,191,35,0,0         ; vbroadcastss  0x23bf(%rip),%ymm13        # 49ec <_sk_callback_hsw+0x2d8>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,185,35,0,0         ; vbroadcastss  0x23b9(%rip),%ymm13        # 49f4 <_sk_callback_hsw+0x2dc>
+  DB  196,98,125,24,45,181,35,0,0         ; vbroadcastss  0x23b5(%rip),%ymm13        # 49f0 <_sk_callback_hsw+0x2dc>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,175,35,0,0         ; vbroadcastss  0x23af(%rip),%ymm13        # 49f8 <_sk_callback_hsw+0x2e0>
+  DB  196,98,125,24,45,171,35,0,0         ; vbroadcastss  0x23ab(%rip),%ymm13        # 49f4 <_sk_callback_hsw+0x2e0>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,165,35,0,0         ; vbroadcastss  0x23a5(%rip),%ymm11        # 49fc <_sk_callback_hsw+0x2e4>
+  DB  196,98,125,24,29,161,35,0,0         ; vbroadcastss  0x23a1(%rip),%ymm11        # 49f8 <_sk_callback_hsw+0x2e4>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,155,35,0,0         ; vbroadcastss  0x239b(%rip),%ymm12        # 4a00 <_sk_callback_hsw+0x2e8>
+  DB  196,98,125,24,37,151,35,0,0         ; vbroadcastss  0x2397(%rip),%ymm12        # 49fc <_sk_callback_hsw+0x2e8>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,145,35,0,0         ; vbroadcastss  0x2391(%rip),%ymm12        # 4a04 <_sk_callback_hsw+0x2ec>
+  DB  196,98,125,24,37,141,35,0,0         ; vbroadcastss  0x238d(%rip),%ymm12        # 4a00 <_sk_callback_hsw+0x2ec>
   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  196,98,125,24,29,114,35,0,0         ; vbroadcastss  0x2372(%rip),%ymm11        # 4a08 <_sk_callback_hsw+0x2f0>
+  DB  196,98,125,24,29,110,35,0,0         ; vbroadcastss  0x236e(%rip),%ymm11        # 4a04 <_sk_callback_hsw+0x2f0>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,104,35,0,0         ; vbroadcastss  0x2368(%rip),%ymm11        # 4a0c <_sk_callback_hsw+0x2f4>
+  DB  196,98,125,24,29,100,35,0,0         ; vbroadcastss  0x2364(%rip),%ymm11        # 4a08 <_sk_callback_hsw+0x2f4>
   DB  196,98,45,172,218                   ; vfnmadd213ps  %ymm2,%ymm10,%ymm11
-  DB  196,226,125,24,21,94,35,0,0         ; vbroadcastss  0x235e(%rip),%ymm2        # 4a10 <_sk_callback_hsw+0x2f8>
+  DB  196,226,125,24,21,90,35,0,0         ; vbroadcastss  0x235a(%rip),%ymm2        # 4a0c <_sk_callback_hsw+0x2f8>
   DB  196,193,108,92,210                  ; vsubps        %ymm10,%ymm2,%ymm2
-  DB  196,98,125,24,21,84,35,0,0          ; vbroadcastss  0x2354(%rip),%ymm10        # 4a14 <_sk_callback_hsw+0x2fc>
+  DB  196,98,125,24,21,80,35,0,0          ; vbroadcastss  0x2350(%rip),%ymm10        # 4a10 <_sk_callback_hsw+0x2fc>
   DB  197,172,94,210                      ; vdivps        %ymm2,%ymm10,%ymm2
   DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
-  DB  196,98,125,24,21,71,35,0,0          ; vbroadcastss  0x2347(%rip),%ymm10        # 4a18 <_sk_callback_hsw+0x300>
+  DB  196,98,125,24,21,67,35,0,0          ; vbroadcastss  0x2343(%rip),%ymm10        # 4a14 <_sk_callback_hsw+0x300>
   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
@@ -2321,7 +2321,7 @@ _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  196,98,125,24,5,30,35,0,0           ; vbroadcastss  0x231e(%rip),%ymm8        # 4a1c <_sk_callback_hsw+0x304>
+  DB  196,98,125,24,5,26,35,0,0           ; vbroadcastss  0x231a(%rip),%ymm8        # 4a18 <_sk_callback_hsw+0x304>
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2339,33 +2339,33 @@ _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  196,98,125,24,37,214,34,0,0         ; vbroadcastss  0x22d6(%rip),%ymm12        # 4a20 <_sk_callback_hsw+0x308>
-  DB  196,98,125,24,45,209,34,0,0         ; vbroadcastss  0x22d1(%rip),%ymm13        # 4a24 <_sk_callback_hsw+0x30c>
+  DB  196,98,125,24,37,210,34,0,0         ; vbroadcastss  0x22d2(%rip),%ymm12        # 4a1c <_sk_callback_hsw+0x308>
+  DB  196,98,125,24,45,205,34,0,0         ; vbroadcastss  0x22cd(%rip),%ymm13        # 4a20 <_sk_callback_hsw+0x30c>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,199,34,0,0         ; vbroadcastss  0x22c7(%rip),%ymm13        # 4a28 <_sk_callback_hsw+0x310>
+  DB  196,98,125,24,45,195,34,0,0         ; vbroadcastss  0x22c3(%rip),%ymm13        # 4a24 <_sk_callback_hsw+0x310>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,189,34,0,0         ; vbroadcastss  0x22bd(%rip),%ymm13        # 4a2c <_sk_callback_hsw+0x314>
+  DB  196,98,125,24,45,185,34,0,0         ; vbroadcastss  0x22b9(%rip),%ymm13        # 4a28 <_sk_callback_hsw+0x314>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,179,34,0,0         ; vbroadcastss  0x22b3(%rip),%ymm11        # 4a30 <_sk_callback_hsw+0x318>
+  DB  196,98,125,24,29,175,34,0,0         ; vbroadcastss  0x22af(%rip),%ymm11        # 4a2c <_sk_callback_hsw+0x318>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,169,34,0,0         ; vbroadcastss  0x22a9(%rip),%ymm12        # 4a34 <_sk_callback_hsw+0x31c>
+  DB  196,98,125,24,37,165,34,0,0         ; vbroadcastss  0x22a5(%rip),%ymm12        # 4a30 <_sk_callback_hsw+0x31c>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,159,34,0,0         ; vbroadcastss  0x229f(%rip),%ymm12        # 4a38 <_sk_callback_hsw+0x320>
+  DB  196,98,125,24,37,155,34,0,0         ; vbroadcastss  0x229b(%rip),%ymm12        # 4a34 <_sk_callback_hsw+0x320>
   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  196,98,125,24,29,128,34,0,0         ; vbroadcastss  0x2280(%rip),%ymm11        # 4a3c <_sk_callback_hsw+0x324>
+  DB  196,98,125,24,29,124,34,0,0         ; vbroadcastss  0x227c(%rip),%ymm11        # 4a38 <_sk_callback_hsw+0x324>
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,118,34,0,0         ; vbroadcastss  0x2276(%rip),%ymm11        # 4a40 <_sk_callback_hsw+0x328>
+  DB  196,98,125,24,29,114,34,0,0         ; vbroadcastss  0x2272(%rip),%ymm11        # 4a3c <_sk_callback_hsw+0x328>
   DB  196,98,45,172,219                   ; vfnmadd213ps  %ymm3,%ymm10,%ymm11
-  DB  196,226,125,24,29,108,34,0,0        ; vbroadcastss  0x226c(%rip),%ymm3        # 4a44 <_sk_callback_hsw+0x32c>
+  DB  196,226,125,24,29,104,34,0,0        ; vbroadcastss  0x2268(%rip),%ymm3        # 4a40 <_sk_callback_hsw+0x32c>
   DB  196,193,100,92,218                  ; vsubps        %ymm10,%ymm3,%ymm3
-  DB  196,98,125,24,21,98,34,0,0          ; vbroadcastss  0x2262(%rip),%ymm10        # 4a48 <_sk_callback_hsw+0x330>
+  DB  196,98,125,24,21,94,34,0,0          ; vbroadcastss  0x225e(%rip),%ymm10        # 4a44 <_sk_callback_hsw+0x330>
   DB  197,172,94,219                      ; vdivps        %ymm3,%ymm10,%ymm3
   DB  197,164,88,219                      ; vaddps        %ymm3,%ymm11,%ymm3
-  DB  196,98,125,24,21,85,34,0,0          ; vbroadcastss  0x2255(%rip),%ymm10        # 4a4c <_sk_callback_hsw+0x334>
+  DB  196,98,125,24,21,81,34,0,0          ; vbroadcastss  0x2251(%rip),%ymm10        # 4a48 <_sk_callback_hsw+0x334>
   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
@@ -2373,33 +2373,33 @@ _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  196,98,125,24,5,44,34,0,0           ; vbroadcastss  0x222c(%rip),%ymm8        # 4a50 <_sk_callback_hsw+0x338>
+  DB  196,98,125,24,5,40,34,0,0           ; vbroadcastss  0x2228(%rip),%ymm8        # 4a4c <_sk_callback_hsw+0x338>
   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  196,98,125,24,5,30,34,0,0           ; vbroadcastss  0x221e(%rip),%ymm8        # 4a54 <_sk_callback_hsw+0x33c>
-  DB  196,98,125,24,13,25,34,0,0          ; vbroadcastss  0x2219(%rip),%ymm9        # 4a58 <_sk_callback_hsw+0x340>
-  DB  196,98,125,24,21,20,34,0,0          ; vbroadcastss  0x2214(%rip),%ymm10        # 4a5c <_sk_callback_hsw+0x344>
+  DB  196,98,125,24,5,26,34,0,0           ; vbroadcastss  0x221a(%rip),%ymm8        # 4a50 <_sk_callback_hsw+0x33c>
+  DB  196,98,125,24,13,21,34,0,0          ; vbroadcastss  0x2215(%rip),%ymm9        # 4a54 <_sk_callback_hsw+0x340>
+  DB  196,98,125,24,21,16,34,0,0          ; vbroadcastss  0x2210(%rip),%ymm10        # 4a58 <_sk_callback_hsw+0x344>
   DB  196,194,53,168,202                  ; vfmadd213ps   %ymm10,%ymm9,%ymm1
   DB  196,194,53,168,210                  ; vfmadd213ps   %ymm10,%ymm9,%ymm2
-  DB  196,98,125,24,13,5,34,0,0           ; vbroadcastss  0x2205(%rip),%ymm9        # 4a60 <_sk_callback_hsw+0x348>
+  DB  196,98,125,24,13,1,34,0,0           ; vbroadcastss  0x2201(%rip),%ymm9        # 4a5c <_sk_callback_hsw+0x348>
   DB  196,66,125,184,200                  ; vfmadd231ps   %ymm8,%ymm0,%ymm9
-  DB  196,226,125,24,5,251,33,0,0         ; vbroadcastss  0x21fb(%rip),%ymm0        # 4a64 <_sk_callback_hsw+0x34c>
+  DB  196,226,125,24,5,247,33,0,0         ; vbroadcastss  0x21f7(%rip),%ymm0        # 4a60 <_sk_callback_hsw+0x34c>
   DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
-  DB  196,98,125,24,5,242,33,0,0          ; vbroadcastss  0x21f2(%rip),%ymm8        # 4a68 <_sk_callback_hsw+0x350>
+  DB  196,98,125,24,5,238,33,0,0          ; vbroadcastss  0x21ee(%rip),%ymm8        # 4a64 <_sk_callback_hsw+0x350>
   DB  196,98,117,168,192                  ; vfmadd213ps   %ymm0,%ymm1,%ymm8
-  DB  196,98,125,24,13,232,33,0,0         ; vbroadcastss  0x21e8(%rip),%ymm9        # 4a6c <_sk_callback_hsw+0x354>
+  DB  196,98,125,24,13,228,33,0,0         ; vbroadcastss  0x21e4(%rip),%ymm9        # 4a68 <_sk_callback_hsw+0x354>
   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  196,226,125,24,21,213,33,0,0        ; vbroadcastss  0x21d5(%rip),%ymm2        # 4a70 <_sk_callback_hsw+0x358>
+  DB  196,226,125,24,21,209,33,0,0        ; vbroadcastss  0x21d1(%rip),%ymm2        # 4a6c <_sk_callback_hsw+0x358>
   DB  197,108,194,209,1                   ; vcmpltps      %ymm1,%ymm2,%ymm10
-  DB  196,98,125,24,29,203,33,0,0         ; vbroadcastss  0x21cb(%rip),%ymm11        # 4a74 <_sk_callback_hsw+0x35c>
+  DB  196,98,125,24,29,199,33,0,0         ; vbroadcastss  0x21c7(%rip),%ymm11        # 4a70 <_sk_callback_hsw+0x35c>
   DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
-  DB  196,98,125,24,37,193,33,0,0         ; vbroadcastss  0x21c1(%rip),%ymm12        # 4a78 <_sk_callback_hsw+0x360>
+  DB  196,98,125,24,37,189,33,0,0         ; vbroadcastss  0x21bd(%rip),%ymm12        # 4a74 <_sk_callback_hsw+0x360>
   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
@@ -2414,9 +2414,9 @@ _sk_lab_to_xyz_hsw LABEL PROC
   DB  196,65,52,88,203                    ; vaddps        %ymm11,%ymm9,%ymm9
   DB  196,65,52,89,204                    ; vmulps        %ymm12,%ymm9,%ymm9
   DB  196,227,53,74,208,32                ; vblendvps     %ymm2,%ymm0,%ymm9,%ymm2
-  DB  196,226,125,24,5,118,33,0,0         ; vbroadcastss  0x2176(%rip),%ymm0        # 4a7c <_sk_callback_hsw+0x364>
+  DB  196,226,125,24,5,114,33,0,0         ; vbroadcastss  0x2172(%rip),%ymm0        # 4a78 <_sk_callback_hsw+0x364>
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,109,33,0,0          ; vbroadcastss  0x216d(%rip),%ymm8        # 4a80 <_sk_callback_hsw+0x368>
+  DB  196,98,125,24,5,105,33,0,0          ; vbroadcastss  0x2169(%rip),%ymm8        # 4a7c <_sk_callback_hsw+0x368>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2432,7 +2432,7 @@ _sk_load_a8_hsw LABEL PROC
   DB  197,250,126,0                       ; vmovq         (%rax),%xmm0
   DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,66,33,0,0         ; vbroadcastss  0x2142(%rip),%ymm1        # 4a84 <_sk_callback_hsw+0x36c>
+  DB  196,226,125,24,13,62,33,0,0         ; vbroadcastss  0x213e(%rip),%ymm1        # 4a80 <_sk_callback_hsw+0x36c>
   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
@@ -2495,7 +2495,7 @@ _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  196,226,125,24,13,77,32,0,0         ; vbroadcastss  0x204d(%rip),%ymm1        # 4a88 <_sk_callback_hsw+0x370>
+  DB  196,226,125,24,13,73,32,0,0         ; vbroadcastss  0x2049(%rip),%ymm1        # 4a84 <_sk_callback_hsw+0x370>
   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
@@ -2511,7 +2511,7 @@ PUBLIC _sk_store_a8_hsw
 _sk_store_a8_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,40,32,0,0           ; vbroadcastss  0x2028(%rip),%ymm8        # 4a8c <_sk_callback_hsw+0x374>
+  DB  196,98,125,24,5,36,32,0,0           ; vbroadcastss  0x2024(%rip),%ymm8        # 4a88 <_sk_callback_hsw+0x374>
   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
@@ -2576,10 +2576,10 @@ _sk_load_g8_hsw LABEL PROC
   DB  197,250,126,0                       ; vmovq         (%rax),%xmm0
   DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,94,31,0,0         ; vbroadcastss  0x1f5e(%rip),%ymm1        # 4a90 <_sk_callback_hsw+0x378>
+  DB  196,226,125,24,13,90,31,0,0         ; vbroadcastss  0x1f5a(%rip),%ymm1        # 4a8c <_sk_callback_hsw+0x378>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,83,31,0,0         ; vbroadcastss  0x1f53(%rip),%ymm3        # 4a94 <_sk_callback_hsw+0x37c>
+  DB  196,226,125,24,29,79,31,0,0         ; vbroadcastss  0x1f4f(%rip),%ymm3        # 4a90 <_sk_callback_hsw+0x37c>
   DB  76,137,193                          ; mov           %r8,%rcx
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
@@ -2639,10 +2639,10 @@ _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  196,226,125,24,13,104,30,0,0        ; vbroadcastss  0x1e68(%rip),%ymm1        # 4a98 <_sk_callback_hsw+0x380>
+  DB  196,226,125,24,13,100,30,0,0        ; vbroadcastss  0x1e64(%rip),%ymm1        # 4a94 <_sk_callback_hsw+0x380>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,93,30,0,0         ; vbroadcastss  0x1e5d(%rip),%ymm3        # 4a9c <_sk_callback_hsw+0x384>
+  DB  196,226,125,24,29,89,30,0,0         ; vbroadcastss  0x1e59(%rip),%ymm3        # 4a98 <_sk_callback_hsw+0x384>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  91                                  ; pop           %rbx
@@ -2696,14 +2696,14 @@ _sk_gather_i8_hsw LABEL PROC
   DB  73,139,64,8                         ; mov           0x8(%r8),%rax
   DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
   DB  196,226,117,144,28,128              ; vpgatherdd    %ymm1,(%rax,%ymm0,4),%ymm3
-  DB  197,229,219,5,113,31,0,0            ; vpand         0x1f71(%rip),%ymm3,%ymm0        # 4c80 <_sk_callback_hsw+0x568>
+  DB  197,229,219,5,113,31,0,0            ; vpand         0x1f71(%rip),%ymm3,%ymm0        # 4c80 <_sk_callback_hsw+0x56c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,132,29,0,0          ; vbroadcastss  0x1d84(%rip),%ymm8        # 4aa0 <_sk_callback_hsw+0x388>
+  DB  196,98,125,24,5,128,29,0,0          ; vbroadcastss  0x1d80(%rip),%ymm8        # 4a9c <_sk_callback_hsw+0x388>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,118,31,0,0         ; vpshufb       0x1f76(%rip),%ymm3,%ymm1        # 4ca0 <_sk_callback_hsw+0x588>
+  DB  196,226,101,0,13,118,31,0,0         ; vpshufb       0x1f76(%rip),%ymm3,%ymm1        # 4ca0 <_sk_callback_hsw+0x58c>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,132,31,0,0         ; vpshufb       0x1f84(%rip),%ymm3,%ymm2        # 4cc0 <_sk_callback_hsw+0x5a8>
+  DB  196,226,101,0,21,132,31,0,0         ; vpshufb       0x1f84(%rip),%ymm3,%ymm2        # 4cc0 <_sk_callback_hsw+0x5ac>
   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
@@ -2725,23 +2725,23 @@ _sk_load_565_hsw LABEL PROC
   DB  117,114                             ; jne           2ddc <_sk_load_565_hsw+0x7c>
   DB  196,193,122,111,4,122               ; vmovdqu       (%r10,%rdi,2),%xmm0
   DB  196,226,125,51,208                  ; vpmovzxwd     %xmm0,%ymm2
-  DB  196,226,125,88,5,38,29,0,0          ; vpbroadcastd  0x1d26(%rip),%ymm0        # 4aa4 <_sk_callback_hsw+0x38c>
+  DB  196,226,125,88,5,34,29,0,0          ; vpbroadcastd  0x1d22(%rip),%ymm0        # 4aa0 <_sk_callback_hsw+0x38c>
   DB  197,237,219,192                     ; vpand         %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,25,29,0,0         ; vbroadcastss  0x1d19(%rip),%ymm1        # 4aa8 <_sk_callback_hsw+0x390>
+  DB  196,226,125,24,13,21,29,0,0         ; vbroadcastss  0x1d15(%rip),%ymm1        # 4aa4 <_sk_callback_hsw+0x390>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,16,29,0,0         ; vpbroadcastd  0x1d10(%rip),%ymm1        # 4aac <_sk_callback_hsw+0x394>
+  DB  196,226,125,88,13,12,29,0,0         ; vpbroadcastd  0x1d0c(%rip),%ymm1        # 4aa8 <_sk_callback_hsw+0x394>
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,3,29,0,0          ; vbroadcastss  0x1d03(%rip),%ymm3        # 4ab0 <_sk_callback_hsw+0x398>
+  DB  196,226,125,24,29,255,28,0,0        ; vbroadcastss  0x1cff(%rip),%ymm3        # 4aac <_sk_callback_hsw+0x398>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,88,29,250,28,0,0        ; vpbroadcastd  0x1cfa(%rip),%ymm3        # 4ab4 <_sk_callback_hsw+0x39c>
+  DB  196,226,125,88,29,246,28,0,0        ; vpbroadcastd  0x1cf6(%rip),%ymm3        # 4ab0 <_sk_callback_hsw+0x39c>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,237,28,0,0        ; vbroadcastss  0x1ced(%rip),%ymm3        # 4ab8 <_sk_callback_hsw+0x3a0>
+  DB  196,226,125,24,29,233,28,0,0        ; vbroadcastss  0x1ce9(%rip),%ymm3        # 4ab4 <_sk_callback_hsw+0x3a0>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,226,28,0,0        ; vbroadcastss  0x1ce2(%rip),%ymm3        # 4abc <_sk_callback_hsw+0x3a4>
+  DB  196,226,125,24,29,222,28,0,0        ; vbroadcastss  0x1cde(%rip),%ymm3        # 4ab8 <_sk_callback_hsw+0x3a4>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,200                          ; mov           %ecx,%r8d
   DB  65,128,224,7                        ; and           $0x7,%r8b
@@ -2830,23 +2830,23 @@ _sk_gather_565_hsw LABEL PROC
   DB  65,15,183,4,88                      ; movzwl        (%r8,%rbx,2),%eax
   DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
   DB  196,226,125,51,208                  ; vpmovzxwd     %xmm0,%ymm2
-  DB  196,226,125,88,5,165,27,0,0         ; vpbroadcastd  0x1ba5(%rip),%ymm0        # 4ac0 <_sk_callback_hsw+0x3a8>
+  DB  196,226,125,88,5,161,27,0,0         ; vpbroadcastd  0x1ba1(%rip),%ymm0        # 4abc <_sk_callback_hsw+0x3a8>
   DB  197,237,219,192                     ; vpand         %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,152,27,0,0        ; vbroadcastss  0x1b98(%rip),%ymm1        # 4ac4 <_sk_callback_hsw+0x3ac>
+  DB  196,226,125,24,13,148,27,0,0        ; vbroadcastss  0x1b94(%rip),%ymm1        # 4ac0 <_sk_callback_hsw+0x3ac>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,143,27,0,0        ; vpbroadcastd  0x1b8f(%rip),%ymm1        # 4ac8 <_sk_callback_hsw+0x3b0>
+  DB  196,226,125,88,13,139,27,0,0        ; vpbroadcastd  0x1b8b(%rip),%ymm1        # 4ac4 <_sk_callback_hsw+0x3b0>
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,130,27,0,0        ; vbroadcastss  0x1b82(%rip),%ymm3        # 4acc <_sk_callback_hsw+0x3b4>
+  DB  196,226,125,24,29,126,27,0,0        ; vbroadcastss  0x1b7e(%rip),%ymm3        # 4ac8 <_sk_callback_hsw+0x3b4>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,88,29,121,27,0,0        ; vpbroadcastd  0x1b79(%rip),%ymm3        # 4ad0 <_sk_callback_hsw+0x3b8>
+  DB  196,226,125,88,29,117,27,0,0        ; vpbroadcastd  0x1b75(%rip),%ymm3        # 4acc <_sk_callback_hsw+0x3b8>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,108,27,0,0        ; vbroadcastss  0x1b6c(%rip),%ymm3        # 4ad4 <_sk_callback_hsw+0x3bc>
+  DB  196,226,125,24,29,104,27,0,0        ; vbroadcastss  0x1b68(%rip),%ymm3        # 4ad0 <_sk_callback_hsw+0x3bc>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,97,27,0,0         ; vbroadcastss  0x1b61(%rip),%ymm3        # 4ad8 <_sk_callback_hsw+0x3c0>
+  DB  196,226,125,24,29,93,27,0,0         ; vbroadcastss  0x1b5d(%rip),%ymm3        # 4ad4 <_sk_callback_hsw+0x3c0>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,94                               ; pop           %r14
@@ -2857,11 +2857,11 @@ PUBLIC _sk_store_565_hsw
 _sk_store_565_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,78,27,0,0           ; vbroadcastss  0x1b4e(%rip),%ymm8        # 4adc <_sk_callback_hsw+0x3c4>
+  DB  196,98,125,24,5,74,27,0,0           ; vbroadcastss  0x1b4a(%rip),%ymm8        # 4ad8 <_sk_callback_hsw+0x3c4>
   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  196,98,125,24,21,57,27,0,0          ; vbroadcastss  0x1b39(%rip),%ymm10        # 4ae0 <_sk_callback_hsw+0x3c8>
+  DB  196,98,125,24,21,53,27,0,0          ; vbroadcastss  0x1b35(%rip),%ymm10        # 4adc <_sk_callback_hsw+0x3c8>
   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
@@ -2927,25 +2927,25 @@ _sk_load_4444_hsw LABEL PROC
   DB  15,133,138,0,0,0                    ; jne           30f8 <_sk_load_4444_hsw+0x98>
   DB  196,193,122,111,4,122               ; vmovdqu       (%r10,%rdi,2),%xmm0
   DB  196,226,125,51,216                  ; vpmovzxwd     %xmm0,%ymm3
-  DB  196,226,125,88,5,98,26,0,0          ; vpbroadcastd  0x1a62(%rip),%ymm0        # 4ae4 <_sk_callback_hsw+0x3cc>
+  DB  196,226,125,88,5,94,26,0,0          ; vpbroadcastd  0x1a5e(%rip),%ymm0        # 4ae0 <_sk_callback_hsw+0x3cc>
   DB  197,229,219,192                     ; vpand         %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,85,26,0,0         ; vbroadcastss  0x1a55(%rip),%ymm1        # 4ae8 <_sk_callback_hsw+0x3d0>
+  DB  196,226,125,24,13,81,26,0,0         ; vbroadcastss  0x1a51(%rip),%ymm1        # 4ae4 <_sk_callback_hsw+0x3d0>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,76,26,0,0         ; vpbroadcastd  0x1a4c(%rip),%ymm1        # 4aec <_sk_callback_hsw+0x3d4>
+  DB  196,226,125,88,13,72,26,0,0         ; vpbroadcastd  0x1a48(%rip),%ymm1        # 4ae8 <_sk_callback_hsw+0x3d4>
   DB  197,229,219,201                     ; vpand         %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,63,26,0,0         ; vbroadcastss  0x1a3f(%rip),%ymm2        # 4af0 <_sk_callback_hsw+0x3d8>
+  DB  196,226,125,24,21,59,26,0,0         ; vbroadcastss  0x1a3b(%rip),%ymm2        # 4aec <_sk_callback_hsw+0x3d8>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,88,21,54,26,0,0         ; vpbroadcastd  0x1a36(%rip),%ymm2        # 4af4 <_sk_callback_hsw+0x3dc>
+  DB  196,226,125,88,21,50,26,0,0         ; vpbroadcastd  0x1a32(%rip),%ymm2        # 4af0 <_sk_callback_hsw+0x3dc>
   DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,41,26,0,0           ; vbroadcastss  0x1a29(%rip),%ymm8        # 4af8 <_sk_callback_hsw+0x3e0>
+  DB  196,98,125,24,5,37,26,0,0           ; vbroadcastss  0x1a25(%rip),%ymm8        # 4af4 <_sk_callback_hsw+0x3e0>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,88,5,31,26,0,0           ; vpbroadcastd  0x1a1f(%rip),%ymm8        # 4afc <_sk_callback_hsw+0x3e4>
+  DB  196,98,125,88,5,27,26,0,0           ; vpbroadcastd  0x1a1b(%rip),%ymm8        # 4af8 <_sk_callback_hsw+0x3e4>
   DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,17,26,0,0           ; vbroadcastss  0x1a11(%rip),%ymm8        # 4b00 <_sk_callback_hsw+0x3e8>
+  DB  196,98,125,24,5,13,26,0,0           ; vbroadcastss  0x1a0d(%rip),%ymm8        # 4afc <_sk_callback_hsw+0x3e8>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -3036,25 +3036,25 @@ _sk_gather_4444_hsw LABEL PROC
   DB  65,15,183,4,88                      ; movzwl        (%r8,%rbx,2),%eax
   DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
   DB  196,226,125,51,216                  ; vpmovzxwd     %xmm0,%ymm3
-  DB  196,226,125,88,5,201,24,0,0         ; vpbroadcastd  0x18c9(%rip),%ymm0        # 4b04 <_sk_callback_hsw+0x3ec>
+  DB  196,226,125,88,5,197,24,0,0         ; vpbroadcastd  0x18c5(%rip),%ymm0        # 4b00 <_sk_callback_hsw+0x3ec>
   DB  197,229,219,192                     ; vpand         %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,188,24,0,0        ; vbroadcastss  0x18bc(%rip),%ymm1        # 4b08 <_sk_callback_hsw+0x3f0>
+  DB  196,226,125,24,13,184,24,0,0        ; vbroadcastss  0x18b8(%rip),%ymm1        # 4b04 <_sk_callback_hsw+0x3f0>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,179,24,0,0        ; vpbroadcastd  0x18b3(%rip),%ymm1        # 4b0c <_sk_callback_hsw+0x3f4>
+  DB  196,226,125,88,13,175,24,0,0        ; vpbroadcastd  0x18af(%rip),%ymm1        # 4b08 <_sk_callback_hsw+0x3f4>
   DB  197,229,219,201                     ; vpand         %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,166,24,0,0        ; vbroadcastss  0x18a6(%rip),%ymm2        # 4b10 <_sk_callback_hsw+0x3f8>
+  DB  196,226,125,24,21,162,24,0,0        ; vbroadcastss  0x18a2(%rip),%ymm2        # 4b0c <_sk_callback_hsw+0x3f8>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,88,21,157,24,0,0        ; vpbroadcastd  0x189d(%rip),%ymm2        # 4b14 <_sk_callback_hsw+0x3fc>
+  DB  196,226,125,88,21,153,24,0,0        ; vpbroadcastd  0x1899(%rip),%ymm2        # 4b10 <_sk_callback_hsw+0x3fc>
   DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,144,24,0,0          ; vbroadcastss  0x1890(%rip),%ymm8        # 4b18 <_sk_callback_hsw+0x400>
+  DB  196,98,125,24,5,140,24,0,0          ; vbroadcastss  0x188c(%rip),%ymm8        # 4b14 <_sk_callback_hsw+0x400>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,88,5,134,24,0,0          ; vpbroadcastd  0x1886(%rip),%ymm8        # 4b1c <_sk_callback_hsw+0x404>
+  DB  196,98,125,88,5,130,24,0,0          ; vpbroadcastd  0x1882(%rip),%ymm8        # 4b18 <_sk_callback_hsw+0x404>
   DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,120,24,0,0          ; vbroadcastss  0x1878(%rip),%ymm8        # 4b20 <_sk_callback_hsw+0x408>
+  DB  196,98,125,24,5,116,24,0,0          ; vbroadcastss  0x1874(%rip),%ymm8        # 4b1c <_sk_callback_hsw+0x408>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -3067,7 +3067,7 @@ PUBLIC _sk_store_4444_hsw
 _sk_store_4444_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,94,24,0,0           ; vbroadcastss  0x185e(%rip),%ymm8        # 4b24 <_sk_callback_hsw+0x40c>
+  DB  196,98,125,24,5,90,24,0,0           ; vbroadcastss  0x185a(%rip),%ymm8        # 4b20 <_sk_callback_hsw+0x40c>
   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
@@ -3141,14 +3141,14 @@ _sk_load_8888_hsw LABEL PROC
   DB  77,133,192                          ; test          %r8,%r8
   DB  117,88                              ; jne           3411 <_sk_load_8888_hsw+0x6d>
   DB  196,193,126,111,25                  ; vmovdqu       (%r9),%ymm3
-  DB  197,229,219,5,26,25,0,0             ; vpand         0x191a(%rip),%ymm3,%ymm0        # 4ce0 <_sk_callback_hsw+0x5c8>
+  DB  197,229,219,5,26,25,0,0             ; vpand         0x191a(%rip),%ymm3,%ymm0        # 4ce0 <_sk_callback_hsw+0x5cc>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,85,23,0,0           ; vbroadcastss  0x1755(%rip),%ymm8        # 4b28 <_sk_callback_hsw+0x410>
+  DB  196,98,125,24,5,81,23,0,0           ; vbroadcastss  0x1751(%rip),%ymm8        # 4b24 <_sk_callback_hsw+0x410>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,31,25,0,0          ; vpshufb       0x191f(%rip),%ymm3,%ymm1        # 4d00 <_sk_callback_hsw+0x5e8>
+  DB  196,226,101,0,13,31,25,0,0          ; vpshufb       0x191f(%rip),%ymm3,%ymm1        # 4d00 <_sk_callback_hsw+0x5ec>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,45,25,0,0          ; vpshufb       0x192d(%rip),%ymm3,%ymm2        # 4d20 <_sk_callback_hsw+0x608>
+  DB  196,226,101,0,21,45,25,0,0          ; vpshufb       0x192d(%rip),%ymm3,%ymm2        # 4d20 <_sk_callback_hsw+0x60c>
   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
@@ -3178,14 +3178,14 @@ _sk_gather_8888_hsw LABEL PROC
   DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
   DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
   DB  196,194,117,144,28,128              ; vpgatherdd    %ymm1,(%r8,%ymm0,4),%ymm3
-  DB  197,229,219,5,219,24,0,0            ; vpand         0x18db(%rip),%ymm3,%ymm0        # 4d40 <_sk_callback_hsw+0x628>
+  DB  197,229,219,5,219,24,0,0            ; vpand         0x18db(%rip),%ymm3,%ymm0        # 4d40 <_sk_callback_hsw+0x62c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,186,22,0,0          ; vbroadcastss  0x16ba(%rip),%ymm8        # 4b2c <_sk_callback_hsw+0x414>
+  DB  196,98,125,24,5,182,22,0,0          ; vbroadcastss  0x16b6(%rip),%ymm8        # 4b28 <_sk_callback_hsw+0x414>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,224,24,0,0         ; vpshufb       0x18e0(%rip),%ymm3,%ymm1        # 4d60 <_sk_callback_hsw+0x648>
+  DB  196,226,101,0,13,224,24,0,0         ; vpshufb       0x18e0(%rip),%ymm3,%ymm1        # 4d60 <_sk_callback_hsw+0x64c>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,238,24,0,0         ; vpshufb       0x18ee(%rip),%ymm3,%ymm2        # 4d80 <_sk_callback_hsw+0x668>
+  DB  196,226,101,0,21,238,24,0,0         ; vpshufb       0x18ee(%rip),%ymm3,%ymm2        # 4d80 <_sk_callback_hsw+0x66c>
   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
@@ -3200,7 +3200,7 @@ _sk_store_8888_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
   DB  76,3,8                              ; add           (%rax),%r9
-  DB  196,98,125,24,5,106,22,0,0          ; vbroadcastss  0x166a(%rip),%ymm8        # 4b30 <_sk_callback_hsw+0x418>
+  DB  196,98,125,24,5,102,22,0,0          ; vbroadcastss  0x1666(%rip),%ymm8        # 4b2c <_sk_callback_hsw+0x418>
   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
@@ -3389,7 +3389,7 @@ _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  196,98,125,24,21,97,19,0,0          ; vbroadcastss  0x1361(%rip),%ymm10        # 4b34 <_sk_callback_hsw+0x41c>
+  DB  196,98,125,24,21,93,19,0,0          ; vbroadcastss  0x135d(%rip),%ymm10        # 4b30 <_sk_callback_hsw+0x41c>
   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
@@ -3471,7 +3471,7 @@ _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  196,98,125,24,21,242,17,0,0         ; vbroadcastss  0x11f2(%rip),%ymm10        # 4b38 <_sk_callback_hsw+0x420>
+  DB  196,98,125,24,21,238,17,0,0         ; vbroadcastss  0x11ee(%rip),%ymm10        # 4b34 <_sk_callback_hsw+0x420>
   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
@@ -3488,7 +3488,7 @@ _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  196,226,125,24,29,166,17,0,0        ; vbroadcastss  0x11a6(%rip),%ymm3        # 4b3c <_sk_callback_hsw+0x424>
+  DB  196,226,125,24,29,162,17,0,0        ; vbroadcastss  0x11a2(%rip),%ymm3        # 4b38 <_sk_callback_hsw+0x424>
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,4,64                ; vmovd         (%r8,%rax,2),%xmm0
   DB  196,193,121,196,68,64,4,2           ; vpinsrw       $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
@@ -3529,7 +3529,7 @@ _sk_store_u16_be_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  72,141,4,189,0,0,0,0                ; lea           0x0(,%rdi,4),%rax
-  DB  196,98,125,24,5,227,16,0,0          ; vbroadcastss  0x10e3(%rip),%ymm8        # 4b40 <_sk_callback_hsw+0x428>
+  DB  196,98,125,24,5,223,16,0,0          ; vbroadcastss  0x10df(%rip),%ymm8        # 4b3c <_sk_callback_hsw+0x428>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,67,125,25,202,1                 ; vextractf128  $0x1,%ymm9,%xmm10
@@ -3771,11 +3771,11 @@ _sk_mirror_y_hsw LABEL PROC
 
 PUBLIC _sk_luminance_to_alpha_hsw
 _sk_luminance_to_alpha_hsw LABEL PROC
-  DB  196,226,125,24,29,51,13,0,0         ; vbroadcastss  0xd33(%rip),%ymm3        # 4b44 <_sk_callback_hsw+0x42c>
-  DB  196,98,125,24,5,46,13,0,0           ; vbroadcastss  0xd2e(%rip),%ymm8        # 4b48 <_sk_callback_hsw+0x430>
+  DB  196,226,125,24,29,47,13,0,0         ; vbroadcastss  0xd2f(%rip),%ymm3        # 4b40 <_sk_callback_hsw+0x42c>
+  DB  196,98,125,24,5,42,13,0,0           ; vbroadcastss  0xd2a(%rip),%ymm8        # 4b44 <_sk_callback_hsw+0x430>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  196,226,125,184,203                 ; vfmadd231ps   %ymm3,%ymm0,%ymm1
-  DB  196,226,125,24,29,31,13,0,0         ; vbroadcastss  0xd1f(%rip),%ymm3        # 4b4c <_sk_callback_hsw+0x434>
+  DB  196,226,125,24,29,27,13,0,0         ; vbroadcastss  0xd1b(%rip),%ymm3        # 4b48 <_sk_callback_hsw+0x434>
   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
@@ -3978,7 +3978,7 @@ _sk_gradient_hsw LABEL PROC
   DB  76,139,72,72                        ; mov           0x48(%rax),%r9
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  65,186,1,0,0,0                      ; mov           $0x1,%r10d
-  DB  196,226,125,24,21,201,9,0,0         ; vbroadcastss  0x9c9(%rip),%ymm2        # 4b50 <_sk_callback_hsw+0x438>
+  DB  196,226,125,24,21,197,9,0,0         ; vbroadcastss  0x9c5(%rip),%ymm2        # 4b4c <_sk_callback_hsw+0x438>
   DB  196,65,53,239,201                   ; vpxor         %ymm9,%ymm9,%ymm9
   DB  196,130,125,24,28,145               ; vbroadcastss  (%r9,%r10,4),%ymm3
   DB  197,228,194,216,2                   ; vcmpleps      %ymm0,%ymm3,%ymm3
@@ -4069,24 +4069,24 @@ _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  196,98,125,24,45,72,8,0,0           ; vbroadcastss  0x848(%rip),%ymm13        # 4b54 <_sk_callback_hsw+0x43c>
-  DB  196,98,125,24,53,67,8,0,0           ; vbroadcastss  0x843(%rip),%ymm14        # 4b58 <_sk_callback_hsw+0x440>
+  DB  196,98,125,24,45,68,8,0,0           ; vbroadcastss  0x844(%rip),%ymm13        # 4b50 <_sk_callback_hsw+0x43c>
+  DB  196,98,125,24,53,63,8,0,0           ; vbroadcastss  0x83f(%rip),%ymm14        # 4b54 <_sk_callback_hsw+0x440>
   DB  196,66,29,184,245                   ; vfmadd231ps   %ymm13,%ymm12,%ymm14
-  DB  196,98,125,24,45,57,8,0,0           ; vbroadcastss  0x839(%rip),%ymm13        # 4b5c <_sk_callback_hsw+0x444>
+  DB  196,98,125,24,45,53,8,0,0           ; vbroadcastss  0x835(%rip),%ymm13        # 4b58 <_sk_callback_hsw+0x444>
   DB  196,66,29,184,238                   ; vfmadd231ps   %ymm14,%ymm12,%ymm13
-  DB  196,98,125,24,53,47,8,0,0           ; vbroadcastss  0x82f(%rip),%ymm14        # 4b60 <_sk_callback_hsw+0x448>
+  DB  196,98,125,24,53,43,8,0,0           ; vbroadcastss  0x82b(%rip),%ymm14        # 4b5c <_sk_callback_hsw+0x448>
   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  196,98,125,24,21,26,8,0,0           ; vbroadcastss  0x81a(%rip),%ymm10        # 4b64 <_sk_callback_hsw+0x44c>
+  DB  196,98,125,24,21,22,8,0,0           ; vbroadcastss  0x816(%rip),%ymm10        # 4b60 <_sk_callback_hsw+0x44c>
   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  196,98,125,24,21,4,8,0,0            ; vbroadcastss  0x804(%rip),%ymm10        # 4b68 <_sk_callback_hsw+0x450>
+  DB  196,98,125,24,21,0,8,0,0            ; vbroadcastss  0x800(%rip),%ymm10        # 4b64 <_sk_callback_hsw+0x450>
   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  196,98,125,24,21,238,7,0,0          ; vbroadcastss  0x7ee(%rip),%ymm10        # 4b6c <_sk_callback_hsw+0x454>
+  DB  196,98,125,24,21,234,7,0,0          ; vbroadcastss  0x7ea(%rip),%ymm10        # 4b68 <_sk_callback_hsw+0x454>
   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
@@ -4098,15 +4098,14 @@ PUBLIC _sk_xy_to_radius_hsw
 _sk_xy_to_radius_hsw LABEL PROC
   DB  197,116,89,193                      ; vmulps        %ymm1,%ymm1,%ymm8
   DB  196,98,125,184,192                  ; vfmadd231ps   %ymm0,%ymm0,%ymm8
-  DB  196,193,124,82,192                  ; vrsqrtps      %ymm8,%ymm0
-  DB  197,252,83,192                      ; vrcpps        %ymm0,%ymm0
+  DB  196,193,124,81,192                  ; vsqrtps       %ymm8,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_save_xy_hsw
 _sk_save_xy_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,183,7,0,0           ; vbroadcastss  0x7b7(%rip),%ymm8        # 4b70 <_sk_callback_hsw+0x458>
+  DB  196,98,125,24,5,183,7,0,0           ; vbroadcastss  0x7b7(%rip),%ymm8        # 4b6c <_sk_callback_hsw+0x458>
   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
@@ -4136,9 +4135,9 @@ _sk_accumulate_hsw LABEL PROC
 PUBLIC _sk_bilinear_nx_hsw
 _sk_bilinear_nx_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,75,7,0,0           ; vbroadcastss  0x74b(%rip),%ymm0        # 4b74 <_sk_callback_hsw+0x45c>
+  DB  196,226,125,24,5,75,7,0,0           ; vbroadcastss  0x74b(%rip),%ymm0        # 4b70 <_sk_callback_hsw+0x45c>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,66,7,0,0            ; vbroadcastss  0x742(%rip),%ymm8        # 4b78 <_sk_callback_hsw+0x460>
+  DB  196,98,125,24,5,66,7,0,0            ; vbroadcastss  0x742(%rip),%ymm8        # 4b74 <_sk_callback_hsw+0x460>
   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
@@ -4147,7 +4146,7 @@ _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  196,226,125,24,5,42,7,0,0           ; vbroadcastss  0x72a(%rip),%ymm0        # 4b7c <_sk_callback_hsw+0x464>
+  DB  196,226,125,24,5,42,7,0,0           ; vbroadcastss  0x72a(%rip),%ymm0        # 4b78 <_sk_callback_hsw+0x464>
   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)
@@ -4157,9 +4156,9 @@ _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  196,226,125,24,13,14,7,0,0          ; vbroadcastss  0x70e(%rip),%ymm1        # 4b80 <_sk_callback_hsw+0x468>
+  DB  196,226,125,24,13,14,7,0,0          ; vbroadcastss  0x70e(%rip),%ymm1        # 4b7c <_sk_callback_hsw+0x468>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,4,7,0,0             ; vbroadcastss  0x704(%rip),%ymm8        # 4b84 <_sk_callback_hsw+0x46c>
+  DB  196,98,125,24,5,4,7,0,0             ; vbroadcastss  0x704(%rip),%ymm8        # 4b80 <_sk_callback_hsw+0x46c>
   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
@@ -4168,7 +4167,7 @@ _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  196,226,125,24,13,236,6,0,0         ; vbroadcastss  0x6ec(%rip),%ymm1        # 4b88 <_sk_callback_hsw+0x470>
+  DB  196,226,125,24,13,236,6,0,0         ; vbroadcastss  0x6ec(%rip),%ymm1        # 4b84 <_sk_callback_hsw+0x470>
   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)
@@ -4178,13 +4177,13 @@ _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  196,226,125,24,5,207,6,0,0          ; vbroadcastss  0x6cf(%rip),%ymm0        # 4b8c <_sk_callback_hsw+0x474>
+  DB  196,226,125,24,5,207,6,0,0          ; vbroadcastss  0x6cf(%rip),%ymm0        # 4b88 <_sk_callback_hsw+0x474>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,198,6,0,0           ; vbroadcastss  0x6c6(%rip),%ymm8        # 4b90 <_sk_callback_hsw+0x478>
+  DB  196,98,125,24,5,198,6,0,0           ; vbroadcastss  0x6c6(%rip),%ymm8        # 4b8c <_sk_callback_hsw+0x478>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,183,6,0,0          ; vbroadcastss  0x6b7(%rip),%ymm10        # 4b94 <_sk_callback_hsw+0x47c>
-  DB  196,98,125,24,29,178,6,0,0          ; vbroadcastss  0x6b2(%rip),%ymm11        # 4b98 <_sk_callback_hsw+0x480>
+  DB  196,98,125,24,21,183,6,0,0          ; vbroadcastss  0x6b7(%rip),%ymm10        # 4b90 <_sk_callback_hsw+0x47c>
+  DB  196,98,125,24,29,178,6,0,0          ; vbroadcastss  0x6b2(%rip),%ymm11        # 4b94 <_sk_callback_hsw+0x480>
   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)
@@ -4194,16 +4193,16 @@ _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  196,226,125,24,5,149,6,0,0          ; vbroadcastss  0x695(%rip),%ymm0        # 4b9c <_sk_callback_hsw+0x484>
+  DB  196,226,125,24,5,149,6,0,0          ; vbroadcastss  0x695(%rip),%ymm0        # 4b98 <_sk_callback_hsw+0x484>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,140,6,0,0           ; vbroadcastss  0x68c(%rip),%ymm8        # 4ba0 <_sk_callback_hsw+0x488>
+  DB  196,98,125,24,5,140,6,0,0           ; vbroadcastss  0x68c(%rip),%ymm8        # 4b9c <_sk_callback_hsw+0x488>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,130,6,0,0          ; vbroadcastss  0x682(%rip),%ymm9        # 4ba4 <_sk_callback_hsw+0x48c>
-  DB  196,98,125,24,21,125,6,0,0          ; vbroadcastss  0x67d(%rip),%ymm10        # 4ba8 <_sk_callback_hsw+0x490>
+  DB  196,98,125,24,13,130,6,0,0          ; vbroadcastss  0x682(%rip),%ymm9        # 4ba0 <_sk_callback_hsw+0x48c>
+  DB  196,98,125,24,21,125,6,0,0          ; vbroadcastss  0x67d(%rip),%ymm10        # 4ba4 <_sk_callback_hsw+0x490>
   DB  196,66,61,168,209                   ; vfmadd213ps   %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,13,115,6,0,0          ; vbroadcastss  0x673(%rip),%ymm9        # 4bac <_sk_callback_hsw+0x494>
+  DB  196,98,125,24,13,115,6,0,0          ; vbroadcastss  0x673(%rip),%ymm9        # 4ba8 <_sk_callback_hsw+0x494>
   DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
-  DB  196,98,125,24,21,105,6,0,0          ; vbroadcastss  0x669(%rip),%ymm10        # 4bb0 <_sk_callback_hsw+0x498>
+  DB  196,98,125,24,21,105,6,0,0          ; vbroadcastss  0x669(%rip),%ymm10        # 4bac <_sk_callback_hsw+0x498>
   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
@@ -4212,14 +4211,14 @@ _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  196,98,125,24,5,81,6,0,0            ; vbroadcastss  0x651(%rip),%ymm8        # 4bb4 <_sk_callback_hsw+0x49c>
+  DB  196,98,125,24,5,81,6,0,0            ; vbroadcastss  0x651(%rip),%ymm8        # 4bb0 <_sk_callback_hsw+0x49c>
   DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
   DB  197,124,16,72,64                    ; vmovups       0x40(%rax),%ymm9
-  DB  196,98,125,24,21,67,6,0,0           ; vbroadcastss  0x643(%rip),%ymm10        # 4bb8 <_sk_callback_hsw+0x4a0>
-  DB  196,98,125,24,29,62,6,0,0           ; vbroadcastss  0x63e(%rip),%ymm11        # 4bbc <_sk_callback_hsw+0x4a4>
+  DB  196,98,125,24,21,67,6,0,0           ; vbroadcastss  0x643(%rip),%ymm10        # 4bb4 <_sk_callback_hsw+0x4a0>
+  DB  196,98,125,24,29,62,6,0,0           ; vbroadcastss  0x63e(%rip),%ymm11        # 4bb8 <_sk_callback_hsw+0x4a4>
   DB  196,66,53,168,218                   ; vfmadd213ps   %ymm10,%ymm9,%ymm11
   DB  196,66,53,168,216                   ; vfmadd213ps   %ymm8,%ymm9,%ymm11
-  DB  196,98,125,24,5,47,6,0,0            ; vbroadcastss  0x62f(%rip),%ymm8        # 4bc0 <_sk_callback_hsw+0x4a8>
+  DB  196,98,125,24,5,47,6,0,0            ; vbroadcastss  0x62f(%rip),%ymm8        # 4bbc <_sk_callback_hsw+0x4a8>
   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
@@ -4228,12 +4227,12 @@ _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  196,226,125,24,5,23,6,0,0           ; vbroadcastss  0x617(%rip),%ymm0        # 4bc4 <_sk_callback_hsw+0x4ac>
+  DB  196,226,125,24,5,23,6,0,0           ; vbroadcastss  0x617(%rip),%ymm0        # 4bc0 <_sk_callback_hsw+0x4ac>
   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  196,98,125,24,21,4,6,0,0            ; vbroadcastss  0x604(%rip),%ymm10        # 4bc8 <_sk_callback_hsw+0x4b0>
-  DB  196,98,125,24,29,255,5,0,0          ; vbroadcastss  0x5ff(%rip),%ymm11        # 4bcc <_sk_callback_hsw+0x4b4>
+  DB  196,98,125,24,21,4,6,0,0            ; vbroadcastss  0x604(%rip),%ymm10        # 4bc4 <_sk_callback_hsw+0x4b0>
+  DB  196,98,125,24,29,255,5,0,0          ; vbroadcastss  0x5ff(%rip),%ymm11        # 4bc8 <_sk_callback_hsw+0x4b4>
   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)
@@ -4243,13 +4242,13 @@ _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  196,226,125,24,13,226,5,0,0         ; vbroadcastss  0x5e2(%rip),%ymm1        # 4bd0 <_sk_callback_hsw+0x4b8>
+  DB  196,226,125,24,13,226,5,0,0         ; vbroadcastss  0x5e2(%rip),%ymm1        # 4bcc <_sk_callback_hsw+0x4b8>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,216,5,0,0           ; vbroadcastss  0x5d8(%rip),%ymm8        # 4bd4 <_sk_callback_hsw+0x4bc>
+  DB  196,98,125,24,5,216,5,0,0           ; vbroadcastss  0x5d8(%rip),%ymm8        # 4bd0 <_sk_callback_hsw+0x4bc>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,201,5,0,0          ; vbroadcastss  0x5c9(%rip),%ymm10        # 4bd8 <_sk_callback_hsw+0x4c0>
-  DB  196,98,125,24,29,196,5,0,0          ; vbroadcastss  0x5c4(%rip),%ymm11        # 4bdc <_sk_callback_hsw+0x4c4>
+  DB  196,98,125,24,21,201,5,0,0          ; vbroadcastss  0x5c9(%rip),%ymm10        # 4bd4 <_sk_callback_hsw+0x4c0>
+  DB  196,98,125,24,29,196,5,0,0          ; vbroadcastss  0x5c4(%rip),%ymm11        # 4bd8 <_sk_callback_hsw+0x4c4>
   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)
@@ -4259,16 +4258,16 @@ _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  196,226,125,24,13,167,5,0,0         ; vbroadcastss  0x5a7(%rip),%ymm1        # 4be0 <_sk_callback_hsw+0x4c8>
+  DB  196,226,125,24,13,167,5,0,0         ; vbroadcastss  0x5a7(%rip),%ymm1        # 4bdc <_sk_callback_hsw+0x4c8>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,157,5,0,0           ; vbroadcastss  0x59d(%rip),%ymm8        # 4be4 <_sk_callback_hsw+0x4cc>
+  DB  196,98,125,24,5,157,5,0,0           ; vbroadcastss  0x59d(%rip),%ymm8        # 4be0 <_sk_callback_hsw+0x4cc>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,147,5,0,0          ; vbroadcastss  0x593(%rip),%ymm9        # 4be8 <_sk_callback_hsw+0x4d0>
-  DB  196,98,125,24,21,142,5,0,0          ; vbroadcastss  0x58e(%rip),%ymm10        # 4bec <_sk_callback_hsw+0x4d4>
+  DB  196,98,125,24,13,147,5,0,0          ; vbroadcastss  0x593(%rip),%ymm9        # 4be4 <_sk_callback_hsw+0x4d0>
+  DB  196,98,125,24,21,142,5,0,0          ; vbroadcastss  0x58e(%rip),%ymm10        # 4be8 <_sk_callback_hsw+0x4d4>
   DB  196,66,61,168,209                   ; vfmadd213ps   %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,13,132,5,0,0          ; vbroadcastss  0x584(%rip),%ymm9        # 4bf0 <_sk_callback_hsw+0x4d8>
+  DB  196,98,125,24,13,132,5,0,0          ; vbroadcastss  0x584(%rip),%ymm9        # 4bec <_sk_callback_hsw+0x4d8>
   DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
-  DB  196,98,125,24,21,122,5,0,0          ; vbroadcastss  0x57a(%rip),%ymm10        # 4bf4 <_sk_callback_hsw+0x4dc>
+  DB  196,98,125,24,21,122,5,0,0          ; vbroadcastss  0x57a(%rip),%ymm10        # 4bf0 <_sk_callback_hsw+0x4dc>
   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
@@ -4277,14 +4276,14 @@ _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  196,98,125,24,5,98,5,0,0            ; vbroadcastss  0x562(%rip),%ymm8        # 4bf8 <_sk_callback_hsw+0x4e0>
+  DB  196,98,125,24,5,98,5,0,0            ; vbroadcastss  0x562(%rip),%ymm8        # 4bf4 <_sk_callback_hsw+0x4e0>
   DB  197,188,88,72,32                    ; vaddps        0x20(%rax),%ymm8,%ymm1
   DB  197,124,16,72,96                    ; vmovups       0x60(%rax),%ymm9
-  DB  196,98,125,24,21,83,5,0,0           ; vbroadcastss  0x553(%rip),%ymm10        # 4bfc <_sk_callback_hsw+0x4e4>
-  DB  196,98,125,24,29,78,5,0,0           ; vbroadcastss  0x54e(%rip),%ymm11        # 4c00 <_sk_callback_hsw+0x4e8>
+  DB  196,98,125,24,21,83,5,0,0           ; vbroadcastss  0x553(%rip),%ymm10        # 4bf8 <_sk_callback_hsw+0x4e4>
+  DB  196,98,125,24,29,78,5,0,0           ; vbroadcastss  0x54e(%rip),%ymm11        # 4bfc <_sk_callback_hsw+0x4e8>
   DB  196,66,53,168,218                   ; vfmadd213ps   %ymm10,%ymm9,%ymm11
   DB  196,66,53,168,216                   ; vfmadd213ps   %ymm8,%ymm9,%ymm11
-  DB  196,98,125,24,5,63,5,0,0            ; vbroadcastss  0x53f(%rip),%ymm8        # 4c04 <_sk_callback_hsw+0x4ec>
+  DB  196,98,125,24,5,63,5,0,0            ; vbroadcastss  0x53f(%rip),%ymm8        # 4c00 <_sk_callback_hsw+0x4ec>
   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
@@ -4293,12 +4292,12 @@ _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  196,226,125,24,13,39,5,0,0          ; vbroadcastss  0x527(%rip),%ymm1        # 4c08 <_sk_callback_hsw+0x4f0>
+  DB  196,226,125,24,13,39,5,0,0          ; vbroadcastss  0x527(%rip),%ymm1        # 4c04 <_sk_callback_hsw+0x4f0>
   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  196,98,125,24,21,19,5,0,0           ; vbroadcastss  0x513(%rip),%ymm10        # 4c0c <_sk_callback_hsw+0x4f4>
-  DB  196,98,125,24,29,14,5,0,0           ; vbroadcastss  0x50e(%rip),%ymm11        # 4c10 <_sk_callback_hsw+0x4f8>
+  DB  196,98,125,24,21,19,5,0,0           ; vbroadcastss  0x513(%rip),%ymm10        # 4c08 <_sk_callback_hsw+0x4f4>
+  DB  196,98,125,24,29,14,5,0,0           ; vbroadcastss  0x50e(%rip),%ymm11        # 4c0c <_sk_callback_hsw+0x4f8>
   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)
@@ -4412,25 +4411,25 @@ 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 48e5 <.literal4+0xb1>
+  DB  71,225,61                           ; rex.RXB       loope 48e1 <.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 48f5 <.literal4+0xc1>
+  DB  71,225,61                           ; rex.RXB       loope 48f1 <.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 4905 <.literal4+0xd1>
+  DB  71,225,61                           ; rex.RXB       loope 4901 <.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 4915 <.literal4+0xe1>
+  DB  71,225,61                           ; rex.RXB       loope 4911 <.literal4+0xe1>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -4480,7 +4479,7 @@ 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        4965 <.literal4+0x131>
+  DB  224,7                               ; loopne        4961 <.literal4+0x131>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -4496,10 +4495,10 @@ 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            498c <.literal4+0x158>
+  DB  127,0                               ; jg            4988 <.literal4+0x158>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4a05 <.literal4+0x1d1>
+  DB  119,115                             ; ja            4a01 <.literal4+0x1d1>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4513,10 +4512,10 @@ 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            49c0 <.literal4+0x18c>
+  DB  127,0                               ; jg            49bc <.literal4+0x18c>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4a39 <.literal4+0x205>
+  DB  119,115                             ; ja            4a35 <.literal4+0x205>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4530,10 +4529,10 @@ 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            49f4 <.literal4+0x1c0>
+  DB  127,0                               ; jg            49f0 <.literal4+0x1c0>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4a6d <.literal4+0x239>
+  DB  119,115                             ; ja            4a69 <.literal4+0x239>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4547,10 +4546,10 @@ 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            4a28 <.literal4+0x1f4>
+  DB  127,0                               ; jg            4a24 <.literal4+0x1f4>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4aa1 <.literal4+0x26d>
+  DB  119,115                             ; ja            4a9d <.literal4+0x26d>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4563,7 +4562,7 @@ 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            4a9f <.literal4+0x26b>
+  DB  127,67                              ; jg            4a9b <.literal4+0x26b>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,195                               ; add           %al,%bl
   DB  0,0                                 ; add           %al,(%rax)
@@ -4575,10 +4574,10 @@ ALIGN 4
   DB  190,80,128,3,62                     ; mov           $0x3e038050,%esi
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           4abf <.literal4+0x28b>
+  DB  118,63                              ; jbe           4abb <.literal4+0x28b>
   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            4ad3 <.literal4+0x29f>
+  DB  127,67                              ; jg            4acf <.literal4+0x29f>
   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)
@@ -4587,7 +4586,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        4ab5 <.literal4+0x281>
+  DB  224,7                               ; loopne        4ab1 <.literal4+0x281>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -4599,7 +4598,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        4ad1 <.literal4+0x29d>
+  DB  224,7                               ; loopne        4acd <.literal4+0x29d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -4610,7 +4609,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            4b26 <.literal4+0x2f2>
+  DB  124,66                              ; jl            4b22 <.literal4+0x2f2>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  137,136,136,55,0,15                 ; mov           %ecx,0xf003788(%rax)
@@ -4628,9 +4627,9 @@ 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            4b69 <.literal4+0x335>
+  DB  112,65                              ; jo            4b65 <.literal4+0x335>
   DB  129,128,128,59,129,128,128,59,0,0   ; addl          $0x3b80,-0x7f7ec480(%rax)
-  DB  127,67                              ; jg            4b77 <.literal4+0x343>
+  DB  127,67                              ; jg            4b73 <.literal4+0x343>
   DB  128,0,128                           ; addb          $0x80,(%rax)
   DB  55                                  ; (bad)
   DB  128,0,128                           ; addb          $0x80,(%rax)
@@ -4638,7 +4637,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  255                                 ; (bad)
-  DB  127,71                              ; jg            4b8b <.literal4+0x357>
+  DB  127,71                              ; jg            4b87 <.literal4+0x357>
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,89                               ; ds            pop %rcx
@@ -4738,16 +4737,16 @@ 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)        # a004c48 <_sk_callback_hsw+0xa000530>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004c48 <_sk_callback_hsw+0xa000534>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004c50 <_sk_callback_hsw+0x12000538>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004c50 <_sk_callback_hsw+0x1200053c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004c58 <_sk_callback_hsw+0x1a000540>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004c58 <_sk_callback_hsw+0x1a000544>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004c60 <_sk_callback_hsw+0x3000548>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004c60 <_sk_callback_hsw+0x300054c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -4790,16 +4789,16 @@ 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)        # a004ca8 <_sk_callback_hsw+0xa000590>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004ca8 <_sk_callback_hsw+0xa000594>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004cb0 <_sk_callback_hsw+0x12000598>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004cb0 <_sk_callback_hsw+0x1200059c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004cb8 <_sk_callback_hsw+0x1a0005a0>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004cb8 <_sk_callback_hsw+0x1a0005a4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004cc0 <_sk_callback_hsw+0x30005a8>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004cc0 <_sk_callback_hsw+0x30005ac>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -4842,16 +4841,16 @@ 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)        # a004d08 <_sk_callback_hsw+0xa0005f0>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004d08 <_sk_callback_hsw+0xa0005f4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004d10 <_sk_callback_hsw+0x120005f8>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004d10 <_sk_callback_hsw+0x120005fc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004d18 <_sk_callback_hsw+0x1a000600>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004d18 <_sk_callback_hsw+0x1a000604>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d20 <_sk_callback_hsw+0x3000608>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d20 <_sk_callback_hsw+0x300060c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -4894,16 +4893,16 @@ 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)        # a004d68 <_sk_callback_hsw+0xa000650>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004d68 <_sk_callback_hsw+0xa000654>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004d70 <_sk_callback_hsw+0x12000658>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004d70 <_sk_callback_hsw+0x1200065c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004d78 <_sk_callback_hsw+0x1a000660>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004d78 <_sk_callback_hsw+0x1a000664>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d80 <_sk_callback_hsw+0x3000668>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d80 <_sk_callback_hsw+0x300066c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -5045,14 +5044,14 @@ _sk_seed_shader_avx LABEL PROC
   DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
   DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,224,98,0,0        ; vbroadcastss  0x62e0(%rip),%ymm1        # 6440 <_sk_callback_avx+0x11a>
+  DB  196,226,125,24,13,220,98,0,0        ; vbroadcastss  0x62dc(%rip),%ymm1        # 643c <_sk_callback_avx+0x11a>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  197,252,88,2                        ; vaddps        (%rdx),%ymm0,%ymm0
   DB  196,226,125,24,16                   ; vbroadcastss  (%rax),%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  197,236,88,201                      ; vaddps        %ymm1,%ymm2,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,21,196,98,0,0        ; vbroadcastss  0x62c4(%rip),%ymm2        # 6444 <_sk_callback_avx+0x11e>
+  DB  196,226,125,24,21,192,98,0,0        ; vbroadcastss  0x62c0(%rip),%ymm2        # 6440 <_sk_callback_avx+0x11e>
   DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
   DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
   DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
@@ -5072,7 +5071,7 @@ _sk_dither_avx LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  196,66,125,24,8                     ; vbroadcastss  (%r8),%ymm9
   DB  196,65,60,87,209                    ; vxorps        %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,29,124,98,0,0         ; vbroadcastss  0x627c(%rip),%ymm11        # 6448 <_sk_callback_avx+0x122>
+  DB  196,98,125,24,29,120,98,0,0         ; vbroadcastss  0x6278(%rip),%ymm11        # 6444 <_sk_callback_avx+0x122>
   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
@@ -5083,8 +5082,8 @@ _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  196,98,125,24,37,61,98,0,0          ; vbroadcastss  0x623d(%rip),%ymm12        # 644c <_sk_callback_avx+0x126>
-  DB  196,98,125,24,45,56,98,0,0          ; vbroadcastss  0x6238(%rip),%ymm13        # 6450 <_sk_callback_avx+0x12a>
+  DB  196,98,125,24,37,57,98,0,0          ; vbroadcastss  0x6239(%rip),%ymm12        # 6448 <_sk_callback_avx+0x126>
+  DB  196,98,125,24,45,52,98,0,0          ; vbroadcastss  0x6234(%rip),%ymm13        # 644c <_sk_callback_avx+0x12a>
   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
@@ -5111,9 +5110,9 @@ _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  196,98,125,24,13,163,97,0,0         ; vbroadcastss  0x61a3(%rip),%ymm9        # 6454 <_sk_callback_avx+0x12e>
+  DB  196,98,125,24,13,159,97,0,0         ; vbroadcastss  0x619f(%rip),%ymm9        # 6450 <_sk_callback_avx+0x12e>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,153,97,0,0         ; vbroadcastss  0x6199(%rip),%ymm9        # 6458 <_sk_callback_avx+0x132>
+  DB  196,98,125,24,13,149,97,0,0         ; vbroadcastss  0x6195(%rip),%ymm9        # 6454 <_sk_callback_avx+0x132>
   DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
   DB  196,98,125,24,72,8                  ; vbroadcastss  0x8(%rax),%ymm9
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
@@ -5165,7 +5164,7 @@ _sk_clear_avx LABEL PROC
 PUBLIC _sk_srcatop_avx
 _sk_srcatop_avx LABEL PROC
   DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
-  DB  196,98,125,24,5,13,97,0,0           ; vbroadcastss  0x610d(%rip),%ymm8        # 645c <_sk_callback_avx+0x136>
+  DB  196,98,125,24,5,9,97,0,0            ; vbroadcastss  0x6109(%rip),%ymm8        # 6458 <_sk_callback_avx+0x136>
   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
@@ -5184,7 +5183,7 @@ _sk_srcatop_avx LABEL PROC
 PUBLIC _sk_dstatop_avx
 _sk_dstatop_avx LABEL PROC
   DB  197,100,89,196                      ; vmulps        %ymm4,%ymm3,%ymm8
-  DB  196,98,125,24,13,207,96,0,0         ; vbroadcastss  0x60cf(%rip),%ymm9        # 6460 <_sk_callback_avx+0x13a>
+  DB  196,98,125,24,13,203,96,0,0         ; vbroadcastss  0x60cb(%rip),%ymm9        # 645c <_sk_callback_avx+0x13a>
   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
@@ -5220,7 +5219,7 @@ _sk_dstin_avx LABEL PROC
 
 PUBLIC _sk_srcout_avx
 _sk_srcout_avx LABEL PROC
-  DB  196,98,125,24,5,110,96,0,0          ; vbroadcastss  0x606e(%rip),%ymm8        # 6464 <_sk_callback_avx+0x13e>
+  DB  196,98,125,24,5,106,96,0,0          ; vbroadcastss  0x606a(%rip),%ymm8        # 6460 <_sk_callback_avx+0x13e>
   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
@@ -5231,7 +5230,7 @@ _sk_srcout_avx LABEL PROC
 
 PUBLIC _sk_dstout_avx
 _sk_dstout_avx LABEL PROC
-  DB  196,226,125,24,5,81,96,0,0          ; vbroadcastss  0x6051(%rip),%ymm0        # 6468 <_sk_callback_avx+0x142>
+  DB  196,226,125,24,5,77,96,0,0          ; vbroadcastss  0x604d(%rip),%ymm0        # 6464 <_sk_callback_avx+0x142>
   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
@@ -5242,7 +5241,7 @@ _sk_dstout_avx LABEL PROC
 
 PUBLIC _sk_srcover_avx
 _sk_srcover_avx LABEL PROC
-  DB  196,98,125,24,5,52,96,0,0           ; vbroadcastss  0x6034(%rip),%ymm8        # 646c <_sk_callback_avx+0x146>
+  DB  196,98,125,24,5,48,96,0,0           ; vbroadcastss  0x6030(%rip),%ymm8        # 6468 <_sk_callback_avx+0x146>
   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
@@ -5257,7 +5256,7 @@ _sk_srcover_avx LABEL PROC
 
 PUBLIC _sk_dstover_avx
 _sk_dstover_avx LABEL PROC
-  DB  196,98,125,24,5,7,96,0,0            ; vbroadcastss  0x6007(%rip),%ymm8        # 6470 <_sk_callback_avx+0x14a>
+  DB  196,98,125,24,5,3,96,0,0            ; vbroadcastss  0x6003(%rip),%ymm8        # 646c <_sk_callback_avx+0x14a>
   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
@@ -5281,7 +5280,7 @@ _sk_modulate_avx LABEL PROC
 
 PUBLIC _sk_multiply_avx
 _sk_multiply_avx LABEL PROC
-  DB  196,98,125,24,5,198,95,0,0          ; vbroadcastss  0x5fc6(%rip),%ymm8        # 6474 <_sk_callback_avx+0x14e>
+  DB  196,98,125,24,5,194,95,0,0          ; vbroadcastss  0x5fc2(%rip),%ymm8        # 6470 <_sk_callback_avx+0x14e>
   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
@@ -5335,7 +5334,7 @@ _sk_screen_avx LABEL PROC
 
 PUBLIC _sk_xor__avx
 _sk_xor__avx LABEL PROC
-  DB  196,98,125,24,5,21,95,0,0           ; vbroadcastss  0x5f15(%rip),%ymm8        # 6478 <_sk_callback_avx+0x152>
+  DB  196,98,125,24,5,17,95,0,0           ; vbroadcastss  0x5f11(%rip),%ymm8        # 6474 <_sk_callback_avx+0x152>
   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
@@ -5370,7 +5369,7 @@ _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  196,98,125,24,5,149,94,0,0          ; vbroadcastss  0x5e95(%rip),%ymm8        # 647c <_sk_callback_avx+0x156>
+  DB  196,98,125,24,5,145,94,0,0          ; vbroadcastss  0x5e91(%rip),%ymm8        # 6478 <_sk_callback_avx+0x156>
   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
@@ -5394,7 +5393,7 @@ _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  196,98,125,24,5,65,94,0,0           ; vbroadcastss  0x5e41(%rip),%ymm8        # 6480 <_sk_callback_avx+0x15a>
+  DB  196,98,125,24,5,61,94,0,0           ; vbroadcastss  0x5e3d(%rip),%ymm8        # 647c <_sk_callback_avx+0x15a>
   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
@@ -5421,7 +5420,7 @@ _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  196,98,125,24,5,225,93,0,0          ; vbroadcastss  0x5de1(%rip),%ymm8        # 6484 <_sk_callback_avx+0x15e>
+  DB  196,98,125,24,5,221,93,0,0          ; vbroadcastss  0x5ddd(%rip),%ymm8        # 6480 <_sk_callback_avx+0x15e>
   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
@@ -5442,7 +5441,7 @@ _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  196,98,125,24,5,156,93,0,0          ; vbroadcastss  0x5d9c(%rip),%ymm8        # 6488 <_sk_callback_avx+0x162>
+  DB  196,98,125,24,5,152,93,0,0          ; vbroadcastss  0x5d98(%rip),%ymm8        # 6484 <_sk_callback_avx+0x162>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
   DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
@@ -5451,7 +5450,7 @@ _sk_exclusion_avx LABEL PROC
 
 PUBLIC _sk_colorburn_avx
 _sk_colorburn_avx LABEL PROC
-  DB  196,98,125,24,5,135,93,0,0          ; vbroadcastss  0x5d87(%rip),%ymm8        # 648c <_sk_callback_avx+0x166>
+  DB  196,98,125,24,5,131,93,0,0          ; vbroadcastss  0x5d83(%rip),%ymm8        # 6488 <_sk_callback_avx+0x166>
   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
@@ -5511,7 +5510,7 @@ _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  196,98,125,24,13,131,92,0,0         ; vbroadcastss  0x5c83(%rip),%ymm9        # 6490 <_sk_callback_avx+0x16a>
+  DB  196,98,125,24,13,127,92,0,0         ; vbroadcastss  0x5c7f(%rip),%ymm9        # 648c <_sk_callback_avx+0x16a>
   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
@@ -5566,7 +5565,7 @@ _sk_colordodge_avx LABEL PROC
 
 PUBLIC _sk_hardlight_avx
 _sk_hardlight_avx LABEL PROC
-  DB  196,98,125,24,5,149,91,0,0          ; vbroadcastss  0x5b95(%rip),%ymm8        # 6494 <_sk_callback_avx+0x16e>
+  DB  196,98,125,24,5,145,91,0,0          ; vbroadcastss  0x5b91(%rip),%ymm8        # 6490 <_sk_callback_avx+0x16e>
   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
@@ -5619,7 +5618,7 @@ _sk_hardlight_avx LABEL PROC
 
 PUBLIC _sk_overlay_avx
 _sk_overlay_avx LABEL PROC
-  DB  196,98,125,24,5,190,90,0,0          ; vbroadcastss  0x5abe(%rip),%ymm8        # 6498 <_sk_callback_avx+0x172>
+  DB  196,98,125,24,5,186,90,0,0          ; vbroadcastss  0x5aba(%rip),%ymm8        # 6494 <_sk_callback_avx+0x172>
   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
@@ -5684,10 +5683,10 @@ _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  196,98,125,24,29,177,89,0,0         ; vbroadcastss  0x59b1(%rip),%ymm11        # 64a0 <_sk_callback_avx+0x17a>
+  DB  196,98,125,24,29,173,89,0,0         ; vbroadcastss  0x59ad(%rip),%ymm11        # 649c <_sk_callback_avx+0x17a>
   DB  196,65,28,88,235                    ; vaddps        %ymm11,%ymm12,%ymm13
   DB  196,65,20,89,192                    ; vmulps        %ymm8,%ymm13,%ymm8
-  DB  196,98,125,24,45,162,89,0,0         ; vbroadcastss  0x59a2(%rip),%ymm13        # 64a4 <_sk_callback_avx+0x17e>
+  DB  196,98,125,24,45,158,89,0,0         ; vbroadcastss  0x599e(%rip),%ymm13        # 64a0 <_sk_callback_avx+0x17e>
   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
@@ -5698,7 +5697,7 @@ _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  196,98,125,24,5,96,89,0,0           ; vbroadcastss  0x5960(%rip),%ymm8        # 649c <_sk_callback_avx+0x176>
+  DB  196,98,125,24,5,92,89,0,0           ; vbroadcastss  0x595c(%rip),%ymm8        # 6498 <_sk_callback_avx+0x176>
   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
@@ -5825,12 +5824,12 @@ _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  196,98,125,24,53,42,87,0,0          ; vbroadcastss  0x572a(%rip),%ymm14        # 64a8 <_sk_callback_avx+0x182>
+  DB  196,98,125,24,53,38,87,0,0          ; vbroadcastss  0x5726(%rip),%ymm14        # 64a4 <_sk_callback_avx+0x182>
   DB  196,65,92,89,222                    ; vmulps        %ymm14,%ymm4,%ymm11
-  DB  196,98,125,24,61,32,87,0,0          ; vbroadcastss  0x5720(%rip),%ymm15        # 64ac <_sk_callback_avx+0x186>
+  DB  196,98,125,24,61,28,87,0,0          ; vbroadcastss  0x571c(%rip),%ymm15        # 64a8 <_sk_callback_avx+0x186>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
-  DB  196,226,125,24,5,17,87,0,0          ; vbroadcastss  0x5711(%rip),%ymm0        # 64b0 <_sk_callback_avx+0x18a>
+  DB  196,226,125,24,5,13,87,0,0          ; vbroadcastss  0x570d(%rip),%ymm0        # 64ac <_sk_callback_avx+0x18a>
   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
@@ -5891,7 +5890,7 @@ _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  196,98,125,24,5,234,85,0,0          ; vbroadcastss  0x55ea(%rip),%ymm8        # 64b4 <_sk_callback_avx+0x18e>
+  DB  196,98,125,24,5,230,85,0,0          ; vbroadcastss  0x55e6(%rip),%ymm8        # 64b0 <_sk_callback_avx+0x18e>
   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
@@ -5948,12 +5947,12 @@ _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  196,98,125,24,53,242,84,0,0         ; vbroadcastss  0x54f2(%rip),%ymm14        # 64b8 <_sk_callback_avx+0x192>
+  DB  196,98,125,24,53,238,84,0,0         ; vbroadcastss  0x54ee(%rip),%ymm14        # 64b4 <_sk_callback_avx+0x192>
   DB  196,65,92,89,222                    ; vmulps        %ymm14,%ymm4,%ymm11
-  DB  196,98,125,24,61,232,84,0,0         ; vbroadcastss  0x54e8(%rip),%ymm15        # 64bc <_sk_callback_avx+0x196>
+  DB  196,98,125,24,61,228,84,0,0         ; vbroadcastss  0x54e4(%rip),%ymm15        # 64b8 <_sk_callback_avx+0x196>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
-  DB  196,226,125,24,5,217,84,0,0         ; vbroadcastss  0x54d9(%rip),%ymm0        # 64c0 <_sk_callback_avx+0x19a>
+  DB  196,226,125,24,5,213,84,0,0         ; vbroadcastss  0x54d5(%rip),%ymm0        # 64bc <_sk_callback_avx+0x19a>
   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
@@ -6014,7 +6013,7 @@ _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  196,98,125,24,5,178,83,0,0          ; vbroadcastss  0x53b2(%rip),%ymm8        # 64c4 <_sk_callback_avx+0x19e>
+  DB  196,98,125,24,5,174,83,0,0          ; vbroadcastss  0x53ae(%rip),%ymm8        # 64c0 <_sk_callback_avx+0x19e>
   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
@@ -6043,12 +6042,12 @@ _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  196,98,125,24,45,66,83,0,0          ; vbroadcastss  0x5342(%rip),%ymm13        # 64c8 <_sk_callback_avx+0x1a2>
+  DB  196,98,125,24,45,62,83,0,0          ; vbroadcastss  0x533e(%rip),%ymm13        # 64c4 <_sk_callback_avx+0x1a2>
   DB  196,65,92,89,213                    ; vmulps        %ymm13,%ymm4,%ymm10
-  DB  196,98,125,24,53,56,83,0,0          ; vbroadcastss  0x5338(%rip),%ymm14        # 64cc <_sk_callback_avx+0x1a6>
+  DB  196,98,125,24,53,52,83,0,0          ; vbroadcastss  0x5334(%rip),%ymm14        # 64c8 <_sk_callback_avx+0x1a6>
   DB  196,65,84,89,222                    ; vmulps        %ymm14,%ymm5,%ymm11
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,61,41,83,0,0          ; vbroadcastss  0x5329(%rip),%ymm15        # 64d0 <_sk_callback_avx+0x1aa>
+  DB  196,98,125,24,61,37,83,0,0          ; vbroadcastss  0x5325(%rip),%ymm15        # 64cc <_sk_callback_avx+0x1aa>
   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
@@ -6111,7 +6110,7 @@ _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  196,226,125,24,5,240,81,0,0         ; vbroadcastss  0x51f0(%rip),%ymm0        # 64d4 <_sk_callback_avx+0x1ae>
+  DB  196,226,125,24,5,236,81,0,0         ; vbroadcastss  0x51ec(%rip),%ymm0        # 64d0 <_sk_callback_avx+0x1ae>
   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
@@ -6141,12 +6140,12 @@ _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  196,98,125,24,45,124,81,0,0         ; vbroadcastss  0x517c(%rip),%ymm13        # 64d8 <_sk_callback_avx+0x1b2>
+  DB  196,98,125,24,45,120,81,0,0         ; vbroadcastss  0x5178(%rip),%ymm13        # 64d4 <_sk_callback_avx+0x1b2>
   DB  196,65,108,89,213                   ; vmulps        %ymm13,%ymm2,%ymm10
-  DB  196,98,125,24,53,114,81,0,0         ; vbroadcastss  0x5172(%rip),%ymm14        # 64dc <_sk_callback_avx+0x1b6>
+  DB  196,98,125,24,53,110,81,0,0         ; vbroadcastss  0x516e(%rip),%ymm14        # 64d8 <_sk_callback_avx+0x1b6>
   DB  196,65,116,89,222                   ; vmulps        %ymm14,%ymm1,%ymm11
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,61,99,81,0,0          ; vbroadcastss  0x5163(%rip),%ymm15        # 64e0 <_sk_callback_avx+0x1ba>
+  DB  196,98,125,24,61,95,81,0,0          ; vbroadcastss  0x515f(%rip),%ymm15        # 64dc <_sk_callback_avx+0x1ba>
   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
@@ -6209,7 +6208,7 @@ _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  196,226,125,24,5,42,80,0,0          ; vbroadcastss  0x502a(%rip),%ymm0        # 64e4 <_sk_callback_avx+0x1be>
+  DB  196,226,125,24,5,38,80,0,0          ; vbroadcastss  0x5026(%rip),%ymm0        # 64e0 <_sk_callback_avx+0x1be>
   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
@@ -6242,7 +6241,7 @@ _sk_clamp_0_avx LABEL PROC
 
 PUBLIC _sk_clamp_1_avx
 _sk_clamp_1_avx LABEL PROC
-  DB  196,98,125,24,5,186,79,0,0          ; vbroadcastss  0x4fba(%rip),%ymm8        # 64e8 <_sk_callback_avx+0x1c2>
+  DB  196,98,125,24,5,182,79,0,0          ; vbroadcastss  0x4fb6(%rip),%ymm8        # 64e4 <_sk_callback_avx+0x1c2>
   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
@@ -6252,7 +6251,7 @@ _sk_clamp_1_avx LABEL PROC
 
 PUBLIC _sk_clamp_a_avx
 _sk_clamp_a_avx LABEL PROC
-  DB  196,98,125,24,5,157,79,0,0          ; vbroadcastss  0x4f9d(%rip),%ymm8        # 64ec <_sk_callback_avx+0x1c6>
+  DB  196,98,125,24,5,153,79,0,0          ; vbroadcastss  0x4f99(%rip),%ymm8        # 64e8 <_sk_callback_avx+0x1c6>
   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
@@ -6324,7 +6323,7 @@ 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  196,98,125,24,21,229,78,0,0         ; vbroadcastss  0x4ee5(%rip),%ymm10        # 64f0 <_sk_callback_avx+0x1ca>
+  DB  196,98,125,24,21,225,78,0,0         ; vbroadcastss  0x4ee1(%rip),%ymm10        # 64ec <_sk_callback_avx+0x1ca>
   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
@@ -6335,17 +6334,17 @@ _sk_unpremul_avx LABEL PROC
 
 PUBLIC _sk_from_srgb_avx
 _sk_from_srgb_avx LABEL PROC
-  DB  196,98,125,24,5,198,78,0,0          ; vbroadcastss  0x4ec6(%rip),%ymm8        # 64f4 <_sk_callback_avx+0x1ce>
+  DB  196,98,125,24,5,194,78,0,0          ; vbroadcastss  0x4ec2(%rip),%ymm8        # 64f0 <_sk_callback_avx+0x1ce>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
-  DB  196,98,125,24,29,184,78,0,0         ; vbroadcastss  0x4eb8(%rip),%ymm11        # 64f8 <_sk_callback_avx+0x1d2>
+  DB  196,98,125,24,29,180,78,0,0         ; vbroadcastss  0x4eb4(%rip),%ymm11        # 64f4 <_sk_callback_avx+0x1d2>
   DB  196,65,124,89,227                   ; vmulps        %ymm11,%ymm0,%ymm12
-  DB  196,98,125,24,45,174,78,0,0         ; vbroadcastss  0x4eae(%rip),%ymm13        # 64fc <_sk_callback_avx+0x1d6>
+  DB  196,98,125,24,45,170,78,0,0         ; vbroadcastss  0x4eaa(%rip),%ymm13        # 64f8 <_sk_callback_avx+0x1d6>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,44,89,212                    ; vmulps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,159,78,0,0         ; vbroadcastss  0x4e9f(%rip),%ymm12        # 6500 <_sk_callback_avx+0x1da>
+  DB  196,98,125,24,37,155,78,0,0         ; vbroadcastss  0x4e9b(%rip),%ymm12        # 64fc <_sk_callback_avx+0x1da>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,53,149,78,0,0         ; vbroadcastss  0x4e95(%rip),%ymm14        # 6504 <_sk_callback_avx+0x1de>
+  DB  196,98,125,24,53,145,78,0,0         ; vbroadcastss  0x4e91(%rip),%ymm14        # 6500 <_sk_callback_avx+0x1de>
   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
@@ -6372,18 +6371,18 @@ _sk_to_srgb_avx LABEL PROC
   DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
   DB  196,65,124,83,200                   ; vrcpps        %ymm8,%ymm9
   DB  196,65,124,82,208                   ; vrsqrtps      %ymm8,%ymm10
-  DB  196,98,125,24,5,32,78,0,0           ; vbroadcastss  0x4e20(%rip),%ymm8        # 6508 <_sk_callback_avx+0x1e2>
+  DB  196,98,125,24,5,28,78,0,0           ; vbroadcastss  0x4e1c(%rip),%ymm8        # 6504 <_sk_callback_avx+0x1e2>
   DB  196,65,124,89,216                   ; vmulps        %ymm8,%ymm0,%ymm11
-  DB  196,98,125,24,37,22,78,0,0          ; vbroadcastss  0x4e16(%rip),%ymm12        # 650c <_sk_callback_avx+0x1e6>
+  DB  196,98,125,24,37,18,78,0,0          ; vbroadcastss  0x4e12(%rip),%ymm12        # 6508 <_sk_callback_avx+0x1e6>
   DB  196,65,52,89,204                    ; vmulps        %ymm12,%ymm9,%ymm9
-  DB  196,98,125,24,45,12,78,0,0          ; vbroadcastss  0x4e0c(%rip),%ymm13        # 6510 <_sk_callback_avx+0x1ea>
+  DB  196,98,125,24,45,8,78,0,0           ; vbroadcastss  0x4e08(%rip),%ymm13        # 650c <_sk_callback_avx+0x1ea>
   DB  196,65,52,88,205                    ; vaddps        %ymm13,%ymm9,%ymm9
-  DB  196,98,125,24,53,2,78,0,0           ; vbroadcastss  0x4e02(%rip),%ymm14        # 6514 <_sk_callback_avx+0x1ee>
+  DB  196,98,125,24,53,254,77,0,0         ; vbroadcastss  0x4dfe(%rip),%ymm14        # 6510 <_sk_callback_avx+0x1ee>
   DB  196,65,44,89,214                    ; vmulps        %ymm14,%ymm10,%ymm10
   DB  196,65,44,88,201                    ; vaddps        %ymm9,%ymm10,%ymm9
-  DB  196,98,125,24,21,243,77,0,0         ; vbroadcastss  0x4df3(%rip),%ymm10        # 6518 <_sk_callback_avx+0x1f2>
+  DB  196,98,125,24,21,239,77,0,0         ; vbroadcastss  0x4def(%rip),%ymm10        # 6514 <_sk_callback_avx+0x1f2>
   DB  196,65,44,93,201                    ; vminps        %ymm9,%ymm10,%ymm9
-  DB  196,98,125,24,61,233,77,0,0         ; vbroadcastss  0x4de9(%rip),%ymm15        # 651c <_sk_callback_avx+0x1f6>
+  DB  196,98,125,24,61,229,77,0,0         ; vbroadcastss  0x4de5(%rip),%ymm15        # 6518 <_sk_callback_avx+0x1f6>
   DB  196,193,124,194,199,1               ; vcmpltps      %ymm15,%ymm0,%ymm0
   DB  196,195,53,74,195,0                 ; vblendvps     %ymm0,%ymm11,%ymm9,%ymm0
   DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
@@ -6418,7 +6417,7 @@ _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  196,98,125,24,29,79,77,0,0          ; vbroadcastss  0x4d4f(%rip),%ymm11        # 6520 <_sk_callback_avx+0x1fa>
+  DB  196,98,125,24,29,75,77,0,0          ; vbroadcastss  0x4d4b(%rip),%ymm11        # 651c <_sk_callback_avx+0x1fa>
   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
@@ -6428,19 +6427,19 @@ _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  196,98,125,24,29,40,77,0,0          ; vbroadcastss  0x4d28(%rip),%ymm11        # 652c <_sk_callback_avx+0x206>
+  DB  196,98,125,24,29,36,77,0,0          ; vbroadcastss  0x4d24(%rip),%ymm11        # 6528 <_sk_callback_avx+0x206>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,22,77,0,0          ; vbroadcastss  0x4d16(%rip),%ymm11        # 6528 <_sk_callback_avx+0x202>
+  DB  196,98,125,24,29,18,77,0,0          ; vbroadcastss  0x4d12(%rip),%ymm11        # 6524 <_sk_callback_avx+0x202>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
   DB  196,227,117,74,202,224              ; vblendvps     %ymm14,%ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,254,76,0,0        ; vbroadcastss  0x4cfe(%rip),%ymm2        # 6524 <_sk_callback_avx+0x1fe>
+  DB  196,226,125,24,21,250,76,0,0        ; vbroadcastss  0x4cfa(%rip),%ymm2        # 6520 <_sk_callback_avx+0x1fe>
   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  196,98,125,24,37,229,76,0,0         ; vbroadcastss  0x4ce5(%rip),%ymm12        # 6534 <_sk_callback_avx+0x20e>
+  DB  196,98,125,24,37,225,76,0,0         ; vbroadcastss  0x4ce1(%rip),%ymm12        # 6530 <_sk_callback_avx+0x20e>
   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
@@ -6450,7 +6449,7 @@ _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  196,98,125,24,5,168,76,0,0          ; vbroadcastss  0x4ca8(%rip),%ymm8        # 6530 <_sk_callback_avx+0x20a>
+  DB  196,98,125,24,5,164,76,0,0          ; vbroadcastss  0x4ca4(%rip),%ymm8        # 652c <_sk_callback_avx+0x20a>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -6465,7 +6464,7 @@ _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  196,98,125,24,5,111,76,0,0          ; vbroadcastss  0x4c6f(%rip),%ymm8        # 6538 <_sk_callback_avx+0x212>
+  DB  196,98,125,24,5,107,76,0,0          ; vbroadcastss  0x4c6b(%rip),%ymm8        # 6534 <_sk_callback_avx+0x212>
   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
@@ -6473,23 +6472,23 @@ _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  196,98,125,24,29,73,76,0,0          ; vbroadcastss  0x4c49(%rip),%ymm11        # 653c <_sk_callback_avx+0x216>
+  DB  196,98,125,24,29,69,76,0,0          ; vbroadcastss  0x4c45(%rip),%ymm11        # 6538 <_sk_callback_avx+0x216>
   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  196,98,125,24,37,51,76,0,0          ; vbroadcastss  0x4c33(%rip),%ymm12        # 6544 <_sk_callback_avx+0x21e>
+  DB  196,98,125,24,37,47,76,0,0          ; vbroadcastss  0x4c2f(%rip),%ymm12        # 6540 <_sk_callback_avx+0x21e>
   DB  196,193,4,89,196                    ; vmulps        %ymm12,%ymm15,%ymm0
-  DB  196,98,125,24,45,41,76,0,0          ; vbroadcastss  0x4c29(%rip),%ymm13        # 6548 <_sk_callback_avx+0x222>
+  DB  196,98,125,24,45,37,76,0,0          ; vbroadcastss  0x4c25(%rip),%ymm13        # 6544 <_sk_callback_avx+0x222>
   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  196,226,125,24,13,10,76,0,0         ; vbroadcastss  0x4c0a(%rip),%ymm1        # 6540 <_sk_callback_avx+0x21a>
+  DB  196,226,125,24,13,6,76,0,0          ; vbroadcastss  0x4c06(%rip),%ymm1        # 653c <_sk_callback_avx+0x21a>
   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  196,98,125,24,53,245,75,0,0         ; vbroadcastss  0x4bf5(%rip),%ymm14        # 654c <_sk_callback_avx+0x226>
+  DB  196,98,125,24,53,241,75,0,0         ; vbroadcastss  0x4bf1(%rip),%ymm14        # 6548 <_sk_callback_avx+0x226>
   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
@@ -6508,7 +6507,7 @@ _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  196,226,125,24,5,151,75,0,0         ; vbroadcastss  0x4b97(%rip),%ymm0        # 6550 <_sk_callback_avx+0x22a>
+  DB  196,226,125,24,5,147,75,0,0         ; vbroadcastss  0x4b93(%rip),%ymm0        # 654c <_sk_callback_avx+0x22a>
   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
@@ -6563,7 +6562,7 @@ _sk_scale_u8_avx LABEL PROC
   DB  196,66,121,49,192                   ; vpmovzxbd     %xmm8,%xmm8
   DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,186,74,0,0         ; vbroadcastss  0x4aba(%rip),%ymm9        # 6554 <_sk_callback_avx+0x22e>
+  DB  196,98,125,24,13,182,74,0,0         ; vbroadcastss  0x4ab6(%rip),%ymm9        # 6550 <_sk_callback_avx+0x22e>
   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
@@ -6618,7 +6617,7 @@ _sk_lerp_u8_avx LABEL PROC
   DB  196,66,121,49,192                   ; vpmovzxbd     %xmm8,%xmm8
   DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,6,74,0,0           ; vbroadcastss  0x4a06(%rip),%ymm9        # 6558 <_sk_callback_avx+0x232>
+  DB  196,98,125,24,13,2,74,0,0           ; vbroadcastss  0x4a02(%rip),%ymm9        # 6554 <_sk_callback_avx+0x232>
   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
@@ -6659,20 +6658,20 @@ _sk_lerp_565_avx LABEL PROC
   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  196,98,125,24,13,112,73,0,0         ; vbroadcastss  0x4970(%rip),%ymm9        # 655c <_sk_callback_avx+0x236>
+  DB  196,98,125,24,13,108,73,0,0         ; vbroadcastss  0x496c(%rip),%ymm9        # 6558 <_sk_callback_avx+0x236>
   DB  196,65,60,84,201                    ; vandps        %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,97,73,0,0          ; vbroadcastss  0x4961(%rip),%ymm10        # 6560 <_sk_callback_avx+0x23a>
+  DB  196,98,125,24,21,93,73,0,0          ; vbroadcastss  0x495d(%rip),%ymm10        # 655c <_sk_callback_avx+0x23a>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,21,87,73,0,0          ; vbroadcastss  0x4957(%rip),%ymm10        # 6564 <_sk_callback_avx+0x23e>
+  DB  196,98,125,24,21,83,73,0,0          ; vbroadcastss  0x4953(%rip),%ymm10        # 6560 <_sk_callback_avx+0x23e>
   DB  196,65,60,84,210                    ; vandps        %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,72,73,0,0          ; vbroadcastss  0x4948(%rip),%ymm11        # 6568 <_sk_callback_avx+0x242>
+  DB  196,98,125,24,29,68,73,0,0          ; vbroadcastss  0x4944(%rip),%ymm11        # 6564 <_sk_callback_avx+0x242>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,29,62,73,0,0          ; vbroadcastss  0x493e(%rip),%ymm11        # 656c <_sk_callback_avx+0x246>
+  DB  196,98,125,24,29,58,73,0,0          ; vbroadcastss  0x493a(%rip),%ymm11        # 6568 <_sk_callback_avx+0x246>
   DB  196,65,60,84,195                    ; vandps        %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,47,73,0,0          ; vbroadcastss  0x492f(%rip),%ymm11        # 6570 <_sk_callback_avx+0x24a>
+  DB  196,98,125,24,29,43,73,0,0          ; vbroadcastss  0x492b(%rip),%ymm11        # 656c <_sk_callback_avx+0x24a>
   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
@@ -6752,7 +6751,7 @@ _sk_load_tables_avx LABEL PROC
   DB  65,85                               ; push          %r13
   DB  65,84                               ; push          %r12
   DB  83                                  ; push          %rbx
-  DB  197,124,40,13,22,75,0,0             ; vmovaps       0x4b16(%rip),%ymm9        # 6860 <_sk_callback_avx+0x53a>
+  DB  197,124,40,13,22,75,0,0             ; vmovaps       0x4b16(%rip),%ymm9        # 6860 <_sk_callback_avx+0x53e>
   DB  196,193,60,84,193                   ; vandps        %ymm9,%ymm8,%ymm0
   DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
   DB  69,137,203                          ; mov           %r9d,%r11d
@@ -6844,7 +6843,7 @@ _sk_load_tables_avx LABEL PROC
   DB  196,193,97,114,210,24               ; vpsrld        $0x18,%xmm10,%xmm3
   DB  196,227,61,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,59,70,0,0           ; vbroadcastss  0x463b(%rip),%ymm8        # 6574 <_sk_callback_avx+0x24e>
+  DB  196,98,125,24,5,55,70,0,0           ; vbroadcastss  0x4637(%rip),%ymm8        # 6570 <_sk_callback_avx+0x24e>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -6934,7 +6933,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
   DB  197,177,108,208                     ; vpunpcklqdq   %xmm0,%xmm9,%xmm2
   DB  197,177,109,200                     ; vpunpckhqdq   %xmm0,%xmm9,%xmm1
   DB  196,65,57,108,212                   ; vpunpcklqdq   %xmm12,%xmm8,%xmm10
-  DB  197,121,111,29,86,72,0,0            ; vmovdqa       0x4856(%rip),%xmm11        # 68e0 <_sk_callback_avx+0x5ba>
+  DB  197,121,111,29,86,72,0,0            ; vmovdqa       0x4856(%rip),%xmm11        # 68e0 <_sk_callback_avx+0x5be>
   DB  196,193,105,219,195                 ; vpand         %xmm11,%xmm2,%xmm0
   DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
   DB  196,193,121,105,209                 ; vpunpckhwd    %xmm9,%xmm0,%xmm2
@@ -7033,7 +7032,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
   DB  196,226,121,51,219                  ; vpmovzxwd     %xmm3,%xmm3
   DB  196,195,101,24,216,1                ; vinsertf128   $0x1,%xmm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,236,66,0,0          ; vbroadcastss  0x42ec(%rip),%ymm8        # 6578 <_sk_callback_avx+0x252>
+  DB  196,98,125,24,5,232,66,0,0          ; vbroadcastss  0x42e8(%rip),%ymm8        # 6574 <_sk_callback_avx+0x252>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -7103,7 +7102,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
   DB  197,185,108,202                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm1
   DB  197,185,109,210                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm2
   DB  197,121,108,195                     ; vpunpcklqdq   %xmm3,%xmm0,%xmm8
-  DB  197,121,111,13,79,69,0,0            ; vmovdqa       0x454f(%rip),%xmm9        # 68f0 <_sk_callback_avx+0x5ca>
+  DB  197,121,111,13,79,69,0,0            ; vmovdqa       0x454f(%rip),%xmm9        # 68f0 <_sk_callback_avx+0x5ce>
   DB  196,193,113,219,193                 ; vpand         %xmm9,%xmm1,%xmm0
   DB  196,65,41,239,210                   ; vpxor         %xmm10,%xmm10,%xmm10
   DB  196,193,121,105,202                 ; vpunpckhwd    %xmm10,%xmm0,%xmm1
@@ -7195,7 +7194,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
   DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
   DB  196,195,109,24,208,1                ; vinsertf128   $0x1,%xmm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,254,63,0,0        ; vbroadcastss  0x3ffe(%rip),%ymm3        # 657c <_sk_callback_avx+0x256>
+  DB  196,226,125,24,29,250,63,0,0        ; vbroadcastss  0x3ffa(%rip),%ymm3        # 6578 <_sk_callback_avx+0x256>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,93                               ; pop           %r13
@@ -7246,7 +7245,7 @@ _sk_byte_tables_avx LABEL PROC
   DB  65,84                               ; push          %r12
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,50,63,0,0           ; vbroadcastss  0x3f32(%rip),%ymm8        # 6580 <_sk_callback_avx+0x25a>
+  DB  196,98,125,24,5,46,63,0,0           ; vbroadcastss  0x3f2e(%rip),%ymm8        # 657c <_sk_callback_avx+0x25a>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
   DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
@@ -7283,7 +7282,7 @@ _sk_byte_tables_avx LABEL PROC
   DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
   DB  196,227,53,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,128,62,0,0         ; vbroadcastss  0x3e80(%rip),%ymm9        # 6584 <_sk_callback_avx+0x25e>
+  DB  196,98,125,24,13,124,62,0,0         ; vbroadcastss  0x3e7c(%rip),%ymm9        # 6580 <_sk_callback_avx+0x25e>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
@@ -7443,7 +7442,7 @@ _sk_byte_tables_rgb_avx LABEL PROC
   DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
   DB  196,227,53,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,166,59,0,0         ; vbroadcastss  0x3ba6(%rip),%ymm9        # 6588 <_sk_callback_avx+0x262>
+  DB  196,98,125,24,13,162,59,0,0         ; vbroadcastss  0x3ba2(%rip),%ymm9        # 6584 <_sk_callback_avx+0x262>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
   DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
@@ -7730,36 +7729,36 @@ _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  196,98,125,24,37,4,55,0,0           ; vbroadcastss  0x3704(%rip),%ymm12        # 658c <_sk_callback_avx+0x266>
+  DB  196,98,125,24,37,0,55,0,0           ; vbroadcastss  0x3700(%rip),%ymm12        # 6588 <_sk_callback_avx+0x266>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,250,54,0,0         ; vbroadcastss  0x36fa(%rip),%ymm12        # 6590 <_sk_callback_avx+0x26a>
+  DB  196,98,125,24,37,246,54,0,0         ; vbroadcastss  0x36f6(%rip),%ymm12        # 658c <_sk_callback_avx+0x26a>
   DB  196,193,124,84,196                  ; vandps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,240,54,0,0         ; vbroadcastss  0x36f0(%rip),%ymm12        # 6594 <_sk_callback_avx+0x26e>
+  DB  196,98,125,24,37,236,54,0,0         ; vbroadcastss  0x36ec(%rip),%ymm12        # 6590 <_sk_callback_avx+0x26e>
   DB  196,193,124,86,196                  ; vorps         %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,230,54,0,0         ; vbroadcastss  0x36e6(%rip),%ymm12        # 6598 <_sk_callback_avx+0x272>
+  DB  196,98,125,24,37,226,54,0,0         ; vbroadcastss  0x36e2(%rip),%ymm12        # 6594 <_sk_callback_avx+0x272>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,220,54,0,0         ; vbroadcastss  0x36dc(%rip),%ymm12        # 659c <_sk_callback_avx+0x276>
+  DB  196,98,125,24,37,216,54,0,0         ; vbroadcastss  0x36d8(%rip),%ymm12        # 6598 <_sk_callback_avx+0x276>
   DB  196,65,124,89,228                   ; vmulps        %ymm12,%ymm0,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,205,54,0,0         ; vbroadcastss  0x36cd(%rip),%ymm12        # 65a0 <_sk_callback_avx+0x27a>
+  DB  196,98,125,24,37,201,54,0,0         ; vbroadcastss  0x36c9(%rip),%ymm12        # 659c <_sk_callback_avx+0x27a>
   DB  196,193,124,88,196                  ; vaddps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,195,54,0,0         ; vbroadcastss  0x36c3(%rip),%ymm12        # 65a4 <_sk_callback_avx+0x27e>
+  DB  196,98,125,24,37,191,54,0,0         ; vbroadcastss  0x36bf(%rip),%ymm12        # 65a0 <_sk_callback_avx+0x27e>
   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  196,98,125,24,29,167,54,0,0         ; vbroadcastss  0x36a7(%rip),%ymm11        # 65a8 <_sk_callback_avx+0x282>
+  DB  196,98,125,24,29,163,54,0,0         ; vbroadcastss  0x36a3(%rip),%ymm11        # 65a4 <_sk_callback_avx+0x282>
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,157,54,0,0         ; vbroadcastss  0x369d(%rip),%ymm11        # 65ac <_sk_callback_avx+0x286>
+  DB  196,98,125,24,29,153,54,0,0         ; vbroadcastss  0x3699(%rip),%ymm11        # 65a8 <_sk_callback_avx+0x286>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,124,92,195                  ; vsubps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,142,54,0,0         ; vbroadcastss  0x368e(%rip),%ymm11        # 65b0 <_sk_callback_avx+0x28a>
+  DB  196,98,125,24,29,138,54,0,0         ; vbroadcastss  0x368a(%rip),%ymm11        # 65ac <_sk_callback_avx+0x28a>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,132,54,0,0         ; vbroadcastss  0x3684(%rip),%ymm11        # 65b4 <_sk_callback_avx+0x28e>
+  DB  196,98,125,24,29,128,54,0,0         ; vbroadcastss  0x3680(%rip),%ymm11        # 65b0 <_sk_callback_avx+0x28e>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
-  DB  196,98,125,24,21,117,54,0,0         ; vbroadcastss  0x3675(%rip),%ymm10        # 65b8 <_sk_callback_avx+0x292>
+  DB  196,98,125,24,21,113,54,0,0         ; vbroadcastss  0x3671(%rip),%ymm10        # 65b4 <_sk_callback_avx+0x292>
   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
@@ -7767,7 +7766,7 @@ _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  196,98,125,24,5,76,54,0,0           ; vbroadcastss  0x364c(%rip),%ymm8        # 65bc <_sk_callback_avx+0x296>
+  DB  196,98,125,24,5,72,54,0,0           ; vbroadcastss  0x3648(%rip),%ymm8        # 65b8 <_sk_callback_avx+0x296>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -7787,36 +7786,36 @@ _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  196,98,125,24,37,253,53,0,0         ; vbroadcastss  0x35fd(%rip),%ymm12        # 65c0 <_sk_callback_avx+0x29a>
+  DB  196,98,125,24,37,249,53,0,0         ; vbroadcastss  0x35f9(%rip),%ymm12        # 65bc <_sk_callback_avx+0x29a>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,243,53,0,0         ; vbroadcastss  0x35f3(%rip),%ymm12        # 65c4 <_sk_callback_avx+0x29e>
+  DB  196,98,125,24,37,239,53,0,0         ; vbroadcastss  0x35ef(%rip),%ymm12        # 65c0 <_sk_callback_avx+0x29e>
   DB  196,193,116,84,204                  ; vandps        %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,233,53,0,0         ; vbroadcastss  0x35e9(%rip),%ymm12        # 65c8 <_sk_callback_avx+0x2a2>
+  DB  196,98,125,24,37,229,53,0,0         ; vbroadcastss  0x35e5(%rip),%ymm12        # 65c4 <_sk_callback_avx+0x2a2>
   DB  196,193,116,86,204                  ; vorps         %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,223,53,0,0         ; vbroadcastss  0x35df(%rip),%ymm12        # 65cc <_sk_callback_avx+0x2a6>
+  DB  196,98,125,24,37,219,53,0,0         ; vbroadcastss  0x35db(%rip),%ymm12        # 65c8 <_sk_callback_avx+0x2a6>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,213,53,0,0         ; vbroadcastss  0x35d5(%rip),%ymm12        # 65d0 <_sk_callback_avx+0x2aa>
+  DB  196,98,125,24,37,209,53,0,0         ; vbroadcastss  0x35d1(%rip),%ymm12        # 65cc <_sk_callback_avx+0x2aa>
   DB  196,65,116,89,228                   ; vmulps        %ymm12,%ymm1,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,198,53,0,0         ; vbroadcastss  0x35c6(%rip),%ymm12        # 65d4 <_sk_callback_avx+0x2ae>
+  DB  196,98,125,24,37,194,53,0,0         ; vbroadcastss  0x35c2(%rip),%ymm12        # 65d0 <_sk_callback_avx+0x2ae>
   DB  196,193,116,88,204                  ; vaddps        %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,188,53,0,0         ; vbroadcastss  0x35bc(%rip),%ymm12        # 65d8 <_sk_callback_avx+0x2b2>
+  DB  196,98,125,24,37,184,53,0,0         ; vbroadcastss  0x35b8(%rip),%ymm12        # 65d4 <_sk_callback_avx+0x2b2>
   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  196,98,125,24,29,160,53,0,0         ; vbroadcastss  0x35a0(%rip),%ymm11        # 65dc <_sk_callback_avx+0x2b6>
+  DB  196,98,125,24,29,156,53,0,0         ; vbroadcastss  0x359c(%rip),%ymm11        # 65d8 <_sk_callback_avx+0x2b6>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,150,53,0,0         ; vbroadcastss  0x3596(%rip),%ymm11        # 65e0 <_sk_callback_avx+0x2ba>
+  DB  196,98,125,24,29,146,53,0,0         ; vbroadcastss  0x3592(%rip),%ymm11        # 65dc <_sk_callback_avx+0x2ba>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,116,92,203                  ; vsubps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,135,53,0,0         ; vbroadcastss  0x3587(%rip),%ymm11        # 65e4 <_sk_callback_avx+0x2be>
+  DB  196,98,125,24,29,131,53,0,0         ; vbroadcastss  0x3583(%rip),%ymm11        # 65e0 <_sk_callback_avx+0x2be>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,125,53,0,0         ; vbroadcastss  0x357d(%rip),%ymm11        # 65e8 <_sk_callback_avx+0x2c2>
+  DB  196,98,125,24,29,121,53,0,0         ; vbroadcastss  0x3579(%rip),%ymm11        # 65e4 <_sk_callback_avx+0x2c2>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,21,110,53,0,0         ; vbroadcastss  0x356e(%rip),%ymm10        # 65ec <_sk_callback_avx+0x2c6>
+  DB  196,98,125,24,21,106,53,0,0         ; vbroadcastss  0x356a(%rip),%ymm10        # 65e8 <_sk_callback_avx+0x2c6>
   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
@@ -7824,7 +7823,7 @@ _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  196,98,125,24,5,69,53,0,0           ; vbroadcastss  0x3545(%rip),%ymm8        # 65f0 <_sk_callback_avx+0x2ca>
+  DB  196,98,125,24,5,65,53,0,0           ; vbroadcastss  0x3541(%rip),%ymm8        # 65ec <_sk_callback_avx+0x2ca>
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -7844,36 +7843,36 @@ _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  196,98,125,24,37,246,52,0,0         ; vbroadcastss  0x34f6(%rip),%ymm12        # 65f4 <_sk_callback_avx+0x2ce>
+  DB  196,98,125,24,37,242,52,0,0         ; vbroadcastss  0x34f2(%rip),%ymm12        # 65f0 <_sk_callback_avx+0x2ce>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,236,52,0,0         ; vbroadcastss  0x34ec(%rip),%ymm12        # 65f8 <_sk_callback_avx+0x2d2>
+  DB  196,98,125,24,37,232,52,0,0         ; vbroadcastss  0x34e8(%rip),%ymm12        # 65f4 <_sk_callback_avx+0x2d2>
   DB  196,193,108,84,212                  ; vandps        %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,226,52,0,0         ; vbroadcastss  0x34e2(%rip),%ymm12        # 65fc <_sk_callback_avx+0x2d6>
+  DB  196,98,125,24,37,222,52,0,0         ; vbroadcastss  0x34de(%rip),%ymm12        # 65f8 <_sk_callback_avx+0x2d6>
   DB  196,193,108,86,212                  ; vorps         %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,216,52,0,0         ; vbroadcastss  0x34d8(%rip),%ymm12        # 6600 <_sk_callback_avx+0x2da>
+  DB  196,98,125,24,37,212,52,0,0         ; vbroadcastss  0x34d4(%rip),%ymm12        # 65fc <_sk_callback_avx+0x2da>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,206,52,0,0         ; vbroadcastss  0x34ce(%rip),%ymm12        # 6604 <_sk_callback_avx+0x2de>
+  DB  196,98,125,24,37,202,52,0,0         ; vbroadcastss  0x34ca(%rip),%ymm12        # 6600 <_sk_callback_avx+0x2de>
   DB  196,65,108,89,228                   ; vmulps        %ymm12,%ymm2,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,191,52,0,0         ; vbroadcastss  0x34bf(%rip),%ymm12        # 6608 <_sk_callback_avx+0x2e2>
+  DB  196,98,125,24,37,187,52,0,0         ; vbroadcastss  0x34bb(%rip),%ymm12        # 6604 <_sk_callback_avx+0x2e2>
   DB  196,193,108,88,212                  ; vaddps        %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,181,52,0,0         ; vbroadcastss  0x34b5(%rip),%ymm12        # 660c <_sk_callback_avx+0x2e6>
+  DB  196,98,125,24,37,177,52,0,0         ; vbroadcastss  0x34b1(%rip),%ymm12        # 6608 <_sk_callback_avx+0x2e6>
   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  196,98,125,24,29,153,52,0,0         ; vbroadcastss  0x3499(%rip),%ymm11        # 6610 <_sk_callback_avx+0x2ea>
+  DB  196,98,125,24,29,149,52,0,0         ; vbroadcastss  0x3495(%rip),%ymm11        # 660c <_sk_callback_avx+0x2ea>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,143,52,0,0         ; vbroadcastss  0x348f(%rip),%ymm11        # 6614 <_sk_callback_avx+0x2ee>
+  DB  196,98,125,24,29,139,52,0,0         ; vbroadcastss  0x348b(%rip),%ymm11        # 6610 <_sk_callback_avx+0x2ee>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,108,92,211                  ; vsubps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,128,52,0,0         ; vbroadcastss  0x3480(%rip),%ymm11        # 6618 <_sk_callback_avx+0x2f2>
+  DB  196,98,125,24,29,124,52,0,0         ; vbroadcastss  0x347c(%rip),%ymm11        # 6614 <_sk_callback_avx+0x2f2>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,118,52,0,0         ; vbroadcastss  0x3476(%rip),%ymm11        # 661c <_sk_callback_avx+0x2f6>
+  DB  196,98,125,24,29,114,52,0,0         ; vbroadcastss  0x3472(%rip),%ymm11        # 6618 <_sk_callback_avx+0x2f6>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
-  DB  196,98,125,24,21,103,52,0,0         ; vbroadcastss  0x3467(%rip),%ymm10        # 6620 <_sk_callback_avx+0x2fa>
+  DB  196,98,125,24,21,99,52,0,0          ; vbroadcastss  0x3463(%rip),%ymm10        # 661c <_sk_callback_avx+0x2fa>
   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
@@ -7881,7 +7880,7 @@ _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  196,98,125,24,5,62,52,0,0           ; vbroadcastss  0x343e(%rip),%ymm8        # 6624 <_sk_callback_avx+0x2fe>
+  DB  196,98,125,24,5,58,52,0,0           ; vbroadcastss  0x343a(%rip),%ymm8        # 6620 <_sk_callback_avx+0x2fe>
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -7901,36 +7900,36 @@ _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  196,98,125,24,37,239,51,0,0         ; vbroadcastss  0x33ef(%rip),%ymm12        # 6628 <_sk_callback_avx+0x302>
+  DB  196,98,125,24,37,235,51,0,0         ; vbroadcastss  0x33eb(%rip),%ymm12        # 6624 <_sk_callback_avx+0x302>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,229,51,0,0         ; vbroadcastss  0x33e5(%rip),%ymm12        # 662c <_sk_callback_avx+0x306>
+  DB  196,98,125,24,37,225,51,0,0         ; vbroadcastss  0x33e1(%rip),%ymm12        # 6628 <_sk_callback_avx+0x306>
   DB  196,193,100,84,220                  ; vandps        %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,219,51,0,0         ; vbroadcastss  0x33db(%rip),%ymm12        # 6630 <_sk_callback_avx+0x30a>
+  DB  196,98,125,24,37,215,51,0,0         ; vbroadcastss  0x33d7(%rip),%ymm12        # 662c <_sk_callback_avx+0x30a>
   DB  196,193,100,86,220                  ; vorps         %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,209,51,0,0         ; vbroadcastss  0x33d1(%rip),%ymm12        # 6634 <_sk_callback_avx+0x30e>
+  DB  196,98,125,24,37,205,51,0,0         ; vbroadcastss  0x33cd(%rip),%ymm12        # 6630 <_sk_callback_avx+0x30e>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,199,51,0,0         ; vbroadcastss  0x33c7(%rip),%ymm12        # 6638 <_sk_callback_avx+0x312>
+  DB  196,98,125,24,37,195,51,0,0         ; vbroadcastss  0x33c3(%rip),%ymm12        # 6634 <_sk_callback_avx+0x312>
   DB  196,65,100,89,228                   ; vmulps        %ymm12,%ymm3,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,184,51,0,0         ; vbroadcastss  0x33b8(%rip),%ymm12        # 663c <_sk_callback_avx+0x316>
+  DB  196,98,125,24,37,180,51,0,0         ; vbroadcastss  0x33b4(%rip),%ymm12        # 6638 <_sk_callback_avx+0x316>
   DB  196,193,100,88,220                  ; vaddps        %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,174,51,0,0         ; vbroadcastss  0x33ae(%rip),%ymm12        # 6640 <_sk_callback_avx+0x31a>
+  DB  196,98,125,24,37,170,51,0,0         ; vbroadcastss  0x33aa(%rip),%ymm12        # 663c <_sk_callback_avx+0x31a>
   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  196,98,125,24,29,146,51,0,0         ; vbroadcastss  0x3392(%rip),%ymm11        # 6644 <_sk_callback_avx+0x31e>
+  DB  196,98,125,24,29,142,51,0,0         ; vbroadcastss  0x338e(%rip),%ymm11        # 6640 <_sk_callback_avx+0x31e>
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,136,51,0,0         ; vbroadcastss  0x3388(%rip),%ymm11        # 6648 <_sk_callback_avx+0x322>
+  DB  196,98,125,24,29,132,51,0,0         ; vbroadcastss  0x3384(%rip),%ymm11        # 6644 <_sk_callback_avx+0x322>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,100,92,219                  ; vsubps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,121,51,0,0         ; vbroadcastss  0x3379(%rip),%ymm11        # 664c <_sk_callback_avx+0x326>
+  DB  196,98,125,24,29,117,51,0,0         ; vbroadcastss  0x3375(%rip),%ymm11        # 6648 <_sk_callback_avx+0x326>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,111,51,0,0         ; vbroadcastss  0x336f(%rip),%ymm11        # 6650 <_sk_callback_avx+0x32a>
+  DB  196,98,125,24,29,107,51,0,0         ; vbroadcastss  0x336b(%rip),%ymm11        # 664c <_sk_callback_avx+0x32a>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,100,88,218                  ; vaddps        %ymm10,%ymm3,%ymm3
-  DB  196,98,125,24,21,96,51,0,0          ; vbroadcastss  0x3360(%rip),%ymm10        # 6654 <_sk_callback_avx+0x32e>
+  DB  196,98,125,24,21,92,51,0,0          ; vbroadcastss  0x335c(%rip),%ymm10        # 6650 <_sk_callback_avx+0x32e>
   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
@@ -7938,38 +7937,38 @@ _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  196,98,125,24,5,55,51,0,0           ; vbroadcastss  0x3337(%rip),%ymm8        # 6658 <_sk_callback_avx+0x332>
+  DB  196,98,125,24,5,51,51,0,0           ; vbroadcastss  0x3333(%rip),%ymm8        # 6654 <_sk_callback_avx+0x332>
   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  196,98,125,24,5,41,51,0,0           ; vbroadcastss  0x3329(%rip),%ymm8        # 665c <_sk_callback_avx+0x336>
+  DB  196,98,125,24,5,37,51,0,0           ; vbroadcastss  0x3325(%rip),%ymm8        # 6658 <_sk_callback_avx+0x336>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,31,51,0,0           ; vbroadcastss  0x331f(%rip),%ymm8        # 6660 <_sk_callback_avx+0x33a>
+  DB  196,98,125,24,5,27,51,0,0           ; vbroadcastss  0x331b(%rip),%ymm8        # 665c <_sk_callback_avx+0x33a>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,98,125,24,13,21,51,0,0          ; vbroadcastss  0x3315(%rip),%ymm9        # 6664 <_sk_callback_avx+0x33e>
+  DB  196,98,125,24,13,17,51,0,0          ; vbroadcastss  0x3311(%rip),%ymm9        # 6660 <_sk_callback_avx+0x33e>
   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  196,98,125,24,5,1,51,0,0            ; vbroadcastss  0x3301(%rip),%ymm8        # 6668 <_sk_callback_avx+0x342>
+  DB  196,98,125,24,5,253,50,0,0          ; vbroadcastss  0x32fd(%rip),%ymm8        # 6664 <_sk_callback_avx+0x342>
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,247,50,0,0          ; vbroadcastss  0x32f7(%rip),%ymm8        # 666c <_sk_callback_avx+0x346>
+  DB  196,98,125,24,5,243,50,0,0          ; vbroadcastss  0x32f3(%rip),%ymm8        # 6668 <_sk_callback_avx+0x346>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,237,50,0,0          ; vbroadcastss  0x32ed(%rip),%ymm8        # 6670 <_sk_callback_avx+0x34a>
+  DB  196,98,125,24,5,233,50,0,0          ; vbroadcastss  0x32e9(%rip),%ymm8        # 666c <_sk_callback_avx+0x34a>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  197,252,88,201                      ; vaddps        %ymm1,%ymm0,%ymm1
-  DB  196,98,125,24,5,223,50,0,0          ; vbroadcastss  0x32df(%rip),%ymm8        # 6674 <_sk_callback_avx+0x34e>
+  DB  196,98,125,24,5,219,50,0,0          ; vbroadcastss  0x32db(%rip),%ymm8        # 6670 <_sk_callback_avx+0x34e>
   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  196,98,125,24,13,200,50,0,0         ; vbroadcastss  0x32c8(%rip),%ymm9        # 6678 <_sk_callback_avx+0x352>
+  DB  196,98,125,24,13,196,50,0,0         ; vbroadcastss  0x32c4(%rip),%ymm9        # 6674 <_sk_callback_avx+0x352>
   DB  196,65,52,194,208,1                 ; vcmpltps      %ymm8,%ymm9,%ymm10
-  DB  196,98,125,24,29,189,50,0,0         ; vbroadcastss  0x32bd(%rip),%ymm11        # 667c <_sk_callback_avx+0x356>
+  DB  196,98,125,24,29,185,50,0,0         ; vbroadcastss  0x32b9(%rip),%ymm11        # 6678 <_sk_callback_avx+0x356>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,37,179,50,0,0         ; vbroadcastss  0x32b3(%rip),%ymm12        # 6680 <_sk_callback_avx+0x35a>
+  DB  196,98,125,24,37,175,50,0,0         ; vbroadcastss  0x32af(%rip),%ymm12        # 667c <_sk_callback_avx+0x35a>
   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
@@ -7984,9 +7983,9 @@ _sk_lab_to_xyz_avx LABEL PROC
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
   DB  196,193,108,89,212                  ; vmulps        %ymm12,%ymm2,%ymm2
   DB  196,227,109,74,208,144              ; vblendvps     %ymm9,%ymm0,%ymm2,%ymm2
-  DB  196,226,125,24,5,105,50,0,0         ; vbroadcastss  0x3269(%rip),%ymm0        # 6684 <_sk_callback_avx+0x35e>
+  DB  196,226,125,24,5,101,50,0,0         ; vbroadcastss  0x3265(%rip),%ymm0        # 6680 <_sk_callback_avx+0x35e>
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,96,50,0,0           ; vbroadcastss  0x3260(%rip),%ymm8        # 6688 <_sk_callback_avx+0x362>
+  DB  196,98,125,24,5,92,50,0,0           ; vbroadcastss  0x325c(%rip),%ymm8        # 6684 <_sk_callback_avx+0x362>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -8005,7 +8004,7 @@ _sk_load_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  196,226,125,24,13,36,50,0,0         ; vbroadcastss  0x3224(%rip),%ymm1        # 668c <_sk_callback_avx+0x366>
+  DB  196,226,125,24,13,32,50,0,0         ; vbroadcastss  0x3220(%rip),%ymm1        # 6688 <_sk_callback_avx+0x366>
   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
@@ -8072,7 +8071,7 @@ _sk_gather_a8_avx LABEL PROC
   DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
   DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,25,49,0,0         ; vbroadcastss  0x3119(%rip),%ymm1        # 6690 <_sk_callback_avx+0x36a>
+  DB  196,226,125,24,13,21,49,0,0         ; vbroadcastss  0x3115(%rip),%ymm1        # 668c <_sk_callback_avx+0x36a>
   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
@@ -8088,7 +8087,7 @@ PUBLIC _sk_store_a8_avx
 _sk_store_a8_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,244,48,0,0          ; vbroadcastss  0x30f4(%rip),%ymm8        # 6694 <_sk_callback_avx+0x36e>
+  DB  196,98,125,24,5,240,48,0,0          ; vbroadcastss  0x30f0(%rip),%ymm8        # 6690 <_sk_callback_avx+0x36e>
   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
@@ -8156,10 +8155,10 @@ _sk_load_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  196,226,125,24,13,25,48,0,0         ; vbroadcastss  0x3019(%rip),%ymm1        # 6698 <_sk_callback_avx+0x372>
+  DB  196,226,125,24,13,21,48,0,0         ; vbroadcastss  0x3015(%rip),%ymm1        # 6694 <_sk_callback_avx+0x372>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,14,48,0,0         ; vbroadcastss  0x300e(%rip),%ymm3        # 669c <_sk_callback_avx+0x376>
+  DB  196,226,125,24,29,10,48,0,0         ; vbroadcastss  0x300a(%rip),%ymm3        # 6698 <_sk_callback_avx+0x376>
   DB  76,137,193                          ; mov           %r8,%rcx
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
@@ -8223,10 +8222,10 @@ _sk_gather_g8_avx LABEL PROC
   DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
   DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,13,47,0,0         ; vbroadcastss  0x2f0d(%rip),%ymm1        # 66a0 <_sk_callback_avx+0x37a>
+  DB  196,226,125,24,13,9,47,0,0          ; vbroadcastss  0x2f09(%rip),%ymm1        # 669c <_sk_callback_avx+0x37a>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,2,47,0,0          ; vbroadcastss  0x2f02(%rip),%ymm3        # 66a4 <_sk_callback_avx+0x37e>
+  DB  196,226,125,24,29,254,46,0,0        ; vbroadcastss  0x2efe(%rip),%ymm3        # 66a0 <_sk_callback_avx+0x37e>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  91                                  ; pop           %rbx
@@ -8304,10 +8303,10 @@ _sk_gather_i8_avx LABEL PROC
   DB  196,163,121,34,4,163,2              ; vpinsrd       $0x2,(%rbx,%r12,4),%xmm0,%xmm0
   DB  196,163,121,34,28,19,3              ; vpinsrd       $0x3,(%rbx,%r10,1),%xmm0,%xmm3
   DB  196,227,61,24,195,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  DB  197,124,40,21,146,47,0,0            ; vmovaps       0x2f92(%rip),%ymm10        # 6880 <_sk_callback_avx+0x55a>
+  DB  197,124,40,21,146,47,0,0            ; vmovaps       0x2f92(%rip),%ymm10        # 6880 <_sk_callback_avx+0x55e>
   DB  196,193,124,84,194                  ; vandps        %ymm10,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,168,45,0,0         ; vbroadcastss  0x2da8(%rip),%ymm9        # 66a8 <_sk_callback_avx+0x382>
+  DB  196,98,125,24,13,164,45,0,0         ; vbroadcastss  0x2da4(%rip),%ymm9        # 66a4 <_sk_callback_avx+0x382>
   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
@@ -8345,23 +8344,23 @@ _sk_load_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  196,226,125,24,5,18,45,0,0          ; vbroadcastss  0x2d12(%rip),%ymm0        # 66ac <_sk_callback_avx+0x386>
+  DB  196,226,125,24,5,14,45,0,0          ; vbroadcastss  0x2d0e(%rip),%ymm0        # 66a8 <_sk_callback_avx+0x386>
   DB  197,236,84,192                      ; vandps        %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,5,45,0,0          ; vbroadcastss  0x2d05(%rip),%ymm1        # 66b0 <_sk_callback_avx+0x38a>
+  DB  196,226,125,24,13,1,45,0,0          ; vbroadcastss  0x2d01(%rip),%ymm1        # 66ac <_sk_callback_avx+0x38a>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,252,44,0,0        ; vbroadcastss  0x2cfc(%rip),%ymm1        # 66b4 <_sk_callback_avx+0x38e>
+  DB  196,226,125,24,13,248,44,0,0        ; vbroadcastss  0x2cf8(%rip),%ymm1        # 66b0 <_sk_callback_avx+0x38e>
   DB  197,236,84,201                      ; vandps        %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,239,44,0,0        ; vbroadcastss  0x2cef(%rip),%ymm3        # 66b8 <_sk_callback_avx+0x392>
+  DB  196,226,125,24,29,235,44,0,0        ; vbroadcastss  0x2ceb(%rip),%ymm3        # 66b4 <_sk_callback_avx+0x392>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,24,29,230,44,0,0        ; vbroadcastss  0x2ce6(%rip),%ymm3        # 66bc <_sk_callback_avx+0x396>
+  DB  196,226,125,24,29,226,44,0,0        ; vbroadcastss  0x2ce2(%rip),%ymm3        # 66b8 <_sk_callback_avx+0x396>
   DB  197,236,84,211                      ; vandps        %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,217,44,0,0        ; vbroadcastss  0x2cd9(%rip),%ymm3        # 66c0 <_sk_callback_avx+0x39a>
+  DB  196,226,125,24,29,213,44,0,0        ; vbroadcastss  0x2cd5(%rip),%ymm3        # 66bc <_sk_callback_avx+0x39a>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,206,44,0,0        ; vbroadcastss  0x2cce(%rip),%ymm3        # 66c4 <_sk_callback_avx+0x39e>
+  DB  196,226,125,24,29,202,44,0,0        ; vbroadcastss  0x2cca(%rip),%ymm3        # 66c0 <_sk_callback_avx+0x39e>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,200                          ; mov           %ecx,%r8d
   DB  65,128,224,7                        ; and           $0x7,%r8b
@@ -8458,23 +8457,23 @@ _sk_gather_565_avx LABEL PROC
   DB  197,249,105,201                     ; vpunpckhwd    %xmm1,%xmm0,%xmm1
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,125,24,209,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm2
-  DB  196,226,125,24,5,110,43,0,0         ; vbroadcastss  0x2b6e(%rip),%ymm0        # 66c8 <_sk_callback_avx+0x3a2>
+  DB  196,226,125,24,5,106,43,0,0         ; vbroadcastss  0x2b6a(%rip),%ymm0        # 66c4 <_sk_callback_avx+0x3a2>
   DB  197,236,84,192                      ; vandps        %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,97,43,0,0         ; vbroadcastss  0x2b61(%rip),%ymm1        # 66cc <_sk_callback_avx+0x3a6>
+  DB  196,226,125,24,13,93,43,0,0         ; vbroadcastss  0x2b5d(%rip),%ymm1        # 66c8 <_sk_callback_avx+0x3a6>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,88,43,0,0         ; vbroadcastss  0x2b58(%rip),%ymm1        # 66d0 <_sk_callback_avx+0x3aa>
+  DB  196,226,125,24,13,84,43,0,0         ; vbroadcastss  0x2b54(%rip),%ymm1        # 66cc <_sk_callback_avx+0x3aa>
   DB  197,236,84,201                      ; vandps        %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,75,43,0,0         ; vbroadcastss  0x2b4b(%rip),%ymm3        # 66d4 <_sk_callback_avx+0x3ae>
+  DB  196,226,125,24,29,71,43,0,0         ; vbroadcastss  0x2b47(%rip),%ymm3        # 66d0 <_sk_callback_avx+0x3ae>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,24,29,66,43,0,0         ; vbroadcastss  0x2b42(%rip),%ymm3        # 66d8 <_sk_callback_avx+0x3b2>
+  DB  196,226,125,24,29,62,43,0,0         ; vbroadcastss  0x2b3e(%rip),%ymm3        # 66d4 <_sk_callback_avx+0x3b2>
   DB  197,236,84,211                      ; vandps        %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,53,43,0,0         ; vbroadcastss  0x2b35(%rip),%ymm3        # 66dc <_sk_callback_avx+0x3b6>
+  DB  196,226,125,24,29,49,43,0,0         ; vbroadcastss  0x2b31(%rip),%ymm3        # 66d8 <_sk_callback_avx+0x3b6>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,42,43,0,0         ; vbroadcastss  0x2b2a(%rip),%ymm3        # 66e0 <_sk_callback_avx+0x3ba>
+  DB  196,226,125,24,29,38,43,0,0         ; vbroadcastss  0x2b26(%rip),%ymm3        # 66dc <_sk_callback_avx+0x3ba>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,94                               ; pop           %r14
@@ -8486,14 +8485,14 @@ PUBLIC _sk_store_565_avx
 _sk_store_565_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,22,43,0,0           ; vbroadcastss  0x2b16(%rip),%ymm8        # 66e4 <_sk_callback_avx+0x3be>
+  DB  196,98,125,24,5,18,43,0,0           ; vbroadcastss  0x2b12(%rip),%ymm8        # 66e0 <_sk_callback_avx+0x3be>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,193,41,114,241,11               ; 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  196,98,125,24,21,239,42,0,0         ; vbroadcastss  0x2aef(%rip),%ymm10        # 66e8 <_sk_callback_avx+0x3c2>
+  DB  196,98,125,24,21,235,42,0,0         ; vbroadcastss  0x2aeb(%rip),%ymm10        # 66e4 <_sk_callback_avx+0x3c2>
   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
@@ -8565,25 +8564,25 @@ _sk_load_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  196,226,125,24,5,248,41,0,0         ; vbroadcastss  0x29f8(%rip),%ymm0        # 66ec <_sk_callback_avx+0x3c6>
+  DB  196,226,125,24,5,244,41,0,0         ; vbroadcastss  0x29f4(%rip),%ymm0        # 66e8 <_sk_callback_avx+0x3c6>
   DB  197,228,84,192                      ; vandps        %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,235,41,0,0        ; vbroadcastss  0x29eb(%rip),%ymm1        # 66f0 <_sk_callback_avx+0x3ca>
+  DB  196,226,125,24,13,231,41,0,0        ; vbroadcastss  0x29e7(%rip),%ymm1        # 66ec <_sk_callback_avx+0x3ca>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,226,41,0,0        ; vbroadcastss  0x29e2(%rip),%ymm1        # 66f4 <_sk_callback_avx+0x3ce>
+  DB  196,226,125,24,13,222,41,0,0        ; vbroadcastss  0x29de(%rip),%ymm1        # 66f0 <_sk_callback_avx+0x3ce>
   DB  197,228,84,201                      ; vandps        %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,213,41,0,0        ; vbroadcastss  0x29d5(%rip),%ymm2        # 66f8 <_sk_callback_avx+0x3d2>
+  DB  196,226,125,24,21,209,41,0,0        ; vbroadcastss  0x29d1(%rip),%ymm2        # 66f4 <_sk_callback_avx+0x3d2>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,204,41,0,0        ; vbroadcastss  0x29cc(%rip),%ymm2        # 66fc <_sk_callback_avx+0x3d6>
+  DB  196,226,125,24,21,200,41,0,0        ; vbroadcastss  0x29c8(%rip),%ymm2        # 66f8 <_sk_callback_avx+0x3d6>
   DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,191,41,0,0          ; vbroadcastss  0x29bf(%rip),%ymm8        # 6700 <_sk_callback_avx+0x3da>
+  DB  196,98,125,24,5,187,41,0,0          ; vbroadcastss  0x29bb(%rip),%ymm8        # 66fc <_sk_callback_avx+0x3da>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,181,41,0,0          ; vbroadcastss  0x29b5(%rip),%ymm8        # 6704 <_sk_callback_avx+0x3de>
+  DB  196,98,125,24,5,177,41,0,0          ; vbroadcastss  0x29b1(%rip),%ymm8        # 6700 <_sk_callback_avx+0x3de>
   DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,167,41,0,0          ; vbroadcastss  0x29a7(%rip),%ymm8        # 6708 <_sk_callback_avx+0x3e2>
+  DB  196,98,125,24,5,163,41,0,0          ; vbroadcastss  0x29a3(%rip),%ymm8        # 6704 <_sk_callback_avx+0x3e2>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -8683,25 +8682,25 @@ _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  196,226,125,24,5,62,40,0,0          ; vbroadcastss  0x283e(%rip),%ymm0        # 670c <_sk_callback_avx+0x3e6>
+  DB  196,226,125,24,5,58,40,0,0          ; vbroadcastss  0x283a(%rip),%ymm0        # 6708 <_sk_callback_avx+0x3e6>
   DB  197,228,84,192                      ; vandps        %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,49,40,0,0         ; vbroadcastss  0x2831(%rip),%ymm1        # 6710 <_sk_callback_avx+0x3ea>
+  DB  196,226,125,24,13,45,40,0,0         ; vbroadcastss  0x282d(%rip),%ymm1        # 670c <_sk_callback_avx+0x3ea>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,40,40,0,0         ; vbroadcastss  0x2828(%rip),%ymm1        # 6714 <_sk_callback_avx+0x3ee>
+  DB  196,226,125,24,13,36,40,0,0         ; vbroadcastss  0x2824(%rip),%ymm1        # 6710 <_sk_callback_avx+0x3ee>
   DB  197,228,84,201                      ; vandps        %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,27,40,0,0         ; vbroadcastss  0x281b(%rip),%ymm2        # 6718 <_sk_callback_avx+0x3f2>
+  DB  196,226,125,24,21,23,40,0,0         ; vbroadcastss  0x2817(%rip),%ymm2        # 6714 <_sk_callback_avx+0x3f2>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,18,40,0,0         ; vbroadcastss  0x2812(%rip),%ymm2        # 671c <_sk_callback_avx+0x3f6>
+  DB  196,226,125,24,21,14,40,0,0         ; vbroadcastss  0x280e(%rip),%ymm2        # 6718 <_sk_callback_avx+0x3f6>
   DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,5,40,0,0            ; vbroadcastss  0x2805(%rip),%ymm8        # 6720 <_sk_callback_avx+0x3fa>
+  DB  196,98,125,24,5,1,40,0,0            ; vbroadcastss  0x2801(%rip),%ymm8        # 671c <_sk_callback_avx+0x3fa>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,251,39,0,0          ; vbroadcastss  0x27fb(%rip),%ymm8        # 6724 <_sk_callback_avx+0x3fe>
+  DB  196,98,125,24,5,247,39,0,0          ; vbroadcastss  0x27f7(%rip),%ymm8        # 6720 <_sk_callback_avx+0x3fe>
   DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,237,39,0,0          ; vbroadcastss  0x27ed(%rip),%ymm8        # 6728 <_sk_callback_avx+0x402>
+  DB  196,98,125,24,5,233,39,0,0          ; vbroadcastss  0x27e9(%rip),%ymm8        # 6724 <_sk_callback_avx+0x402>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -8715,7 +8714,7 @@ PUBLIC _sk_store_4444_avx
 _sk_store_4444_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,210,39,0,0          ; vbroadcastss  0x27d2(%rip),%ymm8        # 672c <_sk_callback_avx+0x406>
+  DB  196,98,125,24,5,206,39,0,0          ; vbroadcastss  0x27ce(%rip),%ymm8        # 6728 <_sk_callback_avx+0x406>
   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
@@ -8794,10 +8793,10 @@ _sk_load_8888_avx LABEL PROC
   DB  72,133,201                          ; test          %rcx,%rcx
   DB  15,133,135,0,0,0                    ; jne           4101 <_sk_load_8888_avx+0x95>
   DB  196,65,124,16,12,186                ; vmovups       (%r10,%rdi,4),%ymm9
-  DB  197,124,40,21,24,40,0,0             ; vmovaps       0x2818(%rip),%ymm10        # 68a0 <_sk_callback_avx+0x57a>
+  DB  197,124,40,21,24,40,0,0             ; vmovaps       0x2818(%rip),%ymm10        # 68a0 <_sk_callback_avx+0x57e>
   DB  196,193,52,84,194                   ; vandps        %ymm10,%ymm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,150,38,0,0          ; vbroadcastss  0x2696(%rip),%ymm8        # 6730 <_sk_callback_avx+0x40a>
+  DB  196,98,125,24,5,146,38,0,0          ; vbroadcastss  0x2692(%rip),%ymm8        # 672c <_sk_callback_avx+0x40a>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  196,193,113,114,209,8               ; vpsrld        $0x8,%xmm9,%xmm1
   DB  196,99,125,25,203,1                 ; vextractf128  $0x1,%ymm9,%xmm3
@@ -8910,10 +8909,10 @@ _sk_gather_8888_avx LABEL PROC
   DB  196,131,121,34,4,152,2              ; vpinsrd       $0x2,(%r8,%r11,4),%xmm0,%xmm0
   DB  196,131,121,34,28,144,3             ; vpinsrd       $0x3,(%r8,%r10,4),%xmm0,%xmm3
   DB  196,227,61,24,195,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  DB  197,124,40,21,66,38,0,0             ; vmovaps       0x2642(%rip),%ymm10        # 68c0 <_sk_callback_avx+0x59a>
+  DB  197,124,40,21,66,38,0,0             ; vmovaps       0x2642(%rip),%ymm10        # 68c0 <_sk_callback_avx+0x59e>
   DB  196,193,124,84,194                  ; vandps        %ymm10,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,164,36,0,0         ; vbroadcastss  0x24a4(%rip),%ymm9        # 6734 <_sk_callback_avx+0x40e>
+  DB  196,98,125,24,13,160,36,0,0         ; vbroadcastss  0x24a0(%rip),%ymm9        # 6730 <_sk_callback_avx+0x40e>
   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
@@ -8943,7 +8942,7 @@ PUBLIC _sk_store_8888_avx
 _sk_store_8888_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,16                           ; mov           (%rax),%r10
-  DB  196,98,125,24,5,50,36,0,0           ; vbroadcastss  0x2432(%rip),%ymm8        # 6738 <_sk_callback_avx+0x412>
+  DB  196,98,125,24,5,46,36,0,0           ; vbroadcastss  0x242e(%rip),%ymm8        # 6734 <_sk_callback_avx+0x412>
   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
@@ -9046,13 +9045,13 @@ _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  196,98,125,24,37,151,34,0,0         ; vbroadcastss  0x2297(%rip),%ymm12        # 673c <_sk_callback_avx+0x416>
+  DB  196,98,125,24,37,147,34,0,0         ; vbroadcastss  0x2293(%rip),%ymm12        # 6738 <_sk_callback_avx+0x416>
   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  196,98,121,24,29,131,34,0,0         ; vbroadcastss  0x2283(%rip),%xmm11        # 6740 <_sk_callback_avx+0x41a>
+  DB  196,98,121,24,29,127,34,0,0         ; vbroadcastss  0x227f(%rip),%xmm11        # 673c <_sk_callback_avx+0x41a>
   DB  196,193,8,87,219                    ; vxorps        %xmm11,%xmm14,%xmm3
-  DB  196,98,121,24,45,121,34,0,0         ; vbroadcastss  0x2279(%rip),%xmm13        # 6744 <_sk_callback_avx+0x41e>
+  DB  196,98,121,24,45,117,34,0,0         ; vbroadcastss  0x2275(%rip),%xmm13        # 6740 <_sk_callback_avx+0x41e>
   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
@@ -9066,7 +9065,7 @@ _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  196,226,121,24,29,47,34,0,0         ; vbroadcastss  0x222f(%rip),%xmm3        # 6748 <_sk_callback_avx+0x422>
+  DB  196,226,121,24,29,43,34,0,0         ; vbroadcastss  0x222b(%rip),%xmm3        # 6744 <_sk_callback_avx+0x422>
   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
@@ -9243,13 +9242,13 @@ _sk_gather_f16_avx LABEL PROC
   DB  197,249,105,210                     ; vpunpckhwd    %xmm2,%xmm0,%xmm2
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,125,24,194,1                ; vinsertf128   $0x1,%xmm2,%ymm0,%ymm0
-  DB  196,98,125,24,37,239,30,0,0         ; vbroadcastss  0x1eef(%rip),%ymm12        # 674c <_sk_callback_avx+0x426>
+  DB  196,98,125,24,37,235,30,0,0         ; vbroadcastss  0x1eeb(%rip),%ymm12        # 6748 <_sk_callback_avx+0x426>
   DB  196,193,124,84,212                  ; vandps        %ymm12,%ymm0,%ymm2
   DB  197,252,87,194                      ; vxorps        %ymm2,%ymm0,%ymm0
   DB  196,195,125,25,198,1                ; vextractf128  $0x1,%ymm0,%xmm14
-  DB  196,98,121,24,29,219,30,0,0         ; vbroadcastss  0x1edb(%rip),%xmm11        # 6750 <_sk_callback_avx+0x42a>
+  DB  196,98,121,24,29,215,30,0,0         ; vbroadcastss  0x1ed7(%rip),%xmm11        # 674c <_sk_callback_avx+0x42a>
   DB  196,193,8,87,219                    ; vxorps        %xmm11,%xmm14,%xmm3
-  DB  196,98,121,24,45,209,30,0,0         ; vbroadcastss  0x1ed1(%rip),%xmm13        # 6754 <_sk_callback_avx+0x42e>
+  DB  196,98,121,24,45,205,30,0,0         ; vbroadcastss  0x1ecd(%rip),%xmm13        # 6750 <_sk_callback_avx+0x42e>
   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
@@ -9263,7 +9262,7 @@ _sk_gather_f16_avx LABEL PROC
   DB  196,227,125,24,195,1                ; vinsertf128   $0x1,%xmm3,%ymm0,%ymm0
   DB  197,252,86,194                      ; vorps         %ymm2,%ymm0,%ymm0
   DB  196,227,125,25,194,1                ; vextractf128  $0x1,%ymm0,%xmm2
-  DB  196,226,121,24,29,135,30,0,0        ; vbroadcastss  0x1e87(%rip),%xmm3        # 6758 <_sk_callback_avx+0x432>
+  DB  196,226,121,24,29,131,30,0,0        ; vbroadcastss  0x1e83(%rip),%xmm3        # 6754 <_sk_callback_avx+0x432>
   DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
   DB  197,249,254,195                     ; vpaddd        %xmm3,%xmm0,%xmm0
   DB  196,227,125,24,194,1                ; vinsertf128   $0x1,%xmm2,%ymm0,%ymm0
@@ -9365,12 +9364,12 @@ _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  196,98,125,24,13,148,28,0,0         ; vbroadcastss  0x1c94(%rip),%ymm9        # 675c <_sk_callback_avx+0x436>
+  DB  196,98,125,24,13,144,28,0,0         ; vbroadcastss  0x1c90(%rip),%ymm9        # 6758 <_sk_callback_avx+0x436>
   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  196,98,121,24,5,122,28,0,0          ; vbroadcastss  0x1c7a(%rip),%xmm8        # 6760 <_sk_callback_avx+0x43a>
+  DB  196,98,121,24,5,118,28,0,0          ; vbroadcastss  0x1c76(%rip),%xmm8        # 675c <_sk_callback_avx+0x43a>
   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
@@ -9380,7 +9379,7 @@ _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  196,98,125,24,21,65,28,0,0          ; vbroadcastss  0x1c41(%rip),%ymm10        # 6764 <_sk_callback_avx+0x43e>
+  DB  196,98,125,24,21,61,28,0,0          ; vbroadcastss  0x1c3d(%rip),%ymm10        # 6760 <_sk_callback_avx+0x43e>
   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
@@ -9523,7 +9522,7 @@ _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  196,98,125,24,29,144,25,0,0         ; vbroadcastss  0x1990(%rip),%ymm11        # 6768 <_sk_callback_avx+0x442>
+  DB  196,98,125,24,29,140,25,0,0         ; vbroadcastss  0x198c(%rip),%ymm11        # 6764 <_sk_callback_avx+0x442>
   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
@@ -9614,7 +9613,7 @@ _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  196,98,125,24,29,240,23,0,0         ; vbroadcastss  0x17f0(%rip),%ymm11        # 676c <_sk_callback_avx+0x446>
+  DB  196,98,125,24,29,236,23,0,0         ; vbroadcastss  0x17ec(%rip),%ymm11        # 6768 <_sk_callback_avx+0x446>
   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
@@ -9635,7 +9634,7 @@ _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  196,226,125,24,29,141,23,0,0        ; vbroadcastss  0x178d(%rip),%ymm3        # 6770 <_sk_callback_avx+0x44a>
+  DB  196,226,125,24,29,137,23,0,0        ; vbroadcastss  0x1789(%rip),%ymm3        # 676c <_sk_callback_avx+0x44a>
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,4,64                ; vmovd         (%r8,%rax,2),%xmm0
   DB  196,193,121,196,68,64,4,2           ; vpinsrw       $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
@@ -9676,7 +9675,7 @@ _sk_store_u16_be_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  72,141,4,189,0,0,0,0                ; lea           0x0(,%rdi,4),%rax
-  DB  196,98,125,24,5,202,22,0,0          ; vbroadcastss  0x16ca(%rip),%ymm8        # 6774 <_sk_callback_avx+0x44e>
+  DB  196,98,125,24,5,198,22,0,0          ; vbroadcastss  0x16c6(%rip),%ymm8        # 6770 <_sk_callback_avx+0x44e>
   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
@@ -9924,12 +9923,12 @@ _sk_mirror_y_avx LABEL PROC
 
 PUBLIC _sk_luminance_to_alpha_avx
 _sk_luminance_to_alpha_avx LABEL PROC
-  DB  196,226,125,24,29,239,18,0,0        ; vbroadcastss  0x12ef(%rip),%ymm3        # 6778 <_sk_callback_avx+0x452>
+  DB  196,226,125,24,29,235,18,0,0        ; vbroadcastss  0x12eb(%rip),%ymm3        # 6774 <_sk_callback_avx+0x452>
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,125,24,29,230,18,0,0        ; vbroadcastss  0x12e6(%rip),%ymm3        # 677c <_sk_callback_avx+0x456>
+  DB  196,226,125,24,29,226,18,0,0        ; vbroadcastss  0x12e2(%rip),%ymm3        # 6778 <_sk_callback_avx+0x456>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,217,18,0,0        ; vbroadcastss  0x12d9(%rip),%ymm1        # 6780 <_sk_callback_avx+0x45a>
+  DB  196,226,125,24,13,213,18,0,0        ; vbroadcastss  0x12d5(%rip),%ymm1        # 677c <_sk_callback_avx+0x45a>
   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
@@ -10272,7 +10271,7 @@ _sk_gradient_avx LABEL PROC
   DB  73,255,200                          ; dec           %r8
   DB  72,131,195,4                        ; add           $0x4,%rbx
   DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
-  DB  196,98,125,24,21,180,12,0,0         ; vbroadcastss  0xcb4(%rip),%ymm10        # 6784 <_sk_callback_avx+0x45e>
+  DB  196,98,125,24,21,176,12,0,0         ; vbroadcastss  0xcb0(%rip),%ymm10        # 6780 <_sk_callback_avx+0x45e>
   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
@@ -10462,27 +10461,27 @@ _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  196,98,125,24,45,216,8,0,0          ; vbroadcastss  0x8d8(%rip),%ymm13        # 6788 <_sk_callback_avx+0x462>
+  DB  196,98,125,24,45,212,8,0,0          ; vbroadcastss  0x8d4(%rip),%ymm13        # 6784 <_sk_callback_avx+0x462>
   DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
-  DB  196,98,125,24,53,206,8,0,0          ; vbroadcastss  0x8ce(%rip),%ymm14        # 678c <_sk_callback_avx+0x466>
+  DB  196,98,125,24,53,202,8,0,0          ; vbroadcastss  0x8ca(%rip),%ymm14        # 6788 <_sk_callback_avx+0x466>
   DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
   DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
-  DB  196,98,125,24,53,191,8,0,0          ; vbroadcastss  0x8bf(%rip),%ymm14        # 6790 <_sk_callback_avx+0x46a>
+  DB  196,98,125,24,53,187,8,0,0          ; vbroadcastss  0x8bb(%rip),%ymm14        # 678c <_sk_callback_avx+0x46a>
   DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
   DB  196,65,28,89,229                    ; vmulps        %ymm13,%ymm12,%ymm12
-  DB  196,98,125,24,45,176,8,0,0          ; vbroadcastss  0x8b0(%rip),%ymm13        # 6794 <_sk_callback_avx+0x46e>
+  DB  196,98,125,24,45,172,8,0,0          ; vbroadcastss  0x8ac(%rip),%ymm13        # 6790 <_sk_callback_avx+0x46e>
   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  196,98,125,24,21,155,8,0,0          ; vbroadcastss  0x89b(%rip),%ymm10        # 6798 <_sk_callback_avx+0x472>
+  DB  196,98,125,24,21,151,8,0,0          ; vbroadcastss  0x897(%rip),%ymm10        # 6794 <_sk_callback_avx+0x472>
   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  196,98,125,24,21,133,8,0,0          ; vbroadcastss  0x885(%rip),%ymm10        # 679c <_sk_callback_avx+0x476>
+  DB  196,98,125,24,21,129,8,0,0          ; vbroadcastss  0x881(%rip),%ymm10        # 6798 <_sk_callback_avx+0x476>
   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  196,98,125,24,21,111,8,0,0          ; vbroadcastss  0x86f(%rip),%ymm10        # 67a0 <_sk_callback_avx+0x47a>
+  DB  196,98,125,24,21,107,8,0,0          ; vbroadcastss  0x86b(%rip),%ymm10        # 679c <_sk_callback_avx+0x47a>
   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
@@ -10495,15 +10494,14 @@ _sk_xy_to_radius_avx LABEL PROC
   DB  197,252,89,192                      ; vmulps        %ymm0,%ymm0,%ymm0
   DB  197,116,89,193                      ; vmulps        %ymm1,%ymm1,%ymm8
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
-  DB  197,252,82,192                      ; vrsqrtps      %ymm0,%ymm0
-  DB  197,252,83,192                      ; vrcpps        %ymm0,%ymm0
+  DB  197,252,81,192                      ; vsqrtps       %ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_save_xy_avx
 _sk_save_xy_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,53,8,0,0            ; vbroadcastss  0x835(%rip),%ymm8        # 67a4 <_sk_callback_avx+0x47e>
+  DB  196,98,125,24,5,53,8,0,0            ; vbroadcastss  0x835(%rip),%ymm8        # 67a0 <_sk_callback_avx+0x47e>
   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
@@ -10536,9 +10534,9 @@ _sk_accumulate_avx LABEL PROC
 PUBLIC _sk_bilinear_nx_avx
 _sk_bilinear_nx_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,193,7,0,0          ; vbroadcastss  0x7c1(%rip),%ymm0        # 67a8 <_sk_callback_avx+0x482>
+  DB  196,226,125,24,5,193,7,0,0          ; vbroadcastss  0x7c1(%rip),%ymm0        # 67a4 <_sk_callback_avx+0x482>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,184,7,0,0           ; vbroadcastss  0x7b8(%rip),%ymm8        # 67ac <_sk_callback_avx+0x486>
+  DB  196,98,125,24,5,184,7,0,0           ; vbroadcastss  0x7b8(%rip),%ymm8        # 67a8 <_sk_callback_avx+0x486>
   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
@@ -10547,7 +10545,7 @@ _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  196,226,125,24,5,160,7,0,0          ; vbroadcastss  0x7a0(%rip),%ymm0        # 67b0 <_sk_callback_avx+0x48a>
+  DB  196,226,125,24,5,160,7,0,0          ; vbroadcastss  0x7a0(%rip),%ymm0        # 67ac <_sk_callback_avx+0x48a>
   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)
@@ -10557,9 +10555,9 @@ _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  196,226,125,24,13,132,7,0,0         ; vbroadcastss  0x784(%rip),%ymm1        # 67b4 <_sk_callback_avx+0x48e>
+  DB  196,226,125,24,13,132,7,0,0         ; vbroadcastss  0x784(%rip),%ymm1        # 67b0 <_sk_callback_avx+0x48e>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,122,7,0,0           ; vbroadcastss  0x77a(%rip),%ymm8        # 67b8 <_sk_callback_avx+0x492>
+  DB  196,98,125,24,5,122,7,0,0           ; vbroadcastss  0x77a(%rip),%ymm8        # 67b4 <_sk_callback_avx+0x492>
   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
@@ -10568,7 +10566,7 @@ _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  196,226,125,24,13,98,7,0,0          ; vbroadcastss  0x762(%rip),%ymm1        # 67bc <_sk_callback_avx+0x496>
+  DB  196,226,125,24,13,98,7,0,0          ; vbroadcastss  0x762(%rip),%ymm1        # 67b8 <_sk_callback_avx+0x496>
   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)
@@ -10578,14 +10576,14 @@ _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  196,226,125,24,5,69,7,0,0           ; vbroadcastss  0x745(%rip),%ymm0        # 67c0 <_sk_callback_avx+0x49a>
+  DB  196,226,125,24,5,69,7,0,0           ; vbroadcastss  0x745(%rip),%ymm0        # 67bc <_sk_callback_avx+0x49a>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,60,7,0,0            ; vbroadcastss  0x73c(%rip),%ymm8        # 67c4 <_sk_callback_avx+0x49e>
+  DB  196,98,125,24,5,60,7,0,0            ; vbroadcastss  0x73c(%rip),%ymm8        # 67c0 <_sk_callback_avx+0x49e>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,45,7,0,0           ; vbroadcastss  0x72d(%rip),%ymm10        # 67c8 <_sk_callback_avx+0x4a2>
+  DB  196,98,125,24,21,45,7,0,0           ; vbroadcastss  0x72d(%rip),%ymm10        # 67c4 <_sk_callback_avx+0x4a2>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,35,7,0,0           ; vbroadcastss  0x723(%rip),%ymm10        # 67cc <_sk_callback_avx+0x4a6>
+  DB  196,98,125,24,21,35,7,0,0           ; vbroadcastss  0x723(%rip),%ymm10        # 67c8 <_sk_callback_avx+0x4a6>
   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)
@@ -10595,19 +10593,19 @@ _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  196,226,125,24,5,6,7,0,0            ; vbroadcastss  0x706(%rip),%ymm0        # 67d0 <_sk_callback_avx+0x4aa>
+  DB  196,226,125,24,5,6,7,0,0            ; vbroadcastss  0x706(%rip),%ymm0        # 67cc <_sk_callback_avx+0x4aa>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,253,6,0,0           ; vbroadcastss  0x6fd(%rip),%ymm8        # 67d4 <_sk_callback_avx+0x4ae>
+  DB  196,98,125,24,5,253,6,0,0           ; vbroadcastss  0x6fd(%rip),%ymm8        # 67d0 <_sk_callback_avx+0x4ae>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,243,6,0,0          ; vbroadcastss  0x6f3(%rip),%ymm9        # 67d8 <_sk_callback_avx+0x4b2>
+  DB  196,98,125,24,13,243,6,0,0          ; vbroadcastss  0x6f3(%rip),%ymm9        # 67d4 <_sk_callback_avx+0x4b2>
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,233,6,0,0          ; vbroadcastss  0x6e9(%rip),%ymm10        # 67dc <_sk_callback_avx+0x4b6>
+  DB  196,98,125,24,21,233,6,0,0          ; vbroadcastss  0x6e9(%rip),%ymm10        # 67d8 <_sk_callback_avx+0x4b6>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,218,6,0,0          ; vbroadcastss  0x6da(%rip),%ymm10        # 67e0 <_sk_callback_avx+0x4ba>
+  DB  196,98,125,24,21,218,6,0,0          ; vbroadcastss  0x6da(%rip),%ymm10        # 67dc <_sk_callback_avx+0x4ba>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,203,6,0,0          ; vbroadcastss  0x6cb(%rip),%ymm9        # 67e4 <_sk_callback_avx+0x4be>
+  DB  196,98,125,24,13,203,6,0,0          ; vbroadcastss  0x6cb(%rip),%ymm9        # 67e0 <_sk_callback_avx+0x4be>
   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
@@ -10616,17 +10614,17 @@ _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  196,98,125,24,5,179,6,0,0           ; vbroadcastss  0x6b3(%rip),%ymm8        # 67e8 <_sk_callback_avx+0x4c2>
+  DB  196,98,125,24,5,179,6,0,0           ; vbroadcastss  0x6b3(%rip),%ymm8        # 67e4 <_sk_callback_avx+0x4c2>
   DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
   DB  197,124,16,72,64                    ; vmovups       0x40(%rax),%ymm9
-  DB  196,98,125,24,21,165,6,0,0          ; vbroadcastss  0x6a5(%rip),%ymm10        # 67ec <_sk_callback_avx+0x4c6>
+  DB  196,98,125,24,21,165,6,0,0          ; vbroadcastss  0x6a5(%rip),%ymm10        # 67e8 <_sk_callback_avx+0x4c6>
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
-  DB  196,98,125,24,29,155,6,0,0          ; vbroadcastss  0x69b(%rip),%ymm11        # 67f0 <_sk_callback_avx+0x4ca>
+  DB  196,98,125,24,29,155,6,0,0          ; vbroadcastss  0x69b(%rip),%ymm11        # 67ec <_sk_callback_avx+0x4ca>
   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  196,98,125,24,13,130,6,0,0          ; vbroadcastss  0x682(%rip),%ymm9        # 67f4 <_sk_callback_avx+0x4ce>
+  DB  196,98,125,24,13,130,6,0,0          ; vbroadcastss  0x682(%rip),%ymm9        # 67f0 <_sk_callback_avx+0x4ce>
   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
@@ -10635,13 +10633,13 @@ _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  196,226,125,24,5,106,6,0,0          ; vbroadcastss  0x66a(%rip),%ymm0        # 67f8 <_sk_callback_avx+0x4d2>
+  DB  196,226,125,24,5,106,6,0,0          ; vbroadcastss  0x66a(%rip),%ymm0        # 67f4 <_sk_callback_avx+0x4d2>
   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  196,98,125,24,21,87,6,0,0           ; vbroadcastss  0x657(%rip),%ymm10        # 67fc <_sk_callback_avx+0x4d6>
+  DB  196,98,125,24,21,87,6,0,0           ; vbroadcastss  0x657(%rip),%ymm10        # 67f8 <_sk_callback_avx+0x4d6>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,77,6,0,0           ; vbroadcastss  0x64d(%rip),%ymm10        # 6800 <_sk_callback_avx+0x4da>
+  DB  196,98,125,24,21,77,6,0,0           ; vbroadcastss  0x64d(%rip),%ymm10        # 67fc <_sk_callback_avx+0x4da>
   DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
   DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
@@ -10651,14 +10649,14 @@ _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  196,226,125,24,13,48,6,0,0          ; vbroadcastss  0x630(%rip),%ymm1        # 6804 <_sk_callback_avx+0x4de>
+  DB  196,226,125,24,13,48,6,0,0          ; vbroadcastss  0x630(%rip),%ymm1        # 6800 <_sk_callback_avx+0x4de>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,38,6,0,0            ; vbroadcastss  0x626(%rip),%ymm8        # 6808 <_sk_callback_avx+0x4e2>
+  DB  196,98,125,24,5,38,6,0,0            ; vbroadcastss  0x626(%rip),%ymm8        # 6804 <_sk_callback_avx+0x4e2>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,23,6,0,0           ; vbroadcastss  0x617(%rip),%ymm10        # 680c <_sk_callback_avx+0x4e6>
+  DB  196,98,125,24,21,23,6,0,0           ; vbroadcastss  0x617(%rip),%ymm10        # 6808 <_sk_callback_avx+0x4e6>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,13,6,0,0           ; vbroadcastss  0x60d(%rip),%ymm10        # 6810 <_sk_callback_avx+0x4ea>
+  DB  196,98,125,24,21,13,6,0,0           ; vbroadcastss  0x60d(%rip),%ymm10        # 680c <_sk_callback_avx+0x4ea>
   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)
@@ -10668,19 +10666,19 @@ _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  196,226,125,24,13,240,5,0,0         ; vbroadcastss  0x5f0(%rip),%ymm1        # 6814 <_sk_callback_avx+0x4ee>
+  DB  196,226,125,24,13,240,5,0,0         ; vbroadcastss  0x5f0(%rip),%ymm1        # 6810 <_sk_callback_avx+0x4ee>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,230,5,0,0           ; vbroadcastss  0x5e6(%rip),%ymm8        # 6818 <_sk_callback_avx+0x4f2>
+  DB  196,98,125,24,5,230,5,0,0           ; vbroadcastss  0x5e6(%rip),%ymm8        # 6814 <_sk_callback_avx+0x4f2>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,220,5,0,0          ; vbroadcastss  0x5dc(%rip),%ymm9        # 681c <_sk_callback_avx+0x4f6>
+  DB  196,98,125,24,13,220,5,0,0          ; vbroadcastss  0x5dc(%rip),%ymm9        # 6818 <_sk_callback_avx+0x4f6>
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,210,5,0,0          ; vbroadcastss  0x5d2(%rip),%ymm10        # 6820 <_sk_callback_avx+0x4fa>
+  DB  196,98,125,24,21,210,5,0,0          ; vbroadcastss  0x5d2(%rip),%ymm10        # 681c <_sk_callback_avx+0x4fa>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,195,5,0,0          ; vbroadcastss  0x5c3(%rip),%ymm10        # 6824 <_sk_callback_avx+0x4fe>
+  DB  196,98,125,24,21,195,5,0,0          ; vbroadcastss  0x5c3(%rip),%ymm10        # 6820 <_sk_callback_avx+0x4fe>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,180,5,0,0          ; vbroadcastss  0x5b4(%rip),%ymm9        # 6828 <_sk_callback_avx+0x502>
+  DB  196,98,125,24,13,180,5,0,0          ; vbroadcastss  0x5b4(%rip),%ymm9        # 6824 <_sk_callback_avx+0x502>
   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
@@ -10689,17 +10687,17 @@ _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  196,98,125,24,5,156,5,0,0           ; vbroadcastss  0x59c(%rip),%ymm8        # 682c <_sk_callback_avx+0x506>
+  DB  196,98,125,24,5,156,5,0,0           ; vbroadcastss  0x59c(%rip),%ymm8        # 6828 <_sk_callback_avx+0x506>
   DB  197,188,88,72,32                    ; vaddps        0x20(%rax),%ymm8,%ymm1
   DB  197,124,16,72,96                    ; vmovups       0x60(%rax),%ymm9
-  DB  196,98,125,24,21,141,5,0,0          ; vbroadcastss  0x58d(%rip),%ymm10        # 6830 <_sk_callback_avx+0x50a>
+  DB  196,98,125,24,21,141,5,0,0          ; vbroadcastss  0x58d(%rip),%ymm10        # 682c <_sk_callback_avx+0x50a>
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
-  DB  196,98,125,24,29,131,5,0,0          ; vbroadcastss  0x583(%rip),%ymm11        # 6834 <_sk_callback_avx+0x50e>
+  DB  196,98,125,24,29,131,5,0,0          ; vbroadcastss  0x583(%rip),%ymm11        # 6830 <_sk_callback_avx+0x50e>
   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  196,98,125,24,13,106,5,0,0          ; vbroadcastss  0x56a(%rip),%ymm9        # 6838 <_sk_callback_avx+0x512>
+  DB  196,98,125,24,13,106,5,0,0          ; vbroadcastss  0x56a(%rip),%ymm9        # 6834 <_sk_callback_avx+0x512>
   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
@@ -10708,13 +10706,13 @@ _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  196,226,125,24,13,82,5,0,0          ; vbroadcastss  0x552(%rip),%ymm1        # 683c <_sk_callback_avx+0x516>
+  DB  196,226,125,24,13,82,5,0,0          ; vbroadcastss  0x552(%rip),%ymm1        # 6838 <_sk_callback_avx+0x516>
   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  196,98,125,24,21,62,5,0,0           ; vbroadcastss  0x53e(%rip),%ymm10        # 6840 <_sk_callback_avx+0x51a>
+  DB  196,98,125,24,21,62,5,0,0           ; vbroadcastss  0x53e(%rip),%ymm10        # 683c <_sk_callback_avx+0x51a>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,52,5,0,0           ; vbroadcastss  0x534(%rip),%ymm10        # 6844 <_sk_callback_avx+0x51e>
+  DB  196,98,125,24,21,52,5,0,0           ; vbroadcastss  0x534(%rip),%ymm10        # 6840 <_sk_callback_avx+0x51e>
   DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
   DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
@@ -10828,25 +10826,25 @@ 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 64f1 <.literal4+0xb1>
+  DB  71,225,61                           ; rex.RXB       loope 64ed <.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 6501 <.literal4+0xc1>
+  DB  71,225,61                           ; rex.RXB       loope 64fd <.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 6511 <.literal4+0xd1>
+  DB  71,225,61                           ; rex.RXB       loope 650d <.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 6521 <.literal4+0xe1>
+  DB  71,225,61                           ; rex.RXB       loope 651d <.literal4+0xe1>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -10896,7 +10894,7 @@ 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        656d <.literal4+0x12d>
+  DB  224,7                               ; loopne        6569 <.literal4+0x12d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -10912,10 +10910,10 @@ 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            6594 <.literal4+0x154>
+  DB  127,0                               ; jg            6590 <.literal4+0x154>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            660d <.literal4+0x1cd>
+  DB  119,115                             ; ja            6609 <.literal4+0x1cd>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -10929,10 +10927,10 @@ 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            65c8 <.literal4+0x188>
+  DB  127,0                               ; jg            65c4 <.literal4+0x188>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            6641 <.literal4+0x201>
+  DB  119,115                             ; ja            663d <.literal4+0x201>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -10946,10 +10944,10 @@ 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            65fc <.literal4+0x1bc>
+  DB  127,0                               ; jg            65f8 <.literal4+0x1bc>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            6675 <.literal4+0x235>
+  DB  119,115                             ; ja            6671 <.literal4+0x235>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -10963,10 +10961,10 @@ 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            6630 <.literal4+0x1f0>
+  DB  127,0                               ; jg            662c <.literal4+0x1f0>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            66a9 <.literal4+0x269>
+  DB  119,115                             ; ja            66a5 <.literal4+0x269>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -10979,7 +10977,7 @@ 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            66a7 <.literal4+0x267>
+  DB  127,67                              ; jg            66a3 <.literal4+0x267>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,195                               ; add           %al,%bl
   DB  0,0                                 ; add           %al,(%rax)
@@ -10991,10 +10989,10 @@ ALIGN 4
   DB  190,80,128,3,62                     ; mov           $0x3e038050,%esi
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           66c7 <.literal4+0x287>
+  DB  118,63                              ; jbe           66c3 <.literal4+0x287>
   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            66db <.literal4+0x29b>
+  DB  127,67                              ; jg            66d7 <.literal4+0x29b>
   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)
@@ -11003,7 +11001,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        66bd <.literal4+0x27d>
+  DB  224,7                               ; loopne        66b9 <.literal4+0x27d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -11015,7 +11013,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        66d9 <.literal4+0x299>
+  DB  224,7                               ; loopne        66d5 <.literal4+0x299>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -11026,7 +11024,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            672e <.literal4+0x2ee>
+  DB  124,66                              ; jl            672a <.literal4+0x2ee>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  137,136,136,55,0,15                 ; mov           %ecx,0xf003788(%rax)
@@ -11044,9 +11042,9 @@ 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            6771 <.literal4+0x331>
+  DB  112,65                              ; jo            676d <.literal4+0x331>
   DB  129,128,128,59,129,128,128,59,0,0   ; addl          $0x3b80,-0x7f7ec480(%rax)
-  DB  127,67                              ; jg            677f <.literal4+0x33f>
+  DB  127,67                              ; jg            677b <.literal4+0x33f>
   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)
@@ -11062,7 +11060,7 @@ 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            67bf <.literal4+0x37f>
+  DB  127,71                              ; jg            67bb <.literal4+0x37f>
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,89                               ; ds            pop %rcx
@@ -11313,7 +11311,7 @@ _sk_seed_shader_sse41 LABEL PROC
   DB  102,15,110,199                      ; movd          %edi,%xmm0
   DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
   DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
-  DB  15,40,21,113,70,0,0                 ; movaps        0x4671(%rip),%xmm2        # 4780 <_sk_callback_sse41+0xad>
+  DB  15,40,21,113,70,0,0                 ; movaps        0x4671(%rip),%xmm2        # 4780 <_sk_callback_sse41+0xb0>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  15,16,2                             ; movups        (%rdx),%xmm0
   DB  15,88,193                           ; addps         %xmm1,%xmm0
@@ -11322,7 +11320,7 @@ _sk_seed_shader_sse41 LABEL PROC
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,21,96,70,0,0                  ; movaps        0x4660(%rip),%xmm2        # 4790 <_sk_callback_sse41+0xbd>
+  DB  15,40,21,96,70,0,0                  ; movaps        0x4660(%rip),%xmm2        # 4790 <_sk_callback_sse41+0xc0>
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
@@ -11343,14 +11341,14 @@ _sk_dither_sse41 LABEL PROC
   DB  102,68,15,110,1                     ; movd          (%rcx),%xmm8
   DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
   DB  102,69,15,239,193                   ; pxor          %xmm9,%xmm8
-  DB  102,68,15,111,21,37,70,0,0          ; movdqa        0x4625(%rip),%xmm10        # 47a0 <_sk_callback_sse41+0xcd>
+  DB  102,68,15,111,21,37,70,0,0          ; movdqa        0x4625(%rip),%xmm10        # 47a0 <_sk_callback_sse41+0xd0>
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,219,218                   ; pand          %xmm10,%xmm11
   DB  102,65,15,114,243,5                 ; pslld         $0x5,%xmm11
   DB  102,69,15,219,209                   ; pand          %xmm9,%xmm10
   DB  102,65,15,114,242,4                 ; pslld         $0x4,%xmm10
-  DB  102,68,15,111,37,17,70,0,0          ; movdqa        0x4611(%rip),%xmm12        # 47b0 <_sk_callback_sse41+0xdd>
-  DB  102,68,15,111,45,24,70,0,0          ; movdqa        0x4618(%rip),%xmm13        # 47c0 <_sk_callback_sse41+0xed>
+  DB  102,68,15,111,37,17,70,0,0          ; movdqa        0x4611(%rip),%xmm12        # 47b0 <_sk_callback_sse41+0xe0>
+  DB  102,68,15,111,45,24,70,0,0          ; movdqa        0x4618(%rip),%xmm13        # 47c0 <_sk_callback_sse41+0xf0>
   DB  102,69,15,111,240                   ; movdqa        %xmm8,%xmm14
   DB  102,69,15,219,245                   ; pand          %xmm13,%xmm14
   DB  102,65,15,114,246,2                 ; pslld         $0x2,%xmm14
@@ -11366,8 +11364,8 @@ _sk_dither_sse41 LABEL PROC
   DB  102,69,15,235,245                   ; por           %xmm13,%xmm14
   DB  102,69,15,235,240                   ; por           %xmm8,%xmm14
   DB  69,15,91,198                        ; cvtdq2ps      %xmm14,%xmm8
-  DB  68,15,89,5,211,69,0,0               ; mulps         0x45d3(%rip),%xmm8        # 47d0 <_sk_callback_sse41+0xfd>
-  DB  68,15,88,5,219,69,0,0               ; addps         0x45db(%rip),%xmm8        # 47e0 <_sk_callback_sse41+0x10d>
+  DB  68,15,89,5,211,69,0,0               ; mulps         0x45d3(%rip),%xmm8        # 47d0 <_sk_callback_sse41+0x100>
+  DB  68,15,88,5,219,69,0,0               ; addps         0x45db(%rip),%xmm8        # 47e0 <_sk_callback_sse41+0x110>
   DB  243,68,15,16,72,8                   ; movss         0x8(%rax),%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
@@ -11423,7 +11421,7 @@ _sk_clear_sse41 LABEL PROC
 PUBLIC _sk_srcatop_sse41
 _sk_srcatop_sse41 LABEL PROC
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  68,15,40,5,94,69,0,0                ; movaps        0x455e(%rip),%xmm8        # 47f0 <_sk_callback_sse41+0x11d>
+  DB  68,15,40,5,94,69,0,0                ; movaps        0x455e(%rip),%xmm8        # 47f0 <_sk_callback_sse41+0x120>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -11446,7 +11444,7 @@ PUBLIC _sk_dstatop_sse41
 _sk_dstatop_sse41 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
   DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
-  DB  68,15,40,13,33,69,0,0               ; movaps        0x4521(%rip),%xmm9        # 4800 <_sk_callback_sse41+0x12d>
+  DB  68,15,40,13,33,69,0,0               ; movaps        0x4521(%rip),%xmm9        # 4800 <_sk_callback_sse41+0x130>
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
@@ -11487,7 +11485,7 @@ _sk_dstin_sse41 LABEL PROC
 
 PUBLIC _sk_srcout_sse41
 _sk_srcout_sse41 LABEL PROC
-  DB  68,15,40,5,197,68,0,0               ; movaps        0x44c5(%rip),%xmm8        # 4810 <_sk_callback_sse41+0x13d>
+  DB  68,15,40,5,197,68,0,0               ; movaps        0x44c5(%rip),%xmm8        # 4810 <_sk_callback_sse41+0x140>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
@@ -11498,7 +11496,7 @@ _sk_srcout_sse41 LABEL PROC
 
 PUBLIC _sk_dstout_sse41
 _sk_dstout_sse41 LABEL PROC
-  DB  68,15,40,5,181,68,0,0               ; movaps        0x44b5(%rip),%xmm8        # 4820 <_sk_callback_sse41+0x14d>
+  DB  68,15,40,5,181,68,0,0               ; movaps        0x44b5(%rip),%xmm8        # 4820 <_sk_callback_sse41+0x150>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -11513,7 +11511,7 @@ _sk_dstout_sse41 LABEL PROC
 
 PUBLIC _sk_srcover_sse41
 _sk_srcover_sse41 LABEL PROC
-  DB  68,15,40,5,152,68,0,0               ; movaps        0x4498(%rip),%xmm8        # 4830 <_sk_callback_sse41+0x15d>
+  DB  68,15,40,5,152,68,0,0               ; movaps        0x4498(%rip),%xmm8        # 4830 <_sk_callback_sse41+0x160>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -11531,7 +11529,7 @@ _sk_srcover_sse41 LABEL PROC
 
 PUBLIC _sk_dstover_sse41
 _sk_dstover_sse41 LABEL PROC
-  DB  68,15,40,5,108,68,0,0               ; movaps        0x446c(%rip),%xmm8        # 4840 <_sk_callback_sse41+0x16d>
+  DB  68,15,40,5,108,68,0,0               ; movaps        0x446c(%rip),%xmm8        # 4840 <_sk_callback_sse41+0x170>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -11555,7 +11553,7 @@ _sk_modulate_sse41 LABEL PROC
 
 PUBLIC _sk_multiply_sse41
 _sk_multiply_sse41 LABEL PROC
-  DB  68,15,40,5,64,68,0,0                ; movaps        0x4440(%rip),%xmm8        # 4850 <_sk_callback_sse41+0x17d>
+  DB  68,15,40,5,64,68,0,0                ; movaps        0x4440(%rip),%xmm8        # 4850 <_sk_callback_sse41+0x180>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
@@ -11625,7 +11623,7 @@ _sk_screen_sse41 LABEL PROC
 PUBLIC _sk_xor__sse41
 _sk_xor__sse41 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
-  DB  15,40,29,113,67,0,0                 ; movaps        0x4371(%rip),%xmm3        # 4860 <_sk_callback_sse41+0x18d>
+  DB  15,40,29,113,67,0,0                 ; movaps        0x4371(%rip),%xmm3        # 4860 <_sk_callback_sse41+0x190>
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
@@ -11671,7 +11669,7 @@ _sk_darken_sse41 LABEL PROC
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,220,66,0,0                 ; movaps        0x42dc(%rip),%xmm2        # 4870 <_sk_callback_sse41+0x19d>
+  DB  15,40,21,220,66,0,0                 ; movaps        0x42dc(%rip),%xmm2        # 4870 <_sk_callback_sse41+0x1a0>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -11703,7 +11701,7 @@ _sk_lighten_sse41 LABEL PROC
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,129,66,0,0                 ; movaps        0x4281(%rip),%xmm2        # 4880 <_sk_callback_sse41+0x1ad>
+  DB  15,40,21,129,66,0,0                 ; movaps        0x4281(%rip),%xmm2        # 4880 <_sk_callback_sse41+0x1b0>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -11738,7 +11736,7 @@ _sk_difference_sse41 LABEL PROC
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,27,66,0,0                  ; movaps        0x421b(%rip),%xmm2        # 4890 <_sk_callback_sse41+0x1bd>
+  DB  15,40,21,27,66,0,0                  ; movaps        0x421b(%rip),%xmm2        # 4890 <_sk_callback_sse41+0x1c0>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -11763,7 +11761,7 @@ _sk_exclusion_sse41 LABEL PROC
   DB  15,89,214                           ; mulps         %xmm6,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,202                        ; subps         %xmm2,%xmm9
-  DB  15,40,13,220,65,0,0                 ; movaps        0x41dc(%rip),%xmm1        # 48a0 <_sk_callback_sse41+0x1cd>
+  DB  15,40,13,220,65,0,0                 ; movaps        0x41dc(%rip),%xmm1        # 48a0 <_sk_callback_sse41+0x1d0>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,207                           ; mulps         %xmm7,%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -11775,7 +11773,7 @@ _sk_exclusion_sse41 LABEL PROC
 PUBLIC _sk_colorburn_sse41
 _sk_colorburn_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,203,65,0,0              ; movaps        0x41cb(%rip),%xmm10        # 48b0 <_sk_callback_sse41+0x1dd>
+  DB  68,15,40,21,203,65,0,0              ; movaps        0x41cb(%rip),%xmm10        # 48b0 <_sk_callback_sse41+0x1e0>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
@@ -11855,7 +11853,7 @@ _sk_colorburn_sse41 LABEL PROC
 PUBLIC _sk_colordodge_sse41
 _sk_colordodge_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,169,64,0,0              ; movaps        0x40a9(%rip),%xmm10        # 48c0 <_sk_callback_sse41+0x1ed>
+  DB  68,15,40,21,169,64,0,0              ; movaps        0x40a9(%rip),%xmm10        # 48c0 <_sk_callback_sse41+0x1f0>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -11936,7 +11934,7 @@ _sk_hardlight_sse41 LABEL PROC
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
-  DB  68,15,40,21,127,63,0,0              ; movaps        0x3f7f(%rip),%xmm10        # 48d0 <_sk_callback_sse41+0x1fd>
+  DB  68,15,40,21,127,63,0,0              ; movaps        0x3f7f(%rip),%xmm10        # 48d0 <_sk_callback_sse41+0x200>
   DB  65,15,40,234                        ; movaps        %xmm10,%xmm5
   DB  15,92,239                           ; subps         %xmm7,%xmm5
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
@@ -12018,7 +12016,7 @@ PUBLIC _sk_overlay_sse41
 _sk_overlay_sse41 LABEL PROC
   DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
-  DB  68,15,40,21,97,62,0,0               ; movaps        0x3e61(%rip),%xmm10        # 48e0 <_sk_callback_sse41+0x20d>
+  DB  68,15,40,21,97,62,0,0               ; movaps        0x3e61(%rip),%xmm10        # 48e0 <_sk_callback_sse41+0x210>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
@@ -12102,7 +12100,7 @@ _sk_softlight_sse41 LABEL PROC
   DB  15,40,198                           ; movaps        %xmm6,%xmm0
   DB  15,94,199                           ; divps         %xmm7,%xmm0
   DB  65,15,84,193                        ; andps         %xmm9,%xmm0
-  DB  15,40,13,52,61,0,0                  ; movaps        0x3d34(%rip),%xmm1        # 48f0 <_sk_callback_sse41+0x21d>
+  DB  15,40,13,52,61,0,0                  ; movaps        0x3d34(%rip),%xmm1        # 48f0 <_sk_callback_sse41+0x220>
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  68,15,92,208                        ; subps         %xmm0,%xmm10
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
@@ -12115,10 +12113,10 @@ _sk_softlight_sse41 LABEL PROC
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,89,210                           ; mulps         %xmm2,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
-  DB  68,15,40,45,18,61,0,0               ; movaps        0x3d12(%rip),%xmm13        # 4900 <_sk_callback_sse41+0x22d>
+  DB  68,15,40,45,18,61,0,0               ; movaps        0x3d12(%rip),%xmm13        # 4900 <_sk_callback_sse41+0x230>
   DB  69,15,88,245                        ; addps         %xmm13,%xmm14
   DB  68,15,89,242                        ; mulps         %xmm2,%xmm14
-  DB  68,15,40,37,18,61,0,0               ; movaps        0x3d12(%rip),%xmm12        # 4910 <_sk_callback_sse41+0x23d>
+  DB  68,15,40,37,18,61,0,0               ; movaps        0x3d12(%rip),%xmm12        # 4910 <_sk_callback_sse41+0x240>
   DB  69,15,89,252                        ; mulps         %xmm12,%xmm15
   DB  69,15,88,254                        ; addps         %xmm14,%xmm15
   DB  15,40,198                           ; movaps        %xmm6,%xmm0
@@ -12304,12 +12302,12 @@ _sk_hue_sse41 LABEL PROC
   DB  68,15,84,208                        ; andps         %xmm0,%xmm10
   DB  15,84,200                           ; andps         %xmm0,%xmm1
   DB  68,15,84,232                        ; andps         %xmm0,%xmm13
-  DB  15,40,5,120,58,0,0                  ; movaps        0x3a78(%rip),%xmm0        # 4920 <_sk_callback_sse41+0x24d>
+  DB  15,40,5,120,58,0,0                  ; movaps        0x3a78(%rip),%xmm0        # 4920 <_sk_callback_sse41+0x250>
   DB  68,15,89,224                        ; mulps         %xmm0,%xmm12
-  DB  15,40,21,125,58,0,0                 ; movaps        0x3a7d(%rip),%xmm2        # 4930 <_sk_callback_sse41+0x25d>
+  DB  15,40,21,125,58,0,0                 ; movaps        0x3a7d(%rip),%xmm2        # 4930 <_sk_callback_sse41+0x260>
   DB  15,89,250                           ; mulps         %xmm2,%xmm7
   DB  65,15,88,252                        ; addps         %xmm12,%xmm7
-  DB  68,15,40,53,126,58,0,0              ; movaps        0x3a7e(%rip),%xmm14        # 4940 <_sk_callback_sse41+0x26d>
+  DB  68,15,40,53,126,58,0,0              ; movaps        0x3a7e(%rip),%xmm14        # 4940 <_sk_callback_sse41+0x270>
   DB  68,15,40,252                        ; movaps        %xmm4,%xmm15
   DB  69,15,89,254                        ; mulps         %xmm14,%xmm15
   DB  68,15,88,255                        ; addps         %xmm7,%xmm15
@@ -12392,7 +12390,7 @@ _sk_hue_sse41 LABEL PROC
   DB  65,15,88,214                        ; addps         %xmm14,%xmm2
   DB  15,40,196                           ; movaps        %xmm4,%xmm0
   DB  102,15,56,20,202                    ; blendvps      %xmm0,%xmm2,%xmm1
-  DB  68,15,40,13,67,57,0,0               ; movaps        0x3943(%rip),%xmm9        # 4950 <_sk_callback_sse41+0x27d>
+  DB  68,15,40,13,67,57,0,0               ; movaps        0x3943(%rip),%xmm9        # 4950 <_sk_callback_sse41+0x280>
   DB  65,15,40,225                        ; movaps        %xmm9,%xmm4
   DB  15,92,229                           ; subps         %xmm5,%xmm4
   DB  15,40,68,36,48                      ; movaps        0x30(%rsp),%xmm0
@@ -12486,14 +12484,14 @@ _sk_saturation_sse41 LABEL PROC
   DB  68,15,84,215                        ; andps         %xmm7,%xmm10
   DB  68,15,84,223                        ; andps         %xmm7,%xmm11
   DB  68,15,84,199                        ; andps         %xmm7,%xmm8
-  DB  15,40,21,246,55,0,0                 ; movaps        0x37f6(%rip),%xmm2        # 4960 <_sk_callback_sse41+0x28d>
+  DB  15,40,21,246,55,0,0                 ; movaps        0x37f6(%rip),%xmm2        # 4960 <_sk_callback_sse41+0x290>
   DB  15,40,221                           ; movaps        %xmm5,%xmm3
   DB  15,89,218                           ; mulps         %xmm2,%xmm3
-  DB  15,40,13,249,55,0,0                 ; movaps        0x37f9(%rip),%xmm1        # 4970 <_sk_callback_sse41+0x29d>
+  DB  15,40,13,249,55,0,0                 ; movaps        0x37f9(%rip),%xmm1        # 4970 <_sk_callback_sse41+0x2a0>
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
   DB  15,89,249                           ; mulps         %xmm1,%xmm7
   DB  15,88,251                           ; addps         %xmm3,%xmm7
-  DB  68,15,40,45,248,55,0,0              ; movaps        0x37f8(%rip),%xmm13        # 4980 <_sk_callback_sse41+0x2ad>
+  DB  68,15,40,45,248,55,0,0              ; movaps        0x37f8(%rip),%xmm13        # 4980 <_sk_callback_sse41+0x2b0>
   DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
   DB  68,15,88,247                        ; addps         %xmm7,%xmm14
   DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
@@ -12574,7 +12572,7 @@ _sk_saturation_sse41 LABEL PROC
   DB  65,15,88,253                        ; addps         %xmm13,%xmm7
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  102,68,15,56,20,223                 ; blendvps      %xmm0,%xmm7,%xmm11
-  DB  68,15,40,13,190,54,0,0              ; movaps        0x36be(%rip),%xmm9        # 4990 <_sk_callback_sse41+0x2bd>
+  DB  68,15,40,13,190,54,0,0              ; movaps        0x36be(%rip),%xmm9        # 4990 <_sk_callback_sse41+0x2c0>
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  68,15,92,204                        ; subps         %xmm4,%xmm9
   DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
@@ -12629,14 +12627,14 @@ _sk_color_sse41 LABEL PROC
   DB  15,40,231                           ; movaps        %xmm7,%xmm4
   DB  68,15,89,244                        ; mulps         %xmm4,%xmm14
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
-  DB  68,15,40,13,3,54,0,0                ; movaps        0x3603(%rip),%xmm9        # 49a0 <_sk_callback_sse41+0x2cd>
+  DB  68,15,40,13,3,54,0,0                ; movaps        0x3603(%rip),%xmm9        # 49a0 <_sk_callback_sse41+0x2d0>
   DB  65,15,40,250                        ; movaps        %xmm10,%xmm7
   DB  65,15,89,249                        ; mulps         %xmm9,%xmm7
-  DB  68,15,40,21,3,54,0,0                ; movaps        0x3603(%rip),%xmm10        # 49b0 <_sk_callback_sse41+0x2dd>
+  DB  68,15,40,21,3,54,0,0                ; movaps        0x3603(%rip),%xmm10        # 49b0 <_sk_callback_sse41+0x2e0>
   DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
   DB  65,15,89,218                        ; mulps         %xmm10,%xmm3
   DB  15,88,223                           ; addps         %xmm7,%xmm3
-  DB  68,15,40,29,0,54,0,0                ; movaps        0x3600(%rip),%xmm11        # 49c0 <_sk_callback_sse41+0x2ed>
+  DB  68,15,40,29,0,54,0,0                ; movaps        0x3600(%rip),%xmm11        # 49c0 <_sk_callback_sse41+0x2f0>
   DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
   DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
   DB  68,15,88,235                        ; addps         %xmm3,%xmm13
@@ -12721,7 +12719,7 @@ _sk_color_sse41 LABEL PROC
   DB  65,15,88,251                        ; addps         %xmm11,%xmm7
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  102,15,56,20,207                    ; blendvps      %xmm0,%xmm7,%xmm1
-  DB  68,15,40,13,188,52,0,0              ; movaps        0x34bc(%rip),%xmm9        # 49d0 <_sk_callback_sse41+0x2fd>
+  DB  68,15,40,13,188,52,0,0              ; movaps        0x34bc(%rip),%xmm9        # 49d0 <_sk_callback_sse41+0x300>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
@@ -12773,13 +12771,13 @@ _sk_luminosity_sse41 LABEL PROC
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
-  DB  68,15,40,5,14,52,0,0                ; movaps        0x340e(%rip),%xmm8        # 49e0 <_sk_callback_sse41+0x30d>
+  DB  68,15,40,5,14,52,0,0                ; movaps        0x340e(%rip),%xmm8        # 49e0 <_sk_callback_sse41+0x310>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
-  DB  68,15,40,21,18,52,0,0               ; movaps        0x3412(%rip),%xmm10        # 49f0 <_sk_callback_sse41+0x31d>
+  DB  68,15,40,21,18,52,0,0               ; movaps        0x3412(%rip),%xmm10        # 49f0 <_sk_callback_sse41+0x320>
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
   DB  65,15,89,234                        ; mulps         %xmm10,%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
-  DB  68,15,40,37,16,52,0,0               ; movaps        0x3410(%rip),%xmm12        # 4a00 <_sk_callback_sse41+0x32d>
+  DB  68,15,40,37,16,52,0,0               ; movaps        0x3410(%rip),%xmm12        # 4a00 <_sk_callback_sse41+0x330>
   DB  68,15,40,242                        ; movaps        %xmm2,%xmm14
   DB  69,15,89,244                        ; mulps         %xmm12,%xmm14
   DB  68,15,88,245                        ; addps         %xmm5,%xmm14
@@ -12864,7 +12862,7 @@ _sk_luminosity_sse41 LABEL PROC
   DB  65,15,88,244                        ; addps         %xmm12,%xmm6
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  102,68,15,56,20,206                 ; blendvps      %xmm0,%xmm6,%xmm9
-  DB  15,40,5,198,50,0,0                  ; movaps        0x32c6(%rip),%xmm0        # 4a10 <_sk_callback_sse41+0x33d>
+  DB  15,40,5,198,50,0,0                  ; movaps        0x32c6(%rip),%xmm0        # 4a10 <_sk_callback_sse41+0x340>
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,92,215                           ; subps         %xmm7,%xmm2
   DB  15,89,226                           ; mulps         %xmm2,%xmm4
@@ -12910,7 +12908,7 @@ _sk_clamp_0_sse41 LABEL PROC
 
 PUBLIC _sk_clamp_1_sse41
 _sk_clamp_1_sse41 LABEL PROC
-  DB  68,15,40,5,70,50,0,0                ; movaps        0x3246(%rip),%xmm8        # 4a20 <_sk_callback_sse41+0x34d>
+  DB  68,15,40,5,70,50,0,0                ; movaps        0x3246(%rip),%xmm8        # 4a20 <_sk_callback_sse41+0x350>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
   DB  65,15,93,208                        ; minps         %xmm8,%xmm2
@@ -12920,7 +12918,7 @@ _sk_clamp_1_sse41 LABEL PROC
 
 PUBLIC _sk_clamp_a_sse41
 _sk_clamp_a_sse41 LABEL PROC
-  DB  15,93,29,59,50,0,0                  ; minps         0x323b(%rip),%xmm3        # 4a30 <_sk_callback_sse41+0x35d>
+  DB  15,93,29,59,50,0,0                  ; minps         0x323b(%rip),%xmm3        # 4a30 <_sk_callback_sse41+0x360>
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -12993,7 +12991,7 @@ _sk_premul_sse41 LABEL PROC
 PUBLIC _sk_unpremul_sse41
 _sk_unpremul_sse41 LABEL PROC
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
-  DB  68,15,40,13,166,49,0,0              ; movaps        0x31a6(%rip),%xmm9        # 4a40 <_sk_callback_sse41+0x36d>
+  DB  68,15,40,13,166,49,0,0              ; movaps        0x31a6(%rip),%xmm9        # 4a40 <_sk_callback_sse41+0x370>
   DB  68,15,94,203                        ; divps         %xmm3,%xmm9
   DB  68,15,194,195,4                     ; cmpneqps      %xmm3,%xmm8
   DB  69,15,84,193                        ; andps         %xmm9,%xmm8
@@ -13005,20 +13003,20 @@ _sk_unpremul_sse41 LABEL PROC
 
 PUBLIC _sk_from_srgb_sse41
 _sk_from_srgb_sse41 LABEL PROC
-  DB  68,15,40,29,145,49,0,0              ; movaps        0x3191(%rip),%xmm11        # 4a50 <_sk_callback_sse41+0x37d>
+  DB  68,15,40,29,145,49,0,0              ; movaps        0x3191(%rip),%xmm11        # 4a50 <_sk_callback_sse41+0x380>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
   DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
   DB  69,15,89,210                        ; mulps         %xmm10,%xmm10
-  DB  68,15,40,37,137,49,0,0              ; movaps        0x3189(%rip),%xmm12        # 4a60 <_sk_callback_sse41+0x38d>
+  DB  68,15,40,37,137,49,0,0              ; movaps        0x3189(%rip),%xmm12        # 4a60 <_sk_callback_sse41+0x390>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
-  DB  68,15,40,45,137,49,0,0              ; movaps        0x3189(%rip),%xmm13        # 4a70 <_sk_callback_sse41+0x39d>
+  DB  68,15,40,45,137,49,0,0              ; movaps        0x3189(%rip),%xmm13        # 4a70 <_sk_callback_sse41+0x3a0>
   DB  69,15,88,197                        ; addps         %xmm13,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
-  DB  68,15,40,53,137,49,0,0              ; movaps        0x3189(%rip),%xmm14        # 4a80 <_sk_callback_sse41+0x3ad>
+  DB  68,15,40,53,137,49,0,0              ; movaps        0x3189(%rip),%xmm14        # 4a80 <_sk_callback_sse41+0x3b0>
   DB  69,15,88,198                        ; addps         %xmm14,%xmm8
-  DB  68,15,40,61,141,49,0,0              ; movaps        0x318d(%rip),%xmm15        # 4a90 <_sk_callback_sse41+0x3bd>
+  DB  68,15,40,61,141,49,0,0              ; movaps        0x318d(%rip),%xmm15        # 4a90 <_sk_callback_sse41+0x3c0>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
@@ -13062,20 +13060,20 @@ _sk_to_srgb_sse41 LABEL PROC
   DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
   DB  69,15,83,200                        ; rcpps         %xmm8,%xmm9
   DB  69,15,82,208                        ; rsqrtps       %xmm8,%xmm10
-  DB  68,15,40,29,250,48,0,0              ; movaps        0x30fa(%rip),%xmm11        # 4aa0 <_sk_callback_sse41+0x3cd>
+  DB  68,15,40,29,250,48,0,0              ; movaps        0x30fa(%rip),%xmm11        # 4aa0 <_sk_callback_sse41+0x3d0>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
-  DB  68,15,40,37,251,48,0,0              ; movaps        0x30fb(%rip),%xmm12        # 4ab0 <_sk_callback_sse41+0x3dd>
+  DB  68,15,40,37,251,48,0,0              ; movaps        0x30fb(%rip),%xmm12        # 4ab0 <_sk_callback_sse41+0x3e0>
   DB  69,15,89,204                        ; mulps         %xmm12,%xmm9
-  DB  68,15,40,45,255,48,0,0              ; movaps        0x30ff(%rip),%xmm13        # 4ac0 <_sk_callback_sse41+0x3ed>
+  DB  68,15,40,45,255,48,0,0              ; movaps        0x30ff(%rip),%xmm13        # 4ac0 <_sk_callback_sse41+0x3f0>
   DB  69,15,88,205                        ; addps         %xmm13,%xmm9
-  DB  68,15,40,53,3,49,0,0                ; movaps        0x3103(%rip),%xmm14        # 4ad0 <_sk_callback_sse41+0x3fd>
+  DB  68,15,40,53,3,49,0,0                ; movaps        0x3103(%rip),%xmm14        # 4ad0 <_sk_callback_sse41+0x400>
   DB  69,15,89,214                        ; mulps         %xmm14,%xmm10
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
-  DB  68,15,40,5,3,49,0,0                 ; movaps        0x3103(%rip),%xmm8        # 4ae0 <_sk_callback_sse41+0x40d>
+  DB  68,15,40,5,3,49,0,0                 ; movaps        0x3103(%rip),%xmm8        # 4ae0 <_sk_callback_sse41+0x410>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,93,202                        ; minps         %xmm10,%xmm9
-  DB  68,15,40,61,3,49,0,0                ; movaps        0x3103(%rip),%xmm15        # 4af0 <_sk_callback_sse41+0x41d>
+  DB  68,15,40,61,3,49,0,0                ; movaps        0x3103(%rip),%xmm15        # 4af0 <_sk_callback_sse41+0x420>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  102,68,15,56,20,201                 ; blendvps      %xmm0,%xmm1,%xmm9
   DB  15,82,194                           ; rsqrtps       %xmm2,%xmm0
@@ -13128,7 +13126,7 @@ _sk_rgb_to_hsl_sse41 LABEL PROC
   DB  68,15,93,226                        ; minps         %xmm2,%xmm12
   DB  65,15,40,203                        ; movaps        %xmm11,%xmm1
   DB  65,15,92,204                        ; subps         %xmm12,%xmm1
-  DB  68,15,40,53,81,48,0,0               ; movaps        0x3051(%rip),%xmm14        # 4b00 <_sk_callback_sse41+0x42d>
+  DB  68,15,40,53,81,48,0,0               ; movaps        0x3051(%rip),%xmm14        # 4b00 <_sk_callback_sse41+0x430>
   DB  68,15,94,241                        ; divps         %xmm1,%xmm14
   DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
   DB  69,15,194,208,0                     ; cmpeqps       %xmm8,%xmm10
@@ -13137,27 +13135,27 @@ _sk_rgb_to_hsl_sse41 LABEL PROC
   DB  65,15,89,198                        ; mulps         %xmm14,%xmm0
   DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
   DB  68,15,194,250,1                     ; cmpltps       %xmm2,%xmm15
-  DB  68,15,84,61,56,48,0,0               ; andps         0x3038(%rip),%xmm15        # 4b10 <_sk_callback_sse41+0x43d>
+  DB  68,15,84,61,56,48,0,0               ; andps         0x3038(%rip),%xmm15        # 4b10 <_sk_callback_sse41+0x440>
   DB  68,15,88,248                        ; addps         %xmm0,%xmm15
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  65,15,194,193,0                     ; cmpeqps       %xmm9,%xmm0
   DB  65,15,92,208                        ; subps         %xmm8,%xmm2
   DB  65,15,89,214                        ; mulps         %xmm14,%xmm2
-  DB  68,15,40,45,43,48,0,0               ; movaps        0x302b(%rip),%xmm13        # 4b20 <_sk_callback_sse41+0x44d>
+  DB  68,15,40,45,43,48,0,0               ; movaps        0x302b(%rip),%xmm13        # 4b20 <_sk_callback_sse41+0x450>
   DB  65,15,88,213                        ; addps         %xmm13,%xmm2
   DB  69,15,92,193                        ; subps         %xmm9,%xmm8
   DB  69,15,89,198                        ; mulps         %xmm14,%xmm8
-  DB  68,15,88,5,39,48,0,0                ; addps         0x3027(%rip),%xmm8        # 4b30 <_sk_callback_sse41+0x45d>
+  DB  68,15,88,5,39,48,0,0                ; addps         0x3027(%rip),%xmm8        # 4b30 <_sk_callback_sse41+0x460>
   DB  102,68,15,56,20,194                 ; blendvps      %xmm0,%xmm2,%xmm8
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  102,69,15,56,20,199                 ; blendvps      %xmm0,%xmm15,%xmm8
-  DB  68,15,89,5,31,48,0,0                ; mulps         0x301f(%rip),%xmm8        # 4b40 <_sk_callback_sse41+0x46d>
+  DB  68,15,89,5,31,48,0,0                ; mulps         0x301f(%rip),%xmm8        # 4b40 <_sk_callback_sse41+0x470>
   DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
   DB  69,15,194,204,4                     ; cmpneqps      %xmm12,%xmm9
   DB  69,15,84,193                        ; andps         %xmm9,%xmm8
   DB  69,15,92,235                        ; subps         %xmm11,%xmm13
   DB  69,15,88,220                        ; addps         %xmm12,%xmm11
-  DB  15,40,5,19,48,0,0                   ; movaps        0x3013(%rip),%xmm0        # 4b50 <_sk_callback_sse41+0x47d>
+  DB  15,40,5,19,48,0,0                   ; movaps        0x3013(%rip),%xmm0        # 4b50 <_sk_callback_sse41+0x480>
   DB  65,15,40,211                        ; movaps        %xmm11,%xmm2
   DB  15,89,208                           ; mulps         %xmm0,%xmm2
   DB  15,194,194,1                        ; cmpltps       %xmm2,%xmm0
@@ -13178,7 +13176,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC
   DB  15,41,100,36,32                     ; movaps        %xmm4,0x20(%rsp)
   DB  15,41,92,36,16                      ; movaps        %xmm3,0x10(%rsp)
   DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
-  DB  68,15,40,13,213,47,0,0              ; movaps        0x2fd5(%rip),%xmm9        # 4b60 <_sk_callback_sse41+0x48d>
+  DB  68,15,40,13,213,47,0,0              ; movaps        0x2fd5(%rip),%xmm9        # 4b60 <_sk_callback_sse41+0x490>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,194,194,2                        ; cmpleps       %xmm2,%xmm0
   DB  15,40,217                           ; movaps        %xmm1,%xmm3
@@ -13191,19 +13189,19 @@ _sk_hsl_to_rgb_sse41 LABEL PROC
   DB  15,41,20,36                         ; movaps        %xmm2,(%rsp)
   DB  69,15,88,192                        ; addps         %xmm8,%xmm8
   DB  68,15,92,197                        ; subps         %xmm5,%xmm8
-  DB  68,15,40,53,177,47,0,0              ; movaps        0x2fb1(%rip),%xmm14        # 4b70 <_sk_callback_sse41+0x49d>
+  DB  68,15,40,53,177,47,0,0              ; movaps        0x2fb1(%rip),%xmm14        # 4b70 <_sk_callback_sse41+0x4a0>
   DB  69,15,88,242                        ; addps         %xmm10,%xmm14
   DB  102,65,15,58,8,198,1                ; roundps       $0x1,%xmm14,%xmm0
   DB  68,15,92,240                        ; subps         %xmm0,%xmm14
-  DB  68,15,40,29,170,47,0,0              ; movaps        0x2faa(%rip),%xmm11        # 4b80 <_sk_callback_sse41+0x4ad>
+  DB  68,15,40,29,170,47,0,0              ; movaps        0x2faa(%rip),%xmm11        # 4b80 <_sk_callback_sse41+0x4b0>
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  65,15,194,198,2                     ; cmpleps       %xmm14,%xmm0
   DB  15,40,245                           ; movaps        %xmm5,%xmm6
   DB  65,15,92,240                        ; subps         %xmm8,%xmm6
-  DB  15,40,61,163,47,0,0                 ; movaps        0x2fa3(%rip),%xmm7        # 4b90 <_sk_callback_sse41+0x4bd>
+  DB  15,40,61,163,47,0,0                 ; movaps        0x2fa3(%rip),%xmm7        # 4b90 <_sk_callback_sse41+0x4c0>
   DB  69,15,40,238                        ; movaps        %xmm14,%xmm13
   DB  68,15,89,239                        ; mulps         %xmm7,%xmm13
-  DB  15,40,29,164,47,0,0                 ; movaps        0x2fa4(%rip),%xmm3        # 4ba0 <_sk_callback_sse41+0x4cd>
+  DB  15,40,29,164,47,0,0                 ; movaps        0x2fa4(%rip),%xmm3        # 4ba0 <_sk_callback_sse41+0x4d0>
   DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  68,15,89,230                        ; mulps         %xmm6,%xmm12
@@ -13213,7 +13211,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC
   DB  65,15,194,198,2                     ; cmpleps       %xmm14,%xmm0
   DB  68,15,40,253                        ; movaps        %xmm5,%xmm15
   DB  102,69,15,56,20,252                 ; blendvps      %xmm0,%xmm12,%xmm15
-  DB  68,15,40,37,131,47,0,0              ; movaps        0x2f83(%rip),%xmm12        # 4bb0 <_sk_callback_sse41+0x4dd>
+  DB  68,15,40,37,131,47,0,0              ; movaps        0x2f83(%rip),%xmm12        # 4bb0 <_sk_callback_sse41+0x4e0>
   DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
   DB  65,15,194,198,2                     ; cmpleps       %xmm14,%xmm0
   DB  68,15,89,238                        ; mulps         %xmm6,%xmm13
@@ -13247,7 +13245,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC
   DB  65,15,40,198                        ; movaps        %xmm14,%xmm0
   DB  15,40,20,36                         ; movaps        (%rsp),%xmm2
   DB  102,15,56,20,202                    ; blendvps      %xmm0,%xmm2,%xmm1
-  DB  68,15,88,21,252,46,0,0              ; addps         0x2efc(%rip),%xmm10        # 4bc0 <_sk_callback_sse41+0x4ed>
+  DB  68,15,88,21,252,46,0,0              ; addps         0x2efc(%rip),%xmm10        # 4bc0 <_sk_callback_sse41+0x4f0>
   DB  102,65,15,58,8,194,1                ; roundps       $0x1,%xmm10,%xmm0
   DB  68,15,92,208                        ; subps         %xmm0,%xmm10
   DB  69,15,194,218,2                     ; cmpleps       %xmm10,%xmm11
@@ -13296,7 +13294,7 @@ _sk_scale_u8_sse41 LABEL PROC
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,68,15,56,49,4,56                ; pmovzxbd      (%rax,%rdi,1),%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,85,46,0,0                ; mulps         0x2e55(%rip),%xmm8        # 4bd0 <_sk_callback_sse41+0x4fd>
+  DB  68,15,89,5,85,46,0,0                ; mulps         0x2e55(%rip),%xmm8        # 4bd0 <_sk_callback_sse41+0x500>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
@@ -13330,7 +13328,7 @@ _sk_lerp_u8_sse41 LABEL PROC
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,68,15,56,49,4,56                ; pmovzxbd      (%rax,%rdi,1),%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,1,46,0,0                 ; mulps         0x2e01(%rip),%xmm8        # 4be0 <_sk_callback_sse41+0x50d>
+  DB  68,15,89,5,1,46,0,0                 ; mulps         0x2e01(%rip),%xmm8        # 4be0 <_sk_callback_sse41+0x510>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -13351,17 +13349,17 @@ _sk_lerp_565_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,68,15,56,51,20,120              ; pmovzxwd      (%rax,%rdi,2),%xmm10
-  DB  102,68,15,111,5,208,45,0,0          ; movdqa        0x2dd0(%rip),%xmm8        # 4bf0 <_sk_callback_sse41+0x51d>
+  DB  102,68,15,111,5,208,45,0,0          ; movdqa        0x2dd0(%rip),%xmm8        # 4bf0 <_sk_callback_sse41+0x520>
   DB  102,69,15,219,194                   ; pand          %xmm10,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,207,45,0,0               ; mulps         0x2dcf(%rip),%xmm8        # 4c00 <_sk_callback_sse41+0x52d>
-  DB  102,68,15,111,13,214,45,0,0         ; movdqa        0x2dd6(%rip),%xmm9        # 4c10 <_sk_callback_sse41+0x53d>
+  DB  68,15,89,5,207,45,0,0               ; mulps         0x2dcf(%rip),%xmm8        # 4c00 <_sk_callback_sse41+0x530>
+  DB  102,68,15,111,13,214,45,0,0         ; movdqa        0x2dd6(%rip),%xmm9        # 4c10 <_sk_callback_sse41+0x540>
   DB  102,69,15,219,202                   ; pand          %xmm10,%xmm9
   DB  69,15,91,201                        ; cvtdq2ps      %xmm9,%xmm9
-  DB  68,15,89,13,213,45,0,0              ; mulps         0x2dd5(%rip),%xmm9        # 4c20 <_sk_callback_sse41+0x54d>
-  DB  102,68,15,219,21,220,45,0,0         ; pand          0x2ddc(%rip),%xmm10        # 4c30 <_sk_callback_sse41+0x55d>
+  DB  68,15,89,13,213,45,0,0              ; mulps         0x2dd5(%rip),%xmm9        # 4c20 <_sk_callback_sse41+0x550>
+  DB  102,68,15,219,21,220,45,0,0         ; pand          0x2ddc(%rip),%xmm10        # 4c30 <_sk_callback_sse41+0x560>
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
-  DB  68,15,89,21,224,45,0,0              ; mulps         0x2de0(%rip),%xmm10        # 4c40 <_sk_callback_sse41+0x56d>
+  DB  68,15,89,21,224,45,0,0              ; mulps         0x2de0(%rip),%xmm10        # 4c40 <_sk_callback_sse41+0x570>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -13390,7 +13388,7 @@ _sk_load_tables_sse41 LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  76,139,72,8                         ; mov           0x8(%rax),%r9
   DB  243,69,15,111,4,184                 ; movdqu        (%r8,%rdi,4),%xmm8
-  DB  102,15,111,5,145,45,0,0             ; movdqa        0x2d91(%rip),%xmm0        # 4c50 <_sk_callback_sse41+0x57d>
+  DB  102,15,111,5,145,45,0,0             ; movdqa        0x2d91(%rip),%xmm0        # 4c50 <_sk_callback_sse41+0x580>
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,73,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%r8
   DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
@@ -13405,7 +13403,7 @@ _sk_load_tables_sse41 LABEL PROC
   DB  102,15,58,33,193,48                 ; insertps      $0x30,%xmm1,%xmm0
   DB  76,139,64,16                        ; mov           0x10(%rax),%r8
   DB  102,65,15,111,200                   ; movdqa        %xmm8,%xmm1
-  DB  102,15,56,0,13,76,45,0,0            ; pshufb        0x2d4c(%rip),%xmm1        # 4c60 <_sk_callback_sse41+0x58d>
+  DB  102,15,56,0,13,76,45,0,0            ; pshufb        0x2d4c(%rip),%xmm1        # 4c60 <_sk_callback_sse41+0x590>
   DB  102,73,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%r9
   DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
   DB  68,15,182,209                       ; movzbl        %cl,%r10d
@@ -13420,7 +13418,7 @@ _sk_load_tables_sse41 LABEL PROC
   DB  102,15,58,33,202,48                 ; insertps      $0x30,%xmm2,%xmm1
   DB  76,139,64,24                        ; mov           0x18(%rax),%r8
   DB  102,65,15,111,208                   ; movdqa        %xmm8,%xmm2
-  DB  102,15,56,0,21,8,45,0,0             ; pshufb        0x2d08(%rip),%xmm2        # 4c70 <_sk_callback_sse41+0x59d>
+  DB  102,15,56,0,21,8,45,0,0             ; pshufb        0x2d08(%rip),%xmm2        # 4c70 <_sk_callback_sse41+0x5a0>
   DB  102,72,15,58,22,209,1               ; pextrq        $0x1,%xmm2,%rcx
   DB  102,72,15,126,208                   ; movq          %xmm2,%rax
   DB  68,15,182,200                       ; movzbl        %al,%r9d
@@ -13435,7 +13433,7 @@ _sk_load_tables_sse41 LABEL PROC
   DB  102,15,58,33,211,48                 ; insertps      $0x30,%xmm3,%xmm2
   DB  102,65,15,114,208,24                ; psrld         $0x18,%xmm8
   DB  65,15,91,216                        ; cvtdq2ps      %xmm8,%xmm3
-  DB  15,89,29,197,44,0,0                 ; mulps         0x2cc5(%rip),%xmm3        # 4c80 <_sk_callback_sse41+0x5ad>
+  DB  15,89,29,197,44,0,0                 ; mulps         0x2cc5(%rip),%xmm3        # 4c80 <_sk_callback_sse41+0x5b0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -13452,7 +13450,7 @@ _sk_load_tables_u16_be_sse41 LABEL PROC
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
   DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
-  DB  102,68,15,111,5,152,44,0,0          ; movdqa        0x2c98(%rip),%xmm8        # 4c90 <_sk_callback_sse41+0x5bd>
+  DB  102,68,15,111,5,152,44,0,0          ; movdqa        0x2c98(%rip),%xmm8        # 4c90 <_sk_callback_sse41+0x5c0>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
@@ -13469,7 +13467,7 @@ _sk_load_tables_u16_be_sse41 LABEL PROC
   DB  243,67,15,16,20,8                   ; movss         (%r8,%r9,1),%xmm2
   DB  102,15,58,33,194,48                 ; insertps      $0x30,%xmm2,%xmm0
   DB  76,139,64,16                        ; mov           0x10(%rax),%r8
-  DB  102,15,56,0,13,75,44,0,0            ; pshufb        0x2c4b(%rip),%xmm1        # 4ca0 <_sk_callback_sse41+0x5cd>
+  DB  102,15,56,0,13,75,44,0,0            ; pshufb        0x2c4b(%rip),%xmm1        # 4ca0 <_sk_callback_sse41+0x5d0>
   DB  102,15,56,51,201                    ; pmovzxwd      %xmm1,%xmm1
   DB  102,73,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%r9
   DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
@@ -13505,7 +13503,7 @@ _sk_load_tables_u16_be_sse41 LABEL PROC
   DB  102,65,15,235,216                   ; por           %xmm8,%xmm3
   DB  102,15,56,51,219                    ; pmovzxwd      %xmm3,%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,153,43,0,0                 ; mulps         0x2b99(%rip),%xmm3        # 4cb0 <_sk_callback_sse41+0x5dd>
+  DB  15,89,29,153,43,0,0                 ; mulps         0x2b99(%rip),%xmm3        # 4cb0 <_sk_callback_sse41+0x5e0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -13525,7 +13523,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC
   DB  102,68,15,97,200                    ; punpcklwd     %xmm0,%xmm9
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
-  DB  102,68,15,111,5,91,43,0,0           ; movdqa        0x2b5b(%rip),%xmm8        # 4cc0 <_sk_callback_sse41+0x5ed>
+  DB  102,68,15,111,5,91,43,0,0           ; movdqa        0x2b5b(%rip),%xmm8        # 4cc0 <_sk_callback_sse41+0x5f0>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
@@ -13542,7 +13540,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC
   DB  243,67,15,16,28,8                   ; movss         (%r8,%r9,1),%xmm3
   DB  102,15,58,33,195,48                 ; insertps      $0x30,%xmm3,%xmm0
   DB  76,139,64,16                        ; mov           0x10(%rax),%r8
-  DB  102,15,56,0,13,14,43,0,0            ; pshufb        0x2b0e(%rip),%xmm1        # 4cd0 <_sk_callback_sse41+0x5fd>
+  DB  102,15,56,0,13,14,43,0,0            ; pshufb        0x2b0e(%rip),%xmm1        # 4cd0 <_sk_callback_sse41+0x600>
   DB  102,15,56,51,201                    ; pmovzxwd      %xmm1,%xmm1
   DB  102,73,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%r9
   DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
@@ -13573,7 +13571,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC
   DB  243,65,15,16,28,8                   ; movss         (%r8,%rcx,1),%xmm3
   DB  102,15,58,33,211,48                 ; insertps      $0x30,%xmm3,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,121,42,0,0                 ; movaps        0x2a79(%rip),%xmm3        # 4ce0 <_sk_callback_sse41+0x60d>
+  DB  15,40,29,121,42,0,0                 ; movaps        0x2a79(%rip),%xmm3        # 4ce0 <_sk_callback_sse41+0x610>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_byte_tables_sse41
@@ -13581,7 +13579,7 @@ _sk_byte_tables_sse41 LABEL PROC
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,122,42,0,0               ; movaps        0x2a7a(%rip),%xmm8        # 4cf0 <_sk_callback_sse41+0x61d>
+  DB  68,15,40,5,122,42,0,0               ; movaps        0x2a7a(%rip),%xmm8        # 4cf0 <_sk_callback_sse41+0x620>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
@@ -13600,7 +13598,7 @@ _sk_byte_tables_sse41 LABEL PROC
   DB  102,15,58,32,193,3                  ; pinsrb        $0x3,%ecx,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,13,43,42,0,0               ; movaps        0x2a2b(%rip),%xmm9        # 4d00 <_sk_callback_sse41+0x62d>
+  DB  68,15,40,13,43,42,0,0               ; movaps        0x2a2b(%rip),%xmm9        # 4d00 <_sk_callback_sse41+0x630>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -13689,7 +13687,7 @@ _sk_byte_tables_rgb_sse41 LABEL PROC
   DB  102,15,58,32,193,3                  ; pinsrb        $0x3,%ecx,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,13,179,40,0,0              ; movaps        0x28b3(%rip),%xmm9        # 4d10 <_sk_callback_sse41+0x63d>
+  DB  68,15,40,13,179,40,0,0              ; movaps        0x28b3(%rip),%xmm9        # 4d10 <_sk_callback_sse41+0x640>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -13856,31 +13854,31 @@ _sk_parametric_r_sse41 LABEL PROC
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,194                        ; cvtdq2ps      %xmm10,%xmm8
-  DB  68,15,89,5,10,38,0,0                ; mulps         0x260a(%rip),%xmm8        # 4d20 <_sk_callback_sse41+0x64d>
-  DB  68,15,84,21,18,38,0,0               ; andps         0x2612(%rip),%xmm10        # 4d30 <_sk_callback_sse41+0x65d>
-  DB  68,15,86,21,26,38,0,0               ; orps          0x261a(%rip),%xmm10        # 4d40 <_sk_callback_sse41+0x66d>
-  DB  68,15,88,5,34,38,0,0                ; addps         0x2622(%rip),%xmm8        # 4d50 <_sk_callback_sse41+0x67d>
-  DB  68,15,40,37,42,38,0,0               ; movaps        0x262a(%rip),%xmm12        # 4d60 <_sk_callback_sse41+0x68d>
+  DB  68,15,89,5,10,38,0,0                ; mulps         0x260a(%rip),%xmm8        # 4d20 <_sk_callback_sse41+0x650>
+  DB  68,15,84,21,18,38,0,0               ; andps         0x2612(%rip),%xmm10        # 4d30 <_sk_callback_sse41+0x660>
+  DB  68,15,86,21,26,38,0,0               ; orps          0x261a(%rip),%xmm10        # 4d40 <_sk_callback_sse41+0x670>
+  DB  68,15,88,5,34,38,0,0                ; addps         0x2622(%rip),%xmm8        # 4d50 <_sk_callback_sse41+0x680>
+  DB  68,15,40,37,42,38,0,0               ; movaps        0x262a(%rip),%xmm12        # 4d60 <_sk_callback_sse41+0x690>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,196                        ; subps         %xmm12,%xmm8
-  DB  68,15,88,21,42,38,0,0               ; addps         0x262a(%rip),%xmm10        # 4d70 <_sk_callback_sse41+0x69d>
-  DB  68,15,40,37,50,38,0,0               ; movaps        0x2632(%rip),%xmm12        # 4d80 <_sk_callback_sse41+0x6ad>
+  DB  68,15,88,21,42,38,0,0               ; addps         0x262a(%rip),%xmm10        # 4d70 <_sk_callback_sse41+0x6a0>
+  DB  68,15,40,37,50,38,0,0               ; movaps        0x2632(%rip),%xmm12        # 4d80 <_sk_callback_sse41+0x6b0>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,196                        ; subps         %xmm12,%xmm8
   DB  69,15,89,195                        ; mulps         %xmm11,%xmm8
   DB  102,69,15,58,8,208,1                ; roundps       $0x1,%xmm8,%xmm10
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,5,31,38,0,0                ; addps         0x261f(%rip),%xmm8        # 4d90 <_sk_callback_sse41+0x6bd>
-  DB  68,15,40,21,39,38,0,0               ; movaps        0x2627(%rip),%xmm10        # 4da0 <_sk_callback_sse41+0x6cd>
+  DB  68,15,88,5,31,38,0,0                ; addps         0x261f(%rip),%xmm8        # 4d90 <_sk_callback_sse41+0x6c0>
+  DB  68,15,40,21,39,38,0,0               ; movaps        0x2627(%rip),%xmm10        # 4da0 <_sk_callback_sse41+0x6d0>
   DB  69,15,89,211                        ; mulps         %xmm11,%xmm10
   DB  69,15,92,194                        ; subps         %xmm10,%xmm8
-  DB  68,15,40,21,39,38,0,0               ; movaps        0x2627(%rip),%xmm10        # 4db0 <_sk_callback_sse41+0x6dd>
+  DB  68,15,40,21,39,38,0,0               ; movaps        0x2627(%rip),%xmm10        # 4db0 <_sk_callback_sse41+0x6e0>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  68,15,40,29,43,38,0,0               ; movaps        0x262b(%rip),%xmm11        # 4dc0 <_sk_callback_sse41+0x6ed>
+  DB  68,15,40,29,43,38,0,0               ; movaps        0x262b(%rip),%xmm11        # 4dc0 <_sk_callback_sse41+0x6f0>
   DB  69,15,94,218                        ; divps         %xmm10,%xmm11
   DB  69,15,88,216                        ; addps         %xmm8,%xmm11
-  DB  68,15,89,29,43,38,0,0               ; mulps         0x262b(%rip),%xmm11        # 4dd0 <_sk_callback_sse41+0x6fd>
+  DB  68,15,89,29,43,38,0,0               ; mulps         0x262b(%rip),%xmm11        # 4dd0 <_sk_callback_sse41+0x700>
   DB  102,69,15,91,211                    ; cvtps2dq      %xmm11,%xmm10
   DB  243,68,15,16,64,20                  ; movss         0x14(%rax),%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
@@ -13888,7 +13886,7 @@ _sk_parametric_r_sse41 LABEL PROC
   DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
-  DB  68,15,93,5,18,38,0,0                ; minps         0x2612(%rip),%xmm8        # 4de0 <_sk_callback_sse41+0x70d>
+  DB  68,15,93,5,18,38,0,0                ; minps         0x2612(%rip),%xmm8        # 4de0 <_sk_callback_sse41+0x710>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -13916,31 +13914,31 @@ _sk_parametric_g_sse41 LABEL PROC
   DB  68,15,88,217                        ; addps         %xmm1,%xmm11
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
-  DB  68,15,89,37,179,37,0,0              ; mulps         0x25b3(%rip),%xmm12        # 4df0 <_sk_callback_sse41+0x71d>
-  DB  68,15,84,29,187,37,0,0              ; andps         0x25bb(%rip),%xmm11        # 4e00 <_sk_callback_sse41+0x72d>
-  DB  68,15,86,29,195,37,0,0              ; orps          0x25c3(%rip),%xmm11        # 4e10 <_sk_callback_sse41+0x73d>
-  DB  68,15,88,37,203,37,0,0              ; addps         0x25cb(%rip),%xmm12        # 4e20 <_sk_callback_sse41+0x74d>
-  DB  15,40,13,212,37,0,0                 ; movaps        0x25d4(%rip),%xmm1        # 4e30 <_sk_callback_sse41+0x75d>
+  DB  68,15,89,37,179,37,0,0              ; mulps         0x25b3(%rip),%xmm12        # 4df0 <_sk_callback_sse41+0x720>
+  DB  68,15,84,29,187,37,0,0              ; andps         0x25bb(%rip),%xmm11        # 4e00 <_sk_callback_sse41+0x730>
+  DB  68,15,86,29,195,37,0,0              ; orps          0x25c3(%rip),%xmm11        # 4e10 <_sk_callback_sse41+0x740>
+  DB  68,15,88,37,203,37,0,0              ; addps         0x25cb(%rip),%xmm12        # 4e20 <_sk_callback_sse41+0x750>
+  DB  15,40,13,212,37,0,0                 ; movaps        0x25d4(%rip),%xmm1        # 4e30 <_sk_callback_sse41+0x760>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
-  DB  68,15,88,29,212,37,0,0              ; addps         0x25d4(%rip),%xmm11        # 4e40 <_sk_callback_sse41+0x76d>
-  DB  15,40,13,221,37,0,0                 ; movaps        0x25dd(%rip),%xmm1        # 4e50 <_sk_callback_sse41+0x77d>
+  DB  68,15,88,29,212,37,0,0              ; addps         0x25d4(%rip),%xmm11        # 4e40 <_sk_callback_sse41+0x770>
+  DB  15,40,13,221,37,0,0                 ; movaps        0x25dd(%rip),%xmm1        # 4e50 <_sk_callback_sse41+0x780>
   DB  65,15,94,203                        ; divps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  102,69,15,58,8,212,1                ; roundps       $0x1,%xmm12,%xmm10
   DB  69,15,40,220                        ; movaps        %xmm12,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,37,202,37,0,0              ; addps         0x25ca(%rip),%xmm12        # 4e60 <_sk_callback_sse41+0x78d>
-  DB  15,40,13,211,37,0,0                 ; movaps        0x25d3(%rip),%xmm1        # 4e70 <_sk_callback_sse41+0x79d>
+  DB  68,15,88,37,202,37,0,0              ; addps         0x25ca(%rip),%xmm12        # 4e60 <_sk_callback_sse41+0x790>
+  DB  15,40,13,211,37,0,0                 ; movaps        0x25d3(%rip),%xmm1        # 4e70 <_sk_callback_sse41+0x7a0>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
-  DB  68,15,40,21,211,37,0,0              ; movaps        0x25d3(%rip),%xmm10        # 4e80 <_sk_callback_sse41+0x7ad>
+  DB  68,15,40,21,211,37,0,0              ; movaps        0x25d3(%rip),%xmm10        # 4e80 <_sk_callback_sse41+0x7b0>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,13,216,37,0,0                 ; movaps        0x25d8(%rip),%xmm1        # 4e90 <_sk_callback_sse41+0x7bd>
+  DB  15,40,13,216,37,0,0                 ; movaps        0x25d8(%rip),%xmm1        # 4e90 <_sk_callback_sse41+0x7c0>
   DB  65,15,94,202                        ; divps         %xmm10,%xmm1
   DB  65,15,88,204                        ; addps         %xmm12,%xmm1
-  DB  15,89,13,217,37,0,0                 ; mulps         0x25d9(%rip),%xmm1        # 4ea0 <_sk_callback_sse41+0x7cd>
+  DB  15,89,13,217,37,0,0                 ; mulps         0x25d9(%rip),%xmm1        # 4ea0 <_sk_callback_sse41+0x7d0>
   DB  102,68,15,91,209                    ; cvtps2dq      %xmm1,%xmm10
   DB  243,15,16,72,20                     ; movss         0x14(%rax),%xmm1
   DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
@@ -13948,7 +13946,7 @@ _sk_parametric_g_sse41 LABEL PROC
   DB  102,65,15,56,20,201                 ; blendvps      %xmm0,%xmm9,%xmm1
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,95,200                           ; maxps         %xmm0,%xmm1
-  DB  15,93,13,196,37,0,0                 ; minps         0x25c4(%rip),%xmm1        # 4eb0 <_sk_callback_sse41+0x7dd>
+  DB  15,93,13,196,37,0,0                 ; minps         0x25c4(%rip),%xmm1        # 4eb0 <_sk_callback_sse41+0x7e0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -13976,31 +13974,31 @@ _sk_parametric_b_sse41 LABEL PROC
   DB  68,15,88,218                        ; addps         %xmm2,%xmm11
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
-  DB  68,15,89,37,101,37,0,0              ; mulps         0x2565(%rip),%xmm12        # 4ec0 <_sk_callback_sse41+0x7ed>
-  DB  68,15,84,29,109,37,0,0              ; andps         0x256d(%rip),%xmm11        # 4ed0 <_sk_callback_sse41+0x7fd>
-  DB  68,15,86,29,117,37,0,0              ; orps          0x2575(%rip),%xmm11        # 4ee0 <_sk_callback_sse41+0x80d>
-  DB  68,15,88,37,125,37,0,0              ; addps         0x257d(%rip),%xmm12        # 4ef0 <_sk_callback_sse41+0x81d>
-  DB  15,40,21,134,37,0,0                 ; movaps        0x2586(%rip),%xmm2        # 4f00 <_sk_callback_sse41+0x82d>
+  DB  68,15,89,37,101,37,0,0              ; mulps         0x2565(%rip),%xmm12        # 4ec0 <_sk_callback_sse41+0x7f0>
+  DB  68,15,84,29,109,37,0,0              ; andps         0x256d(%rip),%xmm11        # 4ed0 <_sk_callback_sse41+0x800>
+  DB  68,15,86,29,117,37,0,0              ; orps          0x2575(%rip),%xmm11        # 4ee0 <_sk_callback_sse41+0x810>
+  DB  68,15,88,37,125,37,0,0              ; addps         0x257d(%rip),%xmm12        # 4ef0 <_sk_callback_sse41+0x820>
+  DB  15,40,21,134,37,0,0                 ; movaps        0x2586(%rip),%xmm2        # 4f00 <_sk_callback_sse41+0x830>
   DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
-  DB  68,15,88,29,134,37,0,0              ; addps         0x2586(%rip),%xmm11        # 4f10 <_sk_callback_sse41+0x83d>
-  DB  15,40,21,143,37,0,0                 ; movaps        0x258f(%rip),%xmm2        # 4f20 <_sk_callback_sse41+0x84d>
+  DB  68,15,88,29,134,37,0,0              ; addps         0x2586(%rip),%xmm11        # 4f10 <_sk_callback_sse41+0x840>
+  DB  15,40,21,143,37,0,0                 ; movaps        0x258f(%rip),%xmm2        # 4f20 <_sk_callback_sse41+0x850>
   DB  65,15,94,211                        ; divps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  102,69,15,58,8,212,1                ; roundps       $0x1,%xmm12,%xmm10
   DB  69,15,40,220                        ; movaps        %xmm12,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,37,124,37,0,0              ; addps         0x257c(%rip),%xmm12        # 4f30 <_sk_callback_sse41+0x85d>
-  DB  15,40,21,133,37,0,0                 ; movaps        0x2585(%rip),%xmm2        # 4f40 <_sk_callback_sse41+0x86d>
+  DB  68,15,88,37,124,37,0,0              ; addps         0x257c(%rip),%xmm12        # 4f30 <_sk_callback_sse41+0x860>
+  DB  15,40,21,133,37,0,0                 ; movaps        0x2585(%rip),%xmm2        # 4f40 <_sk_callback_sse41+0x870>
   DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
-  DB  68,15,40,21,133,37,0,0              ; movaps        0x2585(%rip),%xmm10        # 4f50 <_sk_callback_sse41+0x87d>
+  DB  68,15,40,21,133,37,0,0              ; movaps        0x2585(%rip),%xmm10        # 4f50 <_sk_callback_sse41+0x880>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,21,138,37,0,0                 ; movaps        0x258a(%rip),%xmm2        # 4f60 <_sk_callback_sse41+0x88d>
+  DB  15,40,21,138,37,0,0                 ; movaps        0x258a(%rip),%xmm2        # 4f60 <_sk_callback_sse41+0x890>
   DB  65,15,94,210                        ; divps         %xmm10,%xmm2
   DB  65,15,88,212                        ; addps         %xmm12,%xmm2
-  DB  15,89,21,139,37,0,0                 ; mulps         0x258b(%rip),%xmm2        # 4f70 <_sk_callback_sse41+0x89d>
+  DB  15,89,21,139,37,0,0                 ; mulps         0x258b(%rip),%xmm2        # 4f70 <_sk_callback_sse41+0x8a0>
   DB  102,68,15,91,210                    ; cvtps2dq      %xmm2,%xmm10
   DB  243,15,16,80,20                     ; movss         0x14(%rax),%xmm2
   DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
@@ -14008,7 +14006,7 @@ _sk_parametric_b_sse41 LABEL PROC
   DB  102,65,15,56,20,209                 ; blendvps      %xmm0,%xmm9,%xmm2
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,95,208                           ; maxps         %xmm0,%xmm2
-  DB  15,93,21,118,37,0,0                 ; minps         0x2576(%rip),%xmm2        # 4f80 <_sk_callback_sse41+0x8ad>
+  DB  15,93,21,118,37,0,0                 ; minps         0x2576(%rip),%xmm2        # 4f80 <_sk_callback_sse41+0x8b0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -14036,31 +14034,31 @@ _sk_parametric_a_sse41 LABEL PROC
   DB  68,15,88,219                        ; addps         %xmm3,%xmm11
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
-  DB  68,15,89,37,23,37,0,0               ; mulps         0x2517(%rip),%xmm12        # 4f90 <_sk_callback_sse41+0x8bd>
-  DB  68,15,84,29,31,37,0,0               ; andps         0x251f(%rip),%xmm11        # 4fa0 <_sk_callback_sse41+0x8cd>
-  DB  68,15,86,29,39,37,0,0               ; orps          0x2527(%rip),%xmm11        # 4fb0 <_sk_callback_sse41+0x8dd>
-  DB  68,15,88,37,47,37,0,0               ; addps         0x252f(%rip),%xmm12        # 4fc0 <_sk_callback_sse41+0x8ed>
-  DB  15,40,29,56,37,0,0                  ; movaps        0x2538(%rip),%xmm3        # 4fd0 <_sk_callback_sse41+0x8fd>
+  DB  68,15,89,37,23,37,0,0               ; mulps         0x2517(%rip),%xmm12        # 4f90 <_sk_callback_sse41+0x8c0>
+  DB  68,15,84,29,31,37,0,0               ; andps         0x251f(%rip),%xmm11        # 4fa0 <_sk_callback_sse41+0x8d0>
+  DB  68,15,86,29,39,37,0,0               ; orps          0x2527(%rip),%xmm11        # 4fb0 <_sk_callback_sse41+0x8e0>
+  DB  68,15,88,37,47,37,0,0               ; addps         0x252f(%rip),%xmm12        # 4fc0 <_sk_callback_sse41+0x8f0>
+  DB  15,40,29,56,37,0,0                  ; movaps        0x2538(%rip),%xmm3        # 4fd0 <_sk_callback_sse41+0x900>
   DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
-  DB  68,15,88,29,56,37,0,0               ; addps         0x2538(%rip),%xmm11        # 4fe0 <_sk_callback_sse41+0x90d>
-  DB  15,40,29,65,37,0,0                  ; movaps        0x2541(%rip),%xmm3        # 4ff0 <_sk_callback_sse41+0x91d>
+  DB  68,15,88,29,56,37,0,0               ; addps         0x2538(%rip),%xmm11        # 4fe0 <_sk_callback_sse41+0x910>
+  DB  15,40,29,65,37,0,0                  ; movaps        0x2541(%rip),%xmm3        # 4ff0 <_sk_callback_sse41+0x920>
   DB  65,15,94,219                        ; divps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  102,69,15,58,8,212,1                ; roundps       $0x1,%xmm12,%xmm10
   DB  69,15,40,220                        ; movaps        %xmm12,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,37,46,37,0,0               ; addps         0x252e(%rip),%xmm12        # 5000 <_sk_callback_sse41+0x92d>
-  DB  15,40,29,55,37,0,0                  ; movaps        0x2537(%rip),%xmm3        # 5010 <_sk_callback_sse41+0x93d>
+  DB  68,15,88,37,46,37,0,0               ; addps         0x252e(%rip),%xmm12        # 5000 <_sk_callback_sse41+0x930>
+  DB  15,40,29,55,37,0,0                  ; movaps        0x2537(%rip),%xmm3        # 5010 <_sk_callback_sse41+0x940>
   DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
-  DB  68,15,40,21,55,37,0,0               ; movaps        0x2537(%rip),%xmm10        # 5020 <_sk_callback_sse41+0x94d>
+  DB  68,15,40,21,55,37,0,0               ; movaps        0x2537(%rip),%xmm10        # 5020 <_sk_callback_sse41+0x950>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,29,60,37,0,0                  ; movaps        0x253c(%rip),%xmm3        # 5030 <_sk_callback_sse41+0x95d>
+  DB  15,40,29,60,37,0,0                  ; movaps        0x253c(%rip),%xmm3        # 5030 <_sk_callback_sse41+0x960>
   DB  65,15,94,218                        ; divps         %xmm10,%xmm3
   DB  65,15,88,220                        ; addps         %xmm12,%xmm3
-  DB  15,89,29,61,37,0,0                  ; mulps         0x253d(%rip),%xmm3        # 5040 <_sk_callback_sse41+0x96d>
+  DB  15,89,29,61,37,0,0                  ; mulps         0x253d(%rip),%xmm3        # 5040 <_sk_callback_sse41+0x970>
   DB  102,68,15,91,211                    ; cvtps2dq      %xmm3,%xmm10
   DB  243,15,16,88,20                     ; movss         0x14(%rax),%xmm3
   DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
@@ -14068,7 +14066,7 @@ _sk_parametric_a_sse41 LABEL PROC
   DB  102,65,15,56,20,217                 ; blendvps      %xmm0,%xmm9,%xmm3
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,95,216                           ; maxps         %xmm0,%xmm3
-  DB  15,93,29,40,37,0,0                  ; minps         0x2528(%rip),%xmm3        # 5050 <_sk_callback_sse41+0x97d>
+  DB  15,93,29,40,37,0,0                  ; minps         0x2528(%rip),%xmm3        # 5050 <_sk_callback_sse41+0x980>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -14076,29 +14074,29 @@ _sk_parametric_a_sse41 LABEL PROC
 PUBLIC _sk_lab_to_xyz_sse41
 _sk_lab_to_xyz_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,89,5,36,37,0,0                ; mulps         0x2524(%rip),%xmm8        # 5060 <_sk_callback_sse41+0x98d>
-  DB  68,15,40,13,44,37,0,0               ; movaps        0x252c(%rip),%xmm9        # 5070 <_sk_callback_sse41+0x99d>
+  DB  68,15,89,5,36,37,0,0                ; mulps         0x2524(%rip),%xmm8        # 5060 <_sk_callback_sse41+0x990>
+  DB  68,15,40,13,44,37,0,0               ; movaps        0x252c(%rip),%xmm9        # 5070 <_sk_callback_sse41+0x9a0>
   DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
-  DB  15,40,5,49,37,0,0                   ; movaps        0x2531(%rip),%xmm0        # 5080 <_sk_callback_sse41+0x9ad>
+  DB  15,40,5,49,37,0,0                   ; movaps        0x2531(%rip),%xmm0        # 5080 <_sk_callback_sse41+0x9b0>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
-  DB  68,15,88,5,47,37,0,0                ; addps         0x252f(%rip),%xmm8        # 5090 <_sk_callback_sse41+0x9bd>
-  DB  68,15,89,5,55,37,0,0                ; mulps         0x2537(%rip),%xmm8        # 50a0 <_sk_callback_sse41+0x9cd>
-  DB  15,89,13,64,37,0,0                  ; mulps         0x2540(%rip),%xmm1        # 50b0 <_sk_callback_sse41+0x9dd>
+  DB  68,15,88,5,47,37,0,0                ; addps         0x252f(%rip),%xmm8        # 5090 <_sk_callback_sse41+0x9c0>
+  DB  68,15,89,5,55,37,0,0                ; mulps         0x2537(%rip),%xmm8        # 50a0 <_sk_callback_sse41+0x9d0>
+  DB  15,89,13,64,37,0,0                  ; mulps         0x2540(%rip),%xmm1        # 50b0 <_sk_callback_sse41+0x9e0>
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  15,89,21,69,37,0,0                  ; mulps         0x2545(%rip),%xmm2        # 50c0 <_sk_callback_sse41+0x9ed>
+  DB  15,89,21,69,37,0,0                  ; mulps         0x2545(%rip),%xmm2        # 50c0 <_sk_callback_sse41+0x9f0>
   DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
   DB  68,15,92,210                        ; subps         %xmm2,%xmm10
   DB  68,15,40,217                        ; movaps        %xmm1,%xmm11
   DB  69,15,89,219                        ; mulps         %xmm11,%xmm11
   DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
-  DB  68,15,40,13,57,37,0,0               ; movaps        0x2539(%rip),%xmm9        # 50d0 <_sk_callback_sse41+0x9fd>
+  DB  68,15,40,13,57,37,0,0               ; movaps        0x2539(%rip),%xmm9        # 50d0 <_sk_callback_sse41+0xa00>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  65,15,194,195,1                     ; cmpltps       %xmm11,%xmm0
-  DB  15,40,21,57,37,0,0                  ; movaps        0x2539(%rip),%xmm2        # 50e0 <_sk_callback_sse41+0xa0d>
+  DB  15,40,21,57,37,0,0                  ; movaps        0x2539(%rip),%xmm2        # 50e0 <_sk_callback_sse41+0xa10>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
-  DB  68,15,40,37,62,37,0,0               ; movaps        0x253e(%rip),%xmm12        # 50f0 <_sk_callback_sse41+0xa1d>
+  DB  68,15,40,37,62,37,0,0               ; movaps        0x253e(%rip),%xmm12        # 50f0 <_sk_callback_sse41+0xa20>
   DB  65,15,89,204                        ; mulps         %xmm12,%xmm1
   DB  102,65,15,56,20,203                 ; blendvps      %xmm0,%xmm11,%xmm1
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
@@ -14117,8 +14115,8 @@ _sk_lab_to_xyz_sse41 LABEL PROC
   DB  65,15,89,212                        ; mulps         %xmm12,%xmm2
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  102,65,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm2
-  DB  15,89,13,247,36,0,0                 ; mulps         0x24f7(%rip),%xmm1        # 5100 <_sk_callback_sse41+0xa2d>
-  DB  15,89,21,0,37,0,0                   ; mulps         0x2500(%rip),%xmm2        # 5110 <_sk_callback_sse41+0xa3d>
+  DB  15,89,13,247,36,0,0                 ; mulps         0x24f7(%rip),%xmm1        # 5100 <_sk_callback_sse41+0xa30>
+  DB  15,89,21,0,37,0,0                   ; mulps         0x2500(%rip),%xmm2        # 5110 <_sk_callback_sse41+0xa40>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
@@ -14130,7 +14128,7 @@ _sk_load_a8_sse41 LABEL PROC
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,15,56,49,4,56                   ; pmovzxbd      (%rax,%rdi,1),%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,240,36,0,0                 ; mulps         0x24f0(%rip),%xmm3        # 5120 <_sk_callback_sse41+0xa4d>
+  DB  15,89,29,240,36,0,0                 ; mulps         0x24f0(%rip),%xmm3        # 5120 <_sk_callback_sse41+0xa50>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
@@ -14161,7 +14159,7 @@ _sk_gather_a8_sse41 LABEL PROC
   DB  102,15,58,32,192,3                  ; pinsrb        $0x3,%eax,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,132,36,0,0                 ; mulps         0x2484(%rip),%xmm3        # 5130 <_sk_callback_sse41+0xa5d>
+  DB  15,89,29,132,36,0,0                 ; mulps         0x2484(%rip),%xmm3        # 5130 <_sk_callback_sse41+0xa60>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -14172,7 +14170,7 @@ PUBLIC _sk_store_a8_sse41
 _sk_store_a8_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,120,36,0,0               ; movaps        0x2478(%rip),%xmm8        # 5140 <_sk_callback_sse41+0xa6d>
+  DB  68,15,40,5,120,36,0,0               ; movaps        0x2478(%rip),%xmm8        # 5140 <_sk_callback_sse41+0xa70>
   DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
   DB  102,69,15,56,43,192                 ; packusdw      %xmm8,%xmm8
@@ -14187,9 +14185,9 @@ _sk_load_g8_sse41 LABEL PROC
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,15,56,49,4,56                   ; pmovzxbd      (%rax,%rdi,1),%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,85,36,0,0                   ; mulps         0x2455(%rip),%xmm0        # 5150 <_sk_callback_sse41+0xa7d>
+  DB  15,89,5,85,36,0,0                   ; mulps         0x2455(%rip),%xmm0        # 5150 <_sk_callback_sse41+0xa80>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,92,36,0,0                  ; movaps        0x245c(%rip),%xmm3        # 5160 <_sk_callback_sse41+0xa8d>
+  DB  15,40,29,92,36,0,0                  ; movaps        0x245c(%rip),%xmm3        # 5160 <_sk_callback_sse41+0xa90>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -14218,9 +14216,9 @@ _sk_gather_g8_sse41 LABEL PROC
   DB  102,15,58,32,192,3                  ; pinsrb        $0x3,%eax,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,245,35,0,0                  ; mulps         0x23f5(%rip),%xmm0        # 5170 <_sk_callback_sse41+0xa9d>
+  DB  15,89,5,245,35,0,0                  ; mulps         0x23f5(%rip),%xmm0        # 5170 <_sk_callback_sse41+0xaa0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,252,35,0,0                 ; movaps        0x23fc(%rip),%xmm3        # 5180 <_sk_callback_sse41+0xaad>
+  DB  15,40,29,252,35,0,0                 ; movaps        0x23fc(%rip),%xmm3        # 5180 <_sk_callback_sse41+0xab0>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -14263,17 +14261,17 @@ _sk_gather_i8_sse41 LABEL PROC
   DB  102,15,58,34,28,8,1                 ; pinsrd        $0x1,(%rax,%rcx,1),%xmm3
   DB  102,66,15,58,34,28,144,2            ; pinsrd        $0x2,(%rax,%r10,4),%xmm3
   DB  102,66,15,58,34,28,8,3              ; pinsrd        $0x3,(%rax,%r9,1),%xmm3
-  DB  102,15,111,5,83,35,0,0              ; movdqa        0x2353(%rip),%xmm0        # 5190 <_sk_callback_sse41+0xabd>
+  DB  102,15,111,5,83,35,0,0              ; movdqa        0x2353(%rip),%xmm0        # 5190 <_sk_callback_sse41+0xac0>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,84,35,0,0                ; movaps        0x2354(%rip),%xmm8        # 51a0 <_sk_callback_sse41+0xacd>
+  DB  68,15,40,5,84,35,0,0                ; movaps        0x2354(%rip),%xmm8        # 51a0 <_sk_callback_sse41+0xad0>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
-  DB  102,15,56,0,13,83,35,0,0            ; pshufb        0x2353(%rip),%xmm1        # 51b0 <_sk_callback_sse41+0xadd>
+  DB  102,15,56,0,13,83,35,0,0            ; pshufb        0x2353(%rip),%xmm1        # 51b0 <_sk_callback_sse41+0xae0>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
-  DB  102,15,56,0,21,79,35,0,0            ; pshufb        0x234f(%rip),%xmm2        # 51c0 <_sk_callback_sse41+0xaed>
+  DB  102,15,56,0,21,79,35,0,0            ; pshufb        0x234f(%rip),%xmm2        # 51c0 <_sk_callback_sse41+0xaf0>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -14287,19 +14285,19 @@ _sk_load_565_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,15,56,51,20,120                 ; pmovzxwd      (%rax,%rdi,2),%xmm2
-  DB  102,15,111,5,53,35,0,0              ; movdqa        0x2335(%rip),%xmm0        # 51d0 <_sk_callback_sse41+0xafd>
+  DB  102,15,111,5,53,35,0,0              ; movdqa        0x2335(%rip),%xmm0        # 51d0 <_sk_callback_sse41+0xb00>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,55,35,0,0                   ; mulps         0x2337(%rip),%xmm0        # 51e0 <_sk_callback_sse41+0xb0d>
-  DB  102,15,111,13,63,35,0,0             ; movdqa        0x233f(%rip),%xmm1        # 51f0 <_sk_callback_sse41+0xb1d>
+  DB  15,89,5,55,35,0,0                   ; mulps         0x2337(%rip),%xmm0        # 51e0 <_sk_callback_sse41+0xb10>
+  DB  102,15,111,13,63,35,0,0             ; movdqa        0x233f(%rip),%xmm1        # 51f0 <_sk_callback_sse41+0xb20>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,65,35,0,0                  ; mulps         0x2341(%rip),%xmm1        # 5200 <_sk_callback_sse41+0xb2d>
-  DB  102,15,219,21,73,35,0,0             ; pand          0x2349(%rip),%xmm2        # 5210 <_sk_callback_sse41+0xb3d>
+  DB  15,89,13,65,35,0,0                  ; mulps         0x2341(%rip),%xmm1        # 5200 <_sk_callback_sse41+0xb30>
+  DB  102,15,219,21,73,35,0,0             ; pand          0x2349(%rip),%xmm2        # 5210 <_sk_callback_sse41+0xb40>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,79,35,0,0                  ; mulps         0x234f(%rip),%xmm2        # 5220 <_sk_callback_sse41+0xb4d>
+  DB  15,89,21,79,35,0,0                  ; mulps         0x234f(%rip),%xmm2        # 5220 <_sk_callback_sse41+0xb50>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,86,35,0,0                  ; movaps        0x2356(%rip),%xmm3        # 5230 <_sk_callback_sse41+0xb5d>
+  DB  15,40,29,86,35,0,0                  ; movaps        0x2356(%rip),%xmm3        # 5230 <_sk_callback_sse41+0xb60>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_gather_565_sse41
@@ -14325,31 +14323,31 @@ _sk_gather_565_sse41 LABEL PROC
   DB  65,15,183,4,65                      ; movzwl        (%r9,%rax,2),%eax
   DB  102,15,196,192,3                    ; pinsrw        $0x3,%eax,%xmm0
   DB  102,15,56,51,208                    ; pmovzxwd      %xmm0,%xmm2
-  DB  102,15,111,5,251,34,0,0             ; movdqa        0x22fb(%rip),%xmm0        # 5240 <_sk_callback_sse41+0xb6d>
+  DB  102,15,111,5,251,34,0,0             ; movdqa        0x22fb(%rip),%xmm0        # 5240 <_sk_callback_sse41+0xb70>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,253,34,0,0                  ; mulps         0x22fd(%rip),%xmm0        # 5250 <_sk_callback_sse41+0xb7d>
-  DB  102,15,111,13,5,35,0,0              ; movdqa        0x2305(%rip),%xmm1        # 5260 <_sk_callback_sse41+0xb8d>
+  DB  15,89,5,253,34,0,0                  ; mulps         0x22fd(%rip),%xmm0        # 5250 <_sk_callback_sse41+0xb80>
+  DB  102,15,111,13,5,35,0,0              ; movdqa        0x2305(%rip),%xmm1        # 5260 <_sk_callback_sse41+0xb90>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,7,35,0,0                   ; mulps         0x2307(%rip),%xmm1        # 5270 <_sk_callback_sse41+0xb9d>
-  DB  102,15,219,21,15,35,0,0             ; pand          0x230f(%rip),%xmm2        # 5280 <_sk_callback_sse41+0xbad>
+  DB  15,89,13,7,35,0,0                   ; mulps         0x2307(%rip),%xmm1        # 5270 <_sk_callback_sse41+0xba0>
+  DB  102,15,219,21,15,35,0,0             ; pand          0x230f(%rip),%xmm2        # 5280 <_sk_callback_sse41+0xbb0>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,21,35,0,0                  ; mulps         0x2315(%rip),%xmm2        # 5290 <_sk_callback_sse41+0xbbd>
+  DB  15,89,21,21,35,0,0                  ; mulps         0x2315(%rip),%xmm2        # 5290 <_sk_callback_sse41+0xbc0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,28,35,0,0                  ; movaps        0x231c(%rip),%xmm3        # 52a0 <_sk_callback_sse41+0xbcd>
+  DB  15,40,29,28,35,0,0                  ; movaps        0x231c(%rip),%xmm3        # 52a0 <_sk_callback_sse41+0xbd0>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_store_565_sse41
 _sk_store_565_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,29,35,0,0                ; movaps        0x231d(%rip),%xmm8        # 52b0 <_sk_callback_sse41+0xbdd>
+  DB  68,15,40,5,29,35,0,0                ; movaps        0x231d(%rip),%xmm8        # 52b0 <_sk_callback_sse41+0xbe0>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
   DB  102,65,15,114,241,11                ; pslld         $0xb,%xmm9
-  DB  68,15,40,21,18,35,0,0               ; movaps        0x2312(%rip),%xmm10        # 52c0 <_sk_callback_sse41+0xbed>
+  DB  68,15,40,21,18,35,0,0               ; movaps        0x2312(%rip),%xmm10        # 52c0 <_sk_callback_sse41+0xbf0>
   DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
   DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
   DB  102,65,15,114,242,5                 ; pslld         $0x5,%xmm10
@@ -14367,21 +14365,21 @@ _sk_load_4444_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  102,15,56,51,28,120                 ; pmovzxwd      (%rax,%rdi,2),%xmm3
-  DB  102,15,111,5,221,34,0,0             ; movdqa        0x22dd(%rip),%xmm0        # 52d0 <_sk_callback_sse41+0xbfd>
+  DB  102,15,111,5,221,34,0,0             ; movdqa        0x22dd(%rip),%xmm0        # 52d0 <_sk_callback_sse41+0xc00>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,223,34,0,0                  ; mulps         0x22df(%rip),%xmm0        # 52e0 <_sk_callback_sse41+0xc0d>
-  DB  102,15,111,13,231,34,0,0            ; movdqa        0x22e7(%rip),%xmm1        # 52f0 <_sk_callback_sse41+0xc1d>
+  DB  15,89,5,223,34,0,0                  ; mulps         0x22df(%rip),%xmm0        # 52e0 <_sk_callback_sse41+0xc10>
+  DB  102,15,111,13,231,34,0,0            ; movdqa        0x22e7(%rip),%xmm1        # 52f0 <_sk_callback_sse41+0xc20>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,233,34,0,0                 ; mulps         0x22e9(%rip),%xmm1        # 5300 <_sk_callback_sse41+0xc2d>
-  DB  102,15,111,21,241,34,0,0            ; movdqa        0x22f1(%rip),%xmm2        # 5310 <_sk_callback_sse41+0xc3d>
+  DB  15,89,13,233,34,0,0                 ; mulps         0x22e9(%rip),%xmm1        # 5300 <_sk_callback_sse41+0xc30>
+  DB  102,15,111,21,241,34,0,0            ; movdqa        0x22f1(%rip),%xmm2        # 5310 <_sk_callback_sse41+0xc40>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,243,34,0,0                 ; mulps         0x22f3(%rip),%xmm2        # 5320 <_sk_callback_sse41+0xc4d>
-  DB  102,15,219,29,251,34,0,0            ; pand          0x22fb(%rip),%xmm3        # 5330 <_sk_callback_sse41+0xc5d>
+  DB  15,89,21,243,34,0,0                 ; mulps         0x22f3(%rip),%xmm2        # 5320 <_sk_callback_sse41+0xc50>
+  DB  102,15,219,29,251,34,0,0            ; pand          0x22fb(%rip),%xmm3        # 5330 <_sk_callback_sse41+0xc60>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,1,35,0,0                   ; mulps         0x2301(%rip),%xmm3        # 5340 <_sk_callback_sse41+0xc6d>
+  DB  15,89,29,1,35,0,0                   ; mulps         0x2301(%rip),%xmm3        # 5340 <_sk_callback_sse41+0xc70>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -14408,21 +14406,21 @@ _sk_gather_4444_sse41 LABEL PROC
   DB  65,15,183,4,65                      ; movzwl        (%r9,%rax,2),%eax
   DB  102,15,196,192,3                    ; pinsrw        $0x3,%eax,%xmm0
   DB  102,15,56,51,216                    ; pmovzxwd      %xmm0,%xmm3
-  DB  102,15,111,5,164,34,0,0             ; movdqa        0x22a4(%rip),%xmm0        # 5350 <_sk_callback_sse41+0xc7d>
+  DB  102,15,111,5,164,34,0,0             ; movdqa        0x22a4(%rip),%xmm0        # 5350 <_sk_callback_sse41+0xc80>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,166,34,0,0                  ; mulps         0x22a6(%rip),%xmm0        # 5360 <_sk_callback_sse41+0xc8d>
-  DB  102,15,111,13,174,34,0,0            ; movdqa        0x22ae(%rip),%xmm1        # 5370 <_sk_callback_sse41+0xc9d>
+  DB  15,89,5,166,34,0,0                  ; mulps         0x22a6(%rip),%xmm0        # 5360 <_sk_callback_sse41+0xc90>
+  DB  102,15,111,13,174,34,0,0            ; movdqa        0x22ae(%rip),%xmm1        # 5370 <_sk_callback_sse41+0xca0>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,176,34,0,0                 ; mulps         0x22b0(%rip),%xmm1        # 5380 <_sk_callback_sse41+0xcad>
-  DB  102,15,111,21,184,34,0,0            ; movdqa        0x22b8(%rip),%xmm2        # 5390 <_sk_callback_sse41+0xcbd>
+  DB  15,89,13,176,34,0,0                 ; mulps         0x22b0(%rip),%xmm1        # 5380 <_sk_callback_sse41+0xcb0>
+  DB  102,15,111,21,184,34,0,0            ; movdqa        0x22b8(%rip),%xmm2        # 5390 <_sk_callback_sse41+0xcc0>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,186,34,0,0                 ; mulps         0x22ba(%rip),%xmm2        # 53a0 <_sk_callback_sse41+0xccd>
-  DB  102,15,219,29,194,34,0,0            ; pand          0x22c2(%rip),%xmm3        # 53b0 <_sk_callback_sse41+0xcdd>
+  DB  15,89,21,186,34,0,0                 ; mulps         0x22ba(%rip),%xmm2        # 53a0 <_sk_callback_sse41+0xcd0>
+  DB  102,15,219,29,194,34,0,0            ; pand          0x22c2(%rip),%xmm3        # 53b0 <_sk_callback_sse41+0xce0>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,200,34,0,0                 ; mulps         0x22c8(%rip),%xmm3        # 53c0 <_sk_callback_sse41+0xced>
+  DB  15,89,29,200,34,0,0                 ; mulps         0x22c8(%rip),%xmm3        # 53c0 <_sk_callback_sse41+0xcf0>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -14430,7 +14428,7 @@ PUBLIC _sk_store_4444_sse41
 _sk_store_4444_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,199,34,0,0               ; movaps        0x22c7(%rip),%xmm8        # 53d0 <_sk_callback_sse41+0xcfd>
+  DB  68,15,40,5,199,34,0,0               ; movaps        0x22c7(%rip),%xmm8        # 53d0 <_sk_callback_sse41+0xd00>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -14458,17 +14456,17 @@ _sk_load_8888_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  15,16,28,184                        ; movups        (%rax,%rdi,4),%xmm3
-  DB  15,40,5,102,34,0,0                  ; movaps        0x2266(%rip),%xmm0        # 53e0 <_sk_callback_sse41+0xd0d>
+  DB  15,40,5,102,34,0,0                  ; movaps        0x2266(%rip),%xmm0        # 53e0 <_sk_callback_sse41+0xd10>
   DB  15,84,195                           ; andps         %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,104,34,0,0               ; movaps        0x2268(%rip),%xmm8        # 53f0 <_sk_callback_sse41+0xd1d>
+  DB  68,15,40,5,104,34,0,0               ; movaps        0x2268(%rip),%xmm8        # 53f0 <_sk_callback_sse41+0xd20>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,40,203                           ; movaps        %xmm3,%xmm1
-  DB  102,15,56,0,13,104,34,0,0           ; pshufb        0x2268(%rip),%xmm1        # 5400 <_sk_callback_sse41+0xd2d>
+  DB  102,15,56,0,13,104,34,0,0           ; pshufb        0x2268(%rip),%xmm1        # 5400 <_sk_callback_sse41+0xd30>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  15,40,211                           ; movaps        %xmm3,%xmm2
-  DB  102,15,56,0,21,101,34,0,0           ; pshufb        0x2265(%rip),%xmm2        # 5410 <_sk_callback_sse41+0xd3d>
+  DB  102,15,56,0,21,101,34,0,0           ; pshufb        0x2265(%rip),%xmm2        # 5410 <_sk_callback_sse41+0xd40>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -14497,17 +14495,17 @@ _sk_gather_8888_sse41 LABEL PROC
   DB  102,65,15,58,34,28,129,1            ; pinsrd        $0x1,(%r9,%rax,4),%xmm3
   DB  102,67,15,58,34,28,145,2            ; pinsrd        $0x2,(%r9,%r10,4),%xmm3
   DB  102,65,15,58,34,28,137,3            ; pinsrd        $0x3,(%r9,%rcx,4),%xmm3
-  DB  102,15,111,5,254,33,0,0             ; movdqa        0x21fe(%rip),%xmm0        # 5420 <_sk_callback_sse41+0xd4d>
+  DB  102,15,111,5,254,33,0,0             ; movdqa        0x21fe(%rip),%xmm0        # 5420 <_sk_callback_sse41+0xd50>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,255,33,0,0               ; movaps        0x21ff(%rip),%xmm8        # 5430 <_sk_callback_sse41+0xd5d>
+  DB  68,15,40,5,255,33,0,0               ; movaps        0x21ff(%rip),%xmm8        # 5430 <_sk_callback_sse41+0xd60>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
-  DB  102,15,56,0,13,254,33,0,0           ; pshufb        0x21fe(%rip),%xmm1        # 5440 <_sk_callback_sse41+0xd6d>
+  DB  102,15,56,0,13,254,33,0,0           ; pshufb        0x21fe(%rip),%xmm1        # 5440 <_sk_callback_sse41+0xd70>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
-  DB  102,15,56,0,21,250,33,0,0           ; pshufb        0x21fa(%rip),%xmm2        # 5450 <_sk_callback_sse41+0xd7d>
+  DB  102,15,56,0,21,250,33,0,0           ; pshufb        0x21fa(%rip),%xmm2        # 5450 <_sk_callback_sse41+0xd80>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -14520,7 +14518,7 @@ PUBLIC _sk_store_8888_sse41
 _sk_store_8888_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,230,33,0,0               ; movaps        0x21e6(%rip),%xmm8        # 5460 <_sk_callback_sse41+0xd8d>
+  DB  68,15,40,5,230,33,0,0               ; movaps        0x21e6(%rip),%xmm8        # 5460 <_sk_callback_sse41+0xd90>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -14555,18 +14553,18 @@ _sk_load_f16_sse41 LABEL PROC
   DB  102,68,15,97,216                    ; punpcklwd     %xmm0,%xmm11
   DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
   DB  102,65,15,56,51,203                 ; pmovzxwd      %xmm11,%xmm1
-  DB  102,68,15,111,5,95,33,0,0           ; movdqa        0x215f(%rip),%xmm8        # 5470 <_sk_callback_sse41+0xd9d>
+  DB  102,68,15,111,5,95,33,0,0           ; movdqa        0x215f(%rip),%xmm8        # 5470 <_sk_callback_sse41+0xda0>
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
   DB  102,15,239,202                      ; pxor          %xmm2,%xmm1
-  DB  102,15,111,29,90,33,0,0             ; movdqa        0x215a(%rip),%xmm3        # 5480 <_sk_callback_sse41+0xdad>
+  DB  102,15,111,29,90,33,0,0             ; movdqa        0x215a(%rip),%xmm3        # 5480 <_sk_callback_sse41+0xdb0>
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,56,63,195                    ; pmaxud        %xmm3,%xmm0
   DB  102,15,118,193                      ; pcmpeqd       %xmm1,%xmm0
   DB  102,15,114,241,13                   ; pslld         $0xd,%xmm1
   DB  102,15,235,202                      ; por           %xmm2,%xmm1
-  DB  102,68,15,111,21,70,33,0,0          ; movdqa        0x2146(%rip),%xmm10        # 5490 <_sk_callback_sse41+0xdbd>
+  DB  102,68,15,111,21,70,33,0,0          ; movdqa        0x2146(%rip),%xmm10        # 5490 <_sk_callback_sse41+0xdc0>
   DB  102,65,15,254,202                   ; paddd         %xmm10,%xmm1
   DB  102,15,219,193                      ; pand          %xmm1,%xmm0
   DB  102,65,15,115,219,8                 ; psrldq        $0x8,%xmm11
@@ -14637,18 +14635,18 @@ _sk_gather_f16_sse41 LABEL PROC
   DB  102,68,15,97,218                    ; punpcklwd     %xmm2,%xmm11
   DB  102,68,15,105,202                   ; punpckhwd     %xmm2,%xmm9
   DB  102,65,15,56,51,203                 ; pmovzxwd      %xmm11,%xmm1
-  DB  102,68,15,111,5,4,32,0,0            ; movdqa        0x2004(%rip),%xmm8        # 54a0 <_sk_callback_sse41+0xdcd>
+  DB  102,68,15,111,5,4,32,0,0            ; movdqa        0x2004(%rip),%xmm8        # 54a0 <_sk_callback_sse41+0xdd0>
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
   DB  102,15,239,202                      ; pxor          %xmm2,%xmm1
-  DB  102,15,111,29,255,31,0,0            ; movdqa        0x1fff(%rip),%xmm3        # 54b0 <_sk_callback_sse41+0xddd>
+  DB  102,15,111,29,255,31,0,0            ; movdqa        0x1fff(%rip),%xmm3        # 54b0 <_sk_callback_sse41+0xde0>
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,56,63,195                    ; pmaxud        %xmm3,%xmm0
   DB  102,15,118,193                      ; pcmpeqd       %xmm1,%xmm0
   DB  102,15,114,241,13                   ; pslld         $0xd,%xmm1
   DB  102,15,235,202                      ; por           %xmm2,%xmm1
-  DB  102,68,15,111,21,235,31,0,0         ; movdqa        0x1feb(%rip),%xmm10        # 54c0 <_sk_callback_sse41+0xded>
+  DB  102,68,15,111,21,235,31,0,0         ; movdqa        0x1feb(%rip),%xmm10        # 54c0 <_sk_callback_sse41+0xdf0>
   DB  102,65,15,254,202                   ; paddd         %xmm10,%xmm1
   DB  102,15,219,193                      ; pand          %xmm1,%xmm0
   DB  102,65,15,115,219,8                 ; psrldq        $0x8,%xmm11
@@ -14694,17 +14692,17 @@ PUBLIC _sk_store_f16_sse41
 _sk_store_f16_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  102,68,15,111,21,33,31,0,0          ; movdqa        0x1f21(%rip),%xmm10        # 54d0 <_sk_callback_sse41+0xdfd>
+  DB  102,68,15,111,21,33,31,0,0          ; movdqa        0x1f21(%rip),%xmm10        # 54d0 <_sk_callback_sse41+0xe00>
   DB  102,68,15,111,224                   ; movdqa        %xmm0,%xmm12
   DB  102,68,15,111,232                   ; movdqa        %xmm0,%xmm13
   DB  102,69,15,219,234                   ; pand          %xmm10,%xmm13
   DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
-  DB  102,68,15,111,13,20,31,0,0          ; movdqa        0x1f14(%rip),%xmm9        # 54e0 <_sk_callback_sse41+0xe0d>
+  DB  102,68,15,111,13,20,31,0,0          ; movdqa        0x1f14(%rip),%xmm9        # 54e0 <_sk_callback_sse41+0xe10>
   DB  102,65,15,114,213,16                ; psrld         $0x10,%xmm13
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
   DB  102,69,15,102,196                   ; pcmpgtd       %xmm12,%xmm8
   DB  102,65,15,114,212,13                ; psrld         $0xd,%xmm12
-  DB  102,68,15,111,29,5,31,0,0           ; movdqa        0x1f05(%rip),%xmm11        # 54f0 <_sk_callback_sse41+0xe1d>
+  DB  102,68,15,111,29,5,31,0,0           ; movdqa        0x1f05(%rip),%xmm11        # 54f0 <_sk_callback_sse41+0xe20>
   DB  102,69,15,235,235                   ; por           %xmm11,%xmm13
   DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
   DB  102,69,15,223,197                   ; pandn         %xmm13,%xmm8
@@ -14772,7 +14770,7 @@ _sk_load_u16_be_sse41 LABEL PROC
   DB  102,15,235,200                      ; por           %xmm0,%xmm1
   DB  102,15,56,51,193                    ; pmovzxwd      %xmm1,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,212,29,0,0               ; movaps        0x1dd4(%rip),%xmm8        # 5500 <_sk_callback_sse41+0xe2d>
+  DB  68,15,40,5,212,29,0,0               ; movaps        0x1dd4(%rip),%xmm8        # 5500 <_sk_callback_sse41+0xe30>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -14822,7 +14820,7 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC
   DB  102,15,235,193                      ; por           %xmm1,%xmm0
   DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,21,29,0,0                ; movaps        0x1d15(%rip),%xmm8        # 5510 <_sk_callback_sse41+0xe3d>
+  DB  68,15,40,5,21,29,0,0                ; movaps        0x1d15(%rip),%xmm8        # 5510 <_sk_callback_sse41+0xe40>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -14839,14 +14837,14 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,220,28,0,0                 ; movaps        0x1cdc(%rip),%xmm3        # 5520 <_sk_callback_sse41+0xe4d>
+  DB  15,40,29,220,28,0,0                 ; movaps        0x1cdc(%rip),%xmm3        # 5520 <_sk_callback_sse41+0xe50>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_store_u16_be_sse41
 _sk_store_u16_be_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,13,221,28,0,0              ; movaps        0x1cdd(%rip),%xmm9        # 5530 <_sk_callback_sse41+0xe5d>
+  DB  68,15,40,13,221,28,0,0              ; movaps        0x1cdd(%rip),%xmm9        # 5530 <_sk_callback_sse41+0xe60>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
@@ -15039,10 +15037,10 @@ _sk_mirror_y_sse41 LABEL PROC
 PUBLIC _sk_luminance_to_alpha_sse41
 _sk_luminance_to_alpha_sse41 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
-  DB  15,89,5,57,26,0,0                   ; mulps         0x1a39(%rip),%xmm0        # 5540 <_sk_callback_sse41+0xe6d>
-  DB  15,89,13,66,26,0,0                  ; mulps         0x1a42(%rip),%xmm1        # 5550 <_sk_callback_sse41+0xe7d>
+  DB  15,89,5,57,26,0,0                   ; mulps         0x1a39(%rip),%xmm0        # 5540 <_sk_callback_sse41+0xe70>
+  DB  15,89,13,66,26,0,0                  ; mulps         0x1a42(%rip),%xmm1        # 5550 <_sk_callback_sse41+0xe80>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,29,72,26,0,0                  ; mulps         0x1a48(%rip),%xmm3        # 5560 <_sk_callback_sse41+0xe8d>
+  DB  15,89,29,72,26,0,0                  ; mulps         0x1a48(%rip),%xmm3        # 5560 <_sk_callback_sse41+0xe90>
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -15354,7 +15352,7 @@ _sk_gradient_sse41 LABEL PROC
   DB  73,255,200                          ; dec           %r8
   DB  72,131,193,4                        ; add           $0x4,%rcx
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
-  DB  15,40,21,253,20,0,0                 ; movaps        0x14fd(%rip),%xmm2        # 5570 <_sk_callback_sse41+0xe9d>
+  DB  15,40,21,253,20,0,0                 ; movaps        0x14fd(%rip),%xmm2        # 5570 <_sk_callback_sse41+0xea0>
   DB  243,15,16,25                        ; movss         (%rcx),%xmm3
   DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
   DB  15,194,216,2                        ; cmpleps       %xmm0,%xmm3
@@ -15489,26 +15487,26 @@ _sk_xy_to_unit_angle_sse41 LABEL PROC
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
   DB  69,15,89,237                        ; mulps         %xmm13,%xmm13
-  DB  68,15,40,21,159,18,0,0              ; movaps        0x129f(%rip),%xmm10        # 5580 <_sk_callback_sse41+0xead>
+  DB  68,15,40,21,159,18,0,0              ; movaps        0x129f(%rip),%xmm10        # 5580 <_sk_callback_sse41+0xeb0>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,163,18,0,0              ; addps         0x12a3(%rip),%xmm10        # 5590 <_sk_callback_sse41+0xebd>
+  DB  68,15,88,21,163,18,0,0              ; addps         0x12a3(%rip),%xmm10        # 5590 <_sk_callback_sse41+0xec0>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,167,18,0,0              ; addps         0x12a7(%rip),%xmm10        # 55a0 <_sk_callback_sse41+0xecd>
+  DB  68,15,88,21,167,18,0,0              ; addps         0x12a7(%rip),%xmm10        # 55a0 <_sk_callback_sse41+0xed0>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,171,18,0,0              ; addps         0x12ab(%rip),%xmm10        # 55b0 <_sk_callback_sse41+0xedd>
+  DB  68,15,88,21,171,18,0,0              ; addps         0x12ab(%rip),%xmm10        # 55b0 <_sk_callback_sse41+0xee0>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,194,195,1                     ; cmpltps       %xmm11,%xmm0
-  DB  68,15,40,29,170,18,0,0              ; movaps        0x12aa(%rip),%xmm11        # 55c0 <_sk_callback_sse41+0xeed>
+  DB  68,15,40,29,170,18,0,0              ; movaps        0x12aa(%rip),%xmm11        # 55c0 <_sk_callback_sse41+0xef0>
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  102,69,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm10
   DB  69,15,194,200,1                     ; cmpltps       %xmm8,%xmm9
-  DB  68,15,40,29,163,18,0,0              ; movaps        0x12a3(%rip),%xmm11        # 55d0 <_sk_callback_sse41+0xefd>
+  DB  68,15,40,29,163,18,0,0              ; movaps        0x12a3(%rip),%xmm11        # 55d0 <_sk_callback_sse41+0xf00>
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  102,69,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm10
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  65,15,194,192,1                     ; cmpltps       %xmm8,%xmm0
-  DB  68,15,40,13,149,18,0,0              ; movaps        0x1295(%rip),%xmm9        # 55e0 <_sk_callback_sse41+0xf0d>
+  DB  68,15,40,13,149,18,0,0              ; movaps        0x1295(%rip),%xmm9        # 55e0 <_sk_callback_sse41+0xf10>
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
   DB  102,69,15,56,20,209                 ; blendvps      %xmm0,%xmm9,%xmm10
   DB  69,15,194,194,7                     ; cmpordps      %xmm10,%xmm8
@@ -15523,15 +15521,14 @@ _sk_xy_to_radius_sse41 LABEL PROC
   DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
   DB  68,15,88,192                        ; addps         %xmm0,%xmm8
-  DB  65,15,82,192                        ; rsqrtps       %xmm8,%xmm0
-  DB  15,83,192                           ; rcpps         %xmm0,%xmm0
+  DB  65,15,81,192                        ; sqrtps        %xmm8,%xmm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_save_xy_sse41
 _sk_save_xy_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,102,18,0,0               ; movaps        0x1266(%rip),%xmm8        # 55f0 <_sk_callback_sse41+0xf1d>
+  DB  68,15,40,5,105,18,0,0               ; movaps        0x1269(%rip),%xmm8        # 55f0 <_sk_callback_sse41+0xf20>
   DB  15,17,0                             ; movups        %xmm0,(%rax)
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,88,200                        ; addps         %xmm8,%xmm9
@@ -15571,8 +15568,8 @@ _sk_bilinear_nx_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,232,17,0,0                  ; addps         0x11e8(%rip),%xmm0        # 5600 <_sk_callback_sse41+0xf2d>
-  DB  68,15,40,13,240,17,0,0              ; movaps        0x11f0(%rip),%xmm9        # 5610 <_sk_callback_sse41+0xf3d>
+  DB  15,88,5,235,17,0,0                  ; addps         0x11eb(%rip),%xmm0        # 5600 <_sk_callback_sse41+0xf30>
+  DB  68,15,40,13,243,17,0,0              ; movaps        0x11f3(%rip),%xmm9        # 5610 <_sk_callback_sse41+0xf40>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -15583,7 +15580,7 @@ _sk_bilinear_px_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,223,17,0,0                  ; addps         0x11df(%rip),%xmm0        # 5620 <_sk_callback_sse41+0xf4d>
+  DB  15,88,5,226,17,0,0                  ; addps         0x11e2(%rip),%xmm0        # 5620 <_sk_callback_sse41+0xf50>
   DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -15593,8 +15590,8 @@ _sk_bilinear_ny_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,209,17,0,0                 ; addps         0x11d1(%rip),%xmm1        # 5630 <_sk_callback_sse41+0xf5d>
-  DB  68,15,40,13,217,17,0,0              ; movaps        0x11d9(%rip),%xmm9        # 5640 <_sk_callback_sse41+0xf6d>
+  DB  15,88,13,212,17,0,0                 ; addps         0x11d4(%rip),%xmm1        # 5630 <_sk_callback_sse41+0xf60>
+  DB  68,15,40,13,220,17,0,0              ; movaps        0x11dc(%rip),%xmm9        # 5640 <_sk_callback_sse41+0xf70>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -15605,7 +15602,7 @@ _sk_bilinear_py_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,199,17,0,0                 ; addps         0x11c7(%rip),%xmm1        # 5650 <_sk_callback_sse41+0xf7d>
+  DB  15,88,13,202,17,0,0                 ; addps         0x11ca(%rip),%xmm1        # 5650 <_sk_callback_sse41+0xf80>
   DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -15615,13 +15612,13 @@ _sk_bicubic_n3x_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,186,17,0,0                  ; addps         0x11ba(%rip),%xmm0        # 5660 <_sk_callback_sse41+0xf8d>
-  DB  68,15,40,13,194,17,0,0              ; movaps        0x11c2(%rip),%xmm9        # 5670 <_sk_callback_sse41+0xf9d>
+  DB  15,88,5,189,17,0,0                  ; addps         0x11bd(%rip),%xmm0        # 5660 <_sk_callback_sse41+0xf90>
+  DB  68,15,40,13,197,17,0,0              ; movaps        0x11c5(%rip),%xmm9        # 5670 <_sk_callback_sse41+0xfa0>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,190,17,0,0              ; mulps         0x11be(%rip),%xmm9        # 5680 <_sk_callback_sse41+0xfad>
-  DB  68,15,88,13,198,17,0,0              ; addps         0x11c6(%rip),%xmm9        # 5690 <_sk_callback_sse41+0xfbd>
+  DB  68,15,89,13,193,17,0,0              ; mulps         0x11c1(%rip),%xmm9        # 5680 <_sk_callback_sse41+0xfb0>
+  DB  68,15,88,13,201,17,0,0              ; addps         0x11c9(%rip),%xmm9        # 5690 <_sk_callback_sse41+0xfc0>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -15632,16 +15629,16 @@ _sk_bicubic_n1x_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,181,17,0,0                  ; addps         0x11b5(%rip),%xmm0        # 56a0 <_sk_callback_sse41+0xfcd>
-  DB  68,15,40,13,189,17,0,0              ; movaps        0x11bd(%rip),%xmm9        # 56b0 <_sk_callback_sse41+0xfdd>
+  DB  15,88,5,184,17,0,0                  ; addps         0x11b8(%rip),%xmm0        # 56a0 <_sk_callback_sse41+0xfd0>
+  DB  68,15,40,13,192,17,0,0              ; movaps        0x11c0(%rip),%xmm9        # 56b0 <_sk_callback_sse41+0xfe0>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,193,17,0,0               ; movaps        0x11c1(%rip),%xmm8        # 56c0 <_sk_callback_sse41+0xfed>
+  DB  68,15,40,5,196,17,0,0               ; movaps        0x11c4(%rip),%xmm8        # 56c0 <_sk_callback_sse41+0xff0>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,197,17,0,0               ; addps         0x11c5(%rip),%xmm8        # 56d0 <_sk_callback_sse41+0xffd>
+  DB  68,15,88,5,200,17,0,0               ; addps         0x11c8(%rip),%xmm8        # 56d0 <_sk_callback_sse41+0x1000>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,201,17,0,0               ; addps         0x11c9(%rip),%xmm8        # 56e0 <_sk_callback_sse41+0x100d>
+  DB  68,15,88,5,204,17,0,0               ; addps         0x11cc(%rip),%xmm8        # 56e0 <_sk_callback_sse41+0x1010>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,205,17,0,0               ; addps         0x11cd(%rip),%xmm8        # 56f0 <_sk_callback_sse41+0x101d>
+  DB  68,15,88,5,208,17,0,0               ; addps         0x11d0(%rip),%xmm8        # 56f0 <_sk_callback_sse41+0x1020>
   DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -15649,17 +15646,17 @@ _sk_bicubic_n1x_sse41 LABEL PROC
 PUBLIC _sk_bicubic_p1x_sse41
 _sk_bicubic_p1x_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,199,17,0,0               ; movaps        0x11c7(%rip),%xmm8        # 5700 <_sk_callback_sse41+0x102d>
+  DB  68,15,40,5,202,17,0,0               ; movaps        0x11ca(%rip),%xmm8        # 5700 <_sk_callback_sse41+0x1030>
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,21,195,17,0,0              ; movaps        0x11c3(%rip),%xmm10        # 5710 <_sk_callback_sse41+0x103d>
+  DB  68,15,40,21,198,17,0,0              ; movaps        0x11c6(%rip),%xmm10        # 5710 <_sk_callback_sse41+0x1040>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,199,17,0,0              ; addps         0x11c7(%rip),%xmm10        # 5720 <_sk_callback_sse41+0x104d>
+  DB  68,15,88,21,202,17,0,0              ; addps         0x11ca(%rip),%xmm10        # 5720 <_sk_callback_sse41+0x1050>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,195,17,0,0              ; addps         0x11c3(%rip),%xmm10        # 5730 <_sk_callback_sse41+0x105d>
+  DB  68,15,88,21,198,17,0,0              ; addps         0x11c6(%rip),%xmm10        # 5730 <_sk_callback_sse41+0x1060>
   DB  68,15,17,144,128,0,0,0              ; movups        %xmm10,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -15669,11 +15666,11 @@ _sk_bicubic_p3x_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,182,17,0,0                  ; addps         0x11b6(%rip),%xmm0        # 5740 <_sk_callback_sse41+0x106d>
+  DB  15,88,5,185,17,0,0                  ; addps         0x11b9(%rip),%xmm0        # 5740 <_sk_callback_sse41+0x1070>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,182,17,0,0               ; mulps         0x11b6(%rip),%xmm8        # 5750 <_sk_callback_sse41+0x107d>
-  DB  68,15,88,5,190,17,0,0               ; addps         0x11be(%rip),%xmm8        # 5760 <_sk_callback_sse41+0x108d>
+  DB  68,15,89,5,185,17,0,0               ; mulps         0x11b9(%rip),%xmm8        # 5750 <_sk_callback_sse41+0x1080>
+  DB  68,15,88,5,193,17,0,0               ; addps         0x11c1(%rip),%xmm8        # 5760 <_sk_callback_sse41+0x1090>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -15684,13 +15681,13 @@ _sk_bicubic_n3y_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,172,17,0,0                 ; addps         0x11ac(%rip),%xmm1        # 5770 <_sk_callback_sse41+0x109d>
-  DB  68,15,40,13,180,17,0,0              ; movaps        0x11b4(%rip),%xmm9        # 5780 <_sk_callback_sse41+0x10ad>
+  DB  15,88,13,175,17,0,0                 ; addps         0x11af(%rip),%xmm1        # 5770 <_sk_callback_sse41+0x10a0>
+  DB  68,15,40,13,183,17,0,0              ; movaps        0x11b7(%rip),%xmm9        # 5780 <_sk_callback_sse41+0x10b0>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,176,17,0,0              ; mulps         0x11b0(%rip),%xmm9        # 5790 <_sk_callback_sse41+0x10bd>
-  DB  68,15,88,13,184,17,0,0              ; addps         0x11b8(%rip),%xmm9        # 57a0 <_sk_callback_sse41+0x10cd>
+  DB  68,15,89,13,179,17,0,0              ; mulps         0x11b3(%rip),%xmm9        # 5790 <_sk_callback_sse41+0x10c0>
+  DB  68,15,88,13,187,17,0,0              ; addps         0x11bb(%rip),%xmm9        # 57a0 <_sk_callback_sse41+0x10d0>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -15701,16 +15698,16 @@ _sk_bicubic_n1y_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,166,17,0,0                 ; addps         0x11a6(%rip),%xmm1        # 57b0 <_sk_callback_sse41+0x10dd>
-  DB  68,15,40,13,174,17,0,0              ; movaps        0x11ae(%rip),%xmm9        # 57c0 <_sk_callback_sse41+0x10ed>
+  DB  15,88,13,169,17,0,0                 ; addps         0x11a9(%rip),%xmm1        # 57b0 <_sk_callback_sse41+0x10e0>
+  DB  68,15,40,13,177,17,0,0              ; movaps        0x11b1(%rip),%xmm9        # 57c0 <_sk_callback_sse41+0x10f0>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,178,17,0,0               ; movaps        0x11b2(%rip),%xmm8        # 57d0 <_sk_callback_sse41+0x10fd>
+  DB  68,15,40,5,181,17,0,0               ; movaps        0x11b5(%rip),%xmm8        # 57d0 <_sk_callback_sse41+0x1100>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,182,17,0,0               ; addps         0x11b6(%rip),%xmm8        # 57e0 <_sk_callback_sse41+0x110d>
+  DB  68,15,88,5,185,17,0,0               ; addps         0x11b9(%rip),%xmm8        # 57e0 <_sk_callback_sse41+0x1110>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,186,17,0,0               ; addps         0x11ba(%rip),%xmm8        # 57f0 <_sk_callback_sse41+0x111d>
+  DB  68,15,88,5,189,17,0,0               ; addps         0x11bd(%rip),%xmm8        # 57f0 <_sk_callback_sse41+0x1120>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,190,17,0,0               ; addps         0x11be(%rip),%xmm8        # 5800 <_sk_callback_sse41+0x112d>
+  DB  68,15,88,5,193,17,0,0               ; addps         0x11c1(%rip),%xmm8        # 5800 <_sk_callback_sse41+0x1130>
   DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -15718,17 +15715,17 @@ _sk_bicubic_n1y_sse41 LABEL PROC
 PUBLIC _sk_bicubic_p1y_sse41
 _sk_bicubic_p1y_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,184,17,0,0               ; movaps        0x11b8(%rip),%xmm8        # 5810 <_sk_callback_sse41+0x113d>
+  DB  68,15,40,5,187,17,0,0               ; movaps        0x11bb(%rip),%xmm8        # 5810 <_sk_callback_sse41+0x1140>
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  68,15,40,21,179,17,0,0              ; movaps        0x11b3(%rip),%xmm10        # 5820 <_sk_callback_sse41+0x114d>
+  DB  68,15,40,21,182,17,0,0              ; movaps        0x11b6(%rip),%xmm10        # 5820 <_sk_callback_sse41+0x1150>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,183,17,0,0              ; addps         0x11b7(%rip),%xmm10        # 5830 <_sk_callback_sse41+0x115d>
+  DB  68,15,88,21,186,17,0,0              ; addps         0x11ba(%rip),%xmm10        # 5830 <_sk_callback_sse41+0x1160>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,179,17,0,0              ; addps         0x11b3(%rip),%xmm10        # 5840 <_sk_callback_sse41+0x116d>
+  DB  68,15,88,21,182,17,0,0              ; addps         0x11b6(%rip),%xmm10        # 5840 <_sk_callback_sse41+0x1170>
   DB  68,15,17,144,160,0,0,0              ; movups        %xmm10,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -15738,11 +15735,11 @@ _sk_bicubic_p3y_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,165,17,0,0                 ; addps         0x11a5(%rip),%xmm1        # 5850 <_sk_callback_sse41+0x117d>
+  DB  15,88,13,168,17,0,0                 ; addps         0x11a8(%rip),%xmm1        # 5850 <_sk_callback_sse41+0x1180>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,165,17,0,0               ; mulps         0x11a5(%rip),%xmm8        # 5860 <_sk_callback_sse41+0x118d>
-  DB  68,15,88,5,173,17,0,0               ; addps         0x11ad(%rip),%xmm8        # 5870 <_sk_callback_sse41+0x119d>
+  DB  68,15,89,5,168,17,0,0               ; mulps         0x11a8(%rip),%xmm8        # 5860 <_sk_callback_sse41+0x1190>
+  DB  68,15,88,5,176,17,0,0               ; addps         0x11b0(%rip),%xmm8        # 5870 <_sk_callback_sse41+0x11a0>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -16334,10 +16331,10 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004c68 <_sk_callback_sse41+0xa000595>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004c68 <_sk_callback_sse41+0xa000598>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3004c70 <_sk_callback_sse41+0x300059d>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3004c70 <_sk_callback_sse41+0x30005a0>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -16452,7 +16449,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38d6a <_sk_callback_sse41+0xffffffffe9a34697>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38d6a <_sk_callback_sse41+0xffffffffe9a3469a>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16548,7 +16545,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38e3a <_sk_callback_sse41+0xffffffffe9a34767>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38e3a <_sk_callback_sse41+0xffffffffe9a3476a>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16644,7 +16641,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38f0a <_sk_callback_sse41+0xffffffffe9a34837>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38f0a <_sk_callback_sse41+0xffffffffe9a3483a>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16740,7 +16737,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38fda <_sk_callback_sse41+0xffffffffe9a34907>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38fda <_sk_callback_sse41+0xffffffffe9a3490a>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16898,7 +16895,7 @@ ALIGN 16
   DB  5,255,255,255,9                     ; add           $0x9ffffff,%eax
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 30051c0 <_sk_callback_sse41+0x3000aed>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 30051c0 <_sk_callback_sse41+0x3000af0>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -17140,7 +17137,7 @@ ALIGN 16
   DB  5,255,255,255,9                     ; add           $0x9ffffff,%eax
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005410 <_sk_callback_sse41+0x3000d3d>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005410 <_sk_callback_sse41+0x3000d40>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -17167,7 +17164,7 @@ ALIGN 16
   DB  5,255,255,255,9                     ; add           $0x9ffffff,%eax
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005450 <_sk_callback_sse41+0x3000d7d>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005450 <_sk_callback_sse41+0x3000d80>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -17450,7 +17447,7 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e535 <_sk_callback_sse41+0x3d639e62>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e535 <_sk_callback_sse41+0x3d639e65>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -17476,7 +17473,7 @@ ALIGN 16
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e575 <_sk_callback_sse41+0x3d639ea2>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e575 <_sk_callback_sse41+0x3d639ea5>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -17560,7 +17557,7 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e645 <_sk_callback_sse41+0x3d639f72>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e645 <_sk_callback_sse41+0x3d639f75>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -17586,7 +17583,7 @@ ALIGN 16
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e685 <_sk_callback_sse41+0x3d639fb2>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e685 <_sk_callback_sse41+0x3d639fb5>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -17597,11 +17594,11 @@ ALIGN 16
   DB  63                                  ; (bad)
   DB  114,28                              ; jb            587e <.literal16+0x10fe>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5882 <_sk_callback_sse41+0x11af>
+  DB  62,114,28                           ; jb,pt         5882 <_sk_callback_sse41+0x11b2>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5886 <_sk_callback_sse41+0x11b3>
+  DB  62,114,28                           ; jb,pt         5886 <_sk_callback_sse41+0x11b6>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         588a <_sk_callback_sse41+0x11b7>
+  DB  62,114,28                           ; jb,pt         588a <_sk_callback_sse41+0x11ba>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -17692,7 +17689,7 @@ _sk_seed_shader_sse2 LABEL PROC
   DB  102,15,110,199                      ; movd          %edi,%xmm0
   DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
   DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
-  DB  15,40,21,225,74,0,0                 ; movaps        0x4ae1(%rip),%xmm2        # 4bf0 <_sk_callback_sse2+0xb1>
+  DB  15,40,21,225,74,0,0                 ; movaps        0x4ae1(%rip),%xmm2        # 4bf0 <_sk_callback_sse2+0xb4>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  15,16,2                             ; movups        (%rdx),%xmm0
   DB  15,88,193                           ; addps         %xmm1,%xmm0
@@ -17701,7 +17698,7 @@ _sk_seed_shader_sse2 LABEL PROC
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,21,208,74,0,0                 ; movaps        0x4ad0(%rip),%xmm2        # 4c00 <_sk_callback_sse2+0xc1>
+  DB  15,40,21,208,74,0,0                 ; movaps        0x4ad0(%rip),%xmm2        # 4c00 <_sk_callback_sse2+0xc4>
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
@@ -17722,14 +17719,14 @@ _sk_dither_sse2 LABEL PROC
   DB  102,68,15,110,1                     ; movd          (%rcx),%xmm8
   DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
   DB  102,69,15,239,193                   ; pxor          %xmm9,%xmm8
-  DB  102,68,15,111,21,149,74,0,0         ; movdqa        0x4a95(%rip),%xmm10        # 4c10 <_sk_callback_sse2+0xd1>
+  DB  102,68,15,111,21,149,74,0,0         ; movdqa        0x4a95(%rip),%xmm10        # 4c10 <_sk_callback_sse2+0xd4>
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,219,218                   ; pand          %xmm10,%xmm11
   DB  102,65,15,114,243,5                 ; pslld         $0x5,%xmm11
   DB  102,69,15,219,209                   ; pand          %xmm9,%xmm10
   DB  102,65,15,114,242,4                 ; pslld         $0x4,%xmm10
-  DB  102,68,15,111,37,129,74,0,0         ; movdqa        0x4a81(%rip),%xmm12        # 4c20 <_sk_callback_sse2+0xe1>
-  DB  102,68,15,111,45,136,74,0,0         ; movdqa        0x4a88(%rip),%xmm13        # 4c30 <_sk_callback_sse2+0xf1>
+  DB  102,68,15,111,37,129,74,0,0         ; movdqa        0x4a81(%rip),%xmm12        # 4c20 <_sk_callback_sse2+0xe4>
+  DB  102,68,15,111,45,136,74,0,0         ; movdqa        0x4a88(%rip),%xmm13        # 4c30 <_sk_callback_sse2+0xf4>
   DB  102,69,15,111,240                   ; movdqa        %xmm8,%xmm14
   DB  102,69,15,219,245                   ; pand          %xmm13,%xmm14
   DB  102,65,15,114,246,2                 ; pslld         $0x2,%xmm14
@@ -17745,8 +17742,8 @@ _sk_dither_sse2 LABEL PROC
   DB  102,69,15,235,245                   ; por           %xmm13,%xmm14
   DB  102,69,15,235,240                   ; por           %xmm8,%xmm14
   DB  69,15,91,198                        ; cvtdq2ps      %xmm14,%xmm8
-  DB  68,15,89,5,67,74,0,0                ; mulps         0x4a43(%rip),%xmm8        # 4c40 <_sk_callback_sse2+0x101>
-  DB  68,15,88,5,75,74,0,0                ; addps         0x4a4b(%rip),%xmm8        # 4c50 <_sk_callback_sse2+0x111>
+  DB  68,15,89,5,67,74,0,0                ; mulps         0x4a43(%rip),%xmm8        # 4c40 <_sk_callback_sse2+0x104>
+  DB  68,15,88,5,75,74,0,0                ; addps         0x4a4b(%rip),%xmm8        # 4c50 <_sk_callback_sse2+0x114>
   DB  243,68,15,16,72,8                   ; movss         0x8(%rax),%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
@@ -17802,7 +17799,7 @@ _sk_clear_sse2 LABEL PROC
 PUBLIC _sk_srcatop_sse2
 _sk_srcatop_sse2 LABEL PROC
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  68,15,40,5,206,73,0,0               ; movaps        0x49ce(%rip),%xmm8        # 4c60 <_sk_callback_sse2+0x121>
+  DB  68,15,40,5,206,73,0,0               ; movaps        0x49ce(%rip),%xmm8        # 4c60 <_sk_callback_sse2+0x124>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -17825,7 +17822,7 @@ PUBLIC _sk_dstatop_sse2
 _sk_dstatop_sse2 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
   DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
-  DB  68,15,40,13,145,73,0,0              ; movaps        0x4991(%rip),%xmm9        # 4c70 <_sk_callback_sse2+0x131>
+  DB  68,15,40,13,145,73,0,0              ; movaps        0x4991(%rip),%xmm9        # 4c70 <_sk_callback_sse2+0x134>
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
@@ -17866,7 +17863,7 @@ _sk_dstin_sse2 LABEL PROC
 
 PUBLIC _sk_srcout_sse2
 _sk_srcout_sse2 LABEL PROC
-  DB  68,15,40,5,53,73,0,0                ; movaps        0x4935(%rip),%xmm8        # 4c80 <_sk_callback_sse2+0x141>
+  DB  68,15,40,5,53,73,0,0                ; movaps        0x4935(%rip),%xmm8        # 4c80 <_sk_callback_sse2+0x144>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
@@ -17877,7 +17874,7 @@ _sk_srcout_sse2 LABEL PROC
 
 PUBLIC _sk_dstout_sse2
 _sk_dstout_sse2 LABEL PROC
-  DB  68,15,40,5,37,73,0,0                ; movaps        0x4925(%rip),%xmm8        # 4c90 <_sk_callback_sse2+0x151>
+  DB  68,15,40,5,37,73,0,0                ; movaps        0x4925(%rip),%xmm8        # 4c90 <_sk_callback_sse2+0x154>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -17892,7 +17889,7 @@ _sk_dstout_sse2 LABEL PROC
 
 PUBLIC _sk_srcover_sse2
 _sk_srcover_sse2 LABEL PROC
-  DB  68,15,40,5,8,73,0,0                 ; movaps        0x4908(%rip),%xmm8        # 4ca0 <_sk_callback_sse2+0x161>
+  DB  68,15,40,5,8,73,0,0                 ; movaps        0x4908(%rip),%xmm8        # 4ca0 <_sk_callback_sse2+0x164>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -17910,7 +17907,7 @@ _sk_srcover_sse2 LABEL PROC
 
 PUBLIC _sk_dstover_sse2
 _sk_dstover_sse2 LABEL PROC
-  DB  68,15,40,5,220,72,0,0               ; movaps        0x48dc(%rip),%xmm8        # 4cb0 <_sk_callback_sse2+0x171>
+  DB  68,15,40,5,220,72,0,0               ; movaps        0x48dc(%rip),%xmm8        # 4cb0 <_sk_callback_sse2+0x174>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -17934,7 +17931,7 @@ _sk_modulate_sse2 LABEL PROC
 
 PUBLIC _sk_multiply_sse2
 _sk_multiply_sse2 LABEL PROC
-  DB  68,15,40,5,176,72,0,0               ; movaps        0x48b0(%rip),%xmm8        # 4cc0 <_sk_callback_sse2+0x181>
+  DB  68,15,40,5,176,72,0,0               ; movaps        0x48b0(%rip),%xmm8        # 4cc0 <_sk_callback_sse2+0x184>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
@@ -18004,7 +18001,7 @@ _sk_screen_sse2 LABEL PROC
 PUBLIC _sk_xor__sse2
 _sk_xor__sse2 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
-  DB  15,40,29,225,71,0,0                 ; movaps        0x47e1(%rip),%xmm3        # 4cd0 <_sk_callback_sse2+0x191>
+  DB  15,40,29,225,71,0,0                 ; movaps        0x47e1(%rip),%xmm3        # 4cd0 <_sk_callback_sse2+0x194>
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
@@ -18050,7 +18047,7 @@ _sk_darken_sse2 LABEL PROC
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,76,71,0,0                  ; movaps        0x474c(%rip),%xmm2        # 4ce0 <_sk_callback_sse2+0x1a1>
+  DB  15,40,21,76,71,0,0                  ; movaps        0x474c(%rip),%xmm2        # 4ce0 <_sk_callback_sse2+0x1a4>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -18082,7 +18079,7 @@ _sk_lighten_sse2 LABEL PROC
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,241,70,0,0                 ; movaps        0x46f1(%rip),%xmm2        # 4cf0 <_sk_callback_sse2+0x1b1>
+  DB  15,40,21,241,70,0,0                 ; movaps        0x46f1(%rip),%xmm2        # 4cf0 <_sk_callback_sse2+0x1b4>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -18117,7 +18114,7 @@ _sk_difference_sse2 LABEL PROC
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,139,70,0,0                 ; movaps        0x468b(%rip),%xmm2        # 4d00 <_sk_callback_sse2+0x1c1>
+  DB  15,40,21,139,70,0,0                 ; movaps        0x468b(%rip),%xmm2        # 4d00 <_sk_callback_sse2+0x1c4>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -18142,7 +18139,7 @@ _sk_exclusion_sse2 LABEL PROC
   DB  15,89,214                           ; mulps         %xmm6,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,202                        ; subps         %xmm2,%xmm9
-  DB  15,40,13,76,70,0,0                  ; movaps        0x464c(%rip),%xmm1        # 4d10 <_sk_callback_sse2+0x1d1>
+  DB  15,40,13,76,70,0,0                  ; movaps        0x464c(%rip),%xmm1        # 4d10 <_sk_callback_sse2+0x1d4>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,207                           ; mulps         %xmm7,%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -18154,7 +18151,7 @@ _sk_exclusion_sse2 LABEL PROC
 PUBLIC _sk_colorburn_sse2
 _sk_colorburn_sse2 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,59,70,0,0               ; movaps        0x463b(%rip),%xmm10        # 4d20 <_sk_callback_sse2+0x1e1>
+  DB  68,15,40,21,59,70,0,0               ; movaps        0x463b(%rip),%xmm10        # 4d20 <_sk_callback_sse2+0x1e4>
   DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
@@ -18246,7 +18243,7 @@ _sk_colorburn_sse2 LABEL PROC
 PUBLIC _sk_colordodge_sse2
 _sk_colordodge_sse2 LABEL PROC
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
-  DB  68,15,40,21,241,68,0,0              ; movaps        0x44f1(%rip),%xmm10        # 4d30 <_sk_callback_sse2+0x1f1>
+  DB  68,15,40,21,241,68,0,0              ; movaps        0x44f1(%rip),%xmm10        # 4d30 <_sk_callback_sse2+0x1f4>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -18339,7 +18336,7 @@ _sk_hardlight_sse2 LABEL PROC
   DB  15,41,52,36                         ; movaps        %xmm6,(%rsp)
   DB  15,40,245                           ; movaps        %xmm5,%xmm6
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
-  DB  68,15,40,29,163,67,0,0              ; movaps        0x43a3(%rip),%xmm11        # 4d40 <_sk_callback_sse2+0x201>
+  DB  68,15,40,29,163,67,0,0              ; movaps        0x43a3(%rip),%xmm11        # 4d40 <_sk_callback_sse2+0x204>
   DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
@@ -18426,7 +18423,7 @@ PUBLIC _sk_overlay_sse2
 _sk_overlay_sse2 LABEL PROC
   DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
   DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
-  DB  68,15,40,13,110,66,0,0              ; movaps        0x426e(%rip),%xmm9        # 4d50 <_sk_callback_sse2+0x211>
+  DB  68,15,40,13,110,66,0,0              ; movaps        0x426e(%rip),%xmm9        # 4d50 <_sk_callback_sse2+0x214>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -18516,7 +18513,7 @@ _sk_softlight_sse2 LABEL PROC
   DB  68,15,40,213                        ; movaps        %xmm5,%xmm10
   DB  68,15,94,215                        ; divps         %xmm7,%xmm10
   DB  69,15,84,212                        ; andps         %xmm12,%xmm10
-  DB  68,15,40,13,40,65,0,0               ; movaps        0x4128(%rip),%xmm9        # 4d60 <_sk_callback_sse2+0x221>
+  DB  68,15,40,13,40,65,0,0               ; movaps        0x4128(%rip),%xmm9        # 4d60 <_sk_callback_sse2+0x224>
   DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
   DB  69,15,92,250                        ; subps         %xmm10,%xmm15
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -18529,10 +18526,10 @@ _sk_softlight_sse2 LABEL PROC
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  65,15,88,194                        ; addps         %xmm10,%xmm0
-  DB  68,15,40,53,2,65,0,0                ; movaps        0x4102(%rip),%xmm14        # 4d70 <_sk_callback_sse2+0x231>
+  DB  68,15,40,53,2,65,0,0                ; movaps        0x4102(%rip),%xmm14        # 4d70 <_sk_callback_sse2+0x234>
   DB  69,15,88,222                        ; addps         %xmm14,%xmm11
   DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
-  DB  68,15,40,21,2,65,0,0                ; movaps        0x4102(%rip),%xmm10        # 4d80 <_sk_callback_sse2+0x241>
+  DB  68,15,40,21,2,65,0,0                ; movaps        0x4102(%rip),%xmm10        # 4d80 <_sk_callback_sse2+0x244>
   DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
   DB  69,15,88,235                        ; addps         %xmm11,%xmm13
   DB  15,88,228                           ; addps         %xmm4,%xmm4
@@ -18677,7 +18674,7 @@ _sk_hue_sse2 LABEL PROC
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  68,15,89,211                        ; mulps         %xmm3,%xmm10
-  DB  68,15,40,5,62,63,0,0                ; movaps        0x3f3e(%rip),%xmm8        # 4dc0 <_sk_callback_sse2+0x281>
+  DB  68,15,40,5,62,63,0,0                ; movaps        0x3f3e(%rip),%xmm8        # 4dc0 <_sk_callback_sse2+0x284>
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
   DB  15,40,207                           ; movaps        %xmm7,%xmm1
   DB  68,15,92,217                        ; subps         %xmm1,%xmm11
@@ -18723,12 +18720,12 @@ _sk_hue_sse2 LABEL PROC
   DB  69,15,84,206                        ; andps         %xmm14,%xmm9
   DB  69,15,84,214                        ; andps         %xmm14,%xmm10
   DB  65,15,84,214                        ; andps         %xmm14,%xmm2
-  DB  68,15,40,61,82,62,0,0               ; movaps        0x3e52(%rip),%xmm15        # 4d90 <_sk_callback_sse2+0x251>
+  DB  68,15,40,61,82,62,0,0               ; movaps        0x3e52(%rip),%xmm15        # 4d90 <_sk_callback_sse2+0x254>
   DB  65,15,89,231                        ; mulps         %xmm15,%xmm4
-  DB  15,40,5,87,62,0,0                   ; movaps        0x3e57(%rip),%xmm0        # 4da0 <_sk_callback_sse2+0x261>
+  DB  15,40,5,87,62,0,0                   ; movaps        0x3e57(%rip),%xmm0        # 4da0 <_sk_callback_sse2+0x264>
   DB  15,89,240                           ; mulps         %xmm0,%xmm6
   DB  15,88,244                           ; addps         %xmm4,%xmm6
-  DB  68,15,40,53,89,62,0,0               ; movaps        0x3e59(%rip),%xmm14        # 4db0 <_sk_callback_sse2+0x271>
+  DB  68,15,40,53,89,62,0,0               ; movaps        0x3e59(%rip),%xmm14        # 4db0 <_sk_callback_sse2+0x274>
   DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
   DB  69,15,89,238                        ; mulps         %xmm14,%xmm13
   DB  68,15,88,238                        ; addps         %xmm6,%xmm13
@@ -18905,14 +18902,14 @@ _sk_saturation_sse2 LABEL PROC
   DB  68,15,84,211                        ; andps         %xmm3,%xmm10
   DB  68,15,84,203                        ; andps         %xmm3,%xmm9
   DB  15,84,195                           ; andps         %xmm3,%xmm0
-  DB  68,15,40,5,233,59,0,0               ; movaps        0x3be9(%rip),%xmm8        # 4dd0 <_sk_callback_sse2+0x291>
+  DB  68,15,40,5,233,59,0,0               ; movaps        0x3be9(%rip),%xmm8        # 4dd0 <_sk_callback_sse2+0x294>
   DB  15,40,214                           ; movaps        %xmm6,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
-  DB  15,40,13,235,59,0,0                 ; movaps        0x3beb(%rip),%xmm1        # 4de0 <_sk_callback_sse2+0x2a1>
+  DB  15,40,13,235,59,0,0                 ; movaps        0x3beb(%rip),%xmm1        # 4de0 <_sk_callback_sse2+0x2a4>
   DB  15,40,221                           ; movaps        %xmm5,%xmm3
   DB  15,89,217                           ; mulps         %xmm1,%xmm3
   DB  15,88,218                           ; addps         %xmm2,%xmm3
-  DB  68,15,40,37,234,59,0,0              ; movaps        0x3bea(%rip),%xmm12        # 4df0 <_sk_callback_sse2+0x2b1>
+  DB  68,15,40,37,234,59,0,0              ; movaps        0x3bea(%rip),%xmm12        # 4df0 <_sk_callback_sse2+0x2b4>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
   DB  68,15,88,235                        ; addps         %xmm3,%xmm13
   DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
@@ -18957,7 +18954,7 @@ _sk_saturation_sse2 LABEL PROC
   DB  15,40,223                           ; movaps        %xmm7,%xmm3
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,89,221                           ; mulps         %xmm5,%xmm3
-  DB  68,15,40,5,79,59,0,0                ; movaps        0x3b4f(%rip),%xmm8        # 4e00 <_sk_callback_sse2+0x2c1>
+  DB  68,15,40,5,79,59,0,0                ; movaps        0x3b4f(%rip),%xmm8        # 4e00 <_sk_callback_sse2+0x2c4>
   DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  15,88,253                           ; addps         %xmm5,%xmm7
@@ -19058,14 +19055,14 @@ _sk_color_sse2 LABEL PROC
   DB  68,15,40,213                        ; movaps        %xmm5,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
   DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
-  DB  68,15,40,45,231,57,0,0              ; movaps        0x39e7(%rip),%xmm13        # 4e10 <_sk_callback_sse2+0x2d1>
+  DB  68,15,40,45,231,57,0,0              ; movaps        0x39e7(%rip),%xmm13        # 4e10 <_sk_callback_sse2+0x2d4>
   DB  68,15,40,198                        ; movaps        %xmm6,%xmm8
   DB  69,15,89,197                        ; mulps         %xmm13,%xmm8
-  DB  68,15,40,53,231,57,0,0              ; movaps        0x39e7(%rip),%xmm14        # 4e20 <_sk_callback_sse2+0x2e1>
+  DB  68,15,40,53,231,57,0,0              ; movaps        0x39e7(%rip),%xmm14        # 4e20 <_sk_callback_sse2+0x2e4>
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  65,15,89,198                        ; mulps         %xmm14,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,29,227,57,0,0              ; movaps        0x39e3(%rip),%xmm11        # 4e30 <_sk_callback_sse2+0x2f1>
+  DB  68,15,40,29,227,57,0,0              ; movaps        0x39e3(%rip),%xmm11        # 4e30 <_sk_callback_sse2+0x2f4>
   DB  69,15,89,227                        ; mulps         %xmm11,%xmm12
   DB  68,15,88,224                        ; addps         %xmm0,%xmm12
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
@@ -19073,7 +19070,7 @@ _sk_color_sse2 LABEL PROC
   DB  69,15,40,250                        ; movaps        %xmm10,%xmm15
   DB  69,15,89,254                        ; mulps         %xmm14,%xmm15
   DB  68,15,88,248                        ; addps         %xmm0,%xmm15
-  DB  68,15,40,5,207,57,0,0               ; movaps        0x39cf(%rip),%xmm8        # 4e40 <_sk_callback_sse2+0x301>
+  DB  68,15,40,5,207,57,0,0               ; movaps        0x39cf(%rip),%xmm8        # 4e40 <_sk_callback_sse2+0x304>
   DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
   DB  15,92,226                           ; subps         %xmm2,%xmm4
   DB  15,89,252                           ; mulps         %xmm4,%xmm7
@@ -19209,15 +19206,15 @@ _sk_luminosity_sse2 LABEL PROC
   DB  68,15,40,205                        ; movaps        %xmm5,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
   DB  15,89,222                           ; mulps         %xmm6,%xmm3
-  DB  68,15,40,37,225,55,0,0              ; movaps        0x37e1(%rip),%xmm12        # 4e50 <_sk_callback_sse2+0x311>
+  DB  68,15,40,37,225,55,0,0              ; movaps        0x37e1(%rip),%xmm12        # 4e50 <_sk_callback_sse2+0x314>
   DB  68,15,40,199                        ; movaps        %xmm7,%xmm8
   DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
-  DB  68,15,40,45,225,55,0,0              ; movaps        0x37e1(%rip),%xmm13        # 4e60 <_sk_callback_sse2+0x321>
+  DB  68,15,40,45,225,55,0,0              ; movaps        0x37e1(%rip),%xmm13        # 4e60 <_sk_callback_sse2+0x324>
   DB  68,15,40,241                        ; movaps        %xmm1,%xmm14
   DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
   DB  69,15,88,240                        ; addps         %xmm8,%xmm14
-  DB  68,15,40,29,221,55,0,0              ; movaps        0x37dd(%rip),%xmm11        # 4e70 <_sk_callback_sse2+0x331>
-  DB  68,15,40,5,229,55,0,0               ; movaps        0x37e5(%rip),%xmm8        # 4e80 <_sk_callback_sse2+0x341>
+  DB  68,15,40,29,221,55,0,0              ; movaps        0x37dd(%rip),%xmm11        # 4e70 <_sk_callback_sse2+0x334>
+  DB  68,15,40,5,229,55,0,0               ; movaps        0x37e5(%rip),%xmm8        # 4e80 <_sk_callback_sse2+0x344>
   DB  69,15,40,248                        ; movaps        %xmm8,%xmm15
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  68,15,92,248                        ; subps         %xmm0,%xmm15
@@ -19359,7 +19356,7 @@ _sk_clamp_0_sse2 LABEL PROC
 
 PUBLIC _sk_clamp_1_sse2
 _sk_clamp_1_sse2 LABEL PROC
-  DB  68,15,40,5,236,53,0,0               ; movaps        0x35ec(%rip),%xmm8        # 4e90 <_sk_callback_sse2+0x351>
+  DB  68,15,40,5,236,53,0,0               ; movaps        0x35ec(%rip),%xmm8        # 4e90 <_sk_callback_sse2+0x354>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
   DB  65,15,93,208                        ; minps         %xmm8,%xmm2
@@ -19369,7 +19366,7 @@ _sk_clamp_1_sse2 LABEL PROC
 
 PUBLIC _sk_clamp_a_sse2
 _sk_clamp_a_sse2 LABEL PROC
-  DB  15,93,29,225,53,0,0                 ; minps         0x35e1(%rip),%xmm3        # 4ea0 <_sk_callback_sse2+0x361>
+  DB  15,93,29,225,53,0,0                 ; minps         0x35e1(%rip),%xmm3        # 4ea0 <_sk_callback_sse2+0x364>
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -19442,7 +19439,7 @@ _sk_premul_sse2 LABEL PROC
 PUBLIC _sk_unpremul_sse2
 _sk_unpremul_sse2 LABEL PROC
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
-  DB  68,15,40,13,76,53,0,0               ; movaps        0x354c(%rip),%xmm9        # 4eb0 <_sk_callback_sse2+0x371>
+  DB  68,15,40,13,76,53,0,0               ; movaps        0x354c(%rip),%xmm9        # 4eb0 <_sk_callback_sse2+0x374>
   DB  68,15,94,203                        ; divps         %xmm3,%xmm9
   DB  68,15,194,195,4                     ; cmpneqps      %xmm3,%xmm8
   DB  69,15,84,193                        ; andps         %xmm9,%xmm8
@@ -19454,20 +19451,20 @@ _sk_unpremul_sse2 LABEL PROC
 
 PUBLIC _sk_from_srgb_sse2
 _sk_from_srgb_sse2 LABEL PROC
-  DB  68,15,40,5,55,53,0,0                ; movaps        0x3537(%rip),%xmm8        # 4ec0 <_sk_callback_sse2+0x381>
+  DB  68,15,40,5,55,53,0,0                ; movaps        0x3537(%rip),%xmm8        # 4ec0 <_sk_callback_sse2+0x384>
   DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
   DB  69,15,89,232                        ; mulps         %xmm8,%xmm13
   DB  68,15,40,216                        ; movaps        %xmm0,%xmm11
   DB  69,15,89,219                        ; mulps         %xmm11,%xmm11
-  DB  68,15,40,13,47,53,0,0               ; movaps        0x352f(%rip),%xmm9        # 4ed0 <_sk_callback_sse2+0x391>
+  DB  68,15,40,13,47,53,0,0               ; movaps        0x352f(%rip),%xmm9        # 4ed0 <_sk_callback_sse2+0x394>
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
   DB  69,15,89,241                        ; mulps         %xmm9,%xmm14
-  DB  68,15,40,21,47,53,0,0               ; movaps        0x352f(%rip),%xmm10        # 4ee0 <_sk_callback_sse2+0x3a1>
+  DB  68,15,40,21,47,53,0,0               ; movaps        0x352f(%rip),%xmm10        # 4ee0 <_sk_callback_sse2+0x3a4>
   DB  69,15,88,242                        ; addps         %xmm10,%xmm14
   DB  69,15,89,243                        ; mulps         %xmm11,%xmm14
-  DB  68,15,40,29,47,53,0,0               ; movaps        0x352f(%rip),%xmm11        # 4ef0 <_sk_callback_sse2+0x3b1>
+  DB  68,15,40,29,47,53,0,0               ; movaps        0x352f(%rip),%xmm11        # 4ef0 <_sk_callback_sse2+0x3b4>
   DB  69,15,88,243                        ; addps         %xmm11,%xmm14
-  DB  68,15,40,37,51,53,0,0               ; movaps        0x3533(%rip),%xmm12        # 4f00 <_sk_callback_sse2+0x3c1>
+  DB  68,15,40,37,51,53,0,0               ; movaps        0x3533(%rip),%xmm12        # 4f00 <_sk_callback_sse2+0x3c4>
   DB  65,15,194,196,1                     ; cmpltps       %xmm12,%xmm0
   DB  68,15,84,232                        ; andps         %xmm0,%xmm13
   DB  65,15,85,198                        ; andnps        %xmm14,%xmm0
@@ -19504,20 +19501,20 @@ _sk_to_srgb_sse2 LABEL PROC
   DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
   DB  69,15,83,200                        ; rcpps         %xmm8,%xmm9
   DB  69,15,82,232                        ; rsqrtps       %xmm8,%xmm13
-  DB  68,15,40,5,184,52,0,0               ; movaps        0x34b8(%rip),%xmm8        # 4f10 <_sk_callback_sse2+0x3d1>
+  DB  68,15,40,5,184,52,0,0               ; movaps        0x34b8(%rip),%xmm8        # 4f10 <_sk_callback_sse2+0x3d4>
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
   DB  69,15,89,240                        ; mulps         %xmm8,%xmm14
-  DB  68,15,40,21,184,52,0,0              ; movaps        0x34b8(%rip),%xmm10        # 4f20 <_sk_callback_sse2+0x3e1>
+  DB  68,15,40,21,184,52,0,0              ; movaps        0x34b8(%rip),%xmm10        # 4f20 <_sk_callback_sse2+0x3e4>
   DB  69,15,89,202                        ; mulps         %xmm10,%xmm9
-  DB  68,15,40,29,188,52,0,0              ; movaps        0x34bc(%rip),%xmm11        # 4f30 <_sk_callback_sse2+0x3f1>
+  DB  68,15,40,29,188,52,0,0              ; movaps        0x34bc(%rip),%xmm11        # 4f30 <_sk_callback_sse2+0x3f4>
   DB  69,15,88,203                        ; addps         %xmm11,%xmm9
-  DB  68,15,40,37,192,52,0,0              ; movaps        0x34c0(%rip),%xmm12        # 4f40 <_sk_callback_sse2+0x401>
+  DB  68,15,40,37,192,52,0,0              ; movaps        0x34c0(%rip),%xmm12        # 4f40 <_sk_callback_sse2+0x404>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,40,13,192,52,0,0              ; movaps        0x34c0(%rip),%xmm9        # 4f50 <_sk_callback_sse2+0x411>
+  DB  68,15,40,13,192,52,0,0              ; movaps        0x34c0(%rip),%xmm9        # 4f50 <_sk_callback_sse2+0x414>
   DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
   DB  69,15,93,253                        ; minps         %xmm13,%xmm15
-  DB  68,15,40,45,192,52,0,0              ; movaps        0x34c0(%rip),%xmm13        # 4f60 <_sk_callback_sse2+0x421>
+  DB  68,15,40,45,192,52,0,0              ; movaps        0x34c0(%rip),%xmm13        # 4f60 <_sk_callback_sse2+0x424>
   DB  65,15,194,197,1                     ; cmpltps       %xmm13,%xmm0
   DB  68,15,84,240                        ; andps         %xmm0,%xmm14
   DB  65,15,85,199                        ; andnps        %xmm15,%xmm0
@@ -19565,7 +19562,7 @@ _sk_rgb_to_hsl_sse2 LABEL PROC
   DB  68,15,93,218                        ; minps         %xmm2,%xmm11
   DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
   DB  65,15,92,203                        ; subps         %xmm11,%xmm1
-  DB  68,15,40,45,25,52,0,0               ; movaps        0x3419(%rip),%xmm13        # 4f70 <_sk_callback_sse2+0x431>
+  DB  68,15,40,45,25,52,0,0               ; movaps        0x3419(%rip),%xmm13        # 4f70 <_sk_callback_sse2+0x434>
   DB  68,15,94,233                        ; divps         %xmm1,%xmm13
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  65,15,194,192,0                     ; cmpeqps       %xmm8,%xmm0
@@ -19574,30 +19571,30 @@ _sk_rgb_to_hsl_sse2 LABEL PROC
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
   DB  68,15,194,242,1                     ; cmpltps       %xmm2,%xmm14
-  DB  68,15,84,53,255,51,0,0              ; andps         0x33ff(%rip),%xmm14        # 4f80 <_sk_callback_sse2+0x441>
+  DB  68,15,84,53,255,51,0,0              ; andps         0x33ff(%rip),%xmm14        # 4f80 <_sk_callback_sse2+0x444>
   DB  69,15,88,244                        ; addps         %xmm12,%xmm14
   DB  69,15,40,250                        ; movaps        %xmm10,%xmm15
   DB  69,15,194,249,0                     ; cmpeqps       %xmm9,%xmm15
   DB  65,15,92,208                        ; subps         %xmm8,%xmm2
   DB  65,15,89,213                        ; mulps         %xmm13,%xmm2
-  DB  68,15,40,37,242,51,0,0              ; movaps        0x33f2(%rip),%xmm12        # 4f90 <_sk_callback_sse2+0x451>
+  DB  68,15,40,37,242,51,0,0              ; movaps        0x33f2(%rip),%xmm12        # 4f90 <_sk_callback_sse2+0x454>
   DB  65,15,88,212                        ; addps         %xmm12,%xmm2
   DB  69,15,92,193                        ; subps         %xmm9,%xmm8
   DB  69,15,89,197                        ; mulps         %xmm13,%xmm8
-  DB  68,15,88,5,238,51,0,0               ; addps         0x33ee(%rip),%xmm8        # 4fa0 <_sk_callback_sse2+0x461>
+  DB  68,15,88,5,238,51,0,0               ; addps         0x33ee(%rip),%xmm8        # 4fa0 <_sk_callback_sse2+0x464>
   DB  65,15,84,215                        ; andps         %xmm15,%xmm2
   DB  69,15,85,248                        ; andnps        %xmm8,%xmm15
   DB  68,15,86,250                        ; orps          %xmm2,%xmm15
   DB  68,15,84,240                        ; andps         %xmm0,%xmm14
   DB  65,15,85,199                        ; andnps        %xmm15,%xmm0
   DB  65,15,86,198                        ; orps          %xmm14,%xmm0
-  DB  15,89,5,223,51,0,0                  ; mulps         0x33df(%rip),%xmm0        # 4fb0 <_sk_callback_sse2+0x471>
+  DB  15,89,5,223,51,0,0                  ; mulps         0x33df(%rip),%xmm0        # 4fb0 <_sk_callback_sse2+0x474>
   DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
   DB  69,15,194,195,4                     ; cmpneqps      %xmm11,%xmm8
   DB  65,15,84,192                        ; andps         %xmm8,%xmm0
   DB  69,15,92,226                        ; subps         %xmm10,%xmm12
   DB  69,15,88,211                        ; addps         %xmm11,%xmm10
-  DB  68,15,40,13,210,51,0,0              ; movaps        0x33d2(%rip),%xmm9        # 4fc0 <_sk_callback_sse2+0x481>
+  DB  68,15,40,13,210,51,0,0              ; movaps        0x33d2(%rip),%xmm9        # 4fc0 <_sk_callback_sse2+0x484>
   DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  68,15,194,202,1                     ; cmpltps       %xmm2,%xmm9
@@ -19620,7 +19617,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC
   DB  15,41,92,36,32                      ; movaps        %xmm3,0x20(%rsp)
   DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
-  DB  68,15,40,13,141,51,0,0              ; movaps        0x338d(%rip),%xmm9        # 4fd0 <_sk_callback_sse2+0x491>
+  DB  68,15,40,13,141,51,0,0              ; movaps        0x338d(%rip),%xmm9        # 4fd0 <_sk_callback_sse2+0x494>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  69,15,194,211,2                     ; cmpleps       %xmm11,%xmm10
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
@@ -19637,28 +19634,28 @@ _sk_hsl_to_rgb_sse2 LABEL PROC
   DB  69,15,88,211                        ; addps         %xmm11,%xmm10
   DB  69,15,88,219                        ; addps         %xmm11,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  15,40,5,87,51,0,0                   ; movaps        0x3357(%rip),%xmm0        # 4fe0 <_sk_callback_sse2+0x4a1>
+  DB  15,40,5,87,51,0,0                   ; movaps        0x3357(%rip),%xmm0        # 4fe0 <_sk_callback_sse2+0x4a4>
   DB  15,88,198                           ; addps         %xmm6,%xmm0
   DB  243,15,91,200                       ; cvttps2dq     %xmm0,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
   DB  15,194,217,1                        ; cmpltps       %xmm1,%xmm3
-  DB  15,84,29,79,51,0,0                  ; andps         0x334f(%rip),%xmm3        # 4ff0 <_sk_callback_sse2+0x4b1>
+  DB  15,84,29,79,51,0,0                  ; andps         0x334f(%rip),%xmm3        # 4ff0 <_sk_callback_sse2+0x4b4>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,92,193                           ; subps         %xmm1,%xmm0
-  DB  68,15,40,45,81,51,0,0               ; movaps        0x3351(%rip),%xmm13        # 5000 <_sk_callback_sse2+0x4c1>
+  DB  68,15,40,45,81,51,0,0               ; movaps        0x3351(%rip),%xmm13        # 5000 <_sk_callback_sse2+0x4c4>
   DB  69,15,40,197                        ; movaps        %xmm13,%xmm8
   DB  68,15,194,192,2                     ; cmpleps       %xmm0,%xmm8
   DB  69,15,40,242                        ; movaps        %xmm10,%xmm14
   DB  69,15,92,243                        ; subps         %xmm11,%xmm14
   DB  65,15,40,217                        ; movaps        %xmm9,%xmm3
   DB  15,194,216,2                        ; cmpleps       %xmm0,%xmm3
-  DB  15,40,21,97,51,0,0                  ; movaps        0x3361(%rip),%xmm2        # 5030 <_sk_callback_sse2+0x4f1>
+  DB  15,40,21,97,51,0,0                  ; movaps        0x3361(%rip),%xmm2        # 5030 <_sk_callback_sse2+0x4f4>
   DB  68,15,40,250                        ; movaps        %xmm2,%xmm15
   DB  68,15,194,248,2                     ; cmpleps       %xmm0,%xmm15
-  DB  15,40,13,49,51,0,0                  ; movaps        0x3331(%rip),%xmm1        # 5010 <_sk_callback_sse2+0x4d1>
+  DB  15,40,13,49,51,0,0                  ; movaps        0x3331(%rip),%xmm1        # 5010 <_sk_callback_sse2+0x4d4>
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
-  DB  15,40,45,55,51,0,0                  ; movaps        0x3337(%rip),%xmm5        # 5020 <_sk_callback_sse2+0x4e1>
+  DB  15,40,45,55,51,0,0                  ; movaps        0x3337(%rip),%xmm5        # 5020 <_sk_callback_sse2+0x4e4>
   DB  15,40,229                           ; movaps        %xmm5,%xmm4
   DB  15,92,224                           ; subps         %xmm0,%xmm4
   DB  65,15,89,230                        ; mulps         %xmm14,%xmm4
@@ -19681,7 +19678,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  15,40,222                           ; movaps        %xmm6,%xmm3
   DB  15,194,216,1                        ; cmpltps       %xmm0,%xmm3
-  DB  15,84,29,172,50,0,0                 ; andps         0x32ac(%rip),%xmm3        # 4ff0 <_sk_callback_sse2+0x4b1>
+  DB  15,84,29,172,50,0,0                 ; andps         0x32ac(%rip),%xmm3        # 4ff0 <_sk_callback_sse2+0x4b4>
   DB  15,92,195                           ; subps         %xmm3,%xmm0
   DB  68,15,40,230                        ; movaps        %xmm6,%xmm12
   DB  68,15,92,224                        ; subps         %xmm0,%xmm12
@@ -19711,12 +19708,12 @@ _sk_hsl_to_rgb_sse2 LABEL PROC
   DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
   DB  15,40,231                           ; movaps        %xmm7,%xmm4
   DB  15,85,227                           ; andnps        %xmm3,%xmm4
-  DB  15,88,53,133,50,0,0                 ; addps         0x3285(%rip),%xmm6        # 5040 <_sk_callback_sse2+0x501>
+  DB  15,88,53,133,50,0,0                 ; addps         0x3285(%rip),%xmm6        # 5040 <_sk_callback_sse2+0x504>
   DB  243,15,91,198                       ; cvttps2dq     %xmm6,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  15,40,222                           ; movaps        %xmm6,%xmm3
   DB  15,194,216,1                        ; cmpltps       %xmm0,%xmm3
-  DB  15,84,29,32,50,0,0                  ; andps         0x3220(%rip),%xmm3        # 4ff0 <_sk_callback_sse2+0x4b1>
+  DB  15,84,29,32,50,0,0                  ; andps         0x3220(%rip),%xmm3        # 4ff0 <_sk_callback_sse2+0x4b4>
   DB  15,92,195                           ; subps         %xmm3,%xmm0
   DB  15,92,240                           ; subps         %xmm0,%xmm6
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
@@ -19777,7 +19774,7 @@ _sk_scale_u8_sse2 LABEL PROC
   DB  102,69,15,96,193                    ; punpcklbw     %xmm9,%xmm8
   DB  102,69,15,97,193                    ; punpcklwd     %xmm9,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,170,49,0,0               ; mulps         0x31aa(%rip),%xmm8        # 5050 <_sk_callback_sse2+0x511>
+  DB  68,15,89,5,170,49,0,0               ; mulps         0x31aa(%rip),%xmm8        # 5050 <_sk_callback_sse2+0x514>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
@@ -19814,7 +19811,7 @@ _sk_lerp_u8_sse2 LABEL PROC
   DB  102,69,15,96,193                    ; punpcklbw     %xmm9,%xmm8
   DB  102,69,15,97,193                    ; punpcklwd     %xmm9,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,72,49,0,0                ; mulps         0x3148(%rip),%xmm8        # 5060 <_sk_callback_sse2+0x521>
+  DB  68,15,89,5,72,49,0,0                ; mulps         0x3148(%rip),%xmm8        # 5060 <_sk_callback_sse2+0x524>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -19837,17 +19834,17 @@ _sk_lerp_565_sse2 LABEL PROC
   DB  243,68,15,126,20,120                ; movq          (%rax,%rdi,2),%xmm10
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,69,15,97,208                    ; punpcklwd     %xmm8,%xmm10
-  DB  102,68,15,111,5,14,49,0,0           ; movdqa        0x310e(%rip),%xmm8        # 5070 <_sk_callback_sse2+0x531>
+  DB  102,68,15,111,5,14,49,0,0           ; movdqa        0x310e(%rip),%xmm8        # 5070 <_sk_callback_sse2+0x534>
   DB  102,69,15,219,194                   ; pand          %xmm10,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,13,49,0,0                ; mulps         0x310d(%rip),%xmm8        # 5080 <_sk_callback_sse2+0x541>
-  DB  102,68,15,111,13,20,49,0,0          ; movdqa        0x3114(%rip),%xmm9        # 5090 <_sk_callback_sse2+0x551>
+  DB  68,15,89,5,13,49,0,0                ; mulps         0x310d(%rip),%xmm8        # 5080 <_sk_callback_sse2+0x544>
+  DB  102,68,15,111,13,20,49,0,0          ; movdqa        0x3114(%rip),%xmm9        # 5090 <_sk_callback_sse2+0x554>
   DB  102,69,15,219,202                   ; pand          %xmm10,%xmm9
   DB  69,15,91,201                        ; cvtdq2ps      %xmm9,%xmm9
-  DB  68,15,89,13,19,49,0,0               ; mulps         0x3113(%rip),%xmm9        # 50a0 <_sk_callback_sse2+0x561>
-  DB  102,68,15,219,21,26,49,0,0          ; pand          0x311a(%rip),%xmm10        # 50b0 <_sk_callback_sse2+0x571>
+  DB  68,15,89,13,19,49,0,0               ; mulps         0x3113(%rip),%xmm9        # 50a0 <_sk_callback_sse2+0x564>
+  DB  102,68,15,219,21,26,49,0,0          ; pand          0x311a(%rip),%xmm10        # 50b0 <_sk_callback_sse2+0x574>
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
-  DB  68,15,89,21,30,49,0,0               ; mulps         0x311e(%rip),%xmm10        # 50c0 <_sk_callback_sse2+0x581>
+  DB  68,15,89,21,30,49,0,0               ; mulps         0x311e(%rip),%xmm10        # 50c0 <_sk_callback_sse2+0x584>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -19876,7 +19873,7 @@ _sk_load_tables_sse2 LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  76,139,72,8                         ; mov           0x8(%rax),%r9
   DB  243,69,15,111,12,184                ; movdqu        (%r8,%rdi,4),%xmm9
-  DB  102,68,15,111,5,206,48,0,0          ; movdqa        0x30ce(%rip),%xmm8        # 50d0 <_sk_callback_sse2+0x591>
+  DB  102,68,15,111,5,206,48,0,0          ; movdqa        0x30ce(%rip),%xmm8        # 50d0 <_sk_callback_sse2+0x594>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
@@ -19931,7 +19928,7 @@ _sk_load_tables_sse2 LABEL PROC
   DB  65,15,20,208                        ; unpcklps      %xmm8,%xmm2
   DB  102,65,15,114,209,24                ; psrld         $0x18,%xmm9
   DB  65,15,91,217                        ; cvtdq2ps      %xmm9,%xmm3
-  DB  15,89,29,219,47,0,0                 ; mulps         0x2fdb(%rip),%xmm3        # 50e0 <_sk_callback_sse2+0x5a1>
+  DB  15,89,29,219,47,0,0                 ; mulps         0x2fdb(%rip),%xmm3        # 50e0 <_sk_callback_sse2+0x5a4>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -19948,7 +19945,7 @@ _sk_load_tables_u16_be_sse2 LABEL PROC
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
   DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
-  DB  102,68,15,111,21,174,47,0,0         ; movdqa        0x2fae(%rip),%xmm10        # 50f0 <_sk_callback_sse2+0x5b1>
+  DB  102,68,15,111,21,174,47,0,0         ; movdqa        0x2fae(%rip),%xmm10        # 50f0 <_sk_callback_sse2+0x5b4>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,194                   ; pand          %xmm10,%xmm0
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
@@ -20009,7 +20006,7 @@ _sk_load_tables_u16_be_sse2 LABEL PROC
   DB  102,65,15,235,217                   ; por           %xmm9,%xmm3
   DB  102,65,15,97,216                    ; punpcklwd     %xmm8,%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,157,46,0,0                 ; mulps         0x2e9d(%rip),%xmm3        # 5100 <_sk_callback_sse2+0x5c1>
+  DB  15,89,29,157,46,0,0                 ; mulps         0x2e9d(%rip),%xmm3        # 5100 <_sk_callback_sse2+0x5c4>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -20029,7 +20026,7 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC
   DB  102,68,15,97,208                    ; punpcklwd     %xmm0,%xmm10
   DB  102,65,15,111,195                   ; movdqa        %xmm11,%xmm0
   DB  102,65,15,97,194                    ; punpcklwd     %xmm10,%xmm0
-  DB  102,68,15,111,5,93,46,0,0           ; movdqa        0x2e5d(%rip),%xmm8        # 5110 <_sk_callback_sse2+0x5d1>
+  DB  102,68,15,111,5,93,46,0,0           ; movdqa        0x2e5d(%rip),%xmm8        # 5110 <_sk_callback_sse2+0x5d4>
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
@@ -20084,7 +20081,7 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC
   DB  15,20,211                           ; unpcklps      %xmm3,%xmm2
   DB  65,15,20,208                        ; unpcklps      %xmm8,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,108,45,0,0                 ; movaps        0x2d6c(%rip),%xmm3        # 5120 <_sk_callback_sse2+0x5e1>
+  DB  15,40,29,108,45,0,0                 ; movaps        0x2d6c(%rip),%xmm3        # 5120 <_sk_callback_sse2+0x5e4>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_byte_tables_sse2
@@ -20092,7 +20089,7 @@ _sk_byte_tables_sse2 LABEL PROC
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,109,45,0,0               ; movaps        0x2d6d(%rip),%xmm8        # 5130 <_sk_callback_sse2+0x5f1>
+  DB  68,15,40,5,109,45,0,0               ; movaps        0x2d6d(%rip),%xmm8        # 5130 <_sk_callback_sse2+0x5f4>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
@@ -20119,7 +20116,7 @@ _sk_byte_tables_sse2 LABEL PROC
   DB  102,65,15,96,193                    ; punpcklbw     %xmm9,%xmm0
   DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,21,10,45,0,0               ; movaps        0x2d0a(%rip),%xmm10        # 5140 <_sk_callback_sse2+0x601>
+  DB  68,15,40,21,10,45,0,0               ; movaps        0x2d0a(%rip),%xmm10        # 5140 <_sk_callback_sse2+0x604>
   DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -20233,7 +20230,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC
   DB  102,65,15,96,193                    ; punpcklbw     %xmm9,%xmm0
   DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,21,93,43,0,0               ; movaps        0x2b5d(%rip),%xmm10        # 5150 <_sk_callback_sse2+0x611>
+  DB  68,15,40,21,93,43,0,0               ; movaps        0x2b5d(%rip),%xmm10        # 5150 <_sk_callback_sse2+0x614>
   DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -20420,15 +20417,15 @@ _sk_parametric_r_sse2 LABEL PROC
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,156,40,0,0              ; mulps         0x289c(%rip),%xmm9        # 5160 <_sk_callback_sse2+0x621>
-  DB  68,15,84,21,164,40,0,0              ; andps         0x28a4(%rip),%xmm10        # 5170 <_sk_callback_sse2+0x631>
-  DB  68,15,86,21,172,40,0,0              ; orps          0x28ac(%rip),%xmm10        # 5180 <_sk_callback_sse2+0x641>
-  DB  68,15,88,13,180,40,0,0              ; addps         0x28b4(%rip),%xmm9        # 5190 <_sk_callback_sse2+0x651>
-  DB  68,15,40,37,188,40,0,0              ; movaps        0x28bc(%rip),%xmm12        # 51a0 <_sk_callback_sse2+0x661>
+  DB  68,15,89,13,156,40,0,0              ; mulps         0x289c(%rip),%xmm9        # 5160 <_sk_callback_sse2+0x624>
+  DB  68,15,84,21,164,40,0,0              ; andps         0x28a4(%rip),%xmm10        # 5170 <_sk_callback_sse2+0x634>
+  DB  68,15,86,21,172,40,0,0              ; orps          0x28ac(%rip),%xmm10        # 5180 <_sk_callback_sse2+0x644>
+  DB  68,15,88,13,180,40,0,0              ; addps         0x28b4(%rip),%xmm9        # 5190 <_sk_callback_sse2+0x654>
+  DB  68,15,40,37,188,40,0,0              ; movaps        0x28bc(%rip),%xmm12        # 51a0 <_sk_callback_sse2+0x664>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,188,40,0,0              ; addps         0x28bc(%rip),%xmm10        # 51b0 <_sk_callback_sse2+0x671>
-  DB  68,15,40,37,196,40,0,0              ; movaps        0x28c4(%rip),%xmm12        # 51c0 <_sk_callback_sse2+0x681>
+  DB  68,15,88,21,188,40,0,0              ; addps         0x28bc(%rip),%xmm10        # 51b0 <_sk_callback_sse2+0x674>
+  DB  68,15,40,37,196,40,0,0              ; movaps        0x28c4(%rip),%xmm12        # 51c0 <_sk_callback_sse2+0x684>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20436,22 +20433,22 @@ _sk_parametric_r_sse2 LABEL PROC
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,174,40,0,0              ; movaps        0x28ae(%rip),%xmm10        # 51d0 <_sk_callback_sse2+0x691>
+  DB  68,15,40,21,174,40,0,0              ; movaps        0x28ae(%rip),%xmm10        # 51d0 <_sk_callback_sse2+0x694>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,162,40,0,0              ; addps         0x28a2(%rip),%xmm9        # 51e0 <_sk_callback_sse2+0x6a1>
-  DB  68,15,40,37,170,40,0,0              ; movaps        0x28aa(%rip),%xmm12        # 51f0 <_sk_callback_sse2+0x6b1>
+  DB  68,15,88,13,162,40,0,0              ; addps         0x28a2(%rip),%xmm9        # 51e0 <_sk_callback_sse2+0x6a4>
+  DB  68,15,40,37,170,40,0,0              ; movaps        0x28aa(%rip),%xmm12        # 51f0 <_sk_callback_sse2+0x6b4>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,170,40,0,0              ; movaps        0x28aa(%rip),%xmm12        # 5200 <_sk_callback_sse2+0x6c1>
+  DB  68,15,40,37,170,40,0,0              ; movaps        0x28aa(%rip),%xmm12        # 5200 <_sk_callback_sse2+0x6c4>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,174,40,0,0              ; movaps        0x28ae(%rip),%xmm13        # 5210 <_sk_callback_sse2+0x6d1>
+  DB  68,15,40,45,174,40,0,0              ; movaps        0x28ae(%rip),%xmm13        # 5210 <_sk_callback_sse2+0x6d4>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,174,40,0,0              ; mulps         0x28ae(%rip),%xmm13        # 5220 <_sk_callback_sse2+0x6e1>
+  DB  68,15,89,45,174,40,0,0              ; mulps         0x28ae(%rip),%xmm13        # 5220 <_sk_callback_sse2+0x6e4>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -20485,15 +20482,15 @@ _sk_parametric_g_sse2 LABEL PROC
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,46,40,0,0               ; mulps         0x282e(%rip),%xmm9        # 5230 <_sk_callback_sse2+0x6f1>
-  DB  68,15,84,21,54,40,0,0               ; andps         0x2836(%rip),%xmm10        # 5240 <_sk_callback_sse2+0x701>
-  DB  68,15,86,21,62,40,0,0               ; orps          0x283e(%rip),%xmm10        # 5250 <_sk_callback_sse2+0x711>
-  DB  68,15,88,13,70,40,0,0               ; addps         0x2846(%rip),%xmm9        # 5260 <_sk_callback_sse2+0x721>
-  DB  68,15,40,37,78,40,0,0               ; movaps        0x284e(%rip),%xmm12        # 5270 <_sk_callback_sse2+0x731>
+  DB  68,15,89,13,46,40,0,0               ; mulps         0x282e(%rip),%xmm9        # 5230 <_sk_callback_sse2+0x6f4>
+  DB  68,15,84,21,54,40,0,0               ; andps         0x2836(%rip),%xmm10        # 5240 <_sk_callback_sse2+0x704>
+  DB  68,15,86,21,62,40,0,0               ; orps          0x283e(%rip),%xmm10        # 5250 <_sk_callback_sse2+0x714>
+  DB  68,15,88,13,70,40,0,0               ; addps         0x2846(%rip),%xmm9        # 5260 <_sk_callback_sse2+0x724>
+  DB  68,15,40,37,78,40,0,0               ; movaps        0x284e(%rip),%xmm12        # 5270 <_sk_callback_sse2+0x734>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,78,40,0,0               ; addps         0x284e(%rip),%xmm10        # 5280 <_sk_callback_sse2+0x741>
-  DB  68,15,40,37,86,40,0,0               ; movaps        0x2856(%rip),%xmm12        # 5290 <_sk_callback_sse2+0x751>
+  DB  68,15,88,21,78,40,0,0               ; addps         0x284e(%rip),%xmm10        # 5280 <_sk_callback_sse2+0x744>
+  DB  68,15,40,37,86,40,0,0               ; movaps        0x2856(%rip),%xmm12        # 5290 <_sk_callback_sse2+0x754>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20501,22 +20498,22 @@ _sk_parametric_g_sse2 LABEL PROC
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,64,40,0,0               ; movaps        0x2840(%rip),%xmm10        # 52a0 <_sk_callback_sse2+0x761>
+  DB  68,15,40,21,64,40,0,0               ; movaps        0x2840(%rip),%xmm10        # 52a0 <_sk_callback_sse2+0x764>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,52,40,0,0               ; addps         0x2834(%rip),%xmm9        # 52b0 <_sk_callback_sse2+0x771>
-  DB  68,15,40,37,60,40,0,0               ; movaps        0x283c(%rip),%xmm12        # 52c0 <_sk_callback_sse2+0x781>
+  DB  68,15,88,13,52,40,0,0               ; addps         0x2834(%rip),%xmm9        # 52b0 <_sk_callback_sse2+0x774>
+  DB  68,15,40,37,60,40,0,0               ; movaps        0x283c(%rip),%xmm12        # 52c0 <_sk_callback_sse2+0x784>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,60,40,0,0               ; movaps        0x283c(%rip),%xmm12        # 52d0 <_sk_callback_sse2+0x791>
+  DB  68,15,40,37,60,40,0,0               ; movaps        0x283c(%rip),%xmm12        # 52d0 <_sk_callback_sse2+0x794>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,64,40,0,0               ; movaps        0x2840(%rip),%xmm13        # 52e0 <_sk_callback_sse2+0x7a1>
+  DB  68,15,40,45,64,40,0,0               ; movaps        0x2840(%rip),%xmm13        # 52e0 <_sk_callback_sse2+0x7a4>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,64,40,0,0               ; mulps         0x2840(%rip),%xmm13        # 52f0 <_sk_callback_sse2+0x7b1>
+  DB  68,15,89,45,64,40,0,0               ; mulps         0x2840(%rip),%xmm13        # 52f0 <_sk_callback_sse2+0x7b4>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -20550,15 +20547,15 @@ _sk_parametric_b_sse2 LABEL PROC
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,192,39,0,0              ; mulps         0x27c0(%rip),%xmm9        # 5300 <_sk_callback_sse2+0x7c1>
-  DB  68,15,84,21,200,39,0,0              ; andps         0x27c8(%rip),%xmm10        # 5310 <_sk_callback_sse2+0x7d1>
-  DB  68,15,86,21,208,39,0,0              ; orps          0x27d0(%rip),%xmm10        # 5320 <_sk_callback_sse2+0x7e1>
-  DB  68,15,88,13,216,39,0,0              ; addps         0x27d8(%rip),%xmm9        # 5330 <_sk_callback_sse2+0x7f1>
-  DB  68,15,40,37,224,39,0,0              ; movaps        0x27e0(%rip),%xmm12        # 5340 <_sk_callback_sse2+0x801>
+  DB  68,15,89,13,192,39,0,0              ; mulps         0x27c0(%rip),%xmm9        # 5300 <_sk_callback_sse2+0x7c4>
+  DB  68,15,84,21,200,39,0,0              ; andps         0x27c8(%rip),%xmm10        # 5310 <_sk_callback_sse2+0x7d4>
+  DB  68,15,86,21,208,39,0,0              ; orps          0x27d0(%rip),%xmm10        # 5320 <_sk_callback_sse2+0x7e4>
+  DB  68,15,88,13,216,39,0,0              ; addps         0x27d8(%rip),%xmm9        # 5330 <_sk_callback_sse2+0x7f4>
+  DB  68,15,40,37,224,39,0,0              ; movaps        0x27e0(%rip),%xmm12        # 5340 <_sk_callback_sse2+0x804>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,224,39,0,0              ; addps         0x27e0(%rip),%xmm10        # 5350 <_sk_callback_sse2+0x811>
-  DB  68,15,40,37,232,39,0,0              ; movaps        0x27e8(%rip),%xmm12        # 5360 <_sk_callback_sse2+0x821>
+  DB  68,15,88,21,224,39,0,0              ; addps         0x27e0(%rip),%xmm10        # 5350 <_sk_callback_sse2+0x814>
+  DB  68,15,40,37,232,39,0,0              ; movaps        0x27e8(%rip),%xmm12        # 5360 <_sk_callback_sse2+0x824>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20566,22 +20563,22 @@ _sk_parametric_b_sse2 LABEL PROC
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,210,39,0,0              ; movaps        0x27d2(%rip),%xmm10        # 5370 <_sk_callback_sse2+0x831>
+  DB  68,15,40,21,210,39,0,0              ; movaps        0x27d2(%rip),%xmm10        # 5370 <_sk_callback_sse2+0x834>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,198,39,0,0              ; addps         0x27c6(%rip),%xmm9        # 5380 <_sk_callback_sse2+0x841>
-  DB  68,15,40,37,206,39,0,0              ; movaps        0x27ce(%rip),%xmm12        # 5390 <_sk_callback_sse2+0x851>
+  DB  68,15,88,13,198,39,0,0              ; addps         0x27c6(%rip),%xmm9        # 5380 <_sk_callback_sse2+0x844>
+  DB  68,15,40,37,206,39,0,0              ; movaps        0x27ce(%rip),%xmm12        # 5390 <_sk_callback_sse2+0x854>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,206,39,0,0              ; movaps        0x27ce(%rip),%xmm12        # 53a0 <_sk_callback_sse2+0x861>
+  DB  68,15,40,37,206,39,0,0              ; movaps        0x27ce(%rip),%xmm12        # 53a0 <_sk_callback_sse2+0x864>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,210,39,0,0              ; movaps        0x27d2(%rip),%xmm13        # 53b0 <_sk_callback_sse2+0x871>
+  DB  68,15,40,45,210,39,0,0              ; movaps        0x27d2(%rip),%xmm13        # 53b0 <_sk_callback_sse2+0x874>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,210,39,0,0              ; mulps         0x27d2(%rip),%xmm13        # 53c0 <_sk_callback_sse2+0x881>
+  DB  68,15,89,45,210,39,0,0              ; mulps         0x27d2(%rip),%xmm13        # 53c0 <_sk_callback_sse2+0x884>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -20615,15 +20612,15 @@ _sk_parametric_a_sse2 LABEL PROC
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,82,39,0,0               ; mulps         0x2752(%rip),%xmm9        # 53d0 <_sk_callback_sse2+0x891>
-  DB  68,15,84,21,90,39,0,0               ; andps         0x275a(%rip),%xmm10        # 53e0 <_sk_callback_sse2+0x8a1>
-  DB  68,15,86,21,98,39,0,0               ; orps          0x2762(%rip),%xmm10        # 53f0 <_sk_callback_sse2+0x8b1>
-  DB  68,15,88,13,106,39,0,0              ; addps         0x276a(%rip),%xmm9        # 5400 <_sk_callback_sse2+0x8c1>
-  DB  68,15,40,37,114,39,0,0              ; movaps        0x2772(%rip),%xmm12        # 5410 <_sk_callback_sse2+0x8d1>
+  DB  68,15,89,13,82,39,0,0               ; mulps         0x2752(%rip),%xmm9        # 53d0 <_sk_callback_sse2+0x894>
+  DB  68,15,84,21,90,39,0,0               ; andps         0x275a(%rip),%xmm10        # 53e0 <_sk_callback_sse2+0x8a4>
+  DB  68,15,86,21,98,39,0,0               ; orps          0x2762(%rip),%xmm10        # 53f0 <_sk_callback_sse2+0x8b4>
+  DB  68,15,88,13,106,39,0,0              ; addps         0x276a(%rip),%xmm9        # 5400 <_sk_callback_sse2+0x8c4>
+  DB  68,15,40,37,114,39,0,0              ; movaps        0x2772(%rip),%xmm12        # 5410 <_sk_callback_sse2+0x8d4>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,114,39,0,0              ; addps         0x2772(%rip),%xmm10        # 5420 <_sk_callback_sse2+0x8e1>
-  DB  68,15,40,37,122,39,0,0              ; movaps        0x277a(%rip),%xmm12        # 5430 <_sk_callback_sse2+0x8f1>
+  DB  68,15,88,21,114,39,0,0              ; addps         0x2772(%rip),%xmm10        # 5420 <_sk_callback_sse2+0x8e4>
+  DB  68,15,40,37,122,39,0,0              ; movaps        0x277a(%rip),%xmm12        # 5430 <_sk_callback_sse2+0x8f4>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20631,22 +20628,22 @@ _sk_parametric_a_sse2 LABEL PROC
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,100,39,0,0              ; movaps        0x2764(%rip),%xmm10        # 5440 <_sk_callback_sse2+0x901>
+  DB  68,15,40,21,100,39,0,0              ; movaps        0x2764(%rip),%xmm10        # 5440 <_sk_callback_sse2+0x904>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,88,39,0,0               ; addps         0x2758(%rip),%xmm9        # 5450 <_sk_callback_sse2+0x911>
-  DB  68,15,40,37,96,39,0,0               ; movaps        0x2760(%rip),%xmm12        # 5460 <_sk_callback_sse2+0x921>
+  DB  68,15,88,13,88,39,0,0               ; addps         0x2758(%rip),%xmm9        # 5450 <_sk_callback_sse2+0x914>
+  DB  68,15,40,37,96,39,0,0               ; movaps        0x2760(%rip),%xmm12        # 5460 <_sk_callback_sse2+0x924>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,96,39,0,0               ; movaps        0x2760(%rip),%xmm12        # 5470 <_sk_callback_sse2+0x931>
+  DB  68,15,40,37,96,39,0,0               ; movaps        0x2760(%rip),%xmm12        # 5470 <_sk_callback_sse2+0x934>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,100,39,0,0              ; movaps        0x2764(%rip),%xmm13        # 5480 <_sk_callback_sse2+0x941>
+  DB  68,15,40,45,100,39,0,0              ; movaps        0x2764(%rip),%xmm13        # 5480 <_sk_callback_sse2+0x944>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,100,39,0,0              ; mulps         0x2764(%rip),%xmm13        # 5490 <_sk_callback_sse2+0x951>
+  DB  68,15,89,45,100,39,0,0              ; mulps         0x2764(%rip),%xmm13        # 5490 <_sk_callback_sse2+0x954>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -20661,29 +20658,29 @@ _sk_parametric_a_sse2 LABEL PROC
 
 PUBLIC _sk_lab_to_xyz_sse2
 _sk_lab_to_xyz_sse2 LABEL PROC
-  DB  15,89,5,65,39,0,0                   ; mulps         0x2741(%rip),%xmm0        # 54a0 <_sk_callback_sse2+0x961>
-  DB  68,15,40,5,73,39,0,0                ; movaps        0x2749(%rip),%xmm8        # 54b0 <_sk_callback_sse2+0x971>
+  DB  15,89,5,65,39,0,0                   ; mulps         0x2741(%rip),%xmm0        # 54a0 <_sk_callback_sse2+0x964>
+  DB  68,15,40,5,73,39,0,0                ; movaps        0x2749(%rip),%xmm8        # 54b0 <_sk_callback_sse2+0x974>
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
-  DB  68,15,40,13,77,39,0,0               ; movaps        0x274d(%rip),%xmm9        # 54c0 <_sk_callback_sse2+0x981>
+  DB  68,15,40,13,77,39,0,0               ; movaps        0x274d(%rip),%xmm9        # 54c0 <_sk_callback_sse2+0x984>
   DB  65,15,88,201                        ; addps         %xmm9,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  65,15,88,209                        ; addps         %xmm9,%xmm2
-  DB  15,88,5,74,39,0,0                   ; addps         0x274a(%rip),%xmm0        # 54d0 <_sk_callback_sse2+0x991>
-  DB  15,89,5,83,39,0,0                   ; mulps         0x2753(%rip),%xmm0        # 54e0 <_sk_callback_sse2+0x9a1>
-  DB  15,89,13,92,39,0,0                  ; mulps         0x275c(%rip),%xmm1        # 54f0 <_sk_callback_sse2+0x9b1>
+  DB  15,88,5,74,39,0,0                   ; addps         0x274a(%rip),%xmm0        # 54d0 <_sk_callback_sse2+0x994>
+  DB  15,89,5,83,39,0,0                   ; mulps         0x2753(%rip),%xmm0        # 54e0 <_sk_callback_sse2+0x9a4>
+  DB  15,89,13,92,39,0,0                  ; mulps         0x275c(%rip),%xmm1        # 54f0 <_sk_callback_sse2+0x9b4>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,21,98,39,0,0                  ; mulps         0x2762(%rip),%xmm2        # 5500 <_sk_callback_sse2+0x9c1>
+  DB  15,89,21,98,39,0,0                  ; mulps         0x2762(%rip),%xmm2        # 5500 <_sk_callback_sse2+0x9c4>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  68,15,92,202                        ; subps         %xmm2,%xmm9
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
   DB  68,15,89,225                        ; mulps         %xmm1,%xmm12
-  DB  15,40,21,87,39,0,0                  ; movaps        0x2757(%rip),%xmm2        # 5510 <_sk_callback_sse2+0x9d1>
+  DB  15,40,21,87,39,0,0                  ; movaps        0x2757(%rip),%xmm2        # 5510 <_sk_callback_sse2+0x9d4>
   DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
   DB  69,15,194,196,1                     ; cmpltps       %xmm12,%xmm8
-  DB  68,15,40,21,86,39,0,0               ; movaps        0x2756(%rip),%xmm10        # 5520 <_sk_callback_sse2+0x9e1>
+  DB  68,15,40,21,86,39,0,0               ; movaps        0x2756(%rip),%xmm10        # 5520 <_sk_callback_sse2+0x9e4>
   DB  65,15,88,202                        ; addps         %xmm10,%xmm1
-  DB  68,15,40,29,90,39,0,0               ; movaps        0x275a(%rip),%xmm11        # 5530 <_sk_callback_sse2+0x9f1>
+  DB  68,15,40,29,90,39,0,0               ; movaps        0x275a(%rip),%xmm11        # 5530 <_sk_callback_sse2+0x9f4>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  69,15,84,224                        ; andps         %xmm8,%xmm12
   DB  68,15,85,193                        ; andnps        %xmm1,%xmm8
@@ -20707,8 +20704,8 @@ _sk_lab_to_xyz_sse2 LABEL PROC
   DB  15,84,194                           ; andps         %xmm2,%xmm0
   DB  65,15,85,209                        ; andnps        %xmm9,%xmm2
   DB  15,86,208                           ; orps          %xmm0,%xmm2
-  DB  68,15,89,5,10,39,0,0                ; mulps         0x270a(%rip),%xmm8        # 5540 <_sk_callback_sse2+0xa01>
-  DB  15,89,21,19,39,0,0                  ; mulps         0x2713(%rip),%xmm2        # 5550 <_sk_callback_sse2+0xa11>
+  DB  68,15,89,5,10,39,0,0                ; mulps         0x270a(%rip),%xmm8        # 5540 <_sk_callback_sse2+0xa04>
+  DB  15,89,21,19,39,0,0                  ; mulps         0x2713(%rip),%xmm2        # 5550 <_sk_callback_sse2+0xa14>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -20722,7 +20719,7 @@ _sk_load_a8_sse2 LABEL PROC
   DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
   DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,251,38,0,0                 ; mulps         0x26fb(%rip),%xmm3        # 5560 <_sk_callback_sse2+0xa21>
+  DB  15,89,29,251,38,0,0                 ; mulps         0x26fb(%rip),%xmm3        # 5560 <_sk_callback_sse2+0xa24>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -20765,7 +20762,7 @@ _sk_gather_a8_sse2 LABEL PROC
   DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
   DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,106,38,0,0                 ; mulps         0x266a(%rip),%xmm3        # 5570 <_sk_callback_sse2+0xa31>
+  DB  15,89,29,106,38,0,0                 ; mulps         0x266a(%rip),%xmm3        # 5570 <_sk_callback_sse2+0xa34>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -20776,7 +20773,7 @@ PUBLIC _sk_store_a8_sse2
 _sk_store_a8_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,94,38,0,0                ; movaps        0x265e(%rip),%xmm8        # 5580 <_sk_callback_sse2+0xa41>
+  DB  68,15,40,5,94,38,0,0                ; movaps        0x265e(%rip),%xmm8        # 5580 <_sk_callback_sse2+0xa44>
   DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
   DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
@@ -20796,9 +20793,9 @@ _sk_load_g8_sse2 LABEL PROC
   DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
   DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,37,38,0,0                   ; mulps         0x2625(%rip),%xmm0        # 5590 <_sk_callback_sse2+0xa51>
+  DB  15,89,5,37,38,0,0                   ; mulps         0x2625(%rip),%xmm0        # 5590 <_sk_callback_sse2+0xa54>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,44,38,0,0                  ; movaps        0x262c(%rip),%xmm3        # 55a0 <_sk_callback_sse2+0xa61>
+  DB  15,40,29,44,38,0,0                  ; movaps        0x262c(%rip),%xmm3        # 55a0 <_sk_callback_sse2+0xa64>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -20839,9 +20836,9 @@ _sk_gather_g8_sse2 LABEL PROC
   DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
   DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,161,37,0,0                  ; mulps         0x25a1(%rip),%xmm0        # 55b0 <_sk_callback_sse2+0xa71>
+  DB  15,89,5,161,37,0,0                  ; mulps         0x25a1(%rip),%xmm0        # 55b0 <_sk_callback_sse2+0xa74>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,168,37,0,0                 ; movaps        0x25a8(%rip),%xmm3        # 55c0 <_sk_callback_sse2+0xa81>
+  DB  15,40,29,168,37,0,0                 ; movaps        0x25a8(%rip),%xmm3        # 55c0 <_sk_callback_sse2+0xa84>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -20902,11 +20899,11 @@ _sk_gather_i8_sse2 LABEL PROC
   DB  102,67,15,110,12,136                ; movd          (%r8,%r9,4),%xmm1
   DB  102,68,15,98,201                    ; punpckldq     %xmm1,%xmm9
   DB  102,68,15,98,200                    ; punpckldq     %xmm0,%xmm9
-  DB  102,15,111,21,199,36,0,0            ; movdqa        0x24c7(%rip),%xmm2        # 55d0 <_sk_callback_sse2+0xa91>
+  DB  102,15,111,21,199,36,0,0            ; movdqa        0x24c7(%rip),%xmm2        # 55d0 <_sk_callback_sse2+0xa94>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,195,36,0,0               ; movaps        0x24c3(%rip),%xmm8        # 55e0 <_sk_callback_sse2+0xaa1>
+  DB  68,15,40,5,195,36,0,0               ; movaps        0x24c3(%rip),%xmm8        # 55e0 <_sk_callback_sse2+0xaa4>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -20931,19 +20928,19 @@ _sk_load_565_sse2 LABEL PROC
   DB  243,15,126,20,120                   ; movq          (%rax,%rdi,2),%xmm2
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
-  DB  102,15,111,5,121,36,0,0             ; movdqa        0x2479(%rip),%xmm0        # 55f0 <_sk_callback_sse2+0xab1>
+  DB  102,15,111,5,121,36,0,0             ; movdqa        0x2479(%rip),%xmm0        # 55f0 <_sk_callback_sse2+0xab4>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,123,36,0,0                  ; mulps         0x247b(%rip),%xmm0        # 5600 <_sk_callback_sse2+0xac1>
-  DB  102,15,111,13,131,36,0,0            ; movdqa        0x2483(%rip),%xmm1        # 5610 <_sk_callback_sse2+0xad1>
+  DB  15,89,5,123,36,0,0                  ; mulps         0x247b(%rip),%xmm0        # 5600 <_sk_callback_sse2+0xac4>
+  DB  102,15,111,13,131,36,0,0            ; movdqa        0x2483(%rip),%xmm1        # 5610 <_sk_callback_sse2+0xad4>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,133,36,0,0                 ; mulps         0x2485(%rip),%xmm1        # 5620 <_sk_callback_sse2+0xae1>
-  DB  102,15,219,21,141,36,0,0            ; pand          0x248d(%rip),%xmm2        # 5630 <_sk_callback_sse2+0xaf1>
+  DB  15,89,13,133,36,0,0                 ; mulps         0x2485(%rip),%xmm1        # 5620 <_sk_callback_sse2+0xae4>
+  DB  102,15,219,21,141,36,0,0            ; pand          0x248d(%rip),%xmm2        # 5630 <_sk_callback_sse2+0xaf4>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,147,36,0,0                 ; mulps         0x2493(%rip),%xmm2        # 5640 <_sk_callback_sse2+0xb01>
+  DB  15,89,21,147,36,0,0                 ; mulps         0x2493(%rip),%xmm2        # 5640 <_sk_callback_sse2+0xb04>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,154,36,0,0                 ; movaps        0x249a(%rip),%xmm3        # 5650 <_sk_callback_sse2+0xb11>
+  DB  15,40,29,154,36,0,0                 ; movaps        0x249a(%rip),%xmm3        # 5650 <_sk_callback_sse2+0xb14>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_gather_565_sse2
@@ -20976,31 +20973,31 @@ _sk_gather_565_sse2 LABEL PROC
   DB  102,15,196,208,3                    ; pinsrw        $0x3,%eax,%xmm2
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
-  DB  102,15,111,5,35,36,0,0              ; movdqa        0x2423(%rip),%xmm0        # 5660 <_sk_callback_sse2+0xb21>
+  DB  102,15,111,5,35,36,0,0              ; movdqa        0x2423(%rip),%xmm0        # 5660 <_sk_callback_sse2+0xb24>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,37,36,0,0                   ; mulps         0x2425(%rip),%xmm0        # 5670 <_sk_callback_sse2+0xb31>
-  DB  102,15,111,13,45,36,0,0             ; movdqa        0x242d(%rip),%xmm1        # 5680 <_sk_callback_sse2+0xb41>
+  DB  15,89,5,37,36,0,0                   ; mulps         0x2425(%rip),%xmm0        # 5670 <_sk_callback_sse2+0xb34>
+  DB  102,15,111,13,45,36,0,0             ; movdqa        0x242d(%rip),%xmm1        # 5680 <_sk_callback_sse2+0xb44>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,47,36,0,0                  ; mulps         0x242f(%rip),%xmm1        # 5690 <_sk_callback_sse2+0xb51>
-  DB  102,15,219,21,55,36,0,0             ; pand          0x2437(%rip),%xmm2        # 56a0 <_sk_callback_sse2+0xb61>
+  DB  15,89,13,47,36,0,0                  ; mulps         0x242f(%rip),%xmm1        # 5690 <_sk_callback_sse2+0xb54>
+  DB  102,15,219,21,55,36,0,0             ; pand          0x2437(%rip),%xmm2        # 56a0 <_sk_callback_sse2+0xb64>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,61,36,0,0                  ; mulps         0x243d(%rip),%xmm2        # 56b0 <_sk_callback_sse2+0xb71>
+  DB  15,89,21,61,36,0,0                  ; mulps         0x243d(%rip),%xmm2        # 56b0 <_sk_callback_sse2+0xb74>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,68,36,0,0                  ; movaps        0x2444(%rip),%xmm3        # 56c0 <_sk_callback_sse2+0xb81>
+  DB  15,40,29,68,36,0,0                  ; movaps        0x2444(%rip),%xmm3        # 56c0 <_sk_callback_sse2+0xb84>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_store_565_sse2
 _sk_store_565_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,69,36,0,0                ; movaps        0x2445(%rip),%xmm8        # 56d0 <_sk_callback_sse2+0xb91>
+  DB  68,15,40,5,69,36,0,0                ; movaps        0x2445(%rip),%xmm8        # 56d0 <_sk_callback_sse2+0xb94>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
   DB  102,65,15,114,241,11                ; pslld         $0xb,%xmm9
-  DB  68,15,40,21,58,36,0,0               ; movaps        0x243a(%rip),%xmm10        # 56e0 <_sk_callback_sse2+0xba1>
+  DB  68,15,40,21,58,36,0,0               ; movaps        0x243a(%rip),%xmm10        # 56e0 <_sk_callback_sse2+0xba4>
   DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
   DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
   DB  102,65,15,114,242,5                 ; pslld         $0x5,%xmm10
@@ -21022,21 +21019,21 @@ _sk_load_4444_sse2 LABEL PROC
   DB  243,15,126,28,120                   ; movq          (%rax,%rdi,2),%xmm3
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,97,216                       ; punpcklwd     %xmm0,%xmm3
-  DB  102,15,111,5,243,35,0,0             ; movdqa        0x23f3(%rip),%xmm0        # 56f0 <_sk_callback_sse2+0xbb1>
+  DB  102,15,111,5,243,35,0,0             ; movdqa        0x23f3(%rip),%xmm0        # 56f0 <_sk_callback_sse2+0xbb4>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,245,35,0,0                  ; mulps         0x23f5(%rip),%xmm0        # 5700 <_sk_callback_sse2+0xbc1>
-  DB  102,15,111,13,253,35,0,0            ; movdqa        0x23fd(%rip),%xmm1        # 5710 <_sk_callback_sse2+0xbd1>
+  DB  15,89,5,245,35,0,0                  ; mulps         0x23f5(%rip),%xmm0        # 5700 <_sk_callback_sse2+0xbc4>
+  DB  102,15,111,13,253,35,0,0            ; movdqa        0x23fd(%rip),%xmm1        # 5710 <_sk_callback_sse2+0xbd4>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,255,35,0,0                 ; mulps         0x23ff(%rip),%xmm1        # 5720 <_sk_callback_sse2+0xbe1>
-  DB  102,15,111,21,7,36,0,0              ; movdqa        0x2407(%rip),%xmm2        # 5730 <_sk_callback_sse2+0xbf1>
+  DB  15,89,13,255,35,0,0                 ; mulps         0x23ff(%rip),%xmm1        # 5720 <_sk_callback_sse2+0xbe4>
+  DB  102,15,111,21,7,36,0,0              ; movdqa        0x2407(%rip),%xmm2        # 5730 <_sk_callback_sse2+0xbf4>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,9,36,0,0                   ; mulps         0x2409(%rip),%xmm2        # 5740 <_sk_callback_sse2+0xc01>
-  DB  102,15,219,29,17,36,0,0             ; pand          0x2411(%rip),%xmm3        # 5750 <_sk_callback_sse2+0xc11>
+  DB  15,89,21,9,36,0,0                   ; mulps         0x2409(%rip),%xmm2        # 5740 <_sk_callback_sse2+0xc04>
+  DB  102,15,219,29,17,36,0,0             ; pand          0x2411(%rip),%xmm3        # 5750 <_sk_callback_sse2+0xc14>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,23,36,0,0                  ; mulps         0x2417(%rip),%xmm3        # 5760 <_sk_callback_sse2+0xc21>
+  DB  15,89,29,23,36,0,0                  ; mulps         0x2417(%rip),%xmm3        # 5760 <_sk_callback_sse2+0xc24>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -21070,21 +21067,21 @@ _sk_gather_4444_sse2 LABEL PROC
   DB  102,15,196,216,3                    ; pinsrw        $0x3,%eax,%xmm3
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,97,216                       ; punpcklwd     %xmm0,%xmm3
-  DB  102,15,111,5,158,35,0,0             ; movdqa        0x239e(%rip),%xmm0        # 5770 <_sk_callback_sse2+0xc31>
+  DB  102,15,111,5,158,35,0,0             ; movdqa        0x239e(%rip),%xmm0        # 5770 <_sk_callback_sse2+0xc34>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,160,35,0,0                  ; mulps         0x23a0(%rip),%xmm0        # 5780 <_sk_callback_sse2+0xc41>
-  DB  102,15,111,13,168,35,0,0            ; movdqa        0x23a8(%rip),%xmm1        # 5790 <_sk_callback_sse2+0xc51>
+  DB  15,89,5,160,35,0,0                  ; mulps         0x23a0(%rip),%xmm0        # 5780 <_sk_callback_sse2+0xc44>
+  DB  102,15,111,13,168,35,0,0            ; movdqa        0x23a8(%rip),%xmm1        # 5790 <_sk_callback_sse2+0xc54>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,170,35,0,0                 ; mulps         0x23aa(%rip),%xmm1        # 57a0 <_sk_callback_sse2+0xc61>
-  DB  102,15,111,21,178,35,0,0            ; movdqa        0x23b2(%rip),%xmm2        # 57b0 <_sk_callback_sse2+0xc71>
+  DB  15,89,13,170,35,0,0                 ; mulps         0x23aa(%rip),%xmm1        # 57a0 <_sk_callback_sse2+0xc64>
+  DB  102,15,111,21,178,35,0,0            ; movdqa        0x23b2(%rip),%xmm2        # 57b0 <_sk_callback_sse2+0xc74>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,180,35,0,0                 ; mulps         0x23b4(%rip),%xmm2        # 57c0 <_sk_callback_sse2+0xc81>
-  DB  102,15,219,29,188,35,0,0            ; pand          0x23bc(%rip),%xmm3        # 57d0 <_sk_callback_sse2+0xc91>
+  DB  15,89,21,180,35,0,0                 ; mulps         0x23b4(%rip),%xmm2        # 57c0 <_sk_callback_sse2+0xc84>
+  DB  102,15,219,29,188,35,0,0            ; pand          0x23bc(%rip),%xmm3        # 57d0 <_sk_callback_sse2+0xc94>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,194,35,0,0                 ; mulps         0x23c2(%rip),%xmm3        # 57e0 <_sk_callback_sse2+0xca1>
+  DB  15,89,29,194,35,0,0                 ; mulps         0x23c2(%rip),%xmm3        # 57e0 <_sk_callback_sse2+0xca4>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -21092,7 +21089,7 @@ PUBLIC _sk_store_4444_sse2
 _sk_store_4444_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,193,35,0,0               ; movaps        0x23c1(%rip),%xmm8        # 57f0 <_sk_callback_sse2+0xcb1>
+  DB  68,15,40,5,193,35,0,0               ; movaps        0x23c1(%rip),%xmm8        # 57f0 <_sk_callback_sse2+0xcb4>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -21122,11 +21119,11 @@ _sk_load_8888_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  68,15,16,12,184                     ; movups        (%rax,%rdi,4),%xmm9
-  DB  15,40,21,84,35,0,0                  ; movaps        0x2354(%rip),%xmm2        # 5800 <_sk_callback_sse2+0xcc1>
+  DB  15,40,21,84,35,0,0                  ; movaps        0x2354(%rip),%xmm2        # 5800 <_sk_callback_sse2+0xcc4>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,84,194                           ; andps         %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,82,35,0,0                ; movaps        0x2352(%rip),%xmm8        # 5810 <_sk_callback_sse2+0xcd1>
+  DB  68,15,40,5,82,35,0,0                ; movaps        0x2352(%rip),%xmm8        # 5810 <_sk_callback_sse2+0xcd4>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -21173,11 +21170,11 @@ _sk_gather_8888_sse2 LABEL PROC
   DB  102,67,15,110,12,129                ; movd          (%r9,%r8,4),%xmm1
   DB  102,68,15,98,201                    ; punpckldq     %xmm1,%xmm9
   DB  102,68,15,98,200                    ; punpckldq     %xmm0,%xmm9
-  DB  102,15,111,21,163,34,0,0            ; movdqa        0x22a3(%rip),%xmm2        # 5820 <_sk_callback_sse2+0xce1>
+  DB  102,15,111,21,163,34,0,0            ; movdqa        0x22a3(%rip),%xmm2        # 5820 <_sk_callback_sse2+0xce4>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,159,34,0,0               ; movaps        0x229f(%rip),%xmm8        # 5830 <_sk_callback_sse2+0xcf1>
+  DB  68,15,40,5,159,34,0,0               ; movaps        0x229f(%rip),%xmm8        # 5830 <_sk_callback_sse2+0xcf4>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -21199,7 +21196,7 @@ PUBLIC _sk_store_8888_sse2
 _sk_store_8888_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,5,98,34,0,0                ; movaps        0x2262(%rip),%xmm8        # 5840 <_sk_callback_sse2+0xd01>
+  DB  68,15,40,5,98,34,0,0                ; movaps        0x2262(%rip),%xmm8        # 5840 <_sk_callback_sse2+0xd04>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -21236,7 +21233,7 @@ _sk_load_f16_sse2 LABEL PROC
   DB  102,69,15,239,210                   ; pxor          %xmm10,%xmm10
   DB  102,65,15,111,206                   ; movdqa        %xmm14,%xmm1
   DB  102,65,15,97,202                    ; punpcklwd     %xmm10,%xmm1
-  DB  102,68,15,111,13,210,33,0,0         ; movdqa        0x21d2(%rip),%xmm9        # 5850 <_sk_callback_sse2+0xd11>
+  DB  102,68,15,111,13,210,33,0,0         ; movdqa        0x21d2(%rip),%xmm9        # 5850 <_sk_callback_sse2+0xd14>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
   DB  102,15,239,200                      ; pxor          %xmm0,%xmm1
@@ -21244,11 +21241,11 @@ _sk_load_f16_sse2 LABEL PROC
   DB  102,68,15,111,233                   ; movdqa        %xmm1,%xmm13
   DB  102,65,15,114,245,13                ; pslld         $0xd,%xmm13
   DB  102,68,15,235,232                   ; por           %xmm0,%xmm13
-  DB  102,68,15,111,29,183,33,0,0         ; movdqa        0x21b7(%rip),%xmm11        # 5860 <_sk_callback_sse2+0xd21>
+  DB  102,68,15,111,29,183,33,0,0         ; movdqa        0x21b7(%rip),%xmm11        # 5860 <_sk_callback_sse2+0xd24>
   DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
-  DB  102,68,15,111,37,185,33,0,0         ; movdqa        0x21b9(%rip),%xmm12        # 5870 <_sk_callback_sse2+0xd31>
+  DB  102,68,15,111,37,185,33,0,0         ; movdqa        0x21b9(%rip),%xmm12        # 5870 <_sk_callback_sse2+0xd34>
   DB  102,65,15,239,204                   ; pxor          %xmm12,%xmm1
-  DB  102,15,111,29,188,33,0,0            ; movdqa        0x21bc(%rip),%xmm3        # 5880 <_sk_callback_sse2+0xd41>
+  DB  102,15,111,29,188,33,0,0            ; movdqa        0x21bc(%rip),%xmm3        # 5880 <_sk_callback_sse2+0xd44>
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,102,193                      ; pcmpgtd       %xmm1,%xmm0
   DB  102,65,15,223,197                   ; pandn         %xmm13,%xmm0
@@ -21332,7 +21329,7 @@ _sk_gather_f16_sse2 LABEL PROC
   DB  102,69,15,239,210                   ; pxor          %xmm10,%xmm10
   DB  102,65,15,111,206                   ; movdqa        %xmm14,%xmm1
   DB  102,65,15,97,202                    ; punpcklwd     %xmm10,%xmm1
-  DB  102,68,15,111,13,74,32,0,0          ; movdqa        0x204a(%rip),%xmm9        # 5890 <_sk_callback_sse2+0xd51>
+  DB  102,68,15,111,13,74,32,0,0          ; movdqa        0x204a(%rip),%xmm9        # 5890 <_sk_callback_sse2+0xd54>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
   DB  102,15,239,200                      ; pxor          %xmm0,%xmm1
@@ -21340,11 +21337,11 @@ _sk_gather_f16_sse2 LABEL PROC
   DB  102,68,15,111,233                   ; movdqa        %xmm1,%xmm13
   DB  102,65,15,114,245,13                ; pslld         $0xd,%xmm13
   DB  102,68,15,235,232                   ; por           %xmm0,%xmm13
-  DB  102,68,15,111,29,47,32,0,0          ; movdqa        0x202f(%rip),%xmm11        # 58a0 <_sk_callback_sse2+0xd61>
+  DB  102,68,15,111,29,47,32,0,0          ; movdqa        0x202f(%rip),%xmm11        # 58a0 <_sk_callback_sse2+0xd64>
   DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
-  DB  102,68,15,111,37,49,32,0,0          ; movdqa        0x2031(%rip),%xmm12        # 58b0 <_sk_callback_sse2+0xd71>
+  DB  102,68,15,111,37,49,32,0,0          ; movdqa        0x2031(%rip),%xmm12        # 58b0 <_sk_callback_sse2+0xd74>
   DB  102,65,15,239,204                   ; pxor          %xmm12,%xmm1
-  DB  102,15,111,29,52,32,0,0             ; movdqa        0x2034(%rip),%xmm3        # 58c0 <_sk_callback_sse2+0xd81>
+  DB  102,15,111,29,52,32,0,0             ; movdqa        0x2034(%rip),%xmm3        # 58c0 <_sk_callback_sse2+0xd84>
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,102,193                      ; pcmpgtd       %xmm1,%xmm0
   DB  102,65,15,223,197                   ; pandn         %xmm13,%xmm0
@@ -21395,17 +21392,17 @@ PUBLIC _sk_store_f16_sse2
 _sk_store_f16_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  102,68,15,111,21,92,31,0,0          ; movdqa        0x1f5c(%rip),%xmm10        # 58d0 <_sk_callback_sse2+0xd91>
+  DB  102,68,15,111,21,92,31,0,0          ; movdqa        0x1f5c(%rip),%xmm10        # 58d0 <_sk_callback_sse2+0xd94>
   DB  102,68,15,111,224                   ; movdqa        %xmm0,%xmm12
   DB  102,68,15,111,232                   ; movdqa        %xmm0,%xmm13
   DB  102,69,15,219,234                   ; pand          %xmm10,%xmm13
   DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
-  DB  102,68,15,111,13,79,31,0,0          ; movdqa        0x1f4f(%rip),%xmm9        # 58e0 <_sk_callback_sse2+0xda1>
+  DB  102,68,15,111,13,79,31,0,0          ; movdqa        0x1f4f(%rip),%xmm9        # 58e0 <_sk_callback_sse2+0xda4>
   DB  102,65,15,114,213,16                ; psrld         $0x10,%xmm13
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
   DB  102,69,15,102,196                   ; pcmpgtd       %xmm12,%xmm8
   DB  102,65,15,114,212,13                ; psrld         $0xd,%xmm12
-  DB  102,68,15,111,29,64,31,0,0          ; movdqa        0x1f40(%rip),%xmm11        # 58f0 <_sk_callback_sse2+0xdb1>
+  DB  102,68,15,111,29,64,31,0,0          ; movdqa        0x1f40(%rip),%xmm11        # 58f0 <_sk_callback_sse2+0xdb4>
   DB  102,69,15,235,235                   ; por           %xmm11,%xmm13
   DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
   DB  102,65,15,114,245,16                ; pslld         $0x10,%xmm13
@@ -21482,7 +21479,7 @@ _sk_load_u16_be_sse2 LABEL PROC
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
   DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
-  DB  68,15,40,5,222,29,0,0               ; movaps        0x1dde(%rip),%xmm8        # 5900 <_sk_callback_sse2+0xdc1>
+  DB  68,15,40,5,222,29,0,0               ; movaps        0x1dde(%rip),%xmm8        # 5900 <_sk_callback_sse2+0xdc4>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -21533,7 +21530,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,65,15,97,192                    ; punpcklwd     %xmm8,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,13,26,29,0,0               ; movaps        0x1d1a(%rip),%xmm9        # 5910 <_sk_callback_sse2+0xdd1>
+  DB  68,15,40,13,26,29,0,0               ; movaps        0x1d1a(%rip),%xmm9        # 5910 <_sk_callback_sse2+0xdd4>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -21550,14 +21547,14 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,225,28,0,0                 ; movaps        0x1ce1(%rip),%xmm3        # 5920 <_sk_callback_sse2+0xde1>
+  DB  15,40,29,225,28,0,0                 ; movaps        0x1ce1(%rip),%xmm3        # 5920 <_sk_callback_sse2+0xde4>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_store_u16_be_sse2
 _sk_store_u16_be_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  72,139,0                            ; mov           (%rax),%rax
-  DB  68,15,40,13,226,28,0,0              ; movaps        0x1ce2(%rip),%xmm9        # 5930 <_sk_callback_sse2+0xdf1>
+  DB  68,15,40,13,226,28,0,0              ; movaps        0x1ce2(%rip),%xmm9        # 5930 <_sk_callback_sse2+0xdf4>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
@@ -21693,7 +21690,7 @@ _sk_repeat_x_sse2 LABEL PROC
   DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,84,13,226,26,0,0              ; andps         0x1ae2(%rip),%xmm9        # 5940 <_sk_callback_sse2+0xe01>
+  DB  68,15,84,13,226,26,0,0              ; andps         0x1ae2(%rip),%xmm9        # 5940 <_sk_callback_sse2+0xe04>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
   DB  65,15,92,194                        ; subps         %xmm10,%xmm0
@@ -21711,7 +21708,7 @@ _sk_repeat_y_sse2 LABEL PROC
   DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,84,13,180,26,0,0              ; andps         0x1ab4(%rip),%xmm9        # 5950 <_sk_callback_sse2+0xe11>
+  DB  68,15,84,13,180,26,0,0              ; andps         0x1ab4(%rip),%xmm9        # 5950 <_sk_callback_sse2+0xe14>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
   DB  65,15,92,202                        ; subps         %xmm10,%xmm1
@@ -21733,7 +21730,7 @@ _sk_mirror_x_sse2 LABEL PROC
   DB  243,69,15,91,218                    ; cvttps2dq     %xmm10,%xmm11
   DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
   DB  69,15,194,211,1                     ; cmpltps       %xmm11,%xmm10
-  DB  68,15,84,21,116,26,0,0              ; andps         0x1a74(%rip),%xmm10        # 5960 <_sk_callback_sse2+0xe21>
+  DB  68,15,84,21,116,26,0,0              ; andps         0x1a74(%rip),%xmm10        # 5960 <_sk_callback_sse2+0xe24>
   DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
@@ -21759,7 +21756,7 @@ _sk_mirror_y_sse2 LABEL PROC
   DB  243,69,15,91,218                    ; cvttps2dq     %xmm10,%xmm11
   DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
   DB  69,15,194,211,1                     ; cmpltps       %xmm11,%xmm10
-  DB  68,15,84,21,36,26,0,0               ; andps         0x1a24(%rip),%xmm10        # 5970 <_sk_callback_sse2+0xe31>
+  DB  68,15,84,21,36,26,0,0               ; andps         0x1a24(%rip),%xmm10        # 5970 <_sk_callback_sse2+0xe34>
   DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
@@ -21774,10 +21771,10 @@ _sk_mirror_y_sse2 LABEL PROC
 PUBLIC _sk_luminance_to_alpha_sse2
 _sk_luminance_to_alpha_sse2 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
-  DB  15,89,5,6,26,0,0                    ; mulps         0x1a06(%rip),%xmm0        # 5980 <_sk_callback_sse2+0xe41>
-  DB  15,89,13,15,26,0,0                  ; mulps         0x1a0f(%rip),%xmm1        # 5990 <_sk_callback_sse2+0xe51>
+  DB  15,89,5,6,26,0,0                    ; mulps         0x1a06(%rip),%xmm0        # 5980 <_sk_callback_sse2+0xe44>
+  DB  15,89,13,15,26,0,0                  ; mulps         0x1a0f(%rip),%xmm1        # 5990 <_sk_callback_sse2+0xe54>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,29,21,26,0,0                  ; mulps         0x1a15(%rip),%xmm3        # 59a0 <_sk_callback_sse2+0xe61>
+  DB  15,89,29,21,26,0,0                  ; mulps         0x1a15(%rip),%xmm3        # 59a0 <_sk_callback_sse2+0xe64>
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -22098,7 +22095,7 @@ _sk_gradient_sse2 LABEL PROC
   DB  73,255,200                          ; dec           %r8
   DB  72,131,193,4                        ; add           $0x4,%rcx
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
-  DB  15,40,21,234,20,0,0                 ; movaps        0x14ea(%rip),%xmm2        # 59b0 <_sk_callback_sse2+0xe71>
+  DB  15,40,21,234,20,0,0                 ; movaps        0x14ea(%rip),%xmm2        # 59b0 <_sk_callback_sse2+0xe74>
   DB  243,15,16,25                        ; movss         (%rcx),%xmm3
   DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
   DB  15,194,216,2                        ; cmpleps       %xmm0,%xmm3
@@ -22242,29 +22239,29 @@ _sk_xy_to_unit_angle_sse2 LABEL PROC
   DB  69,15,94,220                        ; divps         %xmm12,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
-  DB  68,15,40,45,172,18,0,0              ; movaps        0x12ac(%rip),%xmm13        # 59c0 <_sk_callback_sse2+0xe81>
+  DB  68,15,40,45,172,18,0,0              ; movaps        0x12ac(%rip),%xmm13        # 59c0 <_sk_callback_sse2+0xe84>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,176,18,0,0              ; addps         0x12b0(%rip),%xmm13        # 59d0 <_sk_callback_sse2+0xe91>
+  DB  68,15,88,45,176,18,0,0              ; addps         0x12b0(%rip),%xmm13        # 59d0 <_sk_callback_sse2+0xe94>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,180,18,0,0              ; addps         0x12b4(%rip),%xmm13        # 59e0 <_sk_callback_sse2+0xea1>
+  DB  68,15,88,45,180,18,0,0              ; addps         0x12b4(%rip),%xmm13        # 59e0 <_sk_callback_sse2+0xea4>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,184,18,0,0              ; addps         0x12b8(%rip),%xmm13        # 59f0 <_sk_callback_sse2+0xeb1>
+  DB  68,15,88,45,184,18,0,0              ; addps         0x12b8(%rip),%xmm13        # 59f0 <_sk_callback_sse2+0xeb4>
   DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,40,21,183,18,0,0              ; movaps        0x12b7(%rip),%xmm10        # 5a00 <_sk_callback_sse2+0xec1>
+  DB  68,15,40,21,183,18,0,0              ; movaps        0x12b7(%rip),%xmm10        # 5a00 <_sk_callback_sse2+0xec4>
   DB  69,15,92,213                        ; subps         %xmm13,%xmm10
   DB  69,15,84,209                        ; andps         %xmm9,%xmm10
   DB  69,15,85,205                        ; andnps        %xmm13,%xmm9
   DB  69,15,86,202                        ; orps          %xmm10,%xmm9
   DB  68,15,194,192,1                     ; cmpltps       %xmm0,%xmm8
-  DB  68,15,40,21,170,18,0,0              ; movaps        0x12aa(%rip),%xmm10        # 5a10 <_sk_callback_sse2+0xed1>
+  DB  68,15,40,21,170,18,0,0              ; movaps        0x12aa(%rip),%xmm10        # 5a10 <_sk_callback_sse2+0xed4>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,84,208                        ; andps         %xmm8,%xmm10
   DB  69,15,85,193                        ; andnps        %xmm9,%xmm8
   DB  69,15,86,194                        ; orps          %xmm10,%xmm8
   DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
   DB  68,15,194,200,1                     ; cmpltps       %xmm0,%xmm9
-  DB  68,15,40,21,153,18,0,0              ; movaps        0x1299(%rip),%xmm10        # 5a20 <_sk_callback_sse2+0xee1>
+  DB  68,15,40,21,153,18,0,0              ; movaps        0x1299(%rip),%xmm10        # 5a20 <_sk_callback_sse2+0xee4>
   DB  69,15,92,208                        ; subps         %xmm8,%xmm10
   DB  69,15,84,209                        ; andps         %xmm9,%xmm10
   DB  69,15,85,200                        ; andnps        %xmm8,%xmm9
@@ -22280,15 +22277,14 @@ _sk_xy_to_radius_sse2 LABEL PROC
   DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
   DB  68,15,88,192                        ; addps         %xmm0,%xmm8
-  DB  65,15,82,192                        ; rsqrtps       %xmm8,%xmm0
-  DB  15,83,192                           ; rcpps         %xmm0,%xmm0
+  DB  65,15,81,192                        ; sqrtps        %xmm8,%xmm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_save_xy_sse2
 _sk_save_xy_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,104,18,0,0               ; movaps        0x1268(%rip),%xmm8        # 5a30 <_sk_callback_sse2+0xef1>
+  DB  68,15,40,5,107,18,0,0               ; movaps        0x126b(%rip),%xmm8        # 5a30 <_sk_callback_sse2+0xef4>
   DB  15,17,0                             ; movups        %xmm0,(%rax)
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,88,200                        ; addps         %xmm8,%xmm9
@@ -22296,7 +22292,7 @@ _sk_save_xy_sse2 LABEL PROC
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
   DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
   DB  69,15,194,218,1                     ; cmpltps       %xmm10,%xmm11
-  DB  68,15,40,37,83,18,0,0               ; movaps        0x1253(%rip),%xmm12        # 5a40 <_sk_callback_sse2+0xf01>
+  DB  68,15,40,37,86,18,0,0               ; movaps        0x1256(%rip),%xmm12        # 5a40 <_sk_callback_sse2+0xf04>
   DB  69,15,84,220                        ; andps         %xmm12,%xmm11
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
@@ -22339,8 +22335,8 @@ _sk_bilinear_nx_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,204,17,0,0                  ; addps         0x11cc(%rip),%xmm0        # 5a50 <_sk_callback_sse2+0xf11>
-  DB  68,15,40,13,212,17,0,0              ; movaps        0x11d4(%rip),%xmm9        # 5a60 <_sk_callback_sse2+0xf21>
+  DB  15,88,5,207,17,0,0                  ; addps         0x11cf(%rip),%xmm0        # 5a50 <_sk_callback_sse2+0xf14>
+  DB  68,15,40,13,215,17,0,0              ; movaps        0x11d7(%rip),%xmm9        # 5a60 <_sk_callback_sse2+0xf24>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -22351,7 +22347,7 @@ _sk_bilinear_px_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,195,17,0,0                  ; addps         0x11c3(%rip),%xmm0        # 5a70 <_sk_callback_sse2+0xf31>
+  DB  15,88,5,198,17,0,0                  ; addps         0x11c6(%rip),%xmm0        # 5a70 <_sk_callback_sse2+0xf34>
   DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -22361,8 +22357,8 @@ _sk_bilinear_ny_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,181,17,0,0                 ; addps         0x11b5(%rip),%xmm1        # 5a80 <_sk_callback_sse2+0xf41>
-  DB  68,15,40,13,189,17,0,0              ; movaps        0x11bd(%rip),%xmm9        # 5a90 <_sk_callback_sse2+0xf51>
+  DB  15,88,13,184,17,0,0                 ; addps         0x11b8(%rip),%xmm1        # 5a80 <_sk_callback_sse2+0xf44>
+  DB  68,15,40,13,192,17,0,0              ; movaps        0x11c0(%rip),%xmm9        # 5a90 <_sk_callback_sse2+0xf54>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -22373,7 +22369,7 @@ _sk_bilinear_py_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,171,17,0,0                 ; addps         0x11ab(%rip),%xmm1        # 5aa0 <_sk_callback_sse2+0xf61>
+  DB  15,88,13,174,17,0,0                 ; addps         0x11ae(%rip),%xmm1        # 5aa0 <_sk_callback_sse2+0xf64>
   DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -22383,13 +22379,13 @@ _sk_bicubic_n3x_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,158,17,0,0                  ; addps         0x119e(%rip),%xmm0        # 5ab0 <_sk_callback_sse2+0xf71>
-  DB  68,15,40,13,166,17,0,0              ; movaps        0x11a6(%rip),%xmm9        # 5ac0 <_sk_callback_sse2+0xf81>
+  DB  15,88,5,161,17,0,0                  ; addps         0x11a1(%rip),%xmm0        # 5ab0 <_sk_callback_sse2+0xf74>
+  DB  68,15,40,13,169,17,0,0              ; movaps        0x11a9(%rip),%xmm9        # 5ac0 <_sk_callback_sse2+0xf84>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,162,17,0,0              ; mulps         0x11a2(%rip),%xmm9        # 5ad0 <_sk_callback_sse2+0xf91>
-  DB  68,15,88,13,170,17,0,0              ; addps         0x11aa(%rip),%xmm9        # 5ae0 <_sk_callback_sse2+0xfa1>
+  DB  68,15,89,13,165,17,0,0              ; mulps         0x11a5(%rip),%xmm9        # 5ad0 <_sk_callback_sse2+0xf94>
+  DB  68,15,88,13,173,17,0,0              ; addps         0x11ad(%rip),%xmm9        # 5ae0 <_sk_callback_sse2+0xfa4>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -22400,16 +22396,16 @@ _sk_bicubic_n1x_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,153,17,0,0                  ; addps         0x1199(%rip),%xmm0        # 5af0 <_sk_callback_sse2+0xfb1>
-  DB  68,15,40,13,161,17,0,0              ; movaps        0x11a1(%rip),%xmm9        # 5b00 <_sk_callback_sse2+0xfc1>
+  DB  15,88,5,156,17,0,0                  ; addps         0x119c(%rip),%xmm0        # 5af0 <_sk_callback_sse2+0xfb4>
+  DB  68,15,40,13,164,17,0,0              ; movaps        0x11a4(%rip),%xmm9        # 5b00 <_sk_callback_sse2+0xfc4>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,165,17,0,0               ; movaps        0x11a5(%rip),%xmm8        # 5b10 <_sk_callback_sse2+0xfd1>
+  DB  68,15,40,5,168,17,0,0               ; movaps        0x11a8(%rip),%xmm8        # 5b10 <_sk_callback_sse2+0xfd4>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,169,17,0,0               ; addps         0x11a9(%rip),%xmm8        # 5b20 <_sk_callback_sse2+0xfe1>
+  DB  68,15,88,5,172,17,0,0               ; addps         0x11ac(%rip),%xmm8        # 5b20 <_sk_callback_sse2+0xfe4>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,173,17,0,0               ; addps         0x11ad(%rip),%xmm8        # 5b30 <_sk_callback_sse2+0xff1>
+  DB  68,15,88,5,176,17,0,0               ; addps         0x11b0(%rip),%xmm8        # 5b30 <_sk_callback_sse2+0xff4>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,177,17,0,0               ; addps         0x11b1(%rip),%xmm8        # 5b40 <_sk_callback_sse2+0x1001>
+  DB  68,15,88,5,180,17,0,0               ; addps         0x11b4(%rip),%xmm8        # 5b40 <_sk_callback_sse2+0x1004>
   DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -22417,17 +22413,17 @@ _sk_bicubic_n1x_sse2 LABEL PROC
 PUBLIC _sk_bicubic_p1x_sse2
 _sk_bicubic_p1x_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,171,17,0,0               ; movaps        0x11ab(%rip),%xmm8        # 5b50 <_sk_callback_sse2+0x1011>
+  DB  68,15,40,5,174,17,0,0               ; movaps        0x11ae(%rip),%xmm8        # 5b50 <_sk_callback_sse2+0x1014>
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,21,167,17,0,0              ; movaps        0x11a7(%rip),%xmm10        # 5b60 <_sk_callback_sse2+0x1021>
+  DB  68,15,40,21,170,17,0,0              ; movaps        0x11aa(%rip),%xmm10        # 5b60 <_sk_callback_sse2+0x1024>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,171,17,0,0              ; addps         0x11ab(%rip),%xmm10        # 5b70 <_sk_callback_sse2+0x1031>
+  DB  68,15,88,21,174,17,0,0              ; addps         0x11ae(%rip),%xmm10        # 5b70 <_sk_callback_sse2+0x1034>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,167,17,0,0              ; addps         0x11a7(%rip),%xmm10        # 5b80 <_sk_callback_sse2+0x1041>
+  DB  68,15,88,21,170,17,0,0              ; addps         0x11aa(%rip),%xmm10        # 5b80 <_sk_callback_sse2+0x1044>
   DB  68,15,17,144,128,0,0,0              ; movups        %xmm10,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -22437,11 +22433,11 @@ _sk_bicubic_p3x_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,64,64                      ; movups        0x40(%rax),%xmm8
-  DB  15,88,5,154,17,0,0                  ; addps         0x119a(%rip),%xmm0        # 5b90 <_sk_callback_sse2+0x1051>
+  DB  15,88,5,157,17,0,0                  ; addps         0x119d(%rip),%xmm0        # 5b90 <_sk_callback_sse2+0x1054>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,154,17,0,0               ; mulps         0x119a(%rip),%xmm8        # 5ba0 <_sk_callback_sse2+0x1061>
-  DB  68,15,88,5,162,17,0,0               ; addps         0x11a2(%rip),%xmm8        # 5bb0 <_sk_callback_sse2+0x1071>
+  DB  68,15,89,5,157,17,0,0               ; mulps         0x119d(%rip),%xmm8        # 5ba0 <_sk_callback_sse2+0x1064>
+  DB  68,15,88,5,165,17,0,0               ; addps         0x11a5(%rip),%xmm8        # 5bb0 <_sk_callback_sse2+0x1074>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -22452,13 +22448,13 @@ _sk_bicubic_n3y_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,144,17,0,0                 ; addps         0x1190(%rip),%xmm1        # 5bc0 <_sk_callback_sse2+0x1081>
-  DB  68,15,40,13,152,17,0,0              ; movaps        0x1198(%rip),%xmm9        # 5bd0 <_sk_callback_sse2+0x1091>
+  DB  15,88,13,147,17,0,0                 ; addps         0x1193(%rip),%xmm1        # 5bc0 <_sk_callback_sse2+0x1084>
+  DB  68,15,40,13,155,17,0,0              ; movaps        0x119b(%rip),%xmm9        # 5bd0 <_sk_callback_sse2+0x1094>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,148,17,0,0              ; mulps         0x1194(%rip),%xmm9        # 5be0 <_sk_callback_sse2+0x10a1>
-  DB  68,15,88,13,156,17,0,0              ; addps         0x119c(%rip),%xmm9        # 5bf0 <_sk_callback_sse2+0x10b1>
+  DB  68,15,89,13,151,17,0,0              ; mulps         0x1197(%rip),%xmm9        # 5be0 <_sk_callback_sse2+0x10a4>
+  DB  68,15,88,13,159,17,0,0              ; addps         0x119f(%rip),%xmm9        # 5bf0 <_sk_callback_sse2+0x10b4>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -22469,16 +22465,16 @@ _sk_bicubic_n1y_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,138,17,0,0                 ; addps         0x118a(%rip),%xmm1        # 5c00 <_sk_callback_sse2+0x10c1>
-  DB  68,15,40,13,146,17,0,0              ; movaps        0x1192(%rip),%xmm9        # 5c10 <_sk_callback_sse2+0x10d1>
+  DB  15,88,13,141,17,0,0                 ; addps         0x118d(%rip),%xmm1        # 5c00 <_sk_callback_sse2+0x10c4>
+  DB  68,15,40,13,149,17,0,0              ; movaps        0x1195(%rip),%xmm9        # 5c10 <_sk_callback_sse2+0x10d4>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,150,17,0,0               ; movaps        0x1196(%rip),%xmm8        # 5c20 <_sk_callback_sse2+0x10e1>
+  DB  68,15,40,5,153,17,0,0               ; movaps        0x1199(%rip),%xmm8        # 5c20 <_sk_callback_sse2+0x10e4>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,154,17,0,0               ; addps         0x119a(%rip),%xmm8        # 5c30 <_sk_callback_sse2+0x10f1>
+  DB  68,15,88,5,157,17,0,0               ; addps         0x119d(%rip),%xmm8        # 5c30 <_sk_callback_sse2+0x10f4>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,158,17,0,0               ; addps         0x119e(%rip),%xmm8        # 5c40 <_sk_callback_sse2+0x1101>
+  DB  68,15,88,5,161,17,0,0               ; addps         0x11a1(%rip),%xmm8        # 5c40 <_sk_callback_sse2+0x1104>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,162,17,0,0               ; addps         0x11a2(%rip),%xmm8        # 5c50 <_sk_callback_sse2+0x1111>
+  DB  68,15,88,5,165,17,0,0               ; addps         0x11a5(%rip),%xmm8        # 5c50 <_sk_callback_sse2+0x1114>
   DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -22486,17 +22482,17 @@ _sk_bicubic_n1y_sse2 LABEL PROC
 PUBLIC _sk_bicubic_p1y_sse2
 _sk_bicubic_p1y_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,156,17,0,0               ; movaps        0x119c(%rip),%xmm8        # 5c60 <_sk_callback_sse2+0x1121>
+  DB  68,15,40,5,159,17,0,0               ; movaps        0x119f(%rip),%xmm8        # 5c60 <_sk_callback_sse2+0x1124>
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  68,15,40,21,151,17,0,0              ; movaps        0x1197(%rip),%xmm10        # 5c70 <_sk_callback_sse2+0x1131>
+  DB  68,15,40,21,154,17,0,0              ; movaps        0x119a(%rip),%xmm10        # 5c70 <_sk_callback_sse2+0x1134>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,155,17,0,0              ; addps         0x119b(%rip),%xmm10        # 5c80 <_sk_callback_sse2+0x1141>
+  DB  68,15,88,21,158,17,0,0              ; addps         0x119e(%rip),%xmm10        # 5c80 <_sk_callback_sse2+0x1144>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,151,17,0,0              ; addps         0x1197(%rip),%xmm10        # 5c90 <_sk_callback_sse2+0x1151>
+  DB  68,15,88,21,154,17,0,0              ; addps         0x119a(%rip),%xmm10        # 5c90 <_sk_callback_sse2+0x1154>
   DB  68,15,17,144,160,0,0,0              ; movups        %xmm10,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -22506,11 +22502,11 @@ _sk_bicubic_p3y_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
   DB  68,15,16,64,96                      ; movups        0x60(%rax),%xmm8
-  DB  15,88,13,137,17,0,0                 ; addps         0x1189(%rip),%xmm1        # 5ca0 <_sk_callback_sse2+0x1161>
+  DB  15,88,13,140,17,0,0                 ; addps         0x118c(%rip),%xmm1        # 5ca0 <_sk_callback_sse2+0x1164>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,137,17,0,0               ; mulps         0x1189(%rip),%xmm8        # 5cb0 <_sk_callback_sse2+0x1171>
-  DB  68,15,88,5,145,17,0,0               ; addps         0x1191(%rip),%xmm8        # 5cc0 <_sk_callback_sse2+0x1181>
+  DB  68,15,89,5,140,17,0,0               ; mulps         0x118c(%rip),%xmm8        # 5cb0 <_sk_callback_sse2+0x1174>
+  DB  68,15,88,5,148,17,0,0               ; addps         0x1194(%rip),%xmm8        # 5cc0 <_sk_callback_sse2+0x1184>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23198,7 +23194,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a391aa <_sk_callback_sse2+0xffffffffe9a3466b>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a391aa <_sk_callback_sse2+0xffffffffe9a3466e>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23293,7 +23289,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3927a <_sk_callback_sse2+0xffffffffe9a3473b>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3927a <_sk_callback_sse2+0xffffffffe9a3473e>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23388,7 +23384,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3934a <_sk_callback_sse2+0xffffffffe9a3480b>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3934a <_sk_callback_sse2+0xffffffffe9a3480e>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23483,7 +23479,7 @@ ALIGN 16
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3941a <_sk_callback_sse2+0xffffffffe9a348db>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3941a <_sk_callback_sse2+0xffffffffe9a348de>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -24166,7 +24162,7 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e985 <_sk_callback_sse2+0x3d639e46>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e985 <_sk_callback_sse2+0x3d639e49>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -24192,7 +24188,7 @@ ALIGN 16
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e9c5 <_sk_callback_sse2+0x3d639e86>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e9c5 <_sk_callback_sse2+0x3d639e89>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -24276,7 +24272,7 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63ea95 <_sk_callback_sse2+0x3d639f56>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63ea95 <_sk_callback_sse2+0x3d639f59>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -24302,7 +24298,7 @@ ALIGN 16
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63ead5 <_sk_callback_sse2+0x3d639f96>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63ead5 <_sk_callback_sse2+0x3d639f99>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -24313,11 +24309,11 @@ ALIGN 16
   DB  63                                  ; (bad)
   DB  114,28                              ; jb            5cce <.literal16+0x10de>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5cd2 <_sk_callback_sse2+0x1193>
+  DB  62,114,28                           ; jb,pt         5cd2 <_sk_callback_sse2+0x1196>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5cd6 <_sk_callback_sse2+0x1197>
+  DB  62,114,28                           ; jb,pt         5cd6 <_sk_callback_sse2+0x119a>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5cda <_sk_callback_sse2+0x119b>
+  DB  62,114,28                           ; jb,pt         5cda <_sk_callback_sse2+0x119e>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
index 8076e59..a486403 100644 (file)
@@ -1147,7 +1147,7 @@ STAGE(xy_to_unit_angle) {
 STAGE(xy_to_radius) {
     F X2 = r * r,
       Y2 = g * g;
-    r = rcp(rsqrt(X2 + Y2));
+    r = sqrt_(X2 + Y2);
 }
 
 STAGE(save_xy) {
index 504aca0..073b946 100644 (file)
@@ -34,6 +34,7 @@
     SI F   floor_(F v)          { return floorf(v); }
     SI F   rcp   (F v)          { return 1.0f / v; }
     SI F   rsqrt (F v)          { return 1.0f / sqrtf(v); }
+    SI F    sqrt_(F v)          { return sqrtf(v); }
     SI U32 round (F v, F scale) { return (uint32_t)lrintf(v*scale); }
     SI U16 pack(U32 v)          { return (U16)v; }
     SI U8  pack(U16 v)          { return  (U8)v; }
@@ -94,6 +95,7 @@
     SI F   floor_(F v)                           { return vrndmq_f32(v);           }
     SI F   rcp   (F v) { auto e = vrecpeq_f32 (v); return vrecpsq_f32 (v,e  ) * e; }
     SI F   rsqrt (F v) { auto e = vrsqrteq_f32(v); return vrsqrtsq_f32(v,e*e) * e; }
+    SI F    sqrt_(F v)                           { return vsqrtq_f32(v); }
     SI U32 round (F v, F scale)                  { return vcvtnq_u32_f32(v*scale); }
     SI U16 pack(U32 v)                           { return __builtin_convertvector(v, U16); }
     SI U8  pack(U16 v)                           { return __builtin_convertvector(v,  U8); }
     SI U16 pack(U32 v)                         { return __builtin_convertvector(v, U16); }
     SI U8  pack(U16 v)                         { return __builtin_convertvector(v,  U8); }
 
+    SI F sqrt_(F v) {
+        auto e = vrsqrte_f32(v);  // Estimate and two refinement steps for e = rsqrt(v).
+        e *= vrsqrts_f32(v,e*e);
+        e *= vrsqrts_f32(v,e*e);
+        return v*e;               // sqrt(v) == v*rsqrt(v).
+    }
+
     SI F if_then_else(I32 c, F t, F e) { return vbsl_f32((U32)c,t,e); }
 
     SI F floor_(F v) {
     SI F   floor_(F v)          { return _mm256_floor_ps(v);    }
     SI F   rcp   (F v)          { return _mm256_rcp_ps  (v);    }
     SI F   rsqrt (F v)          { return _mm256_rsqrt_ps(v);    }
+    SI F    sqrt_(F v)          { return _mm256_sqrt_ps (v);    }
     SI U32 round (F v, F scale) { return _mm256_cvtps_epi32(v*scale); }
 
     SI U16 pack(U32 v) {
     SI F   min(F a, F b)       { return _mm_min_ps(a,b);    }
     SI F   max(F a, F b)       { return _mm_max_ps(a,b);    }
     SI F   abs_(F v)           { return _mm_and_ps(v, 0-v); }
-    SI F   rcp  (F v)          { return _mm_rcp_ps  (v);    }
-    SI F   rsqrt(F v)          { return _mm_rsqrt_ps(v);    }
+    SI F   rcp   (F v)         { return _mm_rcp_ps  (v);    }
+    SI F   rsqrt (F v)         { return _mm_rsqrt_ps(v);    }
+    SI F    sqrt_(F v)         { return _mm_sqrt_ps (v);    }
     SI U32 round(F v, F scale) { return _mm_cvtps_epi32(v*scale); }
 
     SI U16 pack(U32 v) {