retry tilers against 1
authorMike Klein <mtklein@chromium.org>
Tue, 23 May 2017 11:52:01 +0000 (07:52 -0400)
committerSkia Commit-Bot <skia-commit-bot@chromium.org>
Tue, 23 May 2017 12:17:15 +0000 (12:17 +0000)
These weren't the Valgrind problem.

Change-Id: Ic76df460cf0ce7f7ae8155d549f4e92a7f8de040
Reviewed-on: https://skia-review.googlesource.com/17701
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>

src/core/SkRasterPipeline.h
src/effects/gradients/SkGradientShader.cpp
src/jumper/SkJumper_generated.S
src/jumper/SkJumper_generated_win.S
src/jumper/SkJumper_stages.cpp

index 6287d62..3725033 100644 (file)
@@ -90,8 +90,9 @@
     M(parametric_a)                                              \
     M(table_r) M(table_g) M(table_b) M(table_a)                  \
     M(lab_to_xyz)                                                \
-    M(clamp_x) M(mirror_x) M(repeat_x)                           \
-    M(clamp_y) M(mirror_y) M(repeat_y)                           \
+    M(clamp_x)   M(mirror_x)   M(repeat_x)                       \
+    M(clamp_y)   M(mirror_y)   M(repeat_y)                       \
+    M(clamp_x_1) M(mirror_x_1) M(repeat_x_1)                     \
     M(gather_a8) M(gather_g8) M(gather_i8)                       \
     M(gather_565) M(gather_4444) M(gather_8888) M(gather_f16)    \
     M(bilinear_nx) M(bilinear_px) M(bilinear_ny) M(bilinear_py)  \
index 0c6788d..c4a1dad 100644 (file)
@@ -433,15 +433,15 @@ bool SkGradientShaderBase::onAppendStages(SkRasterPipeline* p,
     p->extend(subclass);
 
     switch(fTileMode) {
-        case kMirror_TileMode: p->append(SkRasterPipeline::mirror_x, alloc->make<float>(1)); break;
-        case kRepeat_TileMode: p->append(SkRasterPipeline::repeat_x, alloc->make<float>(1)); break;
+        case kMirror_TileMode: p->append(SkRasterPipeline::mirror_x_1); break;
+        case kRepeat_TileMode: p->append(SkRasterPipeline::repeat_x_1); break;
         case kClamp_TileMode:
             if (!fOrigPos) {
                 // We clamp only when the stops are evenly spaced.
                 // If not, there may be hard stops, and clamping ruins hard stops at 0 and/or 1.
                 // In that case, we must make sure we're using the general "gradient" stage,
                 // which is the only stage that will correctly handle unclamped t.
-                p->append(SkRasterPipeline::clamp_x,  alloc->make<float>(1));
+                p->append(SkRasterPipeline::clamp_x_1);
             }
     }
 
index 143f6b5..27a0b76 100644 (file)
@@ -3379,6 +3379,42 @@ _sk_mirror_y_aarch64:
   .long  0x4eb1f421                          // fmin          v1.4s, v1.4s, v17.4s
   .long  0xd61f0060                          // br            x3
 
+HIDDEN _sk_clamp_x_1_aarch64
+.globl _sk_clamp_x_1_aarch64
+FUNCTION(_sk_clamp_x_1_aarch64)
+_sk_clamp_x_1_aarch64:
+  .long  0xf8408423                          // ldr           x3, [x1], #8
+  .long  0x6f00e410                          // movi          v16.2d, #0x0
+  .long  0x4e20f600                          // fmax          v0.4s, v16.4s, v0.4s
+  .long  0x4f03f610                          // fmov          v16.4s, #1.000000000000000000e+00
+  .long  0x4eb0f400                          // fmin          v0.4s, v0.4s, v16.4s
+  .long  0xd61f0060                          // br            x3
+
+HIDDEN _sk_repeat_x_1_aarch64
+.globl _sk_repeat_x_1_aarch64
+FUNCTION(_sk_repeat_x_1_aarch64)
+_sk_repeat_x_1_aarch64:
+  .long  0xf8408423                          // ldr           x3, [x1], #8
+  .long  0x4e219810                          // frintm        v16.4s, v0.4s
+  .long  0x4eb0d400                          // fsub          v0.4s, v0.4s, v16.4s
+  .long  0xd61f0060                          // br            x3
+
+HIDDEN _sk_mirror_x_1_aarch64
+.globl _sk_mirror_x_1_aarch64
+FUNCTION(_sk_mirror_x_1_aarch64)
+_sk_mirror_x_1_aarch64:
+  .long  0x4f07f610                          // fmov          v16.4s, #-1.000000000000000000e+00
+  .long  0x4f0167f1                          // movi          v17.4s, #0x3f, lsl #24
+  .long  0x4e30d400                          // fadd          v0.4s, v0.4s, v16.4s
+  .long  0x6e31dc11                          // fmul          v17.4s, v0.4s, v17.4s
+  .long  0x4e219a31                          // frintm        v17.4s, v17.4s
+  .long  0x4e31d631                          // fadd          v17.4s, v17.4s, v17.4s
+  .long  0xf8408423                          // ldr           x3, [x1], #8
+  .long  0x4eb1d400                          // fsub          v0.4s, v0.4s, v17.4s
+  .long  0x4e30d400                          // fadd          v0.4s, v0.4s, v16.4s
+  .long  0x4ea0f800                          // fabs          v0.4s, v0.4s
+  .long  0xd61f0060                          // br            x3
+
 HIDDEN _sk_luminance_to_alpha_aarch64
 .globl _sk_luminance_to_alpha_aarch64
 FUNCTION(_sk_luminance_to_alpha_aarch64)
@@ -3673,7 +3709,7 @@ _sk_gradient_aarch64:
   .long  0x6f00e411                          // movi          v17.2d, #0x0
   .long  0xf9400109                          // ldr           x9, [x8]
   .long  0xf100093f                          // cmp           x9, #0x2
-  .long  0x540001c3                          // b.cc          3120 <sk_gradient_aarch64+0x58>  // b.lo, b.ul, b.last
+  .long  0x540001c3                          // b.cc          3174 <sk_gradient_aarch64+0x58>  // b.lo, b.ul, b.last
   .long  0xf940250a                          // ldr           x10, [x8, #72]
   .long  0xd1000529                          // sub           x9, x9, #0x1
   .long  0x6f00e401                          // movi          v1.2d, #0x0
@@ -3684,7 +3720,7 @@ _sk_gradient_aarch64:
   .long  0x6e23e403                          // fcmge         v3.4s, v0.4s, v3.4s
   .long  0x4e221c63                          // and           v3.16b, v3.16b, v2.16b
   .long  0x4ea18461                          // add           v1.4s, v3.4s, v1.4s
-  .long  0xb5ffff69                          // cbnz          x9, 3100 <sk_gradient_aarch64+0x38>
+  .long  0xb5ffff69                          // cbnz          x9, 3154 <sk_gradient_aarch64+0x38>
   .long  0x6f20a431                          // uxtl2         v17.2d, v1.4s
   .long  0x2f20a421                          // uxtl          v1.2d, v1.2s
   .long  0xa940b10a                          // ldp           x10, x12, [x8, #8]
@@ -7874,6 +7910,59 @@ _sk_mirror_y_vfp4:
   .long  0x3f800000                          // .word         0x3f800000
   .long  0x3f800000                          // .word         0x3f800000
 
+HIDDEN _sk_clamp_x_1_vfp4
+.globl _sk_clamp_x_1_vfp4
+FUNCTION(_sk_clamp_x_1_vfp4)
+_sk_clamp_x_1_vfp4:
+  .long  0xf2c00010                          // vmov.i32      d16, #0
+  .long  0xe4913004                          // ldr           r3, [r1], #4
+  .long  0xf2c71f10                          // vmov.f32      d17, #1
+  .long  0xf2400f80                          // vmax.f32      d16, d16, d0
+  .long  0xf2200fa1                          // vmin.f32      d0, d16, d17
+  .long  0xe12fff13                          // bx            r3
+
+HIDDEN _sk_repeat_x_1_vfp4
+.globl _sk_repeat_x_1_vfp4
+FUNCTION(_sk_repeat_x_1_vfp4)
+_sk_repeat_x_1_vfp4:
+  .long  0xf3fb0700                          // vcvt.s32.f32  d16, d0
+  .long  0xeddf3b07                          // vldr          d19, [pc, #28]
+  .long  0xf2c02010                          // vmov.i32      d18, #0
+  .long  0xe4913004                          // ldr           r3, [r1], #4
+  .long  0xf3fb0620                          // vcvt.f32.s32  d16, d16
+  .long  0xf3601e80                          // vcgt.f32      d17, d16, d0
+  .long  0xf35311b2                          // vbsl          d17, d19, d18
+  .long  0xf2600da1                          // vsub.f32      d16, d16, d17
+  .long  0xf2200d20                          // vsub.f32      d0, d0, d16
+  .long  0xe12fff13                          // bx            r3
+  .long  0x3f800000                          // .word         0x3f800000
+  .long  0x3f800000                          // .word         0x3f800000
+
+HIDDEN _sk_mirror_x_1_vfp4
+.globl _sk_mirror_x_1_vfp4
+FUNCTION(_sk_mirror_x_1_vfp4)
+_sk_mirror_x_1_vfp4:
+  .long  0xf3c70f10                          // vmov.f32      d16, #-1
+  .long  0xeddf5b0f                          // vldr          d21, [pc, #60]
+  .long  0xf2c3261f                          // vmov.i32      d18, #1056964608
+  .long  0xe4913004                          // ldr           r3, [r1], #4
+  .long  0xf2401d20                          // vadd.f32      d17, d0, d16
+  .long  0xf2c04010                          // vmov.i32      d20, #0
+  .long  0xf3412db2                          // vmul.f32      d18, d17, d18
+  .long  0xf3fb3722                          // vcvt.s32.f32  d19, d18
+  .long  0xf3fb3623                          // vcvt.f32.s32  d19, d19
+  .long  0xf3632ea2                          // vcgt.f32      d18, d19, d18
+  .long  0xf35521b4                          // vbsl          d18, d21, d20
+  .long  0xf2632da2                          // vsub.f32      d18, d19, d18
+  .long  0xf2422da2                          // vadd.f32      d18, d18, d18
+  .long  0xf2611da2                          // vsub.f32      d17, d17, d18
+  .long  0xf2410da0                          // vadd.f32      d16, d17, d16
+  .long  0xf3b90720                          // vabs.f32      d0, d16
+  .long  0xe12fff13                          // bx            r3
+  .long  0xe320f000                          // nop           {0}
+  .long  0x3f800000                          // .word         0x3f800000
+  .long  0x3f800000                          // .word         0x3f800000
+
 HIDDEN _sk_luminance_to_alpha_vfp4
 .globl _sk_luminance_to_alpha_vfp4
 FUNCTION(_sk_luminance_to_alpha_vfp4)
@@ -8195,7 +8284,7 @@ _sk_gradient_vfp4:
   .long  0xf2c00010                          // vmov.i32      d16, #0
   .long  0xe59c3000                          // ldr           r3, [ip]
   .long  0xe3530002                          // cmp           r3, #2
-  .long  0x3a00000b                          // bcc           36e4 <sk_gradient_vfp4+0x50>
+  .long  0x3a00000b                          // bcc           377c <sk_gradient_vfp4+0x50>
   .long  0xe59c4024                          // ldr           r4, [ip, #36]
   .long  0xf2c01010                          // vmov.i32      d17, #0
   .long  0xf2c02011                          // vmov.i32      d18, #1
@@ -8207,7 +8296,7 @@ _sk_gradient_vfp4:
   .long  0xf3403e23                          // vcge.f32      d19, d0, d19
   .long  0xf35231b1                          // vbsl          d19, d18, d17
   .long  0xf26308a0                          // vadd.i32      d16, d19, d16
-  .long  0x1afffff9                          // bne           36cc <sk_gradient_vfp4+0x38>
+  .long  0x1afffff9                          // bne           3764 <sk_gradient_vfp4+0x38>
   .long  0xee303b90                          // vmov.32       r3, d16[1]
   .long  0xe59c7010                          // ldr           r7, [ip, #16]
   .long  0xee10eb90                          // vmov.32       lr, d16[0]
@@ -8800,14 +8889,14 @@ _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  196,226,125,24,13,81,71,0,0         // vbroadcastss  0x4751(%rip),%ymm1        # 4814 <_sk_callback_hsw+0x127>
+  .byte  196,226,125,24,13,189,71,0,0        // vbroadcastss  0x47bd(%rip),%ymm1        # 4880 <_sk_callback_hsw+0x127>
   .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  196,226,125,24,21,53,71,0,0         // vbroadcastss  0x4735(%rip),%ymm2        # 4818 <_sk_callback_hsw+0x12b>
+  .byte  196,226,125,24,21,161,71,0,0        // vbroadcastss  0x47a1(%rip),%ymm2        # 4884 <_sk_callback_hsw+0x12b>
   .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
@@ -8826,13 +8915,13 @@ _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  196,98,125,88,21,253,70,0,0         // vpbroadcastd  0x46fd(%rip),%ymm10        # 481c <_sk_callback_hsw+0x12f>
+  .byte  196,98,125,88,21,105,71,0,0         // vpbroadcastd  0x4769(%rip),%ymm10        # 4888 <_sk_callback_hsw+0x12f>
   .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  196,98,125,88,37,226,70,0,0         // vpbroadcastd  0x46e2(%rip),%ymm12        # 4820 <_sk_callback_hsw+0x133>
-  .byte  196,98,125,88,45,221,70,0,0         // vpbroadcastd  0x46dd(%rip),%ymm13        # 4824 <_sk_callback_hsw+0x137>
+  .byte  196,98,125,88,37,78,71,0,0          // vpbroadcastd  0x474e(%rip),%ymm12        # 488c <_sk_callback_hsw+0x133>
+  .byte  196,98,125,88,45,73,71,0,0          // vpbroadcastd  0x4749(%rip),%ymm13        # 4890 <_sk_callback_hsw+0x137>
   .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
@@ -8847,8 +8936,8 @@ _sk_dither_hsw:
   .byte  196,65,13,235,201                   // vpor          %ymm9,%ymm14,%ymm9
   .byte  196,65,61,235,193                   // vpor          %ymm9,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,143,70,0,0         // vbroadcastss  0x468f(%rip),%ymm9        # 4828 <_sk_callback_hsw+0x13b>
-  .byte  196,98,125,24,21,138,70,0,0         // vbroadcastss  0x468a(%rip),%ymm10        # 482c <_sk_callback_hsw+0x13f>
+  .byte  196,98,125,24,13,251,70,0,0         // vbroadcastss  0x46fb(%rip),%ymm9        # 4894 <_sk_callback_hsw+0x13b>
+  .byte  196,98,125,24,21,246,70,0,0         // vbroadcastss  0x46f6(%rip),%ymm10        # 4898 <_sk_callback_hsw+0x13f>
   .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
@@ -8917,7 +9006,7 @@ HIDDEN _sk_srcatop_hsw
 FUNCTION(_sk_srcatop_hsw)
 _sk_srcatop_hsw:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,225,69,0,0          // vbroadcastss  0x45e1(%rip),%ymm8        # 4830 <_sk_callback_hsw+0x143>
+  .byte  196,98,125,24,5,77,70,0,0           // vbroadcastss  0x464d(%rip),%ymm8        # 489c <_sk_callback_hsw+0x143>
   .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
@@ -8933,7 +9022,7 @@ HIDDEN _sk_dstatop_hsw
 .globl _sk_dstatop_hsw
 FUNCTION(_sk_dstatop_hsw)
 _sk_dstatop_hsw:
-  .byte  196,98,125,24,5,180,69,0,0          // vbroadcastss  0x45b4(%rip),%ymm8        # 4834 <_sk_callback_hsw+0x147>
+  .byte  196,98,125,24,5,32,70,0,0           // vbroadcastss  0x4620(%rip),%ymm8        # 48a0 <_sk_callback_hsw+0x147>
   .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
@@ -8972,7 +9061,7 @@ HIDDEN _sk_srcout_hsw
 .globl _sk_srcout_hsw
 FUNCTION(_sk_srcout_hsw)
 _sk_srcout_hsw:
-  .byte  196,98,125,24,5,91,69,0,0           // vbroadcastss  0x455b(%rip),%ymm8        # 4838 <_sk_callback_hsw+0x14b>
+  .byte  196,98,125,24,5,199,69,0,0          // vbroadcastss  0x45c7(%rip),%ymm8        # 48a4 <_sk_callback_hsw+0x14b>
   .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
@@ -8985,7 +9074,7 @@ HIDDEN _sk_dstout_hsw
 .globl _sk_dstout_hsw
 FUNCTION(_sk_dstout_hsw)
 _sk_dstout_hsw:
-  .byte  196,226,125,24,5,62,69,0,0          // vbroadcastss  0x453e(%rip),%ymm0        # 483c <_sk_callback_hsw+0x14f>
+  .byte  196,226,125,24,5,170,69,0,0         // vbroadcastss  0x45aa(%rip),%ymm0        # 48a8 <_sk_callback_hsw+0x14f>
   .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
@@ -8998,7 +9087,7 @@ HIDDEN _sk_srcover_hsw
 .globl _sk_srcover_hsw
 FUNCTION(_sk_srcover_hsw)
 _sk_srcover_hsw:
-  .byte  196,98,125,24,5,33,69,0,0           // vbroadcastss  0x4521(%rip),%ymm8        # 4840 <_sk_callback_hsw+0x153>
+  .byte  196,98,125,24,5,141,69,0,0          // vbroadcastss  0x458d(%rip),%ymm8        # 48ac <_sk_callback_hsw+0x153>
   .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
@@ -9011,7 +9100,7 @@ HIDDEN _sk_dstover_hsw
 .globl _sk_dstover_hsw
 FUNCTION(_sk_dstover_hsw)
 _sk_dstover_hsw:
-  .byte  196,98,125,24,5,0,69,0,0            // vbroadcastss  0x4500(%rip),%ymm8        # 4844 <_sk_callback_hsw+0x157>
+  .byte  196,98,125,24,5,108,69,0,0          // vbroadcastss  0x456c(%rip),%ymm8        # 48b0 <_sk_callback_hsw+0x157>
   .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
@@ -9035,7 +9124,7 @@ HIDDEN _sk_multiply_hsw
 .globl _sk_multiply_hsw
 FUNCTION(_sk_multiply_hsw)
 _sk_multiply_hsw:
-  .byte  196,98,125,24,5,203,68,0,0          // vbroadcastss  0x44cb(%rip),%ymm8        # 4848 <_sk_callback_hsw+0x15b>
+  .byte  196,98,125,24,5,55,69,0,0           // vbroadcastss  0x4537(%rip),%ymm8        # 48b4 <_sk_callback_hsw+0x15b>
   .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
@@ -9083,7 +9172,7 @@ HIDDEN _sk_xor__hsw
 .globl _sk_xor__hsw
 FUNCTION(_sk_xor__hsw)
 _sk_xor__hsw:
-  .byte  196,98,125,24,5,70,68,0,0           // vbroadcastss  0x4446(%rip),%ymm8        # 484c <_sk_callback_hsw+0x15f>
+  .byte  196,98,125,24,5,178,68,0,0          // vbroadcastss  0x44b2(%rip),%ymm8        # 48b8 <_sk_callback_hsw+0x15f>
   .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
@@ -9117,7 +9206,7 @@ _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  196,98,125,24,5,206,67,0,0          // vbroadcastss  0x43ce(%rip),%ymm8        # 4850 <_sk_callback_hsw+0x163>
+  .byte  196,98,125,24,5,58,68,0,0           // vbroadcastss  0x443a(%rip),%ymm8        # 48bc <_sk_callback_hsw+0x163>
   .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
@@ -9142,7 +9231,7 @@ _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  196,98,125,24,5,125,67,0,0          // vbroadcastss  0x437d(%rip),%ymm8        # 4854 <_sk_callback_hsw+0x167>
+  .byte  196,98,125,24,5,233,67,0,0          // vbroadcastss  0x43e9(%rip),%ymm8        # 48c0 <_sk_callback_hsw+0x167>
   .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
@@ -9170,7 +9259,7 @@ _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  196,98,125,24,5,32,67,0,0           // vbroadcastss  0x4320(%rip),%ymm8        # 4858 <_sk_callback_hsw+0x16b>
+  .byte  196,98,125,24,5,140,67,0,0          // vbroadcastss  0x438c(%rip),%ymm8        # 48c4 <_sk_callback_hsw+0x16b>
   .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
@@ -9192,7 +9281,7 @@ _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  196,98,125,24,5,222,66,0,0          // vbroadcastss  0x42de(%rip),%ymm8        # 485c <_sk_callback_hsw+0x16f>
+  .byte  196,98,125,24,5,74,67,0,0           // vbroadcastss  0x434a(%rip),%ymm8        # 48c8 <_sk_callback_hsw+0x16f>
   .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
@@ -9202,7 +9291,7 @@ HIDDEN _sk_colorburn_hsw
 .globl _sk_colorburn_hsw
 FUNCTION(_sk_colorburn_hsw)
 _sk_colorburn_hsw:
-  .byte  196,98,125,24,5,204,66,0,0          // vbroadcastss  0x42cc(%rip),%ymm8        # 4860 <_sk_callback_hsw+0x173>
+  .byte  196,98,125,24,5,56,67,0,0           // vbroadcastss  0x4338(%rip),%ymm8        # 48cc <_sk_callback_hsw+0x173>
   .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
@@ -9260,7 +9349,7 @@ HIDDEN _sk_colordodge_hsw
 FUNCTION(_sk_colordodge_hsw)
 _sk_colordodge_hsw:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,13,215,65,0,0         // vbroadcastss  0x41d7(%rip),%ymm9        # 4864 <_sk_callback_hsw+0x177>
+  .byte  196,98,125,24,13,67,66,0,0          // vbroadcastss  0x4243(%rip),%ymm9        # 48d0 <_sk_callback_hsw+0x177>
   .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
@@ -9313,7 +9402,7 @@ HIDDEN _sk_hardlight_hsw
 .globl _sk_hardlight_hsw
 FUNCTION(_sk_hardlight_hsw)
 _sk_hardlight_hsw:
-  .byte  196,98,125,24,5,248,64,0,0          // vbroadcastss  0x40f8(%rip),%ymm8        # 4868 <_sk_callback_hsw+0x17b>
+  .byte  196,98,125,24,5,100,65,0,0          // vbroadcastss  0x4164(%rip),%ymm8        # 48d4 <_sk_callback_hsw+0x17b>
   .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
@@ -9364,7 +9453,7 @@ HIDDEN _sk_overlay_hsw
 .globl _sk_overlay_hsw
 FUNCTION(_sk_overlay_hsw)
 _sk_overlay_hsw:
-  .byte  196,98,125,24,5,48,64,0,0           // vbroadcastss  0x4030(%rip),%ymm8        # 486c <_sk_callback_hsw+0x17f>
+  .byte  196,98,125,24,5,156,64,0,0          // vbroadcastss  0x409c(%rip),%ymm8        # 48d8 <_sk_callback_hsw+0x17f>
   .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
@@ -9425,10 +9514,10 @@ _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  196,98,125,24,29,59,63,0,0          // vbroadcastss  0x3f3b(%rip),%ymm11        # 4874 <_sk_callback_hsw+0x187>
+  .byte  196,98,125,24,29,167,63,0,0         // vbroadcastss  0x3fa7(%rip),%ymm11        # 48e0 <_sk_callback_hsw+0x187>
   .byte  196,65,20,88,227                    // vaddps        %ymm11,%ymm13,%ymm12
   .byte  196,65,28,89,192                    // vmulps        %ymm8,%ymm12,%ymm8
-  .byte  196,98,125,24,37,44,63,0,0          // vbroadcastss  0x3f2c(%rip),%ymm12        # 4878 <_sk_callback_hsw+0x18b>
+  .byte  196,98,125,24,37,152,63,0,0         // vbroadcastss  0x3f98(%rip),%ymm12        # 48e4 <_sk_callback_hsw+0x18b>
   .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
@@ -9438,7 +9527,7 @@ _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  196,98,125,24,5,239,62,0,0          // vbroadcastss  0x3eef(%rip),%ymm8        # 4870 <_sk_callback_hsw+0x183>
+  .byte  196,98,125,24,5,91,63,0,0           // vbroadcastss  0x3f5b(%rip),%ymm8        # 48dc <_sk_callback_hsw+0x183>
   .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
@@ -9551,11 +9640,11 @@ _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  196,98,125,24,53,243,60,0,0         // vbroadcastss  0x3cf3(%rip),%ymm14        # 487c <_sk_callback_hsw+0x18f>
-  .byte  196,98,125,24,61,238,60,0,0         // vbroadcastss  0x3cee(%rip),%ymm15        # 4880 <_sk_callback_hsw+0x193>
+  .byte  196,98,125,24,53,95,61,0,0          // vbroadcastss  0x3d5f(%rip),%ymm14        # 48e8 <_sk_callback_hsw+0x18f>
+  .byte  196,98,125,24,61,90,61,0,0          // vbroadcastss  0x3d5a(%rip),%ymm15        # 48ec <_sk_callback_hsw+0x193>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,66,93,184,238                   // vfmadd231ps   %ymm14,%ymm4,%ymm13
-  .byte  196,226,125,24,5,223,60,0,0         // vbroadcastss  0x3cdf(%rip),%ymm0        # 4884 <_sk_callback_hsw+0x197>
+  .byte  196,226,125,24,5,75,61,0,0          // vbroadcastss  0x3d4b(%rip),%ymm0        # 48f0 <_sk_callback_hsw+0x197>
   .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
@@ -9610,7 +9699,7 @@ _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  196,226,125,24,13,204,59,0,0        // vbroadcastss  0x3bcc(%rip),%ymm1        # 4888 <_sk_callback_hsw+0x19b>
+  .byte  196,226,125,24,13,56,60,0,0         // vbroadcastss  0x3c38(%rip),%ymm1        # 48f4 <_sk_callback_hsw+0x19b>
   .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
@@ -9664,11 +9753,11 @@ _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  196,98,125,24,53,227,58,0,0         // vbroadcastss  0x3ae3(%rip),%ymm14        # 488c <_sk_callback_hsw+0x19f>
-  .byte  196,98,125,24,61,222,58,0,0         // vbroadcastss  0x3ade(%rip),%ymm15        # 4890 <_sk_callback_hsw+0x1a3>
+  .byte  196,98,125,24,53,79,59,0,0          // vbroadcastss  0x3b4f(%rip),%ymm14        # 48f8 <_sk_callback_hsw+0x19f>
+  .byte  196,98,125,24,61,74,59,0,0          // vbroadcastss  0x3b4a(%rip),%ymm15        # 48fc <_sk_callback_hsw+0x1a3>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,66,93,184,238                   // vfmadd231ps   %ymm14,%ymm4,%ymm13
-  .byte  196,226,125,24,5,207,58,0,0         // vbroadcastss  0x3acf(%rip),%ymm0        # 4894 <_sk_callback_hsw+0x1a7>
+  .byte  196,226,125,24,5,59,59,0,0          // vbroadcastss  0x3b3b(%rip),%ymm0        # 4900 <_sk_callback_hsw+0x1a7>
   .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
@@ -9723,7 +9812,7 @@ _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  196,226,125,24,13,188,57,0,0        // vbroadcastss  0x39bc(%rip),%ymm1        # 4898 <_sk_callback_hsw+0x1ab>
+  .byte  196,226,125,24,13,40,58,0,0         // vbroadcastss  0x3a28(%rip),%ymm1        # 4904 <_sk_callback_hsw+0x1ab>
   .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
@@ -9751,11 +9840,11 @@ _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  196,98,125,24,45,85,57,0,0          // vbroadcastss  0x3955(%rip),%ymm13        # 489c <_sk_callback_hsw+0x1af>
-  .byte  196,98,125,24,53,80,57,0,0          // vbroadcastss  0x3950(%rip),%ymm14        # 48a0 <_sk_callback_hsw+0x1b3>
+  .byte  196,98,125,24,45,193,57,0,0         // vbroadcastss  0x39c1(%rip),%ymm13        # 4908 <_sk_callback_hsw+0x1af>
+  .byte  196,98,125,24,53,188,57,0,0         // vbroadcastss  0x39bc(%rip),%ymm14        # 490c <_sk_callback_hsw+0x1b3>
   .byte  196,65,84,89,230                    // vmulps        %ymm14,%ymm5,%ymm12
   .byte  196,66,93,184,229                   // vfmadd231ps   %ymm13,%ymm4,%ymm12
-  .byte  196,98,125,24,61,65,57,0,0          // vbroadcastss  0x3941(%rip),%ymm15        # 48a4 <_sk_callback_hsw+0x1b7>
+  .byte  196,98,125,24,61,173,57,0,0         // vbroadcastss  0x39ad(%rip),%ymm15        # 4910 <_sk_callback_hsw+0x1b7>
   .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
@@ -9811,7 +9900,7 @@ _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  196,226,125,24,5,35,56,0,0          // vbroadcastss  0x3823(%rip),%ymm0        # 48a8 <_sk_callback_hsw+0x1bb>
+  .byte  196,226,125,24,5,143,56,0,0         // vbroadcastss  0x388f(%rip),%ymm0        # 4914 <_sk_callback_hsw+0x1bb>
   .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
@@ -9839,11 +9928,11 @@ _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  196,98,125,24,45,188,55,0,0         // vbroadcastss  0x37bc(%rip),%ymm13        # 48ac <_sk_callback_hsw+0x1bf>
-  .byte  196,98,125,24,53,183,55,0,0         // vbroadcastss  0x37b7(%rip),%ymm14        # 48b0 <_sk_callback_hsw+0x1c3>
+  .byte  196,98,125,24,45,40,56,0,0          // vbroadcastss  0x3828(%rip),%ymm13        # 4918 <_sk_callback_hsw+0x1bf>
+  .byte  196,98,125,24,53,35,56,0,0          // vbroadcastss  0x3823(%rip),%ymm14        # 491c <_sk_callback_hsw+0x1c3>
   .byte  196,65,116,89,230                   // vmulps        %ymm14,%ymm1,%ymm12
   .byte  196,66,109,184,229                  // vfmadd231ps   %ymm13,%ymm2,%ymm12
-  .byte  196,98,125,24,61,168,55,0,0         // vbroadcastss  0x37a8(%rip),%ymm15        # 48b4 <_sk_callback_hsw+0x1c7>
+  .byte  196,98,125,24,61,20,56,0,0          // vbroadcastss  0x3814(%rip),%ymm15        # 4920 <_sk_callback_hsw+0x1c7>
   .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
@@ -9899,7 +9988,7 @@ _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  196,226,125,24,5,138,54,0,0         // vbroadcastss  0x368a(%rip),%ymm0        # 48b8 <_sk_callback_hsw+0x1cb>
+  .byte  196,226,125,24,5,246,54,0,0         // vbroadcastss  0x36f6(%rip),%ymm0        # 4924 <_sk_callback_hsw+0x1cb>
   .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
@@ -9932,7 +10021,7 @@ HIDDEN _sk_clamp_1_hsw
 .globl _sk_clamp_1_hsw
 FUNCTION(_sk_clamp_1_hsw)
 _sk_clamp_1_hsw:
-  .byte  196,98,125,24,5,38,54,0,0           // vbroadcastss  0x3626(%rip),%ymm8        # 48bc <_sk_callback_hsw+0x1cf>
+  .byte  196,98,125,24,5,146,54,0,0          // vbroadcastss  0x3692(%rip),%ymm8        # 4928 <_sk_callback_hsw+0x1cf>
   .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
@@ -9944,7 +10033,7 @@ HIDDEN _sk_clamp_a_hsw
 .globl _sk_clamp_a_hsw
 FUNCTION(_sk_clamp_a_hsw)
 _sk_clamp_a_hsw:
-  .byte  196,98,125,24,5,9,54,0,0            // vbroadcastss  0x3609(%rip),%ymm8        # 48c0 <_sk_callback_hsw+0x1d3>
+  .byte  196,98,125,24,5,117,54,0,0          // vbroadcastss  0x3675(%rip),%ymm8        # 492c <_sk_callback_hsw+0x1d3>
   .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
@@ -10030,7 +10119,7 @@ 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  196,98,125,24,21,81,53,0,0          // vbroadcastss  0x3551(%rip),%ymm10        # 48c4 <_sk_callback_hsw+0x1d7>
+  .byte  196,98,125,24,21,189,53,0,0         // vbroadcastss  0x35bd(%rip),%ymm10        # 4930 <_sk_callback_hsw+0x1d7>
   .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
@@ -10043,16 +10132,16 @@ HIDDEN _sk_from_srgb_hsw
 .globl _sk_from_srgb_hsw
 FUNCTION(_sk_from_srgb_hsw)
 _sk_from_srgb_hsw:
-  .byte  196,98,125,24,5,50,53,0,0           // vbroadcastss  0x3532(%rip),%ymm8        # 48c8 <_sk_callback_hsw+0x1db>
+  .byte  196,98,125,24,5,158,53,0,0          // vbroadcastss  0x359e(%rip),%ymm8        # 4934 <_sk_callback_hsw+0x1db>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,36,53,0,0          // vbroadcastss  0x3524(%rip),%ymm11        # 48cc <_sk_callback_hsw+0x1df>
-  .byte  196,98,125,24,37,31,53,0,0          // vbroadcastss  0x351f(%rip),%ymm12        # 48d0 <_sk_callback_hsw+0x1e3>
+  .byte  196,98,125,24,29,144,53,0,0         // vbroadcastss  0x3590(%rip),%ymm11        # 4938 <_sk_callback_hsw+0x1df>
+  .byte  196,98,125,24,37,139,53,0,0         // vbroadcastss  0x358b(%rip),%ymm12        # 493c <_sk_callback_hsw+0x1e3>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,125,168,235                  // vfmadd213ps   %ymm11,%ymm0,%ymm13
-  .byte  196,98,125,24,53,16,53,0,0          // vbroadcastss  0x3510(%rip),%ymm14        # 48d4 <_sk_callback_hsw+0x1e7>
+  .byte  196,98,125,24,53,124,53,0,0         // vbroadcastss  0x357c(%rip),%ymm14        # 4940 <_sk_callback_hsw+0x1e7>
   .byte  196,66,45,168,238                   // vfmadd213ps   %ymm14,%ymm10,%ymm13
-  .byte  196,98,125,24,21,6,53,0,0           // vbroadcastss  0x3506(%rip),%ymm10        # 48d8 <_sk_callback_hsw+0x1eb>
+  .byte  196,98,125,24,21,114,53,0,0         // vbroadcastss  0x3572(%rip),%ymm10        # 4944 <_sk_callback_hsw+0x1eb>
   .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
@@ -10076,19 +10165,19 @@ HIDDEN _sk_to_srgb_hsw
 FUNCTION(_sk_to_srgb_hsw)
 _sk_to_srgb_hsw:
   .byte  197,124,82,200                      // vrsqrtps      %ymm0,%ymm9
-  .byte  196,98,125,24,5,170,52,0,0          // vbroadcastss  0x34aa(%rip),%ymm8        # 48dc <_sk_callback_hsw+0x1ef>
+  .byte  196,98,125,24,5,22,53,0,0           // vbroadcastss  0x3516(%rip),%ymm8        # 4948 <_sk_callback_hsw+0x1ef>
   .byte  196,65,124,89,208                   // vmulps        %ymm8,%ymm0,%ymm10
-  .byte  196,98,125,24,29,160,52,0,0         // vbroadcastss  0x34a0(%rip),%ymm11        # 48e0 <_sk_callback_hsw+0x1f3>
-  .byte  196,98,125,24,37,155,52,0,0         // vbroadcastss  0x349b(%rip),%ymm12        # 48e4 <_sk_callback_hsw+0x1f7>
+  .byte  196,98,125,24,29,12,53,0,0          // vbroadcastss  0x350c(%rip),%ymm11        # 494c <_sk_callback_hsw+0x1f3>
+  .byte  196,98,125,24,37,7,53,0,0           // vbroadcastss  0x3507(%rip),%ymm12        # 4950 <_sk_callback_hsw+0x1f7>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,53,168,235                   // vfmadd213ps   %ymm11,%ymm9,%ymm13
-  .byte  196,98,125,24,53,140,52,0,0         // vbroadcastss  0x348c(%rip),%ymm14        # 48e8 <_sk_callback_hsw+0x1fb>
+  .byte  196,98,125,24,53,248,52,0,0         // vbroadcastss  0x34f8(%rip),%ymm14        # 4954 <_sk_callback_hsw+0x1fb>
   .byte  196,66,53,168,238                   // vfmadd213ps   %ymm14,%ymm9,%ymm13
-  .byte  196,98,125,24,61,130,52,0,0         // vbroadcastss  0x3482(%rip),%ymm15        # 48ec <_sk_callback_hsw+0x1ff>
+  .byte  196,98,125,24,61,238,52,0,0         // vbroadcastss  0x34ee(%rip),%ymm15        # 4958 <_sk_callback_hsw+0x1ff>
   .byte  196,65,52,88,207                    // vaddps        %ymm15,%ymm9,%ymm9
   .byte  196,65,124,83,201                   // vrcpps        %ymm9,%ymm9
   .byte  196,65,20,89,201                    // vmulps        %ymm9,%ymm13,%ymm9
-  .byte  196,98,125,24,45,110,52,0,0         // vbroadcastss  0x346e(%rip),%ymm13        # 48f0 <_sk_callback_hsw+0x203>
+  .byte  196,98,125,24,45,218,52,0,0         // vbroadcastss  0x34da(%rip),%ymm13        # 495c <_sk_callback_hsw+0x203>
   .byte  196,193,124,194,197,1               // vcmpltps      %ymm13,%ymm0,%ymm0
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  197,124,82,201                      // vrsqrtps      %ymm1,%ymm9
@@ -10122,26 +10211,26 @@ _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  196,98,125,24,29,227,51,0,0         // vbroadcastss  0x33e3(%rip),%ymm11        # 48f4 <_sk_callback_hsw+0x207>
+  .byte  196,98,125,24,29,79,52,0,0          // vbroadcastss  0x344f(%rip),%ymm11        # 4960 <_sk_callback_hsw+0x207>
   .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  196,98,125,24,53,208,51,0,0         // vbroadcastss  0x33d0(%rip),%ymm14        # 48f8 <_sk_callback_hsw+0x20b>
+  .byte  196,98,125,24,53,60,52,0,0          // vbroadcastss  0x343c(%rip),%ymm14        # 4964 <_sk_callback_hsw+0x20b>
   .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  196,98,125,24,53,183,51,0,0         // vbroadcastss  0x33b7(%rip),%ymm14        # 4900 <_sk_callback_hsw+0x213>
+  .byte  196,98,125,24,53,35,52,0,0          // vbroadcastss  0x3423(%rip),%ymm14        # 496c <_sk_callback_hsw+0x213>
   .byte  196,66,37,168,238                   // vfmadd213ps   %ymm14,%ymm11,%ymm13
-  .byte  196,98,125,24,53,165,51,0,0         // vbroadcastss  0x33a5(%rip),%ymm14        # 48fc <_sk_callback_hsw+0x20f>
+  .byte  196,98,125,24,53,17,52,0,0          // vbroadcastss  0x3411(%rip),%ymm14        # 4968 <_sk_callback_hsw+0x20f>
   .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  196,98,125,24,29,136,51,0,0         // vbroadcastss  0x3388(%rip),%ymm11        # 4908 <_sk_callback_hsw+0x21b>
+  .byte  196,98,125,24,29,244,51,0,0         // vbroadcastss  0x33f4(%rip),%ymm11        # 4974 <_sk_callback_hsw+0x21b>
   .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
@@ -10151,7 +10240,7 @@ _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  196,98,125,24,5,75,51,0,0           // vbroadcastss  0x334b(%rip),%ymm8        # 4904 <_sk_callback_hsw+0x217>
+  .byte  196,98,125,24,5,183,51,0,0          // vbroadcastss  0x33b7(%rip),%ymm8        # 4970 <_sk_callback_hsw+0x217>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -10168,30 +10257,30 @@ _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  196,98,125,24,5,24,51,0,0           // vbroadcastss  0x3318(%rip),%ymm8        # 490c <_sk_callback_hsw+0x21f>
+  .byte  196,98,125,24,5,132,51,0,0          // vbroadcastss  0x3384(%rip),%ymm8        # 4978 <_sk_callback_hsw+0x21f>
   .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  196,98,125,24,13,251,50,0,0         // vbroadcastss  0x32fb(%rip),%ymm9        # 4910 <_sk_callback_hsw+0x223>
+  .byte  196,98,125,24,13,103,51,0,0         // vbroadcastss  0x3367(%rip),%ymm9        # 497c <_sk_callback_hsw+0x223>
   .byte  196,66,109,170,202                  // vfmsub213ps   %ymm10,%ymm2,%ymm9
-  .byte  196,98,125,24,29,241,50,0,0         // vbroadcastss  0x32f1(%rip),%ymm11        # 4914 <_sk_callback_hsw+0x227>
+  .byte  196,98,125,24,29,93,51,0,0          // vbroadcastss  0x335d(%rip),%ymm11        # 4980 <_sk_callback_hsw+0x227>
   .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  196,98,125,24,45,219,50,0,0         // vbroadcastss  0x32db(%rip),%ymm13        # 491c <_sk_callback_hsw+0x22f>
+  .byte  196,98,125,24,45,71,51,0,0          // vbroadcastss  0x3347(%rip),%ymm13        # 4988 <_sk_callback_hsw+0x22f>
   .byte  196,193,4,89,197                    // vmulps        %ymm13,%ymm15,%ymm0
-  .byte  196,98,125,24,53,209,50,0,0         // vbroadcastss  0x32d1(%rip),%ymm14        # 4920 <_sk_callback_hsw+0x233>
+  .byte  196,98,125,24,53,61,51,0,0          // vbroadcastss  0x333d(%rip),%ymm14        # 498c <_sk_callback_hsw+0x233>
   .byte  197,12,92,224                       // vsubps        %ymm0,%ymm14,%ymm12
   .byte  196,66,37,168,225                   // vfmadd213ps   %ymm9,%ymm11,%ymm12
-  .byte  196,226,125,24,29,183,50,0,0        // vbroadcastss  0x32b7(%rip),%ymm3        # 4918 <_sk_callback_hsw+0x22b>
+  .byte  196,226,125,24,29,35,51,0,0         // vbroadcastss  0x3323(%rip),%ymm3        # 4984 <_sk_callback_hsw+0x22b>
   .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  196,98,125,24,37,162,50,0,0         // vbroadcastss  0x32a2(%rip),%ymm12        # 4924 <_sk_callback_hsw+0x237>
+  .byte  196,98,125,24,37,14,51,0,0          // vbroadcastss  0x330e(%rip),%ymm12        # 4990 <_sk_callback_hsw+0x237>
   .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
@@ -10207,7 +10296,7 @@ _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  196,226,125,24,5,78,50,0,0          // vbroadcastss  0x324e(%rip),%ymm0        # 4928 <_sk_callback_hsw+0x23b>
+  .byte  196,226,125,24,5,186,50,0,0         // vbroadcastss  0x32ba(%rip),%ymm0        # 4994 <_sk_callback_hsw+0x23b>
   .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
@@ -10261,7 +10350,7 @@ _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  196,98,125,24,13,142,49,0,0         // vbroadcastss  0x318e(%rip),%ymm9        # 492c <_sk_callback_hsw+0x23f>
+  .byte  196,98,125,24,13,250,49,0,0         // vbroadcastss  0x31fa(%rip),%ymm9        # 4998 <_sk_callback_hsw+0x23f>
   .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
@@ -10313,7 +10402,7 @@ _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  196,98,125,24,13,251,48,0,0         // vbroadcastss  0x30fb(%rip),%ymm9        # 4930 <_sk_callback_hsw+0x243>
+  .byte  196,98,125,24,13,103,49,0,0         // vbroadcastss  0x3167(%rip),%ymm9        # 499c <_sk_callback_hsw+0x243>
   .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
@@ -10349,20 +10438,20 @@ _sk_lerp_565_hsw:
   .byte  15,133,169,0,0,0                    // jne           1941 <_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  196,98,125,88,13,136,48,0,0         // vpbroadcastd  0x3088(%rip),%ymm9        # 4934 <_sk_callback_hsw+0x247>
+  .byte  196,98,125,88,13,244,48,0,0         // vpbroadcastd  0x30f4(%rip),%ymm9        # 49a0 <_sk_callback_hsw+0x247>
   .byte  196,65,61,219,201                   // vpand         %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,121,48,0,0         // vbroadcastss  0x3079(%rip),%ymm10        # 4938 <_sk_callback_hsw+0x24b>
+  .byte  196,98,125,24,21,229,48,0,0         // vbroadcastss  0x30e5(%rip),%ymm10        # 49a4 <_sk_callback_hsw+0x24b>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,88,21,111,48,0,0         // vpbroadcastd  0x306f(%rip),%ymm10        # 493c <_sk_callback_hsw+0x24f>
+  .byte  196,98,125,88,21,219,48,0,0         // vpbroadcastd  0x30db(%rip),%ymm10        # 49a8 <_sk_callback_hsw+0x24f>
   .byte  196,65,61,219,210                   // vpand         %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,96,48,0,0          // vbroadcastss  0x3060(%rip),%ymm11        # 4940 <_sk_callback_hsw+0x253>
+  .byte  196,98,125,24,29,204,48,0,0         // vbroadcastss  0x30cc(%rip),%ymm11        # 49ac <_sk_callback_hsw+0x253>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,88,29,86,48,0,0          // vpbroadcastd  0x3056(%rip),%ymm11        # 4944 <_sk_callback_hsw+0x257>
+  .byte  196,98,125,88,29,194,48,0,0         // vpbroadcastd  0x30c2(%rip),%ymm11        # 49b0 <_sk_callback_hsw+0x257>
   .byte  196,65,61,219,195                   // vpand         %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,71,48,0,0          // vbroadcastss  0x3047(%rip),%ymm11        # 4948 <_sk_callback_hsw+0x25b>
+  .byte  196,98,125,24,29,179,48,0,0         // vbroadcastss  0x30b3(%rip),%ymm11        # 49b4 <_sk_callback_hsw+0x25b>
   .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
@@ -10433,21 +10522,21 @@ _sk_load_tables_hsw:
   .byte  77,133,192                          // test          %r8,%r8
   .byte  117,105                             // jne           1a4a <_sk_load_tables_hsw+0x7e>
   .byte  196,193,126,111,25                  // vmovdqu       (%r9),%ymm3
-  .byte  197,229,219,13,18,50,0,0            // vpand         0x3212(%rip),%ymm3,%ymm1        # 4c00 <_sk_callback_hsw+0x513>
+  .byte  197,229,219,13,146,50,0,0           // vpand         0x3292(%rip),%ymm3,%ymm1        # 4c80 <_sk_callback_hsw+0x527>
   .byte  196,65,61,118,192                   // vpcmpeqd      %ymm8,%ymm8,%ymm8
   .byte  72,139,72,8                         // mov           0x8(%rax),%rcx
   .byte  76,139,72,16                        // mov           0x10(%rax),%r9
   .byte  197,237,118,210                     // vpcmpeqd      %ymm2,%ymm2,%ymm2
   .byte  196,226,109,146,4,137               // vgatherdps    %ymm2,(%rcx,%ymm1,4),%ymm0
-  .byte  196,226,101,0,21,18,50,0,0          // vpshufb       0x3212(%rip),%ymm3,%ymm2        # 4c20 <_sk_callback_hsw+0x533>
+  .byte  196,226,101,0,21,146,50,0,0         // vpshufb       0x3292(%rip),%ymm3,%ymm2        # 4ca0 <_sk_callback_hsw+0x547>
   .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
   .byte  196,194,53,146,12,145               // vgatherdps    %ymm9,(%r9,%ymm2,4),%ymm1
   .byte  72,139,64,24                        // mov           0x18(%rax),%rax
-  .byte  196,98,101,0,13,26,50,0,0           // vpshufb       0x321a(%rip),%ymm3,%ymm9        # 4c40 <_sk_callback_hsw+0x553>
+  .byte  196,98,101,0,13,154,50,0,0          // vpshufb       0x329a(%rip),%ymm3,%ymm9        # 4cc0 <_sk_callback_hsw+0x567>
   .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  196,98,125,24,5,14,47,0,0           // vbroadcastss  0x2f0e(%rip),%ymm8        # 494c <_sk_callback_hsw+0x25f>
+  .byte  196,98,125,24,5,122,47,0,0          // vbroadcastss  0x2f7a(%rip),%ymm8        # 49b8 <_sk_callback_hsw+0x25f>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,137,193                          // mov           %r8,%rcx
@@ -10486,7 +10575,7 @@ _sk_load_tables_u16_be_hsw:
   .byte  197,185,108,200                     // vpunpcklqdq   %xmm0,%xmm8,%xmm1
   .byte  197,185,109,208                     // vpunpckhqdq   %xmm0,%xmm8,%xmm2
   .byte  197,49,108,195                      // vpunpcklqdq   %xmm3,%xmm9,%xmm8
-  .byte  197,121,111,21,166,50,0,0           // vmovdqa       0x32a6(%rip),%xmm10        # 4d80 <_sk_callback_hsw+0x693>
+  .byte  197,121,111,21,38,51,0,0            // vmovdqa       0x3326(%rip),%xmm10        # 4e00 <_sk_callback_hsw+0x6a7>
   .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
@@ -10508,7 +10597,7 @@ _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  196,98,125,24,5,7,46,0,0            // vbroadcastss  0x2e07(%rip),%ymm8        # 4950 <_sk_callback_hsw+0x263>
+  .byte  196,98,125,24,5,115,46,0,0          // vbroadcastss  0x2e73(%rip),%ymm8        # 49bc <_sk_callback_hsw+0x263>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -10568,7 +10657,7 @@ _sk_load_tables_rgb_u16_be_hsw:
   .byte  197,185,108,218                     // vpunpcklqdq   %xmm2,%xmm8,%xmm3
   .byte  197,185,109,210                     // vpunpckhqdq   %xmm2,%xmm8,%xmm2
   .byte  197,121,108,193                     // vpunpcklqdq   %xmm1,%xmm0,%xmm8
-  .byte  197,121,111,13,70,49,0,0            // vmovdqa       0x3146(%rip),%xmm9        # 4d90 <_sk_callback_hsw+0x6a3>
+  .byte  197,121,111,13,198,49,0,0           // vmovdqa       0x31c6(%rip),%xmm9        # 4e10 <_sk_callback_hsw+0x6b7>
   .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
@@ -10585,7 +10674,7 @@ _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  196,226,125,24,29,181,44,0,0        // vbroadcastss  0x2cb5(%rip),%ymm3        # 4954 <_sk_callback_hsw+0x267>
+  .byte  196,226,125,24,29,33,45,0,0         // vbroadcastss  0x2d21(%rip),%ymm3        # 49c0 <_sk_callback_hsw+0x267>
   .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
@@ -10632,7 +10721,7 @@ _sk_byte_tables_hsw:
   .byte  65,84                               // push          %r12
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,243,43,0,0          // vbroadcastss  0x2bf3(%rip),%ymm8        # 4958 <_sk_callback_hsw+0x26b>
+  .byte  196,98,125,24,5,95,44,0,0           // vbroadcastss  0x2c5f(%rip),%ymm8        # 49c4 <_sk_callback_hsw+0x26b>
   .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
@@ -10669,7 +10758,7 @@ _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  196,98,125,24,13,68,43,0,0          // vbroadcastss  0x2b44(%rip),%ymm9        # 495c <_sk_callback_hsw+0x26f>
+  .byte  196,98,125,24,13,176,43,0,0         // vbroadcastss  0x2bb0(%rip),%ymm9        # 49c8 <_sk_callback_hsw+0x26f>
   .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
@@ -10830,7 +10919,7 @@ _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  196,98,125,24,13,125,40,0,0         // vbroadcastss  0x287d(%rip),%ymm9        # 4960 <_sk_callback_hsw+0x273>
+  .byte  196,98,125,24,13,233,40,0,0         // vbroadcastss  0x28e9(%rip),%ymm9        # 49cc <_sk_callback_hsw+0x273>
   .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
@@ -10993,33 +11082,33 @@ _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  196,98,125,24,37,48,38,0,0          // vbroadcastss  0x2630(%rip),%ymm12        # 4964 <_sk_callback_hsw+0x277>
-  .byte  196,98,125,24,45,43,38,0,0          // vbroadcastss  0x262b(%rip),%ymm13        # 4968 <_sk_callback_hsw+0x27b>
+  .byte  196,98,125,24,37,156,38,0,0         // vbroadcastss  0x269c(%rip),%ymm12        # 49d0 <_sk_callback_hsw+0x277>
+  .byte  196,98,125,24,45,151,38,0,0         // vbroadcastss  0x2697(%rip),%ymm13        # 49d4 <_sk_callback_hsw+0x27b>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,33,38,0,0          // vbroadcastss  0x2621(%rip),%ymm13        # 496c <_sk_callback_hsw+0x27f>
+  .byte  196,98,125,24,45,141,38,0,0         // vbroadcastss  0x268d(%rip),%ymm13        # 49d8 <_sk_callback_hsw+0x27f>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,23,38,0,0          // vbroadcastss  0x2617(%rip),%ymm13        # 4970 <_sk_callback_hsw+0x283>
+  .byte  196,98,125,24,45,131,38,0,0         // vbroadcastss  0x2683(%rip),%ymm13        # 49dc <_sk_callback_hsw+0x283>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,13,38,0,0          // vbroadcastss  0x260d(%rip),%ymm11        # 4974 <_sk_callback_hsw+0x287>
+  .byte  196,98,125,24,29,121,38,0,0         // vbroadcastss  0x2679(%rip),%ymm11        # 49e0 <_sk_callback_hsw+0x287>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,3,38,0,0           // vbroadcastss  0x2603(%rip),%ymm12        # 4978 <_sk_callback_hsw+0x28b>
+  .byte  196,98,125,24,37,111,38,0,0         // vbroadcastss  0x266f(%rip),%ymm12        # 49e4 <_sk_callback_hsw+0x28b>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,249,37,0,0         // vbroadcastss  0x25f9(%rip),%ymm12        # 497c <_sk_callback_hsw+0x28f>
+  .byte  196,98,125,24,37,101,38,0,0         // vbroadcastss  0x2665(%rip),%ymm12        # 49e8 <_sk_callback_hsw+0x28f>
   .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  196,98,125,24,29,218,37,0,0         // vbroadcastss  0x25da(%rip),%ymm11        # 4980 <_sk_callback_hsw+0x293>
+  .byte  196,98,125,24,29,70,38,0,0          // vbroadcastss  0x2646(%rip),%ymm11        # 49ec <_sk_callback_hsw+0x293>
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,208,37,0,0         // vbroadcastss  0x25d0(%rip),%ymm11        # 4984 <_sk_callback_hsw+0x297>
+  .byte  196,98,125,24,29,60,38,0,0          // vbroadcastss  0x263c(%rip),%ymm11        # 49f0 <_sk_callback_hsw+0x297>
   .byte  196,98,45,172,216                   // vfnmadd213ps  %ymm0,%ymm10,%ymm11
-  .byte  196,226,125,24,5,198,37,0,0         // vbroadcastss  0x25c6(%rip),%ymm0        # 4988 <_sk_callback_hsw+0x29b>
+  .byte  196,226,125,24,5,50,38,0,0          // vbroadcastss  0x2632(%rip),%ymm0        # 49f4 <_sk_callback_hsw+0x29b>
   .byte  196,193,124,92,194                  // vsubps        %ymm10,%ymm0,%ymm0
-  .byte  196,98,125,24,21,188,37,0,0         // vbroadcastss  0x25bc(%rip),%ymm10        # 498c <_sk_callback_hsw+0x29f>
+  .byte  196,98,125,24,21,40,38,0,0          // vbroadcastss  0x2628(%rip),%ymm10        # 49f8 <_sk_callback_hsw+0x29f>
   .byte  197,172,94,192                      // vdivps        %ymm0,%ymm10,%ymm0
   .byte  197,164,88,192                      // vaddps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,21,175,37,0,0         // vbroadcastss  0x25af(%rip),%ymm10        # 4990 <_sk_callback_hsw+0x2a3>
+  .byte  196,98,125,24,21,27,38,0,0          // vbroadcastss  0x261b(%rip),%ymm10        # 49fc <_sk_callback_hsw+0x2a3>
   .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
@@ -11027,7 +11116,7 @@ _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  196,98,125,24,5,134,37,0,0          // vbroadcastss  0x2586(%rip),%ymm8        # 4994 <_sk_callback_hsw+0x2a7>
+  .byte  196,98,125,24,5,242,37,0,0          // vbroadcastss  0x25f2(%rip),%ymm8        # 4a00 <_sk_callback_hsw+0x2a7>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11047,33 +11136,33 @@ _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  196,98,125,24,37,62,37,0,0          // vbroadcastss  0x253e(%rip),%ymm12        # 4998 <_sk_callback_hsw+0x2ab>
-  .byte  196,98,125,24,45,57,37,0,0          // vbroadcastss  0x2539(%rip),%ymm13        # 499c <_sk_callback_hsw+0x2af>
+  .byte  196,98,125,24,37,170,37,0,0         // vbroadcastss  0x25aa(%rip),%ymm12        # 4a04 <_sk_callback_hsw+0x2ab>
+  .byte  196,98,125,24,45,165,37,0,0         // vbroadcastss  0x25a5(%rip),%ymm13        # 4a08 <_sk_callback_hsw+0x2af>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,47,37,0,0          // vbroadcastss  0x252f(%rip),%ymm13        # 49a0 <_sk_callback_hsw+0x2b3>
+  .byte  196,98,125,24,45,155,37,0,0         // vbroadcastss  0x259b(%rip),%ymm13        # 4a0c <_sk_callback_hsw+0x2b3>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,37,37,0,0          // vbroadcastss  0x2525(%rip),%ymm13        # 49a4 <_sk_callback_hsw+0x2b7>
+  .byte  196,98,125,24,45,145,37,0,0         // vbroadcastss  0x2591(%rip),%ymm13        # 4a10 <_sk_callback_hsw+0x2b7>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,27,37,0,0          // vbroadcastss  0x251b(%rip),%ymm11        # 49a8 <_sk_callback_hsw+0x2bb>
+  .byte  196,98,125,24,29,135,37,0,0         // vbroadcastss  0x2587(%rip),%ymm11        # 4a14 <_sk_callback_hsw+0x2bb>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,17,37,0,0          // vbroadcastss  0x2511(%rip),%ymm12        # 49ac <_sk_callback_hsw+0x2bf>
+  .byte  196,98,125,24,37,125,37,0,0         // vbroadcastss  0x257d(%rip),%ymm12        # 4a18 <_sk_callback_hsw+0x2bf>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,7,37,0,0           // vbroadcastss  0x2507(%rip),%ymm12        # 49b0 <_sk_callback_hsw+0x2c3>
+  .byte  196,98,125,24,37,115,37,0,0         // vbroadcastss  0x2573(%rip),%ymm12        # 4a1c <_sk_callback_hsw+0x2c3>
   .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  196,98,125,24,29,232,36,0,0         // vbroadcastss  0x24e8(%rip),%ymm11        # 49b4 <_sk_callback_hsw+0x2c7>
+  .byte  196,98,125,24,29,84,37,0,0          // vbroadcastss  0x2554(%rip),%ymm11        # 4a20 <_sk_callback_hsw+0x2c7>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,222,36,0,0         // vbroadcastss  0x24de(%rip),%ymm11        # 49b8 <_sk_callback_hsw+0x2cb>
+  .byte  196,98,125,24,29,74,37,0,0          // vbroadcastss  0x254a(%rip),%ymm11        # 4a24 <_sk_callback_hsw+0x2cb>
   .byte  196,98,45,172,217                   // vfnmadd213ps  %ymm1,%ymm10,%ymm11
-  .byte  196,226,125,24,13,212,36,0,0        // vbroadcastss  0x24d4(%rip),%ymm1        # 49bc <_sk_callback_hsw+0x2cf>
+  .byte  196,226,125,24,13,64,37,0,0         // vbroadcastss  0x2540(%rip),%ymm1        # 4a28 <_sk_callback_hsw+0x2cf>
   .byte  196,193,116,92,202                  // vsubps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,21,202,36,0,0         // vbroadcastss  0x24ca(%rip),%ymm10        # 49c0 <_sk_callback_hsw+0x2d3>
+  .byte  196,98,125,24,21,54,37,0,0          // vbroadcastss  0x2536(%rip),%ymm10        # 4a2c <_sk_callback_hsw+0x2d3>
   .byte  197,172,94,201                      // vdivps        %ymm1,%ymm10,%ymm1
   .byte  197,164,88,201                      // vaddps        %ymm1,%ymm11,%ymm1
-  .byte  196,98,125,24,21,189,36,0,0         // vbroadcastss  0x24bd(%rip),%ymm10        # 49c4 <_sk_callback_hsw+0x2d7>
+  .byte  196,98,125,24,21,41,37,0,0          // vbroadcastss  0x2529(%rip),%ymm10        # 4a30 <_sk_callback_hsw+0x2d7>
   .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
@@ -11081,7 +11170,7 @@ _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  196,98,125,24,5,148,36,0,0          // vbroadcastss  0x2494(%rip),%ymm8        # 49c8 <_sk_callback_hsw+0x2db>
+  .byte  196,98,125,24,5,0,37,0,0            // vbroadcastss  0x2500(%rip),%ymm8        # 4a34 <_sk_callback_hsw+0x2db>
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11101,33 +11190,33 @@ _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  196,98,125,24,37,76,36,0,0          // vbroadcastss  0x244c(%rip),%ymm12        # 49cc <_sk_callback_hsw+0x2df>
-  .byte  196,98,125,24,45,71,36,0,0          // vbroadcastss  0x2447(%rip),%ymm13        # 49d0 <_sk_callback_hsw+0x2e3>
+  .byte  196,98,125,24,37,184,36,0,0         // vbroadcastss  0x24b8(%rip),%ymm12        # 4a38 <_sk_callback_hsw+0x2df>
+  .byte  196,98,125,24,45,179,36,0,0         // vbroadcastss  0x24b3(%rip),%ymm13        # 4a3c <_sk_callback_hsw+0x2e3>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,61,36,0,0          // vbroadcastss  0x243d(%rip),%ymm13        # 49d4 <_sk_callback_hsw+0x2e7>
+  .byte  196,98,125,24,45,169,36,0,0         // vbroadcastss  0x24a9(%rip),%ymm13        # 4a40 <_sk_callback_hsw+0x2e7>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,51,36,0,0          // vbroadcastss  0x2433(%rip),%ymm13        # 49d8 <_sk_callback_hsw+0x2eb>
+  .byte  196,98,125,24,45,159,36,0,0         // vbroadcastss  0x249f(%rip),%ymm13        # 4a44 <_sk_callback_hsw+0x2eb>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,41,36,0,0          // vbroadcastss  0x2429(%rip),%ymm11        # 49dc <_sk_callback_hsw+0x2ef>
+  .byte  196,98,125,24,29,149,36,0,0         // vbroadcastss  0x2495(%rip),%ymm11        # 4a48 <_sk_callback_hsw+0x2ef>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,31,36,0,0          // vbroadcastss  0x241f(%rip),%ymm12        # 49e0 <_sk_callback_hsw+0x2f3>
+  .byte  196,98,125,24,37,139,36,0,0         // vbroadcastss  0x248b(%rip),%ymm12        # 4a4c <_sk_callback_hsw+0x2f3>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,21,36,0,0          // vbroadcastss  0x2415(%rip),%ymm12        # 49e4 <_sk_callback_hsw+0x2f7>
+  .byte  196,98,125,24,37,129,36,0,0         // vbroadcastss  0x2481(%rip),%ymm12        # 4a50 <_sk_callback_hsw+0x2f7>
   .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  196,98,125,24,29,246,35,0,0         // vbroadcastss  0x23f6(%rip),%ymm11        # 49e8 <_sk_callback_hsw+0x2fb>
+  .byte  196,98,125,24,29,98,36,0,0          // vbroadcastss  0x2462(%rip),%ymm11        # 4a54 <_sk_callback_hsw+0x2fb>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,236,35,0,0         // vbroadcastss  0x23ec(%rip),%ymm11        # 49ec <_sk_callback_hsw+0x2ff>
+  .byte  196,98,125,24,29,88,36,0,0          // vbroadcastss  0x2458(%rip),%ymm11        # 4a58 <_sk_callback_hsw+0x2ff>
   .byte  196,98,45,172,218                   // vfnmadd213ps  %ymm2,%ymm10,%ymm11
-  .byte  196,226,125,24,21,226,35,0,0        // vbroadcastss  0x23e2(%rip),%ymm2        # 49f0 <_sk_callback_hsw+0x303>
+  .byte  196,226,125,24,21,78,36,0,0         // vbroadcastss  0x244e(%rip),%ymm2        # 4a5c <_sk_callback_hsw+0x303>
   .byte  196,193,108,92,210                  // vsubps        %ymm10,%ymm2,%ymm2
-  .byte  196,98,125,24,21,216,35,0,0         // vbroadcastss  0x23d8(%rip),%ymm10        # 49f4 <_sk_callback_hsw+0x307>
+  .byte  196,98,125,24,21,68,36,0,0          // vbroadcastss  0x2444(%rip),%ymm10        # 4a60 <_sk_callback_hsw+0x307>
   .byte  197,172,94,210                      // vdivps        %ymm2,%ymm10,%ymm2
   .byte  197,164,88,210                      // vaddps        %ymm2,%ymm11,%ymm2
-  .byte  196,98,125,24,21,203,35,0,0         // vbroadcastss  0x23cb(%rip),%ymm10        # 49f8 <_sk_callback_hsw+0x30b>
+  .byte  196,98,125,24,21,55,36,0,0          // vbroadcastss  0x2437(%rip),%ymm10        # 4a64 <_sk_callback_hsw+0x30b>
   .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
@@ -11135,7 +11224,7 @@ _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  196,98,125,24,5,162,35,0,0          // vbroadcastss  0x23a2(%rip),%ymm8        # 49fc <_sk_callback_hsw+0x30f>
+  .byte  196,98,125,24,5,14,36,0,0           // vbroadcastss  0x240e(%rip),%ymm8        # 4a68 <_sk_callback_hsw+0x30f>
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11155,33 +11244,33 @@ _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  196,98,125,24,37,90,35,0,0          // vbroadcastss  0x235a(%rip),%ymm12        # 4a00 <_sk_callback_hsw+0x313>
-  .byte  196,98,125,24,45,85,35,0,0          // vbroadcastss  0x2355(%rip),%ymm13        # 4a04 <_sk_callback_hsw+0x317>
+  .byte  196,98,125,24,37,198,35,0,0         // vbroadcastss  0x23c6(%rip),%ymm12        # 4a6c <_sk_callback_hsw+0x313>
+  .byte  196,98,125,24,45,193,35,0,0         // vbroadcastss  0x23c1(%rip),%ymm13        # 4a70 <_sk_callback_hsw+0x317>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,75,35,0,0          // vbroadcastss  0x234b(%rip),%ymm13        # 4a08 <_sk_callback_hsw+0x31b>
+  .byte  196,98,125,24,45,183,35,0,0         // vbroadcastss  0x23b7(%rip),%ymm13        # 4a74 <_sk_callback_hsw+0x31b>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,65,35,0,0          // vbroadcastss  0x2341(%rip),%ymm13        # 4a0c <_sk_callback_hsw+0x31f>
+  .byte  196,98,125,24,45,173,35,0,0         // vbroadcastss  0x23ad(%rip),%ymm13        # 4a78 <_sk_callback_hsw+0x31f>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,55,35,0,0          // vbroadcastss  0x2337(%rip),%ymm11        # 4a10 <_sk_callback_hsw+0x323>
+  .byte  196,98,125,24,29,163,35,0,0         // vbroadcastss  0x23a3(%rip),%ymm11        # 4a7c <_sk_callback_hsw+0x323>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,45,35,0,0          // vbroadcastss  0x232d(%rip),%ymm12        # 4a14 <_sk_callback_hsw+0x327>
+  .byte  196,98,125,24,37,153,35,0,0         // vbroadcastss  0x2399(%rip),%ymm12        # 4a80 <_sk_callback_hsw+0x327>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,35,35,0,0          // vbroadcastss  0x2323(%rip),%ymm12        # 4a18 <_sk_callback_hsw+0x32b>
+  .byte  196,98,125,24,37,143,35,0,0         // vbroadcastss  0x238f(%rip),%ymm12        # 4a84 <_sk_callback_hsw+0x32b>
   .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  196,98,125,24,29,4,35,0,0           // vbroadcastss  0x2304(%rip),%ymm11        # 4a1c <_sk_callback_hsw+0x32f>
+  .byte  196,98,125,24,29,112,35,0,0         // vbroadcastss  0x2370(%rip),%ymm11        # 4a88 <_sk_callback_hsw+0x32f>
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,250,34,0,0         // vbroadcastss  0x22fa(%rip),%ymm11        # 4a20 <_sk_callback_hsw+0x333>
+  .byte  196,98,125,24,29,102,35,0,0         // vbroadcastss  0x2366(%rip),%ymm11        # 4a8c <_sk_callback_hsw+0x333>
   .byte  196,98,45,172,219                   // vfnmadd213ps  %ymm3,%ymm10,%ymm11
-  .byte  196,226,125,24,29,240,34,0,0        // vbroadcastss  0x22f0(%rip),%ymm3        # 4a24 <_sk_callback_hsw+0x337>
+  .byte  196,226,125,24,29,92,35,0,0         // vbroadcastss  0x235c(%rip),%ymm3        # 4a90 <_sk_callback_hsw+0x337>
   .byte  196,193,100,92,218                  // vsubps        %ymm10,%ymm3,%ymm3
-  .byte  196,98,125,24,21,230,34,0,0         // vbroadcastss  0x22e6(%rip),%ymm10        # 4a28 <_sk_callback_hsw+0x33b>
+  .byte  196,98,125,24,21,82,35,0,0          // vbroadcastss  0x2352(%rip),%ymm10        # 4a94 <_sk_callback_hsw+0x33b>
   .byte  197,172,94,219                      // vdivps        %ymm3,%ymm10,%ymm3
   .byte  197,164,88,219                      // vaddps        %ymm3,%ymm11,%ymm3
-  .byte  196,98,125,24,21,217,34,0,0         // vbroadcastss  0x22d9(%rip),%ymm10        # 4a2c <_sk_callback_hsw+0x33f>
+  .byte  196,98,125,24,21,69,35,0,0          // vbroadcastss  0x2345(%rip),%ymm10        # 4a98 <_sk_callback_hsw+0x33f>
   .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
@@ -11189,7 +11278,7 @@ _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  196,98,125,24,5,176,34,0,0          // vbroadcastss  0x22b0(%rip),%ymm8        # 4a30 <_sk_callback_hsw+0x343>
+  .byte  196,98,125,24,5,28,35,0,0           // vbroadcastss  0x231c(%rip),%ymm8        # 4a9c <_sk_callback_hsw+0x343>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11198,26 +11287,26 @@ HIDDEN _sk_lab_to_xyz_hsw
 .globl _sk_lab_to_xyz_hsw
 FUNCTION(_sk_lab_to_xyz_hsw)
 _sk_lab_to_xyz_hsw:
-  .byte  196,98,125,24,5,162,34,0,0          // vbroadcastss  0x22a2(%rip),%ymm8        # 4a34 <_sk_callback_hsw+0x347>
-  .byte  196,98,125,24,13,157,34,0,0         // vbroadcastss  0x229d(%rip),%ymm9        # 4a38 <_sk_callback_hsw+0x34b>
-  .byte  196,98,125,24,21,152,34,0,0         // vbroadcastss  0x2298(%rip),%ymm10        # 4a3c <_sk_callback_hsw+0x34f>
+  .byte  196,98,125,24,5,14,35,0,0           // vbroadcastss  0x230e(%rip),%ymm8        # 4aa0 <_sk_callback_hsw+0x347>
+  .byte  196,98,125,24,13,9,35,0,0           // vbroadcastss  0x2309(%rip),%ymm9        # 4aa4 <_sk_callback_hsw+0x34b>
+  .byte  196,98,125,24,21,4,35,0,0           // vbroadcastss  0x2304(%rip),%ymm10        # 4aa8 <_sk_callback_hsw+0x34f>
   .byte  196,194,53,168,202                  // vfmadd213ps   %ymm10,%ymm9,%ymm1
   .byte  196,194,53,168,210                  // vfmadd213ps   %ymm10,%ymm9,%ymm2
-  .byte  196,98,125,24,13,137,34,0,0         // vbroadcastss  0x2289(%rip),%ymm9        # 4a40 <_sk_callback_hsw+0x353>
+  .byte  196,98,125,24,13,245,34,0,0         // vbroadcastss  0x22f5(%rip),%ymm9        # 4aac <_sk_callback_hsw+0x353>
   .byte  196,66,125,184,200                  // vfmadd231ps   %ymm8,%ymm0,%ymm9
-  .byte  196,226,125,24,5,127,34,0,0         // vbroadcastss  0x227f(%rip),%ymm0        # 4a44 <_sk_callback_hsw+0x357>
+  .byte  196,226,125,24,5,235,34,0,0         // vbroadcastss  0x22eb(%rip),%ymm0        # 4ab0 <_sk_callback_hsw+0x357>
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
-  .byte  196,98,125,24,5,118,34,0,0          // vbroadcastss  0x2276(%rip),%ymm8        # 4a48 <_sk_callback_hsw+0x35b>
+  .byte  196,98,125,24,5,226,34,0,0          // vbroadcastss  0x22e2(%rip),%ymm8        # 4ab4 <_sk_callback_hsw+0x35b>
   .byte  196,98,117,168,192                  // vfmadd213ps   %ymm0,%ymm1,%ymm8
-  .byte  196,98,125,24,13,108,34,0,0         // vbroadcastss  0x226c(%rip),%ymm9        # 4a4c <_sk_callback_hsw+0x35f>
+  .byte  196,98,125,24,13,216,34,0,0         // vbroadcastss  0x22d8(%rip),%ymm9        # 4ab8 <_sk_callback_hsw+0x35f>
   .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  196,226,125,24,21,89,34,0,0         // vbroadcastss  0x2259(%rip),%ymm2        # 4a50 <_sk_callback_hsw+0x363>
+  .byte  196,226,125,24,21,197,34,0,0        // vbroadcastss  0x22c5(%rip),%ymm2        # 4abc <_sk_callback_hsw+0x363>
   .byte  197,108,194,209,1                   // vcmpltps      %ymm1,%ymm2,%ymm10
-  .byte  196,98,125,24,29,79,34,0,0          // vbroadcastss  0x224f(%rip),%ymm11        # 4a54 <_sk_callback_hsw+0x367>
+  .byte  196,98,125,24,29,187,34,0,0         // vbroadcastss  0x22bb(%rip),%ymm11        # 4ac0 <_sk_callback_hsw+0x367>
   .byte  196,65,60,88,195                    // vaddps        %ymm11,%ymm8,%ymm8
-  .byte  196,98,125,24,37,69,34,0,0          // vbroadcastss  0x2245(%rip),%ymm12        # 4a58 <_sk_callback_hsw+0x36b>
+  .byte  196,98,125,24,37,177,34,0,0         // vbroadcastss  0x22b1(%rip),%ymm12        # 4ac4 <_sk_callback_hsw+0x36b>
   .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
@@ -11232,9 +11321,9 @@ _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  196,226,125,24,5,250,33,0,0         // vbroadcastss  0x21fa(%rip),%ymm0        # 4a5c <_sk_callback_hsw+0x36f>
+  .byte  196,226,125,24,5,102,34,0,0         // vbroadcastss  0x2266(%rip),%ymm0        # 4ac8 <_sk_callback_hsw+0x36f>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,241,33,0,0          // vbroadcastss  0x21f1(%rip),%ymm8        # 4a60 <_sk_callback_hsw+0x373>
+  .byte  196,98,125,24,5,93,34,0,0           // vbroadcastss  0x225d(%rip),%ymm8        # 4acc <_sk_callback_hsw+0x373>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11252,7 +11341,7 @@ _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  196,226,125,24,13,198,33,0,0        // vbroadcastss  0x21c6(%rip),%ymm1        # 4a64 <_sk_callback_hsw+0x377>
+  .byte  196,226,125,24,13,50,34,0,0         // vbroadcastss  0x2232(%rip),%ymm1        # 4ad0 <_sk_callback_hsw+0x377>
   .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
@@ -11317,7 +11406,7 @@ _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  196,226,125,24,13,209,32,0,0        // vbroadcastss  0x20d1(%rip),%ymm1        # 4a68 <_sk_callback_hsw+0x37b>
+  .byte  196,226,125,24,13,61,33,0,0         // vbroadcastss  0x213d(%rip),%ymm1        # 4ad4 <_sk_callback_hsw+0x37b>
   .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
@@ -11335,7 +11424,7 @@ FUNCTION(_sk_store_a8_hsw)
 _sk_store_a8_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,172,32,0,0          // vbroadcastss  0x20ac(%rip),%ymm8        # 4a6c <_sk_callback_hsw+0x37f>
+  .byte  196,98,125,24,5,24,33,0,0           // vbroadcastss  0x2118(%rip),%ymm8        # 4ad8 <_sk_callback_hsw+0x37f>
   .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
@@ -11402,10 +11491,10 @@ _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  196,226,125,24,13,226,31,0,0        // vbroadcastss  0x1fe2(%rip),%ymm1        # 4a70 <_sk_callback_hsw+0x383>
+  .byte  196,226,125,24,13,78,32,0,0         // vbroadcastss  0x204e(%rip),%ymm1        # 4adc <_sk_callback_hsw+0x383>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,215,31,0,0        // vbroadcastss  0x1fd7(%rip),%ymm3        # 4a74 <_sk_callback_hsw+0x387>
+  .byte  196,226,125,24,29,67,32,0,0         // vbroadcastss  0x2043(%rip),%ymm3        # 4ae0 <_sk_callback_hsw+0x387>
   .byte  76,137,193                          // mov           %r8,%rcx
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
@@ -11467,10 +11556,10 @@ _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  196,226,125,24,13,236,30,0,0        // vbroadcastss  0x1eec(%rip),%ymm1        # 4a78 <_sk_callback_hsw+0x38b>
+  .byte  196,226,125,24,13,88,31,0,0         // vbroadcastss  0x1f58(%rip),%ymm1        # 4ae4 <_sk_callback_hsw+0x38b>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,225,30,0,0        // vbroadcastss  0x1ee1(%rip),%ymm3        # 4a7c <_sk_callback_hsw+0x38f>
+  .byte  196,226,125,24,29,77,31,0,0         // vbroadcastss  0x1f4d(%rip),%ymm3        # 4ae8 <_sk_callback_hsw+0x38f>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -11526,14 +11615,14 @@ _sk_gather_i8_hsw:
   .byte  73,139,64,8                         // mov           0x8(%r8),%rax
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
   .byte  196,226,117,144,28,128              // vpgatherdd    %ymm1,(%rax,%ymm0,4),%ymm3
-  .byte  197,229,219,5,245,31,0,0            // vpand         0x1ff5(%rip),%ymm3,%ymm0        # 4c60 <_sk_callback_hsw+0x573>
+  .byte  197,229,219,5,117,32,0,0            // vpand         0x2075(%rip),%ymm3,%ymm0        # 4ce0 <_sk_callback_hsw+0x587>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,8,30,0,0            // vbroadcastss  0x1e08(%rip),%ymm8        # 4a80 <_sk_callback_hsw+0x393>
+  .byte  196,98,125,24,5,116,30,0,0          // vbroadcastss  0x1e74(%rip),%ymm8        # 4aec <_sk_callback_hsw+0x393>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,250,31,0,0         // vpshufb       0x1ffa(%rip),%ymm3,%ymm1        # 4c80 <_sk_callback_hsw+0x593>
+  .byte  196,226,101,0,13,122,32,0,0         // vpshufb       0x207a(%rip),%ymm3,%ymm1        # 4d00 <_sk_callback_hsw+0x5a7>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,8,32,0,0           // vpshufb       0x2008(%rip),%ymm3,%ymm2        # 4ca0 <_sk_callback_hsw+0x5b3>
+  .byte  196,226,101,0,21,136,32,0,0         // vpshufb       0x2088(%rip),%ymm3,%ymm2        # 4d20 <_sk_callback_hsw+0x5c7>
   .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
@@ -11557,23 +11646,23 @@ _sk_load_565_hsw:
   .byte  117,114                             // jne           2d38 <_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  196,226,125,88,5,170,29,0,0         // vpbroadcastd  0x1daa(%rip),%ymm0        # 4a84 <_sk_callback_hsw+0x397>
+  .byte  196,226,125,88,5,22,30,0,0          // vpbroadcastd  0x1e16(%rip),%ymm0        # 4af0 <_sk_callback_hsw+0x397>
   .byte  197,237,219,192                     // vpand         %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,157,29,0,0        // vbroadcastss  0x1d9d(%rip),%ymm1        # 4a88 <_sk_callback_hsw+0x39b>
+  .byte  196,226,125,24,13,9,30,0,0          // vbroadcastss  0x1e09(%rip),%ymm1        # 4af4 <_sk_callback_hsw+0x39b>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,148,29,0,0        // vpbroadcastd  0x1d94(%rip),%ymm1        # 4a8c <_sk_callback_hsw+0x39f>
+  .byte  196,226,125,88,13,0,30,0,0          // vpbroadcastd  0x1e00(%rip),%ymm1        # 4af8 <_sk_callback_hsw+0x39f>
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,135,29,0,0        // vbroadcastss  0x1d87(%rip),%ymm3        # 4a90 <_sk_callback_hsw+0x3a3>
+  .byte  196,226,125,24,29,243,29,0,0        // vbroadcastss  0x1df3(%rip),%ymm3        # 4afc <_sk_callback_hsw+0x3a3>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,88,29,126,29,0,0        // vpbroadcastd  0x1d7e(%rip),%ymm3        # 4a94 <_sk_callback_hsw+0x3a7>
+  .byte  196,226,125,88,29,234,29,0,0        // vpbroadcastd  0x1dea(%rip),%ymm3        # 4b00 <_sk_callback_hsw+0x3a7>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,113,29,0,0        // vbroadcastss  0x1d71(%rip),%ymm3        # 4a98 <_sk_callback_hsw+0x3ab>
+  .byte  196,226,125,24,29,221,29,0,0        // vbroadcastss  0x1ddd(%rip),%ymm3        # 4b04 <_sk_callback_hsw+0x3ab>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,102,29,0,0        // vbroadcastss  0x1d66(%rip),%ymm3        # 4a9c <_sk_callback_hsw+0x3af>
+  .byte  196,226,125,24,29,210,29,0,0        // vbroadcastss  0x1dd2(%rip),%ymm3        # 4b08 <_sk_callback_hsw+0x3af>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,200                          // mov           %ecx,%r8d
   .byte  65,128,224,7                        // and           $0x7,%r8b
@@ -11664,23 +11753,23 @@ _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  196,226,125,88,5,41,28,0,0          // vpbroadcastd  0x1c29(%rip),%ymm0        # 4aa0 <_sk_callback_hsw+0x3b3>
+  .byte  196,226,125,88,5,149,28,0,0         // vpbroadcastd  0x1c95(%rip),%ymm0        # 4b0c <_sk_callback_hsw+0x3b3>
   .byte  197,237,219,192                     // vpand         %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,28,28,0,0         // vbroadcastss  0x1c1c(%rip),%ymm1        # 4aa4 <_sk_callback_hsw+0x3b7>
+  .byte  196,226,125,24,13,136,28,0,0        // vbroadcastss  0x1c88(%rip),%ymm1        # 4b10 <_sk_callback_hsw+0x3b7>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,19,28,0,0         // vpbroadcastd  0x1c13(%rip),%ymm1        # 4aa8 <_sk_callback_hsw+0x3bb>
+  .byte  196,226,125,88,13,127,28,0,0        // vpbroadcastd  0x1c7f(%rip),%ymm1        # 4b14 <_sk_callback_hsw+0x3bb>
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,6,28,0,0          // vbroadcastss  0x1c06(%rip),%ymm3        # 4aac <_sk_callback_hsw+0x3bf>
+  .byte  196,226,125,24,29,114,28,0,0        // vbroadcastss  0x1c72(%rip),%ymm3        # 4b18 <_sk_callback_hsw+0x3bf>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,88,29,253,27,0,0        // vpbroadcastd  0x1bfd(%rip),%ymm3        # 4ab0 <_sk_callback_hsw+0x3c3>
+  .byte  196,226,125,88,29,105,28,0,0        // vpbroadcastd  0x1c69(%rip),%ymm3        # 4b1c <_sk_callback_hsw+0x3c3>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,240,27,0,0        // vbroadcastss  0x1bf0(%rip),%ymm3        # 4ab4 <_sk_callback_hsw+0x3c7>
+  .byte  196,226,125,24,29,92,28,0,0         // vbroadcastss  0x1c5c(%rip),%ymm3        # 4b20 <_sk_callback_hsw+0x3c7>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,229,27,0,0        // vbroadcastss  0x1be5(%rip),%ymm3        # 4ab8 <_sk_callback_hsw+0x3cb>
+  .byte  196,226,125,24,29,81,28,0,0         // vbroadcastss  0x1c51(%rip),%ymm3        # 4b24 <_sk_callback_hsw+0x3cb>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -11693,11 +11782,11 @@ FUNCTION(_sk_store_565_hsw)
 _sk_store_565_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,210,27,0,0          // vbroadcastss  0x1bd2(%rip),%ymm8        # 4abc <_sk_callback_hsw+0x3cf>
+  .byte  196,98,125,24,5,62,28,0,0           // vbroadcastss  0x1c3e(%rip),%ymm8        # 4b28 <_sk_callback_hsw+0x3cf>
   .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  196,98,125,24,21,189,27,0,0         // vbroadcastss  0x1bbd(%rip),%ymm10        # 4ac0 <_sk_callback_hsw+0x3d3>
+  .byte  196,98,125,24,21,41,28,0,0          // vbroadcastss  0x1c29(%rip),%ymm10        # 4b2c <_sk_callback_hsw+0x3d3>
   .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
@@ -11765,25 +11854,25 @@ _sk_load_4444_hsw:
   .byte  15,133,138,0,0,0                    // jne           3054 <_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  196,226,125,88,5,230,26,0,0         // vpbroadcastd  0x1ae6(%rip),%ymm0        # 4ac4 <_sk_callback_hsw+0x3d7>
+  .byte  196,226,125,88,5,82,27,0,0          // vpbroadcastd  0x1b52(%rip),%ymm0        # 4b30 <_sk_callback_hsw+0x3d7>
   .byte  197,229,219,192                     // vpand         %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,217,26,0,0        // vbroadcastss  0x1ad9(%rip),%ymm1        # 4ac8 <_sk_callback_hsw+0x3db>
+  .byte  196,226,125,24,13,69,27,0,0         // vbroadcastss  0x1b45(%rip),%ymm1        # 4b34 <_sk_callback_hsw+0x3db>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,208,26,0,0        // vpbroadcastd  0x1ad0(%rip),%ymm1        # 4acc <_sk_callback_hsw+0x3df>
+  .byte  196,226,125,88,13,60,27,0,0         // vpbroadcastd  0x1b3c(%rip),%ymm1        # 4b38 <_sk_callback_hsw+0x3df>
   .byte  197,229,219,201                     // vpand         %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,195,26,0,0        // vbroadcastss  0x1ac3(%rip),%ymm2        # 4ad0 <_sk_callback_hsw+0x3e3>
+  .byte  196,226,125,24,21,47,27,0,0         // vbroadcastss  0x1b2f(%rip),%ymm2        # 4b3c <_sk_callback_hsw+0x3e3>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,88,21,186,26,0,0        // vpbroadcastd  0x1aba(%rip),%ymm2        # 4ad4 <_sk_callback_hsw+0x3e7>
+  .byte  196,226,125,88,21,38,27,0,0         // vpbroadcastd  0x1b26(%rip),%ymm2        # 4b40 <_sk_callback_hsw+0x3e7>
   .byte  197,229,219,210                     // vpand         %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,173,26,0,0          // vbroadcastss  0x1aad(%rip),%ymm8        # 4ad8 <_sk_callback_hsw+0x3eb>
+  .byte  196,98,125,24,5,25,27,0,0           // vbroadcastss  0x1b19(%rip),%ymm8        # 4b44 <_sk_callback_hsw+0x3eb>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,88,5,163,26,0,0          // vpbroadcastd  0x1aa3(%rip),%ymm8        # 4adc <_sk_callback_hsw+0x3ef>
+  .byte  196,98,125,88,5,15,27,0,0           // vpbroadcastd  0x1b0f(%rip),%ymm8        # 4b48 <_sk_callback_hsw+0x3ef>
   .byte  196,193,101,219,216                 // vpand         %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,149,26,0,0          // vbroadcastss  0x1a95(%rip),%ymm8        # 4ae0 <_sk_callback_hsw+0x3f3>
+  .byte  196,98,125,24,5,1,27,0,0            // vbroadcastss  0x1b01(%rip),%ymm8        # 4b4c <_sk_callback_hsw+0x3f3>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -11876,25 +11965,25 @@ _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  196,226,125,88,5,77,25,0,0          // vpbroadcastd  0x194d(%rip),%ymm0        # 4ae4 <_sk_callback_hsw+0x3f7>
+  .byte  196,226,125,88,5,185,25,0,0         // vpbroadcastd  0x19b9(%rip),%ymm0        # 4b50 <_sk_callback_hsw+0x3f7>
   .byte  197,229,219,192                     // vpand         %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,64,25,0,0         // vbroadcastss  0x1940(%rip),%ymm1        # 4ae8 <_sk_callback_hsw+0x3fb>
+  .byte  196,226,125,24,13,172,25,0,0        // vbroadcastss  0x19ac(%rip),%ymm1        # 4b54 <_sk_callback_hsw+0x3fb>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,55,25,0,0         // vpbroadcastd  0x1937(%rip),%ymm1        # 4aec <_sk_callback_hsw+0x3ff>
+  .byte  196,226,125,88,13,163,25,0,0        // vpbroadcastd  0x19a3(%rip),%ymm1        # 4b58 <_sk_callback_hsw+0x3ff>
   .byte  197,229,219,201                     // vpand         %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,42,25,0,0         // vbroadcastss  0x192a(%rip),%ymm2        # 4af0 <_sk_callback_hsw+0x403>
+  .byte  196,226,125,24,21,150,25,0,0        // vbroadcastss  0x1996(%rip),%ymm2        # 4b5c <_sk_callback_hsw+0x403>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,88,21,33,25,0,0         // vpbroadcastd  0x1921(%rip),%ymm2        # 4af4 <_sk_callback_hsw+0x407>
+  .byte  196,226,125,88,21,141,25,0,0        // vpbroadcastd  0x198d(%rip),%ymm2        # 4b60 <_sk_callback_hsw+0x407>
   .byte  197,229,219,210                     // vpand         %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,20,25,0,0           // vbroadcastss  0x1914(%rip),%ymm8        # 4af8 <_sk_callback_hsw+0x40b>
+  .byte  196,98,125,24,5,128,25,0,0          // vbroadcastss  0x1980(%rip),%ymm8        # 4b64 <_sk_callback_hsw+0x40b>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,88,5,10,25,0,0           // vpbroadcastd  0x190a(%rip),%ymm8        # 4afc <_sk_callback_hsw+0x40f>
+  .byte  196,98,125,88,5,118,25,0,0          // vpbroadcastd  0x1976(%rip),%ymm8        # 4b68 <_sk_callback_hsw+0x40f>
   .byte  196,193,101,219,216                 // vpand         %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,252,24,0,0          // vbroadcastss  0x18fc(%rip),%ymm8        # 4b00 <_sk_callback_hsw+0x413>
+  .byte  196,98,125,24,5,104,25,0,0          // vbroadcastss  0x1968(%rip),%ymm8        # 4b6c <_sk_callback_hsw+0x413>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -11909,7 +11998,7 @@ FUNCTION(_sk_store_4444_hsw)
 _sk_store_4444_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,226,24,0,0          // vbroadcastss  0x18e2(%rip),%ymm8        # 4b04 <_sk_callback_hsw+0x417>
+  .byte  196,98,125,24,5,78,25,0,0           // vbroadcastss  0x194e(%rip),%ymm8        # 4b70 <_sk_callback_hsw+0x417>
   .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
@@ -11985,14 +12074,14 @@ _sk_load_8888_hsw:
   .byte  77,133,192                          // test          %r8,%r8
   .byte  117,88                              // jne           336d <_sk_load_8888_hsw+0x6d>
   .byte  196,193,126,111,25                  // vmovdqu       (%r9),%ymm3
-  .byte  197,229,219,5,158,25,0,0            // vpand         0x199e(%rip),%ymm3,%ymm0        # 4cc0 <_sk_callback_hsw+0x5d3>
+  .byte  197,229,219,5,30,26,0,0             // vpand         0x1a1e(%rip),%ymm3,%ymm0        # 4d40 <_sk_callback_hsw+0x5e7>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,217,23,0,0          // vbroadcastss  0x17d9(%rip),%ymm8        # 4b08 <_sk_callback_hsw+0x41b>
+  .byte  196,98,125,24,5,69,24,0,0           // vbroadcastss  0x1845(%rip),%ymm8        # 4b74 <_sk_callback_hsw+0x41b>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,163,25,0,0         // vpshufb       0x19a3(%rip),%ymm3,%ymm1        # 4ce0 <_sk_callback_hsw+0x5f3>
+  .byte  196,226,101,0,13,35,26,0,0          // vpshufb       0x1a23(%rip),%ymm3,%ymm1        # 4d60 <_sk_callback_hsw+0x607>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,177,25,0,0         // vpshufb       0x19b1(%rip),%ymm3,%ymm2        # 4d00 <_sk_callback_hsw+0x613>
+  .byte  196,226,101,0,21,49,26,0,0          // vpshufb       0x1a31(%rip),%ymm3,%ymm2        # 4d80 <_sk_callback_hsw+0x627>
   .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
@@ -12024,14 +12113,14 @@ _sk_gather_8888_hsw:
   .byte  197,245,254,192                     // vpaddd        %ymm0,%ymm1,%ymm0
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
   .byte  196,194,117,144,28,128              // vpgatherdd    %ymm1,(%r8,%ymm0,4),%ymm3
-  .byte  197,229,219,5,95,25,0,0             // vpand         0x195f(%rip),%ymm3,%ymm0        # 4d20 <_sk_callback_hsw+0x633>
+  .byte  197,229,219,5,223,25,0,0            // vpand         0x19df(%rip),%ymm3,%ymm0        # 4da0 <_sk_callback_hsw+0x647>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,62,23,0,0           // vbroadcastss  0x173e(%rip),%ymm8        # 4b0c <_sk_callback_hsw+0x41f>
+  .byte  196,98,125,24,5,170,23,0,0          // vbroadcastss  0x17aa(%rip),%ymm8        # 4b78 <_sk_callback_hsw+0x41f>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,100,25,0,0         // vpshufb       0x1964(%rip),%ymm3,%ymm1        # 4d40 <_sk_callback_hsw+0x653>
+  .byte  196,226,101,0,13,228,25,0,0         // vpshufb       0x19e4(%rip),%ymm3,%ymm1        # 4dc0 <_sk_callback_hsw+0x667>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,114,25,0,0         // vpshufb       0x1972(%rip),%ymm3,%ymm2        # 4d60 <_sk_callback_hsw+0x673>
+  .byte  196,226,101,0,21,242,25,0,0         // vpshufb       0x19f2(%rip),%ymm3,%ymm2        # 4de0 <_sk_callback_hsw+0x687>
   .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
@@ -12048,7 +12137,7 @@ _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  196,98,125,24,5,238,22,0,0          // vbroadcastss  0x16ee(%rip),%ymm8        # 4b10 <_sk_callback_hsw+0x423>
+  .byte  196,98,125,24,5,90,23,0,0           // vbroadcastss  0x175a(%rip),%ymm8        # 4b7c <_sk_callback_hsw+0x423>
   .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
@@ -12245,7 +12334,7 @@ _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  196,98,125,24,21,229,19,0,0         // vbroadcastss  0x13e5(%rip),%ymm10        # 4b14 <_sk_callback_hsw+0x427>
+  .byte  196,98,125,24,21,81,20,0,0          // vbroadcastss  0x1451(%rip),%ymm10        # 4b80 <_sk_callback_hsw+0x427>
   .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
@@ -12329,7 +12418,7 @@ _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  196,98,125,24,21,118,18,0,0         // vbroadcastss  0x1276(%rip),%ymm10        # 4b18 <_sk_callback_hsw+0x42b>
+  .byte  196,98,125,24,21,226,18,0,0         // vbroadcastss  0x12e2(%rip),%ymm10        # 4b84 <_sk_callback_hsw+0x42b>
   .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
@@ -12346,7 +12435,7 @@ _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  196,226,125,24,29,42,18,0,0         // vbroadcastss  0x122a(%rip),%ymm3        # 4b1c <_sk_callback_hsw+0x42f>
+  .byte  196,226,125,24,29,150,18,0,0        // vbroadcastss  0x1296(%rip),%ymm3        # 4b88 <_sk_callback_hsw+0x42f>
   .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
@@ -12389,7 +12478,7 @@ _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  196,98,125,24,5,103,17,0,0          // vbroadcastss  0x1167(%rip),%ymm8        # 4b20 <_sk_callback_hsw+0x433>
+  .byte  196,98,125,24,5,211,17,0,0          // vbroadcastss  0x11d3(%rip),%ymm8        # 4b8c <_sk_callback_hsw+0x433>
   .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
@@ -12645,15 +12734,53 @@ _sk_mirror_y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_clamp_x_1_hsw
+.globl _sk_clamp_x_1_hsw
+FUNCTION(_sk_clamp_x_1_hsw)
+_sk_clamp_x_1_hsw:
+  .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
+  .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
+  .byte  196,98,125,24,5,26,14,0,0           // vbroadcastss  0xe1a(%rip),%ymm8        # 4b90 <_sk_callback_hsw+0x437>
+  .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_repeat_x_1_hsw
+.globl _sk_repeat_x_1_hsw
+FUNCTION(_sk_repeat_x_1_hsw)
+_sk_repeat_x_1_hsw:
+  .byte  196,99,125,8,192,1                  // vroundps      $0x1,%ymm0,%ymm8
+  .byte  196,193,124,92,192                  // vsubps        %ymm8,%ymm0,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_mirror_x_1_hsw
+.globl _sk_mirror_x_1_hsw
+FUNCTION(_sk_mirror_x_1_hsw)
+_sk_mirror_x_1_hsw:
+  .byte  196,98,125,24,5,253,13,0,0          // vbroadcastss  0xdfd(%rip),%ymm8        # 4b94 <_sk_callback_hsw+0x43b>
+  .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
+  .byte  196,98,125,24,13,243,13,0,0         // vbroadcastss  0xdf3(%rip),%ymm9        # 4b98 <_sk_callback_hsw+0x43f>
+  .byte  196,65,124,89,201                   // vmulps        %ymm9,%ymm0,%ymm9
+  .byte  196,67,125,8,201,1                  // vroundps      $0x1,%ymm9,%ymm9
+  .byte  196,65,52,88,201                    // vaddps        %ymm9,%ymm9,%ymm9
+  .byte  196,193,124,92,193                  // vsubps        %ymm9,%ymm0,%ymm0
+  .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
+  .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
+  .byte  197,60,92,192                       // vsubps        %ymm0,%ymm8,%ymm8
+  .byte  197,188,84,192                      // vandps        %ymm0,%ymm8,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_luminance_to_alpha_hsw
 .globl _sk_luminance_to_alpha_hsw
 FUNCTION(_sk_luminance_to_alpha_hsw)
 _sk_luminance_to_alpha_hsw:
-  .byte  196,226,125,24,29,183,13,0,0        // vbroadcastss  0xdb7(%rip),%ymm3        # 4b24 <_sk_callback_hsw+0x437>
-  .byte  196,98,125,24,5,178,13,0,0          // vbroadcastss  0xdb2(%rip),%ymm8        # 4b28 <_sk_callback_hsw+0x43b>
+  .byte  196,226,125,24,29,195,13,0,0        // vbroadcastss  0xdc3(%rip),%ymm3        # 4b9c <_sk_callback_hsw+0x443>
+  .byte  196,98,125,24,5,190,13,0,0          // vbroadcastss  0xdbe(%rip),%ymm8        # 4ba0 <_sk_callback_hsw+0x447>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  196,226,125,184,203                 // vfmadd231ps   %ymm3,%ymm0,%ymm1
-  .byte  196,226,125,24,29,163,13,0,0        // vbroadcastss  0xda3(%rip),%ymm3        # 4b2c <_sk_callback_hsw+0x43f>
+  .byte  196,226,125,24,29,175,13,0,0        // vbroadcastss  0xdaf(%rip),%ymm3        # 4ba4 <_sk_callback_hsw+0x44b>
   .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
@@ -12826,9 +12953,9 @@ _sk_evenly_spaced_gradient_hsw:
   .byte  76,139,64,8                         // mov           0x8(%rax),%r8
   .byte  77,137,202                          // mov           %r9,%r10
   .byte  73,255,202                          // dec           %r10
-  .byte  120,7                               // js            4041 <_sk_evenly_spaced_gradient_hsw+0x18>
+  .byte  120,7                               // js            40ad <_sk_evenly_spaced_gradient_hsw+0x18>
   .byte  196,193,242,42,202                  // vcvtsi2ss     %r10,%xmm1,%xmm1
-  .byte  235,22                              // jmp           4057 <_sk_evenly_spaced_gradient_hsw+0x2e>
+  .byte  235,22                              // jmp           40c3 <_sk_evenly_spaced_gradient_hsw+0x2e>
   .byte  77,137,211                          // mov           %r10,%r11
   .byte  73,209,235                          // shr           %r11
   .byte  65,131,226,1                        // and           $0x1,%r10d
@@ -12839,7 +12966,7 @@ _sk_evenly_spaced_gradient_hsw:
   .byte  197,244,89,200                      // vmulps        %ymm0,%ymm1,%ymm1
   .byte  197,126,91,217                      // vcvttps2dq    %ymm1,%ymm11
   .byte  73,131,249,8                        // cmp           $0x8,%r9
-  .byte  119,70                              // ja            40b0 <_sk_evenly_spaced_gradient_hsw+0x87>
+  .byte  119,70                              // ja            411c <_sk_evenly_spaced_gradient_hsw+0x87>
   .byte  196,66,37,22,0                      // vpermps       (%r8),%ymm11,%ymm8
   .byte  76,139,64,40                        // mov           0x28(%rax),%r8
   .byte  196,66,37,22,8                      // vpermps       (%r8),%ymm11,%ymm9
@@ -12855,7 +12982,7 @@ _sk_evenly_spaced_gradient_hsw:
   .byte  196,194,37,22,24                    // vpermps       (%r8),%ymm11,%ymm3
   .byte  72,139,64,64                        // mov           0x40(%rax),%rax
   .byte  196,98,37,22,40                     // vpermps       (%rax),%ymm11,%ymm13
-  .byte  235,110                             // jmp           411e <_sk_evenly_spaced_gradient_hsw+0xf5>
+  .byte  235,110                             // jmp           418a <_sk_evenly_spaced_gradient_hsw+0xf5>
   .byte  196,65,13,118,246                   // vpcmpeqd      %ymm14,%ymm14,%ymm14
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
   .byte  196,2,117,146,4,152                 // vgatherdps    %ymm1,(%r8,%ymm11,4),%ymm8
@@ -12894,11 +13021,11 @@ _sk_gradient_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  73,131,248,1                        // cmp           $0x1,%r8
-  .byte  15,134,180,0,0,0                    // jbe           41fd <_sk_gradient_hsw+0xc3>
+  .byte  15,134,180,0,0,0                    // jbe           4269 <_sk_gradient_hsw+0xc3>
   .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  196,226,125,24,21,208,9,0,0         // vbroadcastss  0x9d0(%rip),%ymm2        # 4b30 <_sk_callback_hsw+0x443>
+  .byte  196,226,125,24,21,220,9,0,0         // vbroadcastss  0x9dc(%rip),%ymm2        # 4ba8 <_sk_callback_hsw+0x44f>
   .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
@@ -12906,10 +13033,10 @@ _sk_gradient_hsw:
   .byte  196,65,101,254,201                  // vpaddd        %ymm9,%ymm3,%ymm9
   .byte  73,255,194                          // inc           %r10
   .byte  77,57,208                           // cmp           %r10,%r8
-  .byte  117,226                             // jne           4165 <_sk_gradient_hsw+0x2b>
+  .byte  117,226                             // jne           41d1 <_sk_gradient_hsw+0x2b>
   .byte  76,139,72,8                         // mov           0x8(%rax),%r9
   .byte  73,131,248,8                        // cmp           $0x8,%r8
-  .byte  118,121                             // jbe           4206 <_sk_gradient_hsw+0xcc>
+  .byte  118,121                             // jbe           4272 <_sk_gradient_hsw+0xcc>
   .byte  196,65,13,118,246                   // vpcmpeqd      %ymm14,%ymm14,%ymm14
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
   .byte  196,2,117,146,4,137                 // vgatherdps    %ymm1,(%r9,%ymm9,4),%ymm8
@@ -12933,7 +13060,7 @@ _sk_gradient_hsw:
   .byte  196,130,21,146,28,136               // vgatherdps    %ymm13,(%r8,%ymm9,4),%ymm3
   .byte  72,139,64,64                        // mov           0x40(%rax),%rax
   .byte  196,34,13,146,44,136                // vgatherdps    %ymm14,(%rax,%ymm9,4),%ymm13
-  .byte  235,77                              // jmp           424a <_sk_gradient_hsw+0x110>
+  .byte  235,77                              // jmp           42b6 <_sk_gradient_hsw+0x110>
   .byte  76,139,72,8                         // mov           0x8(%rax),%r9
   .byte  196,65,52,87,201                    // vxorps        %ymm9,%ymm9,%ymm9
   .byte  196,66,53,22,1                      // vpermps       (%r9),%ymm9,%ymm8
@@ -12993,24 +13120,24 @@ _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  196,98,125,24,45,79,8,0,0           // vbroadcastss  0x84f(%rip),%ymm13        # 4b34 <_sk_callback_hsw+0x447>
-  .byte  196,98,125,24,53,74,8,0,0           // vbroadcastss  0x84a(%rip),%ymm14        # 4b38 <_sk_callback_hsw+0x44b>
+  .byte  196,98,125,24,45,91,8,0,0           // vbroadcastss  0x85b(%rip),%ymm13        # 4bac <_sk_callback_hsw+0x453>
+  .byte  196,98,125,24,53,86,8,0,0           // vbroadcastss  0x856(%rip),%ymm14        # 4bb0 <_sk_callback_hsw+0x457>
   .byte  196,66,29,184,245                   // vfmadd231ps   %ymm13,%ymm12,%ymm14
-  .byte  196,98,125,24,45,64,8,0,0           // vbroadcastss  0x840(%rip),%ymm13        # 4b3c <_sk_callback_hsw+0x44f>
+  .byte  196,98,125,24,45,76,8,0,0           // vbroadcastss  0x84c(%rip),%ymm13        # 4bb4 <_sk_callback_hsw+0x45b>
   .byte  196,66,29,184,238                   // vfmadd231ps   %ymm14,%ymm12,%ymm13
-  .byte  196,98,125,24,53,54,8,0,0           // vbroadcastss  0x836(%rip),%ymm14        # 4b40 <_sk_callback_hsw+0x453>
+  .byte  196,98,125,24,53,66,8,0,0           // vbroadcastss  0x842(%rip),%ymm14        # 4bb8 <_sk_callback_hsw+0x45f>
   .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  196,98,125,24,21,33,8,0,0           // vbroadcastss  0x821(%rip),%ymm10        # 4b44 <_sk_callback_hsw+0x457>
+  .byte  196,98,125,24,21,45,8,0,0           // vbroadcastss  0x82d(%rip),%ymm10        # 4bbc <_sk_callback_hsw+0x463>
   .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  196,98,125,24,21,11,8,0,0           // vbroadcastss  0x80b(%rip),%ymm10        # 4b48 <_sk_callback_hsw+0x45b>
+  .byte  196,98,125,24,21,23,8,0,0           // vbroadcastss  0x817(%rip),%ymm10        # 4bc0 <_sk_callback_hsw+0x467>
   .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  196,98,125,24,21,245,7,0,0          // vbroadcastss  0x7f5(%rip),%ymm10        # 4b4c <_sk_callback_hsw+0x45f>
+  .byte  196,98,125,24,21,1,8,0,0            // vbroadcastss  0x801(%rip),%ymm10        # 4bc4 <_sk_callback_hsw+0x46b>
   .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
@@ -13033,7 +13160,7 @@ HIDDEN _sk_save_xy_hsw
 FUNCTION(_sk_save_xy_hsw)
 _sk_save_xy_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,194,7,0,0           // vbroadcastss  0x7c2(%rip),%ymm8        # 4b50 <_sk_callback_hsw+0x463>
+  .byte  196,98,125,24,5,206,7,0,0           // vbroadcastss  0x7ce(%rip),%ymm8        # 4bc8 <_sk_callback_hsw+0x46f>
   .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
@@ -13067,9 +13194,9 @@ HIDDEN _sk_bilinear_nx_hsw
 FUNCTION(_sk_bilinear_nx_hsw)
 _sk_bilinear_nx_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,86,7,0,0           // vbroadcastss  0x756(%rip),%ymm0        # 4b54 <_sk_callback_hsw+0x467>
+  .byte  196,226,125,24,5,98,7,0,0           // vbroadcastss  0x762(%rip),%ymm0        # 4bcc <_sk_callback_hsw+0x473>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,77,7,0,0            // vbroadcastss  0x74d(%rip),%ymm8        # 4b58 <_sk_callback_hsw+0x46b>
+  .byte  196,98,125,24,5,89,7,0,0            // vbroadcastss  0x759(%rip),%ymm8        # 4bd0 <_sk_callback_hsw+0x477>
   .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
@@ -13080,7 +13207,7 @@ HIDDEN _sk_bilinear_px_hsw
 FUNCTION(_sk_bilinear_px_hsw)
 _sk_bilinear_px_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,53,7,0,0           // vbroadcastss  0x735(%rip),%ymm0        # 4b5c <_sk_callback_hsw+0x46f>
+  .byte  196,226,125,24,5,65,7,0,0           // vbroadcastss  0x741(%rip),%ymm0        # 4bd4 <_sk_callback_hsw+0x47b>
   .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)
@@ -13092,9 +13219,9 @@ HIDDEN _sk_bilinear_ny_hsw
 FUNCTION(_sk_bilinear_ny_hsw)
 _sk_bilinear_ny_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,25,7,0,0          // vbroadcastss  0x719(%rip),%ymm1        # 4b60 <_sk_callback_hsw+0x473>
+  .byte  196,226,125,24,13,37,7,0,0          // vbroadcastss  0x725(%rip),%ymm1        # 4bd8 <_sk_callback_hsw+0x47f>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,15,7,0,0            // vbroadcastss  0x70f(%rip),%ymm8        # 4b64 <_sk_callback_hsw+0x477>
+  .byte  196,98,125,24,5,27,7,0,0            // vbroadcastss  0x71b(%rip),%ymm8        # 4bdc <_sk_callback_hsw+0x483>
   .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
@@ -13105,7 +13232,7 @@ HIDDEN _sk_bilinear_py_hsw
 FUNCTION(_sk_bilinear_py_hsw)
 _sk_bilinear_py_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,247,6,0,0         // vbroadcastss  0x6f7(%rip),%ymm1        # 4b68 <_sk_callback_hsw+0x47b>
+  .byte  196,226,125,24,13,3,7,0,0           // vbroadcastss  0x703(%rip),%ymm1        # 4be0 <_sk_callback_hsw+0x487>
   .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)
@@ -13117,13 +13244,13 @@ HIDDEN _sk_bicubic_n3x_hsw
 FUNCTION(_sk_bicubic_n3x_hsw)
 _sk_bicubic_n3x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,218,6,0,0          // vbroadcastss  0x6da(%rip),%ymm0        # 4b6c <_sk_callback_hsw+0x47f>
+  .byte  196,226,125,24,5,230,6,0,0          // vbroadcastss  0x6e6(%rip),%ymm0        # 4be4 <_sk_callback_hsw+0x48b>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,209,6,0,0           // vbroadcastss  0x6d1(%rip),%ymm8        # 4b70 <_sk_callback_hsw+0x483>
+  .byte  196,98,125,24,5,221,6,0,0           // vbroadcastss  0x6dd(%rip),%ymm8        # 4be8 <_sk_callback_hsw+0x48f>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,194,6,0,0          // vbroadcastss  0x6c2(%rip),%ymm10        # 4b74 <_sk_callback_hsw+0x487>
-  .byte  196,98,125,24,29,189,6,0,0          // vbroadcastss  0x6bd(%rip),%ymm11        # 4b78 <_sk_callback_hsw+0x48b>
+  .byte  196,98,125,24,21,206,6,0,0          // vbroadcastss  0x6ce(%rip),%ymm10        # 4bec <_sk_callback_hsw+0x493>
+  .byte  196,98,125,24,29,201,6,0,0          // vbroadcastss  0x6c9(%rip),%ymm11        # 4bf0 <_sk_callback_hsw+0x497>
   .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)
@@ -13135,16 +13262,16 @@ HIDDEN _sk_bicubic_n1x_hsw
 FUNCTION(_sk_bicubic_n1x_hsw)
 _sk_bicubic_n1x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,160,6,0,0          // vbroadcastss  0x6a0(%rip),%ymm0        # 4b7c <_sk_callback_hsw+0x48f>
+  .byte  196,226,125,24,5,172,6,0,0          // vbroadcastss  0x6ac(%rip),%ymm0        # 4bf4 <_sk_callback_hsw+0x49b>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,151,6,0,0           // vbroadcastss  0x697(%rip),%ymm8        # 4b80 <_sk_callback_hsw+0x493>
+  .byte  196,98,125,24,5,163,6,0,0           // vbroadcastss  0x6a3(%rip),%ymm8        # 4bf8 <_sk_callback_hsw+0x49f>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,141,6,0,0          // vbroadcastss  0x68d(%rip),%ymm9        # 4b84 <_sk_callback_hsw+0x497>
-  .byte  196,98,125,24,21,136,6,0,0          // vbroadcastss  0x688(%rip),%ymm10        # 4b88 <_sk_callback_hsw+0x49b>
+  .byte  196,98,125,24,13,153,6,0,0          // vbroadcastss  0x699(%rip),%ymm9        # 4bfc <_sk_callback_hsw+0x4a3>
+  .byte  196,98,125,24,21,148,6,0,0          // vbroadcastss  0x694(%rip),%ymm10        # 4c00 <_sk_callback_hsw+0x4a7>
   .byte  196,66,61,168,209                   // vfmadd213ps   %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,13,126,6,0,0          // vbroadcastss  0x67e(%rip),%ymm9        # 4b8c <_sk_callback_hsw+0x49f>
+  .byte  196,98,125,24,13,138,6,0,0          // vbroadcastss  0x68a(%rip),%ymm9        # 4c04 <_sk_callback_hsw+0x4ab>
   .byte  196,66,61,184,202                   // vfmadd231ps   %ymm10,%ymm8,%ymm9
-  .byte  196,98,125,24,21,116,6,0,0          // vbroadcastss  0x674(%rip),%ymm10        # 4b90 <_sk_callback_hsw+0x4a3>
+  .byte  196,98,125,24,21,128,6,0,0          // vbroadcastss  0x680(%rip),%ymm10        # 4c08 <_sk_callback_hsw+0x4af>
   .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
@@ -13155,14 +13282,14 @@ HIDDEN _sk_bicubic_p1x_hsw
 FUNCTION(_sk_bicubic_p1x_hsw)
 _sk_bicubic_p1x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,92,6,0,0            // vbroadcastss  0x65c(%rip),%ymm8        # 4b94 <_sk_callback_hsw+0x4a7>
+  .byte  196,98,125,24,5,104,6,0,0           // vbroadcastss  0x668(%rip),%ymm8        # 4c0c <_sk_callback_hsw+0x4b3>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,72,64                    // vmovups       0x40(%rax),%ymm9
-  .byte  196,98,125,24,21,78,6,0,0           // vbroadcastss  0x64e(%rip),%ymm10        # 4b98 <_sk_callback_hsw+0x4ab>
-  .byte  196,98,125,24,29,73,6,0,0           // vbroadcastss  0x649(%rip),%ymm11        # 4b9c <_sk_callback_hsw+0x4af>
+  .byte  196,98,125,24,21,90,6,0,0           // vbroadcastss  0x65a(%rip),%ymm10        # 4c10 <_sk_callback_hsw+0x4b7>
+  .byte  196,98,125,24,29,85,6,0,0           // vbroadcastss  0x655(%rip),%ymm11        # 4c14 <_sk_callback_hsw+0x4bb>
   .byte  196,66,53,168,218                   // vfmadd213ps   %ymm10,%ymm9,%ymm11
   .byte  196,66,53,168,216                   // vfmadd213ps   %ymm8,%ymm9,%ymm11
-  .byte  196,98,125,24,5,58,6,0,0            // vbroadcastss  0x63a(%rip),%ymm8        # 4ba0 <_sk_callback_hsw+0x4b3>
+  .byte  196,98,125,24,5,70,6,0,0            // vbroadcastss  0x646(%rip),%ymm8        # 4c18 <_sk_callback_hsw+0x4bf>
   .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
@@ -13173,12 +13300,12 @@ HIDDEN _sk_bicubic_p3x_hsw
 FUNCTION(_sk_bicubic_p3x_hsw)
 _sk_bicubic_p3x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,34,6,0,0           // vbroadcastss  0x622(%rip),%ymm0        # 4ba4 <_sk_callback_hsw+0x4b7>
+  .byte  196,226,125,24,5,46,6,0,0           // vbroadcastss  0x62e(%rip),%ymm0        # 4c1c <_sk_callback_hsw+0x4c3>
   .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  196,98,125,24,21,15,6,0,0           // vbroadcastss  0x60f(%rip),%ymm10        # 4ba8 <_sk_callback_hsw+0x4bb>
-  .byte  196,98,125,24,29,10,6,0,0           // vbroadcastss  0x60a(%rip),%ymm11        # 4bac <_sk_callback_hsw+0x4bf>
+  .byte  196,98,125,24,21,27,6,0,0           // vbroadcastss  0x61b(%rip),%ymm10        # 4c20 <_sk_callback_hsw+0x4c7>
+  .byte  196,98,125,24,29,22,6,0,0           // vbroadcastss  0x616(%rip),%ymm11        # 4c24 <_sk_callback_hsw+0x4cb>
   .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)
@@ -13190,13 +13317,13 @@ HIDDEN _sk_bicubic_n3y_hsw
 FUNCTION(_sk_bicubic_n3y_hsw)
 _sk_bicubic_n3y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,237,5,0,0         // vbroadcastss  0x5ed(%rip),%ymm1        # 4bb0 <_sk_callback_hsw+0x4c3>
+  .byte  196,226,125,24,13,249,5,0,0         // vbroadcastss  0x5f9(%rip),%ymm1        # 4c28 <_sk_callback_hsw+0x4cf>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,227,5,0,0           // vbroadcastss  0x5e3(%rip),%ymm8        # 4bb4 <_sk_callback_hsw+0x4c7>
+  .byte  196,98,125,24,5,239,5,0,0           // vbroadcastss  0x5ef(%rip),%ymm8        # 4c2c <_sk_callback_hsw+0x4d3>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,212,5,0,0          // vbroadcastss  0x5d4(%rip),%ymm10        # 4bb8 <_sk_callback_hsw+0x4cb>
-  .byte  196,98,125,24,29,207,5,0,0          // vbroadcastss  0x5cf(%rip),%ymm11        # 4bbc <_sk_callback_hsw+0x4cf>
+  .byte  196,98,125,24,21,224,5,0,0          // vbroadcastss  0x5e0(%rip),%ymm10        # 4c30 <_sk_callback_hsw+0x4d7>
+  .byte  196,98,125,24,29,219,5,0,0          // vbroadcastss  0x5db(%rip),%ymm11        # 4c34 <_sk_callback_hsw+0x4db>
   .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)
@@ -13208,16 +13335,16 @@ HIDDEN _sk_bicubic_n1y_hsw
 FUNCTION(_sk_bicubic_n1y_hsw)
 _sk_bicubic_n1y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,178,5,0,0         // vbroadcastss  0x5b2(%rip),%ymm1        # 4bc0 <_sk_callback_hsw+0x4d3>
+  .byte  196,226,125,24,13,190,5,0,0         // vbroadcastss  0x5be(%rip),%ymm1        # 4c38 <_sk_callback_hsw+0x4df>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,168,5,0,0           // vbroadcastss  0x5a8(%rip),%ymm8        # 4bc4 <_sk_callback_hsw+0x4d7>
+  .byte  196,98,125,24,5,180,5,0,0           // vbroadcastss  0x5b4(%rip),%ymm8        # 4c3c <_sk_callback_hsw+0x4e3>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,158,5,0,0          // vbroadcastss  0x59e(%rip),%ymm9        # 4bc8 <_sk_callback_hsw+0x4db>
-  .byte  196,98,125,24,21,153,5,0,0          // vbroadcastss  0x599(%rip),%ymm10        # 4bcc <_sk_callback_hsw+0x4df>
+  .byte  196,98,125,24,13,170,5,0,0          // vbroadcastss  0x5aa(%rip),%ymm9        # 4c40 <_sk_callback_hsw+0x4e7>
+  .byte  196,98,125,24,21,165,5,0,0          // vbroadcastss  0x5a5(%rip),%ymm10        # 4c44 <_sk_callback_hsw+0x4eb>
   .byte  196,66,61,168,209                   // vfmadd213ps   %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,13,143,5,0,0          // vbroadcastss  0x58f(%rip),%ymm9        # 4bd0 <_sk_callback_hsw+0x4e3>
+  .byte  196,98,125,24,13,155,5,0,0          // vbroadcastss  0x59b(%rip),%ymm9        # 4c48 <_sk_callback_hsw+0x4ef>
   .byte  196,66,61,184,202                   // vfmadd231ps   %ymm10,%ymm8,%ymm9
-  .byte  196,98,125,24,21,133,5,0,0          // vbroadcastss  0x585(%rip),%ymm10        # 4bd4 <_sk_callback_hsw+0x4e7>
+  .byte  196,98,125,24,21,145,5,0,0          // vbroadcastss  0x591(%rip),%ymm10        # 4c4c <_sk_callback_hsw+0x4f3>
   .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
@@ -13228,14 +13355,14 @@ HIDDEN _sk_bicubic_p1y_hsw
 FUNCTION(_sk_bicubic_p1y_hsw)
 _sk_bicubic_p1y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,109,5,0,0           // vbroadcastss  0x56d(%rip),%ymm8        # 4bd8 <_sk_callback_hsw+0x4eb>
+  .byte  196,98,125,24,5,121,5,0,0           // vbroadcastss  0x579(%rip),%ymm8        # 4c50 <_sk_callback_hsw+0x4f7>
   .byte  197,188,88,72,32                    // vaddps        0x20(%rax),%ymm8,%ymm1
   .byte  197,124,16,72,96                    // vmovups       0x60(%rax),%ymm9
-  .byte  196,98,125,24,21,94,5,0,0           // vbroadcastss  0x55e(%rip),%ymm10        # 4bdc <_sk_callback_hsw+0x4ef>
-  .byte  196,98,125,24,29,89,5,0,0           // vbroadcastss  0x559(%rip),%ymm11        # 4be0 <_sk_callback_hsw+0x4f3>
+  .byte  196,98,125,24,21,106,5,0,0          // vbroadcastss  0x56a(%rip),%ymm10        # 4c54 <_sk_callback_hsw+0x4fb>
+  .byte  196,98,125,24,29,101,5,0,0          // vbroadcastss  0x565(%rip),%ymm11        # 4c58 <_sk_callback_hsw+0x4ff>
   .byte  196,66,53,168,218                   // vfmadd213ps   %ymm10,%ymm9,%ymm11
   .byte  196,66,53,168,216                   // vfmadd213ps   %ymm8,%ymm9,%ymm11
-  .byte  196,98,125,24,5,74,5,0,0            // vbroadcastss  0x54a(%rip),%ymm8        # 4be4 <_sk_callback_hsw+0x4f7>
+  .byte  196,98,125,24,5,86,5,0,0            // vbroadcastss  0x556(%rip),%ymm8        # 4c5c <_sk_callback_hsw+0x503>
   .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
@@ -13246,12 +13373,12 @@ HIDDEN _sk_bicubic_p3y_hsw
 FUNCTION(_sk_bicubic_p3y_hsw)
 _sk_bicubic_p3y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,50,5,0,0          // vbroadcastss  0x532(%rip),%ymm1        # 4be8 <_sk_callback_hsw+0x4fb>
+  .byte  196,226,125,24,13,62,5,0,0          // vbroadcastss  0x53e(%rip),%ymm1        # 4c60 <_sk_callback_hsw+0x507>
   .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  196,98,125,24,21,30,5,0,0           // vbroadcastss  0x51e(%rip),%ymm10        # 4bec <_sk_callback_hsw+0x4ff>
-  .byte  196,98,125,24,29,25,5,0,0           // vbroadcastss  0x519(%rip),%ymm11        # 4bf0 <_sk_callback_hsw+0x503>
+  .byte  196,98,125,24,21,42,5,0,0           // vbroadcastss  0x52a(%rip),%ymm10        # 4c64 <_sk_callback_hsw+0x50b>
+  .byte  196,98,125,24,29,37,5,0,0           // vbroadcastss  0x525(%rip),%ymm11        # 4c68 <_sk_callback_hsw+0x50f>
   .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)
@@ -13375,25 +13502,25 @@ 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 48c5 <.literal4+0xb1>
+  .byte  71,225,61                           // rex.RXB       loope 4931 <.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 48d5 <.literal4+0xc1>
+  .byte  71,225,61                           // rex.RXB       loope 4941 <.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 48e5 <.literal4+0xd1>
+  .byte  71,225,61                           // rex.RXB       loope 4951 <.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 48f5 <.literal4+0xe1>
+  .byte  71,225,61                           // rex.RXB       loope 4961 <.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)
@@ -13440,7 +13567,7 @@ 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        4945 <.literal4+0x131>
+  .byte  224,7                               // loopne        49b1 <.literal4+0x131>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -13456,10 +13583,10 @@ 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            496c <.literal4+0x158>
+  .byte  127,0                               // jg            49d8 <.literal4+0x158>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            49e5 <.literal4+0x1d1>
+  .byte  119,115                             // ja            4a51 <.literal4+0x1d1>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13473,10 +13600,10 @@ 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            49a0 <.literal4+0x18c>
+  .byte  127,0                               // jg            4a0c <.literal4+0x18c>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4a19 <.literal4+0x205>
+  .byte  119,115                             // ja            4a85 <.literal4+0x205>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13490,10 +13617,10 @@ 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            49d4 <.literal4+0x1c0>
+  .byte  127,0                               // jg            4a40 <.literal4+0x1c0>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4a4d <.literal4+0x239>
+  .byte  119,115                             // ja            4ab9 <.literal4+0x239>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13507,10 +13634,10 @@ 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            4a08 <.literal4+0x1f4>
+  .byte  127,0                               // jg            4a74 <.literal4+0x1f4>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4a81 <.literal4+0x26d>
+  .byte  119,115                             // ja            4aed <.literal4+0x26d>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -13523,7 +13650,7 @@ 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            4a7f <.literal4+0x26b>
+  .byte  127,67                              // jg            4aeb <.literal4+0x26b>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,195                               // add           %al,%bl
   .byte  0,0                                 // add           %al,(%rax)
@@ -13535,10 +13662,10 @@ BALIGN4
   .byte  190,80,128,3,62                     // mov           $0x3e038050,%esi
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           4a9f <.literal4+0x28b>
+  .byte  118,63                              // jbe           4b0b <.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            4ab3 <.literal4+0x29f>
+  .byte  127,67                              // jg            4b1f <.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)
@@ -13547,7 +13674,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        4a95 <.literal4+0x281>
+  .byte  224,7                               // loopne        4b01 <.literal4+0x281>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -13559,7 +13686,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        4ab1 <.literal4+0x29d>
+  .byte  224,7                               // loopne        4b1d <.literal4+0x29d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -13570,7 +13697,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            4b06 <.literal4+0x2f2>
+  .byte  124,66                              // jl            4b72 <.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)
@@ -13588,9 +13715,9 @@ 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            4b49 <.literal4+0x335>
+  .byte  112,65                              // jo            4bb5 <.literal4+0x335>
   .byte  129,128,128,59,129,128,128,59,0,0   // addl          $0x3b80,-0x7f7ec480(%rax)
-  .byte  127,67                              // jg            4b57 <.literal4+0x343>
+  .byte  127,67                              // jg            4bc3 <.literal4+0x343>
   .byte  128,0,128                           // addb          $0x80,(%rax)
   .byte  55                                  // (bad)
   .byte  128,0,128                           // addb          $0x80,(%rax)
@@ -13598,7 +13725,11 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  255                                 // (bad)
-  .byte  127,71                              // jg            4b6b <.literal4+0x357>
+  .byte  127,71                              // jg            4bd7 <.literal4+0x357>
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,191,0,0,0                     // add           %al,0xbf(%rax)
+  .byte  63                                  // (bad)
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,89                               // ds            pop %rcx
@@ -13698,16 +13829,16 @@ 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)        # a004c28 <_sk_callback_hsw+0xa00053b>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004ca8 <_sk_callback_hsw+0xa00054f>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004c30 <_sk_callback_hsw+0x12000543>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004cb0 <_sk_callback_hsw+0x12000557>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004c38 <_sk_callback_hsw+0x1a00054b>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004cb8 <_sk_callback_hsw+0x1a00055f>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004c40 <_sk_callback_hsw+0x3000553>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004cc0 <_sk_callback_hsw+0x3000567>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13750,16 +13881,16 @@ 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)        # a004c88 <_sk_callback_hsw+0xa00059b>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004d08 <_sk_callback_hsw+0xa0005af>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004c90 <_sk_callback_hsw+0x120005a3>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004d10 <_sk_callback_hsw+0x120005b7>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004c98 <_sk_callback_hsw+0x1a0005ab>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004d18 <_sk_callback_hsw+0x1a0005bf>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004ca0 <_sk_callback_hsw+0x30005b3>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004d20 <_sk_callback_hsw+0x30005c7>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13802,16 +13933,16 @@ 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)        # a004ce8 <_sk_callback_hsw+0xa0005fb>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004d68 <_sk_callback_hsw+0xa00060f>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004cf0 <_sk_callback_hsw+0x12000603>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004d70 <_sk_callback_hsw+0x12000617>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004cf8 <_sk_callback_hsw+0x1a00060b>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004d78 <_sk_callback_hsw+0x1a00061f>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004d00 <_sk_callback_hsw+0x3000613>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004d80 <_sk_callback_hsw+0x3000627>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13854,16 +13985,16 @@ 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)        # a004d48 <_sk_callback_hsw+0xa00065b>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004dc8 <_sk_callback_hsw+0xa00066f>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004d50 <_sk_callback_hsw+0x12000663>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12004dd0 <_sk_callback_hsw+0x12000677>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004d58 <_sk_callback_hsw+0x1a00066b>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a004dd8 <_sk_callback_hsw+0x1a00067f>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004d60 <_sk_callback_hsw+0x3000673>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3004de0 <_sk_callback_hsw+0x3000687>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -13984,14 +14115,14 @@ _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  196,226,125,24,13,95,99,0,0         // vbroadcastss  0x635f(%rip),%ymm1        # 6428 <_sk_callback_avx+0x125>
+  .byte  196,226,125,24,13,203,99,0,0        // vbroadcastss  0x63cb(%rip),%ymm1        # 6494 <_sk_callback_avx+0x125>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  197,252,88,2                        // vaddps        (%rdx),%ymm0,%ymm0
   .byte  196,226,125,24,16                   // vbroadcastss  (%rax),%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  197,236,88,201                      // vaddps        %ymm1,%ymm2,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,21,67,99,0,0         // vbroadcastss  0x6343(%rip),%ymm2        # 642c <_sk_callback_avx+0x129>
+  .byte  196,226,125,24,21,175,99,0,0        // vbroadcastss  0x63af(%rip),%ymm2        # 6498 <_sk_callback_avx+0x129>
   .byte  197,228,87,219                      // vxorps        %ymm3,%ymm3,%ymm3
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
   .byte  197,212,87,237                      // vxorps        %ymm5,%ymm5,%ymm5
@@ -14014,7 +14145,7 @@ _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  196,98,125,24,29,244,98,0,0         // vbroadcastss  0x62f4(%rip),%ymm11        # 6430 <_sk_callback_avx+0x12d>
+  .byte  196,98,125,24,29,96,99,0,0          // vbroadcastss  0x6360(%rip),%ymm11        # 649c <_sk_callback_avx+0x12d>
   .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
@@ -14025,8 +14156,8 @@ _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  196,98,125,24,37,181,98,0,0         // vbroadcastss  0x62b5(%rip),%ymm12        # 6434 <_sk_callback_avx+0x131>
-  .byte  196,98,125,24,45,176,98,0,0         // vbroadcastss  0x62b0(%rip),%ymm13        # 6438 <_sk_callback_avx+0x135>
+  .byte  196,98,125,24,37,33,99,0,0          // vbroadcastss  0x6321(%rip),%ymm12        # 64a0 <_sk_callback_avx+0x131>
+  .byte  196,98,125,24,45,28,99,0,0          // vbroadcastss  0x631c(%rip),%ymm13        # 64a4 <_sk_callback_avx+0x135>
   .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
@@ -14053,9 +14184,9 @@ _sk_dither_avx:
   .byte  196,65,12,86,202                    // vorps         %ymm10,%ymm14,%ymm9
   .byte  196,65,60,86,193                    // vorps         %ymm9,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,27,98,0,0          // vbroadcastss  0x621b(%rip),%ymm9        # 643c <_sk_callback_avx+0x139>
+  .byte  196,98,125,24,13,135,98,0,0         // vbroadcastss  0x6287(%rip),%ymm9        # 64a8 <_sk_callback_avx+0x139>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,17,98,0,0          // vbroadcastss  0x6211(%rip),%ymm9        # 6440 <_sk_callback_avx+0x13d>
+  .byte  196,98,125,24,13,125,98,0,0         // vbroadcastss  0x627d(%rip),%ymm9        # 64ac <_sk_callback_avx+0x13d>
   .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
@@ -14124,7 +14255,7 @@ HIDDEN _sk_srcatop_avx
 FUNCTION(_sk_srcatop_avx)
 _sk_srcatop_avx:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,104,97,0,0          // vbroadcastss  0x6168(%rip),%ymm8        # 6444 <_sk_callback_avx+0x141>
+  .byte  196,98,125,24,5,212,97,0,0          // vbroadcastss  0x61d4(%rip),%ymm8        # 64b0 <_sk_callback_avx+0x141>
   .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
@@ -14145,7 +14276,7 @@ HIDDEN _sk_dstatop_avx
 FUNCTION(_sk_dstatop_avx)
 _sk_dstatop_avx:
   .byte  197,100,89,196                      // vmulps        %ymm4,%ymm3,%ymm8
-  .byte  196,98,125,24,13,42,97,0,0          // vbroadcastss  0x612a(%rip),%ymm9        # 6448 <_sk_callback_avx+0x145>
+  .byte  196,98,125,24,13,150,97,0,0         // vbroadcastss  0x6196(%rip),%ymm9        # 64b4 <_sk_callback_avx+0x145>
   .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
@@ -14187,7 +14318,7 @@ HIDDEN _sk_srcout_avx
 .globl _sk_srcout_avx
 FUNCTION(_sk_srcout_avx)
 _sk_srcout_avx:
-  .byte  196,98,125,24,5,201,96,0,0          // vbroadcastss  0x60c9(%rip),%ymm8        # 644c <_sk_callback_avx+0x149>
+  .byte  196,98,125,24,5,53,97,0,0           // vbroadcastss  0x6135(%rip),%ymm8        # 64b8 <_sk_callback_avx+0x149>
   .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
@@ -14200,7 +14331,7 @@ HIDDEN _sk_dstout_avx
 .globl _sk_dstout_avx
 FUNCTION(_sk_dstout_avx)
 _sk_dstout_avx:
-  .byte  196,226,125,24,5,172,96,0,0         // vbroadcastss  0x60ac(%rip),%ymm0        # 6450 <_sk_callback_avx+0x14d>
+  .byte  196,226,125,24,5,24,97,0,0          // vbroadcastss  0x6118(%rip),%ymm0        # 64bc <_sk_callback_avx+0x14d>
   .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
@@ -14213,7 +14344,7 @@ HIDDEN _sk_srcover_avx
 .globl _sk_srcover_avx
 FUNCTION(_sk_srcover_avx)
 _sk_srcover_avx:
-  .byte  196,98,125,24,5,143,96,0,0          // vbroadcastss  0x608f(%rip),%ymm8        # 6454 <_sk_callback_avx+0x151>
+  .byte  196,98,125,24,5,251,96,0,0          // vbroadcastss  0x60fb(%rip),%ymm8        # 64c0 <_sk_callback_avx+0x151>
   .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
@@ -14230,7 +14361,7 @@ HIDDEN _sk_dstover_avx
 .globl _sk_dstover_avx
 FUNCTION(_sk_dstover_avx)
 _sk_dstover_avx:
-  .byte  196,98,125,24,5,98,96,0,0           // vbroadcastss  0x6062(%rip),%ymm8        # 6458 <_sk_callback_avx+0x155>
+  .byte  196,98,125,24,5,206,96,0,0          // vbroadcastss  0x60ce(%rip),%ymm8        # 64c4 <_sk_callback_avx+0x155>
   .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
@@ -14258,7 +14389,7 @@ HIDDEN _sk_multiply_avx
 .globl _sk_multiply_avx
 FUNCTION(_sk_multiply_avx)
 _sk_multiply_avx:
-  .byte  196,98,125,24,5,33,96,0,0           // vbroadcastss  0x6021(%rip),%ymm8        # 645c <_sk_callback_avx+0x159>
+  .byte  196,98,125,24,5,141,96,0,0          // vbroadcastss  0x608d(%rip),%ymm8        # 64c8 <_sk_callback_avx+0x159>
   .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
@@ -14318,7 +14449,7 @@ HIDDEN _sk_xor__avx
 .globl _sk_xor__avx
 FUNCTION(_sk_xor__avx)
 _sk_xor__avx:
-  .byte  196,98,125,24,5,112,95,0,0          // vbroadcastss  0x5f70(%rip),%ymm8        # 6460 <_sk_callback_avx+0x15d>
+  .byte  196,98,125,24,5,220,95,0,0          // vbroadcastss  0x5fdc(%rip),%ymm8        # 64cc <_sk_callback_avx+0x15d>
   .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
@@ -14355,7 +14486,7 @@ _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  196,98,125,24,5,240,94,0,0          // vbroadcastss  0x5ef0(%rip),%ymm8        # 6464 <_sk_callback_avx+0x161>
+  .byte  196,98,125,24,5,92,95,0,0           // vbroadcastss  0x5f5c(%rip),%ymm8        # 64d0 <_sk_callback_avx+0x161>
   .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
@@ -14381,7 +14512,7 @@ _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  196,98,125,24,5,156,94,0,0          // vbroadcastss  0x5e9c(%rip),%ymm8        # 6468 <_sk_callback_avx+0x165>
+  .byte  196,98,125,24,5,8,95,0,0            // vbroadcastss  0x5f08(%rip),%ymm8        # 64d4 <_sk_callback_avx+0x165>
   .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
@@ -14410,7 +14541,7 @@ _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  196,98,125,24,5,60,94,0,0           // vbroadcastss  0x5e3c(%rip),%ymm8        # 646c <_sk_callback_avx+0x169>
+  .byte  196,98,125,24,5,168,94,0,0          // vbroadcastss  0x5ea8(%rip),%ymm8        # 64d8 <_sk_callback_avx+0x169>
   .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
@@ -14433,7 +14564,7 @@ _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  196,98,125,24,5,247,93,0,0          // vbroadcastss  0x5df7(%rip),%ymm8        # 6470 <_sk_callback_avx+0x16d>
+  .byte  196,98,125,24,5,99,94,0,0           // vbroadcastss  0x5e63(%rip),%ymm8        # 64dc <_sk_callback_avx+0x16d>
   .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
@@ -14444,7 +14575,7 @@ HIDDEN _sk_colorburn_avx
 .globl _sk_colorburn_avx
 FUNCTION(_sk_colorburn_avx)
 _sk_colorburn_avx:
-  .byte  196,98,125,24,5,226,93,0,0          // vbroadcastss  0x5de2(%rip),%ymm8        # 6474 <_sk_callback_avx+0x171>
+  .byte  196,98,125,24,5,78,94,0,0           // vbroadcastss  0x5e4e(%rip),%ymm8        # 64e0 <_sk_callback_avx+0x171>
   .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
@@ -14506,7 +14637,7 @@ HIDDEN _sk_colordodge_avx
 FUNCTION(_sk_colordodge_avx)
 _sk_colordodge_avx:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,13,222,92,0,0         // vbroadcastss  0x5cde(%rip),%ymm9        # 6478 <_sk_callback_avx+0x175>
+  .byte  196,98,125,24,13,74,93,0,0          // vbroadcastss  0x5d4a(%rip),%ymm9        # 64e4 <_sk_callback_avx+0x175>
   .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
@@ -14563,7 +14694,7 @@ HIDDEN _sk_hardlight_avx
 .globl _sk_hardlight_avx
 FUNCTION(_sk_hardlight_avx)
 _sk_hardlight_avx:
-  .byte  196,98,125,24,5,240,91,0,0          // vbroadcastss  0x5bf0(%rip),%ymm8        # 647c <_sk_callback_avx+0x179>
+  .byte  196,98,125,24,5,92,92,0,0           // vbroadcastss  0x5c5c(%rip),%ymm8        # 64e8 <_sk_callback_avx+0x179>
   .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
@@ -14618,7 +14749,7 @@ HIDDEN _sk_overlay_avx
 .globl _sk_overlay_avx
 FUNCTION(_sk_overlay_avx)
 _sk_overlay_avx:
-  .byte  196,98,125,24,5,25,91,0,0           // vbroadcastss  0x5b19(%rip),%ymm8        # 6480 <_sk_callback_avx+0x17d>
+  .byte  196,98,125,24,5,133,91,0,0          // vbroadcastss  0x5b85(%rip),%ymm8        # 64ec <_sk_callback_avx+0x17d>
   .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
@@ -14684,10 +14815,10 @@ _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  196,98,125,24,29,16,90,0,0          // vbroadcastss  0x5a10(%rip),%ymm11        # 6488 <_sk_callback_avx+0x185>
+  .byte  196,98,125,24,29,124,90,0,0         // vbroadcastss  0x5a7c(%rip),%ymm11        # 64f4 <_sk_callback_avx+0x185>
   .byte  196,65,28,88,235                    // vaddps        %ymm11,%ymm12,%ymm13
   .byte  196,65,20,89,192                    // vmulps        %ymm8,%ymm13,%ymm8
-  .byte  196,98,125,24,45,1,90,0,0           // vbroadcastss  0x5a01(%rip),%ymm13        # 648c <_sk_callback_avx+0x189>
+  .byte  196,98,125,24,45,109,90,0,0         // vbroadcastss  0x5a6d(%rip),%ymm13        # 64f8 <_sk_callback_avx+0x189>
   .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
@@ -14698,7 +14829,7 @@ _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  196,98,125,24,5,191,89,0,0          // vbroadcastss  0x59bf(%rip),%ymm8        # 6484 <_sk_callback_avx+0x181>
+  .byte  196,98,125,24,5,43,90,0,0           // vbroadcastss  0x5a2b(%rip),%ymm8        # 64f0 <_sk_callback_avx+0x181>
   .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
@@ -14825,12 +14956,12 @@ _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  196,98,125,24,53,142,87,0,0         // vbroadcastss  0x578e(%rip),%ymm14        # 6490 <_sk_callback_avx+0x18d>
+  .byte  196,98,125,24,53,250,87,0,0         // vbroadcastss  0x57fa(%rip),%ymm14        # 64fc <_sk_callback_avx+0x18d>
   .byte  196,65,92,89,222                    // vmulps        %ymm14,%ymm4,%ymm11
-  .byte  196,98,125,24,61,132,87,0,0         // vbroadcastss  0x5784(%rip),%ymm15        # 6494 <_sk_callback_avx+0x191>
+  .byte  196,98,125,24,61,240,87,0,0         // vbroadcastss  0x57f0(%rip),%ymm15        # 6500 <_sk_callback_avx+0x191>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
-  .byte  196,226,125,24,5,117,87,0,0         // vbroadcastss  0x5775(%rip),%ymm0        # 6498 <_sk_callback_avx+0x195>
+  .byte  196,226,125,24,5,225,87,0,0         // vbroadcastss  0x57e1(%rip),%ymm0        # 6504 <_sk_callback_avx+0x195>
   .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
@@ -14891,7 +15022,7 @@ _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  196,98,125,24,5,78,86,0,0           // vbroadcastss  0x564e(%rip),%ymm8        # 649c <_sk_callback_avx+0x199>
+  .byte  196,98,125,24,5,186,86,0,0          // vbroadcastss  0x56ba(%rip),%ymm8        # 6508 <_sk_callback_avx+0x199>
   .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
@@ -14948,12 +15079,12 @@ _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  196,98,125,24,53,92,85,0,0          // vbroadcastss  0x555c(%rip),%ymm14        # 64a0 <_sk_callback_avx+0x19d>
+  .byte  196,98,125,24,53,200,85,0,0         // vbroadcastss  0x55c8(%rip),%ymm14        # 650c <_sk_callback_avx+0x19d>
   .byte  196,65,92,89,222                    // vmulps        %ymm14,%ymm4,%ymm11
-  .byte  196,98,125,24,61,82,85,0,0          // vbroadcastss  0x5552(%rip),%ymm15        # 64a4 <_sk_callback_avx+0x1a1>
+  .byte  196,98,125,24,61,190,85,0,0         // vbroadcastss  0x55be(%rip),%ymm15        # 6510 <_sk_callback_avx+0x1a1>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
-  .byte  196,226,125,24,5,67,85,0,0          // vbroadcastss  0x5543(%rip),%ymm0        # 64a8 <_sk_callback_avx+0x1a5>
+  .byte  196,226,125,24,5,175,85,0,0         // vbroadcastss  0x55af(%rip),%ymm0        # 6514 <_sk_callback_avx+0x1a5>
   .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
@@ -15014,7 +15145,7 @@ _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  196,98,125,24,5,28,84,0,0           // vbroadcastss  0x541c(%rip),%ymm8        # 64ac <_sk_callback_avx+0x1a9>
+  .byte  196,98,125,24,5,136,84,0,0          // vbroadcastss  0x5488(%rip),%ymm8        # 6518 <_sk_callback_avx+0x1a9>
   .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
@@ -15043,12 +15174,12 @@ _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  196,98,125,24,45,178,83,0,0         // vbroadcastss  0x53b2(%rip),%ymm13        # 64b0 <_sk_callback_avx+0x1ad>
+  .byte  196,98,125,24,45,30,84,0,0          // vbroadcastss  0x541e(%rip),%ymm13        # 651c <_sk_callback_avx+0x1ad>
   .byte  196,65,92,89,213                    // vmulps        %ymm13,%ymm4,%ymm10
-  .byte  196,98,125,24,53,168,83,0,0         // vbroadcastss  0x53a8(%rip),%ymm14        # 64b4 <_sk_callback_avx+0x1b1>
+  .byte  196,98,125,24,53,20,84,0,0          // vbroadcastss  0x5414(%rip),%ymm14        # 6520 <_sk_callback_avx+0x1b1>
   .byte  196,65,84,89,222                    // vmulps        %ymm14,%ymm5,%ymm11
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,61,153,83,0,0         // vbroadcastss  0x5399(%rip),%ymm15        # 64b8 <_sk_callback_avx+0x1b5>
+  .byte  196,98,125,24,61,5,84,0,0           // vbroadcastss  0x5405(%rip),%ymm15        # 6524 <_sk_callback_avx+0x1b5>
   .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
@@ -15111,7 +15242,7 @@ _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  196,226,125,24,5,96,82,0,0          // vbroadcastss  0x5260(%rip),%ymm0        # 64bc <_sk_callback_avx+0x1b9>
+  .byte  196,226,125,24,5,204,82,0,0         // vbroadcastss  0x52cc(%rip),%ymm0        # 6528 <_sk_callback_avx+0x1b9>
   .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
@@ -15141,12 +15272,12 @@ _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  196,98,125,24,45,242,81,0,0         // vbroadcastss  0x51f2(%rip),%ymm13        # 64c0 <_sk_callback_avx+0x1bd>
+  .byte  196,98,125,24,45,94,82,0,0          // vbroadcastss  0x525e(%rip),%ymm13        # 652c <_sk_callback_avx+0x1bd>
   .byte  196,65,108,89,213                   // vmulps        %ymm13,%ymm2,%ymm10
-  .byte  196,98,125,24,53,232,81,0,0         // vbroadcastss  0x51e8(%rip),%ymm14        # 64c4 <_sk_callback_avx+0x1c1>
+  .byte  196,98,125,24,53,84,82,0,0          // vbroadcastss  0x5254(%rip),%ymm14        # 6530 <_sk_callback_avx+0x1c1>
   .byte  196,65,116,89,222                   // vmulps        %ymm14,%ymm1,%ymm11
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,61,217,81,0,0         // vbroadcastss  0x51d9(%rip),%ymm15        # 64c8 <_sk_callback_avx+0x1c5>
+  .byte  196,98,125,24,61,69,82,0,0          // vbroadcastss  0x5245(%rip),%ymm15        # 6534 <_sk_callback_avx+0x1c5>
   .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
@@ -15209,7 +15340,7 @@ _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  196,226,125,24,5,160,80,0,0         // vbroadcastss  0x50a0(%rip),%ymm0        # 64cc <_sk_callback_avx+0x1c9>
+  .byte  196,226,125,24,5,12,81,0,0          // vbroadcastss  0x510c(%rip),%ymm0        # 6538 <_sk_callback_avx+0x1c9>
   .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
@@ -15245,7 +15376,7 @@ HIDDEN _sk_clamp_1_avx
 .globl _sk_clamp_1_avx
 FUNCTION(_sk_clamp_1_avx)
 _sk_clamp_1_avx:
-  .byte  196,98,125,24,5,51,80,0,0           // vbroadcastss  0x5033(%rip),%ymm8        # 64d0 <_sk_callback_avx+0x1cd>
+  .byte  196,98,125,24,5,159,80,0,0          // vbroadcastss  0x509f(%rip),%ymm8        # 653c <_sk_callback_avx+0x1cd>
   .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
@@ -15257,7 +15388,7 @@ HIDDEN _sk_clamp_a_avx
 .globl _sk_clamp_a_avx
 FUNCTION(_sk_clamp_a_avx)
 _sk_clamp_a_avx:
-  .byte  196,98,125,24,5,22,80,0,0           // vbroadcastss  0x5016(%rip),%ymm8        # 64d4 <_sk_callback_avx+0x1d1>
+  .byte  196,98,125,24,5,130,80,0,0          // vbroadcastss  0x5082(%rip),%ymm8        # 6540 <_sk_callback_avx+0x1d1>
   .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
@@ -15343,7 +15474,7 @@ 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  196,98,125,24,21,94,79,0,0          // vbroadcastss  0x4f5e(%rip),%ymm10        # 64d8 <_sk_callback_avx+0x1d5>
+  .byte  196,98,125,24,21,202,79,0,0         // vbroadcastss  0x4fca(%rip),%ymm10        # 6544 <_sk_callback_avx+0x1d5>
   .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
@@ -15356,17 +15487,17 @@ HIDDEN _sk_from_srgb_avx
 .globl _sk_from_srgb_avx
 FUNCTION(_sk_from_srgb_avx)
 _sk_from_srgb_avx:
-  .byte  196,98,125,24,5,63,79,0,0           // vbroadcastss  0x4f3f(%rip),%ymm8        # 64dc <_sk_callback_avx+0x1d9>
+  .byte  196,98,125,24,5,171,79,0,0          // vbroadcastss  0x4fab(%rip),%ymm8        # 6548 <_sk_callback_avx+0x1d9>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,49,79,0,0          // vbroadcastss  0x4f31(%rip),%ymm11        # 64e0 <_sk_callback_avx+0x1dd>
+  .byte  196,98,125,24,29,157,79,0,0         // vbroadcastss  0x4f9d(%rip),%ymm11        # 654c <_sk_callback_avx+0x1dd>
   .byte  196,65,124,89,227                   // vmulps        %ymm11,%ymm0,%ymm12
-  .byte  196,98,125,24,45,39,79,0,0          // vbroadcastss  0x4f27(%rip),%ymm13        # 64e4 <_sk_callback_avx+0x1e1>
+  .byte  196,98,125,24,45,147,79,0,0         // vbroadcastss  0x4f93(%rip),%ymm13        # 6550 <_sk_callback_avx+0x1e1>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,44,89,212                    // vmulps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,24,79,0,0          // vbroadcastss  0x4f18(%rip),%ymm12        # 64e8 <_sk_callback_avx+0x1e5>
+  .byte  196,98,125,24,37,132,79,0,0         // vbroadcastss  0x4f84(%rip),%ymm12        # 6554 <_sk_callback_avx+0x1e5>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,53,14,79,0,0          // vbroadcastss  0x4f0e(%rip),%ymm14        # 64ec <_sk_callback_avx+0x1e9>
+  .byte  196,98,125,24,53,122,79,0,0         // vbroadcastss  0x4f7a(%rip),%ymm14        # 6558 <_sk_callback_avx+0x1e9>
   .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
@@ -15393,20 +15524,20 @@ HIDDEN _sk_to_srgb_avx
 FUNCTION(_sk_to_srgb_avx)
 _sk_to_srgb_avx:
   .byte  197,124,82,200                      // vrsqrtps      %ymm0,%ymm9
-  .byte  196,98,125,24,5,163,78,0,0          // vbroadcastss  0x4ea3(%rip),%ymm8        # 64f0 <_sk_callback_avx+0x1ed>
+  .byte  196,98,125,24,5,15,79,0,0           // vbroadcastss  0x4f0f(%rip),%ymm8        # 655c <_sk_callback_avx+0x1ed>
   .byte  196,65,124,89,208                   // vmulps        %ymm8,%ymm0,%ymm10
-  .byte  196,98,125,24,29,153,78,0,0         // vbroadcastss  0x4e99(%rip),%ymm11        # 64f4 <_sk_callback_avx+0x1f1>
+  .byte  196,98,125,24,29,5,79,0,0           // vbroadcastss  0x4f05(%rip),%ymm11        # 6560 <_sk_callback_avx+0x1f1>
   .byte  196,65,52,89,227                    // vmulps        %ymm11,%ymm9,%ymm12
-  .byte  196,98,125,24,45,143,78,0,0         // vbroadcastss  0x4e8f(%rip),%ymm13        # 64f8 <_sk_callback_avx+0x1f5>
+  .byte  196,98,125,24,45,251,78,0,0         // vbroadcastss  0x4efb(%rip),%ymm13        # 6564 <_sk_callback_avx+0x1f5>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,52,89,228                    // vmulps        %ymm12,%ymm9,%ymm12
-  .byte  196,98,125,24,53,128,78,0,0         // vbroadcastss  0x4e80(%rip),%ymm14        # 64fc <_sk_callback_avx+0x1f9>
+  .byte  196,98,125,24,53,236,78,0,0         // vbroadcastss  0x4eec(%rip),%ymm14        # 6568 <_sk_callback_avx+0x1f9>
   .byte  196,65,28,88,230                    // vaddps        %ymm14,%ymm12,%ymm12
-  .byte  196,98,125,24,61,118,78,0,0         // vbroadcastss  0x4e76(%rip),%ymm15        # 6500 <_sk_callback_avx+0x1fd>
+  .byte  196,98,125,24,61,226,78,0,0         // vbroadcastss  0x4ee2(%rip),%ymm15        # 656c <_sk_callback_avx+0x1fd>
   .byte  196,65,52,88,207                    // vaddps        %ymm15,%ymm9,%ymm9
   .byte  196,65,124,83,201                   // vrcpps        %ymm9,%ymm9
   .byte  196,65,52,89,204                    // vmulps        %ymm12,%ymm9,%ymm9
-  .byte  196,98,125,24,37,98,78,0,0          // vbroadcastss  0x4e62(%rip),%ymm12        # 6504 <_sk_callback_avx+0x201>
+  .byte  196,98,125,24,37,206,78,0,0         // vbroadcastss  0x4ece(%rip),%ymm12        # 6570 <_sk_callback_avx+0x201>
   .byte  196,193,124,194,196,1               // vcmpltps      %ymm12,%ymm0,%ymm0
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  197,124,82,201                      // vrsqrtps      %ymm1,%ymm9
@@ -15443,7 +15574,7 @@ _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  196,98,125,24,29,200,77,0,0         // vbroadcastss  0x4dc8(%rip),%ymm11        # 6508 <_sk_callback_avx+0x205>
+  .byte  196,98,125,24,29,52,78,0,0          // vbroadcastss  0x4e34(%rip),%ymm11        # 6574 <_sk_callback_avx+0x205>
   .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
@@ -15453,19 +15584,19 @@ _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  196,98,125,24,29,161,77,0,0         // vbroadcastss  0x4da1(%rip),%ymm11        # 6514 <_sk_callback_avx+0x211>
+  .byte  196,98,125,24,29,13,78,0,0          // vbroadcastss  0x4e0d(%rip),%ymm11        # 6580 <_sk_callback_avx+0x211>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,143,77,0,0         // vbroadcastss  0x4d8f(%rip),%ymm11        # 6510 <_sk_callback_avx+0x20d>
+  .byte  196,98,125,24,29,251,77,0,0         // vbroadcastss  0x4dfb(%rip),%ymm11        # 657c <_sk_callback_avx+0x20d>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
   .byte  196,227,117,74,202,224              // vblendvps     %ymm14,%ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,119,77,0,0        // vbroadcastss  0x4d77(%rip),%ymm2        # 650c <_sk_callback_avx+0x209>
+  .byte  196,226,125,24,21,227,77,0,0        // vbroadcastss  0x4de3(%rip),%ymm2        # 6578 <_sk_callback_avx+0x209>
   .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  196,98,125,24,37,94,77,0,0          // vbroadcastss  0x4d5e(%rip),%ymm12        # 651c <_sk_callback_avx+0x219>
+  .byte  196,98,125,24,37,202,77,0,0         // vbroadcastss  0x4dca(%rip),%ymm12        # 6588 <_sk_callback_avx+0x219>
   .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
@@ -15475,7 +15606,7 @@ _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  196,98,125,24,5,33,77,0,0           // vbroadcastss  0x4d21(%rip),%ymm8        # 6518 <_sk_callback_avx+0x215>
+  .byte  196,98,125,24,5,141,77,0,0          // vbroadcastss  0x4d8d(%rip),%ymm8        # 6584 <_sk_callback_avx+0x215>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15492,7 +15623,7 @@ _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  196,98,125,24,5,238,76,0,0          // vbroadcastss  0x4cee(%rip),%ymm8        # 6520 <_sk_callback_avx+0x21d>
+  .byte  196,98,125,24,5,90,77,0,0           // vbroadcastss  0x4d5a(%rip),%ymm8        # 658c <_sk_callback_avx+0x21d>
   .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
@@ -15500,23 +15631,23 @@ _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  196,98,125,24,29,200,76,0,0         // vbroadcastss  0x4cc8(%rip),%ymm11        # 6524 <_sk_callback_avx+0x221>
+  .byte  196,98,125,24,29,52,77,0,0          // vbroadcastss  0x4d34(%rip),%ymm11        # 6590 <_sk_callback_avx+0x221>
   .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  196,98,125,24,37,178,76,0,0         // vbroadcastss  0x4cb2(%rip),%ymm12        # 652c <_sk_callback_avx+0x229>
+  .byte  196,98,125,24,37,30,77,0,0          // vbroadcastss  0x4d1e(%rip),%ymm12        # 6598 <_sk_callback_avx+0x229>
   .byte  196,193,4,89,196                    // vmulps        %ymm12,%ymm15,%ymm0
-  .byte  196,98,125,24,45,168,76,0,0         // vbroadcastss  0x4ca8(%rip),%ymm13        # 6530 <_sk_callback_avx+0x22d>
+  .byte  196,98,125,24,45,20,77,0,0          // vbroadcastss  0x4d14(%rip),%ymm13        # 659c <_sk_callback_avx+0x22d>
   .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  196,226,125,24,13,137,76,0,0        // vbroadcastss  0x4c89(%rip),%ymm1        # 6528 <_sk_callback_avx+0x225>
+  .byte  196,226,125,24,13,245,76,0,0        // vbroadcastss  0x4cf5(%rip),%ymm1        # 6594 <_sk_callback_avx+0x225>
   .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  196,98,125,24,53,116,76,0,0         // vbroadcastss  0x4c74(%rip),%ymm14        # 6534 <_sk_callback_avx+0x231>
+  .byte  196,98,125,24,53,224,76,0,0         // vbroadcastss  0x4ce0(%rip),%ymm14        # 65a0 <_sk_callback_avx+0x231>
   .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
@@ -15535,7 +15666,7 @@ _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  196,226,125,24,5,22,76,0,0          // vbroadcastss  0x4c16(%rip),%ymm0        # 6538 <_sk_callback_avx+0x235>
+  .byte  196,226,125,24,5,130,76,0,0         // vbroadcastss  0x4c82(%rip),%ymm0        # 65a4 <_sk_callback_avx+0x235>
   .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
@@ -15594,7 +15725,7 @@ _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  196,98,125,24,13,63,75,0,0          // vbroadcastss  0x4b3f(%rip),%ymm9        # 653c <_sk_callback_avx+0x239>
+  .byte  196,98,125,24,13,171,75,0,0         // vbroadcastss  0x4bab(%rip),%ymm9        # 65a8 <_sk_callback_avx+0x239>
   .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
@@ -15653,7 +15784,7 @@ _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  196,98,125,24,13,139,74,0,0         // vbroadcastss  0x4a8b(%rip),%ymm9        # 6540 <_sk_callback_avx+0x23d>
+  .byte  196,98,125,24,13,247,74,0,0         // vbroadcastss  0x4af7(%rip),%ymm9        # 65ac <_sk_callback_avx+0x23d>
   .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
@@ -15696,20 +15827,20 @@ _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  196,98,125,24,13,245,73,0,0         // vbroadcastss  0x49f5(%rip),%ymm9        # 6544 <_sk_callback_avx+0x241>
+  .byte  196,98,125,24,13,97,74,0,0          // vbroadcastss  0x4a61(%rip),%ymm9        # 65b0 <_sk_callback_avx+0x241>
   .byte  196,65,60,84,201                    // vandps        %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,230,73,0,0         // vbroadcastss  0x49e6(%rip),%ymm10        # 6548 <_sk_callback_avx+0x245>
+  .byte  196,98,125,24,21,82,74,0,0          // vbroadcastss  0x4a52(%rip),%ymm10        # 65b4 <_sk_callback_avx+0x245>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,21,220,73,0,0         // vbroadcastss  0x49dc(%rip),%ymm10        # 654c <_sk_callback_avx+0x249>
+  .byte  196,98,125,24,21,72,74,0,0          // vbroadcastss  0x4a48(%rip),%ymm10        # 65b8 <_sk_callback_avx+0x249>
   .byte  196,65,60,84,210                    // vandps        %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,205,73,0,0         // vbroadcastss  0x49cd(%rip),%ymm11        # 6550 <_sk_callback_avx+0x24d>
+  .byte  196,98,125,24,29,57,74,0,0          // vbroadcastss  0x4a39(%rip),%ymm11        # 65bc <_sk_callback_avx+0x24d>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,29,195,73,0,0         // vbroadcastss  0x49c3(%rip),%ymm11        # 6554 <_sk_callback_avx+0x251>
+  .byte  196,98,125,24,29,47,74,0,0          // vbroadcastss  0x4a2f(%rip),%ymm11        # 65c0 <_sk_callback_avx+0x251>
   .byte  196,65,60,84,195                    // vandps        %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,180,73,0,0         // vbroadcastss  0x49b4(%rip),%ymm11        # 6558 <_sk_callback_avx+0x255>
+  .byte  196,98,125,24,29,32,74,0,0          // vbroadcastss  0x4a20(%rip),%ymm11        # 65c4 <_sk_callback_avx+0x255>
   .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
@@ -15756,7 +15887,7 @@ _sk_lerp_565_avx:
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,225                 // jmpq          ffffffffe2001c74 <_sk_callback_avx+0xffffffffe1ffb971>
+  .byte  233,255,255,255,225                 // jmpq          ffffffffe2001c74 <_sk_callback_avx+0xffffffffe1ffb905>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
@@ -15789,7 +15920,7 @@ _sk_load_tables_avx:
   .byte  65,85                               // push          %r13
   .byte  65,84                               // push          %r12
   .byte  83                                  // push          %rbx
-  .byte  197,124,40,13,146,75,0,0            // vmovaps       0x4b92(%rip),%ymm9        # 6840 <_sk_callback_avx+0x53d>
+  .byte  197,124,40,13,18,76,0,0             // vmovaps       0x4c12(%rip),%ymm9        # 68c0 <_sk_callback_avx+0x551>
   .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
@@ -15881,7 +16012,7 @@ _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  196,98,125,24,5,191,70,0,0          // vbroadcastss  0x46bf(%rip),%ymm8        # 655c <_sk_callback_avx+0x259>
+  .byte  196,98,125,24,5,43,71,0,0           // vbroadcastss  0x472b(%rip),%ymm8        # 65c8 <_sk_callback_avx+0x259>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -15973,7 +16104,7 @@ _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  197,121,111,29,210,72,0,0           // vmovdqa       0x48d2(%rip),%xmm11        # 68c0 <_sk_callback_avx+0x5bd>
+  .byte  197,121,111,29,82,73,0,0            // vmovdqa       0x4952(%rip),%xmm11        # 6940 <_sk_callback_avx+0x5d1>
   .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
@@ -16072,7 +16203,7 @@ _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  196,98,125,24,5,112,67,0,0          // vbroadcastss  0x4370(%rip),%ymm8        # 6560 <_sk_callback_avx+0x25d>
+  .byte  196,98,125,24,5,220,67,0,0          // vbroadcastss  0x43dc(%rip),%ymm8        # 65cc <_sk_callback_avx+0x25d>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -16144,7 +16275,7 @@ _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  197,121,111,13,203,69,0,0           // vmovdqa       0x45cb(%rip),%xmm9        # 68d0 <_sk_callback_avx+0x5cd>
+  .byte  197,121,111,13,75,70,0,0            // vmovdqa       0x464b(%rip),%xmm9        # 6950 <_sk_callback_avx+0x5e1>
   .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
@@ -16236,7 +16367,7 @@ _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  196,226,125,24,29,130,64,0,0        // vbroadcastss  0x4082(%rip),%ymm3        # 6564 <_sk_callback_avx+0x261>
+  .byte  196,226,125,24,29,238,64,0,0        // vbroadcastss  0x40ee(%rip),%ymm3        # 65d0 <_sk_callback_avx+0x261>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,93                               // pop           %r13
@@ -16289,7 +16420,7 @@ _sk_byte_tables_avx:
   .byte  65,84                               // push          %r12
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,182,63,0,0          // vbroadcastss  0x3fb6(%rip),%ymm8        # 6568 <_sk_callback_avx+0x265>
+  .byte  196,98,125,24,5,34,64,0,0           // vbroadcastss  0x4022(%rip),%ymm8        # 65d4 <_sk_callback_avx+0x265>
   .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
@@ -16326,7 +16457,7 @@ _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  196,98,125,24,13,4,63,0,0           // vbroadcastss  0x3f04(%rip),%ymm9        # 656c <_sk_callback_avx+0x269>
+  .byte  196,98,125,24,13,112,63,0,0         // vbroadcastss  0x3f70(%rip),%ymm9        # 65d8 <_sk_callback_avx+0x269>
   .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
@@ -16488,7 +16619,7 @@ _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  196,98,125,24,13,42,60,0,0          // vbroadcastss  0x3c2a(%rip),%ymm9        # 6570 <_sk_callback_avx+0x26d>
+  .byte  196,98,125,24,13,150,60,0,0         // vbroadcastss  0x3c96(%rip),%ymm9        # 65dc <_sk_callback_avx+0x26d>
   .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
@@ -16785,36 +16916,36 @@ _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  196,98,125,24,37,136,55,0,0         // vbroadcastss  0x3788(%rip),%ymm12        # 6574 <_sk_callback_avx+0x271>
+  .byte  196,98,125,24,37,244,55,0,0         // vbroadcastss  0x37f4(%rip),%ymm12        # 65e0 <_sk_callback_avx+0x271>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,126,55,0,0         // vbroadcastss  0x377e(%rip),%ymm12        # 6578 <_sk_callback_avx+0x275>
+  .byte  196,98,125,24,37,234,55,0,0         // vbroadcastss  0x37ea(%rip),%ymm12        # 65e4 <_sk_callback_avx+0x275>
   .byte  196,193,124,84,196                  // vandps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,116,55,0,0         // vbroadcastss  0x3774(%rip),%ymm12        # 657c <_sk_callback_avx+0x279>
+  .byte  196,98,125,24,37,224,55,0,0         // vbroadcastss  0x37e0(%rip),%ymm12        # 65e8 <_sk_callback_avx+0x279>
   .byte  196,193,124,86,196                  // vorps         %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,106,55,0,0         // vbroadcastss  0x376a(%rip),%ymm12        # 6580 <_sk_callback_avx+0x27d>
+  .byte  196,98,125,24,37,214,55,0,0         // vbroadcastss  0x37d6(%rip),%ymm12        # 65ec <_sk_callback_avx+0x27d>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,96,55,0,0          // vbroadcastss  0x3760(%rip),%ymm12        # 6584 <_sk_callback_avx+0x281>
+  .byte  196,98,125,24,37,204,55,0,0         // vbroadcastss  0x37cc(%rip),%ymm12        # 65f0 <_sk_callback_avx+0x281>
   .byte  196,65,124,89,228                   // vmulps        %ymm12,%ymm0,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,81,55,0,0          // vbroadcastss  0x3751(%rip),%ymm12        # 6588 <_sk_callback_avx+0x285>
+  .byte  196,98,125,24,37,189,55,0,0         // vbroadcastss  0x37bd(%rip),%ymm12        # 65f4 <_sk_callback_avx+0x285>
   .byte  196,193,124,88,196                  // vaddps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,71,55,0,0          // vbroadcastss  0x3747(%rip),%ymm12        # 658c <_sk_callback_avx+0x289>
+  .byte  196,98,125,24,37,179,55,0,0         // vbroadcastss  0x37b3(%rip),%ymm12        # 65f8 <_sk_callback_avx+0x289>
   .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  196,98,125,24,29,43,55,0,0          // vbroadcastss  0x372b(%rip),%ymm11        # 6590 <_sk_callback_avx+0x28d>
+  .byte  196,98,125,24,29,151,55,0,0         // vbroadcastss  0x3797(%rip),%ymm11        # 65fc <_sk_callback_avx+0x28d>
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,33,55,0,0          // vbroadcastss  0x3721(%rip),%ymm11        # 6594 <_sk_callback_avx+0x291>
+  .byte  196,98,125,24,29,141,55,0,0         // vbroadcastss  0x378d(%rip),%ymm11        # 6600 <_sk_callback_avx+0x291>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,124,92,195                  // vsubps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,18,55,0,0          // vbroadcastss  0x3712(%rip),%ymm11        # 6598 <_sk_callback_avx+0x295>
+  .byte  196,98,125,24,29,126,55,0,0         // vbroadcastss  0x377e(%rip),%ymm11        # 6604 <_sk_callback_avx+0x295>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,8,55,0,0           // vbroadcastss  0x3708(%rip),%ymm11        # 659c <_sk_callback_avx+0x299>
+  .byte  196,98,125,24,29,116,55,0,0         // vbroadcastss  0x3774(%rip),%ymm11        # 6608 <_sk_callback_avx+0x299>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,124,88,194                  // vaddps        %ymm10,%ymm0,%ymm0
-  .byte  196,98,125,24,21,249,54,0,0         // vbroadcastss  0x36f9(%rip),%ymm10        # 65a0 <_sk_callback_avx+0x29d>
+  .byte  196,98,125,24,21,101,55,0,0         // vbroadcastss  0x3765(%rip),%ymm10        # 660c <_sk_callback_avx+0x29d>
   .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
@@ -16822,7 +16953,7 @@ _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  196,98,125,24,5,208,54,0,0          // vbroadcastss  0x36d0(%rip),%ymm8        # 65a4 <_sk_callback_avx+0x2a1>
+  .byte  196,98,125,24,5,60,55,0,0           // vbroadcastss  0x373c(%rip),%ymm8        # 6610 <_sk_callback_avx+0x2a1>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16844,36 +16975,36 @@ _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  196,98,125,24,37,129,54,0,0         // vbroadcastss  0x3681(%rip),%ymm12        # 65a8 <_sk_callback_avx+0x2a5>
+  .byte  196,98,125,24,37,237,54,0,0         // vbroadcastss  0x36ed(%rip),%ymm12        # 6614 <_sk_callback_avx+0x2a5>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,119,54,0,0         // vbroadcastss  0x3677(%rip),%ymm12        # 65ac <_sk_callback_avx+0x2a9>
+  .byte  196,98,125,24,37,227,54,0,0         // vbroadcastss  0x36e3(%rip),%ymm12        # 6618 <_sk_callback_avx+0x2a9>
   .byte  196,193,116,84,204                  // vandps        %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,109,54,0,0         // vbroadcastss  0x366d(%rip),%ymm12        # 65b0 <_sk_callback_avx+0x2ad>
+  .byte  196,98,125,24,37,217,54,0,0         // vbroadcastss  0x36d9(%rip),%ymm12        # 661c <_sk_callback_avx+0x2ad>
   .byte  196,193,116,86,204                  // vorps         %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,99,54,0,0          // vbroadcastss  0x3663(%rip),%ymm12        # 65b4 <_sk_callback_avx+0x2b1>
+  .byte  196,98,125,24,37,207,54,0,0         // vbroadcastss  0x36cf(%rip),%ymm12        # 6620 <_sk_callback_avx+0x2b1>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,89,54,0,0          // vbroadcastss  0x3659(%rip),%ymm12        # 65b8 <_sk_callback_avx+0x2b5>
+  .byte  196,98,125,24,37,197,54,0,0         // vbroadcastss  0x36c5(%rip),%ymm12        # 6624 <_sk_callback_avx+0x2b5>
   .byte  196,65,116,89,228                   // vmulps        %ymm12,%ymm1,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,74,54,0,0          // vbroadcastss  0x364a(%rip),%ymm12        # 65bc <_sk_callback_avx+0x2b9>
+  .byte  196,98,125,24,37,182,54,0,0         // vbroadcastss  0x36b6(%rip),%ymm12        # 6628 <_sk_callback_avx+0x2b9>
   .byte  196,193,116,88,204                  // vaddps        %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,64,54,0,0          // vbroadcastss  0x3640(%rip),%ymm12        # 65c0 <_sk_callback_avx+0x2bd>
+  .byte  196,98,125,24,37,172,54,0,0         // vbroadcastss  0x36ac(%rip),%ymm12        # 662c <_sk_callback_avx+0x2bd>
   .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  196,98,125,24,29,36,54,0,0          // vbroadcastss  0x3624(%rip),%ymm11        # 65c4 <_sk_callback_avx+0x2c1>
+  .byte  196,98,125,24,29,144,54,0,0         // vbroadcastss  0x3690(%rip),%ymm11        # 6630 <_sk_callback_avx+0x2c1>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,26,54,0,0          // vbroadcastss  0x361a(%rip),%ymm11        # 65c8 <_sk_callback_avx+0x2c5>
+  .byte  196,98,125,24,29,134,54,0,0         // vbroadcastss  0x3686(%rip),%ymm11        # 6634 <_sk_callback_avx+0x2c5>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,116,92,203                  // vsubps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,11,54,0,0          // vbroadcastss  0x360b(%rip),%ymm11        # 65cc <_sk_callback_avx+0x2c9>
+  .byte  196,98,125,24,29,119,54,0,0         // vbroadcastss  0x3677(%rip),%ymm11        # 6638 <_sk_callback_avx+0x2c9>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,1,54,0,0           // vbroadcastss  0x3601(%rip),%ymm11        # 65d0 <_sk_callback_avx+0x2cd>
+  .byte  196,98,125,24,29,109,54,0,0         // vbroadcastss  0x366d(%rip),%ymm11        # 663c <_sk_callback_avx+0x2cd>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,116,88,202                  // vaddps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,21,242,53,0,0         // vbroadcastss  0x35f2(%rip),%ymm10        # 65d4 <_sk_callback_avx+0x2d1>
+  .byte  196,98,125,24,21,94,54,0,0          // vbroadcastss  0x365e(%rip),%ymm10        # 6640 <_sk_callback_avx+0x2d1>
   .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
@@ -16881,7 +17012,7 @@ _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  196,98,125,24,5,201,53,0,0          // vbroadcastss  0x35c9(%rip),%ymm8        # 65d8 <_sk_callback_avx+0x2d5>
+  .byte  196,98,125,24,5,53,54,0,0           // vbroadcastss  0x3635(%rip),%ymm8        # 6644 <_sk_callback_avx+0x2d5>
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16903,36 +17034,36 @@ _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  196,98,125,24,37,122,53,0,0         // vbroadcastss  0x357a(%rip),%ymm12        # 65dc <_sk_callback_avx+0x2d9>
+  .byte  196,98,125,24,37,230,53,0,0         // vbroadcastss  0x35e6(%rip),%ymm12        # 6648 <_sk_callback_avx+0x2d9>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,112,53,0,0         // vbroadcastss  0x3570(%rip),%ymm12        # 65e0 <_sk_callback_avx+0x2dd>
+  .byte  196,98,125,24,37,220,53,0,0         // vbroadcastss  0x35dc(%rip),%ymm12        # 664c <_sk_callback_avx+0x2dd>
   .byte  196,193,108,84,212                  // vandps        %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,102,53,0,0         // vbroadcastss  0x3566(%rip),%ymm12        # 65e4 <_sk_callback_avx+0x2e1>
+  .byte  196,98,125,24,37,210,53,0,0         // vbroadcastss  0x35d2(%rip),%ymm12        # 6650 <_sk_callback_avx+0x2e1>
   .byte  196,193,108,86,212                  // vorps         %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,92,53,0,0          // vbroadcastss  0x355c(%rip),%ymm12        # 65e8 <_sk_callback_avx+0x2e5>
+  .byte  196,98,125,24,37,200,53,0,0         // vbroadcastss  0x35c8(%rip),%ymm12        # 6654 <_sk_callback_avx+0x2e5>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,82,53,0,0          // vbroadcastss  0x3552(%rip),%ymm12        # 65ec <_sk_callback_avx+0x2e9>
+  .byte  196,98,125,24,37,190,53,0,0         // vbroadcastss  0x35be(%rip),%ymm12        # 6658 <_sk_callback_avx+0x2e9>
   .byte  196,65,108,89,228                   // vmulps        %ymm12,%ymm2,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,67,53,0,0          // vbroadcastss  0x3543(%rip),%ymm12        # 65f0 <_sk_callback_avx+0x2ed>
+  .byte  196,98,125,24,37,175,53,0,0         // vbroadcastss  0x35af(%rip),%ymm12        # 665c <_sk_callback_avx+0x2ed>
   .byte  196,193,108,88,212                  // vaddps        %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,57,53,0,0          // vbroadcastss  0x3539(%rip),%ymm12        # 65f4 <_sk_callback_avx+0x2f1>
+  .byte  196,98,125,24,37,165,53,0,0         // vbroadcastss  0x35a5(%rip),%ymm12        # 6660 <_sk_callback_avx+0x2f1>
   .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  196,98,125,24,29,29,53,0,0          // vbroadcastss  0x351d(%rip),%ymm11        # 65f8 <_sk_callback_avx+0x2f5>
+  .byte  196,98,125,24,29,137,53,0,0         // vbroadcastss  0x3589(%rip),%ymm11        # 6664 <_sk_callback_avx+0x2f5>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,19,53,0,0          // vbroadcastss  0x3513(%rip),%ymm11        # 65fc <_sk_callback_avx+0x2f9>
+  .byte  196,98,125,24,29,127,53,0,0         // vbroadcastss  0x357f(%rip),%ymm11        # 6668 <_sk_callback_avx+0x2f9>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,108,92,211                  // vsubps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,4,53,0,0           // vbroadcastss  0x3504(%rip),%ymm11        # 6600 <_sk_callback_avx+0x2fd>
+  .byte  196,98,125,24,29,112,53,0,0         // vbroadcastss  0x3570(%rip),%ymm11        # 666c <_sk_callback_avx+0x2fd>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,250,52,0,0         // vbroadcastss  0x34fa(%rip),%ymm11        # 6604 <_sk_callback_avx+0x301>
+  .byte  196,98,125,24,29,102,53,0,0         // vbroadcastss  0x3566(%rip),%ymm11        # 6670 <_sk_callback_avx+0x301>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,108,88,210                  // vaddps        %ymm10,%ymm2,%ymm2
-  .byte  196,98,125,24,21,235,52,0,0         // vbroadcastss  0x34eb(%rip),%ymm10        # 6608 <_sk_callback_avx+0x305>
+  .byte  196,98,125,24,21,87,53,0,0          // vbroadcastss  0x3557(%rip),%ymm10        # 6674 <_sk_callback_avx+0x305>
   .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
@@ -16940,7 +17071,7 @@ _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  196,98,125,24,5,194,52,0,0          // vbroadcastss  0x34c2(%rip),%ymm8        # 660c <_sk_callback_avx+0x309>
+  .byte  196,98,125,24,5,46,53,0,0           // vbroadcastss  0x352e(%rip),%ymm8        # 6678 <_sk_callback_avx+0x309>
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -16962,36 +17093,36 @@ _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  196,98,125,24,37,115,52,0,0         // vbroadcastss  0x3473(%rip),%ymm12        # 6610 <_sk_callback_avx+0x30d>
+  .byte  196,98,125,24,37,223,52,0,0         // vbroadcastss  0x34df(%rip),%ymm12        # 667c <_sk_callback_avx+0x30d>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,105,52,0,0         // vbroadcastss  0x3469(%rip),%ymm12        # 6614 <_sk_callback_avx+0x311>
+  .byte  196,98,125,24,37,213,52,0,0         // vbroadcastss  0x34d5(%rip),%ymm12        # 6680 <_sk_callback_avx+0x311>
   .byte  196,193,100,84,220                  // vandps        %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,95,52,0,0          // vbroadcastss  0x345f(%rip),%ymm12        # 6618 <_sk_callback_avx+0x315>
+  .byte  196,98,125,24,37,203,52,0,0         // vbroadcastss  0x34cb(%rip),%ymm12        # 6684 <_sk_callback_avx+0x315>
   .byte  196,193,100,86,220                  // vorps         %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,85,52,0,0          // vbroadcastss  0x3455(%rip),%ymm12        # 661c <_sk_callback_avx+0x319>
+  .byte  196,98,125,24,37,193,52,0,0         // vbroadcastss  0x34c1(%rip),%ymm12        # 6688 <_sk_callback_avx+0x319>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,75,52,0,0          // vbroadcastss  0x344b(%rip),%ymm12        # 6620 <_sk_callback_avx+0x31d>
+  .byte  196,98,125,24,37,183,52,0,0         // vbroadcastss  0x34b7(%rip),%ymm12        # 668c <_sk_callback_avx+0x31d>
   .byte  196,65,100,89,228                   // vmulps        %ymm12,%ymm3,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,60,52,0,0          // vbroadcastss  0x343c(%rip),%ymm12        # 6624 <_sk_callback_avx+0x321>
+  .byte  196,98,125,24,37,168,52,0,0         // vbroadcastss  0x34a8(%rip),%ymm12        # 6690 <_sk_callback_avx+0x321>
   .byte  196,193,100,88,220                  // vaddps        %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,50,52,0,0          // vbroadcastss  0x3432(%rip),%ymm12        # 6628 <_sk_callback_avx+0x325>
+  .byte  196,98,125,24,37,158,52,0,0         // vbroadcastss  0x349e(%rip),%ymm12        # 6694 <_sk_callback_avx+0x325>
   .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  196,98,125,24,29,22,52,0,0          // vbroadcastss  0x3416(%rip),%ymm11        # 662c <_sk_callback_avx+0x329>
+  .byte  196,98,125,24,29,130,52,0,0         // vbroadcastss  0x3482(%rip),%ymm11        # 6698 <_sk_callback_avx+0x329>
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,12,52,0,0          // vbroadcastss  0x340c(%rip),%ymm11        # 6630 <_sk_callback_avx+0x32d>
+  .byte  196,98,125,24,29,120,52,0,0         // vbroadcastss  0x3478(%rip),%ymm11        # 669c <_sk_callback_avx+0x32d>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,100,92,219                  // vsubps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,253,51,0,0         // vbroadcastss  0x33fd(%rip),%ymm11        # 6634 <_sk_callback_avx+0x331>
+  .byte  196,98,125,24,29,105,52,0,0         // vbroadcastss  0x3469(%rip),%ymm11        # 66a0 <_sk_callback_avx+0x331>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,243,51,0,0         // vbroadcastss  0x33f3(%rip),%ymm11        # 6638 <_sk_callback_avx+0x335>
+  .byte  196,98,125,24,29,95,52,0,0          // vbroadcastss  0x345f(%rip),%ymm11        # 66a4 <_sk_callback_avx+0x335>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,100,88,218                  // vaddps        %ymm10,%ymm3,%ymm3
-  .byte  196,98,125,24,21,228,51,0,0         // vbroadcastss  0x33e4(%rip),%ymm10        # 663c <_sk_callback_avx+0x339>
+  .byte  196,98,125,24,21,80,52,0,0          // vbroadcastss  0x3450(%rip),%ymm10        # 66a8 <_sk_callback_avx+0x339>
   .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
@@ -16999,7 +17130,7 @@ _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  196,98,125,24,5,187,51,0,0          // vbroadcastss  0x33bb(%rip),%ymm8        # 6640 <_sk_callback_avx+0x33d>
+  .byte  196,98,125,24,5,39,52,0,0           // vbroadcastss  0x3427(%rip),%ymm8        # 66ac <_sk_callback_avx+0x33d>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -17008,31 +17139,31 @@ HIDDEN _sk_lab_to_xyz_avx
 .globl _sk_lab_to_xyz_avx
 FUNCTION(_sk_lab_to_xyz_avx)
 _sk_lab_to_xyz_avx:
-  .byte  196,98,125,24,5,173,51,0,0          // vbroadcastss  0x33ad(%rip),%ymm8        # 6644 <_sk_callback_avx+0x341>
+  .byte  196,98,125,24,5,25,52,0,0           // vbroadcastss  0x3419(%rip),%ymm8        # 66b0 <_sk_callback_avx+0x341>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,163,51,0,0          // vbroadcastss  0x33a3(%rip),%ymm8        # 6648 <_sk_callback_avx+0x345>
+  .byte  196,98,125,24,5,15,52,0,0           // vbroadcastss  0x340f(%rip),%ymm8        # 66b4 <_sk_callback_avx+0x345>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,98,125,24,13,153,51,0,0         // vbroadcastss  0x3399(%rip),%ymm9        # 664c <_sk_callback_avx+0x349>
+  .byte  196,98,125,24,13,5,52,0,0           // vbroadcastss  0x3405(%rip),%ymm9        # 66b8 <_sk_callback_avx+0x349>
   .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  196,98,125,24,5,133,51,0,0          // vbroadcastss  0x3385(%rip),%ymm8        # 6650 <_sk_callback_avx+0x34d>
+  .byte  196,98,125,24,5,241,51,0,0          // vbroadcastss  0x33f1(%rip),%ymm8        # 66bc <_sk_callback_avx+0x34d>
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,123,51,0,0          // vbroadcastss  0x337b(%rip),%ymm8        # 6654 <_sk_callback_avx+0x351>
+  .byte  196,98,125,24,5,231,51,0,0          // vbroadcastss  0x33e7(%rip),%ymm8        # 66c0 <_sk_callback_avx+0x351>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,113,51,0,0          // vbroadcastss  0x3371(%rip),%ymm8        # 6658 <_sk_callback_avx+0x355>
+  .byte  196,98,125,24,5,221,51,0,0          // vbroadcastss  0x33dd(%rip),%ymm8        # 66c4 <_sk_callback_avx+0x355>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  197,252,88,201                      // vaddps        %ymm1,%ymm0,%ymm1
-  .byte  196,98,125,24,5,99,51,0,0           // vbroadcastss  0x3363(%rip),%ymm8        # 665c <_sk_callback_avx+0x359>
+  .byte  196,98,125,24,5,207,51,0,0          // vbroadcastss  0x33cf(%rip),%ymm8        # 66c8 <_sk_callback_avx+0x359>
   .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  196,98,125,24,13,76,51,0,0          // vbroadcastss  0x334c(%rip),%ymm9        # 6660 <_sk_callback_avx+0x35d>
+  .byte  196,98,125,24,13,184,51,0,0         // vbroadcastss  0x33b8(%rip),%ymm9        # 66cc <_sk_callback_avx+0x35d>
   .byte  196,65,52,194,208,1                 // vcmpltps      %ymm8,%ymm9,%ymm10
-  .byte  196,98,125,24,29,65,51,0,0          // vbroadcastss  0x3341(%rip),%ymm11        # 6664 <_sk_callback_avx+0x361>
+  .byte  196,98,125,24,29,173,51,0,0         // vbroadcastss  0x33ad(%rip),%ymm11        # 66d0 <_sk_callback_avx+0x361>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,37,55,51,0,0          // vbroadcastss  0x3337(%rip),%ymm12        # 6668 <_sk_callback_avx+0x365>
+  .byte  196,98,125,24,37,163,51,0,0         // vbroadcastss  0x33a3(%rip),%ymm12        # 66d4 <_sk_callback_avx+0x365>
   .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
@@ -17047,9 +17178,9 @@ _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  196,226,125,24,5,237,50,0,0         // vbroadcastss  0x32ed(%rip),%ymm0        # 666c <_sk_callback_avx+0x369>
+  .byte  196,226,125,24,5,89,51,0,0          // vbroadcastss  0x3359(%rip),%ymm0        # 66d8 <_sk_callback_avx+0x369>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,228,50,0,0          // vbroadcastss  0x32e4(%rip),%ymm8        # 6670 <_sk_callback_avx+0x36d>
+  .byte  196,98,125,24,5,80,51,0,0           // vbroadcastss  0x3350(%rip),%ymm8        # 66dc <_sk_callback_avx+0x36d>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -17070,7 +17201,7 @@ _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  196,226,125,24,13,168,50,0,0        // vbroadcastss  0x32a8(%rip),%ymm1        # 6674 <_sk_callback_avx+0x371>
+  .byte  196,226,125,24,13,20,51,0,0         // vbroadcastss  0x3314(%rip),%ymm1        # 66e0 <_sk_callback_avx+0x371>
   .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
@@ -17139,7 +17270,7 @@ _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  196,226,125,24,13,157,49,0,0        // vbroadcastss  0x319d(%rip),%ymm1        # 6678 <_sk_callback_avx+0x375>
+  .byte  196,226,125,24,13,9,50,0,0          // vbroadcastss  0x3209(%rip),%ymm1        # 66e4 <_sk_callback_avx+0x375>
   .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
@@ -17157,7 +17288,7 @@ FUNCTION(_sk_store_a8_avx)
 _sk_store_a8_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,120,49,0,0          // vbroadcastss  0x3178(%rip),%ymm8        # 667c <_sk_callback_avx+0x379>
+  .byte  196,98,125,24,5,228,49,0,0          // vbroadcastss  0x31e4(%rip),%ymm8        # 66e8 <_sk_callback_avx+0x379>
   .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
@@ -17227,10 +17358,10 @@ _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  196,226,125,24,13,157,48,0,0        // vbroadcastss  0x309d(%rip),%ymm1        # 6680 <_sk_callback_avx+0x37d>
+  .byte  196,226,125,24,13,9,49,0,0          // vbroadcastss  0x3109(%rip),%ymm1        # 66ec <_sk_callback_avx+0x37d>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,146,48,0,0        // vbroadcastss  0x3092(%rip),%ymm3        # 6684 <_sk_callback_avx+0x381>
+  .byte  196,226,125,24,29,254,48,0,0        // vbroadcastss  0x30fe(%rip),%ymm3        # 66f0 <_sk_callback_avx+0x381>
   .byte  76,137,193                          // mov           %r8,%rcx
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
@@ -17296,10 +17427,10 @@ _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  196,226,125,24,13,145,47,0,0        // vbroadcastss  0x2f91(%rip),%ymm1        # 6688 <_sk_callback_avx+0x385>
+  .byte  196,226,125,24,13,253,47,0,0        // vbroadcastss  0x2ffd(%rip),%ymm1        # 66f4 <_sk_callback_avx+0x385>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,134,47,0,0        // vbroadcastss  0x2f86(%rip),%ymm3        # 668c <_sk_callback_avx+0x389>
+  .byte  196,226,125,24,29,242,47,0,0        // vbroadcastss  0x2ff2(%rip),%ymm3        # 66f8 <_sk_callback_avx+0x389>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -17379,10 +17510,10 @@ _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  197,124,40,21,14,48,0,0             // vmovaps       0x300e(%rip),%ymm10        # 6860 <_sk_callback_avx+0x55d>
+  .byte  197,124,40,21,142,48,0,0            // vmovaps       0x308e(%rip),%ymm10        # 68e0 <_sk_callback_avx+0x571>
   .byte  196,193,124,84,194                  // vandps        %ymm10,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,44,46,0,0          // vbroadcastss  0x2e2c(%rip),%ymm9        # 6690 <_sk_callback_avx+0x38d>
+  .byte  196,98,125,24,13,152,46,0,0         // vbroadcastss  0x2e98(%rip),%ymm9        # 66fc <_sk_callback_avx+0x38d>
   .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
@@ -17422,23 +17553,23 @@ _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  196,226,125,24,5,150,45,0,0         // vbroadcastss  0x2d96(%rip),%ymm0        # 6694 <_sk_callback_avx+0x391>
+  .byte  196,226,125,24,5,2,46,0,0           // vbroadcastss  0x2e02(%rip),%ymm0        # 6700 <_sk_callback_avx+0x391>
   .byte  197,236,84,192                      // vandps        %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,137,45,0,0        // vbroadcastss  0x2d89(%rip),%ymm1        # 6698 <_sk_callback_avx+0x395>
+  .byte  196,226,125,24,13,245,45,0,0        // vbroadcastss  0x2df5(%rip),%ymm1        # 6704 <_sk_callback_avx+0x395>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,128,45,0,0        // vbroadcastss  0x2d80(%rip),%ymm1        # 669c <_sk_callback_avx+0x399>
+  .byte  196,226,125,24,13,236,45,0,0        // vbroadcastss  0x2dec(%rip),%ymm1        # 6708 <_sk_callback_avx+0x399>
   .byte  197,236,84,201                      // vandps        %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,115,45,0,0        // vbroadcastss  0x2d73(%rip),%ymm3        # 66a0 <_sk_callback_avx+0x39d>
+  .byte  196,226,125,24,29,223,45,0,0        // vbroadcastss  0x2ddf(%rip),%ymm3        # 670c <_sk_callback_avx+0x39d>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,24,29,106,45,0,0        // vbroadcastss  0x2d6a(%rip),%ymm3        # 66a4 <_sk_callback_avx+0x3a1>
+  .byte  196,226,125,24,29,214,45,0,0        // vbroadcastss  0x2dd6(%rip),%ymm3        # 6710 <_sk_callback_avx+0x3a1>
   .byte  197,236,84,211                      // vandps        %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,93,45,0,0         // vbroadcastss  0x2d5d(%rip),%ymm3        # 66a8 <_sk_callback_avx+0x3a5>
+  .byte  196,226,125,24,29,201,45,0,0        // vbroadcastss  0x2dc9(%rip),%ymm3        # 6714 <_sk_callback_avx+0x3a5>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,82,45,0,0         // vbroadcastss  0x2d52(%rip),%ymm3        # 66ac <_sk_callback_avx+0x3a9>
+  .byte  196,226,125,24,29,190,45,0,0        // vbroadcastss  0x2dbe(%rip),%ymm3        # 6718 <_sk_callback_avx+0x3a9>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,200                          // mov           %ecx,%r8d
   .byte  65,128,224,7                        // and           $0x7,%r8b
@@ -17537,23 +17668,23 @@ _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  196,226,125,24,5,242,43,0,0         // vbroadcastss  0x2bf2(%rip),%ymm0        # 66b0 <_sk_callback_avx+0x3ad>
+  .byte  196,226,125,24,5,94,44,0,0          // vbroadcastss  0x2c5e(%rip),%ymm0        # 671c <_sk_callback_avx+0x3ad>
   .byte  197,236,84,192                      // vandps        %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,229,43,0,0        // vbroadcastss  0x2be5(%rip),%ymm1        # 66b4 <_sk_callback_avx+0x3b1>
+  .byte  196,226,125,24,13,81,44,0,0         // vbroadcastss  0x2c51(%rip),%ymm1        # 6720 <_sk_callback_avx+0x3b1>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,220,43,0,0        // vbroadcastss  0x2bdc(%rip),%ymm1        # 66b8 <_sk_callback_avx+0x3b5>
+  .byte  196,226,125,24,13,72,44,0,0         // vbroadcastss  0x2c48(%rip),%ymm1        # 6724 <_sk_callback_avx+0x3b5>
   .byte  197,236,84,201                      // vandps        %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,207,43,0,0        // vbroadcastss  0x2bcf(%rip),%ymm3        # 66bc <_sk_callback_avx+0x3b9>
+  .byte  196,226,125,24,29,59,44,0,0         // vbroadcastss  0x2c3b(%rip),%ymm3        # 6728 <_sk_callback_avx+0x3b9>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,24,29,198,43,0,0        // vbroadcastss  0x2bc6(%rip),%ymm3        # 66c0 <_sk_callback_avx+0x3bd>
+  .byte  196,226,125,24,29,50,44,0,0         // vbroadcastss  0x2c32(%rip),%ymm3        # 672c <_sk_callback_avx+0x3bd>
   .byte  197,236,84,211                      // vandps        %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,185,43,0,0        // vbroadcastss  0x2bb9(%rip),%ymm3        # 66c4 <_sk_callback_avx+0x3c1>
+  .byte  196,226,125,24,29,37,44,0,0         // vbroadcastss  0x2c25(%rip),%ymm3        # 6730 <_sk_callback_avx+0x3c1>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,174,43,0,0        // vbroadcastss  0x2bae(%rip),%ymm3        # 66c8 <_sk_callback_avx+0x3c5>
+  .byte  196,226,125,24,29,26,44,0,0         // vbroadcastss  0x2c1a(%rip),%ymm3        # 6734 <_sk_callback_avx+0x3c5>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -17567,14 +17698,14 @@ FUNCTION(_sk_store_565_avx)
 _sk_store_565_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,154,43,0,0          // vbroadcastss  0x2b9a(%rip),%ymm8        # 66cc <_sk_callback_avx+0x3c9>
+  .byte  196,98,125,24,5,6,44,0,0            // vbroadcastss  0x2c06(%rip),%ymm8        # 6738 <_sk_callback_avx+0x3c9>
   .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  196,98,125,24,21,115,43,0,0         // vbroadcastss  0x2b73(%rip),%ymm10        # 66d0 <_sk_callback_avx+0x3cd>
+  .byte  196,98,125,24,21,223,43,0,0         // vbroadcastss  0x2bdf(%rip),%ymm10        # 673c <_sk_callback_avx+0x3cd>
   .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
@@ -17648,25 +17779,25 @@ _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  196,226,125,24,5,124,42,0,0         // vbroadcastss  0x2a7c(%rip),%ymm0        # 66d4 <_sk_callback_avx+0x3d1>
+  .byte  196,226,125,24,5,232,42,0,0         // vbroadcastss  0x2ae8(%rip),%ymm0        # 6740 <_sk_callback_avx+0x3d1>
   .byte  197,228,84,192                      // vandps        %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,111,42,0,0        // vbroadcastss  0x2a6f(%rip),%ymm1        # 66d8 <_sk_callback_avx+0x3d5>
+  .byte  196,226,125,24,13,219,42,0,0        // vbroadcastss  0x2adb(%rip),%ymm1        # 6744 <_sk_callback_avx+0x3d5>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,102,42,0,0        // vbroadcastss  0x2a66(%rip),%ymm1        # 66dc <_sk_callback_avx+0x3d9>
+  .byte  196,226,125,24,13,210,42,0,0        // vbroadcastss  0x2ad2(%rip),%ymm1        # 6748 <_sk_callback_avx+0x3d9>
   .byte  197,228,84,201                      // vandps        %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,89,42,0,0         // vbroadcastss  0x2a59(%rip),%ymm2        # 66e0 <_sk_callback_avx+0x3dd>
+  .byte  196,226,125,24,21,197,42,0,0        // vbroadcastss  0x2ac5(%rip),%ymm2        # 674c <_sk_callback_avx+0x3dd>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,80,42,0,0         // vbroadcastss  0x2a50(%rip),%ymm2        # 66e4 <_sk_callback_avx+0x3e1>
+  .byte  196,226,125,24,21,188,42,0,0        // vbroadcastss  0x2abc(%rip),%ymm2        # 6750 <_sk_callback_avx+0x3e1>
   .byte  197,228,84,210                      // vandps        %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,67,42,0,0           // vbroadcastss  0x2a43(%rip),%ymm8        # 66e8 <_sk_callback_avx+0x3e5>
+  .byte  196,98,125,24,5,175,42,0,0          // vbroadcastss  0x2aaf(%rip),%ymm8        # 6754 <_sk_callback_avx+0x3e5>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,57,42,0,0           // vbroadcastss  0x2a39(%rip),%ymm8        # 66ec <_sk_callback_avx+0x3e9>
+  .byte  196,98,125,24,5,165,42,0,0          // vbroadcastss  0x2aa5(%rip),%ymm8        # 6758 <_sk_callback_avx+0x3e9>
   .byte  196,193,100,84,216                  // vandps        %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,43,42,0,0           // vbroadcastss  0x2a2b(%rip),%ymm8        # 66f0 <_sk_callback_avx+0x3ed>
+  .byte  196,98,125,24,5,151,42,0,0          // vbroadcastss  0x2a97(%rip),%ymm8        # 675c <_sk_callback_avx+0x3ed>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -17768,25 +17899,25 @@ _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  196,226,125,24,5,194,40,0,0         // vbroadcastss  0x28c2(%rip),%ymm0        # 66f4 <_sk_callback_avx+0x3f1>
+  .byte  196,226,125,24,5,46,41,0,0          // vbroadcastss  0x292e(%rip),%ymm0        # 6760 <_sk_callback_avx+0x3f1>
   .byte  197,228,84,192                      // vandps        %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,181,40,0,0        // vbroadcastss  0x28b5(%rip),%ymm1        # 66f8 <_sk_callback_avx+0x3f5>
+  .byte  196,226,125,24,13,33,41,0,0         // vbroadcastss  0x2921(%rip),%ymm1        # 6764 <_sk_callback_avx+0x3f5>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,172,40,0,0        // vbroadcastss  0x28ac(%rip),%ymm1        # 66fc <_sk_callback_avx+0x3f9>
+  .byte  196,226,125,24,13,24,41,0,0         // vbroadcastss  0x2918(%rip),%ymm1        # 6768 <_sk_callback_avx+0x3f9>
   .byte  197,228,84,201                      // vandps        %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,159,40,0,0        // vbroadcastss  0x289f(%rip),%ymm2        # 6700 <_sk_callback_avx+0x3fd>
+  .byte  196,226,125,24,21,11,41,0,0         // vbroadcastss  0x290b(%rip),%ymm2        # 676c <_sk_callback_avx+0x3fd>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,150,40,0,0        // vbroadcastss  0x2896(%rip),%ymm2        # 6704 <_sk_callback_avx+0x401>
+  .byte  196,226,125,24,21,2,41,0,0          // vbroadcastss  0x2902(%rip),%ymm2        # 6770 <_sk_callback_avx+0x401>
   .byte  197,228,84,210                      // vandps        %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,137,40,0,0          // vbroadcastss  0x2889(%rip),%ymm8        # 6708 <_sk_callback_avx+0x405>
+  .byte  196,98,125,24,5,245,40,0,0          // vbroadcastss  0x28f5(%rip),%ymm8        # 6774 <_sk_callback_avx+0x405>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,127,40,0,0          // vbroadcastss  0x287f(%rip),%ymm8        # 670c <_sk_callback_avx+0x409>
+  .byte  196,98,125,24,5,235,40,0,0          // vbroadcastss  0x28eb(%rip),%ymm8        # 6778 <_sk_callback_avx+0x409>
   .byte  196,193,100,84,216                  // vandps        %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,113,40,0,0          // vbroadcastss  0x2871(%rip),%ymm8        # 6710 <_sk_callback_avx+0x40d>
+  .byte  196,98,125,24,5,221,40,0,0          // vbroadcastss  0x28dd(%rip),%ymm8        # 677c <_sk_callback_avx+0x40d>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -17802,7 +17933,7 @@ FUNCTION(_sk_store_4444_avx)
 _sk_store_4444_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,86,40,0,0           // vbroadcastss  0x2856(%rip),%ymm8        # 6714 <_sk_callback_avx+0x411>
+  .byte  196,98,125,24,5,194,40,0,0          // vbroadcastss  0x28c2(%rip),%ymm8        # 6780 <_sk_callback_avx+0x411>
   .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
@@ -17883,10 +18014,10 @@ _sk_load_8888_avx:
   .byte  72,133,201                          // test          %rcx,%rcx
   .byte  15,133,135,0,0,0                    // jne           4065 <_sk_load_8888_avx+0x95>
   .byte  196,65,124,16,12,186                // vmovups       (%r10,%rdi,4),%ymm9
-  .byte  197,124,40,21,148,40,0,0            // vmovaps       0x2894(%rip),%ymm10        # 6880 <_sk_callback_avx+0x57d>
+  .byte  197,124,40,21,20,41,0,0             // vmovaps       0x2914(%rip),%ymm10        # 6900 <_sk_callback_avx+0x591>
   .byte  196,193,52,84,194                   // vandps        %ymm10,%ymm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,26,39,0,0           // vbroadcastss  0x271a(%rip),%ymm8        # 6718 <_sk_callback_avx+0x415>
+  .byte  196,98,125,24,5,134,39,0,0          // vbroadcastss  0x2786(%rip),%ymm8        # 6784 <_sk_callback_avx+0x415>
   .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
@@ -18001,10 +18132,10 @@ _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  197,124,40,21,190,38,0,0            // vmovaps       0x26be(%rip),%ymm10        # 68a0 <_sk_callback_avx+0x59d>
+  .byte  197,124,40,21,62,39,0,0             // vmovaps       0x273e(%rip),%ymm10        # 6920 <_sk_callback_avx+0x5b1>
   .byte  196,193,124,84,194                  // vandps        %ymm10,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,40,37,0,0          // vbroadcastss  0x2528(%rip),%ymm9        # 671c <_sk_callback_avx+0x419>
+  .byte  196,98,125,24,13,148,37,0,0         // vbroadcastss  0x2594(%rip),%ymm9        # 6788 <_sk_callback_avx+0x419>
   .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
@@ -18036,7 +18167,7 @@ FUNCTION(_sk_store_8888_avx)
 _sk_store_8888_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  76,139,16                           // mov           (%rax),%r10
-  .byte  196,98,125,24,5,182,36,0,0          // vbroadcastss  0x24b6(%rip),%ymm8        # 6720 <_sk_callback_avx+0x41d>
+  .byte  196,98,125,24,5,34,37,0,0           // vbroadcastss  0x2522(%rip),%ymm8        # 678c <_sk_callback_avx+0x41d>
   .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
@@ -18141,13 +18272,13 @@ _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  196,98,125,24,37,29,35,0,0          // vbroadcastss  0x231d(%rip),%ymm12        # 6724 <_sk_callback_avx+0x421>
+  .byte  196,98,125,24,37,137,35,0,0         // vbroadcastss  0x2389(%rip),%ymm12        # 6790 <_sk_callback_avx+0x421>
   .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  196,98,121,24,29,9,35,0,0           // vbroadcastss  0x2309(%rip),%xmm11        # 6728 <_sk_callback_avx+0x425>
+  .byte  196,98,121,24,29,117,35,0,0         // vbroadcastss  0x2375(%rip),%xmm11        # 6794 <_sk_callback_avx+0x425>
   .byte  196,193,8,87,219                    // vxorps        %xmm11,%xmm14,%xmm3
-  .byte  196,98,121,24,45,255,34,0,0         // vbroadcastss  0x22ff(%rip),%xmm13        # 672c <_sk_callback_avx+0x429>
+  .byte  196,98,121,24,45,107,35,0,0         // vbroadcastss  0x236b(%rip),%xmm13        # 6798 <_sk_callback_avx+0x429>
   .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
@@ -18161,7 +18292,7 @@ _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  196,226,121,24,29,181,34,0,0        // vbroadcastss  0x22b5(%rip),%xmm3        # 6730 <_sk_callback_avx+0x42d>
+  .byte  196,226,121,24,29,33,35,0,0         // vbroadcastss  0x2321(%rip),%xmm3        # 679c <_sk_callback_avx+0x42d>
   .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
@@ -18340,13 +18471,13 @@ _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  196,98,125,24,37,121,31,0,0         // vbroadcastss  0x1f79(%rip),%ymm12        # 6734 <_sk_callback_avx+0x431>
+  .byte  196,98,125,24,37,229,31,0,0         // vbroadcastss  0x1fe5(%rip),%ymm12        # 67a0 <_sk_callback_avx+0x431>
   .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  196,98,121,24,29,101,31,0,0         // vbroadcastss  0x1f65(%rip),%xmm11        # 6738 <_sk_callback_avx+0x435>
+  .byte  196,98,121,24,29,209,31,0,0         // vbroadcastss  0x1fd1(%rip),%xmm11        # 67a4 <_sk_callback_avx+0x435>
   .byte  196,193,8,87,219                    // vxorps        %xmm11,%xmm14,%xmm3
-  .byte  196,98,121,24,45,91,31,0,0          // vbroadcastss  0x1f5b(%rip),%xmm13        # 673c <_sk_callback_avx+0x439>
+  .byte  196,98,121,24,45,199,31,0,0         // vbroadcastss  0x1fc7(%rip),%xmm13        # 67a8 <_sk_callback_avx+0x439>
   .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
@@ -18360,7 +18491,7 @@ _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  196,226,121,24,29,17,31,0,0         // vbroadcastss  0x1f11(%rip),%xmm3        # 6740 <_sk_callback_avx+0x43d>
+  .byte  196,226,121,24,29,125,31,0,0        // vbroadcastss  0x1f7d(%rip),%xmm3        # 67ac <_sk_callback_avx+0x43d>
   .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
@@ -18464,12 +18595,12 @@ _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  196,98,125,24,13,42,29,0,0          // vbroadcastss  0x1d2a(%rip),%ymm9        # 6744 <_sk_callback_avx+0x441>
+  .byte  196,98,125,24,13,150,29,0,0         // vbroadcastss  0x1d96(%rip),%ymm9        # 67b0 <_sk_callback_avx+0x441>
   .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  196,98,121,24,5,15,29,0,0           // vbroadcastss  0x1d0f(%rip),%xmm8        # 6748 <_sk_callback_avx+0x445>
+  .byte  196,98,121,24,5,123,29,0,0          // vbroadcastss  0x1d7b(%rip),%xmm8        # 67b4 <_sk_callback_avx+0x445>
   .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
@@ -18479,7 +18610,7 @@ _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  196,98,125,24,21,214,28,0,0         // vbroadcastss  0x1cd6(%rip),%ymm10        # 674c <_sk_callback_avx+0x449>
+  .byte  196,98,125,24,21,66,29,0,0          // vbroadcastss  0x1d42(%rip),%ymm10        # 67b8 <_sk_callback_avx+0x449>
   .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
@@ -18624,7 +18755,7 @@ _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  196,98,125,24,29,46,26,0,0          // vbroadcastss  0x1a2e(%rip),%ymm11        # 6750 <_sk_callback_avx+0x44d>
+  .byte  196,98,125,24,29,154,26,0,0         // vbroadcastss  0x1a9a(%rip),%ymm11        # 67bc <_sk_callback_avx+0x44d>
   .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
@@ -18717,7 +18848,7 @@ _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  196,98,125,24,29,142,24,0,0         // vbroadcastss  0x188e(%rip),%ymm11        # 6754 <_sk_callback_avx+0x451>
+  .byte  196,98,125,24,29,250,24,0,0         // vbroadcastss  0x18fa(%rip),%ymm11        # 67c0 <_sk_callback_avx+0x451>
   .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
@@ -18738,7 +18869,7 @@ _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  196,226,125,24,29,43,24,0,0         // vbroadcastss  0x182b(%rip),%ymm3        # 6758 <_sk_callback_avx+0x455>
+  .byte  196,226,125,24,29,151,24,0,0        // vbroadcastss  0x1897(%rip),%ymm3        # 67c4 <_sk_callback_avx+0x455>
   .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
@@ -18781,7 +18912,7 @@ _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  196,98,125,24,5,104,23,0,0          // vbroadcastss  0x1768(%rip),%ymm8        # 675c <_sk_callback_avx+0x459>
+  .byte  196,98,125,24,5,212,23,0,0          // vbroadcastss  0x17d4(%rip),%ymm8        # 67c8 <_sk_callback_avx+0x459>
   .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
@@ -19043,16 +19174,54 @@ _sk_mirror_y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_clamp_x_1_avx
+.globl _sk_clamp_x_1_avx
+FUNCTION(_sk_clamp_x_1_avx)
+_sk_clamp_x_1_avx:
+  .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
+  .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
+  .byte  196,98,125,24,5,242,19,0,0          // vbroadcastss  0x13f2(%rip),%ymm8        # 67cc <_sk_callback_avx+0x45d>
+  .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_repeat_x_1_avx
+.globl _sk_repeat_x_1_avx
+FUNCTION(_sk_repeat_x_1_avx)
+_sk_repeat_x_1_avx:
+  .byte  196,99,125,8,192,1                  // vroundps      $0x1,%ymm0,%ymm8
+  .byte  196,193,124,92,192                  // vsubps        %ymm8,%ymm0,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_mirror_x_1_avx
+.globl _sk_mirror_x_1_avx
+FUNCTION(_sk_mirror_x_1_avx)
+_sk_mirror_x_1_avx:
+  .byte  196,98,125,24,5,213,19,0,0          // vbroadcastss  0x13d5(%rip),%ymm8        # 67d0 <_sk_callback_avx+0x461>
+  .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
+  .byte  196,98,125,24,13,203,19,0,0         // vbroadcastss  0x13cb(%rip),%ymm9        # 67d4 <_sk_callback_avx+0x465>
+  .byte  196,65,124,89,201                   // vmulps        %ymm9,%ymm0,%ymm9
+  .byte  196,67,125,8,201,1                  // vroundps      $0x1,%ymm9,%ymm9
+  .byte  196,65,52,88,201                    // vaddps        %ymm9,%ymm9,%ymm9
+  .byte  196,193,124,92,193                  // vsubps        %ymm9,%ymm0,%ymm0
+  .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
+  .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
+  .byte  197,60,92,192                       // vsubps        %ymm0,%ymm8,%ymm8
+  .byte  197,188,84,192                      // vandps        %ymm0,%ymm8,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_luminance_to_alpha_avx
 .globl _sk_luminance_to_alpha_avx
 FUNCTION(_sk_luminance_to_alpha_avx)
 _sk_luminance_to_alpha_avx:
-  .byte  196,226,125,24,29,143,19,0,0        // vbroadcastss  0x138f(%rip),%ymm3        # 6760 <_sk_callback_avx+0x45d>
+  .byte  196,226,125,24,29,155,19,0,0        // vbroadcastss  0x139b(%rip),%ymm3        # 67d8 <_sk_callback_avx+0x469>
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,24,29,134,19,0,0        // vbroadcastss  0x1386(%rip),%ymm3        # 6764 <_sk_callback_avx+0x461>
+  .byte  196,226,125,24,29,146,19,0,0        // vbroadcastss  0x1392(%rip),%ymm3        # 67dc <_sk_callback_avx+0x46d>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,121,19,0,0        // vbroadcastss  0x1379(%rip),%ymm1        # 6768 <_sk_callback_avx+0x465>
+  .byte  196,226,125,24,13,133,19,0,0        // vbroadcastss  0x1385(%rip),%ymm1        # 67e0 <_sk_callback_avx+0x471>
   .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
@@ -19271,9 +19440,9 @@ _sk_evenly_spaced_gradient_avx:
   .byte  72,139,24                           // mov           (%rax),%rbx
   .byte  72,139,104,8                        // mov           0x8(%rax),%rbp
   .byte  72,255,203                          // dec           %rbx
-  .byte  120,7                               // js            5745 <_sk_evenly_spaced_gradient_avx+0x1f>
+  .byte  120,7                               // js            57b1 <_sk_evenly_spaced_gradient_avx+0x1f>
   .byte  196,225,242,42,203                  // vcvtsi2ss     %rbx,%xmm1,%xmm1
-  .byte  235,21                              // jmp           575a <_sk_evenly_spaced_gradient_avx+0x34>
+  .byte  235,21                              // jmp           57c6 <_sk_evenly_spaced_gradient_avx+0x34>
   .byte  73,137,216                          // mov           %rbx,%r8
   .byte  73,209,232                          // shr           %r8
   .byte  131,227,1                           // and           $0x1,%ebx
@@ -19440,12 +19609,12 @@ _sk_gradient_avx:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  73,131,248,2                        // cmp           $0x2,%r8
-  .byte  114,80                              // jb            5ae8 <_sk_gradient_avx+0x69>
+  .byte  114,80                              // jb            5b54 <_sk_gradient_avx+0x69>
   .byte  72,139,88,72                        // mov           0x48(%rax),%rbx
   .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  196,98,125,24,21,187,12,0,0         // vbroadcastss  0xcbb(%rip),%ymm10        # 676c <_sk_callback_avx+0x469>
+  .byte  196,98,125,24,21,199,12,0,0         // vbroadcastss  0xcc7(%rip),%ymm10        # 67e4 <_sk_callback_avx+0x475>
   .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
@@ -19457,7 +19626,7 @@ _sk_gradient_avx:
   .byte  196,227,117,24,202,1                // vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
   .byte  72,131,195,4                        // add           $0x4,%rbx
   .byte  73,255,200                          // dec           %r8
-  .byte  117,205                             // jne           5ab5 <_sk_gradient_avx+0x36>
+  .byte  117,205                             // jne           5b21 <_sk_gradient_avx+0x36>
   .byte  196,195,249,22,200,1                // vpextrq       $0x1,%xmm1,%r8
   .byte  69,137,193                          // mov           %r8d,%r9d
   .byte  73,193,232,32                       // shr           $0x20,%r8
@@ -19639,27 +19808,27 @@ _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  196,98,125,24,45,223,8,0,0          // vbroadcastss  0x8df(%rip),%ymm13        # 6770 <_sk_callback_avx+0x46d>
+  .byte  196,98,125,24,45,235,8,0,0          // vbroadcastss  0x8eb(%rip),%ymm13        # 67e8 <_sk_callback_avx+0x479>
   .byte  196,65,28,89,237                    // vmulps        %ymm13,%ymm12,%ymm13
-  .byte  196,98,125,24,53,213,8,0,0          // vbroadcastss  0x8d5(%rip),%ymm14        # 6774 <_sk_callback_avx+0x471>
+  .byte  196,98,125,24,53,225,8,0,0          // vbroadcastss  0x8e1(%rip),%ymm14        # 67ec <_sk_callback_avx+0x47d>
   .byte  196,65,20,88,238                    // vaddps        %ymm14,%ymm13,%ymm13
   .byte  196,65,28,89,237                    // vmulps        %ymm13,%ymm12,%ymm13
-  .byte  196,98,125,24,53,198,8,0,0          // vbroadcastss  0x8c6(%rip),%ymm14        # 6778 <_sk_callback_avx+0x475>
+  .byte  196,98,125,24,53,210,8,0,0          // vbroadcastss  0x8d2(%rip),%ymm14        # 67f0 <_sk_callback_avx+0x481>
   .byte  196,65,20,88,238                    // vaddps        %ymm14,%ymm13,%ymm13
   .byte  196,65,28,89,229                    // vmulps        %ymm13,%ymm12,%ymm12
-  .byte  196,98,125,24,45,183,8,0,0          // vbroadcastss  0x8b7(%rip),%ymm13        # 677c <_sk_callback_avx+0x479>
+  .byte  196,98,125,24,45,195,8,0,0          // vbroadcastss  0x8c3(%rip),%ymm13        # 67f4 <_sk_callback_avx+0x485>
   .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  196,98,125,24,21,162,8,0,0          // vbroadcastss  0x8a2(%rip),%ymm10        # 6780 <_sk_callback_avx+0x47d>
+  .byte  196,98,125,24,21,174,8,0,0          // vbroadcastss  0x8ae(%rip),%ymm10        # 67f8 <_sk_callback_avx+0x489>
   .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  196,98,125,24,21,140,8,0,0          // vbroadcastss  0x88c(%rip),%ymm10        # 6784 <_sk_callback_avx+0x481>
+  .byte  196,98,125,24,21,152,8,0,0          // vbroadcastss  0x898(%rip),%ymm10        # 67fc <_sk_callback_avx+0x48d>
   .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  196,98,125,24,21,118,8,0,0          // vbroadcastss  0x876(%rip),%ymm10        # 6788 <_sk_callback_avx+0x485>
+  .byte  196,98,125,24,21,130,8,0,0          // vbroadcastss  0x882(%rip),%ymm10        # 6800 <_sk_callback_avx+0x491>
   .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
@@ -19683,7 +19852,7 @@ HIDDEN _sk_save_xy_avx
 FUNCTION(_sk_save_xy_avx)
 _sk_save_xy_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,64,8,0,0            // vbroadcastss  0x840(%rip),%ymm8        # 678c <_sk_callback_avx+0x489>
+  .byte  196,98,125,24,5,76,8,0,0            // vbroadcastss  0x84c(%rip),%ymm8        # 6804 <_sk_callback_avx+0x495>
   .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
@@ -19720,9 +19889,9 @@ HIDDEN _sk_bilinear_nx_avx
 FUNCTION(_sk_bilinear_nx_avx)
 _sk_bilinear_nx_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,204,7,0,0          // vbroadcastss  0x7cc(%rip),%ymm0        # 6790 <_sk_callback_avx+0x48d>
+  .byte  196,226,125,24,5,216,7,0,0          // vbroadcastss  0x7d8(%rip),%ymm0        # 6808 <_sk_callback_avx+0x499>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,195,7,0,0           // vbroadcastss  0x7c3(%rip),%ymm8        # 6794 <_sk_callback_avx+0x491>
+  .byte  196,98,125,24,5,207,7,0,0           // vbroadcastss  0x7cf(%rip),%ymm8        # 680c <_sk_callback_avx+0x49d>
   .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
@@ -19733,7 +19902,7 @@ HIDDEN _sk_bilinear_px_avx
 FUNCTION(_sk_bilinear_px_avx)
 _sk_bilinear_px_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,171,7,0,0          // vbroadcastss  0x7ab(%rip),%ymm0        # 6798 <_sk_callback_avx+0x495>
+  .byte  196,226,125,24,5,183,7,0,0          // vbroadcastss  0x7b7(%rip),%ymm0        # 6810 <_sk_callback_avx+0x4a1>
   .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)
@@ -19745,9 +19914,9 @@ HIDDEN _sk_bilinear_ny_avx
 FUNCTION(_sk_bilinear_ny_avx)
 _sk_bilinear_ny_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,143,7,0,0         // vbroadcastss  0x78f(%rip),%ymm1        # 679c <_sk_callback_avx+0x499>
+  .byte  196,226,125,24,13,155,7,0,0         // vbroadcastss  0x79b(%rip),%ymm1        # 6814 <_sk_callback_avx+0x4a5>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,133,7,0,0           // vbroadcastss  0x785(%rip),%ymm8        # 67a0 <_sk_callback_avx+0x49d>
+  .byte  196,98,125,24,5,145,7,0,0           // vbroadcastss  0x791(%rip),%ymm8        # 6818 <_sk_callback_avx+0x4a9>
   .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
@@ -19758,7 +19927,7 @@ HIDDEN _sk_bilinear_py_avx
 FUNCTION(_sk_bilinear_py_avx)
 _sk_bilinear_py_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,109,7,0,0         // vbroadcastss  0x76d(%rip),%ymm1        # 67a4 <_sk_callback_avx+0x4a1>
+  .byte  196,226,125,24,13,121,7,0,0         // vbroadcastss  0x779(%rip),%ymm1        # 681c <_sk_callback_avx+0x4ad>
   .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)
@@ -19770,14 +19939,14 @@ HIDDEN _sk_bicubic_n3x_avx
 FUNCTION(_sk_bicubic_n3x_avx)
 _sk_bicubic_n3x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,80,7,0,0           // vbroadcastss  0x750(%rip),%ymm0        # 67a8 <_sk_callback_avx+0x4a5>
+  .byte  196,226,125,24,5,92,7,0,0           // vbroadcastss  0x75c(%rip),%ymm0        # 6820 <_sk_callback_avx+0x4b1>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,71,7,0,0            // vbroadcastss  0x747(%rip),%ymm8        # 67ac <_sk_callback_avx+0x4a9>
+  .byte  196,98,125,24,5,83,7,0,0            // vbroadcastss  0x753(%rip),%ymm8        # 6824 <_sk_callback_avx+0x4b5>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,56,7,0,0           // vbroadcastss  0x738(%rip),%ymm10        # 67b0 <_sk_callback_avx+0x4ad>
+  .byte  196,98,125,24,21,68,7,0,0           // vbroadcastss  0x744(%rip),%ymm10        # 6828 <_sk_callback_avx+0x4b9>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,46,7,0,0           // vbroadcastss  0x72e(%rip),%ymm10        # 67b4 <_sk_callback_avx+0x4b1>
+  .byte  196,98,125,24,21,58,7,0,0           // vbroadcastss  0x73a(%rip),%ymm10        # 682c <_sk_callback_avx+0x4bd>
   .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)
@@ -19789,19 +19958,19 @@ HIDDEN _sk_bicubic_n1x_avx
 FUNCTION(_sk_bicubic_n1x_avx)
 _sk_bicubic_n1x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,17,7,0,0           // vbroadcastss  0x711(%rip),%ymm0        # 67b8 <_sk_callback_avx+0x4b5>
+  .byte  196,226,125,24,5,29,7,0,0           // vbroadcastss  0x71d(%rip),%ymm0        # 6830 <_sk_callback_avx+0x4c1>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,8,7,0,0             // vbroadcastss  0x708(%rip),%ymm8        # 67bc <_sk_callback_avx+0x4b9>
+  .byte  196,98,125,24,5,20,7,0,0            // vbroadcastss  0x714(%rip),%ymm8        # 6834 <_sk_callback_avx+0x4c5>
   .byte  197,60,92,64,64                     // vsubps        0x40(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,254,6,0,0          // vbroadcastss  0x6fe(%rip),%ymm9        # 67c0 <_sk_callback_avx+0x4bd>
+  .byte  196,98,125,24,13,10,7,0,0           // vbroadcastss  0x70a(%rip),%ymm9        # 6838 <_sk_callback_avx+0x4c9>
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,244,6,0,0          // vbroadcastss  0x6f4(%rip),%ymm10        # 67c4 <_sk_callback_avx+0x4c1>
+  .byte  196,98,125,24,21,0,7,0,0            // vbroadcastss  0x700(%rip),%ymm10        # 683c <_sk_callback_avx+0x4cd>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,229,6,0,0          // vbroadcastss  0x6e5(%rip),%ymm10        # 67c8 <_sk_callback_avx+0x4c5>
+  .byte  196,98,125,24,21,241,6,0,0          // vbroadcastss  0x6f1(%rip),%ymm10        # 6840 <_sk_callback_avx+0x4d1>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,214,6,0,0          // vbroadcastss  0x6d6(%rip),%ymm9        # 67cc <_sk_callback_avx+0x4c9>
+  .byte  196,98,125,24,13,226,6,0,0          // vbroadcastss  0x6e2(%rip),%ymm9        # 6844 <_sk_callback_avx+0x4d5>
   .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
@@ -19812,17 +19981,17 @@ HIDDEN _sk_bicubic_p1x_avx
 FUNCTION(_sk_bicubic_p1x_avx)
 _sk_bicubic_p1x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,190,6,0,0           // vbroadcastss  0x6be(%rip),%ymm8        # 67d0 <_sk_callback_avx+0x4cd>
+  .byte  196,98,125,24,5,202,6,0,0           // vbroadcastss  0x6ca(%rip),%ymm8        # 6848 <_sk_callback_avx+0x4d9>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,72,64                    // vmovups       0x40(%rax),%ymm9
-  .byte  196,98,125,24,21,176,6,0,0          // vbroadcastss  0x6b0(%rip),%ymm10        # 67d4 <_sk_callback_avx+0x4d1>
+  .byte  196,98,125,24,21,188,6,0,0          // vbroadcastss  0x6bc(%rip),%ymm10        # 684c <_sk_callback_avx+0x4dd>
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
-  .byte  196,98,125,24,29,166,6,0,0          // vbroadcastss  0x6a6(%rip),%ymm11        # 67d8 <_sk_callback_avx+0x4d5>
+  .byte  196,98,125,24,29,178,6,0,0          // vbroadcastss  0x6b2(%rip),%ymm11        # 6850 <_sk_callback_avx+0x4e1>
   .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  196,98,125,24,13,141,6,0,0          // vbroadcastss  0x68d(%rip),%ymm9        # 67dc <_sk_callback_avx+0x4d9>
+  .byte  196,98,125,24,13,153,6,0,0          // vbroadcastss  0x699(%rip),%ymm9        # 6854 <_sk_callback_avx+0x4e5>
   .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
@@ -19833,13 +20002,13 @@ HIDDEN _sk_bicubic_p3x_avx
 FUNCTION(_sk_bicubic_p3x_avx)
 _sk_bicubic_p3x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,117,6,0,0          // vbroadcastss  0x675(%rip),%ymm0        # 67e0 <_sk_callback_avx+0x4dd>
+  .byte  196,226,125,24,5,129,6,0,0          // vbroadcastss  0x681(%rip),%ymm0        # 6858 <_sk_callback_avx+0x4e9>
   .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  196,98,125,24,21,98,6,0,0           // vbroadcastss  0x662(%rip),%ymm10        # 67e4 <_sk_callback_avx+0x4e1>
+  .byte  196,98,125,24,21,110,6,0,0          // vbroadcastss  0x66e(%rip),%ymm10        # 685c <_sk_callback_avx+0x4ed>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,88,6,0,0           // vbroadcastss  0x658(%rip),%ymm10        # 67e8 <_sk_callback_avx+0x4e5>
+  .byte  196,98,125,24,21,100,6,0,0          // vbroadcastss  0x664(%rip),%ymm10        # 6860 <_sk_callback_avx+0x4f1>
   .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)
@@ -19851,14 +20020,14 @@ HIDDEN _sk_bicubic_n3y_avx
 FUNCTION(_sk_bicubic_n3y_avx)
 _sk_bicubic_n3y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,59,6,0,0          // vbroadcastss  0x63b(%rip),%ymm1        # 67ec <_sk_callback_avx+0x4e9>
+  .byte  196,226,125,24,13,71,6,0,0          // vbroadcastss  0x647(%rip),%ymm1        # 6864 <_sk_callback_avx+0x4f5>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,49,6,0,0            // vbroadcastss  0x631(%rip),%ymm8        # 67f0 <_sk_callback_avx+0x4ed>
+  .byte  196,98,125,24,5,61,6,0,0            // vbroadcastss  0x63d(%rip),%ymm8        # 6868 <_sk_callback_avx+0x4f9>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,34,6,0,0           // vbroadcastss  0x622(%rip),%ymm10        # 67f4 <_sk_callback_avx+0x4f1>
+  .byte  196,98,125,24,21,46,6,0,0           // vbroadcastss  0x62e(%rip),%ymm10        # 686c <_sk_callback_avx+0x4fd>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,24,6,0,0           // vbroadcastss  0x618(%rip),%ymm10        # 67f8 <_sk_callback_avx+0x4f5>
+  .byte  196,98,125,24,21,36,6,0,0           // vbroadcastss  0x624(%rip),%ymm10        # 6870 <_sk_callback_avx+0x501>
   .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)
@@ -19870,19 +20039,19 @@ HIDDEN _sk_bicubic_n1y_avx
 FUNCTION(_sk_bicubic_n1y_avx)
 _sk_bicubic_n1y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,251,5,0,0         // vbroadcastss  0x5fb(%rip),%ymm1        # 67fc <_sk_callback_avx+0x4f9>
+  .byte  196,226,125,24,13,7,6,0,0           // vbroadcastss  0x607(%rip),%ymm1        # 6874 <_sk_callback_avx+0x505>
   .byte  197,244,88,72,32                    // vaddps        0x20(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,241,5,0,0           // vbroadcastss  0x5f1(%rip),%ymm8        # 6800 <_sk_callback_avx+0x4fd>
+  .byte  196,98,125,24,5,253,5,0,0           // vbroadcastss  0x5fd(%rip),%ymm8        # 6878 <_sk_callback_avx+0x509>
   .byte  197,60,92,64,96                     // vsubps        0x60(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,231,5,0,0          // vbroadcastss  0x5e7(%rip),%ymm9        # 6804 <_sk_callback_avx+0x501>
+  .byte  196,98,125,24,13,243,5,0,0          // vbroadcastss  0x5f3(%rip),%ymm9        # 687c <_sk_callback_avx+0x50d>
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,221,5,0,0          // vbroadcastss  0x5dd(%rip),%ymm10        # 6808 <_sk_callback_avx+0x505>
+  .byte  196,98,125,24,21,233,5,0,0          // vbroadcastss  0x5e9(%rip),%ymm10        # 6880 <_sk_callback_avx+0x511>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,206,5,0,0          // vbroadcastss  0x5ce(%rip),%ymm10        # 680c <_sk_callback_avx+0x509>
+  .byte  196,98,125,24,21,218,5,0,0          // vbroadcastss  0x5da(%rip),%ymm10        # 6884 <_sk_callback_avx+0x515>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,191,5,0,0          // vbroadcastss  0x5bf(%rip),%ymm9        # 6810 <_sk_callback_avx+0x50d>
+  .byte  196,98,125,24,13,203,5,0,0          // vbroadcastss  0x5cb(%rip),%ymm9        # 6888 <_sk_callback_avx+0x519>
   .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
@@ -19893,17 +20062,17 @@ HIDDEN _sk_bicubic_p1y_avx
 FUNCTION(_sk_bicubic_p1y_avx)
 _sk_bicubic_p1y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,167,5,0,0           // vbroadcastss  0x5a7(%rip),%ymm8        # 6814 <_sk_callback_avx+0x511>
+  .byte  196,98,125,24,5,179,5,0,0           // vbroadcastss  0x5b3(%rip),%ymm8        # 688c <_sk_callback_avx+0x51d>
   .byte  197,188,88,72,32                    // vaddps        0x20(%rax),%ymm8,%ymm1
   .byte  197,124,16,72,96                    // vmovups       0x60(%rax),%ymm9
-  .byte  196,98,125,24,21,152,5,0,0          // vbroadcastss  0x598(%rip),%ymm10        # 6818 <_sk_callback_avx+0x515>
+  .byte  196,98,125,24,21,164,5,0,0          // vbroadcastss  0x5a4(%rip),%ymm10        # 6890 <_sk_callback_avx+0x521>
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
-  .byte  196,98,125,24,29,142,5,0,0          // vbroadcastss  0x58e(%rip),%ymm11        # 681c <_sk_callback_avx+0x519>
+  .byte  196,98,125,24,29,154,5,0,0          // vbroadcastss  0x59a(%rip),%ymm11        # 6894 <_sk_callback_avx+0x525>
   .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  196,98,125,24,13,117,5,0,0          // vbroadcastss  0x575(%rip),%ymm9        # 6820 <_sk_callback_avx+0x51d>
+  .byte  196,98,125,24,13,129,5,0,0          // vbroadcastss  0x581(%rip),%ymm9        # 6898 <_sk_callback_avx+0x529>
   .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
@@ -19914,13 +20083,13 @@ HIDDEN _sk_bicubic_p3y_avx
 FUNCTION(_sk_bicubic_p3y_avx)
 _sk_bicubic_p3y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,93,5,0,0          // vbroadcastss  0x55d(%rip),%ymm1        # 6824 <_sk_callback_avx+0x521>
+  .byte  196,226,125,24,13,105,5,0,0         // vbroadcastss  0x569(%rip),%ymm1        # 689c <_sk_callback_avx+0x52d>
   .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  196,98,125,24,21,73,5,0,0           // vbroadcastss  0x549(%rip),%ymm10        # 6828 <_sk_callback_avx+0x525>
+  .byte  196,98,125,24,21,85,5,0,0           // vbroadcastss  0x555(%rip),%ymm10        # 68a0 <_sk_callback_avx+0x531>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,63,5,0,0           // vbroadcastss  0x53f(%rip),%ymm10        # 682c <_sk_callback_avx+0x529>
+  .byte  196,98,125,24,21,75,5,0,0           // vbroadcastss  0x54b(%rip),%ymm10        # 68a4 <_sk_callback_avx+0x535>
   .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)
@@ -20044,25 +20213,25 @@ 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 64d9 <.literal4+0xb1>
+  .byte  71,225,61                           // rex.RXB       loope 6545 <.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 64e9 <.literal4+0xc1>
+  .byte  71,225,61                           // rex.RXB       loope 6555 <.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 64f9 <.literal4+0xd1>
+  .byte  71,225,61                           // rex.RXB       loope 6565 <.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 6509 <.literal4+0xe1>
+  .byte  71,225,61                           // rex.RXB       loope 6575 <.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)
@@ -20110,7 +20279,7 @@ 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        6555 <.literal4+0x12d>
+  .byte  224,7                               // loopne        65c1 <.literal4+0x12d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -20126,10 +20295,10 @@ 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            657c <.literal4+0x154>
+  .byte  127,0                               // jg            65e8 <.literal4+0x154>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            65f5 <.literal4+0x1cd>
+  .byte  119,115                             // ja            6661 <.literal4+0x1cd>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20143,10 +20312,10 @@ 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            65b0 <.literal4+0x188>
+  .byte  127,0                               // jg            661c <.literal4+0x188>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            6629 <.literal4+0x201>
+  .byte  119,115                             // ja            6695 <.literal4+0x201>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20160,10 +20329,10 @@ 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            65e4 <.literal4+0x1bc>
+  .byte  127,0                               // jg            6650 <.literal4+0x1bc>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            665d <.literal4+0x235>
+  .byte  119,115                             // ja            66c9 <.literal4+0x235>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20177,10 +20346,10 @@ 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            6618 <.literal4+0x1f0>
+  .byte  127,0                               // jg            6684 <.literal4+0x1f0>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            6691 <.literal4+0x269>
+  .byte  119,115                             // ja            66fd <.literal4+0x269>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -20193,7 +20362,7 @@ 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            668f <.literal4+0x267>
+  .byte  127,67                              // jg            66fb <.literal4+0x267>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,195                               // add           %al,%bl
   .byte  0,0                                 // add           %al,(%rax)
@@ -20205,10 +20374,10 @@ BALIGN4
   .byte  190,80,128,3,62                     // mov           $0x3e038050,%esi
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           66af <.literal4+0x287>
+  .byte  118,63                              // jbe           671b <.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            66c3 <.literal4+0x29b>
+  .byte  127,67                              // jg            672f <.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)
@@ -20217,7 +20386,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        66a5 <.literal4+0x27d>
+  .byte  224,7                               // loopne        6711 <.literal4+0x27d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -20229,7 +20398,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        66c1 <.literal4+0x299>
+  .byte  224,7                               // loopne        672d <.literal4+0x299>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -20240,7 +20409,7 @@ BALIGN4
   .byte  0,0                                 // add           %al,(%rax)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            6716 <.literal4+0x2ee>
+  .byte  124,66                              // jl            6782 <.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)
@@ -20258,9 +20427,9 @@ 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            6759 <.literal4+0x331>
+  .byte  112,65                              // jo            67c5 <.literal4+0x331>
   .byte  129,128,128,59,129,128,128,59,0,0   // addl          $0x3b80,-0x7f7ec480(%rax)
-  .byte  127,67                              // jg            6767 <.literal4+0x33f>
+  .byte  127,67                              // jg            67d3 <.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)
@@ -20276,7 +20445,11 @@ 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            67a7 <.literal4+0x37f>
+  .byte  127,71                              // jg            6813 <.literal4+0x37f>
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,191,0,0,0                     // add           %al,0xbf(%rax)
+  .byte  63                                  // (bad)
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,89                               // ds            pop %rcx
@@ -20506,7 +20679,7 @@ _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  15,40,21,84,71,0,0                  // movaps        0x4754(%rip),%xmm2        # 47d0 <_sk_callback_sse41+0xe0>
+  .byte  15,40,21,180,71,0,0                 // movaps        0x47b4(%rip),%xmm2        # 4830 <_sk_callback_sse41+0xde>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  15,16,2                             // movups        (%rdx),%xmm0
   .byte  15,88,193                           // addps         %xmm1,%xmm0
@@ -20515,7 +20688,7 @@ _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  15,40,21,67,71,0,0                  // movaps        0x4743(%rip),%xmm2        # 47e0 <_sk_callback_sse41+0xf0>
+  .byte  15,40,21,163,71,0,0                 // movaps        0x47a3(%rip),%xmm2        # 4840 <_sk_callback_sse41+0xee>
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
@@ -20536,14 +20709,14 @@ _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  102,68,15,111,21,14,71,0,0          // movdqa        0x470e(%rip),%xmm10        # 47f0 <_sk_callback_sse41+0x100>
+  .byte  102,68,15,111,21,110,71,0,0         // movdqa        0x476e(%rip),%xmm10        # 4850 <_sk_callback_sse41+0xfe>
   .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  102,68,15,111,37,250,70,0,0         // movdqa        0x46fa(%rip),%xmm12        # 4800 <_sk_callback_sse41+0x110>
-  .byte  102,68,15,111,45,1,71,0,0           // movdqa        0x4701(%rip),%xmm13        # 4810 <_sk_callback_sse41+0x120>
+  .byte  102,68,15,111,37,90,71,0,0          // movdqa        0x475a(%rip),%xmm12        # 4860 <_sk_callback_sse41+0x10e>
+  .byte  102,68,15,111,45,97,71,0,0          // movdqa        0x4761(%rip),%xmm13        # 4870 <_sk_callback_sse41+0x11e>
   .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
@@ -20559,8 +20732,8 @@ _sk_dither_sse41:
   .byte  102,69,15,235,198                   // por           %xmm14,%xmm8
   .byte  102,69,15,235,197                   // por           %xmm13,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,188,70,0,0               // mulps         0x46bc(%rip),%xmm8        # 4820 <_sk_callback_sse41+0x130>
-  .byte  68,15,88,5,196,70,0,0               // addps         0x46c4(%rip),%xmm8        # 4830 <_sk_callback_sse41+0x140>
+  .byte  68,15,89,5,28,71,0,0                // mulps         0x471c(%rip),%xmm8        # 4880 <_sk_callback_sse41+0x12e>
+  .byte  68,15,88,5,36,71,0,0                // addps         0x4724(%rip),%xmm8        # 4890 <_sk_callback_sse41+0x13e>
   .byte  243,68,15,16,80,8                   // movss         0x8(%rax),%xmm10
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
@@ -20637,7 +20810,7 @@ HIDDEN _sk_srcatop_sse41
 FUNCTION(_sk_srcatop_sse41)
 _sk_srcatop_sse41:
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  68,15,40,5,29,70,0,0                // movaps        0x461d(%rip),%xmm8        # 4840 <_sk_callback_sse41+0x150>
+  .byte  68,15,40,5,125,70,0,0               // movaps        0x467d(%rip),%xmm8        # 48a0 <_sk_callback_sse41+0x14e>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -20662,7 +20835,7 @@ FUNCTION(_sk_dstatop_sse41)
 _sk_dstatop_sse41:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
   .byte  68,15,89,196                        // mulps         %xmm4,%xmm8
-  .byte  68,15,40,13,224,69,0,0              // movaps        0x45e0(%rip),%xmm9        # 4850 <_sk_callback_sse41+0x160>
+  .byte  68,15,40,13,64,70,0,0               // movaps        0x4640(%rip),%xmm9        # 48b0 <_sk_callback_sse41+0x15e>
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
@@ -20709,7 +20882,7 @@ HIDDEN _sk_srcout_sse41
 .globl _sk_srcout_sse41
 FUNCTION(_sk_srcout_sse41)
 _sk_srcout_sse41:
-  .byte  68,15,40,5,132,69,0,0               // movaps        0x4584(%rip),%xmm8        # 4860 <_sk_callback_sse41+0x170>
+  .byte  68,15,40,5,228,69,0,0               // movaps        0x45e4(%rip),%xmm8        # 48c0 <_sk_callback_sse41+0x16e>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
@@ -20722,7 +20895,7 @@ HIDDEN _sk_dstout_sse41
 .globl _sk_dstout_sse41
 FUNCTION(_sk_dstout_sse41)
 _sk_dstout_sse41:
-  .byte  68,15,40,5,116,69,0,0               // movaps        0x4574(%rip),%xmm8        # 4870 <_sk_callback_sse41+0x180>
+  .byte  68,15,40,5,212,69,0,0               // movaps        0x45d4(%rip),%xmm8        # 48d0 <_sk_callback_sse41+0x17e>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -20739,7 +20912,7 @@ HIDDEN _sk_srcover_sse41
 .globl _sk_srcover_sse41
 FUNCTION(_sk_srcover_sse41)
 _sk_srcover_sse41:
-  .byte  68,15,40,5,87,69,0,0                // movaps        0x4557(%rip),%xmm8        # 4880 <_sk_callback_sse41+0x190>
+  .byte  68,15,40,5,183,69,0,0               // movaps        0x45b7(%rip),%xmm8        # 48e0 <_sk_callback_sse41+0x18e>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -20759,7 +20932,7 @@ HIDDEN _sk_dstover_sse41
 .globl _sk_dstover_sse41
 FUNCTION(_sk_dstover_sse41)
 _sk_dstover_sse41:
-  .byte  68,15,40,5,43,69,0,0                // movaps        0x452b(%rip),%xmm8        # 4890 <_sk_callback_sse41+0x1a0>
+  .byte  68,15,40,5,139,69,0,0               // movaps        0x458b(%rip),%xmm8        # 48f0 <_sk_callback_sse41+0x19e>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -20787,7 +20960,7 @@ HIDDEN _sk_multiply_sse41
 .globl _sk_multiply_sse41
 FUNCTION(_sk_multiply_sse41)
 _sk_multiply_sse41:
-  .byte  68,15,40,5,255,68,0,0               // movaps        0x44ff(%rip),%xmm8        # 48a0 <_sk_callback_sse41+0x1b0>
+  .byte  68,15,40,5,95,69,0,0                // movaps        0x455f(%rip),%xmm8        # 4900 <_sk_callback_sse41+0x1ae>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
@@ -20863,7 +21036,7 @@ HIDDEN _sk_xor__sse41
 FUNCTION(_sk_xor__sse41)
 _sk_xor__sse41:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
-  .byte  15,40,29,48,68,0,0                  // movaps        0x4430(%rip),%xmm3        # 48b0 <_sk_callback_sse41+0x1c0>
+  .byte  15,40,29,144,68,0,0                 // movaps        0x4490(%rip),%xmm3        # 4910 <_sk_callback_sse41+0x1be>
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
@@ -20911,7 +21084,7 @@ _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  15,40,21,155,67,0,0                 // movaps        0x439b(%rip),%xmm2        # 48c0 <_sk_callback_sse41+0x1d0>
+  .byte  15,40,21,251,67,0,0                 // movaps        0x43fb(%rip),%xmm2        # 4920 <_sk_callback_sse41+0x1ce>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -20945,7 +21118,7 @@ _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  15,40,21,64,67,0,0                  // movaps        0x4340(%rip),%xmm2        # 48d0 <_sk_callback_sse41+0x1e0>
+  .byte  15,40,21,160,67,0,0                 // movaps        0x43a0(%rip),%xmm2        # 4930 <_sk_callback_sse41+0x1de>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -20982,7 +21155,7 @@ _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  15,40,21,218,66,0,0                 // movaps        0x42da(%rip),%xmm2        # 48e0 <_sk_callback_sse41+0x1f0>
+  .byte  15,40,21,58,67,0,0                  // movaps        0x433a(%rip),%xmm2        # 4940 <_sk_callback_sse41+0x1ee>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -21009,7 +21182,7 @@ _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  15,40,13,155,66,0,0                 // movaps        0x429b(%rip),%xmm1        # 48f0 <_sk_callback_sse41+0x200>
+  .byte  15,40,13,251,66,0,0                 // movaps        0x42fb(%rip),%xmm1        # 4950 <_sk_callback_sse41+0x1fe>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,207                           // mulps         %xmm7,%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -21023,7 +21196,7 @@ HIDDEN _sk_colorburn_sse41
 FUNCTION(_sk_colorburn_sse41)
 _sk_colorburn_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,138,66,0,0              // movaps        0x428a(%rip),%xmm10        # 4900 <_sk_callback_sse41+0x210>
+  .byte  68,15,40,21,234,66,0,0              // movaps        0x42ea(%rip),%xmm10        # 4960 <_sk_callback_sse41+0x20e>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,203                        // movaps        %xmm11,%xmm9
@@ -21105,7 +21278,7 @@ HIDDEN _sk_colordodge_sse41
 FUNCTION(_sk_colordodge_sse41)
 _sk_colordodge_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,104,65,0,0              // movaps        0x4168(%rip),%xmm10        # 4910 <_sk_callback_sse41+0x220>
+  .byte  68,15,40,21,200,65,0,0              // movaps        0x41c8(%rip),%xmm10        # 4970 <_sk_callback_sse41+0x21e>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -21187,7 +21360,7 @@ _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  68,15,40,21,65,64,0,0               // movaps        0x4041(%rip),%xmm10        # 4920 <_sk_callback_sse41+0x230>
+  .byte  68,15,40,21,161,64,0,0              // movaps        0x40a1(%rip),%xmm10        # 4980 <_sk_callback_sse41+0x22e>
   .byte  65,15,40,234                        // movaps        %xmm10,%xmm5
   .byte  15,92,239                           // subps         %xmm7,%xmm5
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
@@ -21270,7 +21443,7 @@ FUNCTION(_sk_overlay_sse41)
 _sk_overlay_sse41:
   .byte  68,15,40,201                        // movaps        %xmm1,%xmm9
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
-  .byte  68,15,40,21,38,63,0,0               // movaps        0x3f26(%rip),%xmm10        # 4930 <_sk_callback_sse41+0x240>
+  .byte  68,15,40,21,134,63,0,0              // movaps        0x3f86(%rip),%xmm10        # 4990 <_sk_callback_sse41+0x23e>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
@@ -21355,7 +21528,7 @@ _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  15,40,13,253,61,0,0                 // movaps        0x3dfd(%rip),%xmm1        # 4940 <_sk_callback_sse41+0x250>
+  .byte  15,40,13,93,62,0,0                  // movaps        0x3e5d(%rip),%xmm1        # 49a0 <_sk_callback_sse41+0x24e>
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  68,15,92,208                        // subps         %xmm0,%xmm10
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
@@ -21368,10 +21541,10 @@ _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  68,15,40,45,219,61,0,0              // movaps        0x3ddb(%rip),%xmm13        # 4950 <_sk_callback_sse41+0x260>
+  .byte  68,15,40,45,59,62,0,0               // movaps        0x3e3b(%rip),%xmm13        # 49b0 <_sk_callback_sse41+0x25e>
   .byte  69,15,88,245                        // addps         %xmm13,%xmm14
   .byte  68,15,89,242                        // mulps         %xmm2,%xmm14
-  .byte  68,15,40,37,219,61,0,0              // movaps        0x3ddb(%rip),%xmm12        # 4960 <_sk_callback_sse41+0x270>
+  .byte  68,15,40,37,59,62,0,0               // movaps        0x3e3b(%rip),%xmm12        # 49c0 <_sk_callback_sse41+0x26e>
   .byte  69,15,89,252                        // mulps         %xmm12,%xmm15
   .byte  69,15,88,254                        // addps         %xmm14,%xmm15
   .byte  15,40,198                           // movaps        %xmm6,%xmm0
@@ -21557,12 +21730,12 @@ _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  15,40,5,70,59,0,0                   // movaps        0x3b46(%rip),%xmm0        # 4970 <_sk_callback_sse41+0x280>
+  .byte  15,40,5,166,59,0,0                  // movaps        0x3ba6(%rip),%xmm0        # 49d0 <_sk_callback_sse41+0x27e>
   .byte  68,15,89,224                        // mulps         %xmm0,%xmm12
-  .byte  15,40,21,75,59,0,0                  // movaps        0x3b4b(%rip),%xmm2        # 4980 <_sk_callback_sse41+0x290>
+  .byte  15,40,21,171,59,0,0                 // movaps        0x3bab(%rip),%xmm2        # 49e0 <_sk_callback_sse41+0x28e>
   .byte  15,89,250                           // mulps         %xmm2,%xmm7
   .byte  65,15,88,252                        // addps         %xmm12,%xmm7
-  .byte  68,15,40,53,76,59,0,0               // movaps        0x3b4c(%rip),%xmm14        # 4990 <_sk_callback_sse41+0x2a0>
+  .byte  68,15,40,53,172,59,0,0              // movaps        0x3bac(%rip),%xmm14        # 49f0 <_sk_callback_sse41+0x29e>
   .byte  68,15,40,252                        // movaps        %xmm4,%xmm15
   .byte  69,15,89,254                        // mulps         %xmm14,%xmm15
   .byte  68,15,88,255                        // addps         %xmm7,%xmm15
@@ -21645,7 +21818,7 @@ _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  68,15,40,13,16,58,0,0               // movaps        0x3a10(%rip),%xmm9        # 49a0 <_sk_callback_sse41+0x2b0>
+  .byte  68,15,40,13,112,58,0,0              // movaps        0x3a70(%rip),%xmm9        # 4a00 <_sk_callback_sse41+0x2ae>
   .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
@@ -21739,14 +21912,14 @@ _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  15,40,21,202,56,0,0                 // movaps        0x38ca(%rip),%xmm2        # 49b0 <_sk_callback_sse41+0x2c0>
+  .byte  15,40,21,42,57,0,0                  // movaps        0x392a(%rip),%xmm2        # 4a10 <_sk_callback_sse41+0x2be>
   .byte  15,40,221                           // movaps        %xmm5,%xmm3
   .byte  15,89,218                           // mulps         %xmm2,%xmm3
-  .byte  15,40,13,205,56,0,0                 // movaps        0x38cd(%rip),%xmm1        # 49c0 <_sk_callback_sse41+0x2d0>
+  .byte  15,40,13,45,57,0,0                  // movaps        0x392d(%rip),%xmm1        # 4a20 <_sk_callback_sse41+0x2ce>
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
   .byte  15,89,249                           // mulps         %xmm1,%xmm7
   .byte  15,88,251                           // addps         %xmm3,%xmm7
-  .byte  68,15,40,45,204,56,0,0              // movaps        0x38cc(%rip),%xmm13        # 49d0 <_sk_callback_sse41+0x2e0>
+  .byte  68,15,40,45,44,57,0,0               // movaps        0x392c(%rip),%xmm13        # 4a30 <_sk_callback_sse41+0x2de>
   .byte  69,15,89,245                        // mulps         %xmm13,%xmm14
   .byte  68,15,88,247                        // addps         %xmm7,%xmm14
   .byte  65,15,40,218                        // movaps        %xmm10,%xmm3
@@ -21827,7 +22000,7 @@ _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  68,15,40,13,146,55,0,0              // movaps        0x3792(%rip),%xmm9        # 49e0 <_sk_callback_sse41+0x2f0>
+  .byte  68,15,40,13,242,55,0,0              // movaps        0x37f2(%rip),%xmm9        # 4a40 <_sk_callback_sse41+0x2ee>
   .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
@@ -21882,14 +22055,14 @@ _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  68,15,40,13,221,54,0,0              // movaps        0x36dd(%rip),%xmm9        # 49f0 <_sk_callback_sse41+0x300>
+  .byte  68,15,40,13,61,55,0,0               // movaps        0x373d(%rip),%xmm9        # 4a50 <_sk_callback_sse41+0x2fe>
   .byte  65,15,40,250                        // movaps        %xmm10,%xmm7
   .byte  65,15,89,249                        // mulps         %xmm9,%xmm7
-  .byte  68,15,40,21,221,54,0,0              // movaps        0x36dd(%rip),%xmm10        # 4a00 <_sk_callback_sse41+0x310>
+  .byte  68,15,40,21,61,55,0,0               // movaps        0x373d(%rip),%xmm10        # 4a60 <_sk_callback_sse41+0x30e>
   .byte  65,15,40,219                        // movaps        %xmm11,%xmm3
   .byte  65,15,89,218                        // mulps         %xmm10,%xmm3
   .byte  15,88,223                           // addps         %xmm7,%xmm3
-  .byte  68,15,40,29,218,54,0,0              // movaps        0x36da(%rip),%xmm11        # 4a10 <_sk_callback_sse41+0x320>
+  .byte  68,15,40,29,58,55,0,0               // movaps        0x373a(%rip),%xmm11        # 4a70 <_sk_callback_sse41+0x31e>
   .byte  69,15,40,236                        // movaps        %xmm12,%xmm13
   .byte  69,15,89,235                        // mulps         %xmm11,%xmm13
   .byte  68,15,88,235                        // addps         %xmm3,%xmm13
@@ -21974,7 +22147,7 @@ _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  68,15,40,13,150,53,0,0              // movaps        0x3596(%rip),%xmm9        # 4a20 <_sk_callback_sse41+0x330>
+  .byte  68,15,40,13,246,53,0,0              // movaps        0x35f6(%rip),%xmm9        # 4a80 <_sk_callback_sse41+0x32e>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  68,15,89,192                        // mulps         %xmm0,%xmm8
@@ -22026,13 +22199,13 @@ _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  68,15,40,5,238,52,0,0               // movaps        0x34ee(%rip),%xmm8        # 4a30 <_sk_callback_sse41+0x340>
+  .byte  68,15,40,5,78,53,0,0                // movaps        0x354e(%rip),%xmm8        # 4a90 <_sk_callback_sse41+0x33e>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
-  .byte  68,15,40,21,242,52,0,0              // movaps        0x34f2(%rip),%xmm10        # 4a40 <_sk_callback_sse41+0x350>
+  .byte  68,15,40,21,82,53,0,0               // movaps        0x3552(%rip),%xmm10        # 4aa0 <_sk_callback_sse41+0x34e>
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
   .byte  65,15,89,234                        // mulps         %xmm10,%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
-  .byte  68,15,40,37,240,52,0,0              // movaps        0x34f0(%rip),%xmm12        # 4a50 <_sk_callback_sse41+0x360>
+  .byte  68,15,40,37,80,53,0,0               // movaps        0x3550(%rip),%xmm12        # 4ab0 <_sk_callback_sse41+0x35e>
   .byte  68,15,40,242                        // movaps        %xmm2,%xmm14
   .byte  69,15,89,244                        // mulps         %xmm12,%xmm14
   .byte  68,15,88,245                        // addps         %xmm5,%xmm14
@@ -22117,7 +22290,7 @@ _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  15,40,5,166,51,0,0                  // movaps        0x33a6(%rip),%xmm0        # 4a60 <_sk_callback_sse41+0x370>
+  .byte  15,40,5,6,52,0,0                    // movaps        0x3406(%rip),%xmm0        # 4ac0 <_sk_callback_sse41+0x36e>
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,92,215                           // subps         %xmm7,%xmm2
   .byte  15,89,226                           // mulps         %xmm2,%xmm4
@@ -22166,7 +22339,7 @@ HIDDEN _sk_clamp_1_sse41
 .globl _sk_clamp_1_sse41
 FUNCTION(_sk_clamp_1_sse41)
 _sk_clamp_1_sse41:
-  .byte  68,15,40,5,41,51,0,0                // movaps        0x3329(%rip),%xmm8        # 4a70 <_sk_callback_sse41+0x380>
+  .byte  68,15,40,5,137,51,0,0               // movaps        0x3389(%rip),%xmm8        # 4ad0 <_sk_callback_sse41+0x37e>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
   .byte  65,15,93,208                        // minps         %xmm8,%xmm2
@@ -22178,7 +22351,7 @@ HIDDEN _sk_clamp_a_sse41
 .globl _sk_clamp_a_sse41
 FUNCTION(_sk_clamp_a_sse41)
 _sk_clamp_a_sse41:
-  .byte  15,93,29,30,51,0,0                  // minps         0x331e(%rip),%xmm3        # 4a80 <_sk_callback_sse41+0x390>
+  .byte  15,93,29,126,51,0,0                 // minps         0x337e(%rip),%xmm3        # 4ae0 <_sk_callback_sse41+0x38e>
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -22265,7 +22438,7 @@ HIDDEN _sk_unpremul_sse41
 FUNCTION(_sk_unpremul_sse41)
 _sk_unpremul_sse41:
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
-  .byte  68,15,40,13,137,50,0,0              // movaps        0x3289(%rip),%xmm9        # 4a90 <_sk_callback_sse41+0x3a0>
+  .byte  68,15,40,13,233,50,0,0              // movaps        0x32e9(%rip),%xmm9        # 4af0 <_sk_callback_sse41+0x39e>
   .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
@@ -22279,20 +22452,20 @@ HIDDEN _sk_from_srgb_sse41
 .globl _sk_from_srgb_sse41
 FUNCTION(_sk_from_srgb_sse41)
 _sk_from_srgb_sse41:
-  .byte  68,15,40,29,116,50,0,0              // movaps        0x3274(%rip),%xmm11        # 4aa0 <_sk_callback_sse41+0x3b0>
+  .byte  68,15,40,29,212,50,0,0              // movaps        0x32d4(%rip),%xmm11        # 4b00 <_sk_callback_sse41+0x3ae>
   .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  68,15,40,37,108,50,0,0              // movaps        0x326c(%rip),%xmm12        # 4ab0 <_sk_callback_sse41+0x3c0>
+  .byte  68,15,40,37,204,50,0,0              // movaps        0x32cc(%rip),%xmm12        # 4b10 <_sk_callback_sse41+0x3be>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,196                        // mulps         %xmm12,%xmm8
-  .byte  68,15,40,45,108,50,0,0              // movaps        0x326c(%rip),%xmm13        # 4ac0 <_sk_callback_sse41+0x3d0>
+  .byte  68,15,40,45,204,50,0,0              // movaps        0x32cc(%rip),%xmm13        # 4b20 <_sk_callback_sse41+0x3ce>
   .byte  69,15,88,197                        // addps         %xmm13,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
-  .byte  68,15,40,53,108,50,0,0              // movaps        0x326c(%rip),%xmm14        # 4ad0 <_sk_callback_sse41+0x3e0>
+  .byte  68,15,40,53,204,50,0,0              // movaps        0x32cc(%rip),%xmm14        # 4b30 <_sk_callback_sse41+0x3de>
   .byte  69,15,88,198                        // addps         %xmm14,%xmm8
-  .byte  68,15,40,61,112,50,0,0              // movaps        0x3270(%rip),%xmm15        # 4ae0 <_sk_callback_sse41+0x3f0>
+  .byte  68,15,40,61,208,50,0,0              // movaps        0x32d0(%rip),%xmm15        # 4b40 <_sk_callback_sse41+0x3ee>
   .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
@@ -22335,22 +22508,22 @@ _sk_to_srgb_sse41:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  15,40,209                           // movaps        %xmm1,%xmm2
   .byte  68,15,82,192                        // rsqrtps       %xmm0,%xmm8
-  .byte  68,15,40,29,232,49,0,0              // movaps        0x31e8(%rip),%xmm11        # 4af0 <_sk_callback_sse41+0x400>
+  .byte  68,15,40,29,72,50,0,0               // movaps        0x3248(%rip),%xmm11        # 4b50 <_sk_callback_sse41+0x3fe>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
-  .byte  68,15,40,37,232,49,0,0              // movaps        0x31e8(%rip),%xmm12        # 4b00 <_sk_callback_sse41+0x410>
+  .byte  68,15,40,37,72,50,0,0               // movaps        0x3248(%rip),%xmm12        # 4b60 <_sk_callback_sse41+0x40e>
   .byte  69,15,40,248                        // movaps        %xmm8,%xmm15
   .byte  69,15,89,252                        // mulps         %xmm12,%xmm15
-  .byte  68,15,40,21,232,49,0,0              // movaps        0x31e8(%rip),%xmm10        # 4b10 <_sk_callback_sse41+0x420>
+  .byte  68,15,40,21,72,50,0,0               // movaps        0x3248(%rip),%xmm10        # 4b70 <_sk_callback_sse41+0x41e>
   .byte  69,15,88,250                        // addps         %xmm10,%xmm15
   .byte  69,15,89,248                        // mulps         %xmm8,%xmm15
-  .byte  68,15,40,45,232,49,0,0              // movaps        0x31e8(%rip),%xmm13        # 4b20 <_sk_callback_sse41+0x430>
+  .byte  68,15,40,45,72,50,0,0               // movaps        0x3248(%rip),%xmm13        # 4b80 <_sk_callback_sse41+0x42e>
   .byte  69,15,88,253                        // addps         %xmm13,%xmm15
-  .byte  68,15,40,53,236,49,0,0              // movaps        0x31ec(%rip),%xmm14        # 4b30 <_sk_callback_sse41+0x440>
+  .byte  68,15,40,53,76,50,0,0               // movaps        0x324c(%rip),%xmm14        # 4b90 <_sk_callback_sse41+0x43e>
   .byte  69,15,88,198                        // addps         %xmm14,%xmm8
   .byte  69,15,83,192                        // rcpps         %xmm8,%xmm8
   .byte  69,15,89,199                        // mulps         %xmm15,%xmm8
-  .byte  68,15,40,61,232,49,0,0              // movaps        0x31e8(%rip),%xmm15        # 4b40 <_sk_callback_sse41+0x450>
+  .byte  68,15,40,61,72,50,0,0               // movaps        0x3248(%rip),%xmm15        # 4ba0 <_sk_callback_sse41+0x44e>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  102,69,15,56,20,193                 // blendvps      %xmm0,%xmm9,%xmm8
   .byte  68,15,82,202                        // rsqrtps       %xmm2,%xmm9
@@ -22403,7 +22576,7 @@ _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  68,15,40,53,53,49,0,0               // movaps        0x3135(%rip),%xmm14        # 4b50 <_sk_callback_sse41+0x460>
+  .byte  68,15,40,53,149,49,0,0              // movaps        0x3195(%rip),%xmm14        # 4bb0 <_sk_callback_sse41+0x45e>
   .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
@@ -22412,27 +22585,27 @@ _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  68,15,84,61,28,49,0,0               // andps         0x311c(%rip),%xmm15        # 4b60 <_sk_callback_sse41+0x470>
+  .byte  68,15,84,61,124,49,0,0              // andps         0x317c(%rip),%xmm15        # 4bc0 <_sk_callback_sse41+0x46e>
   .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  68,15,40,45,15,49,0,0               // movaps        0x310f(%rip),%xmm13        # 4b70 <_sk_callback_sse41+0x480>
+  .byte  68,15,40,45,111,49,0,0              // movaps        0x316f(%rip),%xmm13        # 4bd0 <_sk_callback_sse41+0x47e>
   .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  68,15,88,5,11,49,0,0                // addps         0x310b(%rip),%xmm8        # 4b80 <_sk_callback_sse41+0x490>
+  .byte  68,15,88,5,107,49,0,0               // addps         0x316b(%rip),%xmm8        # 4be0 <_sk_callback_sse41+0x48e>
   .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  68,15,89,5,3,49,0,0                 // mulps         0x3103(%rip),%xmm8        # 4b90 <_sk_callback_sse41+0x4a0>
+  .byte  68,15,89,5,99,49,0,0                // mulps         0x3163(%rip),%xmm8        # 4bf0 <_sk_callback_sse41+0x49e>
   .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  15,40,5,247,48,0,0                  // movaps        0x30f7(%rip),%xmm0        # 4ba0 <_sk_callback_sse41+0x4b0>
+  .byte  15,40,5,87,49,0,0                   // movaps        0x3157(%rip),%xmm0        # 4c00 <_sk_callback_sse41+0x4ae>
   .byte  65,15,40,211                        // movaps        %xmm11,%xmm2
   .byte  15,89,208                           // mulps         %xmm0,%xmm2
   .byte  15,194,194,1                        // cmpltps       %xmm2,%xmm0
@@ -22454,7 +22627,7 @@ _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  68,15,40,13,189,48,0,0              // movaps        0x30bd(%rip),%xmm9        # 4bb0 <_sk_callback_sse41+0x4c0>
+  .byte  68,15,40,13,29,49,0,0               // movaps        0x311d(%rip),%xmm9        # 4c10 <_sk_callback_sse41+0x4be>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,194,194,2                        // cmpleps       %xmm2,%xmm0
   .byte  15,40,217                           // movaps        %xmm1,%xmm3
@@ -22467,19 +22640,19 @@ _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  68,15,40,53,152,48,0,0              // movaps        0x3098(%rip),%xmm14        # 4bc0 <_sk_callback_sse41+0x4d0>
+  .byte  68,15,40,53,248,48,0,0              // movaps        0x30f8(%rip),%xmm14        # 4c20 <_sk_callback_sse41+0x4ce>
   .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  68,15,40,29,145,48,0,0              // movaps        0x3091(%rip),%xmm11        # 4bd0 <_sk_callback_sse41+0x4e0>
+  .byte  68,15,40,29,241,48,0,0              // movaps        0x30f1(%rip),%xmm11        # 4c30 <_sk_callback_sse41+0x4de>
   .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  15,40,61,138,48,0,0                 // movaps        0x308a(%rip),%xmm7        # 4be0 <_sk_callback_sse41+0x4f0>
+  .byte  15,40,61,234,48,0,0                 // movaps        0x30ea(%rip),%xmm7        # 4c40 <_sk_callback_sse41+0x4ee>
   .byte  69,15,40,238                        // movaps        %xmm14,%xmm13
   .byte  68,15,89,239                        // mulps         %xmm7,%xmm13
-  .byte  15,40,29,139,48,0,0                 // movaps        0x308b(%rip),%xmm3        # 4bf0 <_sk_callback_sse41+0x500>
+  .byte  15,40,29,235,48,0,0                 // movaps        0x30eb(%rip),%xmm3        # 4c50 <_sk_callback_sse41+0x4fe>
   .byte  68,15,40,227                        // movaps        %xmm3,%xmm12
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  68,15,89,230                        // mulps         %xmm6,%xmm12
@@ -22489,7 +22662,7 @@ _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  68,15,40,37,106,48,0,0              // movaps        0x306a(%rip),%xmm12        # 4c00 <_sk_callback_sse41+0x510>
+  .byte  68,15,40,37,202,48,0,0              // movaps        0x30ca(%rip),%xmm12        # 4c60 <_sk_callback_sse41+0x50e>
   .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
@@ -22523,7 +22696,7 @@ _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  68,15,88,21,226,47,0,0              // addps         0x2fe2(%rip),%xmm10        # 4c10 <_sk_callback_sse41+0x520>
+  .byte  68,15,88,21,66,48,0,0               // addps         0x3042(%rip),%xmm10        # 4c70 <_sk_callback_sse41+0x51e>
   .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
@@ -22575,7 +22748,7 @@ _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  68,15,89,5,63,47,0,0                // mulps         0x2f3f(%rip),%xmm8        # 4c20 <_sk_callback_sse41+0x530>
+  .byte  68,15,89,5,159,47,0,0               // mulps         0x2f9f(%rip),%xmm8        # 4c80 <_sk_callback_sse41+0x52e>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
@@ -22613,7 +22786,7 @@ _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  68,15,89,5,235,46,0,0               // mulps         0x2eeb(%rip),%xmm8        # 4c30 <_sk_callback_sse41+0x540>
+  .byte  68,15,89,5,75,47,0,0                // mulps         0x2f4b(%rip),%xmm8        # 4c90 <_sk_callback_sse41+0x53e>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -22636,17 +22809,17 @@ _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  102,68,15,111,5,186,46,0,0          // movdqa        0x2eba(%rip),%xmm8        # 4c40 <_sk_callback_sse41+0x550>
+  .byte  102,68,15,111,5,26,47,0,0           // movdqa        0x2f1a(%rip),%xmm8        # 4ca0 <_sk_callback_sse41+0x54e>
   .byte  102,69,15,219,194                   // pand          %xmm10,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,185,46,0,0               // mulps         0x2eb9(%rip),%xmm8        # 4c50 <_sk_callback_sse41+0x560>
-  .byte  102,68,15,111,13,192,46,0,0         // movdqa        0x2ec0(%rip),%xmm9        # 4c60 <_sk_callback_sse41+0x570>
+  .byte  68,15,89,5,25,47,0,0                // mulps         0x2f19(%rip),%xmm8        # 4cb0 <_sk_callback_sse41+0x55e>
+  .byte  102,68,15,111,13,32,47,0,0          // movdqa        0x2f20(%rip),%xmm9        # 4cc0 <_sk_callback_sse41+0x56e>
   .byte  102,69,15,219,202                   // pand          %xmm10,%xmm9
   .byte  69,15,91,201                        // cvtdq2ps      %xmm9,%xmm9
-  .byte  68,15,89,13,191,46,0,0              // mulps         0x2ebf(%rip),%xmm9        # 4c70 <_sk_callback_sse41+0x580>
-  .byte  102,68,15,219,21,198,46,0,0         // pand          0x2ec6(%rip),%xmm10        # 4c80 <_sk_callback_sse41+0x590>
+  .byte  68,15,89,13,31,47,0,0               // mulps         0x2f1f(%rip),%xmm9        # 4cd0 <_sk_callback_sse41+0x57e>
+  .byte  102,68,15,219,21,38,47,0,0          // pand          0x2f26(%rip),%xmm10        # 4ce0 <_sk_callback_sse41+0x58e>
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
-  .byte  68,15,89,21,202,46,0,0              // mulps         0x2eca(%rip),%xmm10        # 4c90 <_sk_callback_sse41+0x5a0>
+  .byte  68,15,89,21,42,47,0,0               // mulps         0x2f2a(%rip),%xmm10        # 4cf0 <_sk_callback_sse41+0x59e>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -22677,7 +22850,7 @@ _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  102,15,111,5,123,46,0,0             // movdqa        0x2e7b(%rip),%xmm0        # 4ca0 <_sk_callback_sse41+0x5b0>
+  .byte  102,15,111,5,219,46,0,0             // movdqa        0x2edb(%rip),%xmm0        # 4d00 <_sk_callback_sse41+0x5ae>
   .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
@@ -22692,7 +22865,7 @@ _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  102,15,56,0,13,54,46,0,0            // pshufb        0x2e36(%rip),%xmm1        # 4cb0 <_sk_callback_sse41+0x5c0>
+  .byte  102,15,56,0,13,150,46,0,0           // pshufb        0x2e96(%rip),%xmm1        # 4d10 <_sk_callback_sse41+0x5be>
   .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
@@ -22707,7 +22880,7 @@ _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  102,15,56,0,21,242,45,0,0           // pshufb        0x2df2(%rip),%xmm2        # 4cc0 <_sk_callback_sse41+0x5d0>
+  .byte  102,15,56,0,21,82,46,0,0            // pshufb        0x2e52(%rip),%xmm2        # 4d20 <_sk_callback_sse41+0x5ce>
   .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
@@ -22722,7 +22895,7 @@ _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  15,89,29,175,45,0,0                 // mulps         0x2daf(%rip),%xmm3        # 4cd0 <_sk_callback_sse41+0x5e0>
+  .byte  15,89,29,15,46,0,0                  // mulps         0x2e0f(%rip),%xmm3        # 4d30 <_sk_callback_sse41+0x5de>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -22741,7 +22914,7 @@ _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  102,68,15,111,5,130,45,0,0          // movdqa        0x2d82(%rip),%xmm8        # 4ce0 <_sk_callback_sse41+0x5f0>
+  .byte  102,68,15,111,5,226,45,0,0          // movdqa        0x2de2(%rip),%xmm8        # 4d40 <_sk_callback_sse41+0x5ee>
   .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
@@ -22758,7 +22931,7 @@ _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  102,15,56,0,13,53,45,0,0            // pshufb        0x2d35(%rip),%xmm1        # 4cf0 <_sk_callback_sse41+0x600>
+  .byte  102,15,56,0,13,149,45,0,0           // pshufb        0x2d95(%rip),%xmm1        # 4d50 <_sk_callback_sse41+0x5fe>
   .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
@@ -22794,7 +22967,7 @@ _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  15,89,29,131,44,0,0                 // mulps         0x2c83(%rip),%xmm3        # 4d00 <_sk_callback_sse41+0x610>
+  .byte  15,89,29,227,44,0,0                 // mulps         0x2ce3(%rip),%xmm3        # 4d60 <_sk_callback_sse41+0x60e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -22816,7 +22989,7 @@ _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  102,68,15,111,5,69,44,0,0           // movdqa        0x2c45(%rip),%xmm8        # 4d10 <_sk_callback_sse41+0x620>
+  .byte  102,68,15,111,5,165,44,0,0          // movdqa        0x2ca5(%rip),%xmm8        # 4d70 <_sk_callback_sse41+0x61e>
   .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
@@ -22833,7 +23006,7 @@ _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  102,15,56,0,13,248,43,0,0           // pshufb        0x2bf8(%rip),%xmm1        # 4d20 <_sk_callback_sse41+0x630>
+  .byte  102,15,56,0,13,88,44,0,0            // pshufb        0x2c58(%rip),%xmm1        # 4d80 <_sk_callback_sse41+0x62e>
   .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
@@ -22864,7 +23037,7 @@ _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  15,40,29,99,43,0,0                  // movaps        0x2b63(%rip),%xmm3        # 4d30 <_sk_callback_sse41+0x640>
+  .byte  15,40,29,195,43,0,0                 // movaps        0x2bc3(%rip),%xmm3        # 4d90 <_sk_callback_sse41+0x63e>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_byte_tables_sse41
@@ -22874,7 +23047,7 @@ _sk_byte_tables_sse41:
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,100,43,0,0               // movaps        0x2b64(%rip),%xmm8        # 4d40 <_sk_callback_sse41+0x650>
+  .byte  68,15,40,5,196,43,0,0               // movaps        0x2bc4(%rip),%xmm8        # 4da0 <_sk_callback_sse41+0x64e>
   .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
@@ -22893,7 +23066,7 @@ _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  68,15,40,13,21,43,0,0               // movaps        0x2b15(%rip),%xmm9        # 4d50 <_sk_callback_sse41+0x660>
+  .byte  68,15,40,13,117,43,0,0              // movaps        0x2b75(%rip),%xmm9        # 4db0 <_sk_callback_sse41+0x65e>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -22984,7 +23157,7 @@ _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  68,15,40,13,157,41,0,0              // movaps        0x299d(%rip),%xmm9        # 4d60 <_sk_callback_sse41+0x670>
+  .byte  68,15,40,13,253,41,0,0              // movaps        0x29fd(%rip),%xmm9        # 4dc0 <_sk_callback_sse41+0x66e>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -23161,31 +23334,31 @@ _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  68,15,89,5,244,38,0,0               // mulps         0x26f4(%rip),%xmm8        # 4d70 <_sk_callback_sse41+0x680>
-  .byte  68,15,84,21,252,38,0,0              // andps         0x26fc(%rip),%xmm10        # 4d80 <_sk_callback_sse41+0x690>
-  .byte  68,15,86,21,4,39,0,0                // orps          0x2704(%rip),%xmm10        # 4d90 <_sk_callback_sse41+0x6a0>
-  .byte  68,15,88,5,12,39,0,0                // addps         0x270c(%rip),%xmm8        # 4da0 <_sk_callback_sse41+0x6b0>
-  .byte  68,15,40,37,20,39,0,0               // movaps        0x2714(%rip),%xmm12        # 4db0 <_sk_callback_sse41+0x6c0>
+  .byte  68,15,89,5,84,39,0,0                // mulps         0x2754(%rip),%xmm8        # 4dd0 <_sk_callback_sse41+0x67e>
+  .byte  68,15,84,21,92,39,0,0               // andps         0x275c(%rip),%xmm10        # 4de0 <_sk_callback_sse41+0x68e>
+  .byte  68,15,86,21,100,39,0,0              // orps          0x2764(%rip),%xmm10        # 4df0 <_sk_callback_sse41+0x69e>
+  .byte  68,15,88,5,108,39,0,0               // addps         0x276c(%rip),%xmm8        # 4e00 <_sk_callback_sse41+0x6ae>
+  .byte  68,15,40,37,116,39,0,0              // movaps        0x2774(%rip),%xmm12        # 4e10 <_sk_callback_sse41+0x6be>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,196                        // subps         %xmm12,%xmm8
-  .byte  68,15,88,21,20,39,0,0               // addps         0x2714(%rip),%xmm10        # 4dc0 <_sk_callback_sse41+0x6d0>
-  .byte  68,15,40,37,28,39,0,0               // movaps        0x271c(%rip),%xmm12        # 4dd0 <_sk_callback_sse41+0x6e0>
+  .byte  68,15,88,21,116,39,0,0              // addps         0x2774(%rip),%xmm10        # 4e20 <_sk_callback_sse41+0x6ce>
+  .byte  68,15,40,37,124,39,0,0              // movaps        0x277c(%rip),%xmm12        # 4e30 <_sk_callback_sse41+0x6de>
   .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  68,15,88,5,9,39,0,0                 // addps         0x2709(%rip),%xmm8        # 4de0 <_sk_callback_sse41+0x6f0>
-  .byte  68,15,40,21,17,39,0,0               // movaps        0x2711(%rip),%xmm10        # 4df0 <_sk_callback_sse41+0x700>
+  .byte  68,15,88,5,105,39,0,0               // addps         0x2769(%rip),%xmm8        # 4e40 <_sk_callback_sse41+0x6ee>
+  .byte  68,15,40,21,113,39,0,0              // movaps        0x2771(%rip),%xmm10        # 4e50 <_sk_callback_sse41+0x6fe>
   .byte  69,15,89,211                        // mulps         %xmm11,%xmm10
   .byte  69,15,92,194                        // subps         %xmm10,%xmm8
-  .byte  68,15,40,21,17,39,0,0               // movaps        0x2711(%rip),%xmm10        # 4e00 <_sk_callback_sse41+0x710>
+  .byte  68,15,40,21,113,39,0,0              // movaps        0x2771(%rip),%xmm10        # 4e60 <_sk_callback_sse41+0x70e>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  68,15,40,29,21,39,0,0               // movaps        0x2715(%rip),%xmm11        # 4e10 <_sk_callback_sse41+0x720>
+  .byte  68,15,40,29,117,39,0,0              // movaps        0x2775(%rip),%xmm11        # 4e70 <_sk_callback_sse41+0x71e>
   .byte  69,15,94,218                        // divps         %xmm10,%xmm11
   .byte  69,15,88,216                        // addps         %xmm8,%xmm11
-  .byte  68,15,89,29,21,39,0,0               // mulps         0x2715(%rip),%xmm11        # 4e20 <_sk_callback_sse41+0x730>
+  .byte  68,15,89,29,117,39,0,0              // mulps         0x2775(%rip),%xmm11        # 4e80 <_sk_callback_sse41+0x72e>
   .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
@@ -23193,7 +23366,7 @@ _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  68,15,93,5,252,38,0,0               // minps         0x26fc(%rip),%xmm8        # 4e30 <_sk_callback_sse41+0x740>
+  .byte  68,15,93,5,92,39,0,0                // minps         0x275c(%rip),%xmm8        # 4e90 <_sk_callback_sse41+0x73e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23223,31 +23396,31 @@ _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  68,15,89,37,157,38,0,0              // mulps         0x269d(%rip),%xmm12        # 4e40 <_sk_callback_sse41+0x750>
-  .byte  68,15,84,29,165,38,0,0              // andps         0x26a5(%rip),%xmm11        # 4e50 <_sk_callback_sse41+0x760>
-  .byte  68,15,86,29,173,38,0,0              // orps          0x26ad(%rip),%xmm11        # 4e60 <_sk_callback_sse41+0x770>
-  .byte  68,15,88,37,181,38,0,0              // addps         0x26b5(%rip),%xmm12        # 4e70 <_sk_callback_sse41+0x780>
-  .byte  15,40,13,190,38,0,0                 // movaps        0x26be(%rip),%xmm1        # 4e80 <_sk_callback_sse41+0x790>
+  .byte  68,15,89,37,253,38,0,0              // mulps         0x26fd(%rip),%xmm12        # 4ea0 <_sk_callback_sse41+0x74e>
+  .byte  68,15,84,29,5,39,0,0                // andps         0x2705(%rip),%xmm11        # 4eb0 <_sk_callback_sse41+0x75e>
+  .byte  68,15,86,29,13,39,0,0               // orps          0x270d(%rip),%xmm11        # 4ec0 <_sk_callback_sse41+0x76e>
+  .byte  68,15,88,37,21,39,0,0               // addps         0x2715(%rip),%xmm12        # 4ed0 <_sk_callback_sse41+0x77e>
+  .byte  15,40,13,30,39,0,0                  // movaps        0x271e(%rip),%xmm1        # 4ee0 <_sk_callback_sse41+0x78e>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
-  .byte  68,15,88,29,190,38,0,0              // addps         0x26be(%rip),%xmm11        # 4e90 <_sk_callback_sse41+0x7a0>
-  .byte  15,40,13,199,38,0,0                 // movaps        0x26c7(%rip),%xmm1        # 4ea0 <_sk_callback_sse41+0x7b0>
+  .byte  68,15,88,29,30,39,0,0               // addps         0x271e(%rip),%xmm11        # 4ef0 <_sk_callback_sse41+0x79e>
+  .byte  15,40,13,39,39,0,0                  // movaps        0x2727(%rip),%xmm1        # 4f00 <_sk_callback_sse41+0x7ae>
   .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  68,15,88,37,180,38,0,0              // addps         0x26b4(%rip),%xmm12        # 4eb0 <_sk_callback_sse41+0x7c0>
-  .byte  15,40,13,189,38,0,0                 // movaps        0x26bd(%rip),%xmm1        # 4ec0 <_sk_callback_sse41+0x7d0>
+  .byte  68,15,88,37,20,39,0,0               // addps         0x2714(%rip),%xmm12        # 4f10 <_sk_callback_sse41+0x7be>
+  .byte  15,40,13,29,39,0,0                  // movaps        0x271d(%rip),%xmm1        # 4f20 <_sk_callback_sse41+0x7ce>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
-  .byte  68,15,40,21,189,38,0,0              // movaps        0x26bd(%rip),%xmm10        # 4ed0 <_sk_callback_sse41+0x7e0>
+  .byte  68,15,40,21,29,39,0,0               // movaps        0x271d(%rip),%xmm10        # 4f30 <_sk_callback_sse41+0x7de>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,13,194,38,0,0                 // movaps        0x26c2(%rip),%xmm1        # 4ee0 <_sk_callback_sse41+0x7f0>
+  .byte  15,40,13,34,39,0,0                  // movaps        0x2722(%rip),%xmm1        # 4f40 <_sk_callback_sse41+0x7ee>
   .byte  65,15,94,202                        // divps         %xmm10,%xmm1
   .byte  65,15,88,204                        // addps         %xmm12,%xmm1
-  .byte  15,89,13,195,38,0,0                 // mulps         0x26c3(%rip),%xmm1        # 4ef0 <_sk_callback_sse41+0x800>
+  .byte  15,89,13,35,39,0,0                  // mulps         0x2723(%rip),%xmm1        # 4f50 <_sk_callback_sse41+0x7fe>
   .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
@@ -23255,7 +23428,7 @@ _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  15,93,13,174,38,0,0                 // minps         0x26ae(%rip),%xmm1        # 4f00 <_sk_callback_sse41+0x810>
+  .byte  15,93,13,14,39,0,0                  // minps         0x270e(%rip),%xmm1        # 4f60 <_sk_callback_sse41+0x80e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23285,31 +23458,31 @@ _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  68,15,89,37,79,38,0,0               // mulps         0x264f(%rip),%xmm12        # 4f10 <_sk_callback_sse41+0x820>
-  .byte  68,15,84,29,87,38,0,0               // andps         0x2657(%rip),%xmm11        # 4f20 <_sk_callback_sse41+0x830>
-  .byte  68,15,86,29,95,38,0,0               // orps          0x265f(%rip),%xmm11        # 4f30 <_sk_callback_sse41+0x840>
-  .byte  68,15,88,37,103,38,0,0              // addps         0x2667(%rip),%xmm12        # 4f40 <_sk_callback_sse41+0x850>
-  .byte  15,40,21,112,38,0,0                 // movaps        0x2670(%rip),%xmm2        # 4f50 <_sk_callback_sse41+0x860>
+  .byte  68,15,89,37,175,38,0,0              // mulps         0x26af(%rip),%xmm12        # 4f70 <_sk_callback_sse41+0x81e>
+  .byte  68,15,84,29,183,38,0,0              // andps         0x26b7(%rip),%xmm11        # 4f80 <_sk_callback_sse41+0x82e>
+  .byte  68,15,86,29,191,38,0,0              // orps          0x26bf(%rip),%xmm11        # 4f90 <_sk_callback_sse41+0x83e>
+  .byte  68,15,88,37,199,38,0,0              // addps         0x26c7(%rip),%xmm12        # 4fa0 <_sk_callback_sse41+0x84e>
+  .byte  15,40,21,208,38,0,0                 // movaps        0x26d0(%rip),%xmm2        # 4fb0 <_sk_callback_sse41+0x85e>
   .byte  65,15,89,211                        // mulps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
-  .byte  68,15,88,29,112,38,0,0              // addps         0x2670(%rip),%xmm11        # 4f60 <_sk_callback_sse41+0x870>
-  .byte  15,40,21,121,38,0,0                 // movaps        0x2679(%rip),%xmm2        # 4f70 <_sk_callback_sse41+0x880>
+  .byte  68,15,88,29,208,38,0,0              // addps         0x26d0(%rip),%xmm11        # 4fc0 <_sk_callback_sse41+0x86e>
+  .byte  15,40,21,217,38,0,0                 // movaps        0x26d9(%rip),%xmm2        # 4fd0 <_sk_callback_sse41+0x87e>
   .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  68,15,88,37,102,38,0,0              // addps         0x2666(%rip),%xmm12        # 4f80 <_sk_callback_sse41+0x890>
-  .byte  15,40,21,111,38,0,0                 // movaps        0x266f(%rip),%xmm2        # 4f90 <_sk_callback_sse41+0x8a0>
+  .byte  68,15,88,37,198,38,0,0              // addps         0x26c6(%rip),%xmm12        # 4fe0 <_sk_callback_sse41+0x88e>
+  .byte  15,40,21,207,38,0,0                 // movaps        0x26cf(%rip),%xmm2        # 4ff0 <_sk_callback_sse41+0x89e>
   .byte  65,15,89,211                        // mulps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
-  .byte  68,15,40,21,111,38,0,0              // movaps        0x266f(%rip),%xmm10        # 4fa0 <_sk_callback_sse41+0x8b0>
+  .byte  68,15,40,21,207,38,0,0              // movaps        0x26cf(%rip),%xmm10        # 5000 <_sk_callback_sse41+0x8ae>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,21,116,38,0,0                 // movaps        0x2674(%rip),%xmm2        # 4fb0 <_sk_callback_sse41+0x8c0>
+  .byte  15,40,21,212,38,0,0                 // movaps        0x26d4(%rip),%xmm2        # 5010 <_sk_callback_sse41+0x8be>
   .byte  65,15,94,210                        // divps         %xmm10,%xmm2
   .byte  65,15,88,212                        // addps         %xmm12,%xmm2
-  .byte  15,89,21,117,38,0,0                 // mulps         0x2675(%rip),%xmm2        # 4fc0 <_sk_callback_sse41+0x8d0>
+  .byte  15,89,21,213,38,0,0                 // mulps         0x26d5(%rip),%xmm2        # 5020 <_sk_callback_sse41+0x8ce>
   .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
@@ -23317,7 +23490,7 @@ _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  15,93,21,96,38,0,0                  // minps         0x2660(%rip),%xmm2        # 4fd0 <_sk_callback_sse41+0x8e0>
+  .byte  15,93,21,192,38,0,0                 // minps         0x26c0(%rip),%xmm2        # 5030 <_sk_callback_sse41+0x8de>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23347,31 +23520,31 @@ _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  68,15,89,37,1,38,0,0                // mulps         0x2601(%rip),%xmm12        # 4fe0 <_sk_callback_sse41+0x8f0>
-  .byte  68,15,84,29,9,38,0,0                // andps         0x2609(%rip),%xmm11        # 4ff0 <_sk_callback_sse41+0x900>
-  .byte  68,15,86,29,17,38,0,0               // orps          0x2611(%rip),%xmm11        # 5000 <_sk_callback_sse41+0x910>
-  .byte  68,15,88,37,25,38,0,0               // addps         0x2619(%rip),%xmm12        # 5010 <_sk_callback_sse41+0x920>
-  .byte  15,40,29,34,38,0,0                  // movaps        0x2622(%rip),%xmm3        # 5020 <_sk_callback_sse41+0x930>
+  .byte  68,15,89,37,97,38,0,0               // mulps         0x2661(%rip),%xmm12        # 5040 <_sk_callback_sse41+0x8ee>
+  .byte  68,15,84,29,105,38,0,0              // andps         0x2669(%rip),%xmm11        # 5050 <_sk_callback_sse41+0x8fe>
+  .byte  68,15,86,29,113,38,0,0              // orps          0x2671(%rip),%xmm11        # 5060 <_sk_callback_sse41+0x90e>
+  .byte  68,15,88,37,121,38,0,0              // addps         0x2679(%rip),%xmm12        # 5070 <_sk_callback_sse41+0x91e>
+  .byte  15,40,29,130,38,0,0                 // movaps        0x2682(%rip),%xmm3        # 5080 <_sk_callback_sse41+0x92e>
   .byte  65,15,89,219                        // mulps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
-  .byte  68,15,88,29,34,38,0,0               // addps         0x2622(%rip),%xmm11        # 5030 <_sk_callback_sse41+0x940>
-  .byte  15,40,29,43,38,0,0                  // movaps        0x262b(%rip),%xmm3        # 5040 <_sk_callback_sse41+0x950>
+  .byte  68,15,88,29,130,38,0,0              // addps         0x2682(%rip),%xmm11        # 5090 <_sk_callback_sse41+0x93e>
+  .byte  15,40,29,139,38,0,0                 // movaps        0x268b(%rip),%xmm3        # 50a0 <_sk_callback_sse41+0x94e>
   .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  68,15,88,37,24,38,0,0               // addps         0x2618(%rip),%xmm12        # 5050 <_sk_callback_sse41+0x960>
-  .byte  15,40,29,33,38,0,0                  // movaps        0x2621(%rip),%xmm3        # 5060 <_sk_callback_sse41+0x970>
+  .byte  68,15,88,37,120,38,0,0              // addps         0x2678(%rip),%xmm12        # 50b0 <_sk_callback_sse41+0x95e>
+  .byte  15,40,29,129,38,0,0                 // movaps        0x2681(%rip),%xmm3        # 50c0 <_sk_callback_sse41+0x96e>
   .byte  65,15,89,219                        // mulps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
-  .byte  68,15,40,21,33,38,0,0               // movaps        0x2621(%rip),%xmm10        # 5070 <_sk_callback_sse41+0x980>
+  .byte  68,15,40,21,129,38,0,0              // movaps        0x2681(%rip),%xmm10        # 50d0 <_sk_callback_sse41+0x97e>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,29,38,38,0,0                  // movaps        0x2626(%rip),%xmm3        # 5080 <_sk_callback_sse41+0x990>
+  .byte  15,40,29,134,38,0,0                 // movaps        0x2686(%rip),%xmm3        # 50e0 <_sk_callback_sse41+0x98e>
   .byte  65,15,94,218                        // divps         %xmm10,%xmm3
   .byte  65,15,88,220                        // addps         %xmm12,%xmm3
-  .byte  15,89,29,39,38,0,0                  // mulps         0x2627(%rip),%xmm3        # 5090 <_sk_callback_sse41+0x9a0>
+  .byte  15,89,29,135,38,0,0                 // mulps         0x2687(%rip),%xmm3        # 50f0 <_sk_callback_sse41+0x99e>
   .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
@@ -23379,7 +23552,7 @@ _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  15,93,29,18,38,0,0                  // minps         0x2612(%rip),%xmm3        # 50a0 <_sk_callback_sse41+0x9b0>
+  .byte  15,93,29,114,38,0,0                 // minps         0x2672(%rip),%xmm3        # 5100 <_sk_callback_sse41+0x9ae>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -23389,29 +23562,29 @@ 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  68,15,89,5,14,38,0,0                // mulps         0x260e(%rip),%xmm8        # 50b0 <_sk_callback_sse41+0x9c0>
-  .byte  68,15,40,13,22,38,0,0               // movaps        0x2616(%rip),%xmm9        # 50c0 <_sk_callback_sse41+0x9d0>
+  .byte  68,15,89,5,110,38,0,0               // mulps         0x266e(%rip),%xmm8        # 5110 <_sk_callback_sse41+0x9be>
+  .byte  68,15,40,13,118,38,0,0              // movaps        0x2676(%rip),%xmm9        # 5120 <_sk_callback_sse41+0x9ce>
   .byte  65,15,89,201                        // mulps         %xmm9,%xmm1
-  .byte  15,40,5,27,38,0,0                   // movaps        0x261b(%rip),%xmm0        # 50d0 <_sk_callback_sse41+0x9e0>
+  .byte  15,40,5,123,38,0,0                  // movaps        0x267b(%rip),%xmm0        # 5130 <_sk_callback_sse41+0x9de>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
-  .byte  68,15,88,5,25,38,0,0                // addps         0x2619(%rip),%xmm8        # 50e0 <_sk_callback_sse41+0x9f0>
-  .byte  68,15,89,5,33,38,0,0                // mulps         0x2621(%rip),%xmm8        # 50f0 <_sk_callback_sse41+0xa00>
-  .byte  15,89,13,42,38,0,0                  // mulps         0x262a(%rip),%xmm1        # 5100 <_sk_callback_sse41+0xa10>
+  .byte  68,15,88,5,121,38,0,0               // addps         0x2679(%rip),%xmm8        # 5140 <_sk_callback_sse41+0x9ee>
+  .byte  68,15,89,5,129,38,0,0               // mulps         0x2681(%rip),%xmm8        # 5150 <_sk_callback_sse41+0x9fe>
+  .byte  15,89,13,138,38,0,0                 // mulps         0x268a(%rip),%xmm1        # 5160 <_sk_callback_sse41+0xa0e>
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  15,89,21,47,38,0,0                  // mulps         0x262f(%rip),%xmm2        # 5110 <_sk_callback_sse41+0xa20>
+  .byte  15,89,21,143,38,0,0                 // mulps         0x268f(%rip),%xmm2        # 5170 <_sk_callback_sse41+0xa1e>
   .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  68,15,40,13,35,38,0,0               // movaps        0x2623(%rip),%xmm9        # 5120 <_sk_callback_sse41+0xa30>
+  .byte  68,15,40,13,131,38,0,0              // movaps        0x2683(%rip),%xmm9        # 5180 <_sk_callback_sse41+0xa2e>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  65,15,194,195,1                     // cmpltps       %xmm11,%xmm0
-  .byte  15,40,21,35,38,0,0                  // movaps        0x2623(%rip),%xmm2        # 5130 <_sk_callback_sse41+0xa40>
+  .byte  15,40,21,131,38,0,0                 // movaps        0x2683(%rip),%xmm2        # 5190 <_sk_callback_sse41+0xa3e>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
-  .byte  68,15,40,37,40,38,0,0               // movaps        0x2628(%rip),%xmm12        # 5140 <_sk_callback_sse41+0xa50>
+  .byte  68,15,40,37,136,38,0,0              // movaps        0x2688(%rip),%xmm12        # 51a0 <_sk_callback_sse41+0xa4e>
   .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
@@ -23430,8 +23603,8 @@ _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  15,89,13,225,37,0,0                 // mulps         0x25e1(%rip),%xmm1        # 5150 <_sk_callback_sse41+0xa60>
-  .byte  15,89,21,234,37,0,0                 // mulps         0x25ea(%rip),%xmm2        # 5160 <_sk_callback_sse41+0xa70>
+  .byte  15,89,13,65,38,0,0                  // mulps         0x2641(%rip),%xmm1        # 51b0 <_sk_callback_sse41+0xa5e>
+  .byte  15,89,21,74,38,0,0                  // mulps         0x264a(%rip),%xmm2        # 51c0 <_sk_callback_sse41+0xa6e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  65,15,40,200                        // movaps        %xmm8,%xmm1
@@ -23445,7 +23618,7 @@ _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  15,89,29,218,37,0,0                 // mulps         0x25da(%rip),%xmm3        # 5170 <_sk_callback_sse41+0xa80>
+  .byte  15,89,29,58,38,0,0                  // mulps         0x263a(%rip),%xmm3        # 51d0 <_sk_callback_sse41+0xa7e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
@@ -23478,7 +23651,7 @@ _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  15,89,29,110,37,0,0                 // mulps         0x256e(%rip),%xmm3        # 5180 <_sk_callback_sse41+0xa90>
+  .byte  15,89,29,206,37,0,0                 // mulps         0x25ce(%rip),%xmm3        # 51e0 <_sk_callback_sse41+0xa8e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -23491,7 +23664,7 @@ FUNCTION(_sk_store_a8_sse41)
 _sk_store_a8_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,98,37,0,0                // movaps        0x2562(%rip),%xmm8        # 5190 <_sk_callback_sse41+0xaa0>
+  .byte  68,15,40,5,194,37,0,0               // movaps        0x25c2(%rip),%xmm8        # 51f0 <_sk_callback_sse41+0xa9e>
   .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
@@ -23508,9 +23681,9 @@ _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  15,89,5,63,37,0,0                   // mulps         0x253f(%rip),%xmm0        # 51a0 <_sk_callback_sse41+0xab0>
+  .byte  15,89,5,159,37,0,0                  // mulps         0x259f(%rip),%xmm0        # 5200 <_sk_callback_sse41+0xaae>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,70,37,0,0                  // movaps        0x2546(%rip),%xmm3        # 51b0 <_sk_callback_sse41+0xac0>
+  .byte  15,40,29,166,37,0,0                 // movaps        0x25a6(%rip),%xmm3        # 5210 <_sk_callback_sse41+0xabe>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -23541,9 +23714,9 @@ _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  15,89,5,223,36,0,0                  // mulps         0x24df(%rip),%xmm0        # 51c0 <_sk_callback_sse41+0xad0>
+  .byte  15,89,5,63,37,0,0                   // mulps         0x253f(%rip),%xmm0        # 5220 <_sk_callback_sse41+0xace>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,230,36,0,0                 // movaps        0x24e6(%rip),%xmm3        # 51d0 <_sk_callback_sse41+0xae0>
+  .byte  15,40,29,70,37,0,0                  // movaps        0x2546(%rip),%xmm3        # 5230 <_sk_callback_sse41+0xade>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -23588,17 +23761,17 @@ _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  102,15,111,5,61,36,0,0              // movdqa        0x243d(%rip),%xmm0        # 51e0 <_sk_callback_sse41+0xaf0>
+  .byte  102,15,111,5,157,36,0,0             // movdqa        0x249d(%rip),%xmm0        # 5240 <_sk_callback_sse41+0xaee>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,62,36,0,0                // movaps        0x243e(%rip),%xmm8        # 51f0 <_sk_callback_sse41+0xb00>
+  .byte  68,15,40,5,158,36,0,0               // movaps        0x249e(%rip),%xmm8        # 5250 <_sk_callback_sse41+0xafe>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
-  .byte  102,15,56,0,13,61,36,0,0            // pshufb        0x243d(%rip),%xmm1        # 5200 <_sk_callback_sse41+0xb10>
+  .byte  102,15,56,0,13,157,36,0,0           // pshufb        0x249d(%rip),%xmm1        # 5260 <_sk_callback_sse41+0xb0e>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,211                      // movdqa        %xmm3,%xmm2
-  .byte  102,15,56,0,21,57,36,0,0            // pshufb        0x2439(%rip),%xmm2        # 5210 <_sk_callback_sse41+0xb20>
+  .byte  102,15,56,0,21,153,36,0,0           // pshufb        0x2499(%rip),%xmm2        # 5270 <_sk_callback_sse41+0xb1e>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -23614,19 +23787,19 @@ _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  102,15,111,5,31,36,0,0              // movdqa        0x241f(%rip),%xmm0        # 5220 <_sk_callback_sse41+0xb30>
+  .byte  102,15,111,5,127,36,0,0             // movdqa        0x247f(%rip),%xmm0        # 5280 <_sk_callback_sse41+0xb2e>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,33,36,0,0                   // mulps         0x2421(%rip),%xmm0        # 5230 <_sk_callback_sse41+0xb40>
-  .byte  102,15,111,13,41,36,0,0             // movdqa        0x2429(%rip),%xmm1        # 5240 <_sk_callback_sse41+0xb50>
+  .byte  15,89,5,129,36,0,0                  // mulps         0x2481(%rip),%xmm0        # 5290 <_sk_callback_sse41+0xb3e>
+  .byte  102,15,111,13,137,36,0,0            // movdqa        0x2489(%rip),%xmm1        # 52a0 <_sk_callback_sse41+0xb4e>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,43,36,0,0                  // mulps         0x242b(%rip),%xmm1        # 5250 <_sk_callback_sse41+0xb60>
-  .byte  102,15,219,21,51,36,0,0             // pand          0x2433(%rip),%xmm2        # 5260 <_sk_callback_sse41+0xb70>
+  .byte  15,89,13,139,36,0,0                 // mulps         0x248b(%rip),%xmm1        # 52b0 <_sk_callback_sse41+0xb5e>
+  .byte  102,15,219,21,147,36,0,0            // pand          0x2493(%rip),%xmm2        # 52c0 <_sk_callback_sse41+0xb6e>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,57,36,0,0                  // mulps         0x2439(%rip),%xmm2        # 5270 <_sk_callback_sse41+0xb80>
+  .byte  15,89,21,153,36,0,0                 // mulps         0x2499(%rip),%xmm2        # 52d0 <_sk_callback_sse41+0xb7e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,64,36,0,0                  // movaps        0x2440(%rip),%xmm3        # 5280 <_sk_callback_sse41+0xb90>
+  .byte  15,40,29,160,36,0,0                 // movaps        0x24a0(%rip),%xmm3        # 52e0 <_sk_callback_sse41+0xb8e>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_gather_565_sse41
@@ -23654,19 +23827,19 @@ _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  102,15,111,5,229,35,0,0             // movdqa        0x23e5(%rip),%xmm0        # 5290 <_sk_callback_sse41+0xba0>
+  .byte  102,15,111,5,69,36,0,0              // movdqa        0x2445(%rip),%xmm0        # 52f0 <_sk_callback_sse41+0xb9e>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,231,35,0,0                  // mulps         0x23e7(%rip),%xmm0        # 52a0 <_sk_callback_sse41+0xbb0>
-  .byte  102,15,111,13,239,35,0,0            // movdqa        0x23ef(%rip),%xmm1        # 52b0 <_sk_callback_sse41+0xbc0>
+  .byte  15,89,5,71,36,0,0                   // mulps         0x2447(%rip),%xmm0        # 5300 <_sk_callback_sse41+0xbae>
+  .byte  102,15,111,13,79,36,0,0             // movdqa        0x244f(%rip),%xmm1        # 5310 <_sk_callback_sse41+0xbbe>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,241,35,0,0                 // mulps         0x23f1(%rip),%xmm1        # 52c0 <_sk_callback_sse41+0xbd0>
-  .byte  102,15,219,21,249,35,0,0            // pand          0x23f9(%rip),%xmm2        # 52d0 <_sk_callback_sse41+0xbe0>
+  .byte  15,89,13,81,36,0,0                  // mulps         0x2451(%rip),%xmm1        # 5320 <_sk_callback_sse41+0xbce>
+  .byte  102,15,219,21,89,36,0,0             // pand          0x2459(%rip),%xmm2        # 5330 <_sk_callback_sse41+0xbde>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,255,35,0,0                 // mulps         0x23ff(%rip),%xmm2        # 52e0 <_sk_callback_sse41+0xbf0>
+  .byte  15,89,21,95,36,0,0                  // mulps         0x245f(%rip),%xmm2        # 5340 <_sk_callback_sse41+0xbee>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,6,36,0,0                   // movaps        0x2406(%rip),%xmm3        # 52f0 <_sk_callback_sse41+0xc00>
+  .byte  15,40,29,102,36,0,0                 // movaps        0x2466(%rip),%xmm3        # 5350 <_sk_callback_sse41+0xbfe>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_565_sse41
@@ -23675,12 +23848,12 @@ FUNCTION(_sk_store_565_sse41)
 _sk_store_565_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,7,36,0,0                 // movaps        0x2407(%rip),%xmm8        # 5300 <_sk_callback_sse41+0xc10>
+  .byte  68,15,40,5,103,36,0,0               // movaps        0x2467(%rip),%xmm8        # 5360 <_sk_callback_sse41+0xc0e>
   .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  68,15,40,21,252,35,0,0              // movaps        0x23fc(%rip),%xmm10        # 5310 <_sk_callback_sse41+0xc20>
+  .byte  68,15,40,21,92,36,0,0               // movaps        0x245c(%rip),%xmm10        # 5370 <_sk_callback_sse41+0xc1e>
   .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
@@ -23700,21 +23873,21 @@ _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  102,15,111,5,199,35,0,0             // movdqa        0x23c7(%rip),%xmm0        # 5320 <_sk_callback_sse41+0xc30>
+  .byte  102,15,111,5,39,36,0,0              // movdqa        0x2427(%rip),%xmm0        # 5380 <_sk_callback_sse41+0xc2e>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,201,35,0,0                  // mulps         0x23c9(%rip),%xmm0        # 5330 <_sk_callback_sse41+0xc40>
-  .byte  102,15,111,13,209,35,0,0            // movdqa        0x23d1(%rip),%xmm1        # 5340 <_sk_callback_sse41+0xc50>
+  .byte  15,89,5,41,36,0,0                   // mulps         0x2429(%rip),%xmm0        # 5390 <_sk_callback_sse41+0xc3e>
+  .byte  102,15,111,13,49,36,0,0             // movdqa        0x2431(%rip),%xmm1        # 53a0 <_sk_callback_sse41+0xc4e>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,211,35,0,0                 // mulps         0x23d3(%rip),%xmm1        # 5350 <_sk_callback_sse41+0xc60>
-  .byte  102,15,111,21,219,35,0,0            // movdqa        0x23db(%rip),%xmm2        # 5360 <_sk_callback_sse41+0xc70>
+  .byte  15,89,13,51,36,0,0                  // mulps         0x2433(%rip),%xmm1        # 53b0 <_sk_callback_sse41+0xc5e>
+  .byte  102,15,111,21,59,36,0,0             // movdqa        0x243b(%rip),%xmm2        # 53c0 <_sk_callback_sse41+0xc6e>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,221,35,0,0                 // mulps         0x23dd(%rip),%xmm2        # 5370 <_sk_callback_sse41+0xc80>
-  .byte  102,15,219,29,229,35,0,0            // pand          0x23e5(%rip),%xmm3        # 5380 <_sk_callback_sse41+0xc90>
+  .byte  15,89,21,61,36,0,0                  // mulps         0x243d(%rip),%xmm2        # 53d0 <_sk_callback_sse41+0xc7e>
+  .byte  102,15,219,29,69,36,0,0             // pand          0x2445(%rip),%xmm3        # 53e0 <_sk_callback_sse41+0xc8e>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,235,35,0,0                 // mulps         0x23eb(%rip),%xmm3        # 5390 <_sk_callback_sse41+0xca0>
+  .byte  15,89,29,75,36,0,0                  // mulps         0x244b(%rip),%xmm3        # 53f0 <_sk_callback_sse41+0xc9e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -23743,21 +23916,21 @@ _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  102,15,111,5,142,35,0,0             // movdqa        0x238e(%rip),%xmm0        # 53a0 <_sk_callback_sse41+0xcb0>
+  .byte  102,15,111,5,238,35,0,0             // movdqa        0x23ee(%rip),%xmm0        # 5400 <_sk_callback_sse41+0xcae>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,144,35,0,0                  // mulps         0x2390(%rip),%xmm0        # 53b0 <_sk_callback_sse41+0xcc0>
-  .byte  102,15,111,13,152,35,0,0            // movdqa        0x2398(%rip),%xmm1        # 53c0 <_sk_callback_sse41+0xcd0>
+  .byte  15,89,5,240,35,0,0                  // mulps         0x23f0(%rip),%xmm0        # 5410 <_sk_callback_sse41+0xcbe>
+  .byte  102,15,111,13,248,35,0,0            // movdqa        0x23f8(%rip),%xmm1        # 5420 <_sk_callback_sse41+0xcce>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,154,35,0,0                 // mulps         0x239a(%rip),%xmm1        # 53d0 <_sk_callback_sse41+0xce0>
-  .byte  102,15,111,21,162,35,0,0            // movdqa        0x23a2(%rip),%xmm2        # 53e0 <_sk_callback_sse41+0xcf0>
+  .byte  15,89,13,250,35,0,0                 // mulps         0x23fa(%rip),%xmm1        # 5430 <_sk_callback_sse41+0xcde>
+  .byte  102,15,111,21,2,36,0,0              // movdqa        0x2402(%rip),%xmm2        # 5440 <_sk_callback_sse41+0xcee>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,164,35,0,0                 // mulps         0x23a4(%rip),%xmm2        # 53f0 <_sk_callback_sse41+0xd00>
-  .byte  102,15,219,29,172,35,0,0            // pand          0x23ac(%rip),%xmm3        # 5400 <_sk_callback_sse41+0xd10>
+  .byte  15,89,21,4,36,0,0                   // mulps         0x2404(%rip),%xmm2        # 5450 <_sk_callback_sse41+0xcfe>
+  .byte  102,15,219,29,12,36,0,0             // pand          0x240c(%rip),%xmm3        # 5460 <_sk_callback_sse41+0xd0e>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,178,35,0,0                 // mulps         0x23b2(%rip),%xmm3        # 5410 <_sk_callback_sse41+0xd20>
+  .byte  15,89,29,18,36,0,0                  // mulps         0x2412(%rip),%xmm3        # 5470 <_sk_callback_sse41+0xd1e>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -23767,7 +23940,7 @@ FUNCTION(_sk_store_4444_sse41)
 _sk_store_4444_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,177,35,0,0               // movaps        0x23b1(%rip),%xmm8        # 5420 <_sk_callback_sse41+0xd30>
+  .byte  68,15,40,5,17,36,0,0                // movaps        0x2411(%rip),%xmm8        # 5480 <_sk_callback_sse41+0xd2e>
   .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
@@ -23797,17 +23970,17 @@ _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  15,40,5,80,35,0,0                   // movaps        0x2350(%rip),%xmm0        # 5430 <_sk_callback_sse41+0xd40>
+  .byte  15,40,5,176,35,0,0                  // movaps        0x23b0(%rip),%xmm0        # 5490 <_sk_callback_sse41+0xd3e>
   .byte  15,84,195                           // andps         %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,82,35,0,0                // movaps        0x2352(%rip),%xmm8        # 5440 <_sk_callback_sse41+0xd50>
+  .byte  68,15,40,5,178,35,0,0               // movaps        0x23b2(%rip),%xmm8        # 54a0 <_sk_callback_sse41+0xd4e>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,40,203                           // movaps        %xmm3,%xmm1
-  .byte  102,15,56,0,13,82,35,0,0            // pshufb        0x2352(%rip),%xmm1        # 5450 <_sk_callback_sse41+0xd60>
+  .byte  102,15,56,0,13,178,35,0,0           // pshufb        0x23b2(%rip),%xmm1        # 54b0 <_sk_callback_sse41+0xd5e>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  15,40,211                           // movaps        %xmm3,%xmm2
-  .byte  102,15,56,0,21,79,35,0,0            // pshufb        0x234f(%rip),%xmm2        # 5460 <_sk_callback_sse41+0xd70>
+  .byte  102,15,56,0,21,175,35,0,0           // pshufb        0x23af(%rip),%xmm2        # 54c0 <_sk_callback_sse41+0xd6e>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -23838,17 +24011,17 @@ _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  102,15,111,5,232,34,0,0             // movdqa        0x22e8(%rip),%xmm0        # 5470 <_sk_callback_sse41+0xd80>
+  .byte  102,15,111,5,72,35,0,0              // movdqa        0x2348(%rip),%xmm0        # 54d0 <_sk_callback_sse41+0xd7e>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,233,34,0,0               // movaps        0x22e9(%rip),%xmm8        # 5480 <_sk_callback_sse41+0xd90>
+  .byte  68,15,40,5,73,35,0,0                // movaps        0x2349(%rip),%xmm8        # 54e0 <_sk_callback_sse41+0xd8e>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
-  .byte  102,15,56,0,13,232,34,0,0           // pshufb        0x22e8(%rip),%xmm1        # 5490 <_sk_callback_sse41+0xda0>
+  .byte  102,15,56,0,13,72,35,0,0            // pshufb        0x2348(%rip),%xmm1        # 54f0 <_sk_callback_sse41+0xd9e>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,211                      // movdqa        %xmm3,%xmm2
-  .byte  102,15,56,0,21,228,34,0,0           // pshufb        0x22e4(%rip),%xmm2        # 54a0 <_sk_callback_sse41+0xdb0>
+  .byte  102,15,56,0,21,68,35,0,0            // pshufb        0x2344(%rip),%xmm2        # 5500 <_sk_callback_sse41+0xdae>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -23863,7 +24036,7 @@ FUNCTION(_sk_store_8888_sse41)
 _sk_store_8888_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,208,34,0,0               // movaps        0x22d0(%rip),%xmm8        # 54b0 <_sk_callback_sse41+0xdc0>
+  .byte  68,15,40,5,48,35,0,0                // movaps        0x2330(%rip),%xmm8        # 5510 <_sk_callback_sse41+0xdbe>
   .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
@@ -23900,18 +24073,18 @@ _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  102,68,15,111,5,73,34,0,0           // movdqa        0x2249(%rip),%xmm8        # 54c0 <_sk_callback_sse41+0xdd0>
+  .byte  102,68,15,111,5,169,34,0,0          // movdqa        0x22a9(%rip),%xmm8        # 5520 <_sk_callback_sse41+0xdce>
   .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  102,15,111,29,68,34,0,0             // movdqa        0x2244(%rip),%xmm3        # 54d0 <_sk_callback_sse41+0xde0>
+  .byte  102,15,111,29,164,34,0,0            // movdqa        0x22a4(%rip),%xmm3        # 5530 <_sk_callback_sse41+0xdde>
   .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  102,68,15,111,21,48,34,0,0          // movdqa        0x2230(%rip),%xmm10        # 54e0 <_sk_callback_sse41+0xdf0>
+  .byte  102,68,15,111,21,144,34,0,0         // movdqa        0x2290(%rip),%xmm10        # 5540 <_sk_callback_sse41+0xdee>
   .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
@@ -23984,18 +24157,18 @@ _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  102,68,15,111,5,238,32,0,0          // movdqa        0x20ee(%rip),%xmm8        # 54f0 <_sk_callback_sse41+0xe00>
+  .byte  102,68,15,111,5,78,33,0,0           // movdqa        0x214e(%rip),%xmm8        # 5550 <_sk_callback_sse41+0xdfe>
   .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  102,15,111,29,233,32,0,0            // movdqa        0x20e9(%rip),%xmm3        # 5500 <_sk_callback_sse41+0xe10>
+  .byte  102,15,111,29,73,33,0,0             // movdqa        0x2149(%rip),%xmm3        # 5560 <_sk_callback_sse41+0xe0e>
   .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  102,68,15,111,21,213,32,0,0         // movdqa        0x20d5(%rip),%xmm10        # 5510 <_sk_callback_sse41+0xe20>
+  .byte  102,68,15,111,21,53,33,0,0          // movdqa        0x2135(%rip),%xmm10        # 5570 <_sk_callback_sse41+0xe1e>
   .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
@@ -24043,17 +24216,17 @@ FUNCTION(_sk_store_f16_sse41)
 _sk_store_f16_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  102,68,15,111,21,11,32,0,0          // movdqa        0x200b(%rip),%xmm10        # 5520 <_sk_callback_sse41+0xe30>
+  .byte  102,68,15,111,21,107,32,0,0         // movdqa        0x206b(%rip),%xmm10        # 5580 <_sk_callback_sse41+0xe2e>
   .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  102,68,15,111,13,254,31,0,0         // movdqa        0x1ffe(%rip),%xmm9        # 5530 <_sk_callback_sse41+0xe40>
+  .byte  102,68,15,111,13,94,32,0,0          // movdqa        0x205e(%rip),%xmm9        # 5590 <_sk_callback_sse41+0xe3e>
   .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  102,68,15,111,29,239,31,0,0         // movdqa        0x1fef(%rip),%xmm11        # 5540 <_sk_callback_sse41+0xe50>
+  .byte  102,68,15,111,29,79,32,0,0          // movdqa        0x204f(%rip),%xmm11        # 55a0 <_sk_callback_sse41+0xe4e>
   .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
@@ -24123,7 +24296,7 @@ _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  68,15,40,5,190,30,0,0               // movaps        0x1ebe(%rip),%xmm8        # 5550 <_sk_callback_sse41+0xe60>
+  .byte  68,15,40,5,30,31,0,0                // movaps        0x1f1e(%rip),%xmm8        # 55b0 <_sk_callback_sse41+0xe5e>
   .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
@@ -24175,7 +24348,7 @@ _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  68,15,40,5,255,29,0,0               // movaps        0x1dff(%rip),%xmm8        # 5560 <_sk_callback_sse41+0xe70>
+  .byte  68,15,40,5,95,30,0,0                // movaps        0x1e5f(%rip),%xmm8        # 55c0 <_sk_callback_sse41+0xe6e>
   .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
@@ -24192,7 +24365,7 @@ _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  15,40,29,198,29,0,0                 // movaps        0x1dc6(%rip),%xmm3        # 5570 <_sk_callback_sse41+0xe80>
+  .byte  15,40,29,38,30,0,0                  // movaps        0x1e26(%rip),%xmm3        # 55d0 <_sk_callback_sse41+0xe7e>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_u16_be_sse41
@@ -24201,7 +24374,7 @@ 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  68,15,40,13,199,29,0,0              // movaps        0x1dc7(%rip),%xmm9        # 5580 <_sk_callback_sse41+0xe90>
+  .byte  68,15,40,13,39,30,0,0               // movaps        0x1e27(%rip),%xmm9        # 55e0 <_sk_callback_sse41+0xe8e>
   .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
@@ -24407,15 +24580,53 @@ _sk_mirror_y_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_clamp_x_1_sse41
+.globl _sk_clamp_x_1_sse41
+FUNCTION(_sk_clamp_x_1_sse41)
+_sk_clamp_x_1_sse41:
+  .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
+  .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
+  .byte  68,15,93,5,125,27,0,0               // minps         0x1b7d(%rip),%xmm8        # 55f0 <_sk_callback_sse41+0xe9e>
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_repeat_x_1_sse41
+.globl _sk_repeat_x_1_sse41
+FUNCTION(_sk_repeat_x_1_sse41)
+_sk_repeat_x_1_sse41:
+  .byte  102,68,15,58,8,192,1                // roundps       $0x1,%xmm0,%xmm8
+  .byte  65,15,92,192                        // subps         %xmm8,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_mirror_x_1_sse41
+.globl _sk_mirror_x_1_sse41
+FUNCTION(_sk_mirror_x_1_sse41)
+_sk_mirror_x_1_sse41:
+  .byte  68,15,40,5,110,27,0,0               // movaps        0x1b6e(%rip),%xmm8        # 5600 <_sk_callback_sse41+0xeae>
+  .byte  65,15,88,192                        // addps         %xmm8,%xmm0
+  .byte  68,15,40,13,114,27,0,0              // movaps        0x1b72(%rip),%xmm9        # 5610 <_sk_callback_sse41+0xebe>
+  .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
+  .byte  102,69,15,58,8,201,1                // roundps       $0x1,%xmm9,%xmm9
+  .byte  69,15,88,201                        // addps         %xmm9,%xmm9
+  .byte  65,15,92,193                        // subps         %xmm9,%xmm0
+  .byte  65,15,88,192                        // addps         %xmm8,%xmm0
+  .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
+  .byte  68,15,92,192                        // subps         %xmm0,%xmm8
+  .byte  65,15,84,192                        // andps         %xmm8,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_luminance_to_alpha_sse41
 .globl _sk_luminance_to_alpha_sse41
 FUNCTION(_sk_luminance_to_alpha_sse41)
 _sk_luminance_to_alpha_sse41:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
-  .byte  15,89,5,35,27,0,0                   // mulps         0x1b23(%rip),%xmm0        # 5590 <_sk_callback_sse41+0xea0>
-  .byte  15,89,13,44,27,0,0                  // mulps         0x1b2c(%rip),%xmm1        # 55a0 <_sk_callback_sse41+0xeb0>
+  .byte  15,89,5,81,27,0,0                   // mulps         0x1b51(%rip),%xmm0        # 5620 <_sk_callback_sse41+0xece>
+  .byte  15,89,13,90,27,0,0                  // mulps         0x1b5a(%rip),%xmm1        # 5630 <_sk_callback_sse41+0xede>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,29,50,27,0,0                  // mulps         0x1b32(%rip),%xmm3        # 55b0 <_sk_callback_sse41+0xec0>
+  .byte  15,89,29,96,27,0,0                  // mulps         0x1b60(%rip),%xmm3        # 5640 <_sk_callback_sse41+0xeee>
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -24691,9 +24902,9 @@ _sk_evenly_spaced_gradient_sse41:
   .byte  72,139,8                            // mov           (%rax),%rcx
   .byte  76,139,88,8                         // mov           0x8(%rax),%r11
   .byte  72,255,201                          // dec           %rcx
-  .byte  120,7                               // js            3eb7 <_sk_evenly_spaced_gradient_sse41+0x15>
+  .byte  120,7                               // js            3f19 <_sk_evenly_spaced_gradient_sse41+0x15>
   .byte  243,72,15,42,201                    // cvtsi2ss      %rcx,%xmm1
-  .byte  235,21                              // jmp           3ecc <_sk_evenly_spaced_gradient_sse41+0x2a>
+  .byte  235,21                              // jmp           3f2e <_sk_evenly_spaced_gradient_sse41+0x2a>
   .byte  73,137,200                          // mov           %rcx,%r8
   .byte  73,209,232                          // shr           %r8
   .byte  131,225,1                           // and           $0x1,%ecx
@@ -24784,12 +24995,12 @@ _sk_gradient_sse41:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  73,131,248,2                        // cmp           $0x2,%r8
-  .byte  114,50                              // jb            40af <_sk_gradient_sse41+0x41>
+  .byte  114,50                              // jb            4111 <_sk_gradient_sse41+0x41>
   .byte  72,139,72,72                        // mov           0x48(%rax),%rcx
   .byte  73,255,200                          // dec           %r8
   .byte  72,131,193,4                        // add           $0x4,%rcx
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
-  .byte  15,40,21,45,21,0,0                  // movaps        0x152d(%rip),%xmm2        # 55c0 <_sk_callback_sse41+0xed0>
+  .byte  15,40,21,91,21,0,0                  // movaps        0x155b(%rip),%xmm2        # 5650 <_sk_callback_sse41+0xefe>
   .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
@@ -24797,7 +25008,7 @@ _sk_gradient_sse41:
   .byte  102,15,254,203                      // paddd         %xmm3,%xmm1
   .byte  72,131,193,4                        // add           $0x4,%rcx
   .byte  73,255,200                          // dec           %r8
-  .byte  117,228                             // jne           4093 <_sk_gradient_sse41+0x25>
+  .byte  117,228                             // jne           40f5 <_sk_gradient_sse41+0x25>
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  102,73,15,58,22,201,1               // pextrq        $0x1,%xmm1,%r9
@@ -24928,26 +25139,26 @@ _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  68,15,40,21,207,18,0,0              // movaps        0x12cf(%rip),%xmm10        # 55d0 <_sk_callback_sse41+0xee0>
+  .byte  68,15,40,21,253,18,0,0              // movaps        0x12fd(%rip),%xmm10        # 5660 <_sk_callback_sse41+0xf0e>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,211,18,0,0              // addps         0x12d3(%rip),%xmm10        # 55e0 <_sk_callback_sse41+0xef0>
+  .byte  68,15,88,21,1,19,0,0                // addps         0x1301(%rip),%xmm10        # 5670 <_sk_callback_sse41+0xf1e>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,215,18,0,0              // addps         0x12d7(%rip),%xmm10        # 55f0 <_sk_callback_sse41+0xf00>
+  .byte  68,15,88,21,5,19,0,0                // addps         0x1305(%rip),%xmm10        # 5680 <_sk_callback_sse41+0xf2e>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,219,18,0,0              // addps         0x12db(%rip),%xmm10        # 5600 <_sk_callback_sse41+0xf10>
+  .byte  68,15,88,21,9,19,0,0                // addps         0x1309(%rip),%xmm10        # 5690 <_sk_callback_sse41+0xf3e>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,194,195,1                     // cmpltps       %xmm11,%xmm0
-  .byte  68,15,40,29,218,18,0,0              // movaps        0x12da(%rip),%xmm11        # 5610 <_sk_callback_sse41+0xf20>
+  .byte  68,15,40,29,8,19,0,0                // movaps        0x1308(%rip),%xmm11        # 56a0 <_sk_callback_sse41+0xf4e>
   .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  68,15,40,29,211,18,0,0              // movaps        0x12d3(%rip),%xmm11        # 5620 <_sk_callback_sse41+0xf30>
+  .byte  68,15,40,29,1,19,0,0                // movaps        0x1301(%rip),%xmm11        # 56b0 <_sk_callback_sse41+0xf5e>
   .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  68,15,40,13,197,18,0,0              // movaps        0x12c5(%rip),%xmm9        # 5630 <_sk_callback_sse41+0xf40>
+  .byte  68,15,40,13,243,18,0,0              // movaps        0x12f3(%rip),%xmm9        # 56c0 <_sk_callback_sse41+0xf6e>
   .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
@@ -24973,7 +25184,7 @@ 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        # 5640 <_sk_callback_sse41+0xf50>
+  .byte  68,15,40,5,199,18,0,0               // movaps        0x12c7(%rip),%xmm8        # 56d0 <_sk_callback_sse41+0xf7e>
   .byte  15,17,0                             // movups        %xmm0,(%rax)
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,88,200                        // addps         %xmm8,%xmm9
@@ -25017,8 +25228,8 @@ _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        # 5650 <_sk_callback_sse41+0xf60>
-  .byte  68,15,40,13,35,18,0,0               // movaps        0x1223(%rip),%xmm9        # 5660 <_sk_callback_sse41+0xf70>
+  .byte  15,88,5,73,18,0,0                   // addps         0x1249(%rip),%xmm0        # 56e0 <_sk_callback_sse41+0xf8e>
+  .byte  68,15,40,13,81,18,0,0               // movaps        0x1251(%rip),%xmm9        # 56f0 <_sk_callback_sse41+0xf9e>
   .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
@@ -25031,7 +25242,7 @@ _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        # 5670 <_sk_callback_sse41+0xf80>
+  .byte  15,88,5,64,18,0,0                   // addps         0x1240(%rip),%xmm0        # 5700 <_sk_callback_sse41+0xfae>
   .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
@@ -25043,8 +25254,8 @@ _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        # 5680 <_sk_callback_sse41+0xf90>
-  .byte  68,15,40,13,12,18,0,0               // movaps        0x120c(%rip),%xmm9        # 5690 <_sk_callback_sse41+0xfa0>
+  .byte  15,88,13,50,18,0,0                  // addps         0x1232(%rip),%xmm1        # 5710 <_sk_callback_sse41+0xfbe>
+  .byte  68,15,40,13,58,18,0,0               // movaps        0x123a(%rip),%xmm9        # 5720 <_sk_callback_sse41+0xfce>
   .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
@@ -25057,7 +25268,7 @@ _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        # 56a0 <_sk_callback_sse41+0xfb0>
+  .byte  15,88,13,40,18,0,0                  // addps         0x1228(%rip),%xmm1        # 5730 <_sk_callback_sse41+0xfde>
   .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
@@ -25069,13 +25280,13 @@ _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        # 56b0 <_sk_callback_sse41+0xfc0>
-  .byte  68,15,40,13,245,17,0,0              // movaps        0x11f5(%rip),%xmm9        # 56c0 <_sk_callback_sse41+0xfd0>
+  .byte  15,88,5,27,18,0,0                   // addps         0x121b(%rip),%xmm0        # 5740 <_sk_callback_sse41+0xfee>
+  .byte  68,15,40,13,35,18,0,0               // movaps        0x1223(%rip),%xmm9        # 5750 <_sk_callback_sse41+0xffe>
   .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        # 56d0 <_sk_callback_sse41+0xfe0>
-  .byte  68,15,88,13,249,17,0,0              // addps         0x11f9(%rip),%xmm9        # 56e0 <_sk_callback_sse41+0xff0>
+  .byte  68,15,89,13,31,18,0,0               // mulps         0x121f(%rip),%xmm9        # 5760 <_sk_callback_sse41+0x100e>
+  .byte  68,15,88,13,39,18,0,0               // addps         0x1227(%rip),%xmm9        # 5770 <_sk_callback_sse41+0x101e>
   .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
@@ -25088,16 +25299,16 @@ _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        # 56f0 <_sk_callback_sse41+0x1000>
-  .byte  68,15,40,13,240,17,0,0              // movaps        0x11f0(%rip),%xmm9        # 5700 <_sk_callback_sse41+0x1010>
+  .byte  15,88,5,22,18,0,0                   // addps         0x1216(%rip),%xmm0        # 5780 <_sk_callback_sse41+0x102e>
+  .byte  68,15,40,13,30,18,0,0               // movaps        0x121e(%rip),%xmm9        # 5790 <_sk_callback_sse41+0x103e>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,244,17,0,0               // movaps        0x11f4(%rip),%xmm8        # 5710 <_sk_callback_sse41+0x1020>
+  .byte  68,15,40,5,34,18,0,0                // movaps        0x1222(%rip),%xmm8        # 57a0 <_sk_callback_sse41+0x104e>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,248,17,0,0               // addps         0x11f8(%rip),%xmm8        # 5720 <_sk_callback_sse41+0x1030>
+  .byte  68,15,88,5,38,18,0,0                // addps         0x1226(%rip),%xmm8        # 57b0 <_sk_callback_sse41+0x105e>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,252,17,0,0               // addps         0x11fc(%rip),%xmm8        # 5730 <_sk_callback_sse41+0x1040>
+  .byte  68,15,88,5,42,18,0,0                // addps         0x122a(%rip),%xmm8        # 57c0 <_sk_callback_sse41+0x106e>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,0,18,0,0                 // addps         0x1200(%rip),%xmm8        # 5740 <_sk_callback_sse41+0x1050>
+  .byte  68,15,88,5,46,18,0,0                // addps         0x122e(%rip),%xmm8        # 57d0 <_sk_callback_sse41+0x107e>
   .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
@@ -25107,17 +25318,17 @@ 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        # 5750 <_sk_callback_sse41+0x1060>
+  .byte  68,15,40,5,40,18,0,0                // movaps        0x1228(%rip),%xmm8        # 57e0 <_sk_callback_sse41+0x108e>
   .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        # 5760 <_sk_callback_sse41+0x1070>
+  .byte  68,15,40,21,36,18,0,0               // movaps        0x1224(%rip),%xmm10        # 57f0 <_sk_callback_sse41+0x109e>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,250,17,0,0              // addps         0x11fa(%rip),%xmm10        # 5770 <_sk_callback_sse41+0x1080>
+  .byte  68,15,88,21,40,18,0,0               // addps         0x1228(%rip),%xmm10        # 5800 <_sk_callback_sse41+0x10ae>
   .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        # 5780 <_sk_callback_sse41+0x1090>
+  .byte  68,15,88,21,36,18,0,0               // addps         0x1224(%rip),%xmm10        # 5810 <_sk_callback_sse41+0x10be>
   .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
@@ -25129,11 +25340,11 @@ _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        # 5790 <_sk_callback_sse41+0x10a0>
+  .byte  15,88,5,23,18,0,0                   // addps         0x1217(%rip),%xmm0        # 5820 <_sk_callback_sse41+0x10ce>
   .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        # 57a0 <_sk_callback_sse41+0x10b0>
-  .byte  68,15,88,5,241,17,0,0               // addps         0x11f1(%rip),%xmm8        # 57b0 <_sk_callback_sse41+0x10c0>
+  .byte  68,15,89,5,23,18,0,0                // mulps         0x1217(%rip),%xmm8        # 5830 <_sk_callback_sse41+0x10de>
+  .byte  68,15,88,5,31,18,0,0                // addps         0x121f(%rip),%xmm8        # 5840 <_sk_callback_sse41+0x10ee>
   .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
@@ -25146,13 +25357,13 @@ _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        # 57c0 <_sk_callback_sse41+0x10d0>
-  .byte  68,15,40,13,231,17,0,0              // movaps        0x11e7(%rip),%xmm9        # 57d0 <_sk_callback_sse41+0x10e0>
+  .byte  15,88,13,13,18,0,0                  // addps         0x120d(%rip),%xmm1        # 5850 <_sk_callback_sse41+0x10fe>
+  .byte  68,15,40,13,21,18,0,0               // movaps        0x1215(%rip),%xmm9        # 5860 <_sk_callback_sse41+0x110e>
   .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        # 57e0 <_sk_callback_sse41+0x10f0>
-  .byte  68,15,88,13,235,17,0,0              // addps         0x11eb(%rip),%xmm9        # 57f0 <_sk_callback_sse41+0x1100>
+  .byte  68,15,89,13,17,18,0,0               // mulps         0x1211(%rip),%xmm9        # 5870 <_sk_callback_sse41+0x111e>
+  .byte  68,15,88,13,25,18,0,0               // addps         0x1219(%rip),%xmm9        # 5880 <_sk_callback_sse41+0x112e>
   .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
@@ -25165,16 +25376,16 @@ _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        # 5800 <_sk_callback_sse41+0x1110>
-  .byte  68,15,40,13,225,17,0,0              // movaps        0x11e1(%rip),%xmm9        # 5810 <_sk_callback_sse41+0x1120>
+  .byte  15,88,13,7,18,0,0                   // addps         0x1207(%rip),%xmm1        # 5890 <_sk_callback_sse41+0x113e>
+  .byte  68,15,40,13,15,18,0,0               // movaps        0x120f(%rip),%xmm9        # 58a0 <_sk_callback_sse41+0x114e>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,229,17,0,0               // movaps        0x11e5(%rip),%xmm8        # 5820 <_sk_callback_sse41+0x1130>
+  .byte  68,15,40,5,19,18,0,0                // movaps        0x1213(%rip),%xmm8        # 58b0 <_sk_callback_sse41+0x115e>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,233,17,0,0               // addps         0x11e9(%rip),%xmm8        # 5830 <_sk_callback_sse41+0x1140>
+  .byte  68,15,88,5,23,18,0,0                // addps         0x1217(%rip),%xmm8        # 58c0 <_sk_callback_sse41+0x116e>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,237,17,0,0               // addps         0x11ed(%rip),%xmm8        # 5840 <_sk_callback_sse41+0x1150>
+  .byte  68,15,88,5,27,18,0,0                // addps         0x121b(%rip),%xmm8        # 58d0 <_sk_callback_sse41+0x117e>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,241,17,0,0               // addps         0x11f1(%rip),%xmm8        # 5850 <_sk_callback_sse41+0x1160>
+  .byte  68,15,88,5,31,18,0,0                // addps         0x121f(%rip),%xmm8        # 58e0 <_sk_callback_sse41+0x118e>
   .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
@@ -25184,17 +25395,17 @@ 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        # 5860 <_sk_callback_sse41+0x1170>
+  .byte  68,15,40,5,25,18,0,0                // movaps        0x1219(%rip),%xmm8        # 58f0 <_sk_callback_sse41+0x119e>
   .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        # 5870 <_sk_callback_sse41+0x1180>
+  .byte  68,15,40,21,20,18,0,0               // movaps        0x1214(%rip),%xmm10        # 5900 <_sk_callback_sse41+0x11ae>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,234,17,0,0              // addps         0x11ea(%rip),%xmm10        # 5880 <_sk_callback_sse41+0x1190>
+  .byte  68,15,88,21,24,18,0,0               // addps         0x1218(%rip),%xmm10        # 5910 <_sk_callback_sse41+0x11be>
   .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        # 5890 <_sk_callback_sse41+0x11a0>
+  .byte  68,15,88,21,20,18,0,0               // addps         0x1214(%rip),%xmm10        # 5920 <_sk_callback_sse41+0x11ce>
   .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
@@ -25206,11 +25417,11 @@ _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        # 58a0 <_sk_callback_sse41+0x11b0>
+  .byte  15,88,13,6,18,0,0                   // addps         0x1206(%rip),%xmm1        # 5930 <_sk_callback_sse41+0x11de>
   .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        # 58b0 <_sk_callback_sse41+0x11c0>
-  .byte  68,15,88,5,224,17,0,0               // addps         0x11e0(%rip),%xmm8        # 58c0 <_sk_callback_sse41+0x11d0>
+  .byte  68,15,89,5,6,18,0,0                 // mulps         0x1206(%rip),%xmm8        # 5940 <_sk_callback_sse41+0x11ee>
+  .byte  68,15,88,5,14,18,0,0                // addps         0x120e(%rip),%xmm8        # 5950 <_sk_callback_sse41+0x11fe>
   .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
@@ -25429,11 +25640,11 @@ BALIGN16
   .byte  128,191,0,0,128,191,0               // cmpb          $0x0,-0x40800000(%rdi)
   .byte  0,224                               // add           %ah,%al
   .byte  64,0,0                              // add           %al,(%rax)
-  .byte  224,64                              // loopne        49a8 <.literal16+0x1d8>
+  .byte  224,64                              // loopne        4a08 <.literal16+0x1d8>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        49ac <.literal16+0x1dc>
+  .byte  224,64                              // loopne        4a0c <.literal16+0x1dc>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        49b0 <.literal16+0x1e0>
+  .byte  224,64                              // loopne        4a10 <.literal16+0x1e0>
   .byte  154                                 // (bad)
   .byte  153                                 // cltd
   .byte  153                                 // cltd
@@ -25453,13 +25664,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 49d1 <.literal16+0x201>
+  .byte  71,225,61                           // rex.RXB       loope 4a31 <.literal16+0x201>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 49d5 <.literal16+0x205>
+  .byte  71,225,61                           // rex.RXB       loope 4a35 <.literal16+0x205>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 49d9 <.literal16+0x209>
+  .byte  71,225,61                           // rex.RXB       loope 4a39 <.literal16+0x209>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 49dd <.literal16+0x20d>
+  .byte  71,225,61                           // rex.RXB       loope 4a3d <.literal16+0x20d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -25484,13 +25695,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a11 <.literal16+0x241>
+  .byte  71,225,61                           // rex.RXB       loope 4a71 <.literal16+0x241>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a15 <.literal16+0x245>
+  .byte  71,225,61                           // rex.RXB       loope 4a75 <.literal16+0x245>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a19 <.literal16+0x249>
+  .byte  71,225,61                           // rex.RXB       loope 4a79 <.literal16+0x249>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a1d <.literal16+0x24d>
+  .byte  71,225,61                           // rex.RXB       loope 4a7d <.literal16+0x24d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -25515,13 +25726,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a51 <.literal16+0x281>
+  .byte  71,225,61                           // rex.RXB       loope 4ab1 <.literal16+0x281>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a55 <.literal16+0x285>
+  .byte  71,225,61                           // rex.RXB       loope 4ab5 <.literal16+0x285>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a59 <.literal16+0x289>
+  .byte  71,225,61                           // rex.RXB       loope 4ab9 <.literal16+0x289>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a5d <.literal16+0x28d>
+  .byte  71,225,61                           // rex.RXB       loope 4abd <.literal16+0x28d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -25546,13 +25757,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a91 <.literal16+0x2c1>
+  .byte  71,225,61                           // rex.RXB       loope 4af1 <.literal16+0x2c1>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a95 <.literal16+0x2c5>
+  .byte  71,225,61                           // rex.RXB       loope 4af5 <.literal16+0x2c5>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a99 <.literal16+0x2c9>
+  .byte  71,225,61                           // rex.RXB       loope 4af9 <.literal16+0x2c9>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4a9d <.literal16+0x2cd>
+  .byte  71,225,61                           // rex.RXB       loope 4afd <.literal16+0x2cd>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -25768,13 +25979,13 @@ BALIGN16
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        4c69 <.literal16+0x499>
+  .byte  224,7                               // loopne        4cc9 <.literal16+0x499>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        4c6d <.literal16+0x49d>
+  .byte  224,7                               // loopne        4ccd <.literal16+0x49d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        4c71 <.literal16+0x4a1>
+  .byte  224,7                               // loopne        4cd1 <.literal16+0x4a1>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        4c75 <.literal16+0x4a5>
+  .byte  224,7                               // loopne        4cd5 <.literal16+0x4a5>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -25808,10 +26019,10 @@ 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)        # a004cb8 <_sk_callback_sse41+0xa0005c8>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a004d18 <_sk_callback_sse41+0xa0005c6>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3004cc0 <_sk_callback_sse41+0x30005d0>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3004d20 <_sk_callback_sse41+0x30005ce>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -25866,11 +26077,11 @@ BALIGN16
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,127,67                            // add           %bh,0x43(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            4d8b <.literal16+0x5bb>
+  .byte  127,67                              // jg            4deb <.literal16+0x5bb>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            4d8f <.literal16+0x5bf>
+  .byte  127,67                              // jg            4def <.literal16+0x5bf>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            4d93 <.literal16+0x5c3>
+  .byte  127,67                              // jg            4df3 <.literal16+0x5c3>
   .byte  129,128,128,59,129,128,128,59,129,128// addl          $0x80813b80,-0x7f7ec480(%rax)
   .byte  128,59,129                          // cmpb          $0x81,(%rbx)
   .byte  128,128,59,129,128,128,59           // addb          $0x3b,-0x7f7f7ec5(%rax)
@@ -25885,16 +26096,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4d84 <.literal16+0x5b4>
+  .byte  127,0                               // jg            4de4 <.literal16+0x5b4>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4d88 <.literal16+0x5b8>
+  .byte  127,0                               // jg            4de8 <.literal16+0x5b8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4d8c <.literal16+0x5bc>
+  .byte  127,0                               // jg            4dec <.literal16+0x5bc>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4d90 <.literal16+0x5c0>
+  .byte  127,0                               // jg            4df0 <.literal16+0x5c0>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -25903,7 +26114,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4e15 <.literal16+0x645>
+  .byte  119,115                             // ja            4e75 <.literal16+0x645>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -25914,7 +26125,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           4d79 <.literal16+0x5a9>
+  .byte  117,191                             // jne           4dd9 <.literal16+0x5a9>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -25926,7 +26137,7 @@ 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          ffffffffe9a38dba <_sk_callback_sse41+0xffffffffe9a346ca>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38e1a <_sk_callback_sse41+0xffffffffe9a346c8>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -25981,16 +26192,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4e54 <.literal16+0x684>
+  .byte  127,0                               // jg            4eb4 <.literal16+0x684>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4e58 <.literal16+0x688>
+  .byte  127,0                               // jg            4eb8 <.literal16+0x688>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4e5c <.literal16+0x68c>
+  .byte  127,0                               // jg            4ebc <.literal16+0x68c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4e60 <.literal16+0x690>
+  .byte  127,0                               // jg            4ec0 <.literal16+0x690>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -25999,7 +26210,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4ee5 <.literal16+0x715>
+  .byte  119,115                             // ja            4f45 <.literal16+0x715>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -26010,7 +26221,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           4e49 <.literal16+0x679>
+  .byte  117,191                             // jne           4ea9 <.literal16+0x679>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -26022,7 +26233,7 @@ 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          ffffffffe9a38e8a <_sk_callback_sse41+0xffffffffe9a3479a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38eea <_sk_callback_sse41+0xffffffffe9a34798>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -26077,16 +26288,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4f24 <.literal16+0x754>
+  .byte  127,0                               // jg            4f84 <.literal16+0x754>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4f28 <.literal16+0x758>
+  .byte  127,0                               // jg            4f88 <.literal16+0x758>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4f2c <.literal16+0x75c>
+  .byte  127,0                               // jg            4f8c <.literal16+0x75c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4f30 <.literal16+0x760>
+  .byte  127,0                               // jg            4f90 <.literal16+0x760>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -26095,7 +26306,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            4fb5 <.literal16+0x7e5>
+  .byte  119,115                             // ja            5015 <.literal16+0x7e5>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -26106,7 +26317,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           4f19 <.literal16+0x749>
+  .byte  117,191                             // jne           4f79 <.literal16+0x749>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -26118,7 +26329,7 @@ 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          ffffffffe9a38f5a <_sk_callback_sse41+0xffffffffe9a3486a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a38fba <_sk_callback_sse41+0xffffffffe9a34868>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -26173,16 +26384,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4ff4 <.literal16+0x824>
+  .byte  127,0                               // jg            5054 <.literal16+0x824>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4ff8 <.literal16+0x828>
+  .byte  127,0                               // jg            5058 <.literal16+0x828>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            4ffc <.literal16+0x82c>
+  .byte  127,0                               // jg            505c <.literal16+0x82c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5000 <.literal16+0x830>
+  .byte  127,0                               // jg            5060 <.literal16+0x830>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -26191,7 +26402,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            5085 <.literal16+0x8b5>
+  .byte  119,115                             // ja            50e5 <.literal16+0x8b5>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -26202,7 +26413,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           4fe9 <.literal16+0x819>
+  .byte  117,191                             // jne           5049 <.literal16+0x819>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -26214,7 +26425,7 @@ 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          ffffffffe9a3902a <_sk_callback_sse41+0xffffffffe9a3493a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3908a <_sk_callback_sse41+0xffffffffe9a34938>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -26265,13 +26476,13 @@ BALIGN16
   .byte  200,66,0,0                          // enterq        $0x42,$0x0
   .byte  200,66,0,0                          // enterq        $0x42,$0x0
   .byte  200,66,0,0                          // enterq        $0x42,$0x0
-  .byte  127,67                              // jg            5107 <.literal16+0x937>
+  .byte  127,67                              // jg            5167 <.literal16+0x937>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            510b <.literal16+0x93b>
+  .byte  127,67                              // jg            516b <.literal16+0x93b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            510f <.literal16+0x93f>
+  .byte  127,67                              // jg            516f <.literal16+0x93f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            5113 <.literal16+0x943>
+  .byte  127,67                              // jg            5173 <.literal16+0x943>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,195                               // add           %al,%bl
   .byte  0,0                                 // add           %al,(%rax)
@@ -26318,16 +26529,16 @@ BALIGN16
   .byte  128,3,62                            // addb          $0x3e,(%rbx)
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           5193 <.literal16+0x9c3>
+  .byte  118,63                              // jbe           51f3 <.literal16+0x9c3>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           5197 <.literal16+0x9c7>
+  .byte  118,63                              // jbe           51f7 <.literal16+0x9c7>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           519b <.literal16+0x9cb>
+  .byte  118,63                              // jbe           51fb <.literal16+0x9cb>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           519f <.literal16+0x9cf>
+  .byte  118,63                              // jbe           51ff <.literal16+0x9cf>
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
@@ -26339,11 +26550,11 @@ BALIGN16
   .byte  128,59,0                            // cmpb          $0x0,(%rbx)
   .byte  0,127,67                            // add           %bh,0x43(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            51db <.literal16+0xa0b>
+  .byte  127,67                              // jg            523b <.literal16+0xa0b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            51df <.literal16+0xa0f>
+  .byte  127,67                              // jg            523f <.literal16+0xa0f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            51e3 <.literal16+0xa13>
+  .byte  127,67                              // jg            5243 <.literal16+0xa13>
   .byte  129,128,128,59,129,128,128,59,129,128// addl          $0x80813b80,-0x7f7ec480(%rax)
   .byte  128,59,129                          // cmpb          $0x81,(%rbx)
   .byte  128,128,59,0,0,128,63               // addb          $0x3f,-0x7fffffc5(%rax)
@@ -26372,7 +26583,7 @@ 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)        # 3005210 <_sk_callback_sse41+0x3000b20>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3005270 <_sk_callback_sse41+0x3000b1e>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -26401,13 +26612,13 @@ BALIGN16
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        5249 <.literal16+0xa79>
+  .byte  224,7                               // loopne        52a9 <.literal16+0xa79>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        524d <.literal16+0xa7d>
+  .byte  224,7                               // loopne        52ad <.literal16+0xa7d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        5251 <.literal16+0xa81>
+  .byte  224,7                               // loopne        52b1 <.literal16+0xa81>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        5255 <.literal16+0xa85>
+  .byte  224,7                               // loopne        52b5 <.literal16+0xa85>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -26453,13 +26664,13 @@ BALIGN16
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        52b9 <.literal16+0xae9>
+  .byte  224,7                               // loopne        5319 <.literal16+0xae9>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        52bd <.literal16+0xaed>
+  .byte  224,7                               // loopne        531d <.literal16+0xaed>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        52c1 <.literal16+0xaf1>
+  .byte  224,7                               // loopne        5321 <.literal16+0xaf1>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        52c5 <.literal16+0xaf5>
+  .byte  224,7                               // loopne        5325 <.literal16+0xaf5>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -26497,13 +26708,13 @@ BALIGN16
   .byte  65,0,0                              // add           %al,(%r8)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            5356 <.literal16+0xb86>
+  .byte  124,66                              // jl            53b6 <.literal16+0xb86>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            535a <.literal16+0xb8a>
+  .byte  124,66                              // jl            53ba <.literal16+0xb8a>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            535e <.literal16+0xb8e>
+  .byte  124,66                              // jl            53be <.literal16+0xb8e>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            5362 <.literal16+0xb92>
+  .byte  124,66                              // jl            53c2 <.literal16+0xb92>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,240                               // add           %dh,%al
@@ -26593,13 +26804,13 @@ BALIGN16
   .byte  136,136,61,137,136,136              // mov           %cl,-0x777776c3(%rax)
   .byte  61,137,136,136,61                   // cmp           $0x3d888889,%eax
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            5465 <.literal16+0xc95>
+  .byte  112,65                              // jo            54c5 <.literal16+0xc95>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            5469 <.literal16+0xc99>
+  .byte  112,65                              // jo            54c9 <.literal16+0xc99>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            546d <.literal16+0xc9d>
+  .byte  112,65                              // jo            54cd <.literal16+0xc9d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            5471 <.literal16+0xca1>
+  .byte  112,65                              // jo            54d1 <.literal16+0xca1>
   .byte  255,0                               // incl          (%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  255,0                               // incl          (%rax)
@@ -26614,7 +26825,7 @@ 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)        # 3005460 <_sk_callback_sse41+0x3000d70>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 30054c0 <_sk_callback_sse41+0x3000d6e>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -26641,7 +26852,7 @@ 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)        # 30054a0 <_sk_callback_sse41+0x3000db0>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3005500 <_sk_callback_sse41+0x3000dae>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -26656,11 +26867,11 @@ BALIGN16
   .byte  255,0                               // incl          (%rax)
   .byte  0,127,67                            // add           %bh,0x43(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            54fb <.literal16+0xd2b>
+  .byte  127,67                              // jg            555b <.literal16+0xd2b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            54ff <.literal16+0xd2f>
+  .byte  127,67                              // jg            555f <.literal16+0xd2f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            5503 <.literal16+0xd33>
+  .byte  127,67                              // jg            5563 <.literal16+0xd33>
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%rax)
@@ -26736,13 +26947,29 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  255                                 // (bad)
-  .byte  127,71                              // jg            55cb <.literal16+0xdfb>
+  .byte  127,71                              // jg            562b <.literal16+0xdfb>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            55cf <.literal16+0xdff>
+  .byte  127,71                              // jg            562f <.literal16+0xdff>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            55d3 <.literal16+0xe03>
+  .byte  127,71                              // jg            5633 <.literal16+0xe03>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            55d7 <.literal16+0xe07>
+  .byte  127,71                              // jg            5637 <.literal16+0xe07>
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
+  .byte  63                                  // (bad)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,191,0,0,128                   // add           %al,-0x7fffff41(%rax)
+  .byte  191,0,0,128,191                     // mov           $0xbf800000,%edi
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,191,0,0,0,63,0                  // cmpb          $0x0,0x3f000000(%rdi)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  63                                  // (bad)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,63                                // add           %bh,(%rdi)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,63                                // add           %bh,(%rdi)
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,208                              // ds            (bad)
@@ -26876,11 +27103,11 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,114                          // cmpb          $0x72,(%rdi)
   .byte  28,199                              // sbb           $0xc7,%al
-  .byte  62,114,28                           // jb,pt         56f2 <.literal16+0xf22>
+  .byte  62,114,28                           // jb,pt         5782 <.literal16+0xf52>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         56f6 <.literal16+0xf26>
+  .byte  62,114,28                           // jb,pt         5786 <.literal16+0xf56>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         56fa <.literal16+0xf2a>
+  .byte  62,114,28                           // jb,pt         578a <.literal16+0xf5a>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -26924,7 +27151,7 @@ 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        # 3d63e585 <_sk_callback_sse41+0x3d639e95>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e615 <_sk_callback_sse41+0x3d639ec3>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -26950,7 +27177,7 @@ 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        # 3d63e5c5 <_sk_callback_sse41+0x3d639ed5>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e655 <_sk_callback_sse41+0x3d639f03>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -26959,13 +27186,13 @@ BALIGN16
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
-  .byte  114,28                              // jb            57be <.literal16+0xfee>
+  .byte  114,28                              // jb            584e <.literal16+0x101e>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         57c2 <.literal16+0xff2>
+  .byte  62,114,28                           // jb,pt         5852 <.literal16+0x1022>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         57c6 <.literal16+0xff6>
+  .byte  62,114,28                           // jb,pt         5856 <.literal16+0x1026>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         57ca <.literal16+0xffa>
+  .byte  62,114,28                           // jb,pt         585a <.literal16+0x102a>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -26986,11 +27213,11 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,114                          // cmpb          $0x72,(%rdi)
   .byte  28,199                              // sbb           $0xc7,%al
-  .byte  62,114,28                           // jb,pt         5802 <.literal16+0x1032>
+  .byte  62,114,28                           // jb,pt         5892 <.literal16+0x1062>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5806 <.literal16+0x1036>
+  .byte  62,114,28                           // jb,pt         5896 <.literal16+0x1066>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         580a <.literal16+0x103a>
+  .byte  62,114,28                           // jb,pt         589a <.literal16+0x106a>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -27034,7 +27261,7 @@ 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        # 3d63e695 <_sk_callback_sse41+0x3d639fa5>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e725 <_sk_callback_sse41+0x3d639fd3>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -27060,7 +27287,7 @@ 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        # 3d63e6d5 <_sk_callback_sse41+0x3d639fe5>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63e765 <_sk_callback_sse41+0x3d63a013>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -27069,13 +27296,13 @@ BALIGN16
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
-  .byte  114,28                              // jb            58ce <.literal16+0x10fe>
+  .byte  114,28                              // jb            595e <.literal16+0x112e>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         58d2 <_sk_callback_sse41+0x11e2>
+  .byte  62,114,28                           // jb,pt         5962 <_sk_callback_sse41+0x1210>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         58d6 <_sk_callback_sse41+0x11e6>
+  .byte  62,114,28                           // jb,pt         5966 <_sk_callback_sse41+0x1214>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         58da <_sk_callback_sse41+0x11ea>
+  .byte  62,114,28                           // jb,pt         596a <_sk_callback_sse41+0x1218>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -27145,7 +27372,7 @@ _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  15,40,21,196,75,0,0                 // movaps        0x4bc4(%rip),%xmm2        # 4c40 <_sk_callback_sse2+0xe1>
+  .byte  15,40,21,68,76,0,0                  // movaps        0x4c44(%rip),%xmm2        # 4cc0 <_sk_callback_sse2+0xd5>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  15,16,2                             // movups        (%rdx),%xmm0
   .byte  15,88,193                           // addps         %xmm1,%xmm0
@@ -27154,7 +27381,7 @@ _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  15,40,21,179,75,0,0                 // movaps        0x4bb3(%rip),%xmm2        # 4c50 <_sk_callback_sse2+0xf1>
+  .byte  15,40,21,51,76,0,0                  // movaps        0x4c33(%rip),%xmm2        # 4cd0 <_sk_callback_sse2+0xe5>
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
@@ -27175,14 +27402,14 @@ _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  102,68,15,111,21,126,75,0,0         // movdqa        0x4b7e(%rip),%xmm10        # 4c60 <_sk_callback_sse2+0x101>
+  .byte  102,68,15,111,21,254,75,0,0         // movdqa        0x4bfe(%rip),%xmm10        # 4ce0 <_sk_callback_sse2+0xf5>
   .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  102,68,15,111,37,106,75,0,0         // movdqa        0x4b6a(%rip),%xmm12        # 4c70 <_sk_callback_sse2+0x111>
-  .byte  102,68,15,111,45,113,75,0,0         // movdqa        0x4b71(%rip),%xmm13        # 4c80 <_sk_callback_sse2+0x121>
+  .byte  102,68,15,111,37,234,75,0,0         // movdqa        0x4bea(%rip),%xmm12        # 4cf0 <_sk_callback_sse2+0x105>
+  .byte  102,68,15,111,45,241,75,0,0         // movdqa        0x4bf1(%rip),%xmm13        # 4d00 <_sk_callback_sse2+0x115>
   .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
@@ -27198,8 +27425,8 @@ _sk_dither_sse2:
   .byte  102,69,15,235,198                   // por           %xmm14,%xmm8
   .byte  102,69,15,235,197                   // por           %xmm13,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,44,75,0,0                // mulps         0x4b2c(%rip),%xmm8        # 4c90 <_sk_callback_sse2+0x131>
-  .byte  68,15,88,5,52,75,0,0                // addps         0x4b34(%rip),%xmm8        # 4ca0 <_sk_callback_sse2+0x141>
+  .byte  68,15,89,5,172,75,0,0               // mulps         0x4bac(%rip),%xmm8        # 4d10 <_sk_callback_sse2+0x125>
+  .byte  68,15,88,5,180,75,0,0               // addps         0x4bb4(%rip),%xmm8        # 4d20 <_sk_callback_sse2+0x135>
   .byte  243,68,15,16,80,8                   // movss         0x8(%rax),%xmm10
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
@@ -27276,7 +27503,7 @@ HIDDEN _sk_srcatop_sse2
 FUNCTION(_sk_srcatop_sse2)
 _sk_srcatop_sse2:
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  68,15,40,5,141,74,0,0               // movaps        0x4a8d(%rip),%xmm8        # 4cb0 <_sk_callback_sse2+0x151>
+  .byte  68,15,40,5,13,75,0,0                // movaps        0x4b0d(%rip),%xmm8        # 4d30 <_sk_callback_sse2+0x145>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -27301,7 +27528,7 @@ FUNCTION(_sk_dstatop_sse2)
 _sk_dstatop_sse2:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
   .byte  68,15,89,196                        // mulps         %xmm4,%xmm8
-  .byte  68,15,40,13,80,74,0,0               // movaps        0x4a50(%rip),%xmm9        # 4cc0 <_sk_callback_sse2+0x161>
+  .byte  68,15,40,13,208,74,0,0              // movaps        0x4ad0(%rip),%xmm9        # 4d40 <_sk_callback_sse2+0x155>
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
@@ -27348,7 +27575,7 @@ HIDDEN _sk_srcout_sse2
 .globl _sk_srcout_sse2
 FUNCTION(_sk_srcout_sse2)
 _sk_srcout_sse2:
-  .byte  68,15,40,5,244,73,0,0               // movaps        0x49f4(%rip),%xmm8        # 4cd0 <_sk_callback_sse2+0x171>
+  .byte  68,15,40,5,116,74,0,0               // movaps        0x4a74(%rip),%xmm8        # 4d50 <_sk_callback_sse2+0x165>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
@@ -27361,7 +27588,7 @@ HIDDEN _sk_dstout_sse2
 .globl _sk_dstout_sse2
 FUNCTION(_sk_dstout_sse2)
 _sk_dstout_sse2:
-  .byte  68,15,40,5,228,73,0,0               // movaps        0x49e4(%rip),%xmm8        # 4ce0 <_sk_callback_sse2+0x181>
+  .byte  68,15,40,5,100,74,0,0               // movaps        0x4a64(%rip),%xmm8        # 4d60 <_sk_callback_sse2+0x175>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -27378,7 +27605,7 @@ HIDDEN _sk_srcover_sse2
 .globl _sk_srcover_sse2
 FUNCTION(_sk_srcover_sse2)
 _sk_srcover_sse2:
-  .byte  68,15,40,5,199,73,0,0               // movaps        0x49c7(%rip),%xmm8        # 4cf0 <_sk_callback_sse2+0x191>
+  .byte  68,15,40,5,71,74,0,0                // movaps        0x4a47(%rip),%xmm8        # 4d70 <_sk_callback_sse2+0x185>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -27398,7 +27625,7 @@ HIDDEN _sk_dstover_sse2
 .globl _sk_dstover_sse2
 FUNCTION(_sk_dstover_sse2)
 _sk_dstover_sse2:
-  .byte  68,15,40,5,155,73,0,0               // movaps        0x499b(%rip),%xmm8        # 4d00 <_sk_callback_sse2+0x1a1>
+  .byte  68,15,40,5,27,74,0,0                // movaps        0x4a1b(%rip),%xmm8        # 4d80 <_sk_callback_sse2+0x195>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -27426,7 +27653,7 @@ HIDDEN _sk_multiply_sse2
 .globl _sk_multiply_sse2
 FUNCTION(_sk_multiply_sse2)
 _sk_multiply_sse2:
-  .byte  68,15,40,5,111,73,0,0               // movaps        0x496f(%rip),%xmm8        # 4d10 <_sk_callback_sse2+0x1b1>
+  .byte  68,15,40,5,239,73,0,0               // movaps        0x49ef(%rip),%xmm8        # 4d90 <_sk_callback_sse2+0x1a5>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
@@ -27502,7 +27729,7 @@ HIDDEN _sk_xor__sse2
 FUNCTION(_sk_xor__sse2)
 _sk_xor__sse2:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
-  .byte  15,40,29,160,72,0,0                 // movaps        0x48a0(%rip),%xmm3        # 4d20 <_sk_callback_sse2+0x1c1>
+  .byte  15,40,29,32,73,0,0                  // movaps        0x4920(%rip),%xmm3        # 4da0 <_sk_callback_sse2+0x1b5>
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
@@ -27550,7 +27777,7 @@ _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  15,40,21,11,72,0,0                  // movaps        0x480b(%rip),%xmm2        # 4d30 <_sk_callback_sse2+0x1d1>
+  .byte  15,40,21,139,72,0,0                 // movaps        0x488b(%rip),%xmm2        # 4db0 <_sk_callback_sse2+0x1c5>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -27584,7 +27811,7 @@ _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  15,40,21,176,71,0,0                 // movaps        0x47b0(%rip),%xmm2        # 4d40 <_sk_callback_sse2+0x1e1>
+  .byte  15,40,21,48,72,0,0                  // movaps        0x4830(%rip),%xmm2        # 4dc0 <_sk_callback_sse2+0x1d5>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -27621,7 +27848,7 @@ _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  15,40,21,74,71,0,0                  // movaps        0x474a(%rip),%xmm2        # 4d50 <_sk_callback_sse2+0x1f1>
+  .byte  15,40,21,202,71,0,0                 // movaps        0x47ca(%rip),%xmm2        # 4dd0 <_sk_callback_sse2+0x1e5>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -27648,7 +27875,7 @@ _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  15,40,13,11,71,0,0                  // movaps        0x470b(%rip),%xmm1        # 4d60 <_sk_callback_sse2+0x201>
+  .byte  15,40,13,139,71,0,0                 // movaps        0x478b(%rip),%xmm1        # 4de0 <_sk_callback_sse2+0x1f5>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,207                           // mulps         %xmm7,%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -27662,7 +27889,7 @@ HIDDEN _sk_colorburn_sse2
 FUNCTION(_sk_colorburn_sse2)
 _sk_colorburn_sse2:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,250,70,0,0              // movaps        0x46fa(%rip),%xmm10        # 4d70 <_sk_callback_sse2+0x211>
+  .byte  68,15,40,21,122,71,0,0              // movaps        0x477a(%rip),%xmm10        # 4df0 <_sk_callback_sse2+0x205>
   .byte  69,15,40,202                        // movaps        %xmm10,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,217                        // movaps        %xmm9,%xmm11
@@ -27756,7 +27983,7 @@ HIDDEN _sk_colordodge_sse2
 FUNCTION(_sk_colordodge_sse2)
 _sk_colordodge_sse2:
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
-  .byte  68,15,40,21,176,69,0,0              // movaps        0x45b0(%rip),%xmm10        # 4d80 <_sk_callback_sse2+0x221>
+  .byte  68,15,40,21,48,70,0,0               // movaps        0x4630(%rip),%xmm10        # 4e00 <_sk_callback_sse2+0x215>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -27850,7 +28077,7 @@ _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  68,15,40,29,101,68,0,0              // movaps        0x4465(%rip),%xmm11        # 4d90 <_sk_callback_sse2+0x231>
+  .byte  68,15,40,29,229,68,0,0              // movaps        0x44e5(%rip),%xmm11        # 4e10 <_sk_callback_sse2+0x225>
   .byte  69,15,40,211                        // movaps        %xmm11,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,194                        // movaps        %xmm10,%xmm8
@@ -27938,7 +28165,7 @@ FUNCTION(_sk_overlay_sse2)
 _sk_overlay_sse2:
   .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
   .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
-  .byte  68,15,40,13,51,67,0,0               // movaps        0x4333(%rip),%xmm9        # 4da0 <_sk_callback_sse2+0x241>
+  .byte  68,15,40,13,179,67,0,0              // movaps        0x43b3(%rip),%xmm9        # 4e20 <_sk_callback_sse2+0x235>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -28029,7 +28256,7 @@ _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  68,15,40,13,240,65,0,0              // movaps        0x41f0(%rip),%xmm9        # 4db0 <_sk_callback_sse2+0x251>
+  .byte  68,15,40,13,112,66,0,0              // movaps        0x4270(%rip),%xmm9        # 4e30 <_sk_callback_sse2+0x245>
   .byte  69,15,40,249                        // movaps        %xmm9,%xmm15
   .byte  69,15,92,250                        // subps         %xmm10,%xmm15
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -28042,10 +28269,10 @@ _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  68,15,40,53,202,65,0,0              // movaps        0x41ca(%rip),%xmm14        # 4dc0 <_sk_callback_sse2+0x261>
+  .byte  68,15,40,53,74,66,0,0               // movaps        0x424a(%rip),%xmm14        # 4e40 <_sk_callback_sse2+0x255>
   .byte  69,15,88,222                        // addps         %xmm14,%xmm11
   .byte  68,15,89,216                        // mulps         %xmm0,%xmm11
-  .byte  68,15,40,21,202,65,0,0              // movaps        0x41ca(%rip),%xmm10        # 4dd0 <_sk_callback_sse2+0x271>
+  .byte  68,15,40,21,74,66,0,0               // movaps        0x424a(%rip),%xmm10        # 4e50 <_sk_callback_sse2+0x265>
   .byte  69,15,89,234                        // mulps         %xmm10,%xmm13
   .byte  69,15,88,235                        // addps         %xmm11,%xmm13
   .byte  15,88,228                           // addps         %xmm4,%xmm4
@@ -28190,7 +28417,7 @@ _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  68,15,40,5,13,64,0,0                // movaps        0x400d(%rip),%xmm8        # 4e10 <_sk_callback_sse2+0x2b1>
+  .byte  68,15,40,5,141,64,0,0               // movaps        0x408d(%rip),%xmm8        # 4e90 <_sk_callback_sse2+0x2a5>
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
   .byte  15,40,207                           // movaps        %xmm7,%xmm1
   .byte  68,15,92,217                        // subps         %xmm1,%xmm11
@@ -28236,12 +28463,12 @@ _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  68,15,40,61,33,63,0,0               // movaps        0x3f21(%rip),%xmm15        # 4de0 <_sk_callback_sse2+0x281>
+  .byte  68,15,40,61,161,63,0,0              // movaps        0x3fa1(%rip),%xmm15        # 4e60 <_sk_callback_sse2+0x275>
   .byte  65,15,89,231                        // mulps         %xmm15,%xmm4
-  .byte  15,40,5,38,63,0,0                   // movaps        0x3f26(%rip),%xmm0        # 4df0 <_sk_callback_sse2+0x291>
+  .byte  15,40,5,166,63,0,0                  // movaps        0x3fa6(%rip),%xmm0        # 4e70 <_sk_callback_sse2+0x285>
   .byte  15,89,240                           // mulps         %xmm0,%xmm6
   .byte  15,88,244                           // addps         %xmm4,%xmm6
-  .byte  68,15,40,53,40,63,0,0               // movaps        0x3f28(%rip),%xmm14        # 4e00 <_sk_callback_sse2+0x2a1>
+  .byte  68,15,40,53,168,63,0,0              // movaps        0x3fa8(%rip),%xmm14        # 4e80 <_sk_callback_sse2+0x295>
   .byte  68,15,40,239                        // movaps        %xmm7,%xmm13
   .byte  69,15,89,238                        // mulps         %xmm14,%xmm13
   .byte  68,15,88,238                        // addps         %xmm6,%xmm13
@@ -28418,14 +28645,14 @@ _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  68,15,40,5,189,60,0,0               // movaps        0x3cbd(%rip),%xmm8        # 4e20 <_sk_callback_sse2+0x2c1>
+  .byte  68,15,40,5,61,61,0,0                // movaps        0x3d3d(%rip),%xmm8        # 4ea0 <_sk_callback_sse2+0x2b5>
   .byte  15,40,214                           // movaps        %xmm6,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
-  .byte  15,40,13,191,60,0,0                 // movaps        0x3cbf(%rip),%xmm1        # 4e30 <_sk_callback_sse2+0x2d1>
+  .byte  15,40,13,63,61,0,0                  // movaps        0x3d3f(%rip),%xmm1        # 4eb0 <_sk_callback_sse2+0x2c5>
   .byte  15,40,221                           // movaps        %xmm5,%xmm3
   .byte  15,89,217                           // mulps         %xmm1,%xmm3
   .byte  15,88,218                           // addps         %xmm2,%xmm3
-  .byte  68,15,40,37,190,60,0,0              // movaps        0x3cbe(%rip),%xmm12        # 4e40 <_sk_callback_sse2+0x2e1>
+  .byte  68,15,40,37,62,61,0,0               // movaps        0x3d3e(%rip),%xmm12        # 4ec0 <_sk_callback_sse2+0x2d5>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
   .byte  68,15,88,235                        // addps         %xmm3,%xmm13
   .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
@@ -28470,7 +28697,7 @@ _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  68,15,40,5,35,60,0,0                // movaps        0x3c23(%rip),%xmm8        # 4e50 <_sk_callback_sse2+0x2f1>
+  .byte  68,15,40,5,163,60,0,0               // movaps        0x3ca3(%rip),%xmm8        # 4ed0 <_sk_callback_sse2+0x2e5>
   .byte  65,15,40,224                        // movaps        %xmm8,%xmm4
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  15,88,253                           // addps         %xmm5,%xmm7
@@ -28571,14 +28798,14 @@ _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  68,15,40,45,193,58,0,0              // movaps        0x3ac1(%rip),%xmm13        # 4e60 <_sk_callback_sse2+0x301>
+  .byte  68,15,40,45,65,59,0,0               // movaps        0x3b41(%rip),%xmm13        # 4ee0 <_sk_callback_sse2+0x2f5>
   .byte  68,15,40,198                        // movaps        %xmm6,%xmm8
   .byte  69,15,89,197                        // mulps         %xmm13,%xmm8
-  .byte  68,15,40,53,193,58,0,0              // movaps        0x3ac1(%rip),%xmm14        # 4e70 <_sk_callback_sse2+0x311>
+  .byte  68,15,40,53,65,59,0,0               // movaps        0x3b41(%rip),%xmm14        # 4ef0 <_sk_callback_sse2+0x305>
   .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  68,15,40,29,189,58,0,0              // movaps        0x3abd(%rip),%xmm11        # 4e80 <_sk_callback_sse2+0x321>
+  .byte  68,15,40,29,61,59,0,0               // movaps        0x3b3d(%rip),%xmm11        # 4f00 <_sk_callback_sse2+0x315>
   .byte  69,15,89,227                        // mulps         %xmm11,%xmm12
   .byte  68,15,88,224                        // addps         %xmm0,%xmm12
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
@@ -28586,7 +28813,7 @@ _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  68,15,40,5,169,58,0,0               // movaps        0x3aa9(%rip),%xmm8        # 4e90 <_sk_callback_sse2+0x331>
+  .byte  68,15,40,5,41,59,0,0                // movaps        0x3b29(%rip),%xmm8        # 4f10 <_sk_callback_sse2+0x325>
   .byte  65,15,40,224                        // movaps        %xmm8,%xmm4
   .byte  15,92,226                           // subps         %xmm2,%xmm4
   .byte  15,89,252                           // mulps         %xmm4,%xmm7
@@ -28722,15 +28949,15 @@ _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  68,15,40,37,192,56,0,0              // movaps        0x38c0(%rip),%xmm12        # 4ea0 <_sk_callback_sse2+0x341>
+  .byte  68,15,40,37,64,57,0,0               // movaps        0x3940(%rip),%xmm12        # 4f20 <_sk_callback_sse2+0x335>
   .byte  68,15,40,199                        // movaps        %xmm7,%xmm8
   .byte  69,15,89,196                        // mulps         %xmm12,%xmm8
-  .byte  68,15,40,45,192,56,0,0              // movaps        0x38c0(%rip),%xmm13        # 4eb0 <_sk_callback_sse2+0x351>
+  .byte  68,15,40,45,64,57,0,0               // movaps        0x3940(%rip),%xmm13        # 4f30 <_sk_callback_sse2+0x345>
   .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  68,15,40,29,188,56,0,0              // movaps        0x38bc(%rip),%xmm11        # 4ec0 <_sk_callback_sse2+0x361>
-  .byte  68,15,40,5,196,56,0,0               // movaps        0x38c4(%rip),%xmm8        # 4ed0 <_sk_callback_sse2+0x371>
+  .byte  68,15,40,29,60,57,0,0               // movaps        0x393c(%rip),%xmm11        # 4f40 <_sk_callback_sse2+0x355>
+  .byte  68,15,40,5,68,57,0,0                // movaps        0x3944(%rip),%xmm8        # 4f50 <_sk_callback_sse2+0x365>
   .byte  69,15,40,248                        // movaps        %xmm8,%xmm15
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  68,15,92,248                        // subps         %xmm0,%xmm15
@@ -28875,7 +29102,7 @@ HIDDEN _sk_clamp_1_sse2
 .globl _sk_clamp_1_sse2
 FUNCTION(_sk_clamp_1_sse2)
 _sk_clamp_1_sse2:
-  .byte  68,15,40,5,205,54,0,0               // movaps        0x36cd(%rip),%xmm8        # 4ee0 <_sk_callback_sse2+0x381>
+  .byte  68,15,40,5,77,55,0,0                // movaps        0x374d(%rip),%xmm8        # 4f60 <_sk_callback_sse2+0x375>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
   .byte  65,15,93,208                        // minps         %xmm8,%xmm2
@@ -28887,7 +29114,7 @@ HIDDEN _sk_clamp_a_sse2
 .globl _sk_clamp_a_sse2
 FUNCTION(_sk_clamp_a_sse2)
 _sk_clamp_a_sse2:
-  .byte  15,93,29,194,54,0,0                 // minps         0x36c2(%rip),%xmm3        # 4ef0 <_sk_callback_sse2+0x391>
+  .byte  15,93,29,66,55,0,0                  // minps         0x3742(%rip),%xmm3        # 4f70 <_sk_callback_sse2+0x385>
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -28974,7 +29201,7 @@ HIDDEN _sk_unpremul_sse2
 FUNCTION(_sk_unpremul_sse2)
 _sk_unpremul_sse2:
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
-  .byte  68,15,40,13,45,54,0,0               // movaps        0x362d(%rip),%xmm9        # 4f00 <_sk_callback_sse2+0x3a1>
+  .byte  68,15,40,13,173,54,0,0              // movaps        0x36ad(%rip),%xmm9        # 4f80 <_sk_callback_sse2+0x395>
   .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
@@ -28988,20 +29215,20 @@ HIDDEN _sk_from_srgb_sse2
 .globl _sk_from_srgb_sse2
 FUNCTION(_sk_from_srgb_sse2)
 _sk_from_srgb_sse2:
-  .byte  68,15,40,5,24,54,0,0                // movaps        0x3618(%rip),%xmm8        # 4f10 <_sk_callback_sse2+0x3b1>
+  .byte  68,15,40,5,152,54,0,0               // movaps        0x3698(%rip),%xmm8        # 4f90 <_sk_callback_sse2+0x3a5>
   .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  68,15,40,13,16,54,0,0               // movaps        0x3610(%rip),%xmm9        # 4f20 <_sk_callback_sse2+0x3c1>
+  .byte  68,15,40,13,144,54,0,0              // movaps        0x3690(%rip),%xmm9        # 4fa0 <_sk_callback_sse2+0x3b5>
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
   .byte  69,15,89,241                        // mulps         %xmm9,%xmm14
-  .byte  68,15,40,21,16,54,0,0               // movaps        0x3610(%rip),%xmm10        # 4f30 <_sk_callback_sse2+0x3d1>
+  .byte  68,15,40,21,144,54,0,0              // movaps        0x3690(%rip),%xmm10        # 4fb0 <_sk_callback_sse2+0x3c5>
   .byte  69,15,88,242                        // addps         %xmm10,%xmm14
   .byte  69,15,89,243                        // mulps         %xmm11,%xmm14
-  .byte  68,15,40,29,16,54,0,0               // movaps        0x3610(%rip),%xmm11        # 4f40 <_sk_callback_sse2+0x3e1>
+  .byte  68,15,40,29,144,54,0,0              // movaps        0x3690(%rip),%xmm11        # 4fc0 <_sk_callback_sse2+0x3d5>
   .byte  69,15,88,243                        // addps         %xmm11,%xmm14
-  .byte  68,15,40,37,20,54,0,0               // movaps        0x3614(%rip),%xmm12        # 4f50 <_sk_callback_sse2+0x3f1>
+  .byte  68,15,40,37,148,54,0,0              // movaps        0x3694(%rip),%xmm12        # 4fd0 <_sk_callback_sse2+0x3e5>
   .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
@@ -29038,22 +29265,22 @@ HIDDEN _sk_to_srgb_sse2
 FUNCTION(_sk_to_srgb_sse2)
 _sk_to_srgb_sse2:
   .byte  68,15,82,232                        // rsqrtps       %xmm0,%xmm13
-  .byte  68,15,40,5,161,53,0,0               // movaps        0x35a1(%rip),%xmm8        # 4f60 <_sk_callback_sse2+0x401>
+  .byte  68,15,40,5,33,54,0,0                // movaps        0x3621(%rip),%xmm8        # 4fe0 <_sk_callback_sse2+0x3f5>
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
   .byte  69,15,89,240                        // mulps         %xmm8,%xmm14
-  .byte  68,15,40,13,161,53,0,0              // movaps        0x35a1(%rip),%xmm9        # 4f70 <_sk_callback_sse2+0x411>
+  .byte  68,15,40,13,33,54,0,0               // movaps        0x3621(%rip),%xmm9        # 4ff0 <_sk_callback_sse2+0x405>
   .byte  69,15,40,253                        // movaps        %xmm13,%xmm15
   .byte  69,15,89,249                        // mulps         %xmm9,%xmm15
-  .byte  68,15,40,21,161,53,0,0              // movaps        0x35a1(%rip),%xmm10        # 4f80 <_sk_callback_sse2+0x421>
+  .byte  68,15,40,21,33,54,0,0               // movaps        0x3621(%rip),%xmm10        # 5000 <_sk_callback_sse2+0x415>
   .byte  69,15,88,250                        // addps         %xmm10,%xmm15
   .byte  69,15,89,253                        // mulps         %xmm13,%xmm15
-  .byte  68,15,40,29,161,53,0,0              // movaps        0x35a1(%rip),%xmm11        # 4f90 <_sk_callback_sse2+0x431>
+  .byte  68,15,40,29,33,54,0,0               // movaps        0x3621(%rip),%xmm11        # 5010 <_sk_callback_sse2+0x425>
   .byte  69,15,88,251                        // addps         %xmm11,%xmm15
-  .byte  68,15,40,37,165,53,0,0              // movaps        0x35a5(%rip),%xmm12        # 4fa0 <_sk_callback_sse2+0x441>
+  .byte  68,15,40,37,37,54,0,0               // movaps        0x3625(%rip),%xmm12        # 5020 <_sk_callback_sse2+0x435>
   .byte  69,15,88,236                        // addps         %xmm12,%xmm13
   .byte  69,15,83,237                        // rcpps         %xmm13,%xmm13
   .byte  69,15,89,239                        // mulps         %xmm15,%xmm13
-  .byte  68,15,40,61,161,53,0,0              // movaps        0x35a1(%rip),%xmm15        # 4fb0 <_sk_callback_sse2+0x451>
+  .byte  68,15,40,61,33,54,0,0               // movaps        0x3621(%rip),%xmm15        # 5030 <_sk_callback_sse2+0x445>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  68,15,84,240                        // andps         %xmm0,%xmm14
   .byte  65,15,85,197                        // andnps        %xmm13,%xmm0
@@ -29103,7 +29330,7 @@ _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  68,15,40,45,250,52,0,0              // movaps        0x34fa(%rip),%xmm13        # 4fc0 <_sk_callback_sse2+0x461>
+  .byte  68,15,40,45,122,53,0,0              // movaps        0x357a(%rip),%xmm13        # 5040 <_sk_callback_sse2+0x455>
   .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
@@ -29112,30 +29339,30 @@ _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  68,15,84,53,224,52,0,0              // andps         0x34e0(%rip),%xmm14        # 4fd0 <_sk_callback_sse2+0x471>
+  .byte  68,15,84,53,96,53,0,0               // andps         0x3560(%rip),%xmm14        # 5050 <_sk_callback_sse2+0x465>
   .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  68,15,40,37,211,52,0,0              // movaps        0x34d3(%rip),%xmm12        # 4fe0 <_sk_callback_sse2+0x481>
+  .byte  68,15,40,37,83,53,0,0               // movaps        0x3553(%rip),%xmm12        # 5060 <_sk_callback_sse2+0x475>
   .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  68,15,88,5,207,52,0,0               // addps         0x34cf(%rip),%xmm8        # 4ff0 <_sk_callback_sse2+0x491>
+  .byte  68,15,88,5,79,53,0,0                // addps         0x354f(%rip),%xmm8        # 5070 <_sk_callback_sse2+0x485>
   .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  15,89,5,192,52,0,0                  // mulps         0x34c0(%rip),%xmm0        # 5000 <_sk_callback_sse2+0x4a1>
+  .byte  15,89,5,64,53,0,0                   // mulps         0x3540(%rip),%xmm0        # 5080 <_sk_callback_sse2+0x495>
   .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  68,15,40,13,179,52,0,0              // movaps        0x34b3(%rip),%xmm9        # 5010 <_sk_callback_sse2+0x4b1>
+  .byte  68,15,40,13,51,53,0,0               // movaps        0x3533(%rip),%xmm9        # 5090 <_sk_callback_sse2+0x4a5>
   .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
@@ -29159,7 +29386,7 @@ _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  68,15,40,13,114,52,0,0              // movaps        0x3472(%rip),%xmm9        # 5020 <_sk_callback_sse2+0x4c1>
+  .byte  68,15,40,13,242,52,0,0              // movaps        0x34f2(%rip),%xmm9        # 50a0 <_sk_callback_sse2+0x4b5>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  69,15,194,211,2                     // cmpleps       %xmm11,%xmm10
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
@@ -29176,28 +29403,28 @@ _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  15,40,5,59,52,0,0                   // movaps        0x343b(%rip),%xmm0        # 5030 <_sk_callback_sse2+0x4d1>
+  .byte  15,40,5,187,52,0,0                  // movaps        0x34bb(%rip),%xmm0        # 50b0 <_sk_callback_sse2+0x4c5>
   .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  15,84,29,51,52,0,0                  // andps         0x3433(%rip),%xmm3        # 5040 <_sk_callback_sse2+0x4e1>
+  .byte  15,84,29,179,52,0,0                 // andps         0x34b3(%rip),%xmm3        # 50c0 <_sk_callback_sse2+0x4d5>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,92,193                           // subps         %xmm1,%xmm0
-  .byte  68,15,40,45,53,52,0,0               // movaps        0x3435(%rip),%xmm13        # 5050 <_sk_callback_sse2+0x4f1>
+  .byte  68,15,40,45,181,52,0,0              // movaps        0x34b5(%rip),%xmm13        # 50d0 <_sk_callback_sse2+0x4e5>
   .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  15,40,21,69,52,0,0                  // movaps        0x3445(%rip),%xmm2        # 5080 <_sk_callback_sse2+0x521>
+  .byte  15,40,21,197,52,0,0                 // movaps        0x34c5(%rip),%xmm2        # 5100 <_sk_callback_sse2+0x515>
   .byte  68,15,40,250                        // movaps        %xmm2,%xmm15
   .byte  68,15,194,248,2                     // cmpleps       %xmm0,%xmm15
-  .byte  15,40,13,21,52,0,0                  // movaps        0x3415(%rip),%xmm1        # 5060 <_sk_callback_sse2+0x501>
+  .byte  15,40,13,149,52,0,0                 // movaps        0x3495(%rip),%xmm1        # 50e0 <_sk_callback_sse2+0x4f5>
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
-  .byte  15,40,45,27,52,0,0                  // movaps        0x341b(%rip),%xmm5        # 5070 <_sk_callback_sse2+0x511>
+  .byte  15,40,45,155,52,0,0                 // movaps        0x349b(%rip),%xmm5        # 50f0 <_sk_callback_sse2+0x505>
   .byte  15,40,229                           // movaps        %xmm5,%xmm4
   .byte  15,92,224                           // subps         %xmm0,%xmm4
   .byte  65,15,89,230                        // mulps         %xmm14,%xmm4
@@ -29220,7 +29447,7 @@ _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  15,84,29,144,51,0,0                 // andps         0x3390(%rip),%xmm3        # 5040 <_sk_callback_sse2+0x4e1>
+  .byte  15,84,29,16,52,0,0                  // andps         0x3410(%rip),%xmm3        # 50c0 <_sk_callback_sse2+0x4d5>
   .byte  15,92,195                           // subps         %xmm3,%xmm0
   .byte  68,15,40,230                        // movaps        %xmm6,%xmm12
   .byte  68,15,92,224                        // subps         %xmm0,%xmm12
@@ -29250,12 +29477,12 @@ _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  15,88,53,104,51,0,0                 // addps         0x3368(%rip),%xmm6        # 5090 <_sk_callback_sse2+0x531>
+  .byte  15,88,53,232,51,0,0                 // addps         0x33e8(%rip),%xmm6        # 5110 <_sk_callback_sse2+0x525>
   .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  15,84,29,3,51,0,0                   // andps         0x3303(%rip),%xmm3        # 5040 <_sk_callback_sse2+0x4e1>
+  .byte  15,84,29,131,51,0,0                 // andps         0x3383(%rip),%xmm3        # 50c0 <_sk_callback_sse2+0x4d5>
   .byte  15,92,195                           // subps         %xmm3,%xmm0
   .byte  15,92,240                           // subps         %xmm0,%xmm6
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
@@ -29319,7 +29546,7 @@ _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  68,15,89,5,145,50,0,0               // mulps         0x3291(%rip),%xmm8        # 50a0 <_sk_callback_sse2+0x541>
+  .byte  68,15,89,5,17,51,0,0                // mulps         0x3311(%rip),%xmm8        # 5120 <_sk_callback_sse2+0x535>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
@@ -29360,7 +29587,7 @@ _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  68,15,89,5,47,50,0,0                // mulps         0x322f(%rip),%xmm8        # 50b0 <_sk_callback_sse2+0x551>
+  .byte  68,15,89,5,175,50,0,0               // mulps         0x32af(%rip),%xmm8        # 5130 <_sk_callback_sse2+0x545>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -29385,17 +29612,17 @@ _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  102,68,15,111,5,245,49,0,0          // movdqa        0x31f5(%rip),%xmm8        # 50c0 <_sk_callback_sse2+0x561>
+  .byte  102,68,15,111,5,117,50,0,0          // movdqa        0x3275(%rip),%xmm8        # 5140 <_sk_callback_sse2+0x555>
   .byte  102,69,15,219,194                   // pand          %xmm10,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,244,49,0,0               // mulps         0x31f4(%rip),%xmm8        # 50d0 <_sk_callback_sse2+0x571>
-  .byte  102,68,15,111,13,251,49,0,0         // movdqa        0x31fb(%rip),%xmm9        # 50e0 <_sk_callback_sse2+0x581>
+  .byte  68,15,89,5,116,50,0,0               // mulps         0x3274(%rip),%xmm8        # 5150 <_sk_callback_sse2+0x565>
+  .byte  102,68,15,111,13,123,50,0,0         // movdqa        0x327b(%rip),%xmm9        # 5160 <_sk_callback_sse2+0x575>
   .byte  102,69,15,219,202                   // pand          %xmm10,%xmm9
   .byte  69,15,91,201                        // cvtdq2ps      %xmm9,%xmm9
-  .byte  68,15,89,13,250,49,0,0              // mulps         0x31fa(%rip),%xmm9        # 50f0 <_sk_callback_sse2+0x591>
-  .byte  102,68,15,219,21,1,50,0,0           // pand          0x3201(%rip),%xmm10        # 5100 <_sk_callback_sse2+0x5a1>
+  .byte  68,15,89,13,122,50,0,0              // mulps         0x327a(%rip),%xmm9        # 5170 <_sk_callback_sse2+0x585>
+  .byte  102,68,15,219,21,129,50,0,0         // pand          0x3281(%rip),%xmm10        # 5180 <_sk_callback_sse2+0x595>
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
-  .byte  68,15,89,21,5,50,0,0                // mulps         0x3205(%rip),%xmm10        # 5110 <_sk_callback_sse2+0x5b1>
+  .byte  68,15,89,21,133,50,0,0              // mulps         0x3285(%rip),%xmm10        # 5190 <_sk_callback_sse2+0x5a5>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -29426,7 +29653,7 @@ _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  102,68,15,111,5,181,49,0,0          // movdqa        0x31b5(%rip),%xmm8        # 5120 <_sk_callback_sse2+0x5c1>
+  .byte  102,68,15,111,5,53,50,0,0           // movdqa        0x3235(%rip),%xmm8        # 51a0 <_sk_callback_sse2+0x5b5>
   .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
@@ -29481,7 +29708,7 @@ _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  15,89,29,194,48,0,0                 // mulps         0x30c2(%rip),%xmm3        # 5130 <_sk_callback_sse2+0x5d1>
+  .byte  15,89,29,66,49,0,0                  // mulps         0x3142(%rip),%xmm3        # 51b0 <_sk_callback_sse2+0x5c5>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -29500,7 +29727,7 @@ _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  102,68,15,111,21,149,48,0,0         // movdqa        0x3095(%rip),%xmm10        # 5140 <_sk_callback_sse2+0x5e1>
+  .byte  102,68,15,111,21,21,49,0,0          // movdqa        0x3115(%rip),%xmm10        # 51c0 <_sk_callback_sse2+0x5d5>
   .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
@@ -29561,7 +29788,7 @@ _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  15,89,29,132,47,0,0                 // mulps         0x2f84(%rip),%xmm3        # 5150 <_sk_callback_sse2+0x5f1>
+  .byte  15,89,29,4,48,0,0                   // mulps         0x3004(%rip),%xmm3        # 51d0 <_sk_callback_sse2+0x5e5>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -29583,7 +29810,7 @@ _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  102,68,15,111,5,68,47,0,0           // movdqa        0x2f44(%rip),%xmm8        # 5160 <_sk_callback_sse2+0x601>
+  .byte  102,68,15,111,5,196,47,0,0          // movdqa        0x2fc4(%rip),%xmm8        # 51e0 <_sk_callback_sse2+0x5f5>
   .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
@@ -29638,7 +29865,7 @@ _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  15,40,29,83,46,0,0                  // movaps        0x2e53(%rip),%xmm3        # 5170 <_sk_callback_sse2+0x611>
+  .byte  15,40,29,211,46,0,0                 // movaps        0x2ed3(%rip),%xmm3        # 51f0 <_sk_callback_sse2+0x605>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_byte_tables_sse2
@@ -29648,7 +29875,7 @@ _sk_byte_tables_sse2:
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,84,46,0,0                // movaps        0x2e54(%rip),%xmm8        # 5180 <_sk_callback_sse2+0x621>
+  .byte  68,15,40,5,212,46,0,0               // movaps        0x2ed4(%rip),%xmm8        # 5200 <_sk_callback_sse2+0x615>
   .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
@@ -29675,7 +29902,7 @@ _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  68,15,40,21,241,45,0,0              // movaps        0x2df1(%rip),%xmm10        # 5190 <_sk_callback_sse2+0x631>
+  .byte  68,15,40,21,113,46,0,0              // movaps        0x2e71(%rip),%xmm10        # 5210 <_sk_callback_sse2+0x625>
   .byte  65,15,89,194                        // mulps         %xmm10,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -29791,7 +30018,7 @@ _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  68,15,40,21,68,44,0,0               // movaps        0x2c44(%rip),%xmm10        # 51a0 <_sk_callback_sse2+0x641>
+  .byte  68,15,40,21,196,44,0,0              // movaps        0x2cc4(%rip),%xmm10        # 5220 <_sk_callback_sse2+0x635>
   .byte  65,15,89,194                        // mulps         %xmm10,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -29988,15 +30215,15 @@ _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  68,15,89,13,131,41,0,0              // mulps         0x2983(%rip),%xmm9        # 51b0 <_sk_callback_sse2+0x651>
-  .byte  68,15,84,21,139,41,0,0              // andps         0x298b(%rip),%xmm10        # 51c0 <_sk_callback_sse2+0x661>
-  .byte  68,15,86,21,147,41,0,0              // orps          0x2993(%rip),%xmm10        # 51d0 <_sk_callback_sse2+0x671>
-  .byte  68,15,88,13,155,41,0,0              // addps         0x299b(%rip),%xmm9        # 51e0 <_sk_callback_sse2+0x681>
-  .byte  68,15,40,37,163,41,0,0              // movaps        0x29a3(%rip),%xmm12        # 51f0 <_sk_callback_sse2+0x691>
+  .byte  68,15,89,13,3,42,0,0                // mulps         0x2a03(%rip),%xmm9        # 5230 <_sk_callback_sse2+0x645>
+  .byte  68,15,84,21,11,42,0,0               // andps         0x2a0b(%rip),%xmm10        # 5240 <_sk_callback_sse2+0x655>
+  .byte  68,15,86,21,19,42,0,0               // orps          0x2a13(%rip),%xmm10        # 5250 <_sk_callback_sse2+0x665>
+  .byte  68,15,88,13,27,42,0,0               // addps         0x2a1b(%rip),%xmm9        # 5260 <_sk_callback_sse2+0x675>
+  .byte  68,15,40,37,35,42,0,0               // movaps        0x2a23(%rip),%xmm12        # 5270 <_sk_callback_sse2+0x685>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,163,41,0,0              // addps         0x29a3(%rip),%xmm10        # 5200 <_sk_callback_sse2+0x6a1>
-  .byte  68,15,40,37,171,41,0,0              // movaps        0x29ab(%rip),%xmm12        # 5210 <_sk_callback_sse2+0x6b1>
+  .byte  68,15,88,21,35,42,0,0               // addps         0x2a23(%rip),%xmm10        # 5280 <_sk_callback_sse2+0x695>
+  .byte  68,15,40,37,43,42,0,0               // movaps        0x2a2b(%rip),%xmm12        # 5290 <_sk_callback_sse2+0x6a5>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -30004,22 +30231,22 @@ _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  68,15,40,21,149,41,0,0              // movaps        0x2995(%rip),%xmm10        # 5220 <_sk_callback_sse2+0x6c1>
+  .byte  68,15,40,21,21,42,0,0               // movaps        0x2a15(%rip),%xmm10        # 52a0 <_sk_callback_sse2+0x6b5>
   .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  68,15,88,13,137,41,0,0              // addps         0x2989(%rip),%xmm9        # 5230 <_sk_callback_sse2+0x6d1>
-  .byte  68,15,40,37,145,41,0,0              // movaps        0x2991(%rip),%xmm12        # 5240 <_sk_callback_sse2+0x6e1>
+  .byte  68,15,88,13,9,42,0,0                // addps         0x2a09(%rip),%xmm9        # 52b0 <_sk_callback_sse2+0x6c5>
+  .byte  68,15,40,37,17,42,0,0               // movaps        0x2a11(%rip),%xmm12        # 52c0 <_sk_callback_sse2+0x6d5>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,145,41,0,0              // movaps        0x2991(%rip),%xmm12        # 5250 <_sk_callback_sse2+0x6f1>
+  .byte  68,15,40,37,17,42,0,0               // movaps        0x2a11(%rip),%xmm12        # 52d0 <_sk_callback_sse2+0x6e5>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,149,41,0,0              // movaps        0x2995(%rip),%xmm13        # 5260 <_sk_callback_sse2+0x701>
+  .byte  68,15,40,45,21,42,0,0               // movaps        0x2a15(%rip),%xmm13        # 52e0 <_sk_callback_sse2+0x6f5>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,149,41,0,0              // mulps         0x2995(%rip),%xmm13        # 5270 <_sk_callback_sse2+0x711>
+  .byte  68,15,89,45,21,42,0,0               // mulps         0x2a15(%rip),%xmm13        # 52f0 <_sk_callback_sse2+0x705>
   .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
@@ -30055,15 +30282,15 @@ _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  68,15,89,13,21,41,0,0               // mulps         0x2915(%rip),%xmm9        # 5280 <_sk_callback_sse2+0x721>
-  .byte  68,15,84,21,29,41,0,0               // andps         0x291d(%rip),%xmm10        # 5290 <_sk_callback_sse2+0x731>
-  .byte  68,15,86,21,37,41,0,0               // orps          0x2925(%rip),%xmm10        # 52a0 <_sk_callback_sse2+0x741>
-  .byte  68,15,88,13,45,41,0,0               // addps         0x292d(%rip),%xmm9        # 52b0 <_sk_callback_sse2+0x751>
-  .byte  68,15,40,37,53,41,0,0               // movaps        0x2935(%rip),%xmm12        # 52c0 <_sk_callback_sse2+0x761>
+  .byte  68,15,89,13,149,41,0,0              // mulps         0x2995(%rip),%xmm9        # 5300 <_sk_callback_sse2+0x715>
+  .byte  68,15,84,21,157,41,0,0              // andps         0x299d(%rip),%xmm10        # 5310 <_sk_callback_sse2+0x725>
+  .byte  68,15,86,21,165,41,0,0              // orps          0x29a5(%rip),%xmm10        # 5320 <_sk_callback_sse2+0x735>
+  .byte  68,15,88,13,173,41,0,0              // addps         0x29ad(%rip),%xmm9        # 5330 <_sk_callback_sse2+0x745>
+  .byte  68,15,40,37,181,41,0,0              // movaps        0x29b5(%rip),%xmm12        # 5340 <_sk_callback_sse2+0x755>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,53,41,0,0               // addps         0x2935(%rip),%xmm10        # 52d0 <_sk_callback_sse2+0x771>
-  .byte  68,15,40,37,61,41,0,0               // movaps        0x293d(%rip),%xmm12        # 52e0 <_sk_callback_sse2+0x781>
+  .byte  68,15,88,21,181,41,0,0              // addps         0x29b5(%rip),%xmm10        # 5350 <_sk_callback_sse2+0x765>
+  .byte  68,15,40,37,189,41,0,0              // movaps        0x29bd(%rip),%xmm12        # 5360 <_sk_callback_sse2+0x775>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -30071,22 +30298,22 @@ _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  68,15,40,21,39,41,0,0               // movaps        0x2927(%rip),%xmm10        # 52f0 <_sk_callback_sse2+0x791>
+  .byte  68,15,40,21,167,41,0,0              // movaps        0x29a7(%rip),%xmm10        # 5370 <_sk_callback_sse2+0x785>
   .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  68,15,88,13,27,41,0,0               // addps         0x291b(%rip),%xmm9        # 5300 <_sk_callback_sse2+0x7a1>
-  .byte  68,15,40,37,35,41,0,0               // movaps        0x2923(%rip),%xmm12        # 5310 <_sk_callback_sse2+0x7b1>
+  .byte  68,15,88,13,155,41,0,0              // addps         0x299b(%rip),%xmm9        # 5380 <_sk_callback_sse2+0x795>
+  .byte  68,15,40,37,163,41,0,0              // movaps        0x29a3(%rip),%xmm12        # 5390 <_sk_callback_sse2+0x7a5>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,35,41,0,0               // movaps        0x2923(%rip),%xmm12        # 5320 <_sk_callback_sse2+0x7c1>
+  .byte  68,15,40,37,163,41,0,0              // movaps        0x29a3(%rip),%xmm12        # 53a0 <_sk_callback_sse2+0x7b5>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,39,41,0,0               // movaps        0x2927(%rip),%xmm13        # 5330 <_sk_callback_sse2+0x7d1>
+  .byte  68,15,40,45,167,41,0,0              // movaps        0x29a7(%rip),%xmm13        # 53b0 <_sk_callback_sse2+0x7c5>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,39,41,0,0               // mulps         0x2927(%rip),%xmm13        # 5340 <_sk_callback_sse2+0x7e1>
+  .byte  68,15,89,45,167,41,0,0              // mulps         0x29a7(%rip),%xmm13        # 53c0 <_sk_callback_sse2+0x7d5>
   .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
@@ -30122,15 +30349,15 @@ _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  68,15,89,13,167,40,0,0              // mulps         0x28a7(%rip),%xmm9        # 5350 <_sk_callback_sse2+0x7f1>
-  .byte  68,15,84,21,175,40,0,0              // andps         0x28af(%rip),%xmm10        # 5360 <_sk_callback_sse2+0x801>
-  .byte  68,15,86,21,183,40,0,0              // orps          0x28b7(%rip),%xmm10        # 5370 <_sk_callback_sse2+0x811>
-  .byte  68,15,88,13,191,40,0,0              // addps         0x28bf(%rip),%xmm9        # 5380 <_sk_callback_sse2+0x821>
-  .byte  68,15,40,37,199,40,0,0              // movaps        0x28c7(%rip),%xmm12        # 5390 <_sk_callback_sse2+0x831>
+  .byte  68,15,89,13,39,41,0,0               // mulps         0x2927(%rip),%xmm9        # 53d0 <_sk_callback_sse2+0x7e5>
+  .byte  68,15,84,21,47,41,0,0               // andps         0x292f(%rip),%xmm10        # 53e0 <_sk_callback_sse2+0x7f5>
+  .byte  68,15,86,21,55,41,0,0               // orps          0x2937(%rip),%xmm10        # 53f0 <_sk_callback_sse2+0x805>
+  .byte  68,15,88,13,63,41,0,0               // addps         0x293f(%rip),%xmm9        # 5400 <_sk_callback_sse2+0x815>
+  .byte  68,15,40,37,71,41,0,0               // movaps        0x2947(%rip),%xmm12        # 5410 <_sk_callback_sse2+0x825>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,199,40,0,0              // addps         0x28c7(%rip),%xmm10        # 53a0 <_sk_callback_sse2+0x841>
-  .byte  68,15,40,37,207,40,0,0              // movaps        0x28cf(%rip),%xmm12        # 53b0 <_sk_callback_sse2+0x851>
+  .byte  68,15,88,21,71,41,0,0               // addps         0x2947(%rip),%xmm10        # 5420 <_sk_callback_sse2+0x835>
+  .byte  68,15,40,37,79,41,0,0               // movaps        0x294f(%rip),%xmm12        # 5430 <_sk_callback_sse2+0x845>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -30138,22 +30365,22 @@ _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  68,15,40,21,185,40,0,0              // movaps        0x28b9(%rip),%xmm10        # 53c0 <_sk_callback_sse2+0x861>
+  .byte  68,15,40,21,57,41,0,0               // movaps        0x2939(%rip),%xmm10        # 5440 <_sk_callback_sse2+0x855>
   .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  68,15,88,13,173,40,0,0              // addps         0x28ad(%rip),%xmm9        # 53d0 <_sk_callback_sse2+0x871>
-  .byte  68,15,40,37,181,40,0,0              // movaps        0x28b5(%rip),%xmm12        # 53e0 <_sk_callback_sse2+0x881>
+  .byte  68,15,88,13,45,41,0,0               // addps         0x292d(%rip),%xmm9        # 5450 <_sk_callback_sse2+0x865>
+  .byte  68,15,40,37,53,41,0,0               // movaps        0x2935(%rip),%xmm12        # 5460 <_sk_callback_sse2+0x875>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,181,40,0,0              // movaps        0x28b5(%rip),%xmm12        # 53f0 <_sk_callback_sse2+0x891>
+  .byte  68,15,40,37,53,41,0,0               // movaps        0x2935(%rip),%xmm12        # 5470 <_sk_callback_sse2+0x885>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,185,40,0,0              // movaps        0x28b9(%rip),%xmm13        # 5400 <_sk_callback_sse2+0x8a1>
+  .byte  68,15,40,45,57,41,0,0               // movaps        0x2939(%rip),%xmm13        # 5480 <_sk_callback_sse2+0x895>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,185,40,0,0              // mulps         0x28b9(%rip),%xmm13        # 5410 <_sk_callback_sse2+0x8b1>
+  .byte  68,15,89,45,57,41,0,0               // mulps         0x2939(%rip),%xmm13        # 5490 <_sk_callback_sse2+0x8a5>
   .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
@@ -30189,15 +30416,15 @@ _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  68,15,89,13,57,40,0,0               // mulps         0x2839(%rip),%xmm9        # 5420 <_sk_callback_sse2+0x8c1>
-  .byte  68,15,84,21,65,40,0,0               // andps         0x2841(%rip),%xmm10        # 5430 <_sk_callback_sse2+0x8d1>
-  .byte  68,15,86,21,73,40,0,0               // orps          0x2849(%rip),%xmm10        # 5440 <_sk_callback_sse2+0x8e1>
-  .byte  68,15,88,13,81,40,0,0               // addps         0x2851(%rip),%xmm9        # 5450 <_sk_callback_sse2+0x8f1>
-  .byte  68,15,40,37,89,40,0,0               // movaps        0x2859(%rip),%xmm12        # 5460 <_sk_callback_sse2+0x901>
+  .byte  68,15,89,13,185,40,0,0              // mulps         0x28b9(%rip),%xmm9        # 54a0 <_sk_callback_sse2+0x8b5>
+  .byte  68,15,84,21,193,40,0,0              // andps         0x28c1(%rip),%xmm10        # 54b0 <_sk_callback_sse2+0x8c5>
+  .byte  68,15,86,21,201,40,0,0              // orps          0x28c9(%rip),%xmm10        # 54c0 <_sk_callback_sse2+0x8d5>
+  .byte  68,15,88,13,209,40,0,0              // addps         0x28d1(%rip),%xmm9        # 54d0 <_sk_callback_sse2+0x8e5>
+  .byte  68,15,40,37,217,40,0,0              // movaps        0x28d9(%rip),%xmm12        # 54e0 <_sk_callback_sse2+0x8f5>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,89,40,0,0               // addps         0x2859(%rip),%xmm10        # 5470 <_sk_callback_sse2+0x911>
-  .byte  68,15,40,37,97,40,0,0               // movaps        0x2861(%rip),%xmm12        # 5480 <_sk_callback_sse2+0x921>
+  .byte  68,15,88,21,217,40,0,0              // addps         0x28d9(%rip),%xmm10        # 54f0 <_sk_callback_sse2+0x905>
+  .byte  68,15,40,37,225,40,0,0              // movaps        0x28e1(%rip),%xmm12        # 5500 <_sk_callback_sse2+0x915>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -30205,22 +30432,22 @@ _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  68,15,40,21,75,40,0,0               // movaps        0x284b(%rip),%xmm10        # 5490 <_sk_callback_sse2+0x931>
+  .byte  68,15,40,21,203,40,0,0              // movaps        0x28cb(%rip),%xmm10        # 5510 <_sk_callback_sse2+0x925>
   .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  68,15,88,13,63,40,0,0               // addps         0x283f(%rip),%xmm9        # 54a0 <_sk_callback_sse2+0x941>
-  .byte  68,15,40,37,71,40,0,0               // movaps        0x2847(%rip),%xmm12        # 54b0 <_sk_callback_sse2+0x951>
+  .byte  68,15,88,13,191,40,0,0              // addps         0x28bf(%rip),%xmm9        # 5520 <_sk_callback_sse2+0x935>
+  .byte  68,15,40,37,199,40,0,0              // movaps        0x28c7(%rip),%xmm12        # 5530 <_sk_callback_sse2+0x945>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,71,40,0,0               // movaps        0x2847(%rip),%xmm12        # 54c0 <_sk_callback_sse2+0x961>
+  .byte  68,15,40,37,199,40,0,0              // movaps        0x28c7(%rip),%xmm12        # 5540 <_sk_callback_sse2+0x955>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,75,40,0,0               // movaps        0x284b(%rip),%xmm13        # 54d0 <_sk_callback_sse2+0x971>
+  .byte  68,15,40,45,203,40,0,0              // movaps        0x28cb(%rip),%xmm13        # 5550 <_sk_callback_sse2+0x965>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,75,40,0,0               // mulps         0x284b(%rip),%xmm13        # 54e0 <_sk_callback_sse2+0x981>
+  .byte  68,15,89,45,203,40,0,0              // mulps         0x28cb(%rip),%xmm13        # 5560 <_sk_callback_sse2+0x975>
   .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
@@ -30237,29 +30464,29 @@ HIDDEN _sk_lab_to_xyz_sse2
 .globl _sk_lab_to_xyz_sse2
 FUNCTION(_sk_lab_to_xyz_sse2)
 _sk_lab_to_xyz_sse2:
-  .byte  15,89,5,40,40,0,0                   // mulps         0x2828(%rip),%xmm0        # 54f0 <_sk_callback_sse2+0x991>
-  .byte  68,15,40,5,48,40,0,0                // movaps        0x2830(%rip),%xmm8        # 5500 <_sk_callback_sse2+0x9a1>
+  .byte  15,89,5,168,40,0,0                  // mulps         0x28a8(%rip),%xmm0        # 5570 <_sk_callback_sse2+0x985>
+  .byte  68,15,40,5,176,40,0,0               // movaps        0x28b0(%rip),%xmm8        # 5580 <_sk_callback_sse2+0x995>
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
-  .byte  68,15,40,13,52,40,0,0               // movaps        0x2834(%rip),%xmm9        # 5510 <_sk_callback_sse2+0x9b1>
+  .byte  68,15,40,13,180,40,0,0              // movaps        0x28b4(%rip),%xmm9        # 5590 <_sk_callback_sse2+0x9a5>
   .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  15,88,5,49,40,0,0                   // addps         0x2831(%rip),%xmm0        # 5520 <_sk_callback_sse2+0x9c1>
-  .byte  15,89,5,58,40,0,0                   // mulps         0x283a(%rip),%xmm0        # 5530 <_sk_callback_sse2+0x9d1>
-  .byte  15,89,13,67,40,0,0                  // mulps         0x2843(%rip),%xmm1        # 5540 <_sk_callback_sse2+0x9e1>
+  .byte  15,88,5,177,40,0,0                  // addps         0x28b1(%rip),%xmm0        # 55a0 <_sk_callback_sse2+0x9b5>
+  .byte  15,89,5,186,40,0,0                  // mulps         0x28ba(%rip),%xmm0        # 55b0 <_sk_callback_sse2+0x9c5>
+  .byte  15,89,13,195,40,0,0                 // mulps         0x28c3(%rip),%xmm1        # 55c0 <_sk_callback_sse2+0x9d5>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,21,73,40,0,0                  // mulps         0x2849(%rip),%xmm2        # 5550 <_sk_callback_sse2+0x9f1>
+  .byte  15,89,21,201,40,0,0                 // mulps         0x28c9(%rip),%xmm2        # 55d0 <_sk_callback_sse2+0x9e5>
   .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  15,40,21,62,40,0,0                  // movaps        0x283e(%rip),%xmm2        # 5560 <_sk_callback_sse2+0xa01>
+  .byte  15,40,21,190,40,0,0                 // movaps        0x28be(%rip),%xmm2        # 55e0 <_sk_callback_sse2+0x9f5>
   .byte  68,15,40,194                        // movaps        %xmm2,%xmm8
   .byte  69,15,194,196,1                     // cmpltps       %xmm12,%xmm8
-  .byte  68,15,40,21,61,40,0,0               // movaps        0x283d(%rip),%xmm10        # 5570 <_sk_callback_sse2+0xa11>
+  .byte  68,15,40,21,189,40,0,0              // movaps        0x28bd(%rip),%xmm10        # 55f0 <_sk_callback_sse2+0xa05>
   .byte  65,15,88,202                        // addps         %xmm10,%xmm1
-  .byte  68,15,40,29,65,40,0,0               // movaps        0x2841(%rip),%xmm11        # 5580 <_sk_callback_sse2+0xa21>
+  .byte  68,15,40,29,193,40,0,0              // movaps        0x28c1(%rip),%xmm11        # 5600 <_sk_callback_sse2+0xa15>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  69,15,84,224                        // andps         %xmm8,%xmm12
   .byte  68,15,85,193                        // andnps        %xmm1,%xmm8
@@ -30283,8 +30510,8 @@ _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  68,15,89,5,241,39,0,0               // mulps         0x27f1(%rip),%xmm8        # 5590 <_sk_callback_sse2+0xa31>
-  .byte  15,89,21,250,39,0,0                 // mulps         0x27fa(%rip),%xmm2        # 55a0 <_sk_callback_sse2+0xa41>
+  .byte  68,15,89,5,113,40,0,0               // mulps         0x2871(%rip),%xmm8        # 5610 <_sk_callback_sse2+0xa25>
+  .byte  15,89,21,122,40,0,0                 // mulps         0x287a(%rip),%xmm2        # 5620 <_sk_callback_sse2+0xa35>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -30300,7 +30527,7 @@ _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  15,89,29,226,39,0,0                 // mulps         0x27e2(%rip),%xmm3        # 55b0 <_sk_callback_sse2+0xa51>
+  .byte  15,89,29,98,40,0,0                  // mulps         0x2862(%rip),%xmm3        # 5630 <_sk_callback_sse2+0xa45>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -30345,7 +30572,7 @@ _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  15,89,29,81,39,0,0                  // mulps         0x2751(%rip),%xmm3        # 55c0 <_sk_callback_sse2+0xa61>
+  .byte  15,89,29,209,39,0,0                 // mulps         0x27d1(%rip),%xmm3        # 5640 <_sk_callback_sse2+0xa55>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -30358,7 +30585,7 @@ FUNCTION(_sk_store_a8_sse2)
 _sk_store_a8_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,69,39,0,0                // movaps        0x2745(%rip),%xmm8        # 55d0 <_sk_callback_sse2+0xa71>
+  .byte  68,15,40,5,197,39,0,0               // movaps        0x27c5(%rip),%xmm8        # 5650 <_sk_callback_sse2+0xa65>
   .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
@@ -30380,9 +30607,9 @@ _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  15,89,5,12,39,0,0                   // mulps         0x270c(%rip),%xmm0        # 55e0 <_sk_callback_sse2+0xa81>
+  .byte  15,89,5,140,39,0,0                  // mulps         0x278c(%rip),%xmm0        # 5660 <_sk_callback_sse2+0xa75>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,19,39,0,0                  // movaps        0x2713(%rip),%xmm3        # 55f0 <_sk_callback_sse2+0xa91>
+  .byte  15,40,29,147,39,0,0                 // movaps        0x2793(%rip),%xmm3        # 5670 <_sk_callback_sse2+0xa85>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -30425,9 +30652,9 @@ _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  15,89,5,136,38,0,0                  // mulps         0x2688(%rip),%xmm0        # 5600 <_sk_callback_sse2+0xaa1>
+  .byte  15,89,5,8,39,0,0                    // mulps         0x2708(%rip),%xmm0        # 5680 <_sk_callback_sse2+0xa95>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,143,38,0,0                 // movaps        0x268f(%rip),%xmm3        # 5610 <_sk_callback_sse2+0xab1>
+  .byte  15,40,29,15,39,0,0                  // movaps        0x270f(%rip),%xmm3        # 5690 <_sk_callback_sse2+0xaa5>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -30490,11 +30717,11 @@ _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  102,15,111,21,174,37,0,0            // movdqa        0x25ae(%rip),%xmm2        # 5620 <_sk_callback_sse2+0xac1>
+  .byte  102,15,111,21,46,38,0,0             // movdqa        0x262e(%rip),%xmm2        # 56a0 <_sk_callback_sse2+0xab5>
   .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  68,15,40,5,170,37,0,0               // movaps        0x25aa(%rip),%xmm8        # 5630 <_sk_callback_sse2+0xad1>
+  .byte  68,15,40,5,42,38,0,0                // movaps        0x262a(%rip),%xmm8        # 56b0 <_sk_callback_sse2+0xac5>
   .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
@@ -30521,19 +30748,19 @@ _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  102,15,111,5,96,37,0,0              // movdqa        0x2560(%rip),%xmm0        # 5640 <_sk_callback_sse2+0xae1>
+  .byte  102,15,111,5,224,37,0,0             // movdqa        0x25e0(%rip),%xmm0        # 56c0 <_sk_callback_sse2+0xad5>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,98,37,0,0                   // mulps         0x2562(%rip),%xmm0        # 5650 <_sk_callback_sse2+0xaf1>
-  .byte  102,15,111,13,106,37,0,0            // movdqa        0x256a(%rip),%xmm1        # 5660 <_sk_callback_sse2+0xb01>
+  .byte  15,89,5,226,37,0,0                  // mulps         0x25e2(%rip),%xmm0        # 56d0 <_sk_callback_sse2+0xae5>
+  .byte  102,15,111,13,234,37,0,0            // movdqa        0x25ea(%rip),%xmm1        # 56e0 <_sk_callback_sse2+0xaf5>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,108,37,0,0                 // mulps         0x256c(%rip),%xmm1        # 5670 <_sk_callback_sse2+0xb11>
-  .byte  102,15,219,21,116,37,0,0            // pand          0x2574(%rip),%xmm2        # 5680 <_sk_callback_sse2+0xb21>
+  .byte  15,89,13,236,37,0,0                 // mulps         0x25ec(%rip),%xmm1        # 56f0 <_sk_callback_sse2+0xb05>
+  .byte  102,15,219,21,244,37,0,0            // pand          0x25f4(%rip),%xmm2        # 5700 <_sk_callback_sse2+0xb15>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,122,37,0,0                 // mulps         0x257a(%rip),%xmm2        # 5690 <_sk_callback_sse2+0xb31>
+  .byte  15,89,21,250,37,0,0                 // mulps         0x25fa(%rip),%xmm2        # 5710 <_sk_callback_sse2+0xb25>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,129,37,0,0                 // movaps        0x2581(%rip),%xmm3        # 56a0 <_sk_callback_sse2+0xb41>
+  .byte  15,40,29,1,38,0,0                   // movaps        0x2601(%rip),%xmm3        # 5720 <_sk_callback_sse2+0xb35>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_gather_565_sse2
@@ -30568,19 +30795,19 @@ _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  102,15,111,5,10,37,0,0              // movdqa        0x250a(%rip),%xmm0        # 56b0 <_sk_callback_sse2+0xb51>
+  .byte  102,15,111,5,138,37,0,0             // movdqa        0x258a(%rip),%xmm0        # 5730 <_sk_callback_sse2+0xb45>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,12,37,0,0                   // mulps         0x250c(%rip),%xmm0        # 56c0 <_sk_callback_sse2+0xb61>
-  .byte  102,15,111,13,20,37,0,0             // movdqa        0x2514(%rip),%xmm1        # 56d0 <_sk_callback_sse2+0xb71>
+  .byte  15,89,5,140,37,0,0                  // mulps         0x258c(%rip),%xmm0        # 5740 <_sk_callback_sse2+0xb55>
+  .byte  102,15,111,13,148,37,0,0            // movdqa        0x2594(%rip),%xmm1        # 5750 <_sk_callback_sse2+0xb65>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,22,37,0,0                  // mulps         0x2516(%rip),%xmm1        # 56e0 <_sk_callback_sse2+0xb81>
-  .byte  102,15,219,21,30,37,0,0             // pand          0x251e(%rip),%xmm2        # 56f0 <_sk_callback_sse2+0xb91>
+  .byte  15,89,13,150,37,0,0                 // mulps         0x2596(%rip),%xmm1        # 5760 <_sk_callback_sse2+0xb75>
+  .byte  102,15,219,21,158,37,0,0            // pand          0x259e(%rip),%xmm2        # 5770 <_sk_callback_sse2+0xb85>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,36,37,0,0                  // mulps         0x2524(%rip),%xmm2        # 5700 <_sk_callback_sse2+0xba1>
+  .byte  15,89,21,164,37,0,0                 // mulps         0x25a4(%rip),%xmm2        # 5780 <_sk_callback_sse2+0xb95>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,43,37,0,0                  // movaps        0x252b(%rip),%xmm3        # 5710 <_sk_callback_sse2+0xbb1>
+  .byte  15,40,29,171,37,0,0                 // movaps        0x25ab(%rip),%xmm3        # 5790 <_sk_callback_sse2+0xba5>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_565_sse2
@@ -30589,12 +30816,12 @@ FUNCTION(_sk_store_565_sse2)
 _sk_store_565_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,44,37,0,0                // movaps        0x252c(%rip),%xmm8        # 5720 <_sk_callback_sse2+0xbc1>
+  .byte  68,15,40,5,172,37,0,0               // movaps        0x25ac(%rip),%xmm8        # 57a0 <_sk_callback_sse2+0xbb5>
   .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  68,15,40,21,33,37,0,0               // movaps        0x2521(%rip),%xmm10        # 5730 <_sk_callback_sse2+0xbd1>
+  .byte  68,15,40,21,161,37,0,0              // movaps        0x25a1(%rip),%xmm10        # 57b0 <_sk_callback_sse2+0xbc5>
   .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
@@ -30618,21 +30845,21 @@ _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  102,15,111,5,218,36,0,0             // movdqa        0x24da(%rip),%xmm0        # 5740 <_sk_callback_sse2+0xbe1>
+  .byte  102,15,111,5,90,37,0,0              // movdqa        0x255a(%rip),%xmm0        # 57c0 <_sk_callback_sse2+0xbd5>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,220,36,0,0                  // mulps         0x24dc(%rip),%xmm0        # 5750 <_sk_callback_sse2+0xbf1>
-  .byte  102,15,111,13,228,36,0,0            // movdqa        0x24e4(%rip),%xmm1        # 5760 <_sk_callback_sse2+0xc01>
+  .byte  15,89,5,92,37,0,0                   // mulps         0x255c(%rip),%xmm0        # 57d0 <_sk_callback_sse2+0xbe5>
+  .byte  102,15,111,13,100,37,0,0            // movdqa        0x2564(%rip),%xmm1        # 57e0 <_sk_callback_sse2+0xbf5>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,230,36,0,0                 // mulps         0x24e6(%rip),%xmm1        # 5770 <_sk_callback_sse2+0xc11>
-  .byte  102,15,111,21,238,36,0,0            // movdqa        0x24ee(%rip),%xmm2        # 5780 <_sk_callback_sse2+0xc21>
+  .byte  15,89,13,102,37,0,0                 // mulps         0x2566(%rip),%xmm1        # 57f0 <_sk_callback_sse2+0xc05>
+  .byte  102,15,111,21,110,37,0,0            // movdqa        0x256e(%rip),%xmm2        # 5800 <_sk_callback_sse2+0xc15>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,240,36,0,0                 // mulps         0x24f0(%rip),%xmm2        # 5790 <_sk_callback_sse2+0xc31>
-  .byte  102,15,219,29,248,36,0,0            // pand          0x24f8(%rip),%xmm3        # 57a0 <_sk_callback_sse2+0xc41>
+  .byte  15,89,21,112,37,0,0                 // mulps         0x2570(%rip),%xmm2        # 5810 <_sk_callback_sse2+0xc25>
+  .byte  102,15,219,29,120,37,0,0            // pand          0x2578(%rip),%xmm3        # 5820 <_sk_callback_sse2+0xc35>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,254,36,0,0                 // mulps         0x24fe(%rip),%xmm3        # 57b0 <_sk_callback_sse2+0xc51>
+  .byte  15,89,29,126,37,0,0                 // mulps         0x257e(%rip),%xmm3        # 5830 <_sk_callback_sse2+0xc45>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -30668,21 +30895,21 @@ _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  102,15,111,5,133,36,0,0             // movdqa        0x2485(%rip),%xmm0        # 57c0 <_sk_callback_sse2+0xc61>
+  .byte  102,15,111,5,5,37,0,0               // movdqa        0x2505(%rip),%xmm0        # 5840 <_sk_callback_sse2+0xc55>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,135,36,0,0                  // mulps         0x2487(%rip),%xmm0        # 57d0 <_sk_callback_sse2+0xc71>
-  .byte  102,15,111,13,143,36,0,0            // movdqa        0x248f(%rip),%xmm1        # 57e0 <_sk_callback_sse2+0xc81>
+  .byte  15,89,5,7,37,0,0                    // mulps         0x2507(%rip),%xmm0        # 5850 <_sk_callback_sse2+0xc65>
+  .byte  102,15,111,13,15,37,0,0             // movdqa        0x250f(%rip),%xmm1        # 5860 <_sk_callback_sse2+0xc75>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,145,36,0,0                 // mulps         0x2491(%rip),%xmm1        # 57f0 <_sk_callback_sse2+0xc91>
-  .byte  102,15,111,21,153,36,0,0            // movdqa        0x2499(%rip),%xmm2        # 5800 <_sk_callback_sse2+0xca1>
+  .byte  15,89,13,17,37,0,0                  // mulps         0x2511(%rip),%xmm1        # 5870 <_sk_callback_sse2+0xc85>
+  .byte  102,15,111,21,25,37,0,0             // movdqa        0x2519(%rip),%xmm2        # 5880 <_sk_callback_sse2+0xc95>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,155,36,0,0                 // mulps         0x249b(%rip),%xmm2        # 5810 <_sk_callback_sse2+0xcb1>
-  .byte  102,15,219,29,163,36,0,0            // pand          0x24a3(%rip),%xmm3        # 5820 <_sk_callback_sse2+0xcc1>
+  .byte  15,89,21,27,37,0,0                  // mulps         0x251b(%rip),%xmm2        # 5890 <_sk_callback_sse2+0xca5>
+  .byte  102,15,219,29,35,37,0,0             // pand          0x2523(%rip),%xmm3        # 58a0 <_sk_callback_sse2+0xcb5>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,169,36,0,0                 // mulps         0x24a9(%rip),%xmm3        # 5830 <_sk_callback_sse2+0xcd1>
+  .byte  15,89,29,41,37,0,0                  // mulps         0x2529(%rip),%xmm3        # 58b0 <_sk_callback_sse2+0xcc5>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -30692,7 +30919,7 @@ FUNCTION(_sk_store_4444_sse2)
 _sk_store_4444_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,168,36,0,0               // movaps        0x24a8(%rip),%xmm8        # 5840 <_sk_callback_sse2+0xce1>
+  .byte  68,15,40,5,40,37,0,0                // movaps        0x2528(%rip),%xmm8        # 58c0 <_sk_callback_sse2+0xcd5>
   .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
@@ -30724,11 +30951,11 @@ _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  15,40,21,59,36,0,0                  // movaps        0x243b(%rip),%xmm2        # 5850 <_sk_callback_sse2+0xcf1>
+  .byte  15,40,21,187,36,0,0                 // movaps        0x24bb(%rip),%xmm2        # 58d0 <_sk_callback_sse2+0xce5>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,84,194                           // andps         %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,57,36,0,0                // movaps        0x2439(%rip),%xmm8        # 5860 <_sk_callback_sse2+0xd01>
+  .byte  68,15,40,5,185,36,0,0               // movaps        0x24b9(%rip),%xmm8        # 58e0 <_sk_callback_sse2+0xcf5>
   .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
@@ -30777,11 +31004,11 @@ _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  102,15,111,21,138,35,0,0            // movdqa        0x238a(%rip),%xmm2        # 5870 <_sk_callback_sse2+0xd11>
+  .byte  102,15,111,21,10,36,0,0             // movdqa        0x240a(%rip),%xmm2        # 58f0 <_sk_callback_sse2+0xd05>
   .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  68,15,40,5,134,35,0,0               // movaps        0x2386(%rip),%xmm8        # 5880 <_sk_callback_sse2+0xd21>
+  .byte  68,15,40,5,6,36,0,0                 // movaps        0x2406(%rip),%xmm8        # 5900 <_sk_callback_sse2+0xd15>
   .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
@@ -30805,7 +31032,7 @@ FUNCTION(_sk_store_8888_sse2)
 _sk_store_8888_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  68,15,40,5,73,35,0,0                // movaps        0x2349(%rip),%xmm8        # 5890 <_sk_callback_sse2+0xd31>
+  .byte  68,15,40,5,201,35,0,0               // movaps        0x23c9(%rip),%xmm8        # 5910 <_sk_callback_sse2+0xd25>
   .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
@@ -30844,7 +31071,7 @@ _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  102,68,15,111,13,185,34,0,0         // movdqa        0x22b9(%rip),%xmm9        # 58a0 <_sk_callback_sse2+0xd41>
+  .byte  102,68,15,111,13,57,35,0,0          // movdqa        0x2339(%rip),%xmm9        # 5920 <_sk_callback_sse2+0xd35>
   .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
@@ -30852,11 +31079,11 @@ _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  102,68,15,111,29,158,34,0,0         // movdqa        0x229e(%rip),%xmm11        # 58b0 <_sk_callback_sse2+0xd51>
+  .byte  102,68,15,111,29,30,35,0,0          // movdqa        0x231e(%rip),%xmm11        # 5930 <_sk_callback_sse2+0xd45>
   .byte  102,69,15,254,235                   // paddd         %xmm11,%xmm13
-  .byte  102,68,15,111,37,160,34,0,0         // movdqa        0x22a0(%rip),%xmm12        # 58c0 <_sk_callback_sse2+0xd61>
+  .byte  102,68,15,111,37,32,35,0,0          // movdqa        0x2320(%rip),%xmm12        # 5940 <_sk_callback_sse2+0xd55>
   .byte  102,65,15,239,204                   // pxor          %xmm12,%xmm1
-  .byte  102,15,111,29,163,34,0,0            // movdqa        0x22a3(%rip),%xmm3        # 58d0 <_sk_callback_sse2+0xd71>
+  .byte  102,15,111,29,35,35,0,0             // movdqa        0x2323(%rip),%xmm3        # 5950 <_sk_callback_sse2+0xd65>
   .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
@@ -30942,7 +31169,7 @@ _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  102,68,15,111,13,49,33,0,0          // movdqa        0x2131(%rip),%xmm9        # 58e0 <_sk_callback_sse2+0xd81>
+  .byte  102,68,15,111,13,177,33,0,0         // movdqa        0x21b1(%rip),%xmm9        # 5960 <_sk_callback_sse2+0xd75>
   .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
@@ -30950,11 +31177,11 @@ _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  102,68,15,111,29,22,33,0,0          // movdqa        0x2116(%rip),%xmm11        # 58f0 <_sk_callback_sse2+0xd91>
+  .byte  102,68,15,111,29,150,33,0,0         // movdqa        0x2196(%rip),%xmm11        # 5970 <_sk_callback_sse2+0xd85>
   .byte  102,69,15,254,235                   // paddd         %xmm11,%xmm13
-  .byte  102,68,15,111,37,24,33,0,0          // movdqa        0x2118(%rip),%xmm12        # 5900 <_sk_callback_sse2+0xda1>
+  .byte  102,68,15,111,37,152,33,0,0         // movdqa        0x2198(%rip),%xmm12        # 5980 <_sk_callback_sse2+0xd95>
   .byte  102,65,15,239,204                   // pxor          %xmm12,%xmm1
-  .byte  102,15,111,29,27,33,0,0             // movdqa        0x211b(%rip),%xmm3        # 5910 <_sk_callback_sse2+0xdb1>
+  .byte  102,15,111,29,155,33,0,0            // movdqa        0x219b(%rip),%xmm3        # 5990 <_sk_callback_sse2+0xda5>
   .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
@@ -31007,17 +31234,17 @@ FUNCTION(_sk_store_f16_sse2)
 _sk_store_f16_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  102,68,15,111,21,67,32,0,0          // movdqa        0x2043(%rip),%xmm10        # 5920 <_sk_callback_sse2+0xdc1>
+  .byte  102,68,15,111,21,195,32,0,0         // movdqa        0x20c3(%rip),%xmm10        # 59a0 <_sk_callback_sse2+0xdb5>
   .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  102,68,15,111,13,54,32,0,0          // movdqa        0x2036(%rip),%xmm9        # 5930 <_sk_callback_sse2+0xdd1>
+  .byte  102,68,15,111,13,182,32,0,0         // movdqa        0x20b6(%rip),%xmm9        # 59b0 <_sk_callback_sse2+0xdc5>
   .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  102,68,15,111,29,39,32,0,0          // movdqa        0x2027(%rip),%xmm11        # 5940 <_sk_callback_sse2+0xde1>
+  .byte  102,68,15,111,29,167,32,0,0         // movdqa        0x20a7(%rip),%xmm11        # 59c0 <_sk_callback_sse2+0xdd5>
   .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
@@ -31096,7 +31323,7 @@ _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  68,15,40,5,197,30,0,0               // movaps        0x1ec5(%rip),%xmm8        # 5950 <_sk_callback_sse2+0xdf1>
+  .byte  68,15,40,5,69,31,0,0                // movaps        0x1f45(%rip),%xmm8        # 59d0 <_sk_callback_sse2+0xde5>
   .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
@@ -31149,7 +31376,7 @@ _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  68,15,40,13,1,30,0,0                // movaps        0x1e01(%rip),%xmm9        # 5960 <_sk_callback_sse2+0xe01>
+  .byte  68,15,40,13,129,30,0,0              // movaps        0x1e81(%rip),%xmm9        # 59e0 <_sk_callback_sse2+0xdf5>
   .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
@@ -31166,7 +31393,7 @@ _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  15,40,29,200,29,0,0                 // movaps        0x1dc8(%rip),%xmm3        # 5970 <_sk_callback_sse2+0xe11>
+  .byte  15,40,29,72,30,0,0                  // movaps        0x1e48(%rip),%xmm3        # 59f0 <_sk_callback_sse2+0xe05>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_u16_be_sse2
@@ -31175,7 +31402,7 @@ 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  68,15,40,13,201,29,0,0              // movaps        0x1dc9(%rip),%xmm9        # 5980 <_sk_callback_sse2+0xe21>
+  .byte  68,15,40,13,73,30,0,0               // movaps        0x1e49(%rip),%xmm9        # 5a00 <_sk_callback_sse2+0xe15>
   .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
@@ -31321,7 +31548,7 @@ _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  68,15,84,13,201,27,0,0              // andps         0x1bc9(%rip),%xmm9        # 5990 <_sk_callback_sse2+0xe31>
+  .byte  68,15,84,13,73,28,0,0               // andps         0x1c49(%rip),%xmm9        # 5a10 <_sk_callback_sse2+0xe25>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
   .byte  65,15,92,194                        // subps         %xmm10,%xmm0
@@ -31341,7 +31568,7 @@ _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  68,15,84,13,155,27,0,0              // andps         0x1b9b(%rip),%xmm9        # 59a0 <_sk_callback_sse2+0xe41>
+  .byte  68,15,84,13,27,28,0,0               // andps         0x1c1b(%rip),%xmm9        # 5a20 <_sk_callback_sse2+0xe35>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
   .byte  65,15,92,202                        // subps         %xmm10,%xmm1
@@ -31365,7 +31592,7 @@ _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  68,15,84,21,91,27,0,0               // andps         0x1b5b(%rip),%xmm10        # 59b0 <_sk_callback_sse2+0xe51>
+  .byte  68,15,84,21,219,27,0,0              // andps         0x1bdb(%rip),%xmm10        # 5a30 <_sk_callback_sse2+0xe45>
   .byte  69,15,87,228                        // xorps         %xmm12,%xmm12
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
@@ -31393,7 +31620,7 @@ _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  68,15,84,21,11,27,0,0               // andps         0x1b0b(%rip),%xmm10        # 59c0 <_sk_callback_sse2+0xe61>
+  .byte  68,15,84,21,139,27,0,0              // andps         0x1b8b(%rip),%xmm10        # 5a40 <_sk_callback_sse2+0xe55>
   .byte  69,15,87,228                        // xorps         %xmm12,%xmm12
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
@@ -31405,15 +31632,62 @@ _sk_mirror_y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_clamp_x_1_sse2
+.globl _sk_clamp_x_1_sse2
+FUNCTION(_sk_clamp_x_1_sse2)
+_sk_clamp_x_1_sse2:
+  .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
+  .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
+  .byte  68,15,93,5,103,27,0,0               // minps         0x1b67(%rip),%xmm8        # 5a50 <_sk_callback_sse2+0xe65>
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_repeat_x_1_sse2
+.globl _sk_repeat_x_1_sse2
+FUNCTION(_sk_repeat_x_1_sse2)
+_sk_repeat_x_1_sse2:
+  .byte  243,68,15,91,192                    // cvttps2dq     %xmm0,%xmm8
+  .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
+  .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
+  .byte  69,15,194,200,1                     // cmpltps       %xmm8,%xmm9
+  .byte  68,15,84,13,85,27,0,0               // andps         0x1b55(%rip),%xmm9        # 5a60 <_sk_callback_sse2+0xe75>
+  .byte  69,15,92,193                        // subps         %xmm9,%xmm8
+  .byte  65,15,92,192                        // subps         %xmm8,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_mirror_x_1_sse2
+.globl _sk_mirror_x_1_sse2
+FUNCTION(_sk_mirror_x_1_sse2)
+_sk_mirror_x_1_sse2:
+  .byte  68,15,40,5,81,27,0,0                // movaps        0x1b51(%rip),%xmm8        # 5a70 <_sk_callback_sse2+0xe85>
+  .byte  65,15,88,192                        // addps         %xmm8,%xmm0
+  .byte  68,15,40,13,85,27,0,0               // movaps        0x1b55(%rip),%xmm9        # 5a80 <_sk_callback_sse2+0xe95>
+  .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
+  .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  68,15,84,13,75,27,0,0               // andps         0x1b4b(%rip),%xmm9        # 5a90 <_sk_callback_sse2+0xea5>
+  .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
+  .byte  69,15,92,209                        // subps         %xmm9,%xmm10
+  .byte  69,15,88,210                        // addps         %xmm10,%xmm10
+  .byte  65,15,92,194                        // subps         %xmm10,%xmm0
+  .byte  65,15,88,192                        // addps         %xmm8,%xmm0
+  .byte  68,15,92,216                        // subps         %xmm0,%xmm11
+  .byte  65,15,84,195                        // andps         %xmm11,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_luminance_to_alpha_sse2
 .globl _sk_luminance_to_alpha_sse2
 FUNCTION(_sk_luminance_to_alpha_sse2)
 _sk_luminance_to_alpha_sse2:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
-  .byte  15,89,5,237,26,0,0                  // mulps         0x1aed(%rip),%xmm0        # 59d0 <_sk_callback_sse2+0xe71>
-  .byte  15,89,13,246,26,0,0                 // mulps         0x1af6(%rip),%xmm1        # 59e0 <_sk_callback_sse2+0xe81>
+  .byte  15,89,5,49,27,0,0                   // mulps         0x1b31(%rip),%xmm0        # 5aa0 <_sk_callback_sse2+0xeb5>
+  .byte  15,89,13,58,27,0,0                  // mulps         0x1b3a(%rip),%xmm1        # 5ab0 <_sk_callback_sse2+0xec5>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,29,252,26,0,0                 // mulps         0x1afc(%rip),%xmm3        # 59f0 <_sk_callback_sse2+0xe91>
+  .byte  15,89,29,64,27,0,0                  // mulps         0x1b40(%rip),%xmm3        # 5ac0 <_sk_callback_sse2+0xed5>
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -31689,9 +31963,9 @@ _sk_evenly_spaced_gradient_sse2:
   .byte  72,139,8                            // mov           (%rax),%rcx
   .byte  76,139,88,8                         // mov           0x8(%rax),%r11
   .byte  72,255,201                          // dec           %rcx
-  .byte  120,7                               // js            432d <_sk_evenly_spaced_gradient_sse2+0x15>
+  .byte  120,7                               // js            43b9 <_sk_evenly_spaced_gradient_sse2+0x15>
   .byte  243,72,15,42,201                    // cvtsi2ss      %rcx,%xmm1
-  .byte  235,21                              // jmp           4342 <_sk_evenly_spaced_gradient_sse2+0x2a>
+  .byte  235,21                              // jmp           43ce <_sk_evenly_spaced_gradient_sse2+0x2a>
   .byte  73,137,200                          // mov           %rcx,%r8
   .byte  73,209,232                          // shr           %r8
   .byte  131,225,1                           // and           $0x1,%ecx
@@ -31791,12 +32065,12 @@ _sk_gradient_sse2:
   .byte  76,139,0                            // mov           (%rax),%r8
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  73,131,248,2                        // cmp           $0x2,%r8
-  .byte  114,50                              // jb            4505 <_sk_gradient_sse2+0x41>
+  .byte  114,50                              // jb            4591 <_sk_gradient_sse2+0x41>
   .byte  72,139,72,72                        // mov           0x48(%rax),%rcx
   .byte  73,255,200                          // dec           %r8
   .byte  72,131,193,4                        // add           $0x4,%rcx
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
-  .byte  15,40,21,23,21,0,0                  // movaps        0x1517(%rip),%xmm2        # 5a00 <_sk_callback_sse2+0xea1>
+  .byte  15,40,21,91,21,0,0                  // movaps        0x155b(%rip),%xmm2        # 5ad0 <_sk_callback_sse2+0xee5>
   .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
@@ -31804,7 +32078,7 @@ _sk_gradient_sse2:
   .byte  102,15,254,203                      // paddd         %xmm3,%xmm1
   .byte  72,131,193,4                        // add           $0x4,%rcx
   .byte  73,255,200                          // dec           %r8
-  .byte  117,228                             // jne           44e9 <_sk_gradient_sse2+0x25>
+  .byte  117,228                             // jne           4575 <_sk_gradient_sse2+0x25>
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
@@ -31944,29 +32218,29 @@ _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  68,15,40,45,217,18,0,0              // movaps        0x12d9(%rip),%xmm13        # 5a10 <_sk_callback_sse2+0xeb1>
+  .byte  68,15,40,45,29,19,0,0               // movaps        0x131d(%rip),%xmm13        # 5ae0 <_sk_callback_sse2+0xef5>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,221,18,0,0              // addps         0x12dd(%rip),%xmm13        # 5a20 <_sk_callback_sse2+0xec1>
+  .byte  68,15,88,45,33,19,0,0               // addps         0x1321(%rip),%xmm13        # 5af0 <_sk_callback_sse2+0xf05>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,225,18,0,0              // addps         0x12e1(%rip),%xmm13        # 5a30 <_sk_callback_sse2+0xed1>
+  .byte  68,15,88,45,37,19,0,0               // addps         0x1325(%rip),%xmm13        # 5b00 <_sk_callback_sse2+0xf15>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,229,18,0,0              // addps         0x12e5(%rip),%xmm13        # 5a40 <_sk_callback_sse2+0xee1>
+  .byte  68,15,88,45,41,19,0,0               // addps         0x1329(%rip),%xmm13        # 5b10 <_sk_callback_sse2+0xf25>
   .byte  69,15,89,235                        // mulps         %xmm11,%xmm13
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,40,21,228,18,0,0              // movaps        0x12e4(%rip),%xmm10        # 5a50 <_sk_callback_sse2+0xef1>
+  .byte  68,15,40,21,40,19,0,0               // movaps        0x1328(%rip),%xmm10        # 5b20 <_sk_callback_sse2+0xf35>
   .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  68,15,40,21,215,18,0,0              // movaps        0x12d7(%rip),%xmm10        # 5a60 <_sk_callback_sse2+0xf01>
+  .byte  68,15,40,21,27,19,0,0               // movaps        0x131b(%rip),%xmm10        # 5b30 <_sk_callback_sse2+0xf45>
   .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  68,15,40,21,198,18,0,0              // movaps        0x12c6(%rip),%xmm10        # 5a70 <_sk_callback_sse2+0xf11>
+  .byte  68,15,40,21,10,19,0,0               // movaps        0x130a(%rip),%xmm10        # 5b40 <_sk_callback_sse2+0xf55>
   .byte  69,15,92,208                        // subps         %xmm8,%xmm10
   .byte  69,15,84,209                        // andps         %xmm9,%xmm10
   .byte  69,15,85,200                        // andnps        %xmm8,%xmm9
@@ -31993,7 +32267,7 @@ HIDDEN _sk_save_xy_sse2
 FUNCTION(_sk_save_xy_sse2)
 _sk_save_xy_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,152,18,0,0               // movaps        0x1298(%rip),%xmm8        # 5a80 <_sk_callback_sse2+0xf21>
+  .byte  68,15,40,5,220,18,0,0               // movaps        0x12dc(%rip),%xmm8        # 5b50 <_sk_callback_sse2+0xf65>
   .byte  15,17,0                             // movups        %xmm0,(%rax)
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,88,200                        // addps         %xmm8,%xmm9
@@ -32001,7 +32275,7 @@ _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  68,15,40,37,131,18,0,0              // movaps        0x1283(%rip),%xmm12        # 5a90 <_sk_callback_sse2+0xf31>
+  .byte  68,15,40,37,199,18,0,0              // movaps        0x12c7(%rip),%xmm12        # 5b60 <_sk_callback_sse2+0xf75>
   .byte  69,15,84,220                        // andps         %xmm12,%xmm11
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
@@ -32048,8 +32322,8 @@ _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  15,88,5,252,17,0,0                  // addps         0x11fc(%rip),%xmm0        # 5aa0 <_sk_callback_sse2+0xf41>
-  .byte  68,15,40,13,4,18,0,0                // movaps        0x1204(%rip),%xmm9        # 5ab0 <_sk_callback_sse2+0xf51>
+  .byte  15,88,5,64,18,0,0                   // addps         0x1240(%rip),%xmm0        # 5b70 <_sk_callback_sse2+0xf85>
+  .byte  68,15,40,13,72,18,0,0               // movaps        0x1248(%rip),%xmm9        # 5b80 <_sk_callback_sse2+0xf95>
   .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
@@ -32062,7 +32336,7 @@ _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  15,88,5,243,17,0,0                  // addps         0x11f3(%rip),%xmm0        # 5ac0 <_sk_callback_sse2+0xf61>
+  .byte  15,88,5,55,18,0,0                   // addps         0x1237(%rip),%xmm0        # 5b90 <_sk_callback_sse2+0xfa5>
   .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
@@ -32074,8 +32348,8 @@ _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  15,88,13,229,17,0,0                 // addps         0x11e5(%rip),%xmm1        # 5ad0 <_sk_callback_sse2+0xf71>
-  .byte  68,15,40,13,237,17,0,0              // movaps        0x11ed(%rip),%xmm9        # 5ae0 <_sk_callback_sse2+0xf81>
+  .byte  15,88,13,41,18,0,0                  // addps         0x1229(%rip),%xmm1        # 5ba0 <_sk_callback_sse2+0xfb5>
+  .byte  68,15,40,13,49,18,0,0               // movaps        0x1231(%rip),%xmm9        # 5bb0 <_sk_callback_sse2+0xfc5>
   .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
@@ -32088,7 +32362,7 @@ _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  15,88,13,219,17,0,0                 // addps         0x11db(%rip),%xmm1        # 5af0 <_sk_callback_sse2+0xf91>
+  .byte  15,88,13,31,18,0,0                  // addps         0x121f(%rip),%xmm1        # 5bc0 <_sk_callback_sse2+0xfd5>
   .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
@@ -32100,13 +32374,13 @@ _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  15,88,5,206,17,0,0                  // addps         0x11ce(%rip),%xmm0        # 5b00 <_sk_callback_sse2+0xfa1>
-  .byte  68,15,40,13,214,17,0,0              // movaps        0x11d6(%rip),%xmm9        # 5b10 <_sk_callback_sse2+0xfb1>
+  .byte  15,88,5,18,18,0,0                   // addps         0x1212(%rip),%xmm0        # 5bd0 <_sk_callback_sse2+0xfe5>
+  .byte  68,15,40,13,26,18,0,0               // movaps        0x121a(%rip),%xmm9        # 5be0 <_sk_callback_sse2+0xff5>
   .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  68,15,89,13,210,17,0,0              // mulps         0x11d2(%rip),%xmm9        # 5b20 <_sk_callback_sse2+0xfc1>
-  .byte  68,15,88,13,218,17,0,0              // addps         0x11da(%rip),%xmm9        # 5b30 <_sk_callback_sse2+0xfd1>
+  .byte  68,15,89,13,22,18,0,0               // mulps         0x1216(%rip),%xmm9        # 5bf0 <_sk_callback_sse2+0x1005>
+  .byte  68,15,88,13,30,18,0,0               // addps         0x121e(%rip),%xmm9        # 5c00 <_sk_callback_sse2+0x1015>
   .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
@@ -32119,16 +32393,16 @@ _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  15,88,5,201,17,0,0                  // addps         0x11c9(%rip),%xmm0        # 5b40 <_sk_callback_sse2+0xfe1>
-  .byte  68,15,40,13,209,17,0,0              // movaps        0x11d1(%rip),%xmm9        # 5b50 <_sk_callback_sse2+0xff1>
+  .byte  15,88,5,13,18,0,0                   // addps         0x120d(%rip),%xmm0        # 5c10 <_sk_callback_sse2+0x1025>
+  .byte  68,15,40,13,21,18,0,0               // movaps        0x1215(%rip),%xmm9        # 5c20 <_sk_callback_sse2+0x1035>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,213,17,0,0               // movaps        0x11d5(%rip),%xmm8        # 5b60 <_sk_callback_sse2+0x1001>
+  .byte  68,15,40,5,25,18,0,0                // movaps        0x1219(%rip),%xmm8        # 5c30 <_sk_callback_sse2+0x1045>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,217,17,0,0               // addps         0x11d9(%rip),%xmm8        # 5b70 <_sk_callback_sse2+0x1011>
+  .byte  68,15,88,5,29,18,0,0                // addps         0x121d(%rip),%xmm8        # 5c40 <_sk_callback_sse2+0x1055>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,221,17,0,0               // addps         0x11dd(%rip),%xmm8        # 5b80 <_sk_callback_sse2+0x1021>
+  .byte  68,15,88,5,33,18,0,0                // addps         0x1221(%rip),%xmm8        # 5c50 <_sk_callback_sse2+0x1065>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,225,17,0,0               // addps         0x11e1(%rip),%xmm8        # 5b90 <_sk_callback_sse2+0x1031>
+  .byte  68,15,88,5,37,18,0,0                // addps         0x1225(%rip),%xmm8        # 5c60 <_sk_callback_sse2+0x1075>
   .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
@@ -32138,17 +32412,17 @@ HIDDEN _sk_bicubic_p1x_sse2
 FUNCTION(_sk_bicubic_p1x_sse2)
 _sk_bicubic_p1x_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,219,17,0,0               // movaps        0x11db(%rip),%xmm8        # 5ba0 <_sk_callback_sse2+0x1041>
+  .byte  68,15,40,5,31,18,0,0                // movaps        0x121f(%rip),%xmm8        # 5c70 <_sk_callback_sse2+0x1085>
   .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  68,15,40,21,215,17,0,0              // movaps        0x11d7(%rip),%xmm10        # 5bb0 <_sk_callback_sse2+0x1051>
+  .byte  68,15,40,21,27,18,0,0               // movaps        0x121b(%rip),%xmm10        # 5c80 <_sk_callback_sse2+0x1095>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,219,17,0,0              // addps         0x11db(%rip),%xmm10        # 5bc0 <_sk_callback_sse2+0x1061>
+  .byte  68,15,88,21,31,18,0,0               // addps         0x121f(%rip),%xmm10        # 5c90 <_sk_callback_sse2+0x10a5>
   .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  68,15,88,21,215,17,0,0              // addps         0x11d7(%rip),%xmm10        # 5bd0 <_sk_callback_sse2+0x1071>
+  .byte  68,15,88,21,27,18,0,0               // addps         0x121b(%rip),%xmm10        # 5ca0 <_sk_callback_sse2+0x10b5>
   .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
@@ -32160,11 +32434,11 @@ _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  15,88,5,202,17,0,0                  // addps         0x11ca(%rip),%xmm0        # 5be0 <_sk_callback_sse2+0x1081>
+  .byte  15,88,5,14,18,0,0                   // addps         0x120e(%rip),%xmm0        # 5cb0 <_sk_callback_sse2+0x10c5>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,202,17,0,0               // mulps         0x11ca(%rip),%xmm8        # 5bf0 <_sk_callback_sse2+0x1091>
-  .byte  68,15,88,5,210,17,0,0               // addps         0x11d2(%rip),%xmm8        # 5c00 <_sk_callback_sse2+0x10a1>
+  .byte  68,15,89,5,14,18,0,0                // mulps         0x120e(%rip),%xmm8        # 5cc0 <_sk_callback_sse2+0x10d5>
+  .byte  68,15,88,5,22,18,0,0                // addps         0x1216(%rip),%xmm8        # 5cd0 <_sk_callback_sse2+0x10e5>
   .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
@@ -32177,13 +32451,13 @@ _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  15,88,13,192,17,0,0                 // addps         0x11c0(%rip),%xmm1        # 5c10 <_sk_callback_sse2+0x10b1>
-  .byte  68,15,40,13,200,17,0,0              // movaps        0x11c8(%rip),%xmm9        # 5c20 <_sk_callback_sse2+0x10c1>
+  .byte  15,88,13,4,18,0,0                   // addps         0x1204(%rip),%xmm1        # 5ce0 <_sk_callback_sse2+0x10f5>
+  .byte  68,15,40,13,12,18,0,0               // movaps        0x120c(%rip),%xmm9        # 5cf0 <_sk_callback_sse2+0x1105>
   .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  68,15,89,13,196,17,0,0              // mulps         0x11c4(%rip),%xmm9        # 5c30 <_sk_callback_sse2+0x10d1>
-  .byte  68,15,88,13,204,17,0,0              // addps         0x11cc(%rip),%xmm9        # 5c40 <_sk_callback_sse2+0x10e1>
+  .byte  68,15,89,13,8,18,0,0                // mulps         0x1208(%rip),%xmm9        # 5d00 <_sk_callback_sse2+0x1115>
+  .byte  68,15,88,13,16,18,0,0               // addps         0x1210(%rip),%xmm9        # 5d10 <_sk_callback_sse2+0x1125>
   .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
@@ -32196,16 +32470,16 @@ _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  15,88,13,186,17,0,0                 // addps         0x11ba(%rip),%xmm1        # 5c50 <_sk_callback_sse2+0x10f1>
-  .byte  68,15,40,13,194,17,0,0              // movaps        0x11c2(%rip),%xmm9        # 5c60 <_sk_callback_sse2+0x1101>
+  .byte  15,88,13,254,17,0,0                 // addps         0x11fe(%rip),%xmm1        # 5d20 <_sk_callback_sse2+0x1135>
+  .byte  68,15,40,13,6,18,0,0                // movaps        0x1206(%rip),%xmm9        # 5d30 <_sk_callback_sse2+0x1145>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,198,17,0,0               // movaps        0x11c6(%rip),%xmm8        # 5c70 <_sk_callback_sse2+0x1111>
+  .byte  68,15,40,5,10,18,0,0                // movaps        0x120a(%rip),%xmm8        # 5d40 <_sk_callback_sse2+0x1155>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,202,17,0,0               // addps         0x11ca(%rip),%xmm8        # 5c80 <_sk_callback_sse2+0x1121>
+  .byte  68,15,88,5,14,18,0,0                // addps         0x120e(%rip),%xmm8        # 5d50 <_sk_callback_sse2+0x1165>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,206,17,0,0               // addps         0x11ce(%rip),%xmm8        # 5c90 <_sk_callback_sse2+0x1131>
+  .byte  68,15,88,5,18,18,0,0                // addps         0x1212(%rip),%xmm8        # 5d60 <_sk_callback_sse2+0x1175>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,210,17,0,0               // addps         0x11d2(%rip),%xmm8        # 5ca0 <_sk_callback_sse2+0x1141>
+  .byte  68,15,88,5,22,18,0,0                // addps         0x1216(%rip),%xmm8        # 5d70 <_sk_callback_sse2+0x1185>
   .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
@@ -32215,17 +32489,17 @@ HIDDEN _sk_bicubic_p1y_sse2
 FUNCTION(_sk_bicubic_p1y_sse2)
 _sk_bicubic_p1y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,204,17,0,0               // movaps        0x11cc(%rip),%xmm8        # 5cb0 <_sk_callback_sse2+0x1151>
+  .byte  68,15,40,5,16,18,0,0                // movaps        0x1210(%rip),%xmm8        # 5d80 <_sk_callback_sse2+0x1195>
   .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  68,15,40,21,199,17,0,0              // movaps        0x11c7(%rip),%xmm10        # 5cc0 <_sk_callback_sse2+0x1161>
+  .byte  68,15,40,21,11,18,0,0               // movaps        0x120b(%rip),%xmm10        # 5d90 <_sk_callback_sse2+0x11a5>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,203,17,0,0              // addps         0x11cb(%rip),%xmm10        # 5cd0 <_sk_callback_sse2+0x1171>
+  .byte  68,15,88,21,15,18,0,0               // addps         0x120f(%rip),%xmm10        # 5da0 <_sk_callback_sse2+0x11b5>
   .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  68,15,88,21,199,17,0,0              // addps         0x11c7(%rip),%xmm10        # 5ce0 <_sk_callback_sse2+0x1181>
+  .byte  68,15,88,21,11,18,0,0               // addps         0x120b(%rip),%xmm10        # 5db0 <_sk_callback_sse2+0x11c5>
   .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
@@ -32237,11 +32511,11 @@ _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  15,88,13,185,17,0,0                 // addps         0x11b9(%rip),%xmm1        # 5cf0 <_sk_callback_sse2+0x1191>
+  .byte  15,88,13,253,17,0,0                 // addps         0x11fd(%rip),%xmm1        # 5dc0 <_sk_callback_sse2+0x11d5>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,185,17,0,0               // mulps         0x11b9(%rip),%xmm8        # 5d00 <_sk_callback_sse2+0x11a1>
-  .byte  68,15,88,5,193,17,0,0               // addps         0x11c1(%rip),%xmm8        # 5d10 <_sk_callback_sse2+0x11b1>
+  .byte  68,15,89,5,253,17,0,0               // mulps         0x11fd(%rip),%xmm8        # 5dd0 <_sk_callback_sse2+0x11e5>
+  .byte  68,15,88,5,5,18,0,0                 // addps         0x1205(%rip),%xmm8        # 5de0 <_sk_callback_sse2+0x11f5>
   .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
@@ -32460,11 +32734,11 @@ BALIGN16
   .byte  128,191,0,0,128,191,0               // cmpb          $0x0,-0x40800000(%rdi)
   .byte  0,224                               // add           %ah,%al
   .byte  64,0,0                              // add           %al,(%rax)
-  .byte  224,64                              // loopne        4e18 <.literal16+0x1d8>
+  .byte  224,64                              // loopne        4e98 <.literal16+0x1d8>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        4e1c <.literal16+0x1dc>
+  .byte  224,64                              // loopne        4e9c <.literal16+0x1dc>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        4e20 <.literal16+0x1e0>
+  .byte  224,64                              // loopne        4ea0 <.literal16+0x1e0>
   .byte  154                                 // (bad)
   .byte  153                                 // cltd
   .byte  153                                 // cltd
@@ -32484,13 +32758,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e41 <.literal16+0x201>
+  .byte  71,225,61                           // rex.RXB       loope 4ec1 <.literal16+0x201>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e45 <.literal16+0x205>
+  .byte  71,225,61                           // rex.RXB       loope 4ec5 <.literal16+0x205>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e49 <.literal16+0x209>
+  .byte  71,225,61                           // rex.RXB       loope 4ec9 <.literal16+0x209>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e4d <.literal16+0x20d>
+  .byte  71,225,61                           // rex.RXB       loope 4ecd <.literal16+0x20d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -32515,13 +32789,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e81 <.literal16+0x241>
+  .byte  71,225,61                           // rex.RXB       loope 4f01 <.literal16+0x241>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e85 <.literal16+0x245>
+  .byte  71,225,61                           // rex.RXB       loope 4f05 <.literal16+0x245>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e89 <.literal16+0x249>
+  .byte  71,225,61                           // rex.RXB       loope 4f09 <.literal16+0x249>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4e8d <.literal16+0x24d>
+  .byte  71,225,61                           // rex.RXB       loope 4f0d <.literal16+0x24d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -32546,13 +32820,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4ec1 <.literal16+0x281>
+  .byte  71,225,61                           // rex.RXB       loope 4f41 <.literal16+0x281>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4ec5 <.literal16+0x285>
+  .byte  71,225,61                           // rex.RXB       loope 4f45 <.literal16+0x285>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4ec9 <.literal16+0x289>
+  .byte  71,225,61                           // rex.RXB       loope 4f49 <.literal16+0x289>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4ecd <.literal16+0x28d>
+  .byte  71,225,61                           // rex.RXB       loope 4f4d <.literal16+0x28d>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -32577,13 +32851,13 @@ BALIGN16
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4f01 <.literal16+0x2c1>
+  .byte  71,225,61                           // rex.RXB       loope 4f81 <.literal16+0x2c1>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4f05 <.literal16+0x2c5>
+  .byte  71,225,61                           // rex.RXB       loope 4f85 <.literal16+0x2c5>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4f09 <.literal16+0x2c9>
+  .byte  71,225,61                           // rex.RXB       loope 4f89 <.literal16+0x2c9>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 4f0d <.literal16+0x2cd>
+  .byte  71,225,61                           // rex.RXB       loope 4f8d <.literal16+0x2cd>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -32804,13 +33078,13 @@ BALIGN16
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        50e9 <.literal16+0x4a9>
+  .byte  224,7                               // loopne        5169 <.literal16+0x4a9>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        50ed <.literal16+0x4ad>
+  .byte  224,7                               // loopne        516d <.literal16+0x4ad>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        50f1 <.literal16+0x4b1>
+  .byte  224,7                               // loopne        5171 <.literal16+0x4b1>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        50f5 <.literal16+0x4b5>
+  .byte  224,7                               // loopne        5175 <.literal16+0x4b5>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -32875,11 +33149,11 @@ BALIGN16
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,127,67                            // add           %bh,0x43(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            51cb <.literal16+0x58b>
+  .byte  127,67                              // jg            524b <.literal16+0x58b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            51cf <.literal16+0x58f>
+  .byte  127,67                              // jg            524f <.literal16+0x58f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            51d3 <.literal16+0x593>
+  .byte  127,67                              // jg            5253 <.literal16+0x593>
   .byte  129,128,128,59,129,128,128,59,129,128// addl          $0x80813b80,-0x7f7ec480(%rax)
   .byte  128,59,129                          // cmpb          $0x81,(%rbx)
   .byte  128,128,59,129,128,128,59           // addb          $0x3b,-0x7f7f7ec5(%rax)
@@ -32894,16 +33168,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            51c4 <.literal16+0x584>
+  .byte  127,0                               // jg            5244 <.literal16+0x584>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            51c8 <.literal16+0x588>
+  .byte  127,0                               // jg            5248 <.literal16+0x588>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            51cc <.literal16+0x58c>
+  .byte  127,0                               // jg            524c <.literal16+0x58c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            51d0 <.literal16+0x590>
+  .byte  127,0                               // jg            5250 <.literal16+0x590>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -32912,7 +33186,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            5255 <.literal16+0x615>
+  .byte  119,115                             // ja            52d5 <.literal16+0x615>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -32923,7 +33197,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           51b9 <.literal16+0x579>
+  .byte  117,191                             // jne           5239 <.literal16+0x579>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -32935,7 +33209,7 @@ 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          ffffffffe9a391fa <_sk_callback_sse2+0xffffffffe9a3469b>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3927a <_sk_callback_sse2+0xffffffffe9a3468f>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -32989,16 +33263,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5294 <.literal16+0x654>
+  .byte  127,0                               // jg            5314 <.literal16+0x654>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5298 <.literal16+0x658>
+  .byte  127,0                               // jg            5318 <.literal16+0x658>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            529c <.literal16+0x65c>
+  .byte  127,0                               // jg            531c <.literal16+0x65c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            52a0 <.literal16+0x660>
+  .byte  127,0                               // jg            5320 <.literal16+0x660>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -33007,7 +33281,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            5325 <.literal16+0x6e5>
+  .byte  119,115                             // ja            53a5 <.literal16+0x6e5>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -33018,7 +33292,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           5289 <.literal16+0x649>
+  .byte  117,191                             // jne           5309 <.literal16+0x649>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -33030,7 +33304,7 @@ 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          ffffffffe9a392ca <_sk_callback_sse2+0xffffffffe9a3476b>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3934a <_sk_callback_sse2+0xffffffffe9a3475f>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -33084,16 +33358,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5364 <.literal16+0x724>
+  .byte  127,0                               // jg            53e4 <.literal16+0x724>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5368 <.literal16+0x728>
+  .byte  127,0                               // jg            53e8 <.literal16+0x728>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            536c <.literal16+0x72c>
+  .byte  127,0                               // jg            53ec <.literal16+0x72c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5370 <.literal16+0x730>
+  .byte  127,0                               // jg            53f0 <.literal16+0x730>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -33102,7 +33376,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            53f5 <.literal16+0x7b5>
+  .byte  119,115                             // ja            5475 <.literal16+0x7b5>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -33113,7 +33387,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           5359 <.literal16+0x719>
+  .byte  117,191                             // jne           53d9 <.literal16+0x719>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -33125,7 +33399,7 @@ 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          ffffffffe9a3939a <_sk_callback_sse2+0xffffffffe9a3483b>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a3941a <_sk_callback_sse2+0xffffffffe9a3482f>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -33179,16 +33453,16 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5434 <.literal16+0x7f4>
+  .byte  127,0                               // jg            54b4 <.literal16+0x7f4>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5438 <.literal16+0x7f8>
+  .byte  127,0                               // jg            54b8 <.literal16+0x7f8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            543c <.literal16+0x7fc>
+  .byte  127,0                               // jg            54bc <.literal16+0x7fc>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            5440 <.literal16+0x800>
+  .byte  127,0                               // jg            54c0 <.literal16+0x800>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -33197,7 +33471,7 @@ BALIGN16
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,63                                // add           %bh,(%rdi)
-  .byte  119,115                             // ja            54c5 <.literal16+0x885>
+  .byte  119,115                             // ja            5545 <.literal16+0x885>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -33208,7 +33482,7 @@ BALIGN16
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           5429 <.literal16+0x7e9>
+  .byte  117,191                             // jne           54a9 <.literal16+0x7e9>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -33220,7 +33494,7 @@ 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          ffffffffe9a3946a <_sk_callback_sse2+0xffffffffe9a3490b>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a394ea <_sk_callback_sse2+0xffffffffe9a348ff>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
@@ -33270,13 +33544,13 @@ BALIGN16
   .byte  200,66,0,0                          // enterq        $0x42,$0x0
   .byte  200,66,0,0                          // enterq        $0x42,$0x0
   .byte  200,66,0,0                          // enterq        $0x42,$0x0
-  .byte  127,67                              // jg            5547 <.literal16+0x907>
+  .byte  127,67                              // jg            55c7 <.literal16+0x907>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            554b <.literal16+0x90b>
+  .byte  127,67                              // jg            55cb <.literal16+0x90b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            554f <.literal16+0x90f>
+  .byte  127,67                              // jg            55cf <.literal16+0x90f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            5553 <.literal16+0x913>
+  .byte  127,67                              // jg            55d3 <.literal16+0x913>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,195                               // add           %al,%bl
   .byte  0,0                                 // add           %al,(%rax)
@@ -33323,16 +33597,16 @@ BALIGN16
   .byte  128,3,62                            // addb          $0x3e,(%rbx)
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           55d3 <.literal16+0x993>
+  .byte  118,63                              // jbe           5653 <.literal16+0x993>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           55d7 <.literal16+0x997>
+  .byte  118,63                              // jbe           5657 <.literal16+0x997>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           55db <.literal16+0x99b>
+  .byte  118,63                              // jbe           565b <.literal16+0x99b>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           55df <.literal16+0x99f>
+  .byte  118,63                              // jbe           565f <.literal16+0x99f>
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
@@ -33344,11 +33618,11 @@ BALIGN16
   .byte  128,59,0                            // cmpb          $0x0,(%rbx)
   .byte  0,127,67                            // add           %bh,0x43(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            561b <.literal16+0x9db>
+  .byte  127,67                              // jg            569b <.literal16+0x9db>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            561f <.literal16+0x9df>
+  .byte  127,67                              // jg            569f <.literal16+0x9df>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            5623 <.literal16+0x9e3>
+  .byte  127,67                              // jg            56a3 <.literal16+0x9e3>
   .byte  129,128,128,59,129,128,128,59,129,128// addl          $0x80813b80,-0x7f7ec480(%rax)
   .byte  128,59,129                          // cmpb          $0x81,(%rbx)
   .byte  128,128,59,0,0,128,63               // addb          $0x3f,-0x7fffffc5(%rax)
@@ -33388,13 +33662,13 @@ BALIGN16
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        5669 <.literal16+0xa29>
+  .byte  224,7                               // loopne        56e9 <.literal16+0xa29>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        566d <.literal16+0xa2d>
+  .byte  224,7                               // loopne        56ed <.literal16+0xa2d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        5671 <.literal16+0xa31>
+  .byte  224,7                               // loopne        56f1 <.literal16+0xa31>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        5675 <.literal16+0xa35>
+  .byte  224,7                               // loopne        56f5 <.literal16+0xa35>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -33440,13 +33714,13 @@ BALIGN16
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        56d9 <.literal16+0xa99>
+  .byte  224,7                               // loopne        5759 <.literal16+0xa99>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        56dd <.literal16+0xa9d>
+  .byte  224,7                               // loopne        575d <.literal16+0xa9d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        56e1 <.literal16+0xaa1>
+  .byte  224,7                               // loopne        5761 <.literal16+0xaa1>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        56e5 <.literal16+0xaa5>
+  .byte  224,7                               // loopne        5765 <.literal16+0xaa5>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -33484,13 +33758,13 @@ BALIGN16
   .byte  65,0,0                              // add           %al,(%r8)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            5776 <.literal16+0xb36>
+  .byte  124,66                              // jl            57f6 <.literal16+0xb36>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            577a <.literal16+0xb3a>
+  .byte  124,66                              // jl            57fa <.literal16+0xb3a>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            577e <.literal16+0xb3e>
+  .byte  124,66                              // jl            57fe <.literal16+0xb3e>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            5782 <.literal16+0xb42>
+  .byte  124,66                              // jl            5802 <.literal16+0xb42>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,240                               // add           %dh,%al
@@ -33580,13 +33854,13 @@ BALIGN16
   .byte  136,136,61,137,136,136              // mov           %cl,-0x777776c3(%rax)
   .byte  61,137,136,136,61                   // cmp           $0x3d888889,%eax
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            5885 <.literal16+0xc45>
+  .byte  112,65                              // jo            5905 <.literal16+0xc45>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            5889 <.literal16+0xc49>
+  .byte  112,65                              // jo            5909 <.literal16+0xc49>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            588d <.literal16+0xc4d>
+  .byte  112,65                              // jo            590d <.literal16+0xc4d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            5891 <.literal16+0xc51>
+  .byte  112,65                              // jo            5911 <.literal16+0xc51>
   .byte  255,0                               // incl          (%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  255,0                               // incl          (%rax)
@@ -33608,11 +33882,11 @@ BALIGN16
   .byte  128,59,129                          // cmpb          $0x81,(%rbx)
   .byte  128,128,59,0,0,127,67               // addb          $0x43,0x7f00003b(%rax)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            58db <.literal16+0xc9b>
+  .byte  127,67                              // jg            595b <.literal16+0xc9b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            58df <.literal16+0xc9f>
+  .byte  127,67                              // jg            595f <.literal16+0xc9f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            58e3 <.literal16+0xca3>
+  .byte  127,67                              // jg            5963 <.literal16+0xca3>
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%rax)
@@ -33688,13 +33962,21 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  255                                 // (bad)
-  .byte  127,71                              // jg            59cb <.literal16+0xd8b>
+  .byte  127,71                              // jg            5a4b <.literal16+0xd8b>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            59cf <.literal16+0xd8f>
+  .byte  127,71                              // jg            5a4f <.literal16+0xd8f>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            59d3 <.literal16+0xd93>
+  .byte  127,71                              // jg            5a53 <.literal16+0xd93>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            59d7 <.literal16+0xd97>
+  .byte  127,71                              // jg            5a57 <.literal16+0xd97>
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
+  .byte  63                                  // (bad)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
+  .byte  63                                  // (bad)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,0                            // cmpb          $0x0,(%rdi)
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
@@ -33716,6 +33998,24 @@ BALIGN16
   .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
   .byte  63                                  // (bad)
   .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
+  .byte  63                                  // (bad)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,191,0,0,128,191,0               // cmpb          $0x0,-0x40800000(%rdi)
+  .byte  0,128,191,0,0,128                   // add           %al,-0x7fffff41(%rax)
+  .byte  191,0,0,0,63                        // mov           $0x3f000000,%edi
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,63                                // add           %bh,(%rdi)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,63                                // add           %bh,(%rdi)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,63                                // add           %bh,(%rdi)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  128,63,0                            // cmpb          $0x0,(%rdi)
+  .byte  0,128,63,0,0,128                    // add           %al,-0x7fffffc1(%rax)
+  .byte  63                                  // (bad)
+  .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,208                          // cmpb          $0xd0,(%rdi)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,208                              // ds            (bad)
@@ -33855,11 +34155,11 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,114                          // cmpb          $0x72,(%rdi)
   .byte  28,199                              // sbb           $0xc7,%al
-  .byte  62,114,28                           // jb,pt         5b42 <.literal16+0xf02>
+  .byte  62,114,28                           // jb,pt         5c12 <.literal16+0xf52>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5b46 <.literal16+0xf06>
+  .byte  62,114,28                           // jb,pt         5c16 <.literal16+0xf56>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5b4a <.literal16+0xf0a>
+  .byte  62,114,28                           // jb,pt         5c1a <.literal16+0xf5a>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -33903,7 +34203,7 @@ 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        # 3d63e9d5 <_sk_callback_sse2+0x3d639e76>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63eaa5 <_sk_callback_sse2+0x3d639eba>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -33929,7 +34229,7 @@ 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        # 3d63ea15 <_sk_callback_sse2+0x3d639eb6>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63eae5 <_sk_callback_sse2+0x3d639efa>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -33938,13 +34238,13 @@ BALIGN16
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
-  .byte  114,28                              // jb            5c0e <.literal16+0xfce>
+  .byte  114,28                              // jb            5cde <.literal16+0x101e>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c12 <.literal16+0xfd2>
+  .byte  62,114,28                           // jb,pt         5ce2 <.literal16+0x1022>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c16 <.literal16+0xfd6>
+  .byte  62,114,28                           // jb,pt         5ce6 <.literal16+0x1026>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c1a <.literal16+0xfda>
+  .byte  62,114,28                           // jb,pt         5cea <.literal16+0x102a>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -33965,11 +34265,11 @@ BALIGN16
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,63,114                          // cmpb          $0x72,(%rdi)
   .byte  28,199                              // sbb           $0xc7,%al
-  .byte  62,114,28                           // jb,pt         5c52 <.literal16+0x1012>
+  .byte  62,114,28                           // jb,pt         5d22 <.literal16+0x1062>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c56 <.literal16+0x1016>
+  .byte  62,114,28                           // jb,pt         5d26 <.literal16+0x1066>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5c5a <.literal16+0x101a>
+  .byte  62,114,28                           // jb,pt         5d2a <.literal16+0x106a>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
@@ -34013,7 +34313,7 @@ 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        # 3d63eae5 <_sk_callback_sse2+0x3d639f86>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63ebb5 <_sk_callback_sse2+0x3d639fca>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
@@ -34039,7 +34339,7 @@ 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        # 3d63eb25 <_sk_callback_sse2+0x3d639fc6>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d63ebf5 <_sk_callback_sse2+0x3d63a00a>
   .byte  57,142,99,61,0,0                    // cmp           %ecx,0x3d63(%rsi)
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
@@ -34048,13 +34348,13 @@ BALIGN16
   .byte  192,63,0                            // sarb          $0x0,(%rdi)
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
-  .byte  114,28                              // jb            5d1e <.literal16+0x10de>
+  .byte  114,28                              // jb            5dee <.literal16+0x112e>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5d22 <_sk_callback_sse2+0x11c3>
+  .byte  62,114,28                           // jb,pt         5df2 <_sk_callback_sse2+0x1207>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5d26 <_sk_callback_sse2+0x11c7>
+  .byte  62,114,28                           // jb,pt         5df6 <_sk_callback_sse2+0x120b>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         5d2a <_sk_callback_sse2+0x11cb>
+  .byte  62,114,28                           // jb,pt         5dfa <_sk_callback_sse2+0x120f>
   .byte  199                                 // (bad)
   .byte  62,171                              // ds            stos %eax,%es:(%rdi)
   .byte  170                                 // stos          %al,%es:(%rdi)
index 268cd26..d537c90 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  196,226,125,24,13,110,71,0,0        ; vbroadcastss  0x476e(%rip),%ymm1        # 48c8 <_sk_callback_hsw+0x11b>
+  DB  196,226,125,24,13,218,71,0,0        ; vbroadcastss  0x47da(%rip),%ymm1        # 4934 <_sk_callback_hsw+0x11b>
   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  196,226,125,24,21,82,71,0,0         ; vbroadcastss  0x4752(%rip),%ymm2        # 48cc <_sk_callback_hsw+0x11f>
+  DB  196,226,125,24,21,190,71,0,0        ; vbroadcastss  0x47be(%rip),%ymm2        # 4938 <_sk_callback_hsw+0x11f>
   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
@@ -130,13 +130,13 @@ _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  196,98,125,88,21,26,71,0,0          ; vpbroadcastd  0x471a(%rip),%ymm10        # 48d0 <_sk_callback_hsw+0x123>
+  DB  196,98,125,88,21,134,71,0,0         ; vpbroadcastd  0x4786(%rip),%ymm10        # 493c <_sk_callback_hsw+0x123>
   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  196,98,125,88,37,255,70,0,0         ; vpbroadcastd  0x46ff(%rip),%ymm12        # 48d4 <_sk_callback_hsw+0x127>
-  DB  196,98,125,88,45,250,70,0,0         ; vpbroadcastd  0x46fa(%rip),%ymm13        # 48d8 <_sk_callback_hsw+0x12b>
+  DB  196,98,125,88,37,107,71,0,0         ; vpbroadcastd  0x476b(%rip),%ymm12        # 4940 <_sk_callback_hsw+0x127>
+  DB  196,98,125,88,45,102,71,0,0         ; vpbroadcastd  0x4766(%rip),%ymm13        # 4944 <_sk_callback_hsw+0x12b>
   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
@@ -151,8 +151,8 @@ _sk_dither_hsw LABEL PROC
   DB  196,65,13,235,201                   ; vpor          %ymm9,%ymm14,%ymm9
   DB  196,65,61,235,193                   ; vpor          %ymm9,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,172,70,0,0         ; vbroadcastss  0x46ac(%rip),%ymm9        # 48dc <_sk_callback_hsw+0x12f>
-  DB  196,98,125,24,21,167,70,0,0         ; vbroadcastss  0x46a7(%rip),%ymm10        # 48e0 <_sk_callback_hsw+0x133>
+  DB  196,98,125,24,13,24,71,0,0          ; vbroadcastss  0x4718(%rip),%ymm9        # 4948 <_sk_callback_hsw+0x12f>
+  DB  196,98,125,24,21,19,71,0,0          ; vbroadcastss  0x4713(%rip),%ymm10        # 494c <_sk_callback_hsw+0x133>
   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
@@ -211,7 +211,7 @@ _sk_clear_hsw LABEL PROC
 PUBLIC _sk_srcatop_hsw
 _sk_srcatop_hsw LABEL PROC
   DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
-  DB  196,98,125,24,5,254,69,0,0          ; vbroadcastss  0x45fe(%rip),%ymm8        # 48e4 <_sk_callback_hsw+0x137>
+  DB  196,98,125,24,5,106,70,0,0          ; vbroadcastss  0x466a(%rip),%ymm8        # 4950 <_sk_callback_hsw+0x137>
   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
@@ -225,7 +225,7 @@ _sk_srcatop_hsw LABEL PROC
 
 PUBLIC _sk_dstatop_hsw
 _sk_dstatop_hsw LABEL PROC
-  DB  196,98,125,24,5,209,69,0,0          ; vbroadcastss  0x45d1(%rip),%ymm8        # 48e8 <_sk_callback_hsw+0x13b>
+  DB  196,98,125,24,5,61,70,0,0           ; vbroadcastss  0x463d(%rip),%ymm8        # 4954 <_sk_callback_hsw+0x13b>
   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
@@ -258,7 +258,7 @@ _sk_dstin_hsw LABEL PROC
 
 PUBLIC _sk_srcout_hsw
 _sk_srcout_hsw LABEL PROC
-  DB  196,98,125,24,5,120,69,0,0          ; vbroadcastss  0x4578(%rip),%ymm8        # 48ec <_sk_callback_hsw+0x13f>
+  DB  196,98,125,24,5,228,69,0,0          ; vbroadcastss  0x45e4(%rip),%ymm8        # 4958 <_sk_callback_hsw+0x13f>
   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
@@ -269,7 +269,7 @@ _sk_srcout_hsw LABEL PROC
 
 PUBLIC _sk_dstout_hsw
 _sk_dstout_hsw LABEL PROC
-  DB  196,226,125,24,5,91,69,0,0          ; vbroadcastss  0x455b(%rip),%ymm0        # 48f0 <_sk_callback_hsw+0x143>
+  DB  196,226,125,24,5,199,69,0,0         ; vbroadcastss  0x45c7(%rip),%ymm0        # 495c <_sk_callback_hsw+0x143>
   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
@@ -280,7 +280,7 @@ _sk_dstout_hsw LABEL PROC
 
 PUBLIC _sk_srcover_hsw
 _sk_srcover_hsw LABEL PROC
-  DB  196,98,125,24,5,62,69,0,0           ; vbroadcastss  0x453e(%rip),%ymm8        # 48f4 <_sk_callback_hsw+0x147>
+  DB  196,98,125,24,5,170,69,0,0          ; vbroadcastss  0x45aa(%rip),%ymm8        # 4960 <_sk_callback_hsw+0x147>
   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
@@ -291,7 +291,7 @@ _sk_srcover_hsw LABEL PROC
 
 PUBLIC _sk_dstover_hsw
 _sk_dstover_hsw LABEL PROC
-  DB  196,98,125,24,5,29,69,0,0           ; vbroadcastss  0x451d(%rip),%ymm8        # 48f8 <_sk_callback_hsw+0x14b>
+  DB  196,98,125,24,5,137,69,0,0          ; vbroadcastss  0x4589(%rip),%ymm8        # 4964 <_sk_callback_hsw+0x14b>
   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
@@ -311,7 +311,7 @@ _sk_modulate_hsw LABEL PROC
 
 PUBLIC _sk_multiply_hsw
 _sk_multiply_hsw LABEL PROC
-  DB  196,98,125,24,5,232,68,0,0          ; vbroadcastss  0x44e8(%rip),%ymm8        # 48fc <_sk_callback_hsw+0x14f>
+  DB  196,98,125,24,5,84,69,0,0           ; vbroadcastss  0x4554(%rip),%ymm8        # 4968 <_sk_callback_hsw+0x14f>
   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
@@ -353,7 +353,7 @@ _sk_screen_hsw LABEL PROC
 
 PUBLIC _sk_xor__hsw
 _sk_xor__hsw LABEL PROC
-  DB  196,98,125,24,5,99,68,0,0           ; vbroadcastss  0x4463(%rip),%ymm8        # 4900 <_sk_callback_hsw+0x153>
+  DB  196,98,125,24,5,207,68,0,0          ; vbroadcastss  0x44cf(%rip),%ymm8        # 496c <_sk_callback_hsw+0x153>
   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
@@ -385,7 +385,7 @@ _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  196,98,125,24,5,235,67,0,0          ; vbroadcastss  0x43eb(%rip),%ymm8        # 4904 <_sk_callback_hsw+0x157>
+  DB  196,98,125,24,5,87,68,0,0           ; vbroadcastss  0x4457(%rip),%ymm8        # 4970 <_sk_callback_hsw+0x157>
   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
@@ -408,7 +408,7 @@ _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  196,98,125,24,5,154,67,0,0          ; vbroadcastss  0x439a(%rip),%ymm8        # 4908 <_sk_callback_hsw+0x15b>
+  DB  196,98,125,24,5,6,68,0,0            ; vbroadcastss  0x4406(%rip),%ymm8        # 4974 <_sk_callback_hsw+0x15b>
   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
@@ -434,7 +434,7 @@ _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  196,98,125,24,5,61,67,0,0           ; vbroadcastss  0x433d(%rip),%ymm8        # 490c <_sk_callback_hsw+0x15f>
+  DB  196,98,125,24,5,169,67,0,0          ; vbroadcastss  0x43a9(%rip),%ymm8        # 4978 <_sk_callback_hsw+0x15f>
   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
@@ -454,7 +454,7 @@ _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  196,98,125,24,5,251,66,0,0          ; vbroadcastss  0x42fb(%rip),%ymm8        # 4910 <_sk_callback_hsw+0x163>
+  DB  196,98,125,24,5,103,67,0,0          ; vbroadcastss  0x4367(%rip),%ymm8        # 497c <_sk_callback_hsw+0x163>
   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
@@ -462,7 +462,7 @@ _sk_exclusion_hsw LABEL PROC
 
 PUBLIC _sk_colorburn_hsw
 _sk_colorburn_hsw LABEL PROC
-  DB  196,98,125,24,5,233,66,0,0          ; vbroadcastss  0x42e9(%rip),%ymm8        # 4914 <_sk_callback_hsw+0x167>
+  DB  196,98,125,24,5,85,67,0,0           ; vbroadcastss  0x4355(%rip),%ymm8        # 4980 <_sk_callback_hsw+0x167>
   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
@@ -518,7 +518,7 @@ _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  196,98,125,24,13,244,65,0,0         ; vbroadcastss  0x41f4(%rip),%ymm9        # 4918 <_sk_callback_hsw+0x16b>
+  DB  196,98,125,24,13,96,66,0,0          ; vbroadcastss  0x4260(%rip),%ymm9        # 4984 <_sk_callback_hsw+0x16b>
   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
@@ -569,7 +569,7 @@ _sk_colordodge_hsw LABEL PROC
 
 PUBLIC _sk_hardlight_hsw
 _sk_hardlight_hsw LABEL PROC
-  DB  196,98,125,24,5,21,65,0,0           ; vbroadcastss  0x4115(%rip),%ymm8        # 491c <_sk_callback_hsw+0x16f>
+  DB  196,98,125,24,5,129,65,0,0          ; vbroadcastss  0x4181(%rip),%ymm8        # 4988 <_sk_callback_hsw+0x16f>
   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
@@ -618,7 +618,7 @@ _sk_hardlight_hsw LABEL PROC
 
 PUBLIC _sk_overlay_hsw
 _sk_overlay_hsw LABEL PROC
-  DB  196,98,125,24,5,77,64,0,0           ; vbroadcastss  0x404d(%rip),%ymm8        # 4920 <_sk_callback_hsw+0x173>
+  DB  196,98,125,24,5,185,64,0,0          ; vbroadcastss  0x40b9(%rip),%ymm8        # 498c <_sk_callback_hsw+0x173>
   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
@@ -678,10 +678,10 @@ _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  196,98,125,24,29,84,63,0,0          ; vbroadcastss  0x3f54(%rip),%ymm11        # 4928 <_sk_callback_hsw+0x17b>
+  DB  196,98,125,24,29,192,63,0,0         ; vbroadcastss  0x3fc0(%rip),%ymm11        # 4994 <_sk_callback_hsw+0x17b>
   DB  196,65,20,88,227                    ; vaddps        %ymm11,%ymm13,%ymm12
   DB  196,65,28,89,192                    ; vmulps        %ymm8,%ymm12,%ymm8
-  DB  196,98,125,24,37,69,63,0,0          ; vbroadcastss  0x3f45(%rip),%ymm12        # 492c <_sk_callback_hsw+0x17f>
+  DB  196,98,125,24,37,177,63,0,0         ; vbroadcastss  0x3fb1(%rip),%ymm12        # 4998 <_sk_callback_hsw+0x17f>
   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
@@ -691,7 +691,7 @@ _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  196,98,125,24,5,8,63,0,0            ; vbroadcastss  0x3f08(%rip),%ymm8        # 4924 <_sk_callback_hsw+0x177>
+  DB  196,98,125,24,5,116,63,0,0          ; vbroadcastss  0x3f74(%rip),%ymm8        # 4990 <_sk_callback_hsw+0x177>
   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
@@ -804,11 +804,11 @@ _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  196,98,125,24,53,7,61,0,0           ; vbroadcastss  0x3d07(%rip),%ymm14        # 4930 <_sk_callback_hsw+0x183>
-  DB  196,98,125,24,61,2,61,0,0           ; vbroadcastss  0x3d02(%rip),%ymm15        # 4934 <_sk_callback_hsw+0x187>
+  DB  196,98,125,24,53,115,61,0,0         ; vbroadcastss  0x3d73(%rip),%ymm14        # 499c <_sk_callback_hsw+0x183>
+  DB  196,98,125,24,61,110,61,0,0         ; vbroadcastss  0x3d6e(%rip),%ymm15        # 49a0 <_sk_callback_hsw+0x187>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,66,93,184,238                   ; vfmadd231ps   %ymm14,%ymm4,%ymm13
-  DB  196,226,125,24,5,243,60,0,0         ; vbroadcastss  0x3cf3(%rip),%ymm0        # 4938 <_sk_callback_hsw+0x18b>
+  DB  196,226,125,24,5,95,61,0,0          ; vbroadcastss  0x3d5f(%rip),%ymm0        # 49a4 <_sk_callback_hsw+0x18b>
   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
@@ -863,7 +863,7 @@ _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  196,226,125,24,13,224,59,0,0        ; vbroadcastss  0x3be0(%rip),%ymm1        # 493c <_sk_callback_hsw+0x18f>
+  DB  196,226,125,24,13,76,60,0,0         ; vbroadcastss  0x3c4c(%rip),%ymm1        # 49a8 <_sk_callback_hsw+0x18f>
   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
@@ -917,11 +917,11 @@ _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  196,98,125,24,53,241,58,0,0         ; vbroadcastss  0x3af1(%rip),%ymm14        # 4940 <_sk_callback_hsw+0x193>
-  DB  196,98,125,24,61,236,58,0,0         ; vbroadcastss  0x3aec(%rip),%ymm15        # 4944 <_sk_callback_hsw+0x197>
+  DB  196,98,125,24,53,93,59,0,0          ; vbroadcastss  0x3b5d(%rip),%ymm14        # 49ac <_sk_callback_hsw+0x193>
+  DB  196,98,125,24,61,88,59,0,0          ; vbroadcastss  0x3b58(%rip),%ymm15        # 49b0 <_sk_callback_hsw+0x197>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,66,93,184,238                   ; vfmadd231ps   %ymm14,%ymm4,%ymm13
-  DB  196,226,125,24,5,221,58,0,0         ; vbroadcastss  0x3add(%rip),%ymm0        # 4948 <_sk_callback_hsw+0x19b>
+  DB  196,226,125,24,5,73,59,0,0          ; vbroadcastss  0x3b49(%rip),%ymm0        # 49b4 <_sk_callback_hsw+0x19b>
   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
@@ -976,7 +976,7 @@ _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  196,226,125,24,13,202,57,0,0        ; vbroadcastss  0x39ca(%rip),%ymm1        # 494c <_sk_callback_hsw+0x19f>
+  DB  196,226,125,24,13,54,58,0,0         ; vbroadcastss  0x3a36(%rip),%ymm1        # 49b8 <_sk_callback_hsw+0x19f>
   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
@@ -1004,11 +1004,11 @@ _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  196,98,125,24,45,93,57,0,0          ; vbroadcastss  0x395d(%rip),%ymm13        # 4950 <_sk_callback_hsw+0x1a3>
-  DB  196,98,125,24,53,88,57,0,0          ; vbroadcastss  0x3958(%rip),%ymm14        # 4954 <_sk_callback_hsw+0x1a7>
+  DB  196,98,125,24,45,201,57,0,0         ; vbroadcastss  0x39c9(%rip),%ymm13        # 49bc <_sk_callback_hsw+0x1a3>
+  DB  196,98,125,24,53,196,57,0,0         ; vbroadcastss  0x39c4(%rip),%ymm14        # 49c0 <_sk_callback_hsw+0x1a7>
   DB  196,65,84,89,230                    ; vmulps        %ymm14,%ymm5,%ymm12
   DB  196,66,93,184,229                   ; vfmadd231ps   %ymm13,%ymm4,%ymm12
-  DB  196,98,125,24,61,73,57,0,0          ; vbroadcastss  0x3949(%rip),%ymm15        # 4958 <_sk_callback_hsw+0x1ab>
+  DB  196,98,125,24,61,181,57,0,0         ; vbroadcastss  0x39b5(%rip),%ymm15        # 49c4 <_sk_callback_hsw+0x1ab>
   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
@@ -1064,7 +1064,7 @@ _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  196,226,125,24,5,43,56,0,0          ; vbroadcastss  0x382b(%rip),%ymm0        # 495c <_sk_callback_hsw+0x1af>
+  DB  196,226,125,24,5,151,56,0,0         ; vbroadcastss  0x3897(%rip),%ymm0        # 49c8 <_sk_callback_hsw+0x1af>
   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
@@ -1092,11 +1092,11 @@ _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  196,98,125,24,45,190,55,0,0         ; vbroadcastss  0x37be(%rip),%ymm13        # 4960 <_sk_callback_hsw+0x1b3>
-  DB  196,98,125,24,53,185,55,0,0         ; vbroadcastss  0x37b9(%rip),%ymm14        # 4964 <_sk_callback_hsw+0x1b7>
+  DB  196,98,125,24,45,42,56,0,0          ; vbroadcastss  0x382a(%rip),%ymm13        # 49cc <_sk_callback_hsw+0x1b3>
+  DB  196,98,125,24,53,37,56,0,0          ; vbroadcastss  0x3825(%rip),%ymm14        # 49d0 <_sk_callback_hsw+0x1b7>
   DB  196,65,116,89,230                   ; vmulps        %ymm14,%ymm1,%ymm12
   DB  196,66,109,184,229                  ; vfmadd231ps   %ymm13,%ymm2,%ymm12
-  DB  196,98,125,24,61,170,55,0,0         ; vbroadcastss  0x37aa(%rip),%ymm15        # 4968 <_sk_callback_hsw+0x1bb>
+  DB  196,98,125,24,61,22,56,0,0          ; vbroadcastss  0x3816(%rip),%ymm15        # 49d4 <_sk_callback_hsw+0x1bb>
   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
@@ -1152,7 +1152,7 @@ _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  196,226,125,24,5,140,54,0,0         ; vbroadcastss  0x368c(%rip),%ymm0        # 496c <_sk_callback_hsw+0x1bf>
+  DB  196,226,125,24,5,248,54,0,0         ; vbroadcastss  0x36f8(%rip),%ymm0        # 49d8 <_sk_callback_hsw+0x1bf>
   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
@@ -1182,7 +1182,7 @@ _sk_clamp_0_hsw LABEL PROC
 
 PUBLIC _sk_clamp_1_hsw
 _sk_clamp_1_hsw LABEL PROC
-  DB  196,98,125,24,5,37,54,0,0           ; vbroadcastss  0x3625(%rip),%ymm8        # 4970 <_sk_callback_hsw+0x1c3>
+  DB  196,98,125,24,5,145,54,0,0          ; vbroadcastss  0x3691(%rip),%ymm8        # 49dc <_sk_callback_hsw+0x1c3>
   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
@@ -1192,7 +1192,7 @@ _sk_clamp_1_hsw LABEL PROC
 
 PUBLIC _sk_clamp_a_hsw
 _sk_clamp_a_hsw LABEL PROC
-  DB  196,98,125,24,5,8,54,0,0            ; vbroadcastss  0x3608(%rip),%ymm8        # 4974 <_sk_callback_hsw+0x1c7>
+  DB  196,98,125,24,5,116,54,0,0          ; vbroadcastss  0x3674(%rip),%ymm8        # 49e0 <_sk_callback_hsw+0x1c7>
   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
@@ -1264,7 +1264,7 @@ 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  196,98,125,24,21,80,53,0,0          ; vbroadcastss  0x3550(%rip),%ymm10        # 4978 <_sk_callback_hsw+0x1cb>
+  DB  196,98,125,24,21,188,53,0,0         ; vbroadcastss  0x35bc(%rip),%ymm10        # 49e4 <_sk_callback_hsw+0x1cb>
   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
@@ -1275,16 +1275,16 @@ _sk_unpremul_hsw LABEL PROC
 
 PUBLIC _sk_from_srgb_hsw
 _sk_from_srgb_hsw LABEL PROC
-  DB  196,98,125,24,5,49,53,0,0           ; vbroadcastss  0x3531(%rip),%ymm8        # 497c <_sk_callback_hsw+0x1cf>
+  DB  196,98,125,24,5,157,53,0,0          ; vbroadcastss  0x359d(%rip),%ymm8        # 49e8 <_sk_callback_hsw+0x1cf>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
-  DB  196,98,125,24,29,35,53,0,0          ; vbroadcastss  0x3523(%rip),%ymm11        # 4980 <_sk_callback_hsw+0x1d3>
-  DB  196,98,125,24,37,30,53,0,0          ; vbroadcastss  0x351e(%rip),%ymm12        # 4984 <_sk_callback_hsw+0x1d7>
+  DB  196,98,125,24,29,143,53,0,0         ; vbroadcastss  0x358f(%rip),%ymm11        # 49ec <_sk_callback_hsw+0x1d3>
+  DB  196,98,125,24,37,138,53,0,0         ; vbroadcastss  0x358a(%rip),%ymm12        # 49f0 <_sk_callback_hsw+0x1d7>
   DB  196,65,124,40,236                   ; vmovaps       %ymm12,%ymm13
   DB  196,66,125,168,235                  ; vfmadd213ps   %ymm11,%ymm0,%ymm13
-  DB  196,98,125,24,53,15,53,0,0          ; vbroadcastss  0x350f(%rip),%ymm14        # 4988 <_sk_callback_hsw+0x1db>
+  DB  196,98,125,24,53,123,53,0,0         ; vbroadcastss  0x357b(%rip),%ymm14        # 49f4 <_sk_callback_hsw+0x1db>
   DB  196,66,45,168,238                   ; vfmadd213ps   %ymm14,%ymm10,%ymm13
-  DB  196,98,125,24,21,5,53,0,0           ; vbroadcastss  0x3505(%rip),%ymm10        # 498c <_sk_callback_hsw+0x1df>
+  DB  196,98,125,24,21,113,53,0,0         ; vbroadcastss  0x3571(%rip),%ymm10        # 49f8 <_sk_callback_hsw+0x1df>
   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
@@ -1306,19 +1306,19 @@ _sk_from_srgb_hsw LABEL PROC
 PUBLIC _sk_to_srgb_hsw
 _sk_to_srgb_hsw LABEL PROC
   DB  197,124,82,200                      ; vrsqrtps      %ymm0,%ymm9
-  DB  196,98,125,24,5,169,52,0,0          ; vbroadcastss  0x34a9(%rip),%ymm8        # 4990 <_sk_callback_hsw+0x1e3>
+  DB  196,98,125,24,5,21,53,0,0           ; vbroadcastss  0x3515(%rip),%ymm8        # 49fc <_sk_callback_hsw+0x1e3>
   DB  196,65,124,89,208                   ; vmulps        %ymm8,%ymm0,%ymm10
-  DB  196,98,125,24,29,159,52,0,0         ; vbroadcastss  0x349f(%rip),%ymm11        # 4994 <_sk_callback_hsw+0x1e7>
-  DB  196,98,125,24,37,154,52,0,0         ; vbroadcastss  0x349a(%rip),%ymm12        # 4998 <_sk_callback_hsw+0x1eb>
+  DB  196,98,125,24,29,11,53,0,0          ; vbroadcastss  0x350b(%rip),%ymm11        # 4a00 <_sk_callback_hsw+0x1e7>
+  DB  196,98,125,24,37,6,53,0,0           ; vbroadcastss  0x3506(%rip),%ymm12        # 4a04 <_sk_callback_hsw+0x1eb>
   DB  196,65,124,40,236                   ; vmovaps       %ymm12,%ymm13
   DB  196,66,53,168,235                   ; vfmadd213ps   %ymm11,%ymm9,%ymm13
-  DB  196,98,125,24,53,139,52,0,0         ; vbroadcastss  0x348b(%rip),%ymm14        # 499c <_sk_callback_hsw+0x1ef>
+  DB  196,98,125,24,53,247,52,0,0         ; vbroadcastss  0x34f7(%rip),%ymm14        # 4a08 <_sk_callback_hsw+0x1ef>
   DB  196,66,53,168,238                   ; vfmadd213ps   %ymm14,%ymm9,%ymm13
-  DB  196,98,125,24,61,129,52,0,0         ; vbroadcastss  0x3481(%rip),%ymm15        # 49a0 <_sk_callback_hsw+0x1f3>
+  DB  196,98,125,24,61,237,52,0,0         ; vbroadcastss  0x34ed(%rip),%ymm15        # 4a0c <_sk_callback_hsw+0x1f3>
   DB  196,65,52,88,207                    ; vaddps        %ymm15,%ymm9,%ymm9
   DB  196,65,124,83,201                   ; vrcpps        %ymm9,%ymm9
   DB  196,65,20,89,201                    ; vmulps        %ymm9,%ymm13,%ymm9
-  DB  196,98,125,24,45,109,52,0,0         ; vbroadcastss  0x346d(%rip),%ymm13        # 49a4 <_sk_callback_hsw+0x1f7>
+  DB  196,98,125,24,45,217,52,0,0         ; vbroadcastss  0x34d9(%rip),%ymm13        # 4a10 <_sk_callback_hsw+0x1f7>
   DB  196,193,124,194,197,1               ; vcmpltps      %ymm13,%ymm0,%ymm0
   DB  196,195,53,74,194,0                 ; vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
@@ -1350,26 +1350,26 @@ _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  196,98,125,24,29,226,51,0,0         ; vbroadcastss  0x33e2(%rip),%ymm11        # 49a8 <_sk_callback_hsw+0x1fb>
+  DB  196,98,125,24,29,78,52,0,0          ; vbroadcastss  0x344e(%rip),%ymm11        # 4a14 <_sk_callback_hsw+0x1fb>
   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  196,98,125,24,53,207,51,0,0         ; vbroadcastss  0x33cf(%rip),%ymm14        # 49ac <_sk_callback_hsw+0x1ff>
+  DB  196,98,125,24,53,59,52,0,0          ; vbroadcastss  0x343b(%rip),%ymm14        # 4a18 <_sk_callback_hsw+0x1ff>
   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  196,98,125,24,53,182,51,0,0         ; vbroadcastss  0x33b6(%rip),%ymm14        # 49b4 <_sk_callback_hsw+0x207>
+  DB  196,98,125,24,53,34,52,0,0          ; vbroadcastss  0x3422(%rip),%ymm14        # 4a20 <_sk_callback_hsw+0x207>
   DB  196,66,37,168,238                   ; vfmadd213ps   %ymm14,%ymm11,%ymm13
-  DB  196,98,125,24,53,164,51,0,0         ; vbroadcastss  0x33a4(%rip),%ymm14        # 49b0 <_sk_callback_hsw+0x203>
+  DB  196,98,125,24,53,16,52,0,0          ; vbroadcastss  0x3410(%rip),%ymm14        # 4a1c <_sk_callback_hsw+0x203>
   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  196,98,125,24,29,135,51,0,0         ; vbroadcastss  0x3387(%rip),%ymm11        # 49bc <_sk_callback_hsw+0x20f>
+  DB  196,98,125,24,29,243,51,0,0         ; vbroadcastss  0x33f3(%rip),%ymm11        # 4a28 <_sk_callback_hsw+0x20f>
   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
@@ -1379,7 +1379,7 @@ _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  196,98,125,24,5,74,51,0,0           ; vbroadcastss  0x334a(%rip),%ymm8        # 49b8 <_sk_callback_hsw+0x20b>
+  DB  196,98,125,24,5,182,51,0,0          ; vbroadcastss  0x33b6(%rip),%ymm8        # 4a24 <_sk_callback_hsw+0x20b>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -1394,30 +1394,30 @@ _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  196,98,125,24,5,17,51,0,0           ; vbroadcastss  0x3311(%rip),%ymm8        # 49c0 <_sk_callback_hsw+0x213>
+  DB  196,98,125,24,5,125,51,0,0          ; vbroadcastss  0x337d(%rip),%ymm8        # 4a2c <_sk_callback_hsw+0x213>
   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  196,98,125,24,13,244,50,0,0         ; vbroadcastss  0x32f4(%rip),%ymm9        # 49c4 <_sk_callback_hsw+0x217>
+  DB  196,98,125,24,13,96,51,0,0          ; vbroadcastss  0x3360(%rip),%ymm9        # 4a30 <_sk_callback_hsw+0x217>
   DB  196,66,109,170,202                  ; vfmsub213ps   %ymm10,%ymm2,%ymm9
-  DB  196,98,125,24,29,234,50,0,0         ; vbroadcastss  0x32ea(%rip),%ymm11        # 49c8 <_sk_callback_hsw+0x21b>
+  DB  196,98,125,24,29,86,51,0,0          ; vbroadcastss  0x3356(%rip),%ymm11        # 4a34 <_sk_callback_hsw+0x21b>
   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  196,98,125,24,45,212,50,0,0         ; vbroadcastss  0x32d4(%rip),%ymm13        # 49d0 <_sk_callback_hsw+0x223>
+  DB  196,98,125,24,45,64,51,0,0          ; vbroadcastss  0x3340(%rip),%ymm13        # 4a3c <_sk_callback_hsw+0x223>
   DB  196,193,4,89,197                    ; vmulps        %ymm13,%ymm15,%ymm0
-  DB  196,98,125,24,53,202,50,0,0         ; vbroadcastss  0x32ca(%rip),%ymm14        # 49d4 <_sk_callback_hsw+0x227>
+  DB  196,98,125,24,53,54,51,0,0          ; vbroadcastss  0x3336(%rip),%ymm14        # 4a40 <_sk_callback_hsw+0x227>
   DB  197,12,92,224                       ; vsubps        %ymm0,%ymm14,%ymm12
   DB  196,66,37,168,225                   ; vfmadd213ps   %ymm9,%ymm11,%ymm12
-  DB  196,226,125,24,29,176,50,0,0        ; vbroadcastss  0x32b0(%rip),%ymm3        # 49cc <_sk_callback_hsw+0x21f>
+  DB  196,226,125,24,29,28,51,0,0         ; vbroadcastss  0x331c(%rip),%ymm3        # 4a38 <_sk_callback_hsw+0x21f>
   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  196,98,125,24,37,155,50,0,0         ; vbroadcastss  0x329b(%rip),%ymm12        # 49d8 <_sk_callback_hsw+0x22b>
+  DB  196,98,125,24,37,7,51,0,0           ; vbroadcastss  0x3307(%rip),%ymm12        # 4a44 <_sk_callback_hsw+0x22b>
   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
@@ -1433,7 +1433,7 @@ _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  196,226,125,24,5,71,50,0,0          ; vbroadcastss  0x3247(%rip),%ymm0        # 49dc <_sk_callback_hsw+0x22f>
+  DB  196,226,125,24,5,179,50,0,0         ; vbroadcastss  0x32b3(%rip),%ymm0        # 4a48 <_sk_callback_hsw+0x22f>
   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
@@ -1483,7 +1483,7 @@ _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  196,98,125,24,13,129,49,0,0         ; vbroadcastss  0x3181(%rip),%ymm9        # 49e0 <_sk_callback_hsw+0x233>
+  DB  196,98,125,24,13,237,49,0,0         ; vbroadcastss  0x31ed(%rip),%ymm9        # 4a4c <_sk_callback_hsw+0x233>
   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
@@ -1531,7 +1531,7 @@ _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  196,98,125,24,13,238,48,0,0         ; vbroadcastss  0x30ee(%rip),%ymm9        # 49e4 <_sk_callback_hsw+0x237>
+  DB  196,98,125,24,13,90,49,0,0          ; vbroadcastss  0x315a(%rip),%ymm9        # 4a50 <_sk_callback_hsw+0x237>
   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
@@ -1565,20 +1565,20 @@ _sk_lerp_565_hsw LABEL PROC
   DB  15,133,169,0,0,0                    ; jne           1a02 <_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  196,98,125,88,13,123,48,0,0         ; vpbroadcastd  0x307b(%rip),%ymm9        # 49e8 <_sk_callback_hsw+0x23b>
+  DB  196,98,125,88,13,231,48,0,0         ; vpbroadcastd  0x30e7(%rip),%ymm9        # 4a54 <_sk_callback_hsw+0x23b>
   DB  196,65,61,219,201                   ; vpand         %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,108,48,0,0         ; vbroadcastss  0x306c(%rip),%ymm10        # 49ec <_sk_callback_hsw+0x23f>
+  DB  196,98,125,24,21,216,48,0,0         ; vbroadcastss  0x30d8(%rip),%ymm10        # 4a58 <_sk_callback_hsw+0x23f>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,88,21,98,48,0,0          ; vpbroadcastd  0x3062(%rip),%ymm10        # 49f0 <_sk_callback_hsw+0x243>
+  DB  196,98,125,88,21,206,48,0,0         ; vpbroadcastd  0x30ce(%rip),%ymm10        # 4a5c <_sk_callback_hsw+0x243>
   DB  196,65,61,219,210                   ; vpand         %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,83,48,0,0          ; vbroadcastss  0x3053(%rip),%ymm11        # 49f4 <_sk_callback_hsw+0x247>
+  DB  196,98,125,24,29,191,48,0,0         ; vbroadcastss  0x30bf(%rip),%ymm11        # 4a60 <_sk_callback_hsw+0x247>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,88,29,73,48,0,0          ; vpbroadcastd  0x3049(%rip),%ymm11        # 49f8 <_sk_callback_hsw+0x24b>
+  DB  196,98,125,88,29,181,48,0,0         ; vpbroadcastd  0x30b5(%rip),%ymm11        # 4a64 <_sk_callback_hsw+0x24b>
   DB  196,65,61,219,195                   ; vpand         %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,58,48,0,0          ; vbroadcastss  0x303a(%rip),%ymm11        # 49fc <_sk_callback_hsw+0x24f>
+  DB  196,98,125,24,29,166,48,0,0         ; vbroadcastss  0x30a6(%rip),%ymm11        # 4a68 <_sk_callback_hsw+0x24f>
   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
@@ -1648,21 +1648,21 @@ _sk_load_tables_hsw LABEL PROC
   DB  77,133,192                          ; test          %r8,%r8
   DB  117,105                             ; jne           1b0a <_sk_load_tables_hsw+0x7e>
   DB  196,193,126,111,25                  ; vmovdqu       (%r9),%ymm3
-  DB  197,229,219,13,18,50,0,0            ; vpand         0x3212(%rip),%ymm3,%ymm1        # 4cc0 <_sk_callback_hsw+0x513>
+  DB  197,229,219,13,114,50,0,0           ; vpand         0x3272(%rip),%ymm3,%ymm1        # 4d20 <_sk_callback_hsw+0x507>
   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  196,226,101,0,21,18,50,0,0          ; vpshufb       0x3212(%rip),%ymm3,%ymm2        # 4ce0 <_sk_callback_hsw+0x533>
+  DB  196,226,101,0,21,114,50,0,0         ; vpshufb       0x3272(%rip),%ymm3,%ymm2        # 4d40 <_sk_callback_hsw+0x527>
   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  196,98,101,0,13,26,50,0,0           ; vpshufb       0x321a(%rip),%ymm3,%ymm9        # 4d00 <_sk_callback_hsw+0x553>
+  DB  196,98,101,0,13,122,50,0,0          ; vpshufb       0x327a(%rip),%ymm3,%ymm9        # 4d60 <_sk_callback_hsw+0x547>
   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  196,98,125,24,5,2,47,0,0            ; vbroadcastss  0x2f02(%rip),%ymm8        # 4a00 <_sk_callback_hsw+0x253>
+  DB  196,98,125,24,5,110,47,0,0          ; vbroadcastss  0x2f6e(%rip),%ymm8        # 4a6c <_sk_callback_hsw+0x253>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,137,193                          ; mov           %r8,%rcx
@@ -1699,7 +1699,7 @@ _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  197,121,111,21,166,50,0,0           ; vmovdqa       0x32a6(%rip),%xmm10        # 4e40 <_sk_callback_hsw+0x693>
+  DB  197,121,111,21,6,51,0,0             ; vmovdqa       0x3306(%rip),%xmm10        # 4ea0 <_sk_callback_hsw+0x687>
   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
@@ -1721,7 +1721,7 @@ _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  196,98,125,24,5,251,45,0,0          ; vbroadcastss  0x2dfb(%rip),%ymm8        # 4a04 <_sk_callback_hsw+0x257>
+  DB  196,98,125,24,5,103,46,0,0          ; vbroadcastss  0x2e67(%rip),%ymm8        # 4a70 <_sk_callback_hsw+0x257>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -1779,7 +1779,7 @@ _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  197,121,111,13,70,49,0,0            ; vmovdqa       0x3146(%rip),%xmm9        # 4e50 <_sk_callback_hsw+0x6a3>
+  DB  197,121,111,13,166,49,0,0           ; vmovdqa       0x31a6(%rip),%xmm9        # 4eb0 <_sk_callback_hsw+0x697>
   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
@@ -1796,7 +1796,7 @@ _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  196,226,125,24,29,169,44,0,0        ; vbroadcastss  0x2ca9(%rip),%ymm3        # 4a08 <_sk_callback_hsw+0x25b>
+  DB  196,226,125,24,29,21,45,0,0         ; vbroadcastss  0x2d15(%rip),%ymm3        # 4a74 <_sk_callback_hsw+0x25b>
   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
@@ -1841,7 +1841,7 @@ _sk_byte_tables_hsw LABEL PROC
   DB  65,84                               ; push          %r12
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,231,43,0,0          ; vbroadcastss  0x2be7(%rip),%ymm8        # 4a0c <_sk_callback_hsw+0x25f>
+  DB  196,98,125,24,5,83,44,0,0           ; vbroadcastss  0x2c53(%rip),%ymm8        # 4a78 <_sk_callback_hsw+0x25f>
   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
@@ -1878,7 +1878,7 @@ _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  196,98,125,24,13,56,43,0,0          ; vbroadcastss  0x2b38(%rip),%ymm9        # 4a10 <_sk_callback_hsw+0x263>
+  DB  196,98,125,24,13,164,43,0,0         ; vbroadcastss  0x2ba4(%rip),%ymm9        # 4a7c <_sk_callback_hsw+0x263>
   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
@@ -2037,7 +2037,7 @@ _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  196,98,125,24,13,113,40,0,0         ; vbroadcastss  0x2871(%rip),%ymm9        # 4a14 <_sk_callback_hsw+0x267>
+  DB  196,98,125,24,13,221,40,0,0         ; vbroadcastss  0x28dd(%rip),%ymm9        # 4a80 <_sk_callback_hsw+0x267>
   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
@@ -2190,33 +2190,33 @@ _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  196,98,125,24,37,36,38,0,0          ; vbroadcastss  0x2624(%rip),%ymm12        # 4a18 <_sk_callback_hsw+0x26b>
-  DB  196,98,125,24,45,31,38,0,0          ; vbroadcastss  0x261f(%rip),%ymm13        # 4a1c <_sk_callback_hsw+0x26f>
+  DB  196,98,125,24,37,144,38,0,0         ; vbroadcastss  0x2690(%rip),%ymm12        # 4a84 <_sk_callback_hsw+0x26b>
+  DB  196,98,125,24,45,139,38,0,0         ; vbroadcastss  0x268b(%rip),%ymm13        # 4a88 <_sk_callback_hsw+0x26f>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,21,38,0,0          ; vbroadcastss  0x2615(%rip),%ymm13        # 4a20 <_sk_callback_hsw+0x273>
+  DB  196,98,125,24,45,129,38,0,0         ; vbroadcastss  0x2681(%rip),%ymm13        # 4a8c <_sk_callback_hsw+0x273>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,11,38,0,0          ; vbroadcastss  0x260b(%rip),%ymm13        # 4a24 <_sk_callback_hsw+0x277>
+  DB  196,98,125,24,45,119,38,0,0         ; vbroadcastss  0x2677(%rip),%ymm13        # 4a90 <_sk_callback_hsw+0x277>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,1,38,0,0           ; vbroadcastss  0x2601(%rip),%ymm11        # 4a28 <_sk_callback_hsw+0x27b>
+  DB  196,98,125,24,29,109,38,0,0         ; vbroadcastss  0x266d(%rip),%ymm11        # 4a94 <_sk_callback_hsw+0x27b>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,247,37,0,0         ; vbroadcastss  0x25f7(%rip),%ymm12        # 4a2c <_sk_callback_hsw+0x27f>
+  DB  196,98,125,24,37,99,38,0,0          ; vbroadcastss  0x2663(%rip),%ymm12        # 4a98 <_sk_callback_hsw+0x27f>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,237,37,0,0         ; vbroadcastss  0x25ed(%rip),%ymm12        # 4a30 <_sk_callback_hsw+0x283>
+  DB  196,98,125,24,37,89,38,0,0          ; vbroadcastss  0x2659(%rip),%ymm12        # 4a9c <_sk_callback_hsw+0x283>
   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  196,98,125,24,29,206,37,0,0         ; vbroadcastss  0x25ce(%rip),%ymm11        # 4a34 <_sk_callback_hsw+0x287>
+  DB  196,98,125,24,29,58,38,0,0          ; vbroadcastss  0x263a(%rip),%ymm11        # 4aa0 <_sk_callback_hsw+0x287>
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,196,37,0,0         ; vbroadcastss  0x25c4(%rip),%ymm11        # 4a38 <_sk_callback_hsw+0x28b>
+  DB  196,98,125,24,29,48,38,0,0          ; vbroadcastss  0x2630(%rip),%ymm11        # 4aa4 <_sk_callback_hsw+0x28b>
   DB  196,98,45,172,216                   ; vfnmadd213ps  %ymm0,%ymm10,%ymm11
-  DB  196,226,125,24,5,186,37,0,0         ; vbroadcastss  0x25ba(%rip),%ymm0        # 4a3c <_sk_callback_hsw+0x28f>
+  DB  196,226,125,24,5,38,38,0,0          ; vbroadcastss  0x2626(%rip),%ymm0        # 4aa8 <_sk_callback_hsw+0x28f>
   DB  196,193,124,92,194                  ; vsubps        %ymm10,%ymm0,%ymm0
-  DB  196,98,125,24,21,176,37,0,0         ; vbroadcastss  0x25b0(%rip),%ymm10        # 4a40 <_sk_callback_hsw+0x293>
+  DB  196,98,125,24,21,28,38,0,0          ; vbroadcastss  0x261c(%rip),%ymm10        # 4aac <_sk_callback_hsw+0x293>
   DB  197,172,94,192                      ; vdivps        %ymm0,%ymm10,%ymm0
   DB  197,164,88,192                      ; vaddps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,21,163,37,0,0         ; vbroadcastss  0x25a3(%rip),%ymm10        # 4a44 <_sk_callback_hsw+0x297>
+  DB  196,98,125,24,21,15,38,0,0          ; vbroadcastss  0x260f(%rip),%ymm10        # 4ab0 <_sk_callback_hsw+0x297>
   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
@@ -2224,7 +2224,7 @@ _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  196,98,125,24,5,122,37,0,0          ; vbroadcastss  0x257a(%rip),%ymm8        # 4a48 <_sk_callback_hsw+0x29b>
+  DB  196,98,125,24,5,230,37,0,0          ; vbroadcastss  0x25e6(%rip),%ymm8        # 4ab4 <_sk_callback_hsw+0x29b>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2242,33 +2242,33 @@ _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  196,98,125,24,37,50,37,0,0          ; vbroadcastss  0x2532(%rip),%ymm12        # 4a4c <_sk_callback_hsw+0x29f>
-  DB  196,98,125,24,45,45,37,0,0          ; vbroadcastss  0x252d(%rip),%ymm13        # 4a50 <_sk_callback_hsw+0x2a3>
+  DB  196,98,125,24,37,158,37,0,0         ; vbroadcastss  0x259e(%rip),%ymm12        # 4ab8 <_sk_callback_hsw+0x29f>
+  DB  196,98,125,24,45,153,37,0,0         ; vbroadcastss  0x2599(%rip),%ymm13        # 4abc <_sk_callback_hsw+0x2a3>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,35,37,0,0          ; vbroadcastss  0x2523(%rip),%ymm13        # 4a54 <_sk_callback_hsw+0x2a7>
+  DB  196,98,125,24,45,143,37,0,0         ; vbroadcastss  0x258f(%rip),%ymm13        # 4ac0 <_sk_callback_hsw+0x2a7>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,25,37,0,0          ; vbroadcastss  0x2519(%rip),%ymm13        # 4a58 <_sk_callback_hsw+0x2ab>
+  DB  196,98,125,24,45,133,37,0,0         ; vbroadcastss  0x2585(%rip),%ymm13        # 4ac4 <_sk_callback_hsw+0x2ab>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,15,37,0,0          ; vbroadcastss  0x250f(%rip),%ymm11        # 4a5c <_sk_callback_hsw+0x2af>
+  DB  196,98,125,24,29,123,37,0,0         ; vbroadcastss  0x257b(%rip),%ymm11        # 4ac8 <_sk_callback_hsw+0x2af>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,5,37,0,0           ; vbroadcastss  0x2505(%rip),%ymm12        # 4a60 <_sk_callback_hsw+0x2b3>
+  DB  196,98,125,24,37,113,37,0,0         ; vbroadcastss  0x2571(%rip),%ymm12        # 4acc <_sk_callback_hsw+0x2b3>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,251,36,0,0         ; vbroadcastss  0x24fb(%rip),%ymm12        # 4a64 <_sk_callback_hsw+0x2b7>
+  DB  196,98,125,24,37,103,37,0,0         ; vbroadcastss  0x2567(%rip),%ymm12        # 4ad0 <_sk_callback_hsw+0x2b7>
   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  196,98,125,24,29,220,36,0,0         ; vbroadcastss  0x24dc(%rip),%ymm11        # 4a68 <_sk_callback_hsw+0x2bb>
+  DB  196,98,125,24,29,72,37,0,0          ; vbroadcastss  0x2548(%rip),%ymm11        # 4ad4 <_sk_callback_hsw+0x2bb>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,210,36,0,0         ; vbroadcastss  0x24d2(%rip),%ymm11        # 4a6c <_sk_callback_hsw+0x2bf>
+  DB  196,98,125,24,29,62,37,0,0          ; vbroadcastss  0x253e(%rip),%ymm11        # 4ad8 <_sk_callback_hsw+0x2bf>
   DB  196,98,45,172,217                   ; vfnmadd213ps  %ymm1,%ymm10,%ymm11
-  DB  196,226,125,24,13,200,36,0,0        ; vbroadcastss  0x24c8(%rip),%ymm1        # 4a70 <_sk_callback_hsw+0x2c3>
+  DB  196,226,125,24,13,52,37,0,0         ; vbroadcastss  0x2534(%rip),%ymm1        # 4adc <_sk_callback_hsw+0x2c3>
   DB  196,193,116,92,202                  ; vsubps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,21,190,36,0,0         ; vbroadcastss  0x24be(%rip),%ymm10        # 4a74 <_sk_callback_hsw+0x2c7>
+  DB  196,98,125,24,21,42,37,0,0          ; vbroadcastss  0x252a(%rip),%ymm10        # 4ae0 <_sk_callback_hsw+0x2c7>
   DB  197,172,94,201                      ; vdivps        %ymm1,%ymm10,%ymm1
   DB  197,164,88,201                      ; vaddps        %ymm1,%ymm11,%ymm1
-  DB  196,98,125,24,21,177,36,0,0         ; vbroadcastss  0x24b1(%rip),%ymm10        # 4a78 <_sk_callback_hsw+0x2cb>
+  DB  196,98,125,24,21,29,37,0,0          ; vbroadcastss  0x251d(%rip),%ymm10        # 4ae4 <_sk_callback_hsw+0x2cb>
   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
@@ -2276,7 +2276,7 @@ _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  196,98,125,24,5,136,36,0,0          ; vbroadcastss  0x2488(%rip),%ymm8        # 4a7c <_sk_callback_hsw+0x2cf>
+  DB  196,98,125,24,5,244,36,0,0          ; vbroadcastss  0x24f4(%rip),%ymm8        # 4ae8 <_sk_callback_hsw+0x2cf>
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2294,33 +2294,33 @@ _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  196,98,125,24,37,64,36,0,0          ; vbroadcastss  0x2440(%rip),%ymm12        # 4a80 <_sk_callback_hsw+0x2d3>
-  DB  196,98,125,24,45,59,36,0,0          ; vbroadcastss  0x243b(%rip),%ymm13        # 4a84 <_sk_callback_hsw+0x2d7>
+  DB  196,98,125,24,37,172,36,0,0         ; vbroadcastss  0x24ac(%rip),%ymm12        # 4aec <_sk_callback_hsw+0x2d3>
+  DB  196,98,125,24,45,167,36,0,0         ; vbroadcastss  0x24a7(%rip),%ymm13        # 4af0 <_sk_callback_hsw+0x2d7>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,49,36,0,0          ; vbroadcastss  0x2431(%rip),%ymm13        # 4a88 <_sk_callback_hsw+0x2db>
+  DB  196,98,125,24,45,157,36,0,0         ; vbroadcastss  0x249d(%rip),%ymm13        # 4af4 <_sk_callback_hsw+0x2db>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,39,36,0,0          ; vbroadcastss  0x2427(%rip),%ymm13        # 4a8c <_sk_callback_hsw+0x2df>
+  DB  196,98,125,24,45,147,36,0,0         ; vbroadcastss  0x2493(%rip),%ymm13        # 4af8 <_sk_callback_hsw+0x2df>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,29,36,0,0          ; vbroadcastss  0x241d(%rip),%ymm11        # 4a90 <_sk_callback_hsw+0x2e3>
+  DB  196,98,125,24,29,137,36,0,0         ; vbroadcastss  0x2489(%rip),%ymm11        # 4afc <_sk_callback_hsw+0x2e3>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,19,36,0,0          ; vbroadcastss  0x2413(%rip),%ymm12        # 4a94 <_sk_callback_hsw+0x2e7>
+  DB  196,98,125,24,37,127,36,0,0         ; vbroadcastss  0x247f(%rip),%ymm12        # 4b00 <_sk_callback_hsw+0x2e7>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,9,36,0,0           ; vbroadcastss  0x2409(%rip),%ymm12        # 4a98 <_sk_callback_hsw+0x2eb>
+  DB  196,98,125,24,37,117,36,0,0         ; vbroadcastss  0x2475(%rip),%ymm12        # 4b04 <_sk_callback_hsw+0x2eb>
   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  196,98,125,24,29,234,35,0,0         ; vbroadcastss  0x23ea(%rip),%ymm11        # 4a9c <_sk_callback_hsw+0x2ef>
+  DB  196,98,125,24,29,86,36,0,0          ; vbroadcastss  0x2456(%rip),%ymm11        # 4b08 <_sk_callback_hsw+0x2ef>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,224,35,0,0         ; vbroadcastss  0x23e0(%rip),%ymm11        # 4aa0 <_sk_callback_hsw+0x2f3>
+  DB  196,98,125,24,29,76,36,0,0          ; vbroadcastss  0x244c(%rip),%ymm11        # 4b0c <_sk_callback_hsw+0x2f3>
   DB  196,98,45,172,218                   ; vfnmadd213ps  %ymm2,%ymm10,%ymm11
-  DB  196,226,125,24,21,214,35,0,0        ; vbroadcastss  0x23d6(%rip),%ymm2        # 4aa4 <_sk_callback_hsw+0x2f7>
+  DB  196,226,125,24,21,66,36,0,0         ; vbroadcastss  0x2442(%rip),%ymm2        # 4b10 <_sk_callback_hsw+0x2f7>
   DB  196,193,108,92,210                  ; vsubps        %ymm10,%ymm2,%ymm2
-  DB  196,98,125,24,21,204,35,0,0         ; vbroadcastss  0x23cc(%rip),%ymm10        # 4aa8 <_sk_callback_hsw+0x2fb>
+  DB  196,98,125,24,21,56,36,0,0          ; vbroadcastss  0x2438(%rip),%ymm10        # 4b14 <_sk_callback_hsw+0x2fb>
   DB  197,172,94,210                      ; vdivps        %ymm2,%ymm10,%ymm2
   DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
-  DB  196,98,125,24,21,191,35,0,0         ; vbroadcastss  0x23bf(%rip),%ymm10        # 4aac <_sk_callback_hsw+0x2ff>
+  DB  196,98,125,24,21,43,36,0,0          ; vbroadcastss  0x242b(%rip),%ymm10        # 4b18 <_sk_callback_hsw+0x2ff>
   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
@@ -2328,7 +2328,7 @@ _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  196,98,125,24,5,150,35,0,0          ; vbroadcastss  0x2396(%rip),%ymm8        # 4ab0 <_sk_callback_hsw+0x303>
+  DB  196,98,125,24,5,2,36,0,0            ; vbroadcastss  0x2402(%rip),%ymm8        # 4b1c <_sk_callback_hsw+0x303>
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2346,33 +2346,33 @@ _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  196,98,125,24,37,78,35,0,0          ; vbroadcastss  0x234e(%rip),%ymm12        # 4ab4 <_sk_callback_hsw+0x307>
-  DB  196,98,125,24,45,73,35,0,0          ; vbroadcastss  0x2349(%rip),%ymm13        # 4ab8 <_sk_callback_hsw+0x30b>
+  DB  196,98,125,24,37,186,35,0,0         ; vbroadcastss  0x23ba(%rip),%ymm12        # 4b20 <_sk_callback_hsw+0x307>
+  DB  196,98,125,24,45,181,35,0,0         ; vbroadcastss  0x23b5(%rip),%ymm13        # 4b24 <_sk_callback_hsw+0x30b>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,63,35,0,0          ; vbroadcastss  0x233f(%rip),%ymm13        # 4abc <_sk_callback_hsw+0x30f>
+  DB  196,98,125,24,45,171,35,0,0         ; vbroadcastss  0x23ab(%rip),%ymm13        # 4b28 <_sk_callback_hsw+0x30f>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,53,35,0,0          ; vbroadcastss  0x2335(%rip),%ymm13        # 4ac0 <_sk_callback_hsw+0x313>
+  DB  196,98,125,24,45,161,35,0,0         ; vbroadcastss  0x23a1(%rip),%ymm13        # 4b2c <_sk_callback_hsw+0x313>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,43,35,0,0          ; vbroadcastss  0x232b(%rip),%ymm11        # 4ac4 <_sk_callback_hsw+0x317>
+  DB  196,98,125,24,29,151,35,0,0         ; vbroadcastss  0x2397(%rip),%ymm11        # 4b30 <_sk_callback_hsw+0x317>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,33,35,0,0          ; vbroadcastss  0x2321(%rip),%ymm12        # 4ac8 <_sk_callback_hsw+0x31b>
+  DB  196,98,125,24,37,141,35,0,0         ; vbroadcastss  0x238d(%rip),%ymm12        # 4b34 <_sk_callback_hsw+0x31b>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,23,35,0,0          ; vbroadcastss  0x2317(%rip),%ymm12        # 4acc <_sk_callback_hsw+0x31f>
+  DB  196,98,125,24,37,131,35,0,0         ; vbroadcastss  0x2383(%rip),%ymm12        # 4b38 <_sk_callback_hsw+0x31f>
   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  196,98,125,24,29,248,34,0,0         ; vbroadcastss  0x22f8(%rip),%ymm11        # 4ad0 <_sk_callback_hsw+0x323>
+  DB  196,98,125,24,29,100,35,0,0         ; vbroadcastss  0x2364(%rip),%ymm11        # 4b3c <_sk_callback_hsw+0x323>
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,238,34,0,0         ; vbroadcastss  0x22ee(%rip),%ymm11        # 4ad4 <_sk_callback_hsw+0x327>
+  DB  196,98,125,24,29,90,35,0,0          ; vbroadcastss  0x235a(%rip),%ymm11        # 4b40 <_sk_callback_hsw+0x327>
   DB  196,98,45,172,219                   ; vfnmadd213ps  %ymm3,%ymm10,%ymm11
-  DB  196,226,125,24,29,228,34,0,0        ; vbroadcastss  0x22e4(%rip),%ymm3        # 4ad8 <_sk_callback_hsw+0x32b>
+  DB  196,226,125,24,29,80,35,0,0         ; vbroadcastss  0x2350(%rip),%ymm3        # 4b44 <_sk_callback_hsw+0x32b>
   DB  196,193,100,92,218                  ; vsubps        %ymm10,%ymm3,%ymm3
-  DB  196,98,125,24,21,218,34,0,0         ; vbroadcastss  0x22da(%rip),%ymm10        # 4adc <_sk_callback_hsw+0x32f>
+  DB  196,98,125,24,21,70,35,0,0          ; vbroadcastss  0x2346(%rip),%ymm10        # 4b48 <_sk_callback_hsw+0x32f>
   DB  197,172,94,219                      ; vdivps        %ymm3,%ymm10,%ymm3
   DB  197,164,88,219                      ; vaddps        %ymm3,%ymm11,%ymm3
-  DB  196,98,125,24,21,205,34,0,0         ; vbroadcastss  0x22cd(%rip),%ymm10        # 4ae0 <_sk_callback_hsw+0x333>
+  DB  196,98,125,24,21,57,35,0,0          ; vbroadcastss  0x2339(%rip),%ymm10        # 4b4c <_sk_callback_hsw+0x333>
   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
@@ -2380,33 +2380,33 @@ _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  196,98,125,24,5,164,34,0,0          ; vbroadcastss  0x22a4(%rip),%ymm8        # 4ae4 <_sk_callback_hsw+0x337>
+  DB  196,98,125,24,5,16,35,0,0           ; vbroadcastss  0x2310(%rip),%ymm8        # 4b50 <_sk_callback_hsw+0x337>
   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  196,98,125,24,5,150,34,0,0          ; vbroadcastss  0x2296(%rip),%ymm8        # 4ae8 <_sk_callback_hsw+0x33b>
-  DB  196,98,125,24,13,145,34,0,0         ; vbroadcastss  0x2291(%rip),%ymm9        # 4aec <_sk_callback_hsw+0x33f>
-  DB  196,98,125,24,21,140,34,0,0         ; vbroadcastss  0x228c(%rip),%ymm10        # 4af0 <_sk_callback_hsw+0x343>
+  DB  196,98,125,24,5,2,35,0,0            ; vbroadcastss  0x2302(%rip),%ymm8        # 4b54 <_sk_callback_hsw+0x33b>
+  DB  196,98,125,24,13,253,34,0,0         ; vbroadcastss  0x22fd(%rip),%ymm9        # 4b58 <_sk_callback_hsw+0x33f>
+  DB  196,98,125,24,21,248,34,0,0         ; vbroadcastss  0x22f8(%rip),%ymm10        # 4b5c <_sk_callback_hsw+0x343>
   DB  196,194,53,168,202                  ; vfmadd213ps   %ymm10,%ymm9,%ymm1
   DB  196,194,53,168,210                  ; vfmadd213ps   %ymm10,%ymm9,%ymm2
-  DB  196,98,125,24,13,125,34,0,0         ; vbroadcastss  0x227d(%rip),%ymm9        # 4af4 <_sk_callback_hsw+0x347>
+  DB  196,98,125,24,13,233,34,0,0         ; vbroadcastss  0x22e9(%rip),%ymm9        # 4b60 <_sk_callback_hsw+0x347>
   DB  196,66,125,184,200                  ; vfmadd231ps   %ymm8,%ymm0,%ymm9
-  DB  196,226,125,24,5,115,34,0,0         ; vbroadcastss  0x2273(%rip),%ymm0        # 4af8 <_sk_callback_hsw+0x34b>
+  DB  196,226,125,24,5,223,34,0,0         ; vbroadcastss  0x22df(%rip),%ymm0        # 4b64 <_sk_callback_hsw+0x34b>
   DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
-  DB  196,98,125,24,5,106,34,0,0          ; vbroadcastss  0x226a(%rip),%ymm8        # 4afc <_sk_callback_hsw+0x34f>
+  DB  196,98,125,24,5,214,34,0,0          ; vbroadcastss  0x22d6(%rip),%ymm8        # 4b68 <_sk_callback_hsw+0x34f>
   DB  196,98,117,168,192                  ; vfmadd213ps   %ymm0,%ymm1,%ymm8
-  DB  196,98,125,24,13,96,34,0,0          ; vbroadcastss  0x2260(%rip),%ymm9        # 4b00 <_sk_callback_hsw+0x353>
+  DB  196,98,125,24,13,204,34,0,0         ; vbroadcastss  0x22cc(%rip),%ymm9        # 4b6c <_sk_callback_hsw+0x353>
   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  196,226,125,24,21,77,34,0,0         ; vbroadcastss  0x224d(%rip),%ymm2        # 4b04 <_sk_callback_hsw+0x357>
+  DB  196,226,125,24,21,185,34,0,0        ; vbroadcastss  0x22b9(%rip),%ymm2        # 4b70 <_sk_callback_hsw+0x357>
   DB  197,108,194,209,1                   ; vcmpltps      %ymm1,%ymm2,%ymm10
-  DB  196,98,125,24,29,67,34,0,0          ; vbroadcastss  0x2243(%rip),%ymm11        # 4b08 <_sk_callback_hsw+0x35b>
+  DB  196,98,125,24,29,175,34,0,0         ; vbroadcastss  0x22af(%rip),%ymm11        # 4b74 <_sk_callback_hsw+0x35b>
   DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
-  DB  196,98,125,24,37,57,34,0,0          ; vbroadcastss  0x2239(%rip),%ymm12        # 4b0c <_sk_callback_hsw+0x35f>
+  DB  196,98,125,24,37,165,34,0,0         ; vbroadcastss  0x22a5(%rip),%ymm12        # 4b78 <_sk_callback_hsw+0x35f>
   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
@@ -2421,9 +2421,9 @@ _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  196,226,125,24,5,238,33,0,0         ; vbroadcastss  0x21ee(%rip),%ymm0        # 4b10 <_sk_callback_hsw+0x363>
+  DB  196,226,125,24,5,90,34,0,0          ; vbroadcastss  0x225a(%rip),%ymm0        # 4b7c <_sk_callback_hsw+0x363>
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,229,33,0,0          ; vbroadcastss  0x21e5(%rip),%ymm8        # 4b14 <_sk_callback_hsw+0x367>
+  DB  196,98,125,24,5,81,34,0,0           ; vbroadcastss  0x2251(%rip),%ymm8        # 4b80 <_sk_callback_hsw+0x367>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2439,7 +2439,7 @@ _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  196,226,125,24,13,186,33,0,0        ; vbroadcastss  0x21ba(%rip),%ymm1        # 4b18 <_sk_callback_hsw+0x36b>
+  DB  196,226,125,24,13,38,34,0,0         ; vbroadcastss  0x2226(%rip),%ymm1        # 4b84 <_sk_callback_hsw+0x36b>
   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
@@ -2502,7 +2502,7 @@ _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  196,226,125,24,13,197,32,0,0        ; vbroadcastss  0x20c5(%rip),%ymm1        # 4b1c <_sk_callback_hsw+0x36f>
+  DB  196,226,125,24,13,49,33,0,0         ; vbroadcastss  0x2131(%rip),%ymm1        # 4b88 <_sk_callback_hsw+0x36f>
   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
@@ -2518,7 +2518,7 @@ 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  196,98,125,24,5,160,32,0,0          ; vbroadcastss  0x20a0(%rip),%ymm8        # 4b20 <_sk_callback_hsw+0x373>
+  DB  196,98,125,24,5,12,33,0,0           ; vbroadcastss  0x210c(%rip),%ymm8        # 4b8c <_sk_callback_hsw+0x373>
   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
@@ -2583,10 +2583,10 @@ _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  196,226,125,24,13,214,31,0,0        ; vbroadcastss  0x1fd6(%rip),%ymm1        # 4b24 <_sk_callback_hsw+0x377>
+  DB  196,226,125,24,13,66,32,0,0         ; vbroadcastss  0x2042(%rip),%ymm1        # 4b90 <_sk_callback_hsw+0x377>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,203,31,0,0        ; vbroadcastss  0x1fcb(%rip),%ymm3        # 4b28 <_sk_callback_hsw+0x37b>
+  DB  196,226,125,24,29,55,32,0,0         ; vbroadcastss  0x2037(%rip),%ymm3        # 4b94 <_sk_callback_hsw+0x37b>
   DB  76,137,193                          ; mov           %r8,%rcx
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
@@ -2646,10 +2646,10 @@ _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  196,226,125,24,13,224,30,0,0        ; vbroadcastss  0x1ee0(%rip),%ymm1        # 4b2c <_sk_callback_hsw+0x37f>
+  DB  196,226,125,24,13,76,31,0,0         ; vbroadcastss  0x1f4c(%rip),%ymm1        # 4b98 <_sk_callback_hsw+0x37f>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,213,30,0,0        ; vbroadcastss  0x1ed5(%rip),%ymm3        # 4b30 <_sk_callback_hsw+0x383>
+  DB  196,226,125,24,29,65,31,0,0         ; vbroadcastss  0x1f41(%rip),%ymm3        # 4b9c <_sk_callback_hsw+0x383>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  91                                  ; pop           %rbx
@@ -2703,14 +2703,14 @@ _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  197,229,219,5,245,31,0,0            ; vpand         0x1ff5(%rip),%ymm3,%ymm0        # 4d20 <_sk_callback_hsw+0x573>
+  DB  197,229,219,5,85,32,0,0             ; vpand         0x2055(%rip),%ymm3,%ymm0        # 4d80 <_sk_callback_hsw+0x567>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,252,29,0,0          ; vbroadcastss  0x1dfc(%rip),%ymm8        # 4b34 <_sk_callback_hsw+0x387>
+  DB  196,98,125,24,5,104,30,0,0          ; vbroadcastss  0x1e68(%rip),%ymm8        # 4ba0 <_sk_callback_hsw+0x387>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,250,31,0,0         ; vpshufb       0x1ffa(%rip),%ymm3,%ymm1        # 4d40 <_sk_callback_hsw+0x593>
+  DB  196,226,101,0,13,90,32,0,0          ; vpshufb       0x205a(%rip),%ymm3,%ymm1        # 4da0 <_sk_callback_hsw+0x587>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,8,32,0,0           ; vpshufb       0x2008(%rip),%ymm3,%ymm2        # 4d60 <_sk_callback_hsw+0x5b3>
+  DB  196,226,101,0,21,104,32,0,0         ; vpshufb       0x2068(%rip),%ymm3,%ymm2        # 4dc0 <_sk_callback_hsw+0x5a7>
   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
@@ -2732,23 +2732,23 @@ _sk_load_565_hsw LABEL PROC
   DB  117,114                             ; jne           2df8 <_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  196,226,125,88,5,158,29,0,0         ; vpbroadcastd  0x1d9e(%rip),%ymm0        # 4b38 <_sk_callback_hsw+0x38b>
+  DB  196,226,125,88,5,10,30,0,0          ; vpbroadcastd  0x1e0a(%rip),%ymm0        # 4ba4 <_sk_callback_hsw+0x38b>
   DB  197,237,219,192                     ; vpand         %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,145,29,0,0        ; vbroadcastss  0x1d91(%rip),%ymm1        # 4b3c <_sk_callback_hsw+0x38f>
+  DB  196,226,125,24,13,253,29,0,0        ; vbroadcastss  0x1dfd(%rip),%ymm1        # 4ba8 <_sk_callback_hsw+0x38f>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,136,29,0,0        ; vpbroadcastd  0x1d88(%rip),%ymm1        # 4b40 <_sk_callback_hsw+0x393>
+  DB  196,226,125,88,13,244,29,0,0        ; vpbroadcastd  0x1df4(%rip),%ymm1        # 4bac <_sk_callback_hsw+0x393>
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,123,29,0,0        ; vbroadcastss  0x1d7b(%rip),%ymm3        # 4b44 <_sk_callback_hsw+0x397>
+  DB  196,226,125,24,29,231,29,0,0        ; vbroadcastss  0x1de7(%rip),%ymm3        # 4bb0 <_sk_callback_hsw+0x397>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,88,29,114,29,0,0        ; vpbroadcastd  0x1d72(%rip),%ymm3        # 4b48 <_sk_callback_hsw+0x39b>
+  DB  196,226,125,88,29,222,29,0,0        ; vpbroadcastd  0x1dde(%rip),%ymm3        # 4bb4 <_sk_callback_hsw+0x39b>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,101,29,0,0        ; vbroadcastss  0x1d65(%rip),%ymm3        # 4b4c <_sk_callback_hsw+0x39f>
+  DB  196,226,125,24,29,209,29,0,0        ; vbroadcastss  0x1dd1(%rip),%ymm3        # 4bb8 <_sk_callback_hsw+0x39f>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,90,29,0,0         ; vbroadcastss  0x1d5a(%rip),%ymm3        # 4b50 <_sk_callback_hsw+0x3a3>
+  DB  196,226,125,24,29,198,29,0,0        ; vbroadcastss  0x1dc6(%rip),%ymm3        # 4bbc <_sk_callback_hsw+0x3a3>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,200                          ; mov           %ecx,%r8d
   DB  65,128,224,7                        ; and           $0x7,%r8b
@@ -2837,23 +2837,23 @@ _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  196,226,125,88,5,29,28,0,0          ; vpbroadcastd  0x1c1d(%rip),%ymm0        # 4b54 <_sk_callback_hsw+0x3a7>
+  DB  196,226,125,88,5,137,28,0,0         ; vpbroadcastd  0x1c89(%rip),%ymm0        # 4bc0 <_sk_callback_hsw+0x3a7>
   DB  197,237,219,192                     ; vpand         %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,16,28,0,0         ; vbroadcastss  0x1c10(%rip),%ymm1        # 4b58 <_sk_callback_hsw+0x3ab>
+  DB  196,226,125,24,13,124,28,0,0        ; vbroadcastss  0x1c7c(%rip),%ymm1        # 4bc4 <_sk_callback_hsw+0x3ab>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,7,28,0,0          ; vpbroadcastd  0x1c07(%rip),%ymm1        # 4b5c <_sk_callback_hsw+0x3af>
+  DB  196,226,125,88,13,115,28,0,0        ; vpbroadcastd  0x1c73(%rip),%ymm1        # 4bc8 <_sk_callback_hsw+0x3af>
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,250,27,0,0        ; vbroadcastss  0x1bfa(%rip),%ymm3        # 4b60 <_sk_callback_hsw+0x3b3>
+  DB  196,226,125,24,29,102,28,0,0        ; vbroadcastss  0x1c66(%rip),%ymm3        # 4bcc <_sk_callback_hsw+0x3b3>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,88,29,241,27,0,0        ; vpbroadcastd  0x1bf1(%rip),%ymm3        # 4b64 <_sk_callback_hsw+0x3b7>
+  DB  196,226,125,88,29,93,28,0,0         ; vpbroadcastd  0x1c5d(%rip),%ymm3        # 4bd0 <_sk_callback_hsw+0x3b7>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,228,27,0,0        ; vbroadcastss  0x1be4(%rip),%ymm3        # 4b68 <_sk_callback_hsw+0x3bb>
+  DB  196,226,125,24,29,80,28,0,0         ; vbroadcastss  0x1c50(%rip),%ymm3        # 4bd4 <_sk_callback_hsw+0x3bb>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,217,27,0,0        ; vbroadcastss  0x1bd9(%rip),%ymm3        # 4b6c <_sk_callback_hsw+0x3bf>
+  DB  196,226,125,24,29,69,28,0,0         ; vbroadcastss  0x1c45(%rip),%ymm3        # 4bd8 <_sk_callback_hsw+0x3bf>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,94                               ; pop           %r14
@@ -2864,11 +2864,11 @@ 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  196,98,125,24,5,198,27,0,0          ; vbroadcastss  0x1bc6(%rip),%ymm8        # 4b70 <_sk_callback_hsw+0x3c3>
+  DB  196,98,125,24,5,50,28,0,0           ; vbroadcastss  0x1c32(%rip),%ymm8        # 4bdc <_sk_callback_hsw+0x3c3>
   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  196,98,125,24,21,177,27,0,0         ; vbroadcastss  0x1bb1(%rip),%ymm10        # 4b74 <_sk_callback_hsw+0x3c7>
+  DB  196,98,125,24,21,29,28,0,0          ; vbroadcastss  0x1c1d(%rip),%ymm10        # 4be0 <_sk_callback_hsw+0x3c7>
   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
@@ -2934,25 +2934,25 @@ _sk_load_4444_hsw LABEL PROC
   DB  15,133,138,0,0,0                    ; jne           3114 <_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  196,226,125,88,5,218,26,0,0         ; vpbroadcastd  0x1ada(%rip),%ymm0        # 4b78 <_sk_callback_hsw+0x3cb>
+  DB  196,226,125,88,5,70,27,0,0          ; vpbroadcastd  0x1b46(%rip),%ymm0        # 4be4 <_sk_callback_hsw+0x3cb>
   DB  197,229,219,192                     ; vpand         %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,205,26,0,0        ; vbroadcastss  0x1acd(%rip),%ymm1        # 4b7c <_sk_callback_hsw+0x3cf>
+  DB  196,226,125,24,13,57,27,0,0         ; vbroadcastss  0x1b39(%rip),%ymm1        # 4be8 <_sk_callback_hsw+0x3cf>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,196,26,0,0        ; vpbroadcastd  0x1ac4(%rip),%ymm1        # 4b80 <_sk_callback_hsw+0x3d3>
+  DB  196,226,125,88,13,48,27,0,0         ; vpbroadcastd  0x1b30(%rip),%ymm1        # 4bec <_sk_callback_hsw+0x3d3>
   DB  197,229,219,201                     ; vpand         %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,183,26,0,0        ; vbroadcastss  0x1ab7(%rip),%ymm2        # 4b84 <_sk_callback_hsw+0x3d7>
+  DB  196,226,125,24,21,35,27,0,0         ; vbroadcastss  0x1b23(%rip),%ymm2        # 4bf0 <_sk_callback_hsw+0x3d7>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,88,21,174,26,0,0        ; vpbroadcastd  0x1aae(%rip),%ymm2        # 4b88 <_sk_callback_hsw+0x3db>
+  DB  196,226,125,88,21,26,27,0,0         ; vpbroadcastd  0x1b1a(%rip),%ymm2        # 4bf4 <_sk_callback_hsw+0x3db>
   DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,161,26,0,0          ; vbroadcastss  0x1aa1(%rip),%ymm8        # 4b8c <_sk_callback_hsw+0x3df>
+  DB  196,98,125,24,5,13,27,0,0           ; vbroadcastss  0x1b0d(%rip),%ymm8        # 4bf8 <_sk_callback_hsw+0x3df>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,88,5,151,26,0,0          ; vpbroadcastd  0x1a97(%rip),%ymm8        # 4b90 <_sk_callback_hsw+0x3e3>
+  DB  196,98,125,88,5,3,27,0,0            ; vpbroadcastd  0x1b03(%rip),%ymm8        # 4bfc <_sk_callback_hsw+0x3e3>
   DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,137,26,0,0          ; vbroadcastss  0x1a89(%rip),%ymm8        # 4b94 <_sk_callback_hsw+0x3e7>
+  DB  196,98,125,24,5,245,26,0,0          ; vbroadcastss  0x1af5(%rip),%ymm8        # 4c00 <_sk_callback_hsw+0x3e7>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -3043,25 +3043,25 @@ _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  196,226,125,88,5,65,25,0,0          ; vpbroadcastd  0x1941(%rip),%ymm0        # 4b98 <_sk_callback_hsw+0x3eb>
+  DB  196,226,125,88,5,173,25,0,0         ; vpbroadcastd  0x19ad(%rip),%ymm0        # 4c04 <_sk_callback_hsw+0x3eb>
   DB  197,229,219,192                     ; vpand         %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,52,25,0,0         ; vbroadcastss  0x1934(%rip),%ymm1        # 4b9c <_sk_callback_hsw+0x3ef>
+  DB  196,226,125,24,13,160,25,0,0        ; vbroadcastss  0x19a0(%rip),%ymm1        # 4c08 <_sk_callback_hsw+0x3ef>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,43,25,0,0         ; vpbroadcastd  0x192b(%rip),%ymm1        # 4ba0 <_sk_callback_hsw+0x3f3>
+  DB  196,226,125,88,13,151,25,0,0        ; vpbroadcastd  0x1997(%rip),%ymm1        # 4c0c <_sk_callback_hsw+0x3f3>
   DB  197,229,219,201                     ; vpand         %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,30,25,0,0         ; vbroadcastss  0x191e(%rip),%ymm2        # 4ba4 <_sk_callback_hsw+0x3f7>
+  DB  196,226,125,24,21,138,25,0,0        ; vbroadcastss  0x198a(%rip),%ymm2        # 4c10 <_sk_callback_hsw+0x3f7>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,88,21,21,25,0,0         ; vpbroadcastd  0x1915(%rip),%ymm2        # 4ba8 <_sk_callback_hsw+0x3fb>
+  DB  196,226,125,88,21,129,25,0,0        ; vpbroadcastd  0x1981(%rip),%ymm2        # 4c14 <_sk_callback_hsw+0x3fb>
   DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,8,25,0,0            ; vbroadcastss  0x1908(%rip),%ymm8        # 4bac <_sk_callback_hsw+0x3ff>
+  DB  196,98,125,24,5,116,25,0,0          ; vbroadcastss  0x1974(%rip),%ymm8        # 4c18 <_sk_callback_hsw+0x3ff>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,88,5,254,24,0,0          ; vpbroadcastd  0x18fe(%rip),%ymm8        # 4bb0 <_sk_callback_hsw+0x403>
+  DB  196,98,125,88,5,106,25,0,0          ; vpbroadcastd  0x196a(%rip),%ymm8        # 4c1c <_sk_callback_hsw+0x403>
   DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,240,24,0,0          ; vbroadcastss  0x18f0(%rip),%ymm8        # 4bb4 <_sk_callback_hsw+0x407>
+  DB  196,98,125,24,5,92,25,0,0           ; vbroadcastss  0x195c(%rip),%ymm8        # 4c20 <_sk_callback_hsw+0x407>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -3074,7 +3074,7 @@ 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  196,98,125,24,5,214,24,0,0          ; vbroadcastss  0x18d6(%rip),%ymm8        # 4bb8 <_sk_callback_hsw+0x40b>
+  DB  196,98,125,24,5,66,25,0,0           ; vbroadcastss  0x1942(%rip),%ymm8        # 4c24 <_sk_callback_hsw+0x40b>
   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
@@ -3148,14 +3148,14 @@ _sk_load_8888_hsw LABEL PROC
   DB  77,133,192                          ; test          %r8,%r8
   DB  117,88                              ; jne           342d <_sk_load_8888_hsw+0x6d>
   DB  196,193,126,111,25                  ; vmovdqu       (%r9),%ymm3
-  DB  197,229,219,5,158,25,0,0            ; vpand         0x199e(%rip),%ymm3,%ymm0        # 4d80 <_sk_callback_hsw+0x5d3>
+  DB  197,229,219,5,254,25,0,0            ; vpand         0x19fe(%rip),%ymm3,%ymm0        # 4de0 <_sk_callback_hsw+0x5c7>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,205,23,0,0          ; vbroadcastss  0x17cd(%rip),%ymm8        # 4bbc <_sk_callback_hsw+0x40f>
+  DB  196,98,125,24,5,57,24,0,0           ; vbroadcastss  0x1839(%rip),%ymm8        # 4c28 <_sk_callback_hsw+0x40f>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,163,25,0,0         ; vpshufb       0x19a3(%rip),%ymm3,%ymm1        # 4da0 <_sk_callback_hsw+0x5f3>
+  DB  196,226,101,0,13,3,26,0,0           ; vpshufb       0x1a03(%rip),%ymm3,%ymm1        # 4e00 <_sk_callback_hsw+0x5e7>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,177,25,0,0         ; vpshufb       0x19b1(%rip),%ymm3,%ymm2        # 4dc0 <_sk_callback_hsw+0x613>
+  DB  196,226,101,0,21,17,26,0,0          ; vpshufb       0x1a11(%rip),%ymm3,%ymm2        # 4e20 <_sk_callback_hsw+0x607>
   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
@@ -3185,14 +3185,14 @@ _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  197,229,219,5,95,25,0,0             ; vpand         0x195f(%rip),%ymm3,%ymm0        # 4de0 <_sk_callback_hsw+0x633>
+  DB  197,229,219,5,191,25,0,0            ; vpand         0x19bf(%rip),%ymm3,%ymm0        # 4e40 <_sk_callback_hsw+0x627>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,50,23,0,0           ; vbroadcastss  0x1732(%rip),%ymm8        # 4bc0 <_sk_callback_hsw+0x413>
+  DB  196,98,125,24,5,158,23,0,0          ; vbroadcastss  0x179e(%rip),%ymm8        # 4c2c <_sk_callback_hsw+0x413>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,100,25,0,0         ; vpshufb       0x1964(%rip),%ymm3,%ymm1        # 4e00 <_sk_callback_hsw+0x653>
+  DB  196,226,101,0,13,196,25,0,0         ; vpshufb       0x19c4(%rip),%ymm3,%ymm1        # 4e60 <_sk_callback_hsw+0x647>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,114,25,0,0         ; vpshufb       0x1972(%rip),%ymm3,%ymm2        # 4e20 <_sk_callback_hsw+0x673>
+  DB  196,226,101,0,21,210,25,0,0         ; vpshufb       0x19d2(%rip),%ymm3,%ymm2        # 4e80 <_sk_callback_hsw+0x667>
   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
@@ -3207,7 +3207,7 @@ _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  196,98,125,24,5,226,22,0,0          ; vbroadcastss  0x16e2(%rip),%ymm8        # 4bc4 <_sk_callback_hsw+0x417>
+  DB  196,98,125,24,5,78,23,0,0           ; vbroadcastss  0x174e(%rip),%ymm8        # 4c30 <_sk_callback_hsw+0x417>
   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
@@ -3396,7 +3396,7 @@ _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  196,98,125,24,21,217,19,0,0         ; vbroadcastss  0x13d9(%rip),%ymm10        # 4bc8 <_sk_callback_hsw+0x41b>
+  DB  196,98,125,24,21,69,20,0,0          ; vbroadcastss  0x1445(%rip),%ymm10        # 4c34 <_sk_callback_hsw+0x41b>
   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
@@ -3478,7 +3478,7 @@ _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  196,98,125,24,21,106,18,0,0         ; vbroadcastss  0x126a(%rip),%ymm10        # 4bcc <_sk_callback_hsw+0x41f>
+  DB  196,98,125,24,21,214,18,0,0         ; vbroadcastss  0x12d6(%rip),%ymm10        # 4c38 <_sk_callback_hsw+0x41f>
   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
@@ -3495,7 +3495,7 @@ _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  196,226,125,24,29,30,18,0,0         ; vbroadcastss  0x121e(%rip),%ymm3        # 4bd0 <_sk_callback_hsw+0x423>
+  DB  196,226,125,24,29,138,18,0,0        ; vbroadcastss  0x128a(%rip),%ymm3        # 4c3c <_sk_callback_hsw+0x423>
   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
@@ -3536,7 +3536,7 @@ _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  196,98,125,24,5,91,17,0,0           ; vbroadcastss  0x115b(%rip),%ymm8        # 4bd4 <_sk_callback_hsw+0x427>
+  DB  196,98,125,24,5,199,17,0,0          ; vbroadcastss  0x11c7(%rip),%ymm8        # 4c40 <_sk_callback_hsw+0x427>
   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
@@ -3776,13 +3776,45 @@ _sk_mirror_y_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_clamp_x_1_hsw
+_sk_clamp_x_1_hsw LABEL PROC
+  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
+  DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
+  DB  196,98,125,24,5,14,14,0,0           ; vbroadcastss  0xe0e(%rip),%ymm8        # 4c44 <_sk_callback_hsw+0x42b>
+  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_repeat_x_1_hsw
+_sk_repeat_x_1_hsw LABEL PROC
+  DB  196,99,125,8,192,1                  ; vroundps      $0x1,%ymm0,%ymm8
+  DB  196,193,124,92,192                  ; vsubps        %ymm8,%ymm0,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_mirror_x_1_hsw
+_sk_mirror_x_1_hsw LABEL PROC
+  DB  196,98,125,24,5,241,13,0,0          ; vbroadcastss  0xdf1(%rip),%ymm8        # 4c48 <_sk_callback_hsw+0x42f>
+  DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
+  DB  196,98,125,24,13,231,13,0,0         ; vbroadcastss  0xde7(%rip),%ymm9        # 4c4c <_sk_callback_hsw+0x433>
+  DB  196,65,124,89,201                   ; vmulps        %ymm9,%ymm0,%ymm9
+  DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
+  DB  196,65,52,88,201                    ; vaddps        %ymm9,%ymm9,%ymm9
+  DB  196,193,124,92,193                  ; vsubps        %ymm9,%ymm0,%ymm0
+  DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
+  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
+  DB  197,60,92,192                       ; vsubps        %ymm0,%ymm8,%ymm8
+  DB  197,188,84,192                      ; vandps        %ymm0,%ymm8,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_luminance_to_alpha_hsw
 _sk_luminance_to_alpha_hsw LABEL PROC
-  DB  196,226,125,24,29,171,13,0,0        ; vbroadcastss  0xdab(%rip),%ymm3        # 4bd8 <_sk_callback_hsw+0x42b>
-  DB  196,98,125,24,5,166,13,0,0          ; vbroadcastss  0xda6(%rip),%ymm8        # 4bdc <_sk_callback_hsw+0x42f>
+  DB  196,226,125,24,29,183,13,0,0        ; vbroadcastss  0xdb7(%rip),%ymm3        # 4c50 <_sk_callback_hsw+0x437>
+  DB  196,98,125,24,5,178,13,0,0          ; vbroadcastss  0xdb2(%rip),%ymm8        # 4c54 <_sk_callback_hsw+0x43b>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  196,226,125,184,203                 ; vfmadd231ps   %ymm3,%ymm0,%ymm1
-  DB  196,226,125,24,29,151,13,0,0        ; vbroadcastss  0xd97(%rip),%ymm3        # 4be0 <_sk_callback_hsw+0x433>
+  DB  196,226,125,24,29,163,13,0,0        ; vbroadcastss  0xda3(%rip),%ymm3        # 4c58 <_sk_callback_hsw+0x43f>
   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
@@ -3943,9 +3975,9 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
   DB  76,139,64,8                         ; mov           0x8(%rax),%r8
   DB  77,137,202                          ; mov           %r9,%r10
   DB  73,255,202                          ; dec           %r10
-  DB  120,7                               ; js            4101 <_sk_evenly_spaced_gradient_hsw+0x18>
+  DB  120,7                               ; js            416d <_sk_evenly_spaced_gradient_hsw+0x18>
   DB  196,193,242,42,202                  ; vcvtsi2ss     %r10,%xmm1,%xmm1
-  DB  235,22                              ; jmp           4117 <_sk_evenly_spaced_gradient_hsw+0x2e>
+  DB  235,22                              ; jmp           4183 <_sk_evenly_spaced_gradient_hsw+0x2e>
   DB  77,137,211                          ; mov           %r10,%r11
   DB  73,209,235                          ; shr           %r11
   DB  65,131,226,1                        ; and           $0x1,%r10d
@@ -3956,7 +3988,7 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
   DB  197,244,89,200                      ; vmulps        %ymm0,%ymm1,%ymm1
   DB  197,126,91,217                      ; vcvttps2dq    %ymm1,%ymm11
   DB  73,131,249,8                        ; cmp           $0x8,%r9
-  DB  119,70                              ; ja            4170 <_sk_evenly_spaced_gradient_hsw+0x87>
+  DB  119,70                              ; ja            41dc <_sk_evenly_spaced_gradient_hsw+0x87>
   DB  196,66,37,22,0                      ; vpermps       (%r8),%ymm11,%ymm8
   DB  76,139,64,40                        ; mov           0x28(%rax),%r8
   DB  196,66,37,22,8                      ; vpermps       (%r8),%ymm11,%ymm9
@@ -3972,7 +4004,7 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC
   DB  196,194,37,22,24                    ; vpermps       (%r8),%ymm11,%ymm3
   DB  72,139,64,64                        ; mov           0x40(%rax),%rax
   DB  196,98,37,22,40                     ; vpermps       (%rax),%ymm11,%ymm13
-  DB  235,110                             ; jmp           41de <_sk_evenly_spaced_gradient_hsw+0xf5>
+  DB  235,110                             ; jmp           424a <_sk_evenly_spaced_gradient_hsw+0xf5>
   DB  196,65,13,118,246                   ; vpcmpeqd      %ymm14,%ymm14,%ymm14
   DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
   DB  196,2,117,146,4,152                 ; vgatherdps    %ymm1,(%r8,%ymm11,4),%ymm8
@@ -4009,11 +4041,11 @@ _sk_gradient_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  73,131,248,1                        ; cmp           $0x1,%r8
-  DB  15,134,180,0,0,0                    ; jbe           42bd <_sk_gradient_hsw+0xc3>
+  DB  15,134,180,0,0,0                    ; jbe           4329 <_sk_gradient_hsw+0xc3>
   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  196,226,125,24,21,196,9,0,0         ; vbroadcastss  0x9c4(%rip),%ymm2        # 4be4 <_sk_callback_hsw+0x437>
+  DB  196,226,125,24,21,208,9,0,0         ; vbroadcastss  0x9d0(%rip),%ymm2        # 4c5c <_sk_callback_hsw+0x443>
   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
@@ -4021,10 +4053,10 @@ _sk_gradient_hsw LABEL PROC
   DB  196,65,101,254,201                  ; vpaddd        %ymm9,%ymm3,%ymm9
   DB  73,255,194                          ; inc           %r10
   DB  77,57,208                           ; cmp           %r10,%r8
-  DB  117,226                             ; jne           4225 <_sk_gradient_hsw+0x2b>
+  DB  117,226                             ; jne           4291 <_sk_gradient_hsw+0x2b>
   DB  76,139,72,8                         ; mov           0x8(%rax),%r9
   DB  73,131,248,8                        ; cmp           $0x8,%r8
-  DB  118,121                             ; jbe           42c6 <_sk_gradient_hsw+0xcc>
+  DB  118,121                             ; jbe           4332 <_sk_gradient_hsw+0xcc>
   DB  196,65,13,118,246                   ; vpcmpeqd      %ymm14,%ymm14,%ymm14
   DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
   DB  196,2,117,146,4,137                 ; vgatherdps    %ymm1,(%r9,%ymm9,4),%ymm8
@@ -4048,7 +4080,7 @@ _sk_gradient_hsw LABEL PROC
   DB  196,130,21,146,28,136               ; vgatherdps    %ymm13,(%r8,%ymm9,4),%ymm3
   DB  72,139,64,64                        ; mov           0x40(%rax),%rax
   DB  196,34,13,146,44,136                ; vgatherdps    %ymm14,(%rax,%ymm9,4),%ymm13
-  DB  235,77                              ; jmp           430a <_sk_gradient_hsw+0x110>
+  DB  235,77                              ; jmp           4376 <_sk_gradient_hsw+0x110>
   DB  76,139,72,8                         ; mov           0x8(%rax),%r9
   DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
   DB  196,66,53,22,1                      ; vpermps       (%r9),%ymm9,%ymm8
@@ -4104,24 +4136,24 @@ _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  196,98,125,24,45,67,8,0,0           ; vbroadcastss  0x843(%rip),%ymm13        # 4be8 <_sk_callback_hsw+0x43b>
-  DB  196,98,125,24,53,62,8,0,0           ; vbroadcastss  0x83e(%rip),%ymm14        # 4bec <_sk_callback_hsw+0x43f>
+  DB  196,98,125,24,45,79,8,0,0           ; vbroadcastss  0x84f(%rip),%ymm13        # 4c60 <_sk_callback_hsw+0x447>
+  DB  196,98,125,24,53,74,8,0,0           ; vbroadcastss  0x84a(%rip),%ymm14        # 4c64 <_sk_callback_hsw+0x44b>
   DB  196,66,29,184,245                   ; vfmadd231ps   %ymm13,%ymm12,%ymm14
-  DB  196,98,125,24,45,52,8,0,0           ; vbroadcastss  0x834(%rip),%ymm13        # 4bf0 <_sk_callback_hsw+0x443>
+  DB  196,98,125,24,45,64,8,0,0           ; vbroadcastss  0x840(%rip),%ymm13        # 4c68 <_sk_callback_hsw+0x44f>
   DB  196,66,29,184,238                   ; vfmadd231ps   %ymm14,%ymm12,%ymm13
-  DB  196,98,125,24,53,42,8,0,0           ; vbroadcastss  0x82a(%rip),%ymm14        # 4bf4 <_sk_callback_hsw+0x447>
+  DB  196,98,125,24,53,54,8,0,0           ; vbroadcastss  0x836(%rip),%ymm14        # 4c6c <_sk_callback_hsw+0x453>
   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  196,98,125,24,21,21,8,0,0           ; vbroadcastss  0x815(%rip),%ymm10        # 4bf8 <_sk_callback_hsw+0x44b>
+  DB  196,98,125,24,21,33,8,0,0           ; vbroadcastss  0x821(%rip),%ymm10        # 4c70 <_sk_callback_hsw+0x457>
   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  196,98,125,24,21,255,7,0,0          ; vbroadcastss  0x7ff(%rip),%ymm10        # 4bfc <_sk_callback_hsw+0x44f>
+  DB  196,98,125,24,21,11,8,0,0           ; vbroadcastss  0x80b(%rip),%ymm10        # 4c74 <_sk_callback_hsw+0x45b>
   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  196,98,125,24,21,233,7,0,0          ; vbroadcastss  0x7e9(%rip),%ymm10        # 4c00 <_sk_callback_hsw+0x453>
+  DB  196,98,125,24,21,245,7,0,0          ; vbroadcastss  0x7f5(%rip),%ymm10        # 4c78 <_sk_callback_hsw+0x45f>
   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
@@ -4140,7 +4172,7 @@ _sk_xy_to_radius_hsw LABEL PROC
 PUBLIC _sk_save_xy_hsw
 _sk_save_xy_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,182,7,0,0           ; vbroadcastss  0x7b6(%rip),%ymm8        # 4c04 <_sk_callback_hsw+0x457>
+  DB  196,98,125,24,5,194,7,0,0           ; vbroadcastss  0x7c2(%rip),%ymm8        # 4c7c <_sk_callback_hsw+0x463>
   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
@@ -4170,9 +4202,9 @@ _sk_accumulate_hsw LABEL PROC
 PUBLIC _sk_bilinear_nx_hsw
 _sk_bilinear_nx_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,74,7,0,0           ; vbroadcastss  0x74a(%rip),%ymm0        # 4c08 <_sk_callback_hsw+0x45b>
+  DB  196,226,125,24,5,86,7,0,0           ; vbroadcastss  0x756(%rip),%ymm0        # 4c80 <_sk_callback_hsw+0x467>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,65,7,0,0            ; vbroadcastss  0x741(%rip),%ymm8        # 4c0c <_sk_callback_hsw+0x45f>
+  DB  196,98,125,24,5,77,7,0,0            ; vbroadcastss  0x74d(%rip),%ymm8        # 4c84 <_sk_callback_hsw+0x46b>
   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
@@ -4181,7 +4213,7 @@ _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  196,226,125,24,5,41,7,0,0           ; vbroadcastss  0x729(%rip),%ymm0        # 4c10 <_sk_callback_hsw+0x463>
+  DB  196,226,125,24,5,53,7,0,0           ; vbroadcastss  0x735(%rip),%ymm0        # 4c88 <_sk_callback_hsw+0x46f>
   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)
@@ -4191,9 +4223,9 @@ _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  196,226,125,24,13,13,7,0,0          ; vbroadcastss  0x70d(%rip),%ymm1        # 4c14 <_sk_callback_hsw+0x467>
+  DB  196,226,125,24,13,25,7,0,0          ; vbroadcastss  0x719(%rip),%ymm1        # 4c8c <_sk_callback_hsw+0x473>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,3,7,0,0             ; vbroadcastss  0x703(%rip),%ymm8        # 4c18 <_sk_callback_hsw+0x46b>
+  DB  196,98,125,24,5,15,7,0,0            ; vbroadcastss  0x70f(%rip),%ymm8        # 4c90 <_sk_callback_hsw+0x477>
   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
@@ -4202,7 +4234,7 @@ _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  196,226,125,24,13,235,6,0,0         ; vbroadcastss  0x6eb(%rip),%ymm1        # 4c1c <_sk_callback_hsw+0x46f>
+  DB  196,226,125,24,13,247,6,0,0         ; vbroadcastss  0x6f7(%rip),%ymm1        # 4c94 <_sk_callback_hsw+0x47b>
   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)
@@ -4212,13 +4244,13 @@ _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  196,226,125,24,5,206,6,0,0          ; vbroadcastss  0x6ce(%rip),%ymm0        # 4c20 <_sk_callback_hsw+0x473>
+  DB  196,226,125,24,5,218,6,0,0          ; vbroadcastss  0x6da(%rip),%ymm0        # 4c98 <_sk_callback_hsw+0x47f>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,197,6,0,0           ; vbroadcastss  0x6c5(%rip),%ymm8        # 4c24 <_sk_callback_hsw+0x477>
+  DB  196,98,125,24,5,209,6,0,0           ; vbroadcastss  0x6d1(%rip),%ymm8        # 4c9c <_sk_callback_hsw+0x483>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,182,6,0,0          ; vbroadcastss  0x6b6(%rip),%ymm10        # 4c28 <_sk_callback_hsw+0x47b>
-  DB  196,98,125,24,29,177,6,0,0          ; vbroadcastss  0x6b1(%rip),%ymm11        # 4c2c <_sk_callback_hsw+0x47f>
+  DB  196,98,125,24,21,194,6,0,0          ; vbroadcastss  0x6c2(%rip),%ymm10        # 4ca0 <_sk_callback_hsw+0x487>
+  DB  196,98,125,24,29,189,6,0,0          ; vbroadcastss  0x6bd(%rip),%ymm11        # 4ca4 <_sk_callback_hsw+0x48b>
   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)
@@ -4228,16 +4260,16 @@ _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  196,226,125,24,5,148,6,0,0          ; vbroadcastss  0x694(%rip),%ymm0        # 4c30 <_sk_callback_hsw+0x483>
+  DB  196,226,125,24,5,160,6,0,0          ; vbroadcastss  0x6a0(%rip),%ymm0        # 4ca8 <_sk_callback_hsw+0x48f>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,139,6,0,0           ; vbroadcastss  0x68b(%rip),%ymm8        # 4c34 <_sk_callback_hsw+0x487>
+  DB  196,98,125,24,5,151,6,0,0           ; vbroadcastss  0x697(%rip),%ymm8        # 4cac <_sk_callback_hsw+0x493>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,129,6,0,0          ; vbroadcastss  0x681(%rip),%ymm9        # 4c38 <_sk_callback_hsw+0x48b>
-  DB  196,98,125,24,21,124,6,0,0          ; vbroadcastss  0x67c(%rip),%ymm10        # 4c3c <_sk_callback_hsw+0x48f>
+  DB  196,98,125,24,13,141,6,0,0          ; vbroadcastss  0x68d(%rip),%ymm9        # 4cb0 <_sk_callback_hsw+0x497>
+  DB  196,98,125,24,21,136,6,0,0          ; vbroadcastss  0x688(%rip),%ymm10        # 4cb4 <_sk_callback_hsw+0x49b>
   DB  196,66,61,168,209                   ; vfmadd213ps   %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,13,114,6,0,0          ; vbroadcastss  0x672(%rip),%ymm9        # 4c40 <_sk_callback_hsw+0x493>
+  DB  196,98,125,24,13,126,6,0,0          ; vbroadcastss  0x67e(%rip),%ymm9        # 4cb8 <_sk_callback_hsw+0x49f>
   DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
-  DB  196,98,125,24,21,104,6,0,0          ; vbroadcastss  0x668(%rip),%ymm10        # 4c44 <_sk_callback_hsw+0x497>
+  DB  196,98,125,24,21,116,6,0,0          ; vbroadcastss  0x674(%rip),%ymm10        # 4cbc <_sk_callback_hsw+0x4a3>
   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
@@ -4246,14 +4278,14 @@ _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  196,98,125,24,5,80,6,0,0            ; vbroadcastss  0x650(%rip),%ymm8        # 4c48 <_sk_callback_hsw+0x49b>
+  DB  196,98,125,24,5,92,6,0,0            ; vbroadcastss  0x65c(%rip),%ymm8        # 4cc0 <_sk_callback_hsw+0x4a7>
   DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
   DB  197,124,16,72,64                    ; vmovups       0x40(%rax),%ymm9
-  DB  196,98,125,24,21,66,6,0,0           ; vbroadcastss  0x642(%rip),%ymm10        # 4c4c <_sk_callback_hsw+0x49f>
-  DB  196,98,125,24,29,61,6,0,0           ; vbroadcastss  0x63d(%rip),%ymm11        # 4c50 <_sk_callback_hsw+0x4a3>
+  DB  196,98,125,24,21,78,6,0,0           ; vbroadcastss  0x64e(%rip),%ymm10        # 4cc4 <_sk_callback_hsw+0x4ab>
+  DB  196,98,125,24,29,73,6,0,0           ; vbroadcastss  0x649(%rip),%ymm11        # 4cc8 <_sk_callback_hsw+0x4af>
   DB  196,66,53,168,218                   ; vfmadd213ps   %ymm10,%ymm9,%ymm11
   DB  196,66,53,168,216                   ; vfmadd213ps   %ymm8,%ymm9,%ymm11
-  DB  196,98,125,24,5,46,6,0,0            ; vbroadcastss  0x62e(%rip),%ymm8        # 4c54 <_sk_callback_hsw+0x4a7>
+  DB  196,98,125,24,5,58,6,0,0            ; vbroadcastss  0x63a(%rip),%ymm8        # 4ccc <_sk_callback_hsw+0x4b3>
   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
@@ -4262,12 +4294,12 @@ _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  196,226,125,24,5,22,6,0,0           ; vbroadcastss  0x616(%rip),%ymm0        # 4c58 <_sk_callback_hsw+0x4ab>
+  DB  196,226,125,24,5,34,6,0,0           ; vbroadcastss  0x622(%rip),%ymm0        # 4cd0 <_sk_callback_hsw+0x4b7>
   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  196,98,125,24,21,3,6,0,0            ; vbroadcastss  0x603(%rip),%ymm10        # 4c5c <_sk_callback_hsw+0x4af>
-  DB  196,98,125,24,29,254,5,0,0          ; vbroadcastss  0x5fe(%rip),%ymm11        # 4c60 <_sk_callback_hsw+0x4b3>
+  DB  196,98,125,24,21,15,6,0,0           ; vbroadcastss  0x60f(%rip),%ymm10        # 4cd4 <_sk_callback_hsw+0x4bb>
+  DB  196,98,125,24,29,10,6,0,0           ; vbroadcastss  0x60a(%rip),%ymm11        # 4cd8 <_sk_callback_hsw+0x4bf>
   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)
@@ -4277,13 +4309,13 @@ _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  196,226,125,24,13,225,5,0,0         ; vbroadcastss  0x5e1(%rip),%ymm1        # 4c64 <_sk_callback_hsw+0x4b7>
+  DB  196,226,125,24,13,237,5,0,0         ; vbroadcastss  0x5ed(%rip),%ymm1        # 4cdc <_sk_callback_hsw+0x4c3>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,215,5,0,0           ; vbroadcastss  0x5d7(%rip),%ymm8        # 4c68 <_sk_callback_hsw+0x4bb>
+  DB  196,98,125,24,5,227,5,0,0           ; vbroadcastss  0x5e3(%rip),%ymm8        # 4ce0 <_sk_callback_hsw+0x4c7>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,200,5,0,0          ; vbroadcastss  0x5c8(%rip),%ymm10        # 4c6c <_sk_callback_hsw+0x4bf>
-  DB  196,98,125,24,29,195,5,0,0          ; vbroadcastss  0x5c3(%rip),%ymm11        # 4c70 <_sk_callback_hsw+0x4c3>
+  DB  196,98,125,24,21,212,5,0,0          ; vbroadcastss  0x5d4(%rip),%ymm10        # 4ce4 <_sk_callback_hsw+0x4cb>
+  DB  196,98,125,24,29,207,5,0,0          ; vbroadcastss  0x5cf(%rip),%ymm11        # 4ce8 <_sk_callback_hsw+0x4cf>
   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)
@@ -4293,16 +4325,16 @@ _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  196,226,125,24,13,166,5,0,0         ; vbroadcastss  0x5a6(%rip),%ymm1        # 4c74 <_sk_callback_hsw+0x4c7>
+  DB  196,226,125,24,13,178,5,0,0         ; vbroadcastss  0x5b2(%rip),%ymm1        # 4cec <_sk_callback_hsw+0x4d3>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,156,5,0,0           ; vbroadcastss  0x59c(%rip),%ymm8        # 4c78 <_sk_callback_hsw+0x4cb>
+  DB  196,98,125,24,5,168,5,0,0           ; vbroadcastss  0x5a8(%rip),%ymm8        # 4cf0 <_sk_callback_hsw+0x4d7>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,146,5,0,0          ; vbroadcastss  0x592(%rip),%ymm9        # 4c7c <_sk_callback_hsw+0x4cf>
-  DB  196,98,125,24,21,141,5,0,0          ; vbroadcastss  0x58d(%rip),%ymm10        # 4c80 <_sk_callback_hsw+0x4d3>
+  DB  196,98,125,24,13,158,5,0,0          ; vbroadcastss  0x59e(%rip),%ymm9        # 4cf4 <_sk_callback_hsw+0x4db>
+  DB  196,98,125,24,21,153,5,0,0          ; vbroadcastss  0x599(%rip),%ymm10        # 4cf8 <_sk_callback_hsw+0x4df>
   DB  196,66,61,168,209                   ; vfmadd213ps   %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,13,131,5,0,0          ; vbroadcastss  0x583(%rip),%ymm9        # 4c84 <_sk_callback_hsw+0x4d7>
+  DB  196,98,125,24,13,143,5,0,0          ; vbroadcastss  0x58f(%rip),%ymm9        # 4cfc <_sk_callback_hsw+0x4e3>
   DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
-  DB  196,98,125,24,21,121,5,0,0          ; vbroadcastss  0x579(%rip),%ymm10        # 4c88 <_sk_callback_hsw+0x4db>
+  DB  196,98,125,24,21,133,5,0,0          ; vbroadcastss  0x585(%rip),%ymm10        # 4d00 <_sk_callback_hsw+0x4e7>
   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
@@ -4311,14 +4343,14 @@ _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  196,98,125,24,5,97,5,0,0            ; vbroadcastss  0x561(%rip),%ymm8        # 4c8c <_sk_callback_hsw+0x4df>
+  DB  196,98,125,24,5,109,5,0,0           ; vbroadcastss  0x56d(%rip),%ymm8        # 4d04 <_sk_callback_hsw+0x4eb>
   DB  197,188,88,72,32                    ; vaddps        0x20(%rax),%ymm8,%ymm1
   DB  197,124,16,72,96                    ; vmovups       0x60(%rax),%ymm9
-  DB  196,98,125,24,21,82,5,0,0           ; vbroadcastss  0x552(%rip),%ymm10        # 4c90 <_sk_callback_hsw+0x4e3>
-  DB  196,98,125,24,29,77,5,0,0           ; vbroadcastss  0x54d(%rip),%ymm11        # 4c94 <_sk_callback_hsw+0x4e7>
+  DB  196,98,125,24,21,94,5,0,0           ; vbroadcastss  0x55e(%rip),%ymm10        # 4d08 <_sk_callback_hsw+0x4ef>
+  DB  196,98,125,24,29,89,5,0,0           ; vbroadcastss  0x559(%rip),%ymm11        # 4d0c <_sk_callback_hsw+0x4f3>
   DB  196,66,53,168,218                   ; vfmadd213ps   %ymm10,%ymm9,%ymm11
   DB  196,66,53,168,216                   ; vfmadd213ps   %ymm8,%ymm9,%ymm11
-  DB  196,98,125,24,5,62,5,0,0            ; vbroadcastss  0x53e(%rip),%ymm8        # 4c98 <_sk_callback_hsw+0x4eb>
+  DB  196,98,125,24,5,74,5,0,0            ; vbroadcastss  0x54a(%rip),%ymm8        # 4d10 <_sk_callback_hsw+0x4f7>
   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
@@ -4327,12 +4359,12 @@ _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  196,226,125,24,13,38,5,0,0          ; vbroadcastss  0x526(%rip),%ymm1        # 4c9c <_sk_callback_hsw+0x4ef>
+  DB  196,226,125,24,13,50,5,0,0          ; vbroadcastss  0x532(%rip),%ymm1        # 4d14 <_sk_callback_hsw+0x4fb>
   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  196,98,125,24,21,18,5,0,0           ; vbroadcastss  0x512(%rip),%ymm10        # 4ca0 <_sk_callback_hsw+0x4f3>
-  DB  196,98,125,24,29,13,5,0,0           ; vbroadcastss  0x50d(%rip),%ymm11        # 4ca4 <_sk_callback_hsw+0x4f7>
+  DB  196,98,125,24,21,30,5,0,0           ; vbroadcastss  0x51e(%rip),%ymm10        # 4d18 <_sk_callback_hsw+0x4ff>
+  DB  196,98,125,24,29,25,5,0,0           ; vbroadcastss  0x519(%rip),%ymm11        # 4d1c <_sk_callback_hsw+0x503>
   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)
@@ -4446,25 +4478,25 @@ 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 4979 <.literal4+0xb1>
+  DB  71,225,61                           ; rex.RXB       loope 49e5 <.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 4989 <.literal4+0xc1>
+  DB  71,225,61                           ; rex.RXB       loope 49f5 <.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 4999 <.literal4+0xd1>
+  DB  71,225,61                           ; rex.RXB       loope 4a05 <.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 49a9 <.literal4+0xe1>
+  DB  71,225,61                           ; rex.RXB       loope 4a15 <.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)
@@ -4511,7 +4543,7 @@ 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        49f9 <.literal4+0x131>
+  DB  224,7                               ; loopne        4a65 <.literal4+0x131>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -4527,10 +4559,10 @@ 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            4a20 <.literal4+0x158>
+  DB  127,0                               ; jg            4a8c <.literal4+0x158>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4a99 <.literal4+0x1d1>
+  DB  119,115                             ; ja            4b05 <.literal4+0x1d1>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4544,10 +4576,10 @@ 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            4a54 <.literal4+0x18c>
+  DB  127,0                               ; jg            4ac0 <.literal4+0x18c>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4acd <.literal4+0x205>
+  DB  119,115                             ; ja            4b39 <.literal4+0x205>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4561,10 +4593,10 @@ 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            4a88 <.literal4+0x1c0>
+  DB  127,0                               ; jg            4af4 <.literal4+0x1c0>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4b01 <.literal4+0x239>
+  DB  119,115                             ; ja            4b6d <.literal4+0x239>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4578,10 +4610,10 @@ 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            4abc <.literal4+0x1f4>
+  DB  127,0                               ; jg            4b28 <.literal4+0x1f4>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4b35 <.literal4+0x26d>
+  DB  119,115                             ; ja            4ba1 <.literal4+0x26d>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -4594,7 +4626,7 @@ 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            4b33 <.literal4+0x26b>
+  DB  127,67                              ; jg            4b9f <.literal4+0x26b>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,195                               ; add           %al,%bl
   DB  0,0                                 ; add           %al,(%rax)
@@ -4606,10 +4638,10 @@ ALIGN 4
   DB  190,80,128,3,62                     ; mov           $0x3e038050,%esi
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           4b53 <.literal4+0x28b>
+  DB  118,63                              ; jbe           4bbf <.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            4b67 <.literal4+0x29f>
+  DB  127,67                              ; jg            4bd3 <.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)
@@ -4618,7 +4650,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        4b49 <.literal4+0x281>
+  DB  224,7                               ; loopne        4bb5 <.literal4+0x281>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -4630,7 +4662,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        4b65 <.literal4+0x29d>
+  DB  224,7                               ; loopne        4bd1 <.literal4+0x29d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -4641,7 +4673,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            4bba <.literal4+0x2f2>
+  DB  124,66                              ; jl            4c26 <.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)
@@ -4659,9 +4691,9 @@ 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            4bfd <.literal4+0x335>
+  DB  112,65                              ; jo            4c69 <.literal4+0x335>
   DB  129,128,128,59,129,128,128,59,0,0   ; addl          $0x3b80,-0x7f7ec480(%rax)
-  DB  127,67                              ; jg            4c0b <.literal4+0x343>
+  DB  127,67                              ; jg            4c77 <.literal4+0x343>
   DB  128,0,128                           ; addb          $0x80,(%rax)
   DB  55                                  ; (bad)
   DB  128,0,128                           ; addb          $0x80,(%rax)
@@ -4669,7 +4701,11 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  255                                 ; (bad)
-  DB  127,71                              ; jg            4c1f <.literal4+0x357>
+  DB  127,71                              ; jg            4c8b <.literal4+0x357>
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,191,0,0,0                     ; add           %al,0xbf(%rax)
+  DB  63                                  ; (bad)
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,89                               ; ds            pop %rcx
@@ -4769,16 +4805,16 @@ 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)        # a004ce8 <_sk_callback_hsw+0xa00053b>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004d48 <_sk_callback_hsw+0xa00052f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004cf0 <_sk_callback_hsw+0x12000543>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004d50 <_sk_callback_hsw+0x12000537>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004cf8 <_sk_callback_hsw+0x1a00054b>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004d58 <_sk_callback_hsw+0x1a00053f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d00 <_sk_callback_hsw+0x3000553>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d60 <_sk_callback_hsw+0x3000547>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -4821,16 +4857,16 @@ 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)        # a004d48 <_sk_callback_hsw+0xa00059b>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004da8 <_sk_callback_hsw+0xa00058f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004d50 <_sk_callback_hsw+0x120005a3>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004db0 <_sk_callback_hsw+0x12000597>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004d58 <_sk_callback_hsw+0x1a0005ab>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004db8 <_sk_callback_hsw+0x1a00059f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004d60 <_sk_callback_hsw+0x30005b3>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004dc0 <_sk_callback_hsw+0x30005a7>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -4873,16 +4909,16 @@ 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)        # a004da8 <_sk_callback_hsw+0xa0005fb>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004e08 <_sk_callback_hsw+0xa0005ef>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004db0 <_sk_callback_hsw+0x12000603>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004e10 <_sk_callback_hsw+0x120005f7>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004db8 <_sk_callback_hsw+0x1a00060b>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004e18 <_sk_callback_hsw+0x1a0005ff>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004dc0 <_sk_callback_hsw+0x3000613>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004e20 <_sk_callback_hsw+0x3000607>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -4925,16 +4961,16 @@ 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)        # a004e08 <_sk_callback_hsw+0xa00065b>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004e68 <_sk_callback_hsw+0xa00064f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004e10 <_sk_callback_hsw+0x12000663>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 12004e70 <_sk_callback_hsw+0x12000657>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004e18 <_sk_callback_hsw+0x1a00066b>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a004e78 <_sk_callback_hsw+0x1a00065f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004e20 <_sk_callback_hsw+0x3000673>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 3004e80 <_sk_callback_hsw+0x3000667>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -5076,14 +5112,14 @@ _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  196,226,125,24,13,152,99,0,0        ; vbroadcastss  0x6398(%rip),%ymm1        # 64f8 <_sk_callback_avx+0x119>
+  DB  196,226,125,24,13,4,100,0,0         ; vbroadcastss  0x6404(%rip),%ymm1        # 6564 <_sk_callback_avx+0x119>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  197,252,88,2                        ; vaddps        (%rdx),%ymm0,%ymm0
   DB  196,226,125,24,16                   ; vbroadcastss  (%rax),%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  197,236,88,201                      ; vaddps        %ymm1,%ymm2,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,21,124,99,0,0        ; vbroadcastss  0x637c(%rip),%ymm2        # 64fc <_sk_callback_avx+0x11d>
+  DB  196,226,125,24,21,232,99,0,0        ; vbroadcastss  0x63e8(%rip),%ymm2        # 6568 <_sk_callback_avx+0x11d>
   DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
   DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
   DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
@@ -5104,7 +5140,7 @@ _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  196,98,125,24,29,45,99,0,0          ; vbroadcastss  0x632d(%rip),%ymm11        # 6500 <_sk_callback_avx+0x121>
+  DB  196,98,125,24,29,153,99,0,0         ; vbroadcastss  0x6399(%rip),%ymm11        # 656c <_sk_callback_avx+0x121>
   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
@@ -5115,8 +5151,8 @@ _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  196,98,125,24,37,238,98,0,0         ; vbroadcastss  0x62ee(%rip),%ymm12        # 6504 <_sk_callback_avx+0x125>
-  DB  196,98,125,24,45,233,98,0,0         ; vbroadcastss  0x62e9(%rip),%ymm13        # 6508 <_sk_callback_avx+0x129>
+  DB  196,98,125,24,37,90,99,0,0          ; vbroadcastss  0x635a(%rip),%ymm12        # 6570 <_sk_callback_avx+0x125>
+  DB  196,98,125,24,45,85,99,0,0          ; vbroadcastss  0x6355(%rip),%ymm13        # 6574 <_sk_callback_avx+0x129>
   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
@@ -5143,9 +5179,9 @@ _sk_dither_avx LABEL PROC
   DB  196,65,12,86,202                    ; vorps         %ymm10,%ymm14,%ymm9
   DB  196,65,60,86,193                    ; vorps         %ymm9,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,84,98,0,0          ; vbroadcastss  0x6254(%rip),%ymm9        # 650c <_sk_callback_avx+0x12d>
+  DB  196,98,125,24,13,192,98,0,0         ; vbroadcastss  0x62c0(%rip),%ymm9        # 6578 <_sk_callback_avx+0x12d>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,74,98,0,0          ; vbroadcastss  0x624a(%rip),%ymm9        # 6510 <_sk_callback_avx+0x131>
+  DB  196,98,125,24,13,182,98,0,0         ; vbroadcastss  0x62b6(%rip),%ymm9        # 657c <_sk_callback_avx+0x131>
   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
@@ -5204,7 +5240,7 @@ _sk_clear_avx LABEL PROC
 PUBLIC _sk_srcatop_avx
 _sk_srcatop_avx LABEL PROC
   DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
-  DB  196,98,125,24,5,161,97,0,0          ; vbroadcastss  0x61a1(%rip),%ymm8        # 6514 <_sk_callback_avx+0x135>
+  DB  196,98,125,24,5,13,98,0,0           ; vbroadcastss  0x620d(%rip),%ymm8        # 6580 <_sk_callback_avx+0x135>
   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
@@ -5223,7 +5259,7 @@ _sk_srcatop_avx LABEL PROC
 PUBLIC _sk_dstatop_avx
 _sk_dstatop_avx LABEL PROC
   DB  197,100,89,196                      ; vmulps        %ymm4,%ymm3,%ymm8
-  DB  196,98,125,24,13,99,97,0,0          ; vbroadcastss  0x6163(%rip),%ymm9        # 6518 <_sk_callback_avx+0x139>
+  DB  196,98,125,24,13,207,97,0,0         ; vbroadcastss  0x61cf(%rip),%ymm9        # 6584 <_sk_callback_avx+0x139>
   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
@@ -5259,7 +5295,7 @@ _sk_dstin_avx LABEL PROC
 
 PUBLIC _sk_srcout_avx
 _sk_srcout_avx LABEL PROC
-  DB  196,98,125,24,5,2,97,0,0            ; vbroadcastss  0x6102(%rip),%ymm8        # 651c <_sk_callback_avx+0x13d>
+  DB  196,98,125,24,5,110,97,0,0          ; vbroadcastss  0x616e(%rip),%ymm8        # 6588 <_sk_callback_avx+0x13d>
   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
@@ -5270,7 +5306,7 @@ _sk_srcout_avx LABEL PROC
 
 PUBLIC _sk_dstout_avx
 _sk_dstout_avx LABEL PROC
-  DB  196,226,125,24,5,229,96,0,0         ; vbroadcastss  0x60e5(%rip),%ymm0        # 6520 <_sk_callback_avx+0x141>
+  DB  196,226,125,24,5,81,97,0,0          ; vbroadcastss  0x6151(%rip),%ymm0        # 658c <_sk_callback_avx+0x141>
   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
@@ -5281,7 +5317,7 @@ _sk_dstout_avx LABEL PROC
 
 PUBLIC _sk_srcover_avx
 _sk_srcover_avx LABEL PROC
-  DB  196,98,125,24,5,200,96,0,0          ; vbroadcastss  0x60c8(%rip),%ymm8        # 6524 <_sk_callback_avx+0x145>
+  DB  196,98,125,24,5,52,97,0,0           ; vbroadcastss  0x6134(%rip),%ymm8        # 6590 <_sk_callback_avx+0x145>
   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
@@ -5296,7 +5332,7 @@ _sk_srcover_avx LABEL PROC
 
 PUBLIC _sk_dstover_avx
 _sk_dstover_avx LABEL PROC
-  DB  196,98,125,24,5,155,96,0,0          ; vbroadcastss  0x609b(%rip),%ymm8        # 6528 <_sk_callback_avx+0x149>
+  DB  196,98,125,24,5,7,97,0,0            ; vbroadcastss  0x6107(%rip),%ymm8        # 6594 <_sk_callback_avx+0x149>
   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
@@ -5320,7 +5356,7 @@ _sk_modulate_avx LABEL PROC
 
 PUBLIC _sk_multiply_avx
 _sk_multiply_avx LABEL PROC
-  DB  196,98,125,24,5,90,96,0,0           ; vbroadcastss  0x605a(%rip),%ymm8        # 652c <_sk_callback_avx+0x14d>
+  DB  196,98,125,24,5,198,96,0,0          ; vbroadcastss  0x60c6(%rip),%ymm8        # 6598 <_sk_callback_avx+0x14d>
   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
@@ -5374,7 +5410,7 @@ _sk_screen_avx LABEL PROC
 
 PUBLIC _sk_xor__avx
 _sk_xor__avx LABEL PROC
-  DB  196,98,125,24,5,169,95,0,0          ; vbroadcastss  0x5fa9(%rip),%ymm8        # 6530 <_sk_callback_avx+0x151>
+  DB  196,98,125,24,5,21,96,0,0           ; vbroadcastss  0x6015(%rip),%ymm8        # 659c <_sk_callback_avx+0x151>
   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
@@ -5409,7 +5445,7 @@ _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  196,98,125,24,5,41,95,0,0           ; vbroadcastss  0x5f29(%rip),%ymm8        # 6534 <_sk_callback_avx+0x155>
+  DB  196,98,125,24,5,149,95,0,0          ; vbroadcastss  0x5f95(%rip),%ymm8        # 65a0 <_sk_callback_avx+0x155>
   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
@@ -5433,7 +5469,7 @@ _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  196,98,125,24,5,213,94,0,0          ; vbroadcastss  0x5ed5(%rip),%ymm8        # 6538 <_sk_callback_avx+0x159>
+  DB  196,98,125,24,5,65,95,0,0           ; vbroadcastss  0x5f41(%rip),%ymm8        # 65a4 <_sk_callback_avx+0x159>
   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
@@ -5460,7 +5496,7 @@ _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  196,98,125,24,5,117,94,0,0          ; vbroadcastss  0x5e75(%rip),%ymm8        # 653c <_sk_callback_avx+0x15d>
+  DB  196,98,125,24,5,225,94,0,0          ; vbroadcastss  0x5ee1(%rip),%ymm8        # 65a8 <_sk_callback_avx+0x15d>
   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
@@ -5481,7 +5517,7 @@ _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  196,98,125,24,5,48,94,0,0           ; vbroadcastss  0x5e30(%rip),%ymm8        # 6540 <_sk_callback_avx+0x161>
+  DB  196,98,125,24,5,156,94,0,0          ; vbroadcastss  0x5e9c(%rip),%ymm8        # 65ac <_sk_callback_avx+0x161>
   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
@@ -5490,7 +5526,7 @@ _sk_exclusion_avx LABEL PROC
 
 PUBLIC _sk_colorburn_avx
 _sk_colorburn_avx LABEL PROC
-  DB  196,98,125,24,5,27,94,0,0           ; vbroadcastss  0x5e1b(%rip),%ymm8        # 6544 <_sk_callback_avx+0x165>
+  DB  196,98,125,24,5,135,94,0,0          ; vbroadcastss  0x5e87(%rip),%ymm8        # 65b0 <_sk_callback_avx+0x165>
   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
@@ -5550,7 +5586,7 @@ _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  196,98,125,24,13,23,93,0,0          ; vbroadcastss  0x5d17(%rip),%ymm9        # 6548 <_sk_callback_avx+0x169>
+  DB  196,98,125,24,13,131,93,0,0         ; vbroadcastss  0x5d83(%rip),%ymm9        # 65b4 <_sk_callback_avx+0x169>
   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
@@ -5605,7 +5641,7 @@ _sk_colordodge_avx LABEL PROC
 
 PUBLIC _sk_hardlight_avx
 _sk_hardlight_avx LABEL PROC
-  DB  196,98,125,24,5,41,92,0,0           ; vbroadcastss  0x5c29(%rip),%ymm8        # 654c <_sk_callback_avx+0x16d>
+  DB  196,98,125,24,5,149,92,0,0          ; vbroadcastss  0x5c95(%rip),%ymm8        # 65b8 <_sk_callback_avx+0x16d>
   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
@@ -5658,7 +5694,7 @@ _sk_hardlight_avx LABEL PROC
 
 PUBLIC _sk_overlay_avx
 _sk_overlay_avx LABEL PROC
-  DB  196,98,125,24,5,82,91,0,0           ; vbroadcastss  0x5b52(%rip),%ymm8        # 6550 <_sk_callback_avx+0x171>
+  DB  196,98,125,24,5,190,91,0,0          ; vbroadcastss  0x5bbe(%rip),%ymm8        # 65bc <_sk_callback_avx+0x171>
   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
@@ -5723,10 +5759,10 @@ _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  196,98,125,24,29,69,90,0,0          ; vbroadcastss  0x5a45(%rip),%ymm11        # 6558 <_sk_callback_avx+0x179>
+  DB  196,98,125,24,29,177,90,0,0         ; vbroadcastss  0x5ab1(%rip),%ymm11        # 65c4 <_sk_callback_avx+0x179>
   DB  196,65,28,88,235                    ; vaddps        %ymm11,%ymm12,%ymm13
   DB  196,65,20,89,192                    ; vmulps        %ymm8,%ymm13,%ymm8
-  DB  196,98,125,24,45,54,90,0,0          ; vbroadcastss  0x5a36(%rip),%ymm13        # 655c <_sk_callback_avx+0x17d>
+  DB  196,98,125,24,45,162,90,0,0         ; vbroadcastss  0x5aa2(%rip),%ymm13        # 65c8 <_sk_callback_avx+0x17d>
   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
@@ -5737,7 +5773,7 @@ _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  196,98,125,24,5,244,89,0,0          ; vbroadcastss  0x59f4(%rip),%ymm8        # 6554 <_sk_callback_avx+0x175>
+  DB  196,98,125,24,5,96,90,0,0           ; vbroadcastss  0x5a60(%rip),%ymm8        # 65c0 <_sk_callback_avx+0x175>
   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
@@ -5864,12 +5900,12 @@ _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  196,98,125,24,53,190,87,0,0         ; vbroadcastss  0x57be(%rip),%ymm14        # 6560 <_sk_callback_avx+0x181>
+  DB  196,98,125,24,53,42,88,0,0          ; vbroadcastss  0x582a(%rip),%ymm14        # 65cc <_sk_callback_avx+0x181>
   DB  196,65,92,89,222                    ; vmulps        %ymm14,%ymm4,%ymm11
-  DB  196,98,125,24,61,180,87,0,0         ; vbroadcastss  0x57b4(%rip),%ymm15        # 6564 <_sk_callback_avx+0x185>
+  DB  196,98,125,24,61,32,88,0,0          ; vbroadcastss  0x5820(%rip),%ymm15        # 65d0 <_sk_callback_avx+0x185>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
-  DB  196,226,125,24,5,165,87,0,0         ; vbroadcastss  0x57a5(%rip),%ymm0        # 6568 <_sk_callback_avx+0x189>
+  DB  196,226,125,24,5,17,88,0,0          ; vbroadcastss  0x5811(%rip),%ymm0        # 65d4 <_sk_callback_avx+0x189>
   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
@@ -5930,7 +5966,7 @@ _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  196,98,125,24,5,126,86,0,0          ; vbroadcastss  0x567e(%rip),%ymm8        # 656c <_sk_callback_avx+0x18d>
+  DB  196,98,125,24,5,234,86,0,0          ; vbroadcastss  0x56ea(%rip),%ymm8        # 65d8 <_sk_callback_avx+0x18d>
   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
@@ -5987,12 +6023,12 @@ _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  196,98,125,24,53,134,85,0,0         ; vbroadcastss  0x5586(%rip),%ymm14        # 6570 <_sk_callback_avx+0x191>
+  DB  196,98,125,24,53,242,85,0,0         ; vbroadcastss  0x55f2(%rip),%ymm14        # 65dc <_sk_callback_avx+0x191>
   DB  196,65,92,89,222                    ; vmulps        %ymm14,%ymm4,%ymm11
-  DB  196,98,125,24,61,124,85,0,0         ; vbroadcastss  0x557c(%rip),%ymm15        # 6574 <_sk_callback_avx+0x195>
+  DB  196,98,125,24,61,232,85,0,0         ; vbroadcastss  0x55e8(%rip),%ymm15        # 65e0 <_sk_callback_avx+0x195>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
-  DB  196,226,125,24,5,109,85,0,0         ; vbroadcastss  0x556d(%rip),%ymm0        # 6578 <_sk_callback_avx+0x199>
+  DB  196,226,125,24,5,217,85,0,0         ; vbroadcastss  0x55d9(%rip),%ymm0        # 65e4 <_sk_callback_avx+0x199>
   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
@@ -6053,7 +6089,7 @@ _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  196,98,125,24,5,70,84,0,0           ; vbroadcastss  0x5446(%rip),%ymm8        # 657c <_sk_callback_avx+0x19d>
+  DB  196,98,125,24,5,178,84,0,0          ; vbroadcastss  0x54b2(%rip),%ymm8        # 65e8 <_sk_callback_avx+0x19d>
   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
@@ -6082,12 +6118,12 @@ _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  196,98,125,24,45,214,83,0,0         ; vbroadcastss  0x53d6(%rip),%ymm13        # 6580 <_sk_callback_avx+0x1a1>
+  DB  196,98,125,24,45,66,84,0,0          ; vbroadcastss  0x5442(%rip),%ymm13        # 65ec <_sk_callback_avx+0x1a1>
   DB  196,65,92,89,213                    ; vmulps        %ymm13,%ymm4,%ymm10
-  DB  196,98,125,24,53,204,83,0,0         ; vbroadcastss  0x53cc(%rip),%ymm14        # 6584 <_sk_callback_avx+0x1a5>
+  DB  196,98,125,24,53,56,84,0,0          ; vbroadcastss  0x5438(%rip),%ymm14        # 65f0 <_sk_callback_avx+0x1a5>
   DB  196,65,84,89,222                    ; vmulps        %ymm14,%ymm5,%ymm11
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,61,189,83,0,0         ; vbroadcastss  0x53bd(%rip),%ymm15        # 6588 <_sk_callback_avx+0x1a9>
+  DB  196,98,125,24,61,41,84,0,0          ; vbroadcastss  0x5429(%rip),%ymm15        # 65f4 <_sk_callback_avx+0x1a9>
   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
@@ -6150,7 +6186,7 @@ _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  196,226,125,24,5,132,82,0,0         ; vbroadcastss  0x5284(%rip),%ymm0        # 658c <_sk_callback_avx+0x1ad>
+  DB  196,226,125,24,5,240,82,0,0         ; vbroadcastss  0x52f0(%rip),%ymm0        # 65f8 <_sk_callback_avx+0x1ad>
   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
@@ -6180,12 +6216,12 @@ _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  196,98,125,24,45,16,82,0,0          ; vbroadcastss  0x5210(%rip),%ymm13        # 6590 <_sk_callback_avx+0x1b1>
+  DB  196,98,125,24,45,124,82,0,0         ; vbroadcastss  0x527c(%rip),%ymm13        # 65fc <_sk_callback_avx+0x1b1>
   DB  196,65,108,89,213                   ; vmulps        %ymm13,%ymm2,%ymm10
-  DB  196,98,125,24,53,6,82,0,0           ; vbroadcastss  0x5206(%rip),%ymm14        # 6594 <_sk_callback_avx+0x1b5>
+  DB  196,98,125,24,53,114,82,0,0         ; vbroadcastss  0x5272(%rip),%ymm14        # 6600 <_sk_callback_avx+0x1b5>
   DB  196,65,116,89,222                   ; vmulps        %ymm14,%ymm1,%ymm11
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,61,247,81,0,0         ; vbroadcastss  0x51f7(%rip),%ymm15        # 6598 <_sk_callback_avx+0x1b9>
+  DB  196,98,125,24,61,99,82,0,0          ; vbroadcastss  0x5263(%rip),%ymm15        # 6604 <_sk_callback_avx+0x1b9>
   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
@@ -6248,7 +6284,7 @@ _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  196,226,125,24,5,190,80,0,0         ; vbroadcastss  0x50be(%rip),%ymm0        # 659c <_sk_callback_avx+0x1bd>
+  DB  196,226,125,24,5,42,81,0,0          ; vbroadcastss  0x512a(%rip),%ymm0        # 6608 <_sk_callback_avx+0x1bd>
   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
@@ -6281,7 +6317,7 @@ _sk_clamp_0_avx LABEL PROC
 
 PUBLIC _sk_clamp_1_avx
 _sk_clamp_1_avx LABEL PROC
-  DB  196,98,125,24,5,78,80,0,0           ; vbroadcastss  0x504e(%rip),%ymm8        # 65a0 <_sk_callback_avx+0x1c1>
+  DB  196,98,125,24,5,186,80,0,0          ; vbroadcastss  0x50ba(%rip),%ymm8        # 660c <_sk_callback_avx+0x1c1>
   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
@@ -6291,7 +6327,7 @@ _sk_clamp_1_avx LABEL PROC
 
 PUBLIC _sk_clamp_a_avx
 _sk_clamp_a_avx LABEL PROC
-  DB  196,98,125,24,5,49,80,0,0           ; vbroadcastss  0x5031(%rip),%ymm8        # 65a4 <_sk_callback_avx+0x1c5>
+  DB  196,98,125,24,5,157,80,0,0          ; vbroadcastss  0x509d(%rip),%ymm8        # 6610 <_sk_callback_avx+0x1c5>
   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
@@ -6363,7 +6399,7 @@ 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  196,98,125,24,21,121,79,0,0         ; vbroadcastss  0x4f79(%rip),%ymm10        # 65a8 <_sk_callback_avx+0x1c9>
+  DB  196,98,125,24,21,229,79,0,0         ; vbroadcastss  0x4fe5(%rip),%ymm10        # 6614 <_sk_callback_avx+0x1c9>
   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
@@ -6374,17 +6410,17 @@ _sk_unpremul_avx LABEL PROC
 
 PUBLIC _sk_from_srgb_avx
 _sk_from_srgb_avx LABEL PROC
-  DB  196,98,125,24,5,90,79,0,0           ; vbroadcastss  0x4f5a(%rip),%ymm8        # 65ac <_sk_callback_avx+0x1cd>
+  DB  196,98,125,24,5,198,79,0,0          ; vbroadcastss  0x4fc6(%rip),%ymm8        # 6618 <_sk_callback_avx+0x1cd>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
-  DB  196,98,125,24,29,76,79,0,0          ; vbroadcastss  0x4f4c(%rip),%ymm11        # 65b0 <_sk_callback_avx+0x1d1>
+  DB  196,98,125,24,29,184,79,0,0         ; vbroadcastss  0x4fb8(%rip),%ymm11        # 661c <_sk_callback_avx+0x1d1>
   DB  196,65,124,89,227                   ; vmulps        %ymm11,%ymm0,%ymm12
-  DB  196,98,125,24,45,66,79,0,0          ; vbroadcastss  0x4f42(%rip),%ymm13        # 65b4 <_sk_callback_avx+0x1d5>
+  DB  196,98,125,24,45,174,79,0,0         ; vbroadcastss  0x4fae(%rip),%ymm13        # 6620 <_sk_callback_avx+0x1d5>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,44,89,212                    ; vmulps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,51,79,0,0          ; vbroadcastss  0x4f33(%rip),%ymm12        # 65b8 <_sk_callback_avx+0x1d9>
+  DB  196,98,125,24,37,159,79,0,0         ; vbroadcastss  0x4f9f(%rip),%ymm12        # 6624 <_sk_callback_avx+0x1d9>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,53,41,79,0,0          ; vbroadcastss  0x4f29(%rip),%ymm14        # 65bc <_sk_callback_avx+0x1dd>
+  DB  196,98,125,24,53,149,79,0,0         ; vbroadcastss  0x4f95(%rip),%ymm14        # 6628 <_sk_callback_avx+0x1dd>
   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
@@ -6409,20 +6445,20 @@ _sk_from_srgb_avx LABEL PROC
 PUBLIC _sk_to_srgb_avx
 _sk_to_srgb_avx LABEL PROC
   DB  197,124,82,200                      ; vrsqrtps      %ymm0,%ymm9
-  DB  196,98,125,24,5,190,78,0,0          ; vbroadcastss  0x4ebe(%rip),%ymm8        # 65c0 <_sk_callback_avx+0x1e1>
+  DB  196,98,125,24,5,42,79,0,0           ; vbroadcastss  0x4f2a(%rip),%ymm8        # 662c <_sk_callback_avx+0x1e1>
   DB  196,65,124,89,208                   ; vmulps        %ymm8,%ymm0,%ymm10
-  DB  196,98,125,24,29,180,78,0,0         ; vbroadcastss  0x4eb4(%rip),%ymm11        # 65c4 <_sk_callback_avx+0x1e5>
+  DB  196,98,125,24,29,32,79,0,0          ; vbroadcastss  0x4f20(%rip),%ymm11        # 6630 <_sk_callback_avx+0x1e5>
   DB  196,65,52,89,227                    ; vmulps        %ymm11,%ymm9,%ymm12
-  DB  196,98,125,24,45,170,78,0,0         ; vbroadcastss  0x4eaa(%rip),%ymm13        # 65c8 <_sk_callback_avx+0x1e9>
+  DB  196,98,125,24,45,22,79,0,0          ; vbroadcastss  0x4f16(%rip),%ymm13        # 6634 <_sk_callback_avx+0x1e9>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,52,89,228                    ; vmulps        %ymm12,%ymm9,%ymm12
-  DB  196,98,125,24,53,155,78,0,0         ; vbroadcastss  0x4e9b(%rip),%ymm14        # 65cc <_sk_callback_avx+0x1ed>
+  DB  196,98,125,24,53,7,79,0,0           ; vbroadcastss  0x4f07(%rip),%ymm14        # 6638 <_sk_callback_avx+0x1ed>
   DB  196,65,28,88,230                    ; vaddps        %ymm14,%ymm12,%ymm12
-  DB  196,98,125,24,61,145,78,0,0         ; vbroadcastss  0x4e91(%rip),%ymm15        # 65d0 <_sk_callback_avx+0x1f1>
+  DB  196,98,125,24,61,253,78,0,0         ; vbroadcastss  0x4efd(%rip),%ymm15        # 663c <_sk_callback_avx+0x1f1>
   DB  196,65,52,88,207                    ; vaddps        %ymm15,%ymm9,%ymm9
   DB  196,65,124,83,201                   ; vrcpps        %ymm9,%ymm9
   DB  196,65,52,89,204                    ; vmulps        %ymm12,%ymm9,%ymm9
-  DB  196,98,125,24,37,125,78,0,0         ; vbroadcastss  0x4e7d(%rip),%ymm12        # 65d4 <_sk_callback_avx+0x1f5>
+  DB  196,98,125,24,37,233,78,0,0         ; vbroadcastss  0x4ee9(%rip),%ymm12        # 6640 <_sk_callback_avx+0x1f5>
   DB  196,193,124,194,196,1               ; vcmpltps      %ymm12,%ymm0,%ymm0
   DB  196,195,53,74,194,0                 ; vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
@@ -6457,7 +6493,7 @@ _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  196,98,125,24,29,227,77,0,0         ; vbroadcastss  0x4de3(%rip),%ymm11        # 65d8 <_sk_callback_avx+0x1f9>
+  DB  196,98,125,24,29,79,78,0,0          ; vbroadcastss  0x4e4f(%rip),%ymm11        # 6644 <_sk_callback_avx+0x1f9>
   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
@@ -6467,19 +6503,19 @@ _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  196,98,125,24,29,188,77,0,0         ; vbroadcastss  0x4dbc(%rip),%ymm11        # 65e4 <_sk_callback_avx+0x205>
+  DB  196,98,125,24,29,40,78,0,0          ; vbroadcastss  0x4e28(%rip),%ymm11        # 6650 <_sk_callback_avx+0x205>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,170,77,0,0         ; vbroadcastss  0x4daa(%rip),%ymm11        # 65e0 <_sk_callback_avx+0x201>
+  DB  196,98,125,24,29,22,78,0,0          ; vbroadcastss  0x4e16(%rip),%ymm11        # 664c <_sk_callback_avx+0x201>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
   DB  196,227,117,74,202,224              ; vblendvps     %ymm14,%ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,146,77,0,0        ; vbroadcastss  0x4d92(%rip),%ymm2        # 65dc <_sk_callback_avx+0x1fd>
+  DB  196,226,125,24,21,254,77,0,0        ; vbroadcastss  0x4dfe(%rip),%ymm2        # 6648 <_sk_callback_avx+0x1fd>
   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  196,98,125,24,37,121,77,0,0         ; vbroadcastss  0x4d79(%rip),%ymm12        # 65ec <_sk_callback_avx+0x20d>
+  DB  196,98,125,24,37,229,77,0,0         ; vbroadcastss  0x4de5(%rip),%ymm12        # 6658 <_sk_callback_avx+0x20d>
   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
@@ -6489,7 +6525,7 @@ _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  196,98,125,24,5,60,77,0,0           ; vbroadcastss  0x4d3c(%rip),%ymm8        # 65e8 <_sk_callback_avx+0x209>
+  DB  196,98,125,24,5,168,77,0,0          ; vbroadcastss  0x4da8(%rip),%ymm8        # 6654 <_sk_callback_avx+0x209>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -6504,7 +6540,7 @@ _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  196,98,125,24,5,3,77,0,0            ; vbroadcastss  0x4d03(%rip),%ymm8        # 65f0 <_sk_callback_avx+0x211>
+  DB  196,98,125,24,5,111,77,0,0          ; vbroadcastss  0x4d6f(%rip),%ymm8        # 665c <_sk_callback_avx+0x211>
   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
@@ -6512,23 +6548,23 @@ _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  196,98,125,24,29,221,76,0,0         ; vbroadcastss  0x4cdd(%rip),%ymm11        # 65f4 <_sk_callback_avx+0x215>
+  DB  196,98,125,24,29,73,77,0,0          ; vbroadcastss  0x4d49(%rip),%ymm11        # 6660 <_sk_callback_avx+0x215>
   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  196,98,125,24,37,199,76,0,0         ; vbroadcastss  0x4cc7(%rip),%ymm12        # 65fc <_sk_callback_avx+0x21d>
+  DB  196,98,125,24,37,51,77,0,0          ; vbroadcastss  0x4d33(%rip),%ymm12        # 6668 <_sk_callback_avx+0x21d>
   DB  196,193,4,89,196                    ; vmulps        %ymm12,%ymm15,%ymm0
-  DB  196,98,125,24,45,189,76,0,0         ; vbroadcastss  0x4cbd(%rip),%ymm13        # 6600 <_sk_callback_avx+0x221>
+  DB  196,98,125,24,45,41,77,0,0          ; vbroadcastss  0x4d29(%rip),%ymm13        # 666c <_sk_callback_avx+0x221>
   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  196,226,125,24,13,158,76,0,0        ; vbroadcastss  0x4c9e(%rip),%ymm1        # 65f8 <_sk_callback_avx+0x219>
+  DB  196,226,125,24,13,10,77,0,0         ; vbroadcastss  0x4d0a(%rip),%ymm1        # 6664 <_sk_callback_avx+0x219>
   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  196,98,125,24,53,137,76,0,0         ; vbroadcastss  0x4c89(%rip),%ymm14        # 6604 <_sk_callback_avx+0x225>
+  DB  196,98,125,24,53,245,76,0,0         ; vbroadcastss  0x4cf5(%rip),%ymm14        # 6670 <_sk_callback_avx+0x225>
   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
@@ -6547,7 +6583,7 @@ _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  196,226,125,24,5,43,76,0,0          ; vbroadcastss  0x4c2b(%rip),%ymm0        # 6608 <_sk_callback_avx+0x229>
+  DB  196,226,125,24,5,151,76,0,0         ; vbroadcastss  0x4c97(%rip),%ymm0        # 6674 <_sk_callback_avx+0x229>
   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
@@ -6602,7 +6638,7 @@ _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  196,98,125,24,13,78,75,0,0          ; vbroadcastss  0x4b4e(%rip),%ymm9        # 660c <_sk_callback_avx+0x22d>
+  DB  196,98,125,24,13,186,75,0,0         ; vbroadcastss  0x4bba(%rip),%ymm9        # 6678 <_sk_callback_avx+0x22d>
   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
@@ -6657,7 +6693,7 @@ _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  196,98,125,24,13,154,74,0,0         ; vbroadcastss  0x4a9a(%rip),%ymm9        # 6610 <_sk_callback_avx+0x231>
+  DB  196,98,125,24,13,6,75,0,0           ; vbroadcastss  0x4b06(%rip),%ymm9        # 667c <_sk_callback_avx+0x231>
   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
@@ -6698,20 +6734,20 @@ _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  196,98,125,24,13,4,74,0,0           ; vbroadcastss  0x4a04(%rip),%ymm9        # 6614 <_sk_callback_avx+0x235>
+  DB  196,98,125,24,13,112,74,0,0         ; vbroadcastss  0x4a70(%rip),%ymm9        # 6680 <_sk_callback_avx+0x235>
   DB  196,65,60,84,201                    ; vandps        %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,245,73,0,0         ; vbroadcastss  0x49f5(%rip),%ymm10        # 6618 <_sk_callback_avx+0x239>
+  DB  196,98,125,24,21,97,74,0,0          ; vbroadcastss  0x4a61(%rip),%ymm10        # 6684 <_sk_callback_avx+0x239>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,21,235,73,0,0         ; vbroadcastss  0x49eb(%rip),%ymm10        # 661c <_sk_callback_avx+0x23d>
+  DB  196,98,125,24,21,87,74,0,0          ; vbroadcastss  0x4a57(%rip),%ymm10        # 6688 <_sk_callback_avx+0x23d>
   DB  196,65,60,84,210                    ; vandps        %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,220,73,0,0         ; vbroadcastss  0x49dc(%rip),%ymm11        # 6620 <_sk_callback_avx+0x241>
+  DB  196,98,125,24,29,72,74,0,0          ; vbroadcastss  0x4a48(%rip),%ymm11        # 668c <_sk_callback_avx+0x241>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,29,210,73,0,0         ; vbroadcastss  0x49d2(%rip),%ymm11        # 6624 <_sk_callback_avx+0x245>
+  DB  196,98,125,24,29,62,74,0,0          ; vbroadcastss  0x4a3e(%rip),%ymm11        # 6690 <_sk_callback_avx+0x245>
   DB  196,65,60,84,195                    ; vandps        %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,195,73,0,0         ; vbroadcastss  0x49c3(%rip),%ymm11        # 6628 <_sk_callback_avx+0x249>
+  DB  196,98,125,24,29,47,74,0,0          ; vbroadcastss  0x4a2f(%rip),%ymm11        # 6694 <_sk_callback_avx+0x249>
   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
@@ -6791,7 +6827,7 @@ _sk_load_tables_avx LABEL PROC
   DB  65,85                               ; push          %r13
   DB  65,84                               ; push          %r12
   DB  83                                  ; push          %rbx
-  DB  197,124,40,13,146,75,0,0            ; vmovaps       0x4b92(%rip),%ymm9        # 6900 <_sk_callback_avx+0x521>
+  DB  197,124,40,13,18,76,0,0             ; vmovaps       0x4c12(%rip),%ymm9        # 6980 <_sk_callback_avx+0x535>
   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
@@ -6883,7 +6919,7 @@ _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  196,98,125,24,5,207,70,0,0          ; vbroadcastss  0x46cf(%rip),%ymm8        # 662c <_sk_callback_avx+0x24d>
+  DB  196,98,125,24,5,59,71,0,0           ; vbroadcastss  0x473b(%rip),%ymm8        # 6698 <_sk_callback_avx+0x24d>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -6973,7 +7009,7 @@ _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  197,121,111,29,210,72,0,0           ; vmovdqa       0x48d2(%rip),%xmm11        # 6980 <_sk_callback_avx+0x5a1>
+  DB  197,121,111,29,82,73,0,0            ; vmovdqa       0x4952(%rip),%xmm11        # 6a00 <_sk_callback_avx+0x5b5>
   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
@@ -7072,7 +7108,7 @@ _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  196,98,125,24,5,128,67,0,0          ; vbroadcastss  0x4380(%rip),%ymm8        # 6630 <_sk_callback_avx+0x251>
+  DB  196,98,125,24,5,236,67,0,0          ; vbroadcastss  0x43ec(%rip),%ymm8        # 669c <_sk_callback_avx+0x251>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -7142,7 +7178,7 @@ _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  197,121,111,13,203,69,0,0           ; vmovdqa       0x45cb(%rip),%xmm9        # 6990 <_sk_callback_avx+0x5b1>
+  DB  197,121,111,13,75,70,0,0            ; vmovdqa       0x464b(%rip),%xmm9        # 6a10 <_sk_callback_avx+0x5c5>
   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
@@ -7234,7 +7270,7 @@ _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  196,226,125,24,29,146,64,0,0        ; vbroadcastss  0x4092(%rip),%ymm3        # 6634 <_sk_callback_avx+0x255>
+  DB  196,226,125,24,29,254,64,0,0        ; vbroadcastss  0x40fe(%rip),%ymm3        # 66a0 <_sk_callback_avx+0x255>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,93                               ; pop           %r13
@@ -7285,7 +7321,7 @@ _sk_byte_tables_avx LABEL PROC
   DB  65,84                               ; push          %r12
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,198,63,0,0          ; vbroadcastss  0x3fc6(%rip),%ymm8        # 6638 <_sk_callback_avx+0x259>
+  DB  196,98,125,24,5,50,64,0,0           ; vbroadcastss  0x4032(%rip),%ymm8        # 66a4 <_sk_callback_avx+0x259>
   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
@@ -7322,7 +7358,7 @@ _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  196,98,125,24,13,20,63,0,0          ; vbroadcastss  0x3f14(%rip),%ymm9        # 663c <_sk_callback_avx+0x25d>
+  DB  196,98,125,24,13,128,63,0,0         ; vbroadcastss  0x3f80(%rip),%ymm9        # 66a8 <_sk_callback_avx+0x25d>
   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
@@ -7482,7 +7518,7 @@ _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  196,98,125,24,13,58,60,0,0          ; vbroadcastss  0x3c3a(%rip),%ymm9        # 6640 <_sk_callback_avx+0x261>
+  DB  196,98,125,24,13,166,60,0,0         ; vbroadcastss  0x3ca6(%rip),%ymm9        # 66ac <_sk_callback_avx+0x261>
   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
@@ -7769,36 +7805,36 @@ _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  196,98,125,24,37,152,55,0,0         ; vbroadcastss  0x3798(%rip),%ymm12        # 6644 <_sk_callback_avx+0x265>
+  DB  196,98,125,24,37,4,56,0,0           ; vbroadcastss  0x3804(%rip),%ymm12        # 66b0 <_sk_callback_avx+0x265>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,142,55,0,0         ; vbroadcastss  0x378e(%rip),%ymm12        # 6648 <_sk_callback_avx+0x269>
+  DB  196,98,125,24,37,250,55,0,0         ; vbroadcastss  0x37fa(%rip),%ymm12        # 66b4 <_sk_callback_avx+0x269>
   DB  196,193,124,84,196                  ; vandps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,132,55,0,0         ; vbroadcastss  0x3784(%rip),%ymm12        # 664c <_sk_callback_avx+0x26d>
+  DB  196,98,125,24,37,240,55,0,0         ; vbroadcastss  0x37f0(%rip),%ymm12        # 66b8 <_sk_callback_avx+0x26d>
   DB  196,193,124,86,196                  ; vorps         %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,122,55,0,0         ; vbroadcastss  0x377a(%rip),%ymm12        # 6650 <_sk_callback_avx+0x271>
+  DB  196,98,125,24,37,230,55,0,0         ; vbroadcastss  0x37e6(%rip),%ymm12        # 66bc <_sk_callback_avx+0x271>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,112,55,0,0         ; vbroadcastss  0x3770(%rip),%ymm12        # 6654 <_sk_callback_avx+0x275>
+  DB  196,98,125,24,37,220,55,0,0         ; vbroadcastss  0x37dc(%rip),%ymm12        # 66c0 <_sk_callback_avx+0x275>
   DB  196,65,124,89,228                   ; vmulps        %ymm12,%ymm0,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,97,55,0,0          ; vbroadcastss  0x3761(%rip),%ymm12        # 6658 <_sk_callback_avx+0x279>
+  DB  196,98,125,24,37,205,55,0,0         ; vbroadcastss  0x37cd(%rip),%ymm12        # 66c4 <_sk_callback_avx+0x279>
   DB  196,193,124,88,196                  ; vaddps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,87,55,0,0          ; vbroadcastss  0x3757(%rip),%ymm12        # 665c <_sk_callback_avx+0x27d>
+  DB  196,98,125,24,37,195,55,0,0         ; vbroadcastss  0x37c3(%rip),%ymm12        # 66c8 <_sk_callback_avx+0x27d>
   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  196,98,125,24,29,59,55,0,0          ; vbroadcastss  0x373b(%rip),%ymm11        # 6660 <_sk_callback_avx+0x281>
+  DB  196,98,125,24,29,167,55,0,0         ; vbroadcastss  0x37a7(%rip),%ymm11        # 66cc <_sk_callback_avx+0x281>
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,49,55,0,0          ; vbroadcastss  0x3731(%rip),%ymm11        # 6664 <_sk_callback_avx+0x285>
+  DB  196,98,125,24,29,157,55,0,0         ; vbroadcastss  0x379d(%rip),%ymm11        # 66d0 <_sk_callback_avx+0x285>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,124,92,195                  ; vsubps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,34,55,0,0          ; vbroadcastss  0x3722(%rip),%ymm11        # 6668 <_sk_callback_avx+0x289>
+  DB  196,98,125,24,29,142,55,0,0         ; vbroadcastss  0x378e(%rip),%ymm11        # 66d4 <_sk_callback_avx+0x289>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,24,55,0,0          ; vbroadcastss  0x3718(%rip),%ymm11        # 666c <_sk_callback_avx+0x28d>
+  DB  196,98,125,24,29,132,55,0,0         ; vbroadcastss  0x3784(%rip),%ymm11        # 66d8 <_sk_callback_avx+0x28d>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
-  DB  196,98,125,24,21,9,55,0,0           ; vbroadcastss  0x3709(%rip),%ymm10        # 6670 <_sk_callback_avx+0x291>
+  DB  196,98,125,24,21,117,55,0,0         ; vbroadcastss  0x3775(%rip),%ymm10        # 66dc <_sk_callback_avx+0x291>
   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
@@ -7806,7 +7842,7 @@ _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  196,98,125,24,5,224,54,0,0          ; vbroadcastss  0x36e0(%rip),%ymm8        # 6674 <_sk_callback_avx+0x295>
+  DB  196,98,125,24,5,76,55,0,0           ; vbroadcastss  0x374c(%rip),%ymm8        # 66e0 <_sk_callback_avx+0x295>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -7826,36 +7862,36 @@ _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  196,98,125,24,37,145,54,0,0         ; vbroadcastss  0x3691(%rip),%ymm12        # 6678 <_sk_callback_avx+0x299>
+  DB  196,98,125,24,37,253,54,0,0         ; vbroadcastss  0x36fd(%rip),%ymm12        # 66e4 <_sk_callback_avx+0x299>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,135,54,0,0         ; vbroadcastss  0x3687(%rip),%ymm12        # 667c <_sk_callback_avx+0x29d>
+  DB  196,98,125,24,37,243,54,0,0         ; vbroadcastss  0x36f3(%rip),%ymm12        # 66e8 <_sk_callback_avx+0x29d>
   DB  196,193,116,84,204                  ; vandps        %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,125,54,0,0         ; vbroadcastss  0x367d(%rip),%ymm12        # 6680 <_sk_callback_avx+0x2a1>
+  DB  196,98,125,24,37,233,54,0,0         ; vbroadcastss  0x36e9(%rip),%ymm12        # 66ec <_sk_callback_avx+0x2a1>
   DB  196,193,116,86,204                  ; vorps         %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,115,54,0,0         ; vbroadcastss  0x3673(%rip),%ymm12        # 6684 <_sk_callback_avx+0x2a5>
+  DB  196,98,125,24,37,223,54,0,0         ; vbroadcastss  0x36df(%rip),%ymm12        # 66f0 <_sk_callback_avx+0x2a5>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,105,54,0,0         ; vbroadcastss  0x3669(%rip),%ymm12        # 6688 <_sk_callback_avx+0x2a9>
+  DB  196,98,125,24,37,213,54,0,0         ; vbroadcastss  0x36d5(%rip),%ymm12        # 66f4 <_sk_callback_avx+0x2a9>
   DB  196,65,116,89,228                   ; vmulps        %ymm12,%ymm1,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,90,54,0,0          ; vbroadcastss  0x365a(%rip),%ymm12        # 668c <_sk_callback_avx+0x2ad>
+  DB  196,98,125,24,37,198,54,0,0         ; vbroadcastss  0x36c6(%rip),%ymm12        # 66f8 <_sk_callback_avx+0x2ad>
   DB  196,193,116,88,204                  ; vaddps        %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,80,54,0,0          ; vbroadcastss  0x3650(%rip),%ymm12        # 6690 <_sk_callback_avx+0x2b1>
+  DB  196,98,125,24,37,188,54,0,0         ; vbroadcastss  0x36bc(%rip),%ymm12        # 66fc <_sk_callback_avx+0x2b1>
   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  196,98,125,24,29,52,54,0,0          ; vbroadcastss  0x3634(%rip),%ymm11        # 6694 <_sk_callback_avx+0x2b5>
+  DB  196,98,125,24,29,160,54,0,0         ; vbroadcastss  0x36a0(%rip),%ymm11        # 6700 <_sk_callback_avx+0x2b5>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,42,54,0,0          ; vbroadcastss  0x362a(%rip),%ymm11        # 6698 <_sk_callback_avx+0x2b9>
+  DB  196,98,125,24,29,150,54,0,0         ; vbroadcastss  0x3696(%rip),%ymm11        # 6704 <_sk_callback_avx+0x2b9>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,116,92,203                  ; vsubps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,27,54,0,0          ; vbroadcastss  0x361b(%rip),%ymm11        # 669c <_sk_callback_avx+0x2bd>
+  DB  196,98,125,24,29,135,54,0,0         ; vbroadcastss  0x3687(%rip),%ymm11        # 6708 <_sk_callback_avx+0x2bd>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,17,54,0,0          ; vbroadcastss  0x3611(%rip),%ymm11        # 66a0 <_sk_callback_avx+0x2c1>
+  DB  196,98,125,24,29,125,54,0,0         ; vbroadcastss  0x367d(%rip),%ymm11        # 670c <_sk_callback_avx+0x2c1>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,21,2,54,0,0           ; vbroadcastss  0x3602(%rip),%ymm10        # 66a4 <_sk_callback_avx+0x2c5>
+  DB  196,98,125,24,21,110,54,0,0         ; vbroadcastss  0x366e(%rip),%ymm10        # 6710 <_sk_callback_avx+0x2c5>
   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
@@ -7863,7 +7899,7 @@ _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  196,98,125,24,5,217,53,0,0          ; vbroadcastss  0x35d9(%rip),%ymm8        # 66a8 <_sk_callback_avx+0x2c9>
+  DB  196,98,125,24,5,69,54,0,0           ; vbroadcastss  0x3645(%rip),%ymm8        # 6714 <_sk_callback_avx+0x2c9>
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -7883,36 +7919,36 @@ _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  196,98,125,24,37,138,53,0,0         ; vbroadcastss  0x358a(%rip),%ymm12        # 66ac <_sk_callback_avx+0x2cd>
+  DB  196,98,125,24,37,246,53,0,0         ; vbroadcastss  0x35f6(%rip),%ymm12        # 6718 <_sk_callback_avx+0x2cd>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,128,53,0,0         ; vbroadcastss  0x3580(%rip),%ymm12        # 66b0 <_sk_callback_avx+0x2d1>
+  DB  196,98,125,24,37,236,53,0,0         ; vbroadcastss  0x35ec(%rip),%ymm12        # 671c <_sk_callback_avx+0x2d1>
   DB  196,193,108,84,212                  ; vandps        %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,118,53,0,0         ; vbroadcastss  0x3576(%rip),%ymm12        # 66b4 <_sk_callback_avx+0x2d5>
+  DB  196,98,125,24,37,226,53,0,0         ; vbroadcastss  0x35e2(%rip),%ymm12        # 6720 <_sk_callback_avx+0x2d5>
   DB  196,193,108,86,212                  ; vorps         %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,108,53,0,0         ; vbroadcastss  0x356c(%rip),%ymm12        # 66b8 <_sk_callback_avx+0x2d9>
+  DB  196,98,125,24,37,216,53,0,0         ; vbroadcastss  0x35d8(%rip),%ymm12        # 6724 <_sk_callback_avx+0x2d9>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,98,53,0,0          ; vbroadcastss  0x3562(%rip),%ymm12        # 66bc <_sk_callback_avx+0x2dd>
+  DB  196,98,125,24,37,206,53,0,0         ; vbroadcastss  0x35ce(%rip),%ymm12        # 6728 <_sk_callback_avx+0x2dd>
   DB  196,65,108,89,228                   ; vmulps        %ymm12,%ymm2,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,83,53,0,0          ; vbroadcastss  0x3553(%rip),%ymm12        # 66c0 <_sk_callback_avx+0x2e1>
+  DB  196,98,125,24,37,191,53,0,0         ; vbroadcastss  0x35bf(%rip),%ymm12        # 672c <_sk_callback_avx+0x2e1>
   DB  196,193,108,88,212                  ; vaddps        %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,73,53,0,0          ; vbroadcastss  0x3549(%rip),%ymm12        # 66c4 <_sk_callback_avx+0x2e5>
+  DB  196,98,125,24,37,181,53,0,0         ; vbroadcastss  0x35b5(%rip),%ymm12        # 6730 <_sk_callback_avx+0x2e5>
   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  196,98,125,24,29,45,53,0,0          ; vbroadcastss  0x352d(%rip),%ymm11        # 66c8 <_sk_callback_avx+0x2e9>
+  DB  196,98,125,24,29,153,53,0,0         ; vbroadcastss  0x3599(%rip),%ymm11        # 6734 <_sk_callback_avx+0x2e9>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,35,53,0,0          ; vbroadcastss  0x3523(%rip),%ymm11        # 66cc <_sk_callback_avx+0x2ed>
+  DB  196,98,125,24,29,143,53,0,0         ; vbroadcastss  0x358f(%rip),%ymm11        # 6738 <_sk_callback_avx+0x2ed>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,108,92,211                  ; vsubps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,20,53,0,0          ; vbroadcastss  0x3514(%rip),%ymm11        # 66d0 <_sk_callback_avx+0x2f1>
+  DB  196,98,125,24,29,128,53,0,0         ; vbroadcastss  0x3580(%rip),%ymm11        # 673c <_sk_callback_avx+0x2f1>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,10,53,0,0          ; vbroadcastss  0x350a(%rip),%ymm11        # 66d4 <_sk_callback_avx+0x2f5>
+  DB  196,98,125,24,29,118,53,0,0         ; vbroadcastss  0x3576(%rip),%ymm11        # 6740 <_sk_callback_avx+0x2f5>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
-  DB  196,98,125,24,21,251,52,0,0         ; vbroadcastss  0x34fb(%rip),%ymm10        # 66d8 <_sk_callback_avx+0x2f9>
+  DB  196,98,125,24,21,103,53,0,0         ; vbroadcastss  0x3567(%rip),%ymm10        # 6744 <_sk_callback_avx+0x2f9>
   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
@@ -7920,7 +7956,7 @@ _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  196,98,125,24,5,210,52,0,0          ; vbroadcastss  0x34d2(%rip),%ymm8        # 66dc <_sk_callback_avx+0x2fd>
+  DB  196,98,125,24,5,62,53,0,0           ; vbroadcastss  0x353e(%rip),%ymm8        # 6748 <_sk_callback_avx+0x2fd>
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -7940,36 +7976,36 @@ _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  196,98,125,24,37,131,52,0,0         ; vbroadcastss  0x3483(%rip),%ymm12        # 66e0 <_sk_callback_avx+0x301>
+  DB  196,98,125,24,37,239,52,0,0         ; vbroadcastss  0x34ef(%rip),%ymm12        # 674c <_sk_callback_avx+0x301>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,121,52,0,0         ; vbroadcastss  0x3479(%rip),%ymm12        # 66e4 <_sk_callback_avx+0x305>
+  DB  196,98,125,24,37,229,52,0,0         ; vbroadcastss  0x34e5(%rip),%ymm12        # 6750 <_sk_callback_avx+0x305>
   DB  196,193,100,84,220                  ; vandps        %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,111,52,0,0         ; vbroadcastss  0x346f(%rip),%ymm12        # 66e8 <_sk_callback_avx+0x309>
+  DB  196,98,125,24,37,219,52,0,0         ; vbroadcastss  0x34db(%rip),%ymm12        # 6754 <_sk_callback_avx+0x309>
   DB  196,193,100,86,220                  ; vorps         %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,101,52,0,0         ; vbroadcastss  0x3465(%rip),%ymm12        # 66ec <_sk_callback_avx+0x30d>
+  DB  196,98,125,24,37,209,52,0,0         ; vbroadcastss  0x34d1(%rip),%ymm12        # 6758 <_sk_callback_avx+0x30d>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,91,52,0,0          ; vbroadcastss  0x345b(%rip),%ymm12        # 66f0 <_sk_callback_avx+0x311>
+  DB  196,98,125,24,37,199,52,0,0         ; vbroadcastss  0x34c7(%rip),%ymm12        # 675c <_sk_callback_avx+0x311>
   DB  196,65,100,89,228                   ; vmulps        %ymm12,%ymm3,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,76,52,0,0          ; vbroadcastss  0x344c(%rip),%ymm12        # 66f4 <_sk_callback_avx+0x315>
+  DB  196,98,125,24,37,184,52,0,0         ; vbroadcastss  0x34b8(%rip),%ymm12        # 6760 <_sk_callback_avx+0x315>
   DB  196,193,100,88,220                  ; vaddps        %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,66,52,0,0          ; vbroadcastss  0x3442(%rip),%ymm12        # 66f8 <_sk_callback_avx+0x319>
+  DB  196,98,125,24,37,174,52,0,0         ; vbroadcastss  0x34ae(%rip),%ymm12        # 6764 <_sk_callback_avx+0x319>
   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  196,98,125,24,29,38,52,0,0          ; vbroadcastss  0x3426(%rip),%ymm11        # 66fc <_sk_callback_avx+0x31d>
+  DB  196,98,125,24,29,146,52,0,0         ; vbroadcastss  0x3492(%rip),%ymm11        # 6768 <_sk_callback_avx+0x31d>
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,28,52,0,0          ; vbroadcastss  0x341c(%rip),%ymm11        # 6700 <_sk_callback_avx+0x321>
+  DB  196,98,125,24,29,136,52,0,0         ; vbroadcastss  0x3488(%rip),%ymm11        # 676c <_sk_callback_avx+0x321>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,100,92,219                  ; vsubps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,13,52,0,0          ; vbroadcastss  0x340d(%rip),%ymm11        # 6704 <_sk_callback_avx+0x325>
+  DB  196,98,125,24,29,121,52,0,0         ; vbroadcastss  0x3479(%rip),%ymm11        # 6770 <_sk_callback_avx+0x325>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,3,52,0,0           ; vbroadcastss  0x3403(%rip),%ymm11        # 6708 <_sk_callback_avx+0x329>
+  DB  196,98,125,24,29,111,52,0,0         ; vbroadcastss  0x346f(%rip),%ymm11        # 6774 <_sk_callback_avx+0x329>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,100,88,218                  ; vaddps        %ymm10,%ymm3,%ymm3
-  DB  196,98,125,24,21,244,51,0,0         ; vbroadcastss  0x33f4(%rip),%ymm10        # 670c <_sk_callback_avx+0x32d>
+  DB  196,98,125,24,21,96,52,0,0          ; vbroadcastss  0x3460(%rip),%ymm10        # 6778 <_sk_callback_avx+0x32d>
   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
@@ -7977,38 +8013,38 @@ _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  196,98,125,24,5,203,51,0,0          ; vbroadcastss  0x33cb(%rip),%ymm8        # 6710 <_sk_callback_avx+0x331>
+  DB  196,98,125,24,5,55,52,0,0           ; vbroadcastss  0x3437(%rip),%ymm8        # 677c <_sk_callback_avx+0x331>
   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  196,98,125,24,5,189,51,0,0          ; vbroadcastss  0x33bd(%rip),%ymm8        # 6714 <_sk_callback_avx+0x335>
+  DB  196,98,125,24,5,41,52,0,0           ; vbroadcastss  0x3429(%rip),%ymm8        # 6780 <_sk_callback_avx+0x335>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,179,51,0,0          ; vbroadcastss  0x33b3(%rip),%ymm8        # 6718 <_sk_callback_avx+0x339>
+  DB  196,98,125,24,5,31,52,0,0           ; vbroadcastss  0x341f(%rip),%ymm8        # 6784 <_sk_callback_avx+0x339>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,98,125,24,13,169,51,0,0         ; vbroadcastss  0x33a9(%rip),%ymm9        # 671c <_sk_callback_avx+0x33d>
+  DB  196,98,125,24,13,21,52,0,0          ; vbroadcastss  0x3415(%rip),%ymm9        # 6788 <_sk_callback_avx+0x33d>
   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  196,98,125,24,5,149,51,0,0          ; vbroadcastss  0x3395(%rip),%ymm8        # 6720 <_sk_callback_avx+0x341>
+  DB  196,98,125,24,5,1,52,0,0            ; vbroadcastss  0x3401(%rip),%ymm8        # 678c <_sk_callback_avx+0x341>
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,139,51,0,0          ; vbroadcastss  0x338b(%rip),%ymm8        # 6724 <_sk_callback_avx+0x345>
+  DB  196,98,125,24,5,247,51,0,0          ; vbroadcastss  0x33f7(%rip),%ymm8        # 6790 <_sk_callback_avx+0x345>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,129,51,0,0          ; vbroadcastss  0x3381(%rip),%ymm8        # 6728 <_sk_callback_avx+0x349>
+  DB  196,98,125,24,5,237,51,0,0          ; vbroadcastss  0x33ed(%rip),%ymm8        # 6794 <_sk_callback_avx+0x349>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  197,252,88,201                      ; vaddps        %ymm1,%ymm0,%ymm1
-  DB  196,98,125,24,5,115,51,0,0          ; vbroadcastss  0x3373(%rip),%ymm8        # 672c <_sk_callback_avx+0x34d>
+  DB  196,98,125,24,5,223,51,0,0          ; vbroadcastss  0x33df(%rip),%ymm8        # 6798 <_sk_callback_avx+0x34d>
   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  196,98,125,24,13,92,51,0,0          ; vbroadcastss  0x335c(%rip),%ymm9        # 6730 <_sk_callback_avx+0x351>
+  DB  196,98,125,24,13,200,51,0,0         ; vbroadcastss  0x33c8(%rip),%ymm9        # 679c <_sk_callback_avx+0x351>
   DB  196,65,52,194,208,1                 ; vcmpltps      %ymm8,%ymm9,%ymm10
-  DB  196,98,125,24,29,81,51,0,0          ; vbroadcastss  0x3351(%rip),%ymm11        # 6734 <_sk_callback_avx+0x355>
+  DB  196,98,125,24,29,189,51,0,0         ; vbroadcastss  0x33bd(%rip),%ymm11        # 67a0 <_sk_callback_avx+0x355>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,37,71,51,0,0          ; vbroadcastss  0x3347(%rip),%ymm12        # 6738 <_sk_callback_avx+0x359>
+  DB  196,98,125,24,37,179,51,0,0         ; vbroadcastss  0x33b3(%rip),%ymm12        # 67a4 <_sk_callback_avx+0x359>
   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
@@ -8023,9 +8059,9 @@ _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  196,226,125,24,5,253,50,0,0         ; vbroadcastss  0x32fd(%rip),%ymm0        # 673c <_sk_callback_avx+0x35d>
+  DB  196,226,125,24,5,105,51,0,0         ; vbroadcastss  0x3369(%rip),%ymm0        # 67a8 <_sk_callback_avx+0x35d>
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,244,50,0,0          ; vbroadcastss  0x32f4(%rip),%ymm8        # 6740 <_sk_callback_avx+0x361>
+  DB  196,98,125,24,5,96,51,0,0           ; vbroadcastss  0x3360(%rip),%ymm8        # 67ac <_sk_callback_avx+0x361>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -8044,7 +8080,7 @@ _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  196,226,125,24,13,184,50,0,0        ; vbroadcastss  0x32b8(%rip),%ymm1        # 6744 <_sk_callback_avx+0x365>
+  DB  196,226,125,24,13,36,51,0,0         ; vbroadcastss  0x3324(%rip),%ymm1        # 67b0 <_sk_callback_avx+0x365>
   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
@@ -8111,7 +8147,7 @@ _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  196,226,125,24,13,173,49,0,0        ; vbroadcastss  0x31ad(%rip),%ymm1        # 6748 <_sk_callback_avx+0x369>
+  DB  196,226,125,24,13,25,50,0,0         ; vbroadcastss  0x3219(%rip),%ymm1        # 67b4 <_sk_callback_avx+0x369>
   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
@@ -8127,7 +8163,7 @@ 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  196,98,125,24,5,136,49,0,0          ; vbroadcastss  0x3188(%rip),%ymm8        # 674c <_sk_callback_avx+0x36d>
+  DB  196,98,125,24,5,244,49,0,0          ; vbroadcastss  0x31f4(%rip),%ymm8        # 67b8 <_sk_callback_avx+0x36d>
   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
@@ -8195,10 +8231,10 @@ _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  196,226,125,24,13,173,48,0,0        ; vbroadcastss  0x30ad(%rip),%ymm1        # 6750 <_sk_callback_avx+0x371>
+  DB  196,226,125,24,13,25,49,0,0         ; vbroadcastss  0x3119(%rip),%ymm1        # 67bc <_sk_callback_avx+0x371>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,162,48,0,0        ; vbroadcastss  0x30a2(%rip),%ymm3        # 6754 <_sk_callback_avx+0x375>
+  DB  196,226,125,24,29,14,49,0,0         ; vbroadcastss  0x310e(%rip),%ymm3        # 67c0 <_sk_callback_avx+0x375>
   DB  76,137,193                          ; mov           %r8,%rcx
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
@@ -8262,10 +8298,10 @@ _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  196,226,125,24,13,161,47,0,0        ; vbroadcastss  0x2fa1(%rip),%ymm1        # 6758 <_sk_callback_avx+0x379>
+  DB  196,226,125,24,13,13,48,0,0         ; vbroadcastss  0x300d(%rip),%ymm1        # 67c4 <_sk_callback_avx+0x379>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,150,47,0,0        ; vbroadcastss  0x2f96(%rip),%ymm3        # 675c <_sk_callback_avx+0x37d>
+  DB  196,226,125,24,29,2,48,0,0          ; vbroadcastss  0x3002(%rip),%ymm3        # 67c8 <_sk_callback_avx+0x37d>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  91                                  ; pop           %rbx
@@ -8343,10 +8379,10 @@ _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  197,124,40,21,14,48,0,0             ; vmovaps       0x300e(%rip),%ymm10        # 6920 <_sk_callback_avx+0x541>
+  DB  197,124,40,21,142,48,0,0            ; vmovaps       0x308e(%rip),%ymm10        # 69a0 <_sk_callback_avx+0x555>
   DB  196,193,124,84,194                  ; vandps        %ymm10,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,60,46,0,0          ; vbroadcastss  0x2e3c(%rip),%ymm9        # 6760 <_sk_callback_avx+0x381>
+  DB  196,98,125,24,13,168,46,0,0         ; vbroadcastss  0x2ea8(%rip),%ymm9        # 67cc <_sk_callback_avx+0x381>
   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
@@ -8384,23 +8420,23 @@ _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  196,226,125,24,5,166,45,0,0         ; vbroadcastss  0x2da6(%rip),%ymm0        # 6764 <_sk_callback_avx+0x385>
+  DB  196,226,125,24,5,18,46,0,0          ; vbroadcastss  0x2e12(%rip),%ymm0        # 67d0 <_sk_callback_avx+0x385>
   DB  197,236,84,192                      ; vandps        %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,153,45,0,0        ; vbroadcastss  0x2d99(%rip),%ymm1        # 6768 <_sk_callback_avx+0x389>
+  DB  196,226,125,24,13,5,46,0,0          ; vbroadcastss  0x2e05(%rip),%ymm1        # 67d4 <_sk_callback_avx+0x389>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,144,45,0,0        ; vbroadcastss  0x2d90(%rip),%ymm1        # 676c <_sk_callback_avx+0x38d>
+  DB  196,226,125,24,13,252,45,0,0        ; vbroadcastss  0x2dfc(%rip),%ymm1        # 67d8 <_sk_callback_avx+0x38d>
   DB  197,236,84,201                      ; vandps        %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,131,45,0,0        ; vbroadcastss  0x2d83(%rip),%ymm3        # 6770 <_sk_callback_avx+0x391>
+  DB  196,226,125,24,29,239,45,0,0        ; vbroadcastss  0x2def(%rip),%ymm3        # 67dc <_sk_callback_avx+0x391>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,24,29,122,45,0,0        ; vbroadcastss  0x2d7a(%rip),%ymm3        # 6774 <_sk_callback_avx+0x395>
+  DB  196,226,125,24,29,230,45,0,0        ; vbroadcastss  0x2de6(%rip),%ymm3        # 67e0 <_sk_callback_avx+0x395>
   DB  197,236,84,211                      ; vandps        %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,109,45,0,0        ; vbroadcastss  0x2d6d(%rip),%ymm3        # 6778 <_sk_callback_avx+0x399>
+  DB  196,226,125,24,29,217,45,0,0        ; vbroadcastss  0x2dd9(%rip),%ymm3        # 67e4 <_sk_callback_avx+0x399>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,98,45,0,0         ; vbroadcastss  0x2d62(%rip),%ymm3        # 677c <_sk_callback_avx+0x39d>
+  DB  196,226,125,24,29,206,45,0,0        ; vbroadcastss  0x2dce(%rip),%ymm3        # 67e8 <_sk_callback_avx+0x39d>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,200                          ; mov           %ecx,%r8d
   DB  65,128,224,7                        ; and           $0x7,%r8b
@@ -8497,23 +8533,23 @@ _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  196,226,125,24,5,2,44,0,0           ; vbroadcastss  0x2c02(%rip),%ymm0        # 6780 <_sk_callback_avx+0x3a1>
+  DB  196,226,125,24,5,110,44,0,0         ; vbroadcastss  0x2c6e(%rip),%ymm0        # 67ec <_sk_callback_avx+0x3a1>
   DB  197,236,84,192                      ; vandps        %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,245,43,0,0        ; vbroadcastss  0x2bf5(%rip),%ymm1        # 6784 <_sk_callback_avx+0x3a5>
+  DB  196,226,125,24,13,97,44,0,0         ; vbroadcastss  0x2c61(%rip),%ymm1        # 67f0 <_sk_callback_avx+0x3a5>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,236,43,0,0        ; vbroadcastss  0x2bec(%rip),%ymm1        # 6788 <_sk_callback_avx+0x3a9>
+  DB  196,226,125,24,13,88,44,0,0         ; vbroadcastss  0x2c58(%rip),%ymm1        # 67f4 <_sk_callback_avx+0x3a9>
   DB  197,236,84,201                      ; vandps        %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,223,43,0,0        ; vbroadcastss  0x2bdf(%rip),%ymm3        # 678c <_sk_callback_avx+0x3ad>
+  DB  196,226,125,24,29,75,44,0,0         ; vbroadcastss  0x2c4b(%rip),%ymm3        # 67f8 <_sk_callback_avx+0x3ad>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,24,29,214,43,0,0        ; vbroadcastss  0x2bd6(%rip),%ymm3        # 6790 <_sk_callback_avx+0x3b1>
+  DB  196,226,125,24,29,66,44,0,0         ; vbroadcastss  0x2c42(%rip),%ymm3        # 67fc <_sk_callback_avx+0x3b1>
   DB  197,236,84,211                      ; vandps        %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,201,43,0,0        ; vbroadcastss  0x2bc9(%rip),%ymm3        # 6794 <_sk_callback_avx+0x3b5>
+  DB  196,226,125,24,29,53,44,0,0         ; vbroadcastss  0x2c35(%rip),%ymm3        # 6800 <_sk_callback_avx+0x3b5>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,190,43,0,0        ; vbroadcastss  0x2bbe(%rip),%ymm3        # 6798 <_sk_callback_avx+0x3b9>
+  DB  196,226,125,24,29,42,44,0,0         ; vbroadcastss  0x2c2a(%rip),%ymm3        # 6804 <_sk_callback_avx+0x3b9>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,94                               ; pop           %r14
@@ -8525,14 +8561,14 @@ 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  196,98,125,24,5,170,43,0,0          ; vbroadcastss  0x2baa(%rip),%ymm8        # 679c <_sk_callback_avx+0x3bd>
+  DB  196,98,125,24,5,22,44,0,0           ; vbroadcastss  0x2c16(%rip),%ymm8        # 6808 <_sk_callback_avx+0x3bd>
   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  196,98,125,24,21,131,43,0,0         ; vbroadcastss  0x2b83(%rip),%ymm10        # 67a0 <_sk_callback_avx+0x3c1>
+  DB  196,98,125,24,21,239,43,0,0         ; vbroadcastss  0x2bef(%rip),%ymm10        # 680c <_sk_callback_avx+0x3c1>
   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
@@ -8604,25 +8640,25 @@ _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  196,226,125,24,5,140,42,0,0         ; vbroadcastss  0x2a8c(%rip),%ymm0        # 67a4 <_sk_callback_avx+0x3c5>
+  DB  196,226,125,24,5,248,42,0,0         ; vbroadcastss  0x2af8(%rip),%ymm0        # 6810 <_sk_callback_avx+0x3c5>
   DB  197,228,84,192                      ; vandps        %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,127,42,0,0        ; vbroadcastss  0x2a7f(%rip),%ymm1        # 67a8 <_sk_callback_avx+0x3c9>
+  DB  196,226,125,24,13,235,42,0,0        ; vbroadcastss  0x2aeb(%rip),%ymm1        # 6814 <_sk_callback_avx+0x3c9>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,118,42,0,0        ; vbroadcastss  0x2a76(%rip),%ymm1        # 67ac <_sk_callback_avx+0x3cd>
+  DB  196,226,125,24,13,226,42,0,0        ; vbroadcastss  0x2ae2(%rip),%ymm1        # 6818 <_sk_callback_avx+0x3cd>
   DB  197,228,84,201                      ; vandps        %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,105,42,0,0        ; vbroadcastss  0x2a69(%rip),%ymm2        # 67b0 <_sk_callback_avx+0x3d1>
+  DB  196,226,125,24,21,213,42,0,0        ; vbroadcastss  0x2ad5(%rip),%ymm2        # 681c <_sk_callback_avx+0x3d1>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,96,42,0,0         ; vbroadcastss  0x2a60(%rip),%ymm2        # 67b4 <_sk_callback_avx+0x3d5>
+  DB  196,226,125,24,21,204,42,0,0        ; vbroadcastss  0x2acc(%rip),%ymm2        # 6820 <_sk_callback_avx+0x3d5>
   DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,83,42,0,0           ; vbroadcastss  0x2a53(%rip),%ymm8        # 67b8 <_sk_callback_avx+0x3d9>
+  DB  196,98,125,24,5,191,42,0,0          ; vbroadcastss  0x2abf(%rip),%ymm8        # 6824 <_sk_callback_avx+0x3d9>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,73,42,0,0           ; vbroadcastss  0x2a49(%rip),%ymm8        # 67bc <_sk_callback_avx+0x3dd>
+  DB  196,98,125,24,5,181,42,0,0          ; vbroadcastss  0x2ab5(%rip),%ymm8        # 6828 <_sk_callback_avx+0x3dd>
   DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,59,42,0,0           ; vbroadcastss  0x2a3b(%rip),%ymm8        # 67c0 <_sk_callback_avx+0x3e1>
+  DB  196,98,125,24,5,167,42,0,0          ; vbroadcastss  0x2aa7(%rip),%ymm8        # 682c <_sk_callback_avx+0x3e1>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -8722,25 +8758,25 @@ _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  196,226,125,24,5,210,40,0,0         ; vbroadcastss  0x28d2(%rip),%ymm0        # 67c4 <_sk_callback_avx+0x3e5>
+  DB  196,226,125,24,5,62,41,0,0          ; vbroadcastss  0x293e(%rip),%ymm0        # 6830 <_sk_callback_avx+0x3e5>
   DB  197,228,84,192                      ; vandps        %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,197,40,0,0        ; vbroadcastss  0x28c5(%rip),%ymm1        # 67c8 <_sk_callback_avx+0x3e9>
+  DB  196,226,125,24,13,49,41,0,0         ; vbroadcastss  0x2931(%rip),%ymm1        # 6834 <_sk_callback_avx+0x3e9>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,188,40,0,0        ; vbroadcastss  0x28bc(%rip),%ymm1        # 67cc <_sk_callback_avx+0x3ed>
+  DB  196,226,125,24,13,40,41,0,0         ; vbroadcastss  0x2928(%rip),%ymm1        # 6838 <_sk_callback_avx+0x3ed>
   DB  197,228,84,201                      ; vandps        %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,175,40,0,0        ; vbroadcastss  0x28af(%rip),%ymm2        # 67d0 <_sk_callback_avx+0x3f1>
+  DB  196,226,125,24,21,27,41,0,0         ; vbroadcastss  0x291b(%rip),%ymm2        # 683c <_sk_callback_avx+0x3f1>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,166,40,0,0        ; vbroadcastss  0x28a6(%rip),%ymm2        # 67d4 <_sk_callback_avx+0x3f5>
+  DB  196,226,125,24,21,18,41,0,0         ; vbroadcastss  0x2912(%rip),%ymm2        # 6840 <_sk_callback_avx+0x3f5>
   DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,153,40,0,0          ; vbroadcastss  0x2899(%rip),%ymm8        # 67d8 <_sk_callback_avx+0x3f9>
+  DB  196,98,125,24,5,5,41,0,0            ; vbroadcastss  0x2905(%rip),%ymm8        # 6844 <_sk_callback_avx+0x3f9>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,143,40,0,0          ; vbroadcastss  0x288f(%rip),%ymm8        # 67dc <_sk_callback_avx+0x3fd>
+  DB  196,98,125,24,5,251,40,0,0          ; vbroadcastss  0x28fb(%rip),%ymm8        # 6848 <_sk_callback_avx+0x3fd>
   DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,129,40,0,0          ; vbroadcastss  0x2881(%rip),%ymm8        # 67e0 <_sk_callback_avx+0x401>
+  DB  196,98,125,24,5,237,40,0,0          ; vbroadcastss  0x28ed(%rip),%ymm8        # 684c <_sk_callback_avx+0x401>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -8754,7 +8790,7 @@ 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  196,98,125,24,5,102,40,0,0          ; vbroadcastss  0x2866(%rip),%ymm8        # 67e4 <_sk_callback_avx+0x405>
+  DB  196,98,125,24,5,210,40,0,0          ; vbroadcastss  0x28d2(%rip),%ymm8        # 6850 <_sk_callback_avx+0x405>
   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
@@ -8833,10 +8869,10 @@ _sk_load_8888_avx LABEL PROC
   DB  72,133,201                          ; test          %rcx,%rcx
   DB  15,133,135,0,0,0                    ; jne           4125 <_sk_load_8888_avx+0x95>
   DB  196,65,124,16,12,186                ; vmovups       (%r10,%rdi,4),%ymm9
-  DB  197,124,40,21,148,40,0,0            ; vmovaps       0x2894(%rip),%ymm10        # 6940 <_sk_callback_avx+0x561>
+  DB  197,124,40,21,20,41,0,0             ; vmovaps       0x2914(%rip),%ymm10        # 69c0 <_sk_callback_avx+0x575>
   DB  196,193,52,84,194                   ; vandps        %ymm10,%ymm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,42,39,0,0           ; vbroadcastss  0x272a(%rip),%ymm8        # 67e8 <_sk_callback_avx+0x409>
+  DB  196,98,125,24,5,150,39,0,0          ; vbroadcastss  0x2796(%rip),%ymm8        # 6854 <_sk_callback_avx+0x409>
   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
@@ -8949,10 +8985,10 @@ _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  197,124,40,21,190,38,0,0            ; vmovaps       0x26be(%rip),%ymm10        # 6960 <_sk_callback_avx+0x581>
+  DB  197,124,40,21,62,39,0,0             ; vmovaps       0x273e(%rip),%ymm10        # 69e0 <_sk_callback_avx+0x595>
   DB  196,193,124,84,194                  ; vandps        %ymm10,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,56,37,0,0          ; vbroadcastss  0x2538(%rip),%ymm9        # 67ec <_sk_callback_avx+0x40d>
+  DB  196,98,125,24,13,164,37,0,0         ; vbroadcastss  0x25a4(%rip),%ymm9        # 6858 <_sk_callback_avx+0x40d>
   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
@@ -8982,7 +9018,7 @@ 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  196,98,125,24,5,198,36,0,0          ; vbroadcastss  0x24c6(%rip),%ymm8        # 67f0 <_sk_callback_avx+0x411>
+  DB  196,98,125,24,5,50,37,0,0           ; vbroadcastss  0x2532(%rip),%ymm8        # 685c <_sk_callback_avx+0x411>
   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
@@ -9085,13 +9121,13 @@ _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  196,98,125,24,37,43,35,0,0          ; vbroadcastss  0x232b(%rip),%ymm12        # 67f4 <_sk_callback_avx+0x415>
+  DB  196,98,125,24,37,151,35,0,0         ; vbroadcastss  0x2397(%rip),%ymm12        # 6860 <_sk_callback_avx+0x415>
   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  196,98,121,24,29,23,35,0,0          ; vbroadcastss  0x2317(%rip),%xmm11        # 67f8 <_sk_callback_avx+0x419>
+  DB  196,98,121,24,29,131,35,0,0         ; vbroadcastss  0x2383(%rip),%xmm11        # 6864 <_sk_callback_avx+0x419>
   DB  196,193,8,87,219                    ; vxorps        %xmm11,%xmm14,%xmm3
-  DB  196,98,121,24,45,13,35,0,0          ; vbroadcastss  0x230d(%rip),%xmm13        # 67fc <_sk_callback_avx+0x41d>
+  DB  196,98,121,24,45,121,35,0,0         ; vbroadcastss  0x2379(%rip),%xmm13        # 6868 <_sk_callback_avx+0x41d>
   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
@@ -9105,7 +9141,7 @@ _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  196,226,121,24,29,195,34,0,0        ; vbroadcastss  0x22c3(%rip),%xmm3        # 6800 <_sk_callback_avx+0x421>
+  DB  196,226,121,24,29,47,35,0,0         ; vbroadcastss  0x232f(%rip),%xmm3        # 686c <_sk_callback_avx+0x421>
   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
@@ -9282,13 +9318,13 @@ _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  196,98,125,24,37,131,31,0,0         ; vbroadcastss  0x1f83(%rip),%ymm12        # 6804 <_sk_callback_avx+0x425>
+  DB  196,98,125,24,37,239,31,0,0         ; vbroadcastss  0x1fef(%rip),%ymm12        # 6870 <_sk_callback_avx+0x425>
   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  196,98,121,24,29,111,31,0,0         ; vbroadcastss  0x1f6f(%rip),%xmm11        # 6808 <_sk_callback_avx+0x429>
+  DB  196,98,121,24,29,219,31,0,0         ; vbroadcastss  0x1fdb(%rip),%xmm11        # 6874 <_sk_callback_avx+0x429>
   DB  196,193,8,87,219                    ; vxorps        %xmm11,%xmm14,%xmm3
-  DB  196,98,121,24,45,101,31,0,0         ; vbroadcastss  0x1f65(%rip),%xmm13        # 680c <_sk_callback_avx+0x42d>
+  DB  196,98,121,24,45,209,31,0,0         ; vbroadcastss  0x1fd1(%rip),%xmm13        # 6878 <_sk_callback_avx+0x42d>
   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
@@ -9302,7 +9338,7 @@ _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  196,226,121,24,29,27,31,0,0         ; vbroadcastss  0x1f1b(%rip),%xmm3        # 6810 <_sk_callback_avx+0x431>
+  DB  196,226,121,24,29,135,31,0,0        ; vbroadcastss  0x1f87(%rip),%xmm3        # 687c <_sk_callback_avx+0x431>
   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
@@ -9404,12 +9440,12 @@ _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  196,98,125,24,13,40,29,0,0          ; vbroadcastss  0x1d28(%rip),%ymm9        # 6814 <_sk_callback_avx+0x435>
+  DB  196,98,125,24,13,148,29,0,0         ; vbroadcastss  0x1d94(%rip),%ymm9        # 6880 <_sk_callback_avx+0x435>
   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  196,98,121,24,5,14,29,0,0           ; vbroadcastss  0x1d0e(%rip),%xmm8        # 6818 <_sk_callback_avx+0x439>
+  DB  196,98,121,24,5,122,29,0,0          ; vbroadcastss  0x1d7a(%rip),%xmm8        # 6884 <_sk_callback_avx+0x439>
   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
@@ -9419,7 +9455,7 @@ _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  196,98,125,24,21,213,28,0,0         ; vbroadcastss  0x1cd5(%rip),%ymm10        # 681c <_sk_callback_avx+0x43d>
+  DB  196,98,125,24,21,65,29,0,0          ; vbroadcastss  0x1d41(%rip),%ymm10        # 6888 <_sk_callback_avx+0x43d>
   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
@@ -9562,7 +9598,7 @@ _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  196,98,125,24,29,36,26,0,0          ; vbroadcastss  0x1a24(%rip),%ymm11        # 6820 <_sk_callback_avx+0x441>
+  DB  196,98,125,24,29,144,26,0,0         ; vbroadcastss  0x1a90(%rip),%ymm11        # 688c <_sk_callback_avx+0x441>
   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
@@ -9653,7 +9689,7 @@ _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  196,98,125,24,29,132,24,0,0         ; vbroadcastss  0x1884(%rip),%ymm11        # 6824 <_sk_callback_avx+0x445>
+  DB  196,98,125,24,29,240,24,0,0         ; vbroadcastss  0x18f0(%rip),%ymm11        # 6890 <_sk_callback_avx+0x445>
   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
@@ -9674,7 +9710,7 @@ _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  196,226,125,24,29,33,24,0,0         ; vbroadcastss  0x1821(%rip),%ymm3        # 6828 <_sk_callback_avx+0x449>
+  DB  196,226,125,24,29,141,24,0,0        ; vbroadcastss  0x188d(%rip),%ymm3        # 6894 <_sk_callback_avx+0x449>
   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
@@ -9715,7 +9751,7 @@ _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  196,98,125,24,5,94,23,0,0           ; vbroadcastss  0x175e(%rip),%ymm8        # 682c <_sk_callback_avx+0x44d>
+  DB  196,98,125,24,5,202,23,0,0          ; vbroadcastss  0x17ca(%rip),%ymm8        # 6898 <_sk_callback_avx+0x44d>
   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
@@ -9961,14 +9997,46 @@ _sk_mirror_y_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_clamp_x_1_avx
+_sk_clamp_x_1_avx LABEL PROC
+  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
+  DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
+  DB  196,98,125,24,5,230,19,0,0          ; vbroadcastss  0x13e6(%rip),%ymm8        # 689c <_sk_callback_avx+0x451>
+  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_repeat_x_1_avx
+_sk_repeat_x_1_avx LABEL PROC
+  DB  196,99,125,8,192,1                  ; vroundps      $0x1,%ymm0,%ymm8
+  DB  196,193,124,92,192                  ; vsubps        %ymm8,%ymm0,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_mirror_x_1_avx
+_sk_mirror_x_1_avx LABEL PROC
+  DB  196,98,125,24,5,201,19,0,0          ; vbroadcastss  0x13c9(%rip),%ymm8        # 68a0 <_sk_callback_avx+0x455>
+  DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
+  DB  196,98,125,24,13,191,19,0,0         ; vbroadcastss  0x13bf(%rip),%ymm9        # 68a4 <_sk_callback_avx+0x459>
+  DB  196,65,124,89,201                   ; vmulps        %ymm9,%ymm0,%ymm9
+  DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
+  DB  196,65,52,88,201                    ; vaddps        %ymm9,%ymm9,%ymm9
+  DB  196,193,124,92,193                  ; vsubps        %ymm9,%ymm0,%ymm0
+  DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
+  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
+  DB  197,60,92,192                       ; vsubps        %ymm0,%ymm8,%ymm8
+  DB  197,188,84,192                      ; vandps        %ymm0,%ymm8,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_luminance_to_alpha_avx
 _sk_luminance_to_alpha_avx LABEL PROC
-  DB  196,226,125,24,29,131,19,0,0        ; vbroadcastss  0x1383(%rip),%ymm3        # 6830 <_sk_callback_avx+0x451>
+  DB  196,226,125,24,29,143,19,0,0        ; vbroadcastss  0x138f(%rip),%ymm3        # 68a8 <_sk_callback_avx+0x45d>
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,125,24,29,122,19,0,0        ; vbroadcastss  0x137a(%rip),%ymm3        # 6834 <_sk_callback_avx+0x455>
+  DB  196,226,125,24,29,134,19,0,0        ; vbroadcastss  0x1386(%rip),%ymm3        # 68ac <_sk_callback_avx+0x461>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,109,19,0,0        ; vbroadcastss  0x136d(%rip),%ymm1        # 6838 <_sk_callback_avx+0x459>
+  DB  196,226,125,24,13,121,19,0,0        ; vbroadcastss  0x1379(%rip),%ymm1        # 68b0 <_sk_callback_avx+0x465>
   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
@@ -10175,9 +10243,9 @@ _sk_evenly_spaced_gradient_avx LABEL PROC
   DB  72,139,24                           ; mov           (%rax),%rbx
   DB  72,139,104,8                        ; mov           0x8(%rax),%rbp
   DB  72,255,203                          ; dec           %rbx
-  DB  120,7                               ; js            5821 <_sk_evenly_spaced_gradient_avx+0x1f>
+  DB  120,7                               ; js            588d <_sk_evenly_spaced_gradient_avx+0x1f>
   DB  196,225,242,42,203                  ; vcvtsi2ss     %rbx,%xmm1,%xmm1
-  DB  235,21                              ; jmp           5836 <_sk_evenly_spaced_gradient_avx+0x34>
+  DB  235,21                              ; jmp           58a2 <_sk_evenly_spaced_gradient_avx+0x34>
   DB  73,137,216                          ; mov           %rbx,%r8
   DB  73,209,232                          ; shr           %r8
   DB  131,227,1                           ; and           $0x1,%ebx
@@ -10342,12 +10410,12 @@ _sk_gradient_avx LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  73,131,248,2                        ; cmp           $0x2,%r8
-  DB  114,80                              ; jb            5bc4 <_sk_gradient_avx+0x69>
+  DB  114,80                              ; jb            5c30 <_sk_gradient_avx+0x69>
   DB  72,139,88,72                        ; mov           0x48(%rax),%rbx
   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  196,98,125,24,21,175,12,0,0         ; vbroadcastss  0xcaf(%rip),%ymm10        # 683c <_sk_callback_avx+0x45d>
+  DB  196,98,125,24,21,187,12,0,0         ; vbroadcastss  0xcbb(%rip),%ymm10        # 68b4 <_sk_callback_avx+0x469>
   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
@@ -10359,7 +10427,7 @@ _sk_gradient_avx LABEL PROC
   DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
   DB  72,131,195,4                        ; add           $0x4,%rbx
   DB  73,255,200                          ; dec           %r8
-  DB  117,205                             ; jne           5b91 <_sk_gradient_avx+0x36>
+  DB  117,205                             ; jne           5bfd <_sk_gradient_avx+0x36>
   DB  196,195,249,22,200,1                ; vpextrq       $0x1,%xmm1,%r8
   DB  69,137,193                          ; mov           %r8d,%r9d
   DB  73,193,232,32                       ; shr           $0x20,%r8
@@ -10537,27 +10605,27 @@ _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  196,98,125,24,45,211,8,0,0          ; vbroadcastss  0x8d3(%rip),%ymm13        # 6840 <_sk_callback_avx+0x461>
+  DB  196,98,125,24,45,223,8,0,0          ; vbroadcastss  0x8df(%rip),%ymm13        # 68b8 <_sk_callback_avx+0x46d>
   DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
-  DB  196,98,125,24,53,201,8,0,0          ; vbroadcastss  0x8c9(%rip),%ymm14        # 6844 <_sk_callback_avx+0x465>
+  DB  196,98,125,24,53,213,8,0,0          ; vbroadcastss  0x8d5(%rip),%ymm14        # 68bc <_sk_callback_avx+0x471>
   DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
   DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
-  DB  196,98,125,24,53,186,8,0,0          ; vbroadcastss  0x8ba(%rip),%ymm14        # 6848 <_sk_callback_avx+0x469>
+  DB  196,98,125,24,53,198,8,0,0          ; vbroadcastss  0x8c6(%rip),%ymm14        # 68c0 <_sk_callback_avx+0x475>
   DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
   DB  196,65,28,89,229                    ; vmulps        %ymm13,%ymm12,%ymm12
-  DB  196,98,125,24,45,171,8,0,0          ; vbroadcastss  0x8ab(%rip),%ymm13        # 684c <_sk_callback_avx+0x46d>
+  DB  196,98,125,24,45,183,8,0,0          ; vbroadcastss  0x8b7(%rip),%ymm13        # 68c4 <_sk_callback_avx+0x479>
   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  196,98,125,24,21,150,8,0,0          ; vbroadcastss  0x896(%rip),%ymm10        # 6850 <_sk_callback_avx+0x471>
+  DB  196,98,125,24,21,162,8,0,0          ; vbroadcastss  0x8a2(%rip),%ymm10        # 68c8 <_sk_callback_avx+0x47d>
   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  196,98,125,24,21,128,8,0,0          ; vbroadcastss  0x880(%rip),%ymm10        # 6854 <_sk_callback_avx+0x475>
+  DB  196,98,125,24,21,140,8,0,0          ; vbroadcastss  0x88c(%rip),%ymm10        # 68cc <_sk_callback_avx+0x481>
   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  196,98,125,24,21,106,8,0,0          ; vbroadcastss  0x86a(%rip),%ymm10        # 6858 <_sk_callback_avx+0x479>
+  DB  196,98,125,24,21,118,8,0,0          ; vbroadcastss  0x876(%rip),%ymm10        # 68d0 <_sk_callback_avx+0x485>
   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
@@ -10577,7 +10645,7 @@ _sk_xy_to_radius_avx LABEL PROC
 PUBLIC _sk_save_xy_avx
 _sk_save_xy_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,52,8,0,0            ; vbroadcastss  0x834(%rip),%ymm8        # 685c <_sk_callback_avx+0x47d>
+  DB  196,98,125,24,5,64,8,0,0            ; vbroadcastss  0x840(%rip),%ymm8        # 68d4 <_sk_callback_avx+0x489>
   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
@@ -10610,9 +10678,9 @@ _sk_accumulate_avx LABEL PROC
 PUBLIC _sk_bilinear_nx_avx
 _sk_bilinear_nx_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,192,7,0,0          ; vbroadcastss  0x7c0(%rip),%ymm0        # 6860 <_sk_callback_avx+0x481>
+  DB  196,226,125,24,5,204,7,0,0          ; vbroadcastss  0x7cc(%rip),%ymm0        # 68d8 <_sk_callback_avx+0x48d>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,183,7,0,0           ; vbroadcastss  0x7b7(%rip),%ymm8        # 6864 <_sk_callback_avx+0x485>
+  DB  196,98,125,24,5,195,7,0,0           ; vbroadcastss  0x7c3(%rip),%ymm8        # 68dc <_sk_callback_avx+0x491>
   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
@@ -10621,7 +10689,7 @@ _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  196,226,125,24,5,159,7,0,0          ; vbroadcastss  0x79f(%rip),%ymm0        # 6868 <_sk_callback_avx+0x489>
+  DB  196,226,125,24,5,171,7,0,0          ; vbroadcastss  0x7ab(%rip),%ymm0        # 68e0 <_sk_callback_avx+0x495>
   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)
@@ -10631,9 +10699,9 @@ _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  196,226,125,24,13,131,7,0,0         ; vbroadcastss  0x783(%rip),%ymm1        # 686c <_sk_callback_avx+0x48d>
+  DB  196,226,125,24,13,143,7,0,0         ; vbroadcastss  0x78f(%rip),%ymm1        # 68e4 <_sk_callback_avx+0x499>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,121,7,0,0           ; vbroadcastss  0x779(%rip),%ymm8        # 6870 <_sk_callback_avx+0x491>
+  DB  196,98,125,24,5,133,7,0,0           ; vbroadcastss  0x785(%rip),%ymm8        # 68e8 <_sk_callback_avx+0x49d>
   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
@@ -10642,7 +10710,7 @@ _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  196,226,125,24,13,97,7,0,0          ; vbroadcastss  0x761(%rip),%ymm1        # 6874 <_sk_callback_avx+0x495>
+  DB  196,226,125,24,13,109,7,0,0         ; vbroadcastss  0x76d(%rip),%ymm1        # 68ec <_sk_callback_avx+0x4a1>
   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)
@@ -10652,14 +10720,14 @@ _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  196,226,125,24,5,68,7,0,0           ; vbroadcastss  0x744(%rip),%ymm0        # 6878 <_sk_callback_avx+0x499>
+  DB  196,226,125,24,5,80,7,0,0           ; vbroadcastss  0x750(%rip),%ymm0        # 68f0 <_sk_callback_avx+0x4a5>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,59,7,0,0            ; vbroadcastss  0x73b(%rip),%ymm8        # 687c <_sk_callback_avx+0x49d>
+  DB  196,98,125,24,5,71,7,0,0            ; vbroadcastss  0x747(%rip),%ymm8        # 68f4 <_sk_callback_avx+0x4a9>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,44,7,0,0           ; vbroadcastss  0x72c(%rip),%ymm10        # 6880 <_sk_callback_avx+0x4a1>
+  DB  196,98,125,24,21,56,7,0,0           ; vbroadcastss  0x738(%rip),%ymm10        # 68f8 <_sk_callback_avx+0x4ad>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,34,7,0,0           ; vbroadcastss  0x722(%rip),%ymm10        # 6884 <_sk_callback_avx+0x4a5>
+  DB  196,98,125,24,21,46,7,0,0           ; vbroadcastss  0x72e(%rip),%ymm10        # 68fc <_sk_callback_avx+0x4b1>
   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)
@@ -10669,19 +10737,19 @@ _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  196,226,125,24,5,5,7,0,0            ; vbroadcastss  0x705(%rip),%ymm0        # 6888 <_sk_callback_avx+0x4a9>
+  DB  196,226,125,24,5,17,7,0,0           ; vbroadcastss  0x711(%rip),%ymm0        # 6900 <_sk_callback_avx+0x4b5>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,252,6,0,0           ; vbroadcastss  0x6fc(%rip),%ymm8        # 688c <_sk_callback_avx+0x4ad>
+  DB  196,98,125,24,5,8,7,0,0             ; vbroadcastss  0x708(%rip),%ymm8        # 6904 <_sk_callback_avx+0x4b9>
   DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,242,6,0,0          ; vbroadcastss  0x6f2(%rip),%ymm9        # 6890 <_sk_callback_avx+0x4b1>
+  DB  196,98,125,24,13,254,6,0,0          ; vbroadcastss  0x6fe(%rip),%ymm9        # 6908 <_sk_callback_avx+0x4bd>
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,232,6,0,0          ; vbroadcastss  0x6e8(%rip),%ymm10        # 6894 <_sk_callback_avx+0x4b5>
+  DB  196,98,125,24,21,244,6,0,0          ; vbroadcastss  0x6f4(%rip),%ymm10        # 690c <_sk_callback_avx+0x4c1>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,217,6,0,0          ; vbroadcastss  0x6d9(%rip),%ymm10        # 6898 <_sk_callback_avx+0x4b9>
+  DB  196,98,125,24,21,229,6,0,0          ; vbroadcastss  0x6e5(%rip),%ymm10        # 6910 <_sk_callback_avx+0x4c5>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,202,6,0,0          ; vbroadcastss  0x6ca(%rip),%ymm9        # 689c <_sk_callback_avx+0x4bd>
+  DB  196,98,125,24,13,214,6,0,0          ; vbroadcastss  0x6d6(%rip),%ymm9        # 6914 <_sk_callback_avx+0x4c9>
   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
@@ -10690,17 +10758,17 @@ _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  196,98,125,24,5,178,6,0,0           ; vbroadcastss  0x6b2(%rip),%ymm8        # 68a0 <_sk_callback_avx+0x4c1>
+  DB  196,98,125,24,5,190,6,0,0           ; vbroadcastss  0x6be(%rip),%ymm8        # 6918 <_sk_callback_avx+0x4cd>
   DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
   DB  197,124,16,72,64                    ; vmovups       0x40(%rax),%ymm9
-  DB  196,98,125,24,21,164,6,0,0          ; vbroadcastss  0x6a4(%rip),%ymm10        # 68a4 <_sk_callback_avx+0x4c5>
+  DB  196,98,125,24,21,176,6,0,0          ; vbroadcastss  0x6b0(%rip),%ymm10        # 691c <_sk_callback_avx+0x4d1>
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
-  DB  196,98,125,24,29,154,6,0,0          ; vbroadcastss  0x69a(%rip),%ymm11        # 68a8 <_sk_callback_avx+0x4c9>
+  DB  196,98,125,24,29,166,6,0,0          ; vbroadcastss  0x6a6(%rip),%ymm11        # 6920 <_sk_callback_avx+0x4d5>
   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  196,98,125,24,13,129,6,0,0          ; vbroadcastss  0x681(%rip),%ymm9        # 68ac <_sk_callback_avx+0x4cd>
+  DB  196,98,125,24,13,141,6,0,0          ; vbroadcastss  0x68d(%rip),%ymm9        # 6924 <_sk_callback_avx+0x4d9>
   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
@@ -10709,13 +10777,13 @@ _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  196,226,125,24,5,105,6,0,0          ; vbroadcastss  0x669(%rip),%ymm0        # 68b0 <_sk_callback_avx+0x4d1>
+  DB  196,226,125,24,5,117,6,0,0          ; vbroadcastss  0x675(%rip),%ymm0        # 6928 <_sk_callback_avx+0x4dd>
   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  196,98,125,24,21,86,6,0,0           ; vbroadcastss  0x656(%rip),%ymm10        # 68b4 <_sk_callback_avx+0x4d5>
+  DB  196,98,125,24,21,98,6,0,0           ; vbroadcastss  0x662(%rip),%ymm10        # 692c <_sk_callback_avx+0x4e1>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,76,6,0,0           ; vbroadcastss  0x64c(%rip),%ymm10        # 68b8 <_sk_callback_avx+0x4d9>
+  DB  196,98,125,24,21,88,6,0,0           ; vbroadcastss  0x658(%rip),%ymm10        # 6930 <_sk_callback_avx+0x4e5>
   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)
@@ -10725,14 +10793,14 @@ _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  196,226,125,24,13,47,6,0,0          ; vbroadcastss  0x62f(%rip),%ymm1        # 68bc <_sk_callback_avx+0x4dd>
+  DB  196,226,125,24,13,59,6,0,0          ; vbroadcastss  0x63b(%rip),%ymm1        # 6934 <_sk_callback_avx+0x4e9>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,37,6,0,0            ; vbroadcastss  0x625(%rip),%ymm8        # 68c0 <_sk_callback_avx+0x4e1>
+  DB  196,98,125,24,5,49,6,0,0            ; vbroadcastss  0x631(%rip),%ymm8        # 6938 <_sk_callback_avx+0x4ed>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,22,6,0,0           ; vbroadcastss  0x616(%rip),%ymm10        # 68c4 <_sk_callback_avx+0x4e5>
+  DB  196,98,125,24,21,34,6,0,0           ; vbroadcastss  0x622(%rip),%ymm10        # 693c <_sk_callback_avx+0x4f1>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,12,6,0,0           ; vbroadcastss  0x60c(%rip),%ymm10        # 68c8 <_sk_callback_avx+0x4e9>
+  DB  196,98,125,24,21,24,6,0,0           ; vbroadcastss  0x618(%rip),%ymm10        # 6940 <_sk_callback_avx+0x4f5>
   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)
@@ -10742,19 +10810,19 @@ _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  196,226,125,24,13,239,5,0,0         ; vbroadcastss  0x5ef(%rip),%ymm1        # 68cc <_sk_callback_avx+0x4ed>
+  DB  196,226,125,24,13,251,5,0,0         ; vbroadcastss  0x5fb(%rip),%ymm1        # 6944 <_sk_callback_avx+0x4f9>
   DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,229,5,0,0           ; vbroadcastss  0x5e5(%rip),%ymm8        # 68d0 <_sk_callback_avx+0x4f1>
+  DB  196,98,125,24,5,241,5,0,0           ; vbroadcastss  0x5f1(%rip),%ymm8        # 6948 <_sk_callback_avx+0x4fd>
   DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,219,5,0,0          ; vbroadcastss  0x5db(%rip),%ymm9        # 68d4 <_sk_callback_avx+0x4f5>
+  DB  196,98,125,24,13,231,5,0,0          ; vbroadcastss  0x5e7(%rip),%ymm9        # 694c <_sk_callback_avx+0x501>
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,209,5,0,0          ; vbroadcastss  0x5d1(%rip),%ymm10        # 68d8 <_sk_callback_avx+0x4f9>
+  DB  196,98,125,24,21,221,5,0,0          ; vbroadcastss  0x5dd(%rip),%ymm10        # 6950 <_sk_callback_avx+0x505>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,194,5,0,0          ; vbroadcastss  0x5c2(%rip),%ymm10        # 68dc <_sk_callback_avx+0x4fd>
+  DB  196,98,125,24,21,206,5,0,0          ; vbroadcastss  0x5ce(%rip),%ymm10        # 6954 <_sk_callback_avx+0x509>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,179,5,0,0          ; vbroadcastss  0x5b3(%rip),%ymm9        # 68e0 <_sk_callback_avx+0x501>
+  DB  196,98,125,24,13,191,5,0,0          ; vbroadcastss  0x5bf(%rip),%ymm9        # 6958 <_sk_callback_avx+0x50d>
   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
@@ -10763,17 +10831,17 @@ _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  196,98,125,24,5,155,5,0,0           ; vbroadcastss  0x59b(%rip),%ymm8        # 68e4 <_sk_callback_avx+0x505>
+  DB  196,98,125,24,5,167,5,0,0           ; vbroadcastss  0x5a7(%rip),%ymm8        # 695c <_sk_callback_avx+0x511>
   DB  197,188,88,72,32                    ; vaddps        0x20(%rax),%ymm8,%ymm1
   DB  197,124,16,72,96                    ; vmovups       0x60(%rax),%ymm9
-  DB  196,98,125,24,21,140,5,0,0          ; vbroadcastss  0x58c(%rip),%ymm10        # 68e8 <_sk_callback_avx+0x509>
+  DB  196,98,125,24,21,152,5,0,0          ; vbroadcastss  0x598(%rip),%ymm10        # 6960 <_sk_callback_avx+0x515>
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
-  DB  196,98,125,24,29,130,5,0,0          ; vbroadcastss  0x582(%rip),%ymm11        # 68ec <_sk_callback_avx+0x50d>
+  DB  196,98,125,24,29,142,5,0,0          ; vbroadcastss  0x58e(%rip),%ymm11        # 6964 <_sk_callback_avx+0x519>
   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  196,98,125,24,13,105,5,0,0          ; vbroadcastss  0x569(%rip),%ymm9        # 68f0 <_sk_callback_avx+0x511>
+  DB  196,98,125,24,13,117,5,0,0          ; vbroadcastss  0x575(%rip),%ymm9        # 6968 <_sk_callback_avx+0x51d>
   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
@@ -10782,13 +10850,13 @@ _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  196,226,125,24,13,81,5,0,0          ; vbroadcastss  0x551(%rip),%ymm1        # 68f4 <_sk_callback_avx+0x515>
+  DB  196,226,125,24,13,93,5,0,0          ; vbroadcastss  0x55d(%rip),%ymm1        # 696c <_sk_callback_avx+0x521>
   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  196,98,125,24,21,61,5,0,0           ; vbroadcastss  0x53d(%rip),%ymm10        # 68f8 <_sk_callback_avx+0x519>
+  DB  196,98,125,24,21,73,5,0,0           ; vbroadcastss  0x549(%rip),%ymm10        # 6970 <_sk_callback_avx+0x525>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,51,5,0,0           ; vbroadcastss  0x533(%rip),%ymm10        # 68fc <_sk_callback_avx+0x51d>
+  DB  196,98,125,24,21,63,5,0,0           ; vbroadcastss  0x53f(%rip),%ymm10        # 6974 <_sk_callback_avx+0x529>
   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)
@@ -10902,25 +10970,25 @@ 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 65a9 <.literal4+0xb1>
+  DB  71,225,61                           ; rex.RXB       loope 6615 <.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 65b9 <.literal4+0xc1>
+  DB  71,225,61                           ; rex.RXB       loope 6625 <.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 65c9 <.literal4+0xd1>
+  DB  71,225,61                           ; rex.RXB       loope 6635 <.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 65d9 <.literal4+0xe1>
+  DB  71,225,61                           ; rex.RXB       loope 6645 <.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)
@@ -10968,7 +11036,7 @@ 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        6625 <.literal4+0x12d>
+  DB  224,7                               ; loopne        6691 <.literal4+0x12d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -10984,10 +11052,10 @@ 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            664c <.literal4+0x154>
+  DB  127,0                               ; jg            66b8 <.literal4+0x154>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            66c5 <.literal4+0x1cd>
+  DB  119,115                             ; ja            6731 <.literal4+0x1cd>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -11001,10 +11069,10 @@ 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            6680 <.literal4+0x188>
+  DB  127,0                               ; jg            66ec <.literal4+0x188>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            66f9 <.literal4+0x201>
+  DB  119,115                             ; ja            6765 <.literal4+0x201>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -11018,10 +11086,10 @@ 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            66b4 <.literal4+0x1bc>
+  DB  127,0                               ; jg            6720 <.literal4+0x1bc>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            672d <.literal4+0x235>
+  DB  119,115                             ; ja            6799 <.literal4+0x235>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -11035,10 +11103,10 @@ 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            66e8 <.literal4+0x1f0>
+  DB  127,0                               ; jg            6754 <.literal4+0x1f0>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            6761 <.literal4+0x269>
+  DB  119,115                             ; ja            67cd <.literal4+0x269>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -11051,7 +11119,7 @@ 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            675f <.literal4+0x267>
+  DB  127,67                              ; jg            67cb <.literal4+0x267>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,195                               ; add           %al,%bl
   DB  0,0                                 ; add           %al,(%rax)
@@ -11063,10 +11131,10 @@ ALIGN 4
   DB  190,80,128,3,62                     ; mov           $0x3e038050,%esi
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           677f <.literal4+0x287>
+  DB  118,63                              ; jbe           67eb <.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            6793 <.literal4+0x29b>
+  DB  127,67                              ; jg            67ff <.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)
@@ -11075,7 +11143,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        6775 <.literal4+0x27d>
+  DB  224,7                               ; loopne        67e1 <.literal4+0x27d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -11087,7 +11155,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        6791 <.literal4+0x299>
+  DB  224,7                               ; loopne        67fd <.literal4+0x299>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -11098,7 +11166,7 @@ ALIGN 4
   DB  0,0                                 ; add           %al,(%rax)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            67e6 <.literal4+0x2ee>
+  DB  124,66                              ; jl            6852 <.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)
@@ -11116,9 +11184,9 @@ 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            6829 <.literal4+0x331>
+  DB  112,65                              ; jo            6895 <.literal4+0x331>
   DB  129,128,128,59,129,128,128,59,0,0   ; addl          $0x3b80,-0x7f7ec480(%rax)
-  DB  127,67                              ; jg            6837 <.literal4+0x33f>
+  DB  127,67                              ; jg            68a3 <.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)
@@ -11134,7 +11202,11 @@ 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            6877 <.literal4+0x37f>
+  DB  127,71                              ; jg            68e3 <.literal4+0x37f>
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,191,0,0,0                     ; add           %al,0xbf(%rax)
+  DB  63                                  ; (bad)
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,89                               ; ds            pop %rcx
@@ -11385,7 +11457,7 @@ _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  15,40,21,81,71,0,0                  ; movaps        0x4751(%rip),%xmm2        # 4860 <_sk_callback_sse41+0xad>
+  DB  15,40,21,177,71,0,0                 ; movaps        0x47b1(%rip),%xmm2        # 48c0 <_sk_callback_sse41+0xab>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  15,16,2                             ; movups        (%rdx),%xmm0
   DB  15,88,193                           ; addps         %xmm1,%xmm0
@@ -11394,7 +11466,7 @@ _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  15,40,21,64,71,0,0                  ; movaps        0x4740(%rip),%xmm2        # 4870 <_sk_callback_sse41+0xbd>
+  DB  15,40,21,160,71,0,0                 ; movaps        0x47a0(%rip),%xmm2        # 48d0 <_sk_callback_sse41+0xbb>
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
@@ -11413,14 +11485,14 @@ _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  102,68,15,111,21,11,71,0,0          ; movdqa        0x470b(%rip),%xmm10        # 4880 <_sk_callback_sse41+0xcd>
+  DB  102,68,15,111,21,107,71,0,0         ; movdqa        0x476b(%rip),%xmm10        # 48e0 <_sk_callback_sse41+0xcb>
   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  102,68,15,111,37,247,70,0,0         ; movdqa        0x46f7(%rip),%xmm12        # 4890 <_sk_callback_sse41+0xdd>
-  DB  102,68,15,111,45,254,70,0,0         ; movdqa        0x46fe(%rip),%xmm13        # 48a0 <_sk_callback_sse41+0xed>
+  DB  102,68,15,111,37,87,71,0,0          ; movdqa        0x4757(%rip),%xmm12        # 48f0 <_sk_callback_sse41+0xdb>
+  DB  102,68,15,111,45,94,71,0,0          ; movdqa        0x475e(%rip),%xmm13        # 4900 <_sk_callback_sse41+0xeb>
   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
@@ -11436,8 +11508,8 @@ _sk_dither_sse41 LABEL PROC
   DB  102,69,15,235,198                   ; por           %xmm14,%xmm8
   DB  102,69,15,235,197                   ; por           %xmm13,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,185,70,0,0               ; mulps         0x46b9(%rip),%xmm8        # 48b0 <_sk_callback_sse41+0xfd>
-  DB  68,15,88,5,193,70,0,0               ; addps         0x46c1(%rip),%xmm8        # 48c0 <_sk_callback_sse41+0x10d>
+  DB  68,15,89,5,25,71,0,0                ; mulps         0x4719(%rip),%xmm8        # 4910 <_sk_callback_sse41+0xfb>
+  DB  68,15,88,5,33,71,0,0                ; addps         0x4721(%rip),%xmm8        # 4920 <_sk_callback_sse41+0x10b>
   DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
@@ -11504,7 +11576,7 @@ _sk_clear_sse41 LABEL PROC
 PUBLIC _sk_srcatop_sse41
 _sk_srcatop_sse41 LABEL PROC
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  68,15,40,5,26,70,0,0                ; movaps        0x461a(%rip),%xmm8        # 48d0 <_sk_callback_sse41+0x11d>
+  DB  68,15,40,5,122,70,0,0               ; movaps        0x467a(%rip),%xmm8        # 4930 <_sk_callback_sse41+0x11b>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -11527,7 +11599,7 @@ 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  68,15,40,13,221,69,0,0              ; movaps        0x45dd(%rip),%xmm9        # 48e0 <_sk_callback_sse41+0x12d>
+  DB  68,15,40,13,61,70,0,0               ; movaps        0x463d(%rip),%xmm9        # 4940 <_sk_callback_sse41+0x12b>
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
@@ -11568,7 +11640,7 @@ _sk_dstin_sse41 LABEL PROC
 
 PUBLIC _sk_srcout_sse41
 _sk_srcout_sse41 LABEL PROC
-  DB  68,15,40,5,129,69,0,0               ; movaps        0x4581(%rip),%xmm8        # 48f0 <_sk_callback_sse41+0x13d>
+  DB  68,15,40,5,225,69,0,0               ; movaps        0x45e1(%rip),%xmm8        # 4950 <_sk_callback_sse41+0x13b>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
@@ -11579,7 +11651,7 @@ _sk_srcout_sse41 LABEL PROC
 
 PUBLIC _sk_dstout_sse41
 _sk_dstout_sse41 LABEL PROC
-  DB  68,15,40,5,113,69,0,0               ; movaps        0x4571(%rip),%xmm8        # 4900 <_sk_callback_sse41+0x14d>
+  DB  68,15,40,5,209,69,0,0               ; movaps        0x45d1(%rip),%xmm8        # 4960 <_sk_callback_sse41+0x14b>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -11594,7 +11666,7 @@ _sk_dstout_sse41 LABEL PROC
 
 PUBLIC _sk_srcover_sse41
 _sk_srcover_sse41 LABEL PROC
-  DB  68,15,40,5,84,69,0,0                ; movaps        0x4554(%rip),%xmm8        # 4910 <_sk_callback_sse41+0x15d>
+  DB  68,15,40,5,180,69,0,0               ; movaps        0x45b4(%rip),%xmm8        # 4970 <_sk_callback_sse41+0x15b>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -11612,7 +11684,7 @@ _sk_srcover_sse41 LABEL PROC
 
 PUBLIC _sk_dstover_sse41
 _sk_dstover_sse41 LABEL PROC
-  DB  68,15,40,5,40,69,0,0                ; movaps        0x4528(%rip),%xmm8        # 4920 <_sk_callback_sse41+0x16d>
+  DB  68,15,40,5,136,69,0,0               ; movaps        0x4588(%rip),%xmm8        # 4980 <_sk_callback_sse41+0x16b>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -11636,7 +11708,7 @@ _sk_modulate_sse41 LABEL PROC
 
 PUBLIC _sk_multiply_sse41
 _sk_multiply_sse41 LABEL PROC
-  DB  68,15,40,5,252,68,0,0               ; movaps        0x44fc(%rip),%xmm8        # 4930 <_sk_callback_sse41+0x17d>
+  DB  68,15,40,5,92,69,0,0                ; movaps        0x455c(%rip),%xmm8        # 4990 <_sk_callback_sse41+0x17b>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
@@ -11706,7 +11778,7 @@ _sk_screen_sse41 LABEL PROC
 PUBLIC _sk_xor__sse41
 _sk_xor__sse41 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
-  DB  15,40,29,45,68,0,0                  ; movaps        0x442d(%rip),%xmm3        # 4940 <_sk_callback_sse41+0x18d>
+  DB  15,40,29,141,68,0,0                 ; movaps        0x448d(%rip),%xmm3        # 49a0 <_sk_callback_sse41+0x18b>
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
@@ -11752,7 +11824,7 @@ _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  15,40,21,152,67,0,0                 ; movaps        0x4398(%rip),%xmm2        # 4950 <_sk_callback_sse41+0x19d>
+  DB  15,40,21,248,67,0,0                 ; movaps        0x43f8(%rip),%xmm2        # 49b0 <_sk_callback_sse41+0x19b>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -11784,7 +11856,7 @@ _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  15,40,21,61,67,0,0                  ; movaps        0x433d(%rip),%xmm2        # 4960 <_sk_callback_sse41+0x1ad>
+  DB  15,40,21,157,67,0,0                 ; movaps        0x439d(%rip),%xmm2        # 49c0 <_sk_callback_sse41+0x1ab>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -11819,7 +11891,7 @@ _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  15,40,21,215,66,0,0                 ; movaps        0x42d7(%rip),%xmm2        # 4970 <_sk_callback_sse41+0x1bd>
+  DB  15,40,21,55,67,0,0                  ; movaps        0x4337(%rip),%xmm2        # 49d0 <_sk_callback_sse41+0x1bb>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -11844,7 +11916,7 @@ _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  15,40,13,152,66,0,0                 ; movaps        0x4298(%rip),%xmm1        # 4980 <_sk_callback_sse41+0x1cd>
+  DB  15,40,13,248,66,0,0                 ; movaps        0x42f8(%rip),%xmm1        # 49e0 <_sk_callback_sse41+0x1cb>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,207                           ; mulps         %xmm7,%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -11856,7 +11928,7 @@ _sk_exclusion_sse41 LABEL PROC
 PUBLIC _sk_colorburn_sse41
 _sk_colorburn_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,135,66,0,0              ; movaps        0x4287(%rip),%xmm10        # 4990 <_sk_callback_sse41+0x1dd>
+  DB  68,15,40,21,231,66,0,0              ; movaps        0x42e7(%rip),%xmm10        # 49f0 <_sk_callback_sse41+0x1db>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
@@ -11936,7 +12008,7 @@ _sk_colorburn_sse41 LABEL PROC
 PUBLIC _sk_colordodge_sse41
 _sk_colordodge_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,101,65,0,0              ; movaps        0x4165(%rip),%xmm10        # 49a0 <_sk_callback_sse41+0x1ed>
+  DB  68,15,40,21,197,65,0,0              ; movaps        0x41c5(%rip),%xmm10        # 4a00 <_sk_callback_sse41+0x1eb>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -12017,7 +12089,7 @@ _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  68,15,40,21,59,64,0,0               ; movaps        0x403b(%rip),%xmm10        # 49b0 <_sk_callback_sse41+0x1fd>
+  DB  68,15,40,21,155,64,0,0              ; movaps        0x409b(%rip),%xmm10        # 4a10 <_sk_callback_sse41+0x1fb>
   DB  65,15,40,234                        ; movaps        %xmm10,%xmm5
   DB  15,92,239                           ; subps         %xmm7,%xmm5
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
@@ -12099,7 +12171,7 @@ 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  68,15,40,21,29,63,0,0               ; movaps        0x3f1d(%rip),%xmm10        # 49c0 <_sk_callback_sse41+0x20d>
+  DB  68,15,40,21,125,63,0,0              ; movaps        0x3f7d(%rip),%xmm10        # 4a20 <_sk_callback_sse41+0x20b>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
@@ -12183,7 +12255,7 @@ _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  15,40,13,240,61,0,0                 ; movaps        0x3df0(%rip),%xmm1        # 49d0 <_sk_callback_sse41+0x21d>
+  DB  15,40,13,80,62,0,0                  ; movaps        0x3e50(%rip),%xmm1        # 4a30 <_sk_callback_sse41+0x21b>
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  68,15,92,208                        ; subps         %xmm0,%xmm10
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
@@ -12196,10 +12268,10 @@ _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  68,15,40,45,206,61,0,0              ; movaps        0x3dce(%rip),%xmm13        # 49e0 <_sk_callback_sse41+0x22d>
+  DB  68,15,40,45,46,62,0,0               ; movaps        0x3e2e(%rip),%xmm13        # 4a40 <_sk_callback_sse41+0x22b>
   DB  69,15,88,245                        ; addps         %xmm13,%xmm14
   DB  68,15,89,242                        ; mulps         %xmm2,%xmm14
-  DB  68,15,40,37,206,61,0,0              ; movaps        0x3dce(%rip),%xmm12        # 49f0 <_sk_callback_sse41+0x23d>
+  DB  68,15,40,37,46,62,0,0               ; movaps        0x3e2e(%rip),%xmm12        # 4a50 <_sk_callback_sse41+0x23b>
   DB  69,15,89,252                        ; mulps         %xmm12,%xmm15
   DB  69,15,88,254                        ; addps         %xmm14,%xmm15
   DB  15,40,198                           ; movaps        %xmm6,%xmm0
@@ -12385,12 +12457,12 @@ _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  15,40,5,52,59,0,0                   ; movaps        0x3b34(%rip),%xmm0        # 4a00 <_sk_callback_sse41+0x24d>
+  DB  15,40,5,148,59,0,0                  ; movaps        0x3b94(%rip),%xmm0        # 4a60 <_sk_callback_sse41+0x24b>
   DB  68,15,89,224                        ; mulps         %xmm0,%xmm12
-  DB  15,40,21,57,59,0,0                  ; movaps        0x3b39(%rip),%xmm2        # 4a10 <_sk_callback_sse41+0x25d>
+  DB  15,40,21,153,59,0,0                 ; movaps        0x3b99(%rip),%xmm2        # 4a70 <_sk_callback_sse41+0x25b>
   DB  15,89,250                           ; mulps         %xmm2,%xmm7
   DB  65,15,88,252                        ; addps         %xmm12,%xmm7
-  DB  68,15,40,53,58,59,0,0               ; movaps        0x3b3a(%rip),%xmm14        # 4a20 <_sk_callback_sse41+0x26d>
+  DB  68,15,40,53,154,59,0,0              ; movaps        0x3b9a(%rip),%xmm14        # 4a80 <_sk_callback_sse41+0x26b>
   DB  68,15,40,252                        ; movaps        %xmm4,%xmm15
   DB  69,15,89,254                        ; mulps         %xmm14,%xmm15
   DB  68,15,88,255                        ; addps         %xmm7,%xmm15
@@ -12473,7 +12545,7 @@ _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  68,15,40,13,255,57,0,0              ; movaps        0x39ff(%rip),%xmm9        # 4a30 <_sk_callback_sse41+0x27d>
+  DB  68,15,40,13,95,58,0,0               ; movaps        0x3a5f(%rip),%xmm9        # 4a90 <_sk_callback_sse41+0x27b>
   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
@@ -12567,14 +12639,14 @@ _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  15,40,21,178,56,0,0                 ; movaps        0x38b2(%rip),%xmm2        # 4a40 <_sk_callback_sse41+0x28d>
+  DB  15,40,21,18,57,0,0                  ; movaps        0x3912(%rip),%xmm2        # 4aa0 <_sk_callback_sse41+0x28b>
   DB  15,40,221                           ; movaps        %xmm5,%xmm3
   DB  15,89,218                           ; mulps         %xmm2,%xmm3
-  DB  15,40,13,181,56,0,0                 ; movaps        0x38b5(%rip),%xmm1        # 4a50 <_sk_callback_sse41+0x29d>
+  DB  15,40,13,21,57,0,0                  ; movaps        0x3915(%rip),%xmm1        # 4ab0 <_sk_callback_sse41+0x29b>
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
   DB  15,89,249                           ; mulps         %xmm1,%xmm7
   DB  15,88,251                           ; addps         %xmm3,%xmm7
-  DB  68,15,40,45,180,56,0,0              ; movaps        0x38b4(%rip),%xmm13        # 4a60 <_sk_callback_sse41+0x2ad>
+  DB  68,15,40,45,20,57,0,0               ; movaps        0x3914(%rip),%xmm13        # 4ac0 <_sk_callback_sse41+0x2ab>
   DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
   DB  68,15,88,247                        ; addps         %xmm7,%xmm14
   DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
@@ -12655,7 +12727,7 @@ _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  68,15,40,13,122,55,0,0              ; movaps        0x377a(%rip),%xmm9        # 4a70 <_sk_callback_sse41+0x2bd>
+  DB  68,15,40,13,218,55,0,0              ; movaps        0x37da(%rip),%xmm9        # 4ad0 <_sk_callback_sse41+0x2bb>
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  68,15,92,204                        ; subps         %xmm4,%xmm9
   DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
@@ -12710,14 +12782,14 @@ _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  68,15,40,13,191,54,0,0              ; movaps        0x36bf(%rip),%xmm9        # 4a80 <_sk_callback_sse41+0x2cd>
+  DB  68,15,40,13,31,55,0,0               ; movaps        0x371f(%rip),%xmm9        # 4ae0 <_sk_callback_sse41+0x2cb>
   DB  65,15,40,250                        ; movaps        %xmm10,%xmm7
   DB  65,15,89,249                        ; mulps         %xmm9,%xmm7
-  DB  68,15,40,21,191,54,0,0              ; movaps        0x36bf(%rip),%xmm10        # 4a90 <_sk_callback_sse41+0x2dd>
+  DB  68,15,40,21,31,55,0,0               ; movaps        0x371f(%rip),%xmm10        # 4af0 <_sk_callback_sse41+0x2db>
   DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
   DB  65,15,89,218                        ; mulps         %xmm10,%xmm3
   DB  15,88,223                           ; addps         %xmm7,%xmm3
-  DB  68,15,40,29,188,54,0,0              ; movaps        0x36bc(%rip),%xmm11        # 4aa0 <_sk_callback_sse41+0x2ed>
+  DB  68,15,40,29,28,55,0,0               ; movaps        0x371c(%rip),%xmm11        # 4b00 <_sk_callback_sse41+0x2eb>
   DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
   DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
   DB  68,15,88,235                        ; addps         %xmm3,%xmm13
@@ -12802,7 +12874,7 @@ _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  68,15,40,13,120,53,0,0              ; movaps        0x3578(%rip),%xmm9        # 4ab0 <_sk_callback_sse41+0x2fd>
+  DB  68,15,40,13,216,53,0,0              ; movaps        0x35d8(%rip),%xmm9        # 4b10 <_sk_callback_sse41+0x2fb>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
@@ -12854,13 +12926,13 @@ _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  68,15,40,5,202,52,0,0               ; movaps        0x34ca(%rip),%xmm8        # 4ac0 <_sk_callback_sse41+0x30d>
+  DB  68,15,40,5,42,53,0,0                ; movaps        0x352a(%rip),%xmm8        # 4b20 <_sk_callback_sse41+0x30b>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
-  DB  68,15,40,21,206,52,0,0              ; movaps        0x34ce(%rip),%xmm10        # 4ad0 <_sk_callback_sse41+0x31d>
+  DB  68,15,40,21,46,53,0,0               ; movaps        0x352e(%rip),%xmm10        # 4b30 <_sk_callback_sse41+0x31b>
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
   DB  65,15,89,234                        ; mulps         %xmm10,%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
-  DB  68,15,40,37,204,52,0,0              ; movaps        0x34cc(%rip),%xmm12        # 4ae0 <_sk_callback_sse41+0x32d>
+  DB  68,15,40,37,44,53,0,0               ; movaps        0x352c(%rip),%xmm12        # 4b40 <_sk_callback_sse41+0x32b>
   DB  68,15,40,242                        ; movaps        %xmm2,%xmm14
   DB  69,15,89,244                        ; mulps         %xmm12,%xmm14
   DB  68,15,88,245                        ; addps         %xmm5,%xmm14
@@ -12945,7 +13017,7 @@ _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  15,40,5,130,51,0,0                  ; movaps        0x3382(%rip),%xmm0        # 4af0 <_sk_callback_sse41+0x33d>
+  DB  15,40,5,226,51,0,0                  ; movaps        0x33e2(%rip),%xmm0        # 4b50 <_sk_callback_sse41+0x33b>
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,92,215                           ; subps         %xmm7,%xmm2
   DB  15,89,226                           ; mulps         %xmm2,%xmm4
@@ -12991,7 +13063,7 @@ _sk_clamp_0_sse41 LABEL PROC
 
 PUBLIC _sk_clamp_1_sse41
 _sk_clamp_1_sse41 LABEL PROC
-  DB  68,15,40,5,2,51,0,0                 ; movaps        0x3302(%rip),%xmm8        # 4b00 <_sk_callback_sse41+0x34d>
+  DB  68,15,40,5,98,51,0,0                ; movaps        0x3362(%rip),%xmm8        # 4b60 <_sk_callback_sse41+0x34b>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
   DB  65,15,93,208                        ; minps         %xmm8,%xmm2
@@ -13001,7 +13073,7 @@ _sk_clamp_1_sse41 LABEL PROC
 
 PUBLIC _sk_clamp_a_sse41
 _sk_clamp_a_sse41 LABEL PROC
-  DB  15,93,29,247,50,0,0                 ; minps         0x32f7(%rip),%xmm3        # 4b10 <_sk_callback_sse41+0x35d>
+  DB  15,93,29,87,51,0,0                  ; minps         0x3357(%rip),%xmm3        # 4b70 <_sk_callback_sse41+0x35b>
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -13074,7 +13146,7 @@ _sk_premul_sse41 LABEL PROC
 PUBLIC _sk_unpremul_sse41
 _sk_unpremul_sse41 LABEL PROC
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
-  DB  68,15,40,13,98,50,0,0               ; movaps        0x3262(%rip),%xmm9        # 4b20 <_sk_callback_sse41+0x36d>
+  DB  68,15,40,13,194,50,0,0              ; movaps        0x32c2(%rip),%xmm9        # 4b80 <_sk_callback_sse41+0x36b>
   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
@@ -13086,20 +13158,20 @@ _sk_unpremul_sse41 LABEL PROC
 
 PUBLIC _sk_from_srgb_sse41
 _sk_from_srgb_sse41 LABEL PROC
-  DB  68,15,40,29,77,50,0,0               ; movaps        0x324d(%rip),%xmm11        # 4b30 <_sk_callback_sse41+0x37d>
+  DB  68,15,40,29,173,50,0,0              ; movaps        0x32ad(%rip),%xmm11        # 4b90 <_sk_callback_sse41+0x37b>
   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  68,15,40,37,69,50,0,0               ; movaps        0x3245(%rip),%xmm12        # 4b40 <_sk_callback_sse41+0x38d>
+  DB  68,15,40,37,165,50,0,0              ; movaps        0x32a5(%rip),%xmm12        # 4ba0 <_sk_callback_sse41+0x38b>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
-  DB  68,15,40,45,69,50,0,0               ; movaps        0x3245(%rip),%xmm13        # 4b50 <_sk_callback_sse41+0x39d>
+  DB  68,15,40,45,165,50,0,0              ; movaps        0x32a5(%rip),%xmm13        # 4bb0 <_sk_callback_sse41+0x39b>
   DB  69,15,88,197                        ; addps         %xmm13,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
-  DB  68,15,40,53,69,50,0,0               ; movaps        0x3245(%rip),%xmm14        # 4b60 <_sk_callback_sse41+0x3ad>
+  DB  68,15,40,53,165,50,0,0              ; movaps        0x32a5(%rip),%xmm14        # 4bc0 <_sk_callback_sse41+0x3ab>
   DB  69,15,88,198                        ; addps         %xmm14,%xmm8
-  DB  68,15,40,61,73,50,0,0               ; movaps        0x3249(%rip),%xmm15        # 4b70 <_sk_callback_sse41+0x3bd>
+  DB  68,15,40,61,169,50,0,0              ; movaps        0x32a9(%rip),%xmm15        # 4bd0 <_sk_callback_sse41+0x3bb>
   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
@@ -13141,22 +13213,22 @@ _sk_to_srgb_sse41 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  15,40,209                           ; movaps        %xmm1,%xmm2
   DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
-  DB  68,15,40,29,190,49,0,0              ; movaps        0x31be(%rip),%xmm11        # 4b80 <_sk_callback_sse41+0x3cd>
+  DB  68,15,40,29,30,50,0,0               ; movaps        0x321e(%rip),%xmm11        # 4be0 <_sk_callback_sse41+0x3cb>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
-  DB  68,15,40,37,190,49,0,0              ; movaps        0x31be(%rip),%xmm12        # 4b90 <_sk_callback_sse41+0x3dd>
+  DB  68,15,40,37,30,50,0,0               ; movaps        0x321e(%rip),%xmm12        # 4bf0 <_sk_callback_sse41+0x3db>
   DB  69,15,40,248                        ; movaps        %xmm8,%xmm15
   DB  69,15,89,252                        ; mulps         %xmm12,%xmm15
-  DB  68,15,40,21,190,49,0,0              ; movaps        0x31be(%rip),%xmm10        # 4ba0 <_sk_callback_sse41+0x3ed>
+  DB  68,15,40,21,30,50,0,0               ; movaps        0x321e(%rip),%xmm10        # 4c00 <_sk_callback_sse41+0x3eb>
   DB  69,15,88,250                        ; addps         %xmm10,%xmm15
   DB  69,15,89,248                        ; mulps         %xmm8,%xmm15
-  DB  68,15,40,45,190,49,0,0              ; movaps        0x31be(%rip),%xmm13        # 4bb0 <_sk_callback_sse41+0x3fd>
+  DB  68,15,40,45,30,50,0,0               ; movaps        0x321e(%rip),%xmm13        # 4c10 <_sk_callback_sse41+0x3fb>
   DB  69,15,88,253                        ; addps         %xmm13,%xmm15
-  DB  68,15,40,53,194,49,0,0              ; movaps        0x31c2(%rip),%xmm14        # 4bc0 <_sk_callback_sse41+0x40d>
+  DB  68,15,40,53,34,50,0,0               ; movaps        0x3222(%rip),%xmm14        # 4c20 <_sk_callback_sse41+0x40b>
   DB  69,15,88,198                        ; addps         %xmm14,%xmm8
   DB  69,15,83,192                        ; rcpps         %xmm8,%xmm8
   DB  69,15,89,199                        ; mulps         %xmm15,%xmm8
-  DB  68,15,40,61,190,49,0,0              ; movaps        0x31be(%rip),%xmm15        # 4bd0 <_sk_callback_sse41+0x41d>
+  DB  68,15,40,61,30,50,0,0               ; movaps        0x321e(%rip),%xmm15        # 4c30 <_sk_callback_sse41+0x41b>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
   DB  68,15,82,202                        ; rsqrtps       %xmm2,%xmm9
@@ -13208,7 +13280,7 @@ _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  68,15,40,53,8,49,0,0                ; movaps        0x3108(%rip),%xmm14        # 4be0 <_sk_callback_sse41+0x42d>
+  DB  68,15,40,53,104,49,0,0              ; movaps        0x3168(%rip),%xmm14        # 4c40 <_sk_callback_sse41+0x42b>
   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
@@ -13217,27 +13289,27 @@ _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  68,15,84,61,239,48,0,0              ; andps         0x30ef(%rip),%xmm15        # 4bf0 <_sk_callback_sse41+0x43d>
+  DB  68,15,84,61,79,49,0,0               ; andps         0x314f(%rip),%xmm15        # 4c50 <_sk_callback_sse41+0x43b>
   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  68,15,40,45,226,48,0,0              ; movaps        0x30e2(%rip),%xmm13        # 4c00 <_sk_callback_sse41+0x44d>
+  DB  68,15,40,45,66,49,0,0               ; movaps        0x3142(%rip),%xmm13        # 4c60 <_sk_callback_sse41+0x44b>
   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  68,15,88,5,222,48,0,0               ; addps         0x30de(%rip),%xmm8        # 4c10 <_sk_callback_sse41+0x45d>
+  DB  68,15,88,5,62,49,0,0                ; addps         0x313e(%rip),%xmm8        # 4c70 <_sk_callback_sse41+0x45b>
   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  68,15,89,5,214,48,0,0               ; mulps         0x30d6(%rip),%xmm8        # 4c20 <_sk_callback_sse41+0x46d>
+  DB  68,15,89,5,54,49,0,0                ; mulps         0x3136(%rip),%xmm8        # 4c80 <_sk_callback_sse41+0x46b>
   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  15,40,5,202,48,0,0                  ; movaps        0x30ca(%rip),%xmm0        # 4c30 <_sk_callback_sse41+0x47d>
+  DB  15,40,5,42,49,0,0                   ; movaps        0x312a(%rip),%xmm0        # 4c90 <_sk_callback_sse41+0x47b>
   DB  65,15,40,211                        ; movaps        %xmm11,%xmm2
   DB  15,89,208                           ; mulps         %xmm0,%xmm2
   DB  15,194,194,1                        ; cmpltps       %xmm2,%xmm0
@@ -13258,7 +13330,7 @@ _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  68,15,40,13,140,48,0,0              ; movaps        0x308c(%rip),%xmm9        # 4c40 <_sk_callback_sse41+0x48d>
+  DB  68,15,40,13,236,48,0,0              ; movaps        0x30ec(%rip),%xmm9        # 4ca0 <_sk_callback_sse41+0x48b>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,194,194,2                        ; cmpleps       %xmm2,%xmm0
   DB  15,40,217                           ; movaps        %xmm1,%xmm3
@@ -13271,19 +13343,19 @@ _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  68,15,40,53,104,48,0,0              ; movaps        0x3068(%rip),%xmm14        # 4c50 <_sk_callback_sse41+0x49d>
+  DB  68,15,40,53,200,48,0,0              ; movaps        0x30c8(%rip),%xmm14        # 4cb0 <_sk_callback_sse41+0x49b>
   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  68,15,40,29,97,48,0,0               ; movaps        0x3061(%rip),%xmm11        # 4c60 <_sk_callback_sse41+0x4ad>
+  DB  68,15,40,29,193,48,0,0              ; movaps        0x30c1(%rip),%xmm11        # 4cc0 <_sk_callback_sse41+0x4ab>
   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  15,40,61,90,48,0,0                  ; movaps        0x305a(%rip),%xmm7        # 4c70 <_sk_callback_sse41+0x4bd>
+  DB  15,40,61,186,48,0,0                 ; movaps        0x30ba(%rip),%xmm7        # 4cd0 <_sk_callback_sse41+0x4bb>
   DB  69,15,40,238                        ; movaps        %xmm14,%xmm13
   DB  68,15,89,239                        ; mulps         %xmm7,%xmm13
-  DB  15,40,29,91,48,0,0                  ; movaps        0x305b(%rip),%xmm3        # 4c80 <_sk_callback_sse41+0x4cd>
+  DB  15,40,29,187,48,0,0                 ; movaps        0x30bb(%rip),%xmm3        # 4ce0 <_sk_callback_sse41+0x4cb>
   DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  68,15,89,230                        ; mulps         %xmm6,%xmm12
@@ -13293,7 +13365,7 @@ _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  68,15,40,37,58,48,0,0               ; movaps        0x303a(%rip),%xmm12        # 4c90 <_sk_callback_sse41+0x4dd>
+  DB  68,15,40,37,154,48,0,0              ; movaps        0x309a(%rip),%xmm12        # 4cf0 <_sk_callback_sse41+0x4db>
   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
@@ -13327,7 +13399,7 @@ _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  68,15,88,21,179,47,0,0              ; addps         0x2fb3(%rip),%xmm10        # 4ca0 <_sk_callback_sse41+0x4ed>
+  DB  68,15,88,21,19,48,0,0               ; addps         0x3013(%rip),%xmm10        # 4d00 <_sk_callback_sse41+0x4eb>
   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
@@ -13376,7 +13448,7 @@ _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  68,15,89,5,12,47,0,0                ; mulps         0x2f0c(%rip),%xmm8        # 4cb0 <_sk_callback_sse41+0x4fd>
+  DB  68,15,89,5,108,47,0,0               ; mulps         0x2f6c(%rip),%xmm8        # 4d10 <_sk_callback_sse41+0x4fb>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
@@ -13410,7 +13482,7 @@ _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  68,15,89,5,184,46,0,0               ; mulps         0x2eb8(%rip),%xmm8        # 4cc0 <_sk_callback_sse41+0x50d>
+  DB  68,15,89,5,24,47,0,0                ; mulps         0x2f18(%rip),%xmm8        # 4d20 <_sk_callback_sse41+0x50b>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -13431,17 +13503,17 @@ _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  102,68,15,111,5,135,46,0,0          ; movdqa        0x2e87(%rip),%xmm8        # 4cd0 <_sk_callback_sse41+0x51d>
+  DB  102,68,15,111,5,231,46,0,0          ; movdqa        0x2ee7(%rip),%xmm8        # 4d30 <_sk_callback_sse41+0x51b>
   DB  102,69,15,219,194                   ; pand          %xmm10,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,134,46,0,0               ; mulps         0x2e86(%rip),%xmm8        # 4ce0 <_sk_callback_sse41+0x52d>
-  DB  102,68,15,111,13,141,46,0,0         ; movdqa        0x2e8d(%rip),%xmm9        # 4cf0 <_sk_callback_sse41+0x53d>
+  DB  68,15,89,5,230,46,0,0               ; mulps         0x2ee6(%rip),%xmm8        # 4d40 <_sk_callback_sse41+0x52b>
+  DB  102,68,15,111,13,237,46,0,0         ; movdqa        0x2eed(%rip),%xmm9        # 4d50 <_sk_callback_sse41+0x53b>
   DB  102,69,15,219,202                   ; pand          %xmm10,%xmm9
   DB  69,15,91,201                        ; cvtdq2ps      %xmm9,%xmm9
-  DB  68,15,89,13,140,46,0,0              ; mulps         0x2e8c(%rip),%xmm9        # 4d00 <_sk_callback_sse41+0x54d>
-  DB  102,68,15,219,21,147,46,0,0         ; pand          0x2e93(%rip),%xmm10        # 4d10 <_sk_callback_sse41+0x55d>
+  DB  68,15,89,13,236,46,0,0              ; mulps         0x2eec(%rip),%xmm9        # 4d60 <_sk_callback_sse41+0x54b>
+  DB  102,68,15,219,21,243,46,0,0         ; pand          0x2ef3(%rip),%xmm10        # 4d70 <_sk_callback_sse41+0x55b>
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
-  DB  68,15,89,21,151,46,0,0              ; mulps         0x2e97(%rip),%xmm10        # 4d20 <_sk_callback_sse41+0x56d>
+  DB  68,15,89,21,247,46,0,0              ; mulps         0x2ef7(%rip),%xmm10        # 4d80 <_sk_callback_sse41+0x56b>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -13470,7 +13542,7 @@ _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  102,15,111,5,72,46,0,0              ; movdqa        0x2e48(%rip),%xmm0        # 4d30 <_sk_callback_sse41+0x57d>
+  DB  102,15,111,5,168,46,0,0             ; movdqa        0x2ea8(%rip),%xmm0        # 4d90 <_sk_callback_sse41+0x57b>
   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
@@ -13485,7 +13557,7 @@ _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  102,15,56,0,13,3,46,0,0             ; pshufb        0x2e03(%rip),%xmm1        # 4d40 <_sk_callback_sse41+0x58d>
+  DB  102,15,56,0,13,99,46,0,0            ; pshufb        0x2e63(%rip),%xmm1        # 4da0 <_sk_callback_sse41+0x58b>
   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
@@ -13500,7 +13572,7 @@ _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  102,15,56,0,21,191,45,0,0           ; pshufb        0x2dbf(%rip),%xmm2        # 4d50 <_sk_callback_sse41+0x59d>
+  DB  102,15,56,0,21,31,46,0,0            ; pshufb        0x2e1f(%rip),%xmm2        # 4db0 <_sk_callback_sse41+0x59b>
   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
@@ -13515,7 +13587,7 @@ _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  15,89,29,124,45,0,0                 ; mulps         0x2d7c(%rip),%xmm3        # 4d60 <_sk_callback_sse41+0x5ad>
+  DB  15,89,29,220,45,0,0                 ; mulps         0x2ddc(%rip),%xmm3        # 4dc0 <_sk_callback_sse41+0x5ab>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -13532,7 +13604,7 @@ _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  102,68,15,111,5,79,45,0,0           ; movdqa        0x2d4f(%rip),%xmm8        # 4d70 <_sk_callback_sse41+0x5bd>
+  DB  102,68,15,111,5,175,45,0,0          ; movdqa        0x2daf(%rip),%xmm8        # 4dd0 <_sk_callback_sse41+0x5bb>
   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
@@ -13549,7 +13621,7 @@ _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  102,15,56,0,13,2,45,0,0             ; pshufb        0x2d02(%rip),%xmm1        # 4d80 <_sk_callback_sse41+0x5cd>
+  DB  102,15,56,0,13,98,45,0,0            ; pshufb        0x2d62(%rip),%xmm1        # 4de0 <_sk_callback_sse41+0x5cb>
   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
@@ -13585,7 +13657,7 @@ _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  15,89,29,80,44,0,0                  ; mulps         0x2c50(%rip),%xmm3        # 4d90 <_sk_callback_sse41+0x5dd>
+  DB  15,89,29,176,44,0,0                 ; mulps         0x2cb0(%rip),%xmm3        # 4df0 <_sk_callback_sse41+0x5db>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -13605,7 +13677,7 @@ _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  102,68,15,111,5,18,44,0,0           ; movdqa        0x2c12(%rip),%xmm8        # 4da0 <_sk_callback_sse41+0x5ed>
+  DB  102,68,15,111,5,114,44,0,0          ; movdqa        0x2c72(%rip),%xmm8        # 4e00 <_sk_callback_sse41+0x5eb>
   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
@@ -13622,7 +13694,7 @@ _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  102,15,56,0,13,197,43,0,0           ; pshufb        0x2bc5(%rip),%xmm1        # 4db0 <_sk_callback_sse41+0x5fd>
+  DB  102,15,56,0,13,37,44,0,0            ; pshufb        0x2c25(%rip),%xmm1        # 4e10 <_sk_callback_sse41+0x5fb>
   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
@@ -13653,7 +13725,7 @@ _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  15,40,29,48,43,0,0                  ; movaps        0x2b30(%rip),%xmm3        # 4dc0 <_sk_callback_sse41+0x60d>
+  DB  15,40,29,144,43,0,0                 ; movaps        0x2b90(%rip),%xmm3        # 4e20 <_sk_callback_sse41+0x60b>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_byte_tables_sse41
@@ -13661,7 +13733,7 @@ _sk_byte_tables_sse41 LABEL PROC
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,49,43,0,0                ; movaps        0x2b31(%rip),%xmm8        # 4dd0 <_sk_callback_sse41+0x61d>
+  DB  68,15,40,5,145,43,0,0               ; movaps        0x2b91(%rip),%xmm8        # 4e30 <_sk_callback_sse41+0x61b>
   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
@@ -13680,7 +13752,7 @@ _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  68,15,40,13,226,42,0,0              ; movaps        0x2ae2(%rip),%xmm9        # 4de0 <_sk_callback_sse41+0x62d>
+  DB  68,15,40,13,66,43,0,0               ; movaps        0x2b42(%rip),%xmm9        # 4e40 <_sk_callback_sse41+0x62b>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -13769,7 +13841,7 @@ _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  68,15,40,13,106,41,0,0              ; movaps        0x296a(%rip),%xmm9        # 4df0 <_sk_callback_sse41+0x63d>
+  DB  68,15,40,13,202,41,0,0              ; movaps        0x29ca(%rip),%xmm9        # 4e50 <_sk_callback_sse41+0x63b>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -13936,31 +14008,31 @@ _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  68,15,89,5,193,38,0,0               ; mulps         0x26c1(%rip),%xmm8        # 4e00 <_sk_callback_sse41+0x64d>
-  DB  68,15,84,21,201,38,0,0              ; andps         0x26c9(%rip),%xmm10        # 4e10 <_sk_callback_sse41+0x65d>
-  DB  68,15,86,21,209,38,0,0              ; orps          0x26d1(%rip),%xmm10        # 4e20 <_sk_callback_sse41+0x66d>
-  DB  68,15,88,5,217,38,0,0               ; addps         0x26d9(%rip),%xmm8        # 4e30 <_sk_callback_sse41+0x67d>
-  DB  68,15,40,37,225,38,0,0              ; movaps        0x26e1(%rip),%xmm12        # 4e40 <_sk_callback_sse41+0x68d>
+  DB  68,15,89,5,33,39,0,0                ; mulps         0x2721(%rip),%xmm8        # 4e60 <_sk_callback_sse41+0x64b>
+  DB  68,15,84,21,41,39,0,0               ; andps         0x2729(%rip),%xmm10        # 4e70 <_sk_callback_sse41+0x65b>
+  DB  68,15,86,21,49,39,0,0               ; orps          0x2731(%rip),%xmm10        # 4e80 <_sk_callback_sse41+0x66b>
+  DB  68,15,88,5,57,39,0,0                ; addps         0x2739(%rip),%xmm8        # 4e90 <_sk_callback_sse41+0x67b>
+  DB  68,15,40,37,65,39,0,0               ; movaps        0x2741(%rip),%xmm12        # 4ea0 <_sk_callback_sse41+0x68b>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,196                        ; subps         %xmm12,%xmm8
-  DB  68,15,88,21,225,38,0,0              ; addps         0x26e1(%rip),%xmm10        # 4e50 <_sk_callback_sse41+0x69d>
-  DB  68,15,40,37,233,38,0,0              ; movaps        0x26e9(%rip),%xmm12        # 4e60 <_sk_callback_sse41+0x6ad>
+  DB  68,15,88,21,65,39,0,0               ; addps         0x2741(%rip),%xmm10        # 4eb0 <_sk_callback_sse41+0x69b>
+  DB  68,15,40,37,73,39,0,0               ; movaps        0x2749(%rip),%xmm12        # 4ec0 <_sk_callback_sse41+0x6ab>
   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  68,15,88,5,214,38,0,0               ; addps         0x26d6(%rip),%xmm8        # 4e70 <_sk_callback_sse41+0x6bd>
-  DB  68,15,40,21,222,38,0,0              ; movaps        0x26de(%rip),%xmm10        # 4e80 <_sk_callback_sse41+0x6cd>
+  DB  68,15,88,5,54,39,0,0                ; addps         0x2736(%rip),%xmm8        # 4ed0 <_sk_callback_sse41+0x6bb>
+  DB  68,15,40,21,62,39,0,0               ; movaps        0x273e(%rip),%xmm10        # 4ee0 <_sk_callback_sse41+0x6cb>
   DB  69,15,89,211                        ; mulps         %xmm11,%xmm10
   DB  69,15,92,194                        ; subps         %xmm10,%xmm8
-  DB  68,15,40,21,222,38,0,0              ; movaps        0x26de(%rip),%xmm10        # 4e90 <_sk_callback_sse41+0x6dd>
+  DB  68,15,40,21,62,39,0,0               ; movaps        0x273e(%rip),%xmm10        # 4ef0 <_sk_callback_sse41+0x6db>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  68,15,40,29,226,38,0,0              ; movaps        0x26e2(%rip),%xmm11        # 4ea0 <_sk_callback_sse41+0x6ed>
+  DB  68,15,40,29,66,39,0,0               ; movaps        0x2742(%rip),%xmm11        # 4f00 <_sk_callback_sse41+0x6eb>
   DB  69,15,94,218                        ; divps         %xmm10,%xmm11
   DB  69,15,88,216                        ; addps         %xmm8,%xmm11
-  DB  68,15,89,29,226,38,0,0              ; mulps         0x26e2(%rip),%xmm11        # 4eb0 <_sk_callback_sse41+0x6fd>
+  DB  68,15,89,29,66,39,0,0               ; mulps         0x2742(%rip),%xmm11        # 4f10 <_sk_callback_sse41+0x6fb>
   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
@@ -13968,7 +14040,7 @@ _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  68,15,93,5,201,38,0,0               ; minps         0x26c9(%rip),%xmm8        # 4ec0 <_sk_callback_sse41+0x70d>
+  DB  68,15,93,5,41,39,0,0                ; minps         0x2729(%rip),%xmm8        # 4f20 <_sk_callback_sse41+0x70b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -13996,31 +14068,31 @@ _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  68,15,89,37,106,38,0,0              ; mulps         0x266a(%rip),%xmm12        # 4ed0 <_sk_callback_sse41+0x71d>
-  DB  68,15,84,29,114,38,0,0              ; andps         0x2672(%rip),%xmm11        # 4ee0 <_sk_callback_sse41+0x72d>
-  DB  68,15,86,29,122,38,0,0              ; orps          0x267a(%rip),%xmm11        # 4ef0 <_sk_callback_sse41+0x73d>
-  DB  68,15,88,37,130,38,0,0              ; addps         0x2682(%rip),%xmm12        # 4f00 <_sk_callback_sse41+0x74d>
-  DB  15,40,13,139,38,0,0                 ; movaps        0x268b(%rip),%xmm1        # 4f10 <_sk_callback_sse41+0x75d>
+  DB  68,15,89,37,202,38,0,0              ; mulps         0x26ca(%rip),%xmm12        # 4f30 <_sk_callback_sse41+0x71b>
+  DB  68,15,84,29,210,38,0,0              ; andps         0x26d2(%rip),%xmm11        # 4f40 <_sk_callback_sse41+0x72b>
+  DB  68,15,86,29,218,38,0,0              ; orps          0x26da(%rip),%xmm11        # 4f50 <_sk_callback_sse41+0x73b>
+  DB  68,15,88,37,226,38,0,0              ; addps         0x26e2(%rip),%xmm12        # 4f60 <_sk_callback_sse41+0x74b>
+  DB  15,40,13,235,38,0,0                 ; movaps        0x26eb(%rip),%xmm1        # 4f70 <_sk_callback_sse41+0x75b>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
-  DB  68,15,88,29,139,38,0,0              ; addps         0x268b(%rip),%xmm11        # 4f20 <_sk_callback_sse41+0x76d>
-  DB  15,40,13,148,38,0,0                 ; movaps        0x2694(%rip),%xmm1        # 4f30 <_sk_callback_sse41+0x77d>
+  DB  68,15,88,29,235,38,0,0              ; addps         0x26eb(%rip),%xmm11        # 4f80 <_sk_callback_sse41+0x76b>
+  DB  15,40,13,244,38,0,0                 ; movaps        0x26f4(%rip),%xmm1        # 4f90 <_sk_callback_sse41+0x77b>
   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  68,15,88,37,129,38,0,0              ; addps         0x2681(%rip),%xmm12        # 4f40 <_sk_callback_sse41+0x78d>
-  DB  15,40,13,138,38,0,0                 ; movaps        0x268a(%rip),%xmm1        # 4f50 <_sk_callback_sse41+0x79d>
+  DB  68,15,88,37,225,38,0,0              ; addps         0x26e1(%rip),%xmm12        # 4fa0 <_sk_callback_sse41+0x78b>
+  DB  15,40,13,234,38,0,0                 ; movaps        0x26ea(%rip),%xmm1        # 4fb0 <_sk_callback_sse41+0x79b>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
-  DB  68,15,40,21,138,38,0,0              ; movaps        0x268a(%rip),%xmm10        # 4f60 <_sk_callback_sse41+0x7ad>
+  DB  68,15,40,21,234,38,0,0              ; movaps        0x26ea(%rip),%xmm10        # 4fc0 <_sk_callback_sse41+0x7ab>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,13,143,38,0,0                 ; movaps        0x268f(%rip),%xmm1        # 4f70 <_sk_callback_sse41+0x7bd>
+  DB  15,40,13,239,38,0,0                 ; movaps        0x26ef(%rip),%xmm1        # 4fd0 <_sk_callback_sse41+0x7bb>
   DB  65,15,94,202                        ; divps         %xmm10,%xmm1
   DB  65,15,88,204                        ; addps         %xmm12,%xmm1
-  DB  15,89,13,144,38,0,0                 ; mulps         0x2690(%rip),%xmm1        # 4f80 <_sk_callback_sse41+0x7cd>
+  DB  15,89,13,240,38,0,0                 ; mulps         0x26f0(%rip),%xmm1        # 4fe0 <_sk_callback_sse41+0x7cb>
   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
@@ -14028,7 +14100,7 @@ _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  15,93,13,123,38,0,0                 ; minps         0x267b(%rip),%xmm1        # 4f90 <_sk_callback_sse41+0x7dd>
+  DB  15,93,13,219,38,0,0                 ; minps         0x26db(%rip),%xmm1        # 4ff0 <_sk_callback_sse41+0x7db>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -14056,31 +14128,31 @@ _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  68,15,89,37,28,38,0,0               ; mulps         0x261c(%rip),%xmm12        # 4fa0 <_sk_callback_sse41+0x7ed>
-  DB  68,15,84,29,36,38,0,0               ; andps         0x2624(%rip),%xmm11        # 4fb0 <_sk_callback_sse41+0x7fd>
-  DB  68,15,86,29,44,38,0,0               ; orps          0x262c(%rip),%xmm11        # 4fc0 <_sk_callback_sse41+0x80d>
-  DB  68,15,88,37,52,38,0,0               ; addps         0x2634(%rip),%xmm12        # 4fd0 <_sk_callback_sse41+0x81d>
-  DB  15,40,21,61,38,0,0                  ; movaps        0x263d(%rip),%xmm2        # 4fe0 <_sk_callback_sse41+0x82d>
+  DB  68,15,89,37,124,38,0,0              ; mulps         0x267c(%rip),%xmm12        # 5000 <_sk_callback_sse41+0x7eb>
+  DB  68,15,84,29,132,38,0,0              ; andps         0x2684(%rip),%xmm11        # 5010 <_sk_callback_sse41+0x7fb>
+  DB  68,15,86,29,140,38,0,0              ; orps          0x268c(%rip),%xmm11        # 5020 <_sk_callback_sse41+0x80b>
+  DB  68,15,88,37,148,38,0,0              ; addps         0x2694(%rip),%xmm12        # 5030 <_sk_callback_sse41+0x81b>
+  DB  15,40,21,157,38,0,0                 ; movaps        0x269d(%rip),%xmm2        # 5040 <_sk_callback_sse41+0x82b>
   DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
-  DB  68,15,88,29,61,38,0,0               ; addps         0x263d(%rip),%xmm11        # 4ff0 <_sk_callback_sse41+0x83d>
-  DB  15,40,21,70,38,0,0                  ; movaps        0x2646(%rip),%xmm2        # 5000 <_sk_callback_sse41+0x84d>
+  DB  68,15,88,29,157,38,0,0              ; addps         0x269d(%rip),%xmm11        # 5050 <_sk_callback_sse41+0x83b>
+  DB  15,40,21,166,38,0,0                 ; movaps        0x26a6(%rip),%xmm2        # 5060 <_sk_callback_sse41+0x84b>
   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  68,15,88,37,51,38,0,0               ; addps         0x2633(%rip),%xmm12        # 5010 <_sk_callback_sse41+0x85d>
-  DB  15,40,21,60,38,0,0                  ; movaps        0x263c(%rip),%xmm2        # 5020 <_sk_callback_sse41+0x86d>
+  DB  68,15,88,37,147,38,0,0              ; addps         0x2693(%rip),%xmm12        # 5070 <_sk_callback_sse41+0x85b>
+  DB  15,40,21,156,38,0,0                 ; movaps        0x269c(%rip),%xmm2        # 5080 <_sk_callback_sse41+0x86b>
   DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
-  DB  68,15,40,21,60,38,0,0               ; movaps        0x263c(%rip),%xmm10        # 5030 <_sk_callback_sse41+0x87d>
+  DB  68,15,40,21,156,38,0,0              ; movaps        0x269c(%rip),%xmm10        # 5090 <_sk_callback_sse41+0x87b>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,21,65,38,0,0                  ; movaps        0x2641(%rip),%xmm2        # 5040 <_sk_callback_sse41+0x88d>
+  DB  15,40,21,161,38,0,0                 ; movaps        0x26a1(%rip),%xmm2        # 50a0 <_sk_callback_sse41+0x88b>
   DB  65,15,94,210                        ; divps         %xmm10,%xmm2
   DB  65,15,88,212                        ; addps         %xmm12,%xmm2
-  DB  15,89,21,66,38,0,0                  ; mulps         0x2642(%rip),%xmm2        # 5050 <_sk_callback_sse41+0x89d>
+  DB  15,89,21,162,38,0,0                 ; mulps         0x26a2(%rip),%xmm2        # 50b0 <_sk_callback_sse41+0x89b>
   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
@@ -14088,7 +14160,7 @@ _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  15,93,21,45,38,0,0                  ; minps         0x262d(%rip),%xmm2        # 5060 <_sk_callback_sse41+0x8ad>
+  DB  15,93,21,141,38,0,0                 ; minps         0x268d(%rip),%xmm2        # 50c0 <_sk_callback_sse41+0x8ab>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -14116,31 +14188,31 @@ _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  68,15,89,37,206,37,0,0              ; mulps         0x25ce(%rip),%xmm12        # 5070 <_sk_callback_sse41+0x8bd>
-  DB  68,15,84,29,214,37,0,0              ; andps         0x25d6(%rip),%xmm11        # 5080 <_sk_callback_sse41+0x8cd>
-  DB  68,15,86,29,222,37,0,0              ; orps          0x25de(%rip),%xmm11        # 5090 <_sk_callback_sse41+0x8dd>
-  DB  68,15,88,37,230,37,0,0              ; addps         0x25e6(%rip),%xmm12        # 50a0 <_sk_callback_sse41+0x8ed>
-  DB  15,40,29,239,37,0,0                 ; movaps        0x25ef(%rip),%xmm3        # 50b0 <_sk_callback_sse41+0x8fd>
+  DB  68,15,89,37,46,38,0,0               ; mulps         0x262e(%rip),%xmm12        # 50d0 <_sk_callback_sse41+0x8bb>
+  DB  68,15,84,29,54,38,0,0               ; andps         0x2636(%rip),%xmm11        # 50e0 <_sk_callback_sse41+0x8cb>
+  DB  68,15,86,29,62,38,0,0               ; orps          0x263e(%rip),%xmm11        # 50f0 <_sk_callback_sse41+0x8db>
+  DB  68,15,88,37,70,38,0,0               ; addps         0x2646(%rip),%xmm12        # 5100 <_sk_callback_sse41+0x8eb>
+  DB  15,40,29,79,38,0,0                  ; movaps        0x264f(%rip),%xmm3        # 5110 <_sk_callback_sse41+0x8fb>
   DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
-  DB  68,15,88,29,239,37,0,0              ; addps         0x25ef(%rip),%xmm11        # 50c0 <_sk_callback_sse41+0x90d>
-  DB  15,40,29,248,37,0,0                 ; movaps        0x25f8(%rip),%xmm3        # 50d0 <_sk_callback_sse41+0x91d>
+  DB  68,15,88,29,79,38,0,0               ; addps         0x264f(%rip),%xmm11        # 5120 <_sk_callback_sse41+0x90b>
+  DB  15,40,29,88,38,0,0                  ; movaps        0x2658(%rip),%xmm3        # 5130 <_sk_callback_sse41+0x91b>
   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  68,15,88,37,229,37,0,0              ; addps         0x25e5(%rip),%xmm12        # 50e0 <_sk_callback_sse41+0x92d>
-  DB  15,40,29,238,37,0,0                 ; movaps        0x25ee(%rip),%xmm3        # 50f0 <_sk_callback_sse41+0x93d>
+  DB  68,15,88,37,69,38,0,0               ; addps         0x2645(%rip),%xmm12        # 5140 <_sk_callback_sse41+0x92b>
+  DB  15,40,29,78,38,0,0                  ; movaps        0x264e(%rip),%xmm3        # 5150 <_sk_callback_sse41+0x93b>
   DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
-  DB  68,15,40,21,238,37,0,0              ; movaps        0x25ee(%rip),%xmm10        # 5100 <_sk_callback_sse41+0x94d>
+  DB  68,15,40,21,78,38,0,0               ; movaps        0x264e(%rip),%xmm10        # 5160 <_sk_callback_sse41+0x94b>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,29,243,37,0,0                 ; movaps        0x25f3(%rip),%xmm3        # 5110 <_sk_callback_sse41+0x95d>
+  DB  15,40,29,83,38,0,0                  ; movaps        0x2653(%rip),%xmm3        # 5170 <_sk_callback_sse41+0x95b>
   DB  65,15,94,218                        ; divps         %xmm10,%xmm3
   DB  65,15,88,220                        ; addps         %xmm12,%xmm3
-  DB  15,89,29,244,37,0,0                 ; mulps         0x25f4(%rip),%xmm3        # 5120 <_sk_callback_sse41+0x96d>
+  DB  15,89,29,84,38,0,0                  ; mulps         0x2654(%rip),%xmm3        # 5180 <_sk_callback_sse41+0x96b>
   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
@@ -14148,7 +14220,7 @@ _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  15,93,29,223,37,0,0                 ; minps         0x25df(%rip),%xmm3        # 5130 <_sk_callback_sse41+0x97d>
+  DB  15,93,29,63,38,0,0                  ; minps         0x263f(%rip),%xmm3        # 5190 <_sk_callback_sse41+0x97b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -14156,29 +14228,29 @@ _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  68,15,89,5,219,37,0,0               ; mulps         0x25db(%rip),%xmm8        # 5140 <_sk_callback_sse41+0x98d>
-  DB  68,15,40,13,227,37,0,0              ; movaps        0x25e3(%rip),%xmm9        # 5150 <_sk_callback_sse41+0x99d>
+  DB  68,15,89,5,59,38,0,0                ; mulps         0x263b(%rip),%xmm8        # 51a0 <_sk_callback_sse41+0x98b>
+  DB  68,15,40,13,67,38,0,0               ; movaps        0x2643(%rip),%xmm9        # 51b0 <_sk_callback_sse41+0x99b>
   DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
-  DB  15,40,5,232,37,0,0                  ; movaps        0x25e8(%rip),%xmm0        # 5160 <_sk_callback_sse41+0x9ad>
+  DB  15,40,5,72,38,0,0                   ; movaps        0x2648(%rip),%xmm0        # 51c0 <_sk_callback_sse41+0x9ab>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
-  DB  68,15,88,5,230,37,0,0               ; addps         0x25e6(%rip),%xmm8        # 5170 <_sk_callback_sse41+0x9bd>
-  DB  68,15,89,5,238,37,0,0               ; mulps         0x25ee(%rip),%xmm8        # 5180 <_sk_callback_sse41+0x9cd>
-  DB  15,89,13,247,37,0,0                 ; mulps         0x25f7(%rip),%xmm1        # 5190 <_sk_callback_sse41+0x9dd>
+  DB  68,15,88,5,70,38,0,0                ; addps         0x2646(%rip),%xmm8        # 51d0 <_sk_callback_sse41+0x9bb>
+  DB  68,15,89,5,78,38,0,0                ; mulps         0x264e(%rip),%xmm8        # 51e0 <_sk_callback_sse41+0x9cb>
+  DB  15,89,13,87,38,0,0                  ; mulps         0x2657(%rip),%xmm1        # 51f0 <_sk_callback_sse41+0x9db>
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  15,89,21,252,37,0,0                 ; mulps         0x25fc(%rip),%xmm2        # 51a0 <_sk_callback_sse41+0x9ed>
+  DB  15,89,21,92,38,0,0                  ; mulps         0x265c(%rip),%xmm2        # 5200 <_sk_callback_sse41+0x9eb>
   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  68,15,40,13,240,37,0,0              ; movaps        0x25f0(%rip),%xmm9        # 51b0 <_sk_callback_sse41+0x9fd>
+  DB  68,15,40,13,80,38,0,0               ; movaps        0x2650(%rip),%xmm9        # 5210 <_sk_callback_sse41+0x9fb>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  65,15,194,195,1                     ; cmpltps       %xmm11,%xmm0
-  DB  15,40,21,240,37,0,0                 ; movaps        0x25f0(%rip),%xmm2        # 51c0 <_sk_callback_sse41+0xa0d>
+  DB  15,40,21,80,38,0,0                  ; movaps        0x2650(%rip),%xmm2        # 5220 <_sk_callback_sse41+0xa0b>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
-  DB  68,15,40,37,245,37,0,0              ; movaps        0x25f5(%rip),%xmm12        # 51d0 <_sk_callback_sse41+0xa1d>
+  DB  68,15,40,37,85,38,0,0               ; movaps        0x2655(%rip),%xmm12        # 5230 <_sk_callback_sse41+0xa1b>
   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
@@ -14197,8 +14269,8 @@ _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  15,89,13,174,37,0,0                 ; mulps         0x25ae(%rip),%xmm1        # 51e0 <_sk_callback_sse41+0xa2d>
-  DB  15,89,21,183,37,0,0                 ; mulps         0x25b7(%rip),%xmm2        # 51f0 <_sk_callback_sse41+0xa3d>
+  DB  15,89,13,14,38,0,0                  ; mulps         0x260e(%rip),%xmm1        # 5240 <_sk_callback_sse41+0xa2b>
+  DB  15,89,21,23,38,0,0                  ; mulps         0x2617(%rip),%xmm2        # 5250 <_sk_callback_sse41+0xa3b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
@@ -14210,7 +14282,7 @@ _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  15,89,29,167,37,0,0                 ; mulps         0x25a7(%rip),%xmm3        # 5200 <_sk_callback_sse41+0xa4d>
+  DB  15,89,29,7,38,0,0                   ; mulps         0x2607(%rip),%xmm3        # 5260 <_sk_callback_sse41+0xa4b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
@@ -14241,7 +14313,7 @@ _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  15,89,29,59,37,0,0                  ; mulps         0x253b(%rip),%xmm3        # 5210 <_sk_callback_sse41+0xa5d>
+  DB  15,89,29,155,37,0,0                 ; mulps         0x259b(%rip),%xmm3        # 5270 <_sk_callback_sse41+0xa5b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -14252,7 +14324,7 @@ 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  68,15,40,5,47,37,0,0                ; movaps        0x252f(%rip),%xmm8        # 5220 <_sk_callback_sse41+0xa6d>
+  DB  68,15,40,5,143,37,0,0               ; movaps        0x258f(%rip),%xmm8        # 5280 <_sk_callback_sse41+0xa6b>
   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
@@ -14267,9 +14339,9 @@ _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  15,89,5,12,37,0,0                   ; mulps         0x250c(%rip),%xmm0        # 5230 <_sk_callback_sse41+0xa7d>
+  DB  15,89,5,108,37,0,0                  ; mulps         0x256c(%rip),%xmm0        # 5290 <_sk_callback_sse41+0xa7b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,19,37,0,0                  ; movaps        0x2513(%rip),%xmm3        # 5240 <_sk_callback_sse41+0xa8d>
+  DB  15,40,29,115,37,0,0                 ; movaps        0x2573(%rip),%xmm3        # 52a0 <_sk_callback_sse41+0xa8b>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -14298,9 +14370,9 @@ _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  15,89,5,172,36,0,0                  ; mulps         0x24ac(%rip),%xmm0        # 5250 <_sk_callback_sse41+0xa9d>
+  DB  15,89,5,12,37,0,0                   ; mulps         0x250c(%rip),%xmm0        # 52b0 <_sk_callback_sse41+0xa9b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,179,36,0,0                 ; movaps        0x24b3(%rip),%xmm3        # 5260 <_sk_callback_sse41+0xaad>
+  DB  15,40,29,19,37,0,0                  ; movaps        0x2513(%rip),%xmm3        # 52c0 <_sk_callback_sse41+0xaab>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -14343,17 +14415,17 @@ _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  102,15,111,5,10,36,0,0              ; movdqa        0x240a(%rip),%xmm0        # 5270 <_sk_callback_sse41+0xabd>
+  DB  102,15,111,5,106,36,0,0             ; movdqa        0x246a(%rip),%xmm0        # 52d0 <_sk_callback_sse41+0xabb>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,11,36,0,0                ; movaps        0x240b(%rip),%xmm8        # 5280 <_sk_callback_sse41+0xacd>
+  DB  68,15,40,5,107,36,0,0               ; movaps        0x246b(%rip),%xmm8        # 52e0 <_sk_callback_sse41+0xacb>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
-  DB  102,15,56,0,13,10,36,0,0            ; pshufb        0x240a(%rip),%xmm1        # 5290 <_sk_callback_sse41+0xadd>
+  DB  102,15,56,0,13,106,36,0,0           ; pshufb        0x246a(%rip),%xmm1        # 52f0 <_sk_callback_sse41+0xadb>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
-  DB  102,15,56,0,21,6,36,0,0             ; pshufb        0x2406(%rip),%xmm2        # 52a0 <_sk_callback_sse41+0xaed>
+  DB  102,15,56,0,21,102,36,0,0           ; pshufb        0x2466(%rip),%xmm2        # 5300 <_sk_callback_sse41+0xaeb>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -14367,19 +14439,19 @@ _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  102,15,111,5,236,35,0,0             ; movdqa        0x23ec(%rip),%xmm0        # 52b0 <_sk_callback_sse41+0xafd>
+  DB  102,15,111,5,76,36,0,0              ; movdqa        0x244c(%rip),%xmm0        # 5310 <_sk_callback_sse41+0xafb>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,238,35,0,0                  ; mulps         0x23ee(%rip),%xmm0        # 52c0 <_sk_callback_sse41+0xb0d>
-  DB  102,15,111,13,246,35,0,0            ; movdqa        0x23f6(%rip),%xmm1        # 52d0 <_sk_callback_sse41+0xb1d>
+  DB  15,89,5,78,36,0,0                   ; mulps         0x244e(%rip),%xmm0        # 5320 <_sk_callback_sse41+0xb0b>
+  DB  102,15,111,13,86,36,0,0             ; movdqa        0x2456(%rip),%xmm1        # 5330 <_sk_callback_sse41+0xb1b>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,248,35,0,0                 ; mulps         0x23f8(%rip),%xmm1        # 52e0 <_sk_callback_sse41+0xb2d>
-  DB  102,15,219,21,0,36,0,0              ; pand          0x2400(%rip),%xmm2        # 52f0 <_sk_callback_sse41+0xb3d>
+  DB  15,89,13,88,36,0,0                  ; mulps         0x2458(%rip),%xmm1        # 5340 <_sk_callback_sse41+0xb2b>
+  DB  102,15,219,21,96,36,0,0             ; pand          0x2460(%rip),%xmm2        # 5350 <_sk_callback_sse41+0xb3b>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,6,36,0,0                   ; mulps         0x2406(%rip),%xmm2        # 5300 <_sk_callback_sse41+0xb4d>
+  DB  15,89,21,102,36,0,0                 ; mulps         0x2466(%rip),%xmm2        # 5360 <_sk_callback_sse41+0xb4b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,13,36,0,0                  ; movaps        0x240d(%rip),%xmm3        # 5310 <_sk_callback_sse41+0xb5d>
+  DB  15,40,29,109,36,0,0                 ; movaps        0x246d(%rip),%xmm3        # 5370 <_sk_callback_sse41+0xb5b>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_gather_565_sse41
@@ -14405,31 +14477,31 @@ _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  102,15,111,5,178,35,0,0             ; movdqa        0x23b2(%rip),%xmm0        # 5320 <_sk_callback_sse41+0xb6d>
+  DB  102,15,111,5,18,36,0,0              ; movdqa        0x2412(%rip),%xmm0        # 5380 <_sk_callback_sse41+0xb6b>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,180,35,0,0                  ; mulps         0x23b4(%rip),%xmm0        # 5330 <_sk_callback_sse41+0xb7d>
-  DB  102,15,111,13,188,35,0,0            ; movdqa        0x23bc(%rip),%xmm1        # 5340 <_sk_callback_sse41+0xb8d>
+  DB  15,89,5,20,36,0,0                   ; mulps         0x2414(%rip),%xmm0        # 5390 <_sk_callback_sse41+0xb7b>
+  DB  102,15,111,13,28,36,0,0             ; movdqa        0x241c(%rip),%xmm1        # 53a0 <_sk_callback_sse41+0xb8b>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,190,35,0,0                 ; mulps         0x23be(%rip),%xmm1        # 5350 <_sk_callback_sse41+0xb9d>
-  DB  102,15,219,21,198,35,0,0            ; pand          0x23c6(%rip),%xmm2        # 5360 <_sk_callback_sse41+0xbad>
+  DB  15,89,13,30,36,0,0                  ; mulps         0x241e(%rip),%xmm1        # 53b0 <_sk_callback_sse41+0xb9b>
+  DB  102,15,219,21,38,36,0,0             ; pand          0x2426(%rip),%xmm2        # 53c0 <_sk_callback_sse41+0xbab>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,204,35,0,0                 ; mulps         0x23cc(%rip),%xmm2        # 5370 <_sk_callback_sse41+0xbbd>
+  DB  15,89,21,44,36,0,0                  ; mulps         0x242c(%rip),%xmm2        # 53d0 <_sk_callback_sse41+0xbbb>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,211,35,0,0                 ; movaps        0x23d3(%rip),%xmm3        # 5380 <_sk_callback_sse41+0xbcd>
+  DB  15,40,29,51,36,0,0                  ; movaps        0x2433(%rip),%xmm3        # 53e0 <_sk_callback_sse41+0xbcb>
   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  68,15,40,5,212,35,0,0               ; movaps        0x23d4(%rip),%xmm8        # 5390 <_sk_callback_sse41+0xbdd>
+  DB  68,15,40,5,52,36,0,0                ; movaps        0x2434(%rip),%xmm8        # 53f0 <_sk_callback_sse41+0xbdb>
   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  68,15,40,21,201,35,0,0              ; movaps        0x23c9(%rip),%xmm10        # 53a0 <_sk_callback_sse41+0xbed>
+  DB  68,15,40,21,41,36,0,0               ; movaps        0x2429(%rip),%xmm10        # 5400 <_sk_callback_sse41+0xbeb>
   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
@@ -14447,21 +14519,21 @@ _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  102,15,111,5,148,35,0,0             ; movdqa        0x2394(%rip),%xmm0        # 53b0 <_sk_callback_sse41+0xbfd>
+  DB  102,15,111,5,244,35,0,0             ; movdqa        0x23f4(%rip),%xmm0        # 5410 <_sk_callback_sse41+0xbfb>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,150,35,0,0                  ; mulps         0x2396(%rip),%xmm0        # 53c0 <_sk_callback_sse41+0xc0d>
-  DB  102,15,111,13,158,35,0,0            ; movdqa        0x239e(%rip),%xmm1        # 53d0 <_sk_callback_sse41+0xc1d>
+  DB  15,89,5,246,35,0,0                  ; mulps         0x23f6(%rip),%xmm0        # 5420 <_sk_callback_sse41+0xc0b>
+  DB  102,15,111,13,254,35,0,0            ; movdqa        0x23fe(%rip),%xmm1        # 5430 <_sk_callback_sse41+0xc1b>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,160,35,0,0                 ; mulps         0x23a0(%rip),%xmm1        # 53e0 <_sk_callback_sse41+0xc2d>
-  DB  102,15,111,21,168,35,0,0            ; movdqa        0x23a8(%rip),%xmm2        # 53f0 <_sk_callback_sse41+0xc3d>
+  DB  15,89,13,0,36,0,0                   ; mulps         0x2400(%rip),%xmm1        # 5440 <_sk_callback_sse41+0xc2b>
+  DB  102,15,111,21,8,36,0,0              ; movdqa        0x2408(%rip),%xmm2        # 5450 <_sk_callback_sse41+0xc3b>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,170,35,0,0                 ; mulps         0x23aa(%rip),%xmm2        # 5400 <_sk_callback_sse41+0xc4d>
-  DB  102,15,219,29,178,35,0,0            ; pand          0x23b2(%rip),%xmm3        # 5410 <_sk_callback_sse41+0xc5d>
+  DB  15,89,21,10,36,0,0                  ; mulps         0x240a(%rip),%xmm2        # 5460 <_sk_callback_sse41+0xc4b>
+  DB  102,15,219,29,18,36,0,0             ; pand          0x2412(%rip),%xmm3        # 5470 <_sk_callback_sse41+0xc5b>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,184,35,0,0                 ; mulps         0x23b8(%rip),%xmm3        # 5420 <_sk_callback_sse41+0xc6d>
+  DB  15,89,29,24,36,0,0                  ; mulps         0x2418(%rip),%xmm3        # 5480 <_sk_callback_sse41+0xc6b>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -14488,21 +14560,21 @@ _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  102,15,111,5,91,35,0,0              ; movdqa        0x235b(%rip),%xmm0        # 5430 <_sk_callback_sse41+0xc7d>
+  DB  102,15,111,5,187,35,0,0             ; movdqa        0x23bb(%rip),%xmm0        # 5490 <_sk_callback_sse41+0xc7b>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,93,35,0,0                   ; mulps         0x235d(%rip),%xmm0        # 5440 <_sk_callback_sse41+0xc8d>
-  DB  102,15,111,13,101,35,0,0            ; movdqa        0x2365(%rip),%xmm1        # 5450 <_sk_callback_sse41+0xc9d>
+  DB  15,89,5,189,35,0,0                  ; mulps         0x23bd(%rip),%xmm0        # 54a0 <_sk_callback_sse41+0xc8b>
+  DB  102,15,111,13,197,35,0,0            ; movdqa        0x23c5(%rip),%xmm1        # 54b0 <_sk_callback_sse41+0xc9b>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,103,35,0,0                 ; mulps         0x2367(%rip),%xmm1        # 5460 <_sk_callback_sse41+0xcad>
-  DB  102,15,111,21,111,35,0,0            ; movdqa        0x236f(%rip),%xmm2        # 5470 <_sk_callback_sse41+0xcbd>
+  DB  15,89,13,199,35,0,0                 ; mulps         0x23c7(%rip),%xmm1        # 54c0 <_sk_callback_sse41+0xcab>
+  DB  102,15,111,21,207,35,0,0            ; movdqa        0x23cf(%rip),%xmm2        # 54d0 <_sk_callback_sse41+0xcbb>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,113,35,0,0                 ; mulps         0x2371(%rip),%xmm2        # 5480 <_sk_callback_sse41+0xccd>
-  DB  102,15,219,29,121,35,0,0            ; pand          0x2379(%rip),%xmm3        # 5490 <_sk_callback_sse41+0xcdd>
+  DB  15,89,21,209,35,0,0                 ; mulps         0x23d1(%rip),%xmm2        # 54e0 <_sk_callback_sse41+0xccb>
+  DB  102,15,219,29,217,35,0,0            ; pand          0x23d9(%rip),%xmm3        # 54f0 <_sk_callback_sse41+0xcdb>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,127,35,0,0                 ; mulps         0x237f(%rip),%xmm3        # 54a0 <_sk_callback_sse41+0xced>
+  DB  15,89,29,223,35,0,0                 ; mulps         0x23df(%rip),%xmm3        # 5500 <_sk_callback_sse41+0xceb>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -14510,7 +14582,7 @@ 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  68,15,40,5,126,35,0,0               ; movaps        0x237e(%rip),%xmm8        # 54b0 <_sk_callback_sse41+0xcfd>
+  DB  68,15,40,5,222,35,0,0               ; movaps        0x23de(%rip),%xmm8        # 5510 <_sk_callback_sse41+0xcfb>
   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
@@ -14538,17 +14610,17 @@ _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  15,40,5,29,35,0,0                   ; movaps        0x231d(%rip),%xmm0        # 54c0 <_sk_callback_sse41+0xd0d>
+  DB  15,40,5,125,35,0,0                  ; movaps        0x237d(%rip),%xmm0        # 5520 <_sk_callback_sse41+0xd0b>
   DB  15,84,195                           ; andps         %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,31,35,0,0                ; movaps        0x231f(%rip),%xmm8        # 54d0 <_sk_callback_sse41+0xd1d>
+  DB  68,15,40,5,127,35,0,0               ; movaps        0x237f(%rip),%xmm8        # 5530 <_sk_callback_sse41+0xd1b>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,40,203                           ; movaps        %xmm3,%xmm1
-  DB  102,15,56,0,13,31,35,0,0            ; pshufb        0x231f(%rip),%xmm1        # 54e0 <_sk_callback_sse41+0xd2d>
+  DB  102,15,56,0,13,127,35,0,0           ; pshufb        0x237f(%rip),%xmm1        # 5540 <_sk_callback_sse41+0xd2b>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  15,40,211                           ; movaps        %xmm3,%xmm2
-  DB  102,15,56,0,21,28,35,0,0            ; pshufb        0x231c(%rip),%xmm2        # 54f0 <_sk_callback_sse41+0xd3d>
+  DB  102,15,56,0,21,124,35,0,0           ; pshufb        0x237c(%rip),%xmm2        # 5550 <_sk_callback_sse41+0xd3b>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -14577,17 +14649,17 @@ _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  102,15,111,5,181,34,0,0             ; movdqa        0x22b5(%rip),%xmm0        # 5500 <_sk_callback_sse41+0xd4d>
+  DB  102,15,111,5,21,35,0,0              ; movdqa        0x2315(%rip),%xmm0        # 5560 <_sk_callback_sse41+0xd4b>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,182,34,0,0               ; movaps        0x22b6(%rip),%xmm8        # 5510 <_sk_callback_sse41+0xd5d>
+  DB  68,15,40,5,22,35,0,0                ; movaps        0x2316(%rip),%xmm8        # 5570 <_sk_callback_sse41+0xd5b>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
-  DB  102,15,56,0,13,181,34,0,0           ; pshufb        0x22b5(%rip),%xmm1        # 5520 <_sk_callback_sse41+0xd6d>
+  DB  102,15,56,0,13,21,35,0,0            ; pshufb        0x2315(%rip),%xmm1        # 5580 <_sk_callback_sse41+0xd6b>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
-  DB  102,15,56,0,21,177,34,0,0           ; pshufb        0x22b1(%rip),%xmm2        # 5530 <_sk_callback_sse41+0xd7d>
+  DB  102,15,56,0,21,17,35,0,0            ; pshufb        0x2311(%rip),%xmm2        # 5590 <_sk_callback_sse41+0xd7b>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -14600,7 +14672,7 @@ 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  68,15,40,5,157,34,0,0               ; movaps        0x229d(%rip),%xmm8        # 5540 <_sk_callback_sse41+0xd8d>
+  DB  68,15,40,5,253,34,0,0               ; movaps        0x22fd(%rip),%xmm8        # 55a0 <_sk_callback_sse41+0xd8b>
   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
@@ -14635,18 +14707,18 @@ _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  102,68,15,111,5,22,34,0,0           ; movdqa        0x2216(%rip),%xmm8        # 5550 <_sk_callback_sse41+0xd9d>
+  DB  102,68,15,111,5,118,34,0,0          ; movdqa        0x2276(%rip),%xmm8        # 55b0 <_sk_callback_sse41+0xd9b>
   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  102,15,111,29,17,34,0,0             ; movdqa        0x2211(%rip),%xmm3        # 5560 <_sk_callback_sse41+0xdad>
+  DB  102,15,111,29,113,34,0,0            ; movdqa        0x2271(%rip),%xmm3        # 55c0 <_sk_callback_sse41+0xdab>
   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  102,68,15,111,21,253,33,0,0         ; movdqa        0x21fd(%rip),%xmm10        # 5570 <_sk_callback_sse41+0xdbd>
+  DB  102,68,15,111,21,93,34,0,0          ; movdqa        0x225d(%rip),%xmm10        # 55d0 <_sk_callback_sse41+0xdbb>
   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
@@ -14717,18 +14789,18 @@ _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  102,68,15,111,5,187,32,0,0          ; movdqa        0x20bb(%rip),%xmm8        # 5580 <_sk_callback_sse41+0xdcd>
+  DB  102,68,15,111,5,27,33,0,0           ; movdqa        0x211b(%rip),%xmm8        # 55e0 <_sk_callback_sse41+0xdcb>
   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  102,15,111,29,182,32,0,0            ; movdqa        0x20b6(%rip),%xmm3        # 5590 <_sk_callback_sse41+0xddd>
+  DB  102,15,111,29,22,33,0,0             ; movdqa        0x2116(%rip),%xmm3        # 55f0 <_sk_callback_sse41+0xddb>
   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  102,68,15,111,21,162,32,0,0         ; movdqa        0x20a2(%rip),%xmm10        # 55a0 <_sk_callback_sse41+0xded>
+  DB  102,68,15,111,21,2,33,0,0           ; movdqa        0x2102(%rip),%xmm10        # 5600 <_sk_callback_sse41+0xdeb>
   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
@@ -14774,17 +14846,17 @@ 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  102,68,15,111,21,216,31,0,0         ; movdqa        0x1fd8(%rip),%xmm10        # 55b0 <_sk_callback_sse41+0xdfd>
+  DB  102,68,15,111,21,56,32,0,0          ; movdqa        0x2038(%rip),%xmm10        # 5610 <_sk_callback_sse41+0xdfb>
   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  102,68,15,111,13,203,31,0,0         ; movdqa        0x1fcb(%rip),%xmm9        # 55c0 <_sk_callback_sse41+0xe0d>
+  DB  102,68,15,111,13,43,32,0,0          ; movdqa        0x202b(%rip),%xmm9        # 5620 <_sk_callback_sse41+0xe0b>
   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  102,68,15,111,29,188,31,0,0         ; movdqa        0x1fbc(%rip),%xmm11        # 55d0 <_sk_callback_sse41+0xe1d>
+  DB  102,68,15,111,29,28,32,0,0          ; movdqa        0x201c(%rip),%xmm11        # 5630 <_sk_callback_sse41+0xe1b>
   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
@@ -14852,7 +14924,7 @@ _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  68,15,40,5,139,30,0,0               ; movaps        0x1e8b(%rip),%xmm8        # 55e0 <_sk_callback_sse41+0xe2d>
+  DB  68,15,40,5,235,30,0,0               ; movaps        0x1eeb(%rip),%xmm8        # 5640 <_sk_callback_sse41+0xe2b>
   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
@@ -14902,7 +14974,7 @@ _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  68,15,40,5,204,29,0,0               ; movaps        0x1dcc(%rip),%xmm8        # 55f0 <_sk_callback_sse41+0xe3d>
+  DB  68,15,40,5,44,30,0,0                ; movaps        0x1e2c(%rip),%xmm8        # 5650 <_sk_callback_sse41+0xe3b>
   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
@@ -14919,14 +14991,14 @@ _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  15,40,29,147,29,0,0                 ; movaps        0x1d93(%rip),%xmm3        # 5600 <_sk_callback_sse41+0xe4d>
+  DB  15,40,29,243,29,0,0                 ; movaps        0x1df3(%rip),%xmm3        # 5660 <_sk_callback_sse41+0xe4b>
   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  68,15,40,13,148,29,0,0              ; movaps        0x1d94(%rip),%xmm9        # 5610 <_sk_callback_sse41+0xe5d>
+  DB  68,15,40,13,244,29,0,0              ; movaps        0x1df4(%rip),%xmm9        # 5670 <_sk_callback_sse41+0xe5b>
   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
@@ -15116,13 +15188,45 @@ _sk_mirror_y_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_clamp_x_1_sse41
+_sk_clamp_x_1_sse41 LABEL PROC
+  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
+  DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
+  DB  68,15,93,5,74,27,0,0                ; minps         0x1b4a(%rip),%xmm8        # 5680 <_sk_callback_sse41+0xe6b>
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_repeat_x_1_sse41
+_sk_repeat_x_1_sse41 LABEL PROC
+  DB  102,68,15,58,8,192,1                ; roundps       $0x1,%xmm0,%xmm8
+  DB  65,15,92,192                        ; subps         %xmm8,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_mirror_x_1_sse41
+_sk_mirror_x_1_sse41 LABEL PROC
+  DB  68,15,40,5,59,27,0,0                ; movaps        0x1b3b(%rip),%xmm8        # 5690 <_sk_callback_sse41+0xe7b>
+  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
+  DB  68,15,40,13,63,27,0,0               ; movaps        0x1b3f(%rip),%xmm9        # 56a0 <_sk_callback_sse41+0xe8b>
+  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
+  DB  102,69,15,58,8,201,1                ; roundps       $0x1,%xmm9,%xmm9
+  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
+  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
+  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
+  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
+  DB  68,15,92,192                        ; subps         %xmm0,%xmm8
+  DB  65,15,84,192                        ; andps         %xmm8,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_luminance_to_alpha_sse41
 _sk_luminance_to_alpha_sse41 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
-  DB  15,89,5,240,26,0,0                  ; mulps         0x1af0(%rip),%xmm0        # 5620 <_sk_callback_sse41+0xe6d>
-  DB  15,89,13,249,26,0,0                 ; mulps         0x1af9(%rip),%xmm1        # 5630 <_sk_callback_sse41+0xe7d>
+  DB  15,89,5,30,27,0,0                   ; mulps         0x1b1e(%rip),%xmm0        # 56b0 <_sk_callback_sse41+0xe9b>
+  DB  15,89,13,39,27,0,0                  ; mulps         0x1b27(%rip),%xmm1        # 56c0 <_sk_callback_sse41+0xeab>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,29,255,26,0,0                 ; mulps         0x1aff(%rip),%xmm3        # 5640 <_sk_callback_sse41+0xe8d>
+  DB  15,89,29,45,27,0,0                  ; mulps         0x1b2d(%rip),%xmm3        # 56d0 <_sk_callback_sse41+0xebb>
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -15386,9 +15490,9 @@ _sk_evenly_spaced_gradient_sse41 LABEL PROC
   DB  72,139,8                            ; mov           (%rax),%rcx
   DB  76,139,88,8                         ; mov           0x8(%rax),%r11
   DB  72,255,201                          ; dec           %rcx
-  DB  120,7                               ; js            3f7a <_sk_evenly_spaced_gradient_sse41+0x15>
+  DB  120,7                               ; js            3fdc <_sk_evenly_spaced_gradient_sse41+0x15>
   DB  243,72,15,42,201                    ; cvtsi2ss      %rcx,%xmm1
-  DB  235,21                              ; jmp           3f8f <_sk_evenly_spaced_gradient_sse41+0x2a>
+  DB  235,21                              ; jmp           3ff1 <_sk_evenly_spaced_gradient_sse41+0x2a>
   DB  73,137,200                          ; mov           %rcx,%r8
   DB  73,209,232                          ; shr           %r8
   DB  131,225,1                           ; and           $0x1,%ecx
@@ -15477,12 +15581,12 @@ _sk_gradient_sse41 LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  73,131,248,2                        ; cmp           $0x2,%r8
-  DB  114,50                              ; jb            4172 <_sk_gradient_sse41+0x41>
+  DB  114,50                              ; jb            41d4 <_sk_gradient_sse41+0x41>
   DB  72,139,72,72                        ; mov           0x48(%rax),%rcx
   DB  73,255,200                          ; dec           %r8
   DB  72,131,193,4                        ; add           $0x4,%rcx
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
-  DB  15,40,21,250,20,0,0                 ; movaps        0x14fa(%rip),%xmm2        # 5650 <_sk_callback_sse41+0xe9d>
+  DB  15,40,21,40,21,0,0                  ; movaps        0x1528(%rip),%xmm2        # 56e0 <_sk_callback_sse41+0xecb>
   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
@@ -15490,7 +15594,7 @@ _sk_gradient_sse41 LABEL PROC
   DB  102,15,254,203                      ; paddd         %xmm3,%xmm1
   DB  72,131,193,4                        ; add           $0x4,%rcx
   DB  73,255,200                          ; dec           %r8
-  DB  117,228                             ; jne           4156 <_sk_gradient_sse41+0x25>
+  DB  117,228                             ; jne           41b8 <_sk_gradient_sse41+0x25>
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  102,73,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%r9
@@ -15617,26 +15721,26 @@ _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  68,15,40,21,156,18,0,0              ; movaps        0x129c(%rip),%xmm10        # 5660 <_sk_callback_sse41+0xead>
+  DB  68,15,40,21,202,18,0,0              ; movaps        0x12ca(%rip),%xmm10        # 56f0 <_sk_callback_sse41+0xedb>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,160,18,0,0              ; addps         0x12a0(%rip),%xmm10        # 5670 <_sk_callback_sse41+0xebd>
+  DB  68,15,88,21,206,18,0,0              ; addps         0x12ce(%rip),%xmm10        # 5700 <_sk_callback_sse41+0xeeb>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,164,18,0,0              ; addps         0x12a4(%rip),%xmm10        # 5680 <_sk_callback_sse41+0xecd>
+  DB  68,15,88,21,210,18,0,0              ; addps         0x12d2(%rip),%xmm10        # 5710 <_sk_callback_sse41+0xefb>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,168,18,0,0              ; addps         0x12a8(%rip),%xmm10        # 5690 <_sk_callback_sse41+0xedd>
+  DB  68,15,88,21,214,18,0,0              ; addps         0x12d6(%rip),%xmm10        # 5720 <_sk_callback_sse41+0xf0b>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,194,195,1                     ; cmpltps       %xmm11,%xmm0
-  DB  68,15,40,29,167,18,0,0              ; movaps        0x12a7(%rip),%xmm11        # 56a0 <_sk_callback_sse41+0xeed>
+  DB  68,15,40,29,213,18,0,0              ; movaps        0x12d5(%rip),%xmm11        # 5730 <_sk_callback_sse41+0xf1b>
   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  68,15,40,29,160,18,0,0              ; movaps        0x12a0(%rip),%xmm11        # 56b0 <_sk_callback_sse41+0xefd>
+  DB  68,15,40,29,206,18,0,0              ; movaps        0x12ce(%rip),%xmm11        # 5740 <_sk_callback_sse41+0xf2b>
   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  68,15,40,13,146,18,0,0              ; movaps        0x1292(%rip),%xmm9        # 56c0 <_sk_callback_sse41+0xf0d>
+  DB  68,15,40,13,192,18,0,0              ; movaps        0x12c0(%rip),%xmm9        # 5750 <_sk_callback_sse41+0xf3b>
   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
@@ -15658,7 +15762,7 @@ _sk_xy_to_radius_sse41 LABEL PROC
 PUBLIC _sk_save_xy_sse41
 _sk_save_xy_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,102,18,0,0               ; movaps        0x1266(%rip),%xmm8        # 56d0 <_sk_callback_sse41+0xf1d>
+  DB  68,15,40,5,148,18,0,0               ; movaps        0x1294(%rip),%xmm8        # 5760 <_sk_callback_sse41+0xf4b>
   DB  15,17,0                             ; movups        %xmm0,(%rax)
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,88,200                        ; addps         %xmm8,%xmm9
@@ -15698,8 +15802,8 @@ _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        # 56e0 <_sk_callback_sse41+0xf2d>
-  DB  68,15,40,13,240,17,0,0              ; movaps        0x11f0(%rip),%xmm9        # 56f0 <_sk_callback_sse41+0xf3d>
+  DB  15,88,5,22,18,0,0                   ; addps         0x1216(%rip),%xmm0        # 5770 <_sk_callback_sse41+0xf5b>
+  DB  68,15,40,13,30,18,0,0               ; movaps        0x121e(%rip),%xmm9        # 5780 <_sk_callback_sse41+0xf6b>
   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
@@ -15710,7 +15814,7 @@ _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        # 5700 <_sk_callback_sse41+0xf4d>
+  DB  15,88,5,13,18,0,0                   ; addps         0x120d(%rip),%xmm0        # 5790 <_sk_callback_sse41+0xf7b>
   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
@@ -15720,8 +15824,8 @@ _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        # 5710 <_sk_callback_sse41+0xf5d>
-  DB  68,15,40,13,217,17,0,0              ; movaps        0x11d9(%rip),%xmm9        # 5720 <_sk_callback_sse41+0xf6d>
+  DB  15,88,13,255,17,0,0                 ; addps         0x11ff(%rip),%xmm1        # 57a0 <_sk_callback_sse41+0xf8b>
+  DB  68,15,40,13,7,18,0,0                ; movaps        0x1207(%rip),%xmm9        # 57b0 <_sk_callback_sse41+0xf9b>
   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
@@ -15732,7 +15836,7 @@ _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        # 5730 <_sk_callback_sse41+0xf7d>
+  DB  15,88,13,245,17,0,0                 ; addps         0x11f5(%rip),%xmm1        # 57c0 <_sk_callback_sse41+0xfab>
   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
@@ -15742,13 +15846,13 @@ _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        # 5740 <_sk_callback_sse41+0xf8d>
-  DB  68,15,40,13,194,17,0,0              ; movaps        0x11c2(%rip),%xmm9        # 5750 <_sk_callback_sse41+0xf9d>
+  DB  15,88,5,232,17,0,0                  ; addps         0x11e8(%rip),%xmm0        # 57d0 <_sk_callback_sse41+0xfbb>
+  DB  68,15,40,13,240,17,0,0              ; movaps        0x11f0(%rip),%xmm9        # 57e0 <_sk_callback_sse41+0xfcb>
   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        # 5760 <_sk_callback_sse41+0xfad>
-  DB  68,15,88,13,198,17,0,0              ; addps         0x11c6(%rip),%xmm9        # 5770 <_sk_callback_sse41+0xfbd>
+  DB  68,15,89,13,236,17,0,0              ; mulps         0x11ec(%rip),%xmm9        # 57f0 <_sk_callback_sse41+0xfdb>
+  DB  68,15,88,13,244,17,0,0              ; addps         0x11f4(%rip),%xmm9        # 5800 <_sk_callback_sse41+0xfeb>
   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
@@ -15759,16 +15863,16 @@ _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        # 5780 <_sk_callback_sse41+0xfcd>
-  DB  68,15,40,13,189,17,0,0              ; movaps        0x11bd(%rip),%xmm9        # 5790 <_sk_callback_sse41+0xfdd>
+  DB  15,88,5,227,17,0,0                  ; addps         0x11e3(%rip),%xmm0        # 5810 <_sk_callback_sse41+0xffb>
+  DB  68,15,40,13,235,17,0,0              ; movaps        0x11eb(%rip),%xmm9        # 5820 <_sk_callback_sse41+0x100b>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,193,17,0,0               ; movaps        0x11c1(%rip),%xmm8        # 57a0 <_sk_callback_sse41+0xfed>
+  DB  68,15,40,5,239,17,0,0               ; movaps        0x11ef(%rip),%xmm8        # 5830 <_sk_callback_sse41+0x101b>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,197,17,0,0               ; addps         0x11c5(%rip),%xmm8        # 57b0 <_sk_callback_sse41+0xffd>
+  DB  68,15,88,5,243,17,0,0               ; addps         0x11f3(%rip),%xmm8        # 5840 <_sk_callback_sse41+0x102b>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,201,17,0,0               ; addps         0x11c9(%rip),%xmm8        # 57c0 <_sk_callback_sse41+0x100d>
+  DB  68,15,88,5,247,17,0,0               ; addps         0x11f7(%rip),%xmm8        # 5850 <_sk_callback_sse41+0x103b>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,205,17,0,0               ; addps         0x11cd(%rip),%xmm8        # 57d0 <_sk_callback_sse41+0x101d>
+  DB  68,15,88,5,251,17,0,0               ; addps         0x11fb(%rip),%xmm8        # 5860 <_sk_callback_sse41+0x104b>
   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
@@ -15776,17 +15880,17 @@ _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        # 57e0 <_sk_callback_sse41+0x102d>
+  DB  68,15,40,5,245,17,0,0               ; movaps        0x11f5(%rip),%xmm8        # 5870 <_sk_callback_sse41+0x105b>
   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        # 57f0 <_sk_callback_sse41+0x103d>
+  DB  68,15,40,21,241,17,0,0              ; movaps        0x11f1(%rip),%xmm10        # 5880 <_sk_callback_sse41+0x106b>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,199,17,0,0              ; addps         0x11c7(%rip),%xmm10        # 5800 <_sk_callback_sse41+0x104d>
+  DB  68,15,88,21,245,17,0,0              ; addps         0x11f5(%rip),%xmm10        # 5890 <_sk_callback_sse41+0x107b>
   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        # 5810 <_sk_callback_sse41+0x105d>
+  DB  68,15,88,21,241,17,0,0              ; addps         0x11f1(%rip),%xmm10        # 58a0 <_sk_callback_sse41+0x108b>
   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
@@ -15796,11 +15900,11 @@ _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        # 5820 <_sk_callback_sse41+0x106d>
+  DB  15,88,5,228,17,0,0                  ; addps         0x11e4(%rip),%xmm0        # 58b0 <_sk_callback_sse41+0x109b>
   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        # 5830 <_sk_callback_sse41+0x107d>
-  DB  68,15,88,5,190,17,0,0               ; addps         0x11be(%rip),%xmm8        # 5840 <_sk_callback_sse41+0x108d>
+  DB  68,15,89,5,228,17,0,0               ; mulps         0x11e4(%rip),%xmm8        # 58c0 <_sk_callback_sse41+0x10ab>
+  DB  68,15,88,5,236,17,0,0               ; addps         0x11ec(%rip),%xmm8        # 58d0 <_sk_callback_sse41+0x10bb>
   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
@@ -15811,13 +15915,13 @@ _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        # 5850 <_sk_callback_sse41+0x109d>
-  DB  68,15,40,13,180,17,0,0              ; movaps        0x11b4(%rip),%xmm9        # 5860 <_sk_callback_sse41+0x10ad>
+  DB  15,88,13,218,17,0,0                 ; addps         0x11da(%rip),%xmm1        # 58e0 <_sk_callback_sse41+0x10cb>
+  DB  68,15,40,13,226,17,0,0              ; movaps        0x11e2(%rip),%xmm9        # 58f0 <_sk_callback_sse41+0x10db>
   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        # 5870 <_sk_callback_sse41+0x10bd>
-  DB  68,15,88,13,184,17,0,0              ; addps         0x11b8(%rip),%xmm9        # 5880 <_sk_callback_sse41+0x10cd>
+  DB  68,15,89,13,222,17,0,0              ; mulps         0x11de(%rip),%xmm9        # 5900 <_sk_callback_sse41+0x10eb>
+  DB  68,15,88,13,230,17,0,0              ; addps         0x11e6(%rip),%xmm9        # 5910 <_sk_callback_sse41+0x10fb>
   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
@@ -15828,16 +15932,16 @@ _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        # 5890 <_sk_callback_sse41+0x10dd>
-  DB  68,15,40,13,174,17,0,0              ; movaps        0x11ae(%rip),%xmm9        # 58a0 <_sk_callback_sse41+0x10ed>
+  DB  15,88,13,212,17,0,0                 ; addps         0x11d4(%rip),%xmm1        # 5920 <_sk_callback_sse41+0x110b>
+  DB  68,15,40,13,220,17,0,0              ; movaps        0x11dc(%rip),%xmm9        # 5930 <_sk_callback_sse41+0x111b>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,178,17,0,0               ; movaps        0x11b2(%rip),%xmm8        # 58b0 <_sk_callback_sse41+0x10fd>
+  DB  68,15,40,5,224,17,0,0               ; movaps        0x11e0(%rip),%xmm8        # 5940 <_sk_callback_sse41+0x112b>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,182,17,0,0               ; addps         0x11b6(%rip),%xmm8        # 58c0 <_sk_callback_sse41+0x110d>
+  DB  68,15,88,5,228,17,0,0               ; addps         0x11e4(%rip),%xmm8        # 5950 <_sk_callback_sse41+0x113b>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,186,17,0,0               ; addps         0x11ba(%rip),%xmm8        # 58d0 <_sk_callback_sse41+0x111d>
+  DB  68,15,88,5,232,17,0,0               ; addps         0x11e8(%rip),%xmm8        # 5960 <_sk_callback_sse41+0x114b>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,190,17,0,0               ; addps         0x11be(%rip),%xmm8        # 58e0 <_sk_callback_sse41+0x112d>
+  DB  68,15,88,5,236,17,0,0               ; addps         0x11ec(%rip),%xmm8        # 5970 <_sk_callback_sse41+0x115b>
   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
@@ -15845,17 +15949,17 @@ _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        # 58f0 <_sk_callback_sse41+0x113d>
+  DB  68,15,40,5,230,17,0,0               ; movaps        0x11e6(%rip),%xmm8        # 5980 <_sk_callback_sse41+0x116b>
   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        # 5900 <_sk_callback_sse41+0x114d>
+  DB  68,15,40,21,225,17,0,0              ; movaps        0x11e1(%rip),%xmm10        # 5990 <_sk_callback_sse41+0x117b>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,183,17,0,0              ; addps         0x11b7(%rip),%xmm10        # 5910 <_sk_callback_sse41+0x115d>
+  DB  68,15,88,21,229,17,0,0              ; addps         0x11e5(%rip),%xmm10        # 59a0 <_sk_callback_sse41+0x118b>
   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        # 5920 <_sk_callback_sse41+0x116d>
+  DB  68,15,88,21,225,17,0,0              ; addps         0x11e1(%rip),%xmm10        # 59b0 <_sk_callback_sse41+0x119b>
   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
@@ -15865,11 +15969,11 @@ _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        # 5930 <_sk_callback_sse41+0x117d>
+  DB  15,88,13,211,17,0,0                 ; addps         0x11d3(%rip),%xmm1        # 59c0 <_sk_callback_sse41+0x11ab>
   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        # 5940 <_sk_callback_sse41+0x118d>
-  DB  68,15,88,5,173,17,0,0               ; addps         0x11ad(%rip),%xmm8        # 5950 <_sk_callback_sse41+0x119d>
+  DB  68,15,89,5,211,17,0,0               ; mulps         0x11d3(%rip),%xmm8        # 59d0 <_sk_callback_sse41+0x11bb>
+  DB  68,15,88,5,219,17,0,0               ; addps         0x11db(%rip),%xmm8        # 59e0 <_sk_callback_sse41+0x11cb>
   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
@@ -16074,11 +16178,11 @@ ALIGN 16
   DB  128,191,0,0,128,191,0               ; cmpb          $0x0,-0x40800000(%rdi)
   DB  0,224                               ; add           %ah,%al
   DB  64,0,0                              ; add           %al,(%rax)
-  DB  224,64                              ; loopne        4a38 <.literal16+0x1d8>
+  DB  224,64                              ; loopne        4a98 <.literal16+0x1d8>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        4a3c <.literal16+0x1dc>
+  DB  224,64                              ; loopne        4a9c <.literal16+0x1dc>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        4a40 <.literal16+0x1e0>
+  DB  224,64                              ; loopne        4aa0 <.literal16+0x1e0>
   DB  154                                 ; (bad)
   DB  153                                 ; cltd
   DB  153                                 ; cltd
@@ -16098,13 +16202,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4a61 <.literal16+0x201>
+  DB  71,225,61                           ; rex.RXB       loope 4ac1 <.literal16+0x201>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4a65 <.literal16+0x205>
+  DB  71,225,61                           ; rex.RXB       loope 4ac5 <.literal16+0x205>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4a69 <.literal16+0x209>
+  DB  71,225,61                           ; rex.RXB       loope 4ac9 <.literal16+0x209>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4a6d <.literal16+0x20d>
+  DB  71,225,61                           ; rex.RXB       loope 4acd <.literal16+0x20d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -16129,13 +16233,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4aa1 <.literal16+0x241>
+  DB  71,225,61                           ; rex.RXB       loope 4b01 <.literal16+0x241>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4aa5 <.literal16+0x245>
+  DB  71,225,61                           ; rex.RXB       loope 4b05 <.literal16+0x245>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4aa9 <.literal16+0x249>
+  DB  71,225,61                           ; rex.RXB       loope 4b09 <.literal16+0x249>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4aad <.literal16+0x24d>
+  DB  71,225,61                           ; rex.RXB       loope 4b0d <.literal16+0x24d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -16160,13 +16264,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4ae1 <.literal16+0x281>
+  DB  71,225,61                           ; rex.RXB       loope 4b41 <.literal16+0x281>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4ae5 <.literal16+0x285>
+  DB  71,225,61                           ; rex.RXB       loope 4b45 <.literal16+0x285>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4ae9 <.literal16+0x289>
+  DB  71,225,61                           ; rex.RXB       loope 4b49 <.literal16+0x289>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4aed <.literal16+0x28d>
+  DB  71,225,61                           ; rex.RXB       loope 4b4d <.literal16+0x28d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -16191,13 +16295,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4b21 <.literal16+0x2c1>
+  DB  71,225,61                           ; rex.RXB       loope 4b81 <.literal16+0x2c1>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4b25 <.literal16+0x2c5>
+  DB  71,225,61                           ; rex.RXB       loope 4b85 <.literal16+0x2c5>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4b29 <.literal16+0x2c9>
+  DB  71,225,61                           ; rex.RXB       loope 4b89 <.literal16+0x2c9>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4b2d <.literal16+0x2cd>
+  DB  71,225,61                           ; rex.RXB       loope 4b8d <.literal16+0x2cd>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -16413,13 +16517,13 @@ ALIGN 16
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        4cf9 <.literal16+0x499>
+  DB  224,7                               ; loopne        4d59 <.literal16+0x499>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        4cfd <.literal16+0x49d>
+  DB  224,7                               ; loopne        4d5d <.literal16+0x49d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        4d01 <.literal16+0x4a1>
+  DB  224,7                               ; loopne        4d61 <.literal16+0x4a1>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        4d05 <.literal16+0x4a5>
+  DB  224,7                               ; loopne        4d65 <.literal16+0x4a5>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -16453,10 +16557,10 @@ 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)        # a004d48 <_sk_callback_sse41+0xa000595>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a004da8 <_sk_callback_sse41+0xa000593>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3004d50 <_sk_callback_sse41+0x300059d>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3004db0 <_sk_callback_sse41+0x300059b>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -16511,11 +16615,11 @@ ALIGN 16
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,127,67                            ; add           %bh,0x43(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            4e1b <.literal16+0x5bb>
+  DB  127,67                              ; jg            4e7b <.literal16+0x5bb>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            4e1f <.literal16+0x5bf>
+  DB  127,67                              ; jg            4e7f <.literal16+0x5bf>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            4e23 <.literal16+0x5c3>
+  DB  127,67                              ; jg            4e83 <.literal16+0x5c3>
   DB  129,128,128,59,129,128,128,59,129,128; addl          $0x80813b80,-0x7f7ec480(%rax)
   DB  128,59,129                          ; cmpb          $0x81,(%rbx)
   DB  128,128,59,129,128,128,59           ; addb          $0x3b,-0x7f7f7ec5(%rax)
@@ -16530,16 +16634,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4e14 <.literal16+0x5b4>
+  DB  127,0                               ; jg            4e74 <.literal16+0x5b4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4e18 <.literal16+0x5b8>
+  DB  127,0                               ; jg            4e78 <.literal16+0x5b8>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4e1c <.literal16+0x5bc>
+  DB  127,0                               ; jg            4e7c <.literal16+0x5bc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4e20 <.literal16+0x5c0>
+  DB  127,0                               ; jg            4e80 <.literal16+0x5c0>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -16548,7 +16652,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4ea5 <.literal16+0x645>
+  DB  119,115                             ; ja            4f05 <.literal16+0x645>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -16559,7 +16663,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           4e09 <.literal16+0x5a9>
+  DB  117,191                             ; jne           4e69 <.literal16+0x5a9>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -16571,7 +16675,7 @@ 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          ffffffffe9a38e4a <_sk_callback_sse41+0xffffffffe9a34697>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38eaa <_sk_callback_sse41+0xffffffffe9a34695>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16626,16 +16730,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4ee4 <.literal16+0x684>
+  DB  127,0                               ; jg            4f44 <.literal16+0x684>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4ee8 <.literal16+0x688>
+  DB  127,0                               ; jg            4f48 <.literal16+0x688>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4eec <.literal16+0x68c>
+  DB  127,0                               ; jg            4f4c <.literal16+0x68c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4ef0 <.literal16+0x690>
+  DB  127,0                               ; jg            4f50 <.literal16+0x690>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -16644,7 +16748,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            4f75 <.literal16+0x715>
+  DB  119,115                             ; ja            4fd5 <.literal16+0x715>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -16655,7 +16759,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           4ed9 <.literal16+0x679>
+  DB  117,191                             ; jne           4f39 <.literal16+0x679>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -16667,7 +16771,7 @@ 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          ffffffffe9a38f1a <_sk_callback_sse41+0xffffffffe9a34767>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a38f7a <_sk_callback_sse41+0xffffffffe9a34765>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16722,16 +16826,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4fb4 <.literal16+0x754>
+  DB  127,0                               ; jg            5014 <.literal16+0x754>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4fb8 <.literal16+0x758>
+  DB  127,0                               ; jg            5018 <.literal16+0x758>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4fbc <.literal16+0x75c>
+  DB  127,0                               ; jg            501c <.literal16+0x75c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            4fc0 <.literal16+0x760>
+  DB  127,0                               ; jg            5020 <.literal16+0x760>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -16740,7 +16844,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            5045 <.literal16+0x7e5>
+  DB  119,115                             ; ja            50a5 <.literal16+0x7e5>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -16751,7 +16855,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           4fa9 <.literal16+0x749>
+  DB  117,191                             ; jne           5009 <.literal16+0x749>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -16763,7 +16867,7 @@ 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          ffffffffe9a38fea <_sk_callback_sse41+0xffffffffe9a34837>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3904a <_sk_callback_sse41+0xffffffffe9a34835>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16818,16 +16922,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5084 <.literal16+0x824>
+  DB  127,0                               ; jg            50e4 <.literal16+0x824>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5088 <.literal16+0x828>
+  DB  127,0                               ; jg            50e8 <.literal16+0x828>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            508c <.literal16+0x82c>
+  DB  127,0                               ; jg            50ec <.literal16+0x82c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5090 <.literal16+0x830>
+  DB  127,0                               ; jg            50f0 <.literal16+0x830>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -16836,7 +16940,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            5115 <.literal16+0x8b5>
+  DB  119,115                             ; ja            5175 <.literal16+0x8b5>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -16847,7 +16951,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           5079 <.literal16+0x819>
+  DB  117,191                             ; jne           50d9 <.literal16+0x819>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -16859,7 +16963,7 @@ 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          ffffffffe9a390ba <_sk_callback_sse41+0xffffffffe9a34907>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3911a <_sk_callback_sse41+0xffffffffe9a34905>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -16910,13 +17014,13 @@ ALIGN 16
   DB  200,66,0,0                          ; enterq        $0x42,$0x0
   DB  200,66,0,0                          ; enterq        $0x42,$0x0
   DB  200,66,0,0                          ; enterq        $0x42,$0x0
-  DB  127,67                              ; jg            5197 <.literal16+0x937>
+  DB  127,67                              ; jg            51f7 <.literal16+0x937>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            519b <.literal16+0x93b>
+  DB  127,67                              ; jg            51fb <.literal16+0x93b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            519f <.literal16+0x93f>
+  DB  127,67                              ; jg            51ff <.literal16+0x93f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            51a3 <.literal16+0x943>
+  DB  127,67                              ; jg            5203 <.literal16+0x943>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,195                               ; add           %al,%bl
   DB  0,0                                 ; add           %al,(%rax)
@@ -16963,16 +17067,16 @@ ALIGN 16
   DB  128,3,62                            ; addb          $0x3e,(%rbx)
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           5223 <.literal16+0x9c3>
+  DB  118,63                              ; jbe           5283 <.literal16+0x9c3>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           5227 <.literal16+0x9c7>
+  DB  118,63                              ; jbe           5287 <.literal16+0x9c7>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           522b <.literal16+0x9cb>
+  DB  118,63                              ; jbe           528b <.literal16+0x9cb>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           522f <.literal16+0x9cf>
+  DB  118,63                              ; jbe           528f <.literal16+0x9cf>
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
@@ -16984,11 +17088,11 @@ ALIGN 16
   DB  128,59,0                            ; cmpb          $0x0,(%rbx)
   DB  0,127,67                            ; add           %bh,0x43(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            526b <.literal16+0xa0b>
+  DB  127,67                              ; jg            52cb <.literal16+0xa0b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            526f <.literal16+0xa0f>
+  DB  127,67                              ; jg            52cf <.literal16+0xa0f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            5273 <.literal16+0xa13>
+  DB  127,67                              ; jg            52d3 <.literal16+0xa13>
   DB  129,128,128,59,129,128,128,59,129,128; addl          $0x80813b80,-0x7f7ec480(%rax)
   DB  128,59,129                          ; cmpb          $0x81,(%rbx)
   DB  128,128,59,0,0,128,63               ; addb          $0x3f,-0x7fffffc5(%rax)
@@ -17017,7 +17121,7 @@ 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)        # 30052a0 <_sk_callback_sse41+0x3000aed>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005300 <_sk_callback_sse41+0x3000aeb>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -17046,13 +17150,13 @@ ALIGN 16
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        52d9 <.literal16+0xa79>
+  DB  224,7                               ; loopne        5339 <.literal16+0xa79>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        52dd <.literal16+0xa7d>
+  DB  224,7                               ; loopne        533d <.literal16+0xa7d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        52e1 <.literal16+0xa81>
+  DB  224,7                               ; loopne        5341 <.literal16+0xa81>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        52e5 <.literal16+0xa85>
+  DB  224,7                               ; loopne        5345 <.literal16+0xa85>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -17098,13 +17202,13 @@ ALIGN 16
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        5349 <.literal16+0xae9>
+  DB  224,7                               ; loopne        53a9 <.literal16+0xae9>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        534d <.literal16+0xaed>
+  DB  224,7                               ; loopne        53ad <.literal16+0xaed>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5351 <.literal16+0xaf1>
+  DB  224,7                               ; loopne        53b1 <.literal16+0xaf1>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5355 <.literal16+0xaf5>
+  DB  224,7                               ; loopne        53b5 <.literal16+0xaf5>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -17142,13 +17246,13 @@ ALIGN 16
   DB  65,0,0                              ; add           %al,(%r8)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            53e6 <.literal16+0xb86>
+  DB  124,66                              ; jl            5446 <.literal16+0xb86>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            53ea <.literal16+0xb8a>
+  DB  124,66                              ; jl            544a <.literal16+0xb8a>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            53ee <.literal16+0xb8e>
+  DB  124,66                              ; jl            544e <.literal16+0xb8e>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            53f2 <.literal16+0xb92>
+  DB  124,66                              ; jl            5452 <.literal16+0xb92>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,240                               ; add           %dh,%al
@@ -17238,13 +17342,13 @@ ALIGN 16
   DB  136,136,61,137,136,136              ; mov           %cl,-0x777776c3(%rax)
   DB  61,137,136,136,61                   ; cmp           $0x3d888889,%eax
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            54f5 <.literal16+0xc95>
+  DB  112,65                              ; jo            5555 <.literal16+0xc95>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            54f9 <.literal16+0xc99>
+  DB  112,65                              ; jo            5559 <.literal16+0xc99>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            54fd <.literal16+0xc9d>
+  DB  112,65                              ; jo            555d <.literal16+0xc9d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            5501 <.literal16+0xca1>
+  DB  112,65                              ; jo            5561 <.literal16+0xca1>
   DB  255,0                               ; incl          (%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  255,0                               ; incl          (%rax)
@@ -17259,7 +17363,7 @@ 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)        # 30054f0 <_sk_callback_sse41+0x3000d3d>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005550 <_sk_callback_sse41+0x3000d3b>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -17286,7 +17390,7 @@ 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)        # 3005530 <_sk_callback_sse41+0x3000d7d>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 3005590 <_sk_callback_sse41+0x3000d7b>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -17301,11 +17405,11 @@ ALIGN 16
   DB  255,0                               ; incl          (%rax)
   DB  0,127,67                            ; add           %bh,0x43(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            558b <.literal16+0xd2b>
+  DB  127,67                              ; jg            55eb <.literal16+0xd2b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            558f <.literal16+0xd2f>
+  DB  127,67                              ; jg            55ef <.literal16+0xd2f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            5593 <.literal16+0xd33>
+  DB  127,67                              ; jg            55f3 <.literal16+0xd33>
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%rax)
@@ -17381,13 +17485,29 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  255                                 ; (bad)
-  DB  127,71                              ; jg            565b <.literal16+0xdfb>
+  DB  127,71                              ; jg            56bb <.literal16+0xdfb>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            565f <.literal16+0xdff>
+  DB  127,71                              ; jg            56bf <.literal16+0xdff>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            5663 <.literal16+0xe03>
+  DB  127,71                              ; jg            56c3 <.literal16+0xe03>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            5667 <.literal16+0xe07>
+  DB  127,71                              ; jg            56c7 <.literal16+0xe07>
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
+  DB  63                                  ; (bad)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,191,0,0,128                   ; add           %al,-0x7fffff41(%rax)
+  DB  191,0,0,128,191                     ; mov           $0xbf800000,%edi
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,191,0,0,0,63,0                  ; cmpb          $0x0,0x3f000000(%rdi)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  63                                  ; (bad)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,63                                ; add           %bh,(%rdi)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,63                                ; add           %bh,(%rdi)
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,208                              ; ds            (bad)
@@ -17521,11 +17641,11 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,114                          ; cmpb          $0x72,(%rdi)
   DB  28,199                              ; sbb           $0xc7,%al
-  DB  62,114,28                           ; jb,pt         5782 <.literal16+0xf22>
+  DB  62,114,28                           ; jb,pt         5812 <.literal16+0xf52>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5786 <.literal16+0xf26>
+  DB  62,114,28                           ; jb,pt         5816 <.literal16+0xf56>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         578a <.literal16+0xf2a>
+  DB  62,114,28                           ; jb,pt         581a <.literal16+0xf5a>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -17569,7 +17689,7 @@ 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        # 3d63e615 <_sk_callback_sse41+0x3d639e62>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e6a5 <_sk_callback_sse41+0x3d639e90>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -17595,7 +17715,7 @@ 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        # 3d63e655 <_sk_callback_sse41+0x3d639ea2>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e6e5 <_sk_callback_sse41+0x3d639ed0>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -17604,13 +17724,13 @@ ALIGN 16
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
-  DB  114,28                              ; jb            584e <.literal16+0xfee>
+  DB  114,28                              ; jb            58de <.literal16+0x101e>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5852 <.literal16+0xff2>
+  DB  62,114,28                           ; jb,pt         58e2 <.literal16+0x1022>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5856 <.literal16+0xff6>
+  DB  62,114,28                           ; jb,pt         58e6 <.literal16+0x1026>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         585a <.literal16+0xffa>
+  DB  62,114,28                           ; jb,pt         58ea <.literal16+0x102a>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -17631,11 +17751,11 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,114                          ; cmpb          $0x72,(%rdi)
   DB  28,199                              ; sbb           $0xc7,%al
-  DB  62,114,28                           ; jb,pt         5892 <.literal16+0x1032>
+  DB  62,114,28                           ; jb,pt         5922 <.literal16+0x1062>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5896 <.literal16+0x1036>
+  DB  62,114,28                           ; jb,pt         5926 <.literal16+0x1066>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         589a <.literal16+0x103a>
+  DB  62,114,28                           ; jb,pt         592a <.literal16+0x106a>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -17679,7 +17799,7 @@ 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        # 3d63e725 <_sk_callback_sse41+0x3d639f72>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e7b5 <_sk_callback_sse41+0x3d639fa0>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -17705,7 +17825,7 @@ 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        # 3d63e765 <_sk_callback_sse41+0x3d639fb2>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63e7f5 <_sk_callback_sse41+0x3d639fe0>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -17714,13 +17834,13 @@ ALIGN 16
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
-  DB  114,28                              ; jb            595e <.literal16+0x10fe>
+  DB  114,28                              ; jb            59ee <.literal16+0x112e>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5962 <_sk_callback_sse41+0x11af>
+  DB  62,114,28                           ; jb,pt         59f2 <_sk_callback_sse41+0x11dd>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5966 <_sk_callback_sse41+0x11b3>
+  DB  62,114,28                           ; jb,pt         59f6 <_sk_callback_sse41+0x11e1>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         596a <_sk_callback_sse41+0x11b7>
+  DB  62,114,28                           ; jb,pt         59fa <_sk_callback_sse41+0x11e5>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -17811,7 +17931,7 @@ _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  15,40,21,193,75,0,0                 ; movaps        0x4bc1(%rip),%xmm2        # 4cd0 <_sk_callback_sse2+0xb6>
+  DB  15,40,21,81,76,0,0                  ; movaps        0x4c51(%rip),%xmm2        # 4d60 <_sk_callback_sse2+0xba>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  15,16,2                             ; movups        (%rdx),%xmm0
   DB  15,88,193                           ; addps         %xmm1,%xmm0
@@ -17820,7 +17940,7 @@ _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  15,40,21,176,75,0,0                 ; movaps        0x4bb0(%rip),%xmm2        # 4ce0 <_sk_callback_sse2+0xc6>
+  DB  15,40,21,64,76,0,0                  ; movaps        0x4c40(%rip),%xmm2        # 4d70 <_sk_callback_sse2+0xca>
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
@@ -17839,14 +17959,14 @@ _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  102,68,15,111,21,123,75,0,0         ; movdqa        0x4b7b(%rip),%xmm10        # 4cf0 <_sk_callback_sse2+0xd6>
+  DB  102,68,15,111,21,11,76,0,0          ; movdqa        0x4c0b(%rip),%xmm10        # 4d80 <_sk_callback_sse2+0xda>
   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  102,68,15,111,37,103,75,0,0         ; movdqa        0x4b67(%rip),%xmm12        # 4d00 <_sk_callback_sse2+0xe6>
-  DB  102,68,15,111,45,110,75,0,0         ; movdqa        0x4b6e(%rip),%xmm13        # 4d10 <_sk_callback_sse2+0xf6>
+  DB  102,68,15,111,37,247,75,0,0         ; movdqa        0x4bf7(%rip),%xmm12        # 4d90 <_sk_callback_sse2+0xea>
+  DB  102,68,15,111,45,254,75,0,0         ; movdqa        0x4bfe(%rip),%xmm13        # 4da0 <_sk_callback_sse2+0xfa>
   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
@@ -17862,8 +17982,8 @@ _sk_dither_sse2 LABEL PROC
   DB  102,69,15,235,198                   ; por           %xmm14,%xmm8
   DB  102,69,15,235,197                   ; por           %xmm13,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,41,75,0,0                ; mulps         0x4b29(%rip),%xmm8        # 4d20 <_sk_callback_sse2+0x106>
-  DB  68,15,88,5,49,75,0,0                ; addps         0x4b31(%rip),%xmm8        # 4d30 <_sk_callback_sse2+0x116>
+  DB  68,15,89,5,185,75,0,0               ; mulps         0x4bb9(%rip),%xmm8        # 4db0 <_sk_callback_sse2+0x10a>
+  DB  68,15,88,5,193,75,0,0               ; addps         0x4bc1(%rip),%xmm8        # 4dc0 <_sk_callback_sse2+0x11a>
   DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
@@ -17930,7 +18050,7 @@ _sk_clear_sse2 LABEL PROC
 PUBLIC _sk_srcatop_sse2
 _sk_srcatop_sse2 LABEL PROC
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  68,15,40,5,138,74,0,0               ; movaps        0x4a8a(%rip),%xmm8        # 4d40 <_sk_callback_sse2+0x126>
+  DB  68,15,40,5,26,75,0,0                ; movaps        0x4b1a(%rip),%xmm8        # 4dd0 <_sk_callback_sse2+0x12a>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -17953,7 +18073,7 @@ 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  68,15,40,13,77,74,0,0               ; movaps        0x4a4d(%rip),%xmm9        # 4d50 <_sk_callback_sse2+0x136>
+  DB  68,15,40,13,221,74,0,0              ; movaps        0x4add(%rip),%xmm9        # 4de0 <_sk_callback_sse2+0x13a>
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
@@ -17994,7 +18114,7 @@ _sk_dstin_sse2 LABEL PROC
 
 PUBLIC _sk_srcout_sse2
 _sk_srcout_sse2 LABEL PROC
-  DB  68,15,40,5,241,73,0,0               ; movaps        0x49f1(%rip),%xmm8        # 4d60 <_sk_callback_sse2+0x146>
+  DB  68,15,40,5,129,74,0,0               ; movaps        0x4a81(%rip),%xmm8        # 4df0 <_sk_callback_sse2+0x14a>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
@@ -18005,7 +18125,7 @@ _sk_srcout_sse2 LABEL PROC
 
 PUBLIC _sk_dstout_sse2
 _sk_dstout_sse2 LABEL PROC
-  DB  68,15,40,5,225,73,0,0               ; movaps        0x49e1(%rip),%xmm8        # 4d70 <_sk_callback_sse2+0x156>
+  DB  68,15,40,5,113,74,0,0               ; movaps        0x4a71(%rip),%xmm8        # 4e00 <_sk_callback_sse2+0x15a>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -18020,7 +18140,7 @@ _sk_dstout_sse2 LABEL PROC
 
 PUBLIC _sk_srcover_sse2
 _sk_srcover_sse2 LABEL PROC
-  DB  68,15,40,5,196,73,0,0               ; movaps        0x49c4(%rip),%xmm8        # 4d80 <_sk_callback_sse2+0x166>
+  DB  68,15,40,5,84,74,0,0                ; movaps        0x4a54(%rip),%xmm8        # 4e10 <_sk_callback_sse2+0x16a>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -18038,7 +18158,7 @@ _sk_srcover_sse2 LABEL PROC
 
 PUBLIC _sk_dstover_sse2
 _sk_dstover_sse2 LABEL PROC
-  DB  68,15,40,5,152,73,0,0               ; movaps        0x4998(%rip),%xmm8        # 4d90 <_sk_callback_sse2+0x176>
+  DB  68,15,40,5,40,74,0,0                ; movaps        0x4a28(%rip),%xmm8        # 4e20 <_sk_callback_sse2+0x17a>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -18062,7 +18182,7 @@ _sk_modulate_sse2 LABEL PROC
 
 PUBLIC _sk_multiply_sse2
 _sk_multiply_sse2 LABEL PROC
-  DB  68,15,40,5,108,73,0,0               ; movaps        0x496c(%rip),%xmm8        # 4da0 <_sk_callback_sse2+0x186>
+  DB  68,15,40,5,252,73,0,0               ; movaps        0x49fc(%rip),%xmm8        # 4e30 <_sk_callback_sse2+0x18a>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
@@ -18132,7 +18252,7 @@ _sk_screen_sse2 LABEL PROC
 PUBLIC _sk_xor__sse2
 _sk_xor__sse2 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
-  DB  15,40,29,157,72,0,0                 ; movaps        0x489d(%rip),%xmm3        # 4db0 <_sk_callback_sse2+0x196>
+  DB  15,40,29,45,73,0,0                  ; movaps        0x492d(%rip),%xmm3        # 4e40 <_sk_callback_sse2+0x19a>
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
@@ -18178,7 +18298,7 @@ _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  15,40,21,8,72,0,0                   ; movaps        0x4808(%rip),%xmm2        # 4dc0 <_sk_callback_sse2+0x1a6>
+  DB  15,40,21,152,72,0,0                 ; movaps        0x4898(%rip),%xmm2        # 4e50 <_sk_callback_sse2+0x1aa>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -18210,7 +18330,7 @@ _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  15,40,21,173,71,0,0                 ; movaps        0x47ad(%rip),%xmm2        # 4dd0 <_sk_callback_sse2+0x1b6>
+  DB  15,40,21,61,72,0,0                  ; movaps        0x483d(%rip),%xmm2        # 4e60 <_sk_callback_sse2+0x1ba>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -18245,7 +18365,7 @@ _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  15,40,21,71,71,0,0                  ; movaps        0x4747(%rip),%xmm2        # 4de0 <_sk_callback_sse2+0x1c6>
+  DB  15,40,21,215,71,0,0                 ; movaps        0x47d7(%rip),%xmm2        # 4e70 <_sk_callback_sse2+0x1ca>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -18270,7 +18390,7 @@ _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  15,40,13,8,71,0,0                   ; movaps        0x4708(%rip),%xmm1        # 4df0 <_sk_callback_sse2+0x1d6>
+  DB  15,40,13,152,71,0,0                 ; movaps        0x4798(%rip),%xmm1        # 4e80 <_sk_callback_sse2+0x1da>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,207                           ; mulps         %xmm7,%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -18282,7 +18402,7 @@ _sk_exclusion_sse2 LABEL PROC
 PUBLIC _sk_colorburn_sse2
 _sk_colorburn_sse2 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,247,70,0,0              ; movaps        0x46f7(%rip),%xmm10        # 4e00 <_sk_callback_sse2+0x1e6>
+  DB  68,15,40,21,135,71,0,0              ; movaps        0x4787(%rip),%xmm10        # 4e90 <_sk_callback_sse2+0x1ea>
   DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
@@ -18374,7 +18494,7 @@ _sk_colorburn_sse2 LABEL PROC
 PUBLIC _sk_colordodge_sse2
 _sk_colordodge_sse2 LABEL PROC
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
-  DB  68,15,40,21,173,69,0,0              ; movaps        0x45ad(%rip),%xmm10        # 4e10 <_sk_callback_sse2+0x1f6>
+  DB  68,15,40,21,61,70,0,0               ; movaps        0x463d(%rip),%xmm10        # 4ea0 <_sk_callback_sse2+0x1fa>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -18467,7 +18587,7 @@ _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  68,15,40,29,95,68,0,0               ; movaps        0x445f(%rip),%xmm11        # 4e20 <_sk_callback_sse2+0x206>
+  DB  68,15,40,29,239,68,0,0              ; movaps        0x44ef(%rip),%xmm11        # 4eb0 <_sk_callback_sse2+0x20a>
   DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
@@ -18554,7 +18674,7 @@ 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  68,15,40,13,42,67,0,0               ; movaps        0x432a(%rip),%xmm9        # 4e30 <_sk_callback_sse2+0x216>
+  DB  68,15,40,13,186,67,0,0              ; movaps        0x43ba(%rip),%xmm9        # 4ec0 <_sk_callback_sse2+0x21a>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -18644,7 +18764,7 @@ _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  68,15,40,13,228,65,0,0              ; movaps        0x41e4(%rip),%xmm9        # 4e40 <_sk_callback_sse2+0x226>
+  DB  68,15,40,13,116,66,0,0              ; movaps        0x4274(%rip),%xmm9        # 4ed0 <_sk_callback_sse2+0x22a>
   DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
   DB  69,15,92,250                        ; subps         %xmm10,%xmm15
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -18657,10 +18777,10 @@ _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  68,15,40,53,190,65,0,0              ; movaps        0x41be(%rip),%xmm14        # 4e50 <_sk_callback_sse2+0x236>
+  DB  68,15,40,53,78,66,0,0               ; movaps        0x424e(%rip),%xmm14        # 4ee0 <_sk_callback_sse2+0x23a>
   DB  69,15,88,222                        ; addps         %xmm14,%xmm11
   DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
-  DB  68,15,40,21,190,65,0,0              ; movaps        0x41be(%rip),%xmm10        # 4e60 <_sk_callback_sse2+0x246>
+  DB  68,15,40,21,78,66,0,0               ; movaps        0x424e(%rip),%xmm10        # 4ef0 <_sk_callback_sse2+0x24a>
   DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
   DB  69,15,88,235                        ; addps         %xmm11,%xmm13
   DB  15,88,228                           ; addps         %xmm4,%xmm4
@@ -18805,7 +18925,7 @@ _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  68,15,40,5,250,63,0,0               ; movaps        0x3ffa(%rip),%xmm8        # 4ea0 <_sk_callback_sse2+0x286>
+  DB  68,15,40,5,138,64,0,0               ; movaps        0x408a(%rip),%xmm8        # 4f30 <_sk_callback_sse2+0x28a>
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
   DB  15,40,207                           ; movaps        %xmm7,%xmm1
   DB  68,15,92,217                        ; subps         %xmm1,%xmm11
@@ -18851,12 +18971,12 @@ _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  68,15,40,61,14,63,0,0               ; movaps        0x3f0e(%rip),%xmm15        # 4e70 <_sk_callback_sse2+0x256>
+  DB  68,15,40,61,158,63,0,0              ; movaps        0x3f9e(%rip),%xmm15        # 4f00 <_sk_callback_sse2+0x25a>
   DB  65,15,89,231                        ; mulps         %xmm15,%xmm4
-  DB  15,40,5,19,63,0,0                   ; movaps        0x3f13(%rip),%xmm0        # 4e80 <_sk_callback_sse2+0x266>
+  DB  15,40,5,163,63,0,0                  ; movaps        0x3fa3(%rip),%xmm0        # 4f10 <_sk_callback_sse2+0x26a>
   DB  15,89,240                           ; mulps         %xmm0,%xmm6
   DB  15,88,244                           ; addps         %xmm4,%xmm6
-  DB  68,15,40,53,21,63,0,0               ; movaps        0x3f15(%rip),%xmm14        # 4e90 <_sk_callback_sse2+0x276>
+  DB  68,15,40,53,165,63,0,0              ; movaps        0x3fa5(%rip),%xmm14        # 4f20 <_sk_callback_sse2+0x27a>
   DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
   DB  69,15,89,238                        ; mulps         %xmm14,%xmm13
   DB  68,15,88,238                        ; addps         %xmm6,%xmm13
@@ -19033,14 +19153,14 @@ _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  68,15,40,5,165,60,0,0               ; movaps        0x3ca5(%rip),%xmm8        # 4eb0 <_sk_callback_sse2+0x296>
+  DB  68,15,40,5,53,61,0,0                ; movaps        0x3d35(%rip),%xmm8        # 4f40 <_sk_callback_sse2+0x29a>
   DB  15,40,214                           ; movaps        %xmm6,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
-  DB  15,40,13,167,60,0,0                 ; movaps        0x3ca7(%rip),%xmm1        # 4ec0 <_sk_callback_sse2+0x2a6>
+  DB  15,40,13,55,61,0,0                  ; movaps        0x3d37(%rip),%xmm1        # 4f50 <_sk_callback_sse2+0x2aa>
   DB  15,40,221                           ; movaps        %xmm5,%xmm3
   DB  15,89,217                           ; mulps         %xmm1,%xmm3
   DB  15,88,218                           ; addps         %xmm2,%xmm3
-  DB  68,15,40,37,166,60,0,0              ; movaps        0x3ca6(%rip),%xmm12        # 4ed0 <_sk_callback_sse2+0x2b6>
+  DB  68,15,40,37,54,61,0,0               ; movaps        0x3d36(%rip),%xmm12        # 4f60 <_sk_callback_sse2+0x2ba>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
   DB  68,15,88,235                        ; addps         %xmm3,%xmm13
   DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
@@ -19085,7 +19205,7 @@ _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  68,15,40,5,11,60,0,0                ; movaps        0x3c0b(%rip),%xmm8        # 4ee0 <_sk_callback_sse2+0x2c6>
+  DB  68,15,40,5,155,60,0,0               ; movaps        0x3c9b(%rip),%xmm8        # 4f70 <_sk_callback_sse2+0x2ca>
   DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  15,88,253                           ; addps         %xmm5,%xmm7
@@ -19186,14 +19306,14 @@ _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  68,15,40,45,163,58,0,0              ; movaps        0x3aa3(%rip),%xmm13        # 4ef0 <_sk_callback_sse2+0x2d6>
+  DB  68,15,40,45,51,59,0,0               ; movaps        0x3b33(%rip),%xmm13        # 4f80 <_sk_callback_sse2+0x2da>
   DB  68,15,40,198                        ; movaps        %xmm6,%xmm8
   DB  69,15,89,197                        ; mulps         %xmm13,%xmm8
-  DB  68,15,40,53,163,58,0,0              ; movaps        0x3aa3(%rip),%xmm14        # 4f00 <_sk_callback_sse2+0x2e6>
+  DB  68,15,40,53,51,59,0,0               ; movaps        0x3b33(%rip),%xmm14        # 4f90 <_sk_callback_sse2+0x2ea>
   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  68,15,40,29,159,58,0,0              ; movaps        0x3a9f(%rip),%xmm11        # 4f10 <_sk_callback_sse2+0x2f6>
+  DB  68,15,40,29,47,59,0,0               ; movaps        0x3b2f(%rip),%xmm11        # 4fa0 <_sk_callback_sse2+0x2fa>
   DB  69,15,89,227                        ; mulps         %xmm11,%xmm12
   DB  68,15,88,224                        ; addps         %xmm0,%xmm12
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
@@ -19201,7 +19321,7 @@ _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  68,15,40,5,139,58,0,0               ; movaps        0x3a8b(%rip),%xmm8        # 4f20 <_sk_callback_sse2+0x306>
+  DB  68,15,40,5,27,59,0,0                ; movaps        0x3b1b(%rip),%xmm8        # 4fb0 <_sk_callback_sse2+0x30a>
   DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
   DB  15,92,226                           ; subps         %xmm2,%xmm4
   DB  15,89,252                           ; mulps         %xmm4,%xmm7
@@ -19337,15 +19457,15 @@ _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  68,15,40,37,157,56,0,0              ; movaps        0x389d(%rip),%xmm12        # 4f30 <_sk_callback_sse2+0x316>
+  DB  68,15,40,37,45,57,0,0               ; movaps        0x392d(%rip),%xmm12        # 4fc0 <_sk_callback_sse2+0x31a>
   DB  68,15,40,199                        ; movaps        %xmm7,%xmm8
   DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
-  DB  68,15,40,45,157,56,0,0              ; movaps        0x389d(%rip),%xmm13        # 4f40 <_sk_callback_sse2+0x326>
+  DB  68,15,40,45,45,57,0,0               ; movaps        0x392d(%rip),%xmm13        # 4fd0 <_sk_callback_sse2+0x32a>
   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  68,15,40,29,153,56,0,0              ; movaps        0x3899(%rip),%xmm11        # 4f50 <_sk_callback_sse2+0x336>
-  DB  68,15,40,5,161,56,0,0               ; movaps        0x38a1(%rip),%xmm8        # 4f60 <_sk_callback_sse2+0x346>
+  DB  68,15,40,29,41,57,0,0               ; movaps        0x3929(%rip),%xmm11        # 4fe0 <_sk_callback_sse2+0x33a>
+  DB  68,15,40,5,49,57,0,0                ; movaps        0x3931(%rip),%xmm8        # 4ff0 <_sk_callback_sse2+0x34a>
   DB  69,15,40,248                        ; movaps        %xmm8,%xmm15
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  68,15,92,248                        ; subps         %xmm0,%xmm15
@@ -19487,7 +19607,7 @@ _sk_clamp_0_sse2 LABEL PROC
 
 PUBLIC _sk_clamp_1_sse2
 _sk_clamp_1_sse2 LABEL PROC
-  DB  68,15,40,5,168,54,0,0               ; movaps        0x36a8(%rip),%xmm8        # 4f70 <_sk_callback_sse2+0x356>
+  DB  68,15,40,5,56,55,0,0                ; movaps        0x3738(%rip),%xmm8        # 5000 <_sk_callback_sse2+0x35a>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
   DB  65,15,93,208                        ; minps         %xmm8,%xmm2
@@ -19497,7 +19617,7 @@ _sk_clamp_1_sse2 LABEL PROC
 
 PUBLIC _sk_clamp_a_sse2
 _sk_clamp_a_sse2 LABEL PROC
-  DB  15,93,29,157,54,0,0                 ; minps         0x369d(%rip),%xmm3        # 4f80 <_sk_callback_sse2+0x366>
+  DB  15,93,29,45,55,0,0                  ; minps         0x372d(%rip),%xmm3        # 5010 <_sk_callback_sse2+0x36a>
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -19570,7 +19690,7 @@ _sk_premul_sse2 LABEL PROC
 PUBLIC _sk_unpremul_sse2
 _sk_unpremul_sse2 LABEL PROC
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
-  DB  68,15,40,13,8,54,0,0                ; movaps        0x3608(%rip),%xmm9        # 4f90 <_sk_callback_sse2+0x376>
+  DB  68,15,40,13,152,54,0,0              ; movaps        0x3698(%rip),%xmm9        # 5020 <_sk_callback_sse2+0x37a>
   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
@@ -19582,20 +19702,20 @@ _sk_unpremul_sse2 LABEL PROC
 
 PUBLIC _sk_from_srgb_sse2
 _sk_from_srgb_sse2 LABEL PROC
-  DB  68,15,40,5,243,53,0,0               ; movaps        0x35f3(%rip),%xmm8        # 4fa0 <_sk_callback_sse2+0x386>
+  DB  68,15,40,5,131,54,0,0               ; movaps        0x3683(%rip),%xmm8        # 5030 <_sk_callback_sse2+0x38a>
   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  68,15,40,13,235,53,0,0              ; movaps        0x35eb(%rip),%xmm9        # 4fb0 <_sk_callback_sse2+0x396>
+  DB  68,15,40,13,123,54,0,0              ; movaps        0x367b(%rip),%xmm9        # 5040 <_sk_callback_sse2+0x39a>
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
   DB  69,15,89,241                        ; mulps         %xmm9,%xmm14
-  DB  68,15,40,21,235,53,0,0              ; movaps        0x35eb(%rip),%xmm10        # 4fc0 <_sk_callback_sse2+0x3a6>
+  DB  68,15,40,21,123,54,0,0              ; movaps        0x367b(%rip),%xmm10        # 5050 <_sk_callback_sse2+0x3aa>
   DB  69,15,88,242                        ; addps         %xmm10,%xmm14
   DB  69,15,89,243                        ; mulps         %xmm11,%xmm14
-  DB  68,15,40,29,235,53,0,0              ; movaps        0x35eb(%rip),%xmm11        # 4fd0 <_sk_callback_sse2+0x3b6>
+  DB  68,15,40,29,123,54,0,0              ; movaps        0x367b(%rip),%xmm11        # 5060 <_sk_callback_sse2+0x3ba>
   DB  69,15,88,243                        ; addps         %xmm11,%xmm14
-  DB  68,15,40,37,239,53,0,0              ; movaps        0x35ef(%rip),%xmm12        # 4fe0 <_sk_callback_sse2+0x3c6>
+  DB  68,15,40,37,127,54,0,0              ; movaps        0x367f(%rip),%xmm12        # 5070 <_sk_callback_sse2+0x3ca>
   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
@@ -19630,22 +19750,22 @@ _sk_from_srgb_sse2 LABEL PROC
 PUBLIC _sk_to_srgb_sse2
 _sk_to_srgb_sse2 LABEL PROC
   DB  68,15,82,232                        ; rsqrtps       %xmm0,%xmm13
-  DB  68,15,40,5,124,53,0,0               ; movaps        0x357c(%rip),%xmm8        # 4ff0 <_sk_callback_sse2+0x3d6>
+  DB  68,15,40,5,12,54,0,0                ; movaps        0x360c(%rip),%xmm8        # 5080 <_sk_callback_sse2+0x3da>
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
   DB  69,15,89,240                        ; mulps         %xmm8,%xmm14
-  DB  68,15,40,13,124,53,0,0              ; movaps        0x357c(%rip),%xmm9        # 5000 <_sk_callback_sse2+0x3e6>
+  DB  68,15,40,13,12,54,0,0               ; movaps        0x360c(%rip),%xmm9        # 5090 <_sk_callback_sse2+0x3ea>
   DB  69,15,40,253                        ; movaps        %xmm13,%xmm15
   DB  69,15,89,249                        ; mulps         %xmm9,%xmm15
-  DB  68,15,40,21,124,53,0,0              ; movaps        0x357c(%rip),%xmm10        # 5010 <_sk_callback_sse2+0x3f6>
+  DB  68,15,40,21,12,54,0,0               ; movaps        0x360c(%rip),%xmm10        # 50a0 <_sk_callback_sse2+0x3fa>
   DB  69,15,88,250                        ; addps         %xmm10,%xmm15
   DB  69,15,89,253                        ; mulps         %xmm13,%xmm15
-  DB  68,15,40,29,124,53,0,0              ; movaps        0x357c(%rip),%xmm11        # 5020 <_sk_callback_sse2+0x406>
+  DB  68,15,40,29,12,54,0,0               ; movaps        0x360c(%rip),%xmm11        # 50b0 <_sk_callback_sse2+0x40a>
   DB  69,15,88,251                        ; addps         %xmm11,%xmm15
-  DB  68,15,40,37,128,53,0,0              ; movaps        0x3580(%rip),%xmm12        # 5030 <_sk_callback_sse2+0x416>
+  DB  68,15,40,37,16,54,0,0               ; movaps        0x3610(%rip),%xmm12        # 50c0 <_sk_callback_sse2+0x41a>
   DB  69,15,88,236                        ; addps         %xmm12,%xmm13
   DB  69,15,83,237                        ; rcpps         %xmm13,%xmm13
   DB  69,15,89,239                        ; mulps         %xmm15,%xmm13
-  DB  68,15,40,61,124,53,0,0              ; movaps        0x357c(%rip),%xmm15        # 5040 <_sk_callback_sse2+0x426>
+  DB  68,15,40,61,12,54,0,0               ; movaps        0x360c(%rip),%xmm15        # 50d0 <_sk_callback_sse2+0x42a>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  68,15,84,240                        ; andps         %xmm0,%xmm14
   DB  65,15,85,197                        ; andnps        %xmm13,%xmm0
@@ -19693,7 +19813,7 @@ _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  68,15,40,45,213,52,0,0              ; movaps        0x34d5(%rip),%xmm13        # 5050 <_sk_callback_sse2+0x436>
+  DB  68,15,40,45,101,53,0,0              ; movaps        0x3565(%rip),%xmm13        # 50e0 <_sk_callback_sse2+0x43a>
   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
@@ -19702,30 +19822,30 @@ _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  68,15,84,53,187,52,0,0              ; andps         0x34bb(%rip),%xmm14        # 5060 <_sk_callback_sse2+0x446>
+  DB  68,15,84,53,75,53,0,0               ; andps         0x354b(%rip),%xmm14        # 50f0 <_sk_callback_sse2+0x44a>
   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  68,15,40,37,174,52,0,0              ; movaps        0x34ae(%rip),%xmm12        # 5070 <_sk_callback_sse2+0x456>
+  DB  68,15,40,37,62,53,0,0               ; movaps        0x353e(%rip),%xmm12        # 5100 <_sk_callback_sse2+0x45a>
   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  68,15,88,5,170,52,0,0               ; addps         0x34aa(%rip),%xmm8        # 5080 <_sk_callback_sse2+0x466>
+  DB  68,15,88,5,58,53,0,0                ; addps         0x353a(%rip),%xmm8        # 5110 <_sk_callback_sse2+0x46a>
   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  15,89,5,155,52,0,0                  ; mulps         0x349b(%rip),%xmm0        # 5090 <_sk_callback_sse2+0x476>
+  DB  15,89,5,43,53,0,0                   ; mulps         0x352b(%rip),%xmm0        # 5120 <_sk_callback_sse2+0x47a>
   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  68,15,40,13,142,52,0,0              ; movaps        0x348e(%rip),%xmm9        # 50a0 <_sk_callback_sse2+0x486>
+  DB  68,15,40,13,30,53,0,0               ; movaps        0x351e(%rip),%xmm9        # 5130 <_sk_callback_sse2+0x48a>
   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
@@ -19748,7 +19868,7 @@ _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  68,15,40,13,73,52,0,0               ; movaps        0x3449(%rip),%xmm9        # 50b0 <_sk_callback_sse2+0x496>
+  DB  68,15,40,13,217,52,0,0              ; movaps        0x34d9(%rip),%xmm9        # 5140 <_sk_callback_sse2+0x49a>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  69,15,194,211,2                     ; cmpleps       %xmm11,%xmm10
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
@@ -19765,28 +19885,28 @@ _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  15,40,5,19,52,0,0                   ; movaps        0x3413(%rip),%xmm0        # 50c0 <_sk_callback_sse2+0x4a6>
+  DB  15,40,5,163,52,0,0                  ; movaps        0x34a3(%rip),%xmm0        # 5150 <_sk_callback_sse2+0x4aa>
   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  15,84,29,11,52,0,0                  ; andps         0x340b(%rip),%xmm3        # 50d0 <_sk_callback_sse2+0x4b6>
+  DB  15,84,29,155,52,0,0                 ; andps         0x349b(%rip),%xmm3        # 5160 <_sk_callback_sse2+0x4ba>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,92,193                           ; subps         %xmm1,%xmm0
-  DB  68,15,40,45,13,52,0,0               ; movaps        0x340d(%rip),%xmm13        # 50e0 <_sk_callback_sse2+0x4c6>
+  DB  68,15,40,45,157,52,0,0              ; movaps        0x349d(%rip),%xmm13        # 5170 <_sk_callback_sse2+0x4ca>
   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  15,40,21,29,52,0,0                  ; movaps        0x341d(%rip),%xmm2        # 5110 <_sk_callback_sse2+0x4f6>
+  DB  15,40,21,173,52,0,0                 ; movaps        0x34ad(%rip),%xmm2        # 51a0 <_sk_callback_sse2+0x4fa>
   DB  68,15,40,250                        ; movaps        %xmm2,%xmm15
   DB  68,15,194,248,2                     ; cmpleps       %xmm0,%xmm15
-  DB  15,40,13,237,51,0,0                 ; movaps        0x33ed(%rip),%xmm1        # 50f0 <_sk_callback_sse2+0x4d6>
+  DB  15,40,13,125,52,0,0                 ; movaps        0x347d(%rip),%xmm1        # 5180 <_sk_callback_sse2+0x4da>
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
-  DB  15,40,45,243,51,0,0                 ; movaps        0x33f3(%rip),%xmm5        # 5100 <_sk_callback_sse2+0x4e6>
+  DB  15,40,45,131,52,0,0                 ; movaps        0x3483(%rip),%xmm5        # 5190 <_sk_callback_sse2+0x4ea>
   DB  15,40,229                           ; movaps        %xmm5,%xmm4
   DB  15,92,224                           ; subps         %xmm0,%xmm4
   DB  65,15,89,230                        ; mulps         %xmm14,%xmm4
@@ -19809,7 +19929,7 @@ _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  15,84,29,104,51,0,0                 ; andps         0x3368(%rip),%xmm3        # 50d0 <_sk_callback_sse2+0x4b6>
+  DB  15,84,29,248,51,0,0                 ; andps         0x33f8(%rip),%xmm3        # 5160 <_sk_callback_sse2+0x4ba>
   DB  15,92,195                           ; subps         %xmm3,%xmm0
   DB  68,15,40,230                        ; movaps        %xmm6,%xmm12
   DB  68,15,92,224                        ; subps         %xmm0,%xmm12
@@ -19839,12 +19959,12 @@ _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  15,88,53,65,51,0,0                  ; addps         0x3341(%rip),%xmm6        # 5120 <_sk_callback_sse2+0x506>
+  DB  15,88,53,209,51,0,0                 ; addps         0x33d1(%rip),%xmm6        # 51b0 <_sk_callback_sse2+0x50a>
   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  15,84,29,220,50,0,0                 ; andps         0x32dc(%rip),%xmm3        # 50d0 <_sk_callback_sse2+0x4b6>
+  DB  15,84,29,108,51,0,0                 ; andps         0x336c(%rip),%xmm3        # 5160 <_sk_callback_sse2+0x4ba>
   DB  15,92,195                           ; subps         %xmm3,%xmm0
   DB  15,92,240                           ; subps         %xmm0,%xmm6
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
@@ -19905,7 +20025,7 @@ _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  68,15,89,5,102,50,0,0               ; mulps         0x3266(%rip),%xmm8        # 5130 <_sk_callback_sse2+0x516>
+  DB  68,15,89,5,246,50,0,0               ; mulps         0x32f6(%rip),%xmm8        # 51c0 <_sk_callback_sse2+0x51a>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
@@ -19942,7 +20062,7 @@ _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  68,15,89,5,4,50,0,0                 ; mulps         0x3204(%rip),%xmm8        # 5140 <_sk_callback_sse2+0x526>
+  DB  68,15,89,5,148,50,0,0               ; mulps         0x3294(%rip),%xmm8        # 51d0 <_sk_callback_sse2+0x52a>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -19965,17 +20085,17 @@ _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  102,68,15,111,5,202,49,0,0          ; movdqa        0x31ca(%rip),%xmm8        # 5150 <_sk_callback_sse2+0x536>
+  DB  102,68,15,111,5,90,50,0,0           ; movdqa        0x325a(%rip),%xmm8        # 51e0 <_sk_callback_sse2+0x53a>
   DB  102,69,15,219,194                   ; pand          %xmm10,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,201,49,0,0               ; mulps         0x31c9(%rip),%xmm8        # 5160 <_sk_callback_sse2+0x546>
-  DB  102,68,15,111,13,208,49,0,0         ; movdqa        0x31d0(%rip),%xmm9        # 5170 <_sk_callback_sse2+0x556>
+  DB  68,15,89,5,89,50,0,0                ; mulps         0x3259(%rip),%xmm8        # 51f0 <_sk_callback_sse2+0x54a>
+  DB  102,68,15,111,13,96,50,0,0          ; movdqa        0x3260(%rip),%xmm9        # 5200 <_sk_callback_sse2+0x55a>
   DB  102,69,15,219,202                   ; pand          %xmm10,%xmm9
   DB  69,15,91,201                        ; cvtdq2ps      %xmm9,%xmm9
-  DB  68,15,89,13,207,49,0,0              ; mulps         0x31cf(%rip),%xmm9        # 5180 <_sk_callback_sse2+0x566>
-  DB  102,68,15,219,21,214,49,0,0         ; pand          0x31d6(%rip),%xmm10        # 5190 <_sk_callback_sse2+0x576>
+  DB  68,15,89,13,95,50,0,0               ; mulps         0x325f(%rip),%xmm9        # 5210 <_sk_callback_sse2+0x56a>
+  DB  102,68,15,219,21,102,50,0,0         ; pand          0x3266(%rip),%xmm10        # 5220 <_sk_callback_sse2+0x57a>
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
-  DB  68,15,89,21,218,49,0,0              ; mulps         0x31da(%rip),%xmm10        # 51a0 <_sk_callback_sse2+0x586>
+  DB  68,15,89,21,106,50,0,0              ; mulps         0x326a(%rip),%xmm10        # 5230 <_sk_callback_sse2+0x58a>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -20004,7 +20124,7 @@ _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  102,68,15,111,5,138,49,0,0          ; movdqa        0x318a(%rip),%xmm8        # 51b0 <_sk_callback_sse2+0x596>
+  DB  102,68,15,111,5,26,50,0,0           ; movdqa        0x321a(%rip),%xmm8        # 5240 <_sk_callback_sse2+0x59a>
   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
@@ -20059,7 +20179,7 @@ _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  15,89,29,151,48,0,0                 ; mulps         0x3097(%rip),%xmm3        # 51c0 <_sk_callback_sse2+0x5a6>
+  DB  15,89,29,39,49,0,0                  ; mulps         0x3127(%rip),%xmm3        # 5250 <_sk_callback_sse2+0x5aa>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -20076,7 +20196,7 @@ _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  102,68,15,111,21,106,48,0,0         ; movdqa        0x306a(%rip),%xmm10        # 51d0 <_sk_callback_sse2+0x5b6>
+  DB  102,68,15,111,21,250,48,0,0         ; movdqa        0x30fa(%rip),%xmm10        # 5260 <_sk_callback_sse2+0x5ba>
   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
@@ -20137,7 +20257,7 @@ _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  15,89,29,89,47,0,0                  ; mulps         0x2f59(%rip),%xmm3        # 51e0 <_sk_callback_sse2+0x5c6>
+  DB  15,89,29,233,47,0,0                 ; mulps         0x2fe9(%rip),%xmm3        # 5270 <_sk_callback_sse2+0x5ca>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -20157,7 +20277,7 @@ _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  102,68,15,111,5,25,47,0,0           ; movdqa        0x2f19(%rip),%xmm8        # 51f0 <_sk_callback_sse2+0x5d6>
+  DB  102,68,15,111,5,169,47,0,0          ; movdqa        0x2fa9(%rip),%xmm8        # 5280 <_sk_callback_sse2+0x5da>
   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
@@ -20212,7 +20332,7 @@ _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  15,40,29,40,46,0,0                  ; movaps        0x2e28(%rip),%xmm3        # 5200 <_sk_callback_sse2+0x5e6>
+  DB  15,40,29,184,46,0,0                 ; movaps        0x2eb8(%rip),%xmm3        # 5290 <_sk_callback_sse2+0x5ea>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_byte_tables_sse2
@@ -20220,7 +20340,7 @@ _sk_byte_tables_sse2 LABEL PROC
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,41,46,0,0                ; movaps        0x2e29(%rip),%xmm8        # 5210 <_sk_callback_sse2+0x5f6>
+  DB  68,15,40,5,185,46,0,0               ; movaps        0x2eb9(%rip),%xmm8        # 52a0 <_sk_callback_sse2+0x5fa>
   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
@@ -20247,7 +20367,7 @@ _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  68,15,40,21,198,45,0,0              ; movaps        0x2dc6(%rip),%xmm10        # 5220 <_sk_callback_sse2+0x606>
+  DB  68,15,40,21,86,46,0,0               ; movaps        0x2e56(%rip),%xmm10        # 52b0 <_sk_callback_sse2+0x60a>
   DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -20361,7 +20481,7 @@ _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  68,15,40,21,25,44,0,0               ; movaps        0x2c19(%rip),%xmm10        # 5230 <_sk_callback_sse2+0x616>
+  DB  68,15,40,21,169,44,0,0              ; movaps        0x2ca9(%rip),%xmm10        # 52c0 <_sk_callback_sse2+0x61a>
   DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -20548,15 +20668,15 @@ _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  68,15,89,13,88,41,0,0               ; mulps         0x2958(%rip),%xmm9        # 5240 <_sk_callback_sse2+0x626>
-  DB  68,15,84,21,96,41,0,0               ; andps         0x2960(%rip),%xmm10        # 5250 <_sk_callback_sse2+0x636>
-  DB  68,15,86,21,104,41,0,0              ; orps          0x2968(%rip),%xmm10        # 5260 <_sk_callback_sse2+0x646>
-  DB  68,15,88,13,112,41,0,0              ; addps         0x2970(%rip),%xmm9        # 5270 <_sk_callback_sse2+0x656>
-  DB  68,15,40,37,120,41,0,0              ; movaps        0x2978(%rip),%xmm12        # 5280 <_sk_callback_sse2+0x666>
+  DB  68,15,89,13,232,41,0,0              ; mulps         0x29e8(%rip),%xmm9        # 52d0 <_sk_callback_sse2+0x62a>
+  DB  68,15,84,21,240,41,0,0              ; andps         0x29f0(%rip),%xmm10        # 52e0 <_sk_callback_sse2+0x63a>
+  DB  68,15,86,21,248,41,0,0              ; orps          0x29f8(%rip),%xmm10        # 52f0 <_sk_callback_sse2+0x64a>
+  DB  68,15,88,13,0,42,0,0                ; addps         0x2a00(%rip),%xmm9        # 5300 <_sk_callback_sse2+0x65a>
+  DB  68,15,40,37,8,42,0,0                ; movaps        0x2a08(%rip),%xmm12        # 5310 <_sk_callback_sse2+0x66a>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,120,41,0,0              ; addps         0x2978(%rip),%xmm10        # 5290 <_sk_callback_sse2+0x676>
-  DB  68,15,40,37,128,41,0,0              ; movaps        0x2980(%rip),%xmm12        # 52a0 <_sk_callback_sse2+0x686>
+  DB  68,15,88,21,8,42,0,0                ; addps         0x2a08(%rip),%xmm10        # 5320 <_sk_callback_sse2+0x67a>
+  DB  68,15,40,37,16,42,0,0               ; movaps        0x2a10(%rip),%xmm12        # 5330 <_sk_callback_sse2+0x68a>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20564,22 +20684,22 @@ _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  68,15,40,21,106,41,0,0              ; movaps        0x296a(%rip),%xmm10        # 52b0 <_sk_callback_sse2+0x696>
+  DB  68,15,40,21,250,41,0,0              ; movaps        0x29fa(%rip),%xmm10        # 5340 <_sk_callback_sse2+0x69a>
   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  68,15,88,13,94,41,0,0               ; addps         0x295e(%rip),%xmm9        # 52c0 <_sk_callback_sse2+0x6a6>
-  DB  68,15,40,37,102,41,0,0              ; movaps        0x2966(%rip),%xmm12        # 52d0 <_sk_callback_sse2+0x6b6>
+  DB  68,15,88,13,238,41,0,0              ; addps         0x29ee(%rip),%xmm9        # 5350 <_sk_callback_sse2+0x6aa>
+  DB  68,15,40,37,246,41,0,0              ; movaps        0x29f6(%rip),%xmm12        # 5360 <_sk_callback_sse2+0x6ba>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,102,41,0,0              ; movaps        0x2966(%rip),%xmm12        # 52e0 <_sk_callback_sse2+0x6c6>
+  DB  68,15,40,37,246,41,0,0              ; movaps        0x29f6(%rip),%xmm12        # 5370 <_sk_callback_sse2+0x6ca>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,106,41,0,0              ; movaps        0x296a(%rip),%xmm13        # 52f0 <_sk_callback_sse2+0x6d6>
+  DB  68,15,40,45,250,41,0,0              ; movaps        0x29fa(%rip),%xmm13        # 5380 <_sk_callback_sse2+0x6da>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,106,41,0,0              ; mulps         0x296a(%rip),%xmm13        # 5300 <_sk_callback_sse2+0x6e6>
+  DB  68,15,89,45,250,41,0,0              ; mulps         0x29fa(%rip),%xmm13        # 5390 <_sk_callback_sse2+0x6ea>
   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
@@ -20613,15 +20733,15 @@ _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  68,15,89,13,234,40,0,0              ; mulps         0x28ea(%rip),%xmm9        # 5310 <_sk_callback_sse2+0x6f6>
-  DB  68,15,84,21,242,40,0,0              ; andps         0x28f2(%rip),%xmm10        # 5320 <_sk_callback_sse2+0x706>
-  DB  68,15,86,21,250,40,0,0              ; orps          0x28fa(%rip),%xmm10        # 5330 <_sk_callback_sse2+0x716>
-  DB  68,15,88,13,2,41,0,0                ; addps         0x2902(%rip),%xmm9        # 5340 <_sk_callback_sse2+0x726>
-  DB  68,15,40,37,10,41,0,0               ; movaps        0x290a(%rip),%xmm12        # 5350 <_sk_callback_sse2+0x736>
+  DB  68,15,89,13,122,41,0,0              ; mulps         0x297a(%rip),%xmm9        # 53a0 <_sk_callback_sse2+0x6fa>
+  DB  68,15,84,21,130,41,0,0              ; andps         0x2982(%rip),%xmm10        # 53b0 <_sk_callback_sse2+0x70a>
+  DB  68,15,86,21,138,41,0,0              ; orps          0x298a(%rip),%xmm10        # 53c0 <_sk_callback_sse2+0x71a>
+  DB  68,15,88,13,146,41,0,0              ; addps         0x2992(%rip),%xmm9        # 53d0 <_sk_callback_sse2+0x72a>
+  DB  68,15,40,37,154,41,0,0              ; movaps        0x299a(%rip),%xmm12        # 53e0 <_sk_callback_sse2+0x73a>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,10,41,0,0               ; addps         0x290a(%rip),%xmm10        # 5360 <_sk_callback_sse2+0x746>
-  DB  68,15,40,37,18,41,0,0               ; movaps        0x2912(%rip),%xmm12        # 5370 <_sk_callback_sse2+0x756>
+  DB  68,15,88,21,154,41,0,0              ; addps         0x299a(%rip),%xmm10        # 53f0 <_sk_callback_sse2+0x74a>
+  DB  68,15,40,37,162,41,0,0              ; movaps        0x29a2(%rip),%xmm12        # 5400 <_sk_callback_sse2+0x75a>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20629,22 +20749,22 @@ _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  68,15,40,21,252,40,0,0              ; movaps        0x28fc(%rip),%xmm10        # 5380 <_sk_callback_sse2+0x766>
+  DB  68,15,40,21,140,41,0,0              ; movaps        0x298c(%rip),%xmm10        # 5410 <_sk_callback_sse2+0x76a>
   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  68,15,88,13,240,40,0,0              ; addps         0x28f0(%rip),%xmm9        # 5390 <_sk_callback_sse2+0x776>
-  DB  68,15,40,37,248,40,0,0              ; movaps        0x28f8(%rip),%xmm12        # 53a0 <_sk_callback_sse2+0x786>
+  DB  68,15,88,13,128,41,0,0              ; addps         0x2980(%rip),%xmm9        # 5420 <_sk_callback_sse2+0x77a>
+  DB  68,15,40,37,136,41,0,0              ; movaps        0x2988(%rip),%xmm12        # 5430 <_sk_callback_sse2+0x78a>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,248,40,0,0              ; movaps        0x28f8(%rip),%xmm12        # 53b0 <_sk_callback_sse2+0x796>
+  DB  68,15,40,37,136,41,0,0              ; movaps        0x2988(%rip),%xmm12        # 5440 <_sk_callback_sse2+0x79a>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,252,40,0,0              ; movaps        0x28fc(%rip),%xmm13        # 53c0 <_sk_callback_sse2+0x7a6>
+  DB  68,15,40,45,140,41,0,0              ; movaps        0x298c(%rip),%xmm13        # 5450 <_sk_callback_sse2+0x7aa>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,252,40,0,0              ; mulps         0x28fc(%rip),%xmm13        # 53d0 <_sk_callback_sse2+0x7b6>
+  DB  68,15,89,45,140,41,0,0              ; mulps         0x298c(%rip),%xmm13        # 5460 <_sk_callback_sse2+0x7ba>
   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
@@ -20678,15 +20798,15 @@ _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  68,15,89,13,124,40,0,0              ; mulps         0x287c(%rip),%xmm9        # 53e0 <_sk_callback_sse2+0x7c6>
-  DB  68,15,84,21,132,40,0,0              ; andps         0x2884(%rip),%xmm10        # 53f0 <_sk_callback_sse2+0x7d6>
-  DB  68,15,86,21,140,40,0,0              ; orps          0x288c(%rip),%xmm10        # 5400 <_sk_callback_sse2+0x7e6>
-  DB  68,15,88,13,148,40,0,0              ; addps         0x2894(%rip),%xmm9        # 5410 <_sk_callback_sse2+0x7f6>
-  DB  68,15,40,37,156,40,0,0              ; movaps        0x289c(%rip),%xmm12        # 5420 <_sk_callback_sse2+0x806>
+  DB  68,15,89,13,12,41,0,0               ; mulps         0x290c(%rip),%xmm9        # 5470 <_sk_callback_sse2+0x7ca>
+  DB  68,15,84,21,20,41,0,0               ; andps         0x2914(%rip),%xmm10        # 5480 <_sk_callback_sse2+0x7da>
+  DB  68,15,86,21,28,41,0,0               ; orps          0x291c(%rip),%xmm10        # 5490 <_sk_callback_sse2+0x7ea>
+  DB  68,15,88,13,36,41,0,0               ; addps         0x2924(%rip),%xmm9        # 54a0 <_sk_callback_sse2+0x7fa>
+  DB  68,15,40,37,44,41,0,0               ; movaps        0x292c(%rip),%xmm12        # 54b0 <_sk_callback_sse2+0x80a>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,156,40,0,0              ; addps         0x289c(%rip),%xmm10        # 5430 <_sk_callback_sse2+0x816>
-  DB  68,15,40,37,164,40,0,0              ; movaps        0x28a4(%rip),%xmm12        # 5440 <_sk_callback_sse2+0x826>
+  DB  68,15,88,21,44,41,0,0               ; addps         0x292c(%rip),%xmm10        # 54c0 <_sk_callback_sse2+0x81a>
+  DB  68,15,40,37,52,41,0,0               ; movaps        0x2934(%rip),%xmm12        # 54d0 <_sk_callback_sse2+0x82a>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20694,22 +20814,22 @@ _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  68,15,40,21,142,40,0,0              ; movaps        0x288e(%rip),%xmm10        # 5450 <_sk_callback_sse2+0x836>
+  DB  68,15,40,21,30,41,0,0               ; movaps        0x291e(%rip),%xmm10        # 54e0 <_sk_callback_sse2+0x83a>
   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  68,15,88,13,130,40,0,0              ; addps         0x2882(%rip),%xmm9        # 5460 <_sk_callback_sse2+0x846>
-  DB  68,15,40,37,138,40,0,0              ; movaps        0x288a(%rip),%xmm12        # 5470 <_sk_callback_sse2+0x856>
+  DB  68,15,88,13,18,41,0,0               ; addps         0x2912(%rip),%xmm9        # 54f0 <_sk_callback_sse2+0x84a>
+  DB  68,15,40,37,26,41,0,0               ; movaps        0x291a(%rip),%xmm12        # 5500 <_sk_callback_sse2+0x85a>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,138,40,0,0              ; movaps        0x288a(%rip),%xmm12        # 5480 <_sk_callback_sse2+0x866>
+  DB  68,15,40,37,26,41,0,0               ; movaps        0x291a(%rip),%xmm12        # 5510 <_sk_callback_sse2+0x86a>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,142,40,0,0              ; movaps        0x288e(%rip),%xmm13        # 5490 <_sk_callback_sse2+0x876>
+  DB  68,15,40,45,30,41,0,0               ; movaps        0x291e(%rip),%xmm13        # 5520 <_sk_callback_sse2+0x87a>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,142,40,0,0              ; mulps         0x288e(%rip),%xmm13        # 54a0 <_sk_callback_sse2+0x886>
+  DB  68,15,89,45,30,41,0,0               ; mulps         0x291e(%rip),%xmm13        # 5530 <_sk_callback_sse2+0x88a>
   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
@@ -20743,15 +20863,15 @@ _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  68,15,89,13,14,40,0,0               ; mulps         0x280e(%rip),%xmm9        # 54b0 <_sk_callback_sse2+0x896>
-  DB  68,15,84,21,22,40,0,0               ; andps         0x2816(%rip),%xmm10        # 54c0 <_sk_callback_sse2+0x8a6>
-  DB  68,15,86,21,30,40,0,0               ; orps          0x281e(%rip),%xmm10        # 54d0 <_sk_callback_sse2+0x8b6>
-  DB  68,15,88,13,38,40,0,0               ; addps         0x2826(%rip),%xmm9        # 54e0 <_sk_callback_sse2+0x8c6>
-  DB  68,15,40,37,46,40,0,0               ; movaps        0x282e(%rip),%xmm12        # 54f0 <_sk_callback_sse2+0x8d6>
+  DB  68,15,89,13,158,40,0,0              ; mulps         0x289e(%rip),%xmm9        # 5540 <_sk_callback_sse2+0x89a>
+  DB  68,15,84,21,166,40,0,0              ; andps         0x28a6(%rip),%xmm10        # 5550 <_sk_callback_sse2+0x8aa>
+  DB  68,15,86,21,174,40,0,0              ; orps          0x28ae(%rip),%xmm10        # 5560 <_sk_callback_sse2+0x8ba>
+  DB  68,15,88,13,182,40,0,0              ; addps         0x28b6(%rip),%xmm9        # 5570 <_sk_callback_sse2+0x8ca>
+  DB  68,15,40,37,190,40,0,0              ; movaps        0x28be(%rip),%xmm12        # 5580 <_sk_callback_sse2+0x8da>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,46,40,0,0               ; addps         0x282e(%rip),%xmm10        # 5500 <_sk_callback_sse2+0x8e6>
-  DB  68,15,40,37,54,40,0,0               ; movaps        0x2836(%rip),%xmm12        # 5510 <_sk_callback_sse2+0x8f6>
+  DB  68,15,88,21,190,40,0,0              ; addps         0x28be(%rip),%xmm10        # 5590 <_sk_callback_sse2+0x8ea>
+  DB  68,15,40,37,198,40,0,0              ; movaps        0x28c6(%rip),%xmm12        # 55a0 <_sk_callback_sse2+0x8fa>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -20759,22 +20879,22 @@ _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  68,15,40,21,32,40,0,0               ; movaps        0x2820(%rip),%xmm10        # 5520 <_sk_callback_sse2+0x906>
+  DB  68,15,40,21,176,40,0,0              ; movaps        0x28b0(%rip),%xmm10        # 55b0 <_sk_callback_sse2+0x90a>
   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  68,15,88,13,20,40,0,0               ; addps         0x2814(%rip),%xmm9        # 5530 <_sk_callback_sse2+0x916>
-  DB  68,15,40,37,28,40,0,0               ; movaps        0x281c(%rip),%xmm12        # 5540 <_sk_callback_sse2+0x926>
+  DB  68,15,88,13,164,40,0,0              ; addps         0x28a4(%rip),%xmm9        # 55c0 <_sk_callback_sse2+0x91a>
+  DB  68,15,40,37,172,40,0,0              ; movaps        0x28ac(%rip),%xmm12        # 55d0 <_sk_callback_sse2+0x92a>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,28,40,0,0               ; movaps        0x281c(%rip),%xmm12        # 5550 <_sk_callback_sse2+0x936>
+  DB  68,15,40,37,172,40,0,0              ; movaps        0x28ac(%rip),%xmm12        # 55e0 <_sk_callback_sse2+0x93a>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,32,40,0,0               ; movaps        0x2820(%rip),%xmm13        # 5560 <_sk_callback_sse2+0x946>
+  DB  68,15,40,45,176,40,0,0              ; movaps        0x28b0(%rip),%xmm13        # 55f0 <_sk_callback_sse2+0x94a>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,32,40,0,0               ; mulps         0x2820(%rip),%xmm13        # 5570 <_sk_callback_sse2+0x956>
+  DB  68,15,89,45,176,40,0,0              ; mulps         0x28b0(%rip),%xmm13        # 5600 <_sk_callback_sse2+0x95a>
   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
@@ -20789,29 +20909,29 @@ _sk_parametric_a_sse2 LABEL PROC
 
 PUBLIC _sk_lab_to_xyz_sse2
 _sk_lab_to_xyz_sse2 LABEL PROC
-  DB  15,89,5,253,39,0,0                  ; mulps         0x27fd(%rip),%xmm0        # 5580 <_sk_callback_sse2+0x966>
-  DB  68,15,40,5,5,40,0,0                 ; movaps        0x2805(%rip),%xmm8        # 5590 <_sk_callback_sse2+0x976>
+  DB  15,89,5,141,40,0,0                  ; mulps         0x288d(%rip),%xmm0        # 5610 <_sk_callback_sse2+0x96a>
+  DB  68,15,40,5,149,40,0,0               ; movaps        0x2895(%rip),%xmm8        # 5620 <_sk_callback_sse2+0x97a>
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
-  DB  68,15,40,13,9,40,0,0                ; movaps        0x2809(%rip),%xmm9        # 55a0 <_sk_callback_sse2+0x986>
+  DB  68,15,40,13,153,40,0,0              ; movaps        0x2899(%rip),%xmm9        # 5630 <_sk_callback_sse2+0x98a>
   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  15,88,5,6,40,0,0                    ; addps         0x2806(%rip),%xmm0        # 55b0 <_sk_callback_sse2+0x996>
-  DB  15,89,5,15,40,0,0                   ; mulps         0x280f(%rip),%xmm0        # 55c0 <_sk_callback_sse2+0x9a6>
-  DB  15,89,13,24,40,0,0                  ; mulps         0x2818(%rip),%xmm1        # 55d0 <_sk_callback_sse2+0x9b6>
+  DB  15,88,5,150,40,0,0                  ; addps         0x2896(%rip),%xmm0        # 5640 <_sk_callback_sse2+0x99a>
+  DB  15,89,5,159,40,0,0                  ; mulps         0x289f(%rip),%xmm0        # 5650 <_sk_callback_sse2+0x9aa>
+  DB  15,89,13,168,40,0,0                 ; mulps         0x28a8(%rip),%xmm1        # 5660 <_sk_callback_sse2+0x9ba>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,21,30,40,0,0                  ; mulps         0x281e(%rip),%xmm2        # 55e0 <_sk_callback_sse2+0x9c6>
+  DB  15,89,21,174,40,0,0                 ; mulps         0x28ae(%rip),%xmm2        # 5670 <_sk_callback_sse2+0x9ca>
   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  15,40,21,19,40,0,0                  ; movaps        0x2813(%rip),%xmm2        # 55f0 <_sk_callback_sse2+0x9d6>
+  DB  15,40,21,163,40,0,0                 ; movaps        0x28a3(%rip),%xmm2        # 5680 <_sk_callback_sse2+0x9da>
   DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
   DB  69,15,194,196,1                     ; cmpltps       %xmm12,%xmm8
-  DB  68,15,40,21,18,40,0,0               ; movaps        0x2812(%rip),%xmm10        # 5600 <_sk_callback_sse2+0x9e6>
+  DB  68,15,40,21,162,40,0,0              ; movaps        0x28a2(%rip),%xmm10        # 5690 <_sk_callback_sse2+0x9ea>
   DB  65,15,88,202                        ; addps         %xmm10,%xmm1
-  DB  68,15,40,29,22,40,0,0               ; movaps        0x2816(%rip),%xmm11        # 5610 <_sk_callback_sse2+0x9f6>
+  DB  68,15,40,29,166,40,0,0              ; movaps        0x28a6(%rip),%xmm11        # 56a0 <_sk_callback_sse2+0x9fa>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  69,15,84,224                        ; andps         %xmm8,%xmm12
   DB  68,15,85,193                        ; andnps        %xmm1,%xmm8
@@ -20835,8 +20955,8 @@ _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  68,15,89,5,198,39,0,0               ; mulps         0x27c6(%rip),%xmm8        # 5620 <_sk_callback_sse2+0xa06>
-  DB  15,89,21,207,39,0,0                 ; mulps         0x27cf(%rip),%xmm2        # 5630 <_sk_callback_sse2+0xa16>
+  DB  68,15,89,5,86,40,0,0                ; mulps         0x2856(%rip),%xmm8        # 56b0 <_sk_callback_sse2+0xa0a>
+  DB  15,89,21,95,40,0,0                  ; mulps         0x285f(%rip),%xmm2        # 56c0 <_sk_callback_sse2+0xa1a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -20850,7 +20970,7 @@ _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  15,89,29,183,39,0,0                 ; mulps         0x27b7(%rip),%xmm3        # 5640 <_sk_callback_sse2+0xa26>
+  DB  15,89,29,71,40,0,0                  ; mulps         0x2847(%rip),%xmm3        # 56d0 <_sk_callback_sse2+0xa2a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -20893,7 +21013,7 @@ _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  15,89,29,38,39,0,0                  ; mulps         0x2726(%rip),%xmm3        # 5650 <_sk_callback_sse2+0xa36>
+  DB  15,89,29,182,39,0,0                 ; mulps         0x27b6(%rip),%xmm3        # 56e0 <_sk_callback_sse2+0xa3a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -20904,7 +21024,7 @@ 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  68,15,40,5,26,39,0,0                ; movaps        0x271a(%rip),%xmm8        # 5660 <_sk_callback_sse2+0xa46>
+  DB  68,15,40,5,170,39,0,0               ; movaps        0x27aa(%rip),%xmm8        # 56f0 <_sk_callback_sse2+0xa4a>
   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
@@ -20924,9 +21044,9 @@ _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  15,89,5,225,38,0,0                  ; mulps         0x26e1(%rip),%xmm0        # 5670 <_sk_callback_sse2+0xa56>
+  DB  15,89,5,113,39,0,0                  ; mulps         0x2771(%rip),%xmm0        # 5700 <_sk_callback_sse2+0xa5a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,232,38,0,0                 ; movaps        0x26e8(%rip),%xmm3        # 5680 <_sk_callback_sse2+0xa66>
+  DB  15,40,29,120,39,0,0                 ; movaps        0x2778(%rip),%xmm3        # 5710 <_sk_callback_sse2+0xa6a>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -20967,9 +21087,9 @@ _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  15,89,5,93,38,0,0                   ; mulps         0x265d(%rip),%xmm0        # 5690 <_sk_callback_sse2+0xa76>
+  DB  15,89,5,237,38,0,0                  ; mulps         0x26ed(%rip),%xmm0        # 5720 <_sk_callback_sse2+0xa7a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,100,38,0,0                 ; movaps        0x2664(%rip),%xmm3        # 56a0 <_sk_callback_sse2+0xa86>
+  DB  15,40,29,244,38,0,0                 ; movaps        0x26f4(%rip),%xmm3        # 5730 <_sk_callback_sse2+0xa8a>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -21030,11 +21150,11 @@ _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  102,15,111,21,131,37,0,0            ; movdqa        0x2583(%rip),%xmm2        # 56b0 <_sk_callback_sse2+0xa96>
+  DB  102,15,111,21,19,38,0,0             ; movdqa        0x2613(%rip),%xmm2        # 5740 <_sk_callback_sse2+0xa9a>
   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  68,15,40,5,127,37,0,0               ; movaps        0x257f(%rip),%xmm8        # 56c0 <_sk_callback_sse2+0xaa6>
+  DB  68,15,40,5,15,38,0,0                ; movaps        0x260f(%rip),%xmm8        # 5750 <_sk_callback_sse2+0xaaa>
   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
@@ -21059,19 +21179,19 @@ _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  102,15,111,5,53,37,0,0              ; movdqa        0x2535(%rip),%xmm0        # 56d0 <_sk_callback_sse2+0xab6>
+  DB  102,15,111,5,197,37,0,0             ; movdqa        0x25c5(%rip),%xmm0        # 5760 <_sk_callback_sse2+0xaba>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,55,37,0,0                   ; mulps         0x2537(%rip),%xmm0        # 56e0 <_sk_callback_sse2+0xac6>
-  DB  102,15,111,13,63,37,0,0             ; movdqa        0x253f(%rip),%xmm1        # 56f0 <_sk_callback_sse2+0xad6>
+  DB  15,89,5,199,37,0,0                  ; mulps         0x25c7(%rip),%xmm0        # 5770 <_sk_callback_sse2+0xaca>
+  DB  102,15,111,13,207,37,0,0            ; movdqa        0x25cf(%rip),%xmm1        # 5780 <_sk_callback_sse2+0xada>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,65,37,0,0                  ; mulps         0x2541(%rip),%xmm1        # 5700 <_sk_callback_sse2+0xae6>
-  DB  102,15,219,21,73,37,0,0             ; pand          0x2549(%rip),%xmm2        # 5710 <_sk_callback_sse2+0xaf6>
+  DB  15,89,13,209,37,0,0                 ; mulps         0x25d1(%rip),%xmm1        # 5790 <_sk_callback_sse2+0xaea>
+  DB  102,15,219,21,217,37,0,0            ; pand          0x25d9(%rip),%xmm2        # 57a0 <_sk_callback_sse2+0xafa>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,79,37,0,0                  ; mulps         0x254f(%rip),%xmm2        # 5720 <_sk_callback_sse2+0xb06>
+  DB  15,89,21,223,37,0,0                 ; mulps         0x25df(%rip),%xmm2        # 57b0 <_sk_callback_sse2+0xb0a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,86,37,0,0                  ; movaps        0x2556(%rip),%xmm3        # 5730 <_sk_callback_sse2+0xb16>
+  DB  15,40,29,230,37,0,0                 ; movaps        0x25e6(%rip),%xmm3        # 57c0 <_sk_callback_sse2+0xb1a>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_gather_565_sse2
@@ -21104,31 +21224,31 @@ _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  102,15,111,5,223,36,0,0             ; movdqa        0x24df(%rip),%xmm0        # 5740 <_sk_callback_sse2+0xb26>
+  DB  102,15,111,5,111,37,0,0             ; movdqa        0x256f(%rip),%xmm0        # 57d0 <_sk_callback_sse2+0xb2a>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,225,36,0,0                  ; mulps         0x24e1(%rip),%xmm0        # 5750 <_sk_callback_sse2+0xb36>
-  DB  102,15,111,13,233,36,0,0            ; movdqa        0x24e9(%rip),%xmm1        # 5760 <_sk_callback_sse2+0xb46>
+  DB  15,89,5,113,37,0,0                  ; mulps         0x2571(%rip),%xmm0        # 57e0 <_sk_callback_sse2+0xb3a>
+  DB  102,15,111,13,121,37,0,0            ; movdqa        0x2579(%rip),%xmm1        # 57f0 <_sk_callback_sse2+0xb4a>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,235,36,0,0                 ; mulps         0x24eb(%rip),%xmm1        # 5770 <_sk_callback_sse2+0xb56>
-  DB  102,15,219,21,243,36,0,0            ; pand          0x24f3(%rip),%xmm2        # 5780 <_sk_callback_sse2+0xb66>
+  DB  15,89,13,123,37,0,0                 ; mulps         0x257b(%rip),%xmm1        # 5800 <_sk_callback_sse2+0xb5a>
+  DB  102,15,219,21,131,37,0,0            ; pand          0x2583(%rip),%xmm2        # 5810 <_sk_callback_sse2+0xb6a>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,249,36,0,0                 ; mulps         0x24f9(%rip),%xmm2        # 5790 <_sk_callback_sse2+0xb76>
+  DB  15,89,21,137,37,0,0                 ; mulps         0x2589(%rip),%xmm2        # 5820 <_sk_callback_sse2+0xb7a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,0,37,0,0                   ; movaps        0x2500(%rip),%xmm3        # 57a0 <_sk_callback_sse2+0xb86>
+  DB  15,40,29,144,37,0,0                 ; movaps        0x2590(%rip),%xmm3        # 5830 <_sk_callback_sse2+0xb8a>
   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  68,15,40,5,1,37,0,0                 ; movaps        0x2501(%rip),%xmm8        # 57b0 <_sk_callback_sse2+0xb96>
+  DB  68,15,40,5,145,37,0,0               ; movaps        0x2591(%rip),%xmm8        # 5840 <_sk_callback_sse2+0xb9a>
   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  68,15,40,21,246,36,0,0              ; movaps        0x24f6(%rip),%xmm10        # 57c0 <_sk_callback_sse2+0xba6>
+  DB  68,15,40,21,134,37,0,0              ; movaps        0x2586(%rip),%xmm10        # 5850 <_sk_callback_sse2+0xbaa>
   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
@@ -21150,21 +21270,21 @@ _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  102,15,111,5,175,36,0,0             ; movdqa        0x24af(%rip),%xmm0        # 57d0 <_sk_callback_sse2+0xbb6>
+  DB  102,15,111,5,63,37,0,0              ; movdqa        0x253f(%rip),%xmm0        # 5860 <_sk_callback_sse2+0xbba>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,177,36,0,0                  ; mulps         0x24b1(%rip),%xmm0        # 57e0 <_sk_callback_sse2+0xbc6>
-  DB  102,15,111,13,185,36,0,0            ; movdqa        0x24b9(%rip),%xmm1        # 57f0 <_sk_callback_sse2+0xbd6>
+  DB  15,89,5,65,37,0,0                   ; mulps         0x2541(%rip),%xmm0        # 5870 <_sk_callback_sse2+0xbca>
+  DB  102,15,111,13,73,37,0,0             ; movdqa        0x2549(%rip),%xmm1        # 5880 <_sk_callback_sse2+0xbda>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,187,36,0,0                 ; mulps         0x24bb(%rip),%xmm1        # 5800 <_sk_callback_sse2+0xbe6>
-  DB  102,15,111,21,195,36,0,0            ; movdqa        0x24c3(%rip),%xmm2        # 5810 <_sk_callback_sse2+0xbf6>
+  DB  15,89,13,75,37,0,0                  ; mulps         0x254b(%rip),%xmm1        # 5890 <_sk_callback_sse2+0xbea>
+  DB  102,15,111,21,83,37,0,0             ; movdqa        0x2553(%rip),%xmm2        # 58a0 <_sk_callback_sse2+0xbfa>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,197,36,0,0                 ; mulps         0x24c5(%rip),%xmm2        # 5820 <_sk_callback_sse2+0xc06>
-  DB  102,15,219,29,205,36,0,0            ; pand          0x24cd(%rip),%xmm3        # 5830 <_sk_callback_sse2+0xc16>
+  DB  15,89,21,85,37,0,0                  ; mulps         0x2555(%rip),%xmm2        # 58b0 <_sk_callback_sse2+0xc0a>
+  DB  102,15,219,29,93,37,0,0             ; pand          0x255d(%rip),%xmm3        # 58c0 <_sk_callback_sse2+0xc1a>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,211,36,0,0                 ; mulps         0x24d3(%rip),%xmm3        # 5840 <_sk_callback_sse2+0xc26>
+  DB  15,89,29,99,37,0,0                  ; mulps         0x2563(%rip),%xmm3        # 58d0 <_sk_callback_sse2+0xc2a>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -21198,21 +21318,21 @@ _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  102,15,111,5,90,36,0,0              ; movdqa        0x245a(%rip),%xmm0        # 5850 <_sk_callback_sse2+0xc36>
+  DB  102,15,111,5,234,36,0,0             ; movdqa        0x24ea(%rip),%xmm0        # 58e0 <_sk_callback_sse2+0xc3a>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,92,36,0,0                   ; mulps         0x245c(%rip),%xmm0        # 5860 <_sk_callback_sse2+0xc46>
-  DB  102,15,111,13,100,36,0,0            ; movdqa        0x2464(%rip),%xmm1        # 5870 <_sk_callback_sse2+0xc56>
+  DB  15,89,5,236,36,0,0                  ; mulps         0x24ec(%rip),%xmm0        # 58f0 <_sk_callback_sse2+0xc4a>
+  DB  102,15,111,13,244,36,0,0            ; movdqa        0x24f4(%rip),%xmm1        # 5900 <_sk_callback_sse2+0xc5a>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,102,36,0,0                 ; mulps         0x2466(%rip),%xmm1        # 5880 <_sk_callback_sse2+0xc66>
-  DB  102,15,111,21,110,36,0,0            ; movdqa        0x246e(%rip),%xmm2        # 5890 <_sk_callback_sse2+0xc76>
+  DB  15,89,13,246,36,0,0                 ; mulps         0x24f6(%rip),%xmm1        # 5910 <_sk_callback_sse2+0xc6a>
+  DB  102,15,111,21,254,36,0,0            ; movdqa        0x24fe(%rip),%xmm2        # 5920 <_sk_callback_sse2+0xc7a>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,112,36,0,0                 ; mulps         0x2470(%rip),%xmm2        # 58a0 <_sk_callback_sse2+0xc86>
-  DB  102,15,219,29,120,36,0,0            ; pand          0x2478(%rip),%xmm3        # 58b0 <_sk_callback_sse2+0xc96>
+  DB  15,89,21,0,37,0,0                   ; mulps         0x2500(%rip),%xmm2        # 5930 <_sk_callback_sse2+0xc8a>
+  DB  102,15,219,29,8,37,0,0              ; pand          0x2508(%rip),%xmm3        # 5940 <_sk_callback_sse2+0xc9a>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,126,36,0,0                 ; mulps         0x247e(%rip),%xmm3        # 58c0 <_sk_callback_sse2+0xca6>
+  DB  15,89,29,14,37,0,0                  ; mulps         0x250e(%rip),%xmm3        # 5950 <_sk_callback_sse2+0xcaa>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -21220,7 +21340,7 @@ 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  68,15,40,5,125,36,0,0               ; movaps        0x247d(%rip),%xmm8        # 58d0 <_sk_callback_sse2+0xcb6>
+  DB  68,15,40,5,13,37,0,0                ; movaps        0x250d(%rip),%xmm8        # 5960 <_sk_callback_sse2+0xcba>
   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
@@ -21250,11 +21370,11 @@ _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  15,40,21,16,36,0,0                  ; movaps        0x2410(%rip),%xmm2        # 58e0 <_sk_callback_sse2+0xcc6>
+  DB  15,40,21,160,36,0,0                 ; movaps        0x24a0(%rip),%xmm2        # 5970 <_sk_callback_sse2+0xcca>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,84,194                           ; andps         %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,14,36,0,0                ; movaps        0x240e(%rip),%xmm8        # 58f0 <_sk_callback_sse2+0xcd6>
+  DB  68,15,40,5,158,36,0,0               ; movaps        0x249e(%rip),%xmm8        # 5980 <_sk_callback_sse2+0xcda>
   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
@@ -21301,11 +21421,11 @@ _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  102,15,111,21,95,35,0,0             ; movdqa        0x235f(%rip),%xmm2        # 5900 <_sk_callback_sse2+0xce6>
+  DB  102,15,111,21,239,35,0,0            ; movdqa        0x23ef(%rip),%xmm2        # 5990 <_sk_callback_sse2+0xcea>
   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  68,15,40,5,91,35,0,0                ; movaps        0x235b(%rip),%xmm8        # 5910 <_sk_callback_sse2+0xcf6>
+  DB  68,15,40,5,235,35,0,0               ; movaps        0x23eb(%rip),%xmm8        # 59a0 <_sk_callback_sse2+0xcfa>
   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
@@ -21327,7 +21447,7 @@ 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  68,15,40,5,30,35,0,0                ; movaps        0x231e(%rip),%xmm8        # 5920 <_sk_callback_sse2+0xd06>
+  DB  68,15,40,5,174,35,0,0               ; movaps        0x23ae(%rip),%xmm8        # 59b0 <_sk_callback_sse2+0xd0a>
   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
@@ -21364,7 +21484,7 @@ _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  102,68,15,111,13,142,34,0,0         ; movdqa        0x228e(%rip),%xmm9        # 5930 <_sk_callback_sse2+0xd16>
+  DB  102,68,15,111,13,30,35,0,0          ; movdqa        0x231e(%rip),%xmm9        # 59c0 <_sk_callback_sse2+0xd1a>
   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
@@ -21372,11 +21492,11 @@ _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  102,68,15,111,29,115,34,0,0         ; movdqa        0x2273(%rip),%xmm11        # 5940 <_sk_callback_sse2+0xd26>
+  DB  102,68,15,111,29,3,35,0,0           ; movdqa        0x2303(%rip),%xmm11        # 59d0 <_sk_callback_sse2+0xd2a>
   DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
-  DB  102,68,15,111,37,117,34,0,0         ; movdqa        0x2275(%rip),%xmm12        # 5950 <_sk_callback_sse2+0xd36>
+  DB  102,68,15,111,37,5,35,0,0           ; movdqa        0x2305(%rip),%xmm12        # 59e0 <_sk_callback_sse2+0xd3a>
   DB  102,65,15,239,204                   ; pxor          %xmm12,%xmm1
-  DB  102,15,111,29,120,34,0,0            ; movdqa        0x2278(%rip),%xmm3        # 5960 <_sk_callback_sse2+0xd46>
+  DB  102,15,111,29,8,35,0,0              ; movdqa        0x2308(%rip),%xmm3        # 59f0 <_sk_callback_sse2+0xd4a>
   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
@@ -21460,7 +21580,7 @@ _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  102,68,15,111,13,6,33,0,0           ; movdqa        0x2106(%rip),%xmm9        # 5970 <_sk_callback_sse2+0xd56>
+  DB  102,68,15,111,13,150,33,0,0         ; movdqa        0x2196(%rip),%xmm9        # 5a00 <_sk_callback_sse2+0xd5a>
   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
@@ -21468,11 +21588,11 @@ _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  102,68,15,111,29,235,32,0,0         ; movdqa        0x20eb(%rip),%xmm11        # 5980 <_sk_callback_sse2+0xd66>
+  DB  102,68,15,111,29,123,33,0,0         ; movdqa        0x217b(%rip),%xmm11        # 5a10 <_sk_callback_sse2+0xd6a>
   DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
-  DB  102,68,15,111,37,237,32,0,0         ; movdqa        0x20ed(%rip),%xmm12        # 5990 <_sk_callback_sse2+0xd76>
+  DB  102,68,15,111,37,125,33,0,0         ; movdqa        0x217d(%rip),%xmm12        # 5a20 <_sk_callback_sse2+0xd7a>
   DB  102,65,15,239,204                   ; pxor          %xmm12,%xmm1
-  DB  102,15,111,29,240,32,0,0            ; movdqa        0x20f0(%rip),%xmm3        # 59a0 <_sk_callback_sse2+0xd86>
+  DB  102,15,111,29,128,33,0,0            ; movdqa        0x2180(%rip),%xmm3        # 5a30 <_sk_callback_sse2+0xd8a>
   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
@@ -21523,17 +21643,17 @@ 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  102,68,15,111,21,24,32,0,0          ; movdqa        0x2018(%rip),%xmm10        # 59b0 <_sk_callback_sse2+0xd96>
+  DB  102,68,15,111,21,168,32,0,0         ; movdqa        0x20a8(%rip),%xmm10        # 5a40 <_sk_callback_sse2+0xd9a>
   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  102,68,15,111,13,11,32,0,0          ; movdqa        0x200b(%rip),%xmm9        # 59c0 <_sk_callback_sse2+0xda6>
+  DB  102,68,15,111,13,155,32,0,0         ; movdqa        0x209b(%rip),%xmm9        # 5a50 <_sk_callback_sse2+0xdaa>
   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  102,68,15,111,29,252,31,0,0         ; movdqa        0x1ffc(%rip),%xmm11        # 59d0 <_sk_callback_sse2+0xdb6>
+  DB  102,68,15,111,29,140,32,0,0         ; movdqa        0x208c(%rip),%xmm11        # 5a60 <_sk_callback_sse2+0xdba>
   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
@@ -21610,7 +21730,7 @@ _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  68,15,40,5,154,30,0,0               ; movaps        0x1e9a(%rip),%xmm8        # 59e0 <_sk_callback_sse2+0xdc6>
+  DB  68,15,40,5,42,31,0,0                ; movaps        0x1f2a(%rip),%xmm8        # 5a70 <_sk_callback_sse2+0xdca>
   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
@@ -21661,7 +21781,7 @@ _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  68,15,40,13,214,29,0,0              ; movaps        0x1dd6(%rip),%xmm9        # 59f0 <_sk_callback_sse2+0xdd6>
+  DB  68,15,40,13,102,30,0,0              ; movaps        0x1e66(%rip),%xmm9        # 5a80 <_sk_callback_sse2+0xdda>
   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
@@ -21678,14 +21798,14 @@ _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  15,40,29,157,29,0,0                 ; movaps        0x1d9d(%rip),%xmm3        # 5a00 <_sk_callback_sse2+0xde6>
+  DB  15,40,29,45,30,0,0                  ; movaps        0x1e2d(%rip),%xmm3        # 5a90 <_sk_callback_sse2+0xdea>
   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  68,15,40,13,158,29,0,0              ; movaps        0x1d9e(%rip),%xmm9        # 5a10 <_sk_callback_sse2+0xdf6>
+  DB  68,15,40,13,46,30,0,0               ; movaps        0x1e2e(%rip),%xmm9        # 5aa0 <_sk_callback_sse2+0xdfa>
   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
@@ -21821,7 +21941,7 @@ _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  68,15,84,13,158,27,0,0              ; andps         0x1b9e(%rip),%xmm9        # 5a20 <_sk_callback_sse2+0xe06>
+  DB  68,15,84,13,46,28,0,0               ; andps         0x1c2e(%rip),%xmm9        # 5ab0 <_sk_callback_sse2+0xe0a>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
   DB  65,15,92,194                        ; subps         %xmm10,%xmm0
@@ -21839,7 +21959,7 @@ _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  68,15,84,13,112,27,0,0              ; andps         0x1b70(%rip),%xmm9        # 5a30 <_sk_callback_sse2+0xe16>
+  DB  68,15,84,13,0,28,0,0                ; andps         0x1c00(%rip),%xmm9        # 5ac0 <_sk_callback_sse2+0xe1a>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
   DB  65,15,92,202                        ; subps         %xmm10,%xmm1
@@ -21861,7 +21981,7 @@ _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  68,15,84,21,48,27,0,0               ; andps         0x1b30(%rip),%xmm10        # 5a40 <_sk_callback_sse2+0xe26>
+  DB  68,15,84,21,192,27,0,0              ; andps         0x1bc0(%rip),%xmm10        # 5ad0 <_sk_callback_sse2+0xe2a>
   DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
@@ -21887,7 +22007,7 @@ _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  68,15,84,21,224,26,0,0              ; andps         0x1ae0(%rip),%xmm10        # 5a50 <_sk_callback_sse2+0xe36>
+  DB  68,15,84,21,112,27,0,0              ; andps         0x1b70(%rip),%xmm10        # 5ae0 <_sk_callback_sse2+0xe3a>
   DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
@@ -21899,13 +22019,54 @@ _sk_mirror_y_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_clamp_x_1_sse2
+_sk_clamp_x_1_sse2 LABEL PROC
+  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
+  DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
+  DB  68,15,93,5,76,27,0,0                ; minps         0x1b4c(%rip),%xmm8        # 5af0 <_sk_callback_sse2+0xe4a>
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_repeat_x_1_sse2
+_sk_repeat_x_1_sse2 LABEL PROC
+  DB  243,68,15,91,192                    ; cvttps2dq     %xmm0,%xmm8
+  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
+  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
+  DB  69,15,194,200,1                     ; cmpltps       %xmm8,%xmm9
+  DB  68,15,84,13,58,27,0,0               ; andps         0x1b3a(%rip),%xmm9        # 5b00 <_sk_callback_sse2+0xe5a>
+  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
+  DB  65,15,92,192                        ; subps         %xmm8,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_mirror_x_1_sse2
+_sk_mirror_x_1_sse2 LABEL PROC
+  DB  68,15,40,5,54,27,0,0                ; movaps        0x1b36(%rip),%xmm8        # 5b10 <_sk_callback_sse2+0xe6a>
+  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
+  DB  68,15,40,13,58,27,0,0               ; movaps        0x1b3a(%rip),%xmm9        # 5b20 <_sk_callback_sse2+0xe7a>
+  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
+  DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
+  DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
+  DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
+  DB  68,15,84,13,48,27,0,0               ; andps         0x1b30(%rip),%xmm9        # 5b30 <_sk_callback_sse2+0xe8a>
+  DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
+  DB  69,15,92,209                        ; subps         %xmm9,%xmm10
+  DB  69,15,88,210                        ; addps         %xmm10,%xmm10
+  DB  65,15,92,194                        ; subps         %xmm10,%xmm0
+  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
+  DB  68,15,92,216                        ; subps         %xmm0,%xmm11
+  DB  65,15,84,195                        ; andps         %xmm11,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_luminance_to_alpha_sse2
 _sk_luminance_to_alpha_sse2 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
-  DB  15,89,5,194,26,0,0                  ; mulps         0x1ac2(%rip),%xmm0        # 5a60 <_sk_callback_sse2+0xe46>
-  DB  15,89,13,203,26,0,0                 ; mulps         0x1acb(%rip),%xmm1        # 5a70 <_sk_callback_sse2+0xe56>
+  DB  15,89,5,22,27,0,0                   ; mulps         0x1b16(%rip),%xmm0        # 5b40 <_sk_callback_sse2+0xe9a>
+  DB  15,89,13,31,27,0,0                  ; mulps         0x1b1f(%rip),%xmm1        # 5b50 <_sk_callback_sse2+0xeaa>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,29,209,26,0,0                 ; mulps         0x1ad1(%rip),%xmm3        # 5a80 <_sk_callback_sse2+0xe66>
+  DB  15,89,29,37,27,0,0                  ; mulps         0x1b25(%rip),%xmm3        # 5b60 <_sk_callback_sse2+0xeba>
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -22169,9 +22330,9 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC
   DB  72,139,8                            ; mov           (%rax),%rcx
   DB  76,139,88,8                         ; mov           0x8(%rax),%r11
   DB  72,255,201                          ; dec           %rcx
-  DB  120,7                               ; js            43e8 <_sk_evenly_spaced_gradient_sse2+0x15>
+  DB  120,7                               ; js            4474 <_sk_evenly_spaced_gradient_sse2+0x15>
   DB  243,72,15,42,201                    ; cvtsi2ss      %rcx,%xmm1
-  DB  235,21                              ; jmp           43fd <_sk_evenly_spaced_gradient_sse2+0x2a>
+  DB  235,21                              ; jmp           4489 <_sk_evenly_spaced_gradient_sse2+0x2a>
   DB  73,137,200                          ; mov           %rcx,%r8
   DB  73,209,232                          ; shr           %r8
   DB  131,225,1                           ; and           $0x1,%ecx
@@ -22269,12 +22430,12 @@ _sk_gradient_sse2 LABEL PROC
   DB  76,139,0                            ; mov           (%rax),%r8
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  73,131,248,2                        ; cmp           $0x2,%r8
-  DB  114,50                              ; jb            45c0 <_sk_gradient_sse2+0x41>
+  DB  114,50                              ; jb            464c <_sk_gradient_sse2+0x41>
   DB  72,139,72,72                        ; mov           0x48(%rax),%rcx
   DB  73,255,200                          ; dec           %r8
   DB  72,131,193,4                        ; add           $0x4,%rcx
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
-  DB  15,40,21,236,20,0,0                 ; movaps        0x14ec(%rip),%xmm2        # 5a90 <_sk_callback_sse2+0xe76>
+  DB  15,40,21,64,21,0,0                  ; movaps        0x1540(%rip),%xmm2        # 5b70 <_sk_callback_sse2+0xeca>
   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
@@ -22282,7 +22443,7 @@ _sk_gradient_sse2 LABEL PROC
   DB  102,15,254,203                      ; paddd         %xmm3,%xmm1
   DB  72,131,193,4                        ; add           $0x4,%rcx
   DB  73,255,200                          ; dec           %r8
-  DB  117,228                             ; jne           45a4 <_sk_gradient_sse2+0x25>
+  DB  117,228                             ; jne           4630 <_sk_gradient_sse2+0x25>
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
@@ -22418,29 +22579,29 @@ _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  68,15,40,45,174,18,0,0              ; movaps        0x12ae(%rip),%xmm13        # 5aa0 <_sk_callback_sse2+0xe86>
+  DB  68,15,40,45,2,19,0,0                ; movaps        0x1302(%rip),%xmm13        # 5b80 <_sk_callback_sse2+0xeda>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,178,18,0,0              ; addps         0x12b2(%rip),%xmm13        # 5ab0 <_sk_callback_sse2+0xe96>
+  DB  68,15,88,45,6,19,0,0                ; addps         0x1306(%rip),%xmm13        # 5b90 <_sk_callback_sse2+0xeea>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,182,18,0,0              ; addps         0x12b6(%rip),%xmm13        # 5ac0 <_sk_callback_sse2+0xea6>
+  DB  68,15,88,45,10,19,0,0               ; addps         0x130a(%rip),%xmm13        # 5ba0 <_sk_callback_sse2+0xefa>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,186,18,0,0              ; addps         0x12ba(%rip),%xmm13        # 5ad0 <_sk_callback_sse2+0xeb6>
+  DB  68,15,88,45,14,19,0,0               ; addps         0x130e(%rip),%xmm13        # 5bb0 <_sk_callback_sse2+0xf0a>
   DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,40,21,185,18,0,0              ; movaps        0x12b9(%rip),%xmm10        # 5ae0 <_sk_callback_sse2+0xec6>
+  DB  68,15,40,21,13,19,0,0               ; movaps        0x130d(%rip),%xmm10        # 5bc0 <_sk_callback_sse2+0xf1a>
   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  68,15,40,21,172,18,0,0              ; movaps        0x12ac(%rip),%xmm10        # 5af0 <_sk_callback_sse2+0xed6>
+  DB  68,15,40,21,0,19,0,0                ; movaps        0x1300(%rip),%xmm10        # 5bd0 <_sk_callback_sse2+0xf2a>
   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  68,15,40,21,155,18,0,0              ; movaps        0x129b(%rip),%xmm10        # 5b00 <_sk_callback_sse2+0xee6>
+  DB  68,15,40,21,239,18,0,0              ; movaps        0x12ef(%rip),%xmm10        # 5be0 <_sk_callback_sse2+0xf3a>
   DB  69,15,92,208                        ; subps         %xmm8,%xmm10
   DB  69,15,84,209                        ; andps         %xmm9,%xmm10
   DB  69,15,85,200                        ; andnps        %xmm8,%xmm9
@@ -22463,7 +22624,7 @@ _sk_xy_to_radius_sse2 LABEL PROC
 PUBLIC _sk_save_xy_sse2
 _sk_save_xy_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,109,18,0,0               ; movaps        0x126d(%rip),%xmm8        # 5b10 <_sk_callback_sse2+0xef6>
+  DB  68,15,40,5,193,18,0,0               ; movaps        0x12c1(%rip),%xmm8        # 5bf0 <_sk_callback_sse2+0xf4a>
   DB  15,17,0                             ; movups        %xmm0,(%rax)
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,88,200                        ; addps         %xmm8,%xmm9
@@ -22471,7 +22632,7 @@ _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  68,15,40,37,88,18,0,0               ; movaps        0x1258(%rip),%xmm12        # 5b20 <_sk_callback_sse2+0xf06>
+  DB  68,15,40,37,172,18,0,0              ; movaps        0x12ac(%rip),%xmm12        # 5c00 <_sk_callback_sse2+0xf5a>
   DB  69,15,84,220                        ; andps         %xmm12,%xmm11
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
@@ -22514,8 +22675,8 @@ _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  15,88,5,209,17,0,0                  ; addps         0x11d1(%rip),%xmm0        # 5b30 <_sk_callback_sse2+0xf16>
-  DB  68,15,40,13,217,17,0,0              ; movaps        0x11d9(%rip),%xmm9        # 5b40 <_sk_callback_sse2+0xf26>
+  DB  15,88,5,37,18,0,0                   ; addps         0x1225(%rip),%xmm0        # 5c10 <_sk_callback_sse2+0xf6a>
+  DB  68,15,40,13,45,18,0,0               ; movaps        0x122d(%rip),%xmm9        # 5c20 <_sk_callback_sse2+0xf7a>
   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
@@ -22526,7 +22687,7 @@ _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  15,88,5,200,17,0,0                  ; addps         0x11c8(%rip),%xmm0        # 5b50 <_sk_callback_sse2+0xf36>
+  DB  15,88,5,28,18,0,0                   ; addps         0x121c(%rip),%xmm0        # 5c30 <_sk_callback_sse2+0xf8a>
   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
@@ -22536,8 +22697,8 @@ _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  15,88,13,186,17,0,0                 ; addps         0x11ba(%rip),%xmm1        # 5b60 <_sk_callback_sse2+0xf46>
-  DB  68,15,40,13,194,17,0,0              ; movaps        0x11c2(%rip),%xmm9        # 5b70 <_sk_callback_sse2+0xf56>
+  DB  15,88,13,14,18,0,0                  ; addps         0x120e(%rip),%xmm1        # 5c40 <_sk_callback_sse2+0xf9a>
+  DB  68,15,40,13,22,18,0,0               ; movaps        0x1216(%rip),%xmm9        # 5c50 <_sk_callback_sse2+0xfaa>
   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
@@ -22548,7 +22709,7 @@ _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  15,88,13,176,17,0,0                 ; addps         0x11b0(%rip),%xmm1        # 5b80 <_sk_callback_sse2+0xf66>
+  DB  15,88,13,4,18,0,0                   ; addps         0x1204(%rip),%xmm1        # 5c60 <_sk_callback_sse2+0xfba>
   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
@@ -22558,13 +22719,13 @@ _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  15,88,5,163,17,0,0                  ; addps         0x11a3(%rip),%xmm0        # 5b90 <_sk_callback_sse2+0xf76>
-  DB  68,15,40,13,171,17,0,0              ; movaps        0x11ab(%rip),%xmm9        # 5ba0 <_sk_callback_sse2+0xf86>
+  DB  15,88,5,247,17,0,0                  ; addps         0x11f7(%rip),%xmm0        # 5c70 <_sk_callback_sse2+0xfca>
+  DB  68,15,40,13,255,17,0,0              ; movaps        0x11ff(%rip),%xmm9        # 5c80 <_sk_callback_sse2+0xfda>
   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  68,15,89,13,167,17,0,0              ; mulps         0x11a7(%rip),%xmm9        # 5bb0 <_sk_callback_sse2+0xf96>
-  DB  68,15,88,13,175,17,0,0              ; addps         0x11af(%rip),%xmm9        # 5bc0 <_sk_callback_sse2+0xfa6>
+  DB  68,15,89,13,251,17,0,0              ; mulps         0x11fb(%rip),%xmm9        # 5c90 <_sk_callback_sse2+0xfea>
+  DB  68,15,88,13,3,18,0,0                ; addps         0x1203(%rip),%xmm9        # 5ca0 <_sk_callback_sse2+0xffa>
   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
@@ -22575,16 +22736,16 @@ _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  15,88,5,158,17,0,0                  ; addps         0x119e(%rip),%xmm0        # 5bd0 <_sk_callback_sse2+0xfb6>
-  DB  68,15,40,13,166,17,0,0              ; movaps        0x11a6(%rip),%xmm9        # 5be0 <_sk_callback_sse2+0xfc6>
+  DB  15,88,5,242,17,0,0                  ; addps         0x11f2(%rip),%xmm0        # 5cb0 <_sk_callback_sse2+0x100a>
+  DB  68,15,40,13,250,17,0,0              ; movaps        0x11fa(%rip),%xmm9        # 5cc0 <_sk_callback_sse2+0x101a>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,170,17,0,0               ; movaps        0x11aa(%rip),%xmm8        # 5bf0 <_sk_callback_sse2+0xfd6>
+  DB  68,15,40,5,254,17,0,0               ; movaps        0x11fe(%rip),%xmm8        # 5cd0 <_sk_callback_sse2+0x102a>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,174,17,0,0               ; addps         0x11ae(%rip),%xmm8        # 5c00 <_sk_callback_sse2+0xfe6>
+  DB  68,15,88,5,2,18,0,0                 ; addps         0x1202(%rip),%xmm8        # 5ce0 <_sk_callback_sse2+0x103a>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,178,17,0,0               ; addps         0x11b2(%rip),%xmm8        # 5c10 <_sk_callback_sse2+0xff6>
+  DB  68,15,88,5,6,18,0,0                 ; addps         0x1206(%rip),%xmm8        # 5cf0 <_sk_callback_sse2+0x104a>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,182,17,0,0               ; addps         0x11b6(%rip),%xmm8        # 5c20 <_sk_callback_sse2+0x1006>
+  DB  68,15,88,5,10,18,0,0                ; addps         0x120a(%rip),%xmm8        # 5d00 <_sk_callback_sse2+0x105a>
   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
@@ -22592,17 +22753,17 @@ _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  68,15,40,5,176,17,0,0               ; movaps        0x11b0(%rip),%xmm8        # 5c30 <_sk_callback_sse2+0x1016>
+  DB  68,15,40,5,4,18,0,0                 ; movaps        0x1204(%rip),%xmm8        # 5d10 <_sk_callback_sse2+0x106a>
   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  68,15,40,21,172,17,0,0              ; movaps        0x11ac(%rip),%xmm10        # 5c40 <_sk_callback_sse2+0x1026>
+  DB  68,15,40,21,0,18,0,0                ; movaps        0x1200(%rip),%xmm10        # 5d20 <_sk_callback_sse2+0x107a>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,176,17,0,0              ; addps         0x11b0(%rip),%xmm10        # 5c50 <_sk_callback_sse2+0x1036>
+  DB  68,15,88,21,4,18,0,0                ; addps         0x1204(%rip),%xmm10        # 5d30 <_sk_callback_sse2+0x108a>
   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  68,15,88,21,172,17,0,0              ; addps         0x11ac(%rip),%xmm10        # 5c60 <_sk_callback_sse2+0x1046>
+  DB  68,15,88,21,0,18,0,0                ; addps         0x1200(%rip),%xmm10        # 5d40 <_sk_callback_sse2+0x109a>
   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
@@ -22612,11 +22773,11 @@ _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  15,88,5,159,17,0,0                  ; addps         0x119f(%rip),%xmm0        # 5c70 <_sk_callback_sse2+0x1056>
+  DB  15,88,5,243,17,0,0                  ; addps         0x11f3(%rip),%xmm0        # 5d50 <_sk_callback_sse2+0x10aa>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,159,17,0,0               ; mulps         0x119f(%rip),%xmm8        # 5c80 <_sk_callback_sse2+0x1066>
-  DB  68,15,88,5,167,17,0,0               ; addps         0x11a7(%rip),%xmm8        # 5c90 <_sk_callback_sse2+0x1076>
+  DB  68,15,89,5,243,17,0,0               ; mulps         0x11f3(%rip),%xmm8        # 5d60 <_sk_callback_sse2+0x10ba>
+  DB  68,15,88,5,251,17,0,0               ; addps         0x11fb(%rip),%xmm8        # 5d70 <_sk_callback_sse2+0x10ca>
   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
@@ -22627,13 +22788,13 @@ _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  15,88,13,149,17,0,0                 ; addps         0x1195(%rip),%xmm1        # 5ca0 <_sk_callback_sse2+0x1086>
-  DB  68,15,40,13,157,17,0,0              ; movaps        0x119d(%rip),%xmm9        # 5cb0 <_sk_callback_sse2+0x1096>
+  DB  15,88,13,233,17,0,0                 ; addps         0x11e9(%rip),%xmm1        # 5d80 <_sk_callback_sse2+0x10da>
+  DB  68,15,40,13,241,17,0,0              ; movaps        0x11f1(%rip),%xmm9        # 5d90 <_sk_callback_sse2+0x10ea>
   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  68,15,89,13,153,17,0,0              ; mulps         0x1199(%rip),%xmm9        # 5cc0 <_sk_callback_sse2+0x10a6>
-  DB  68,15,88,13,161,17,0,0              ; addps         0x11a1(%rip),%xmm9        # 5cd0 <_sk_callback_sse2+0x10b6>
+  DB  68,15,89,13,237,17,0,0              ; mulps         0x11ed(%rip),%xmm9        # 5da0 <_sk_callback_sse2+0x10fa>
+  DB  68,15,88,13,245,17,0,0              ; addps         0x11f5(%rip),%xmm9        # 5db0 <_sk_callback_sse2+0x110a>
   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
@@ -22644,16 +22805,16 @@ _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  15,88,13,143,17,0,0                 ; addps         0x118f(%rip),%xmm1        # 5ce0 <_sk_callback_sse2+0x10c6>
-  DB  68,15,40,13,151,17,0,0              ; movaps        0x1197(%rip),%xmm9        # 5cf0 <_sk_callback_sse2+0x10d6>
+  DB  15,88,13,227,17,0,0                 ; addps         0x11e3(%rip),%xmm1        # 5dc0 <_sk_callback_sse2+0x111a>
+  DB  68,15,40,13,235,17,0,0              ; movaps        0x11eb(%rip),%xmm9        # 5dd0 <_sk_callback_sse2+0x112a>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,155,17,0,0               ; movaps        0x119b(%rip),%xmm8        # 5d00 <_sk_callback_sse2+0x10e6>
+  DB  68,15,40,5,239,17,0,0               ; movaps        0x11ef(%rip),%xmm8        # 5de0 <_sk_callback_sse2+0x113a>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,159,17,0,0               ; addps         0x119f(%rip),%xmm8        # 5d10 <_sk_callback_sse2+0x10f6>
+  DB  68,15,88,5,243,17,0,0               ; addps         0x11f3(%rip),%xmm8        # 5df0 <_sk_callback_sse2+0x114a>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,163,17,0,0               ; addps         0x11a3(%rip),%xmm8        # 5d20 <_sk_callback_sse2+0x1106>
+  DB  68,15,88,5,247,17,0,0               ; addps         0x11f7(%rip),%xmm8        # 5e00 <_sk_callback_sse2+0x115a>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,167,17,0,0               ; addps         0x11a7(%rip),%xmm8        # 5d30 <_sk_callback_sse2+0x1116>
+  DB  68,15,88,5,251,17,0,0               ; addps         0x11fb(%rip),%xmm8        # 5e10 <_sk_callback_sse2+0x116a>
   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
@@ -22661,17 +22822,17 @@ _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  68,15,40,5,161,17,0,0               ; movaps        0x11a1(%rip),%xmm8        # 5d40 <_sk_callback_sse2+0x1126>
+  DB  68,15,40,5,245,17,0,0               ; movaps        0x11f5(%rip),%xmm8        # 5e20 <_sk_callback_sse2+0x117a>
   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  68,15,40,21,156,17,0,0              ; movaps        0x119c(%rip),%xmm10        # 5d50 <_sk_callback_sse2+0x1136>
+  DB  68,15,40,21,240,17,0,0              ; movaps        0x11f0(%rip),%xmm10        # 5e30 <_sk_callback_sse2+0x118a>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,160,17,0,0              ; addps         0x11a0(%rip),%xmm10        # 5d60 <_sk_callback_sse2+0x1146>
+  DB  68,15,88,21,244,17,0,0              ; addps         0x11f4(%rip),%xmm10        # 5e40 <_sk_callback_sse2+0x119a>
   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  68,15,88,21,156,17,0,0              ; addps         0x119c(%rip),%xmm10        # 5d70 <_sk_callback_sse2+0x1156>
+  DB  68,15,88,21,240,17,0,0              ; addps         0x11f0(%rip),%xmm10        # 5e50 <_sk_callback_sse2+0x11aa>
   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
@@ -22681,11 +22842,11 @@ _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  15,88,13,142,17,0,0                 ; addps         0x118e(%rip),%xmm1        # 5d80 <_sk_callback_sse2+0x1166>
+  DB  15,88,13,226,17,0,0                 ; addps         0x11e2(%rip),%xmm1        # 5e60 <_sk_callback_sse2+0x11ba>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,142,17,0,0               ; mulps         0x118e(%rip),%xmm8        # 5d90 <_sk_callback_sse2+0x1176>
-  DB  68,15,88,5,150,17,0,0               ; addps         0x1196(%rip),%xmm8        # 5da0 <_sk_callback_sse2+0x1186>
+  DB  68,15,89,5,226,17,0,0               ; mulps         0x11e2(%rip),%xmm8        # 5e70 <_sk_callback_sse2+0x11ca>
+  DB  68,15,88,5,234,17,0,0               ; addps         0x11ea(%rip),%xmm8        # 5e80 <_sk_callback_sse2+0x11da>
   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
@@ -22890,11 +23051,11 @@ ALIGN 16
   DB  128,191,0,0,128,191,0               ; cmpb          $0x0,-0x40800000(%rdi)
   DB  0,224                               ; add           %ah,%al
   DB  64,0,0                              ; add           %al,(%rax)
-  DB  224,64                              ; loopne        4ea8 <.literal16+0x1d8>
+  DB  224,64                              ; loopne        4f38 <.literal16+0x1d8>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        4eac <.literal16+0x1dc>
+  DB  224,64                              ; loopne        4f3c <.literal16+0x1dc>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        4eb0 <.literal16+0x1e0>
+  DB  224,64                              ; loopne        4f40 <.literal16+0x1e0>
   DB  154                                 ; (bad)
   DB  153                                 ; cltd
   DB  153                                 ; cltd
@@ -22914,13 +23075,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4ed1 <.literal16+0x201>
+  DB  71,225,61                           ; rex.RXB       loope 4f61 <.literal16+0x201>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4ed5 <.literal16+0x205>
+  DB  71,225,61                           ; rex.RXB       loope 4f65 <.literal16+0x205>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4ed9 <.literal16+0x209>
+  DB  71,225,61                           ; rex.RXB       loope 4f69 <.literal16+0x209>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4edd <.literal16+0x20d>
+  DB  71,225,61                           ; rex.RXB       loope 4f6d <.literal16+0x20d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -22945,13 +23106,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f11 <.literal16+0x241>
+  DB  71,225,61                           ; rex.RXB       loope 4fa1 <.literal16+0x241>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f15 <.literal16+0x245>
+  DB  71,225,61                           ; rex.RXB       loope 4fa5 <.literal16+0x245>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f19 <.literal16+0x249>
+  DB  71,225,61                           ; rex.RXB       loope 4fa9 <.literal16+0x249>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f1d <.literal16+0x24d>
+  DB  71,225,61                           ; rex.RXB       loope 4fad <.literal16+0x24d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -22976,13 +23137,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f51 <.literal16+0x281>
+  DB  71,225,61                           ; rex.RXB       loope 4fe1 <.literal16+0x281>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f55 <.literal16+0x285>
+  DB  71,225,61                           ; rex.RXB       loope 4fe5 <.literal16+0x285>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f59 <.literal16+0x289>
+  DB  71,225,61                           ; rex.RXB       loope 4fe9 <.literal16+0x289>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f5d <.literal16+0x28d>
+  DB  71,225,61                           ; rex.RXB       loope 4fed <.literal16+0x28d>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -23007,13 +23168,13 @@ ALIGN 16
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f91 <.literal16+0x2c1>
+  DB  71,225,61                           ; rex.RXB       loope 5021 <.literal16+0x2c1>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f95 <.literal16+0x2c5>
+  DB  71,225,61                           ; rex.RXB       loope 5025 <.literal16+0x2c5>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f99 <.literal16+0x2c9>
+  DB  71,225,61                           ; rex.RXB       loope 5029 <.literal16+0x2c9>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 4f9d <.literal16+0x2cd>
+  DB  71,225,61                           ; rex.RXB       loope 502d <.literal16+0x2cd>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -23234,13 +23395,13 @@ ALIGN 16
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        5179 <.literal16+0x4a9>
+  DB  224,7                               ; loopne        5209 <.literal16+0x4a9>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        517d <.literal16+0x4ad>
+  DB  224,7                               ; loopne        520d <.literal16+0x4ad>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5181 <.literal16+0x4b1>
+  DB  224,7                               ; loopne        5211 <.literal16+0x4b1>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5185 <.literal16+0x4b5>
+  DB  224,7                               ; loopne        5215 <.literal16+0x4b5>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -23305,11 +23466,11 @@ ALIGN 16
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,127,67                            ; add           %bh,0x43(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            525b <.literal16+0x58b>
+  DB  127,67                              ; jg            52eb <.literal16+0x58b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            525f <.literal16+0x58f>
+  DB  127,67                              ; jg            52ef <.literal16+0x58f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            5263 <.literal16+0x593>
+  DB  127,67                              ; jg            52f3 <.literal16+0x593>
   DB  129,128,128,59,129,128,128,59,129,128; addl          $0x80813b80,-0x7f7ec480(%rax)
   DB  128,59,129                          ; cmpb          $0x81,(%rbx)
   DB  128,128,59,129,128,128,59           ; addb          $0x3b,-0x7f7f7ec5(%rax)
@@ -23324,16 +23485,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5254 <.literal16+0x584>
+  DB  127,0                               ; jg            52e4 <.literal16+0x584>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5258 <.literal16+0x588>
+  DB  127,0                               ; jg            52e8 <.literal16+0x588>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            525c <.literal16+0x58c>
+  DB  127,0                               ; jg            52ec <.literal16+0x58c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5260 <.literal16+0x590>
+  DB  127,0                               ; jg            52f0 <.literal16+0x590>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -23342,7 +23503,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            52e5 <.literal16+0x615>
+  DB  119,115                             ; ja            5375 <.literal16+0x615>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -23353,7 +23514,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           5249 <.literal16+0x579>
+  DB  117,191                             ; jne           52d9 <.literal16+0x579>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -23365,7 +23526,7 @@ 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          ffffffffe9a3928a <_sk_callback_sse2+0xffffffffe9a34670>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3931a <_sk_callback_sse2+0xffffffffe9a34674>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23419,16 +23580,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5324 <.literal16+0x654>
+  DB  127,0                               ; jg            53b4 <.literal16+0x654>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5328 <.literal16+0x658>
+  DB  127,0                               ; jg            53b8 <.literal16+0x658>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            532c <.literal16+0x65c>
+  DB  127,0                               ; jg            53bc <.literal16+0x65c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5330 <.literal16+0x660>
+  DB  127,0                               ; jg            53c0 <.literal16+0x660>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -23437,7 +23598,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            53b5 <.literal16+0x6e5>
+  DB  119,115                             ; ja            5445 <.literal16+0x6e5>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -23448,7 +23609,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           5319 <.literal16+0x649>
+  DB  117,191                             ; jne           53a9 <.literal16+0x649>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -23460,7 +23621,7 @@ 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          ffffffffe9a3935a <_sk_callback_sse2+0xffffffffe9a34740>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a393ea <_sk_callback_sse2+0xffffffffe9a34744>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23514,16 +23675,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            53f4 <.literal16+0x724>
+  DB  127,0                               ; jg            5484 <.literal16+0x724>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            53f8 <.literal16+0x728>
+  DB  127,0                               ; jg            5488 <.literal16+0x728>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            53fc <.literal16+0x72c>
+  DB  127,0                               ; jg            548c <.literal16+0x72c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            5400 <.literal16+0x730>
+  DB  127,0                               ; jg            5490 <.literal16+0x730>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -23532,7 +23693,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            5485 <.literal16+0x7b5>
+  DB  119,115                             ; ja            5515 <.literal16+0x7b5>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -23543,7 +23704,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           53e9 <.literal16+0x719>
+  DB  117,191                             ; jne           5479 <.literal16+0x719>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -23555,7 +23716,7 @@ 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          ffffffffe9a3942a <_sk_callback_sse2+0xffffffffe9a34810>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a394ba <_sk_callback_sse2+0xffffffffe9a34814>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23609,16 +23770,16 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            54c4 <.literal16+0x7f4>
+  DB  127,0                               ; jg            5554 <.literal16+0x7f4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            54c8 <.literal16+0x7f8>
+  DB  127,0                               ; jg            5558 <.literal16+0x7f8>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            54cc <.literal16+0x7fc>
+  DB  127,0                               ; jg            555c <.literal16+0x7fc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            54d0 <.literal16+0x800>
+  DB  127,0                               ; jg            5560 <.literal16+0x800>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -23627,7 +23788,7 @@ ALIGN 16
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,63                                ; add           %bh,(%rdi)
-  DB  119,115                             ; ja            5555 <.literal16+0x885>
+  DB  119,115                             ; ja            55e5 <.literal16+0x885>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -23638,7 +23799,7 @@ ALIGN 16
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           54b9 <.literal16+0x7e9>
+  DB  117,191                             ; jne           5549 <.literal16+0x7e9>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -23650,7 +23811,7 @@ 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          ffffffffe9a394fa <_sk_callback_sse2+0xffffffffe9a348e0>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a3958a <_sk_callback_sse2+0xffffffffe9a348e4>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
@@ -23700,13 +23861,13 @@ ALIGN 16
   DB  200,66,0,0                          ; enterq        $0x42,$0x0
   DB  200,66,0,0                          ; enterq        $0x42,$0x0
   DB  200,66,0,0                          ; enterq        $0x42,$0x0
-  DB  127,67                              ; jg            55d7 <.literal16+0x907>
+  DB  127,67                              ; jg            5667 <.literal16+0x907>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            55db <.literal16+0x90b>
+  DB  127,67                              ; jg            566b <.literal16+0x90b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            55df <.literal16+0x90f>
+  DB  127,67                              ; jg            566f <.literal16+0x90f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            55e3 <.literal16+0x913>
+  DB  127,67                              ; jg            5673 <.literal16+0x913>
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,195                               ; add           %al,%bl
   DB  0,0                                 ; add           %al,(%rax)
@@ -23753,16 +23914,16 @@ ALIGN 16
   DB  128,3,62                            ; addb          $0x3e,(%rbx)
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           5663 <.literal16+0x993>
+  DB  118,63                              ; jbe           56f3 <.literal16+0x993>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           5667 <.literal16+0x997>
+  DB  118,63                              ; jbe           56f7 <.literal16+0x997>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           566b <.literal16+0x99b>
+  DB  118,63                              ; jbe           56fb <.literal16+0x99b>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           566f <.literal16+0x99f>
+  DB  118,63                              ; jbe           56ff <.literal16+0x99f>
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
@@ -23774,11 +23935,11 @@ ALIGN 16
   DB  128,59,0                            ; cmpb          $0x0,(%rbx)
   DB  0,127,67                            ; add           %bh,0x43(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            56ab <.literal16+0x9db>
+  DB  127,67                              ; jg            573b <.literal16+0x9db>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            56af <.literal16+0x9df>
+  DB  127,67                              ; jg            573f <.literal16+0x9df>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            56b3 <.literal16+0x9e3>
+  DB  127,67                              ; jg            5743 <.literal16+0x9e3>
   DB  129,128,128,59,129,128,128,59,129,128; addl          $0x80813b80,-0x7f7ec480(%rax)
   DB  128,59,129                          ; cmpb          $0x81,(%rbx)
   DB  128,128,59,0,0,128,63               ; addb          $0x3f,-0x7fffffc5(%rax)
@@ -23818,13 +23979,13 @@ ALIGN 16
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        56f9 <.literal16+0xa29>
+  DB  224,7                               ; loopne        5789 <.literal16+0xa29>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        56fd <.literal16+0xa2d>
+  DB  224,7                               ; loopne        578d <.literal16+0xa2d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5701 <.literal16+0xa31>
+  DB  224,7                               ; loopne        5791 <.literal16+0xa31>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5705 <.literal16+0xa35>
+  DB  224,7                               ; loopne        5795 <.literal16+0xa35>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -23870,13 +24031,13 @@ ALIGN 16
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        5769 <.literal16+0xa99>
+  DB  224,7                               ; loopne        57f9 <.literal16+0xa99>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        576d <.literal16+0xa9d>
+  DB  224,7                               ; loopne        57fd <.literal16+0xa9d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5771 <.literal16+0xaa1>
+  DB  224,7                               ; loopne        5801 <.literal16+0xaa1>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        5775 <.literal16+0xaa5>
+  DB  224,7                               ; loopne        5805 <.literal16+0xaa5>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -23914,13 +24075,13 @@ ALIGN 16
   DB  65,0,0                              ; add           %al,(%r8)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            5806 <.literal16+0xb36>
+  DB  124,66                              ; jl            5896 <.literal16+0xb36>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            580a <.literal16+0xb3a>
+  DB  124,66                              ; jl            589a <.literal16+0xb3a>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            580e <.literal16+0xb3e>
+  DB  124,66                              ; jl            589e <.literal16+0xb3e>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            5812 <.literal16+0xb42>
+  DB  124,66                              ; jl            58a2 <.literal16+0xb42>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,240                               ; add           %dh,%al
@@ -24010,13 +24171,13 @@ ALIGN 16
   DB  136,136,61,137,136,136              ; mov           %cl,-0x777776c3(%rax)
   DB  61,137,136,136,61                   ; cmp           $0x3d888889,%eax
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            5915 <.literal16+0xc45>
+  DB  112,65                              ; jo            59a5 <.literal16+0xc45>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            5919 <.literal16+0xc49>
+  DB  112,65                              ; jo            59a9 <.literal16+0xc49>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            591d <.literal16+0xc4d>
+  DB  112,65                              ; jo            59ad <.literal16+0xc4d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            5921 <.literal16+0xc51>
+  DB  112,65                              ; jo            59b1 <.literal16+0xc51>
   DB  255,0                               ; incl          (%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  255,0                               ; incl          (%rax)
@@ -24038,11 +24199,11 @@ ALIGN 16
   DB  128,59,129                          ; cmpb          $0x81,(%rbx)
   DB  128,128,59,0,0,127,67               ; addb          $0x43,0x7f00003b(%rax)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            596b <.literal16+0xc9b>
+  DB  127,67                              ; jg            59fb <.literal16+0xc9b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            596f <.literal16+0xc9f>
+  DB  127,67                              ; jg            59ff <.literal16+0xc9f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            5973 <.literal16+0xca3>
+  DB  127,67                              ; jg            5a03 <.literal16+0xca3>
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%rax)
@@ -24118,13 +24279,13 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  255                                 ; (bad)
-  DB  127,71                              ; jg            5a5b <.literal16+0xd8b>
+  DB  127,71                              ; jg            5aeb <.literal16+0xd8b>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            5a5f <.literal16+0xd8f>
+  DB  127,71                              ; jg            5aef <.literal16+0xd8f>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            5a63 <.literal16+0xd93>
+  DB  127,71                              ; jg            5af3 <.literal16+0xd93>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            5a67 <.literal16+0xd97>
+  DB  127,71                              ; jg            5af7 <.literal16+0xd97>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,0                            ; cmpb          $0x0,(%rdi)
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
@@ -24146,6 +24307,32 @@ ALIGN 16
   DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
   DB  63                                  ; (bad)
   DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
+  DB  63                                  ; (bad)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
+  DB  63                                  ; (bad)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
+  DB  63                                  ; (bad)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,191,0,0,128,191,0               ; cmpb          $0x0,-0x40800000(%rdi)
+  DB  0,128,191,0,0,128                   ; add           %al,-0x7fffff41(%rax)
+  DB  191,0,0,0,63                        ; mov           $0x3f000000,%edi
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,63                                ; add           %bh,(%rdi)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,63                                ; add           %bh,(%rdi)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,63                                ; add           %bh,(%rdi)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  128,63,0                            ; cmpb          $0x0,(%rdi)
+  DB  0,128,63,0,0,128                    ; add           %al,-0x7fffffc1(%rax)
+  DB  63                                  ; (bad)
+  DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,208                          ; cmpb          $0xd0,(%rdi)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,208                              ; ds            (bad)
@@ -24285,11 +24472,11 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,114                          ; cmpb          $0x72,(%rdi)
   DB  28,199                              ; sbb           $0xc7,%al
-  DB  62,114,28                           ; jb,pt         5bd2 <.literal16+0xf02>
+  DB  62,114,28                           ; jb,pt         5cb2 <.literal16+0xf52>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5bd6 <.literal16+0xf06>
+  DB  62,114,28                           ; jb,pt         5cb6 <.literal16+0xf56>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5bda <.literal16+0xf0a>
+  DB  62,114,28                           ; jb,pt         5cba <.literal16+0xf5a>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -24333,7 +24520,7 @@ 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        # 3d63ea65 <_sk_callback_sse2+0x3d639e4b>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63eb45 <_sk_callback_sse2+0x3d639e9f>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -24359,7 +24546,7 @@ 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        # 3d63eaa5 <_sk_callback_sse2+0x3d639e8b>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63eb85 <_sk_callback_sse2+0x3d639edf>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -24368,13 +24555,13 @@ ALIGN 16
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
-  DB  114,28                              ; jb            5c9e <.literal16+0xfce>
+  DB  114,28                              ; jb            5d7e <.literal16+0x101e>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5ca2 <.literal16+0xfd2>
+  DB  62,114,28                           ; jb,pt         5d82 <.literal16+0x1022>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5ca6 <.literal16+0xfd6>
+  DB  62,114,28                           ; jb,pt         5d86 <.literal16+0x1026>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5caa <.literal16+0xfda>
+  DB  62,114,28                           ; jb,pt         5d8a <.literal16+0x102a>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -24395,11 +24582,11 @@ ALIGN 16
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,63,114                          ; cmpb          $0x72,(%rdi)
   DB  28,199                              ; sbb           $0xc7,%al
-  DB  62,114,28                           ; jb,pt         5ce2 <.literal16+0x1012>
+  DB  62,114,28                           ; jb,pt         5dc2 <.literal16+0x1062>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5ce6 <.literal16+0x1016>
+  DB  62,114,28                           ; jb,pt         5dc6 <.literal16+0x1066>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5cea <.literal16+0x101a>
+  DB  62,114,28                           ; jb,pt         5dca <.literal16+0x106a>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
@@ -24443,7 +24630,7 @@ 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        # 3d63eb75 <_sk_callback_sse2+0x3d639f5b>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63ec55 <_sk_callback_sse2+0x3d639faf>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
@@ -24469,7 +24656,7 @@ 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        # 3d63ebb5 <_sk_callback_sse2+0x3d639f9b>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d63ec95 <_sk_callback_sse2+0x3d639fef>
   DB  57,142,99,61,0,0                    ; cmp           %ecx,0x3d63(%rsi)
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
@@ -24478,13 +24665,13 @@ ALIGN 16
   DB  192,63,0                            ; sarb          $0x0,(%rdi)
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
-  DB  114,28                              ; jb            5dae <.literal16+0x10de>
+  DB  114,28                              ; jb            5e8e <.literal16+0x112e>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5db2 <_sk_callback_sse2+0x1198>
+  DB  62,114,28                           ; jb,pt         5e92 <_sk_callback_sse2+0x11ec>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5db6 <_sk_callback_sse2+0x119c>
+  DB  62,114,28                           ; jb,pt         5e96 <_sk_callback_sse2+0x11f0>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         5dba <_sk_callback_sse2+0x11a0>
+  DB  62,114,28                           ; jb,pt         5e9a <_sk_callback_sse2+0x11f4>
   DB  199                                 ; (bad)
   DB  62,171                              ; ds            stos %eax,%es:(%rdi)
   DB  170                                 ; stos          %al,%es:(%rdi)
index fa11869..5909931 100644 (file)
@@ -1012,6 +1012,10 @@ STAGE(repeat_y) { g = repeat(g, *(const float*)ctx); }
 STAGE(mirror_x) { r = mirror(r, *(const float*)ctx); }
 STAGE(mirror_y) { g = mirror(g, *(const float*)ctx); }
 
+STAGE( clamp_x_1) { r = min(max(0, r), 1); }
+STAGE(repeat_x_1) { r = r - floor_(r); }
+STAGE(mirror_x_1) { r = abs_( (r-1.0f) - two(floor_((r-1.0f)*0.5f)) - 1.0f ); }
+
 STAGE(luminance_to_alpha) {
     a = r*0.2126f + g*0.7152f + b*0.0722f;
     r = g = b = 0;