aboutsummaryrefslogtreecommitdiffhomepage
diff options
context:
space:
mode:
authorGravatar Mike Klein <mtklein@chromium.org>2017-05-04 12:42:52 -0400
committerGravatar Skia Commit-Bot <skia-commit-bot@chromium.org>2017-05-08 15:24:01 +0000
commitbb33833ed25c30007e4ea3cd3de6df728407f94e (patch)
treef4d3df11e408b8131c9472a7499dcc6de6c12801
parent05814de6ba5087ad71f189d6413246ef1d518e4b (diff)
jumper, finish blend modes
I've decided to ignore our existing CPU implementations and start from scratch, mostly referencing the GL ES 3.2 spec and w3 spec. This implementation ought to look a lot like the reference implementation I've written in gm/hsl.cpp, with the addition of handling alpha: unpremul, blend, re-premul with a simple SrcOver alpha. Change-Id: I38cf6be2dc66a6f46d7b18b91847f6933d2fab62 Reviewed-on: https://skia-review.googlesource.com/15316 Reviewed-by: Herb Derby <herb@google.com> Commit-Queue: Mike Klein <mtklein@chromium.org>
-rw-r--r--src/core/SkBlendModePriv.h2
-rw-r--r--src/core/SkModeColorFilter.cpp6
-rw-r--r--src/core/SkRasterPipeline.h1
-rw-r--r--src/core/SkRasterPipelineBlitter.cpp4
-rw-r--r--src/core/SkXfermode.cpp17
-rw-r--r--src/jumper/SkJumper_generated.S6161
-rw-r--r--src/jumper/SkJumper_generated_win.S5250
-rw-r--r--src/jumper/SkJumper_stages.cpp110
8 files changed, 8447 insertions, 3104 deletions
diff --git a/src/core/SkBlendModePriv.h b/src/core/SkBlendModePriv.h
index 0d0589c29c..887c0dc482 100644
--- a/src/core/SkBlendModePriv.h
+++ b/src/core/SkBlendModePriv.h
@@ -13,7 +13,7 @@
bool SkBlendMode_SupportsCoverageAsAlpha(SkBlendMode);
bool SkBlendMode_CanOverflow(SkBlendMode);
-bool SkBlendMode_AppendStages(SkBlendMode, SkRasterPipeline* = nullptr);
+void SkBlendMode_AppendStages(SkBlendMode, SkRasterPipeline*);
#if SK_SUPPORT_GPU
#include "GrXferProcessor.h"
diff --git a/src/core/SkModeColorFilter.cpp b/src/core/SkModeColorFilter.cpp
index a5ec5e08ac..33d7977592 100644
--- a/src/core/SkModeColorFilter.cpp
+++ b/src/core/SkModeColorFilter.cpp
@@ -98,10 +98,10 @@ bool SkModeColorFilter::onAppendStages(SkRasterPipeline* p,
p->append(SkRasterPipeline::move_src_dst);
p->append(SkRasterPipeline::constant_color, color);
auto mode = (SkBlendMode)fMode;
- if (!SkBlendMode_AppendStages(mode, p)) {
- return false;
+ SkBlendMode_AppendStages(mode, p);
+ if (SkBlendMode_CanOverflow(mode)) {
+ p->append(SkRasterPipeline::clamp_a);
}
- if (SkBlendMode_CanOverflow(mode)) { p->append(SkRasterPipeline::clamp_a); }
return true;
}
diff --git a/src/core/SkRasterPipeline.h b/src/core/SkRasterPipeline.h
index 361e833621..1099777197 100644
--- a/src/core/SkRasterPipeline.h
+++ b/src/core/SkRasterPipeline.h
@@ -80,6 +80,7 @@
M(clear) M(modulate) M(multiply) M(plus_) M(screen) M(xor_) \
M(colorburn) M(colordodge) M(darken) M(difference) \
M(exclusion) M(hardlight) M(lighten) M(overlay) M(softlight) \
+ M(hue) M(saturation) M(color) M(luminosity) \
M(luminance_to_alpha) \
M(matrix_2x3) M(matrix_3x4) M(matrix_4x5) \
M(matrix_perspective) \
diff --git a/src/core/SkRasterPipelineBlitter.cpp b/src/core/SkRasterPipelineBlitter.cpp
index d18decd1cd..57a377bb19 100644
--- a/src/core/SkRasterPipelineBlitter.cpp
+++ b/src/core/SkRasterPipelineBlitter.cpp
@@ -104,7 +104,7 @@ SkBlitter* SkRasterPipelineBlitter::Create(const SkPixmap& dst,
SkColorFilter* colorFilter = paint.getColorFilter();
// TODO: all temporary
- if (!supported(dst.info()) || !SkBlendMode_AppendStages(*blend)) {
+ if (!supported(dst.info())) {
return nullptr;
}
@@ -219,7 +219,7 @@ void SkRasterPipelineBlitter::append_store(SkRasterPipeline* p) const {
}
void SkRasterPipelineBlitter::append_blend(SkRasterPipeline* p) const {
- SkAssertResult(SkBlendMode_AppendStages(fBlend, p));
+ SkBlendMode_AppendStages(fBlend, p);
}
void SkRasterPipelineBlitter::maybe_clamp(SkRasterPipeline* p) const {
diff --git a/src/core/SkXfermode.cpp b/src/core/SkXfermode.cpp
index 6f470f6dbe..11be44b3e5 100644
--- a/src/core/SkXfermode.cpp
+++ b/src/core/SkXfermode.cpp
@@ -1484,11 +1484,11 @@ const GrXPFactory* SkBlendMode_AsXPFactory(SkBlendMode mode) {
bool SkBlendMode_CanOverflow(SkBlendMode mode) { return mode == SkBlendMode::kPlus; }
-bool SkBlendMode_AppendStages(SkBlendMode mode, SkRasterPipeline* p) {
+void SkBlendMode_AppendStages(SkBlendMode mode, SkRasterPipeline* p) {
auto stage = SkRasterPipeline::srcover;
switch (mode) {
case SkBlendMode::kClear: stage = SkRasterPipeline::clear; break;
- case SkBlendMode::kSrc: return true; // This stage is a no-op.
+ case SkBlendMode::kSrc: return; // This stage is a no-op.
case SkBlendMode::kDst: stage = SkRasterPipeline::move_dst_src; break;
case SkBlendMode::kSrcOver: stage = SkRasterPipeline::srcover; break;
case SkBlendMode::kDstOver: stage = SkRasterPipeline::dstover; break;
@@ -1514,13 +1514,10 @@ bool SkBlendMode_AppendStages(SkBlendMode mode, SkRasterPipeline* p) {
case SkBlendMode::kExclusion: stage = SkRasterPipeline::exclusion; break;
case SkBlendMode::kMultiply: stage = SkRasterPipeline::multiply; break;
- case SkBlendMode::kHue:
- case SkBlendMode::kSaturation:
- case SkBlendMode::kColor:
- case SkBlendMode::kLuminosity: return false; // TODO
+ case SkBlendMode::kHue: stage = SkRasterPipeline::hue; break;
+ case SkBlendMode::kSaturation: stage = SkRasterPipeline::saturation; break;
+ case SkBlendMode::kColor: stage = SkRasterPipeline::color; break;
+ case SkBlendMode::kLuminosity: stage = SkRasterPipeline::luminosity; break;
}
- if (p) {
- p->append(stage);
- }
- return true;
+ p->append(stage);
}
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S
index 72860b734b..fce6fe9cf7 100644
--- a/src/jumper/SkJumper_generated.S
+++ b/src/jumper/SkJumper_generated.S
@@ -783,6 +783,408 @@ _sk_softlight_aarch64:
.long 0x4e27ce43 // fmla v3.4s, v18.4s, v7.4s
.long 0xd61f0060 // br x3
+HIDDEN _sk_hue_aarch64
+.globl _sk_hue_aarch64
+FUNCTION(_sk_hue_aarch64)
+_sk_hue_aarch64:
+ .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
+ .long 0x4ea0d873 // fcmeq v19.4s, v3.4s, #0.0
+ .long 0x6e23fe58 // fdiv v24.4s, v18.4s, v3.4s
+ .long 0x4e731f13 // bic v19.16b, v24.16b, v19.16b
+ .long 0x6e21de61 // fmul v1.4s, v19.4s, v1.4s
+ .long 0x6e22de62 // fmul v2.4s, v19.4s, v2.4s
+ .long 0x4e26f4b5 // fmax v21.4s, v5.4s, v6.4s
+ .long 0x4ea6f4b6 // fmin v22.4s, v5.4s, v6.4s
+ .long 0x6e20de60 // fmul v0.4s, v19.4s, v0.4s
+ .long 0x4ea2f433 // fmin v19.4s, v1.4s, v2.4s
+ .long 0x52a7d328 // mov w8, #0x3e990000
+ .long 0x4e35f495 // fmax v21.4s, v4.4s, v21.4s
+ .long 0x4eb6f496 // fmin v22.4s, v4.4s, v22.4s
+ .long 0x4e22f438 // fmax v24.4s, v1.4s, v2.4s
+ .long 0x4eb3f413 // fmin v19.4s, v0.4s, v19.4s
+ .long 0x72933348 // movk w8, #0x999a
+ .long 0x4eb6d6b5 // fsub v21.4s, v21.4s, v22.4s
+ .long 0x4e38f418 // fmax v24.4s, v0.4s, v24.4s
+ .long 0x4eb3d421 // fsub v1.4s, v1.4s, v19.4s
+ .long 0x4e040d14 // dup v20.4s, w8
+ .long 0x52a7e2e8 // mov w8, #0x3f170000
+ .long 0x4eb3d718 // fsub v24.4s, v24.4s, v19.4s
+ .long 0x4eb3d400 // fsub v0.4s, v0.4s, v19.4s
+ .long 0x4eb3d442 // fsub v2.4s, v2.4s, v19.4s
+ .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s
+ .long 0x728147a8 // movk w8, #0xa3d
+ .long 0x4ea0db13 // fcmeq v19.4s, v24.4s, #0.0
+ .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s
+ .long 0x6e22dea2 // fmul v2.4s, v21.4s, v2.4s
+ .long 0x6e38fc21 // fdiv v1.4s, v1.4s, v24.4s
+ .long 0x4e040d17 // dup v23.4s, w8
+ .long 0x52a7bc28 // mov w8, #0x3de10000
+ .long 0x6e38fc00 // fdiv v0.4s, v0.4s, v24.4s
+ .long 0x6e38fc42 // fdiv v2.4s, v2.4s, v24.4s
+ .long 0x4e731c21 // bic v1.16b, v1.16b, v19.16b
+ .long 0x7288f5c8 // movk w8, #0x47ae
+ .long 0x4e27d470 // fadd v16.4s, v3.4s, v7.4s
+ .long 0x6e37dcb6 // fmul v22.4s, v5.4s, v23.4s
+ .long 0x4e731c00 // bic v0.16b, v0.16b, v19.16b
+ .long 0x4e731c42 // bic v2.16b, v2.16b, v19.16b
+ .long 0x6e37dc33 // fmul v19.4s, v1.4s, v23.4s
+ .long 0x4ea7cc70 // fmls v16.4s, v3.4s, v7.4s
+ .long 0x4e040d03 // dup v3.4s, w8
+ .long 0x4e24ce96 // fmla v22.4s, v20.4s, v4.4s
+ .long 0x4e20ce93 // fmla v19.4s, v20.4s, v0.4s
+ .long 0x4e26cc76 // fmla v22.4s, v3.4s, v6.4s
+ .long 0x4e22cc73 // fmla v19.4s, v3.4s, v2.4s
+ .long 0x4eb3d6d3 // fsub v19.4s, v22.4s, v19.4s
+ .long 0x4e33d421 // fadd v1.4s, v1.4s, v19.4s
+ .long 0x4e33d400 // fadd v0.4s, v0.4s, v19.4s
+ .long 0x6e37dc36 // fmul v22.4s, v1.4s, v23.4s
+ .long 0x4e33d442 // fadd v2.4s, v2.4s, v19.4s
+ .long 0x4e20ce96 // fmla v22.4s, v20.4s, v0.4s
+ .long 0x4ea2f433 // fmin v19.4s, v1.4s, v2.4s
+ .long 0x4e22cc76 // fmla v22.4s, v3.4s, v2.4s
+ .long 0x4eb3f413 // fmin v19.4s, v0.4s, v19.4s
+ .long 0x4eb6d417 // fsub v23.4s, v0.4s, v22.4s
+ .long 0x6ea0ca63 // fcmge v3.4s, v19.4s, #0.0
+ .long 0x4eb3d6d3 // fsub v19.4s, v22.4s, v19.4s
+ .long 0x6e37ded7 // fmul v23.4s, v22.4s, v23.4s
+ .long 0x6e33fef7 // fdiv v23.4s, v23.4s, v19.4s
+ .long 0x4e22f435 // fmax v21.4s, v1.4s, v2.4s
+ .long 0x4ea31c78 // mov v24.16b, v3.16b
+ .long 0x4e37d6d7 // fadd v23.4s, v22.4s, v23.4s
+ .long 0x4e35f415 // fmax v21.4s, v0.4s, v21.4s
+ .long 0x6e771c18 // bsl v24.16b, v0.16b, v23.16b
+ .long 0x4eb6d420 // fsub v0.4s, v1.4s, v22.4s
+ .long 0x6e20dec0 // fmul v0.4s, v22.4s, v0.4s
+ .long 0x6e33fc00 // fdiv v0.4s, v0.4s, v19.4s
+ .long 0x4ea31c77 // mov v23.16b, v3.16b
+ .long 0x4e20d6c0 // fadd v0.4s, v22.4s, v0.4s
+ .long 0x6e601c37 // bsl v23.16b, v1.16b, v0.16b
+ .long 0x4eb6d441 // fsub v1.4s, v2.4s, v22.4s
+ .long 0x6e21dec1 // fmul v1.4s, v22.4s, v1.4s
+ .long 0x6e33fc21 // fdiv v1.4s, v1.4s, v19.4s
+ .long 0x4e21d6c1 // fadd v1.4s, v22.4s, v1.4s
+ .long 0x6e611c43 // bsl v3.16b, v2.16b, v1.16b
+ .long 0x6eb2e6b4 // fcmgt v20.4s, v21.4s, v18.4s
+ .long 0x4eb6d652 // fsub v18.4s, v18.4s, v22.4s
+ .long 0x4eb6d701 // fsub v1.4s, v24.4s, v22.4s
+ .long 0x4eb6d6e2 // fsub v2.4s, v23.4s, v22.4s
+ .long 0x4eb6d473 // fsub v19.4s, v3.4s, v22.4s
+ .long 0x4eb6d6b5 // fsub v21.4s, v21.4s, v22.4s
+ .long 0x6e21de41 // fmul v1.4s, v18.4s, v1.4s
+ .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s
+ .long 0x6e33de52 // fmul v18.4s, v18.4s, v19.4s
+ .long 0x6e35fc21 // fdiv v1.4s, v1.4s, v21.4s
+ .long 0x6e35fc42 // fdiv v2.4s, v2.4s, v21.4s
+ .long 0x6e35fe52 // fdiv v18.4s, v18.4s, v21.4s
+ .long 0xf8408423 // ldr x3, [x1], #8
+ .long 0x4eb41e80 // mov v0.16b, v20.16b
+ .long 0x4eb41e93 // mov v19.16b, v20.16b
+ .long 0x4e21d6c1 // fadd v1.4s, v22.4s, v1.4s
+ .long 0x4e22d6c2 // fadd v2.4s, v22.4s, v2.4s
+ .long 0x4e32d6d2 // fadd v18.4s, v22.4s, v18.4s
+ .long 0x6f00e411 // movi v17.2d, #0x0
+ .long 0x6e781c20 // bsl v0.16b, v1.16b, v24.16b
+ .long 0x6e771c53 // bsl v19.16b, v2.16b, v23.16b
+ .long 0x6e631e54 // bsl v20.16b, v18.16b, v3.16b
+ .long 0x4e31f400 // fmax v0.4s, v0.4s, v17.4s
+ .long 0x4e31f661 // fmax v1.4s, v19.4s, v17.4s
+ .long 0x4e31f682 // fmax v2.4s, v20.4s, v17.4s
+ .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
+ .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
+ .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
+ .long 0x4eb01e03 // mov v3.16b, v16.16b
+ .long 0xd61f0060 // br x3
+
+HIDDEN _sk_saturation_aarch64
+.globl _sk_saturation_aarch64
+FUNCTION(_sk_saturation_aarch64)
+_sk_saturation_aarch64:
+ .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
+ .long 0x4ea0d8f3 // fcmeq v19.4s, v7.4s, #0.0
+ .long 0x4e22f434 // fmax v20.4s, v1.4s, v2.4s
+ .long 0x4ea2f422 // fmin v2.4s, v1.4s, v2.4s
+ .long 0x4e27d470 // fadd v16.4s, v3.4s, v7.4s
+ .long 0x6e27fe56 // fdiv v22.4s, v18.4s, v7.4s
+ .long 0x4e34f414 // fmax v20.4s, v0.4s, v20.4s
+ .long 0x4ea2f400 // fmin v0.4s, v0.4s, v2.4s
+ .long 0x4ea7cc70 // fmls v16.4s, v3.4s, v7.4s
+ .long 0x4e731ec3 // bic v3.16b, v22.16b, v19.16b
+ .long 0x4ea0d680 // fsub v0.4s, v20.4s, v0.4s
+ .long 0x6e24dc74 // fmul v20.4s, v3.4s, v4.4s
+ .long 0x6e25dc76 // fmul v22.4s, v3.4s, v5.4s
+ .long 0x6e26dc63 // fmul v3.4s, v3.4s, v6.4s
+ .long 0x4ea3f6d7 // fmin v23.4s, v22.4s, v3.4s
+ .long 0x52a7d328 // mov w8, #0x3e990000
+ .long 0x4e23f6d8 // fmax v24.4s, v22.4s, v3.4s
+ .long 0x4eb7f697 // fmin v23.4s, v20.4s, v23.4s
+ .long 0x72933348 // movk w8, #0x999a
+ .long 0x4e38f698 // fmax v24.4s, v20.4s, v24.4s
+ .long 0x4eb7d694 // fsub v20.4s, v20.4s, v23.4s
+ .long 0x4eb7d6d6 // fsub v22.4s, v22.4s, v23.4s
+ .long 0x4e040d01 // dup v1.4s, w8
+ .long 0x52a7e2e8 // mov w8, #0x3f170000
+ .long 0x4eb7d718 // fsub v24.4s, v24.4s, v23.4s
+ .long 0x4eb7d463 // fsub v3.4s, v3.4s, v23.4s
+ .long 0x6e34dc14 // fmul v20.4s, v0.4s, v20.4s
+ .long 0x6e36dc16 // fmul v22.4s, v0.4s, v22.4s
+ .long 0x728147a8 // movk w8, #0xa3d
+ .long 0x4ea0db17 // fcmeq v23.4s, v24.4s, #0.0
+ .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s
+ .long 0x6e38fe83 // fdiv v3.4s, v20.4s, v24.4s
+ .long 0x6e38fed4 // fdiv v20.4s, v22.4s, v24.4s
+ .long 0x4e040d15 // dup v21.4s, w8
+ .long 0x52a7bc28 // mov w8, #0x3de10000
+ .long 0x4e771e94 // bic v20.16b, v20.16b, v23.16b
+ .long 0x7288f5c8 // movk w8, #0x47ae
+ .long 0x6e35dcb3 // fmul v19.4s, v5.4s, v21.4s
+ .long 0x6e38fc00 // fdiv v0.4s, v0.4s, v24.4s
+ .long 0x4e771c63 // bic v3.16b, v3.16b, v23.16b
+ .long 0x6e35de96 // fmul v22.4s, v20.4s, v21.4s
+ .long 0x4e040d02 // dup v2.4s, w8
+ .long 0x4e24cc33 // fmla v19.4s, v1.4s, v4.4s
+ .long 0x4e771c00 // bic v0.16b, v0.16b, v23.16b
+ .long 0x4e23cc36 // fmla v22.4s, v1.4s, v3.4s
+ .long 0x4e26cc53 // fmla v19.4s, v2.4s, v6.4s
+ .long 0x4e20cc56 // fmla v22.4s, v2.4s, v0.4s
+ .long 0x4eb6d673 // fsub v19.4s, v19.4s, v22.4s
+ .long 0x4e33d694 // fadd v20.4s, v20.4s, v19.4s
+ .long 0x4e33d463 // fadd v3.4s, v3.4s, v19.4s
+ .long 0x6e35de95 // fmul v21.4s, v20.4s, v21.4s
+ .long 0x4e33d400 // fadd v0.4s, v0.4s, v19.4s
+ .long 0x4e23cc35 // fmla v21.4s, v1.4s, v3.4s
+ .long 0x4ea0f693 // fmin v19.4s, v20.4s, v0.4s
+ .long 0x4e20cc55 // fmla v21.4s, v2.4s, v0.4s
+ .long 0x4eb3f473 // fmin v19.4s, v3.4s, v19.4s
+ .long 0x4eb5d477 // fsub v23.4s, v3.4s, v21.4s
+ .long 0x6ea0ca61 // fcmge v1.4s, v19.4s, #0.0
+ .long 0x4eb3d6b3 // fsub v19.4s, v21.4s, v19.4s
+ .long 0x6e37deb7 // fmul v23.4s, v21.4s, v23.4s
+ .long 0x6e33fef7 // fdiv v23.4s, v23.4s, v19.4s
+ .long 0x4e20f696 // fmax v22.4s, v20.4s, v0.4s
+ .long 0x4ea11c38 // mov v24.16b, v1.16b
+ .long 0x4e37d6b7 // fadd v23.4s, v21.4s, v23.4s
+ .long 0x4e36f476 // fmax v22.4s, v3.4s, v22.4s
+ .long 0x6e771c78 // bsl v24.16b, v3.16b, v23.16b
+ .long 0x4eb5d683 // fsub v3.4s, v20.4s, v21.4s
+ .long 0x6e23dea3 // fmul v3.4s, v21.4s, v3.4s
+ .long 0x6e33fc63 // fdiv v3.4s, v3.4s, v19.4s
+ .long 0x4ea11c37 // mov v23.16b, v1.16b
+ .long 0x4e23d6a3 // fadd v3.4s, v21.4s, v3.4s
+ .long 0x6e631e97 // bsl v23.16b, v20.16b, v3.16b
+ .long 0x4eb5d414 // fsub v20.4s, v0.4s, v21.4s
+ .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s
+ .long 0x6e33fe93 // fdiv v19.4s, v20.4s, v19.4s
+ .long 0x4e33d6b3 // fadd v19.4s, v21.4s, v19.4s
+ .long 0x6e731c01 // bsl v1.16b, v0.16b, v19.16b
+ .long 0x6eb2e6c2 // fcmgt v2.4s, v22.4s, v18.4s
+ .long 0x4eb5d652 // fsub v18.4s, v18.4s, v21.4s
+ .long 0x4eb5d700 // fsub v0.4s, v24.4s, v21.4s
+ .long 0x4eb5d6f3 // fsub v19.4s, v23.4s, v21.4s
+ .long 0x4eb5d434 // fsub v20.4s, v1.4s, v21.4s
+ .long 0x4eb5d6d6 // fsub v22.4s, v22.4s, v21.4s
+ .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s
+ .long 0x6e33de53 // fmul v19.4s, v18.4s, v19.4s
+ .long 0x6e34de52 // fmul v18.4s, v18.4s, v20.4s
+ .long 0x6e36fc00 // fdiv v0.4s, v0.4s, v22.4s
+ .long 0x6e36fe73 // fdiv v19.4s, v19.4s, v22.4s
+ .long 0x6e36fe52 // fdiv v18.4s, v18.4s, v22.4s
+ .long 0xf8408423 // ldr x3, [x1], #8
+ .long 0x4ea21c43 // mov v3.16b, v2.16b
+ .long 0x4ea21c54 // mov v20.16b, v2.16b
+ .long 0x4e20d6a0 // fadd v0.4s, v21.4s, v0.4s
+ .long 0x4e33d6b3 // fadd v19.4s, v21.4s, v19.4s
+ .long 0x4e32d6b2 // fadd v18.4s, v21.4s, v18.4s
+ .long 0x6f00e411 // movi v17.2d, #0x0
+ .long 0x6e781c03 // bsl v3.16b, v0.16b, v24.16b
+ .long 0x6e771e74 // bsl v20.16b, v19.16b, v23.16b
+ .long 0x6e611e42 // bsl v2.16b, v18.16b, v1.16b
+ .long 0x4e31f460 // fmax v0.4s, v3.4s, v17.4s
+ .long 0x4e31f681 // fmax v1.4s, v20.4s, v17.4s
+ .long 0x4e31f442 // fmax v2.4s, v2.4s, v17.4s
+ .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
+ .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
+ .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
+ .long 0x4eb01e03 // mov v3.16b, v16.16b
+ .long 0xd61f0060 // br x3
+
+HIDDEN _sk_color_aarch64
+.globl _sk_color_aarch64
+FUNCTION(_sk_color_aarch64)
+_sk_color_aarch64:
+ .long 0x52a7d328 // mov w8, #0x3e990000
+ .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
+ .long 0x72933348 // movk w8, #0x999a
+ .long 0x4ea0d874 // fcmeq v20.4s, v3.4s, #0.0
+ .long 0x4e040d13 // dup v19.4s, w8
+ .long 0x52a7e2e8 // mov w8, #0x3f170000
+ .long 0x4e27d470 // fadd v16.4s, v3.4s, v7.4s
+ .long 0x6e23fe56 // fdiv v22.4s, v18.4s, v3.4s
+ .long 0x728147a8 // movk w8, #0xa3d
+ .long 0x4ea7cc70 // fmls v16.4s, v3.4s, v7.4s
+ .long 0x4e741ec3 // bic v3.16b, v22.16b, v20.16b
+ .long 0x4e040d15 // dup v21.4s, w8
+ .long 0x52a7bc28 // mov w8, #0x3de10000
+ .long 0x6e21dc61 // fmul v1.4s, v3.4s, v1.4s
+ .long 0x7288f5c8 // movk w8, #0x47ae
+ .long 0x6e35dcb4 // fmul v20.4s, v5.4s, v21.4s
+ .long 0x6e20dc60 // fmul v0.4s, v3.4s, v0.4s
+ .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s
+ .long 0x6e35dc23 // fmul v3.4s, v1.4s, v21.4s
+ .long 0x4e040d17 // dup v23.4s, w8
+ .long 0x4e24ce74 // fmla v20.4s, v19.4s, v4.4s
+ .long 0x4e20ce63 // fmla v3.4s, v19.4s, v0.4s
+ .long 0x4e26cef4 // fmla v20.4s, v23.4s, v6.4s
+ .long 0x4e22cee3 // fmla v3.4s, v23.4s, v2.4s
+ .long 0x4ea3d683 // fsub v3.4s, v20.4s, v3.4s
+ .long 0x4e23d421 // fadd v1.4s, v1.4s, v3.4s
+ .long 0x4e23d400 // fadd v0.4s, v0.4s, v3.4s
+ .long 0x6e35dc35 // fmul v21.4s, v1.4s, v21.4s
+ .long 0x4e23d442 // fadd v2.4s, v2.4s, v3.4s
+ .long 0x4e20ce75 // fmla v21.4s, v19.4s, v0.4s
+ .long 0x4ea2f423 // fmin v3.4s, v1.4s, v2.4s
+ .long 0x4e22cef5 // fmla v21.4s, v23.4s, v2.4s
+ .long 0x4ea3f403 // fmin v3.4s, v0.4s, v3.4s
+ .long 0x4eb5d417 // fsub v23.4s, v0.4s, v21.4s
+ .long 0x6ea0c873 // fcmge v19.4s, v3.4s, #0.0
+ .long 0x4ea3d6a3 // fsub v3.4s, v21.4s, v3.4s
+ .long 0x6e37deb7 // fmul v23.4s, v21.4s, v23.4s
+ .long 0x6e23fef7 // fdiv v23.4s, v23.4s, v3.4s
+ .long 0x4e22f434 // fmax v20.4s, v1.4s, v2.4s
+ .long 0x4eb31e78 // mov v24.16b, v19.16b
+ .long 0x4e37d6b7 // fadd v23.4s, v21.4s, v23.4s
+ .long 0x4e34f414 // fmax v20.4s, v0.4s, v20.4s
+ .long 0x6e771c18 // bsl v24.16b, v0.16b, v23.16b
+ .long 0x4eb5d420 // fsub v0.4s, v1.4s, v21.4s
+ .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s
+ .long 0x6e23fc00 // fdiv v0.4s, v0.4s, v3.4s
+ .long 0x4eb31e77 // mov v23.16b, v19.16b
+ .long 0x4e20d6a0 // fadd v0.4s, v21.4s, v0.4s
+ .long 0x6e601c37 // bsl v23.16b, v1.16b, v0.16b
+ .long 0x4eb5d441 // fsub v1.4s, v2.4s, v21.4s
+ .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s
+ .long 0x6e23fc21 // fdiv v1.4s, v1.4s, v3.4s
+ .long 0x4e21d6a1 // fadd v1.4s, v21.4s, v1.4s
+ .long 0x6e611c53 // bsl v19.16b, v2.16b, v1.16b
+ .long 0x6eb2e696 // fcmgt v22.4s, v20.4s, v18.4s
+ .long 0x4eb5d652 // fsub v18.4s, v18.4s, v21.4s
+ .long 0x4eb5d701 // fsub v1.4s, v24.4s, v21.4s
+ .long 0x4eb5d6e2 // fsub v2.4s, v23.4s, v21.4s
+ .long 0x4eb5d663 // fsub v3.4s, v19.4s, v21.4s
+ .long 0x4eb5d694 // fsub v20.4s, v20.4s, v21.4s
+ .long 0x6e21de41 // fmul v1.4s, v18.4s, v1.4s
+ .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s
+ .long 0x6e23de43 // fmul v3.4s, v18.4s, v3.4s
+ .long 0x6e34fc21 // fdiv v1.4s, v1.4s, v20.4s
+ .long 0x6e34fc42 // fdiv v2.4s, v2.4s, v20.4s
+ .long 0x6e34fc63 // fdiv v3.4s, v3.4s, v20.4s
+ .long 0xf8408423 // ldr x3, [x1], #8
+ .long 0x4eb61ec0 // mov v0.16b, v22.16b
+ .long 0x4eb61ed2 // mov v18.16b, v22.16b
+ .long 0x4e21d6a1 // fadd v1.4s, v21.4s, v1.4s
+ .long 0x4e22d6a2 // fadd v2.4s, v21.4s, v2.4s
+ .long 0x4e23d6a3 // fadd v3.4s, v21.4s, v3.4s
+ .long 0x6f00e411 // movi v17.2d, #0x0
+ .long 0x6e781c20 // bsl v0.16b, v1.16b, v24.16b
+ .long 0x6e771c52 // bsl v18.16b, v2.16b, v23.16b
+ .long 0x6e731c76 // bsl v22.16b, v3.16b, v19.16b
+ .long 0x4e31f400 // fmax v0.4s, v0.4s, v17.4s
+ .long 0x4e31f641 // fmax v1.4s, v18.4s, v17.4s
+ .long 0x4e31f6c2 // fmax v2.4s, v22.4s, v17.4s
+ .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
+ .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
+ .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
+ .long 0x4eb01e03 // mov v3.16b, v16.16b
+ .long 0xd61f0060 // br x3
+
+HIDDEN _sk_luminosity_aarch64
+.globl _sk_luminosity_aarch64
+FUNCTION(_sk_luminosity_aarch64)
+_sk_luminosity_aarch64:
+ .long 0x52a7d328 // mov w8, #0x3e990000
+ .long 0x72933348 // movk w8, #0x999a
+ .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
+ .long 0x4e040d13 // dup v19.4s, w8
+ .long 0x52a7e2e8 // mov w8, #0x3f170000
+ .long 0x4ea0d8f4 // fcmeq v20.4s, v7.4s, #0.0
+ .long 0x728147a8 // movk w8, #0xa3d
+ .long 0x4e27d470 // fadd v16.4s, v3.4s, v7.4s
+ .long 0x6e27fe56 // fdiv v22.4s, v18.4s, v7.4s
+ .long 0x4e040d15 // dup v21.4s, w8
+ .long 0x4ea7cc70 // fmls v16.4s, v3.4s, v7.4s
+ .long 0x4e741ec3 // bic v3.16b, v22.16b, v20.16b
+ .long 0x52a7bc28 // mov w8, #0x3de10000
+ .long 0x6e35dc21 // fmul v1.4s, v1.4s, v21.4s
+ .long 0x6e25dc76 // fmul v22.4s, v3.4s, v5.4s
+ .long 0x7288f5c8 // movk w8, #0x47ae
+ .long 0x6e24dc74 // fmul v20.4s, v3.4s, v4.4s
+ .long 0x4e20ce61 // fmla v1.4s, v19.4s, v0.4s
+ .long 0x6e35dec0 // fmul v0.4s, v22.4s, v21.4s
+ .long 0x4e040d17 // dup v23.4s, w8
+ .long 0x6e26dc63 // fmul v3.4s, v3.4s, v6.4s
+ .long 0x4e34ce60 // fmla v0.4s, v19.4s, v20.4s
+ .long 0x4e22cee1 // fmla v1.4s, v23.4s, v2.4s
+ .long 0x4e23cee0 // fmla v0.4s, v23.4s, v3.4s
+ .long 0x4ea0d420 // fsub v0.4s, v1.4s, v0.4s
+ .long 0x4e20d6c2 // fadd v2.4s, v22.4s, v0.4s
+ .long 0x4e20d681 // fadd v1.4s, v20.4s, v0.4s
+ .long 0x6e35dc55 // fmul v21.4s, v2.4s, v21.4s
+ .long 0x4e20d460 // fadd v0.4s, v3.4s, v0.4s
+ .long 0x4e21ce75 // fmla v21.4s, v19.4s, v1.4s
+ .long 0x4ea0f443 // fmin v3.4s, v2.4s, v0.4s
+ .long 0x4e20cef5 // fmla v21.4s, v23.4s, v0.4s
+ .long 0x4ea3f423 // fmin v3.4s, v1.4s, v3.4s
+ .long 0x4eb5d437 // fsub v23.4s, v1.4s, v21.4s
+ .long 0x6ea0c873 // fcmge v19.4s, v3.4s, #0.0
+ .long 0x4ea3d6a3 // fsub v3.4s, v21.4s, v3.4s
+ .long 0x6e37deb7 // fmul v23.4s, v21.4s, v23.4s
+ .long 0x6e23fef7 // fdiv v23.4s, v23.4s, v3.4s
+ .long 0x4e20f454 // fmax v20.4s, v2.4s, v0.4s
+ .long 0x4eb31e78 // mov v24.16b, v19.16b
+ .long 0x4e37d6b7 // fadd v23.4s, v21.4s, v23.4s
+ .long 0x4e34f434 // fmax v20.4s, v1.4s, v20.4s
+ .long 0x6e771c38 // bsl v24.16b, v1.16b, v23.16b
+ .long 0x4eb5d441 // fsub v1.4s, v2.4s, v21.4s
+ .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s
+ .long 0x6e23fc21 // fdiv v1.4s, v1.4s, v3.4s
+ .long 0x4eb31e77 // mov v23.16b, v19.16b
+ .long 0x4e21d6a1 // fadd v1.4s, v21.4s, v1.4s
+ .long 0x6e611c57 // bsl v23.16b, v2.16b, v1.16b
+ .long 0x4eb5d402 // fsub v2.4s, v0.4s, v21.4s
+ .long 0x6e22dea2 // fmul v2.4s, v21.4s, v2.4s
+ .long 0x6e23fc42 // fdiv v2.4s, v2.4s, v3.4s
+ .long 0x4e22d6a2 // fadd v2.4s, v21.4s, v2.4s
+ .long 0x6e621c13 // bsl v19.16b, v0.16b, v2.16b
+ .long 0x6eb2e696 // fcmgt v22.4s, v20.4s, v18.4s
+ .long 0x4eb5d652 // fsub v18.4s, v18.4s, v21.4s
+ .long 0x4eb5d700 // fsub v0.4s, v24.4s, v21.4s
+ .long 0x4eb5d6e2 // fsub v2.4s, v23.4s, v21.4s
+ .long 0x4eb5d663 // fsub v3.4s, v19.4s, v21.4s
+ .long 0x4eb5d694 // fsub v20.4s, v20.4s, v21.4s
+ .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s
+ .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s
+ .long 0x6e23de43 // fmul v3.4s, v18.4s, v3.4s
+ .long 0x6e34fc00 // fdiv v0.4s, v0.4s, v20.4s
+ .long 0x6e34fc42 // fdiv v2.4s, v2.4s, v20.4s
+ .long 0x6e34fc63 // fdiv v3.4s, v3.4s, v20.4s
+ .long 0xf8408423 // ldr x3, [x1], #8
+ .long 0x4eb61ec1 // mov v1.16b, v22.16b
+ .long 0x4eb61ed2 // mov v18.16b, v22.16b
+ .long 0x4e20d6a0 // fadd v0.4s, v21.4s, v0.4s
+ .long 0x4e22d6a2 // fadd v2.4s, v21.4s, v2.4s
+ .long 0x4e23d6a3 // fadd v3.4s, v21.4s, v3.4s
+ .long 0x6f00e411 // movi v17.2d, #0x0
+ .long 0x6e781c01 // bsl v1.16b, v0.16b, v24.16b
+ .long 0x6e771c52 // bsl v18.16b, v2.16b, v23.16b
+ .long 0x6e731c76 // bsl v22.16b, v3.16b, v19.16b
+ .long 0x4e31f420 // fmax v0.4s, v1.4s, v17.4s
+ .long 0x4e31f641 // fmax v1.4s, v18.4s, v17.4s
+ .long 0x4e31f6c2 // fmax v2.4s, v22.4s, v17.4s
+ .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
+ .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
+ .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
+ .long 0x4eb01e03 // mov v3.16b, v16.16b
+ .long 0xd61f0060 // br x3
+
HIDDEN _sk_clamp_0_aarch64
.globl _sk_clamp_0_aarch64
FUNCTION(_sk_clamp_0_aarch64)
@@ -2250,9 +2652,9 @@ FUNCTION(_sk_gather_i8_aarch64)
_sk_gather_i8_aarch64:
.long 0xaa0103e8 // mov x8, x1
.long 0xf8408429 // ldr x9, [x1], #8
- .long 0xb4000069 // cbz x9, 1d9c <sk_gather_i8_aarch64+0x14>
+ .long 0xb4000069 // cbz x9, 2394 <sk_gather_i8_aarch64+0x14>
.long 0xaa0903ea // mov x10, x9
- .long 0x14000003 // b 1da4 <sk_gather_i8_aarch64+0x1c>
+ .long 0x14000003 // b 239c <sk_gather_i8_aarch64+0x1c>
.long 0xf940050a // ldr x10, [x8, #8]
.long 0x91004101 // add x1, x8, #0x10
.long 0xf8410548 // ldr x8, [x10], #16
@@ -3101,7 +3503,7 @@ _sk_linear_gradient_aarch64:
.long 0x4d40c902 // ld1r {v2.4s}, [x8]
.long 0xf9400128 // ldr x8, [x9]
.long 0x4d40c943 // ld1r {v3.4s}, [x10]
- .long 0xb40006c8 // cbz x8, 2970 <sk_linear_gradient_aarch64+0x100>
+ .long 0xb40006c8 // cbz x8, 2f68 <sk_linear_gradient_aarch64+0x100>
.long 0x6dbf23e9 // stp d9, d8, [sp, #-16]!
.long 0xf9400529 // ldr x9, [x9, #8]
.long 0x6f00e413 // movi v19.2d, #0x0
@@ -3152,9 +3554,9 @@ _sk_linear_gradient_aarch64:
.long 0xd1000508 // sub x8, x8, #0x1
.long 0x6e771fd0 // bsl v16.16b, v30.16b, v23.16b
.long 0x91009129 // add x9, x9, #0x24
- .long 0xb5fffaa8 // cbnz x8, 28b8 <sk_linear_gradient_aarch64+0x48>
+ .long 0xb5fffaa8 // cbnz x8, 2eb0 <sk_linear_gradient_aarch64+0x48>
.long 0x6cc123e9 // ldp d9, d8, [sp], #16
- .long 0x14000005 // b 2980 <sk_linear_gradient_aarch64+0x110>
+ .long 0x14000005 // b 2f78 <sk_linear_gradient_aarch64+0x110>
.long 0x6f00e414 // movi v20.2d, #0x0
.long 0x6f00e412 // movi v18.2d, #0x0
.long 0x6f00e411 // movi v17.2d, #0x0
@@ -4393,6 +4795,482 @@ _sk_softlight_vfp4:
.long 0xecbd8b06 // vpop {d8-d10}
.long 0xe12fff13 // bx r3
+HIDDEN _sk_hue_vfp4
+.globl _sk_hue_vfp4
+FUNCTION(_sk_hue_vfp4)
+_sk_hue_vfp4:
+ .long 0xed2d8b04 // vpush {d8-d9}
+ .long 0xeeb78a00 // vmov.f32 s16, #112
+ .long 0xf3f91503 // vceq.f32 d17, d3, #0
+ .long 0xf2c00010 // vmov.i32 d16, #0
+ .long 0xeddf9b78 // vldr d25, [pc, #480]
+ .long 0xf2655f06 // vmin.f32 d21, d5, d6
+ .long 0xe4913004 // ldr r3, [r1], #4
+ .long 0xf2456f06 // vmax.f32 d22, d5, d6
+ .long 0xf346bd39 // vmul.f32 d27, d6, d25
+ .long 0xf2645f25 // vmin.f32 d21, d4, d21
+ .long 0xeec89a23 // vdiv.f32 s19, s16, s7
+ .long 0xee889a03 // vdiv.f32 s18, s16, s6
+ .long 0xf3501199 // vbsl d17, d16, d9
+ .long 0xf2446f26 // vmax.f32 d22, d4, d22
+ .long 0xf3412d92 // vmul.f32 d18, d17, d2
+ .long 0xf3413d91 // vmul.f32 d19, d17, d1
+ .long 0xf3411d90 // vmul.f32 d17, d17, d0
+ .long 0xf2665da5 // vsub.f32 d21, d22, d21
+ .long 0xf2634fa2 // vmin.f32 d20, d19, d18
+ .long 0xf2437fa2 // vmax.f32 d23, d19, d18
+ .long 0xf2614fa4 // vmin.f32 d20, d17, d20
+ .long 0xf2417fa7 // vmax.f32 d23, d17, d23
+ .long 0xf2633da4 // vsub.f32 d19, d19, d20
+ .long 0xf2270da4 // vsub.f32 d0, d23, d20
+ .long 0xf2611da4 // vsub.f32 d17, d17, d20
+ .long 0xf2622da4 // vsub.f32 d18, d18, d20
+ .long 0xeddf4b60 // vldr d20, [pc, #384]
+ .long 0xf3051db3 // vmul.f32 d1, d21, d19
+ .long 0xf3447d34 // vmul.f32 d23, d4, d20
+ .long 0xf3052db1 // vmul.f32 d2, d21, d17
+ .long 0xeddf1b5a // vldr d17, [pc, #360]
+ .long 0xf3058db2 // vmul.f32 d8, d21, d18
+ .long 0xf3f92500 // vceq.f32 d18, d0, #0
+ .long 0xf26231b2 // vorr d19, d18, d18
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf26251b2 // vorr d21, d18, d18
+ .long 0xf3503199 // vbsl d19, d16, d9
+ .long 0xf3505191 // vbsl d21, d16, d1
+ .long 0xf3456d31 // vmul.f32 d22, d5, d17
+ .long 0xf3438db1 // vmul.f32 d24, d19, d17
+ .long 0xf345adb4 // vmul.f32 d26, d21, d20
+ .long 0xf3502192 // vbsl d18, d16, d2
+ .long 0xf2476da6 // vadd.f32 d22, d23, d22
+ .long 0xf3427db9 // vmul.f32 d23, d18, d25
+ .long 0xf24a8da8 // vadd.f32 d24, d26, d24
+ .long 0xf2466dab // vadd.f32 d22, d22, d27
+ .long 0xf2487da7 // vadd.f32 d23, d24, d23
+ .long 0xf2666da7 // vsub.f32 d22, d22, d23
+ .long 0xf2433da6 // vadd.f32 d19, d19, d22
+ .long 0xf2455da6 // vadd.f32 d21, d21, d22
+ .long 0xf2422da6 // vadd.f32 d18, d18, d22
+ .long 0xf3431db1 // vmul.f32 d17, d19, d17
+ .long 0xf3454db4 // vmul.f32 d20, d21, d20
+ .long 0xf3426db9 // vmul.f32 d22, d18, d25
+ .long 0xf2441da1 // vadd.f32 d17, d20, d17
+ .long 0xf2634fa2 // vmin.f32 d20, d19, d18
+ .long 0xf2461da1 // vadd.f32 d17, d22, d17
+ .long 0xf2654fa4 // vmin.f32 d20, d21, d20
+ .long 0xf2636da1 // vsub.f32 d22, d19, d17
+ .long 0xf2210da4 // vsub.f32 d0, d17, d20
+ .long 0xf3f944a4 // vcge.f32 d20, d20, #0
+ .long 0xf2627da1 // vsub.f32 d23, d18, d17
+ .long 0xf2658da1 // vsub.f32 d24, d21, d17
+ .long 0xf264a1b4 // vorr d26, d20, d20
+ .long 0xf3011db6 // vmul.f32 d1, d17, d22
+ .long 0xf2436fa2 // vmax.f32 d22, d19, d18
+ .long 0xf3012db7 // vmul.f32 d2, d17, d23
+ .long 0xf3018db8 // vmul.f32 d8, d17, d24
+ .long 0xf264b1b4 // vorr d27, d20, d20
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2419d82 // vadd.f32 d25, d17, d2
+ .long 0xf355a1b9 // vbsl d26, d21, d25
+ .long 0xf2418d81 // vadd.f32 d24, d17, d1
+ .long 0xf2c79f10 // vmov.f32 d25, #1
+ .long 0xf2417d89 // vadd.f32 d23, d17, d9
+ .long 0xf35241b8 // vbsl d20, d18, d24
+ .long 0xf26a2da1 // vsub.f32 d18, d26, d17
+ .long 0xf353b1b7 // vbsl d27, d19, d23
+ .long 0xf2453fa6 // vmax.f32 d19, d21, d22
+ .long 0xf2695da1 // vsub.f32 d21, d25, d17
+ .long 0xf26b6da1 // vsub.f32 d22, d27, d17
+ .long 0xf2647da1 // vsub.f32 d23, d20, d17
+ .long 0xf2230da1 // vsub.f32 d0, d19, d17
+ .long 0xf3051db2 // vmul.f32 d1, d21, d18
+ .long 0xf3632ea9 // vcgt.f32 d18, d19, d25
+ .long 0xf3052db6 // vmul.f32 d2, d21, d22
+ .long 0xf3058db7 // vmul.f32 d8, d21, d23
+ .long 0xf3436d17 // vmul.f32 d22, d3, d7
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2413d89 // vadd.f32 d19, d17, d9
+ .long 0xf26281b2 // vorr d24, d18, d18
+ .long 0xf2415d81 // vadd.f32 d21, d17, d1
+ .long 0xf2411d82 // vadd.f32 d17, d17, d2
+ .long 0xf35381ba // vbsl d24, d19, d26
+ .long 0xf26231b2 // vorr d19, d18, d18
+ .long 0xf2437d07 // vadd.f32 d23, d3, d7
+ .long 0xf35531bb // vbsl d19, d21, d27
+ .long 0xf35121b4 // vbsl d18, d17, d20
+ .long 0xf2481fa0 // vmax.f32 d17, d24, d16
+ .long 0xf2433fa0 // vmax.f32 d19, d19, d16
+ .long 0xf2420fa0 // vmax.f32 d16, d18, d16
+ .long 0xf2273da6 // vsub.f32 d3, d23, d22
+ .long 0xf3030d31 // vmul.f32 d0, d3, d17
+ .long 0xf3031d33 // vmul.f32 d1, d3, d19
+ .long 0xf3032d30 // vmul.f32 d2, d3, d16
+ .long 0xecbd8b04 // vpop {d8-d9}
+ .long 0xe12fff13 // bx r3
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3de147ae // .word 0x3de147ae
+ .long 0x3de147ae // .word 0x3de147ae
+
+HIDDEN _sk_saturation_vfp4
+.globl _sk_saturation_vfp4
+FUNCTION(_sk_saturation_vfp4)
+_sk_saturation_vfp4:
+ .long 0xed2d8b04 // vpush {d8-d9}
+ .long 0xeeb78a00 // vmov.f32 s16, #112
+ .long 0xf3f91507 // vceq.f32 d17, d7, #0
+ .long 0xf2c00010 // vmov.i32 d16, #0
+ .long 0xeddf9b78 // vldr d25, [pc, #480]
+ .long 0xf2615f02 // vmin.f32 d21, d1, d2
+ .long 0xe4913004 // ldr r3, [r1], #4
+ .long 0xf2416f02 // vmax.f32 d22, d1, d2
+ .long 0xf346bd39 // vmul.f32 d27, d6, d25
+ .long 0xf2605f25 // vmin.f32 d21, d0, d21
+ .long 0xeec89a27 // vdiv.f32 s19, s16, s15
+ .long 0xee889a07 // vdiv.f32 s18, s16, s14
+ .long 0xf3501199 // vbsl d17, d16, d9
+ .long 0xf2406f26 // vmax.f32 d22, d0, d22
+ .long 0xf3412d96 // vmul.f32 d18, d17, d6
+ .long 0xf3413d95 // vmul.f32 d19, d17, d5
+ .long 0xf3411d94 // vmul.f32 d17, d17, d4
+ .long 0xf2665da5 // vsub.f32 d21, d22, d21
+ .long 0xf2634fa2 // vmin.f32 d20, d19, d18
+ .long 0xf2437fa2 // vmax.f32 d23, d19, d18
+ .long 0xf2614fa4 // vmin.f32 d20, d17, d20
+ .long 0xf2417fa7 // vmax.f32 d23, d17, d23
+ .long 0xf2633da4 // vsub.f32 d19, d19, d20
+ .long 0xf2270da4 // vsub.f32 d0, d23, d20
+ .long 0xf2611da4 // vsub.f32 d17, d17, d20
+ .long 0xf2622da4 // vsub.f32 d18, d18, d20
+ .long 0xeddf4b60 // vldr d20, [pc, #384]
+ .long 0xf3051db3 // vmul.f32 d1, d21, d19
+ .long 0xf3447d34 // vmul.f32 d23, d4, d20
+ .long 0xf3052db1 // vmul.f32 d2, d21, d17
+ .long 0xeddf1b5a // vldr d17, [pc, #360]
+ .long 0xf3058db2 // vmul.f32 d8, d21, d18
+ .long 0xf3f92500 // vceq.f32 d18, d0, #0
+ .long 0xf26231b2 // vorr d19, d18, d18
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf26251b2 // vorr d21, d18, d18
+ .long 0xf3503199 // vbsl d19, d16, d9
+ .long 0xf3505191 // vbsl d21, d16, d1
+ .long 0xf3456d31 // vmul.f32 d22, d5, d17
+ .long 0xf3438db1 // vmul.f32 d24, d19, d17
+ .long 0xf345adb4 // vmul.f32 d26, d21, d20
+ .long 0xf3502192 // vbsl d18, d16, d2
+ .long 0xf2476da6 // vadd.f32 d22, d23, d22
+ .long 0xf3427db9 // vmul.f32 d23, d18, d25
+ .long 0xf24a8da8 // vadd.f32 d24, d26, d24
+ .long 0xf2466dab // vadd.f32 d22, d22, d27
+ .long 0xf2487da7 // vadd.f32 d23, d24, d23
+ .long 0xf2666da7 // vsub.f32 d22, d22, d23
+ .long 0xf2433da6 // vadd.f32 d19, d19, d22
+ .long 0xf2455da6 // vadd.f32 d21, d21, d22
+ .long 0xf2422da6 // vadd.f32 d18, d18, d22
+ .long 0xf3431db1 // vmul.f32 d17, d19, d17
+ .long 0xf3454db4 // vmul.f32 d20, d21, d20
+ .long 0xf3426db9 // vmul.f32 d22, d18, d25
+ .long 0xf2441da1 // vadd.f32 d17, d20, d17
+ .long 0xf2634fa2 // vmin.f32 d20, d19, d18
+ .long 0xf2461da1 // vadd.f32 d17, d22, d17
+ .long 0xf2654fa4 // vmin.f32 d20, d21, d20
+ .long 0xf2636da1 // vsub.f32 d22, d19, d17
+ .long 0xf2210da4 // vsub.f32 d0, d17, d20
+ .long 0xf3f944a4 // vcge.f32 d20, d20, #0
+ .long 0xf2627da1 // vsub.f32 d23, d18, d17
+ .long 0xf2658da1 // vsub.f32 d24, d21, d17
+ .long 0xf264a1b4 // vorr d26, d20, d20
+ .long 0xf3011db6 // vmul.f32 d1, d17, d22
+ .long 0xf2436fa2 // vmax.f32 d22, d19, d18
+ .long 0xf3012db7 // vmul.f32 d2, d17, d23
+ .long 0xf3018db8 // vmul.f32 d8, d17, d24
+ .long 0xf264b1b4 // vorr d27, d20, d20
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2419d82 // vadd.f32 d25, d17, d2
+ .long 0xf355a1b9 // vbsl d26, d21, d25
+ .long 0xf2418d81 // vadd.f32 d24, d17, d1
+ .long 0xf2c79f10 // vmov.f32 d25, #1
+ .long 0xf2417d89 // vadd.f32 d23, d17, d9
+ .long 0xf35241b8 // vbsl d20, d18, d24
+ .long 0xf26a2da1 // vsub.f32 d18, d26, d17
+ .long 0xf353b1b7 // vbsl d27, d19, d23
+ .long 0xf2453fa6 // vmax.f32 d19, d21, d22
+ .long 0xf2695da1 // vsub.f32 d21, d25, d17
+ .long 0xf26b6da1 // vsub.f32 d22, d27, d17
+ .long 0xf2647da1 // vsub.f32 d23, d20, d17
+ .long 0xf2230da1 // vsub.f32 d0, d19, d17
+ .long 0xf3051db2 // vmul.f32 d1, d21, d18
+ .long 0xf3632ea9 // vcgt.f32 d18, d19, d25
+ .long 0xf3052db6 // vmul.f32 d2, d21, d22
+ .long 0xf3058db7 // vmul.f32 d8, d21, d23
+ .long 0xf3436d17 // vmul.f32 d22, d3, d7
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2413d89 // vadd.f32 d19, d17, d9
+ .long 0xf26281b2 // vorr d24, d18, d18
+ .long 0xf2415d81 // vadd.f32 d21, d17, d1
+ .long 0xf2411d82 // vadd.f32 d17, d17, d2
+ .long 0xf35381ba // vbsl d24, d19, d26
+ .long 0xf26231b2 // vorr d19, d18, d18
+ .long 0xf2437d07 // vadd.f32 d23, d3, d7
+ .long 0xf35531bb // vbsl d19, d21, d27
+ .long 0xf35121b4 // vbsl d18, d17, d20
+ .long 0xf2481fa0 // vmax.f32 d17, d24, d16
+ .long 0xf2433fa0 // vmax.f32 d19, d19, d16
+ .long 0xf2420fa0 // vmax.f32 d16, d18, d16
+ .long 0xf2273da6 // vsub.f32 d3, d23, d22
+ .long 0xf3030d31 // vmul.f32 d0, d3, d17
+ .long 0xf3031d33 // vmul.f32 d1, d3, d19
+ .long 0xf3032d30 // vmul.f32 d2, d3, d16
+ .long 0xecbd8b04 // vpop {d8-d9}
+ .long 0xe12fff13 // bx r3
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3de147ae // .word 0x3de147ae
+ .long 0x3de147ae // .word 0x3de147ae
+
+HIDDEN _sk_color_vfp4
+.globl _sk_color_vfp4
+FUNCTION(_sk_color_vfp4)
+_sk_color_vfp4:
+ .long 0xed2d8b04 // vpush {d8-d9}
+ .long 0xeeb78a00 // vmov.f32 s16, #112
+ .long 0xf3f91503 // vceq.f32 d17, d3, #0
+ .long 0xf2c00010 // vmov.i32 d16, #0
+ .long 0xeddf2b58 // vldr d18, [pc, #352]
+ .long 0xeddf4b59 // vldr d20, [pc, #356]
+ .long 0xf3456d32 // vmul.f32 d22, d5, d18
+ .long 0xeddf8b59 // vldr d24, [pc, #356]
+ .long 0xf3447d34 // vmul.f32 d23, d4, d20
+ .long 0xe4913004 // ldr r3, [r1], #4
+ .long 0xf346bd38 // vmul.f32 d27, d6, d24
+ .long 0xeec89a23 // vdiv.f32 s19, s16, s7
+ .long 0xee889a03 // vdiv.f32 s18, s16, s6
+ .long 0xf3501199 // vbsl d17, d16, d9
+ .long 0xf2476da6 // vadd.f32 d22, d23, d22
+ .long 0xf3413d91 // vmul.f32 d19, d17, d1
+ .long 0xf3415d90 // vmul.f32 d21, d17, d0
+ .long 0xf3411d92 // vmul.f32 d17, d17, d2
+ .long 0xf2466dab // vadd.f32 d22, d22, d27
+ .long 0xf3439db2 // vmul.f32 d25, d19, d18
+ .long 0xf345adb4 // vmul.f32 d26, d21, d20
+ .long 0xf3417db8 // vmul.f32 d23, d17, d24
+ .long 0xf24a9da9 // vadd.f32 d25, d26, d25
+ .long 0xf2477da9 // vadd.f32 d23, d23, d25
+ .long 0xf2666da7 // vsub.f32 d22, d22, d23
+ .long 0xf2433da6 // vadd.f32 d19, d19, d22
+ .long 0xf2455da6 // vadd.f32 d21, d21, d22
+ .long 0xf2416da6 // vadd.f32 d22, d17, d22
+ .long 0xf3431db2 // vmul.f32 d17, d19, d18
+ .long 0xf3452db4 // vmul.f32 d18, d21, d20
+ .long 0xf3464db8 // vmul.f32 d20, d22, d24
+ .long 0xf2421da1 // vadd.f32 d17, d18, d17
+ .long 0xf2632fa6 // vmin.f32 d18, d19, d22
+ .long 0xf2441da1 // vadd.f32 d17, d20, d17
+ .long 0xf2652fa2 // vmin.f32 d18, d21, d18
+ .long 0xf2634da1 // vsub.f32 d20, d19, d17
+ .long 0xf2210da2 // vsub.f32 d0, d17, d18
+ .long 0xf3f924a2 // vcge.f32 d18, d18, #0
+ .long 0xf2667da1 // vsub.f32 d23, d22, d17
+ .long 0xf2658da1 // vsub.f32 d24, d21, d17
+ .long 0xf262a1b2 // vorr d26, d18, d18
+ .long 0xf3011db4 // vmul.f32 d1, d17, d20
+ .long 0xf2434fa6 // vmax.f32 d20, d19, d22
+ .long 0xf3012db7 // vmul.f32 d2, d17, d23
+ .long 0xf3018db8 // vmul.f32 d8, d17, d24
+ .long 0xf262b1b2 // vorr d27, d18, d18
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2419d82 // vadd.f32 d25, d17, d2
+ .long 0xf355a1b9 // vbsl d26, d21, d25
+ .long 0xf2417d89 // vadd.f32 d23, d17, d9
+ .long 0xf2c79f10 // vmov.f32 d25, #1
+ .long 0xf353b1b7 // vbsl d27, d19, d23
+ .long 0xf2453fa4 // vmax.f32 d19, d21, d20
+ .long 0xf26a4da1 // vsub.f32 d20, d26, d17
+ .long 0xf2695da1 // vsub.f32 d21, d25, d17
+ .long 0xf2418d81 // vadd.f32 d24, d17, d1
+ .long 0xf2230da1 // vsub.f32 d0, d19, d17
+ .long 0xf35621b8 // vbsl d18, d22, d24
+ .long 0xf3051db4 // vmul.f32 d1, d21, d20
+ .long 0xf26b6da1 // vsub.f32 d22, d27, d17
+ .long 0xf2627da1 // vsub.f32 d23, d18, d17
+ .long 0xf3633ea9 // vcgt.f32 d19, d19, d25
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xf3052db6 // vmul.f32 d2, d21, d22
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xf3058db7 // vmul.f32 d8, d21, d23
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2414d89 // vadd.f32 d20, d17, d9
+ .long 0xf26381b3 // vorr d24, d19, d19
+ .long 0xf2415d81 // vadd.f32 d21, d17, d1
+ .long 0xf2411d82 // vadd.f32 d17, d17, d2
+ .long 0xf35481ba // vbsl d24, d20, d26
+ .long 0xf26341b3 // vorr d20, d19, d19
+ .long 0xf3436d17 // vmul.f32 d22, d3, d7
+ .long 0xf2437d07 // vadd.f32 d23, d3, d7
+ .long 0xf35131b2 // vbsl d19, d17, d18
+ .long 0xf35541bb // vbsl d20, d21, d27
+ .long 0xf2481fa0 // vmax.f32 d17, d24, d16
+ .long 0xf2442fa0 // vmax.f32 d18, d20, d16
+ .long 0xf2430fa0 // vmax.f32 d16, d19, d16
+ .long 0xf2273da6 // vsub.f32 d3, d23, d22
+ .long 0xf3030d31 // vmul.f32 d0, d3, d17
+ .long 0xf3031d32 // vmul.f32 d1, d3, d18
+ .long 0xf3032d30 // vmul.f32 d2, d3, d16
+ .long 0xecbd8b04 // vpop {d8-d9}
+ .long 0xe12fff13 // bx r3
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3de147ae // .word 0x3de147ae
+ .long 0x3de147ae // .word 0x3de147ae
+
+HIDDEN _sk_luminosity_vfp4
+.globl _sk_luminosity_vfp4
+FUNCTION(_sk_luminosity_vfp4)
+_sk_luminosity_vfp4:
+ .long 0xed2d8b04 // vpush {d8-d9}
+ .long 0xeeb78a00 // vmov.f32 s16, #112
+ .long 0xf3f91507 // vceq.f32 d17, d7, #0
+ .long 0xf2c00010 // vmov.i32 d16, #0
+ .long 0xeddf2b58 // vldr d18, [pc, #352]
+ .long 0xeddf4b59 // vldr d20, [pc, #356]
+ .long 0xf3416d32 // vmul.f32 d22, d1, d18
+ .long 0xeddf8b59 // vldr d24, [pc, #356]
+ .long 0xf3407d34 // vmul.f32 d23, d0, d20
+ .long 0xe4913004 // ldr r3, [r1], #4
+ .long 0xf342bd38 // vmul.f32 d27, d2, d24
+ .long 0xeec89a27 // vdiv.f32 s19, s16, s15
+ .long 0xee889a07 // vdiv.f32 s18, s16, s14
+ .long 0xf3501199 // vbsl d17, d16, d9
+ .long 0xf2476da6 // vadd.f32 d22, d23, d22
+ .long 0xf3413d95 // vmul.f32 d19, d17, d5
+ .long 0xf3415d94 // vmul.f32 d21, d17, d4
+ .long 0xf3411d96 // vmul.f32 d17, d17, d6
+ .long 0xf2466dab // vadd.f32 d22, d22, d27
+ .long 0xf3439db2 // vmul.f32 d25, d19, d18
+ .long 0xf345adb4 // vmul.f32 d26, d21, d20
+ .long 0xf3417db8 // vmul.f32 d23, d17, d24
+ .long 0xf24a9da9 // vadd.f32 d25, d26, d25
+ .long 0xf2477da9 // vadd.f32 d23, d23, d25
+ .long 0xf2666da7 // vsub.f32 d22, d22, d23
+ .long 0xf2433da6 // vadd.f32 d19, d19, d22
+ .long 0xf2455da6 // vadd.f32 d21, d21, d22
+ .long 0xf2416da6 // vadd.f32 d22, d17, d22
+ .long 0xf3431db2 // vmul.f32 d17, d19, d18
+ .long 0xf3452db4 // vmul.f32 d18, d21, d20
+ .long 0xf3464db8 // vmul.f32 d20, d22, d24
+ .long 0xf2421da1 // vadd.f32 d17, d18, d17
+ .long 0xf2632fa6 // vmin.f32 d18, d19, d22
+ .long 0xf2441da1 // vadd.f32 d17, d20, d17
+ .long 0xf2652fa2 // vmin.f32 d18, d21, d18
+ .long 0xf2634da1 // vsub.f32 d20, d19, d17
+ .long 0xf2210da2 // vsub.f32 d0, d17, d18
+ .long 0xf3f924a2 // vcge.f32 d18, d18, #0
+ .long 0xf2667da1 // vsub.f32 d23, d22, d17
+ .long 0xf2658da1 // vsub.f32 d24, d21, d17
+ .long 0xf262a1b2 // vorr d26, d18, d18
+ .long 0xf3011db4 // vmul.f32 d1, d17, d20
+ .long 0xf2434fa6 // vmax.f32 d20, d19, d22
+ .long 0xf3012db7 // vmul.f32 d2, d17, d23
+ .long 0xf3018db8 // vmul.f32 d8, d17, d24
+ .long 0xf262b1b2 // vorr d27, d18, d18
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2419d82 // vadd.f32 d25, d17, d2
+ .long 0xf355a1b9 // vbsl d26, d21, d25
+ .long 0xf2417d89 // vadd.f32 d23, d17, d9
+ .long 0xf2c79f10 // vmov.f32 d25, #1
+ .long 0xf353b1b7 // vbsl d27, d19, d23
+ .long 0xf2453fa4 // vmax.f32 d19, d21, d20
+ .long 0xf26a4da1 // vsub.f32 d20, d26, d17
+ .long 0xf2695da1 // vsub.f32 d21, d25, d17
+ .long 0xf2418d81 // vadd.f32 d24, d17, d1
+ .long 0xf2230da1 // vsub.f32 d0, d19, d17
+ .long 0xf35621b8 // vbsl d18, d22, d24
+ .long 0xf3051db4 // vmul.f32 d1, d21, d20
+ .long 0xf26b6da1 // vsub.f32 d22, d27, d17
+ .long 0xf2627da1 // vsub.f32 d23, d18, d17
+ .long 0xf3633ea9 // vcgt.f32 d19, d19, d25
+ .long 0xeec19aa0 // vdiv.f32 s19, s3, s1
+ .long 0xf3052db6 // vmul.f32 d2, d21, d22
+ .long 0xee819a00 // vdiv.f32 s18, s2, s0
+ .long 0xeec21aa0 // vdiv.f32 s3, s5, s1
+ .long 0xf3058db7 // vmul.f32 d8, d21, d23
+ .long 0xee821a00 // vdiv.f32 s2, s4, s0
+ .long 0xeec82aa0 // vdiv.f32 s5, s17, s1
+ .long 0xee882a00 // vdiv.f32 s4, s16, s0
+ .long 0xf2414d89 // vadd.f32 d20, d17, d9
+ .long 0xf26381b3 // vorr d24, d19, d19
+ .long 0xf2415d81 // vadd.f32 d21, d17, d1
+ .long 0xf2411d82 // vadd.f32 d17, d17, d2
+ .long 0xf35481ba // vbsl d24, d20, d26
+ .long 0xf26341b3 // vorr d20, d19, d19
+ .long 0xf3436d17 // vmul.f32 d22, d3, d7
+ .long 0xf2437d07 // vadd.f32 d23, d3, d7
+ .long 0xf35131b2 // vbsl d19, d17, d18
+ .long 0xf35541bb // vbsl d20, d21, d27
+ .long 0xf2481fa0 // vmax.f32 d17, d24, d16
+ .long 0xf2442fa0 // vmax.f32 d18, d20, d16
+ .long 0xf2430fa0 // vmax.f32 d16, d19, d16
+ .long 0xf2273da6 // vsub.f32 d3, d23, d22
+ .long 0xf3030d31 // vmul.f32 d0, d3, d17
+ .long 0xf3031d32 // vmul.f32 d1, d3, d18
+ .long 0xf3032d30 // vmul.f32 d2, d3, d16
+ .long 0xecbd8b04 // vpop {d8-d9}
+ .long 0xe12fff13 // bx r3
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3f170a3d // .word 0x3f170a3d
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3e99999a // .word 0x3e99999a
+ .long 0x3de147ae // .word 0x3de147ae
+ .long 0x3de147ae // .word 0x3de147ae
+
HIDDEN _sk_clamp_0_vfp4
.globl _sk_clamp_0_vfp4
FUNCTION(_sk_clamp_0_vfp4)
@@ -6946,7 +7824,7 @@ _sk_linear_gradient_vfp4:
.long 0xe494c00c // ldr ip, [r4], #12
.long 0xf4a41c9f // vld1.32 {d1[]}, [r4 :32]
.long 0xe35c0000 // cmp ip, #0
- .long 0x0a000036 // beq 2e38 <sk_linear_gradient_vfp4+0x110>
+ .long 0x0a000036 // beq 3558 <sk_linear_gradient_vfp4+0x110>
.long 0xe59e3004 // ldr r3, [lr, #4]
.long 0xf2c01010 // vmov.i32 d17, #0
.long 0xf2c07010 // vmov.i32 d23, #0
@@ -6996,12 +7874,12 @@ _sk_linear_gradient_vfp4:
.long 0xf26371b3 // vorr d23, d19, d19
.long 0xf26481b4 // vorr d24, d20, d20
.long 0xf26561b5 // vorr d22, d21, d21
- .long 0x1affffd3 // bne 2d74 <sk_linear_gradient_vfp4+0x4c>
+ .long 0x1affffd3 // bne 3494 <sk_linear_gradient_vfp4+0x4c>
.long 0xf26c01bc // vorr d16, d28, d28
.long 0xf22b11bb // vorr d1, d27, d27
.long 0xf22a21ba // vorr d2, d26, d26
.long 0xf22931b9 // vorr d3, d25, d25
- .long 0xea000003 // b 2e48 <sk_linear_gradient_vfp4+0x120>
+ .long 0xea000003 // b 3568 <sk_linear_gradient_vfp4+0x120>
.long 0xf2c05010 // vmov.i32 d21, #0
.long 0xf2c04010 // vmov.i32 d20, #0
.long 0xf2c03010 // vmov.i32 d19, #0
@@ -7534,14 +8412,14 @@ _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 196,226,125,24,13,169,61,0,0 // vbroadcastss 0x3da9(%rip),%ymm1 # 3e6c <_sk_callback_hsw+0x127>
+ .byte 196,226,125,24,13,41,68,0,0 // vbroadcastss 0x4429(%rip),%ymm1 # 44ec <_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 196,226,125,24,21,141,61,0,0 // vbroadcastss 0x3d8d(%rip),%ymm2 # 3e70 <_sk_callback_hsw+0x12b>
+ .byte 196,226,125,24,21,13,68,0,0 // vbroadcastss 0x440d(%rip),%ymm2 # 44f0 <_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
@@ -7562,13 +8440,13 @@ _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 196,98,125,88,21,76,61,0,0 // vpbroadcastd 0x3d4c(%rip),%ymm10 # 3e74 <_sk_callback_hsw+0x12f>
+ .byte 196,98,125,88,21,204,67,0,0 // vpbroadcastd 0x43cc(%rip),%ymm10 # 44f4 <_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 196,98,125,88,37,49,61,0,0 // vpbroadcastd 0x3d31(%rip),%ymm12 # 3e78 <_sk_callback_hsw+0x133>
- .byte 196,98,125,88,45,44,61,0,0 // vpbroadcastd 0x3d2c(%rip),%ymm13 # 3e7c <_sk_callback_hsw+0x137>
+ .byte 196,98,125,88,37,177,67,0,0 // vpbroadcastd 0x43b1(%rip),%ymm12 # 44f8 <_sk_callback_hsw+0x133>
+ .byte 196,98,125,88,45,172,67,0,0 // vpbroadcastd 0x43ac(%rip),%ymm13 # 44fc <_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
@@ -7583,8 +8461,8 @@ _sk_dither_hsw:
.byte 196,65,61,235,194 // vpor %ymm10,%ymm8,%ymm8
.byte 196,65,61,235,193 // vpor %ymm9,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,222,60,0,0 // vbroadcastss 0x3cde(%rip),%ymm9 # 3e80 <_sk_callback_hsw+0x13b>
- .byte 196,98,125,24,21,217,60,0,0 // vbroadcastss 0x3cd9(%rip),%ymm10 # 3e84 <_sk_callback_hsw+0x13f>
+ .byte 196,98,125,24,13,94,67,0,0 // vbroadcastss 0x435e(%rip),%ymm9 # 4500 <_sk_callback_hsw+0x13b>
+ .byte 196,98,125,24,21,89,67,0,0 // vbroadcastss 0x4359(%rip),%ymm10 # 4504 <_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
@@ -7622,7 +8500,7 @@ HIDDEN _sk_srcatop_hsw
FUNCTION(_sk_srcatop_hsw)
_sk_srcatop_hsw:
.byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
- .byte 196,98,125,24,5,127,60,0,0 // vbroadcastss 0x3c7f(%rip),%ymm8 # 3e88 <_sk_callback_hsw+0x143>
+ .byte 196,98,125,24,5,255,66,0,0 // vbroadcastss 0x42ff(%rip),%ymm8 # 4508 <_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
@@ -7638,7 +8516,7 @@ HIDDEN _sk_dstatop_hsw
.globl _sk_dstatop_hsw
FUNCTION(_sk_dstatop_hsw)
_sk_dstatop_hsw:
- .byte 196,98,125,24,5,82,60,0,0 // vbroadcastss 0x3c52(%rip),%ymm8 # 3e8c <_sk_callback_hsw+0x147>
+ .byte 196,98,125,24,5,210,66,0,0 // vbroadcastss 0x42d2(%rip),%ymm8 # 450c <_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
@@ -7677,7 +8555,7 @@ HIDDEN _sk_srcout_hsw
.globl _sk_srcout_hsw
FUNCTION(_sk_srcout_hsw)
_sk_srcout_hsw:
- .byte 196,98,125,24,5,249,59,0,0 // vbroadcastss 0x3bf9(%rip),%ymm8 # 3e90 <_sk_callback_hsw+0x14b>
+ .byte 196,98,125,24,5,121,66,0,0 // vbroadcastss 0x4279(%rip),%ymm8 # 4510 <_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
@@ -7690,7 +8568,7 @@ HIDDEN _sk_dstout_hsw
.globl _sk_dstout_hsw
FUNCTION(_sk_dstout_hsw)
_sk_dstout_hsw:
- .byte 196,226,125,24,5,220,59,0,0 // vbroadcastss 0x3bdc(%rip),%ymm0 # 3e94 <_sk_callback_hsw+0x14f>
+ .byte 196,226,125,24,5,92,66,0,0 // vbroadcastss 0x425c(%rip),%ymm0 # 4514 <_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
@@ -7703,7 +8581,7 @@ HIDDEN _sk_srcover_hsw
.globl _sk_srcover_hsw
FUNCTION(_sk_srcover_hsw)
_sk_srcover_hsw:
- .byte 196,98,125,24,5,191,59,0,0 // vbroadcastss 0x3bbf(%rip),%ymm8 # 3e98 <_sk_callback_hsw+0x153>
+ .byte 196,98,125,24,5,63,66,0,0 // vbroadcastss 0x423f(%rip),%ymm8 # 4518 <_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
@@ -7716,7 +8594,7 @@ HIDDEN _sk_dstover_hsw
.globl _sk_dstover_hsw
FUNCTION(_sk_dstover_hsw)
_sk_dstover_hsw:
- .byte 196,98,125,24,5,158,59,0,0 // vbroadcastss 0x3b9e(%rip),%ymm8 # 3e9c <_sk_callback_hsw+0x157>
+ .byte 196,98,125,24,5,30,66,0,0 // vbroadcastss 0x421e(%rip),%ymm8 # 451c <_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
@@ -7740,7 +8618,7 @@ HIDDEN _sk_multiply_hsw
.globl _sk_multiply_hsw
FUNCTION(_sk_multiply_hsw)
_sk_multiply_hsw:
- .byte 196,98,125,24,5,105,59,0,0 // vbroadcastss 0x3b69(%rip),%ymm8 # 3ea0 <_sk_callback_hsw+0x15b>
+ .byte 196,98,125,24,5,233,65,0,0 // vbroadcastss 0x41e9(%rip),%ymm8 # 4520 <_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
@@ -7788,7 +8666,7 @@ HIDDEN _sk_xor__hsw
.globl _sk_xor__hsw
FUNCTION(_sk_xor__hsw)
_sk_xor__hsw:
- .byte 196,98,125,24,5,228,58,0,0 // vbroadcastss 0x3ae4(%rip),%ymm8 # 3ea4 <_sk_callback_hsw+0x15f>
+ .byte 196,98,125,24,5,100,65,0,0 // vbroadcastss 0x4164(%rip),%ymm8 # 4524 <_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
@@ -7822,7 +8700,7 @@ _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 196,98,125,24,5,108,58,0,0 // vbroadcastss 0x3a6c(%rip),%ymm8 # 3ea8 <_sk_callback_hsw+0x163>
+ .byte 196,98,125,24,5,236,64,0,0 // vbroadcastss 0x40ec(%rip),%ymm8 # 4528 <_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
@@ -7847,7 +8725,7 @@ _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 196,98,125,24,5,27,58,0,0 // vbroadcastss 0x3a1b(%rip),%ymm8 # 3eac <_sk_callback_hsw+0x167>
+ .byte 196,98,125,24,5,155,64,0,0 // vbroadcastss 0x409b(%rip),%ymm8 # 452c <_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
@@ -7875,7 +8753,7 @@ _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 196,98,125,24,5,190,57,0,0 // vbroadcastss 0x39be(%rip),%ymm8 # 3eb0 <_sk_callback_hsw+0x16b>
+ .byte 196,98,125,24,5,62,64,0,0 // vbroadcastss 0x403e(%rip),%ymm8 # 4530 <_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
@@ -7897,7 +8775,7 @@ _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 196,98,125,24,5,124,57,0,0 // vbroadcastss 0x397c(%rip),%ymm8 # 3eb4 <_sk_callback_hsw+0x16f>
+ .byte 196,98,125,24,5,252,63,0,0 // vbroadcastss 0x3ffc(%rip),%ymm8 # 4534 <_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
@@ -7907,7 +8785,7 @@ HIDDEN _sk_colorburn_hsw
.globl _sk_colorburn_hsw
FUNCTION(_sk_colorburn_hsw)
_sk_colorburn_hsw:
- .byte 196,98,125,24,5,106,57,0,0 // vbroadcastss 0x396a(%rip),%ymm8 # 3eb8 <_sk_callback_hsw+0x173>
+ .byte 196,98,125,24,5,234,63,0,0 // vbroadcastss 0x3fea(%rip),%ymm8 # 4538 <_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
@@ -7965,7 +8843,7 @@ HIDDEN _sk_colordodge_hsw
FUNCTION(_sk_colordodge_hsw)
_sk_colordodge_hsw:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
- .byte 196,98,125,24,13,117,56,0,0 // vbroadcastss 0x3875(%rip),%ymm9 # 3ebc <_sk_callback_hsw+0x177>
+ .byte 196,98,125,24,13,245,62,0,0 // vbroadcastss 0x3ef5(%rip),%ymm9 # 453c <_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
@@ -8018,7 +8896,7 @@ HIDDEN _sk_hardlight_hsw
.globl _sk_hardlight_hsw
FUNCTION(_sk_hardlight_hsw)
_sk_hardlight_hsw:
- .byte 196,98,125,24,5,150,55,0,0 // vbroadcastss 0x3796(%rip),%ymm8 # 3ec0 <_sk_callback_hsw+0x17b>
+ .byte 196,98,125,24,5,22,62,0,0 // vbroadcastss 0x3e16(%rip),%ymm8 # 4540 <_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
@@ -8069,7 +8947,7 @@ HIDDEN _sk_overlay_hsw
.globl _sk_overlay_hsw
FUNCTION(_sk_overlay_hsw)
_sk_overlay_hsw:
- .byte 196,98,125,24,5,206,54,0,0 // vbroadcastss 0x36ce(%rip),%ymm8 # 3ec4 <_sk_callback_hsw+0x17f>
+ .byte 196,98,125,24,5,78,61,0,0 // vbroadcastss 0x3d4e(%rip),%ymm8 # 4544 <_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
@@ -8130,10 +9008,10 @@ _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 196,98,125,24,29,217,53,0,0 // vbroadcastss 0x35d9(%rip),%ymm11 # 3ecc <_sk_callback_hsw+0x187>
+ .byte 196,98,125,24,29,89,60,0,0 // vbroadcastss 0x3c59(%rip),%ymm11 # 454c <_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 196,98,125,24,37,202,53,0,0 // vbroadcastss 0x35ca(%rip),%ymm12 # 3ed0 <_sk_callback_hsw+0x18b>
+ .byte 196,98,125,24,37,74,60,0,0 // vbroadcastss 0x3c4a(%rip),%ymm12 # 4550 <_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
@@ -8143,7 +9021,7 @@ _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 196,98,125,24,5,141,53,0,0 // vbroadcastss 0x358d(%rip),%ymm8 # 3ec8 <_sk_callback_hsw+0x183>
+ .byte 196,98,125,24,5,13,60,0,0 // vbroadcastss 0x3c0d(%rip),%ymm8 # 4548 <_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
@@ -8220,6 +9098,364 @@ _sk_softlight_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_hue_hsw
+.globl _sk_hue_hsw
+FUNCTION(_sk_hue_hsw)
+_sk_hue_hsw:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,100,194,208,0 // vcmpeqps %ymm8,%ymm3,%ymm10
+ .byte 196,98,125,24,13,164,58,0,0 // vbroadcastss 0x3aa4(%rip),%ymm9 # 4554 <_sk_callback_hsw+0x18f>
+ .byte 197,52,94,219 // vdivps %ymm3,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
+ .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
+ .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
+ .byte 197,84,95,214 // vmaxps %ymm6,%ymm5,%ymm10
+ .byte 196,65,92,95,210 // vmaxps %ymm10,%ymm4,%ymm10
+ .byte 197,84,93,222 // vminps %ymm6,%ymm5,%ymm11
+ .byte 196,65,92,93,219 // vminps %ymm11,%ymm4,%ymm11
+ .byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10
+ .byte 197,116,93,218 // vminps %ymm2,%ymm1,%ymm11
+ .byte 196,65,124,93,219 // vminps %ymm11,%ymm0,%ymm11
+ .byte 197,116,95,226 // vmaxps %ymm2,%ymm1,%ymm12
+ .byte 196,65,124,95,228 // vmaxps %ymm12,%ymm0,%ymm12
+ .byte 196,65,28,92,227 // vsubps %ymm11,%ymm12,%ymm12
+ .byte 196,65,28,194,232,0 // vcmpeqps %ymm8,%ymm12,%ymm13
+ .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0
+ .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
+ .byte 196,193,124,94,196 // vdivps %ymm12,%ymm0,%ymm0
+ .byte 196,195,125,74,192,208 // vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1
+ .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
+ .byte 196,193,116,94,204 // vdivps %ymm12,%ymm1,%ymm1
+ .byte 196,195,117,74,200,208 // vblendvps %ymm13,%ymm8,%ymm1,%ymm1
+ .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2
+ .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
+ .byte 196,193,108,94,212 // vdivps %ymm12,%ymm2,%ymm2
+ .byte 196,195,109,74,208,208 // vblendvps %ymm13,%ymm8,%ymm2,%ymm2
+ .byte 196,98,125,24,21,25,58,0,0 // vbroadcastss 0x3a19(%rip),%ymm10 # 4558 <_sk_callback_hsw+0x193>
+ .byte 196,98,125,24,29,20,58,0,0 // vbroadcastss 0x3a14(%rip),%ymm11 # 455c <_sk_callback_hsw+0x197>
+ .byte 196,65,84,89,227 // vmulps %ymm11,%ymm5,%ymm12
+ .byte 196,66,93,184,226 // vfmadd231ps %ymm10,%ymm4,%ymm12
+ .byte 196,98,125,24,45,5,58,0,0 // vbroadcastss 0x3a05(%rip),%ymm13 # 4560 <_sk_callback_hsw+0x19b>
+ .byte 196,66,77,184,229 // vfmadd231ps %ymm13,%ymm6,%ymm12
+ .byte 196,65,116,89,243 // vmulps %ymm11,%ymm1,%ymm14
+ .byte 196,66,125,184,242 // vfmadd231ps %ymm10,%ymm0,%ymm14
+ .byte 196,66,109,184,245 // vfmadd231ps %ymm13,%ymm2,%ymm14
+ .byte 196,65,28,92,230 // vsubps %ymm14,%ymm12,%ymm12
+ .byte 196,65,124,88,244 // vaddps %ymm12,%ymm0,%ymm14
+ .byte 196,65,116,88,252 // vaddps %ymm12,%ymm1,%ymm15
+ .byte 196,193,108,88,196 // vaddps %ymm12,%ymm2,%ymm0
+ .byte 197,132,93,200 // vminps %ymm0,%ymm15,%ymm1
+ .byte 197,140,93,209 // vminps %ymm1,%ymm14,%ymm2
+ .byte 197,132,95,200 // vmaxps %ymm0,%ymm15,%ymm1
+ .byte 197,12,95,225 // vmaxps %ymm1,%ymm14,%ymm12
+ .byte 196,193,4,89,203 // vmulps %ymm11,%ymm15,%ymm1
+ .byte 196,194,13,184,202 // vfmadd231ps %ymm10,%ymm14,%ymm1
+ .byte 196,194,125,184,205 // vfmadd231ps %ymm13,%ymm0,%ymm1
+ .byte 197,60,194,210,2 // vcmpleps %ymm2,%ymm8,%ymm10
+ .byte 197,12,92,217 // vsubps %ymm1,%ymm14,%ymm11
+ .byte 196,65,116,89,219 // vmulps %ymm11,%ymm1,%ymm11
+ .byte 197,244,92,210 // vsubps %ymm2,%ymm1,%ymm2
+ .byte 197,36,94,218 // vdivps %ymm2,%ymm11,%ymm11
+ .byte 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11
+ .byte 196,67,37,74,222,160 // vblendvps %ymm10,%ymm14,%ymm11,%ymm11
+ .byte 196,65,52,194,236,1 // vcmpltps %ymm12,%ymm9,%ymm13
+ .byte 197,36,92,241 // vsubps %ymm1,%ymm11,%ymm14
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,246 // vmulps %ymm14,%ymm9,%ymm14
+ .byte 197,28,92,225 // vsubps %ymm1,%ymm12,%ymm12
+ .byte 196,65,12,94,244 // vdivps %ymm12,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,37,74,222,208 // vblendvps %ymm13,%ymm14,%ymm11,%ymm11
+ .byte 196,65,36,95,216 // vmaxps %ymm8,%ymm11,%ymm11
+ .byte 197,4,92,241 // vsubps %ymm1,%ymm15,%ymm14
+ .byte 196,65,116,89,246 // vmulps %ymm14,%ymm1,%ymm14
+ .byte 197,12,94,242 // vdivps %ymm2,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,13,74,247,160 // vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ .byte 197,12,92,249 // vsubps %ymm1,%ymm14,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,252 // vdivps %ymm12,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,13,74,247,208 // vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ .byte 196,65,12,95,240 // vmaxps %ymm8,%ymm14,%ymm14
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 197,132,94,210 // vdivps %ymm2,%ymm15,%ymm2
+ .byte 197,244,88,210 // vaddps %ymm2,%ymm1,%ymm2
+ .byte 196,227,109,74,192,160 // vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
+ .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
+ .byte 196,193,108,94,212 // vdivps %ymm12,%ymm2,%ymm2
+ .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
+ .byte 196,227,125,74,193,208 // vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ .byte 196,193,124,95,208 // vmaxps %ymm8,%ymm0,%ymm2
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 196,226,69,172,216 // vfnmadd213ps %ymm0,%ymm7,%ymm3
+ .byte 196,193,100,89,195 // vmulps %ymm11,%ymm3,%ymm0
+ .byte 196,193,100,89,206 // vmulps %ymm14,%ymm3,%ymm1
+ .byte 197,228,89,210 // vmulps %ymm2,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_saturation_hsw
+.globl _sk_saturation_hsw
+FUNCTION(_sk_saturation_hsw)
+_sk_saturation_hsw:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,68,194,208,0 // vcmpeqps %ymm8,%ymm7,%ymm10
+ .byte 196,98,125,24,13,221,56,0,0 // vbroadcastss 0x38dd(%rip),%ymm9 # 4564 <_sk_callback_hsw+0x19f>
+ .byte 197,52,94,223 // vdivps %ymm7,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,44,89,220 // vmulps %ymm4,%ymm10,%ymm11
+ .byte 197,44,89,229 // vmulps %ymm5,%ymm10,%ymm12
+ .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10
+ .byte 197,116,95,234 // vmaxps %ymm2,%ymm1,%ymm13
+ .byte 196,65,124,95,237 // vmaxps %ymm13,%ymm0,%ymm13
+ .byte 197,244,93,202 // vminps %ymm2,%ymm1,%ymm1
+ .byte 197,252,93,193 // vminps %ymm1,%ymm0,%ymm0
+ .byte 197,148,92,192 // vsubps %ymm0,%ymm13,%ymm0
+ .byte 196,193,28,93,202 // vminps %ymm10,%ymm12,%ymm1
+ .byte 197,164,93,201 // vminps %ymm1,%ymm11,%ymm1
+ .byte 196,193,28,95,210 // vmaxps %ymm10,%ymm12,%ymm2
+ .byte 197,164,95,210 // vmaxps %ymm2,%ymm11,%ymm2
+ .byte 197,236,92,209 // vsubps %ymm1,%ymm2,%ymm2
+ .byte 196,65,108,194,232,0 // vcmpeqps %ymm8,%ymm2,%ymm13
+ .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11
+ .byte 196,65,124,89,219 // vmulps %ymm11,%ymm0,%ymm11
+ .byte 197,36,94,218 // vdivps %ymm2,%ymm11,%ymm11
+ .byte 196,67,37,74,216,208 // vblendvps %ymm13,%ymm8,%ymm11,%ymm11
+ .byte 197,28,92,225 // vsubps %ymm1,%ymm12,%ymm12
+ .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12
+ .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12
+ .byte 196,67,29,74,224,208 // vblendvps %ymm13,%ymm8,%ymm12,%ymm12
+ .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1
+ .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
+ .byte 197,252,94,194 // vdivps %ymm2,%ymm0,%ymm0
+ .byte 196,195,125,74,192,208 // vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ .byte 196,226,125,24,21,89,56,0,0 // vbroadcastss 0x3859(%rip),%ymm2 # 4568 <_sk_callback_hsw+0x1a3>
+ .byte 196,226,125,24,13,84,56,0,0 // vbroadcastss 0x3854(%rip),%ymm1 # 456c <_sk_callback_hsw+0x1a7>
+ .byte 197,84,89,209 // vmulps %ymm1,%ymm5,%ymm10
+ .byte 196,98,93,184,210 // vfmadd231ps %ymm2,%ymm4,%ymm10
+ .byte 196,98,125,24,45,70,56,0,0 // vbroadcastss 0x3846(%rip),%ymm13 # 4570 <_sk_callback_hsw+0x1ab>
+ .byte 196,66,77,184,213 // vfmadd231ps %ymm13,%ymm6,%ymm10
+ .byte 197,28,89,241 // vmulps %ymm1,%ymm12,%ymm14
+ .byte 196,98,37,184,242 // vfmadd231ps %ymm2,%ymm11,%ymm14
+ .byte 196,66,125,184,245 // vfmadd231ps %ymm13,%ymm0,%ymm14
+ .byte 196,65,44,92,214 // vsubps %ymm14,%ymm10,%ymm10
+ .byte 196,65,36,88,218 // vaddps %ymm10,%ymm11,%ymm11
+ .byte 196,65,28,88,226 // vaddps %ymm10,%ymm12,%ymm12
+ .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
+ .byte 197,28,93,208 // vminps %ymm0,%ymm12,%ymm10
+ .byte 196,65,36,93,210 // vminps %ymm10,%ymm11,%ymm10
+ .byte 197,28,95,240 // vmaxps %ymm0,%ymm12,%ymm14
+ .byte 196,65,36,95,246 // vmaxps %ymm14,%ymm11,%ymm14
+ .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1
+ .byte 196,226,37,184,202 // vfmadd231ps %ymm2,%ymm11,%ymm1
+ .byte 196,194,125,184,205 // vfmadd231ps %ymm13,%ymm0,%ymm1
+ .byte 196,193,60,194,210,2 // vcmpleps %ymm10,%ymm8,%ymm2
+ .byte 197,36,92,233 // vsubps %ymm1,%ymm11,%ymm13
+ .byte 196,65,116,89,237 // vmulps %ymm13,%ymm1,%ymm13
+ .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
+ .byte 196,65,20,94,234 // vdivps %ymm10,%ymm13,%ymm13
+ .byte 196,65,116,88,237 // vaddps %ymm13,%ymm1,%ymm13
+ .byte 196,67,21,74,219,32 // vblendvps %ymm2,%ymm11,%ymm13,%ymm11
+ .byte 196,65,52,194,238,1 // vcmpltps %ymm14,%ymm9,%ymm13
+ .byte 197,36,92,249 // vsubps %ymm1,%ymm11,%ymm15
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 197,12,92,241 // vsubps %ymm1,%ymm14,%ymm14
+ .byte 196,65,4,94,254 // vdivps %ymm14,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,37,74,223,208 // vblendvps %ymm13,%ymm15,%ymm11,%ymm11
+ .byte 196,65,36,95,216 // vmaxps %ymm8,%ymm11,%ymm11
+ .byte 197,28,92,249 // vsubps %ymm1,%ymm12,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 196,65,4,94,250 // vdivps %ymm10,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,5,74,228,32 // vblendvps %ymm2,%ymm12,%ymm15,%ymm12
+ .byte 197,28,92,249 // vsubps %ymm1,%ymm12,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,254 // vdivps %ymm14,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,29,74,231,208 // vblendvps %ymm13,%ymm15,%ymm12,%ymm12
+ .byte 196,65,28,95,224 // vmaxps %ymm8,%ymm12,%ymm12
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 196,65,4,94,210 // vdivps %ymm10,%ymm15,%ymm10
+ .byte 196,65,116,88,210 // vaddps %ymm10,%ymm1,%ymm10
+ .byte 196,227,45,74,192,32 // vblendvps %ymm2,%ymm0,%ymm10,%ymm0
+ .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
+ .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
+ .byte 196,193,108,94,214 // vdivps %ymm14,%ymm2,%ymm2
+ .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
+ .byte 196,227,125,74,193,208 // vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ .byte 196,193,124,95,208 // vmaxps %ymm8,%ymm0,%ymm2
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 196,226,69,172,216 // vfnmadd213ps %ymm0,%ymm7,%ymm3
+ .byte 196,193,100,89,195 // vmulps %ymm11,%ymm3,%ymm0
+ .byte 196,193,100,89,204 // vmulps %ymm12,%ymm3,%ymm1
+ .byte 197,228,89,210 // vmulps %ymm2,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_color_hsw
+.globl _sk_color_hsw
+FUNCTION(_sk_color_hsw)
+_sk_color_hsw:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,100,194,208,0 // vcmpeqps %ymm8,%ymm3,%ymm10
+ .byte 196,98,125,24,13,24,55,0,0 // vbroadcastss 0x3718(%rip),%ymm9 # 4574 <_sk_callback_hsw+0x1af>
+ .byte 197,52,94,219 // vdivps %ymm3,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
+ .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
+ .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
+ .byte 196,98,125,24,21,253,54,0,0 // vbroadcastss 0x36fd(%rip),%ymm10 # 4578 <_sk_callback_hsw+0x1b3>
+ .byte 196,98,125,24,29,248,54,0,0 // vbroadcastss 0x36f8(%rip),%ymm11 # 457c <_sk_callback_hsw+0x1b7>
+ .byte 196,65,84,89,227 // vmulps %ymm11,%ymm5,%ymm12
+ .byte 196,66,93,184,226 // vfmadd231ps %ymm10,%ymm4,%ymm12
+ .byte 196,98,125,24,45,233,54,0,0 // vbroadcastss 0x36e9(%rip),%ymm13 # 4580 <_sk_callback_hsw+0x1bb>
+ .byte 196,66,77,184,229 // vfmadd231ps %ymm13,%ymm6,%ymm12
+ .byte 196,65,116,89,243 // vmulps %ymm11,%ymm1,%ymm14
+ .byte 196,66,125,184,242 // vfmadd231ps %ymm10,%ymm0,%ymm14
+ .byte 196,66,109,184,245 // vfmadd231ps %ymm13,%ymm2,%ymm14
+ .byte 196,65,28,92,230 // vsubps %ymm14,%ymm12,%ymm12
+ .byte 196,65,124,88,244 // vaddps %ymm12,%ymm0,%ymm14
+ .byte 196,65,116,88,252 // vaddps %ymm12,%ymm1,%ymm15
+ .byte 196,193,108,88,196 // vaddps %ymm12,%ymm2,%ymm0
+ .byte 197,132,93,200 // vminps %ymm0,%ymm15,%ymm1
+ .byte 197,140,93,209 // vminps %ymm1,%ymm14,%ymm2
+ .byte 197,132,95,200 // vmaxps %ymm0,%ymm15,%ymm1
+ .byte 197,12,95,225 // vmaxps %ymm1,%ymm14,%ymm12
+ .byte 196,193,4,89,203 // vmulps %ymm11,%ymm15,%ymm1
+ .byte 196,194,13,184,202 // vfmadd231ps %ymm10,%ymm14,%ymm1
+ .byte 196,194,125,184,205 // vfmadd231ps %ymm13,%ymm0,%ymm1
+ .byte 197,60,194,210,2 // vcmpleps %ymm2,%ymm8,%ymm10
+ .byte 197,12,92,217 // vsubps %ymm1,%ymm14,%ymm11
+ .byte 196,65,116,89,219 // vmulps %ymm11,%ymm1,%ymm11
+ .byte 197,244,92,210 // vsubps %ymm2,%ymm1,%ymm2
+ .byte 197,36,94,218 // vdivps %ymm2,%ymm11,%ymm11
+ .byte 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11
+ .byte 196,67,37,74,222,160 // vblendvps %ymm10,%ymm14,%ymm11,%ymm11
+ .byte 196,65,52,194,236,1 // vcmpltps %ymm12,%ymm9,%ymm13
+ .byte 197,36,92,241 // vsubps %ymm1,%ymm11,%ymm14
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,246 // vmulps %ymm14,%ymm9,%ymm14
+ .byte 197,28,92,225 // vsubps %ymm1,%ymm12,%ymm12
+ .byte 196,65,12,94,244 // vdivps %ymm12,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,37,74,222,208 // vblendvps %ymm13,%ymm14,%ymm11,%ymm11
+ .byte 196,65,36,95,216 // vmaxps %ymm8,%ymm11,%ymm11
+ .byte 197,4,92,241 // vsubps %ymm1,%ymm15,%ymm14
+ .byte 196,65,116,89,246 // vmulps %ymm14,%ymm1,%ymm14
+ .byte 197,12,94,242 // vdivps %ymm2,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,13,74,247,160 // vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ .byte 197,12,92,249 // vsubps %ymm1,%ymm14,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,252 // vdivps %ymm12,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,13,74,247,208 // vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ .byte 196,65,12,95,240 // vmaxps %ymm8,%ymm14,%ymm14
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 197,132,94,210 // vdivps %ymm2,%ymm15,%ymm2
+ .byte 197,244,88,210 // vaddps %ymm2,%ymm1,%ymm2
+ .byte 196,227,109,74,192,160 // vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
+ .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
+ .byte 196,193,108,94,212 // vdivps %ymm12,%ymm2,%ymm2
+ .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
+ .byte 196,227,125,74,193,208 // vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ .byte 196,193,124,95,208 // vmaxps %ymm8,%ymm0,%ymm2
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 196,226,69,172,216 // vfnmadd213ps %ymm0,%ymm7,%ymm3
+ .byte 196,193,100,89,195 // vmulps %ymm11,%ymm3,%ymm0
+ .byte 196,193,100,89,206 // vmulps %ymm14,%ymm3,%ymm1
+ .byte 197,228,89,210 // vmulps %ymm2,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_luminosity_hsw
+.globl _sk_luminosity_hsw
+FUNCTION(_sk_luminosity_hsw)
+_sk_luminosity_hsw:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,68,194,208,0 // vcmpeqps %ymm8,%ymm7,%ymm10
+ .byte 196,98,125,24,13,193,53,0,0 // vbroadcastss 0x35c1(%rip),%ymm9 # 4584 <_sk_callback_hsw+0x1bf>
+ .byte 197,52,94,223 // vdivps %ymm7,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,44,89,220 // vmulps %ymm4,%ymm10,%ymm11
+ .byte 197,44,89,229 // vmulps %ymm5,%ymm10,%ymm12
+ .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10
+ .byte 196,98,125,24,45,166,53,0,0 // vbroadcastss 0x35a6(%rip),%ymm13 # 4588 <_sk_callback_hsw+0x1c3>
+ .byte 196,98,125,24,53,161,53,0,0 // vbroadcastss 0x35a1(%rip),%ymm14 # 458c <_sk_callback_hsw+0x1c7>
+ .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1
+ .byte 196,226,21,168,193 // vfmadd213ps %ymm1,%ymm13,%ymm0
+ .byte 196,98,125,24,61,146,53,0,0 // vbroadcastss 0x3592(%rip),%ymm15 # 4590 <_sk_callback_hsw+0x1cb>
+ .byte 196,226,5,168,208 // vfmadd213ps %ymm0,%ymm15,%ymm2
+ .byte 196,193,28,89,198 // vmulps %ymm14,%ymm12,%ymm0
+ .byte 196,194,37,184,197 // vfmadd231ps %ymm13,%ymm11,%ymm0
+ .byte 196,194,45,184,199 // vfmadd231ps %ymm15,%ymm10,%ymm0
+ .byte 197,236,92,192 // vsubps %ymm0,%ymm2,%ymm0
+ .byte 197,164,88,208 // vaddps %ymm0,%ymm11,%ymm2
+ .byte 197,28,88,216 // vaddps %ymm0,%ymm12,%ymm11
+ .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0
+ .byte 197,164,93,200 // vminps %ymm0,%ymm11,%ymm1
+ .byte 197,108,93,209 // vminps %ymm1,%ymm2,%ymm10
+ .byte 197,164,95,200 // vmaxps %ymm0,%ymm11,%ymm1
+ .byte 197,108,95,225 // vmaxps %ymm1,%ymm2,%ymm12
+ .byte 196,193,36,89,206 // vmulps %ymm14,%ymm11,%ymm1
+ .byte 196,194,109,184,205 // vfmadd231ps %ymm13,%ymm2,%ymm1
+ .byte 196,194,125,184,207 // vfmadd231ps %ymm15,%ymm0,%ymm1
+ .byte 196,65,60,194,234,2 // vcmpleps %ymm10,%ymm8,%ymm13
+ .byte 197,108,92,241 // vsubps %ymm1,%ymm2,%ymm14
+ .byte 196,65,116,89,246 // vmulps %ymm14,%ymm1,%ymm14
+ .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
+ .byte 196,65,12,94,242 // vdivps %ymm10,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2
+ .byte 196,65,52,194,244,1 // vcmpltps %ymm12,%ymm9,%ymm14
+ .byte 197,108,92,249 // vsubps %ymm1,%ymm2,%ymm15
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 197,28,92,225 // vsubps %ymm1,%ymm12,%ymm12
+ .byte 196,65,4,94,252 // vdivps %ymm12,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,195,109,74,215,224 // vblendvps %ymm14,%ymm15,%ymm2,%ymm2
+ .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
+ .byte 197,36,92,249 // vsubps %ymm1,%ymm11,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 196,65,4,94,250 // vdivps %ymm10,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,5,74,219,208 // vblendvps %ymm13,%ymm11,%ymm15,%ymm11
+ .byte 197,36,92,249 // vsubps %ymm1,%ymm11,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,252 // vdivps %ymm12,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,37,74,223,224 // vblendvps %ymm14,%ymm15,%ymm11,%ymm11
+ .byte 196,65,36,95,216 // vmaxps %ymm8,%ymm11,%ymm11
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 196,65,4,94,210 // vdivps %ymm10,%ymm15,%ymm10
+ .byte 196,65,116,88,210 // vaddps %ymm10,%ymm1,%ymm10
+ .byte 196,227,45,74,192,208 // vblendvps %ymm13,%ymm0,%ymm10,%ymm0
+ .byte 197,124,92,209 // vsubps %ymm1,%ymm0,%ymm10
+ .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
+ .byte 196,65,52,94,204 // vdivps %ymm12,%ymm9,%ymm9
+ .byte 196,193,116,88,201 // vaddps %ymm9,%ymm1,%ymm1
+ .byte 196,227,125,74,193,224 // vblendvps %ymm14,%ymm1,%ymm0,%ymm0
+ .byte 196,65,124,95,192 // vmaxps %ymm8,%ymm0,%ymm8
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 196,226,69,172,216 // vfnmadd213ps %ymm0,%ymm7,%ymm3
+ .byte 197,228,89,194 // vmulps %ymm2,%ymm3,%ymm0
+ .byte 196,193,100,89,203 // vmulps %ymm11,%ymm3,%ymm1
+ .byte 196,193,100,89,208 // vmulps %ymm8,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_clamp_0_hsw
.globl _sk_clamp_0_hsw
FUNCTION(_sk_clamp_0_hsw)
@@ -8236,7 +9472,7 @@ HIDDEN _sk_clamp_1_hsw
.globl _sk_clamp_1_hsw
FUNCTION(_sk_clamp_1_hsw)
_sk_clamp_1_hsw:
- .byte 196,98,125,24,5,18,52,0,0 // vbroadcastss 0x3412(%rip),%ymm8 # 3ed4 <_sk_callback_hsw+0x18f>
+ .byte 196,98,125,24,5,84,52,0,0 // vbroadcastss 0x3454(%rip),%ymm8 # 4594 <_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
@@ -8248,7 +9484,7 @@ HIDDEN _sk_clamp_a_hsw
.globl _sk_clamp_a_hsw
FUNCTION(_sk_clamp_a_hsw)
_sk_clamp_a_hsw:
- .byte 196,98,125,24,5,245,51,0,0 // vbroadcastss 0x33f5(%rip),%ymm8 # 3ed8 <_sk_callback_hsw+0x193>
+ .byte 196,98,125,24,5,55,52,0,0 // vbroadcastss 0x3437(%rip),%ymm8 # 4598 <_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
@@ -8334,7 +9570,7 @@ 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 196,98,125,24,21,61,51,0,0 // vbroadcastss 0x333d(%rip),%ymm10 # 3edc <_sk_callback_hsw+0x197>
+ .byte 196,98,125,24,21,127,51,0,0 // vbroadcastss 0x337f(%rip),%ymm10 # 459c <_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
@@ -8347,16 +9583,16 @@ HIDDEN _sk_from_srgb_hsw
.globl _sk_from_srgb_hsw
FUNCTION(_sk_from_srgb_hsw)
_sk_from_srgb_hsw:
- .byte 196,98,125,24,5,30,51,0,0 // vbroadcastss 0x331e(%rip),%ymm8 # 3ee0 <_sk_callback_hsw+0x19b>
+ .byte 196,98,125,24,5,96,51,0,0 // vbroadcastss 0x3360(%rip),%ymm8 # 45a0 <_sk_callback_hsw+0x1db>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
- .byte 196,98,125,24,29,16,51,0,0 // vbroadcastss 0x3310(%rip),%ymm11 # 3ee4 <_sk_callback_hsw+0x19f>
- .byte 196,98,125,24,37,11,51,0,0 // vbroadcastss 0x330b(%rip),%ymm12 # 3ee8 <_sk_callback_hsw+0x1a3>
+ .byte 196,98,125,24,29,82,51,0,0 // vbroadcastss 0x3352(%rip),%ymm11 # 45a4 <_sk_callback_hsw+0x1df>
+ .byte 196,98,125,24,37,77,51,0,0 // vbroadcastss 0x334d(%rip),%ymm12 # 45a8 <_sk_callback_hsw+0x1e3>
.byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13
.byte 196,66,125,168,235 // vfmadd213ps %ymm11,%ymm0,%ymm13
- .byte 196,98,125,24,53,252,50,0,0 // vbroadcastss 0x32fc(%rip),%ymm14 # 3eec <_sk_callback_hsw+0x1a7>
+ .byte 196,98,125,24,53,62,51,0,0 // vbroadcastss 0x333e(%rip),%ymm14 # 45ac <_sk_callback_hsw+0x1e7>
.byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13
- .byte 196,98,125,24,21,242,50,0,0 // vbroadcastss 0x32f2(%rip),%ymm10 # 3ef0 <_sk_callback_hsw+0x1ab>
+ .byte 196,98,125,24,21,52,51,0,0 // vbroadcastss 0x3334(%rip),%ymm10 # 45b0 <_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
@@ -8382,16 +9618,16 @@ _sk_to_srgb_hsw:
.byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8
.byte 196,65,124,83,200 // vrcpps %ymm8,%ymm9
.byte 196,65,124,82,208 // vrsqrtps %ymm8,%ymm10
- .byte 196,98,125,24,5,140,50,0,0 // vbroadcastss 0x328c(%rip),%ymm8 # 3ef4 <_sk_callback_hsw+0x1af>
+ .byte 196,98,125,24,5,206,50,0,0 // vbroadcastss 0x32ce(%rip),%ymm8 # 45b4 <_sk_callback_hsw+0x1ef>
.byte 196,65,124,89,216 // vmulps %ymm8,%ymm0,%ymm11
- .byte 196,98,125,24,37,130,50,0,0 // vbroadcastss 0x3282(%rip),%ymm12 # 3ef8 <_sk_callback_hsw+0x1b3>
- .byte 196,98,125,24,45,125,50,0,0 // vbroadcastss 0x327d(%rip),%ymm13 # 3efc <_sk_callback_hsw+0x1b7>
+ .byte 196,98,125,24,37,196,50,0,0 // vbroadcastss 0x32c4(%rip),%ymm12 # 45b8 <_sk_callback_hsw+0x1f3>
+ .byte 196,98,125,24,45,191,50,0,0 // vbroadcastss 0x32bf(%rip),%ymm13 # 45bc <_sk_callback_hsw+0x1f7>
.byte 196,66,21,168,204 // vfmadd213ps %ymm12,%ymm13,%ymm9
- .byte 196,98,125,24,53,115,50,0,0 // vbroadcastss 0x3273(%rip),%ymm14 # 3f00 <_sk_callback_hsw+0x1bb>
+ .byte 196,98,125,24,53,181,50,0,0 // vbroadcastss 0x32b5(%rip),%ymm14 # 45c0 <_sk_callback_hsw+0x1fb>
.byte 196,66,13,184,202 // vfmadd231ps %ymm10,%ymm14,%ymm9
- .byte 196,98,125,24,21,105,50,0,0 // vbroadcastss 0x3269(%rip),%ymm10 # 3f04 <_sk_callback_hsw+0x1bf>
+ .byte 196,98,125,24,21,171,50,0,0 // vbroadcastss 0x32ab(%rip),%ymm10 # 45c4 <_sk_callback_hsw+0x1ff>
.byte 196,65,44,93,201 // vminps %ymm9,%ymm10,%ymm9
- .byte 196,98,125,24,61,95,50,0,0 // vbroadcastss 0x325f(%rip),%ymm15 # 3f08 <_sk_callback_hsw+0x1c3>
+ .byte 196,98,125,24,61,161,50,0,0 // vbroadcastss 0x32a1(%rip),%ymm15 # 45c8 <_sk_callback_hsw+0x203>
.byte 196,193,124,194,199,1 // vcmpltps %ymm15,%ymm0,%ymm0
.byte 196,195,53,74,195,0 // vblendvps %ymm0,%ymm11,%ymm9,%ymm0
.byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9
@@ -8424,26 +9660,26 @@ _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 196,98,125,24,29,217,49,0,0 // vbroadcastss 0x31d9(%rip),%ymm11 # 3f0c <_sk_callback_hsw+0x1c7>
+ .byte 196,98,125,24,29,27,50,0,0 // vbroadcastss 0x321b(%rip),%ymm11 # 45cc <_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 196,98,125,24,53,198,49,0,0 // vbroadcastss 0x31c6(%rip),%ymm14 # 3f10 <_sk_callback_hsw+0x1cb>
+ .byte 196,98,125,24,53,8,50,0,0 // vbroadcastss 0x3208(%rip),%ymm14 # 45d0 <_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 196,98,125,24,53,173,49,0,0 // vbroadcastss 0x31ad(%rip),%ymm14 # 3f18 <_sk_callback_hsw+0x1d3>
+ .byte 196,98,125,24,53,239,49,0,0 // vbroadcastss 0x31ef(%rip),%ymm14 # 45d8 <_sk_callback_hsw+0x213>
.byte 196,66,37,168,238 // vfmadd213ps %ymm14,%ymm11,%ymm13
- .byte 196,98,125,24,53,155,49,0,0 // vbroadcastss 0x319b(%rip),%ymm14 # 3f14 <_sk_callback_hsw+0x1cf>
+ .byte 196,98,125,24,53,221,49,0,0 // vbroadcastss 0x31dd(%rip),%ymm14 # 45d4 <_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 196,98,125,24,29,126,49,0,0 // vbroadcastss 0x317e(%rip),%ymm11 # 3f20 <_sk_callback_hsw+0x1db>
+ .byte 196,98,125,24,29,192,49,0,0 // vbroadcastss 0x31c0(%rip),%ymm11 # 45e0 <_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
@@ -8453,7 +9689,7 @@ _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 196,98,125,24,5,65,49,0,0 // vbroadcastss 0x3141(%rip),%ymm8 # 3f1c <_sk_callback_hsw+0x1d7>
+ .byte 196,98,125,24,5,131,49,0,0 // vbroadcastss 0x3183(%rip),%ymm8 # 45dc <_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
@@ -8470,30 +9706,30 @@ _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 196,98,125,24,5,14,49,0,0 // vbroadcastss 0x310e(%rip),%ymm8 # 3f24 <_sk_callback_hsw+0x1df>
+ .byte 196,98,125,24,5,80,49,0,0 // vbroadcastss 0x3150(%rip),%ymm8 # 45e4 <_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 196,98,125,24,13,241,48,0,0 // vbroadcastss 0x30f1(%rip),%ymm9 # 3f28 <_sk_callback_hsw+0x1e3>
+ .byte 196,98,125,24,13,51,49,0,0 // vbroadcastss 0x3133(%rip),%ymm9 # 45e8 <_sk_callback_hsw+0x223>
.byte 196,66,109,170,202 // vfmsub213ps %ymm10,%ymm2,%ymm9
- .byte 196,98,125,24,29,231,48,0,0 // vbroadcastss 0x30e7(%rip),%ymm11 # 3f2c <_sk_callback_hsw+0x1e7>
+ .byte 196,98,125,24,29,41,49,0,0 // vbroadcastss 0x3129(%rip),%ymm11 # 45ec <_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 196,98,125,24,45,209,48,0,0 // vbroadcastss 0x30d1(%rip),%ymm13 # 3f34 <_sk_callback_hsw+0x1ef>
+ .byte 196,98,125,24,45,19,49,0,0 // vbroadcastss 0x3113(%rip),%ymm13 # 45f4 <_sk_callback_hsw+0x22f>
.byte 196,193,4,89,197 // vmulps %ymm13,%ymm15,%ymm0
- .byte 196,98,125,24,53,199,48,0,0 // vbroadcastss 0x30c7(%rip),%ymm14 # 3f38 <_sk_callback_hsw+0x1f3>
+ .byte 196,98,125,24,53,9,49,0,0 // vbroadcastss 0x3109(%rip),%ymm14 # 45f8 <_sk_callback_hsw+0x233>
.byte 197,12,92,224 // vsubps %ymm0,%ymm14,%ymm12
.byte 196,66,37,168,225 // vfmadd213ps %ymm9,%ymm11,%ymm12
- .byte 196,226,125,24,29,173,48,0,0 // vbroadcastss 0x30ad(%rip),%ymm3 # 3f30 <_sk_callback_hsw+0x1eb>
+ .byte 196,226,125,24,29,239,48,0,0 // vbroadcastss 0x30ef(%rip),%ymm3 # 45f0 <_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 196,98,125,24,37,152,48,0,0 // vbroadcastss 0x3098(%rip),%ymm12 # 3f3c <_sk_callback_hsw+0x1f7>
+ .byte 196,98,125,24,37,218,48,0,0 // vbroadcastss 0x30da(%rip),%ymm12 # 45fc <_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
@@ -8509,7 +9745,7 @@ _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 196,226,125,24,5,68,48,0,0 // vbroadcastss 0x3044(%rip),%ymm0 # 3f40 <_sk_callback_hsw+0x1fb>
+ .byte 196,226,125,24,5,134,48,0,0 // vbroadcastss 0x3086(%rip),%ymm0 # 4600 <_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
@@ -8559,11 +9795,11 @@ _sk_scale_u8_hsw:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,51 // jne fdc <_sk_scale_u8_hsw+0x43>
+ .byte 117,51 // jne 165a <_sk_scale_u8_hsw+0x43>
.byte 197,122,126,0 // vmovq (%rax),%xmm8
.byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,132,47,0,0 // vbroadcastss 0x2f84(%rip),%ymm9 # 3f44 <_sk_callback_hsw+0x1ff>
+ .byte 196,98,125,24,13,198,47,0,0 // vbroadcastss 0x2fc6(%rip),%ymm9 # 4604 <_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
@@ -8581,9 +9817,9 @@ _sk_scale_u8_hsw:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne fe4 <_sk_scale_u8_hsw+0x4b>
+ .byte 117,234 // jne 1662 <_sk_scale_u8_hsw+0x4b>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 235,172 // jmp fad <_sk_scale_u8_hsw+0x14>
+ .byte 235,172 // jmp 162b <_sk_scale_u8_hsw+0x14>
HIDDEN _sk_lerp_1_float_hsw
.globl _sk_lerp_1_float_hsw
@@ -8611,11 +9847,11 @@ _sk_lerp_u8_hsw:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,71 // jne 1087 <_sk_lerp_u8_hsw+0x57>
+ .byte 117,71 // jne 1705 <_sk_lerp_u8_hsw+0x57>
.byte 197,122,126,0 // vmovq (%rax),%xmm8
.byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,241,46,0,0 // vbroadcastss 0x2ef1(%rip),%ymm9 # 3f48 <_sk_callback_hsw+0x203>
+ .byte 196,98,125,24,13,51,47,0,0 // vbroadcastss 0x2f33(%rip),%ymm9 # 4608 <_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
@@ -8637,9 +9873,9 @@ _sk_lerp_u8_hsw:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 108f <_sk_lerp_u8_hsw+0x5f>
+ .byte 117,234 // jne 170d <_sk_lerp_u8_hsw+0x5f>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 235,152 // jmp 1044 <_sk_lerp_u8_hsw+0x14>
+ .byte 235,152 // jmp 16c2 <_sk_lerp_u8_hsw+0x14>
HIDDEN _sk_lerp_565_hsw
.globl _sk_lerp_565_hsw
@@ -8648,23 +9884,23 @@ _sk_lerp_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,149,0,0,0 // jne 114f <_sk_lerp_565_hsw+0xa3>
+ .byte 15,133,149,0,0,0 // jne 17cd <_sk_lerp_565_hsw+0xa3>
.byte 196,193,122,111,28,122 // vmovdqu (%r10,%rdi,2),%xmm3
.byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3
- .byte 196,98,125,88,5,126,46,0,0 // vpbroadcastd 0x2e7e(%rip),%ymm8 # 3f4c <_sk_callback_hsw+0x207>
+ .byte 196,98,125,88,5,192,46,0,0 // vpbroadcastd 0x2ec0(%rip),%ymm8 # 460c <_sk_callback_hsw+0x247>
.byte 196,65,101,219,192 // vpand %ymm8,%ymm3,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,111,46,0,0 // vbroadcastss 0x2e6f(%rip),%ymm9 # 3f50 <_sk_callback_hsw+0x20b>
+ .byte 196,98,125,24,13,177,46,0,0 // vbroadcastss 0x2eb1(%rip),%ymm9 # 4610 <_sk_callback_hsw+0x24b>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,88,13,101,46,0,0 // vpbroadcastd 0x2e65(%rip),%ymm9 # 3f54 <_sk_callback_hsw+0x20f>
+ .byte 196,98,125,88,13,167,46,0,0 // vpbroadcastd 0x2ea7(%rip),%ymm9 # 4614 <_sk_callback_hsw+0x24f>
.byte 196,65,101,219,201 // vpand %ymm9,%ymm3,%ymm9
.byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9
- .byte 196,98,125,24,21,86,46,0,0 // vbroadcastss 0x2e56(%rip),%ymm10 # 3f58 <_sk_callback_hsw+0x213>
+ .byte 196,98,125,24,21,152,46,0,0 // vbroadcastss 0x2e98(%rip),%ymm10 # 4618 <_sk_callback_hsw+0x253>
.byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,88,21,76,46,0,0 // vpbroadcastd 0x2e4c(%rip),%ymm10 # 3f5c <_sk_callback_hsw+0x217>
+ .byte 196,98,125,88,21,142,46,0,0 // vpbroadcastd 0x2e8e(%rip),%ymm10 # 461c <_sk_callback_hsw+0x257>
.byte 196,193,101,219,218 // vpand %ymm10,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,21,62,46,0,0 // vbroadcastss 0x2e3e(%rip),%ymm10 # 3f60 <_sk_callback_hsw+0x21b>
+ .byte 196,98,125,24,21,128,46,0,0 // vbroadcastss 0x2e80(%rip),%ymm10 # 4620 <_sk_callback_hsw+0x25b>
.byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
.byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0
@@ -8673,16 +9909,16 @@ _sk_lerp_565_hsw:
.byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
.byte 196,226,101,168,214 // vfmadd213ps %ymm6,%ymm3,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,23,46,0,0 // vbroadcastss 0x2e17(%rip),%ymm3 # 3f64 <_sk_callback_hsw+0x21f>
+ .byte 196,226,125,24,29,89,46,0,0 // vbroadcastss 0x2e59(%rip),%ymm3 # 4624 <_sk_callback_hsw+0x25f>
.byte 255,224 // jmpq *%rax
.byte 65,137,200 // mov %ecx,%r8d
.byte 65,128,224,7 // and $0x7,%r8b
.byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,89,255,255,255 // ja 10c0 <_sk_lerp_565_hsw+0x14>
+ .byte 15,135,89,255,255,255 // ja 173e <_sk_lerp_565_hsw+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 11bc <_sk_lerp_565_hsw+0x110>
+ .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 183c <_sk_lerp_565_hsw+0x112>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -8694,27 +9930,26 @@ _sk_lerp_565_hsw:
.byte 196,193,97,196,92,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3
.byte 196,193,97,196,92,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3
.byte 196,193,97,196,28,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3
- .byte 233,5,255,255,255 // jmpq 10c0 <_sk_lerp_565_hsw+0x14>
- .byte 144 // nop
- .byte 243,255 // repz (bad)
+ .byte 233,5,255,255,255 // jmpq 173e <_sk_lerp_565_hsw+0x14>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 241 // icebp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 235,255 // jmp 11c1 <_sk_lerp_565_hsw+0x115>
.byte 255 // (bad)
- .byte 255,227 // jmpq *%rbx
+ .byte 233,255,255,255,225 // jmpq ffffffffe2001844 <_sk_callback_hsw+0xffffffffe1ffd47f>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 219,255 // (bad)
+ .byte 217,255 // fcos
.byte 255 // (bad)
- .byte 255,211 // callq *%rbx
+ .byte 255,209 // callq *%rcx
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,203 // dec %ebx
+ .byte 255,201 // dec %ecx
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 191 // .byte 0xbf
+ .byte 189 // .byte 0xbd
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -8728,23 +9963,23 @@ _sk_load_tables_hsw:
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
.byte 76,3,8 // add (%rax),%r9
.byte 77,133,192 // test %r8,%r8
- .byte 117,105 // jne 1256 <_sk_load_tables_hsw+0x7e>
+ .byte 117,105 // jne 18d6 <_sk_load_tables_hsw+0x7e>
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
- .byte 197,229,219,13,38,48,0,0 // vpand 0x3026(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x4db>
+ .byte 197,229,219,13,102,48,0,0 // vpand 0x3066(%rip),%ymm3,%ymm1 # 48e0 <_sk_callback_hsw+0x51b>
.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 196,226,101,0,21,38,48,0,0 // vpshufb 0x3026(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x4fb>
+ .byte 196,226,101,0,21,102,48,0,0 // vpshufb 0x3066(%rip),%ymm3,%ymm2 # 4900 <_sk_callback_hsw+0x53b>
.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 196,98,101,0,13,46,48,0,0 // vpshufb 0x302e(%rip),%ymm3,%ymm9 # 4260 <_sk_callback_hsw+0x51b>
+ .byte 196,98,101,0,13,110,48,0,0 // vpshufb 0x306e(%rip),%ymm3,%ymm9 # 4920 <_sk_callback_hsw+0x55b>
.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 196,98,125,24,5,30,45,0,0 // vbroadcastss 0x2d1e(%rip),%ymm8 # 3f68 <_sk_callback_hsw+0x223>
+ .byte 196,98,125,24,5,94,45,0,0 // vbroadcastss 0x2d5e(%rip),%ymm8 # 4628 <_sk_callback_hsw+0x263>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,193 // mov %r8,%rcx
@@ -8757,7 +9992,7 @@ _sk_load_tables_hsw:
.byte 196,193,249,110,194 // vmovq %r10,%xmm0
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
.byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
- .byte 233,115,255,255,255 // jmpq 11f2 <_sk_load_tables_hsw+0x1a>
+ .byte 233,115,255,255,255 // jmpq 1872 <_sk_load_tables_hsw+0x1a>
HIDDEN _sk_load_tables_u16_be_hsw
.globl _sk_load_tables_u16_be_hsw
@@ -8767,7 +10002,7 @@ _sk_load_tables_u16_be_hsw:
.byte 76,139,0 // mov (%rax),%r8
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,201,0,0,0 // jne 135e <_sk_load_tables_u16_be_hsw+0xdf>
+ .byte 15,133,201,0,0,0 // jne 19de <_sk_load_tables_u16_be_hsw+0xdf>
.byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8
.byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2
.byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3
@@ -8783,7 +10018,7 @@ _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 197,121,111,21,186,48,0,0 // vmovdqa 0x30ba(%rip),%xmm10 # 43a0 <_sk_callback_hsw+0x65b>
+ .byte 197,121,111,21,250,48,0,0 // vmovdqa 0x30fa(%rip),%xmm10 # 4a60 <_sk_callback_hsw+0x69b>
.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
@@ -8805,36 +10040,36 @@ _sk_load_tables_u16_be_hsw:
.byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3
.byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,23,44,0,0 // vbroadcastss 0x2c17(%rip),%ymm8 # 3f6c <_sk_callback_hsw+0x227>
+ .byte 196,98,125,24,5,87,44,0,0 // vbroadcastss 0x2c57(%rip),%ymm8 # 462c <_sk_callback_hsw+0x267>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
.byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,85 // je 13c4 <_sk_load_tables_u16_be_hsw+0x145>
+ .byte 116,85 // je 1a44 <_sk_load_tables_u16_be_hsw+0x145>
.byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,72 // jb 13c4 <_sk_load_tables_u16_be_hsw+0x145>
+ .byte 114,72 // jb 1a44 <_sk_load_tables_u16_be_hsw+0x145>
.byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,72 // je 13d1 <_sk_load_tables_u16_be_hsw+0x152>
+ .byte 116,72 // je 1a51 <_sk_load_tables_u16_be_hsw+0x152>
.byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,59 // jb 13d1 <_sk_load_tables_u16_be_hsw+0x152>
+ .byte 114,59 // jb 1a51 <_sk_load_tables_u16_be_hsw+0x152>
.byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,9,255,255,255 // je 12b0 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 15,132,9,255,255,255 // je 1930 <_sk_load_tables_u16_be_hsw+0x31>
.byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,248,254,255,255 // jb 12b0 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 15,130,248,254,255,255 // jb 1930 <_sk_load_tables_u16_be_hsw+0x31>
.byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9
- .byte 233,236,254,255,255 // jmpq 12b0 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,236,254,255,255 // jmpq 1930 <_sk_load_tables_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,223,254,255,255 // jmpq 12b0 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,223,254,255,255 // jmpq 1930 <_sk_load_tables_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,214,254,255,255 // jmpq 12b0 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,214,254,255,255 // jmpq 1930 <_sk_load_tables_u16_be_hsw+0x31>
HIDDEN _sk_load_tables_rgb_u16_be_hsw
.globl _sk_load_tables_rgb_u16_be_hsw
@@ -8844,7 +10079,7 @@ _sk_load_tables_rgb_u16_be_hsw:
.byte 76,139,0 // mov (%rax),%r8
.byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,193,0,0,0 // jne 14ad <_sk_load_tables_rgb_u16_be_hsw+0xd3>
+ .byte 15,133,193,0,0,0 // jne 1b2d <_sk_load_tables_rgb_u16_be_hsw+0xd3>
.byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0
.byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2
.byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1
@@ -8865,7 +10100,7 @@ _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 197,121,111,13,90,47,0,0 // vmovdqa 0x2f5a(%rip),%xmm9 # 43b0 <_sk_callback_hsw+0x66b>
+ .byte 197,121,111,13,154,47,0,0 // vmovdqa 0x2f9a(%rip),%xmm9 # 4a70 <_sk_callback_hsw+0x6ab>
.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
@@ -8882,41 +10117,41 @@ _sk_load_tables_rgb_u16_be_hsw:
.byte 196,98,125,51,194 // vpmovzxwd %xmm2,%ymm8
.byte 196,162,101,146,20,128 // vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,197,42,0,0 // vbroadcastss 0x2ac5(%rip),%ymm3 # 3f70 <_sk_callback_hsw+0x22b>
+ .byte 196,226,125,24,29,5,43,0,0 // vbroadcastss 0x2b05(%rip),%ymm3 # 4630 <_sk_callback_hsw+0x26b>
.byte 255,224 // jmpq *%rax
.byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0
.byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 117,5 // jne 14c6 <_sk_load_tables_rgb_u16_be_hsw+0xec>
- .byte 233,90,255,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,5 // jne 1b46 <_sk_load_tables_rgb_u16_be_hsw+0xec>
+ .byte 233,90,255,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
.byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1
.byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,26 // jb 14f5 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
+ .byte 114,26 // jb 1b75 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
.byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1
.byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 117,10 // jne 14fa <_sk_load_tables_rgb_u16_be_hsw+0x120>
- .byte 233,43,255,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,38,255,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 1b7a <_sk_load_tables_rgb_u16_be_hsw+0x120>
+ .byte 233,43,255,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,38,255,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
.byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1
.byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,26 // jb 1529 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
+ .byte 114,26 // jb 1ba9 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
.byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1
.byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 117,10 // jne 152e <_sk_load_tables_rgb_u16_be_hsw+0x154>
- .byte 233,247,254,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,242,254,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 1bae <_sk_load_tables_rgb_u16_be_hsw+0x154>
+ .byte 233,247,254,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,242,254,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
.byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3
.byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,20 // jb 1557 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
+ .byte 114,20 // jb 1bd7 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
.byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3
.byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
- .byte 233,201,254,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,196,254,255,255 // jmpq 1420 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,201,254,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,196,254,255,255 // jmpq 1aa0 <_sk_load_tables_rgb_u16_be_hsw+0x46>
HIDDEN _sk_byte_tables_hsw
.globl _sk_byte_tables_hsw
@@ -8929,7 +10164,7 @@ _sk_byte_tables_hsw:
.byte 65,84 // push %r12
.byte 83 // push %rbx
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,3,42,0,0 // vbroadcastss 0x2a03(%rip),%ymm8 # 3f74 <_sk_callback_hsw+0x22f>
+ .byte 196,98,125,24,5,67,42,0,0 // vbroadcastss 0x2a43(%rip),%ymm8 # 4634 <_sk_callback_hsw+0x26f>
.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
@@ -8966,7 +10201,7 @@ _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 196,98,125,24,13,84,41,0,0 // vbroadcastss 0x2954(%rip),%ymm9 # 3f78 <_sk_callback_hsw+0x233>
+ .byte 196,98,125,24,13,148,41,0,0 // vbroadcastss 0x2994(%rip),%ymm9 # 4638 <_sk_callback_hsw+0x273>
.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
@@ -9127,7 +10362,7 @@ _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 196,98,125,24,13,141,38,0,0 // vbroadcastss 0x268d(%rip),%ymm9 # 3f7c <_sk_callback_hsw+0x237>
+ .byte 196,98,125,24,13,205,38,0,0 // vbroadcastss 0x26cd(%rip),%ymm9 # 463c <_sk_callback_hsw+0x277>
.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
@@ -9290,33 +10525,33 @@ _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 196,98,125,24,37,64,36,0,0 // vbroadcastss 0x2440(%rip),%ymm12 # 3f80 <_sk_callback_hsw+0x23b>
- .byte 196,98,125,24,45,59,36,0,0 // vbroadcastss 0x243b(%rip),%ymm13 # 3f84 <_sk_callback_hsw+0x23f>
+ .byte 196,98,125,24,37,128,36,0,0 // vbroadcastss 0x2480(%rip),%ymm12 # 4640 <_sk_callback_hsw+0x27b>
+ .byte 196,98,125,24,45,123,36,0,0 // vbroadcastss 0x247b(%rip),%ymm13 # 4644 <_sk_callback_hsw+0x27f>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,49,36,0,0 // vbroadcastss 0x2431(%rip),%ymm13 # 3f88 <_sk_callback_hsw+0x243>
+ .byte 196,98,125,24,45,113,36,0,0 // vbroadcastss 0x2471(%rip),%ymm13 # 4648 <_sk_callback_hsw+0x283>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,39,36,0,0 // vbroadcastss 0x2427(%rip),%ymm13 # 3f8c <_sk_callback_hsw+0x247>
+ .byte 196,98,125,24,45,103,36,0,0 // vbroadcastss 0x2467(%rip),%ymm13 # 464c <_sk_callback_hsw+0x287>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,29,36,0,0 // vbroadcastss 0x241d(%rip),%ymm11 # 3f90 <_sk_callback_hsw+0x24b>
+ .byte 196,98,125,24,29,93,36,0,0 // vbroadcastss 0x245d(%rip),%ymm11 # 4650 <_sk_callback_hsw+0x28b>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,19,36,0,0 // vbroadcastss 0x2413(%rip),%ymm12 # 3f94 <_sk_callback_hsw+0x24f>
+ .byte 196,98,125,24,37,83,36,0,0 // vbroadcastss 0x2453(%rip),%ymm12 # 4654 <_sk_callback_hsw+0x28f>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,9,36,0,0 // vbroadcastss 0x2409(%rip),%ymm12 # 3f98 <_sk_callback_hsw+0x253>
+ .byte 196,98,125,24,37,73,36,0,0 // vbroadcastss 0x2449(%rip),%ymm12 # 4658 <_sk_callback_hsw+0x293>
.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 196,98,125,24,29,234,35,0,0 // vbroadcastss 0x23ea(%rip),%ymm11 # 3f9c <_sk_callback_hsw+0x257>
+ .byte 196,98,125,24,29,42,36,0,0 // vbroadcastss 0x242a(%rip),%ymm11 # 465c <_sk_callback_hsw+0x297>
.byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,224,35,0,0 // vbroadcastss 0x23e0(%rip),%ymm11 # 3fa0 <_sk_callback_hsw+0x25b>
+ .byte 196,98,125,24,29,32,36,0,0 // vbroadcastss 0x2420(%rip),%ymm11 # 4660 <_sk_callback_hsw+0x29b>
.byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11
- .byte 196,226,125,24,5,214,35,0,0 // vbroadcastss 0x23d6(%rip),%ymm0 # 3fa4 <_sk_callback_hsw+0x25f>
+ .byte 196,226,125,24,5,22,36,0,0 // vbroadcastss 0x2416(%rip),%ymm0 # 4664 <_sk_callback_hsw+0x29f>
.byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0
- .byte 196,98,125,24,21,204,35,0,0 // vbroadcastss 0x23cc(%rip),%ymm10 # 3fa8 <_sk_callback_hsw+0x263>
+ .byte 196,98,125,24,21,12,36,0,0 // vbroadcastss 0x240c(%rip),%ymm10 # 4668 <_sk_callback_hsw+0x2a3>
.byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0
.byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
- .byte 196,98,125,24,21,191,35,0,0 // vbroadcastss 0x23bf(%rip),%ymm10 # 3fac <_sk_callback_hsw+0x267>
+ .byte 196,98,125,24,21,255,35,0,0 // vbroadcastss 0x23ff(%rip),%ymm10 # 466c <_sk_callback_hsw+0x2a7>
.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
@@ -9324,7 +10559,7 @@ _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 196,98,125,24,5,150,35,0,0 // vbroadcastss 0x2396(%rip),%ymm8 # 3fb0 <_sk_callback_hsw+0x26b>
+ .byte 196,98,125,24,5,214,35,0,0 // vbroadcastss 0x23d6(%rip),%ymm8 # 4670 <_sk_callback_hsw+0x2ab>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9344,33 +10579,33 @@ _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 196,98,125,24,37,78,35,0,0 // vbroadcastss 0x234e(%rip),%ymm12 # 3fb4 <_sk_callback_hsw+0x26f>
- .byte 196,98,125,24,45,73,35,0,0 // vbroadcastss 0x2349(%rip),%ymm13 # 3fb8 <_sk_callback_hsw+0x273>
+ .byte 196,98,125,24,37,142,35,0,0 // vbroadcastss 0x238e(%rip),%ymm12 # 4674 <_sk_callback_hsw+0x2af>
+ .byte 196,98,125,24,45,137,35,0,0 // vbroadcastss 0x2389(%rip),%ymm13 # 4678 <_sk_callback_hsw+0x2b3>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,63,35,0,0 // vbroadcastss 0x233f(%rip),%ymm13 # 3fbc <_sk_callback_hsw+0x277>
+ .byte 196,98,125,24,45,127,35,0,0 // vbroadcastss 0x237f(%rip),%ymm13 # 467c <_sk_callback_hsw+0x2b7>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,53,35,0,0 // vbroadcastss 0x2335(%rip),%ymm13 # 3fc0 <_sk_callback_hsw+0x27b>
+ .byte 196,98,125,24,45,117,35,0,0 // vbroadcastss 0x2375(%rip),%ymm13 # 4680 <_sk_callback_hsw+0x2bb>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,43,35,0,0 // vbroadcastss 0x232b(%rip),%ymm11 # 3fc4 <_sk_callback_hsw+0x27f>
+ .byte 196,98,125,24,29,107,35,0,0 // vbroadcastss 0x236b(%rip),%ymm11 # 4684 <_sk_callback_hsw+0x2bf>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,33,35,0,0 // vbroadcastss 0x2321(%rip),%ymm12 # 3fc8 <_sk_callback_hsw+0x283>
+ .byte 196,98,125,24,37,97,35,0,0 // vbroadcastss 0x2361(%rip),%ymm12 # 4688 <_sk_callback_hsw+0x2c3>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,23,35,0,0 // vbroadcastss 0x2317(%rip),%ymm12 # 3fcc <_sk_callback_hsw+0x287>
+ .byte 196,98,125,24,37,87,35,0,0 // vbroadcastss 0x2357(%rip),%ymm12 # 468c <_sk_callback_hsw+0x2c7>
.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 196,98,125,24,29,248,34,0,0 // vbroadcastss 0x22f8(%rip),%ymm11 # 3fd0 <_sk_callback_hsw+0x28b>
+ .byte 196,98,125,24,29,56,35,0,0 // vbroadcastss 0x2338(%rip),%ymm11 # 4690 <_sk_callback_hsw+0x2cb>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,238,34,0,0 // vbroadcastss 0x22ee(%rip),%ymm11 # 3fd4 <_sk_callback_hsw+0x28f>
+ .byte 196,98,125,24,29,46,35,0,0 // vbroadcastss 0x232e(%rip),%ymm11 # 4694 <_sk_callback_hsw+0x2cf>
.byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11
- .byte 196,226,125,24,13,228,34,0,0 // vbroadcastss 0x22e4(%rip),%ymm1 # 3fd8 <_sk_callback_hsw+0x293>
+ .byte 196,226,125,24,13,36,35,0,0 // vbroadcastss 0x2324(%rip),%ymm1 # 4698 <_sk_callback_hsw+0x2d3>
.byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1
- .byte 196,98,125,24,21,218,34,0,0 // vbroadcastss 0x22da(%rip),%ymm10 # 3fdc <_sk_callback_hsw+0x297>
+ .byte 196,98,125,24,21,26,35,0,0 // vbroadcastss 0x231a(%rip),%ymm10 # 469c <_sk_callback_hsw+0x2d7>
.byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
.byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
- .byte 196,98,125,24,21,205,34,0,0 // vbroadcastss 0x22cd(%rip),%ymm10 # 3fe0 <_sk_callback_hsw+0x29b>
+ .byte 196,98,125,24,21,13,35,0,0 // vbroadcastss 0x230d(%rip),%ymm10 # 46a0 <_sk_callback_hsw+0x2db>
.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
@@ -9378,7 +10613,7 @@ _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 196,98,125,24,5,164,34,0,0 // vbroadcastss 0x22a4(%rip),%ymm8 # 3fe4 <_sk_callback_hsw+0x29f>
+ .byte 196,98,125,24,5,228,34,0,0 // vbroadcastss 0x22e4(%rip),%ymm8 # 46a4 <_sk_callback_hsw+0x2df>
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9398,33 +10633,33 @@ _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 196,98,125,24,37,92,34,0,0 // vbroadcastss 0x225c(%rip),%ymm12 # 3fe8 <_sk_callback_hsw+0x2a3>
- .byte 196,98,125,24,45,87,34,0,0 // vbroadcastss 0x2257(%rip),%ymm13 # 3fec <_sk_callback_hsw+0x2a7>
+ .byte 196,98,125,24,37,156,34,0,0 // vbroadcastss 0x229c(%rip),%ymm12 # 46a8 <_sk_callback_hsw+0x2e3>
+ .byte 196,98,125,24,45,151,34,0,0 // vbroadcastss 0x2297(%rip),%ymm13 # 46ac <_sk_callback_hsw+0x2e7>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,77,34,0,0 // vbroadcastss 0x224d(%rip),%ymm13 # 3ff0 <_sk_callback_hsw+0x2ab>
+ .byte 196,98,125,24,45,141,34,0,0 // vbroadcastss 0x228d(%rip),%ymm13 # 46b0 <_sk_callback_hsw+0x2eb>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,67,34,0,0 // vbroadcastss 0x2243(%rip),%ymm13 # 3ff4 <_sk_callback_hsw+0x2af>
+ .byte 196,98,125,24,45,131,34,0,0 // vbroadcastss 0x2283(%rip),%ymm13 # 46b4 <_sk_callback_hsw+0x2ef>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,57,34,0,0 // vbroadcastss 0x2239(%rip),%ymm11 # 3ff8 <_sk_callback_hsw+0x2b3>
+ .byte 196,98,125,24,29,121,34,0,0 // vbroadcastss 0x2279(%rip),%ymm11 # 46b8 <_sk_callback_hsw+0x2f3>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,47,34,0,0 // vbroadcastss 0x222f(%rip),%ymm12 # 3ffc <_sk_callback_hsw+0x2b7>
+ .byte 196,98,125,24,37,111,34,0,0 // vbroadcastss 0x226f(%rip),%ymm12 # 46bc <_sk_callback_hsw+0x2f7>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,37,34,0,0 // vbroadcastss 0x2225(%rip),%ymm12 # 4000 <_sk_callback_hsw+0x2bb>
+ .byte 196,98,125,24,37,101,34,0,0 // vbroadcastss 0x2265(%rip),%ymm12 # 46c0 <_sk_callback_hsw+0x2fb>
.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 196,98,125,24,29,6,34,0,0 // vbroadcastss 0x2206(%rip),%ymm11 # 4004 <_sk_callback_hsw+0x2bf>
+ .byte 196,98,125,24,29,70,34,0,0 // vbroadcastss 0x2246(%rip),%ymm11 # 46c4 <_sk_callback_hsw+0x2ff>
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,252,33,0,0 // vbroadcastss 0x21fc(%rip),%ymm11 # 4008 <_sk_callback_hsw+0x2c3>
+ .byte 196,98,125,24,29,60,34,0,0 // vbroadcastss 0x223c(%rip),%ymm11 # 46c8 <_sk_callback_hsw+0x303>
.byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11
- .byte 196,226,125,24,21,242,33,0,0 // vbroadcastss 0x21f2(%rip),%ymm2 # 400c <_sk_callback_hsw+0x2c7>
+ .byte 196,226,125,24,21,50,34,0,0 // vbroadcastss 0x2232(%rip),%ymm2 # 46cc <_sk_callback_hsw+0x307>
.byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2
- .byte 196,98,125,24,21,232,33,0,0 // vbroadcastss 0x21e8(%rip),%ymm10 # 4010 <_sk_callback_hsw+0x2cb>
+ .byte 196,98,125,24,21,40,34,0,0 // vbroadcastss 0x2228(%rip),%ymm10 # 46d0 <_sk_callback_hsw+0x30b>
.byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2
.byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
- .byte 196,98,125,24,21,219,33,0,0 // vbroadcastss 0x21db(%rip),%ymm10 # 4014 <_sk_callback_hsw+0x2cf>
+ .byte 196,98,125,24,21,27,34,0,0 // vbroadcastss 0x221b(%rip),%ymm10 # 46d4 <_sk_callback_hsw+0x30f>
.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
@@ -9432,7 +10667,7 @@ _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 196,98,125,24,5,178,33,0,0 // vbroadcastss 0x21b2(%rip),%ymm8 # 4018 <_sk_callback_hsw+0x2d3>
+ .byte 196,98,125,24,5,242,33,0,0 // vbroadcastss 0x21f2(%rip),%ymm8 # 46d8 <_sk_callback_hsw+0x313>
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9452,33 +10687,33 @@ _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 196,98,125,24,37,106,33,0,0 // vbroadcastss 0x216a(%rip),%ymm12 # 401c <_sk_callback_hsw+0x2d7>
- .byte 196,98,125,24,45,101,33,0,0 // vbroadcastss 0x2165(%rip),%ymm13 # 4020 <_sk_callback_hsw+0x2db>
+ .byte 196,98,125,24,37,170,33,0,0 // vbroadcastss 0x21aa(%rip),%ymm12 # 46dc <_sk_callback_hsw+0x317>
+ .byte 196,98,125,24,45,165,33,0,0 // vbroadcastss 0x21a5(%rip),%ymm13 # 46e0 <_sk_callback_hsw+0x31b>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,91,33,0,0 // vbroadcastss 0x215b(%rip),%ymm13 # 4024 <_sk_callback_hsw+0x2df>
+ .byte 196,98,125,24,45,155,33,0,0 // vbroadcastss 0x219b(%rip),%ymm13 # 46e4 <_sk_callback_hsw+0x31f>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,81,33,0,0 // vbroadcastss 0x2151(%rip),%ymm13 # 4028 <_sk_callback_hsw+0x2e3>
+ .byte 196,98,125,24,45,145,33,0,0 // vbroadcastss 0x2191(%rip),%ymm13 # 46e8 <_sk_callback_hsw+0x323>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,71,33,0,0 // vbroadcastss 0x2147(%rip),%ymm11 # 402c <_sk_callback_hsw+0x2e7>
+ .byte 196,98,125,24,29,135,33,0,0 // vbroadcastss 0x2187(%rip),%ymm11 # 46ec <_sk_callback_hsw+0x327>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,61,33,0,0 // vbroadcastss 0x213d(%rip),%ymm12 # 4030 <_sk_callback_hsw+0x2eb>
+ .byte 196,98,125,24,37,125,33,0,0 // vbroadcastss 0x217d(%rip),%ymm12 # 46f0 <_sk_callback_hsw+0x32b>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,51,33,0,0 // vbroadcastss 0x2133(%rip),%ymm12 # 4034 <_sk_callback_hsw+0x2ef>
+ .byte 196,98,125,24,37,115,33,0,0 // vbroadcastss 0x2173(%rip),%ymm12 # 46f4 <_sk_callback_hsw+0x32f>
.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 196,98,125,24,29,20,33,0,0 // vbroadcastss 0x2114(%rip),%ymm11 # 4038 <_sk_callback_hsw+0x2f3>
+ .byte 196,98,125,24,29,84,33,0,0 // vbroadcastss 0x2154(%rip),%ymm11 # 46f8 <_sk_callback_hsw+0x333>
.byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,10,33,0,0 // vbroadcastss 0x210a(%rip),%ymm11 # 403c <_sk_callback_hsw+0x2f7>
+ .byte 196,98,125,24,29,74,33,0,0 // vbroadcastss 0x214a(%rip),%ymm11 # 46fc <_sk_callback_hsw+0x337>
.byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11
- .byte 196,226,125,24,29,0,33,0,0 // vbroadcastss 0x2100(%rip),%ymm3 # 4040 <_sk_callback_hsw+0x2fb>
+ .byte 196,226,125,24,29,64,33,0,0 // vbroadcastss 0x2140(%rip),%ymm3 # 4700 <_sk_callback_hsw+0x33b>
.byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3
- .byte 196,98,125,24,21,246,32,0,0 // vbroadcastss 0x20f6(%rip),%ymm10 # 4044 <_sk_callback_hsw+0x2ff>
+ .byte 196,98,125,24,21,54,33,0,0 // vbroadcastss 0x2136(%rip),%ymm10 # 4704 <_sk_callback_hsw+0x33f>
.byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3
.byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
- .byte 196,98,125,24,21,233,32,0,0 // vbroadcastss 0x20e9(%rip),%ymm10 # 4048 <_sk_callback_hsw+0x303>
+ .byte 196,98,125,24,21,41,33,0,0 // vbroadcastss 0x2129(%rip),%ymm10 # 4708 <_sk_callback_hsw+0x343>
.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
@@ -9486,7 +10721,7 @@ _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 196,98,125,24,5,192,32,0,0 // vbroadcastss 0x20c0(%rip),%ymm8 # 404c <_sk_callback_hsw+0x307>
+ .byte 196,98,125,24,5,0,33,0,0 // vbroadcastss 0x2100(%rip),%ymm8 # 470c <_sk_callback_hsw+0x347>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9495,26 +10730,26 @@ HIDDEN _sk_lab_to_xyz_hsw
.globl _sk_lab_to_xyz_hsw
FUNCTION(_sk_lab_to_xyz_hsw)
_sk_lab_to_xyz_hsw:
- .byte 196,98,125,24,5,178,32,0,0 // vbroadcastss 0x20b2(%rip),%ymm8 # 4050 <_sk_callback_hsw+0x30b>
- .byte 196,98,125,24,13,173,32,0,0 // vbroadcastss 0x20ad(%rip),%ymm9 # 4054 <_sk_callback_hsw+0x30f>
- .byte 196,98,125,24,21,168,32,0,0 // vbroadcastss 0x20a8(%rip),%ymm10 # 4058 <_sk_callback_hsw+0x313>
+ .byte 196,98,125,24,5,242,32,0,0 // vbroadcastss 0x20f2(%rip),%ymm8 # 4710 <_sk_callback_hsw+0x34b>
+ .byte 196,98,125,24,13,237,32,0,0 // vbroadcastss 0x20ed(%rip),%ymm9 # 4714 <_sk_callback_hsw+0x34f>
+ .byte 196,98,125,24,21,232,32,0,0 // vbroadcastss 0x20e8(%rip),%ymm10 # 4718 <_sk_callback_hsw+0x353>
.byte 196,194,53,168,202 // vfmadd213ps %ymm10,%ymm9,%ymm1
.byte 196,194,53,168,210 // vfmadd213ps %ymm10,%ymm9,%ymm2
- .byte 196,98,125,24,13,153,32,0,0 // vbroadcastss 0x2099(%rip),%ymm9 # 405c <_sk_callback_hsw+0x317>
+ .byte 196,98,125,24,13,217,32,0,0 // vbroadcastss 0x20d9(%rip),%ymm9 # 471c <_sk_callback_hsw+0x357>
.byte 196,66,125,184,200 // vfmadd231ps %ymm8,%ymm0,%ymm9
- .byte 196,226,125,24,5,143,32,0,0 // vbroadcastss 0x208f(%rip),%ymm0 # 4060 <_sk_callback_hsw+0x31b>
+ .byte 196,226,125,24,5,207,32,0,0 // vbroadcastss 0x20cf(%rip),%ymm0 # 4720 <_sk_callback_hsw+0x35b>
.byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
- .byte 196,98,125,24,5,134,32,0,0 // vbroadcastss 0x2086(%rip),%ymm8 # 4064 <_sk_callback_hsw+0x31f>
+ .byte 196,98,125,24,5,198,32,0,0 // vbroadcastss 0x20c6(%rip),%ymm8 # 4724 <_sk_callback_hsw+0x35f>
.byte 196,98,117,168,192 // vfmadd213ps %ymm0,%ymm1,%ymm8
- .byte 196,98,125,24,13,124,32,0,0 // vbroadcastss 0x207c(%rip),%ymm9 # 4068 <_sk_callback_hsw+0x323>
+ .byte 196,98,125,24,13,188,32,0,0 // vbroadcastss 0x20bc(%rip),%ymm9 # 4728 <_sk_callback_hsw+0x363>
.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 196,226,125,24,21,105,32,0,0 // vbroadcastss 0x2069(%rip),%ymm2 # 406c <_sk_callback_hsw+0x327>
+ .byte 196,226,125,24,21,169,32,0,0 // vbroadcastss 0x20a9(%rip),%ymm2 # 472c <_sk_callback_hsw+0x367>
.byte 197,108,194,209,1 // vcmpltps %ymm1,%ymm2,%ymm10
- .byte 196,98,125,24,29,95,32,0,0 // vbroadcastss 0x205f(%rip),%ymm11 # 4070 <_sk_callback_hsw+0x32b>
+ .byte 196,98,125,24,29,159,32,0,0 // vbroadcastss 0x209f(%rip),%ymm11 # 4730 <_sk_callback_hsw+0x36b>
.byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
- .byte 196,98,125,24,37,85,32,0,0 // vbroadcastss 0x2055(%rip),%ymm12 # 4074 <_sk_callback_hsw+0x32f>
+ .byte 196,98,125,24,37,149,32,0,0 // vbroadcastss 0x2095(%rip),%ymm12 # 4734 <_sk_callback_hsw+0x36f>
.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
@@ -9529,9 +10764,9 @@ _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 196,226,125,24,5,10,32,0,0 // vbroadcastss 0x200a(%rip),%ymm0 # 4078 <_sk_callback_hsw+0x333>
+ .byte 196,226,125,24,5,74,32,0,0 // vbroadcastss 0x204a(%rip),%ymm0 # 4738 <_sk_callback_hsw+0x373>
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,1,32,0,0 // vbroadcastss 0x2001(%rip),%ymm8 # 407c <_sk_callback_hsw+0x337>
+ .byte 196,98,125,24,5,65,32,0,0 // vbroadcastss 0x2041(%rip),%ymm8 # 473c <_sk_callback_hsw+0x377>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9545,11 +10780,11 @@ _sk_load_a8_hsw:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,45 // jne 20c1 <_sk_load_a8_hsw+0x3d>
+ .byte 117,45 // jne 2741 <_sk_load_a8_hsw+0x3d>
.byte 197,250,126,0 // vmovq (%rax),%xmm0
.byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,214,31,0,0 // vbroadcastss 0x1fd6(%rip),%ymm1 # 4080 <_sk_callback_hsw+0x33b>
+ .byte 196,226,125,24,13,22,32,0,0 // vbroadcastss 0x2016(%rip),%ymm1 # 4740 <_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
@@ -9566,9 +10801,9 @@ _sk_load_a8_hsw:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 20c9 <_sk_load_a8_hsw+0x45>
+ .byte 117,234 // jne 2749 <_sk_load_a8_hsw+0x45>
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
- .byte 235,178 // jmp 2098 <_sk_load_a8_hsw+0x14>
+ .byte 235,178 // jmp 2718 <_sk_load_a8_hsw+0x14>
HIDDEN _sk_gather_a8_hsw
.globl _sk_gather_a8_hsw
@@ -9614,7 +10849,7 @@ _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 196,226,125,24,13,225,30,0,0 // vbroadcastss 0x1ee1(%rip),%ymm1 # 4084 <_sk_callback_hsw+0x33f>
+ .byte 196,226,125,24,13,33,31,0,0 // vbroadcastss 0x1f21(%rip),%ymm1 # 4744 <_sk_callback_hsw+0x37f>
.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
@@ -9632,14 +10867,14 @@ FUNCTION(_sk_store_a8_hsw)
_sk_store_a8_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,188,30,0,0 // vbroadcastss 0x1ebc(%rip),%ymm8 # 4088 <_sk_callback_hsw+0x343>
+ .byte 196,98,125,24,5,252,30,0,0 // vbroadcastss 0x1efc(%rip),%ymm8 # 4748 <_sk_callback_hsw+0x383>
.byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8
.byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 21f5 <_sk_store_a8_hsw+0x37>
+ .byte 117,10 // jne 2875 <_sk_store_a8_hsw+0x37>
.byte 196,65,123,17,4,58 // vmovsd %xmm8,(%r10,%rdi,1)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9647,10 +10882,10 @@ _sk_store_a8_hsw:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 21f1 <_sk_store_a8_hsw+0x33>
+ .byte 119,236 // ja 2871 <_sk_store_a8_hsw+0x33>
.byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 2258 <_sk_store_a8_hsw+0x9a>
+ .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 28d8 <_sk_store_a8_hsw+0x9a>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -9661,7 +10896,7 @@ _sk_store_a8_hsw:
.byte 196,67,121,20,68,58,2,4 // vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1)
.byte 196,67,121,20,68,58,1,2 // vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1)
.byte 196,67,121,20,4,58,0 // vpextrb $0x0,%xmm8,(%r10,%rdi,1)
- .byte 235,154 // jmp 21f1 <_sk_store_a8_hsw+0x33>
+ .byte 235,154 // jmp 2871 <_sk_store_a8_hsw+0x33>
.byte 144 // nop
.byte 246,255 // idiv %bh
.byte 255 // (bad)
@@ -9695,14 +10930,14 @@ _sk_load_g8_hsw:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,50 // jne 22b6 <_sk_load_g8_hsw+0x42>
+ .byte 117,50 // jne 2936 <_sk_load_g8_hsw+0x42>
.byte 197,250,126,0 // vmovq (%rax),%xmm0
.byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,242,29,0,0 // vbroadcastss 0x1df2(%rip),%ymm1 # 408c <_sk_callback_hsw+0x347>
+ .byte 196,226,125,24,13,50,30,0,0 // vbroadcastss 0x1e32(%rip),%ymm1 # 474c <_sk_callback_hsw+0x387>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,231,29,0,0 // vbroadcastss 0x1de7(%rip),%ymm3 # 4090 <_sk_callback_hsw+0x34b>
+ .byte 196,226,125,24,29,39,30,0,0 // vbroadcastss 0x1e27(%rip),%ymm3 # 4750 <_sk_callback_hsw+0x38b>
.byte 76,137,193 // mov %r8,%rcx
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
@@ -9716,9 +10951,9 @@ _sk_load_g8_hsw:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 22be <_sk_load_g8_hsw+0x4a>
+ .byte 117,234 // jne 293e <_sk_load_g8_hsw+0x4a>
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
- .byte 235,173 // jmp 2288 <_sk_load_g8_hsw+0x14>
+ .byte 235,173 // jmp 2908 <_sk_load_g8_hsw+0x14>
HIDDEN _sk_gather_g8_hsw
.globl _sk_gather_g8_hsw
@@ -9764,10 +10999,10 @@ _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 196,226,125,24,13,252,28,0,0 // vbroadcastss 0x1cfc(%rip),%ymm1 # 4094 <_sk_callback_hsw+0x34f>
+ .byte 196,226,125,24,13,60,29,0,0 // vbroadcastss 0x1d3c(%rip),%ymm1 # 4754 <_sk_callback_hsw+0x38f>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,241,28,0,0 // vbroadcastss 0x1cf1(%rip),%ymm3 # 4098 <_sk_callback_hsw+0x353>
+ .byte 196,226,125,24,29,49,29,0,0 // vbroadcastss 0x1d31(%rip),%ymm3 # 4758 <_sk_callback_hsw+0x393>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 91 // pop %rbx
@@ -9783,9 +11018,9 @@ _sk_gather_i8_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,137,192 // mov %rax,%r8
.byte 77,133,192 // test %r8,%r8
- .byte 116,5 // je 23c7 <_sk_gather_i8_hsw+0xf>
+ .byte 116,5 // je 2a47 <_sk_gather_i8_hsw+0xf>
.byte 76,137,192 // mov %r8,%rax
- .byte 235,2 // jmp 23c9 <_sk_gather_i8_hsw+0x11>
+ .byte 235,2 // jmp 2a49 <_sk_gather_i8_hsw+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,87 // push %r15
.byte 65,86 // push %r14
@@ -9823,14 +11058,14 @@ _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 197,229,219,5,9,30,0,0 // vpand 0x1e09(%rip),%ymm3,%ymm0 # 4280 <_sk_callback_hsw+0x53b>
+ .byte 197,229,219,5,73,30,0,0 // vpand 0x1e49(%rip),%ymm3,%ymm0 # 4940 <_sk_callback_hsw+0x57b>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,24,28,0,0 // vbroadcastss 0x1c18(%rip),%ymm8 # 409c <_sk_callback_hsw+0x357>
+ .byte 196,98,125,24,5,88,28,0,0 // vbroadcastss 0x1c58(%rip),%ymm8 # 475c <_sk_callback_hsw+0x397>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,14,30,0,0 // vpshufb 0x1e0e(%rip),%ymm3,%ymm1 # 42a0 <_sk_callback_hsw+0x55b>
+ .byte 196,226,101,0,13,78,30,0,0 // vpshufb 0x1e4e(%rip),%ymm3,%ymm1 # 4960 <_sk_callback_hsw+0x59b>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,28,30,0,0 // vpshufb 0x1e1c(%rip),%ymm3,%ymm2 # 42c0 <_sk_callback_hsw+0x57b>
+ .byte 196,226,101,0,21,92,30,0,0 // vpshufb 0x1e5c(%rip),%ymm3,%ymm2 # 4980 <_sk_callback_hsw+0x5bb>
.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
@@ -9851,35 +11086,35 @@ _sk_load_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,114 // jne 2544 <_sk_load_565_hsw+0x7c>
+ .byte 117,114 // jne 2bc4 <_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 196,226,125,88,5,186,27,0,0 // vpbroadcastd 0x1bba(%rip),%ymm0 # 40a0 <_sk_callback_hsw+0x35b>
+ .byte 196,226,125,88,5,250,27,0,0 // vpbroadcastd 0x1bfa(%rip),%ymm0 # 4760 <_sk_callback_hsw+0x39b>
.byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,173,27,0,0 // vbroadcastss 0x1bad(%rip),%ymm1 # 40a4 <_sk_callback_hsw+0x35f>
+ .byte 196,226,125,24,13,237,27,0,0 // vbroadcastss 0x1bed(%rip),%ymm1 # 4764 <_sk_callback_hsw+0x39f>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,164,27,0,0 // vpbroadcastd 0x1ba4(%rip),%ymm1 # 40a8 <_sk_callback_hsw+0x363>
+ .byte 196,226,125,88,13,228,27,0,0 // vpbroadcastd 0x1be4(%rip),%ymm1 # 4768 <_sk_callback_hsw+0x3a3>
.byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,151,27,0,0 // vbroadcastss 0x1b97(%rip),%ymm3 # 40ac <_sk_callback_hsw+0x367>
+ .byte 196,226,125,24,29,215,27,0,0 // vbroadcastss 0x1bd7(%rip),%ymm3 # 476c <_sk_callback_hsw+0x3a7>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,88,29,142,27,0,0 // vpbroadcastd 0x1b8e(%rip),%ymm3 # 40b0 <_sk_callback_hsw+0x36b>
+ .byte 196,226,125,88,29,206,27,0,0 // vpbroadcastd 0x1bce(%rip),%ymm3 # 4770 <_sk_callback_hsw+0x3ab>
.byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,129,27,0,0 // vbroadcastss 0x1b81(%rip),%ymm3 # 40b4 <_sk_callback_hsw+0x36f>
+ .byte 196,226,125,24,29,193,27,0,0 // vbroadcastss 0x1bc1(%rip),%ymm3 # 4774 <_sk_callback_hsw+0x3af>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,118,27,0,0 // vbroadcastss 0x1b76(%rip),%ymm3 # 40b8 <_sk_callback_hsw+0x373>
+ .byte 196,226,125,24,29,182,27,0,0 // vbroadcastss 0x1bb6(%rip),%ymm3 # 4778 <_sk_callback_hsw+0x3b3>
.byte 255,224 // jmpq *%rax
.byte 65,137,200 // mov %ecx,%r8d
.byte 65,128,224,7 // and $0x7,%r8b
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,128 // ja 24d8 <_sk_load_565_hsw+0x10>
+ .byte 119,128 // ja 2b58 <_sk_load_565_hsw+0x10>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 25ac <_sk_load_565_hsw+0xe4>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2c2c <_sk_load_565_hsw+0xe4>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -9891,7 +11126,7 @@ _sk_load_565_hsw:
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- .byte 233,44,255,255,255 // jmpq 24d8 <_sk_load_565_hsw+0x10>
+ .byte 233,44,255,255,255 // jmpq 2b58 <_sk_load_565_hsw+0x10>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -9961,23 +11196,23 @@ _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 196,226,125,88,5,57,26,0,0 // vpbroadcastd 0x1a39(%rip),%ymm0 # 40bc <_sk_callback_hsw+0x377>
+ .byte 196,226,125,88,5,121,26,0,0 // vpbroadcastd 0x1a79(%rip),%ymm0 # 477c <_sk_callback_hsw+0x3b7>
.byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,44,26,0,0 // vbroadcastss 0x1a2c(%rip),%ymm1 # 40c0 <_sk_callback_hsw+0x37b>
+ .byte 196,226,125,24,13,108,26,0,0 // vbroadcastss 0x1a6c(%rip),%ymm1 # 4780 <_sk_callback_hsw+0x3bb>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,35,26,0,0 // vpbroadcastd 0x1a23(%rip),%ymm1 # 40c4 <_sk_callback_hsw+0x37f>
+ .byte 196,226,125,88,13,99,26,0,0 // vpbroadcastd 0x1a63(%rip),%ymm1 # 4784 <_sk_callback_hsw+0x3bf>
.byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,22,26,0,0 // vbroadcastss 0x1a16(%rip),%ymm3 # 40c8 <_sk_callback_hsw+0x383>
+ .byte 196,226,125,24,29,86,26,0,0 // vbroadcastss 0x1a56(%rip),%ymm3 # 4788 <_sk_callback_hsw+0x3c3>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,88,29,13,26,0,0 // vpbroadcastd 0x1a0d(%rip),%ymm3 # 40cc <_sk_callback_hsw+0x387>
+ .byte 196,226,125,88,29,77,26,0,0 // vpbroadcastd 0x1a4d(%rip),%ymm3 # 478c <_sk_callback_hsw+0x3c7>
.byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,0,26,0,0 // vbroadcastss 0x1a00(%rip),%ymm3 # 40d0 <_sk_callback_hsw+0x38b>
+ .byte 196,226,125,24,29,64,26,0,0 // vbroadcastss 0x1a40(%rip),%ymm3 # 4790 <_sk_callback_hsw+0x3cb>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,245,25,0,0 // vbroadcastss 0x19f5(%rip),%ymm3 # 40d4 <_sk_callback_hsw+0x38f>
+ .byte 196,226,125,24,29,53,26,0,0 // vbroadcastss 0x1a35(%rip),%ymm3 # 4794 <_sk_callback_hsw+0x3cf>
.byte 91 // pop %rbx
.byte 65,92 // pop %r12
.byte 65,94 // pop %r14
@@ -9990,11 +11225,11 @@ FUNCTION(_sk_store_565_hsw)
_sk_store_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,226,25,0,0 // vbroadcastss 0x19e2(%rip),%ymm8 # 40d8 <_sk_callback_hsw+0x393>
+ .byte 196,98,125,24,5,34,26,0,0 // vbroadcastss 0x1a22(%rip),%ymm8 # 4798 <_sk_callback_hsw+0x3d3>
.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 196,98,125,24,21,205,25,0,0 // vbroadcastss 0x19cd(%rip),%ymm10 # 40dc <_sk_callback_hsw+0x397>
+ .byte 196,98,125,24,21,13,26,0,0 // vbroadcastss 0x1a0d(%rip),%ymm10 # 479c <_sk_callback_hsw+0x3d7>
.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
@@ -10005,7 +11240,7 @@ _sk_store_565_hsw:
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 274d <_sk_store_565_hsw+0x65>
+ .byte 117,10 // jne 2dcd <_sk_store_565_hsw+0x65>
.byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -10013,9 +11248,9 @@ _sk_store_565_hsw:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 2749 <_sk_store_565_hsw+0x61>
+ .byte 119,236 // ja 2dc9 <_sk_store_565_hsw+0x61>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 27ac <_sk_store_565_hsw+0xc4>
+ .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2e2c <_sk_store_565_hsw+0xc4>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -10026,7 +11261,7 @@ _sk_store_565_hsw:
.byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
.byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
.byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- .byte 235,159 // jmp 2749 <_sk_store_565_hsw+0x61>
+ .byte 235,159 // jmp 2dc9 <_sk_store_565_hsw+0x61>
.byte 102,144 // xchg %ax,%ax
.byte 245 // cmc
.byte 255 // (bad)
@@ -10059,28 +11294,28 @@ _sk_load_4444_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,138,0,0,0 // jne 2860 <_sk_load_4444_hsw+0x98>
+ .byte 15,133,138,0,0,0 // jne 2ee0 <_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 196,226,125,88,5,246,24,0,0 // vpbroadcastd 0x18f6(%rip),%ymm0 # 40e0 <_sk_callback_hsw+0x39b>
+ .byte 196,226,125,88,5,54,25,0,0 // vpbroadcastd 0x1936(%rip),%ymm0 # 47a0 <_sk_callback_hsw+0x3db>
.byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,233,24,0,0 // vbroadcastss 0x18e9(%rip),%ymm1 # 40e4 <_sk_callback_hsw+0x39f>
+ .byte 196,226,125,24,13,41,25,0,0 // vbroadcastss 0x1929(%rip),%ymm1 # 47a4 <_sk_callback_hsw+0x3df>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,224,24,0,0 // vpbroadcastd 0x18e0(%rip),%ymm1 # 40e8 <_sk_callback_hsw+0x3a3>
+ .byte 196,226,125,88,13,32,25,0,0 // vpbroadcastd 0x1920(%rip),%ymm1 # 47a8 <_sk_callback_hsw+0x3e3>
.byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,211,24,0,0 // vbroadcastss 0x18d3(%rip),%ymm2 # 40ec <_sk_callback_hsw+0x3a7>
+ .byte 196,226,125,24,21,19,25,0,0 // vbroadcastss 0x1913(%rip),%ymm2 # 47ac <_sk_callback_hsw+0x3e7>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,88,21,202,24,0,0 // vpbroadcastd 0x18ca(%rip),%ymm2 # 40f0 <_sk_callback_hsw+0x3ab>
+ .byte 196,226,125,88,21,10,25,0,0 // vpbroadcastd 0x190a(%rip),%ymm2 # 47b0 <_sk_callback_hsw+0x3eb>
.byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,189,24,0,0 // vbroadcastss 0x18bd(%rip),%ymm8 # 40f4 <_sk_callback_hsw+0x3af>
+ .byte 196,98,125,24,5,253,24,0,0 // vbroadcastss 0x18fd(%rip),%ymm8 # 47b4 <_sk_callback_hsw+0x3ef>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,88,5,179,24,0,0 // vpbroadcastd 0x18b3(%rip),%ymm8 # 40f8 <_sk_callback_hsw+0x3b3>
+ .byte 196,98,125,88,5,243,24,0,0 // vpbroadcastd 0x18f3(%rip),%ymm8 # 47b8 <_sk_callback_hsw+0x3f3>
.byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,165,24,0,0 // vbroadcastss 0x18a5(%rip),%ymm8 # 40fc <_sk_callback_hsw+0x3b7>
+ .byte 196,98,125,24,5,229,24,0,0 // vbroadcastss 0x18e5(%rip),%ymm8 # 47bc <_sk_callback_hsw+0x3f7>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -10089,9 +11324,9 @@ _sk_load_4444_hsw:
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,100,255,255,255 // ja 27dc <_sk_load_4444_hsw+0x14>
+ .byte 15,135,100,255,255,255 // ja 2e5c <_sk_load_4444_hsw+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 28cc <_sk_load_4444_hsw+0x104>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2f4c <_sk_load_4444_hsw+0x104>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -10103,7 +11338,7 @@ _sk_load_4444_hsw:
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- .byte 233,16,255,255,255 // jmpq 27dc <_sk_load_4444_hsw+0x14>
+ .byte 233,16,255,255,255 // jmpq 2e5c <_sk_load_4444_hsw+0x14>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -10173,25 +11408,25 @@ _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 196,226,125,88,5,93,23,0,0 // vpbroadcastd 0x175d(%rip),%ymm0 # 4100 <_sk_callback_hsw+0x3bb>
+ .byte 196,226,125,88,5,157,23,0,0 // vpbroadcastd 0x179d(%rip),%ymm0 # 47c0 <_sk_callback_hsw+0x3fb>
.byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,80,23,0,0 // vbroadcastss 0x1750(%rip),%ymm1 # 4104 <_sk_callback_hsw+0x3bf>
+ .byte 196,226,125,24,13,144,23,0,0 // vbroadcastss 0x1790(%rip),%ymm1 # 47c4 <_sk_callback_hsw+0x3ff>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,71,23,0,0 // vpbroadcastd 0x1747(%rip),%ymm1 # 4108 <_sk_callback_hsw+0x3c3>
+ .byte 196,226,125,88,13,135,23,0,0 // vpbroadcastd 0x1787(%rip),%ymm1 # 47c8 <_sk_callback_hsw+0x403>
.byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,58,23,0,0 // vbroadcastss 0x173a(%rip),%ymm2 # 410c <_sk_callback_hsw+0x3c7>
+ .byte 196,226,125,24,21,122,23,0,0 // vbroadcastss 0x177a(%rip),%ymm2 # 47cc <_sk_callback_hsw+0x407>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,88,21,49,23,0,0 // vpbroadcastd 0x1731(%rip),%ymm2 # 4110 <_sk_callback_hsw+0x3cb>
+ .byte 196,226,125,88,21,113,23,0,0 // vpbroadcastd 0x1771(%rip),%ymm2 # 47d0 <_sk_callback_hsw+0x40b>
.byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,36,23,0,0 // vbroadcastss 0x1724(%rip),%ymm8 # 4114 <_sk_callback_hsw+0x3cf>
+ .byte 196,98,125,24,5,100,23,0,0 // vbroadcastss 0x1764(%rip),%ymm8 # 47d4 <_sk_callback_hsw+0x40f>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,88,5,26,23,0,0 // vpbroadcastd 0x171a(%rip),%ymm8 # 4118 <_sk_callback_hsw+0x3d3>
+ .byte 196,98,125,88,5,90,23,0,0 // vpbroadcastd 0x175a(%rip),%ymm8 # 47d8 <_sk_callback_hsw+0x413>
.byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,12,23,0,0 // vbroadcastss 0x170c(%rip),%ymm8 # 411c <_sk_callback_hsw+0x3d7>
+ .byte 196,98,125,24,5,76,23,0,0 // vbroadcastss 0x174c(%rip),%ymm8 # 47dc <_sk_callback_hsw+0x417>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 91 // pop %rbx
@@ -10206,7 +11441,7 @@ FUNCTION(_sk_store_4444_hsw)
_sk_store_4444_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,242,22,0,0 // vbroadcastss 0x16f2(%rip),%ymm8 # 4120 <_sk_callback_hsw+0x3db>
+ .byte 196,98,125,24,5,50,23,0,0 // vbroadcastss 0x1732(%rip),%ymm8 # 47e0 <_sk_callback_hsw+0x41b>
.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
@@ -10224,7 +11459,7 @@ _sk_store_4444_hsw:
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 2a91 <_sk_store_4444_hsw+0x71>
+ .byte 117,10 // jne 3111 <_sk_store_4444_hsw+0x71>
.byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -10232,9 +11467,9 @@ _sk_store_4444_hsw:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 2a8d <_sk_store_4444_hsw+0x6d>
+ .byte 119,236 // ja 310d <_sk_store_4444_hsw+0x6d>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2af0 <_sk_store_4444_hsw+0xd0>
+ .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 3170 <_sk_store_4444_hsw+0xd0>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -10245,7 +11480,7 @@ _sk_store_4444_hsw:
.byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
.byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
.byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- .byte 235,159 // jmp 2a8d <_sk_store_4444_hsw+0x6d>
+ .byte 235,159 // jmp 310d <_sk_store_4444_hsw+0x6d>
.byte 102,144 // xchg %ax,%ax
.byte 245 // cmc
.byte 255 // (bad)
@@ -10280,16 +11515,16 @@ _sk_load_8888_hsw:
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
.byte 76,3,8 // add (%rax),%r9
.byte 77,133,192 // test %r8,%r8
- .byte 117,88 // jne 2b79 <_sk_load_8888_hsw+0x6d>
+ .byte 117,88 // jne 31f9 <_sk_load_8888_hsw+0x6d>
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
- .byte 197,229,219,5,178,23,0,0 // vpand 0x17b2(%rip),%ymm3,%ymm0 # 42e0 <_sk_callback_hsw+0x59b>
+ .byte 197,229,219,5,242,23,0,0 // vpand 0x17f2(%rip),%ymm3,%ymm0 # 49a0 <_sk_callback_hsw+0x5db>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,233,21,0,0 // vbroadcastss 0x15e9(%rip),%ymm8 # 4124 <_sk_callback_hsw+0x3df>
+ .byte 196,98,125,24,5,41,22,0,0 // vbroadcastss 0x1629(%rip),%ymm8 # 47e4 <_sk_callback_hsw+0x41f>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,183,23,0,0 // vpshufb 0x17b7(%rip),%ymm3,%ymm1 # 4300 <_sk_callback_hsw+0x5bb>
+ .byte 196,226,101,0,13,247,23,0,0 // vpshufb 0x17f7(%rip),%ymm3,%ymm1 # 49c0 <_sk_callback_hsw+0x5fb>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,197,23,0,0 // vpshufb 0x17c5(%rip),%ymm3,%ymm2 # 4320 <_sk_callback_hsw+0x5db>
+ .byte 196,226,101,0,21,5,24,0,0 // vpshufb 0x1805(%rip),%ymm3,%ymm2 # 49e0 <_sk_callback_hsw+0x61b>
.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
@@ -10306,7 +11541,7 @@ _sk_load_8888_hsw:
.byte 196,225,249,110,192 // vmovq %rax,%xmm0
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
.byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
- .byte 235,135 // jmp 2b26 <_sk_load_8888_hsw+0x1a>
+ .byte 235,135 // jmp 31a6 <_sk_load_8888_hsw+0x1a>
HIDDEN _sk_gather_8888_hsw
.globl _sk_gather_8888_hsw
@@ -10321,14 +11556,14 @@ _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 197,229,219,5,115,23,0,0 // vpand 0x1773(%rip),%ymm3,%ymm0 # 4340 <_sk_callback_hsw+0x5fb>
+ .byte 197,229,219,5,179,23,0,0 // vpand 0x17b3(%rip),%ymm3,%ymm0 # 4a00 <_sk_callback_hsw+0x63b>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,78,21,0,0 // vbroadcastss 0x154e(%rip),%ymm8 # 4128 <_sk_callback_hsw+0x3e3>
+ .byte 196,98,125,24,5,142,21,0,0 // vbroadcastss 0x158e(%rip),%ymm8 # 47e8 <_sk_callback_hsw+0x423>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,120,23,0,0 // vpshufb 0x1778(%rip),%ymm3,%ymm1 # 4360 <_sk_callback_hsw+0x61b>
+ .byte 196,226,101,0,13,184,23,0,0 // vpshufb 0x17b8(%rip),%ymm3,%ymm1 # 4a20 <_sk_callback_hsw+0x65b>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,134,23,0,0 // vpshufb 0x1786(%rip),%ymm3,%ymm2 # 4380 <_sk_callback_hsw+0x63b>
+ .byte 196,226,101,0,21,198,23,0,0 // vpshufb 0x17c6(%rip),%ymm3,%ymm2 # 4a40 <_sk_callback_hsw+0x67b>
.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
@@ -10345,7 +11580,7 @@ _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 196,98,125,24,5,254,20,0,0 // vbroadcastss 0x14fe(%rip),%ymm8 # 412c <_sk_callback_hsw+0x3e7>
+ .byte 196,98,125,24,5,62,21,0,0 // vbroadcastss 0x153e(%rip),%ymm8 # 47ec <_sk_callback_hsw+0x427>
.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
@@ -10361,7 +11596,7 @@ _sk_store_8888_hsw:
.byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8
.byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
.byte 77,133,192 // test %r8,%r8
- .byte 117,12 // jne 2c88 <_sk_store_8888_hsw+0x73>
+ .byte 117,12 // jne 3308 <_sk_store_8888_hsw+0x73>
.byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,193 // mov %r8,%rcx
@@ -10374,7 +11609,7 @@ _sk_store_8888_hsw:
.byte 196,97,249,110,200 // vmovq %rax,%xmm9
.byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9
.byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9)
- .byte 235,211 // jmp 2c81 <_sk_store_8888_hsw+0x6c>
+ .byte 235,211 // jmp 3301 <_sk_store_8888_hsw+0x6c>
HIDDEN _sk_load_f16_hsw
.globl _sk_load_f16_hsw
@@ -10383,7 +11618,7 @@ _sk_load_f16_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,97 // jne 2d19 <_sk_load_f16_hsw+0x6b>
+ .byte 117,97 // jne 3399 <_sk_load_f16_hsw+0x6b>
.byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8
.byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
.byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3
@@ -10409,29 +11644,29 @@ _sk_load_f16_hsw:
.byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,79 // je 2d78 <_sk_load_f16_hsw+0xca>
+ .byte 116,79 // je 33f8 <_sk_load_f16_hsw+0xca>
.byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,67 // jb 2d78 <_sk_load_f16_hsw+0xca>
+ .byte 114,67 // jb 33f8 <_sk_load_f16_hsw+0xca>
.byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,68 // je 2d85 <_sk_load_f16_hsw+0xd7>
+ .byte 116,68 // je 3405 <_sk_load_f16_hsw+0xd7>
.byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,56 // jb 2d85 <_sk_load_f16_hsw+0xd7>
+ .byte 114,56 // jb 3405 <_sk_load_f16_hsw+0xd7>
.byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,114,255,255,255 // je 2ccf <_sk_load_f16_hsw+0x21>
+ .byte 15,132,114,255,255,255 // je 334f <_sk_load_f16_hsw+0x21>
.byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,98,255,255,255 // jb 2ccf <_sk_load_f16_hsw+0x21>
+ .byte 15,130,98,255,255,255 // jb 334f <_sk_load_f16_hsw+0x21>
.byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9
- .byte 233,87,255,255,255 // jmpq 2ccf <_sk_load_f16_hsw+0x21>
+ .byte 233,87,255,255,255 // jmpq 334f <_sk_load_f16_hsw+0x21>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,74,255,255,255 // jmpq 2ccf <_sk_load_f16_hsw+0x21>
+ .byte 233,74,255,255,255 // jmpq 334f <_sk_load_f16_hsw+0x21>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,65,255,255,255 // jmpq 2ccf <_sk_load_f16_hsw+0x21>
+ .byte 233,65,255,255,255 // jmpq 334f <_sk_load_f16_hsw+0x21>
HIDDEN _sk_gather_f16_hsw
.globl _sk_gather_f16_hsw
@@ -10489,7 +11724,7 @@ _sk_store_f16_hsw:
.byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9
.byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,27 // jne 2e7d <_sk_store_f16_hsw+0x65>
+ .byte 117,27 // jne 34fd <_sk_store_f16_hsw+0x65>
.byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8)
.byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8)
.byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -10498,22 +11733,22 @@ _sk_store_f16_hsw:
.byte 255,224 // jmpq *%rax
.byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,241 // je 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 116,241 // je 34f9 <_sk_store_f16_hsw+0x61>
.byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,229 // jb 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 114,229 // jb 34f9 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
- .byte 116,221 // je 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 116,221 // je 34f9 <_sk_store_f16_hsw+0x61>
.byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,209 // jb 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 114,209 // jb 34f9 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
- .byte 116,201 // je 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 116,201 // je 34f9 <_sk_store_f16_hsw+0x61>
.byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,189 // jb 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 114,189 // jb 34f9 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
- .byte 235,181 // jmp 2e79 <_sk_store_f16_hsw+0x61>
+ .byte 235,181 // jmp 34f9 <_sk_store_f16_hsw+0x61>
HIDDEN _sk_load_u16_be_hsw
.globl _sk_load_u16_be_hsw
@@ -10523,7 +11758,7 @@ _sk_load_u16_be_hsw:
.byte 76,139,0 // mov (%rax),%r8
.byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,204,0,0,0 // jne 2fa6 <_sk_load_u16_be_hsw+0xe2>
+ .byte 15,133,204,0,0,0 // jne 3626 <_sk_load_u16_be_hsw+0xe2>
.byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8
.byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2
.byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3
@@ -10542,7 +11777,7 @@ _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 196,98,125,24,21,245,17,0,0 // vbroadcastss 0x11f5(%rip),%ymm10 # 4130 <_sk_callback_hsw+0x3eb>
+ .byte 196,98,125,24,21,53,18,0,0 // vbroadcastss 0x1235(%rip),%ymm10 # 47f0 <_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
@@ -10570,29 +11805,29 @@ _sk_load_u16_be_hsw:
.byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,85 // je 300c <_sk_load_u16_be_hsw+0x148>
+ .byte 116,85 // je 368c <_sk_load_u16_be_hsw+0x148>
.byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,72 // jb 300c <_sk_load_u16_be_hsw+0x148>
+ .byte 114,72 // jb 368c <_sk_load_u16_be_hsw+0x148>
.byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,72 // je 3019 <_sk_load_u16_be_hsw+0x155>
+ .byte 116,72 // je 3699 <_sk_load_u16_be_hsw+0x155>
.byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,59 // jb 3019 <_sk_load_u16_be_hsw+0x155>
+ .byte 114,59 // jb 3699 <_sk_load_u16_be_hsw+0x155>
.byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,6,255,255,255 // je 2ef5 <_sk_load_u16_be_hsw+0x31>
+ .byte 15,132,6,255,255,255 // je 3575 <_sk_load_u16_be_hsw+0x31>
.byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,245,254,255,255 // jb 2ef5 <_sk_load_u16_be_hsw+0x31>
+ .byte 15,130,245,254,255,255 // jb 3575 <_sk_load_u16_be_hsw+0x31>
.byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9
- .byte 233,233,254,255,255 // jmpq 2ef5 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,233,254,255,255 // jmpq 3575 <_sk_load_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,220,254,255,255 // jmpq 2ef5 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,220,254,255,255 // jmpq 3575 <_sk_load_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,211,254,255,255 // jmpq 2ef5 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,211,254,255,255 // jmpq 3575 <_sk_load_u16_be_hsw+0x31>
HIDDEN _sk_load_rgb_u16_be_hsw
.globl _sk_load_rgb_u16_be_hsw
@@ -10602,7 +11837,7 @@ _sk_load_rgb_u16_be_hsw:
.byte 76,139,0 // mov (%rax),%r8
.byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,204,0,0,0 // jne 3100 <_sk_load_rgb_u16_be_hsw+0xde>
+ .byte 15,133,204,0,0,0 // jne 3780 <_sk_load_rgb_u16_be_hsw+0xde>
.byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0
.byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2
.byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1
@@ -10626,7 +11861,7 @@ _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 196,98,125,24,21,134,16,0,0 // vbroadcastss 0x1086(%rip),%ymm10 # 4134 <_sk_callback_hsw+0x3ef>
+ .byte 196,98,125,24,21,198,16,0,0 // vbroadcastss 0x10c6(%rip),%ymm10 # 47f4 <_sk_callback_hsw+0x42f>
.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
@@ -10643,41 +11878,41 @@ _sk_load_rgb_u16_be_hsw:
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,58,16,0,0 // vbroadcastss 0x103a(%rip),%ymm3 # 4138 <_sk_callback_hsw+0x3f3>
+ .byte 196,226,125,24,29,122,16,0,0 // vbroadcastss 0x107a(%rip),%ymm3 # 47f8 <_sk_callback_hsw+0x433>
.byte 255,224 // jmpq *%rax
.byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0
.byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 117,5 // jne 3119 <_sk_load_rgb_u16_be_hsw+0xf7>
- .byte 233,79,255,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,5 // jne 3799 <_sk_load_rgb_u16_be_hsw+0xf7>
+ .byte 233,79,255,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
.byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1
.byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,26 // jb 3148 <_sk_load_rgb_u16_be_hsw+0x126>
+ .byte 114,26 // jb 37c8 <_sk_load_rgb_u16_be_hsw+0x126>
.byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1
.byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 117,10 // jne 314d <_sk_load_rgb_u16_be_hsw+0x12b>
- .byte 233,32,255,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,27,255,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 37cd <_sk_load_rgb_u16_be_hsw+0x12b>
+ .byte 233,32,255,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,27,255,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
.byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1
.byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,26 // jb 317c <_sk_load_rgb_u16_be_hsw+0x15a>
+ .byte 114,26 // jb 37fc <_sk_load_rgb_u16_be_hsw+0x15a>
.byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1
.byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 117,10 // jne 3181 <_sk_load_rgb_u16_be_hsw+0x15f>
- .byte 233,236,254,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,231,254,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 3801 <_sk_load_rgb_u16_be_hsw+0x15f>
+ .byte 233,236,254,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,231,254,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
.byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3
.byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,20 // jb 31aa <_sk_load_rgb_u16_be_hsw+0x188>
+ .byte 114,20 // jb 382a <_sk_load_rgb_u16_be_hsw+0x188>
.byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3
.byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
- .byte 233,190,254,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,185,254,255,255 // jmpq 3068 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,190,254,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,185,254,255,255 // jmpq 36e8 <_sk_load_rgb_u16_be_hsw+0x46>
HIDDEN _sk_store_u16_be_hsw
.globl _sk_store_u16_be_hsw
@@ -10686,7 +11921,7 @@ _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 196,98,125,24,5,119,15,0,0 // vbroadcastss 0xf77(%rip),%ymm8 # 413c <_sk_callback_hsw+0x3f7>
+ .byte 196,98,125,24,5,183,15,0,0 // vbroadcastss 0xfb7(%rip),%ymm8 # 47fc <_sk_callback_hsw+0x437>
.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
@@ -10724,7 +11959,7 @@ _sk_store_u16_be_hsw:
.byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9
.byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,31 // jne 32a9 <_sk_store_u16_be_hsw+0xfa>
+ .byte 117,31 // jne 3929 <_sk_store_u16_be_hsw+0xfa>
.byte 196,65,120,17,28,64 // vmovups %xmm11,(%r8,%rax,2)
.byte 196,65,120,17,84,64,16 // vmovups %xmm10,0x10(%r8,%rax,2)
.byte 196,65,120,17,76,64,32 // vmovups %xmm9,0x20(%r8,%rax,2)
@@ -10733,22 +11968,22 @@ _sk_store_u16_be_hsw:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,214,28,64 // vmovq %xmm11,(%r8,%rax,2)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,240 // je 3925 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,23,92,64,8 // vmovhpd %xmm11,0x8(%r8,%rax,2)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,227 // jb 3925 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,84,64,16 // vmovq %xmm10,0x10(%r8,%rax,2)
- .byte 116,218 // je 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,218 // je 3925 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,23,84,64,24 // vmovhpd %xmm10,0x18(%r8,%rax,2)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,205 // jb 3925 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,76,64,32 // vmovq %xmm9,0x20(%r8,%rax,2)
- .byte 116,196 // je 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,196 // je 3925 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,23,76,64,40 // vmovhpd %xmm9,0x28(%r8,%rax,2)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,183 // jb 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,183 // jb 3925 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,68,64,48 // vmovq %xmm8,0x30(%r8,%rax,2)
- .byte 235,174 // jmp 32a5 <_sk_store_u16_be_hsw+0xf6>
+ .byte 235,174 // jmp 3925 <_sk_store_u16_be_hsw+0xf6>
HIDDEN _sk_load_f32_hsw
.globl _sk_load_f32_hsw
@@ -10756,10 +11991,10 @@ FUNCTION(_sk_load_f32_hsw)
_sk_load_f32_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 119,110 // ja 336d <_sk_load_f32_hsw+0x76>
+ .byte 119,110 // ja 39ed <_sk_load_f32_hsw+0x76>
.byte 76,139,0 // mov (%rax),%r8
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
- .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 3398 <_sk_load_f32_hsw+0xa1>
+ .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 3a18 <_sk_load_f32_hsw+0xa1>
.byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -10820,7 +12055,7 @@ _sk_store_f32_hsw:
.byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
.byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,55 // jne 3425 <_sk_store_f32_hsw+0x6d>
+ .byte 117,55 // jne 3aa5 <_sk_store_f32_hsw+0x6d>
.byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
.byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
.byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
@@ -10833,22 +12068,22 @@ _sk_store_f32_hsw:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 3421 <_sk_store_f32_hsw+0x69>
+ .byte 116,240 // je 3aa1 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 3421 <_sk_store_f32_hsw+0x69>
+ .byte 114,227 // jb 3aa1 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
- .byte 116,218 // je 3421 <_sk_store_f32_hsw+0x69>
+ .byte 116,218 // je 3aa1 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 3421 <_sk_store_f32_hsw+0x69>
+ .byte 114,205 // jb 3aa1 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
- .byte 116,195 // je 3421 <_sk_store_f32_hsw+0x69>
+ .byte 116,195 // je 3aa1 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,181 // jb 3421 <_sk_store_f32_hsw+0x69>
+ .byte 114,181 // jb 3aa1 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
- .byte 235,171 // jmp 3421 <_sk_store_f32_hsw+0x69>
+ .byte 235,171 // jmp 3aa1 <_sk_store_f32_hsw+0x69>
HIDDEN _sk_clamp_x_hsw
.globl _sk_clamp_x_hsw
@@ -10958,11 +12193,11 @@ HIDDEN _sk_luminance_to_alpha_hsw
.globl _sk_luminance_to_alpha_hsw
FUNCTION(_sk_luminance_to_alpha_hsw)
_sk_luminance_to_alpha_hsw:
- .byte 196,226,125,24,29,145,11,0,0 // vbroadcastss 0xb91(%rip),%ymm3 # 4140 <_sk_callback_hsw+0x3fb>
- .byte 196,98,125,24,5,140,11,0,0 // vbroadcastss 0xb8c(%rip),%ymm8 # 4144 <_sk_callback_hsw+0x3ff>
+ .byte 196,226,125,24,29,209,11,0,0 // vbroadcastss 0xbd1(%rip),%ymm3 # 4800 <_sk_callback_hsw+0x43b>
+ .byte 196,98,125,24,5,204,11,0,0 // vbroadcastss 0xbcc(%rip),%ymm8 # 4804 <_sk_callback_hsw+0x43f>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
.byte 196,226,125,184,203 // vfmadd231ps %ymm3,%ymm0,%ymm1
- .byte 196,226,125,24,29,125,11,0,0 // vbroadcastss 0xb7d(%rip),%ymm3 # 4148 <_sk_callback_hsw+0x403>
+ .byte 196,226,125,24,29,189,11,0,0 // vbroadcastss 0xbbd(%rip),%ymm3 # 4808 <_sk_callback_hsw+0x443>
.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
@@ -11107,7 +12342,7 @@ _sk_linear_gradient_hsw:
.byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9
.byte 76,139,0 // mov (%rax),%r8
.byte 77,133,192 // test %r8,%r8
- .byte 15,132,143,0,0,0 // je 38a3 <_sk_linear_gradient_hsw+0xb5>
+ .byte 15,132,143,0,0,0 // je 3f23 <_sk_linear_gradient_hsw+0xb5>
.byte 72,139,64,8 // mov 0x8(%rax),%rax
.byte 72,131,192,32 // add $0x20,%rax
.byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12
@@ -11134,8 +12369,8 @@ _sk_linear_gradient_hsw:
.byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9
.byte 72,131,192,36 // add $0x24,%rax
.byte 73,255,200 // dec %r8
- .byte 117,140 // jne 382d <_sk_linear_gradient_hsw+0x3f>
- .byte 235,17 // jmp 38b4 <_sk_linear_gradient_hsw+0xc6>
+ .byte 117,140 // jne 3ead <_sk_linear_gradient_hsw+0x3f>
+ .byte 235,17 // jmp 3f34 <_sk_linear_gradient_hsw+0xc6>
.byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
.byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
.byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
@@ -11182,24 +12417,24 @@ _sk_xy_to_polar_unit_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 196,98,125,24,45,253,7,0,0 // vbroadcastss 0x7fd(%rip),%ymm13 # 414c <_sk_callback_hsw+0x407>
- .byte 196,98,125,24,53,248,7,0,0 // vbroadcastss 0x7f8(%rip),%ymm14 # 4150 <_sk_callback_hsw+0x40b>
+ .byte 196,98,125,24,45,61,8,0,0 // vbroadcastss 0x83d(%rip),%ymm13 # 480c <_sk_callback_hsw+0x447>
+ .byte 196,98,125,24,53,56,8,0,0 // vbroadcastss 0x838(%rip),%ymm14 # 4810 <_sk_callback_hsw+0x44b>
.byte 196,66,29,184,245 // vfmadd231ps %ymm13,%ymm12,%ymm14
- .byte 196,98,125,24,45,238,7,0,0 // vbroadcastss 0x7ee(%rip),%ymm13 # 4154 <_sk_callback_hsw+0x40f>
+ .byte 196,98,125,24,45,46,8,0,0 // vbroadcastss 0x82e(%rip),%ymm13 # 4814 <_sk_callback_hsw+0x44f>
.byte 196,66,29,184,238 // vfmadd231ps %ymm14,%ymm12,%ymm13
- .byte 196,98,125,24,53,228,7,0,0 // vbroadcastss 0x7e4(%rip),%ymm14 # 4158 <_sk_callback_hsw+0x413>
+ .byte 196,98,125,24,53,36,8,0,0 // vbroadcastss 0x824(%rip),%ymm14 # 4818 <_sk_callback_hsw+0x453>
.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 196,98,125,24,21,207,7,0,0 // vbroadcastss 0x7cf(%rip),%ymm10 # 415c <_sk_callback_hsw+0x417>
+ .byte 196,98,125,24,21,15,8,0,0 // vbroadcastss 0x80f(%rip),%ymm10 # 481c <_sk_callback_hsw+0x457>
.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 196,98,125,24,21,185,7,0,0 // vbroadcastss 0x7b9(%rip),%ymm10 # 4160 <_sk_callback_hsw+0x41b>
+ .byte 196,98,125,24,21,249,7,0,0 // vbroadcastss 0x7f9(%rip),%ymm10 # 4820 <_sk_callback_hsw+0x45b>
.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 196,98,125,24,21,163,7,0,0 // vbroadcastss 0x7a3(%rip),%ymm10 # 4164 <_sk_callback_hsw+0x41f>
+ .byte 196,98,125,24,21,227,7,0,0 // vbroadcastss 0x7e3(%rip),%ymm10 # 4824 <_sk_callback_hsw+0x45f>
.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
@@ -11212,7 +12447,7 @@ HIDDEN _sk_save_xy_hsw
FUNCTION(_sk_save_xy_hsw)
_sk_save_xy_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,130,7,0,0 // vbroadcastss 0x782(%rip),%ymm8 # 4168 <_sk_callback_hsw+0x423>
+ .byte 196,98,125,24,5,194,7,0,0 // vbroadcastss 0x7c2(%rip),%ymm8 # 4828 <_sk_callback_hsw+0x463>
.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
@@ -11246,9 +12481,9 @@ HIDDEN _sk_bilinear_nx_hsw
FUNCTION(_sk_bilinear_nx_hsw)
_sk_bilinear_nx_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,22,7,0,0 // vbroadcastss 0x716(%rip),%ymm0 # 416c <_sk_callback_hsw+0x427>
+ .byte 196,226,125,24,5,86,7,0,0 // vbroadcastss 0x756(%rip),%ymm0 # 482c <_sk_callback_hsw+0x467>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,13,7,0,0 // vbroadcastss 0x70d(%rip),%ymm8 # 4170 <_sk_callback_hsw+0x42b>
+ .byte 196,98,125,24,5,77,7,0,0 // vbroadcastss 0x74d(%rip),%ymm8 # 4830 <_sk_callback_hsw+0x46b>
.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
@@ -11259,7 +12494,7 @@ HIDDEN _sk_bilinear_px_hsw
FUNCTION(_sk_bilinear_px_hsw)
_sk_bilinear_px_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,245,6,0,0 // vbroadcastss 0x6f5(%rip),%ymm0 # 4174 <_sk_callback_hsw+0x42f>
+ .byte 196,226,125,24,5,53,7,0,0 // vbroadcastss 0x735(%rip),%ymm0 # 4834 <_sk_callback_hsw+0x46f>
.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)
@@ -11271,9 +12506,9 @@ HIDDEN _sk_bilinear_ny_hsw
FUNCTION(_sk_bilinear_ny_hsw)
_sk_bilinear_ny_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,217,6,0,0 // vbroadcastss 0x6d9(%rip),%ymm1 # 4178 <_sk_callback_hsw+0x433>
+ .byte 196,226,125,24,13,25,7,0,0 // vbroadcastss 0x719(%rip),%ymm1 # 4838 <_sk_callback_hsw+0x473>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,207,6,0,0 // vbroadcastss 0x6cf(%rip),%ymm8 # 417c <_sk_callback_hsw+0x437>
+ .byte 196,98,125,24,5,15,7,0,0 // vbroadcastss 0x70f(%rip),%ymm8 # 483c <_sk_callback_hsw+0x477>
.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
@@ -11284,7 +12519,7 @@ HIDDEN _sk_bilinear_py_hsw
FUNCTION(_sk_bilinear_py_hsw)
_sk_bilinear_py_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,183,6,0,0 // vbroadcastss 0x6b7(%rip),%ymm1 # 4180 <_sk_callback_hsw+0x43b>
+ .byte 196,226,125,24,13,247,6,0,0 // vbroadcastss 0x6f7(%rip),%ymm1 # 4840 <_sk_callback_hsw+0x47b>
.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)
@@ -11296,13 +12531,13 @@ HIDDEN _sk_bicubic_n3x_hsw
FUNCTION(_sk_bicubic_n3x_hsw)
_sk_bicubic_n3x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,154,6,0,0 // vbroadcastss 0x69a(%rip),%ymm0 # 4184 <_sk_callback_hsw+0x43f>
+ .byte 196,226,125,24,5,218,6,0,0 // vbroadcastss 0x6da(%rip),%ymm0 # 4844 <_sk_callback_hsw+0x47f>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,145,6,0,0 // vbroadcastss 0x691(%rip),%ymm8 # 4188 <_sk_callback_hsw+0x443>
+ .byte 196,98,125,24,5,209,6,0,0 // vbroadcastss 0x6d1(%rip),%ymm8 # 4848 <_sk_callback_hsw+0x483>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,130,6,0,0 // vbroadcastss 0x682(%rip),%ymm10 # 418c <_sk_callback_hsw+0x447>
- .byte 196,98,125,24,29,125,6,0,0 // vbroadcastss 0x67d(%rip),%ymm11 # 4190 <_sk_callback_hsw+0x44b>
+ .byte 196,98,125,24,21,194,6,0,0 // vbroadcastss 0x6c2(%rip),%ymm10 # 484c <_sk_callback_hsw+0x487>
+ .byte 196,98,125,24,29,189,6,0,0 // vbroadcastss 0x6bd(%rip),%ymm11 # 4850 <_sk_callback_hsw+0x48b>
.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)
@@ -11314,16 +12549,16 @@ HIDDEN _sk_bicubic_n1x_hsw
FUNCTION(_sk_bicubic_n1x_hsw)
_sk_bicubic_n1x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,96,6,0,0 // vbroadcastss 0x660(%rip),%ymm0 # 4194 <_sk_callback_hsw+0x44f>
+ .byte 196,226,125,24,5,160,6,0,0 // vbroadcastss 0x6a0(%rip),%ymm0 # 4854 <_sk_callback_hsw+0x48f>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,87,6,0,0 // vbroadcastss 0x657(%rip),%ymm8 # 4198 <_sk_callback_hsw+0x453>
+ .byte 196,98,125,24,5,151,6,0,0 // vbroadcastss 0x697(%rip),%ymm8 # 4858 <_sk_callback_hsw+0x493>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,77,6,0,0 // vbroadcastss 0x64d(%rip),%ymm9 # 419c <_sk_callback_hsw+0x457>
- .byte 196,98,125,24,21,72,6,0,0 // vbroadcastss 0x648(%rip),%ymm10 # 41a0 <_sk_callback_hsw+0x45b>
+ .byte 196,98,125,24,13,141,6,0,0 // vbroadcastss 0x68d(%rip),%ymm9 # 485c <_sk_callback_hsw+0x497>
+ .byte 196,98,125,24,21,136,6,0,0 // vbroadcastss 0x688(%rip),%ymm10 # 4860 <_sk_callback_hsw+0x49b>
.byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,13,62,6,0,0 // vbroadcastss 0x63e(%rip),%ymm9 # 41a4 <_sk_callback_hsw+0x45f>
+ .byte 196,98,125,24,13,126,6,0,0 // vbroadcastss 0x67e(%rip),%ymm9 # 4864 <_sk_callback_hsw+0x49f>
.byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
- .byte 196,98,125,24,21,52,6,0,0 // vbroadcastss 0x634(%rip),%ymm10 # 41a8 <_sk_callback_hsw+0x463>
+ .byte 196,98,125,24,21,116,6,0,0 // vbroadcastss 0x674(%rip),%ymm10 # 4868 <_sk_callback_hsw+0x4a3>
.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
@@ -11334,14 +12569,14 @@ HIDDEN _sk_bicubic_p1x_hsw
FUNCTION(_sk_bicubic_p1x_hsw)
_sk_bicubic_p1x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,28,6,0,0 // vbroadcastss 0x61c(%rip),%ymm8 # 41ac <_sk_callback_hsw+0x467>
+ .byte 196,98,125,24,5,92,6,0,0 // vbroadcastss 0x65c(%rip),%ymm8 # 486c <_sk_callback_hsw+0x4a7>
.byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
.byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
- .byte 196,98,125,24,21,14,6,0,0 // vbroadcastss 0x60e(%rip),%ymm10 # 41b0 <_sk_callback_hsw+0x46b>
- .byte 196,98,125,24,29,9,6,0,0 // vbroadcastss 0x609(%rip),%ymm11 # 41b4 <_sk_callback_hsw+0x46f>
+ .byte 196,98,125,24,21,78,6,0,0 // vbroadcastss 0x64e(%rip),%ymm10 # 4870 <_sk_callback_hsw+0x4ab>
+ .byte 196,98,125,24,29,73,6,0,0 // vbroadcastss 0x649(%rip),%ymm11 # 4874 <_sk_callback_hsw+0x4af>
.byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11
.byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11
- .byte 196,98,125,24,5,250,5,0,0 // vbroadcastss 0x5fa(%rip),%ymm8 # 41b8 <_sk_callback_hsw+0x473>
+ .byte 196,98,125,24,5,58,6,0,0 // vbroadcastss 0x63a(%rip),%ymm8 # 4878 <_sk_callback_hsw+0x4b3>
.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
@@ -11352,12 +12587,12 @@ HIDDEN _sk_bicubic_p3x_hsw
FUNCTION(_sk_bicubic_p3x_hsw)
_sk_bicubic_p3x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,226,5,0,0 // vbroadcastss 0x5e2(%rip),%ymm0 # 41bc <_sk_callback_hsw+0x477>
+ .byte 196,226,125,24,5,34,6,0,0 // vbroadcastss 0x622(%rip),%ymm0 # 487c <_sk_callback_hsw+0x4b7>
.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 196,98,125,24,21,207,5,0,0 // vbroadcastss 0x5cf(%rip),%ymm10 # 41c0 <_sk_callback_hsw+0x47b>
- .byte 196,98,125,24,29,202,5,0,0 // vbroadcastss 0x5ca(%rip),%ymm11 # 41c4 <_sk_callback_hsw+0x47f>
+ .byte 196,98,125,24,21,15,6,0,0 // vbroadcastss 0x60f(%rip),%ymm10 # 4880 <_sk_callback_hsw+0x4bb>
+ .byte 196,98,125,24,29,10,6,0,0 // vbroadcastss 0x60a(%rip),%ymm11 # 4884 <_sk_callback_hsw+0x4bf>
.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)
@@ -11369,13 +12604,13 @@ HIDDEN _sk_bicubic_n3y_hsw
FUNCTION(_sk_bicubic_n3y_hsw)
_sk_bicubic_n3y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,173,5,0,0 // vbroadcastss 0x5ad(%rip),%ymm1 # 41c8 <_sk_callback_hsw+0x483>
+ .byte 196,226,125,24,13,237,5,0,0 // vbroadcastss 0x5ed(%rip),%ymm1 # 4888 <_sk_callback_hsw+0x4c3>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,163,5,0,0 // vbroadcastss 0x5a3(%rip),%ymm8 # 41cc <_sk_callback_hsw+0x487>
+ .byte 196,98,125,24,5,227,5,0,0 // vbroadcastss 0x5e3(%rip),%ymm8 # 488c <_sk_callback_hsw+0x4c7>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,148,5,0,0 // vbroadcastss 0x594(%rip),%ymm10 # 41d0 <_sk_callback_hsw+0x48b>
- .byte 196,98,125,24,29,143,5,0,0 // vbroadcastss 0x58f(%rip),%ymm11 # 41d4 <_sk_callback_hsw+0x48f>
+ .byte 196,98,125,24,21,212,5,0,0 // vbroadcastss 0x5d4(%rip),%ymm10 # 4890 <_sk_callback_hsw+0x4cb>
+ .byte 196,98,125,24,29,207,5,0,0 // vbroadcastss 0x5cf(%rip),%ymm11 # 4894 <_sk_callback_hsw+0x4cf>
.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)
@@ -11387,16 +12622,16 @@ HIDDEN _sk_bicubic_n1y_hsw
FUNCTION(_sk_bicubic_n1y_hsw)
_sk_bicubic_n1y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,114,5,0,0 // vbroadcastss 0x572(%rip),%ymm1 # 41d8 <_sk_callback_hsw+0x493>
+ .byte 196,226,125,24,13,178,5,0,0 // vbroadcastss 0x5b2(%rip),%ymm1 # 4898 <_sk_callback_hsw+0x4d3>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,104,5,0,0 // vbroadcastss 0x568(%rip),%ymm8 # 41dc <_sk_callback_hsw+0x497>
+ .byte 196,98,125,24,5,168,5,0,0 // vbroadcastss 0x5a8(%rip),%ymm8 # 489c <_sk_callback_hsw+0x4d7>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm9 # 41e0 <_sk_callback_hsw+0x49b>
- .byte 196,98,125,24,21,89,5,0,0 // vbroadcastss 0x559(%rip),%ymm10 # 41e4 <_sk_callback_hsw+0x49f>
+ .byte 196,98,125,24,13,158,5,0,0 // vbroadcastss 0x59e(%rip),%ymm9 # 48a0 <_sk_callback_hsw+0x4db>
+ .byte 196,98,125,24,21,153,5,0,0 // vbroadcastss 0x599(%rip),%ymm10 # 48a4 <_sk_callback_hsw+0x4df>
.byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,13,79,5,0,0 // vbroadcastss 0x54f(%rip),%ymm9 # 41e8 <_sk_callback_hsw+0x4a3>
+ .byte 196,98,125,24,13,143,5,0,0 // vbroadcastss 0x58f(%rip),%ymm9 # 48a8 <_sk_callback_hsw+0x4e3>
.byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
- .byte 196,98,125,24,21,69,5,0,0 // vbroadcastss 0x545(%rip),%ymm10 # 41ec <_sk_callback_hsw+0x4a7>
+ .byte 196,98,125,24,21,133,5,0,0 // vbroadcastss 0x585(%rip),%ymm10 # 48ac <_sk_callback_hsw+0x4e7>
.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
@@ -11407,14 +12642,14 @@ HIDDEN _sk_bicubic_p1y_hsw
FUNCTION(_sk_bicubic_p1y_hsw)
_sk_bicubic_p1y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,45,5,0,0 // vbroadcastss 0x52d(%rip),%ymm8 # 41f0 <_sk_callback_hsw+0x4ab>
+ .byte 196,98,125,24,5,109,5,0,0 // vbroadcastss 0x56d(%rip),%ymm8 # 48b0 <_sk_callback_hsw+0x4eb>
.byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
.byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
- .byte 196,98,125,24,21,30,5,0,0 // vbroadcastss 0x51e(%rip),%ymm10 # 41f4 <_sk_callback_hsw+0x4af>
- .byte 196,98,125,24,29,25,5,0,0 // vbroadcastss 0x519(%rip),%ymm11 # 41f8 <_sk_callback_hsw+0x4b3>
+ .byte 196,98,125,24,21,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm10 # 48b4 <_sk_callback_hsw+0x4ef>
+ .byte 196,98,125,24,29,89,5,0,0 // vbroadcastss 0x559(%rip),%ymm11 # 48b8 <_sk_callback_hsw+0x4f3>
.byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11
.byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11
- .byte 196,98,125,24,5,10,5,0,0 // vbroadcastss 0x50a(%rip),%ymm8 # 41fc <_sk_callback_hsw+0x4b7>
+ .byte 196,98,125,24,5,74,5,0,0 // vbroadcastss 0x54a(%rip),%ymm8 # 48bc <_sk_callback_hsw+0x4f7>
.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
@@ -11425,12 +12660,12 @@ HIDDEN _sk_bicubic_p3y_hsw
FUNCTION(_sk_bicubic_p3y_hsw)
_sk_bicubic_p3y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,242,4,0,0 // vbroadcastss 0x4f2(%rip),%ymm1 # 4200 <_sk_callback_hsw+0x4bb>
+ .byte 196,226,125,24,13,50,5,0,0 // vbroadcastss 0x532(%rip),%ymm1 # 48c0 <_sk_callback_hsw+0x4fb>
.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 196,98,125,24,21,222,4,0,0 // vbroadcastss 0x4de(%rip),%ymm10 # 4204 <_sk_callback_hsw+0x4bf>
- .byte 196,98,125,24,29,217,4,0,0 // vbroadcastss 0x4d9(%rip),%ymm11 # 4208 <_sk_callback_hsw+0x4c3>
+ .byte 196,98,125,24,21,30,5,0,0 // vbroadcastss 0x51e(%rip),%ymm10 # 48c4 <_sk_callback_hsw+0x4ff>
+ .byte 196,98,125,24,29,25,5,0,0 // vbroadcastss 0x519(%rip),%ymm11 # 48c8 <_sk_callback_hsw+0x503>
.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)
@@ -11551,6 +12786,30 @@ BALIGN4
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,191,0,0,224 // add %al,-0x1fffff41(%rax)
.byte 64,0,0 // 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 45a1 <.literal4+0xb5>
+ .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 45b1 <.literal4+0xc5>
+ .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 45c1 <.literal4+0xd5>
+ .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 45d1 <.literal4+0xe5>
+ .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)
@@ -11598,7 +12857,7 @@ 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 3f5d <.literal4+0xf1>
+ .byte 224,7 // loopne 461d <.literal4+0x131>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -11612,10 +12871,10 @@ 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 3f88 <.literal4+0x11c>
+ .byte 127,0 // jg 4648 <.literal4+0x15c>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4001 <.literal4+0x195>
+ .byte 119,115 // ja 46c1 <.literal4+0x1d5>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11629,10 +12888,10 @@ 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 3fbc <.literal4+0x150>
+ .byte 127,0 // jg 467c <.literal4+0x190>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4035 <.literal4+0x1c9>
+ .byte 119,115 // ja 46f5 <.literal4+0x209>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11646,10 +12905,10 @@ 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 3ff0 <.literal4+0x184>
+ .byte 127,0 // jg 46b0 <.literal4+0x1c4>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4069 <.literal4+0x1fd>
+ .byte 119,115 // ja 4729 <.literal4+0x23d>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11663,10 +12922,10 @@ 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 4024 <.literal4+0x1b8>
+ .byte 127,0 // jg 46e4 <.literal4+0x1f8>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 409d <.literal4+0x231>
+ .byte 119,115 // ja 475d <.literal4+0x271>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11679,7 +12938,7 @@ 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 409b <.literal4+0x22f>
+ .byte 127,67 // jg 475b <.literal4+0x26f>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -11691,10 +12950,10 @@ BALIGN4
.byte 190,80,128,3,62 // mov $0x3e038050,%esi
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 40bb <.literal4+0x24f>
+ .byte 118,63 // jbe 477b <.literal4+0x28f>
.byte 246,64,83,63 // testb $0x3f,0x53(%rax)
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 40cf <.literal4+0x263>
+ .byte 127,67 // jg 478f <.literal4+0x2a3>
.byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax)
.byte 128,59,0 // cmpb $0x0,(%rbx)
.byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax)
@@ -11703,7 +12962,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 40b1 <.literal4+0x245>
+ .byte 224,7 // loopne 4771 <.literal4+0x285>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -11715,7 +12974,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 40cd <.literal4+0x261>
+ .byte 224,7 // loopne 478d <.literal4+0x2a1>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -11726,7 +12985,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 4122 <.literal4+0x2b6>
+ .byte 124,66 // jl 47e2 <.literal4+0x2f6>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax)
@@ -11744,9 +13003,9 @@ 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 4165 <.literal4+0x2f9>
+ .byte 112,65 // jo 4825 <.literal4+0x339>
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 4173 <.literal4+0x307>
+ .byte 127,67 // jg 4833 <.literal4+0x347>
.byte 128,0,128 // addb $0x80,(%rax)
.byte 55 // (bad)
.byte 128,0,128 // addb $0x80,(%rax)
@@ -11754,7 +13013,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 255 // (bad)
- .byte 127,71 // jg 4187 <.literal4+0x31b>
+ .byte 127,71 // jg 4847 <.literal4+0x35b>
.byte 208 // (bad)
.byte 179,89 // mov $0x59,%bl
.byte 62,89 // ds pop %rcx
@@ -11851,16 +13110,16 @@ 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) # a004248 <_sk_callback_hsw+0xa000503>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004908 <_sk_callback_hsw+0xa000543>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004250 <_sk_callback_hsw+0x1200050b>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004910 <_sk_callback_hsw+0x1200054b>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004258 <_sk_callback_hsw+0x1a000513>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004918 <_sk_callback_hsw+0x1a000553>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004260 <_sk_callback_hsw+0x300051b>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004920 <_sk_callback_hsw+0x300055b>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -11903,16 +13162,16 @@ 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) # a0042a8 <_sk_callback_hsw+0xa000563>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004968 <_sk_callback_hsw+0xa0005a3>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120042b0 <_sk_callback_hsw+0x1200056b>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004970 <_sk_callback_hsw+0x120005ab>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0042b8 <_sk_callback_hsw+0x1a000573>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004978 <_sk_callback_hsw+0x1a0005b3>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30042c0 <_sk_callback_hsw+0x300057b>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004980 <_sk_callback_hsw+0x30005bb>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -11955,16 +13214,16 @@ 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) # a004308 <_sk_callback_hsw+0xa0005c3>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0049c8 <_sk_callback_hsw+0xa000603>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004310 <_sk_callback_hsw+0x120005cb>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120049d0 <_sk_callback_hsw+0x1200060b>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004318 <_sk_callback_hsw+0x1a0005d3>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0049d8 <_sk_callback_hsw+0x1a000613>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004320 <_sk_callback_hsw+0x30005db>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30049e0 <_sk_callback_hsw+0x300061b>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -12007,16 +13266,16 @@ 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) # a004368 <_sk_callback_hsw+0xa000623>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004a28 <_sk_callback_hsw+0xa000663>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004370 <_sk_callback_hsw+0x1200062b>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004a30 <_sk_callback_hsw+0x1200066b>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004378 <_sk_callback_hsw+0x1a000633>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004a38 <_sk_callback_hsw+0x1a000673>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004380 <_sk_callback_hsw+0x300063b>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004a40 <_sk_callback_hsw+0x300067b>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -12137,14 +13396,14 @@ _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 196,226,125,24,13,203,84,0,0 // vbroadcastss 0x54cb(%rip),%ymm1 # 5594 <_sk_callback_avx+0x125>
+ .byte 196,226,125,24,13,191,91,0,0 // vbroadcastss 0x5bbf(%rip),%ymm1 # 5c88 <_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 196,226,125,24,21,175,84,0,0 // vbroadcastss 0x54af(%rip),%ymm2 # 5598 <_sk_callback_avx+0x129>
+ .byte 196,226,125,24,21,163,91,0,0 // vbroadcastss 0x5ba3(%rip),%ymm2 # 5c8c <_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
@@ -12166,7 +13425,7 @@ _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 196,98,125,24,29,103,84,0,0 // vbroadcastss 0x5467(%rip),%ymm11 # 559c <_sk_callback_avx+0x12d>
+ .byte 196,98,125,24,29,91,91,0,0 // vbroadcastss 0x5b5b(%rip),%ymm11 # 5c90 <_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
@@ -12177,8 +13436,8 @@ _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 196,98,125,24,37,40,84,0,0 // vbroadcastss 0x5428(%rip),%ymm12 # 55a0 <_sk_callback_avx+0x131>
- .byte 196,98,125,24,45,35,84,0,0 // vbroadcastss 0x5423(%rip),%ymm13 # 55a4 <_sk_callback_avx+0x135>
+ .byte 196,98,125,24,37,28,91,0,0 // vbroadcastss 0x5b1c(%rip),%ymm12 # 5c94 <_sk_callback_avx+0x131>
+ .byte 196,98,125,24,45,23,91,0,0 // vbroadcastss 0x5b17(%rip),%ymm13 # 5c98 <_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
@@ -12205,9 +13464,9 @@ _sk_dither_avx:
.byte 196,65,60,86,193 // vorps %ymm9,%ymm8,%ymm8
.byte 196,65,60,86,194 // vorps %ymm10,%ymm8,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,142,83,0,0 // vbroadcastss 0x538e(%rip),%ymm9 # 55a8 <_sk_callback_avx+0x139>
+ .byte 196,98,125,24,13,130,90,0,0 // vbroadcastss 0x5a82(%rip),%ymm9 # 5c9c <_sk_callback_avx+0x139>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,132,83,0,0 // vbroadcastss 0x5384(%rip),%ymm9 # 55ac <_sk_callback_avx+0x13d>
+ .byte 196,98,125,24,13,120,90,0,0 // vbroadcastss 0x5a78(%rip),%ymm9 # 5ca0 <_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
@@ -12245,7 +13504,7 @@ HIDDEN _sk_srcatop_avx
FUNCTION(_sk_srcatop_avx)
_sk_srcatop_avx:
.byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
- .byte 196,98,125,24,5,42,83,0,0 // vbroadcastss 0x532a(%rip),%ymm8 # 55b0 <_sk_callback_avx+0x141>
+ .byte 196,98,125,24,5,30,90,0,0 // vbroadcastss 0x5a1e(%rip),%ymm8 # 5ca4 <_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
@@ -12266,7 +13525,7 @@ HIDDEN _sk_dstatop_avx
FUNCTION(_sk_dstatop_avx)
_sk_dstatop_avx:
.byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8
- .byte 196,98,125,24,13,236,82,0,0 // vbroadcastss 0x52ec(%rip),%ymm9 # 55b4 <_sk_callback_avx+0x145>
+ .byte 196,98,125,24,13,224,89,0,0 // vbroadcastss 0x59e0(%rip),%ymm9 # 5ca8 <_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
@@ -12308,7 +13567,7 @@ HIDDEN _sk_srcout_avx
.globl _sk_srcout_avx
FUNCTION(_sk_srcout_avx)
_sk_srcout_avx:
- .byte 196,98,125,24,5,139,82,0,0 // vbroadcastss 0x528b(%rip),%ymm8 # 55b8 <_sk_callback_avx+0x149>
+ .byte 196,98,125,24,5,127,89,0,0 // vbroadcastss 0x597f(%rip),%ymm8 # 5cac <_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
@@ -12321,7 +13580,7 @@ HIDDEN _sk_dstout_avx
.globl _sk_dstout_avx
FUNCTION(_sk_dstout_avx)
_sk_dstout_avx:
- .byte 196,226,125,24,5,110,82,0,0 // vbroadcastss 0x526e(%rip),%ymm0 # 55bc <_sk_callback_avx+0x14d>
+ .byte 196,226,125,24,5,98,89,0,0 // vbroadcastss 0x5962(%rip),%ymm0 # 5cb0 <_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
@@ -12334,7 +13593,7 @@ HIDDEN _sk_srcover_avx
.globl _sk_srcover_avx
FUNCTION(_sk_srcover_avx)
_sk_srcover_avx:
- .byte 196,98,125,24,5,81,82,0,0 // vbroadcastss 0x5251(%rip),%ymm8 # 55c0 <_sk_callback_avx+0x151>
+ .byte 196,98,125,24,5,69,89,0,0 // vbroadcastss 0x5945(%rip),%ymm8 # 5cb4 <_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
@@ -12351,7 +13610,7 @@ HIDDEN _sk_dstover_avx
.globl _sk_dstover_avx
FUNCTION(_sk_dstover_avx)
_sk_dstover_avx:
- .byte 196,98,125,24,5,36,82,0,0 // vbroadcastss 0x5224(%rip),%ymm8 # 55c4 <_sk_callback_avx+0x155>
+ .byte 196,98,125,24,5,24,89,0,0 // vbroadcastss 0x5918(%rip),%ymm8 # 5cb8 <_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
@@ -12379,7 +13638,7 @@ HIDDEN _sk_multiply_avx
.globl _sk_multiply_avx
FUNCTION(_sk_multiply_avx)
_sk_multiply_avx:
- .byte 196,98,125,24,5,227,81,0,0 // vbroadcastss 0x51e3(%rip),%ymm8 # 55c8 <_sk_callback_avx+0x159>
+ .byte 196,98,125,24,5,215,88,0,0 // vbroadcastss 0x58d7(%rip),%ymm8 # 5cbc <_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
@@ -12439,7 +13698,7 @@ HIDDEN _sk_xor__avx
.globl _sk_xor__avx
FUNCTION(_sk_xor__avx)
_sk_xor__avx:
- .byte 196,98,125,24,5,50,81,0,0 // vbroadcastss 0x5132(%rip),%ymm8 # 55cc <_sk_callback_avx+0x15d>
+ .byte 196,98,125,24,5,38,88,0,0 // vbroadcastss 0x5826(%rip),%ymm8 # 5cc0 <_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
@@ -12476,7 +13735,7 @@ _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 196,98,125,24,5,178,80,0,0 // vbroadcastss 0x50b2(%rip),%ymm8 # 55d0 <_sk_callback_avx+0x161>
+ .byte 196,98,125,24,5,166,87,0,0 // vbroadcastss 0x57a6(%rip),%ymm8 # 5cc4 <_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
@@ -12502,7 +13761,7 @@ _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 196,98,125,24,5,94,80,0,0 // vbroadcastss 0x505e(%rip),%ymm8 # 55d4 <_sk_callback_avx+0x165>
+ .byte 196,98,125,24,5,82,87,0,0 // vbroadcastss 0x5752(%rip),%ymm8 # 5cc8 <_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
@@ -12531,7 +13790,7 @@ _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 196,98,125,24,5,254,79,0,0 // vbroadcastss 0x4ffe(%rip),%ymm8 # 55d8 <_sk_callback_avx+0x169>
+ .byte 196,98,125,24,5,242,86,0,0 // vbroadcastss 0x56f2(%rip),%ymm8 # 5ccc <_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
@@ -12554,7 +13813,7 @@ _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 196,98,125,24,5,185,79,0,0 // vbroadcastss 0x4fb9(%rip),%ymm8 # 55dc <_sk_callback_avx+0x16d>
+ .byte 196,98,125,24,5,173,86,0,0 // vbroadcastss 0x56ad(%rip),%ymm8 # 5cd0 <_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
@@ -12565,7 +13824,7 @@ HIDDEN _sk_colorburn_avx
.globl _sk_colorburn_avx
FUNCTION(_sk_colorburn_avx)
_sk_colorburn_avx:
- .byte 196,98,125,24,5,164,79,0,0 // vbroadcastss 0x4fa4(%rip),%ymm8 # 55e0 <_sk_callback_avx+0x171>
+ .byte 196,98,125,24,5,152,86,0,0 // vbroadcastss 0x5698(%rip),%ymm8 # 5cd4 <_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
@@ -12627,7 +13886,7 @@ HIDDEN _sk_colordodge_avx
FUNCTION(_sk_colordodge_avx)
_sk_colordodge_avx:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
- .byte 196,98,125,24,13,160,78,0,0 // vbroadcastss 0x4ea0(%rip),%ymm9 # 55e4 <_sk_callback_avx+0x175>
+ .byte 196,98,125,24,13,148,85,0,0 // vbroadcastss 0x5594(%rip),%ymm9 # 5cd8 <_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
@@ -12684,7 +13943,7 @@ HIDDEN _sk_hardlight_avx
.globl _sk_hardlight_avx
FUNCTION(_sk_hardlight_avx)
_sk_hardlight_avx:
- .byte 196,98,125,24,5,178,77,0,0 // vbroadcastss 0x4db2(%rip),%ymm8 # 55e8 <_sk_callback_avx+0x179>
+ .byte 196,98,125,24,5,166,84,0,0 // vbroadcastss 0x54a6(%rip),%ymm8 # 5cdc <_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
@@ -12739,7 +13998,7 @@ HIDDEN _sk_overlay_avx
.globl _sk_overlay_avx
FUNCTION(_sk_overlay_avx)
_sk_overlay_avx:
- .byte 196,98,125,24,5,219,76,0,0 // vbroadcastss 0x4cdb(%rip),%ymm8 # 55ec <_sk_callback_avx+0x17d>
+ .byte 196,98,125,24,5,207,83,0,0 // vbroadcastss 0x53cf(%rip),%ymm8 # 5ce0 <_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
@@ -12805,10 +14064,10 @@ _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 196,98,125,24,29,210,75,0,0 // vbroadcastss 0x4bd2(%rip),%ymm11 # 55f4 <_sk_callback_avx+0x185>
+ .byte 196,98,125,24,29,198,82,0,0 // vbroadcastss 0x52c6(%rip),%ymm11 # 5ce8 <_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 196,98,125,24,45,195,75,0,0 // vbroadcastss 0x4bc3(%rip),%ymm13 # 55f8 <_sk_callback_avx+0x189>
+ .byte 196,98,125,24,45,183,82,0,0 // vbroadcastss 0x52b7(%rip),%ymm13 # 5cec <_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
@@ -12819,7 +14078,7 @@ _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 196,98,125,24,5,129,75,0,0 // vbroadcastss 0x4b81(%rip),%ymm8 # 55f0 <_sk_callback_avx+0x181>
+ .byte 196,98,125,24,5,117,82,0,0 // vbroadcastss 0x5275(%rip),%ymm8 # 5ce4 <_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
@@ -12910,6 +14169,392 @@ _sk_softlight_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_hue_avx
+.globl _sk_hue_avx
+FUNCTION(_sk_hue_avx)
+_sk_hue_avx:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,100,194,208,0 // vcmpeqps %ymm8,%ymm3,%ymm10
+ .byte 196,98,125,24,13,215,80,0,0 // vbroadcastss 0x50d7(%rip),%ymm9 # 5cf0 <_sk_callback_avx+0x18d>
+ .byte 197,52,94,219 // vdivps %ymm3,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
+ .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
+ .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
+ .byte 197,84,95,214 // vmaxps %ymm6,%ymm5,%ymm10
+ .byte 196,65,92,95,210 // vmaxps %ymm10,%ymm4,%ymm10
+ .byte 197,84,93,222 // vminps %ymm6,%ymm5,%ymm11
+ .byte 196,65,92,93,219 // vminps %ymm11,%ymm4,%ymm11
+ .byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10
+ .byte 197,116,93,218 // vminps %ymm2,%ymm1,%ymm11
+ .byte 196,65,124,93,219 // vminps %ymm11,%ymm0,%ymm11
+ .byte 197,116,95,226 // vmaxps %ymm2,%ymm1,%ymm12
+ .byte 196,65,124,95,228 // vmaxps %ymm12,%ymm0,%ymm12
+ .byte 196,65,28,92,227 // vsubps %ymm11,%ymm12,%ymm12
+ .byte 196,65,28,194,232,0 // vcmpeqps %ymm8,%ymm12,%ymm13
+ .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0
+ .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
+ .byte 196,193,124,94,196 // vdivps %ymm12,%ymm0,%ymm0
+ .byte 196,195,125,74,192,208 // vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1
+ .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
+ .byte 196,193,116,94,204 // vdivps %ymm12,%ymm1,%ymm1
+ .byte 196,195,117,74,200,208 // vblendvps %ymm13,%ymm8,%ymm1,%ymm1
+ .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2
+ .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
+ .byte 196,193,108,94,212 // vdivps %ymm12,%ymm2,%ymm2
+ .byte 196,195,109,74,208,208 // vblendvps %ymm13,%ymm8,%ymm2,%ymm2
+ .byte 196,98,125,24,21,76,80,0,0 // vbroadcastss 0x504c(%rip),%ymm10 # 5cf4 <_sk_callback_avx+0x191>
+ .byte 196,65,92,89,218 // vmulps %ymm10,%ymm4,%ymm11
+ .byte 196,98,125,24,37,66,80,0,0 // vbroadcastss 0x5042(%rip),%ymm12 # 5cf8 <_sk_callback_avx+0x195>
+ .byte 196,65,84,89,236 // vmulps %ymm12,%ymm5,%ymm13
+ .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11
+ .byte 196,98,125,24,45,51,80,0,0 // vbroadcastss 0x5033(%rip),%ymm13 # 5cfc <_sk_callback_avx+0x199>
+ .byte 196,65,76,89,245 // vmulps %ymm13,%ymm6,%ymm14
+ .byte 196,65,36,88,222 // vaddps %ymm14,%ymm11,%ymm11
+ .byte 196,65,124,89,242 // vmulps %ymm10,%ymm0,%ymm14
+ .byte 196,65,116,89,252 // vmulps %ymm12,%ymm1,%ymm15
+ .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14
+ .byte 196,65,108,89,253 // vmulps %ymm13,%ymm2,%ymm15
+ .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14
+ .byte 196,65,36,92,222 // vsubps %ymm14,%ymm11,%ymm11
+ .byte 196,65,124,88,243 // vaddps %ymm11,%ymm0,%ymm14
+ .byte 196,65,116,88,251 // vaddps %ymm11,%ymm1,%ymm15
+ .byte 196,193,108,88,195 // vaddps %ymm11,%ymm2,%ymm0
+ .byte 197,132,93,200 // vminps %ymm0,%ymm15,%ymm1
+ .byte 197,140,93,209 // vminps %ymm1,%ymm14,%ymm2
+ .byte 197,132,95,200 // vmaxps %ymm0,%ymm15,%ymm1
+ .byte 197,12,95,217 // vmaxps %ymm1,%ymm14,%ymm11
+ .byte 196,193,12,89,202 // vmulps %ymm10,%ymm14,%ymm1
+ .byte 196,65,4,89,212 // vmulps %ymm12,%ymm15,%ymm10
+ .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
+ .byte 196,65,124,89,213 // vmulps %ymm13,%ymm0,%ymm10
+ .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1
+ .byte 197,60,194,210,2 // vcmpleps %ymm2,%ymm8,%ymm10
+ .byte 197,12,92,225 // vsubps %ymm1,%ymm14,%ymm12
+ .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12
+ .byte 197,244,92,210 // vsubps %ymm2,%ymm1,%ymm2
+ .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12
+ .byte 196,65,116,88,228 // vaddps %ymm12,%ymm1,%ymm12
+ .byte 196,67,29,74,230,160 // vblendvps %ymm10,%ymm14,%ymm12,%ymm12
+ .byte 196,65,52,194,235,1 // vcmpltps %ymm11,%ymm9,%ymm13
+ .byte 197,28,92,241 // vsubps %ymm1,%ymm12,%ymm14
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,246 // vmulps %ymm14,%ymm9,%ymm14
+ .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11
+ .byte 196,65,12,94,243 // vdivps %ymm11,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,29,74,230,208 // vblendvps %ymm13,%ymm14,%ymm12,%ymm12
+ .byte 196,65,28,95,224 // vmaxps %ymm8,%ymm12,%ymm12
+ .byte 197,4,92,241 // vsubps %ymm1,%ymm15,%ymm14
+ .byte 196,65,116,89,246 // vmulps %ymm14,%ymm1,%ymm14
+ .byte 197,12,94,242 // vdivps %ymm2,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,13,74,247,160 // vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ .byte 197,12,92,249 // vsubps %ymm1,%ymm14,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,251 // vdivps %ymm11,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,13,74,247,208 // vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ .byte 196,65,12,95,240 // vmaxps %ymm8,%ymm14,%ymm14
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 197,132,94,210 // vdivps %ymm2,%ymm15,%ymm2
+ .byte 197,244,88,210 // vaddps %ymm2,%ymm1,%ymm2
+ .byte 196,227,109,74,192,160 // vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
+ .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
+ .byte 196,193,108,94,211 // vdivps %ymm11,%ymm2,%ymm2
+ .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
+ .byte 196,227,125,74,193,208 // vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ .byte 196,193,124,95,208 // vmaxps %ymm8,%ymm0,%ymm2
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 197,228,89,207 // vmulps %ymm7,%ymm3,%ymm1
+ .byte 197,252,92,217 // vsubps %ymm1,%ymm0,%ymm3
+ .byte 196,193,100,89,196 // vmulps %ymm12,%ymm3,%ymm0
+ .byte 196,193,100,89,206 // vmulps %ymm14,%ymm3,%ymm1
+ .byte 197,228,89,210 // vmulps %ymm2,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_saturation_avx
+.globl _sk_saturation_avx
+FUNCTION(_sk_saturation_avx)
+_sk_saturation_avx:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,68,194,208,0 // vcmpeqps %ymm8,%ymm7,%ymm10
+ .byte 196,98,125,24,13,240,78,0,0 // vbroadcastss 0x4ef0(%rip),%ymm9 # 5d00 <_sk_callback_avx+0x19d>
+ .byte 197,52,94,223 // vdivps %ymm7,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,44,89,220 // vmulps %ymm4,%ymm10,%ymm11
+ .byte 197,44,89,229 // vmulps %ymm5,%ymm10,%ymm12
+ .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10
+ .byte 197,116,95,234 // vmaxps %ymm2,%ymm1,%ymm13
+ .byte 196,65,124,95,237 // vmaxps %ymm13,%ymm0,%ymm13
+ .byte 197,244,93,202 // vminps %ymm2,%ymm1,%ymm1
+ .byte 197,252,93,193 // vminps %ymm1,%ymm0,%ymm0
+ .byte 197,148,92,192 // vsubps %ymm0,%ymm13,%ymm0
+ .byte 196,193,28,93,202 // vminps %ymm10,%ymm12,%ymm1
+ .byte 197,164,93,201 // vminps %ymm1,%ymm11,%ymm1
+ .byte 196,193,28,95,210 // vmaxps %ymm10,%ymm12,%ymm2
+ .byte 197,164,95,210 // vmaxps %ymm2,%ymm11,%ymm2
+ .byte 197,236,92,209 // vsubps %ymm1,%ymm2,%ymm2
+ .byte 196,65,108,194,232,0 // vcmpeqps %ymm8,%ymm2,%ymm13
+ .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11
+ .byte 196,65,124,89,219 // vmulps %ymm11,%ymm0,%ymm11
+ .byte 197,36,94,218 // vdivps %ymm2,%ymm11,%ymm11
+ .byte 196,67,37,74,216,208 // vblendvps %ymm13,%ymm8,%ymm11,%ymm11
+ .byte 197,28,92,225 // vsubps %ymm1,%ymm12,%ymm12
+ .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12
+ .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12
+ .byte 196,67,29,74,224,208 // vblendvps %ymm13,%ymm8,%ymm12,%ymm12
+ .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1
+ .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
+ .byte 197,252,94,194 // vdivps %ymm2,%ymm0,%ymm0
+ .byte 196,195,125,74,192,208 // vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ .byte 196,226,125,24,13,108,78,0,0 // vbroadcastss 0x4e6c(%rip),%ymm1 # 5d04 <_sk_callback_avx+0x1a1>
+ .byte 197,220,89,209 // vmulps %ymm1,%ymm4,%ymm2
+ .byte 196,98,125,24,21,99,78,0,0 // vbroadcastss 0x4e63(%rip),%ymm10 # 5d08 <_sk_callback_avx+0x1a5>
+ .byte 196,65,84,89,234 // vmulps %ymm10,%ymm5,%ymm13
+ .byte 196,193,108,88,213 // vaddps %ymm13,%ymm2,%ymm2
+ .byte 196,98,125,24,45,84,78,0,0 // vbroadcastss 0x4e54(%rip),%ymm13 # 5d0c <_sk_callback_avx+0x1a9>
+ .byte 196,65,76,89,245 // vmulps %ymm13,%ymm6,%ymm14
+ .byte 196,193,108,88,214 // vaddps %ymm14,%ymm2,%ymm2
+ .byte 197,36,89,241 // vmulps %ymm1,%ymm11,%ymm14
+ .byte 196,65,28,89,250 // vmulps %ymm10,%ymm12,%ymm15
+ .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14
+ .byte 196,65,124,89,253 // vmulps %ymm13,%ymm0,%ymm15
+ .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14
+ .byte 196,193,108,92,214 // vsubps %ymm14,%ymm2,%ymm2
+ .byte 197,36,88,218 // vaddps %ymm2,%ymm11,%ymm11
+ .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12
+ .byte 197,252,88,194 // vaddps %ymm2,%ymm0,%ymm0
+ .byte 197,156,93,208 // vminps %ymm0,%ymm12,%ymm2
+ .byte 197,164,93,210 // vminps %ymm2,%ymm11,%ymm2
+ .byte 197,28,95,240 // vmaxps %ymm0,%ymm12,%ymm14
+ .byte 196,65,36,95,246 // vmaxps %ymm14,%ymm11,%ymm14
+ .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1
+ .byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10
+ .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
+ .byte 196,65,124,89,213 // vmulps %ymm13,%ymm0,%ymm10
+ .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1
+ .byte 197,60,194,210,2 // vcmpleps %ymm2,%ymm8,%ymm10
+ .byte 197,36,92,233 // vsubps %ymm1,%ymm11,%ymm13
+ .byte 196,65,116,89,237 // vmulps %ymm13,%ymm1,%ymm13
+ .byte 197,244,92,210 // vsubps %ymm2,%ymm1,%ymm2
+ .byte 197,20,94,234 // vdivps %ymm2,%ymm13,%ymm13
+ .byte 196,65,116,88,237 // vaddps %ymm13,%ymm1,%ymm13
+ .byte 196,67,21,74,219,160 // vblendvps %ymm10,%ymm11,%ymm13,%ymm11
+ .byte 196,65,52,194,238,1 // vcmpltps %ymm14,%ymm9,%ymm13
+ .byte 197,36,92,249 // vsubps %ymm1,%ymm11,%ymm15
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 197,12,92,241 // vsubps %ymm1,%ymm14,%ymm14
+ .byte 196,65,4,94,254 // vdivps %ymm14,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,37,74,223,208 // vblendvps %ymm13,%ymm15,%ymm11,%ymm11
+ .byte 196,65,36,95,216 // vmaxps %ymm8,%ymm11,%ymm11
+ .byte 197,28,92,249 // vsubps %ymm1,%ymm12,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 197,4,94,250 // vdivps %ymm2,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,5,74,228,160 // vblendvps %ymm10,%ymm12,%ymm15,%ymm12
+ .byte 197,28,92,249 // vsubps %ymm1,%ymm12,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,254 // vdivps %ymm14,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,29,74,231,208 // vblendvps %ymm13,%ymm15,%ymm12,%ymm12
+ .byte 196,65,28,95,224 // vmaxps %ymm8,%ymm12,%ymm12
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 197,132,94,210 // vdivps %ymm2,%ymm15,%ymm2
+ .byte 197,244,88,210 // vaddps %ymm2,%ymm1,%ymm2
+ .byte 196,227,109,74,192,160 // vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
+ .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
+ .byte 196,193,108,94,214 // vdivps %ymm14,%ymm2,%ymm2
+ .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
+ .byte 196,227,125,74,193,208 // vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ .byte 196,193,124,95,208 // vmaxps %ymm8,%ymm0,%ymm2
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 197,228,89,207 // vmulps %ymm7,%ymm3,%ymm1
+ .byte 197,252,92,217 // vsubps %ymm1,%ymm0,%ymm3
+ .byte 196,193,100,89,195 // vmulps %ymm11,%ymm3,%ymm0
+ .byte 196,193,100,89,204 // vmulps %ymm12,%ymm3,%ymm1
+ .byte 197,228,89,210 // vmulps %ymm2,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_color_avx
+.globl _sk_color_avx
+FUNCTION(_sk_color_avx)
+_sk_color_avx:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,100,194,208,0 // vcmpeqps %ymm8,%ymm3,%ymm10
+ .byte 196,98,125,24,13,21,77,0,0 // vbroadcastss 0x4d15(%rip),%ymm9 # 5d10 <_sk_callback_avx+0x1ad>
+ .byte 197,52,94,219 // vdivps %ymm3,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
+ .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
+ .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
+ .byte 196,98,125,24,21,250,76,0,0 // vbroadcastss 0x4cfa(%rip),%ymm10 # 5d14 <_sk_callback_avx+0x1b1>
+ .byte 196,65,92,89,218 // vmulps %ymm10,%ymm4,%ymm11
+ .byte 196,98,125,24,37,240,76,0,0 // vbroadcastss 0x4cf0(%rip),%ymm12 # 5d18 <_sk_callback_avx+0x1b5>
+ .byte 196,65,84,89,236 // vmulps %ymm12,%ymm5,%ymm13
+ .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11
+ .byte 196,98,125,24,45,225,76,0,0 // vbroadcastss 0x4ce1(%rip),%ymm13 # 5d1c <_sk_callback_avx+0x1b9>
+ .byte 196,65,76,89,245 // vmulps %ymm13,%ymm6,%ymm14
+ .byte 196,65,36,88,222 // vaddps %ymm14,%ymm11,%ymm11
+ .byte 196,65,124,89,242 // vmulps %ymm10,%ymm0,%ymm14
+ .byte 196,65,116,89,252 // vmulps %ymm12,%ymm1,%ymm15
+ .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14
+ .byte 196,65,108,89,253 // vmulps %ymm13,%ymm2,%ymm15
+ .byte 196,65,4,88,246 // vaddps %ymm14,%ymm15,%ymm14
+ .byte 196,65,36,92,222 // vsubps %ymm14,%ymm11,%ymm11
+ .byte 196,65,124,88,243 // vaddps %ymm11,%ymm0,%ymm14
+ .byte 196,65,116,88,251 // vaddps %ymm11,%ymm1,%ymm15
+ .byte 196,193,108,88,195 // vaddps %ymm11,%ymm2,%ymm0
+ .byte 197,132,93,200 // vminps %ymm0,%ymm15,%ymm1
+ .byte 197,140,93,209 // vminps %ymm1,%ymm14,%ymm2
+ .byte 197,132,95,200 // vmaxps %ymm0,%ymm15,%ymm1
+ .byte 197,12,95,217 // vmaxps %ymm1,%ymm14,%ymm11
+ .byte 196,193,12,89,202 // vmulps %ymm10,%ymm14,%ymm1
+ .byte 196,65,4,89,212 // vmulps %ymm12,%ymm15,%ymm10
+ .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
+ .byte 196,65,124,89,213 // vmulps %ymm13,%ymm0,%ymm10
+ .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1
+ .byte 197,60,194,210,2 // vcmpleps %ymm2,%ymm8,%ymm10
+ .byte 197,12,92,225 // vsubps %ymm1,%ymm14,%ymm12
+ .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12
+ .byte 197,244,92,210 // vsubps %ymm2,%ymm1,%ymm2
+ .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12
+ .byte 196,65,116,88,228 // vaddps %ymm12,%ymm1,%ymm12
+ .byte 196,67,29,74,230,160 // vblendvps %ymm10,%ymm14,%ymm12,%ymm12
+ .byte 196,65,52,194,235,1 // vcmpltps %ymm11,%ymm9,%ymm13
+ .byte 197,28,92,241 // vsubps %ymm1,%ymm12,%ymm14
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,246 // vmulps %ymm14,%ymm9,%ymm14
+ .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11
+ .byte 196,65,12,94,243 // vdivps %ymm11,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,29,74,230,208 // vblendvps %ymm13,%ymm14,%ymm12,%ymm12
+ .byte 196,65,28,95,224 // vmaxps %ymm8,%ymm12,%ymm12
+ .byte 197,4,92,241 // vsubps %ymm1,%ymm15,%ymm14
+ .byte 196,65,116,89,246 // vmulps %ymm14,%ymm1,%ymm14
+ .byte 197,12,94,242 // vdivps %ymm2,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,67,13,74,247,160 // vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ .byte 197,12,92,249 // vsubps %ymm1,%ymm14,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,251 // vdivps %ymm11,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,13,74,247,208 // vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ .byte 196,65,12,95,240 // vmaxps %ymm8,%ymm14,%ymm14
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 197,132,94,210 // vdivps %ymm2,%ymm15,%ymm2
+ .byte 197,244,88,210 // vaddps %ymm2,%ymm1,%ymm2
+ .byte 196,227,109,74,192,160 // vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
+ .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
+ .byte 196,193,108,94,211 // vdivps %ymm11,%ymm2,%ymm2
+ .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
+ .byte 196,227,125,74,193,208 // vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ .byte 196,193,124,95,208 // vmaxps %ymm8,%ymm0,%ymm2
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 197,228,89,207 // vmulps %ymm7,%ymm3,%ymm1
+ .byte 197,252,92,217 // vsubps %ymm1,%ymm0,%ymm3
+ .byte 196,193,100,89,196 // vmulps %ymm12,%ymm3,%ymm0
+ .byte 196,193,100,89,206 // vmulps %ymm14,%ymm3,%ymm1
+ .byte 197,228,89,210 // vmulps %ymm2,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_luminosity_avx
+.globl _sk_luminosity_avx
+FUNCTION(_sk_luminosity_avx)
+_sk_luminosity_avx:
+ .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
+ .byte 196,65,68,194,208,0 // vcmpeqps %ymm8,%ymm7,%ymm10
+ .byte 196,98,125,24,13,158,75,0,0 // vbroadcastss 0x4b9e(%rip),%ymm9 # 5d20 <_sk_callback_avx+0x1bd>
+ .byte 197,52,94,223 // vdivps %ymm7,%ymm9,%ymm11
+ .byte 196,67,37,74,208,160 // vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ .byte 197,44,89,220 // vmulps %ymm4,%ymm10,%ymm11
+ .byte 197,44,89,229 // vmulps %ymm5,%ymm10,%ymm12
+ .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10
+ .byte 196,98,125,24,45,131,75,0,0 // vbroadcastss 0x4b83(%rip),%ymm13 # 5d24 <_sk_callback_avx+0x1c1>
+ .byte 196,193,124,89,197 // vmulps %ymm13,%ymm0,%ymm0
+ .byte 196,98,125,24,53,121,75,0,0 // vbroadcastss 0x4b79(%rip),%ymm14 # 5d28 <_sk_callback_avx+0x1c5>
+ .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1
+ .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
+ .byte 196,226,125,24,13,107,75,0,0 // vbroadcastss 0x4b6b(%rip),%ymm1 # 5d2c <_sk_callback_avx+0x1c9>
+ .byte 197,236,89,209 // vmulps %ymm1,%ymm2,%ymm2
+ .byte 197,252,88,194 // vaddps %ymm2,%ymm0,%ymm0
+ .byte 196,193,36,89,213 // vmulps %ymm13,%ymm11,%ymm2
+ .byte 196,65,28,89,254 // vmulps %ymm14,%ymm12,%ymm15
+ .byte 196,193,108,88,215 // vaddps %ymm15,%ymm2,%ymm2
+ .byte 197,44,89,249 // vmulps %ymm1,%ymm10,%ymm15
+ .byte 197,132,88,210 // vaddps %ymm2,%ymm15,%ymm2
+ .byte 197,252,92,194 // vsubps %ymm2,%ymm0,%ymm0
+ .byte 197,164,88,208 // vaddps %ymm0,%ymm11,%ymm2
+ .byte 197,28,88,216 // vaddps %ymm0,%ymm12,%ymm11
+ .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0
+ .byte 197,36,93,208 // vminps %ymm0,%ymm11,%ymm10
+ .byte 196,65,108,93,210 // vminps %ymm10,%ymm2,%ymm10
+ .byte 197,36,95,224 // vmaxps %ymm0,%ymm11,%ymm12
+ .byte 196,65,108,95,228 // vmaxps %ymm12,%ymm2,%ymm12
+ .byte 196,65,108,89,237 // vmulps %ymm13,%ymm2,%ymm13
+ .byte 196,65,36,89,246 // vmulps %ymm14,%ymm11,%ymm14
+ .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
+ .byte 197,252,89,201 // vmulps %ymm1,%ymm0,%ymm1
+ .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1
+ .byte 196,65,60,194,234,2 // vcmpleps %ymm10,%ymm8,%ymm13
+ .byte 197,108,92,241 // vsubps %ymm1,%ymm2,%ymm14
+ .byte 196,65,116,89,246 // vmulps %ymm14,%ymm1,%ymm14
+ .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
+ .byte 196,65,12,94,242 // vdivps %ymm10,%ymm14,%ymm14
+ .byte 196,65,116,88,246 // vaddps %ymm14,%ymm1,%ymm14
+ .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2
+ .byte 196,65,52,194,244,1 // vcmpltps %ymm12,%ymm9,%ymm14
+ .byte 197,108,92,249 // vsubps %ymm1,%ymm2,%ymm15
+ .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 197,28,92,225 // vsubps %ymm1,%ymm12,%ymm12
+ .byte 196,65,4,94,252 // vdivps %ymm12,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,195,109,74,215,224 // vblendvps %ymm14,%ymm15,%ymm2,%ymm2
+ .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
+ .byte 197,36,92,249 // vsubps %ymm1,%ymm11,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 196,65,4,94,250 // vdivps %ymm10,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,5,74,219,208 // vblendvps %ymm13,%ymm11,%ymm15,%ymm11
+ .byte 197,36,92,249 // vsubps %ymm1,%ymm11,%ymm15
+ .byte 196,65,52,89,255 // vmulps %ymm15,%ymm9,%ymm15
+ .byte 196,65,4,94,252 // vdivps %ymm12,%ymm15,%ymm15
+ .byte 196,65,116,88,255 // vaddps %ymm15,%ymm1,%ymm15
+ .byte 196,67,37,74,223,224 // vblendvps %ymm14,%ymm15,%ymm11,%ymm11
+ .byte 196,65,36,95,216 // vmaxps %ymm8,%ymm11,%ymm11
+ .byte 197,124,92,249 // vsubps %ymm1,%ymm0,%ymm15
+ .byte 196,65,116,89,255 // vmulps %ymm15,%ymm1,%ymm15
+ .byte 196,65,4,94,210 // vdivps %ymm10,%ymm15,%ymm10
+ .byte 196,65,116,88,210 // vaddps %ymm10,%ymm1,%ymm10
+ .byte 196,227,45,74,192,208 // vblendvps %ymm13,%ymm0,%ymm10,%ymm0
+ .byte 197,124,92,209 // vsubps %ymm1,%ymm0,%ymm10
+ .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
+ .byte 196,65,52,94,204 // vdivps %ymm12,%ymm9,%ymm9
+ .byte 196,193,116,88,201 // vaddps %ymm9,%ymm1,%ymm1
+ .byte 196,227,125,74,193,224 // vblendvps %ymm14,%ymm1,%ymm0,%ymm0
+ .byte 196,65,124,95,192 // vmaxps %ymm8,%ymm0,%ymm8
+ .byte 197,228,88,199 // vaddps %ymm7,%ymm3,%ymm0
+ .byte 197,228,89,207 // vmulps %ymm7,%ymm3,%ymm1
+ .byte 197,252,92,217 // vsubps %ymm1,%ymm0,%ymm3
+ .byte 197,228,89,194 // vmulps %ymm2,%ymm3,%ymm0
+ .byte 196,193,100,89,203 // vmulps %ymm11,%ymm3,%ymm1
+ .byte 196,193,100,89,208 // vmulps %ymm8,%ymm3,%ymm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_clamp_0_avx
.globl _sk_clamp_0_avx
FUNCTION(_sk_clamp_0_avx)
@@ -12926,7 +14571,7 @@ HIDDEN _sk_clamp_1_avx
.globl _sk_clamp_1_avx
FUNCTION(_sk_clamp_1_avx)
_sk_clamp_1_avx:
- .byte 196,98,125,24,5,209,73,0,0 // vbroadcastss 0x49d1(%rip),%ymm8 # 55fc <_sk_callback_avx+0x18d>
+ .byte 196,98,125,24,5,20,74,0,0 // vbroadcastss 0x4a14(%rip),%ymm8 # 5d30 <_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
@@ -12938,7 +14583,7 @@ HIDDEN _sk_clamp_a_avx
.globl _sk_clamp_a_avx
FUNCTION(_sk_clamp_a_avx)
_sk_clamp_a_avx:
- .byte 196,98,125,24,5,180,73,0,0 // vbroadcastss 0x49b4(%rip),%ymm8 # 5600 <_sk_callback_avx+0x191>
+ .byte 196,98,125,24,5,247,73,0,0 // vbroadcastss 0x49f7(%rip),%ymm8 # 5d34 <_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
@@ -13024,7 +14669,7 @@ 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 196,98,125,24,21,252,72,0,0 // vbroadcastss 0x48fc(%rip),%ymm10 # 5604 <_sk_callback_avx+0x195>
+ .byte 196,98,125,24,21,63,73,0,0 // vbroadcastss 0x493f(%rip),%ymm10 # 5d38 <_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
@@ -13037,17 +14682,17 @@ HIDDEN _sk_from_srgb_avx
.globl _sk_from_srgb_avx
FUNCTION(_sk_from_srgb_avx)
_sk_from_srgb_avx:
- .byte 196,98,125,24,5,221,72,0,0 // vbroadcastss 0x48dd(%rip),%ymm8 # 5608 <_sk_callback_avx+0x199>
+ .byte 196,98,125,24,5,32,73,0,0 // vbroadcastss 0x4920(%rip),%ymm8 # 5d3c <_sk_callback_avx+0x1d9>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
- .byte 196,98,125,24,29,207,72,0,0 // vbroadcastss 0x48cf(%rip),%ymm11 # 560c <_sk_callback_avx+0x19d>
+ .byte 196,98,125,24,29,18,73,0,0 // vbroadcastss 0x4912(%rip),%ymm11 # 5d40 <_sk_callback_avx+0x1dd>
.byte 196,65,124,89,227 // vmulps %ymm11,%ymm0,%ymm12
- .byte 196,98,125,24,45,197,72,0,0 // vbroadcastss 0x48c5(%rip),%ymm13 # 5610 <_sk_callback_avx+0x1a1>
+ .byte 196,98,125,24,45,8,73,0,0 // vbroadcastss 0x4908(%rip),%ymm13 # 5d44 <_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 196,98,125,24,37,182,72,0,0 // vbroadcastss 0x48b6(%rip),%ymm12 # 5614 <_sk_callback_avx+0x1a5>
+ .byte 196,98,125,24,37,249,72,0,0 // vbroadcastss 0x48f9(%rip),%ymm12 # 5d48 <_sk_callback_avx+0x1e5>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,53,172,72,0,0 // vbroadcastss 0x48ac(%rip),%ymm14 # 5618 <_sk_callback_avx+0x1a9>
+ .byte 196,98,125,24,53,239,72,0,0 // vbroadcastss 0x48ef(%rip),%ymm14 # 5d4c <_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
@@ -13076,18 +14721,18 @@ _sk_to_srgb_avx:
.byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8
.byte 196,65,124,83,200 // vrcpps %ymm8,%ymm9
.byte 196,65,124,82,208 // vrsqrtps %ymm8,%ymm10
- .byte 196,98,125,24,5,55,72,0,0 // vbroadcastss 0x4837(%rip),%ymm8 # 561c <_sk_callback_avx+0x1ad>
+ .byte 196,98,125,24,5,122,72,0,0 // vbroadcastss 0x487a(%rip),%ymm8 # 5d50 <_sk_callback_avx+0x1ed>
.byte 196,65,124,89,216 // vmulps %ymm8,%ymm0,%ymm11
- .byte 196,98,125,24,37,45,72,0,0 // vbroadcastss 0x482d(%rip),%ymm12 # 5620 <_sk_callback_avx+0x1b1>
+ .byte 196,98,125,24,37,112,72,0,0 // vbroadcastss 0x4870(%rip),%ymm12 # 5d54 <_sk_callback_avx+0x1f1>
.byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9
- .byte 196,98,125,24,45,35,72,0,0 // vbroadcastss 0x4823(%rip),%ymm13 # 5624 <_sk_callback_avx+0x1b5>
+ .byte 196,98,125,24,45,102,72,0,0 // vbroadcastss 0x4866(%rip),%ymm13 # 5d58 <_sk_callback_avx+0x1f5>
.byte 196,65,52,88,205 // vaddps %ymm13,%ymm9,%ymm9
- .byte 196,98,125,24,53,25,72,0,0 // vbroadcastss 0x4819(%rip),%ymm14 # 5628 <_sk_callback_avx+0x1b9>
+ .byte 196,98,125,24,53,92,72,0,0 // vbroadcastss 0x485c(%rip),%ymm14 # 5d5c <_sk_callback_avx+0x1f9>
.byte 196,65,44,89,214 // vmulps %ymm14,%ymm10,%ymm10
.byte 196,65,44,88,201 // vaddps %ymm9,%ymm10,%ymm9
- .byte 196,98,125,24,21,10,72,0,0 // vbroadcastss 0x480a(%rip),%ymm10 # 562c <_sk_callback_avx+0x1bd>
+ .byte 196,98,125,24,21,77,72,0,0 // vbroadcastss 0x484d(%rip),%ymm10 # 5d60 <_sk_callback_avx+0x1fd>
.byte 196,65,44,93,201 // vminps %ymm9,%ymm10,%ymm9
- .byte 196,98,125,24,61,0,72,0,0 // vbroadcastss 0x4800(%rip),%ymm15 # 5630 <_sk_callback_avx+0x1c1>
+ .byte 196,98,125,24,61,67,72,0,0 // vbroadcastss 0x4843(%rip),%ymm15 # 5d64 <_sk_callback_avx+0x201>
.byte 196,193,124,194,199,1 // vcmpltps %ymm15,%ymm0,%ymm0
.byte 196,195,53,74,195,0 // vblendvps %ymm0,%ymm11,%ymm9,%ymm0
.byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9
@@ -13124,7 +14769,7 @@ _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 196,98,125,24,29,102,71,0,0 // vbroadcastss 0x4766(%rip),%ymm11 # 5634 <_sk_callback_avx+0x1c5>
+ .byte 196,98,125,24,29,169,71,0,0 // vbroadcastss 0x47a9(%rip),%ymm11 # 5d68 <_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
@@ -13134,19 +14779,19 @@ _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 196,98,125,24,29,63,71,0,0 // vbroadcastss 0x473f(%rip),%ymm11 # 5640 <_sk_callback_avx+0x1d1>
+ .byte 196,98,125,24,29,130,71,0,0 // vbroadcastss 0x4782(%rip),%ymm11 # 5d74 <_sk_callback_avx+0x211>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,45,71,0,0 // vbroadcastss 0x472d(%rip),%ymm11 # 563c <_sk_callback_avx+0x1cd>
+ .byte 196,98,125,24,29,112,71,0,0 // vbroadcastss 0x4770(%rip),%ymm11 # 5d70 <_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 196,226,125,24,21,21,71,0,0 // vbroadcastss 0x4715(%rip),%ymm2 # 5638 <_sk_callback_avx+0x1c9>
+ .byte 196,226,125,24,21,88,71,0,0 // vbroadcastss 0x4758(%rip),%ymm2 # 5d6c <_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 196,98,125,24,37,252,70,0,0 // vbroadcastss 0x46fc(%rip),%ymm12 # 5648 <_sk_callback_avx+0x1d9>
+ .byte 196,98,125,24,37,63,71,0,0 // vbroadcastss 0x473f(%rip),%ymm12 # 5d7c <_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
@@ -13156,7 +14801,7 @@ _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 196,98,125,24,5,191,70,0,0 // vbroadcastss 0x46bf(%rip),%ymm8 # 5644 <_sk_callback_avx+0x1d5>
+ .byte 196,98,125,24,5,2,71,0,0 // vbroadcastss 0x4702(%rip),%ymm8 # 5d78 <_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
@@ -13173,7 +14818,7 @@ _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 196,98,125,24,5,140,70,0,0 // vbroadcastss 0x468c(%rip),%ymm8 # 564c <_sk_callback_avx+0x1dd>
+ .byte 196,98,125,24,5,207,70,0,0 // vbroadcastss 0x46cf(%rip),%ymm8 # 5d80 <_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
@@ -13181,23 +14826,23 @@ _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 196,98,125,24,29,102,70,0,0 // vbroadcastss 0x4666(%rip),%ymm11 # 5650 <_sk_callback_avx+0x1e1>
+ .byte 196,98,125,24,29,169,70,0,0 // vbroadcastss 0x46a9(%rip),%ymm11 # 5d84 <_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 196,98,125,24,37,80,70,0,0 // vbroadcastss 0x4650(%rip),%ymm12 # 5658 <_sk_callback_avx+0x1e9>
+ .byte 196,98,125,24,37,147,70,0,0 // vbroadcastss 0x4693(%rip),%ymm12 # 5d8c <_sk_callback_avx+0x229>
.byte 196,193,4,89,196 // vmulps %ymm12,%ymm15,%ymm0
- .byte 196,98,125,24,45,70,70,0,0 // vbroadcastss 0x4646(%rip),%ymm13 # 565c <_sk_callback_avx+0x1ed>
+ .byte 196,98,125,24,45,137,70,0,0 // vbroadcastss 0x4689(%rip),%ymm13 # 5d90 <_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 196,226,125,24,13,39,70,0,0 // vbroadcastss 0x4627(%rip),%ymm1 # 5654 <_sk_callback_avx+0x1e5>
+ .byte 196,226,125,24,13,106,70,0,0 // vbroadcastss 0x466a(%rip),%ymm1 # 5d88 <_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 196,98,125,24,53,18,70,0,0 // vbroadcastss 0x4612(%rip),%ymm14 # 5660 <_sk_callback_avx+0x1f1>
+ .byte 196,98,125,24,53,85,70,0,0 // vbroadcastss 0x4655(%rip),%ymm14 # 5d94 <_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
@@ -13216,7 +14861,7 @@ _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 196,226,125,24,5,180,69,0,0 // vbroadcastss 0x45b4(%rip),%ymm0 # 5664 <_sk_callback_avx+0x1f5>
+ .byte 196,226,125,24,5,247,69,0,0 // vbroadcastss 0x45f7(%rip),%ymm0 # 5d98 <_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
@@ -13268,14 +14913,14 @@ _sk_scale_u8_avx:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,68 // jne 11a7 <_sk_scale_u8_avx+0x54>
+ .byte 117,68 // jne 1898 <_sk_scale_u8_avx+0x54>
.byte 197,122,126,0 // vmovq (%rax),%xmm8
.byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9
.byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8
.byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8
.byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,221,68,0,0 // vbroadcastss 0x44dd(%rip),%ymm9 # 5668 <_sk_callback_avx+0x1f9>
+ .byte 196,98,125,24,13,32,69,0,0 // vbroadcastss 0x4520(%rip),%ymm9 # 5d9c <_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
@@ -13293,9 +14938,9 @@ _sk_scale_u8_avx:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 11af <_sk_scale_u8_avx+0x5c>
+ .byte 117,234 // jne 18a0 <_sk_scale_u8_avx+0x5c>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 235,155 // jmp 1167 <_sk_scale_u8_avx+0x14>
+ .byte 235,155 // jmp 1858 <_sk_scale_u8_avx+0x14>
HIDDEN _sk_lerp_1_float_avx
.globl _sk_lerp_1_float_avx
@@ -13327,14 +14972,14 @@ _sk_lerp_u8_avx:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,104 // jne 1283 <_sk_lerp_u8_avx+0x78>
+ .byte 117,104 // jne 1974 <_sk_lerp_u8_avx+0x78>
.byte 197,122,126,0 // vmovq (%rax),%xmm8
.byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9
.byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8
.byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8
.byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,41,68,0,0 // vbroadcastss 0x4429(%rip),%ymm9 # 566c <_sk_callback_avx+0x1fd>
+ .byte 196,98,125,24,13,108,68,0,0 // vbroadcastss 0x446c(%rip),%ymm9 # 5da0 <_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
@@ -13360,9 +15005,9 @@ _sk_lerp_u8_avx:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 128b <_sk_lerp_u8_avx+0x80>
+ .byte 117,234 // jne 197c <_sk_lerp_u8_avx+0x80>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 233,116,255,255,255 // jmpq 121f <_sk_lerp_u8_avx+0x14>
+ .byte 233,116,255,255,255 // jmpq 1910 <_sk_lerp_u8_avx+0x14>
HIDDEN _sk_lerp_565_avx
.globl _sk_lerp_565_avx
@@ -13371,26 +15016,26 @@ _sk_lerp_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,174,0,0,0 // jne 1367 <_sk_lerp_565_avx+0xbc>
+ .byte 15,133,174,0,0,0 // jne 1a58 <_sk_lerp_565_avx+0xbc>
.byte 196,65,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm8
.byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3
.byte 197,185,105,219 // vpunpckhwd %xmm3,%xmm8,%xmm3
.byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8
.byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
- .byte 196,98,125,24,5,149,67,0,0 // vbroadcastss 0x4395(%rip),%ymm8 # 5670 <_sk_callback_avx+0x201>
+ .byte 196,98,125,24,5,216,67,0,0 // vbroadcastss 0x43d8(%rip),%ymm8 # 5da4 <_sk_callback_avx+0x241>
.byte 196,65,100,84,192 // vandps %ymm8,%ymm3,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,134,67,0,0 // vbroadcastss 0x4386(%rip),%ymm9 # 5674 <_sk_callback_avx+0x205>
+ .byte 196,98,125,24,13,201,67,0,0 // vbroadcastss 0x43c9(%rip),%ymm9 # 5da8 <_sk_callback_avx+0x245>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,124,67,0,0 // vbroadcastss 0x437c(%rip),%ymm9 # 5678 <_sk_callback_avx+0x209>
+ .byte 196,98,125,24,13,191,67,0,0 // vbroadcastss 0x43bf(%rip),%ymm9 # 5dac <_sk_callback_avx+0x249>
.byte 196,65,100,84,201 // vandps %ymm9,%ymm3,%ymm9
.byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9
- .byte 196,98,125,24,21,109,67,0,0 // vbroadcastss 0x436d(%rip),%ymm10 # 567c <_sk_callback_avx+0x20d>
+ .byte 196,98,125,24,21,176,67,0,0 // vbroadcastss 0x43b0(%rip),%ymm10 # 5db0 <_sk_callback_avx+0x24d>
.byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,24,21,99,67,0,0 // vbroadcastss 0x4363(%rip),%ymm10 # 5680 <_sk_callback_avx+0x211>
+ .byte 196,98,125,24,21,166,67,0,0 // vbroadcastss 0x43a6(%rip),%ymm10 # 5db4 <_sk_callback_avx+0x251>
.byte 196,193,100,84,218 // vandps %ymm10,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,21,85,67,0,0 // vbroadcastss 0x4355(%rip),%ymm10 # 5684 <_sk_callback_avx+0x215>
+ .byte 196,98,125,24,21,152,67,0,0 // vbroadcastss 0x4398(%rip),%ymm10 # 5db8 <_sk_callback_avx+0x255>
.byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3
.byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
@@ -13402,16 +15047,16 @@ _sk_lerp_565_avx:
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,35,67,0,0 // vbroadcastss 0x4323(%rip),%ymm3 # 5688 <_sk_callback_avx+0x219>
+ .byte 196,226,125,24,29,102,67,0,0 // vbroadcastss 0x4366(%rip),%ymm3 # 5dbc <_sk_callback_avx+0x259>
.byte 255,224 // jmpq *%rax
.byte 65,137,200 // mov %ecx,%r8d
.byte 65,128,224,7 // and $0x7,%r8b
.byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,63,255,255,255 // ja 12bf <_sk_lerp_565_avx+0x14>
+ .byte 15,135,63,255,255,255 // ja 19b0 <_sk_lerp_565_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 13d4 <_sk_lerp_565_avx+0x129>
+ .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 1ac8 <_sk_lerp_565_avx+0x12c>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -13423,27 +15068,26 @@ _sk_lerp_565_avx:
.byte 196,65,57,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8
.byte 196,65,57,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8
.byte 196,65,57,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8
- .byte 233,235,254,255,255 // jmpq 12bf <_sk_lerp_565_avx+0x14>
- .byte 244 // hlt
- .byte 255 // (bad)
+ .byte 233,235,254,255,255 // jmpq 19b0 <_sk_lerp_565_avx+0x14>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 241 // icebp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 236 // in (%dx),%al
.byte 255 // (bad)
+ .byte 233,255,255,255,225 // jmpq ffffffffe2001ad0 <_sk_callback_avx+0xffffffffe1ffbf6d>
.byte 255 // (bad)
- .byte 255,228 // jmpq *%rsp
.byte 255 // (bad)
.byte 255 // (bad)
+ .byte 217,255 // fcos
.byte 255 // (bad)
- .byte 220,255 // fdivr %st,%st(7)
+ .byte 255,209 // callq *%rcx
.byte 255 // (bad)
- .byte 255,212 // callq *%rsp
.byte 255 // (bad)
+ .byte 255,201 // dec %ecx
.byte 255 // (bad)
- .byte 255,204 // dec %esp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,192 // inc %eax
+ .byte 189 // .byte 0xbd
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -13455,7 +15099,7 @@ _sk_load_tables_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,0 // mov (%rax),%r8
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,26,2,0,0 // jne 1618 <_sk_load_tables_avx+0x228>
+ .byte 15,133,26,2,0,0 // jne 1d0c <_sk_load_tables_avx+0x228>
.byte 196,65,124,16,4,184 // vmovups (%r8,%rdi,4),%ymm8
.byte 85 // push %rbp
.byte 65,87 // push %r15
@@ -13463,7 +15107,7 @@ _sk_load_tables_avx:
.byte 65,85 // push %r13
.byte 65,84 // push %r12
.byte 83 // push %rbx
- .byte 197,124,40,13,74,69,0,0 // vmovaps 0x454a(%rip),%ymm9 # 5960 <_sk_callback_avx+0x4f1>
+ .byte 197,124,40,13,150,69,0,0 // vmovaps 0x4596(%rip),%ymm9 # 60a0 <_sk_callback_avx+0x53d>
.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
@@ -13555,7 +15199,7 @@ _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 196,98,125,24,5,135,64,0,0 // vbroadcastss 0x4087(%rip),%ymm8 # 568c <_sk_callback_avx+0x21d>
+ .byte 196,98,125,24,5,199,64,0,0 // vbroadcastss 0x40c7(%rip),%ymm8 # 5dc0 <_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
@@ -13570,9 +15214,9 @@ _sk_load_tables_avx:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
.byte 65,254,201 // dec %r9b
.byte 65,128,249,6 // cmp $0x6,%r9b
- .byte 15,135,211,253,255,255 // ja 1404 <_sk_load_tables_avx+0x14>
+ .byte 15,135,211,253,255,255 // ja 1af8 <_sk_load_tables_avx+0x14>
.byte 69,15,182,201 // movzbl %r9b,%r9d
- .byte 76,141,21,140,0,0,0 // lea 0x8c(%rip),%r10 # 16c8 <_sk_load_tables_avx+0x2d8>
+ .byte 76,141,21,140,0,0,0 // lea 0x8c(%rip),%r10 # 1dbc <_sk_load_tables_avx+0x2d8>
.byte 79,99,12,138 // movslq (%r10,%r9,4),%r9
.byte 77,1,209 // add %r10,%r9
.byte 65,255,225 // jmpq *%r9
@@ -13595,7 +15239,7 @@ _sk_load_tables_avx:
.byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8
.byte 196,195,57,34,4,184,0 // vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0
.byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8
- .byte 233,62,253,255,255 // jmpq 1404 <_sk_load_tables_avx+0x14>
+ .byte 233,62,253,255,255 // jmpq 1af8 <_sk_load_tables_avx+0x14>
.byte 102,144 // xchg %ax,%ax
.byte 236 // in (%dx),%al
.byte 255 // (bad)
@@ -13613,7 +15257,7 @@ _sk_load_tables_avx:
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 126,255 // jle 16e1 <_sk_load_tables_avx+0x2f1>
+ .byte 126,255 // jle 1dd5 <_sk_load_tables_avx+0x2f1>
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -13625,7 +15269,7 @@ _sk_load_tables_u16_be_avx:
.byte 76,139,0 // mov (%rax),%r8
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,113,2,0,0 // jne 196b <_sk_load_tables_u16_be_avx+0x287>
+ .byte 15,133,113,2,0,0 // jne 205f <_sk_load_tables_u16_be_avx+0x287>
.byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8
.byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2
.byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3
@@ -13647,7 +15291,7 @@ _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 197,121,111,29,138,66,0,0 // vmovdqa 0x428a(%rip),%xmm11 # 59e0 <_sk_callback_avx+0x571>
+ .byte 197,121,111,29,214,66,0,0 // vmovdqa 0x42d6(%rip),%xmm11 # 6120 <_sk_callback_avx+0x5bd>
.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
@@ -13746,7 +15390,7 @@ _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 196,98,125,24,5,56,61,0,0 // vbroadcastss 0x3d38(%rip),%ymm8 # 5690 <_sk_callback_avx+0x221>
+ .byte 196,98,125,24,5,120,61,0,0 // vbroadcastss 0x3d78(%rip),%ymm8 # 5dc4 <_sk_callback_avx+0x261>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 91 // pop %rbx
@@ -13759,29 +15403,29 @@ _sk_load_tables_u16_be_avx:
.byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,85 // je 19d1 <_sk_load_tables_u16_be_avx+0x2ed>
+ .byte 116,85 // je 20c5 <_sk_load_tables_u16_be_avx+0x2ed>
.byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,72 // jb 19d1 <_sk_load_tables_u16_be_avx+0x2ed>
+ .byte 114,72 // jb 20c5 <_sk_load_tables_u16_be_avx+0x2ed>
.byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,72 // je 19de <_sk_load_tables_u16_be_avx+0x2fa>
+ .byte 116,72 // je 20d2 <_sk_load_tables_u16_be_avx+0x2fa>
.byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,59 // jb 19de <_sk_load_tables_u16_be_avx+0x2fa>
+ .byte 114,59 // jb 20d2 <_sk_load_tables_u16_be_avx+0x2fa>
.byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,97,253,255,255 // je 1715 <_sk_load_tables_u16_be_avx+0x31>
+ .byte 15,132,97,253,255,255 // je 1e09 <_sk_load_tables_u16_be_avx+0x31>
.byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,80,253,255,255 // jb 1715 <_sk_load_tables_u16_be_avx+0x31>
+ .byte 15,130,80,253,255,255 // jb 1e09 <_sk_load_tables_u16_be_avx+0x31>
.byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9
- .byte 233,68,253,255,255 // jmpq 1715 <_sk_load_tables_u16_be_avx+0x31>
+ .byte 233,68,253,255,255 // jmpq 1e09 <_sk_load_tables_u16_be_avx+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,55,253,255,255 // jmpq 1715 <_sk_load_tables_u16_be_avx+0x31>
+ .byte 233,55,253,255,255 // jmpq 1e09 <_sk_load_tables_u16_be_avx+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,46,253,255,255 // jmpq 1715 <_sk_load_tables_u16_be_avx+0x31>
+ .byte 233,46,253,255,255 // jmpq 1e09 <_sk_load_tables_u16_be_avx+0x31>
HIDDEN _sk_load_tables_rgb_u16_be_avx
.globl _sk_load_tables_rgb_u16_be_avx
@@ -13791,7 +15435,7 @@ _sk_load_tables_rgb_u16_be_avx:
.byte 76,139,0 // mov (%rax),%r8
.byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,93,2,0,0 // jne 1c56 <_sk_load_tables_rgb_u16_be_avx+0x26f>
+ .byte 15,133,93,2,0,0 // jne 234a <_sk_load_tables_rgb_u16_be_avx+0x26f>
.byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0
.byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2
.byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1
@@ -13818,7 +15462,7 @@ _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 197,121,111,13,131,63,0,0 // vmovdqa 0x3f83(%rip),%xmm9 # 59f0 <_sk_callback_avx+0x581>
+ .byte 197,121,111,13,207,63,0,0 // vmovdqa 0x3fcf(%rip),%xmm9 # 6130 <_sk_callback_avx+0x5cd>
.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
@@ -13910,7 +15554,7 @@ _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 196,226,125,24,29,74,58,0,0 // vbroadcastss 0x3a4a(%rip),%ymm3 # 5694 <_sk_callback_avx+0x225>
+ .byte 196,226,125,24,29,138,58,0,0 // vbroadcastss 0x3a8a(%rip),%ymm3 # 5dc8 <_sk_callback_avx+0x265>
.byte 91 // pop %rbx
.byte 65,92 // pop %r12
.byte 65,93 // pop %r13
@@ -13921,36 +15565,36 @@ _sk_load_tables_rgb_u16_be_avx:
.byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0
.byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 117,5 // jne 1c6f <_sk_load_tables_rgb_u16_be_avx+0x288>
- .byte 233,190,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 117,5 // jne 2363 <_sk_load_tables_rgb_u16_be_avx+0x288>
+ .byte 233,190,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
.byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1
.byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,26 // jb 1c9e <_sk_load_tables_rgb_u16_be_avx+0x2b7>
+ .byte 114,26 // jb 2392 <_sk_load_tables_rgb_u16_be_avx+0x2b7>
.byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1
.byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 117,10 // jne 1ca3 <_sk_load_tables_rgb_u16_be_avx+0x2bc>
- .byte 233,143,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
- .byte 233,138,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 117,10 // jne 2397 <_sk_load_tables_rgb_u16_be_avx+0x2bc>
+ .byte 233,143,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 233,138,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
.byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1
.byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,26 // jb 1cd2 <_sk_load_tables_rgb_u16_be_avx+0x2eb>
+ .byte 114,26 // jb 23c6 <_sk_load_tables_rgb_u16_be_avx+0x2eb>
.byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1
.byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 117,10 // jne 1cd7 <_sk_load_tables_rgb_u16_be_avx+0x2f0>
- .byte 233,91,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
- .byte 233,86,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 117,10 // jne 23cb <_sk_load_tables_rgb_u16_be_avx+0x2f0>
+ .byte 233,91,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 233,86,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
.byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3
.byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,20 // jb 1d00 <_sk_load_tables_rgb_u16_be_avx+0x319>
+ .byte 114,20 // jb 23f4 <_sk_load_tables_rgb_u16_be_avx+0x319>
.byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3
.byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
- .byte 233,45,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
- .byte 233,40,253,255,255 // jmpq 1a2d <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 233,45,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ .byte 233,40,253,255,255 // jmpq 2121 <_sk_load_tables_rgb_u16_be_avx+0x46>
HIDDEN _sk_byte_tables_avx
.globl _sk_byte_tables_avx
@@ -13963,7 +15607,7 @@ _sk_byte_tables_avx:
.byte 65,84 // push %r12
.byte 83 // push %rbx
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,126,57,0,0 // vbroadcastss 0x397e(%rip),%ymm8 # 5698 <_sk_callback_avx+0x229>
+ .byte 196,98,125,24,5,190,57,0,0 // vbroadcastss 0x39be(%rip),%ymm8 # 5dcc <_sk_callback_avx+0x269>
.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
@@ -14000,7 +15644,7 @@ _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 196,98,125,24,13,204,56,0,0 // vbroadcastss 0x38cc(%rip),%ymm9 # 569c <_sk_callback_avx+0x22d>
+ .byte 196,98,125,24,13,12,57,0,0 // vbroadcastss 0x390c(%rip),%ymm9 # 5dd0 <_sk_callback_avx+0x26d>
.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
@@ -14162,7 +15806,7 @@ _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 196,98,125,24,13,242,53,0,0 // vbroadcastss 0x35f2(%rip),%ymm9 # 56a0 <_sk_callback_avx+0x231>
+ .byte 196,98,125,24,13,50,54,0,0 // vbroadcastss 0x3632(%rip),%ymm9 # 5dd4 <_sk_callback_avx+0x271>
.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
@@ -14459,36 +16103,36 @@ _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 196,98,125,24,37,80,49,0,0 // vbroadcastss 0x3150(%rip),%ymm12 # 56a4 <_sk_callback_avx+0x235>
+ .byte 196,98,125,24,37,144,49,0,0 // vbroadcastss 0x3190(%rip),%ymm12 # 5dd8 <_sk_callback_avx+0x275>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,70,49,0,0 // vbroadcastss 0x3146(%rip),%ymm12 # 56a8 <_sk_callback_avx+0x239>
+ .byte 196,98,125,24,37,134,49,0,0 // vbroadcastss 0x3186(%rip),%ymm12 # 5ddc <_sk_callback_avx+0x279>
.byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0
- .byte 196,98,125,24,37,60,49,0,0 // vbroadcastss 0x313c(%rip),%ymm12 # 56ac <_sk_callback_avx+0x23d>
+ .byte 196,98,125,24,37,124,49,0,0 // vbroadcastss 0x317c(%rip),%ymm12 # 5de0 <_sk_callback_avx+0x27d>
.byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0
- .byte 196,98,125,24,37,50,49,0,0 // vbroadcastss 0x3132(%rip),%ymm12 # 56b0 <_sk_callback_avx+0x241>
+ .byte 196,98,125,24,37,114,49,0,0 // vbroadcastss 0x3172(%rip),%ymm12 # 5de4 <_sk_callback_avx+0x281>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,40,49,0,0 // vbroadcastss 0x3128(%rip),%ymm12 # 56b4 <_sk_callback_avx+0x245>
+ .byte 196,98,125,24,37,104,49,0,0 // vbroadcastss 0x3168(%rip),%ymm12 # 5de8 <_sk_callback_avx+0x285>
.byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,25,49,0,0 // vbroadcastss 0x3119(%rip),%ymm12 # 56b8 <_sk_callback_avx+0x249>
+ .byte 196,98,125,24,37,89,49,0,0 // vbroadcastss 0x3159(%rip),%ymm12 # 5dec <_sk_callback_avx+0x289>
.byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0
- .byte 196,98,125,24,37,15,49,0,0 // vbroadcastss 0x310f(%rip),%ymm12 # 56bc <_sk_callback_avx+0x24d>
+ .byte 196,98,125,24,37,79,49,0,0 // vbroadcastss 0x314f(%rip),%ymm12 # 5df0 <_sk_callback_avx+0x28d>
.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 196,98,125,24,29,243,48,0,0 // vbroadcastss 0x30f3(%rip),%ymm11 # 56c0 <_sk_callback_avx+0x251>
+ .byte 196,98,125,24,29,51,49,0,0 // vbroadcastss 0x3133(%rip),%ymm11 # 5df4 <_sk_callback_avx+0x291>
.byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,233,48,0,0 // vbroadcastss 0x30e9(%rip),%ymm11 # 56c4 <_sk_callback_avx+0x255>
+ .byte 196,98,125,24,29,41,49,0,0 // vbroadcastss 0x3129(%rip),%ymm11 # 5df8 <_sk_callback_avx+0x295>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,218,48,0,0 // vbroadcastss 0x30da(%rip),%ymm11 # 56c8 <_sk_callback_avx+0x259>
+ .byte 196,98,125,24,29,26,49,0,0 // vbroadcastss 0x311a(%rip),%ymm11 # 5dfc <_sk_callback_avx+0x299>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,208,48,0,0 // vbroadcastss 0x30d0(%rip),%ymm11 # 56cc <_sk_callback_avx+0x25d>
+ .byte 196,98,125,24,29,16,49,0,0 // vbroadcastss 0x3110(%rip),%ymm11 # 5e00 <_sk_callback_avx+0x29d>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
- .byte 196,98,125,24,21,193,48,0,0 // vbroadcastss 0x30c1(%rip),%ymm10 # 56d0 <_sk_callback_avx+0x261>
+ .byte 196,98,125,24,21,1,49,0,0 // vbroadcastss 0x3101(%rip),%ymm10 # 5e04 <_sk_callback_avx+0x2a1>
.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
@@ -14496,7 +16140,7 @@ _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 196,98,125,24,5,152,48,0,0 // vbroadcastss 0x3098(%rip),%ymm8 # 56d4 <_sk_callback_avx+0x265>
+ .byte 196,98,125,24,5,216,48,0,0 // vbroadcastss 0x30d8(%rip),%ymm8 # 5e08 <_sk_callback_avx+0x2a5>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -14518,36 +16162,36 @@ _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 196,98,125,24,37,73,48,0,0 // vbroadcastss 0x3049(%rip),%ymm12 # 56d8 <_sk_callback_avx+0x269>
+ .byte 196,98,125,24,37,137,48,0,0 // vbroadcastss 0x3089(%rip),%ymm12 # 5e0c <_sk_callback_avx+0x2a9>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,63,48,0,0 // vbroadcastss 0x303f(%rip),%ymm12 # 56dc <_sk_callback_avx+0x26d>
+ .byte 196,98,125,24,37,127,48,0,0 // vbroadcastss 0x307f(%rip),%ymm12 # 5e10 <_sk_callback_avx+0x2ad>
.byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1
- .byte 196,98,125,24,37,53,48,0,0 // vbroadcastss 0x3035(%rip),%ymm12 # 56e0 <_sk_callback_avx+0x271>
+ .byte 196,98,125,24,37,117,48,0,0 // vbroadcastss 0x3075(%rip),%ymm12 # 5e14 <_sk_callback_avx+0x2b1>
.byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1
- .byte 196,98,125,24,37,43,48,0,0 // vbroadcastss 0x302b(%rip),%ymm12 # 56e4 <_sk_callback_avx+0x275>
+ .byte 196,98,125,24,37,107,48,0,0 // vbroadcastss 0x306b(%rip),%ymm12 # 5e18 <_sk_callback_avx+0x2b5>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,33,48,0,0 // vbroadcastss 0x3021(%rip),%ymm12 # 56e8 <_sk_callback_avx+0x279>
+ .byte 196,98,125,24,37,97,48,0,0 // vbroadcastss 0x3061(%rip),%ymm12 # 5e1c <_sk_callback_avx+0x2b9>
.byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,18,48,0,0 // vbroadcastss 0x3012(%rip),%ymm12 # 56ec <_sk_callback_avx+0x27d>
+ .byte 196,98,125,24,37,82,48,0,0 // vbroadcastss 0x3052(%rip),%ymm12 # 5e20 <_sk_callback_avx+0x2bd>
.byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1
- .byte 196,98,125,24,37,8,48,0,0 // vbroadcastss 0x3008(%rip),%ymm12 # 56f0 <_sk_callback_avx+0x281>
+ .byte 196,98,125,24,37,72,48,0,0 // vbroadcastss 0x3048(%rip),%ymm12 # 5e24 <_sk_callback_avx+0x2c1>
.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 196,98,125,24,29,236,47,0,0 // vbroadcastss 0x2fec(%rip),%ymm11 # 56f4 <_sk_callback_avx+0x285>
+ .byte 196,98,125,24,29,44,48,0,0 // vbroadcastss 0x302c(%rip),%ymm11 # 5e28 <_sk_callback_avx+0x2c5>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,226,47,0,0 // vbroadcastss 0x2fe2(%rip),%ymm11 # 56f8 <_sk_callback_avx+0x289>
+ .byte 196,98,125,24,29,34,48,0,0 // vbroadcastss 0x3022(%rip),%ymm11 # 5e2c <_sk_callback_avx+0x2c9>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,211,47,0,0 // vbroadcastss 0x2fd3(%rip),%ymm11 # 56fc <_sk_callback_avx+0x28d>
+ .byte 196,98,125,24,29,19,48,0,0 // vbroadcastss 0x3013(%rip),%ymm11 # 5e30 <_sk_callback_avx+0x2cd>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,201,47,0,0 // vbroadcastss 0x2fc9(%rip),%ymm11 # 5700 <_sk_callback_avx+0x291>
+ .byte 196,98,125,24,29,9,48,0,0 // vbroadcastss 0x3009(%rip),%ymm11 # 5e34 <_sk_callback_avx+0x2d1>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
- .byte 196,98,125,24,21,186,47,0,0 // vbroadcastss 0x2fba(%rip),%ymm10 # 5704 <_sk_callback_avx+0x295>
+ .byte 196,98,125,24,21,250,47,0,0 // vbroadcastss 0x2ffa(%rip),%ymm10 # 5e38 <_sk_callback_avx+0x2d5>
.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
@@ -14555,7 +16199,7 @@ _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 196,98,125,24,5,145,47,0,0 // vbroadcastss 0x2f91(%rip),%ymm8 # 5708 <_sk_callback_avx+0x299>
+ .byte 196,98,125,24,5,209,47,0,0 // vbroadcastss 0x2fd1(%rip),%ymm8 # 5e3c <_sk_callback_avx+0x2d9>
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -14577,36 +16221,36 @@ _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 196,98,125,24,37,66,47,0,0 // vbroadcastss 0x2f42(%rip),%ymm12 # 570c <_sk_callback_avx+0x29d>
+ .byte 196,98,125,24,37,130,47,0,0 // vbroadcastss 0x2f82(%rip),%ymm12 # 5e40 <_sk_callback_avx+0x2dd>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,56,47,0,0 // vbroadcastss 0x2f38(%rip),%ymm12 # 5710 <_sk_callback_avx+0x2a1>
+ .byte 196,98,125,24,37,120,47,0,0 // vbroadcastss 0x2f78(%rip),%ymm12 # 5e44 <_sk_callback_avx+0x2e1>
.byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2
- .byte 196,98,125,24,37,46,47,0,0 // vbroadcastss 0x2f2e(%rip),%ymm12 # 5714 <_sk_callback_avx+0x2a5>
+ .byte 196,98,125,24,37,110,47,0,0 // vbroadcastss 0x2f6e(%rip),%ymm12 # 5e48 <_sk_callback_avx+0x2e5>
.byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2
- .byte 196,98,125,24,37,36,47,0,0 // vbroadcastss 0x2f24(%rip),%ymm12 # 5718 <_sk_callback_avx+0x2a9>
+ .byte 196,98,125,24,37,100,47,0,0 // vbroadcastss 0x2f64(%rip),%ymm12 # 5e4c <_sk_callback_avx+0x2e9>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,26,47,0,0 // vbroadcastss 0x2f1a(%rip),%ymm12 # 571c <_sk_callback_avx+0x2ad>
+ .byte 196,98,125,24,37,90,47,0,0 // vbroadcastss 0x2f5a(%rip),%ymm12 # 5e50 <_sk_callback_avx+0x2ed>
.byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,11,47,0,0 // vbroadcastss 0x2f0b(%rip),%ymm12 # 5720 <_sk_callback_avx+0x2b1>
+ .byte 196,98,125,24,37,75,47,0,0 // vbroadcastss 0x2f4b(%rip),%ymm12 # 5e54 <_sk_callback_avx+0x2f1>
.byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2
- .byte 196,98,125,24,37,1,47,0,0 // vbroadcastss 0x2f01(%rip),%ymm12 # 5724 <_sk_callback_avx+0x2b5>
+ .byte 196,98,125,24,37,65,47,0,0 // vbroadcastss 0x2f41(%rip),%ymm12 # 5e58 <_sk_callback_avx+0x2f5>
.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 196,98,125,24,29,229,46,0,0 // vbroadcastss 0x2ee5(%rip),%ymm11 # 5728 <_sk_callback_avx+0x2b9>
+ .byte 196,98,125,24,29,37,47,0,0 // vbroadcastss 0x2f25(%rip),%ymm11 # 5e5c <_sk_callback_avx+0x2f9>
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,219,46,0,0 // vbroadcastss 0x2edb(%rip),%ymm11 # 572c <_sk_callback_avx+0x2bd>
+ .byte 196,98,125,24,29,27,47,0,0 // vbroadcastss 0x2f1b(%rip),%ymm11 # 5e60 <_sk_callback_avx+0x2fd>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,204,46,0,0 // vbroadcastss 0x2ecc(%rip),%ymm11 # 5730 <_sk_callback_avx+0x2c1>
+ .byte 196,98,125,24,29,12,47,0,0 // vbroadcastss 0x2f0c(%rip),%ymm11 # 5e64 <_sk_callback_avx+0x301>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,194,46,0,0 // vbroadcastss 0x2ec2(%rip),%ymm11 # 5734 <_sk_callback_avx+0x2c5>
+ .byte 196,98,125,24,29,2,47,0,0 // vbroadcastss 0x2f02(%rip),%ymm11 # 5e68 <_sk_callback_avx+0x305>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
- .byte 196,98,125,24,21,179,46,0,0 // vbroadcastss 0x2eb3(%rip),%ymm10 # 5738 <_sk_callback_avx+0x2c9>
+ .byte 196,98,125,24,21,243,46,0,0 // vbroadcastss 0x2ef3(%rip),%ymm10 # 5e6c <_sk_callback_avx+0x309>
.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
@@ -14614,7 +16258,7 @@ _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 196,98,125,24,5,138,46,0,0 // vbroadcastss 0x2e8a(%rip),%ymm8 # 573c <_sk_callback_avx+0x2cd>
+ .byte 196,98,125,24,5,202,46,0,0 // vbroadcastss 0x2eca(%rip),%ymm8 # 5e70 <_sk_callback_avx+0x30d>
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -14636,36 +16280,36 @@ _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 196,98,125,24,37,59,46,0,0 // vbroadcastss 0x2e3b(%rip),%ymm12 # 5740 <_sk_callback_avx+0x2d1>
+ .byte 196,98,125,24,37,123,46,0,0 // vbroadcastss 0x2e7b(%rip),%ymm12 # 5e74 <_sk_callback_avx+0x311>
.byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,49,46,0,0 // vbroadcastss 0x2e31(%rip),%ymm12 # 5744 <_sk_callback_avx+0x2d5>
+ .byte 196,98,125,24,37,113,46,0,0 // vbroadcastss 0x2e71(%rip),%ymm12 # 5e78 <_sk_callback_avx+0x315>
.byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3
- .byte 196,98,125,24,37,39,46,0,0 // vbroadcastss 0x2e27(%rip),%ymm12 # 5748 <_sk_callback_avx+0x2d9>
+ .byte 196,98,125,24,37,103,46,0,0 // vbroadcastss 0x2e67(%rip),%ymm12 # 5e7c <_sk_callback_avx+0x319>
.byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3
- .byte 196,98,125,24,37,29,46,0,0 // vbroadcastss 0x2e1d(%rip),%ymm12 # 574c <_sk_callback_avx+0x2dd>
+ .byte 196,98,125,24,37,93,46,0,0 // vbroadcastss 0x2e5d(%rip),%ymm12 # 5e80 <_sk_callback_avx+0x31d>
.byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,19,46,0,0 // vbroadcastss 0x2e13(%rip),%ymm12 # 5750 <_sk_callback_avx+0x2e1>
+ .byte 196,98,125,24,37,83,46,0,0 // vbroadcastss 0x2e53(%rip),%ymm12 # 5e84 <_sk_callback_avx+0x321>
.byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12
.byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
- .byte 196,98,125,24,37,4,46,0,0 // vbroadcastss 0x2e04(%rip),%ymm12 # 5754 <_sk_callback_avx+0x2e5>
+ .byte 196,98,125,24,37,68,46,0,0 // vbroadcastss 0x2e44(%rip),%ymm12 # 5e88 <_sk_callback_avx+0x325>
.byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3
- .byte 196,98,125,24,37,250,45,0,0 // vbroadcastss 0x2dfa(%rip),%ymm12 # 5758 <_sk_callback_avx+0x2e9>
+ .byte 196,98,125,24,37,58,46,0,0 // vbroadcastss 0x2e3a(%rip),%ymm12 # 5e8c <_sk_callback_avx+0x329>
.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 196,98,125,24,29,222,45,0,0 // vbroadcastss 0x2dde(%rip),%ymm11 # 575c <_sk_callback_avx+0x2ed>
+ .byte 196,98,125,24,29,30,46,0,0 // vbroadcastss 0x2e1e(%rip),%ymm11 # 5e90 <_sk_callback_avx+0x32d>
.byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,212,45,0,0 // vbroadcastss 0x2dd4(%rip),%ymm11 # 5760 <_sk_callback_avx+0x2f1>
+ .byte 196,98,125,24,29,20,46,0,0 // vbroadcastss 0x2e14(%rip),%ymm11 # 5e94 <_sk_callback_avx+0x331>
.byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
.byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,197,45,0,0 // vbroadcastss 0x2dc5(%rip),%ymm11 # 5764 <_sk_callback_avx+0x2f5>
+ .byte 196,98,125,24,29,5,46,0,0 // vbroadcastss 0x2e05(%rip),%ymm11 # 5e98 <_sk_callback_avx+0x335>
.byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
- .byte 196,98,125,24,29,187,45,0,0 // vbroadcastss 0x2dbb(%rip),%ymm11 # 5768 <_sk_callback_avx+0x2f9>
+ .byte 196,98,125,24,29,251,45,0,0 // vbroadcastss 0x2dfb(%rip),%ymm11 # 5e9c <_sk_callback_avx+0x339>
.byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
.byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3
- .byte 196,98,125,24,21,172,45,0,0 // vbroadcastss 0x2dac(%rip),%ymm10 # 576c <_sk_callback_avx+0x2fd>
+ .byte 196,98,125,24,21,236,45,0,0 // vbroadcastss 0x2dec(%rip),%ymm10 # 5ea0 <_sk_callback_avx+0x33d>
.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
@@ -14673,7 +16317,7 @@ _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 196,98,125,24,5,131,45,0,0 // vbroadcastss 0x2d83(%rip),%ymm8 # 5770 <_sk_callback_avx+0x301>
+ .byte 196,98,125,24,5,195,45,0,0 // vbroadcastss 0x2dc3(%rip),%ymm8 # 5ea4 <_sk_callback_avx+0x341>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -14682,31 +16326,31 @@ HIDDEN _sk_lab_to_xyz_avx
.globl _sk_lab_to_xyz_avx
FUNCTION(_sk_lab_to_xyz_avx)
_sk_lab_to_xyz_avx:
- .byte 196,98,125,24,5,117,45,0,0 // vbroadcastss 0x2d75(%rip),%ymm8 # 5774 <_sk_callback_avx+0x305>
+ .byte 196,98,125,24,5,181,45,0,0 // vbroadcastss 0x2db5(%rip),%ymm8 # 5ea8 <_sk_callback_avx+0x345>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,107,45,0,0 // vbroadcastss 0x2d6b(%rip),%ymm8 # 5778 <_sk_callback_avx+0x309>
+ .byte 196,98,125,24,5,171,45,0,0 // vbroadcastss 0x2dab(%rip),%ymm8 # 5eac <_sk_callback_avx+0x349>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,98,125,24,13,97,45,0,0 // vbroadcastss 0x2d61(%rip),%ymm9 # 577c <_sk_callback_avx+0x30d>
+ .byte 196,98,125,24,13,161,45,0,0 // vbroadcastss 0x2da1(%rip),%ymm9 # 5eb0 <_sk_callback_avx+0x34d>
.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 196,98,125,24,5,77,45,0,0 // vbroadcastss 0x2d4d(%rip),%ymm8 # 5780 <_sk_callback_avx+0x311>
+ .byte 196,98,125,24,5,141,45,0,0 // vbroadcastss 0x2d8d(%rip),%ymm8 # 5eb4 <_sk_callback_avx+0x351>
.byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,67,45,0,0 // vbroadcastss 0x2d43(%rip),%ymm8 # 5784 <_sk_callback_avx+0x315>
+ .byte 196,98,125,24,5,131,45,0,0 // vbroadcastss 0x2d83(%rip),%ymm8 # 5eb8 <_sk_callback_avx+0x355>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,98,125,24,5,57,45,0,0 // vbroadcastss 0x2d39(%rip),%ymm8 # 5788 <_sk_callback_avx+0x319>
+ .byte 196,98,125,24,5,121,45,0,0 // vbroadcastss 0x2d79(%rip),%ymm8 # 5ebc <_sk_callback_avx+0x359>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
.byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1
- .byte 196,98,125,24,5,43,45,0,0 // vbroadcastss 0x2d2b(%rip),%ymm8 # 578c <_sk_callback_avx+0x31d>
+ .byte 196,98,125,24,5,107,45,0,0 // vbroadcastss 0x2d6b(%rip),%ymm8 # 5ec0 <_sk_callback_avx+0x35d>
.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 196,98,125,24,13,20,45,0,0 // vbroadcastss 0x2d14(%rip),%ymm9 # 5790 <_sk_callback_avx+0x321>
+ .byte 196,98,125,24,13,84,45,0,0 // vbroadcastss 0x2d54(%rip),%ymm9 # 5ec4 <_sk_callback_avx+0x361>
.byte 196,65,52,194,208,1 // vcmpltps %ymm8,%ymm9,%ymm10
- .byte 196,98,125,24,29,9,45,0,0 // vbroadcastss 0x2d09(%rip),%ymm11 # 5794 <_sk_callback_avx+0x325>
+ .byte 196,98,125,24,29,73,45,0,0 // vbroadcastss 0x2d49(%rip),%ymm11 # 5ec8 <_sk_callback_avx+0x365>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,37,255,44,0,0 // vbroadcastss 0x2cff(%rip),%ymm12 # 5798 <_sk_callback_avx+0x329>
+ .byte 196,98,125,24,37,63,45,0,0 // vbroadcastss 0x2d3f(%rip),%ymm12 # 5ecc <_sk_callback_avx+0x369>
.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
@@ -14721,9 +16365,9 @@ _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 196,226,125,24,5,181,44,0,0 // vbroadcastss 0x2cb5(%rip),%ymm0 # 579c <_sk_callback_avx+0x32d>
+ .byte 196,226,125,24,5,245,44,0,0 // vbroadcastss 0x2cf5(%rip),%ymm0 # 5ed0 <_sk_callback_avx+0x36d>
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,172,44,0,0 // vbroadcastss 0x2cac(%rip),%ymm8 # 57a0 <_sk_callback_avx+0x331>
+ .byte 196,98,125,24,5,236,44,0,0 // vbroadcastss 0x2cec(%rip),%ymm8 # 5ed4 <_sk_callback_avx+0x371>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -14737,14 +16381,14 @@ _sk_load_a8_avx:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,62 // jne 2b4b <_sk_load_a8_avx+0x4e>
+ .byte 117,62 // jne 323f <_sk_load_a8_avx+0x4e>
.byte 197,250,126,0 // vmovq (%rax),%xmm0
.byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
.byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
.byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
.byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,112,44,0,0 // vbroadcastss 0x2c70(%rip),%ymm1 # 57a4 <_sk_callback_avx+0x335>
+ .byte 196,226,125,24,13,176,44,0,0 // vbroadcastss 0x2cb0(%rip),%ymm1 # 5ed8 <_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
@@ -14761,9 +16405,9 @@ _sk_load_a8_avx:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 2b53 <_sk_load_a8_avx+0x56>
+ .byte 117,234 // jne 3247 <_sk_load_a8_avx+0x56>
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
- .byte 235,161 // jmp 2b11 <_sk_load_a8_avx+0x14>
+ .byte 235,161 // jmp 3205 <_sk_load_a8_avx+0x14>
HIDDEN _sk_gather_a8_avx
.globl _sk_gather_a8_avx
@@ -14813,7 +16457,7 @@ _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 196,226,125,24,13,101,43,0,0 // vbroadcastss 0x2b65(%rip),%ymm1 # 57a8 <_sk_callback_avx+0x339>
+ .byte 196,226,125,24,13,165,43,0,0 // vbroadcastss 0x2ba5(%rip),%ymm1 # 5edc <_sk_callback_avx+0x379>
.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
@@ -14831,14 +16475,14 @@ FUNCTION(_sk_store_a8_avx)
_sk_store_a8_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,64,43,0,0 // vbroadcastss 0x2b40(%rip),%ymm8 # 57ac <_sk_callback_avx+0x33d>
+ .byte 196,98,125,24,5,128,43,0,0 // vbroadcastss 0x2b80(%rip),%ymm8 # 5ee0 <_sk_callback_avx+0x37d>
.byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8
.byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 2c95 <_sk_store_a8_avx+0x37>
+ .byte 117,10 // jne 3389 <_sk_store_a8_avx+0x37>
.byte 196,65,123,17,4,58 // vmovsd %xmm8,(%r10,%rdi,1)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -14846,10 +16490,10 @@ _sk_store_a8_avx:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 2c91 <_sk_store_a8_avx+0x33>
+ .byte 119,236 // ja 3385 <_sk_store_a8_avx+0x33>
.byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 2cf8 <_sk_store_a8_avx+0x9a>
+ .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 33ec <_sk_store_a8_avx+0x9a>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -14860,7 +16504,7 @@ _sk_store_a8_avx:
.byte 196,67,121,20,68,58,2,4 // vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1)
.byte 196,67,121,20,68,58,1,2 // vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1)
.byte 196,67,121,20,4,58,0 // vpextrb $0x0,%xmm8,(%r10,%rdi,1)
- .byte 235,154 // jmp 2c91 <_sk_store_a8_avx+0x33>
+ .byte 235,154 // jmp 3385 <_sk_store_a8_avx+0x33>
.byte 144 // nop
.byte 246,255 // idiv %bh
.byte 255 // (bad)
@@ -14894,17 +16538,17 @@ _sk_load_g8_avx:
.byte 72,139,0 // mov (%rax),%rax
.byte 72,1,248 // add %rdi,%rax
.byte 77,133,192 // test %r8,%r8
- .byte 117,67 // jne 2d67 <_sk_load_g8_avx+0x53>
+ .byte 117,67 // jne 345b <_sk_load_g8_avx+0x53>
.byte 197,250,126,0 // vmovq (%rax),%xmm0
.byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
.byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
.byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
.byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,101,42,0,0 // vbroadcastss 0x2a65(%rip),%ymm1 # 57b0 <_sk_callback_avx+0x341>
+ .byte 196,226,125,24,13,165,42,0,0 // vbroadcastss 0x2aa5(%rip),%ymm1 # 5ee4 <_sk_callback_avx+0x381>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,90,42,0,0 // vbroadcastss 0x2a5a(%rip),%ymm3 # 57b4 <_sk_callback_avx+0x345>
+ .byte 196,226,125,24,29,154,42,0,0 // vbroadcastss 0x2a9a(%rip),%ymm3 # 5ee8 <_sk_callback_avx+0x385>
.byte 76,137,193 // mov %r8,%rcx
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
@@ -14918,9 +16562,9 @@ _sk_load_g8_avx:
.byte 77,9,217 // or %r11,%r9
.byte 72,131,193,8 // add $0x8,%rcx
.byte 73,255,202 // dec %r10
- .byte 117,234 // jne 2d6f <_sk_load_g8_avx+0x5b>
+ .byte 117,234 // jne 3463 <_sk_load_g8_avx+0x5b>
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
- .byte 235,156 // jmp 2d28 <_sk_load_g8_avx+0x14>
+ .byte 235,156 // jmp 341c <_sk_load_g8_avx+0x14>
HIDDEN _sk_gather_g8_avx
.globl _sk_gather_g8_avx
@@ -14970,10 +16614,10 @@ _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 196,226,125,24,13,89,41,0,0 // vbroadcastss 0x2959(%rip),%ymm1 # 57b8 <_sk_callback_avx+0x349>
+ .byte 196,226,125,24,13,153,41,0,0 // vbroadcastss 0x2999(%rip),%ymm1 # 5eec <_sk_callback_avx+0x389>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,78,41,0,0 // vbroadcastss 0x294e(%rip),%ymm3 # 57bc <_sk_callback_avx+0x34d>
+ .byte 196,226,125,24,29,142,41,0,0 // vbroadcastss 0x298e(%rip),%ymm3 # 5ef0 <_sk_callback_avx+0x38d>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 91 // pop %rbx
@@ -14989,9 +16633,9 @@ _sk_gather_i8_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,137,192 // mov %rax,%r8
.byte 77,133,192 // test %r8,%r8
- .byte 116,5 // je 2e8e <_sk_gather_i8_avx+0xf>
+ .byte 116,5 // je 3582 <_sk_gather_i8_avx+0xf>
.byte 76,137,192 // mov %r8,%rax
- .byte 235,2 // jmp 2e90 <_sk_gather_i8_avx+0x11>
+ .byte 235,2 // jmp 3584 <_sk_gather_i8_avx+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,87 // push %r15
.byte 65,86 // push %r14
@@ -15053,10 +16697,10 @@ _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 197,124,40,21,198,41,0,0 // vmovaps 0x29c6(%rip),%ymm10 # 5980 <_sk_callback_avx+0x511>
+ .byte 197,124,40,21,18,42,0,0 // vmovaps 0x2a12(%rip),%ymm10 # 60c0 <_sk_callback_avx+0x55d>
.byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,13,244,39,0,0 // vbroadcastss 0x27f4(%rip),%ymm9 # 57c0 <_sk_callback_avx+0x351>
+ .byte 196,98,125,24,13,52,40,0,0 // vbroadcastss 0x2834(%rip),%ymm9 # 5ef4 <_sk_callback_avx+0x391>
.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
@@ -15090,38 +16734,38 @@ _sk_load_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,128,0,0,0 // jne 30c4 <_sk_load_565_avx+0x8e>
+ .byte 15,133,128,0,0,0 // jne 37b8 <_sk_load_565_avx+0x8e>
.byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
.byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
- .byte 196,226,125,24,5,94,39,0,0 // vbroadcastss 0x275e(%rip),%ymm0 # 57c4 <_sk_callback_avx+0x355>
+ .byte 196,226,125,24,5,158,39,0,0 // vbroadcastss 0x279e(%rip),%ymm0 # 5ef8 <_sk_callback_avx+0x395>
.byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,81,39,0,0 // vbroadcastss 0x2751(%rip),%ymm1 # 57c8 <_sk_callback_avx+0x359>
+ .byte 196,226,125,24,13,145,39,0,0 // vbroadcastss 0x2791(%rip),%ymm1 # 5efc <_sk_callback_avx+0x399>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,72,39,0,0 // vbroadcastss 0x2748(%rip),%ymm1 # 57cc <_sk_callback_avx+0x35d>
+ .byte 196,226,125,24,13,136,39,0,0 // vbroadcastss 0x2788(%rip),%ymm1 # 5f00 <_sk_callback_avx+0x39d>
.byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,59,39,0,0 // vbroadcastss 0x273b(%rip),%ymm3 # 57d0 <_sk_callback_avx+0x361>
+ .byte 196,226,125,24,29,123,39,0,0 // vbroadcastss 0x277b(%rip),%ymm3 # 5f04 <_sk_callback_avx+0x3a1>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,24,29,50,39,0,0 // vbroadcastss 0x2732(%rip),%ymm3 # 57d4 <_sk_callback_avx+0x365>
+ .byte 196,226,125,24,29,114,39,0,0 // vbroadcastss 0x2772(%rip),%ymm3 # 5f08 <_sk_callback_avx+0x3a5>
.byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,37,39,0,0 // vbroadcastss 0x2725(%rip),%ymm3 # 57d8 <_sk_callback_avx+0x369>
+ .byte 196,226,125,24,29,101,39,0,0 // vbroadcastss 0x2765(%rip),%ymm3 # 5f0c <_sk_callback_avx+0x3a9>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,26,39,0,0 // vbroadcastss 0x271a(%rip),%ymm3 # 57dc <_sk_callback_avx+0x36d>
+ .byte 196,226,125,24,29,90,39,0,0 // vbroadcastss 0x275a(%rip),%ymm3 # 5f10 <_sk_callback_avx+0x3ad>
.byte 255,224 // jmpq *%rax
.byte 65,137,200 // mov %ecx,%r8d
.byte 65,128,224,7 // and $0x7,%r8b
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,110,255,255,255 // ja 304a <_sk_load_565_avx+0x14>
+ .byte 15,135,110,255,255,255 // ja 373e <_sk_load_565_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 3130 <_sk_load_565_avx+0xfa>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 3824 <_sk_load_565_avx+0xfa>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -15133,7 +16777,7 @@ _sk_load_565_avx:
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- .byte 233,26,255,255,255 // jmpq 304a <_sk_load_565_avx+0x14>
+ .byte 233,26,255,255,255 // jmpq 373e <_sk_load_565_avx+0x14>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -15211,23 +16855,23 @@ _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 196,226,125,24,5,186,37,0,0 // vbroadcastss 0x25ba(%rip),%ymm0 # 57e0 <_sk_callback_avx+0x371>
+ .byte 196,226,125,24,5,250,37,0,0 // vbroadcastss 0x25fa(%rip),%ymm0 # 5f14 <_sk_callback_avx+0x3b1>
.byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,173,37,0,0 // vbroadcastss 0x25ad(%rip),%ymm1 # 57e4 <_sk_callback_avx+0x375>
+ .byte 196,226,125,24,13,237,37,0,0 // vbroadcastss 0x25ed(%rip),%ymm1 # 5f18 <_sk_callback_avx+0x3b5>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,164,37,0,0 // vbroadcastss 0x25a4(%rip),%ymm1 # 57e8 <_sk_callback_avx+0x379>
+ .byte 196,226,125,24,13,228,37,0,0 // vbroadcastss 0x25e4(%rip),%ymm1 # 5f1c <_sk_callback_avx+0x3b9>
.byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,151,37,0,0 // vbroadcastss 0x2597(%rip),%ymm3 # 57ec <_sk_callback_avx+0x37d>
+ .byte 196,226,125,24,29,215,37,0,0 // vbroadcastss 0x25d7(%rip),%ymm3 # 5f20 <_sk_callback_avx+0x3bd>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,24,29,142,37,0,0 // vbroadcastss 0x258e(%rip),%ymm3 # 57f0 <_sk_callback_avx+0x381>
+ .byte 196,226,125,24,29,206,37,0,0 // vbroadcastss 0x25ce(%rip),%ymm3 # 5f24 <_sk_callback_avx+0x3c1>
.byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,129,37,0,0 // vbroadcastss 0x2581(%rip),%ymm3 # 57f4 <_sk_callback_avx+0x385>
+ .byte 196,226,125,24,29,193,37,0,0 // vbroadcastss 0x25c1(%rip),%ymm3 # 5f28 <_sk_callback_avx+0x3c5>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,118,37,0,0 // vbroadcastss 0x2576(%rip),%ymm3 # 57f8 <_sk_callback_avx+0x389>
+ .byte 196,226,125,24,29,182,37,0,0 // vbroadcastss 0x25b6(%rip),%ymm3 # 5f2c <_sk_callback_avx+0x3c9>
.byte 91 // pop %rbx
.byte 65,92 // pop %r12
.byte 65,94 // pop %r14
@@ -15241,14 +16885,14 @@ FUNCTION(_sk_store_565_avx)
_sk_store_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,98,37,0,0 // vbroadcastss 0x2562(%rip),%ymm8 # 57fc <_sk_callback_avx+0x38d>
+ .byte 196,98,125,24,5,162,37,0,0 // vbroadcastss 0x25a2(%rip),%ymm8 # 5f30 <_sk_callback_avx+0x3cd>
.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 196,98,125,24,21,59,37,0,0 // vbroadcastss 0x253b(%rip),%ymm10 # 5800 <_sk_callback_avx+0x391>
+ .byte 196,98,125,24,21,123,37,0,0 // vbroadcastss 0x257b(%rip),%ymm10 # 5f34 <_sk_callback_avx+0x3d1>
.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
@@ -15262,7 +16906,7 @@ _sk_store_565_avx:
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 3315 <_sk_store_565_avx+0x89>
+ .byte 117,10 // jne 3a09 <_sk_store_565_avx+0x89>
.byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -15270,9 +16914,9 @@ _sk_store_565_avx:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 3311 <_sk_store_565_avx+0x85>
+ .byte 119,236 // ja 3a05 <_sk_store_565_avx+0x85>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 3374 <_sk_store_565_avx+0xe8>
+ .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 3a68 <_sk_store_565_avx+0xe8>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -15283,7 +16927,7 @@ _sk_store_565_avx:
.byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
.byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
.byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- .byte 235,159 // jmp 3311 <_sk_store_565_avx+0x85>
+ .byte 235,159 // jmp 3a05 <_sk_store_565_avx+0x85>
.byte 102,144 // xchg %ax,%ax
.byte 245 // cmc
.byte 255 // (bad)
@@ -15316,31 +16960,31 @@ _sk_load_4444_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,152,0,0,0 // jne 3436 <_sk_load_4444_avx+0xa6>
+ .byte 15,133,152,0,0,0 // jne 3b2a <_sk_load_4444_avx+0xa6>
.byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
.byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
.byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
.byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3
- .byte 196,226,125,24,5,68,36,0,0 // vbroadcastss 0x2444(%rip),%ymm0 # 5804 <_sk_callback_avx+0x395>
+ .byte 196,226,125,24,5,132,36,0,0 // vbroadcastss 0x2484(%rip),%ymm0 # 5f38 <_sk_callback_avx+0x3d5>
.byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,55,36,0,0 // vbroadcastss 0x2437(%rip),%ymm1 # 5808 <_sk_callback_avx+0x399>
+ .byte 196,226,125,24,13,119,36,0,0 // vbroadcastss 0x2477(%rip),%ymm1 # 5f3c <_sk_callback_avx+0x3d9>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,46,36,0,0 // vbroadcastss 0x242e(%rip),%ymm1 # 580c <_sk_callback_avx+0x39d>
+ .byte 196,226,125,24,13,110,36,0,0 // vbroadcastss 0x246e(%rip),%ymm1 # 5f40 <_sk_callback_avx+0x3dd>
.byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,33,36,0,0 // vbroadcastss 0x2421(%rip),%ymm2 # 5810 <_sk_callback_avx+0x3a1>
+ .byte 196,226,125,24,21,97,36,0,0 // vbroadcastss 0x2461(%rip),%ymm2 # 5f44 <_sk_callback_avx+0x3e1>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,24,21,24,36,0,0 // vbroadcastss 0x2418(%rip),%ymm2 # 5814 <_sk_callback_avx+0x3a5>
+ .byte 196,226,125,24,21,88,36,0,0 // vbroadcastss 0x2458(%rip),%ymm2 # 5f48 <_sk_callback_avx+0x3e5>
.byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,11,36,0,0 // vbroadcastss 0x240b(%rip),%ymm8 # 5818 <_sk_callback_avx+0x3a9>
+ .byte 196,98,125,24,5,75,36,0,0 // vbroadcastss 0x244b(%rip),%ymm8 # 5f4c <_sk_callback_avx+0x3e9>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,1,36,0,0 // vbroadcastss 0x2401(%rip),%ymm8 # 581c <_sk_callback_avx+0x3ad>
+ .byte 196,98,125,24,5,65,36,0,0 // vbroadcastss 0x2441(%rip),%ymm8 # 5f50 <_sk_callback_avx+0x3ed>
.byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,243,35,0,0 // vbroadcastss 0x23f3(%rip),%ymm8 # 5820 <_sk_callback_avx+0x3b1>
+ .byte 196,98,125,24,5,51,36,0,0 // vbroadcastss 0x2433(%rip),%ymm8 # 5f54 <_sk_callback_avx+0x3f1>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -15349,9 +16993,9 @@ _sk_load_4444_avx:
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,86,255,255,255 // ja 33a4 <_sk_load_4444_avx+0x14>
+ .byte 15,135,86,255,255,255 // ja 3a98 <_sk_load_4444_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 34a4 <_sk_load_4444_avx+0x114>
+ .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3b98 <_sk_load_4444_avx+0x114>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -15363,7 +17007,7 @@ _sk_load_4444_avx:
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- .byte 233,2,255,255,255 // jmpq 33a4 <_sk_load_4444_avx+0x14>
+ .byte 233,2,255,255,255 // jmpq 3a98 <_sk_load_4444_avx+0x14>
.byte 102,144 // xchg %ax,%ax
.byte 242,255 // repnz (bad)
.byte 255 // (bad)
@@ -15442,25 +17086,25 @@ _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 196,226,125,24,5,138,34,0,0 // vbroadcastss 0x228a(%rip),%ymm0 # 5824 <_sk_callback_avx+0x3b5>
+ .byte 196,226,125,24,5,202,34,0,0 // vbroadcastss 0x22ca(%rip),%ymm0 # 5f58 <_sk_callback_avx+0x3f5>
.byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,125,34,0,0 // vbroadcastss 0x227d(%rip),%ymm1 # 5828 <_sk_callback_avx+0x3b9>
+ .byte 196,226,125,24,13,189,34,0,0 // vbroadcastss 0x22bd(%rip),%ymm1 # 5f5c <_sk_callback_avx+0x3f9>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,116,34,0,0 // vbroadcastss 0x2274(%rip),%ymm1 # 582c <_sk_callback_avx+0x3bd>
+ .byte 196,226,125,24,13,180,34,0,0 // vbroadcastss 0x22b4(%rip),%ymm1 # 5f60 <_sk_callback_avx+0x3fd>
.byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,103,34,0,0 // vbroadcastss 0x2267(%rip),%ymm2 # 5830 <_sk_callback_avx+0x3c1>
+ .byte 196,226,125,24,21,167,34,0,0 // vbroadcastss 0x22a7(%rip),%ymm2 # 5f64 <_sk_callback_avx+0x401>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,24,21,94,34,0,0 // vbroadcastss 0x225e(%rip),%ymm2 # 5834 <_sk_callback_avx+0x3c5>
+ .byte 196,226,125,24,21,158,34,0,0 // vbroadcastss 0x229e(%rip),%ymm2 # 5f68 <_sk_callback_avx+0x405>
.byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,81,34,0,0 // vbroadcastss 0x2251(%rip),%ymm8 # 5838 <_sk_callback_avx+0x3c9>
+ .byte 196,98,125,24,5,145,34,0,0 // vbroadcastss 0x2291(%rip),%ymm8 # 5f6c <_sk_callback_avx+0x409>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,24,5,71,34,0,0 // vbroadcastss 0x2247(%rip),%ymm8 # 583c <_sk_callback_avx+0x3cd>
+ .byte 196,98,125,24,5,135,34,0,0 // vbroadcastss 0x2287(%rip),%ymm8 # 5f70 <_sk_callback_avx+0x40d>
.byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,57,34,0,0 // vbroadcastss 0x2239(%rip),%ymm8 # 5840 <_sk_callback_avx+0x3d1>
+ .byte 196,98,125,24,5,121,34,0,0 // vbroadcastss 0x2279(%rip),%ymm8 # 5f74 <_sk_callback_avx+0x411>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 91 // pop %rbx
@@ -15476,7 +17120,7 @@ FUNCTION(_sk_store_4444_avx)
_sk_store_4444_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,30,34,0,0 // vbroadcastss 0x221e(%rip),%ymm8 # 5844 <_sk_callback_avx+0x3d5>
+ .byte 196,98,125,24,5,94,34,0,0 // vbroadcastss 0x225e(%rip),%ymm8 # 5f78 <_sk_callback_avx+0x415>
.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
@@ -15503,7 +17147,7 @@ _sk_store_4444_avx:
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 36bf <_sk_store_4444_avx+0xa7>
+ .byte 117,10 // jne 3db3 <_sk_store_4444_avx+0xa7>
.byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -15511,9 +17155,9 @@ _sk_store_4444_avx:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 36bb <_sk_store_4444_avx+0xa3>
+ .byte 119,236 // ja 3daf <_sk_store_4444_avx+0xa3>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,66,0,0,0 // lea 0x42(%rip),%r9 # 371c <_sk_store_4444_avx+0x104>
+ .byte 76,141,13,66,0,0,0 // lea 0x42(%rip),%r9 # 3e10 <_sk_store_4444_avx+0x104>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -15524,7 +17168,7 @@ _sk_store_4444_avx:
.byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
.byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
.byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- .byte 235,159 // jmp 36bb <_sk_store_4444_avx+0xa3>
+ .byte 235,159 // jmp 3daf <_sk_store_4444_avx+0xa3>
.byte 247,255 // idiv %edi
.byte 255 // (bad)
.byte 255 // (bad)
@@ -15555,12 +17199,12 @@ _sk_load_8888_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,135,0,0,0 // jne 37cd <_sk_load_8888_avx+0x95>
+ .byte 15,133,135,0,0,0 // jne 3ec1 <_sk_load_8888_avx+0x95>
.byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9
- .byte 197,124,40,21,76,34,0,0 // vmovaps 0x224c(%rip),%ymm10 # 59a0 <_sk_callback_avx+0x531>
+ .byte 197,124,40,21,152,34,0,0 // vmovaps 0x2298(%rip),%ymm10 # 60e0 <_sk_callback_avx+0x57d>
.byte 196,193,52,84,194 // vandps %ymm10,%ymm9,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,226,32,0,0 // vbroadcastss 0x20e2(%rip),%ymm8 # 5848 <_sk_callback_avx+0x3d9>
+ .byte 196,98,125,24,5,34,33,0,0 // vbroadcastss 0x2122(%rip),%ymm8 # 5f7c <_sk_callback_avx+0x419>
.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
@@ -15587,9 +17231,9 @@ _sk_load_8888_avx:
.byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,102,255,255,255 // ja 374c <_sk_load_8888_avx+0x14>
+ .byte 15,135,102,255,255,255 // ja 3e40 <_sk_load_8888_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,139,0,0,0 // lea 0x8b(%rip),%r9 # 387c <_sk_load_8888_avx+0x144>
+ .byte 76,141,13,139,0,0,0 // lea 0x8b(%rip),%r9 # 3f70 <_sk_load_8888_avx+0x144>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -15612,7 +17256,7 @@ _sk_load_8888_avx:
.byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
.byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0
.byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
- .byte 233,210,254,255,255 // jmpq 374c <_sk_load_8888_avx+0x14>
+ .byte 233,210,254,255,255 // jmpq 3e40 <_sk_load_8888_avx+0x14>
.byte 102,144 // xchg %ax,%ax
.byte 236 // in (%dx),%al
.byte 255 // (bad)
@@ -15630,7 +17274,7 @@ _sk_load_8888_avx:
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 126,255 // jle 3895 <_sk_load_8888_avx+0x15d>
+ .byte 126,255 // jle 3f89 <_sk_load_8888_avx+0x15d>
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -15675,10 +17319,10 @@ _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 197,124,40,21,118,32,0,0 // vmovaps 0x2076(%rip),%ymm10 # 59c0 <_sk_callback_avx+0x551>
+ .byte 197,124,40,21,194,32,0,0 // vmovaps 0x20c2(%rip),%ymm10 # 6100 <_sk_callback_avx+0x59d>
.byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,13,240,30,0,0 // vbroadcastss 0x1ef0(%rip),%ymm9 # 584c <_sk_callback_avx+0x3dd>
+ .byte 196,98,125,24,13,48,31,0,0 // vbroadcastss 0x1f30(%rip),%ymm9 # 5f80 <_sk_callback_avx+0x41d>
.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
@@ -15710,7 +17354,7 @@ FUNCTION(_sk_store_8888_avx)
_sk_store_8888_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,126,30,0,0 // vbroadcastss 0x1e7e(%rip),%ymm8 # 5850 <_sk_callback_avx+0x3e1>
+ .byte 196,98,125,24,5,190,30,0,0 // vbroadcastss 0x1ebe(%rip),%ymm8 # 5f84 <_sk_callback_avx+0x421>
.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
@@ -15735,7 +17379,7 @@ _sk_store_8888_avx:
.byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8
.byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 3a60 <_sk_store_8888_avx+0x9c>
+ .byte 117,10 // jne 4154 <_sk_store_8888_avx+0x9c>
.byte 196,65,124,17,4,186 // vmovups %ymm8,(%r10,%rdi,4)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -15743,9 +17387,9 @@ _sk_store_8888_avx:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 3a5c <_sk_store_8888_avx+0x98>
+ .byte 119,236 // ja 4150 <_sk_store_8888_avx+0x98>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,85,0,0,0 // lea 0x55(%rip),%r9 # 3ad0 <_sk_store_8888_avx+0x10c>
+ .byte 76,141,13,85,0,0,0 // lea 0x55(%rip),%r9 # 41c4 <_sk_store_8888_avx+0x10c>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -15759,7 +17403,7 @@ _sk_store_8888_avx:
.byte 196,67,121,22,68,186,8,2 // vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4)
.byte 196,67,121,22,68,186,4,1 // vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4)
.byte 196,65,121,126,4,186 // vmovd %xmm8,(%r10,%rdi,4)
- .byte 235,143 // jmp 3a5c <_sk_store_8888_avx+0x98>
+ .byte 235,143 // jmp 4150 <_sk_store_8888_avx+0x98>
.byte 15,31,0 // nopl (%rax)
.byte 245 // cmc
.byte 255 // (bad)
@@ -15797,7 +17441,7 @@ _sk_load_f16_avx:
.byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp)
.byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp)
.byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp)
- .byte 15,133,141,2,0,0 // jne 3da3 <_sk_load_f16_avx+0x2b7>
+ .byte 15,133,141,2,0,0 // jne 4497 <_sk_load_f16_avx+0x2b7>
.byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8
.byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
.byte 197,249,16,76,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm1
@@ -15815,13 +17459,13 @@ _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 196,98,125,24,37,229,28,0,0 // vbroadcastss 0x1ce5(%rip),%ymm12 # 5854 <_sk_callback_avx+0x3e5>
+ .byte 196,98,125,24,37,37,29,0,0 // vbroadcastss 0x1d25(%rip),%ymm12 # 5f88 <_sk_callback_avx+0x425>
.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 196,98,121,24,29,209,28,0,0 // vbroadcastss 0x1cd1(%rip),%xmm11 # 5858 <_sk_callback_avx+0x3e9>
+ .byte 196,98,121,24,29,17,29,0,0 // vbroadcastss 0x1d11(%rip),%xmm11 # 5f8c <_sk_callback_avx+0x429>
.byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3
- .byte 196,98,121,24,45,199,28,0,0 // vbroadcastss 0x1cc7(%rip),%xmm13 # 585c <_sk_callback_avx+0x3ed>
+ .byte 196,98,121,24,45,7,29,0,0 // vbroadcastss 0x1d07(%rip),%xmm13 # 5f90 <_sk_callback_avx+0x42d>
.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
@@ -15835,7 +17479,7 @@ _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 196,226,121,24,29,125,28,0,0 // vbroadcastss 0x1c7d(%rip),%xmm3 # 5860 <_sk_callback_avx+0x3f1>
+ .byte 196,226,121,24,29,189,28,0,0 // vbroadcastss 0x1cbd(%rip),%xmm3 # 5f94 <_sk_callback_avx+0x431>
.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
@@ -15928,29 +17572,29 @@ _sk_load_f16_avx:
.byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,79 // je 3e02 <_sk_load_f16_avx+0x316>
+ .byte 116,79 // je 44f6 <_sk_load_f16_avx+0x316>
.byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,67 // jb 3e02 <_sk_load_f16_avx+0x316>
+ .byte 114,67 // jb 44f6 <_sk_load_f16_avx+0x316>
.byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,68 // je 3e0f <_sk_load_f16_avx+0x323>
+ .byte 116,68 // je 4503 <_sk_load_f16_avx+0x323>
.byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,56 // jb 3e0f <_sk_load_f16_avx+0x323>
+ .byte 114,56 // jb 4503 <_sk_load_f16_avx+0x323>
.byte 197,251,16,76,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm1
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,70,253,255,255 // je 3b2d <_sk_load_f16_avx+0x41>
+ .byte 15,132,70,253,255,255 // je 4221 <_sk_load_f16_avx+0x41>
.byte 197,241,22,76,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,54,253,255,255 // jb 3b2d <_sk_load_f16_avx+0x41>
+ .byte 15,130,54,253,255,255 // jb 4221 <_sk_load_f16_avx+0x41>
.byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9
- .byte 233,43,253,255,255 // jmpq 3b2d <_sk_load_f16_avx+0x41>
+ .byte 233,43,253,255,255 // jmpq 4221 <_sk_load_f16_avx+0x41>
.byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,30,253,255,255 // jmpq 3b2d <_sk_load_f16_avx+0x41>
+ .byte 233,30,253,255,255 // jmpq 4221 <_sk_load_f16_avx+0x41>
.byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1
- .byte 233,21,253,255,255 // jmpq 3b2d <_sk_load_f16_avx+0x41>
+ .byte 233,21,253,255,255 // jmpq 4221 <_sk_load_f16_avx+0x41>
HIDDEN _sk_gather_f16_avx
.globl _sk_gather_f16_avx
@@ -16014,13 +17658,13 @@ _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 196,98,125,24,37,65,25,0,0 // vbroadcastss 0x1941(%rip),%ymm12 # 5864 <_sk_callback_avx+0x3f5>
+ .byte 196,98,125,24,37,129,25,0,0 // vbroadcastss 0x1981(%rip),%ymm12 # 5f98 <_sk_callback_avx+0x435>
.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 196,98,121,24,29,45,25,0,0 // vbroadcastss 0x192d(%rip),%xmm11 # 5868 <_sk_callback_avx+0x3f9>
+ .byte 196,98,121,24,29,109,25,0,0 // vbroadcastss 0x196d(%rip),%xmm11 # 5f9c <_sk_callback_avx+0x439>
.byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3
- .byte 196,98,121,24,45,35,25,0,0 // vbroadcastss 0x1923(%rip),%xmm13 # 586c <_sk_callback_avx+0x3fd>
+ .byte 196,98,121,24,45,99,25,0,0 // vbroadcastss 0x1963(%rip),%xmm13 # 5fa0 <_sk_callback_avx+0x43d>
.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
@@ -16034,7 +17678,7 @@ _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 196,226,121,24,29,217,24,0,0 // vbroadcastss 0x18d9(%rip),%xmm3 # 5870 <_sk_callback_avx+0x401>
+ .byte 196,226,121,24,29,25,25,0,0 // vbroadcastss 0x1919(%rip),%xmm3 # 5fa4 <_sk_callback_avx+0x441>
.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
@@ -16138,12 +17782,12 @@ _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 196,98,125,24,13,242,22,0,0 // vbroadcastss 0x16f2(%rip),%ymm9 # 5874 <_sk_callback_avx+0x405>
+ .byte 196,98,125,24,13,50,23,0,0 // vbroadcastss 0x1732(%rip),%ymm9 # 5fa8 <_sk_callback_avx+0x445>
.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 196,98,121,24,5,215,22,0,0 // vbroadcastss 0x16d7(%rip),%xmm8 # 5878 <_sk_callback_avx+0x409>
+ .byte 196,98,121,24,5,23,23,0,0 // vbroadcastss 0x1717(%rip),%xmm8 # 5fac <_sk_callback_avx+0x449>
.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
@@ -16153,7 +17797,7 @@ _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 196,98,125,24,21,158,22,0,0 // vbroadcastss 0x169e(%rip),%ymm10 # 587c <_sk_callback_avx+0x40d>
+ .byte 196,98,125,24,21,222,22,0,0 // vbroadcastss 0x16de(%rip),%ymm10 # 5fb0 <_sk_callback_avx+0x44d>
.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
@@ -16235,7 +17879,7 @@ _sk_store_f16_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,66 // jne 43bc <_sk_store_f16_avx+0x25e>
+ .byte 117,66 // jne 4ab0 <_sk_store_f16_avx+0x25e>
.byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8)
.byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8)
.byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -16251,22 +17895,22 @@ _sk_store_f16_avx:
.byte 255,224 // jmpq *%rax
.byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,202 // je 4391 <_sk_store_f16_avx+0x233>
+ .byte 116,202 // je 4a85 <_sk_store_f16_avx+0x233>
.byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,190 // jb 4391 <_sk_store_f16_avx+0x233>
+ .byte 114,190 // jb 4a85 <_sk_store_f16_avx+0x233>
.byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
- .byte 116,182 // je 4391 <_sk_store_f16_avx+0x233>
+ .byte 116,182 // je 4a85 <_sk_store_f16_avx+0x233>
.byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,170 // jb 4391 <_sk_store_f16_avx+0x233>
+ .byte 114,170 // jb 4a85 <_sk_store_f16_avx+0x233>
.byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
- .byte 116,162 // je 4391 <_sk_store_f16_avx+0x233>
+ .byte 116,162 // je 4a85 <_sk_store_f16_avx+0x233>
.byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,150 // jb 4391 <_sk_store_f16_avx+0x233>
+ .byte 114,150 // jb 4a85 <_sk_store_f16_avx+0x233>
.byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
- .byte 235,142 // jmp 4391 <_sk_store_f16_avx+0x233>
+ .byte 235,142 // jmp 4a85 <_sk_store_f16_avx+0x233>
HIDDEN _sk_load_u16_be_avx
.globl _sk_load_u16_be_avx
@@ -16276,7 +17920,7 @@ _sk_load_u16_be_avx:
.byte 76,139,0 // mov (%rax),%r8
.byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,253,0,0,0 // jne 4516 <_sk_load_u16_be_avx+0x113>
+ .byte 15,133,253,0,0,0 // jne 4c0a <_sk_load_u16_be_avx+0x113>
.byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8
.byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2
.byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3
@@ -16298,7 +17942,7 @@ _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 196,98,125,24,29,246,19,0,0 // vbroadcastss 0x13f6(%rip),%ymm11 # 5880 <_sk_callback_avx+0x411>
+ .byte 196,98,125,24,29,54,20,0,0 // vbroadcastss 0x1436(%rip),%ymm11 # 5fb4 <_sk_callback_avx+0x451>
.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
@@ -16332,29 +17976,29 @@ _sk_load_u16_be_avx:
.byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,85 // je 457c <_sk_load_u16_be_avx+0x179>
+ .byte 116,85 // je 4c70 <_sk_load_u16_be_avx+0x179>
.byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,72 // jb 457c <_sk_load_u16_be_avx+0x179>
+ .byte 114,72 // jb 4c70 <_sk_load_u16_be_avx+0x179>
.byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,72 // je 4589 <_sk_load_u16_be_avx+0x186>
+ .byte 116,72 // je 4c7d <_sk_load_u16_be_avx+0x186>
.byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,59 // jb 4589 <_sk_load_u16_be_avx+0x186>
+ .byte 114,59 // jb 4c7d <_sk_load_u16_be_avx+0x186>
.byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,213,254,255,255 // je 4434 <_sk_load_u16_be_avx+0x31>
+ .byte 15,132,213,254,255,255 // je 4b28 <_sk_load_u16_be_avx+0x31>
.byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,196,254,255,255 // jb 4434 <_sk_load_u16_be_avx+0x31>
+ .byte 15,130,196,254,255,255 // jb 4b28 <_sk_load_u16_be_avx+0x31>
.byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9
- .byte 233,184,254,255,255 // jmpq 4434 <_sk_load_u16_be_avx+0x31>
+ .byte 233,184,254,255,255 // jmpq 4b28 <_sk_load_u16_be_avx+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,171,254,255,255 // jmpq 4434 <_sk_load_u16_be_avx+0x31>
+ .byte 233,171,254,255,255 // jmpq 4b28 <_sk_load_u16_be_avx+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,162,254,255,255 // jmpq 4434 <_sk_load_u16_be_avx+0x31>
+ .byte 233,162,254,255,255 // jmpq 4b28 <_sk_load_u16_be_avx+0x31>
HIDDEN _sk_load_rgb_u16_be_avx
.globl _sk_load_rgb_u16_be_avx
@@ -16364,7 +18008,7 @@ _sk_load_rgb_u16_be_avx:
.byte 76,139,0 // mov (%rax),%r8
.byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,243,0,0,0 // jne 4697 <_sk_load_rgb_u16_be_avx+0x105>
+ .byte 15,133,243,0,0,0 // jne 4d8b <_sk_load_rgb_u16_be_avx+0x105>
.byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0
.byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2
.byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1
@@ -16391,7 +18035,7 @@ _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 196,98,125,24,29,86,18,0,0 // vbroadcastss 0x1256(%rip),%ymm11 # 5884 <_sk_callback_avx+0x415>
+ .byte 196,98,125,24,29,150,18,0,0 // vbroadcastss 0x1296(%rip),%ymm11 # 5fb8 <_sk_callback_avx+0x455>
.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
@@ -16412,41 +18056,41 @@ _sk_load_rgb_u16_be_avx:
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
.byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,243,17,0,0 // vbroadcastss 0x11f3(%rip),%ymm3 # 5888 <_sk_callback_avx+0x419>
+ .byte 196,226,125,24,29,51,18,0,0 // vbroadcastss 0x1233(%rip),%ymm3 # 5fbc <_sk_callback_avx+0x459>
.byte 255,224 // jmpq *%rax
.byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0
.byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 117,5 // jne 46b0 <_sk_load_rgb_u16_be_avx+0x11e>
- .byte 233,40,255,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 117,5 // jne 4da4 <_sk_load_rgb_u16_be_avx+0x11e>
+ .byte 233,40,255,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
.byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1
.byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,26 // jb 46df <_sk_load_rgb_u16_be_avx+0x14d>
+ .byte 114,26 // jb 4dd3 <_sk_load_rgb_u16_be_avx+0x14d>
.byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1
.byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 117,10 // jne 46e4 <_sk_load_rgb_u16_be_avx+0x152>
- .byte 233,249,254,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
- .byte 233,244,254,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 117,10 // jne 4dd8 <_sk_load_rgb_u16_be_avx+0x152>
+ .byte 233,249,254,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,244,254,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
.byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1
.byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,26 // jb 4713 <_sk_load_rgb_u16_be_avx+0x181>
+ .byte 114,26 // jb 4e07 <_sk_load_rgb_u16_be_avx+0x181>
.byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1
.byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 117,10 // jne 4718 <_sk_load_rgb_u16_be_avx+0x186>
- .byte 233,197,254,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
- .byte 233,192,254,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 117,10 // jne 4e0c <_sk_load_rgb_u16_be_avx+0x186>
+ .byte 233,197,254,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,192,254,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
.byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3
.byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,20 // jb 4741 <_sk_load_rgb_u16_be_avx+0x1af>
+ .byte 114,20 // jb 4e35 <_sk_load_rgb_u16_be_avx+0x1af>
.byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3
.byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
- .byte 233,151,254,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
- .byte 233,146,254,255,255 // jmpq 45d8 <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,151,254,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
+ .byte 233,146,254,255,255 // jmpq 4ccc <_sk_load_rgb_u16_be_avx+0x46>
HIDDEN _sk_store_u16_be_avx
.globl _sk_store_u16_be_avx
@@ -16455,7 +18099,7 @@ _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 196,98,125,24,5,48,17,0,0 // vbroadcastss 0x1130(%rip),%ymm8 # 588c <_sk_callback_avx+0x41d>
+ .byte 196,98,125,24,5,112,17,0,0 // vbroadcastss 0x1170(%rip),%ymm8 # 5fc0 <_sk_callback_avx+0x45d>
.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
@@ -16493,7 +18137,7 @@ _sk_store_u16_be_avx:
.byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9
.byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,31 // jne 4840 <_sk_store_u16_be_avx+0xfa>
+ .byte 117,31 // jne 4f34 <_sk_store_u16_be_avx+0xfa>
.byte 196,65,120,17,28,64 // vmovups %xmm11,(%r8,%rax,2)
.byte 196,65,120,17,84,64,16 // vmovups %xmm10,0x10(%r8,%rax,2)
.byte 196,65,120,17,76,64,32 // vmovups %xmm9,0x20(%r8,%rax,2)
@@ -16502,22 +18146,22 @@ _sk_store_u16_be_avx:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,214,28,64 // vmovq %xmm11,(%r8,%rax,2)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 116,240 // je 4f30 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,23,92,64,8 // vmovhpd %xmm11,0x8(%r8,%rax,2)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 114,227 // jb 4f30 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,214,84,64,16 // vmovq %xmm10,0x10(%r8,%rax,2)
- .byte 116,218 // je 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 116,218 // je 4f30 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,23,84,64,24 // vmovhpd %xmm10,0x18(%r8,%rax,2)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 114,205 // jb 4f30 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,214,76,64,32 // vmovq %xmm9,0x20(%r8,%rax,2)
- .byte 116,196 // je 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 116,196 // je 4f30 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,23,76,64,40 // vmovhpd %xmm9,0x28(%r8,%rax,2)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,183 // jb 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 114,183 // jb 4f30 <_sk_store_u16_be_avx+0xf6>
.byte 196,65,121,214,68,64,48 // vmovq %xmm8,0x30(%r8,%rax,2)
- .byte 235,174 // jmp 483c <_sk_store_u16_be_avx+0xf6>
+ .byte 235,174 // jmp 4f30 <_sk_store_u16_be_avx+0xf6>
HIDDEN _sk_load_f32_avx
.globl _sk_load_f32_avx
@@ -16525,10 +18169,10 @@ FUNCTION(_sk_load_f32_avx)
_sk_load_f32_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 119,110 // ja 4904 <_sk_load_f32_avx+0x76>
+ .byte 119,110 // ja 4ff8 <_sk_load_f32_avx+0x76>
.byte 76,139,0 // mov (%rax),%r8
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
- .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 492c <_sk_load_f32_avx+0x9e>
+ .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 5020 <_sk_load_f32_avx+0x9e>
.byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax
.byte 76,1,208 // add %r10,%rax
.byte 255,224 // jmpq *%rax
@@ -16587,7 +18231,7 @@ _sk_store_f32_avx:
.byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
.byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,55 // jne 49b9 <_sk_store_f32_avx+0x6d>
+ .byte 117,55 // jne 50ad <_sk_store_f32_avx+0x6d>
.byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
.byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
.byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
@@ -16600,22 +18244,22 @@ _sk_store_f32_avx:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 49b5 <_sk_store_f32_avx+0x69>
+ .byte 116,240 // je 50a9 <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 49b5 <_sk_store_f32_avx+0x69>
+ .byte 114,227 // jb 50a9 <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
- .byte 116,218 // je 49b5 <_sk_store_f32_avx+0x69>
+ .byte 116,218 // je 50a9 <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 49b5 <_sk_store_f32_avx+0x69>
+ .byte 114,205 // jb 50a9 <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
- .byte 116,195 // je 49b5 <_sk_store_f32_avx+0x69>
+ .byte 116,195 // je 50a9 <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,181 // jb 49b5 <_sk_store_f32_avx+0x69>
+ .byte 114,181 // jb 50a9 <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
- .byte 235,171 // jmp 49b5 <_sk_store_f32_avx+0x69>
+ .byte 235,171 // jmp 50a9 <_sk_store_f32_avx+0x69>
HIDDEN _sk_clamp_x_avx
.globl _sk_clamp_x_avx
@@ -16751,12 +18395,12 @@ HIDDEN _sk_luminance_to_alpha_avx
.globl _sk_luminance_to_alpha_avx
FUNCTION(_sk_luminance_to_alpha_avx)
_sk_luminance_to_alpha_avx:
- .byte 196,226,125,24,29,187,12,0,0 // vbroadcastss 0xcbb(%rip),%ymm3 # 5890 <_sk_callback_avx+0x421>
+ .byte 196,226,125,24,29,251,12,0,0 // vbroadcastss 0xcfb(%rip),%ymm3 # 5fc4 <_sk_callback_avx+0x461>
.byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
- .byte 196,226,125,24,29,178,12,0,0 // vbroadcastss 0xcb2(%rip),%ymm3 # 5894 <_sk_callback_avx+0x425>
+ .byte 196,226,125,24,29,242,12,0,0 // vbroadcastss 0xcf2(%rip),%ymm3 # 5fc8 <_sk_callback_avx+0x465>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
.byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,24,13,165,12,0,0 // vbroadcastss 0xca5(%rip),%ymm1 # 5898 <_sk_callback_avx+0x429>
+ .byte 196,226,125,24,13,229,12,0,0 // vbroadcastss 0xce5(%rip),%ymm1 # 5fcc <_sk_callback_avx+0x469>
.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
@@ -16934,7 +18578,7 @@ _sk_linear_gradient_avx:
.byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3
.byte 76,139,0 // mov (%rax),%r8
.byte 77,133,192 // test %r8,%r8
- .byte 15,132,146,0,0,0 // je 4f49 <_sk_linear_gradient_avx+0xb8>
+ .byte 15,132,146,0,0,0 // je 563d <_sk_linear_gradient_avx+0xb8>
.byte 72,139,64,8 // mov 0x8(%rax),%rax
.byte 72,131,192,32 // add $0x20,%rax
.byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12
@@ -16961,8 +18605,8 @@ _sk_linear_gradient_avx:
.byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3
.byte 72,131,192,36 // add $0x24,%rax
.byte 73,255,200 // dec %r8
- .byte 117,140 // jne 4ed3 <_sk_linear_gradient_avx+0x42>
- .byte 235,20 // jmp 4f5d <_sk_linear_gradient_avx+0xcc>
+ .byte 117,140 // jne 55c7 <_sk_linear_gradient_avx+0x42>
+ .byte 235,20 // jmp 5651 <_sk_linear_gradient_avx+0xcc>
.byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11
.byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
.byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
@@ -17017,27 +18661,27 @@ _sk_xy_to_polar_unit_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 196,98,125,24,45,138,8,0,0 // vbroadcastss 0x88a(%rip),%ymm13 # 589c <_sk_callback_avx+0x42d>
+ .byte 196,98,125,24,45,202,8,0,0 // vbroadcastss 0x8ca(%rip),%ymm13 # 5fd0 <_sk_callback_avx+0x46d>
.byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13
- .byte 196,98,125,24,53,128,8,0,0 // vbroadcastss 0x880(%rip),%ymm14 # 58a0 <_sk_callback_avx+0x431>
+ .byte 196,98,125,24,53,192,8,0,0 // vbroadcastss 0x8c0(%rip),%ymm14 # 5fd4 <_sk_callback_avx+0x471>
.byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
.byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13
- .byte 196,98,125,24,53,113,8,0,0 // vbroadcastss 0x871(%rip),%ymm14 # 58a4 <_sk_callback_avx+0x435>
+ .byte 196,98,125,24,53,177,8,0,0 // vbroadcastss 0x8b1(%rip),%ymm14 # 5fd8 <_sk_callback_avx+0x475>
.byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
.byte 196,65,28,89,229 // vmulps %ymm13,%ymm12,%ymm12
- .byte 196,98,125,24,45,98,8,0,0 // vbroadcastss 0x862(%rip),%ymm13 # 58a8 <_sk_callback_avx+0x439>
+ .byte 196,98,125,24,45,162,8,0,0 // vbroadcastss 0x8a2(%rip),%ymm13 # 5fdc <_sk_callback_avx+0x479>
.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 196,98,125,24,21,77,8,0,0 // vbroadcastss 0x84d(%rip),%ymm10 # 58ac <_sk_callback_avx+0x43d>
+ .byte 196,98,125,24,21,141,8,0,0 // vbroadcastss 0x88d(%rip),%ymm10 # 5fe0 <_sk_callback_avx+0x47d>
.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 196,98,125,24,21,55,8,0,0 // vbroadcastss 0x837(%rip),%ymm10 # 58b0 <_sk_callback_avx+0x441>
+ .byte 196,98,125,24,21,119,8,0,0 // vbroadcastss 0x877(%rip),%ymm10 # 5fe4 <_sk_callback_avx+0x481>
.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 196,98,125,24,21,33,8,0,0 // vbroadcastss 0x821(%rip),%ymm10 # 58b4 <_sk_callback_avx+0x445>
+ .byte 196,98,125,24,21,97,8,0,0 // vbroadcastss 0x861(%rip),%ymm10 # 5fe8 <_sk_callback_avx+0x485>
.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
@@ -17050,7 +18694,7 @@ HIDDEN _sk_save_xy_avx
FUNCTION(_sk_save_xy_avx)
_sk_save_xy_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,0,8,0,0 // vbroadcastss 0x800(%rip),%ymm8 # 58b8 <_sk_callback_avx+0x449>
+ .byte 196,98,125,24,5,64,8,0,0 // vbroadcastss 0x840(%rip),%ymm8 # 5fec <_sk_callback_avx+0x489>
.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
@@ -17087,9 +18731,9 @@ HIDDEN _sk_bilinear_nx_avx
FUNCTION(_sk_bilinear_nx_avx)
_sk_bilinear_nx_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,140,7,0,0 // vbroadcastss 0x78c(%rip),%ymm0 # 58bc <_sk_callback_avx+0x44d>
+ .byte 196,226,125,24,5,204,7,0,0 // vbroadcastss 0x7cc(%rip),%ymm0 # 5ff0 <_sk_callback_avx+0x48d>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,131,7,0,0 // vbroadcastss 0x783(%rip),%ymm8 # 58c0 <_sk_callback_avx+0x451>
+ .byte 196,98,125,24,5,195,7,0,0 // vbroadcastss 0x7c3(%rip),%ymm8 # 5ff4 <_sk_callback_avx+0x491>
.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
@@ -17100,7 +18744,7 @@ HIDDEN _sk_bilinear_px_avx
FUNCTION(_sk_bilinear_px_avx)
_sk_bilinear_px_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,107,7,0,0 // vbroadcastss 0x76b(%rip),%ymm0 # 58c4 <_sk_callback_avx+0x455>
+ .byte 196,226,125,24,5,171,7,0,0 // vbroadcastss 0x7ab(%rip),%ymm0 # 5ff8 <_sk_callback_avx+0x495>
.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)
@@ -17112,9 +18756,9 @@ HIDDEN _sk_bilinear_ny_avx
FUNCTION(_sk_bilinear_ny_avx)
_sk_bilinear_ny_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,79,7,0,0 // vbroadcastss 0x74f(%rip),%ymm1 # 58c8 <_sk_callback_avx+0x459>
+ .byte 196,226,125,24,13,143,7,0,0 // vbroadcastss 0x78f(%rip),%ymm1 # 5ffc <_sk_callback_avx+0x499>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,69,7,0,0 // vbroadcastss 0x745(%rip),%ymm8 # 58cc <_sk_callback_avx+0x45d>
+ .byte 196,98,125,24,5,133,7,0,0 // vbroadcastss 0x785(%rip),%ymm8 # 6000 <_sk_callback_avx+0x49d>
.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
@@ -17125,7 +18769,7 @@ HIDDEN _sk_bilinear_py_avx
FUNCTION(_sk_bilinear_py_avx)
_sk_bilinear_py_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,45,7,0,0 // vbroadcastss 0x72d(%rip),%ymm1 # 58d0 <_sk_callback_avx+0x461>
+ .byte 196,226,125,24,13,109,7,0,0 // vbroadcastss 0x76d(%rip),%ymm1 # 6004 <_sk_callback_avx+0x4a1>
.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)
@@ -17137,14 +18781,14 @@ HIDDEN _sk_bicubic_n3x_avx
FUNCTION(_sk_bicubic_n3x_avx)
_sk_bicubic_n3x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,16,7,0,0 // vbroadcastss 0x710(%rip),%ymm0 # 58d4 <_sk_callback_avx+0x465>
+ .byte 196,226,125,24,5,80,7,0,0 // vbroadcastss 0x750(%rip),%ymm0 # 6008 <_sk_callback_avx+0x4a5>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,7,7,0,0 // vbroadcastss 0x707(%rip),%ymm8 # 58d8 <_sk_callback_avx+0x469>
+ .byte 196,98,125,24,5,71,7,0,0 // vbroadcastss 0x747(%rip),%ymm8 # 600c <_sk_callback_avx+0x4a9>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,248,6,0,0 // vbroadcastss 0x6f8(%rip),%ymm10 # 58dc <_sk_callback_avx+0x46d>
+ .byte 196,98,125,24,21,56,7,0,0 // vbroadcastss 0x738(%rip),%ymm10 # 6010 <_sk_callback_avx+0x4ad>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,238,6,0,0 // vbroadcastss 0x6ee(%rip),%ymm10 # 58e0 <_sk_callback_avx+0x471>
+ .byte 196,98,125,24,21,46,7,0,0 // vbroadcastss 0x72e(%rip),%ymm10 # 6014 <_sk_callback_avx+0x4b1>
.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)
@@ -17156,19 +18800,19 @@ HIDDEN _sk_bicubic_n1x_avx
FUNCTION(_sk_bicubic_n1x_avx)
_sk_bicubic_n1x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,209,6,0,0 // vbroadcastss 0x6d1(%rip),%ymm0 # 58e4 <_sk_callback_avx+0x475>
+ .byte 196,226,125,24,5,17,7,0,0 // vbroadcastss 0x711(%rip),%ymm0 # 6018 <_sk_callback_avx+0x4b5>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,200,6,0,0 // vbroadcastss 0x6c8(%rip),%ymm8 # 58e8 <_sk_callback_avx+0x479>
+ .byte 196,98,125,24,5,8,7,0,0 // vbroadcastss 0x708(%rip),%ymm8 # 601c <_sk_callback_avx+0x4b9>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,190,6,0,0 // vbroadcastss 0x6be(%rip),%ymm9 # 58ec <_sk_callback_avx+0x47d>
+ .byte 196,98,125,24,13,254,6,0,0 // vbroadcastss 0x6fe(%rip),%ymm9 # 6020 <_sk_callback_avx+0x4bd>
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,180,6,0,0 // vbroadcastss 0x6b4(%rip),%ymm10 # 58f0 <_sk_callback_avx+0x481>
+ .byte 196,98,125,24,21,244,6,0,0 // vbroadcastss 0x6f4(%rip),%ymm10 # 6024 <_sk_callback_avx+0x4c1>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,165,6,0,0 // vbroadcastss 0x6a5(%rip),%ymm10 # 58f4 <_sk_callback_avx+0x485>
+ .byte 196,98,125,24,21,229,6,0,0 // vbroadcastss 0x6e5(%rip),%ymm10 # 6028 <_sk_callback_avx+0x4c5>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,150,6,0,0 // vbroadcastss 0x696(%rip),%ymm9 # 58f8 <_sk_callback_avx+0x489>
+ .byte 196,98,125,24,13,214,6,0,0 // vbroadcastss 0x6d6(%rip),%ymm9 # 602c <_sk_callback_avx+0x4c9>
.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
@@ -17179,17 +18823,17 @@ HIDDEN _sk_bicubic_p1x_avx
FUNCTION(_sk_bicubic_p1x_avx)
_sk_bicubic_p1x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,126,6,0,0 // vbroadcastss 0x67e(%rip),%ymm8 # 58fc <_sk_callback_avx+0x48d>
+ .byte 196,98,125,24,5,190,6,0,0 // vbroadcastss 0x6be(%rip),%ymm8 # 6030 <_sk_callback_avx+0x4cd>
.byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
.byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
- .byte 196,98,125,24,21,112,6,0,0 // vbroadcastss 0x670(%rip),%ymm10 # 5900 <_sk_callback_avx+0x491>
+ .byte 196,98,125,24,21,176,6,0,0 // vbroadcastss 0x6b0(%rip),%ymm10 # 6034 <_sk_callback_avx+0x4d1>
.byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
- .byte 196,98,125,24,29,102,6,0,0 // vbroadcastss 0x666(%rip),%ymm11 # 5904 <_sk_callback_avx+0x495>
+ .byte 196,98,125,24,29,166,6,0,0 // vbroadcastss 0x6a6(%rip),%ymm11 # 6038 <_sk_callback_avx+0x4d5>
.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 196,98,125,24,13,77,6,0,0 // vbroadcastss 0x64d(%rip),%ymm9 # 5908 <_sk_callback_avx+0x499>
+ .byte 196,98,125,24,13,141,6,0,0 // vbroadcastss 0x68d(%rip),%ymm9 # 603c <_sk_callback_avx+0x4d9>
.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
@@ -17200,13 +18844,13 @@ HIDDEN _sk_bicubic_p3x_avx
FUNCTION(_sk_bicubic_p3x_avx)
_sk_bicubic_p3x_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,53,6,0,0 // vbroadcastss 0x635(%rip),%ymm0 # 590c <_sk_callback_avx+0x49d>
+ .byte 196,226,125,24,5,117,6,0,0 // vbroadcastss 0x675(%rip),%ymm0 # 6040 <_sk_callback_avx+0x4dd>
.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 196,98,125,24,21,34,6,0,0 // vbroadcastss 0x622(%rip),%ymm10 # 5910 <_sk_callback_avx+0x4a1>
+ .byte 196,98,125,24,21,98,6,0,0 // vbroadcastss 0x662(%rip),%ymm10 # 6044 <_sk_callback_avx+0x4e1>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,24,6,0,0 // vbroadcastss 0x618(%rip),%ymm10 # 5914 <_sk_callback_avx+0x4a5>
+ .byte 196,98,125,24,21,88,6,0,0 // vbroadcastss 0x658(%rip),%ymm10 # 6048 <_sk_callback_avx+0x4e5>
.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)
@@ -17218,14 +18862,14 @@ HIDDEN _sk_bicubic_n3y_avx
FUNCTION(_sk_bicubic_n3y_avx)
_sk_bicubic_n3y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,251,5,0,0 // vbroadcastss 0x5fb(%rip),%ymm1 # 5918 <_sk_callback_avx+0x4a9>
+ .byte 196,226,125,24,13,59,6,0,0 // vbroadcastss 0x63b(%rip),%ymm1 # 604c <_sk_callback_avx+0x4e9>
.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 # 591c <_sk_callback_avx+0x4ad>
+ .byte 196,98,125,24,5,49,6,0,0 // vbroadcastss 0x631(%rip),%ymm8 # 6050 <_sk_callback_avx+0x4ed>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,226,5,0,0 // vbroadcastss 0x5e2(%rip),%ymm10 # 5920 <_sk_callback_avx+0x4b1>
+ .byte 196,98,125,24,21,34,6,0,0 // vbroadcastss 0x622(%rip),%ymm10 # 6054 <_sk_callback_avx+0x4f1>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,216,5,0,0 // vbroadcastss 0x5d8(%rip),%ymm10 # 5924 <_sk_callback_avx+0x4b5>
+ .byte 196,98,125,24,21,24,6,0,0 // vbroadcastss 0x618(%rip),%ymm10 # 6058 <_sk_callback_avx+0x4f5>
.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)
@@ -17237,19 +18881,19 @@ HIDDEN _sk_bicubic_n1y_avx
FUNCTION(_sk_bicubic_n1y_avx)
_sk_bicubic_n1y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,187,5,0,0 // vbroadcastss 0x5bb(%rip),%ymm1 # 5928 <_sk_callback_avx+0x4b9>
+ .byte 196,226,125,24,13,251,5,0,0 // vbroadcastss 0x5fb(%rip),%ymm1 # 605c <_sk_callback_avx+0x4f9>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,177,5,0,0 // vbroadcastss 0x5b1(%rip),%ymm8 # 592c <_sk_callback_avx+0x4bd>
+ .byte 196,98,125,24,5,241,5,0,0 // vbroadcastss 0x5f1(%rip),%ymm8 # 6060 <_sk_callback_avx+0x4fd>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,167,5,0,0 // vbroadcastss 0x5a7(%rip),%ymm9 # 5930 <_sk_callback_avx+0x4c1>
+ .byte 196,98,125,24,13,231,5,0,0 // vbroadcastss 0x5e7(%rip),%ymm9 # 6064 <_sk_callback_avx+0x501>
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,157,5,0,0 // vbroadcastss 0x59d(%rip),%ymm10 # 5934 <_sk_callback_avx+0x4c5>
+ .byte 196,98,125,24,21,221,5,0,0 // vbroadcastss 0x5dd(%rip),%ymm10 # 6068 <_sk_callback_avx+0x505>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
- .byte 196,98,125,24,21,142,5,0,0 // vbroadcastss 0x58e(%rip),%ymm10 # 5938 <_sk_callback_avx+0x4c9>
+ .byte 196,98,125,24,21,206,5,0,0 // vbroadcastss 0x5ce(%rip),%ymm10 # 606c <_sk_callback_avx+0x509>
.byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,127,5,0,0 // vbroadcastss 0x57f(%rip),%ymm9 # 593c <_sk_callback_avx+0x4cd>
+ .byte 196,98,125,24,13,191,5,0,0 // vbroadcastss 0x5bf(%rip),%ymm9 # 6070 <_sk_callback_avx+0x50d>
.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
@@ -17260,17 +18904,17 @@ HIDDEN _sk_bicubic_p1y_avx
FUNCTION(_sk_bicubic_p1y_avx)
_sk_bicubic_p1y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,103,5,0,0 // vbroadcastss 0x567(%rip),%ymm8 # 5940 <_sk_callback_avx+0x4d1>
+ .byte 196,98,125,24,5,167,5,0,0 // vbroadcastss 0x5a7(%rip),%ymm8 # 6074 <_sk_callback_avx+0x511>
.byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
.byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
- .byte 196,98,125,24,21,88,5,0,0 // vbroadcastss 0x558(%rip),%ymm10 # 5944 <_sk_callback_avx+0x4d5>
+ .byte 196,98,125,24,21,152,5,0,0 // vbroadcastss 0x598(%rip),%ymm10 # 6078 <_sk_callback_avx+0x515>
.byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
- .byte 196,98,125,24,29,78,5,0,0 // vbroadcastss 0x54e(%rip),%ymm11 # 5948 <_sk_callback_avx+0x4d9>
+ .byte 196,98,125,24,29,142,5,0,0 // vbroadcastss 0x58e(%rip),%ymm11 # 607c <_sk_callback_avx+0x519>
.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 196,98,125,24,13,53,5,0,0 // vbroadcastss 0x535(%rip),%ymm9 # 594c <_sk_callback_avx+0x4dd>
+ .byte 196,98,125,24,13,117,5,0,0 // vbroadcastss 0x575(%rip),%ymm9 # 6080 <_sk_callback_avx+0x51d>
.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
@@ -17281,13 +18925,13 @@ HIDDEN _sk_bicubic_p3y_avx
FUNCTION(_sk_bicubic_p3y_avx)
_sk_bicubic_p3y_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,29,5,0,0 // vbroadcastss 0x51d(%rip),%ymm1 # 5950 <_sk_callback_avx+0x4e1>
+ .byte 196,226,125,24,13,93,5,0,0 // vbroadcastss 0x55d(%rip),%ymm1 # 6084 <_sk_callback_avx+0x521>
.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 196,98,125,24,21,9,5,0,0 // vbroadcastss 0x509(%rip),%ymm10 # 5954 <_sk_callback_avx+0x4e5>
+ .byte 196,98,125,24,21,73,5,0,0 // vbroadcastss 0x549(%rip),%ymm10 # 6088 <_sk_callback_avx+0x525>
.byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
- .byte 196,98,125,24,21,255,4,0,0 // vbroadcastss 0x4ff(%rip),%ymm10 # 5958 <_sk_callback_avx+0x4e9>
+ .byte 196,98,125,24,21,63,5,0,0 // vbroadcastss 0x53f(%rip),%ymm10 # 608c <_sk_callback_avx+0x529>
.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)
@@ -17408,6 +19052,30 @@ BALIGN4
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,191,0,0,224 // add %al,-0x1fffff41(%rax)
.byte 64,0,0 // 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 5d3d <.literal4+0xb5>
+ .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 5d4d <.literal4+0xc5>
+ .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 5d5d <.literal4+0xd5>
+ .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 5d6d <.literal4+0xe5>
+ .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)
@@ -17455,7 +19123,7 @@ 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 5681 <.literal4+0xed>
+ .byte 224,7 // loopne 5db5 <.literal4+0x12d>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -17469,10 +19137,10 @@ 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 56ac <.literal4+0x118>
+ .byte 127,0 // jg 5de0 <.literal4+0x158>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 5725 <.literal4+0x191>
+ .byte 119,115 // ja 5e59 <.literal4+0x1d1>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -17486,10 +19154,10 @@ 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 56e0 <.literal4+0x14c>
+ .byte 127,0 // jg 5e14 <.literal4+0x18c>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 5759 <.literal4+0x1c5>
+ .byte 119,115 // ja 5e8d <.literal4+0x205>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -17503,10 +19171,10 @@ 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 5714 <.literal4+0x180>
+ .byte 127,0 // jg 5e48 <.literal4+0x1c0>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 578d <.literal4+0x1f9>
+ .byte 119,115 // ja 5ec1 <.literal4+0x239>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -17520,10 +19188,10 @@ 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 5748 <.literal4+0x1b4>
+ .byte 127,0 // jg 5e7c <.literal4+0x1f4>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 57c1 <.literal4+0x22d>
+ .byte 119,115 // ja 5ef5 <.literal4+0x26d>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -17536,7 +19204,7 @@ 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 57bf <.literal4+0x22b>
+ .byte 127,67 // jg 5ef3 <.literal4+0x26b>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -17548,10 +19216,10 @@ BALIGN4
.byte 190,80,128,3,62 // mov $0x3e038050,%esi
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 57df <.literal4+0x24b>
+ .byte 118,63 // jbe 5f13 <.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 57f3 <.literal4+0x25f>
+ .byte 127,67 // jg 5f27 <.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)
@@ -17560,7 +19228,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 57d5 <.literal4+0x241>
+ .byte 224,7 // loopne 5f09 <.literal4+0x281>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -17572,7 +19240,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 57f1 <.literal4+0x25d>
+ .byte 224,7 // loopne 5f25 <.literal4+0x29d>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -17583,7 +19251,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 5846 <.literal4+0x2b2>
+ .byte 124,66 // jl 5f7a <.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)
@@ -17601,9 +19269,9 @@ 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 5889 <.literal4+0x2f5>
+ .byte 112,65 // jo 5fbd <.literal4+0x335>
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 5897 <.literal4+0x303>
+ .byte 127,67 // jg 5fcb <.literal4+0x343>
.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)
@@ -17619,7 +19287,7 @@ BALIGN4
.byte 0,128,55,0,0,128 // add %al,-0x7fffffc9(%rax)
.byte 63 // (bad)
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 58d7 <.literal4+0x343>
+ .byte 127,71 // jg 600b <.literal4+0x383>
.byte 208 // (bad)
.byte 179,89 // mov $0x59,%bl
.byte 62,89 // ds pop %rcx
@@ -17846,7 +19514,7 @@ _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 15,40,21,68,58,0,0 // movaps 0x3a44(%rip),%xmm2 # 3ac0 <_sk_callback_sse41+0xe0>
+ .byte 15,40,21,180,66,0,0 // movaps 0x42b4(%rip),%xmm2 # 4330 <_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
@@ -17855,7 +19523,7 @@ _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 15,40,21,51,58,0,0 // movaps 0x3a33(%rip),%xmm2 # 3ad0 <_sk_callback_sse41+0xf0>
+ .byte 15,40,21,163,66,0,0 // movaps 0x42a3(%rip),%xmm2 # 4340 <_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
@@ -17878,14 +19546,14 @@ _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 102,68,15,111,21,248,57,0,0 // movdqa 0x39f8(%rip),%xmm10 # 3ae0 <_sk_callback_sse41+0x100>
+ .byte 102,68,15,111,21,104,66,0,0 // movdqa 0x4268(%rip),%xmm10 # 4350 <_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 102,68,15,111,37,228,57,0,0 // movdqa 0x39e4(%rip),%xmm12 # 3af0 <_sk_callback_sse41+0x110>
- .byte 102,68,15,111,45,235,57,0,0 // movdqa 0x39eb(%rip),%xmm13 # 3b00 <_sk_callback_sse41+0x120>
+ .byte 102,68,15,111,37,84,66,0,0 // movdqa 0x4254(%rip),%xmm12 # 4360 <_sk_callback_sse41+0x10e>
+ .byte 102,68,15,111,45,91,66,0,0 // movdqa 0x425b(%rip),%xmm13 # 4370 <_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
@@ -17901,8 +19569,8 @@ _sk_dither_sse41:
.byte 102,69,15,235,245 // por %xmm13,%xmm14
.byte 102,69,15,235,240 // por %xmm8,%xmm14
.byte 69,15,91,198 // cvtdq2ps %xmm14,%xmm8
- .byte 68,15,89,5,166,57,0,0 // mulps 0x39a6(%rip),%xmm8 # 3b10 <_sk_callback_sse41+0x130>
- .byte 68,15,88,5,174,57,0,0 // addps 0x39ae(%rip),%xmm8 # 3b20 <_sk_callback_sse41+0x140>
+ .byte 68,15,89,5,22,66,0,0 // mulps 0x4216(%rip),%xmm8 # 4380 <_sk_callback_sse41+0x12e>
+ .byte 68,15,88,5,30,66,0,0 // addps 0x421e(%rip),%xmm8 # 4390 <_sk_callback_sse41+0x13e>
.byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9
.byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
.byte 69,15,89,200 // mulps %xmm8,%xmm9
@@ -17944,7 +19612,7 @@ HIDDEN _sk_srcatop_sse41
FUNCTION(_sk_srcatop_sse41)
_sk_srcatop_sse41:
.byte 15,89,199 // mulps %xmm7,%xmm0
- .byte 68,15,40,5,91,57,0,0 // movaps 0x395b(%rip),%xmm8 # 3b30 <_sk_callback_sse41+0x150>
+ .byte 68,15,40,5,203,65,0,0 // movaps 0x41cb(%rip),%xmm8 # 43a0 <_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
@@ -17969,7 +19637,7 @@ FUNCTION(_sk_dstatop_sse41)
_sk_dstatop_sse41:
.byte 68,15,40,195 // movaps %xmm3,%xmm8
.byte 68,15,89,196 // mulps %xmm4,%xmm8
- .byte 68,15,40,13,30,57,0,0 // movaps 0x391e(%rip),%xmm9 # 3b40 <_sk_callback_sse41+0x160>
+ .byte 68,15,40,13,142,65,0,0 // movaps 0x418e(%rip),%xmm9 # 43b0 <_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
@@ -18016,7 +19684,7 @@ HIDDEN _sk_srcout_sse41
.globl _sk_srcout_sse41
FUNCTION(_sk_srcout_sse41)
_sk_srcout_sse41:
- .byte 68,15,40,5,194,56,0,0 // movaps 0x38c2(%rip),%xmm8 # 3b50 <_sk_callback_sse41+0x170>
+ .byte 68,15,40,5,50,65,0,0 // movaps 0x4132(%rip),%xmm8 # 43c0 <_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
@@ -18029,7 +19697,7 @@ HIDDEN _sk_dstout_sse41
.globl _sk_dstout_sse41
FUNCTION(_sk_dstout_sse41)
_sk_dstout_sse41:
- .byte 68,15,40,5,178,56,0,0 // movaps 0x38b2(%rip),%xmm8 # 3b60 <_sk_callback_sse41+0x180>
+ .byte 68,15,40,5,34,65,0,0 // movaps 0x4122(%rip),%xmm8 # 43d0 <_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
@@ -18046,7 +19714,7 @@ HIDDEN _sk_srcover_sse41
.globl _sk_srcover_sse41
FUNCTION(_sk_srcover_sse41)
_sk_srcover_sse41:
- .byte 68,15,40,5,149,56,0,0 // movaps 0x3895(%rip),%xmm8 # 3b70 <_sk_callback_sse41+0x190>
+ .byte 68,15,40,5,5,65,0,0 // movaps 0x4105(%rip),%xmm8 # 43e0 <_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
@@ -18066,7 +19734,7 @@ HIDDEN _sk_dstover_sse41
.globl _sk_dstover_sse41
FUNCTION(_sk_dstover_sse41)
_sk_dstover_sse41:
- .byte 68,15,40,5,105,56,0,0 // movaps 0x3869(%rip),%xmm8 # 3b80 <_sk_callback_sse41+0x1a0>
+ .byte 68,15,40,5,217,64,0,0 // movaps 0x40d9(%rip),%xmm8 # 43f0 <_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
@@ -18094,7 +19762,7 @@ HIDDEN _sk_multiply_sse41
.globl _sk_multiply_sse41
FUNCTION(_sk_multiply_sse41)
_sk_multiply_sse41:
- .byte 68,15,40,5,61,56,0,0 // movaps 0x383d(%rip),%xmm8 # 3b90 <_sk_callback_sse41+0x1b0>
+ .byte 68,15,40,5,173,64,0,0 // movaps 0x40ad(%rip),%xmm8 # 4400 <_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
@@ -18170,7 +19838,7 @@ HIDDEN _sk_xor__sse41
FUNCTION(_sk_xor__sse41)
_sk_xor__sse41:
.byte 68,15,40,195 // movaps %xmm3,%xmm8
- .byte 15,40,29,110,55,0,0 // movaps 0x376e(%rip),%xmm3 # 3ba0 <_sk_callback_sse41+0x1c0>
+ .byte 15,40,29,222,63,0,0 // movaps 0x3fde(%rip),%xmm3 # 4410 <_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
@@ -18218,7 +19886,7 @@ _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 15,40,21,217,54,0,0 // movaps 0x36d9(%rip),%xmm2 # 3bb0 <_sk_callback_sse41+0x1d0>
+ .byte 15,40,21,73,63,0,0 // movaps 0x3f49(%rip),%xmm2 # 4420 <_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
@@ -18252,7 +19920,7 @@ _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 15,40,21,126,54,0,0 // movaps 0x367e(%rip),%xmm2 # 3bc0 <_sk_callback_sse41+0x1e0>
+ .byte 15,40,21,238,62,0,0 // movaps 0x3eee(%rip),%xmm2 # 4430 <_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
@@ -18289,7 +19957,7 @@ _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 15,40,21,24,54,0,0 // movaps 0x3618(%rip),%xmm2 # 3bd0 <_sk_callback_sse41+0x1f0>
+ .byte 15,40,21,136,62,0,0 // movaps 0x3e88(%rip),%xmm2 # 4440 <_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
@@ -18316,7 +19984,7 @@ _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 15,40,13,217,53,0,0 // movaps 0x35d9(%rip),%xmm1 # 3be0 <_sk_callback_sse41+0x200>
+ .byte 15,40,13,73,62,0,0 // movaps 0x3e49(%rip),%xmm1 # 4450 <_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
@@ -18330,7 +19998,7 @@ HIDDEN _sk_colorburn_sse41
FUNCTION(_sk_colorburn_sse41)
_sk_colorburn_sse41:
.byte 68,15,40,192 // movaps %xmm0,%xmm8
- .byte 68,15,40,21,200,53,0,0 // movaps 0x35c8(%rip),%xmm10 # 3bf0 <_sk_callback_sse41+0x210>
+ .byte 68,15,40,21,56,62,0,0 // movaps 0x3e38(%rip),%xmm10 # 4460 <_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
@@ -18412,7 +20080,7 @@ HIDDEN _sk_colordodge_sse41
FUNCTION(_sk_colordodge_sse41)
_sk_colordodge_sse41:
.byte 68,15,40,192 // movaps %xmm0,%xmm8
- .byte 68,15,40,21,166,52,0,0 // movaps 0x34a6(%rip),%xmm10 # 3c00 <_sk_callback_sse41+0x220>
+ .byte 68,15,40,21,22,61,0,0 // movaps 0x3d16(%rip),%xmm10 # 4470 <_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
@@ -18494,7 +20162,7 @@ _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 68,15,40,21,127,51,0,0 // movaps 0x337f(%rip),%xmm10 # 3c10 <_sk_callback_sse41+0x230>
+ .byte 68,15,40,21,239,59,0,0 // movaps 0x3bef(%rip),%xmm10 # 4480 <_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
@@ -18577,7 +20245,7 @@ FUNCTION(_sk_overlay_sse41)
_sk_overlay_sse41:
.byte 68,15,40,201 // movaps %xmm1,%xmm9
.byte 68,15,40,240 // movaps %xmm0,%xmm14
- .byte 68,15,40,21,100,50,0,0 // movaps 0x3264(%rip),%xmm10 # 3c20 <_sk_callback_sse41+0x240>
+ .byte 68,15,40,21,212,58,0,0 // movaps 0x3ad4(%rip),%xmm10 # 4490 <_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
@@ -18662,7 +20330,7 @@ _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 15,40,13,59,49,0,0 // movaps 0x313b(%rip),%xmm1 # 3c30 <_sk_callback_sse41+0x250>
+ .byte 15,40,13,171,57,0,0 // movaps 0x39ab(%rip),%xmm1 # 44a0 <_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
@@ -18675,10 +20343,10 @@ _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 68,15,40,45,25,49,0,0 // movaps 0x3119(%rip),%xmm13 # 3c40 <_sk_callback_sse41+0x260>
+ .byte 68,15,40,45,137,57,0,0 // movaps 0x3989(%rip),%xmm13 # 44b0 <_sk_callback_sse41+0x25e>
.byte 69,15,88,245 // addps %xmm13,%xmm14
.byte 68,15,89,242 // mulps %xmm2,%xmm14
- .byte 68,15,40,37,25,49,0,0 // movaps 0x3119(%rip),%xmm12 # 3c50 <_sk_callback_sse41+0x270>
+ .byte 68,15,40,37,137,57,0,0 // movaps 0x3989(%rip),%xmm12 # 44c0 <_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
@@ -18810,6 +20478,567 @@ _sk_softlight_sse41:
.byte 65,15,40,209 // movaps %xmm9,%xmm2
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_hue_sse41
+.globl _sk_hue_sse41
+FUNCTION(_sk_hue_sse41)
+_sk_hue_sse41:
+ .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp)
+ .byte 68,15,40,214 // movaps %xmm6,%xmm10
+ .byte 68,15,40,229 // movaps %xmm5,%xmm12
+ .byte 15,40,252 // movaps %xmm4,%xmm7
+ .byte 15,40,243 // movaps %xmm3,%xmm6
+ .byte 68,15,40,192 // movaps %xmm0,%xmm8
+ .byte 69,15,87,246 // xorps %xmm14,%xmm14
+ .byte 68,15,40,45,152,55,0,0 // movaps 0x3798(%rip),%xmm13 # 44d0 <_sk_callback_sse41+0x27e>
+ .byte 65,15,40,221 // movaps %xmm13,%xmm3
+ .byte 15,94,222 // divps %xmm6,%xmm3
+ .byte 15,40,198 // movaps %xmm6,%xmm0
+ .byte 65,15,194,198,4 // cmpneqps %xmm14,%xmm0
+ .byte 15,84,195 // andps %xmm3,%xmm0
+ .byte 68,15,89,192 // mulps %xmm0,%xmm8
+ .byte 15,89,200 // mulps %xmm0,%xmm1
+ .byte 15,89,194 // mulps %xmm2,%xmm0
+ .byte 69,15,40,204 // movaps %xmm12,%xmm9
+ .byte 69,15,40,220 // movaps %xmm12,%xmm11
+ .byte 65,15,40,228 // movaps %xmm12,%xmm4
+ .byte 65,15,40,234 // movaps %xmm10,%xmm5
+ .byte 15,95,229 // maxps %xmm5,%xmm4
+ .byte 15,40,215 // movaps %xmm7,%xmm2
+ .byte 68,15,40,215 // movaps %xmm7,%xmm10
+ .byte 15,40,223 // movaps %xmm7,%xmm3
+ .byte 15,95,220 // maxps %xmm4,%xmm3
+ .byte 68,15,93,205 // minps %xmm5,%xmm9
+ .byte 15,41,108,36,216 // movaps %xmm5,-0x28(%rsp)
+ .byte 65,15,93,209 // minps %xmm9,%xmm2
+ .byte 15,92,218 // subps %xmm2,%xmm3
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 15,93,208 // minps %xmm0,%xmm2
+ .byte 65,15,40,224 // movaps %xmm8,%xmm4
+ .byte 15,93,226 // minps %xmm2,%xmm4
+ .byte 68,15,40,201 // movaps %xmm1,%xmm9
+ .byte 68,15,95,200 // maxps %xmm0,%xmm9
+ .byte 65,15,40,208 // movaps %xmm8,%xmm2
+ .byte 65,15,95,209 // maxps %xmm9,%xmm2
+ .byte 15,92,212 // subps %xmm4,%xmm2
+ .byte 68,15,92,196 // subps %xmm4,%xmm8
+ .byte 68,15,89,195 // mulps %xmm3,%xmm8
+ .byte 68,15,94,194 // divps %xmm2,%xmm8
+ .byte 15,92,204 // subps %xmm4,%xmm1
+ .byte 15,89,203 // mulps %xmm3,%xmm1
+ .byte 15,94,202 // divps %xmm2,%xmm1
+ .byte 15,92,196 // subps %xmm4,%xmm0
+ .byte 15,89,195 // mulps %xmm3,%xmm0
+ .byte 15,94,194 // divps %xmm2,%xmm0
+ .byte 65,15,194,214,4 // cmpneqps %xmm14,%xmm2
+ .byte 68,15,84,194 // andps %xmm2,%xmm8
+ .byte 15,84,202 // andps %xmm2,%xmm1
+ .byte 15,84,194 // andps %xmm2,%xmm0
+ .byte 68,15,40,13,7,55,0,0 // movaps 0x3707(%rip),%xmm9 # 44e0 <_sk_callback_sse41+0x28e>
+ .byte 69,15,89,209 // mulps %xmm9,%xmm10
+ .byte 15,40,29,12,55,0,0 // movaps 0x370c(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0x29e>
+ .byte 68,15,89,219 // mulps %xmm3,%xmm11
+ .byte 69,15,88,218 // addps %xmm10,%xmm11
+ .byte 68,15,40,53,12,55,0,0 // movaps 0x370c(%rip),%xmm14 # 4500 <_sk_callback_sse41+0x2ae>
+ .byte 68,15,40,253 // movaps %xmm5,%xmm15
+ .byte 69,15,89,254 // mulps %xmm14,%xmm15
+ .byte 69,15,88,251 // addps %xmm11,%xmm15
+ .byte 65,15,40,224 // movaps %xmm8,%xmm4
+ .byte 65,15,89,225 // mulps %xmm9,%xmm4
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 15,89,211 // mulps %xmm3,%xmm2
+ .byte 15,88,212 // addps %xmm4,%xmm2
+ .byte 15,40,224 // movaps %xmm0,%xmm4
+ .byte 65,15,89,230 // mulps %xmm14,%xmm4
+ .byte 15,88,226 // addps %xmm2,%xmm4
+ .byte 68,15,92,252 // subps %xmm4,%xmm15
+ .byte 69,15,88,199 // addps %xmm15,%xmm8
+ .byte 65,15,88,207 // addps %xmm15,%xmm1
+ .byte 68,15,88,248 // addps %xmm0,%xmm15
+ .byte 15,40,193 // movaps %xmm1,%xmm0
+ .byte 65,15,93,199 // minps %xmm15,%xmm0
+ .byte 65,15,40,208 // movaps %xmm8,%xmm2
+ .byte 15,93,208 // minps %xmm0,%xmm2
+ .byte 15,40,225 // movaps %xmm1,%xmm4
+ .byte 65,15,95,231 // maxps %xmm15,%xmm4
+ .byte 69,15,89,200 // mulps %xmm8,%xmm9
+ .byte 15,89,217 // mulps %xmm1,%xmm3
+ .byte 65,15,88,217 // addps %xmm9,%xmm3
+ .byte 69,15,89,247 // mulps %xmm15,%xmm14
+ .byte 68,15,88,243 // addps %xmm3,%xmm14
+ .byte 15,87,237 // xorps %xmm5,%xmm5
+ .byte 15,194,234,2 // cmpleps %xmm2,%xmm5
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 69,15,92,206 // subps %xmm14,%xmm9
+ .byte 69,15,89,206 // mulps %xmm14,%xmm9
+ .byte 65,15,40,222 // movaps %xmm14,%xmm3
+ .byte 15,92,218 // subps %xmm2,%xmm3
+ .byte 68,15,94,203 // divps %xmm3,%xmm9
+ .byte 69,15,88,206 // addps %xmm14,%xmm9
+ .byte 15,40,197 // movaps %xmm5,%xmm0
+ .byte 102,69,15,56,20,200 // blendvps %xmm0,%xmm8,%xmm9
+ .byte 68,15,95,196 // maxps %xmm4,%xmm8
+ .byte 69,15,40,213 // movaps %xmm13,%xmm10
+ .byte 69,15,194,208,1 // cmpltps %xmm8,%xmm10
+ .byte 65,15,40,209 // movaps %xmm9,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 69,15,92,238 // subps %xmm14,%xmm13
+ .byte 65,15,89,213 // mulps %xmm13,%xmm2
+ .byte 69,15,92,198 // subps %xmm14,%xmm8
+ .byte 65,15,94,208 // divps %xmm8,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 102,68,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm9
+ .byte 68,15,40,217 // movaps %xmm1,%xmm11
+ .byte 69,15,92,222 // subps %xmm14,%xmm11
+ .byte 69,15,89,222 // mulps %xmm14,%xmm11
+ .byte 68,15,94,219 // divps %xmm3,%xmm11
+ .byte 69,15,88,222 // addps %xmm14,%xmm11
+ .byte 15,40,197 // movaps %xmm5,%xmm0
+ .byte 102,68,15,56,20,217 // blendvps %xmm0,%xmm1,%xmm11
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 65,15,89,205 // mulps %xmm13,%xmm1
+ .byte 65,15,94,200 // divps %xmm8,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 102,68,15,56,20,217 // blendvps %xmm0,%xmm1,%xmm11
+ .byte 65,15,40,215 // movaps %xmm15,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 15,94,211 // divps %xmm3,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 15,40,197 // movaps %xmm5,%xmm0
+ .byte 102,65,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm2
+ .byte 15,40,202 // movaps %xmm2,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 65,15,89,205 // mulps %xmm13,%xmm1
+ .byte 65,15,94,200 // divps %xmm8,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 102,15,56,20,209 // blendvps %xmm0,%xmm1,%xmm2
+ .byte 15,87,192 // xorps %xmm0,%xmm0
+ .byte 68,15,95,200 // maxps %xmm0,%xmm9
+ .byte 68,15,95,216 // maxps %xmm0,%xmm11
+ .byte 15,95,208 // maxps %xmm0,%xmm2
+ .byte 68,15,40,198 // movaps %xmm6,%xmm8
+ .byte 68,15,40,84,36,232 // movaps -0x18(%rsp),%xmm10
+ .byte 69,15,88,194 // addps %xmm10,%xmm8
+ .byte 65,15,89,242 // mulps %xmm10,%xmm6
+ .byte 68,15,92,198 // subps %xmm6,%xmm8
+ .byte 69,15,89,200 // mulps %xmm8,%xmm9
+ .byte 69,15,89,216 // mulps %xmm8,%xmm11
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,231 // movaps %xmm7,%xmm4
+ .byte 65,15,40,236 // movaps %xmm12,%xmm5
+ .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6
+ .byte 65,15,40,193 // movaps %xmm9,%xmm0
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,40,216 // movaps %xmm8,%xmm3
+ .byte 65,15,40,250 // movaps %xmm10,%xmm7
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_saturation_sse41
+.globl _sk_saturation_sse41
+FUNCTION(_sk_saturation_sse41)
+_sk_saturation_sse41:
+ .byte 68,15,40,214 // movaps %xmm6,%xmm10
+ .byte 68,15,40,220 // movaps %xmm4,%xmm11
+ .byte 15,40,243 // movaps %xmm3,%xmm6
+ .byte 69,15,87,246 // xorps %xmm14,%xmm14
+ .byte 68,15,40,37,133,53,0,0 // movaps 0x3585(%rip),%xmm12 # 4510 <_sk_callback_sse41+0x2be>
+ .byte 65,15,40,220 // movaps %xmm12,%xmm3
+ .byte 15,94,223 // divps %xmm7,%xmm3
+ .byte 68,15,40,199 // movaps %xmm7,%xmm8
+ .byte 69,15,194,198,4 // cmpneqps %xmm14,%xmm8
+ .byte 68,15,84,195 // andps %xmm3,%xmm8
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 69,15,89,203 // mulps %xmm11,%xmm9
+ .byte 69,15,40,232 // movaps %xmm8,%xmm13
+ .byte 68,15,89,237 // mulps %xmm5,%xmm13
+ .byte 15,41,108,36,232 // movaps %xmm5,-0x18(%rsp)
+ .byte 69,15,89,194 // mulps %xmm10,%xmm8
+ .byte 68,15,41,84,36,216 // movaps %xmm10,-0x28(%rsp)
+ .byte 15,40,225 // movaps %xmm1,%xmm4
+ .byte 15,95,226 // maxps %xmm2,%xmm4
+ .byte 15,40,216 // movaps %xmm0,%xmm3
+ .byte 15,95,220 // maxps %xmm4,%xmm3
+ .byte 15,93,202 // minps %xmm2,%xmm1
+ .byte 15,93,193 // minps %xmm1,%xmm0
+ .byte 15,92,216 // subps %xmm0,%xmm3
+ .byte 65,15,40,197 // movaps %xmm13,%xmm0
+ .byte 65,15,93,192 // minps %xmm8,%xmm0
+ .byte 65,15,40,201 // movaps %xmm9,%xmm1
+ .byte 15,93,200 // minps %xmm0,%xmm1
+ .byte 65,15,40,197 // movaps %xmm13,%xmm0
+ .byte 65,15,95,192 // maxps %xmm8,%xmm0
+ .byte 65,15,40,209 // movaps %xmm9,%xmm2
+ .byte 15,95,208 // maxps %xmm0,%xmm2
+ .byte 15,92,209 // subps %xmm1,%xmm2
+ .byte 68,15,92,201 // subps %xmm1,%xmm9
+ .byte 68,15,89,203 // mulps %xmm3,%xmm9
+ .byte 68,15,94,202 // divps %xmm2,%xmm9
+ .byte 68,15,92,233 // subps %xmm1,%xmm13
+ .byte 68,15,89,235 // mulps %xmm3,%xmm13
+ .byte 68,15,94,234 // divps %xmm2,%xmm13
+ .byte 68,15,92,193 // subps %xmm1,%xmm8
+ .byte 68,15,89,195 // mulps %xmm3,%xmm8
+ .byte 68,15,94,194 // divps %xmm2,%xmm8
+ .byte 65,15,194,214,4 // cmpneqps %xmm14,%xmm2
+ .byte 68,15,84,202 // andps %xmm2,%xmm9
+ .byte 68,15,84,234 // andps %xmm2,%xmm13
+ .byte 68,15,84,194 // andps %xmm2,%xmm8
+ .byte 15,40,13,240,52,0,0 // movaps 0x34f0(%rip),%xmm1 # 4520 <_sk_callback_sse41+0x2ce>
+ .byte 65,15,40,211 // movaps %xmm11,%xmm2
+ .byte 15,89,209 // mulps %xmm1,%xmm2
+ .byte 15,40,5,242,52,0,0 // movaps 0x34f2(%rip),%xmm0 # 4530 <_sk_callback_sse41+0x2de>
+ .byte 15,40,221 // movaps %xmm5,%xmm3
+ .byte 15,89,216 // mulps %xmm0,%xmm3
+ .byte 15,88,218 // addps %xmm2,%xmm3
+ .byte 68,15,40,53,241,52,0,0 // movaps 0x34f1(%rip),%xmm14 # 4540 <_sk_callback_sse41+0x2ee>
+ .byte 69,15,40,250 // movaps %xmm10,%xmm15
+ .byte 69,15,89,254 // mulps %xmm14,%xmm15
+ .byte 68,15,88,251 // addps %xmm3,%xmm15
+ .byte 65,15,40,209 // movaps %xmm9,%xmm2
+ .byte 15,89,209 // mulps %xmm1,%xmm2
+ .byte 65,15,40,221 // movaps %xmm13,%xmm3
+ .byte 15,89,216 // mulps %xmm0,%xmm3
+ .byte 15,88,218 // addps %xmm2,%xmm3
+ .byte 65,15,40,208 // movaps %xmm8,%xmm2
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 15,88,211 // addps %xmm3,%xmm2
+ .byte 68,15,92,250 // subps %xmm2,%xmm15
+ .byte 69,15,88,207 // addps %xmm15,%xmm9
+ .byte 69,15,88,239 // addps %xmm15,%xmm13
+ .byte 69,15,88,248 // addps %xmm8,%xmm15
+ .byte 65,15,40,213 // movaps %xmm13,%xmm2
+ .byte 65,15,93,215 // minps %xmm15,%xmm2
+ .byte 65,15,40,225 // movaps %xmm9,%xmm4
+ .byte 15,93,226 // minps %xmm2,%xmm4
+ .byte 65,15,40,213 // movaps %xmm13,%xmm2
+ .byte 65,15,95,215 // maxps %xmm15,%xmm2
+ .byte 65,15,89,201 // mulps %xmm9,%xmm1
+ .byte 65,15,89,197 // mulps %xmm13,%xmm0
+ .byte 15,88,193 // addps %xmm1,%xmm0
+ .byte 69,15,89,247 // mulps %xmm15,%xmm14
+ .byte 68,15,88,240 // addps %xmm0,%xmm14
+ .byte 15,87,237 // xorps %xmm5,%xmm5
+ .byte 15,194,236,2 // cmpleps %xmm4,%xmm5
+ .byte 69,15,40,209 // movaps %xmm9,%xmm10
+ .byte 69,15,92,214 // subps %xmm14,%xmm10
+ .byte 69,15,89,214 // mulps %xmm14,%xmm10
+ .byte 65,15,40,222 // movaps %xmm14,%xmm3
+ .byte 15,92,220 // subps %xmm4,%xmm3
+ .byte 68,15,94,211 // divps %xmm3,%xmm10
+ .byte 69,15,88,214 // addps %xmm14,%xmm10
+ .byte 15,40,197 // movaps %xmm5,%xmm0
+ .byte 102,69,15,56,20,209 // blendvps %xmm0,%xmm9,%xmm10
+ .byte 68,15,95,202 // maxps %xmm2,%xmm9
+ .byte 69,15,40,196 // movaps %xmm12,%xmm8
+ .byte 69,15,194,193,1 // cmpltps %xmm9,%xmm8
+ .byte 65,15,40,202 // movaps %xmm10,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 69,15,92,230 // subps %xmm14,%xmm12
+ .byte 65,15,89,204 // mulps %xmm12,%xmm1
+ .byte 69,15,92,206 // subps %xmm14,%xmm9
+ .byte 65,15,94,201 // divps %xmm9,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 65,15,40,192 // movaps %xmm8,%xmm0
+ .byte 102,68,15,56,20,209 // blendvps %xmm0,%xmm1,%xmm10
+ .byte 65,15,40,205 // movaps %xmm13,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 65,15,89,206 // mulps %xmm14,%xmm1
+ .byte 15,94,203 // divps %xmm3,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 15,40,197 // movaps %xmm5,%xmm0
+ .byte 102,65,15,56,20,205 // blendvps %xmm0,%xmm13,%xmm1
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 65,15,89,212 // mulps %xmm12,%xmm2
+ .byte 65,15,94,209 // divps %xmm9,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 65,15,40,192 // movaps %xmm8,%xmm0
+ .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1
+ .byte 65,15,40,215 // movaps %xmm15,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 15,94,211 // divps %xmm3,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 15,40,197 // movaps %xmm5,%xmm0
+ .byte 102,65,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm2
+ .byte 15,40,218 // movaps %xmm2,%xmm3
+ .byte 65,15,92,222 // subps %xmm14,%xmm3
+ .byte 65,15,89,220 // mulps %xmm12,%xmm3
+ .byte 65,15,94,217 // divps %xmm9,%xmm3
+ .byte 65,15,88,222 // addps %xmm14,%xmm3
+ .byte 65,15,40,192 // movaps %xmm8,%xmm0
+ .byte 102,15,56,20,211 // blendvps %xmm0,%xmm3,%xmm2
+ .byte 15,87,192 // xorps %xmm0,%xmm0
+ .byte 68,15,95,208 // maxps %xmm0,%xmm10
+ .byte 15,95,200 // maxps %xmm0,%xmm1
+ .byte 15,95,208 // maxps %xmm0,%xmm2
+ .byte 68,15,40,198 // movaps %xmm6,%xmm8
+ .byte 68,15,88,199 // addps %xmm7,%xmm8
+ .byte 15,89,247 // mulps %xmm7,%xmm6
+ .byte 68,15,92,198 // subps %xmm6,%xmm8
+ .byte 69,15,89,208 // mulps %xmm8,%xmm10
+ .byte 65,15,89,200 // mulps %xmm8,%xmm1
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 65,15,40,227 // movaps %xmm11,%xmm4
+ .byte 15,40,108,36,232 // movaps -0x18(%rsp),%xmm5
+ .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 65,15,40,216 // movaps %xmm8,%xmm3
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_color_sse41
+.globl _sk_color_sse41
+FUNCTION(_sk_color_sse41)
+_sk_color_sse41:
+ .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp)
+ .byte 68,15,40,221 // movaps %xmm5,%xmm11
+ .byte 68,15,41,92,36,232 // movaps %xmm11,-0x18(%rsp)
+ .byte 15,40,236 // movaps %xmm4,%xmm5
+ .byte 15,40,227 // movaps %xmm3,%xmm4
+ .byte 68,15,40,192 // movaps %xmm0,%xmm8
+ .byte 69,15,87,201 // xorps %xmm9,%xmm9
+ .byte 68,15,40,45,106,51,0,0 // movaps 0x336a(%rip),%xmm13 # 4550 <_sk_callback_sse41+0x2fe>
+ .byte 65,15,40,197 // movaps %xmm13,%xmm0
+ .byte 15,94,196 // divps %xmm4,%xmm0
+ .byte 65,15,194,217,4 // cmpneqps %xmm9,%xmm3
+ .byte 15,84,216 // andps %xmm0,%xmm3
+ .byte 68,15,89,195 // mulps %xmm3,%xmm8
+ .byte 15,89,203 // mulps %xmm3,%xmm1
+ .byte 15,89,218 // mulps %xmm2,%xmm3
+ .byte 68,15,40,13,89,51,0,0 // movaps 0x3359(%rip),%xmm9 # 4560 <_sk_callback_sse41+0x30e>
+ .byte 15,40,213 // movaps %xmm5,%xmm2
+ .byte 65,15,89,209 // mulps %xmm9,%xmm2
+ .byte 68,15,40,21,90,51,0,0 // movaps 0x335a(%rip),%xmm10 # 4570 <_sk_callback_sse41+0x31e>
+ .byte 69,15,89,218 // mulps %xmm10,%xmm11
+ .byte 68,15,88,218 // addps %xmm2,%xmm11
+ .byte 68,15,40,53,90,51,0,0 // movaps 0x335a(%rip),%xmm14 # 4580 <_sk_callback_sse41+0x32e>
+ .byte 68,15,40,254 // movaps %xmm6,%xmm15
+ .byte 69,15,89,254 // mulps %xmm14,%xmm15
+ .byte 69,15,88,251 // addps %xmm11,%xmm15
+ .byte 65,15,40,192 // movaps %xmm8,%xmm0
+ .byte 65,15,89,193 // mulps %xmm9,%xmm0
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 65,15,89,210 // mulps %xmm10,%xmm2
+ .byte 15,88,208 // addps %xmm0,%xmm2
+ .byte 15,40,195 // movaps %xmm3,%xmm0
+ .byte 65,15,89,198 // mulps %xmm14,%xmm0
+ .byte 15,88,194 // addps %xmm2,%xmm0
+ .byte 68,15,92,248 // subps %xmm0,%xmm15
+ .byte 69,15,88,199 // addps %xmm15,%xmm8
+ .byte 65,15,88,207 // addps %xmm15,%xmm1
+ .byte 68,15,88,251 // addps %xmm3,%xmm15
+ .byte 15,40,193 // movaps %xmm1,%xmm0
+ .byte 65,15,93,199 // minps %xmm15,%xmm0
+ .byte 65,15,40,208 // movaps %xmm8,%xmm2
+ .byte 15,93,208 // minps %xmm0,%xmm2
+ .byte 15,40,241 // movaps %xmm1,%xmm6
+ .byte 65,15,95,247 // maxps %xmm15,%xmm6
+ .byte 69,15,89,200 // mulps %xmm8,%xmm9
+ .byte 68,15,89,209 // mulps %xmm1,%xmm10
+ .byte 69,15,88,209 // addps %xmm9,%xmm10
+ .byte 69,15,89,247 // mulps %xmm15,%xmm14
+ .byte 69,15,88,242 // addps %xmm10,%xmm14
+ .byte 15,87,219 // xorps %xmm3,%xmm3
+ .byte 15,194,218,2 // cmpleps %xmm2,%xmm3
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 69,15,92,206 // subps %xmm14,%xmm9
+ .byte 69,15,89,206 // mulps %xmm14,%xmm9
+ .byte 69,15,40,214 // movaps %xmm14,%xmm10
+ .byte 68,15,92,210 // subps %xmm2,%xmm10
+ .byte 69,15,94,202 // divps %xmm10,%xmm9
+ .byte 69,15,88,206 // addps %xmm14,%xmm9
+ .byte 15,40,195 // movaps %xmm3,%xmm0
+ .byte 102,69,15,56,20,200 // blendvps %xmm0,%xmm8,%xmm9
+ .byte 68,15,95,198 // maxps %xmm6,%xmm8
+ .byte 69,15,40,229 // movaps %xmm13,%xmm12
+ .byte 69,15,194,224,1 // cmpltps %xmm8,%xmm12
+ .byte 65,15,40,209 // movaps %xmm9,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 69,15,92,238 // subps %xmm14,%xmm13
+ .byte 65,15,89,213 // mulps %xmm13,%xmm2
+ .byte 69,15,92,198 // subps %xmm14,%xmm8
+ .byte 65,15,94,208 // divps %xmm8,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 65,15,40,196 // movaps %xmm12,%xmm0
+ .byte 102,68,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm9
+ .byte 68,15,40,217 // movaps %xmm1,%xmm11
+ .byte 69,15,92,222 // subps %xmm14,%xmm11
+ .byte 69,15,89,222 // mulps %xmm14,%xmm11
+ .byte 69,15,94,218 // divps %xmm10,%xmm11
+ .byte 69,15,88,222 // addps %xmm14,%xmm11
+ .byte 15,40,195 // movaps %xmm3,%xmm0
+ .byte 102,68,15,56,20,217 // blendvps %xmm0,%xmm1,%xmm11
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 65,15,89,205 // mulps %xmm13,%xmm1
+ .byte 65,15,94,200 // divps %xmm8,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 65,15,40,196 // movaps %xmm12,%xmm0
+ .byte 102,68,15,56,20,217 // blendvps %xmm0,%xmm1,%xmm11
+ .byte 65,15,40,215 // movaps %xmm15,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 65,15,94,210 // divps %xmm10,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 15,40,195 // movaps %xmm3,%xmm0
+ .byte 102,65,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm2
+ .byte 15,40,202 // movaps %xmm2,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 65,15,89,205 // mulps %xmm13,%xmm1
+ .byte 65,15,94,200 // divps %xmm8,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 65,15,40,196 // movaps %xmm12,%xmm0
+ .byte 102,15,56,20,209 // blendvps %xmm0,%xmm1,%xmm2
+ .byte 15,87,192 // xorps %xmm0,%xmm0
+ .byte 68,15,95,200 // maxps %xmm0,%xmm9
+ .byte 68,15,95,216 // maxps %xmm0,%xmm11
+ .byte 15,95,208 // maxps %xmm0,%xmm2
+ .byte 68,15,40,196 // movaps %xmm4,%xmm8
+ .byte 68,15,88,199 // addps %xmm7,%xmm8
+ .byte 15,89,231 // mulps %xmm7,%xmm4
+ .byte 68,15,92,196 // subps %xmm4,%xmm8
+ .byte 69,15,89,200 // mulps %xmm8,%xmm9
+ .byte 69,15,89,216 // mulps %xmm8,%xmm11
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,229 // movaps %xmm5,%xmm4
+ .byte 15,40,108,36,232 // movaps -0x18(%rsp),%xmm5
+ .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6
+ .byte 65,15,40,193 // movaps %xmm9,%xmm0
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,40,216 // movaps %xmm8,%xmm3
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_luminosity_sse41
+.globl _sk_luminosity_sse41
+FUNCTION(_sk_luminosity_sse41)
+_sk_luminosity_sse41:
+ .byte 68,15,40,206 // movaps %xmm6,%xmm9
+ .byte 68,15,41,76,36,216 // movaps %xmm9,-0x28(%rsp)
+ .byte 68,15,40,245 // movaps %xmm5,%xmm14
+ .byte 68,15,41,116,36,232 // movaps %xmm14,-0x18(%rsp)
+ .byte 15,40,244 // movaps %xmm4,%xmm6
+ .byte 15,40,235 // movaps %xmm3,%xmm5
+ .byte 69,15,87,228 // xorps %xmm12,%xmm12
+ .byte 68,15,40,45,202,49,0,0 // movaps 0x31ca(%rip),%xmm13 # 4590 <_sk_callback_sse41+0x33e>
+ .byte 69,15,40,197 // movaps %xmm13,%xmm8
+ .byte 68,15,94,199 // divps %xmm7,%xmm8
+ .byte 15,40,223 // movaps %xmm7,%xmm3
+ .byte 65,15,194,220,4 // cmpneqps %xmm12,%xmm3
+ .byte 65,15,84,216 // andps %xmm8,%xmm3
+ .byte 68,15,40,211 // movaps %xmm3,%xmm10
+ .byte 68,15,89,214 // mulps %xmm6,%xmm10
+ .byte 68,15,40,219 // movaps %xmm3,%xmm11
+ .byte 69,15,89,222 // mulps %xmm14,%xmm11
+ .byte 65,15,89,217 // mulps %xmm9,%xmm3
+ .byte 68,15,40,5,170,49,0,0 // movaps 0x31aa(%rip),%xmm8 # 45a0 <_sk_callback_sse41+0x34e>
+ .byte 65,15,89,192 // mulps %xmm8,%xmm0
+ .byte 68,15,40,13,174,49,0,0 // movaps 0x31ae(%rip),%xmm9 # 45b0 <_sk_callback_sse41+0x35e>
+ .byte 65,15,89,201 // mulps %xmm9,%xmm1
+ .byte 15,88,200 // addps %xmm0,%xmm1
+ .byte 68,15,40,53,175,49,0,0 // movaps 0x31af(%rip),%xmm14 # 45c0 <_sk_callback_sse41+0x36e>
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 15,88,209 // addps %xmm1,%xmm2
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 65,15,89,192 // mulps %xmm8,%xmm0
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,89,201 // mulps %xmm9,%xmm1
+ .byte 15,88,200 // addps %xmm0,%xmm1
+ .byte 15,40,195 // movaps %xmm3,%xmm0
+ .byte 65,15,89,198 // mulps %xmm14,%xmm0
+ .byte 15,88,193 // addps %xmm1,%xmm0
+ .byte 15,92,208 // subps %xmm0,%xmm2
+ .byte 68,15,88,210 // addps %xmm2,%xmm10
+ .byte 68,15,88,218 // addps %xmm2,%xmm11
+ .byte 15,88,211 // addps %xmm3,%xmm2
+ .byte 65,15,40,195 // movaps %xmm11,%xmm0
+ .byte 15,93,194 // minps %xmm2,%xmm0
+ .byte 65,15,40,202 // movaps %xmm10,%xmm1
+ .byte 15,93,200 // minps %xmm0,%xmm1
+ .byte 65,15,40,219 // movaps %xmm11,%xmm3
+ .byte 15,95,218 // maxps %xmm2,%xmm3
+ .byte 69,15,89,194 // mulps %xmm10,%xmm8
+ .byte 69,15,89,203 // mulps %xmm11,%xmm9
+ .byte 69,15,88,200 // addps %xmm8,%xmm9
+ .byte 68,15,89,242 // mulps %xmm2,%xmm14
+ .byte 69,15,88,241 // addps %xmm9,%xmm14
+ .byte 15,87,228 // xorps %xmm4,%xmm4
+ .byte 15,194,225,2 // cmpleps %xmm1,%xmm4
+ .byte 69,15,40,194 // movaps %xmm10,%xmm8
+ .byte 69,15,92,198 // subps %xmm14,%xmm8
+ .byte 69,15,89,198 // mulps %xmm14,%xmm8
+ .byte 69,15,40,206 // movaps %xmm14,%xmm9
+ .byte 68,15,92,201 // subps %xmm1,%xmm9
+ .byte 69,15,94,193 // divps %xmm9,%xmm8
+ .byte 69,15,88,198 // addps %xmm14,%xmm8
+ .byte 15,40,196 // movaps %xmm4,%xmm0
+ .byte 102,69,15,56,20,194 // blendvps %xmm0,%xmm10,%xmm8
+ .byte 69,15,40,250 // movaps %xmm10,%xmm15
+ .byte 68,15,95,251 // maxps %xmm3,%xmm15
+ .byte 69,15,40,213 // movaps %xmm13,%xmm10
+ .byte 69,15,194,215,1 // cmpltps %xmm15,%xmm10
+ .byte 65,15,40,200 // movaps %xmm8,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 69,15,92,238 // subps %xmm14,%xmm13
+ .byte 65,15,89,205 // mulps %xmm13,%xmm1
+ .byte 69,15,92,254 // subps %xmm14,%xmm15
+ .byte 65,15,94,207 // divps %xmm15,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 102,68,15,56,20,193 // blendvps %xmm0,%xmm1,%xmm8
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,92,206 // subps %xmm14,%xmm1
+ .byte 65,15,89,206 // mulps %xmm14,%xmm1
+ .byte 65,15,94,201 // divps %xmm9,%xmm1
+ .byte 65,15,88,206 // addps %xmm14,%xmm1
+ .byte 15,40,196 // movaps %xmm4,%xmm0
+ .byte 102,65,15,56,20,203 // blendvps %xmm0,%xmm11,%xmm1
+ .byte 15,40,217 // movaps %xmm1,%xmm3
+ .byte 65,15,92,222 // subps %xmm14,%xmm3
+ .byte 65,15,89,221 // mulps %xmm13,%xmm3
+ .byte 65,15,94,223 // divps %xmm15,%xmm3
+ .byte 65,15,88,222 // addps %xmm14,%xmm3
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 102,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm1
+ .byte 68,15,40,218 // movaps %xmm2,%xmm11
+ .byte 69,15,92,222 // subps %xmm14,%xmm11
+ .byte 69,15,89,222 // mulps %xmm14,%xmm11
+ .byte 69,15,94,217 // divps %xmm9,%xmm11
+ .byte 69,15,88,222 // addps %xmm14,%xmm11
+ .byte 15,40,196 // movaps %xmm4,%xmm0
+ .byte 102,68,15,56,20,218 // blendvps %xmm0,%xmm2,%xmm11
+ .byte 65,15,40,211 // movaps %xmm11,%xmm2
+ .byte 65,15,92,214 // subps %xmm14,%xmm2
+ .byte 65,15,89,213 // mulps %xmm13,%xmm2
+ .byte 65,15,94,215 // divps %xmm15,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 65,15,40,194 // movaps %xmm10,%xmm0
+ .byte 102,68,15,56,20,218 // blendvps %xmm0,%xmm2,%xmm11
+ .byte 69,15,95,196 // maxps %xmm12,%xmm8
+ .byte 65,15,95,204 // maxps %xmm12,%xmm1
+ .byte 69,15,95,220 // maxps %xmm12,%xmm11
+ .byte 68,15,40,205 // movaps %xmm5,%xmm9
+ .byte 68,15,88,207 // addps %xmm7,%xmm9
+ .byte 15,89,239 // mulps %xmm7,%xmm5
+ .byte 68,15,92,205 // subps %xmm5,%xmm9
+ .byte 69,15,89,193 // mulps %xmm9,%xmm8
+ .byte 65,15,89,201 // mulps %xmm9,%xmm1
+ .byte 69,15,89,217 // mulps %xmm9,%xmm11
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 65,15,40,192 // movaps %xmm8,%xmm0
+ .byte 65,15,40,211 // movaps %xmm11,%xmm2
+ .byte 65,15,40,217 // movaps %xmm9,%xmm3
+ .byte 15,40,230 // movaps %xmm6,%xmm4
+ .byte 15,40,108,36,232 // movaps -0x18(%rsp),%xmm5
+ .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_clamp_0_sse41
.globl _sk_clamp_0_sse41
FUNCTION(_sk_clamp_0_sse41)
@@ -18826,7 +21055,7 @@ HIDDEN _sk_clamp_1_sse41
.globl _sk_clamp_1_sse41
FUNCTION(_sk_clamp_1_sse41)
_sk_clamp_1_sse41:
- .byte 68,15,40,5,43,47,0,0 // movaps 0x2f2b(%rip),%xmm8 # 3c60 <_sk_callback_sse41+0x280>
+ .byte 68,15,40,5,41,48,0,0 // movaps 0x3029(%rip),%xmm8 # 45d0 <_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
@@ -18838,7 +21067,7 @@ HIDDEN _sk_clamp_a_sse41
.globl _sk_clamp_a_sse41
FUNCTION(_sk_clamp_a_sse41)
_sk_clamp_a_sse41:
- .byte 15,93,29,32,47,0,0 // minps 0x2f20(%rip),%xmm3 # 3c70 <_sk_callback_sse41+0x290>
+ .byte 15,93,29,30,48,0,0 // minps 0x301e(%rip),%xmm3 # 45e0 <_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
@@ -18925,7 +21154,7 @@ HIDDEN _sk_unpremul_sse41
FUNCTION(_sk_unpremul_sse41)
_sk_unpremul_sse41:
.byte 69,15,87,192 // xorps %xmm8,%xmm8
- .byte 68,15,40,13,139,46,0,0 // movaps 0x2e8b(%rip),%xmm9 # 3c80 <_sk_callback_sse41+0x2a0>
+ .byte 68,15,40,13,137,47,0,0 // movaps 0x2f89(%rip),%xmm9 # 45f0 <_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
@@ -18939,20 +21168,20 @@ HIDDEN _sk_from_srgb_sse41
.globl _sk_from_srgb_sse41
FUNCTION(_sk_from_srgb_sse41)
_sk_from_srgb_sse41:
- .byte 68,15,40,29,118,46,0,0 // movaps 0x2e76(%rip),%xmm11 # 3c90 <_sk_callback_sse41+0x2b0>
+ .byte 68,15,40,29,116,47,0,0 // movaps 0x2f74(%rip),%xmm11 # 4600 <_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 68,15,40,37,110,46,0,0 // movaps 0x2e6e(%rip),%xmm12 # 3ca0 <_sk_callback_sse41+0x2c0>
+ .byte 68,15,40,37,108,47,0,0 // movaps 0x2f6c(%rip),%xmm12 # 4610 <_sk_callback_sse41+0x3be>
.byte 68,15,40,192 // movaps %xmm0,%xmm8
.byte 69,15,89,196 // mulps %xmm12,%xmm8
- .byte 68,15,40,45,110,46,0,0 // movaps 0x2e6e(%rip),%xmm13 # 3cb0 <_sk_callback_sse41+0x2d0>
+ .byte 68,15,40,45,108,47,0,0 // movaps 0x2f6c(%rip),%xmm13 # 4620 <_sk_callback_sse41+0x3ce>
.byte 69,15,88,197 // addps %xmm13,%xmm8
.byte 69,15,89,194 // mulps %xmm10,%xmm8
- .byte 68,15,40,53,110,46,0,0 // movaps 0x2e6e(%rip),%xmm14 # 3cc0 <_sk_callback_sse41+0x2e0>
+ .byte 68,15,40,53,108,47,0,0 // movaps 0x2f6c(%rip),%xmm14 # 4630 <_sk_callback_sse41+0x3de>
.byte 69,15,88,198 // addps %xmm14,%xmm8
- .byte 68,15,40,61,114,46,0,0 // movaps 0x2e72(%rip),%xmm15 # 3cd0 <_sk_callback_sse41+0x2f0>
+ .byte 68,15,40,61,112,47,0,0 // movaps 0x2f70(%rip),%xmm15 # 4640 <_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
@@ -18997,20 +21226,20 @@ _sk_to_srgb_sse41:
.byte 68,15,82,192 // rsqrtps %xmm0,%xmm8
.byte 69,15,83,200 // rcpps %xmm8,%xmm9
.byte 69,15,82,208 // rsqrtps %xmm8,%xmm10
- .byte 68,15,40,29,226,45,0,0 // movaps 0x2de2(%rip),%xmm11 # 3ce0 <_sk_callback_sse41+0x300>
+ .byte 68,15,40,29,224,46,0,0 // movaps 0x2ee0(%rip),%xmm11 # 4650 <_sk_callback_sse41+0x3fe>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 65,15,89,203 // mulps %xmm11,%xmm1
- .byte 68,15,40,37,227,45,0,0 // movaps 0x2de3(%rip),%xmm12 # 3cf0 <_sk_callback_sse41+0x310>
+ .byte 68,15,40,37,225,46,0,0 // movaps 0x2ee1(%rip),%xmm12 # 4660 <_sk_callback_sse41+0x40e>
.byte 69,15,89,204 // mulps %xmm12,%xmm9
- .byte 68,15,40,45,231,45,0,0 // movaps 0x2de7(%rip),%xmm13 # 3d00 <_sk_callback_sse41+0x320>
+ .byte 68,15,40,45,229,46,0,0 // movaps 0x2ee5(%rip),%xmm13 # 4670 <_sk_callback_sse41+0x41e>
.byte 69,15,88,205 // addps %xmm13,%xmm9
- .byte 68,15,40,53,235,45,0,0 // movaps 0x2deb(%rip),%xmm14 # 3d10 <_sk_callback_sse41+0x330>
+ .byte 68,15,40,53,233,46,0,0 // movaps 0x2ee9(%rip),%xmm14 # 4680 <_sk_callback_sse41+0x42e>
.byte 69,15,89,214 // mulps %xmm14,%xmm10
.byte 69,15,88,209 // addps %xmm9,%xmm10
- .byte 68,15,40,5,235,45,0,0 // movaps 0x2deb(%rip),%xmm8 # 3d20 <_sk_callback_sse41+0x340>
+ .byte 68,15,40,5,233,46,0,0 // movaps 0x2ee9(%rip),%xmm8 # 4690 <_sk_callback_sse41+0x43e>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,93,202 // minps %xmm10,%xmm9
- .byte 68,15,40,61,235,45,0,0 // movaps 0x2deb(%rip),%xmm15 # 3d30 <_sk_callback_sse41+0x350>
+ .byte 68,15,40,61,233,46,0,0 // movaps 0x2ee9(%rip),%xmm15 # 46a0 <_sk_callback_sse41+0x44e>
.byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0
.byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9
.byte 15,82,194 // rsqrtps %xmm2,%xmm0
@@ -19064,7 +21293,7 @@ _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 68,15,40,53,60,45,0,0 // movaps 0x2d3c(%rip),%xmm14 # 3d40 <_sk_callback_sse41+0x360>
+ .byte 68,15,40,53,58,46,0,0 // movaps 0x2e3a(%rip),%xmm14 # 46b0 <_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
@@ -19073,27 +21302,27 @@ _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 68,15,84,61,35,45,0,0 // andps 0x2d23(%rip),%xmm15 # 3d50 <_sk_callback_sse41+0x370>
+ .byte 68,15,84,61,33,46,0,0 // andps 0x2e21(%rip),%xmm15 # 46c0 <_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 68,15,40,45,22,45,0,0 // movaps 0x2d16(%rip),%xmm13 # 3d60 <_sk_callback_sse41+0x380>
+ .byte 68,15,40,45,20,46,0,0 // movaps 0x2e14(%rip),%xmm13 # 46d0 <_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 68,15,88,5,18,45,0,0 // addps 0x2d12(%rip),%xmm8 # 3d70 <_sk_callback_sse41+0x390>
+ .byte 68,15,88,5,16,46,0,0 // addps 0x2e10(%rip),%xmm8 # 46e0 <_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 68,15,89,5,10,45,0,0 // mulps 0x2d0a(%rip),%xmm8 # 3d80 <_sk_callback_sse41+0x3a0>
+ .byte 68,15,89,5,8,46,0,0 // mulps 0x2e08(%rip),%xmm8 # 46f0 <_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 15,40,5,254,44,0,0 // movaps 0x2cfe(%rip),%xmm0 # 3d90 <_sk_callback_sse41+0x3b0>
+ .byte 15,40,5,252,45,0,0 // movaps 0x2dfc(%rip),%xmm0 # 4700 <_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
@@ -19115,7 +21344,7 @@ _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 68,15,40,13,196,44,0,0 // movaps 0x2cc4(%rip),%xmm9 # 3da0 <_sk_callback_sse41+0x3c0>
+ .byte 68,15,40,13,194,45,0,0 // movaps 0x2dc2(%rip),%xmm9 # 4710 <_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
@@ -19128,19 +21357,19 @@ _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 68,15,40,53,159,44,0,0 // movaps 0x2c9f(%rip),%xmm14 # 3db0 <_sk_callback_sse41+0x3d0>
+ .byte 68,15,40,53,157,45,0,0 // movaps 0x2d9d(%rip),%xmm14 # 4720 <_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 68,15,40,29,152,44,0,0 // movaps 0x2c98(%rip),%xmm11 # 3dc0 <_sk_callback_sse41+0x3e0>
+ .byte 68,15,40,29,150,45,0,0 // movaps 0x2d96(%rip),%xmm11 # 4730 <_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 15,40,61,145,44,0,0 // movaps 0x2c91(%rip),%xmm7 # 3dd0 <_sk_callback_sse41+0x3f0>
+ .byte 15,40,61,143,45,0,0 // movaps 0x2d8f(%rip),%xmm7 # 4740 <_sk_callback_sse41+0x4ee>
.byte 69,15,40,238 // movaps %xmm14,%xmm13
.byte 68,15,89,239 // mulps %xmm7,%xmm13
- .byte 15,40,29,146,44,0,0 // movaps 0x2c92(%rip),%xmm3 # 3de0 <_sk_callback_sse41+0x400>
+ .byte 15,40,29,144,45,0,0 // movaps 0x2d90(%rip),%xmm3 # 4750 <_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
@@ -19150,7 +21379,7 @@ _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 68,15,40,37,113,44,0,0 // movaps 0x2c71(%rip),%xmm12 # 3df0 <_sk_callback_sse41+0x410>
+ .byte 68,15,40,37,111,45,0,0 // movaps 0x2d6f(%rip),%xmm12 # 4760 <_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
@@ -19184,7 +21413,7 @@ _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 68,15,88,21,233,43,0,0 // addps 0x2be9(%rip),%xmm10 # 3e00 <_sk_callback_sse41+0x420>
+ .byte 68,15,88,21,231,44,0,0 // addps 0x2ce7(%rip),%xmm10 # 4770 <_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
@@ -19236,7 +21465,7 @@ _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 68,15,89,5,70,43,0,0 // mulps 0x2b46(%rip),%xmm8 # 3e10 <_sk_callback_sse41+0x430>
+ .byte 68,15,89,5,68,44,0,0 // mulps 0x2c44(%rip),%xmm8 # 4780 <_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
@@ -19274,7 +21503,7 @@ _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 68,15,89,5,242,42,0,0 // mulps 0x2af2(%rip),%xmm8 # 3e20 <_sk_callback_sse41+0x440>
+ .byte 68,15,89,5,240,43,0,0 // mulps 0x2bf0(%rip),%xmm8 # 4790 <_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
@@ -19297,17 +21526,17 @@ _sk_lerp_565_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 102,68,15,56,51,4,120 // pmovzxwd (%rax,%rdi,2),%xmm8
- .byte 102,15,111,29,194,42,0,0 // movdqa 0x2ac2(%rip),%xmm3 # 3e30 <_sk_callback_sse41+0x450>
+ .byte 102,15,111,29,192,43,0,0 // movdqa 0x2bc0(%rip),%xmm3 # 47a0 <_sk_callback_sse41+0x54e>
.byte 102,65,15,219,216 // pand %xmm8,%xmm3
.byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
- .byte 68,15,89,13,193,42,0,0 // mulps 0x2ac1(%rip),%xmm9 # 3e40 <_sk_callback_sse41+0x460>
- .byte 102,15,111,29,201,42,0,0 // movdqa 0x2ac9(%rip),%xmm3 # 3e50 <_sk_callback_sse41+0x470>
+ .byte 68,15,89,13,191,43,0,0 // mulps 0x2bbf(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0x55e>
+ .byte 102,15,111,29,199,43,0,0 // movdqa 0x2bc7(%rip),%xmm3 # 47c0 <_sk_callback_sse41+0x56e>
.byte 102,65,15,219,216 // pand %xmm8,%xmm3
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,202,42,0,0 // mulps 0x2aca(%rip),%xmm3 # 3e60 <_sk_callback_sse41+0x480>
- .byte 102,68,15,219,5,209,42,0,0 // pand 0x2ad1(%rip),%xmm8 # 3e70 <_sk_callback_sse41+0x490>
+ .byte 15,89,29,200,43,0,0 // mulps 0x2bc8(%rip),%xmm3 # 47d0 <_sk_callback_sse41+0x57e>
+ .byte 102,68,15,219,5,207,43,0,0 // pand 0x2bcf(%rip),%xmm8 # 47e0 <_sk_callback_sse41+0x58e>
.byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
- .byte 68,15,89,5,213,42,0,0 // mulps 0x2ad5(%rip),%xmm8 # 3e80 <_sk_callback_sse41+0x4a0>
+ .byte 68,15,89,5,211,43,0,0 // mulps 0x2bd3(%rip),%xmm8 # 47f0 <_sk_callback_sse41+0x59e>
.byte 15,92,196 // subps %xmm4,%xmm0
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -19318,7 +21547,7 @@ _sk_lerp_565_sse41:
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 15,88,214 // addps %xmm6,%xmm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,191,42,0,0 // movaps 0x2abf(%rip),%xmm3 # 3e90 <_sk_callback_sse41+0x4b0>
+ .byte 15,40,29,189,43,0,0 // movaps 0x2bbd(%rip),%xmm3 # 4800 <_sk_callback_sse41+0x5ae>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_load_tables_sse41
@@ -19329,7 +21558,7 @@ _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 102,15,111,5,182,42,0,0 // movdqa 0x2ab6(%rip),%xmm0 # 3ea0 <_sk_callback_sse41+0x4c0>
+ .byte 102,15,111,5,180,43,0,0 // movdqa 0x2bb4(%rip),%xmm0 # 4810 <_sk_callback_sse41+0x5be>
.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
@@ -19344,7 +21573,7 @@ _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 102,15,56,0,13,113,42,0,0 // pshufb 0x2a71(%rip),%xmm1 # 3eb0 <_sk_callback_sse41+0x4d0>
+ .byte 102,15,56,0,13,111,43,0,0 // pshufb 0x2b6f(%rip),%xmm1 # 4820 <_sk_callback_sse41+0x5ce>
.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
@@ -19359,7 +21588,7 @@ _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 102,15,56,0,21,45,42,0,0 // pshufb 0x2a2d(%rip),%xmm2 # 3ec0 <_sk_callback_sse41+0x4e0>
+ .byte 102,15,56,0,21,43,43,0,0 // pshufb 0x2b2b(%rip),%xmm2 # 4830 <_sk_callback_sse41+0x5de>
.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
@@ -19374,7 +21603,7 @@ _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 15,89,29,234,41,0,0 // mulps 0x29ea(%rip),%xmm3 # 3ed0 <_sk_callback_sse41+0x4f0>
+ .byte 15,89,29,232,42,0,0 // mulps 0x2ae8(%rip),%xmm3 # 4840 <_sk_callback_sse41+0x5ee>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -19393,7 +21622,7 @@ _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 102,68,15,111,5,189,41,0,0 // movdqa 0x29bd(%rip),%xmm8 # 3ee0 <_sk_callback_sse41+0x500>
+ .byte 102,68,15,111,5,187,42,0,0 // movdqa 0x2abb(%rip),%xmm8 # 4850 <_sk_callback_sse41+0x5fe>
.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
@@ -19410,7 +21639,7 @@ _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 102,15,56,0,13,112,41,0,0 // pshufb 0x2970(%rip),%xmm1 # 3ef0 <_sk_callback_sse41+0x510>
+ .byte 102,15,56,0,13,110,42,0,0 // pshufb 0x2a6e(%rip),%xmm1 # 4860 <_sk_callback_sse41+0x60e>
.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
@@ -19446,7 +21675,7 @@ _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 15,89,29,190,40,0,0 // mulps 0x28be(%rip),%xmm3 # 3f00 <_sk_callback_sse41+0x520>
+ .byte 15,89,29,188,41,0,0 // mulps 0x29bc(%rip),%xmm3 # 4870 <_sk_callback_sse41+0x61e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -19468,7 +21697,7 @@ _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 102,68,15,111,5,128,40,0,0 // movdqa 0x2880(%rip),%xmm8 # 3f10 <_sk_callback_sse41+0x530>
+ .byte 102,68,15,111,5,126,41,0,0 // movdqa 0x297e(%rip),%xmm8 # 4880 <_sk_callback_sse41+0x62e>
.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
@@ -19485,7 +21714,7 @@ _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 102,15,56,0,13,51,40,0,0 // pshufb 0x2833(%rip),%xmm1 # 3f20 <_sk_callback_sse41+0x540>
+ .byte 102,15,56,0,13,49,41,0,0 // pshufb 0x2931(%rip),%xmm1 # 4890 <_sk_callback_sse41+0x63e>
.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
@@ -19516,7 +21745,7 @@ _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 15,40,29,158,39,0,0 // movaps 0x279e(%rip),%xmm3 # 3f30 <_sk_callback_sse41+0x550>
+ .byte 15,40,29,156,40,0,0 // movaps 0x289c(%rip),%xmm3 # 48a0 <_sk_callback_sse41+0x64e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_byte_tables_sse41
@@ -19526,7 +21755,7 @@ _sk_byte_tables_sse41:
.byte 65,86 // push %r14
.byte 83 // push %rbx
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,159,39,0,0 // movaps 0x279f(%rip),%xmm8 # 3f40 <_sk_callback_sse41+0x560>
+ .byte 68,15,40,5,157,40,0,0 // movaps 0x289d(%rip),%xmm8 # 48b0 <_sk_callback_sse41+0x65e>
.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
@@ -19545,7 +21774,7 @@ _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 68,15,40,13,80,39,0,0 // movaps 0x2750(%rip),%xmm9 # 3f50 <_sk_callback_sse41+0x570>
+ .byte 68,15,40,13,78,40,0,0 // movaps 0x284e(%rip),%xmm9 # 48c0 <_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
@@ -19636,7 +21865,7 @@ _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 68,15,40,13,216,37,0,0 // movaps 0x25d8(%rip),%xmm9 # 3f60 <_sk_callback_sse41+0x580>
+ .byte 68,15,40,13,214,38,0,0 // movaps 0x26d6(%rip),%xmm9 # 48d0 <_sk_callback_sse41+0x67e>
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
@@ -19813,31 +22042,31 @@ _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 68,15,89,5,47,35,0,0 // mulps 0x232f(%rip),%xmm8 # 3f70 <_sk_callback_sse41+0x590>
- .byte 68,15,84,21,55,35,0,0 // andps 0x2337(%rip),%xmm10 # 3f80 <_sk_callback_sse41+0x5a0>
- .byte 68,15,86,21,63,35,0,0 // orps 0x233f(%rip),%xmm10 # 3f90 <_sk_callback_sse41+0x5b0>
- .byte 68,15,88,5,71,35,0,0 // addps 0x2347(%rip),%xmm8 # 3fa0 <_sk_callback_sse41+0x5c0>
- .byte 68,15,40,37,79,35,0,0 // movaps 0x234f(%rip),%xmm12 # 3fb0 <_sk_callback_sse41+0x5d0>
+ .byte 68,15,89,5,45,36,0,0 // mulps 0x242d(%rip),%xmm8 # 48e0 <_sk_callback_sse41+0x68e>
+ .byte 68,15,84,21,53,36,0,0 // andps 0x2435(%rip),%xmm10 # 48f0 <_sk_callback_sse41+0x69e>
+ .byte 68,15,86,21,61,36,0,0 // orps 0x243d(%rip),%xmm10 # 4900 <_sk_callback_sse41+0x6ae>
+ .byte 68,15,88,5,69,36,0,0 // addps 0x2445(%rip),%xmm8 # 4910 <_sk_callback_sse41+0x6be>
+ .byte 68,15,40,37,77,36,0,0 // movaps 0x244d(%rip),%xmm12 # 4920 <_sk_callback_sse41+0x6ce>
.byte 69,15,89,226 // mulps %xmm10,%xmm12
.byte 69,15,92,196 // subps %xmm12,%xmm8
- .byte 68,15,88,21,79,35,0,0 // addps 0x234f(%rip),%xmm10 # 3fc0 <_sk_callback_sse41+0x5e0>
- .byte 68,15,40,37,87,35,0,0 // movaps 0x2357(%rip),%xmm12 # 3fd0 <_sk_callback_sse41+0x5f0>
+ .byte 68,15,88,21,77,36,0,0 // addps 0x244d(%rip),%xmm10 # 4930 <_sk_callback_sse41+0x6de>
+ .byte 68,15,40,37,85,36,0,0 // movaps 0x2455(%rip),%xmm12 # 4940 <_sk_callback_sse41+0x6ee>
.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 68,15,88,5,68,35,0,0 // addps 0x2344(%rip),%xmm8 # 3fe0 <_sk_callback_sse41+0x600>
- .byte 68,15,40,21,76,35,0,0 // movaps 0x234c(%rip),%xmm10 # 3ff0 <_sk_callback_sse41+0x610>
+ .byte 68,15,88,5,66,36,0,0 // addps 0x2442(%rip),%xmm8 # 4950 <_sk_callback_sse41+0x6fe>
+ .byte 68,15,40,21,74,36,0,0 // movaps 0x244a(%rip),%xmm10 # 4960 <_sk_callback_sse41+0x70e>
.byte 69,15,89,211 // mulps %xmm11,%xmm10
.byte 69,15,92,194 // subps %xmm10,%xmm8
- .byte 68,15,40,21,76,35,0,0 // movaps 0x234c(%rip),%xmm10 # 4000 <_sk_callback_sse41+0x620>
+ .byte 68,15,40,21,74,36,0,0 // movaps 0x244a(%rip),%xmm10 # 4970 <_sk_callback_sse41+0x71e>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 68,15,40,29,80,35,0,0 // movaps 0x2350(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x630>
+ .byte 68,15,40,29,78,36,0,0 // movaps 0x244e(%rip),%xmm11 # 4980 <_sk_callback_sse41+0x72e>
.byte 69,15,94,218 // divps %xmm10,%xmm11
.byte 69,15,88,216 // addps %xmm8,%xmm11
- .byte 68,15,89,29,80,35,0,0 // mulps 0x2350(%rip),%xmm11 # 4020 <_sk_callback_sse41+0x640>
+ .byte 68,15,89,29,78,36,0,0 // mulps 0x244e(%rip),%xmm11 # 4990 <_sk_callback_sse41+0x73e>
.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
@@ -19845,7 +22074,7 @@ _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 68,15,93,5,55,35,0,0 // minps 0x2337(%rip),%xmm8 # 4030 <_sk_callback_sse41+0x650>
+ .byte 68,15,93,5,53,36,0,0 // minps 0x2435(%rip),%xmm8 # 49a0 <_sk_callback_sse41+0x74e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19875,31 +22104,31 @@ _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 68,15,89,37,216,34,0,0 // mulps 0x22d8(%rip),%xmm12 # 4040 <_sk_callback_sse41+0x660>
- .byte 68,15,84,29,224,34,0,0 // andps 0x22e0(%rip),%xmm11 # 4050 <_sk_callback_sse41+0x670>
- .byte 68,15,86,29,232,34,0,0 // orps 0x22e8(%rip),%xmm11 # 4060 <_sk_callback_sse41+0x680>
- .byte 68,15,88,37,240,34,0,0 // addps 0x22f0(%rip),%xmm12 # 4070 <_sk_callback_sse41+0x690>
- .byte 15,40,13,249,34,0,0 // movaps 0x22f9(%rip),%xmm1 # 4080 <_sk_callback_sse41+0x6a0>
+ .byte 68,15,89,37,214,35,0,0 // mulps 0x23d6(%rip),%xmm12 # 49b0 <_sk_callback_sse41+0x75e>
+ .byte 68,15,84,29,222,35,0,0 // andps 0x23de(%rip),%xmm11 # 49c0 <_sk_callback_sse41+0x76e>
+ .byte 68,15,86,29,230,35,0,0 // orps 0x23e6(%rip),%xmm11 # 49d0 <_sk_callback_sse41+0x77e>
+ .byte 68,15,88,37,238,35,0,0 // addps 0x23ee(%rip),%xmm12 # 49e0 <_sk_callback_sse41+0x78e>
+ .byte 15,40,13,247,35,0,0 // movaps 0x23f7(%rip),%xmm1 # 49f0 <_sk_callback_sse41+0x79e>
.byte 65,15,89,203 // mulps %xmm11,%xmm1
.byte 68,15,92,225 // subps %xmm1,%xmm12
- .byte 68,15,88,29,249,34,0,0 // addps 0x22f9(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x6b0>
- .byte 15,40,13,2,35,0,0 // movaps 0x2302(%rip),%xmm1 # 40a0 <_sk_callback_sse41+0x6c0>
+ .byte 68,15,88,29,247,35,0,0 // addps 0x23f7(%rip),%xmm11 # 4a00 <_sk_callback_sse41+0x7ae>
+ .byte 15,40,13,0,36,0,0 // movaps 0x2400(%rip),%xmm1 # 4a10 <_sk_callback_sse41+0x7be>
.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 68,15,88,37,239,34,0,0 // addps 0x22ef(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x6d0>
- .byte 15,40,13,248,34,0,0 // movaps 0x22f8(%rip),%xmm1 # 40c0 <_sk_callback_sse41+0x6e0>
+ .byte 68,15,88,37,237,35,0,0 // addps 0x23ed(%rip),%xmm12 # 4a20 <_sk_callback_sse41+0x7ce>
+ .byte 15,40,13,246,35,0,0 // movaps 0x23f6(%rip),%xmm1 # 4a30 <_sk_callback_sse41+0x7de>
.byte 65,15,89,203 // mulps %xmm11,%xmm1
.byte 68,15,92,225 // subps %xmm1,%xmm12
- .byte 68,15,40,21,248,34,0,0 // movaps 0x22f8(%rip),%xmm10 # 40d0 <_sk_callback_sse41+0x6f0>
+ .byte 68,15,40,21,246,35,0,0 // movaps 0x23f6(%rip),%xmm10 # 4a40 <_sk_callback_sse41+0x7ee>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 15,40,13,253,34,0,0 // movaps 0x22fd(%rip),%xmm1 # 40e0 <_sk_callback_sse41+0x700>
+ .byte 15,40,13,251,35,0,0 // movaps 0x23fb(%rip),%xmm1 # 4a50 <_sk_callback_sse41+0x7fe>
.byte 65,15,94,202 // divps %xmm10,%xmm1
.byte 65,15,88,204 // addps %xmm12,%xmm1
- .byte 15,89,13,254,34,0,0 // mulps 0x22fe(%rip),%xmm1 # 40f0 <_sk_callback_sse41+0x710>
+ .byte 15,89,13,252,35,0,0 // mulps 0x23fc(%rip),%xmm1 # 4a60 <_sk_callback_sse41+0x80e>
.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
@@ -19907,7 +22136,7 @@ _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 15,93,13,233,34,0,0 // minps 0x22e9(%rip),%xmm1 # 4100 <_sk_callback_sse41+0x720>
+ .byte 15,93,13,231,35,0,0 // minps 0x23e7(%rip),%xmm1 # 4a70 <_sk_callback_sse41+0x81e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19937,31 +22166,31 @@ _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 68,15,89,37,138,34,0,0 // mulps 0x228a(%rip),%xmm12 # 4110 <_sk_callback_sse41+0x730>
- .byte 68,15,84,29,146,34,0,0 // andps 0x2292(%rip),%xmm11 # 4120 <_sk_callback_sse41+0x740>
- .byte 68,15,86,29,154,34,0,0 // orps 0x229a(%rip),%xmm11 # 4130 <_sk_callback_sse41+0x750>
- .byte 68,15,88,37,162,34,0,0 // addps 0x22a2(%rip),%xmm12 # 4140 <_sk_callback_sse41+0x760>
- .byte 15,40,21,171,34,0,0 // movaps 0x22ab(%rip),%xmm2 # 4150 <_sk_callback_sse41+0x770>
+ .byte 68,15,89,37,136,35,0,0 // mulps 0x2388(%rip),%xmm12 # 4a80 <_sk_callback_sse41+0x82e>
+ .byte 68,15,84,29,144,35,0,0 // andps 0x2390(%rip),%xmm11 # 4a90 <_sk_callback_sse41+0x83e>
+ .byte 68,15,86,29,152,35,0,0 // orps 0x2398(%rip),%xmm11 # 4aa0 <_sk_callback_sse41+0x84e>
+ .byte 68,15,88,37,160,35,0,0 // addps 0x23a0(%rip),%xmm12 # 4ab0 <_sk_callback_sse41+0x85e>
+ .byte 15,40,21,169,35,0,0 // movaps 0x23a9(%rip),%xmm2 # 4ac0 <_sk_callback_sse41+0x86e>
.byte 65,15,89,211 // mulps %xmm11,%xmm2
.byte 68,15,92,226 // subps %xmm2,%xmm12
- .byte 68,15,88,29,171,34,0,0 // addps 0x22ab(%rip),%xmm11 # 4160 <_sk_callback_sse41+0x780>
- .byte 15,40,21,180,34,0,0 // movaps 0x22b4(%rip),%xmm2 # 4170 <_sk_callback_sse41+0x790>
+ .byte 68,15,88,29,169,35,0,0 // addps 0x23a9(%rip),%xmm11 # 4ad0 <_sk_callback_sse41+0x87e>
+ .byte 15,40,21,178,35,0,0 // movaps 0x23b2(%rip),%xmm2 # 4ae0 <_sk_callback_sse41+0x88e>
.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 68,15,88,37,161,34,0,0 // addps 0x22a1(%rip),%xmm12 # 4180 <_sk_callback_sse41+0x7a0>
- .byte 15,40,21,170,34,0,0 // movaps 0x22aa(%rip),%xmm2 # 4190 <_sk_callback_sse41+0x7b0>
+ .byte 68,15,88,37,159,35,0,0 // addps 0x239f(%rip),%xmm12 # 4af0 <_sk_callback_sse41+0x89e>
+ .byte 15,40,21,168,35,0,0 // movaps 0x23a8(%rip),%xmm2 # 4b00 <_sk_callback_sse41+0x8ae>
.byte 65,15,89,211 // mulps %xmm11,%xmm2
.byte 68,15,92,226 // subps %xmm2,%xmm12
- .byte 68,15,40,21,170,34,0,0 // movaps 0x22aa(%rip),%xmm10 # 41a0 <_sk_callback_sse41+0x7c0>
+ .byte 68,15,40,21,168,35,0,0 // movaps 0x23a8(%rip),%xmm10 # 4b10 <_sk_callback_sse41+0x8be>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 15,40,21,175,34,0,0 // movaps 0x22af(%rip),%xmm2 # 41b0 <_sk_callback_sse41+0x7d0>
+ .byte 15,40,21,173,35,0,0 // movaps 0x23ad(%rip),%xmm2 # 4b20 <_sk_callback_sse41+0x8ce>
.byte 65,15,94,210 // divps %xmm10,%xmm2
.byte 65,15,88,212 // addps %xmm12,%xmm2
- .byte 15,89,21,176,34,0,0 // mulps 0x22b0(%rip),%xmm2 # 41c0 <_sk_callback_sse41+0x7e0>
+ .byte 15,89,21,174,35,0,0 // mulps 0x23ae(%rip),%xmm2 # 4b30 <_sk_callback_sse41+0x8de>
.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
@@ -19969,7 +22198,7 @@ _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 15,93,21,155,34,0,0 // minps 0x229b(%rip),%xmm2 # 41d0 <_sk_callback_sse41+0x7f0>
+ .byte 15,93,21,153,35,0,0 // minps 0x2399(%rip),%xmm2 # 4b40 <_sk_callback_sse41+0x8ee>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19999,31 +22228,31 @@ _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 68,15,89,37,60,34,0,0 // mulps 0x223c(%rip),%xmm12 # 41e0 <_sk_callback_sse41+0x800>
- .byte 68,15,84,29,68,34,0,0 // andps 0x2244(%rip),%xmm11 # 41f0 <_sk_callback_sse41+0x810>
- .byte 68,15,86,29,76,34,0,0 // orps 0x224c(%rip),%xmm11 # 4200 <_sk_callback_sse41+0x820>
- .byte 68,15,88,37,84,34,0,0 // addps 0x2254(%rip),%xmm12 # 4210 <_sk_callback_sse41+0x830>
- .byte 15,40,29,93,34,0,0 // movaps 0x225d(%rip),%xmm3 # 4220 <_sk_callback_sse41+0x840>
+ .byte 68,15,89,37,58,35,0,0 // mulps 0x233a(%rip),%xmm12 # 4b50 <_sk_callback_sse41+0x8fe>
+ .byte 68,15,84,29,66,35,0,0 // andps 0x2342(%rip),%xmm11 # 4b60 <_sk_callback_sse41+0x90e>
+ .byte 68,15,86,29,74,35,0,0 // orps 0x234a(%rip),%xmm11 # 4b70 <_sk_callback_sse41+0x91e>
+ .byte 68,15,88,37,82,35,0,0 // addps 0x2352(%rip),%xmm12 # 4b80 <_sk_callback_sse41+0x92e>
+ .byte 15,40,29,91,35,0,0 // movaps 0x235b(%rip),%xmm3 # 4b90 <_sk_callback_sse41+0x93e>
.byte 65,15,89,219 // mulps %xmm11,%xmm3
.byte 68,15,92,227 // subps %xmm3,%xmm12
- .byte 68,15,88,29,93,34,0,0 // addps 0x225d(%rip),%xmm11 # 4230 <_sk_callback_sse41+0x850>
- .byte 15,40,29,102,34,0,0 // movaps 0x2266(%rip),%xmm3 # 4240 <_sk_callback_sse41+0x860>
+ .byte 68,15,88,29,91,35,0,0 // addps 0x235b(%rip),%xmm11 # 4ba0 <_sk_callback_sse41+0x94e>
+ .byte 15,40,29,100,35,0,0 // movaps 0x2364(%rip),%xmm3 # 4bb0 <_sk_callback_sse41+0x95e>
.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 68,15,88,37,83,34,0,0 // addps 0x2253(%rip),%xmm12 # 4250 <_sk_callback_sse41+0x870>
- .byte 15,40,29,92,34,0,0 // movaps 0x225c(%rip),%xmm3 # 4260 <_sk_callback_sse41+0x880>
+ .byte 68,15,88,37,81,35,0,0 // addps 0x2351(%rip),%xmm12 # 4bc0 <_sk_callback_sse41+0x96e>
+ .byte 15,40,29,90,35,0,0 // movaps 0x235a(%rip),%xmm3 # 4bd0 <_sk_callback_sse41+0x97e>
.byte 65,15,89,219 // mulps %xmm11,%xmm3
.byte 68,15,92,227 // subps %xmm3,%xmm12
- .byte 68,15,40,21,92,34,0,0 // movaps 0x225c(%rip),%xmm10 # 4270 <_sk_callback_sse41+0x890>
+ .byte 68,15,40,21,90,35,0,0 // movaps 0x235a(%rip),%xmm10 # 4be0 <_sk_callback_sse41+0x98e>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 15,40,29,97,34,0,0 // movaps 0x2261(%rip),%xmm3 # 4280 <_sk_callback_sse41+0x8a0>
+ .byte 15,40,29,95,35,0,0 // movaps 0x235f(%rip),%xmm3 # 4bf0 <_sk_callback_sse41+0x99e>
.byte 65,15,94,218 // divps %xmm10,%xmm3
.byte 65,15,88,220 // addps %xmm12,%xmm3
- .byte 15,89,29,98,34,0,0 // mulps 0x2262(%rip),%xmm3 # 4290 <_sk_callback_sse41+0x8b0>
+ .byte 15,89,29,96,35,0,0 // mulps 0x2360(%rip),%xmm3 # 4c00 <_sk_callback_sse41+0x9ae>
.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
@@ -20031,7 +22260,7 @@ _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 15,93,29,77,34,0,0 // minps 0x224d(%rip),%xmm3 # 42a0 <_sk_callback_sse41+0x8c0>
+ .byte 15,93,29,75,35,0,0 // minps 0x234b(%rip),%xmm3 # 4c10 <_sk_callback_sse41+0x9be>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -20041,29 +22270,29 @@ 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 68,15,89,5,73,34,0,0 // mulps 0x2249(%rip),%xmm8 # 42b0 <_sk_callback_sse41+0x8d0>
- .byte 68,15,40,13,81,34,0,0 // movaps 0x2251(%rip),%xmm9 # 42c0 <_sk_callback_sse41+0x8e0>
+ .byte 68,15,89,5,71,35,0,0 // mulps 0x2347(%rip),%xmm8 # 4c20 <_sk_callback_sse41+0x9ce>
+ .byte 68,15,40,13,79,35,0,0 // movaps 0x234f(%rip),%xmm9 # 4c30 <_sk_callback_sse41+0x9de>
.byte 65,15,89,201 // mulps %xmm9,%xmm1
- .byte 15,40,5,86,34,0,0 // movaps 0x2256(%rip),%xmm0 # 42d0 <_sk_callback_sse41+0x8f0>
+ .byte 15,40,5,84,35,0,0 // movaps 0x2354(%rip),%xmm0 # 4c40 <_sk_callback_sse41+0x9ee>
.byte 15,88,200 // addps %xmm0,%xmm1
.byte 65,15,89,209 // mulps %xmm9,%xmm2
.byte 15,88,208 // addps %xmm0,%xmm2
- .byte 68,15,88,5,84,34,0,0 // addps 0x2254(%rip),%xmm8 # 42e0 <_sk_callback_sse41+0x900>
- .byte 68,15,89,5,92,34,0,0 // mulps 0x225c(%rip),%xmm8 # 42f0 <_sk_callback_sse41+0x910>
- .byte 15,89,13,101,34,0,0 // mulps 0x2265(%rip),%xmm1 # 4300 <_sk_callback_sse41+0x920>
+ .byte 68,15,88,5,82,35,0,0 // addps 0x2352(%rip),%xmm8 # 4c50 <_sk_callback_sse41+0x9fe>
+ .byte 68,15,89,5,90,35,0,0 // mulps 0x235a(%rip),%xmm8 # 4c60 <_sk_callback_sse41+0xa0e>
+ .byte 15,89,13,99,35,0,0 // mulps 0x2363(%rip),%xmm1 # 4c70 <_sk_callback_sse41+0xa1e>
.byte 65,15,88,200 // addps %xmm8,%xmm1
- .byte 15,89,21,106,34,0,0 // mulps 0x226a(%rip),%xmm2 # 4310 <_sk_callback_sse41+0x930>
+ .byte 15,89,21,104,35,0,0 // mulps 0x2368(%rip),%xmm2 # 4c80 <_sk_callback_sse41+0xa2e>
.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 68,15,40,13,94,34,0,0 // movaps 0x225e(%rip),%xmm9 # 4320 <_sk_callback_sse41+0x940>
+ .byte 68,15,40,13,92,35,0,0 // movaps 0x235c(%rip),%xmm9 # 4c90 <_sk_callback_sse41+0xa3e>
.byte 65,15,40,193 // movaps %xmm9,%xmm0
.byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0
- .byte 15,40,21,94,34,0,0 // movaps 0x225e(%rip),%xmm2 # 4330 <_sk_callback_sse41+0x950>
+ .byte 15,40,21,92,35,0,0 // movaps 0x235c(%rip),%xmm2 # 4ca0 <_sk_callback_sse41+0xa4e>
.byte 15,88,202 // addps %xmm2,%xmm1
- .byte 68,15,40,37,99,34,0,0 // movaps 0x2263(%rip),%xmm12 # 4340 <_sk_callback_sse41+0x960>
+ .byte 68,15,40,37,97,35,0,0 // movaps 0x2361(%rip),%xmm12 # 4cb0 <_sk_callback_sse41+0xa5e>
.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
@@ -20082,8 +22311,8 @@ _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 15,89,13,28,34,0,0 // mulps 0x221c(%rip),%xmm1 # 4350 <_sk_callback_sse41+0x970>
- .byte 15,89,21,37,34,0,0 // mulps 0x2225(%rip),%xmm2 # 4360 <_sk_callback_sse41+0x980>
+ .byte 15,89,13,26,35,0,0 // mulps 0x231a(%rip),%xmm1 # 4cc0 <_sk_callback_sse41+0xa6e>
+ .byte 15,89,21,35,35,0,0 // mulps 0x2323(%rip),%xmm2 # 4cd0 <_sk_callback_sse41+0xa7e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,40,193 // movaps %xmm1,%xmm0
.byte 65,15,40,200 // movaps %xmm8,%xmm1
@@ -20097,7 +22326,7 @@ _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 15,89,29,21,34,0,0 // mulps 0x2215(%rip),%xmm3 # 4370 <_sk_callback_sse41+0x990>
+ .byte 15,89,29,19,35,0,0 // mulps 0x2313(%rip),%xmm3 # 4ce0 <_sk_callback_sse41+0xa8e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
.byte 15,87,201 // xorps %xmm1,%xmm1
@@ -20130,7 +22359,7 @@ _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 15,89,29,169,33,0,0 // mulps 0x21a9(%rip),%xmm3 # 4380 <_sk_callback_sse41+0x9a0>
+ .byte 15,89,29,167,34,0,0 // mulps 0x22a7(%rip),%xmm3 # 4cf0 <_sk_callback_sse41+0xa9e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
.byte 102,15,239,201 // pxor %xmm1,%xmm1
@@ -20143,7 +22372,7 @@ FUNCTION(_sk_store_a8_sse41)
_sk_store_a8_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,157,33,0,0 // movaps 0x219d(%rip),%xmm8 # 4390 <_sk_callback_sse41+0x9b0>
+ .byte 68,15,40,5,155,34,0,0 // movaps 0x229b(%rip),%xmm8 # 4d00 <_sk_callback_sse41+0xaae>
.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
@@ -20160,9 +22389,9 @@ _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 15,89,5,122,33,0,0 // mulps 0x217a(%rip),%xmm0 # 43a0 <_sk_callback_sse41+0x9c0>
+ .byte 15,89,5,120,34,0,0 // mulps 0x2278(%rip),%xmm0 # 4d10 <_sk_callback_sse41+0xabe>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,129,33,0,0 // movaps 0x2181(%rip),%xmm3 # 43b0 <_sk_callback_sse41+0x9d0>
+ .byte 15,40,29,127,34,0,0 // movaps 0x227f(%rip),%xmm3 # 4d20 <_sk_callback_sse41+0xace>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 15,40,208 // movaps %xmm0,%xmm2
.byte 255,224 // jmpq *%rax
@@ -20193,9 +22422,9 @@ _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 15,89,5,26,33,0,0 // mulps 0x211a(%rip),%xmm0 # 43c0 <_sk_callback_sse41+0x9e0>
+ .byte 15,89,5,24,34,0,0 // mulps 0x2218(%rip),%xmm0 # 4d30 <_sk_callback_sse41+0xade>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,33,33,0,0 // movaps 0x2121(%rip),%xmm3 # 43d0 <_sk_callback_sse41+0x9f0>
+ .byte 15,40,29,31,34,0,0 // movaps 0x221f(%rip),%xmm3 # 4d40 <_sk_callback_sse41+0xaee>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 15,40,208 // movaps %xmm0,%xmm2
.byte 255,224 // jmpq *%rax
@@ -20207,9 +22436,9 @@ _sk_gather_i8_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,137,192 // mov %rax,%r8
.byte 77,133,192 // test %r8,%r8
- .byte 116,5 // je 22c6 <_sk_gather_i8_sse41+0xf>
+ .byte 116,5 // je 2b38 <_sk_gather_i8_sse41+0xf>
.byte 76,137,192 // mov %r8,%rax
- .byte 235,2 // jmp 22c8 <_sk_gather_i8_sse41+0x11>
+ .byte 235,2 // jmp 2b3a <_sk_gather_i8_sse41+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
@@ -20240,17 +22469,17 @@ _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 102,15,111,5,120,32,0,0 // movdqa 0x2078(%rip),%xmm0 # 43e0 <_sk_callback_sse41+0xa00>
+ .byte 102,15,111,5,118,33,0,0 // movdqa 0x2176(%rip),%xmm0 # 4d50 <_sk_callback_sse41+0xafe>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,121,32,0,0 // movaps 0x2079(%rip),%xmm8 # 43f0 <_sk_callback_sse41+0xa10>
+ .byte 68,15,40,5,119,33,0,0 // movaps 0x2177(%rip),%xmm8 # 4d60 <_sk_callback_sse41+0xb0e>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 102,15,111,203 // movdqa %xmm3,%xmm1
- .byte 102,15,56,0,13,120,32,0,0 // pshufb 0x2078(%rip),%xmm1 # 4400 <_sk_callback_sse41+0xa20>
+ .byte 102,15,56,0,13,118,33,0,0 // pshufb 0x2176(%rip),%xmm1 # 4d70 <_sk_callback_sse41+0xb1e>
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,111,211 // movdqa %xmm3,%xmm2
- .byte 102,15,56,0,21,116,32,0,0 // pshufb 0x2074(%rip),%xmm2 # 4410 <_sk_callback_sse41+0xa30>
+ .byte 102,15,56,0,21,114,33,0,0 // pshufb 0x2172(%rip),%xmm2 # 4d80 <_sk_callback_sse41+0xb2e>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 102,15,114,211,24 // psrld $0x18,%xmm3
@@ -20266,19 +22495,19 @@ _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 102,15,111,5,90,32,0,0 // movdqa 0x205a(%rip),%xmm0 # 4420 <_sk_callback_sse41+0xa40>
+ .byte 102,15,111,5,88,33,0,0 // movdqa 0x2158(%rip),%xmm0 # 4d90 <_sk_callback_sse41+0xb3e>
.byte 102,15,219,194 // pand %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,92,32,0,0 // mulps 0x205c(%rip),%xmm0 # 4430 <_sk_callback_sse41+0xa50>
- .byte 102,15,111,13,100,32,0,0 // movdqa 0x2064(%rip),%xmm1 # 4440 <_sk_callback_sse41+0xa60>
+ .byte 15,89,5,90,33,0,0 // mulps 0x215a(%rip),%xmm0 # 4da0 <_sk_callback_sse41+0xb4e>
+ .byte 102,15,111,13,98,33,0,0 // movdqa 0x2162(%rip),%xmm1 # 4db0 <_sk_callback_sse41+0xb5e>
.byte 102,15,219,202 // pand %xmm2,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,102,32,0,0 // mulps 0x2066(%rip),%xmm1 # 4450 <_sk_callback_sse41+0xa70>
- .byte 102,15,219,21,110,32,0,0 // pand 0x206e(%rip),%xmm2 # 4460 <_sk_callback_sse41+0xa80>
+ .byte 15,89,13,100,33,0,0 // mulps 0x2164(%rip),%xmm1 # 4dc0 <_sk_callback_sse41+0xb6e>
+ .byte 102,15,219,21,108,33,0,0 // pand 0x216c(%rip),%xmm2 # 4dd0 <_sk_callback_sse41+0xb7e>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,116,32,0,0 // mulps 0x2074(%rip),%xmm2 # 4470 <_sk_callback_sse41+0xa90>
+ .byte 15,89,21,114,33,0,0 // mulps 0x2172(%rip),%xmm2 # 4de0 <_sk_callback_sse41+0xb8e>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,123,32,0,0 // movaps 0x207b(%rip),%xmm3 # 4480 <_sk_callback_sse41+0xaa0>
+ .byte 15,40,29,121,33,0,0 // movaps 0x2179(%rip),%xmm3 # 4df0 <_sk_callback_sse41+0xb9e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_gather_565_sse41
@@ -20306,19 +22535,19 @@ _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 102,15,111,5,32,32,0,0 // movdqa 0x2020(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xab0>
+ .byte 102,15,111,5,30,33,0,0 // movdqa 0x211e(%rip),%xmm0 # 4e00 <_sk_callback_sse41+0xbae>
.byte 102,15,219,194 // pand %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,34,32,0,0 // mulps 0x2022(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xac0>
- .byte 102,15,111,13,42,32,0,0 // movdqa 0x202a(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xad0>
+ .byte 15,89,5,32,33,0,0 // mulps 0x2120(%rip),%xmm0 # 4e10 <_sk_callback_sse41+0xbbe>
+ .byte 102,15,111,13,40,33,0,0 // movdqa 0x2128(%rip),%xmm1 # 4e20 <_sk_callback_sse41+0xbce>
.byte 102,15,219,202 // pand %xmm2,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,44,32,0,0 // mulps 0x202c(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xae0>
- .byte 102,15,219,21,52,32,0,0 // pand 0x2034(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xaf0>
+ .byte 15,89,13,42,33,0,0 // mulps 0x212a(%rip),%xmm1 # 4e30 <_sk_callback_sse41+0xbde>
+ .byte 102,15,219,21,50,33,0,0 // pand 0x2132(%rip),%xmm2 # 4e40 <_sk_callback_sse41+0xbee>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,58,32,0,0 // mulps 0x203a(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xb00>
+ .byte 15,89,21,56,33,0,0 // mulps 0x2138(%rip),%xmm2 # 4e50 <_sk_callback_sse41+0xbfe>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,65,32,0,0 // movaps 0x2041(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xb10>
+ .byte 15,40,29,63,33,0,0 // movaps 0x213f(%rip),%xmm3 # 4e60 <_sk_callback_sse41+0xc0e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_565_sse41
@@ -20327,12 +22556,12 @@ FUNCTION(_sk_store_565_sse41)
_sk_store_565_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,66,32,0,0 // movaps 0x2042(%rip),%xmm8 # 4500 <_sk_callback_sse41+0xb20>
+ .byte 68,15,40,5,64,33,0,0 // movaps 0x2140(%rip),%xmm8 # 4e70 <_sk_callback_sse41+0xc1e>
.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 68,15,40,21,55,32,0,0 // movaps 0x2037(%rip),%xmm10 # 4510 <_sk_callback_sse41+0xb30>
+ .byte 68,15,40,21,53,33,0,0 // movaps 0x2135(%rip),%xmm10 # 4e80 <_sk_callback_sse41+0xc2e>
.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
@@ -20352,21 +22581,21 @@ _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 102,15,111,5,2,32,0,0 // movdqa 0x2002(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xb40>
+ .byte 102,15,111,5,0,33,0,0 // movdqa 0x2100(%rip),%xmm0 # 4e90 <_sk_callback_sse41+0xc3e>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,4,32,0,0 // mulps 0x2004(%rip),%xmm0 # 4530 <_sk_callback_sse41+0xb50>
- .byte 102,15,111,13,12,32,0,0 // movdqa 0x200c(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xb60>
+ .byte 15,89,5,2,33,0,0 // mulps 0x2102(%rip),%xmm0 # 4ea0 <_sk_callback_sse41+0xc4e>
+ .byte 102,15,111,13,10,33,0,0 // movdqa 0x210a(%rip),%xmm1 # 4eb0 <_sk_callback_sse41+0xc5e>
.byte 102,15,219,203 // pand %xmm3,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,14,32,0,0 // mulps 0x200e(%rip),%xmm1 # 4550 <_sk_callback_sse41+0xb70>
- .byte 102,15,111,21,22,32,0,0 // movdqa 0x2016(%rip),%xmm2 # 4560 <_sk_callback_sse41+0xb80>
+ .byte 15,89,13,12,33,0,0 // mulps 0x210c(%rip),%xmm1 # 4ec0 <_sk_callback_sse41+0xc6e>
+ .byte 102,15,111,21,20,33,0,0 // movdqa 0x2114(%rip),%xmm2 # 4ed0 <_sk_callback_sse41+0xc7e>
.byte 102,15,219,211 // pand %xmm3,%xmm2
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,24,32,0,0 // mulps 0x2018(%rip),%xmm2 # 4570 <_sk_callback_sse41+0xb90>
- .byte 102,15,219,29,32,32,0,0 // pand 0x2020(%rip),%xmm3 # 4580 <_sk_callback_sse41+0xba0>
+ .byte 15,89,21,22,33,0,0 // mulps 0x2116(%rip),%xmm2 # 4ee0 <_sk_callback_sse41+0xc8e>
+ .byte 102,15,219,29,30,33,0,0 // pand 0x211e(%rip),%xmm3 # 4ef0 <_sk_callback_sse41+0xc9e>
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,38,32,0,0 // mulps 0x2026(%rip),%xmm3 # 4590 <_sk_callback_sse41+0xbb0>
+ .byte 15,89,29,36,33,0,0 // mulps 0x2124(%rip),%xmm3 # 4f00 <_sk_callback_sse41+0xcae>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -20395,21 +22624,21 @@ _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 102,15,111,5,201,31,0,0 // movdqa 0x1fc9(%rip),%xmm0 # 45a0 <_sk_callback_sse41+0xbc0>
+ .byte 102,15,111,5,199,32,0,0 // movdqa 0x20c7(%rip),%xmm0 # 4f10 <_sk_callback_sse41+0xcbe>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,203,31,0,0 // mulps 0x1fcb(%rip),%xmm0 # 45b0 <_sk_callback_sse41+0xbd0>
- .byte 102,15,111,13,211,31,0,0 // movdqa 0x1fd3(%rip),%xmm1 # 45c0 <_sk_callback_sse41+0xbe0>
+ .byte 15,89,5,201,32,0,0 // mulps 0x20c9(%rip),%xmm0 # 4f20 <_sk_callback_sse41+0xcce>
+ .byte 102,15,111,13,209,32,0,0 // movdqa 0x20d1(%rip),%xmm1 # 4f30 <_sk_callback_sse41+0xcde>
.byte 102,15,219,203 // pand %xmm3,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,213,31,0,0 // mulps 0x1fd5(%rip),%xmm1 # 45d0 <_sk_callback_sse41+0xbf0>
- .byte 102,15,111,21,221,31,0,0 // movdqa 0x1fdd(%rip),%xmm2 # 45e0 <_sk_callback_sse41+0xc00>
+ .byte 15,89,13,211,32,0,0 // mulps 0x20d3(%rip),%xmm1 # 4f40 <_sk_callback_sse41+0xcee>
+ .byte 102,15,111,21,219,32,0,0 // movdqa 0x20db(%rip),%xmm2 # 4f50 <_sk_callback_sse41+0xcfe>
.byte 102,15,219,211 // pand %xmm3,%xmm2
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,223,31,0,0 // mulps 0x1fdf(%rip),%xmm2 # 45f0 <_sk_callback_sse41+0xc10>
- .byte 102,15,219,29,231,31,0,0 // pand 0x1fe7(%rip),%xmm3 # 4600 <_sk_callback_sse41+0xc20>
+ .byte 15,89,21,221,32,0,0 // mulps 0x20dd(%rip),%xmm2 # 4f60 <_sk_callback_sse41+0xd0e>
+ .byte 102,15,219,29,229,32,0,0 // pand 0x20e5(%rip),%xmm3 # 4f70 <_sk_callback_sse41+0xd1e>
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,237,31,0,0 // mulps 0x1fed(%rip),%xmm3 # 4610 <_sk_callback_sse41+0xc30>
+ .byte 15,89,29,235,32,0,0 // mulps 0x20eb(%rip),%xmm3 # 4f80 <_sk_callback_sse41+0xd2e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -20419,7 +22648,7 @@ FUNCTION(_sk_store_4444_sse41)
_sk_store_4444_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,236,31,0,0 // movaps 0x1fec(%rip),%xmm8 # 4620 <_sk_callback_sse41+0xc40>
+ .byte 68,15,40,5,234,32,0,0 // movaps 0x20ea(%rip),%xmm8 # 4f90 <_sk_callback_sse41+0xd3e>
.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
@@ -20449,17 +22678,17 @@ _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 15,40,5,139,31,0,0 // movaps 0x1f8b(%rip),%xmm0 # 4630 <_sk_callback_sse41+0xc50>
+ .byte 15,40,5,137,32,0,0 // movaps 0x2089(%rip),%xmm0 # 4fa0 <_sk_callback_sse41+0xd4e>
.byte 15,84,195 // andps %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,141,31,0,0 // movaps 0x1f8d(%rip),%xmm8 # 4640 <_sk_callback_sse41+0xc60>
+ .byte 68,15,40,5,139,32,0,0 // movaps 0x208b(%rip),%xmm8 # 4fb0 <_sk_callback_sse41+0xd5e>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 15,40,203 // movaps %xmm3,%xmm1
- .byte 102,15,56,0,13,141,31,0,0 // pshufb 0x1f8d(%rip),%xmm1 # 4650 <_sk_callback_sse41+0xc70>
+ .byte 102,15,56,0,13,139,32,0,0 // pshufb 0x208b(%rip),%xmm1 # 4fc0 <_sk_callback_sse41+0xd6e>
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 15,40,211 // movaps %xmm3,%xmm2
- .byte 102,15,56,0,21,138,31,0,0 // pshufb 0x1f8a(%rip),%xmm2 # 4660 <_sk_callback_sse41+0xc80>
+ .byte 102,15,56,0,21,136,32,0,0 // pshufb 0x2088(%rip),%xmm2 # 4fd0 <_sk_callback_sse41+0xd7e>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 102,15,114,211,24 // psrld $0x18,%xmm3
@@ -20490,17 +22719,17 @@ _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 102,15,111,5,35,31,0,0 // movdqa 0x1f23(%rip),%xmm0 # 4670 <_sk_callback_sse41+0xc90>
+ .byte 102,15,111,5,33,32,0,0 // movdqa 0x2021(%rip),%xmm0 # 4fe0 <_sk_callback_sse41+0xd8e>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,36,31,0,0 // movaps 0x1f24(%rip),%xmm8 # 4680 <_sk_callback_sse41+0xca0>
+ .byte 68,15,40,5,34,32,0,0 // movaps 0x2022(%rip),%xmm8 # 4ff0 <_sk_callback_sse41+0xd9e>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 102,15,111,203 // movdqa %xmm3,%xmm1
- .byte 102,15,56,0,13,35,31,0,0 // pshufb 0x1f23(%rip),%xmm1 # 4690 <_sk_callback_sse41+0xcb0>
+ .byte 102,15,56,0,13,33,32,0,0 // pshufb 0x2021(%rip),%xmm1 # 5000 <_sk_callback_sse41+0xdae>
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,111,211 // movdqa %xmm3,%xmm2
- .byte 102,15,56,0,21,31,31,0,0 // pshufb 0x1f1f(%rip),%xmm2 # 46a0 <_sk_callback_sse41+0xcc0>
+ .byte 102,15,56,0,21,29,32,0,0 // pshufb 0x201d(%rip),%xmm2 # 5010 <_sk_callback_sse41+0xdbe>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 102,15,114,211,24 // psrld $0x18,%xmm3
@@ -20515,7 +22744,7 @@ FUNCTION(_sk_store_8888_sse41)
_sk_store_8888_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,11,31,0,0 // movaps 0x1f0b(%rip),%xmm8 # 46b0 <_sk_callback_sse41+0xcd0>
+ .byte 68,15,40,5,9,32,0,0 // movaps 0x2009(%rip),%xmm8 # 5020 <_sk_callback_sse41+0xdce>
.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
@@ -20552,18 +22781,18 @@ _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 102,68,15,111,5,132,30,0,0 // movdqa 0x1e84(%rip),%xmm8 # 46c0 <_sk_callback_sse41+0xce0>
+ .byte 102,68,15,111,5,130,31,0,0 // movdqa 0x1f82(%rip),%xmm8 # 5030 <_sk_callback_sse41+0xdde>
.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 102,15,111,29,127,30,0,0 // movdqa 0x1e7f(%rip),%xmm3 # 46d0 <_sk_callback_sse41+0xcf0>
+ .byte 102,15,111,29,125,31,0,0 // movdqa 0x1f7d(%rip),%xmm3 # 5040 <_sk_callback_sse41+0xdee>
.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 102,68,15,111,21,107,30,0,0 // movdqa 0x1e6b(%rip),%xmm10 # 46e0 <_sk_callback_sse41+0xd00>
+ .byte 102,68,15,111,21,105,31,0,0 // movdqa 0x1f69(%rip),%xmm10 # 5050 <_sk_callback_sse41+0xdfe>
.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
@@ -20636,18 +22865,18 @@ _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 102,68,15,111,5,41,29,0,0 // movdqa 0x1d29(%rip),%xmm8 # 46f0 <_sk_callback_sse41+0xd10>
+ .byte 102,68,15,111,5,39,30,0,0 // movdqa 0x1e27(%rip),%xmm8 # 5060 <_sk_callback_sse41+0xe0e>
.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 102,15,111,29,36,29,0,0 // movdqa 0x1d24(%rip),%xmm3 # 4700 <_sk_callback_sse41+0xd20>
+ .byte 102,15,111,29,34,30,0,0 // movdqa 0x1e22(%rip),%xmm3 # 5070 <_sk_callback_sse41+0xe1e>
.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 102,68,15,111,21,16,29,0,0 // movdqa 0x1d10(%rip),%xmm10 # 4710 <_sk_callback_sse41+0xd30>
+ .byte 102,68,15,111,21,14,30,0,0 // movdqa 0x1e0e(%rip),%xmm10 # 5080 <_sk_callback_sse41+0xe2e>
.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
@@ -20695,17 +22924,17 @@ FUNCTION(_sk_store_f16_sse41)
_sk_store_f16_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 102,68,15,111,21,70,28,0,0 // movdqa 0x1c46(%rip),%xmm10 # 4720 <_sk_callback_sse41+0xd40>
+ .byte 102,68,15,111,21,68,29,0,0 // movdqa 0x1d44(%rip),%xmm10 # 5090 <_sk_callback_sse41+0xe3e>
.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 102,68,15,111,13,57,28,0,0 // movdqa 0x1c39(%rip),%xmm9 # 4730 <_sk_callback_sse41+0xd50>
+ .byte 102,68,15,111,13,55,29,0,0 // movdqa 0x1d37(%rip),%xmm9 # 50a0 <_sk_callback_sse41+0xe4e>
.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 102,68,15,111,29,42,28,0,0 // movdqa 0x1c2a(%rip),%xmm11 # 4740 <_sk_callback_sse41+0xd60>
+ .byte 102,68,15,111,29,40,29,0,0 // movdqa 0x1d28(%rip),%xmm11 # 50b0 <_sk_callback_sse41+0xe5e>
.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
@@ -20775,7 +23004,7 @@ _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 68,15,40,5,249,26,0,0 // movaps 0x1af9(%rip),%xmm8 # 4750 <_sk_callback_sse41+0xd70>
+ .byte 68,15,40,5,247,27,0,0 // movaps 0x1bf7(%rip),%xmm8 # 50c0 <_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
@@ -20827,7 +23056,7 @@ _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 68,15,40,5,58,26,0,0 // movaps 0x1a3a(%rip),%xmm8 # 4760 <_sk_callback_sse41+0xd80>
+ .byte 68,15,40,5,56,27,0,0 // movaps 0x1b38(%rip),%xmm8 # 50d0 <_sk_callback_sse41+0xe7e>
.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
@@ -20844,7 +23073,7 @@ _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 15,40,29,1,26,0,0 // movaps 0x1a01(%rip),%xmm3 # 4770 <_sk_callback_sse41+0xd90>
+ .byte 15,40,29,255,26,0,0 // movaps 0x1aff(%rip),%xmm3 # 50e0 <_sk_callback_sse41+0xe8e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_u16_be_sse41
@@ -20853,7 +23082,7 @@ 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 68,15,40,13,2,26,0,0 // movaps 0x1a02(%rip),%xmm9 # 4780 <_sk_callback_sse41+0xda0>
+ .byte 68,15,40,13,0,27,0,0 // movaps 0x1b00(%rip),%xmm9 # 50f0 <_sk_callback_sse41+0xe9e>
.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
@@ -21076,10 +23305,10 @@ HIDDEN _sk_luminance_to_alpha_sse41
FUNCTION(_sk_luminance_to_alpha_sse41)
_sk_luminance_to_alpha_sse41:
.byte 15,40,218 // movaps %xmm2,%xmm3
- .byte 15,89,5,32,23,0,0 // mulps 0x1720(%rip),%xmm0 # 4790 <_sk_callback_sse41+0xdb0>
- .byte 15,89,13,41,23,0,0 // mulps 0x1729(%rip),%xmm1 # 47a0 <_sk_callback_sse41+0xdc0>
+ .byte 15,89,5,30,24,0,0 // mulps 0x181e(%rip),%xmm0 # 5100 <_sk_callback_sse41+0xeae>
+ .byte 15,89,13,39,24,0,0 // mulps 0x1827(%rip),%xmm1 # 5110 <_sk_callback_sse41+0xebe>
.byte 15,88,200 // addps %xmm0,%xmm1
- .byte 15,89,29,47,23,0,0 // mulps 0x172f(%rip),%xmm3 # 47b0 <_sk_callback_sse41+0xdd0>
+ .byte 15,89,29,45,24,0,0 // mulps 0x182d(%rip),%xmm3 # 5120 <_sk_callback_sse41+0xece>
.byte 15,88,217 // addps %xmm1,%xmm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
@@ -21312,7 +23541,7 @@ _sk_linear_gradient_sse41:
.byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
.byte 72,139,8 // mov (%rax),%rcx
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,132,254,0,0,0 // je 3523 <_sk_linear_gradient_sse41+0x138>
+ .byte 15,132,254,0,0,0 // je 3d95 <_sk_linear_gradient_sse41+0x138>
.byte 15,41,100,36,168 // movaps %xmm4,-0x58(%rsp)
.byte 15,41,108,36,184 // movaps %xmm5,-0x48(%rsp)
.byte 15,41,116,36,200 // movaps %xmm6,-0x38(%rsp)
@@ -21362,12 +23591,12 @@ _sk_linear_gradient_sse41:
.byte 15,40,196 // movaps %xmm4,%xmm0
.byte 72,131,192,36 // add $0x24,%rax
.byte 72,255,201 // dec %rcx
- .byte 15,133,65,255,255,255 // jne 344e <_sk_linear_gradient_sse41+0x63>
+ .byte 15,133,65,255,255,255 // jne 3cc0 <_sk_linear_gradient_sse41+0x63>
.byte 15,40,124,36,216 // movaps -0x28(%rsp),%xmm7
.byte 15,40,116,36,200 // movaps -0x38(%rsp),%xmm6
.byte 15,40,108,36,184 // movaps -0x48(%rsp),%xmm5
.byte 15,40,100,36,168 // movaps -0x58(%rsp),%xmm4
- .byte 235,13 // jmp 3530 <_sk_linear_gradient_sse41+0x145>
+ .byte 235,13 // jmp 3da2 <_sk_linear_gradient_sse41+0x145>
.byte 15,87,201 // xorps %xmm1,%xmm1
.byte 15,87,210 // xorps %xmm2,%xmm2
.byte 15,87,219 // xorps %xmm3,%xmm3
@@ -21436,26 +23665,26 @@ _sk_xy_to_polar_unit_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 68,15,40,21,184,17,0,0 // movaps 0x11b8(%rip),%xmm10 # 47c0 <_sk_callback_sse41+0xde0>
+ .byte 68,15,40,21,182,18,0,0 // movaps 0x12b6(%rip),%xmm10 # 5130 <_sk_callback_sse41+0xede>
.byte 69,15,89,213 // mulps %xmm13,%xmm10
- .byte 68,15,88,21,188,17,0,0 // addps 0x11bc(%rip),%xmm10 # 47d0 <_sk_callback_sse41+0xdf0>
+ .byte 68,15,88,21,186,18,0,0 // addps 0x12ba(%rip),%xmm10 # 5140 <_sk_callback_sse41+0xeee>
.byte 69,15,89,213 // mulps %xmm13,%xmm10
- .byte 68,15,88,21,192,17,0,0 // addps 0x11c0(%rip),%xmm10 # 47e0 <_sk_callback_sse41+0xe00>
+ .byte 68,15,88,21,190,18,0,0 // addps 0x12be(%rip),%xmm10 # 5150 <_sk_callback_sse41+0xefe>
.byte 69,15,89,213 // mulps %xmm13,%xmm10
- .byte 68,15,88,21,196,17,0,0 // addps 0x11c4(%rip),%xmm10 # 47f0 <_sk_callback_sse41+0xe10>
+ .byte 68,15,88,21,194,18,0,0 // addps 0x12c2(%rip),%xmm10 # 5160 <_sk_callback_sse41+0xf0e>
.byte 69,15,89,212 // mulps %xmm12,%xmm10
.byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0
- .byte 68,15,40,29,195,17,0,0 // movaps 0x11c3(%rip),%xmm11 # 4800 <_sk_callback_sse41+0xe20>
+ .byte 68,15,40,29,193,18,0,0 // movaps 0x12c1(%rip),%xmm11 # 5170 <_sk_callback_sse41+0xf1e>
.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 68,15,40,29,188,17,0,0 // movaps 0x11bc(%rip),%xmm11 # 4810 <_sk_callback_sse41+0xe30>
+ .byte 68,15,40,29,186,18,0,0 // movaps 0x12ba(%rip),%xmm11 # 5180 <_sk_callback_sse41+0xf2e>
.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 68,15,40,13,174,17,0,0 // movaps 0x11ae(%rip),%xmm9 # 4820 <_sk_callback_sse41+0xe40>
+ .byte 68,15,40,13,172,18,0,0 // movaps 0x12ac(%rip),%xmm9 # 5190 <_sk_callback_sse41+0xf3e>
.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
@@ -21469,7 +23698,7 @@ HIDDEN _sk_save_xy_sse41
FUNCTION(_sk_save_xy_sse41)
_sk_save_xy_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,153,17,0,0 // movaps 0x1199(%rip),%xmm8 # 4830 <_sk_callback_sse41+0xe50>
+ .byte 68,15,40,5,151,18,0,0 // movaps 0x1297(%rip),%xmm8 # 51a0 <_sk_callback_sse41+0xf4e>
.byte 15,17,0 // movups %xmm0,(%rax)
.byte 68,15,40,200 // movaps %xmm0,%xmm9
.byte 69,15,88,200 // addps %xmm8,%xmm9
@@ -21513,8 +23742,8 @@ _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 15,88,5,27,17,0,0 // addps 0x111b(%rip),%xmm0 # 4840 <_sk_callback_sse41+0xe60>
- .byte 68,15,40,13,35,17,0,0 // movaps 0x1123(%rip),%xmm9 # 4850 <_sk_callback_sse41+0xe70>
+ .byte 15,88,5,25,18,0,0 // addps 0x1219(%rip),%xmm0 # 51b0 <_sk_callback_sse41+0xf5e>
+ .byte 68,15,40,13,33,18,0,0 // movaps 0x1221(%rip),%xmm9 # 51c0 <_sk_callback_sse41+0xf6e>
.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
@@ -21527,7 +23756,7 @@ _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 15,88,5,18,17,0,0 // addps 0x1112(%rip),%xmm0 # 4860 <_sk_callback_sse41+0xe80>
+ .byte 15,88,5,16,18,0,0 // addps 0x1210(%rip),%xmm0 # 51d0 <_sk_callback_sse41+0xf7e>
.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
@@ -21539,8 +23768,8 @@ _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 15,88,13,4,17,0,0 // addps 0x1104(%rip),%xmm1 # 4870 <_sk_callback_sse41+0xe90>
- .byte 68,15,40,13,12,17,0,0 // movaps 0x110c(%rip),%xmm9 # 4880 <_sk_callback_sse41+0xea0>
+ .byte 15,88,13,2,18,0,0 // addps 0x1202(%rip),%xmm1 # 51e0 <_sk_callback_sse41+0xf8e>
+ .byte 68,15,40,13,10,18,0,0 // movaps 0x120a(%rip),%xmm9 # 51f0 <_sk_callback_sse41+0xf9e>
.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
@@ -21553,7 +23782,7 @@ _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 15,88,13,250,16,0,0 // addps 0x10fa(%rip),%xmm1 # 4890 <_sk_callback_sse41+0xeb0>
+ .byte 15,88,13,248,17,0,0 // addps 0x11f8(%rip),%xmm1 # 5200 <_sk_callback_sse41+0xfae>
.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
@@ -21565,13 +23794,13 @@ _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 15,88,5,237,16,0,0 // addps 0x10ed(%rip),%xmm0 # 48a0 <_sk_callback_sse41+0xec0>
- .byte 68,15,40,13,245,16,0,0 // movaps 0x10f5(%rip),%xmm9 # 48b0 <_sk_callback_sse41+0xed0>
+ .byte 15,88,5,235,17,0,0 // addps 0x11eb(%rip),%xmm0 # 5210 <_sk_callback_sse41+0xfbe>
+ .byte 68,15,40,13,243,17,0,0 // movaps 0x11f3(%rip),%xmm9 # 5220 <_sk_callback_sse41+0xfce>
.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 68,15,89,13,241,16,0,0 // mulps 0x10f1(%rip),%xmm9 # 48c0 <_sk_callback_sse41+0xee0>
- .byte 68,15,88,13,249,16,0,0 // addps 0x10f9(%rip),%xmm9 # 48d0 <_sk_callback_sse41+0xef0>
+ .byte 68,15,89,13,239,17,0,0 // mulps 0x11ef(%rip),%xmm9 # 5230 <_sk_callback_sse41+0xfde>
+ .byte 68,15,88,13,247,17,0,0 // addps 0x11f7(%rip),%xmm9 # 5240 <_sk_callback_sse41+0xfee>
.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
@@ -21584,16 +23813,16 @@ _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 15,88,5,232,16,0,0 // addps 0x10e8(%rip),%xmm0 # 48e0 <_sk_callback_sse41+0xf00>
- .byte 68,15,40,13,240,16,0,0 // movaps 0x10f0(%rip),%xmm9 # 48f0 <_sk_callback_sse41+0xf10>
+ .byte 15,88,5,230,17,0,0 // addps 0x11e6(%rip),%xmm0 # 5250 <_sk_callback_sse41+0xffe>
+ .byte 68,15,40,13,238,17,0,0 // movaps 0x11ee(%rip),%xmm9 # 5260 <_sk_callback_sse41+0x100e>
.byte 69,15,92,200 // subps %xmm8,%xmm9
- .byte 68,15,40,5,244,16,0,0 // movaps 0x10f4(%rip),%xmm8 # 4900 <_sk_callback_sse41+0xf20>
+ .byte 68,15,40,5,242,17,0,0 // movaps 0x11f2(%rip),%xmm8 # 5270 <_sk_callback_sse41+0x101e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,248,16,0,0 // addps 0x10f8(%rip),%xmm8 # 4910 <_sk_callback_sse41+0xf30>
+ .byte 68,15,88,5,246,17,0,0 // addps 0x11f6(%rip),%xmm8 # 5280 <_sk_callback_sse41+0x102e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,252,16,0,0 // addps 0x10fc(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xf40>
+ .byte 68,15,88,5,250,17,0,0 // addps 0x11fa(%rip),%xmm8 # 5290 <_sk_callback_sse41+0x103e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,0,17,0,0 // addps 0x1100(%rip),%xmm8 # 4930 <_sk_callback_sse41+0xf50>
+ .byte 68,15,88,5,254,17,0,0 // addps 0x11fe(%rip),%xmm8 # 52a0 <_sk_callback_sse41+0x104e>
.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
@@ -21603,17 +23832,17 @@ HIDDEN _sk_bicubic_p1x_sse41
FUNCTION(_sk_bicubic_p1x_sse41)
_sk_bicubic_p1x_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,250,16,0,0 // movaps 0x10fa(%rip),%xmm8 # 4940 <_sk_callback_sse41+0xf60>
+ .byte 68,15,40,5,248,17,0,0 // movaps 0x11f8(%rip),%xmm8 # 52b0 <_sk_callback_sse41+0x105e>
.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 68,15,40,21,246,16,0,0 // movaps 0x10f6(%rip),%xmm10 # 4950 <_sk_callback_sse41+0xf70>
+ .byte 68,15,40,21,244,17,0,0 // movaps 0x11f4(%rip),%xmm10 # 52c0 <_sk_callback_sse41+0x106e>
.byte 69,15,89,209 // mulps %xmm9,%xmm10
- .byte 68,15,88,21,250,16,0,0 // addps 0x10fa(%rip),%xmm10 # 4960 <_sk_callback_sse41+0xf80>
+ .byte 68,15,88,21,248,17,0,0 // addps 0x11f8(%rip),%xmm10 # 52d0 <_sk_callback_sse41+0x107e>
.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 68,15,88,21,246,16,0,0 // addps 0x10f6(%rip),%xmm10 # 4970 <_sk_callback_sse41+0xf90>
+ .byte 68,15,88,21,244,17,0,0 // addps 0x11f4(%rip),%xmm10 # 52e0 <_sk_callback_sse41+0x108e>
.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
@@ -21625,11 +23854,11 @@ _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 15,88,5,233,16,0,0 // addps 0x10e9(%rip),%xmm0 # 4980 <_sk_callback_sse41+0xfa0>
+ .byte 15,88,5,231,17,0,0 // addps 0x11e7(%rip),%xmm0 # 52f0 <_sk_callback_sse41+0x109e>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,89,201 // mulps %xmm9,%xmm9
- .byte 68,15,89,5,233,16,0,0 // mulps 0x10e9(%rip),%xmm8 # 4990 <_sk_callback_sse41+0xfb0>
- .byte 68,15,88,5,241,16,0,0 // addps 0x10f1(%rip),%xmm8 # 49a0 <_sk_callback_sse41+0xfc0>
+ .byte 68,15,89,5,231,17,0,0 // mulps 0x11e7(%rip),%xmm8 # 5300 <_sk_callback_sse41+0x10ae>
+ .byte 68,15,88,5,239,17,0,0 // addps 0x11ef(%rip),%xmm8 # 5310 <_sk_callback_sse41+0x10be>
.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
@@ -21642,13 +23871,13 @@ _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 15,88,13,223,16,0,0 // addps 0x10df(%rip),%xmm1 # 49b0 <_sk_callback_sse41+0xfd0>
- .byte 68,15,40,13,231,16,0,0 // movaps 0x10e7(%rip),%xmm9 # 49c0 <_sk_callback_sse41+0xfe0>
+ .byte 15,88,13,221,17,0,0 // addps 0x11dd(%rip),%xmm1 # 5320 <_sk_callback_sse41+0x10ce>
+ .byte 68,15,40,13,229,17,0,0 // movaps 0x11e5(%rip),%xmm9 # 5330 <_sk_callback_sse41+0x10de>
.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 68,15,89,13,227,16,0,0 // mulps 0x10e3(%rip),%xmm9 # 49d0 <_sk_callback_sse41+0xff0>
- .byte 68,15,88,13,235,16,0,0 // addps 0x10eb(%rip),%xmm9 # 49e0 <_sk_callback_sse41+0x1000>
+ .byte 68,15,89,13,225,17,0,0 // mulps 0x11e1(%rip),%xmm9 # 5340 <_sk_callback_sse41+0x10ee>
+ .byte 68,15,88,13,233,17,0,0 // addps 0x11e9(%rip),%xmm9 # 5350 <_sk_callback_sse41+0x10fe>
.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
@@ -21661,16 +23890,16 @@ _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 15,88,13,217,16,0,0 // addps 0x10d9(%rip),%xmm1 # 49f0 <_sk_callback_sse41+0x1010>
- .byte 68,15,40,13,225,16,0,0 // movaps 0x10e1(%rip),%xmm9 # 4a00 <_sk_callback_sse41+0x1020>
+ .byte 15,88,13,215,17,0,0 // addps 0x11d7(%rip),%xmm1 # 5360 <_sk_callback_sse41+0x110e>
+ .byte 68,15,40,13,223,17,0,0 // movaps 0x11df(%rip),%xmm9 # 5370 <_sk_callback_sse41+0x111e>
.byte 69,15,92,200 // subps %xmm8,%xmm9
- .byte 68,15,40,5,229,16,0,0 // movaps 0x10e5(%rip),%xmm8 # 4a10 <_sk_callback_sse41+0x1030>
+ .byte 68,15,40,5,227,17,0,0 // movaps 0x11e3(%rip),%xmm8 # 5380 <_sk_callback_sse41+0x112e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,233,16,0,0 // addps 0x10e9(%rip),%xmm8 # 4a20 <_sk_callback_sse41+0x1040>
+ .byte 68,15,88,5,231,17,0,0 // addps 0x11e7(%rip),%xmm8 # 5390 <_sk_callback_sse41+0x113e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,237,16,0,0 // addps 0x10ed(%rip),%xmm8 # 4a30 <_sk_callback_sse41+0x1050>
+ .byte 68,15,88,5,235,17,0,0 // addps 0x11eb(%rip),%xmm8 # 53a0 <_sk_callback_sse41+0x114e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,241,16,0,0 // addps 0x10f1(%rip),%xmm8 # 4a40 <_sk_callback_sse41+0x1060>
+ .byte 68,15,88,5,239,17,0,0 // addps 0x11ef(%rip),%xmm8 # 53b0 <_sk_callback_sse41+0x115e>
.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
@@ -21680,17 +23909,17 @@ HIDDEN _sk_bicubic_p1y_sse41
FUNCTION(_sk_bicubic_p1y_sse41)
_sk_bicubic_p1y_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,235,16,0,0 // movaps 0x10eb(%rip),%xmm8 # 4a50 <_sk_callback_sse41+0x1070>
+ .byte 68,15,40,5,233,17,0,0 // movaps 0x11e9(%rip),%xmm8 # 53c0 <_sk_callback_sse41+0x116e>
.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 68,15,40,21,230,16,0,0 // movaps 0x10e6(%rip),%xmm10 # 4a60 <_sk_callback_sse41+0x1080>
+ .byte 68,15,40,21,228,17,0,0 // movaps 0x11e4(%rip),%xmm10 # 53d0 <_sk_callback_sse41+0x117e>
.byte 69,15,89,209 // mulps %xmm9,%xmm10
- .byte 68,15,88,21,234,16,0,0 // addps 0x10ea(%rip),%xmm10 # 4a70 <_sk_callback_sse41+0x1090>
+ .byte 68,15,88,21,232,17,0,0 // addps 0x11e8(%rip),%xmm10 # 53e0 <_sk_callback_sse41+0x118e>
.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 68,15,88,21,230,16,0,0 // addps 0x10e6(%rip),%xmm10 # 4a80 <_sk_callback_sse41+0x10a0>
+ .byte 68,15,88,21,228,17,0,0 // addps 0x11e4(%rip),%xmm10 # 53f0 <_sk_callback_sse41+0x119e>
.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
@@ -21702,11 +23931,11 @@ _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 15,88,13,216,16,0,0 // addps 0x10d8(%rip),%xmm1 # 4a90 <_sk_callback_sse41+0x10b0>
+ .byte 15,88,13,214,17,0,0 // addps 0x11d6(%rip),%xmm1 # 5400 <_sk_callback_sse41+0x11ae>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,89,201 // mulps %xmm9,%xmm9
- .byte 68,15,89,5,216,16,0,0 // mulps 0x10d8(%rip),%xmm8 # 4aa0 <_sk_callback_sse41+0x10c0>
- .byte 68,15,88,5,224,16,0,0 // addps 0x10e0(%rip),%xmm8 # 4ab0 <_sk_callback_sse41+0x10d0>
+ .byte 68,15,89,5,214,17,0,0 // mulps 0x11d6(%rip),%xmm8 # 5410 <_sk_callback_sse41+0x11be>
+ .byte 68,15,88,5,222,17,0,0 // addps 0x11de(%rip),%xmm8 # 5420 <_sk_callback_sse41+0x11ce>
.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
@@ -21925,11 +24154,135 @@ 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 3c98 <.literal16+0x1d8>
+ .byte 224,64 // loopne 4508 <.literal16+0x1d8>
+ .byte 0,0 // add %al,(%rax)
+ .byte 224,64 // loopne 450c <.literal16+0x1dc>
+ .byte 0,0 // add %al,(%rax)
+ .byte 224,64 // loopne 4510 <.literal16+0x1e0>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4541 <.literal16+0x211>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4545 <.literal16+0x215>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4549 <.literal16+0x219>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 454d <.literal16+0x21d>
.byte 0,0 // add %al,(%rax)
- .byte 224,64 // loopne 3c9c <.literal16+0x1dc>
+ .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 224,64 // loopne 3ca0 <.literal16+0x1e0>
+ .byte 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4581 <.literal16+0x251>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4585 <.literal16+0x255>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4589 <.literal16+0x259>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 458d <.literal16+0x25d>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 45c1 <.literal16+0x291>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 45c5 <.literal16+0x295>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 45c9 <.literal16+0x299>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 45cd <.literal16+0x29d>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4601 <.literal16+0x2d1>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4605 <.literal16+0x2d5>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4609 <.literal16+0x2d9>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 460d <.literal16+0x2dd>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
@@ -22148,13 +24501,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 3e59 <.literal16+0x399>
+ .byte 224,7 // loopne 47c9 <.literal16+0x499>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 3e5d <.literal16+0x39d>
+ .byte 224,7 // loopne 47cd <.literal16+0x49d>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 3e61 <.literal16+0x3a1>
+ .byte 224,7 // loopne 47d1 <.literal16+0x4a1>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 3e65 <.literal16+0x3a5>
+ .byte 224,7 // loopne 47d5 <.literal16+0x4a5>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -22194,10 +24547,10 @@ BALIGN16
.byte 0,1 // add %al,(%rcx)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a003eb8 <_sk_callback_sse41+0xa0004d8>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004828 <_sk_callback_sse41+0xa0005d6>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3003ec0 <_sk_callback_sse41+0x30004e0>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004830 <_sk_callback_sse41+0x30005de>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -22252,11 +24605,11 @@ 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 3f8b <.literal16+0x4cb>
+ .byte 127,67 // jg 48fb <.literal16+0x5cb>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 3f8f <.literal16+0x4cf>
+ .byte 127,67 // jg 48ff <.literal16+0x5cf>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 3f93 <.literal16+0x4d3>
+ .byte 127,67 // jg 4903 <.literal16+0x5d3>
.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)
@@ -22271,16 +24624,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 3f84 <.literal16+0x4c4>
+ .byte 127,0 // jg 48f4 <.literal16+0x5c4>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3f88 <.literal16+0x4c8>
+ .byte 127,0 // jg 48f8 <.literal16+0x5c8>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3f8c <.literal16+0x4cc>
+ .byte 127,0 // jg 48fc <.literal16+0x5cc>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3f90 <.literal16+0x4d0>
+ .byte 127,0 // jg 4900 <.literal16+0x5d0>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22289,7 +24642,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4015 <.literal16+0x555>
+ .byte 119,115 // ja 4985 <.literal16+0x655>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -22300,7 +24653,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 3f79 <.literal16+0x4b9>
+ .byte 117,191 // jne 48e9 <.literal16+0x5b9>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -22312,7 +24665,7 @@ BALIGN16
.byte 249 // stc
.byte 68,180,62 // rex.R mov $0x3e,%spl
.byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
- .byte 233,220,63,163,233 // jmpq ffffffffe9a37fba <_sk_callback_sse41+0xffffffffe9a345da>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a3892a <_sk_callback_sse41+0xffffffffe9a346d8>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22367,16 +24720,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4054 <.literal16+0x594>
+ .byte 127,0 // jg 49c4 <.literal16+0x694>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4058 <.literal16+0x598>
+ .byte 127,0 // jg 49c8 <.literal16+0x698>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 405c <.literal16+0x59c>
+ .byte 127,0 // jg 49cc <.literal16+0x69c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4060 <.literal16+0x5a0>
+ .byte 127,0 // jg 49d0 <.literal16+0x6a0>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22385,7 +24738,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 40e5 <.literal16+0x625>
+ .byte 119,115 // ja 4a55 <.literal16+0x725>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -22396,7 +24749,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 4049 <.literal16+0x589>
+ .byte 117,191 // jne 49b9 <.literal16+0x689>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -22408,7 +24761,7 @@ BALIGN16
.byte 249 // stc
.byte 68,180,62 // rex.R mov $0x3e,%spl
.byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
- .byte 233,220,63,163,233 // jmpq ffffffffe9a3808a <_sk_callback_sse41+0xffffffffe9a346aa>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a389fa <_sk_callback_sse41+0xffffffffe9a347a8>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22463,16 +24816,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4124 <.literal16+0x664>
+ .byte 127,0 // jg 4a94 <.literal16+0x764>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4128 <.literal16+0x668>
+ .byte 127,0 // jg 4a98 <.literal16+0x768>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 412c <.literal16+0x66c>
+ .byte 127,0 // jg 4a9c <.literal16+0x76c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4130 <.literal16+0x670>
+ .byte 127,0 // jg 4aa0 <.literal16+0x770>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22481,7 +24834,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 41b5 <.literal16+0x6f5>
+ .byte 119,115 // ja 4b25 <.literal16+0x7f5>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -22492,7 +24845,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 4119 <.literal16+0x659>
+ .byte 117,191 // jne 4a89 <.literal16+0x759>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -22504,7 +24857,7 @@ 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 ffffffffe9a3815a <_sk_callback_sse41+0xffffffffe9a3477a>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a38aca <_sk_callback_sse41+0xffffffffe9a34878>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22559,16 +24912,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 41f4 <.literal16+0x734>
+ .byte 127,0 // jg 4b64 <.literal16+0x834>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 41f8 <.literal16+0x738>
+ .byte 127,0 // jg 4b68 <.literal16+0x838>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 41fc <.literal16+0x73c>
+ .byte 127,0 // jg 4b6c <.literal16+0x83c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4200 <.literal16+0x740>
+ .byte 127,0 // jg 4b70 <.literal16+0x840>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22577,7 +24930,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4285 <.literal16+0x7c5>
+ .byte 119,115 // ja 4bf5 <.literal16+0x8c5>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -22588,7 +24941,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 41e9 <.literal16+0x729>
+ .byte 117,191 // jne 4b59 <.literal16+0x829>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -22600,7 +24953,7 @@ 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 ffffffffe9a3822a <_sk_callback_sse41+0xffffffffe9a3484a>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a38b9a <_sk_callback_sse41+0xffffffffe9a34948>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22651,13 +25004,13 @@ 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 4307 <.literal16+0x847>
+ .byte 127,67 // jg 4c77 <.literal16+0x947>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 430b <.literal16+0x84b>
+ .byte 127,67 // jg 4c7b <.literal16+0x94b>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 430f <.literal16+0x84f>
+ .byte 127,67 // jg 4c7f <.literal16+0x94f>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 4313 <.literal16+0x853>
+ .byte 127,67 // jg 4c83 <.literal16+0x953>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -22704,16 +25057,16 @@ BALIGN16
.byte 128,3,62 // addb $0x3e,(%rbx)
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 4393 <.literal16+0x8d3>
+ .byte 118,63 // jbe 4d03 <.literal16+0x9d3>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 4397 <.literal16+0x8d7>
+ .byte 118,63 // jbe 4d07 <.literal16+0x9d7>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 439b <.literal16+0x8db>
+ .byte 118,63 // jbe 4d0b <.literal16+0x9db>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 439f <.literal16+0x8df>
+ .byte 118,63 // jbe 4d0f <.literal16+0x9df>
.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)
@@ -22725,11 +25078,11 @@ 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 43db <.literal16+0x91b>
+ .byte 127,67 // jg 4d4b <.literal16+0xa1b>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 43df <.literal16+0x91f>
+ .byte 127,67 // jg 4d4f <.literal16+0xa1f>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 43e3 <.literal16+0x923>
+ .byte 127,67 // jg 4d53 <.literal16+0xa23>
.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)
@@ -22758,7 +25111,7 @@ 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) # 3004410 <_sk_callback_sse41+0x3000a30>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004d80 <_sk_callback_sse41+0x3000b2e>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -22787,13 +25140,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4449 <.literal16+0x989>
+ .byte 224,7 // loopne 4db9 <.literal16+0xa89>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 444d <.literal16+0x98d>
+ .byte 224,7 // loopne 4dbd <.literal16+0xa8d>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 4451 <.literal16+0x991>
+ .byte 224,7 // loopne 4dc1 <.literal16+0xa91>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 4455 <.literal16+0x995>
+ .byte 224,7 // loopne 4dc5 <.literal16+0xa95>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -22839,13 +25192,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 44b9 <.literal16+0x9f9>
+ .byte 224,7 // loopne 4e29 <.literal16+0xaf9>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 44bd <.literal16+0x9fd>
+ .byte 224,7 // loopne 4e2d <.literal16+0xafd>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 44c1 <.literal16+0xa01>
+ .byte 224,7 // loopne 4e31 <.literal16+0xb01>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 44c5 <.literal16+0xa05>
+ .byte 224,7 // loopne 4e35 <.literal16+0xb05>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -22883,13 +25236,13 @@ BALIGN16
.byte 65,0,0 // add %al,(%r8)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 4556 <.literal16+0xa96>
+ .byte 124,66 // jl 4ec6 <.literal16+0xb96>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 455a <.literal16+0xa9a>
+ .byte 124,66 // jl 4eca <.literal16+0xb9a>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 455e <.literal16+0xa9e>
+ .byte 124,66 // jl 4ece <.literal16+0xb9e>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 4562 <.literal16+0xaa2>
+ .byte 124,66 // jl 4ed2 <.literal16+0xba2>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 0,240 // add %dh,%al
@@ -22979,13 +25332,13 @@ 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 4665 <.literal16+0xba5>
+ .byte 112,65 // jo 4fd5 <.literal16+0xca5>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4669 <.literal16+0xba9>
+ .byte 112,65 // jo 4fd9 <.literal16+0xca9>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 466d <.literal16+0xbad>
+ .byte 112,65 // jo 4fdd <.literal16+0xcad>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4671 <.literal16+0xbb1>
+ .byte 112,65 // jo 4fe1 <.literal16+0xcb1>
.byte 255,0 // incl (%rax)
.byte 0,0 // add %al,(%rax)
.byte 255,0 // incl (%rax)
@@ -23000,7 +25353,7 @@ 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) # 3004660 <_sk_callback_sse41+0x3000c80>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004fd0 <_sk_callback_sse41+0x3000d7e>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -23027,7 +25380,7 @@ 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) # 30046a0 <_sk_callback_sse41+0x3000cc0>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3005010 <_sk_callback_sse41+0x3000dbe>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -23042,11 +25395,11 @@ BALIGN16
.byte 255,0 // incl (%rax)
.byte 0,127,67 // add %bh,0x43(%rdi)
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 46fb <.literal16+0xc3b>
+ .byte 127,67 // jg 506b <.literal16+0xd3b>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 46ff <.literal16+0xc3f>
+ .byte 127,67 // jg 506f <.literal16+0xd3f>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 4703 <.literal16+0xc43>
+ .byte 127,67 // jg 5073 <.literal16+0xd43>
.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)
@@ -23122,13 +25475,13 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 255 // (bad)
- .byte 127,71 // jg 47cb <.literal16+0xd0b>
+ .byte 127,71 // jg 513b <.literal16+0xe0b>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 47cf <.literal16+0xd0f>
+ .byte 127,71 // jg 513f <.literal16+0xe0f>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 47d3 <.literal16+0xd13>
+ .byte 127,71 // jg 5143 <.literal16+0xe13>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 47d7 <.literal16+0xd17>
+ .byte 127,71 // jg 5147 <.literal16+0xe17>
.byte 208 // (bad)
.byte 179,89 // mov $0x59,%bl
.byte 62,208 // ds (bad)
@@ -23254,11 +25607,11 @@ 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 48e2 <.literal16+0xe22>
+ .byte 62,114,28 // jb,pt 5252 <.literal16+0xf22>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 48e6 <.literal16+0xe26>
+ .byte 62,114,28 // jb,pt 5256 <.literal16+0xf26>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 48ea <.literal16+0xe2a>
+ .byte 62,114,28 // jb,pt 525a <.literal16+0xf2a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -23302,7 +25655,7 @@ 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 # 3d63d775 <_sk_callback_sse41+0x3d639d95>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e0e5 <_sk_callback_sse41+0x3d639e93>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -23328,7 +25681,7 @@ 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 # 3d63d7b5 <_sk_callback_sse41+0x3d639dd5>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e125 <_sk_callback_sse41+0x3d639ed3>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
@@ -23337,13 +25690,13 @@ BALIGN16
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
.byte 63 // (bad)
- .byte 114,28 // jb 49ae <.literal16+0xeee>
+ .byte 114,28 // jb 531e <.literal16+0xfee>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 49b2 <.literal16+0xef2>
+ .byte 62,114,28 // jb,pt 5322 <.literal16+0xff2>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 49b6 <.literal16+0xef6>
+ .byte 62,114,28 // jb,pt 5326 <.literal16+0xff6>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 49ba <.literal16+0xefa>
+ .byte 62,114,28 // jb,pt 532a <.literal16+0xffa>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -23364,11 +25717,11 @@ 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 49f2 <.literal16+0xf32>
+ .byte 62,114,28 // jb,pt 5362 <.literal16+0x1032>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 49f6 <.literal16+0xf36>
+ .byte 62,114,28 // jb,pt 5366 <.literal16+0x1036>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 49fa <.literal16+0xf3a>
+ .byte 62,114,28 // jb,pt 536a <.literal16+0x103a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -23412,7 +25765,7 @@ 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 # 3d63d885 <_sk_callback_sse41+0x3d639ea5>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e1f5 <_sk_callback_sse41+0x3d639fa3>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -23438,7 +25791,7 @@ 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 # 3d63d8c5 <_sk_callback_sse41+0x3d639ee5>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e235 <_sk_callback_sse41+0x3d639fe3>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
@@ -23447,13 +25800,13 @@ BALIGN16
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
.byte 63 // (bad)
- .byte 114,28 // jb 4abe <.literal16+0xffe>
+ .byte 114,28 // jb 542e <.literal16+0x10fe>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4ac2 <_sk_callback_sse41+0x10e2>
+ .byte 62,114,28 // jb,pt 5432 <_sk_callback_sse41+0x11e0>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4ac6 <_sk_callback_sse41+0x10e6>
+ .byte 62,114,28 // jb,pt 5436 <_sk_callback_sse41+0x11e4>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4aca <_sk_callback_sse41+0x10ea>
+ .byte 62,114,28 // jb,pt 543a <_sk_callback_sse41+0x11e8>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -23523,7 +25876,7 @@ _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 15,40,21,196,62,0,0 // movaps 0x3ec4(%rip),%xmm2 # 3f40 <_sk_callback_sse2+0xe4>
+ .byte 15,40,21,148,71,0,0 // movaps 0x4794(%rip),%xmm2 # 4810 <_sk_callback_sse2+0xde>
.byte 15,88,202 // addps %xmm2,%xmm1
.byte 15,16,2 // movups (%rdx),%xmm0
.byte 15,88,193 // addps %xmm1,%xmm0
@@ -23532,7 +25885,7 @@ _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 15,40,21,179,62,0,0 // movaps 0x3eb3(%rip),%xmm2 # 3f50 <_sk_callback_sse2+0xf4>
+ .byte 15,40,21,131,71,0,0 // movaps 0x4783(%rip),%xmm2 # 4820 <_sk_callback_sse2+0xee>
.byte 15,87,219 // xorps %xmm3,%xmm3
.byte 15,87,228 // xorps %xmm4,%xmm4
.byte 15,87,237 // xorps %xmm5,%xmm5
@@ -23555,14 +25908,14 @@ _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 102,68,15,111,21,120,62,0,0 // movdqa 0x3e78(%rip),%xmm10 # 3f60 <_sk_callback_sse2+0x104>
+ .byte 102,68,15,111,21,72,71,0,0 // movdqa 0x4748(%rip),%xmm10 # 4830 <_sk_callback_sse2+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 102,68,15,111,37,100,62,0,0 // movdqa 0x3e64(%rip),%xmm12 # 3f70 <_sk_callback_sse2+0x114>
- .byte 102,68,15,111,45,107,62,0,0 // movdqa 0x3e6b(%rip),%xmm13 # 3f80 <_sk_callback_sse2+0x124>
+ .byte 102,68,15,111,37,52,71,0,0 // movdqa 0x4734(%rip),%xmm12 # 4840 <_sk_callback_sse2+0x10e>
+ .byte 102,68,15,111,45,59,71,0,0 // movdqa 0x473b(%rip),%xmm13 # 4850 <_sk_callback_sse2+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
@@ -23578,8 +25931,8 @@ _sk_dither_sse2:
.byte 102,69,15,235,245 // por %xmm13,%xmm14
.byte 102,69,15,235,240 // por %xmm8,%xmm14
.byte 69,15,91,198 // cvtdq2ps %xmm14,%xmm8
- .byte 68,15,89,5,38,62,0,0 // mulps 0x3e26(%rip),%xmm8 # 3f90 <_sk_callback_sse2+0x134>
- .byte 68,15,88,5,46,62,0,0 // addps 0x3e2e(%rip),%xmm8 # 3fa0 <_sk_callback_sse2+0x144>
+ .byte 68,15,89,5,246,70,0,0 // mulps 0x46f6(%rip),%xmm8 # 4860 <_sk_callback_sse2+0x12e>
+ .byte 68,15,88,5,254,70,0,0 // addps 0x46fe(%rip),%xmm8 # 4870 <_sk_callback_sse2+0x13e>
.byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9
.byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
.byte 69,15,89,200 // mulps %xmm8,%xmm9
@@ -23621,7 +25974,7 @@ HIDDEN _sk_srcatop_sse2
FUNCTION(_sk_srcatop_sse2)
_sk_srcatop_sse2:
.byte 15,89,199 // mulps %xmm7,%xmm0
- .byte 68,15,40,5,219,61,0,0 // movaps 0x3ddb(%rip),%xmm8 # 3fb0 <_sk_callback_sse2+0x154>
+ .byte 68,15,40,5,171,70,0,0 // movaps 0x46ab(%rip),%xmm8 # 4880 <_sk_callback_sse2+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
@@ -23646,7 +25999,7 @@ FUNCTION(_sk_dstatop_sse2)
_sk_dstatop_sse2:
.byte 68,15,40,195 // movaps %xmm3,%xmm8
.byte 68,15,89,196 // mulps %xmm4,%xmm8
- .byte 68,15,40,13,158,61,0,0 // movaps 0x3d9e(%rip),%xmm9 # 3fc0 <_sk_callback_sse2+0x164>
+ .byte 68,15,40,13,110,70,0,0 // movaps 0x466e(%rip),%xmm9 # 4890 <_sk_callback_sse2+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
@@ -23693,7 +26046,7 @@ HIDDEN _sk_srcout_sse2
.globl _sk_srcout_sse2
FUNCTION(_sk_srcout_sse2)
_sk_srcout_sse2:
- .byte 68,15,40,5,66,61,0,0 // movaps 0x3d42(%rip),%xmm8 # 3fd0 <_sk_callback_sse2+0x174>
+ .byte 68,15,40,5,18,70,0,0 // movaps 0x4612(%rip),%xmm8 # 48a0 <_sk_callback_sse2+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
@@ -23706,7 +26059,7 @@ HIDDEN _sk_dstout_sse2
.globl _sk_dstout_sse2
FUNCTION(_sk_dstout_sse2)
_sk_dstout_sse2:
- .byte 68,15,40,5,50,61,0,0 // movaps 0x3d32(%rip),%xmm8 # 3fe0 <_sk_callback_sse2+0x184>
+ .byte 68,15,40,5,2,70,0,0 // movaps 0x4602(%rip),%xmm8 # 48b0 <_sk_callback_sse2+0x17e>
.byte 68,15,92,195 // subps %xmm3,%xmm8
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 15,89,196 // mulps %xmm4,%xmm0
@@ -23723,7 +26076,7 @@ HIDDEN _sk_srcover_sse2
.globl _sk_srcover_sse2
FUNCTION(_sk_srcover_sse2)
_sk_srcover_sse2:
- .byte 68,15,40,5,21,61,0,0 // movaps 0x3d15(%rip),%xmm8 # 3ff0 <_sk_callback_sse2+0x194>
+ .byte 68,15,40,5,229,69,0,0 // movaps 0x45e5(%rip),%xmm8 # 48c0 <_sk_callback_sse2+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
@@ -23743,7 +26096,7 @@ HIDDEN _sk_dstover_sse2
.globl _sk_dstover_sse2
FUNCTION(_sk_dstover_sse2)
_sk_dstover_sse2:
- .byte 68,15,40,5,233,60,0,0 // movaps 0x3ce9(%rip),%xmm8 # 4000 <_sk_callback_sse2+0x1a4>
+ .byte 68,15,40,5,185,69,0,0 // movaps 0x45b9(%rip),%xmm8 # 48d0 <_sk_callback_sse2+0x19e>
.byte 68,15,92,199 // subps %xmm7,%xmm8
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -23771,7 +26124,7 @@ HIDDEN _sk_multiply_sse2
.globl _sk_multiply_sse2
FUNCTION(_sk_multiply_sse2)
_sk_multiply_sse2:
- .byte 68,15,40,5,189,60,0,0 // movaps 0x3cbd(%rip),%xmm8 # 4010 <_sk_callback_sse2+0x1b4>
+ .byte 68,15,40,5,141,69,0,0 // movaps 0x458d(%rip),%xmm8 # 48e0 <_sk_callback_sse2+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
@@ -23847,7 +26200,7 @@ HIDDEN _sk_xor__sse2
FUNCTION(_sk_xor__sse2)
_sk_xor__sse2:
.byte 68,15,40,195 // movaps %xmm3,%xmm8
- .byte 15,40,29,238,59,0,0 // movaps 0x3bee(%rip),%xmm3 # 4020 <_sk_callback_sse2+0x1c4>
+ .byte 15,40,29,190,68,0,0 // movaps 0x44be(%rip),%xmm3 # 48f0 <_sk_callback_sse2+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
@@ -23895,7 +26248,7 @@ _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 15,40,21,89,59,0,0 // movaps 0x3b59(%rip),%xmm2 # 4030 <_sk_callback_sse2+0x1d4>
+ .byte 15,40,21,41,68,0,0 // movaps 0x4429(%rip),%xmm2 # 4900 <_sk_callback_sse2+0x1ce>
.byte 15,92,211 // subps %xmm3,%xmm2
.byte 15,89,215 // mulps %xmm7,%xmm2
.byte 15,88,218 // addps %xmm2,%xmm3
@@ -23929,7 +26282,7 @@ _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 15,40,21,254,58,0,0 // movaps 0x3afe(%rip),%xmm2 # 4040 <_sk_callback_sse2+0x1e4>
+ .byte 15,40,21,206,67,0,0 // movaps 0x43ce(%rip),%xmm2 # 4910 <_sk_callback_sse2+0x1de>
.byte 15,92,211 // subps %xmm3,%xmm2
.byte 15,89,215 // mulps %xmm7,%xmm2
.byte 15,88,218 // addps %xmm2,%xmm3
@@ -23966,7 +26319,7 @@ _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 15,40,21,152,58,0,0 // movaps 0x3a98(%rip),%xmm2 # 4050 <_sk_callback_sse2+0x1f4>
+ .byte 15,40,21,104,67,0,0 // movaps 0x4368(%rip),%xmm2 # 4920 <_sk_callback_sse2+0x1ee>
.byte 15,92,211 // subps %xmm3,%xmm2
.byte 15,89,215 // mulps %xmm7,%xmm2
.byte 15,88,218 // addps %xmm2,%xmm3
@@ -23993,7 +26346,7 @@ _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 15,40,13,89,58,0,0 // movaps 0x3a59(%rip),%xmm1 # 4060 <_sk_callback_sse2+0x204>
+ .byte 15,40,13,41,67,0,0 // movaps 0x4329(%rip),%xmm1 # 4930 <_sk_callback_sse2+0x1fe>
.byte 15,92,203 // subps %xmm3,%xmm1
.byte 15,89,207 // mulps %xmm7,%xmm1
.byte 15,88,217 // addps %xmm1,%xmm3
@@ -24007,7 +26360,7 @@ HIDDEN _sk_colorburn_sse2
FUNCTION(_sk_colorburn_sse2)
_sk_colorburn_sse2:
.byte 68,15,40,192 // movaps %xmm0,%xmm8
- .byte 68,15,40,21,72,58,0,0 // movaps 0x3a48(%rip),%xmm10 # 4070 <_sk_callback_sse2+0x214>
+ .byte 68,15,40,21,24,67,0,0 // movaps 0x4318(%rip),%xmm10 # 4940 <_sk_callback_sse2+0x20e>
.byte 69,15,40,202 // movaps %xmm10,%xmm9
.byte 68,15,92,207 // subps %xmm7,%xmm9
.byte 69,15,40,217 // movaps %xmm9,%xmm11
@@ -24101,7 +26454,7 @@ HIDDEN _sk_colordodge_sse2
FUNCTION(_sk_colordodge_sse2)
_sk_colordodge_sse2:
.byte 68,15,40,200 // movaps %xmm0,%xmm9
- .byte 68,15,40,21,254,56,0,0 // movaps 0x38fe(%rip),%xmm10 # 4080 <_sk_callback_sse2+0x224>
+ .byte 68,15,40,21,206,65,0,0 // movaps 0x41ce(%rip),%xmm10 # 4950 <_sk_callback_sse2+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
@@ -24195,7 +26548,7 @@ _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 68,15,40,29,179,55,0,0 // movaps 0x37b3(%rip),%xmm11 # 4090 <_sk_callback_sse2+0x234>
+ .byte 68,15,40,29,131,64,0,0 // movaps 0x4083(%rip),%xmm11 # 4960 <_sk_callback_sse2+0x22e>
.byte 69,15,40,211 // movaps %xmm11,%xmm10
.byte 68,15,92,215 // subps %xmm7,%xmm10
.byte 69,15,40,194 // movaps %xmm10,%xmm8
@@ -24283,7 +26636,7 @@ FUNCTION(_sk_overlay_sse2)
_sk_overlay_sse2:
.byte 68,15,40,193 // movaps %xmm1,%xmm8
.byte 68,15,40,232 // movaps %xmm0,%xmm13
- .byte 68,15,40,13,129,54,0,0 // movaps 0x3681(%rip),%xmm9 # 40a0 <_sk_callback_sse2+0x244>
+ .byte 68,15,40,13,81,63,0,0 // movaps 0x3f51(%rip),%xmm9 # 4970 <_sk_callback_sse2+0x23e>
.byte 69,15,40,209 // movaps %xmm9,%xmm10
.byte 68,15,92,215 // subps %xmm7,%xmm10
.byte 69,15,40,218 // movaps %xmm10,%xmm11
@@ -24374,7 +26727,7 @@ _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 68,15,40,13,62,53,0,0 // movaps 0x353e(%rip),%xmm9 # 40b0 <_sk_callback_sse2+0x254>
+ .byte 68,15,40,13,14,62,0,0 // movaps 0x3e0e(%rip),%xmm9 # 4980 <_sk_callback_sse2+0x24e>
.byte 69,15,40,249 // movaps %xmm9,%xmm15
.byte 69,15,92,250 // subps %xmm10,%xmm15
.byte 69,15,40,218 // movaps %xmm10,%xmm11
@@ -24387,10 +26740,10 @@ _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 68,15,40,53,24,53,0,0 // movaps 0x3518(%rip),%xmm14 # 40c0 <_sk_callback_sse2+0x264>
+ .byte 68,15,40,53,232,61,0,0 // movaps 0x3de8(%rip),%xmm14 # 4990 <_sk_callback_sse2+0x25e>
.byte 69,15,88,222 // addps %xmm14,%xmm11
.byte 68,15,89,216 // mulps %xmm0,%xmm11
- .byte 68,15,40,21,24,53,0,0 // movaps 0x3518(%rip),%xmm10 # 40d0 <_sk_callback_sse2+0x274>
+ .byte 68,15,40,21,232,61,0,0 // movaps 0x3de8(%rip),%xmm10 # 49a0 <_sk_callback_sse2+0x26e>
.byte 69,15,89,234 // mulps %xmm10,%xmm13
.byte 69,15,88,235 // addps %xmm11,%xmm13
.byte 15,88,228 // addps %xmm4,%xmm4
@@ -24526,6 +26879,596 @@ _sk_softlight_sse2:
.byte 65,15,40,200 // movaps %xmm8,%xmm1
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_hue_sse2
+.globl _sk_hue_sse2
+FUNCTION(_sk_hue_sse2)
+_sk_hue_sse2:
+ .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp)
+ .byte 15,40,254 // movaps %xmm6,%xmm7
+ .byte 15,40,245 // movaps %xmm5,%xmm6
+ .byte 15,40,236 // movaps %xmm4,%xmm5
+ .byte 15,40,227 // movaps %xmm3,%xmm4
+ .byte 69,15,87,192 // xorps %xmm8,%xmm8
+ .byte 68,15,40,13,250,59,0,0 // movaps 0x3bfa(%rip),%xmm9 # 49b0 <_sk_callback_sse2+0x27e>
+ .byte 69,15,40,209 // movaps %xmm9,%xmm10
+ .byte 68,15,94,212 // divps %xmm4,%xmm10
+ .byte 68,15,40,228 // movaps %xmm4,%xmm12
+ .byte 69,15,194,224,4 // cmpneqps %xmm8,%xmm12
+ .byte 69,15,84,226 // andps %xmm10,%xmm12
+ .byte 65,15,89,196 // mulps %xmm12,%xmm0
+ .byte 65,15,89,204 // mulps %xmm12,%xmm1
+ .byte 68,15,89,226 // mulps %xmm2,%xmm12
+ .byte 68,15,40,214 // movaps %xmm6,%xmm10
+ .byte 68,15,40,238 // movaps %xmm6,%xmm13
+ .byte 68,15,40,254 // movaps %xmm6,%xmm15
+ .byte 68,15,95,255 // maxps %xmm7,%xmm15
+ .byte 68,15,40,221 // movaps %xmm5,%xmm11
+ .byte 68,15,40,245 // movaps %xmm5,%xmm14
+ .byte 15,40,213 // movaps %xmm5,%xmm2
+ .byte 65,15,95,215 // maxps %xmm15,%xmm2
+ .byte 68,15,93,215 // minps %xmm7,%xmm10
+ .byte 69,15,93,218 // minps %xmm10,%xmm11
+ .byte 65,15,92,211 // subps %xmm11,%xmm2
+ .byte 68,15,40,209 // movaps %xmm1,%xmm10
+ .byte 69,15,93,212 // minps %xmm12,%xmm10
+ .byte 68,15,40,216 // movaps %xmm0,%xmm11
+ .byte 69,15,93,218 // minps %xmm10,%xmm11
+ .byte 68,15,40,209 // movaps %xmm1,%xmm10
+ .byte 69,15,95,212 // maxps %xmm12,%xmm10
+ .byte 68,15,40,248 // movaps %xmm0,%xmm15
+ .byte 69,15,95,250 // maxps %xmm10,%xmm15
+ .byte 69,15,92,251 // subps %xmm11,%xmm15
+ .byte 65,15,92,195 // subps %xmm11,%xmm0
+ .byte 15,89,194 // mulps %xmm2,%xmm0
+ .byte 65,15,94,199 // divps %xmm15,%xmm0
+ .byte 65,15,92,203 // subps %xmm11,%xmm1
+ .byte 15,89,202 // mulps %xmm2,%xmm1
+ .byte 65,15,94,207 // divps %xmm15,%xmm1
+ .byte 69,15,92,227 // subps %xmm11,%xmm12
+ .byte 68,15,89,226 // mulps %xmm2,%xmm12
+ .byte 69,15,94,231 // divps %xmm15,%xmm12
+ .byte 69,15,194,248,4 // cmpneqps %xmm8,%xmm15
+ .byte 65,15,84,199 // andps %xmm15,%xmm0
+ .byte 65,15,84,207 // andps %xmm15,%xmm1
+ .byte 69,15,84,231 // andps %xmm15,%xmm12
+ .byte 68,15,40,61,95,59,0,0 // movaps 0x3b5f(%rip),%xmm15 # 49c0 <_sk_callback_sse2+0x28e>
+ .byte 69,15,89,247 // mulps %xmm15,%xmm14
+ .byte 15,40,29,100,59,0,0 // movaps 0x3b64(%rip),%xmm3 # 49d0 <_sk_callback_sse2+0x29e>
+ .byte 68,15,89,235 // mulps %xmm3,%xmm13
+ .byte 69,15,88,238 // addps %xmm14,%xmm13
+ .byte 68,15,40,21,100,59,0,0 // movaps 0x3b64(%rip),%xmm10 # 49e0 <_sk_callback_sse2+0x2ae>
+ .byte 68,15,40,223 // movaps %xmm7,%xmm11
+ .byte 69,15,89,218 // mulps %xmm10,%xmm11
+ .byte 69,15,88,221 // addps %xmm13,%xmm11
+ .byte 68,15,40,232 // movaps %xmm0,%xmm13
+ .byte 69,15,89,239 // mulps %xmm15,%xmm13
+ .byte 68,15,40,241 // movaps %xmm1,%xmm14
+ .byte 68,15,89,243 // mulps %xmm3,%xmm14
+ .byte 69,15,88,245 // addps %xmm13,%xmm14
+ .byte 65,15,40,212 // movaps %xmm12,%xmm2
+ .byte 65,15,89,210 // mulps %xmm10,%xmm2
+ .byte 65,15,88,214 // addps %xmm14,%xmm2
+ .byte 68,15,92,218 // subps %xmm2,%xmm11
+ .byte 65,15,88,195 // addps %xmm11,%xmm0
+ .byte 65,15,88,203 // addps %xmm11,%xmm1
+ .byte 69,15,88,220 // addps %xmm12,%xmm11
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 65,15,93,211 // minps %xmm11,%xmm2
+ .byte 68,15,40,232 // movaps %xmm0,%xmm13
+ .byte 68,15,93,234 // minps %xmm2,%xmm13
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 65,15,95,211 // maxps %xmm11,%xmm2
+ .byte 68,15,40,224 // movaps %xmm0,%xmm12
+ .byte 68,15,95,226 // maxps %xmm2,%xmm12
+ .byte 68,15,89,248 // mulps %xmm0,%xmm15
+ .byte 15,89,217 // mulps %xmm1,%xmm3
+ .byte 65,15,88,223 // addps %xmm15,%xmm3
+ .byte 69,15,89,211 // mulps %xmm11,%xmm10
+ .byte 68,15,88,211 // addps %xmm3,%xmm10
+ .byte 15,87,210 // xorps %xmm2,%xmm2
+ .byte 65,15,194,213,2 // cmpleps %xmm13,%xmm2
+ .byte 15,40,216 // movaps %xmm0,%xmm3
+ .byte 65,15,92,218 // subps %xmm10,%xmm3
+ .byte 65,15,89,218 // mulps %xmm10,%xmm3
+ .byte 69,15,40,242 // movaps %xmm10,%xmm14
+ .byte 69,15,92,245 // subps %xmm13,%xmm14
+ .byte 65,15,94,222 // divps %xmm14,%xmm3
+ .byte 65,15,88,218 // addps %xmm10,%xmm3
+ .byte 68,15,40,234 // movaps %xmm2,%xmm13
+ .byte 68,15,85,235 // andnps %xmm3,%xmm13
+ .byte 15,84,194 // andps %xmm2,%xmm0
+ .byte 65,15,86,197 // orps %xmm13,%xmm0
+ .byte 69,15,40,233 // movaps %xmm9,%xmm13
+ .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13
+ .byte 65,15,40,221 // movaps %xmm13,%xmm3
+ .byte 15,85,216 // andnps %xmm0,%xmm3
+ .byte 65,15,92,194 // subps %xmm10,%xmm0
+ .byte 69,15,92,202 // subps %xmm10,%xmm9
+ .byte 65,15,89,193 // mulps %xmm9,%xmm0
+ .byte 69,15,92,226 // subps %xmm10,%xmm12
+ .byte 65,15,94,196 // divps %xmm12,%xmm0
+ .byte 65,15,88,194 // addps %xmm10,%xmm0
+ .byte 65,15,84,197 // andps %xmm13,%xmm0
+ .byte 15,86,195 // orps %xmm3,%xmm0
+ .byte 15,40,217 // movaps %xmm1,%xmm3
+ .byte 65,15,92,218 // subps %xmm10,%xmm3
+ .byte 65,15,89,218 // mulps %xmm10,%xmm3
+ .byte 65,15,94,222 // divps %xmm14,%xmm3
+ .byte 65,15,88,218 // addps %xmm10,%xmm3
+ .byte 68,15,40,250 // movaps %xmm2,%xmm15
+ .byte 68,15,85,251 // andnps %xmm3,%xmm15
+ .byte 15,84,202 // andps %xmm2,%xmm1
+ .byte 65,15,86,207 // orps %xmm15,%xmm1
+ .byte 65,15,40,221 // movaps %xmm13,%xmm3
+ .byte 15,85,217 // andnps %xmm1,%xmm3
+ .byte 65,15,92,202 // subps %xmm10,%xmm1
+ .byte 65,15,89,201 // mulps %xmm9,%xmm1
+ .byte 65,15,94,204 // divps %xmm12,%xmm1
+ .byte 65,15,88,202 // addps %xmm10,%xmm1
+ .byte 65,15,84,205 // andps %xmm13,%xmm1
+ .byte 15,86,203 // orps %xmm3,%xmm1
+ .byte 65,15,40,219 // movaps %xmm11,%xmm3
+ .byte 65,15,92,218 // subps %xmm10,%xmm3
+ .byte 65,15,89,218 // mulps %xmm10,%xmm3
+ .byte 65,15,94,222 // divps %xmm14,%xmm3
+ .byte 65,15,88,218 // addps %xmm10,%xmm3
+ .byte 68,15,84,218 // andps %xmm2,%xmm11
+ .byte 15,85,211 // andnps %xmm3,%xmm2
+ .byte 65,15,86,211 // orps %xmm11,%xmm2
+ .byte 65,15,40,221 // movaps %xmm13,%xmm3
+ .byte 15,85,218 // andnps %xmm2,%xmm3
+ .byte 65,15,92,210 // subps %xmm10,%xmm2
+ .byte 65,15,89,209 // mulps %xmm9,%xmm2
+ .byte 65,15,94,212 // divps %xmm12,%xmm2
+ .byte 65,15,88,210 // addps %xmm10,%xmm2
+ .byte 65,15,84,213 // andps %xmm13,%xmm2
+ .byte 15,86,211 // orps %xmm3,%xmm2
+ .byte 65,15,95,192 // maxps %xmm8,%xmm0
+ .byte 65,15,95,200 // maxps %xmm8,%xmm1
+ .byte 65,15,95,208 // maxps %xmm8,%xmm2
+ .byte 68,15,40,196 // movaps %xmm4,%xmm8
+ .byte 68,15,40,76,36,232 // movaps -0x18(%rsp),%xmm9
+ .byte 69,15,88,193 // addps %xmm9,%xmm8
+ .byte 65,15,89,225 // mulps %xmm9,%xmm4
+ .byte 68,15,92,196 // subps %xmm4,%xmm8
+ .byte 65,15,89,192 // mulps %xmm8,%xmm0
+ .byte 65,15,89,200 // mulps %xmm8,%xmm1
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,229 // movaps %xmm5,%xmm4
+ .byte 15,40,238 // movaps %xmm6,%xmm5
+ .byte 15,40,247 // movaps %xmm7,%xmm6
+ .byte 65,15,40,216 // movaps %xmm8,%xmm3
+ .byte 65,15,40,249 // movaps %xmm9,%xmm7
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_saturation_sse2
+.globl _sk_saturation_sse2
+FUNCTION(_sk_saturation_sse2)
+_sk_saturation_sse2:
+ .byte 68,15,40,214 // movaps %xmm6,%xmm10
+ .byte 15,40,245 // movaps %xmm5,%xmm6
+ .byte 15,40,236 // movaps %xmm4,%xmm5
+ .byte 15,40,227 // movaps %xmm3,%xmm4
+ .byte 68,15,40,193 // movaps %xmm1,%xmm8
+ .byte 68,15,40,200 // movaps %xmm0,%xmm9
+ .byte 15,87,201 // xorps %xmm1,%xmm1
+ .byte 68,15,40,29,198,57,0,0 // movaps 0x39c6(%rip),%xmm11 # 49f0 <_sk_callback_sse2+0x2be>
+ .byte 65,15,40,195 // movaps %xmm11,%xmm0
+ .byte 15,94,199 // divps %xmm7,%xmm0
+ .byte 68,15,40,231 // movaps %xmm7,%xmm12
+ .byte 68,15,194,225,4 // cmpneqps %xmm1,%xmm12
+ .byte 68,15,84,224 // andps %xmm0,%xmm12
+ .byte 65,15,40,196 // movaps %xmm12,%xmm0
+ .byte 15,89,197 // mulps %xmm5,%xmm0
+ .byte 69,15,40,252 // movaps %xmm12,%xmm15
+ .byte 68,15,89,254 // mulps %xmm6,%xmm15
+ .byte 69,15,89,226 // mulps %xmm10,%xmm12
+ .byte 69,15,40,240 // movaps %xmm8,%xmm14
+ .byte 68,15,95,242 // maxps %xmm2,%xmm14
+ .byte 69,15,40,233 // movaps %xmm9,%xmm13
+ .byte 69,15,95,238 // maxps %xmm14,%xmm13
+ .byte 68,15,93,194 // minps %xmm2,%xmm8
+ .byte 69,15,93,200 // minps %xmm8,%xmm9
+ .byte 69,15,92,233 // subps %xmm9,%xmm13
+ .byte 65,15,40,215 // movaps %xmm15,%xmm2
+ .byte 65,15,93,212 // minps %xmm12,%xmm2
+ .byte 68,15,40,192 // movaps %xmm0,%xmm8
+ .byte 68,15,93,194 // minps %xmm2,%xmm8
+ .byte 69,15,40,207 // movaps %xmm15,%xmm9
+ .byte 69,15,95,204 // maxps %xmm12,%xmm9
+ .byte 15,40,208 // movaps %xmm0,%xmm2
+ .byte 65,15,95,209 // maxps %xmm9,%xmm2
+ .byte 65,15,92,208 // subps %xmm8,%xmm2
+ .byte 65,15,92,192 // subps %xmm8,%xmm0
+ .byte 65,15,89,197 // mulps %xmm13,%xmm0
+ .byte 15,94,194 // divps %xmm2,%xmm0
+ .byte 69,15,92,248 // subps %xmm8,%xmm15
+ .byte 69,15,89,253 // mulps %xmm13,%xmm15
+ .byte 68,15,94,250 // divps %xmm2,%xmm15
+ .byte 69,15,92,224 // subps %xmm8,%xmm12
+ .byte 69,15,89,229 // mulps %xmm13,%xmm12
+ .byte 68,15,94,226 // divps %xmm2,%xmm12
+ .byte 15,194,209,4 // cmpneqps %xmm1,%xmm2
+ .byte 15,84,194 // andps %xmm2,%xmm0
+ .byte 68,15,84,250 // andps %xmm2,%xmm15
+ .byte 68,15,84,226 // andps %xmm2,%xmm12
+ .byte 68,15,40,45,54,57,0,0 // movaps 0x3936(%rip),%xmm13 # 4a00 <_sk_callback_sse2+0x2ce>
+ .byte 68,15,40,197 // movaps %xmm5,%xmm8
+ .byte 69,15,89,197 // mulps %xmm13,%xmm8
+ .byte 68,15,40,53,54,57,0,0 // movaps 0x3936(%rip),%xmm14 # 4a10 <_sk_callback_sse2+0x2de>
+ .byte 15,40,214 // movaps %xmm6,%xmm2
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 65,15,88,208 // addps %xmm8,%xmm2
+ .byte 68,15,40,5,51,57,0,0 // movaps 0x3933(%rip),%xmm8 # 4a20 <_sk_callback_sse2+0x2ee>
+ .byte 69,15,40,202 // movaps %xmm10,%xmm9
+ .byte 69,15,89,200 // mulps %xmm8,%xmm9
+ .byte 68,15,88,202 // addps %xmm2,%xmm9
+ .byte 15,40,208 // movaps %xmm0,%xmm2
+ .byte 65,15,89,213 // mulps %xmm13,%xmm2
+ .byte 65,15,40,207 // movaps %xmm15,%xmm1
+ .byte 65,15,89,206 // mulps %xmm14,%xmm1
+ .byte 15,88,202 // addps %xmm2,%xmm1
+ .byte 65,15,40,212 // movaps %xmm12,%xmm2
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 15,88,209 // addps %xmm1,%xmm2
+ .byte 68,15,92,202 // subps %xmm2,%xmm9
+ .byte 65,15,88,193 // addps %xmm9,%xmm0
+ .byte 69,15,88,249 // addps %xmm9,%xmm15
+ .byte 69,15,88,204 // addps %xmm12,%xmm9
+ .byte 65,15,40,207 // movaps %xmm15,%xmm1
+ .byte 65,15,93,201 // minps %xmm9,%xmm1
+ .byte 68,15,40,224 // movaps %xmm0,%xmm12
+ .byte 68,15,93,225 // minps %xmm1,%xmm12
+ .byte 65,15,40,215 // movaps %xmm15,%xmm2
+ .byte 65,15,95,209 // maxps %xmm9,%xmm2
+ .byte 15,40,216 // movaps %xmm0,%xmm3
+ .byte 15,95,218 // maxps %xmm2,%xmm3
+ .byte 68,15,89,232 // mulps %xmm0,%xmm13
+ .byte 69,15,89,247 // mulps %xmm15,%xmm14
+ .byte 69,15,88,245 // addps %xmm13,%xmm14
+ .byte 69,15,89,193 // mulps %xmm9,%xmm8
+ .byte 69,15,88,198 // addps %xmm14,%xmm8
+ .byte 15,87,210 // xorps %xmm2,%xmm2
+ .byte 65,15,194,212,2 // cmpleps %xmm12,%xmm2
+ .byte 68,15,40,232 // movaps %xmm0,%xmm13
+ .byte 69,15,92,232 // subps %xmm8,%xmm13
+ .byte 69,15,89,232 // mulps %xmm8,%xmm13
+ .byte 69,15,40,240 // movaps %xmm8,%xmm14
+ .byte 69,15,92,244 // subps %xmm12,%xmm14
+ .byte 69,15,94,238 // divps %xmm14,%xmm13
+ .byte 69,15,88,232 // addps %xmm8,%xmm13
+ .byte 68,15,40,226 // movaps %xmm2,%xmm12
+ .byte 69,15,85,229 // andnps %xmm13,%xmm12
+ .byte 15,84,194 // andps %xmm2,%xmm0
+ .byte 65,15,86,196 // orps %xmm12,%xmm0
+ .byte 69,15,40,235 // movaps %xmm11,%xmm13
+ .byte 68,15,194,235,1 // cmpltps %xmm3,%xmm13
+ .byte 69,15,40,229 // movaps %xmm13,%xmm12
+ .byte 68,15,85,224 // andnps %xmm0,%xmm12
+ .byte 65,15,92,192 // subps %xmm8,%xmm0
+ .byte 69,15,92,216 // subps %xmm8,%xmm11
+ .byte 65,15,89,195 // mulps %xmm11,%xmm0
+ .byte 65,15,92,216 // subps %xmm8,%xmm3
+ .byte 15,94,195 // divps %xmm3,%xmm0
+ .byte 65,15,88,192 // addps %xmm8,%xmm0
+ .byte 65,15,84,197 // andps %xmm13,%xmm0
+ .byte 65,15,86,196 // orps %xmm12,%xmm0
+ .byte 69,15,40,231 // movaps %xmm15,%xmm12
+ .byte 69,15,92,224 // subps %xmm8,%xmm12
+ .byte 69,15,89,224 // mulps %xmm8,%xmm12
+ .byte 69,15,94,230 // divps %xmm14,%xmm12
+ .byte 69,15,88,224 // addps %xmm8,%xmm12
+ .byte 15,40,202 // movaps %xmm2,%xmm1
+ .byte 65,15,85,204 // andnps %xmm12,%xmm1
+ .byte 68,15,84,250 // andps %xmm2,%xmm15
+ .byte 68,15,86,249 // orps %xmm1,%xmm15
+ .byte 65,15,40,205 // movaps %xmm13,%xmm1
+ .byte 65,15,85,207 // andnps %xmm15,%xmm1
+ .byte 69,15,92,248 // subps %xmm8,%xmm15
+ .byte 69,15,89,251 // mulps %xmm11,%xmm15
+ .byte 68,15,94,251 // divps %xmm3,%xmm15
+ .byte 69,15,88,248 // addps %xmm8,%xmm15
+ .byte 69,15,84,253 // andps %xmm13,%xmm15
+ .byte 68,15,86,249 // orps %xmm1,%xmm15
+ .byte 65,15,40,201 // movaps %xmm9,%xmm1
+ .byte 65,15,92,200 // subps %xmm8,%xmm1
+ .byte 65,15,89,200 // mulps %xmm8,%xmm1
+ .byte 65,15,94,206 // divps %xmm14,%xmm1
+ .byte 65,15,88,200 // addps %xmm8,%xmm1
+ .byte 68,15,84,202 // andps %xmm2,%xmm9
+ .byte 15,85,209 // andnps %xmm1,%xmm2
+ .byte 65,15,86,209 // orps %xmm9,%xmm2
+ .byte 65,15,40,205 // movaps %xmm13,%xmm1
+ .byte 15,85,202 // andnps %xmm2,%xmm1
+ .byte 65,15,92,208 // subps %xmm8,%xmm2
+ .byte 65,15,89,211 // mulps %xmm11,%xmm2
+ .byte 15,94,211 // divps %xmm3,%xmm2
+ .byte 65,15,88,208 // addps %xmm8,%xmm2
+ .byte 65,15,84,213 // andps %xmm13,%xmm2
+ .byte 15,86,209 // orps %xmm1,%xmm2
+ .byte 15,87,201 // xorps %xmm1,%xmm1
+ .byte 15,95,193 // maxps %xmm1,%xmm0
+ .byte 68,15,95,249 // maxps %xmm1,%xmm15
+ .byte 15,95,209 // maxps %xmm1,%xmm2
+ .byte 68,15,40,196 // movaps %xmm4,%xmm8
+ .byte 68,15,88,199 // addps %xmm7,%xmm8
+ .byte 15,89,231 // mulps %xmm7,%xmm4
+ .byte 68,15,92,196 // subps %xmm4,%xmm8
+ .byte 65,15,89,192 // mulps %xmm8,%xmm0
+ .byte 69,15,89,248 // mulps %xmm8,%xmm15
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,229 // movaps %xmm5,%xmm4
+ .byte 15,40,238 // movaps %xmm6,%xmm5
+ .byte 65,15,40,242 // movaps %xmm10,%xmm6
+ .byte 65,15,40,207 // movaps %xmm15,%xmm1
+ .byte 65,15,40,216 // movaps %xmm8,%xmm3
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_color_sse2
+.globl _sk_color_sse2
+FUNCTION(_sk_color_sse2)
+_sk_color_sse2:
+ .byte 68,15,40,221 // movaps %xmm5,%xmm11
+ .byte 68,15,41,92,36,232 // movaps %xmm11,-0x18(%rsp)
+ .byte 15,40,236 // movaps %xmm4,%xmm5
+ .byte 15,40,227 // movaps %xmm3,%xmm4
+ .byte 68,15,40,249 // movaps %xmm1,%xmm15
+ .byte 69,15,87,192 // xorps %xmm8,%xmm8
+ .byte 68,15,40,13,152,55,0,0 // movaps 0x3798(%rip),%xmm9 # 4a30 <_sk_callback_sse2+0x2fe>
+ .byte 69,15,40,209 // movaps %xmm9,%xmm10
+ .byte 68,15,94,212 // divps %xmm4,%xmm10
+ .byte 68,15,40,228 // movaps %xmm4,%xmm12
+ .byte 69,15,194,224,4 // cmpneqps %xmm8,%xmm12
+ .byte 69,15,84,226 // andps %xmm10,%xmm12
+ .byte 65,15,89,196 // mulps %xmm12,%xmm0
+ .byte 69,15,89,252 // mulps %xmm12,%xmm15
+ .byte 68,15,89,226 // mulps %xmm2,%xmm12
+ .byte 68,15,40,45,127,55,0,0 // movaps 0x377f(%rip),%xmm13 # 4a40 <_sk_callback_sse2+0x30e>
+ .byte 68,15,40,213 // movaps %xmm5,%xmm10
+ .byte 69,15,89,213 // mulps %xmm13,%xmm10
+ .byte 68,15,40,53,127,55,0,0 // movaps 0x377f(%rip),%xmm14 # 4a50 <_sk_callback_sse2+0x31e>
+ .byte 65,15,40,211 // movaps %xmm11,%xmm2
+ .byte 65,15,89,214 // mulps %xmm14,%xmm2
+ .byte 65,15,88,210 // addps %xmm10,%xmm2
+ .byte 68,15,40,21,123,55,0,0 // movaps 0x377b(%rip),%xmm10 # 4a60 <_sk_callback_sse2+0x32e>
+ .byte 68,15,40,222 // movaps %xmm6,%xmm11
+ .byte 69,15,89,218 // mulps %xmm10,%xmm11
+ .byte 68,15,88,218 // addps %xmm2,%xmm11
+ .byte 15,40,208 // movaps %xmm0,%xmm2
+ .byte 65,15,89,213 // mulps %xmm13,%xmm2
+ .byte 65,15,40,207 // movaps %xmm15,%xmm1
+ .byte 65,15,89,206 // mulps %xmm14,%xmm1
+ .byte 15,88,202 // addps %xmm2,%xmm1
+ .byte 65,15,40,212 // movaps %xmm12,%xmm2
+ .byte 65,15,89,210 // mulps %xmm10,%xmm2
+ .byte 15,88,209 // addps %xmm1,%xmm2
+ .byte 68,15,92,218 // subps %xmm2,%xmm11
+ .byte 65,15,88,195 // addps %xmm11,%xmm0
+ .byte 69,15,88,251 // addps %xmm11,%xmm15
+ .byte 69,15,88,220 // addps %xmm12,%xmm11
+ .byte 65,15,40,207 // movaps %xmm15,%xmm1
+ .byte 65,15,93,203 // minps %xmm11,%xmm1
+ .byte 68,15,40,224 // movaps %xmm0,%xmm12
+ .byte 68,15,93,225 // minps %xmm1,%xmm12
+ .byte 65,15,40,215 // movaps %xmm15,%xmm2
+ .byte 65,15,95,211 // maxps %xmm11,%xmm2
+ .byte 15,40,216 // movaps %xmm0,%xmm3
+ .byte 15,95,218 // maxps %xmm2,%xmm3
+ .byte 68,15,89,232 // mulps %xmm0,%xmm13
+ .byte 69,15,89,247 // mulps %xmm15,%xmm14
+ .byte 69,15,88,245 // addps %xmm13,%xmm14
+ .byte 69,15,89,211 // mulps %xmm11,%xmm10
+ .byte 69,15,88,214 // addps %xmm14,%xmm10
+ .byte 15,87,210 // xorps %xmm2,%xmm2
+ .byte 65,15,194,212,2 // cmpleps %xmm12,%xmm2
+ .byte 68,15,40,232 // movaps %xmm0,%xmm13
+ .byte 69,15,92,234 // subps %xmm10,%xmm13
+ .byte 69,15,89,234 // mulps %xmm10,%xmm13
+ .byte 69,15,40,242 // movaps %xmm10,%xmm14
+ .byte 69,15,92,244 // subps %xmm12,%xmm14
+ .byte 69,15,94,238 // divps %xmm14,%xmm13
+ .byte 69,15,88,234 // addps %xmm10,%xmm13
+ .byte 68,15,40,226 // movaps %xmm2,%xmm12
+ .byte 69,15,85,229 // andnps %xmm13,%xmm12
+ .byte 15,84,194 // andps %xmm2,%xmm0
+ .byte 65,15,86,196 // orps %xmm12,%xmm0
+ .byte 69,15,40,233 // movaps %xmm9,%xmm13
+ .byte 68,15,194,235,1 // cmpltps %xmm3,%xmm13
+ .byte 69,15,40,229 // movaps %xmm13,%xmm12
+ .byte 68,15,85,224 // andnps %xmm0,%xmm12
+ .byte 65,15,92,194 // subps %xmm10,%xmm0
+ .byte 69,15,92,202 // subps %xmm10,%xmm9
+ .byte 65,15,89,193 // mulps %xmm9,%xmm0
+ .byte 65,15,92,218 // subps %xmm10,%xmm3
+ .byte 15,94,195 // divps %xmm3,%xmm0
+ .byte 65,15,88,194 // addps %xmm10,%xmm0
+ .byte 65,15,84,197 // andps %xmm13,%xmm0
+ .byte 65,15,86,196 // orps %xmm12,%xmm0
+ .byte 69,15,40,231 // movaps %xmm15,%xmm12
+ .byte 69,15,92,226 // subps %xmm10,%xmm12
+ .byte 69,15,89,226 // mulps %xmm10,%xmm12
+ .byte 69,15,94,230 // divps %xmm14,%xmm12
+ .byte 69,15,88,226 // addps %xmm10,%xmm12
+ .byte 15,40,202 // movaps %xmm2,%xmm1
+ .byte 65,15,85,204 // andnps %xmm12,%xmm1
+ .byte 68,15,84,250 // andps %xmm2,%xmm15
+ .byte 68,15,86,249 // orps %xmm1,%xmm15
+ .byte 65,15,40,205 // movaps %xmm13,%xmm1
+ .byte 65,15,85,207 // andnps %xmm15,%xmm1
+ .byte 69,15,92,250 // subps %xmm10,%xmm15
+ .byte 69,15,89,249 // mulps %xmm9,%xmm15
+ .byte 68,15,94,251 // divps %xmm3,%xmm15
+ .byte 69,15,88,250 // addps %xmm10,%xmm15
+ .byte 69,15,84,253 // andps %xmm13,%xmm15
+ .byte 68,15,86,249 // orps %xmm1,%xmm15
+ .byte 65,15,40,203 // movaps %xmm11,%xmm1
+ .byte 65,15,92,202 // subps %xmm10,%xmm1
+ .byte 65,15,89,202 // mulps %xmm10,%xmm1
+ .byte 65,15,94,206 // divps %xmm14,%xmm1
+ .byte 65,15,88,202 // addps %xmm10,%xmm1
+ .byte 68,15,84,218 // andps %xmm2,%xmm11
+ .byte 15,85,209 // andnps %xmm1,%xmm2
+ .byte 65,15,86,211 // orps %xmm11,%xmm2
+ .byte 65,15,40,205 // movaps %xmm13,%xmm1
+ .byte 15,85,202 // andnps %xmm2,%xmm1
+ .byte 65,15,92,210 // subps %xmm10,%xmm2
+ .byte 65,15,89,209 // mulps %xmm9,%xmm2
+ .byte 15,94,211 // divps %xmm3,%xmm2
+ .byte 65,15,88,210 // addps %xmm10,%xmm2
+ .byte 65,15,84,213 // andps %xmm13,%xmm2
+ .byte 15,86,209 // orps %xmm1,%xmm2
+ .byte 65,15,95,192 // maxps %xmm8,%xmm0
+ .byte 69,15,95,248 // maxps %xmm8,%xmm15
+ .byte 65,15,95,208 // maxps %xmm8,%xmm2
+ .byte 68,15,40,196 // movaps %xmm4,%xmm8
+ .byte 68,15,88,199 // addps %xmm7,%xmm8
+ .byte 15,89,231 // mulps %xmm7,%xmm4
+ .byte 68,15,92,196 // subps %xmm4,%xmm8
+ .byte 65,15,89,192 // mulps %xmm8,%xmm0
+ .byte 69,15,89,248 // mulps %xmm8,%xmm15
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,229 // movaps %xmm5,%xmm4
+ .byte 15,40,108,36,232 // movaps -0x18(%rsp),%xmm5
+ .byte 65,15,40,207 // movaps %xmm15,%xmm1
+ .byte 65,15,40,216 // movaps %xmm8,%xmm3
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_luminosity_sse2
+.globl _sk_luminosity_sse2
+FUNCTION(_sk_luminosity_sse2)
+_sk_luminosity_sse2:
+ .byte 68,15,40,238 // movaps %xmm6,%xmm13
+ .byte 68,15,41,108,36,232 // movaps %xmm13,-0x18(%rsp)
+ .byte 15,40,245 // movaps %xmm5,%xmm6
+ .byte 15,40,236 // movaps %xmm4,%xmm5
+ .byte 15,40,227 // movaps %xmm3,%xmm4
+ .byte 15,40,218 // movaps %xmm2,%xmm3
+ .byte 68,15,40,193 // movaps %xmm1,%xmm8
+ .byte 68,15,40,200 // movaps %xmm0,%xmm9
+ .byte 69,15,87,210 // xorps %xmm10,%xmm10
+ .byte 68,15,40,29,217,53,0,0 // movaps 0x35d9(%rip),%xmm11 # 4a70 <_sk_callback_sse2+0x33e>
+ .byte 65,15,40,195 // movaps %xmm11,%xmm0
+ .byte 15,94,199 // divps %xmm7,%xmm0
+ .byte 68,15,40,231 // movaps %xmm7,%xmm12
+ .byte 69,15,194,226,4 // cmpneqps %xmm10,%xmm12
+ .byte 68,15,84,224 // andps %xmm0,%xmm12
+ .byte 65,15,40,196 // movaps %xmm12,%xmm0
+ .byte 15,89,197 // mulps %xmm5,%xmm0
+ .byte 65,15,40,204 // movaps %xmm12,%xmm1
+ .byte 15,89,206 // mulps %xmm6,%xmm1
+ .byte 69,15,89,229 // mulps %xmm13,%xmm12
+ .byte 68,15,40,53,187,53,0,0 // movaps 0x35bb(%rip),%xmm14 # 4a80 <_sk_callback_sse2+0x34e>
+ .byte 69,15,89,206 // mulps %xmm14,%xmm9
+ .byte 68,15,40,45,191,53,0,0 // movaps 0x35bf(%rip),%xmm13 # 4a90 <_sk_callback_sse2+0x35e>
+ .byte 69,15,89,197 // mulps %xmm13,%xmm8
+ .byte 69,15,88,193 // addps %xmm9,%xmm8
+ .byte 68,15,40,13,191,53,0,0 // movaps 0x35bf(%rip),%xmm9 # 4aa0 <_sk_callback_sse2+0x36e>
+ .byte 65,15,89,217 // mulps %xmm9,%xmm3
+ .byte 65,15,88,216 // addps %xmm8,%xmm3
+ .byte 68,15,40,192 // movaps %xmm0,%xmm8
+ .byte 69,15,89,198 // mulps %xmm14,%xmm8
+ .byte 68,15,40,249 // movaps %xmm1,%xmm15
+ .byte 69,15,89,253 // mulps %xmm13,%xmm15
+ .byte 69,15,88,248 // addps %xmm8,%xmm15
+ .byte 69,15,40,196 // movaps %xmm12,%xmm8
+ .byte 69,15,89,193 // mulps %xmm9,%xmm8
+ .byte 69,15,88,199 // addps %xmm15,%xmm8
+ .byte 65,15,92,216 // subps %xmm8,%xmm3
+ .byte 15,88,195 // addps %xmm3,%xmm0
+ .byte 15,88,203 // addps %xmm3,%xmm1
+ .byte 65,15,88,220 // addps %xmm12,%xmm3
+ .byte 68,15,40,193 // movaps %xmm1,%xmm8
+ .byte 68,15,93,195 // minps %xmm3,%xmm8
+ .byte 68,15,40,248 // movaps %xmm0,%xmm15
+ .byte 69,15,93,248 // minps %xmm8,%xmm15
+ .byte 68,15,40,193 // movaps %xmm1,%xmm8
+ .byte 68,15,95,195 // maxps %xmm3,%xmm8
+ .byte 68,15,40,224 // movaps %xmm0,%xmm12
+ .byte 69,15,95,224 // maxps %xmm8,%xmm12
+ .byte 68,15,89,240 // mulps %xmm0,%xmm14
+ .byte 68,15,89,233 // mulps %xmm1,%xmm13
+ .byte 69,15,88,238 // addps %xmm14,%xmm13
+ .byte 68,15,89,203 // mulps %xmm3,%xmm9
+ .byte 69,15,88,205 // addps %xmm13,%xmm9
+ .byte 69,15,87,192 // xorps %xmm8,%xmm8
+ .byte 69,15,194,199,2 // cmpleps %xmm15,%xmm8
+ .byte 68,15,40,232 // movaps %xmm0,%xmm13
+ .byte 69,15,92,233 // subps %xmm9,%xmm13
+ .byte 69,15,89,233 // mulps %xmm9,%xmm13
+ .byte 69,15,40,241 // movaps %xmm9,%xmm14
+ .byte 69,15,92,247 // subps %xmm15,%xmm14
+ .byte 69,15,94,238 // divps %xmm14,%xmm13
+ .byte 69,15,88,233 // addps %xmm9,%xmm13
+ .byte 69,15,40,248 // movaps %xmm8,%xmm15
+ .byte 69,15,85,253 // andnps %xmm13,%xmm15
+ .byte 65,15,84,192 // andps %xmm8,%xmm0
+ .byte 65,15,86,199 // orps %xmm15,%xmm0
+ .byte 69,15,40,235 // movaps %xmm11,%xmm13
+ .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13
+ .byte 69,15,40,253 // movaps %xmm13,%xmm15
+ .byte 68,15,85,248 // andnps %xmm0,%xmm15
+ .byte 65,15,92,193 // subps %xmm9,%xmm0
+ .byte 69,15,92,217 // subps %xmm9,%xmm11
+ .byte 65,15,89,195 // mulps %xmm11,%xmm0
+ .byte 69,15,92,225 // subps %xmm9,%xmm12
+ .byte 65,15,94,196 // divps %xmm12,%xmm0
+ .byte 65,15,88,193 // addps %xmm9,%xmm0
+ .byte 65,15,84,197 // andps %xmm13,%xmm0
+ .byte 65,15,86,199 // orps %xmm15,%xmm0
+ .byte 68,15,40,249 // movaps %xmm1,%xmm15
+ .byte 69,15,92,249 // subps %xmm9,%xmm15
+ .byte 69,15,89,249 // mulps %xmm9,%xmm15
+ .byte 69,15,94,254 // divps %xmm14,%xmm15
+ .byte 69,15,88,249 // addps %xmm9,%xmm15
+ .byte 65,15,40,208 // movaps %xmm8,%xmm2
+ .byte 65,15,85,215 // andnps %xmm15,%xmm2
+ .byte 65,15,84,200 // andps %xmm8,%xmm1
+ .byte 15,86,202 // orps %xmm2,%xmm1
+ .byte 65,15,40,213 // movaps %xmm13,%xmm2
+ .byte 15,85,209 // andnps %xmm1,%xmm2
+ .byte 65,15,92,201 // subps %xmm9,%xmm1
+ .byte 65,15,89,203 // mulps %xmm11,%xmm1
+ .byte 65,15,94,204 // divps %xmm12,%xmm1
+ .byte 65,15,88,201 // addps %xmm9,%xmm1
+ .byte 65,15,84,205 // andps %xmm13,%xmm1
+ .byte 15,86,202 // orps %xmm2,%xmm1
+ .byte 15,40,211 // movaps %xmm3,%xmm2
+ .byte 65,15,92,209 // subps %xmm9,%xmm2
+ .byte 65,15,89,209 // mulps %xmm9,%xmm2
+ .byte 65,15,94,214 // divps %xmm14,%xmm2
+ .byte 65,15,88,209 // addps %xmm9,%xmm2
+ .byte 65,15,84,216 // andps %xmm8,%xmm3
+ .byte 68,15,85,194 // andnps %xmm2,%xmm8
+ .byte 68,15,86,195 // orps %xmm3,%xmm8
+ .byte 65,15,40,213 // movaps %xmm13,%xmm2
+ .byte 65,15,85,208 // andnps %xmm8,%xmm2
+ .byte 69,15,92,193 // subps %xmm9,%xmm8
+ .byte 69,15,89,195 // mulps %xmm11,%xmm8
+ .byte 69,15,94,196 // divps %xmm12,%xmm8
+ .byte 69,15,88,193 // addps %xmm9,%xmm8
+ .byte 69,15,84,197 // andps %xmm13,%xmm8
+ .byte 68,15,86,194 // orps %xmm2,%xmm8
+ .byte 65,15,95,194 // maxps %xmm10,%xmm0
+ .byte 65,15,95,202 // maxps %xmm10,%xmm1
+ .byte 69,15,95,194 // maxps %xmm10,%xmm8
+ .byte 68,15,40,204 // movaps %xmm4,%xmm9
+ .byte 68,15,88,207 // addps %xmm7,%xmm9
+ .byte 15,89,231 // mulps %xmm7,%xmm4
+ .byte 68,15,92,204 // subps %xmm4,%xmm9
+ .byte 65,15,89,193 // mulps %xmm9,%xmm0
+ .byte 65,15,89,201 // mulps %xmm9,%xmm1
+ .byte 69,15,89,193 // mulps %xmm9,%xmm8
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 65,15,40,208 // movaps %xmm8,%xmm2
+ .byte 65,15,40,217 // movaps %xmm9,%xmm3
+ .byte 15,40,229 // movaps %xmm5,%xmm4
+ .byte 15,40,238 // movaps %xmm6,%xmm5
+ .byte 15,40,116,36,232 // movaps -0x18(%rsp),%xmm6
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_clamp_0_sse2
.globl _sk_clamp_0_sse2
FUNCTION(_sk_clamp_0_sse2)
@@ -24542,7 +27485,7 @@ HIDDEN _sk_clamp_1_sse2
.globl _sk_clamp_1_sse2
FUNCTION(_sk_clamp_1_sse2)
_sk_clamp_1_sse2:
- .byte 68,15,40,5,39,51,0,0 // movaps 0x3327(%rip),%xmm8 # 40e0 <_sk_callback_sse2+0x284>
+ .byte 68,15,40,5,33,52,0,0 // movaps 0x3421(%rip),%xmm8 # 4ab0 <_sk_callback_sse2+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
@@ -24554,7 +27497,7 @@ HIDDEN _sk_clamp_a_sse2
.globl _sk_clamp_a_sse2
FUNCTION(_sk_clamp_a_sse2)
_sk_clamp_a_sse2:
- .byte 15,93,29,28,51,0,0 // minps 0x331c(%rip),%xmm3 # 40f0 <_sk_callback_sse2+0x294>
+ .byte 15,93,29,22,52,0,0 // minps 0x3416(%rip),%xmm3 # 4ac0 <_sk_callback_sse2+0x38e>
.byte 15,93,195 // minps %xmm3,%xmm0
.byte 15,93,203 // minps %xmm3,%xmm1
.byte 15,93,211 // minps %xmm3,%xmm2
@@ -24641,7 +27584,7 @@ HIDDEN _sk_unpremul_sse2
FUNCTION(_sk_unpremul_sse2)
_sk_unpremul_sse2:
.byte 69,15,87,192 // xorps %xmm8,%xmm8
- .byte 68,15,40,13,135,50,0,0 // movaps 0x3287(%rip),%xmm9 # 4100 <_sk_callback_sse2+0x2a4>
+ .byte 68,15,40,13,129,51,0,0 // movaps 0x3381(%rip),%xmm9 # 4ad0 <_sk_callback_sse2+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
@@ -24655,20 +27598,20 @@ HIDDEN _sk_from_srgb_sse2
.globl _sk_from_srgb_sse2
FUNCTION(_sk_from_srgb_sse2)
_sk_from_srgb_sse2:
- .byte 68,15,40,5,114,50,0,0 // movaps 0x3272(%rip),%xmm8 # 4110 <_sk_callback_sse2+0x2b4>
+ .byte 68,15,40,5,108,51,0,0 // movaps 0x336c(%rip),%xmm8 # 4ae0 <_sk_callback_sse2+0x3ae>
.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 68,15,40,13,106,50,0,0 // movaps 0x326a(%rip),%xmm9 # 4120 <_sk_callback_sse2+0x2c4>
+ .byte 68,15,40,13,100,51,0,0 // movaps 0x3364(%rip),%xmm9 # 4af0 <_sk_callback_sse2+0x3be>
.byte 68,15,40,240 // movaps %xmm0,%xmm14
.byte 69,15,89,241 // mulps %xmm9,%xmm14
- .byte 68,15,40,21,106,50,0,0 // movaps 0x326a(%rip),%xmm10 # 4130 <_sk_callback_sse2+0x2d4>
+ .byte 68,15,40,21,100,51,0,0 // movaps 0x3364(%rip),%xmm10 # 4b00 <_sk_callback_sse2+0x3ce>
.byte 69,15,88,242 // addps %xmm10,%xmm14
.byte 69,15,89,243 // mulps %xmm11,%xmm14
- .byte 68,15,40,29,106,50,0,0 // movaps 0x326a(%rip),%xmm11 # 4140 <_sk_callback_sse2+0x2e4>
+ .byte 68,15,40,29,100,51,0,0 // movaps 0x3364(%rip),%xmm11 # 4b10 <_sk_callback_sse2+0x3de>
.byte 69,15,88,243 // addps %xmm11,%xmm14
- .byte 68,15,40,37,110,50,0,0 // movaps 0x326e(%rip),%xmm12 # 4150 <_sk_callback_sse2+0x2f4>
+ .byte 68,15,40,37,104,51,0,0 // movaps 0x3368(%rip),%xmm12 # 4b20 <_sk_callback_sse2+0x3ee>
.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
@@ -24707,20 +27650,20 @@ _sk_to_srgb_sse2:
.byte 68,15,82,192 // rsqrtps %xmm0,%xmm8
.byte 69,15,83,200 // rcpps %xmm8,%xmm9
.byte 69,15,82,232 // rsqrtps %xmm8,%xmm13
- .byte 68,15,40,5,243,49,0,0 // movaps 0x31f3(%rip),%xmm8 # 4160 <_sk_callback_sse2+0x304>
+ .byte 68,15,40,5,237,50,0,0 // movaps 0x32ed(%rip),%xmm8 # 4b30 <_sk_callback_sse2+0x3fe>
.byte 68,15,40,240 // movaps %xmm0,%xmm14
.byte 69,15,89,240 // mulps %xmm8,%xmm14
- .byte 68,15,40,21,243,49,0,0 // movaps 0x31f3(%rip),%xmm10 # 4170 <_sk_callback_sse2+0x314>
+ .byte 68,15,40,21,237,50,0,0 // movaps 0x32ed(%rip),%xmm10 # 4b40 <_sk_callback_sse2+0x40e>
.byte 69,15,89,202 // mulps %xmm10,%xmm9
- .byte 68,15,40,29,247,49,0,0 // movaps 0x31f7(%rip),%xmm11 # 4180 <_sk_callback_sse2+0x324>
+ .byte 68,15,40,29,241,50,0,0 // movaps 0x32f1(%rip),%xmm11 # 4b50 <_sk_callback_sse2+0x41e>
.byte 69,15,88,203 // addps %xmm11,%xmm9
- .byte 68,15,40,37,251,49,0,0 // movaps 0x31fb(%rip),%xmm12 # 4190 <_sk_callback_sse2+0x334>
+ .byte 68,15,40,37,245,50,0,0 // movaps 0x32f5(%rip),%xmm12 # 4b60 <_sk_callback_sse2+0x42e>
.byte 69,15,89,236 // mulps %xmm12,%xmm13
.byte 69,15,88,233 // addps %xmm9,%xmm13
- .byte 68,15,40,13,251,49,0,0 // movaps 0x31fb(%rip),%xmm9 # 41a0 <_sk_callback_sse2+0x344>
+ .byte 68,15,40,13,245,50,0,0 // movaps 0x32f5(%rip),%xmm9 # 4b70 <_sk_callback_sse2+0x43e>
.byte 69,15,40,249 // movaps %xmm9,%xmm15
.byte 69,15,93,253 // minps %xmm13,%xmm15
- .byte 68,15,40,45,251,49,0,0 // movaps 0x31fb(%rip),%xmm13 # 41b0 <_sk_callback_sse2+0x354>
+ .byte 68,15,40,45,245,50,0,0 // movaps 0x32f5(%rip),%xmm13 # 4b80 <_sk_callback_sse2+0x44e>
.byte 65,15,194,197,1 // cmpltps %xmm13,%xmm0
.byte 68,15,84,240 // andps %xmm0,%xmm14
.byte 65,15,85,199 // andnps %xmm15,%xmm0
@@ -24770,7 +27713,7 @@ _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 68,15,40,45,84,49,0,0 // movaps 0x3154(%rip),%xmm13 # 41c0 <_sk_callback_sse2+0x364>
+ .byte 68,15,40,45,78,50,0,0 // movaps 0x324e(%rip),%xmm13 # 4b90 <_sk_callback_sse2+0x45e>
.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
@@ -24779,30 +27722,30 @@ _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 68,15,84,53,58,49,0,0 // andps 0x313a(%rip),%xmm14 # 41d0 <_sk_callback_sse2+0x374>
+ .byte 68,15,84,53,52,50,0,0 // andps 0x3234(%rip),%xmm14 # 4ba0 <_sk_callback_sse2+0x46e>
.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 68,15,40,37,45,49,0,0 // movaps 0x312d(%rip),%xmm12 # 41e0 <_sk_callback_sse2+0x384>
+ .byte 68,15,40,37,39,50,0,0 // movaps 0x3227(%rip),%xmm12 # 4bb0 <_sk_callback_sse2+0x47e>
.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 68,15,88,5,41,49,0,0 // addps 0x3129(%rip),%xmm8 # 41f0 <_sk_callback_sse2+0x394>
+ .byte 68,15,88,5,35,50,0,0 // addps 0x3223(%rip),%xmm8 # 4bc0 <_sk_callback_sse2+0x48e>
.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 15,89,5,26,49,0,0 // mulps 0x311a(%rip),%xmm0 # 4200 <_sk_callback_sse2+0x3a4>
+ .byte 15,89,5,20,50,0,0 // mulps 0x3214(%rip),%xmm0 # 4bd0 <_sk_callback_sse2+0x49e>
.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 68,15,40,13,13,49,0,0 // movaps 0x310d(%rip),%xmm9 # 4210 <_sk_callback_sse2+0x3b4>
+ .byte 68,15,40,13,7,50,0,0 // movaps 0x3207(%rip),%xmm9 # 4be0 <_sk_callback_sse2+0x4ae>
.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
@@ -24826,7 +27769,7 @@ _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 68,15,40,13,204,48,0,0 // movaps 0x30cc(%rip),%xmm9 # 4220 <_sk_callback_sse2+0x3c4>
+ .byte 68,15,40,13,198,49,0,0 // movaps 0x31c6(%rip),%xmm9 # 4bf0 <_sk_callback_sse2+0x4be>
.byte 69,15,40,209 // movaps %xmm9,%xmm10
.byte 69,15,194,211,2 // cmpleps %xmm11,%xmm10
.byte 15,40,193 // movaps %xmm1,%xmm0
@@ -24843,28 +27786,28 @@ _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 15,40,5,149,48,0,0 // movaps 0x3095(%rip),%xmm0 # 4230 <_sk_callback_sse2+0x3d4>
+ .byte 15,40,5,143,49,0,0 // movaps 0x318f(%rip),%xmm0 # 4c00 <_sk_callback_sse2+0x4ce>
.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 15,84,29,141,48,0,0 // andps 0x308d(%rip),%xmm3 # 4240 <_sk_callback_sse2+0x3e4>
+ .byte 15,84,29,135,49,0,0 // andps 0x3187(%rip),%xmm3 # 4c10 <_sk_callback_sse2+0x4de>
.byte 15,92,203 // subps %xmm3,%xmm1
.byte 15,92,193 // subps %xmm1,%xmm0
- .byte 68,15,40,45,143,48,0,0 // movaps 0x308f(%rip),%xmm13 # 4250 <_sk_callback_sse2+0x3f4>
+ .byte 68,15,40,45,137,49,0,0 // movaps 0x3189(%rip),%xmm13 # 4c20 <_sk_callback_sse2+0x4ee>
.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 15,40,21,159,48,0,0 // movaps 0x309f(%rip),%xmm2 # 4280 <_sk_callback_sse2+0x424>
+ .byte 15,40,21,153,49,0,0 // movaps 0x3199(%rip),%xmm2 # 4c50 <_sk_callback_sse2+0x51e>
.byte 68,15,40,250 // movaps %xmm2,%xmm15
.byte 68,15,194,248,2 // cmpleps %xmm0,%xmm15
- .byte 15,40,13,111,48,0,0 // movaps 0x306f(%rip),%xmm1 # 4260 <_sk_callback_sse2+0x404>
+ .byte 15,40,13,105,49,0,0 // movaps 0x3169(%rip),%xmm1 # 4c30 <_sk_callback_sse2+0x4fe>
.byte 15,89,193 // mulps %xmm1,%xmm0
- .byte 15,40,45,117,48,0,0 // movaps 0x3075(%rip),%xmm5 # 4270 <_sk_callback_sse2+0x414>
+ .byte 15,40,45,111,49,0,0 // movaps 0x316f(%rip),%xmm5 # 4c40 <_sk_callback_sse2+0x50e>
.byte 15,40,229 // movaps %xmm5,%xmm4
.byte 15,92,224 // subps %xmm0,%xmm4
.byte 65,15,89,230 // mulps %xmm14,%xmm4
@@ -24887,7 +27830,7 @@ _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 15,84,29,234,47,0,0 // andps 0x2fea(%rip),%xmm3 # 4240 <_sk_callback_sse2+0x3e4>
+ .byte 15,84,29,228,48,0,0 // andps 0x30e4(%rip),%xmm3 # 4c10 <_sk_callback_sse2+0x4de>
.byte 15,92,195 // subps %xmm3,%xmm0
.byte 68,15,40,230 // movaps %xmm6,%xmm12
.byte 68,15,92,224 // subps %xmm0,%xmm12
@@ -24917,12 +27860,12 @@ _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 15,88,53,194,47,0,0 // addps 0x2fc2(%rip),%xmm6 # 4290 <_sk_callback_sse2+0x434>
+ .byte 15,88,53,188,48,0,0 // addps 0x30bc(%rip),%xmm6 # 4c60 <_sk_callback_sse2+0x52e>
.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 15,84,29,93,47,0,0 // andps 0x2f5d(%rip),%xmm3 # 4240 <_sk_callback_sse2+0x3e4>
+ .byte 15,84,29,87,48,0,0 // andps 0x3057(%rip),%xmm3 # 4c10 <_sk_callback_sse2+0x4de>
.byte 15,92,195 // subps %xmm3,%xmm0
.byte 15,92,240 // subps %xmm0,%xmm6
.byte 15,89,206 // mulps %xmm6,%xmm1
@@ -24986,7 +27929,7 @@ _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 68,15,89,5,235,46,0,0 // mulps 0x2eeb(%rip),%xmm8 # 42a0 <_sk_callback_sse2+0x444>
+ .byte 68,15,89,5,229,47,0,0 // mulps 0x2fe5(%rip),%xmm8 # 4c70 <_sk_callback_sse2+0x53e>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 65,15,89,208 // mulps %xmm8,%xmm2
@@ -25027,7 +27970,7 @@ _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 68,15,89,5,137,46,0,0 // mulps 0x2e89(%rip),%xmm8 # 42b0 <_sk_callback_sse2+0x454>
+ .byte 68,15,89,5,131,47,0,0 // mulps 0x2f83(%rip),%xmm8 # 4c80 <_sk_callback_sse2+0x54e>
.byte 15,92,196 // subps %xmm4,%xmm0
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -25052,17 +27995,17 @@ _sk_lerp_565_sse2:
.byte 243,68,15,126,4,120 // movq (%rax,%rdi,2),%xmm8
.byte 102,15,239,219 // pxor %xmm3,%xmm3
.byte 102,68,15,97,195 // punpcklwd %xmm3,%xmm8
- .byte 102,15,111,29,81,46,0,0 // movdqa 0x2e51(%rip),%xmm3 # 42c0 <_sk_callback_sse2+0x464>
+ .byte 102,15,111,29,75,47,0,0 // movdqa 0x2f4b(%rip),%xmm3 # 4c90 <_sk_callback_sse2+0x55e>
.byte 102,65,15,219,216 // pand %xmm8,%xmm3
.byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
- .byte 68,15,89,13,80,46,0,0 // mulps 0x2e50(%rip),%xmm9 # 42d0 <_sk_callback_sse2+0x474>
- .byte 102,15,111,29,88,46,0,0 // movdqa 0x2e58(%rip),%xmm3 # 42e0 <_sk_callback_sse2+0x484>
+ .byte 68,15,89,13,74,47,0,0 // mulps 0x2f4a(%rip),%xmm9 # 4ca0 <_sk_callback_sse2+0x56e>
+ .byte 102,15,111,29,82,47,0,0 // movdqa 0x2f52(%rip),%xmm3 # 4cb0 <_sk_callback_sse2+0x57e>
.byte 102,65,15,219,216 // pand %xmm8,%xmm3
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,89,46,0,0 // mulps 0x2e59(%rip),%xmm3 # 42f0 <_sk_callback_sse2+0x494>
- .byte 102,68,15,219,5,96,46,0,0 // pand 0x2e60(%rip),%xmm8 # 4300 <_sk_callback_sse2+0x4a4>
+ .byte 15,89,29,83,47,0,0 // mulps 0x2f53(%rip),%xmm3 # 4cc0 <_sk_callback_sse2+0x58e>
+ .byte 102,68,15,219,5,90,47,0,0 // pand 0x2f5a(%rip),%xmm8 # 4cd0 <_sk_callback_sse2+0x59e>
.byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
- .byte 68,15,89,5,100,46,0,0 // mulps 0x2e64(%rip),%xmm8 # 4310 <_sk_callback_sse2+0x4b4>
+ .byte 68,15,89,5,94,47,0,0 // mulps 0x2f5e(%rip),%xmm8 # 4ce0 <_sk_callback_sse2+0x5ae>
.byte 15,92,196 // subps %xmm4,%xmm0
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -25073,7 +28016,7 @@ _sk_lerp_565_sse2:
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 15,88,214 // addps %xmm6,%xmm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,78,46,0,0 // movaps 0x2e4e(%rip),%xmm3 # 4320 <_sk_callback_sse2+0x4c4>
+ .byte 15,40,29,72,47,0,0 // movaps 0x2f48(%rip),%xmm3 # 4cf0 <_sk_callback_sse2+0x5be>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_load_tables_sse2
@@ -25084,7 +28027,7 @@ _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 102,68,15,111,5,68,46,0,0 // movdqa 0x2e44(%rip),%xmm8 # 4330 <_sk_callback_sse2+0x4d4>
+ .byte 102,68,15,111,5,62,47,0,0 // movdqa 0x2f3e(%rip),%xmm8 # 4d00 <_sk_callback_sse2+0x5ce>
.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
@@ -25139,7 +28082,7 @@ _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 15,89,29,81,45,0,0 // mulps 0x2d51(%rip),%xmm3 # 4340 <_sk_callback_sse2+0x4e4>
+ .byte 15,89,29,75,46,0,0 // mulps 0x2e4b(%rip),%xmm3 # 4d10 <_sk_callback_sse2+0x5de>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -25158,7 +28101,7 @@ _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 102,68,15,111,21,36,45,0,0 // movdqa 0x2d24(%rip),%xmm10 # 4350 <_sk_callback_sse2+0x4f4>
+ .byte 102,68,15,111,21,30,46,0,0 // movdqa 0x2e1e(%rip),%xmm10 # 4d20 <_sk_callback_sse2+0x5ee>
.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
@@ -25219,7 +28162,7 @@ _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 15,89,29,19,44,0,0 // mulps 0x2c13(%rip),%xmm3 # 4360 <_sk_callback_sse2+0x504>
+ .byte 15,89,29,13,45,0,0 // mulps 0x2d0d(%rip),%xmm3 # 4d30 <_sk_callback_sse2+0x5fe>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -25241,7 +28184,7 @@ _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 102,68,15,111,5,211,43,0,0 // movdqa 0x2bd3(%rip),%xmm8 # 4370 <_sk_callback_sse2+0x514>
+ .byte 102,68,15,111,5,205,44,0,0 // movdqa 0x2ccd(%rip),%xmm8 # 4d40 <_sk_callback_sse2+0x60e>
.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
@@ -25296,7 +28239,7 @@ _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 15,40,29,226,42,0,0 // movaps 0x2ae2(%rip),%xmm3 # 4380 <_sk_callback_sse2+0x524>
+ .byte 15,40,29,220,43,0,0 // movaps 0x2bdc(%rip),%xmm3 # 4d50 <_sk_callback_sse2+0x61e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_byte_tables_sse2
@@ -25306,7 +28249,7 @@ _sk_byte_tables_sse2:
.byte 65,86 // push %r14
.byte 83 // push %rbx
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,227,42,0,0 // movaps 0x2ae3(%rip),%xmm8 # 4390 <_sk_callback_sse2+0x534>
+ .byte 68,15,40,5,221,43,0,0 // movaps 0x2bdd(%rip),%xmm8 # 4d60 <_sk_callback_sse2+0x62e>
.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
@@ -25333,7 +28276,7 @@ _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 68,15,40,21,128,42,0,0 // movaps 0x2a80(%rip),%xmm10 # 43a0 <_sk_callback_sse2+0x544>
+ .byte 68,15,40,21,122,43,0,0 // movaps 0x2b7a(%rip),%xmm10 # 4d70 <_sk_callback_sse2+0x63e>
.byte 65,15,89,194 // mulps %xmm10,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
@@ -25449,7 +28392,7 @@ _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 68,15,40,21,211,40,0,0 // movaps 0x28d3(%rip),%xmm10 # 43b0 <_sk_callback_sse2+0x554>
+ .byte 68,15,40,21,205,41,0,0 // movaps 0x29cd(%rip),%xmm10 # 4d80 <_sk_callback_sse2+0x64e>
.byte 65,15,89,194 // mulps %xmm10,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
@@ -25646,15 +28589,15 @@ _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 68,15,89,13,18,38,0,0 // mulps 0x2612(%rip),%xmm9 # 43c0 <_sk_callback_sse2+0x564>
- .byte 68,15,84,21,26,38,0,0 // andps 0x261a(%rip),%xmm10 # 43d0 <_sk_callback_sse2+0x574>
- .byte 68,15,86,21,34,38,0,0 // orps 0x2622(%rip),%xmm10 # 43e0 <_sk_callback_sse2+0x584>
- .byte 68,15,88,13,42,38,0,0 // addps 0x262a(%rip),%xmm9 # 43f0 <_sk_callback_sse2+0x594>
- .byte 68,15,40,37,50,38,0,0 // movaps 0x2632(%rip),%xmm12 # 4400 <_sk_callback_sse2+0x5a4>
+ .byte 68,15,89,13,12,39,0,0 // mulps 0x270c(%rip),%xmm9 # 4d90 <_sk_callback_sse2+0x65e>
+ .byte 68,15,84,21,20,39,0,0 // andps 0x2714(%rip),%xmm10 # 4da0 <_sk_callback_sse2+0x66e>
+ .byte 68,15,86,21,28,39,0,0 // orps 0x271c(%rip),%xmm10 # 4db0 <_sk_callback_sse2+0x67e>
+ .byte 68,15,88,13,36,39,0,0 // addps 0x2724(%rip),%xmm9 # 4dc0 <_sk_callback_sse2+0x68e>
+ .byte 68,15,40,37,44,39,0,0 // movaps 0x272c(%rip),%xmm12 # 4dd0 <_sk_callback_sse2+0x69e>
.byte 69,15,89,226 // mulps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,88,21,50,38,0,0 // addps 0x2632(%rip),%xmm10 # 4410 <_sk_callback_sse2+0x5b4>
- .byte 68,15,40,37,58,38,0,0 // movaps 0x263a(%rip),%xmm12 # 4420 <_sk_callback_sse2+0x5c4>
+ .byte 68,15,88,21,44,39,0,0 // addps 0x272c(%rip),%xmm10 # 4de0 <_sk_callback_sse2+0x6ae>
+ .byte 68,15,40,37,52,39,0,0 // movaps 0x2734(%rip),%xmm12 # 4df0 <_sk_callback_sse2+0x6be>
.byte 69,15,94,226 // divps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
.byte 69,15,89,203 // mulps %xmm11,%xmm9
@@ -25662,22 +28605,22 @@ _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 68,15,40,21,36,38,0,0 // movaps 0x2624(%rip),%xmm10 # 4430 <_sk_callback_sse2+0x5d4>
+ .byte 68,15,40,21,30,39,0,0 // movaps 0x271e(%rip),%xmm10 # 4e00 <_sk_callback_sse2+0x6ce>
.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 68,15,88,13,24,38,0,0 // addps 0x2618(%rip),%xmm9 # 4440 <_sk_callback_sse2+0x5e4>
- .byte 68,15,40,37,32,38,0,0 // movaps 0x2620(%rip),%xmm12 # 4450 <_sk_callback_sse2+0x5f4>
+ .byte 68,15,88,13,18,39,0,0 // addps 0x2712(%rip),%xmm9 # 4e10 <_sk_callback_sse2+0x6de>
+ .byte 68,15,40,37,26,39,0,0 // movaps 0x271a(%rip),%xmm12 # 4e20 <_sk_callback_sse2+0x6ee>
.byte 69,15,89,229 // mulps %xmm13,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,40,37,32,38,0,0 // movaps 0x2620(%rip),%xmm12 # 4460 <_sk_callback_sse2+0x604>
+ .byte 68,15,40,37,26,39,0,0 // movaps 0x271a(%rip),%xmm12 # 4e30 <_sk_callback_sse2+0x6fe>
.byte 69,15,92,229 // subps %xmm13,%xmm12
- .byte 68,15,40,45,36,38,0,0 // movaps 0x2624(%rip),%xmm13 # 4470 <_sk_callback_sse2+0x614>
+ .byte 68,15,40,45,30,39,0,0 // movaps 0x271e(%rip),%xmm13 # 4e40 <_sk_callback_sse2+0x70e>
.byte 69,15,94,236 // divps %xmm12,%xmm13
.byte 69,15,88,233 // addps %xmm9,%xmm13
- .byte 68,15,89,45,36,38,0,0 // mulps 0x2624(%rip),%xmm13 # 4480 <_sk_callback_sse2+0x624>
+ .byte 68,15,89,45,30,39,0,0 // mulps 0x271e(%rip),%xmm13 # 4e50 <_sk_callback_sse2+0x71e>
.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
@@ -25713,15 +28656,15 @@ _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 68,15,89,13,164,37,0,0 // mulps 0x25a4(%rip),%xmm9 # 4490 <_sk_callback_sse2+0x634>
- .byte 68,15,84,21,172,37,0,0 // andps 0x25ac(%rip),%xmm10 # 44a0 <_sk_callback_sse2+0x644>
- .byte 68,15,86,21,180,37,0,0 // orps 0x25b4(%rip),%xmm10 # 44b0 <_sk_callback_sse2+0x654>
- .byte 68,15,88,13,188,37,0,0 // addps 0x25bc(%rip),%xmm9 # 44c0 <_sk_callback_sse2+0x664>
- .byte 68,15,40,37,196,37,0,0 // movaps 0x25c4(%rip),%xmm12 # 44d0 <_sk_callback_sse2+0x674>
+ .byte 68,15,89,13,158,38,0,0 // mulps 0x269e(%rip),%xmm9 # 4e60 <_sk_callback_sse2+0x72e>
+ .byte 68,15,84,21,166,38,0,0 // andps 0x26a6(%rip),%xmm10 # 4e70 <_sk_callback_sse2+0x73e>
+ .byte 68,15,86,21,174,38,0,0 // orps 0x26ae(%rip),%xmm10 # 4e80 <_sk_callback_sse2+0x74e>
+ .byte 68,15,88,13,182,38,0,0 // addps 0x26b6(%rip),%xmm9 # 4e90 <_sk_callback_sse2+0x75e>
+ .byte 68,15,40,37,190,38,0,0 // movaps 0x26be(%rip),%xmm12 # 4ea0 <_sk_callback_sse2+0x76e>
.byte 69,15,89,226 // mulps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,88,21,196,37,0,0 // addps 0x25c4(%rip),%xmm10 # 44e0 <_sk_callback_sse2+0x684>
- .byte 68,15,40,37,204,37,0,0 // movaps 0x25cc(%rip),%xmm12 # 44f0 <_sk_callback_sse2+0x694>
+ .byte 68,15,88,21,190,38,0,0 // addps 0x26be(%rip),%xmm10 # 4eb0 <_sk_callback_sse2+0x77e>
+ .byte 68,15,40,37,198,38,0,0 // movaps 0x26c6(%rip),%xmm12 # 4ec0 <_sk_callback_sse2+0x78e>
.byte 69,15,94,226 // divps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
.byte 69,15,89,203 // mulps %xmm11,%xmm9
@@ -25729,22 +28672,22 @@ _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 68,15,40,21,182,37,0,0 // movaps 0x25b6(%rip),%xmm10 # 4500 <_sk_callback_sse2+0x6a4>
+ .byte 68,15,40,21,176,38,0,0 // movaps 0x26b0(%rip),%xmm10 # 4ed0 <_sk_callback_sse2+0x79e>
.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 68,15,88,13,170,37,0,0 // addps 0x25aa(%rip),%xmm9 # 4510 <_sk_callback_sse2+0x6b4>
- .byte 68,15,40,37,178,37,0,0 // movaps 0x25b2(%rip),%xmm12 # 4520 <_sk_callback_sse2+0x6c4>
+ .byte 68,15,88,13,164,38,0,0 // addps 0x26a4(%rip),%xmm9 # 4ee0 <_sk_callback_sse2+0x7ae>
+ .byte 68,15,40,37,172,38,0,0 // movaps 0x26ac(%rip),%xmm12 # 4ef0 <_sk_callback_sse2+0x7be>
.byte 69,15,89,229 // mulps %xmm13,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,40,37,178,37,0,0 // movaps 0x25b2(%rip),%xmm12 # 4530 <_sk_callback_sse2+0x6d4>
+ .byte 68,15,40,37,172,38,0,0 // movaps 0x26ac(%rip),%xmm12 # 4f00 <_sk_callback_sse2+0x7ce>
.byte 69,15,92,229 // subps %xmm13,%xmm12
- .byte 68,15,40,45,182,37,0,0 // movaps 0x25b6(%rip),%xmm13 # 4540 <_sk_callback_sse2+0x6e4>
+ .byte 68,15,40,45,176,38,0,0 // movaps 0x26b0(%rip),%xmm13 # 4f10 <_sk_callback_sse2+0x7de>
.byte 69,15,94,236 // divps %xmm12,%xmm13
.byte 69,15,88,233 // addps %xmm9,%xmm13
- .byte 68,15,89,45,182,37,0,0 // mulps 0x25b6(%rip),%xmm13 # 4550 <_sk_callback_sse2+0x6f4>
+ .byte 68,15,89,45,176,38,0,0 // mulps 0x26b0(%rip),%xmm13 # 4f20 <_sk_callback_sse2+0x7ee>
.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
@@ -25780,15 +28723,15 @@ _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 68,15,89,13,54,37,0,0 // mulps 0x2536(%rip),%xmm9 # 4560 <_sk_callback_sse2+0x704>
- .byte 68,15,84,21,62,37,0,0 // andps 0x253e(%rip),%xmm10 # 4570 <_sk_callback_sse2+0x714>
- .byte 68,15,86,21,70,37,0,0 // orps 0x2546(%rip),%xmm10 # 4580 <_sk_callback_sse2+0x724>
- .byte 68,15,88,13,78,37,0,0 // addps 0x254e(%rip),%xmm9 # 4590 <_sk_callback_sse2+0x734>
- .byte 68,15,40,37,86,37,0,0 // movaps 0x2556(%rip),%xmm12 # 45a0 <_sk_callback_sse2+0x744>
+ .byte 68,15,89,13,48,38,0,0 // mulps 0x2630(%rip),%xmm9 # 4f30 <_sk_callback_sse2+0x7fe>
+ .byte 68,15,84,21,56,38,0,0 // andps 0x2638(%rip),%xmm10 # 4f40 <_sk_callback_sse2+0x80e>
+ .byte 68,15,86,21,64,38,0,0 // orps 0x2640(%rip),%xmm10 # 4f50 <_sk_callback_sse2+0x81e>
+ .byte 68,15,88,13,72,38,0,0 // addps 0x2648(%rip),%xmm9 # 4f60 <_sk_callback_sse2+0x82e>
+ .byte 68,15,40,37,80,38,0,0 // movaps 0x2650(%rip),%xmm12 # 4f70 <_sk_callback_sse2+0x83e>
.byte 69,15,89,226 // mulps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,88,21,86,37,0,0 // addps 0x2556(%rip),%xmm10 # 45b0 <_sk_callback_sse2+0x754>
- .byte 68,15,40,37,94,37,0,0 // movaps 0x255e(%rip),%xmm12 # 45c0 <_sk_callback_sse2+0x764>
+ .byte 68,15,88,21,80,38,0,0 // addps 0x2650(%rip),%xmm10 # 4f80 <_sk_callback_sse2+0x84e>
+ .byte 68,15,40,37,88,38,0,0 // movaps 0x2658(%rip),%xmm12 # 4f90 <_sk_callback_sse2+0x85e>
.byte 69,15,94,226 // divps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
.byte 69,15,89,203 // mulps %xmm11,%xmm9
@@ -25796,22 +28739,22 @@ _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 68,15,40,21,72,37,0,0 // movaps 0x2548(%rip),%xmm10 # 45d0 <_sk_callback_sse2+0x774>
+ .byte 68,15,40,21,66,38,0,0 // movaps 0x2642(%rip),%xmm10 # 4fa0 <_sk_callback_sse2+0x86e>
.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 68,15,88,13,60,37,0,0 // addps 0x253c(%rip),%xmm9 # 45e0 <_sk_callback_sse2+0x784>
- .byte 68,15,40,37,68,37,0,0 // movaps 0x2544(%rip),%xmm12 # 45f0 <_sk_callback_sse2+0x794>
+ .byte 68,15,88,13,54,38,0,0 // addps 0x2636(%rip),%xmm9 # 4fb0 <_sk_callback_sse2+0x87e>
+ .byte 68,15,40,37,62,38,0,0 // movaps 0x263e(%rip),%xmm12 # 4fc0 <_sk_callback_sse2+0x88e>
.byte 69,15,89,229 // mulps %xmm13,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,40,37,68,37,0,0 // movaps 0x2544(%rip),%xmm12 # 4600 <_sk_callback_sse2+0x7a4>
+ .byte 68,15,40,37,62,38,0,0 // movaps 0x263e(%rip),%xmm12 # 4fd0 <_sk_callback_sse2+0x89e>
.byte 69,15,92,229 // subps %xmm13,%xmm12
- .byte 68,15,40,45,72,37,0,0 // movaps 0x2548(%rip),%xmm13 # 4610 <_sk_callback_sse2+0x7b4>
+ .byte 68,15,40,45,66,38,0,0 // movaps 0x2642(%rip),%xmm13 # 4fe0 <_sk_callback_sse2+0x8ae>
.byte 69,15,94,236 // divps %xmm12,%xmm13
.byte 69,15,88,233 // addps %xmm9,%xmm13
- .byte 68,15,89,45,72,37,0,0 // mulps 0x2548(%rip),%xmm13 # 4620 <_sk_callback_sse2+0x7c4>
+ .byte 68,15,89,45,66,38,0,0 // mulps 0x2642(%rip),%xmm13 # 4ff0 <_sk_callback_sse2+0x8be>
.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
@@ -25847,15 +28790,15 @@ _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 68,15,89,13,200,36,0,0 // mulps 0x24c8(%rip),%xmm9 # 4630 <_sk_callback_sse2+0x7d4>
- .byte 68,15,84,21,208,36,0,0 // andps 0x24d0(%rip),%xmm10 # 4640 <_sk_callback_sse2+0x7e4>
- .byte 68,15,86,21,216,36,0,0 // orps 0x24d8(%rip),%xmm10 # 4650 <_sk_callback_sse2+0x7f4>
- .byte 68,15,88,13,224,36,0,0 // addps 0x24e0(%rip),%xmm9 # 4660 <_sk_callback_sse2+0x804>
- .byte 68,15,40,37,232,36,0,0 // movaps 0x24e8(%rip),%xmm12 # 4670 <_sk_callback_sse2+0x814>
+ .byte 68,15,89,13,194,37,0,0 // mulps 0x25c2(%rip),%xmm9 # 5000 <_sk_callback_sse2+0x8ce>
+ .byte 68,15,84,21,202,37,0,0 // andps 0x25ca(%rip),%xmm10 # 5010 <_sk_callback_sse2+0x8de>
+ .byte 68,15,86,21,210,37,0,0 // orps 0x25d2(%rip),%xmm10 # 5020 <_sk_callback_sse2+0x8ee>
+ .byte 68,15,88,13,218,37,0,0 // addps 0x25da(%rip),%xmm9 # 5030 <_sk_callback_sse2+0x8fe>
+ .byte 68,15,40,37,226,37,0,0 // movaps 0x25e2(%rip),%xmm12 # 5040 <_sk_callback_sse2+0x90e>
.byte 69,15,89,226 // mulps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,88,21,232,36,0,0 // addps 0x24e8(%rip),%xmm10 # 4680 <_sk_callback_sse2+0x824>
- .byte 68,15,40,37,240,36,0,0 // movaps 0x24f0(%rip),%xmm12 # 4690 <_sk_callback_sse2+0x834>
+ .byte 68,15,88,21,226,37,0,0 // addps 0x25e2(%rip),%xmm10 # 5050 <_sk_callback_sse2+0x91e>
+ .byte 68,15,40,37,234,37,0,0 // movaps 0x25ea(%rip),%xmm12 # 5060 <_sk_callback_sse2+0x92e>
.byte 69,15,94,226 // divps %xmm10,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
.byte 69,15,89,203 // mulps %xmm11,%xmm9
@@ -25863,22 +28806,22 @@ _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 68,15,40,21,218,36,0,0 // movaps 0x24da(%rip),%xmm10 # 46a0 <_sk_callback_sse2+0x844>
+ .byte 68,15,40,21,212,37,0,0 // movaps 0x25d4(%rip),%xmm10 # 5070 <_sk_callback_sse2+0x93e>
.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 68,15,88,13,206,36,0,0 // addps 0x24ce(%rip),%xmm9 # 46b0 <_sk_callback_sse2+0x854>
- .byte 68,15,40,37,214,36,0,0 // movaps 0x24d6(%rip),%xmm12 # 46c0 <_sk_callback_sse2+0x864>
+ .byte 68,15,88,13,200,37,0,0 // addps 0x25c8(%rip),%xmm9 # 5080 <_sk_callback_sse2+0x94e>
+ .byte 68,15,40,37,208,37,0,0 // movaps 0x25d0(%rip),%xmm12 # 5090 <_sk_callback_sse2+0x95e>
.byte 69,15,89,229 // mulps %xmm13,%xmm12
.byte 69,15,92,204 // subps %xmm12,%xmm9
- .byte 68,15,40,37,214,36,0,0 // movaps 0x24d6(%rip),%xmm12 # 46d0 <_sk_callback_sse2+0x874>
+ .byte 68,15,40,37,208,37,0,0 // movaps 0x25d0(%rip),%xmm12 # 50a0 <_sk_callback_sse2+0x96e>
.byte 69,15,92,229 // subps %xmm13,%xmm12
- .byte 68,15,40,45,218,36,0,0 // movaps 0x24da(%rip),%xmm13 # 46e0 <_sk_callback_sse2+0x884>
+ .byte 68,15,40,45,212,37,0,0 // movaps 0x25d4(%rip),%xmm13 # 50b0 <_sk_callback_sse2+0x97e>
.byte 69,15,94,236 // divps %xmm12,%xmm13
.byte 69,15,88,233 // addps %xmm9,%xmm13
- .byte 68,15,89,45,218,36,0,0 // mulps 0x24da(%rip),%xmm13 # 46f0 <_sk_callback_sse2+0x894>
+ .byte 68,15,89,45,212,37,0,0 // mulps 0x25d4(%rip),%xmm13 # 50c0 <_sk_callback_sse2+0x98e>
.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
@@ -25895,29 +28838,29 @@ HIDDEN _sk_lab_to_xyz_sse2
.globl _sk_lab_to_xyz_sse2
FUNCTION(_sk_lab_to_xyz_sse2)
_sk_lab_to_xyz_sse2:
- .byte 15,89,5,183,36,0,0 // mulps 0x24b7(%rip),%xmm0 # 4700 <_sk_callback_sse2+0x8a4>
- .byte 68,15,40,5,191,36,0,0 // movaps 0x24bf(%rip),%xmm8 # 4710 <_sk_callback_sse2+0x8b4>
+ .byte 15,89,5,177,37,0,0 // mulps 0x25b1(%rip),%xmm0 # 50d0 <_sk_callback_sse2+0x99e>
+ .byte 68,15,40,5,185,37,0,0 // movaps 0x25b9(%rip),%xmm8 # 50e0 <_sk_callback_sse2+0x9ae>
.byte 65,15,89,200 // mulps %xmm8,%xmm1
- .byte 68,15,40,13,195,36,0,0 // movaps 0x24c3(%rip),%xmm9 # 4720 <_sk_callback_sse2+0x8c4>
+ .byte 68,15,40,13,189,37,0,0 // movaps 0x25bd(%rip),%xmm9 # 50f0 <_sk_callback_sse2+0x9be>
.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 15,88,5,192,36,0,0 // addps 0x24c0(%rip),%xmm0 # 4730 <_sk_callback_sse2+0x8d4>
- .byte 15,89,5,201,36,0,0 // mulps 0x24c9(%rip),%xmm0 # 4740 <_sk_callback_sse2+0x8e4>
- .byte 15,89,13,210,36,0,0 // mulps 0x24d2(%rip),%xmm1 # 4750 <_sk_callback_sse2+0x8f4>
+ .byte 15,88,5,186,37,0,0 // addps 0x25ba(%rip),%xmm0 # 5100 <_sk_callback_sse2+0x9ce>
+ .byte 15,89,5,195,37,0,0 // mulps 0x25c3(%rip),%xmm0 # 5110 <_sk_callback_sse2+0x9de>
+ .byte 15,89,13,204,37,0,0 // mulps 0x25cc(%rip),%xmm1 # 5120 <_sk_callback_sse2+0x9ee>
.byte 15,88,200 // addps %xmm0,%xmm1
- .byte 15,89,21,216,36,0,0 // mulps 0x24d8(%rip),%xmm2 # 4760 <_sk_callback_sse2+0x904>
+ .byte 15,89,21,210,37,0,0 // mulps 0x25d2(%rip),%xmm2 # 5130 <_sk_callback_sse2+0x9fe>
.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 15,40,21,205,36,0,0 // movaps 0x24cd(%rip),%xmm2 # 4770 <_sk_callback_sse2+0x914>
+ .byte 15,40,21,199,37,0,0 // movaps 0x25c7(%rip),%xmm2 # 5140 <_sk_callback_sse2+0xa0e>
.byte 68,15,40,194 // movaps %xmm2,%xmm8
.byte 69,15,194,196,1 // cmpltps %xmm12,%xmm8
- .byte 68,15,40,21,204,36,0,0 // movaps 0x24cc(%rip),%xmm10 # 4780 <_sk_callback_sse2+0x924>
+ .byte 68,15,40,21,198,37,0,0 // movaps 0x25c6(%rip),%xmm10 # 5150 <_sk_callback_sse2+0xa1e>
.byte 65,15,88,202 // addps %xmm10,%xmm1
- .byte 68,15,40,29,208,36,0,0 // movaps 0x24d0(%rip),%xmm11 # 4790 <_sk_callback_sse2+0x934>
+ .byte 68,15,40,29,202,37,0,0 // movaps 0x25ca(%rip),%xmm11 # 5160 <_sk_callback_sse2+0xa2e>
.byte 65,15,89,203 // mulps %xmm11,%xmm1
.byte 69,15,84,224 // andps %xmm8,%xmm12
.byte 68,15,85,193 // andnps %xmm1,%xmm8
@@ -25941,8 +28884,8 @@ _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 68,15,89,5,128,36,0,0 // mulps 0x2480(%rip),%xmm8 # 47a0 <_sk_callback_sse2+0x944>
- .byte 15,89,21,137,36,0,0 // mulps 0x2489(%rip),%xmm2 # 47b0 <_sk_callback_sse2+0x954>
+ .byte 68,15,89,5,122,37,0,0 // mulps 0x257a(%rip),%xmm8 # 5170 <_sk_callback_sse2+0xa3e>
+ .byte 15,89,21,131,37,0,0 // mulps 0x2583(%rip),%xmm2 # 5180 <_sk_callback_sse2+0xa4e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -25958,7 +28901,7 @@ _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 15,89,29,113,36,0,0 // mulps 0x2471(%rip),%xmm3 # 47c0 <_sk_callback_sse2+0x964>
+ .byte 15,89,29,107,37,0,0 // mulps 0x256b(%rip),%xmm3 # 5190 <_sk_callback_sse2+0xa5e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
.byte 102,15,239,201 // pxor %xmm1,%xmm1
@@ -26003,7 +28946,7 @@ _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 15,89,29,224,35,0,0 // mulps 0x23e0(%rip),%xmm3 # 47d0 <_sk_callback_sse2+0x974>
+ .byte 15,89,29,218,36,0,0 // mulps 0x24da(%rip),%xmm3 # 51a0 <_sk_callback_sse2+0xa6e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
.byte 102,15,239,201 // pxor %xmm1,%xmm1
@@ -26016,7 +28959,7 @@ FUNCTION(_sk_store_a8_sse2)
_sk_store_a8_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,212,35,0,0 // movaps 0x23d4(%rip),%xmm8 # 47e0 <_sk_callback_sse2+0x984>
+ .byte 68,15,40,5,206,36,0,0 // movaps 0x24ce(%rip),%xmm8 # 51b0 <_sk_callback_sse2+0xa7e>
.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
@@ -26038,9 +28981,9 @@ _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 15,89,5,155,35,0,0 // mulps 0x239b(%rip),%xmm0 # 47f0 <_sk_callback_sse2+0x994>
+ .byte 15,89,5,149,36,0,0 // mulps 0x2495(%rip),%xmm0 # 51c0 <_sk_callback_sse2+0xa8e>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,162,35,0,0 // movaps 0x23a2(%rip),%xmm3 # 4800 <_sk_callback_sse2+0x9a4>
+ .byte 15,40,29,156,36,0,0 // movaps 0x249c(%rip),%xmm3 # 51d0 <_sk_callback_sse2+0xa9e>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 15,40,208 // movaps %xmm0,%xmm2
.byte 255,224 // jmpq *%rax
@@ -26083,9 +29026,9 @@ _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 15,89,5,23,35,0,0 // mulps 0x2317(%rip),%xmm0 # 4810 <_sk_callback_sse2+0x9b4>
+ .byte 15,89,5,17,36,0,0 // mulps 0x2411(%rip),%xmm0 # 51e0 <_sk_callback_sse2+0xaae>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,30,35,0,0 // movaps 0x231e(%rip),%xmm3 # 4820 <_sk_callback_sse2+0x9c4>
+ .byte 15,40,29,24,36,0,0 // movaps 0x2418(%rip),%xmm3 # 51f0 <_sk_callback_sse2+0xabe>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 15,40,208 // movaps %xmm0,%xmm2
.byte 255,224 // jmpq *%rax
@@ -26097,9 +29040,9 @@ _sk_gather_i8_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 73,137,192 // mov %rax,%r8
.byte 77,133,192 // test %r8,%r8
- .byte 116,5 // je 2519 <_sk_gather_i8_sse2+0xf>
+ .byte 116,5 // je 2def <_sk_gather_i8_sse2+0xf>
.byte 76,137,192 // mov %r8,%rax
- .byte 235,2 // jmp 251b <_sk_gather_i8_sse2+0x11>
+ .byte 235,2 // jmp 2df1 <_sk_gather_i8_sse2+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
@@ -26148,11 +29091,11 @@ _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 102,15,111,21,61,34,0,0 // movdqa 0x223d(%rip),%xmm2 # 4830 <_sk_callback_sse2+0x9d4>
+ .byte 102,15,111,21,55,35,0,0 // movdqa 0x2337(%rip),%xmm2 # 5200 <_sk_callback_sse2+0xace>
.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 68,15,40,5,57,34,0,0 // movaps 0x2239(%rip),%xmm8 # 4840 <_sk_callback_sse2+0x9e4>
+ .byte 68,15,40,5,51,35,0,0 // movaps 0x2333(%rip),%xmm8 # 5210 <_sk_callback_sse2+0xade>
.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
@@ -26179,19 +29122,19 @@ _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 102,15,111,5,239,33,0,0 // movdqa 0x21ef(%rip),%xmm0 # 4850 <_sk_callback_sse2+0x9f4>
+ .byte 102,15,111,5,233,34,0,0 // movdqa 0x22e9(%rip),%xmm0 # 5220 <_sk_callback_sse2+0xaee>
.byte 102,15,219,194 // pand %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,241,33,0,0 // mulps 0x21f1(%rip),%xmm0 # 4860 <_sk_callback_sse2+0xa04>
- .byte 102,15,111,13,249,33,0,0 // movdqa 0x21f9(%rip),%xmm1 # 4870 <_sk_callback_sse2+0xa14>
+ .byte 15,89,5,235,34,0,0 // mulps 0x22eb(%rip),%xmm0 # 5230 <_sk_callback_sse2+0xafe>
+ .byte 102,15,111,13,243,34,0,0 // movdqa 0x22f3(%rip),%xmm1 # 5240 <_sk_callback_sse2+0xb0e>
.byte 102,15,219,202 // pand %xmm2,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,251,33,0,0 // mulps 0x21fb(%rip),%xmm1 # 4880 <_sk_callback_sse2+0xa24>
- .byte 102,15,219,21,3,34,0,0 // pand 0x2203(%rip),%xmm2 # 4890 <_sk_callback_sse2+0xa34>
+ .byte 15,89,13,245,34,0,0 // mulps 0x22f5(%rip),%xmm1 # 5250 <_sk_callback_sse2+0xb1e>
+ .byte 102,15,219,21,253,34,0,0 // pand 0x22fd(%rip),%xmm2 # 5260 <_sk_callback_sse2+0xb2e>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,9,34,0,0 // mulps 0x2209(%rip),%xmm2 # 48a0 <_sk_callback_sse2+0xa44>
+ .byte 15,89,21,3,35,0,0 // mulps 0x2303(%rip),%xmm2 # 5270 <_sk_callback_sse2+0xb3e>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,16,34,0,0 // movaps 0x2210(%rip),%xmm3 # 48b0 <_sk_callback_sse2+0xa54>
+ .byte 15,40,29,10,35,0,0 // movaps 0x230a(%rip),%xmm3 # 5280 <_sk_callback_sse2+0xb4e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_gather_565_sse2
@@ -26226,19 +29169,19 @@ _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 102,15,111,5,153,33,0,0 // movdqa 0x2199(%rip),%xmm0 # 48c0 <_sk_callback_sse2+0xa64>
+ .byte 102,15,111,5,147,34,0,0 // movdqa 0x2293(%rip),%xmm0 # 5290 <_sk_callback_sse2+0xb5e>
.byte 102,15,219,194 // pand %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,155,33,0,0 // mulps 0x219b(%rip),%xmm0 # 48d0 <_sk_callback_sse2+0xa74>
- .byte 102,15,111,13,163,33,0,0 // movdqa 0x21a3(%rip),%xmm1 # 48e0 <_sk_callback_sse2+0xa84>
+ .byte 15,89,5,149,34,0,0 // mulps 0x2295(%rip),%xmm0 # 52a0 <_sk_callback_sse2+0xb6e>
+ .byte 102,15,111,13,157,34,0,0 // movdqa 0x229d(%rip),%xmm1 # 52b0 <_sk_callback_sse2+0xb7e>
.byte 102,15,219,202 // pand %xmm2,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,165,33,0,0 // mulps 0x21a5(%rip),%xmm1 # 48f0 <_sk_callback_sse2+0xa94>
- .byte 102,15,219,21,173,33,0,0 // pand 0x21ad(%rip),%xmm2 # 4900 <_sk_callback_sse2+0xaa4>
+ .byte 15,89,13,159,34,0,0 // mulps 0x229f(%rip),%xmm1 # 52c0 <_sk_callback_sse2+0xb8e>
+ .byte 102,15,219,21,167,34,0,0 // pand 0x22a7(%rip),%xmm2 # 52d0 <_sk_callback_sse2+0xb9e>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,179,33,0,0 // mulps 0x21b3(%rip),%xmm2 # 4910 <_sk_callback_sse2+0xab4>
+ .byte 15,89,21,173,34,0,0 // mulps 0x22ad(%rip),%xmm2 # 52e0 <_sk_callback_sse2+0xbae>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,186,33,0,0 // movaps 0x21ba(%rip),%xmm3 # 4920 <_sk_callback_sse2+0xac4>
+ .byte 15,40,29,180,34,0,0 // movaps 0x22b4(%rip),%xmm3 # 52f0 <_sk_callback_sse2+0xbbe>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_565_sse2
@@ -26247,12 +29190,12 @@ FUNCTION(_sk_store_565_sse2)
_sk_store_565_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,187,33,0,0 // movaps 0x21bb(%rip),%xmm8 # 4930 <_sk_callback_sse2+0xad4>
+ .byte 68,15,40,5,181,34,0,0 // movaps 0x22b5(%rip),%xmm8 # 5300 <_sk_callback_sse2+0xbce>
.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 68,15,40,21,176,33,0,0 // movaps 0x21b0(%rip),%xmm10 # 4940 <_sk_callback_sse2+0xae4>
+ .byte 68,15,40,21,170,34,0,0 // movaps 0x22aa(%rip),%xmm10 # 5310 <_sk_callback_sse2+0xbde>
.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
@@ -26276,21 +29219,21 @@ _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 102,15,111,5,105,33,0,0 // movdqa 0x2169(%rip),%xmm0 # 4950 <_sk_callback_sse2+0xaf4>
+ .byte 102,15,111,5,99,34,0,0 // movdqa 0x2263(%rip),%xmm0 # 5320 <_sk_callback_sse2+0xbee>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,107,33,0,0 // mulps 0x216b(%rip),%xmm0 # 4960 <_sk_callback_sse2+0xb04>
- .byte 102,15,111,13,115,33,0,0 // movdqa 0x2173(%rip),%xmm1 # 4970 <_sk_callback_sse2+0xb14>
+ .byte 15,89,5,101,34,0,0 // mulps 0x2265(%rip),%xmm0 # 5330 <_sk_callback_sse2+0xbfe>
+ .byte 102,15,111,13,109,34,0,0 // movdqa 0x226d(%rip),%xmm1 # 5340 <_sk_callback_sse2+0xc0e>
.byte 102,15,219,203 // pand %xmm3,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,117,33,0,0 // mulps 0x2175(%rip),%xmm1 # 4980 <_sk_callback_sse2+0xb24>
- .byte 102,15,111,21,125,33,0,0 // movdqa 0x217d(%rip),%xmm2 # 4990 <_sk_callback_sse2+0xb34>
+ .byte 15,89,13,111,34,0,0 // mulps 0x226f(%rip),%xmm1 # 5350 <_sk_callback_sse2+0xc1e>
+ .byte 102,15,111,21,119,34,0,0 // movdqa 0x2277(%rip),%xmm2 # 5360 <_sk_callback_sse2+0xc2e>
.byte 102,15,219,211 // pand %xmm3,%xmm2
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,127,33,0,0 // mulps 0x217f(%rip),%xmm2 # 49a0 <_sk_callback_sse2+0xb44>
- .byte 102,15,219,29,135,33,0,0 // pand 0x2187(%rip),%xmm3 # 49b0 <_sk_callback_sse2+0xb54>
+ .byte 15,89,21,121,34,0,0 // mulps 0x2279(%rip),%xmm2 # 5370 <_sk_callback_sse2+0xc3e>
+ .byte 102,15,219,29,129,34,0,0 // pand 0x2281(%rip),%xmm3 # 5380 <_sk_callback_sse2+0xc4e>
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,141,33,0,0 // mulps 0x218d(%rip),%xmm3 # 49c0 <_sk_callback_sse2+0xb64>
+ .byte 15,89,29,135,34,0,0 // mulps 0x2287(%rip),%xmm3 # 5390 <_sk_callback_sse2+0xc5e>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -26326,21 +29269,21 @@ _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 102,15,111,5,20,33,0,0 // movdqa 0x2114(%rip),%xmm0 # 49d0 <_sk_callback_sse2+0xb74>
+ .byte 102,15,111,5,14,34,0,0 // movdqa 0x220e(%rip),%xmm0 # 53a0 <_sk_callback_sse2+0xc6e>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,22,33,0,0 // mulps 0x2116(%rip),%xmm0 # 49e0 <_sk_callback_sse2+0xb84>
- .byte 102,15,111,13,30,33,0,0 // movdqa 0x211e(%rip),%xmm1 # 49f0 <_sk_callback_sse2+0xb94>
+ .byte 15,89,5,16,34,0,0 // mulps 0x2210(%rip),%xmm0 # 53b0 <_sk_callback_sse2+0xc7e>
+ .byte 102,15,111,13,24,34,0,0 // movdqa 0x2218(%rip),%xmm1 # 53c0 <_sk_callback_sse2+0xc8e>
.byte 102,15,219,203 // pand %xmm3,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,32,33,0,0 // mulps 0x2120(%rip),%xmm1 # 4a00 <_sk_callback_sse2+0xba4>
- .byte 102,15,111,21,40,33,0,0 // movdqa 0x2128(%rip),%xmm2 # 4a10 <_sk_callback_sse2+0xbb4>
+ .byte 15,89,13,26,34,0,0 // mulps 0x221a(%rip),%xmm1 # 53d0 <_sk_callback_sse2+0xc9e>
+ .byte 102,15,111,21,34,34,0,0 // movdqa 0x2222(%rip),%xmm2 # 53e0 <_sk_callback_sse2+0xcae>
.byte 102,15,219,211 // pand %xmm3,%xmm2
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,42,33,0,0 // mulps 0x212a(%rip),%xmm2 # 4a20 <_sk_callback_sse2+0xbc4>
- .byte 102,15,219,29,50,33,0,0 // pand 0x2132(%rip),%xmm3 # 4a30 <_sk_callback_sse2+0xbd4>
+ .byte 15,89,21,36,34,0,0 // mulps 0x2224(%rip),%xmm2 # 53f0 <_sk_callback_sse2+0xcbe>
+ .byte 102,15,219,29,44,34,0,0 // pand 0x222c(%rip),%xmm3 # 5400 <_sk_callback_sse2+0xcce>
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,56,33,0,0 // mulps 0x2138(%rip),%xmm3 # 4a40 <_sk_callback_sse2+0xbe4>
+ .byte 15,89,29,50,34,0,0 // mulps 0x2232(%rip),%xmm3 # 5410 <_sk_callback_sse2+0xcde>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -26350,7 +29293,7 @@ FUNCTION(_sk_store_4444_sse2)
_sk_store_4444_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,55,33,0,0 // movaps 0x2137(%rip),%xmm8 # 4a50 <_sk_callback_sse2+0xbf4>
+ .byte 68,15,40,5,49,34,0,0 // movaps 0x2231(%rip),%xmm8 # 5420 <_sk_callback_sse2+0xcee>
.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
@@ -26382,11 +29325,11 @@ _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 15,40,21,202,32,0,0 // movaps 0x20ca(%rip),%xmm2 # 4a60 <_sk_callback_sse2+0xc04>
+ .byte 15,40,21,196,33,0,0 // movaps 0x21c4(%rip),%xmm2 # 5430 <_sk_callback_sse2+0xcfe>
.byte 65,15,40,193 // movaps %xmm9,%xmm0
.byte 15,84,194 // andps %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,200,32,0,0 // movaps 0x20c8(%rip),%xmm8 # 4a70 <_sk_callback_sse2+0xc14>
+ .byte 68,15,40,5,194,33,0,0 // movaps 0x21c2(%rip),%xmm8 # 5440 <_sk_callback_sse2+0xd0e>
.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
@@ -26435,11 +29378,11 @@ _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 102,15,111,21,25,32,0,0 // movdqa 0x2019(%rip),%xmm2 # 4a80 <_sk_callback_sse2+0xc24>
+ .byte 102,15,111,21,19,33,0,0 // movdqa 0x2113(%rip),%xmm2 # 5450 <_sk_callback_sse2+0xd1e>
.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 68,15,40,5,21,32,0,0 // movaps 0x2015(%rip),%xmm8 # 4a90 <_sk_callback_sse2+0xc34>
+ .byte 68,15,40,5,15,33,0,0 // movaps 0x210f(%rip),%xmm8 # 5460 <_sk_callback_sse2+0xd2e>
.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
@@ -26463,7 +29406,7 @@ FUNCTION(_sk_store_8888_sse2)
_sk_store_8888_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,216,31,0,0 // movaps 0x1fd8(%rip),%xmm8 # 4aa0 <_sk_callback_sse2+0xc44>
+ .byte 68,15,40,5,210,32,0,0 // movaps 0x20d2(%rip),%xmm8 # 5470 <_sk_callback_sse2+0xd3e>
.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
@@ -26502,7 +29445,7 @@ _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 102,68,15,111,13,72,31,0,0 // movdqa 0x1f48(%rip),%xmm9 # 4ab0 <_sk_callback_sse2+0xc54>
+ .byte 102,68,15,111,13,66,32,0,0 // movdqa 0x2042(%rip),%xmm9 # 5480 <_sk_callback_sse2+0xd4e>
.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
@@ -26510,11 +29453,11 @@ _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 102,68,15,111,29,45,31,0,0 // movdqa 0x1f2d(%rip),%xmm11 # 4ac0 <_sk_callback_sse2+0xc64>
+ .byte 102,68,15,111,29,39,32,0,0 // movdqa 0x2027(%rip),%xmm11 # 5490 <_sk_callback_sse2+0xd5e>
.byte 102,69,15,254,235 // paddd %xmm11,%xmm13
- .byte 102,68,15,111,37,47,31,0,0 // movdqa 0x1f2f(%rip),%xmm12 # 4ad0 <_sk_callback_sse2+0xc74>
+ .byte 102,68,15,111,37,41,32,0,0 // movdqa 0x2029(%rip),%xmm12 # 54a0 <_sk_callback_sse2+0xd6e>
.byte 102,65,15,239,204 // pxor %xmm12,%xmm1
- .byte 102,15,111,29,50,31,0,0 // movdqa 0x1f32(%rip),%xmm3 # 4ae0 <_sk_callback_sse2+0xc84>
+ .byte 102,15,111,29,44,32,0,0 // movdqa 0x202c(%rip),%xmm3 # 54b0 <_sk_callback_sse2+0xd7e>
.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
@@ -26600,7 +29543,7 @@ _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 102,68,15,111,13,192,29,0,0 // movdqa 0x1dc0(%rip),%xmm9 # 4af0 <_sk_callback_sse2+0xc94>
+ .byte 102,68,15,111,13,186,30,0,0 // movdqa 0x1eba(%rip),%xmm9 # 54c0 <_sk_callback_sse2+0xd8e>
.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
@@ -26608,11 +29551,11 @@ _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 102,68,15,111,29,165,29,0,0 // movdqa 0x1da5(%rip),%xmm11 # 4b00 <_sk_callback_sse2+0xca4>
+ .byte 102,68,15,111,29,159,30,0,0 // movdqa 0x1e9f(%rip),%xmm11 # 54d0 <_sk_callback_sse2+0xd9e>
.byte 102,69,15,254,235 // paddd %xmm11,%xmm13
- .byte 102,68,15,111,37,167,29,0,0 // movdqa 0x1da7(%rip),%xmm12 # 4b10 <_sk_callback_sse2+0xcb4>
+ .byte 102,68,15,111,37,161,30,0,0 // movdqa 0x1ea1(%rip),%xmm12 # 54e0 <_sk_callback_sse2+0xdae>
.byte 102,65,15,239,204 // pxor %xmm12,%xmm1
- .byte 102,15,111,29,170,29,0,0 // movdqa 0x1daa(%rip),%xmm3 # 4b20 <_sk_callback_sse2+0xcc4>
+ .byte 102,15,111,29,164,30,0,0 // movdqa 0x1ea4(%rip),%xmm3 # 54f0 <_sk_callback_sse2+0xdbe>
.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
@@ -26665,17 +29608,17 @@ FUNCTION(_sk_store_f16_sse2)
_sk_store_f16_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 102,68,15,111,21,210,28,0,0 // movdqa 0x1cd2(%rip),%xmm10 # 4b30 <_sk_callback_sse2+0xcd4>
+ .byte 102,68,15,111,21,204,29,0,0 // movdqa 0x1dcc(%rip),%xmm10 # 5500 <_sk_callback_sse2+0xdce>
.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 102,68,15,111,13,197,28,0,0 // movdqa 0x1cc5(%rip),%xmm9 # 4b40 <_sk_callback_sse2+0xce4>
+ .byte 102,68,15,111,13,191,29,0,0 // movdqa 0x1dbf(%rip),%xmm9 # 5510 <_sk_callback_sse2+0xdde>
.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 102,68,15,111,29,182,28,0,0 // movdqa 0x1cb6(%rip),%xmm11 # 4b50 <_sk_callback_sse2+0xcf4>
+ .byte 102,68,15,111,29,176,29,0,0 // movdqa 0x1db0(%rip),%xmm11 # 5520 <_sk_callback_sse2+0xdee>
.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
@@ -26754,7 +29697,7 @@ _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 68,15,40,5,84,27,0,0 // movaps 0x1b54(%rip),%xmm8 # 4b60 <_sk_callback_sse2+0xd04>
+ .byte 68,15,40,5,78,28,0,0 // movaps 0x1c4e(%rip),%xmm8 # 5530 <_sk_callback_sse2+0xdfe>
.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
@@ -26807,7 +29750,7 @@ _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 68,15,40,13,144,26,0,0 // movaps 0x1a90(%rip),%xmm9 # 4b70 <_sk_callback_sse2+0xd14>
+ .byte 68,15,40,13,138,27,0,0 // movaps 0x1b8a(%rip),%xmm9 # 5540 <_sk_callback_sse2+0xe0e>
.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
@@ -26824,7 +29767,7 @@ _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 15,40,29,87,26,0,0 // movaps 0x1a57(%rip),%xmm3 # 4b80 <_sk_callback_sse2+0xd24>
+ .byte 15,40,29,81,27,0,0 // movaps 0x1b51(%rip),%xmm3 # 5550 <_sk_callback_sse2+0xe1e>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_u16_be_sse2
@@ -26833,7 +29776,7 @@ 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 68,15,40,13,88,26,0,0 // movaps 0x1a58(%rip),%xmm9 # 4b90 <_sk_callback_sse2+0xd34>
+ .byte 68,15,40,13,82,27,0,0 // movaps 0x1b52(%rip),%xmm9 # 5560 <_sk_callback_sse2+0xe2e>
.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
@@ -26983,7 +29926,7 @@ _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 68,15,84,13,66,24,0,0 // andps 0x1842(%rip),%xmm9 # 4ba0 <_sk_callback_sse2+0xd44>
+ .byte 68,15,84,13,60,25,0,0 // andps 0x193c(%rip),%xmm9 # 5570 <_sk_callback_sse2+0xe3e>
.byte 69,15,92,209 // subps %xmm9,%xmm10
.byte 69,15,89,208 // mulps %xmm8,%xmm10
.byte 65,15,92,194 // subps %xmm10,%xmm0
@@ -27005,7 +29948,7 @@ _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 68,15,84,13,10,24,0,0 // andps 0x180a(%rip),%xmm9 # 4bb0 <_sk_callback_sse2+0xd54>
+ .byte 68,15,84,13,4,25,0,0 // andps 0x1904(%rip),%xmm9 # 5580 <_sk_callback_sse2+0xe4e>
.byte 69,15,92,209 // subps %xmm9,%xmm10
.byte 69,15,89,208 // mulps %xmm8,%xmm10
.byte 65,15,92,202 // subps %xmm10,%xmm1
@@ -27031,7 +29974,7 @@ _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 68,15,84,21,192,23,0,0 // andps 0x17c0(%rip),%xmm10 # 4bc0 <_sk_callback_sse2+0xd64>
+ .byte 68,15,84,21,186,24,0,0 // andps 0x18ba(%rip),%xmm10 # 5590 <_sk_callback_sse2+0xe5e>
.byte 69,15,87,228 // xorps %xmm12,%xmm12
.byte 69,15,92,218 // subps %xmm10,%xmm11
.byte 69,15,89,216 // mulps %xmm8,%xmm11
@@ -27061,7 +30004,7 @@ _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 68,15,84,21,102,23,0,0 // andps 0x1766(%rip),%xmm10 # 4bd0 <_sk_callback_sse2+0xd74>
+ .byte 68,15,84,21,96,24,0,0 // andps 0x1860(%rip),%xmm10 # 55a0 <_sk_callback_sse2+0xe6e>
.byte 69,15,87,228 // xorps %xmm12,%xmm12
.byte 69,15,92,218 // subps %xmm10,%xmm11
.byte 69,15,89,216 // mulps %xmm8,%xmm11
@@ -27080,10 +30023,10 @@ HIDDEN _sk_luminance_to_alpha_sse2
FUNCTION(_sk_luminance_to_alpha_sse2)
_sk_luminance_to_alpha_sse2:
.byte 15,40,218 // movaps %xmm2,%xmm3
- .byte 15,89,5,62,23,0,0 // mulps 0x173e(%rip),%xmm0 # 4be0 <_sk_callback_sse2+0xd84>
- .byte 15,89,13,71,23,0,0 // mulps 0x1747(%rip),%xmm1 # 4bf0 <_sk_callback_sse2+0xd94>
+ .byte 15,89,5,56,24,0,0 // mulps 0x1838(%rip),%xmm0 # 55b0 <_sk_callback_sse2+0xe7e>
+ .byte 15,89,13,65,24,0,0 // mulps 0x1841(%rip),%xmm1 # 55c0 <_sk_callback_sse2+0xe8e>
.byte 15,88,200 // addps %xmm0,%xmm1
- .byte 15,89,29,77,23,0,0 // mulps 0x174d(%rip),%xmm3 # 4c00 <_sk_callback_sse2+0xda4>
+ .byte 15,89,29,71,24,0,0 // mulps 0x1847(%rip),%xmm3 # 55d0 <_sk_callback_sse2+0xe9e>
.byte 15,88,217 // addps %xmm1,%xmm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
@@ -27316,7 +30259,7 @@ _sk_linear_gradient_sse2:
.byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
.byte 72,139,8 // mov (%rax),%rcx
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,132,15,1,0,0 // je 3966 <_sk_linear_gradient_sse2+0x149>
+ .byte 15,132,15,1,0,0 // je 423c <_sk_linear_gradient_sse2+0x149>
.byte 72,139,64,8 // mov 0x8(%rax),%rax
.byte 72,131,192,32 // add $0x20,%rax
.byte 69,15,87,192 // xorps %xmm8,%xmm8
@@ -27377,8 +30320,8 @@ _sk_linear_gradient_sse2:
.byte 69,15,86,231 // orps %xmm15,%xmm12
.byte 72,131,192,36 // add $0x24,%rax
.byte 72,255,201 // dec %rcx
- .byte 15,133,8,255,255,255 // jne 386c <_sk_linear_gradient_sse2+0x4f>
- .byte 235,13 // jmp 3973 <_sk_linear_gradient_sse2+0x156>
+ .byte 15,133,8,255,255,255 // jne 4142 <_sk_linear_gradient_sse2+0x4f>
+ .byte 235,13 // jmp 4249 <_sk_linear_gradient_sse2+0x156>
.byte 15,87,201 // xorps %xmm1,%xmm1
.byte 15,87,210 // xorps %xmm2,%xmm2
.byte 15,87,219 // xorps %xmm3,%xmm3
@@ -27447,29 +30390,29 @@ _sk_xy_to_polar_unit_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 68,15,40,45,197,17,0,0 // movaps 0x11c5(%rip),%xmm13 # 4c10 <_sk_callback_sse2+0xdb4>
+ .byte 68,15,40,45,191,18,0,0 // movaps 0x12bf(%rip),%xmm13 # 55e0 <_sk_callback_sse2+0xeae>
.byte 69,15,89,236 // mulps %xmm12,%xmm13
- .byte 68,15,88,45,201,17,0,0 // addps 0x11c9(%rip),%xmm13 # 4c20 <_sk_callback_sse2+0xdc4>
+ .byte 68,15,88,45,195,18,0,0 // addps 0x12c3(%rip),%xmm13 # 55f0 <_sk_callback_sse2+0xebe>
.byte 69,15,89,236 // mulps %xmm12,%xmm13
- .byte 68,15,88,45,205,17,0,0 // addps 0x11cd(%rip),%xmm13 # 4c30 <_sk_callback_sse2+0xdd4>
+ .byte 68,15,88,45,199,18,0,0 // addps 0x12c7(%rip),%xmm13 # 5600 <_sk_callback_sse2+0xece>
.byte 69,15,89,236 // mulps %xmm12,%xmm13
- .byte 68,15,88,45,209,17,0,0 // addps 0x11d1(%rip),%xmm13 # 4c40 <_sk_callback_sse2+0xde4>
+ .byte 68,15,88,45,203,18,0,0 // addps 0x12cb(%rip),%xmm13 # 5610 <_sk_callback_sse2+0xede>
.byte 69,15,89,235 // mulps %xmm11,%xmm13
.byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9
- .byte 68,15,40,21,208,17,0,0 // movaps 0x11d0(%rip),%xmm10 # 4c50 <_sk_callback_sse2+0xdf4>
+ .byte 68,15,40,21,202,18,0,0 // movaps 0x12ca(%rip),%xmm10 # 5620 <_sk_callback_sse2+0xeee>
.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 68,15,40,21,195,17,0,0 // movaps 0x11c3(%rip),%xmm10 # 4c60 <_sk_callback_sse2+0xe04>
+ .byte 68,15,40,21,189,18,0,0 // movaps 0x12bd(%rip),%xmm10 # 5630 <_sk_callback_sse2+0xefe>
.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 68,15,40,21,178,17,0,0 // movaps 0x11b2(%rip),%xmm10 # 4c70 <_sk_callback_sse2+0xe14>
+ .byte 68,15,40,21,172,18,0,0 // movaps 0x12ac(%rip),%xmm10 # 5640 <_sk_callback_sse2+0xf0e>
.byte 69,15,92,208 // subps %xmm8,%xmm10
.byte 69,15,84,209 // andps %xmm9,%xmm10
.byte 69,15,85,200 // andnps %xmm8,%xmm9
@@ -27484,7 +30427,7 @@ HIDDEN _sk_save_xy_sse2
FUNCTION(_sk_save_xy_sse2)
_sk_save_xy_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,155,17,0,0 // movaps 0x119b(%rip),%xmm8 # 4c80 <_sk_callback_sse2+0xe24>
+ .byte 68,15,40,5,149,18,0,0 // movaps 0x1295(%rip),%xmm8 # 5650 <_sk_callback_sse2+0xf1e>
.byte 15,17,0 // movups %xmm0,(%rax)
.byte 68,15,40,200 // movaps %xmm0,%xmm9
.byte 69,15,88,200 // addps %xmm8,%xmm9
@@ -27492,7 +30435,7 @@ _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 68,15,40,37,134,17,0,0 // movaps 0x1186(%rip),%xmm12 # 4c90 <_sk_callback_sse2+0xe34>
+ .byte 68,15,40,37,128,18,0,0 // movaps 0x1280(%rip),%xmm12 # 5660 <_sk_callback_sse2+0xf2e>
.byte 69,15,84,220 // andps %xmm12,%xmm11
.byte 69,15,92,211 // subps %xmm11,%xmm10
.byte 69,15,92,202 // subps %xmm10,%xmm9
@@ -27539,8 +30482,8 @@ _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 15,88,5,255,16,0,0 // addps 0x10ff(%rip),%xmm0 # 4ca0 <_sk_callback_sse2+0xe44>
- .byte 68,15,40,13,7,17,0,0 // movaps 0x1107(%rip),%xmm9 # 4cb0 <_sk_callback_sse2+0xe54>
+ .byte 15,88,5,249,17,0,0 // addps 0x11f9(%rip),%xmm0 # 5670 <_sk_callback_sse2+0xf3e>
+ .byte 68,15,40,13,1,18,0,0 // movaps 0x1201(%rip),%xmm9 # 5680 <_sk_callback_sse2+0xf4e>
.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
@@ -27553,7 +30496,7 @@ _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 15,88,5,246,16,0,0 // addps 0x10f6(%rip),%xmm0 # 4cc0 <_sk_callback_sse2+0xe64>
+ .byte 15,88,5,240,17,0,0 // addps 0x11f0(%rip),%xmm0 # 5690 <_sk_callback_sse2+0xf5e>
.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
@@ -27565,8 +30508,8 @@ _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 15,88,13,232,16,0,0 // addps 0x10e8(%rip),%xmm1 # 4cd0 <_sk_callback_sse2+0xe74>
- .byte 68,15,40,13,240,16,0,0 // movaps 0x10f0(%rip),%xmm9 # 4ce0 <_sk_callback_sse2+0xe84>
+ .byte 15,88,13,226,17,0,0 // addps 0x11e2(%rip),%xmm1 # 56a0 <_sk_callback_sse2+0xf6e>
+ .byte 68,15,40,13,234,17,0,0 // movaps 0x11ea(%rip),%xmm9 # 56b0 <_sk_callback_sse2+0xf7e>
.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
@@ -27579,7 +30522,7 @@ _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 15,88,13,222,16,0,0 // addps 0x10de(%rip),%xmm1 # 4cf0 <_sk_callback_sse2+0xe94>
+ .byte 15,88,13,216,17,0,0 // addps 0x11d8(%rip),%xmm1 # 56c0 <_sk_callback_sse2+0xf8e>
.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
@@ -27591,13 +30534,13 @@ _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 15,88,5,209,16,0,0 // addps 0x10d1(%rip),%xmm0 # 4d00 <_sk_callback_sse2+0xea4>
- .byte 68,15,40,13,217,16,0,0 // movaps 0x10d9(%rip),%xmm9 # 4d10 <_sk_callback_sse2+0xeb4>
+ .byte 15,88,5,203,17,0,0 // addps 0x11cb(%rip),%xmm0 # 56d0 <_sk_callback_sse2+0xf9e>
+ .byte 68,15,40,13,211,17,0,0 // movaps 0x11d3(%rip),%xmm9 # 56e0 <_sk_callback_sse2+0xfae>
.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 68,15,89,13,213,16,0,0 // mulps 0x10d5(%rip),%xmm9 # 4d20 <_sk_callback_sse2+0xec4>
- .byte 68,15,88,13,221,16,0,0 // addps 0x10dd(%rip),%xmm9 # 4d30 <_sk_callback_sse2+0xed4>
+ .byte 68,15,89,13,207,17,0,0 // mulps 0x11cf(%rip),%xmm9 # 56f0 <_sk_callback_sse2+0xfbe>
+ .byte 68,15,88,13,215,17,0,0 // addps 0x11d7(%rip),%xmm9 # 5700 <_sk_callback_sse2+0xfce>
.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
@@ -27610,16 +30553,16 @@ _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 15,88,5,204,16,0,0 // addps 0x10cc(%rip),%xmm0 # 4d40 <_sk_callback_sse2+0xee4>
- .byte 68,15,40,13,212,16,0,0 // movaps 0x10d4(%rip),%xmm9 # 4d50 <_sk_callback_sse2+0xef4>
+ .byte 15,88,5,198,17,0,0 // addps 0x11c6(%rip),%xmm0 # 5710 <_sk_callback_sse2+0xfde>
+ .byte 68,15,40,13,206,17,0,0 // movaps 0x11ce(%rip),%xmm9 # 5720 <_sk_callback_sse2+0xfee>
.byte 69,15,92,200 // subps %xmm8,%xmm9
- .byte 68,15,40,5,216,16,0,0 // movaps 0x10d8(%rip),%xmm8 # 4d60 <_sk_callback_sse2+0xf04>
+ .byte 68,15,40,5,210,17,0,0 // movaps 0x11d2(%rip),%xmm8 # 5730 <_sk_callback_sse2+0xffe>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,220,16,0,0 // addps 0x10dc(%rip),%xmm8 # 4d70 <_sk_callback_sse2+0xf14>
+ .byte 68,15,88,5,214,17,0,0 // addps 0x11d6(%rip),%xmm8 # 5740 <_sk_callback_sse2+0x100e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,224,16,0,0 // addps 0x10e0(%rip),%xmm8 # 4d80 <_sk_callback_sse2+0xf24>
+ .byte 68,15,88,5,218,17,0,0 // addps 0x11da(%rip),%xmm8 # 5750 <_sk_callback_sse2+0x101e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,228,16,0,0 // addps 0x10e4(%rip),%xmm8 # 4d90 <_sk_callback_sse2+0xf34>
+ .byte 68,15,88,5,222,17,0,0 // addps 0x11de(%rip),%xmm8 # 5760 <_sk_callback_sse2+0x102e>
.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
@@ -27629,17 +30572,17 @@ HIDDEN _sk_bicubic_p1x_sse2
FUNCTION(_sk_bicubic_p1x_sse2)
_sk_bicubic_p1x_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,222,16,0,0 // movaps 0x10de(%rip),%xmm8 # 4da0 <_sk_callback_sse2+0xf44>
+ .byte 68,15,40,5,216,17,0,0 // movaps 0x11d8(%rip),%xmm8 # 5770 <_sk_callback_sse2+0x103e>
.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 68,15,40,21,218,16,0,0 // movaps 0x10da(%rip),%xmm10 # 4db0 <_sk_callback_sse2+0xf54>
+ .byte 68,15,40,21,212,17,0,0 // movaps 0x11d4(%rip),%xmm10 # 5780 <_sk_callback_sse2+0x104e>
.byte 69,15,89,209 // mulps %xmm9,%xmm10
- .byte 68,15,88,21,222,16,0,0 // addps 0x10de(%rip),%xmm10 # 4dc0 <_sk_callback_sse2+0xf64>
+ .byte 68,15,88,21,216,17,0,0 // addps 0x11d8(%rip),%xmm10 # 5790 <_sk_callback_sse2+0x105e>
.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 68,15,88,21,218,16,0,0 // addps 0x10da(%rip),%xmm10 # 4dd0 <_sk_callback_sse2+0xf74>
+ .byte 68,15,88,21,212,17,0,0 // addps 0x11d4(%rip),%xmm10 # 57a0 <_sk_callback_sse2+0x106e>
.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
@@ -27651,11 +30594,11 @@ _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 15,88,5,205,16,0,0 // addps 0x10cd(%rip),%xmm0 # 4de0 <_sk_callback_sse2+0xf84>
+ .byte 15,88,5,199,17,0,0 // addps 0x11c7(%rip),%xmm0 # 57b0 <_sk_callback_sse2+0x107e>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,89,201 // mulps %xmm9,%xmm9
- .byte 68,15,89,5,205,16,0,0 // mulps 0x10cd(%rip),%xmm8 # 4df0 <_sk_callback_sse2+0xf94>
- .byte 68,15,88,5,213,16,0,0 // addps 0x10d5(%rip),%xmm8 # 4e00 <_sk_callback_sse2+0xfa4>
+ .byte 68,15,89,5,199,17,0,0 // mulps 0x11c7(%rip),%xmm8 # 57c0 <_sk_callback_sse2+0x108e>
+ .byte 68,15,88,5,207,17,0,0 // addps 0x11cf(%rip),%xmm8 # 57d0 <_sk_callback_sse2+0x109e>
.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
@@ -27668,13 +30611,13 @@ _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 15,88,13,195,16,0,0 // addps 0x10c3(%rip),%xmm1 # 4e10 <_sk_callback_sse2+0xfb4>
- .byte 68,15,40,13,203,16,0,0 // movaps 0x10cb(%rip),%xmm9 # 4e20 <_sk_callback_sse2+0xfc4>
+ .byte 15,88,13,189,17,0,0 // addps 0x11bd(%rip),%xmm1 # 57e0 <_sk_callback_sse2+0x10ae>
+ .byte 68,15,40,13,197,17,0,0 // movaps 0x11c5(%rip),%xmm9 # 57f0 <_sk_callback_sse2+0x10be>
.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 68,15,89,13,199,16,0,0 // mulps 0x10c7(%rip),%xmm9 # 4e30 <_sk_callback_sse2+0xfd4>
- .byte 68,15,88,13,207,16,0,0 // addps 0x10cf(%rip),%xmm9 # 4e40 <_sk_callback_sse2+0xfe4>
+ .byte 68,15,89,13,193,17,0,0 // mulps 0x11c1(%rip),%xmm9 # 5800 <_sk_callback_sse2+0x10ce>
+ .byte 68,15,88,13,201,17,0,0 // addps 0x11c9(%rip),%xmm9 # 5810 <_sk_callback_sse2+0x10de>
.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
@@ -27687,16 +30630,16 @@ _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 15,88,13,189,16,0,0 // addps 0x10bd(%rip),%xmm1 # 4e50 <_sk_callback_sse2+0xff4>
- .byte 68,15,40,13,197,16,0,0 // movaps 0x10c5(%rip),%xmm9 # 4e60 <_sk_callback_sse2+0x1004>
+ .byte 15,88,13,183,17,0,0 // addps 0x11b7(%rip),%xmm1 # 5820 <_sk_callback_sse2+0x10ee>
+ .byte 68,15,40,13,191,17,0,0 // movaps 0x11bf(%rip),%xmm9 # 5830 <_sk_callback_sse2+0x10fe>
.byte 69,15,92,200 // subps %xmm8,%xmm9
- .byte 68,15,40,5,201,16,0,0 // movaps 0x10c9(%rip),%xmm8 # 4e70 <_sk_callback_sse2+0x1014>
+ .byte 68,15,40,5,195,17,0,0 // movaps 0x11c3(%rip),%xmm8 # 5840 <_sk_callback_sse2+0x110e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,205,16,0,0 // addps 0x10cd(%rip),%xmm8 # 4e80 <_sk_callback_sse2+0x1024>
+ .byte 68,15,88,5,199,17,0,0 // addps 0x11c7(%rip),%xmm8 # 5850 <_sk_callback_sse2+0x111e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,209,16,0,0 // addps 0x10d1(%rip),%xmm8 # 4e90 <_sk_callback_sse2+0x1034>
+ .byte 68,15,88,5,203,17,0,0 // addps 0x11cb(%rip),%xmm8 # 5860 <_sk_callback_sse2+0x112e>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,213,16,0,0 // addps 0x10d5(%rip),%xmm8 # 4ea0 <_sk_callback_sse2+0x1044>
+ .byte 68,15,88,5,207,17,0,0 // addps 0x11cf(%rip),%xmm8 # 5870 <_sk_callback_sse2+0x113e>
.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
@@ -27706,17 +30649,17 @@ HIDDEN _sk_bicubic_p1y_sse2
FUNCTION(_sk_bicubic_p1y_sse2)
_sk_bicubic_p1y_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,207,16,0,0 // movaps 0x10cf(%rip),%xmm8 # 4eb0 <_sk_callback_sse2+0x1054>
+ .byte 68,15,40,5,201,17,0,0 // movaps 0x11c9(%rip),%xmm8 # 5880 <_sk_callback_sse2+0x114e>
.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 68,15,40,21,202,16,0,0 // movaps 0x10ca(%rip),%xmm10 # 4ec0 <_sk_callback_sse2+0x1064>
+ .byte 68,15,40,21,196,17,0,0 // movaps 0x11c4(%rip),%xmm10 # 5890 <_sk_callback_sse2+0x115e>
.byte 69,15,89,209 // mulps %xmm9,%xmm10
- .byte 68,15,88,21,206,16,0,0 // addps 0x10ce(%rip),%xmm10 # 4ed0 <_sk_callback_sse2+0x1074>
+ .byte 68,15,88,21,200,17,0,0 // addps 0x11c8(%rip),%xmm10 # 58a0 <_sk_callback_sse2+0x116e>
.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 68,15,88,21,202,16,0,0 // addps 0x10ca(%rip),%xmm10 # 4ee0 <_sk_callback_sse2+0x1084>
+ .byte 68,15,88,21,196,17,0,0 // addps 0x11c4(%rip),%xmm10 # 58b0 <_sk_callback_sse2+0x117e>
.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
@@ -27728,11 +30671,11 @@ _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 15,88,13,188,16,0,0 // addps 0x10bc(%rip),%xmm1 # 4ef0 <_sk_callback_sse2+0x1094>
+ .byte 15,88,13,182,17,0,0 // addps 0x11b6(%rip),%xmm1 # 58c0 <_sk_callback_sse2+0x118e>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,89,201 // mulps %xmm9,%xmm9
- .byte 68,15,89,5,188,16,0,0 // mulps 0x10bc(%rip),%xmm8 # 4f00 <_sk_callback_sse2+0x10a4>
- .byte 68,15,88,5,196,16,0,0 // addps 0x10c4(%rip),%xmm8 # 4f10 <_sk_callback_sse2+0x10b4>
+ .byte 68,15,89,5,182,17,0,0 // mulps 0x11b6(%rip),%xmm8 # 58d0 <_sk_callback_sse2+0x119e>
+ .byte 68,15,88,5,190,17,0,0 // addps 0x11be(%rip),%xmm8 # 58e0 <_sk_callback_sse2+0x11ae>
.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
@@ -27951,11 +30894,135 @@ 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 4118 <.literal16+0x1d8>
+ .byte 224,64 // loopne 49e8 <.literal16+0x1d8>
+ .byte 0,0 // add %al,(%rax)
+ .byte 224,64 // loopne 49ec <.literal16+0x1dc>
.byte 0,0 // add %al,(%rax)
- .byte 224,64 // loopne 411c <.literal16+0x1dc>
+ .byte 224,64 // loopne 49f0 <.literal16+0x1e0>
.byte 0,0 // add %al,(%rax)
- .byte 224,64 // loopne 4120 <.literal16+0x1e0>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a21 <.literal16+0x211>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a25 <.literal16+0x215>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a29 <.literal16+0x219>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a2d <.literal16+0x21d>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a61 <.literal16+0x251>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a65 <.literal16+0x255>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a69 <.literal16+0x259>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4a6d <.literal16+0x25d>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4aa1 <.literal16+0x291>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4aa5 <.literal16+0x295>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4aa9 <.literal16+0x299>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4aad <.literal16+0x29d>
+ .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 128,63,154 // cmpb $0x9a,(%rdi)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,154 // ds (bad)
+ .byte 153 // cltd
+ .byte 153 // cltd
+ .byte 62,61,10,23,63,61 // ds cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 61,10,23,63,61 // cmp $0x3d3f170a,%eax
+ .byte 10,23 // or (%rdi),%dl
+ .byte 63 // (bad)
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4ae1 <.literal16+0x2d1>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4ae5 <.literal16+0x2d5>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4ae9 <.literal16+0x2d9>
+ .byte 174 // scas %es:(%rdi),%al
+ .byte 71,225,61 // rex.RXB loope 4aed <.literal16+0x2dd>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
@@ -28179,13 +31246,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 42e9 <.literal16+0x3a9>
+ .byte 224,7 // loopne 4cb9 <.literal16+0x4a9>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 42ed <.literal16+0x3ad>
+ .byte 224,7 // loopne 4cbd <.literal16+0x4ad>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 42f1 <.literal16+0x3b1>
+ .byte 224,7 // loopne 4cc1 <.literal16+0x4b1>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 42f5 <.literal16+0x3b5>
+ .byte 224,7 // loopne 4cc5 <.literal16+0x4b5>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -28254,11 +31321,11 @@ 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 43db <.literal16+0x49b>
+ .byte 127,67 // jg 4dab <.literal16+0x59b>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 43df <.literal16+0x49f>
+ .byte 127,67 // jg 4daf <.literal16+0x59f>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 43e3 <.literal16+0x4a3>
+ .byte 127,67 // jg 4db3 <.literal16+0x5a3>
.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)
@@ -28273,16 +31340,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 43d4 <.literal16+0x494>
+ .byte 127,0 // jg 4da4 <.literal16+0x594>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 43d8 <.literal16+0x498>
+ .byte 127,0 // jg 4da8 <.literal16+0x598>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 43dc <.literal16+0x49c>
+ .byte 127,0 // jg 4dac <.literal16+0x59c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 43e0 <.literal16+0x4a0>
+ .byte 127,0 // jg 4db0 <.literal16+0x5a0>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -28291,7 +31358,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4465 <.literal16+0x525>
+ .byte 119,115 // ja 4e35 <.literal16+0x625>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -28302,7 +31369,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 43c9 <.literal16+0x489>
+ .byte 117,191 // jne 4d99 <.literal16+0x589>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -28314,7 +31381,7 @@ 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 ffffffffe9a3840a <_sk_callback_sse2+0xffffffffe9a345ae>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a38dda <_sk_callback_sse2+0xffffffffe9a346a8>
.byte 220,63 // fdivrl (%rdi)
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
@@ -28368,16 +31435,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 44a4 <.literal16+0x564>
+ .byte 127,0 // jg 4e74 <.literal16+0x664>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 44a8 <.literal16+0x568>
+ .byte 127,0 // jg 4e78 <.literal16+0x668>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 44ac <.literal16+0x56c>
+ .byte 127,0 // jg 4e7c <.literal16+0x66c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 44b0 <.literal16+0x570>
+ .byte 127,0 // jg 4e80 <.literal16+0x670>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -28386,7 +31453,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4535 <.literal16+0x5f5>
+ .byte 119,115 // ja 4f05 <.literal16+0x6f5>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -28397,7 +31464,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 4499 <.literal16+0x559>
+ .byte 117,191 // jne 4e69 <.literal16+0x659>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -28409,7 +31476,7 @@ 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 ffffffffe9a384da <_sk_callback_sse2+0xffffffffe9a3467e>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a38eaa <_sk_callback_sse2+0xffffffffe9a34778>
.byte 220,63 // fdivrl (%rdi)
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
@@ -28463,16 +31530,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4574 <.literal16+0x634>
+ .byte 127,0 // jg 4f44 <.literal16+0x734>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4578 <.literal16+0x638>
+ .byte 127,0 // jg 4f48 <.literal16+0x738>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 457c <.literal16+0x63c>
+ .byte 127,0 // jg 4f4c <.literal16+0x73c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4580 <.literal16+0x640>
+ .byte 127,0 // jg 4f50 <.literal16+0x740>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -28481,7 +31548,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4605 <.literal16+0x6c5>
+ .byte 119,115 // ja 4fd5 <.literal16+0x7c5>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -28492,7 +31559,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 4569 <.literal16+0x629>
+ .byte 117,191 // jne 4f39 <.literal16+0x729>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -28504,7 +31571,7 @@ 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 ffffffffe9a385aa <_sk_callback_sse2+0xffffffffe9a3474e>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a38f7a <_sk_callback_sse2+0xffffffffe9a34848>
.byte 220,63 // fdivrl (%rdi)
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
@@ -28558,16 +31625,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4644 <.literal16+0x704>
+ .byte 127,0 // jg 5014 <.literal16+0x804>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4648 <.literal16+0x708>
+ .byte 127,0 // jg 5018 <.literal16+0x808>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 464c <.literal16+0x70c>
+ .byte 127,0 // jg 501c <.literal16+0x80c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4650 <.literal16+0x710>
+ .byte 127,0 // jg 5020 <.literal16+0x810>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -28576,7 +31643,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 46d5 <.literal16+0x795>
+ .byte 119,115 // ja 50a5 <.literal16+0x895>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -28587,7 +31654,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 4639 <.literal16+0x6f9>
+ .byte 117,191 // jne 5009 <.literal16+0x7f9>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -28599,7 +31666,7 @@ 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 ffffffffe9a3867a <_sk_callback_sse2+0xffffffffe9a3481e>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a3904a <_sk_callback_sse2+0xffffffffe9a34918>
.byte 220,63 // fdivrl (%rdi)
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
@@ -28649,13 +31716,13 @@ 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 4757 <.literal16+0x817>
+ .byte 127,67 // jg 5127 <.literal16+0x917>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 475b <.literal16+0x81b>
+ .byte 127,67 // jg 512b <.literal16+0x91b>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 475f <.literal16+0x81f>
+ .byte 127,67 // jg 512f <.literal16+0x91f>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 4763 <.literal16+0x823>
+ .byte 127,67 // jg 5133 <.literal16+0x923>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -28702,16 +31769,16 @@ BALIGN16
.byte 128,3,62 // addb $0x3e,(%rbx)
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 47e3 <.literal16+0x8a3>
+ .byte 118,63 // jbe 51b3 <.literal16+0x9a3>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 47e7 <.literal16+0x8a7>
+ .byte 118,63 // jbe 51b7 <.literal16+0x9a7>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 47eb <.literal16+0x8ab>
+ .byte 118,63 // jbe 51bb <.literal16+0x9ab>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 47ef <.literal16+0x8af>
+ .byte 118,63 // jbe 51bf <.literal16+0x9af>
.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)
@@ -28723,11 +31790,11 @@ 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 482b <.literal16+0x8eb>
+ .byte 127,67 // jg 51fb <.literal16+0x9eb>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 482f <.literal16+0x8ef>
+ .byte 127,67 // jg 51ff <.literal16+0x9ef>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 4833 <.literal16+0x8f3>
+ .byte 127,67 // jg 5203 <.literal16+0x9f3>
.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)
@@ -28767,13 +31834,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4879 <.literal16+0x939>
+ .byte 224,7 // loopne 5249 <.literal16+0xa39>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 487d <.literal16+0x93d>
+ .byte 224,7 // loopne 524d <.literal16+0xa3d>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 4881 <.literal16+0x941>
+ .byte 224,7 // loopne 5251 <.literal16+0xa41>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 4885 <.literal16+0x945>
+ .byte 224,7 // loopne 5255 <.literal16+0xa45>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -28819,13 +31886,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 48e9 <.literal16+0x9a9>
+ .byte 224,7 // loopne 52b9 <.literal16+0xaa9>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 48ed <.literal16+0x9ad>
+ .byte 224,7 // loopne 52bd <.literal16+0xaad>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 48f1 <.literal16+0x9b1>
+ .byte 224,7 // loopne 52c1 <.literal16+0xab1>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 48f5 <.literal16+0x9b5>
+ .byte 224,7 // loopne 52c5 <.literal16+0xab5>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -28863,13 +31930,13 @@ BALIGN16
.byte 65,0,0 // add %al,(%r8)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 4986 <.literal16+0xa46>
+ .byte 124,66 // jl 5356 <.literal16+0xb46>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 498a <.literal16+0xa4a>
+ .byte 124,66 // jl 535a <.literal16+0xb4a>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 498e <.literal16+0xa4e>
+ .byte 124,66 // jl 535e <.literal16+0xb4e>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 4992 <.literal16+0xa52>
+ .byte 124,66 // jl 5362 <.literal16+0xb52>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 0,240 // add %dh,%al
@@ -28959,13 +32026,13 @@ 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 4a95 <.literal16+0xb55>
+ .byte 112,65 // jo 5465 <.literal16+0xc55>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4a99 <.literal16+0xb59>
+ .byte 112,65 // jo 5469 <.literal16+0xc59>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4a9d <.literal16+0xb5d>
+ .byte 112,65 // jo 546d <.literal16+0xc5d>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4aa1 <.literal16+0xb61>
+ .byte 112,65 // jo 5471 <.literal16+0xc61>
.byte 255,0 // incl (%rax)
.byte 0,0 // add %al,(%rax)
.byte 255,0 // incl (%rax)
@@ -28987,11 +32054,11 @@ 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 4aeb <.literal16+0xbab>
+ .byte 127,67 // jg 54bb <.literal16+0xcab>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 4aef <.literal16+0xbaf>
+ .byte 127,67 // jg 54bf <.literal16+0xcaf>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 4af3 <.literal16+0xbb3>
+ .byte 127,67 // jg 54c3 <.literal16+0xcb3>
.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)
@@ -29067,13 +32134,13 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 255 // (bad)
- .byte 127,71 // jg 4bdb <.literal16+0xc9b>
+ .byte 127,71 // jg 55ab <.literal16+0xd9b>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 4bdf <.literal16+0xc9f>
+ .byte 127,71 // jg 55af <.literal16+0xd9f>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 4be3 <.literal16+0xca3>
+ .byte 127,71 // jg 55b3 <.literal16+0xda3>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 4be7 <.literal16+0xca7>
+ .byte 127,71 // jg 55b7 <.literal16+0xda7>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
@@ -29226,11 +32293,11 @@ 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 4d42 <.literal16+0xe02>
+ .byte 62,114,28 // jb,pt 5712 <.literal16+0xf02>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4d46 <.literal16+0xe06>
+ .byte 62,114,28 // jb,pt 5716 <.literal16+0xf06>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4d4a <.literal16+0xe0a>
+ .byte 62,114,28 // jb,pt 571a <.literal16+0xf0a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -29274,7 +32341,7 @@ 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 # 3d63dbd5 <_sk_callback_sse2+0x3d639d79>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e5a5 <_sk_callback_sse2+0x3d639e73>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -29300,7 +32367,7 @@ 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 # 3d63dc15 <_sk_callback_sse2+0x3d639db9>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e5e5 <_sk_callback_sse2+0x3d639eb3>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
@@ -29309,13 +32376,13 @@ BALIGN16
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
.byte 63 // (bad)
- .byte 114,28 // jb 4e0e <.literal16+0xece>
+ .byte 114,28 // jb 57de <.literal16+0xfce>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4e12 <.literal16+0xed2>
+ .byte 62,114,28 // jb,pt 57e2 <.literal16+0xfd2>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4e16 <.literal16+0xed6>
+ .byte 62,114,28 // jb,pt 57e6 <.literal16+0xfd6>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4e1a <.literal16+0xeda>
+ .byte 62,114,28 // jb,pt 57ea <.literal16+0xfda>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -29336,11 +32403,11 @@ 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 4e52 <.literal16+0xf12>
+ .byte 62,114,28 // jb,pt 5822 <.literal16+0x1012>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4e56 <.literal16+0xf16>
+ .byte 62,114,28 // jb,pt 5826 <.literal16+0x1016>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4e5a <.literal16+0xf1a>
+ .byte 62,114,28 // jb,pt 582a <.literal16+0x101a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -29384,7 +32451,7 @@ 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 # 3d63dce5 <_sk_callback_sse2+0x3d639e89>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e6b5 <_sk_callback_sse2+0x3d639f83>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -29410,7 +32477,7 @@ 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 # 3d63dd25 <_sk_callback_sse2+0x3d639ec9>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63e6f5 <_sk_callback_sse2+0x3d639fc3>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
@@ -29419,13 +32486,13 @@ BALIGN16
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
.byte 63 // (bad)
- .byte 114,28 // jb 4f1e <.literal16+0xfde>
+ .byte 114,28 // jb 58ee <.literal16+0x10de>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4f22 <_sk_callback_sse2+0x10c6>
+ .byte 62,114,28 // jb,pt 58f2 <_sk_callback_sse2+0x11c0>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4f26 <_sk_callback_sse2+0x10ca>
+ .byte 62,114,28 // jb,pt 58f6 <_sk_callback_sse2+0x11c4>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4f2a <_sk_callback_sse2+0x10ce>
+ .byte 62,114,28 // jb,pt 58fa <_sk_callback_sse2+0x11c8>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index 74fcd7bb56..2aaa5c6f90 100644
--- a/src/jumper/SkJumper_generated_win.S
+++ b/src/jumper/SkJumper_generated_win.S
@@ -106,14 +106,14 @@ _sk_seed_shader_hsw LABEL PROC
DB 197,249,110,199 ; vmovd %edi,%xmm0
DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,174,61,0,0 ; vbroadcastss 0x3dae(%rip),%ymm1 # 3f08 <_sk_callback_hsw+0x11b>
+ DB 196,226,125,24,13,46,68,0,0 ; vbroadcastss 0x442e(%rip),%ymm1 # 4588 <_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 196,226,125,24,21,146,61,0,0 ; vbroadcastss 0x3d92(%rip),%ymm2 # 3f0c <_sk_callback_hsw+0x11f>
+ DB 196,226,125,24,21,18,68,0,0 ; vbroadcastss 0x4412(%rip),%ymm2 # 458c <_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
@@ -132,13 +132,13 @@ _sk_dither_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 196,66,125,88,8 ; vpbroadcastd (%r8),%ymm9
DB 196,65,61,239,201 ; vpxor %ymm9,%ymm8,%ymm9
- DB 196,98,125,88,21,81,61,0,0 ; vpbroadcastd 0x3d51(%rip),%ymm10 # 3f10 <_sk_callback_hsw+0x123>
+ DB 196,98,125,88,21,209,67,0,0 ; vpbroadcastd 0x43d1(%rip),%ymm10 # 4590 <_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 196,98,125,88,37,54,61,0,0 ; vpbroadcastd 0x3d36(%rip),%ymm12 # 3f14 <_sk_callback_hsw+0x127>
- DB 196,98,125,88,45,49,61,0,0 ; vpbroadcastd 0x3d31(%rip),%ymm13 # 3f18 <_sk_callback_hsw+0x12b>
+ DB 196,98,125,88,37,182,67,0,0 ; vpbroadcastd 0x43b6(%rip),%ymm12 # 4594 <_sk_callback_hsw+0x127>
+ DB 196,98,125,88,45,177,67,0,0 ; vpbroadcastd 0x43b1(%rip),%ymm13 # 4598 <_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
@@ -153,8 +153,8 @@ _sk_dither_hsw LABEL PROC
DB 196,65,61,235,194 ; vpor %ymm10,%ymm8,%ymm8
DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,227,60,0,0 ; vbroadcastss 0x3ce3(%rip),%ymm9 # 3f1c <_sk_callback_hsw+0x12f>
- DB 196,98,125,24,21,222,60,0,0 ; vbroadcastss 0x3cde(%rip),%ymm10 # 3f20 <_sk_callback_hsw+0x133>
+ DB 196,98,125,24,13,99,67,0,0 ; vbroadcastss 0x4363(%rip),%ymm9 # 459c <_sk_callback_hsw+0x12f>
+ DB 196,98,125,24,21,94,67,0,0 ; vbroadcastss 0x435e(%rip),%ymm10 # 45a0 <_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
@@ -186,7 +186,7 @@ _sk_clear_hsw LABEL PROC
PUBLIC _sk_srcatop_hsw
_sk_srcatop_hsw LABEL PROC
DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0
- DB 196,98,125,24,5,132,60,0,0 ; vbroadcastss 0x3c84(%rip),%ymm8 # 3f24 <_sk_callback_hsw+0x137>
+ DB 196,98,125,24,5,4,67,0,0 ; vbroadcastss 0x4304(%rip),%ymm8 # 45a4 <_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
@@ -200,7 +200,7 @@ _sk_srcatop_hsw LABEL PROC
PUBLIC _sk_dstatop_hsw
_sk_dstatop_hsw LABEL PROC
- DB 196,98,125,24,5,87,60,0,0 ; vbroadcastss 0x3c57(%rip),%ymm8 # 3f28 <_sk_callback_hsw+0x13b>
+ DB 196,98,125,24,5,215,66,0,0 ; vbroadcastss 0x42d7(%rip),%ymm8 # 45a8 <_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
@@ -233,7 +233,7 @@ _sk_dstin_hsw LABEL PROC
PUBLIC _sk_srcout_hsw
_sk_srcout_hsw LABEL PROC
- DB 196,98,125,24,5,254,59,0,0 ; vbroadcastss 0x3bfe(%rip),%ymm8 # 3f2c <_sk_callback_hsw+0x13f>
+ DB 196,98,125,24,5,126,66,0,0 ; vbroadcastss 0x427e(%rip),%ymm8 # 45ac <_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
@@ -244,7 +244,7 @@ _sk_srcout_hsw LABEL PROC
PUBLIC _sk_dstout_hsw
_sk_dstout_hsw LABEL PROC
- DB 196,226,125,24,5,225,59,0,0 ; vbroadcastss 0x3be1(%rip),%ymm0 # 3f30 <_sk_callback_hsw+0x143>
+ DB 196,226,125,24,5,97,66,0,0 ; vbroadcastss 0x4261(%rip),%ymm0 # 45b0 <_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
@@ -255,7 +255,7 @@ _sk_dstout_hsw LABEL PROC
PUBLIC _sk_srcover_hsw
_sk_srcover_hsw LABEL PROC
- DB 196,98,125,24,5,196,59,0,0 ; vbroadcastss 0x3bc4(%rip),%ymm8 # 3f34 <_sk_callback_hsw+0x147>
+ DB 196,98,125,24,5,68,66,0,0 ; vbroadcastss 0x4244(%rip),%ymm8 # 45b4 <_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
@@ -266,7 +266,7 @@ _sk_srcover_hsw LABEL PROC
PUBLIC _sk_dstover_hsw
_sk_dstover_hsw LABEL PROC
- DB 196,98,125,24,5,163,59,0,0 ; vbroadcastss 0x3ba3(%rip),%ymm8 # 3f38 <_sk_callback_hsw+0x14b>
+ DB 196,98,125,24,5,35,66,0,0 ; vbroadcastss 0x4223(%rip),%ymm8 # 45b8 <_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
@@ -286,7 +286,7 @@ _sk_modulate_hsw LABEL PROC
PUBLIC _sk_multiply_hsw
_sk_multiply_hsw LABEL PROC
- DB 196,98,125,24,5,110,59,0,0 ; vbroadcastss 0x3b6e(%rip),%ymm8 # 3f3c <_sk_callback_hsw+0x14f>
+ DB 196,98,125,24,5,238,65,0,0 ; vbroadcastss 0x41ee(%rip),%ymm8 # 45bc <_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
@@ -328,7 +328,7 @@ _sk_screen_hsw LABEL PROC
PUBLIC _sk_xor__hsw
_sk_xor__hsw LABEL PROC
- DB 196,98,125,24,5,233,58,0,0 ; vbroadcastss 0x3ae9(%rip),%ymm8 # 3f40 <_sk_callback_hsw+0x153>
+ DB 196,98,125,24,5,105,65,0,0 ; vbroadcastss 0x4169(%rip),%ymm8 # 45c0 <_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
@@ -360,7 +360,7 @@ _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 196,98,125,24,5,113,58,0,0 ; vbroadcastss 0x3a71(%rip),%ymm8 # 3f44 <_sk_callback_hsw+0x157>
+ DB 196,98,125,24,5,241,64,0,0 ; vbroadcastss 0x40f1(%rip),%ymm8 # 45c4 <_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
@@ -383,7 +383,7 @@ _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 196,98,125,24,5,32,58,0,0 ; vbroadcastss 0x3a20(%rip),%ymm8 # 3f48 <_sk_callback_hsw+0x15b>
+ DB 196,98,125,24,5,160,64,0,0 ; vbroadcastss 0x40a0(%rip),%ymm8 # 45c8 <_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
@@ -409,7 +409,7 @@ _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 196,98,125,24,5,195,57,0,0 ; vbroadcastss 0x39c3(%rip),%ymm8 # 3f4c <_sk_callback_hsw+0x15f>
+ DB 196,98,125,24,5,67,64,0,0 ; vbroadcastss 0x4043(%rip),%ymm8 # 45cc <_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
@@ -429,7 +429,7 @@ _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 196,98,125,24,5,129,57,0,0 ; vbroadcastss 0x3981(%rip),%ymm8 # 3f50 <_sk_callback_hsw+0x163>
+ DB 196,98,125,24,5,1,64,0,0 ; vbroadcastss 0x4001(%rip),%ymm8 # 45d0 <_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
@@ -437,7 +437,7 @@ _sk_exclusion_hsw LABEL PROC
PUBLIC _sk_colorburn_hsw
_sk_colorburn_hsw LABEL PROC
- DB 196,98,125,24,5,111,57,0,0 ; vbroadcastss 0x396f(%rip),%ymm8 # 3f54 <_sk_callback_hsw+0x167>
+ DB 196,98,125,24,5,239,63,0,0 ; vbroadcastss 0x3fef(%rip),%ymm8 # 45d4 <_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
@@ -493,7 +493,7 @@ _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 196,98,125,24,13,122,56,0,0 ; vbroadcastss 0x387a(%rip),%ymm9 # 3f58 <_sk_callback_hsw+0x16b>
+ DB 196,98,125,24,13,250,62,0,0 ; vbroadcastss 0x3efa(%rip),%ymm9 # 45d8 <_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
@@ -544,7 +544,7 @@ _sk_colordodge_hsw LABEL PROC
PUBLIC _sk_hardlight_hsw
_sk_hardlight_hsw LABEL PROC
- DB 196,98,125,24,5,155,55,0,0 ; vbroadcastss 0x379b(%rip),%ymm8 # 3f5c <_sk_callback_hsw+0x16f>
+ DB 196,98,125,24,5,27,62,0,0 ; vbroadcastss 0x3e1b(%rip),%ymm8 # 45dc <_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
@@ -593,7 +593,7 @@ _sk_hardlight_hsw LABEL PROC
PUBLIC _sk_overlay_hsw
_sk_overlay_hsw LABEL PROC
- DB 196,98,125,24,5,211,54,0,0 ; vbroadcastss 0x36d3(%rip),%ymm8 # 3f60 <_sk_callback_hsw+0x173>
+ DB 196,98,125,24,5,83,61,0,0 ; vbroadcastss 0x3d53(%rip),%ymm8 # 45e0 <_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
@@ -653,10 +653,10 @@ _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 196,98,125,24,29,218,53,0,0 ; vbroadcastss 0x35da(%rip),%ymm11 # 3f68 <_sk_callback_hsw+0x17b>
+ DB 196,98,125,24,29,90,60,0,0 ; vbroadcastss 0x3c5a(%rip),%ymm11 # 45e8 <_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 196,98,125,24,37,203,53,0,0 ; vbroadcastss 0x35cb(%rip),%ymm12 # 3f6c <_sk_callback_hsw+0x17f>
+ DB 196,98,125,24,37,75,60,0,0 ; vbroadcastss 0x3c4b(%rip),%ymm12 # 45ec <_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
@@ -666,7 +666,7 @@ _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 196,98,125,24,5,142,53,0,0 ; vbroadcastss 0x358e(%rip),%ymm8 # 3f64 <_sk_callback_hsw+0x177>
+ DB 196,98,125,24,5,14,60,0,0 ; vbroadcastss 0x3c0e(%rip),%ymm8 # 45e4 <_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
@@ -744,6 +744,356 @@ _sk_softlight_hsw LABEL PROC
DB 72,131,196,88 ; add $0x58,%rsp
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_hue_hsw
+_sk_hue_hsw LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,100,194,208,0 ; vcmpeqps %ymm8,%ymm3,%ymm10
+ DB 196,98,125,24,13,163,58,0,0 ; vbroadcastss 0x3aa3(%rip),%ymm9 # 45f0 <_sk_callback_hsw+0x183>
+ DB 197,52,94,219 ; vdivps %ymm3,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
+ DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
+ DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
+ DB 197,84,95,214 ; vmaxps %ymm6,%ymm5,%ymm10
+ DB 196,65,92,95,210 ; vmaxps %ymm10,%ymm4,%ymm10
+ DB 197,84,93,222 ; vminps %ymm6,%ymm5,%ymm11
+ DB 196,65,92,93,219 ; vminps %ymm11,%ymm4,%ymm11
+ DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10
+ DB 197,116,93,218 ; vminps %ymm2,%ymm1,%ymm11
+ DB 196,65,124,93,219 ; vminps %ymm11,%ymm0,%ymm11
+ DB 197,116,95,226 ; vmaxps %ymm2,%ymm1,%ymm12
+ DB 196,65,124,95,228 ; vmaxps %ymm12,%ymm0,%ymm12
+ DB 196,65,28,92,227 ; vsubps %ymm11,%ymm12,%ymm12
+ DB 196,65,28,194,232,0 ; vcmpeqps %ymm8,%ymm12,%ymm13
+ DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0
+ DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
+ DB 196,193,124,94,196 ; vdivps %ymm12,%ymm0,%ymm0
+ DB 196,195,125,74,192,208 ; vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1
+ DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
+ DB 196,193,116,94,204 ; vdivps %ymm12,%ymm1,%ymm1
+ DB 196,195,117,74,200,208 ; vblendvps %ymm13,%ymm8,%ymm1,%ymm1
+ DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2
+ DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
+ DB 196,193,108,94,212 ; vdivps %ymm12,%ymm2,%ymm2
+ DB 196,195,109,74,208,208 ; vblendvps %ymm13,%ymm8,%ymm2,%ymm2
+ DB 196,98,125,24,21,24,58,0,0 ; vbroadcastss 0x3a18(%rip),%ymm10 # 45f4 <_sk_callback_hsw+0x187>
+ DB 196,98,125,24,29,19,58,0,0 ; vbroadcastss 0x3a13(%rip),%ymm11 # 45f8 <_sk_callback_hsw+0x18b>
+ DB 196,65,84,89,227 ; vmulps %ymm11,%ymm5,%ymm12
+ DB 196,66,93,184,226 ; vfmadd231ps %ymm10,%ymm4,%ymm12
+ DB 196,98,125,24,45,4,58,0,0 ; vbroadcastss 0x3a04(%rip),%ymm13 # 45fc <_sk_callback_hsw+0x18f>
+ DB 196,66,77,184,229 ; vfmadd231ps %ymm13,%ymm6,%ymm12
+ DB 196,65,116,89,243 ; vmulps %ymm11,%ymm1,%ymm14
+ DB 196,66,125,184,242 ; vfmadd231ps %ymm10,%ymm0,%ymm14
+ DB 196,66,109,184,245 ; vfmadd231ps %ymm13,%ymm2,%ymm14
+ DB 196,65,28,92,230 ; vsubps %ymm14,%ymm12,%ymm12
+ DB 196,65,124,88,244 ; vaddps %ymm12,%ymm0,%ymm14
+ DB 196,65,116,88,252 ; vaddps %ymm12,%ymm1,%ymm15
+ DB 196,193,108,88,196 ; vaddps %ymm12,%ymm2,%ymm0
+ DB 197,132,93,200 ; vminps %ymm0,%ymm15,%ymm1
+ DB 197,140,93,209 ; vminps %ymm1,%ymm14,%ymm2
+ DB 197,132,95,200 ; vmaxps %ymm0,%ymm15,%ymm1
+ DB 197,12,95,225 ; vmaxps %ymm1,%ymm14,%ymm12
+ DB 196,193,4,89,203 ; vmulps %ymm11,%ymm15,%ymm1
+ DB 196,194,13,184,202 ; vfmadd231ps %ymm10,%ymm14,%ymm1
+ DB 196,194,125,184,205 ; vfmadd231ps %ymm13,%ymm0,%ymm1
+ DB 197,60,194,210,2 ; vcmpleps %ymm2,%ymm8,%ymm10
+ DB 197,12,92,217 ; vsubps %ymm1,%ymm14,%ymm11
+ DB 196,65,116,89,219 ; vmulps %ymm11,%ymm1,%ymm11
+ DB 197,244,92,210 ; vsubps %ymm2,%ymm1,%ymm2
+ DB 197,36,94,218 ; vdivps %ymm2,%ymm11,%ymm11
+ DB 196,65,116,88,219 ; vaddps %ymm11,%ymm1,%ymm11
+ DB 196,67,37,74,222,160 ; vblendvps %ymm10,%ymm14,%ymm11,%ymm11
+ DB 196,65,52,194,236,1 ; vcmpltps %ymm12,%ymm9,%ymm13
+ DB 197,36,92,241 ; vsubps %ymm1,%ymm11,%ymm14
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,246 ; vmulps %ymm14,%ymm9,%ymm14
+ DB 197,28,92,225 ; vsubps %ymm1,%ymm12,%ymm12
+ DB 196,65,12,94,244 ; vdivps %ymm12,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,37,74,222,208 ; vblendvps %ymm13,%ymm14,%ymm11,%ymm11
+ DB 196,65,36,95,216 ; vmaxps %ymm8,%ymm11,%ymm11
+ DB 197,4,92,241 ; vsubps %ymm1,%ymm15,%ymm14
+ DB 196,65,116,89,246 ; vmulps %ymm14,%ymm1,%ymm14
+ DB 197,12,94,242 ; vdivps %ymm2,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,13,74,247,160 ; vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ DB 197,12,92,249 ; vsubps %ymm1,%ymm14,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,252 ; vdivps %ymm12,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,13,74,247,208 ; vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ DB 196,65,12,95,240 ; vmaxps %ymm8,%ymm14,%ymm14
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 197,132,94,210 ; vdivps %ymm2,%ymm15,%ymm2
+ DB 197,244,88,210 ; vaddps %ymm2,%ymm1,%ymm2
+ DB 196,227,109,74,192,160 ; vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,94,212 ; vdivps %ymm12,%ymm2,%ymm2
+ DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1
+ DB 196,227,125,74,193,208 ; vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ DB 196,193,124,95,208 ; vmaxps %ymm8,%ymm0,%ymm2
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 196,226,69,172,216 ; vfnmadd213ps %ymm0,%ymm7,%ymm3
+ DB 196,193,100,89,195 ; vmulps %ymm11,%ymm3,%ymm0
+ DB 196,193,100,89,206 ; vmulps %ymm14,%ymm3,%ymm1
+ DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_saturation_hsw
+_sk_saturation_hsw LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,68,194,208,0 ; vcmpeqps %ymm8,%ymm7,%ymm10
+ DB 196,98,125,24,13,220,56,0,0 ; vbroadcastss 0x38dc(%rip),%ymm9 # 4600 <_sk_callback_hsw+0x193>
+ DB 197,52,94,223 ; vdivps %ymm7,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,44,89,220 ; vmulps %ymm4,%ymm10,%ymm11
+ DB 197,44,89,229 ; vmulps %ymm5,%ymm10,%ymm12
+ DB 197,44,89,214 ; vmulps %ymm6,%ymm10,%ymm10
+ DB 197,116,95,234 ; vmaxps %ymm2,%ymm1,%ymm13
+ DB 196,65,124,95,237 ; vmaxps %ymm13,%ymm0,%ymm13
+ DB 197,244,93,202 ; vminps %ymm2,%ymm1,%ymm1
+ DB 197,252,93,193 ; vminps %ymm1,%ymm0,%ymm0
+ DB 197,148,92,192 ; vsubps %ymm0,%ymm13,%ymm0
+ DB 196,193,28,93,202 ; vminps %ymm10,%ymm12,%ymm1
+ DB 197,164,93,201 ; vminps %ymm1,%ymm11,%ymm1
+ DB 196,193,28,95,210 ; vmaxps %ymm10,%ymm12,%ymm2
+ DB 197,164,95,210 ; vmaxps %ymm2,%ymm11,%ymm2
+ DB 197,236,92,209 ; vsubps %ymm1,%ymm2,%ymm2
+ DB 196,65,108,194,232,0 ; vcmpeqps %ymm8,%ymm2,%ymm13
+ DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11
+ DB 196,65,124,89,219 ; vmulps %ymm11,%ymm0,%ymm11
+ DB 197,36,94,218 ; vdivps %ymm2,%ymm11,%ymm11
+ DB 196,67,37,74,216,208 ; vblendvps %ymm13,%ymm8,%ymm11,%ymm11
+ DB 197,28,92,225 ; vsubps %ymm1,%ymm12,%ymm12
+ DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12
+ DB 197,28,94,226 ; vdivps %ymm2,%ymm12,%ymm12
+ DB 196,67,29,74,224,208 ; vblendvps %ymm13,%ymm8,%ymm12,%ymm12
+ DB 197,172,92,201 ; vsubps %ymm1,%ymm10,%ymm1
+ DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
+ DB 197,252,94,194 ; vdivps %ymm2,%ymm0,%ymm0
+ DB 196,195,125,74,192,208 ; vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ DB 196,226,125,24,21,88,56,0,0 ; vbroadcastss 0x3858(%rip),%ymm2 # 4604 <_sk_callback_hsw+0x197>
+ DB 196,226,125,24,13,83,56,0,0 ; vbroadcastss 0x3853(%rip),%ymm1 # 4608 <_sk_callback_hsw+0x19b>
+ DB 197,84,89,209 ; vmulps %ymm1,%ymm5,%ymm10
+ DB 196,98,93,184,210 ; vfmadd231ps %ymm2,%ymm4,%ymm10
+ DB 196,98,125,24,45,69,56,0,0 ; vbroadcastss 0x3845(%rip),%ymm13 # 460c <_sk_callback_hsw+0x19f>
+ DB 196,66,77,184,213 ; vfmadd231ps %ymm13,%ymm6,%ymm10
+ DB 197,28,89,241 ; vmulps %ymm1,%ymm12,%ymm14
+ DB 196,98,37,184,242 ; vfmadd231ps %ymm2,%ymm11,%ymm14
+ DB 196,66,125,184,245 ; vfmadd231ps %ymm13,%ymm0,%ymm14
+ DB 196,65,44,92,214 ; vsubps %ymm14,%ymm10,%ymm10
+ DB 196,65,36,88,218 ; vaddps %ymm10,%ymm11,%ymm11
+ DB 196,65,28,88,226 ; vaddps %ymm10,%ymm12,%ymm12
+ DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
+ DB 197,28,93,208 ; vminps %ymm0,%ymm12,%ymm10
+ DB 196,65,36,93,210 ; vminps %ymm10,%ymm11,%ymm10
+ DB 197,28,95,240 ; vmaxps %ymm0,%ymm12,%ymm14
+ DB 196,65,36,95,246 ; vmaxps %ymm14,%ymm11,%ymm14
+ DB 197,156,89,201 ; vmulps %ymm1,%ymm12,%ymm1
+ DB 196,226,37,184,202 ; vfmadd231ps %ymm2,%ymm11,%ymm1
+ DB 196,194,125,184,205 ; vfmadd231ps %ymm13,%ymm0,%ymm1
+ DB 196,193,60,194,210,2 ; vcmpleps %ymm10,%ymm8,%ymm2
+ DB 197,36,92,233 ; vsubps %ymm1,%ymm11,%ymm13
+ DB 196,65,116,89,237 ; vmulps %ymm13,%ymm1,%ymm13
+ DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
+ DB 196,65,20,94,234 ; vdivps %ymm10,%ymm13,%ymm13
+ DB 196,65,116,88,237 ; vaddps %ymm13,%ymm1,%ymm13
+ DB 196,67,21,74,219,32 ; vblendvps %ymm2,%ymm11,%ymm13,%ymm11
+ DB 196,65,52,194,238,1 ; vcmpltps %ymm14,%ymm9,%ymm13
+ DB 197,36,92,249 ; vsubps %ymm1,%ymm11,%ymm15
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 197,12,92,241 ; vsubps %ymm1,%ymm14,%ymm14
+ DB 196,65,4,94,254 ; vdivps %ymm14,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,37,74,223,208 ; vblendvps %ymm13,%ymm15,%ymm11,%ymm11
+ DB 196,65,36,95,216 ; vmaxps %ymm8,%ymm11,%ymm11
+ DB 197,28,92,249 ; vsubps %ymm1,%ymm12,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 196,65,4,94,250 ; vdivps %ymm10,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,5,74,228,32 ; vblendvps %ymm2,%ymm12,%ymm15,%ymm12
+ DB 197,28,92,249 ; vsubps %ymm1,%ymm12,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,254 ; vdivps %ymm14,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,29,74,231,208 ; vblendvps %ymm13,%ymm15,%ymm12,%ymm12
+ DB 196,65,28,95,224 ; vmaxps %ymm8,%ymm12,%ymm12
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 196,65,4,94,210 ; vdivps %ymm10,%ymm15,%ymm10
+ DB 196,65,116,88,210 ; vaddps %ymm10,%ymm1,%ymm10
+ DB 196,227,45,74,192,32 ; vblendvps %ymm2,%ymm0,%ymm10,%ymm0
+ DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,94,214 ; vdivps %ymm14,%ymm2,%ymm2
+ DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1
+ DB 196,227,125,74,193,208 ; vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ DB 196,193,124,95,208 ; vmaxps %ymm8,%ymm0,%ymm2
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 196,226,69,172,216 ; vfnmadd213ps %ymm0,%ymm7,%ymm3
+ DB 196,193,100,89,195 ; vmulps %ymm11,%ymm3,%ymm0
+ DB 196,193,100,89,204 ; vmulps %ymm12,%ymm3,%ymm1
+ DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_color_hsw
+_sk_color_hsw LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,100,194,208,0 ; vcmpeqps %ymm8,%ymm3,%ymm10
+ DB 196,98,125,24,13,23,55,0,0 ; vbroadcastss 0x3717(%rip),%ymm9 # 4610 <_sk_callback_hsw+0x1a3>
+ DB 197,52,94,219 ; vdivps %ymm3,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
+ DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
+ DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
+ DB 196,98,125,24,21,252,54,0,0 ; vbroadcastss 0x36fc(%rip),%ymm10 # 4614 <_sk_callback_hsw+0x1a7>
+ DB 196,98,125,24,29,247,54,0,0 ; vbroadcastss 0x36f7(%rip),%ymm11 # 4618 <_sk_callback_hsw+0x1ab>
+ DB 196,65,84,89,227 ; vmulps %ymm11,%ymm5,%ymm12
+ DB 196,66,93,184,226 ; vfmadd231ps %ymm10,%ymm4,%ymm12
+ DB 196,98,125,24,45,232,54,0,0 ; vbroadcastss 0x36e8(%rip),%ymm13 # 461c <_sk_callback_hsw+0x1af>
+ DB 196,66,77,184,229 ; vfmadd231ps %ymm13,%ymm6,%ymm12
+ DB 196,65,116,89,243 ; vmulps %ymm11,%ymm1,%ymm14
+ DB 196,66,125,184,242 ; vfmadd231ps %ymm10,%ymm0,%ymm14
+ DB 196,66,109,184,245 ; vfmadd231ps %ymm13,%ymm2,%ymm14
+ DB 196,65,28,92,230 ; vsubps %ymm14,%ymm12,%ymm12
+ DB 196,65,124,88,244 ; vaddps %ymm12,%ymm0,%ymm14
+ DB 196,65,116,88,252 ; vaddps %ymm12,%ymm1,%ymm15
+ DB 196,193,108,88,196 ; vaddps %ymm12,%ymm2,%ymm0
+ DB 197,132,93,200 ; vminps %ymm0,%ymm15,%ymm1
+ DB 197,140,93,209 ; vminps %ymm1,%ymm14,%ymm2
+ DB 197,132,95,200 ; vmaxps %ymm0,%ymm15,%ymm1
+ DB 197,12,95,225 ; vmaxps %ymm1,%ymm14,%ymm12
+ DB 196,193,4,89,203 ; vmulps %ymm11,%ymm15,%ymm1
+ DB 196,194,13,184,202 ; vfmadd231ps %ymm10,%ymm14,%ymm1
+ DB 196,194,125,184,205 ; vfmadd231ps %ymm13,%ymm0,%ymm1
+ DB 197,60,194,210,2 ; vcmpleps %ymm2,%ymm8,%ymm10
+ DB 197,12,92,217 ; vsubps %ymm1,%ymm14,%ymm11
+ DB 196,65,116,89,219 ; vmulps %ymm11,%ymm1,%ymm11
+ DB 197,244,92,210 ; vsubps %ymm2,%ymm1,%ymm2
+ DB 197,36,94,218 ; vdivps %ymm2,%ymm11,%ymm11
+ DB 196,65,116,88,219 ; vaddps %ymm11,%ymm1,%ymm11
+ DB 196,67,37,74,222,160 ; vblendvps %ymm10,%ymm14,%ymm11,%ymm11
+ DB 196,65,52,194,236,1 ; vcmpltps %ymm12,%ymm9,%ymm13
+ DB 197,36,92,241 ; vsubps %ymm1,%ymm11,%ymm14
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,246 ; vmulps %ymm14,%ymm9,%ymm14
+ DB 197,28,92,225 ; vsubps %ymm1,%ymm12,%ymm12
+ DB 196,65,12,94,244 ; vdivps %ymm12,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,37,74,222,208 ; vblendvps %ymm13,%ymm14,%ymm11,%ymm11
+ DB 196,65,36,95,216 ; vmaxps %ymm8,%ymm11,%ymm11
+ DB 197,4,92,241 ; vsubps %ymm1,%ymm15,%ymm14
+ DB 196,65,116,89,246 ; vmulps %ymm14,%ymm1,%ymm14
+ DB 197,12,94,242 ; vdivps %ymm2,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,13,74,247,160 ; vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ DB 197,12,92,249 ; vsubps %ymm1,%ymm14,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,252 ; vdivps %ymm12,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,13,74,247,208 ; vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ DB 196,65,12,95,240 ; vmaxps %ymm8,%ymm14,%ymm14
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 197,132,94,210 ; vdivps %ymm2,%ymm15,%ymm2
+ DB 197,244,88,210 ; vaddps %ymm2,%ymm1,%ymm2
+ DB 196,227,109,74,192,160 ; vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,94,212 ; vdivps %ymm12,%ymm2,%ymm2
+ DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1
+ DB 196,227,125,74,193,208 ; vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ DB 196,193,124,95,208 ; vmaxps %ymm8,%ymm0,%ymm2
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 196,226,69,172,216 ; vfnmadd213ps %ymm0,%ymm7,%ymm3
+ DB 196,193,100,89,195 ; vmulps %ymm11,%ymm3,%ymm0
+ DB 196,193,100,89,206 ; vmulps %ymm14,%ymm3,%ymm1
+ DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_luminosity_hsw
+_sk_luminosity_hsw LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,68,194,208,0 ; vcmpeqps %ymm8,%ymm7,%ymm10
+ DB 196,98,125,24,13,192,53,0,0 ; vbroadcastss 0x35c0(%rip),%ymm9 # 4620 <_sk_callback_hsw+0x1b3>
+ DB 197,52,94,223 ; vdivps %ymm7,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,44,89,220 ; vmulps %ymm4,%ymm10,%ymm11
+ DB 197,44,89,229 ; vmulps %ymm5,%ymm10,%ymm12
+ DB 197,44,89,214 ; vmulps %ymm6,%ymm10,%ymm10
+ DB 196,98,125,24,45,165,53,0,0 ; vbroadcastss 0x35a5(%rip),%ymm13 # 4624 <_sk_callback_hsw+0x1b7>
+ DB 196,98,125,24,53,160,53,0,0 ; vbroadcastss 0x35a0(%rip),%ymm14 # 4628 <_sk_callback_hsw+0x1bb>
+ DB 196,193,116,89,206 ; vmulps %ymm14,%ymm1,%ymm1
+ DB 196,226,21,168,193 ; vfmadd213ps %ymm1,%ymm13,%ymm0
+ DB 196,98,125,24,61,145,53,0,0 ; vbroadcastss 0x3591(%rip),%ymm15 # 462c <_sk_callback_hsw+0x1bf>
+ DB 196,226,5,168,208 ; vfmadd213ps %ymm0,%ymm15,%ymm2
+ DB 196,193,28,89,198 ; vmulps %ymm14,%ymm12,%ymm0
+ DB 196,194,37,184,197 ; vfmadd231ps %ymm13,%ymm11,%ymm0
+ DB 196,194,45,184,199 ; vfmadd231ps %ymm15,%ymm10,%ymm0
+ DB 197,236,92,192 ; vsubps %ymm0,%ymm2,%ymm0
+ DB 197,164,88,208 ; vaddps %ymm0,%ymm11,%ymm2
+ DB 197,28,88,216 ; vaddps %ymm0,%ymm12,%ymm11
+ DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0
+ DB 197,164,93,200 ; vminps %ymm0,%ymm11,%ymm1
+ DB 197,108,93,209 ; vminps %ymm1,%ymm2,%ymm10
+ DB 197,164,95,200 ; vmaxps %ymm0,%ymm11,%ymm1
+ DB 197,108,95,225 ; vmaxps %ymm1,%ymm2,%ymm12
+ DB 196,193,36,89,206 ; vmulps %ymm14,%ymm11,%ymm1
+ DB 196,194,109,184,205 ; vfmadd231ps %ymm13,%ymm2,%ymm1
+ DB 196,194,125,184,207 ; vfmadd231ps %ymm15,%ymm0,%ymm1
+ DB 196,65,60,194,234,2 ; vcmpleps %ymm10,%ymm8,%ymm13
+ DB 197,108,92,241 ; vsubps %ymm1,%ymm2,%ymm14
+ DB 196,65,116,89,246 ; vmulps %ymm14,%ymm1,%ymm14
+ DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
+ DB 196,65,12,94,242 ; vdivps %ymm10,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2
+ DB 196,65,52,194,244,1 ; vcmpltps %ymm12,%ymm9,%ymm14
+ DB 197,108,92,249 ; vsubps %ymm1,%ymm2,%ymm15
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 197,28,92,225 ; vsubps %ymm1,%ymm12,%ymm12
+ DB 196,65,4,94,252 ; vdivps %ymm12,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,195,109,74,215,224 ; vblendvps %ymm14,%ymm15,%ymm2,%ymm2
+ DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2
+ DB 197,36,92,249 ; vsubps %ymm1,%ymm11,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 196,65,4,94,250 ; vdivps %ymm10,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,5,74,219,208 ; vblendvps %ymm13,%ymm11,%ymm15,%ymm11
+ DB 197,36,92,249 ; vsubps %ymm1,%ymm11,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,252 ; vdivps %ymm12,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,37,74,223,224 ; vblendvps %ymm14,%ymm15,%ymm11,%ymm11
+ DB 196,65,36,95,216 ; vmaxps %ymm8,%ymm11,%ymm11
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 196,65,4,94,210 ; vdivps %ymm10,%ymm15,%ymm10
+ DB 196,65,116,88,210 ; vaddps %ymm10,%ymm1,%ymm10
+ DB 196,227,45,74,192,208 ; vblendvps %ymm13,%ymm0,%ymm10,%ymm0
+ DB 197,124,92,209 ; vsubps %ymm1,%ymm0,%ymm10
+ DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
+ DB 196,65,52,94,204 ; vdivps %ymm12,%ymm9,%ymm9
+ DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1
+ DB 196,227,125,74,193,224 ; vblendvps %ymm14,%ymm1,%ymm0,%ymm0
+ DB 196,65,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm8
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 196,226,69,172,216 ; vfnmadd213ps %ymm0,%ymm7,%ymm3
+ DB 197,228,89,194 ; vmulps %ymm2,%ymm3,%ymm0
+ DB 196,193,100,89,203 ; vmulps %ymm11,%ymm3,%ymm1
+ DB 196,193,100,89,208 ; vmulps %ymm8,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_clamp_0_hsw
_sk_clamp_0_hsw LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
@@ -756,7 +1106,7 @@ _sk_clamp_0_hsw LABEL PROC
PUBLIC _sk_clamp_1_hsw
_sk_clamp_1_hsw LABEL PROC
- DB 196,98,125,24,5,17,52,0,0 ; vbroadcastss 0x3411(%rip),%ymm8 # 3f70 <_sk_callback_hsw+0x183>
+ DB 196,98,125,24,5,83,52,0,0 ; vbroadcastss 0x3453(%rip),%ymm8 # 4630 <_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
@@ -766,7 +1116,7 @@ _sk_clamp_1_hsw LABEL PROC
PUBLIC _sk_clamp_a_hsw
_sk_clamp_a_hsw LABEL PROC
- DB 196,98,125,24,5,244,51,0,0 ; vbroadcastss 0x33f4(%rip),%ymm8 # 3f74 <_sk_callback_hsw+0x187>
+ DB 196,98,125,24,5,54,52,0,0 ; vbroadcastss 0x3436(%rip),%ymm8 # 4634 <_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
@@ -838,7 +1188,7 @@ 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 196,98,125,24,21,60,51,0,0 ; vbroadcastss 0x333c(%rip),%ymm10 # 3f78 <_sk_callback_hsw+0x18b>
+ DB 196,98,125,24,21,126,51,0,0 ; vbroadcastss 0x337e(%rip),%ymm10 # 4638 <_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
@@ -849,16 +1199,16 @@ _sk_unpremul_hsw LABEL PROC
PUBLIC _sk_from_srgb_hsw
_sk_from_srgb_hsw LABEL PROC
- DB 196,98,125,24,5,29,51,0,0 ; vbroadcastss 0x331d(%rip),%ymm8 # 3f7c <_sk_callback_hsw+0x18f>
+ DB 196,98,125,24,5,95,51,0,0 ; vbroadcastss 0x335f(%rip),%ymm8 # 463c <_sk_callback_hsw+0x1cf>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10
- DB 196,98,125,24,29,15,51,0,0 ; vbroadcastss 0x330f(%rip),%ymm11 # 3f80 <_sk_callback_hsw+0x193>
- DB 196,98,125,24,37,10,51,0,0 ; vbroadcastss 0x330a(%rip),%ymm12 # 3f84 <_sk_callback_hsw+0x197>
+ DB 196,98,125,24,29,81,51,0,0 ; vbroadcastss 0x3351(%rip),%ymm11 # 4640 <_sk_callback_hsw+0x1d3>
+ DB 196,98,125,24,37,76,51,0,0 ; vbroadcastss 0x334c(%rip),%ymm12 # 4644 <_sk_callback_hsw+0x1d7>
DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13
DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13
- DB 196,98,125,24,53,251,50,0,0 ; vbroadcastss 0x32fb(%rip),%ymm14 # 3f88 <_sk_callback_hsw+0x19b>
+ DB 196,98,125,24,53,61,51,0,0 ; vbroadcastss 0x333d(%rip),%ymm14 # 4648 <_sk_callback_hsw+0x1db>
DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13
- DB 196,98,125,24,21,241,50,0,0 ; vbroadcastss 0x32f1(%rip),%ymm10 # 3f8c <_sk_callback_hsw+0x19f>
+ DB 196,98,125,24,21,51,51,0,0 ; vbroadcastss 0x3333(%rip),%ymm10 # 464c <_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
@@ -882,16 +1232,16 @@ _sk_to_srgb_hsw LABEL PROC
DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8
DB 196,65,124,83,200 ; vrcpps %ymm8,%ymm9
DB 196,65,124,82,208 ; vrsqrtps %ymm8,%ymm10
- DB 196,98,125,24,5,139,50,0,0 ; vbroadcastss 0x328b(%rip),%ymm8 # 3f90 <_sk_callback_hsw+0x1a3>
+ DB 196,98,125,24,5,205,50,0,0 ; vbroadcastss 0x32cd(%rip),%ymm8 # 4650 <_sk_callback_hsw+0x1e3>
DB 196,65,124,89,216 ; vmulps %ymm8,%ymm0,%ymm11
- DB 196,98,125,24,37,129,50,0,0 ; vbroadcastss 0x3281(%rip),%ymm12 # 3f94 <_sk_callback_hsw+0x1a7>
- DB 196,98,125,24,45,124,50,0,0 ; vbroadcastss 0x327c(%rip),%ymm13 # 3f98 <_sk_callback_hsw+0x1ab>
+ DB 196,98,125,24,37,195,50,0,0 ; vbroadcastss 0x32c3(%rip),%ymm12 # 4654 <_sk_callback_hsw+0x1e7>
+ DB 196,98,125,24,45,190,50,0,0 ; vbroadcastss 0x32be(%rip),%ymm13 # 4658 <_sk_callback_hsw+0x1eb>
DB 196,66,21,168,204 ; vfmadd213ps %ymm12,%ymm13,%ymm9
- DB 196,98,125,24,53,114,50,0,0 ; vbroadcastss 0x3272(%rip),%ymm14 # 3f9c <_sk_callback_hsw+0x1af>
+ DB 196,98,125,24,53,180,50,0,0 ; vbroadcastss 0x32b4(%rip),%ymm14 # 465c <_sk_callback_hsw+0x1ef>
DB 196,66,13,184,202 ; vfmadd231ps %ymm10,%ymm14,%ymm9
- DB 196,98,125,24,21,104,50,0,0 ; vbroadcastss 0x3268(%rip),%ymm10 # 3fa0 <_sk_callback_hsw+0x1b3>
+ DB 196,98,125,24,21,170,50,0,0 ; vbroadcastss 0x32aa(%rip),%ymm10 # 4660 <_sk_callback_hsw+0x1f3>
DB 196,65,44,93,201 ; vminps %ymm9,%ymm10,%ymm9
- DB 196,98,125,24,61,94,50,0,0 ; vbroadcastss 0x325e(%rip),%ymm15 # 3fa4 <_sk_callback_hsw+0x1b7>
+ DB 196,98,125,24,61,160,50,0,0 ; vbroadcastss 0x32a0(%rip),%ymm15 # 4664 <_sk_callback_hsw+0x1f7>
DB 196,193,124,194,199,1 ; vcmpltps %ymm15,%ymm0,%ymm0
DB 196,195,53,74,195,0 ; vblendvps %ymm0,%ymm11,%ymm9,%ymm0
DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9
@@ -922,26 +1272,26 @@ _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 196,98,125,24,29,216,49,0,0 ; vbroadcastss 0x31d8(%rip),%ymm11 # 3fa8 <_sk_callback_hsw+0x1bb>
+ DB 196,98,125,24,29,26,50,0,0 ; vbroadcastss 0x321a(%rip),%ymm11 # 4668 <_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 196,98,125,24,53,197,49,0,0 ; vbroadcastss 0x31c5(%rip),%ymm14 # 3fac <_sk_callback_hsw+0x1bf>
+ DB 196,98,125,24,53,7,50,0,0 ; vbroadcastss 0x3207(%rip),%ymm14 # 466c <_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 196,98,125,24,53,172,49,0,0 ; vbroadcastss 0x31ac(%rip),%ymm14 # 3fb4 <_sk_callback_hsw+0x1c7>
+ DB 196,98,125,24,53,238,49,0,0 ; vbroadcastss 0x31ee(%rip),%ymm14 # 4674 <_sk_callback_hsw+0x207>
DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13
- DB 196,98,125,24,53,154,49,0,0 ; vbroadcastss 0x319a(%rip),%ymm14 # 3fb0 <_sk_callback_hsw+0x1c3>
+ DB 196,98,125,24,53,220,49,0,0 ; vbroadcastss 0x31dc(%rip),%ymm14 # 4670 <_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 196,98,125,24,29,125,49,0,0 ; vbroadcastss 0x317d(%rip),%ymm11 # 3fbc <_sk_callback_hsw+0x1cf>
+ DB 196,98,125,24,29,191,49,0,0 ; vbroadcastss 0x31bf(%rip),%ymm11 # 467c <_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
@@ -951,7 +1301,7 @@ _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 196,98,125,24,5,64,49,0,0 ; vbroadcastss 0x3140(%rip),%ymm8 # 3fb8 <_sk_callback_hsw+0x1cb>
+ DB 196,98,125,24,5,130,49,0,0 ; vbroadcastss 0x3182(%rip),%ymm8 # 4678 <_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
@@ -966,30 +1316,30 @@ _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 196,98,125,24,5,7,49,0,0 ; vbroadcastss 0x3107(%rip),%ymm8 # 3fc0 <_sk_callback_hsw+0x1d3>
+ DB 196,98,125,24,5,73,49,0,0 ; vbroadcastss 0x3149(%rip),%ymm8 # 4680 <_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 196,98,125,24,13,234,48,0,0 ; vbroadcastss 0x30ea(%rip),%ymm9 # 3fc4 <_sk_callback_hsw+0x1d7>
+ DB 196,98,125,24,13,44,49,0,0 ; vbroadcastss 0x312c(%rip),%ymm9 # 4684 <_sk_callback_hsw+0x217>
DB 196,66,109,170,202 ; vfmsub213ps %ymm10,%ymm2,%ymm9
- DB 196,98,125,24,29,224,48,0,0 ; vbroadcastss 0x30e0(%rip),%ymm11 # 3fc8 <_sk_callback_hsw+0x1db>
+ DB 196,98,125,24,29,34,49,0,0 ; vbroadcastss 0x3122(%rip),%ymm11 # 4688 <_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 196,98,125,24,45,202,48,0,0 ; vbroadcastss 0x30ca(%rip),%ymm13 # 3fd0 <_sk_callback_hsw+0x1e3>
+ DB 196,98,125,24,45,12,49,0,0 ; vbroadcastss 0x310c(%rip),%ymm13 # 4690 <_sk_callback_hsw+0x223>
DB 196,193,4,89,197 ; vmulps %ymm13,%ymm15,%ymm0
- DB 196,98,125,24,53,192,48,0,0 ; vbroadcastss 0x30c0(%rip),%ymm14 # 3fd4 <_sk_callback_hsw+0x1e7>
+ DB 196,98,125,24,53,2,49,0,0 ; vbroadcastss 0x3102(%rip),%ymm14 # 4694 <_sk_callback_hsw+0x227>
DB 197,12,92,224 ; vsubps %ymm0,%ymm14,%ymm12
DB 196,66,37,168,225 ; vfmadd213ps %ymm9,%ymm11,%ymm12
- DB 196,226,125,24,29,166,48,0,0 ; vbroadcastss 0x30a6(%rip),%ymm3 # 3fcc <_sk_callback_hsw+0x1df>
+ DB 196,226,125,24,29,232,48,0,0 ; vbroadcastss 0x30e8(%rip),%ymm3 # 468c <_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 196,98,125,24,37,145,48,0,0 ; vbroadcastss 0x3091(%rip),%ymm12 # 3fd8 <_sk_callback_hsw+0x1eb>
+ DB 196,98,125,24,37,211,48,0,0 ; vbroadcastss 0x30d3(%rip),%ymm12 # 4698 <_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
@@ -1005,7 +1355,7 @@ _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 196,226,125,24,5,61,48,0,0 ; vbroadcastss 0x303d(%rip),%ymm0 # 3fdc <_sk_callback_hsw+0x1ef>
+ DB 196,226,125,24,5,127,48,0,0 ; vbroadcastss 0x307f(%rip),%ymm0 # 469c <_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
@@ -1051,11 +1401,11 @@ _sk_scale_u8_hsw LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,51 ; jne 1085 <_sk_scale_u8_hsw+0x43>
+ DB 117,51 ; jne 1703 <_sk_scale_u8_hsw+0x43>
DB 197,122,126,0 ; vmovq (%rax),%xmm8
DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,119,47,0,0 ; vbroadcastss 0x2f77(%rip),%ymm9 # 3fe0 <_sk_callback_hsw+0x1f3>
+ DB 196,98,125,24,13,185,47,0,0 ; vbroadcastss 0x2fb9(%rip),%ymm9 # 46a0 <_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
@@ -1073,9 +1423,9 @@ _sk_scale_u8_hsw LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 108d <_sk_scale_u8_hsw+0x4b>
+ DB 117,234 ; jne 170b <_sk_scale_u8_hsw+0x4b>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 235,172 ; jmp 1056 <_sk_scale_u8_hsw+0x14>
+ DB 235,172 ; jmp 16d4 <_sk_scale_u8_hsw+0x14>
PUBLIC _sk_lerp_1_float_hsw
_sk_lerp_1_float_hsw LABEL PROC
@@ -1099,11 +1449,11 @@ _sk_lerp_u8_hsw LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,71 ; jne 1130 <_sk_lerp_u8_hsw+0x57>
+ DB 117,71 ; jne 17ae <_sk_lerp_u8_hsw+0x57>
DB 197,122,126,0 ; vmovq (%rax),%xmm8
DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,228,46,0,0 ; vbroadcastss 0x2ee4(%rip),%ymm9 # 3fe4 <_sk_callback_hsw+0x1f7>
+ DB 196,98,125,24,13,38,47,0,0 ; vbroadcastss 0x2f26(%rip),%ymm9 # 46a4 <_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
@@ -1125,32 +1475,32 @@ _sk_lerp_u8_hsw LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 1138 <_sk_lerp_u8_hsw+0x5f>
+ DB 117,234 ; jne 17b6 <_sk_lerp_u8_hsw+0x5f>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 235,152 ; jmp 10ed <_sk_lerp_u8_hsw+0x14>
+ DB 235,152 ; jmp 176b <_sk_lerp_u8_hsw+0x14>
PUBLIC _sk_lerp_565_hsw
_sk_lerp_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,149,0,0,0 ; jne 11f8 <_sk_lerp_565_hsw+0xa3>
+ DB 15,133,149,0,0,0 ; jne 1876 <_sk_lerp_565_hsw+0xa3>
DB 196,193,122,111,28,122 ; vmovdqu (%r10,%rdi,2),%xmm3
DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3
- DB 196,98,125,88,5,113,46,0,0 ; vpbroadcastd 0x2e71(%rip),%ymm8 # 3fe8 <_sk_callback_hsw+0x1fb>
+ DB 196,98,125,88,5,179,46,0,0 ; vpbroadcastd 0x2eb3(%rip),%ymm8 # 46a8 <_sk_callback_hsw+0x23b>
DB 196,65,101,219,192 ; vpand %ymm8,%ymm3,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,98,46,0,0 ; vbroadcastss 0x2e62(%rip),%ymm9 # 3fec <_sk_callback_hsw+0x1ff>
+ DB 196,98,125,24,13,164,46,0,0 ; vbroadcastss 0x2ea4(%rip),%ymm9 # 46ac <_sk_callback_hsw+0x23f>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,88,13,88,46,0,0 ; vpbroadcastd 0x2e58(%rip),%ymm9 # 3ff0 <_sk_callback_hsw+0x203>
+ DB 196,98,125,88,13,154,46,0,0 ; vpbroadcastd 0x2e9a(%rip),%ymm9 # 46b0 <_sk_callback_hsw+0x243>
DB 196,65,101,219,201 ; vpand %ymm9,%ymm3,%ymm9
DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9
- DB 196,98,125,24,21,73,46,0,0 ; vbroadcastss 0x2e49(%rip),%ymm10 # 3ff4 <_sk_callback_hsw+0x207>
+ DB 196,98,125,24,21,139,46,0,0 ; vbroadcastss 0x2e8b(%rip),%ymm10 # 46b4 <_sk_callback_hsw+0x247>
DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
- DB 196,98,125,88,21,63,46,0,0 ; vpbroadcastd 0x2e3f(%rip),%ymm10 # 3ff8 <_sk_callback_hsw+0x20b>
+ DB 196,98,125,88,21,129,46,0,0 ; vpbroadcastd 0x2e81(%rip),%ymm10 # 46b8 <_sk_callback_hsw+0x24b>
DB 196,193,101,219,218 ; vpand %ymm10,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,21,49,46,0,0 ; vbroadcastss 0x2e31(%rip),%ymm10 # 3ffc <_sk_callback_hsw+0x20f>
+ DB 196,98,125,24,21,115,46,0,0 ; vbroadcastss 0x2e73(%rip),%ymm10 # 46bc <_sk_callback_hsw+0x24f>
DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0
@@ -1159,16 +1509,16 @@ _sk_lerp_565_hsw LABEL PROC
DB 197,236,92,214 ; vsubps %ymm6,%ymm2,%ymm2
DB 196,226,101,168,214 ; vfmadd213ps %ymm6,%ymm3,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,10,46,0,0 ; vbroadcastss 0x2e0a(%rip),%ymm3 # 4000 <_sk_callback_hsw+0x213>
+ DB 196,226,125,24,29,76,46,0,0 ; vbroadcastss 0x2e4c(%rip),%ymm3 # 46c0 <_sk_callback_hsw+0x253>
DB 255,224 ; jmpq *%rax
DB 65,137,200 ; mov %ecx,%r8d
DB 65,128,224,7 ; and $0x7,%r8b
DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,89,255,255,255 ; ja 1169 <_sk_lerp_565_hsw+0x14>
+ DB 15,135,89,255,255,255 ; ja 17e7 <_sk_lerp_565_hsw+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 1264 <_sk_lerp_565_hsw+0x10f>
+ DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 18e4 <_sk_lerp_565_hsw+0x111>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -1180,27 +1530,28 @@ _sk_lerp_565_hsw LABEL PROC
DB 196,193,97,196,92,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3
DB 196,193,97,196,92,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3
DB 196,193,97,196,28,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3
- DB 233,5,255,255,255 ; jmpq 1169 <_sk_lerp_565_hsw+0x14>
- DB 244 ; hlt
+ DB 233,5,255,255,255 ; jmpq 17e7 <_sk_lerp_565_hsw+0x14>
+ DB 102,144 ; xchg %ax,%ax
+ DB 242,255 ; repnz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 234 ; (bad)
DB 255 ; (bad)
- DB 236 ; in (%dx),%al
DB 255 ; (bad)
+ DB 255,226 ; jmpq *%rdx
DB 255 ; (bad)
- DB 255,228 ; jmpq *%rsp
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 218,255 ; (bad)
DB 255 ; (bad)
- DB 220,255 ; fdivr %st,%st(7)
+ DB 255,210 ; callq *%rdx
DB 255 ; (bad)
- DB 255,212 ; callq *%rsp
DB 255 ; (bad)
+ DB 255,202 ; dec %edx
DB 255 ; (bad)
- DB 255,204 ; dec %esp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,192 ; inc %eax
+ DB 190 ; .byte 0xbe
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -1212,23 +1563,23 @@ _sk_load_tables_hsw LABEL PROC
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 76,3,8 ; add (%rax),%r9
DB 77,133,192 ; test %r8,%r8
- DB 117,105 ; jne 12fe <_sk_load_tables_hsw+0x7e>
+ DB 117,105 ; jne 197e <_sk_load_tables_hsw+0x7e>
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
- DB 197,229,219,13,30,48,0,0 ; vpand 0x301e(%rip),%ymm3,%ymm1 # 42c0 <_sk_callback_hsw+0x4d3>
+ DB 197,229,219,13,94,48,0,0 ; vpand 0x305e(%rip),%ymm3,%ymm1 # 4980 <_sk_callback_hsw+0x513>
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 196,226,101,0,21,30,48,0,0 ; vpshufb 0x301e(%rip),%ymm3,%ymm2 # 42e0 <_sk_callback_hsw+0x4f3>
+ DB 196,226,101,0,21,94,48,0,0 ; vpshufb 0x305e(%rip),%ymm3,%ymm2 # 49a0 <_sk_callback_hsw+0x533>
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 196,98,101,0,13,38,48,0,0 ; vpshufb 0x3026(%rip),%ymm3,%ymm9 # 4300 <_sk_callback_hsw+0x513>
+ DB 196,98,101,0,13,102,48,0,0 ; vpshufb 0x3066(%rip),%ymm3,%ymm9 # 49c0 <_sk_callback_hsw+0x553>
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 196,98,125,24,5,18,45,0,0 ; vbroadcastss 0x2d12(%rip),%ymm8 # 4004 <_sk_callback_hsw+0x217>
+ DB 196,98,125,24,5,82,45,0,0 ; vbroadcastss 0x2d52(%rip),%ymm8 # 46c4 <_sk_callback_hsw+0x257>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,193 ; mov %r8,%rcx
@@ -1241,7 +1592,7 @@ _sk_load_tables_hsw LABEL PROC
DB 196,193,249,110,194 ; vmovq %r10,%xmm0
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
- DB 233,115,255,255,255 ; jmpq 129a <_sk_load_tables_hsw+0x1a>
+ DB 233,115,255,255,255 ; jmpq 191a <_sk_load_tables_hsw+0x1a>
PUBLIC _sk_load_tables_u16_be_hsw
_sk_load_tables_u16_be_hsw LABEL PROC
@@ -1249,7 +1600,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,201,0,0,0 ; jne 1406 <_sk_load_tables_u16_be_hsw+0xdf>
+ DB 15,133,201,0,0,0 ; jne 1a86 <_sk_load_tables_u16_be_hsw+0xdf>
DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8
DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2
DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3
@@ -1265,7 +1616,7 @@ _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 197,121,111,21,178,48,0,0 ; vmovdqa 0x30b2(%rip),%xmm10 # 4440 <_sk_callback_hsw+0x653>
+ DB 197,121,111,21,242,48,0,0 ; vmovdqa 0x30f2(%rip),%xmm10 # 4b00 <_sk_callback_hsw+0x693>
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
@@ -1287,36 +1638,36 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3
DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,11,44,0,0 ; vbroadcastss 0x2c0b(%rip),%ymm8 # 4008 <_sk_callback_hsw+0x21b>
+ DB 196,98,125,24,5,75,44,0,0 ; vbroadcastss 0x2c4b(%rip),%ymm8 # 46c8 <_sk_callback_hsw+0x25b>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 146c <_sk_load_tables_u16_be_hsw+0x145>
+ DB 116,85 ; je 1aec <_sk_load_tables_u16_be_hsw+0x145>
DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 146c <_sk_load_tables_u16_be_hsw+0x145>
+ DB 114,72 ; jb 1aec <_sk_load_tables_u16_be_hsw+0x145>
DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 1479 <_sk_load_tables_u16_be_hsw+0x152>
+ DB 116,72 ; je 1af9 <_sk_load_tables_u16_be_hsw+0x152>
DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 1479 <_sk_load_tables_u16_be_hsw+0x152>
+ DB 114,59 ; jb 1af9 <_sk_load_tables_u16_be_hsw+0x152>
DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,9,255,255,255 ; je 1358 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,132,9,255,255,255 ; je 19d8 <_sk_load_tables_u16_be_hsw+0x31>
DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,248,254,255,255 ; jb 1358 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,130,248,254,255,255 ; jb 19d8 <_sk_load_tables_u16_be_hsw+0x31>
DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9
- DB 233,236,254,255,255 ; jmpq 1358 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,236,254,255,255 ; jmpq 19d8 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,223,254,255,255 ; jmpq 1358 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,223,254,255,255 ; jmpq 19d8 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,214,254,255,255 ; jmpq 1358 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,214,254,255,255 ; jmpq 19d8 <_sk_load_tables_u16_be_hsw+0x31>
PUBLIC _sk_load_tables_rgb_u16_be_hsw
_sk_load_tables_rgb_u16_be_hsw LABEL PROC
@@ -1324,7 +1675,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,193,0,0,0 ; jne 1555 <_sk_load_tables_rgb_u16_be_hsw+0xd3>
+ DB 15,133,193,0,0,0 ; jne 1bd5 <_sk_load_tables_rgb_u16_be_hsw+0xd3>
DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0
DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2
DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1
@@ -1345,7 +1696,7 @@ _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 197,121,111,13,82,47,0,0 ; vmovdqa 0x2f52(%rip),%xmm9 # 4450 <_sk_callback_hsw+0x663>
+ DB 197,121,111,13,146,47,0,0 ; vmovdqa 0x2f92(%rip),%xmm9 # 4b10 <_sk_callback_hsw+0x6a3>
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
@@ -1362,41 +1713,41 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8
DB 196,162,101,146,20,128 ; vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,185,42,0,0 ; vbroadcastss 0x2ab9(%rip),%ymm3 # 400c <_sk_callback_hsw+0x21f>
+ DB 196,226,125,24,29,249,42,0,0 ; vbroadcastss 0x2af9(%rip),%ymm3 # 46cc <_sk_callback_hsw+0x25f>
DB 255,224 ; jmpq *%rax
DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0
DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 156e <_sk_load_tables_rgb_u16_be_hsw+0xec>
- DB 233,90,255,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 1bee <_sk_load_tables_rgb_u16_be_hsw+0xec>
+ DB 233,90,255,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1
DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 159d <_sk_load_tables_rgb_u16_be_hsw+0x11b>
+ DB 114,26 ; jb 1c1d <_sk_load_tables_rgb_u16_be_hsw+0x11b>
DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1
DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 15a2 <_sk_load_tables_rgb_u16_be_hsw+0x120>
- DB 233,43,255,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,38,255,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 1c22 <_sk_load_tables_rgb_u16_be_hsw+0x120>
+ DB 233,43,255,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,38,255,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1
DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 15d1 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
+ DB 114,26 ; jb 1c51 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1
DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 15d6 <_sk_load_tables_rgb_u16_be_hsw+0x154>
- DB 233,247,254,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,242,254,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 1c56 <_sk_load_tables_rgb_u16_be_hsw+0x154>
+ DB 233,247,254,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,242,254,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3
DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 15ff <_sk_load_tables_rgb_u16_be_hsw+0x17d>
+ DB 114,20 ; jb 1c7f <_sk_load_tables_rgb_u16_be_hsw+0x17d>
DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3
DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
- DB 233,201,254,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,196,254,255,255 ; jmpq 14c8 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,201,254,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,196,254,255,255 ; jmpq 1b48 <_sk_load_tables_rgb_u16_be_hsw+0x46>
PUBLIC _sk_byte_tables_hsw
_sk_byte_tables_hsw LABEL PROC
@@ -1407,7 +1758,7 @@ _sk_byte_tables_hsw LABEL PROC
DB 65,84 ; push %r12
DB 83 ; push %rbx
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,247,41,0,0 ; vbroadcastss 0x29f7(%rip),%ymm8 # 4010 <_sk_callback_hsw+0x223>
+ DB 196,98,125,24,5,55,42,0,0 ; vbroadcastss 0x2a37(%rip),%ymm8 # 46d0 <_sk_callback_hsw+0x263>
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
@@ -1444,7 +1795,7 @@ _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 196,98,125,24,13,72,41,0,0 ; vbroadcastss 0x2948(%rip),%ymm9 # 4014 <_sk_callback_hsw+0x227>
+ DB 196,98,125,24,13,136,41,0,0 ; vbroadcastss 0x2988(%rip),%ymm9 # 46d4 <_sk_callback_hsw+0x267>
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
@@ -1603,7 +1954,7 @@ _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 196,98,125,24,13,129,38,0,0 ; vbroadcastss 0x2681(%rip),%ymm9 # 4018 <_sk_callback_hsw+0x22b>
+ DB 196,98,125,24,13,193,38,0,0 ; vbroadcastss 0x26c1(%rip),%ymm9 # 46d8 <_sk_callback_hsw+0x26b>
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
@@ -1756,33 +2107,33 @@ _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 196,98,125,24,37,52,36,0,0 ; vbroadcastss 0x2434(%rip),%ymm12 # 401c <_sk_callback_hsw+0x22f>
- DB 196,98,125,24,45,47,36,0,0 ; vbroadcastss 0x242f(%rip),%ymm13 # 4020 <_sk_callback_hsw+0x233>
+ DB 196,98,125,24,37,116,36,0,0 ; vbroadcastss 0x2474(%rip),%ymm12 # 46dc <_sk_callback_hsw+0x26f>
+ DB 196,98,125,24,45,111,36,0,0 ; vbroadcastss 0x246f(%rip),%ymm13 # 46e0 <_sk_callback_hsw+0x273>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,37,36,0,0 ; vbroadcastss 0x2425(%rip),%ymm13 # 4024 <_sk_callback_hsw+0x237>
+ DB 196,98,125,24,45,101,36,0,0 ; vbroadcastss 0x2465(%rip),%ymm13 # 46e4 <_sk_callback_hsw+0x277>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,27,36,0,0 ; vbroadcastss 0x241b(%rip),%ymm13 # 4028 <_sk_callback_hsw+0x23b>
+ DB 196,98,125,24,45,91,36,0,0 ; vbroadcastss 0x245b(%rip),%ymm13 # 46e8 <_sk_callback_hsw+0x27b>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,17,36,0,0 ; vbroadcastss 0x2411(%rip),%ymm11 # 402c <_sk_callback_hsw+0x23f>
+ DB 196,98,125,24,29,81,36,0,0 ; vbroadcastss 0x2451(%rip),%ymm11 # 46ec <_sk_callback_hsw+0x27f>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,7,36,0,0 ; vbroadcastss 0x2407(%rip),%ymm12 # 4030 <_sk_callback_hsw+0x243>
+ DB 196,98,125,24,37,71,36,0,0 ; vbroadcastss 0x2447(%rip),%ymm12 # 46f0 <_sk_callback_hsw+0x283>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,253,35,0,0 ; vbroadcastss 0x23fd(%rip),%ymm12 # 4034 <_sk_callback_hsw+0x247>
+ DB 196,98,125,24,37,61,36,0,0 ; vbroadcastss 0x243d(%rip),%ymm12 # 46f4 <_sk_callback_hsw+0x287>
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 196,98,125,24,29,222,35,0,0 ; vbroadcastss 0x23de(%rip),%ymm11 # 4038 <_sk_callback_hsw+0x24b>
+ DB 196,98,125,24,29,30,36,0,0 ; vbroadcastss 0x241e(%rip),%ymm11 # 46f8 <_sk_callback_hsw+0x28b>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,212,35,0,0 ; vbroadcastss 0x23d4(%rip),%ymm11 # 403c <_sk_callback_hsw+0x24f>
+ DB 196,98,125,24,29,20,36,0,0 ; vbroadcastss 0x2414(%rip),%ymm11 # 46fc <_sk_callback_hsw+0x28f>
DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11
- DB 196,226,125,24,5,202,35,0,0 ; vbroadcastss 0x23ca(%rip),%ymm0 # 4040 <_sk_callback_hsw+0x253>
+ DB 196,226,125,24,5,10,36,0,0 ; vbroadcastss 0x240a(%rip),%ymm0 # 4700 <_sk_callback_hsw+0x293>
DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,192,35,0,0 ; vbroadcastss 0x23c0(%rip),%ymm10 # 4044 <_sk_callback_hsw+0x257>
+ DB 196,98,125,24,21,0,36,0,0 ; vbroadcastss 0x2400(%rip),%ymm10 # 4704 <_sk_callback_hsw+0x297>
DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0
DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0
- DB 196,98,125,24,21,179,35,0,0 ; vbroadcastss 0x23b3(%rip),%ymm10 # 4048 <_sk_callback_hsw+0x25b>
+ DB 196,98,125,24,21,243,35,0,0 ; vbroadcastss 0x23f3(%rip),%ymm10 # 4708 <_sk_callback_hsw+0x29b>
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
@@ -1790,7 +2141,7 @@ _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 196,98,125,24,5,138,35,0,0 ; vbroadcastss 0x238a(%rip),%ymm8 # 404c <_sk_callback_hsw+0x25f>
+ DB 196,98,125,24,5,202,35,0,0 ; vbroadcastss 0x23ca(%rip),%ymm8 # 470c <_sk_callback_hsw+0x29f>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1808,33 +2159,33 @@ _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 196,98,125,24,37,66,35,0,0 ; vbroadcastss 0x2342(%rip),%ymm12 # 4050 <_sk_callback_hsw+0x263>
- DB 196,98,125,24,45,61,35,0,0 ; vbroadcastss 0x233d(%rip),%ymm13 # 4054 <_sk_callback_hsw+0x267>
+ DB 196,98,125,24,37,130,35,0,0 ; vbroadcastss 0x2382(%rip),%ymm12 # 4710 <_sk_callback_hsw+0x2a3>
+ DB 196,98,125,24,45,125,35,0,0 ; vbroadcastss 0x237d(%rip),%ymm13 # 4714 <_sk_callback_hsw+0x2a7>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,51,35,0,0 ; vbroadcastss 0x2333(%rip),%ymm13 # 4058 <_sk_callback_hsw+0x26b>
+ DB 196,98,125,24,45,115,35,0,0 ; vbroadcastss 0x2373(%rip),%ymm13 # 4718 <_sk_callback_hsw+0x2ab>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,41,35,0,0 ; vbroadcastss 0x2329(%rip),%ymm13 # 405c <_sk_callback_hsw+0x26f>
+ DB 196,98,125,24,45,105,35,0,0 ; vbroadcastss 0x2369(%rip),%ymm13 # 471c <_sk_callback_hsw+0x2af>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,31,35,0,0 ; vbroadcastss 0x231f(%rip),%ymm11 # 4060 <_sk_callback_hsw+0x273>
+ DB 196,98,125,24,29,95,35,0,0 ; vbroadcastss 0x235f(%rip),%ymm11 # 4720 <_sk_callback_hsw+0x2b3>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,21,35,0,0 ; vbroadcastss 0x2315(%rip),%ymm12 # 4064 <_sk_callback_hsw+0x277>
+ DB 196,98,125,24,37,85,35,0,0 ; vbroadcastss 0x2355(%rip),%ymm12 # 4724 <_sk_callback_hsw+0x2b7>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,11,35,0,0 ; vbroadcastss 0x230b(%rip),%ymm12 # 4068 <_sk_callback_hsw+0x27b>
+ DB 196,98,125,24,37,75,35,0,0 ; vbroadcastss 0x234b(%rip),%ymm12 # 4728 <_sk_callback_hsw+0x2bb>
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 196,98,125,24,29,236,34,0,0 ; vbroadcastss 0x22ec(%rip),%ymm11 # 406c <_sk_callback_hsw+0x27f>
+ DB 196,98,125,24,29,44,35,0,0 ; vbroadcastss 0x232c(%rip),%ymm11 # 472c <_sk_callback_hsw+0x2bf>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,226,34,0,0 ; vbroadcastss 0x22e2(%rip),%ymm11 # 4070 <_sk_callback_hsw+0x283>
+ DB 196,98,125,24,29,34,35,0,0 ; vbroadcastss 0x2322(%rip),%ymm11 # 4730 <_sk_callback_hsw+0x2c3>
DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11
- DB 196,226,125,24,13,216,34,0,0 ; vbroadcastss 0x22d8(%rip),%ymm1 # 4074 <_sk_callback_hsw+0x287>
+ DB 196,226,125,24,13,24,35,0,0 ; vbroadcastss 0x2318(%rip),%ymm1 # 4734 <_sk_callback_hsw+0x2c7>
DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,206,34,0,0 ; vbroadcastss 0x22ce(%rip),%ymm10 # 4078 <_sk_callback_hsw+0x28b>
+ DB 196,98,125,24,21,14,35,0,0 ; vbroadcastss 0x230e(%rip),%ymm10 # 4738 <_sk_callback_hsw+0x2cb>
DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1
DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1
- DB 196,98,125,24,21,193,34,0,0 ; vbroadcastss 0x22c1(%rip),%ymm10 # 407c <_sk_callback_hsw+0x28f>
+ DB 196,98,125,24,21,1,35,0,0 ; vbroadcastss 0x2301(%rip),%ymm10 # 473c <_sk_callback_hsw+0x2cf>
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
@@ -1842,7 +2193,7 @@ _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 196,98,125,24,5,152,34,0,0 ; vbroadcastss 0x2298(%rip),%ymm8 # 4080 <_sk_callback_hsw+0x293>
+ DB 196,98,125,24,5,216,34,0,0 ; vbroadcastss 0x22d8(%rip),%ymm8 # 4740 <_sk_callback_hsw+0x2d3>
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1860,33 +2211,33 @@ _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 196,98,125,24,37,80,34,0,0 ; vbroadcastss 0x2250(%rip),%ymm12 # 4084 <_sk_callback_hsw+0x297>
- DB 196,98,125,24,45,75,34,0,0 ; vbroadcastss 0x224b(%rip),%ymm13 # 4088 <_sk_callback_hsw+0x29b>
+ DB 196,98,125,24,37,144,34,0,0 ; vbroadcastss 0x2290(%rip),%ymm12 # 4744 <_sk_callback_hsw+0x2d7>
+ DB 196,98,125,24,45,139,34,0,0 ; vbroadcastss 0x228b(%rip),%ymm13 # 4748 <_sk_callback_hsw+0x2db>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,65,34,0,0 ; vbroadcastss 0x2241(%rip),%ymm13 # 408c <_sk_callback_hsw+0x29f>
+ DB 196,98,125,24,45,129,34,0,0 ; vbroadcastss 0x2281(%rip),%ymm13 # 474c <_sk_callback_hsw+0x2df>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,55,34,0,0 ; vbroadcastss 0x2237(%rip),%ymm13 # 4090 <_sk_callback_hsw+0x2a3>
+ DB 196,98,125,24,45,119,34,0,0 ; vbroadcastss 0x2277(%rip),%ymm13 # 4750 <_sk_callback_hsw+0x2e3>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,45,34,0,0 ; vbroadcastss 0x222d(%rip),%ymm11 # 4094 <_sk_callback_hsw+0x2a7>
+ DB 196,98,125,24,29,109,34,0,0 ; vbroadcastss 0x226d(%rip),%ymm11 # 4754 <_sk_callback_hsw+0x2e7>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,35,34,0,0 ; vbroadcastss 0x2223(%rip),%ymm12 # 4098 <_sk_callback_hsw+0x2ab>
+ DB 196,98,125,24,37,99,34,0,0 ; vbroadcastss 0x2263(%rip),%ymm12 # 4758 <_sk_callback_hsw+0x2eb>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,25,34,0,0 ; vbroadcastss 0x2219(%rip),%ymm12 # 409c <_sk_callback_hsw+0x2af>
+ DB 196,98,125,24,37,89,34,0,0 ; vbroadcastss 0x2259(%rip),%ymm12 # 475c <_sk_callback_hsw+0x2ef>
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 196,98,125,24,29,250,33,0,0 ; vbroadcastss 0x21fa(%rip),%ymm11 # 40a0 <_sk_callback_hsw+0x2b3>
+ DB 196,98,125,24,29,58,34,0,0 ; vbroadcastss 0x223a(%rip),%ymm11 # 4760 <_sk_callback_hsw+0x2f3>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,240,33,0,0 ; vbroadcastss 0x21f0(%rip),%ymm11 # 40a4 <_sk_callback_hsw+0x2b7>
+ DB 196,98,125,24,29,48,34,0,0 ; vbroadcastss 0x2230(%rip),%ymm11 # 4764 <_sk_callback_hsw+0x2f7>
DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11
- DB 196,226,125,24,21,230,33,0,0 ; vbroadcastss 0x21e6(%rip),%ymm2 # 40a8 <_sk_callback_hsw+0x2bb>
+ DB 196,226,125,24,21,38,34,0,0 ; vbroadcastss 0x2226(%rip),%ymm2 # 4768 <_sk_callback_hsw+0x2fb>
DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,220,33,0,0 ; vbroadcastss 0x21dc(%rip),%ymm10 # 40ac <_sk_callback_hsw+0x2bf>
+ DB 196,98,125,24,21,28,34,0,0 ; vbroadcastss 0x221c(%rip),%ymm10 # 476c <_sk_callback_hsw+0x2ff>
DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2
DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2
- DB 196,98,125,24,21,207,33,0,0 ; vbroadcastss 0x21cf(%rip),%ymm10 # 40b0 <_sk_callback_hsw+0x2c3>
+ DB 196,98,125,24,21,15,34,0,0 ; vbroadcastss 0x220f(%rip),%ymm10 # 4770 <_sk_callback_hsw+0x303>
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
@@ -1894,7 +2245,7 @@ _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 196,98,125,24,5,166,33,0,0 ; vbroadcastss 0x21a6(%rip),%ymm8 # 40b4 <_sk_callback_hsw+0x2c7>
+ DB 196,98,125,24,5,230,33,0,0 ; vbroadcastss 0x21e6(%rip),%ymm8 # 4774 <_sk_callback_hsw+0x307>
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1912,33 +2263,33 @@ _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 196,98,125,24,37,94,33,0,0 ; vbroadcastss 0x215e(%rip),%ymm12 # 40b8 <_sk_callback_hsw+0x2cb>
- DB 196,98,125,24,45,89,33,0,0 ; vbroadcastss 0x2159(%rip),%ymm13 # 40bc <_sk_callback_hsw+0x2cf>
+ DB 196,98,125,24,37,158,33,0,0 ; vbroadcastss 0x219e(%rip),%ymm12 # 4778 <_sk_callback_hsw+0x30b>
+ DB 196,98,125,24,45,153,33,0,0 ; vbroadcastss 0x2199(%rip),%ymm13 # 477c <_sk_callback_hsw+0x30f>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,79,33,0,0 ; vbroadcastss 0x214f(%rip),%ymm13 # 40c0 <_sk_callback_hsw+0x2d3>
+ DB 196,98,125,24,45,143,33,0,0 ; vbroadcastss 0x218f(%rip),%ymm13 # 4780 <_sk_callback_hsw+0x313>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,69,33,0,0 ; vbroadcastss 0x2145(%rip),%ymm13 # 40c4 <_sk_callback_hsw+0x2d7>
+ DB 196,98,125,24,45,133,33,0,0 ; vbroadcastss 0x2185(%rip),%ymm13 # 4784 <_sk_callback_hsw+0x317>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,59,33,0,0 ; vbroadcastss 0x213b(%rip),%ymm11 # 40c8 <_sk_callback_hsw+0x2db>
+ DB 196,98,125,24,29,123,33,0,0 ; vbroadcastss 0x217b(%rip),%ymm11 # 4788 <_sk_callback_hsw+0x31b>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,49,33,0,0 ; vbroadcastss 0x2131(%rip),%ymm12 # 40cc <_sk_callback_hsw+0x2df>
+ DB 196,98,125,24,37,113,33,0,0 ; vbroadcastss 0x2171(%rip),%ymm12 # 478c <_sk_callback_hsw+0x31f>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,39,33,0,0 ; vbroadcastss 0x2127(%rip),%ymm12 # 40d0 <_sk_callback_hsw+0x2e3>
+ DB 196,98,125,24,37,103,33,0,0 ; vbroadcastss 0x2167(%rip),%ymm12 # 4790 <_sk_callback_hsw+0x323>
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 196,98,125,24,29,8,33,0,0 ; vbroadcastss 0x2108(%rip),%ymm11 # 40d4 <_sk_callback_hsw+0x2e7>
+ DB 196,98,125,24,29,72,33,0,0 ; vbroadcastss 0x2148(%rip),%ymm11 # 4794 <_sk_callback_hsw+0x327>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,254,32,0,0 ; vbroadcastss 0x20fe(%rip),%ymm11 # 40d8 <_sk_callback_hsw+0x2eb>
+ DB 196,98,125,24,29,62,33,0,0 ; vbroadcastss 0x213e(%rip),%ymm11 # 4798 <_sk_callback_hsw+0x32b>
DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11
- DB 196,226,125,24,29,244,32,0,0 ; vbroadcastss 0x20f4(%rip),%ymm3 # 40dc <_sk_callback_hsw+0x2ef>
+ DB 196,226,125,24,29,52,33,0,0 ; vbroadcastss 0x2134(%rip),%ymm3 # 479c <_sk_callback_hsw+0x32f>
DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,234,32,0,0 ; vbroadcastss 0x20ea(%rip),%ymm10 # 40e0 <_sk_callback_hsw+0x2f3>
+ DB 196,98,125,24,21,42,33,0,0 ; vbroadcastss 0x212a(%rip),%ymm10 # 47a0 <_sk_callback_hsw+0x333>
DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3
DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3
- DB 196,98,125,24,21,221,32,0,0 ; vbroadcastss 0x20dd(%rip),%ymm10 # 40e4 <_sk_callback_hsw+0x2f7>
+ DB 196,98,125,24,21,29,33,0,0 ; vbroadcastss 0x211d(%rip),%ymm10 # 47a4 <_sk_callback_hsw+0x337>
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
@@ -1946,33 +2297,33 @@ _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 196,98,125,24,5,180,32,0,0 ; vbroadcastss 0x20b4(%rip),%ymm8 # 40e8 <_sk_callback_hsw+0x2fb>
+ DB 196,98,125,24,5,244,32,0,0 ; vbroadcastss 0x20f4(%rip),%ymm8 # 47a8 <_sk_callback_hsw+0x33b>
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 196,98,125,24,5,166,32,0,0 ; vbroadcastss 0x20a6(%rip),%ymm8 # 40ec <_sk_callback_hsw+0x2ff>
- DB 196,98,125,24,13,161,32,0,0 ; vbroadcastss 0x20a1(%rip),%ymm9 # 40f0 <_sk_callback_hsw+0x303>
- DB 196,98,125,24,21,156,32,0,0 ; vbroadcastss 0x209c(%rip),%ymm10 # 40f4 <_sk_callback_hsw+0x307>
+ DB 196,98,125,24,5,230,32,0,0 ; vbroadcastss 0x20e6(%rip),%ymm8 # 47ac <_sk_callback_hsw+0x33f>
+ DB 196,98,125,24,13,225,32,0,0 ; vbroadcastss 0x20e1(%rip),%ymm9 # 47b0 <_sk_callback_hsw+0x343>
+ DB 196,98,125,24,21,220,32,0,0 ; vbroadcastss 0x20dc(%rip),%ymm10 # 47b4 <_sk_callback_hsw+0x347>
DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1
DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2
- DB 196,98,125,24,13,141,32,0,0 ; vbroadcastss 0x208d(%rip),%ymm9 # 40f8 <_sk_callback_hsw+0x30b>
+ DB 196,98,125,24,13,205,32,0,0 ; vbroadcastss 0x20cd(%rip),%ymm9 # 47b8 <_sk_callback_hsw+0x34b>
DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9
- DB 196,226,125,24,5,131,32,0,0 ; vbroadcastss 0x2083(%rip),%ymm0 # 40fc <_sk_callback_hsw+0x30f>
+ DB 196,226,125,24,5,195,32,0,0 ; vbroadcastss 0x20c3(%rip),%ymm0 # 47bc <_sk_callback_hsw+0x34f>
DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0
- DB 196,98,125,24,5,122,32,0,0 ; vbroadcastss 0x207a(%rip),%ymm8 # 4100 <_sk_callback_hsw+0x313>
+ DB 196,98,125,24,5,186,32,0,0 ; vbroadcastss 0x20ba(%rip),%ymm8 # 47c0 <_sk_callback_hsw+0x353>
DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8
- DB 196,98,125,24,13,112,32,0,0 ; vbroadcastss 0x2070(%rip),%ymm9 # 4104 <_sk_callback_hsw+0x317>
+ DB 196,98,125,24,13,176,32,0,0 ; vbroadcastss 0x20b0(%rip),%ymm9 # 47c4 <_sk_callback_hsw+0x357>
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 196,226,125,24,21,93,32,0,0 ; vbroadcastss 0x205d(%rip),%ymm2 # 4108 <_sk_callback_hsw+0x31b>
+ DB 196,226,125,24,21,157,32,0,0 ; vbroadcastss 0x209d(%rip),%ymm2 # 47c8 <_sk_callback_hsw+0x35b>
DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10
- DB 196,98,125,24,29,83,32,0,0 ; vbroadcastss 0x2053(%rip),%ymm11 # 410c <_sk_callback_hsw+0x31f>
+ DB 196,98,125,24,29,147,32,0,0 ; vbroadcastss 0x2093(%rip),%ymm11 # 47cc <_sk_callback_hsw+0x35f>
DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8
- DB 196,98,125,24,37,73,32,0,0 ; vbroadcastss 0x2049(%rip),%ymm12 # 4110 <_sk_callback_hsw+0x323>
+ DB 196,98,125,24,37,137,32,0,0 ; vbroadcastss 0x2089(%rip),%ymm12 # 47d0 <_sk_callback_hsw+0x363>
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
@@ -1987,9 +2338,9 @@ _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 196,226,125,24,5,254,31,0,0 ; vbroadcastss 0x1ffe(%rip),%ymm0 # 4114 <_sk_callback_hsw+0x327>
+ DB 196,226,125,24,5,62,32,0,0 ; vbroadcastss 0x203e(%rip),%ymm0 # 47d4 <_sk_callback_hsw+0x367>
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,245,31,0,0 ; vbroadcastss 0x1ff5(%rip),%ymm8 # 4118 <_sk_callback_hsw+0x32b>
+ DB 196,98,125,24,5,53,32,0,0 ; vbroadcastss 0x2035(%rip),%ymm8 # 47d8 <_sk_callback_hsw+0x36b>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2001,11 +2352,11 @@ _sk_load_a8_hsw LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,45 ; jne 2169 <_sk_load_a8_hsw+0x3d>
+ DB 117,45 ; jne 27e9 <_sk_load_a8_hsw+0x3d>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,202,31,0,0 ; vbroadcastss 0x1fca(%rip),%ymm1 # 411c <_sk_callback_hsw+0x32f>
+ DB 196,226,125,24,13,10,32,0,0 ; vbroadcastss 0x200a(%rip),%ymm1 # 47dc <_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
@@ -2022,9 +2373,9 @@ _sk_load_a8_hsw LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 2171 <_sk_load_a8_hsw+0x45>
+ DB 117,234 ; jne 27f1 <_sk_load_a8_hsw+0x45>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,178 ; jmp 2140 <_sk_load_a8_hsw+0x14>
+ DB 235,178 ; jmp 27c0 <_sk_load_a8_hsw+0x14>
PUBLIC _sk_gather_a8_hsw
_sk_gather_a8_hsw LABEL PROC
@@ -2068,7 +2419,7 @@ _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 196,226,125,24,13,213,30,0,0 ; vbroadcastss 0x1ed5(%rip),%ymm1 # 4120 <_sk_callback_hsw+0x333>
+ DB 196,226,125,24,13,21,31,0,0 ; vbroadcastss 0x1f15(%rip),%ymm1 # 47e0 <_sk_callback_hsw+0x373>
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
@@ -2084,14 +2435,14 @@ PUBLIC _sk_store_a8_hsw
_sk_store_a8_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
- DB 196,98,125,24,5,176,30,0,0 ; vbroadcastss 0x1eb0(%rip),%ymm8 # 4124 <_sk_callback_hsw+0x337>
+ DB 196,98,125,24,5,240,30,0,0 ; vbroadcastss 0x1ef0(%rip),%ymm8 # 47e4 <_sk_callback_hsw+0x377>
DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8
DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 229d <_sk_store_a8_hsw+0x37>
+ DB 117,10 ; jne 291d <_sk_store_a8_hsw+0x37>
DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2099,10 +2450,10 @@ _sk_store_a8_hsw LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 2299 <_sk_store_a8_hsw+0x33>
+ DB 119,236 ; ja 2919 <_sk_store_a8_hsw+0x33>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 2300 <_sk_store_a8_hsw+0x9a>
+ DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 2980 <_sk_store_a8_hsw+0x9a>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -2113,7 +2464,7 @@ _sk_store_a8_hsw LABEL PROC
DB 196,67,121,20,68,58,2,4 ; vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1)
DB 196,67,121,20,68,58,1,2 ; vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1)
DB 196,67,121,20,4,58,0 ; vpextrb $0x0,%xmm8,(%r10,%rdi,1)
- DB 235,154 ; jmp 2299 <_sk_store_a8_hsw+0x33>
+ DB 235,154 ; jmp 2919 <_sk_store_a8_hsw+0x33>
DB 144 ; nop
DB 246,255 ; idiv %bh
DB 255 ; (bad)
@@ -2145,14 +2496,14 @@ _sk_load_g8_hsw LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,50 ; jne 235e <_sk_load_g8_hsw+0x42>
+ DB 117,50 ; jne 29de <_sk_load_g8_hsw+0x42>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,230,29,0,0 ; vbroadcastss 0x1de6(%rip),%ymm1 # 4128 <_sk_callback_hsw+0x33b>
+ DB 196,226,125,24,13,38,30,0,0 ; vbroadcastss 0x1e26(%rip),%ymm1 # 47e8 <_sk_callback_hsw+0x37b>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,219,29,0,0 ; vbroadcastss 0x1ddb(%rip),%ymm3 # 412c <_sk_callback_hsw+0x33f>
+ DB 196,226,125,24,29,27,30,0,0 ; vbroadcastss 0x1e1b(%rip),%ymm3 # 47ec <_sk_callback_hsw+0x37f>
DB 76,137,193 ; mov %r8,%rcx
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
@@ -2166,9 +2517,9 @@ _sk_load_g8_hsw LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 2366 <_sk_load_g8_hsw+0x4a>
+ DB 117,234 ; jne 29e6 <_sk_load_g8_hsw+0x4a>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,173 ; jmp 2330 <_sk_load_g8_hsw+0x14>
+ DB 235,173 ; jmp 29b0 <_sk_load_g8_hsw+0x14>
PUBLIC _sk_gather_g8_hsw
_sk_gather_g8_hsw LABEL PROC
@@ -2212,10 +2563,10 @@ _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 196,226,125,24,13,240,28,0,0 ; vbroadcastss 0x1cf0(%rip),%ymm1 # 4130 <_sk_callback_hsw+0x343>
+ DB 196,226,125,24,13,48,29,0,0 ; vbroadcastss 0x1d30(%rip),%ymm1 # 47f0 <_sk_callback_hsw+0x383>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,229,28,0,0 ; vbroadcastss 0x1ce5(%rip),%ymm3 # 4134 <_sk_callback_hsw+0x347>
+ DB 196,226,125,24,29,37,29,0,0 ; vbroadcastss 0x1d25(%rip),%ymm3 # 47f4 <_sk_callback_hsw+0x387>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 91 ; pop %rbx
@@ -2229,9 +2580,9 @@ _sk_gather_i8_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,137,192 ; mov %rax,%r8
DB 77,133,192 ; test %r8,%r8
- DB 116,5 ; je 246f <_sk_gather_i8_hsw+0xf>
+ DB 116,5 ; je 2aef <_sk_gather_i8_hsw+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 2471 <_sk_gather_i8_hsw+0x11>
+ DB 235,2 ; jmp 2af1 <_sk_gather_i8_hsw+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -2269,14 +2620,14 @@ _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 197,229,219,5,1,30,0,0 ; vpand 0x1e01(%rip),%ymm3,%ymm0 # 4320 <_sk_callback_hsw+0x533>
+ DB 197,229,219,5,65,30,0,0 ; vpand 0x1e41(%rip),%ymm3,%ymm0 # 49e0 <_sk_callback_hsw+0x573>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,12,28,0,0 ; vbroadcastss 0x1c0c(%rip),%ymm8 # 4138 <_sk_callback_hsw+0x34b>
+ DB 196,98,125,24,5,76,28,0,0 ; vbroadcastss 0x1c4c(%rip),%ymm8 # 47f8 <_sk_callback_hsw+0x38b>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,6,30,0,0 ; vpshufb 0x1e06(%rip),%ymm3,%ymm1 # 4340 <_sk_callback_hsw+0x553>
+ DB 196,226,101,0,13,70,30,0,0 ; vpshufb 0x1e46(%rip),%ymm3,%ymm1 # 4a00 <_sk_callback_hsw+0x593>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,20,30,0,0 ; vpshufb 0x1e14(%rip),%ymm3,%ymm2 # 4360 <_sk_callback_hsw+0x573>
+ DB 196,226,101,0,21,84,30,0,0 ; vpshufb 0x1e54(%rip),%ymm3,%ymm2 # 4a20 <_sk_callback_hsw+0x5b3>
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
@@ -2295,35 +2646,35 @@ _sk_load_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 117,114 ; jne 25ec <_sk_load_565_hsw+0x7c>
+ DB 117,114 ; jne 2c6c <_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 196,226,125,88,5,174,27,0,0 ; vpbroadcastd 0x1bae(%rip),%ymm0 # 413c <_sk_callback_hsw+0x34f>
+ DB 196,226,125,88,5,238,27,0,0 ; vpbroadcastd 0x1bee(%rip),%ymm0 # 47fc <_sk_callback_hsw+0x38f>
DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,161,27,0,0 ; vbroadcastss 0x1ba1(%rip),%ymm1 # 4140 <_sk_callback_hsw+0x353>
+ DB 196,226,125,24,13,225,27,0,0 ; vbroadcastss 0x1be1(%rip),%ymm1 # 4800 <_sk_callback_hsw+0x393>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,152,27,0,0 ; vpbroadcastd 0x1b98(%rip),%ymm1 # 4144 <_sk_callback_hsw+0x357>
+ DB 196,226,125,88,13,216,27,0,0 ; vpbroadcastd 0x1bd8(%rip),%ymm1 # 4804 <_sk_callback_hsw+0x397>
DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,139,27,0,0 ; vbroadcastss 0x1b8b(%rip),%ymm3 # 4148 <_sk_callback_hsw+0x35b>
+ DB 196,226,125,24,29,203,27,0,0 ; vbroadcastss 0x1bcb(%rip),%ymm3 # 4808 <_sk_callback_hsw+0x39b>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,88,29,130,27,0,0 ; vpbroadcastd 0x1b82(%rip),%ymm3 # 414c <_sk_callback_hsw+0x35f>
+ DB 196,226,125,88,29,194,27,0,0 ; vpbroadcastd 0x1bc2(%rip),%ymm3 # 480c <_sk_callback_hsw+0x39f>
DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,117,27,0,0 ; vbroadcastss 0x1b75(%rip),%ymm3 # 4150 <_sk_callback_hsw+0x363>
+ DB 196,226,125,24,29,181,27,0,0 ; vbroadcastss 0x1bb5(%rip),%ymm3 # 4810 <_sk_callback_hsw+0x3a3>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,106,27,0,0 ; vbroadcastss 0x1b6a(%rip),%ymm3 # 4154 <_sk_callback_hsw+0x367>
+ DB 196,226,125,24,29,170,27,0,0 ; vbroadcastss 0x1baa(%rip),%ymm3 # 4814 <_sk_callback_hsw+0x3a7>
DB 255,224 ; jmpq *%rax
DB 65,137,200 ; mov %ecx,%r8d
DB 65,128,224,7 ; and $0x7,%r8b
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,128 ; ja 2580 <_sk_load_565_hsw+0x10>
+ DB 119,128 ; ja 2c00 <_sk_load_565_hsw+0x10>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2654 <_sk_load_565_hsw+0xe4>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2cd4 <_sk_load_565_hsw+0xe4>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -2335,7 +2686,7 @@ _sk_load_565_hsw LABEL PROC
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- DB 233,44,255,255,255 ; jmpq 2580 <_sk_load_565_hsw+0x10>
+ DB 233,44,255,255,255 ; jmpq 2c00 <_sk_load_565_hsw+0x10>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2403,23 +2754,23 @@ _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 196,226,125,88,5,45,26,0,0 ; vpbroadcastd 0x1a2d(%rip),%ymm0 # 4158 <_sk_callback_hsw+0x36b>
+ DB 196,226,125,88,5,109,26,0,0 ; vpbroadcastd 0x1a6d(%rip),%ymm0 # 4818 <_sk_callback_hsw+0x3ab>
DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,32,26,0,0 ; vbroadcastss 0x1a20(%rip),%ymm1 # 415c <_sk_callback_hsw+0x36f>
+ DB 196,226,125,24,13,96,26,0,0 ; vbroadcastss 0x1a60(%rip),%ymm1 # 481c <_sk_callback_hsw+0x3af>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,23,26,0,0 ; vpbroadcastd 0x1a17(%rip),%ymm1 # 4160 <_sk_callback_hsw+0x373>
+ DB 196,226,125,88,13,87,26,0,0 ; vpbroadcastd 0x1a57(%rip),%ymm1 # 4820 <_sk_callback_hsw+0x3b3>
DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,10,26,0,0 ; vbroadcastss 0x1a0a(%rip),%ymm3 # 4164 <_sk_callback_hsw+0x377>
+ DB 196,226,125,24,29,74,26,0,0 ; vbroadcastss 0x1a4a(%rip),%ymm3 # 4824 <_sk_callback_hsw+0x3b7>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,88,29,1,26,0,0 ; vpbroadcastd 0x1a01(%rip),%ymm3 # 4168 <_sk_callback_hsw+0x37b>
+ DB 196,226,125,88,29,65,26,0,0 ; vpbroadcastd 0x1a41(%rip),%ymm3 # 4828 <_sk_callback_hsw+0x3bb>
DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,244,25,0,0 ; vbroadcastss 0x19f4(%rip),%ymm3 # 416c <_sk_callback_hsw+0x37f>
+ DB 196,226,125,24,29,52,26,0,0 ; vbroadcastss 0x1a34(%rip),%ymm3 # 482c <_sk_callback_hsw+0x3bf>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,233,25,0,0 ; vbroadcastss 0x19e9(%rip),%ymm3 # 4170 <_sk_callback_hsw+0x383>
+ DB 196,226,125,24,29,41,26,0,0 ; vbroadcastss 0x1a29(%rip),%ymm3 # 4830 <_sk_callback_hsw+0x3c3>
DB 91 ; pop %rbx
DB 65,92 ; pop %r12
DB 65,94 ; pop %r14
@@ -2430,11 +2781,11 @@ 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 196,98,125,24,5,214,25,0,0 ; vbroadcastss 0x19d6(%rip),%ymm8 # 4174 <_sk_callback_hsw+0x387>
+ DB 196,98,125,24,5,22,26,0,0 ; vbroadcastss 0x1a16(%rip),%ymm8 # 4834 <_sk_callback_hsw+0x3c7>
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 196,98,125,24,21,193,25,0,0 ; vbroadcastss 0x19c1(%rip),%ymm10 # 4178 <_sk_callback_hsw+0x38b>
+ DB 196,98,125,24,21,1,26,0,0 ; vbroadcastss 0x1a01(%rip),%ymm10 # 4838 <_sk_callback_hsw+0x3cb>
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
@@ -2445,7 +2796,7 @@ _sk_store_565_hsw LABEL PROC
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 27f5 <_sk_store_565_hsw+0x65>
+ DB 117,10 ; jne 2e75 <_sk_store_565_hsw+0x65>
DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2453,9 +2804,9 @@ _sk_store_565_hsw LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 27f1 <_sk_store_565_hsw+0x61>
+ DB 119,236 ; ja 2e71 <_sk_store_565_hsw+0x61>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2854 <_sk_store_565_hsw+0xc4>
+ DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2ed4 <_sk_store_565_hsw+0xc4>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -2466,7 +2817,7 @@ _sk_store_565_hsw LABEL PROC
DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- DB 235,159 ; jmp 27f1 <_sk_store_565_hsw+0x61>
+ DB 235,159 ; jmp 2e71 <_sk_store_565_hsw+0x61>
DB 102,144 ; xchg %ax,%ax
DB 245 ; cmc
DB 255 ; (bad)
@@ -2497,28 +2848,28 @@ _sk_load_4444_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,138,0,0,0 ; jne 2908 <_sk_load_4444_hsw+0x98>
+ DB 15,133,138,0,0,0 ; jne 2f88 <_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 196,226,125,88,5,234,24,0,0 ; vpbroadcastd 0x18ea(%rip),%ymm0 # 417c <_sk_callback_hsw+0x38f>
+ DB 196,226,125,88,5,42,25,0,0 ; vpbroadcastd 0x192a(%rip),%ymm0 # 483c <_sk_callback_hsw+0x3cf>
DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,221,24,0,0 ; vbroadcastss 0x18dd(%rip),%ymm1 # 4180 <_sk_callback_hsw+0x393>
+ DB 196,226,125,24,13,29,25,0,0 ; vbroadcastss 0x191d(%rip),%ymm1 # 4840 <_sk_callback_hsw+0x3d3>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,212,24,0,0 ; vpbroadcastd 0x18d4(%rip),%ymm1 # 4184 <_sk_callback_hsw+0x397>
+ DB 196,226,125,88,13,20,25,0,0 ; vpbroadcastd 0x1914(%rip),%ymm1 # 4844 <_sk_callback_hsw+0x3d7>
DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,199,24,0,0 ; vbroadcastss 0x18c7(%rip),%ymm2 # 4188 <_sk_callback_hsw+0x39b>
+ DB 196,226,125,24,21,7,25,0,0 ; vbroadcastss 0x1907(%rip),%ymm2 # 4848 <_sk_callback_hsw+0x3db>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,88,21,190,24,0,0 ; vpbroadcastd 0x18be(%rip),%ymm2 # 418c <_sk_callback_hsw+0x39f>
+ DB 196,226,125,88,21,254,24,0,0 ; vpbroadcastd 0x18fe(%rip),%ymm2 # 484c <_sk_callback_hsw+0x3df>
DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,177,24,0,0 ; vbroadcastss 0x18b1(%rip),%ymm8 # 4190 <_sk_callback_hsw+0x3a3>
+ DB 196,98,125,24,5,241,24,0,0 ; vbroadcastss 0x18f1(%rip),%ymm8 # 4850 <_sk_callback_hsw+0x3e3>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,88,5,167,24,0,0 ; vpbroadcastd 0x18a7(%rip),%ymm8 # 4194 <_sk_callback_hsw+0x3a7>
+ DB 196,98,125,88,5,231,24,0,0 ; vpbroadcastd 0x18e7(%rip),%ymm8 # 4854 <_sk_callback_hsw+0x3e7>
DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,153,24,0,0 ; vbroadcastss 0x1899(%rip),%ymm8 # 4198 <_sk_callback_hsw+0x3ab>
+ DB 196,98,125,24,5,217,24,0,0 ; vbroadcastss 0x18d9(%rip),%ymm8 # 4858 <_sk_callback_hsw+0x3eb>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2527,9 +2878,9 @@ _sk_load_4444_hsw LABEL PROC
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,100,255,255,255 ; ja 2884 <_sk_load_4444_hsw+0x14>
+ DB 15,135,100,255,255,255 ; ja 2f04 <_sk_load_4444_hsw+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2974 <_sk_load_4444_hsw+0x104>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2ff4 <_sk_load_4444_hsw+0x104>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -2541,7 +2892,7 @@ _sk_load_4444_hsw LABEL PROC
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- DB 233,16,255,255,255 ; jmpq 2884 <_sk_load_4444_hsw+0x14>
+ DB 233,16,255,255,255 ; jmpq 2f04 <_sk_load_4444_hsw+0x14>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2609,25 +2960,25 @@ _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 196,226,125,88,5,81,23,0,0 ; vpbroadcastd 0x1751(%rip),%ymm0 # 419c <_sk_callback_hsw+0x3af>
+ DB 196,226,125,88,5,145,23,0,0 ; vpbroadcastd 0x1791(%rip),%ymm0 # 485c <_sk_callback_hsw+0x3ef>
DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,68,23,0,0 ; vbroadcastss 0x1744(%rip),%ymm1 # 41a0 <_sk_callback_hsw+0x3b3>
+ DB 196,226,125,24,13,132,23,0,0 ; vbroadcastss 0x1784(%rip),%ymm1 # 4860 <_sk_callback_hsw+0x3f3>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,59,23,0,0 ; vpbroadcastd 0x173b(%rip),%ymm1 # 41a4 <_sk_callback_hsw+0x3b7>
+ DB 196,226,125,88,13,123,23,0,0 ; vpbroadcastd 0x177b(%rip),%ymm1 # 4864 <_sk_callback_hsw+0x3f7>
DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,46,23,0,0 ; vbroadcastss 0x172e(%rip),%ymm2 # 41a8 <_sk_callback_hsw+0x3bb>
+ DB 196,226,125,24,21,110,23,0,0 ; vbroadcastss 0x176e(%rip),%ymm2 # 4868 <_sk_callback_hsw+0x3fb>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,88,21,37,23,0,0 ; vpbroadcastd 0x1725(%rip),%ymm2 # 41ac <_sk_callback_hsw+0x3bf>
+ DB 196,226,125,88,21,101,23,0,0 ; vpbroadcastd 0x1765(%rip),%ymm2 # 486c <_sk_callback_hsw+0x3ff>
DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,24,23,0,0 ; vbroadcastss 0x1718(%rip),%ymm8 # 41b0 <_sk_callback_hsw+0x3c3>
+ DB 196,98,125,24,5,88,23,0,0 ; vbroadcastss 0x1758(%rip),%ymm8 # 4870 <_sk_callback_hsw+0x403>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,88,5,14,23,0,0 ; vpbroadcastd 0x170e(%rip),%ymm8 # 41b4 <_sk_callback_hsw+0x3c7>
+ DB 196,98,125,88,5,78,23,0,0 ; vpbroadcastd 0x174e(%rip),%ymm8 # 4874 <_sk_callback_hsw+0x407>
DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,0,23,0,0 ; vbroadcastss 0x1700(%rip),%ymm8 # 41b8 <_sk_callback_hsw+0x3cb>
+ DB 196,98,125,24,5,64,23,0,0 ; vbroadcastss 0x1740(%rip),%ymm8 # 4878 <_sk_callback_hsw+0x40b>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 91 ; pop %rbx
@@ -2640,7 +2991,7 @@ 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 196,98,125,24,5,230,22,0,0 ; vbroadcastss 0x16e6(%rip),%ymm8 # 41bc <_sk_callback_hsw+0x3cf>
+ DB 196,98,125,24,5,38,23,0,0 ; vbroadcastss 0x1726(%rip),%ymm8 # 487c <_sk_callback_hsw+0x40f>
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
@@ -2658,7 +3009,7 @@ _sk_store_4444_hsw LABEL PROC
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 2b39 <_sk_store_4444_hsw+0x71>
+ DB 117,10 ; jne 31b9 <_sk_store_4444_hsw+0x71>
DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2666,9 +3017,9 @@ _sk_store_4444_hsw LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 2b35 <_sk_store_4444_hsw+0x6d>
+ DB 119,236 ; ja 31b5 <_sk_store_4444_hsw+0x6d>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2b98 <_sk_store_4444_hsw+0xd0>
+ DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3218 <_sk_store_4444_hsw+0xd0>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -2679,7 +3030,7 @@ _sk_store_4444_hsw LABEL PROC
DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- DB 235,159 ; jmp 2b35 <_sk_store_4444_hsw+0x6d>
+ DB 235,159 ; jmp 31b5 <_sk_store_4444_hsw+0x6d>
DB 102,144 ; xchg %ax,%ax
DB 245 ; cmc
DB 255 ; (bad)
@@ -2712,16 +3063,16 @@ _sk_load_8888_hsw LABEL PROC
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 76,3,8 ; add (%rax),%r9
DB 77,133,192 ; test %r8,%r8
- DB 117,88 ; jne 2c21 <_sk_load_8888_hsw+0x6d>
+ DB 117,88 ; jne 32a1 <_sk_load_8888_hsw+0x6d>
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
- DB 197,229,219,5,170,23,0,0 ; vpand 0x17aa(%rip),%ymm3,%ymm0 # 4380 <_sk_callback_hsw+0x593>
+ DB 197,229,219,5,234,23,0,0 ; vpand 0x17ea(%rip),%ymm3,%ymm0 # 4a40 <_sk_callback_hsw+0x5d3>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,221,21,0,0 ; vbroadcastss 0x15dd(%rip),%ymm8 # 41c0 <_sk_callback_hsw+0x3d3>
+ DB 196,98,125,24,5,29,22,0,0 ; vbroadcastss 0x161d(%rip),%ymm8 # 4880 <_sk_callback_hsw+0x413>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,175,23,0,0 ; vpshufb 0x17af(%rip),%ymm3,%ymm1 # 43a0 <_sk_callback_hsw+0x5b3>
+ DB 196,226,101,0,13,239,23,0,0 ; vpshufb 0x17ef(%rip),%ymm3,%ymm1 # 4a60 <_sk_callback_hsw+0x5f3>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,189,23,0,0 ; vpshufb 0x17bd(%rip),%ymm3,%ymm2 # 43c0 <_sk_callback_hsw+0x5d3>
+ DB 196,226,101,0,21,253,23,0,0 ; vpshufb 0x17fd(%rip),%ymm3,%ymm2 # 4a80 <_sk_callback_hsw+0x613>
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
@@ -2738,7 +3089,7 @@ _sk_load_8888_hsw LABEL PROC
DB 196,225,249,110,192 ; vmovq %rax,%xmm0
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
- DB 235,135 ; jmp 2bce <_sk_load_8888_hsw+0x1a>
+ DB 235,135 ; jmp 324e <_sk_load_8888_hsw+0x1a>
PUBLIC _sk_gather_8888_hsw
_sk_gather_8888_hsw LABEL PROC
@@ -2751,14 +3102,14 @@ _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 197,229,219,5,107,23,0,0 ; vpand 0x176b(%rip),%ymm3,%ymm0 # 43e0 <_sk_callback_hsw+0x5f3>
+ DB 197,229,219,5,171,23,0,0 ; vpand 0x17ab(%rip),%ymm3,%ymm0 # 4aa0 <_sk_callback_hsw+0x633>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,66,21,0,0 ; vbroadcastss 0x1542(%rip),%ymm8 # 41c4 <_sk_callback_hsw+0x3d7>
+ DB 196,98,125,24,5,130,21,0,0 ; vbroadcastss 0x1582(%rip),%ymm8 # 4884 <_sk_callback_hsw+0x417>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,112,23,0,0 ; vpshufb 0x1770(%rip),%ymm3,%ymm1 # 4400 <_sk_callback_hsw+0x613>
+ DB 196,226,101,0,13,176,23,0,0 ; vpshufb 0x17b0(%rip),%ymm3,%ymm1 # 4ac0 <_sk_callback_hsw+0x653>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,126,23,0,0 ; vpshufb 0x177e(%rip),%ymm3,%ymm2 # 4420 <_sk_callback_hsw+0x633>
+ DB 196,226,101,0,21,190,23,0,0 ; vpshufb 0x17be(%rip),%ymm3,%ymm2 # 4ae0 <_sk_callback_hsw+0x673>
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
@@ -2773,7 +3124,7 @@ _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 196,98,125,24,5,242,20,0,0 ; vbroadcastss 0x14f2(%rip),%ymm8 # 41c8 <_sk_callback_hsw+0x3db>
+ DB 196,98,125,24,5,50,21,0,0 ; vbroadcastss 0x1532(%rip),%ymm8 # 4888 <_sk_callback_hsw+0x41b>
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
@@ -2789,7 +3140,7 @@ _sk_store_8888_hsw LABEL PROC
DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8
DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8
DB 77,133,192 ; test %r8,%r8
- DB 117,12 ; jne 2d30 <_sk_store_8888_hsw+0x73>
+ DB 117,12 ; jne 33b0 <_sk_store_8888_hsw+0x73>
DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,193 ; mov %r8,%rcx
@@ -2802,14 +3153,14 @@ _sk_store_8888_hsw LABEL PROC
DB 196,97,249,110,200 ; vmovq %rax,%xmm9
DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9
DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9)
- DB 235,211 ; jmp 2d29 <_sk_store_8888_hsw+0x6c>
+ DB 235,211 ; jmp 33a9 <_sk_store_8888_hsw+0x6c>
PUBLIC _sk_load_f16_hsw
_sk_load_f16_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 117,97 ; jne 2dc1 <_sk_load_f16_hsw+0x6b>
+ DB 117,97 ; jne 3441 <_sk_load_f16_hsw+0x6b>
DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8
DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2
DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3
@@ -2835,29 +3186,29 @@ _sk_load_f16_hsw LABEL PROC
DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,79 ; je 2e20 <_sk_load_f16_hsw+0xca>
+ DB 116,79 ; je 34a0 <_sk_load_f16_hsw+0xca>
DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,67 ; jb 2e20 <_sk_load_f16_hsw+0xca>
+ DB 114,67 ; jb 34a0 <_sk_load_f16_hsw+0xca>
DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,68 ; je 2e2d <_sk_load_f16_hsw+0xd7>
+ DB 116,68 ; je 34ad <_sk_load_f16_hsw+0xd7>
DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,56 ; jb 2e2d <_sk_load_f16_hsw+0xd7>
+ DB 114,56 ; jb 34ad <_sk_load_f16_hsw+0xd7>
DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,114,255,255,255 ; je 2d77 <_sk_load_f16_hsw+0x21>
+ DB 15,132,114,255,255,255 ; je 33f7 <_sk_load_f16_hsw+0x21>
DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,98,255,255,255 ; jb 2d77 <_sk_load_f16_hsw+0x21>
+ DB 15,130,98,255,255,255 ; jb 33f7 <_sk_load_f16_hsw+0x21>
DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9
- DB 233,87,255,255,255 ; jmpq 2d77 <_sk_load_f16_hsw+0x21>
+ DB 233,87,255,255,255 ; jmpq 33f7 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,74,255,255,255 ; jmpq 2d77 <_sk_load_f16_hsw+0x21>
+ DB 233,74,255,255,255 ; jmpq 33f7 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,65,255,255,255 ; jmpq 2d77 <_sk_load_f16_hsw+0x21>
+ DB 233,65,255,255,255 ; jmpq 33f7 <_sk_load_f16_hsw+0x21>
PUBLIC _sk_gather_f16_hsw
_sk_gather_f16_hsw LABEL PROC
@@ -2911,7 +3262,7 @@ _sk_store_f16_hsw LABEL PROC
DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9
DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,27 ; jne 2f25 <_sk_store_f16_hsw+0x65>
+ DB 117,27 ; jne 35a5 <_sk_store_f16_hsw+0x65>
DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8)
DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8)
DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -2920,22 +3271,22 @@ _sk_store_f16_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,241 ; je 2f21 <_sk_store_f16_hsw+0x61>
+ DB 116,241 ; je 35a1 <_sk_store_f16_hsw+0x61>
DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,229 ; jb 2f21 <_sk_store_f16_hsw+0x61>
+ DB 114,229 ; jb 35a1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,221 ; je 2f21 <_sk_store_f16_hsw+0x61>
+ DB 116,221 ; je 35a1 <_sk_store_f16_hsw+0x61>
DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,209 ; jb 2f21 <_sk_store_f16_hsw+0x61>
+ DB 114,209 ; jb 35a1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,201 ; je 2f21 <_sk_store_f16_hsw+0x61>
+ DB 116,201 ; je 35a1 <_sk_store_f16_hsw+0x61>
DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,189 ; jb 2f21 <_sk_store_f16_hsw+0x61>
+ DB 114,189 ; jb 35a1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,181 ; jmp 2f21 <_sk_store_f16_hsw+0x61>
+ DB 235,181 ; jmp 35a1 <_sk_store_f16_hsw+0x61>
PUBLIC _sk_load_u16_be_hsw
_sk_load_u16_be_hsw LABEL PROC
@@ -2943,7 +3294,7 @@ _sk_load_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,204,0,0,0 ; jne 304e <_sk_load_u16_be_hsw+0xe2>
+ DB 15,133,204,0,0,0 ; jne 36ce <_sk_load_u16_be_hsw+0xe2>
DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8
DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2
DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3
@@ -2962,7 +3313,7 @@ _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 196,98,125,24,21,233,17,0,0 ; vbroadcastss 0x11e9(%rip),%ymm10 # 41cc <_sk_callback_hsw+0x3df>
+ DB 196,98,125,24,21,41,18,0,0 ; vbroadcastss 0x1229(%rip),%ymm10 # 488c <_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
@@ -2990,29 +3341,29 @@ _sk_load_u16_be_hsw LABEL PROC
DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 30b4 <_sk_load_u16_be_hsw+0x148>
+ DB 116,85 ; je 3734 <_sk_load_u16_be_hsw+0x148>
DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 30b4 <_sk_load_u16_be_hsw+0x148>
+ DB 114,72 ; jb 3734 <_sk_load_u16_be_hsw+0x148>
DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 30c1 <_sk_load_u16_be_hsw+0x155>
+ DB 116,72 ; je 3741 <_sk_load_u16_be_hsw+0x155>
DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 30c1 <_sk_load_u16_be_hsw+0x155>
+ DB 114,59 ; jb 3741 <_sk_load_u16_be_hsw+0x155>
DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,6,255,255,255 ; je 2f9d <_sk_load_u16_be_hsw+0x31>
+ DB 15,132,6,255,255,255 ; je 361d <_sk_load_u16_be_hsw+0x31>
DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,245,254,255,255 ; jb 2f9d <_sk_load_u16_be_hsw+0x31>
+ DB 15,130,245,254,255,255 ; jb 361d <_sk_load_u16_be_hsw+0x31>
DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9
- DB 233,233,254,255,255 ; jmpq 2f9d <_sk_load_u16_be_hsw+0x31>
+ DB 233,233,254,255,255 ; jmpq 361d <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,220,254,255,255 ; jmpq 2f9d <_sk_load_u16_be_hsw+0x31>
+ DB 233,220,254,255,255 ; jmpq 361d <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,211,254,255,255 ; jmpq 2f9d <_sk_load_u16_be_hsw+0x31>
+ DB 233,211,254,255,255 ; jmpq 361d <_sk_load_u16_be_hsw+0x31>
PUBLIC _sk_load_rgb_u16_be_hsw
_sk_load_rgb_u16_be_hsw LABEL PROC
@@ -3020,7 +3371,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,204,0,0,0 ; jne 31a8 <_sk_load_rgb_u16_be_hsw+0xde>
+ DB 15,133,204,0,0,0 ; jne 3828 <_sk_load_rgb_u16_be_hsw+0xde>
DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0
DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2
DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1
@@ -3044,7 +3395,7 @@ _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 196,98,125,24,21,122,16,0,0 ; vbroadcastss 0x107a(%rip),%ymm10 # 41d0 <_sk_callback_hsw+0x3e3>
+ DB 196,98,125,24,21,186,16,0,0 ; vbroadcastss 0x10ba(%rip),%ymm10 # 4890 <_sk_callback_hsw+0x423>
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
@@ -3061,48 +3412,48 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,46,16,0,0 ; vbroadcastss 0x102e(%rip),%ymm3 # 41d4 <_sk_callback_hsw+0x3e7>
+ DB 196,226,125,24,29,110,16,0,0 ; vbroadcastss 0x106e(%rip),%ymm3 # 4894 <_sk_callback_hsw+0x427>
DB 255,224 ; jmpq *%rax
DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0
DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 31c1 <_sk_load_rgb_u16_be_hsw+0xf7>
- DB 233,79,255,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 3841 <_sk_load_rgb_u16_be_hsw+0xf7>
+ DB 233,79,255,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1
DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 31f0 <_sk_load_rgb_u16_be_hsw+0x126>
+ DB 114,26 ; jb 3870 <_sk_load_rgb_u16_be_hsw+0x126>
DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1
DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 31f5 <_sk_load_rgb_u16_be_hsw+0x12b>
- DB 233,32,255,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,27,255,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 3875 <_sk_load_rgb_u16_be_hsw+0x12b>
+ DB 233,32,255,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,27,255,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1
DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 3224 <_sk_load_rgb_u16_be_hsw+0x15a>
+ DB 114,26 ; jb 38a4 <_sk_load_rgb_u16_be_hsw+0x15a>
DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1
DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 3229 <_sk_load_rgb_u16_be_hsw+0x15f>
- DB 233,236,254,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,231,254,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 38a9 <_sk_load_rgb_u16_be_hsw+0x15f>
+ DB 233,236,254,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,231,254,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3
DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 3252 <_sk_load_rgb_u16_be_hsw+0x188>
+ DB 114,20 ; jb 38d2 <_sk_load_rgb_u16_be_hsw+0x188>
DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3
DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
- DB 233,190,254,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,185,254,255,255 ; jmpq 3110 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,190,254,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,185,254,255,255 ; jmpq 3790 <_sk_load_rgb_u16_be_hsw+0x46>
PUBLIC _sk_store_u16_be_hsw
_sk_store_u16_be_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax
- DB 196,98,125,24,5,107,15,0,0 ; vbroadcastss 0xf6b(%rip),%ymm8 # 41d8 <_sk_callback_hsw+0x3eb>
+ DB 196,98,125,24,5,171,15,0,0 ; vbroadcastss 0xfab(%rip),%ymm8 # 4898 <_sk_callback_hsw+0x42b>
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
@@ -3140,7 +3491,7 @@ _sk_store_u16_be_hsw LABEL PROC
DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9
DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,31 ; jne 3351 <_sk_store_u16_be_hsw+0xfa>
+ DB 117,31 ; jne 39d1 <_sk_store_u16_be_hsw+0xfa>
DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2)
DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2)
DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2)
@@ -3149,31 +3500,31 @@ _sk_store_u16_be_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 116,240 ; je 39cd <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 114,227 ; jb 39cd <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2)
- DB 116,218 ; je 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 116,218 ; je 39cd <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 114,205 ; jb 39cd <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2)
- DB 116,196 ; je 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 116,196 ; je 39cd <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,183 ; jb 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 114,183 ; jb 39cd <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2)
- DB 235,174 ; jmp 334d <_sk_store_u16_be_hsw+0xf6>
+ DB 235,174 ; jmp 39cd <_sk_store_u16_be_hsw+0xf6>
PUBLIC _sk_load_f32_hsw
_sk_load_f32_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 119,110 ; ja 3415 <_sk_load_f32_hsw+0x76>
+ DB 119,110 ; ja 3a95 <_sk_load_f32_hsw+0x76>
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
- DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 3440 <_sk_load_f32_hsw+0xa1>
+ DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 3ac0 <_sk_load_f32_hsw+0xa1>
DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -3232,7 +3583,7 @@ _sk_store_f32_hsw LABEL PROC
DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8
DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11
DB 72,133,201 ; test %rcx,%rcx
- DB 117,55 ; jne 34cd <_sk_store_f32_hsw+0x6d>
+ DB 117,55 ; jne 3b4d <_sk_store_f32_hsw+0x6d>
DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
@@ -3245,22 +3596,22 @@ _sk_store_f32_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 34c9 <_sk_store_f32_hsw+0x69>
+ DB 116,240 ; je 3b49 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 34c9 <_sk_store_f32_hsw+0x69>
+ DB 114,227 ; jb 3b49 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 34c9 <_sk_store_f32_hsw+0x69>
+ DB 116,218 ; je 3b49 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 34c9 <_sk_store_f32_hsw+0x69>
+ DB 114,205 ; jb 3b49 <_sk_store_f32_hsw+0x69>
DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
- DB 116,195 ; je 34c9 <_sk_store_f32_hsw+0x69>
+ DB 116,195 ; je 3b49 <_sk_store_f32_hsw+0x69>
DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,181 ; jb 34c9 <_sk_store_f32_hsw+0x69>
+ DB 114,181 ; jb 3b49 <_sk_store_f32_hsw+0x69>
DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
- DB 235,171 ; jmp 34c9 <_sk_store_f32_hsw+0x69>
+ DB 235,171 ; jmp 3b49 <_sk_store_f32_hsw+0x69>
PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
@@ -3356,11 +3707,11 @@ _sk_mirror_y_hsw LABEL PROC
PUBLIC _sk_luminance_to_alpha_hsw
_sk_luminance_to_alpha_hsw LABEL PROC
- DB 196,226,125,24,29,133,11,0,0 ; vbroadcastss 0xb85(%rip),%ymm3 # 41dc <_sk_callback_hsw+0x3ef>
- DB 196,98,125,24,5,128,11,0,0 ; vbroadcastss 0xb80(%rip),%ymm8 # 41e0 <_sk_callback_hsw+0x3f3>
+ DB 196,226,125,24,29,197,11,0,0 ; vbroadcastss 0xbc5(%rip),%ymm3 # 489c <_sk_callback_hsw+0x42f>
+ DB 196,98,125,24,5,192,11,0,0 ; vbroadcastss 0xbc0(%rip),%ymm8 # 48a0 <_sk_callback_hsw+0x433>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1
- DB 196,226,125,24,29,113,11,0,0 ; vbroadcastss 0xb71(%rip),%ymm3 # 41e4 <_sk_callback_hsw+0x3f7>
+ DB 196,226,125,24,29,177,11,0,0 ; vbroadcastss 0xbb1(%rip),%ymm3 # 48a4 <_sk_callback_hsw+0x437>
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
@@ -3495,7 +3846,7 @@ _sk_linear_gradient_hsw LABEL PROC
DB 196,98,125,24,72,28 ; vbroadcastss 0x1c(%rax),%ymm9
DB 76,139,0 ; mov (%rax),%r8
DB 77,133,192 ; test %r8,%r8
- DB 15,132,143,0,0,0 ; je 394b <_sk_linear_gradient_hsw+0xb5>
+ DB 15,132,143,0,0,0 ; je 3fcb <_sk_linear_gradient_hsw+0xb5>
DB 72,139,64,8 ; mov 0x8(%rax),%rax
DB 72,131,192,32 ; add $0x20,%rax
DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12
@@ -3522,8 +3873,8 @@ _sk_linear_gradient_hsw LABEL PROC
DB 196,67,13,74,201,208 ; vblendvps %ymm13,%ymm9,%ymm14,%ymm9
DB 72,131,192,36 ; add $0x24,%rax
DB 73,255,200 ; dec %r8
- DB 117,140 ; jne 38d5 <_sk_linear_gradient_hsw+0x3f>
- DB 235,17 ; jmp 395c <_sk_linear_gradient_hsw+0xc6>
+ DB 117,140 ; jne 3f55 <_sk_linear_gradient_hsw+0x3f>
+ DB 235,17 ; jmp 3fdc <_sk_linear_gradient_hsw+0xc6>
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2
DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3
@@ -3566,24 +3917,24 @@ _sk_xy_to_polar_unit_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 196,98,125,24,45,241,7,0,0 ; vbroadcastss 0x7f1(%rip),%ymm13 # 41e8 <_sk_callback_hsw+0x3fb>
- DB 196,98,125,24,53,236,7,0,0 ; vbroadcastss 0x7ec(%rip),%ymm14 # 41ec <_sk_callback_hsw+0x3ff>
+ DB 196,98,125,24,45,49,8,0,0 ; vbroadcastss 0x831(%rip),%ymm13 # 48a8 <_sk_callback_hsw+0x43b>
+ DB 196,98,125,24,53,44,8,0,0 ; vbroadcastss 0x82c(%rip),%ymm14 # 48ac <_sk_callback_hsw+0x43f>
DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14
- DB 196,98,125,24,45,226,7,0,0 ; vbroadcastss 0x7e2(%rip),%ymm13 # 41f0 <_sk_callback_hsw+0x403>
+ DB 196,98,125,24,45,34,8,0,0 ; vbroadcastss 0x822(%rip),%ymm13 # 48b0 <_sk_callback_hsw+0x443>
DB 196,66,29,184,238 ; vfmadd231ps %ymm14,%ymm12,%ymm13
- DB 196,98,125,24,53,216,7,0,0 ; vbroadcastss 0x7d8(%rip),%ymm14 # 41f4 <_sk_callback_hsw+0x407>
+ DB 196,98,125,24,53,24,8,0,0 ; vbroadcastss 0x818(%rip),%ymm14 # 48b4 <_sk_callback_hsw+0x447>
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 196,98,125,24,21,195,7,0,0 ; vbroadcastss 0x7c3(%rip),%ymm10 # 41f8 <_sk_callback_hsw+0x40b>
+ DB 196,98,125,24,21,3,8,0,0 ; vbroadcastss 0x803(%rip),%ymm10 # 48b8 <_sk_callback_hsw+0x44b>
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 196,98,125,24,21,173,7,0,0 ; vbroadcastss 0x7ad(%rip),%ymm10 # 41fc <_sk_callback_hsw+0x40f>
+ DB 196,98,125,24,21,237,7,0,0 ; vbroadcastss 0x7ed(%rip),%ymm10 # 48bc <_sk_callback_hsw+0x44f>
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 196,98,125,24,21,151,7,0,0 ; vbroadcastss 0x797(%rip),%ymm10 # 4200 <_sk_callback_hsw+0x413>
+ DB 196,98,125,24,21,215,7,0,0 ; vbroadcastss 0x7d7(%rip),%ymm10 # 48c0 <_sk_callback_hsw+0x453>
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
@@ -3594,7 +3945,7 @@ _sk_xy_to_polar_unit_hsw LABEL PROC
PUBLIC _sk_save_xy_hsw
_sk_save_xy_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,118,7,0,0 ; vbroadcastss 0x776(%rip),%ymm8 # 4204 <_sk_callback_hsw+0x417>
+ DB 196,98,125,24,5,182,7,0,0 ; vbroadcastss 0x7b6(%rip),%ymm8 # 48c4 <_sk_callback_hsw+0x457>
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
@@ -3624,9 +3975,9 @@ _sk_accumulate_hsw LABEL PROC
PUBLIC _sk_bilinear_nx_hsw
_sk_bilinear_nx_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,10,7,0,0 ; vbroadcastss 0x70a(%rip),%ymm0 # 4208 <_sk_callback_hsw+0x41b>
+ DB 196,226,125,24,5,74,7,0,0 ; vbroadcastss 0x74a(%rip),%ymm0 # 48c8 <_sk_callback_hsw+0x45b>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,1,7,0,0 ; vbroadcastss 0x701(%rip),%ymm8 # 420c <_sk_callback_hsw+0x41f>
+ DB 196,98,125,24,5,65,7,0,0 ; vbroadcastss 0x741(%rip),%ymm8 # 48cc <_sk_callback_hsw+0x45f>
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
@@ -3635,7 +3986,7 @@ _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 196,226,125,24,5,233,6,0,0 ; vbroadcastss 0x6e9(%rip),%ymm0 # 4210 <_sk_callback_hsw+0x423>
+ DB 196,226,125,24,5,41,7,0,0 ; vbroadcastss 0x729(%rip),%ymm0 # 48d0 <_sk_callback_hsw+0x463>
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)
@@ -3645,9 +3996,9 @@ _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 196,226,125,24,13,205,6,0,0 ; vbroadcastss 0x6cd(%rip),%ymm1 # 4214 <_sk_callback_hsw+0x427>
+ DB 196,226,125,24,13,13,7,0,0 ; vbroadcastss 0x70d(%rip),%ymm1 # 48d4 <_sk_callback_hsw+0x467>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,195,6,0,0 ; vbroadcastss 0x6c3(%rip),%ymm8 # 4218 <_sk_callback_hsw+0x42b>
+ DB 196,98,125,24,5,3,7,0,0 ; vbroadcastss 0x703(%rip),%ymm8 # 48d8 <_sk_callback_hsw+0x46b>
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
@@ -3656,7 +4007,7 @@ _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 196,226,125,24,13,171,6,0,0 ; vbroadcastss 0x6ab(%rip),%ymm1 # 421c <_sk_callback_hsw+0x42f>
+ DB 196,226,125,24,13,235,6,0,0 ; vbroadcastss 0x6eb(%rip),%ymm1 # 48dc <_sk_callback_hsw+0x46f>
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)
@@ -3666,13 +4017,13 @@ _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 196,226,125,24,5,142,6,0,0 ; vbroadcastss 0x68e(%rip),%ymm0 # 4220 <_sk_callback_hsw+0x433>
+ DB 196,226,125,24,5,206,6,0,0 ; vbroadcastss 0x6ce(%rip),%ymm0 # 48e0 <_sk_callback_hsw+0x473>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,133,6,0,0 ; vbroadcastss 0x685(%rip),%ymm8 # 4224 <_sk_callback_hsw+0x437>
+ DB 196,98,125,24,5,197,6,0,0 ; vbroadcastss 0x6c5(%rip),%ymm8 # 48e4 <_sk_callback_hsw+0x477>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,118,6,0,0 ; vbroadcastss 0x676(%rip),%ymm10 # 4228 <_sk_callback_hsw+0x43b>
- DB 196,98,125,24,29,113,6,0,0 ; vbroadcastss 0x671(%rip),%ymm11 # 422c <_sk_callback_hsw+0x43f>
+ DB 196,98,125,24,21,182,6,0,0 ; vbroadcastss 0x6b6(%rip),%ymm10 # 48e8 <_sk_callback_hsw+0x47b>
+ DB 196,98,125,24,29,177,6,0,0 ; vbroadcastss 0x6b1(%rip),%ymm11 # 48ec <_sk_callback_hsw+0x47f>
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)
@@ -3682,16 +4033,16 @@ _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 196,226,125,24,5,84,6,0,0 ; vbroadcastss 0x654(%rip),%ymm0 # 4230 <_sk_callback_hsw+0x443>
+ DB 196,226,125,24,5,148,6,0,0 ; vbroadcastss 0x694(%rip),%ymm0 # 48f0 <_sk_callback_hsw+0x483>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,75,6,0,0 ; vbroadcastss 0x64b(%rip),%ymm8 # 4234 <_sk_callback_hsw+0x447>
+ DB 196,98,125,24,5,139,6,0,0 ; vbroadcastss 0x68b(%rip),%ymm8 # 48f4 <_sk_callback_hsw+0x487>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,65,6,0,0 ; vbroadcastss 0x641(%rip),%ymm9 # 4238 <_sk_callback_hsw+0x44b>
- DB 196,98,125,24,21,60,6,0,0 ; vbroadcastss 0x63c(%rip),%ymm10 # 423c <_sk_callback_hsw+0x44f>
+ DB 196,98,125,24,13,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm9 # 48f8 <_sk_callback_hsw+0x48b>
+ DB 196,98,125,24,21,124,6,0,0 ; vbroadcastss 0x67c(%rip),%ymm10 # 48fc <_sk_callback_hsw+0x48f>
DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,13,50,6,0,0 ; vbroadcastss 0x632(%rip),%ymm9 # 4240 <_sk_callback_hsw+0x453>
+ DB 196,98,125,24,13,114,6,0,0 ; vbroadcastss 0x672(%rip),%ymm9 # 4900 <_sk_callback_hsw+0x493>
DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9
- DB 196,98,125,24,21,40,6,0,0 ; vbroadcastss 0x628(%rip),%ymm10 # 4244 <_sk_callback_hsw+0x457>
+ DB 196,98,125,24,21,104,6,0,0 ; vbroadcastss 0x668(%rip),%ymm10 # 4904 <_sk_callback_hsw+0x497>
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
@@ -3700,14 +4051,14 @@ _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 196,98,125,24,5,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm8 # 4248 <_sk_callback_hsw+0x45b>
+ DB 196,98,125,24,5,80,6,0,0 ; vbroadcastss 0x650(%rip),%ymm8 # 4908 <_sk_callback_hsw+0x49b>
DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0
DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9
- DB 196,98,125,24,21,2,6,0,0 ; vbroadcastss 0x602(%rip),%ymm10 # 424c <_sk_callback_hsw+0x45f>
- DB 196,98,125,24,29,253,5,0,0 ; vbroadcastss 0x5fd(%rip),%ymm11 # 4250 <_sk_callback_hsw+0x463>
+ DB 196,98,125,24,21,66,6,0,0 ; vbroadcastss 0x642(%rip),%ymm10 # 490c <_sk_callback_hsw+0x49f>
+ DB 196,98,125,24,29,61,6,0,0 ; vbroadcastss 0x63d(%rip),%ymm11 # 4910 <_sk_callback_hsw+0x4a3>
DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11
DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11
- DB 196,98,125,24,5,238,5,0,0 ; vbroadcastss 0x5ee(%rip),%ymm8 # 4254 <_sk_callback_hsw+0x467>
+ DB 196,98,125,24,5,46,6,0,0 ; vbroadcastss 0x62e(%rip),%ymm8 # 4914 <_sk_callback_hsw+0x4a7>
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
@@ -3716,12 +4067,12 @@ _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 196,226,125,24,5,214,5,0,0 ; vbroadcastss 0x5d6(%rip),%ymm0 # 4258 <_sk_callback_hsw+0x46b>
+ DB 196,226,125,24,5,22,6,0,0 ; vbroadcastss 0x616(%rip),%ymm0 # 4918 <_sk_callback_hsw+0x4ab>
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 196,98,125,24,21,195,5,0,0 ; vbroadcastss 0x5c3(%rip),%ymm10 # 425c <_sk_callback_hsw+0x46f>
- DB 196,98,125,24,29,190,5,0,0 ; vbroadcastss 0x5be(%rip),%ymm11 # 4260 <_sk_callback_hsw+0x473>
+ DB 196,98,125,24,21,3,6,0,0 ; vbroadcastss 0x603(%rip),%ymm10 # 491c <_sk_callback_hsw+0x4af>
+ DB 196,98,125,24,29,254,5,0,0 ; vbroadcastss 0x5fe(%rip),%ymm11 # 4920 <_sk_callback_hsw+0x4b3>
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)
@@ -3731,13 +4082,13 @@ _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 196,226,125,24,13,161,5,0,0 ; vbroadcastss 0x5a1(%rip),%ymm1 # 4264 <_sk_callback_hsw+0x477>
+ DB 196,226,125,24,13,225,5,0,0 ; vbroadcastss 0x5e1(%rip),%ymm1 # 4924 <_sk_callback_hsw+0x4b7>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,151,5,0,0 ; vbroadcastss 0x597(%rip),%ymm8 # 4268 <_sk_callback_hsw+0x47b>
+ DB 196,98,125,24,5,215,5,0,0 ; vbroadcastss 0x5d7(%rip),%ymm8 # 4928 <_sk_callback_hsw+0x4bb>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,136,5,0,0 ; vbroadcastss 0x588(%rip),%ymm10 # 426c <_sk_callback_hsw+0x47f>
- DB 196,98,125,24,29,131,5,0,0 ; vbroadcastss 0x583(%rip),%ymm11 # 4270 <_sk_callback_hsw+0x483>
+ DB 196,98,125,24,21,200,5,0,0 ; vbroadcastss 0x5c8(%rip),%ymm10 # 492c <_sk_callback_hsw+0x4bf>
+ DB 196,98,125,24,29,195,5,0,0 ; vbroadcastss 0x5c3(%rip),%ymm11 # 4930 <_sk_callback_hsw+0x4c3>
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)
@@ -3747,16 +4098,16 @@ _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 196,226,125,24,13,102,5,0,0 ; vbroadcastss 0x566(%rip),%ymm1 # 4274 <_sk_callback_hsw+0x487>
+ DB 196,226,125,24,13,166,5,0,0 ; vbroadcastss 0x5a6(%rip),%ymm1 # 4934 <_sk_callback_hsw+0x4c7>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,92,5,0,0 ; vbroadcastss 0x55c(%rip),%ymm8 # 4278 <_sk_callback_hsw+0x48b>
+ DB 196,98,125,24,5,156,5,0,0 ; vbroadcastss 0x59c(%rip),%ymm8 # 4938 <_sk_callback_hsw+0x4cb>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,82,5,0,0 ; vbroadcastss 0x552(%rip),%ymm9 # 427c <_sk_callback_hsw+0x48f>
- DB 196,98,125,24,21,77,5,0,0 ; vbroadcastss 0x54d(%rip),%ymm10 # 4280 <_sk_callback_hsw+0x493>
+ DB 196,98,125,24,13,146,5,0,0 ; vbroadcastss 0x592(%rip),%ymm9 # 493c <_sk_callback_hsw+0x4cf>
+ DB 196,98,125,24,21,141,5,0,0 ; vbroadcastss 0x58d(%rip),%ymm10 # 4940 <_sk_callback_hsw+0x4d3>
DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,13,67,5,0,0 ; vbroadcastss 0x543(%rip),%ymm9 # 4284 <_sk_callback_hsw+0x497>
+ DB 196,98,125,24,13,131,5,0,0 ; vbroadcastss 0x583(%rip),%ymm9 # 4944 <_sk_callback_hsw+0x4d7>
DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9
- DB 196,98,125,24,21,57,5,0,0 ; vbroadcastss 0x539(%rip),%ymm10 # 4288 <_sk_callback_hsw+0x49b>
+ DB 196,98,125,24,21,121,5,0,0 ; vbroadcastss 0x579(%rip),%ymm10 # 4948 <_sk_callback_hsw+0x4db>
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
@@ -3765,14 +4116,14 @@ _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 196,98,125,24,5,33,5,0,0 ; vbroadcastss 0x521(%rip),%ymm8 # 428c <_sk_callback_hsw+0x49f>
+ DB 196,98,125,24,5,97,5,0,0 ; vbroadcastss 0x561(%rip),%ymm8 # 494c <_sk_callback_hsw+0x4df>
DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1
DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9
- DB 196,98,125,24,21,18,5,0,0 ; vbroadcastss 0x512(%rip),%ymm10 # 4290 <_sk_callback_hsw+0x4a3>
- DB 196,98,125,24,29,13,5,0,0 ; vbroadcastss 0x50d(%rip),%ymm11 # 4294 <_sk_callback_hsw+0x4a7>
+ DB 196,98,125,24,21,82,5,0,0 ; vbroadcastss 0x552(%rip),%ymm10 # 4950 <_sk_callback_hsw+0x4e3>
+ DB 196,98,125,24,29,77,5,0,0 ; vbroadcastss 0x54d(%rip),%ymm11 # 4954 <_sk_callback_hsw+0x4e7>
DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11
DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11
- DB 196,98,125,24,5,254,4,0,0 ; vbroadcastss 0x4fe(%rip),%ymm8 # 4298 <_sk_callback_hsw+0x4ab>
+ DB 196,98,125,24,5,62,5,0,0 ; vbroadcastss 0x53e(%rip),%ymm8 # 4958 <_sk_callback_hsw+0x4eb>
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
@@ -3781,12 +4132,12 @@ _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 196,226,125,24,13,230,4,0,0 ; vbroadcastss 0x4e6(%rip),%ymm1 # 429c <_sk_callback_hsw+0x4af>
+ DB 196,226,125,24,13,38,5,0,0 ; vbroadcastss 0x526(%rip),%ymm1 # 495c <_sk_callback_hsw+0x4ef>
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 196,98,125,24,21,210,4,0,0 ; vbroadcastss 0x4d2(%rip),%ymm10 # 42a0 <_sk_callback_hsw+0x4b3>
- DB 196,98,125,24,29,205,4,0,0 ; vbroadcastss 0x4cd(%rip),%ymm11 # 42a4 <_sk_callback_hsw+0x4b7>
+ DB 196,98,125,24,21,18,5,0,0 ; vbroadcastss 0x512(%rip),%ymm10 # 4960 <_sk_callback_hsw+0x4f3>
+ DB 196,98,125,24,29,13,5,0,0 ; vbroadcastss 0x50d(%rip),%ymm11 # 4964 <_sk_callback_hsw+0x4f7>
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)
@@ -3897,6 +4248,30 @@ ALIGN 4
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,191,0,0,224 ; add %al,-0x1fffff41(%rax)
DB 64,0,0 ; 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 463d <.literal4+0xb5>
+ 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 464d <.literal4+0xc5>
+ 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 465d <.literal4+0xd5>
+ 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 466d <.literal4+0xe5>
+ 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)
@@ -3944,7 +4319,7 @@ 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 3ff9 <.literal4+0xf1>
+ DB 224,7 ; loopne 46b9 <.literal4+0x131>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -3958,10 +4333,10 @@ 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 4024 <.literal4+0x11c>
+ DB 127,0 ; jg 46e4 <.literal4+0x15c>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 409d <.literal4+0x195>
+ DB 119,115 ; ja 475d <.literal4+0x1d5>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -3975,10 +4350,10 @@ 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 4058 <.literal4+0x150>
+ DB 127,0 ; jg 4718 <.literal4+0x190>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 40d1 <.literal4+0x1c9>
+ DB 119,115 ; ja 4791 <.literal4+0x209>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -3992,10 +4367,10 @@ 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 408c <.literal4+0x184>
+ DB 127,0 ; jg 474c <.literal4+0x1c4>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4105 <.literal4+0x1fd>
+ DB 119,115 ; ja 47c5 <.literal4+0x23d>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4009,10 +4384,10 @@ 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 40c0 <.literal4+0x1b8>
+ DB 127,0 ; jg 4780 <.literal4+0x1f8>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4139 <.literal4+0x231>
+ DB 119,115 ; ja 47f9 <.literal4+0x271>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4025,7 +4400,7 @@ 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 4137 <.literal4+0x22f>
+ DB 127,67 ; jg 47f7 <.literal4+0x26f>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -4037,10 +4412,10 @@ ALIGN 4
DB 190,80,128,3,62 ; mov $0x3e038050,%esi
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4157 <.literal4+0x24f>
+ DB 118,63 ; jbe 4817 <.literal4+0x28f>
DB 246,64,83,63 ; testb $0x3f,0x53(%rax)
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 416b <.literal4+0x263>
+ DB 127,67 ; jg 482b <.literal4+0x2a3>
DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax)
DB 128,59,0 ; cmpb $0x0,(%rbx)
DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax)
@@ -4049,7 +4424,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 414d <.literal4+0x245>
+ DB 224,7 ; loopne 480d <.literal4+0x285>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -4061,7 +4436,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4169 <.literal4+0x261>
+ DB 224,7 ; loopne 4829 <.literal4+0x2a1>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -4072,7 +4447,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 41be <.literal4+0x2b6>
+ DB 124,66 ; jl 487e <.literal4+0x2f6>
DB 0,240 ; add %dh,%al
DB 0,0 ; add %al,(%rax)
DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax)
@@ -4090,9 +4465,9 @@ 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 4201 <.literal4+0x2f9>
+ DB 112,65 ; jo 48c1 <.literal4+0x339>
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 420f <.literal4+0x307>
+ DB 127,67 ; jg 48cf <.literal4+0x347>
DB 128,0,128 ; addb $0x80,(%rax)
DB 55 ; (bad)
DB 128,0,128 ; addb $0x80,(%rax)
@@ -4100,7 +4475,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 255 ; (bad)
- DB 127,71 ; jg 4223 <.literal4+0x31b>
+ DB 127,71 ; jg 48e3 <.literal4+0x35b>
DB 208 ; (bad)
DB 179,89 ; mov $0x59,%bl
DB 62,89 ; ds pop %rcx
@@ -4197,16 +4572,16 @@ 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) # a0042e8 <_sk_callback_hsw+0xa0004fb>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0049a8 <_sk_callback_hsw+0xa00053b>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120042f0 <_sk_callback_hsw+0x12000503>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120049b0 <_sk_callback_hsw+0x12000543>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0042f8 <_sk_callback_hsw+0x1a00050b>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0049b8 <_sk_callback_hsw+0x1a00054b>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004300 <_sk_callback_hsw+0x3000513>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30049c0 <_sk_callback_hsw+0x3000553>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4249,16 +4624,16 @@ 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) # a004348 <_sk_callback_hsw+0xa00055b>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004a08 <_sk_callback_hsw+0xa00059b>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004350 <_sk_callback_hsw+0x12000563>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004a10 <_sk_callback_hsw+0x120005a3>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004358 <_sk_callback_hsw+0x1a00056b>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004a18 <_sk_callback_hsw+0x1a0005ab>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004360 <_sk_callback_hsw+0x3000573>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004a20 <_sk_callback_hsw+0x30005b3>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4301,16 +4676,16 @@ 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) # a0043a8 <_sk_callback_hsw+0xa0005bb>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004a68 <_sk_callback_hsw+0xa0005fb>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120043b0 <_sk_callback_hsw+0x120005c3>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004a70 <_sk_callback_hsw+0x12000603>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0043b8 <_sk_callback_hsw+0x1a0005cb>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004a78 <_sk_callback_hsw+0x1a00060b>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30043c0 <_sk_callback_hsw+0x30005d3>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004a80 <_sk_callback_hsw+0x3000613>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4353,16 +4728,16 @@ 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) # a004408 <_sk_callback_hsw+0xa00061b>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004ac8 <_sk_callback_hsw+0xa00065b>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004410 <_sk_callback_hsw+0x12000623>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004ad0 <_sk_callback_hsw+0x12000663>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004418 <_sk_callback_hsw+0x1a00062b>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004ad8 <_sk_callback_hsw+0x1a00066b>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004420 <_sk_callback_hsw+0x3000633>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004ae0 <_sk_callback_hsw+0x3000673>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4504,14 +4879,14 @@ _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 196,226,125,24,13,240,84,0,0 ; vbroadcastss 0x54f0(%rip),%ymm1 # 5650 <_sk_callback_avx+0x119>
+ DB 196,226,125,24,13,224,91,0,0 ; vbroadcastss 0x5be0(%rip),%ymm1 # 5d40 <_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 196,226,125,24,21,212,84,0,0 ; vbroadcastss 0x54d4(%rip),%ymm2 # 5654 <_sk_callback_avx+0x11d>
+ DB 196,226,125,24,21,196,91,0,0 ; vbroadcastss 0x5bc4(%rip),%ymm2 # 5d44 <_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
@@ -4531,7 +4906,7 @@ _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 196,98,125,24,29,140,84,0,0 ; vbroadcastss 0x548c(%rip),%ymm11 # 5658 <_sk_callback_avx+0x121>
+ DB 196,98,125,24,29,124,91,0,0 ; vbroadcastss 0x5b7c(%rip),%ymm11 # 5d48 <_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
@@ -4542,8 +4917,8 @@ _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 196,98,125,24,37,77,84,0,0 ; vbroadcastss 0x544d(%rip),%ymm12 # 565c <_sk_callback_avx+0x125>
- DB 196,98,125,24,45,72,84,0,0 ; vbroadcastss 0x5448(%rip),%ymm13 # 5660 <_sk_callback_avx+0x129>
+ DB 196,98,125,24,37,61,91,0,0 ; vbroadcastss 0x5b3d(%rip),%ymm12 # 5d4c <_sk_callback_avx+0x125>
+ DB 196,98,125,24,45,56,91,0,0 ; vbroadcastss 0x5b38(%rip),%ymm13 # 5d50 <_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
@@ -4570,9 +4945,9 @@ _sk_dither_avx LABEL PROC
DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8
DB 196,65,60,86,194 ; vorps %ymm10,%ymm8,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,179,83,0,0 ; vbroadcastss 0x53b3(%rip),%ymm9 # 5664 <_sk_callback_avx+0x12d>
+ DB 196,98,125,24,13,163,90,0,0 ; vbroadcastss 0x5aa3(%rip),%ymm9 # 5d54 <_sk_callback_avx+0x12d>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,169,83,0,0 ; vbroadcastss 0x53a9(%rip),%ymm9 # 5668 <_sk_callback_avx+0x131>
+ DB 196,98,125,24,13,153,90,0,0 ; vbroadcastss 0x5a99(%rip),%ymm9 # 5d58 <_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
@@ -4604,7 +4979,7 @@ _sk_clear_avx LABEL PROC
PUBLIC _sk_srcatop_avx
_sk_srcatop_avx LABEL PROC
DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0
- DB 196,98,125,24,5,79,83,0,0 ; vbroadcastss 0x534f(%rip),%ymm8 # 566c <_sk_callback_avx+0x135>
+ DB 196,98,125,24,5,63,90,0,0 ; vbroadcastss 0x5a3f(%rip),%ymm8 # 5d5c <_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
@@ -4623,7 +4998,7 @@ _sk_srcatop_avx LABEL PROC
PUBLIC _sk_dstatop_avx
_sk_dstatop_avx LABEL PROC
DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8
- DB 196,98,125,24,13,17,83,0,0 ; vbroadcastss 0x5311(%rip),%ymm9 # 5670 <_sk_callback_avx+0x139>
+ DB 196,98,125,24,13,1,90,0,0 ; vbroadcastss 0x5a01(%rip),%ymm9 # 5d60 <_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
@@ -4659,7 +5034,7 @@ _sk_dstin_avx LABEL PROC
PUBLIC _sk_srcout_avx
_sk_srcout_avx LABEL PROC
- DB 196,98,125,24,5,176,82,0,0 ; vbroadcastss 0x52b0(%rip),%ymm8 # 5674 <_sk_callback_avx+0x13d>
+ DB 196,98,125,24,5,160,89,0,0 ; vbroadcastss 0x59a0(%rip),%ymm8 # 5d64 <_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
@@ -4670,7 +5045,7 @@ _sk_srcout_avx LABEL PROC
PUBLIC _sk_dstout_avx
_sk_dstout_avx LABEL PROC
- DB 196,226,125,24,5,147,82,0,0 ; vbroadcastss 0x5293(%rip),%ymm0 # 5678 <_sk_callback_avx+0x141>
+ DB 196,226,125,24,5,131,89,0,0 ; vbroadcastss 0x5983(%rip),%ymm0 # 5d68 <_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
@@ -4681,7 +5056,7 @@ _sk_dstout_avx LABEL PROC
PUBLIC _sk_srcover_avx
_sk_srcover_avx LABEL PROC
- DB 196,98,125,24,5,118,82,0,0 ; vbroadcastss 0x5276(%rip),%ymm8 # 567c <_sk_callback_avx+0x145>
+ DB 196,98,125,24,5,102,89,0,0 ; vbroadcastss 0x5966(%rip),%ymm8 # 5d6c <_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
@@ -4696,7 +5071,7 @@ _sk_srcover_avx LABEL PROC
PUBLIC _sk_dstover_avx
_sk_dstover_avx LABEL PROC
- DB 196,98,125,24,5,73,82,0,0 ; vbroadcastss 0x5249(%rip),%ymm8 # 5680 <_sk_callback_avx+0x149>
+ DB 196,98,125,24,5,57,89,0,0 ; vbroadcastss 0x5939(%rip),%ymm8 # 5d70 <_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
@@ -4720,7 +5095,7 @@ _sk_modulate_avx LABEL PROC
PUBLIC _sk_multiply_avx
_sk_multiply_avx LABEL PROC
- DB 196,98,125,24,5,8,82,0,0 ; vbroadcastss 0x5208(%rip),%ymm8 # 5684 <_sk_callback_avx+0x14d>
+ DB 196,98,125,24,5,248,88,0,0 ; vbroadcastss 0x58f8(%rip),%ymm8 # 5d74 <_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
@@ -4774,7 +5149,7 @@ _sk_screen_avx LABEL PROC
PUBLIC _sk_xor__avx
_sk_xor__avx LABEL PROC
- DB 196,98,125,24,5,87,81,0,0 ; vbroadcastss 0x5157(%rip),%ymm8 # 5688 <_sk_callback_avx+0x151>
+ DB 196,98,125,24,5,71,88,0,0 ; vbroadcastss 0x5847(%rip),%ymm8 # 5d78 <_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
@@ -4809,7 +5184,7 @@ _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 196,98,125,24,5,215,80,0,0 ; vbroadcastss 0x50d7(%rip),%ymm8 # 568c <_sk_callback_avx+0x155>
+ DB 196,98,125,24,5,199,87,0,0 ; vbroadcastss 0x57c7(%rip),%ymm8 # 5d7c <_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
@@ -4833,7 +5208,7 @@ _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 196,98,125,24,5,131,80,0,0 ; vbroadcastss 0x5083(%rip),%ymm8 # 5690 <_sk_callback_avx+0x159>
+ DB 196,98,125,24,5,115,87,0,0 ; vbroadcastss 0x5773(%rip),%ymm8 # 5d80 <_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
@@ -4860,7 +5235,7 @@ _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 196,98,125,24,5,35,80,0,0 ; vbroadcastss 0x5023(%rip),%ymm8 # 5694 <_sk_callback_avx+0x15d>
+ DB 196,98,125,24,5,19,87,0,0 ; vbroadcastss 0x5713(%rip),%ymm8 # 5d84 <_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
@@ -4881,7 +5256,7 @@ _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 196,98,125,24,5,222,79,0,0 ; vbroadcastss 0x4fde(%rip),%ymm8 # 5698 <_sk_callback_avx+0x161>
+ DB 196,98,125,24,5,206,86,0,0 ; vbroadcastss 0x56ce(%rip),%ymm8 # 5d88 <_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
@@ -4890,7 +5265,7 @@ _sk_exclusion_avx LABEL PROC
PUBLIC _sk_colorburn_avx
_sk_colorburn_avx LABEL PROC
- DB 196,98,125,24,5,201,79,0,0 ; vbroadcastss 0x4fc9(%rip),%ymm8 # 569c <_sk_callback_avx+0x165>
+ DB 196,98,125,24,5,185,86,0,0 ; vbroadcastss 0x56b9(%rip),%ymm8 # 5d8c <_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
@@ -4950,7 +5325,7 @@ _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 196,98,125,24,13,197,78,0,0 ; vbroadcastss 0x4ec5(%rip),%ymm9 # 56a0 <_sk_callback_avx+0x169>
+ DB 196,98,125,24,13,181,85,0,0 ; vbroadcastss 0x55b5(%rip),%ymm9 # 5d90 <_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
@@ -5005,7 +5380,7 @@ _sk_colordodge_avx LABEL PROC
PUBLIC _sk_hardlight_avx
_sk_hardlight_avx LABEL PROC
- DB 196,98,125,24,5,215,77,0,0 ; vbroadcastss 0x4dd7(%rip),%ymm8 # 56a4 <_sk_callback_avx+0x16d>
+ DB 196,98,125,24,5,199,84,0,0 ; vbroadcastss 0x54c7(%rip),%ymm8 # 5d94 <_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
@@ -5058,7 +5433,7 @@ _sk_hardlight_avx LABEL PROC
PUBLIC _sk_overlay_avx
_sk_overlay_avx LABEL PROC
- DB 196,98,125,24,5,0,77,0,0 ; vbroadcastss 0x4d00(%rip),%ymm8 # 56a8 <_sk_callback_avx+0x171>
+ DB 196,98,125,24,5,240,83,0,0 ; vbroadcastss 0x53f0(%rip),%ymm8 # 5d98 <_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
@@ -5123,10 +5498,10 @@ _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 196,98,125,24,29,243,75,0,0 ; vbroadcastss 0x4bf3(%rip),%ymm11 # 56b0 <_sk_callback_avx+0x179>
+ DB 196,98,125,24,29,227,82,0,0 ; vbroadcastss 0x52e3(%rip),%ymm11 # 5da0 <_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 196,98,125,24,45,228,75,0,0 ; vbroadcastss 0x4be4(%rip),%ymm13 # 56b4 <_sk_callback_avx+0x17d>
+ DB 196,98,125,24,45,212,82,0,0 ; vbroadcastss 0x52d4(%rip),%ymm13 # 5da4 <_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
@@ -5137,7 +5512,7 @@ _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 196,98,125,24,5,162,75,0,0 ; vbroadcastss 0x4ba2(%rip),%ymm8 # 56ac <_sk_callback_avx+0x175>
+ DB 196,98,125,24,5,146,82,0,0 ; vbroadcastss 0x5292(%rip),%ymm8 # 5d9c <_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
@@ -5229,6 +5604,384 @@ _sk_softlight_avx LABEL PROC
DB 72,131,196,88 ; add $0x58,%rsp
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_hue_avx
+_sk_hue_avx LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,100,194,208,0 ; vcmpeqps %ymm8,%ymm3,%ymm10
+ DB 196,98,125,24,13,242,80,0,0 ; vbroadcastss 0x50f2(%rip),%ymm9 # 5da8 <_sk_callback_avx+0x181>
+ DB 197,52,94,219 ; vdivps %ymm3,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
+ DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
+ DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
+ DB 197,84,95,214 ; vmaxps %ymm6,%ymm5,%ymm10
+ DB 196,65,92,95,210 ; vmaxps %ymm10,%ymm4,%ymm10
+ DB 197,84,93,222 ; vminps %ymm6,%ymm5,%ymm11
+ DB 196,65,92,93,219 ; vminps %ymm11,%ymm4,%ymm11
+ DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10
+ DB 197,116,93,218 ; vminps %ymm2,%ymm1,%ymm11
+ DB 196,65,124,93,219 ; vminps %ymm11,%ymm0,%ymm11
+ DB 197,116,95,226 ; vmaxps %ymm2,%ymm1,%ymm12
+ DB 196,65,124,95,228 ; vmaxps %ymm12,%ymm0,%ymm12
+ DB 196,65,28,92,227 ; vsubps %ymm11,%ymm12,%ymm12
+ DB 196,65,28,194,232,0 ; vcmpeqps %ymm8,%ymm12,%ymm13
+ DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0
+ DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
+ DB 196,193,124,94,196 ; vdivps %ymm12,%ymm0,%ymm0
+ DB 196,195,125,74,192,208 ; vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1
+ DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
+ DB 196,193,116,94,204 ; vdivps %ymm12,%ymm1,%ymm1
+ DB 196,195,117,74,200,208 ; vblendvps %ymm13,%ymm8,%ymm1,%ymm1
+ DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2
+ DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
+ DB 196,193,108,94,212 ; vdivps %ymm12,%ymm2,%ymm2
+ DB 196,195,109,74,208,208 ; vblendvps %ymm13,%ymm8,%ymm2,%ymm2
+ DB 196,98,125,24,21,103,80,0,0 ; vbroadcastss 0x5067(%rip),%ymm10 # 5dac <_sk_callback_avx+0x185>
+ DB 196,65,92,89,218 ; vmulps %ymm10,%ymm4,%ymm11
+ DB 196,98,125,24,37,93,80,0,0 ; vbroadcastss 0x505d(%rip),%ymm12 # 5db0 <_sk_callback_avx+0x189>
+ DB 196,65,84,89,236 ; vmulps %ymm12,%ymm5,%ymm13
+ DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11
+ DB 196,98,125,24,45,78,80,0,0 ; vbroadcastss 0x504e(%rip),%ymm13 # 5db4 <_sk_callback_avx+0x18d>
+ DB 196,65,76,89,245 ; vmulps %ymm13,%ymm6,%ymm14
+ DB 196,65,36,88,222 ; vaddps %ymm14,%ymm11,%ymm11
+ DB 196,65,124,89,242 ; vmulps %ymm10,%ymm0,%ymm14
+ DB 196,65,116,89,252 ; vmulps %ymm12,%ymm1,%ymm15
+ DB 196,65,12,88,247 ; vaddps %ymm15,%ymm14,%ymm14
+ DB 196,65,108,89,253 ; vmulps %ymm13,%ymm2,%ymm15
+ DB 196,65,12,88,247 ; vaddps %ymm15,%ymm14,%ymm14
+ DB 196,65,36,92,222 ; vsubps %ymm14,%ymm11,%ymm11
+ DB 196,65,124,88,243 ; vaddps %ymm11,%ymm0,%ymm14
+ DB 196,65,116,88,251 ; vaddps %ymm11,%ymm1,%ymm15
+ DB 196,193,108,88,195 ; vaddps %ymm11,%ymm2,%ymm0
+ DB 197,132,93,200 ; vminps %ymm0,%ymm15,%ymm1
+ DB 197,140,93,209 ; vminps %ymm1,%ymm14,%ymm2
+ DB 197,132,95,200 ; vmaxps %ymm0,%ymm15,%ymm1
+ DB 197,12,95,217 ; vmaxps %ymm1,%ymm14,%ymm11
+ DB 196,193,12,89,202 ; vmulps %ymm10,%ymm14,%ymm1
+ DB 196,65,4,89,212 ; vmulps %ymm12,%ymm15,%ymm10
+ DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
+ DB 196,65,124,89,213 ; vmulps %ymm13,%ymm0,%ymm10
+ DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1
+ DB 197,60,194,210,2 ; vcmpleps %ymm2,%ymm8,%ymm10
+ DB 197,12,92,225 ; vsubps %ymm1,%ymm14,%ymm12
+ DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12
+ DB 197,244,92,210 ; vsubps %ymm2,%ymm1,%ymm2
+ DB 197,28,94,226 ; vdivps %ymm2,%ymm12,%ymm12
+ DB 196,65,116,88,228 ; vaddps %ymm12,%ymm1,%ymm12
+ DB 196,67,29,74,230,160 ; vblendvps %ymm10,%ymm14,%ymm12,%ymm12
+ DB 196,65,52,194,235,1 ; vcmpltps %ymm11,%ymm9,%ymm13
+ DB 197,28,92,241 ; vsubps %ymm1,%ymm12,%ymm14
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,246 ; vmulps %ymm14,%ymm9,%ymm14
+ DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11
+ DB 196,65,12,94,243 ; vdivps %ymm11,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,29,74,230,208 ; vblendvps %ymm13,%ymm14,%ymm12,%ymm12
+ DB 196,65,28,95,224 ; vmaxps %ymm8,%ymm12,%ymm12
+ DB 197,4,92,241 ; vsubps %ymm1,%ymm15,%ymm14
+ DB 196,65,116,89,246 ; vmulps %ymm14,%ymm1,%ymm14
+ DB 197,12,94,242 ; vdivps %ymm2,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,13,74,247,160 ; vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ DB 197,12,92,249 ; vsubps %ymm1,%ymm14,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,251 ; vdivps %ymm11,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,13,74,247,208 ; vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ DB 196,65,12,95,240 ; vmaxps %ymm8,%ymm14,%ymm14
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 197,132,94,210 ; vdivps %ymm2,%ymm15,%ymm2
+ DB 197,244,88,210 ; vaddps %ymm2,%ymm1,%ymm2
+ DB 196,227,109,74,192,160 ; vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,94,211 ; vdivps %ymm11,%ymm2,%ymm2
+ DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1
+ DB 196,227,125,74,193,208 ; vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ DB 196,193,124,95,208 ; vmaxps %ymm8,%ymm0,%ymm2
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 197,228,89,207 ; vmulps %ymm7,%ymm3,%ymm1
+ DB 197,252,92,217 ; vsubps %ymm1,%ymm0,%ymm3
+ DB 196,193,100,89,196 ; vmulps %ymm12,%ymm3,%ymm0
+ DB 196,193,100,89,206 ; vmulps %ymm14,%ymm3,%ymm1
+ DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_saturation_avx
+_sk_saturation_avx LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,68,194,208,0 ; vcmpeqps %ymm8,%ymm7,%ymm10
+ DB 196,98,125,24,13,11,79,0,0 ; vbroadcastss 0x4f0b(%rip),%ymm9 # 5db8 <_sk_callback_avx+0x191>
+ DB 197,52,94,223 ; vdivps %ymm7,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,44,89,220 ; vmulps %ymm4,%ymm10,%ymm11
+ DB 197,44,89,229 ; vmulps %ymm5,%ymm10,%ymm12
+ DB 197,44,89,214 ; vmulps %ymm6,%ymm10,%ymm10
+ DB 197,116,95,234 ; vmaxps %ymm2,%ymm1,%ymm13
+ DB 196,65,124,95,237 ; vmaxps %ymm13,%ymm0,%ymm13
+ DB 197,244,93,202 ; vminps %ymm2,%ymm1,%ymm1
+ DB 197,252,93,193 ; vminps %ymm1,%ymm0,%ymm0
+ DB 197,148,92,192 ; vsubps %ymm0,%ymm13,%ymm0
+ DB 196,193,28,93,202 ; vminps %ymm10,%ymm12,%ymm1
+ DB 197,164,93,201 ; vminps %ymm1,%ymm11,%ymm1
+ DB 196,193,28,95,210 ; vmaxps %ymm10,%ymm12,%ymm2
+ DB 197,164,95,210 ; vmaxps %ymm2,%ymm11,%ymm2
+ DB 197,236,92,209 ; vsubps %ymm1,%ymm2,%ymm2
+ DB 196,65,108,194,232,0 ; vcmpeqps %ymm8,%ymm2,%ymm13
+ DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11
+ DB 196,65,124,89,219 ; vmulps %ymm11,%ymm0,%ymm11
+ DB 197,36,94,218 ; vdivps %ymm2,%ymm11,%ymm11
+ DB 196,67,37,74,216,208 ; vblendvps %ymm13,%ymm8,%ymm11,%ymm11
+ DB 197,28,92,225 ; vsubps %ymm1,%ymm12,%ymm12
+ DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12
+ DB 197,28,94,226 ; vdivps %ymm2,%ymm12,%ymm12
+ DB 196,67,29,74,224,208 ; vblendvps %ymm13,%ymm8,%ymm12,%ymm12
+ DB 197,172,92,201 ; vsubps %ymm1,%ymm10,%ymm1
+ DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
+ DB 197,252,94,194 ; vdivps %ymm2,%ymm0,%ymm0
+ DB 196,195,125,74,192,208 ; vblendvps %ymm13,%ymm8,%ymm0,%ymm0
+ DB 196,226,125,24,13,135,78,0,0 ; vbroadcastss 0x4e87(%rip),%ymm1 # 5dbc <_sk_callback_avx+0x195>
+ DB 197,220,89,209 ; vmulps %ymm1,%ymm4,%ymm2
+ DB 196,98,125,24,21,126,78,0,0 ; vbroadcastss 0x4e7e(%rip),%ymm10 # 5dc0 <_sk_callback_avx+0x199>
+ DB 196,65,84,89,234 ; vmulps %ymm10,%ymm5,%ymm13
+ DB 196,193,108,88,213 ; vaddps %ymm13,%ymm2,%ymm2
+ DB 196,98,125,24,45,111,78,0,0 ; vbroadcastss 0x4e6f(%rip),%ymm13 # 5dc4 <_sk_callback_avx+0x19d>
+ DB 196,65,76,89,245 ; vmulps %ymm13,%ymm6,%ymm14
+ DB 196,193,108,88,214 ; vaddps %ymm14,%ymm2,%ymm2
+ DB 197,36,89,241 ; vmulps %ymm1,%ymm11,%ymm14
+ DB 196,65,28,89,250 ; vmulps %ymm10,%ymm12,%ymm15
+ DB 196,65,12,88,247 ; vaddps %ymm15,%ymm14,%ymm14
+ DB 196,65,124,89,253 ; vmulps %ymm13,%ymm0,%ymm15
+ DB 196,65,12,88,247 ; vaddps %ymm15,%ymm14,%ymm14
+ DB 196,193,108,92,214 ; vsubps %ymm14,%ymm2,%ymm2
+ DB 197,36,88,218 ; vaddps %ymm2,%ymm11,%ymm11
+ DB 197,28,88,226 ; vaddps %ymm2,%ymm12,%ymm12
+ DB 197,252,88,194 ; vaddps %ymm2,%ymm0,%ymm0
+ DB 197,156,93,208 ; vminps %ymm0,%ymm12,%ymm2
+ DB 197,164,93,210 ; vminps %ymm2,%ymm11,%ymm2
+ DB 197,28,95,240 ; vmaxps %ymm0,%ymm12,%ymm14
+ DB 196,65,36,95,246 ; vmaxps %ymm14,%ymm11,%ymm14
+ DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1
+ DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10
+ DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
+ DB 196,65,124,89,213 ; vmulps %ymm13,%ymm0,%ymm10
+ DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1
+ DB 197,60,194,210,2 ; vcmpleps %ymm2,%ymm8,%ymm10
+ DB 197,36,92,233 ; vsubps %ymm1,%ymm11,%ymm13
+ DB 196,65,116,89,237 ; vmulps %ymm13,%ymm1,%ymm13
+ DB 197,244,92,210 ; vsubps %ymm2,%ymm1,%ymm2
+ DB 197,20,94,234 ; vdivps %ymm2,%ymm13,%ymm13
+ DB 196,65,116,88,237 ; vaddps %ymm13,%ymm1,%ymm13
+ DB 196,67,21,74,219,160 ; vblendvps %ymm10,%ymm11,%ymm13,%ymm11
+ DB 196,65,52,194,238,1 ; vcmpltps %ymm14,%ymm9,%ymm13
+ DB 197,36,92,249 ; vsubps %ymm1,%ymm11,%ymm15
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 197,12,92,241 ; vsubps %ymm1,%ymm14,%ymm14
+ DB 196,65,4,94,254 ; vdivps %ymm14,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,37,74,223,208 ; vblendvps %ymm13,%ymm15,%ymm11,%ymm11
+ DB 196,65,36,95,216 ; vmaxps %ymm8,%ymm11,%ymm11
+ DB 197,28,92,249 ; vsubps %ymm1,%ymm12,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 197,4,94,250 ; vdivps %ymm2,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,5,74,228,160 ; vblendvps %ymm10,%ymm12,%ymm15,%ymm12
+ DB 197,28,92,249 ; vsubps %ymm1,%ymm12,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,254 ; vdivps %ymm14,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,29,74,231,208 ; vblendvps %ymm13,%ymm15,%ymm12,%ymm12
+ DB 196,65,28,95,224 ; vmaxps %ymm8,%ymm12,%ymm12
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 197,132,94,210 ; vdivps %ymm2,%ymm15,%ymm2
+ DB 197,244,88,210 ; vaddps %ymm2,%ymm1,%ymm2
+ DB 196,227,109,74,192,160 ; vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,94,214 ; vdivps %ymm14,%ymm2,%ymm2
+ DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1
+ DB 196,227,125,74,193,208 ; vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ DB 196,193,124,95,208 ; vmaxps %ymm8,%ymm0,%ymm2
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 197,228,89,207 ; vmulps %ymm7,%ymm3,%ymm1
+ DB 197,252,92,217 ; vsubps %ymm1,%ymm0,%ymm3
+ DB 196,193,100,89,195 ; vmulps %ymm11,%ymm3,%ymm0
+ DB 196,193,100,89,204 ; vmulps %ymm12,%ymm3,%ymm1
+ DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_color_avx
+_sk_color_avx LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,100,194,208,0 ; vcmpeqps %ymm8,%ymm3,%ymm10
+ DB 196,98,125,24,13,48,77,0,0 ; vbroadcastss 0x4d30(%rip),%ymm9 # 5dc8 <_sk_callback_avx+0x1a1>
+ DB 197,52,94,219 ; vdivps %ymm3,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
+ DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
+ DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
+ DB 196,98,125,24,21,21,77,0,0 ; vbroadcastss 0x4d15(%rip),%ymm10 # 5dcc <_sk_callback_avx+0x1a5>
+ DB 196,65,92,89,218 ; vmulps %ymm10,%ymm4,%ymm11
+ DB 196,98,125,24,37,11,77,0,0 ; vbroadcastss 0x4d0b(%rip),%ymm12 # 5dd0 <_sk_callback_avx+0x1a9>
+ DB 196,65,84,89,236 ; vmulps %ymm12,%ymm5,%ymm13
+ DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11
+ DB 196,98,125,24,45,252,76,0,0 ; vbroadcastss 0x4cfc(%rip),%ymm13 # 5dd4 <_sk_callback_avx+0x1ad>
+ DB 196,65,76,89,245 ; vmulps %ymm13,%ymm6,%ymm14
+ DB 196,65,36,88,222 ; vaddps %ymm14,%ymm11,%ymm11
+ DB 196,65,124,89,242 ; vmulps %ymm10,%ymm0,%ymm14
+ DB 196,65,116,89,252 ; vmulps %ymm12,%ymm1,%ymm15
+ DB 196,65,12,88,247 ; vaddps %ymm15,%ymm14,%ymm14
+ DB 196,65,108,89,253 ; vmulps %ymm13,%ymm2,%ymm15
+ DB 196,65,4,88,246 ; vaddps %ymm14,%ymm15,%ymm14
+ DB 196,65,36,92,222 ; vsubps %ymm14,%ymm11,%ymm11
+ DB 196,65,124,88,243 ; vaddps %ymm11,%ymm0,%ymm14
+ DB 196,65,116,88,251 ; vaddps %ymm11,%ymm1,%ymm15
+ DB 196,193,108,88,195 ; vaddps %ymm11,%ymm2,%ymm0
+ DB 197,132,93,200 ; vminps %ymm0,%ymm15,%ymm1
+ DB 197,140,93,209 ; vminps %ymm1,%ymm14,%ymm2
+ DB 197,132,95,200 ; vmaxps %ymm0,%ymm15,%ymm1
+ DB 197,12,95,217 ; vmaxps %ymm1,%ymm14,%ymm11
+ DB 196,193,12,89,202 ; vmulps %ymm10,%ymm14,%ymm1
+ DB 196,65,4,89,212 ; vmulps %ymm12,%ymm15,%ymm10
+ DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
+ DB 196,65,124,89,213 ; vmulps %ymm13,%ymm0,%ymm10
+ DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1
+ DB 197,60,194,210,2 ; vcmpleps %ymm2,%ymm8,%ymm10
+ DB 197,12,92,225 ; vsubps %ymm1,%ymm14,%ymm12
+ DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12
+ DB 197,244,92,210 ; vsubps %ymm2,%ymm1,%ymm2
+ DB 197,28,94,226 ; vdivps %ymm2,%ymm12,%ymm12
+ DB 196,65,116,88,228 ; vaddps %ymm12,%ymm1,%ymm12
+ DB 196,67,29,74,230,160 ; vblendvps %ymm10,%ymm14,%ymm12,%ymm12
+ DB 196,65,52,194,235,1 ; vcmpltps %ymm11,%ymm9,%ymm13
+ DB 197,28,92,241 ; vsubps %ymm1,%ymm12,%ymm14
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,246 ; vmulps %ymm14,%ymm9,%ymm14
+ DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11
+ DB 196,65,12,94,243 ; vdivps %ymm11,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,29,74,230,208 ; vblendvps %ymm13,%ymm14,%ymm12,%ymm12
+ DB 196,65,28,95,224 ; vmaxps %ymm8,%ymm12,%ymm12
+ DB 197,4,92,241 ; vsubps %ymm1,%ymm15,%ymm14
+ DB 196,65,116,89,246 ; vmulps %ymm14,%ymm1,%ymm14
+ DB 197,12,94,242 ; vdivps %ymm2,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,67,13,74,247,160 ; vblendvps %ymm10,%ymm15,%ymm14,%ymm14
+ DB 197,12,92,249 ; vsubps %ymm1,%ymm14,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,251 ; vdivps %ymm11,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,13,74,247,208 ; vblendvps %ymm13,%ymm15,%ymm14,%ymm14
+ DB 196,65,12,95,240 ; vmaxps %ymm8,%ymm14,%ymm14
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 197,132,94,210 ; vdivps %ymm2,%ymm15,%ymm2
+ DB 197,244,88,210 ; vaddps %ymm2,%ymm1,%ymm2
+ DB 196,227,109,74,192,160 ; vblendvps %ymm10,%ymm0,%ymm2,%ymm0
+ DB 197,252,92,209 ; vsubps %ymm1,%ymm0,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,94,211 ; vdivps %ymm11,%ymm2,%ymm2
+ DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1
+ DB 196,227,125,74,193,208 ; vblendvps %ymm13,%ymm1,%ymm0,%ymm0
+ DB 196,193,124,95,208 ; vmaxps %ymm8,%ymm0,%ymm2
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 197,228,89,207 ; vmulps %ymm7,%ymm3,%ymm1
+ DB 197,252,92,217 ; vsubps %ymm1,%ymm0,%ymm3
+ DB 196,193,100,89,196 ; vmulps %ymm12,%ymm3,%ymm0
+ DB 196,193,100,89,206 ; vmulps %ymm14,%ymm3,%ymm1
+ DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_luminosity_avx
+_sk_luminosity_avx LABEL PROC
+ DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
+ DB 196,65,68,194,208,0 ; vcmpeqps %ymm8,%ymm7,%ymm10
+ DB 196,98,125,24,13,185,75,0,0 ; vbroadcastss 0x4bb9(%rip),%ymm9 # 5dd8 <_sk_callback_avx+0x1b1>
+ DB 197,52,94,223 ; vdivps %ymm7,%ymm9,%ymm11
+ DB 196,67,37,74,208,160 ; vblendvps %ymm10,%ymm8,%ymm11,%ymm10
+ DB 197,44,89,220 ; vmulps %ymm4,%ymm10,%ymm11
+ DB 197,44,89,229 ; vmulps %ymm5,%ymm10,%ymm12
+ DB 197,44,89,214 ; vmulps %ymm6,%ymm10,%ymm10
+ DB 196,98,125,24,45,158,75,0,0 ; vbroadcastss 0x4b9e(%rip),%ymm13 # 5ddc <_sk_callback_avx+0x1b5>
+ DB 196,193,124,89,197 ; vmulps %ymm13,%ymm0,%ymm0
+ DB 196,98,125,24,53,148,75,0,0 ; vbroadcastss 0x4b94(%rip),%ymm14 # 5de0 <_sk_callback_avx+0x1b9>
+ DB 196,193,116,89,206 ; vmulps %ymm14,%ymm1,%ymm1
+ DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
+ DB 196,226,125,24,13,134,75,0,0 ; vbroadcastss 0x4b86(%rip),%ymm1 # 5de4 <_sk_callback_avx+0x1bd>
+ DB 197,236,89,209 ; vmulps %ymm1,%ymm2,%ymm2
+ DB 197,252,88,194 ; vaddps %ymm2,%ymm0,%ymm0
+ DB 196,193,36,89,213 ; vmulps %ymm13,%ymm11,%ymm2
+ DB 196,65,28,89,254 ; vmulps %ymm14,%ymm12,%ymm15
+ DB 196,193,108,88,215 ; vaddps %ymm15,%ymm2,%ymm2
+ DB 197,44,89,249 ; vmulps %ymm1,%ymm10,%ymm15
+ DB 197,132,88,210 ; vaddps %ymm2,%ymm15,%ymm2
+ DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0
+ DB 197,164,88,208 ; vaddps %ymm0,%ymm11,%ymm2
+ DB 197,28,88,216 ; vaddps %ymm0,%ymm12,%ymm11
+ DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0
+ DB 197,36,93,208 ; vminps %ymm0,%ymm11,%ymm10
+ DB 196,65,108,93,210 ; vminps %ymm10,%ymm2,%ymm10
+ DB 197,36,95,224 ; vmaxps %ymm0,%ymm11,%ymm12
+ DB 196,65,108,95,228 ; vmaxps %ymm12,%ymm2,%ymm12
+ DB 196,65,108,89,237 ; vmulps %ymm13,%ymm2,%ymm13
+ DB 196,65,36,89,246 ; vmulps %ymm14,%ymm11,%ymm14
+ DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13
+ DB 197,252,89,201 ; vmulps %ymm1,%ymm0,%ymm1
+ DB 196,193,116,88,205 ; vaddps %ymm13,%ymm1,%ymm1
+ DB 196,65,60,194,234,2 ; vcmpleps %ymm10,%ymm8,%ymm13
+ DB 197,108,92,241 ; vsubps %ymm1,%ymm2,%ymm14
+ DB 196,65,116,89,246 ; vmulps %ymm14,%ymm1,%ymm14
+ DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
+ DB 196,65,12,94,242 ; vdivps %ymm10,%ymm14,%ymm14
+ DB 196,65,116,88,246 ; vaddps %ymm14,%ymm1,%ymm14
+ DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2
+ DB 196,65,52,194,244,1 ; vcmpltps %ymm12,%ymm9,%ymm14
+ DB 197,108,92,249 ; vsubps %ymm1,%ymm2,%ymm15
+ DB 197,52,92,201 ; vsubps %ymm1,%ymm9,%ymm9
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 197,28,92,225 ; vsubps %ymm1,%ymm12,%ymm12
+ DB 196,65,4,94,252 ; vdivps %ymm12,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,195,109,74,215,224 ; vblendvps %ymm14,%ymm15,%ymm2,%ymm2
+ DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2
+ DB 197,36,92,249 ; vsubps %ymm1,%ymm11,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 196,65,4,94,250 ; vdivps %ymm10,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,5,74,219,208 ; vblendvps %ymm13,%ymm11,%ymm15,%ymm11
+ DB 197,36,92,249 ; vsubps %ymm1,%ymm11,%ymm15
+ DB 196,65,52,89,255 ; vmulps %ymm15,%ymm9,%ymm15
+ DB 196,65,4,94,252 ; vdivps %ymm12,%ymm15,%ymm15
+ DB 196,65,116,88,255 ; vaddps %ymm15,%ymm1,%ymm15
+ DB 196,67,37,74,223,224 ; vblendvps %ymm14,%ymm15,%ymm11,%ymm11
+ DB 196,65,36,95,216 ; vmaxps %ymm8,%ymm11,%ymm11
+ DB 197,124,92,249 ; vsubps %ymm1,%ymm0,%ymm15
+ DB 196,65,116,89,255 ; vmulps %ymm15,%ymm1,%ymm15
+ DB 196,65,4,94,210 ; vdivps %ymm10,%ymm15,%ymm10
+ DB 196,65,116,88,210 ; vaddps %ymm10,%ymm1,%ymm10
+ DB 196,227,45,74,192,208 ; vblendvps %ymm13,%ymm0,%ymm10,%ymm0
+ DB 197,124,92,209 ; vsubps %ymm1,%ymm0,%ymm10
+ DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
+ DB 196,65,52,94,204 ; vdivps %ymm12,%ymm9,%ymm9
+ DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1
+ DB 196,227,125,74,193,224 ; vblendvps %ymm14,%ymm1,%ymm0,%ymm0
+ DB 196,65,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm8
+ DB 197,228,88,199 ; vaddps %ymm7,%ymm3,%ymm0
+ DB 197,228,89,207 ; vmulps %ymm7,%ymm3,%ymm1
+ DB 197,252,92,217 ; vsubps %ymm1,%ymm0,%ymm3
+ DB 197,228,89,194 ; vmulps %ymm2,%ymm3,%ymm0
+ DB 196,193,100,89,203 ; vmulps %ymm11,%ymm3,%ymm1
+ DB 196,193,100,89,208 ; vmulps %ymm8,%ymm3,%ymm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_clamp_0_avx
_sk_clamp_0_avx LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
@@ -5241,7 +5994,7 @@ _sk_clamp_0_avx LABEL PROC
PUBLIC _sk_clamp_1_avx
_sk_clamp_1_avx LABEL PROC
- DB 196,98,125,24,5,240,73,0,0 ; vbroadcastss 0x49f0(%rip),%ymm8 # 56b8 <_sk_callback_avx+0x181>
+ DB 196,98,125,24,5,47,74,0,0 ; vbroadcastss 0x4a2f(%rip),%ymm8 # 5de8 <_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
@@ -5251,7 +6004,7 @@ _sk_clamp_1_avx LABEL PROC
PUBLIC _sk_clamp_a_avx
_sk_clamp_a_avx LABEL PROC
- DB 196,98,125,24,5,211,73,0,0 ; vbroadcastss 0x49d3(%rip),%ymm8 # 56bc <_sk_callback_avx+0x185>
+ DB 196,98,125,24,5,18,74,0,0 ; vbroadcastss 0x4a12(%rip),%ymm8 # 5dec <_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
@@ -5323,7 +6076,7 @@ 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 196,98,125,24,21,27,73,0,0 ; vbroadcastss 0x491b(%rip),%ymm10 # 56c0 <_sk_callback_avx+0x189>
+ DB 196,98,125,24,21,90,73,0,0 ; vbroadcastss 0x495a(%rip),%ymm10 # 5df0 <_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
@@ -5334,17 +6087,17 @@ _sk_unpremul_avx LABEL PROC
PUBLIC _sk_from_srgb_avx
_sk_from_srgb_avx LABEL PROC
- DB 196,98,125,24,5,252,72,0,0 ; vbroadcastss 0x48fc(%rip),%ymm8 # 56c4 <_sk_callback_avx+0x18d>
+ DB 196,98,125,24,5,59,73,0,0 ; vbroadcastss 0x493b(%rip),%ymm8 # 5df4 <_sk_callback_avx+0x1cd>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10
- DB 196,98,125,24,29,238,72,0,0 ; vbroadcastss 0x48ee(%rip),%ymm11 # 56c8 <_sk_callback_avx+0x191>
+ DB 196,98,125,24,29,45,73,0,0 ; vbroadcastss 0x492d(%rip),%ymm11 # 5df8 <_sk_callback_avx+0x1d1>
DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12
- DB 196,98,125,24,45,228,72,0,0 ; vbroadcastss 0x48e4(%rip),%ymm13 # 56cc <_sk_callback_avx+0x195>
+ DB 196,98,125,24,45,35,73,0,0 ; vbroadcastss 0x4923(%rip),%ymm13 # 5dfc <_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 196,98,125,24,37,213,72,0,0 ; vbroadcastss 0x48d5(%rip),%ymm12 # 56d0 <_sk_callback_avx+0x199>
+ DB 196,98,125,24,37,20,73,0,0 ; vbroadcastss 0x4914(%rip),%ymm12 # 5e00 <_sk_callback_avx+0x1d9>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,53,203,72,0,0 ; vbroadcastss 0x48cb(%rip),%ymm14 # 56d4 <_sk_callback_avx+0x19d>
+ DB 196,98,125,24,53,10,73,0,0 ; vbroadcastss 0x490a(%rip),%ymm14 # 5e04 <_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
@@ -5371,18 +6124,18 @@ _sk_to_srgb_avx LABEL PROC
DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8
DB 196,65,124,83,200 ; vrcpps %ymm8,%ymm9
DB 196,65,124,82,208 ; vrsqrtps %ymm8,%ymm10
- DB 196,98,125,24,5,86,72,0,0 ; vbroadcastss 0x4856(%rip),%ymm8 # 56d8 <_sk_callback_avx+0x1a1>
+ DB 196,98,125,24,5,149,72,0,0 ; vbroadcastss 0x4895(%rip),%ymm8 # 5e08 <_sk_callback_avx+0x1e1>
DB 196,65,124,89,216 ; vmulps %ymm8,%ymm0,%ymm11
- DB 196,98,125,24,37,76,72,0,0 ; vbroadcastss 0x484c(%rip),%ymm12 # 56dc <_sk_callback_avx+0x1a5>
+ DB 196,98,125,24,37,139,72,0,0 ; vbroadcastss 0x488b(%rip),%ymm12 # 5e0c <_sk_callback_avx+0x1e5>
DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9
- DB 196,98,125,24,45,66,72,0,0 ; vbroadcastss 0x4842(%rip),%ymm13 # 56e0 <_sk_callback_avx+0x1a9>
+ DB 196,98,125,24,45,129,72,0,0 ; vbroadcastss 0x4881(%rip),%ymm13 # 5e10 <_sk_callback_avx+0x1e9>
DB 196,65,52,88,205 ; vaddps %ymm13,%ymm9,%ymm9
- DB 196,98,125,24,53,56,72,0,0 ; vbroadcastss 0x4838(%rip),%ymm14 # 56e4 <_sk_callback_avx+0x1ad>
+ DB 196,98,125,24,53,119,72,0,0 ; vbroadcastss 0x4877(%rip),%ymm14 # 5e14 <_sk_callback_avx+0x1ed>
DB 196,65,44,89,214 ; vmulps %ymm14,%ymm10,%ymm10
DB 196,65,44,88,201 ; vaddps %ymm9,%ymm10,%ymm9
- DB 196,98,125,24,21,41,72,0,0 ; vbroadcastss 0x4829(%rip),%ymm10 # 56e8 <_sk_callback_avx+0x1b1>
+ DB 196,98,125,24,21,104,72,0,0 ; vbroadcastss 0x4868(%rip),%ymm10 # 5e18 <_sk_callback_avx+0x1f1>
DB 196,65,44,93,201 ; vminps %ymm9,%ymm10,%ymm9
- DB 196,98,125,24,61,31,72,0,0 ; vbroadcastss 0x481f(%rip),%ymm15 # 56ec <_sk_callback_avx+0x1b5>
+ DB 196,98,125,24,61,94,72,0,0 ; vbroadcastss 0x485e(%rip),%ymm15 # 5e1c <_sk_callback_avx+0x1f5>
DB 196,193,124,194,199,1 ; vcmpltps %ymm15,%ymm0,%ymm0
DB 196,195,53,74,195,0 ; vblendvps %ymm0,%ymm11,%ymm9,%ymm0
DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9
@@ -5417,7 +6170,7 @@ _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 196,98,125,24,29,133,71,0,0 ; vbroadcastss 0x4785(%rip),%ymm11 # 56f0 <_sk_callback_avx+0x1b9>
+ DB 196,98,125,24,29,196,71,0,0 ; vbroadcastss 0x47c4(%rip),%ymm11 # 5e20 <_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
@@ -5427,19 +6180,19 @@ _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 196,98,125,24,29,94,71,0,0 ; vbroadcastss 0x475e(%rip),%ymm11 # 56fc <_sk_callback_avx+0x1c5>
+ DB 196,98,125,24,29,157,71,0,0 ; vbroadcastss 0x479d(%rip),%ymm11 # 5e2c <_sk_callback_avx+0x205>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,76,71,0,0 ; vbroadcastss 0x474c(%rip),%ymm11 # 56f8 <_sk_callback_avx+0x1c1>
+ DB 196,98,125,24,29,139,71,0,0 ; vbroadcastss 0x478b(%rip),%ymm11 # 5e28 <_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 196,226,125,24,21,52,71,0,0 ; vbroadcastss 0x4734(%rip),%ymm2 # 56f4 <_sk_callback_avx+0x1bd>
+ DB 196,226,125,24,21,115,71,0,0 ; vbroadcastss 0x4773(%rip),%ymm2 # 5e24 <_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 196,98,125,24,37,27,71,0,0 ; vbroadcastss 0x471b(%rip),%ymm12 # 5704 <_sk_callback_avx+0x1cd>
+ DB 196,98,125,24,37,90,71,0,0 ; vbroadcastss 0x475a(%rip),%ymm12 # 5e34 <_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
@@ -5449,7 +6202,7 @@ _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 196,98,125,24,5,222,70,0,0 ; vbroadcastss 0x46de(%rip),%ymm8 # 5700 <_sk_callback_avx+0x1c9>
+ DB 196,98,125,24,5,29,71,0,0 ; vbroadcastss 0x471d(%rip),%ymm8 # 5e30 <_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
@@ -5464,7 +6217,7 @@ _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 196,98,125,24,5,165,70,0,0 ; vbroadcastss 0x46a5(%rip),%ymm8 # 5708 <_sk_callback_avx+0x1d1>
+ DB 196,98,125,24,5,228,70,0,0 ; vbroadcastss 0x46e4(%rip),%ymm8 # 5e38 <_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
@@ -5472,23 +6225,23 @@ _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 196,98,125,24,29,127,70,0,0 ; vbroadcastss 0x467f(%rip),%ymm11 # 570c <_sk_callback_avx+0x1d5>
+ DB 196,98,125,24,29,190,70,0,0 ; vbroadcastss 0x46be(%rip),%ymm11 # 5e3c <_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 196,98,125,24,37,105,70,0,0 ; vbroadcastss 0x4669(%rip),%ymm12 # 5714 <_sk_callback_avx+0x1dd>
+ DB 196,98,125,24,37,168,70,0,0 ; vbroadcastss 0x46a8(%rip),%ymm12 # 5e44 <_sk_callback_avx+0x21d>
DB 196,193,4,89,196 ; vmulps %ymm12,%ymm15,%ymm0
- DB 196,98,125,24,45,95,70,0,0 ; vbroadcastss 0x465f(%rip),%ymm13 # 5718 <_sk_callback_avx+0x1e1>
+ DB 196,98,125,24,45,158,70,0,0 ; vbroadcastss 0x469e(%rip),%ymm13 # 5e48 <_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 196,226,125,24,13,64,70,0,0 ; vbroadcastss 0x4640(%rip),%ymm1 # 5710 <_sk_callback_avx+0x1d9>
+ DB 196,226,125,24,13,127,70,0,0 ; vbroadcastss 0x467f(%rip),%ymm1 # 5e40 <_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 196,98,125,24,53,43,70,0,0 ; vbroadcastss 0x462b(%rip),%ymm14 # 571c <_sk_callback_avx+0x1e5>
+ DB 196,98,125,24,53,106,70,0,0 ; vbroadcastss 0x466a(%rip),%ymm14 # 5e4c <_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
@@ -5507,7 +6260,7 @@ _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 196,226,125,24,5,205,69,0,0 ; vbroadcastss 0x45cd(%rip),%ymm0 # 5720 <_sk_callback_avx+0x1e9>
+ DB 196,226,125,24,5,12,70,0,0 ; vbroadcastss 0x460c(%rip),%ymm0 # 5e50 <_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
@@ -5555,14 +6308,14 @@ _sk_scale_u8_avx LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,68 ; jne 1250 <_sk_scale_u8_avx+0x54>
+ DB 117,68 ; jne 1941 <_sk_scale_u8_avx+0x54>
DB 197,122,126,0 ; vmovq (%rax),%xmm8
DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9
DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8
DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8
DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,240,68,0,0 ; vbroadcastss 0x44f0(%rip),%ymm9 # 5724 <_sk_callback_avx+0x1ed>
+ DB 196,98,125,24,13,47,69,0,0 ; vbroadcastss 0x452f(%rip),%ymm9 # 5e54 <_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
@@ -5580,9 +6333,9 @@ _sk_scale_u8_avx LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 1258 <_sk_scale_u8_avx+0x5c>
+ DB 117,234 ; jne 1949 <_sk_scale_u8_avx+0x5c>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 235,155 ; jmp 1210 <_sk_scale_u8_avx+0x14>
+ DB 235,155 ; jmp 1901 <_sk_scale_u8_avx+0x14>
PUBLIC _sk_lerp_1_float_avx
_sk_lerp_1_float_avx LABEL PROC
@@ -5610,14 +6363,14 @@ _sk_lerp_u8_avx LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,104 ; jne 132c <_sk_lerp_u8_avx+0x78>
+ DB 117,104 ; jne 1a1d <_sk_lerp_u8_avx+0x78>
DB 197,122,126,0 ; vmovq (%rax),%xmm8
DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9
DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8
DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8
DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,60,68,0,0 ; vbroadcastss 0x443c(%rip),%ymm9 # 5728 <_sk_callback_avx+0x1f1>
+ DB 196,98,125,24,13,123,68,0,0 ; vbroadcastss 0x447b(%rip),%ymm9 # 5e58 <_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
@@ -5643,35 +6396,35 @@ _sk_lerp_u8_avx LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 1334 <_sk_lerp_u8_avx+0x80>
+ DB 117,234 ; jne 1a25 <_sk_lerp_u8_avx+0x80>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 233,116,255,255,255 ; jmpq 12c8 <_sk_lerp_u8_avx+0x14>
+ DB 233,116,255,255,255 ; jmpq 19b9 <_sk_lerp_u8_avx+0x14>
PUBLIC _sk_lerp_565_avx
_sk_lerp_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,174,0,0,0 ; jne 1410 <_sk_lerp_565_avx+0xbc>
+ DB 15,133,174,0,0,0 ; jne 1b01 <_sk_lerp_565_avx+0xbc>
DB 196,65,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm8
DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3
DB 197,185,105,219 ; vpunpckhwd %xmm3,%xmm8,%xmm3
DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8
DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
- DB 196,98,125,24,5,168,67,0,0 ; vbroadcastss 0x43a8(%rip),%ymm8 # 572c <_sk_callback_avx+0x1f5>
+ DB 196,98,125,24,5,231,67,0,0 ; vbroadcastss 0x43e7(%rip),%ymm8 # 5e5c <_sk_callback_avx+0x235>
DB 196,65,100,84,192 ; vandps %ymm8,%ymm3,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,153,67,0,0 ; vbroadcastss 0x4399(%rip),%ymm9 # 5730 <_sk_callback_avx+0x1f9>
+ DB 196,98,125,24,13,216,67,0,0 ; vbroadcastss 0x43d8(%rip),%ymm9 # 5e60 <_sk_callback_avx+0x239>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,143,67,0,0 ; vbroadcastss 0x438f(%rip),%ymm9 # 5734 <_sk_callback_avx+0x1fd>
+ DB 196,98,125,24,13,206,67,0,0 ; vbroadcastss 0x43ce(%rip),%ymm9 # 5e64 <_sk_callback_avx+0x23d>
DB 196,65,100,84,201 ; vandps %ymm9,%ymm3,%ymm9
DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9
- DB 196,98,125,24,21,128,67,0,0 ; vbroadcastss 0x4380(%rip),%ymm10 # 5738 <_sk_callback_avx+0x201>
+ DB 196,98,125,24,21,191,67,0,0 ; vbroadcastss 0x43bf(%rip),%ymm10 # 5e68 <_sk_callback_avx+0x241>
DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
- DB 196,98,125,24,21,118,67,0,0 ; vbroadcastss 0x4376(%rip),%ymm10 # 573c <_sk_callback_avx+0x205>
+ DB 196,98,125,24,21,181,67,0,0 ; vbroadcastss 0x43b5(%rip),%ymm10 # 5e6c <_sk_callback_avx+0x245>
DB 196,193,100,84,218 ; vandps %ymm10,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,21,104,67,0,0 ; vbroadcastss 0x4368(%rip),%ymm10 # 5740 <_sk_callback_avx+0x209>
+ DB 196,98,125,24,21,167,67,0,0 ; vbroadcastss 0x43a7(%rip),%ymm10 # 5e70 <_sk_callback_avx+0x249>
DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
@@ -5683,16 +6436,16 @@ _sk_lerp_565_avx LABEL PROC
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 197,236,88,214 ; vaddps %ymm6,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,54,67,0,0 ; vbroadcastss 0x4336(%rip),%ymm3 # 5744 <_sk_callback_avx+0x20d>
+ DB 196,226,125,24,29,117,67,0,0 ; vbroadcastss 0x4375(%rip),%ymm3 # 5e74 <_sk_callback_avx+0x24d>
DB 255,224 ; jmpq *%rax
DB 65,137,200 ; mov %ecx,%r8d
DB 65,128,224,7 ; and $0x7,%r8b
DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,63,255,255,255 ; ja 1368 <_sk_lerp_565_avx+0x14>
+ DB 15,135,63,255,255,255 ; ja 1a59 <_sk_lerp_565_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1480 <_sk_lerp_565_avx+0x12c>
+ DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 1b70 <_sk_lerp_565_avx+0x12b>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -5704,26 +6457,28 @@ _sk_lerp_565_avx LABEL PROC
DB 196,65,57,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8
DB 196,65,57,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8
DB 196,65,57,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8
- DB 233,235,254,255,255 ; jmpq 1368 <_sk_lerp_565_avx+0x14>
- DB 15,31,0 ; nopl (%rax)
- DB 241 ; icebp
+ DB 233,235,254,255,255 ; jmpq 1a59 <_sk_lerp_565_avx+0x14>
+ DB 102,144 ; xchg %ax,%ax
+ DB 242,255 ; repnz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 234 ; (bad)
+ DB 255 ; (bad)
DB 255 ; (bad)
- DB 233,255,255,255,225 ; jmpq ffffffffe2001488 <_sk_callback_avx+0xffffffffe1ffbf51>
+ DB 255,226 ; jmpq *%rdx
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 217,255 ; fcos
+ DB 218,255 ; (bad)
DB 255 ; (bad)
- DB 255,209 ; callq *%rcx
+ DB 255,210 ; callq *%rdx
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,201 ; dec %ecx
+ DB 255,202 ; dec %edx
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 189 ; .byte 0xbd
+ DB 190 ; .byte 0xbe
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -5733,7 +6488,7 @@ _sk_load_tables_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,0 ; mov (%rax),%r8
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,26,2,0,0 ; jne 16c4 <_sk_load_tables_avx+0x228>
+ DB 15,133,26,2,0,0 ; jne 1db4 <_sk_load_tables_avx+0x228>
DB 196,65,124,16,4,184 ; vmovups (%r8,%rdi,4),%ymm8
DB 85 ; push %rbp
DB 65,87 ; push %r15
@@ -5741,7 +6496,7 @@ _sk_load_tables_avx LABEL PROC
DB 65,85 ; push %r13
DB 65,84 ; push %r12
DB 83 ; push %rbx
- DB 197,124,40,13,94,69,0,0 ; vmovaps 0x455e(%rip),%ymm9 # 5a20 <_sk_callback_avx+0x4e9>
+ DB 197,124,40,13,174,69,0,0 ; vmovaps 0x45ae(%rip),%ymm9 # 6160 <_sk_callback_avx+0x539>
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
@@ -5833,7 +6588,7 @@ _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 196,98,125,24,5,151,64,0,0 ; vbroadcastss 0x4097(%rip),%ymm8 # 5748 <_sk_callback_avx+0x211>
+ DB 196,98,125,24,5,215,64,0,0 ; vbroadcastss 0x40d7(%rip),%ymm8 # 5e78 <_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
@@ -5848,9 +6603,9 @@ _sk_load_tables_avx LABEL PROC
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
DB 65,254,201 ; dec %r9b
DB 65,128,249,6 ; cmp $0x6,%r9b
- DB 15,135,211,253,255,255 ; ja 14b0 <_sk_load_tables_avx+0x14>
+ DB 15,135,211,253,255,255 ; ja 1ba0 <_sk_load_tables_avx+0x14>
DB 69,15,182,201 ; movzbl %r9b,%r9d
- DB 76,141,21,140,0,0,0 ; lea 0x8c(%rip),%r10 # 1774 <_sk_load_tables_avx+0x2d8>
+ DB 76,141,21,140,0,0,0 ; lea 0x8c(%rip),%r10 # 1e64 <_sk_load_tables_avx+0x2d8>
DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9
DB 77,1,209 ; add %r10,%r9
DB 65,255,225 ; jmpq *%r9
@@ -5873,7 +6628,7 @@ _sk_load_tables_avx LABEL PROC
DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8
DB 196,195,57,34,4,184,0 ; vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0
DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8
- DB 233,62,253,255,255 ; jmpq 14b0 <_sk_load_tables_avx+0x14>
+ DB 233,62,253,255,255 ; jmpq 1ba0 <_sk_load_tables_avx+0x14>
DB 102,144 ; xchg %ax,%ax
DB 236 ; in (%dx),%al
DB 255 ; (bad)
@@ -5891,7 +6646,7 @@ _sk_load_tables_avx LABEL PROC
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 126,255 ; jle 178d <_sk_load_tables_avx+0x2f1>
+ DB 126,255 ; jle 1e7d <_sk_load_tables_avx+0x2f1>
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -5901,7 +6656,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,113,2,0,0 ; jne 1a17 <_sk_load_tables_u16_be_avx+0x287>
+ DB 15,133,113,2,0,0 ; jne 2107 <_sk_load_tables_u16_be_avx+0x287>
DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8
DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2
DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3
@@ -5923,7 +6678,7 @@ _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 197,121,111,29,158,66,0,0 ; vmovdqa 0x429e(%rip),%xmm11 # 5aa0 <_sk_callback_avx+0x569>
+ DB 197,121,111,29,238,66,0,0 ; vmovdqa 0x42ee(%rip),%xmm11 # 61e0 <_sk_callback_avx+0x5b9>
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
@@ -6022,7 +6777,7 @@ _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 196,98,125,24,5,72,61,0,0 ; vbroadcastss 0x3d48(%rip),%ymm8 # 574c <_sk_callback_avx+0x215>
+ DB 196,98,125,24,5,136,61,0,0 ; vbroadcastss 0x3d88(%rip),%ymm8 # 5e7c <_sk_callback_avx+0x255>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 91 ; pop %rbx
@@ -6035,29 +6790,29 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 1a7d <_sk_load_tables_u16_be_avx+0x2ed>
+ DB 116,85 ; je 216d <_sk_load_tables_u16_be_avx+0x2ed>
DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 1a7d <_sk_load_tables_u16_be_avx+0x2ed>
+ DB 114,72 ; jb 216d <_sk_load_tables_u16_be_avx+0x2ed>
DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 1a8a <_sk_load_tables_u16_be_avx+0x2fa>
+ DB 116,72 ; je 217a <_sk_load_tables_u16_be_avx+0x2fa>
DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 1a8a <_sk_load_tables_u16_be_avx+0x2fa>
+ DB 114,59 ; jb 217a <_sk_load_tables_u16_be_avx+0x2fa>
DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,97,253,255,255 ; je 17c1 <_sk_load_tables_u16_be_avx+0x31>
+ DB 15,132,97,253,255,255 ; je 1eb1 <_sk_load_tables_u16_be_avx+0x31>
DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,80,253,255,255 ; jb 17c1 <_sk_load_tables_u16_be_avx+0x31>
+ DB 15,130,80,253,255,255 ; jb 1eb1 <_sk_load_tables_u16_be_avx+0x31>
DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9
- DB 233,68,253,255,255 ; jmpq 17c1 <_sk_load_tables_u16_be_avx+0x31>
+ DB 233,68,253,255,255 ; jmpq 1eb1 <_sk_load_tables_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,55,253,255,255 ; jmpq 17c1 <_sk_load_tables_u16_be_avx+0x31>
+ DB 233,55,253,255,255 ; jmpq 1eb1 <_sk_load_tables_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,46,253,255,255 ; jmpq 17c1 <_sk_load_tables_u16_be_avx+0x31>
+ DB 233,46,253,255,255 ; jmpq 1eb1 <_sk_load_tables_u16_be_avx+0x31>
PUBLIC _sk_load_tables_rgb_u16_be_avx
_sk_load_tables_rgb_u16_be_avx LABEL PROC
@@ -6065,7 +6820,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,93,2,0,0 ; jne 1d02 <_sk_load_tables_rgb_u16_be_avx+0x26f>
+ DB 15,133,93,2,0,0 ; jne 23f2 <_sk_load_tables_rgb_u16_be_avx+0x26f>
DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0
DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2
DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1
@@ -6092,7 +6847,7 @@ _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 197,121,111,13,151,63,0,0 ; vmovdqa 0x3f97(%rip),%xmm9 # 5ab0 <_sk_callback_avx+0x579>
+ DB 197,121,111,13,231,63,0,0 ; vmovdqa 0x3fe7(%rip),%xmm9 # 61f0 <_sk_callback_avx+0x5c9>
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
@@ -6184,7 +6939,7 @@ _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 196,226,125,24,29,90,58,0,0 ; vbroadcastss 0x3a5a(%rip),%ymm3 # 5750 <_sk_callback_avx+0x219>
+ DB 196,226,125,24,29,154,58,0,0 ; vbroadcastss 0x3a9a(%rip),%ymm3 # 5e80 <_sk_callback_avx+0x259>
DB 91 ; pop %rbx
DB 65,92 ; pop %r12
DB 65,93 ; pop %r13
@@ -6195,36 +6950,36 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0
DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 1d1b <_sk_load_tables_rgb_u16_be_avx+0x288>
- DB 233,190,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 117,5 ; jne 240b <_sk_load_tables_rgb_u16_be_avx+0x288>
+ DB 233,190,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1
DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 1d4a <_sk_load_tables_rgb_u16_be_avx+0x2b7>
+ DB 114,26 ; jb 243a <_sk_load_tables_rgb_u16_be_avx+0x2b7>
DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1
DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 1d4f <_sk_load_tables_rgb_u16_be_avx+0x2bc>
- DB 233,143,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
- DB 233,138,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 243f <_sk_load_tables_rgb_u16_be_avx+0x2bc>
+ DB 233,143,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,138,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1
DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 1d7e <_sk_load_tables_rgb_u16_be_avx+0x2eb>
+ DB 114,26 ; jb 246e <_sk_load_tables_rgb_u16_be_avx+0x2eb>
DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1
DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 1d83 <_sk_load_tables_rgb_u16_be_avx+0x2f0>
- DB 233,91,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
- DB 233,86,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 2473 <_sk_load_tables_rgb_u16_be_avx+0x2f0>
+ DB 233,91,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,86,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3
DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 1dac <_sk_load_tables_rgb_u16_be_avx+0x319>
+ DB 114,20 ; jb 249c <_sk_load_tables_rgb_u16_be_avx+0x319>
DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3
DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
- DB 233,45,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
- DB 233,40,253,255,255 ; jmpq 1ad9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,45,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,40,253,255,255 ; jmpq 21c9 <_sk_load_tables_rgb_u16_be_avx+0x46>
PUBLIC _sk_byte_tables_avx
_sk_byte_tables_avx LABEL PROC
@@ -6235,7 +6990,7 @@ _sk_byte_tables_avx LABEL PROC
DB 65,84 ; push %r12
DB 83 ; push %rbx
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,142,57,0,0 ; vbroadcastss 0x398e(%rip),%ymm8 # 5754 <_sk_callback_avx+0x21d>
+ DB 196,98,125,24,5,206,57,0,0 ; vbroadcastss 0x39ce(%rip),%ymm8 # 5e84 <_sk_callback_avx+0x25d>
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
@@ -6272,7 +7027,7 @@ _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 196,98,125,24,13,220,56,0,0 ; vbroadcastss 0x38dc(%rip),%ymm9 # 5758 <_sk_callback_avx+0x221>
+ DB 196,98,125,24,13,28,57,0,0 ; vbroadcastss 0x391c(%rip),%ymm9 # 5e88 <_sk_callback_avx+0x261>
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
@@ -6432,7 +7187,7 @@ _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 196,98,125,24,13,2,54,0,0 ; vbroadcastss 0x3602(%rip),%ymm9 # 575c <_sk_callback_avx+0x225>
+ DB 196,98,125,24,13,66,54,0,0 ; vbroadcastss 0x3642(%rip),%ymm9 # 5e8c <_sk_callback_avx+0x265>
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
@@ -6719,36 +7474,36 @@ _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 196,98,125,24,37,96,49,0,0 ; vbroadcastss 0x3160(%rip),%ymm12 # 5760 <_sk_callback_avx+0x229>
+ DB 196,98,125,24,37,160,49,0,0 ; vbroadcastss 0x31a0(%rip),%ymm12 # 5e90 <_sk_callback_avx+0x269>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,86,49,0,0 ; vbroadcastss 0x3156(%rip),%ymm12 # 5764 <_sk_callback_avx+0x22d>
+ DB 196,98,125,24,37,150,49,0,0 ; vbroadcastss 0x3196(%rip),%ymm12 # 5e94 <_sk_callback_avx+0x26d>
DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,76,49,0,0 ; vbroadcastss 0x314c(%rip),%ymm12 # 5768 <_sk_callback_avx+0x231>
+ DB 196,98,125,24,37,140,49,0,0 ; vbroadcastss 0x318c(%rip),%ymm12 # 5e98 <_sk_callback_avx+0x271>
DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,66,49,0,0 ; vbroadcastss 0x3142(%rip),%ymm12 # 576c <_sk_callback_avx+0x235>
+ DB 196,98,125,24,37,130,49,0,0 ; vbroadcastss 0x3182(%rip),%ymm12 # 5e9c <_sk_callback_avx+0x275>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,56,49,0,0 ; vbroadcastss 0x3138(%rip),%ymm12 # 5770 <_sk_callback_avx+0x239>
+ DB 196,98,125,24,37,120,49,0,0 ; vbroadcastss 0x3178(%rip),%ymm12 # 5ea0 <_sk_callback_avx+0x279>
DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,41,49,0,0 ; vbroadcastss 0x3129(%rip),%ymm12 # 5774 <_sk_callback_avx+0x23d>
+ DB 196,98,125,24,37,105,49,0,0 ; vbroadcastss 0x3169(%rip),%ymm12 # 5ea4 <_sk_callback_avx+0x27d>
DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,31,49,0,0 ; vbroadcastss 0x311f(%rip),%ymm12 # 5778 <_sk_callback_avx+0x241>
+ DB 196,98,125,24,37,95,49,0,0 ; vbroadcastss 0x315f(%rip),%ymm12 # 5ea8 <_sk_callback_avx+0x281>
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 196,98,125,24,29,3,49,0,0 ; vbroadcastss 0x3103(%rip),%ymm11 # 577c <_sk_callback_avx+0x245>
+ DB 196,98,125,24,29,67,49,0,0 ; vbroadcastss 0x3143(%rip),%ymm11 # 5eac <_sk_callback_avx+0x285>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,249,48,0,0 ; vbroadcastss 0x30f9(%rip),%ymm11 # 5780 <_sk_callback_avx+0x249>
+ DB 196,98,125,24,29,57,49,0,0 ; vbroadcastss 0x3139(%rip),%ymm11 # 5eb0 <_sk_callback_avx+0x289>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,234,48,0,0 ; vbroadcastss 0x30ea(%rip),%ymm11 # 5784 <_sk_callback_avx+0x24d>
+ DB 196,98,125,24,29,42,49,0,0 ; vbroadcastss 0x312a(%rip),%ymm11 # 5eb4 <_sk_callback_avx+0x28d>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,224,48,0,0 ; vbroadcastss 0x30e0(%rip),%ymm11 # 5788 <_sk_callback_avx+0x251>
+ DB 196,98,125,24,29,32,49,0,0 ; vbroadcastss 0x3120(%rip),%ymm11 # 5eb8 <_sk_callback_avx+0x291>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,209,48,0,0 ; vbroadcastss 0x30d1(%rip),%ymm10 # 578c <_sk_callback_avx+0x255>
+ DB 196,98,125,24,21,17,49,0,0 ; vbroadcastss 0x3111(%rip),%ymm10 # 5ebc <_sk_callback_avx+0x295>
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
@@ -6756,7 +7511,7 @@ _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 196,98,125,24,5,168,48,0,0 ; vbroadcastss 0x30a8(%rip),%ymm8 # 5790 <_sk_callback_avx+0x259>
+ DB 196,98,125,24,5,232,48,0,0 ; vbroadcastss 0x30e8(%rip),%ymm8 # 5ec0 <_sk_callback_avx+0x299>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -6776,36 +7531,36 @@ _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 196,98,125,24,37,89,48,0,0 ; vbroadcastss 0x3059(%rip),%ymm12 # 5794 <_sk_callback_avx+0x25d>
+ DB 196,98,125,24,37,153,48,0,0 ; vbroadcastss 0x3099(%rip),%ymm12 # 5ec4 <_sk_callback_avx+0x29d>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,79,48,0,0 ; vbroadcastss 0x304f(%rip),%ymm12 # 5798 <_sk_callback_avx+0x261>
+ DB 196,98,125,24,37,143,48,0,0 ; vbroadcastss 0x308f(%rip),%ymm12 # 5ec8 <_sk_callback_avx+0x2a1>
DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,69,48,0,0 ; vbroadcastss 0x3045(%rip),%ymm12 # 579c <_sk_callback_avx+0x265>
+ DB 196,98,125,24,37,133,48,0,0 ; vbroadcastss 0x3085(%rip),%ymm12 # 5ecc <_sk_callback_avx+0x2a5>
DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,59,48,0,0 ; vbroadcastss 0x303b(%rip),%ymm12 # 57a0 <_sk_callback_avx+0x269>
+ DB 196,98,125,24,37,123,48,0,0 ; vbroadcastss 0x307b(%rip),%ymm12 # 5ed0 <_sk_callback_avx+0x2a9>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,49,48,0,0 ; vbroadcastss 0x3031(%rip),%ymm12 # 57a4 <_sk_callback_avx+0x26d>
+ DB 196,98,125,24,37,113,48,0,0 ; vbroadcastss 0x3071(%rip),%ymm12 # 5ed4 <_sk_callback_avx+0x2ad>
DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,34,48,0,0 ; vbroadcastss 0x3022(%rip),%ymm12 # 57a8 <_sk_callback_avx+0x271>
+ DB 196,98,125,24,37,98,48,0,0 ; vbroadcastss 0x3062(%rip),%ymm12 # 5ed8 <_sk_callback_avx+0x2b1>
DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,24,48,0,0 ; vbroadcastss 0x3018(%rip),%ymm12 # 57ac <_sk_callback_avx+0x275>
+ DB 196,98,125,24,37,88,48,0,0 ; vbroadcastss 0x3058(%rip),%ymm12 # 5edc <_sk_callback_avx+0x2b5>
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 196,98,125,24,29,252,47,0,0 ; vbroadcastss 0x2ffc(%rip),%ymm11 # 57b0 <_sk_callback_avx+0x279>
+ DB 196,98,125,24,29,60,48,0,0 ; vbroadcastss 0x303c(%rip),%ymm11 # 5ee0 <_sk_callback_avx+0x2b9>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,242,47,0,0 ; vbroadcastss 0x2ff2(%rip),%ymm11 # 57b4 <_sk_callback_avx+0x27d>
+ DB 196,98,125,24,29,50,48,0,0 ; vbroadcastss 0x3032(%rip),%ymm11 # 5ee4 <_sk_callback_avx+0x2bd>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,227,47,0,0 ; vbroadcastss 0x2fe3(%rip),%ymm11 # 57b8 <_sk_callback_avx+0x281>
+ DB 196,98,125,24,29,35,48,0,0 ; vbroadcastss 0x3023(%rip),%ymm11 # 5ee8 <_sk_callback_avx+0x2c1>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,217,47,0,0 ; vbroadcastss 0x2fd9(%rip),%ymm11 # 57bc <_sk_callback_avx+0x285>
+ DB 196,98,125,24,29,25,48,0,0 ; vbroadcastss 0x3019(%rip),%ymm11 # 5eec <_sk_callback_avx+0x2c5>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,202,47,0,0 ; vbroadcastss 0x2fca(%rip),%ymm10 # 57c0 <_sk_callback_avx+0x289>
+ DB 196,98,125,24,21,10,48,0,0 ; vbroadcastss 0x300a(%rip),%ymm10 # 5ef0 <_sk_callback_avx+0x2c9>
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
@@ -6813,7 +7568,7 @@ _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 196,98,125,24,5,161,47,0,0 ; vbroadcastss 0x2fa1(%rip),%ymm8 # 57c4 <_sk_callback_avx+0x28d>
+ DB 196,98,125,24,5,225,47,0,0 ; vbroadcastss 0x2fe1(%rip),%ymm8 # 5ef4 <_sk_callback_avx+0x2cd>
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -6833,36 +7588,36 @@ _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 196,98,125,24,37,82,47,0,0 ; vbroadcastss 0x2f52(%rip),%ymm12 # 57c8 <_sk_callback_avx+0x291>
+ DB 196,98,125,24,37,146,47,0,0 ; vbroadcastss 0x2f92(%rip),%ymm12 # 5ef8 <_sk_callback_avx+0x2d1>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,72,47,0,0 ; vbroadcastss 0x2f48(%rip),%ymm12 # 57cc <_sk_callback_avx+0x295>
+ DB 196,98,125,24,37,136,47,0,0 ; vbroadcastss 0x2f88(%rip),%ymm12 # 5efc <_sk_callback_avx+0x2d5>
DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,62,47,0,0 ; vbroadcastss 0x2f3e(%rip),%ymm12 # 57d0 <_sk_callback_avx+0x299>
+ DB 196,98,125,24,37,126,47,0,0 ; vbroadcastss 0x2f7e(%rip),%ymm12 # 5f00 <_sk_callback_avx+0x2d9>
DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,52,47,0,0 ; vbroadcastss 0x2f34(%rip),%ymm12 # 57d4 <_sk_callback_avx+0x29d>
+ DB 196,98,125,24,37,116,47,0,0 ; vbroadcastss 0x2f74(%rip),%ymm12 # 5f04 <_sk_callback_avx+0x2dd>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,42,47,0,0 ; vbroadcastss 0x2f2a(%rip),%ymm12 # 57d8 <_sk_callback_avx+0x2a1>
+ DB 196,98,125,24,37,106,47,0,0 ; vbroadcastss 0x2f6a(%rip),%ymm12 # 5f08 <_sk_callback_avx+0x2e1>
DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,27,47,0,0 ; vbroadcastss 0x2f1b(%rip),%ymm12 # 57dc <_sk_callback_avx+0x2a5>
+ DB 196,98,125,24,37,91,47,0,0 ; vbroadcastss 0x2f5b(%rip),%ymm12 # 5f0c <_sk_callback_avx+0x2e5>
DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,17,47,0,0 ; vbroadcastss 0x2f11(%rip),%ymm12 # 57e0 <_sk_callback_avx+0x2a9>
+ DB 196,98,125,24,37,81,47,0,0 ; vbroadcastss 0x2f51(%rip),%ymm12 # 5f10 <_sk_callback_avx+0x2e9>
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 196,98,125,24,29,245,46,0,0 ; vbroadcastss 0x2ef5(%rip),%ymm11 # 57e4 <_sk_callback_avx+0x2ad>
+ DB 196,98,125,24,29,53,47,0,0 ; vbroadcastss 0x2f35(%rip),%ymm11 # 5f14 <_sk_callback_avx+0x2ed>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,235,46,0,0 ; vbroadcastss 0x2eeb(%rip),%ymm11 # 57e8 <_sk_callback_avx+0x2b1>
+ DB 196,98,125,24,29,43,47,0,0 ; vbroadcastss 0x2f2b(%rip),%ymm11 # 5f18 <_sk_callback_avx+0x2f1>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,220,46,0,0 ; vbroadcastss 0x2edc(%rip),%ymm11 # 57ec <_sk_callback_avx+0x2b5>
+ DB 196,98,125,24,29,28,47,0,0 ; vbroadcastss 0x2f1c(%rip),%ymm11 # 5f1c <_sk_callback_avx+0x2f5>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,210,46,0,0 ; vbroadcastss 0x2ed2(%rip),%ymm11 # 57f0 <_sk_callback_avx+0x2b9>
+ DB 196,98,125,24,29,18,47,0,0 ; vbroadcastss 0x2f12(%rip),%ymm11 # 5f20 <_sk_callback_avx+0x2f9>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,195,46,0,0 ; vbroadcastss 0x2ec3(%rip),%ymm10 # 57f4 <_sk_callback_avx+0x2bd>
+ DB 196,98,125,24,21,3,47,0,0 ; vbroadcastss 0x2f03(%rip),%ymm10 # 5f24 <_sk_callback_avx+0x2fd>
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
@@ -6870,7 +7625,7 @@ _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 196,98,125,24,5,154,46,0,0 ; vbroadcastss 0x2e9a(%rip),%ymm8 # 57f8 <_sk_callback_avx+0x2c1>
+ DB 196,98,125,24,5,218,46,0,0 ; vbroadcastss 0x2eda(%rip),%ymm8 # 5f28 <_sk_callback_avx+0x301>
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -6890,36 +7645,36 @@ _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 196,98,125,24,37,75,46,0,0 ; vbroadcastss 0x2e4b(%rip),%ymm12 # 57fc <_sk_callback_avx+0x2c5>
+ DB 196,98,125,24,37,139,46,0,0 ; vbroadcastss 0x2e8b(%rip),%ymm12 # 5f2c <_sk_callback_avx+0x305>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,65,46,0,0 ; vbroadcastss 0x2e41(%rip),%ymm12 # 5800 <_sk_callback_avx+0x2c9>
+ DB 196,98,125,24,37,129,46,0,0 ; vbroadcastss 0x2e81(%rip),%ymm12 # 5f30 <_sk_callback_avx+0x309>
DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,55,46,0,0 ; vbroadcastss 0x2e37(%rip),%ymm12 # 5804 <_sk_callback_avx+0x2cd>
+ DB 196,98,125,24,37,119,46,0,0 ; vbroadcastss 0x2e77(%rip),%ymm12 # 5f34 <_sk_callback_avx+0x30d>
DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,45,46,0,0 ; vbroadcastss 0x2e2d(%rip),%ymm12 # 5808 <_sk_callback_avx+0x2d1>
+ DB 196,98,125,24,37,109,46,0,0 ; vbroadcastss 0x2e6d(%rip),%ymm12 # 5f38 <_sk_callback_avx+0x311>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,35,46,0,0 ; vbroadcastss 0x2e23(%rip),%ymm12 # 580c <_sk_callback_avx+0x2d5>
+ DB 196,98,125,24,37,99,46,0,0 ; vbroadcastss 0x2e63(%rip),%ymm12 # 5f3c <_sk_callback_avx+0x315>
DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,20,46,0,0 ; vbroadcastss 0x2e14(%rip),%ymm12 # 5810 <_sk_callback_avx+0x2d9>
+ DB 196,98,125,24,37,84,46,0,0 ; vbroadcastss 0x2e54(%rip),%ymm12 # 5f40 <_sk_callback_avx+0x319>
DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,10,46,0,0 ; vbroadcastss 0x2e0a(%rip),%ymm12 # 5814 <_sk_callback_avx+0x2dd>
+ DB 196,98,125,24,37,74,46,0,0 ; vbroadcastss 0x2e4a(%rip),%ymm12 # 5f44 <_sk_callback_avx+0x31d>
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 196,98,125,24,29,238,45,0,0 ; vbroadcastss 0x2dee(%rip),%ymm11 # 5818 <_sk_callback_avx+0x2e1>
+ DB 196,98,125,24,29,46,46,0,0 ; vbroadcastss 0x2e2e(%rip),%ymm11 # 5f48 <_sk_callback_avx+0x321>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,228,45,0,0 ; vbroadcastss 0x2de4(%rip),%ymm11 # 581c <_sk_callback_avx+0x2e5>
+ DB 196,98,125,24,29,36,46,0,0 ; vbroadcastss 0x2e24(%rip),%ymm11 # 5f4c <_sk_callback_avx+0x325>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,213,45,0,0 ; vbroadcastss 0x2dd5(%rip),%ymm11 # 5820 <_sk_callback_avx+0x2e9>
+ DB 196,98,125,24,29,21,46,0,0 ; vbroadcastss 0x2e15(%rip),%ymm11 # 5f50 <_sk_callback_avx+0x329>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,203,45,0,0 ; vbroadcastss 0x2dcb(%rip),%ymm11 # 5824 <_sk_callback_avx+0x2ed>
+ DB 196,98,125,24,29,11,46,0,0 ; vbroadcastss 0x2e0b(%rip),%ymm11 # 5f54 <_sk_callback_avx+0x32d>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,188,45,0,0 ; vbroadcastss 0x2dbc(%rip),%ymm10 # 5828 <_sk_callback_avx+0x2f1>
+ DB 196,98,125,24,21,252,45,0,0 ; vbroadcastss 0x2dfc(%rip),%ymm10 # 5f58 <_sk_callback_avx+0x331>
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
@@ -6927,38 +7682,38 @@ _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 196,98,125,24,5,147,45,0,0 ; vbroadcastss 0x2d93(%rip),%ymm8 # 582c <_sk_callback_avx+0x2f5>
+ DB 196,98,125,24,5,211,45,0,0 ; vbroadcastss 0x2dd3(%rip),%ymm8 # 5f5c <_sk_callback_avx+0x335>
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 196,98,125,24,5,133,45,0,0 ; vbroadcastss 0x2d85(%rip),%ymm8 # 5830 <_sk_callback_avx+0x2f9>
+ DB 196,98,125,24,5,197,45,0,0 ; vbroadcastss 0x2dc5(%rip),%ymm8 # 5f60 <_sk_callback_avx+0x339>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,123,45,0,0 ; vbroadcastss 0x2d7b(%rip),%ymm8 # 5834 <_sk_callback_avx+0x2fd>
+ DB 196,98,125,24,5,187,45,0,0 ; vbroadcastss 0x2dbb(%rip),%ymm8 # 5f64 <_sk_callback_avx+0x33d>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,98,125,24,13,113,45,0,0 ; vbroadcastss 0x2d71(%rip),%ymm9 # 5838 <_sk_callback_avx+0x301>
+ DB 196,98,125,24,13,177,45,0,0 ; vbroadcastss 0x2db1(%rip),%ymm9 # 5f68 <_sk_callback_avx+0x341>
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 196,98,125,24,5,93,45,0,0 ; vbroadcastss 0x2d5d(%rip),%ymm8 # 583c <_sk_callback_avx+0x305>
+ DB 196,98,125,24,5,157,45,0,0 ; vbroadcastss 0x2d9d(%rip),%ymm8 # 5f6c <_sk_callback_avx+0x345>
DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,83,45,0,0 ; vbroadcastss 0x2d53(%rip),%ymm8 # 5840 <_sk_callback_avx+0x309>
+ DB 196,98,125,24,5,147,45,0,0 ; vbroadcastss 0x2d93(%rip),%ymm8 # 5f70 <_sk_callback_avx+0x349>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,98,125,24,5,73,45,0,0 ; vbroadcastss 0x2d49(%rip),%ymm8 # 5844 <_sk_callback_avx+0x30d>
+ DB 196,98,125,24,5,137,45,0,0 ; vbroadcastss 0x2d89(%rip),%ymm8 # 5f74 <_sk_callback_avx+0x34d>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1
- DB 196,98,125,24,5,59,45,0,0 ; vbroadcastss 0x2d3b(%rip),%ymm8 # 5848 <_sk_callback_avx+0x311>
+ DB 196,98,125,24,5,123,45,0,0 ; vbroadcastss 0x2d7b(%rip),%ymm8 # 5f78 <_sk_callback_avx+0x351>
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 196,98,125,24,13,36,45,0,0 ; vbroadcastss 0x2d24(%rip),%ymm9 # 584c <_sk_callback_avx+0x315>
+ DB 196,98,125,24,13,100,45,0,0 ; vbroadcastss 0x2d64(%rip),%ymm9 # 5f7c <_sk_callback_avx+0x355>
DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10
- DB 196,98,125,24,29,25,45,0,0 ; vbroadcastss 0x2d19(%rip),%ymm11 # 5850 <_sk_callback_avx+0x319>
+ DB 196,98,125,24,29,89,45,0,0 ; vbroadcastss 0x2d59(%rip),%ymm11 # 5f80 <_sk_callback_avx+0x359>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,37,15,45,0,0 ; vbroadcastss 0x2d0f(%rip),%ymm12 # 5854 <_sk_callback_avx+0x31d>
+ DB 196,98,125,24,37,79,45,0,0 ; vbroadcastss 0x2d4f(%rip),%ymm12 # 5f84 <_sk_callback_avx+0x35d>
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
@@ -6973,9 +7728,9 @@ _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 196,226,125,24,5,197,44,0,0 ; vbroadcastss 0x2cc5(%rip),%ymm0 # 5858 <_sk_callback_avx+0x321>
+ DB 196,226,125,24,5,5,45,0,0 ; vbroadcastss 0x2d05(%rip),%ymm0 # 5f88 <_sk_callback_avx+0x361>
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,188,44,0,0 ; vbroadcastss 0x2cbc(%rip),%ymm8 # 585c <_sk_callback_avx+0x325>
+ DB 196,98,125,24,5,252,44,0,0 ; vbroadcastss 0x2cfc(%rip),%ymm8 # 5f8c <_sk_callback_avx+0x365>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -6987,14 +7742,14 @@ _sk_load_a8_avx LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,62 ; jne 2bf7 <_sk_load_a8_avx+0x4e>
+ DB 117,62 ; jne 32e7 <_sk_load_a8_avx+0x4e>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,128,44,0,0 ; vbroadcastss 0x2c80(%rip),%ymm1 # 5860 <_sk_callback_avx+0x329>
+ DB 196,226,125,24,13,192,44,0,0 ; vbroadcastss 0x2cc0(%rip),%ymm1 # 5f90 <_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
@@ -7011,9 +7766,9 @@ _sk_load_a8_avx LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 2bff <_sk_load_a8_avx+0x56>
+ DB 117,234 ; jne 32ef <_sk_load_a8_avx+0x56>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,161 ; jmp 2bbd <_sk_load_a8_avx+0x14>
+ DB 235,161 ; jmp 32ad <_sk_load_a8_avx+0x14>
PUBLIC _sk_gather_a8_avx
_sk_gather_a8_avx LABEL PROC
@@ -7061,7 +7816,7 @@ _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 196,226,125,24,13,117,43,0,0 ; vbroadcastss 0x2b75(%rip),%ymm1 # 5864 <_sk_callback_avx+0x32d>
+ DB 196,226,125,24,13,181,43,0,0 ; vbroadcastss 0x2bb5(%rip),%ymm1 # 5f94 <_sk_callback_avx+0x36d>
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
@@ -7077,14 +7832,14 @@ PUBLIC _sk_store_a8_avx
_sk_store_a8_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
- DB 196,98,125,24,5,80,43,0,0 ; vbroadcastss 0x2b50(%rip),%ymm8 # 5868 <_sk_callback_avx+0x331>
+ DB 196,98,125,24,5,144,43,0,0 ; vbroadcastss 0x2b90(%rip),%ymm8 # 5f98 <_sk_callback_avx+0x371>
DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8
DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 2d41 <_sk_store_a8_avx+0x37>
+ DB 117,10 ; jne 3431 <_sk_store_a8_avx+0x37>
DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7092,10 +7847,10 @@ _sk_store_a8_avx LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 2d3d <_sk_store_a8_avx+0x33>
+ DB 119,236 ; ja 342d <_sk_store_a8_avx+0x33>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 2da4 <_sk_store_a8_avx+0x9a>
+ DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 3494 <_sk_store_a8_avx+0x9a>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7106,7 +7861,7 @@ _sk_store_a8_avx LABEL PROC
DB 196,67,121,20,68,58,2,4 ; vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1)
DB 196,67,121,20,68,58,1,2 ; vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1)
DB 196,67,121,20,4,58,0 ; vpextrb $0x0,%xmm8,(%r10,%rdi,1)
- DB 235,154 ; jmp 2d3d <_sk_store_a8_avx+0x33>
+ DB 235,154 ; jmp 342d <_sk_store_a8_avx+0x33>
DB 144 ; nop
DB 246,255 ; idiv %bh
DB 255 ; (bad)
@@ -7138,17 +7893,17 @@ _sk_load_g8_avx LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,67 ; jne 2e13 <_sk_load_g8_avx+0x53>
+ DB 117,67 ; jne 3503 <_sk_load_g8_avx+0x53>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,117,42,0,0 ; vbroadcastss 0x2a75(%rip),%ymm1 # 586c <_sk_callback_avx+0x335>
+ DB 196,226,125,24,13,181,42,0,0 ; vbroadcastss 0x2ab5(%rip),%ymm1 # 5f9c <_sk_callback_avx+0x375>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,106,42,0,0 ; vbroadcastss 0x2a6a(%rip),%ymm3 # 5870 <_sk_callback_avx+0x339>
+ DB 196,226,125,24,29,170,42,0,0 ; vbroadcastss 0x2aaa(%rip),%ymm3 # 5fa0 <_sk_callback_avx+0x379>
DB 76,137,193 ; mov %r8,%rcx
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
@@ -7162,9 +7917,9 @@ _sk_load_g8_avx LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 2e1b <_sk_load_g8_avx+0x5b>
+ DB 117,234 ; jne 350b <_sk_load_g8_avx+0x5b>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,156 ; jmp 2dd4 <_sk_load_g8_avx+0x14>
+ DB 235,156 ; jmp 34c4 <_sk_load_g8_avx+0x14>
PUBLIC _sk_gather_g8_avx
_sk_gather_g8_avx LABEL PROC
@@ -7212,10 +7967,10 @@ _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 196,226,125,24,13,105,41,0,0 ; vbroadcastss 0x2969(%rip),%ymm1 # 5874 <_sk_callback_avx+0x33d>
+ DB 196,226,125,24,13,169,41,0,0 ; vbroadcastss 0x29a9(%rip),%ymm1 # 5fa4 <_sk_callback_avx+0x37d>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,94,41,0,0 ; vbroadcastss 0x295e(%rip),%ymm3 # 5878 <_sk_callback_avx+0x341>
+ DB 196,226,125,24,29,158,41,0,0 ; vbroadcastss 0x299e(%rip),%ymm3 # 5fa8 <_sk_callback_avx+0x381>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 91 ; pop %rbx
@@ -7229,9 +7984,9 @@ _sk_gather_i8_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,137,192 ; mov %rax,%r8
DB 77,133,192 ; test %r8,%r8
- DB 116,5 ; je 2f3a <_sk_gather_i8_avx+0xf>
+ DB 116,5 ; je 362a <_sk_gather_i8_avx+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 2f3c <_sk_gather_i8_avx+0x11>
+ DB 235,2 ; jmp 362c <_sk_gather_i8_avx+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -7293,10 +8048,10 @@ _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 197,124,40,21,218,41,0,0 ; vmovaps 0x29da(%rip),%ymm10 # 5a40 <_sk_callback_avx+0x509>
+ DB 197,124,40,21,42,42,0,0 ; vmovaps 0x2a2a(%rip),%ymm10 # 6180 <_sk_callback_avx+0x559>
DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,13,4,40,0,0 ; vbroadcastss 0x2804(%rip),%ymm9 # 587c <_sk_callback_avx+0x345>
+ DB 196,98,125,24,13,68,40,0,0 ; vbroadcastss 0x2844(%rip),%ymm9 # 5fac <_sk_callback_avx+0x385>
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
@@ -7328,38 +8083,38 @@ _sk_load_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,128,0,0,0 ; jne 3170 <_sk_load_565_avx+0x8e>
+ DB 15,133,128,0,0,0 ; jne 3860 <_sk_load_565_avx+0x8e>
DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
- DB 196,226,125,24,5,110,39,0,0 ; vbroadcastss 0x276e(%rip),%ymm0 # 5880 <_sk_callback_avx+0x349>
+ DB 196,226,125,24,5,174,39,0,0 ; vbroadcastss 0x27ae(%rip),%ymm0 # 5fb0 <_sk_callback_avx+0x389>
DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,97,39,0,0 ; vbroadcastss 0x2761(%rip),%ymm1 # 5884 <_sk_callback_avx+0x34d>
+ DB 196,226,125,24,13,161,39,0,0 ; vbroadcastss 0x27a1(%rip),%ymm1 # 5fb4 <_sk_callback_avx+0x38d>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,88,39,0,0 ; vbroadcastss 0x2758(%rip),%ymm1 # 5888 <_sk_callback_avx+0x351>
+ DB 196,226,125,24,13,152,39,0,0 ; vbroadcastss 0x2798(%rip),%ymm1 # 5fb8 <_sk_callback_avx+0x391>
DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,75,39,0,0 ; vbroadcastss 0x274b(%rip),%ymm3 # 588c <_sk_callback_avx+0x355>
+ DB 196,226,125,24,29,139,39,0,0 ; vbroadcastss 0x278b(%rip),%ymm3 # 5fbc <_sk_callback_avx+0x395>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,24,29,66,39,0,0 ; vbroadcastss 0x2742(%rip),%ymm3 # 5890 <_sk_callback_avx+0x359>
+ DB 196,226,125,24,29,130,39,0,0 ; vbroadcastss 0x2782(%rip),%ymm3 # 5fc0 <_sk_callback_avx+0x399>
DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,53,39,0,0 ; vbroadcastss 0x2735(%rip),%ymm3 # 5894 <_sk_callback_avx+0x35d>
+ DB 196,226,125,24,29,117,39,0,0 ; vbroadcastss 0x2775(%rip),%ymm3 # 5fc4 <_sk_callback_avx+0x39d>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,42,39,0,0 ; vbroadcastss 0x272a(%rip),%ymm3 # 5898 <_sk_callback_avx+0x361>
+ DB 196,226,125,24,29,106,39,0,0 ; vbroadcastss 0x276a(%rip),%ymm3 # 5fc8 <_sk_callback_avx+0x3a1>
DB 255,224 ; jmpq *%rax
DB 65,137,200 ; mov %ecx,%r8d
DB 65,128,224,7 ; and $0x7,%r8b
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,110,255,255,255 ; ja 30f6 <_sk_load_565_avx+0x14>
+ DB 15,135,110,255,255,255 ; ja 37e6 <_sk_load_565_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 31dc <_sk_load_565_avx+0xfa>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 38cc <_sk_load_565_avx+0xfa>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7371,7 +8126,7 @@ _sk_load_565_avx LABEL PROC
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- DB 233,26,255,255,255 ; jmpq 30f6 <_sk_load_565_avx+0x14>
+ DB 233,26,255,255,255 ; jmpq 37e6 <_sk_load_565_avx+0x14>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -7447,23 +8202,23 @@ _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 196,226,125,24,5,202,37,0,0 ; vbroadcastss 0x25ca(%rip),%ymm0 # 589c <_sk_callback_avx+0x365>
+ DB 196,226,125,24,5,10,38,0,0 ; vbroadcastss 0x260a(%rip),%ymm0 # 5fcc <_sk_callback_avx+0x3a5>
DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,189,37,0,0 ; vbroadcastss 0x25bd(%rip),%ymm1 # 58a0 <_sk_callback_avx+0x369>
+ DB 196,226,125,24,13,253,37,0,0 ; vbroadcastss 0x25fd(%rip),%ymm1 # 5fd0 <_sk_callback_avx+0x3a9>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,180,37,0,0 ; vbroadcastss 0x25b4(%rip),%ymm1 # 58a4 <_sk_callback_avx+0x36d>
+ DB 196,226,125,24,13,244,37,0,0 ; vbroadcastss 0x25f4(%rip),%ymm1 # 5fd4 <_sk_callback_avx+0x3ad>
DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,167,37,0,0 ; vbroadcastss 0x25a7(%rip),%ymm3 # 58a8 <_sk_callback_avx+0x371>
+ DB 196,226,125,24,29,231,37,0,0 ; vbroadcastss 0x25e7(%rip),%ymm3 # 5fd8 <_sk_callback_avx+0x3b1>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,24,29,158,37,0,0 ; vbroadcastss 0x259e(%rip),%ymm3 # 58ac <_sk_callback_avx+0x375>
+ DB 196,226,125,24,29,222,37,0,0 ; vbroadcastss 0x25de(%rip),%ymm3 # 5fdc <_sk_callback_avx+0x3b5>
DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,145,37,0,0 ; vbroadcastss 0x2591(%rip),%ymm3 # 58b0 <_sk_callback_avx+0x379>
+ DB 196,226,125,24,29,209,37,0,0 ; vbroadcastss 0x25d1(%rip),%ymm3 # 5fe0 <_sk_callback_avx+0x3b9>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,134,37,0,0 ; vbroadcastss 0x2586(%rip),%ymm3 # 58b4 <_sk_callback_avx+0x37d>
+ DB 196,226,125,24,29,198,37,0,0 ; vbroadcastss 0x25c6(%rip),%ymm3 # 5fe4 <_sk_callback_avx+0x3bd>
DB 91 ; pop %rbx
DB 65,92 ; pop %r12
DB 65,94 ; pop %r14
@@ -7475,14 +8230,14 @@ 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 196,98,125,24,5,114,37,0,0 ; vbroadcastss 0x2572(%rip),%ymm8 # 58b8 <_sk_callback_avx+0x381>
+ DB 196,98,125,24,5,178,37,0,0 ; vbroadcastss 0x25b2(%rip),%ymm8 # 5fe8 <_sk_callback_avx+0x3c1>
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 196,98,125,24,21,75,37,0,0 ; vbroadcastss 0x254b(%rip),%ymm10 # 58bc <_sk_callback_avx+0x385>
+ DB 196,98,125,24,21,139,37,0,0 ; vbroadcastss 0x258b(%rip),%ymm10 # 5fec <_sk_callback_avx+0x3c5>
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
@@ -7496,7 +8251,7 @@ _sk_store_565_avx LABEL PROC
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 33c1 <_sk_store_565_avx+0x89>
+ DB 117,10 ; jne 3ab1 <_sk_store_565_avx+0x89>
DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7504,9 +8259,9 @@ _sk_store_565_avx LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 33bd <_sk_store_565_avx+0x85>
+ DB 119,236 ; ja 3aad <_sk_store_565_avx+0x85>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3420 <_sk_store_565_avx+0xe8>
+ DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3b10 <_sk_store_565_avx+0xe8>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7517,7 +8272,7 @@ _sk_store_565_avx LABEL PROC
DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- DB 235,159 ; jmp 33bd <_sk_store_565_avx+0x85>
+ DB 235,159 ; jmp 3aad <_sk_store_565_avx+0x85>
DB 102,144 ; xchg %ax,%ax
DB 245 ; cmc
DB 255 ; (bad)
@@ -7548,31 +8303,31 @@ _sk_load_4444_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,152,0,0,0 ; jne 34e2 <_sk_load_4444_avx+0xa6>
+ DB 15,133,152,0,0,0 ; jne 3bd2 <_sk_load_4444_avx+0xa6>
DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3
- DB 196,226,125,24,5,84,36,0,0 ; vbroadcastss 0x2454(%rip),%ymm0 # 58c0 <_sk_callback_avx+0x389>
+ DB 196,226,125,24,5,148,36,0,0 ; vbroadcastss 0x2494(%rip),%ymm0 # 5ff0 <_sk_callback_avx+0x3c9>
DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,71,36,0,0 ; vbroadcastss 0x2447(%rip),%ymm1 # 58c4 <_sk_callback_avx+0x38d>
+ DB 196,226,125,24,13,135,36,0,0 ; vbroadcastss 0x2487(%rip),%ymm1 # 5ff4 <_sk_callback_avx+0x3cd>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,62,36,0,0 ; vbroadcastss 0x243e(%rip),%ymm1 # 58c8 <_sk_callback_avx+0x391>
+ DB 196,226,125,24,13,126,36,0,0 ; vbroadcastss 0x247e(%rip),%ymm1 # 5ff8 <_sk_callback_avx+0x3d1>
DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,49,36,0,0 ; vbroadcastss 0x2431(%rip),%ymm2 # 58cc <_sk_callback_avx+0x395>
+ DB 196,226,125,24,21,113,36,0,0 ; vbroadcastss 0x2471(%rip),%ymm2 # 5ffc <_sk_callback_avx+0x3d5>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,24,21,40,36,0,0 ; vbroadcastss 0x2428(%rip),%ymm2 # 58d0 <_sk_callback_avx+0x399>
+ DB 196,226,125,24,21,104,36,0,0 ; vbroadcastss 0x2468(%rip),%ymm2 # 6000 <_sk_callback_avx+0x3d9>
DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,27,36,0,0 ; vbroadcastss 0x241b(%rip),%ymm8 # 58d4 <_sk_callback_avx+0x39d>
+ DB 196,98,125,24,5,91,36,0,0 ; vbroadcastss 0x245b(%rip),%ymm8 # 6004 <_sk_callback_avx+0x3dd>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,17,36,0,0 ; vbroadcastss 0x2411(%rip),%ymm8 # 58d8 <_sk_callback_avx+0x3a1>
+ DB 196,98,125,24,5,81,36,0,0 ; vbroadcastss 0x2451(%rip),%ymm8 # 6008 <_sk_callback_avx+0x3e1>
DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,3,36,0,0 ; vbroadcastss 0x2403(%rip),%ymm8 # 58dc <_sk_callback_avx+0x3a5>
+ DB 196,98,125,24,5,67,36,0,0 ; vbroadcastss 0x2443(%rip),%ymm8 # 600c <_sk_callback_avx+0x3e5>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7581,9 +8336,9 @@ _sk_load_4444_avx LABEL PROC
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,86,255,255,255 ; ja 3450 <_sk_load_4444_avx+0x14>
+ DB 15,135,86,255,255,255 ; ja 3b40 <_sk_load_4444_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 3550 <_sk_load_4444_avx+0x114>
+ DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 3c40 <_sk_load_4444_avx+0x114>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7595,7 +8350,7 @@ _sk_load_4444_avx LABEL PROC
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- DB 233,2,255,255,255 ; jmpq 3450 <_sk_load_4444_avx+0x14>
+ DB 233,2,255,255,255 ; jmpq 3b40 <_sk_load_4444_avx+0x14>
DB 102,144 ; xchg %ax,%ax
DB 242,255 ; repnz (bad)
DB 255 ; (bad)
@@ -7672,25 +8427,25 @@ _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 196,226,125,24,5,154,34,0,0 ; vbroadcastss 0x229a(%rip),%ymm0 # 58e0 <_sk_callback_avx+0x3a9>
+ DB 196,226,125,24,5,218,34,0,0 ; vbroadcastss 0x22da(%rip),%ymm0 # 6010 <_sk_callback_avx+0x3e9>
DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,141,34,0,0 ; vbroadcastss 0x228d(%rip),%ymm1 # 58e4 <_sk_callback_avx+0x3ad>
+ DB 196,226,125,24,13,205,34,0,0 ; vbroadcastss 0x22cd(%rip),%ymm1 # 6014 <_sk_callback_avx+0x3ed>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,132,34,0,0 ; vbroadcastss 0x2284(%rip),%ymm1 # 58e8 <_sk_callback_avx+0x3b1>
+ DB 196,226,125,24,13,196,34,0,0 ; vbroadcastss 0x22c4(%rip),%ymm1 # 6018 <_sk_callback_avx+0x3f1>
DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,119,34,0,0 ; vbroadcastss 0x2277(%rip),%ymm2 # 58ec <_sk_callback_avx+0x3b5>
+ DB 196,226,125,24,21,183,34,0,0 ; vbroadcastss 0x22b7(%rip),%ymm2 # 601c <_sk_callback_avx+0x3f5>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,24,21,110,34,0,0 ; vbroadcastss 0x226e(%rip),%ymm2 # 58f0 <_sk_callback_avx+0x3b9>
+ DB 196,226,125,24,21,174,34,0,0 ; vbroadcastss 0x22ae(%rip),%ymm2 # 6020 <_sk_callback_avx+0x3f9>
DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,97,34,0,0 ; vbroadcastss 0x2261(%rip),%ymm8 # 58f4 <_sk_callback_avx+0x3bd>
+ DB 196,98,125,24,5,161,34,0,0 ; vbroadcastss 0x22a1(%rip),%ymm8 # 6024 <_sk_callback_avx+0x3fd>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,24,5,87,34,0,0 ; vbroadcastss 0x2257(%rip),%ymm8 # 58f8 <_sk_callback_avx+0x3c1>
+ DB 196,98,125,24,5,151,34,0,0 ; vbroadcastss 0x2297(%rip),%ymm8 # 6028 <_sk_callback_avx+0x401>
DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,73,34,0,0 ; vbroadcastss 0x2249(%rip),%ymm8 # 58fc <_sk_callback_avx+0x3c5>
+ DB 196,98,125,24,5,137,34,0,0 ; vbroadcastss 0x2289(%rip),%ymm8 # 602c <_sk_callback_avx+0x405>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 91 ; pop %rbx
@@ -7704,7 +8459,7 @@ 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 196,98,125,24,5,46,34,0,0 ; vbroadcastss 0x222e(%rip),%ymm8 # 5900 <_sk_callback_avx+0x3c9>
+ DB 196,98,125,24,5,110,34,0,0 ; vbroadcastss 0x226e(%rip),%ymm8 # 6030 <_sk_callback_avx+0x409>
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
@@ -7731,7 +8486,7 @@ _sk_store_4444_avx LABEL PROC
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 376b <_sk_store_4444_avx+0xa7>
+ DB 117,10 ; jne 3e5b <_sk_store_4444_avx+0xa7>
DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7739,9 +8494,9 @@ _sk_store_4444_avx LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 3767 <_sk_store_4444_avx+0xa3>
+ DB 119,236 ; ja 3e57 <_sk_store_4444_avx+0xa3>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,66,0,0,0 ; lea 0x42(%rip),%r9 # 37c8 <_sk_store_4444_avx+0x104>
+ DB 76,141,13,66,0,0,0 ; lea 0x42(%rip),%r9 # 3eb8 <_sk_store_4444_avx+0x104>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7752,7 +8507,7 @@ _sk_store_4444_avx LABEL PROC
DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2)
DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2)
DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2)
- DB 235,159 ; jmp 3767 <_sk_store_4444_avx+0xa3>
+ DB 235,159 ; jmp 3e57 <_sk_store_4444_avx+0xa3>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -7781,12 +8536,12 @@ _sk_load_8888_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,135,0,0,0 ; jne 3879 <_sk_load_8888_avx+0x95>
+ DB 15,133,135,0,0,0 ; jne 3f69 <_sk_load_8888_avx+0x95>
DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9
- DB 197,124,40,21,96,34,0,0 ; vmovaps 0x2260(%rip),%ymm10 # 5a60 <_sk_callback_avx+0x529>
+ DB 197,124,40,21,176,34,0,0 ; vmovaps 0x22b0(%rip),%ymm10 # 61a0 <_sk_callback_avx+0x579>
DB 196,193,52,84,194 ; vandps %ymm10,%ymm9,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,242,32,0,0 ; vbroadcastss 0x20f2(%rip),%ymm8 # 5904 <_sk_callback_avx+0x3cd>
+ DB 196,98,125,24,5,50,33,0,0 ; vbroadcastss 0x2132(%rip),%ymm8 # 6034 <_sk_callback_avx+0x40d>
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
@@ -7813,9 +8568,9 @@ _sk_load_8888_avx LABEL PROC
DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,102,255,255,255 ; ja 37f8 <_sk_load_8888_avx+0x14>
+ DB 15,135,102,255,255,255 ; ja 3ee8 <_sk_load_8888_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,139,0,0,0 ; lea 0x8b(%rip),%r9 # 3928 <_sk_load_8888_avx+0x144>
+ DB 76,141,13,139,0,0,0 ; lea 0x8b(%rip),%r9 # 4018 <_sk_load_8888_avx+0x144>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7838,7 +8593,7 @@ _sk_load_8888_avx LABEL PROC
DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9
DB 196,195,49,34,4,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0
DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9
- DB 233,210,254,255,255 ; jmpq 37f8 <_sk_load_8888_avx+0x14>
+ DB 233,210,254,255,255 ; jmpq 3ee8 <_sk_load_8888_avx+0x14>
DB 102,144 ; xchg %ax,%ax
DB 236 ; in (%dx),%al
DB 255 ; (bad)
@@ -7856,7 +8611,7 @@ _sk_load_8888_avx LABEL PROC
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 126,255 ; jle 3941 <_sk_load_8888_avx+0x15d>
+ DB 126,255 ; jle 4031 <_sk_load_8888_avx+0x15d>
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -7899,10 +8654,10 @@ _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 197,124,40,21,138,32,0,0 ; vmovaps 0x208a(%rip),%ymm10 # 5a80 <_sk_callback_avx+0x549>
+ DB 197,124,40,21,218,32,0,0 ; vmovaps 0x20da(%rip),%ymm10 # 61c0 <_sk_callback_avx+0x599>
DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,13,0,31,0,0 ; vbroadcastss 0x1f00(%rip),%ymm9 # 5908 <_sk_callback_avx+0x3d1>
+ DB 196,98,125,24,13,64,31,0,0 ; vbroadcastss 0x1f40(%rip),%ymm9 # 6038 <_sk_callback_avx+0x411>
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
@@ -7932,7 +8687,7 @@ 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 196,98,125,24,5,142,30,0,0 ; vbroadcastss 0x1e8e(%rip),%ymm8 # 590c <_sk_callback_avx+0x3d5>
+ DB 196,98,125,24,5,206,30,0,0 ; vbroadcastss 0x1ece(%rip),%ymm8 # 603c <_sk_callback_avx+0x415>
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
@@ -7957,7 +8712,7 @@ _sk_store_8888_avx LABEL PROC
DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8
DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 3b0c <_sk_store_8888_avx+0x9c>
+ DB 117,10 ; jne 41fc <_sk_store_8888_avx+0x9c>
DB 196,65,124,17,4,186 ; vmovups %ymm8,(%r10,%rdi,4)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -7965,9 +8720,9 @@ _sk_store_8888_avx LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 3b08 <_sk_store_8888_avx+0x98>
+ DB 119,236 ; ja 41f8 <_sk_store_8888_avx+0x98>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,85,0,0,0 ; lea 0x55(%rip),%r9 # 3b7c <_sk_store_8888_avx+0x10c>
+ DB 76,141,13,85,0,0,0 ; lea 0x55(%rip),%r9 # 426c <_sk_store_8888_avx+0x10c>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7981,7 +8736,7 @@ _sk_store_8888_avx LABEL PROC
DB 196,67,121,22,68,186,8,2 ; vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4)
DB 196,67,121,22,68,186,4,1 ; vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4)
DB 196,65,121,126,4,186 ; vmovd %xmm8,(%r10,%rdi,4)
- DB 235,143 ; jmp 3b08 <_sk_store_8888_avx+0x98>
+ DB 235,143 ; jmp 41f8 <_sk_store_8888_avx+0x98>
DB 15,31,0 ; nopl (%rax)
DB 245 ; cmc
DB 255 ; (bad)
@@ -8017,7 +8772,7 @@ _sk_load_f16_avx LABEL PROC
DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp)
DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp)
DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp)
- DB 15,133,143,2,0,0 ; jne 3e53 <_sk_load_f16_avx+0x2bb>
+ DB 15,133,143,2,0,0 ; jne 4543 <_sk_load_f16_avx+0x2bb>
DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8
DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2
DB 197,249,16,76,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm1
@@ -8035,13 +8790,13 @@ _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 196,98,125,24,37,243,28,0,0 ; vbroadcastss 0x1cf3(%rip),%ymm12 # 5910 <_sk_callback_avx+0x3d9>
+ DB 196,98,125,24,37,51,29,0,0 ; vbroadcastss 0x1d33(%rip),%ymm12 # 6040 <_sk_callback_avx+0x419>
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 196,98,121,24,29,223,28,0,0 ; vbroadcastss 0x1cdf(%rip),%xmm11 # 5914 <_sk_callback_avx+0x3dd>
+ DB 196,98,121,24,29,31,29,0,0 ; vbroadcastss 0x1d1f(%rip),%xmm11 # 6044 <_sk_callback_avx+0x41d>
DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3
- DB 196,98,121,24,45,213,28,0,0 ; vbroadcastss 0x1cd5(%rip),%xmm13 # 5918 <_sk_callback_avx+0x3e1>
+ DB 196,98,121,24,45,21,29,0,0 ; vbroadcastss 0x1d15(%rip),%xmm13 # 6048 <_sk_callback_avx+0x421>
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
@@ -8055,7 +8810,7 @@ _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 196,226,121,24,29,139,28,0,0 ; vbroadcastss 0x1c8b(%rip),%xmm3 # 591c <_sk_callback_avx+0x3e5>
+ DB 196,226,121,24,29,203,28,0,0 ; vbroadcastss 0x1ccb(%rip),%xmm3 # 604c <_sk_callback_avx+0x425>
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
@@ -8148,29 +8903,29 @@ _sk_load_f16_avx LABEL PROC
DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,79 ; je 3eb2 <_sk_load_f16_avx+0x31a>
+ DB 116,79 ; je 45a2 <_sk_load_f16_avx+0x31a>
DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,67 ; jb 3eb2 <_sk_load_f16_avx+0x31a>
+ DB 114,67 ; jb 45a2 <_sk_load_f16_avx+0x31a>
DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,68 ; je 3ebf <_sk_load_f16_avx+0x327>
+ DB 116,68 ; je 45af <_sk_load_f16_avx+0x327>
DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,56 ; jb 3ebf <_sk_load_f16_avx+0x327>
+ DB 114,56 ; jb 45af <_sk_load_f16_avx+0x327>
DB 197,251,16,76,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,68,253,255,255 ; je 3bdb <_sk_load_f16_avx+0x43>
+ DB 15,132,68,253,255,255 ; je 42cb <_sk_load_f16_avx+0x43>
DB 197,241,22,76,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,52,253,255,255 ; jb 3bdb <_sk_load_f16_avx+0x43>
+ DB 15,130,52,253,255,255 ; jb 42cb <_sk_load_f16_avx+0x43>
DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9
- DB 233,41,253,255,255 ; jmpq 3bdb <_sk_load_f16_avx+0x43>
+ DB 233,41,253,255,255 ; jmpq 42cb <_sk_load_f16_avx+0x43>
DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,28,253,255,255 ; jmpq 3bdb <_sk_load_f16_avx+0x43>
+ DB 233,28,253,255,255 ; jmpq 42cb <_sk_load_f16_avx+0x43>
DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1
- DB 233,19,253,255,255 ; jmpq 3bdb <_sk_load_f16_avx+0x43>
+ DB 233,19,253,255,255 ; jmpq 42cb <_sk_load_f16_avx+0x43>
PUBLIC _sk_gather_f16_avx
_sk_gather_f16_avx LABEL PROC
@@ -8232,13 +8987,13 @@ _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 196,98,125,24,37,75,25,0,0 ; vbroadcastss 0x194b(%rip),%ymm12 # 5920 <_sk_callback_avx+0x3e9>
+ DB 196,98,125,24,37,139,25,0,0 ; vbroadcastss 0x198b(%rip),%ymm12 # 6050 <_sk_callback_avx+0x429>
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 196,98,121,24,29,55,25,0,0 ; vbroadcastss 0x1937(%rip),%xmm11 # 5924 <_sk_callback_avx+0x3ed>
+ DB 196,98,121,24,29,119,25,0,0 ; vbroadcastss 0x1977(%rip),%xmm11 # 6054 <_sk_callback_avx+0x42d>
DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3
- DB 196,98,121,24,45,45,25,0,0 ; vbroadcastss 0x192d(%rip),%xmm13 # 5928 <_sk_callback_avx+0x3f1>
+ DB 196,98,121,24,45,109,25,0,0 ; vbroadcastss 0x196d(%rip),%xmm13 # 6058 <_sk_callback_avx+0x431>
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
@@ -8252,7 +9007,7 @@ _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 196,226,121,24,29,227,24,0,0 ; vbroadcastss 0x18e3(%rip),%xmm3 # 592c <_sk_callback_avx+0x3f5>
+ DB 196,226,121,24,29,35,25,0,0 ; vbroadcastss 0x1923(%rip),%xmm3 # 605c <_sk_callback_avx+0x435>
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
@@ -8354,12 +9109,12 @@ _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 196,98,125,24,13,240,22,0,0 ; vbroadcastss 0x16f0(%rip),%ymm9 # 5930 <_sk_callback_avx+0x3f9>
+ DB 196,98,125,24,13,48,23,0,0 ; vbroadcastss 0x1730(%rip),%ymm9 # 6060 <_sk_callback_avx+0x439>
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 196,98,121,24,5,214,22,0,0 ; vbroadcastss 0x16d6(%rip),%xmm8 # 5934 <_sk_callback_avx+0x3fd>
+ DB 196,98,121,24,5,22,23,0,0 ; vbroadcastss 0x1716(%rip),%xmm8 # 6064 <_sk_callback_avx+0x43d>
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
@@ -8369,7 +9124,7 @@ _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 196,98,125,24,21,157,22,0,0 ; vbroadcastss 0x169d(%rip),%ymm10 # 5938 <_sk_callback_avx+0x401>
+ DB 196,98,125,24,21,221,22,0,0 ; vbroadcastss 0x16dd(%rip),%ymm10 # 6068 <_sk_callback_avx+0x441>
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
@@ -8451,7 +9206,7 @@ _sk_store_f16_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 117,75 ; jne 4482 <_sk_store_f16_avx+0x270>
+ DB 117,75 ; jne 4b72 <_sk_store_f16_avx+0x270>
DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8)
DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8)
DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -8467,22 +9222,22 @@ _sk_store_f16_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,193 ; je 444e <_sk_store_f16_avx+0x23c>
+ DB 116,193 ; je 4b3e <_sk_store_f16_avx+0x23c>
DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,181 ; jb 444e <_sk_store_f16_avx+0x23c>
+ DB 114,181 ; jb 4b3e <_sk_store_f16_avx+0x23c>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,173 ; je 444e <_sk_store_f16_avx+0x23c>
+ DB 116,173 ; je 4b3e <_sk_store_f16_avx+0x23c>
DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,161 ; jb 444e <_sk_store_f16_avx+0x23c>
+ DB 114,161 ; jb 4b3e <_sk_store_f16_avx+0x23c>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,153 ; je 444e <_sk_store_f16_avx+0x23c>
+ DB 116,153 ; je 4b3e <_sk_store_f16_avx+0x23c>
DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,141 ; jb 444e <_sk_store_f16_avx+0x23c>
+ DB 114,141 ; jb 4b3e <_sk_store_f16_avx+0x23c>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,133 ; jmp 444e <_sk_store_f16_avx+0x23c>
+ DB 235,133 ; jmp 4b3e <_sk_store_f16_avx+0x23c>
PUBLIC _sk_load_u16_be_avx
_sk_load_u16_be_avx LABEL PROC
@@ -8490,7 +9245,7 @@ _sk_load_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,253,0,0,0 ; jne 45dc <_sk_load_u16_be_avx+0x113>
+ DB 15,133,253,0,0,0 ; jne 4ccc <_sk_load_u16_be_avx+0x113>
DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8
DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2
DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3
@@ -8512,7 +9267,7 @@ _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 196,98,125,24,29,236,19,0,0 ; vbroadcastss 0x13ec(%rip),%ymm11 # 593c <_sk_callback_avx+0x405>
+ DB 196,98,125,24,29,44,20,0,0 ; vbroadcastss 0x142c(%rip),%ymm11 # 606c <_sk_callback_avx+0x445>
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
@@ -8546,29 +9301,29 @@ _sk_load_u16_be_avx LABEL PROC
DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 4642 <_sk_load_u16_be_avx+0x179>
+ DB 116,85 ; je 4d32 <_sk_load_u16_be_avx+0x179>
DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 4642 <_sk_load_u16_be_avx+0x179>
+ DB 114,72 ; jb 4d32 <_sk_load_u16_be_avx+0x179>
DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 464f <_sk_load_u16_be_avx+0x186>
+ DB 116,72 ; je 4d3f <_sk_load_u16_be_avx+0x186>
DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 464f <_sk_load_u16_be_avx+0x186>
+ DB 114,59 ; jb 4d3f <_sk_load_u16_be_avx+0x186>
DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,213,254,255,255 ; je 44fa <_sk_load_u16_be_avx+0x31>
+ DB 15,132,213,254,255,255 ; je 4bea <_sk_load_u16_be_avx+0x31>
DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,196,254,255,255 ; jb 44fa <_sk_load_u16_be_avx+0x31>
+ DB 15,130,196,254,255,255 ; jb 4bea <_sk_load_u16_be_avx+0x31>
DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9
- DB 233,184,254,255,255 ; jmpq 44fa <_sk_load_u16_be_avx+0x31>
+ DB 233,184,254,255,255 ; jmpq 4bea <_sk_load_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,171,254,255,255 ; jmpq 44fa <_sk_load_u16_be_avx+0x31>
+ DB 233,171,254,255,255 ; jmpq 4bea <_sk_load_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,162,254,255,255 ; jmpq 44fa <_sk_load_u16_be_avx+0x31>
+ DB 233,162,254,255,255 ; jmpq 4bea <_sk_load_u16_be_avx+0x31>
PUBLIC _sk_load_rgb_u16_be_avx
_sk_load_rgb_u16_be_avx LABEL PROC
@@ -8576,7 +9331,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,243,0,0,0 ; jne 475d <_sk_load_rgb_u16_be_avx+0x105>
+ DB 15,133,243,0,0,0 ; jne 4e4d <_sk_load_rgb_u16_be_avx+0x105>
DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0
DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2
DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1
@@ -8603,7 +9358,7 @@ _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 196,98,125,24,29,76,18,0,0 ; vbroadcastss 0x124c(%rip),%ymm11 # 5940 <_sk_callback_avx+0x409>
+ DB 196,98,125,24,29,140,18,0,0 ; vbroadcastss 0x128c(%rip),%ymm11 # 6070 <_sk_callback_avx+0x449>
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
@@ -8624,48 +9379,48 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,233,17,0,0 ; vbroadcastss 0x11e9(%rip),%ymm3 # 5944 <_sk_callback_avx+0x40d>
+ DB 196,226,125,24,29,41,18,0,0 ; vbroadcastss 0x1229(%rip),%ymm3 # 6074 <_sk_callback_avx+0x44d>
DB 255,224 ; jmpq *%rax
DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0
DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 4776 <_sk_load_rgb_u16_be_avx+0x11e>
- DB 233,40,255,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,5 ; jne 4e66 <_sk_load_rgb_u16_be_avx+0x11e>
+ DB 233,40,255,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1
DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 47a5 <_sk_load_rgb_u16_be_avx+0x14d>
+ DB 114,26 ; jb 4e95 <_sk_load_rgb_u16_be_avx+0x14d>
DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1
DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 47aa <_sk_load_rgb_u16_be_avx+0x152>
- DB 233,249,254,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,244,254,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 4e9a <_sk_load_rgb_u16_be_avx+0x152>
+ DB 233,249,254,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,244,254,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1
DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 47d9 <_sk_load_rgb_u16_be_avx+0x181>
+ DB 114,26 ; jb 4ec9 <_sk_load_rgb_u16_be_avx+0x181>
DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1
DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 47de <_sk_load_rgb_u16_be_avx+0x186>
- DB 233,197,254,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,192,254,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 4ece <_sk_load_rgb_u16_be_avx+0x186>
+ DB 233,197,254,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,192,254,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3
DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 4807 <_sk_load_rgb_u16_be_avx+0x1af>
+ DB 114,20 ; jb 4ef7 <_sk_load_rgb_u16_be_avx+0x1af>
DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3
DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
- DB 233,151,254,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,146,254,255,255 ; jmpq 469e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,151,254,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,146,254,255,255 ; jmpq 4d8e <_sk_load_rgb_u16_be_avx+0x46>
PUBLIC _sk_store_u16_be_avx
_sk_store_u16_be_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax
- DB 196,98,125,24,5,38,17,0,0 ; vbroadcastss 0x1126(%rip),%ymm8 # 5948 <_sk_callback_avx+0x411>
+ DB 196,98,125,24,5,102,17,0,0 ; vbroadcastss 0x1166(%rip),%ymm8 # 6078 <_sk_callback_avx+0x451>
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
@@ -8703,7 +9458,7 @@ _sk_store_u16_be_avx LABEL PROC
DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9
DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,31 ; jne 4906 <_sk_store_u16_be_avx+0xfa>
+ DB 117,31 ; jne 4ff6 <_sk_store_u16_be_avx+0xfa>
DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2)
DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2)
DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2)
@@ -8712,31 +9467,31 @@ _sk_store_u16_be_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 116,240 ; je 4ff2 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 114,227 ; jb 4ff2 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2)
- DB 116,218 ; je 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 116,218 ; je 4ff2 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 114,205 ; jb 4ff2 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2)
- DB 116,196 ; je 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 116,196 ; je 4ff2 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,183 ; jb 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 114,183 ; jb 4ff2 <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2)
- DB 235,174 ; jmp 4902 <_sk_store_u16_be_avx+0xf6>
+ DB 235,174 ; jmp 4ff2 <_sk_store_u16_be_avx+0xf6>
PUBLIC _sk_load_f32_avx
_sk_load_f32_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 119,110 ; ja 49ca <_sk_load_f32_avx+0x76>
+ DB 119,110 ; ja 50ba <_sk_load_f32_avx+0x76>
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
- DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 49f4 <_sk_load_f32_avx+0xa0>
+ DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 50e4 <_sk_load_f32_avx+0xa0>
DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8793,7 +9548,7 @@ _sk_store_f32_avx LABEL PROC
DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8
DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11
DB 72,133,201 ; test %rcx,%rcx
- DB 117,55 ; jne 4a81 <_sk_store_f32_avx+0x6d>
+ DB 117,55 ; jne 5171 <_sk_store_f32_avx+0x6d>
DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
@@ -8806,22 +9561,22 @@ _sk_store_f32_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 4a7d <_sk_store_f32_avx+0x69>
+ DB 116,240 ; je 516d <_sk_store_f32_avx+0x69>
DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 4a7d <_sk_store_f32_avx+0x69>
+ DB 114,227 ; jb 516d <_sk_store_f32_avx+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 4a7d <_sk_store_f32_avx+0x69>
+ DB 116,218 ; je 516d <_sk_store_f32_avx+0x69>
DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 4a7d <_sk_store_f32_avx+0x69>
+ DB 114,205 ; jb 516d <_sk_store_f32_avx+0x69>
DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
- DB 116,195 ; je 4a7d <_sk_store_f32_avx+0x69>
+ DB 116,195 ; je 516d <_sk_store_f32_avx+0x69>
DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,181 ; jb 4a7d <_sk_store_f32_avx+0x69>
+ DB 114,181 ; jb 516d <_sk_store_f32_avx+0x69>
DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
- DB 235,171 ; jmp 4a7d <_sk_store_f32_avx+0x69>
+ DB 235,171 ; jmp 516d <_sk_store_f32_avx+0x69>
PUBLIC _sk_clamp_x_avx
_sk_clamp_x_avx LABEL PROC
@@ -8943,12 +9698,12 @@ _sk_mirror_y_avx LABEL PROC
PUBLIC _sk_luminance_to_alpha_avx
_sk_luminance_to_alpha_avx LABEL PROC
- DB 196,226,125,24,29,175,12,0,0 ; vbroadcastss 0xcaf(%rip),%ymm3 # 594c <_sk_callback_avx+0x415>
+ DB 196,226,125,24,29,239,12,0,0 ; vbroadcastss 0xcef(%rip),%ymm3 # 607c <_sk_callback_avx+0x455>
DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0
- DB 196,226,125,24,29,166,12,0,0 ; vbroadcastss 0xca6(%rip),%ymm3 # 5950 <_sk_callback_avx+0x419>
+ DB 196,226,125,24,29,230,12,0,0 ; vbroadcastss 0xce6(%rip),%ymm3 # 6080 <_sk_callback_avx+0x459>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
- DB 196,226,125,24,13,153,12,0,0 ; vbroadcastss 0xc99(%rip),%ymm1 # 5954 <_sk_callback_avx+0x41d>
+ DB 196,226,125,24,13,217,12,0,0 ; vbroadcastss 0xcd9(%rip),%ymm1 # 6084 <_sk_callback_avx+0x45d>
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
@@ -9116,7 +9871,7 @@ _sk_linear_gradient_avx LABEL PROC
DB 196,226,125,24,88,28 ; vbroadcastss 0x1c(%rax),%ymm3
DB 76,139,0 ; mov (%rax),%r8
DB 77,133,192 ; test %r8,%r8
- DB 15,132,146,0,0,0 ; je 5011 <_sk_linear_gradient_avx+0xb8>
+ DB 15,132,146,0,0,0 ; je 5701 <_sk_linear_gradient_avx+0xb8>
DB 72,139,64,8 ; mov 0x8(%rax),%rax
DB 72,131,192,32 ; add $0x20,%rax
DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12
@@ -9143,8 +9898,8 @@ _sk_linear_gradient_avx LABEL PROC
DB 196,227,13,74,219,208 ; vblendvps %ymm13,%ymm3,%ymm14,%ymm3
DB 72,131,192,36 ; add $0x24,%rax
DB 73,255,200 ; dec %r8
- DB 117,140 ; jne 4f9b <_sk_linear_gradient_avx+0x42>
- DB 235,20 ; jmp 5025 <_sk_linear_gradient_avx+0xcc>
+ DB 117,140 ; jne 568b <_sk_linear_gradient_avx+0x42>
+ DB 235,20 ; jmp 5715 <_sk_linear_gradient_avx+0xcc>
DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11
DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10
DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9
@@ -9195,27 +9950,27 @@ _sk_xy_to_polar_unit_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 196,98,125,24,45,126,8,0,0 ; vbroadcastss 0x87e(%rip),%ymm13 # 5958 <_sk_callback_avx+0x421>
+ DB 196,98,125,24,45,190,8,0,0 ; vbroadcastss 0x8be(%rip),%ymm13 # 6088 <_sk_callback_avx+0x461>
DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13
- DB 196,98,125,24,53,116,8,0,0 ; vbroadcastss 0x874(%rip),%ymm14 # 595c <_sk_callback_avx+0x425>
+ DB 196,98,125,24,53,180,8,0,0 ; vbroadcastss 0x8b4(%rip),%ymm14 # 608c <_sk_callback_avx+0x465>
DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13
DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13
- DB 196,98,125,24,53,101,8,0,0 ; vbroadcastss 0x865(%rip),%ymm14 # 5960 <_sk_callback_avx+0x429>
+ DB 196,98,125,24,53,165,8,0,0 ; vbroadcastss 0x8a5(%rip),%ymm14 # 6090 <_sk_callback_avx+0x469>
DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13
DB 196,65,28,89,229 ; vmulps %ymm13,%ymm12,%ymm12
- DB 196,98,125,24,45,86,8,0,0 ; vbroadcastss 0x856(%rip),%ymm13 # 5964 <_sk_callback_avx+0x42d>
+ DB 196,98,125,24,45,150,8,0,0 ; vbroadcastss 0x896(%rip),%ymm13 # 6094 <_sk_callback_avx+0x46d>
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 196,98,125,24,21,65,8,0,0 ; vbroadcastss 0x841(%rip),%ymm10 # 5968 <_sk_callback_avx+0x431>
+ DB 196,98,125,24,21,129,8,0,0 ; vbroadcastss 0x881(%rip),%ymm10 # 6098 <_sk_callback_avx+0x471>
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 196,98,125,24,21,43,8,0,0 ; vbroadcastss 0x82b(%rip),%ymm10 # 596c <_sk_callback_avx+0x435>
+ DB 196,98,125,24,21,107,8,0,0 ; vbroadcastss 0x86b(%rip),%ymm10 # 609c <_sk_callback_avx+0x475>
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 196,98,125,24,21,21,8,0,0 ; vbroadcastss 0x815(%rip),%ymm10 # 5970 <_sk_callback_avx+0x439>
+ DB 196,98,125,24,21,85,8,0,0 ; vbroadcastss 0x855(%rip),%ymm10 # 60a0 <_sk_callback_avx+0x479>
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
@@ -9226,7 +9981,7 @@ _sk_xy_to_polar_unit_avx LABEL PROC
PUBLIC _sk_save_xy_avx
_sk_save_xy_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,244,7,0,0 ; vbroadcastss 0x7f4(%rip),%ymm8 # 5974 <_sk_callback_avx+0x43d>
+ DB 196,98,125,24,5,52,8,0,0 ; vbroadcastss 0x834(%rip),%ymm8 # 60a4 <_sk_callback_avx+0x47d>
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
@@ -9259,9 +10014,9 @@ _sk_accumulate_avx LABEL PROC
PUBLIC _sk_bilinear_nx_avx
_sk_bilinear_nx_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,128,7,0,0 ; vbroadcastss 0x780(%rip),%ymm0 # 5978 <_sk_callback_avx+0x441>
+ DB 196,226,125,24,5,192,7,0,0 ; vbroadcastss 0x7c0(%rip),%ymm0 # 60a8 <_sk_callback_avx+0x481>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,119,7,0,0 ; vbroadcastss 0x777(%rip),%ymm8 # 597c <_sk_callback_avx+0x445>
+ DB 196,98,125,24,5,183,7,0,0 ; vbroadcastss 0x7b7(%rip),%ymm8 # 60ac <_sk_callback_avx+0x485>
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
@@ -9270,7 +10025,7 @@ _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 196,226,125,24,5,95,7,0,0 ; vbroadcastss 0x75f(%rip),%ymm0 # 5980 <_sk_callback_avx+0x449>
+ DB 196,226,125,24,5,159,7,0,0 ; vbroadcastss 0x79f(%rip),%ymm0 # 60b0 <_sk_callback_avx+0x489>
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)
@@ -9280,9 +10035,9 @@ _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 196,226,125,24,13,67,7,0,0 ; vbroadcastss 0x743(%rip),%ymm1 # 5984 <_sk_callback_avx+0x44d>
+ DB 196,226,125,24,13,131,7,0,0 ; vbroadcastss 0x783(%rip),%ymm1 # 60b4 <_sk_callback_avx+0x48d>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,57,7,0,0 ; vbroadcastss 0x739(%rip),%ymm8 # 5988 <_sk_callback_avx+0x451>
+ DB 196,98,125,24,5,121,7,0,0 ; vbroadcastss 0x779(%rip),%ymm8 # 60b8 <_sk_callback_avx+0x491>
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
@@ -9291,7 +10046,7 @@ _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 196,226,125,24,13,33,7,0,0 ; vbroadcastss 0x721(%rip),%ymm1 # 598c <_sk_callback_avx+0x455>
+ DB 196,226,125,24,13,97,7,0,0 ; vbroadcastss 0x761(%rip),%ymm1 # 60bc <_sk_callback_avx+0x495>
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)
@@ -9301,14 +10056,14 @@ _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 196,226,125,24,5,4,7,0,0 ; vbroadcastss 0x704(%rip),%ymm0 # 5990 <_sk_callback_avx+0x459>
+ DB 196,226,125,24,5,68,7,0,0 ; vbroadcastss 0x744(%rip),%ymm0 # 60c0 <_sk_callback_avx+0x499>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,251,6,0,0 ; vbroadcastss 0x6fb(%rip),%ymm8 # 5994 <_sk_callback_avx+0x45d>
+ DB 196,98,125,24,5,59,7,0,0 ; vbroadcastss 0x73b(%rip),%ymm8 # 60c4 <_sk_callback_avx+0x49d>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,236,6,0,0 ; vbroadcastss 0x6ec(%rip),%ymm10 # 5998 <_sk_callback_avx+0x461>
+ DB 196,98,125,24,21,44,7,0,0 ; vbroadcastss 0x72c(%rip),%ymm10 # 60c8 <_sk_callback_avx+0x4a1>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,226,6,0,0 ; vbroadcastss 0x6e2(%rip),%ymm10 # 599c <_sk_callback_avx+0x465>
+ DB 196,98,125,24,21,34,7,0,0 ; vbroadcastss 0x722(%rip),%ymm10 # 60cc <_sk_callback_avx+0x4a5>
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)
@@ -9318,19 +10073,19 @@ _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 196,226,125,24,5,197,6,0,0 ; vbroadcastss 0x6c5(%rip),%ymm0 # 59a0 <_sk_callback_avx+0x469>
+ DB 196,226,125,24,5,5,7,0,0 ; vbroadcastss 0x705(%rip),%ymm0 # 60d0 <_sk_callback_avx+0x4a9>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,188,6,0,0 ; vbroadcastss 0x6bc(%rip),%ymm8 # 59a4 <_sk_callback_avx+0x46d>
+ DB 196,98,125,24,5,252,6,0,0 ; vbroadcastss 0x6fc(%rip),%ymm8 # 60d4 <_sk_callback_avx+0x4ad>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,178,6,0,0 ; vbroadcastss 0x6b2(%rip),%ymm9 # 59a8 <_sk_callback_avx+0x471>
+ DB 196,98,125,24,13,242,6,0,0 ; vbroadcastss 0x6f2(%rip),%ymm9 # 60d8 <_sk_callback_avx+0x4b1>
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,168,6,0,0 ; vbroadcastss 0x6a8(%rip),%ymm10 # 59ac <_sk_callback_avx+0x475>
+ DB 196,98,125,24,21,232,6,0,0 ; vbroadcastss 0x6e8(%rip),%ymm10 # 60dc <_sk_callback_avx+0x4b5>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,153,6,0,0 ; vbroadcastss 0x699(%rip),%ymm10 # 59b0 <_sk_callback_avx+0x479>
+ DB 196,98,125,24,21,217,6,0,0 ; vbroadcastss 0x6d9(%rip),%ymm10 # 60e0 <_sk_callback_avx+0x4b9>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,138,6,0,0 ; vbroadcastss 0x68a(%rip),%ymm9 # 59b4 <_sk_callback_avx+0x47d>
+ DB 196,98,125,24,13,202,6,0,0 ; vbroadcastss 0x6ca(%rip),%ymm9 # 60e4 <_sk_callback_avx+0x4bd>
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
@@ -9339,17 +10094,17 @@ _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 196,98,125,24,5,114,6,0,0 ; vbroadcastss 0x672(%rip),%ymm8 # 59b8 <_sk_callback_avx+0x481>
+ DB 196,98,125,24,5,178,6,0,0 ; vbroadcastss 0x6b2(%rip),%ymm8 # 60e8 <_sk_callback_avx+0x4c1>
DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0
DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9
- DB 196,98,125,24,21,100,6,0,0 ; vbroadcastss 0x664(%rip),%ymm10 # 59bc <_sk_callback_avx+0x485>
+ DB 196,98,125,24,21,164,6,0,0 ; vbroadcastss 0x6a4(%rip),%ymm10 # 60ec <_sk_callback_avx+0x4c5>
DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10
- DB 196,98,125,24,29,90,6,0,0 ; vbroadcastss 0x65a(%rip),%ymm11 # 59c0 <_sk_callback_avx+0x489>
+ DB 196,98,125,24,29,154,6,0,0 ; vbroadcastss 0x69a(%rip),%ymm11 # 60f0 <_sk_callback_avx+0x4c9>
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 196,98,125,24,13,65,6,0,0 ; vbroadcastss 0x641(%rip),%ymm9 # 59c4 <_sk_callback_avx+0x48d>
+ DB 196,98,125,24,13,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm9 # 60f4 <_sk_callback_avx+0x4cd>
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
@@ -9358,13 +10113,13 @@ _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 196,226,125,24,5,41,6,0,0 ; vbroadcastss 0x629(%rip),%ymm0 # 59c8 <_sk_callback_avx+0x491>
+ DB 196,226,125,24,5,105,6,0,0 ; vbroadcastss 0x669(%rip),%ymm0 # 60f8 <_sk_callback_avx+0x4d1>
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 196,98,125,24,21,22,6,0,0 ; vbroadcastss 0x616(%rip),%ymm10 # 59cc <_sk_callback_avx+0x495>
+ DB 196,98,125,24,21,86,6,0,0 ; vbroadcastss 0x656(%rip),%ymm10 # 60fc <_sk_callback_avx+0x4d5>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,12,6,0,0 ; vbroadcastss 0x60c(%rip),%ymm10 # 59d0 <_sk_callback_avx+0x499>
+ DB 196,98,125,24,21,76,6,0,0 ; vbroadcastss 0x64c(%rip),%ymm10 # 6100 <_sk_callback_avx+0x4d9>
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)
@@ -9374,14 +10129,14 @@ _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 196,226,125,24,13,239,5,0,0 ; vbroadcastss 0x5ef(%rip),%ymm1 # 59d4 <_sk_callback_avx+0x49d>
+ DB 196,226,125,24,13,47,6,0,0 ; vbroadcastss 0x62f(%rip),%ymm1 # 6104 <_sk_callback_avx+0x4dd>
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 # 59d8 <_sk_callback_avx+0x4a1>
+ DB 196,98,125,24,5,37,6,0,0 ; vbroadcastss 0x625(%rip),%ymm8 # 6108 <_sk_callback_avx+0x4e1>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,214,5,0,0 ; vbroadcastss 0x5d6(%rip),%ymm10 # 59dc <_sk_callback_avx+0x4a5>
+ DB 196,98,125,24,21,22,6,0,0 ; vbroadcastss 0x616(%rip),%ymm10 # 610c <_sk_callback_avx+0x4e5>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,204,5,0,0 ; vbroadcastss 0x5cc(%rip),%ymm10 # 59e0 <_sk_callback_avx+0x4a9>
+ DB 196,98,125,24,21,12,6,0,0 ; vbroadcastss 0x60c(%rip),%ymm10 # 6110 <_sk_callback_avx+0x4e9>
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)
@@ -9391,19 +10146,19 @@ _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 196,226,125,24,13,175,5,0,0 ; vbroadcastss 0x5af(%rip),%ymm1 # 59e4 <_sk_callback_avx+0x4ad>
+ DB 196,226,125,24,13,239,5,0,0 ; vbroadcastss 0x5ef(%rip),%ymm1 # 6114 <_sk_callback_avx+0x4ed>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,165,5,0,0 ; vbroadcastss 0x5a5(%rip),%ymm8 # 59e8 <_sk_callback_avx+0x4b1>
+ DB 196,98,125,24,5,229,5,0,0 ; vbroadcastss 0x5e5(%rip),%ymm8 # 6118 <_sk_callback_avx+0x4f1>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,155,5,0,0 ; vbroadcastss 0x59b(%rip),%ymm9 # 59ec <_sk_callback_avx+0x4b5>
+ DB 196,98,125,24,13,219,5,0,0 ; vbroadcastss 0x5db(%rip),%ymm9 # 611c <_sk_callback_avx+0x4f5>
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,145,5,0,0 ; vbroadcastss 0x591(%rip),%ymm10 # 59f0 <_sk_callback_avx+0x4b9>
+ DB 196,98,125,24,21,209,5,0,0 ; vbroadcastss 0x5d1(%rip),%ymm10 # 6120 <_sk_callback_avx+0x4f9>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9
- DB 196,98,125,24,21,130,5,0,0 ; vbroadcastss 0x582(%rip),%ymm10 # 59f4 <_sk_callback_avx+0x4bd>
+ DB 196,98,125,24,21,194,5,0,0 ; vbroadcastss 0x5c2(%rip),%ymm10 # 6124 <_sk_callback_avx+0x4fd>
DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,115,5,0,0 ; vbroadcastss 0x573(%rip),%ymm9 # 59f8 <_sk_callback_avx+0x4c1>
+ DB 196,98,125,24,13,179,5,0,0 ; vbroadcastss 0x5b3(%rip),%ymm9 # 6128 <_sk_callback_avx+0x501>
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
@@ -9412,17 +10167,17 @@ _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 196,98,125,24,5,91,5,0,0 ; vbroadcastss 0x55b(%rip),%ymm8 # 59fc <_sk_callback_avx+0x4c5>
+ DB 196,98,125,24,5,155,5,0,0 ; vbroadcastss 0x59b(%rip),%ymm8 # 612c <_sk_callback_avx+0x505>
DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1
DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9
- DB 196,98,125,24,21,76,5,0,0 ; vbroadcastss 0x54c(%rip),%ymm10 # 5a00 <_sk_callback_avx+0x4c9>
+ DB 196,98,125,24,21,140,5,0,0 ; vbroadcastss 0x58c(%rip),%ymm10 # 6130 <_sk_callback_avx+0x509>
DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10
- DB 196,98,125,24,29,66,5,0,0 ; vbroadcastss 0x542(%rip),%ymm11 # 5a04 <_sk_callback_avx+0x4cd>
+ DB 196,98,125,24,29,130,5,0,0 ; vbroadcastss 0x582(%rip),%ymm11 # 6134 <_sk_callback_avx+0x50d>
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 196,98,125,24,13,41,5,0,0 ; vbroadcastss 0x529(%rip),%ymm9 # 5a08 <_sk_callback_avx+0x4d1>
+ DB 196,98,125,24,13,105,5,0,0 ; vbroadcastss 0x569(%rip),%ymm9 # 6138 <_sk_callback_avx+0x511>
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
@@ -9431,13 +10186,13 @@ _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 196,226,125,24,13,17,5,0,0 ; vbroadcastss 0x511(%rip),%ymm1 # 5a0c <_sk_callback_avx+0x4d5>
+ DB 196,226,125,24,13,81,5,0,0 ; vbroadcastss 0x551(%rip),%ymm1 # 613c <_sk_callback_avx+0x515>
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 196,98,125,24,21,253,4,0,0 ; vbroadcastss 0x4fd(%rip),%ymm10 # 5a10 <_sk_callback_avx+0x4d9>
+ DB 196,98,125,24,21,61,5,0,0 ; vbroadcastss 0x53d(%rip),%ymm10 # 6140 <_sk_callback_avx+0x519>
DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8
- DB 196,98,125,24,21,243,4,0,0 ; vbroadcastss 0x4f3(%rip),%ymm10 # 5a14 <_sk_callback_avx+0x4dd>
+ DB 196,98,125,24,21,51,5,0,0 ; vbroadcastss 0x533(%rip),%ymm10 # 6144 <_sk_callback_avx+0x51d>
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)
@@ -9548,6 +10303,30 @@ ALIGN 4
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,191,0,0,224 ; add %al,-0x1fffff41(%rax)
DB 64,0,0 ; 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 5df5 <.literal4+0xb5>
+ 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 5e05 <.literal4+0xc5>
+ 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 5e15 <.literal4+0xd5>
+ 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 5e25 <.literal4+0xe5>
+ 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)
@@ -9595,7 +10374,7 @@ 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 573d <.literal4+0xed>
+ DB 224,7 ; loopne 5e6d <.literal4+0x12d>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -9609,10 +10388,10 @@ 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 5768 <.literal4+0x118>
+ DB 127,0 ; jg 5e98 <.literal4+0x158>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 57e1 <.literal4+0x191>
+ DB 119,115 ; ja 5f11 <.literal4+0x1d1>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9626,10 +10405,10 @@ 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 579c <.literal4+0x14c>
+ DB 127,0 ; jg 5ecc <.literal4+0x18c>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 5815 <.literal4+0x1c5>
+ DB 119,115 ; ja 5f45 <.literal4+0x205>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9643,10 +10422,10 @@ 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 57d0 <.literal4+0x180>
+ DB 127,0 ; jg 5f00 <.literal4+0x1c0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 5849 <.literal4+0x1f9>
+ DB 119,115 ; ja 5f79 <.literal4+0x239>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9660,10 +10439,10 @@ 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 5804 <.literal4+0x1b4>
+ DB 127,0 ; jg 5f34 <.literal4+0x1f4>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 587d <.literal4+0x22d>
+ DB 119,115 ; ja 5fad <.literal4+0x26d>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9676,7 +10455,7 @@ 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 587b <.literal4+0x22b>
+ DB 127,67 ; jg 5fab <.literal4+0x26b>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -9688,10 +10467,10 @@ ALIGN 4
DB 190,80,128,3,62 ; mov $0x3e038050,%esi
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 589b <.literal4+0x24b>
+ DB 118,63 ; jbe 5fcb <.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 58af <.literal4+0x25f>
+ DB 127,67 ; jg 5fdf <.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)
@@ -9700,7 +10479,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 5891 <.literal4+0x241>
+ DB 224,7 ; loopne 5fc1 <.literal4+0x281>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -9712,7 +10491,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 58ad <.literal4+0x25d>
+ DB 224,7 ; loopne 5fdd <.literal4+0x29d>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -9723,7 +10502,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 5902 <.literal4+0x2b2>
+ DB 124,66 ; jl 6032 <.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)
@@ -9741,9 +10520,9 @@ 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 5945 <.literal4+0x2f5>
+ DB 112,65 ; jo 6075 <.literal4+0x335>
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 5953 <.literal4+0x303>
+ DB 127,67 ; jg 6083 <.literal4+0x343>
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)
@@ -9759,7 +10538,7 @@ ALIGN 4
DB 0,128,55,0,0,128 ; add %al,-0x7fffffc9(%rax)
DB 63 ; (bad)
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 5993 <.literal4+0x343>
+ DB 127,71 ; jg 60c3 <.literal4+0x383>
DB 208 ; (bad)
DB 179,89 ; mov $0x59,%bl
DB 62,89 ; ds pop %rcx
@@ -10007,7 +10786,7 @@ _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 15,40,21,49,58,0,0 ; movaps 0x3a31(%rip),%xmm2 # 3b40 <_sk_callback_sse41+0xaf>
+ DB 15,40,21,193,66,0,0 ; movaps 0x42c1(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0xb5>
DB 15,88,202 ; addps %xmm2,%xmm1
DB 15,16,2 ; movups (%rdx),%xmm0
DB 15,88,193 ; addps %xmm1,%xmm0
@@ -10016,7 +10795,7 @@ _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 15,40,21,32,58,0,0 ; movaps 0x3a20(%rip),%xmm2 # 3b50 <_sk_callback_sse41+0xbf>
+ DB 15,40,21,176,66,0,0 ; movaps 0x42b0(%rip),%xmm2 # 43e0 <_sk_callback_sse41+0xc5>
DB 15,87,219 ; xorps %xmm3,%xmm3
DB 15,87,228 ; xorps %xmm4,%xmm4
DB 15,87,237 ; xorps %xmm5,%xmm5
@@ -10037,14 +10816,14 @@ _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 102,68,15,111,21,229,57,0,0 ; movdqa 0x39e5(%rip),%xmm10 # 3b60 <_sk_callback_sse41+0xcf>
+ DB 102,68,15,111,21,117,66,0,0 ; movdqa 0x4275(%rip),%xmm10 # 43f0 <_sk_callback_sse41+0xd5>
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 102,68,15,111,37,209,57,0,0 ; movdqa 0x39d1(%rip),%xmm12 # 3b70 <_sk_callback_sse41+0xdf>
- DB 102,68,15,111,45,216,57,0,0 ; movdqa 0x39d8(%rip),%xmm13 # 3b80 <_sk_callback_sse41+0xef>
+ DB 102,68,15,111,37,97,66,0,0 ; movdqa 0x4261(%rip),%xmm12 # 4400 <_sk_callback_sse41+0xe5>
+ DB 102,68,15,111,45,104,66,0,0 ; movdqa 0x4268(%rip),%xmm13 # 4410 <_sk_callback_sse41+0xf5>
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
@@ -10060,8 +10839,8 @@ _sk_dither_sse41 LABEL PROC
DB 102,69,15,235,245 ; por %xmm13,%xmm14
DB 102,69,15,235,240 ; por %xmm8,%xmm14
DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8
- DB 68,15,89,5,147,57,0,0 ; mulps 0x3993(%rip),%xmm8 # 3b90 <_sk_callback_sse41+0xff>
- DB 68,15,88,5,155,57,0,0 ; addps 0x399b(%rip),%xmm8 # 3ba0 <_sk_callback_sse41+0x10f>
+ DB 68,15,89,5,35,66,0,0 ; mulps 0x4223(%rip),%xmm8 # 4420 <_sk_callback_sse41+0x105>
+ DB 68,15,88,5,43,66,0,0 ; addps 0x422b(%rip),%xmm8 # 4430 <_sk_callback_sse41+0x115>
DB 243,68,15,16,72,8 ; movss 0x8(%rax),%xmm9
DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9
DB 69,15,89,200 ; mulps %xmm8,%xmm9
@@ -10097,7 +10876,7 @@ _sk_clear_sse41 LABEL PROC
PUBLIC _sk_srcatop_sse41
_sk_srcatop_sse41 LABEL PROC
DB 15,89,199 ; mulps %xmm7,%xmm0
- DB 68,15,40,5,72,57,0,0 ; movaps 0x3948(%rip),%xmm8 # 3bb0 <_sk_callback_sse41+0x11f>
+ DB 68,15,40,5,216,65,0,0 ; movaps 0x41d8(%rip),%xmm8 # 4440 <_sk_callback_sse41+0x125>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,89,204 ; mulps %xmm4,%xmm9
@@ -10120,7 +10899,7 @@ 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 68,15,40,13,11,57,0,0 ; movaps 0x390b(%rip),%xmm9 # 3bc0 <_sk_callback_sse41+0x12f>
+ DB 68,15,40,13,155,65,0,0 ; movaps 0x419b(%rip),%xmm9 # 4450 <_sk_callback_sse41+0x135>
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,88,192 ; addps %xmm8,%xmm0
@@ -10161,7 +10940,7 @@ _sk_dstin_sse41 LABEL PROC
PUBLIC _sk_srcout_sse41
_sk_srcout_sse41 LABEL PROC
- DB 68,15,40,5,175,56,0,0 ; movaps 0x38af(%rip),%xmm8 # 3bd0 <_sk_callback_sse41+0x13f>
+ DB 68,15,40,5,63,65,0,0 ; movaps 0x413f(%rip),%xmm8 # 4460 <_sk_callback_sse41+0x145>
DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
@@ -10172,7 +10951,7 @@ _sk_srcout_sse41 LABEL PROC
PUBLIC _sk_dstout_sse41
_sk_dstout_sse41 LABEL PROC
- DB 68,15,40,5,159,56,0,0 ; movaps 0x389f(%rip),%xmm8 # 3be0 <_sk_callback_sse41+0x14f>
+ DB 68,15,40,5,47,65,0,0 ; movaps 0x412f(%rip),%xmm8 # 4470 <_sk_callback_sse41+0x155>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 15,89,196 ; mulps %xmm4,%xmm0
@@ -10187,7 +10966,7 @@ _sk_dstout_sse41 LABEL PROC
PUBLIC _sk_srcover_sse41
_sk_srcover_sse41 LABEL PROC
- DB 68,15,40,5,130,56,0,0 ; movaps 0x3882(%rip),%xmm8 # 3bf0 <_sk_callback_sse41+0x15f>
+ DB 68,15,40,5,18,65,0,0 ; movaps 0x4112(%rip),%xmm8 # 4480 <_sk_callback_sse41+0x165>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,89,204 ; mulps %xmm4,%xmm9
@@ -10205,7 +10984,7 @@ _sk_srcover_sse41 LABEL PROC
PUBLIC _sk_dstover_sse41
_sk_dstover_sse41 LABEL PROC
- DB 68,15,40,5,86,56,0,0 ; movaps 0x3856(%rip),%xmm8 # 3c00 <_sk_callback_sse41+0x16f>
+ DB 68,15,40,5,230,64,0,0 ; movaps 0x40e6(%rip),%xmm8 # 4490 <_sk_callback_sse41+0x175>
DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -10229,7 +11008,7 @@ _sk_modulate_sse41 LABEL PROC
PUBLIC _sk_multiply_sse41
_sk_multiply_sse41 LABEL PROC
- DB 68,15,40,5,42,56,0,0 ; movaps 0x382a(%rip),%xmm8 # 3c10 <_sk_callback_sse41+0x17f>
+ DB 68,15,40,5,186,64,0,0 ; movaps 0x40ba(%rip),%xmm8 # 44a0 <_sk_callback_sse41+0x185>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 69,15,40,209 ; movaps %xmm9,%xmm10
@@ -10299,7 +11078,7 @@ _sk_screen_sse41 LABEL PROC
PUBLIC _sk_xor__sse41
_sk_xor__sse41 LABEL PROC
DB 68,15,40,195 ; movaps %xmm3,%xmm8
- DB 15,40,29,91,55,0,0 ; movaps 0x375b(%rip),%xmm3 # 3c20 <_sk_callback_sse41+0x18f>
+ DB 15,40,29,235,63,0,0 ; movaps 0x3feb(%rip),%xmm3 # 44b0 <_sk_callback_sse41+0x195>
DB 68,15,40,203 ; movaps %xmm3,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 65,15,89,193 ; mulps %xmm9,%xmm0
@@ -10345,7 +11124,7 @@ _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 15,40,21,198,54,0,0 ; movaps 0x36c6(%rip),%xmm2 # 3c30 <_sk_callback_sse41+0x19f>
+ DB 15,40,21,86,63,0,0 ; movaps 0x3f56(%rip),%xmm2 # 44c0 <_sk_callback_sse41+0x1a5>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -10377,7 +11156,7 @@ _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 15,40,21,107,54,0,0 ; movaps 0x366b(%rip),%xmm2 # 3c40 <_sk_callback_sse41+0x1af>
+ DB 15,40,21,251,62,0,0 ; movaps 0x3efb(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0x1b5>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -10412,7 +11191,7 @@ _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 15,40,21,5,54,0,0 ; movaps 0x3605(%rip),%xmm2 # 3c50 <_sk_callback_sse41+0x1bf>
+ DB 15,40,21,149,62,0,0 ; movaps 0x3e95(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0x1c5>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -10437,7 +11216,7 @@ _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 15,40,13,198,53,0,0 ; movaps 0x35c6(%rip),%xmm1 # 3c60 <_sk_callback_sse41+0x1cf>
+ DB 15,40,13,86,62,0,0 ; movaps 0x3e56(%rip),%xmm1 # 44f0 <_sk_callback_sse41+0x1d5>
DB 15,92,203 ; subps %xmm3,%xmm1
DB 15,89,207 ; mulps %xmm7,%xmm1
DB 15,88,217 ; addps %xmm1,%xmm3
@@ -10449,7 +11228,7 @@ _sk_exclusion_sse41 LABEL PROC
PUBLIC _sk_colorburn_sse41
_sk_colorburn_sse41 LABEL PROC
DB 68,15,40,192 ; movaps %xmm0,%xmm8
- DB 68,15,40,21,181,53,0,0 ; movaps 0x35b5(%rip),%xmm10 # 3c70 <_sk_callback_sse41+0x1df>
+ DB 68,15,40,21,69,62,0,0 ; movaps 0x3e45(%rip),%xmm10 # 4500 <_sk_callback_sse41+0x1e5>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 69,15,40,203 ; movaps %xmm11,%xmm9
@@ -10529,7 +11308,7 @@ _sk_colorburn_sse41 LABEL PROC
PUBLIC _sk_colordodge_sse41
_sk_colordodge_sse41 LABEL PROC
DB 68,15,40,192 ; movaps %xmm0,%xmm8
- DB 68,15,40,21,147,52,0,0 ; movaps 0x3493(%rip),%xmm10 # 3c80 <_sk_callback_sse41+0x1ef>
+ DB 68,15,40,21,35,61,0,0 ; movaps 0x3d23(%rip),%xmm10 # 4510 <_sk_callback_sse41+0x1f5>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 69,15,40,227 ; movaps %xmm11,%xmm12
@@ -10610,7 +11389,7 @@ _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 68,15,40,21,105,51,0,0 ; movaps 0x3369(%rip),%xmm10 # 3c90 <_sk_callback_sse41+0x1ff>
+ DB 68,15,40,21,249,59,0,0 ; movaps 0x3bf9(%rip),%xmm10 # 4520 <_sk_callback_sse41+0x205>
DB 65,15,40,234 ; movaps %xmm10,%xmm5
DB 15,92,239 ; subps %xmm7,%xmm5
DB 15,40,197 ; movaps %xmm5,%xmm0
@@ -10692,7 +11471,7 @@ 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 68,15,40,21,75,50,0,0 ; movaps 0x324b(%rip),%xmm10 # 3ca0 <_sk_callback_sse41+0x20f>
+ DB 68,15,40,21,219,58,0,0 ; movaps 0x3adb(%rip),%xmm10 # 4530 <_sk_callback_sse41+0x215>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 65,15,40,195 ; movaps %xmm11,%xmm0
@@ -10776,7 +11555,7 @@ _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 15,40,13,30,49,0,0 ; movaps 0x311e(%rip),%xmm1 # 3cb0 <_sk_callback_sse41+0x21f>
+ DB 15,40,13,174,57,0,0 ; movaps 0x39ae(%rip),%xmm1 # 4540 <_sk_callback_sse41+0x225>
DB 68,15,40,209 ; movaps %xmm1,%xmm10
DB 68,15,92,208 ; subps %xmm0,%xmm10
DB 68,15,40,240 ; movaps %xmm0,%xmm14
@@ -10789,10 +11568,10 @@ _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 68,15,40,45,252,48,0,0 ; movaps 0x30fc(%rip),%xmm13 # 3cc0 <_sk_callback_sse41+0x22f>
+ DB 68,15,40,45,140,57,0,0 ; movaps 0x398c(%rip),%xmm13 # 4550 <_sk_callback_sse41+0x235>
DB 69,15,88,245 ; addps %xmm13,%xmm14
DB 68,15,89,242 ; mulps %xmm2,%xmm14
- DB 68,15,40,37,252,48,0,0 ; movaps 0x30fc(%rip),%xmm12 # 3cd0 <_sk_callback_sse41+0x23f>
+ DB 68,15,40,37,140,57,0,0 ; movaps 0x398c(%rip),%xmm12 # 4560 <_sk_callback_sse41+0x245>
DB 69,15,89,252 ; mulps %xmm12,%xmm15
DB 69,15,88,254 ; addps %xmm14,%xmm15
DB 15,40,198 ; movaps %xmm6,%xmm0
@@ -10925,6 +11704,567 @@ _sk_softlight_sse41 LABEL PROC
DB 72,131,196,72 ; add $0x48,%rsp
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_hue_sse41
+_sk_hue_sse41 LABEL PROC
+ DB 72,131,236,40 ; sub $0x28,%rsp
+ DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp)
+ DB 68,15,40,214 ; movaps %xmm6,%xmm10
+ DB 68,15,40,229 ; movaps %xmm5,%xmm12
+ DB 15,40,252 ; movaps %xmm4,%xmm7
+ DB 15,40,243 ; movaps %xmm3,%xmm6
+ DB 68,15,40,192 ; movaps %xmm0,%xmm8
+ DB 69,15,87,246 ; xorps %xmm14,%xmm14
+ DB 68,15,40,45,149,55,0,0 ; movaps 0x3795(%rip),%xmm13 # 4570 <_sk_callback_sse41+0x255>
+ DB 65,15,40,221 ; movaps %xmm13,%xmm3
+ DB 15,94,222 ; divps %xmm6,%xmm3
+ DB 15,40,198 ; movaps %xmm6,%xmm0
+ DB 65,15,194,198,4 ; cmpneqps %xmm14,%xmm0
+ DB 15,84,195 ; andps %xmm3,%xmm0
+ DB 68,15,89,192 ; mulps %xmm0,%xmm8
+ DB 15,89,200 ; mulps %xmm0,%xmm1
+ DB 15,89,194 ; mulps %xmm2,%xmm0
+ DB 69,15,40,204 ; movaps %xmm12,%xmm9
+ DB 69,15,40,220 ; movaps %xmm12,%xmm11
+ DB 65,15,40,228 ; movaps %xmm12,%xmm4
+ DB 65,15,40,234 ; movaps %xmm10,%xmm5
+ DB 15,95,229 ; maxps %xmm5,%xmm4
+ DB 15,40,215 ; movaps %xmm7,%xmm2
+ DB 68,15,40,215 ; movaps %xmm7,%xmm10
+ DB 15,40,223 ; movaps %xmm7,%xmm3
+ DB 15,95,220 ; maxps %xmm4,%xmm3
+ DB 68,15,93,205 ; minps %xmm5,%xmm9
+ DB 15,41,44,36 ; movaps %xmm5,(%rsp)
+ DB 65,15,93,209 ; minps %xmm9,%xmm2
+ DB 15,92,218 ; subps %xmm2,%xmm3
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 15,93,208 ; minps %xmm0,%xmm2
+ DB 65,15,40,224 ; movaps %xmm8,%xmm4
+ DB 15,93,226 ; minps %xmm2,%xmm4
+ DB 68,15,40,201 ; movaps %xmm1,%xmm9
+ DB 68,15,95,200 ; maxps %xmm0,%xmm9
+ DB 65,15,40,208 ; movaps %xmm8,%xmm2
+ DB 65,15,95,209 ; maxps %xmm9,%xmm2
+ DB 15,92,212 ; subps %xmm4,%xmm2
+ DB 68,15,92,196 ; subps %xmm4,%xmm8
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 68,15,94,194 ; divps %xmm2,%xmm8
+ DB 15,92,204 ; subps %xmm4,%xmm1
+ DB 15,89,203 ; mulps %xmm3,%xmm1
+ DB 15,94,202 ; divps %xmm2,%xmm1
+ DB 15,92,196 ; subps %xmm4,%xmm0
+ DB 15,89,195 ; mulps %xmm3,%xmm0
+ DB 15,94,194 ; divps %xmm2,%xmm0
+ DB 65,15,194,214,4 ; cmpneqps %xmm14,%xmm2
+ DB 68,15,84,194 ; andps %xmm2,%xmm8
+ DB 15,84,202 ; andps %xmm2,%xmm1
+ DB 15,84,194 ; andps %xmm2,%xmm0
+ DB 68,15,40,13,5,55,0,0 ; movaps 0x3705(%rip),%xmm9 # 4580 <_sk_callback_sse41+0x265>
+ DB 69,15,89,209 ; mulps %xmm9,%xmm10
+ DB 15,40,29,10,55,0,0 ; movaps 0x370a(%rip),%xmm3 # 4590 <_sk_callback_sse41+0x275>
+ DB 68,15,89,219 ; mulps %xmm3,%xmm11
+ DB 69,15,88,218 ; addps %xmm10,%xmm11
+ DB 68,15,40,53,10,55,0,0 ; movaps 0x370a(%rip),%xmm14 # 45a0 <_sk_callback_sse41+0x285>
+ DB 68,15,40,253 ; movaps %xmm5,%xmm15
+ DB 69,15,89,254 ; mulps %xmm14,%xmm15
+ DB 69,15,88,251 ; addps %xmm11,%xmm15
+ DB 65,15,40,224 ; movaps %xmm8,%xmm4
+ DB 65,15,89,225 ; mulps %xmm9,%xmm4
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 15,89,211 ; mulps %xmm3,%xmm2
+ DB 15,88,212 ; addps %xmm4,%xmm2
+ DB 15,40,224 ; movaps %xmm0,%xmm4
+ DB 65,15,89,230 ; mulps %xmm14,%xmm4
+ DB 15,88,226 ; addps %xmm2,%xmm4
+ DB 68,15,92,252 ; subps %xmm4,%xmm15
+ DB 69,15,88,199 ; addps %xmm15,%xmm8
+ DB 65,15,88,207 ; addps %xmm15,%xmm1
+ DB 68,15,88,248 ; addps %xmm0,%xmm15
+ DB 15,40,193 ; movaps %xmm1,%xmm0
+ DB 65,15,93,199 ; minps %xmm15,%xmm0
+ DB 65,15,40,208 ; movaps %xmm8,%xmm2
+ DB 15,93,208 ; minps %xmm0,%xmm2
+ DB 15,40,225 ; movaps %xmm1,%xmm4
+ DB 65,15,95,231 ; maxps %xmm15,%xmm4
+ DB 69,15,89,200 ; mulps %xmm8,%xmm9
+ DB 15,89,217 ; mulps %xmm1,%xmm3
+ DB 65,15,88,217 ; addps %xmm9,%xmm3
+ DB 69,15,89,247 ; mulps %xmm15,%xmm14
+ DB 68,15,88,243 ; addps %xmm3,%xmm14
+ DB 15,87,237 ; xorps %xmm5,%xmm5
+ DB 15,194,234,2 ; cmpleps %xmm2,%xmm5
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 69,15,92,206 ; subps %xmm14,%xmm9
+ DB 69,15,89,206 ; mulps %xmm14,%xmm9
+ DB 65,15,40,222 ; movaps %xmm14,%xmm3
+ DB 15,92,218 ; subps %xmm2,%xmm3
+ DB 68,15,94,203 ; divps %xmm3,%xmm9
+ DB 69,15,88,206 ; addps %xmm14,%xmm9
+ DB 15,40,197 ; movaps %xmm5,%xmm0
+ DB 102,69,15,56,20,200 ; blendvps %xmm0,%xmm8,%xmm9
+ DB 68,15,95,196 ; maxps %xmm4,%xmm8
+ DB 69,15,40,213 ; movaps %xmm13,%xmm10
+ DB 69,15,194,208,1 ; cmpltps %xmm8,%xmm10
+ DB 65,15,40,209 ; movaps %xmm9,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 69,15,92,238 ; subps %xmm14,%xmm13
+ DB 65,15,89,213 ; mulps %xmm13,%xmm2
+ DB 69,15,92,198 ; subps %xmm14,%xmm8
+ DB 65,15,94,208 ; divps %xmm8,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 102,68,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm9
+ DB 68,15,40,217 ; movaps %xmm1,%xmm11
+ DB 69,15,92,222 ; subps %xmm14,%xmm11
+ DB 69,15,89,222 ; mulps %xmm14,%xmm11
+ DB 68,15,94,219 ; divps %xmm3,%xmm11
+ DB 69,15,88,222 ; addps %xmm14,%xmm11
+ DB 15,40,197 ; movaps %xmm5,%xmm0
+ DB 102,68,15,56,20,217 ; blendvps %xmm0,%xmm1,%xmm11
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 65,15,89,205 ; mulps %xmm13,%xmm1
+ DB 65,15,94,200 ; divps %xmm8,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 102,68,15,56,20,217 ; blendvps %xmm0,%xmm1,%xmm11
+ DB 65,15,40,215 ; movaps %xmm15,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 15,94,211 ; divps %xmm3,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 15,40,197 ; movaps %xmm5,%xmm0
+ DB 102,65,15,56,20,215 ; blendvps %xmm0,%xmm15,%xmm2
+ DB 15,40,202 ; movaps %xmm2,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 65,15,89,205 ; mulps %xmm13,%xmm1
+ DB 65,15,94,200 ; divps %xmm8,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 102,15,56,20,209 ; blendvps %xmm0,%xmm1,%xmm2
+ DB 15,87,192 ; xorps %xmm0,%xmm0
+ DB 68,15,95,200 ; maxps %xmm0,%xmm9
+ DB 68,15,95,216 ; maxps %xmm0,%xmm11
+ DB 15,95,208 ; maxps %xmm0,%xmm2
+ DB 68,15,40,198 ; movaps %xmm6,%xmm8
+ DB 68,15,40,84,36,16 ; movaps 0x10(%rsp),%xmm10
+ DB 69,15,88,194 ; addps %xmm10,%xmm8
+ DB 65,15,89,242 ; mulps %xmm10,%xmm6
+ DB 68,15,92,198 ; subps %xmm6,%xmm8
+ DB 69,15,89,200 ; mulps %xmm8,%xmm9
+ DB 69,15,89,216 ; mulps %xmm8,%xmm11
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,231 ; movaps %xmm7,%xmm4
+ DB 65,15,40,236 ; movaps %xmm12,%xmm5
+ DB 15,40,52,36 ; movaps (%rsp),%xmm6
+ DB 65,15,40,193 ; movaps %xmm9,%xmm0
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,40,216 ; movaps %xmm8,%xmm3
+ DB 65,15,40,250 ; movaps %xmm10,%xmm7
+ DB 72,131,196,40 ; add $0x28,%rsp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_saturation_sse41
+_sk_saturation_sse41 LABEL PROC
+ DB 72,131,236,40 ; sub $0x28,%rsp
+ DB 68,15,40,214 ; movaps %xmm6,%xmm10
+ DB 68,15,40,220 ; movaps %xmm4,%xmm11
+ DB 15,40,243 ; movaps %xmm3,%xmm6
+ DB 69,15,87,246 ; xorps %xmm14,%xmm14
+ DB 68,15,40,37,124,53,0,0 ; movaps 0x357c(%rip),%xmm12 # 45b0 <_sk_callback_sse41+0x295>
+ DB 65,15,40,220 ; movaps %xmm12,%xmm3
+ DB 15,94,223 ; divps %xmm7,%xmm3
+ DB 68,15,40,199 ; movaps %xmm7,%xmm8
+ DB 69,15,194,198,4 ; cmpneqps %xmm14,%xmm8
+ DB 68,15,84,195 ; andps %xmm3,%xmm8
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 69,15,89,203 ; mulps %xmm11,%xmm9
+ DB 69,15,40,232 ; movaps %xmm8,%xmm13
+ DB 68,15,89,237 ; mulps %xmm5,%xmm13
+ DB 15,41,108,36,16 ; movaps %xmm5,0x10(%rsp)
+ DB 69,15,89,194 ; mulps %xmm10,%xmm8
+ DB 68,15,41,20,36 ; movaps %xmm10,(%rsp)
+ DB 15,40,225 ; movaps %xmm1,%xmm4
+ DB 15,95,226 ; maxps %xmm2,%xmm4
+ DB 15,40,216 ; movaps %xmm0,%xmm3
+ DB 15,95,220 ; maxps %xmm4,%xmm3
+ DB 15,93,202 ; minps %xmm2,%xmm1
+ DB 15,93,193 ; minps %xmm1,%xmm0
+ DB 15,92,216 ; subps %xmm0,%xmm3
+ DB 65,15,40,197 ; movaps %xmm13,%xmm0
+ DB 65,15,93,192 ; minps %xmm8,%xmm0
+ DB 65,15,40,201 ; movaps %xmm9,%xmm1
+ DB 15,93,200 ; minps %xmm0,%xmm1
+ DB 65,15,40,197 ; movaps %xmm13,%xmm0
+ DB 65,15,95,192 ; maxps %xmm8,%xmm0
+ DB 65,15,40,209 ; movaps %xmm9,%xmm2
+ DB 15,95,208 ; maxps %xmm0,%xmm2
+ DB 15,92,209 ; subps %xmm1,%xmm2
+ DB 68,15,92,201 ; subps %xmm1,%xmm9
+ DB 68,15,89,203 ; mulps %xmm3,%xmm9
+ DB 68,15,94,202 ; divps %xmm2,%xmm9
+ DB 68,15,92,233 ; subps %xmm1,%xmm13
+ DB 68,15,89,235 ; mulps %xmm3,%xmm13
+ DB 68,15,94,234 ; divps %xmm2,%xmm13
+ DB 68,15,92,193 ; subps %xmm1,%xmm8
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 68,15,94,194 ; divps %xmm2,%xmm8
+ DB 65,15,194,214,4 ; cmpneqps %xmm14,%xmm2
+ DB 68,15,84,202 ; andps %xmm2,%xmm9
+ DB 68,15,84,234 ; andps %xmm2,%xmm13
+ DB 68,15,84,194 ; andps %xmm2,%xmm8
+ DB 15,40,13,232,52,0,0 ; movaps 0x34e8(%rip),%xmm1 # 45c0 <_sk_callback_sse41+0x2a5>
+ DB 65,15,40,211 ; movaps %xmm11,%xmm2
+ DB 15,89,209 ; mulps %xmm1,%xmm2
+ DB 15,40,5,234,52,0,0 ; movaps 0x34ea(%rip),%xmm0 # 45d0 <_sk_callback_sse41+0x2b5>
+ DB 15,40,221 ; movaps %xmm5,%xmm3
+ DB 15,89,216 ; mulps %xmm0,%xmm3
+ DB 15,88,218 ; addps %xmm2,%xmm3
+ DB 68,15,40,53,233,52,0,0 ; movaps 0x34e9(%rip),%xmm14 # 45e0 <_sk_callback_sse41+0x2c5>
+ DB 69,15,40,250 ; movaps %xmm10,%xmm15
+ DB 69,15,89,254 ; mulps %xmm14,%xmm15
+ DB 68,15,88,251 ; addps %xmm3,%xmm15
+ DB 65,15,40,209 ; movaps %xmm9,%xmm2
+ DB 15,89,209 ; mulps %xmm1,%xmm2
+ DB 65,15,40,221 ; movaps %xmm13,%xmm3
+ DB 15,89,216 ; mulps %xmm0,%xmm3
+ DB 15,88,218 ; addps %xmm2,%xmm3
+ DB 65,15,40,208 ; movaps %xmm8,%xmm2
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 15,88,211 ; addps %xmm3,%xmm2
+ DB 68,15,92,250 ; subps %xmm2,%xmm15
+ DB 69,15,88,207 ; addps %xmm15,%xmm9
+ DB 69,15,88,239 ; addps %xmm15,%xmm13
+ DB 69,15,88,248 ; addps %xmm8,%xmm15
+ DB 65,15,40,213 ; movaps %xmm13,%xmm2
+ DB 65,15,93,215 ; minps %xmm15,%xmm2
+ DB 65,15,40,225 ; movaps %xmm9,%xmm4
+ DB 15,93,226 ; minps %xmm2,%xmm4
+ DB 65,15,40,213 ; movaps %xmm13,%xmm2
+ DB 65,15,95,215 ; maxps %xmm15,%xmm2
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 65,15,89,197 ; mulps %xmm13,%xmm0
+ DB 15,88,193 ; addps %xmm1,%xmm0
+ DB 69,15,89,247 ; mulps %xmm15,%xmm14
+ DB 68,15,88,240 ; addps %xmm0,%xmm14
+ DB 15,87,237 ; xorps %xmm5,%xmm5
+ DB 15,194,236,2 ; cmpleps %xmm4,%xmm5
+ DB 69,15,40,209 ; movaps %xmm9,%xmm10
+ DB 69,15,92,214 ; subps %xmm14,%xmm10
+ DB 69,15,89,214 ; mulps %xmm14,%xmm10
+ DB 65,15,40,222 ; movaps %xmm14,%xmm3
+ DB 15,92,220 ; subps %xmm4,%xmm3
+ DB 68,15,94,211 ; divps %xmm3,%xmm10
+ DB 69,15,88,214 ; addps %xmm14,%xmm10
+ DB 15,40,197 ; movaps %xmm5,%xmm0
+ DB 102,69,15,56,20,209 ; blendvps %xmm0,%xmm9,%xmm10
+ DB 68,15,95,202 ; maxps %xmm2,%xmm9
+ DB 69,15,40,196 ; movaps %xmm12,%xmm8
+ DB 69,15,194,193,1 ; cmpltps %xmm9,%xmm8
+ DB 65,15,40,202 ; movaps %xmm10,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 69,15,92,230 ; subps %xmm14,%xmm12
+ DB 65,15,89,204 ; mulps %xmm12,%xmm1
+ DB 69,15,92,206 ; subps %xmm14,%xmm9
+ DB 65,15,94,201 ; divps %xmm9,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 65,15,40,192 ; movaps %xmm8,%xmm0
+ DB 102,68,15,56,20,209 ; blendvps %xmm0,%xmm1,%xmm10
+ DB 65,15,40,205 ; movaps %xmm13,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 65,15,89,206 ; mulps %xmm14,%xmm1
+ DB 15,94,203 ; divps %xmm3,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 15,40,197 ; movaps %xmm5,%xmm0
+ DB 102,65,15,56,20,205 ; blendvps %xmm0,%xmm13,%xmm1
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 65,15,89,212 ; mulps %xmm12,%xmm2
+ DB 65,15,94,209 ; divps %xmm9,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 65,15,40,192 ; movaps %xmm8,%xmm0
+ DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1
+ DB 65,15,40,215 ; movaps %xmm15,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 15,94,211 ; divps %xmm3,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 15,40,197 ; movaps %xmm5,%xmm0
+ DB 102,65,15,56,20,215 ; blendvps %xmm0,%xmm15,%xmm2
+ DB 15,40,218 ; movaps %xmm2,%xmm3
+ DB 65,15,92,222 ; subps %xmm14,%xmm3
+ DB 65,15,89,220 ; mulps %xmm12,%xmm3
+ DB 65,15,94,217 ; divps %xmm9,%xmm3
+ DB 65,15,88,222 ; addps %xmm14,%xmm3
+ DB 65,15,40,192 ; movaps %xmm8,%xmm0
+ DB 102,15,56,20,211 ; blendvps %xmm0,%xmm3,%xmm2
+ DB 15,87,192 ; xorps %xmm0,%xmm0
+ DB 68,15,95,208 ; maxps %xmm0,%xmm10
+ DB 15,95,200 ; maxps %xmm0,%xmm1
+ DB 15,95,208 ; maxps %xmm0,%xmm2
+ DB 68,15,40,198 ; movaps %xmm6,%xmm8
+ DB 68,15,88,199 ; addps %xmm7,%xmm8
+ DB 15,89,247 ; mulps %xmm7,%xmm6
+ DB 68,15,92,198 ; subps %xmm6,%xmm8
+ DB 69,15,89,208 ; mulps %xmm8,%xmm10
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 65,15,40,227 ; movaps %xmm11,%xmm4
+ DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5
+ DB 15,40,52,36 ; movaps (%rsp),%xmm6
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 65,15,40,216 ; movaps %xmm8,%xmm3
+ DB 72,131,196,40 ; add $0x28,%rsp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_color_sse41
+_sk_color_sse41 LABEL PROC
+ DB 72,131,236,40 ; sub $0x28,%rsp
+ DB 15,41,52,36 ; movaps %xmm6,(%rsp)
+ DB 68,15,40,221 ; movaps %xmm5,%xmm11
+ DB 68,15,41,92,36,16 ; movaps %xmm11,0x10(%rsp)
+ DB 15,40,236 ; movaps %xmm4,%xmm5
+ DB 15,40,227 ; movaps %xmm3,%xmm4
+ DB 68,15,40,192 ; movaps %xmm0,%xmm8
+ DB 69,15,87,201 ; xorps %xmm9,%xmm9
+ DB 68,15,40,45,92,51,0,0 ; movaps 0x335c(%rip),%xmm13 # 45f0 <_sk_callback_sse41+0x2d5>
+ DB 65,15,40,197 ; movaps %xmm13,%xmm0
+ DB 15,94,196 ; divps %xmm4,%xmm0
+ DB 65,15,194,217,4 ; cmpneqps %xmm9,%xmm3
+ DB 15,84,216 ; andps %xmm0,%xmm3
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 15,89,203 ; mulps %xmm3,%xmm1
+ DB 15,89,218 ; mulps %xmm2,%xmm3
+ DB 68,15,40,13,75,51,0,0 ; movaps 0x334b(%rip),%xmm9 # 4600 <_sk_callback_sse41+0x2e5>
+ DB 15,40,213 ; movaps %xmm5,%xmm2
+ DB 65,15,89,209 ; mulps %xmm9,%xmm2
+ DB 68,15,40,21,76,51,0,0 ; movaps 0x334c(%rip),%xmm10 # 4610 <_sk_callback_sse41+0x2f5>
+ DB 69,15,89,218 ; mulps %xmm10,%xmm11
+ DB 68,15,88,218 ; addps %xmm2,%xmm11
+ DB 68,15,40,53,76,51,0,0 ; movaps 0x334c(%rip),%xmm14 # 4620 <_sk_callback_sse41+0x305>
+ DB 68,15,40,254 ; movaps %xmm6,%xmm15
+ DB 69,15,89,254 ; mulps %xmm14,%xmm15
+ DB 69,15,88,251 ; addps %xmm11,%xmm15
+ DB 65,15,40,192 ; movaps %xmm8,%xmm0
+ DB 65,15,89,193 ; mulps %xmm9,%xmm0
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 65,15,89,210 ; mulps %xmm10,%xmm2
+ DB 15,88,208 ; addps %xmm0,%xmm2
+ DB 15,40,195 ; movaps %xmm3,%xmm0
+ DB 65,15,89,198 ; mulps %xmm14,%xmm0
+ DB 15,88,194 ; addps %xmm2,%xmm0
+ DB 68,15,92,248 ; subps %xmm0,%xmm15
+ DB 69,15,88,199 ; addps %xmm15,%xmm8
+ DB 65,15,88,207 ; addps %xmm15,%xmm1
+ DB 68,15,88,251 ; addps %xmm3,%xmm15
+ DB 15,40,193 ; movaps %xmm1,%xmm0
+ DB 65,15,93,199 ; minps %xmm15,%xmm0
+ DB 65,15,40,208 ; movaps %xmm8,%xmm2
+ DB 15,93,208 ; minps %xmm0,%xmm2
+ DB 15,40,241 ; movaps %xmm1,%xmm6
+ DB 65,15,95,247 ; maxps %xmm15,%xmm6
+ DB 69,15,89,200 ; mulps %xmm8,%xmm9
+ DB 68,15,89,209 ; mulps %xmm1,%xmm10
+ DB 69,15,88,209 ; addps %xmm9,%xmm10
+ DB 69,15,89,247 ; mulps %xmm15,%xmm14
+ DB 69,15,88,242 ; addps %xmm10,%xmm14
+ DB 15,87,219 ; xorps %xmm3,%xmm3
+ DB 15,194,218,2 ; cmpleps %xmm2,%xmm3
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 69,15,92,206 ; subps %xmm14,%xmm9
+ DB 69,15,89,206 ; mulps %xmm14,%xmm9
+ DB 69,15,40,214 ; movaps %xmm14,%xmm10
+ DB 68,15,92,210 ; subps %xmm2,%xmm10
+ DB 69,15,94,202 ; divps %xmm10,%xmm9
+ DB 69,15,88,206 ; addps %xmm14,%xmm9
+ DB 15,40,195 ; movaps %xmm3,%xmm0
+ DB 102,69,15,56,20,200 ; blendvps %xmm0,%xmm8,%xmm9
+ DB 68,15,95,198 ; maxps %xmm6,%xmm8
+ DB 69,15,40,229 ; movaps %xmm13,%xmm12
+ DB 69,15,194,224,1 ; cmpltps %xmm8,%xmm12
+ DB 65,15,40,209 ; movaps %xmm9,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 69,15,92,238 ; subps %xmm14,%xmm13
+ DB 65,15,89,213 ; mulps %xmm13,%xmm2
+ DB 69,15,92,198 ; subps %xmm14,%xmm8
+ DB 65,15,94,208 ; divps %xmm8,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 65,15,40,196 ; movaps %xmm12,%xmm0
+ DB 102,68,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm9
+ DB 68,15,40,217 ; movaps %xmm1,%xmm11
+ DB 69,15,92,222 ; subps %xmm14,%xmm11
+ DB 69,15,89,222 ; mulps %xmm14,%xmm11
+ DB 69,15,94,218 ; divps %xmm10,%xmm11
+ DB 69,15,88,222 ; addps %xmm14,%xmm11
+ DB 15,40,195 ; movaps %xmm3,%xmm0
+ DB 102,68,15,56,20,217 ; blendvps %xmm0,%xmm1,%xmm11
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 65,15,89,205 ; mulps %xmm13,%xmm1
+ DB 65,15,94,200 ; divps %xmm8,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 65,15,40,196 ; movaps %xmm12,%xmm0
+ DB 102,68,15,56,20,217 ; blendvps %xmm0,%xmm1,%xmm11
+ DB 65,15,40,215 ; movaps %xmm15,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 65,15,94,210 ; divps %xmm10,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 15,40,195 ; movaps %xmm3,%xmm0
+ DB 102,65,15,56,20,215 ; blendvps %xmm0,%xmm15,%xmm2
+ DB 15,40,202 ; movaps %xmm2,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 65,15,89,205 ; mulps %xmm13,%xmm1
+ DB 65,15,94,200 ; divps %xmm8,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 65,15,40,196 ; movaps %xmm12,%xmm0
+ DB 102,15,56,20,209 ; blendvps %xmm0,%xmm1,%xmm2
+ DB 15,87,192 ; xorps %xmm0,%xmm0
+ DB 68,15,95,200 ; maxps %xmm0,%xmm9
+ DB 68,15,95,216 ; maxps %xmm0,%xmm11
+ DB 15,95,208 ; maxps %xmm0,%xmm2
+ DB 68,15,40,196 ; movaps %xmm4,%xmm8
+ DB 68,15,88,199 ; addps %xmm7,%xmm8
+ DB 15,89,231 ; mulps %xmm7,%xmm4
+ DB 68,15,92,196 ; subps %xmm4,%xmm8
+ DB 69,15,89,200 ; mulps %xmm8,%xmm9
+ DB 69,15,89,216 ; mulps %xmm8,%xmm11
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,229 ; movaps %xmm5,%xmm4
+ DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5
+ DB 15,40,52,36 ; movaps (%rsp),%xmm6
+ DB 65,15,40,193 ; movaps %xmm9,%xmm0
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,40,216 ; movaps %xmm8,%xmm3
+ DB 72,131,196,40 ; add $0x28,%rsp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_luminosity_sse41
+_sk_luminosity_sse41 LABEL PROC
+ DB 72,131,236,40 ; sub $0x28,%rsp
+ DB 68,15,40,206 ; movaps %xmm6,%xmm9
+ DB 68,15,41,12,36 ; movaps %xmm9,(%rsp)
+ DB 68,15,40,245 ; movaps %xmm5,%xmm14
+ DB 68,15,41,116,36,16 ; movaps %xmm14,0x10(%rsp)
+ DB 15,40,244 ; movaps %xmm4,%xmm6
+ DB 15,40,235 ; movaps %xmm3,%xmm5
+ DB 69,15,87,228 ; xorps %xmm12,%xmm12
+ DB 68,15,40,45,182,49,0,0 ; movaps 0x31b6(%rip),%xmm13 # 4630 <_sk_callback_sse41+0x315>
+ DB 69,15,40,197 ; movaps %xmm13,%xmm8
+ DB 68,15,94,199 ; divps %xmm7,%xmm8
+ DB 15,40,223 ; movaps %xmm7,%xmm3
+ DB 65,15,194,220,4 ; cmpneqps %xmm12,%xmm3
+ DB 65,15,84,216 ; andps %xmm8,%xmm3
+ DB 68,15,40,211 ; movaps %xmm3,%xmm10
+ DB 68,15,89,214 ; mulps %xmm6,%xmm10
+ DB 68,15,40,219 ; movaps %xmm3,%xmm11
+ DB 69,15,89,222 ; mulps %xmm14,%xmm11
+ DB 65,15,89,217 ; mulps %xmm9,%xmm3
+ DB 68,15,40,5,150,49,0,0 ; movaps 0x3196(%rip),%xmm8 # 4640 <_sk_callback_sse41+0x325>
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 68,15,40,13,154,49,0,0 ; movaps 0x319a(%rip),%xmm9 # 4650 <_sk_callback_sse41+0x335>
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 15,88,200 ; addps %xmm0,%xmm1
+ DB 68,15,40,53,155,49,0,0 ; movaps 0x319b(%rip),%xmm14 # 4660 <_sk_callback_sse41+0x345>
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 15,88,209 ; addps %xmm1,%xmm2
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 15,88,200 ; addps %xmm0,%xmm1
+ DB 15,40,195 ; movaps %xmm3,%xmm0
+ DB 65,15,89,198 ; mulps %xmm14,%xmm0
+ DB 15,88,193 ; addps %xmm1,%xmm0
+ DB 15,92,208 ; subps %xmm0,%xmm2
+ DB 68,15,88,210 ; addps %xmm2,%xmm10
+ DB 68,15,88,218 ; addps %xmm2,%xmm11
+ DB 15,88,211 ; addps %xmm3,%xmm2
+ DB 65,15,40,195 ; movaps %xmm11,%xmm0
+ DB 15,93,194 ; minps %xmm2,%xmm0
+ DB 65,15,40,202 ; movaps %xmm10,%xmm1
+ DB 15,93,200 ; minps %xmm0,%xmm1
+ DB 65,15,40,219 ; movaps %xmm11,%xmm3
+ DB 15,95,218 ; maxps %xmm2,%xmm3
+ DB 69,15,89,194 ; mulps %xmm10,%xmm8
+ DB 69,15,89,203 ; mulps %xmm11,%xmm9
+ DB 69,15,88,200 ; addps %xmm8,%xmm9
+ DB 68,15,89,242 ; mulps %xmm2,%xmm14
+ DB 69,15,88,241 ; addps %xmm9,%xmm14
+ DB 15,87,228 ; xorps %xmm4,%xmm4
+ DB 15,194,225,2 ; cmpleps %xmm1,%xmm4
+ DB 69,15,40,194 ; movaps %xmm10,%xmm8
+ DB 69,15,92,198 ; subps %xmm14,%xmm8
+ DB 69,15,89,198 ; mulps %xmm14,%xmm8
+ DB 69,15,40,206 ; movaps %xmm14,%xmm9
+ DB 68,15,92,201 ; subps %xmm1,%xmm9
+ DB 69,15,94,193 ; divps %xmm9,%xmm8
+ DB 69,15,88,198 ; addps %xmm14,%xmm8
+ DB 15,40,196 ; movaps %xmm4,%xmm0
+ DB 102,69,15,56,20,194 ; blendvps %xmm0,%xmm10,%xmm8
+ DB 69,15,40,250 ; movaps %xmm10,%xmm15
+ DB 68,15,95,251 ; maxps %xmm3,%xmm15
+ DB 69,15,40,213 ; movaps %xmm13,%xmm10
+ DB 69,15,194,215,1 ; cmpltps %xmm15,%xmm10
+ DB 65,15,40,200 ; movaps %xmm8,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 69,15,92,238 ; subps %xmm14,%xmm13
+ DB 65,15,89,205 ; mulps %xmm13,%xmm1
+ DB 69,15,92,254 ; subps %xmm14,%xmm15
+ DB 65,15,94,207 ; divps %xmm15,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 102,68,15,56,20,193 ; blendvps %xmm0,%xmm1,%xmm8
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,92,206 ; subps %xmm14,%xmm1
+ DB 65,15,89,206 ; mulps %xmm14,%xmm1
+ DB 65,15,94,201 ; divps %xmm9,%xmm1
+ DB 65,15,88,206 ; addps %xmm14,%xmm1
+ DB 15,40,196 ; movaps %xmm4,%xmm0
+ DB 102,65,15,56,20,203 ; blendvps %xmm0,%xmm11,%xmm1
+ DB 15,40,217 ; movaps %xmm1,%xmm3
+ DB 65,15,92,222 ; subps %xmm14,%xmm3
+ DB 65,15,89,221 ; mulps %xmm13,%xmm3
+ DB 65,15,94,223 ; divps %xmm15,%xmm3
+ DB 65,15,88,222 ; addps %xmm14,%xmm3
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 102,15,56,20,203 ; blendvps %xmm0,%xmm3,%xmm1
+ DB 68,15,40,218 ; movaps %xmm2,%xmm11
+ DB 69,15,92,222 ; subps %xmm14,%xmm11
+ DB 69,15,89,222 ; mulps %xmm14,%xmm11
+ DB 69,15,94,217 ; divps %xmm9,%xmm11
+ DB 69,15,88,222 ; addps %xmm14,%xmm11
+ DB 15,40,196 ; movaps %xmm4,%xmm0
+ DB 102,68,15,56,20,218 ; blendvps %xmm0,%xmm2,%xmm11
+ DB 65,15,40,211 ; movaps %xmm11,%xmm2
+ DB 65,15,92,214 ; subps %xmm14,%xmm2
+ DB 65,15,89,213 ; mulps %xmm13,%xmm2
+ DB 65,15,94,215 ; divps %xmm15,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 65,15,40,194 ; movaps %xmm10,%xmm0
+ DB 102,68,15,56,20,218 ; blendvps %xmm0,%xmm2,%xmm11
+ DB 69,15,95,196 ; maxps %xmm12,%xmm8
+ DB 65,15,95,204 ; maxps %xmm12,%xmm1
+ DB 69,15,95,220 ; maxps %xmm12,%xmm11
+ DB 68,15,40,205 ; movaps %xmm5,%xmm9
+ DB 68,15,88,207 ; addps %xmm7,%xmm9
+ DB 15,89,239 ; mulps %xmm7,%xmm5
+ DB 68,15,92,205 ; subps %xmm5,%xmm9
+ DB 69,15,89,193 ; mulps %xmm9,%xmm8
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 69,15,89,217 ; mulps %xmm9,%xmm11
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 65,15,40,192 ; movaps %xmm8,%xmm0
+ DB 65,15,40,211 ; movaps %xmm11,%xmm2
+ DB 65,15,40,217 ; movaps %xmm9,%xmm3
+ DB 15,40,230 ; movaps %xmm6,%xmm4
+ DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5
+ DB 15,40,52,36 ; movaps (%rsp),%xmm6
+ DB 72,131,196,40 ; add $0x28,%rsp
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_clamp_0_sse41
_sk_clamp_0_sse41 LABEL PROC
DB 69,15,87,192 ; xorps %xmm8,%xmm8
@@ -10937,7 +12277,7 @@ _sk_clamp_0_sse41 LABEL PROC
PUBLIC _sk_clamp_1_sse41
_sk_clamp_1_sse41 LABEL PROC
- DB 68,15,40,5,12,47,0,0 ; movaps 0x2f0c(%rip),%xmm8 # 3ce0 <_sk_callback_sse41+0x24f>
+ DB 68,15,40,5,18,48,0,0 ; movaps 0x3012(%rip),%xmm8 # 4670 <_sk_callback_sse41+0x355>
DB 65,15,93,192 ; minps %xmm8,%xmm0
DB 65,15,93,200 ; minps %xmm8,%xmm1
DB 65,15,93,208 ; minps %xmm8,%xmm2
@@ -10947,7 +12287,7 @@ _sk_clamp_1_sse41 LABEL PROC
PUBLIC _sk_clamp_a_sse41
_sk_clamp_a_sse41 LABEL PROC
- DB 15,93,29,1,47,0,0 ; minps 0x2f01(%rip),%xmm3 # 3cf0 <_sk_callback_sse41+0x25f>
+ DB 15,93,29,7,48,0,0 ; minps 0x3007(%rip),%xmm3 # 4680 <_sk_callback_sse41+0x365>
DB 15,93,195 ; minps %xmm3,%xmm0
DB 15,93,203 ; minps %xmm3,%xmm1
DB 15,93,211 ; minps %xmm3,%xmm2
@@ -11020,7 +12360,7 @@ _sk_premul_sse41 LABEL PROC
PUBLIC _sk_unpremul_sse41
_sk_unpremul_sse41 LABEL PROC
DB 69,15,87,192 ; xorps %xmm8,%xmm8
- DB 68,15,40,13,108,46,0,0 ; movaps 0x2e6c(%rip),%xmm9 # 3d00 <_sk_callback_sse41+0x26f>
+ DB 68,15,40,13,114,47,0,0 ; movaps 0x2f72(%rip),%xmm9 # 4690 <_sk_callback_sse41+0x375>
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
@@ -11032,20 +12372,20 @@ _sk_unpremul_sse41 LABEL PROC
PUBLIC _sk_from_srgb_sse41
_sk_from_srgb_sse41 LABEL PROC
- DB 68,15,40,29,87,46,0,0 ; movaps 0x2e57(%rip),%xmm11 # 3d10 <_sk_callback_sse41+0x27f>
+ DB 68,15,40,29,93,47,0,0 ; movaps 0x2f5d(%rip),%xmm11 # 46a0 <_sk_callback_sse41+0x385>
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 68,15,40,37,79,46,0,0 ; movaps 0x2e4f(%rip),%xmm12 # 3d20 <_sk_callback_sse41+0x28f>
+ DB 68,15,40,37,85,47,0,0 ; movaps 0x2f55(%rip),%xmm12 # 46b0 <_sk_callback_sse41+0x395>
DB 68,15,40,192 ; movaps %xmm0,%xmm8
DB 69,15,89,196 ; mulps %xmm12,%xmm8
- DB 68,15,40,45,79,46,0,0 ; movaps 0x2e4f(%rip),%xmm13 # 3d30 <_sk_callback_sse41+0x29f>
+ DB 68,15,40,45,85,47,0,0 ; movaps 0x2f55(%rip),%xmm13 # 46c0 <_sk_callback_sse41+0x3a5>
DB 69,15,88,197 ; addps %xmm13,%xmm8
DB 69,15,89,194 ; mulps %xmm10,%xmm8
- DB 68,15,40,53,79,46,0,0 ; movaps 0x2e4f(%rip),%xmm14 # 3d40 <_sk_callback_sse41+0x2af>
+ DB 68,15,40,53,85,47,0,0 ; movaps 0x2f55(%rip),%xmm14 # 46d0 <_sk_callback_sse41+0x3b5>
DB 69,15,88,198 ; addps %xmm14,%xmm8
- DB 68,15,40,61,83,46,0,0 ; movaps 0x2e53(%rip),%xmm15 # 3d50 <_sk_callback_sse41+0x2bf>
+ DB 68,15,40,61,89,47,0,0 ; movaps 0x2f59(%rip),%xmm15 # 46e0 <_sk_callback_sse41+0x3c5>
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
@@ -11089,20 +12429,20 @@ _sk_to_srgb_sse41 LABEL PROC
DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8
DB 69,15,83,200 ; rcpps %xmm8,%xmm9
DB 69,15,82,208 ; rsqrtps %xmm8,%xmm10
- DB 68,15,40,29,192,45,0,0 ; movaps 0x2dc0(%rip),%xmm11 # 3d60 <_sk_callback_sse41+0x2cf>
+ DB 68,15,40,29,198,46,0,0 ; movaps 0x2ec6(%rip),%xmm11 # 46f0 <_sk_callback_sse41+0x3d5>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 65,15,89,203 ; mulps %xmm11,%xmm1
- DB 68,15,40,37,193,45,0,0 ; movaps 0x2dc1(%rip),%xmm12 # 3d70 <_sk_callback_sse41+0x2df>
+ DB 68,15,40,37,199,46,0,0 ; movaps 0x2ec7(%rip),%xmm12 # 4700 <_sk_callback_sse41+0x3e5>
DB 69,15,89,204 ; mulps %xmm12,%xmm9
- DB 68,15,40,45,197,45,0,0 ; movaps 0x2dc5(%rip),%xmm13 # 3d80 <_sk_callback_sse41+0x2ef>
+ DB 68,15,40,45,203,46,0,0 ; movaps 0x2ecb(%rip),%xmm13 # 4710 <_sk_callback_sse41+0x3f5>
DB 69,15,88,205 ; addps %xmm13,%xmm9
- DB 68,15,40,53,201,45,0,0 ; movaps 0x2dc9(%rip),%xmm14 # 3d90 <_sk_callback_sse41+0x2ff>
+ DB 68,15,40,53,207,46,0,0 ; movaps 0x2ecf(%rip),%xmm14 # 4720 <_sk_callback_sse41+0x405>
DB 69,15,89,214 ; mulps %xmm14,%xmm10
DB 69,15,88,209 ; addps %xmm9,%xmm10
- DB 68,15,40,5,201,45,0,0 ; movaps 0x2dc9(%rip),%xmm8 # 3da0 <_sk_callback_sse41+0x30f>
+ DB 68,15,40,5,207,46,0,0 ; movaps 0x2ecf(%rip),%xmm8 # 4730 <_sk_callback_sse41+0x415>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,93,202 ; minps %xmm10,%xmm9
- DB 68,15,40,61,201,45,0,0 ; movaps 0x2dc9(%rip),%xmm15 # 3db0 <_sk_callback_sse41+0x31f>
+ DB 68,15,40,61,207,46,0,0 ; movaps 0x2ecf(%rip),%xmm15 # 4740 <_sk_callback_sse41+0x425>
DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0
DB 102,68,15,56,20,201 ; blendvps %xmm0,%xmm1,%xmm9
DB 15,82,194 ; rsqrtps %xmm2,%xmm0
@@ -11155,7 +12495,7 @@ _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 68,15,40,53,23,45,0,0 ; movaps 0x2d17(%rip),%xmm14 # 3dc0 <_sk_callback_sse41+0x32f>
+ DB 68,15,40,53,29,46,0,0 ; movaps 0x2e1d(%rip),%xmm14 # 4750 <_sk_callback_sse41+0x435>
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
@@ -11164,27 +12504,27 @@ _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 68,15,84,61,254,44,0,0 ; andps 0x2cfe(%rip),%xmm15 # 3dd0 <_sk_callback_sse41+0x33f>
+ DB 68,15,84,61,4,46,0,0 ; andps 0x2e04(%rip),%xmm15 # 4760 <_sk_callback_sse41+0x445>
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 68,15,40,45,241,44,0,0 ; movaps 0x2cf1(%rip),%xmm13 # 3de0 <_sk_callback_sse41+0x34f>
+ DB 68,15,40,45,247,45,0,0 ; movaps 0x2df7(%rip),%xmm13 # 4770 <_sk_callback_sse41+0x455>
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 68,15,88,5,237,44,0,0 ; addps 0x2ced(%rip),%xmm8 # 3df0 <_sk_callback_sse41+0x35f>
+ DB 68,15,88,5,243,45,0,0 ; addps 0x2df3(%rip),%xmm8 # 4780 <_sk_callback_sse41+0x465>
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 68,15,89,5,229,44,0,0 ; mulps 0x2ce5(%rip),%xmm8 # 3e00 <_sk_callback_sse41+0x36f>
+ DB 68,15,89,5,235,45,0,0 ; mulps 0x2deb(%rip),%xmm8 # 4790 <_sk_callback_sse41+0x475>
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 15,40,5,217,44,0,0 ; movaps 0x2cd9(%rip),%xmm0 # 3e10 <_sk_callback_sse41+0x37f>
+ DB 15,40,5,223,45,0,0 ; movaps 0x2ddf(%rip),%xmm0 # 47a0 <_sk_callback_sse41+0x485>
DB 65,15,40,211 ; movaps %xmm11,%xmm2
DB 15,89,208 ; mulps %xmm0,%xmm2
DB 15,194,194,1 ; cmpltps %xmm2,%xmm0
@@ -11205,7 +12545,7 @@ _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 68,15,40,13,155,44,0,0 ; movaps 0x2c9b(%rip),%xmm9 # 3e20 <_sk_callback_sse41+0x38f>
+ DB 68,15,40,13,161,45,0,0 ; movaps 0x2da1(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0x495>
DB 65,15,40,193 ; movaps %xmm9,%xmm0
DB 15,194,194,2 ; cmpleps %xmm2,%xmm0
DB 15,40,217 ; movaps %xmm1,%xmm3
@@ -11218,19 +12558,19 @@ _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 68,15,40,53,119,44,0,0 ; movaps 0x2c77(%rip),%xmm14 # 3e30 <_sk_callback_sse41+0x39f>
+ DB 68,15,40,53,125,45,0,0 ; movaps 0x2d7d(%rip),%xmm14 # 47c0 <_sk_callback_sse41+0x4a5>
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 68,15,40,29,112,44,0,0 ; movaps 0x2c70(%rip),%xmm11 # 3e40 <_sk_callback_sse41+0x3af>
+ DB 68,15,40,29,118,45,0,0 ; movaps 0x2d76(%rip),%xmm11 # 47d0 <_sk_callback_sse41+0x4b5>
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 15,40,61,105,44,0,0 ; movaps 0x2c69(%rip),%xmm7 # 3e50 <_sk_callback_sse41+0x3bf>
+ DB 15,40,61,111,45,0,0 ; movaps 0x2d6f(%rip),%xmm7 # 47e0 <_sk_callback_sse41+0x4c5>
DB 69,15,40,238 ; movaps %xmm14,%xmm13
DB 68,15,89,239 ; mulps %xmm7,%xmm13
- DB 15,40,29,106,44,0,0 ; movaps 0x2c6a(%rip),%xmm3 # 3e60 <_sk_callback_sse41+0x3cf>
+ DB 15,40,29,112,45,0,0 ; movaps 0x2d70(%rip),%xmm3 # 47f0 <_sk_callback_sse41+0x4d5>
DB 68,15,40,227 ; movaps %xmm3,%xmm12
DB 69,15,92,229 ; subps %xmm13,%xmm12
DB 68,15,89,230 ; mulps %xmm6,%xmm12
@@ -11240,7 +12580,7 @@ _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 68,15,40,37,73,44,0,0 ; movaps 0x2c49(%rip),%xmm12 # 3e70 <_sk_callback_sse41+0x3df>
+ DB 68,15,40,37,79,45,0,0 ; movaps 0x2d4f(%rip),%xmm12 # 4800 <_sk_callback_sse41+0x4e5>
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
@@ -11274,7 +12614,7 @@ _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 68,15,88,21,194,43,0,0 ; addps 0x2bc2(%rip),%xmm10 # 3e80 <_sk_callback_sse41+0x3ef>
+ DB 68,15,88,21,200,44,0,0 ; addps 0x2cc8(%rip),%xmm10 # 4810 <_sk_callback_sse41+0x4f5>
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
@@ -11323,7 +12663,7 @@ _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 68,15,89,5,27,43,0,0 ; mulps 0x2b1b(%rip),%xmm8 # 3e90 <_sk_callback_sse41+0x3ff>
+ DB 68,15,89,5,33,44,0,0 ; mulps 0x2c21(%rip),%xmm8 # 4820 <_sk_callback_sse41+0x505>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 65,15,89,208 ; mulps %xmm8,%xmm2
@@ -11357,7 +12697,7 @@ _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 68,15,89,5,199,42,0,0 ; mulps 0x2ac7(%rip),%xmm8 # 3ea0 <_sk_callback_sse41+0x40f>
+ DB 68,15,89,5,205,43,0,0 ; mulps 0x2bcd(%rip),%xmm8 # 4830 <_sk_callback_sse41+0x515>
DB 15,92,196 ; subps %xmm4,%xmm0
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -11378,17 +12718,17 @@ _sk_lerp_565_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 102,68,15,56,51,4,120 ; pmovzxwd (%rax,%rdi,2),%xmm8
- DB 102,15,111,29,151,42,0,0 ; movdqa 0x2a97(%rip),%xmm3 # 3eb0 <_sk_callback_sse41+0x41f>
+ DB 102,15,111,29,157,43,0,0 ; movdqa 0x2b9d(%rip),%xmm3 # 4840 <_sk_callback_sse41+0x525>
DB 102,65,15,219,216 ; pand %xmm8,%xmm3
DB 68,15,91,203 ; cvtdq2ps %xmm3,%xmm9
- DB 68,15,89,13,150,42,0,0 ; mulps 0x2a96(%rip),%xmm9 # 3ec0 <_sk_callback_sse41+0x42f>
- DB 102,15,111,29,158,42,0,0 ; movdqa 0x2a9e(%rip),%xmm3 # 3ed0 <_sk_callback_sse41+0x43f>
+ DB 68,15,89,13,156,43,0,0 ; mulps 0x2b9c(%rip),%xmm9 # 4850 <_sk_callback_sse41+0x535>
+ DB 102,15,111,29,164,43,0,0 ; movdqa 0x2ba4(%rip),%xmm3 # 4860 <_sk_callback_sse41+0x545>
DB 102,65,15,219,216 ; pand %xmm8,%xmm3
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,159,42,0,0 ; mulps 0x2a9f(%rip),%xmm3 # 3ee0 <_sk_callback_sse41+0x44f>
- DB 102,68,15,219,5,166,42,0,0 ; pand 0x2aa6(%rip),%xmm8 # 3ef0 <_sk_callback_sse41+0x45f>
+ DB 15,89,29,165,43,0,0 ; mulps 0x2ba5(%rip),%xmm3 # 4870 <_sk_callback_sse41+0x555>
+ DB 102,68,15,219,5,172,43,0,0 ; pand 0x2bac(%rip),%xmm8 # 4880 <_sk_callback_sse41+0x565>
DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
- DB 68,15,89,5,170,42,0,0 ; mulps 0x2aaa(%rip),%xmm8 # 3f00 <_sk_callback_sse41+0x46f>
+ DB 68,15,89,5,176,43,0,0 ; mulps 0x2bb0(%rip),%xmm8 # 4890 <_sk_callback_sse41+0x575>
DB 15,92,196 ; subps %xmm4,%xmm0
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -11399,7 +12739,7 @@ _sk_lerp_565_sse41 LABEL PROC
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 15,88,214 ; addps %xmm6,%xmm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,148,42,0,0 ; movaps 0x2a94(%rip),%xmm3 # 3f10 <_sk_callback_sse41+0x47f>
+ DB 15,40,29,154,43,0,0 ; movaps 0x2b9a(%rip),%xmm3 # 48a0 <_sk_callback_sse41+0x585>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_load_tables_sse41
@@ -11408,7 +12748,7 @@ _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 102,15,111,5,139,42,0,0 ; movdqa 0x2a8b(%rip),%xmm0 # 3f20 <_sk_callback_sse41+0x48f>
+ DB 102,15,111,5,145,43,0,0 ; movdqa 0x2b91(%rip),%xmm0 # 48b0 <_sk_callback_sse41+0x595>
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
@@ -11423,7 +12763,7 @@ _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 102,15,56,0,13,70,42,0,0 ; pshufb 0x2a46(%rip),%xmm1 # 3f30 <_sk_callback_sse41+0x49f>
+ DB 102,15,56,0,13,76,43,0,0 ; pshufb 0x2b4c(%rip),%xmm1 # 48c0 <_sk_callback_sse41+0x5a5>
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
@@ -11438,7 +12778,7 @@ _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 102,15,56,0,21,2,42,0,0 ; pshufb 0x2a02(%rip),%xmm2 # 3f40 <_sk_callback_sse41+0x4af>
+ DB 102,15,56,0,21,8,43,0,0 ; pshufb 0x2b08(%rip),%xmm2 # 48d0 <_sk_callback_sse41+0x5b5>
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
@@ -11453,7 +12793,7 @@ _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 15,89,29,191,41,0,0 ; mulps 0x29bf(%rip),%xmm3 # 3f50 <_sk_callback_sse41+0x4bf>
+ DB 15,89,29,197,42,0,0 ; mulps 0x2ac5(%rip),%xmm3 # 48e0 <_sk_callback_sse41+0x5c5>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -11470,7 +12810,7 @@ _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 102,68,15,111,5,146,41,0,0 ; movdqa 0x2992(%rip),%xmm8 # 3f60 <_sk_callback_sse41+0x4cf>
+ DB 102,68,15,111,5,152,42,0,0 ; movdqa 0x2a98(%rip),%xmm8 # 48f0 <_sk_callback_sse41+0x5d5>
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
@@ -11487,7 +12827,7 @@ _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 102,15,56,0,13,69,41,0,0 ; pshufb 0x2945(%rip),%xmm1 # 3f70 <_sk_callback_sse41+0x4df>
+ DB 102,15,56,0,13,75,42,0,0 ; pshufb 0x2a4b(%rip),%xmm1 # 4900 <_sk_callback_sse41+0x5e5>
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
@@ -11523,7 +12863,7 @@ _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 15,89,29,147,40,0,0 ; mulps 0x2893(%rip),%xmm3 # 3f80 <_sk_callback_sse41+0x4ef>
+ DB 15,89,29,153,41,0,0 ; mulps 0x2999(%rip),%xmm3 # 4910 <_sk_callback_sse41+0x5f5>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -11543,7 +12883,7 @@ _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 102,68,15,111,5,85,40,0,0 ; movdqa 0x2855(%rip),%xmm8 # 3f90 <_sk_callback_sse41+0x4ff>
+ DB 102,68,15,111,5,91,41,0,0 ; movdqa 0x295b(%rip),%xmm8 # 4920 <_sk_callback_sse41+0x605>
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
@@ -11560,7 +12900,7 @@ _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 102,15,56,0,13,8,40,0,0 ; pshufb 0x2808(%rip),%xmm1 # 3fa0 <_sk_callback_sse41+0x50f>
+ DB 102,15,56,0,13,14,41,0,0 ; pshufb 0x290e(%rip),%xmm1 # 4930 <_sk_callback_sse41+0x615>
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
@@ -11591,7 +12931,7 @@ _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 15,40,29,115,39,0,0 ; movaps 0x2773(%rip),%xmm3 # 3fb0 <_sk_callback_sse41+0x51f>
+ DB 15,40,29,121,40,0,0 ; movaps 0x2879(%rip),%xmm3 # 4940 <_sk_callback_sse41+0x625>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_byte_tables_sse41
@@ -11599,7 +12939,7 @@ _sk_byte_tables_sse41 LABEL PROC
DB 65,86 ; push %r14
DB 83 ; push %rbx
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 68,15,40,5,116,39,0,0 ; movaps 0x2774(%rip),%xmm8 # 3fc0 <_sk_callback_sse41+0x52f>
+ DB 68,15,40,5,122,40,0,0 ; movaps 0x287a(%rip),%xmm8 # 4950 <_sk_callback_sse41+0x635>
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
@@ -11618,7 +12958,7 @@ _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 68,15,40,13,37,39,0,0 ; movaps 0x2725(%rip),%xmm9 # 3fd0 <_sk_callback_sse41+0x53f>
+ DB 68,15,40,13,43,40,0,0 ; movaps 0x282b(%rip),%xmm9 # 4960 <_sk_callback_sse41+0x645>
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
@@ -11707,7 +13047,7 @@ _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 68,15,40,13,173,37,0,0 ; movaps 0x25ad(%rip),%xmm9 # 3fe0 <_sk_callback_sse41+0x54f>
+ DB 68,15,40,13,179,38,0,0 ; movaps 0x26b3(%rip),%xmm9 # 4970 <_sk_callback_sse41+0x655>
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
@@ -11874,31 +13214,31 @@ _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 68,15,89,5,4,35,0,0 ; mulps 0x2304(%rip),%xmm8 # 3ff0 <_sk_callback_sse41+0x55f>
- DB 68,15,84,21,12,35,0,0 ; andps 0x230c(%rip),%xmm10 # 4000 <_sk_callback_sse41+0x56f>
- DB 68,15,86,21,20,35,0,0 ; orps 0x2314(%rip),%xmm10 # 4010 <_sk_callback_sse41+0x57f>
- DB 68,15,88,5,28,35,0,0 ; addps 0x231c(%rip),%xmm8 # 4020 <_sk_callback_sse41+0x58f>
- DB 68,15,40,37,36,35,0,0 ; movaps 0x2324(%rip),%xmm12 # 4030 <_sk_callback_sse41+0x59f>
+ DB 68,15,89,5,10,36,0,0 ; mulps 0x240a(%rip),%xmm8 # 4980 <_sk_callback_sse41+0x665>
+ DB 68,15,84,21,18,36,0,0 ; andps 0x2412(%rip),%xmm10 # 4990 <_sk_callback_sse41+0x675>
+ DB 68,15,86,21,26,36,0,0 ; orps 0x241a(%rip),%xmm10 # 49a0 <_sk_callback_sse41+0x685>
+ DB 68,15,88,5,34,36,0,0 ; addps 0x2422(%rip),%xmm8 # 49b0 <_sk_callback_sse41+0x695>
+ DB 68,15,40,37,42,36,0,0 ; movaps 0x242a(%rip),%xmm12 # 49c0 <_sk_callback_sse41+0x6a5>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,196 ; subps %xmm12,%xmm8
- DB 68,15,88,21,36,35,0,0 ; addps 0x2324(%rip),%xmm10 # 4040 <_sk_callback_sse41+0x5af>
- DB 68,15,40,37,44,35,0,0 ; movaps 0x232c(%rip),%xmm12 # 4050 <_sk_callback_sse41+0x5bf>
+ DB 68,15,88,21,42,36,0,0 ; addps 0x242a(%rip),%xmm10 # 49d0 <_sk_callback_sse41+0x6b5>
+ DB 68,15,40,37,50,36,0,0 ; movaps 0x2432(%rip),%xmm12 # 49e0 <_sk_callback_sse41+0x6c5>
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 68,15,88,5,25,35,0,0 ; addps 0x2319(%rip),%xmm8 # 4060 <_sk_callback_sse41+0x5cf>
- DB 68,15,40,21,33,35,0,0 ; movaps 0x2321(%rip),%xmm10 # 4070 <_sk_callback_sse41+0x5df>
+ DB 68,15,88,5,31,36,0,0 ; addps 0x241f(%rip),%xmm8 # 49f0 <_sk_callback_sse41+0x6d5>
+ DB 68,15,40,21,39,36,0,0 ; movaps 0x2427(%rip),%xmm10 # 4a00 <_sk_callback_sse41+0x6e5>
DB 69,15,89,211 ; mulps %xmm11,%xmm10
DB 69,15,92,194 ; subps %xmm10,%xmm8
- DB 68,15,40,21,33,35,0,0 ; movaps 0x2321(%rip),%xmm10 # 4080 <_sk_callback_sse41+0x5ef>
+ DB 68,15,40,21,39,36,0,0 ; movaps 0x2427(%rip),%xmm10 # 4a10 <_sk_callback_sse41+0x6f5>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 68,15,40,29,37,35,0,0 ; movaps 0x2325(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x5ff>
+ DB 68,15,40,29,43,36,0,0 ; movaps 0x242b(%rip),%xmm11 # 4a20 <_sk_callback_sse41+0x705>
DB 69,15,94,218 ; divps %xmm10,%xmm11
DB 69,15,88,216 ; addps %xmm8,%xmm11
- DB 68,15,89,29,37,35,0,0 ; mulps 0x2325(%rip),%xmm11 # 40a0 <_sk_callback_sse41+0x60f>
+ DB 68,15,89,29,43,36,0,0 ; mulps 0x242b(%rip),%xmm11 # 4a30 <_sk_callback_sse41+0x715>
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
@@ -11906,7 +13246,7 @@ _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 68,15,93,5,12,35,0,0 ; minps 0x230c(%rip),%xmm8 # 40b0 <_sk_callback_sse41+0x61f>
+ DB 68,15,93,5,18,36,0,0 ; minps 0x2412(%rip),%xmm8 # 4a40 <_sk_callback_sse41+0x725>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -11934,31 +13274,31 @@ _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 68,15,89,37,173,34,0,0 ; mulps 0x22ad(%rip),%xmm12 # 40c0 <_sk_callback_sse41+0x62f>
- DB 68,15,84,29,181,34,0,0 ; andps 0x22b5(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x63f>
- DB 68,15,86,29,189,34,0,0 ; orps 0x22bd(%rip),%xmm11 # 40e0 <_sk_callback_sse41+0x64f>
- DB 68,15,88,37,197,34,0,0 ; addps 0x22c5(%rip),%xmm12 # 40f0 <_sk_callback_sse41+0x65f>
- DB 15,40,13,206,34,0,0 ; movaps 0x22ce(%rip),%xmm1 # 4100 <_sk_callback_sse41+0x66f>
+ DB 68,15,89,37,179,35,0,0 ; mulps 0x23b3(%rip),%xmm12 # 4a50 <_sk_callback_sse41+0x735>
+ DB 68,15,84,29,187,35,0,0 ; andps 0x23bb(%rip),%xmm11 # 4a60 <_sk_callback_sse41+0x745>
+ DB 68,15,86,29,195,35,0,0 ; orps 0x23c3(%rip),%xmm11 # 4a70 <_sk_callback_sse41+0x755>
+ DB 68,15,88,37,203,35,0,0 ; addps 0x23cb(%rip),%xmm12 # 4a80 <_sk_callback_sse41+0x765>
+ DB 15,40,13,212,35,0,0 ; movaps 0x23d4(%rip),%xmm1 # 4a90 <_sk_callback_sse41+0x775>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
- DB 68,15,88,29,206,34,0,0 ; addps 0x22ce(%rip),%xmm11 # 4110 <_sk_callback_sse41+0x67f>
- DB 15,40,13,215,34,0,0 ; movaps 0x22d7(%rip),%xmm1 # 4120 <_sk_callback_sse41+0x68f>
+ DB 68,15,88,29,212,35,0,0 ; addps 0x23d4(%rip),%xmm11 # 4aa0 <_sk_callback_sse41+0x785>
+ DB 15,40,13,221,35,0,0 ; movaps 0x23dd(%rip),%xmm1 # 4ab0 <_sk_callback_sse41+0x795>
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 68,15,88,37,196,34,0,0 ; addps 0x22c4(%rip),%xmm12 # 4130 <_sk_callback_sse41+0x69f>
- DB 15,40,13,205,34,0,0 ; movaps 0x22cd(%rip),%xmm1 # 4140 <_sk_callback_sse41+0x6af>
+ DB 68,15,88,37,202,35,0,0 ; addps 0x23ca(%rip),%xmm12 # 4ac0 <_sk_callback_sse41+0x7a5>
+ DB 15,40,13,211,35,0,0 ; movaps 0x23d3(%rip),%xmm1 # 4ad0 <_sk_callback_sse41+0x7b5>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
- DB 68,15,40,21,205,34,0,0 ; movaps 0x22cd(%rip),%xmm10 # 4150 <_sk_callback_sse41+0x6bf>
+ DB 68,15,40,21,211,35,0,0 ; movaps 0x23d3(%rip),%xmm10 # 4ae0 <_sk_callback_sse41+0x7c5>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,13,210,34,0,0 ; movaps 0x22d2(%rip),%xmm1 # 4160 <_sk_callback_sse41+0x6cf>
+ DB 15,40,13,216,35,0,0 ; movaps 0x23d8(%rip),%xmm1 # 4af0 <_sk_callback_sse41+0x7d5>
DB 65,15,94,202 ; divps %xmm10,%xmm1
DB 65,15,88,204 ; addps %xmm12,%xmm1
- DB 15,89,13,211,34,0,0 ; mulps 0x22d3(%rip),%xmm1 # 4170 <_sk_callback_sse41+0x6df>
+ DB 15,89,13,217,35,0,0 ; mulps 0x23d9(%rip),%xmm1 # 4b00 <_sk_callback_sse41+0x7e5>
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
@@ -11966,7 +13306,7 @@ _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 15,93,13,190,34,0,0 ; minps 0x22be(%rip),%xmm1 # 4180 <_sk_callback_sse41+0x6ef>
+ DB 15,93,13,196,35,0,0 ; minps 0x23c4(%rip),%xmm1 # 4b10 <_sk_callback_sse41+0x7f5>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -11994,31 +13334,31 @@ _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 68,15,89,37,95,34,0,0 ; mulps 0x225f(%rip),%xmm12 # 4190 <_sk_callback_sse41+0x6ff>
- DB 68,15,84,29,103,34,0,0 ; andps 0x2267(%rip),%xmm11 # 41a0 <_sk_callback_sse41+0x70f>
- DB 68,15,86,29,111,34,0,0 ; orps 0x226f(%rip),%xmm11 # 41b0 <_sk_callback_sse41+0x71f>
- DB 68,15,88,37,119,34,0,0 ; addps 0x2277(%rip),%xmm12 # 41c0 <_sk_callback_sse41+0x72f>
- DB 15,40,21,128,34,0,0 ; movaps 0x2280(%rip),%xmm2 # 41d0 <_sk_callback_sse41+0x73f>
+ DB 68,15,89,37,101,35,0,0 ; mulps 0x2365(%rip),%xmm12 # 4b20 <_sk_callback_sse41+0x805>
+ DB 68,15,84,29,109,35,0,0 ; andps 0x236d(%rip),%xmm11 # 4b30 <_sk_callback_sse41+0x815>
+ DB 68,15,86,29,117,35,0,0 ; orps 0x2375(%rip),%xmm11 # 4b40 <_sk_callback_sse41+0x825>
+ DB 68,15,88,37,125,35,0,0 ; addps 0x237d(%rip),%xmm12 # 4b50 <_sk_callback_sse41+0x835>
+ DB 15,40,21,134,35,0,0 ; movaps 0x2386(%rip),%xmm2 # 4b60 <_sk_callback_sse41+0x845>
DB 65,15,89,211 ; mulps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
- DB 68,15,88,29,128,34,0,0 ; addps 0x2280(%rip),%xmm11 # 41e0 <_sk_callback_sse41+0x74f>
- DB 15,40,21,137,34,0,0 ; movaps 0x2289(%rip),%xmm2 # 41f0 <_sk_callback_sse41+0x75f>
+ DB 68,15,88,29,134,35,0,0 ; addps 0x2386(%rip),%xmm11 # 4b70 <_sk_callback_sse41+0x855>
+ DB 15,40,21,143,35,0,0 ; movaps 0x238f(%rip),%xmm2 # 4b80 <_sk_callback_sse41+0x865>
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 68,15,88,37,118,34,0,0 ; addps 0x2276(%rip),%xmm12 # 4200 <_sk_callback_sse41+0x76f>
- DB 15,40,21,127,34,0,0 ; movaps 0x227f(%rip),%xmm2 # 4210 <_sk_callback_sse41+0x77f>
+ DB 68,15,88,37,124,35,0,0 ; addps 0x237c(%rip),%xmm12 # 4b90 <_sk_callback_sse41+0x875>
+ DB 15,40,21,133,35,0,0 ; movaps 0x2385(%rip),%xmm2 # 4ba0 <_sk_callback_sse41+0x885>
DB 65,15,89,211 ; mulps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
- DB 68,15,40,21,127,34,0,0 ; movaps 0x227f(%rip),%xmm10 # 4220 <_sk_callback_sse41+0x78f>
+ DB 68,15,40,21,133,35,0,0 ; movaps 0x2385(%rip),%xmm10 # 4bb0 <_sk_callback_sse41+0x895>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,21,132,34,0,0 ; movaps 0x2284(%rip),%xmm2 # 4230 <_sk_callback_sse41+0x79f>
+ DB 15,40,21,138,35,0,0 ; movaps 0x238a(%rip),%xmm2 # 4bc0 <_sk_callback_sse41+0x8a5>
DB 65,15,94,210 ; divps %xmm10,%xmm2
DB 65,15,88,212 ; addps %xmm12,%xmm2
- DB 15,89,21,133,34,0,0 ; mulps 0x2285(%rip),%xmm2 # 4240 <_sk_callback_sse41+0x7af>
+ DB 15,89,21,139,35,0,0 ; mulps 0x238b(%rip),%xmm2 # 4bd0 <_sk_callback_sse41+0x8b5>
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
@@ -12026,7 +13366,7 @@ _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 15,93,21,112,34,0,0 ; minps 0x2270(%rip),%xmm2 # 4250 <_sk_callback_sse41+0x7bf>
+ DB 15,93,21,118,35,0,0 ; minps 0x2376(%rip),%xmm2 # 4be0 <_sk_callback_sse41+0x8c5>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -12054,31 +13394,31 @@ _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 68,15,89,37,17,34,0,0 ; mulps 0x2211(%rip),%xmm12 # 4260 <_sk_callback_sse41+0x7cf>
- DB 68,15,84,29,25,34,0,0 ; andps 0x2219(%rip),%xmm11 # 4270 <_sk_callback_sse41+0x7df>
- DB 68,15,86,29,33,34,0,0 ; orps 0x2221(%rip),%xmm11 # 4280 <_sk_callback_sse41+0x7ef>
- DB 68,15,88,37,41,34,0,0 ; addps 0x2229(%rip),%xmm12 # 4290 <_sk_callback_sse41+0x7ff>
- DB 15,40,29,50,34,0,0 ; movaps 0x2232(%rip),%xmm3 # 42a0 <_sk_callback_sse41+0x80f>
+ DB 68,15,89,37,23,35,0,0 ; mulps 0x2317(%rip),%xmm12 # 4bf0 <_sk_callback_sse41+0x8d5>
+ DB 68,15,84,29,31,35,0,0 ; andps 0x231f(%rip),%xmm11 # 4c00 <_sk_callback_sse41+0x8e5>
+ DB 68,15,86,29,39,35,0,0 ; orps 0x2327(%rip),%xmm11 # 4c10 <_sk_callback_sse41+0x8f5>
+ DB 68,15,88,37,47,35,0,0 ; addps 0x232f(%rip),%xmm12 # 4c20 <_sk_callback_sse41+0x905>
+ DB 15,40,29,56,35,0,0 ; movaps 0x2338(%rip),%xmm3 # 4c30 <_sk_callback_sse41+0x915>
DB 65,15,89,219 ; mulps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
- DB 68,15,88,29,50,34,0,0 ; addps 0x2232(%rip),%xmm11 # 42b0 <_sk_callback_sse41+0x81f>
- DB 15,40,29,59,34,0,0 ; movaps 0x223b(%rip),%xmm3 # 42c0 <_sk_callback_sse41+0x82f>
+ DB 68,15,88,29,56,35,0,0 ; addps 0x2338(%rip),%xmm11 # 4c40 <_sk_callback_sse41+0x925>
+ DB 15,40,29,65,35,0,0 ; movaps 0x2341(%rip),%xmm3 # 4c50 <_sk_callback_sse41+0x935>
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 68,15,88,37,40,34,0,0 ; addps 0x2228(%rip),%xmm12 # 42d0 <_sk_callback_sse41+0x83f>
- DB 15,40,29,49,34,0,0 ; movaps 0x2231(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x84f>
+ DB 68,15,88,37,46,35,0,0 ; addps 0x232e(%rip),%xmm12 # 4c60 <_sk_callback_sse41+0x945>
+ DB 15,40,29,55,35,0,0 ; movaps 0x2337(%rip),%xmm3 # 4c70 <_sk_callback_sse41+0x955>
DB 65,15,89,219 ; mulps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
- DB 68,15,40,21,49,34,0,0 ; movaps 0x2231(%rip),%xmm10 # 42f0 <_sk_callback_sse41+0x85f>
+ DB 68,15,40,21,55,35,0,0 ; movaps 0x2337(%rip),%xmm10 # 4c80 <_sk_callback_sse41+0x965>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,29,54,34,0,0 ; movaps 0x2236(%rip),%xmm3 # 4300 <_sk_callback_sse41+0x86f>
+ DB 15,40,29,60,35,0,0 ; movaps 0x233c(%rip),%xmm3 # 4c90 <_sk_callback_sse41+0x975>
DB 65,15,94,218 ; divps %xmm10,%xmm3
DB 65,15,88,220 ; addps %xmm12,%xmm3
- DB 15,89,29,55,34,0,0 ; mulps 0x2237(%rip),%xmm3 # 4310 <_sk_callback_sse41+0x87f>
+ DB 15,89,29,61,35,0,0 ; mulps 0x233d(%rip),%xmm3 # 4ca0 <_sk_callback_sse41+0x985>
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
@@ -12086,7 +13426,7 @@ _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 15,93,29,34,34,0,0 ; minps 0x2222(%rip),%xmm3 # 4320 <_sk_callback_sse41+0x88f>
+ DB 15,93,29,40,35,0,0 ; minps 0x2328(%rip),%xmm3 # 4cb0 <_sk_callback_sse41+0x995>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -12094,29 +13434,29 @@ _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 68,15,89,5,30,34,0,0 ; mulps 0x221e(%rip),%xmm8 # 4330 <_sk_callback_sse41+0x89f>
- DB 68,15,40,13,38,34,0,0 ; movaps 0x2226(%rip),%xmm9 # 4340 <_sk_callback_sse41+0x8af>
+ DB 68,15,89,5,36,35,0,0 ; mulps 0x2324(%rip),%xmm8 # 4cc0 <_sk_callback_sse41+0x9a5>
+ DB 68,15,40,13,44,35,0,0 ; movaps 0x232c(%rip),%xmm9 # 4cd0 <_sk_callback_sse41+0x9b5>
DB 65,15,89,201 ; mulps %xmm9,%xmm1
- DB 15,40,5,43,34,0,0 ; movaps 0x222b(%rip),%xmm0 # 4350 <_sk_callback_sse41+0x8bf>
+ DB 15,40,5,49,35,0,0 ; movaps 0x2331(%rip),%xmm0 # 4ce0 <_sk_callback_sse41+0x9c5>
DB 15,88,200 ; addps %xmm0,%xmm1
DB 65,15,89,209 ; mulps %xmm9,%xmm2
DB 15,88,208 ; addps %xmm0,%xmm2
- DB 68,15,88,5,41,34,0,0 ; addps 0x2229(%rip),%xmm8 # 4360 <_sk_callback_sse41+0x8cf>
- DB 68,15,89,5,49,34,0,0 ; mulps 0x2231(%rip),%xmm8 # 4370 <_sk_callback_sse41+0x8df>
- DB 15,89,13,58,34,0,0 ; mulps 0x223a(%rip),%xmm1 # 4380 <_sk_callback_sse41+0x8ef>
+ DB 68,15,88,5,47,35,0,0 ; addps 0x232f(%rip),%xmm8 # 4cf0 <_sk_callback_sse41+0x9d5>
+ DB 68,15,89,5,55,35,0,0 ; mulps 0x2337(%rip),%xmm8 # 4d00 <_sk_callback_sse41+0x9e5>
+ DB 15,89,13,64,35,0,0 ; mulps 0x2340(%rip),%xmm1 # 4d10 <_sk_callback_sse41+0x9f5>
DB 65,15,88,200 ; addps %xmm8,%xmm1
- DB 15,89,21,63,34,0,0 ; mulps 0x223f(%rip),%xmm2 # 4390 <_sk_callback_sse41+0x8ff>
+ DB 15,89,21,69,35,0,0 ; mulps 0x2345(%rip),%xmm2 # 4d20 <_sk_callback_sse41+0xa05>
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 68,15,40,13,51,34,0,0 ; movaps 0x2233(%rip),%xmm9 # 43a0 <_sk_callback_sse41+0x90f>
+ DB 68,15,40,13,57,35,0,0 ; movaps 0x2339(%rip),%xmm9 # 4d30 <_sk_callback_sse41+0xa15>
DB 65,15,40,193 ; movaps %xmm9,%xmm0
DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0
- DB 15,40,21,51,34,0,0 ; movaps 0x2233(%rip),%xmm2 # 43b0 <_sk_callback_sse41+0x91f>
+ DB 15,40,21,57,35,0,0 ; movaps 0x2339(%rip),%xmm2 # 4d40 <_sk_callback_sse41+0xa25>
DB 15,88,202 ; addps %xmm2,%xmm1
- DB 68,15,40,37,56,34,0,0 ; movaps 0x2238(%rip),%xmm12 # 43c0 <_sk_callback_sse41+0x92f>
+ DB 68,15,40,37,62,35,0,0 ; movaps 0x233e(%rip),%xmm12 # 4d50 <_sk_callback_sse41+0xa35>
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
@@ -12135,8 +13475,8 @@ _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 15,89,13,241,33,0,0 ; mulps 0x21f1(%rip),%xmm1 # 43d0 <_sk_callback_sse41+0x93f>
- DB 15,89,21,250,33,0,0 ; mulps 0x21fa(%rip),%xmm2 # 43e0 <_sk_callback_sse41+0x94f>
+ DB 15,89,13,247,34,0,0 ; mulps 0x22f7(%rip),%xmm1 # 4d60 <_sk_callback_sse41+0xa45>
+ DB 15,89,21,0,35,0,0 ; mulps 0x2300(%rip),%xmm2 # 4d70 <_sk_callback_sse41+0xa55>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,40,193 ; movaps %xmm1,%xmm0
DB 65,15,40,200 ; movaps %xmm8,%xmm1
@@ -12148,7 +13488,7 @@ _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 15,89,29,234,33,0,0 ; mulps 0x21ea(%rip),%xmm3 # 43f0 <_sk_callback_sse41+0x95f>
+ DB 15,89,29,240,34,0,0 ; mulps 0x22f0(%rip),%xmm3 # 4d80 <_sk_callback_sse41+0xa65>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
DB 15,87,201 ; xorps %xmm1,%xmm1
@@ -12179,7 +13519,7 @@ _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 15,89,29,126,33,0,0 ; mulps 0x217e(%rip),%xmm3 # 4400 <_sk_callback_sse41+0x96f>
+ DB 15,89,29,132,34,0,0 ; mulps 0x2284(%rip),%xmm3 # 4d90 <_sk_callback_sse41+0xa75>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
DB 102,15,239,201 ; pxor %xmm1,%xmm1
@@ -12190,7 +13530,7 @@ 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 68,15,40,5,114,33,0,0 ; movaps 0x2172(%rip),%xmm8 # 4410 <_sk_callback_sse41+0x97f>
+ DB 68,15,40,5,120,34,0,0 ; movaps 0x2278(%rip),%xmm8 # 4da0 <_sk_callback_sse41+0xa85>
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
@@ -12205,9 +13545,9 @@ _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 15,89,5,79,33,0,0 ; mulps 0x214f(%rip),%xmm0 # 4420 <_sk_callback_sse41+0x98f>
+ DB 15,89,5,85,34,0,0 ; mulps 0x2255(%rip),%xmm0 # 4db0 <_sk_callback_sse41+0xa95>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,86,33,0,0 ; movaps 0x2156(%rip),%xmm3 # 4430 <_sk_callback_sse41+0x99f>
+ DB 15,40,29,92,34,0,0 ; movaps 0x225c(%rip),%xmm3 # 4dc0 <_sk_callback_sse41+0xaa5>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 15,40,208 ; movaps %xmm0,%xmm2
DB 255,224 ; jmpq *%rax
@@ -12236,9 +13576,9 @@ _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 15,89,5,239,32,0,0 ; mulps 0x20ef(%rip),%xmm0 # 4440 <_sk_callback_sse41+0x9af>
+ DB 15,89,5,245,33,0,0 ; mulps 0x21f5(%rip),%xmm0 # 4dd0 <_sk_callback_sse41+0xab5>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,246,32,0,0 ; movaps 0x20f6(%rip),%xmm3 # 4450 <_sk_callback_sse41+0x9bf>
+ DB 15,40,29,252,33,0,0 ; movaps 0x21fc(%rip),%xmm3 # 4de0 <_sk_callback_sse41+0xac5>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 15,40,208 ; movaps %xmm0,%xmm2
DB 255,224 ; jmpq *%rax
@@ -12248,9 +13588,9 @@ _sk_gather_i8_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,137,192 ; mov %rax,%r8
DB 77,133,192 ; test %r8,%r8
- DB 116,5 ; je 2371 <_sk_gather_i8_sse41+0xf>
+ DB 116,5 ; je 2bfb <_sk_gather_i8_sse41+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 2373 <_sk_gather_i8_sse41+0x11>
+ DB 235,2 ; jmp 2bfd <_sk_gather_i8_sse41+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1
@@ -12281,17 +13621,17 @@ _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 102,15,111,5,77,32,0,0 ; movdqa 0x204d(%rip),%xmm0 # 4460 <_sk_callback_sse41+0x9cf>
+ DB 102,15,111,5,83,33,0,0 ; movdqa 0x2153(%rip),%xmm0 # 4df0 <_sk_callback_sse41+0xad5>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,78,32,0,0 ; movaps 0x204e(%rip),%xmm8 # 4470 <_sk_callback_sse41+0x9df>
+ DB 68,15,40,5,84,33,0,0 ; movaps 0x2154(%rip),%xmm8 # 4e00 <_sk_callback_sse41+0xae5>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 102,15,111,203 ; movdqa %xmm3,%xmm1
- DB 102,15,56,0,13,77,32,0,0 ; pshufb 0x204d(%rip),%xmm1 # 4480 <_sk_callback_sse41+0x9ef>
+ DB 102,15,56,0,13,83,33,0,0 ; pshufb 0x2153(%rip),%xmm1 # 4e10 <_sk_callback_sse41+0xaf5>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,111,211 ; movdqa %xmm3,%xmm2
- DB 102,15,56,0,21,73,32,0,0 ; pshufb 0x2049(%rip),%xmm2 # 4490 <_sk_callback_sse41+0x9ff>
+ DB 102,15,56,0,21,79,33,0,0 ; pshufb 0x214f(%rip),%xmm2 # 4e20 <_sk_callback_sse41+0xb05>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12305,19 +13645,19 @@ _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 102,15,111,5,47,32,0,0 ; movdqa 0x202f(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xa0f>
+ DB 102,15,111,5,53,33,0,0 ; movdqa 0x2135(%rip),%xmm0 # 4e30 <_sk_callback_sse41+0xb15>
DB 102,15,219,194 ; pand %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,49,32,0,0 ; mulps 0x2031(%rip),%xmm0 # 44b0 <_sk_callback_sse41+0xa1f>
- DB 102,15,111,13,57,32,0,0 ; movdqa 0x2039(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xa2f>
+ DB 15,89,5,55,33,0,0 ; mulps 0x2137(%rip),%xmm0 # 4e40 <_sk_callback_sse41+0xb25>
+ DB 102,15,111,13,63,33,0,0 ; movdqa 0x213f(%rip),%xmm1 # 4e50 <_sk_callback_sse41+0xb35>
DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,59,32,0,0 ; mulps 0x203b(%rip),%xmm1 # 44d0 <_sk_callback_sse41+0xa3f>
- DB 102,15,219,21,67,32,0,0 ; pand 0x2043(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xa4f>
+ DB 15,89,13,65,33,0,0 ; mulps 0x2141(%rip),%xmm1 # 4e60 <_sk_callback_sse41+0xb45>
+ DB 102,15,219,21,73,33,0,0 ; pand 0x2149(%rip),%xmm2 # 4e70 <_sk_callback_sse41+0xb55>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,73,32,0,0 ; mulps 0x2049(%rip),%xmm2 # 44f0 <_sk_callback_sse41+0xa5f>
+ DB 15,89,21,79,33,0,0 ; mulps 0x214f(%rip),%xmm2 # 4e80 <_sk_callback_sse41+0xb65>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,80,32,0,0 ; movaps 0x2050(%rip),%xmm3 # 4500 <_sk_callback_sse41+0xa6f>
+ DB 15,40,29,86,33,0,0 ; movaps 0x2156(%rip),%xmm3 # 4e90 <_sk_callback_sse41+0xb75>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_gather_565_sse41
@@ -12343,31 +13683,31 @@ _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 102,15,111,5,245,31,0,0 ; movdqa 0x1ff5(%rip),%xmm0 # 4510 <_sk_callback_sse41+0xa7f>
+ DB 102,15,111,5,251,32,0,0 ; movdqa 0x20fb(%rip),%xmm0 # 4ea0 <_sk_callback_sse41+0xb85>
DB 102,15,219,194 ; pand %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,247,31,0,0 ; mulps 0x1ff7(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xa8f>
- DB 102,15,111,13,255,31,0,0 ; movdqa 0x1fff(%rip),%xmm1 # 4530 <_sk_callback_sse41+0xa9f>
+ DB 15,89,5,253,32,0,0 ; mulps 0x20fd(%rip),%xmm0 # 4eb0 <_sk_callback_sse41+0xb95>
+ DB 102,15,111,13,5,33,0,0 ; movdqa 0x2105(%rip),%xmm1 # 4ec0 <_sk_callback_sse41+0xba5>
DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,1,32,0,0 ; mulps 0x2001(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xaaf>
- DB 102,15,219,21,9,32,0,0 ; pand 0x2009(%rip),%xmm2 # 4550 <_sk_callback_sse41+0xabf>
+ DB 15,89,13,7,33,0,0 ; mulps 0x2107(%rip),%xmm1 # 4ed0 <_sk_callback_sse41+0xbb5>
+ DB 102,15,219,21,15,33,0,0 ; pand 0x210f(%rip),%xmm2 # 4ee0 <_sk_callback_sse41+0xbc5>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,15,32,0,0 ; mulps 0x200f(%rip),%xmm2 # 4560 <_sk_callback_sse41+0xacf>
+ DB 15,89,21,21,33,0,0 ; mulps 0x2115(%rip),%xmm2 # 4ef0 <_sk_callback_sse41+0xbd5>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,22,32,0,0 ; movaps 0x2016(%rip),%xmm3 # 4570 <_sk_callback_sse41+0xadf>
+ DB 15,40,29,28,33,0,0 ; movaps 0x211c(%rip),%xmm3 # 4f00 <_sk_callback_sse41+0xbe5>
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 68,15,40,5,23,32,0,0 ; movaps 0x2017(%rip),%xmm8 # 4580 <_sk_callback_sse41+0xaef>
+ DB 68,15,40,5,29,33,0,0 ; movaps 0x211d(%rip),%xmm8 # 4f10 <_sk_callback_sse41+0xbf5>
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 68,15,40,21,12,32,0,0 ; movaps 0x200c(%rip),%xmm10 # 4590 <_sk_callback_sse41+0xaff>
+ DB 68,15,40,21,18,33,0,0 ; movaps 0x2112(%rip),%xmm10 # 4f20 <_sk_callback_sse41+0xc05>
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
@@ -12385,21 +13725,21 @@ _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 102,15,111,5,215,31,0,0 ; movdqa 0x1fd7(%rip),%xmm0 # 45a0 <_sk_callback_sse41+0xb0f>
+ DB 102,15,111,5,221,32,0,0 ; movdqa 0x20dd(%rip),%xmm0 # 4f30 <_sk_callback_sse41+0xc15>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,217,31,0,0 ; mulps 0x1fd9(%rip),%xmm0 # 45b0 <_sk_callback_sse41+0xb1f>
- DB 102,15,111,13,225,31,0,0 ; movdqa 0x1fe1(%rip),%xmm1 # 45c0 <_sk_callback_sse41+0xb2f>
+ DB 15,89,5,223,32,0,0 ; mulps 0x20df(%rip),%xmm0 # 4f40 <_sk_callback_sse41+0xc25>
+ DB 102,15,111,13,231,32,0,0 ; movdqa 0x20e7(%rip),%xmm1 # 4f50 <_sk_callback_sse41+0xc35>
DB 102,15,219,203 ; pand %xmm3,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,227,31,0,0 ; mulps 0x1fe3(%rip),%xmm1 # 45d0 <_sk_callback_sse41+0xb3f>
- DB 102,15,111,21,235,31,0,0 ; movdqa 0x1feb(%rip),%xmm2 # 45e0 <_sk_callback_sse41+0xb4f>
+ DB 15,89,13,233,32,0,0 ; mulps 0x20e9(%rip),%xmm1 # 4f60 <_sk_callback_sse41+0xc45>
+ DB 102,15,111,21,241,32,0,0 ; movdqa 0x20f1(%rip),%xmm2 # 4f70 <_sk_callback_sse41+0xc55>
DB 102,15,219,211 ; pand %xmm3,%xmm2
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,237,31,0,0 ; mulps 0x1fed(%rip),%xmm2 # 45f0 <_sk_callback_sse41+0xb5f>
- DB 102,15,219,29,245,31,0,0 ; pand 0x1ff5(%rip),%xmm3 # 4600 <_sk_callback_sse41+0xb6f>
+ DB 15,89,21,243,32,0,0 ; mulps 0x20f3(%rip),%xmm2 # 4f80 <_sk_callback_sse41+0xc65>
+ DB 102,15,219,29,251,32,0,0 ; pand 0x20fb(%rip),%xmm3 # 4f90 <_sk_callback_sse41+0xc75>
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,251,31,0,0 ; mulps 0x1ffb(%rip),%xmm3 # 4610 <_sk_callback_sse41+0xb7f>
+ DB 15,89,29,1,33,0,0 ; mulps 0x2101(%rip),%xmm3 # 4fa0 <_sk_callback_sse41+0xc85>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -12426,21 +13766,21 @@ _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 102,15,111,5,158,31,0,0 ; movdqa 0x1f9e(%rip),%xmm0 # 4620 <_sk_callback_sse41+0xb8f>
+ DB 102,15,111,5,164,32,0,0 ; movdqa 0x20a4(%rip),%xmm0 # 4fb0 <_sk_callback_sse41+0xc95>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,160,31,0,0 ; mulps 0x1fa0(%rip),%xmm0 # 4630 <_sk_callback_sse41+0xb9f>
- DB 102,15,111,13,168,31,0,0 ; movdqa 0x1fa8(%rip),%xmm1 # 4640 <_sk_callback_sse41+0xbaf>
+ DB 15,89,5,166,32,0,0 ; mulps 0x20a6(%rip),%xmm0 # 4fc0 <_sk_callback_sse41+0xca5>
+ DB 102,15,111,13,174,32,0,0 ; movdqa 0x20ae(%rip),%xmm1 # 4fd0 <_sk_callback_sse41+0xcb5>
DB 102,15,219,203 ; pand %xmm3,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,170,31,0,0 ; mulps 0x1faa(%rip),%xmm1 # 4650 <_sk_callback_sse41+0xbbf>
- DB 102,15,111,21,178,31,0,0 ; movdqa 0x1fb2(%rip),%xmm2 # 4660 <_sk_callback_sse41+0xbcf>
+ DB 15,89,13,176,32,0,0 ; mulps 0x20b0(%rip),%xmm1 # 4fe0 <_sk_callback_sse41+0xcc5>
+ DB 102,15,111,21,184,32,0,0 ; movdqa 0x20b8(%rip),%xmm2 # 4ff0 <_sk_callback_sse41+0xcd5>
DB 102,15,219,211 ; pand %xmm3,%xmm2
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,180,31,0,0 ; mulps 0x1fb4(%rip),%xmm2 # 4670 <_sk_callback_sse41+0xbdf>
- DB 102,15,219,29,188,31,0,0 ; pand 0x1fbc(%rip),%xmm3 # 4680 <_sk_callback_sse41+0xbef>
+ DB 15,89,21,186,32,0,0 ; mulps 0x20ba(%rip),%xmm2 # 5000 <_sk_callback_sse41+0xce5>
+ DB 102,15,219,29,194,32,0,0 ; pand 0x20c2(%rip),%xmm3 # 5010 <_sk_callback_sse41+0xcf5>
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,194,31,0,0 ; mulps 0x1fc2(%rip),%xmm3 # 4690 <_sk_callback_sse41+0xbff>
+ DB 15,89,29,200,32,0,0 ; mulps 0x20c8(%rip),%xmm3 # 5020 <_sk_callback_sse41+0xd05>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -12448,7 +13788,7 @@ 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 68,15,40,5,193,31,0,0 ; movaps 0x1fc1(%rip),%xmm8 # 46a0 <_sk_callback_sse41+0xc0f>
+ DB 68,15,40,5,199,32,0,0 ; movaps 0x20c7(%rip),%xmm8 # 5030 <_sk_callback_sse41+0xd15>
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
@@ -12476,17 +13816,17 @@ _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 15,40,5,96,31,0,0 ; movaps 0x1f60(%rip),%xmm0 # 46b0 <_sk_callback_sse41+0xc1f>
+ DB 15,40,5,102,32,0,0 ; movaps 0x2066(%rip),%xmm0 # 5040 <_sk_callback_sse41+0xd25>
DB 15,84,195 ; andps %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,98,31,0,0 ; movaps 0x1f62(%rip),%xmm8 # 46c0 <_sk_callback_sse41+0xc2f>
+ DB 68,15,40,5,104,32,0,0 ; movaps 0x2068(%rip),%xmm8 # 5050 <_sk_callback_sse41+0xd35>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,40,203 ; movaps %xmm3,%xmm1
- DB 102,15,56,0,13,98,31,0,0 ; pshufb 0x1f62(%rip),%xmm1 # 46d0 <_sk_callback_sse41+0xc3f>
+ DB 102,15,56,0,13,104,32,0,0 ; pshufb 0x2068(%rip),%xmm1 # 5060 <_sk_callback_sse41+0xd45>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 15,40,211 ; movaps %xmm3,%xmm2
- DB 102,15,56,0,21,95,31,0,0 ; pshufb 0x1f5f(%rip),%xmm2 # 46e0 <_sk_callback_sse41+0xc4f>
+ DB 102,15,56,0,21,101,32,0,0 ; pshufb 0x2065(%rip),%xmm2 # 5070 <_sk_callback_sse41+0xd55>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12515,17 +13855,17 @@ _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 102,15,111,5,248,30,0,0 ; movdqa 0x1ef8(%rip),%xmm0 # 46f0 <_sk_callback_sse41+0xc5f>
+ DB 102,15,111,5,254,31,0,0 ; movdqa 0x1ffe(%rip),%xmm0 # 5080 <_sk_callback_sse41+0xd65>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,249,30,0,0 ; movaps 0x1ef9(%rip),%xmm8 # 4700 <_sk_callback_sse41+0xc6f>
+ DB 68,15,40,5,255,31,0,0 ; movaps 0x1fff(%rip),%xmm8 # 5090 <_sk_callback_sse41+0xd75>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 102,15,111,203 ; movdqa %xmm3,%xmm1
- DB 102,15,56,0,13,248,30,0,0 ; pshufb 0x1ef8(%rip),%xmm1 # 4710 <_sk_callback_sse41+0xc7f>
+ DB 102,15,56,0,13,254,31,0,0 ; pshufb 0x1ffe(%rip),%xmm1 # 50a0 <_sk_callback_sse41+0xd85>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,111,211 ; movdqa %xmm3,%xmm2
- DB 102,15,56,0,21,244,30,0,0 ; pshufb 0x1ef4(%rip),%xmm2 # 4720 <_sk_callback_sse41+0xc8f>
+ DB 102,15,56,0,21,250,31,0,0 ; pshufb 0x1ffa(%rip),%xmm2 # 50b0 <_sk_callback_sse41+0xd95>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12538,7 +13878,7 @@ 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 68,15,40,5,224,30,0,0 ; movaps 0x1ee0(%rip),%xmm8 # 4730 <_sk_callback_sse41+0xc9f>
+ DB 68,15,40,5,230,31,0,0 ; movaps 0x1fe6(%rip),%xmm8 # 50c0 <_sk_callback_sse41+0xda5>
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
@@ -12573,18 +13913,18 @@ _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 102,68,15,111,5,89,30,0,0 ; movdqa 0x1e59(%rip),%xmm8 # 4740 <_sk_callback_sse41+0xcaf>
+ DB 102,68,15,111,5,95,31,0,0 ; movdqa 0x1f5f(%rip),%xmm8 # 50d0 <_sk_callback_sse41+0xdb5>
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 102,15,111,29,84,30,0,0 ; movdqa 0x1e54(%rip),%xmm3 # 4750 <_sk_callback_sse41+0xcbf>
+ DB 102,15,111,29,90,31,0,0 ; movdqa 0x1f5a(%rip),%xmm3 # 50e0 <_sk_callback_sse41+0xdc5>
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 102,68,15,111,21,64,30,0,0 ; movdqa 0x1e40(%rip),%xmm10 # 4760 <_sk_callback_sse41+0xccf>
+ DB 102,68,15,111,21,70,31,0,0 ; movdqa 0x1f46(%rip),%xmm10 # 50f0 <_sk_callback_sse41+0xdd5>
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
@@ -12655,18 +13995,18 @@ _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 102,68,15,111,5,254,28,0,0 ; movdqa 0x1cfe(%rip),%xmm8 # 4770 <_sk_callback_sse41+0xcdf>
+ DB 102,68,15,111,5,4,30,0,0 ; movdqa 0x1e04(%rip),%xmm8 # 5100 <_sk_callback_sse41+0xde5>
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 102,15,111,29,249,28,0,0 ; movdqa 0x1cf9(%rip),%xmm3 # 4780 <_sk_callback_sse41+0xcef>
+ DB 102,15,111,29,255,29,0,0 ; movdqa 0x1dff(%rip),%xmm3 # 5110 <_sk_callback_sse41+0xdf5>
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 102,68,15,111,21,229,28,0,0 ; movdqa 0x1ce5(%rip),%xmm10 # 4790 <_sk_callback_sse41+0xcff>
+ DB 102,68,15,111,21,235,29,0,0 ; movdqa 0x1deb(%rip),%xmm10 # 5120 <_sk_callback_sse41+0xe05>
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
@@ -12712,17 +14052,17 @@ 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 102,68,15,111,21,27,28,0,0 ; movdqa 0x1c1b(%rip),%xmm10 # 47a0 <_sk_callback_sse41+0xd0f>
+ DB 102,68,15,111,21,33,29,0,0 ; movdqa 0x1d21(%rip),%xmm10 # 5130 <_sk_callback_sse41+0xe15>
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 102,68,15,111,13,14,28,0,0 ; movdqa 0x1c0e(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0xd1f>
+ DB 102,68,15,111,13,20,29,0,0 ; movdqa 0x1d14(%rip),%xmm9 # 5140 <_sk_callback_sse41+0xe25>
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 102,68,15,111,29,255,27,0,0 ; movdqa 0x1bff(%rip),%xmm11 # 47c0 <_sk_callback_sse41+0xd2f>
+ DB 102,68,15,111,29,5,29,0,0 ; movdqa 0x1d05(%rip),%xmm11 # 5150 <_sk_callback_sse41+0xe35>
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
@@ -12790,7 +14130,7 @@ _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 68,15,40,5,206,26,0,0 ; movaps 0x1ace(%rip),%xmm8 # 47d0 <_sk_callback_sse41+0xd3f>
+ DB 68,15,40,5,212,27,0,0 ; movaps 0x1bd4(%rip),%xmm8 # 5160 <_sk_callback_sse41+0xe45>
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
@@ -12840,7 +14180,7 @@ _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 68,15,40,5,15,26,0,0 ; movaps 0x1a0f(%rip),%xmm8 # 47e0 <_sk_callback_sse41+0xd4f>
+ DB 68,15,40,5,21,27,0,0 ; movaps 0x1b15(%rip),%xmm8 # 5170 <_sk_callback_sse41+0xe55>
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
@@ -12857,14 +14197,14 @@ _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 15,40,29,214,25,0,0 ; movaps 0x19d6(%rip),%xmm3 # 47f0 <_sk_callback_sse41+0xd5f>
+ DB 15,40,29,220,26,0,0 ; movaps 0x1adc(%rip),%xmm3 # 5180 <_sk_callback_sse41+0xe65>
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 68,15,40,13,215,25,0,0 ; movaps 0x19d7(%rip),%xmm9 # 4800 <_sk_callback_sse41+0xd6f>
+ DB 68,15,40,13,221,26,0,0 ; movaps 0x1add(%rip),%xmm9 # 5190 <_sk_callback_sse41+0xe75>
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
@@ -13069,10 +14409,10 @@ _sk_mirror_y_sse41 LABEL PROC
PUBLIC _sk_luminance_to_alpha_sse41
_sk_luminance_to_alpha_sse41 LABEL PROC
DB 15,40,218 ; movaps %xmm2,%xmm3
- DB 15,89,5,245,22,0,0 ; mulps 0x16f5(%rip),%xmm0 # 4810 <_sk_callback_sse41+0xd7f>
- DB 15,89,13,254,22,0,0 ; mulps 0x16fe(%rip),%xmm1 # 4820 <_sk_callback_sse41+0xd8f>
+ DB 15,89,5,251,23,0,0 ; mulps 0x17fb(%rip),%xmm0 # 51a0 <_sk_callback_sse41+0xe85>
+ DB 15,89,13,4,24,0,0 ; mulps 0x1804(%rip),%xmm1 # 51b0 <_sk_callback_sse41+0xe95>
DB 15,88,200 ; addps %xmm0,%xmm1
- DB 15,89,29,4,23,0,0 ; mulps 0x1704(%rip),%xmm3 # 4830 <_sk_callback_sse41+0xd9f>
+ DB 15,89,29,10,24,0,0 ; mulps 0x180a(%rip),%xmm3 # 51c0 <_sk_callback_sse41+0xea5>
DB 15,88,217 ; addps %xmm1,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
@@ -13295,7 +14635,7 @@ _sk_linear_gradient_sse41 LABEL PROC
DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13
DB 72,139,8 ; mov (%rax),%rcx
DB 72,133,201 ; test %rcx,%rcx
- DB 15,132,4,1,0,0 ; je 35d4 <_sk_linear_gradient_sse41+0x13e>
+ DB 15,132,4,1,0,0 ; je 3e5e <_sk_linear_gradient_sse41+0x13e>
DB 72,131,236,88 ; sub $0x58,%rsp
DB 15,41,36,36 ; movaps %xmm4,(%rsp)
DB 15,41,108,36,16 ; movaps %xmm5,0x10(%rsp)
@@ -13346,13 +14686,13 @@ _sk_linear_gradient_sse41 LABEL PROC
DB 15,40,196 ; movaps %xmm4,%xmm0
DB 72,131,192,36 ; add $0x24,%rax
DB 72,255,201 ; dec %rcx
- DB 15,133,65,255,255,255 ; jne 34fc <_sk_linear_gradient_sse41+0x66>
+ DB 15,133,65,255,255,255 ; jne 3d86 <_sk_linear_gradient_sse41+0x66>
DB 15,40,124,36,48 ; movaps 0x30(%rsp),%xmm7
DB 15,40,116,36,32 ; movaps 0x20(%rsp),%xmm6
DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5
DB 15,40,36,36 ; movaps (%rsp),%xmm4
DB 72,131,196,88 ; add $0x58,%rsp
- DB 235,13 ; jmp 35e1 <_sk_linear_gradient_sse41+0x14b>
+ DB 235,13 ; jmp 3e6b <_sk_linear_gradient_sse41+0x14b>
DB 15,87,201 ; xorps %xmm1,%xmm1
DB 15,87,210 ; xorps %xmm2,%xmm2
DB 15,87,219 ; xorps %xmm3,%xmm3
@@ -13417,26 +14757,26 @@ _sk_xy_to_polar_unit_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 68,15,40,21,135,17,0,0 ; movaps 0x1187(%rip),%xmm10 # 4840 <_sk_callback_sse41+0xdaf>
+ DB 68,15,40,21,141,18,0,0 ; movaps 0x128d(%rip),%xmm10 # 51d0 <_sk_callback_sse41+0xeb5>
DB 69,15,89,213 ; mulps %xmm13,%xmm10
- DB 68,15,88,21,139,17,0,0 ; addps 0x118b(%rip),%xmm10 # 4850 <_sk_callback_sse41+0xdbf>
+ DB 68,15,88,21,145,18,0,0 ; addps 0x1291(%rip),%xmm10 # 51e0 <_sk_callback_sse41+0xec5>
DB 69,15,89,213 ; mulps %xmm13,%xmm10
- DB 68,15,88,21,143,17,0,0 ; addps 0x118f(%rip),%xmm10 # 4860 <_sk_callback_sse41+0xdcf>
+ DB 68,15,88,21,149,18,0,0 ; addps 0x1295(%rip),%xmm10 # 51f0 <_sk_callback_sse41+0xed5>
DB 69,15,89,213 ; mulps %xmm13,%xmm10
- DB 68,15,88,21,147,17,0,0 ; addps 0x1193(%rip),%xmm10 # 4870 <_sk_callback_sse41+0xddf>
+ DB 68,15,88,21,153,18,0,0 ; addps 0x1299(%rip),%xmm10 # 5200 <_sk_callback_sse41+0xee5>
DB 69,15,89,212 ; mulps %xmm12,%xmm10
DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0
- DB 68,15,40,29,146,17,0,0 ; movaps 0x1192(%rip),%xmm11 # 4880 <_sk_callback_sse41+0xdef>
+ DB 68,15,40,29,152,18,0,0 ; movaps 0x1298(%rip),%xmm11 # 5210 <_sk_callback_sse41+0xef5>
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 68,15,40,29,139,17,0,0 ; movaps 0x118b(%rip),%xmm11 # 4890 <_sk_callback_sse41+0xdff>
+ DB 68,15,40,29,145,18,0,0 ; movaps 0x1291(%rip),%xmm11 # 5220 <_sk_callback_sse41+0xf05>
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 68,15,40,13,125,17,0,0 ; movaps 0x117d(%rip),%xmm9 # 48a0 <_sk_callback_sse41+0xe0f>
+ DB 68,15,40,13,131,18,0,0 ; movaps 0x1283(%rip),%xmm9 # 5230 <_sk_callback_sse41+0xf15>
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
@@ -13448,7 +14788,7 @@ _sk_xy_to_polar_unit_sse41 LABEL PROC
PUBLIC _sk_save_xy_sse41
_sk_save_xy_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 68,15,40,5,104,17,0,0 ; movaps 0x1168(%rip),%xmm8 # 48b0 <_sk_callback_sse41+0xe1f>
+ DB 68,15,40,5,110,18,0,0 ; movaps 0x126e(%rip),%xmm8 # 5240 <_sk_callback_sse41+0xf25>
DB 15,17,0 ; movups %xmm0,(%rax)
DB 68,15,40,200 ; movaps %xmm0,%xmm9
DB 69,15,88,200 ; addps %xmm8,%xmm9
@@ -13488,8 +14828,8 @@ _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 15,88,5,234,16,0,0 ; addps 0x10ea(%rip),%xmm0 # 48c0 <_sk_callback_sse41+0xe2f>
- DB 68,15,40,13,242,16,0,0 ; movaps 0x10f2(%rip),%xmm9 # 48d0 <_sk_callback_sse41+0xe3f>
+ DB 15,88,5,240,17,0,0 ; addps 0x11f0(%rip),%xmm0 # 5250 <_sk_callback_sse41+0xf35>
+ DB 68,15,40,13,248,17,0,0 ; movaps 0x11f8(%rip),%xmm9 # 5260 <_sk_callback_sse41+0xf45>
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
@@ -13500,7 +14840,7 @@ _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 15,88,5,225,16,0,0 ; addps 0x10e1(%rip),%xmm0 # 48e0 <_sk_callback_sse41+0xe4f>
+ DB 15,88,5,231,17,0,0 ; addps 0x11e7(%rip),%xmm0 # 5270 <_sk_callback_sse41+0xf55>
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
@@ -13510,8 +14850,8 @@ _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 15,88,13,211,16,0,0 ; addps 0x10d3(%rip),%xmm1 # 48f0 <_sk_callback_sse41+0xe5f>
- DB 68,15,40,13,219,16,0,0 ; movaps 0x10db(%rip),%xmm9 # 4900 <_sk_callback_sse41+0xe6f>
+ DB 15,88,13,217,17,0,0 ; addps 0x11d9(%rip),%xmm1 # 5280 <_sk_callback_sse41+0xf65>
+ DB 68,15,40,13,225,17,0,0 ; movaps 0x11e1(%rip),%xmm9 # 5290 <_sk_callback_sse41+0xf75>
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
@@ -13522,7 +14862,7 @@ _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 15,88,13,201,16,0,0 ; addps 0x10c9(%rip),%xmm1 # 4910 <_sk_callback_sse41+0xe7f>
+ DB 15,88,13,207,17,0,0 ; addps 0x11cf(%rip),%xmm1 # 52a0 <_sk_callback_sse41+0xf85>
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
@@ -13532,13 +14872,13 @@ _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 15,88,5,188,16,0,0 ; addps 0x10bc(%rip),%xmm0 # 4920 <_sk_callback_sse41+0xe8f>
- DB 68,15,40,13,196,16,0,0 ; movaps 0x10c4(%rip),%xmm9 # 4930 <_sk_callback_sse41+0xe9f>
+ DB 15,88,5,194,17,0,0 ; addps 0x11c2(%rip),%xmm0 # 52b0 <_sk_callback_sse41+0xf95>
+ DB 68,15,40,13,202,17,0,0 ; movaps 0x11ca(%rip),%xmm9 # 52c0 <_sk_callback_sse41+0xfa5>
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 68,15,89,13,192,16,0,0 ; mulps 0x10c0(%rip),%xmm9 # 4940 <_sk_callback_sse41+0xeaf>
- DB 68,15,88,13,200,16,0,0 ; addps 0x10c8(%rip),%xmm9 # 4950 <_sk_callback_sse41+0xebf>
+ DB 68,15,89,13,198,17,0,0 ; mulps 0x11c6(%rip),%xmm9 # 52d0 <_sk_callback_sse41+0xfb5>
+ DB 68,15,88,13,206,17,0,0 ; addps 0x11ce(%rip),%xmm9 # 52e0 <_sk_callback_sse41+0xfc5>
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
@@ -13549,16 +14889,16 @@ _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 15,88,5,183,16,0,0 ; addps 0x10b7(%rip),%xmm0 # 4960 <_sk_callback_sse41+0xecf>
- DB 68,15,40,13,191,16,0,0 ; movaps 0x10bf(%rip),%xmm9 # 4970 <_sk_callback_sse41+0xedf>
+ DB 15,88,5,189,17,0,0 ; addps 0x11bd(%rip),%xmm0 # 52f0 <_sk_callback_sse41+0xfd5>
+ DB 68,15,40,13,197,17,0,0 ; movaps 0x11c5(%rip),%xmm9 # 5300 <_sk_callback_sse41+0xfe5>
DB 69,15,92,200 ; subps %xmm8,%xmm9
- DB 68,15,40,5,195,16,0,0 ; movaps 0x10c3(%rip),%xmm8 # 4980 <_sk_callback_sse41+0xeef>
+ DB 68,15,40,5,201,17,0,0 ; movaps 0x11c9(%rip),%xmm8 # 5310 <_sk_callback_sse41+0xff5>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,199,16,0,0 ; addps 0x10c7(%rip),%xmm8 # 4990 <_sk_callback_sse41+0xeff>
+ DB 68,15,88,5,205,17,0,0 ; addps 0x11cd(%rip),%xmm8 # 5320 <_sk_callback_sse41+0x1005>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,203,16,0,0 ; addps 0x10cb(%rip),%xmm8 # 49a0 <_sk_callback_sse41+0xf0f>
+ DB 68,15,88,5,209,17,0,0 ; addps 0x11d1(%rip),%xmm8 # 5330 <_sk_callback_sse41+0x1015>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,207,16,0,0 ; addps 0x10cf(%rip),%xmm8 # 49b0 <_sk_callback_sse41+0xf1f>
+ DB 68,15,88,5,213,17,0,0 ; addps 0x11d5(%rip),%xmm8 # 5340 <_sk_callback_sse41+0x1025>
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
@@ -13566,17 +14906,17 @@ _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 68,15,40,5,201,16,0,0 ; movaps 0x10c9(%rip),%xmm8 # 49c0 <_sk_callback_sse41+0xf2f>
+ DB 68,15,40,5,207,17,0,0 ; movaps 0x11cf(%rip),%xmm8 # 5350 <_sk_callback_sse41+0x1035>
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 68,15,40,21,197,16,0,0 ; movaps 0x10c5(%rip),%xmm10 # 49d0 <_sk_callback_sse41+0xf3f>
+ DB 68,15,40,21,203,17,0,0 ; movaps 0x11cb(%rip),%xmm10 # 5360 <_sk_callback_sse41+0x1045>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,201,16,0,0 ; addps 0x10c9(%rip),%xmm10 # 49e0 <_sk_callback_sse41+0xf4f>
+ DB 68,15,88,21,207,17,0,0 ; addps 0x11cf(%rip),%xmm10 # 5370 <_sk_callback_sse41+0x1055>
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 68,15,88,21,197,16,0,0 ; addps 0x10c5(%rip),%xmm10 # 49f0 <_sk_callback_sse41+0xf5f>
+ DB 68,15,88,21,203,17,0,0 ; addps 0x11cb(%rip),%xmm10 # 5380 <_sk_callback_sse41+0x1065>
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
@@ -13586,11 +14926,11 @@ _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 15,88,5,184,16,0,0 ; addps 0x10b8(%rip),%xmm0 # 4a00 <_sk_callback_sse41+0xf6f>
+ DB 15,88,5,190,17,0,0 ; addps 0x11be(%rip),%xmm0 # 5390 <_sk_callback_sse41+0x1075>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,89,201 ; mulps %xmm9,%xmm9
- DB 68,15,89,5,184,16,0,0 ; mulps 0x10b8(%rip),%xmm8 # 4a10 <_sk_callback_sse41+0xf7f>
- DB 68,15,88,5,192,16,0,0 ; addps 0x10c0(%rip),%xmm8 # 4a20 <_sk_callback_sse41+0xf8f>
+ DB 68,15,89,5,190,17,0,0 ; mulps 0x11be(%rip),%xmm8 # 53a0 <_sk_callback_sse41+0x1085>
+ DB 68,15,88,5,198,17,0,0 ; addps 0x11c6(%rip),%xmm8 # 53b0 <_sk_callback_sse41+0x1095>
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
@@ -13601,13 +14941,13 @@ _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 15,88,13,174,16,0,0 ; addps 0x10ae(%rip),%xmm1 # 4a30 <_sk_callback_sse41+0xf9f>
- DB 68,15,40,13,182,16,0,0 ; movaps 0x10b6(%rip),%xmm9 # 4a40 <_sk_callback_sse41+0xfaf>
+ DB 15,88,13,180,17,0,0 ; addps 0x11b4(%rip),%xmm1 # 53c0 <_sk_callback_sse41+0x10a5>
+ DB 68,15,40,13,188,17,0,0 ; movaps 0x11bc(%rip),%xmm9 # 53d0 <_sk_callback_sse41+0x10b5>
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 68,15,89,13,178,16,0,0 ; mulps 0x10b2(%rip),%xmm9 # 4a50 <_sk_callback_sse41+0xfbf>
- DB 68,15,88,13,186,16,0,0 ; addps 0x10ba(%rip),%xmm9 # 4a60 <_sk_callback_sse41+0xfcf>
+ DB 68,15,89,13,184,17,0,0 ; mulps 0x11b8(%rip),%xmm9 # 53e0 <_sk_callback_sse41+0x10c5>
+ DB 68,15,88,13,192,17,0,0 ; addps 0x11c0(%rip),%xmm9 # 53f0 <_sk_callback_sse41+0x10d5>
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
@@ -13618,16 +14958,16 @@ _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 15,88,13,168,16,0,0 ; addps 0x10a8(%rip),%xmm1 # 4a70 <_sk_callback_sse41+0xfdf>
- DB 68,15,40,13,176,16,0,0 ; movaps 0x10b0(%rip),%xmm9 # 4a80 <_sk_callback_sse41+0xfef>
+ DB 15,88,13,174,17,0,0 ; addps 0x11ae(%rip),%xmm1 # 5400 <_sk_callback_sse41+0x10e5>
+ DB 68,15,40,13,182,17,0,0 ; movaps 0x11b6(%rip),%xmm9 # 5410 <_sk_callback_sse41+0x10f5>
DB 69,15,92,200 ; subps %xmm8,%xmm9
- DB 68,15,40,5,180,16,0,0 ; movaps 0x10b4(%rip),%xmm8 # 4a90 <_sk_callback_sse41+0xfff>
+ DB 68,15,40,5,186,17,0,0 ; movaps 0x11ba(%rip),%xmm8 # 5420 <_sk_callback_sse41+0x1105>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,184,16,0,0 ; addps 0x10b8(%rip),%xmm8 # 4aa0 <_sk_callback_sse41+0x100f>
+ DB 68,15,88,5,190,17,0,0 ; addps 0x11be(%rip),%xmm8 # 5430 <_sk_callback_sse41+0x1115>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,188,16,0,0 ; addps 0x10bc(%rip),%xmm8 # 4ab0 <_sk_callback_sse41+0x101f>
+ DB 68,15,88,5,194,17,0,0 ; addps 0x11c2(%rip),%xmm8 # 5440 <_sk_callback_sse41+0x1125>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,192,16,0,0 ; addps 0x10c0(%rip),%xmm8 # 4ac0 <_sk_callback_sse41+0x102f>
+ DB 68,15,88,5,198,17,0,0 ; addps 0x11c6(%rip),%xmm8 # 5450 <_sk_callback_sse41+0x1135>
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
@@ -13635,17 +14975,17 @@ _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 68,15,40,5,186,16,0,0 ; movaps 0x10ba(%rip),%xmm8 # 4ad0 <_sk_callback_sse41+0x103f>
+ DB 68,15,40,5,192,17,0,0 ; movaps 0x11c0(%rip),%xmm8 # 5460 <_sk_callback_sse41+0x1145>
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 68,15,40,21,181,16,0,0 ; movaps 0x10b5(%rip),%xmm10 # 4ae0 <_sk_callback_sse41+0x104f>
+ DB 68,15,40,21,187,17,0,0 ; movaps 0x11bb(%rip),%xmm10 # 5470 <_sk_callback_sse41+0x1155>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,185,16,0,0 ; addps 0x10b9(%rip),%xmm10 # 4af0 <_sk_callback_sse41+0x105f>
+ DB 68,15,88,21,191,17,0,0 ; addps 0x11bf(%rip),%xmm10 # 5480 <_sk_callback_sse41+0x1165>
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 68,15,88,21,181,16,0,0 ; addps 0x10b5(%rip),%xmm10 # 4b00 <_sk_callback_sse41+0x106f>
+ DB 68,15,88,21,187,17,0,0 ; addps 0x11bb(%rip),%xmm10 # 5490 <_sk_callback_sse41+0x1175>
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
@@ -13655,11 +14995,11 @@ _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 15,88,13,167,16,0,0 ; addps 0x10a7(%rip),%xmm1 # 4b10 <_sk_callback_sse41+0x107f>
+ DB 15,88,13,173,17,0,0 ; addps 0x11ad(%rip),%xmm1 # 54a0 <_sk_callback_sse41+0x1185>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,89,201 ; mulps %xmm9,%xmm9
- DB 68,15,89,5,167,16,0,0 ; mulps 0x10a7(%rip),%xmm8 # 4b20 <_sk_callback_sse41+0x108f>
- DB 68,15,88,5,175,16,0,0 ; addps 0x10af(%rip),%xmm8 # 4b30 <_sk_callback_sse41+0x109f>
+ DB 68,15,89,5,173,17,0,0 ; mulps 0x11ad(%rip),%xmm8 # 54b0 <_sk_callback_sse41+0x1195>
+ DB 68,15,88,5,181,17,0,0 ; addps 0x11b5(%rip),%xmm8 # 54c0 <_sk_callback_sse41+0x11a5>
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
@@ -13864,11 +15204,135 @@ 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 3d18 <.literal16+0x1d8>
+ DB 224,64 ; loopne 45a8 <.literal16+0x1d8>
DB 0,0 ; add %al,(%rax)
- DB 224,64 ; loopne 3d1c <.literal16+0x1dc>
+ DB 224,64 ; loopne 45ac <.literal16+0x1dc>
DB 0,0 ; add %al,(%rax)
- DB 224,64 ; loopne 3d20 <.literal16+0x1e0>
+ DB 224,64 ; loopne 45b0 <.literal16+0x1e0>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 45e1 <.literal16+0x211>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 45e5 <.literal16+0x215>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 45e9 <.literal16+0x219>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 45ed <.literal16+0x21d>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4621 <.literal16+0x251>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4625 <.literal16+0x255>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4629 <.literal16+0x259>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 462d <.literal16+0x25d>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4661 <.literal16+0x291>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4665 <.literal16+0x295>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4669 <.literal16+0x299>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 466d <.literal16+0x29d>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 46a1 <.literal16+0x2d1>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 46a5 <.literal16+0x2d5>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 46a9 <.literal16+0x2d9>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 46ad <.literal16+0x2dd>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax)
@@ -14087,13 +15551,13 @@ ALIGN 16
DB 132,55 ; test %dh,(%rdi)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 3ed9 <.literal16+0x399>
+ DB 224,7 ; loopne 4869 <.literal16+0x499>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 3edd <.literal16+0x39d>
+ DB 224,7 ; loopne 486d <.literal16+0x49d>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 3ee1 <.literal16+0x3a1>
+ DB 224,7 ; loopne 4871 <.literal16+0x4a1>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 3ee5 <.literal16+0x3a5>
+ DB 224,7 ; loopne 4875 <.literal16+0x4a5>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -14133,10 +15597,10 @@ ALIGN 16
DB 0,1 ; add %al,(%rcx)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a003f38 <_sk_callback_sse41+0xa0004a7>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0048c8 <_sk_callback_sse41+0xa0005ad>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3003f40 <_sk_callback_sse41+0x30004af>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30048d0 <_sk_callback_sse41+0x30005b5>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14191,11 +15655,11 @@ 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 400b <.literal16+0x4cb>
+ DB 127,67 ; jg 499b <.literal16+0x5cb>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 400f <.literal16+0x4cf>
+ DB 127,67 ; jg 499f <.literal16+0x5cf>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4013 <.literal16+0x4d3>
+ DB 127,67 ; jg 49a3 <.literal16+0x5d3>
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)
@@ -14210,16 +15674,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4004 <.literal16+0x4c4>
+ DB 127,0 ; jg 4994 <.literal16+0x5c4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4008 <.literal16+0x4c8>
+ DB 127,0 ; jg 4998 <.literal16+0x5c8>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 400c <.literal16+0x4cc>
+ DB 127,0 ; jg 499c <.literal16+0x5cc>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4010 <.literal16+0x4d0>
+ DB 127,0 ; jg 49a0 <.literal16+0x5d0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14228,7 +15692,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4095 <.literal16+0x555>
+ DB 119,115 ; ja 4a25 <.literal16+0x655>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14239,7 +15703,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 3ff9 <.literal16+0x4b9>
+ DB 117,191 ; jne 4989 <.literal16+0x5b9>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14251,7 +15715,7 @@ 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 ffffffffe9a3803a <_sk_callback_sse41+0xffffffffe9a345a9>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a389ca <_sk_callback_sse41+0xffffffffe9a346af>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14306,16 +15770,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 40d4 <.literal16+0x594>
+ DB 127,0 ; jg 4a64 <.literal16+0x694>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 40d8 <.literal16+0x598>
+ DB 127,0 ; jg 4a68 <.literal16+0x698>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 40dc <.literal16+0x59c>
+ DB 127,0 ; jg 4a6c <.literal16+0x69c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 40e0 <.literal16+0x5a0>
+ DB 127,0 ; jg 4a70 <.literal16+0x6a0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14324,7 +15788,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4165 <.literal16+0x625>
+ DB 119,115 ; ja 4af5 <.literal16+0x725>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14335,7 +15799,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 40c9 <.literal16+0x589>
+ DB 117,191 ; jne 4a59 <.literal16+0x689>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14347,7 +15811,7 @@ 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 ffffffffe9a3810a <_sk_callback_sse41+0xffffffffe9a34679>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a38a9a <_sk_callback_sse41+0xffffffffe9a3477f>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14402,16 +15866,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 41a4 <.literal16+0x664>
+ DB 127,0 ; jg 4b34 <.literal16+0x764>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 41a8 <.literal16+0x668>
+ DB 127,0 ; jg 4b38 <.literal16+0x768>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 41ac <.literal16+0x66c>
+ DB 127,0 ; jg 4b3c <.literal16+0x76c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 41b0 <.literal16+0x670>
+ DB 127,0 ; jg 4b40 <.literal16+0x770>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14420,7 +15884,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4235 <.literal16+0x6f5>
+ DB 119,115 ; ja 4bc5 <.literal16+0x7f5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14431,7 +15895,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4199 <.literal16+0x659>
+ DB 117,191 ; jne 4b29 <.literal16+0x759>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14443,7 +15907,7 @@ 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 ffffffffe9a381da <_sk_callback_sse41+0xffffffffe9a34749>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a38b6a <_sk_callback_sse41+0xffffffffe9a3484f>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14498,16 +15962,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4274 <.literal16+0x734>
+ DB 127,0 ; jg 4c04 <.literal16+0x834>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4278 <.literal16+0x738>
+ DB 127,0 ; jg 4c08 <.literal16+0x838>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 427c <.literal16+0x73c>
+ DB 127,0 ; jg 4c0c <.literal16+0x83c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4280 <.literal16+0x740>
+ DB 127,0 ; jg 4c10 <.literal16+0x840>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14516,7 +15980,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4305 <.literal16+0x7c5>
+ DB 119,115 ; ja 4c95 <.literal16+0x8c5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14527,7 +15991,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4269 <.literal16+0x729>
+ DB 117,191 ; jne 4bf9 <.literal16+0x829>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14539,7 +16003,7 @@ 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 ffffffffe9a382aa <_sk_callback_sse41+0xffffffffe9a34819>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a38c3a <_sk_callback_sse41+0xffffffffe9a3491f>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14590,13 +16054,13 @@ 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 4387 <.literal16+0x847>
+ DB 127,67 ; jg 4d17 <.literal16+0x947>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 438b <.literal16+0x84b>
+ DB 127,67 ; jg 4d1b <.literal16+0x94b>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 438f <.literal16+0x84f>
+ DB 127,67 ; jg 4d1f <.literal16+0x94f>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4393 <.literal16+0x853>
+ DB 127,67 ; jg 4d23 <.literal16+0x953>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -14643,16 +16107,16 @@ ALIGN 16
DB 128,3,62 ; addb $0x3e,(%rbx)
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4413 <.literal16+0x8d3>
+ DB 118,63 ; jbe 4da3 <.literal16+0x9d3>
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4417 <.literal16+0x8d7>
+ DB 118,63 ; jbe 4da7 <.literal16+0x9d7>
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 441b <.literal16+0x8db>
+ DB 118,63 ; jbe 4dab <.literal16+0x9db>
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 441f <.literal16+0x8df>
+ DB 118,63 ; jbe 4daf <.literal16+0x9df>
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)
@@ -14664,11 +16128,11 @@ 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 445b <.literal16+0x91b>
+ DB 127,67 ; jg 4deb <.literal16+0xa1b>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 445f <.literal16+0x91f>
+ DB 127,67 ; jg 4def <.literal16+0xa1f>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4463 <.literal16+0x923>
+ DB 127,67 ; jg 4df3 <.literal16+0xa23>
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)
@@ -14697,7 +16161,7 @@ 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) # 3004490 <_sk_callback_sse41+0x30009ff>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004e20 <_sk_callback_sse41+0x3000b05>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14726,13 +16190,13 @@ ALIGN 16
DB 132,55 ; test %dh,(%rdi)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 44c9 <.literal16+0x989>
+ DB 224,7 ; loopne 4e59 <.literal16+0xa89>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 44cd <.literal16+0x98d>
+ DB 224,7 ; loopne 4e5d <.literal16+0xa8d>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 44d1 <.literal16+0x991>
+ DB 224,7 ; loopne 4e61 <.literal16+0xa91>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 44d5 <.literal16+0x995>
+ DB 224,7 ; loopne 4e65 <.literal16+0xa95>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -14778,13 +16242,13 @@ ALIGN 16
DB 132,55 ; test %dh,(%rdi)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4539 <.literal16+0x9f9>
+ DB 224,7 ; loopne 4ec9 <.literal16+0xaf9>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 453d <.literal16+0x9fd>
+ DB 224,7 ; loopne 4ecd <.literal16+0xafd>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 4541 <.literal16+0xa01>
+ DB 224,7 ; loopne 4ed1 <.literal16+0xb01>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 4545 <.literal16+0xa05>
+ DB 224,7 ; loopne 4ed5 <.literal16+0xb05>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -14822,13 +16286,13 @@ ALIGN 16
DB 65,0,0 ; add %al,(%r8)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 45d6 <.literal16+0xa96>
+ DB 124,66 ; jl 4f66 <.literal16+0xb96>
DB 0,0 ; add %al,(%rax)
- DB 124,66 ; jl 45da <.literal16+0xa9a>
+ DB 124,66 ; jl 4f6a <.literal16+0xb9a>
DB 0,0 ; add %al,(%rax)
- DB 124,66 ; jl 45de <.literal16+0xa9e>
+ DB 124,66 ; jl 4f6e <.literal16+0xb9e>
DB 0,0 ; add %al,(%rax)
- DB 124,66 ; jl 45e2 <.literal16+0xaa2>
+ DB 124,66 ; jl 4f72 <.literal16+0xba2>
DB 0,240 ; add %dh,%al
DB 0,0 ; add %al,(%rax)
DB 0,240 ; add %dh,%al
@@ -14918,13 +16382,13 @@ 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 46e5 <.literal16+0xba5>
+ DB 112,65 ; jo 5075 <.literal16+0xca5>
DB 0,0 ; add %al,(%rax)
- DB 112,65 ; jo 46e9 <.literal16+0xba9>
+ DB 112,65 ; jo 5079 <.literal16+0xca9>
DB 0,0 ; add %al,(%rax)
- DB 112,65 ; jo 46ed <.literal16+0xbad>
+ DB 112,65 ; jo 507d <.literal16+0xcad>
DB 0,0 ; add %al,(%rax)
- DB 112,65 ; jo 46f1 <.literal16+0xbb1>
+ DB 112,65 ; jo 5081 <.literal16+0xcb1>
DB 255,0 ; incl (%rax)
DB 0,0 ; add %al,(%rax)
DB 255,0 ; incl (%rax)
@@ -14939,7 +16403,7 @@ 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) # 30046e0 <_sk_callback_sse41+0x3000c4f>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005070 <_sk_callback_sse41+0x3000d55>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14966,7 +16430,7 @@ 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) # 3004720 <_sk_callback_sse41+0x3000c8f>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30050b0 <_sk_callback_sse41+0x3000d95>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14981,11 +16445,11 @@ 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 477b <.literal16+0xc3b>
+ DB 127,67 ; jg 510b <.literal16+0xd3b>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 477f <.literal16+0xc3f>
+ DB 127,67 ; jg 510f <.literal16+0xd3f>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4783 <.literal16+0xc43>
+ DB 127,67 ; jg 5113 <.literal16+0xd43>
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)
@@ -15061,13 +16525,13 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 255 ; (bad)
- DB 127,71 ; jg 484b <.literal16+0xd0b>
+ DB 127,71 ; jg 51db <.literal16+0xe0b>
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 484f <.literal16+0xd0f>
+ DB 127,71 ; jg 51df <.literal16+0xe0f>
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 4853 <.literal16+0xd13>
+ DB 127,71 ; jg 51e3 <.literal16+0xe13>
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 4857 <.literal16+0xd17>
+ DB 127,71 ; jg 51e7 <.literal16+0xe17>
DB 208 ; (bad)
DB 179,89 ; mov $0x59,%bl
DB 62,208 ; ds (bad)
@@ -15193,11 +16657,11 @@ 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 4962 <.literal16+0xe22>
+ DB 62,114,28 ; jb,pt 52f2 <.literal16+0xf22>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4966 <.literal16+0xe26>
+ DB 62,114,28 ; jb,pt 52f6 <.literal16+0xf26>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 496a <.literal16+0xe2a>
+ DB 62,114,28 ; jb,pt 52fa <.literal16+0xf2a>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -15241,7 +16705,7 @@ 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 # 3d63d7f5 <_sk_callback_sse41+0x3d639d64>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e185 <_sk_callback_sse41+0x3d639e6a>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -15267,7 +16731,7 @@ 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 # 3d63d835 <_sk_callback_sse41+0x3d639da4>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e1c5 <_sk_callback_sse41+0x3d639eaa>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
@@ -15276,13 +16740,13 @@ ALIGN 16
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
DB 63 ; (bad)
- DB 114,28 ; jb 4a2e <.literal16+0xeee>
+ DB 114,28 ; jb 53be <.literal16+0xfee>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4a32 <.literal16+0xef2>
+ DB 62,114,28 ; jb,pt 53c2 <.literal16+0xff2>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4a36 <.literal16+0xef6>
+ DB 62,114,28 ; jb,pt 53c6 <.literal16+0xff6>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4a3a <.literal16+0xefa>
+ DB 62,114,28 ; jb,pt 53ca <.literal16+0xffa>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -15303,11 +16767,11 @@ 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 4a72 <.literal16+0xf32>
+ DB 62,114,28 ; jb,pt 5402 <.literal16+0x1032>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4a76 <.literal16+0xf36>
+ DB 62,114,28 ; jb,pt 5406 <.literal16+0x1036>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4a7a <.literal16+0xf3a>
+ DB 62,114,28 ; jb,pt 540a <.literal16+0x103a>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -15351,7 +16815,7 @@ 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 # 3d63d905 <_sk_callback_sse41+0x3d639e74>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e295 <_sk_callback_sse41+0x3d639f7a>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -15377,7 +16841,7 @@ 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 # 3d63d945 <_sk_callback_sse41+0x3d639eb4>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e2d5 <_sk_callback_sse41+0x3d639fba>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
@@ -15386,13 +16850,13 @@ ALIGN 16
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
DB 63 ; (bad)
- DB 114,28 ; jb 4b3e <.literal16+0xffe>
+ DB 114,28 ; jb 54ce <.literal16+0x10fe>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4b42 <_sk_callback_sse41+0x10b1>
+ DB 62,114,28 ; jb,pt 54d2 <_sk_callback_sse41+0x11b7>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4b46 <_sk_callback_sse41+0x10b5>
+ DB 62,114,28 ; jb,pt 54d6 <_sk_callback_sse41+0x11bb>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4b4a <_sk_callback_sse41+0x10b9>
+ DB 62,114,28 ; jb,pt 54da <_sk_callback_sse41+0x11bf>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -15483,7 +16947,7 @@ _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 15,40,21,161,62,0,0 ; movaps 0x3ea1(%rip),%xmm2 # 3fb0 <_sk_callback_sse2+0xaf>
+ DB 15,40,21,145,71,0,0 ; movaps 0x4791(%rip),%xmm2 # 48a0 <_sk_callback_sse2+0xb7>
DB 15,88,202 ; addps %xmm2,%xmm1
DB 15,16,2 ; movups (%rdx),%xmm0
DB 15,88,193 ; addps %xmm1,%xmm0
@@ -15492,7 +16956,7 @@ _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 15,40,21,144,62,0,0 ; movaps 0x3e90(%rip),%xmm2 # 3fc0 <_sk_callback_sse2+0xbf>
+ DB 15,40,21,128,71,0,0 ; movaps 0x4780(%rip),%xmm2 # 48b0 <_sk_callback_sse2+0xc7>
DB 15,87,219 ; xorps %xmm3,%xmm3
DB 15,87,228 ; xorps %xmm4,%xmm4
DB 15,87,237 ; xorps %xmm5,%xmm5
@@ -15513,14 +16977,14 @@ _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 102,68,15,111,21,85,62,0,0 ; movdqa 0x3e55(%rip),%xmm10 # 3fd0 <_sk_callback_sse2+0xcf>
+ DB 102,68,15,111,21,69,71,0,0 ; movdqa 0x4745(%rip),%xmm10 # 48c0 <_sk_callback_sse2+0xd7>
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 102,68,15,111,37,65,62,0,0 ; movdqa 0x3e41(%rip),%xmm12 # 3fe0 <_sk_callback_sse2+0xdf>
- DB 102,68,15,111,45,72,62,0,0 ; movdqa 0x3e48(%rip),%xmm13 # 3ff0 <_sk_callback_sse2+0xef>
+ DB 102,68,15,111,37,49,71,0,0 ; movdqa 0x4731(%rip),%xmm12 # 48d0 <_sk_callback_sse2+0xe7>
+ DB 102,68,15,111,45,56,71,0,0 ; movdqa 0x4738(%rip),%xmm13 # 48e0 <_sk_callback_sse2+0xf7>
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
@@ -15536,8 +17000,8 @@ _sk_dither_sse2 LABEL PROC
DB 102,69,15,235,245 ; por %xmm13,%xmm14
DB 102,69,15,235,240 ; por %xmm8,%xmm14
DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8
- DB 68,15,89,5,3,62,0,0 ; mulps 0x3e03(%rip),%xmm8 # 4000 <_sk_callback_sse2+0xff>
- DB 68,15,88,5,11,62,0,0 ; addps 0x3e0b(%rip),%xmm8 # 4010 <_sk_callback_sse2+0x10f>
+ DB 68,15,89,5,243,70,0,0 ; mulps 0x46f3(%rip),%xmm8 # 48f0 <_sk_callback_sse2+0x107>
+ DB 68,15,88,5,251,70,0,0 ; addps 0x46fb(%rip),%xmm8 # 4900 <_sk_callback_sse2+0x117>
DB 243,68,15,16,72,8 ; movss 0x8(%rax),%xmm9
DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9
DB 69,15,89,200 ; mulps %xmm8,%xmm9
@@ -15573,7 +17037,7 @@ _sk_clear_sse2 LABEL PROC
PUBLIC _sk_srcatop_sse2
_sk_srcatop_sse2 LABEL PROC
DB 15,89,199 ; mulps %xmm7,%xmm0
- DB 68,15,40,5,184,61,0,0 ; movaps 0x3db8(%rip),%xmm8 # 4020 <_sk_callback_sse2+0x11f>
+ DB 68,15,40,5,168,70,0,0 ; movaps 0x46a8(%rip),%xmm8 # 4910 <_sk_callback_sse2+0x127>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,89,204 ; mulps %xmm4,%xmm9
@@ -15596,7 +17060,7 @@ 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 68,15,40,13,123,61,0,0 ; movaps 0x3d7b(%rip),%xmm9 # 4030 <_sk_callback_sse2+0x12f>
+ DB 68,15,40,13,107,70,0,0 ; movaps 0x466b(%rip),%xmm9 # 4920 <_sk_callback_sse2+0x137>
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,88,192 ; addps %xmm8,%xmm0
@@ -15637,7 +17101,7 @@ _sk_dstin_sse2 LABEL PROC
PUBLIC _sk_srcout_sse2
_sk_srcout_sse2 LABEL PROC
- DB 68,15,40,5,31,61,0,0 ; movaps 0x3d1f(%rip),%xmm8 # 4040 <_sk_callback_sse2+0x13f>
+ DB 68,15,40,5,15,70,0,0 ; movaps 0x460f(%rip),%xmm8 # 4930 <_sk_callback_sse2+0x147>
DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
@@ -15648,7 +17112,7 @@ _sk_srcout_sse2 LABEL PROC
PUBLIC _sk_dstout_sse2
_sk_dstout_sse2 LABEL PROC
- DB 68,15,40,5,15,61,0,0 ; movaps 0x3d0f(%rip),%xmm8 # 4050 <_sk_callback_sse2+0x14f>
+ DB 68,15,40,5,255,69,0,0 ; movaps 0x45ff(%rip),%xmm8 # 4940 <_sk_callback_sse2+0x157>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 15,89,196 ; mulps %xmm4,%xmm0
@@ -15663,7 +17127,7 @@ _sk_dstout_sse2 LABEL PROC
PUBLIC _sk_srcover_sse2
_sk_srcover_sse2 LABEL PROC
- DB 68,15,40,5,242,60,0,0 ; movaps 0x3cf2(%rip),%xmm8 # 4060 <_sk_callback_sse2+0x15f>
+ DB 68,15,40,5,226,69,0,0 ; movaps 0x45e2(%rip),%xmm8 # 4950 <_sk_callback_sse2+0x167>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,89,204 ; mulps %xmm4,%xmm9
@@ -15681,7 +17145,7 @@ _sk_srcover_sse2 LABEL PROC
PUBLIC _sk_dstover_sse2
_sk_dstover_sse2 LABEL PROC
- DB 68,15,40,5,198,60,0,0 ; movaps 0x3cc6(%rip),%xmm8 # 4070 <_sk_callback_sse2+0x16f>
+ DB 68,15,40,5,182,69,0,0 ; movaps 0x45b6(%rip),%xmm8 # 4960 <_sk_callback_sse2+0x177>
DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -15705,7 +17169,7 @@ _sk_modulate_sse2 LABEL PROC
PUBLIC _sk_multiply_sse2
_sk_multiply_sse2 LABEL PROC
- DB 68,15,40,5,154,60,0,0 ; movaps 0x3c9a(%rip),%xmm8 # 4080 <_sk_callback_sse2+0x17f>
+ DB 68,15,40,5,138,69,0,0 ; movaps 0x458a(%rip),%xmm8 # 4970 <_sk_callback_sse2+0x187>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 69,15,40,209 ; movaps %xmm9,%xmm10
@@ -15775,7 +17239,7 @@ _sk_screen_sse2 LABEL PROC
PUBLIC _sk_xor__sse2
_sk_xor__sse2 LABEL PROC
DB 68,15,40,195 ; movaps %xmm3,%xmm8
- DB 15,40,29,203,59,0,0 ; movaps 0x3bcb(%rip),%xmm3 # 4090 <_sk_callback_sse2+0x18f>
+ DB 15,40,29,187,68,0,0 ; movaps 0x44bb(%rip),%xmm3 # 4980 <_sk_callback_sse2+0x197>
DB 68,15,40,203 ; movaps %xmm3,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 65,15,89,193 ; mulps %xmm9,%xmm0
@@ -15821,7 +17285,7 @@ _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 15,40,21,54,59,0,0 ; movaps 0x3b36(%rip),%xmm2 # 40a0 <_sk_callback_sse2+0x19f>
+ DB 15,40,21,38,68,0,0 ; movaps 0x4426(%rip),%xmm2 # 4990 <_sk_callback_sse2+0x1a7>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -15853,7 +17317,7 @@ _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 15,40,21,219,58,0,0 ; movaps 0x3adb(%rip),%xmm2 # 40b0 <_sk_callback_sse2+0x1af>
+ DB 15,40,21,203,67,0,0 ; movaps 0x43cb(%rip),%xmm2 # 49a0 <_sk_callback_sse2+0x1b7>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -15888,7 +17352,7 @@ _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 15,40,21,117,58,0,0 ; movaps 0x3a75(%rip),%xmm2 # 40c0 <_sk_callback_sse2+0x1bf>
+ DB 15,40,21,101,67,0,0 ; movaps 0x4365(%rip),%xmm2 # 49b0 <_sk_callback_sse2+0x1c7>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -15913,7 +17377,7 @@ _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 15,40,13,54,58,0,0 ; movaps 0x3a36(%rip),%xmm1 # 40d0 <_sk_callback_sse2+0x1cf>
+ DB 15,40,13,38,67,0,0 ; movaps 0x4326(%rip),%xmm1 # 49c0 <_sk_callback_sse2+0x1d7>
DB 15,92,203 ; subps %xmm3,%xmm1
DB 15,89,207 ; mulps %xmm7,%xmm1
DB 15,88,217 ; addps %xmm1,%xmm3
@@ -15925,7 +17389,7 @@ _sk_exclusion_sse2 LABEL PROC
PUBLIC _sk_colorburn_sse2
_sk_colorburn_sse2 LABEL PROC
DB 68,15,40,192 ; movaps %xmm0,%xmm8
- DB 68,15,40,21,37,58,0,0 ; movaps 0x3a25(%rip),%xmm10 # 40e0 <_sk_callback_sse2+0x1df>
+ DB 68,15,40,21,21,67,0,0 ; movaps 0x4315(%rip),%xmm10 # 49d0 <_sk_callback_sse2+0x1e7>
DB 69,15,40,202 ; movaps %xmm10,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 69,15,40,217 ; movaps %xmm9,%xmm11
@@ -16017,7 +17481,7 @@ _sk_colorburn_sse2 LABEL PROC
PUBLIC _sk_colordodge_sse2
_sk_colordodge_sse2 LABEL PROC
DB 68,15,40,200 ; movaps %xmm0,%xmm9
- DB 68,15,40,21,219,56,0,0 ; movaps 0x38db(%rip),%xmm10 # 40f0 <_sk_callback_sse2+0x1ef>
+ DB 68,15,40,21,203,65,0,0 ; movaps 0x41cb(%rip),%xmm10 # 49e0 <_sk_callback_sse2+0x1f7>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 69,15,40,227 ; movaps %xmm11,%xmm12
@@ -16110,7 +17574,7 @@ _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 68,15,40,29,141,55,0,0 ; movaps 0x378d(%rip),%xmm11 # 4100 <_sk_callback_sse2+0x1ff>
+ DB 68,15,40,29,125,64,0,0 ; movaps 0x407d(%rip),%xmm11 # 49f0 <_sk_callback_sse2+0x207>
DB 69,15,40,211 ; movaps %xmm11,%xmm10
DB 68,15,92,215 ; subps %xmm7,%xmm10
DB 69,15,40,194 ; movaps %xmm10,%xmm8
@@ -16197,7 +17661,7 @@ 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 68,15,40,13,88,54,0,0 ; movaps 0x3658(%rip),%xmm9 # 4110 <_sk_callback_sse2+0x20f>
+ DB 68,15,40,13,72,63,0,0 ; movaps 0x3f48(%rip),%xmm9 # 4a00 <_sk_callback_sse2+0x217>
DB 69,15,40,209 ; movaps %xmm9,%xmm10
DB 68,15,92,215 ; subps %xmm7,%xmm10
DB 69,15,40,218 ; movaps %xmm10,%xmm11
@@ -16287,7 +17751,7 @@ _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 68,15,40,13,18,53,0,0 ; movaps 0x3512(%rip),%xmm9 # 4120 <_sk_callback_sse2+0x21f>
+ DB 68,15,40,13,2,62,0,0 ; movaps 0x3e02(%rip),%xmm9 # 4a10 <_sk_callback_sse2+0x227>
DB 69,15,40,249 ; movaps %xmm9,%xmm15
DB 69,15,92,250 ; subps %xmm10,%xmm15
DB 69,15,40,218 ; movaps %xmm10,%xmm11
@@ -16300,10 +17764,10 @@ _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 68,15,40,53,236,52,0,0 ; movaps 0x34ec(%rip),%xmm14 # 4130 <_sk_callback_sse2+0x22f>
+ DB 68,15,40,53,220,61,0,0 ; movaps 0x3ddc(%rip),%xmm14 # 4a20 <_sk_callback_sse2+0x237>
DB 69,15,88,222 ; addps %xmm14,%xmm11
DB 68,15,89,216 ; mulps %xmm0,%xmm11
- DB 68,15,40,21,236,52,0,0 ; movaps 0x34ec(%rip),%xmm10 # 4140 <_sk_callback_sse2+0x23f>
+ DB 68,15,40,21,220,61,0,0 ; movaps 0x3ddc(%rip),%xmm10 # 4a30 <_sk_callback_sse2+0x247>
DB 69,15,89,234 ; mulps %xmm10,%xmm13
DB 69,15,88,235 ; addps %xmm11,%xmm13
DB 15,88,228 ; addps %xmm4,%xmm4
@@ -16440,6 +17904,594 @@ _sk_softlight_sse2 LABEL PROC
DB 72,131,196,40 ; add $0x28,%rsp
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_hue_sse2
+_sk_hue_sse2 LABEL PROC
+ DB 72,131,236,24 ; sub $0x18,%rsp
+ DB 15,41,60,36 ; movaps %xmm7,(%rsp)
+ DB 15,40,254 ; movaps %xmm6,%xmm7
+ DB 15,40,245 ; movaps %xmm5,%xmm6
+ DB 15,40,236 ; movaps %xmm4,%xmm5
+ DB 15,40,227 ; movaps %xmm3,%xmm4
+ DB 69,15,87,192 ; xorps %xmm8,%xmm8
+ DB 68,15,40,13,232,59,0,0 ; movaps 0x3be8(%rip),%xmm9 # 4a40 <_sk_callback_sse2+0x257>
+ DB 69,15,40,209 ; movaps %xmm9,%xmm10
+ DB 68,15,94,212 ; divps %xmm4,%xmm10
+ DB 68,15,40,228 ; movaps %xmm4,%xmm12
+ DB 69,15,194,224,4 ; cmpneqps %xmm8,%xmm12
+ DB 69,15,84,226 ; andps %xmm10,%xmm12
+ DB 65,15,89,196 ; mulps %xmm12,%xmm0
+ DB 65,15,89,204 ; mulps %xmm12,%xmm1
+ DB 68,15,89,226 ; mulps %xmm2,%xmm12
+ DB 68,15,40,214 ; movaps %xmm6,%xmm10
+ DB 68,15,40,238 ; movaps %xmm6,%xmm13
+ DB 68,15,40,254 ; movaps %xmm6,%xmm15
+ DB 68,15,95,255 ; maxps %xmm7,%xmm15
+ DB 68,15,40,221 ; movaps %xmm5,%xmm11
+ DB 68,15,40,245 ; movaps %xmm5,%xmm14
+ DB 15,40,213 ; movaps %xmm5,%xmm2
+ DB 65,15,95,215 ; maxps %xmm15,%xmm2
+ DB 68,15,93,215 ; minps %xmm7,%xmm10
+ DB 69,15,93,218 ; minps %xmm10,%xmm11
+ DB 65,15,92,211 ; subps %xmm11,%xmm2
+ DB 68,15,40,209 ; movaps %xmm1,%xmm10
+ DB 69,15,93,212 ; minps %xmm12,%xmm10
+ DB 68,15,40,216 ; movaps %xmm0,%xmm11
+ DB 69,15,93,218 ; minps %xmm10,%xmm11
+ DB 68,15,40,209 ; movaps %xmm1,%xmm10
+ DB 69,15,95,212 ; maxps %xmm12,%xmm10
+ DB 68,15,40,248 ; movaps %xmm0,%xmm15
+ DB 69,15,95,250 ; maxps %xmm10,%xmm15
+ DB 69,15,92,251 ; subps %xmm11,%xmm15
+ DB 65,15,92,195 ; subps %xmm11,%xmm0
+ DB 15,89,194 ; mulps %xmm2,%xmm0
+ DB 65,15,94,199 ; divps %xmm15,%xmm0
+ DB 65,15,92,203 ; subps %xmm11,%xmm1
+ DB 15,89,202 ; mulps %xmm2,%xmm1
+ DB 65,15,94,207 ; divps %xmm15,%xmm1
+ DB 69,15,92,227 ; subps %xmm11,%xmm12
+ DB 68,15,89,226 ; mulps %xmm2,%xmm12
+ DB 69,15,94,231 ; divps %xmm15,%xmm12
+ DB 69,15,194,248,4 ; cmpneqps %xmm8,%xmm15
+ DB 65,15,84,199 ; andps %xmm15,%xmm0
+ DB 65,15,84,207 ; andps %xmm15,%xmm1
+ DB 69,15,84,231 ; andps %xmm15,%xmm12
+ DB 68,15,40,61,77,59,0,0 ; movaps 0x3b4d(%rip),%xmm15 # 4a50 <_sk_callback_sse2+0x267>
+ DB 69,15,89,247 ; mulps %xmm15,%xmm14
+ DB 15,40,29,82,59,0,0 ; movaps 0x3b52(%rip),%xmm3 # 4a60 <_sk_callback_sse2+0x277>
+ DB 68,15,89,235 ; mulps %xmm3,%xmm13
+ DB 69,15,88,238 ; addps %xmm14,%xmm13
+ DB 68,15,40,21,82,59,0,0 ; movaps 0x3b52(%rip),%xmm10 # 4a70 <_sk_callback_sse2+0x287>
+ DB 68,15,40,223 ; movaps %xmm7,%xmm11
+ DB 69,15,89,218 ; mulps %xmm10,%xmm11
+ DB 69,15,88,221 ; addps %xmm13,%xmm11
+ DB 68,15,40,232 ; movaps %xmm0,%xmm13
+ DB 69,15,89,239 ; mulps %xmm15,%xmm13
+ DB 68,15,40,241 ; movaps %xmm1,%xmm14
+ DB 68,15,89,243 ; mulps %xmm3,%xmm14
+ DB 69,15,88,245 ; addps %xmm13,%xmm14
+ DB 65,15,40,212 ; movaps %xmm12,%xmm2
+ DB 65,15,89,210 ; mulps %xmm10,%xmm2
+ DB 65,15,88,214 ; addps %xmm14,%xmm2
+ DB 68,15,92,218 ; subps %xmm2,%xmm11
+ DB 65,15,88,195 ; addps %xmm11,%xmm0
+ DB 65,15,88,203 ; addps %xmm11,%xmm1
+ DB 69,15,88,220 ; addps %xmm12,%xmm11
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 65,15,93,211 ; minps %xmm11,%xmm2
+ DB 68,15,40,232 ; movaps %xmm0,%xmm13
+ DB 68,15,93,234 ; minps %xmm2,%xmm13
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 65,15,95,211 ; maxps %xmm11,%xmm2
+ DB 68,15,40,224 ; movaps %xmm0,%xmm12
+ DB 68,15,95,226 ; maxps %xmm2,%xmm12
+ DB 68,15,89,248 ; mulps %xmm0,%xmm15
+ DB 15,89,217 ; mulps %xmm1,%xmm3
+ DB 65,15,88,223 ; addps %xmm15,%xmm3
+ DB 69,15,89,211 ; mulps %xmm11,%xmm10
+ DB 68,15,88,211 ; addps %xmm3,%xmm10
+ DB 15,87,210 ; xorps %xmm2,%xmm2
+ DB 65,15,194,213,2 ; cmpleps %xmm13,%xmm2
+ DB 15,40,216 ; movaps %xmm0,%xmm3
+ DB 65,15,92,218 ; subps %xmm10,%xmm3
+ DB 65,15,89,218 ; mulps %xmm10,%xmm3
+ DB 69,15,40,242 ; movaps %xmm10,%xmm14
+ DB 69,15,92,245 ; subps %xmm13,%xmm14
+ DB 65,15,94,222 ; divps %xmm14,%xmm3
+ DB 65,15,88,218 ; addps %xmm10,%xmm3
+ DB 68,15,40,234 ; movaps %xmm2,%xmm13
+ DB 68,15,85,235 ; andnps %xmm3,%xmm13
+ DB 15,84,194 ; andps %xmm2,%xmm0
+ DB 65,15,86,197 ; orps %xmm13,%xmm0
+ DB 69,15,40,233 ; movaps %xmm9,%xmm13
+ DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13
+ DB 65,15,40,221 ; movaps %xmm13,%xmm3
+ DB 15,85,216 ; andnps %xmm0,%xmm3
+ DB 65,15,92,194 ; subps %xmm10,%xmm0
+ DB 69,15,92,202 ; subps %xmm10,%xmm9
+ DB 65,15,89,193 ; mulps %xmm9,%xmm0
+ DB 69,15,92,226 ; subps %xmm10,%xmm12
+ DB 65,15,94,196 ; divps %xmm12,%xmm0
+ DB 65,15,88,194 ; addps %xmm10,%xmm0
+ DB 65,15,84,197 ; andps %xmm13,%xmm0
+ DB 15,86,195 ; orps %xmm3,%xmm0
+ DB 15,40,217 ; movaps %xmm1,%xmm3
+ DB 65,15,92,218 ; subps %xmm10,%xmm3
+ DB 65,15,89,218 ; mulps %xmm10,%xmm3
+ DB 65,15,94,222 ; divps %xmm14,%xmm3
+ DB 65,15,88,218 ; addps %xmm10,%xmm3
+ DB 68,15,40,250 ; movaps %xmm2,%xmm15
+ DB 68,15,85,251 ; andnps %xmm3,%xmm15
+ DB 15,84,202 ; andps %xmm2,%xmm1
+ DB 65,15,86,207 ; orps %xmm15,%xmm1
+ DB 65,15,40,221 ; movaps %xmm13,%xmm3
+ DB 15,85,217 ; andnps %xmm1,%xmm3
+ DB 65,15,92,202 ; subps %xmm10,%xmm1
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 65,15,94,204 ; divps %xmm12,%xmm1
+ DB 65,15,88,202 ; addps %xmm10,%xmm1
+ DB 65,15,84,205 ; andps %xmm13,%xmm1
+ DB 15,86,203 ; orps %xmm3,%xmm1
+ DB 65,15,40,219 ; movaps %xmm11,%xmm3
+ DB 65,15,92,218 ; subps %xmm10,%xmm3
+ DB 65,15,89,218 ; mulps %xmm10,%xmm3
+ DB 65,15,94,222 ; divps %xmm14,%xmm3
+ DB 65,15,88,218 ; addps %xmm10,%xmm3
+ DB 68,15,84,218 ; andps %xmm2,%xmm11
+ DB 15,85,211 ; andnps %xmm3,%xmm2
+ DB 65,15,86,211 ; orps %xmm11,%xmm2
+ DB 65,15,40,221 ; movaps %xmm13,%xmm3
+ DB 15,85,218 ; andnps %xmm2,%xmm3
+ DB 65,15,92,210 ; subps %xmm10,%xmm2
+ DB 65,15,89,209 ; mulps %xmm9,%xmm2
+ DB 65,15,94,212 ; divps %xmm12,%xmm2
+ DB 65,15,88,210 ; addps %xmm10,%xmm2
+ DB 65,15,84,213 ; andps %xmm13,%xmm2
+ DB 15,86,211 ; orps %xmm3,%xmm2
+ DB 65,15,95,192 ; maxps %xmm8,%xmm0
+ DB 65,15,95,200 ; maxps %xmm8,%xmm1
+ DB 65,15,95,208 ; maxps %xmm8,%xmm2
+ DB 68,15,40,196 ; movaps %xmm4,%xmm8
+ DB 68,15,40,12,36 ; movaps (%rsp),%xmm9
+ DB 69,15,88,193 ; addps %xmm9,%xmm8
+ DB 65,15,89,225 ; mulps %xmm9,%xmm4
+ DB 68,15,92,196 ; subps %xmm4,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,229 ; movaps %xmm5,%xmm4
+ DB 15,40,238 ; movaps %xmm6,%xmm5
+ DB 15,40,247 ; movaps %xmm7,%xmm6
+ DB 65,15,40,216 ; movaps %xmm8,%xmm3
+ DB 65,15,40,249 ; movaps %xmm9,%xmm7
+ DB 72,131,196,24 ; add $0x18,%rsp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_saturation_sse2
+_sk_saturation_sse2 LABEL PROC
+ DB 68,15,40,214 ; movaps %xmm6,%xmm10
+ DB 15,40,245 ; movaps %xmm5,%xmm6
+ DB 15,40,236 ; movaps %xmm4,%xmm5
+ DB 15,40,227 ; movaps %xmm3,%xmm4
+ DB 68,15,40,193 ; movaps %xmm1,%xmm8
+ DB 68,15,40,200 ; movaps %xmm0,%xmm9
+ DB 15,87,201 ; xorps %xmm1,%xmm1
+ DB 68,15,40,29,177,57,0,0 ; movaps 0x39b1(%rip),%xmm11 # 4a80 <_sk_callback_sse2+0x297>
+ DB 65,15,40,195 ; movaps %xmm11,%xmm0
+ DB 15,94,199 ; divps %xmm7,%xmm0
+ DB 68,15,40,231 ; movaps %xmm7,%xmm12
+ DB 68,15,194,225,4 ; cmpneqps %xmm1,%xmm12
+ DB 68,15,84,224 ; andps %xmm0,%xmm12
+ DB 65,15,40,196 ; movaps %xmm12,%xmm0
+ DB 15,89,197 ; mulps %xmm5,%xmm0
+ DB 69,15,40,252 ; movaps %xmm12,%xmm15
+ DB 68,15,89,254 ; mulps %xmm6,%xmm15
+ DB 69,15,89,226 ; mulps %xmm10,%xmm12
+ DB 69,15,40,240 ; movaps %xmm8,%xmm14
+ DB 68,15,95,242 ; maxps %xmm2,%xmm14
+ DB 69,15,40,233 ; movaps %xmm9,%xmm13
+ DB 69,15,95,238 ; maxps %xmm14,%xmm13
+ DB 68,15,93,194 ; minps %xmm2,%xmm8
+ DB 69,15,93,200 ; minps %xmm8,%xmm9
+ DB 69,15,92,233 ; subps %xmm9,%xmm13
+ DB 65,15,40,215 ; movaps %xmm15,%xmm2
+ DB 65,15,93,212 ; minps %xmm12,%xmm2
+ DB 68,15,40,192 ; movaps %xmm0,%xmm8
+ DB 68,15,93,194 ; minps %xmm2,%xmm8
+ DB 69,15,40,207 ; movaps %xmm15,%xmm9
+ DB 69,15,95,204 ; maxps %xmm12,%xmm9
+ DB 15,40,208 ; movaps %xmm0,%xmm2
+ DB 65,15,95,209 ; maxps %xmm9,%xmm2
+ DB 65,15,92,208 ; subps %xmm8,%xmm2
+ DB 65,15,92,192 ; subps %xmm8,%xmm0
+ DB 65,15,89,197 ; mulps %xmm13,%xmm0
+ DB 15,94,194 ; divps %xmm2,%xmm0
+ DB 69,15,92,248 ; subps %xmm8,%xmm15
+ DB 69,15,89,253 ; mulps %xmm13,%xmm15
+ DB 68,15,94,250 ; divps %xmm2,%xmm15
+ DB 69,15,92,224 ; subps %xmm8,%xmm12
+ DB 69,15,89,229 ; mulps %xmm13,%xmm12
+ DB 68,15,94,226 ; divps %xmm2,%xmm12
+ DB 15,194,209,4 ; cmpneqps %xmm1,%xmm2
+ DB 15,84,194 ; andps %xmm2,%xmm0
+ DB 68,15,84,250 ; andps %xmm2,%xmm15
+ DB 68,15,84,226 ; andps %xmm2,%xmm12
+ DB 68,15,40,45,33,57,0,0 ; movaps 0x3921(%rip),%xmm13 # 4a90 <_sk_callback_sse2+0x2a7>
+ DB 68,15,40,197 ; movaps %xmm5,%xmm8
+ DB 69,15,89,197 ; mulps %xmm13,%xmm8
+ DB 68,15,40,53,33,57,0,0 ; movaps 0x3921(%rip),%xmm14 # 4aa0 <_sk_callback_sse2+0x2b7>
+ DB 15,40,214 ; movaps %xmm6,%xmm2
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 65,15,88,208 ; addps %xmm8,%xmm2
+ DB 68,15,40,5,30,57,0,0 ; movaps 0x391e(%rip),%xmm8 # 4ab0 <_sk_callback_sse2+0x2c7>
+ DB 69,15,40,202 ; movaps %xmm10,%xmm9
+ DB 69,15,89,200 ; mulps %xmm8,%xmm9
+ DB 68,15,88,202 ; addps %xmm2,%xmm9
+ DB 15,40,208 ; movaps %xmm0,%xmm2
+ DB 65,15,89,213 ; mulps %xmm13,%xmm2
+ DB 65,15,40,207 ; movaps %xmm15,%xmm1
+ DB 65,15,89,206 ; mulps %xmm14,%xmm1
+ DB 15,88,202 ; addps %xmm2,%xmm1
+ DB 65,15,40,212 ; movaps %xmm12,%xmm2
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 15,88,209 ; addps %xmm1,%xmm2
+ DB 68,15,92,202 ; subps %xmm2,%xmm9
+ DB 65,15,88,193 ; addps %xmm9,%xmm0
+ DB 69,15,88,249 ; addps %xmm9,%xmm15
+ DB 69,15,88,204 ; addps %xmm12,%xmm9
+ DB 65,15,40,207 ; movaps %xmm15,%xmm1
+ DB 65,15,93,201 ; minps %xmm9,%xmm1
+ DB 68,15,40,224 ; movaps %xmm0,%xmm12
+ DB 68,15,93,225 ; minps %xmm1,%xmm12
+ DB 65,15,40,215 ; movaps %xmm15,%xmm2
+ DB 65,15,95,209 ; maxps %xmm9,%xmm2
+ DB 15,40,216 ; movaps %xmm0,%xmm3
+ DB 15,95,218 ; maxps %xmm2,%xmm3
+ DB 68,15,89,232 ; mulps %xmm0,%xmm13
+ DB 69,15,89,247 ; mulps %xmm15,%xmm14
+ DB 69,15,88,245 ; addps %xmm13,%xmm14
+ DB 69,15,89,193 ; mulps %xmm9,%xmm8
+ DB 69,15,88,198 ; addps %xmm14,%xmm8
+ DB 15,87,210 ; xorps %xmm2,%xmm2
+ DB 65,15,194,212,2 ; cmpleps %xmm12,%xmm2
+ DB 68,15,40,232 ; movaps %xmm0,%xmm13
+ DB 69,15,92,232 ; subps %xmm8,%xmm13
+ DB 69,15,89,232 ; mulps %xmm8,%xmm13
+ DB 69,15,40,240 ; movaps %xmm8,%xmm14
+ DB 69,15,92,244 ; subps %xmm12,%xmm14
+ DB 69,15,94,238 ; divps %xmm14,%xmm13
+ DB 69,15,88,232 ; addps %xmm8,%xmm13
+ DB 68,15,40,226 ; movaps %xmm2,%xmm12
+ DB 69,15,85,229 ; andnps %xmm13,%xmm12
+ DB 15,84,194 ; andps %xmm2,%xmm0
+ DB 65,15,86,196 ; orps %xmm12,%xmm0
+ DB 69,15,40,235 ; movaps %xmm11,%xmm13
+ DB 68,15,194,235,1 ; cmpltps %xmm3,%xmm13
+ DB 69,15,40,229 ; movaps %xmm13,%xmm12
+ DB 68,15,85,224 ; andnps %xmm0,%xmm12
+ DB 65,15,92,192 ; subps %xmm8,%xmm0
+ DB 69,15,92,216 ; subps %xmm8,%xmm11
+ DB 65,15,89,195 ; mulps %xmm11,%xmm0
+ DB 65,15,92,216 ; subps %xmm8,%xmm3
+ DB 15,94,195 ; divps %xmm3,%xmm0
+ DB 65,15,88,192 ; addps %xmm8,%xmm0
+ DB 65,15,84,197 ; andps %xmm13,%xmm0
+ DB 65,15,86,196 ; orps %xmm12,%xmm0
+ DB 69,15,40,231 ; movaps %xmm15,%xmm12
+ DB 69,15,92,224 ; subps %xmm8,%xmm12
+ DB 69,15,89,224 ; mulps %xmm8,%xmm12
+ DB 69,15,94,230 ; divps %xmm14,%xmm12
+ DB 69,15,88,224 ; addps %xmm8,%xmm12
+ DB 15,40,202 ; movaps %xmm2,%xmm1
+ DB 65,15,85,204 ; andnps %xmm12,%xmm1
+ DB 68,15,84,250 ; andps %xmm2,%xmm15
+ DB 68,15,86,249 ; orps %xmm1,%xmm15
+ DB 65,15,40,205 ; movaps %xmm13,%xmm1
+ DB 65,15,85,207 ; andnps %xmm15,%xmm1
+ DB 69,15,92,248 ; subps %xmm8,%xmm15
+ DB 69,15,89,251 ; mulps %xmm11,%xmm15
+ DB 68,15,94,251 ; divps %xmm3,%xmm15
+ DB 69,15,88,248 ; addps %xmm8,%xmm15
+ DB 69,15,84,253 ; andps %xmm13,%xmm15
+ DB 68,15,86,249 ; orps %xmm1,%xmm15
+ DB 65,15,40,201 ; movaps %xmm9,%xmm1
+ DB 65,15,92,200 ; subps %xmm8,%xmm1
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 65,15,94,206 ; divps %xmm14,%xmm1
+ DB 65,15,88,200 ; addps %xmm8,%xmm1
+ DB 68,15,84,202 ; andps %xmm2,%xmm9
+ DB 15,85,209 ; andnps %xmm1,%xmm2
+ DB 65,15,86,209 ; orps %xmm9,%xmm2
+ DB 65,15,40,205 ; movaps %xmm13,%xmm1
+ DB 15,85,202 ; andnps %xmm2,%xmm1
+ DB 65,15,92,208 ; subps %xmm8,%xmm2
+ DB 65,15,89,211 ; mulps %xmm11,%xmm2
+ DB 15,94,211 ; divps %xmm3,%xmm2
+ DB 65,15,88,208 ; addps %xmm8,%xmm2
+ DB 65,15,84,213 ; andps %xmm13,%xmm2
+ DB 15,86,209 ; orps %xmm1,%xmm2
+ DB 15,87,201 ; xorps %xmm1,%xmm1
+ DB 15,95,193 ; maxps %xmm1,%xmm0
+ DB 68,15,95,249 ; maxps %xmm1,%xmm15
+ DB 15,95,209 ; maxps %xmm1,%xmm2
+ DB 68,15,40,196 ; movaps %xmm4,%xmm8
+ DB 68,15,88,199 ; addps %xmm7,%xmm8
+ DB 15,89,231 ; mulps %xmm7,%xmm4
+ DB 68,15,92,196 ; subps %xmm4,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 69,15,89,248 ; mulps %xmm8,%xmm15
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,229 ; movaps %xmm5,%xmm4
+ DB 15,40,238 ; movaps %xmm6,%xmm5
+ DB 65,15,40,242 ; movaps %xmm10,%xmm6
+ DB 65,15,40,207 ; movaps %xmm15,%xmm1
+ DB 65,15,40,216 ; movaps %xmm8,%xmm3
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_color_sse2
+_sk_color_sse2 LABEL PROC
+ DB 72,131,236,24 ; sub $0x18,%rsp
+ DB 68,15,40,221 ; movaps %xmm5,%xmm11
+ DB 68,15,41,28,36 ; movaps %xmm11,(%rsp)
+ DB 15,40,236 ; movaps %xmm4,%xmm5
+ DB 15,40,227 ; movaps %xmm3,%xmm4
+ DB 68,15,40,249 ; movaps %xmm1,%xmm15
+ DB 69,15,87,192 ; xorps %xmm8,%xmm8
+ DB 68,15,40,13,128,55,0,0 ; movaps 0x3780(%rip),%xmm9 # 4ac0 <_sk_callback_sse2+0x2d7>
+ DB 69,15,40,209 ; movaps %xmm9,%xmm10
+ DB 68,15,94,212 ; divps %xmm4,%xmm10
+ DB 68,15,40,228 ; movaps %xmm4,%xmm12
+ DB 69,15,194,224,4 ; cmpneqps %xmm8,%xmm12
+ DB 69,15,84,226 ; andps %xmm10,%xmm12
+ DB 65,15,89,196 ; mulps %xmm12,%xmm0
+ DB 69,15,89,252 ; mulps %xmm12,%xmm15
+ DB 68,15,89,226 ; mulps %xmm2,%xmm12
+ DB 68,15,40,45,103,55,0,0 ; movaps 0x3767(%rip),%xmm13 # 4ad0 <_sk_callback_sse2+0x2e7>
+ DB 68,15,40,213 ; movaps %xmm5,%xmm10
+ DB 69,15,89,213 ; mulps %xmm13,%xmm10
+ DB 68,15,40,53,103,55,0,0 ; movaps 0x3767(%rip),%xmm14 # 4ae0 <_sk_callback_sse2+0x2f7>
+ DB 65,15,40,211 ; movaps %xmm11,%xmm2
+ DB 65,15,89,214 ; mulps %xmm14,%xmm2
+ DB 65,15,88,210 ; addps %xmm10,%xmm2
+ DB 68,15,40,21,99,55,0,0 ; movaps 0x3763(%rip),%xmm10 # 4af0 <_sk_callback_sse2+0x307>
+ DB 68,15,40,222 ; movaps %xmm6,%xmm11
+ DB 69,15,89,218 ; mulps %xmm10,%xmm11
+ DB 68,15,88,218 ; addps %xmm2,%xmm11
+ DB 15,40,208 ; movaps %xmm0,%xmm2
+ DB 65,15,89,213 ; mulps %xmm13,%xmm2
+ DB 65,15,40,207 ; movaps %xmm15,%xmm1
+ DB 65,15,89,206 ; mulps %xmm14,%xmm1
+ DB 15,88,202 ; addps %xmm2,%xmm1
+ DB 65,15,40,212 ; movaps %xmm12,%xmm2
+ DB 65,15,89,210 ; mulps %xmm10,%xmm2
+ DB 15,88,209 ; addps %xmm1,%xmm2
+ DB 68,15,92,218 ; subps %xmm2,%xmm11
+ DB 65,15,88,195 ; addps %xmm11,%xmm0
+ DB 69,15,88,251 ; addps %xmm11,%xmm15
+ DB 69,15,88,220 ; addps %xmm12,%xmm11
+ DB 65,15,40,207 ; movaps %xmm15,%xmm1
+ DB 65,15,93,203 ; minps %xmm11,%xmm1
+ DB 68,15,40,224 ; movaps %xmm0,%xmm12
+ DB 68,15,93,225 ; minps %xmm1,%xmm12
+ DB 65,15,40,215 ; movaps %xmm15,%xmm2
+ DB 65,15,95,211 ; maxps %xmm11,%xmm2
+ DB 15,40,216 ; movaps %xmm0,%xmm3
+ DB 15,95,218 ; maxps %xmm2,%xmm3
+ DB 68,15,89,232 ; mulps %xmm0,%xmm13
+ DB 69,15,89,247 ; mulps %xmm15,%xmm14
+ DB 69,15,88,245 ; addps %xmm13,%xmm14
+ DB 69,15,89,211 ; mulps %xmm11,%xmm10
+ DB 69,15,88,214 ; addps %xmm14,%xmm10
+ DB 15,87,210 ; xorps %xmm2,%xmm2
+ DB 65,15,194,212,2 ; cmpleps %xmm12,%xmm2
+ DB 68,15,40,232 ; movaps %xmm0,%xmm13
+ DB 69,15,92,234 ; subps %xmm10,%xmm13
+ DB 69,15,89,234 ; mulps %xmm10,%xmm13
+ DB 69,15,40,242 ; movaps %xmm10,%xmm14
+ DB 69,15,92,244 ; subps %xmm12,%xmm14
+ DB 69,15,94,238 ; divps %xmm14,%xmm13
+ DB 69,15,88,234 ; addps %xmm10,%xmm13
+ DB 68,15,40,226 ; movaps %xmm2,%xmm12
+ DB 69,15,85,229 ; andnps %xmm13,%xmm12
+ DB 15,84,194 ; andps %xmm2,%xmm0
+ DB 65,15,86,196 ; orps %xmm12,%xmm0
+ DB 69,15,40,233 ; movaps %xmm9,%xmm13
+ DB 68,15,194,235,1 ; cmpltps %xmm3,%xmm13
+ DB 69,15,40,229 ; movaps %xmm13,%xmm12
+ DB 68,15,85,224 ; andnps %xmm0,%xmm12
+ DB 65,15,92,194 ; subps %xmm10,%xmm0
+ DB 69,15,92,202 ; subps %xmm10,%xmm9
+ DB 65,15,89,193 ; mulps %xmm9,%xmm0
+ DB 65,15,92,218 ; subps %xmm10,%xmm3
+ DB 15,94,195 ; divps %xmm3,%xmm0
+ DB 65,15,88,194 ; addps %xmm10,%xmm0
+ DB 65,15,84,197 ; andps %xmm13,%xmm0
+ DB 65,15,86,196 ; orps %xmm12,%xmm0
+ DB 69,15,40,231 ; movaps %xmm15,%xmm12
+ DB 69,15,92,226 ; subps %xmm10,%xmm12
+ DB 69,15,89,226 ; mulps %xmm10,%xmm12
+ DB 69,15,94,230 ; divps %xmm14,%xmm12
+ DB 69,15,88,226 ; addps %xmm10,%xmm12
+ DB 15,40,202 ; movaps %xmm2,%xmm1
+ DB 65,15,85,204 ; andnps %xmm12,%xmm1
+ DB 68,15,84,250 ; andps %xmm2,%xmm15
+ DB 68,15,86,249 ; orps %xmm1,%xmm15
+ DB 65,15,40,205 ; movaps %xmm13,%xmm1
+ DB 65,15,85,207 ; andnps %xmm15,%xmm1
+ DB 69,15,92,250 ; subps %xmm10,%xmm15
+ DB 69,15,89,249 ; mulps %xmm9,%xmm15
+ DB 68,15,94,251 ; divps %xmm3,%xmm15
+ DB 69,15,88,250 ; addps %xmm10,%xmm15
+ DB 69,15,84,253 ; andps %xmm13,%xmm15
+ DB 68,15,86,249 ; orps %xmm1,%xmm15
+ DB 65,15,40,203 ; movaps %xmm11,%xmm1
+ DB 65,15,92,202 ; subps %xmm10,%xmm1
+ DB 65,15,89,202 ; mulps %xmm10,%xmm1
+ DB 65,15,94,206 ; divps %xmm14,%xmm1
+ DB 65,15,88,202 ; addps %xmm10,%xmm1
+ DB 68,15,84,218 ; andps %xmm2,%xmm11
+ DB 15,85,209 ; andnps %xmm1,%xmm2
+ DB 65,15,86,211 ; orps %xmm11,%xmm2
+ DB 65,15,40,205 ; movaps %xmm13,%xmm1
+ DB 15,85,202 ; andnps %xmm2,%xmm1
+ DB 65,15,92,210 ; subps %xmm10,%xmm2
+ DB 65,15,89,209 ; mulps %xmm9,%xmm2
+ DB 15,94,211 ; divps %xmm3,%xmm2
+ DB 65,15,88,210 ; addps %xmm10,%xmm2
+ DB 65,15,84,213 ; andps %xmm13,%xmm2
+ DB 15,86,209 ; orps %xmm1,%xmm2
+ DB 65,15,95,192 ; maxps %xmm8,%xmm0
+ DB 69,15,95,248 ; maxps %xmm8,%xmm15
+ DB 65,15,95,208 ; maxps %xmm8,%xmm2
+ DB 68,15,40,196 ; movaps %xmm4,%xmm8
+ DB 68,15,88,199 ; addps %xmm7,%xmm8
+ DB 15,89,231 ; mulps %xmm7,%xmm4
+ DB 68,15,92,196 ; subps %xmm4,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 69,15,89,248 ; mulps %xmm8,%xmm15
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,229 ; movaps %xmm5,%xmm4
+ DB 15,40,44,36 ; movaps (%rsp),%xmm5
+ DB 65,15,40,207 ; movaps %xmm15,%xmm1
+ DB 65,15,40,216 ; movaps %xmm8,%xmm3
+ DB 72,131,196,24 ; add $0x18,%rsp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_luminosity_sse2
+_sk_luminosity_sse2 LABEL PROC
+ DB 72,131,236,24 ; sub $0x18,%rsp
+ DB 68,15,40,238 ; movaps %xmm6,%xmm13
+ DB 68,15,41,44,36 ; movaps %xmm13,(%rsp)
+ DB 15,40,245 ; movaps %xmm5,%xmm6
+ DB 15,40,236 ; movaps %xmm4,%xmm5
+ DB 15,40,227 ; movaps %xmm3,%xmm4
+ DB 15,40,218 ; movaps %xmm2,%xmm3
+ DB 68,15,40,193 ; movaps %xmm1,%xmm8
+ DB 68,15,40,200 ; movaps %xmm0,%xmm9
+ DB 69,15,87,210 ; xorps %xmm10,%xmm10
+ DB 68,15,40,29,187,53,0,0 ; movaps 0x35bb(%rip),%xmm11 # 4b00 <_sk_callback_sse2+0x317>
+ DB 65,15,40,195 ; movaps %xmm11,%xmm0
+ DB 15,94,199 ; divps %xmm7,%xmm0
+ DB 68,15,40,231 ; movaps %xmm7,%xmm12
+ DB 69,15,194,226,4 ; cmpneqps %xmm10,%xmm12
+ DB 68,15,84,224 ; andps %xmm0,%xmm12
+ DB 65,15,40,196 ; movaps %xmm12,%xmm0
+ DB 15,89,197 ; mulps %xmm5,%xmm0
+ DB 65,15,40,204 ; movaps %xmm12,%xmm1
+ DB 15,89,206 ; mulps %xmm6,%xmm1
+ DB 69,15,89,229 ; mulps %xmm13,%xmm12
+ DB 68,15,40,53,157,53,0,0 ; movaps 0x359d(%rip),%xmm14 # 4b10 <_sk_callback_sse2+0x327>
+ DB 69,15,89,206 ; mulps %xmm14,%xmm9
+ DB 68,15,40,45,161,53,0,0 ; movaps 0x35a1(%rip),%xmm13 # 4b20 <_sk_callback_sse2+0x337>
+ DB 69,15,89,197 ; mulps %xmm13,%xmm8
+ DB 69,15,88,193 ; addps %xmm9,%xmm8
+ DB 68,15,40,13,161,53,0,0 ; movaps 0x35a1(%rip),%xmm9 # 4b30 <_sk_callback_sse2+0x347>
+ DB 65,15,89,217 ; mulps %xmm9,%xmm3
+ DB 65,15,88,216 ; addps %xmm8,%xmm3
+ DB 68,15,40,192 ; movaps %xmm0,%xmm8
+ DB 69,15,89,198 ; mulps %xmm14,%xmm8
+ DB 68,15,40,249 ; movaps %xmm1,%xmm15
+ DB 69,15,89,253 ; mulps %xmm13,%xmm15
+ DB 69,15,88,248 ; addps %xmm8,%xmm15
+ DB 69,15,40,196 ; movaps %xmm12,%xmm8
+ DB 69,15,89,193 ; mulps %xmm9,%xmm8
+ DB 69,15,88,199 ; addps %xmm15,%xmm8
+ DB 65,15,92,216 ; subps %xmm8,%xmm3
+ DB 15,88,195 ; addps %xmm3,%xmm0
+ DB 15,88,203 ; addps %xmm3,%xmm1
+ DB 65,15,88,220 ; addps %xmm12,%xmm3
+ DB 68,15,40,193 ; movaps %xmm1,%xmm8
+ DB 68,15,93,195 ; minps %xmm3,%xmm8
+ DB 68,15,40,248 ; movaps %xmm0,%xmm15
+ DB 69,15,93,248 ; minps %xmm8,%xmm15
+ DB 68,15,40,193 ; movaps %xmm1,%xmm8
+ DB 68,15,95,195 ; maxps %xmm3,%xmm8
+ DB 68,15,40,224 ; movaps %xmm0,%xmm12
+ DB 69,15,95,224 ; maxps %xmm8,%xmm12
+ DB 68,15,89,240 ; mulps %xmm0,%xmm14
+ DB 68,15,89,233 ; mulps %xmm1,%xmm13
+ DB 69,15,88,238 ; addps %xmm14,%xmm13
+ DB 68,15,89,203 ; mulps %xmm3,%xmm9
+ DB 69,15,88,205 ; addps %xmm13,%xmm9
+ DB 69,15,87,192 ; xorps %xmm8,%xmm8
+ DB 69,15,194,199,2 ; cmpleps %xmm15,%xmm8
+ DB 68,15,40,232 ; movaps %xmm0,%xmm13
+ DB 69,15,92,233 ; subps %xmm9,%xmm13
+ DB 69,15,89,233 ; mulps %xmm9,%xmm13
+ DB 69,15,40,241 ; movaps %xmm9,%xmm14
+ DB 69,15,92,247 ; subps %xmm15,%xmm14
+ DB 69,15,94,238 ; divps %xmm14,%xmm13
+ DB 69,15,88,233 ; addps %xmm9,%xmm13
+ DB 69,15,40,248 ; movaps %xmm8,%xmm15
+ DB 69,15,85,253 ; andnps %xmm13,%xmm15
+ DB 65,15,84,192 ; andps %xmm8,%xmm0
+ DB 65,15,86,199 ; orps %xmm15,%xmm0
+ DB 69,15,40,235 ; movaps %xmm11,%xmm13
+ DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13
+ DB 69,15,40,253 ; movaps %xmm13,%xmm15
+ DB 68,15,85,248 ; andnps %xmm0,%xmm15
+ DB 65,15,92,193 ; subps %xmm9,%xmm0
+ DB 69,15,92,217 ; subps %xmm9,%xmm11
+ DB 65,15,89,195 ; mulps %xmm11,%xmm0
+ DB 69,15,92,225 ; subps %xmm9,%xmm12
+ DB 65,15,94,196 ; divps %xmm12,%xmm0
+ DB 65,15,88,193 ; addps %xmm9,%xmm0
+ DB 65,15,84,197 ; andps %xmm13,%xmm0
+ DB 65,15,86,199 ; orps %xmm15,%xmm0
+ DB 68,15,40,249 ; movaps %xmm1,%xmm15
+ DB 69,15,92,249 ; subps %xmm9,%xmm15
+ DB 69,15,89,249 ; mulps %xmm9,%xmm15
+ DB 69,15,94,254 ; divps %xmm14,%xmm15
+ DB 69,15,88,249 ; addps %xmm9,%xmm15
+ DB 65,15,40,208 ; movaps %xmm8,%xmm2
+ DB 65,15,85,215 ; andnps %xmm15,%xmm2
+ DB 65,15,84,200 ; andps %xmm8,%xmm1
+ DB 15,86,202 ; orps %xmm2,%xmm1
+ DB 65,15,40,213 ; movaps %xmm13,%xmm2
+ DB 15,85,209 ; andnps %xmm1,%xmm2
+ DB 65,15,92,201 ; subps %xmm9,%xmm1
+ DB 65,15,89,203 ; mulps %xmm11,%xmm1
+ DB 65,15,94,204 ; divps %xmm12,%xmm1
+ DB 65,15,88,201 ; addps %xmm9,%xmm1
+ DB 65,15,84,205 ; andps %xmm13,%xmm1
+ DB 15,86,202 ; orps %xmm2,%xmm1
+ DB 15,40,211 ; movaps %xmm3,%xmm2
+ DB 65,15,92,209 ; subps %xmm9,%xmm2
+ DB 65,15,89,209 ; mulps %xmm9,%xmm2
+ DB 65,15,94,214 ; divps %xmm14,%xmm2
+ DB 65,15,88,209 ; addps %xmm9,%xmm2
+ DB 65,15,84,216 ; andps %xmm8,%xmm3
+ DB 68,15,85,194 ; andnps %xmm2,%xmm8
+ DB 68,15,86,195 ; orps %xmm3,%xmm8
+ DB 65,15,40,213 ; movaps %xmm13,%xmm2
+ DB 65,15,85,208 ; andnps %xmm8,%xmm2
+ DB 69,15,92,193 ; subps %xmm9,%xmm8
+ DB 69,15,89,195 ; mulps %xmm11,%xmm8
+ DB 69,15,94,196 ; divps %xmm12,%xmm8
+ DB 69,15,88,193 ; addps %xmm9,%xmm8
+ DB 69,15,84,197 ; andps %xmm13,%xmm8
+ DB 68,15,86,194 ; orps %xmm2,%xmm8
+ DB 65,15,95,194 ; maxps %xmm10,%xmm0
+ DB 65,15,95,202 ; maxps %xmm10,%xmm1
+ DB 69,15,95,194 ; maxps %xmm10,%xmm8
+ DB 68,15,40,204 ; movaps %xmm4,%xmm9
+ DB 68,15,88,207 ; addps %xmm7,%xmm9
+ DB 15,89,231 ; mulps %xmm7,%xmm4
+ DB 68,15,92,204 ; subps %xmm4,%xmm9
+ DB 65,15,89,193 ; mulps %xmm9,%xmm0
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 69,15,89,193 ; mulps %xmm9,%xmm8
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 65,15,40,208 ; movaps %xmm8,%xmm2
+ DB 65,15,40,217 ; movaps %xmm9,%xmm3
+ DB 15,40,229 ; movaps %xmm5,%xmm4
+ DB 15,40,238 ; movaps %xmm6,%xmm5
+ DB 15,40,52,36 ; movaps (%rsp),%xmm6
+ DB 72,131,196,24 ; add $0x18,%rsp
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_clamp_0_sse2
_sk_clamp_0_sse2 LABEL PROC
DB 69,15,87,192 ; xorps %xmm8,%xmm8
@@ -16452,7 +18504,7 @@ _sk_clamp_0_sse2 LABEL PROC
PUBLIC _sk_clamp_1_sse2
_sk_clamp_1_sse2 LABEL PROC
- DB 68,15,40,5,248,50,0,0 ; movaps 0x32f8(%rip),%xmm8 # 4150 <_sk_callback_sse2+0x24f>
+ DB 68,15,40,5,0,52,0,0 ; movaps 0x3400(%rip),%xmm8 # 4b40 <_sk_callback_sse2+0x357>
DB 65,15,93,192 ; minps %xmm8,%xmm0
DB 65,15,93,200 ; minps %xmm8,%xmm1
DB 65,15,93,208 ; minps %xmm8,%xmm2
@@ -16462,7 +18514,7 @@ _sk_clamp_1_sse2 LABEL PROC
PUBLIC _sk_clamp_a_sse2
_sk_clamp_a_sse2 LABEL PROC
- DB 15,93,29,237,50,0,0 ; minps 0x32ed(%rip),%xmm3 # 4160 <_sk_callback_sse2+0x25f>
+ DB 15,93,29,245,51,0,0 ; minps 0x33f5(%rip),%xmm3 # 4b50 <_sk_callback_sse2+0x367>
DB 15,93,195 ; minps %xmm3,%xmm0
DB 15,93,203 ; minps %xmm3,%xmm1
DB 15,93,211 ; minps %xmm3,%xmm2
@@ -16535,7 +18587,7 @@ _sk_premul_sse2 LABEL PROC
PUBLIC _sk_unpremul_sse2
_sk_unpremul_sse2 LABEL PROC
DB 69,15,87,192 ; xorps %xmm8,%xmm8
- DB 68,15,40,13,88,50,0,0 ; movaps 0x3258(%rip),%xmm9 # 4170 <_sk_callback_sse2+0x26f>
+ DB 68,15,40,13,96,51,0,0 ; movaps 0x3360(%rip),%xmm9 # 4b60 <_sk_callback_sse2+0x377>
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
@@ -16547,20 +18599,20 @@ _sk_unpremul_sse2 LABEL PROC
PUBLIC _sk_from_srgb_sse2
_sk_from_srgb_sse2 LABEL PROC
- DB 68,15,40,5,67,50,0,0 ; movaps 0x3243(%rip),%xmm8 # 4180 <_sk_callback_sse2+0x27f>
+ DB 68,15,40,5,75,51,0,0 ; movaps 0x334b(%rip),%xmm8 # 4b70 <_sk_callback_sse2+0x387>
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 68,15,40,13,59,50,0,0 ; movaps 0x323b(%rip),%xmm9 # 4190 <_sk_callback_sse2+0x28f>
+ DB 68,15,40,13,67,51,0,0 ; movaps 0x3343(%rip),%xmm9 # 4b80 <_sk_callback_sse2+0x397>
DB 68,15,40,240 ; movaps %xmm0,%xmm14
DB 69,15,89,241 ; mulps %xmm9,%xmm14
- DB 68,15,40,21,59,50,0,0 ; movaps 0x323b(%rip),%xmm10 # 41a0 <_sk_callback_sse2+0x29f>
+ DB 68,15,40,21,67,51,0,0 ; movaps 0x3343(%rip),%xmm10 # 4b90 <_sk_callback_sse2+0x3a7>
DB 69,15,88,242 ; addps %xmm10,%xmm14
DB 69,15,89,243 ; mulps %xmm11,%xmm14
- DB 68,15,40,29,59,50,0,0 ; movaps 0x323b(%rip),%xmm11 # 41b0 <_sk_callback_sse2+0x2af>
+ DB 68,15,40,29,67,51,0,0 ; movaps 0x3343(%rip),%xmm11 # 4ba0 <_sk_callback_sse2+0x3b7>
DB 69,15,88,243 ; addps %xmm11,%xmm14
- DB 68,15,40,37,63,50,0,0 ; movaps 0x323f(%rip),%xmm12 # 41c0 <_sk_callback_sse2+0x2bf>
+ DB 68,15,40,37,71,51,0,0 ; movaps 0x3347(%rip),%xmm12 # 4bb0 <_sk_callback_sse2+0x3c7>
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
@@ -16597,20 +18649,20 @@ _sk_to_srgb_sse2 LABEL PROC
DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8
DB 69,15,83,200 ; rcpps %xmm8,%xmm9
DB 69,15,82,232 ; rsqrtps %xmm8,%xmm13
- DB 68,15,40,5,196,49,0,0 ; movaps 0x31c4(%rip),%xmm8 # 41d0 <_sk_callback_sse2+0x2cf>
+ DB 68,15,40,5,204,50,0,0 ; movaps 0x32cc(%rip),%xmm8 # 4bc0 <_sk_callback_sse2+0x3d7>
DB 68,15,40,240 ; movaps %xmm0,%xmm14
DB 69,15,89,240 ; mulps %xmm8,%xmm14
- DB 68,15,40,21,196,49,0,0 ; movaps 0x31c4(%rip),%xmm10 # 41e0 <_sk_callback_sse2+0x2df>
+ DB 68,15,40,21,204,50,0,0 ; movaps 0x32cc(%rip),%xmm10 # 4bd0 <_sk_callback_sse2+0x3e7>
DB 69,15,89,202 ; mulps %xmm10,%xmm9
- DB 68,15,40,29,200,49,0,0 ; movaps 0x31c8(%rip),%xmm11 # 41f0 <_sk_callback_sse2+0x2ef>
+ DB 68,15,40,29,208,50,0,0 ; movaps 0x32d0(%rip),%xmm11 # 4be0 <_sk_callback_sse2+0x3f7>
DB 69,15,88,203 ; addps %xmm11,%xmm9
- DB 68,15,40,37,204,49,0,0 ; movaps 0x31cc(%rip),%xmm12 # 4200 <_sk_callback_sse2+0x2ff>
+ DB 68,15,40,37,212,50,0,0 ; movaps 0x32d4(%rip),%xmm12 # 4bf0 <_sk_callback_sse2+0x407>
DB 69,15,89,236 ; mulps %xmm12,%xmm13
DB 69,15,88,233 ; addps %xmm9,%xmm13
- DB 68,15,40,13,204,49,0,0 ; movaps 0x31cc(%rip),%xmm9 # 4210 <_sk_callback_sse2+0x30f>
+ DB 68,15,40,13,212,50,0,0 ; movaps 0x32d4(%rip),%xmm9 # 4c00 <_sk_callback_sse2+0x417>
DB 69,15,40,249 ; movaps %xmm9,%xmm15
DB 69,15,93,253 ; minps %xmm13,%xmm15
- DB 68,15,40,45,204,49,0,0 ; movaps 0x31cc(%rip),%xmm13 # 4220 <_sk_callback_sse2+0x31f>
+ DB 68,15,40,45,212,50,0,0 ; movaps 0x32d4(%rip),%xmm13 # 4c10 <_sk_callback_sse2+0x427>
DB 65,15,194,197,1 ; cmpltps %xmm13,%xmm0
DB 68,15,84,240 ; andps %xmm0,%xmm14
DB 65,15,85,199 ; andnps %xmm15,%xmm0
@@ -16658,7 +18710,7 @@ _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 68,15,40,45,37,49,0,0 ; movaps 0x3125(%rip),%xmm13 # 4230 <_sk_callback_sse2+0x32f>
+ DB 68,15,40,45,45,50,0,0 ; movaps 0x322d(%rip),%xmm13 # 4c20 <_sk_callback_sse2+0x437>
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
@@ -16667,30 +18719,30 @@ _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 68,15,84,53,11,49,0,0 ; andps 0x310b(%rip),%xmm14 # 4240 <_sk_callback_sse2+0x33f>
+ DB 68,15,84,53,19,50,0,0 ; andps 0x3213(%rip),%xmm14 # 4c30 <_sk_callback_sse2+0x447>
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 68,15,40,37,254,48,0,0 ; movaps 0x30fe(%rip),%xmm12 # 4250 <_sk_callback_sse2+0x34f>
+ DB 68,15,40,37,6,50,0,0 ; movaps 0x3206(%rip),%xmm12 # 4c40 <_sk_callback_sse2+0x457>
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 68,15,88,5,250,48,0,0 ; addps 0x30fa(%rip),%xmm8 # 4260 <_sk_callback_sse2+0x35f>
+ DB 68,15,88,5,2,50,0,0 ; addps 0x3202(%rip),%xmm8 # 4c50 <_sk_callback_sse2+0x467>
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 15,89,5,235,48,0,0 ; mulps 0x30eb(%rip),%xmm0 # 4270 <_sk_callback_sse2+0x36f>
+ DB 15,89,5,243,49,0,0 ; mulps 0x31f3(%rip),%xmm0 # 4c60 <_sk_callback_sse2+0x477>
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 68,15,40,13,222,48,0,0 ; movaps 0x30de(%rip),%xmm9 # 4280 <_sk_callback_sse2+0x37f>
+ DB 68,15,40,13,230,49,0,0 ; movaps 0x31e6(%rip),%xmm9 # 4c70 <_sk_callback_sse2+0x487>
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
@@ -16713,7 +18765,7 @@ _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 68,15,40,13,153,48,0,0 ; movaps 0x3099(%rip),%xmm9 # 4290 <_sk_callback_sse2+0x38f>
+ DB 68,15,40,13,161,49,0,0 ; movaps 0x31a1(%rip),%xmm9 # 4c80 <_sk_callback_sse2+0x497>
DB 69,15,40,209 ; movaps %xmm9,%xmm10
DB 69,15,194,211,2 ; cmpleps %xmm11,%xmm10
DB 15,40,193 ; movaps %xmm1,%xmm0
@@ -16730,28 +18782,28 @@ _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 15,40,5,99,48,0,0 ; movaps 0x3063(%rip),%xmm0 # 42a0 <_sk_callback_sse2+0x39f>
+ DB 15,40,5,107,49,0,0 ; movaps 0x316b(%rip),%xmm0 # 4c90 <_sk_callback_sse2+0x4a7>
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 15,84,29,91,48,0,0 ; andps 0x305b(%rip),%xmm3 # 42b0 <_sk_callback_sse2+0x3af>
+ DB 15,84,29,99,49,0,0 ; andps 0x3163(%rip),%xmm3 # 4ca0 <_sk_callback_sse2+0x4b7>
DB 15,92,203 ; subps %xmm3,%xmm1
DB 15,92,193 ; subps %xmm1,%xmm0
- DB 68,15,40,45,93,48,0,0 ; movaps 0x305d(%rip),%xmm13 # 42c0 <_sk_callback_sse2+0x3bf>
+ DB 68,15,40,45,101,49,0,0 ; movaps 0x3165(%rip),%xmm13 # 4cb0 <_sk_callback_sse2+0x4c7>
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 15,40,21,109,48,0,0 ; movaps 0x306d(%rip),%xmm2 # 42f0 <_sk_callback_sse2+0x3ef>
+ DB 15,40,21,117,49,0,0 ; movaps 0x3175(%rip),%xmm2 # 4ce0 <_sk_callback_sse2+0x4f7>
DB 68,15,40,250 ; movaps %xmm2,%xmm15
DB 68,15,194,248,2 ; cmpleps %xmm0,%xmm15
- DB 15,40,13,61,48,0,0 ; movaps 0x303d(%rip),%xmm1 # 42d0 <_sk_callback_sse2+0x3cf>
+ DB 15,40,13,69,49,0,0 ; movaps 0x3145(%rip),%xmm1 # 4cc0 <_sk_callback_sse2+0x4d7>
DB 15,89,193 ; mulps %xmm1,%xmm0
- DB 15,40,45,67,48,0,0 ; movaps 0x3043(%rip),%xmm5 # 42e0 <_sk_callback_sse2+0x3df>
+ DB 15,40,45,75,49,0,0 ; movaps 0x314b(%rip),%xmm5 # 4cd0 <_sk_callback_sse2+0x4e7>
DB 15,40,229 ; movaps %xmm5,%xmm4
DB 15,92,224 ; subps %xmm0,%xmm4
DB 65,15,89,230 ; mulps %xmm14,%xmm4
@@ -16774,7 +18826,7 @@ _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 15,84,29,184,47,0,0 ; andps 0x2fb8(%rip),%xmm3 # 42b0 <_sk_callback_sse2+0x3af>
+ DB 15,84,29,192,48,0,0 ; andps 0x30c0(%rip),%xmm3 # 4ca0 <_sk_callback_sse2+0x4b7>
DB 15,92,195 ; subps %xmm3,%xmm0
DB 68,15,40,230 ; movaps %xmm6,%xmm12
DB 68,15,92,224 ; subps %xmm0,%xmm12
@@ -16804,12 +18856,12 @@ _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 15,88,53,145,47,0,0 ; addps 0x2f91(%rip),%xmm6 # 4300 <_sk_callback_sse2+0x3ff>
+ DB 15,88,53,153,48,0,0 ; addps 0x3099(%rip),%xmm6 # 4cf0 <_sk_callback_sse2+0x507>
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 15,84,29,44,47,0,0 ; andps 0x2f2c(%rip),%xmm3 # 42b0 <_sk_callback_sse2+0x3af>
+ DB 15,84,29,52,48,0,0 ; andps 0x3034(%rip),%xmm3 # 4ca0 <_sk_callback_sse2+0x4b7>
DB 15,92,195 ; subps %xmm3,%xmm0
DB 15,92,240 ; subps %xmm0,%xmm6
DB 15,89,206 ; mulps %xmm6,%xmm1
@@ -16870,7 +18922,7 @@ _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 68,15,89,5,182,46,0,0 ; mulps 0x2eb6(%rip),%xmm8 # 4310 <_sk_callback_sse2+0x40f>
+ DB 68,15,89,5,190,47,0,0 ; mulps 0x2fbe(%rip),%xmm8 # 4d00 <_sk_callback_sse2+0x517>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 65,15,89,208 ; mulps %xmm8,%xmm2
@@ -16907,7 +18959,7 @@ _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 68,15,89,5,84,46,0,0 ; mulps 0x2e54(%rip),%xmm8 # 4320 <_sk_callback_sse2+0x41f>
+ DB 68,15,89,5,92,47,0,0 ; mulps 0x2f5c(%rip),%xmm8 # 4d10 <_sk_callback_sse2+0x527>
DB 15,92,196 ; subps %xmm4,%xmm0
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -16930,17 +18982,17 @@ _sk_lerp_565_sse2 LABEL PROC
DB 243,68,15,126,4,120 ; movq (%rax,%rdi,2),%xmm8
DB 102,15,239,219 ; pxor %xmm3,%xmm3
DB 102,68,15,97,195 ; punpcklwd %xmm3,%xmm8
- DB 102,15,111,29,28,46,0,0 ; movdqa 0x2e1c(%rip),%xmm3 # 4330 <_sk_callback_sse2+0x42f>
+ DB 102,15,111,29,36,47,0,0 ; movdqa 0x2f24(%rip),%xmm3 # 4d20 <_sk_callback_sse2+0x537>
DB 102,65,15,219,216 ; pand %xmm8,%xmm3
DB 68,15,91,203 ; cvtdq2ps %xmm3,%xmm9
- DB 68,15,89,13,27,46,0,0 ; mulps 0x2e1b(%rip),%xmm9 # 4340 <_sk_callback_sse2+0x43f>
- DB 102,15,111,29,35,46,0,0 ; movdqa 0x2e23(%rip),%xmm3 # 4350 <_sk_callback_sse2+0x44f>
+ DB 68,15,89,13,35,47,0,0 ; mulps 0x2f23(%rip),%xmm9 # 4d30 <_sk_callback_sse2+0x547>
+ DB 102,15,111,29,43,47,0,0 ; movdqa 0x2f2b(%rip),%xmm3 # 4d40 <_sk_callback_sse2+0x557>
DB 102,65,15,219,216 ; pand %xmm8,%xmm3
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,36,46,0,0 ; mulps 0x2e24(%rip),%xmm3 # 4360 <_sk_callback_sse2+0x45f>
- DB 102,68,15,219,5,43,46,0,0 ; pand 0x2e2b(%rip),%xmm8 # 4370 <_sk_callback_sse2+0x46f>
+ DB 15,89,29,44,47,0,0 ; mulps 0x2f2c(%rip),%xmm3 # 4d50 <_sk_callback_sse2+0x567>
+ DB 102,68,15,219,5,51,47,0,0 ; pand 0x2f33(%rip),%xmm8 # 4d60 <_sk_callback_sse2+0x577>
DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
- DB 68,15,89,5,47,46,0,0 ; mulps 0x2e2f(%rip),%xmm8 # 4380 <_sk_callback_sse2+0x47f>
+ DB 68,15,89,5,55,47,0,0 ; mulps 0x2f37(%rip),%xmm8 # 4d70 <_sk_callback_sse2+0x587>
DB 15,92,196 ; subps %xmm4,%xmm0
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -16951,7 +19003,7 @@ _sk_lerp_565_sse2 LABEL PROC
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 15,88,214 ; addps %xmm6,%xmm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,25,46,0,0 ; movaps 0x2e19(%rip),%xmm3 # 4390 <_sk_callback_sse2+0x48f>
+ DB 15,40,29,33,47,0,0 ; movaps 0x2f21(%rip),%xmm3 # 4d80 <_sk_callback_sse2+0x597>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_load_tables_sse2
@@ -16960,7 +19012,7 @@ _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 102,68,15,111,5,15,46,0,0 ; movdqa 0x2e0f(%rip),%xmm8 # 43a0 <_sk_callback_sse2+0x49f>
+ DB 102,68,15,111,5,23,47,0,0 ; movdqa 0x2f17(%rip),%xmm8 # 4d90 <_sk_callback_sse2+0x5a7>
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
@@ -17015,7 +19067,7 @@ _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 15,89,29,28,45,0,0 ; mulps 0x2d1c(%rip),%xmm3 # 43b0 <_sk_callback_sse2+0x4af>
+ DB 15,89,29,36,46,0,0 ; mulps 0x2e24(%rip),%xmm3 # 4da0 <_sk_callback_sse2+0x5b7>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -17032,7 +19084,7 @@ _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 102,68,15,111,21,239,44,0,0 ; movdqa 0x2cef(%rip),%xmm10 # 43c0 <_sk_callback_sse2+0x4bf>
+ DB 102,68,15,111,21,247,45,0,0 ; movdqa 0x2df7(%rip),%xmm10 # 4db0 <_sk_callback_sse2+0x5c7>
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
@@ -17093,7 +19145,7 @@ _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 15,89,29,222,43,0,0 ; mulps 0x2bde(%rip),%xmm3 # 43d0 <_sk_callback_sse2+0x4cf>
+ DB 15,89,29,230,44,0,0 ; mulps 0x2ce6(%rip),%xmm3 # 4dc0 <_sk_callback_sse2+0x5d7>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -17113,7 +19165,7 @@ _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 102,68,15,111,5,158,43,0,0 ; movdqa 0x2b9e(%rip),%xmm8 # 43e0 <_sk_callback_sse2+0x4df>
+ DB 102,68,15,111,5,166,44,0,0 ; movdqa 0x2ca6(%rip),%xmm8 # 4dd0 <_sk_callback_sse2+0x5e7>
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
@@ -17168,7 +19220,7 @@ _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 15,40,29,173,42,0,0 ; movaps 0x2aad(%rip),%xmm3 # 43f0 <_sk_callback_sse2+0x4ef>
+ DB 15,40,29,181,43,0,0 ; movaps 0x2bb5(%rip),%xmm3 # 4de0 <_sk_callback_sse2+0x5f7>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_byte_tables_sse2
@@ -17176,7 +19228,7 @@ _sk_byte_tables_sse2 LABEL PROC
DB 65,86 ; push %r14
DB 83 ; push %rbx
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 68,15,40,5,174,42,0,0 ; movaps 0x2aae(%rip),%xmm8 # 4400 <_sk_callback_sse2+0x4ff>
+ DB 68,15,40,5,182,43,0,0 ; movaps 0x2bb6(%rip),%xmm8 # 4df0 <_sk_callback_sse2+0x607>
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
@@ -17203,7 +19255,7 @@ _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 68,15,40,21,75,42,0,0 ; movaps 0x2a4b(%rip),%xmm10 # 4410 <_sk_callback_sse2+0x50f>
+ DB 68,15,40,21,83,43,0,0 ; movaps 0x2b53(%rip),%xmm10 # 4e00 <_sk_callback_sse2+0x617>
DB 65,15,89,194 ; mulps %xmm10,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
@@ -17317,7 +19369,7 @@ _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 68,15,40,21,158,40,0,0 ; movaps 0x289e(%rip),%xmm10 # 4420 <_sk_callback_sse2+0x51f>
+ DB 68,15,40,21,166,41,0,0 ; movaps 0x29a6(%rip),%xmm10 # 4e10 <_sk_callback_sse2+0x627>
DB 65,15,89,194 ; mulps %xmm10,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
@@ -17504,15 +19556,15 @@ _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 68,15,89,13,221,37,0,0 ; mulps 0x25dd(%rip),%xmm9 # 4430 <_sk_callback_sse2+0x52f>
- DB 68,15,84,21,229,37,0,0 ; andps 0x25e5(%rip),%xmm10 # 4440 <_sk_callback_sse2+0x53f>
- DB 68,15,86,21,237,37,0,0 ; orps 0x25ed(%rip),%xmm10 # 4450 <_sk_callback_sse2+0x54f>
- DB 68,15,88,13,245,37,0,0 ; addps 0x25f5(%rip),%xmm9 # 4460 <_sk_callback_sse2+0x55f>
- DB 68,15,40,37,253,37,0,0 ; movaps 0x25fd(%rip),%xmm12 # 4470 <_sk_callback_sse2+0x56f>
+ DB 68,15,89,13,229,38,0,0 ; mulps 0x26e5(%rip),%xmm9 # 4e20 <_sk_callback_sse2+0x637>
+ DB 68,15,84,21,237,38,0,0 ; andps 0x26ed(%rip),%xmm10 # 4e30 <_sk_callback_sse2+0x647>
+ DB 68,15,86,21,245,38,0,0 ; orps 0x26f5(%rip),%xmm10 # 4e40 <_sk_callback_sse2+0x657>
+ DB 68,15,88,13,253,38,0,0 ; addps 0x26fd(%rip),%xmm9 # 4e50 <_sk_callback_sse2+0x667>
+ DB 68,15,40,37,5,39,0,0 ; movaps 0x2705(%rip),%xmm12 # 4e60 <_sk_callback_sse2+0x677>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,253,37,0,0 ; addps 0x25fd(%rip),%xmm10 # 4480 <_sk_callback_sse2+0x57f>
- DB 68,15,40,37,5,38,0,0 ; movaps 0x2605(%rip),%xmm12 # 4490 <_sk_callback_sse2+0x58f>
+ DB 68,15,88,21,5,39,0,0 ; addps 0x2705(%rip),%xmm10 # 4e70 <_sk_callback_sse2+0x687>
+ DB 68,15,40,37,13,39,0,0 ; movaps 0x270d(%rip),%xmm12 # 4e80 <_sk_callback_sse2+0x697>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -17520,22 +19572,22 @@ _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 68,15,40,21,239,37,0,0 ; movaps 0x25ef(%rip),%xmm10 # 44a0 <_sk_callback_sse2+0x59f>
+ DB 68,15,40,21,247,38,0,0 ; movaps 0x26f7(%rip),%xmm10 # 4e90 <_sk_callback_sse2+0x6a7>
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 68,15,88,13,227,37,0,0 ; addps 0x25e3(%rip),%xmm9 # 44b0 <_sk_callback_sse2+0x5af>
- DB 68,15,40,37,235,37,0,0 ; movaps 0x25eb(%rip),%xmm12 # 44c0 <_sk_callback_sse2+0x5bf>
+ DB 68,15,88,13,235,38,0,0 ; addps 0x26eb(%rip),%xmm9 # 4ea0 <_sk_callback_sse2+0x6b7>
+ DB 68,15,40,37,243,38,0,0 ; movaps 0x26f3(%rip),%xmm12 # 4eb0 <_sk_callback_sse2+0x6c7>
DB 69,15,89,229 ; mulps %xmm13,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,40,37,235,37,0,0 ; movaps 0x25eb(%rip),%xmm12 # 44d0 <_sk_callback_sse2+0x5cf>
+ DB 68,15,40,37,243,38,0,0 ; movaps 0x26f3(%rip),%xmm12 # 4ec0 <_sk_callback_sse2+0x6d7>
DB 69,15,92,229 ; subps %xmm13,%xmm12
- DB 68,15,40,45,239,37,0,0 ; movaps 0x25ef(%rip),%xmm13 # 44e0 <_sk_callback_sse2+0x5df>
+ DB 68,15,40,45,247,38,0,0 ; movaps 0x26f7(%rip),%xmm13 # 4ed0 <_sk_callback_sse2+0x6e7>
DB 69,15,94,236 ; divps %xmm12,%xmm13
DB 69,15,88,233 ; addps %xmm9,%xmm13
- DB 68,15,89,45,239,37,0,0 ; mulps 0x25ef(%rip),%xmm13 # 44f0 <_sk_callback_sse2+0x5ef>
+ DB 68,15,89,45,247,38,0,0 ; mulps 0x26f7(%rip),%xmm13 # 4ee0 <_sk_callback_sse2+0x6f7>
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
@@ -17569,15 +19621,15 @@ _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 68,15,89,13,111,37,0,0 ; mulps 0x256f(%rip),%xmm9 # 4500 <_sk_callback_sse2+0x5ff>
- DB 68,15,84,21,119,37,0,0 ; andps 0x2577(%rip),%xmm10 # 4510 <_sk_callback_sse2+0x60f>
- DB 68,15,86,21,127,37,0,0 ; orps 0x257f(%rip),%xmm10 # 4520 <_sk_callback_sse2+0x61f>
- DB 68,15,88,13,135,37,0,0 ; addps 0x2587(%rip),%xmm9 # 4530 <_sk_callback_sse2+0x62f>
- DB 68,15,40,37,143,37,0,0 ; movaps 0x258f(%rip),%xmm12 # 4540 <_sk_callback_sse2+0x63f>
+ DB 68,15,89,13,119,38,0,0 ; mulps 0x2677(%rip),%xmm9 # 4ef0 <_sk_callback_sse2+0x707>
+ DB 68,15,84,21,127,38,0,0 ; andps 0x267f(%rip),%xmm10 # 4f00 <_sk_callback_sse2+0x717>
+ DB 68,15,86,21,135,38,0,0 ; orps 0x2687(%rip),%xmm10 # 4f10 <_sk_callback_sse2+0x727>
+ DB 68,15,88,13,143,38,0,0 ; addps 0x268f(%rip),%xmm9 # 4f20 <_sk_callback_sse2+0x737>
+ DB 68,15,40,37,151,38,0,0 ; movaps 0x2697(%rip),%xmm12 # 4f30 <_sk_callback_sse2+0x747>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,143,37,0,0 ; addps 0x258f(%rip),%xmm10 # 4550 <_sk_callback_sse2+0x64f>
- DB 68,15,40,37,151,37,0,0 ; movaps 0x2597(%rip),%xmm12 # 4560 <_sk_callback_sse2+0x65f>
+ DB 68,15,88,21,151,38,0,0 ; addps 0x2697(%rip),%xmm10 # 4f40 <_sk_callback_sse2+0x757>
+ DB 68,15,40,37,159,38,0,0 ; movaps 0x269f(%rip),%xmm12 # 4f50 <_sk_callback_sse2+0x767>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -17585,22 +19637,22 @@ _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 68,15,40,21,129,37,0,0 ; movaps 0x2581(%rip),%xmm10 # 4570 <_sk_callback_sse2+0x66f>
+ DB 68,15,40,21,137,38,0,0 ; movaps 0x2689(%rip),%xmm10 # 4f60 <_sk_callback_sse2+0x777>
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 68,15,88,13,117,37,0,0 ; addps 0x2575(%rip),%xmm9 # 4580 <_sk_callback_sse2+0x67f>
- DB 68,15,40,37,125,37,0,0 ; movaps 0x257d(%rip),%xmm12 # 4590 <_sk_callback_sse2+0x68f>
+ DB 68,15,88,13,125,38,0,0 ; addps 0x267d(%rip),%xmm9 # 4f70 <_sk_callback_sse2+0x787>
+ DB 68,15,40,37,133,38,0,0 ; movaps 0x2685(%rip),%xmm12 # 4f80 <_sk_callback_sse2+0x797>
DB 69,15,89,229 ; mulps %xmm13,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,40,37,125,37,0,0 ; movaps 0x257d(%rip),%xmm12 # 45a0 <_sk_callback_sse2+0x69f>
+ DB 68,15,40,37,133,38,0,0 ; movaps 0x2685(%rip),%xmm12 # 4f90 <_sk_callback_sse2+0x7a7>
DB 69,15,92,229 ; subps %xmm13,%xmm12
- DB 68,15,40,45,129,37,0,0 ; movaps 0x2581(%rip),%xmm13 # 45b0 <_sk_callback_sse2+0x6af>
+ DB 68,15,40,45,137,38,0,0 ; movaps 0x2689(%rip),%xmm13 # 4fa0 <_sk_callback_sse2+0x7b7>
DB 69,15,94,236 ; divps %xmm12,%xmm13
DB 69,15,88,233 ; addps %xmm9,%xmm13
- DB 68,15,89,45,129,37,0,0 ; mulps 0x2581(%rip),%xmm13 # 45c0 <_sk_callback_sse2+0x6bf>
+ DB 68,15,89,45,137,38,0,0 ; mulps 0x2689(%rip),%xmm13 # 4fb0 <_sk_callback_sse2+0x7c7>
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
@@ -17634,15 +19686,15 @@ _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 68,15,89,13,1,37,0,0 ; mulps 0x2501(%rip),%xmm9 # 45d0 <_sk_callback_sse2+0x6cf>
- DB 68,15,84,21,9,37,0,0 ; andps 0x2509(%rip),%xmm10 # 45e0 <_sk_callback_sse2+0x6df>
- DB 68,15,86,21,17,37,0,0 ; orps 0x2511(%rip),%xmm10 # 45f0 <_sk_callback_sse2+0x6ef>
- DB 68,15,88,13,25,37,0,0 ; addps 0x2519(%rip),%xmm9 # 4600 <_sk_callback_sse2+0x6ff>
- DB 68,15,40,37,33,37,0,0 ; movaps 0x2521(%rip),%xmm12 # 4610 <_sk_callback_sse2+0x70f>
+ DB 68,15,89,13,9,38,0,0 ; mulps 0x2609(%rip),%xmm9 # 4fc0 <_sk_callback_sse2+0x7d7>
+ DB 68,15,84,21,17,38,0,0 ; andps 0x2611(%rip),%xmm10 # 4fd0 <_sk_callback_sse2+0x7e7>
+ DB 68,15,86,21,25,38,0,0 ; orps 0x2619(%rip),%xmm10 # 4fe0 <_sk_callback_sse2+0x7f7>
+ DB 68,15,88,13,33,38,0,0 ; addps 0x2621(%rip),%xmm9 # 4ff0 <_sk_callback_sse2+0x807>
+ DB 68,15,40,37,41,38,0,0 ; movaps 0x2629(%rip),%xmm12 # 5000 <_sk_callback_sse2+0x817>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,33,37,0,0 ; addps 0x2521(%rip),%xmm10 # 4620 <_sk_callback_sse2+0x71f>
- DB 68,15,40,37,41,37,0,0 ; movaps 0x2529(%rip),%xmm12 # 4630 <_sk_callback_sse2+0x72f>
+ DB 68,15,88,21,41,38,0,0 ; addps 0x2629(%rip),%xmm10 # 5010 <_sk_callback_sse2+0x827>
+ DB 68,15,40,37,49,38,0,0 ; movaps 0x2631(%rip),%xmm12 # 5020 <_sk_callback_sse2+0x837>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -17650,22 +19702,22 @@ _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 68,15,40,21,19,37,0,0 ; movaps 0x2513(%rip),%xmm10 # 4640 <_sk_callback_sse2+0x73f>
+ DB 68,15,40,21,27,38,0,0 ; movaps 0x261b(%rip),%xmm10 # 5030 <_sk_callback_sse2+0x847>
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 68,15,88,13,7,37,0,0 ; addps 0x2507(%rip),%xmm9 # 4650 <_sk_callback_sse2+0x74f>
- DB 68,15,40,37,15,37,0,0 ; movaps 0x250f(%rip),%xmm12 # 4660 <_sk_callback_sse2+0x75f>
+ DB 68,15,88,13,15,38,0,0 ; addps 0x260f(%rip),%xmm9 # 5040 <_sk_callback_sse2+0x857>
+ DB 68,15,40,37,23,38,0,0 ; movaps 0x2617(%rip),%xmm12 # 5050 <_sk_callback_sse2+0x867>
DB 69,15,89,229 ; mulps %xmm13,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,40,37,15,37,0,0 ; movaps 0x250f(%rip),%xmm12 # 4670 <_sk_callback_sse2+0x76f>
+ DB 68,15,40,37,23,38,0,0 ; movaps 0x2617(%rip),%xmm12 # 5060 <_sk_callback_sse2+0x877>
DB 69,15,92,229 ; subps %xmm13,%xmm12
- DB 68,15,40,45,19,37,0,0 ; movaps 0x2513(%rip),%xmm13 # 4680 <_sk_callback_sse2+0x77f>
+ DB 68,15,40,45,27,38,0,0 ; movaps 0x261b(%rip),%xmm13 # 5070 <_sk_callback_sse2+0x887>
DB 69,15,94,236 ; divps %xmm12,%xmm13
DB 69,15,88,233 ; addps %xmm9,%xmm13
- DB 68,15,89,45,19,37,0,0 ; mulps 0x2513(%rip),%xmm13 # 4690 <_sk_callback_sse2+0x78f>
+ DB 68,15,89,45,27,38,0,0 ; mulps 0x261b(%rip),%xmm13 # 5080 <_sk_callback_sse2+0x897>
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
@@ -17699,15 +19751,15 @@ _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 68,15,89,13,147,36,0,0 ; mulps 0x2493(%rip),%xmm9 # 46a0 <_sk_callback_sse2+0x79f>
- DB 68,15,84,21,155,36,0,0 ; andps 0x249b(%rip),%xmm10 # 46b0 <_sk_callback_sse2+0x7af>
- DB 68,15,86,21,163,36,0,0 ; orps 0x24a3(%rip),%xmm10 # 46c0 <_sk_callback_sse2+0x7bf>
- DB 68,15,88,13,171,36,0,0 ; addps 0x24ab(%rip),%xmm9 # 46d0 <_sk_callback_sse2+0x7cf>
- DB 68,15,40,37,179,36,0,0 ; movaps 0x24b3(%rip),%xmm12 # 46e0 <_sk_callback_sse2+0x7df>
+ DB 68,15,89,13,155,37,0,0 ; mulps 0x259b(%rip),%xmm9 # 5090 <_sk_callback_sse2+0x8a7>
+ DB 68,15,84,21,163,37,0,0 ; andps 0x25a3(%rip),%xmm10 # 50a0 <_sk_callback_sse2+0x8b7>
+ DB 68,15,86,21,171,37,0,0 ; orps 0x25ab(%rip),%xmm10 # 50b0 <_sk_callback_sse2+0x8c7>
+ DB 68,15,88,13,179,37,0,0 ; addps 0x25b3(%rip),%xmm9 # 50c0 <_sk_callback_sse2+0x8d7>
+ DB 68,15,40,37,187,37,0,0 ; movaps 0x25bb(%rip),%xmm12 # 50d0 <_sk_callback_sse2+0x8e7>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,179,36,0,0 ; addps 0x24b3(%rip),%xmm10 # 46f0 <_sk_callback_sse2+0x7ef>
- DB 68,15,40,37,187,36,0,0 ; movaps 0x24bb(%rip),%xmm12 # 4700 <_sk_callback_sse2+0x7ff>
+ DB 68,15,88,21,187,37,0,0 ; addps 0x25bb(%rip),%xmm10 # 50e0 <_sk_callback_sse2+0x8f7>
+ DB 68,15,40,37,195,37,0,0 ; movaps 0x25c3(%rip),%xmm12 # 50f0 <_sk_callback_sse2+0x907>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -17715,22 +19767,22 @@ _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 68,15,40,21,165,36,0,0 ; movaps 0x24a5(%rip),%xmm10 # 4710 <_sk_callback_sse2+0x80f>
+ DB 68,15,40,21,173,37,0,0 ; movaps 0x25ad(%rip),%xmm10 # 5100 <_sk_callback_sse2+0x917>
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 68,15,88,13,153,36,0,0 ; addps 0x2499(%rip),%xmm9 # 4720 <_sk_callback_sse2+0x81f>
- DB 68,15,40,37,161,36,0,0 ; movaps 0x24a1(%rip),%xmm12 # 4730 <_sk_callback_sse2+0x82f>
+ DB 68,15,88,13,161,37,0,0 ; addps 0x25a1(%rip),%xmm9 # 5110 <_sk_callback_sse2+0x927>
+ DB 68,15,40,37,169,37,0,0 ; movaps 0x25a9(%rip),%xmm12 # 5120 <_sk_callback_sse2+0x937>
DB 69,15,89,229 ; mulps %xmm13,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,40,37,161,36,0,0 ; movaps 0x24a1(%rip),%xmm12 # 4740 <_sk_callback_sse2+0x83f>
+ DB 68,15,40,37,169,37,0,0 ; movaps 0x25a9(%rip),%xmm12 # 5130 <_sk_callback_sse2+0x947>
DB 69,15,92,229 ; subps %xmm13,%xmm12
- DB 68,15,40,45,165,36,0,0 ; movaps 0x24a5(%rip),%xmm13 # 4750 <_sk_callback_sse2+0x84f>
+ DB 68,15,40,45,173,37,0,0 ; movaps 0x25ad(%rip),%xmm13 # 5140 <_sk_callback_sse2+0x957>
DB 69,15,94,236 ; divps %xmm12,%xmm13
DB 69,15,88,233 ; addps %xmm9,%xmm13
- DB 68,15,89,45,165,36,0,0 ; mulps 0x24a5(%rip),%xmm13 # 4760 <_sk_callback_sse2+0x85f>
+ DB 68,15,89,45,173,37,0,0 ; mulps 0x25ad(%rip),%xmm13 # 5150 <_sk_callback_sse2+0x967>
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
@@ -17745,29 +19797,29 @@ _sk_parametric_a_sse2 LABEL PROC
PUBLIC _sk_lab_to_xyz_sse2
_sk_lab_to_xyz_sse2 LABEL PROC
- DB 15,89,5,130,36,0,0 ; mulps 0x2482(%rip),%xmm0 # 4770 <_sk_callback_sse2+0x86f>
- DB 68,15,40,5,138,36,0,0 ; movaps 0x248a(%rip),%xmm8 # 4780 <_sk_callback_sse2+0x87f>
+ DB 15,89,5,138,37,0,0 ; mulps 0x258a(%rip),%xmm0 # 5160 <_sk_callback_sse2+0x977>
+ DB 68,15,40,5,146,37,0,0 ; movaps 0x2592(%rip),%xmm8 # 5170 <_sk_callback_sse2+0x987>
DB 65,15,89,200 ; mulps %xmm8,%xmm1
- DB 68,15,40,13,142,36,0,0 ; movaps 0x248e(%rip),%xmm9 # 4790 <_sk_callback_sse2+0x88f>
+ DB 68,15,40,13,150,37,0,0 ; movaps 0x2596(%rip),%xmm9 # 5180 <_sk_callback_sse2+0x997>
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 15,88,5,139,36,0,0 ; addps 0x248b(%rip),%xmm0 # 47a0 <_sk_callback_sse2+0x89f>
- DB 15,89,5,148,36,0,0 ; mulps 0x2494(%rip),%xmm0 # 47b0 <_sk_callback_sse2+0x8af>
- DB 15,89,13,157,36,0,0 ; mulps 0x249d(%rip),%xmm1 # 47c0 <_sk_callback_sse2+0x8bf>
+ DB 15,88,5,147,37,0,0 ; addps 0x2593(%rip),%xmm0 # 5190 <_sk_callback_sse2+0x9a7>
+ DB 15,89,5,156,37,0,0 ; mulps 0x259c(%rip),%xmm0 # 51a0 <_sk_callback_sse2+0x9b7>
+ DB 15,89,13,165,37,0,0 ; mulps 0x25a5(%rip),%xmm1 # 51b0 <_sk_callback_sse2+0x9c7>
DB 15,88,200 ; addps %xmm0,%xmm1
- DB 15,89,21,163,36,0,0 ; mulps 0x24a3(%rip),%xmm2 # 47d0 <_sk_callback_sse2+0x8cf>
+ DB 15,89,21,171,37,0,0 ; mulps 0x25ab(%rip),%xmm2 # 51c0 <_sk_callback_sse2+0x9d7>
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 15,40,21,152,36,0,0 ; movaps 0x2498(%rip),%xmm2 # 47e0 <_sk_callback_sse2+0x8df>
+ DB 15,40,21,160,37,0,0 ; movaps 0x25a0(%rip),%xmm2 # 51d0 <_sk_callback_sse2+0x9e7>
DB 68,15,40,194 ; movaps %xmm2,%xmm8
DB 69,15,194,196,1 ; cmpltps %xmm12,%xmm8
- DB 68,15,40,21,151,36,0,0 ; movaps 0x2497(%rip),%xmm10 # 47f0 <_sk_callback_sse2+0x8ef>
+ DB 68,15,40,21,159,37,0,0 ; movaps 0x259f(%rip),%xmm10 # 51e0 <_sk_callback_sse2+0x9f7>
DB 65,15,88,202 ; addps %xmm10,%xmm1
- DB 68,15,40,29,155,36,0,0 ; movaps 0x249b(%rip),%xmm11 # 4800 <_sk_callback_sse2+0x8ff>
+ DB 68,15,40,29,163,37,0,0 ; movaps 0x25a3(%rip),%xmm11 # 51f0 <_sk_callback_sse2+0xa07>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 69,15,84,224 ; andps %xmm8,%xmm12
DB 68,15,85,193 ; andnps %xmm1,%xmm8
@@ -17791,8 +19843,8 @@ _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 68,15,89,5,75,36,0,0 ; mulps 0x244b(%rip),%xmm8 # 4810 <_sk_callback_sse2+0x90f>
- DB 15,89,21,84,36,0,0 ; mulps 0x2454(%rip),%xmm2 # 4820 <_sk_callback_sse2+0x91f>
+ DB 68,15,89,5,83,37,0,0 ; mulps 0x2553(%rip),%xmm8 # 5200 <_sk_callback_sse2+0xa17>
+ DB 15,89,21,92,37,0,0 ; mulps 0x255c(%rip),%xmm2 # 5210 <_sk_callback_sse2+0xa27>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -17806,7 +19858,7 @@ _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 15,89,29,60,36,0,0 ; mulps 0x243c(%rip),%xmm3 # 4830 <_sk_callback_sse2+0x92f>
+ DB 15,89,29,68,37,0,0 ; mulps 0x2544(%rip),%xmm3 # 5220 <_sk_callback_sse2+0xa37>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
DB 102,15,239,201 ; pxor %xmm1,%xmm1
@@ -17849,7 +19901,7 @@ _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 15,89,29,171,35,0,0 ; mulps 0x23ab(%rip),%xmm3 # 4840 <_sk_callback_sse2+0x93f>
+ DB 15,89,29,179,36,0,0 ; mulps 0x24b3(%rip),%xmm3 # 5230 <_sk_callback_sse2+0xa47>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
DB 102,15,239,201 ; pxor %xmm1,%xmm1
@@ -17860,7 +19912,7 @@ 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 68,15,40,5,159,35,0,0 ; movaps 0x239f(%rip),%xmm8 # 4850 <_sk_callback_sse2+0x94f>
+ DB 68,15,40,5,167,36,0,0 ; movaps 0x24a7(%rip),%xmm8 # 5240 <_sk_callback_sse2+0xa57>
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
@@ -17880,9 +19932,9 @@ _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 15,89,5,102,35,0,0 ; mulps 0x2366(%rip),%xmm0 # 4860 <_sk_callback_sse2+0x95f>
+ DB 15,89,5,110,36,0,0 ; mulps 0x246e(%rip),%xmm0 # 5250 <_sk_callback_sse2+0xa67>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,109,35,0,0 ; movaps 0x236d(%rip),%xmm3 # 4870 <_sk_callback_sse2+0x96f>
+ DB 15,40,29,117,36,0,0 ; movaps 0x2475(%rip),%xmm3 # 5260 <_sk_callback_sse2+0xa77>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 15,40,208 ; movaps %xmm0,%xmm2
DB 255,224 ; jmpq *%rax
@@ -17923,9 +19975,9 @@ _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 15,89,5,226,34,0,0 ; mulps 0x22e2(%rip),%xmm0 # 4880 <_sk_callback_sse2+0x97f>
+ DB 15,89,5,234,35,0,0 ; mulps 0x23ea(%rip),%xmm0 # 5270 <_sk_callback_sse2+0xa87>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,233,34,0,0 ; movaps 0x22e9(%rip),%xmm3 # 4890 <_sk_callback_sse2+0x98f>
+ DB 15,40,29,241,35,0,0 ; movaps 0x23f1(%rip),%xmm3 # 5280 <_sk_callback_sse2+0xa97>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 15,40,208 ; movaps %xmm0,%xmm2
DB 255,224 ; jmpq *%rax
@@ -17935,9 +19987,9 @@ _sk_gather_i8_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 73,137,192 ; mov %rax,%r8
DB 77,133,192 ; test %r8,%r8
- DB 116,5 ; je 25be <_sk_gather_i8_sse2+0xf>
+ DB 116,5 ; je 2ea6 <_sk_gather_i8_sse2+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 25c0 <_sk_gather_i8_sse2+0x11>
+ DB 235,2 ; jmp 2ea8 <_sk_gather_i8_sse2+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1
@@ -17986,11 +20038,11 @@ _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 102,15,111,21,8,34,0,0 ; movdqa 0x2208(%rip),%xmm2 # 48a0 <_sk_callback_sse2+0x99f>
+ DB 102,15,111,21,16,35,0,0 ; movdqa 0x2310(%rip),%xmm2 # 5290 <_sk_callback_sse2+0xaa7>
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 68,15,40,5,4,34,0,0 ; movaps 0x2204(%rip),%xmm8 # 48b0 <_sk_callback_sse2+0x9af>
+ DB 68,15,40,5,12,35,0,0 ; movaps 0x230c(%rip),%xmm8 # 52a0 <_sk_callback_sse2+0xab7>
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
@@ -18015,19 +20067,19 @@ _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 102,15,111,5,186,33,0,0 ; movdqa 0x21ba(%rip),%xmm0 # 48c0 <_sk_callback_sse2+0x9bf>
+ DB 102,15,111,5,194,34,0,0 ; movdqa 0x22c2(%rip),%xmm0 # 52b0 <_sk_callback_sse2+0xac7>
DB 102,15,219,194 ; pand %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,188,33,0,0 ; mulps 0x21bc(%rip),%xmm0 # 48d0 <_sk_callback_sse2+0x9cf>
- DB 102,15,111,13,196,33,0,0 ; movdqa 0x21c4(%rip),%xmm1 # 48e0 <_sk_callback_sse2+0x9df>
+ DB 15,89,5,196,34,0,0 ; mulps 0x22c4(%rip),%xmm0 # 52c0 <_sk_callback_sse2+0xad7>
+ DB 102,15,111,13,204,34,0,0 ; movdqa 0x22cc(%rip),%xmm1 # 52d0 <_sk_callback_sse2+0xae7>
DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,198,33,0,0 ; mulps 0x21c6(%rip),%xmm1 # 48f0 <_sk_callback_sse2+0x9ef>
- DB 102,15,219,21,206,33,0,0 ; pand 0x21ce(%rip),%xmm2 # 4900 <_sk_callback_sse2+0x9ff>
+ DB 15,89,13,206,34,0,0 ; mulps 0x22ce(%rip),%xmm1 # 52e0 <_sk_callback_sse2+0xaf7>
+ DB 102,15,219,21,214,34,0,0 ; pand 0x22d6(%rip),%xmm2 # 52f0 <_sk_callback_sse2+0xb07>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,212,33,0,0 ; mulps 0x21d4(%rip),%xmm2 # 4910 <_sk_callback_sse2+0xa0f>
+ DB 15,89,21,220,34,0,0 ; mulps 0x22dc(%rip),%xmm2 # 5300 <_sk_callback_sse2+0xb17>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,219,33,0,0 ; movaps 0x21db(%rip),%xmm3 # 4920 <_sk_callback_sse2+0xa1f>
+ DB 15,40,29,227,34,0,0 ; movaps 0x22e3(%rip),%xmm3 # 5310 <_sk_callback_sse2+0xb27>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_gather_565_sse2
@@ -18060,31 +20112,31 @@ _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 102,15,111,5,100,33,0,0 ; movdqa 0x2164(%rip),%xmm0 # 4930 <_sk_callback_sse2+0xa2f>
+ DB 102,15,111,5,108,34,0,0 ; movdqa 0x226c(%rip),%xmm0 # 5320 <_sk_callback_sse2+0xb37>
DB 102,15,219,194 ; pand %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,102,33,0,0 ; mulps 0x2166(%rip),%xmm0 # 4940 <_sk_callback_sse2+0xa3f>
- DB 102,15,111,13,110,33,0,0 ; movdqa 0x216e(%rip),%xmm1 # 4950 <_sk_callback_sse2+0xa4f>
+ DB 15,89,5,110,34,0,0 ; mulps 0x226e(%rip),%xmm0 # 5330 <_sk_callback_sse2+0xb47>
+ DB 102,15,111,13,118,34,0,0 ; movdqa 0x2276(%rip),%xmm1 # 5340 <_sk_callback_sse2+0xb57>
DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,112,33,0,0 ; mulps 0x2170(%rip),%xmm1 # 4960 <_sk_callback_sse2+0xa5f>
- DB 102,15,219,21,120,33,0,0 ; pand 0x2178(%rip),%xmm2 # 4970 <_sk_callback_sse2+0xa6f>
+ DB 15,89,13,120,34,0,0 ; mulps 0x2278(%rip),%xmm1 # 5350 <_sk_callback_sse2+0xb67>
+ DB 102,15,219,21,128,34,0,0 ; pand 0x2280(%rip),%xmm2 # 5360 <_sk_callback_sse2+0xb77>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,126,33,0,0 ; mulps 0x217e(%rip),%xmm2 # 4980 <_sk_callback_sse2+0xa7f>
+ DB 15,89,21,134,34,0,0 ; mulps 0x2286(%rip),%xmm2 # 5370 <_sk_callback_sse2+0xb87>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,133,33,0,0 ; movaps 0x2185(%rip),%xmm3 # 4990 <_sk_callback_sse2+0xa8f>
+ DB 15,40,29,141,34,0,0 ; movaps 0x228d(%rip),%xmm3 # 5380 <_sk_callback_sse2+0xb97>
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 68,15,40,5,134,33,0,0 ; movaps 0x2186(%rip),%xmm8 # 49a0 <_sk_callback_sse2+0xa9f>
+ DB 68,15,40,5,142,34,0,0 ; movaps 0x228e(%rip),%xmm8 # 5390 <_sk_callback_sse2+0xba7>
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 68,15,40,21,123,33,0,0 ; movaps 0x217b(%rip),%xmm10 # 49b0 <_sk_callback_sse2+0xaaf>
+ DB 68,15,40,21,131,34,0,0 ; movaps 0x2283(%rip),%xmm10 # 53a0 <_sk_callback_sse2+0xbb7>
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
@@ -18106,21 +20158,21 @@ _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 102,15,111,5,52,33,0,0 ; movdqa 0x2134(%rip),%xmm0 # 49c0 <_sk_callback_sse2+0xabf>
+ DB 102,15,111,5,60,34,0,0 ; movdqa 0x223c(%rip),%xmm0 # 53b0 <_sk_callback_sse2+0xbc7>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,54,33,0,0 ; mulps 0x2136(%rip),%xmm0 # 49d0 <_sk_callback_sse2+0xacf>
- DB 102,15,111,13,62,33,0,0 ; movdqa 0x213e(%rip),%xmm1 # 49e0 <_sk_callback_sse2+0xadf>
+ DB 15,89,5,62,34,0,0 ; mulps 0x223e(%rip),%xmm0 # 53c0 <_sk_callback_sse2+0xbd7>
+ DB 102,15,111,13,70,34,0,0 ; movdqa 0x2246(%rip),%xmm1 # 53d0 <_sk_callback_sse2+0xbe7>
DB 102,15,219,203 ; pand %xmm3,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,64,33,0,0 ; mulps 0x2140(%rip),%xmm1 # 49f0 <_sk_callback_sse2+0xaef>
- DB 102,15,111,21,72,33,0,0 ; movdqa 0x2148(%rip),%xmm2 # 4a00 <_sk_callback_sse2+0xaff>
+ DB 15,89,13,72,34,0,0 ; mulps 0x2248(%rip),%xmm1 # 53e0 <_sk_callback_sse2+0xbf7>
+ DB 102,15,111,21,80,34,0,0 ; movdqa 0x2250(%rip),%xmm2 # 53f0 <_sk_callback_sse2+0xc07>
DB 102,15,219,211 ; pand %xmm3,%xmm2
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,74,33,0,0 ; mulps 0x214a(%rip),%xmm2 # 4a10 <_sk_callback_sse2+0xb0f>
- DB 102,15,219,29,82,33,0,0 ; pand 0x2152(%rip),%xmm3 # 4a20 <_sk_callback_sse2+0xb1f>
+ DB 15,89,21,82,34,0,0 ; mulps 0x2252(%rip),%xmm2 # 5400 <_sk_callback_sse2+0xc17>
+ DB 102,15,219,29,90,34,0,0 ; pand 0x225a(%rip),%xmm3 # 5410 <_sk_callback_sse2+0xc27>
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,88,33,0,0 ; mulps 0x2158(%rip),%xmm3 # 4a30 <_sk_callback_sse2+0xb2f>
+ DB 15,89,29,96,34,0,0 ; mulps 0x2260(%rip),%xmm3 # 5420 <_sk_callback_sse2+0xc37>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -18154,21 +20206,21 @@ _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 102,15,111,5,223,32,0,0 ; movdqa 0x20df(%rip),%xmm0 # 4a40 <_sk_callback_sse2+0xb3f>
+ DB 102,15,111,5,231,33,0,0 ; movdqa 0x21e7(%rip),%xmm0 # 5430 <_sk_callback_sse2+0xc47>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,225,32,0,0 ; mulps 0x20e1(%rip),%xmm0 # 4a50 <_sk_callback_sse2+0xb4f>
- DB 102,15,111,13,233,32,0,0 ; movdqa 0x20e9(%rip),%xmm1 # 4a60 <_sk_callback_sse2+0xb5f>
+ DB 15,89,5,233,33,0,0 ; mulps 0x21e9(%rip),%xmm0 # 5440 <_sk_callback_sse2+0xc57>
+ DB 102,15,111,13,241,33,0,0 ; movdqa 0x21f1(%rip),%xmm1 # 5450 <_sk_callback_sse2+0xc67>
DB 102,15,219,203 ; pand %xmm3,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,235,32,0,0 ; mulps 0x20eb(%rip),%xmm1 # 4a70 <_sk_callback_sse2+0xb6f>
- DB 102,15,111,21,243,32,0,0 ; movdqa 0x20f3(%rip),%xmm2 # 4a80 <_sk_callback_sse2+0xb7f>
+ DB 15,89,13,243,33,0,0 ; mulps 0x21f3(%rip),%xmm1 # 5460 <_sk_callback_sse2+0xc77>
+ DB 102,15,111,21,251,33,0,0 ; movdqa 0x21fb(%rip),%xmm2 # 5470 <_sk_callback_sse2+0xc87>
DB 102,15,219,211 ; pand %xmm3,%xmm2
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,245,32,0,0 ; mulps 0x20f5(%rip),%xmm2 # 4a90 <_sk_callback_sse2+0xb8f>
- DB 102,15,219,29,253,32,0,0 ; pand 0x20fd(%rip),%xmm3 # 4aa0 <_sk_callback_sse2+0xb9f>
+ DB 15,89,21,253,33,0,0 ; mulps 0x21fd(%rip),%xmm2 # 5480 <_sk_callback_sse2+0xc97>
+ DB 102,15,219,29,5,34,0,0 ; pand 0x2205(%rip),%xmm3 # 5490 <_sk_callback_sse2+0xca7>
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,3,33,0,0 ; mulps 0x2103(%rip),%xmm3 # 4ab0 <_sk_callback_sse2+0xbaf>
+ DB 15,89,29,11,34,0,0 ; mulps 0x220b(%rip),%xmm3 # 54a0 <_sk_callback_sse2+0xcb7>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -18176,7 +20228,7 @@ 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 68,15,40,5,2,33,0,0 ; movaps 0x2102(%rip),%xmm8 # 4ac0 <_sk_callback_sse2+0xbbf>
+ DB 68,15,40,5,10,34,0,0 ; movaps 0x220a(%rip),%xmm8 # 54b0 <_sk_callback_sse2+0xcc7>
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
@@ -18206,11 +20258,11 @@ _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 15,40,21,149,32,0,0 ; movaps 0x2095(%rip),%xmm2 # 4ad0 <_sk_callback_sse2+0xbcf>
+ DB 15,40,21,157,33,0,0 ; movaps 0x219d(%rip),%xmm2 # 54c0 <_sk_callback_sse2+0xcd7>
DB 65,15,40,193 ; movaps %xmm9,%xmm0
DB 15,84,194 ; andps %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,147,32,0,0 ; movaps 0x2093(%rip),%xmm8 # 4ae0 <_sk_callback_sse2+0xbdf>
+ DB 68,15,40,5,155,33,0,0 ; movaps 0x219b(%rip),%xmm8 # 54d0 <_sk_callback_sse2+0xce7>
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
@@ -18257,11 +20309,11 @@ _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 102,15,111,21,228,31,0,0 ; movdqa 0x1fe4(%rip),%xmm2 # 4af0 <_sk_callback_sse2+0xbef>
+ DB 102,15,111,21,236,32,0,0 ; movdqa 0x20ec(%rip),%xmm2 # 54e0 <_sk_callback_sse2+0xcf7>
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 68,15,40,5,224,31,0,0 ; movaps 0x1fe0(%rip),%xmm8 # 4b00 <_sk_callback_sse2+0xbff>
+ DB 68,15,40,5,232,32,0,0 ; movaps 0x20e8(%rip),%xmm8 # 54f0 <_sk_callback_sse2+0xd07>
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
@@ -18283,7 +20335,7 @@ 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 68,15,40,5,163,31,0,0 ; movaps 0x1fa3(%rip),%xmm8 # 4b10 <_sk_callback_sse2+0xc0f>
+ DB 68,15,40,5,171,32,0,0 ; movaps 0x20ab(%rip),%xmm8 # 5500 <_sk_callback_sse2+0xd17>
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
@@ -18320,7 +20372,7 @@ _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 102,68,15,111,13,19,31,0,0 ; movdqa 0x1f13(%rip),%xmm9 # 4b20 <_sk_callback_sse2+0xc1f>
+ DB 102,68,15,111,13,27,32,0,0 ; movdqa 0x201b(%rip),%xmm9 # 5510 <_sk_callback_sse2+0xd27>
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
@@ -18328,11 +20380,11 @@ _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 102,68,15,111,29,248,30,0,0 ; movdqa 0x1ef8(%rip),%xmm11 # 4b30 <_sk_callback_sse2+0xc2f>
+ DB 102,68,15,111,29,0,32,0,0 ; movdqa 0x2000(%rip),%xmm11 # 5520 <_sk_callback_sse2+0xd37>
DB 102,69,15,254,235 ; paddd %xmm11,%xmm13
- DB 102,68,15,111,37,250,30,0,0 ; movdqa 0x1efa(%rip),%xmm12 # 4b40 <_sk_callback_sse2+0xc3f>
+ DB 102,68,15,111,37,2,32,0,0 ; movdqa 0x2002(%rip),%xmm12 # 5530 <_sk_callback_sse2+0xd47>
DB 102,65,15,239,204 ; pxor %xmm12,%xmm1
- DB 102,15,111,29,253,30,0,0 ; movdqa 0x1efd(%rip),%xmm3 # 4b50 <_sk_callback_sse2+0xc4f>
+ DB 102,15,111,29,5,32,0,0 ; movdqa 0x2005(%rip),%xmm3 # 5540 <_sk_callback_sse2+0xd57>
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
@@ -18416,7 +20468,7 @@ _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 102,68,15,111,13,139,29,0,0 ; movdqa 0x1d8b(%rip),%xmm9 # 4b60 <_sk_callback_sse2+0xc5f>
+ DB 102,68,15,111,13,147,30,0,0 ; movdqa 0x1e93(%rip),%xmm9 # 5550 <_sk_callback_sse2+0xd67>
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
@@ -18424,11 +20476,11 @@ _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 102,68,15,111,29,112,29,0,0 ; movdqa 0x1d70(%rip),%xmm11 # 4b70 <_sk_callback_sse2+0xc6f>
+ DB 102,68,15,111,29,120,30,0,0 ; movdqa 0x1e78(%rip),%xmm11 # 5560 <_sk_callback_sse2+0xd77>
DB 102,69,15,254,235 ; paddd %xmm11,%xmm13
- DB 102,68,15,111,37,114,29,0,0 ; movdqa 0x1d72(%rip),%xmm12 # 4b80 <_sk_callback_sse2+0xc7f>
+ DB 102,68,15,111,37,122,30,0,0 ; movdqa 0x1e7a(%rip),%xmm12 # 5570 <_sk_callback_sse2+0xd87>
DB 102,65,15,239,204 ; pxor %xmm12,%xmm1
- DB 102,15,111,29,117,29,0,0 ; movdqa 0x1d75(%rip),%xmm3 # 4b90 <_sk_callback_sse2+0xc8f>
+ DB 102,15,111,29,125,30,0,0 ; movdqa 0x1e7d(%rip),%xmm3 # 5580 <_sk_callback_sse2+0xd97>
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
@@ -18479,17 +20531,17 @@ 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 102,68,15,111,21,157,28,0,0 ; movdqa 0x1c9d(%rip),%xmm10 # 4ba0 <_sk_callback_sse2+0xc9f>
+ DB 102,68,15,111,21,165,29,0,0 ; movdqa 0x1da5(%rip),%xmm10 # 5590 <_sk_callback_sse2+0xda7>
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 102,68,15,111,13,144,28,0,0 ; movdqa 0x1c90(%rip),%xmm9 # 4bb0 <_sk_callback_sse2+0xcaf>
+ DB 102,68,15,111,13,152,29,0,0 ; movdqa 0x1d98(%rip),%xmm9 # 55a0 <_sk_callback_sse2+0xdb7>
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 102,68,15,111,29,129,28,0,0 ; movdqa 0x1c81(%rip),%xmm11 # 4bc0 <_sk_callback_sse2+0xcbf>
+ DB 102,68,15,111,29,137,29,0,0 ; movdqa 0x1d89(%rip),%xmm11 # 55b0 <_sk_callback_sse2+0xdc7>
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
@@ -18566,7 +20618,7 @@ _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 68,15,40,5,31,27,0,0 ; movaps 0x1b1f(%rip),%xmm8 # 4bd0 <_sk_callback_sse2+0xccf>
+ DB 68,15,40,5,39,28,0,0 ; movaps 0x1c27(%rip),%xmm8 # 55c0 <_sk_callback_sse2+0xdd7>
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
@@ -18617,7 +20669,7 @@ _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 68,15,40,13,91,26,0,0 ; movaps 0x1a5b(%rip),%xmm9 # 4be0 <_sk_callback_sse2+0xcdf>
+ DB 68,15,40,13,99,27,0,0 ; movaps 0x1b63(%rip),%xmm9 # 55d0 <_sk_callback_sse2+0xde7>
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
@@ -18634,14 +20686,14 @@ _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 15,40,29,34,26,0,0 ; movaps 0x1a22(%rip),%xmm3 # 4bf0 <_sk_callback_sse2+0xcef>
+ DB 15,40,29,42,27,0,0 ; movaps 0x1b2a(%rip),%xmm3 # 55e0 <_sk_callback_sse2+0xdf7>
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 68,15,40,13,35,26,0,0 ; movaps 0x1a23(%rip),%xmm9 # 4c00 <_sk_callback_sse2+0xcff>
+ DB 68,15,40,13,43,27,0,0 ; movaps 0x1b2b(%rip),%xmm9 # 55f0 <_sk_callback_sse2+0xe07>
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
@@ -18781,7 +20833,7 @@ _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 68,15,84,13,13,24,0,0 ; andps 0x180d(%rip),%xmm9 # 4c10 <_sk_callback_sse2+0xd0f>
+ DB 68,15,84,13,21,25,0,0 ; andps 0x1915(%rip),%xmm9 # 5600 <_sk_callback_sse2+0xe17>
DB 69,15,92,209 ; subps %xmm9,%xmm10
DB 69,15,89,208 ; mulps %xmm8,%xmm10
DB 65,15,92,194 ; subps %xmm10,%xmm0
@@ -18801,7 +20853,7 @@ _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 68,15,84,13,213,23,0,0 ; andps 0x17d5(%rip),%xmm9 # 4c20 <_sk_callback_sse2+0xd1f>
+ DB 68,15,84,13,221,24,0,0 ; andps 0x18dd(%rip),%xmm9 # 5610 <_sk_callback_sse2+0xe27>
DB 69,15,92,209 ; subps %xmm9,%xmm10
DB 69,15,89,208 ; mulps %xmm8,%xmm10
DB 65,15,92,202 ; subps %xmm10,%xmm1
@@ -18825,7 +20877,7 @@ _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 68,15,84,21,139,23,0,0 ; andps 0x178b(%rip),%xmm10 # 4c30 <_sk_callback_sse2+0xd2f>
+ DB 68,15,84,21,147,24,0,0 ; andps 0x1893(%rip),%xmm10 # 5620 <_sk_callback_sse2+0xe37>
DB 69,15,87,228 ; xorps %xmm12,%xmm12
DB 69,15,92,218 ; subps %xmm10,%xmm11
DB 69,15,89,216 ; mulps %xmm8,%xmm11
@@ -18853,7 +20905,7 @@ _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 68,15,84,21,49,23,0,0 ; andps 0x1731(%rip),%xmm10 # 4c40 <_sk_callback_sse2+0xd3f>
+ DB 68,15,84,21,57,24,0,0 ; andps 0x1839(%rip),%xmm10 # 5630 <_sk_callback_sse2+0xe47>
DB 69,15,87,228 ; xorps %xmm12,%xmm12
DB 69,15,92,218 ; subps %xmm10,%xmm11
DB 69,15,89,216 ; mulps %xmm8,%xmm11
@@ -18870,10 +20922,10 @@ _sk_mirror_y_sse2 LABEL PROC
PUBLIC _sk_luminance_to_alpha_sse2
_sk_luminance_to_alpha_sse2 LABEL PROC
DB 15,40,218 ; movaps %xmm2,%xmm3
- DB 15,89,5,9,23,0,0 ; mulps 0x1709(%rip),%xmm0 # 4c50 <_sk_callback_sse2+0xd4f>
- DB 15,89,13,18,23,0,0 ; mulps 0x1712(%rip),%xmm1 # 4c60 <_sk_callback_sse2+0xd5f>
+ DB 15,89,5,17,24,0,0 ; mulps 0x1811(%rip),%xmm0 # 5640 <_sk_callback_sse2+0xe57>
+ DB 15,89,13,26,24,0,0 ; mulps 0x181a(%rip),%xmm1 # 5650 <_sk_callback_sse2+0xe67>
DB 15,88,200 ; addps %xmm0,%xmm1
- DB 15,89,29,24,23,0,0 ; mulps 0x1718(%rip),%xmm3 # 4c70 <_sk_callback_sse2+0xd6f>
+ DB 15,89,29,32,24,0,0 ; mulps 0x1820(%rip),%xmm3 # 5660 <_sk_callback_sse2+0xe77>
DB 15,88,217 ; addps %xmm1,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
@@ -19096,7 +21148,7 @@ _sk_linear_gradient_sse2 LABEL PROC
DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12
DB 72,139,8 ; mov (%rax),%rcx
DB 72,133,201 ; test %rcx,%rcx
- DB 15,132,15,1,0,0 ; je 3a0b <_sk_linear_gradient_sse2+0x149>
+ DB 15,132,15,1,0,0 ; je 42f3 <_sk_linear_gradient_sse2+0x149>
DB 72,139,64,8 ; mov 0x8(%rax),%rax
DB 72,131,192,32 ; add $0x20,%rax
DB 69,15,87,192 ; xorps %xmm8,%xmm8
@@ -19157,8 +21209,8 @@ _sk_linear_gradient_sse2 LABEL PROC
DB 69,15,86,231 ; orps %xmm15,%xmm12
DB 72,131,192,36 ; add $0x24,%rax
DB 72,255,201 ; dec %rcx
- DB 15,133,8,255,255,255 ; jne 3911 <_sk_linear_gradient_sse2+0x4f>
- DB 235,13 ; jmp 3a18 <_sk_linear_gradient_sse2+0x156>
+ DB 15,133,8,255,255,255 ; jne 41f9 <_sk_linear_gradient_sse2+0x4f>
+ DB 235,13 ; jmp 4300 <_sk_linear_gradient_sse2+0x156>
DB 15,87,201 ; xorps %xmm1,%xmm1
DB 15,87,210 ; xorps %xmm2,%xmm2
DB 15,87,219 ; xorps %xmm3,%xmm3
@@ -19223,29 +21275,29 @@ _sk_xy_to_polar_unit_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 68,15,40,45,144,17,0,0 ; movaps 0x1190(%rip),%xmm13 # 4c80 <_sk_callback_sse2+0xd7f>
+ DB 68,15,40,45,152,18,0,0 ; movaps 0x1298(%rip),%xmm13 # 5670 <_sk_callback_sse2+0xe87>
DB 69,15,89,236 ; mulps %xmm12,%xmm13
- DB 68,15,88,45,148,17,0,0 ; addps 0x1194(%rip),%xmm13 # 4c90 <_sk_callback_sse2+0xd8f>
+ DB 68,15,88,45,156,18,0,0 ; addps 0x129c(%rip),%xmm13 # 5680 <_sk_callback_sse2+0xe97>
DB 69,15,89,236 ; mulps %xmm12,%xmm13
- DB 68,15,88,45,152,17,0,0 ; addps 0x1198(%rip),%xmm13 # 4ca0 <_sk_callback_sse2+0xd9f>
+ DB 68,15,88,45,160,18,0,0 ; addps 0x12a0(%rip),%xmm13 # 5690 <_sk_callback_sse2+0xea7>
DB 69,15,89,236 ; mulps %xmm12,%xmm13
- DB 68,15,88,45,156,17,0,0 ; addps 0x119c(%rip),%xmm13 # 4cb0 <_sk_callback_sse2+0xdaf>
+ DB 68,15,88,45,164,18,0,0 ; addps 0x12a4(%rip),%xmm13 # 56a0 <_sk_callback_sse2+0xeb7>
DB 69,15,89,235 ; mulps %xmm11,%xmm13
DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9
- DB 68,15,40,21,155,17,0,0 ; movaps 0x119b(%rip),%xmm10 # 4cc0 <_sk_callback_sse2+0xdbf>
+ DB 68,15,40,21,163,18,0,0 ; movaps 0x12a3(%rip),%xmm10 # 56b0 <_sk_callback_sse2+0xec7>
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 68,15,40,21,142,17,0,0 ; movaps 0x118e(%rip),%xmm10 # 4cd0 <_sk_callback_sse2+0xdcf>
+ DB 68,15,40,21,150,18,0,0 ; movaps 0x1296(%rip),%xmm10 # 56c0 <_sk_callback_sse2+0xed7>
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 68,15,40,21,125,17,0,0 ; movaps 0x117d(%rip),%xmm10 # 4ce0 <_sk_callback_sse2+0xddf>
+ DB 68,15,40,21,133,18,0,0 ; movaps 0x1285(%rip),%xmm10 # 56d0 <_sk_callback_sse2+0xee7>
DB 69,15,92,208 ; subps %xmm8,%xmm10
DB 69,15,84,209 ; andps %xmm9,%xmm10
DB 69,15,85,200 ; andnps %xmm8,%xmm9
@@ -19258,7 +21310,7 @@ _sk_xy_to_polar_unit_sse2 LABEL PROC
PUBLIC _sk_save_xy_sse2
_sk_save_xy_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 68,15,40,5,102,17,0,0 ; movaps 0x1166(%rip),%xmm8 # 4cf0 <_sk_callback_sse2+0xdef>
+ DB 68,15,40,5,110,18,0,0 ; movaps 0x126e(%rip),%xmm8 # 56e0 <_sk_callback_sse2+0xef7>
DB 15,17,0 ; movups %xmm0,(%rax)
DB 68,15,40,200 ; movaps %xmm0,%xmm9
DB 69,15,88,200 ; addps %xmm8,%xmm9
@@ -19266,7 +21318,7 @@ _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 68,15,40,37,81,17,0,0 ; movaps 0x1151(%rip),%xmm12 # 4d00 <_sk_callback_sse2+0xdff>
+ DB 68,15,40,37,89,18,0,0 ; movaps 0x1259(%rip),%xmm12 # 56f0 <_sk_callback_sse2+0xf07>
DB 69,15,84,220 ; andps %xmm12,%xmm11
DB 69,15,92,211 ; subps %xmm11,%xmm10
DB 69,15,92,202 ; subps %xmm10,%xmm9
@@ -19309,8 +21361,8 @@ _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 15,88,5,202,16,0,0 ; addps 0x10ca(%rip),%xmm0 # 4d10 <_sk_callback_sse2+0xe0f>
- DB 68,15,40,13,210,16,0,0 ; movaps 0x10d2(%rip),%xmm9 # 4d20 <_sk_callback_sse2+0xe1f>
+ DB 15,88,5,210,17,0,0 ; addps 0x11d2(%rip),%xmm0 # 5700 <_sk_callback_sse2+0xf17>
+ DB 68,15,40,13,218,17,0,0 ; movaps 0x11da(%rip),%xmm9 # 5710 <_sk_callback_sse2+0xf27>
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
@@ -19321,7 +21373,7 @@ _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 15,88,5,193,16,0,0 ; addps 0x10c1(%rip),%xmm0 # 4d30 <_sk_callback_sse2+0xe2f>
+ DB 15,88,5,201,17,0,0 ; addps 0x11c9(%rip),%xmm0 # 5720 <_sk_callback_sse2+0xf37>
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
@@ -19331,8 +21383,8 @@ _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 15,88,13,179,16,0,0 ; addps 0x10b3(%rip),%xmm1 # 4d40 <_sk_callback_sse2+0xe3f>
- DB 68,15,40,13,187,16,0,0 ; movaps 0x10bb(%rip),%xmm9 # 4d50 <_sk_callback_sse2+0xe4f>
+ DB 15,88,13,187,17,0,0 ; addps 0x11bb(%rip),%xmm1 # 5730 <_sk_callback_sse2+0xf47>
+ DB 68,15,40,13,195,17,0,0 ; movaps 0x11c3(%rip),%xmm9 # 5740 <_sk_callback_sse2+0xf57>
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
@@ -19343,7 +21395,7 @@ _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 15,88,13,169,16,0,0 ; addps 0x10a9(%rip),%xmm1 # 4d60 <_sk_callback_sse2+0xe5f>
+ DB 15,88,13,177,17,0,0 ; addps 0x11b1(%rip),%xmm1 # 5750 <_sk_callback_sse2+0xf67>
DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -19353,13 +21405,13 @@ _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 15,88,5,156,16,0,0 ; addps 0x109c(%rip),%xmm0 # 4d70 <_sk_callback_sse2+0xe6f>
- DB 68,15,40,13,164,16,0,0 ; movaps 0x10a4(%rip),%xmm9 # 4d80 <_sk_callback_sse2+0xe7f>
+ DB 15,88,5,164,17,0,0 ; addps 0x11a4(%rip),%xmm0 # 5760 <_sk_callback_sse2+0xf77>
+ DB 68,15,40,13,172,17,0,0 ; movaps 0x11ac(%rip),%xmm9 # 5770 <_sk_callback_sse2+0xf87>
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 68,15,89,13,160,16,0,0 ; mulps 0x10a0(%rip),%xmm9 # 4d90 <_sk_callback_sse2+0xe8f>
- DB 68,15,88,13,168,16,0,0 ; addps 0x10a8(%rip),%xmm9 # 4da0 <_sk_callback_sse2+0xe9f>
+ DB 68,15,89,13,168,17,0,0 ; mulps 0x11a8(%rip),%xmm9 # 5780 <_sk_callback_sse2+0xf97>
+ DB 68,15,88,13,176,17,0,0 ; addps 0x11b0(%rip),%xmm9 # 5790 <_sk_callback_sse2+0xfa7>
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
@@ -19370,16 +21422,16 @@ _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 15,88,5,151,16,0,0 ; addps 0x1097(%rip),%xmm0 # 4db0 <_sk_callback_sse2+0xeaf>
- DB 68,15,40,13,159,16,0,0 ; movaps 0x109f(%rip),%xmm9 # 4dc0 <_sk_callback_sse2+0xebf>
+ DB 15,88,5,159,17,0,0 ; addps 0x119f(%rip),%xmm0 # 57a0 <_sk_callback_sse2+0xfb7>
+ DB 68,15,40,13,167,17,0,0 ; movaps 0x11a7(%rip),%xmm9 # 57b0 <_sk_callback_sse2+0xfc7>
DB 69,15,92,200 ; subps %xmm8,%xmm9
- DB 68,15,40,5,163,16,0,0 ; movaps 0x10a3(%rip),%xmm8 # 4dd0 <_sk_callback_sse2+0xecf>
+ DB 68,15,40,5,171,17,0,0 ; movaps 0x11ab(%rip),%xmm8 # 57c0 <_sk_callback_sse2+0xfd7>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,167,16,0,0 ; addps 0x10a7(%rip),%xmm8 # 4de0 <_sk_callback_sse2+0xedf>
+ DB 68,15,88,5,175,17,0,0 ; addps 0x11af(%rip),%xmm8 # 57d0 <_sk_callback_sse2+0xfe7>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,171,16,0,0 ; addps 0x10ab(%rip),%xmm8 # 4df0 <_sk_callback_sse2+0xeef>
+ DB 68,15,88,5,179,17,0,0 ; addps 0x11b3(%rip),%xmm8 # 57e0 <_sk_callback_sse2+0xff7>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,175,16,0,0 ; addps 0x10af(%rip),%xmm8 # 4e00 <_sk_callback_sse2+0xeff>
+ DB 68,15,88,5,183,17,0,0 ; addps 0x11b7(%rip),%xmm8 # 57f0 <_sk_callback_sse2+0x1007>
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
@@ -19387,17 +21439,17 @@ _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 68,15,40,5,169,16,0,0 ; movaps 0x10a9(%rip),%xmm8 # 4e10 <_sk_callback_sse2+0xf0f>
+ DB 68,15,40,5,177,17,0,0 ; movaps 0x11b1(%rip),%xmm8 # 5800 <_sk_callback_sse2+0x1017>
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 68,15,40,21,165,16,0,0 ; movaps 0x10a5(%rip),%xmm10 # 4e20 <_sk_callback_sse2+0xf1f>
+ DB 68,15,40,21,173,17,0,0 ; movaps 0x11ad(%rip),%xmm10 # 5810 <_sk_callback_sse2+0x1027>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,169,16,0,0 ; addps 0x10a9(%rip),%xmm10 # 4e30 <_sk_callback_sse2+0xf2f>
+ DB 68,15,88,21,177,17,0,0 ; addps 0x11b1(%rip),%xmm10 # 5820 <_sk_callback_sse2+0x1037>
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 68,15,88,21,165,16,0,0 ; addps 0x10a5(%rip),%xmm10 # 4e40 <_sk_callback_sse2+0xf3f>
+ DB 68,15,88,21,173,17,0,0 ; addps 0x11ad(%rip),%xmm10 # 5830 <_sk_callback_sse2+0x1047>
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
@@ -19407,11 +21459,11 @@ _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 15,88,5,152,16,0,0 ; addps 0x1098(%rip),%xmm0 # 4e50 <_sk_callback_sse2+0xf4f>
+ DB 15,88,5,160,17,0,0 ; addps 0x11a0(%rip),%xmm0 # 5840 <_sk_callback_sse2+0x1057>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,89,201 ; mulps %xmm9,%xmm9
- DB 68,15,89,5,152,16,0,0 ; mulps 0x1098(%rip),%xmm8 # 4e60 <_sk_callback_sse2+0xf5f>
- DB 68,15,88,5,160,16,0,0 ; addps 0x10a0(%rip),%xmm8 # 4e70 <_sk_callback_sse2+0xf6f>
+ DB 68,15,89,5,160,17,0,0 ; mulps 0x11a0(%rip),%xmm8 # 5850 <_sk_callback_sse2+0x1067>
+ DB 68,15,88,5,168,17,0,0 ; addps 0x11a8(%rip),%xmm8 # 5860 <_sk_callback_sse2+0x1077>
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
@@ -19422,13 +21474,13 @@ _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 15,88,13,142,16,0,0 ; addps 0x108e(%rip),%xmm1 # 4e80 <_sk_callback_sse2+0xf7f>
- DB 68,15,40,13,150,16,0,0 ; movaps 0x1096(%rip),%xmm9 # 4e90 <_sk_callback_sse2+0xf8f>
+ DB 15,88,13,150,17,0,0 ; addps 0x1196(%rip),%xmm1 # 5870 <_sk_callback_sse2+0x1087>
+ DB 68,15,40,13,158,17,0,0 ; movaps 0x119e(%rip),%xmm9 # 5880 <_sk_callback_sse2+0x1097>
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 68,15,89,13,146,16,0,0 ; mulps 0x1092(%rip),%xmm9 # 4ea0 <_sk_callback_sse2+0xf9f>
- DB 68,15,88,13,154,16,0,0 ; addps 0x109a(%rip),%xmm9 # 4eb0 <_sk_callback_sse2+0xfaf>
+ DB 68,15,89,13,154,17,0,0 ; mulps 0x119a(%rip),%xmm9 # 5890 <_sk_callback_sse2+0x10a7>
+ DB 68,15,88,13,162,17,0,0 ; addps 0x11a2(%rip),%xmm9 # 58a0 <_sk_callback_sse2+0x10b7>
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
@@ -19439,16 +21491,16 @@ _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 15,88,13,136,16,0,0 ; addps 0x1088(%rip),%xmm1 # 4ec0 <_sk_callback_sse2+0xfbf>
- DB 68,15,40,13,144,16,0,0 ; movaps 0x1090(%rip),%xmm9 # 4ed0 <_sk_callback_sse2+0xfcf>
+ DB 15,88,13,144,17,0,0 ; addps 0x1190(%rip),%xmm1 # 58b0 <_sk_callback_sse2+0x10c7>
+ DB 68,15,40,13,152,17,0,0 ; movaps 0x1198(%rip),%xmm9 # 58c0 <_sk_callback_sse2+0x10d7>
DB 69,15,92,200 ; subps %xmm8,%xmm9
- DB 68,15,40,5,148,16,0,0 ; movaps 0x1094(%rip),%xmm8 # 4ee0 <_sk_callback_sse2+0xfdf>
+ DB 68,15,40,5,156,17,0,0 ; movaps 0x119c(%rip),%xmm8 # 58d0 <_sk_callback_sse2+0x10e7>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,152,16,0,0 ; addps 0x1098(%rip),%xmm8 # 4ef0 <_sk_callback_sse2+0xfef>
+ DB 68,15,88,5,160,17,0,0 ; addps 0x11a0(%rip),%xmm8 # 58e0 <_sk_callback_sse2+0x10f7>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,156,16,0,0 ; addps 0x109c(%rip),%xmm8 # 4f00 <_sk_callback_sse2+0xfff>
+ DB 68,15,88,5,164,17,0,0 ; addps 0x11a4(%rip),%xmm8 # 58f0 <_sk_callback_sse2+0x1107>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,160,16,0,0 ; addps 0x10a0(%rip),%xmm8 # 4f10 <_sk_callback_sse2+0x100f>
+ DB 68,15,88,5,168,17,0,0 ; addps 0x11a8(%rip),%xmm8 # 5900 <_sk_callback_sse2+0x1117>
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
@@ -19456,17 +21508,17 @@ _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 68,15,40,5,154,16,0,0 ; movaps 0x109a(%rip),%xmm8 # 4f20 <_sk_callback_sse2+0x101f>
+ DB 68,15,40,5,162,17,0,0 ; movaps 0x11a2(%rip),%xmm8 # 5910 <_sk_callback_sse2+0x1127>
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 68,15,40,21,149,16,0,0 ; movaps 0x1095(%rip),%xmm10 # 4f30 <_sk_callback_sse2+0x102f>
+ DB 68,15,40,21,157,17,0,0 ; movaps 0x119d(%rip),%xmm10 # 5920 <_sk_callback_sse2+0x1137>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,153,16,0,0 ; addps 0x1099(%rip),%xmm10 # 4f40 <_sk_callback_sse2+0x103f>
+ DB 68,15,88,21,161,17,0,0 ; addps 0x11a1(%rip),%xmm10 # 5930 <_sk_callback_sse2+0x1147>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
DB 69,15,88,208 ; addps %xmm8,%xmm10
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,149,16,0,0 ; addps 0x1095(%rip),%xmm10 # 4f50 <_sk_callback_sse2+0x104f>
+ DB 68,15,88,21,157,17,0,0 ; addps 0x119d(%rip),%xmm10 # 5940 <_sk_callback_sse2+0x1157>
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
@@ -19476,11 +21528,11 @@ _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 15,88,13,135,16,0,0 ; addps 0x1087(%rip),%xmm1 # 4f60 <_sk_callback_sse2+0x105f>
+ DB 15,88,13,143,17,0,0 ; addps 0x118f(%rip),%xmm1 # 5950 <_sk_callback_sse2+0x1167>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,89,201 ; mulps %xmm9,%xmm9
- DB 68,15,89,5,135,16,0,0 ; mulps 0x1087(%rip),%xmm8 # 4f70 <_sk_callback_sse2+0x106f>
- DB 68,15,88,5,143,16,0,0 ; addps 0x108f(%rip),%xmm8 # 4f80 <_sk_callback_sse2+0x107f>
+ DB 68,15,89,5,143,17,0,0 ; mulps 0x118f(%rip),%xmm8 # 5960 <_sk_callback_sse2+0x1177>
+ DB 68,15,88,5,151,17,0,0 ; addps 0x1197(%rip),%xmm8 # 5970 <_sk_callback_sse2+0x1187>
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
@@ -19685,11 +21737,135 @@ 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 4188 <.literal16+0x1d8>
+ DB 224,64 ; loopne 4a78 <.literal16+0x1d8>
+ DB 0,0 ; add %al,(%rax)
+ DB 224,64 ; loopne 4a7c <.literal16+0x1dc>
DB 0,0 ; add %al,(%rax)
- DB 224,64 ; loopne 418c <.literal16+0x1dc>
+ DB 224,64 ; loopne 4a80 <.literal16+0x1e0>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4ab1 <.literal16+0x211>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4ab5 <.literal16+0x215>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4ab9 <.literal16+0x219>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4abd <.literal16+0x21d>
DB 0,0 ; add %al,(%rax)
- DB 224,64 ; loopne 4190 <.literal16+0x1e0>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4af1 <.literal16+0x251>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4af5 <.literal16+0x255>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4af9 <.literal16+0x259>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4afd <.literal16+0x25d>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b31 <.literal16+0x291>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b35 <.literal16+0x295>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b39 <.literal16+0x299>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b3d <.literal16+0x29d>
+ 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 128,63,154 ; cmpb $0x9a,(%rdi)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,154 ; ds (bad)
+ DB 153 ; cltd
+ DB 153 ; cltd
+ DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 61,10,23,63,61 ; cmp $0x3d3f170a,%eax
+ DB 10,23 ; or (%rdi),%dl
+ DB 63 ; (bad)
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b71 <.literal16+0x2d1>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b75 <.literal16+0x2d5>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b79 <.literal16+0x2d9>
+ DB 174 ; scas %es:(%rdi),%al
+ DB 71,225,61 ; rex.RXB loope 4b7d <.literal16+0x2dd>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax)
@@ -19913,13 +22089,13 @@ ALIGN 16
DB 132,55 ; test %dh,(%rdi)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4359 <.literal16+0x3a9>
+ DB 224,7 ; loopne 4d49 <.literal16+0x4a9>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 435d <.literal16+0x3ad>
+ DB 224,7 ; loopne 4d4d <.literal16+0x4ad>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 4361 <.literal16+0x3b1>
+ DB 224,7 ; loopne 4d51 <.literal16+0x4b1>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 4365 <.literal16+0x3b5>
+ DB 224,7 ; loopne 4d55 <.literal16+0x4b5>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -19988,11 +22164,11 @@ 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 444b <.literal16+0x49b>
+ DB 127,67 ; jg 4e3b <.literal16+0x59b>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 444f <.literal16+0x49f>
+ DB 127,67 ; jg 4e3f <.literal16+0x59f>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4453 <.literal16+0x4a3>
+ DB 127,67 ; jg 4e43 <.literal16+0x5a3>
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)
@@ -20007,16 +22183,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4444 <.literal16+0x494>
+ DB 127,0 ; jg 4e34 <.literal16+0x594>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4448 <.literal16+0x498>
+ DB 127,0 ; jg 4e38 <.literal16+0x598>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 444c <.literal16+0x49c>
+ DB 127,0 ; jg 4e3c <.literal16+0x59c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4450 <.literal16+0x4a0>
+ DB 127,0 ; jg 4e40 <.literal16+0x5a0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -20025,7 +22201,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 44d5 <.literal16+0x525>
+ DB 119,115 ; ja 4ec5 <.literal16+0x625>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -20036,7 +22212,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4439 <.literal16+0x489>
+ DB 117,191 ; jne 4e29 <.literal16+0x589>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -20048,7 +22224,7 @@ 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 ffffffffe9a3847a <_sk_callback_sse2+0xffffffffe9a34579>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a38e6a <_sk_callback_sse2+0xffffffffe9a34681>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -20102,16 +22278,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4514 <.literal16+0x564>
+ DB 127,0 ; jg 4f04 <.literal16+0x664>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4518 <.literal16+0x568>
+ DB 127,0 ; jg 4f08 <.literal16+0x668>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 451c <.literal16+0x56c>
+ DB 127,0 ; jg 4f0c <.literal16+0x66c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4520 <.literal16+0x570>
+ DB 127,0 ; jg 4f10 <.literal16+0x670>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -20120,7 +22296,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 45a5 <.literal16+0x5f5>
+ DB 119,115 ; ja 4f95 <.literal16+0x6f5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -20131,7 +22307,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4509 <.literal16+0x559>
+ DB 117,191 ; jne 4ef9 <.literal16+0x659>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -20143,7 +22319,7 @@ 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 ffffffffe9a3854a <_sk_callback_sse2+0xffffffffe9a34649>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a38f3a <_sk_callback_sse2+0xffffffffe9a34751>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -20197,16 +22373,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 45e4 <.literal16+0x634>
+ DB 127,0 ; jg 4fd4 <.literal16+0x734>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 45e8 <.literal16+0x638>
+ DB 127,0 ; jg 4fd8 <.literal16+0x738>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 45ec <.literal16+0x63c>
+ DB 127,0 ; jg 4fdc <.literal16+0x73c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 45f0 <.literal16+0x640>
+ DB 127,0 ; jg 4fe0 <.literal16+0x740>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -20215,7 +22391,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4675 <.literal16+0x6c5>
+ DB 119,115 ; ja 5065 <.literal16+0x7c5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -20226,7 +22402,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 45d9 <.literal16+0x629>
+ DB 117,191 ; jne 4fc9 <.literal16+0x729>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -20238,7 +22414,7 @@ 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 ffffffffe9a3861a <_sk_callback_sse2+0xffffffffe9a34719>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a3900a <_sk_callback_sse2+0xffffffffe9a34821>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -20292,16 +22468,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 46b4 <.literal16+0x704>
+ DB 127,0 ; jg 50a4 <.literal16+0x804>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 46b8 <.literal16+0x708>
+ DB 127,0 ; jg 50a8 <.literal16+0x808>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 46bc <.literal16+0x70c>
+ DB 127,0 ; jg 50ac <.literal16+0x80c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 46c0 <.literal16+0x710>
+ DB 127,0 ; jg 50b0 <.literal16+0x810>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -20310,7 +22486,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4745 <.literal16+0x795>
+ DB 119,115 ; ja 5135 <.literal16+0x895>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -20321,7 +22497,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 46a9 <.literal16+0x6f9>
+ DB 117,191 ; jne 5099 <.literal16+0x7f9>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -20333,7 +22509,7 @@ 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 ffffffffe9a386ea <_sk_callback_sse2+0xffffffffe9a347e9>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a390da <_sk_callback_sse2+0xffffffffe9a348f1>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -20383,13 +22559,13 @@ 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 47c7 <.literal16+0x817>
+ DB 127,67 ; jg 51b7 <.literal16+0x917>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 47cb <.literal16+0x81b>
+ DB 127,67 ; jg 51bb <.literal16+0x91b>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 47cf <.literal16+0x81f>
+ DB 127,67 ; jg 51bf <.literal16+0x91f>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 47d3 <.literal16+0x823>
+ DB 127,67 ; jg 51c3 <.literal16+0x923>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -20436,16 +22612,16 @@ ALIGN 16
DB 128,3,62 ; addb $0x3e,(%rbx)
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4853 <.literal16+0x8a3>
+ DB 118,63 ; jbe 5243 <.literal16+0x9a3>
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4857 <.literal16+0x8a7>
+ DB 118,63 ; jbe 5247 <.literal16+0x9a7>
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 485b <.literal16+0x8ab>
+ DB 118,63 ; jbe 524b <.literal16+0x9ab>
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 485f <.literal16+0x8af>
+ DB 118,63 ; jbe 524f <.literal16+0x9af>
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)
@@ -20457,11 +22633,11 @@ 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 489b <.literal16+0x8eb>
+ DB 127,67 ; jg 528b <.literal16+0x9eb>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 489f <.literal16+0x8ef>
+ DB 127,67 ; jg 528f <.literal16+0x9ef>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 48a3 <.literal16+0x8f3>
+ DB 127,67 ; jg 5293 <.literal16+0x9f3>
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)
@@ -20501,13 +22677,13 @@ ALIGN 16
DB 132,55 ; test %dh,(%rdi)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 48e9 <.literal16+0x939>
+ DB 224,7 ; loopne 52d9 <.literal16+0xa39>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 48ed <.literal16+0x93d>
+ DB 224,7 ; loopne 52dd <.literal16+0xa3d>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 48f1 <.literal16+0x941>
+ DB 224,7 ; loopne 52e1 <.literal16+0xa41>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 48f5 <.literal16+0x945>
+ DB 224,7 ; loopne 52e5 <.literal16+0xa45>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -20553,13 +22729,13 @@ ALIGN 16
DB 132,55 ; test %dh,(%rdi)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4959 <.literal16+0x9a9>
+ DB 224,7 ; loopne 5349 <.literal16+0xaa9>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 495d <.literal16+0x9ad>
+ DB 224,7 ; loopne 534d <.literal16+0xaad>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 4961 <.literal16+0x9b1>
+ DB 224,7 ; loopne 5351 <.literal16+0xab1>
DB 0,0 ; add %al,(%rax)
- DB 224,7 ; loopne 4965 <.literal16+0x9b5>
+ DB 224,7 ; loopne 5355 <.literal16+0xab5>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -20597,13 +22773,13 @@ ALIGN 16
DB 65,0,0 ; add %al,(%r8)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 49f6 <.literal16+0xa46>
+ DB 124,66 ; jl 53e6 <.literal16+0xb46>
DB 0,0 ; add %al,(%rax)
- DB 124,66 ; jl 49fa <.literal16+0xa4a>
+ DB 124,66 ; jl 53ea <.literal16+0xb4a>
DB 0,0 ; add %al,(%rax)
- DB 124,66 ; jl 49fe <.literal16+0xa4e>
+ DB 124,66 ; jl 53ee <.literal16+0xb4e>
DB 0,0 ; add %al,(%rax)
- DB 124,66 ; jl 4a02 <.literal16+0xa52>
+ DB 124,66 ; jl 53f2 <.literal16+0xb52>
DB 0,240 ; add %dh,%al
DB 0,0 ; add %al,(%rax)
DB 0,240 ; add %dh,%al
@@ -20693,13 +22869,13 @@ 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 4b05 <.literal16+0xb55>
+ DB 112,65 ; jo 54f5 <.literal16+0xc55>
DB 0,0 ; add %al,(%rax)
- DB 112,65 ; jo 4b09 <.literal16+0xb59>
+ DB 112,65 ; jo 54f9 <.literal16+0xc59>
DB 0,0 ; add %al,(%rax)
- DB 112,65 ; jo 4b0d <.literal16+0xb5d>
+ DB 112,65 ; jo 54fd <.literal16+0xc5d>
DB 0,0 ; add %al,(%rax)
- DB 112,65 ; jo 4b11 <.literal16+0xb61>
+ DB 112,65 ; jo 5501 <.literal16+0xc61>
DB 255,0 ; incl (%rax)
DB 0,0 ; add %al,(%rax)
DB 255,0 ; incl (%rax)
@@ -20721,11 +22897,11 @@ 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 4b5b <.literal16+0xbab>
+ DB 127,67 ; jg 554b <.literal16+0xcab>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4b5f <.literal16+0xbaf>
+ DB 127,67 ; jg 554f <.literal16+0xcaf>
DB 0,0 ; add %al,(%rax)
- DB 127,67 ; jg 4b63 <.literal16+0xbb3>
+ DB 127,67 ; jg 5553 <.literal16+0xcb3>
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)
@@ -20801,13 +22977,13 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 255 ; (bad)
- DB 127,71 ; jg 4c4b <.literal16+0xc9b>
+ DB 127,71 ; jg 563b <.literal16+0xd9b>
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 4c4f <.literal16+0xc9f>
+ DB 127,71 ; jg 563f <.literal16+0xd9f>
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 4c53 <.literal16+0xca3>
+ DB 127,71 ; jg 5643 <.literal16+0xda3>
DB 0,255 ; add %bh,%bh
- DB 127,71 ; jg 4c57 <.literal16+0xca7>
+ DB 127,71 ; jg 5647 <.literal16+0xda7>
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax)
@@ -20960,11 +23136,11 @@ 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 4db2 <.literal16+0xe02>
+ DB 62,114,28 ; jb,pt 57a2 <.literal16+0xf02>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4db6 <.literal16+0xe06>
+ DB 62,114,28 ; jb,pt 57a6 <.literal16+0xf06>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4dba <.literal16+0xe0a>
+ DB 62,114,28 ; jb,pt 57aa <.literal16+0xf0a>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -21008,7 +23184,7 @@ 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 # 3d63dc45 <_sk_callback_sse2+0x3d639d44>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e635 <_sk_callback_sse2+0x3d639e4c>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -21034,7 +23210,7 @@ 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 # 3d63dc85 <_sk_callback_sse2+0x3d639d84>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e675 <_sk_callback_sse2+0x3d639e8c>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
@@ -21043,13 +23219,13 @@ ALIGN 16
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
DB 63 ; (bad)
- DB 114,28 ; jb 4e7e <.literal16+0xece>
+ DB 114,28 ; jb 586e <.literal16+0xfce>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4e82 <.literal16+0xed2>
+ DB 62,114,28 ; jb,pt 5872 <.literal16+0xfd2>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4e86 <.literal16+0xed6>
+ DB 62,114,28 ; jb,pt 5876 <.literal16+0xfd6>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4e8a <.literal16+0xeda>
+ DB 62,114,28 ; jb,pt 587a <.literal16+0xfda>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -21070,11 +23246,11 @@ 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 4ec2 <.literal16+0xf12>
+ DB 62,114,28 ; jb,pt 58b2 <.literal16+0x1012>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4ec6 <.literal16+0xf16>
+ DB 62,114,28 ; jb,pt 58b6 <.literal16+0x1016>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4eca <.literal16+0xf1a>
+ DB 62,114,28 ; jb,pt 58ba <.literal16+0x101a>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -21118,7 +23294,7 @@ 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 # 3d63dd55 <_sk_callback_sse2+0x3d639e54>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e745 <_sk_callback_sse2+0x3d639f5c>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -21144,7 +23320,7 @@ 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 # 3d63dd95 <_sk_callback_sse2+0x3d639e94>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63e785 <_sk_callback_sse2+0x3d639f9c>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
@@ -21153,13 +23329,13 @@ ALIGN 16
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
DB 63 ; (bad)
- DB 114,28 ; jb 4f8e <.literal16+0xfde>
+ DB 114,28 ; jb 597e <.literal16+0x10de>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4f92 <_sk_callback_sse2+0x1091>
+ DB 62,114,28 ; jb,pt 5982 <_sk_callback_sse2+0x1199>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4f96 <_sk_callback_sse2+0x1095>
+ DB 62,114,28 ; jb,pt 5986 <_sk_callback_sse2+0x119d>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 4f9a <_sk_callback_sse2+0x1099>
+ DB 62,114,28 ; jb,pt 598a <_sk_callback_sse2+0x11a1>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp
index e1c245358e..e5b040a742 100644
--- a/src/jumper/SkJumper_stages.cpp
+++ b/src/jumper/SkJumper_stages.cpp
@@ -413,6 +413,109 @@ BLEND_MODE(softlight) {
}
#undef BLEND_MODE
+// We're basing our implemenation of non-separable blend modes on
+// https://www.w3.org/TR/compositing-1/#blendingnonseparable.
+// and
+// https://www.khronos.org/registry/OpenGL/specs/es/3.2/es_spec_3.2.pdf
+// They're equivalent, but ES' math has been better simplified.
+
+SI F max(F r, F g, F b) { return max(r, max(g, b)); }
+SI F min(F r, F g, F b) { return min(r, min(g, b)); }
+
+SI F sat(F r, F g, F b) { return max(r,g,b) - min(r,g,b); }
+SI F lum(F r, F g, F b) { return r*0.30f + g*0.59f + b*0.11f; }
+
+SI void set_sat(F* r, F* g, F* b, F s) {
+ F mn = min(*r,*g,*b),
+ mx = max(*r,*g,*b),
+ sat = mx - mn;
+
+ // Map min channel to 0, max channel to s, and scale the middle proportionally.
+ auto scale = [=](F c) {
+ return if_then_else(sat == 0, 0, (c - mn) * s / sat);
+ };
+ *r = scale(*r);
+ *g = scale(*g);
+ *b = scale(*b);
+}
+SI void clip_color(F* r, F* g, F* b) {
+ F mn = min(*r, *g, *b),
+ mx = max(*r, *g, *b),
+ l = lum(*r, *g, *b);
+
+ auto clip = [=](F c) {
+ c = if_then_else(mn >= 0, c, l + (c - l) * ( l) / (l - mn) );
+ c = if_then_else(mx > 1, l + (c - l) * (1 - l) / (mx - l), c);
+ c = max(c, 0); // Sometimes without this we may dip just a little negative.
+ return c;
+ };
+ *r = clip(*r);
+ *g = clip(*g);
+ *b = clip(*b);
+}
+SI void set_lum(F* r, F* g, F* b, F l) {
+ F diff = l - lum(*r, *g, *b);
+ *r += diff;
+ *g += diff;
+ *b += diff;
+ clip_color(r, g, b);
+}
+
+SI F unpremultiply(F c, F a) {
+ return c * if_then_else(a == 0, 0, 1.0f / a);
+}
+
+STAGE(hue) {
+ F R = unpremultiply(r,a),
+ G = unpremultiply(g,a),
+ B = unpremultiply(b,a);
+
+ set_sat(&R, &G, &B, sat(dr,dg,db));
+ set_lum(&R, &G, &B, lum(dr,dg,db));
+
+ a = a + da - a*da;
+ r = R * a;
+ g = G * a;
+ b = B * a;
+}
+STAGE(saturation) {
+ F R = unpremultiply(dr,da),
+ G = unpremultiply(dg,da),
+ B = unpremultiply(db,da);
+
+ set_sat(&R, &G, &B, sat( r, g, b));
+ set_lum(&R, &G, &B, lum(dr,dg,db)); // (This is not redundant.)
+
+ a = a + da - a*da;
+ r = R * a;
+ g = G * a;
+ b = B * a;
+}
+STAGE(color) {
+ F R = unpremultiply(r,a),
+ G = unpremultiply(g,a),
+ B = unpremultiply(b,a);
+
+ set_lum(&R, &G, &B, lum(dr,dg,db));
+
+ a = a + da - a*da;
+ r = R * a;
+ g = G * a;
+ b = B * a;
+}
+STAGE(luminosity) {
+ F R = unpremultiply(dr,da),
+ G = unpremultiply(dg,da),
+ B = unpremultiply(db,da);
+
+ set_lum(&R, &G, &B, lum(r,g,b));
+
+ a = a + da - a*da;
+ r = R * a;
+ g = G * a;
+ b = B * a;
+}
+
STAGE(clamp_0) {
r = max(r, 0);
g = max(g, 0);
@@ -476,10 +579,9 @@ STAGE(premul) {
b = b * a;
}
STAGE(unpremul) {
- auto scale = if_then_else(a == 0, 0, 1.0f / a);
- r = r * scale;
- g = g * scale;
- b = b * scale;
+ r = unpremultiply(r,a);
+ g = unpremultiply(g,a);
+ b = unpremultiply(b,a);
}
STAGE(from_srgb) {