aboutsummaryrefslogtreecommitdiffhomepage
diff options
context:
space:
mode:
authorGravatar Mike Klein <mtklein@chromium.org>2017-05-01 14:56:04 -0400
committerGravatar Skia Commit-Bot <skia-commit-bot@chromium.org>2017-05-01 19:30:46 +0000
commitb3665f01608dc65374b23f00240582494021a540 (patch)
treee539b3a7be113605093ba142b5e8d0b85b308422
parentfb11acdeefce2fe24e98dc4e4c1d002291604119 (diff)
fix t / t2 confusion in hsl_to_rgb
I think the original[1] SkRasterPipeline_opts.h version had a typo that I faithfully copied over to hsl_to_rgb. In Hue2RGB[2], the scalar equivalent of hue_to_rgb, we mutate t to keep it in 0-1 range[3]. In the SkRasterPipeline_opts.h code we introduced a new value t2 instead, and then used it everywhere, but accidentally typed 't' in the t < 1/6 case. The expression "p + (q - p)*6.0f*t" should have been "p + (q - p)*6.0f*t2". This fixes things by changing t in place, much like Hue2RGB does. The GM doesn't change anywhere, which is troubling. [1] https://skia-review.googlesource.com/c/7460/21/src/opts/SkRasterPipeline_opts.h#808 [2] https://skia-review.googlesource.com/c/7460/21/src/effects/SkHighContrastFilter.cpp#26 [3] I think this whole clamp should probably become "t = fract(t)". Will follow up. Change-Id: I3dcc1a79ffae46830178d931844ee3113f8bdfd1 Reviewed-on: https://skia-review.googlesource.com/14910 Reviewed-by: Brian Osman <brianosman@google.com> Commit-Queue: Mike Klein <mtklein@chromium.org>
-rw-r--r--src/jumper/SkJumper_generated.S2426
-rw-r--r--src/jumper/SkJumper_generated_win.S1999
-rw-r--r--src/jumper/SkJumper_stages.cpp16
3 files changed, 2220 insertions, 2221 deletions
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S
index 9aed8908d5..48530e5576 100644
--- a/src/jumper/SkJumper_generated.S
+++ b/src/jumper/SkJumper_generated.S
@@ -1030,90 +1030,91 @@ FUNCTION(_sk_hsl_to_rgb_aarch64)
_sk_hsl_to_rgb_aarch64:
.long 0x52a7d548 // mov w8, #0x3eaa0000
.long 0x72955568 // movk w8, #0xaaab
- .long 0x4e040d17 // dup v23.4s, w8
+ .long 0x4e040d18 // dup v24.4s, w8
.long 0x52a7c548 // mov w8, #0x3e2a0000
.long 0x72955568 // movk w8, #0xaaab
- .long 0x4e040d13 // dup v19.4s, w8
+ .long 0x4f03f613 // fmov v19.4s, #1.000000000000000000e+00
+ .long 0x4e040d15 // dup v21.4s, w8
.long 0x52a7e548 // mov w8, #0x3f2a0000
- .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
- .long 0x4f07f616 // fmov v22.4s, #-1.000000000000000000e+00
+ .long 0x4f0167f2 // movi v18.4s, #0x3f, lsl #24
+ .long 0x4e22d439 // fadd v25.4s, v1.4s, v2.4s
.long 0x72955568 // movk w8, #0xaaab
- .long 0x4e22d435 // fadd v21.4s, v1.4s, v2.4s
+ .long 0x4e33d43d // fadd v29.4s, v1.4s, v19.4s
+ .long 0x4ea0d830 // fcmeq v16.4s, v1.4s, #0.0
+ .long 0x4f07f616 // fmov v22.4s, #-1.000000000000000000e+00
.long 0x4e040d1a // dup v26.4s, w8
.long 0x52b7d548 // mov w8, #0xbeaa0000
- .long 0x6eb2e41d // fcmgt v29.4s, v0.4s, v18.4s
- .long 0x4e36d41e // fadd v30.4s, v0.4s, v22.4s
- .long 0x4f0167f1 // movi v17.4s, #0x3f, lsl #24
- .long 0x4ea0d830 // fcmeq v16.4s, v1.4s, #0.0
- .long 0x4ea0e819 // fcmlt v25.4s, v0.4s, #0.0
+ .long 0x6ea2e65c // fcmgt v28.4s, v18.4s, v2.4s
+ .long 0x4ea2cc39 // fmls v25.4s, v1.4s, v2.4s
+ .long 0x6e22dfa1 // fmul v1.4s, v29.4s, v2.4s
.long 0x72955568 // movk w8, #0xaaab
- .long 0x4e32d43c // fadd v28.4s, v1.4s, v18.4s
- .long 0x4ea2cc35 // fmls v21.4s, v1.4s, v2.4s
- .long 0x4e32d401 // fadd v1.4s, v0.4s, v18.4s
- .long 0x6e601fdd // bsl v29.16b, v30.16b, v0.16b
- .long 0x4e37d417 // fadd v23.4s, v0.4s, v23.4s
- .long 0x6ea2e63b // fcmgt v27.4s, v17.4s, v2.4s
- .long 0x4e040d1e // dup v30.4s, w8
- .long 0x6e22df9c // fmul v28.4s, v28.4s, v2.4s
- .long 0x6e7d1c39 // bsl v25.16b, v1.16b, v29.16b
- .long 0x6eb2e6e1 // fcmgt v1.4s, v23.4s, v18.4s
- .long 0x4e36d6fd // fadd v29.4s, v23.4s, v22.4s
- .long 0x4e3ed41e // fadd v30.4s, v0.4s, v30.4s
- .long 0x6e751f9b // bsl v27.16b, v28.16b, v21.16b
- .long 0x4ea0eaf5 // fcmlt v21.4s, v23.4s, #0.0
- .long 0x4e32d6fc // fadd v28.4s, v23.4s, v18.4s
- .long 0x6e771fa1 // bsl v1.16b, v29.16b, v23.16b
+ .long 0x6eb3e41f // fcmgt v31.4s, v0.4s, v19.4s
+ .long 0x6e791c3c // bsl v28.16b, v1.16b, v25.16b
+ .long 0x4e36d419 // fadd v25.4s, v0.4s, v22.4s
+ .long 0x4ea0e81b // fcmlt v27.4s, v0.4s, #0.0
+ .long 0x4e33d41e // fadd v30.4s, v0.4s, v19.4s
+ .long 0x6e601f3f // bsl v31.16b, v25.16b, v0.16b
+ .long 0x4e040d19 // dup v25.4s, w8
+ .long 0x4e38d418 // fadd v24.4s, v0.4s, v24.4s
+ .long 0x4e39d400 // fadd v0.4s, v0.4s, v25.4s
+ .long 0x6e7f1fdb // bsl v27.16b, v30.16b, v31.16b
+ .long 0x6eb3e71d // fcmgt v29.4s, v24.4s, v19.4s
+ .long 0x4e36d71e // fadd v30.4s, v24.4s, v22.4s
+ .long 0x6e781fdd // bsl v29.16b, v30.16b, v24.16b
+ .long 0x6eb3e41e // fcmgt v30.4s, v0.4s, v19.4s
+ .long 0x4e36d416 // fadd v22.4s, v0.4s, v22.4s
.long 0x4f026414 // movi v20.4s, #0x40, lsl #24
- .long 0x6e611f95 // bsl v21.16b, v28.16b, v1.16b
- .long 0x4e32d7c1 // fadd v1.4s, v30.4s, v18.4s
- .long 0x6eb2e7d2 // fcmgt v18.4s, v30.4s, v18.4s
- .long 0x4e36d7d6 // fadd v22.4s, v30.4s, v22.4s
- .long 0x6ea0fb7c // fneg v28.4s, v27.4s
- .long 0x4ea0ebdd // fcmlt v29.4s, v30.4s, #0.0
- .long 0x6e7e1ed2 // bsl v18.16b, v22.16b, v30.16b
- .long 0x4e22ce9c // fmla v28.4s, v20.4s, v2.4s
- .long 0x4f00f718 // fmov v24.4s, #6.000000000000000000e+00
- .long 0x6e721c3d // bsl v29.16b, v1.16b, v18.16b
- .long 0x4ebcd761 // fsub v1.4s, v27.4s, v28.4s
- .long 0x4eb5d752 // fsub v18.4s, v26.4s, v21.4s
- .long 0x4ebc1f94 // mov v20.16b, v28.16b
- .long 0x6e38dc38 // fmul v24.4s, v1.4s, v24.4s
- .long 0x4eb9d756 // fsub v22.4s, v26.4s, v25.4s
- .long 0x4ebc1f9f // mov v31.16b, v28.16b
- .long 0x4e32cf14 // fmla v20.4s, v24.4s, v18.4s
- .long 0x4ebc1f81 // mov v1.16b, v28.16b
- .long 0x4ebc1f92 // mov v18.16b, v28.16b
- .long 0x4e38cc1f // fmla v31.4s, v0.4s, v24.4s
- .long 0x4e36cf01 // fmla v1.4s, v24.4s, v22.4s
- .long 0x4ebdd740 // fsub v0.4s, v26.4s, v29.4s
- .long 0x4e3ecf12 // fmla v18.4s, v24.4s, v30.4s
- .long 0x4ebc1f96 // mov v22.16b, v28.16b
- .long 0x6eb5e75e // fcmgt v30.4s, v26.4s, v21.4s
- .long 0x4e20cf16 // fmla v22.4s, v24.4s, v0.4s
- .long 0x6e7c1e9e // bsl v30.16b, v20.16b, v28.16b
+ .long 0x4ea0eb19 // fcmlt v25.4s, v24.4s, #0.0
+ .long 0x4e33d718 // fadd v24.4s, v24.4s, v19.4s
+ .long 0x6e601ede // bsl v30.16b, v22.16b, v0.16b
+ .long 0x4ea0e816 // fcmlt v22.4s, v0.4s, #0.0
+ .long 0x4e33d400 // fadd v0.4s, v0.4s, v19.4s
+ .long 0x6ea0fb93 // fneg v19.4s, v28.4s
+ .long 0x4e22ce93 // fmla v19.4s, v20.4s, v2.4s
+ .long 0x4f00f717 // fmov v23.4s, #6.000000000000000000e+00
+ .long 0x6e7d1f19 // bsl v25.16b, v24.16b, v29.16b
+ .long 0x6e7e1c16 // bsl v22.16b, v0.16b, v30.16b
+ .long 0x4eb3d780 // fsub v0.4s, v28.4s, v19.4s
+ .long 0x4ebbd75d // fsub v29.4s, v26.4s, v27.4s
+ .long 0x4eb9d754 // fsub v20.4s, v26.4s, v25.4s
+ .long 0x4eb31e7e // mov v30.16b, v19.16b
+ .long 0x6e37dc00 // fmul v0.4s, v0.4s, v23.4s
+ .long 0x4eb31e77 // mov v23.16b, v19.16b
+ .long 0x4e34cc1e // fmla v30.4s, v0.4s, v20.4s
+ .long 0x4eb6d754 // fsub v20.4s, v26.4s, v22.4s
+ .long 0x4e3dcc17 // fmla v23.4s, v0.4s, v29.4s
+ .long 0x4eb31e7d // mov v29.16b, v19.16b
+ .long 0x4e34cc1d // fmla v29.4s, v0.4s, v20.4s
.long 0x6eb9e754 // fcmgt v20.4s, v26.4s, v25.4s
- .long 0x6ebde75a // fcmgt v26.4s, v26.4s, v29.4s
- .long 0x6e7c1c34 // bsl v20.16b, v1.16b, v28.16b
- .long 0x6e7c1eda // bsl v26.16b, v22.16b, v28.16b
- .long 0x4e37cf1c // fmla v28.4s, v24.4s, v23.4s
- .long 0x6eb9e637 // fcmgt v23.4s, v17.4s, v25.4s
- .long 0x6eb5e678 // fcmgt v24.4s, v19.4s, v21.4s
- .long 0x6eb5e635 // fcmgt v21.4s, v17.4s, v21.4s
- .long 0x6ebde631 // fcmgt v17.4s, v17.4s, v29.4s
+ .long 0x6e731fd4 // bsl v20.16b, v30.16b, v19.16b
+ .long 0x6ebbe75e // fcmgt v30.4s, v26.4s, v27.4s
+ .long 0x6eb6e75a // fcmgt v26.4s, v26.4s, v22.4s
+ .long 0x6e731efe // bsl v30.16b, v23.16b, v19.16b
+ .long 0x4eb31e77 // mov v23.16b, v19.16b
+ .long 0x6e731fba // bsl v26.16b, v29.16b, v19.16b
+ .long 0x4eb31e7d // mov v29.16b, v19.16b
+ .long 0x6ebbe6b8 // fcmgt v24.4s, v21.4s, v27.4s
+ .long 0x4e3bcc17 // fmla v23.4s, v0.4s, v27.4s
+ .long 0x6ebbe65b // fcmgt v27.4s, v18.4s, v27.4s
+ .long 0x4e36cc1d // fmla v29.4s, v0.4s, v22.4s
+ .long 0x4e39cc13 // fmla v19.4s, v0.4s, v25.4s
+ .long 0x6eb9e6a0 // fcmgt v0.4s, v21.4s, v25.4s
+ .long 0x6eb9e659 // fcmgt v25.4s, v18.4s, v25.4s
+ .long 0x6eb6e652 // fcmgt v18.4s, v18.4s, v22.4s
.long 0xf8408423 // ldr x3, [x1], #8
- .long 0x6eb9e676 // fcmgt v22.4s, v19.4s, v25.4s
- .long 0x6ebde673 // fcmgt v19.4s, v19.4s, v29.4s
- .long 0x6e7a1f71 // bsl v17.16b, v27.16b, v26.16b
- .long 0x6e7e1f75 // bsl v21.16b, v27.16b, v30.16b
- .long 0x6e741f77 // bsl v23.16b, v27.16b, v20.16b
- .long 0x6e711e53 // bsl v19.16b, v18.16b, v17.16b
- .long 0x4eb01e00 // mov v0.16b, v16.16b
+ .long 0x6eb6e6b5 // fcmgt v21.4s, v21.4s, v22.4s
+ .long 0x6e741f99 // bsl v25.16b, v28.16b, v20.16b
+ .long 0x6e7a1f92 // bsl v18.16b, v28.16b, v26.16b
+ .long 0x4eb01e11 // mov v17.16b, v16.16b
+ .long 0x6e7e1f9b // bsl v27.16b, v28.16b, v30.16b
+ .long 0x6e791e60 // bsl v0.16b, v19.16b, v25.16b
+ .long 0x6e721fb5 // bsl v21.16b, v29.16b, v18.16b
.long 0x4eb01e01 // mov v1.16b, v16.16b
- .long 0x6e751f98 // bsl v24.16b, v28.16b, v21.16b
- .long 0x6e771ff6 // bsl v22.16b, v31.16b, v23.16b
- .long 0x6e731c50 // bsl v16.16b, v2.16b, v19.16b
- .long 0x6e781c40 // bsl v0.16b, v2.16b, v24.16b
- .long 0x6e761c41 // bsl v1.16b, v2.16b, v22.16b
+ .long 0x6e7b1ef8 // bsl v24.16b, v23.16b, v27.16b
+ .long 0x6e601c51 // bsl v17.16b, v2.16b, v0.16b
+ .long 0x6e751c50 // bsl v16.16b, v2.16b, v21.16b
+ .long 0x6e781c41 // bsl v1.16b, v2.16b, v24.16b
+ .long 0x4eb11e20 // mov v0.16b, v17.16b
.long 0x4eb01e02 // mov v2.16b, v16.16b
.long 0xd61f0060 // br x3
@@ -2216,9 +2217,9 @@ FUNCTION(_sk_gather_i8_aarch64)
_sk_gather_i8_aarch64:
.long 0xaa0103e8 // mov x8, x1
.long 0xf8408429 // ldr x9, [x1], #8
- .long 0xb4000069 // cbz x9, 1d28 <sk_gather_i8_aarch64+0x14>
+ .long 0xb4000069 // cbz x9, 1d2c <sk_gather_i8_aarch64+0x14>
.long 0xaa0903ea // mov x10, x9
- .long 0x14000003 // b 1d30 <sk_gather_i8_aarch64+0x1c>
+ .long 0x14000003 // b 1d34 <sk_gather_i8_aarch64+0x1c>
.long 0xf940050a // ldr x10, [x8, #8]
.long 0x91004101 // add x1, x8, #0x10
.long 0xf8410548 // ldr x8, [x10], #16
@@ -3067,7 +3068,7 @@ _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, 28fc <sk_linear_gradient_aarch64+0x100>
+ .long 0xb40006c8 // cbz x8, 2900 <sk_linear_gradient_aarch64+0x100>
.long 0x6dbf23e9 // stp d9, d8, [sp, #-16]!
.long 0xf9400529 // ldr x9, [x9, #8]
.long 0x6f00e413 // movi v19.2d, #0x0
@@ -3118,9 +3119,9 @@ _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, 2844 <sk_linear_gradient_aarch64+0x48>
+ .long 0xb5fffaa8 // cbnz x8, 2848 <sk_linear_gradient_aarch64+0x48>
.long 0x6cc123e9 // ldp d9, d8, [sp], #16
- .long 0x14000005 // b 290c <sk_linear_gradient_aarch64+0x110>
+ .long 0x14000005 // b 2910 <sk_linear_gradient_aarch64+0x110>
.long 0x6f00e414 // movi v20.2d, #0x0
.long 0x6f00e412 // movi v18.2d, #0x0
.long 0x6f00e411 // movi v17.2d, #0x0
@@ -4580,16 +4581,18 @@ HIDDEN _sk_hsl_to_rgb_vfp4
.globl _sk_hsl_to_rgb_vfp4
FUNCTION(_sk_hsl_to_rgb_vfp4)
_sk_hsl_to_rgb_vfp4:
- .long 0xf2c72f10 // vmov.f32 d18, #1
- .long 0xeddf0b4f // vldr d16, [pc, #316]
- .long 0xf2c3161f // vmov.i32 d17, #1056964608
- .long 0xeddf9b4f // vldr d25, [pc, #316]
- .long 0xf2415d22 // vadd.f32 d21, d1, d18
- .long 0xe4913004 // ldr r3, [r1], #4
+ .long 0xed2d8b02 // vpush {d8}
+ .long 0xf2c71f10 // vmov.f32 d17, #1
+ .long 0xeddf0b50 // vldr d16, [pc, #320]
+ .long 0xf2c3261f // vmov.i32 d18, #1056964608
+ .long 0xeddf9b50 // vldr d25, [pc, #320]
+ .long 0xf2415d21 // vadd.f32 d21, d1, d17
+ .long 0xed9f8b52 // vldr d8, [pc, #328]
.long 0xf3414d12 // vmul.f32 d20, d1, d2
+ .long 0xe4913004 // ldr r3, [r1], #4
.long 0xf2416d02 // vadd.f32 d22, d1, d2
.long 0xf2407d20 // vadd.f32 d23, d0, d16
- .long 0xf3610e82 // vcgt.f32 d16, d17, d2
+ .long 0xf3620e82 // vcgt.f32 d16, d18, d2
.long 0xf3455d92 // vmul.f32 d21, d21, d2
.long 0xf2664da4 // vsub.f32 d20, d22, d20
.long 0xf2426d02 // vadd.f32 d22, d2, d2
@@ -4597,69 +4600,69 @@ _sk_hsl_to_rgb_vfp4:
.long 0xf35501b4 // vbsl d16, d21, d20
.long 0xf2409d29 // vadd.f32 d25, d0, d25
.long 0xf2408d23 // vadd.f32 d24, d0, d19
- .long 0xf3f9e629 // vclt.f32 d30, d25, #0
- .long 0xf360ae22 // vcgt.f32 d26, d0, d18
+ .long 0xf360ae21 // vcgt.f32 d26, d0, d17
.long 0xf247cda3 // vadd.f32 d28, d23, d19
- .long 0xf367dea2 // vcgt.f32 d29, d23, d18
- .long 0xf240bd22 // vadd.f32 d27, d0, d18
+ .long 0xf367dea1 // vcgt.f32 d29, d23, d17
+ .long 0xf240bd21 // vadd.f32 d27, d0, d17
.long 0xf2666da0 // vsub.f32 d22, d22, d16
- .long 0xf2474da2 // vadd.f32 d20, d23, d18
+ .long 0xf2474da1 // vadd.f32 d20, d23, d17
.long 0xf358a190 // vbsl d26, d24, d0
.long 0xf3f98600 // vclt.f32 d24, d0, #0
- .long 0xf3695ea2 // vcgt.f32 d21, d25, d18
+ .long 0xf3695ea1 // vcgt.f32 d21, d25, d17
.long 0xf2493da3 // vadd.f32 d19, d25, d19
.long 0xf35b81ba // vbsl d24, d27, d26
- .long 0xf3f9a627 // vclt.f32 d26, d23, #0
+ .long 0xeddfbb38 // vldr d27, [pc, #224]
.long 0xf35cd1b7 // vbsl d29, d28, d23
- .long 0xeddfcb35 // vldr d28, [pc, #212]
- .long 0xf2492da2 // vadd.f32 d18, d25, d18
- .long 0xf260bda6 // vsub.f32 d27, d16, d22
- .long 0xf354a1bd // vbsl d26, d20, d29
+ .long 0xf3f97627 // vclt.f32 d23, d23, #0
+ .long 0xf2491da1 // vadd.f32 d17, d25, d17
+ .long 0xf260ada6 // vsub.f32 d26, d16, d22
+ .long 0xf35471bd // vbsl d23, d20, d29
.long 0xf2c14f18 // vmov.f32 d20, #6
.long 0xf35351b9 // vbsl d21, d19, d25
- .long 0xf26cddaa // vsub.f32 d29, d28, d26
- .long 0xf352e1b5 // vbsl d30, d18, d21
- .long 0xf34b2db4 // vmul.f32 d18, d27, d20
- .long 0xf26c3da8 // vsub.f32 d19, d28, d24
- .long 0xf26c4dae // vsub.f32 d20, d28, d30
- .long 0xf36cbeaa // vcgt.f32 d27, d28, d26
- .long 0xf3425dbd // vmul.f32 d21, d18, d29
- .long 0xf3477db2 // vmul.f32 d23, d23, d18
- .long 0xf3423db3 // vmul.f32 d19, d18, d19
- .long 0xf3444db2 // vmul.f32 d20, d20, d18
+ .long 0xf3f99629 // vclt.f32 d25, d25, #0
+ .long 0xf26bcda7 // vsub.f32 d28, d27, d23
+ .long 0xf35191b5 // vbsl d25, d17, d21
+ .long 0xf34a1db4 // vmul.f32 d17, d26, d20
+ .long 0xf26b3da8 // vsub.f32 d19, d27, d24
+ .long 0xf26b4da9 // vsub.f32 d20, d27, d25
+ .long 0xf36beea8 // vcgt.f32 d30, d27, d24
+ .long 0xf3415dbc // vmul.f32 d21, d17, d28
+ .long 0xf36bcea7 // vcgt.f32 d28, d27, d23
+ .long 0xf3413db3 // vmul.f32 d19, d17, d19
+ .long 0xf3444db1 // vmul.f32 d20, d20, d17
.long 0xf2465da5 // vadd.f32 d21, d22, d21
- .long 0xf342dd90 // vmul.f32 d29, d18, d0
- .long 0xf3210eaa // vcgt.f32 d0, d17, d26
- .long 0xf3492db2 // vmul.f32 d18, d25, d18
- .long 0xf355b1b6 // vbsl d27, d21, d22
- .long 0xeddf5b22 // vldr d21, [pc, #136]
- .long 0xf36cfea8 // vcgt.f32 d31, d28, d24
+ .long 0xf341ddb8 // vmul.f32 d29, d17, d24
+ .long 0xf341adb7 // vmul.f32 d26, d17, d23
.long 0xf2463da3 // vadd.f32 d19, d22, d19
- .long 0xf36cceae // vcgt.f32 d28, d28, d30
+ .long 0xf3220ea8 // vcgt.f32 d0, d18, d24
+ .long 0xf3491db1 // vmul.f32 d17, d25, d17
+ .long 0xf355c1b6 // vbsl d28, d21, d22
+ .long 0xf3685e28 // vcgt.f32 d21, d8, d24
+ .long 0xf36bbea9 // vcgt.f32 d27, d27, d25
.long 0xf2464da4 // vadd.f32 d20, d22, d20
- .long 0xf365aeaa // vcgt.f32 d26, d21, d26
- .long 0xf2467da7 // vadd.f32 d23, d22, d23
- .long 0xf3619ea8 // vcgt.f32 d25, d17, d24
- .long 0xf3611eae // vcgt.f32 d17, d17, d30
- .long 0xf31001bb // vbsl d0, d16, d27
- .long 0xf353f1b6 // vbsl d31, d19, d22
- .long 0xf354c1b6 // vbsl d28, d20, d22
- .long 0xf357a190 // vbsl d26, d23, d0
+ .long 0xf2468dad // vadd.f32 d24, d22, d29
+ .long 0xf362fea7 // vcgt.f32 d31, d18, d23
+ .long 0xf3622ea9 // vcgt.f32 d18, d18, d25
+ .long 0xf353e1b6 // vbsl d30, d19, d22
+ .long 0xf3687e27 // vcgt.f32 d23, d8, d23
+ .long 0xf246adaa // vadd.f32 d26, d22, d26
+ .long 0xf31001be // vbsl d0, d16, d30
+ .long 0xf354b1b6 // vbsl d27, d20, d22
+ .long 0xf3585190 // vbsl d21, d24, d0
.long 0xf3b90501 // vceq.f32 d0, d1, #0
- .long 0xf3658ea8 // vcgt.f32 d24, d21, d24
- .long 0xf246ddad // vadd.f32 d29, d22, d29
- .long 0xf3653eae // vcgt.f32 d19, d21, d30
- .long 0xf2462da2 // vadd.f32 d18, d22, d18
- .long 0xf35091bf // vbsl d25, d16, d31
- .long 0xf35011bc // vbsl d17, d16, d28
+ .long 0xf3683e29 // vcgt.f32 d19, d8, d25
+ .long 0xf2461da1 // vadd.f32 d17, d22, d17
+ .long 0xf350f1bc // vbsl d31, d16, d28
+ .long 0xf35021bb // vbsl d18, d16, d27
.long 0xf2600110 // vorr d16, d0, d0
+ .long 0xf35a71bf // vbsl d23, d26, d31
.long 0xf2201110 // vorr d1, d0, d0
- .long 0xf352013a // vbsl d16, d2, d26
- .long 0xf35d81b9 // vbsl d24, d29, d25
- .long 0xf35231b1 // vbsl d19, d18, d17
- .long 0xf3121138 // vbsl d1, d2, d24
+ .long 0xf3520137 // vbsl d16, d2, d23
+ .long 0xf35131b2 // vbsl d19, d17, d18
+ .long 0xf3121135 // vbsl d1, d2, d21
.long 0xf3120133 // vbsl d0, d2, d19
.long 0xf22021b0 // vorr d2, d16, d16
+ .long 0xecbd8b02 // vpop {d8}
.long 0xe12fff13 // bx r3
.long 0xe320f000 // nop {0}
.long 0xbeaaaaab // .word 0xbeaaaaab
@@ -6812,7 +6815,7 @@ _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 2d70 <sk_linear_gradient_vfp4+0x110>
+ .long 0x0a000036 // beq 2d78 <sk_linear_gradient_vfp4+0x110>
.long 0xe59e3004 // ldr r3, [lr, #4]
.long 0xf2c01010 // vmov.i32 d17, #0
.long 0xf2c07010 // vmov.i32 d23, #0
@@ -6862,12 +6865,12 @@ _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 2cac <sk_linear_gradient_vfp4+0x4c>
+ .long 0x1affffd3 // bne 2cb4 <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 2d80 <sk_linear_gradient_vfp4+0x120>
+ .long 0xea000003 // b 2d88 <sk_linear_gradient_vfp4+0x120>
.long 0xf2c05010 // vmov.i32 d21, #0
.long 0xf2c04010 // vmov.i32 d20, #0
.long 0xf2c03010 // vmov.i32 d19, #0
@@ -7347,14 +7350,14 @@ _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 196,226,125,24,13,173,60,0,0 // vbroadcastss 0x3cad(%rip),%ymm1 # 3d70 <_sk_callback_hsw+0x125>
+ .byte 196,226,125,24,13,165,60,0,0 // vbroadcastss 0x3ca5(%rip),%ymm1 # 3d68 <_sk_callback_hsw+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 196,226,125,24,21,145,60,0,0 // vbroadcastss 0x3c91(%rip),%ymm2 # 3d74 <_sk_callback_hsw+0x129>
+ .byte 196,226,125,24,21,137,60,0,0 // vbroadcastss 0x3c89(%rip),%ymm2 # 3d6c <_sk_callback_hsw+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
@@ -7390,7 +7393,7 @@ HIDDEN _sk_srcatop_hsw
FUNCTION(_sk_srcatop_hsw)
_sk_srcatop_hsw:
.byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
- .byte 196,98,125,24,5,65,60,0,0 // vbroadcastss 0x3c41(%rip),%ymm8 # 3d78 <_sk_callback_hsw+0x12d>
+ .byte 196,98,125,24,5,57,60,0,0 // vbroadcastss 0x3c39(%rip),%ymm8 # 3d70 <_sk_callback_hsw+0x12d>
.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
@@ -7406,7 +7409,7 @@ HIDDEN _sk_dstatop_hsw
.globl _sk_dstatop_hsw
FUNCTION(_sk_dstatop_hsw)
_sk_dstatop_hsw:
- .byte 196,98,125,24,5,20,60,0,0 // vbroadcastss 0x3c14(%rip),%ymm8 # 3d7c <_sk_callback_hsw+0x131>
+ .byte 196,98,125,24,5,12,60,0,0 // vbroadcastss 0x3c0c(%rip),%ymm8 # 3d74 <_sk_callback_hsw+0x131>
.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
@@ -7445,7 +7448,7 @@ HIDDEN _sk_srcout_hsw
.globl _sk_srcout_hsw
FUNCTION(_sk_srcout_hsw)
_sk_srcout_hsw:
- .byte 196,98,125,24,5,187,59,0,0 // vbroadcastss 0x3bbb(%rip),%ymm8 # 3d80 <_sk_callback_hsw+0x135>
+ .byte 196,98,125,24,5,179,59,0,0 // vbroadcastss 0x3bb3(%rip),%ymm8 # 3d78 <_sk_callback_hsw+0x135>
.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
@@ -7458,7 +7461,7 @@ HIDDEN _sk_dstout_hsw
.globl _sk_dstout_hsw
FUNCTION(_sk_dstout_hsw)
_sk_dstout_hsw:
- .byte 196,226,125,24,5,158,59,0,0 // vbroadcastss 0x3b9e(%rip),%ymm0 # 3d84 <_sk_callback_hsw+0x139>
+ .byte 196,226,125,24,5,150,59,0,0 // vbroadcastss 0x3b96(%rip),%ymm0 # 3d7c <_sk_callback_hsw+0x139>
.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
@@ -7471,7 +7474,7 @@ HIDDEN _sk_srcover_hsw
.globl _sk_srcover_hsw
FUNCTION(_sk_srcover_hsw)
_sk_srcover_hsw:
- .byte 196,98,125,24,5,129,59,0,0 // vbroadcastss 0x3b81(%rip),%ymm8 # 3d88 <_sk_callback_hsw+0x13d>
+ .byte 196,98,125,24,5,121,59,0,0 // vbroadcastss 0x3b79(%rip),%ymm8 # 3d80 <_sk_callback_hsw+0x13d>
.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
@@ -7484,7 +7487,7 @@ HIDDEN _sk_dstover_hsw
.globl _sk_dstover_hsw
FUNCTION(_sk_dstover_hsw)
_sk_dstover_hsw:
- .byte 196,98,125,24,5,96,59,0,0 // vbroadcastss 0x3b60(%rip),%ymm8 # 3d8c <_sk_callback_hsw+0x141>
+ .byte 196,98,125,24,5,88,59,0,0 // vbroadcastss 0x3b58(%rip),%ymm8 # 3d84 <_sk_callback_hsw+0x141>
.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
@@ -7508,7 +7511,7 @@ HIDDEN _sk_multiply_hsw
.globl _sk_multiply_hsw
FUNCTION(_sk_multiply_hsw)
_sk_multiply_hsw:
- .byte 196,98,125,24,5,43,59,0,0 // vbroadcastss 0x3b2b(%rip),%ymm8 # 3d90 <_sk_callback_hsw+0x145>
+ .byte 196,98,125,24,5,35,59,0,0 // vbroadcastss 0x3b23(%rip),%ymm8 # 3d88 <_sk_callback_hsw+0x145>
.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
@@ -7556,7 +7559,7 @@ HIDDEN _sk_xor__hsw
.globl _sk_xor__hsw
FUNCTION(_sk_xor__hsw)
_sk_xor__hsw:
- .byte 196,98,125,24,5,166,58,0,0 // vbroadcastss 0x3aa6(%rip),%ymm8 # 3d94 <_sk_callback_hsw+0x149>
+ .byte 196,98,125,24,5,158,58,0,0 // vbroadcastss 0x3a9e(%rip),%ymm8 # 3d8c <_sk_callback_hsw+0x149>
.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
@@ -7590,7 +7593,7 @@ _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 196,98,125,24,5,46,58,0,0 // vbroadcastss 0x3a2e(%rip),%ymm8 # 3d98 <_sk_callback_hsw+0x14d>
+ .byte 196,98,125,24,5,38,58,0,0 // vbroadcastss 0x3a26(%rip),%ymm8 # 3d90 <_sk_callback_hsw+0x14d>
.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
@@ -7615,7 +7618,7 @@ _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 196,98,125,24,5,221,57,0,0 // vbroadcastss 0x39dd(%rip),%ymm8 # 3d9c <_sk_callback_hsw+0x151>
+ .byte 196,98,125,24,5,213,57,0,0 // vbroadcastss 0x39d5(%rip),%ymm8 # 3d94 <_sk_callback_hsw+0x151>
.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
@@ -7643,7 +7646,7 @@ _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 196,98,125,24,5,128,57,0,0 // vbroadcastss 0x3980(%rip),%ymm8 # 3da0 <_sk_callback_hsw+0x155>
+ .byte 196,98,125,24,5,120,57,0,0 // vbroadcastss 0x3978(%rip),%ymm8 # 3d98 <_sk_callback_hsw+0x155>
.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
@@ -7665,7 +7668,7 @@ _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 196,98,125,24,5,62,57,0,0 // vbroadcastss 0x393e(%rip),%ymm8 # 3da4 <_sk_callback_hsw+0x159>
+ .byte 196,98,125,24,5,54,57,0,0 // vbroadcastss 0x3936(%rip),%ymm8 # 3d9c <_sk_callback_hsw+0x159>
.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
@@ -7675,7 +7678,7 @@ HIDDEN _sk_colorburn_hsw
.globl _sk_colorburn_hsw
FUNCTION(_sk_colorburn_hsw)
_sk_colorburn_hsw:
- .byte 196,98,125,24,5,44,57,0,0 // vbroadcastss 0x392c(%rip),%ymm8 # 3da8 <_sk_callback_hsw+0x15d>
+ .byte 196,98,125,24,5,36,57,0,0 // vbroadcastss 0x3924(%rip),%ymm8 # 3da0 <_sk_callback_hsw+0x15d>
.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
@@ -7733,7 +7736,7 @@ HIDDEN _sk_colordodge_hsw
FUNCTION(_sk_colordodge_hsw)
_sk_colordodge_hsw:
.byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
- .byte 196,98,125,24,13,55,56,0,0 // vbroadcastss 0x3837(%rip),%ymm9 # 3dac <_sk_callback_hsw+0x161>
+ .byte 196,98,125,24,13,47,56,0,0 // vbroadcastss 0x382f(%rip),%ymm9 # 3da4 <_sk_callback_hsw+0x161>
.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
@@ -7786,7 +7789,7 @@ HIDDEN _sk_hardlight_hsw
.globl _sk_hardlight_hsw
FUNCTION(_sk_hardlight_hsw)
_sk_hardlight_hsw:
- .byte 196,98,125,24,5,88,55,0,0 // vbroadcastss 0x3758(%rip),%ymm8 # 3db0 <_sk_callback_hsw+0x165>
+ .byte 196,98,125,24,5,80,55,0,0 // vbroadcastss 0x3750(%rip),%ymm8 # 3da8 <_sk_callback_hsw+0x165>
.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
@@ -7837,7 +7840,7 @@ HIDDEN _sk_overlay_hsw
.globl _sk_overlay_hsw
FUNCTION(_sk_overlay_hsw)
_sk_overlay_hsw:
- .byte 196,98,125,24,5,144,54,0,0 // vbroadcastss 0x3690(%rip),%ymm8 # 3db4 <_sk_callback_hsw+0x169>
+ .byte 196,98,125,24,5,136,54,0,0 // vbroadcastss 0x3688(%rip),%ymm8 # 3dac <_sk_callback_hsw+0x169>
.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
@@ -7898,10 +7901,10 @@ _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 196,98,125,24,29,155,53,0,0 // vbroadcastss 0x359b(%rip),%ymm11 # 3dbc <_sk_callback_hsw+0x171>
+ .byte 196,98,125,24,29,147,53,0,0 // vbroadcastss 0x3593(%rip),%ymm11 # 3db4 <_sk_callback_hsw+0x171>
.byte 196,65,20,88,227 // vaddps %ymm11,%ymm13,%ymm12
.byte 196,65,28,89,192 // vmulps %ymm8,%ymm12,%ymm8
- .byte 196,98,125,24,37,140,53,0,0 // vbroadcastss 0x358c(%rip),%ymm12 # 3dc0 <_sk_callback_hsw+0x175>
+ .byte 196,98,125,24,37,132,53,0,0 // vbroadcastss 0x3584(%rip),%ymm12 # 3db8 <_sk_callback_hsw+0x175>
.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
@@ -7911,7 +7914,7 @@ _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 196,98,125,24,5,79,53,0,0 // vbroadcastss 0x354f(%rip),%ymm8 # 3db8 <_sk_callback_hsw+0x16d>
+ .byte 196,98,125,24,5,71,53,0,0 // vbroadcastss 0x3547(%rip),%ymm8 # 3db0 <_sk_callback_hsw+0x16d>
.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
@@ -8004,7 +8007,7 @@ HIDDEN _sk_clamp_1_hsw
.globl _sk_clamp_1_hsw
FUNCTION(_sk_clamp_1_hsw)
_sk_clamp_1_hsw:
- .byte 196,98,125,24,5,212,51,0,0 // vbroadcastss 0x33d4(%rip),%ymm8 # 3dc4 <_sk_callback_hsw+0x179>
+ .byte 196,98,125,24,5,204,51,0,0 // vbroadcastss 0x33cc(%rip),%ymm8 # 3dbc <_sk_callback_hsw+0x179>
.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
@@ -8016,7 +8019,7 @@ HIDDEN _sk_clamp_a_hsw
.globl _sk_clamp_a_hsw
FUNCTION(_sk_clamp_a_hsw)
_sk_clamp_a_hsw:
- .byte 196,98,125,24,5,183,51,0,0 // vbroadcastss 0x33b7(%rip),%ymm8 # 3dc8 <_sk_callback_hsw+0x17d>
+ .byte 196,98,125,24,5,175,51,0,0 // vbroadcastss 0x33af(%rip),%ymm8 # 3dc0 <_sk_callback_hsw+0x17d>
.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
@@ -8102,7 +8105,7 @@ 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 196,98,125,24,21,255,50,0,0 // vbroadcastss 0x32ff(%rip),%ymm10 # 3dcc <_sk_callback_hsw+0x181>
+ .byte 196,98,125,24,21,247,50,0,0 // vbroadcastss 0x32f7(%rip),%ymm10 # 3dc4 <_sk_callback_hsw+0x181>
.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
@@ -8115,16 +8118,16 @@ HIDDEN _sk_from_srgb_hsw
.globl _sk_from_srgb_hsw
FUNCTION(_sk_from_srgb_hsw)
_sk_from_srgb_hsw:
- .byte 196,98,125,24,5,224,50,0,0 // vbroadcastss 0x32e0(%rip),%ymm8 # 3dd0 <_sk_callback_hsw+0x185>
+ .byte 196,98,125,24,5,216,50,0,0 // vbroadcastss 0x32d8(%rip),%ymm8 # 3dc8 <_sk_callback_hsw+0x185>
.byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9
.byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
- .byte 196,98,125,24,29,210,50,0,0 // vbroadcastss 0x32d2(%rip),%ymm11 # 3dd4 <_sk_callback_hsw+0x189>
- .byte 196,98,125,24,37,205,50,0,0 // vbroadcastss 0x32cd(%rip),%ymm12 # 3dd8 <_sk_callback_hsw+0x18d>
+ .byte 196,98,125,24,29,202,50,0,0 // vbroadcastss 0x32ca(%rip),%ymm11 # 3dcc <_sk_callback_hsw+0x189>
+ .byte 196,98,125,24,37,197,50,0,0 // vbroadcastss 0x32c5(%rip),%ymm12 # 3dd0 <_sk_callback_hsw+0x18d>
.byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13
.byte 196,66,125,168,235 // vfmadd213ps %ymm11,%ymm0,%ymm13
- .byte 196,98,125,24,53,190,50,0,0 // vbroadcastss 0x32be(%rip),%ymm14 # 3ddc <_sk_callback_hsw+0x191>
+ .byte 196,98,125,24,53,182,50,0,0 // vbroadcastss 0x32b6(%rip),%ymm14 # 3dd4 <_sk_callback_hsw+0x191>
.byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13
- .byte 196,98,125,24,21,180,50,0,0 // vbroadcastss 0x32b4(%rip),%ymm10 # 3de0 <_sk_callback_hsw+0x195>
+ .byte 196,98,125,24,21,172,50,0,0 // vbroadcastss 0x32ac(%rip),%ymm10 # 3dd8 <_sk_callback_hsw+0x195>
.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
@@ -8150,16 +8153,16 @@ _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 196,98,125,24,5,78,50,0,0 // vbroadcastss 0x324e(%rip),%ymm8 # 3de4 <_sk_callback_hsw+0x199>
+ .byte 196,98,125,24,5,70,50,0,0 // vbroadcastss 0x3246(%rip),%ymm8 # 3ddc <_sk_callback_hsw+0x199>
.byte 196,65,124,89,216 // vmulps %ymm8,%ymm0,%ymm11
- .byte 196,98,125,24,37,68,50,0,0 // vbroadcastss 0x3244(%rip),%ymm12 # 3de8 <_sk_callback_hsw+0x19d>
- .byte 196,98,125,24,45,63,50,0,0 // vbroadcastss 0x323f(%rip),%ymm13 # 3dec <_sk_callback_hsw+0x1a1>
+ .byte 196,98,125,24,37,60,50,0,0 // vbroadcastss 0x323c(%rip),%ymm12 # 3de0 <_sk_callback_hsw+0x19d>
+ .byte 196,98,125,24,45,55,50,0,0 // vbroadcastss 0x3237(%rip),%ymm13 # 3de4 <_sk_callback_hsw+0x1a1>
.byte 196,66,21,168,204 // vfmadd213ps %ymm12,%ymm13,%ymm9
- .byte 196,98,125,24,53,53,50,0,0 // vbroadcastss 0x3235(%rip),%ymm14 # 3df0 <_sk_callback_hsw+0x1a5>
+ .byte 196,98,125,24,53,45,50,0,0 // vbroadcastss 0x322d(%rip),%ymm14 # 3de8 <_sk_callback_hsw+0x1a5>
.byte 196,66,13,184,202 // vfmadd231ps %ymm10,%ymm14,%ymm9
- .byte 196,98,125,24,21,43,50,0,0 // vbroadcastss 0x322b(%rip),%ymm10 # 3df4 <_sk_callback_hsw+0x1a9>
+ .byte 196,98,125,24,21,35,50,0,0 // vbroadcastss 0x3223(%rip),%ymm10 # 3dec <_sk_callback_hsw+0x1a9>
.byte 196,65,44,93,201 // vminps %ymm9,%ymm10,%ymm9
- .byte 196,98,125,24,61,33,50,0,0 // vbroadcastss 0x3221(%rip),%ymm15 # 3df8 <_sk_callback_hsw+0x1ad>
+ .byte 196,98,125,24,61,25,50,0,0 // vbroadcastss 0x3219(%rip),%ymm15 # 3df0 <_sk_callback_hsw+0x1ad>
.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
@@ -8192,26 +8195,26 @@ _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 196,98,125,24,29,155,49,0,0 // vbroadcastss 0x319b(%rip),%ymm11 # 3dfc <_sk_callback_hsw+0x1b1>
+ .byte 196,98,125,24,29,147,49,0,0 // vbroadcastss 0x3193(%rip),%ymm11 # 3df4 <_sk_callback_hsw+0x1b1>
.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 196,98,125,24,53,136,49,0,0 // vbroadcastss 0x3188(%rip),%ymm14 # 3e00 <_sk_callback_hsw+0x1b5>
+ .byte 196,98,125,24,53,128,49,0,0 // vbroadcastss 0x3180(%rip),%ymm14 # 3df8 <_sk_callback_hsw+0x1b5>
.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 196,98,125,24,53,111,49,0,0 // vbroadcastss 0x316f(%rip),%ymm14 # 3e08 <_sk_callback_hsw+0x1bd>
+ .byte 196,98,125,24,53,103,49,0,0 // vbroadcastss 0x3167(%rip),%ymm14 # 3e00 <_sk_callback_hsw+0x1bd>
.byte 196,66,37,168,238 // vfmadd213ps %ymm14,%ymm11,%ymm13
- .byte 196,98,125,24,53,93,49,0,0 // vbroadcastss 0x315d(%rip),%ymm14 # 3e04 <_sk_callback_hsw+0x1b9>
+ .byte 196,98,125,24,53,85,49,0,0 // vbroadcastss 0x3155(%rip),%ymm14 # 3dfc <_sk_callback_hsw+0x1b9>
.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 196,98,125,24,29,64,49,0,0 // vbroadcastss 0x3140(%rip),%ymm11 # 3e10 <_sk_callback_hsw+0x1c5>
+ .byte 196,98,125,24,29,56,49,0,0 // vbroadcastss 0x3138(%rip),%ymm11 # 3e08 <_sk_callback_hsw+0x1c5>
.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
@@ -8221,7 +8224,7 @@ _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 196,98,125,24,5,3,49,0,0 // vbroadcastss 0x3103(%rip),%ymm8 # 3e0c <_sk_callback_hsw+0x1c1>
+ .byte 196,98,125,24,5,251,48,0,0 // vbroadcastss 0x30fb(%rip),%ymm8 # 3e04 <_sk_callback_hsw+0x1c1>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -8236,95 +8239,93 @@ _sk_hsl_to_rgb_hsw:
.byte 197,252,17,44,36 // vmovups %ymm5,(%rsp)
.byte 197,252,17,100,36,224 // vmovups %ymm4,-0x20(%rsp)
.byte 197,252,17,92,36,192 // vmovups %ymm3,-0x40(%rsp)
- .byte 197,252,40,234 // vmovaps %ymm2,%ymm5
- .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
- .byte 184,0,0,0,63 // mov $0x3f000000,%eax
- .byte 197,249,110,192 // vmovd %eax,%xmm0
- .byte 196,98,125,88,192 // vpbroadcastd %xmm0,%ymm8
- .byte 196,193,84,194,192,1 // vcmpltps %ymm8,%ymm5,%ymm0
- .byte 196,98,125,24,21,188,48,0,0 // vbroadcastss 0x30bc(%rip),%ymm10 # 3e14 <_sk_callback_hsw+0x1c9>
.byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp)
+ .byte 184,0,0,0,63 // mov $0x3f000000,%eax
+ .byte 197,249,110,216 // vmovd %eax,%xmm3
+ .byte 196,98,125,88,195 // vpbroadcastd %xmm3,%ymm8
+ .byte 196,193,108,194,232,1 // vcmpltps %ymm8,%ymm2,%ymm5
+ .byte 196,98,125,24,21,182,48,0,0 // vbroadcastss 0x30b6(%rip),%ymm10 # 3e0c <_sk_callback_hsw+0x1c9>
.byte 196,193,116,88,218 // vaddps %ymm10,%ymm1,%ymm3
- .byte 197,228,89,221 // vmulps %ymm5,%ymm3,%ymm3
- .byte 197,244,88,229 // vaddps %ymm5,%ymm1,%ymm4
- .byte 196,226,117,188,229 // vfnmadd231ps %ymm5,%ymm1,%ymm4
- .byte 196,99,93,74,203,0 // vblendvps %ymm0,%ymm3,%ymm4,%ymm9
- .byte 196,226,125,24,13,157,48,0,0 // vbroadcastss 0x309d(%rip),%ymm1 # 3e1c <_sk_callback_hsw+0x1d1>
- .byte 197,236,88,241 // vaddps %ymm1,%ymm2,%ymm6
+ .byte 197,228,89,218 // vmulps %ymm2,%ymm3,%ymm3
+ .byte 197,244,88,226 // vaddps %ymm2,%ymm1,%ymm4
+ .byte 196,226,117,188,226 // vfnmadd231ps %ymm2,%ymm1,%ymm4
+ .byte 196,99,93,74,203,80 // vblendvps %ymm5,%ymm3,%ymm4,%ymm9
+ .byte 196,226,125,24,13,157,48,0,0 // vbroadcastss 0x309d(%rip),%ymm1 # 3e14 <_sk_callback_hsw+0x1d1>
+ .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1
.byte 65,184,0,0,0,0 // mov $0x0,%r8d
.byte 184,0,0,128,63 // mov $0x3f800000,%eax
- .byte 197,249,110,200 // vmovd %eax,%xmm1
- .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12
- .byte 197,156,194,206,1 // vcmpltps %ymm6,%ymm12,%ymm1
- .byte 196,98,125,24,45,123,48,0,0 // vbroadcastss 0x307b(%rip),%ymm13 # 3e20 <_sk_callback_hsw+0x1d5>
- .byte 196,193,76,88,221 // vaddps %ymm13,%ymm6,%ymm3
- .byte 196,227,77,74,203,16 // vblendvps %ymm1,%ymm3,%ymm6,%ymm1
- .byte 196,193,121,110,216 // vmovd %r8d,%xmm3
- .byte 196,98,125,88,251 // vpbroadcastd %xmm3,%ymm15
- .byte 196,193,76,194,223,1 // vcmpltps %ymm15,%ymm6,%ymm3
- .byte 196,193,76,88,226 // vaddps %ymm10,%ymm6,%ymm4
- .byte 196,227,117,74,196,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm0
- .byte 196,98,125,24,29,68,48,0,0 // vbroadcastss 0x3044(%rip),%ymm11 # 3e18 <_sk_callback_hsw+0x1cd>
- .byte 196,66,85,170,217 // vfmsub213ps %ymm9,%ymm5,%ymm11
+ .byte 197,249,110,216 // vmovd %eax,%xmm3
+ .byte 196,98,125,88,227 // vpbroadcastd %xmm3,%ymm12
+ .byte 197,156,194,217,1 // vcmpltps %ymm1,%ymm12,%ymm3
+ .byte 196,98,125,24,45,123,48,0,0 // vbroadcastss 0x307b(%rip),%ymm13 # 3e18 <_sk_callback_hsw+0x1d5>
+ .byte 196,193,116,88,229 // vaddps %ymm13,%ymm1,%ymm4
+ .byte 196,227,117,74,220,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm3
+ .byte 196,193,121,110,224 // vmovd %r8d,%xmm4
+ .byte 196,98,125,88,252 // vpbroadcastd %xmm4,%ymm15
+ .byte 196,193,116,194,231,1 // vcmpltps %ymm15,%ymm1,%ymm4
+ .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
+ .byte 196,227,101,74,241,64 // vblendvps %ymm4,%ymm1,%ymm3,%ymm6
+ .byte 196,98,125,24,29,68,48,0,0 // vbroadcastss 0x3044(%rip),%ymm11 # 3e10 <_sk_callback_hsw+0x1cd>
+ .byte 196,66,109,170,217 // vfmsub213ps %ymm9,%ymm2,%ymm11
.byte 196,193,52,92,203 // vsubps %ymm11,%ymm9,%ymm1
- .byte 196,226,125,24,29,61,48,0,0 // vbroadcastss 0x303d(%rip),%ymm3 # 3e24 <_sk_callback_hsw+0x1d9>
+ .byte 196,226,125,24,29,61,48,0,0 // vbroadcastss 0x303d(%rip),%ymm3 # 3e1c <_sk_callback_hsw+0x1d9>
.byte 197,116,89,243 // vmulps %ymm3,%ymm1,%ymm14
.byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
.byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax
.byte 197,249,110,200 // vmovd %eax,%xmm1
- .byte 196,226,125,88,225 // vpbroadcastd %xmm1,%ymm4
- .byte 196,226,125,24,29,32,48,0,0 // vbroadcastss 0x3020(%rip),%ymm3 # 3e28 <_sk_callback_hsw+0x1dd>
- .byte 197,228,92,200 // vsubps %ymm0,%ymm3,%ymm1
+ .byte 196,226,125,88,233 // vpbroadcastd %xmm1,%ymm5
+ .byte 196,226,125,24,37,32,48,0,0 // vbroadcastss 0x3020(%rip),%ymm4 # 3e20 <_sk_callback_hsw+0x1dd>
+ .byte 197,220,92,206 // vsubps %ymm6,%ymm4,%ymm1
.byte 196,194,13,168,203 // vfmadd213ps %ymm11,%ymm14,%ymm1
- .byte 197,252,194,252,1 // vcmpltps %ymm4,%ymm0,%ymm7
+ .byte 197,204,194,253,1 // vcmpltps %ymm5,%ymm6,%ymm7
.byte 196,227,37,74,201,112 // vblendvps %ymm7,%ymm1,%ymm11,%ymm1
- .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7
+ .byte 196,193,76,194,248,1 // vcmpltps %ymm8,%ymm6,%ymm7
.byte 196,195,117,74,249,112 // vblendvps %ymm7,%ymm9,%ymm1,%ymm7
.byte 196,193,121,110,200 // vmovd %r8d,%xmm1
- .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
- .byte 197,252,194,193,1 // vcmpltps %ymm1,%ymm0,%ymm0
+ .byte 196,226,125,88,217 // vpbroadcastd %xmm1,%ymm3
+ .byte 197,204,194,203,1 // vcmpltps %ymm3,%ymm6,%ymm1
.byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6
- .byte 196,227,69,74,198,0 // vblendvps %ymm0,%ymm6,%ymm7,%ymm0
- .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp)
- .byte 197,156,194,194,1 // vcmpltps %ymm2,%ymm12,%ymm0
- .byte 196,193,108,88,253 // vaddps %ymm13,%ymm2,%ymm7
- .byte 196,227,109,74,199,0 // vblendvps %ymm0,%ymm7,%ymm2,%ymm0
- .byte 196,193,108,194,255,1 // vcmpltps %ymm15,%ymm2,%ymm7
- .byte 196,193,108,88,242 // vaddps %ymm10,%ymm2,%ymm6
- .byte 196,227,125,74,198,112 // vblendvps %ymm7,%ymm6,%ymm0,%ymm0
- .byte 197,228,92,240 // vsubps %ymm0,%ymm3,%ymm6
+ .byte 196,227,69,74,206,16 // vblendvps %ymm1,%ymm6,%ymm7,%ymm1
+ .byte 197,252,17,76,36,128 // vmovups %ymm1,-0x80(%rsp)
+ .byte 197,156,194,200,1 // vcmpltps %ymm0,%ymm12,%ymm1
+ .byte 196,193,124,88,253 // vaddps %ymm13,%ymm0,%ymm7
+ .byte 196,227,125,74,207,16 // vblendvps %ymm1,%ymm7,%ymm0,%ymm1
+ .byte 196,193,124,194,255,1 // vcmpltps %ymm15,%ymm0,%ymm7
+ .byte 196,193,124,88,242 // vaddps %ymm10,%ymm0,%ymm6
+ .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1
+ .byte 197,220,92,241 // vsubps %ymm1,%ymm4,%ymm6
.byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6
- .byte 197,252,194,252,1 // vcmpltps %ymm4,%ymm0,%ymm7
+ .byte 197,244,194,253,1 // vcmpltps %ymm5,%ymm1,%ymm7
.byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6
- .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7
+ .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7
.byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6
- .byte 197,252,194,193,1 // vcmpltps %ymm1,%ymm0,%ymm0
- .byte 197,252,40,250 // vmovaps %ymm2,%ymm7
- .byte 196,194,13,168,251 // vfmadd213ps %ymm11,%ymm14,%ymm7
- .byte 196,227,77,74,247,0 // vblendvps %ymm0,%ymm7,%ymm6,%ymm6
- .byte 196,226,125,24,5,134,47,0,0 // vbroadcastss 0x2f86(%rip),%ymm0 # 3e2c <_sk_callback_hsw+0x1e1>
- .byte 197,236,88,192 // vaddps %ymm0,%ymm2,%ymm0
- .byte 197,156,194,208,1 // vcmpltps %ymm0,%ymm12,%ymm2
+ .byte 197,244,194,251,1 // vcmpltps %ymm3,%ymm1,%ymm7
+ .byte 196,194,13,168,203 // vfmadd213ps %ymm11,%ymm14,%ymm1
+ .byte 196,227,77,74,201,112 // vblendvps %ymm7,%ymm1,%ymm6,%ymm1
+ .byte 196,226,125,24,53,138,47,0,0 // vbroadcastss 0x2f8a(%rip),%ymm6 # 3e24 <_sk_callback_hsw+0x1e1>
+ .byte 197,252,88,198 // vaddps %ymm6,%ymm0,%ymm0
+ .byte 197,156,194,240,1 // vcmpltps %ymm0,%ymm12,%ymm6
.byte 196,193,124,88,253 // vaddps %ymm13,%ymm0,%ymm7
- .byte 196,227,125,74,215,32 // vblendvps %ymm2,%ymm7,%ymm0,%ymm2
+ .byte 196,227,125,74,247,96 // vblendvps %ymm6,%ymm7,%ymm0,%ymm6
.byte 196,193,124,194,255,1 // vcmpltps %ymm15,%ymm0,%ymm7
- .byte 196,65,124,88,210 // vaddps %ymm10,%ymm0,%ymm10
- .byte 196,195,109,74,210,112 // vblendvps %ymm7,%ymm10,%ymm2,%ymm2
- .byte 196,194,13,168,195 // vfmadd213ps %ymm11,%ymm14,%ymm0
- .byte 197,228,92,218 // vsubps %ymm2,%ymm3,%ymm3
- .byte 196,194,13,168,219 // vfmadd213ps %ymm11,%ymm14,%ymm3
- .byte 197,236,194,228,1 // vcmpltps %ymm4,%ymm2,%ymm4
- .byte 196,227,37,74,219,64 // vblendvps %ymm4,%ymm3,%ymm11,%ymm3
- .byte 196,193,108,194,224,1 // vcmpltps %ymm8,%ymm2,%ymm4
- .byte 196,195,101,74,217,64 // vblendvps %ymm4,%ymm9,%ymm3,%ymm3
- .byte 197,236,194,201,1 // vcmpltps %ymm1,%ymm2,%ymm1
- .byte 196,227,101,74,208,16 // vblendvps %ymm1,%ymm0,%ymm3,%ymm2
+ .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
+ .byte 196,227,77,74,192,112 // vblendvps %ymm7,%ymm0,%ymm6,%ymm0
+ .byte 197,220,92,224 // vsubps %ymm0,%ymm4,%ymm4
+ .byte 197,252,40,240 // vmovaps %ymm0,%ymm6
+ .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6
+ .byte 196,194,13,168,227 // vfmadd213ps %ymm11,%ymm14,%ymm4
+ .byte 197,252,194,237,1 // vcmpltps %ymm5,%ymm0,%ymm5
+ .byte 196,227,37,74,228,80 // vblendvps %ymm5,%ymm4,%ymm11,%ymm4
+ .byte 196,193,124,194,232,1 // vcmpltps %ymm8,%ymm0,%ymm5
+ .byte 196,195,93,74,225,80 // vblendvps %ymm5,%ymm9,%ymm4,%ymm4
+ .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0
+ .byte 196,227,93,74,222,0 // vblendvps %ymm0,%ymm6,%ymm4,%ymm3
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
- .byte 197,252,194,92,36,160,0 // vcmpeqps -0x60(%rsp),%ymm0,%ymm3
+ .byte 197,252,194,100,36,160,0 // vcmpeqps -0x60(%rsp),%ymm0,%ymm4
.byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0
- .byte 196,227,125,74,197,48 // vblendvps %ymm3,%ymm5,%ymm0,%ymm0
- .byte 196,227,77,74,205,48 // vblendvps %ymm3,%ymm5,%ymm6,%ymm1
- .byte 196,227,109,74,213,48 // vblendvps %ymm3,%ymm5,%ymm2,%ymm2
+ .byte 196,227,125,74,194,64 // vblendvps %ymm4,%ymm2,%ymm0,%ymm0
+ .byte 196,227,117,74,202,64 // vblendvps %ymm4,%ymm2,%ymm1,%ymm1
+ .byte 196,227,101,74,210,64 // vblendvps %ymm4,%ymm2,%ymm3,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,16,92,36,192 // vmovups -0x40(%rsp),%ymm3
.byte 197,252,16,100,36,224 // vmovups -0x20(%rsp),%ymm4
@@ -8356,11 +8357,11 @@ _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 fa1 <_sk_scale_u8_hsw+0x43>
+ .byte 117,51 // jne f99 <_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 196,98,125,24,13,171,46,0,0 // vbroadcastss 0x2eab(%rip),%ymm9 # 3e30 <_sk_callback_hsw+0x1e5>
+ .byte 196,98,125,24,13,171,46,0,0 // vbroadcastss 0x2eab(%rip),%ymm9 # 3e28 <_sk_callback_hsw+0x1e5>
.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
@@ -8378,9 +8379,9 @@ _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 fa9 <_sk_scale_u8_hsw+0x4b>
+ .byte 117,234 // jne fa1 <_sk_scale_u8_hsw+0x4b>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 235,172 // jmp f72 <_sk_scale_u8_hsw+0x14>
+ .byte 235,172 // jmp f6a <_sk_scale_u8_hsw+0x14>
HIDDEN _sk_lerp_1_float_hsw
.globl _sk_lerp_1_float_hsw
@@ -8408,11 +8409,11 @@ _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 104c <_sk_lerp_u8_hsw+0x57>
+ .byte 117,71 // jne 1044 <_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 196,98,125,24,13,24,46,0,0 // vbroadcastss 0x2e18(%rip),%ymm9 # 3e34 <_sk_callback_hsw+0x1e9>
+ .byte 196,98,125,24,13,24,46,0,0 // vbroadcastss 0x2e18(%rip),%ymm9 # 3e2c <_sk_callback_hsw+0x1e9>
.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
@@ -8434,9 +8435,9 @@ _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 1054 <_sk_lerp_u8_hsw+0x5f>
+ .byte 117,234 // jne 104c <_sk_lerp_u8_hsw+0x5f>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 235,152 // jmp 1009 <_sk_lerp_u8_hsw+0x14>
+ .byte 235,152 // jmp 1001 <_sk_lerp_u8_hsw+0x14>
HIDDEN _sk_lerp_565_hsw
.globl _sk_lerp_565_hsw
@@ -8445,23 +8446,23 @@ _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 1114 <_sk_lerp_565_hsw+0xa3>
+ .byte 15,133,149,0,0,0 // jne 110c <_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 196,98,125,88,5,165,45,0,0 // vpbroadcastd 0x2da5(%rip),%ymm8 # 3e38 <_sk_callback_hsw+0x1ed>
+ .byte 196,98,125,88,5,165,45,0,0 // vpbroadcastd 0x2da5(%rip),%ymm8 # 3e30 <_sk_callback_hsw+0x1ed>
.byte 196,65,101,219,192 // vpand %ymm8,%ymm3,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,150,45,0,0 // vbroadcastss 0x2d96(%rip),%ymm9 # 3e3c <_sk_callback_hsw+0x1f1>
+ .byte 196,98,125,24,13,150,45,0,0 // vbroadcastss 0x2d96(%rip),%ymm9 # 3e34 <_sk_callback_hsw+0x1f1>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,88,13,140,45,0,0 // vpbroadcastd 0x2d8c(%rip),%ymm9 # 3e40 <_sk_callback_hsw+0x1f5>
+ .byte 196,98,125,88,13,140,45,0,0 // vpbroadcastd 0x2d8c(%rip),%ymm9 # 3e38 <_sk_callback_hsw+0x1f5>
.byte 196,65,101,219,201 // vpand %ymm9,%ymm3,%ymm9
.byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9
- .byte 196,98,125,24,21,125,45,0,0 // vbroadcastss 0x2d7d(%rip),%ymm10 # 3e44 <_sk_callback_hsw+0x1f9>
+ .byte 196,98,125,24,21,125,45,0,0 // vbroadcastss 0x2d7d(%rip),%ymm10 # 3e3c <_sk_callback_hsw+0x1f9>
.byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,88,21,115,45,0,0 // vpbroadcastd 0x2d73(%rip),%ymm10 # 3e48 <_sk_callback_hsw+0x1fd>
+ .byte 196,98,125,88,21,115,45,0,0 // vpbroadcastd 0x2d73(%rip),%ymm10 # 3e40 <_sk_callback_hsw+0x1fd>
.byte 196,193,101,219,218 // vpand %ymm10,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,21,101,45,0,0 // vbroadcastss 0x2d65(%rip),%ymm10 # 3e4c <_sk_callback_hsw+0x201>
+ .byte 196,98,125,24,21,101,45,0,0 // vbroadcastss 0x2d65(%rip),%ymm10 # 3e44 <_sk_callback_hsw+0x201>
.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
@@ -8470,16 +8471,16 @@ _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 196,226,125,24,29,62,45,0,0 // vbroadcastss 0x2d3e(%rip),%ymm3 # 3e50 <_sk_callback_hsw+0x205>
+ .byte 196,226,125,24,29,62,45,0,0 // vbroadcastss 0x2d3e(%rip),%ymm3 # 3e48 <_sk_callback_hsw+0x205>
.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 1085 <_sk_lerp_565_hsw+0x14>
+ .byte 15,135,89,255,255,255 // ja 107d <_sk_lerp_565_hsw+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 1180 <_sk_lerp_565_hsw+0x10f>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 1178 <_sk_lerp_565_hsw+0x10f>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -8491,7 +8492,7 @@ _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 1085 <_sk_lerp_565_hsw+0x14>
+ .byte 233,5,255,255,255 // jmpq 107d <_sk_lerp_565_hsw+0x14>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -8525,23 +8526,23 @@ _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 121a <_sk_load_tables_hsw+0x7e>
+ .byte 117,105 // jne 1212 <_sk_load_tables_hsw+0x7e>
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
- .byte 197,229,219,13,34,47,0,0 // vpand 0x2f22(%rip),%ymm3,%ymm1 # 40e0 <_sk_callback_hsw+0x495>
+ .byte 197,229,219,13,42,47,0,0 // vpand 0x2f2a(%rip),%ymm3,%ymm1 # 40e0 <_sk_callback_hsw+0x49d>
.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 196,226,101,0,21,34,47,0,0 // vpshufb 0x2f22(%rip),%ymm3,%ymm2 # 4100 <_sk_callback_hsw+0x4b5>
+ .byte 196,226,101,0,21,42,47,0,0 // vpshufb 0x2f2a(%rip),%ymm3,%ymm2 # 4100 <_sk_callback_hsw+0x4bd>
.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 196,98,101,0,13,42,47,0,0 // vpshufb 0x2f2a(%rip),%ymm3,%ymm9 # 4120 <_sk_callback_hsw+0x4d5>
+ .byte 196,98,101,0,13,50,47,0,0 // vpshufb 0x2f32(%rip),%ymm3,%ymm9 # 4120 <_sk_callback_hsw+0x4dd>
.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 196,98,125,24,5,70,44,0,0 // vbroadcastss 0x2c46(%rip),%ymm8 # 3e54 <_sk_callback_hsw+0x209>
+ .byte 196,98,125,24,5,70,44,0,0 // vbroadcastss 0x2c46(%rip),%ymm8 # 3e4c <_sk_callback_hsw+0x209>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,193 // mov %r8,%rcx
@@ -8554,7 +8555,7 @@ _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 11b6 <_sk_load_tables_hsw+0x1a>
+ .byte 233,115,255,255,255 // jmpq 11ae <_sk_load_tables_hsw+0x1a>
HIDDEN _sk_load_tables_u16_be_hsw
.globl _sk_load_tables_u16_be_hsw
@@ -8564,7 +8565,7 @@ _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 1322 <_sk_load_tables_u16_be_hsw+0xdf>
+ .byte 15,133,201,0,0,0 // jne 131a <_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
@@ -8580,7 +8581,7 @@ _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 197,121,111,21,182,47,0,0 // vmovdqa 0x2fb6(%rip),%xmm10 # 4260 <_sk_callback_hsw+0x615>
+ .byte 197,121,111,21,190,47,0,0 // vmovdqa 0x2fbe(%rip),%xmm10 # 4260 <_sk_callback_hsw+0x61d>
.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
@@ -8602,36 +8603,36 @@ _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 196,98,125,24,5,63,43,0,0 // vbroadcastss 0x2b3f(%rip),%ymm8 # 3e58 <_sk_callback_hsw+0x20d>
+ .byte 196,98,125,24,5,63,43,0,0 // vbroadcastss 0x2b3f(%rip),%ymm8 # 3e50 <_sk_callback_hsw+0x20d>
.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 1388 <_sk_load_tables_u16_be_hsw+0x145>
+ .byte 116,85 // je 1380 <_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 1388 <_sk_load_tables_u16_be_hsw+0x145>
+ .byte 114,72 // jb 1380 <_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 1395 <_sk_load_tables_u16_be_hsw+0x152>
+ .byte 116,72 // je 138d <_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 1395 <_sk_load_tables_u16_be_hsw+0x152>
+ .byte 114,59 // jb 138d <_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 1274 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 15,132,9,255,255,255 // je 126c <_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 1274 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 15,130,248,254,255,255 // jb 126c <_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 1274 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,236,254,255,255 // jmpq 126c <_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 1274 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,223,254,255,255 // jmpq 126c <_sk_load_tables_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,214,254,255,255 // jmpq 1274 <_sk_load_tables_u16_be_hsw+0x31>
+ .byte 233,214,254,255,255 // jmpq 126c <_sk_load_tables_u16_be_hsw+0x31>
HIDDEN _sk_load_tables_rgb_u16_be_hsw
.globl _sk_load_tables_rgb_u16_be_hsw
@@ -8641,7 +8642,7 @@ _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 1471 <_sk_load_tables_rgb_u16_be_hsw+0xd3>
+ .byte 15,133,193,0,0,0 // jne 1469 <_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
@@ -8662,7 +8663,7 @@ _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 197,121,111,13,86,46,0,0 // vmovdqa 0x2e56(%rip),%xmm9 # 4270 <_sk_callback_hsw+0x625>
+ .byte 197,121,111,13,94,46,0,0 // vmovdqa 0x2e5e(%rip),%xmm9 # 4270 <_sk_callback_hsw+0x62d>
.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
@@ -8679,41 +8680,41 @@ _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 196,226,125,24,29,237,41,0,0 // vbroadcastss 0x29ed(%rip),%ymm3 # 3e5c <_sk_callback_hsw+0x211>
+ .byte 196,226,125,24,29,237,41,0,0 // vbroadcastss 0x29ed(%rip),%ymm3 # 3e54 <_sk_callback_hsw+0x211>
.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 148a <_sk_load_tables_rgb_u16_be_hsw+0xec>
- .byte 233,90,255,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,5 // jne 1482 <_sk_load_tables_rgb_u16_be_hsw+0xec>
+ .byte 233,90,255,255,255 // jmpq 13dc <_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 14b9 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
+ .byte 114,26 // jb 14b1 <_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 14be <_sk_load_tables_rgb_u16_be_hsw+0x120>
- .byte 233,43,255,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,38,255,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 14b6 <_sk_load_tables_rgb_u16_be_hsw+0x120>
+ .byte 233,43,255,255,255 // jmpq 13dc <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,38,255,255,255 // jmpq 13dc <_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 14ed <_sk_load_tables_rgb_u16_be_hsw+0x14f>
+ .byte 114,26 // jb 14e5 <_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 14f2 <_sk_load_tables_rgb_u16_be_hsw+0x154>
- .byte 233,247,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,242,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 14ea <_sk_load_tables_rgb_u16_be_hsw+0x154>
+ .byte 233,247,254,255,255 // jmpq 13dc <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,242,254,255,255 // jmpq 13dc <_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 151b <_sk_load_tables_rgb_u16_be_hsw+0x17d>
+ .byte 114,20 // jb 1513 <_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 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
- .byte 233,196,254,255,255 // jmpq 13e4 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,201,254,255,255 // jmpq 13dc <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ .byte 233,196,254,255,255 // jmpq 13dc <_sk_load_tables_rgb_u16_be_hsw+0x46>
HIDDEN _sk_byte_tables_hsw
.globl _sk_byte_tables_hsw
@@ -8726,7 +8727,7 @@ _sk_byte_tables_hsw:
.byte 65,84 // push %r12
.byte 83 // push %rbx
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,43,41,0,0 // vbroadcastss 0x292b(%rip),%ymm8 # 3e60 <_sk_callback_hsw+0x215>
+ .byte 196,98,125,24,5,43,41,0,0 // vbroadcastss 0x292b(%rip),%ymm8 # 3e58 <_sk_callback_hsw+0x215>
.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
@@ -8763,7 +8764,7 @@ _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 196,98,125,24,13,124,40,0,0 // vbroadcastss 0x287c(%rip),%ymm9 # 3e64 <_sk_callback_hsw+0x219>
+ .byte 196,98,125,24,13,124,40,0,0 // vbroadcastss 0x287c(%rip),%ymm9 # 3e5c <_sk_callback_hsw+0x219>
.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
@@ -8924,7 +8925,7 @@ _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 196,98,125,24,13,181,37,0,0 // vbroadcastss 0x25b5(%rip),%ymm9 # 3e68 <_sk_callback_hsw+0x21d>
+ .byte 196,98,125,24,13,181,37,0,0 // vbroadcastss 0x25b5(%rip),%ymm9 # 3e60 <_sk_callback_hsw+0x21d>
.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
@@ -9087,33 +9088,33 @@ _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 196,98,125,24,37,104,35,0,0 // vbroadcastss 0x2368(%rip),%ymm12 # 3e6c <_sk_callback_hsw+0x221>
- .byte 196,98,125,24,45,99,35,0,0 // vbroadcastss 0x2363(%rip),%ymm13 # 3e70 <_sk_callback_hsw+0x225>
+ .byte 196,98,125,24,37,104,35,0,0 // vbroadcastss 0x2368(%rip),%ymm12 # 3e64 <_sk_callback_hsw+0x221>
+ .byte 196,98,125,24,45,99,35,0,0 // vbroadcastss 0x2363(%rip),%ymm13 # 3e68 <_sk_callback_hsw+0x225>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,89,35,0,0 // vbroadcastss 0x2359(%rip),%ymm13 # 3e74 <_sk_callback_hsw+0x229>
+ .byte 196,98,125,24,45,89,35,0,0 // vbroadcastss 0x2359(%rip),%ymm13 # 3e6c <_sk_callback_hsw+0x229>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,79,35,0,0 // vbroadcastss 0x234f(%rip),%ymm13 # 3e78 <_sk_callback_hsw+0x22d>
+ .byte 196,98,125,24,45,79,35,0,0 // vbroadcastss 0x234f(%rip),%ymm13 # 3e70 <_sk_callback_hsw+0x22d>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,69,35,0,0 // vbroadcastss 0x2345(%rip),%ymm11 # 3e7c <_sk_callback_hsw+0x231>
+ .byte 196,98,125,24,29,69,35,0,0 // vbroadcastss 0x2345(%rip),%ymm11 # 3e74 <_sk_callback_hsw+0x231>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,59,35,0,0 // vbroadcastss 0x233b(%rip),%ymm12 # 3e80 <_sk_callback_hsw+0x235>
+ .byte 196,98,125,24,37,59,35,0,0 // vbroadcastss 0x233b(%rip),%ymm12 # 3e78 <_sk_callback_hsw+0x235>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,49,35,0,0 // vbroadcastss 0x2331(%rip),%ymm12 # 3e84 <_sk_callback_hsw+0x239>
+ .byte 196,98,125,24,37,49,35,0,0 // vbroadcastss 0x2331(%rip),%ymm12 # 3e7c <_sk_callback_hsw+0x239>
.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 196,98,125,24,29,18,35,0,0 // vbroadcastss 0x2312(%rip),%ymm11 # 3e88 <_sk_callback_hsw+0x23d>
+ .byte 196,98,125,24,29,18,35,0,0 // vbroadcastss 0x2312(%rip),%ymm11 # 3e80 <_sk_callback_hsw+0x23d>
.byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
- .byte 196,98,125,24,29,8,35,0,0 // vbroadcastss 0x2308(%rip),%ymm11 # 3e8c <_sk_callback_hsw+0x241>
+ .byte 196,98,125,24,29,8,35,0,0 // vbroadcastss 0x2308(%rip),%ymm11 # 3e84 <_sk_callback_hsw+0x241>
.byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11
- .byte 196,226,125,24,5,254,34,0,0 // vbroadcastss 0x22fe(%rip),%ymm0 # 3e90 <_sk_callback_hsw+0x245>
+ .byte 196,226,125,24,5,254,34,0,0 // vbroadcastss 0x22fe(%rip),%ymm0 # 3e88 <_sk_callback_hsw+0x245>
.byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0
- .byte 196,98,125,24,21,244,34,0,0 // vbroadcastss 0x22f4(%rip),%ymm10 # 3e94 <_sk_callback_hsw+0x249>
+ .byte 196,98,125,24,21,244,34,0,0 // vbroadcastss 0x22f4(%rip),%ymm10 # 3e8c <_sk_callback_hsw+0x249>
.byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0
.byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
- .byte 196,98,125,24,21,231,34,0,0 // vbroadcastss 0x22e7(%rip),%ymm10 # 3e98 <_sk_callback_hsw+0x24d>
+ .byte 196,98,125,24,21,231,34,0,0 // vbroadcastss 0x22e7(%rip),%ymm10 # 3e90 <_sk_callback_hsw+0x24d>
.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
@@ -9121,7 +9122,7 @@ _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 196,98,125,24,5,190,34,0,0 // vbroadcastss 0x22be(%rip),%ymm8 # 3e9c <_sk_callback_hsw+0x251>
+ .byte 196,98,125,24,5,190,34,0,0 // vbroadcastss 0x22be(%rip),%ymm8 # 3e94 <_sk_callback_hsw+0x251>
.byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9141,33 +9142,33 @@ _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 196,98,125,24,37,118,34,0,0 // vbroadcastss 0x2276(%rip),%ymm12 # 3ea0 <_sk_callback_hsw+0x255>
- .byte 196,98,125,24,45,113,34,0,0 // vbroadcastss 0x2271(%rip),%ymm13 # 3ea4 <_sk_callback_hsw+0x259>
+ .byte 196,98,125,24,37,118,34,0,0 // vbroadcastss 0x2276(%rip),%ymm12 # 3e98 <_sk_callback_hsw+0x255>
+ .byte 196,98,125,24,45,113,34,0,0 // vbroadcastss 0x2271(%rip),%ymm13 # 3e9c <_sk_callback_hsw+0x259>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,103,34,0,0 // vbroadcastss 0x2267(%rip),%ymm13 # 3ea8 <_sk_callback_hsw+0x25d>
+ .byte 196,98,125,24,45,103,34,0,0 // vbroadcastss 0x2267(%rip),%ymm13 # 3ea0 <_sk_callback_hsw+0x25d>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,93,34,0,0 // vbroadcastss 0x225d(%rip),%ymm13 # 3eac <_sk_callback_hsw+0x261>
+ .byte 196,98,125,24,45,93,34,0,0 // vbroadcastss 0x225d(%rip),%ymm13 # 3ea4 <_sk_callback_hsw+0x261>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,83,34,0,0 // vbroadcastss 0x2253(%rip),%ymm11 # 3eb0 <_sk_callback_hsw+0x265>
+ .byte 196,98,125,24,29,83,34,0,0 // vbroadcastss 0x2253(%rip),%ymm11 # 3ea8 <_sk_callback_hsw+0x265>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,73,34,0,0 // vbroadcastss 0x2249(%rip),%ymm12 # 3eb4 <_sk_callback_hsw+0x269>
+ .byte 196,98,125,24,37,73,34,0,0 // vbroadcastss 0x2249(%rip),%ymm12 # 3eac <_sk_callback_hsw+0x269>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,63,34,0,0 // vbroadcastss 0x223f(%rip),%ymm12 # 3eb8 <_sk_callback_hsw+0x26d>
+ .byte 196,98,125,24,37,63,34,0,0 // vbroadcastss 0x223f(%rip),%ymm12 # 3eb0 <_sk_callback_hsw+0x26d>
.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 196,98,125,24,29,32,34,0,0 // vbroadcastss 0x2220(%rip),%ymm11 # 3ebc <_sk_callback_hsw+0x271>
+ .byte 196,98,125,24,29,32,34,0,0 // vbroadcastss 0x2220(%rip),%ymm11 # 3eb4 <_sk_callback_hsw+0x271>
.byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
- .byte 196,98,125,24,29,22,34,0,0 // vbroadcastss 0x2216(%rip),%ymm11 # 3ec0 <_sk_callback_hsw+0x275>
+ .byte 196,98,125,24,29,22,34,0,0 // vbroadcastss 0x2216(%rip),%ymm11 # 3eb8 <_sk_callback_hsw+0x275>
.byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11
- .byte 196,226,125,24,13,12,34,0,0 // vbroadcastss 0x220c(%rip),%ymm1 # 3ec4 <_sk_callback_hsw+0x279>
+ .byte 196,226,125,24,13,12,34,0,0 // vbroadcastss 0x220c(%rip),%ymm1 # 3ebc <_sk_callback_hsw+0x279>
.byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1
- .byte 196,98,125,24,21,2,34,0,0 // vbroadcastss 0x2202(%rip),%ymm10 # 3ec8 <_sk_callback_hsw+0x27d>
+ .byte 196,98,125,24,21,2,34,0,0 // vbroadcastss 0x2202(%rip),%ymm10 # 3ec0 <_sk_callback_hsw+0x27d>
.byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
.byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
- .byte 196,98,125,24,21,245,33,0,0 // vbroadcastss 0x21f5(%rip),%ymm10 # 3ecc <_sk_callback_hsw+0x281>
+ .byte 196,98,125,24,21,245,33,0,0 // vbroadcastss 0x21f5(%rip),%ymm10 # 3ec4 <_sk_callback_hsw+0x281>
.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
@@ -9175,7 +9176,7 @@ _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 196,98,125,24,5,204,33,0,0 // vbroadcastss 0x21cc(%rip),%ymm8 # 3ed0 <_sk_callback_hsw+0x285>
+ .byte 196,98,125,24,5,204,33,0,0 // vbroadcastss 0x21cc(%rip),%ymm8 # 3ec8 <_sk_callback_hsw+0x285>
.byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9195,33 +9196,33 @@ _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 196,98,125,24,37,132,33,0,0 // vbroadcastss 0x2184(%rip),%ymm12 # 3ed4 <_sk_callback_hsw+0x289>
- .byte 196,98,125,24,45,127,33,0,0 // vbroadcastss 0x217f(%rip),%ymm13 # 3ed8 <_sk_callback_hsw+0x28d>
+ .byte 196,98,125,24,37,132,33,0,0 // vbroadcastss 0x2184(%rip),%ymm12 # 3ecc <_sk_callback_hsw+0x289>
+ .byte 196,98,125,24,45,127,33,0,0 // vbroadcastss 0x217f(%rip),%ymm13 # 3ed0 <_sk_callback_hsw+0x28d>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,117,33,0,0 // vbroadcastss 0x2175(%rip),%ymm13 # 3edc <_sk_callback_hsw+0x291>
+ .byte 196,98,125,24,45,117,33,0,0 // vbroadcastss 0x2175(%rip),%ymm13 # 3ed4 <_sk_callback_hsw+0x291>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,107,33,0,0 // vbroadcastss 0x216b(%rip),%ymm13 # 3ee0 <_sk_callback_hsw+0x295>
+ .byte 196,98,125,24,45,107,33,0,0 // vbroadcastss 0x216b(%rip),%ymm13 # 3ed8 <_sk_callback_hsw+0x295>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,97,33,0,0 // vbroadcastss 0x2161(%rip),%ymm11 # 3ee4 <_sk_callback_hsw+0x299>
+ .byte 196,98,125,24,29,97,33,0,0 // vbroadcastss 0x2161(%rip),%ymm11 # 3edc <_sk_callback_hsw+0x299>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,87,33,0,0 // vbroadcastss 0x2157(%rip),%ymm12 # 3ee8 <_sk_callback_hsw+0x29d>
+ .byte 196,98,125,24,37,87,33,0,0 // vbroadcastss 0x2157(%rip),%ymm12 # 3ee0 <_sk_callback_hsw+0x29d>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,77,33,0,0 // vbroadcastss 0x214d(%rip),%ymm12 # 3eec <_sk_callback_hsw+0x2a1>
+ .byte 196,98,125,24,37,77,33,0,0 // vbroadcastss 0x214d(%rip),%ymm12 # 3ee4 <_sk_callback_hsw+0x2a1>
.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 196,98,125,24,29,46,33,0,0 // vbroadcastss 0x212e(%rip),%ymm11 # 3ef0 <_sk_callback_hsw+0x2a5>
+ .byte 196,98,125,24,29,46,33,0,0 // vbroadcastss 0x212e(%rip),%ymm11 # 3ee8 <_sk_callback_hsw+0x2a5>
.byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
- .byte 196,98,125,24,29,36,33,0,0 // vbroadcastss 0x2124(%rip),%ymm11 # 3ef4 <_sk_callback_hsw+0x2a9>
+ .byte 196,98,125,24,29,36,33,0,0 // vbroadcastss 0x2124(%rip),%ymm11 # 3eec <_sk_callback_hsw+0x2a9>
.byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11
- .byte 196,226,125,24,21,26,33,0,0 // vbroadcastss 0x211a(%rip),%ymm2 # 3ef8 <_sk_callback_hsw+0x2ad>
+ .byte 196,226,125,24,21,26,33,0,0 // vbroadcastss 0x211a(%rip),%ymm2 # 3ef0 <_sk_callback_hsw+0x2ad>
.byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2
- .byte 196,98,125,24,21,16,33,0,0 // vbroadcastss 0x2110(%rip),%ymm10 # 3efc <_sk_callback_hsw+0x2b1>
+ .byte 196,98,125,24,21,16,33,0,0 // vbroadcastss 0x2110(%rip),%ymm10 # 3ef4 <_sk_callback_hsw+0x2b1>
.byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2
.byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
- .byte 196,98,125,24,21,3,33,0,0 // vbroadcastss 0x2103(%rip),%ymm10 # 3f00 <_sk_callback_hsw+0x2b5>
+ .byte 196,98,125,24,21,3,33,0,0 // vbroadcastss 0x2103(%rip),%ymm10 # 3ef8 <_sk_callback_hsw+0x2b5>
.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
@@ -9229,7 +9230,7 @@ _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 196,98,125,24,5,218,32,0,0 // vbroadcastss 0x20da(%rip),%ymm8 # 3f04 <_sk_callback_hsw+0x2b9>
+ .byte 196,98,125,24,5,218,32,0,0 // vbroadcastss 0x20da(%rip),%ymm8 # 3efc <_sk_callback_hsw+0x2b9>
.byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9249,33 +9250,33 @@ _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 196,98,125,24,37,146,32,0,0 // vbroadcastss 0x2092(%rip),%ymm12 # 3f08 <_sk_callback_hsw+0x2bd>
- .byte 196,98,125,24,45,141,32,0,0 // vbroadcastss 0x208d(%rip),%ymm13 # 3f0c <_sk_callback_hsw+0x2c1>
+ .byte 196,98,125,24,37,146,32,0,0 // vbroadcastss 0x2092(%rip),%ymm12 # 3f00 <_sk_callback_hsw+0x2bd>
+ .byte 196,98,125,24,45,141,32,0,0 // vbroadcastss 0x208d(%rip),%ymm13 # 3f04 <_sk_callback_hsw+0x2c1>
.byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,131,32,0,0 // vbroadcastss 0x2083(%rip),%ymm13 # 3f10 <_sk_callback_hsw+0x2c5>
+ .byte 196,98,125,24,45,131,32,0,0 // vbroadcastss 0x2083(%rip),%ymm13 # 3f08 <_sk_callback_hsw+0x2c5>
.byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
- .byte 196,98,125,24,45,121,32,0,0 // vbroadcastss 0x2079(%rip),%ymm13 # 3f14 <_sk_callback_hsw+0x2c9>
+ .byte 196,98,125,24,45,121,32,0,0 // vbroadcastss 0x2079(%rip),%ymm13 # 3f0c <_sk_callback_hsw+0x2c9>
.byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
- .byte 196,98,125,24,29,111,32,0,0 // vbroadcastss 0x206f(%rip),%ymm11 # 3f18 <_sk_callback_hsw+0x2cd>
+ .byte 196,98,125,24,29,111,32,0,0 // vbroadcastss 0x206f(%rip),%ymm11 # 3f10 <_sk_callback_hsw+0x2cd>
.byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
- .byte 196,98,125,24,37,101,32,0,0 // vbroadcastss 0x2065(%rip),%ymm12 # 3f1c <_sk_callback_hsw+0x2d1>
+ .byte 196,98,125,24,37,101,32,0,0 // vbroadcastss 0x2065(%rip),%ymm12 # 3f14 <_sk_callback_hsw+0x2d1>
.byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
- .byte 196,98,125,24,37,91,32,0,0 // vbroadcastss 0x205b(%rip),%ymm12 # 3f20 <_sk_callback_hsw+0x2d5>
+ .byte 196,98,125,24,37,91,32,0,0 // vbroadcastss 0x205b(%rip),%ymm12 # 3f18 <_sk_callback_hsw+0x2d5>
.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 196,98,125,24,29,60,32,0,0 // vbroadcastss 0x203c(%rip),%ymm11 # 3f24 <_sk_callback_hsw+0x2d9>
+ .byte 196,98,125,24,29,60,32,0,0 // vbroadcastss 0x203c(%rip),%ymm11 # 3f1c <_sk_callback_hsw+0x2d9>
.byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
- .byte 196,98,125,24,29,50,32,0,0 // vbroadcastss 0x2032(%rip),%ymm11 # 3f28 <_sk_callback_hsw+0x2dd>
+ .byte 196,98,125,24,29,50,32,0,0 // vbroadcastss 0x2032(%rip),%ymm11 # 3f20 <_sk_callback_hsw+0x2dd>
.byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11
- .byte 196,226,125,24,29,40,32,0,0 // vbroadcastss 0x2028(%rip),%ymm3 # 3f2c <_sk_callback_hsw+0x2e1>
+ .byte 196,226,125,24,29,40,32,0,0 // vbroadcastss 0x2028(%rip),%ymm3 # 3f24 <_sk_callback_hsw+0x2e1>
.byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3
- .byte 196,98,125,24,21,30,32,0,0 // vbroadcastss 0x201e(%rip),%ymm10 # 3f30 <_sk_callback_hsw+0x2e5>
+ .byte 196,98,125,24,21,30,32,0,0 // vbroadcastss 0x201e(%rip),%ymm10 # 3f28 <_sk_callback_hsw+0x2e5>
.byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3
.byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
- .byte 196,98,125,24,21,17,32,0,0 // vbroadcastss 0x2011(%rip),%ymm10 # 3f34 <_sk_callback_hsw+0x2e9>
+ .byte 196,98,125,24,21,17,32,0,0 // vbroadcastss 0x2011(%rip),%ymm10 # 3f2c <_sk_callback_hsw+0x2e9>
.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
@@ -9283,7 +9284,7 @@ _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 196,98,125,24,5,232,31,0,0 // vbroadcastss 0x1fe8(%rip),%ymm8 # 3f38 <_sk_callback_hsw+0x2ed>
+ .byte 196,98,125,24,5,232,31,0,0 // vbroadcastss 0x1fe8(%rip),%ymm8 # 3f30 <_sk_callback_hsw+0x2ed>
.byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9292,26 +9293,26 @@ HIDDEN _sk_lab_to_xyz_hsw
.globl _sk_lab_to_xyz_hsw
FUNCTION(_sk_lab_to_xyz_hsw)
_sk_lab_to_xyz_hsw:
- .byte 196,98,125,24,5,218,31,0,0 // vbroadcastss 0x1fda(%rip),%ymm8 # 3f3c <_sk_callback_hsw+0x2f1>
- .byte 196,98,125,24,13,213,31,0,0 // vbroadcastss 0x1fd5(%rip),%ymm9 # 3f40 <_sk_callback_hsw+0x2f5>
- .byte 196,98,125,24,21,208,31,0,0 // vbroadcastss 0x1fd0(%rip),%ymm10 # 3f44 <_sk_callback_hsw+0x2f9>
+ .byte 196,98,125,24,5,218,31,0,0 // vbroadcastss 0x1fda(%rip),%ymm8 # 3f34 <_sk_callback_hsw+0x2f1>
+ .byte 196,98,125,24,13,213,31,0,0 // vbroadcastss 0x1fd5(%rip),%ymm9 # 3f38 <_sk_callback_hsw+0x2f5>
+ .byte 196,98,125,24,21,208,31,0,0 // vbroadcastss 0x1fd0(%rip),%ymm10 # 3f3c <_sk_callback_hsw+0x2f9>
.byte 196,194,53,168,202 // vfmadd213ps %ymm10,%ymm9,%ymm1
.byte 196,194,53,168,210 // vfmadd213ps %ymm10,%ymm9,%ymm2
- .byte 196,98,125,24,13,193,31,0,0 // vbroadcastss 0x1fc1(%rip),%ymm9 # 3f48 <_sk_callback_hsw+0x2fd>
+ .byte 196,98,125,24,13,193,31,0,0 // vbroadcastss 0x1fc1(%rip),%ymm9 # 3f40 <_sk_callback_hsw+0x2fd>
.byte 196,66,125,184,200 // vfmadd231ps %ymm8,%ymm0,%ymm9
- .byte 196,226,125,24,5,183,31,0,0 // vbroadcastss 0x1fb7(%rip),%ymm0 # 3f4c <_sk_callback_hsw+0x301>
+ .byte 196,226,125,24,5,183,31,0,0 // vbroadcastss 0x1fb7(%rip),%ymm0 # 3f44 <_sk_callback_hsw+0x301>
.byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
- .byte 196,98,125,24,5,174,31,0,0 // vbroadcastss 0x1fae(%rip),%ymm8 # 3f50 <_sk_callback_hsw+0x305>
+ .byte 196,98,125,24,5,174,31,0,0 // vbroadcastss 0x1fae(%rip),%ymm8 # 3f48 <_sk_callback_hsw+0x305>
.byte 196,98,117,168,192 // vfmadd213ps %ymm0,%ymm1,%ymm8
- .byte 196,98,125,24,13,164,31,0,0 // vbroadcastss 0x1fa4(%rip),%ymm9 # 3f54 <_sk_callback_hsw+0x309>
+ .byte 196,98,125,24,13,164,31,0,0 // vbroadcastss 0x1fa4(%rip),%ymm9 # 3f4c <_sk_callback_hsw+0x309>
.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 196,226,125,24,21,145,31,0,0 // vbroadcastss 0x1f91(%rip),%ymm2 # 3f58 <_sk_callback_hsw+0x30d>
+ .byte 196,226,125,24,21,145,31,0,0 // vbroadcastss 0x1f91(%rip),%ymm2 # 3f50 <_sk_callback_hsw+0x30d>
.byte 197,108,194,209,1 // vcmpltps %ymm1,%ymm2,%ymm10
- .byte 196,98,125,24,29,135,31,0,0 // vbroadcastss 0x1f87(%rip),%ymm11 # 3f5c <_sk_callback_hsw+0x311>
+ .byte 196,98,125,24,29,135,31,0,0 // vbroadcastss 0x1f87(%rip),%ymm11 # 3f54 <_sk_callback_hsw+0x311>
.byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
- .byte 196,98,125,24,37,125,31,0,0 // vbroadcastss 0x1f7d(%rip),%ymm12 # 3f60 <_sk_callback_hsw+0x315>
+ .byte 196,98,125,24,37,125,31,0,0 // vbroadcastss 0x1f7d(%rip),%ymm12 # 3f58 <_sk_callback_hsw+0x315>
.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
@@ -9326,9 +9327,9 @@ _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 196,226,125,24,5,50,31,0,0 // vbroadcastss 0x1f32(%rip),%ymm0 # 3f64 <_sk_callback_hsw+0x319>
+ .byte 196,226,125,24,5,50,31,0,0 // vbroadcastss 0x1f32(%rip),%ymm0 # 3f5c <_sk_callback_hsw+0x319>
.byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
- .byte 196,98,125,24,5,41,31,0,0 // vbroadcastss 0x1f29(%rip),%ymm8 # 3f68 <_sk_callback_hsw+0x31d>
+ .byte 196,98,125,24,5,41,31,0,0 // vbroadcastss 0x1f29(%rip),%ymm8 # 3f60 <_sk_callback_hsw+0x31d>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9342,11 +9343,11 @@ _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 2085 <_sk_load_a8_hsw+0x3d>
+ .byte 117,45 // jne 207d <_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 196,226,125,24,13,254,30,0,0 // vbroadcastss 0x1efe(%rip),%ymm1 # 3f6c <_sk_callback_hsw+0x321>
+ .byte 196,226,125,24,13,254,30,0,0 // vbroadcastss 0x1efe(%rip),%ymm1 # 3f64 <_sk_callback_hsw+0x321>
.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
@@ -9363,9 +9364,9 @@ _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 208d <_sk_load_a8_hsw+0x45>
+ .byte 117,234 // jne 2085 <_sk_load_a8_hsw+0x45>
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
- .byte 235,178 // jmp 205c <_sk_load_a8_hsw+0x14>
+ .byte 235,178 // jmp 2054 <_sk_load_a8_hsw+0x14>
HIDDEN _sk_gather_a8_hsw
.globl _sk_gather_a8_hsw
@@ -9411,7 +9412,7 @@ _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 196,226,125,24,13,9,30,0,0 // vbroadcastss 0x1e09(%rip),%ymm1 # 3f70 <_sk_callback_hsw+0x325>
+ .byte 196,226,125,24,13,9,30,0,0 // vbroadcastss 0x1e09(%rip),%ymm1 # 3f68 <_sk_callback_hsw+0x325>
.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
@@ -9429,14 +9430,14 @@ FUNCTION(_sk_store_a8_hsw)
_sk_store_a8_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,228,29,0,0 // vbroadcastss 0x1de4(%rip),%ymm8 # 3f74 <_sk_callback_hsw+0x329>
+ .byte 196,98,125,24,5,228,29,0,0 // vbroadcastss 0x1de4(%rip),%ymm8 # 3f6c <_sk_callback_hsw+0x329>
.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 21b9 <_sk_store_a8_hsw+0x37>
+ .byte 117,10 // jne 21b1 <_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
@@ -9444,10 +9445,10 @@ _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 21b5 <_sk_store_a8_hsw+0x33>
+ .byte 119,236 // ja 21ad <_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 # 221c <_sk_store_a8_hsw+0x9a>
+ .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 2214 <_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
@@ -9458,7 +9459,7 @@ _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 21b5 <_sk_store_a8_hsw+0x33>
+ .byte 235,154 // jmp 21ad <_sk_store_a8_hsw+0x33>
.byte 144 // nop
.byte 246,255 // idiv %bh
.byte 255 // (bad)
@@ -9492,14 +9493,14 @@ _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 227a <_sk_load_g8_hsw+0x42>
+ .byte 117,50 // jne 2272 <_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 196,226,125,24,13,26,29,0,0 // vbroadcastss 0x1d1a(%rip),%ymm1 # 3f78 <_sk_callback_hsw+0x32d>
+ .byte 196,226,125,24,13,26,29,0,0 // vbroadcastss 0x1d1a(%rip),%ymm1 # 3f70 <_sk_callback_hsw+0x32d>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,15,29,0,0 // vbroadcastss 0x1d0f(%rip),%ymm3 # 3f7c <_sk_callback_hsw+0x331>
+ .byte 196,226,125,24,29,15,29,0,0 // vbroadcastss 0x1d0f(%rip),%ymm3 # 3f74 <_sk_callback_hsw+0x331>
.byte 76,137,193 // mov %r8,%rcx
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
@@ -9513,9 +9514,9 @@ _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 2282 <_sk_load_g8_hsw+0x4a>
+ .byte 117,234 // jne 227a <_sk_load_g8_hsw+0x4a>
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
- .byte 235,173 // jmp 224c <_sk_load_g8_hsw+0x14>
+ .byte 235,173 // jmp 2244 <_sk_load_g8_hsw+0x14>
HIDDEN _sk_gather_g8_hsw
.globl _sk_gather_g8_hsw
@@ -9561,10 +9562,10 @@ _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 196,226,125,24,13,36,28,0,0 // vbroadcastss 0x1c24(%rip),%ymm1 # 3f80 <_sk_callback_hsw+0x335>
+ .byte 196,226,125,24,13,36,28,0,0 // vbroadcastss 0x1c24(%rip),%ymm1 # 3f78 <_sk_callback_hsw+0x335>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,25,28,0,0 // vbroadcastss 0x1c19(%rip),%ymm3 # 3f84 <_sk_callback_hsw+0x339>
+ .byte 196,226,125,24,29,25,28,0,0 // vbroadcastss 0x1c19(%rip),%ymm3 # 3f7c <_sk_callback_hsw+0x339>
.byte 197,252,40,200 // vmovaps %ymm0,%ymm1
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 91 // pop %rbx
@@ -9580,9 +9581,9 @@ _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 238b <_sk_gather_i8_hsw+0xf>
+ .byte 116,5 // je 2383 <_sk_gather_i8_hsw+0xf>
.byte 76,137,192 // mov %r8,%rax
- .byte 235,2 // jmp 238d <_sk_gather_i8_hsw+0x11>
+ .byte 235,2 // jmp 2385 <_sk_gather_i8_hsw+0x11>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,87 // push %r15
.byte 65,86 // push %r14
@@ -9620,14 +9621,14 @@ _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 197,229,219,5,5,29,0,0 // vpand 0x1d05(%rip),%ymm3,%ymm0 # 4140 <_sk_callback_hsw+0x4f5>
+ .byte 197,229,219,5,13,29,0,0 // vpand 0x1d0d(%rip),%ymm3,%ymm0 # 4140 <_sk_callback_hsw+0x4fd>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,64,27,0,0 // vbroadcastss 0x1b40(%rip),%ymm8 # 3f88 <_sk_callback_hsw+0x33d>
+ .byte 196,98,125,24,5,64,27,0,0 // vbroadcastss 0x1b40(%rip),%ymm8 # 3f80 <_sk_callback_hsw+0x33d>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,10,29,0,0 // vpshufb 0x1d0a(%rip),%ymm3,%ymm1 # 4160 <_sk_callback_hsw+0x515>
+ .byte 196,226,101,0,13,18,29,0,0 // vpshufb 0x1d12(%rip),%ymm3,%ymm1 # 4160 <_sk_callback_hsw+0x51d>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,24,29,0,0 // vpshufb 0x1d18(%rip),%ymm3,%ymm2 # 4180 <_sk_callback_hsw+0x535>
+ .byte 196,226,101,0,21,32,29,0,0 // vpshufb 0x1d20(%rip),%ymm3,%ymm2 # 4180 <_sk_callback_hsw+0x53d>
.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
@@ -9648,35 +9649,35 @@ _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 2508 <_sk_load_565_hsw+0x7c>
+ .byte 117,114 // jne 2500 <_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 196,226,125,88,5,226,26,0,0 // vpbroadcastd 0x1ae2(%rip),%ymm0 # 3f8c <_sk_callback_hsw+0x341>
+ .byte 196,226,125,88,5,226,26,0,0 // vpbroadcastd 0x1ae2(%rip),%ymm0 # 3f84 <_sk_callback_hsw+0x341>
.byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,213,26,0,0 // vbroadcastss 0x1ad5(%rip),%ymm1 # 3f90 <_sk_callback_hsw+0x345>
+ .byte 196,226,125,24,13,213,26,0,0 // vbroadcastss 0x1ad5(%rip),%ymm1 # 3f88 <_sk_callback_hsw+0x345>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,204,26,0,0 // vpbroadcastd 0x1acc(%rip),%ymm1 # 3f94 <_sk_callback_hsw+0x349>
+ .byte 196,226,125,88,13,204,26,0,0 // vpbroadcastd 0x1acc(%rip),%ymm1 # 3f8c <_sk_callback_hsw+0x349>
.byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,191,26,0,0 // vbroadcastss 0x1abf(%rip),%ymm3 # 3f98 <_sk_callback_hsw+0x34d>
+ .byte 196,226,125,24,29,191,26,0,0 // vbroadcastss 0x1abf(%rip),%ymm3 # 3f90 <_sk_callback_hsw+0x34d>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,88,29,182,26,0,0 // vpbroadcastd 0x1ab6(%rip),%ymm3 # 3f9c <_sk_callback_hsw+0x351>
+ .byte 196,226,125,88,29,182,26,0,0 // vpbroadcastd 0x1ab6(%rip),%ymm3 # 3f94 <_sk_callback_hsw+0x351>
.byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,169,26,0,0 // vbroadcastss 0x1aa9(%rip),%ymm3 # 3fa0 <_sk_callback_hsw+0x355>
+ .byte 196,226,125,24,29,169,26,0,0 // vbroadcastss 0x1aa9(%rip),%ymm3 # 3f98 <_sk_callback_hsw+0x355>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,158,26,0,0 // vbroadcastss 0x1a9e(%rip),%ymm3 # 3fa4 <_sk_callback_hsw+0x359>
+ .byte 196,226,125,24,29,158,26,0,0 // vbroadcastss 0x1a9e(%rip),%ymm3 # 3f9c <_sk_callback_hsw+0x359>
.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 249c <_sk_load_565_hsw+0x10>
+ .byte 119,128 // ja 2494 <_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 # 2570 <_sk_load_565_hsw+0xe4>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2568 <_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
@@ -9688,7 +9689,7 @@ _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 249c <_sk_load_565_hsw+0x10>
+ .byte 233,44,255,255,255 // jmpq 2494 <_sk_load_565_hsw+0x10>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -9758,23 +9759,23 @@ _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 196,226,125,88,5,97,25,0,0 // vpbroadcastd 0x1961(%rip),%ymm0 # 3fa8 <_sk_callback_hsw+0x35d>
+ .byte 196,226,125,88,5,97,25,0,0 // vpbroadcastd 0x1961(%rip),%ymm0 # 3fa0 <_sk_callback_hsw+0x35d>
.byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,84,25,0,0 // vbroadcastss 0x1954(%rip),%ymm1 # 3fac <_sk_callback_hsw+0x361>
+ .byte 196,226,125,24,13,84,25,0,0 // vbroadcastss 0x1954(%rip),%ymm1 # 3fa4 <_sk_callback_hsw+0x361>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,75,25,0,0 // vpbroadcastd 0x194b(%rip),%ymm1 # 3fb0 <_sk_callback_hsw+0x365>
+ .byte 196,226,125,88,13,75,25,0,0 // vpbroadcastd 0x194b(%rip),%ymm1 # 3fa8 <_sk_callback_hsw+0x365>
.byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,29,62,25,0,0 // vbroadcastss 0x193e(%rip),%ymm3 # 3fb4 <_sk_callback_hsw+0x369>
+ .byte 196,226,125,24,29,62,25,0,0 // vbroadcastss 0x193e(%rip),%ymm3 # 3fac <_sk_callback_hsw+0x369>
.byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
- .byte 196,226,125,88,29,53,25,0,0 // vpbroadcastd 0x1935(%rip),%ymm3 # 3fb8 <_sk_callback_hsw+0x36d>
+ .byte 196,226,125,88,29,53,25,0,0 // vpbroadcastd 0x1935(%rip),%ymm3 # 3fb0 <_sk_callback_hsw+0x36d>
.byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,226,125,24,29,40,25,0,0 // vbroadcastss 0x1928(%rip),%ymm3 # 3fbc <_sk_callback_hsw+0x371>
+ .byte 196,226,125,24,29,40,25,0,0 // vbroadcastss 0x1928(%rip),%ymm3 # 3fb4 <_sk_callback_hsw+0x371>
.byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,29,29,25,0,0 // vbroadcastss 0x191d(%rip),%ymm3 # 3fc0 <_sk_callback_hsw+0x375>
+ .byte 196,226,125,24,29,29,25,0,0 // vbroadcastss 0x191d(%rip),%ymm3 # 3fb8 <_sk_callback_hsw+0x375>
.byte 91 // pop %rbx
.byte 65,92 // pop %r12
.byte 65,94 // pop %r14
@@ -9787,11 +9788,11 @@ FUNCTION(_sk_store_565_hsw)
_sk_store_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,10,25,0,0 // vbroadcastss 0x190a(%rip),%ymm8 # 3fc4 <_sk_callback_hsw+0x379>
+ .byte 196,98,125,24,5,10,25,0,0 // vbroadcastss 0x190a(%rip),%ymm8 # 3fbc <_sk_callback_hsw+0x379>
.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 196,98,125,24,21,245,24,0,0 // vbroadcastss 0x18f5(%rip),%ymm10 # 3fc8 <_sk_callback_hsw+0x37d>
+ .byte 196,98,125,24,21,245,24,0,0 // vbroadcastss 0x18f5(%rip),%ymm10 # 3fc0 <_sk_callback_hsw+0x37d>
.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
@@ -9802,7 +9803,7 @@ _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 2711 <_sk_store_565_hsw+0x65>
+ .byte 117,10 // jne 2709 <_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
@@ -9810,9 +9811,9 @@ _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 270d <_sk_store_565_hsw+0x61>
+ .byte 119,236 // ja 2705 <_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 # 2770 <_sk_store_565_hsw+0xc4>
+ .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2768 <_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
@@ -9823,7 +9824,7 @@ _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 270d <_sk_store_565_hsw+0x61>
+ .byte 235,159 // jmp 2705 <_sk_store_565_hsw+0x61>
.byte 102,144 // xchg %ax,%ax
.byte 245 // cmc
.byte 255 // (bad)
@@ -9856,28 +9857,28 @@ _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 2824 <_sk_load_4444_hsw+0x98>
+ .byte 15,133,138,0,0,0 // jne 281c <_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 196,226,125,88,5,30,24,0,0 // vpbroadcastd 0x181e(%rip),%ymm0 # 3fcc <_sk_callback_hsw+0x381>
+ .byte 196,226,125,88,5,30,24,0,0 // vpbroadcastd 0x181e(%rip),%ymm0 # 3fc4 <_sk_callback_hsw+0x381>
.byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,17,24,0,0 // vbroadcastss 0x1811(%rip),%ymm1 # 3fd0 <_sk_callback_hsw+0x385>
+ .byte 196,226,125,24,13,17,24,0,0 // vbroadcastss 0x1811(%rip),%ymm1 # 3fc8 <_sk_callback_hsw+0x385>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,8,24,0,0 // vpbroadcastd 0x1808(%rip),%ymm1 # 3fd4 <_sk_callback_hsw+0x389>
+ .byte 196,226,125,88,13,8,24,0,0 // vpbroadcastd 0x1808(%rip),%ymm1 # 3fcc <_sk_callback_hsw+0x389>
.byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,251,23,0,0 // vbroadcastss 0x17fb(%rip),%ymm2 # 3fd8 <_sk_callback_hsw+0x38d>
+ .byte 196,226,125,24,21,251,23,0,0 // vbroadcastss 0x17fb(%rip),%ymm2 # 3fd0 <_sk_callback_hsw+0x38d>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,88,21,242,23,0,0 // vpbroadcastd 0x17f2(%rip),%ymm2 # 3fdc <_sk_callback_hsw+0x391>
+ .byte 196,226,125,88,21,242,23,0,0 // vpbroadcastd 0x17f2(%rip),%ymm2 # 3fd4 <_sk_callback_hsw+0x391>
.byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,229,23,0,0 // vbroadcastss 0x17e5(%rip),%ymm8 # 3fe0 <_sk_callback_hsw+0x395>
+ .byte 196,98,125,24,5,229,23,0,0 // vbroadcastss 0x17e5(%rip),%ymm8 # 3fd8 <_sk_callback_hsw+0x395>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,88,5,219,23,0,0 // vpbroadcastd 0x17db(%rip),%ymm8 # 3fe4 <_sk_callback_hsw+0x399>
+ .byte 196,98,125,88,5,219,23,0,0 // vpbroadcastd 0x17db(%rip),%ymm8 # 3fdc <_sk_callback_hsw+0x399>
.byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,205,23,0,0 // vbroadcastss 0x17cd(%rip),%ymm8 # 3fe8 <_sk_callback_hsw+0x39d>
+ .byte 196,98,125,24,5,205,23,0,0 // vbroadcastss 0x17cd(%rip),%ymm8 # 3fe0 <_sk_callback_hsw+0x39d>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -9886,9 +9887,9 @@ _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 27a0 <_sk_load_4444_hsw+0x14>
+ .byte 15,135,100,255,255,255 // ja 2798 <_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 # 2890 <_sk_load_4444_hsw+0x104>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 2888 <_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
@@ -9900,7 +9901,7 @@ _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 27a0 <_sk_load_4444_hsw+0x14>
+ .byte 233,16,255,255,255 // jmpq 2798 <_sk_load_4444_hsw+0x14>
.byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
@@ -9970,25 +9971,25 @@ _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 196,226,125,88,5,133,22,0,0 // vpbroadcastd 0x1685(%rip),%ymm0 # 3fec <_sk_callback_hsw+0x3a1>
+ .byte 196,226,125,88,5,133,22,0,0 // vpbroadcastd 0x1685(%rip),%ymm0 # 3fe4 <_sk_callback_hsw+0x3a1>
.byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,226,125,24,13,120,22,0,0 // vbroadcastss 0x1678(%rip),%ymm1 # 3ff0 <_sk_callback_hsw+0x3a5>
+ .byte 196,226,125,24,13,120,22,0,0 // vbroadcastss 0x1678(%rip),%ymm1 # 3fe8 <_sk_callback_hsw+0x3a5>
.byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
- .byte 196,226,125,88,13,111,22,0,0 // vpbroadcastd 0x166f(%rip),%ymm1 # 3ff4 <_sk_callback_hsw+0x3a9>
+ .byte 196,226,125,88,13,111,22,0,0 // vpbroadcastd 0x166f(%rip),%ymm1 # 3fec <_sk_callback_hsw+0x3a9>
.byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
- .byte 196,226,125,24,21,98,22,0,0 // vbroadcastss 0x1662(%rip),%ymm2 # 3ff8 <_sk_callback_hsw+0x3ad>
+ .byte 196,226,125,24,21,98,22,0,0 // vbroadcastss 0x1662(%rip),%ymm2 # 3ff0 <_sk_callback_hsw+0x3ad>
.byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
- .byte 196,226,125,88,21,89,22,0,0 // vpbroadcastd 0x1659(%rip),%ymm2 # 3ffc <_sk_callback_hsw+0x3b1>
+ .byte 196,226,125,88,21,89,22,0,0 // vpbroadcastd 0x1659(%rip),%ymm2 # 3ff4 <_sk_callback_hsw+0x3b1>
.byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
.byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
- .byte 196,98,125,24,5,76,22,0,0 // vbroadcastss 0x164c(%rip),%ymm8 # 4000 <_sk_callback_hsw+0x3b5>
+ .byte 196,98,125,24,5,76,22,0,0 // vbroadcastss 0x164c(%rip),%ymm8 # 3ff8 <_sk_callback_hsw+0x3b5>
.byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
- .byte 196,98,125,88,5,66,22,0,0 // vpbroadcastd 0x1642(%rip),%ymm8 # 4004 <_sk_callback_hsw+0x3b9>
+ .byte 196,98,125,88,5,66,22,0,0 // vpbroadcastd 0x1642(%rip),%ymm8 # 3ffc <_sk_callback_hsw+0x3b9>
.byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,5,52,22,0,0 // vbroadcastss 0x1634(%rip),%ymm8 # 4008 <_sk_callback_hsw+0x3bd>
+ .byte 196,98,125,24,5,52,22,0,0 // vbroadcastss 0x1634(%rip),%ymm8 # 4000 <_sk_callback_hsw+0x3bd>
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 91 // pop %rbx
@@ -10003,7 +10004,7 @@ FUNCTION(_sk_store_4444_hsw)
_sk_store_4444_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
- .byte 196,98,125,24,5,26,22,0,0 // vbroadcastss 0x161a(%rip),%ymm8 # 400c <_sk_callback_hsw+0x3c1>
+ .byte 196,98,125,24,5,26,22,0,0 // vbroadcastss 0x161a(%rip),%ymm8 # 4004 <_sk_callback_hsw+0x3c1>
.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
@@ -10021,7 +10022,7 @@ _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 2a55 <_sk_store_4444_hsw+0x71>
+ .byte 117,10 // jne 2a4d <_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
@@ -10029,9 +10030,9 @@ _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 2a51 <_sk_store_4444_hsw+0x6d>
+ .byte 119,236 // ja 2a49 <_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 # 2ab4 <_sk_store_4444_hsw+0xd0>
+ .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 2aac <_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
@@ -10042,7 +10043,7 @@ _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 2a51 <_sk_store_4444_hsw+0x6d>
+ .byte 235,159 // jmp 2a49 <_sk_store_4444_hsw+0x6d>
.byte 102,144 // xchg %ax,%ax
.byte 245 // cmc
.byte 255 // (bad)
@@ -10077,16 +10078,16 @@ _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 2b3d <_sk_load_8888_hsw+0x6d>
+ .byte 117,88 // jne 2b35 <_sk_load_8888_hsw+0x6d>
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
- .byte 197,229,219,5,174,22,0,0 // vpand 0x16ae(%rip),%ymm3,%ymm0 # 41a0 <_sk_callback_hsw+0x555>
+ .byte 197,229,219,5,182,22,0,0 // vpand 0x16b6(%rip),%ymm3,%ymm0 # 41a0 <_sk_callback_hsw+0x55d>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,17,21,0,0 // vbroadcastss 0x1511(%rip),%ymm8 # 4010 <_sk_callback_hsw+0x3c5>
+ .byte 196,98,125,24,5,17,21,0,0 // vbroadcastss 0x1511(%rip),%ymm8 # 4008 <_sk_callback_hsw+0x3c5>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,179,22,0,0 // vpshufb 0x16b3(%rip),%ymm3,%ymm1 # 41c0 <_sk_callback_hsw+0x575>
+ .byte 196,226,101,0,13,187,22,0,0 // vpshufb 0x16bb(%rip),%ymm3,%ymm1 # 41c0 <_sk_callback_hsw+0x57d>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,193,22,0,0 // vpshufb 0x16c1(%rip),%ymm3,%ymm2 # 41e0 <_sk_callback_hsw+0x595>
+ .byte 196,226,101,0,21,201,22,0,0 // vpshufb 0x16c9(%rip),%ymm3,%ymm2 # 41e0 <_sk_callback_hsw+0x59d>
.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
@@ -10103,7 +10104,7 @@ _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 2aea <_sk_load_8888_hsw+0x1a>
+ .byte 235,135 // jmp 2ae2 <_sk_load_8888_hsw+0x1a>
HIDDEN _sk_gather_8888_hsw
.globl _sk_gather_8888_hsw
@@ -10118,14 +10119,14 @@ _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 197,229,219,5,111,22,0,0 // vpand 0x166f(%rip),%ymm3,%ymm0 # 4200 <_sk_callback_hsw+0x5b5>
+ .byte 197,229,219,5,119,22,0,0 // vpand 0x1677(%rip),%ymm3,%ymm0 # 4200 <_sk_callback_hsw+0x5bd>
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
- .byte 196,98,125,24,5,118,20,0,0 // vbroadcastss 0x1476(%rip),%ymm8 # 4014 <_sk_callback_hsw+0x3c9>
+ .byte 196,98,125,24,5,118,20,0,0 // vbroadcastss 0x1476(%rip),%ymm8 # 400c <_sk_callback_hsw+0x3c9>
.byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
- .byte 196,226,101,0,13,116,22,0,0 // vpshufb 0x1674(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x5d5>
+ .byte 196,226,101,0,13,124,22,0,0 // vpshufb 0x167c(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x5dd>
.byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
- .byte 196,226,101,0,21,130,22,0,0 // vpshufb 0x1682(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x5f5>
+ .byte 196,226,101,0,21,138,22,0,0 // vpshufb 0x168a(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x5fd>
.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
@@ -10142,7 +10143,7 @@ _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 196,98,125,24,5,38,20,0,0 // vbroadcastss 0x1426(%rip),%ymm8 # 4018 <_sk_callback_hsw+0x3cd>
+ .byte 196,98,125,24,5,38,20,0,0 // vbroadcastss 0x1426(%rip),%ymm8 # 4010 <_sk_callback_hsw+0x3cd>
.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
@@ -10158,7 +10159,7 @@ _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 2c4c <_sk_store_8888_hsw+0x73>
+ .byte 117,12 // jne 2c44 <_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
@@ -10171,7 +10172,7 @@ _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 2c45 <_sk_store_8888_hsw+0x6c>
+ .byte 235,211 // jmp 2c3d <_sk_store_8888_hsw+0x6c>
HIDDEN _sk_load_f16_hsw
.globl _sk_load_f16_hsw
@@ -10180,7 +10181,7 @@ _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 2cdd <_sk_load_f16_hsw+0x6b>
+ .byte 117,97 // jne 2cd5 <_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
@@ -10206,29 +10207,29 @@ _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 2d3c <_sk_load_f16_hsw+0xca>
+ .byte 116,79 // je 2d34 <_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 2d3c <_sk_load_f16_hsw+0xca>
+ .byte 114,67 // jb 2d34 <_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 2d49 <_sk_load_f16_hsw+0xd7>
+ .byte 116,68 // je 2d41 <_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 2d49 <_sk_load_f16_hsw+0xd7>
+ .byte 114,56 // jb 2d41 <_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 2c93 <_sk_load_f16_hsw+0x21>
+ .byte 15,132,114,255,255,255 // je 2c8b <_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 2c93 <_sk_load_f16_hsw+0x21>
+ .byte 15,130,98,255,255,255 // jb 2c8b <_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 2c93 <_sk_load_f16_hsw+0x21>
+ .byte 233,87,255,255,255 // jmpq 2c8b <_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 2c93 <_sk_load_f16_hsw+0x21>
+ .byte 233,74,255,255,255 // jmpq 2c8b <_sk_load_f16_hsw+0x21>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,65,255,255,255 // jmpq 2c93 <_sk_load_f16_hsw+0x21>
+ .byte 233,65,255,255,255 // jmpq 2c8b <_sk_load_f16_hsw+0x21>
HIDDEN _sk_gather_f16_hsw
.globl _sk_gather_f16_hsw
@@ -10286,7 +10287,7 @@ _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 2e41 <_sk_store_f16_hsw+0x65>
+ .byte 117,27 // jne 2e39 <_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)
@@ -10295,22 +10296,22 @@ _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 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 116,241 // je 2e35 <_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 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 114,229 // jb 2e35 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
- .byte 116,221 // je 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 116,221 // je 2e35 <_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 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 114,209 // jb 2e35 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
- .byte 116,201 // je 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 116,201 // je 2e35 <_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 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 114,189 // jb 2e35 <_sk_store_f16_hsw+0x61>
.byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
- .byte 235,181 // jmp 2e3d <_sk_store_f16_hsw+0x61>
+ .byte 235,181 // jmp 2e35 <_sk_store_f16_hsw+0x61>
HIDDEN _sk_load_u16_be_hsw
.globl _sk_load_u16_be_hsw
@@ -10320,7 +10321,7 @@ _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 2f6a <_sk_load_u16_be_hsw+0xe2>
+ .byte 15,133,204,0,0,0 // jne 2f62 <_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
@@ -10339,7 +10340,7 @@ _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 196,98,125,24,21,29,17,0,0 // vbroadcastss 0x111d(%rip),%ymm10 # 401c <_sk_callback_hsw+0x3d1>
+ .byte 196,98,125,24,21,29,17,0,0 // vbroadcastss 0x111d(%rip),%ymm10 # 4014 <_sk_callback_hsw+0x3d1>
.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
@@ -10367,29 +10368,29 @@ _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 2fd0 <_sk_load_u16_be_hsw+0x148>
+ .byte 116,85 // je 2fc8 <_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 2fd0 <_sk_load_u16_be_hsw+0x148>
+ .byte 114,72 // jb 2fc8 <_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 2fdd <_sk_load_u16_be_hsw+0x155>
+ .byte 116,72 // je 2fd5 <_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 2fdd <_sk_load_u16_be_hsw+0x155>
+ .byte 114,59 // jb 2fd5 <_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 2eb9 <_sk_load_u16_be_hsw+0x31>
+ .byte 15,132,6,255,255,255 // je 2eb1 <_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 2eb9 <_sk_load_u16_be_hsw+0x31>
+ .byte 15,130,245,254,255,255 // jb 2eb1 <_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 2eb9 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,233,254,255,255 // jmpq 2eb1 <_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 2eb9 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,220,254,255,255 // jmpq 2eb1 <_sk_load_u16_be_hsw+0x31>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,211,254,255,255 // jmpq 2eb9 <_sk_load_u16_be_hsw+0x31>
+ .byte 233,211,254,255,255 // jmpq 2eb1 <_sk_load_u16_be_hsw+0x31>
HIDDEN _sk_load_rgb_u16_be_hsw
.globl _sk_load_rgb_u16_be_hsw
@@ -10399,7 +10400,7 @@ _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 30c4 <_sk_load_rgb_u16_be_hsw+0xde>
+ .byte 15,133,204,0,0,0 // jne 30bc <_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
@@ -10423,7 +10424,7 @@ _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 196,98,125,24,21,174,15,0,0 // vbroadcastss 0xfae(%rip),%ymm10 # 4020 <_sk_callback_hsw+0x3d5>
+ .byte 196,98,125,24,21,174,15,0,0 // vbroadcastss 0xfae(%rip),%ymm10 # 4018 <_sk_callback_hsw+0x3d5>
.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
@@ -10440,41 +10441,41 @@ _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 196,226,125,24,29,98,15,0,0 // vbroadcastss 0xf62(%rip),%ymm3 # 4024 <_sk_callback_hsw+0x3d9>
+ .byte 196,226,125,24,29,98,15,0,0 // vbroadcastss 0xf62(%rip),%ymm3 # 401c <_sk_callback_hsw+0x3d9>
.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 30dd <_sk_load_rgb_u16_be_hsw+0xf7>
- .byte 233,79,255,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,5 // jne 30d5 <_sk_load_rgb_u16_be_hsw+0xf7>
+ .byte 233,79,255,255,255 // jmpq 3024 <_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 310c <_sk_load_rgb_u16_be_hsw+0x126>
+ .byte 114,26 // jb 3104 <_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 3111 <_sk_load_rgb_u16_be_hsw+0x12b>
- .byte 233,32,255,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,27,255,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 3109 <_sk_load_rgb_u16_be_hsw+0x12b>
+ .byte 233,32,255,255,255 // jmpq 3024 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,27,255,255,255 // jmpq 3024 <_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 3140 <_sk_load_rgb_u16_be_hsw+0x15a>
+ .byte 114,26 // jb 3138 <_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 3145 <_sk_load_rgb_u16_be_hsw+0x15f>
- .byte 233,236,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,231,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 117,10 // jne 313d <_sk_load_rgb_u16_be_hsw+0x15f>
+ .byte 233,236,254,255,255 // jmpq 3024 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,231,254,255,255 // jmpq 3024 <_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 316e <_sk_load_rgb_u16_be_hsw+0x188>
+ .byte 114,20 // jb 3166 <_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 302c <_sk_load_rgb_u16_be_hsw+0x46>
- .byte 233,185,254,255,255 // jmpq 302c <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,190,254,255,255 // jmpq 3024 <_sk_load_rgb_u16_be_hsw+0x46>
+ .byte 233,185,254,255,255 // jmpq 3024 <_sk_load_rgb_u16_be_hsw+0x46>
HIDDEN _sk_store_u16_be_hsw
.globl _sk_store_u16_be_hsw
@@ -10483,7 +10484,7 @@ _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 196,98,125,24,5,159,14,0,0 // vbroadcastss 0xe9f(%rip),%ymm8 # 4028 <_sk_callback_hsw+0x3dd>
+ .byte 196,98,125,24,5,159,14,0,0 // vbroadcastss 0xe9f(%rip),%ymm8 # 4020 <_sk_callback_hsw+0x3dd>
.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
@@ -10521,7 +10522,7 @@ _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 326d <_sk_store_u16_be_hsw+0xfa>
+ .byte 117,31 // jne 3265 <_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)
@@ -10530,22 +10531,22 @@ _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 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,240 // je 3261 <_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 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,227 // jb 3261 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,84,64,16 // vmovq %xmm10,0x10(%r8,%rax,2)
- .byte 116,218 // je 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,218 // je 3261 <_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 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,205 // jb 3261 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,76,64,32 // vmovq %xmm9,0x20(%r8,%rax,2)
- .byte 116,196 // je 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 116,196 // je 3261 <_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 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 114,183 // jb 3261 <_sk_store_u16_be_hsw+0xf6>
.byte 196,65,121,214,68,64,48 // vmovq %xmm8,0x30(%r8,%rax,2)
- .byte 235,174 // jmp 3269 <_sk_store_u16_be_hsw+0xf6>
+ .byte 235,174 // jmp 3261 <_sk_store_u16_be_hsw+0xf6>
HIDDEN _sk_load_f32_hsw
.globl _sk_load_f32_hsw
@@ -10553,10 +10554,10 @@ 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 3331 <_sk_load_f32_hsw+0x76>
+ .byte 119,110 // ja 3329 <_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 # 335c <_sk_load_f32_hsw+0xa1>
+ .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 3354 <_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
@@ -10617,7 +10618,7 @@ _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 33e9 <_sk_store_f32_hsw+0x6d>
+ .byte 117,55 // jne 33e1 <_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
@@ -10630,22 +10631,22 @@ _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 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 116,240 // je 33dd <_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 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 114,227 // jb 33dd <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
- .byte 116,218 // je 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 116,218 // je 33dd <_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 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 114,205 // jb 33dd <_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 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 116,195 // je 33dd <_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 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 114,181 // jb 33dd <_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 33e5 <_sk_store_f32_hsw+0x69>
+ .byte 235,171 // jmp 33dd <_sk_store_f32_hsw+0x69>
HIDDEN _sk_clamp_x_hsw
.globl _sk_clamp_x_hsw
@@ -10755,11 +10756,11 @@ HIDDEN _sk_luminance_to_alpha_hsw
.globl _sk_luminance_to_alpha_hsw
FUNCTION(_sk_luminance_to_alpha_hsw)
_sk_luminance_to_alpha_hsw:
- .byte 196,226,125,24,29,185,10,0,0 // vbroadcastss 0xab9(%rip),%ymm3 # 402c <_sk_callback_hsw+0x3e1>
- .byte 196,98,125,24,5,180,10,0,0 // vbroadcastss 0xab4(%rip),%ymm8 # 4030 <_sk_callback_hsw+0x3e5>
+ .byte 196,226,125,24,29,185,10,0,0 // vbroadcastss 0xab9(%rip),%ymm3 # 4024 <_sk_callback_hsw+0x3e1>
+ .byte 196,98,125,24,5,180,10,0,0 // vbroadcastss 0xab4(%rip),%ymm8 # 4028 <_sk_callback_hsw+0x3e5>
.byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
.byte 196,226,125,184,203 // vfmadd231ps %ymm3,%ymm0,%ymm1
- .byte 196,226,125,24,29,165,10,0,0 // vbroadcastss 0xaa5(%rip),%ymm3 # 4034 <_sk_callback_hsw+0x3e9>
+ .byte 196,226,125,24,29,165,10,0,0 // vbroadcastss 0xaa5(%rip),%ymm3 # 402c <_sk_callback_hsw+0x3e9>
.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
@@ -10904,7 +10905,7 @@ _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 3867 <_sk_linear_gradient_hsw+0xb5>
+ .byte 15,132,143,0,0,0 // je 385f <_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
@@ -10931,8 +10932,8 @@ _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 37f1 <_sk_linear_gradient_hsw+0x3f>
- .byte 235,17 // jmp 3878 <_sk_linear_gradient_hsw+0xc6>
+ .byte 117,140 // jne 37e9 <_sk_linear_gradient_hsw+0x3f>
+ .byte 235,17 // jmp 3870 <_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
@@ -10971,7 +10972,7 @@ HIDDEN _sk_save_xy_hsw
FUNCTION(_sk_save_xy_hsw)
_sk_save_xy_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,76,7,0,0 // vbroadcastss 0x74c(%rip),%ymm8 # 4038 <_sk_callback_hsw+0x3ed>
+ .byte 196,98,125,24,5,76,7,0,0 // vbroadcastss 0x74c(%rip),%ymm8 # 4030 <_sk_callback_hsw+0x3ed>
.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
@@ -11005,9 +11006,9 @@ HIDDEN _sk_bilinear_nx_hsw
FUNCTION(_sk_bilinear_nx_hsw)
_sk_bilinear_nx_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,224,6,0,0 // vbroadcastss 0x6e0(%rip),%ymm0 # 403c <_sk_callback_hsw+0x3f1>
+ .byte 196,226,125,24,5,224,6,0,0 // vbroadcastss 0x6e0(%rip),%ymm0 # 4034 <_sk_callback_hsw+0x3f1>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,215,6,0,0 // vbroadcastss 0x6d7(%rip),%ymm8 # 4040 <_sk_callback_hsw+0x3f5>
+ .byte 196,98,125,24,5,215,6,0,0 // vbroadcastss 0x6d7(%rip),%ymm8 # 4038 <_sk_callback_hsw+0x3f5>
.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
@@ -11018,7 +11019,7 @@ HIDDEN _sk_bilinear_px_hsw
FUNCTION(_sk_bilinear_px_hsw)
_sk_bilinear_px_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,191,6,0,0 // vbroadcastss 0x6bf(%rip),%ymm0 # 4044 <_sk_callback_hsw+0x3f9>
+ .byte 196,226,125,24,5,191,6,0,0 // vbroadcastss 0x6bf(%rip),%ymm0 # 403c <_sk_callback_hsw+0x3f9>
.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)
@@ -11030,9 +11031,9 @@ HIDDEN _sk_bilinear_ny_hsw
FUNCTION(_sk_bilinear_ny_hsw)
_sk_bilinear_ny_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,163,6,0,0 // vbroadcastss 0x6a3(%rip),%ymm1 # 4048 <_sk_callback_hsw+0x3fd>
+ .byte 196,226,125,24,13,163,6,0,0 // vbroadcastss 0x6a3(%rip),%ymm1 # 4040 <_sk_callback_hsw+0x3fd>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,153,6,0,0 // vbroadcastss 0x699(%rip),%ymm8 # 404c <_sk_callback_hsw+0x401>
+ .byte 196,98,125,24,5,153,6,0,0 // vbroadcastss 0x699(%rip),%ymm8 # 4044 <_sk_callback_hsw+0x401>
.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
@@ -11043,7 +11044,7 @@ HIDDEN _sk_bilinear_py_hsw
FUNCTION(_sk_bilinear_py_hsw)
_sk_bilinear_py_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,129,6,0,0 // vbroadcastss 0x681(%rip),%ymm1 # 4050 <_sk_callback_hsw+0x405>
+ .byte 196,226,125,24,13,129,6,0,0 // vbroadcastss 0x681(%rip),%ymm1 # 4048 <_sk_callback_hsw+0x405>
.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)
@@ -11055,13 +11056,13 @@ HIDDEN _sk_bicubic_n3x_hsw
FUNCTION(_sk_bicubic_n3x_hsw)
_sk_bicubic_n3x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,100,6,0,0 // vbroadcastss 0x664(%rip),%ymm0 # 4054 <_sk_callback_hsw+0x409>
+ .byte 196,226,125,24,5,100,6,0,0 // vbroadcastss 0x664(%rip),%ymm0 # 404c <_sk_callback_hsw+0x409>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,91,6,0,0 // vbroadcastss 0x65b(%rip),%ymm8 # 4058 <_sk_callback_hsw+0x40d>
+ .byte 196,98,125,24,5,91,6,0,0 // vbroadcastss 0x65b(%rip),%ymm8 # 4050 <_sk_callback_hsw+0x40d>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,76,6,0,0 // vbroadcastss 0x64c(%rip),%ymm10 # 405c <_sk_callback_hsw+0x411>
- .byte 196,98,125,24,29,71,6,0,0 // vbroadcastss 0x647(%rip),%ymm11 # 4060 <_sk_callback_hsw+0x415>
+ .byte 196,98,125,24,21,76,6,0,0 // vbroadcastss 0x64c(%rip),%ymm10 # 4054 <_sk_callback_hsw+0x411>
+ .byte 196,98,125,24,29,71,6,0,0 // vbroadcastss 0x647(%rip),%ymm11 # 4058 <_sk_callback_hsw+0x415>
.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)
@@ -11073,16 +11074,16 @@ HIDDEN _sk_bicubic_n1x_hsw
FUNCTION(_sk_bicubic_n1x_hsw)
_sk_bicubic_n1x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,42,6,0,0 // vbroadcastss 0x62a(%rip),%ymm0 # 4064 <_sk_callback_hsw+0x419>
+ .byte 196,226,125,24,5,42,6,0,0 // vbroadcastss 0x62a(%rip),%ymm0 # 405c <_sk_callback_hsw+0x419>
.byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
- .byte 196,98,125,24,5,33,6,0,0 // vbroadcastss 0x621(%rip),%ymm8 # 4068 <_sk_callback_hsw+0x41d>
+ .byte 196,98,125,24,5,33,6,0,0 // vbroadcastss 0x621(%rip),%ymm8 # 4060 <_sk_callback_hsw+0x41d>
.byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,23,6,0,0 // vbroadcastss 0x617(%rip),%ymm9 # 406c <_sk_callback_hsw+0x421>
- .byte 196,98,125,24,21,18,6,0,0 // vbroadcastss 0x612(%rip),%ymm10 # 4070 <_sk_callback_hsw+0x425>
+ .byte 196,98,125,24,13,23,6,0,0 // vbroadcastss 0x617(%rip),%ymm9 # 4064 <_sk_callback_hsw+0x421>
+ .byte 196,98,125,24,21,18,6,0,0 // vbroadcastss 0x612(%rip),%ymm10 # 4068 <_sk_callback_hsw+0x425>
.byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,13,8,6,0,0 // vbroadcastss 0x608(%rip),%ymm9 # 4074 <_sk_callback_hsw+0x429>
+ .byte 196,98,125,24,13,8,6,0,0 // vbroadcastss 0x608(%rip),%ymm9 # 406c <_sk_callback_hsw+0x429>
.byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
- .byte 196,98,125,24,21,254,5,0,0 // vbroadcastss 0x5fe(%rip),%ymm10 # 4078 <_sk_callback_hsw+0x42d>
+ .byte 196,98,125,24,21,254,5,0,0 // vbroadcastss 0x5fe(%rip),%ymm10 # 4070 <_sk_callback_hsw+0x42d>
.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
@@ -11093,14 +11094,14 @@ HIDDEN _sk_bicubic_p1x_hsw
FUNCTION(_sk_bicubic_p1x_hsw)
_sk_bicubic_p1x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,230,5,0,0 // vbroadcastss 0x5e6(%rip),%ymm8 # 407c <_sk_callback_hsw+0x431>
+ .byte 196,98,125,24,5,230,5,0,0 // vbroadcastss 0x5e6(%rip),%ymm8 # 4074 <_sk_callback_hsw+0x431>
.byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
.byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
- .byte 196,98,125,24,21,216,5,0,0 // vbroadcastss 0x5d8(%rip),%ymm10 # 4080 <_sk_callback_hsw+0x435>
- .byte 196,98,125,24,29,211,5,0,0 // vbroadcastss 0x5d3(%rip),%ymm11 # 4084 <_sk_callback_hsw+0x439>
+ .byte 196,98,125,24,21,216,5,0,0 // vbroadcastss 0x5d8(%rip),%ymm10 # 4078 <_sk_callback_hsw+0x435>
+ .byte 196,98,125,24,29,211,5,0,0 // vbroadcastss 0x5d3(%rip),%ymm11 # 407c <_sk_callback_hsw+0x439>
.byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11
.byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11
- .byte 196,98,125,24,5,196,5,0,0 // vbroadcastss 0x5c4(%rip),%ymm8 # 4088 <_sk_callback_hsw+0x43d>
+ .byte 196,98,125,24,5,196,5,0,0 // vbroadcastss 0x5c4(%rip),%ymm8 # 4080 <_sk_callback_hsw+0x43d>
.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
@@ -11111,12 +11112,12 @@ HIDDEN _sk_bicubic_p3x_hsw
FUNCTION(_sk_bicubic_p3x_hsw)
_sk_bicubic_p3x_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,5,172,5,0,0 // vbroadcastss 0x5ac(%rip),%ymm0 # 408c <_sk_callback_hsw+0x441>
+ .byte 196,226,125,24,5,172,5,0,0 // vbroadcastss 0x5ac(%rip),%ymm0 # 4084 <_sk_callback_hsw+0x441>
.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 196,98,125,24,21,153,5,0,0 // vbroadcastss 0x599(%rip),%ymm10 # 4090 <_sk_callback_hsw+0x445>
- .byte 196,98,125,24,29,148,5,0,0 // vbroadcastss 0x594(%rip),%ymm11 # 4094 <_sk_callback_hsw+0x449>
+ .byte 196,98,125,24,21,153,5,0,0 // vbroadcastss 0x599(%rip),%ymm10 # 4088 <_sk_callback_hsw+0x445>
+ .byte 196,98,125,24,29,148,5,0,0 // vbroadcastss 0x594(%rip),%ymm11 # 408c <_sk_callback_hsw+0x449>
.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)
@@ -11128,13 +11129,13 @@ HIDDEN _sk_bicubic_n3y_hsw
FUNCTION(_sk_bicubic_n3y_hsw)
_sk_bicubic_n3y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,119,5,0,0 // vbroadcastss 0x577(%rip),%ymm1 # 4098 <_sk_callback_hsw+0x44d>
+ .byte 196,226,125,24,13,119,5,0,0 // vbroadcastss 0x577(%rip),%ymm1 # 4090 <_sk_callback_hsw+0x44d>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,109,5,0,0 // vbroadcastss 0x56d(%rip),%ymm8 # 409c <_sk_callback_hsw+0x451>
+ .byte 196,98,125,24,5,109,5,0,0 // vbroadcastss 0x56d(%rip),%ymm8 # 4094 <_sk_callback_hsw+0x451>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
.byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
- .byte 196,98,125,24,21,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm10 # 40a0 <_sk_callback_hsw+0x455>
- .byte 196,98,125,24,29,89,5,0,0 // vbroadcastss 0x559(%rip),%ymm11 # 40a4 <_sk_callback_hsw+0x459>
+ .byte 196,98,125,24,21,94,5,0,0 // vbroadcastss 0x55e(%rip),%ymm10 # 4098 <_sk_callback_hsw+0x455>
+ .byte 196,98,125,24,29,89,5,0,0 // vbroadcastss 0x559(%rip),%ymm11 # 409c <_sk_callback_hsw+0x459>
.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)
@@ -11146,16 +11147,16 @@ HIDDEN _sk_bicubic_n1y_hsw
FUNCTION(_sk_bicubic_n1y_hsw)
_sk_bicubic_n1y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,60,5,0,0 // vbroadcastss 0x53c(%rip),%ymm1 # 40a8 <_sk_callback_hsw+0x45d>
+ .byte 196,226,125,24,13,60,5,0,0 // vbroadcastss 0x53c(%rip),%ymm1 # 40a0 <_sk_callback_hsw+0x45d>
.byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
- .byte 196,98,125,24,5,50,5,0,0 // vbroadcastss 0x532(%rip),%ymm8 # 40ac <_sk_callback_hsw+0x461>
+ .byte 196,98,125,24,5,50,5,0,0 // vbroadcastss 0x532(%rip),%ymm8 # 40a4 <_sk_callback_hsw+0x461>
.byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
- .byte 196,98,125,24,13,40,5,0,0 // vbroadcastss 0x528(%rip),%ymm9 # 40b0 <_sk_callback_hsw+0x465>
- .byte 196,98,125,24,21,35,5,0,0 // vbroadcastss 0x523(%rip),%ymm10 # 40b4 <_sk_callback_hsw+0x469>
+ .byte 196,98,125,24,13,40,5,0,0 // vbroadcastss 0x528(%rip),%ymm9 # 40a8 <_sk_callback_hsw+0x465>
+ .byte 196,98,125,24,21,35,5,0,0 // vbroadcastss 0x523(%rip),%ymm10 # 40ac <_sk_callback_hsw+0x469>
.byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10
- .byte 196,98,125,24,13,25,5,0,0 // vbroadcastss 0x519(%rip),%ymm9 # 40b8 <_sk_callback_hsw+0x46d>
+ .byte 196,98,125,24,13,25,5,0,0 // vbroadcastss 0x519(%rip),%ymm9 # 40b0 <_sk_callback_hsw+0x46d>
.byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
- .byte 196,98,125,24,21,15,5,0,0 // vbroadcastss 0x50f(%rip),%ymm10 # 40bc <_sk_callback_hsw+0x471>
+ .byte 196,98,125,24,21,15,5,0,0 // vbroadcastss 0x50f(%rip),%ymm10 # 40b4 <_sk_callback_hsw+0x471>
.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
@@ -11166,14 +11167,14 @@ HIDDEN _sk_bicubic_p1y_hsw
FUNCTION(_sk_bicubic_p1y_hsw)
_sk_bicubic_p1y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,98,125,24,5,247,4,0,0 // vbroadcastss 0x4f7(%rip),%ymm8 # 40c0 <_sk_callback_hsw+0x475>
+ .byte 196,98,125,24,5,247,4,0,0 // vbroadcastss 0x4f7(%rip),%ymm8 # 40b8 <_sk_callback_hsw+0x475>
.byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
.byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
- .byte 196,98,125,24,21,232,4,0,0 // vbroadcastss 0x4e8(%rip),%ymm10 # 40c4 <_sk_callback_hsw+0x479>
- .byte 196,98,125,24,29,227,4,0,0 // vbroadcastss 0x4e3(%rip),%ymm11 # 40c8 <_sk_callback_hsw+0x47d>
+ .byte 196,98,125,24,21,232,4,0,0 // vbroadcastss 0x4e8(%rip),%ymm10 # 40bc <_sk_callback_hsw+0x479>
+ .byte 196,98,125,24,29,227,4,0,0 // vbroadcastss 0x4e3(%rip),%ymm11 # 40c0 <_sk_callback_hsw+0x47d>
.byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11
.byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11
- .byte 196,98,125,24,5,212,4,0,0 // vbroadcastss 0x4d4(%rip),%ymm8 # 40cc <_sk_callback_hsw+0x481>
+ .byte 196,98,125,24,5,212,4,0,0 // vbroadcastss 0x4d4(%rip),%ymm8 # 40c4 <_sk_callback_hsw+0x481>
.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
@@ -11184,12 +11185,12 @@ HIDDEN _sk_bicubic_p3y_hsw
FUNCTION(_sk_bicubic_p3y_hsw)
_sk_bicubic_p3y_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 196,226,125,24,13,188,4,0,0 // vbroadcastss 0x4bc(%rip),%ymm1 # 40d0 <_sk_callback_hsw+0x485>
+ .byte 196,226,125,24,13,188,4,0,0 // vbroadcastss 0x4bc(%rip),%ymm1 # 40c8 <_sk_callback_hsw+0x485>
.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 196,98,125,24,21,168,4,0,0 // vbroadcastss 0x4a8(%rip),%ymm10 # 40d4 <_sk_callback_hsw+0x489>
- .byte 196,98,125,24,29,163,4,0,0 // vbroadcastss 0x4a3(%rip),%ymm11 # 40d8 <_sk_callback_hsw+0x48d>
+ .byte 196,98,125,24,21,168,4,0,0 // vbroadcastss 0x4a8(%rip),%ymm10 # 40cc <_sk_callback_hsw+0x489>
+ .byte 196,98,125,24,29,163,4,0,0 // vbroadcastss 0x4a3(%rip),%ymm11 # 40d0 <_sk_callback_hsw+0x48d>
.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)
@@ -11345,7 +11346,7 @@ 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 3e49 <.literal4+0xd9>
+ .byte 224,7 // loopne 3e41 <.literal4+0xd9>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -11359,10 +11360,10 @@ 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 3e74 <.literal4+0x104>
+ .byte 127,0 // jg 3e6c <.literal4+0x104>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 3eed <.literal4+0x17d>
+ .byte 119,115 // ja 3ee5 <.literal4+0x17d>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11376,10 +11377,10 @@ 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 3ea8 <.literal4+0x138>
+ .byte 127,0 // jg 3ea0 <.literal4+0x138>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 3f21 <.literal4+0x1b1>
+ .byte 119,115 // ja 3f19 <.literal4+0x1b1>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11393,10 +11394,10 @@ 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 3edc <.literal4+0x16c>
+ .byte 127,0 // jg 3ed4 <.literal4+0x16c>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 3f55 <.literal4+0x1e5>
+ .byte 119,115 // ja 3f4d <.literal4+0x1e5>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11410,10 +11411,10 @@ 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 3f10 <.literal4+0x1a0>
+ .byte 127,0 // jg 3f08 <.literal4+0x1a0>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 3f89 <.literal4+0x219>
+ .byte 119,115 // ja 3f81 <.literal4+0x219>
.byte 248 // clc
.byte 194,117,191 // retq $0xbf75
.byte 191,63,249,68,180 // mov $0xb444f93f,%edi
@@ -11426,7 +11427,7 @@ 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 3f87 <.literal4+0x217>
+ .byte 127,67 // jg 3f7f <.literal4+0x217>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -11438,10 +11439,10 @@ BALIGN4
.byte 190,80,128,3,62 // mov $0x3e038050,%esi
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 3fa7 <.literal4+0x237>
+ .byte 118,63 // jbe 3f9f <.literal4+0x237>
.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 3fbb <.literal4+0x24b>
+ .byte 127,67 // jg 3fb3 <.literal4+0x24b>
.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)
@@ -11450,7 +11451,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 3f9d <.literal4+0x22d>
+ .byte 224,7 // loopne 3f95 <.literal4+0x22d>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -11462,7 +11463,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 3fb9 <.literal4+0x249>
+ .byte 224,7 // loopne 3fb1 <.literal4+0x249>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -11473,7 +11474,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 400e <.literal4+0x29e>
+ .byte 124,66 // jl 4006 <.literal4+0x29e>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax)
@@ -11491,9 +11492,9 @@ 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 4051 <.literal4+0x2e1>
+ .byte 112,65 // jo 4049 <.literal4+0x2e1>
.byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax)
- .byte 127,67 // jg 405f <.literal4+0x2ef>
+ .byte 127,67 // jg 4057 <.literal4+0x2ef>
.byte 128,0,128 // addb $0x80,(%rax)
.byte 55 // (bad)
.byte 128,0,128 // addb $0x80,(%rax)
@@ -11501,7 +11502,7 @@ BALIGN4
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 255 // (bad)
- .byte 127,71 // jg 4073 <.literal4+0x303>
+ .byte 127,71 // jg 406b <.literal4+0x303>
.byte 208 // (bad)
.byte 179,89 // mov $0x59,%bl
.byte 62,89 // ds pop %rcx
@@ -11587,16 +11588,16 @@ 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) # a004108 <_sk_callback_hsw+0xa0004bd>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004108 <_sk_callback_hsw+0xa0004c5>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004110 <_sk_callback_hsw+0x120004c5>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004110 <_sk_callback_hsw+0x120004cd>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004118 <_sk_callback_hsw+0x1a0004cd>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004118 <_sk_callback_hsw+0x1a0004d5>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004120 <_sk_callback_hsw+0x30004d5>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004120 <_sk_callback_hsw+0x30004dd>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -11639,16 +11640,16 @@ 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) # a004168 <_sk_callback_hsw+0xa00051d>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004168 <_sk_callback_hsw+0xa000525>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004170 <_sk_callback_hsw+0x12000525>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004170 <_sk_callback_hsw+0x1200052d>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004178 <_sk_callback_hsw+0x1a00052d>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004178 <_sk_callback_hsw+0x1a000535>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004180 <_sk_callback_hsw+0x3000535>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004180 <_sk_callback_hsw+0x300053d>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -11691,16 +11692,16 @@ 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) # a0041c8 <_sk_callback_hsw+0xa00057d>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0041c8 <_sk_callback_hsw+0xa000585>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120041d0 <_sk_callback_hsw+0x12000585>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120041d0 <_sk_callback_hsw+0x1200058d>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0041d8 <_sk_callback_hsw+0x1a00058d>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0041d8 <_sk_callback_hsw+0x1a000595>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30041e0 <_sk_callback_hsw+0x3000595>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30041e0 <_sk_callback_hsw+0x300059d>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -11743,16 +11744,16 @@ 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) # a004228 <_sk_callback_hsw+0xa0005dd>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a004228 <_sk_callback_hsw+0xa0005e5>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004230 <_sk_callback_hsw+0x120005e5>
+ .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004230 <_sk_callback_hsw+0x120005ed>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004238 <_sk_callback_hsw+0x1a0005ed>
+ .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004238 <_sk_callback_hsw+0x1a0005f5>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004240 <_sk_callback_hsw+0x30005f5>
+ .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3004240 <_sk_callback_hsw+0x30005fd>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -12842,23 +12843,23 @@ _sk_hsl_to_rgb_avx:
.byte 197,252,17,44,36 // vmovups %ymm5,(%rsp)
.byte 197,252,17,100,36,224 // vmovups %ymm4,-0x20(%rsp)
.byte 197,252,17,92,36,192 // vmovups %ymm3,-0x40(%rsp)
- .byte 197,252,40,242 // vmovaps %ymm2,%ymm6
- .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp)
+ .byte 197,252,40,234 // vmovaps %ymm2,%ymm5
.byte 197,252,40,208 // vmovaps %ymm0,%ymm2
.byte 184,0,0,0,63 // mov $0x3f000000,%eax
.byte 197,249,110,192 // vmovd %eax,%xmm0
.byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
.byte 196,99,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm8
- .byte 196,193,76,194,192,1 // vcmpltps %ymm8,%ymm6,%ymm0
- .byte 196,98,125,24,21,68,70,0,0 // vbroadcastss 0x4644(%rip),%ymm10 # 54dc <_sk_callback_avx+0x1ca>
+ .byte 196,193,84,194,192,1 // vcmpltps %ymm8,%ymm5,%ymm0
+ .byte 196,98,125,24,21,74,70,0,0 // vbroadcastss 0x464a(%rip),%ymm10 # 54dc <_sk_callback_avx+0x1ca>
+ .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp)
.byte 196,193,116,88,218 // vaddps %ymm10,%ymm1,%ymm3
- .byte 197,228,89,222 // vmulps %ymm6,%ymm3,%ymm3
- .byte 197,244,88,230 // vaddps %ymm6,%ymm1,%ymm4
- .byte 197,244,89,238 // vmulps %ymm6,%ymm1,%ymm5
- .byte 197,220,92,229 // vsubps %ymm5,%ymm4,%ymm4
+ .byte 197,228,89,221 // vmulps %ymm5,%ymm3,%ymm3
+ .byte 197,244,88,229 // vaddps %ymm5,%ymm1,%ymm4
+ .byte 197,244,89,245 // vmulps %ymm5,%ymm1,%ymm6
+ .byte 197,220,92,230 // vsubps %ymm6,%ymm4,%ymm4
.byte 196,99,93,74,203,0 // vblendvps %ymm0,%ymm3,%ymm4,%ymm9
- .byte 196,226,125,24,5,36,70,0,0 // vbroadcastss 0x4624(%rip),%ymm0 # 54e0 <_sk_callback_avx+0x1ce>
- .byte 197,236,88,200 // vaddps %ymm0,%ymm2,%ymm1
+ .byte 196,226,125,24,13,36,70,0,0 // vbroadcastss 0x4624(%rip),%ymm1 # 54e0 <_sk_callback_avx+0x1ce>
+ .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1
.byte 65,184,0,0,0,0 // mov $0x0,%r8d
.byte 184,0,0,128,63 // mov $0x3f800000,%eax
.byte 197,249,110,216 // vmovd %eax,%xmm3
@@ -12872,76 +12873,76 @@ _sk_hsl_to_rgb_avx:
.byte 196,227,121,4,228,0 // vpermilps $0x0,%xmm4,%xmm4
.byte 196,99,93,24,252,1 // vinsertf128 $0x1,%xmm4,%ymm4,%ymm15
.byte 196,193,116,194,231,1 // vcmpltps %ymm15,%ymm1,%ymm4
- .byte 196,193,116,88,234 // vaddps %ymm10,%ymm1,%ymm5
- .byte 196,227,101,74,197,64 // vblendvps %ymm4,%ymm5,%ymm3,%ymm0
- .byte 197,204,88,222 // vaddps %ymm6,%ymm6,%ymm3
- .byte 196,65,100,92,217 // vsubps %ymm9,%ymm3,%ymm11
- .byte 196,193,52,92,219 // vsubps %ymm11,%ymm9,%ymm3
- .byte 196,226,125,24,37,187,69,0,0 // vbroadcastss 0x45bb(%rip),%ymm4 # 54e8 <_sk_callback_avx+0x1d6>
- .byte 197,100,89,236 // vmulps %ymm4,%ymm3,%ymm13
+ .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
+ .byte 196,227,101,74,241,64 // vblendvps %ymm4,%ymm1,%ymm3,%ymm6
+ .byte 197,212,88,205 // vaddps %ymm5,%ymm5,%ymm1
+ .byte 196,65,116,92,217 // vsubps %ymm9,%ymm1,%ymm11
+ .byte 196,193,52,92,203 // vsubps %ymm11,%ymm9,%ymm1
+ .byte 196,226,125,24,29,187,69,0,0 // vbroadcastss 0x45bb(%rip),%ymm3 # 54e8 <_sk_callback_avx+0x1d6>
+ .byte 197,116,89,235 // vmulps %ymm3,%ymm1,%ymm13
.byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
.byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax
- .byte 197,249,110,216 // vmovd %eax,%xmm3
- .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
- .byte 196,227,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm5
- .byte 196,226,125,24,37,151,69,0,0 // vbroadcastss 0x4597(%rip),%ymm4 # 54ec <_sk_callback_avx+0x1da>
- .byte 197,220,92,216 // vsubps %ymm0,%ymm4,%ymm3
- .byte 197,148,89,219 // vmulps %ymm3,%ymm13,%ymm3
- .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
- .byte 197,252,194,253,1 // vcmpltps %ymm5,%ymm0,%ymm7
- .byte 196,227,37,74,219,112 // vblendvps %ymm7,%ymm3,%ymm11,%ymm3
- .byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7
- .byte 196,195,101,74,249,112 // vblendvps %ymm7,%ymm9,%ymm3,%ymm7
- .byte 196,193,121,110,216 // vmovd %r8d,%xmm3
- .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
- .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
- .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0
- .byte 196,193,116,89,205 // vmulps %ymm13,%ymm1,%ymm1
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
+ .byte 196,227,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm4
+ .byte 196,226,125,24,29,151,69,0,0 // vbroadcastss 0x4597(%rip),%ymm3 # 54ec <_sk_callback_avx+0x1da>
+ .byte 197,228,92,206 // vsubps %ymm6,%ymm3,%ymm1
+ .byte 197,148,89,201 // vmulps %ymm1,%ymm13,%ymm1
.byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
- .byte 196,227,69,74,193,0 // vblendvps %ymm0,%ymm1,%ymm7,%ymm0
+ .byte 197,204,194,252,1 // vcmpltps %ymm4,%ymm6,%ymm7
+ .byte 196,227,37,74,201,112 // vblendvps %ymm7,%ymm1,%ymm11,%ymm1
+ .byte 196,193,76,194,248,1 // vcmpltps %ymm8,%ymm6,%ymm7
+ .byte 196,195,117,74,249,112 // vblendvps %ymm7,%ymm9,%ymm1,%ymm7
+ .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
+ .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
+ .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ .byte 197,204,194,193,1 // vcmpltps %ymm1,%ymm6,%ymm0
+ .byte 197,148,89,246 // vmulps %ymm6,%ymm13,%ymm6
+ .byte 197,164,88,246 // vaddps %ymm6,%ymm11,%ymm6
+ .byte 196,227,69,74,198,0 // vblendvps %ymm0,%ymm6,%ymm7,%ymm0
.byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp)
- .byte 197,156,194,202,1 // vcmpltps %ymm2,%ymm12,%ymm1
+ .byte 197,156,194,194,1 // vcmpltps %ymm2,%ymm12,%ymm0
.byte 196,193,108,88,254 // vaddps %ymm14,%ymm2,%ymm7
- .byte 196,227,109,74,207,16 // vblendvps %ymm1,%ymm7,%ymm2,%ymm1
+ .byte 196,227,109,74,199,0 // vblendvps %ymm0,%ymm7,%ymm2,%ymm0
.byte 196,193,108,194,255,1 // vcmpltps %ymm15,%ymm2,%ymm7
- .byte 196,193,108,88,194 // vaddps %ymm10,%ymm2,%ymm0
- .byte 196,227,117,74,192,112 // vblendvps %ymm7,%ymm0,%ymm1,%ymm0
- .byte 197,220,92,200 // vsubps %ymm0,%ymm4,%ymm1
- .byte 197,148,89,201 // vmulps %ymm1,%ymm13,%ymm1
- .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
- .byte 197,252,194,253,1 // vcmpltps %ymm5,%ymm0,%ymm7
- .byte 196,227,37,74,201,112 // vblendvps %ymm7,%ymm1,%ymm11,%ymm1
+ .byte 196,193,108,88,242 // vaddps %ymm10,%ymm2,%ymm6
+ .byte 196,227,125,74,198,112 // vblendvps %ymm7,%ymm6,%ymm0,%ymm0
+ .byte 197,228,92,240 // vsubps %ymm0,%ymm3,%ymm6
+ .byte 197,148,89,246 // vmulps %ymm6,%ymm13,%ymm6
+ .byte 197,164,88,246 // vaddps %ymm6,%ymm11,%ymm6
+ .byte 197,252,194,252,1 // vcmpltps %ymm4,%ymm0,%ymm7
+ .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6
.byte 196,193,124,194,248,1 // vcmpltps %ymm8,%ymm0,%ymm7
- .byte 196,195,117,74,201,112 // vblendvps %ymm7,%ymm9,%ymm1,%ymm1
- .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0
- .byte 197,148,89,250 // vmulps %ymm2,%ymm13,%ymm7
- .byte 197,164,88,255 // vaddps %ymm7,%ymm11,%ymm7
- .byte 196,227,117,74,207,0 // vblendvps %ymm0,%ymm7,%ymm1,%ymm1
- .byte 196,226,125,24,5,237,68,0,0 // vbroadcastss 0x44ed(%rip),%ymm0 # 54f0 <_sk_callback_avx+0x1de>
+ .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6
+ .byte 197,252,194,249,1 // vcmpltps %ymm1,%ymm0,%ymm7
+ .byte 197,148,89,192 // vmulps %ymm0,%ymm13,%ymm0
+ .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
+ .byte 196,227,77,74,240,112 // vblendvps %ymm7,%ymm0,%ymm6,%ymm6
+ .byte 196,226,125,24,5,238,68,0,0 // vbroadcastss 0x44ee(%rip),%ymm0 # 54f0 <_sk_callback_avx+0x1de>
.byte 197,236,88,192 // vaddps %ymm0,%ymm2,%ymm0
.byte 197,156,194,208,1 // vcmpltps %ymm0,%ymm12,%ymm2
.byte 196,193,124,88,254 // vaddps %ymm14,%ymm0,%ymm7
.byte 196,227,125,74,215,32 // vblendvps %ymm2,%ymm7,%ymm0,%ymm2
.byte 196,193,124,194,255,1 // vcmpltps %ymm15,%ymm0,%ymm7
- .byte 196,65,124,88,210 // vaddps %ymm10,%ymm0,%ymm10
- .byte 196,195,109,74,210,112 // vblendvps %ymm7,%ymm10,%ymm2,%ymm2
- .byte 197,236,194,237,1 // vcmpltps %ymm5,%ymm2,%ymm5
- .byte 197,220,92,226 // vsubps %ymm2,%ymm4,%ymm4
- .byte 197,148,89,228 // vmulps %ymm4,%ymm13,%ymm4
- .byte 197,164,88,228 // vaddps %ymm4,%ymm11,%ymm4
- .byte 196,227,37,74,228,80 // vblendvps %ymm5,%ymm4,%ymm11,%ymm4
- .byte 196,193,108,194,232,1 // vcmpltps %ymm8,%ymm2,%ymm5
- .byte 196,195,93,74,225,80 // vblendvps %ymm5,%ymm9,%ymm4,%ymm4
- .byte 197,236,194,211,1 // vcmpltps %ymm3,%ymm2,%ymm2
- .byte 196,193,124,89,197 // vmulps %ymm13,%ymm0,%ymm0
+ .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
+ .byte 196,227,109,74,192,112 // vblendvps %ymm7,%ymm0,%ymm2,%ymm0
+ .byte 197,252,194,212,1 // vcmpltps %ymm4,%ymm0,%ymm2
+ .byte 197,228,92,216 // vsubps %ymm0,%ymm3,%ymm3
+ .byte 197,148,89,219 // vmulps %ymm3,%ymm13,%ymm3
+ .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
+ .byte 196,227,37,74,211,32 // vblendvps %ymm2,%ymm3,%ymm11,%ymm2
+ .byte 196,193,124,194,216,1 // vcmpltps %ymm8,%ymm0,%ymm3
+ .byte 196,195,109,74,209,48 // vblendvps %ymm3,%ymm9,%ymm2,%ymm2
+ .byte 197,252,194,201,1 // vcmpltps %ymm1,%ymm0,%ymm1
+ .byte 197,148,89,192 // vmulps %ymm0,%ymm13,%ymm0
.byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
- .byte 196,227,93,74,208,32 // vblendvps %ymm2,%ymm0,%ymm4,%ymm2
+ .byte 196,227,109,74,208,16 // vblendvps %ymm1,%ymm0,%ymm2,%ymm2
.byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
.byte 197,252,194,92,36,160,0 // vcmpeqps -0x60(%rsp),%ymm0,%ymm3
.byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0
- .byte 196,227,125,74,198,48 // vblendvps %ymm3,%ymm6,%ymm0,%ymm0
- .byte 196,227,117,74,206,48 // vblendvps %ymm3,%ymm6,%ymm1,%ymm1
- .byte 196,227,109,74,214,48 // vblendvps %ymm3,%ymm6,%ymm2,%ymm2
+ .byte 196,227,125,74,197,48 // vblendvps %ymm3,%ymm5,%ymm0,%ymm0
+ .byte 196,227,77,74,205,48 // vblendvps %ymm3,%ymm5,%ymm6,%ymm1
+ .byte 196,227,109,74,213,48 // vblendvps %ymm3,%ymm5,%ymm2,%ymm2
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 197,252,16,92,36,192 // vmovups -0x40(%rsp),%ymm3
.byte 197,252,16,100,36,224 // vmovups -0x20(%rsp),%ymm4
@@ -12973,14 +12974,14 @@ _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 1116 <_sk_scale_u8_avx+0x54>
+ .byte 117,68 // jne 1114 <_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 196,98,125,24,13,250,67,0,0 // vbroadcastss 0x43fa(%rip),%ymm9 # 54f4 <_sk_callback_avx+0x1e2>
+ .byte 196,98,125,24,13,252,67,0,0 // vbroadcastss 0x43fc(%rip),%ymm9 # 54f4 <_sk_callback_avx+0x1e2>
.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
@@ -12998,9 +12999,9 @@ _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 111e <_sk_scale_u8_avx+0x5c>
+ .byte 117,234 // jne 111c <_sk_scale_u8_avx+0x5c>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 235,155 // jmp 10d6 <_sk_scale_u8_avx+0x14>
+ .byte 235,155 // jmp 10d4 <_sk_scale_u8_avx+0x14>
HIDDEN _sk_lerp_1_float_avx
.globl _sk_lerp_1_float_avx
@@ -13032,14 +13033,14 @@ _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 11f2 <_sk_lerp_u8_avx+0x78>
+ .byte 117,104 // jne 11f0 <_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 196,98,125,24,13,70,67,0,0 // vbroadcastss 0x4346(%rip),%ymm9 # 54f8 <_sk_callback_avx+0x1e6>
+ .byte 196,98,125,24,13,72,67,0,0 // vbroadcastss 0x4348(%rip),%ymm9 # 54f8 <_sk_callback_avx+0x1e6>
.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
@@ -13065,9 +13066,9 @@ _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 11fa <_sk_lerp_u8_avx+0x80>
+ .byte 117,234 // jne 11f8 <_sk_lerp_u8_avx+0x80>
.byte 196,65,249,110,193 // vmovq %r9,%xmm8
- .byte 233,116,255,255,255 // jmpq 118e <_sk_lerp_u8_avx+0x14>
+ .byte 233,116,255,255,255 // jmpq 118c <_sk_lerp_u8_avx+0x14>
HIDDEN _sk_lerp_565_avx
.globl _sk_lerp_565_avx
@@ -13076,26 +13077,26 @@ _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 12d6 <_sk_lerp_565_avx+0xbc>
+ .byte 15,133,174,0,0,0 // jne 12d4 <_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 196,98,125,24,5,178,66,0,0 // vbroadcastss 0x42b2(%rip),%ymm8 # 54fc <_sk_callback_avx+0x1ea>
+ .byte 196,98,125,24,5,180,66,0,0 // vbroadcastss 0x42b4(%rip),%ymm8 # 54fc <_sk_callback_avx+0x1ea>
.byte 196,65,100,84,192 // vandps %ymm8,%ymm3,%ymm8
.byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
- .byte 196,98,125,24,13,163,66,0,0 // vbroadcastss 0x42a3(%rip),%ymm9 # 5500 <_sk_callback_avx+0x1ee>
+ .byte 196,98,125,24,13,165,66,0,0 // vbroadcastss 0x42a5(%rip),%ymm9 # 5500 <_sk_callback_avx+0x1ee>
.byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
- .byte 196,98,125,24,13,153,66,0,0 // vbroadcastss 0x4299(%rip),%ymm9 # 5504 <_sk_callback_avx+0x1f2>
+ .byte 196,98,125,24,13,155,66,0,0 // vbroadcastss 0x429b(%rip),%ymm9 # 5504 <_sk_callback_avx+0x1f2>
.byte 196,65,100,84,201 // vandps %ymm9,%ymm3,%ymm9
.byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9
- .byte 196,98,125,24,21,138,66,0,0 // vbroadcastss 0x428a(%rip),%ymm10 # 5508 <_sk_callback_avx+0x1f6>
+ .byte 196,98,125,24,21,140,66,0,0 // vbroadcastss 0x428c(%rip),%ymm10 # 5508 <_sk_callback_avx+0x1f6>
.byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
- .byte 196,98,125,24,21,128,66,0,0 // vbroadcastss 0x4280(%rip),%ymm10 # 550c <_sk_callback_avx+0x1fa>
+ .byte 196,98,125,24,21,130,66,0,0 // vbroadcastss 0x4282(%rip),%ymm10 # 550c <_sk_callback_avx+0x1fa>
.byte 196,193,100,84,218 // vandps %ymm10,%ymm3,%ymm3
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
- .byte 196,98,125,24,21,114,66,0,0 // vbroadcastss 0x4272(%rip),%ymm10 # 5510 <_sk_callback_avx+0x1fe>
+ .byte 196,98,125,24,21,116,66,0,0 // vbroadcastss 0x4274(%rip),%ymm10 # 5510 <_sk_callback_avx+0x1fe>
.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
@@ -13107,16 +13108,16 @@ _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 196,226,125,24,29,64,66,0,0 // vbroadcastss 0x4240(%rip),%ymm3 # 5514 <_sk_callback_avx+0x202>
+ .byte 196,226,125,24,29,66,66,0,0 // vbroadcastss 0x4242(%rip),%ymm3 # 5514 <_sk_callback_avx+0x202>
.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 122e <_sk_lerp_565_avx+0x14>
+ .byte 15,135,63,255,255,255 // ja 122c <_sk_lerp_565_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 1344 <_sk_lerp_565_avx+0x12a>
+ .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 1344 <_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
@@ -13128,27 +13129,26 @@ _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 122e <_sk_lerp_565_avx+0x14>
- .byte 144 // nop
- .byte 243,255 // repz (bad)
+ .byte 233,235,254,255,255 // jmpq 122c <_sk_lerp_565_avx+0x14>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 241 // icebp
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 235,255 // jmp 1349 <_sk_lerp_565_avx+0x12f>
.byte 255 // (bad)
- .byte 255,227 // jmpq *%rbx
+ .byte 233,255,255,255,225 // jmpq ffffffffe200134c <_sk_callback_avx+0xffffffffe1ffc03a>
.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
@@ -17486,7 +17486,7 @@ _sk_seed_shader_sse41:
.byte 102,15,110,199 // movd %edi,%xmm0
.byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
.byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
- .byte 15,40,21,132,57,0,0 // movaps 0x3984(%rip),%xmm2 # 3a00 <_sk_callback_sse41+0xdd>
+ .byte 15,40,21,148,57,0,0 // movaps 0x3994(%rip),%xmm2 # 3a10 <_sk_callback_sse41+0xe3>
.byte 15,88,202 // addps %xmm2,%xmm1
.byte 15,16,2 // movups (%rdx),%xmm0
.byte 15,88,193 // addps %xmm1,%xmm0
@@ -17495,7 +17495,7 @@ _sk_seed_shader_sse41:
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 15,88,202 // addps %xmm2,%xmm1
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,21,115,57,0,0 // movaps 0x3973(%rip),%xmm2 # 3a10 <_sk_callback_sse41+0xed>
+ .byte 15,40,21,131,57,0,0 // movaps 0x3983(%rip),%xmm2 # 3a20 <_sk_callback_sse41+0xf3>
.byte 15,87,219 // xorps %xmm3,%xmm3
.byte 15,87,228 // xorps %xmm4,%xmm4
.byte 15,87,237 // xorps %xmm5,%xmm5
@@ -17535,7 +17535,7 @@ HIDDEN _sk_srcatop_sse41
FUNCTION(_sk_srcatop_sse41)
_sk_srcatop_sse41:
.byte 15,89,199 // mulps %xmm7,%xmm0
- .byte 68,15,40,5,46,57,0,0 // movaps 0x392e(%rip),%xmm8 # 3a20 <_sk_callback_sse41+0xfd>
+ .byte 68,15,40,5,62,57,0,0 // movaps 0x393e(%rip),%xmm8 # 3a30 <_sk_callback_sse41+0x103>
.byte 68,15,92,195 // subps %xmm3,%xmm8
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 68,15,89,204 // mulps %xmm4,%xmm9
@@ -17560,7 +17560,7 @@ FUNCTION(_sk_dstatop_sse41)
_sk_dstatop_sse41:
.byte 68,15,40,195 // movaps %xmm3,%xmm8
.byte 68,15,89,196 // mulps %xmm4,%xmm8
- .byte 68,15,40,13,241,56,0,0 // movaps 0x38f1(%rip),%xmm9 # 3a30 <_sk_callback_sse41+0x10d>
+ .byte 68,15,40,13,1,57,0,0 // movaps 0x3901(%rip),%xmm9 # 3a40 <_sk_callback_sse41+0x113>
.byte 68,15,92,207 // subps %xmm7,%xmm9
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 65,15,88,192 // addps %xmm8,%xmm0
@@ -17607,7 +17607,7 @@ HIDDEN _sk_srcout_sse41
.globl _sk_srcout_sse41
FUNCTION(_sk_srcout_sse41)
_sk_srcout_sse41:
- .byte 68,15,40,5,149,56,0,0 // movaps 0x3895(%rip),%xmm8 # 3a40 <_sk_callback_sse41+0x11d>
+ .byte 68,15,40,5,165,56,0,0 // movaps 0x38a5(%rip),%xmm8 # 3a50 <_sk_callback_sse41+0x123>
.byte 68,15,92,199 // subps %xmm7,%xmm8
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
@@ -17620,7 +17620,7 @@ HIDDEN _sk_dstout_sse41
.globl _sk_dstout_sse41
FUNCTION(_sk_dstout_sse41)
_sk_dstout_sse41:
- .byte 68,15,40,5,133,56,0,0 // movaps 0x3885(%rip),%xmm8 # 3a50 <_sk_callback_sse41+0x12d>
+ .byte 68,15,40,5,149,56,0,0 // movaps 0x3895(%rip),%xmm8 # 3a60 <_sk_callback_sse41+0x133>
.byte 68,15,92,195 // subps %xmm3,%xmm8
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 15,89,196 // mulps %xmm4,%xmm0
@@ -17637,7 +17637,7 @@ HIDDEN _sk_srcover_sse41
.globl _sk_srcover_sse41
FUNCTION(_sk_srcover_sse41)
_sk_srcover_sse41:
- .byte 68,15,40,5,104,56,0,0 // movaps 0x3868(%rip),%xmm8 # 3a60 <_sk_callback_sse41+0x13d>
+ .byte 68,15,40,5,120,56,0,0 // movaps 0x3878(%rip),%xmm8 # 3a70 <_sk_callback_sse41+0x143>
.byte 68,15,92,195 // subps %xmm3,%xmm8
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 68,15,89,204 // mulps %xmm4,%xmm9
@@ -17657,7 +17657,7 @@ HIDDEN _sk_dstover_sse41
.globl _sk_dstover_sse41
FUNCTION(_sk_dstover_sse41)
_sk_dstover_sse41:
- .byte 68,15,40,5,60,56,0,0 // movaps 0x383c(%rip),%xmm8 # 3a70 <_sk_callback_sse41+0x14d>
+ .byte 68,15,40,5,76,56,0,0 // movaps 0x384c(%rip),%xmm8 # 3a80 <_sk_callback_sse41+0x153>
.byte 68,15,92,199 // subps %xmm7,%xmm8
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -17685,7 +17685,7 @@ HIDDEN _sk_multiply_sse41
.globl _sk_multiply_sse41
FUNCTION(_sk_multiply_sse41)
_sk_multiply_sse41:
- .byte 68,15,40,5,16,56,0,0 // movaps 0x3810(%rip),%xmm8 # 3a80 <_sk_callback_sse41+0x15d>
+ .byte 68,15,40,5,32,56,0,0 // movaps 0x3820(%rip),%xmm8 # 3a90 <_sk_callback_sse41+0x163>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 68,15,92,207 // subps %xmm7,%xmm9
.byte 69,15,40,209 // movaps %xmm9,%xmm10
@@ -17761,7 +17761,7 @@ HIDDEN _sk_xor__sse41
FUNCTION(_sk_xor__sse41)
_sk_xor__sse41:
.byte 68,15,40,195 // movaps %xmm3,%xmm8
- .byte 15,40,29,65,55,0,0 // movaps 0x3741(%rip),%xmm3 # 3a90 <_sk_callback_sse41+0x16d>
+ .byte 15,40,29,81,55,0,0 // movaps 0x3751(%rip),%xmm3 # 3aa0 <_sk_callback_sse41+0x173>
.byte 68,15,40,203 // movaps %xmm3,%xmm9
.byte 68,15,92,207 // subps %xmm7,%xmm9
.byte 65,15,89,193 // mulps %xmm9,%xmm0
@@ -17809,7 +17809,7 @@ _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 15,40,21,172,54,0,0 // movaps 0x36ac(%rip),%xmm2 # 3aa0 <_sk_callback_sse41+0x17d>
+ .byte 15,40,21,188,54,0,0 // movaps 0x36bc(%rip),%xmm2 # 3ab0 <_sk_callback_sse41+0x183>
.byte 15,92,211 // subps %xmm3,%xmm2
.byte 15,89,215 // mulps %xmm7,%xmm2
.byte 15,88,218 // addps %xmm2,%xmm3
@@ -17843,7 +17843,7 @@ _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 15,40,21,81,54,0,0 // movaps 0x3651(%rip),%xmm2 # 3ab0 <_sk_callback_sse41+0x18d>
+ .byte 15,40,21,97,54,0,0 // movaps 0x3661(%rip),%xmm2 # 3ac0 <_sk_callback_sse41+0x193>
.byte 15,92,211 // subps %xmm3,%xmm2
.byte 15,89,215 // mulps %xmm7,%xmm2
.byte 15,88,218 // addps %xmm2,%xmm3
@@ -17880,7 +17880,7 @@ _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 15,40,21,235,53,0,0 // movaps 0x35eb(%rip),%xmm2 # 3ac0 <_sk_callback_sse41+0x19d>
+ .byte 15,40,21,251,53,0,0 // movaps 0x35fb(%rip),%xmm2 # 3ad0 <_sk_callback_sse41+0x1a3>
.byte 15,92,211 // subps %xmm3,%xmm2
.byte 15,89,215 // mulps %xmm7,%xmm2
.byte 15,88,218 // addps %xmm2,%xmm3
@@ -17907,7 +17907,7 @@ _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 15,40,13,172,53,0,0 // movaps 0x35ac(%rip),%xmm1 # 3ad0 <_sk_callback_sse41+0x1ad>
+ .byte 15,40,13,188,53,0,0 // movaps 0x35bc(%rip),%xmm1 # 3ae0 <_sk_callback_sse41+0x1b3>
.byte 15,92,203 // subps %xmm3,%xmm1
.byte 15,89,207 // mulps %xmm7,%xmm1
.byte 15,88,217 // addps %xmm1,%xmm3
@@ -17921,7 +17921,7 @@ HIDDEN _sk_colorburn_sse41
FUNCTION(_sk_colorburn_sse41)
_sk_colorburn_sse41:
.byte 68,15,40,192 // movaps %xmm0,%xmm8
- .byte 68,15,40,21,155,53,0,0 // movaps 0x359b(%rip),%xmm10 # 3ae0 <_sk_callback_sse41+0x1bd>
+ .byte 68,15,40,21,171,53,0,0 // movaps 0x35ab(%rip),%xmm10 # 3af0 <_sk_callback_sse41+0x1c3>
.byte 69,15,40,218 // movaps %xmm10,%xmm11
.byte 68,15,92,223 // subps %xmm7,%xmm11
.byte 69,15,40,203 // movaps %xmm11,%xmm9
@@ -18003,7 +18003,7 @@ HIDDEN _sk_colordodge_sse41
FUNCTION(_sk_colordodge_sse41)
_sk_colordodge_sse41:
.byte 68,15,40,192 // movaps %xmm0,%xmm8
- .byte 68,15,40,21,121,52,0,0 // movaps 0x3479(%rip),%xmm10 # 3af0 <_sk_callback_sse41+0x1cd>
+ .byte 68,15,40,21,137,52,0,0 // movaps 0x3489(%rip),%xmm10 # 3b00 <_sk_callback_sse41+0x1d3>
.byte 69,15,40,218 // movaps %xmm10,%xmm11
.byte 68,15,92,223 // subps %xmm7,%xmm11
.byte 69,15,40,227 // movaps %xmm11,%xmm12
@@ -18085,7 +18085,7 @@ _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 68,15,40,21,82,51,0,0 // movaps 0x3352(%rip),%xmm10 # 3b00 <_sk_callback_sse41+0x1dd>
+ .byte 68,15,40,21,98,51,0,0 // movaps 0x3362(%rip),%xmm10 # 3b10 <_sk_callback_sse41+0x1e3>
.byte 65,15,40,234 // movaps %xmm10,%xmm5
.byte 15,92,239 // subps %xmm7,%xmm5
.byte 15,40,197 // movaps %xmm5,%xmm0
@@ -18168,7 +18168,7 @@ FUNCTION(_sk_overlay_sse41)
_sk_overlay_sse41:
.byte 68,15,40,201 // movaps %xmm1,%xmm9
.byte 68,15,40,240 // movaps %xmm0,%xmm14
- .byte 68,15,40,21,55,50,0,0 // movaps 0x3237(%rip),%xmm10 # 3b10 <_sk_callback_sse41+0x1ed>
+ .byte 68,15,40,21,71,50,0,0 // movaps 0x3247(%rip),%xmm10 # 3b20 <_sk_callback_sse41+0x1f3>
.byte 69,15,40,218 // movaps %xmm10,%xmm11
.byte 68,15,92,223 // subps %xmm7,%xmm11
.byte 65,15,40,195 // movaps %xmm11,%xmm0
@@ -18253,7 +18253,7 @@ _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 15,40,13,14,49,0,0 // movaps 0x310e(%rip),%xmm1 # 3b20 <_sk_callback_sse41+0x1fd>
+ .byte 15,40,13,30,49,0,0 // movaps 0x311e(%rip),%xmm1 # 3b30 <_sk_callback_sse41+0x203>
.byte 68,15,40,209 // movaps %xmm1,%xmm10
.byte 68,15,92,208 // subps %xmm0,%xmm10
.byte 68,15,40,240 // movaps %xmm0,%xmm14
@@ -18266,10 +18266,10 @@ _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 68,15,40,45,236,48,0,0 // movaps 0x30ec(%rip),%xmm13 # 3b30 <_sk_callback_sse41+0x20d>
+ .byte 68,15,40,45,252,48,0,0 // movaps 0x30fc(%rip),%xmm13 # 3b40 <_sk_callback_sse41+0x213>
.byte 69,15,88,245 // addps %xmm13,%xmm14
.byte 68,15,89,242 // mulps %xmm2,%xmm14
- .byte 68,15,40,37,236,48,0,0 // movaps 0x30ec(%rip),%xmm12 # 3b40 <_sk_callback_sse41+0x21d>
+ .byte 68,15,40,37,252,48,0,0 // movaps 0x30fc(%rip),%xmm12 # 3b50 <_sk_callback_sse41+0x223>
.byte 69,15,89,252 // mulps %xmm12,%xmm15
.byte 69,15,88,254 // addps %xmm14,%xmm15
.byte 15,40,198 // movaps %xmm6,%xmm0
@@ -18417,7 +18417,7 @@ HIDDEN _sk_clamp_1_sse41
.globl _sk_clamp_1_sse41
FUNCTION(_sk_clamp_1_sse41)
_sk_clamp_1_sse41:
- .byte 68,15,40,5,254,46,0,0 // movaps 0x2efe(%rip),%xmm8 # 3b50 <_sk_callback_sse41+0x22d>
+ .byte 68,15,40,5,14,47,0,0 // movaps 0x2f0e(%rip),%xmm8 # 3b60 <_sk_callback_sse41+0x233>
.byte 65,15,93,192 // minps %xmm8,%xmm0
.byte 65,15,93,200 // minps %xmm8,%xmm1
.byte 65,15,93,208 // minps %xmm8,%xmm2
@@ -18429,7 +18429,7 @@ HIDDEN _sk_clamp_a_sse41
.globl _sk_clamp_a_sse41
FUNCTION(_sk_clamp_a_sse41)
_sk_clamp_a_sse41:
- .byte 15,93,29,243,46,0,0 // minps 0x2ef3(%rip),%xmm3 # 3b60 <_sk_callback_sse41+0x23d>
+ .byte 15,93,29,3,47,0,0 // minps 0x2f03(%rip),%xmm3 # 3b70 <_sk_callback_sse41+0x243>
.byte 15,93,195 // minps %xmm3,%xmm0
.byte 15,93,203 // minps %xmm3,%xmm1
.byte 15,93,211 // minps %xmm3,%xmm2
@@ -18516,7 +18516,7 @@ HIDDEN _sk_unpremul_sse41
FUNCTION(_sk_unpremul_sse41)
_sk_unpremul_sse41:
.byte 69,15,87,192 // xorps %xmm8,%xmm8
- .byte 68,15,40,13,94,46,0,0 // movaps 0x2e5e(%rip),%xmm9 # 3b70 <_sk_callback_sse41+0x24d>
+ .byte 68,15,40,13,110,46,0,0 // movaps 0x2e6e(%rip),%xmm9 # 3b80 <_sk_callback_sse41+0x253>
.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
@@ -18530,20 +18530,20 @@ HIDDEN _sk_from_srgb_sse41
.globl _sk_from_srgb_sse41
FUNCTION(_sk_from_srgb_sse41)
_sk_from_srgb_sse41:
- .byte 68,15,40,29,73,46,0,0 // movaps 0x2e49(%rip),%xmm11 # 3b80 <_sk_callback_sse41+0x25d>
+ .byte 68,15,40,29,89,46,0,0 // movaps 0x2e59(%rip),%xmm11 # 3b90 <_sk_callback_sse41+0x263>
.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 68,15,40,37,65,46,0,0 // movaps 0x2e41(%rip),%xmm12 # 3b90 <_sk_callback_sse41+0x26d>
+ .byte 68,15,40,37,81,46,0,0 // movaps 0x2e51(%rip),%xmm12 # 3ba0 <_sk_callback_sse41+0x273>
.byte 68,15,40,192 // movaps %xmm0,%xmm8
.byte 69,15,89,196 // mulps %xmm12,%xmm8
- .byte 68,15,40,45,65,46,0,0 // movaps 0x2e41(%rip),%xmm13 # 3ba0 <_sk_callback_sse41+0x27d>
+ .byte 68,15,40,45,81,46,0,0 // movaps 0x2e51(%rip),%xmm13 # 3bb0 <_sk_callback_sse41+0x283>
.byte 69,15,88,197 // addps %xmm13,%xmm8
.byte 69,15,89,194 // mulps %xmm10,%xmm8
- .byte 68,15,40,53,65,46,0,0 // movaps 0x2e41(%rip),%xmm14 # 3bb0 <_sk_callback_sse41+0x28d>
+ .byte 68,15,40,53,81,46,0,0 // movaps 0x2e51(%rip),%xmm14 # 3bc0 <_sk_callback_sse41+0x293>
.byte 69,15,88,198 // addps %xmm14,%xmm8
- .byte 68,15,40,61,69,46,0,0 // movaps 0x2e45(%rip),%xmm15 # 3bc0 <_sk_callback_sse41+0x29d>
+ .byte 68,15,40,61,85,46,0,0 // movaps 0x2e55(%rip),%xmm15 # 3bd0 <_sk_callback_sse41+0x2a3>
.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
@@ -18588,20 +18588,20 @@ _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 68,15,40,29,181,45,0,0 // movaps 0x2db5(%rip),%xmm11 # 3bd0 <_sk_callback_sse41+0x2ad>
+ .byte 68,15,40,29,197,45,0,0 // movaps 0x2dc5(%rip),%xmm11 # 3be0 <_sk_callback_sse41+0x2b3>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 65,15,89,203 // mulps %xmm11,%xmm1
- .byte 68,15,40,37,182,45,0,0 // movaps 0x2db6(%rip),%xmm12 # 3be0 <_sk_callback_sse41+0x2bd>
+ .byte 68,15,40,37,198,45,0,0 // movaps 0x2dc6(%rip),%xmm12 # 3bf0 <_sk_callback_sse41+0x2c3>
.byte 69,15,89,204 // mulps %xmm12,%xmm9
- .byte 68,15,40,45,186,45,0,0 // movaps 0x2dba(%rip),%xmm13 # 3bf0 <_sk_callback_sse41+0x2cd>
+ .byte 68,15,40,45,202,45,0,0 // movaps 0x2dca(%rip),%xmm13 # 3c00 <_sk_callback_sse41+0x2d3>
.byte 69,15,88,205 // addps %xmm13,%xmm9
- .byte 68,15,40,53,190,45,0,0 // movaps 0x2dbe(%rip),%xmm14 # 3c00 <_sk_callback_sse41+0x2dd>
+ .byte 68,15,40,53,206,45,0,0 // movaps 0x2dce(%rip),%xmm14 # 3c10 <_sk_callback_sse41+0x2e3>
.byte 69,15,89,214 // mulps %xmm14,%xmm10
.byte 69,15,88,209 // addps %xmm9,%xmm10
- .byte 68,15,40,5,190,45,0,0 // movaps 0x2dbe(%rip),%xmm8 # 3c10 <_sk_callback_sse41+0x2ed>
+ .byte 68,15,40,5,206,45,0,0 // movaps 0x2dce(%rip),%xmm8 # 3c20 <_sk_callback_sse41+0x2f3>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,93,202 // minps %xmm10,%xmm9
- .byte 68,15,40,61,190,45,0,0 // movaps 0x2dbe(%rip),%xmm15 # 3c20 <_sk_callback_sse41+0x2fd>
+ .byte 68,15,40,61,206,45,0,0 // movaps 0x2dce(%rip),%xmm15 # 3c30 <_sk_callback_sse41+0x303>
.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
@@ -18655,7 +18655,7 @@ _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 68,15,40,53,15,45,0,0 // movaps 0x2d0f(%rip),%xmm14 # 3c30 <_sk_callback_sse41+0x30d>
+ .byte 68,15,40,53,31,45,0,0 // movaps 0x2d1f(%rip),%xmm14 # 3c40 <_sk_callback_sse41+0x313>
.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
@@ -18664,27 +18664,27 @@ _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 68,15,84,61,246,44,0,0 // andps 0x2cf6(%rip),%xmm15 # 3c40 <_sk_callback_sse41+0x31d>
+ .byte 68,15,84,61,6,45,0,0 // andps 0x2d06(%rip),%xmm15 # 3c50 <_sk_callback_sse41+0x323>
.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 68,15,40,45,233,44,0,0 // movaps 0x2ce9(%rip),%xmm13 # 3c50 <_sk_callback_sse41+0x32d>
+ .byte 68,15,40,45,249,44,0,0 // movaps 0x2cf9(%rip),%xmm13 # 3c60 <_sk_callback_sse41+0x333>
.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 68,15,88,5,229,44,0,0 // addps 0x2ce5(%rip),%xmm8 # 3c60 <_sk_callback_sse41+0x33d>
+ .byte 68,15,88,5,245,44,0,0 // addps 0x2cf5(%rip),%xmm8 # 3c70 <_sk_callback_sse41+0x343>
.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 68,15,89,5,221,44,0,0 // mulps 0x2cdd(%rip),%xmm8 # 3c70 <_sk_callback_sse41+0x34d>
+ .byte 68,15,89,5,237,44,0,0 // mulps 0x2ced(%rip),%xmm8 # 3c80 <_sk_callback_sse41+0x353>
.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 15,40,5,209,44,0,0 // movaps 0x2cd1(%rip),%xmm0 # 3c80 <_sk_callback_sse41+0x35d>
+ .byte 15,40,5,225,44,0,0 // movaps 0x2ce1(%rip),%xmm0 # 3c90 <_sk_callback_sse41+0x363>
.byte 65,15,40,211 // movaps %xmm11,%xmm2
.byte 15,89,208 // mulps %xmm0,%xmm2
.byte 15,194,194,1 // cmpltps %xmm2,%xmm0
@@ -18713,140 +18713,141 @@ _sk_hsl_to_rgb_sse41:
.byte 15,41,92,36,128 // movaps %xmm3,-0x80(%rsp)
.byte 15,40,194 // movaps %xmm2,%xmm0
.byte 15,194,195,1 // cmpltps %xmm3,%xmm0
- .byte 15,40,45,124,44,0,0 // movaps 0x2c7c(%rip),%xmm5 # 3c90 <_sk_callback_sse41+0x36d>
- .byte 15,40,241 // movaps %xmm1,%xmm6
+ .byte 15,40,45,140,44,0,0 // movaps 0x2c8c(%rip),%xmm5 # 3ca0 <_sk_callback_sse41+0x373>
+ .byte 15,40,249 // movaps %xmm1,%xmm7
.byte 15,40,225 // movaps %xmm1,%xmm4
.byte 15,40,217 // movaps %xmm1,%xmm3
.byte 15,88,221 // addps %xmm5,%xmm3
- .byte 15,40,253 // movaps %xmm5,%xmm7
+ .byte 15,40,245 // movaps %xmm5,%xmm6
.byte 15,89,218 // mulps %xmm2,%xmm3
- .byte 15,88,242 // addps %xmm2,%xmm6
+ .byte 15,88,250 // addps %xmm2,%xmm7
.byte 15,89,226 // mulps %xmm2,%xmm4
.byte 15,40,234 // movaps %xmm2,%xmm5
- .byte 15,92,244 // subps %xmm4,%xmm6
- .byte 102,15,56,20,243 // blendvps %xmm0,%xmm3,%xmm6
- .byte 68,15,40,61,97,44,0,0 // movaps 0x2c61(%rip),%xmm15 # 3ca0 <_sk_callback_sse41+0x37d>
- .byte 69,15,88,251 // addps %xmm11,%xmm15
+ .byte 15,92,252 // subps %xmm4,%xmm7
+ .byte 102,15,56,20,251 // blendvps %xmm0,%xmm3,%xmm7
+ .byte 68,15,40,37,113,44,0,0 // movaps 0x2c71(%rip),%xmm12 # 3cb0 <_sk_callback_sse41+0x383>
+ .byte 69,15,88,227 // addps %xmm11,%xmm12
.byte 184,0,0,0,0 // mov $0x0,%eax
.byte 185,0,0,128,63 // mov $0x3f800000,%ecx
.byte 102,68,15,110,201 // movd %ecx,%xmm9
.byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
.byte 65,15,40,193 // movaps %xmm9,%xmm0
- .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0
- .byte 65,15,40,215 // movaps %xmm15,%xmm2
- .byte 15,88,21,69,44,0,0 // addps 0x2c45(%rip),%xmm2 # 3cb0 <_sk_callback_sse41+0x38d>
- .byte 69,15,40,231 // movaps %xmm15,%xmm12
- .byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12
- .byte 102,15,110,208 // movd %eax,%xmm2
- .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
- .byte 15,41,84,36,160 // movaps %xmm2,-0x60(%rsp)
- .byte 65,15,40,199 // movaps %xmm15,%xmm0
- .byte 15,194,194,1 // cmpltps %xmm2,%xmm0
- .byte 65,15,40,215 // movaps %xmm15,%xmm2
- .byte 15,88,215 // addps %xmm7,%xmm2
+ .byte 65,15,194,196,1 // cmpltps %xmm12,%xmm0
+ .byte 65,15,40,212 // movaps %xmm12,%xmm2
+ .byte 15,88,21,85,44,0,0 // addps 0x2c55(%rip),%xmm2 # 3cc0 <_sk_callback_sse41+0x393>
+ .byte 69,15,40,196 // movaps %xmm12,%xmm8
+ .byte 65,15,40,220 // movaps %xmm12,%xmm3
.byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12
- .byte 15,40,221 // movaps %xmm5,%xmm3
- .byte 15,41,92,36,176 // movaps %xmm3,-0x50(%rsp)
- .byte 68,15,40,195 // movaps %xmm3,%xmm8
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
+ .byte 15,41,68,36,160 // movaps %xmm0,-0x60(%rsp)
+ .byte 68,15,194,192,1 // cmpltps %xmm0,%xmm8
+ .byte 15,88,222 // addps %xmm6,%xmm3
+ .byte 65,15,40,192 // movaps %xmm8,%xmm0
+ .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12
+ .byte 15,40,213 // movaps %xmm5,%xmm2
+ .byte 15,41,84,36,176 // movaps %xmm2,-0x50(%rsp)
+ .byte 68,15,40,194 // movaps %xmm2,%xmm8
.byte 69,15,88,192 // addps %xmm8,%xmm8
- .byte 68,15,92,198 // subps %xmm6,%xmm8
+ .byte 68,15,92,199 // subps %xmm7,%xmm8
.byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax
- .byte 15,40,214 // movaps %xmm6,%xmm2
- .byte 65,15,92,208 // subps %xmm8,%xmm2
- .byte 15,89,21,2,44,0,0 // mulps 0x2c02(%rip),%xmm2 # 3cc0 <_sk_callback_sse41+0x39d>
+ .byte 15,40,247 // movaps %xmm7,%xmm6
+ .byte 65,15,92,240 // subps %xmm8,%xmm6
+ .byte 15,89,53,17,44,0,0 // mulps 0x2c11(%rip),%xmm6 # 3cd0 <_sk_callback_sse41+0x3a3>
.byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx
.byte 102,15,110,193 // movd %ecx,%xmm0
.byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
.byte 15,41,68,36,144 // movaps %xmm0,-0x70(%rsp)
- .byte 15,40,37,249,43,0,0 // movaps 0x2bf9(%rip),%xmm4 # 3cd0 <_sk_callback_sse41+0x3ad>
+ .byte 15,40,37,8,44,0,0 // movaps 0x2c08(%rip),%xmm4 # 3ce0 <_sk_callback_sse41+0x3b3>
.byte 15,40,236 // movaps %xmm4,%xmm5
.byte 65,15,92,236 // subps %xmm12,%xmm5
.byte 69,15,40,236 // movaps %xmm12,%xmm13
+ .byte 69,15,40,252 // movaps %xmm12,%xmm15
.byte 69,15,40,244 // movaps %xmm12,%xmm14
.byte 68,15,194,224,1 // cmpltps %xmm0,%xmm12
- .byte 15,89,234 // mulps %xmm2,%xmm5
+ .byte 15,89,238 // mulps %xmm6,%xmm5
.byte 65,15,88,232 // addps %xmm8,%xmm5
.byte 69,15,40,208 // movaps %xmm8,%xmm10
.byte 65,15,40,196 // movaps %xmm12,%xmm0
.byte 102,68,15,56,20,213 // blendvps %xmm0,%xmm5,%xmm10
- .byte 15,40,108,36,128 // movaps -0x80(%rsp),%xmm5
- .byte 68,15,194,245,1 // cmpltps %xmm5,%xmm14
+ .byte 68,15,194,116,36,128,1 // cmpltps -0x80(%rsp),%xmm14
.byte 65,15,40,198 // movaps %xmm14,%xmm0
- .byte 102,68,15,56,20,214 // blendvps %xmm0,%xmm6,%xmm10
- .byte 102,15,110,248 // movd %eax,%xmm7
- .byte 15,198,255,0 // shufps $0x0,%xmm7,%xmm7
- .byte 68,15,194,239,1 // cmpltps %xmm7,%xmm13
- .byte 68,15,89,250 // mulps %xmm2,%xmm15
+ .byte 102,68,15,56,20,215 // blendvps %xmm0,%xmm7,%xmm10
+ .byte 102,15,110,232 // movd %eax,%xmm5
+ .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5
+ .byte 68,15,194,237,1 // cmpltps %xmm5,%xmm13
+ .byte 68,15,89,254 // mulps %xmm6,%xmm15
.byte 69,15,88,248 // addps %xmm8,%xmm15
.byte 65,15,40,197 // movaps %xmm13,%xmm0
.byte 102,69,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm10
.byte 69,15,87,228 // xorps %xmm12,%xmm12
.byte 68,15,194,225,0 // cmpeqps %xmm1,%xmm12
.byte 65,15,40,196 // movaps %xmm12,%xmm0
- .byte 102,68,15,56,20,211 // blendvps %xmm0,%xmm3,%xmm10
+ .byte 102,68,15,56,20,210 // blendvps %xmm0,%xmm2,%xmm10
.byte 65,15,40,193 // movaps %xmm9,%xmm0
.byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0
.byte 65,15,40,203 // movaps %xmm11,%xmm1
- .byte 15,88,13,86,43,0,0 // addps 0x2b56(%rip),%xmm1 # 3cb0 <_sk_callback_sse41+0x38d>
+ .byte 15,88,13,100,43,0,0 // addps 0x2b64(%rip),%xmm1 # 3cc0 <_sk_callback_sse41+0x393>
.byte 69,15,40,235 // movaps %xmm11,%xmm13
.byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13
.byte 65,15,40,195 // movaps %xmm11,%xmm0
.byte 15,194,68,36,160,1 // cmpltps -0x60(%rsp),%xmm0
.byte 65,15,40,203 // movaps %xmm11,%xmm1
- .byte 15,88,13,23,43,0,0 // addps 0x2b17(%rip),%xmm1 # 3c90 <_sk_callback_sse41+0x36d>
+ .byte 15,88,13,37,43,0,0 // addps 0x2b25(%rip),%xmm1 # 3ca0 <_sk_callback_sse41+0x373>
.byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13
.byte 15,40,220 // movaps %xmm4,%xmm3
.byte 65,15,92,221 // subps %xmm13,%xmm3
+ .byte 65,15,40,213 // movaps %xmm13,%xmm2
.byte 69,15,40,245 // movaps %xmm13,%xmm14
.byte 69,15,40,253 // movaps %xmm13,%xmm15
.byte 68,15,194,108,36,144,1 // cmpltps -0x70(%rsp),%xmm13
- .byte 15,89,218 // mulps %xmm2,%xmm3
+ .byte 15,89,222 // mulps %xmm6,%xmm3
.byte 65,15,88,216 // addps %xmm8,%xmm3
.byte 65,15,40,200 // movaps %xmm8,%xmm1
.byte 65,15,40,197 // movaps %xmm13,%xmm0
.byte 102,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm1
- .byte 68,15,194,253,1 // cmpltps %xmm5,%xmm15
+ .byte 68,15,194,124,36,128,1 // cmpltps -0x80(%rsp),%xmm15
.byte 65,15,40,199 // movaps %xmm15,%xmm0
- .byte 102,15,56,20,206 // blendvps %xmm0,%xmm6,%xmm1
- .byte 68,15,194,247,1 // cmpltps %xmm7,%xmm14
- .byte 15,40,218 // movaps %xmm2,%xmm3
- .byte 65,15,89,219 // mulps %xmm11,%xmm3
- .byte 65,15,88,216 // addps %xmm8,%xmm3
- .byte 65,15,40,198 // movaps %xmm14,%xmm0
- .byte 102,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm1
+ .byte 102,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm1
+ .byte 15,194,213,1 // cmpltps %xmm5,%xmm2
+ .byte 68,15,89,246 // mulps %xmm6,%xmm14
+ .byte 69,15,88,240 // addps %xmm8,%xmm14
+ .byte 15,40,194 // movaps %xmm2,%xmm0
+ .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1
.byte 65,15,40,196 // movaps %xmm12,%xmm0
- .byte 15,40,108,36,176 // movaps -0x50(%rsp),%xmm5
- .byte 102,15,56,20,205 // blendvps %xmm0,%xmm5,%xmm1
- .byte 68,15,88,29,250,42,0,0 // addps 0x2afa(%rip),%xmm11 # 3ce0 <_sk_callback_sse41+0x3bd>
+ .byte 68,15,40,116,36,176 // movaps -0x50(%rsp),%xmm14
+ .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1
+ .byte 68,15,88,29,4,43,0,0 // addps 0x2b04(%rip),%xmm11 # 3cf0 <_sk_callback_sse41+0x3c3>
+ .byte 15,40,21,173,42,0,0 // movaps 0x2aad(%rip),%xmm2 # 3ca0 <_sk_callback_sse41+0x373>
+ .byte 65,15,88,211 // addps %xmm11,%xmm2
.byte 69,15,194,203,1 // cmpltps %xmm11,%xmm9
- .byte 15,40,29,190,42,0,0 // movaps 0x2abe(%rip),%xmm3 # 3cb0 <_sk_callback_sse41+0x38d>
+ .byte 15,40,29,189,42,0,0 // movaps 0x2abd(%rip),%xmm3 # 3cc0 <_sk_callback_sse41+0x393>
.byte 65,15,88,219 // addps %xmm11,%xmm3
.byte 69,15,40,235 // movaps %xmm11,%xmm13
.byte 65,15,40,193 // movaps %xmm9,%xmm0
- .byte 102,68,15,56,20,235 // blendvps %xmm0,%xmm3,%xmm13
- .byte 65,15,40,195 // movaps %xmm11,%xmm0
- .byte 15,194,68,36,160,1 // cmpltps -0x60(%rsp),%xmm0
- .byte 15,40,29,123,42,0,0 // movaps 0x2a7b(%rip),%xmm3 # 3c90 <_sk_callback_sse41+0x36d>
- .byte 65,15,88,219 // addps %xmm11,%xmm3
- .byte 102,68,15,56,20,235 // blendvps %xmm0,%xmm3,%xmm13
- .byte 65,15,92,229 // subps %xmm13,%xmm4
- .byte 69,15,40,205 // movaps %xmm13,%xmm9
- .byte 69,15,40,245 // movaps %xmm13,%xmm14
- .byte 68,15,194,108,36,144,1 // cmpltps -0x70(%rsp),%xmm13
- .byte 68,15,89,218 // mulps %xmm2,%xmm11
- .byte 15,89,226 // mulps %xmm2,%xmm4
- .byte 69,15,88,216 // addps %xmm8,%xmm11
- .byte 65,15,88,224 // addps %xmm8,%xmm4
+ .byte 102,68,15,56,20,219 // blendvps %xmm0,%xmm3,%xmm11
+ .byte 68,15,194,108,36,160,1 // cmpltps -0x60(%rsp),%xmm13
.byte 65,15,40,197 // movaps %xmm13,%xmm0
+ .byte 102,68,15,56,20,218 // blendvps %xmm0,%xmm2,%xmm11
+ .byte 65,15,92,227 // subps %xmm11,%xmm4
+ .byte 69,15,40,203 // movaps %xmm11,%xmm9
+ .byte 65,15,40,211 // movaps %xmm11,%xmm2
+ .byte 69,15,40,235 // movaps %xmm11,%xmm13
+ .byte 68,15,194,92,36,144,1 // cmpltps -0x70(%rsp),%xmm11
+ .byte 15,89,214 // mulps %xmm6,%xmm2
+ .byte 15,89,230 // mulps %xmm6,%xmm4
+ .byte 65,15,88,208 // addps %xmm8,%xmm2
+ .byte 65,15,88,224 // addps %xmm8,%xmm4
+ .byte 65,15,40,195 // movaps %xmm11,%xmm0
.byte 102,68,15,56,20,196 // blendvps %xmm0,%xmm4,%xmm8
- .byte 68,15,194,116,36,128,1 // cmpltps -0x80(%rsp),%xmm14
- .byte 65,15,40,198 // movaps %xmm14,%xmm0
- .byte 102,68,15,56,20,198 // blendvps %xmm0,%xmm6,%xmm8
- .byte 68,15,194,207,1 // cmpltps %xmm7,%xmm9
+ .byte 68,15,194,108,36,128,1 // cmpltps -0x80(%rsp),%xmm13
+ .byte 65,15,40,197 // movaps %xmm13,%xmm0
+ .byte 102,68,15,56,20,199 // blendvps %xmm0,%xmm7,%xmm8
+ .byte 68,15,194,205,1 // cmpltps %xmm5,%xmm9
.byte 65,15,40,193 // movaps %xmm9,%xmm0
- .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8
+ .byte 102,68,15,56,20,194 // blendvps %xmm0,%xmm2,%xmm8
.byte 65,15,40,196 // movaps %xmm12,%xmm0
- .byte 102,68,15,56,20,197 // blendvps %xmm0,%xmm5,%xmm8
+ .byte 102,69,15,56,20,198 // blendvps %xmm0,%xmm14,%xmm8
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,194 // movaps %xmm10,%xmm0
.byte 65,15,40,208 // movaps %xmm8,%xmm2
@@ -18880,7 +18881,7 @@ _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 68,15,89,5,27,42,0,0 // mulps 0x2a1b(%rip),%xmm8 # 3cf0 <_sk_callback_sse41+0x3cd>
+ .byte 68,15,89,5,33,42,0,0 // mulps 0x2a21(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x3d3>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 65,15,89,208 // mulps %xmm8,%xmm2
@@ -18918,7 +18919,7 @@ _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 68,15,89,5,199,41,0,0 // mulps 0x29c7(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x3dd>
+ .byte 68,15,89,5,205,41,0,0 // mulps 0x29cd(%rip),%xmm8 # 3d10 <_sk_callback_sse41+0x3e3>
.byte 15,92,196 // subps %xmm4,%xmm0
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -18941,17 +18942,17 @@ _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 102,15,111,29,151,41,0,0 // movdqa 0x2997(%rip),%xmm3 # 3d10 <_sk_callback_sse41+0x3ed>
+ .byte 102,15,111,29,157,41,0,0 // movdqa 0x299d(%rip),%xmm3 # 3d20 <_sk_callback_sse41+0x3f3>
.byte 102,65,15,219,216 // pand %xmm8,%xmm3
.byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
- .byte 68,15,89,13,150,41,0,0 // mulps 0x2996(%rip),%xmm9 # 3d20 <_sk_callback_sse41+0x3fd>
- .byte 102,15,111,29,158,41,0,0 // movdqa 0x299e(%rip),%xmm3 # 3d30 <_sk_callback_sse41+0x40d>
+ .byte 68,15,89,13,156,41,0,0 // mulps 0x299c(%rip),%xmm9 # 3d30 <_sk_callback_sse41+0x403>
+ .byte 102,15,111,29,164,41,0,0 // movdqa 0x29a4(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x413>
.byte 102,65,15,219,216 // pand %xmm8,%xmm3
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,159,41,0,0 // mulps 0x299f(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x41d>
- .byte 102,68,15,219,5,166,41,0,0 // pand 0x29a6(%rip),%xmm8 # 3d50 <_sk_callback_sse41+0x42d>
+ .byte 15,89,29,165,41,0,0 // mulps 0x29a5(%rip),%xmm3 # 3d50 <_sk_callback_sse41+0x423>
+ .byte 102,68,15,219,5,172,41,0,0 // pand 0x29ac(%rip),%xmm8 # 3d60 <_sk_callback_sse41+0x433>
.byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
- .byte 68,15,89,5,170,41,0,0 // mulps 0x29aa(%rip),%xmm8 # 3d60 <_sk_callback_sse41+0x43d>
+ .byte 68,15,89,5,176,41,0,0 // mulps 0x29b0(%rip),%xmm8 # 3d70 <_sk_callback_sse41+0x443>
.byte 15,92,196 // subps %xmm4,%xmm0
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 15,88,196 // addps %xmm4,%xmm0
@@ -18962,7 +18963,7 @@ _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 15,40,29,148,41,0,0 // movaps 0x2994(%rip),%xmm3 # 3d70 <_sk_callback_sse41+0x44d>
+ .byte 15,40,29,154,41,0,0 // movaps 0x299a(%rip),%xmm3 # 3d80 <_sk_callback_sse41+0x453>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_load_tables_sse41
@@ -18973,7 +18974,7 @@ _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 102,15,111,5,139,41,0,0 // movdqa 0x298b(%rip),%xmm0 # 3d80 <_sk_callback_sse41+0x45d>
+ .byte 102,15,111,5,145,41,0,0 // movdqa 0x2991(%rip),%xmm0 # 3d90 <_sk_callback_sse41+0x463>
.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
@@ -18988,7 +18989,7 @@ _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 102,15,56,0,13,70,41,0,0 // pshufb 0x2946(%rip),%xmm1 # 3d90 <_sk_callback_sse41+0x46d>
+ .byte 102,15,56,0,13,76,41,0,0 // pshufb 0x294c(%rip),%xmm1 # 3da0 <_sk_callback_sse41+0x473>
.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
@@ -19003,7 +19004,7 @@ _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 102,15,56,0,21,2,41,0,0 // pshufb 0x2902(%rip),%xmm2 # 3da0 <_sk_callback_sse41+0x47d>
+ .byte 102,15,56,0,21,8,41,0,0 // pshufb 0x2908(%rip),%xmm2 # 3db0 <_sk_callback_sse41+0x483>
.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
@@ -19018,7 +19019,7 @@ _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 15,89,29,191,40,0,0 // mulps 0x28bf(%rip),%xmm3 # 3db0 <_sk_callback_sse41+0x48d>
+ .byte 15,89,29,197,40,0,0 // mulps 0x28c5(%rip),%xmm3 # 3dc0 <_sk_callback_sse41+0x493>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -19037,7 +19038,7 @@ _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 102,68,15,111,5,146,40,0,0 // movdqa 0x2892(%rip),%xmm8 # 3dc0 <_sk_callback_sse41+0x49d>
+ .byte 102,68,15,111,5,152,40,0,0 // movdqa 0x2898(%rip),%xmm8 # 3dd0 <_sk_callback_sse41+0x4a3>
.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
@@ -19054,7 +19055,7 @@ _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 102,15,56,0,13,69,40,0,0 // pshufb 0x2845(%rip),%xmm1 # 3dd0 <_sk_callback_sse41+0x4ad>
+ .byte 102,15,56,0,13,75,40,0,0 // pshufb 0x284b(%rip),%xmm1 # 3de0 <_sk_callback_sse41+0x4b3>
.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
@@ -19090,7 +19091,7 @@ _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 15,89,29,147,39,0,0 // mulps 0x2793(%rip),%xmm3 # 3de0 <_sk_callback_sse41+0x4bd>
+ .byte 15,89,29,153,39,0,0 // mulps 0x2799(%rip),%xmm3 # 3df0 <_sk_callback_sse41+0x4c3>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -19112,7 +19113,7 @@ _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 102,68,15,111,5,85,39,0,0 // movdqa 0x2755(%rip),%xmm8 # 3df0 <_sk_callback_sse41+0x4cd>
+ .byte 102,68,15,111,5,91,39,0,0 // movdqa 0x275b(%rip),%xmm8 # 3e00 <_sk_callback_sse41+0x4d3>
.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
@@ -19129,7 +19130,7 @@ _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 102,15,56,0,13,8,39,0,0 // pshufb 0x2708(%rip),%xmm1 # 3e00 <_sk_callback_sse41+0x4dd>
+ .byte 102,15,56,0,13,14,39,0,0 // pshufb 0x270e(%rip),%xmm1 # 3e10 <_sk_callback_sse41+0x4e3>
.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
@@ -19160,7 +19161,7 @@ _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 15,40,29,115,38,0,0 // movaps 0x2673(%rip),%xmm3 # 3e10 <_sk_callback_sse41+0x4ed>
+ .byte 15,40,29,121,38,0,0 // movaps 0x2679(%rip),%xmm3 # 3e20 <_sk_callback_sse41+0x4f3>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_byte_tables_sse41
@@ -19170,7 +19171,7 @@ _sk_byte_tables_sse41:
.byte 65,86 // push %r14
.byte 83 // push %rbx
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,116,38,0,0 // movaps 0x2674(%rip),%xmm8 # 3e20 <_sk_callback_sse41+0x4fd>
+ .byte 68,15,40,5,122,38,0,0 // movaps 0x267a(%rip),%xmm8 # 3e30 <_sk_callback_sse41+0x503>
.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
@@ -19189,7 +19190,7 @@ _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 68,15,40,13,37,38,0,0 // movaps 0x2625(%rip),%xmm9 # 3e30 <_sk_callback_sse41+0x50d>
+ .byte 68,15,40,13,43,38,0,0 // movaps 0x262b(%rip),%xmm9 # 3e40 <_sk_callback_sse41+0x513>
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
@@ -19280,7 +19281,7 @@ _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 68,15,40,13,173,36,0,0 // movaps 0x24ad(%rip),%xmm9 # 3e40 <_sk_callback_sse41+0x51d>
+ .byte 68,15,40,13,179,36,0,0 // movaps 0x24b3(%rip),%xmm9 # 3e50 <_sk_callback_sse41+0x523>
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
@@ -19457,31 +19458,31 @@ _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 68,15,89,5,4,34,0,0 // mulps 0x2204(%rip),%xmm8 # 3e50 <_sk_callback_sse41+0x52d>
- .byte 68,15,84,21,12,34,0,0 // andps 0x220c(%rip),%xmm10 # 3e60 <_sk_callback_sse41+0x53d>
- .byte 68,15,86,21,20,34,0,0 // orps 0x2214(%rip),%xmm10 # 3e70 <_sk_callback_sse41+0x54d>
- .byte 68,15,88,5,28,34,0,0 // addps 0x221c(%rip),%xmm8 # 3e80 <_sk_callback_sse41+0x55d>
- .byte 68,15,40,37,36,34,0,0 // movaps 0x2224(%rip),%xmm12 # 3e90 <_sk_callback_sse41+0x56d>
+ .byte 68,15,89,5,10,34,0,0 // mulps 0x220a(%rip),%xmm8 # 3e60 <_sk_callback_sse41+0x533>
+ .byte 68,15,84,21,18,34,0,0 // andps 0x2212(%rip),%xmm10 # 3e70 <_sk_callback_sse41+0x543>
+ .byte 68,15,86,21,26,34,0,0 // orps 0x221a(%rip),%xmm10 # 3e80 <_sk_callback_sse41+0x553>
+ .byte 68,15,88,5,34,34,0,0 // addps 0x2222(%rip),%xmm8 # 3e90 <_sk_callback_sse41+0x563>
+ .byte 68,15,40,37,42,34,0,0 // movaps 0x222a(%rip),%xmm12 # 3ea0 <_sk_callback_sse41+0x573>
.byte 69,15,89,226 // mulps %xmm10,%xmm12
.byte 69,15,92,196 // subps %xmm12,%xmm8
- .byte 68,15,88,21,36,34,0,0 // addps 0x2224(%rip),%xmm10 # 3ea0 <_sk_callback_sse41+0x57d>
- .byte 68,15,40,37,44,34,0,0 // movaps 0x222c(%rip),%xmm12 # 3eb0 <_sk_callback_sse41+0x58d>
+ .byte 68,15,88,21,42,34,0,0 // addps 0x222a(%rip),%xmm10 # 3eb0 <_sk_callback_sse41+0x583>
+ .byte 68,15,40,37,50,34,0,0 // movaps 0x2232(%rip),%xmm12 # 3ec0 <_sk_callback_sse41+0x593>
.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 68,15,88,5,25,34,0,0 // addps 0x2219(%rip),%xmm8 # 3ec0 <_sk_callback_sse41+0x59d>
- .byte 68,15,40,21,33,34,0,0 // movaps 0x2221(%rip),%xmm10 # 3ed0 <_sk_callback_sse41+0x5ad>
+ .byte 68,15,88,5,31,34,0,0 // addps 0x221f(%rip),%xmm8 # 3ed0 <_sk_callback_sse41+0x5a3>
+ .byte 68,15,40,21,39,34,0,0 // movaps 0x2227(%rip),%xmm10 # 3ee0 <_sk_callback_sse41+0x5b3>
.byte 69,15,89,211 // mulps %xmm11,%xmm10
.byte 69,15,92,194 // subps %xmm10,%xmm8
- .byte 68,15,40,21,33,34,0,0 // movaps 0x2221(%rip),%xmm10 # 3ee0 <_sk_callback_sse41+0x5bd>
+ .byte 68,15,40,21,39,34,0,0 // movaps 0x2227(%rip),%xmm10 # 3ef0 <_sk_callback_sse41+0x5c3>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 68,15,40,29,37,34,0,0 // movaps 0x2225(%rip),%xmm11 # 3ef0 <_sk_callback_sse41+0x5cd>
+ .byte 68,15,40,29,43,34,0,0 // movaps 0x222b(%rip),%xmm11 # 3f00 <_sk_callback_sse41+0x5d3>
.byte 69,15,94,218 // divps %xmm10,%xmm11
.byte 69,15,88,216 // addps %xmm8,%xmm11
- .byte 68,15,89,29,37,34,0,0 // mulps 0x2225(%rip),%xmm11 # 3f00 <_sk_callback_sse41+0x5dd>
+ .byte 68,15,89,29,43,34,0,0 // mulps 0x222b(%rip),%xmm11 # 3f10 <_sk_callback_sse41+0x5e3>
.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
@@ -19489,7 +19490,7 @@ _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 68,15,93,5,12,34,0,0 // minps 0x220c(%rip),%xmm8 # 3f10 <_sk_callback_sse41+0x5ed>
+ .byte 68,15,93,5,18,34,0,0 // minps 0x2212(%rip),%xmm8 # 3f20 <_sk_callback_sse41+0x5f3>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19519,31 +19520,31 @@ _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 68,15,89,37,173,33,0,0 // mulps 0x21ad(%rip),%xmm12 # 3f20 <_sk_callback_sse41+0x5fd>
- .byte 68,15,84,29,181,33,0,0 // andps 0x21b5(%rip),%xmm11 # 3f30 <_sk_callback_sse41+0x60d>
- .byte 68,15,86,29,189,33,0,0 // orps 0x21bd(%rip),%xmm11 # 3f40 <_sk_callback_sse41+0x61d>
- .byte 68,15,88,37,197,33,0,0 // addps 0x21c5(%rip),%xmm12 # 3f50 <_sk_callback_sse41+0x62d>
- .byte 15,40,13,206,33,0,0 // movaps 0x21ce(%rip),%xmm1 # 3f60 <_sk_callback_sse41+0x63d>
+ .byte 68,15,89,37,179,33,0,0 // mulps 0x21b3(%rip),%xmm12 # 3f30 <_sk_callback_sse41+0x603>
+ .byte 68,15,84,29,187,33,0,0 // andps 0x21bb(%rip),%xmm11 # 3f40 <_sk_callback_sse41+0x613>
+ .byte 68,15,86,29,195,33,0,0 // orps 0x21c3(%rip),%xmm11 # 3f50 <_sk_callback_sse41+0x623>
+ .byte 68,15,88,37,203,33,0,0 // addps 0x21cb(%rip),%xmm12 # 3f60 <_sk_callback_sse41+0x633>
+ .byte 15,40,13,212,33,0,0 // movaps 0x21d4(%rip),%xmm1 # 3f70 <_sk_callback_sse41+0x643>
.byte 65,15,89,203 // mulps %xmm11,%xmm1
.byte 68,15,92,225 // subps %xmm1,%xmm12
- .byte 68,15,88,29,206,33,0,0 // addps 0x21ce(%rip),%xmm11 # 3f70 <_sk_callback_sse41+0x64d>
- .byte 15,40,13,215,33,0,0 // movaps 0x21d7(%rip),%xmm1 # 3f80 <_sk_callback_sse41+0x65d>
+ .byte 68,15,88,29,212,33,0,0 // addps 0x21d4(%rip),%xmm11 # 3f80 <_sk_callback_sse41+0x653>
+ .byte 15,40,13,221,33,0,0 // movaps 0x21dd(%rip),%xmm1 # 3f90 <_sk_callback_sse41+0x663>
.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 68,15,88,37,196,33,0,0 // addps 0x21c4(%rip),%xmm12 # 3f90 <_sk_callback_sse41+0x66d>
- .byte 15,40,13,205,33,0,0 // movaps 0x21cd(%rip),%xmm1 # 3fa0 <_sk_callback_sse41+0x67d>
+ .byte 68,15,88,37,202,33,0,0 // addps 0x21ca(%rip),%xmm12 # 3fa0 <_sk_callback_sse41+0x673>
+ .byte 15,40,13,211,33,0,0 // movaps 0x21d3(%rip),%xmm1 # 3fb0 <_sk_callback_sse41+0x683>
.byte 65,15,89,203 // mulps %xmm11,%xmm1
.byte 68,15,92,225 // subps %xmm1,%xmm12
- .byte 68,15,40,21,205,33,0,0 // movaps 0x21cd(%rip),%xmm10 # 3fb0 <_sk_callback_sse41+0x68d>
+ .byte 68,15,40,21,211,33,0,0 // movaps 0x21d3(%rip),%xmm10 # 3fc0 <_sk_callback_sse41+0x693>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 15,40,13,210,33,0,0 // movaps 0x21d2(%rip),%xmm1 # 3fc0 <_sk_callback_sse41+0x69d>
+ .byte 15,40,13,216,33,0,0 // movaps 0x21d8(%rip),%xmm1 # 3fd0 <_sk_callback_sse41+0x6a3>
.byte 65,15,94,202 // divps %xmm10,%xmm1
.byte 65,15,88,204 // addps %xmm12,%xmm1
- .byte 15,89,13,211,33,0,0 // mulps 0x21d3(%rip),%xmm1 # 3fd0 <_sk_callback_sse41+0x6ad>
+ .byte 15,89,13,217,33,0,0 // mulps 0x21d9(%rip),%xmm1 # 3fe0 <_sk_callback_sse41+0x6b3>
.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
@@ -19551,7 +19552,7 @@ _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 15,93,13,190,33,0,0 // minps 0x21be(%rip),%xmm1 # 3fe0 <_sk_callback_sse41+0x6bd>
+ .byte 15,93,13,196,33,0,0 // minps 0x21c4(%rip),%xmm1 # 3ff0 <_sk_callback_sse41+0x6c3>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19581,31 +19582,31 @@ _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 68,15,89,37,95,33,0,0 // mulps 0x215f(%rip),%xmm12 # 3ff0 <_sk_callback_sse41+0x6cd>
- .byte 68,15,84,29,103,33,0,0 // andps 0x2167(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x6dd>
- .byte 68,15,86,29,111,33,0,0 // orps 0x216f(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x6ed>
- .byte 68,15,88,37,119,33,0,0 // addps 0x2177(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x6fd>
- .byte 15,40,21,128,33,0,0 // movaps 0x2180(%rip),%xmm2 # 4030 <_sk_callback_sse41+0x70d>
+ .byte 68,15,89,37,101,33,0,0 // mulps 0x2165(%rip),%xmm12 # 4000 <_sk_callback_sse41+0x6d3>
+ .byte 68,15,84,29,109,33,0,0 // andps 0x216d(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x6e3>
+ .byte 68,15,86,29,117,33,0,0 // orps 0x2175(%rip),%xmm11 # 4020 <_sk_callback_sse41+0x6f3>
+ .byte 68,15,88,37,125,33,0,0 // addps 0x217d(%rip),%xmm12 # 4030 <_sk_callback_sse41+0x703>
+ .byte 15,40,21,134,33,0,0 // movaps 0x2186(%rip),%xmm2 # 4040 <_sk_callback_sse41+0x713>
.byte 65,15,89,211 // mulps %xmm11,%xmm2
.byte 68,15,92,226 // subps %xmm2,%xmm12
- .byte 68,15,88,29,128,33,0,0 // addps 0x2180(%rip),%xmm11 # 4040 <_sk_callback_sse41+0x71d>
- .byte 15,40,21,137,33,0,0 // movaps 0x2189(%rip),%xmm2 # 4050 <_sk_callback_sse41+0x72d>
+ .byte 68,15,88,29,134,33,0,0 // addps 0x2186(%rip),%xmm11 # 4050 <_sk_callback_sse41+0x723>
+ .byte 15,40,21,143,33,0,0 // movaps 0x218f(%rip),%xmm2 # 4060 <_sk_callback_sse41+0x733>
.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 68,15,88,37,118,33,0,0 // addps 0x2176(%rip),%xmm12 # 4060 <_sk_callback_sse41+0x73d>
- .byte 15,40,21,127,33,0,0 // movaps 0x217f(%rip),%xmm2 # 4070 <_sk_callback_sse41+0x74d>
+ .byte 68,15,88,37,124,33,0,0 // addps 0x217c(%rip),%xmm12 # 4070 <_sk_callback_sse41+0x743>
+ .byte 15,40,21,133,33,0,0 // movaps 0x2185(%rip),%xmm2 # 4080 <_sk_callback_sse41+0x753>
.byte 65,15,89,211 // mulps %xmm11,%xmm2
.byte 68,15,92,226 // subps %xmm2,%xmm12
- .byte 68,15,40,21,127,33,0,0 // movaps 0x217f(%rip),%xmm10 # 4080 <_sk_callback_sse41+0x75d>
+ .byte 68,15,40,21,133,33,0,0 // movaps 0x2185(%rip),%xmm10 # 4090 <_sk_callback_sse41+0x763>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 15,40,21,132,33,0,0 // movaps 0x2184(%rip),%xmm2 # 4090 <_sk_callback_sse41+0x76d>
+ .byte 15,40,21,138,33,0,0 // movaps 0x218a(%rip),%xmm2 # 40a0 <_sk_callback_sse41+0x773>
.byte 65,15,94,210 // divps %xmm10,%xmm2
.byte 65,15,88,212 // addps %xmm12,%xmm2
- .byte 15,89,21,133,33,0,0 // mulps 0x2185(%rip),%xmm2 # 40a0 <_sk_callback_sse41+0x77d>
+ .byte 15,89,21,139,33,0,0 // mulps 0x218b(%rip),%xmm2 # 40b0 <_sk_callback_sse41+0x783>
.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
@@ -19613,7 +19614,7 @@ _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 15,93,21,112,33,0,0 // minps 0x2170(%rip),%xmm2 # 40b0 <_sk_callback_sse41+0x78d>
+ .byte 15,93,21,118,33,0,0 // minps 0x2176(%rip),%xmm2 # 40c0 <_sk_callback_sse41+0x793>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19643,31 +19644,31 @@ _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 68,15,89,37,17,33,0,0 // mulps 0x2111(%rip),%xmm12 # 40c0 <_sk_callback_sse41+0x79d>
- .byte 68,15,84,29,25,33,0,0 // andps 0x2119(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x7ad>
- .byte 68,15,86,29,33,33,0,0 // orps 0x2121(%rip),%xmm11 # 40e0 <_sk_callback_sse41+0x7bd>
- .byte 68,15,88,37,41,33,0,0 // addps 0x2129(%rip),%xmm12 # 40f0 <_sk_callback_sse41+0x7cd>
- .byte 15,40,29,50,33,0,0 // movaps 0x2132(%rip),%xmm3 # 4100 <_sk_callback_sse41+0x7dd>
+ .byte 68,15,89,37,23,33,0,0 // mulps 0x2117(%rip),%xmm12 # 40d0 <_sk_callback_sse41+0x7a3>
+ .byte 68,15,84,29,31,33,0,0 // andps 0x211f(%rip),%xmm11 # 40e0 <_sk_callback_sse41+0x7b3>
+ .byte 68,15,86,29,39,33,0,0 // orps 0x2127(%rip),%xmm11 # 40f0 <_sk_callback_sse41+0x7c3>
+ .byte 68,15,88,37,47,33,0,0 // addps 0x212f(%rip),%xmm12 # 4100 <_sk_callback_sse41+0x7d3>
+ .byte 15,40,29,56,33,0,0 // movaps 0x2138(%rip),%xmm3 # 4110 <_sk_callback_sse41+0x7e3>
.byte 65,15,89,219 // mulps %xmm11,%xmm3
.byte 68,15,92,227 // subps %xmm3,%xmm12
- .byte 68,15,88,29,50,33,0,0 // addps 0x2132(%rip),%xmm11 # 4110 <_sk_callback_sse41+0x7ed>
- .byte 15,40,29,59,33,0,0 // movaps 0x213b(%rip),%xmm3 # 4120 <_sk_callback_sse41+0x7fd>
+ .byte 68,15,88,29,56,33,0,0 // addps 0x2138(%rip),%xmm11 # 4120 <_sk_callback_sse41+0x7f3>
+ .byte 15,40,29,65,33,0,0 // movaps 0x2141(%rip),%xmm3 # 4130 <_sk_callback_sse41+0x803>
.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 68,15,88,37,40,33,0,0 // addps 0x2128(%rip),%xmm12 # 4130 <_sk_callback_sse41+0x80d>
- .byte 15,40,29,49,33,0,0 // movaps 0x2131(%rip),%xmm3 # 4140 <_sk_callback_sse41+0x81d>
+ .byte 68,15,88,37,46,33,0,0 // addps 0x212e(%rip),%xmm12 # 4140 <_sk_callback_sse41+0x813>
+ .byte 15,40,29,55,33,0,0 // movaps 0x2137(%rip),%xmm3 # 4150 <_sk_callback_sse41+0x823>
.byte 65,15,89,219 // mulps %xmm11,%xmm3
.byte 68,15,92,227 // subps %xmm3,%xmm12
- .byte 68,15,40,21,49,33,0,0 // movaps 0x2131(%rip),%xmm10 # 4150 <_sk_callback_sse41+0x82d>
+ .byte 68,15,40,21,55,33,0,0 // movaps 0x2137(%rip),%xmm10 # 4160 <_sk_callback_sse41+0x833>
.byte 69,15,92,211 // subps %xmm11,%xmm10
- .byte 15,40,29,54,33,0,0 // movaps 0x2136(%rip),%xmm3 # 4160 <_sk_callback_sse41+0x83d>
+ .byte 15,40,29,60,33,0,0 // movaps 0x213c(%rip),%xmm3 # 4170 <_sk_callback_sse41+0x843>
.byte 65,15,94,218 // divps %xmm10,%xmm3
.byte 65,15,88,220 // addps %xmm12,%xmm3
- .byte 15,89,29,55,33,0,0 // mulps 0x2137(%rip),%xmm3 # 4170 <_sk_callback_sse41+0x84d>
+ .byte 15,89,29,61,33,0,0 // mulps 0x213d(%rip),%xmm3 # 4180 <_sk_callback_sse41+0x853>
.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
@@ -19675,7 +19676,7 @@ _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 15,93,29,34,33,0,0 // minps 0x2122(%rip),%xmm3 # 4180 <_sk_callback_sse41+0x85d>
+ .byte 15,93,29,40,33,0,0 // minps 0x2128(%rip),%xmm3 # 4190 <_sk_callback_sse41+0x863>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 65,15,40,192 // movaps %xmm8,%xmm0
.byte 255,224 // jmpq *%rax
@@ -19685,29 +19686,29 @@ 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 68,15,89,5,30,33,0,0 // mulps 0x211e(%rip),%xmm8 # 4190 <_sk_callback_sse41+0x86d>
- .byte 68,15,40,13,38,33,0,0 // movaps 0x2126(%rip),%xmm9 # 41a0 <_sk_callback_sse41+0x87d>
+ .byte 68,15,89,5,36,33,0,0 // mulps 0x2124(%rip),%xmm8 # 41a0 <_sk_callback_sse41+0x873>
+ .byte 68,15,40,13,44,33,0,0 // movaps 0x212c(%rip),%xmm9 # 41b0 <_sk_callback_sse41+0x883>
.byte 65,15,89,201 // mulps %xmm9,%xmm1
- .byte 15,40,5,43,33,0,0 // movaps 0x212b(%rip),%xmm0 # 41b0 <_sk_callback_sse41+0x88d>
+ .byte 15,40,5,49,33,0,0 // movaps 0x2131(%rip),%xmm0 # 41c0 <_sk_callback_sse41+0x893>
.byte 15,88,200 // addps %xmm0,%xmm1
.byte 65,15,89,209 // mulps %xmm9,%xmm2
.byte 15,88,208 // addps %xmm0,%xmm2
- .byte 68,15,88,5,41,33,0,0 // addps 0x2129(%rip),%xmm8 # 41c0 <_sk_callback_sse41+0x89d>
- .byte 68,15,89,5,49,33,0,0 // mulps 0x2131(%rip),%xmm8 # 41d0 <_sk_callback_sse41+0x8ad>
- .byte 15,89,13,58,33,0,0 // mulps 0x213a(%rip),%xmm1 # 41e0 <_sk_callback_sse41+0x8bd>
+ .byte 68,15,88,5,47,33,0,0 // addps 0x212f(%rip),%xmm8 # 41d0 <_sk_callback_sse41+0x8a3>
+ .byte 68,15,89,5,55,33,0,0 // mulps 0x2137(%rip),%xmm8 # 41e0 <_sk_callback_sse41+0x8b3>
+ .byte 15,89,13,64,33,0,0 // mulps 0x2140(%rip),%xmm1 # 41f0 <_sk_callback_sse41+0x8c3>
.byte 65,15,88,200 // addps %xmm8,%xmm1
- .byte 15,89,21,63,33,0,0 // mulps 0x213f(%rip),%xmm2 # 41f0 <_sk_callback_sse41+0x8cd>
+ .byte 15,89,21,69,33,0,0 // mulps 0x2145(%rip),%xmm2 # 4200 <_sk_callback_sse41+0x8d3>
.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 68,15,40,13,51,33,0,0 // movaps 0x2133(%rip),%xmm9 # 4200 <_sk_callback_sse41+0x8dd>
+ .byte 68,15,40,13,57,33,0,0 // movaps 0x2139(%rip),%xmm9 # 4210 <_sk_callback_sse41+0x8e3>
.byte 65,15,40,193 // movaps %xmm9,%xmm0
.byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0
- .byte 15,40,21,51,33,0,0 // movaps 0x2133(%rip),%xmm2 # 4210 <_sk_callback_sse41+0x8ed>
+ .byte 15,40,21,57,33,0,0 // movaps 0x2139(%rip),%xmm2 # 4220 <_sk_callback_sse41+0x8f3>
.byte 15,88,202 // addps %xmm2,%xmm1
- .byte 68,15,40,37,56,33,0,0 // movaps 0x2138(%rip),%xmm12 # 4220 <_sk_callback_sse41+0x8fd>
+ .byte 68,15,40,37,62,33,0,0 // movaps 0x213e(%rip),%xmm12 # 4230 <_sk_callback_sse41+0x903>
.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
@@ -19726,8 +19727,8 @@ _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 15,89,13,241,32,0,0 // mulps 0x20f1(%rip),%xmm1 # 4230 <_sk_callback_sse41+0x90d>
- .byte 15,89,21,250,32,0,0 // mulps 0x20fa(%rip),%xmm2 # 4240 <_sk_callback_sse41+0x91d>
+ .byte 15,89,13,247,32,0,0 // mulps 0x20f7(%rip),%xmm1 # 4240 <_sk_callback_sse41+0x913>
+ .byte 15,89,21,0,33,0,0 // mulps 0x2100(%rip),%xmm2 # 4250 <_sk_callback_sse41+0x923>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,40,193 // movaps %xmm1,%xmm0
.byte 65,15,40,200 // movaps %xmm8,%xmm1
@@ -19741,7 +19742,7 @@ _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 15,89,29,234,32,0,0 // mulps 0x20ea(%rip),%xmm3 # 4250 <_sk_callback_sse41+0x92d>
+ .byte 15,89,29,240,32,0,0 // mulps 0x20f0(%rip),%xmm3 # 4260 <_sk_callback_sse41+0x933>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
.byte 15,87,201 // xorps %xmm1,%xmm1
@@ -19774,7 +19775,7 @@ _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 15,89,29,126,32,0,0 // mulps 0x207e(%rip),%xmm3 # 4260 <_sk_callback_sse41+0x93d>
+ .byte 15,89,29,132,32,0,0 // mulps 0x2084(%rip),%xmm3 # 4270 <_sk_callback_sse41+0x943>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
.byte 102,15,239,201 // pxor %xmm1,%xmm1
@@ -19787,7 +19788,7 @@ FUNCTION(_sk_store_a8_sse41)
_sk_store_a8_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,114,32,0,0 // movaps 0x2072(%rip),%xmm8 # 4270 <_sk_callback_sse41+0x94d>
+ .byte 68,15,40,5,120,32,0,0 // movaps 0x2078(%rip),%xmm8 # 4280 <_sk_callback_sse41+0x953>
.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
@@ -19804,9 +19805,9 @@ _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 15,89,5,79,32,0,0 // mulps 0x204f(%rip),%xmm0 # 4280 <_sk_callback_sse41+0x95d>
+ .byte 15,89,5,85,32,0,0 // mulps 0x2055(%rip),%xmm0 # 4290 <_sk_callback_sse41+0x963>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,86,32,0,0 // movaps 0x2056(%rip),%xmm3 # 4290 <_sk_callback_sse41+0x96d>
+ .byte 15,40,29,92,32,0,0 // movaps 0x205c(%rip),%xmm3 # 42a0 <_sk_callback_sse41+0x973>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 15,40,208 // movaps %xmm0,%xmm2
.byte 255,224 // jmpq *%rax
@@ -19837,9 +19838,9 @@ _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 15,89,5,239,31,0,0 // mulps 0x1fef(%rip),%xmm0 # 42a0 <_sk_callback_sse41+0x97d>
+ .byte 15,89,5,245,31,0,0 // mulps 0x1ff5(%rip),%xmm0 # 42b0 <_sk_callback_sse41+0x983>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,246,31,0,0 // movaps 0x1ff6(%rip),%xmm3 # 42b0 <_sk_callback_sse41+0x98d>
+ .byte 15,40,29,252,31,0,0 // movaps 0x1ffc(%rip),%xmm3 # 42c0 <_sk_callback_sse41+0x993>
.byte 15,40,200 // movaps %xmm0,%xmm1
.byte 15,40,208 // movaps %xmm0,%xmm2
.byte 255,224 // jmpq *%rax
@@ -19851,9 +19852,9 @@ _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 22d1 <_sk_gather_i8_sse41+0xf>
+ .byte 116,5 // je 22db <_sk_gather_i8_sse41+0xf>
.byte 76,137,192 // mov %r8,%rax
- .byte 235,2 // jmp 22d3 <_sk_gather_i8_sse41+0x11>
+ .byte 235,2 // jmp 22dd <_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
@@ -19884,17 +19885,17 @@ _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 102,15,111,5,77,31,0,0 // movdqa 0x1f4d(%rip),%xmm0 # 42c0 <_sk_callback_sse41+0x99d>
+ .byte 102,15,111,5,83,31,0,0 // movdqa 0x1f53(%rip),%xmm0 # 42d0 <_sk_callback_sse41+0x9a3>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,78,31,0,0 // movaps 0x1f4e(%rip),%xmm8 # 42d0 <_sk_callback_sse41+0x9ad>
+ .byte 68,15,40,5,84,31,0,0 // movaps 0x1f54(%rip),%xmm8 # 42e0 <_sk_callback_sse41+0x9b3>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 102,15,111,203 // movdqa %xmm3,%xmm1
- .byte 102,15,56,0,13,77,31,0,0 // pshufb 0x1f4d(%rip),%xmm1 # 42e0 <_sk_callback_sse41+0x9bd>
+ .byte 102,15,56,0,13,83,31,0,0 // pshufb 0x1f53(%rip),%xmm1 # 42f0 <_sk_callback_sse41+0x9c3>
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,111,211 // movdqa %xmm3,%xmm2
- .byte 102,15,56,0,21,73,31,0,0 // pshufb 0x1f49(%rip),%xmm2 # 42f0 <_sk_callback_sse41+0x9cd>
+ .byte 102,15,56,0,21,79,31,0,0 // pshufb 0x1f4f(%rip),%xmm2 # 4300 <_sk_callback_sse41+0x9d3>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 102,15,114,211,24 // psrld $0x18,%xmm3
@@ -19910,19 +19911,19 @@ _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 102,15,111,5,47,31,0,0 // movdqa 0x1f2f(%rip),%xmm0 # 4300 <_sk_callback_sse41+0x9dd>
+ .byte 102,15,111,5,53,31,0,0 // movdqa 0x1f35(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x9e3>
.byte 102,15,219,194 // pand %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,49,31,0,0 // mulps 0x1f31(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x9ed>
- .byte 102,15,111,13,57,31,0,0 // movdqa 0x1f39(%rip),%xmm1 # 4320 <_sk_callback_sse41+0x9fd>
+ .byte 15,89,5,55,31,0,0 // mulps 0x1f37(%rip),%xmm0 # 4320 <_sk_callback_sse41+0x9f3>
+ .byte 102,15,111,13,63,31,0,0 // movdqa 0x1f3f(%rip),%xmm1 # 4330 <_sk_callback_sse41+0xa03>
.byte 102,15,219,202 // pand %xmm2,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,59,31,0,0 // mulps 0x1f3b(%rip),%xmm1 # 4330 <_sk_callback_sse41+0xa0d>
- .byte 102,15,219,21,67,31,0,0 // pand 0x1f43(%rip),%xmm2 # 4340 <_sk_callback_sse41+0xa1d>
+ .byte 15,89,13,65,31,0,0 // mulps 0x1f41(%rip),%xmm1 # 4340 <_sk_callback_sse41+0xa13>
+ .byte 102,15,219,21,73,31,0,0 // pand 0x1f49(%rip),%xmm2 # 4350 <_sk_callback_sse41+0xa23>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,73,31,0,0 // mulps 0x1f49(%rip),%xmm2 # 4350 <_sk_callback_sse41+0xa2d>
+ .byte 15,89,21,79,31,0,0 // mulps 0x1f4f(%rip),%xmm2 # 4360 <_sk_callback_sse41+0xa33>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,80,31,0,0 // movaps 0x1f50(%rip),%xmm3 # 4360 <_sk_callback_sse41+0xa3d>
+ .byte 15,40,29,86,31,0,0 // movaps 0x1f56(%rip),%xmm3 # 4370 <_sk_callback_sse41+0xa43>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_gather_565_sse41
@@ -19950,19 +19951,19 @@ _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 102,15,111,5,245,30,0,0 // movdqa 0x1ef5(%rip),%xmm0 # 4370 <_sk_callback_sse41+0xa4d>
+ .byte 102,15,111,5,251,30,0,0 // movdqa 0x1efb(%rip),%xmm0 # 4380 <_sk_callback_sse41+0xa53>
.byte 102,15,219,194 // pand %xmm2,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,247,30,0,0 // mulps 0x1ef7(%rip),%xmm0 # 4380 <_sk_callback_sse41+0xa5d>
- .byte 102,15,111,13,255,30,0,0 // movdqa 0x1eff(%rip),%xmm1 # 4390 <_sk_callback_sse41+0xa6d>
+ .byte 15,89,5,253,30,0,0 // mulps 0x1efd(%rip),%xmm0 # 4390 <_sk_callback_sse41+0xa63>
+ .byte 102,15,111,13,5,31,0,0 // movdqa 0x1f05(%rip),%xmm1 # 43a0 <_sk_callback_sse41+0xa73>
.byte 102,15,219,202 // pand %xmm2,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,1,31,0,0 // mulps 0x1f01(%rip),%xmm1 # 43a0 <_sk_callback_sse41+0xa7d>
- .byte 102,15,219,21,9,31,0,0 // pand 0x1f09(%rip),%xmm2 # 43b0 <_sk_callback_sse41+0xa8d>
+ .byte 15,89,13,7,31,0,0 // mulps 0x1f07(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0xa83>
+ .byte 102,15,219,21,15,31,0,0 // pand 0x1f0f(%rip),%xmm2 # 43c0 <_sk_callback_sse41+0xa93>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,15,31,0,0 // mulps 0x1f0f(%rip),%xmm2 # 43c0 <_sk_callback_sse41+0xa9d>
+ .byte 15,89,21,21,31,0,0 // mulps 0x1f15(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0xaa3>
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,29,22,31,0,0 // movaps 0x1f16(%rip),%xmm3 # 43d0 <_sk_callback_sse41+0xaad>
+ .byte 15,40,29,28,31,0,0 // movaps 0x1f1c(%rip),%xmm3 # 43e0 <_sk_callback_sse41+0xab3>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_565_sse41
@@ -19971,12 +19972,12 @@ FUNCTION(_sk_store_565_sse41)
_sk_store_565_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,23,31,0,0 // movaps 0x1f17(%rip),%xmm8 # 43e0 <_sk_callback_sse41+0xabd>
+ .byte 68,15,40,5,29,31,0,0 // movaps 0x1f1d(%rip),%xmm8 # 43f0 <_sk_callback_sse41+0xac3>
.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 68,15,40,21,12,31,0,0 // movaps 0x1f0c(%rip),%xmm10 # 43f0 <_sk_callback_sse41+0xacd>
+ .byte 68,15,40,21,18,31,0,0 // movaps 0x1f12(%rip),%xmm10 # 4400 <_sk_callback_sse41+0xad3>
.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
@@ -19996,21 +19997,21 @@ _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 102,15,111,5,215,30,0,0 // movdqa 0x1ed7(%rip),%xmm0 # 4400 <_sk_callback_sse41+0xadd>
+ .byte 102,15,111,5,221,30,0,0 // movdqa 0x1edd(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xae3>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,217,30,0,0 // mulps 0x1ed9(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xaed>
- .byte 102,15,111,13,225,30,0,0 // movdqa 0x1ee1(%rip),%xmm1 # 4420 <_sk_callback_sse41+0xafd>
+ .byte 15,89,5,223,30,0,0 // mulps 0x1edf(%rip),%xmm0 # 4420 <_sk_callback_sse41+0xaf3>
+ .byte 102,15,111,13,231,30,0,0 // movdqa 0x1ee7(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xb03>
.byte 102,15,219,203 // pand %xmm3,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,227,30,0,0 // mulps 0x1ee3(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xb0d>
- .byte 102,15,111,21,235,30,0,0 // movdqa 0x1eeb(%rip),%xmm2 # 4440 <_sk_callback_sse41+0xb1d>
+ .byte 15,89,13,233,30,0,0 // mulps 0x1ee9(%rip),%xmm1 # 4440 <_sk_callback_sse41+0xb13>
+ .byte 102,15,111,21,241,30,0,0 // movdqa 0x1ef1(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xb23>
.byte 102,15,219,211 // pand %xmm3,%xmm2
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,237,30,0,0 // mulps 0x1eed(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xb2d>
- .byte 102,15,219,29,245,30,0,0 // pand 0x1ef5(%rip),%xmm3 # 4460 <_sk_callback_sse41+0xb3d>
+ .byte 15,89,21,243,30,0,0 // mulps 0x1ef3(%rip),%xmm2 # 4460 <_sk_callback_sse41+0xb33>
+ .byte 102,15,219,29,251,30,0,0 // pand 0x1efb(%rip),%xmm3 # 4470 <_sk_callback_sse41+0xb43>
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,251,30,0,0 // mulps 0x1efb(%rip),%xmm3 # 4470 <_sk_callback_sse41+0xb4d>
+ .byte 15,89,29,1,31,0,0 // mulps 0x1f01(%rip),%xmm3 # 4480 <_sk_callback_sse41+0xb53>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -20039,21 +20040,21 @@ _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 102,15,111,5,158,30,0,0 // movdqa 0x1e9e(%rip),%xmm0 # 4480 <_sk_callback_sse41+0xb5d>
+ .byte 102,15,111,5,164,30,0,0 // movdqa 0x1ea4(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xb63>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 15,89,5,160,30,0,0 // mulps 0x1ea0(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xb6d>
- .byte 102,15,111,13,168,30,0,0 // movdqa 0x1ea8(%rip),%xmm1 # 44a0 <_sk_callback_sse41+0xb7d>
+ .byte 15,89,5,166,30,0,0 // mulps 0x1ea6(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xb73>
+ .byte 102,15,111,13,174,30,0,0 // movdqa 0x1eae(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xb83>
.byte 102,15,219,203 // pand %xmm3,%xmm1
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
- .byte 15,89,13,170,30,0,0 // mulps 0x1eaa(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xb8d>
- .byte 102,15,111,21,178,30,0,0 // movdqa 0x1eb2(%rip),%xmm2 # 44c0 <_sk_callback_sse41+0xb9d>
+ .byte 15,89,13,176,30,0,0 // mulps 0x1eb0(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xb93>
+ .byte 102,15,111,21,184,30,0,0 // movdqa 0x1eb8(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xba3>
.byte 102,15,219,211 // pand %xmm3,%xmm2
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
- .byte 15,89,21,180,30,0,0 // mulps 0x1eb4(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xbad>
- .byte 102,15,219,29,188,30,0,0 // pand 0x1ebc(%rip),%xmm3 # 44e0 <_sk_callback_sse41+0xbbd>
+ .byte 15,89,21,186,30,0,0 // mulps 0x1eba(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xbb3>
+ .byte 102,15,219,29,194,30,0,0 // pand 0x1ec2(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xbc3>
.byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
- .byte 15,89,29,194,30,0,0 // mulps 0x1ec2(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xbcd>
+ .byte 15,89,29,200,30,0,0 // mulps 0x1ec8(%rip),%xmm3 # 4500 <_sk_callback_sse41+0xbd3>
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -20063,7 +20064,7 @@ FUNCTION(_sk_store_4444_sse41)
_sk_store_4444_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,193,30,0,0 // movaps 0x1ec1(%rip),%xmm8 # 4500 <_sk_callback_sse41+0xbdd>
+ .byte 68,15,40,5,199,30,0,0 // movaps 0x1ec7(%rip),%xmm8 # 4510 <_sk_callback_sse41+0xbe3>
.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
@@ -20093,17 +20094,17 @@ _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 15,40,5,96,30,0,0 // movaps 0x1e60(%rip),%xmm0 # 4510 <_sk_callback_sse41+0xbed>
+ .byte 15,40,5,102,30,0,0 // movaps 0x1e66(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xbf3>
.byte 15,84,195 // andps %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,98,30,0,0 // movaps 0x1e62(%rip),%xmm8 # 4520 <_sk_callback_sse41+0xbfd>
+ .byte 68,15,40,5,104,30,0,0 // movaps 0x1e68(%rip),%xmm8 # 4530 <_sk_callback_sse41+0xc03>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 15,40,203 // movaps %xmm3,%xmm1
- .byte 102,15,56,0,13,98,30,0,0 // pshufb 0x1e62(%rip),%xmm1 # 4530 <_sk_callback_sse41+0xc0d>
+ .byte 102,15,56,0,13,104,30,0,0 // pshufb 0x1e68(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xc13>
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 15,40,211 // movaps %xmm3,%xmm2
- .byte 102,15,56,0,21,95,30,0,0 // pshufb 0x1e5f(%rip),%xmm2 # 4540 <_sk_callback_sse41+0xc1d>
+ .byte 102,15,56,0,21,101,30,0,0 // pshufb 0x1e65(%rip),%xmm2 # 4550 <_sk_callback_sse41+0xc23>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 102,15,114,211,24 // psrld $0x18,%xmm3
@@ -20134,17 +20135,17 @@ _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 102,15,111,5,248,29,0,0 // movdqa 0x1df8(%rip),%xmm0 # 4550 <_sk_callback_sse41+0xc2d>
+ .byte 102,15,111,5,254,29,0,0 // movdqa 0x1dfe(%rip),%xmm0 # 4560 <_sk_callback_sse41+0xc33>
.byte 102,15,219,195 // pand %xmm3,%xmm0
.byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
- .byte 68,15,40,5,249,29,0,0 // movaps 0x1df9(%rip),%xmm8 # 4560 <_sk_callback_sse41+0xc3d>
+ .byte 68,15,40,5,255,29,0,0 // movaps 0x1dff(%rip),%xmm8 # 4570 <_sk_callback_sse41+0xc43>
.byte 65,15,89,192 // mulps %xmm8,%xmm0
.byte 102,15,111,203 // movdqa %xmm3,%xmm1
- .byte 102,15,56,0,13,248,29,0,0 // pshufb 0x1df8(%rip),%xmm1 # 4570 <_sk_callback_sse41+0xc4d>
+ .byte 102,15,56,0,13,254,29,0,0 // pshufb 0x1dfe(%rip),%xmm1 # 4580 <_sk_callback_sse41+0xc53>
.byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
.byte 65,15,89,200 // mulps %xmm8,%xmm1
.byte 102,15,111,211 // movdqa %xmm3,%xmm2
- .byte 102,15,56,0,21,244,29,0,0 // pshufb 0x1df4(%rip),%xmm2 # 4580 <_sk_callback_sse41+0xc5d>
+ .byte 102,15,56,0,21,250,29,0,0 // pshufb 0x1dfa(%rip),%xmm2 # 4590 <_sk_callback_sse41+0xc63>
.byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
.byte 65,15,89,208 // mulps %xmm8,%xmm2
.byte 102,15,114,211,24 // psrld $0x18,%xmm3
@@ -20159,7 +20160,7 @@ FUNCTION(_sk_store_8888_sse41)
_sk_store_8888_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 68,15,40,5,224,29,0,0 // movaps 0x1de0(%rip),%xmm8 # 4590 <_sk_callback_sse41+0xc6d>
+ .byte 68,15,40,5,230,29,0,0 // movaps 0x1de6(%rip),%xmm8 # 45a0 <_sk_callback_sse41+0xc73>
.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
@@ -20196,18 +20197,18 @@ _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 102,68,15,111,5,89,29,0,0 // movdqa 0x1d59(%rip),%xmm8 # 45a0 <_sk_callback_sse41+0xc7d>
+ .byte 102,68,15,111,5,95,29,0,0 // movdqa 0x1d5f(%rip),%xmm8 # 45b0 <_sk_callback_sse41+0xc83>
.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 102,15,111,29,84,29,0,0 // movdqa 0x1d54(%rip),%xmm3 # 45b0 <_sk_callback_sse41+0xc8d>
+ .byte 102,15,111,29,90,29,0,0 // movdqa 0x1d5a(%rip),%xmm3 # 45c0 <_sk_callback_sse41+0xc93>
.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 102,68,15,111,21,64,29,0,0 // movdqa 0x1d40(%rip),%xmm10 # 45c0 <_sk_callback_sse41+0xc9d>
+ .byte 102,68,15,111,21,70,29,0,0 // movdqa 0x1d46(%rip),%xmm10 # 45d0 <_sk_callback_sse41+0xca3>
.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
@@ -20280,18 +20281,18 @@ _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 102,68,15,111,5,254,27,0,0 // movdqa 0x1bfe(%rip),%xmm8 # 45d0 <_sk_callback_sse41+0xcad>
+ .byte 102,68,15,111,5,4,28,0,0 // movdqa 0x1c04(%rip),%xmm8 # 45e0 <_sk_callback_sse41+0xcb3>
.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 102,15,111,29,249,27,0,0 // movdqa 0x1bf9(%rip),%xmm3 # 45e0 <_sk_callback_sse41+0xcbd>
+ .byte 102,15,111,29,255,27,0,0 // movdqa 0x1bff(%rip),%xmm3 # 45f0 <_sk_callback_sse41+0xcc3>
.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 102,68,15,111,21,229,27,0,0 // movdqa 0x1be5(%rip),%xmm10 # 45f0 <_sk_callback_sse41+0xccd>
+ .byte 102,68,15,111,21,235,27,0,0 // movdqa 0x1beb(%rip),%xmm10 # 4600 <_sk_callback_sse41+0xcd3>
.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
@@ -20339,17 +20340,17 @@ FUNCTION(_sk_store_f16_sse41)
_sk_store_f16_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
- .byte 102,68,15,111,21,27,27,0,0 // movdqa 0x1b1b(%rip),%xmm10 # 4600 <_sk_callback_sse41+0xcdd>
+ .byte 102,68,15,111,21,33,27,0,0 // movdqa 0x1b21(%rip),%xmm10 # 4610 <_sk_callback_sse41+0xce3>
.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 102,68,15,111,13,14,27,0,0 // movdqa 0x1b0e(%rip),%xmm9 # 4610 <_sk_callback_sse41+0xced>
+ .byte 102,68,15,111,13,20,27,0,0 // movdqa 0x1b14(%rip),%xmm9 # 4620 <_sk_callback_sse41+0xcf3>
.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 102,68,15,111,29,255,26,0,0 // movdqa 0x1aff(%rip),%xmm11 # 4620 <_sk_callback_sse41+0xcfd>
+ .byte 102,68,15,111,29,5,27,0,0 // movdqa 0x1b05(%rip),%xmm11 # 4630 <_sk_callback_sse41+0xd03>
.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
@@ -20419,7 +20420,7 @@ _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 68,15,40,5,206,25,0,0 // movaps 0x19ce(%rip),%xmm8 # 4630 <_sk_callback_sse41+0xd0d>
+ .byte 68,15,40,5,212,25,0,0 // movaps 0x19d4(%rip),%xmm8 # 4640 <_sk_callback_sse41+0xd13>
.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
@@ -20471,7 +20472,7 @@ _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 68,15,40,5,15,25,0,0 // movaps 0x190f(%rip),%xmm8 # 4640 <_sk_callback_sse41+0xd1d>
+ .byte 68,15,40,5,21,25,0,0 // movaps 0x1915(%rip),%xmm8 # 4650 <_sk_callback_sse41+0xd23>
.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
@@ -20488,7 +20489,7 @@ _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 15,40,29,214,24,0,0 // movaps 0x18d6(%rip),%xmm3 # 4650 <_sk_callback_sse41+0xd2d>
+ .byte 15,40,29,220,24,0,0 // movaps 0x18dc(%rip),%xmm3 # 4660 <_sk_callback_sse41+0xd33>
.byte 255,224 // jmpq *%rax
HIDDEN _sk_store_u16_be_sse41
@@ -20497,7 +20498,7 @@ 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 68,15,40,13,215,24,0,0 // movaps 0x18d7(%rip),%xmm9 # 4660 <_sk_callback_sse41+0xd3d>
+ .byte 68,15,40,13,221,24,0,0 // movaps 0x18dd(%rip),%xmm9 # 4670 <_sk_callback_sse41+0xd43>
.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
@@ -20720,10 +20721,10 @@ 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 15,89,5,245,21,0,0 // mulps 0x15f5(%rip),%xmm0 # 4670 <_sk_callback_sse41+0xd4d>
- .byte 15,89,13,254,21,0,0 // mulps 0x15fe(%rip),%xmm1 # 4680 <_sk_callback_sse41+0xd5d>
+ .byte 15,89,5,251,21,0,0 // mulps 0x15fb(%rip),%xmm0 # 4680 <_sk_callback_sse41+0xd53>
+ .byte 15,89,13,4,22,0,0 // mulps 0x1604(%rip),%xmm1 # 4690 <_sk_callback_sse41+0xd63>
.byte 15,88,200 // addps %xmm0,%xmm1
- .byte 15,89,29,4,22,0,0 // mulps 0x1604(%rip),%xmm3 # 4690 <_sk_callback_sse41+0xd6d>
+ .byte 15,89,29,10,22,0,0 // mulps 0x160a(%rip),%xmm3 # 46a0 <_sk_callback_sse41+0xd73>
.byte 15,88,217 // addps %xmm1,%xmm3
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 15,87,192 // xorps %xmm0,%xmm0
@@ -20956,7 +20957,7 @@ _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 352e <_sk_linear_gradient_sse41+0x138>
+ .byte 15,132,254,0,0,0 // je 3538 <_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)
@@ -21006,12 +21007,12 @@ _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 3459 <_sk_linear_gradient_sse41+0x63>
+ .byte 15,133,65,255,255,255 // jne 3463 <_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 353b <_sk_linear_gradient_sse41+0x145>
+ .byte 235,13 // jmp 3545 <_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
@@ -21066,7 +21067,7 @@ HIDDEN _sk_save_xy_sse41
FUNCTION(_sk_save_xy_sse41)
_sk_save_xy_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,198,16,0,0 // movaps 0x10c6(%rip),%xmm8 # 46a0 <_sk_callback_sse41+0xd7d>
+ .byte 68,15,40,5,204,16,0,0 // movaps 0x10cc(%rip),%xmm8 # 46b0 <_sk_callback_sse41+0xd83>
.byte 15,17,0 // movups %xmm0,(%rax)
.byte 68,15,40,200 // movaps %xmm0,%xmm9
.byte 69,15,88,200 // addps %xmm8,%xmm9
@@ -21110,8 +21111,8 @@ _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 15,88,5,72,16,0,0 // addps 0x1048(%rip),%xmm0 # 46b0 <_sk_callback_sse41+0xd8d>
- .byte 68,15,40,13,80,16,0,0 // movaps 0x1050(%rip),%xmm9 # 46c0 <_sk_callback_sse41+0xd9d>
+ .byte 15,88,5,78,16,0,0 // addps 0x104e(%rip),%xmm0 # 46c0 <_sk_callback_sse41+0xd93>
+ .byte 68,15,40,13,86,16,0,0 // movaps 0x1056(%rip),%xmm9 # 46d0 <_sk_callback_sse41+0xda3>
.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
@@ -21124,7 +21125,7 @@ _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 15,88,5,63,16,0,0 // addps 0x103f(%rip),%xmm0 # 46d0 <_sk_callback_sse41+0xdad>
+ .byte 15,88,5,69,16,0,0 // addps 0x1045(%rip),%xmm0 # 46e0 <_sk_callback_sse41+0xdb3>
.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
@@ -21136,8 +21137,8 @@ _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 15,88,13,49,16,0,0 // addps 0x1031(%rip),%xmm1 # 46e0 <_sk_callback_sse41+0xdbd>
- .byte 68,15,40,13,57,16,0,0 // movaps 0x1039(%rip),%xmm9 # 46f0 <_sk_callback_sse41+0xdcd>
+ .byte 15,88,13,55,16,0,0 // addps 0x1037(%rip),%xmm1 # 46f0 <_sk_callback_sse41+0xdc3>
+ .byte 68,15,40,13,63,16,0,0 // movaps 0x103f(%rip),%xmm9 # 4700 <_sk_callback_sse41+0xdd3>
.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
@@ -21150,7 +21151,7 @@ _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 15,88,13,39,16,0,0 // addps 0x1027(%rip),%xmm1 # 4700 <_sk_callback_sse41+0xddd>
+ .byte 15,88,13,45,16,0,0 // addps 0x102d(%rip),%xmm1 # 4710 <_sk_callback_sse41+0xde3>
.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
@@ -21162,13 +21163,13 @@ _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 15,88,5,26,16,0,0 // addps 0x101a(%rip),%xmm0 # 4710 <_sk_callback_sse41+0xded>
- .byte 68,15,40,13,34,16,0,0 // movaps 0x1022(%rip),%xmm9 # 4720 <_sk_callback_sse41+0xdfd>
+ .byte 15,88,5,32,16,0,0 // addps 0x1020(%rip),%xmm0 # 4720 <_sk_callback_sse41+0xdf3>
+ .byte 68,15,40,13,40,16,0,0 // movaps 0x1028(%rip),%xmm9 # 4730 <_sk_callback_sse41+0xe03>
.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 68,15,89,13,30,16,0,0 // mulps 0x101e(%rip),%xmm9 # 4730 <_sk_callback_sse41+0xe0d>
- .byte 68,15,88,13,38,16,0,0 // addps 0x1026(%rip),%xmm9 # 4740 <_sk_callback_sse41+0xe1d>
+ .byte 68,15,89,13,36,16,0,0 // mulps 0x1024(%rip),%xmm9 # 4740 <_sk_callback_sse41+0xe13>
+ .byte 68,15,88,13,44,16,0,0 // addps 0x102c(%rip),%xmm9 # 4750 <_sk_callback_sse41+0xe23>
.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
@@ -21181,16 +21182,16 @@ _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 15,88,5,21,16,0,0 // addps 0x1015(%rip),%xmm0 # 4750 <_sk_callback_sse41+0xe2d>
- .byte 68,15,40,13,29,16,0,0 // movaps 0x101d(%rip),%xmm9 # 4760 <_sk_callback_sse41+0xe3d>
+ .byte 15,88,5,27,16,0,0 // addps 0x101b(%rip),%xmm0 # 4760 <_sk_callback_sse41+0xe33>
+ .byte 68,15,40,13,35,16,0,0 // movaps 0x1023(%rip),%xmm9 # 4770 <_sk_callback_sse41+0xe43>
.byte 69,15,92,200 // subps %xmm8,%xmm9
- .byte 68,15,40,5,33,16,0,0 // movaps 0x1021(%rip),%xmm8 # 4770 <_sk_callback_sse41+0xe4d>
+ .byte 68,15,40,5,39,16,0,0 // movaps 0x1027(%rip),%xmm8 # 4780 <_sk_callback_sse41+0xe53>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,37,16,0,0 // addps 0x1025(%rip),%xmm8 # 4780 <_sk_callback_sse41+0xe5d>
+ .byte 68,15,88,5,43,16,0,0 // addps 0x102b(%rip),%xmm8 # 4790 <_sk_callback_sse41+0xe63>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,41,16,0,0 // addps 0x1029(%rip),%xmm8 # 4790 <_sk_callback_sse41+0xe6d>
+ .byte 68,15,88,5,47,16,0,0 // addps 0x102f(%rip),%xmm8 # 47a0 <_sk_callback_sse41+0xe73>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,45,16,0,0 // addps 0x102d(%rip),%xmm8 # 47a0 <_sk_callback_sse41+0xe7d>
+ .byte 68,15,88,5,51,16,0,0 // addps 0x1033(%rip),%xmm8 # 47b0 <_sk_callback_sse41+0xe83>
.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
@@ -21200,17 +21201,17 @@ HIDDEN _sk_bicubic_p1x_sse41
FUNCTION(_sk_bicubic_p1x_sse41)
_sk_bicubic_p1x_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,39,16,0,0 // movaps 0x1027(%rip),%xmm8 # 47b0 <_sk_callback_sse41+0xe8d>
+ .byte 68,15,40,5,45,16,0,0 // movaps 0x102d(%rip),%xmm8 # 47c0 <_sk_callback_sse41+0xe93>
.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 68,15,40,21,35,16,0,0 // movaps 0x1023(%rip),%xmm10 # 47c0 <_sk_callback_sse41+0xe9d>
+ .byte 68,15,40,21,41,16,0,0 // movaps 0x1029(%rip),%xmm10 # 47d0 <_sk_callback_sse41+0xea3>
.byte 69,15,89,209 // mulps %xmm9,%xmm10
- .byte 68,15,88,21,39,16,0,0 // addps 0x1027(%rip),%xmm10 # 47d0 <_sk_callback_sse41+0xead>
+ .byte 68,15,88,21,45,16,0,0 // addps 0x102d(%rip),%xmm10 # 47e0 <_sk_callback_sse41+0xeb3>
.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 68,15,88,21,35,16,0,0 // addps 0x1023(%rip),%xmm10 # 47e0 <_sk_callback_sse41+0xebd>
+ .byte 68,15,88,21,41,16,0,0 // addps 0x1029(%rip),%xmm10 # 47f0 <_sk_callback_sse41+0xec3>
.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
@@ -21222,11 +21223,11 @@ _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 15,88,5,22,16,0,0 // addps 0x1016(%rip),%xmm0 # 47f0 <_sk_callback_sse41+0xecd>
+ .byte 15,88,5,28,16,0,0 // addps 0x101c(%rip),%xmm0 # 4800 <_sk_callback_sse41+0xed3>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,89,201 // mulps %xmm9,%xmm9
- .byte 68,15,89,5,22,16,0,0 // mulps 0x1016(%rip),%xmm8 # 4800 <_sk_callback_sse41+0xedd>
- .byte 68,15,88,5,30,16,0,0 // addps 0x101e(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xeed>
+ .byte 68,15,89,5,28,16,0,0 // mulps 0x101c(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xee3>
+ .byte 68,15,88,5,36,16,0,0 // addps 0x1024(%rip),%xmm8 # 4820 <_sk_callback_sse41+0xef3>
.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
@@ -21239,13 +21240,13 @@ _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 15,88,13,12,16,0,0 // addps 0x100c(%rip),%xmm1 # 4820 <_sk_callback_sse41+0xefd>
- .byte 68,15,40,13,20,16,0,0 // movaps 0x1014(%rip),%xmm9 # 4830 <_sk_callback_sse41+0xf0d>
+ .byte 15,88,13,18,16,0,0 // addps 0x1012(%rip),%xmm1 # 4830 <_sk_callback_sse41+0xf03>
+ .byte 68,15,40,13,26,16,0,0 // movaps 0x101a(%rip),%xmm9 # 4840 <_sk_callback_sse41+0xf13>
.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 68,15,89,13,16,16,0,0 // mulps 0x1010(%rip),%xmm9 # 4840 <_sk_callback_sse41+0xf1d>
- .byte 68,15,88,13,24,16,0,0 // addps 0x1018(%rip),%xmm9 # 4850 <_sk_callback_sse41+0xf2d>
+ .byte 68,15,89,13,22,16,0,0 // mulps 0x1016(%rip),%xmm9 # 4850 <_sk_callback_sse41+0xf23>
+ .byte 68,15,88,13,30,16,0,0 // addps 0x101e(%rip),%xmm9 # 4860 <_sk_callback_sse41+0xf33>
.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
@@ -21258,16 +21259,16 @@ _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 15,88,13,6,16,0,0 // addps 0x1006(%rip),%xmm1 # 4860 <_sk_callback_sse41+0xf3d>
- .byte 68,15,40,13,14,16,0,0 // movaps 0x100e(%rip),%xmm9 # 4870 <_sk_callback_sse41+0xf4d>
+ .byte 15,88,13,12,16,0,0 // addps 0x100c(%rip),%xmm1 # 4870 <_sk_callback_sse41+0xf43>
+ .byte 68,15,40,13,20,16,0,0 // movaps 0x1014(%rip),%xmm9 # 4880 <_sk_callback_sse41+0xf53>
.byte 69,15,92,200 // subps %xmm8,%xmm9
- .byte 68,15,40,5,18,16,0,0 // movaps 0x1012(%rip),%xmm8 # 4880 <_sk_callback_sse41+0xf5d>
+ .byte 68,15,40,5,24,16,0,0 // movaps 0x1018(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xf63>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,22,16,0,0 // addps 0x1016(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xf6d>
+ .byte 68,15,88,5,28,16,0,0 // addps 0x101c(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xf73>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,26,16,0,0 // addps 0x101a(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xf7d>
+ .byte 68,15,88,5,32,16,0,0 // addps 0x1020(%rip),%xmm8 # 48b0 <_sk_callback_sse41+0xf83>
.byte 69,15,89,193 // mulps %xmm9,%xmm8
- .byte 68,15,88,5,30,16,0,0 // addps 0x101e(%rip),%xmm8 # 48b0 <_sk_callback_sse41+0xf8d>
+ .byte 68,15,88,5,36,16,0,0 // addps 0x1024(%rip),%xmm8 # 48c0 <_sk_callback_sse41+0xf93>
.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
@@ -21277,17 +21278,17 @@ HIDDEN _sk_bicubic_p1y_sse41
FUNCTION(_sk_bicubic_p1y_sse41)
_sk_bicubic_p1y_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 68,15,40,5,24,16,0,0 // movaps 0x1018(%rip),%xmm8 # 48c0 <_sk_callback_sse41+0xf9d>
+ .byte 68,15,40,5,30,16,0,0 // movaps 0x101e(%rip),%xmm8 # 48d0 <_sk_callback_sse41+0xfa3>
.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 68,15,40,21,19,16,0,0 // movaps 0x1013(%rip),%xmm10 # 48d0 <_sk_callback_sse41+0xfad>
+ .byte 68,15,40,21,25,16,0,0 // movaps 0x1019(%rip),%xmm10 # 48e0 <_sk_callback_sse41+0xfb3>
.byte 69,15,89,209 // mulps %xmm9,%xmm10
- .byte 68,15,88,21,23,16,0,0 // addps 0x1017(%rip),%xmm10 # 48e0 <_sk_callback_sse41+0xfbd>
+ .byte 68,15,88,21,29,16,0,0 // addps 0x101d(%rip),%xmm10 # 48f0 <_sk_callback_sse41+0xfc3>
.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 68,15,88,21,19,16,0,0 // addps 0x1013(%rip),%xmm10 # 48f0 <_sk_callback_sse41+0xfcd>
+ .byte 68,15,88,21,25,16,0,0 // addps 0x1019(%rip),%xmm10 # 4900 <_sk_callback_sse41+0xfd3>
.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
@@ -21299,11 +21300,11 @@ _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 15,88,13,5,16,0,0 // addps 0x1005(%rip),%xmm1 # 4900 <_sk_callback_sse41+0xfdd>
+ .byte 15,88,13,11,16,0,0 // addps 0x100b(%rip),%xmm1 # 4910 <_sk_callback_sse41+0xfe3>
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,89,201 // mulps %xmm9,%xmm9
- .byte 68,15,89,5,5,16,0,0 // mulps 0x1005(%rip),%xmm8 # 4910 <_sk_callback_sse41+0xfed>
- .byte 68,15,88,5,13,16,0,0 // addps 0x100d(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xffd>
+ .byte 68,15,89,5,11,16,0,0 // mulps 0x100b(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xff3>
+ .byte 68,15,88,5,19,16,0,0 // addps 0x1013(%rip),%xmm8 # 4930 <_sk_callback_sse41+0x1003>
.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
@@ -21488,11 +21489,11 @@ BALIGN16
.byte 0,128,191,0,0,128 // add %al,-0x7fffff41(%rax)
.byte 191,0,0,224,64 // mov $0x40e00000,%edi
.byte 0,0 // add %al,(%rax)
- .byte 224,64 // loopne 3b88 <.literal16+0x188>
+ .byte 224,64 // loopne 3b98 <.literal16+0x188>
.byte 0,0 // add %al,(%rax)
- .byte 224,64 // loopne 3b8c <.literal16+0x18c>
+ .byte 224,64 // loopne 3b9c <.literal16+0x18c>
.byte 0,0 // add %al,(%rax)
- .byte 224,64 // loopne 3b90 <.literal16+0x190>
+ .byte 224,64 // loopne 3ba0 <.literal16+0x190>
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
@@ -21695,13 +21696,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 3d39 <.literal16+0x339>
+ .byte 224,7 // loopne 3d49 <.literal16+0x339>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 3d3d <.literal16+0x33d>
+ .byte 224,7 // loopne 3d4d <.literal16+0x33d>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 3d41 <.literal16+0x341>
+ .byte 224,7 // loopne 3d51 <.literal16+0x341>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 3d45 <.literal16+0x345>
+ .byte 224,7 // loopne 3d55 <.literal16+0x345>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -21741,10 +21742,10 @@ BALIGN16
.byte 0,1 // add %al,(%rcx)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a003d98 <_sk_callback_sse41+0xa000475>
+ .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a003da8 <_sk_callback_sse41+0xa00047b>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3003da0 <_sk_callback_sse41+0x300047d>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3003db0 <_sk_callback_sse41+0x3000483>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -21799,11 +21800,11 @@ 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 3e6b <.literal16+0x46b>
+ .byte 127,67 // jg 3e7b <.literal16+0x46b>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 3e6f <.literal16+0x46f>
+ .byte 127,67 // jg 3e7f <.literal16+0x46f>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 3e73 <.literal16+0x473>
+ .byte 127,67 // jg 3e83 <.literal16+0x473>
.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)
@@ -21818,16 +21819,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 3e64 <.literal16+0x464>
+ .byte 127,0 // jg 3e74 <.literal16+0x464>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3e68 <.literal16+0x468>
+ .byte 127,0 // jg 3e78 <.literal16+0x468>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3e6c <.literal16+0x46c>
+ .byte 127,0 // jg 3e7c <.literal16+0x46c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3e70 <.literal16+0x470>
+ .byte 127,0 // jg 3e80 <.literal16+0x470>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -21836,7 +21837,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 3ef5 <.literal16+0x4f5>
+ .byte 119,115 // ja 3f05 <.literal16+0x4f5>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -21847,7 +21848,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 3e59 <.literal16+0x459>
+ .byte 117,191 // jne 3e69 <.literal16+0x459>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -21859,7 +21860,7 @@ 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 ffffffffe9a37e9a <_sk_callback_sse41+0xffffffffe9a34577>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a37eaa <_sk_callback_sse41+0xffffffffe9a3457d>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -21914,16 +21915,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 3f34 <.literal16+0x534>
+ .byte 127,0 // jg 3f44 <.literal16+0x534>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3f38 <.literal16+0x538>
+ .byte 127,0 // jg 3f48 <.literal16+0x538>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3f3c <.literal16+0x53c>
+ .byte 127,0 // jg 3f4c <.literal16+0x53c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 3f40 <.literal16+0x540>
+ .byte 127,0 // jg 3f50 <.literal16+0x540>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -21932,7 +21933,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 3fc5 <.literal16+0x5c5>
+ .byte 119,115 // ja 3fd5 <.literal16+0x5c5>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -21943,7 +21944,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 3f29 <.literal16+0x529>
+ .byte 117,191 // jne 3f39 <.literal16+0x529>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -21955,7 +21956,7 @@ 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 ffffffffe9a37f6a <_sk_callback_sse41+0xffffffffe9a34647>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a37f7a <_sk_callback_sse41+0xffffffffe9a3464d>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22010,16 +22011,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 4004 <.literal16+0x604>
+ .byte 127,0 // jg 4014 <.literal16+0x604>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4008 <.literal16+0x608>
+ .byte 127,0 // jg 4018 <.literal16+0x608>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 400c <.literal16+0x60c>
+ .byte 127,0 // jg 401c <.literal16+0x60c>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 4010 <.literal16+0x610>
+ .byte 127,0 // jg 4020 <.literal16+0x610>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22028,7 +22029,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4095 <.literal16+0x695>
+ .byte 119,115 // ja 40a5 <.literal16+0x695>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -22039,7 +22040,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 3ff9 <.literal16+0x5f9>
+ .byte 117,191 // jne 4009 <.literal16+0x5f9>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -22051,7 +22052,7 @@ 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 ffffffffe9a3803a <_sk_callback_sse41+0xffffffffe9a34717>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a3804a <_sk_callback_sse41+0xffffffffe9a3471d>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22106,16 +22107,16 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 52,255 // xor $0xff,%al
.byte 255 // (bad)
- .byte 127,0 // jg 40d4 <.literal16+0x6d4>
+ .byte 127,0 // jg 40e4 <.literal16+0x6d4>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 40d8 <.literal16+0x6d8>
+ .byte 127,0 // jg 40e8 <.literal16+0x6d8>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 40dc <.literal16+0x6dc>
+ .byte 127,0 // jg 40ec <.literal16+0x6dc>
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 127,0 // jg 40e0 <.literal16+0x6e0>
+ .byte 127,0 // jg 40f0 <.literal16+0x6e0>
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22124,7 +22125,7 @@ BALIGN16
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
.byte 0,63 // add %bh,(%rdi)
- .byte 119,115 // ja 4165 <.literal16+0x765>
+ .byte 119,115 // ja 4175 <.literal16+0x765>
.byte 248 // clc
.byte 194,119,115 // retq $0x7377
.byte 248 // clc
@@ -22135,7 +22136,7 @@ BALIGN16
.byte 194,117,191 // retq $0xbf75
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
- .byte 117,191 // jne 40c9 <.literal16+0x6c9>
+ .byte 117,191 // jne 40d9 <.literal16+0x6c9>
.byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
.byte 63 // (bad)
.byte 249 // stc
@@ -22147,7 +22148,7 @@ 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 ffffffffe9a3810a <_sk_callback_sse41+0xffffffffe9a347e7>
+ .byte 233,220,63,163,233 // jmpq ffffffffe9a3811a <_sk_callback_sse41+0xffffffffe9a347ed>
.byte 220,63 // fdivrl (%rdi)
.byte 81 // push %rcx
.byte 140,242 // mov %?,%edx
@@ -22198,13 +22199,13 @@ 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 41e7 <.literal16+0x7e7>
+ .byte 127,67 // jg 41f7 <.literal16+0x7e7>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 41eb <.literal16+0x7eb>
+ .byte 127,67 // jg 41fb <.literal16+0x7eb>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 41ef <.literal16+0x7ef>
+ .byte 127,67 // jg 41ff <.literal16+0x7ef>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 41f3 <.literal16+0x7f3>
+ .byte 127,67 // jg 4203 <.literal16+0x7f3>
.byte 0,0 // add %al,(%rax)
.byte 0,195 // add %al,%bl
.byte 0,0 // add %al,(%rax)
@@ -22251,16 +22252,16 @@ BALIGN16
.byte 128,3,62 // addb $0x3e,(%rbx)
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 4273 <.literal16+0x873>
+ .byte 118,63 // jbe 4283 <.literal16+0x873>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 4277 <.literal16+0x877>
+ .byte 118,63 // jbe 4287 <.literal16+0x877>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 427b <.literal16+0x87b>
+ .byte 118,63 // jbe 428b <.literal16+0x87b>
.byte 31 // (bad)
.byte 215 // xlat %ds:(%rbx)
- .byte 118,63 // jbe 427f <.literal16+0x87f>
+ .byte 118,63 // jbe 428f <.literal16+0x87f>
.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)
@@ -22272,11 +22273,11 @@ 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 42bb <.literal16+0x8bb>
+ .byte 127,67 // jg 42cb <.literal16+0x8bb>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 42bf <.literal16+0x8bf>
+ .byte 127,67 // jg 42cf <.literal16+0x8bf>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 42c3 <.literal16+0x8c3>
+ .byte 127,67 // jg 42d3 <.literal16+0x8c3>
.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)
@@ -22305,7 +22306,7 @@ 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) # 30042f0 <_sk_callback_sse41+0x30009cd>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004300 <_sk_callback_sse41+0x30009d3>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -22334,13 +22335,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4329 <.literal16+0x929>
+ .byte 224,7 // loopne 4339 <.literal16+0x929>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 432d <.literal16+0x92d>
+ .byte 224,7 // loopne 433d <.literal16+0x92d>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 4331 <.literal16+0x931>
+ .byte 224,7 // loopne 4341 <.literal16+0x931>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 4335 <.literal16+0x935>
+ .byte 224,7 // loopne 4345 <.literal16+0x935>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -22386,13 +22387,13 @@ BALIGN16
.byte 132,55 // test %dh,(%rdi)
.byte 8,33 // or %ah,(%rcx)
.byte 132,55 // test %dh,(%rdi)
- .byte 224,7 // loopne 4399 <.literal16+0x999>
+ .byte 224,7 // loopne 43a9 <.literal16+0x999>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 439d <.literal16+0x99d>
+ .byte 224,7 // loopne 43ad <.literal16+0x99d>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 43a1 <.literal16+0x9a1>
+ .byte 224,7 // loopne 43b1 <.literal16+0x9a1>
.byte 0,0 // add %al,(%rax)
- .byte 224,7 // loopne 43a5 <.literal16+0x9a5>
+ .byte 224,7 // loopne 43b5 <.literal16+0x9a5>
.byte 0,0 // add %al,(%rax)
.byte 33,8 // and %ecx,(%rax)
.byte 2,58 // add (%rdx),%bh
@@ -22430,13 +22431,13 @@ BALIGN16
.byte 65,0,0 // add %al,(%r8)
.byte 248 // clc
.byte 65,0,0 // add %al,(%r8)
- .byte 124,66 // jl 4436 <.literal16+0xa36>
+ .byte 124,66 // jl 4446 <.literal16+0xa36>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 443a <.literal16+0xa3a>
+ .byte 124,66 // jl 444a <.literal16+0xa3a>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 443e <.literal16+0xa3e>
+ .byte 124,66 // jl 444e <.literal16+0xa3e>
.byte 0,0 // add %al,(%rax)
- .byte 124,66 // jl 4442 <.literal16+0xa42>
+ .byte 124,66 // jl 4452 <.literal16+0xa42>
.byte 0,240 // add %dh,%al
.byte 0,0 // add %al,(%rax)
.byte 0,240 // add %dh,%al
@@ -22526,13 +22527,13 @@ 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 4545 <.literal16+0xb45>
+ .byte 112,65 // jo 4555 <.literal16+0xb45>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4549 <.literal16+0xb49>
+ .byte 112,65 // jo 4559 <.literal16+0xb49>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 454d <.literal16+0xb4d>
+ .byte 112,65 // jo 455d <.literal16+0xb4d>
.byte 0,0 // add %al,(%rax)
- .byte 112,65 // jo 4551 <.literal16+0xb51>
+ .byte 112,65 // jo 4561 <.literal16+0xb51>
.byte 255,0 // incl (%rax)
.byte 0,0 // add %al,(%rax)
.byte 255,0 // incl (%rax)
@@ -22547,7 +22548,7 @@ 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) # 3004540 <_sk_callback_sse41+0x3000c1d>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004550 <_sk_callback_sse41+0x3000c23>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -22574,7 +22575,7 @@ 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) # 3004580 <_sk_callback_sse41+0x3000c5d>
+ .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3004590 <_sk_callback_sse41+0x3000c63>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255,6 // incl (%rsi)
@@ -22589,11 +22590,11 @@ BALIGN16
.byte 255,0 // incl (%rax)
.byte 0,127,67 // add %bh,0x43(%rdi)
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 45db <.literal16+0xbdb>
+ .byte 127,67 // jg 45eb <.literal16+0xbdb>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 45df <.literal16+0xbdf>
+ .byte 127,67 // jg 45ef <.literal16+0xbdf>
.byte 0,0 // add %al,(%rax)
- .byte 127,67 // jg 45e3 <.literal16+0xbe3>
+ .byte 127,67 // jg 45f3 <.literal16+0xbe3>
.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)
@@ -22669,13 +22670,13 @@ BALIGN16
.byte 0,0 // add %al,(%rax)
.byte 128,63,0 // cmpb $0x0,(%rdi)
.byte 255 // (bad)
- .byte 127,71 // jg 46ab <.literal16+0xcab>
+ .byte 127,71 // jg 46bb <.literal16+0xcab>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 46af <.literal16+0xcaf>
+ .byte 127,71 // jg 46bf <.literal16+0xcaf>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 46b3 <.literal16+0xcb3>
+ .byte 127,71 // jg 46c3 <.literal16+0xcb3>
.byte 0,255 // add %bh,%bh
- .byte 127,71 // jg 46b7 <.literal16+0xcb7>
+ .byte 127,71 // jg 46c7 <.literal16+0xcb7>
.byte 208 // (bad)
.byte 179,89 // mov $0x59,%bl
.byte 62,208 // ds (bad)
@@ -22759,11 +22760,11 @@ 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 4752 <.literal16+0xd52>
+ .byte 62,114,28 // jb,pt 4762 <.literal16+0xd52>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4756 <.literal16+0xd56>
+ .byte 62,114,28 // jb,pt 4766 <.literal16+0xd56>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 475a <.literal16+0xd5a>
+ .byte 62,114,28 // jb,pt 476a <.literal16+0xd5a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -22807,7 +22808,7 @@ 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 # 3d63d5e5 <_sk_callback_sse41+0x3d639cc2>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d5f5 <_sk_callback_sse41+0x3d639cc8>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22833,7 +22834,7 @@ 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 # 3d63d625 <_sk_callback_sse41+0x3d639d02>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d635 <_sk_callback_sse41+0x3d639d08>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
@@ -22842,13 +22843,13 @@ BALIGN16
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
.byte 63 // (bad)
- .byte 114,28 // jb 481e <.literal16+0xe1e>
+ .byte 114,28 // jb 482e <.literal16+0xe1e>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4822 <.literal16+0xe22>
+ .byte 62,114,28 // jb,pt 4832 <.literal16+0xe22>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4826 <.literal16+0xe26>
+ .byte 62,114,28 // jb,pt 4836 <.literal16+0xe26>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 482a <.literal16+0xe2a>
+ .byte 62,114,28 // jb,pt 483a <.literal16+0xe2a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -22869,11 +22870,11 @@ 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 4862 <.literal16+0xe62>
+ .byte 62,114,28 // jb,pt 4872 <.literal16+0xe62>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4866 <.literal16+0xe66>
+ .byte 62,114,28 // jb,pt 4876 <.literal16+0xe66>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 486a <.literal16+0xe6a>
+ .byte 62,114,28 // jb,pt 487a <.literal16+0xe6a>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -22917,7 +22918,7 @@ 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 # 3d63d6f5 <_sk_callback_sse41+0x3d639dd2>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d705 <_sk_callback_sse41+0x3d639dd8>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 0,63 // add %bh,(%rdi)
.byte 0,0 // add %al,(%rax)
@@ -22943,7 +22944,7 @@ 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 # 3d63d735 <_sk_callback_sse41+0x3d639e12>
+ .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63d745 <_sk_callback_sse41+0x3d639e18>
.byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi)
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
@@ -22952,13 +22953,13 @@ BALIGN16
.byte 192,63,0 // sarb $0x0,(%rdi)
.byte 0,192 // add %al,%al
.byte 63 // (bad)
- .byte 114,28 // jb 492e <.literal16+0xf2e>
+ .byte 114,28 // jb 493e <.literal16+0xf2e>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4932 <_sk_callback_sse41+0x100f>
+ .byte 62,114,28 // jb,pt 4942 <_sk_callback_sse41+0x1015>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 4936 <_sk_callback_sse41+0x1013>
+ .byte 62,114,28 // jb,pt 4946 <_sk_callback_sse41+0x1019>
.byte 199 // (bad)
- .byte 62,114,28 // jb,pt 493a <_sk_callback_sse41+0x1017>
+ .byte 62,114,28 // jb,pt 494a <_sk_callback_sse41+0x101d>
.byte 199 // (bad)
.byte 62,171 // ds stos %eax,%es:(%rdi)
.byte 170 // stos %al,%es:(%rdi)
@@ -24281,173 +24282,172 @@ _sk_hsl_to_rgb_sse2:
.byte 15,41,108,36,240 // movaps %xmm5,-0x10(%rsp)
.byte 15,41,100,36,224 // movaps %xmm4,-0x20(%rsp)
.byte 15,41,92,36,208 // movaps %xmm3,-0x30(%rsp)
- .byte 68,15,40,210 // movaps %xmm2,%xmm10
- .byte 15,40,217 // movaps %xmm1,%xmm3
- .byte 15,40,232 // movaps %xmm0,%xmm5
+ .byte 68,15,40,226 // movaps %xmm2,%xmm12
+ .byte 15,40,240 // movaps %xmm0,%xmm6
.byte 184,0,0,0,63 // mov $0x3f000000,%eax
.byte 102,15,110,192 // movd %eax,%xmm0
.byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
- .byte 69,15,40,194 // movaps %xmm10,%xmm8
+ .byte 69,15,40,196 // movaps %xmm12,%xmm8
.byte 68,15,194,192,1 // cmpltps %xmm0,%xmm8
- .byte 68,15,40,216 // movaps %xmm0,%xmm11
- .byte 68,15,40,37,255,47,0,0 // movaps 0x2fff(%rip),%xmm12 # 4090 <_sk_callback_sse2+0x369>
- .byte 15,40,195 // movaps %xmm3,%xmm0
- .byte 15,40,211 // movaps %xmm3,%xmm2
- .byte 15,87,201 // xorps %xmm1,%xmm1
- .byte 15,194,203,0 // cmpeqps %xmm3,%xmm1
- .byte 15,41,76,36,176 // movaps %xmm1,-0x50(%rsp)
- .byte 65,15,88,220 // addps %xmm12,%xmm3
- .byte 65,15,89,218 // mulps %xmm10,%xmm3
- .byte 65,15,88,194 // addps %xmm10,%xmm0
- .byte 65,15,89,210 // mulps %xmm10,%xmm2
- .byte 15,92,194 // subps %xmm2,%xmm0
- .byte 65,15,84,216 // andps %xmm8,%xmm3
+ .byte 68,15,40,208 // movaps %xmm0,%xmm10
+ .byte 68,15,41,84,36,176 // movaps %xmm10,-0x50(%rsp)
+ .byte 15,40,61,253,47,0,0 // movaps 0x2ffd(%rip),%xmm7 # 4090 <_sk_callback_sse2+0x369>
+ .byte 15,40,193 // movaps %xmm1,%xmm0
+ .byte 15,40,225 // movaps %xmm1,%xmm4
+ .byte 15,87,210 // xorps %xmm2,%xmm2
+ .byte 15,194,209,0 // cmpeqps %xmm1,%xmm2
+ .byte 15,41,84,36,160 // movaps %xmm2,-0x60(%rsp)
+ .byte 15,88,207 // addps %xmm7,%xmm1
+ .byte 65,15,89,204 // mulps %xmm12,%xmm1
+ .byte 65,15,88,196 // addps %xmm12,%xmm0
+ .byte 65,15,89,228 // mulps %xmm12,%xmm4
+ .byte 15,92,196 // subps %xmm4,%xmm0
+ .byte 65,15,84,200 // andps %xmm8,%xmm1
.byte 68,15,85,192 // andnps %xmm0,%xmm8
- .byte 68,15,86,195 // orps %xmm3,%xmm8
- .byte 15,40,29,215,47,0,0 // movaps 0x2fd7(%rip),%xmm3 # 40a0 <_sk_callback_sse2+0x379>
- .byte 15,88,221 // addps %xmm5,%xmm3
+ .byte 68,15,86,193 // orps %xmm1,%xmm8
+ .byte 15,40,13,214,47,0,0 // movaps 0x2fd6(%rip),%xmm1 # 40a0 <_sk_callback_sse2+0x379>
+ .byte 15,88,206 // addps %xmm6,%xmm1
.byte 184,0,0,0,0 // mov $0x0,%eax
.byte 185,0,0,128,63 // mov $0x3f800000,%ecx
.byte 102,68,15,110,241 // movd %ecx,%xmm14
.byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
- .byte 65,15,40,214 // movaps %xmm14,%xmm2
- .byte 15,194,211,1 // cmpltps %xmm3,%xmm2
- .byte 68,15,40,61,192,47,0,0 // movaps 0x2fc0(%rip),%xmm15 # 40b0 <_sk_callback_sse2+0x389>
- .byte 15,40,195 // movaps %xmm3,%xmm0
- .byte 65,15,88,199 // addps %xmm15,%xmm0
- .byte 15,84,194 // andps %xmm2,%xmm0
- .byte 15,85,211 // andnps %xmm3,%xmm2
- .byte 15,86,208 // orps %xmm0,%xmm2
- .byte 102,15,110,200 // movd %eax,%xmm1
- .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
- .byte 15,41,76,36,128 // movaps %xmm1,-0x80(%rsp)
- .byte 15,40,195 // movaps %xmm3,%xmm0
+ .byte 65,15,40,198 // movaps %xmm14,%xmm0
.byte 15,194,193,1 // cmpltps %xmm1,%xmm0
- .byte 15,40,227 // movaps %xmm3,%xmm4
- .byte 65,15,88,228 // addps %xmm12,%xmm4
+ .byte 68,15,40,61,191,47,0,0 // movaps 0x2fbf(%rip),%xmm15 # 40b0 <_sk_callback_sse2+0x389>
+ .byte 15,40,225 // movaps %xmm1,%xmm4
+ .byte 65,15,88,231 // addps %xmm15,%xmm4
.byte 15,84,224 // andps %xmm0,%xmm4
- .byte 15,85,194 // andnps %xmm2,%xmm0
+ .byte 15,85,193 // andnps %xmm1,%xmm0
.byte 15,86,196 // orps %xmm4,%xmm0
- .byte 69,15,40,234 // movaps %xmm10,%xmm13
+ .byte 102,15,110,208 // movd %eax,%xmm2
+ .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
+ .byte 15,41,84,36,144 // movaps %xmm2,-0x70(%rsp)
+ .byte 15,40,225 // movaps %xmm1,%xmm4
+ .byte 15,194,202,1 // cmpltps %xmm2,%xmm1
+ .byte 15,88,231 // addps %xmm7,%xmm4
+ .byte 15,84,225 // andps %xmm1,%xmm4
+ .byte 15,85,200 // andnps %xmm0,%xmm1
+ .byte 15,86,204 // orps %xmm4,%xmm1
+ .byte 69,15,40,236 // movaps %xmm12,%xmm13
.byte 69,15,88,237 // addps %xmm13,%xmm13
.byte 69,15,92,232 // subps %xmm8,%xmm13
.byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax
.byte 69,15,40,200 // movaps %xmm8,%xmm9
.byte 69,15,92,205 // subps %xmm13,%xmm9
- .byte 68,15,89,13,123,47,0,0 // mulps 0x2f7b(%rip),%xmm9 # 40c0 <_sk_callback_sse2+0x399>
+ .byte 68,15,89,13,126,47,0,0 // mulps 0x2f7e(%rip),%xmm9 # 40c0 <_sk_callback_sse2+0x399>
.byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx
- .byte 102,15,110,249 // movd %ecx,%xmm7
- .byte 15,198,255,0 // shufps $0x0,%xmm7,%xmm7
- .byte 15,41,124,36,144 // movaps %xmm7,-0x70(%rsp)
- .byte 15,40,53,114,47,0,0 // movaps 0x2f72(%rip),%xmm6 # 40d0 <_sk_callback_sse2+0x3a9>
- .byte 15,40,230 // movaps %xmm6,%xmm4
- .byte 15,92,224 // subps %xmm0,%xmm4
- .byte 15,40,208 // movaps %xmm0,%xmm2
- .byte 15,40,200 // movaps %xmm0,%xmm1
- .byte 15,194,199,1 // cmpltps %xmm7,%xmm0
+ .byte 102,15,110,217 // movd %ecx,%xmm3
+ .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
+ .byte 15,41,92,36,128 // movaps %xmm3,-0x80(%rsp)
+ .byte 15,40,45,117,47,0,0 // movaps 0x2f75(%rip),%xmm5 # 40d0 <_sk_callback_sse2+0x3a9>
+ .byte 15,40,229 // movaps %xmm5,%xmm4
+ .byte 15,92,225 // subps %xmm1,%xmm4
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 68,15,40,217 // movaps %xmm1,%xmm11
+ .byte 15,40,193 // movaps %xmm1,%xmm0
+ .byte 15,194,203,1 // cmpltps %xmm3,%xmm1
.byte 65,15,89,225 // mulps %xmm9,%xmm4
.byte 65,15,88,229 // addps %xmm13,%xmm4
- .byte 15,84,224 // andps %xmm0,%xmm4
- .byte 65,15,85,197 // andnps %xmm13,%xmm0
- .byte 15,86,196 // orps %xmm4,%xmm0
- .byte 65,15,40,251 // movaps %xmm11,%xmm7
- .byte 15,41,124,36,160 // movaps %xmm7,-0x60(%rsp)
- .byte 15,194,207,1 // cmpltps %xmm7,%xmm1
- .byte 65,15,40,224 // movaps %xmm8,%xmm4
.byte 15,84,225 // andps %xmm1,%xmm4
- .byte 15,85,200 // andnps %xmm0,%xmm1
+ .byte 65,15,85,205 // andnps %xmm13,%xmm1
.byte 15,86,204 // orps %xmm4,%xmm1
- .byte 102,15,110,224 // movd %eax,%xmm4
- .byte 15,198,228,0 // shufps $0x0,%xmm4,%xmm4
- .byte 15,194,212,1 // cmpltps %xmm4,%xmm2
- .byte 65,15,89,217 // mulps %xmm9,%xmm3
- .byte 65,15,88,221 // addps %xmm13,%xmm3
- .byte 15,84,218 // andps %xmm2,%xmm3
- .byte 15,85,209 // andnps %xmm1,%xmm2
- .byte 15,86,211 // orps %xmm3,%xmm2
- .byte 68,15,40,92,36,176 // movaps -0x50(%rsp),%xmm11
- .byte 65,15,40,195 // movaps %xmm11,%xmm0
+ .byte 65,15,194,194,1 // cmpltps %xmm10,%xmm0
+ .byte 65,15,40,224 // movaps %xmm8,%xmm4
+ .byte 15,84,224 // andps %xmm0,%xmm4
+ .byte 15,85,193 // andnps %xmm1,%xmm0
+ .byte 15,86,196 // orps %xmm4,%xmm0
+ .byte 102,68,15,110,208 // movd %eax,%xmm10
+ .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
+ .byte 65,15,194,210,1 // cmpltps %xmm10,%xmm2
+ .byte 69,15,89,217 // mulps %xmm9,%xmm11
+ .byte 69,15,88,221 // addps %xmm13,%xmm11
+ .byte 68,15,84,218 // andps %xmm2,%xmm11
+ .byte 15,85,208 // andnps %xmm0,%xmm2
+ .byte 65,15,86,211 // orps %xmm11,%xmm2
+ .byte 15,40,68,36,160 // movaps -0x60(%rsp),%xmm0
.byte 15,85,194 // andnps %xmm2,%xmm0
.byte 15,41,68,36,192 // movaps %xmm0,-0x40(%rsp)
.byte 65,15,40,198 // movaps %xmm14,%xmm0
- .byte 15,194,197,1 // cmpltps %xmm5,%xmm0
- .byte 15,40,205 // movaps %xmm5,%xmm1
+ .byte 15,194,198,1 // cmpltps %xmm6,%xmm0
+ .byte 15,40,206 // movaps %xmm6,%xmm1
.byte 65,15,88,207 // addps %xmm15,%xmm1
.byte 15,84,200 // andps %xmm0,%xmm1
- .byte 15,85,197 // andnps %xmm5,%xmm0
+ .byte 15,85,198 // andnps %xmm6,%xmm0
.byte 15,86,193 // orps %xmm1,%xmm0
- .byte 15,40,205 // movaps %xmm5,%xmm1
- .byte 15,194,76,36,128,1 // cmpltps -0x80(%rsp),%xmm1
- .byte 15,40,213 // movaps %xmm5,%xmm2
- .byte 65,15,88,212 // addps %xmm12,%xmm2
+ .byte 15,40,206 // movaps %xmm6,%xmm1
+ .byte 15,194,76,36,144,1 // cmpltps -0x70(%rsp),%xmm1
+ .byte 15,40,214 // movaps %xmm6,%xmm2
+ .byte 15,88,215 // addps %xmm7,%xmm2
.byte 15,84,209 // andps %xmm1,%xmm2
.byte 15,85,200 // andnps %xmm0,%xmm1
.byte 15,86,202 // orps %xmm2,%xmm1
- .byte 15,40,198 // movaps %xmm6,%xmm0
+ .byte 15,40,197 // movaps %xmm5,%xmm0
.byte 15,92,193 // subps %xmm1,%xmm0
- .byte 15,40,209 // movaps %xmm1,%xmm2
.byte 15,40,217 // movaps %xmm1,%xmm3
- .byte 15,194,76,36,144,1 // cmpltps -0x70(%rsp),%xmm1
+ .byte 15,40,225 // movaps %xmm1,%xmm4
+ .byte 15,40,209 // movaps %xmm1,%xmm2
+ .byte 15,194,76,36,128,1 // cmpltps -0x80(%rsp),%xmm1
.byte 65,15,89,193 // mulps %xmm9,%xmm0
.byte 65,15,88,197 // addps %xmm13,%xmm0
.byte 15,84,193 // andps %xmm1,%xmm0
.byte 65,15,85,205 // andnps %xmm13,%xmm1
.byte 15,86,200 // orps %xmm0,%xmm1
- .byte 15,194,223,1 // cmpltps %xmm7,%xmm3
+ .byte 68,15,40,92,36,176 // movaps -0x50(%rsp),%xmm11
+ .byte 65,15,194,211,1 // cmpltps %xmm11,%xmm2
.byte 65,15,40,192 // movaps %xmm8,%xmm0
- .byte 15,84,195 // andps %xmm3,%xmm0
- .byte 15,85,217 // andnps %xmm1,%xmm3
- .byte 15,86,216 // orps %xmm0,%xmm3
- .byte 15,194,212,1 // cmpltps %xmm4,%xmm2
- .byte 65,15,40,193 // movaps %xmm9,%xmm0
- .byte 15,89,197 // mulps %xmm5,%xmm0
- .byte 65,15,88,197 // addps %xmm13,%xmm0
.byte 15,84,194 // andps %xmm2,%xmm0
- .byte 15,85,211 // andnps %xmm3,%xmm2
+ .byte 15,85,209 // andnps %xmm1,%xmm2
.byte 15,86,208 // orps %xmm0,%xmm2
- .byte 65,15,40,219 // movaps %xmm11,%xmm3
+ .byte 65,15,194,218,1 // cmpltps %xmm10,%xmm3
+ .byte 65,15,89,225 // mulps %xmm9,%xmm4
+ .byte 65,15,88,229 // addps %xmm13,%xmm4
+ .byte 15,84,227 // andps %xmm3,%xmm4
.byte 15,85,218 // andnps %xmm2,%xmm3
- .byte 15,88,45,139,46,0,0 // addps 0x2e8b(%rip),%xmm5 # 40e0 <_sk_callback_sse2+0x3b9>
- .byte 15,40,197 // movaps %xmm5,%xmm0
- .byte 15,194,68,36,128,1 // cmpltps -0x80(%rsp),%xmm0
- .byte 68,15,194,245,1 // cmpltps %xmm5,%xmm14
- .byte 68,15,88,253 // addps %xmm5,%xmm15
+ .byte 15,86,220 // orps %xmm4,%xmm3
+ .byte 15,40,100,36,160 // movaps -0x60(%rsp),%xmm4
+ .byte 15,40,204 // movaps %xmm4,%xmm1
+ .byte 15,85,203 // andnps %xmm3,%xmm1
+ .byte 15,88,53,135,46,0,0 // addps 0x2e87(%rip),%xmm6 # 40e0 <_sk_callback_sse2+0x3b9>
+ .byte 15,88,254 // addps %xmm6,%xmm7
+ .byte 68,15,194,246,1 // cmpltps %xmm6,%xmm14
+ .byte 68,15,88,254 // addps %xmm6,%xmm15
.byte 69,15,84,254 // andps %xmm14,%xmm15
- .byte 68,15,85,245 // andnps %xmm5,%xmm14
+ .byte 68,15,85,246 // andnps %xmm6,%xmm14
+ .byte 15,194,116,36,144,1 // cmpltps -0x70(%rsp),%xmm6
.byte 69,15,86,247 // orps %xmm15,%xmm14
- .byte 68,15,88,229 // addps %xmm5,%xmm12
- .byte 68,15,84,224 // andps %xmm0,%xmm12
- .byte 65,15,85,198 // andnps %xmm14,%xmm0
- .byte 65,15,86,196 // orps %xmm12,%xmm0
- .byte 15,40,248 // movaps %xmm0,%xmm7
- .byte 15,194,252,1 // cmpltps %xmm4,%xmm7
- .byte 15,40,200 // movaps %xmm0,%xmm1
- .byte 15,194,76,36,160,1 // cmpltps -0x60(%rsp),%xmm1
- .byte 15,92,240 // subps %xmm0,%xmm6
- .byte 15,194,68,36,144,1 // cmpltps -0x70(%rsp),%xmm0
+ .byte 15,84,254 // andps %xmm6,%xmm7
+ .byte 65,15,85,246 // andnps %xmm14,%xmm6
+ .byte 15,86,247 // orps %xmm7,%xmm6
+ .byte 15,40,254 // movaps %xmm6,%xmm7
+ .byte 65,15,194,250,1 // cmpltps %xmm10,%xmm7
+ .byte 15,40,198 // movaps %xmm6,%xmm0
+ .byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0
+ .byte 15,92,238 // subps %xmm6,%xmm5
+ .byte 15,40,222 // movaps %xmm6,%xmm3
+ .byte 15,194,116,36,128,1 // cmpltps -0x80(%rsp),%xmm6
+ .byte 65,15,89,217 // mulps %xmm9,%xmm3
.byte 65,15,89,233 // mulps %xmm9,%xmm5
- .byte 65,15,89,241 // mulps %xmm9,%xmm6
+ .byte 65,15,88,221 // addps %xmm13,%xmm3
.byte 65,15,88,237 // addps %xmm13,%xmm5
- .byte 65,15,88,245 // addps %xmm13,%xmm6
- .byte 15,84,240 // andps %xmm0,%xmm6
- .byte 65,15,85,197 // andnps %xmm13,%xmm0
- .byte 15,86,198 // orps %xmm6,%xmm0
- .byte 68,15,84,193 // andps %xmm1,%xmm8
- .byte 15,85,200 // andnps %xmm0,%xmm1
- .byte 65,15,86,200 // orps %xmm8,%xmm1
- .byte 15,84,239 // andps %xmm7,%xmm5
- .byte 15,85,249 // andnps %xmm1,%xmm7
- .byte 15,86,253 // orps %xmm5,%xmm7
- .byte 69,15,84,211 // andps %xmm11,%xmm10
- .byte 68,15,85,223 // andnps %xmm7,%xmm11
- .byte 15,40,76,36,192 // movaps -0x40(%rsp),%xmm1
- .byte 65,15,86,202 // orps %xmm10,%xmm1
- .byte 65,15,86,218 // orps %xmm10,%xmm3
- .byte 69,15,86,211 // orps %xmm11,%xmm10
+ .byte 15,84,238 // andps %xmm6,%xmm5
+ .byte 65,15,85,245 // andnps %xmm13,%xmm6
+ .byte 15,86,245 // orps %xmm5,%xmm6
+ .byte 68,15,84,192 // andps %xmm0,%xmm8
+ .byte 15,85,198 // andnps %xmm6,%xmm0
+ .byte 65,15,86,192 // orps %xmm8,%xmm0
+ .byte 15,84,223 // andps %xmm7,%xmm3
+ .byte 15,85,248 // andnps %xmm0,%xmm7
+ .byte 15,86,251 // orps %xmm3,%xmm7
+ .byte 15,40,196 // movaps %xmm4,%xmm0
+ .byte 68,15,84,224 // andps %xmm0,%xmm12
+ .byte 15,85,199 // andnps %xmm7,%xmm0
+ .byte 15,40,84,36,192 // movaps -0x40(%rsp),%xmm2
+ .byte 65,15,86,212 // orps %xmm12,%xmm2
+ .byte 65,15,86,204 // orps %xmm12,%xmm1
+ .byte 68,15,86,224 // orps %xmm0,%xmm12
.byte 72,173 // lods %ds:(%rsi),%rax
- .byte 15,40,193 // movaps %xmm1,%xmm0
- .byte 15,40,203 // movaps %xmm3,%xmm1
- .byte 65,15,40,210 // movaps %xmm10,%xmm2
+ .byte 15,40,194 // movaps %xmm2,%xmm0
+ .byte 65,15,40,212 // movaps %xmm12,%xmm2
.byte 15,40,92,36,208 // movaps -0x30(%rsp),%xmm3
.byte 15,40,100,36,224 // movaps -0x20(%rsp),%xmm4
.byte 15,40,108,36,240 // movaps -0x10(%rsp),%xmm5
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index 488227b4c3..7e067eb89a 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 196,226,125,24,13,194,60,0,0 ; vbroadcastss 0x3cc2(%rip),%ymm1 # 3e1c <_sk_callback_hsw+0x119>
+ DB 196,226,125,24,13,186,60,0,0 ; vbroadcastss 0x3cba(%rip),%ymm1 # 3e14 <_sk_callback_hsw+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 196,226,125,24,21,166,60,0,0 ; vbroadcastss 0x3ca6(%rip),%ymm2 # 3e20 <_sk_callback_hsw+0x11d>
+ DB 196,226,125,24,21,158,60,0,0 ; vbroadcastss 0x3c9e(%rip),%ymm2 # 3e18 <_sk_callback_hsw+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
@@ -143,7 +143,7 @@ _sk_clear_hsw LABEL PROC
PUBLIC _sk_srcatop_hsw
_sk_srcatop_hsw LABEL PROC
DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0
- DB 196,98,125,24,5,86,60,0,0 ; vbroadcastss 0x3c56(%rip),%ymm8 # 3e24 <_sk_callback_hsw+0x121>
+ DB 196,98,125,24,5,78,60,0,0 ; vbroadcastss 0x3c4e(%rip),%ymm8 # 3e1c <_sk_callback_hsw+0x121>
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
@@ -157,7 +157,7 @@ _sk_srcatop_hsw LABEL PROC
PUBLIC _sk_dstatop_hsw
_sk_dstatop_hsw LABEL PROC
- DB 196,98,125,24,5,41,60,0,0 ; vbroadcastss 0x3c29(%rip),%ymm8 # 3e28 <_sk_callback_hsw+0x125>
+ DB 196,98,125,24,5,33,60,0,0 ; vbroadcastss 0x3c21(%rip),%ymm8 # 3e20 <_sk_callback_hsw+0x125>
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
@@ -190,7 +190,7 @@ _sk_dstin_hsw LABEL PROC
PUBLIC _sk_srcout_hsw
_sk_srcout_hsw LABEL PROC
- DB 196,98,125,24,5,208,59,0,0 ; vbroadcastss 0x3bd0(%rip),%ymm8 # 3e2c <_sk_callback_hsw+0x129>
+ DB 196,98,125,24,5,200,59,0,0 ; vbroadcastss 0x3bc8(%rip),%ymm8 # 3e24 <_sk_callback_hsw+0x129>
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
@@ -201,7 +201,7 @@ _sk_srcout_hsw LABEL PROC
PUBLIC _sk_dstout_hsw
_sk_dstout_hsw LABEL PROC
- DB 196,226,125,24,5,179,59,0,0 ; vbroadcastss 0x3bb3(%rip),%ymm0 # 3e30 <_sk_callback_hsw+0x12d>
+ DB 196,226,125,24,5,171,59,0,0 ; vbroadcastss 0x3bab(%rip),%ymm0 # 3e28 <_sk_callback_hsw+0x12d>
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
@@ -212,7 +212,7 @@ _sk_dstout_hsw LABEL PROC
PUBLIC _sk_srcover_hsw
_sk_srcover_hsw LABEL PROC
- DB 196,98,125,24,5,150,59,0,0 ; vbroadcastss 0x3b96(%rip),%ymm8 # 3e34 <_sk_callback_hsw+0x131>
+ DB 196,98,125,24,5,142,59,0,0 ; vbroadcastss 0x3b8e(%rip),%ymm8 # 3e2c <_sk_callback_hsw+0x131>
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
@@ -223,7 +223,7 @@ _sk_srcover_hsw LABEL PROC
PUBLIC _sk_dstover_hsw
_sk_dstover_hsw LABEL PROC
- DB 196,98,125,24,5,117,59,0,0 ; vbroadcastss 0x3b75(%rip),%ymm8 # 3e38 <_sk_callback_hsw+0x135>
+ DB 196,98,125,24,5,109,59,0,0 ; vbroadcastss 0x3b6d(%rip),%ymm8 # 3e30 <_sk_callback_hsw+0x135>
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
@@ -243,7 +243,7 @@ _sk_modulate_hsw LABEL PROC
PUBLIC _sk_multiply_hsw
_sk_multiply_hsw LABEL PROC
- DB 196,98,125,24,5,64,59,0,0 ; vbroadcastss 0x3b40(%rip),%ymm8 # 3e3c <_sk_callback_hsw+0x139>
+ DB 196,98,125,24,5,56,59,0,0 ; vbroadcastss 0x3b38(%rip),%ymm8 # 3e34 <_sk_callback_hsw+0x139>
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
@@ -285,7 +285,7 @@ _sk_screen_hsw LABEL PROC
PUBLIC _sk_xor__hsw
_sk_xor__hsw LABEL PROC
- DB 196,98,125,24,5,187,58,0,0 ; vbroadcastss 0x3abb(%rip),%ymm8 # 3e40 <_sk_callback_hsw+0x13d>
+ DB 196,98,125,24,5,179,58,0,0 ; vbroadcastss 0x3ab3(%rip),%ymm8 # 3e38 <_sk_callback_hsw+0x13d>
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
@@ -317,7 +317,7 @@ _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 196,98,125,24,5,67,58,0,0 ; vbroadcastss 0x3a43(%rip),%ymm8 # 3e44 <_sk_callback_hsw+0x141>
+ DB 196,98,125,24,5,59,58,0,0 ; vbroadcastss 0x3a3b(%rip),%ymm8 # 3e3c <_sk_callback_hsw+0x141>
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
@@ -340,7 +340,7 @@ _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 196,98,125,24,5,242,57,0,0 ; vbroadcastss 0x39f2(%rip),%ymm8 # 3e48 <_sk_callback_hsw+0x145>
+ DB 196,98,125,24,5,234,57,0,0 ; vbroadcastss 0x39ea(%rip),%ymm8 # 3e40 <_sk_callback_hsw+0x145>
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
@@ -366,7 +366,7 @@ _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 196,98,125,24,5,149,57,0,0 ; vbroadcastss 0x3995(%rip),%ymm8 # 3e4c <_sk_callback_hsw+0x149>
+ DB 196,98,125,24,5,141,57,0,0 ; vbroadcastss 0x398d(%rip),%ymm8 # 3e44 <_sk_callback_hsw+0x149>
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
@@ -386,7 +386,7 @@ _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 196,98,125,24,5,83,57,0,0 ; vbroadcastss 0x3953(%rip),%ymm8 # 3e50 <_sk_callback_hsw+0x14d>
+ DB 196,98,125,24,5,75,57,0,0 ; vbroadcastss 0x394b(%rip),%ymm8 # 3e48 <_sk_callback_hsw+0x14d>
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
@@ -394,7 +394,7 @@ _sk_exclusion_hsw LABEL PROC
PUBLIC _sk_colorburn_hsw
_sk_colorburn_hsw LABEL PROC
- DB 196,98,125,24,5,65,57,0,0 ; vbroadcastss 0x3941(%rip),%ymm8 # 3e54 <_sk_callback_hsw+0x151>
+ DB 196,98,125,24,5,57,57,0,0 ; vbroadcastss 0x3939(%rip),%ymm8 # 3e4c <_sk_callback_hsw+0x151>
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
@@ -450,7 +450,7 @@ _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 196,98,125,24,13,76,56,0,0 ; vbroadcastss 0x384c(%rip),%ymm9 # 3e58 <_sk_callback_hsw+0x155>
+ DB 196,98,125,24,13,68,56,0,0 ; vbroadcastss 0x3844(%rip),%ymm9 # 3e50 <_sk_callback_hsw+0x155>
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
@@ -501,7 +501,7 @@ _sk_colordodge_hsw LABEL PROC
PUBLIC _sk_hardlight_hsw
_sk_hardlight_hsw LABEL PROC
- DB 196,98,125,24,5,109,55,0,0 ; vbroadcastss 0x376d(%rip),%ymm8 # 3e5c <_sk_callback_hsw+0x159>
+ DB 196,98,125,24,5,101,55,0,0 ; vbroadcastss 0x3765(%rip),%ymm8 # 3e54 <_sk_callback_hsw+0x159>
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
@@ -550,7 +550,7 @@ _sk_hardlight_hsw LABEL PROC
PUBLIC _sk_overlay_hsw
_sk_overlay_hsw LABEL PROC
- DB 196,98,125,24,5,165,54,0,0 ; vbroadcastss 0x36a5(%rip),%ymm8 # 3e60 <_sk_callback_hsw+0x15d>
+ DB 196,98,125,24,5,157,54,0,0 ; vbroadcastss 0x369d(%rip),%ymm8 # 3e58 <_sk_callback_hsw+0x15d>
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
@@ -610,10 +610,10 @@ _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 196,98,125,24,29,172,53,0,0 ; vbroadcastss 0x35ac(%rip),%ymm11 # 3e68 <_sk_callback_hsw+0x165>
+ DB 196,98,125,24,29,164,53,0,0 ; vbroadcastss 0x35a4(%rip),%ymm11 # 3e60 <_sk_callback_hsw+0x165>
DB 196,65,20,88,227 ; vaddps %ymm11,%ymm13,%ymm12
DB 196,65,28,89,192 ; vmulps %ymm8,%ymm12,%ymm8
- DB 196,98,125,24,37,157,53,0,0 ; vbroadcastss 0x359d(%rip),%ymm12 # 3e6c <_sk_callback_hsw+0x169>
+ DB 196,98,125,24,37,149,53,0,0 ; vbroadcastss 0x3595(%rip),%ymm12 # 3e64 <_sk_callback_hsw+0x169>
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
@@ -623,7 +623,7 @@ _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 196,98,125,24,5,96,53,0,0 ; vbroadcastss 0x3560(%rip),%ymm8 # 3e64 <_sk_callback_hsw+0x161>
+ DB 196,98,125,24,5,88,53,0,0 ; vbroadcastss 0x3558(%rip),%ymm8 # 3e5c <_sk_callback_hsw+0x161>
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
@@ -713,7 +713,7 @@ _sk_clamp_0_hsw LABEL PROC
PUBLIC _sk_clamp_1_hsw
_sk_clamp_1_hsw LABEL PROC
- DB 196,98,125,24,5,227,51,0,0 ; vbroadcastss 0x33e3(%rip),%ymm8 # 3e70 <_sk_callback_hsw+0x16d>
+ DB 196,98,125,24,5,219,51,0,0 ; vbroadcastss 0x33db(%rip),%ymm8 # 3e68 <_sk_callback_hsw+0x16d>
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
@@ -723,7 +723,7 @@ _sk_clamp_1_hsw LABEL PROC
PUBLIC _sk_clamp_a_hsw
_sk_clamp_a_hsw LABEL PROC
- DB 196,98,125,24,5,198,51,0,0 ; vbroadcastss 0x33c6(%rip),%ymm8 # 3e74 <_sk_callback_hsw+0x171>
+ DB 196,98,125,24,5,190,51,0,0 ; vbroadcastss 0x33be(%rip),%ymm8 # 3e6c <_sk_callback_hsw+0x171>
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
@@ -795,7 +795,7 @@ 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 196,98,125,24,21,14,51,0,0 ; vbroadcastss 0x330e(%rip),%ymm10 # 3e78 <_sk_callback_hsw+0x175>
+ DB 196,98,125,24,21,6,51,0,0 ; vbroadcastss 0x3306(%rip),%ymm10 # 3e70 <_sk_callback_hsw+0x175>
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
@@ -806,16 +806,16 @@ _sk_unpremul_hsw LABEL PROC
PUBLIC _sk_from_srgb_hsw
_sk_from_srgb_hsw LABEL PROC
- DB 196,98,125,24,5,239,50,0,0 ; vbroadcastss 0x32ef(%rip),%ymm8 # 3e7c <_sk_callback_hsw+0x179>
+ DB 196,98,125,24,5,231,50,0,0 ; vbroadcastss 0x32e7(%rip),%ymm8 # 3e74 <_sk_callback_hsw+0x179>
DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9
DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10
- DB 196,98,125,24,29,225,50,0,0 ; vbroadcastss 0x32e1(%rip),%ymm11 # 3e80 <_sk_callback_hsw+0x17d>
- DB 196,98,125,24,37,220,50,0,0 ; vbroadcastss 0x32dc(%rip),%ymm12 # 3e84 <_sk_callback_hsw+0x181>
+ DB 196,98,125,24,29,217,50,0,0 ; vbroadcastss 0x32d9(%rip),%ymm11 # 3e78 <_sk_callback_hsw+0x17d>
+ DB 196,98,125,24,37,212,50,0,0 ; vbroadcastss 0x32d4(%rip),%ymm12 # 3e7c <_sk_callback_hsw+0x181>
DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13
DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13
- DB 196,98,125,24,53,205,50,0,0 ; vbroadcastss 0x32cd(%rip),%ymm14 # 3e88 <_sk_callback_hsw+0x185>
+ DB 196,98,125,24,53,197,50,0,0 ; vbroadcastss 0x32c5(%rip),%ymm14 # 3e80 <_sk_callback_hsw+0x185>
DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13
- DB 196,98,125,24,21,195,50,0,0 ; vbroadcastss 0x32c3(%rip),%ymm10 # 3e8c <_sk_callback_hsw+0x189>
+ DB 196,98,125,24,21,187,50,0,0 ; vbroadcastss 0x32bb(%rip),%ymm10 # 3e84 <_sk_callback_hsw+0x189>
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
@@ -839,16 +839,16 @@ _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 196,98,125,24,5,93,50,0,0 ; vbroadcastss 0x325d(%rip),%ymm8 # 3e90 <_sk_callback_hsw+0x18d>
+ DB 196,98,125,24,5,85,50,0,0 ; vbroadcastss 0x3255(%rip),%ymm8 # 3e88 <_sk_callback_hsw+0x18d>
DB 196,65,124,89,216 ; vmulps %ymm8,%ymm0,%ymm11
- DB 196,98,125,24,37,83,50,0,0 ; vbroadcastss 0x3253(%rip),%ymm12 # 3e94 <_sk_callback_hsw+0x191>
- DB 196,98,125,24,45,78,50,0,0 ; vbroadcastss 0x324e(%rip),%ymm13 # 3e98 <_sk_callback_hsw+0x195>
+ DB 196,98,125,24,37,75,50,0,0 ; vbroadcastss 0x324b(%rip),%ymm12 # 3e8c <_sk_callback_hsw+0x191>
+ DB 196,98,125,24,45,70,50,0,0 ; vbroadcastss 0x3246(%rip),%ymm13 # 3e90 <_sk_callback_hsw+0x195>
DB 196,66,21,168,204 ; vfmadd213ps %ymm12,%ymm13,%ymm9
- DB 196,98,125,24,53,68,50,0,0 ; vbroadcastss 0x3244(%rip),%ymm14 # 3e9c <_sk_callback_hsw+0x199>
+ DB 196,98,125,24,53,60,50,0,0 ; vbroadcastss 0x323c(%rip),%ymm14 # 3e94 <_sk_callback_hsw+0x199>
DB 196,66,13,184,202 ; vfmadd231ps %ymm10,%ymm14,%ymm9
- DB 196,98,125,24,21,58,50,0,0 ; vbroadcastss 0x323a(%rip),%ymm10 # 3ea0 <_sk_callback_hsw+0x19d>
+ DB 196,98,125,24,21,50,50,0,0 ; vbroadcastss 0x3232(%rip),%ymm10 # 3e98 <_sk_callback_hsw+0x19d>
DB 196,65,44,93,201 ; vminps %ymm9,%ymm10,%ymm9
- DB 196,98,125,24,61,48,50,0,0 ; vbroadcastss 0x3230(%rip),%ymm15 # 3ea4 <_sk_callback_hsw+0x1a1>
+ DB 196,98,125,24,61,40,50,0,0 ; vbroadcastss 0x3228(%rip),%ymm15 # 3e9c <_sk_callback_hsw+0x1a1>
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
@@ -879,26 +879,26 @@ _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 196,98,125,24,29,170,49,0,0 ; vbroadcastss 0x31aa(%rip),%ymm11 # 3ea8 <_sk_callback_hsw+0x1a5>
+ DB 196,98,125,24,29,162,49,0,0 ; vbroadcastss 0x31a2(%rip),%ymm11 # 3ea0 <_sk_callback_hsw+0x1a5>
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 196,98,125,24,53,151,49,0,0 ; vbroadcastss 0x3197(%rip),%ymm14 # 3eac <_sk_callback_hsw+0x1a9>
+ DB 196,98,125,24,53,143,49,0,0 ; vbroadcastss 0x318f(%rip),%ymm14 # 3ea4 <_sk_callback_hsw+0x1a9>
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 196,98,125,24,53,126,49,0,0 ; vbroadcastss 0x317e(%rip),%ymm14 # 3eb4 <_sk_callback_hsw+0x1b1>
+ DB 196,98,125,24,53,118,49,0,0 ; vbroadcastss 0x3176(%rip),%ymm14 # 3eac <_sk_callback_hsw+0x1b1>
DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13
- DB 196,98,125,24,53,108,49,0,0 ; vbroadcastss 0x316c(%rip),%ymm14 # 3eb0 <_sk_callback_hsw+0x1ad>
+ DB 196,98,125,24,53,100,49,0,0 ; vbroadcastss 0x3164(%rip),%ymm14 # 3ea8 <_sk_callback_hsw+0x1ad>
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 196,98,125,24,29,79,49,0,0 ; vbroadcastss 0x314f(%rip),%ymm11 # 3ebc <_sk_callback_hsw+0x1b9>
+ DB 196,98,125,24,29,71,49,0,0 ; vbroadcastss 0x3147(%rip),%ymm11 # 3eb4 <_sk_callback_hsw+0x1b9>
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
@@ -908,7 +908,7 @@ _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 196,98,125,24,5,18,49,0,0 ; vbroadcastss 0x3112(%rip),%ymm8 # 3eb8 <_sk_callback_hsw+0x1b5>
+ DB 196,98,125,24,5,10,49,0,0 ; vbroadcastss 0x310a(%rip),%ymm8 # 3eb0 <_sk_callback_hsw+0x1b5>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -921,95 +921,93 @@ _sk_hsl_to_rgb_hsw LABEL PROC
DB 197,252,17,172,36,128,0,0,0 ; vmovups %ymm5,0x80(%rsp)
DB 197,252,17,100,36,96 ; vmovups %ymm4,0x60(%rsp)
DB 197,252,17,92,36,64 ; vmovups %ymm3,0x40(%rsp)
- DB 197,252,40,234 ; vmovaps %ymm2,%ymm5
- DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
- DB 184,0,0,0,63 ; mov $0x3f000000,%eax
- DB 197,249,110,192 ; vmovd %eax,%xmm0
- DB 196,98,125,88,192 ; vpbroadcastd %xmm0,%ymm8
- DB 196,193,84,194,192,1 ; vcmpltps %ymm8,%ymm5,%ymm0
- DB 196,98,125,24,21,190,48,0,0 ; vbroadcastss 0x30be(%rip),%ymm10 # 3ec0 <_sk_callback_hsw+0x1bd>
DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp)
+ DB 184,0,0,0,63 ; mov $0x3f000000,%eax
+ DB 197,249,110,216 ; vmovd %eax,%xmm3
+ DB 196,98,125,88,195 ; vpbroadcastd %xmm3,%ymm8
+ DB 196,193,108,194,232,1 ; vcmpltps %ymm8,%ymm2,%ymm5
+ DB 196,98,125,24,21,184,48,0,0 ; vbroadcastss 0x30b8(%rip),%ymm10 # 3eb8 <_sk_callback_hsw+0x1bd>
DB 196,193,116,88,218 ; vaddps %ymm10,%ymm1,%ymm3
- DB 197,228,89,221 ; vmulps %ymm5,%ymm3,%ymm3
- DB 197,244,88,229 ; vaddps %ymm5,%ymm1,%ymm4
- DB 196,226,117,188,229 ; vfnmadd231ps %ymm5,%ymm1,%ymm4
- DB 196,99,93,74,203,0 ; vblendvps %ymm0,%ymm3,%ymm4,%ymm9
- DB 196,226,125,24,13,159,48,0,0 ; vbroadcastss 0x309f(%rip),%ymm1 # 3ec8 <_sk_callback_hsw+0x1c5>
- DB 197,236,88,241 ; vaddps %ymm1,%ymm2,%ymm6
+ DB 197,228,89,218 ; vmulps %ymm2,%ymm3,%ymm3
+ DB 197,244,88,226 ; vaddps %ymm2,%ymm1,%ymm4
+ DB 196,226,117,188,226 ; vfnmadd231ps %ymm2,%ymm1,%ymm4
+ DB 196,99,93,74,203,80 ; vblendvps %ymm5,%ymm3,%ymm4,%ymm9
+ DB 196,226,125,24,13,159,48,0,0 ; vbroadcastss 0x309f(%rip),%ymm1 # 3ec0 <_sk_callback_hsw+0x1c5>
+ DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1
DB 65,184,0,0,0,0 ; mov $0x0,%r8d
DB 184,0,0,128,63 ; mov $0x3f800000,%eax
- DB 197,249,110,200 ; vmovd %eax,%xmm1
- DB 196,98,125,88,225 ; vpbroadcastd %xmm1,%ymm12
- DB 197,156,194,206,1 ; vcmpltps %ymm6,%ymm12,%ymm1
- DB 196,98,125,24,45,125,48,0,0 ; vbroadcastss 0x307d(%rip),%ymm13 # 3ecc <_sk_callback_hsw+0x1c9>
- DB 196,193,76,88,221 ; vaddps %ymm13,%ymm6,%ymm3
- DB 196,227,77,74,203,16 ; vblendvps %ymm1,%ymm3,%ymm6,%ymm1
- DB 196,193,121,110,216 ; vmovd %r8d,%xmm3
- DB 196,98,125,88,251 ; vpbroadcastd %xmm3,%ymm15
- DB 196,193,76,194,223,1 ; vcmpltps %ymm15,%ymm6,%ymm3
- DB 196,193,76,88,226 ; vaddps %ymm10,%ymm6,%ymm4
- DB 196,227,117,74,196,48 ; vblendvps %ymm3,%ymm4,%ymm1,%ymm0
- DB 196,98,125,24,29,70,48,0,0 ; vbroadcastss 0x3046(%rip),%ymm11 # 3ec4 <_sk_callback_hsw+0x1c1>
- DB 196,66,85,170,217 ; vfmsub213ps %ymm9,%ymm5,%ymm11
+ DB 197,249,110,216 ; vmovd %eax,%xmm3
+ DB 196,98,125,88,227 ; vpbroadcastd %xmm3,%ymm12
+ DB 197,156,194,217,1 ; vcmpltps %ymm1,%ymm12,%ymm3
+ DB 196,98,125,24,45,125,48,0,0 ; vbroadcastss 0x307d(%rip),%ymm13 # 3ec4 <_sk_callback_hsw+0x1c9>
+ DB 196,193,116,88,229 ; vaddps %ymm13,%ymm1,%ymm4
+ DB 196,227,117,74,220,48 ; vblendvps %ymm3,%ymm4,%ymm1,%ymm3
+ DB 196,193,121,110,224 ; vmovd %r8d,%xmm4
+ DB 196,98,125,88,252 ; vpbroadcastd %xmm4,%ymm15
+ DB 196,193,116,194,231,1 ; vcmpltps %ymm15,%ymm1,%ymm4
+ DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
+ DB 196,227,101,74,241,64 ; vblendvps %ymm4,%ymm1,%ymm3,%ymm6
+ DB 196,98,125,24,29,70,48,0,0 ; vbroadcastss 0x3046(%rip),%ymm11 # 3ebc <_sk_callback_hsw+0x1c1>
+ DB 196,66,109,170,217 ; vfmsub213ps %ymm9,%ymm2,%ymm11
DB 196,193,52,92,203 ; vsubps %ymm11,%ymm9,%ymm1
- DB 196,226,125,24,29,63,48,0,0 ; vbroadcastss 0x303f(%rip),%ymm3 # 3ed0 <_sk_callback_hsw+0x1cd>
+ DB 196,226,125,24,29,63,48,0,0 ; vbroadcastss 0x303f(%rip),%ymm3 # 3ec8 <_sk_callback_hsw+0x1cd>
DB 197,116,89,243 ; vmulps %ymm3,%ymm1,%ymm14
DB 65,184,171,170,42,62 ; mov $0x3e2aaaab,%r8d
DB 184,171,170,42,63 ; mov $0x3f2aaaab,%eax
DB 197,249,110,200 ; vmovd %eax,%xmm1
- DB 196,226,125,88,225 ; vpbroadcastd %xmm1,%ymm4
- DB 196,226,125,24,29,34,48,0,0 ; vbroadcastss 0x3022(%rip),%ymm3 # 3ed4 <_sk_callback_hsw+0x1d1>
- DB 197,228,92,200 ; vsubps %ymm0,%ymm3,%ymm1
+ DB 196,226,125,88,233 ; vpbroadcastd %xmm1,%ymm5
+ DB 196,226,125,24,37,34,48,0,0 ; vbroadcastss 0x3022(%rip),%ymm4 # 3ecc <_sk_callback_hsw+0x1d1>
+ DB 197,220,92,206 ; vsubps %ymm6,%ymm4,%ymm1
DB 196,194,13,168,203 ; vfmadd213ps %ymm11,%ymm14,%ymm1
- DB 197,252,194,252,1 ; vcmpltps %ymm4,%ymm0,%ymm7
+ DB 197,204,194,253,1 ; vcmpltps %ymm5,%ymm6,%ymm7
DB 196,227,37,74,201,112 ; vblendvps %ymm7,%ymm1,%ymm11,%ymm1
- DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7
+ DB 196,193,76,194,248,1 ; vcmpltps %ymm8,%ymm6,%ymm7
DB 196,195,117,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm1,%ymm7
DB 196,193,121,110,200 ; vmovd %r8d,%xmm1
- DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1
- DB 197,252,194,193,1 ; vcmpltps %ymm1,%ymm0,%ymm0
+ DB 196,226,125,88,217 ; vpbroadcastd %xmm1,%ymm3
+ DB 197,204,194,203,1 ; vcmpltps %ymm3,%ymm6,%ymm1
DB 196,194,13,168,243 ; vfmadd213ps %ymm11,%ymm14,%ymm6
- DB 196,227,69,74,198,0 ; vblendvps %ymm0,%ymm6,%ymm7,%ymm0
- DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp)
- DB 197,156,194,194,1 ; vcmpltps %ymm2,%ymm12,%ymm0
- DB 196,193,108,88,253 ; vaddps %ymm13,%ymm2,%ymm7
- DB 196,227,109,74,199,0 ; vblendvps %ymm0,%ymm7,%ymm2,%ymm0
- DB 196,193,108,194,255,1 ; vcmpltps %ymm15,%ymm2,%ymm7
- DB 196,193,108,88,242 ; vaddps %ymm10,%ymm2,%ymm6
- DB 196,227,125,74,198,112 ; vblendvps %ymm7,%ymm6,%ymm0,%ymm0
- DB 197,228,92,240 ; vsubps %ymm0,%ymm3,%ymm6
+ DB 196,227,69,74,206,16 ; vblendvps %ymm1,%ymm6,%ymm7,%ymm1
+ DB 197,252,17,12,36 ; vmovups %ymm1,(%rsp)
+ DB 197,156,194,200,1 ; vcmpltps %ymm0,%ymm12,%ymm1
+ DB 196,193,124,88,253 ; vaddps %ymm13,%ymm0,%ymm7
+ DB 196,227,125,74,207,16 ; vblendvps %ymm1,%ymm7,%ymm0,%ymm1
+ DB 196,193,124,194,255,1 ; vcmpltps %ymm15,%ymm0,%ymm7
+ DB 196,193,124,88,242 ; vaddps %ymm10,%ymm0,%ymm6
+ DB 196,227,117,74,206,112 ; vblendvps %ymm7,%ymm6,%ymm1,%ymm1
+ DB 197,220,92,241 ; vsubps %ymm1,%ymm4,%ymm6
DB 196,194,13,168,243 ; vfmadd213ps %ymm11,%ymm14,%ymm6
- DB 197,252,194,252,1 ; vcmpltps %ymm4,%ymm0,%ymm7
+ DB 197,244,194,253,1 ; vcmpltps %ymm5,%ymm1,%ymm7
DB 196,227,37,74,246,112 ; vblendvps %ymm7,%ymm6,%ymm11,%ymm6
- DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7
+ DB 196,193,116,194,248,1 ; vcmpltps %ymm8,%ymm1,%ymm7
DB 196,195,77,74,241,112 ; vblendvps %ymm7,%ymm9,%ymm6,%ymm6
- DB 197,252,194,193,1 ; vcmpltps %ymm1,%ymm0,%ymm0
- DB 197,252,40,250 ; vmovaps %ymm2,%ymm7
- DB 196,194,13,168,251 ; vfmadd213ps %ymm11,%ymm14,%ymm7
- DB 196,227,77,74,247,0 ; vblendvps %ymm0,%ymm7,%ymm6,%ymm6
- DB 196,226,125,24,5,137,47,0,0 ; vbroadcastss 0x2f89(%rip),%ymm0 # 3ed8 <_sk_callback_hsw+0x1d5>
- DB 197,236,88,192 ; vaddps %ymm0,%ymm2,%ymm0
- DB 197,156,194,208,1 ; vcmpltps %ymm0,%ymm12,%ymm2
+ DB 197,244,194,251,1 ; vcmpltps %ymm3,%ymm1,%ymm7
+ DB 196,194,13,168,203 ; vfmadd213ps %ymm11,%ymm14,%ymm1
+ DB 196,227,77,74,201,112 ; vblendvps %ymm7,%ymm1,%ymm6,%ymm1
+ DB 196,226,125,24,53,141,47,0,0 ; vbroadcastss 0x2f8d(%rip),%ymm6 # 3ed0 <_sk_callback_hsw+0x1d5>
+ DB 197,252,88,198 ; vaddps %ymm6,%ymm0,%ymm0
+ DB 197,156,194,240,1 ; vcmpltps %ymm0,%ymm12,%ymm6
DB 196,193,124,88,253 ; vaddps %ymm13,%ymm0,%ymm7
- DB 196,227,125,74,215,32 ; vblendvps %ymm2,%ymm7,%ymm0,%ymm2
+ DB 196,227,125,74,247,96 ; vblendvps %ymm6,%ymm7,%ymm0,%ymm6
DB 196,193,124,194,255,1 ; vcmpltps %ymm15,%ymm0,%ymm7
- DB 196,65,124,88,210 ; vaddps %ymm10,%ymm0,%ymm10
- DB 196,195,109,74,210,112 ; vblendvps %ymm7,%ymm10,%ymm2,%ymm2
- DB 196,194,13,168,195 ; vfmadd213ps %ymm11,%ymm14,%ymm0
- DB 197,228,92,218 ; vsubps %ymm2,%ymm3,%ymm3
- DB 196,194,13,168,219 ; vfmadd213ps %ymm11,%ymm14,%ymm3
- DB 197,236,194,228,1 ; vcmpltps %ymm4,%ymm2,%ymm4
- DB 196,227,37,74,219,64 ; vblendvps %ymm4,%ymm3,%ymm11,%ymm3
- DB 196,193,108,194,224,1 ; vcmpltps %ymm8,%ymm2,%ymm4
- DB 196,195,101,74,217,64 ; vblendvps %ymm4,%ymm9,%ymm3,%ymm3
- DB 197,236,194,201,1 ; vcmpltps %ymm1,%ymm2,%ymm1
- DB 196,227,101,74,208,16 ; vblendvps %ymm1,%ymm0,%ymm3,%ymm2
+ DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
+ DB 196,227,77,74,192,112 ; vblendvps %ymm7,%ymm0,%ymm6,%ymm0
+ DB 197,220,92,224 ; vsubps %ymm0,%ymm4,%ymm4
+ DB 197,252,40,240 ; vmovaps %ymm0,%ymm6
+ DB 196,194,13,168,243 ; vfmadd213ps %ymm11,%ymm14,%ymm6
+ DB 196,194,13,168,227 ; vfmadd213ps %ymm11,%ymm14,%ymm4
+ DB 197,252,194,237,1 ; vcmpltps %ymm5,%ymm0,%ymm5
+ DB 196,227,37,74,228,80 ; vblendvps %ymm5,%ymm4,%ymm11,%ymm4
+ DB 196,193,124,194,232,1 ; vcmpltps %ymm8,%ymm0,%ymm5
+ DB 196,195,93,74,225,80 ; vblendvps %ymm5,%ymm9,%ymm4,%ymm4
+ DB 197,252,194,195,1 ; vcmpltps %ymm3,%ymm0,%ymm0
+ DB 196,227,93,74,222,0 ; vblendvps %ymm0,%ymm6,%ymm4,%ymm3
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
- DB 197,252,194,92,36,32,0 ; vcmpeqps 0x20(%rsp),%ymm0,%ymm3
+ DB 197,252,194,100,36,32,0 ; vcmpeqps 0x20(%rsp),%ymm0,%ymm4
DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0
- DB 196,227,125,74,197,48 ; vblendvps %ymm3,%ymm5,%ymm0,%ymm0
- DB 196,227,77,74,205,48 ; vblendvps %ymm3,%ymm5,%ymm6,%ymm1
- DB 196,227,109,74,213,48 ; vblendvps %ymm3,%ymm5,%ymm2,%ymm2
+ DB 196,227,125,74,194,64 ; vblendvps %ymm4,%ymm2,%ymm0,%ymm0
+ DB 196,227,117,74,202,64 ; vblendvps %ymm4,%ymm2,%ymm1,%ymm1
+ DB 196,227,101,74,210,64 ; vblendvps %ymm4,%ymm2,%ymm3,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,16,92,36,64 ; vmovups 0x40(%rsp),%ymm3
DB 197,252,16,100,36,96 ; vmovups 0x60(%rsp),%ymm4
@@ -1037,11 +1035,11 @@ _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 1056 <_sk_scale_u8_hsw+0x43>
+ DB 117,51 ; jne 104e <_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 196,98,125,24,13,162,46,0,0 ; vbroadcastss 0x2ea2(%rip),%ymm9 # 3edc <_sk_callback_hsw+0x1d9>
+ DB 196,98,125,24,13,162,46,0,0 ; vbroadcastss 0x2ea2(%rip),%ymm9 # 3ed4 <_sk_callback_hsw+0x1d9>
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
@@ -1059,9 +1057,9 @@ _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 105e <_sk_scale_u8_hsw+0x4b>
+ DB 117,234 ; jne 1056 <_sk_scale_u8_hsw+0x4b>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 235,172 ; jmp 1027 <_sk_scale_u8_hsw+0x14>
+ DB 235,172 ; jmp 101f <_sk_scale_u8_hsw+0x14>
PUBLIC _sk_lerp_1_float_hsw
_sk_lerp_1_float_hsw LABEL PROC
@@ -1085,11 +1083,11 @@ _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 1101 <_sk_lerp_u8_hsw+0x57>
+ DB 117,71 ; jne 10f9 <_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 196,98,125,24,13,15,46,0,0 ; vbroadcastss 0x2e0f(%rip),%ymm9 # 3ee0 <_sk_callback_hsw+0x1dd>
+ DB 196,98,125,24,13,15,46,0,0 ; vbroadcastss 0x2e0f(%rip),%ymm9 # 3ed8 <_sk_callback_hsw+0x1dd>
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
@@ -1111,32 +1109,32 @@ _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 1109 <_sk_lerp_u8_hsw+0x5f>
+ DB 117,234 ; jne 1101 <_sk_lerp_u8_hsw+0x5f>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 235,152 ; jmp 10be <_sk_lerp_u8_hsw+0x14>
+ DB 235,152 ; jmp 10b6 <_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 11c9 <_sk_lerp_565_hsw+0xa3>
+ DB 15,133,149,0,0,0 ; jne 11c1 <_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 196,98,125,88,5,156,45,0,0 ; vpbroadcastd 0x2d9c(%rip),%ymm8 # 3ee4 <_sk_callback_hsw+0x1e1>
+ DB 196,98,125,88,5,156,45,0,0 ; vpbroadcastd 0x2d9c(%rip),%ymm8 # 3edc <_sk_callback_hsw+0x1e1>
DB 196,65,101,219,192 ; vpand %ymm8,%ymm3,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,141,45,0,0 ; vbroadcastss 0x2d8d(%rip),%ymm9 # 3ee8 <_sk_callback_hsw+0x1e5>
+ DB 196,98,125,24,13,141,45,0,0 ; vbroadcastss 0x2d8d(%rip),%ymm9 # 3ee0 <_sk_callback_hsw+0x1e5>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,88,13,131,45,0,0 ; vpbroadcastd 0x2d83(%rip),%ymm9 # 3eec <_sk_callback_hsw+0x1e9>
+ DB 196,98,125,88,13,131,45,0,0 ; vpbroadcastd 0x2d83(%rip),%ymm9 # 3ee4 <_sk_callback_hsw+0x1e9>
DB 196,65,101,219,201 ; vpand %ymm9,%ymm3,%ymm9
DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9
- DB 196,98,125,24,21,116,45,0,0 ; vbroadcastss 0x2d74(%rip),%ymm10 # 3ef0 <_sk_callback_hsw+0x1ed>
+ DB 196,98,125,24,21,116,45,0,0 ; vbroadcastss 0x2d74(%rip),%ymm10 # 3ee8 <_sk_callback_hsw+0x1ed>
DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
- DB 196,98,125,88,21,106,45,0,0 ; vpbroadcastd 0x2d6a(%rip),%ymm10 # 3ef4 <_sk_callback_hsw+0x1f1>
+ DB 196,98,125,88,21,106,45,0,0 ; vpbroadcastd 0x2d6a(%rip),%ymm10 # 3eec <_sk_callback_hsw+0x1f1>
DB 196,193,101,219,218 ; vpand %ymm10,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,21,92,45,0,0 ; vbroadcastss 0x2d5c(%rip),%ymm10 # 3ef8 <_sk_callback_hsw+0x1f5>
+ DB 196,98,125,24,21,92,45,0,0 ; vbroadcastss 0x2d5c(%rip),%ymm10 # 3ef0 <_sk_callback_hsw+0x1f5>
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
@@ -1145,16 +1143,16 @@ _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 196,226,125,24,29,53,45,0,0 ; vbroadcastss 0x2d35(%rip),%ymm3 # 3efc <_sk_callback_hsw+0x1f9>
+ DB 196,226,125,24,29,53,45,0,0 ; vbroadcastss 0x2d35(%rip),%ymm3 # 3ef4 <_sk_callback_hsw+0x1f9>
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 113a <_sk_lerp_565_hsw+0x14>
+ DB 15,135,89,255,255,255 ; ja 1132 <_sk_lerp_565_hsw+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1238 <_sk_lerp_565_hsw+0x112>
+ DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1230 <_sk_lerp_565_hsw+0x112>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -1166,13 +1164,13 @@ _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 113a <_sk_lerp_565_hsw+0x14>
+ DB 233,5,255,255,255 ; jmpq 1132 <_sk_lerp_565_hsw+0x14>
DB 15,31,0 ; nopl (%rax)
DB 241 ; icebp
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 233,255,255,255,225 ; jmpq ffffffffe2001240 <_sk_callback_hsw+0xffffffffe1ffd53d>
+ DB 233,255,255,255,225 ; jmpq ffffffffe2001238 <_sk_callback_hsw+0xffffffffe1ffd53d>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
@@ -1197,23 +1195,23 @@ _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 12d2 <_sk_load_tables_hsw+0x7e>
+ DB 117,105 ; jne 12ca <_sk_load_tables_hsw+0x7e>
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
- DB 197,229,219,13,42,47,0,0 ; vpand 0x2f2a(%rip),%ymm3,%ymm1 # 41a0 <_sk_callback_hsw+0x49d>
+ DB 197,229,219,13,18,47,0,0 ; vpand 0x2f12(%rip),%ymm3,%ymm1 # 4180 <_sk_callback_hsw+0x485>
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 196,226,101,0,21,42,47,0,0 ; vpshufb 0x2f2a(%rip),%ymm3,%ymm2 # 41c0 <_sk_callback_hsw+0x4bd>
+ DB 196,226,101,0,21,18,47,0,0 ; vpshufb 0x2f12(%rip),%ymm3,%ymm2 # 41a0 <_sk_callback_hsw+0x4a5>
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 196,98,101,0,13,50,47,0,0 ; vpshufb 0x2f32(%rip),%ymm3,%ymm9 # 41e0 <_sk_callback_hsw+0x4dd>
+ DB 196,98,101,0,13,26,47,0,0 ; vpshufb 0x2f1a(%rip),%ymm3,%ymm9 # 41c0 <_sk_callback_hsw+0x4c5>
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 196,98,125,24,5,58,44,0,0 ; vbroadcastss 0x2c3a(%rip),%ymm8 # 3f00 <_sk_callback_hsw+0x1fd>
+ DB 196,98,125,24,5,58,44,0,0 ; vbroadcastss 0x2c3a(%rip),%ymm8 # 3ef8 <_sk_callback_hsw+0x1fd>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,193 ; mov %r8,%rcx
@@ -1226,7 +1224,7 @@ _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 126e <_sk_load_tables_hsw+0x1a>
+ DB 233,115,255,255,255 ; jmpq 1266 <_sk_load_tables_hsw+0x1a>
PUBLIC _sk_load_tables_u16_be_hsw
_sk_load_tables_u16_be_hsw LABEL PROC
@@ -1234,7 +1232,7 @@ _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 13da <_sk_load_tables_u16_be_hsw+0xdf>
+ DB 15,133,201,0,0,0 ; jne 13d2 <_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
@@ -1250,7 +1248,7 @@ _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 197,121,111,21,190,47,0,0 ; vmovdqa 0x2fbe(%rip),%xmm10 # 4320 <_sk_callback_hsw+0x61d>
+ DB 197,121,111,21,166,47,0,0 ; vmovdqa 0x2fa6(%rip),%xmm10 # 4300 <_sk_callback_hsw+0x605>
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
@@ -1272,36 +1270,36 @@ _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 196,98,125,24,5,51,43,0,0 ; vbroadcastss 0x2b33(%rip),%ymm8 # 3f04 <_sk_callback_hsw+0x201>
+ DB 196,98,125,24,5,51,43,0,0 ; vbroadcastss 0x2b33(%rip),%ymm8 # 3efc <_sk_callback_hsw+0x201>
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 1440 <_sk_load_tables_u16_be_hsw+0x145>
+ DB 116,85 ; je 1438 <_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 1440 <_sk_load_tables_u16_be_hsw+0x145>
+ DB 114,72 ; jb 1438 <_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 144d <_sk_load_tables_u16_be_hsw+0x152>
+ DB 116,72 ; je 1445 <_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 144d <_sk_load_tables_u16_be_hsw+0x152>
+ DB 114,59 ; jb 1445 <_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 132c <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,132,9,255,255,255 ; je 1324 <_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 132c <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,130,248,254,255,255 ; jb 1324 <_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 132c <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,236,254,255,255 ; jmpq 1324 <_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 132c <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,223,254,255,255 ; jmpq 1324 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,214,254,255,255 ; jmpq 132c <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,214,254,255,255 ; jmpq 1324 <_sk_load_tables_u16_be_hsw+0x31>
PUBLIC _sk_load_tables_rgb_u16_be_hsw
_sk_load_tables_rgb_u16_be_hsw LABEL PROC
@@ -1309,7 +1307,7 @@ _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 1529 <_sk_load_tables_rgb_u16_be_hsw+0xd3>
+ DB 15,133,193,0,0,0 ; jne 1521 <_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
@@ -1330,7 +1328,7 @@ _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 197,121,111,13,94,46,0,0 ; vmovdqa 0x2e5e(%rip),%xmm9 # 4330 <_sk_callback_hsw+0x62d>
+ DB 197,121,111,13,70,46,0,0 ; vmovdqa 0x2e46(%rip),%xmm9 # 4310 <_sk_callback_hsw+0x615>
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
@@ -1347,41 +1345,41 @@ _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 196,226,125,24,29,225,41,0,0 ; vbroadcastss 0x29e1(%rip),%ymm3 # 3f08 <_sk_callback_hsw+0x205>
+ DB 196,226,125,24,29,225,41,0,0 ; vbroadcastss 0x29e1(%rip),%ymm3 # 3f00 <_sk_callback_hsw+0x205>
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 1542 <_sk_load_tables_rgb_u16_be_hsw+0xec>
- DB 233,90,255,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 153a <_sk_load_tables_rgb_u16_be_hsw+0xec>
+ DB 233,90,255,255,255 ; jmpq 1494 <_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 1571 <_sk_load_tables_rgb_u16_be_hsw+0x11b>
+ DB 114,26 ; jb 1569 <_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 1576 <_sk_load_tables_rgb_u16_be_hsw+0x120>
- DB 233,43,255,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,38,255,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 156e <_sk_load_tables_rgb_u16_be_hsw+0x120>
+ DB 233,43,255,255,255 ; jmpq 1494 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,38,255,255,255 ; jmpq 1494 <_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 15a5 <_sk_load_tables_rgb_u16_be_hsw+0x14f>
+ DB 114,26 ; jb 159d <_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 15aa <_sk_load_tables_rgb_u16_be_hsw+0x154>
- DB 233,247,254,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,242,254,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 15a2 <_sk_load_tables_rgb_u16_be_hsw+0x154>
+ DB 233,247,254,255,255 ; jmpq 1494 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,242,254,255,255 ; jmpq 1494 <_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 15d3 <_sk_load_tables_rgb_u16_be_hsw+0x17d>
+ DB 114,20 ; jb 15cb <_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 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,196,254,255,255 ; jmpq 149c <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,201,254,255,255 ; jmpq 1494 <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,196,254,255,255 ; jmpq 1494 <_sk_load_tables_rgb_u16_be_hsw+0x46>
PUBLIC _sk_byte_tables_hsw
_sk_byte_tables_hsw LABEL PROC
@@ -1392,7 +1390,7 @@ _sk_byte_tables_hsw LABEL PROC
DB 65,84 ; push %r12
DB 83 ; push %rbx
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,31,41,0,0 ; vbroadcastss 0x291f(%rip),%ymm8 # 3f0c <_sk_callback_hsw+0x209>
+ DB 196,98,125,24,5,31,41,0,0 ; vbroadcastss 0x291f(%rip),%ymm8 # 3f04 <_sk_callback_hsw+0x209>
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
@@ -1429,7 +1427,7 @@ _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 196,98,125,24,13,112,40,0,0 ; vbroadcastss 0x2870(%rip),%ymm9 # 3f10 <_sk_callback_hsw+0x20d>
+ DB 196,98,125,24,13,112,40,0,0 ; vbroadcastss 0x2870(%rip),%ymm9 # 3f08 <_sk_callback_hsw+0x20d>
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
@@ -1588,7 +1586,7 @@ _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 196,98,125,24,13,169,37,0,0 ; vbroadcastss 0x25a9(%rip),%ymm9 # 3f14 <_sk_callback_hsw+0x211>
+ DB 196,98,125,24,13,169,37,0,0 ; vbroadcastss 0x25a9(%rip),%ymm9 # 3f0c <_sk_callback_hsw+0x211>
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
@@ -1741,33 +1739,33 @@ _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 196,98,125,24,37,92,35,0,0 ; vbroadcastss 0x235c(%rip),%ymm12 # 3f18 <_sk_callback_hsw+0x215>
- DB 196,98,125,24,45,87,35,0,0 ; vbroadcastss 0x2357(%rip),%ymm13 # 3f1c <_sk_callback_hsw+0x219>
+ DB 196,98,125,24,37,92,35,0,0 ; vbroadcastss 0x235c(%rip),%ymm12 # 3f10 <_sk_callback_hsw+0x215>
+ DB 196,98,125,24,45,87,35,0,0 ; vbroadcastss 0x2357(%rip),%ymm13 # 3f14 <_sk_callback_hsw+0x219>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,77,35,0,0 ; vbroadcastss 0x234d(%rip),%ymm13 # 3f20 <_sk_callback_hsw+0x21d>
+ DB 196,98,125,24,45,77,35,0,0 ; vbroadcastss 0x234d(%rip),%ymm13 # 3f18 <_sk_callback_hsw+0x21d>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,67,35,0,0 ; vbroadcastss 0x2343(%rip),%ymm13 # 3f24 <_sk_callback_hsw+0x221>
+ DB 196,98,125,24,45,67,35,0,0 ; vbroadcastss 0x2343(%rip),%ymm13 # 3f1c <_sk_callback_hsw+0x221>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,57,35,0,0 ; vbroadcastss 0x2339(%rip),%ymm11 # 3f28 <_sk_callback_hsw+0x225>
+ DB 196,98,125,24,29,57,35,0,0 ; vbroadcastss 0x2339(%rip),%ymm11 # 3f20 <_sk_callback_hsw+0x225>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,47,35,0,0 ; vbroadcastss 0x232f(%rip),%ymm12 # 3f2c <_sk_callback_hsw+0x229>
+ DB 196,98,125,24,37,47,35,0,0 ; vbroadcastss 0x232f(%rip),%ymm12 # 3f24 <_sk_callback_hsw+0x229>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,37,35,0,0 ; vbroadcastss 0x2325(%rip),%ymm12 # 3f30 <_sk_callback_hsw+0x22d>
+ DB 196,98,125,24,37,37,35,0,0 ; vbroadcastss 0x2325(%rip),%ymm12 # 3f28 <_sk_callback_hsw+0x22d>
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 196,98,125,24,29,6,35,0,0 ; vbroadcastss 0x2306(%rip),%ymm11 # 3f34 <_sk_callback_hsw+0x231>
+ DB 196,98,125,24,29,6,35,0,0 ; vbroadcastss 0x2306(%rip),%ymm11 # 3f2c <_sk_callback_hsw+0x231>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,252,34,0,0 ; vbroadcastss 0x22fc(%rip),%ymm11 # 3f38 <_sk_callback_hsw+0x235>
+ DB 196,98,125,24,29,252,34,0,0 ; vbroadcastss 0x22fc(%rip),%ymm11 # 3f30 <_sk_callback_hsw+0x235>
DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11
- DB 196,226,125,24,5,242,34,0,0 ; vbroadcastss 0x22f2(%rip),%ymm0 # 3f3c <_sk_callback_hsw+0x239>
+ DB 196,226,125,24,5,242,34,0,0 ; vbroadcastss 0x22f2(%rip),%ymm0 # 3f34 <_sk_callback_hsw+0x239>
DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,232,34,0,0 ; vbroadcastss 0x22e8(%rip),%ymm10 # 3f40 <_sk_callback_hsw+0x23d>
+ DB 196,98,125,24,21,232,34,0,0 ; vbroadcastss 0x22e8(%rip),%ymm10 # 3f38 <_sk_callback_hsw+0x23d>
DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0
DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0
- DB 196,98,125,24,21,219,34,0,0 ; vbroadcastss 0x22db(%rip),%ymm10 # 3f44 <_sk_callback_hsw+0x241>
+ DB 196,98,125,24,21,219,34,0,0 ; vbroadcastss 0x22db(%rip),%ymm10 # 3f3c <_sk_callback_hsw+0x241>
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
@@ -1775,7 +1773,7 @@ _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 196,98,125,24,5,178,34,0,0 ; vbroadcastss 0x22b2(%rip),%ymm8 # 3f48 <_sk_callback_hsw+0x245>
+ DB 196,98,125,24,5,178,34,0,0 ; vbroadcastss 0x22b2(%rip),%ymm8 # 3f40 <_sk_callback_hsw+0x245>
DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1793,33 +1791,33 @@ _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 196,98,125,24,37,106,34,0,0 ; vbroadcastss 0x226a(%rip),%ymm12 # 3f4c <_sk_callback_hsw+0x249>
- DB 196,98,125,24,45,101,34,0,0 ; vbroadcastss 0x2265(%rip),%ymm13 # 3f50 <_sk_callback_hsw+0x24d>
+ DB 196,98,125,24,37,106,34,0,0 ; vbroadcastss 0x226a(%rip),%ymm12 # 3f44 <_sk_callback_hsw+0x249>
+ DB 196,98,125,24,45,101,34,0,0 ; vbroadcastss 0x2265(%rip),%ymm13 # 3f48 <_sk_callback_hsw+0x24d>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,91,34,0,0 ; vbroadcastss 0x225b(%rip),%ymm13 # 3f54 <_sk_callback_hsw+0x251>
+ DB 196,98,125,24,45,91,34,0,0 ; vbroadcastss 0x225b(%rip),%ymm13 # 3f4c <_sk_callback_hsw+0x251>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,81,34,0,0 ; vbroadcastss 0x2251(%rip),%ymm13 # 3f58 <_sk_callback_hsw+0x255>
+ DB 196,98,125,24,45,81,34,0,0 ; vbroadcastss 0x2251(%rip),%ymm13 # 3f50 <_sk_callback_hsw+0x255>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,71,34,0,0 ; vbroadcastss 0x2247(%rip),%ymm11 # 3f5c <_sk_callback_hsw+0x259>
+ DB 196,98,125,24,29,71,34,0,0 ; vbroadcastss 0x2247(%rip),%ymm11 # 3f54 <_sk_callback_hsw+0x259>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,61,34,0,0 ; vbroadcastss 0x223d(%rip),%ymm12 # 3f60 <_sk_callback_hsw+0x25d>
+ DB 196,98,125,24,37,61,34,0,0 ; vbroadcastss 0x223d(%rip),%ymm12 # 3f58 <_sk_callback_hsw+0x25d>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,51,34,0,0 ; vbroadcastss 0x2233(%rip),%ymm12 # 3f64 <_sk_callback_hsw+0x261>
+ DB 196,98,125,24,37,51,34,0,0 ; vbroadcastss 0x2233(%rip),%ymm12 # 3f5c <_sk_callback_hsw+0x261>
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 196,98,125,24,29,20,34,0,0 ; vbroadcastss 0x2214(%rip),%ymm11 # 3f68 <_sk_callback_hsw+0x265>
+ DB 196,98,125,24,29,20,34,0,0 ; vbroadcastss 0x2214(%rip),%ymm11 # 3f60 <_sk_callback_hsw+0x265>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,10,34,0,0 ; vbroadcastss 0x220a(%rip),%ymm11 # 3f6c <_sk_callback_hsw+0x269>
+ DB 196,98,125,24,29,10,34,0,0 ; vbroadcastss 0x220a(%rip),%ymm11 # 3f64 <_sk_callback_hsw+0x269>
DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11
- DB 196,226,125,24,13,0,34,0,0 ; vbroadcastss 0x2200(%rip),%ymm1 # 3f70 <_sk_callback_hsw+0x26d>
+ DB 196,226,125,24,13,0,34,0,0 ; vbroadcastss 0x2200(%rip),%ymm1 # 3f68 <_sk_callback_hsw+0x26d>
DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,246,33,0,0 ; vbroadcastss 0x21f6(%rip),%ymm10 # 3f74 <_sk_callback_hsw+0x271>
+ DB 196,98,125,24,21,246,33,0,0 ; vbroadcastss 0x21f6(%rip),%ymm10 # 3f6c <_sk_callback_hsw+0x271>
DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1
DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1
- DB 196,98,125,24,21,233,33,0,0 ; vbroadcastss 0x21e9(%rip),%ymm10 # 3f78 <_sk_callback_hsw+0x275>
+ DB 196,98,125,24,21,233,33,0,0 ; vbroadcastss 0x21e9(%rip),%ymm10 # 3f70 <_sk_callback_hsw+0x275>
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
@@ -1827,7 +1825,7 @@ _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 196,98,125,24,5,192,33,0,0 ; vbroadcastss 0x21c0(%rip),%ymm8 # 3f7c <_sk_callback_hsw+0x279>
+ DB 196,98,125,24,5,192,33,0,0 ; vbroadcastss 0x21c0(%rip),%ymm8 # 3f74 <_sk_callback_hsw+0x279>
DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1845,33 +1843,33 @@ _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 196,98,125,24,37,120,33,0,0 ; vbroadcastss 0x2178(%rip),%ymm12 # 3f80 <_sk_callback_hsw+0x27d>
- DB 196,98,125,24,45,115,33,0,0 ; vbroadcastss 0x2173(%rip),%ymm13 # 3f84 <_sk_callback_hsw+0x281>
+ DB 196,98,125,24,37,120,33,0,0 ; vbroadcastss 0x2178(%rip),%ymm12 # 3f78 <_sk_callback_hsw+0x27d>
+ DB 196,98,125,24,45,115,33,0,0 ; vbroadcastss 0x2173(%rip),%ymm13 # 3f7c <_sk_callback_hsw+0x281>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,105,33,0,0 ; vbroadcastss 0x2169(%rip),%ymm13 # 3f88 <_sk_callback_hsw+0x285>
+ DB 196,98,125,24,45,105,33,0,0 ; vbroadcastss 0x2169(%rip),%ymm13 # 3f80 <_sk_callback_hsw+0x285>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,95,33,0,0 ; vbroadcastss 0x215f(%rip),%ymm13 # 3f8c <_sk_callback_hsw+0x289>
+ DB 196,98,125,24,45,95,33,0,0 ; vbroadcastss 0x215f(%rip),%ymm13 # 3f84 <_sk_callback_hsw+0x289>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,85,33,0,0 ; vbroadcastss 0x2155(%rip),%ymm11 # 3f90 <_sk_callback_hsw+0x28d>
+ DB 196,98,125,24,29,85,33,0,0 ; vbroadcastss 0x2155(%rip),%ymm11 # 3f88 <_sk_callback_hsw+0x28d>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,75,33,0,0 ; vbroadcastss 0x214b(%rip),%ymm12 # 3f94 <_sk_callback_hsw+0x291>
+ DB 196,98,125,24,37,75,33,0,0 ; vbroadcastss 0x214b(%rip),%ymm12 # 3f8c <_sk_callback_hsw+0x291>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,65,33,0,0 ; vbroadcastss 0x2141(%rip),%ymm12 # 3f98 <_sk_callback_hsw+0x295>
+ DB 196,98,125,24,37,65,33,0,0 ; vbroadcastss 0x2141(%rip),%ymm12 # 3f90 <_sk_callback_hsw+0x295>
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 196,98,125,24,29,34,33,0,0 ; vbroadcastss 0x2122(%rip),%ymm11 # 3f9c <_sk_callback_hsw+0x299>
+ DB 196,98,125,24,29,34,33,0,0 ; vbroadcastss 0x2122(%rip),%ymm11 # 3f94 <_sk_callback_hsw+0x299>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,24,33,0,0 ; vbroadcastss 0x2118(%rip),%ymm11 # 3fa0 <_sk_callback_hsw+0x29d>
+ DB 196,98,125,24,29,24,33,0,0 ; vbroadcastss 0x2118(%rip),%ymm11 # 3f98 <_sk_callback_hsw+0x29d>
DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11
- DB 196,226,125,24,21,14,33,0,0 ; vbroadcastss 0x210e(%rip),%ymm2 # 3fa4 <_sk_callback_hsw+0x2a1>
+ DB 196,226,125,24,21,14,33,0,0 ; vbroadcastss 0x210e(%rip),%ymm2 # 3f9c <_sk_callback_hsw+0x2a1>
DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,4,33,0,0 ; vbroadcastss 0x2104(%rip),%ymm10 # 3fa8 <_sk_callback_hsw+0x2a5>
+ DB 196,98,125,24,21,4,33,0,0 ; vbroadcastss 0x2104(%rip),%ymm10 # 3fa0 <_sk_callback_hsw+0x2a5>
DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2
DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2
- DB 196,98,125,24,21,247,32,0,0 ; vbroadcastss 0x20f7(%rip),%ymm10 # 3fac <_sk_callback_hsw+0x2a9>
+ DB 196,98,125,24,21,247,32,0,0 ; vbroadcastss 0x20f7(%rip),%ymm10 # 3fa4 <_sk_callback_hsw+0x2a9>
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
@@ -1879,7 +1877,7 @@ _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 196,98,125,24,5,206,32,0,0 ; vbroadcastss 0x20ce(%rip),%ymm8 # 3fb0 <_sk_callback_hsw+0x2ad>
+ DB 196,98,125,24,5,206,32,0,0 ; vbroadcastss 0x20ce(%rip),%ymm8 # 3fa8 <_sk_callback_hsw+0x2ad>
DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1897,33 +1895,33 @@ _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 196,98,125,24,37,134,32,0,0 ; vbroadcastss 0x2086(%rip),%ymm12 # 3fb4 <_sk_callback_hsw+0x2b1>
- DB 196,98,125,24,45,129,32,0,0 ; vbroadcastss 0x2081(%rip),%ymm13 # 3fb8 <_sk_callback_hsw+0x2b5>
+ DB 196,98,125,24,37,134,32,0,0 ; vbroadcastss 0x2086(%rip),%ymm12 # 3fac <_sk_callback_hsw+0x2b1>
+ DB 196,98,125,24,45,129,32,0,0 ; vbroadcastss 0x2081(%rip),%ymm13 # 3fb0 <_sk_callback_hsw+0x2b5>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,119,32,0,0 ; vbroadcastss 0x2077(%rip),%ymm13 # 3fbc <_sk_callback_hsw+0x2b9>
+ DB 196,98,125,24,45,119,32,0,0 ; vbroadcastss 0x2077(%rip),%ymm13 # 3fb4 <_sk_callback_hsw+0x2b9>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,109,32,0,0 ; vbroadcastss 0x206d(%rip),%ymm13 # 3fc0 <_sk_callback_hsw+0x2bd>
+ DB 196,98,125,24,45,109,32,0,0 ; vbroadcastss 0x206d(%rip),%ymm13 # 3fb8 <_sk_callback_hsw+0x2bd>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,99,32,0,0 ; vbroadcastss 0x2063(%rip),%ymm11 # 3fc4 <_sk_callback_hsw+0x2c1>
+ DB 196,98,125,24,29,99,32,0,0 ; vbroadcastss 0x2063(%rip),%ymm11 # 3fbc <_sk_callback_hsw+0x2c1>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,89,32,0,0 ; vbroadcastss 0x2059(%rip),%ymm12 # 3fc8 <_sk_callback_hsw+0x2c5>
+ DB 196,98,125,24,37,89,32,0,0 ; vbroadcastss 0x2059(%rip),%ymm12 # 3fc0 <_sk_callback_hsw+0x2c5>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,79,32,0,0 ; vbroadcastss 0x204f(%rip),%ymm12 # 3fcc <_sk_callback_hsw+0x2c9>
+ DB 196,98,125,24,37,79,32,0,0 ; vbroadcastss 0x204f(%rip),%ymm12 # 3fc4 <_sk_callback_hsw+0x2c9>
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 196,98,125,24,29,48,32,0,0 ; vbroadcastss 0x2030(%rip),%ymm11 # 3fd0 <_sk_callback_hsw+0x2cd>
+ DB 196,98,125,24,29,48,32,0,0 ; vbroadcastss 0x2030(%rip),%ymm11 # 3fc8 <_sk_callback_hsw+0x2cd>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,38,32,0,0 ; vbroadcastss 0x2026(%rip),%ymm11 # 3fd4 <_sk_callback_hsw+0x2d1>
+ DB 196,98,125,24,29,38,32,0,0 ; vbroadcastss 0x2026(%rip),%ymm11 # 3fcc <_sk_callback_hsw+0x2d1>
DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11
- DB 196,226,125,24,29,28,32,0,0 ; vbroadcastss 0x201c(%rip),%ymm3 # 3fd8 <_sk_callback_hsw+0x2d5>
+ DB 196,226,125,24,29,28,32,0,0 ; vbroadcastss 0x201c(%rip),%ymm3 # 3fd0 <_sk_callback_hsw+0x2d5>
DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,18,32,0,0 ; vbroadcastss 0x2012(%rip),%ymm10 # 3fdc <_sk_callback_hsw+0x2d9>
+ DB 196,98,125,24,21,18,32,0,0 ; vbroadcastss 0x2012(%rip),%ymm10 # 3fd4 <_sk_callback_hsw+0x2d9>
DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3
DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3
- DB 196,98,125,24,21,5,32,0,0 ; vbroadcastss 0x2005(%rip),%ymm10 # 3fe0 <_sk_callback_hsw+0x2dd>
+ DB 196,98,125,24,21,5,32,0,0 ; vbroadcastss 0x2005(%rip),%ymm10 # 3fd8 <_sk_callback_hsw+0x2dd>
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
@@ -1931,33 +1929,33 @@ _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 196,98,125,24,5,220,31,0,0 ; vbroadcastss 0x1fdc(%rip),%ymm8 # 3fe4 <_sk_callback_hsw+0x2e1>
+ DB 196,98,125,24,5,220,31,0,0 ; vbroadcastss 0x1fdc(%rip),%ymm8 # 3fdc <_sk_callback_hsw+0x2e1>
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 196,98,125,24,5,206,31,0,0 ; vbroadcastss 0x1fce(%rip),%ymm8 # 3fe8 <_sk_callback_hsw+0x2e5>
- DB 196,98,125,24,13,201,31,0,0 ; vbroadcastss 0x1fc9(%rip),%ymm9 # 3fec <_sk_callback_hsw+0x2e9>
- DB 196,98,125,24,21,196,31,0,0 ; vbroadcastss 0x1fc4(%rip),%ymm10 # 3ff0 <_sk_callback_hsw+0x2ed>
+ DB 196,98,125,24,5,206,31,0,0 ; vbroadcastss 0x1fce(%rip),%ymm8 # 3fe0 <_sk_callback_hsw+0x2e5>
+ DB 196,98,125,24,13,201,31,0,0 ; vbroadcastss 0x1fc9(%rip),%ymm9 # 3fe4 <_sk_callback_hsw+0x2e9>
+ DB 196,98,125,24,21,196,31,0,0 ; vbroadcastss 0x1fc4(%rip),%ymm10 # 3fe8 <_sk_callback_hsw+0x2ed>
DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1
DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2
- DB 196,98,125,24,13,181,31,0,0 ; vbroadcastss 0x1fb5(%rip),%ymm9 # 3ff4 <_sk_callback_hsw+0x2f1>
+ DB 196,98,125,24,13,181,31,0,0 ; vbroadcastss 0x1fb5(%rip),%ymm9 # 3fec <_sk_callback_hsw+0x2f1>
DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9
- DB 196,226,125,24,5,171,31,0,0 ; vbroadcastss 0x1fab(%rip),%ymm0 # 3ff8 <_sk_callback_hsw+0x2f5>
+ DB 196,226,125,24,5,171,31,0,0 ; vbroadcastss 0x1fab(%rip),%ymm0 # 3ff0 <_sk_callback_hsw+0x2f5>
DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0
- DB 196,98,125,24,5,162,31,0,0 ; vbroadcastss 0x1fa2(%rip),%ymm8 # 3ffc <_sk_callback_hsw+0x2f9>
+ DB 196,98,125,24,5,162,31,0,0 ; vbroadcastss 0x1fa2(%rip),%ymm8 # 3ff4 <_sk_callback_hsw+0x2f9>
DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8
- DB 196,98,125,24,13,152,31,0,0 ; vbroadcastss 0x1f98(%rip),%ymm9 # 4000 <_sk_callback_hsw+0x2fd>
+ DB 196,98,125,24,13,152,31,0,0 ; vbroadcastss 0x1f98(%rip),%ymm9 # 3ff8 <_sk_callback_hsw+0x2fd>
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 196,226,125,24,21,133,31,0,0 ; vbroadcastss 0x1f85(%rip),%ymm2 # 4004 <_sk_callback_hsw+0x301>
+ DB 196,226,125,24,21,133,31,0,0 ; vbroadcastss 0x1f85(%rip),%ymm2 # 3ffc <_sk_callback_hsw+0x301>
DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10
- DB 196,98,125,24,29,123,31,0,0 ; vbroadcastss 0x1f7b(%rip),%ymm11 # 4008 <_sk_callback_hsw+0x305>
+ DB 196,98,125,24,29,123,31,0,0 ; vbroadcastss 0x1f7b(%rip),%ymm11 # 4000 <_sk_callback_hsw+0x305>
DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8
- DB 196,98,125,24,37,113,31,0,0 ; vbroadcastss 0x1f71(%rip),%ymm12 # 400c <_sk_callback_hsw+0x309>
+ DB 196,98,125,24,37,113,31,0,0 ; vbroadcastss 0x1f71(%rip),%ymm12 # 4004 <_sk_callback_hsw+0x309>
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
@@ -1972,9 +1970,9 @@ _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 196,226,125,24,5,38,31,0,0 ; vbroadcastss 0x1f26(%rip),%ymm0 # 4010 <_sk_callback_hsw+0x30d>
+ DB 196,226,125,24,5,38,31,0,0 ; vbroadcastss 0x1f26(%rip),%ymm0 # 4008 <_sk_callback_hsw+0x30d>
DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
- DB 196,98,125,24,5,29,31,0,0 ; vbroadcastss 0x1f1d(%rip),%ymm8 # 4014 <_sk_callback_hsw+0x311>
+ DB 196,98,125,24,5,29,31,0,0 ; vbroadcastss 0x1f1d(%rip),%ymm8 # 400c <_sk_callback_hsw+0x311>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1986,11 +1984,11 @@ _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 213d <_sk_load_a8_hsw+0x3d>
+ DB 117,45 ; jne 2135 <_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 196,226,125,24,13,242,30,0,0 ; vbroadcastss 0x1ef2(%rip),%ymm1 # 4018 <_sk_callback_hsw+0x315>
+ DB 196,226,125,24,13,242,30,0,0 ; vbroadcastss 0x1ef2(%rip),%ymm1 # 4010 <_sk_callback_hsw+0x315>
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
@@ -2007,9 +2005,9 @@ _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 2145 <_sk_load_a8_hsw+0x45>
+ DB 117,234 ; jne 213d <_sk_load_a8_hsw+0x45>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,178 ; jmp 2114 <_sk_load_a8_hsw+0x14>
+ DB 235,178 ; jmp 210c <_sk_load_a8_hsw+0x14>
PUBLIC _sk_gather_a8_hsw
_sk_gather_a8_hsw LABEL PROC
@@ -2053,7 +2051,7 @@ _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 196,226,125,24,13,253,29,0,0 ; vbroadcastss 0x1dfd(%rip),%ymm1 # 401c <_sk_callback_hsw+0x319>
+ DB 196,226,125,24,13,253,29,0,0 ; vbroadcastss 0x1dfd(%rip),%ymm1 # 4014 <_sk_callback_hsw+0x319>
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
@@ -2069,14 +2067,14 @@ 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 196,98,125,24,5,216,29,0,0 ; vbroadcastss 0x1dd8(%rip),%ymm8 # 4020 <_sk_callback_hsw+0x31d>
+ DB 196,98,125,24,5,216,29,0,0 ; vbroadcastss 0x1dd8(%rip),%ymm8 # 4018 <_sk_callback_hsw+0x31d>
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 2271 <_sk_store_a8_hsw+0x37>
+ DB 117,10 ; jne 2269 <_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
@@ -2084,10 +2082,10 @@ _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 226d <_sk_store_a8_hsw+0x33>
+ DB 119,236 ; ja 2265 <_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 # 22d4 <_sk_store_a8_hsw+0x9a>
+ DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 22cc <_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
@@ -2098,7 +2096,7 @@ _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 226d <_sk_store_a8_hsw+0x33>
+ DB 235,154 ; jmp 2265 <_sk_store_a8_hsw+0x33>
DB 144 ; nop
DB 246,255 ; idiv %bh
DB 255 ; (bad)
@@ -2130,14 +2128,14 @@ _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 2332 <_sk_load_g8_hsw+0x42>
+ DB 117,50 ; jne 232a <_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 196,226,125,24,13,14,29,0,0 ; vbroadcastss 0x1d0e(%rip),%ymm1 # 4024 <_sk_callback_hsw+0x321>
+ DB 196,226,125,24,13,14,29,0,0 ; vbroadcastss 0x1d0e(%rip),%ymm1 # 401c <_sk_callback_hsw+0x321>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,3,29,0,0 ; vbroadcastss 0x1d03(%rip),%ymm3 # 4028 <_sk_callback_hsw+0x325>
+ DB 196,226,125,24,29,3,29,0,0 ; vbroadcastss 0x1d03(%rip),%ymm3 # 4020 <_sk_callback_hsw+0x325>
DB 76,137,193 ; mov %r8,%rcx
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
@@ -2151,9 +2149,9 @@ _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 233a <_sk_load_g8_hsw+0x4a>
+ DB 117,234 ; jne 2332 <_sk_load_g8_hsw+0x4a>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,173 ; jmp 2304 <_sk_load_g8_hsw+0x14>
+ DB 235,173 ; jmp 22fc <_sk_load_g8_hsw+0x14>
PUBLIC _sk_gather_g8_hsw
_sk_gather_g8_hsw LABEL PROC
@@ -2197,10 +2195,10 @@ _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 196,226,125,24,13,24,28,0,0 ; vbroadcastss 0x1c18(%rip),%ymm1 # 402c <_sk_callback_hsw+0x329>
+ DB 196,226,125,24,13,24,28,0,0 ; vbroadcastss 0x1c18(%rip),%ymm1 # 4024 <_sk_callback_hsw+0x329>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,13,28,0,0 ; vbroadcastss 0x1c0d(%rip),%ymm3 # 4030 <_sk_callback_hsw+0x32d>
+ DB 196,226,125,24,29,13,28,0,0 ; vbroadcastss 0x1c0d(%rip),%ymm3 # 4028 <_sk_callback_hsw+0x32d>
DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 91 ; pop %rbx
@@ -2214,9 +2212,9 @@ _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 2443 <_sk_gather_i8_hsw+0xf>
+ DB 116,5 ; je 243b <_sk_gather_i8_hsw+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 2445 <_sk_gather_i8_hsw+0x11>
+ DB 235,2 ; jmp 243d <_sk_gather_i8_hsw+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -2254,14 +2252,14 @@ _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 197,229,219,5,13,29,0,0 ; vpand 0x1d0d(%rip),%ymm3,%ymm0 # 4200 <_sk_callback_hsw+0x4fd>
+ DB 197,229,219,5,245,28,0,0 ; vpand 0x1cf5(%rip),%ymm3,%ymm0 # 41e0 <_sk_callback_hsw+0x4e5>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,52,27,0,0 ; vbroadcastss 0x1b34(%rip),%ymm8 # 4034 <_sk_callback_hsw+0x331>
+ DB 196,98,125,24,5,52,27,0,0 ; vbroadcastss 0x1b34(%rip),%ymm8 # 402c <_sk_callback_hsw+0x331>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,18,29,0,0 ; vpshufb 0x1d12(%rip),%ymm3,%ymm1 # 4220 <_sk_callback_hsw+0x51d>
+ DB 196,226,101,0,13,250,28,0,0 ; vpshufb 0x1cfa(%rip),%ymm3,%ymm1 # 4200 <_sk_callback_hsw+0x505>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,32,29,0,0 ; vpshufb 0x1d20(%rip),%ymm3,%ymm2 # 4240 <_sk_callback_hsw+0x53d>
+ DB 196,226,101,0,21,8,29,0,0 ; vpshufb 0x1d08(%rip),%ymm3,%ymm2 # 4220 <_sk_callback_hsw+0x525>
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
@@ -2280,35 +2278,35 @@ _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 25c0 <_sk_load_565_hsw+0x7c>
+ DB 117,114 ; jne 25b8 <_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 196,226,125,88,5,214,26,0,0 ; vpbroadcastd 0x1ad6(%rip),%ymm0 # 4038 <_sk_callback_hsw+0x335>
+ DB 196,226,125,88,5,214,26,0,0 ; vpbroadcastd 0x1ad6(%rip),%ymm0 # 4030 <_sk_callback_hsw+0x335>
DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,201,26,0,0 ; vbroadcastss 0x1ac9(%rip),%ymm1 # 403c <_sk_callback_hsw+0x339>
+ DB 196,226,125,24,13,201,26,0,0 ; vbroadcastss 0x1ac9(%rip),%ymm1 # 4034 <_sk_callback_hsw+0x339>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,192,26,0,0 ; vpbroadcastd 0x1ac0(%rip),%ymm1 # 4040 <_sk_callback_hsw+0x33d>
+ DB 196,226,125,88,13,192,26,0,0 ; vpbroadcastd 0x1ac0(%rip),%ymm1 # 4038 <_sk_callback_hsw+0x33d>
DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,179,26,0,0 ; vbroadcastss 0x1ab3(%rip),%ymm3 # 4044 <_sk_callback_hsw+0x341>
+ DB 196,226,125,24,29,179,26,0,0 ; vbroadcastss 0x1ab3(%rip),%ymm3 # 403c <_sk_callback_hsw+0x341>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,88,29,170,26,0,0 ; vpbroadcastd 0x1aaa(%rip),%ymm3 # 4048 <_sk_callback_hsw+0x345>
+ DB 196,226,125,88,29,170,26,0,0 ; vpbroadcastd 0x1aaa(%rip),%ymm3 # 4040 <_sk_callback_hsw+0x345>
DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,157,26,0,0 ; vbroadcastss 0x1a9d(%rip),%ymm3 # 404c <_sk_callback_hsw+0x349>
+ DB 196,226,125,24,29,157,26,0,0 ; vbroadcastss 0x1a9d(%rip),%ymm3 # 4044 <_sk_callback_hsw+0x349>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,146,26,0,0 ; vbroadcastss 0x1a92(%rip),%ymm3 # 4050 <_sk_callback_hsw+0x34d>
+ DB 196,226,125,24,29,146,26,0,0 ; vbroadcastss 0x1a92(%rip),%ymm3 # 4048 <_sk_callback_hsw+0x34d>
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 2554 <_sk_load_565_hsw+0x10>
+ DB 119,128 ; ja 254c <_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 # 2628 <_sk_load_565_hsw+0xe4>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2620 <_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
@@ -2320,7 +2318,7 @@ _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 2554 <_sk_load_565_hsw+0x10>
+ DB 233,44,255,255,255 ; jmpq 254c <_sk_load_565_hsw+0x10>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2388,23 +2386,23 @@ _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 196,226,125,88,5,85,25,0,0 ; vpbroadcastd 0x1955(%rip),%ymm0 # 4054 <_sk_callback_hsw+0x351>
+ DB 196,226,125,88,5,85,25,0,0 ; vpbroadcastd 0x1955(%rip),%ymm0 # 404c <_sk_callback_hsw+0x351>
DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,72,25,0,0 ; vbroadcastss 0x1948(%rip),%ymm1 # 4058 <_sk_callback_hsw+0x355>
+ DB 196,226,125,24,13,72,25,0,0 ; vbroadcastss 0x1948(%rip),%ymm1 # 4050 <_sk_callback_hsw+0x355>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,63,25,0,0 ; vpbroadcastd 0x193f(%rip),%ymm1 # 405c <_sk_callback_hsw+0x359>
+ DB 196,226,125,88,13,63,25,0,0 ; vpbroadcastd 0x193f(%rip),%ymm1 # 4054 <_sk_callback_hsw+0x359>
DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,29,50,25,0,0 ; vbroadcastss 0x1932(%rip),%ymm3 # 4060 <_sk_callback_hsw+0x35d>
+ DB 196,226,125,24,29,50,25,0,0 ; vbroadcastss 0x1932(%rip),%ymm3 # 4058 <_sk_callback_hsw+0x35d>
DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1
- DB 196,226,125,88,29,41,25,0,0 ; vpbroadcastd 0x1929(%rip),%ymm3 # 4064 <_sk_callback_hsw+0x361>
+ DB 196,226,125,88,29,41,25,0,0 ; vpbroadcastd 0x1929(%rip),%ymm3 # 405c <_sk_callback_hsw+0x361>
DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,226,125,24,29,28,25,0,0 ; vbroadcastss 0x191c(%rip),%ymm3 # 4068 <_sk_callback_hsw+0x365>
+ DB 196,226,125,24,29,28,25,0,0 ; vbroadcastss 0x191c(%rip),%ymm3 # 4060 <_sk_callback_hsw+0x365>
DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,29,17,25,0,0 ; vbroadcastss 0x1911(%rip),%ymm3 # 406c <_sk_callback_hsw+0x369>
+ DB 196,226,125,24,29,17,25,0,0 ; vbroadcastss 0x1911(%rip),%ymm3 # 4064 <_sk_callback_hsw+0x369>
DB 91 ; pop %rbx
DB 65,92 ; pop %r12
DB 65,94 ; pop %r14
@@ -2415,11 +2413,11 @@ 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 196,98,125,24,5,254,24,0,0 ; vbroadcastss 0x18fe(%rip),%ymm8 # 4070 <_sk_callback_hsw+0x36d>
+ DB 196,98,125,24,5,254,24,0,0 ; vbroadcastss 0x18fe(%rip),%ymm8 # 4068 <_sk_callback_hsw+0x36d>
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 196,98,125,24,21,233,24,0,0 ; vbroadcastss 0x18e9(%rip),%ymm10 # 4074 <_sk_callback_hsw+0x371>
+ DB 196,98,125,24,21,233,24,0,0 ; vbroadcastss 0x18e9(%rip),%ymm10 # 406c <_sk_callback_hsw+0x371>
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
@@ -2430,7 +2428,7 @@ _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 27c9 <_sk_store_565_hsw+0x65>
+ DB 117,10 ; jne 27c1 <_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
@@ -2438,9 +2436,9 @@ _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 27c5 <_sk_store_565_hsw+0x61>
+ DB 119,236 ; ja 27bd <_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 # 2828 <_sk_store_565_hsw+0xc4>
+ DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2820 <_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
@@ -2451,7 +2449,7 @@ _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 27c5 <_sk_store_565_hsw+0x61>
+ DB 235,159 ; jmp 27bd <_sk_store_565_hsw+0x61>
DB 102,144 ; xchg %ax,%ax
DB 245 ; cmc
DB 255 ; (bad)
@@ -2482,28 +2480,28 @@ _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 28dc <_sk_load_4444_hsw+0x98>
+ DB 15,133,138,0,0,0 ; jne 28d4 <_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 196,226,125,88,5,18,24,0,0 ; vpbroadcastd 0x1812(%rip),%ymm0 # 4078 <_sk_callback_hsw+0x375>
+ DB 196,226,125,88,5,18,24,0,0 ; vpbroadcastd 0x1812(%rip),%ymm0 # 4070 <_sk_callback_hsw+0x375>
DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,5,24,0,0 ; vbroadcastss 0x1805(%rip),%ymm1 # 407c <_sk_callback_hsw+0x379>
+ DB 196,226,125,24,13,5,24,0,0 ; vbroadcastss 0x1805(%rip),%ymm1 # 4074 <_sk_callback_hsw+0x379>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,252,23,0,0 ; vpbroadcastd 0x17fc(%rip),%ymm1 # 4080 <_sk_callback_hsw+0x37d>
+ DB 196,226,125,88,13,252,23,0,0 ; vpbroadcastd 0x17fc(%rip),%ymm1 # 4078 <_sk_callback_hsw+0x37d>
DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,239,23,0,0 ; vbroadcastss 0x17ef(%rip),%ymm2 # 4084 <_sk_callback_hsw+0x381>
+ DB 196,226,125,24,21,239,23,0,0 ; vbroadcastss 0x17ef(%rip),%ymm2 # 407c <_sk_callback_hsw+0x381>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,88,21,230,23,0,0 ; vpbroadcastd 0x17e6(%rip),%ymm2 # 4088 <_sk_callback_hsw+0x385>
+ DB 196,226,125,88,21,230,23,0,0 ; vpbroadcastd 0x17e6(%rip),%ymm2 # 4080 <_sk_callback_hsw+0x385>
DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,217,23,0,0 ; vbroadcastss 0x17d9(%rip),%ymm8 # 408c <_sk_callback_hsw+0x389>
+ DB 196,98,125,24,5,217,23,0,0 ; vbroadcastss 0x17d9(%rip),%ymm8 # 4084 <_sk_callback_hsw+0x389>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,88,5,207,23,0,0 ; vpbroadcastd 0x17cf(%rip),%ymm8 # 4090 <_sk_callback_hsw+0x38d>
+ DB 196,98,125,88,5,207,23,0,0 ; vpbroadcastd 0x17cf(%rip),%ymm8 # 4088 <_sk_callback_hsw+0x38d>
DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,193,23,0,0 ; vbroadcastss 0x17c1(%rip),%ymm8 # 4094 <_sk_callback_hsw+0x391>
+ DB 196,98,125,24,5,193,23,0,0 ; vbroadcastss 0x17c1(%rip),%ymm8 # 408c <_sk_callback_hsw+0x391>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -2512,9 +2510,9 @@ _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 2858 <_sk_load_4444_hsw+0x14>
+ DB 15,135,100,255,255,255 ; ja 2850 <_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 # 2948 <_sk_load_4444_hsw+0x104>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 2940 <_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
@@ -2526,7 +2524,7 @@ _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 2858 <_sk_load_4444_hsw+0x14>
+ DB 233,16,255,255,255 ; jmpq 2850 <_sk_load_4444_hsw+0x14>
DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2594,25 +2592,25 @@ _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 196,226,125,88,5,121,22,0,0 ; vpbroadcastd 0x1679(%rip),%ymm0 # 4098 <_sk_callback_hsw+0x395>
+ DB 196,226,125,88,5,121,22,0,0 ; vpbroadcastd 0x1679(%rip),%ymm0 # 4090 <_sk_callback_hsw+0x395>
DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,108,22,0,0 ; vbroadcastss 0x166c(%rip),%ymm1 # 409c <_sk_callback_hsw+0x399>
+ DB 196,226,125,24,13,108,22,0,0 ; vbroadcastss 0x166c(%rip),%ymm1 # 4094 <_sk_callback_hsw+0x399>
DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
- DB 196,226,125,88,13,99,22,0,0 ; vpbroadcastd 0x1663(%rip),%ymm1 # 40a0 <_sk_callback_hsw+0x39d>
+ DB 196,226,125,88,13,99,22,0,0 ; vpbroadcastd 0x1663(%rip),%ymm1 # 4098 <_sk_callback_hsw+0x39d>
DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
- DB 196,226,125,24,21,86,22,0,0 ; vbroadcastss 0x1656(%rip),%ymm2 # 40a4 <_sk_callback_hsw+0x3a1>
+ DB 196,226,125,24,21,86,22,0,0 ; vbroadcastss 0x1656(%rip),%ymm2 # 409c <_sk_callback_hsw+0x3a1>
DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
- DB 196,226,125,88,21,77,22,0,0 ; vpbroadcastd 0x164d(%rip),%ymm2 # 40a8 <_sk_callback_hsw+0x3a5>
+ DB 196,226,125,88,21,77,22,0,0 ; vpbroadcastd 0x164d(%rip),%ymm2 # 40a0 <_sk_callback_hsw+0x3a5>
DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
- DB 196,98,125,24,5,64,22,0,0 ; vbroadcastss 0x1640(%rip),%ymm8 # 40ac <_sk_callback_hsw+0x3a9>
+ DB 196,98,125,24,5,64,22,0,0 ; vbroadcastss 0x1640(%rip),%ymm8 # 40a4 <_sk_callback_hsw+0x3a9>
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
- DB 196,98,125,88,5,54,22,0,0 ; vpbroadcastd 0x1636(%rip),%ymm8 # 40b0 <_sk_callback_hsw+0x3ad>
+ DB 196,98,125,88,5,54,22,0,0 ; vpbroadcastd 0x1636(%rip),%ymm8 # 40a8 <_sk_callback_hsw+0x3ad>
DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,5,40,22,0,0 ; vbroadcastss 0x1628(%rip),%ymm8 # 40b4 <_sk_callback_hsw+0x3b1>
+ DB 196,98,125,24,5,40,22,0,0 ; vbroadcastss 0x1628(%rip),%ymm8 # 40ac <_sk_callback_hsw+0x3b1>
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 91 ; pop %rbx
@@ -2625,7 +2623,7 @@ 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 196,98,125,24,5,14,22,0,0 ; vbroadcastss 0x160e(%rip),%ymm8 # 40b8 <_sk_callback_hsw+0x3b5>
+ DB 196,98,125,24,5,14,22,0,0 ; vbroadcastss 0x160e(%rip),%ymm8 # 40b0 <_sk_callback_hsw+0x3b5>
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
@@ -2643,7 +2641,7 @@ _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 2b0d <_sk_store_4444_hsw+0x71>
+ DB 117,10 ; jne 2b05 <_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
@@ -2651,9 +2649,9 @@ _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 2b09 <_sk_store_4444_hsw+0x6d>
+ DB 119,236 ; ja 2b01 <_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 # 2b6c <_sk_store_4444_hsw+0xd0>
+ DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 2b64 <_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
@@ -2664,7 +2662,7 @@ _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 2b09 <_sk_store_4444_hsw+0x6d>
+ DB 235,159 ; jmp 2b01 <_sk_store_4444_hsw+0x6d>
DB 102,144 ; xchg %ax,%ax
DB 245 ; cmc
DB 255 ; (bad)
@@ -2697,16 +2695,16 @@ _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 2bf5 <_sk_load_8888_hsw+0x6d>
+ DB 117,88 ; jne 2bed <_sk_load_8888_hsw+0x6d>
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
- DB 197,229,219,5,182,22,0,0 ; vpand 0x16b6(%rip),%ymm3,%ymm0 # 4260 <_sk_callback_hsw+0x55d>
+ DB 197,229,219,5,158,22,0,0 ; vpand 0x169e(%rip),%ymm3,%ymm0 # 4240 <_sk_callback_hsw+0x545>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,5,21,0,0 ; vbroadcastss 0x1505(%rip),%ymm8 # 40bc <_sk_callback_hsw+0x3b9>
+ DB 196,98,125,24,5,5,21,0,0 ; vbroadcastss 0x1505(%rip),%ymm8 # 40b4 <_sk_callback_hsw+0x3b9>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,187,22,0,0 ; vpshufb 0x16bb(%rip),%ymm3,%ymm1 # 4280 <_sk_callback_hsw+0x57d>
+ DB 196,226,101,0,13,163,22,0,0 ; vpshufb 0x16a3(%rip),%ymm3,%ymm1 # 4260 <_sk_callback_hsw+0x565>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,201,22,0,0 ; vpshufb 0x16c9(%rip),%ymm3,%ymm2 # 42a0 <_sk_callback_hsw+0x59d>
+ DB 196,226,101,0,21,177,22,0,0 ; vpshufb 0x16b1(%rip),%ymm3,%ymm2 # 4280 <_sk_callback_hsw+0x585>
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
@@ -2723,7 +2721,7 @@ _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 2ba2 <_sk_load_8888_hsw+0x1a>
+ DB 235,135 ; jmp 2b9a <_sk_load_8888_hsw+0x1a>
PUBLIC _sk_gather_8888_hsw
_sk_gather_8888_hsw LABEL PROC
@@ -2736,14 +2734,14 @@ _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 197,229,219,5,119,22,0,0 ; vpand 0x1677(%rip),%ymm3,%ymm0 # 42c0 <_sk_callback_hsw+0x5bd>
+ DB 197,229,219,5,95,22,0,0 ; vpand 0x165f(%rip),%ymm3,%ymm0 # 42a0 <_sk_callback_hsw+0x5a5>
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,98,125,24,5,106,20,0,0 ; vbroadcastss 0x146a(%rip),%ymm8 # 40c0 <_sk_callback_hsw+0x3bd>
+ DB 196,98,125,24,5,106,20,0,0 ; vbroadcastss 0x146a(%rip),%ymm8 # 40b8 <_sk_callback_hsw+0x3bd>
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,226,101,0,13,124,22,0,0 ; vpshufb 0x167c(%rip),%ymm3,%ymm1 # 42e0 <_sk_callback_hsw+0x5dd>
+ DB 196,226,101,0,13,100,22,0,0 ; vpshufb 0x1664(%rip),%ymm3,%ymm1 # 42c0 <_sk_callback_hsw+0x5c5>
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,226,101,0,21,138,22,0,0 ; vpshufb 0x168a(%rip),%ymm3,%ymm2 # 4300 <_sk_callback_hsw+0x5fd>
+ DB 196,226,101,0,21,114,22,0,0 ; vpshufb 0x1672(%rip),%ymm3,%ymm2 # 42e0 <_sk_callback_hsw+0x5e5>
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
@@ -2758,7 +2756,7 @@ _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 196,98,125,24,5,26,20,0,0 ; vbroadcastss 0x141a(%rip),%ymm8 # 40c4 <_sk_callback_hsw+0x3c1>
+ DB 196,98,125,24,5,26,20,0,0 ; vbroadcastss 0x141a(%rip),%ymm8 # 40bc <_sk_callback_hsw+0x3c1>
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
@@ -2774,7 +2772,7 @@ _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 2d04 <_sk_store_8888_hsw+0x73>
+ DB 117,12 ; jne 2cfc <_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
@@ -2787,14 +2785,14 @@ _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 2cfd <_sk_store_8888_hsw+0x6c>
+ DB 235,211 ; jmp 2cf5 <_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 2d95 <_sk_load_f16_hsw+0x6b>
+ DB 117,97 ; jne 2d8d <_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
@@ -2820,29 +2818,29 @@ _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 2df4 <_sk_load_f16_hsw+0xca>
+ DB 116,79 ; je 2dec <_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 2df4 <_sk_load_f16_hsw+0xca>
+ DB 114,67 ; jb 2dec <_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 2e01 <_sk_load_f16_hsw+0xd7>
+ DB 116,68 ; je 2df9 <_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 2e01 <_sk_load_f16_hsw+0xd7>
+ DB 114,56 ; jb 2df9 <_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 2d4b <_sk_load_f16_hsw+0x21>
+ DB 15,132,114,255,255,255 ; je 2d43 <_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 2d4b <_sk_load_f16_hsw+0x21>
+ DB 15,130,98,255,255,255 ; jb 2d43 <_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 2d4b <_sk_load_f16_hsw+0x21>
+ DB 233,87,255,255,255 ; jmpq 2d43 <_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 2d4b <_sk_load_f16_hsw+0x21>
+ DB 233,74,255,255,255 ; jmpq 2d43 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,65,255,255,255 ; jmpq 2d4b <_sk_load_f16_hsw+0x21>
+ DB 233,65,255,255,255 ; jmpq 2d43 <_sk_load_f16_hsw+0x21>
PUBLIC _sk_gather_f16_hsw
_sk_gather_f16_hsw LABEL PROC
@@ -2896,7 +2894,7 @@ _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 2ef9 <_sk_store_f16_hsw+0x65>
+ DB 117,27 ; jne 2ef1 <_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)
@@ -2905,22 +2903,22 @@ _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 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 116,241 ; je 2eed <_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 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 114,229 ; jb 2eed <_sk_store_f16_hsw+0x61>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,221 ; je 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 116,221 ; je 2eed <_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 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 114,209 ; jb 2eed <_sk_store_f16_hsw+0x61>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,201 ; je 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 116,201 ; je 2eed <_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 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 114,189 ; jb 2eed <_sk_store_f16_hsw+0x61>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,181 ; jmp 2ef5 <_sk_store_f16_hsw+0x61>
+ DB 235,181 ; jmp 2eed <_sk_store_f16_hsw+0x61>
PUBLIC _sk_load_u16_be_hsw
_sk_load_u16_be_hsw LABEL PROC
@@ -2928,7 +2926,7 @@ _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 3022 <_sk_load_u16_be_hsw+0xe2>
+ DB 15,133,204,0,0,0 ; jne 301a <_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
@@ -2947,7 +2945,7 @@ _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 196,98,125,24,21,17,17,0,0 ; vbroadcastss 0x1111(%rip),%ymm10 # 40c8 <_sk_callback_hsw+0x3c5>
+ DB 196,98,125,24,21,17,17,0,0 ; vbroadcastss 0x1111(%rip),%ymm10 # 40c0 <_sk_callback_hsw+0x3c5>
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
@@ -2975,29 +2973,29 @@ _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 3088 <_sk_load_u16_be_hsw+0x148>
+ DB 116,85 ; je 3080 <_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 3088 <_sk_load_u16_be_hsw+0x148>
+ DB 114,72 ; jb 3080 <_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 3095 <_sk_load_u16_be_hsw+0x155>
+ DB 116,72 ; je 308d <_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 3095 <_sk_load_u16_be_hsw+0x155>
+ DB 114,59 ; jb 308d <_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 2f71 <_sk_load_u16_be_hsw+0x31>
+ DB 15,132,6,255,255,255 ; je 2f69 <_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 2f71 <_sk_load_u16_be_hsw+0x31>
+ DB 15,130,245,254,255,255 ; jb 2f69 <_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 2f71 <_sk_load_u16_be_hsw+0x31>
+ DB 233,233,254,255,255 ; jmpq 2f69 <_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 2f71 <_sk_load_u16_be_hsw+0x31>
+ DB 233,220,254,255,255 ; jmpq 2f69 <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,211,254,255,255 ; jmpq 2f71 <_sk_load_u16_be_hsw+0x31>
+ DB 233,211,254,255,255 ; jmpq 2f69 <_sk_load_u16_be_hsw+0x31>
PUBLIC _sk_load_rgb_u16_be_hsw
_sk_load_rgb_u16_be_hsw LABEL PROC
@@ -3005,7 +3003,7 @@ _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 317c <_sk_load_rgb_u16_be_hsw+0xde>
+ DB 15,133,204,0,0,0 ; jne 3174 <_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
@@ -3029,7 +3027,7 @@ _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 196,98,125,24,21,162,15,0,0 ; vbroadcastss 0xfa2(%rip),%ymm10 # 40cc <_sk_callback_hsw+0x3c9>
+ DB 196,98,125,24,21,162,15,0,0 ; vbroadcastss 0xfa2(%rip),%ymm10 # 40c4 <_sk_callback_hsw+0x3c9>
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
@@ -3046,48 +3044,48 @@ _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 196,226,125,24,29,86,15,0,0 ; vbroadcastss 0xf56(%rip),%ymm3 # 40d0 <_sk_callback_hsw+0x3cd>
+ DB 196,226,125,24,29,86,15,0,0 ; vbroadcastss 0xf56(%rip),%ymm3 # 40c8 <_sk_callback_hsw+0x3cd>
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 3195 <_sk_load_rgb_u16_be_hsw+0xf7>
- DB 233,79,255,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 318d <_sk_load_rgb_u16_be_hsw+0xf7>
+ DB 233,79,255,255,255 ; jmpq 30dc <_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 31c4 <_sk_load_rgb_u16_be_hsw+0x126>
+ DB 114,26 ; jb 31bc <_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 31c9 <_sk_load_rgb_u16_be_hsw+0x12b>
- DB 233,32,255,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,27,255,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 31c1 <_sk_load_rgb_u16_be_hsw+0x12b>
+ DB 233,32,255,255,255 ; jmpq 30dc <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,27,255,255,255 ; jmpq 30dc <_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 31f8 <_sk_load_rgb_u16_be_hsw+0x15a>
+ DB 114,26 ; jb 31f0 <_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 31fd <_sk_load_rgb_u16_be_hsw+0x15f>
- DB 233,236,254,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,231,254,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 31f5 <_sk_load_rgb_u16_be_hsw+0x15f>
+ DB 233,236,254,255,255 ; jmpq 30dc <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,231,254,255,255 ; jmpq 30dc <_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 3226 <_sk_load_rgb_u16_be_hsw+0x188>
+ DB 114,20 ; jb 321e <_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 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,185,254,255,255 ; jmpq 30e4 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,190,254,255,255 ; jmpq 30dc <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,185,254,255,255 ; jmpq 30dc <_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 196,98,125,24,5,147,14,0,0 ; vbroadcastss 0xe93(%rip),%ymm8 # 40d4 <_sk_callback_hsw+0x3d1>
+ DB 196,98,125,24,5,147,14,0,0 ; vbroadcastss 0xe93(%rip),%ymm8 # 40cc <_sk_callback_hsw+0x3d1>
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
@@ -3125,7 +3123,7 @@ _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 3325 <_sk_store_u16_be_hsw+0xfa>
+ DB 117,31 ; jne 331d <_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)
@@ -3134,31 +3132,31 @@ _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 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 116,240 ; je 3319 <_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 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 114,227 ; jb 3319 <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2)
- DB 116,218 ; je 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 116,218 ; je 3319 <_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 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 114,205 ; jb 3319 <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2)
- DB 116,196 ; je 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 116,196 ; je 3319 <_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 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 114,183 ; jb 3319 <_sk_store_u16_be_hsw+0xf6>
DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2)
- DB 235,174 ; jmp 3321 <_sk_store_u16_be_hsw+0xf6>
+ DB 235,174 ; jmp 3319 <_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 33e9 <_sk_load_f32_hsw+0x76>
+ DB 119,110 ; ja 33e1 <_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 # 3414 <_sk_load_f32_hsw+0xa1>
+ DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 340c <_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
@@ -3217,7 +3215,7 @@ _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 34a1 <_sk_store_f32_hsw+0x6d>
+ DB 117,55 ; jne 3499 <_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
@@ -3230,22 +3228,22 @@ _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 349d <_sk_store_f32_hsw+0x69>
+ DB 116,240 ; je 3495 <_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 349d <_sk_store_f32_hsw+0x69>
+ DB 114,227 ; jb 3495 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 349d <_sk_store_f32_hsw+0x69>
+ DB 116,218 ; je 3495 <_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 349d <_sk_store_f32_hsw+0x69>
+ DB 114,205 ; jb 3495 <_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 349d <_sk_store_f32_hsw+0x69>
+ DB 116,195 ; je 3495 <_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 349d <_sk_store_f32_hsw+0x69>
+ DB 114,181 ; jb 3495 <_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 349d <_sk_store_f32_hsw+0x69>
+ DB 235,171 ; jmp 3495 <_sk_store_f32_hsw+0x69>
PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
@@ -3341,11 +3339,11 @@ _sk_mirror_y_hsw LABEL PROC
PUBLIC _sk_luminance_to_alpha_hsw
_sk_luminance_to_alpha_hsw LABEL PROC
- DB 196,226,125,24,29,173,10,0,0 ; vbroadcastss 0xaad(%rip),%ymm3 # 40d8 <_sk_callback_hsw+0x3d5>
- DB 196,98,125,24,5,168,10,0,0 ; vbroadcastss 0xaa8(%rip),%ymm8 # 40dc <_sk_callback_hsw+0x3d9>
+ DB 196,226,125,24,29,173,10,0,0 ; vbroadcastss 0xaad(%rip),%ymm3 # 40d0 <_sk_callback_hsw+0x3d5>
+ DB 196,98,125,24,5,168,10,0,0 ; vbroadcastss 0xaa8(%rip),%ymm8 # 40d4 <_sk_callback_hsw+0x3d9>
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1
- DB 196,226,125,24,29,153,10,0,0 ; vbroadcastss 0xa99(%rip),%ymm3 # 40e0 <_sk_callback_hsw+0x3dd>
+ DB 196,226,125,24,29,153,10,0,0 ; vbroadcastss 0xa99(%rip),%ymm3 # 40d8 <_sk_callback_hsw+0x3dd>
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
@@ -3480,7 +3478,7 @@ _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 391f <_sk_linear_gradient_hsw+0xb5>
+ DB 15,132,143,0,0,0 ; je 3917 <_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
@@ -3507,8 +3505,8 @@ _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 38a9 <_sk_linear_gradient_hsw+0x3f>
- DB 235,17 ; jmp 3930 <_sk_linear_gradient_hsw+0xc6>
+ DB 117,140 ; jne 38a1 <_sk_linear_gradient_hsw+0x3f>
+ DB 235,17 ; jmp 3928 <_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
@@ -3543,7 +3541,7 @@ _sk_linear_gradient_2stops_hsw LABEL PROC
PUBLIC _sk_save_xy_hsw
_sk_save_xy_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,98,125,24,5,64,7,0,0 ; vbroadcastss 0x740(%rip),%ymm8 # 40e4 <_sk_callback_hsw+0x3e1>
+ DB 196,98,125,24,5,64,7,0,0 ; vbroadcastss 0x740(%rip),%ymm8 # 40dc <_sk_callback_hsw+0x3e1>
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
@@ -3573,9 +3571,9 @@ _sk_accumulate_hsw LABEL PROC
PUBLIC _sk_bilinear_nx_hsw
_sk_bilinear_nx_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,5,212,6,0,0 ; vbroadcastss 0x6d4(%rip),%ymm0 # 40e8 <_sk_callback_hsw+0x3e5>
+ DB 196,226,125,24,5,212,6,0,0 ; vbroadcastss 0x6d4(%rip),%ymm0 # 40e0 <_sk_callback_hsw+0x3e5>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,203,6,0,0 ; vbroadcastss 0x6cb(%rip),%ymm8 # 40ec <_sk_callback_hsw+0x3e9>
+ DB 196,98,125,24,5,203,6,0,0 ; vbroadcastss 0x6cb(%rip),%ymm8 # 40e4 <_sk_callback_hsw+0x3e9>
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
@@ -3584,7 +3582,7 @@ _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 196,226,125,24,5,179,6,0,0 ; vbroadcastss 0x6b3(%rip),%ymm0 # 40f0 <_sk_callback_hsw+0x3ed>
+ DB 196,226,125,24,5,179,6,0,0 ; vbroadcastss 0x6b3(%rip),%ymm0 # 40e8 <_sk_callback_hsw+0x3ed>
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)
@@ -3594,9 +3592,9 @@ _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 196,226,125,24,13,151,6,0,0 ; vbroadcastss 0x697(%rip),%ymm1 # 40f4 <_sk_callback_hsw+0x3f1>
+ DB 196,226,125,24,13,151,6,0,0 ; vbroadcastss 0x697(%rip),%ymm1 # 40ec <_sk_callback_hsw+0x3f1>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,141,6,0,0 ; vbroadcastss 0x68d(%rip),%ymm8 # 40f8 <_sk_callback_hsw+0x3f5>
+ DB 196,98,125,24,5,141,6,0,0 ; vbroadcastss 0x68d(%rip),%ymm8 # 40f0 <_sk_callback_hsw+0x3f5>
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
@@ -3605,7 +3603,7 @@ _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 196,226,125,24,13,117,6,0,0 ; vbroadcastss 0x675(%rip),%ymm1 # 40fc <_sk_callback_hsw+0x3f9>
+ DB 196,226,125,24,13,117,6,0,0 ; vbroadcastss 0x675(%rip),%ymm1 # 40f4 <_sk_callback_hsw+0x3f9>
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)
@@ -3615,13 +3613,13 @@ _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 196,226,125,24,5,88,6,0,0 ; vbroadcastss 0x658(%rip),%ymm0 # 4100 <_sk_callback_hsw+0x3fd>
+ DB 196,226,125,24,5,88,6,0,0 ; vbroadcastss 0x658(%rip),%ymm0 # 40f8 <_sk_callback_hsw+0x3fd>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,79,6,0,0 ; vbroadcastss 0x64f(%rip),%ymm8 # 4104 <_sk_callback_hsw+0x401>
+ DB 196,98,125,24,5,79,6,0,0 ; vbroadcastss 0x64f(%rip),%ymm8 # 40fc <_sk_callback_hsw+0x401>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,64,6,0,0 ; vbroadcastss 0x640(%rip),%ymm10 # 4108 <_sk_callback_hsw+0x405>
- DB 196,98,125,24,29,59,6,0,0 ; vbroadcastss 0x63b(%rip),%ymm11 # 410c <_sk_callback_hsw+0x409>
+ DB 196,98,125,24,21,64,6,0,0 ; vbroadcastss 0x640(%rip),%ymm10 # 4100 <_sk_callback_hsw+0x405>
+ DB 196,98,125,24,29,59,6,0,0 ; vbroadcastss 0x63b(%rip),%ymm11 # 4104 <_sk_callback_hsw+0x409>
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)
@@ -3631,16 +3629,16 @@ _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 196,226,125,24,5,30,6,0,0 ; vbroadcastss 0x61e(%rip),%ymm0 # 4110 <_sk_callback_hsw+0x40d>
+ DB 196,226,125,24,5,30,6,0,0 ; vbroadcastss 0x61e(%rip),%ymm0 # 4108 <_sk_callback_hsw+0x40d>
DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0
- DB 196,98,125,24,5,21,6,0,0 ; vbroadcastss 0x615(%rip),%ymm8 # 4114 <_sk_callback_hsw+0x411>
+ DB 196,98,125,24,5,21,6,0,0 ; vbroadcastss 0x615(%rip),%ymm8 # 410c <_sk_callback_hsw+0x411>
DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,11,6,0,0 ; vbroadcastss 0x60b(%rip),%ymm9 # 4118 <_sk_callback_hsw+0x415>
- DB 196,98,125,24,21,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm10 # 411c <_sk_callback_hsw+0x419>
+ DB 196,98,125,24,13,11,6,0,0 ; vbroadcastss 0x60b(%rip),%ymm9 # 4110 <_sk_callback_hsw+0x415>
+ DB 196,98,125,24,21,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm10 # 4114 <_sk_callback_hsw+0x419>
DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,13,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm9 # 4120 <_sk_callback_hsw+0x41d>
+ DB 196,98,125,24,13,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm9 # 4118 <_sk_callback_hsw+0x41d>
DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9
- DB 196,98,125,24,21,242,5,0,0 ; vbroadcastss 0x5f2(%rip),%ymm10 # 4124 <_sk_callback_hsw+0x421>
+ DB 196,98,125,24,21,242,5,0,0 ; vbroadcastss 0x5f2(%rip),%ymm10 # 411c <_sk_callback_hsw+0x421>
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
@@ -3649,14 +3647,14 @@ _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 196,98,125,24,5,218,5,0,0 ; vbroadcastss 0x5da(%rip),%ymm8 # 4128 <_sk_callback_hsw+0x425>
+ DB 196,98,125,24,5,218,5,0,0 ; vbroadcastss 0x5da(%rip),%ymm8 # 4120 <_sk_callback_hsw+0x425>
DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0
DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9
- DB 196,98,125,24,21,204,5,0,0 ; vbroadcastss 0x5cc(%rip),%ymm10 # 412c <_sk_callback_hsw+0x429>
- DB 196,98,125,24,29,199,5,0,0 ; vbroadcastss 0x5c7(%rip),%ymm11 # 4130 <_sk_callback_hsw+0x42d>
+ DB 196,98,125,24,21,204,5,0,0 ; vbroadcastss 0x5cc(%rip),%ymm10 # 4124 <_sk_callback_hsw+0x429>
+ DB 196,98,125,24,29,199,5,0,0 ; vbroadcastss 0x5c7(%rip),%ymm11 # 4128 <_sk_callback_hsw+0x42d>
DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11
DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11
- DB 196,98,125,24,5,184,5,0,0 ; vbroadcastss 0x5b8(%rip),%ymm8 # 4134 <_sk_callback_hsw+0x431>
+ DB 196,98,125,24,5,184,5,0,0 ; vbroadcastss 0x5b8(%rip),%ymm8 # 412c <_sk_callback_hsw+0x431>
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
@@ -3665,12 +3663,12 @@ _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 196,226,125,24,5,160,5,0,0 ; vbroadcastss 0x5a0(%rip),%ymm0 # 4138 <_sk_callback_hsw+0x435>
+ DB 196,226,125,24,5,160,5,0,0 ; vbroadcastss 0x5a0(%rip),%ymm0 # 4130 <_sk_callback_hsw+0x435>
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 196,98,125,24,21,141,5,0,0 ; vbroadcastss 0x58d(%rip),%ymm10 # 413c <_sk_callback_hsw+0x439>
- DB 196,98,125,24,29,136,5,0,0 ; vbroadcastss 0x588(%rip),%ymm11 # 4140 <_sk_callback_hsw+0x43d>
+ DB 196,98,125,24,21,141,5,0,0 ; vbroadcastss 0x58d(%rip),%ymm10 # 4134 <_sk_callback_hsw+0x439>
+ DB 196,98,125,24,29,136,5,0,0 ; vbroadcastss 0x588(%rip),%ymm11 # 4138 <_sk_callback_hsw+0x43d>
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)
@@ -3680,13 +3678,13 @@ _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 196,226,125,24,13,107,5,0,0 ; vbroadcastss 0x56b(%rip),%ymm1 # 4144 <_sk_callback_hsw+0x441>
+ DB 196,226,125,24,13,107,5,0,0 ; vbroadcastss 0x56b(%rip),%ymm1 # 413c <_sk_callback_hsw+0x441>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,97,5,0,0 ; vbroadcastss 0x561(%rip),%ymm8 # 4148 <_sk_callback_hsw+0x445>
+ DB 196,98,125,24,5,97,5,0,0 ; vbroadcastss 0x561(%rip),%ymm8 # 4140 <_sk_callback_hsw+0x445>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9
- DB 196,98,125,24,21,82,5,0,0 ; vbroadcastss 0x552(%rip),%ymm10 # 414c <_sk_callback_hsw+0x449>
- DB 196,98,125,24,29,77,5,0,0 ; vbroadcastss 0x54d(%rip),%ymm11 # 4150 <_sk_callback_hsw+0x44d>
+ DB 196,98,125,24,21,82,5,0,0 ; vbroadcastss 0x552(%rip),%ymm10 # 4144 <_sk_callback_hsw+0x449>
+ DB 196,98,125,24,29,77,5,0,0 ; vbroadcastss 0x54d(%rip),%ymm11 # 4148 <_sk_callback_hsw+0x44d>
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)
@@ -3696,16 +3694,16 @@ _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 196,226,125,24,13,48,5,0,0 ; vbroadcastss 0x530(%rip),%ymm1 # 4154 <_sk_callback_hsw+0x451>
+ DB 196,226,125,24,13,48,5,0,0 ; vbroadcastss 0x530(%rip),%ymm1 # 414c <_sk_callback_hsw+0x451>
DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1
- DB 196,98,125,24,5,38,5,0,0 ; vbroadcastss 0x526(%rip),%ymm8 # 4158 <_sk_callback_hsw+0x455>
+ DB 196,98,125,24,5,38,5,0,0 ; vbroadcastss 0x526(%rip),%ymm8 # 4150 <_sk_callback_hsw+0x455>
DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8
- DB 196,98,125,24,13,28,5,0,0 ; vbroadcastss 0x51c(%rip),%ymm9 # 415c <_sk_callback_hsw+0x459>
- DB 196,98,125,24,21,23,5,0,0 ; vbroadcastss 0x517(%rip),%ymm10 # 4160 <_sk_callback_hsw+0x45d>
+ DB 196,98,125,24,13,28,5,0,0 ; vbroadcastss 0x51c(%rip),%ymm9 # 4154 <_sk_callback_hsw+0x459>
+ DB 196,98,125,24,21,23,5,0,0 ; vbroadcastss 0x517(%rip),%ymm10 # 4158 <_sk_callback_hsw+0x45d>
DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10
- DB 196,98,125,24,13,13,5,0,0 ; vbroadcastss 0x50d(%rip),%ymm9 # 4164 <_sk_callback_hsw+0x461>
+ DB 196,98,125,24,13,13,5,0,0 ; vbroadcastss 0x50d(%rip),%ymm9 # 415c <_sk_callback_hsw+0x461>
DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9
- DB 196,98,125,24,21,3,5,0,0 ; vbroadcastss 0x503(%rip),%ymm10 # 4168 <_sk_callback_hsw+0x465>
+ DB 196,98,125,24,21,3,5,0,0 ; vbroadcastss 0x503(%rip),%ymm10 # 4160 <_sk_callback_hsw+0x465>
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
@@ -3714,14 +3712,14 @@ _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 196,98,125,24,5,235,4,0,0 ; vbroadcastss 0x4eb(%rip),%ymm8 # 416c <_sk_callback_hsw+0x469>
+ DB 196,98,125,24,5,235,4,0,0 ; vbroadcastss 0x4eb(%rip),%ymm8 # 4164 <_sk_callback_hsw+0x469>
DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1
DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9
- DB 196,98,125,24,21,220,4,0,0 ; vbroadcastss 0x4dc(%rip),%ymm10 # 4170 <_sk_callback_hsw+0x46d>
- DB 196,98,125,24,29,215,4,0,0 ; vbroadcastss 0x4d7(%rip),%ymm11 # 4174 <_sk_callback_hsw+0x471>
+ DB 196,98,125,24,21,220,4,0,0 ; vbroadcastss 0x4dc(%rip),%ymm10 # 4168 <_sk_callback_hsw+0x46d>
+ DB 196,98,125,24,29,215,4,0,0 ; vbroadcastss 0x4d7(%rip),%ymm11 # 416c <_sk_callback_hsw+0x471>
DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11
DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11
- DB 196,98,125,24,5,200,4,0,0 ; vbroadcastss 0x4c8(%rip),%ymm8 # 4178 <_sk_callback_hsw+0x475>
+ DB 196,98,125,24,5,200,4,0,0 ; vbroadcastss 0x4c8(%rip),%ymm8 # 4170 <_sk_callback_hsw+0x475>
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
@@ -3730,12 +3728,12 @@ _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 196,226,125,24,13,176,4,0,0 ; vbroadcastss 0x4b0(%rip),%ymm1 # 417c <_sk_callback_hsw+0x479>
+ DB 196,226,125,24,13,176,4,0,0 ; vbroadcastss 0x4b0(%rip),%ymm1 # 4174 <_sk_callback_hsw+0x479>
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 196,98,125,24,21,156,4,0,0 ; vbroadcastss 0x49c(%rip),%ymm10 # 4180 <_sk_callback_hsw+0x47d>
- DB 196,98,125,24,29,151,4,0,0 ; vbroadcastss 0x497(%rip),%ymm11 # 4184 <_sk_callback_hsw+0x481>
+ DB 196,98,125,24,21,156,4,0,0 ; vbroadcastss 0x49c(%rip),%ymm10 # 4178 <_sk_callback_hsw+0x47d>
+ DB 196,98,125,24,29,151,4,0,0 ; vbroadcastss 0x497(%rip),%ymm11 # 417c <_sk_callback_hsw+0x481>
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)
@@ -3881,7 +3879,7 @@ 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 3ef5 <.literal4+0xd9>
+ DB 224,7 ; loopne 3eed <.literal4+0xd9>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -3895,10 +3893,10 @@ 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 3f20 <.literal4+0x104>
+ DB 127,0 ; jg 3f18 <.literal4+0x104>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 3f99 <.literal4+0x17d>
+ DB 119,115 ; ja 3f91 <.literal4+0x17d>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -3912,10 +3910,10 @@ 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 3f54 <.literal4+0x138>
+ DB 127,0 ; jg 3f4c <.literal4+0x138>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 3fcd <.literal4+0x1b1>
+ DB 119,115 ; ja 3fc5 <.literal4+0x1b1>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -3929,10 +3927,10 @@ 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 3f88 <.literal4+0x16c>
+ DB 127,0 ; jg 3f80 <.literal4+0x16c>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4001 <.literal4+0x1e5>
+ DB 119,115 ; ja 3ff9 <.literal4+0x1e5>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -3946,10 +3944,10 @@ 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 3fbc <.literal4+0x1a0>
+ DB 127,0 ; jg 3fb4 <.literal4+0x1a0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4035 <.literal4+0x219>
+ DB 119,115 ; ja 402d <.literal4+0x219>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -3962,7 +3960,7 @@ 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 4033 <.literal4+0x217>
+ DB 127,67 ; jg 402b <.literal4+0x217>
DB 0,0 ; add %al,(%rax)
DB 0,195 ; add %al,%bl
DB 0,0 ; add %al,(%rax)
@@ -3974,10 +3972,10 @@ ALIGN 4
DB 190,80,128,3,62 ; mov $0x3e038050,%esi
DB 31 ; (bad)
DB 215 ; xlat %ds:(%rbx)
- DB 118,63 ; jbe 4053 <.literal4+0x237>
+ DB 118,63 ; jbe 404b <.literal4+0x237>
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 4067 <.literal4+0x24b>
+ DB 127,67 ; jg 405f <.literal4+0x24b>
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)
@@ -3986,7 +3984,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4049 <.literal4+0x22d>
+ DB 224,7 ; loopne 4041 <.literal4+0x22d>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -3998,7 +3996,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 8,33 ; or %ah,(%rcx)
DB 132,55 ; test %dh,(%rdi)
- DB 224,7 ; loopne 4065 <.literal4+0x249>
+ DB 224,7 ; loopne 405d <.literal4+0x249>
DB 0,0 ; add %al,(%rax)
DB 33,8 ; and %ecx,(%rax)
DB 2,58 ; add (%rdx),%bh
@@ -4009,7 +4007,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 248 ; clc
DB 65,0,0 ; add %al,(%r8)
- DB 124,66 ; jl 40ba <.literal4+0x29e>
+ DB 124,66 ; jl 40b2 <.literal4+0x29e>
DB 0,240 ; add %dh,%al
DB 0,0 ; add %al,(%rax)
DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax)
@@ -4027,9 +4025,9 @@ 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 40fd <.literal4+0x2e1>
+ DB 112,65 ; jo 40f5 <.literal4+0x2e1>
DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax)
- DB 127,67 ; jg 410b <.literal4+0x2ef>
+ DB 127,67 ; jg 4103 <.literal4+0x2ef>
DB 128,0,128 ; addb $0x80,(%rax)
DB 55 ; (bad)
DB 128,0,128 ; addb $0x80,(%rax)
@@ -4037,7 +4035,7 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
DB 255 ; (bad)
- DB 127,71 ; jg 411f <.literal4+0x303>
+ DB 127,71 ; jg 4117 <.literal4+0x303>
DB 208 ; (bad)
DB 179,89 ; mov $0x59,%bl
DB 62,89 ; ds pop %rcx
@@ -4123,16 +4121,16 @@ 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) # a0041c8 <_sk_callback_hsw+0xa0004c5>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0041a8 <_sk_callback_hsw+0xa0004ad>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120041d0 <_sk_callback_hsw+0x120004cd>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120041b0 <_sk_callback_hsw+0x120004b5>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0041d8 <_sk_callback_hsw+0x1a0004d5>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0041b8 <_sk_callback_hsw+0x1a0004bd>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30041e0 <_sk_callback_hsw+0x30004dd>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30041c0 <_sk_callback_hsw+0x30004c5>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4175,16 +4173,16 @@ 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) # a004228 <_sk_callback_hsw+0xa000525>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004208 <_sk_callback_hsw+0xa00050d>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004230 <_sk_callback_hsw+0x1200052d>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004210 <_sk_callback_hsw+0x12000515>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004238 <_sk_callback_hsw+0x1a000535>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004218 <_sk_callback_hsw+0x1a00051d>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004240 <_sk_callback_hsw+0x300053d>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004220 <_sk_callback_hsw+0x3000525>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4227,16 +4225,16 @@ 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) # a004288 <_sk_callback_hsw+0xa000585>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004268 <_sk_callback_hsw+0xa00056d>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004290 <_sk_callback_hsw+0x1200058d>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004270 <_sk_callback_hsw+0x12000575>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004298 <_sk_callback_hsw+0x1a000595>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004278 <_sk_callback_hsw+0x1a00057d>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30042a0 <_sk_callback_hsw+0x300059d>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004280 <_sk_callback_hsw+0x3000585>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -4279,16 +4277,16 @@ 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+0xa0005e5>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0042c8 <_sk_callback_hsw+0xa0005cd>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120042f0 <_sk_callback_hsw+0x120005ed>
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120042d0 <_sk_callback_hsw+0x120005d5>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0042f8 <_sk_callback_hsw+0x1a0005f5>
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0042d8 <_sk_callback_hsw+0x1a0005dd>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004300 <_sk_callback_hsw+0x30005fd>
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30042e0 <_sk_callback_hsw+0x30005e5>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -5325,23 +5323,23 @@ _sk_hsl_to_rgb_avx LABEL PROC
DB 197,252,17,172,36,128,0,0,0 ; vmovups %ymm5,0x80(%rsp)
DB 197,252,17,100,36,96 ; vmovups %ymm4,0x60(%rsp)
DB 197,252,17,92,36,64 ; vmovups %ymm3,0x40(%rsp)
- DB 197,252,40,242 ; vmovaps %ymm2,%ymm6
- DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp)
+ DB 197,252,40,234 ; vmovaps %ymm2,%ymm5
DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
DB 184,0,0,0,63 ; mov $0x3f000000,%eax
DB 197,249,110,192 ; vmovd %eax,%xmm0
DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0
DB 196,99,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm8
- DB 196,193,76,194,192,1 ; vcmpltps %ymm8,%ymm6,%ymm0
- DB 196,98,125,24,21,94,70,0,0 ; vbroadcastss 0x465e(%rip),%ymm10 # 55a0 <_sk_callback_avx+0x1be>
+ DB 196,193,84,194,192,1 ; vcmpltps %ymm8,%ymm5,%ymm0
+ DB 196,98,125,24,21,100,70,0,0 ; vbroadcastss 0x4664(%rip),%ymm10 # 55a0 <_sk_callback_avx+0x1be>
+ DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp)
DB 196,193,116,88,218 ; vaddps %ymm10,%ymm1,%ymm3
- DB 197,228,89,222 ; vmulps %ymm6,%ymm3,%ymm3
- DB 197,244,88,230 ; vaddps %ymm6,%ymm1,%ymm4
- DB 197,244,89,238 ; vmulps %ymm6,%ymm1,%ymm5
- DB 197,220,92,229 ; vsubps %ymm5,%ymm4,%ymm4
+ DB 197,228,89,221 ; vmulps %ymm5,%ymm3,%ymm3
+ DB 197,244,88,229 ; vaddps %ymm5,%ymm1,%ymm4
+ DB 197,244,89,245 ; vmulps %ymm5,%ymm1,%ymm6
+ DB 197,220,92,230 ; vsubps %ymm6,%ymm4,%ymm4
DB 196,99,93,74,203,0 ; vblendvps %ymm0,%ymm3,%ymm4,%ymm9
- DB 196,226,125,24,5,62,70,0,0 ; vbroadcastss 0x463e(%rip),%ymm0 # 55a4 <_sk_callback_avx+0x1c2>
- DB 197,236,88,200 ; vaddps %ymm0,%ymm2,%ymm1
+ DB 196,226,125,24,13,62,70,0,0 ; vbroadcastss 0x463e(%rip),%ymm1 # 55a4 <_sk_callback_avx+0x1c2>
+ DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1
DB 65,184,0,0,0,0 ; mov $0x0,%r8d
DB 184,0,0,128,63 ; mov $0x3f800000,%eax
DB 197,249,110,216 ; vmovd %eax,%xmm3
@@ -5355,76 +5353,76 @@ _sk_hsl_to_rgb_avx LABEL PROC
DB 196,227,121,4,228,0 ; vpermilps $0x0,%xmm4,%xmm4
DB 196,99,93,24,252,1 ; vinsertf128 $0x1,%xmm4,%ymm4,%ymm15
DB 196,193,116,194,231,1 ; vcmpltps %ymm15,%ymm1,%ymm4
- DB 196,193,116,88,234 ; vaddps %ymm10,%ymm1,%ymm5
- DB 196,227,101,74,197,64 ; vblendvps %ymm4,%ymm5,%ymm3,%ymm0
- DB 197,204,88,222 ; vaddps %ymm6,%ymm6,%ymm3
- DB 196,65,100,92,217 ; vsubps %ymm9,%ymm3,%ymm11
- DB 196,193,52,92,219 ; vsubps %ymm11,%ymm9,%ymm3
- DB 196,226,125,24,37,213,69,0,0 ; vbroadcastss 0x45d5(%rip),%ymm4 # 55ac <_sk_callback_avx+0x1ca>
- DB 197,100,89,236 ; vmulps %ymm4,%ymm3,%ymm13
+ DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
+ DB 196,227,101,74,241,64 ; vblendvps %ymm4,%ymm1,%ymm3,%ymm6
+ DB 197,212,88,205 ; vaddps %ymm5,%ymm5,%ymm1
+ DB 196,65,116,92,217 ; vsubps %ymm9,%ymm1,%ymm11
+ DB 196,193,52,92,203 ; vsubps %ymm11,%ymm9,%ymm1
+ DB 196,226,125,24,29,213,69,0,0 ; vbroadcastss 0x45d5(%rip),%ymm3 # 55ac <_sk_callback_avx+0x1ca>
+ DB 197,116,89,235 ; vmulps %ymm3,%ymm1,%ymm13
DB 65,184,171,170,42,62 ; mov $0x3e2aaaab,%r8d
DB 184,171,170,42,63 ; mov $0x3f2aaaab,%eax
- DB 197,249,110,216 ; vmovd %eax,%xmm3
- DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3
- DB 196,227,101,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm5
- DB 196,226,125,24,37,177,69,0,0 ; vbroadcastss 0x45b1(%rip),%ymm4 # 55b0 <_sk_callback_avx+0x1ce>
- DB 197,220,92,216 ; vsubps %ymm0,%ymm4,%ymm3
- DB 197,148,89,219 ; vmulps %ymm3,%ymm13,%ymm3
- DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3
- DB 197,252,194,253,1 ; vcmpltps %ymm5,%ymm0,%ymm7
- DB 196,227,37,74,219,112 ; vblendvps %ymm7,%ymm3,%ymm11,%ymm3
- DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7
- DB 196,195,101,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm3,%ymm7
- DB 196,193,121,110,216 ; vmovd %r8d,%xmm3
- DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3
- DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
- DB 197,252,194,195,1 ; vcmpltps %ymm3,%ymm0,%ymm0
- DB 196,193,116,89,205 ; vmulps %ymm13,%ymm1,%ymm1
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1
+ DB 196,227,117,24,225,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm4
+ DB 196,226,125,24,29,177,69,0,0 ; vbroadcastss 0x45b1(%rip),%ymm3 # 55b0 <_sk_callback_avx+0x1ce>
+ DB 197,228,92,206 ; vsubps %ymm6,%ymm3,%ymm1
+ DB 197,148,89,201 ; vmulps %ymm1,%ymm13,%ymm1
DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1
- DB 196,227,69,74,193,0 ; vblendvps %ymm0,%ymm1,%ymm7,%ymm0
+ DB 197,204,194,252,1 ; vcmpltps %ymm4,%ymm6,%ymm7
+ DB 196,227,37,74,201,112 ; vblendvps %ymm7,%ymm1,%ymm11,%ymm1
+ DB 196,193,76,194,248,1 ; vcmpltps %ymm8,%ymm6,%ymm7
+ DB 196,195,117,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm1,%ymm7
+ DB 196,193,121,110,200 ; vmovd %r8d,%xmm1
+ DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1
+ DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ DB 197,204,194,193,1 ; vcmpltps %ymm1,%ymm6,%ymm0
+ DB 197,148,89,246 ; vmulps %ymm6,%ymm13,%ymm6
+ DB 197,164,88,246 ; vaddps %ymm6,%ymm11,%ymm6
+ DB 196,227,69,74,198,0 ; vblendvps %ymm0,%ymm6,%ymm7,%ymm0
DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp)
- DB 197,156,194,202,1 ; vcmpltps %ymm2,%ymm12,%ymm1
+ DB 197,156,194,194,1 ; vcmpltps %ymm2,%ymm12,%ymm0
DB 196,193,108,88,254 ; vaddps %ymm14,%ymm2,%ymm7
- DB 196,227,109,74,207,16 ; vblendvps %ymm1,%ymm7,%ymm2,%ymm1
+ DB 196,227,109,74,199,0 ; vblendvps %ymm0,%ymm7,%ymm2,%ymm0
DB 196,193,108,194,255,1 ; vcmpltps %ymm15,%ymm2,%ymm7
- DB 196,193,108,88,194 ; vaddps %ymm10,%ymm2,%ymm0
- DB 196,227,117,74,192,112 ; vblendvps %ymm7,%ymm0,%ymm1,%ymm0
- DB 197,220,92,200 ; vsubps %ymm0,%ymm4,%ymm1
- DB 197,148,89,201 ; vmulps %ymm1,%ymm13,%ymm1
- DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1
- DB 197,252,194,253,1 ; vcmpltps %ymm5,%ymm0,%ymm7
- DB 196,227,37,74,201,112 ; vblendvps %ymm7,%ymm1,%ymm11,%ymm1
+ DB 196,193,108,88,242 ; vaddps %ymm10,%ymm2,%ymm6
+ DB 196,227,125,74,198,112 ; vblendvps %ymm7,%ymm6,%ymm0,%ymm0
+ DB 197,228,92,240 ; vsubps %ymm0,%ymm3,%ymm6
+ DB 197,148,89,246 ; vmulps %ymm6,%ymm13,%ymm6
+ DB 197,164,88,246 ; vaddps %ymm6,%ymm11,%ymm6
+ DB 197,252,194,252,1 ; vcmpltps %ymm4,%ymm0,%ymm7
+ DB 196,227,37,74,246,112 ; vblendvps %ymm7,%ymm6,%ymm11,%ymm6
DB 196,193,124,194,248,1 ; vcmpltps %ymm8,%ymm0,%ymm7
- DB 196,195,117,74,201,112 ; vblendvps %ymm7,%ymm9,%ymm1,%ymm1
- DB 197,252,194,195,1 ; vcmpltps %ymm3,%ymm0,%ymm0
- DB 197,148,89,250 ; vmulps %ymm2,%ymm13,%ymm7
- DB 197,164,88,255 ; vaddps %ymm7,%ymm11,%ymm7
- DB 196,227,117,74,207,0 ; vblendvps %ymm0,%ymm7,%ymm1,%ymm1
- DB 196,226,125,24,5,8,69,0,0 ; vbroadcastss 0x4508(%rip),%ymm0 # 55b4 <_sk_callback_avx+0x1d2>
+ DB 196,195,77,74,241,112 ; vblendvps %ymm7,%ymm9,%ymm6,%ymm6
+ DB 197,252,194,249,1 ; vcmpltps %ymm1,%ymm0,%ymm7
+ DB 197,148,89,192 ; vmulps %ymm0,%ymm13,%ymm0
+ DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0
+ DB 196,227,77,74,240,112 ; vblendvps %ymm7,%ymm0,%ymm6,%ymm6
+ DB 196,226,125,24,5,9,69,0,0 ; vbroadcastss 0x4509(%rip),%ymm0 # 55b4 <_sk_callback_avx+0x1d2>
DB 197,236,88,192 ; vaddps %ymm0,%ymm2,%ymm0
DB 197,156,194,208,1 ; vcmpltps %ymm0,%ymm12,%ymm2
DB 196,193,124,88,254 ; vaddps %ymm14,%ymm0,%ymm7
DB 196,227,125,74,215,32 ; vblendvps %ymm2,%ymm7,%ymm0,%ymm2
DB 196,193,124,194,255,1 ; vcmpltps %ymm15,%ymm0,%ymm7
- DB 196,65,124,88,210 ; vaddps %ymm10,%ymm0,%ymm10
- DB 196,195,109,74,210,112 ; vblendvps %ymm7,%ymm10,%ymm2,%ymm2
- DB 197,236,194,237,1 ; vcmpltps %ymm5,%ymm2,%ymm5
- DB 197,220,92,226 ; vsubps %ymm2,%ymm4,%ymm4
- DB 197,148,89,228 ; vmulps %ymm4,%ymm13,%ymm4
- DB 197,164,88,228 ; vaddps %ymm4,%ymm11,%ymm4
- DB 196,227,37,74,228,80 ; vblendvps %ymm5,%ymm4,%ymm11,%ymm4
- DB 196,193,108,194,232,1 ; vcmpltps %ymm8,%ymm2,%ymm5
- DB 196,195,93,74,225,80 ; vblendvps %ymm5,%ymm9,%ymm4,%ymm4
- DB 197,236,194,211,1 ; vcmpltps %ymm3,%ymm2,%ymm2
- DB 196,193,124,89,197 ; vmulps %ymm13,%ymm0,%ymm0
+ DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
+ DB 196,227,109,74,192,112 ; vblendvps %ymm7,%ymm0,%ymm2,%ymm0
+ DB 197,252,194,212,1 ; vcmpltps %ymm4,%ymm0,%ymm2
+ DB 197,228,92,216 ; vsubps %ymm0,%ymm3,%ymm3
+ DB 197,148,89,219 ; vmulps %ymm3,%ymm13,%ymm3
+ DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3
+ DB 196,227,37,74,211,32 ; vblendvps %ymm2,%ymm3,%ymm11,%ymm2
+ DB 196,193,124,194,216,1 ; vcmpltps %ymm8,%ymm0,%ymm3
+ DB 196,195,109,74,209,48 ; vblendvps %ymm3,%ymm9,%ymm2,%ymm2
+ DB 197,252,194,201,1 ; vcmpltps %ymm1,%ymm0,%ymm1
+ DB 197,148,89,192 ; vmulps %ymm0,%ymm13,%ymm0
DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0
- DB 196,227,93,74,208,32 ; vblendvps %ymm2,%ymm0,%ymm4,%ymm2
+ DB 196,227,109,74,208,16 ; vblendvps %ymm1,%ymm0,%ymm2,%ymm2
DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0
DB 197,252,194,92,36,32,0 ; vcmpeqps 0x20(%rsp),%ymm0,%ymm3
DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0
- DB 196,227,125,74,198,48 ; vblendvps %ymm3,%ymm6,%ymm0,%ymm0
- DB 196,227,117,74,206,48 ; vblendvps %ymm3,%ymm6,%ymm1,%ymm1
- DB 196,227,109,74,214,48 ; vblendvps %ymm3,%ymm6,%ymm2,%ymm2
+ DB 196,227,125,74,197,48 ; vblendvps %ymm3,%ymm5,%ymm0,%ymm0
+ DB 196,227,77,74,205,48 ; vblendvps %ymm3,%ymm5,%ymm6,%ymm1
+ DB 196,227,109,74,213,48 ; vblendvps %ymm3,%ymm5,%ymm2,%ymm2
DB 72,173 ; lods %ds:(%rsi),%rax
DB 197,252,16,92,36,64 ; vmovups 0x40(%rsp),%ymm3
DB 197,252,16,100,36,96 ; vmovups 0x60(%rsp),%ymm4
@@ -5452,14 +5450,14 @@ _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 11cb <_sk_scale_u8_avx+0x54>
+ DB 117,68 ; jne 11c9 <_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 196,98,125,24,13,9,68,0,0 ; vbroadcastss 0x4409(%rip),%ymm9 # 55b8 <_sk_callback_avx+0x1d6>
+ DB 196,98,125,24,13,11,68,0,0 ; vbroadcastss 0x440b(%rip),%ymm9 # 55b8 <_sk_callback_avx+0x1d6>
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
@@ -5477,9 +5475,9 @@ _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 11d3 <_sk_scale_u8_avx+0x5c>
+ DB 117,234 ; jne 11d1 <_sk_scale_u8_avx+0x5c>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 235,155 ; jmp 118b <_sk_scale_u8_avx+0x14>
+ DB 235,155 ; jmp 1189 <_sk_scale_u8_avx+0x14>
PUBLIC _sk_lerp_1_float_avx
_sk_lerp_1_float_avx LABEL PROC
@@ -5507,14 +5505,14 @@ _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 12a7 <_sk_lerp_u8_avx+0x78>
+ DB 117,104 ; jne 12a5 <_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 196,98,125,24,13,85,67,0,0 ; vbroadcastss 0x4355(%rip),%ymm9 # 55bc <_sk_callback_avx+0x1da>
+ DB 196,98,125,24,13,87,67,0,0 ; vbroadcastss 0x4357(%rip),%ymm9 # 55bc <_sk_callback_avx+0x1da>
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
@@ -5540,35 +5538,35 @@ _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 12af <_sk_lerp_u8_avx+0x80>
+ DB 117,234 ; jne 12ad <_sk_lerp_u8_avx+0x80>
DB 196,65,249,110,193 ; vmovq %r9,%xmm8
- DB 233,116,255,255,255 ; jmpq 1243 <_sk_lerp_u8_avx+0x14>
+ DB 233,116,255,255,255 ; jmpq 1241 <_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 138b <_sk_lerp_565_avx+0xbc>
+ DB 15,133,174,0,0,0 ; jne 1389 <_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 196,98,125,24,5,193,66,0,0 ; vbroadcastss 0x42c1(%rip),%ymm8 # 55c0 <_sk_callback_avx+0x1de>
+ DB 196,98,125,24,5,195,66,0,0 ; vbroadcastss 0x42c3(%rip),%ymm8 # 55c0 <_sk_callback_avx+0x1de>
DB 196,65,100,84,192 ; vandps %ymm8,%ymm3,%ymm8
DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
- DB 196,98,125,24,13,178,66,0,0 ; vbroadcastss 0x42b2(%rip),%ymm9 # 55c4 <_sk_callback_avx+0x1e2>
+ DB 196,98,125,24,13,180,66,0,0 ; vbroadcastss 0x42b4(%rip),%ymm9 # 55c4 <_sk_callback_avx+0x1e2>
DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8
- DB 196,98,125,24,13,168,66,0,0 ; vbroadcastss 0x42a8(%rip),%ymm9 # 55c8 <_sk_callback_avx+0x1e6>
+ DB 196,98,125,24,13,170,66,0,0 ; vbroadcastss 0x42aa(%rip),%ymm9 # 55c8 <_sk_callback_avx+0x1e6>
DB 196,65,100,84,201 ; vandps %ymm9,%ymm3,%ymm9
DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9
- DB 196,98,125,24,21,153,66,0,0 ; vbroadcastss 0x4299(%rip),%ymm10 # 55cc <_sk_callback_avx+0x1ea>
+ DB 196,98,125,24,21,155,66,0,0 ; vbroadcastss 0x429b(%rip),%ymm10 # 55cc <_sk_callback_avx+0x1ea>
DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9
- DB 196,98,125,24,21,143,66,0,0 ; vbroadcastss 0x428f(%rip),%ymm10 # 55d0 <_sk_callback_avx+0x1ee>
+ DB 196,98,125,24,21,145,66,0,0 ; vbroadcastss 0x4291(%rip),%ymm10 # 55d0 <_sk_callback_avx+0x1ee>
DB 196,193,100,84,218 ; vandps %ymm10,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,98,125,24,21,129,66,0,0 ; vbroadcastss 0x4281(%rip),%ymm10 # 55d4 <_sk_callback_avx+0x1f2>
+ DB 196,98,125,24,21,131,66,0,0 ; vbroadcastss 0x4283(%rip),%ymm10 # 55d4 <_sk_callback_avx+0x1f2>
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
@@ -5580,16 +5578,16 @@ _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 196,226,125,24,29,79,66,0,0 ; vbroadcastss 0x424f(%rip),%ymm3 # 55d8 <_sk_callback_avx+0x1f6>
+ DB 196,226,125,24,29,81,66,0,0 ; vbroadcastss 0x4251(%rip),%ymm3 # 55d8 <_sk_callback_avx+0x1f6>
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 12e3 <_sk_lerp_565_avx+0x14>
+ DB 15,135,63,255,255,255 ; ja 12e1 <_sk_lerp_565_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 13f8 <_sk_lerp_565_avx+0x129>
+ DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 13f8 <_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
@@ -5601,27 +5599,28 @@ _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 12e3 <_sk_lerp_565_avx+0x14>
- DB 244 ; hlt
+ DB 233,235,254,255,255 ; jmpq 12e1 <_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 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
@@ -9842,7 +9841,7 @@ _sk_seed_shader_sse41 LABEL PROC
DB 102,15,110,199 ; movd %edi,%xmm0
DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
- DB 15,40,21,129,57,0,0 ; movaps 0x3981(%rip),%xmm2 # 3a90 <_sk_callback_sse41+0xb7>
+ DB 15,40,21,129,57,0,0 ; movaps 0x3981(%rip),%xmm2 # 3a90 <_sk_callback_sse41+0xae>
DB 15,88,202 ; addps %xmm2,%xmm1
DB 15,16,2 ; movups (%rdx),%xmm0
DB 15,88,193 ; addps %xmm1,%xmm0
@@ -9851,7 +9850,7 @@ _sk_seed_shader_sse41 LABEL PROC
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 15,88,202 ; addps %xmm2,%xmm1
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,21,112,57,0,0 ; movaps 0x3970(%rip),%xmm2 # 3aa0 <_sk_callback_sse41+0xc7>
+ DB 15,40,21,112,57,0,0 ; movaps 0x3970(%rip),%xmm2 # 3aa0 <_sk_callback_sse41+0xbe>
DB 15,87,219 ; xorps %xmm3,%xmm3
DB 15,87,228 ; xorps %xmm4,%xmm4
DB 15,87,237 ; xorps %xmm5,%xmm5
@@ -9885,7 +9884,7 @@ _sk_clear_sse41 LABEL PROC
PUBLIC _sk_srcatop_sse41
_sk_srcatop_sse41 LABEL PROC
DB 15,89,199 ; mulps %xmm7,%xmm0
- DB 68,15,40,5,43,57,0,0 ; movaps 0x392b(%rip),%xmm8 # 3ab0 <_sk_callback_sse41+0xd7>
+ DB 68,15,40,5,43,57,0,0 ; movaps 0x392b(%rip),%xmm8 # 3ab0 <_sk_callback_sse41+0xce>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,89,204 ; mulps %xmm4,%xmm9
@@ -9908,7 +9907,7 @@ 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 68,15,40,13,238,56,0,0 ; movaps 0x38ee(%rip),%xmm9 # 3ac0 <_sk_callback_sse41+0xe7>
+ DB 68,15,40,13,238,56,0,0 ; movaps 0x38ee(%rip),%xmm9 # 3ac0 <_sk_callback_sse41+0xde>
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,88,192 ; addps %xmm8,%xmm0
@@ -9949,7 +9948,7 @@ _sk_dstin_sse41 LABEL PROC
PUBLIC _sk_srcout_sse41
_sk_srcout_sse41 LABEL PROC
- DB 68,15,40,5,146,56,0,0 ; movaps 0x3892(%rip),%xmm8 # 3ad0 <_sk_callback_sse41+0xf7>
+ DB 68,15,40,5,146,56,0,0 ; movaps 0x3892(%rip),%xmm8 # 3ad0 <_sk_callback_sse41+0xee>
DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
@@ -9960,7 +9959,7 @@ _sk_srcout_sse41 LABEL PROC
PUBLIC _sk_dstout_sse41
_sk_dstout_sse41 LABEL PROC
- DB 68,15,40,5,130,56,0,0 ; movaps 0x3882(%rip),%xmm8 # 3ae0 <_sk_callback_sse41+0x107>
+ DB 68,15,40,5,130,56,0,0 ; movaps 0x3882(%rip),%xmm8 # 3ae0 <_sk_callback_sse41+0xfe>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 15,89,196 ; mulps %xmm4,%xmm0
@@ -9975,7 +9974,7 @@ _sk_dstout_sse41 LABEL PROC
PUBLIC _sk_srcover_sse41
_sk_srcover_sse41 LABEL PROC
- DB 68,15,40,5,101,56,0,0 ; movaps 0x3865(%rip),%xmm8 # 3af0 <_sk_callback_sse41+0x117>
+ DB 68,15,40,5,101,56,0,0 ; movaps 0x3865(%rip),%xmm8 # 3af0 <_sk_callback_sse41+0x10e>
DB 68,15,92,195 ; subps %xmm3,%xmm8
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,89,204 ; mulps %xmm4,%xmm9
@@ -9993,7 +9992,7 @@ _sk_srcover_sse41 LABEL PROC
PUBLIC _sk_dstover_sse41
_sk_dstover_sse41 LABEL PROC
- DB 68,15,40,5,57,56,0,0 ; movaps 0x3839(%rip),%xmm8 # 3b00 <_sk_callback_sse41+0x127>
+ DB 68,15,40,5,57,56,0,0 ; movaps 0x3839(%rip),%xmm8 # 3b00 <_sk_callback_sse41+0x11e>
DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -10017,7 +10016,7 @@ _sk_modulate_sse41 LABEL PROC
PUBLIC _sk_multiply_sse41
_sk_multiply_sse41 LABEL PROC
- DB 68,15,40,5,13,56,0,0 ; movaps 0x380d(%rip),%xmm8 # 3b10 <_sk_callback_sse41+0x137>
+ DB 68,15,40,5,13,56,0,0 ; movaps 0x380d(%rip),%xmm8 # 3b10 <_sk_callback_sse41+0x12e>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 69,15,40,209 ; movaps %xmm9,%xmm10
@@ -10087,7 +10086,7 @@ _sk_screen_sse41 LABEL PROC
PUBLIC _sk_xor__sse41
_sk_xor__sse41 LABEL PROC
DB 68,15,40,195 ; movaps %xmm3,%xmm8
- DB 15,40,29,62,55,0,0 ; movaps 0x373e(%rip),%xmm3 # 3b20 <_sk_callback_sse41+0x147>
+ DB 15,40,29,62,55,0,0 ; movaps 0x373e(%rip),%xmm3 # 3b20 <_sk_callback_sse41+0x13e>
DB 68,15,40,203 ; movaps %xmm3,%xmm9
DB 68,15,92,207 ; subps %xmm7,%xmm9
DB 65,15,89,193 ; mulps %xmm9,%xmm0
@@ -10133,7 +10132,7 @@ _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 15,40,21,169,54,0,0 ; movaps 0x36a9(%rip),%xmm2 # 3b30 <_sk_callback_sse41+0x157>
+ DB 15,40,21,169,54,0,0 ; movaps 0x36a9(%rip),%xmm2 # 3b30 <_sk_callback_sse41+0x14e>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -10165,7 +10164,7 @@ _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 15,40,21,78,54,0,0 ; movaps 0x364e(%rip),%xmm2 # 3b40 <_sk_callback_sse41+0x167>
+ DB 15,40,21,78,54,0,0 ; movaps 0x364e(%rip),%xmm2 # 3b40 <_sk_callback_sse41+0x15e>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -10200,7 +10199,7 @@ _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 15,40,21,232,53,0,0 ; movaps 0x35e8(%rip),%xmm2 # 3b50 <_sk_callback_sse41+0x177>
+ DB 15,40,21,232,53,0,0 ; movaps 0x35e8(%rip),%xmm2 # 3b50 <_sk_callback_sse41+0x16e>
DB 15,92,211 ; subps %xmm3,%xmm2
DB 15,89,215 ; mulps %xmm7,%xmm2
DB 15,88,218 ; addps %xmm2,%xmm3
@@ -10225,7 +10224,7 @@ _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 15,40,13,169,53,0,0 ; movaps 0x35a9(%rip),%xmm1 # 3b60 <_sk_callback_sse41+0x187>
+ DB 15,40,13,169,53,0,0 ; movaps 0x35a9(%rip),%xmm1 # 3b60 <_sk_callback_sse41+0x17e>
DB 15,92,203 ; subps %xmm3,%xmm1
DB 15,89,207 ; mulps %xmm7,%xmm1
DB 15,88,217 ; addps %xmm1,%xmm3
@@ -10237,7 +10236,7 @@ _sk_exclusion_sse41 LABEL PROC
PUBLIC _sk_colorburn_sse41
_sk_colorburn_sse41 LABEL PROC
DB 68,15,40,192 ; movaps %xmm0,%xmm8
- DB 68,15,40,21,152,53,0,0 ; movaps 0x3598(%rip),%xmm10 # 3b70 <_sk_callback_sse41+0x197>
+ DB 68,15,40,21,152,53,0,0 ; movaps 0x3598(%rip),%xmm10 # 3b70 <_sk_callback_sse41+0x18e>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 69,15,40,203 ; movaps %xmm11,%xmm9
@@ -10317,7 +10316,7 @@ _sk_colorburn_sse41 LABEL PROC
PUBLIC _sk_colordodge_sse41
_sk_colordodge_sse41 LABEL PROC
DB 68,15,40,192 ; movaps %xmm0,%xmm8
- DB 68,15,40,21,118,52,0,0 ; movaps 0x3476(%rip),%xmm10 # 3b80 <_sk_callback_sse41+0x1a7>
+ DB 68,15,40,21,118,52,0,0 ; movaps 0x3476(%rip),%xmm10 # 3b80 <_sk_callback_sse41+0x19e>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 69,15,40,227 ; movaps %xmm11,%xmm12
@@ -10398,7 +10397,7 @@ _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 68,15,40,21,76,51,0,0 ; movaps 0x334c(%rip),%xmm10 # 3b90 <_sk_callback_sse41+0x1b7>
+ DB 68,15,40,21,76,51,0,0 ; movaps 0x334c(%rip),%xmm10 # 3b90 <_sk_callback_sse41+0x1ae>
DB 65,15,40,234 ; movaps %xmm10,%xmm5
DB 15,92,239 ; subps %xmm7,%xmm5
DB 15,40,197 ; movaps %xmm5,%xmm0
@@ -10480,7 +10479,7 @@ 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 68,15,40,21,46,50,0,0 ; movaps 0x322e(%rip),%xmm10 # 3ba0 <_sk_callback_sse41+0x1c7>
+ DB 68,15,40,21,46,50,0,0 ; movaps 0x322e(%rip),%xmm10 # 3ba0 <_sk_callback_sse41+0x1be>
DB 69,15,40,218 ; movaps %xmm10,%xmm11
DB 68,15,92,223 ; subps %xmm7,%xmm11
DB 65,15,40,195 ; movaps %xmm11,%xmm0
@@ -10564,7 +10563,7 @@ _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 15,40,13,1,49,0,0 ; movaps 0x3101(%rip),%xmm1 # 3bb0 <_sk_callback_sse41+0x1d7>
+ DB 15,40,13,1,49,0,0 ; movaps 0x3101(%rip),%xmm1 # 3bb0 <_sk_callback_sse41+0x1ce>
DB 68,15,40,209 ; movaps %xmm1,%xmm10
DB 68,15,92,208 ; subps %xmm0,%xmm10
DB 68,15,40,240 ; movaps %xmm0,%xmm14
@@ -10577,10 +10576,10 @@ _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 68,15,40,45,223,48,0,0 ; movaps 0x30df(%rip),%xmm13 # 3bc0 <_sk_callback_sse41+0x1e7>
+ DB 68,15,40,45,223,48,0,0 ; movaps 0x30df(%rip),%xmm13 # 3bc0 <_sk_callback_sse41+0x1de>
DB 69,15,88,245 ; addps %xmm13,%xmm14
DB 68,15,89,242 ; mulps %xmm2,%xmm14
- DB 68,15,40,37,223,48,0,0 ; movaps 0x30df(%rip),%xmm12 # 3bd0 <_sk_callback_sse41+0x1f7>
+ DB 68,15,40,37,223,48,0,0 ; movaps 0x30df(%rip),%xmm12 # 3bd0 <_sk_callback_sse41+0x1ee>
DB 69,15,89,252 ; mulps %xmm12,%xmm15
DB 69,15,88,254 ; addps %xmm14,%xmm15
DB 15,40,198 ; movaps %xmm6,%xmm0
@@ -10725,7 +10724,7 @@ _sk_clamp_0_sse41 LABEL PROC
PUBLIC _sk_clamp_1_sse41
_sk_clamp_1_sse41 LABEL PROC
- DB 68,15,40,5,239,46,0,0 ; movaps 0x2eef(%rip),%xmm8 # 3be0 <_sk_callback_sse41+0x207>
+ DB 68,15,40,5,239,46,0,0 ; movaps 0x2eef(%rip),%xmm8 # 3be0 <_sk_callback_sse41+0x1fe>
DB 65,15,93,192 ; minps %xmm8,%xmm0
DB 65,15,93,200 ; minps %xmm8,%xmm1
DB 65,15,93,208 ; minps %xmm8,%xmm2
@@ -10735,7 +10734,7 @@ _sk_clamp_1_sse41 LABEL PROC
PUBLIC _sk_clamp_a_sse41
_sk_clamp_a_sse41 LABEL PROC
- DB 15,93,29,228,46,0,0 ; minps 0x2ee4(%rip),%xmm3 # 3bf0 <_sk_callback_sse41+0x217>
+ DB 15,93,29,228,46,0,0 ; minps 0x2ee4(%rip),%xmm3 # 3bf0 <_sk_callback_sse41+0x20e>
DB 15,93,195 ; minps %xmm3,%xmm0
DB 15,93,203 ; minps %xmm3,%xmm1
DB 15,93,211 ; minps %xmm3,%xmm2
@@ -10808,7 +10807,7 @@ _sk_premul_sse41 LABEL PROC
PUBLIC _sk_unpremul_sse41
_sk_unpremul_sse41 LABEL PROC
DB 69,15,87,192 ; xorps %xmm8,%xmm8
- DB 68,15,40,13,79,46,0,0 ; movaps 0x2e4f(%rip),%xmm9 # 3c00 <_sk_callback_sse41+0x227>
+ DB 68,15,40,13,79,46,0,0 ; movaps 0x2e4f(%rip),%xmm9 # 3c00 <_sk_callback_sse41+0x21e>
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
@@ -10820,20 +10819,20 @@ _sk_unpremul_sse41 LABEL PROC
PUBLIC _sk_from_srgb_sse41
_sk_from_srgb_sse41 LABEL PROC
- DB 68,15,40,29,58,46,0,0 ; movaps 0x2e3a(%rip),%xmm11 # 3c10 <_sk_callback_sse41+0x237>
+ DB 68,15,40,29,58,46,0,0 ; movaps 0x2e3a(%rip),%xmm11 # 3c10 <_sk_callback_sse41+0x22e>
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 68,15,40,37,50,46,0,0 ; movaps 0x2e32(%rip),%xmm12 # 3c20 <_sk_callback_sse41+0x247>
+ DB 68,15,40,37,50,46,0,0 ; movaps 0x2e32(%rip),%xmm12 # 3c20 <_sk_callback_sse41+0x23e>
DB 68,15,40,192 ; movaps %xmm0,%xmm8
DB 69,15,89,196 ; mulps %xmm12,%xmm8
- DB 68,15,40,45,50,46,0,0 ; movaps 0x2e32(%rip),%xmm13 # 3c30 <_sk_callback_sse41+0x257>
+ DB 68,15,40,45,50,46,0,0 ; movaps 0x2e32(%rip),%xmm13 # 3c30 <_sk_callback_sse41+0x24e>
DB 69,15,88,197 ; addps %xmm13,%xmm8
DB 69,15,89,194 ; mulps %xmm10,%xmm8
- DB 68,15,40,53,50,46,0,0 ; movaps 0x2e32(%rip),%xmm14 # 3c40 <_sk_callback_sse41+0x267>
+ DB 68,15,40,53,50,46,0,0 ; movaps 0x2e32(%rip),%xmm14 # 3c40 <_sk_callback_sse41+0x25e>
DB 69,15,88,198 ; addps %xmm14,%xmm8
- DB 68,15,40,61,54,46,0,0 ; movaps 0x2e36(%rip),%xmm15 # 3c50 <_sk_callback_sse41+0x277>
+ DB 68,15,40,61,54,46,0,0 ; movaps 0x2e36(%rip),%xmm15 # 3c50 <_sk_callback_sse41+0x26e>
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
@@ -10877,20 +10876,20 @@ _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 68,15,40,29,163,45,0,0 ; movaps 0x2da3(%rip),%xmm11 # 3c60 <_sk_callback_sse41+0x287>
+ DB 68,15,40,29,163,45,0,0 ; movaps 0x2da3(%rip),%xmm11 # 3c60 <_sk_callback_sse41+0x27e>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 65,15,89,203 ; mulps %xmm11,%xmm1
- DB 68,15,40,37,164,45,0,0 ; movaps 0x2da4(%rip),%xmm12 # 3c70 <_sk_callback_sse41+0x297>
+ DB 68,15,40,37,164,45,0,0 ; movaps 0x2da4(%rip),%xmm12 # 3c70 <_sk_callback_sse41+0x28e>
DB 69,15,89,204 ; mulps %xmm12,%xmm9
- DB 68,15,40,45,168,45,0,0 ; movaps 0x2da8(%rip),%xmm13 # 3c80 <_sk_callback_sse41+0x2a7>
+ DB 68,15,40,45,168,45,0,0 ; movaps 0x2da8(%rip),%xmm13 # 3c80 <_sk_callback_sse41+0x29e>
DB 69,15,88,205 ; addps %xmm13,%xmm9
- DB 68,15,40,53,172,45,0,0 ; movaps 0x2dac(%rip),%xmm14 # 3c90 <_sk_callback_sse41+0x2b7>
+ DB 68,15,40,53,172,45,0,0 ; movaps 0x2dac(%rip),%xmm14 # 3c90 <_sk_callback_sse41+0x2ae>
DB 69,15,89,214 ; mulps %xmm14,%xmm10
DB 69,15,88,209 ; addps %xmm9,%xmm10
- DB 68,15,40,5,172,45,0,0 ; movaps 0x2dac(%rip),%xmm8 # 3ca0 <_sk_callback_sse41+0x2c7>
+ DB 68,15,40,5,172,45,0,0 ; movaps 0x2dac(%rip),%xmm8 # 3ca0 <_sk_callback_sse41+0x2be>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,93,202 ; minps %xmm10,%xmm9
- DB 68,15,40,61,172,45,0,0 ; movaps 0x2dac(%rip),%xmm15 # 3cb0 <_sk_callback_sse41+0x2d7>
+ DB 68,15,40,61,172,45,0,0 ; movaps 0x2dac(%rip),%xmm15 # 3cb0 <_sk_callback_sse41+0x2ce>
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
@@ -10943,7 +10942,7 @@ _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 68,15,40,53,250,44,0,0 ; movaps 0x2cfa(%rip),%xmm14 # 3cc0 <_sk_callback_sse41+0x2e7>
+ DB 68,15,40,53,250,44,0,0 ; movaps 0x2cfa(%rip),%xmm14 # 3cc0 <_sk_callback_sse41+0x2de>
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
@@ -10952,27 +10951,27 @@ _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 68,15,84,61,225,44,0,0 ; andps 0x2ce1(%rip),%xmm15 # 3cd0 <_sk_callback_sse41+0x2f7>
+ DB 68,15,84,61,225,44,0,0 ; andps 0x2ce1(%rip),%xmm15 # 3cd0 <_sk_callback_sse41+0x2ee>
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 68,15,40,45,212,44,0,0 ; movaps 0x2cd4(%rip),%xmm13 # 3ce0 <_sk_callback_sse41+0x307>
+ DB 68,15,40,45,212,44,0,0 ; movaps 0x2cd4(%rip),%xmm13 # 3ce0 <_sk_callback_sse41+0x2fe>
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 68,15,88,5,208,44,0,0 ; addps 0x2cd0(%rip),%xmm8 # 3cf0 <_sk_callback_sse41+0x317>
+ DB 68,15,88,5,208,44,0,0 ; addps 0x2cd0(%rip),%xmm8 # 3cf0 <_sk_callback_sse41+0x30e>
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 68,15,89,5,200,44,0,0 ; mulps 0x2cc8(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x327>
+ DB 68,15,89,5,200,44,0,0 ; mulps 0x2cc8(%rip),%xmm8 # 3d00 <_sk_callback_sse41+0x31e>
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 15,40,5,188,44,0,0 ; movaps 0x2cbc(%rip),%xmm0 # 3d10 <_sk_callback_sse41+0x337>
+ DB 15,40,5,188,44,0,0 ; movaps 0x2cbc(%rip),%xmm0 # 3d10 <_sk_callback_sse41+0x32e>
DB 65,15,40,211 ; movaps %xmm11,%xmm2
DB 15,89,208 ; mulps %xmm0,%xmm2
DB 15,194,194,1 ; cmpltps %xmm2,%xmm0
@@ -10999,140 +10998,141 @@ _sk_hsl_to_rgb_sse41 LABEL PROC
DB 15,41,28,36 ; movaps %xmm3,(%rsp)
DB 15,40,194 ; movaps %xmm2,%xmm0
DB 15,194,195,1 ; cmpltps %xmm3,%xmm0
- DB 15,40,45,97,44,0,0 ; movaps 0x2c61(%rip),%xmm5 # 3d20 <_sk_callback_sse41+0x347>
- DB 15,40,241 ; movaps %xmm1,%xmm6
+ DB 15,40,45,97,44,0,0 ; movaps 0x2c61(%rip),%xmm5 # 3d20 <_sk_callback_sse41+0x33e>
+ DB 15,40,249 ; movaps %xmm1,%xmm7
DB 15,40,225 ; movaps %xmm1,%xmm4
DB 15,40,217 ; movaps %xmm1,%xmm3
DB 15,88,221 ; addps %xmm5,%xmm3
- DB 15,40,253 ; movaps %xmm5,%xmm7
+ DB 15,40,245 ; movaps %xmm5,%xmm6
DB 15,89,218 ; mulps %xmm2,%xmm3
- DB 15,88,242 ; addps %xmm2,%xmm6
+ DB 15,88,250 ; addps %xmm2,%xmm7
DB 15,89,226 ; mulps %xmm2,%xmm4
DB 15,40,234 ; movaps %xmm2,%xmm5
- DB 15,92,244 ; subps %xmm4,%xmm6
- DB 102,15,56,20,243 ; blendvps %xmm0,%xmm3,%xmm6
- DB 68,15,40,61,70,44,0,0 ; movaps 0x2c46(%rip),%xmm15 # 3d30 <_sk_callback_sse41+0x357>
- DB 69,15,88,251 ; addps %xmm11,%xmm15
+ DB 15,92,252 ; subps %xmm4,%xmm7
+ DB 102,15,56,20,251 ; blendvps %xmm0,%xmm3,%xmm7
+ DB 68,15,40,37,70,44,0,0 ; movaps 0x2c46(%rip),%xmm12 # 3d30 <_sk_callback_sse41+0x34e>
+ DB 69,15,88,227 ; addps %xmm11,%xmm12
DB 184,0,0,0,0 ; mov $0x0,%eax
DB 185,0,0,128,63 ; mov $0x3f800000,%ecx
DB 102,68,15,110,201 ; movd %ecx,%xmm9
DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9
DB 65,15,40,193 ; movaps %xmm9,%xmm0
- DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0
- DB 65,15,40,215 ; movaps %xmm15,%xmm2
- DB 15,88,21,42,44,0,0 ; addps 0x2c2a(%rip),%xmm2 # 3d40 <_sk_callback_sse41+0x367>
- DB 69,15,40,231 ; movaps %xmm15,%xmm12
- DB 102,68,15,56,20,226 ; blendvps %xmm0,%xmm2,%xmm12
- DB 102,15,110,208 ; movd %eax,%xmm2
- DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2
- DB 15,41,84,36,32 ; movaps %xmm2,0x20(%rsp)
- DB 65,15,40,199 ; movaps %xmm15,%xmm0
- DB 15,194,194,1 ; cmpltps %xmm2,%xmm0
- DB 65,15,40,215 ; movaps %xmm15,%xmm2
- DB 15,88,215 ; addps %xmm7,%xmm2
+ DB 65,15,194,196,1 ; cmpltps %xmm12,%xmm0
+ DB 65,15,40,212 ; movaps %xmm12,%xmm2
+ DB 15,88,21,42,44,0,0 ; addps 0x2c2a(%rip),%xmm2 # 3d40 <_sk_callback_sse41+0x35e>
+ DB 69,15,40,196 ; movaps %xmm12,%xmm8
+ DB 65,15,40,220 ; movaps %xmm12,%xmm3
DB 102,68,15,56,20,226 ; blendvps %xmm0,%xmm2,%xmm12
- DB 15,40,221 ; movaps %xmm5,%xmm3
- DB 15,41,92,36,48 ; movaps %xmm3,0x30(%rsp)
- DB 68,15,40,195 ; movaps %xmm3,%xmm8
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
+ DB 15,41,68,36,32 ; movaps %xmm0,0x20(%rsp)
+ DB 68,15,194,192,1 ; cmpltps %xmm0,%xmm8
+ DB 15,88,222 ; addps %xmm6,%xmm3
+ DB 65,15,40,192 ; movaps %xmm8,%xmm0
+ DB 102,68,15,56,20,227 ; blendvps %xmm0,%xmm3,%xmm12
+ DB 15,40,213 ; movaps %xmm5,%xmm2
+ DB 15,41,84,36,48 ; movaps %xmm2,0x30(%rsp)
+ DB 68,15,40,194 ; movaps %xmm2,%xmm8
DB 69,15,88,192 ; addps %xmm8,%xmm8
- DB 68,15,92,198 ; subps %xmm6,%xmm8
+ DB 68,15,92,199 ; subps %xmm7,%xmm8
DB 184,171,170,42,62 ; mov $0x3e2aaaab,%eax
- DB 15,40,214 ; movaps %xmm6,%xmm2
- DB 65,15,92,208 ; subps %xmm8,%xmm2
- DB 15,89,21,231,43,0,0 ; mulps 0x2be7(%rip),%xmm2 # 3d50 <_sk_callback_sse41+0x377>
+ DB 15,40,247 ; movaps %xmm7,%xmm6
+ DB 65,15,92,240 ; subps %xmm8,%xmm6
+ DB 15,89,53,230,43,0,0 ; mulps 0x2be6(%rip),%xmm6 # 3d50 <_sk_callback_sse41+0x36e>
DB 185,171,170,42,63 ; mov $0x3f2aaaab,%ecx
DB 102,15,110,193 ; movd %ecx,%xmm0
DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
DB 15,41,68,36,16 ; movaps %xmm0,0x10(%rsp)
- DB 15,40,37,222,43,0,0 ; movaps 0x2bde(%rip),%xmm4 # 3d60 <_sk_callback_sse41+0x387>
+ DB 15,40,37,221,43,0,0 ; movaps 0x2bdd(%rip),%xmm4 # 3d60 <_sk_callback_sse41+0x37e>
DB 15,40,236 ; movaps %xmm4,%xmm5
DB 65,15,92,236 ; subps %xmm12,%xmm5
DB 69,15,40,236 ; movaps %xmm12,%xmm13
+ DB 69,15,40,252 ; movaps %xmm12,%xmm15
DB 69,15,40,244 ; movaps %xmm12,%xmm14
DB 68,15,194,224,1 ; cmpltps %xmm0,%xmm12
- DB 15,89,234 ; mulps %xmm2,%xmm5
+ DB 15,89,238 ; mulps %xmm6,%xmm5
DB 65,15,88,232 ; addps %xmm8,%xmm5
DB 69,15,40,208 ; movaps %xmm8,%xmm10
DB 65,15,40,196 ; movaps %xmm12,%xmm0
DB 102,68,15,56,20,213 ; blendvps %xmm0,%xmm5,%xmm10
- DB 15,40,44,36 ; movaps (%rsp),%xmm5
- DB 68,15,194,245,1 ; cmpltps %xmm5,%xmm14
+ DB 68,15,194,52,36,1 ; cmpltps (%rsp),%xmm14
DB 65,15,40,198 ; movaps %xmm14,%xmm0
- DB 102,68,15,56,20,214 ; blendvps %xmm0,%xmm6,%xmm10
- DB 102,15,110,248 ; movd %eax,%xmm7
- DB 15,198,255,0 ; shufps $0x0,%xmm7,%xmm7
- DB 68,15,194,239,1 ; cmpltps %xmm7,%xmm13
- DB 68,15,89,250 ; mulps %xmm2,%xmm15
+ DB 102,68,15,56,20,215 ; blendvps %xmm0,%xmm7,%xmm10
+ DB 102,15,110,232 ; movd %eax,%xmm5
+ DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5
+ DB 68,15,194,237,1 ; cmpltps %xmm5,%xmm13
+ DB 68,15,89,254 ; mulps %xmm6,%xmm15
DB 69,15,88,248 ; addps %xmm8,%xmm15
DB 65,15,40,197 ; movaps %xmm13,%xmm0
DB 102,69,15,56,20,215 ; blendvps %xmm0,%xmm15,%xmm10
DB 69,15,87,228 ; xorps %xmm12,%xmm12
DB 68,15,194,225,0 ; cmpeqps %xmm1,%xmm12
DB 65,15,40,196 ; movaps %xmm12,%xmm0
- DB 102,68,15,56,20,211 ; blendvps %xmm0,%xmm3,%xmm10
+ DB 102,68,15,56,20,210 ; blendvps %xmm0,%xmm2,%xmm10
DB 65,15,40,193 ; movaps %xmm9,%xmm0
DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0
DB 65,15,40,203 ; movaps %xmm11,%xmm1
- DB 15,88,13,60,43,0,0 ; addps 0x2b3c(%rip),%xmm1 # 3d40 <_sk_callback_sse41+0x367>
+ DB 15,88,13,58,43,0,0 ; addps 0x2b3a(%rip),%xmm1 # 3d40 <_sk_callback_sse41+0x35e>
DB 69,15,40,235 ; movaps %xmm11,%xmm13
DB 102,68,15,56,20,233 ; blendvps %xmm0,%xmm1,%xmm13
DB 65,15,40,195 ; movaps %xmm11,%xmm0
DB 15,194,68,36,32,1 ; cmpltps 0x20(%rsp),%xmm0
DB 65,15,40,203 ; movaps %xmm11,%xmm1
- DB 15,88,13,253,42,0,0 ; addps 0x2afd(%rip),%xmm1 # 3d20 <_sk_callback_sse41+0x347>
+ DB 15,88,13,251,42,0,0 ; addps 0x2afb(%rip),%xmm1 # 3d20 <_sk_callback_sse41+0x33e>
DB 102,68,15,56,20,233 ; blendvps %xmm0,%xmm1,%xmm13
DB 15,40,220 ; movaps %xmm4,%xmm3
DB 65,15,92,221 ; subps %xmm13,%xmm3
+ DB 65,15,40,213 ; movaps %xmm13,%xmm2
DB 69,15,40,245 ; movaps %xmm13,%xmm14
DB 69,15,40,253 ; movaps %xmm13,%xmm15
DB 68,15,194,108,36,16,1 ; cmpltps 0x10(%rsp),%xmm13
- DB 15,89,218 ; mulps %xmm2,%xmm3
+ DB 15,89,222 ; mulps %xmm6,%xmm3
DB 65,15,88,216 ; addps %xmm8,%xmm3
DB 65,15,40,200 ; movaps %xmm8,%xmm1
DB 65,15,40,197 ; movaps %xmm13,%xmm0
DB 102,15,56,20,203 ; blendvps %xmm0,%xmm3,%xmm1
- DB 68,15,194,253,1 ; cmpltps %xmm5,%xmm15
+ DB 68,15,194,60,36,1 ; cmpltps (%rsp),%xmm15
DB 65,15,40,199 ; movaps %xmm15,%xmm0
- DB 102,15,56,20,206 ; blendvps %xmm0,%xmm6,%xmm1
- DB 68,15,194,247,1 ; cmpltps %xmm7,%xmm14
- DB 15,40,218 ; movaps %xmm2,%xmm3
- DB 65,15,89,219 ; mulps %xmm11,%xmm3
- DB 65,15,88,216 ; addps %xmm8,%xmm3
- DB 65,15,40,198 ; movaps %xmm14,%xmm0
- DB 102,15,56,20,203 ; blendvps %xmm0,%xmm3,%xmm1
+ DB 102,15,56,20,207 ; blendvps %xmm0,%xmm7,%xmm1
+ DB 15,194,213,1 ; cmpltps %xmm5,%xmm2
+ DB 68,15,89,246 ; mulps %xmm6,%xmm14
+ DB 69,15,88,240 ; addps %xmm8,%xmm14
+ DB 15,40,194 ; movaps %xmm2,%xmm0
+ DB 102,65,15,56,20,206 ; blendvps %xmm0,%xmm14,%xmm1
DB 65,15,40,196 ; movaps %xmm12,%xmm0
- DB 15,40,108,36,48 ; movaps 0x30(%rsp),%xmm5
- DB 102,15,56,20,205 ; blendvps %xmm0,%xmm5,%xmm1
- DB 68,15,88,29,224,42,0,0 ; addps 0x2ae0(%rip),%xmm11 # 3d70 <_sk_callback_sse41+0x397>
+ DB 68,15,40,116,36,48 ; movaps 0x30(%rsp),%xmm14
+ DB 102,65,15,56,20,206 ; blendvps %xmm0,%xmm14,%xmm1
+ DB 68,15,88,29,219,42,0,0 ; addps 0x2adb(%rip),%xmm11 # 3d70 <_sk_callback_sse41+0x38e>
+ DB 15,40,21,132,42,0,0 ; movaps 0x2a84(%rip),%xmm2 # 3d20 <_sk_callback_sse41+0x33e>
+ DB 65,15,88,211 ; addps %xmm11,%xmm2
DB 69,15,194,203,1 ; cmpltps %xmm11,%xmm9
- DB 15,40,29,164,42,0,0 ; movaps 0x2aa4(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x367>
+ DB 15,40,29,148,42,0,0 ; movaps 0x2a94(%rip),%xmm3 # 3d40 <_sk_callback_sse41+0x35e>
DB 65,15,88,219 ; addps %xmm11,%xmm3
DB 69,15,40,235 ; movaps %xmm11,%xmm13
DB 65,15,40,193 ; movaps %xmm9,%xmm0
- DB 102,68,15,56,20,235 ; blendvps %xmm0,%xmm3,%xmm13
- DB 65,15,40,195 ; movaps %xmm11,%xmm0
- DB 15,194,68,36,32,1 ; cmpltps 0x20(%rsp),%xmm0
- DB 15,40,29,97,42,0,0 ; movaps 0x2a61(%rip),%xmm3 # 3d20 <_sk_callback_sse41+0x347>
- DB 65,15,88,219 ; addps %xmm11,%xmm3
- DB 102,68,15,56,20,235 ; blendvps %xmm0,%xmm3,%xmm13
- DB 65,15,92,229 ; subps %xmm13,%xmm4
- DB 69,15,40,205 ; movaps %xmm13,%xmm9
- DB 69,15,40,245 ; movaps %xmm13,%xmm14
- DB 68,15,194,108,36,16,1 ; cmpltps 0x10(%rsp),%xmm13
- DB 68,15,89,218 ; mulps %xmm2,%xmm11
- DB 15,89,226 ; mulps %xmm2,%xmm4
- DB 69,15,88,216 ; addps %xmm8,%xmm11
- DB 65,15,88,224 ; addps %xmm8,%xmm4
+ DB 102,68,15,56,20,219 ; blendvps %xmm0,%xmm3,%xmm11
+ DB 68,15,194,108,36,32,1 ; cmpltps 0x20(%rsp),%xmm13
DB 65,15,40,197 ; movaps %xmm13,%xmm0
+ DB 102,68,15,56,20,218 ; blendvps %xmm0,%xmm2,%xmm11
+ DB 65,15,92,227 ; subps %xmm11,%xmm4
+ DB 69,15,40,203 ; movaps %xmm11,%xmm9
+ DB 65,15,40,211 ; movaps %xmm11,%xmm2
+ DB 69,15,40,235 ; movaps %xmm11,%xmm13
+ DB 68,15,194,92,36,16,1 ; cmpltps 0x10(%rsp),%xmm11
+ DB 15,89,214 ; mulps %xmm6,%xmm2
+ DB 15,89,230 ; mulps %xmm6,%xmm4
+ DB 65,15,88,208 ; addps %xmm8,%xmm2
+ DB 65,15,88,224 ; addps %xmm8,%xmm4
+ DB 65,15,40,195 ; movaps %xmm11,%xmm0
DB 102,68,15,56,20,196 ; blendvps %xmm0,%xmm4,%xmm8
- DB 68,15,194,52,36,1 ; cmpltps (%rsp),%xmm14
- DB 65,15,40,198 ; movaps %xmm14,%xmm0
- DB 102,68,15,56,20,198 ; blendvps %xmm0,%xmm6,%xmm8
- DB 68,15,194,207,1 ; cmpltps %xmm7,%xmm9
+ DB 68,15,194,44,36,1 ; cmpltps (%rsp),%xmm13
+ DB 65,15,40,197 ; movaps %xmm13,%xmm0
+ DB 102,68,15,56,20,199 ; blendvps %xmm0,%xmm7,%xmm8
+ DB 68,15,194,205,1 ; cmpltps %xmm5,%xmm9
DB 65,15,40,193 ; movaps %xmm9,%xmm0
- DB 102,69,15,56,20,195 ; blendvps %xmm0,%xmm11,%xmm8
+ DB 102,68,15,56,20,194 ; blendvps %xmm0,%xmm2,%xmm8
DB 65,15,40,196 ; movaps %xmm12,%xmm0
- DB 102,68,15,56,20,197 ; blendvps %xmm0,%xmm5,%xmm8
+ DB 102,69,15,56,20,198 ; blendvps %xmm0,%xmm14,%xmm8
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,194 ; movaps %xmm10,%xmm0
DB 65,15,40,208 ; movaps %xmm8,%xmm2
@@ -11162,7 +11162,7 @@ _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 68,15,89,5,251,41,0,0 ; mulps 0x29fb(%rip),%xmm8 # 3d80 <_sk_callback_sse41+0x3a7>
+ DB 68,15,89,5,242,41,0,0 ; mulps 0x29f2(%rip),%xmm8 # 3d80 <_sk_callback_sse41+0x39e>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 65,15,89,208 ; mulps %xmm8,%xmm2
@@ -11196,7 +11196,7 @@ _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 68,15,89,5,167,41,0,0 ; mulps 0x29a7(%rip),%xmm8 # 3d90 <_sk_callback_sse41+0x3b7>
+ DB 68,15,89,5,158,41,0,0 ; mulps 0x299e(%rip),%xmm8 # 3d90 <_sk_callback_sse41+0x3ae>
DB 15,92,196 ; subps %xmm4,%xmm0
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -11217,17 +11217,17 @@ _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 102,15,111,29,119,41,0,0 ; movdqa 0x2977(%rip),%xmm3 # 3da0 <_sk_callback_sse41+0x3c7>
+ DB 102,15,111,29,110,41,0,0 ; movdqa 0x296e(%rip),%xmm3 # 3da0 <_sk_callback_sse41+0x3be>
DB 102,65,15,219,216 ; pand %xmm8,%xmm3
DB 68,15,91,203 ; cvtdq2ps %xmm3,%xmm9
- DB 68,15,89,13,118,41,0,0 ; mulps 0x2976(%rip),%xmm9 # 3db0 <_sk_callback_sse41+0x3d7>
- DB 102,15,111,29,126,41,0,0 ; movdqa 0x297e(%rip),%xmm3 # 3dc0 <_sk_callback_sse41+0x3e7>
+ DB 68,15,89,13,109,41,0,0 ; mulps 0x296d(%rip),%xmm9 # 3db0 <_sk_callback_sse41+0x3ce>
+ DB 102,15,111,29,117,41,0,0 ; movdqa 0x2975(%rip),%xmm3 # 3dc0 <_sk_callback_sse41+0x3de>
DB 102,65,15,219,216 ; pand %xmm8,%xmm3
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,127,41,0,0 ; mulps 0x297f(%rip),%xmm3 # 3dd0 <_sk_callback_sse41+0x3f7>
- DB 102,68,15,219,5,134,41,0,0 ; pand 0x2986(%rip),%xmm8 # 3de0 <_sk_callback_sse41+0x407>
+ DB 15,89,29,118,41,0,0 ; mulps 0x2976(%rip),%xmm3 # 3dd0 <_sk_callback_sse41+0x3ee>
+ DB 102,68,15,219,5,125,41,0,0 ; pand 0x297d(%rip),%xmm8 # 3de0 <_sk_callback_sse41+0x3fe>
DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
- DB 68,15,89,5,138,41,0,0 ; mulps 0x298a(%rip),%xmm8 # 3df0 <_sk_callback_sse41+0x417>
+ DB 68,15,89,5,129,41,0,0 ; mulps 0x2981(%rip),%xmm8 # 3df0 <_sk_callback_sse41+0x40e>
DB 15,92,196 ; subps %xmm4,%xmm0
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 15,88,196 ; addps %xmm4,%xmm0
@@ -11238,7 +11238,7 @@ _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 15,40,29,116,41,0,0 ; movaps 0x2974(%rip),%xmm3 # 3e00 <_sk_callback_sse41+0x427>
+ DB 15,40,29,107,41,0,0 ; movaps 0x296b(%rip),%xmm3 # 3e00 <_sk_callback_sse41+0x41e>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_load_tables_sse41
@@ -11247,7 +11247,7 @@ _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 102,15,111,5,107,41,0,0 ; movdqa 0x296b(%rip),%xmm0 # 3e10 <_sk_callback_sse41+0x437>
+ DB 102,15,111,5,98,41,0,0 ; movdqa 0x2962(%rip),%xmm0 # 3e10 <_sk_callback_sse41+0x42e>
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
@@ -11262,7 +11262,7 @@ _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 102,15,56,0,13,38,41,0,0 ; pshufb 0x2926(%rip),%xmm1 # 3e20 <_sk_callback_sse41+0x447>
+ DB 102,15,56,0,13,29,41,0,0 ; pshufb 0x291d(%rip),%xmm1 # 3e20 <_sk_callback_sse41+0x43e>
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
@@ -11277,7 +11277,7 @@ _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 102,15,56,0,21,226,40,0,0 ; pshufb 0x28e2(%rip),%xmm2 # 3e30 <_sk_callback_sse41+0x457>
+ DB 102,15,56,0,21,217,40,0,0 ; pshufb 0x28d9(%rip),%xmm2 # 3e30 <_sk_callback_sse41+0x44e>
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
@@ -11292,7 +11292,7 @@ _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 15,89,29,159,40,0,0 ; mulps 0x289f(%rip),%xmm3 # 3e40 <_sk_callback_sse41+0x467>
+ DB 15,89,29,150,40,0,0 ; mulps 0x2896(%rip),%xmm3 # 3e40 <_sk_callback_sse41+0x45e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -11309,7 +11309,7 @@ _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 102,68,15,111,5,114,40,0,0 ; movdqa 0x2872(%rip),%xmm8 # 3e50 <_sk_callback_sse41+0x477>
+ DB 102,68,15,111,5,105,40,0,0 ; movdqa 0x2869(%rip),%xmm8 # 3e50 <_sk_callback_sse41+0x46e>
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
@@ -11326,7 +11326,7 @@ _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 102,15,56,0,13,37,40,0,0 ; pshufb 0x2825(%rip),%xmm1 # 3e60 <_sk_callback_sse41+0x487>
+ DB 102,15,56,0,13,28,40,0,0 ; pshufb 0x281c(%rip),%xmm1 # 3e60 <_sk_callback_sse41+0x47e>
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
@@ -11362,7 +11362,7 @@ _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 15,89,29,115,39,0,0 ; mulps 0x2773(%rip),%xmm3 # 3e70 <_sk_callback_sse41+0x497>
+ DB 15,89,29,106,39,0,0 ; mulps 0x276a(%rip),%xmm3 # 3e70 <_sk_callback_sse41+0x48e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -11382,7 +11382,7 @@ _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 102,68,15,111,5,53,39,0,0 ; movdqa 0x2735(%rip),%xmm8 # 3e80 <_sk_callback_sse41+0x4a7>
+ DB 102,68,15,111,5,44,39,0,0 ; movdqa 0x272c(%rip),%xmm8 # 3e80 <_sk_callback_sse41+0x49e>
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
@@ -11399,7 +11399,7 @@ _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 102,15,56,0,13,232,38,0,0 ; pshufb 0x26e8(%rip),%xmm1 # 3e90 <_sk_callback_sse41+0x4b7>
+ DB 102,15,56,0,13,223,38,0,0 ; pshufb 0x26df(%rip),%xmm1 # 3e90 <_sk_callback_sse41+0x4ae>
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
@@ -11430,7 +11430,7 @@ _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 15,40,29,83,38,0,0 ; movaps 0x2653(%rip),%xmm3 # 3ea0 <_sk_callback_sse41+0x4c7>
+ DB 15,40,29,74,38,0,0 ; movaps 0x264a(%rip),%xmm3 # 3ea0 <_sk_callback_sse41+0x4be>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_byte_tables_sse41
@@ -11438,7 +11438,7 @@ _sk_byte_tables_sse41 LABEL PROC
DB 65,86 ; push %r14
DB 83 ; push %rbx
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 68,15,40,5,84,38,0,0 ; movaps 0x2654(%rip),%xmm8 # 3eb0 <_sk_callback_sse41+0x4d7>
+ DB 68,15,40,5,75,38,0,0 ; movaps 0x264b(%rip),%xmm8 # 3eb0 <_sk_callback_sse41+0x4ce>
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
@@ -11457,7 +11457,7 @@ _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 68,15,40,13,5,38,0,0 ; movaps 0x2605(%rip),%xmm9 # 3ec0 <_sk_callback_sse41+0x4e7>
+ DB 68,15,40,13,252,37,0,0 ; movaps 0x25fc(%rip),%xmm9 # 3ec0 <_sk_callback_sse41+0x4de>
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
@@ -11546,7 +11546,7 @@ _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 68,15,40,13,141,36,0,0 ; movaps 0x248d(%rip),%xmm9 # 3ed0 <_sk_callback_sse41+0x4f7>
+ DB 68,15,40,13,132,36,0,0 ; movaps 0x2484(%rip),%xmm9 # 3ed0 <_sk_callback_sse41+0x4ee>
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
@@ -11713,31 +11713,31 @@ _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 68,15,89,5,228,33,0,0 ; mulps 0x21e4(%rip),%xmm8 # 3ee0 <_sk_callback_sse41+0x507>
- DB 68,15,84,21,236,33,0,0 ; andps 0x21ec(%rip),%xmm10 # 3ef0 <_sk_callback_sse41+0x517>
- DB 68,15,86,21,244,33,0,0 ; orps 0x21f4(%rip),%xmm10 # 3f00 <_sk_callback_sse41+0x527>
- DB 68,15,88,5,252,33,0,0 ; addps 0x21fc(%rip),%xmm8 # 3f10 <_sk_callback_sse41+0x537>
- DB 68,15,40,37,4,34,0,0 ; movaps 0x2204(%rip),%xmm12 # 3f20 <_sk_callback_sse41+0x547>
+ DB 68,15,89,5,219,33,0,0 ; mulps 0x21db(%rip),%xmm8 # 3ee0 <_sk_callback_sse41+0x4fe>
+ DB 68,15,84,21,227,33,0,0 ; andps 0x21e3(%rip),%xmm10 # 3ef0 <_sk_callback_sse41+0x50e>
+ DB 68,15,86,21,235,33,0,0 ; orps 0x21eb(%rip),%xmm10 # 3f00 <_sk_callback_sse41+0x51e>
+ DB 68,15,88,5,243,33,0,0 ; addps 0x21f3(%rip),%xmm8 # 3f10 <_sk_callback_sse41+0x52e>
+ DB 68,15,40,37,251,33,0,0 ; movaps 0x21fb(%rip),%xmm12 # 3f20 <_sk_callback_sse41+0x53e>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,196 ; subps %xmm12,%xmm8
- DB 68,15,88,21,4,34,0,0 ; addps 0x2204(%rip),%xmm10 # 3f30 <_sk_callback_sse41+0x557>
- DB 68,15,40,37,12,34,0,0 ; movaps 0x220c(%rip),%xmm12 # 3f40 <_sk_callback_sse41+0x567>
+ DB 68,15,88,21,251,33,0,0 ; addps 0x21fb(%rip),%xmm10 # 3f30 <_sk_callback_sse41+0x54e>
+ DB 68,15,40,37,3,34,0,0 ; movaps 0x2203(%rip),%xmm12 # 3f40 <_sk_callback_sse41+0x55e>
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 68,15,88,5,249,33,0,0 ; addps 0x21f9(%rip),%xmm8 # 3f50 <_sk_callback_sse41+0x577>
- DB 68,15,40,21,1,34,0,0 ; movaps 0x2201(%rip),%xmm10 # 3f60 <_sk_callback_sse41+0x587>
+ DB 68,15,88,5,240,33,0,0 ; addps 0x21f0(%rip),%xmm8 # 3f50 <_sk_callback_sse41+0x56e>
+ DB 68,15,40,21,248,33,0,0 ; movaps 0x21f8(%rip),%xmm10 # 3f60 <_sk_callback_sse41+0x57e>
DB 69,15,89,211 ; mulps %xmm11,%xmm10
DB 69,15,92,194 ; subps %xmm10,%xmm8
- DB 68,15,40,21,1,34,0,0 ; movaps 0x2201(%rip),%xmm10 # 3f70 <_sk_callback_sse41+0x597>
+ DB 68,15,40,21,248,33,0,0 ; movaps 0x21f8(%rip),%xmm10 # 3f70 <_sk_callback_sse41+0x58e>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 68,15,40,29,5,34,0,0 ; movaps 0x2205(%rip),%xmm11 # 3f80 <_sk_callback_sse41+0x5a7>
+ DB 68,15,40,29,252,33,0,0 ; movaps 0x21fc(%rip),%xmm11 # 3f80 <_sk_callback_sse41+0x59e>
DB 69,15,94,218 ; divps %xmm10,%xmm11
DB 69,15,88,216 ; addps %xmm8,%xmm11
- DB 68,15,89,29,5,34,0,0 ; mulps 0x2205(%rip),%xmm11 # 3f90 <_sk_callback_sse41+0x5b7>
+ DB 68,15,89,29,252,33,0,0 ; mulps 0x21fc(%rip),%xmm11 # 3f90 <_sk_callback_sse41+0x5ae>
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
@@ -11745,7 +11745,7 @@ _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 68,15,93,5,236,33,0,0 ; minps 0x21ec(%rip),%xmm8 # 3fa0 <_sk_callback_sse41+0x5c7>
+ DB 68,15,93,5,227,33,0,0 ; minps 0x21e3(%rip),%xmm8 # 3fa0 <_sk_callback_sse41+0x5be>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -11773,31 +11773,31 @@ _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 68,15,89,37,141,33,0,0 ; mulps 0x218d(%rip),%xmm12 # 3fb0 <_sk_callback_sse41+0x5d7>
- DB 68,15,84,29,149,33,0,0 ; andps 0x2195(%rip),%xmm11 # 3fc0 <_sk_callback_sse41+0x5e7>
- DB 68,15,86,29,157,33,0,0 ; orps 0x219d(%rip),%xmm11 # 3fd0 <_sk_callback_sse41+0x5f7>
- DB 68,15,88,37,165,33,0,0 ; addps 0x21a5(%rip),%xmm12 # 3fe0 <_sk_callback_sse41+0x607>
- DB 15,40,13,174,33,0,0 ; movaps 0x21ae(%rip),%xmm1 # 3ff0 <_sk_callback_sse41+0x617>
+ DB 68,15,89,37,132,33,0,0 ; mulps 0x2184(%rip),%xmm12 # 3fb0 <_sk_callback_sse41+0x5ce>
+ DB 68,15,84,29,140,33,0,0 ; andps 0x218c(%rip),%xmm11 # 3fc0 <_sk_callback_sse41+0x5de>
+ DB 68,15,86,29,148,33,0,0 ; orps 0x2194(%rip),%xmm11 # 3fd0 <_sk_callback_sse41+0x5ee>
+ DB 68,15,88,37,156,33,0,0 ; addps 0x219c(%rip),%xmm12 # 3fe0 <_sk_callback_sse41+0x5fe>
+ DB 15,40,13,165,33,0,0 ; movaps 0x21a5(%rip),%xmm1 # 3ff0 <_sk_callback_sse41+0x60e>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
- DB 68,15,88,29,174,33,0,0 ; addps 0x21ae(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x627>
- DB 15,40,13,183,33,0,0 ; movaps 0x21b7(%rip),%xmm1 # 4010 <_sk_callback_sse41+0x637>
+ DB 68,15,88,29,165,33,0,0 ; addps 0x21a5(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x61e>
+ DB 15,40,13,174,33,0,0 ; movaps 0x21ae(%rip),%xmm1 # 4010 <_sk_callback_sse41+0x62e>
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 68,15,88,37,164,33,0,0 ; addps 0x21a4(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x647>
- DB 15,40,13,173,33,0,0 ; movaps 0x21ad(%rip),%xmm1 # 4030 <_sk_callback_sse41+0x657>
+ DB 68,15,88,37,155,33,0,0 ; addps 0x219b(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x63e>
+ DB 15,40,13,164,33,0,0 ; movaps 0x21a4(%rip),%xmm1 # 4030 <_sk_callback_sse41+0x64e>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
- DB 68,15,40,21,173,33,0,0 ; movaps 0x21ad(%rip),%xmm10 # 4040 <_sk_callback_sse41+0x667>
+ DB 68,15,40,21,164,33,0,0 ; movaps 0x21a4(%rip),%xmm10 # 4040 <_sk_callback_sse41+0x65e>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,13,178,33,0,0 ; movaps 0x21b2(%rip),%xmm1 # 4050 <_sk_callback_sse41+0x677>
+ DB 15,40,13,169,33,0,0 ; movaps 0x21a9(%rip),%xmm1 # 4050 <_sk_callback_sse41+0x66e>
DB 65,15,94,202 ; divps %xmm10,%xmm1
DB 65,15,88,204 ; addps %xmm12,%xmm1
- DB 15,89,13,179,33,0,0 ; mulps 0x21b3(%rip),%xmm1 # 4060 <_sk_callback_sse41+0x687>
+ DB 15,89,13,170,33,0,0 ; mulps 0x21aa(%rip),%xmm1 # 4060 <_sk_callback_sse41+0x67e>
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
@@ -11805,7 +11805,7 @@ _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 15,93,13,158,33,0,0 ; minps 0x219e(%rip),%xmm1 # 4070 <_sk_callback_sse41+0x697>
+ DB 15,93,13,149,33,0,0 ; minps 0x2195(%rip),%xmm1 # 4070 <_sk_callback_sse41+0x68e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -11833,31 +11833,31 @@ _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 68,15,89,37,63,33,0,0 ; mulps 0x213f(%rip),%xmm12 # 4080 <_sk_callback_sse41+0x6a7>
- DB 68,15,84,29,71,33,0,0 ; andps 0x2147(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x6b7>
- DB 68,15,86,29,79,33,0,0 ; orps 0x214f(%rip),%xmm11 # 40a0 <_sk_callback_sse41+0x6c7>
- DB 68,15,88,37,87,33,0,0 ; addps 0x2157(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x6d7>
- DB 15,40,21,96,33,0,0 ; movaps 0x2160(%rip),%xmm2 # 40c0 <_sk_callback_sse41+0x6e7>
+ DB 68,15,89,37,54,33,0,0 ; mulps 0x2136(%rip),%xmm12 # 4080 <_sk_callback_sse41+0x69e>
+ DB 68,15,84,29,62,33,0,0 ; andps 0x213e(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x6ae>
+ DB 68,15,86,29,70,33,0,0 ; orps 0x2146(%rip),%xmm11 # 40a0 <_sk_callback_sse41+0x6be>
+ DB 68,15,88,37,78,33,0,0 ; addps 0x214e(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x6ce>
+ DB 15,40,21,87,33,0,0 ; movaps 0x2157(%rip),%xmm2 # 40c0 <_sk_callback_sse41+0x6de>
DB 65,15,89,211 ; mulps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
- DB 68,15,88,29,96,33,0,0 ; addps 0x2160(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x6f7>
- DB 15,40,21,105,33,0,0 ; movaps 0x2169(%rip),%xmm2 # 40e0 <_sk_callback_sse41+0x707>
+ DB 68,15,88,29,87,33,0,0 ; addps 0x2157(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x6ee>
+ DB 15,40,21,96,33,0,0 ; movaps 0x2160(%rip),%xmm2 # 40e0 <_sk_callback_sse41+0x6fe>
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 68,15,88,37,86,33,0,0 ; addps 0x2156(%rip),%xmm12 # 40f0 <_sk_callback_sse41+0x717>
- DB 15,40,21,95,33,0,0 ; movaps 0x215f(%rip),%xmm2 # 4100 <_sk_callback_sse41+0x727>
+ DB 68,15,88,37,77,33,0,0 ; addps 0x214d(%rip),%xmm12 # 40f0 <_sk_callback_sse41+0x70e>
+ DB 15,40,21,86,33,0,0 ; movaps 0x2156(%rip),%xmm2 # 4100 <_sk_callback_sse41+0x71e>
DB 65,15,89,211 ; mulps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
- DB 68,15,40,21,95,33,0,0 ; movaps 0x215f(%rip),%xmm10 # 4110 <_sk_callback_sse41+0x737>
+ DB 68,15,40,21,86,33,0,0 ; movaps 0x2156(%rip),%xmm10 # 4110 <_sk_callback_sse41+0x72e>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,21,100,33,0,0 ; movaps 0x2164(%rip),%xmm2 # 4120 <_sk_callback_sse41+0x747>
+ DB 15,40,21,91,33,0,0 ; movaps 0x215b(%rip),%xmm2 # 4120 <_sk_callback_sse41+0x73e>
DB 65,15,94,210 ; divps %xmm10,%xmm2
DB 65,15,88,212 ; addps %xmm12,%xmm2
- DB 15,89,21,101,33,0,0 ; mulps 0x2165(%rip),%xmm2 # 4130 <_sk_callback_sse41+0x757>
+ DB 15,89,21,92,33,0,0 ; mulps 0x215c(%rip),%xmm2 # 4130 <_sk_callback_sse41+0x74e>
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
@@ -11865,7 +11865,7 @@ _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 15,93,21,80,33,0,0 ; minps 0x2150(%rip),%xmm2 # 4140 <_sk_callback_sse41+0x767>
+ DB 15,93,21,71,33,0,0 ; minps 0x2147(%rip),%xmm2 # 4140 <_sk_callback_sse41+0x75e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -11893,31 +11893,31 @@ _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 68,15,89,37,241,32,0,0 ; mulps 0x20f1(%rip),%xmm12 # 4150 <_sk_callback_sse41+0x777>
- DB 68,15,84,29,249,32,0,0 ; andps 0x20f9(%rip),%xmm11 # 4160 <_sk_callback_sse41+0x787>
- DB 68,15,86,29,1,33,0,0 ; orps 0x2101(%rip),%xmm11 # 4170 <_sk_callback_sse41+0x797>
- DB 68,15,88,37,9,33,0,0 ; addps 0x2109(%rip),%xmm12 # 4180 <_sk_callback_sse41+0x7a7>
- DB 15,40,29,18,33,0,0 ; movaps 0x2112(%rip),%xmm3 # 4190 <_sk_callback_sse41+0x7b7>
+ DB 68,15,89,37,232,32,0,0 ; mulps 0x20e8(%rip),%xmm12 # 4150 <_sk_callback_sse41+0x76e>
+ DB 68,15,84,29,240,32,0,0 ; andps 0x20f0(%rip),%xmm11 # 4160 <_sk_callback_sse41+0x77e>
+ DB 68,15,86,29,248,32,0,0 ; orps 0x20f8(%rip),%xmm11 # 4170 <_sk_callback_sse41+0x78e>
+ DB 68,15,88,37,0,33,0,0 ; addps 0x2100(%rip),%xmm12 # 4180 <_sk_callback_sse41+0x79e>
+ DB 15,40,29,9,33,0,0 ; movaps 0x2109(%rip),%xmm3 # 4190 <_sk_callback_sse41+0x7ae>
DB 65,15,89,219 ; mulps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
- DB 68,15,88,29,18,33,0,0 ; addps 0x2112(%rip),%xmm11 # 41a0 <_sk_callback_sse41+0x7c7>
- DB 15,40,29,27,33,0,0 ; movaps 0x211b(%rip),%xmm3 # 41b0 <_sk_callback_sse41+0x7d7>
+ DB 68,15,88,29,9,33,0,0 ; addps 0x2109(%rip),%xmm11 # 41a0 <_sk_callback_sse41+0x7be>
+ DB 15,40,29,18,33,0,0 ; movaps 0x2112(%rip),%xmm3 # 41b0 <_sk_callback_sse41+0x7ce>
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 68,15,88,37,8,33,0,0 ; addps 0x2108(%rip),%xmm12 # 41c0 <_sk_callback_sse41+0x7e7>
- DB 15,40,29,17,33,0,0 ; movaps 0x2111(%rip),%xmm3 # 41d0 <_sk_callback_sse41+0x7f7>
+ DB 68,15,88,37,255,32,0,0 ; addps 0x20ff(%rip),%xmm12 # 41c0 <_sk_callback_sse41+0x7de>
+ DB 15,40,29,8,33,0,0 ; movaps 0x2108(%rip),%xmm3 # 41d0 <_sk_callback_sse41+0x7ee>
DB 65,15,89,219 ; mulps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
- DB 68,15,40,21,17,33,0,0 ; movaps 0x2111(%rip),%xmm10 # 41e0 <_sk_callback_sse41+0x807>
+ DB 68,15,40,21,8,33,0,0 ; movaps 0x2108(%rip),%xmm10 # 41e0 <_sk_callback_sse41+0x7fe>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,29,22,33,0,0 ; movaps 0x2116(%rip),%xmm3 # 41f0 <_sk_callback_sse41+0x817>
+ DB 15,40,29,13,33,0,0 ; movaps 0x210d(%rip),%xmm3 # 41f0 <_sk_callback_sse41+0x80e>
DB 65,15,94,218 ; divps %xmm10,%xmm3
DB 65,15,88,220 ; addps %xmm12,%xmm3
- DB 15,89,29,23,33,0,0 ; mulps 0x2117(%rip),%xmm3 # 4200 <_sk_callback_sse41+0x827>
+ DB 15,89,29,14,33,0,0 ; mulps 0x210e(%rip),%xmm3 # 4200 <_sk_callback_sse41+0x81e>
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
@@ -11925,7 +11925,7 @@ _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 15,93,29,2,33,0,0 ; minps 0x2102(%rip),%xmm3 # 4210 <_sk_callback_sse41+0x837>
+ DB 15,93,29,249,32,0,0 ; minps 0x20f9(%rip),%xmm3 # 4210 <_sk_callback_sse41+0x82e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,15,40,192 ; movaps %xmm8,%xmm0
DB 255,224 ; jmpq *%rax
@@ -11933,29 +11933,29 @@ _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 68,15,89,5,254,32,0,0 ; mulps 0x20fe(%rip),%xmm8 # 4220 <_sk_callback_sse41+0x847>
- DB 68,15,40,13,6,33,0,0 ; movaps 0x2106(%rip),%xmm9 # 4230 <_sk_callback_sse41+0x857>
+ DB 68,15,89,5,245,32,0,0 ; mulps 0x20f5(%rip),%xmm8 # 4220 <_sk_callback_sse41+0x83e>
+ DB 68,15,40,13,253,32,0,0 ; movaps 0x20fd(%rip),%xmm9 # 4230 <_sk_callback_sse41+0x84e>
DB 65,15,89,201 ; mulps %xmm9,%xmm1
- DB 15,40,5,11,33,0,0 ; movaps 0x210b(%rip),%xmm0 # 4240 <_sk_callback_sse41+0x867>
+ DB 15,40,5,2,33,0,0 ; movaps 0x2102(%rip),%xmm0 # 4240 <_sk_callback_sse41+0x85e>
DB 15,88,200 ; addps %xmm0,%xmm1
DB 65,15,89,209 ; mulps %xmm9,%xmm2
DB 15,88,208 ; addps %xmm0,%xmm2
- DB 68,15,88,5,9,33,0,0 ; addps 0x2109(%rip),%xmm8 # 4250 <_sk_callback_sse41+0x877>
- DB 68,15,89,5,17,33,0,0 ; mulps 0x2111(%rip),%xmm8 # 4260 <_sk_callback_sse41+0x887>
- DB 15,89,13,26,33,0,0 ; mulps 0x211a(%rip),%xmm1 # 4270 <_sk_callback_sse41+0x897>
+ DB 68,15,88,5,0,33,0,0 ; addps 0x2100(%rip),%xmm8 # 4250 <_sk_callback_sse41+0x86e>
+ DB 68,15,89,5,8,33,0,0 ; mulps 0x2108(%rip),%xmm8 # 4260 <_sk_callback_sse41+0x87e>
+ DB 15,89,13,17,33,0,0 ; mulps 0x2111(%rip),%xmm1 # 4270 <_sk_callback_sse41+0x88e>
DB 65,15,88,200 ; addps %xmm8,%xmm1
- DB 15,89,21,31,33,0,0 ; mulps 0x211f(%rip),%xmm2 # 4280 <_sk_callback_sse41+0x8a7>
+ DB 15,89,21,22,33,0,0 ; mulps 0x2116(%rip),%xmm2 # 4280 <_sk_callback_sse41+0x89e>
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 68,15,40,13,19,33,0,0 ; movaps 0x2113(%rip),%xmm9 # 4290 <_sk_callback_sse41+0x8b7>
+ DB 68,15,40,13,10,33,0,0 ; movaps 0x210a(%rip),%xmm9 # 4290 <_sk_callback_sse41+0x8ae>
DB 65,15,40,193 ; movaps %xmm9,%xmm0
DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0
- DB 15,40,21,19,33,0,0 ; movaps 0x2113(%rip),%xmm2 # 42a0 <_sk_callback_sse41+0x8c7>
+ DB 15,40,21,10,33,0,0 ; movaps 0x210a(%rip),%xmm2 # 42a0 <_sk_callback_sse41+0x8be>
DB 15,88,202 ; addps %xmm2,%xmm1
- DB 68,15,40,37,24,33,0,0 ; movaps 0x2118(%rip),%xmm12 # 42b0 <_sk_callback_sse41+0x8d7>
+ DB 68,15,40,37,15,33,0,0 ; movaps 0x210f(%rip),%xmm12 # 42b0 <_sk_callback_sse41+0x8ce>
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
@@ -11974,8 +11974,8 @@ _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 15,89,13,209,32,0,0 ; mulps 0x20d1(%rip),%xmm1 # 42c0 <_sk_callback_sse41+0x8e7>
- DB 15,89,21,218,32,0,0 ; mulps 0x20da(%rip),%xmm2 # 42d0 <_sk_callback_sse41+0x8f7>
+ DB 15,89,13,200,32,0,0 ; mulps 0x20c8(%rip),%xmm1 # 42c0 <_sk_callback_sse41+0x8de>
+ DB 15,89,21,209,32,0,0 ; mulps 0x20d1(%rip),%xmm2 # 42d0 <_sk_callback_sse41+0x8ee>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,40,193 ; movaps %xmm1,%xmm0
DB 65,15,40,200 ; movaps %xmm8,%xmm1
@@ -11987,7 +11987,7 @@ _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 15,89,29,202,32,0,0 ; mulps 0x20ca(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x907>
+ DB 15,89,29,193,32,0,0 ; mulps 0x20c1(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x8fe>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
DB 15,87,201 ; xorps %xmm1,%xmm1
@@ -12018,7 +12018,7 @@ _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 15,89,29,94,32,0,0 ; mulps 0x205e(%rip),%xmm3 # 42f0 <_sk_callback_sse41+0x917>
+ DB 15,89,29,85,32,0,0 ; mulps 0x2055(%rip),%xmm3 # 42f0 <_sk_callback_sse41+0x90e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
DB 102,15,239,201 ; pxor %xmm1,%xmm1
@@ -12029,7 +12029,7 @@ 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 68,15,40,5,82,32,0,0 ; movaps 0x2052(%rip),%xmm8 # 4300 <_sk_callback_sse41+0x927>
+ DB 68,15,40,5,73,32,0,0 ; movaps 0x2049(%rip),%xmm8 # 4300 <_sk_callback_sse41+0x91e>
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
@@ -12044,9 +12044,9 @@ _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 15,89,5,47,32,0,0 ; mulps 0x202f(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x937>
+ DB 15,89,5,38,32,0,0 ; mulps 0x2026(%rip),%xmm0 # 4310 <_sk_callback_sse41+0x92e>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,54,32,0,0 ; movaps 0x2036(%rip),%xmm3 # 4320 <_sk_callback_sse41+0x947>
+ DB 15,40,29,45,32,0,0 ; movaps 0x202d(%rip),%xmm3 # 4320 <_sk_callback_sse41+0x93e>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 15,40,208 ; movaps %xmm0,%xmm2
DB 255,224 ; jmpq *%rax
@@ -12075,9 +12075,9 @@ _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 15,89,5,207,31,0,0 ; mulps 0x1fcf(%rip),%xmm0 # 4330 <_sk_callback_sse41+0x957>
+ DB 15,89,5,198,31,0,0 ; mulps 0x1fc6(%rip),%xmm0 # 4330 <_sk_callback_sse41+0x94e>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,214,31,0,0 ; movaps 0x1fd6(%rip),%xmm3 # 4340 <_sk_callback_sse41+0x967>
+ DB 15,40,29,205,31,0,0 ; movaps 0x1fcd(%rip),%xmm3 # 4340 <_sk_callback_sse41+0x95e>
DB 15,40,200 ; movaps %xmm0,%xmm1
DB 15,40,208 ; movaps %xmm0,%xmm2
DB 255,224 ; jmpq *%rax
@@ -12087,9 +12087,9 @@ _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 2381 <_sk_gather_i8_sse41+0xf>
+ DB 116,5 ; je 238a <_sk_gather_i8_sse41+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 2383 <_sk_gather_i8_sse41+0x11>
+ DB 235,2 ; jmp 238c <_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
@@ -12120,17 +12120,17 @@ _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 102,15,111,5,45,31,0,0 ; movdqa 0x1f2d(%rip),%xmm0 # 4350 <_sk_callback_sse41+0x977>
+ DB 102,15,111,5,36,31,0,0 ; movdqa 0x1f24(%rip),%xmm0 # 4350 <_sk_callback_sse41+0x96e>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,46,31,0,0 ; movaps 0x1f2e(%rip),%xmm8 # 4360 <_sk_callback_sse41+0x987>
+ DB 68,15,40,5,37,31,0,0 ; movaps 0x1f25(%rip),%xmm8 # 4360 <_sk_callback_sse41+0x97e>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 102,15,111,203 ; movdqa %xmm3,%xmm1
- DB 102,15,56,0,13,45,31,0,0 ; pshufb 0x1f2d(%rip),%xmm1 # 4370 <_sk_callback_sse41+0x997>
+ DB 102,15,56,0,13,36,31,0,0 ; pshufb 0x1f24(%rip),%xmm1 # 4370 <_sk_callback_sse41+0x98e>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,111,211 ; movdqa %xmm3,%xmm2
- DB 102,15,56,0,21,41,31,0,0 ; pshufb 0x1f29(%rip),%xmm2 # 4380 <_sk_callback_sse41+0x9a7>
+ DB 102,15,56,0,21,32,31,0,0 ; pshufb 0x1f20(%rip),%xmm2 # 4380 <_sk_callback_sse41+0x99e>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12144,19 +12144,19 @@ _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 102,15,111,5,15,31,0,0 ; movdqa 0x1f0f(%rip),%xmm0 # 4390 <_sk_callback_sse41+0x9b7>
+ DB 102,15,111,5,6,31,0,0 ; movdqa 0x1f06(%rip),%xmm0 # 4390 <_sk_callback_sse41+0x9ae>
DB 102,15,219,194 ; pand %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,17,31,0,0 ; mulps 0x1f11(%rip),%xmm0 # 43a0 <_sk_callback_sse41+0x9c7>
- DB 102,15,111,13,25,31,0,0 ; movdqa 0x1f19(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0x9d7>
+ DB 15,89,5,8,31,0,0 ; mulps 0x1f08(%rip),%xmm0 # 43a0 <_sk_callback_sse41+0x9be>
+ DB 102,15,111,13,16,31,0,0 ; movdqa 0x1f10(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0x9ce>
DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,27,31,0,0 ; mulps 0x1f1b(%rip),%xmm1 # 43c0 <_sk_callback_sse41+0x9e7>
- DB 102,15,219,21,35,31,0,0 ; pand 0x1f23(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0x9f7>
+ DB 15,89,13,18,31,0,0 ; mulps 0x1f12(%rip),%xmm1 # 43c0 <_sk_callback_sse41+0x9de>
+ DB 102,15,219,21,26,31,0,0 ; pand 0x1f1a(%rip),%xmm2 # 43d0 <_sk_callback_sse41+0x9ee>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,41,31,0,0 ; mulps 0x1f29(%rip),%xmm2 # 43e0 <_sk_callback_sse41+0xa07>
+ DB 15,89,21,32,31,0,0 ; mulps 0x1f20(%rip),%xmm2 # 43e0 <_sk_callback_sse41+0x9fe>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,48,31,0,0 ; movaps 0x1f30(%rip),%xmm3 # 43f0 <_sk_callback_sse41+0xa17>
+ DB 15,40,29,39,31,0,0 ; movaps 0x1f27(%rip),%xmm3 # 43f0 <_sk_callback_sse41+0xa0e>
DB 255,224 ; jmpq *%rax
PUBLIC _sk_gather_565_sse41
@@ -12182,31 +12182,31 @@ _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 102,15,111,5,213,30,0,0 ; movdqa 0x1ed5(%rip),%xmm0 # 4400 <_sk_callback_sse41+0xa27>
+ DB 102,15,111,5,204,30,0,0 ; movdqa 0x1ecc(%rip),%xmm0 # 4400 <_sk_callback_sse41+0xa1e>
DB 102,15,219,194 ; pand %xmm2,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,215,30,0,0 ; mulps 0x1ed7(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xa37>
- DB 102,15,111,13,223,30,0,0 ; movdqa 0x1edf(%rip),%xmm1 # 4420 <_sk_callback_sse41+0xa47>
+ DB 15,89,5,206,30,0,0 ; mulps 0x1ece(%rip),%xmm0 # 4410 <_sk_callback_sse41+0xa2e>
+ DB 102,15,111,13,214,30,0,0 ; movdqa 0x1ed6(%rip),%xmm1 # 4420 <_sk_callback_sse41+0xa3e>
DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,225,30,0,0 ; mulps 0x1ee1(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xa57>
- DB 102,15,219,21,233,30,0,0 ; pand 0x1ee9(%rip),%xmm2 # 4440 <_sk_callback_sse41+0xa67>
+ DB 15,89,13,216,30,0,0 ; mulps 0x1ed8(%rip),%xmm1 # 4430 <_sk_callback_sse41+0xa4e>
+ DB 102,15,219,21,224,30,0,0 ; pand 0x1ee0(%rip),%xmm2 # 4440 <_sk_callback_sse41+0xa5e>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,239,30,0,0 ; mulps 0x1eef(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xa77>
+ DB 15,89,21,230,30,0,0 ; mulps 0x1ee6(%rip),%xmm2 # 4450 <_sk_callback_sse41+0xa6e>
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,29,246,30,0,0 ; movaps 0x1ef6(%rip),%xmm3 # 4460 <_sk_callback_sse41+0xa87>
+ DB 15,40,29,237,30,0,0 ; movaps 0x1eed(%rip),%xmm3 # 4460 <_sk_callback_sse41+0xa7e>
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 68,15,40,5,247,30,0,0 ; movaps 0x1ef7(%rip),%xmm8 # 4470 <_sk_callback_sse41+0xa97>
+ DB 68,15,40,5,238,30,0,0 ; movaps 0x1eee(%rip),%xmm8 # 4470 <_sk_callback_sse41+0xa8e>
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 68,15,40,21,236,30,0,0 ; movaps 0x1eec(%rip),%xmm10 # 4480 <_sk_callback_sse41+0xaa7>
+ DB 68,15,40,21,227,30,0,0 ; movaps 0x1ee3(%rip),%xmm10 # 4480 <_sk_callback_sse41+0xa9e>
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
@@ -12224,21 +12224,21 @@ _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 102,15,111,5,183,30,0,0 ; movdqa 0x1eb7(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xab7>
+ DB 102,15,111,5,174,30,0,0 ; movdqa 0x1eae(%rip),%xmm0 # 4490 <_sk_callback_sse41+0xaae>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,185,30,0,0 ; mulps 0x1eb9(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xac7>
- DB 102,15,111,13,193,30,0,0 ; movdqa 0x1ec1(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xad7>
+ DB 15,89,5,176,30,0,0 ; mulps 0x1eb0(%rip),%xmm0 # 44a0 <_sk_callback_sse41+0xabe>
+ DB 102,15,111,13,184,30,0,0 ; movdqa 0x1eb8(%rip),%xmm1 # 44b0 <_sk_callback_sse41+0xace>
DB 102,15,219,203 ; pand %xmm3,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,195,30,0,0 ; mulps 0x1ec3(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xae7>
- DB 102,15,111,21,203,30,0,0 ; movdqa 0x1ecb(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xaf7>
+ DB 15,89,13,186,30,0,0 ; mulps 0x1eba(%rip),%xmm1 # 44c0 <_sk_callback_sse41+0xade>
+ DB 102,15,111,21,194,30,0,0 ; movdqa 0x1ec2(%rip),%xmm2 # 44d0 <_sk_callback_sse41+0xaee>
DB 102,15,219,211 ; pand %xmm3,%xmm2
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,205,30,0,0 ; mulps 0x1ecd(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xb07>
- DB 102,15,219,29,213,30,0,0 ; pand 0x1ed5(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xb17>
+ DB 15,89,21,196,30,0,0 ; mulps 0x1ec4(%rip),%xmm2 # 44e0 <_sk_callback_sse41+0xafe>
+ DB 102,15,219,29,204,30,0,0 ; pand 0x1ecc(%rip),%xmm3 # 44f0 <_sk_callback_sse41+0xb0e>
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,219,30,0,0 ; mulps 0x1edb(%rip),%xmm3 # 4500 <_sk_callback_sse41+0xb27>
+ DB 15,89,29,210,30,0,0 ; mulps 0x1ed2(%rip),%xmm3 # 4500 <_sk_callback_sse41+0xb1e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -12265,21 +12265,21 @@ _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 102,15,111,5,126,30,0,0 ; movdqa 0x1e7e(%rip),%xmm0 # 4510 <_sk_callback_sse41+0xb37>
+ DB 102,15,111,5,117,30,0,0 ; movdqa 0x1e75(%rip),%xmm0 # 4510 <_sk_callback_sse41+0xb2e>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 15,89,5,128,30,0,0 ; mulps 0x1e80(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xb47>
- DB 102,15,111,13,136,30,0,0 ; movdqa 0x1e88(%rip),%xmm1 # 4530 <_sk_callback_sse41+0xb57>
+ DB 15,89,5,119,30,0,0 ; mulps 0x1e77(%rip),%xmm0 # 4520 <_sk_callback_sse41+0xb3e>
+ DB 102,15,111,13,127,30,0,0 ; movdqa 0x1e7f(%rip),%xmm1 # 4530 <_sk_callback_sse41+0xb4e>
DB 102,15,219,203 ; pand %xmm3,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
- DB 15,89,13,138,30,0,0 ; mulps 0x1e8a(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xb67>
- DB 102,15,111,21,146,30,0,0 ; movdqa 0x1e92(%rip),%xmm2 # 4550 <_sk_callback_sse41+0xb77>
+ DB 15,89,13,129,30,0,0 ; mulps 0x1e81(%rip),%xmm1 # 4540 <_sk_callback_sse41+0xb5e>
+ DB 102,15,111,21,137,30,0,0 ; movdqa 0x1e89(%rip),%xmm2 # 4550 <_sk_callback_sse41+0xb6e>
DB 102,15,219,211 ; pand %xmm3,%xmm2
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
- DB 15,89,21,148,30,0,0 ; mulps 0x1e94(%rip),%xmm2 # 4560 <_sk_callback_sse41+0xb87>
- DB 102,15,219,29,156,30,0,0 ; pand 0x1e9c(%rip),%xmm3 # 4570 <_sk_callback_sse41+0xb97>
+ DB 15,89,21,139,30,0,0 ; mulps 0x1e8b(%rip),%xmm2 # 4560 <_sk_callback_sse41+0xb7e>
+ DB 102,15,219,29,147,30,0,0 ; pand 0x1e93(%rip),%xmm3 # 4570 <_sk_callback_sse41+0xb8e>
DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
- DB 15,89,29,162,30,0,0 ; mulps 0x1ea2(%rip),%xmm3 # 4580 <_sk_callback_sse41+0xba7>
+ DB 15,89,29,153,30,0,0 ; mulps 0x1e99(%rip),%xmm3 # 4580 <_sk_callback_sse41+0xb9e>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -12287,7 +12287,7 @@ 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 68,15,40,5,161,30,0,0 ; movaps 0x1ea1(%rip),%xmm8 # 4590 <_sk_callback_sse41+0xbb7>
+ DB 68,15,40,5,152,30,0,0 ; movaps 0x1e98(%rip),%xmm8 # 4590 <_sk_callback_sse41+0xbae>
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
@@ -12315,17 +12315,17 @@ _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 15,40,5,64,30,0,0 ; movaps 0x1e40(%rip),%xmm0 # 45a0 <_sk_callback_sse41+0xbc7>
+ DB 15,40,5,55,30,0,0 ; movaps 0x1e37(%rip),%xmm0 # 45a0 <_sk_callback_sse41+0xbbe>
DB 15,84,195 ; andps %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,66,30,0,0 ; movaps 0x1e42(%rip),%xmm8 # 45b0 <_sk_callback_sse41+0xbd7>
+ DB 68,15,40,5,57,30,0,0 ; movaps 0x1e39(%rip),%xmm8 # 45b0 <_sk_callback_sse41+0xbce>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 15,40,203 ; movaps %xmm3,%xmm1
- DB 102,15,56,0,13,66,30,0,0 ; pshufb 0x1e42(%rip),%xmm1 # 45c0 <_sk_callback_sse41+0xbe7>
+ DB 102,15,56,0,13,57,30,0,0 ; pshufb 0x1e39(%rip),%xmm1 # 45c0 <_sk_callback_sse41+0xbde>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 15,40,211 ; movaps %xmm3,%xmm2
- DB 102,15,56,0,21,63,30,0,0 ; pshufb 0x1e3f(%rip),%xmm2 # 45d0 <_sk_callback_sse41+0xbf7>
+ DB 102,15,56,0,21,54,30,0,0 ; pshufb 0x1e36(%rip),%xmm2 # 45d0 <_sk_callback_sse41+0xbee>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12354,17 +12354,17 @@ _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 102,15,111,5,216,29,0,0 ; movdqa 0x1dd8(%rip),%xmm0 # 45e0 <_sk_callback_sse41+0xc07>
+ DB 102,15,111,5,207,29,0,0 ; movdqa 0x1dcf(%rip),%xmm0 # 45e0 <_sk_callback_sse41+0xbfe>
DB 102,15,219,195 ; pand %xmm3,%xmm0
DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
- DB 68,15,40,5,217,29,0,0 ; movaps 0x1dd9(%rip),%xmm8 # 45f0 <_sk_callback_sse41+0xc17>
+ DB 68,15,40,5,208,29,0,0 ; movaps 0x1dd0(%rip),%xmm8 # 45f0 <_sk_callback_sse41+0xc0e>
DB 65,15,89,192 ; mulps %xmm8,%xmm0
DB 102,15,111,203 ; movdqa %xmm3,%xmm1
- DB 102,15,56,0,13,216,29,0,0 ; pshufb 0x1dd8(%rip),%xmm1 # 4600 <_sk_callback_sse41+0xc27>
+ DB 102,15,56,0,13,207,29,0,0 ; pshufb 0x1dcf(%rip),%xmm1 # 4600 <_sk_callback_sse41+0xc1e>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
DB 102,15,111,211 ; movdqa %xmm3,%xmm2
- DB 102,15,56,0,21,212,29,0,0 ; pshufb 0x1dd4(%rip),%xmm2 # 4610 <_sk_callback_sse41+0xc37>
+ DB 102,15,56,0,21,203,29,0,0 ; pshufb 0x1dcb(%rip),%xmm2 # 4610 <_sk_callback_sse41+0xc2e>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12377,7 +12377,7 @@ 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 68,15,40,5,192,29,0,0 ; movaps 0x1dc0(%rip),%xmm8 # 4620 <_sk_callback_sse41+0xc47>
+ DB 68,15,40,5,183,29,0,0 ; movaps 0x1db7(%rip),%xmm8 # 4620 <_sk_callback_sse41+0xc3e>
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
@@ -12412,18 +12412,18 @@ _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 102,68,15,111,5,57,29,0,0 ; movdqa 0x1d39(%rip),%xmm8 # 4630 <_sk_callback_sse41+0xc57>
+ DB 102,68,15,111,5,48,29,0,0 ; movdqa 0x1d30(%rip),%xmm8 # 4630 <_sk_callback_sse41+0xc4e>
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 102,15,111,29,52,29,0,0 ; movdqa 0x1d34(%rip),%xmm3 # 4640 <_sk_callback_sse41+0xc67>
+ DB 102,15,111,29,43,29,0,0 ; movdqa 0x1d2b(%rip),%xmm3 # 4640 <_sk_callback_sse41+0xc5e>
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 102,68,15,111,21,32,29,0,0 ; movdqa 0x1d20(%rip),%xmm10 # 4650 <_sk_callback_sse41+0xc77>
+ DB 102,68,15,111,21,23,29,0,0 ; movdqa 0x1d17(%rip),%xmm10 # 4650 <_sk_callback_sse41+0xc6e>
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
@@ -12494,18 +12494,18 @@ _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 102,68,15,111,5,222,27,0,0 ; movdqa 0x1bde(%rip),%xmm8 # 4660 <_sk_callback_sse41+0xc87>
+ DB 102,68,15,111,5,213,27,0,0 ; movdqa 0x1bd5(%rip),%xmm8 # 4660 <_sk_callback_sse41+0xc7e>
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 102,15,111,29,217,27,0,0 ; movdqa 0x1bd9(%rip),%xmm3 # 4670 <_sk_callback_sse41+0xc97>
+ DB 102,15,111,29,208,27,0,0 ; movdqa 0x1bd0(%rip),%xmm3 # 4670 <_sk_callback_sse41+0xc8e>
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 102,68,15,111,21,197,27,0,0 ; movdqa 0x1bc5(%rip),%xmm10 # 4680 <_sk_callback_sse41+0xca7>
+ DB 102,68,15,111,21,188,27,0,0 ; movdqa 0x1bbc(%rip),%xmm10 # 4680 <_sk_callback_sse41+0xc9e>
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
@@ -12551,17 +12551,17 @@ 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 102,68,15,111,21,251,26,0,0 ; movdqa 0x1afb(%rip),%xmm10 # 4690 <_sk_callback_sse41+0xcb7>
+ DB 102,68,15,111,21,242,26,0,0 ; movdqa 0x1af2(%rip),%xmm10 # 4690 <_sk_callback_sse41+0xcae>
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 102,68,15,111,13,238,26,0,0 ; movdqa 0x1aee(%rip),%xmm9 # 46a0 <_sk_callback_sse41+0xcc7>
+ DB 102,68,15,111,13,229,26,0,0 ; movdqa 0x1ae5(%rip),%xmm9 # 46a0 <_sk_callback_sse41+0xcbe>
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 102,68,15,111,29,223,26,0,0 ; movdqa 0x1adf(%rip),%xmm11 # 46b0 <_sk_callback_sse41+0xcd7>
+ DB 102,68,15,111,29,214,26,0,0 ; movdqa 0x1ad6(%rip),%xmm11 # 46b0 <_sk_callback_sse41+0xcce>
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
@@ -12629,7 +12629,7 @@ _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 68,15,40,5,174,25,0,0 ; movaps 0x19ae(%rip),%xmm8 # 46c0 <_sk_callback_sse41+0xce7>
+ DB 68,15,40,5,165,25,0,0 ; movaps 0x19a5(%rip),%xmm8 # 46c0 <_sk_callback_sse41+0xcde>
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
@@ -12679,7 +12679,7 @@ _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 68,15,40,5,239,24,0,0 ; movaps 0x18ef(%rip),%xmm8 # 46d0 <_sk_callback_sse41+0xcf7>
+ DB 68,15,40,5,230,24,0,0 ; movaps 0x18e6(%rip),%xmm8 # 46d0 <_sk_callback_sse41+0xcee>
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
@@ -12696,14 +12696,14 @@ _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 15,40,29,182,24,0,0 ; movaps 0x18b6(%rip),%xmm3 # 46e0 <_sk_callback_sse41+0xd07>
+ DB 15,40,29,173,24,0,0 ; movaps 0x18ad(%rip),%xmm3 # 46e0 <_sk_callback_sse41+0xcfe>
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 68,15,40,13,183,24,0,0 ; movaps 0x18b7(%rip),%xmm9 # 46f0 <_sk_callback_sse41+0xd17>
+ DB 68,15,40,13,174,24,0,0 ; movaps 0x18ae(%rip),%xmm9 # 46f0 <_sk_callback_sse41+0xd0e>
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
@@ -12908,10 +12908,10 @@ _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 15,89,5,213,21,0,0 ; mulps 0x15d5(%rip),%xmm0 # 4700 <_sk_callback_sse41+0xd27>
- DB 15,89,13,222,21,0,0 ; mulps 0x15de(%rip),%xmm1 # 4710 <_sk_callback_sse41+0xd37>
+ DB 15,89,5,204,21,0,0 ; mulps 0x15cc(%rip),%xmm0 # 4700 <_sk_callback_sse41+0xd1e>
+ DB 15,89,13,213,21,0,0 ; mulps 0x15d5(%rip),%xmm1 # 4710 <_sk_callback_sse41+0xd2e>
DB 15,88,200 ; addps %xmm0,%xmm1
- DB 15,89,29,228,21,0,0 ; mulps 0x15e4(%rip),%xmm3 # 4720 <_sk_callback_sse41+0xd47>
+ DB 15,89,29,219,21,0,0 ; mulps 0x15db(%rip),%xmm3 # 4720 <_sk_callback_sse41+0xd3e>
DB 15,88,217 ; addps %xmm1,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 15,87,192 ; xorps %xmm0,%xmm0
@@ -13134,7 +13134,7 @@ _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 35e4 <_sk_linear_gradient_sse41+0x13e>
+ DB 15,132,4,1,0,0 ; je 35ed <_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)
@@ -13185,13 +13185,13 @@ _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 350c <_sk_linear_gradient_sse41+0x66>
+ DB 15,133,65,255,255,255 ; jne 3515 <_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 35f1 <_sk_linear_gradient_sse41+0x14b>
+ DB 235,13 ; jmp 35fa <_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
@@ -13242,7 +13242,7 @@ _sk_linear_gradient_2stops_sse41 LABEL PROC
PUBLIC _sk_save_xy_sse41
_sk_save_xy_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 68,15,40,5,160,16,0,0 ; movaps 0x10a0(%rip),%xmm8 # 4730 <_sk_callback_sse41+0xd57>
+ DB 68,15,40,5,151,16,0,0 ; movaps 0x1097(%rip),%xmm8 # 4730 <_sk_callback_sse41+0xd4e>
DB 15,17,0 ; movups %xmm0,(%rax)
DB 68,15,40,200 ; movaps %xmm0,%xmm9
DB 69,15,88,200 ; addps %xmm8,%xmm9
@@ -13282,8 +13282,8 @@ _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 15,88,5,34,16,0,0 ; addps 0x1022(%rip),%xmm0 # 4740 <_sk_callback_sse41+0xd67>
- DB 68,15,40,13,42,16,0,0 ; movaps 0x102a(%rip),%xmm9 # 4750 <_sk_callback_sse41+0xd77>
+ DB 15,88,5,25,16,0,0 ; addps 0x1019(%rip),%xmm0 # 4740 <_sk_callback_sse41+0xd5e>
+ DB 68,15,40,13,33,16,0,0 ; movaps 0x1021(%rip),%xmm9 # 4750 <_sk_callback_sse41+0xd6e>
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
@@ -13294,7 +13294,7 @@ _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 15,88,5,25,16,0,0 ; addps 0x1019(%rip),%xmm0 # 4760 <_sk_callback_sse41+0xd87>
+ DB 15,88,5,16,16,0,0 ; addps 0x1010(%rip),%xmm0 # 4760 <_sk_callback_sse41+0xd7e>
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
@@ -13304,8 +13304,8 @@ _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 15,88,13,11,16,0,0 ; addps 0x100b(%rip),%xmm1 # 4770 <_sk_callback_sse41+0xd97>
- DB 68,15,40,13,19,16,0,0 ; movaps 0x1013(%rip),%xmm9 # 4780 <_sk_callback_sse41+0xda7>
+ DB 15,88,13,2,16,0,0 ; addps 0x1002(%rip),%xmm1 # 4770 <_sk_callback_sse41+0xd8e>
+ DB 68,15,40,13,10,16,0,0 ; movaps 0x100a(%rip),%xmm9 # 4780 <_sk_callback_sse41+0xd9e>
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
@@ -13316,7 +13316,7 @@ _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 15,88,13,1,16,0,0 ; addps 0x1001(%rip),%xmm1 # 4790 <_sk_callback_sse41+0xdb7>
+ DB 15,88,13,248,15,0,0 ; addps 0xff8(%rip),%xmm1 # 4790 <_sk_callback_sse41+0xdae>
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
@@ -13326,13 +13326,13 @@ _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 15,88,5,244,15,0,0 ; addps 0xff4(%rip),%xmm0 # 47a0 <_sk_callback_sse41+0xdc7>
- DB 68,15,40,13,252,15,0,0 ; movaps 0xffc(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0xdd7>
+ DB 15,88,5,235,15,0,0 ; addps 0xfeb(%rip),%xmm0 # 47a0 <_sk_callback_sse41+0xdbe>
+ DB 68,15,40,13,243,15,0,0 ; movaps 0xff3(%rip),%xmm9 # 47b0 <_sk_callback_sse41+0xdce>
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 68,15,89,13,248,15,0,0 ; mulps 0xff8(%rip),%xmm9 # 47c0 <_sk_callback_sse41+0xde7>
- DB 68,15,88,13,0,16,0,0 ; addps 0x1000(%rip),%xmm9 # 47d0 <_sk_callback_sse41+0xdf7>
+ DB 68,15,89,13,239,15,0,0 ; mulps 0xfef(%rip),%xmm9 # 47c0 <_sk_callback_sse41+0xdde>
+ DB 68,15,88,13,247,15,0,0 ; addps 0xff7(%rip),%xmm9 # 47d0 <_sk_callback_sse41+0xdee>
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
@@ -13343,16 +13343,16 @@ _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 15,88,5,239,15,0,0 ; addps 0xfef(%rip),%xmm0 # 47e0 <_sk_callback_sse41+0xe07>
- DB 68,15,40,13,247,15,0,0 ; movaps 0xff7(%rip),%xmm9 # 47f0 <_sk_callback_sse41+0xe17>
+ DB 15,88,5,230,15,0,0 ; addps 0xfe6(%rip),%xmm0 # 47e0 <_sk_callback_sse41+0xdfe>
+ DB 68,15,40,13,238,15,0,0 ; movaps 0xfee(%rip),%xmm9 # 47f0 <_sk_callback_sse41+0xe0e>
DB 69,15,92,200 ; subps %xmm8,%xmm9
- DB 68,15,40,5,251,15,0,0 ; movaps 0xffb(%rip),%xmm8 # 4800 <_sk_callback_sse41+0xe27>
+ DB 68,15,40,5,242,15,0,0 ; movaps 0xff2(%rip),%xmm8 # 4800 <_sk_callback_sse41+0xe1e>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,255,15,0,0 ; addps 0xfff(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xe37>
+ DB 68,15,88,5,246,15,0,0 ; addps 0xff6(%rip),%xmm8 # 4810 <_sk_callback_sse41+0xe2e>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,3,16,0,0 ; addps 0x1003(%rip),%xmm8 # 4820 <_sk_callback_sse41+0xe47>
+ DB 68,15,88,5,250,15,0,0 ; addps 0xffa(%rip),%xmm8 # 4820 <_sk_callback_sse41+0xe3e>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,7,16,0,0 ; addps 0x1007(%rip),%xmm8 # 4830 <_sk_callback_sse41+0xe57>
+ DB 68,15,88,5,254,15,0,0 ; addps 0xffe(%rip),%xmm8 # 4830 <_sk_callback_sse41+0xe4e>
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
@@ -13360,17 +13360,17 @@ _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 68,15,40,5,1,16,0,0 ; movaps 0x1001(%rip),%xmm8 # 4840 <_sk_callback_sse41+0xe67>
+ DB 68,15,40,5,248,15,0,0 ; movaps 0xff8(%rip),%xmm8 # 4840 <_sk_callback_sse41+0xe5e>
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 68,15,40,21,253,15,0,0 ; movaps 0xffd(%rip),%xmm10 # 4850 <_sk_callback_sse41+0xe77>
+ DB 68,15,40,21,244,15,0,0 ; movaps 0xff4(%rip),%xmm10 # 4850 <_sk_callback_sse41+0xe6e>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,1,16,0,0 ; addps 0x1001(%rip),%xmm10 # 4860 <_sk_callback_sse41+0xe87>
+ DB 68,15,88,21,248,15,0,0 ; addps 0xff8(%rip),%xmm10 # 4860 <_sk_callback_sse41+0xe7e>
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 68,15,88,21,253,15,0,0 ; addps 0xffd(%rip),%xmm10 # 4870 <_sk_callback_sse41+0xe97>
+ DB 68,15,88,21,244,15,0,0 ; addps 0xff4(%rip),%xmm10 # 4870 <_sk_callback_sse41+0xe8e>
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
@@ -13380,11 +13380,11 @@ _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 15,88,5,240,15,0,0 ; addps 0xff0(%rip),%xmm0 # 4880 <_sk_callback_sse41+0xea7>
+ DB 15,88,5,231,15,0,0 ; addps 0xfe7(%rip),%xmm0 # 4880 <_sk_callback_sse41+0xe9e>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,89,201 ; mulps %xmm9,%xmm9
- DB 68,15,89,5,240,15,0,0 ; mulps 0xff0(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xeb7>
- DB 68,15,88,5,248,15,0,0 ; addps 0xff8(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xec7>
+ DB 68,15,89,5,231,15,0,0 ; mulps 0xfe7(%rip),%xmm8 # 4890 <_sk_callback_sse41+0xeae>
+ DB 68,15,88,5,239,15,0,0 ; addps 0xfef(%rip),%xmm8 # 48a0 <_sk_callback_sse41+0xebe>
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
@@ -13395,13 +13395,13 @@ _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 15,88,13,230,15,0,0 ; addps 0xfe6(%rip),%xmm1 # 48b0 <_sk_callback_sse41+0xed7>
- DB 68,15,40,13,238,15,0,0 ; movaps 0xfee(%rip),%xmm9 # 48c0 <_sk_callback_sse41+0xee7>
+ DB 15,88,13,221,15,0,0 ; addps 0xfdd(%rip),%xmm1 # 48b0 <_sk_callback_sse41+0xece>
+ DB 68,15,40,13,229,15,0,0 ; movaps 0xfe5(%rip),%xmm9 # 48c0 <_sk_callback_sse41+0xede>
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 68,15,89,13,234,15,0,0 ; mulps 0xfea(%rip),%xmm9 # 48d0 <_sk_callback_sse41+0xef7>
- DB 68,15,88,13,242,15,0,0 ; addps 0xff2(%rip),%xmm9 # 48e0 <_sk_callback_sse41+0xf07>
+ DB 68,15,89,13,225,15,0,0 ; mulps 0xfe1(%rip),%xmm9 # 48d0 <_sk_callback_sse41+0xeee>
+ DB 68,15,88,13,233,15,0,0 ; addps 0xfe9(%rip),%xmm9 # 48e0 <_sk_callback_sse41+0xefe>
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
@@ -13412,16 +13412,16 @@ _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 15,88,13,224,15,0,0 ; addps 0xfe0(%rip),%xmm1 # 48f0 <_sk_callback_sse41+0xf17>
- DB 68,15,40,13,232,15,0,0 ; movaps 0xfe8(%rip),%xmm9 # 4900 <_sk_callback_sse41+0xf27>
+ DB 15,88,13,215,15,0,0 ; addps 0xfd7(%rip),%xmm1 # 48f0 <_sk_callback_sse41+0xf0e>
+ DB 68,15,40,13,223,15,0,0 ; movaps 0xfdf(%rip),%xmm9 # 4900 <_sk_callback_sse41+0xf1e>
DB 69,15,92,200 ; subps %xmm8,%xmm9
- DB 68,15,40,5,236,15,0,0 ; movaps 0xfec(%rip),%xmm8 # 4910 <_sk_callback_sse41+0xf37>
+ DB 68,15,40,5,227,15,0,0 ; movaps 0xfe3(%rip),%xmm8 # 4910 <_sk_callback_sse41+0xf2e>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,240,15,0,0 ; addps 0xff0(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xf47>
+ DB 68,15,88,5,231,15,0,0 ; addps 0xfe7(%rip),%xmm8 # 4920 <_sk_callback_sse41+0xf3e>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,244,15,0,0 ; addps 0xff4(%rip),%xmm8 # 4930 <_sk_callback_sse41+0xf57>
+ DB 68,15,88,5,235,15,0,0 ; addps 0xfeb(%rip),%xmm8 # 4930 <_sk_callback_sse41+0xf4e>
DB 69,15,89,193 ; mulps %xmm9,%xmm8
- DB 68,15,88,5,248,15,0,0 ; addps 0xff8(%rip),%xmm8 # 4940 <_sk_callback_sse41+0xf67>
+ DB 68,15,88,5,239,15,0,0 ; addps 0xfef(%rip),%xmm8 # 4940 <_sk_callback_sse41+0xf5e>
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
@@ -13429,17 +13429,17 @@ _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 68,15,40,5,242,15,0,0 ; movaps 0xff2(%rip),%xmm8 # 4950 <_sk_callback_sse41+0xf77>
+ DB 68,15,40,5,233,15,0,0 ; movaps 0xfe9(%rip),%xmm8 # 4950 <_sk_callback_sse41+0xf6e>
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 68,15,40,21,237,15,0,0 ; movaps 0xfed(%rip),%xmm10 # 4960 <_sk_callback_sse41+0xf87>
+ DB 68,15,40,21,228,15,0,0 ; movaps 0xfe4(%rip),%xmm10 # 4960 <_sk_callback_sse41+0xf7e>
DB 69,15,89,209 ; mulps %xmm9,%xmm10
- DB 68,15,88,21,241,15,0,0 ; addps 0xff1(%rip),%xmm10 # 4970 <_sk_callback_sse41+0xf97>
+ DB 68,15,88,21,232,15,0,0 ; addps 0xfe8(%rip),%xmm10 # 4970 <_sk_callback_sse41+0xf8e>
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 68,15,88,21,237,15,0,0 ; addps 0xfed(%rip),%xmm10 # 4980 <_sk_callback_sse41+0xfa7>
+ DB 68,15,88,21,228,15,0,0 ; addps 0xfe4(%rip),%xmm10 # 4980 <_sk_callback_sse41+0xf9e>
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
@@ -13449,11 +13449,11 @@ _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 15,88,13,223,15,0,0 ; addps 0xfdf(%rip),%xmm1 # 4990 <_sk_callback_sse41+0xfb7>
+ DB 15,88,13,214,15,0,0 ; addps 0xfd6(%rip),%xmm1 # 4990 <_sk_callback_sse41+0xfae>
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,89,201 ; mulps %xmm9,%xmm9
- DB 68,15,89,5,223,15,0,0 ; mulps 0xfdf(%rip),%xmm8 # 49a0 <_sk_callback_sse41+0xfc7>
- DB 68,15,88,5,231,15,0,0 ; addps 0xfe7(%rip),%xmm8 # 49b0 <_sk_callback_sse41+0xfd7>
+ DB 68,15,89,5,214,15,0,0 ; mulps 0xfd6(%rip),%xmm8 # 49a0 <_sk_callback_sse41+0xfbe>
+ DB 68,15,88,5,222,15,0,0 ; addps 0xfde(%rip),%xmm8 # 49b0 <_sk_callback_sse41+0xfce>
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
@@ -13877,10 +13877,10 @@ ALIGN 16
DB 0,1 ; add %al,(%rcx)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a003e28 <_sk_callback_sse41+0xa00044f>
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a003e28 <_sk_callback_sse41+0xa000446>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3003e30 <_sk_callback_sse41+0x3000457>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3003e30 <_sk_callback_sse41+0x300044e>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -13995,7 +13995,7 @@ 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 ffffffffe9a37f2a <_sk_callback_sse41+0xffffffffe9a34551>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a37f2a <_sk_callback_sse41+0xffffffffe9a34548>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14091,7 +14091,7 @@ 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 ffffffffe9a37ffa <_sk_callback_sse41+0xffffffffe9a34621>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a37ffa <_sk_callback_sse41+0xffffffffe9a34618>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14187,7 +14187,7 @@ 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 ffffffffe9a380ca <_sk_callback_sse41+0xffffffffe9a346f1>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a380ca <_sk_callback_sse41+0xffffffffe9a346e8>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14283,7 +14283,7 @@ 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 ffffffffe9a3819a <_sk_callback_sse41+0xffffffffe9a347c1>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a3819a <_sk_callback_sse41+0xffffffffe9a347b8>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14441,7 +14441,7 @@ 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) # 3004380 <_sk_callback_sse41+0x30009a7>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004380 <_sk_callback_sse41+0x300099e>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14683,7 +14683,7 @@ 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) # 30045d0 <_sk_callback_sse41+0x3000bf7>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30045d0 <_sk_callback_sse41+0x3000bee>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14710,7 +14710,7 @@ 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) # 3004610 <_sk_callback_sse41+0x3000c37>
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004610 <_sk_callback_sse41+0x3000c2e>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255,6 ; incl (%rsi)
@@ -14943,7 +14943,7 @@ 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 # 3d63d675 <_sk_callback_sse41+0x3d639c9c>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d675 <_sk_callback_sse41+0x3d639c93>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14969,7 +14969,7 @@ 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 # 3d63d6b5 <_sk_callback_sse41+0x3d639cdc>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d6b5 <_sk_callback_sse41+0x3d639cd3>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
@@ -15053,7 +15053,7 @@ 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 # 3d63d785 <_sk_callback_sse41+0x3d639dac>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d785 <_sk_callback_sse41+0x3d639da3>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -15079,7 +15079,7 @@ 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 # 3d63d7c5 <_sk_callback_sse41+0x3d639dec>
+ DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63d7c5 <_sk_callback_sse41+0x3d639de3>
DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi)
DB 192,63,0 ; sarb $0x0,(%rdi)
DB 0,192 ; add %al,%al
@@ -15090,11 +15090,11 @@ ALIGN 16
DB 63 ; (bad)
DB 114,28 ; jb 49be <.literal16+0xf2e>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 49c2 <_sk_callback_sse41+0xfe9>
+ DB 62,114,28 ; jb,pt 49c2 <_sk_callback_sse41+0xfe0>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 49c6 <_sk_callback_sse41+0xfed>
+ DB 62,114,28 ; jb,pt 49c6 <_sk_callback_sse41+0xfe4>
DB 199 ; (bad)
- DB 62,114,28 ; jb,pt 49ca <_sk_callback_sse41+0xff1>
+ DB 62,114,28 ; jb,pt 49ca <_sk_callback_sse41+0xfe8>
DB 199 ; (bad)
DB 62,171 ; ds stos %eax,%es:(%rdi)
DB 170 ; stos %al,%es:(%rdi)
@@ -16366,173 +16366,172 @@ _sk_hsl_to_rgb_sse2 LABEL PROC
DB 15,41,108,36,112 ; movaps %xmm5,0x70(%rsp)
DB 15,41,100,36,96 ; movaps %xmm4,0x60(%rsp)
DB 15,41,92,36,80 ; movaps %xmm3,0x50(%rsp)
- DB 68,15,40,210 ; movaps %xmm2,%xmm10
- DB 15,40,217 ; movaps %xmm1,%xmm3
- DB 15,40,232 ; movaps %xmm0,%xmm5
+ DB 68,15,40,226 ; movaps %xmm2,%xmm12
+ DB 15,40,240 ; movaps %xmm0,%xmm6
DB 184,0,0,0,63 ; mov $0x3f000000,%eax
DB 102,15,110,192 ; movd %eax,%xmm0
DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
- DB 69,15,40,194 ; movaps %xmm10,%xmm8
+ DB 69,15,40,196 ; movaps %xmm12,%xmm8
DB 68,15,194,192,1 ; cmpltps %xmm0,%xmm8
- DB 68,15,40,216 ; movaps %xmm0,%xmm11
- DB 68,15,40,37,230,47,0,0 ; movaps 0x2fe6(%rip),%xmm12 # 4120 <_sk_callback_sse2+0x349>
- DB 15,40,195 ; movaps %xmm3,%xmm0
- DB 15,40,211 ; movaps %xmm3,%xmm2
- DB 15,87,201 ; xorps %xmm1,%xmm1
- DB 15,194,203,0 ; cmpeqps %xmm3,%xmm1
- DB 15,41,76,36,48 ; movaps %xmm1,0x30(%rsp)
- DB 65,15,88,220 ; addps %xmm12,%xmm3
- DB 65,15,89,218 ; mulps %xmm10,%xmm3
- DB 65,15,88,194 ; addps %xmm10,%xmm0
- DB 65,15,89,210 ; mulps %xmm10,%xmm2
- DB 15,92,194 ; subps %xmm2,%xmm0
- DB 65,15,84,216 ; andps %xmm8,%xmm3
+ DB 68,15,40,208 ; movaps %xmm0,%xmm10
+ DB 68,15,41,84,36,48 ; movaps %xmm10,0x30(%rsp)
+ DB 15,40,61,228,47,0,0 ; movaps 0x2fe4(%rip),%xmm7 # 4120 <_sk_callback_sse2+0x349>
+ DB 15,40,193 ; movaps %xmm1,%xmm0
+ DB 15,40,225 ; movaps %xmm1,%xmm4
+ DB 15,87,210 ; xorps %xmm2,%xmm2
+ DB 15,194,209,0 ; cmpeqps %xmm1,%xmm2
+ DB 15,41,84,36,32 ; movaps %xmm2,0x20(%rsp)
+ DB 15,88,207 ; addps %xmm7,%xmm1
+ DB 65,15,89,204 ; mulps %xmm12,%xmm1
+ DB 65,15,88,196 ; addps %xmm12,%xmm0
+ DB 65,15,89,228 ; mulps %xmm12,%xmm4
+ DB 15,92,196 ; subps %xmm4,%xmm0
+ DB 65,15,84,200 ; andps %xmm8,%xmm1
DB 68,15,85,192 ; andnps %xmm0,%xmm8
- DB 68,15,86,195 ; orps %xmm3,%xmm8
- DB 15,40,29,190,47,0,0 ; movaps 0x2fbe(%rip),%xmm3 # 4130 <_sk_callback_sse2+0x359>
- DB 15,88,221 ; addps %xmm5,%xmm3
+ DB 68,15,86,193 ; orps %xmm1,%xmm8
+ DB 15,40,13,189,47,0,0 ; movaps 0x2fbd(%rip),%xmm1 # 4130 <_sk_callback_sse2+0x359>
+ DB 15,88,206 ; addps %xmm6,%xmm1
DB 184,0,0,0,0 ; mov $0x0,%eax
DB 185,0,0,128,63 ; mov $0x3f800000,%ecx
DB 102,68,15,110,241 ; movd %ecx,%xmm14
DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14
- DB 65,15,40,214 ; movaps %xmm14,%xmm2
- DB 15,194,211,1 ; cmpltps %xmm3,%xmm2
- DB 68,15,40,61,167,47,0,0 ; movaps 0x2fa7(%rip),%xmm15 # 4140 <_sk_callback_sse2+0x369>
- DB 15,40,195 ; movaps %xmm3,%xmm0
- DB 65,15,88,199 ; addps %xmm15,%xmm0
- DB 15,84,194 ; andps %xmm2,%xmm0
- DB 15,85,211 ; andnps %xmm3,%xmm2
- DB 15,86,208 ; orps %xmm0,%xmm2
- DB 102,15,110,200 ; movd %eax,%xmm1
- DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1
- DB 15,41,12,36 ; movaps %xmm1,(%rsp)
- DB 15,40,195 ; movaps %xmm3,%xmm0
+ DB 65,15,40,198 ; movaps %xmm14,%xmm0
DB 15,194,193,1 ; cmpltps %xmm1,%xmm0
- DB 15,40,227 ; movaps %xmm3,%xmm4
- DB 65,15,88,228 ; addps %xmm12,%xmm4
+ DB 68,15,40,61,166,47,0,0 ; movaps 0x2fa6(%rip),%xmm15 # 4140 <_sk_callback_sse2+0x369>
+ DB 15,40,225 ; movaps %xmm1,%xmm4
+ DB 65,15,88,231 ; addps %xmm15,%xmm4
DB 15,84,224 ; andps %xmm0,%xmm4
- DB 15,85,194 ; andnps %xmm2,%xmm0
+ DB 15,85,193 ; andnps %xmm1,%xmm0
DB 15,86,196 ; orps %xmm4,%xmm0
- DB 69,15,40,234 ; movaps %xmm10,%xmm13
+ DB 102,15,110,208 ; movd %eax,%xmm2
+ DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2
+ DB 15,41,84,36,16 ; movaps %xmm2,0x10(%rsp)
+ DB 15,40,225 ; movaps %xmm1,%xmm4
+ DB 15,194,202,1 ; cmpltps %xmm2,%xmm1
+ DB 15,88,231 ; addps %xmm7,%xmm4
+ DB 15,84,225 ; andps %xmm1,%xmm4
+ DB 15,85,200 ; andnps %xmm0,%xmm1
+ DB 15,86,204 ; orps %xmm4,%xmm1
+ DB 69,15,40,236 ; movaps %xmm12,%xmm13
DB 69,15,88,237 ; addps %xmm13,%xmm13
DB 69,15,92,232 ; subps %xmm8,%xmm13
DB 184,171,170,42,62 ; mov $0x3e2aaaab,%eax
DB 69,15,40,200 ; movaps %xmm8,%xmm9
DB 69,15,92,205 ; subps %xmm13,%xmm9
- DB 68,15,89,13,99,47,0,0 ; mulps 0x2f63(%rip),%xmm9 # 4150 <_sk_callback_sse2+0x379>
+ DB 68,15,89,13,101,47,0,0 ; mulps 0x2f65(%rip),%xmm9 # 4150 <_sk_callback_sse2+0x379>
DB 185,171,170,42,63 ; mov $0x3f2aaaab,%ecx
- DB 102,15,110,249 ; movd %ecx,%xmm7
- DB 15,198,255,0 ; shufps $0x0,%xmm7,%xmm7
- DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp)
- DB 15,40,53,90,47,0,0 ; movaps 0x2f5a(%rip),%xmm6 # 4160 <_sk_callback_sse2+0x389>
- DB 15,40,230 ; movaps %xmm6,%xmm4
- DB 15,92,224 ; subps %xmm0,%xmm4
- DB 15,40,208 ; movaps %xmm0,%xmm2
- DB 15,40,200 ; movaps %xmm0,%xmm1
- DB 15,194,199,1 ; cmpltps %xmm7,%xmm0
+ DB 102,15,110,217 ; movd %ecx,%xmm3
+ DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
+ DB 15,41,28,36 ; movaps %xmm3,(%rsp)
+ DB 15,40,45,93,47,0,0 ; movaps 0x2f5d(%rip),%xmm5 # 4160 <_sk_callback_sse2+0x389>
+ DB 15,40,229 ; movaps %xmm5,%xmm4
+ DB 15,92,225 ; subps %xmm1,%xmm4
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 68,15,40,217 ; movaps %xmm1,%xmm11
+ DB 15,40,193 ; movaps %xmm1,%xmm0
+ DB 15,194,203,1 ; cmpltps %xmm3,%xmm1
DB 65,15,89,225 ; mulps %xmm9,%xmm4
DB 65,15,88,229 ; addps %xmm13,%xmm4
- DB 15,84,224 ; andps %xmm0,%xmm4
- DB 65,15,85,197 ; andnps %xmm13,%xmm0
- DB 15,86,196 ; orps %xmm4,%xmm0
- DB 65,15,40,251 ; movaps %xmm11,%xmm7
- DB 15,41,124,36,32 ; movaps %xmm7,0x20(%rsp)
- DB 15,194,207,1 ; cmpltps %xmm7,%xmm1
- DB 65,15,40,224 ; movaps %xmm8,%xmm4
DB 15,84,225 ; andps %xmm1,%xmm4
- DB 15,85,200 ; andnps %xmm0,%xmm1
+ DB 65,15,85,205 ; andnps %xmm13,%xmm1
DB 15,86,204 ; orps %xmm4,%xmm1
- DB 102,15,110,224 ; movd %eax,%xmm4
- DB 15,198,228,0 ; shufps $0x0,%xmm4,%xmm4
- DB 15,194,212,1 ; cmpltps %xmm4,%xmm2
- DB 65,15,89,217 ; mulps %xmm9,%xmm3
- DB 65,15,88,221 ; addps %xmm13,%xmm3
- DB 15,84,218 ; andps %xmm2,%xmm3
- DB 15,85,209 ; andnps %xmm1,%xmm2
- DB 15,86,211 ; orps %xmm3,%xmm2
- DB 68,15,40,92,36,48 ; movaps 0x30(%rsp),%xmm11
- DB 65,15,40,195 ; movaps %xmm11,%xmm0
+ DB 65,15,194,194,1 ; cmpltps %xmm10,%xmm0
+ DB 65,15,40,224 ; movaps %xmm8,%xmm4
+ DB 15,84,224 ; andps %xmm0,%xmm4
+ DB 15,85,193 ; andnps %xmm1,%xmm0
+ DB 15,86,196 ; orps %xmm4,%xmm0
+ DB 102,68,15,110,208 ; movd %eax,%xmm10
+ DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
+ DB 65,15,194,210,1 ; cmpltps %xmm10,%xmm2
+ DB 69,15,89,217 ; mulps %xmm9,%xmm11
+ DB 69,15,88,221 ; addps %xmm13,%xmm11
+ DB 68,15,84,218 ; andps %xmm2,%xmm11
+ DB 15,85,208 ; andnps %xmm0,%xmm2
+ DB 65,15,86,211 ; orps %xmm11,%xmm2
+ DB 15,40,68,36,32 ; movaps 0x20(%rsp),%xmm0
DB 15,85,194 ; andnps %xmm2,%xmm0
DB 15,41,68,36,64 ; movaps %xmm0,0x40(%rsp)
DB 65,15,40,198 ; movaps %xmm14,%xmm0
- DB 15,194,197,1 ; cmpltps %xmm5,%xmm0
- DB 15,40,205 ; movaps %xmm5,%xmm1
+ DB 15,194,198,1 ; cmpltps %xmm6,%xmm0
+ DB 15,40,206 ; movaps %xmm6,%xmm1
DB 65,15,88,207 ; addps %xmm15,%xmm1
DB 15,84,200 ; andps %xmm0,%xmm1
- DB 15,85,197 ; andnps %xmm5,%xmm0
+ DB 15,85,198 ; andnps %xmm6,%xmm0
DB 15,86,193 ; orps %xmm1,%xmm0
- DB 15,40,205 ; movaps %xmm5,%xmm1
- DB 15,194,12,36,1 ; cmpltps (%rsp),%xmm1
- DB 15,40,213 ; movaps %xmm5,%xmm2
- DB 65,15,88,212 ; addps %xmm12,%xmm2
+ DB 15,40,206 ; movaps %xmm6,%xmm1
+ DB 15,194,76,36,16,1 ; cmpltps 0x10(%rsp),%xmm1
+ DB 15,40,214 ; movaps %xmm6,%xmm2
+ DB 15,88,215 ; addps %xmm7,%xmm2
DB 15,84,209 ; andps %xmm1,%xmm2
DB 15,85,200 ; andnps %xmm0,%xmm1
DB 15,86,202 ; orps %xmm2,%xmm1
- DB 15,40,198 ; movaps %xmm6,%xmm0
+ DB 15,40,197 ; movaps %xmm5,%xmm0
DB 15,92,193 ; subps %xmm1,%xmm0
- DB 15,40,209 ; movaps %xmm1,%xmm2
DB 15,40,217 ; movaps %xmm1,%xmm3
- DB 15,194,76,36,16,1 ; cmpltps 0x10(%rsp),%xmm1
+ DB 15,40,225 ; movaps %xmm1,%xmm4
+ DB 15,40,209 ; movaps %xmm1,%xmm2
+ DB 15,194,12,36,1 ; cmpltps (%rsp),%xmm1
DB 65,15,89,193 ; mulps %xmm9,%xmm0
DB 65,15,88,197 ; addps %xmm13,%xmm0
DB 15,84,193 ; andps %xmm1,%xmm0
DB 65,15,85,205 ; andnps %xmm13,%xmm1
DB 15,86,200 ; orps %xmm0,%xmm1
- DB 15,194,223,1 ; cmpltps %xmm7,%xmm3
+ DB 68,15,40,92,36,48 ; movaps 0x30(%rsp),%xmm11
+ DB 65,15,194,211,1 ; cmpltps %xmm11,%xmm2
DB 65,15,40,192 ; movaps %xmm8,%xmm0
- DB 15,84,195 ; andps %xmm3,%xmm0
- DB 15,85,217 ; andnps %xmm1,%xmm3
- DB 15,86,216 ; orps %xmm0,%xmm3
- DB 15,194,212,1 ; cmpltps %xmm4,%xmm2
- DB 65,15,40,193 ; movaps %xmm9,%xmm0
- DB 15,89,197 ; mulps %xmm5,%xmm0
- DB 65,15,88,197 ; addps %xmm13,%xmm0
DB 15,84,194 ; andps %xmm2,%xmm0
- DB 15,85,211 ; andnps %xmm3,%xmm2
+ DB 15,85,209 ; andnps %xmm1,%xmm2
DB 15,86,208 ; orps %xmm0,%xmm2
- DB 65,15,40,219 ; movaps %xmm11,%xmm3
+ DB 65,15,194,218,1 ; cmpltps %xmm10,%xmm3
+ DB 65,15,89,225 ; mulps %xmm9,%xmm4
+ DB 65,15,88,229 ; addps %xmm13,%xmm4
+ DB 15,84,227 ; andps %xmm3,%xmm4
DB 15,85,218 ; andnps %xmm2,%xmm3
- DB 15,88,45,116,46,0,0 ; addps 0x2e74(%rip),%xmm5 # 4170 <_sk_callback_sse2+0x399>
- DB 15,40,197 ; movaps %xmm5,%xmm0
- DB 15,194,4,36,1 ; cmpltps (%rsp),%xmm0
- DB 68,15,194,245,1 ; cmpltps %xmm5,%xmm14
- DB 68,15,88,253 ; addps %xmm5,%xmm15
+ DB 15,86,220 ; orps %xmm4,%xmm3
+ DB 15,40,100,36,32 ; movaps 0x20(%rsp),%xmm4
+ DB 15,40,204 ; movaps %xmm4,%xmm1
+ DB 15,85,203 ; andnps %xmm3,%xmm1
+ DB 15,88,53,112,46,0,0 ; addps 0x2e70(%rip),%xmm6 # 4170 <_sk_callback_sse2+0x399>
+ DB 15,88,254 ; addps %xmm6,%xmm7
+ DB 68,15,194,246,1 ; cmpltps %xmm6,%xmm14
+ DB 68,15,88,254 ; addps %xmm6,%xmm15
DB 69,15,84,254 ; andps %xmm14,%xmm15
- DB 68,15,85,245 ; andnps %xmm5,%xmm14
+ DB 68,15,85,246 ; andnps %xmm6,%xmm14
+ DB 15,194,116,36,16,1 ; cmpltps 0x10(%rsp),%xmm6
DB 69,15,86,247 ; orps %xmm15,%xmm14
- DB 68,15,88,229 ; addps %xmm5,%xmm12
- DB 68,15,84,224 ; andps %xmm0,%xmm12
- DB 65,15,85,198 ; andnps %xmm14,%xmm0
- DB 65,15,86,196 ; orps %xmm12,%xmm0
- DB 15,40,248 ; movaps %xmm0,%xmm7
- DB 15,194,252,1 ; cmpltps %xmm4,%xmm7
- DB 15,40,200 ; movaps %xmm0,%xmm1
- DB 15,194,76,36,32,1 ; cmpltps 0x20(%rsp),%xmm1
- DB 15,92,240 ; subps %xmm0,%xmm6
- DB 15,194,68,36,16,1 ; cmpltps 0x10(%rsp),%xmm0
+ DB 15,84,254 ; andps %xmm6,%xmm7
+ DB 65,15,85,246 ; andnps %xmm14,%xmm6
+ DB 15,86,247 ; orps %xmm7,%xmm6
+ DB 15,40,254 ; movaps %xmm6,%xmm7
+ DB 65,15,194,250,1 ; cmpltps %xmm10,%xmm7
+ DB 15,40,198 ; movaps %xmm6,%xmm0
+ DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0
+ DB 15,92,238 ; subps %xmm6,%xmm5
+ DB 15,40,222 ; movaps %xmm6,%xmm3
+ DB 15,194,52,36,1 ; cmpltps (%rsp),%xmm6
+ DB 65,15,89,217 ; mulps %xmm9,%xmm3
DB 65,15,89,233 ; mulps %xmm9,%xmm5
- DB 65,15,89,241 ; mulps %xmm9,%xmm6
+ DB 65,15,88,221 ; addps %xmm13,%xmm3
DB 65,15,88,237 ; addps %xmm13,%xmm5
- DB 65,15,88,245 ; addps %xmm13,%xmm6
- DB 15,84,240 ; andps %xmm0,%xmm6
- DB 65,15,85,197 ; andnps %xmm13,%xmm0
- DB 15,86,198 ; orps %xmm6,%xmm0
- DB 68,15,84,193 ; andps %xmm1,%xmm8
- DB 15,85,200 ; andnps %xmm0,%xmm1
- DB 65,15,86,200 ; orps %xmm8,%xmm1
- DB 15,84,239 ; andps %xmm7,%xmm5
- DB 15,85,249 ; andnps %xmm1,%xmm7
- DB 15,86,253 ; orps %xmm5,%xmm7
- DB 69,15,84,211 ; andps %xmm11,%xmm10
- DB 68,15,85,223 ; andnps %xmm7,%xmm11
- DB 15,40,76,36,64 ; movaps 0x40(%rsp),%xmm1
- DB 65,15,86,202 ; orps %xmm10,%xmm1
- DB 65,15,86,218 ; orps %xmm10,%xmm3
- DB 69,15,86,211 ; orps %xmm11,%xmm10
+ DB 15,84,238 ; andps %xmm6,%xmm5
+ DB 65,15,85,245 ; andnps %xmm13,%xmm6
+ DB 15,86,245 ; orps %xmm5,%xmm6
+ DB 68,15,84,192 ; andps %xmm0,%xmm8
+ DB 15,85,198 ; andnps %xmm6,%xmm0
+ DB 65,15,86,192 ; orps %xmm8,%xmm0
+ DB 15,84,223 ; andps %xmm7,%xmm3
+ DB 15,85,248 ; andnps %xmm0,%xmm7
+ DB 15,86,251 ; orps %xmm3,%xmm7
+ DB 15,40,196 ; movaps %xmm4,%xmm0
+ DB 68,15,84,224 ; andps %xmm0,%xmm12
+ DB 15,85,199 ; andnps %xmm7,%xmm0
+ DB 15,40,84,36,64 ; movaps 0x40(%rsp),%xmm2
+ DB 65,15,86,212 ; orps %xmm12,%xmm2
+ DB 65,15,86,204 ; orps %xmm12,%xmm1
+ DB 68,15,86,224 ; orps %xmm0,%xmm12
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,193 ; movaps %xmm1,%xmm0
- DB 15,40,203 ; movaps %xmm3,%xmm1
- DB 65,15,40,210 ; movaps %xmm10,%xmm2
+ DB 15,40,194 ; movaps %xmm2,%xmm0
+ DB 65,15,40,212 ; movaps %xmm12,%xmm2
DB 15,40,92,36,80 ; movaps 0x50(%rsp),%xmm3
DB 15,40,100,36,96 ; movaps 0x60(%rsp),%xmm4
DB 15,40,108,36,112 ; movaps 0x70(%rsp),%xmm5
diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp
index 09ebebf8d4..a74cb7badb 100644
--- a/src/jumper/SkJumper_stages.cpp
+++ b/src/jumper/SkJumper_stages.cpp
@@ -508,14 +508,14 @@ STAGE(hsl_to_rgb) {
p = 2.0f*l - q;
auto hue_to_rgb = [&](F t) {
- F t2 = if_then_else(t < 0.0_f, t + 1.0f,
- if_then_else(t > 1.0_f, t - 1.0f,
- t));
-
- return if_then_else(t2 < C(1/6.0f), p + (q-p)*6.0f*t,
- if_then_else(t2 < C(3/6.0f), q,
- if_then_else(t2 < C(4/6.0f), p + (q-p)*6.0f*((4/6.0f) - t2),
- p)));
+ t = if_then_else(t < 0.0_f, t + 1.0f,
+ if_then_else(t > 1.0_f, t - 1.0f,
+ t));
+
+ return if_then_else(t < C(1/6.0f), p + (q-p)*6.0f*t,
+ if_then_else(t < C(3/6.0f), q,
+ if_then_else(t < C(4/6.0f), p + (q-p)*6.0f*((4/6.0f) - t),
+ p)));
};
r = if_then_else(s == 0, l, hue_to_rgb(h + (1/3.0f)));