From 51e46d5e1c48072c08637ba8f875ba249eda2376 Mon Sep 17 00:00:00 2001 From: Mike Reed Date: Fri, 23 Jun 2017 14:21:25 -0400 Subject: [PATCH] use mul_inv instead of div for tiling Bug: skia: Change-Id: I5f88e7923fe204faba8dc5d87454805a4d470d52 Reviewed-on: https://skia-review.googlesource.com/20688 Reviewed-by: Mike Klein Reviewed-by: Florin Malita Commit-Queue: Mike Reed --- src/jumper/SkJumper.h | 5 + src/jumper/SkJumper_generated.S | 3432 ++++++++++++++------------- src/jumper/SkJumper_generated_win.S | 3258 ++++++++++++------------- src/jumper/SkJumper_stages.cpp | 28 +- src/shaders/SkImageShader.cpp | 8 +- 5 files changed, 3422 insertions(+), 3309 deletions(-) diff --git a/src/jumper/SkJumper.h b/src/jumper/SkJumper.h index 8e1dd35300..38ff934aef 100644 --- a/src/jumper/SkJumper.h +++ b/src/jumper/SkJumper.h @@ -71,6 +71,11 @@ struct SkJumper_SamplerCtx { float scaley[SkJumper_kMaxStride]; }; +struct SkJumper_TileCtx { + float scale; + float invScale; // cache of 1/scale +}; + struct SkJumper_CallbackCtx { MAYBE_MSABI void (*fn)(SkJumper_CallbackCtx* self, int active_pixels/*<= SkJumper_kMaxStride*/); diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index 743ad388d6..108c55493e 100644 --- a/src/jumper/SkJumper_generated.S +++ b/src/jumper/SkJumper_generated.S @@ -3776,14 +3776,14 @@ HIDDEN _sk_repeat_x_aarch64 FUNCTION(_sk_repeat_x_aarch64) _sk_repeat_x_aarch64: .long 0xa8c11428 // ldp x8, x5, [x1], #16 + .long 0x4ddfc910 // ld1r {v16.4s}, [x8], #4 + .long 0xbd400111 // ldr s17, [x8] + .long 0x4f919011 // fmul v17.4s, v0.4s, v17.s[0] + .long 0x4e219a31 // frintm v17.4s, v17.4s + .long 0x4eb1ce00 // fmls v0.4s, v16.4s, v17.4s .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff - .long 0xbd400110 // ldr s16, [x8] - .long 0x4e040612 // dup v18.4s, v16.s[0] - .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s - .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s - .long 0x4e219a52 // frintm v18.4s, v18.4s - .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0] - .long 0x4eb1f400 // fmin v0.4s, v0.4s, v17.4s + .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s + .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s .long 0xd61f00a0 // br x5 HIDDEN _sk_repeat_y_aarch64 @@ -3791,14 +3791,14 @@ HIDDEN _sk_repeat_y_aarch64 FUNCTION(_sk_repeat_y_aarch64) _sk_repeat_y_aarch64: .long 0xa8c11428 // ldp x8, x5, [x1], #16 + .long 0x4ddfc910 // ld1r {v16.4s}, [x8], #4 + .long 0xbd400111 // ldr s17, [x8] + .long 0x4f919031 // fmul v17.4s, v1.4s, v17.s[0] + .long 0x4e219a31 // frintm v17.4s, v17.4s + .long 0x4eb1ce01 // fmls v1.4s, v16.4s, v17.4s .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff - .long 0xbd400110 // ldr s16, [x8] - .long 0x4e040612 // dup v18.4s, v16.s[0] - .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s - .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s - .long 0x4e219a52 // frintm v18.4s, v18.4s - .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0] - .long 0x4eb1f421 // fmin v1.4s, v1.4s, v17.4s + .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s + .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s .long 0xd61f00a0 // br x5 HIDDEN _sk_mirror_x_aarch64 @@ -3806,19 +3806,20 @@ HIDDEN _sk_mirror_x_aarch64 FUNCTION(_sk_mirror_x_aarch64) _sk_mirror_x_aarch64: .long 0xa8c11428 // ldp x8, x5, [x1], #16 - .long 0xbd400110 // ldr s16, [x8] - .long 0x4e040611 // dup v17.4s, v16.s[0] - .long 0x1e302a10 // fadd s16, s16, s16 - .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s - .long 0x4e040612 // dup v18.4s, v16.s[0] - .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s - .long 0x4e219a52 // frintm v18.4s, v18.4s - .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0] - .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff - .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s - .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s + .long 0x1e2c1012 // fmov s18, #5.000000000000000000e-01 + .long 0x6f07e7f3 // movi v19.2d, #0xffffffffffffffff + .long 0x2d404111 // ldp s17, s16, [x8] + .long 0x1e320a10 // fmul s16, s16, s18 + .long 0x4e040632 // dup v18.4s, v17.s[0] + .long 0x4eb2d400 // fsub v0.4s, v0.4s, v18.4s + .long 0x4f909010 // fmul v16.4s, v0.4s, v16.s[0] + .long 0x1e312a31 // fadd s17, s17, s17 + .long 0x4e219a10 // frintm v16.4s, v16.4s + .long 0x4f915200 // fmls v0.4s, v16.4s, v17.s[0] + .long 0x4eb2d400 // fsub v0.4s, v0.4s, v18.4s + .long 0x4eb38653 // add v19.4s, v18.4s, v19.4s .long 0x4ea0f800 // fabs v0.4s, v0.4s - .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s + .long 0x4eb3f400 // fmin v0.4s, v0.4s, v19.4s .long 0xd61f00a0 // br x5 HIDDEN _sk_mirror_y_aarch64 @@ -3826,19 +3827,20 @@ HIDDEN _sk_mirror_y_aarch64 FUNCTION(_sk_mirror_y_aarch64) _sk_mirror_y_aarch64: .long 0xa8c11428 // ldp x8, x5, [x1], #16 - .long 0xbd400110 // ldr s16, [x8] - .long 0x4e040611 // dup v17.4s, v16.s[0] - .long 0x1e302a10 // fadd s16, s16, s16 - .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s - .long 0x4e040612 // dup v18.4s, v16.s[0] - .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s - .long 0x4e219a52 // frintm v18.4s, v18.4s - .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0] - .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff - .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s - .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s + .long 0x1e2c1012 // fmov s18, #5.000000000000000000e-01 + .long 0x6f07e7f3 // movi v19.2d, #0xffffffffffffffff + .long 0x2d404111 // ldp s17, s16, [x8] + .long 0x1e320a10 // fmul s16, s16, s18 + .long 0x4e040632 // dup v18.4s, v17.s[0] + .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s + .long 0x4f909030 // fmul v16.4s, v1.4s, v16.s[0] + .long 0x1e312a31 // fadd s17, s17, s17 + .long 0x4e219a10 // frintm v16.4s, v16.4s + .long 0x4f915201 // fmls v1.4s, v16.4s, v17.s[0] + .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s + .long 0x4eb38653 // add v19.4s, v18.4s, v19.4s .long 0x4ea0f821 // fabs v1.4s, v1.4s - .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s + .long 0x4eb3f421 // fmin v1.4s, v1.4s, v19.4s .long 0xd61f00a0 // br x5 HIDDEN _sk_clamp_x_1_aarch64 @@ -4200,7 +4202,7 @@ _sk_gradient_aarch64: .long 0x6f00e411 // movi v17.2d, #0x0 .long 0xf9400109 // ldr x9, [x8] .long 0xf100093f // cmp x9, #0x2 - .long 0x540001c3 // b.cc 38f8 // b.lo, b.ul, b.last + .long 0x540001c3 // b.cc 3900 // b.lo, b.ul, b.last .long 0xf940250a // ldr x10, [x8, #72] .long 0xd1000529 // sub x9, x9, #0x1 .long 0x6f00e401 // movi v1.2d, #0x0 @@ -4211,7 +4213,7 @@ _sk_gradient_aarch64: .long 0x6e23e403 // fcmge v3.4s, v0.4s, v3.4s .long 0x4e221c63 // and v3.16b, v3.16b, v2.16b .long 0x4ea18461 // add v1.4s, v3.4s, v1.4s - .long 0xb5ffff69 // cbnz x9, 38d8 + .long 0xb5ffff69 // cbnz x9, 38e0 .long 0x6f20a431 // uxtl2 v17.2d, v1.4s .long 0x2f20a421 // uxtl v1.2d, v1.2s .long 0xa940b10a // ldp x10, x12, [x8, #8] @@ -8630,29 +8632,27 @@ HIDDEN _sk_repeat_x_vfp4 FUNCTION(_sk_repeat_x_vfp4) _sk_repeat_x_vfp4: .long 0xe92d4800 // push {fp, lr} - .long 0xed2d8b04 // vpush {d8-d9} .long 0xe591e000 // ldr lr, [r1] - .long 0xf2c02010 // vmov.i32 d18, #0 - .long 0xeddf3b12 // vldr d19, [pc, #72] + .long 0xf2c03010 // vmov.i32 d19, #0 + .long 0xeddf4b11 // vldr d20, [pc, #68] .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 - .long 0xed9e8a00 // vldr s16, [lr] - .long 0xeec09a88 // vdiv.f32 s19, s1, s16 - .long 0xee809a08 // vdiv.f32 s18, s0, s16 - .long 0xf3fb0709 // vcvt.s32.f32 d16, d9 - .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 - .long 0xf3601e89 // vcgt.f32 d17, d16, d9 - .long 0xf35311b2 // vbsl d17, d19, d18 - .long 0xf3f42c08 // vdup.32 d18, d8[0] - .long 0xf2600da1 // vsub.f32 d16, d16, d17 - .long 0xf3c71e1f // vmov.i8 d17, #255 - .long 0xf26218a1 // vadd.i32 d17, d18, d17 - .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0] - .long 0xf2600d20 // vsub.f32 d16, d0, d16 - .long 0xf2200fa1 // vmin.f32 d0, d16, d17 - .long 0xecbd8b04 // vpop {d8-d9} + .long 0xf4ee0c9d // vld1.32 {d16[]}, [lr :32]! + .long 0xf4ee1c9f // vld1.32 {d17[]}, [lr :32] + .long 0xf3411d90 // vmul.f32 d17, d17, d0 + .long 0xf3fb2721 // vcvt.s32.f32 d18, d17 + .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 + .long 0xf3621ea1 // vcgt.f32 d17, d18, d17 + .long 0xf35411b3 // vbsl d17, d20, d19 + .long 0xf2621da1 // vsub.f32 d17, d18, d17 + .long 0xf3c72e1f // vmov.i8 d18, #255 + .long 0xf3401db1 // vmul.f32 d17, d16, d17 + .long 0xf26008a2 // vadd.i32 d16, d16, d18 + .long 0xf2601d21 // vsub.f32 d17, d0, d17 + .long 0xf2210fa0 // vmin.f32 d0, d17, d16 .long 0xe8bd4800 // pop {fp, lr} .long 0xe12fff1c // bx ip + .long 0xe320f000 // nop {0} .long 0x3f800000 // .word 0x3f800000 .long 0x3f800000 // .word 0x3f800000 @@ -8661,29 +8661,27 @@ HIDDEN _sk_repeat_y_vfp4 FUNCTION(_sk_repeat_y_vfp4) _sk_repeat_y_vfp4: .long 0xe92d4800 // push {fp, lr} - .long 0xed2d8b04 // vpush {d8-d9} .long 0xe591e000 // ldr lr, [r1] - .long 0xf2c02010 // vmov.i32 d18, #0 - .long 0xeddf3b12 // vldr d19, [pc, #72] + .long 0xf2c03010 // vmov.i32 d19, #0 + .long 0xeddf4b11 // vldr d20, [pc, #68] .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 - .long 0xed9e8a00 // vldr s16, [lr] - .long 0xeec19a88 // vdiv.f32 s19, s3, s16 - .long 0xee819a08 // vdiv.f32 s18, s2, s16 - .long 0xf3fb0709 // vcvt.s32.f32 d16, d9 - .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 - .long 0xf3601e89 // vcgt.f32 d17, d16, d9 - .long 0xf35311b2 // vbsl d17, d19, d18 - .long 0xf3f42c08 // vdup.32 d18, d8[0] - .long 0xf2600da1 // vsub.f32 d16, d16, d17 - .long 0xf3c71e1f // vmov.i8 d17, #255 - .long 0xf26218a1 // vadd.i32 d17, d18, d17 - .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0] - .long 0xf2610d20 // vsub.f32 d16, d1, d16 - .long 0xf2201fa1 // vmin.f32 d1, d16, d17 - .long 0xecbd8b04 // vpop {d8-d9} + .long 0xf4ee0c9d // vld1.32 {d16[]}, [lr :32]! + .long 0xf4ee1c9f // vld1.32 {d17[]}, [lr :32] + .long 0xf3411d91 // vmul.f32 d17, d17, d1 + .long 0xf3fb2721 // vcvt.s32.f32 d18, d17 + .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 + .long 0xf3621ea1 // vcgt.f32 d17, d18, d17 + .long 0xf35411b3 // vbsl d17, d20, d19 + .long 0xf2621da1 // vsub.f32 d17, d18, d17 + .long 0xf3c72e1f // vmov.i8 d18, #255 + .long 0xf3401db1 // vmul.f32 d17, d16, d17 + .long 0xf26008a2 // vadd.i32 d16, d16, d18 + .long 0xf2611d21 // vsub.f32 d17, d1, d17 + .long 0xf2211fa0 // vmin.f32 d1, d17, d16 .long 0xe8bd4800 // pop {fp, lr} .long 0xe12fff1c // bx ip + .long 0xe320f000 // nop {0} .long 0x3f800000 // .word 0x3f800000 .long 0x3f800000 // .word 0x3f800000 @@ -8692,31 +8690,33 @@ HIDDEN _sk_mirror_x_vfp4 FUNCTION(_sk_mirror_x_vfp4) _sk_mirror_x_vfp4: .long 0xe92d4800 // push {fp, lr} - .long 0xed2d8b04 // vpush {d8-d9} + .long 0xed2d8b06 // vpush {d8-d10} + .long 0xeeb68a00 // vmov.f32 s16, #96 .long 0xe591e000 // ldr lr, [r1] - .long 0xf2c03010 // vmov.i32 d19, #0 - .long 0xeddf4b16 // vldr d20, [pc, #88] + .long 0xf2c04010 // vmov.i32 d20, #0 + .long 0xeddf5b17 // vldr d21, [pc, #92] .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 - .long 0xed9e8a00 // vldr s16, [lr] - .long 0xee389a08 // vadd.f32 s18, s16, s16 - .long 0xf3f40c08 // vdup.32 d16, d8[0] - .long 0xf2200d20 // vsub.f32 d0, d0, d16 - .long 0xeec08a89 // vdiv.f32 s17, s1, s18 - .long 0xee808a09 // vdiv.f32 s16, s0, s18 - .long 0xf3fb1708 // vcvt.s32.f32 d17, d8 - .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 - .long 0xf3612e88 // vcgt.f32 d18, d17, d8 - .long 0xf35421b3 // vbsl d18, d20, d19 + .long 0xed9e9a00 // vldr s18, [lr] + .long 0xed9eaa01 // vldr s20, [lr, #4] + .long 0xf3f40c09 // vdup.32 d16, d9[0] + .long 0xf2601d20 // vsub.f32 d17, d0, d16 + .long 0xee2a0a08 // vmul.f32 s0, s20, s16 + .long 0xf2e129c0 // vmul.f32 d18, d17, d0[0] + .long 0xee390a09 // vadd.f32 s0, s18, s18 + .long 0xf3fb3722 // vcvt.s32.f32 d19, d18 + .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 + .long 0xf3632ea2 // vcgt.f32 d18, d19, d18 + .long 0xf35521b4 // vbsl d18, d21, d20 + .long 0xf2632da2 // vsub.f32 d18, d19, d18 + .long 0xf2e229c0 // vmul.f32 d18, d18, d0[0] .long 0xf2611da2 // vsub.f32 d17, d17, d18 .long 0xf3c72e1f // vmov.i8 d18, #255 - .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0] - .long 0xf2601d21 // vsub.f32 d17, d0, d17 .long 0xf2611da0 // vsub.f32 d17, d17, d16 .long 0xf26008a2 // vadd.i32 d16, d16, d18 .long 0xf3f91721 // vabs.f32 d17, d17 .long 0xf2210fa0 // vmin.f32 d0, d17, d16 - .long 0xecbd8b04 // vpop {d8-d9} + .long 0xecbd8b06 // vpop {d8-d10} .long 0xe8bd4800 // pop {fp, lr} .long 0xe12fff1c // bx ip .long 0x3f800000 // .word 0x3f800000 @@ -8727,31 +8727,33 @@ HIDDEN _sk_mirror_y_vfp4 FUNCTION(_sk_mirror_y_vfp4) _sk_mirror_y_vfp4: .long 0xe92d4800 // push {fp, lr} - .long 0xed2d8b04 // vpush {d8-d9} + .long 0xed2d8b06 // vpush {d8-d10} + .long 0xeeb68a00 // vmov.f32 s16, #96 .long 0xe591e000 // ldr lr, [r1] - .long 0xf2c03010 // vmov.i32 d19, #0 - .long 0xeddf4b16 // vldr d20, [pc, #88] + .long 0xf2c04010 // vmov.i32 d20, #0 + .long 0xeddf5b17 // vldr d21, [pc, #92] .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 - .long 0xed9e8a00 // vldr s16, [lr] - .long 0xee389a08 // vadd.f32 s18, s16, s16 - .long 0xf3f40c08 // vdup.32 d16, d8[0] - .long 0xf2211d20 // vsub.f32 d1, d1, d16 - .long 0xeec18a89 // vdiv.f32 s17, s3, s18 - .long 0xee818a09 // vdiv.f32 s16, s2, s18 - .long 0xf3fb1708 // vcvt.s32.f32 d17, d8 - .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 - .long 0xf3612e88 // vcgt.f32 d18, d17, d8 - .long 0xf35421b3 // vbsl d18, d20, d19 + .long 0xed9e9a00 // vldr s18, [lr] + .long 0xed9eaa01 // vldr s20, [lr, #4] + .long 0xf3f40c09 // vdup.32 d16, d9[0] + .long 0xf2611d20 // vsub.f32 d17, d1, d16 + .long 0xee2a1a08 // vmul.f32 s2, s20, s16 + .long 0xf2e129c1 // vmul.f32 d18, d17, d1[0] + .long 0xee391a09 // vadd.f32 s2, s18, s18 + .long 0xf3fb3722 // vcvt.s32.f32 d19, d18 + .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 + .long 0xf3632ea2 // vcgt.f32 d18, d19, d18 + .long 0xf35521b4 // vbsl d18, d21, d20 + .long 0xf2632da2 // vsub.f32 d18, d19, d18 + .long 0xf2e229c1 // vmul.f32 d18, d18, d1[0] .long 0xf2611da2 // vsub.f32 d17, d17, d18 .long 0xf3c72e1f // vmov.i8 d18, #255 - .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0] - .long 0xf2611d21 // vsub.f32 d17, d1, d17 .long 0xf2611da0 // vsub.f32 d17, d17, d16 .long 0xf26008a2 // vadd.i32 d16, d16, d18 .long 0xf3f91721 // vabs.f32 d17, d17 .long 0xf2211fa0 // vmin.f32 d1, d17, d16 - .long 0xecbd8b04 // vpop {d8-d9} + .long 0xecbd8b06 // vpop {d8-d10} .long 0xe8bd4800 // pop {fp, lr} .long 0xe12fff1c // bx ip .long 0x3f800000 // .word 0x3f800000 @@ -9826,7 +9828,7 @@ _sk_seed_shader_hsw: .byte 197,249,110,194 // vmovd %edx,%xmm0 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,222,74,0,0 // vbroadcastss 0x4ade(%rip),%ymm1 # 4bb8 <_sk_callback_hsw+0x142> + .byte 196,226,125,24,13,14,75,0,0 // vbroadcastss 0x4b0e(%rip),%ymm1 # 4be8 <_sk_callback_hsw+0x144> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,88,7 // vaddps (%rdi),%ymm0,%ymm0 .byte 197,249,110,209 // vmovd %ecx,%xmm2 @@ -9834,7 +9836,7 @@ _sk_seed_shader_hsw: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,21,190,74,0,0 // vbroadcastss 0x4abe(%rip),%ymm2 # 4bbc <_sk_callback_hsw+0x146> + .byte 196,226,125,24,21,238,74,0,0 // vbroadcastss 0x4aee(%rip),%ymm2 # 4bec <_sk_callback_hsw+0x148> .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 @@ -9853,13 +9855,13 @@ _sk_dither_hsw: .byte 197,121,110,201 // vmovd %ecx,%xmm9 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 .byte 196,65,53,239,200 // vpxor %ymm8,%ymm9,%ymm9 - .byte 196,98,125,88,21,133,74,0,0 // vpbroadcastd 0x4a85(%rip),%ymm10 # 4bc0 <_sk_callback_hsw+0x14a> + .byte 196,98,125,88,21,181,74,0,0 // vpbroadcastd 0x4ab5(%rip),%ymm10 # 4bf0 <_sk_callback_hsw+0x14c> .byte 196,65,53,219,218 // vpand %ymm10,%ymm9,%ymm11 .byte 196,193,37,114,243,5 // vpslld $0x5,%ymm11,%ymm11 .byte 196,65,61,219,210 // vpand %ymm10,%ymm8,%ymm10 .byte 196,193,45,114,242,4 // vpslld $0x4,%ymm10,%ymm10 - .byte 196,98,125,88,37,106,74,0,0 // vpbroadcastd 0x4a6a(%rip),%ymm12 # 4bc4 <_sk_callback_hsw+0x14e> - .byte 196,98,125,88,45,101,74,0,0 // vpbroadcastd 0x4a65(%rip),%ymm13 # 4bc8 <_sk_callback_hsw+0x152> + .byte 196,98,125,88,37,154,74,0,0 // vpbroadcastd 0x4a9a(%rip),%ymm12 # 4bf4 <_sk_callback_hsw+0x150> + .byte 196,98,125,88,45,149,74,0,0 // vpbroadcastd 0x4a95(%rip),%ymm13 # 4bf8 <_sk_callback_hsw+0x154> .byte 196,65,53,219,245 // vpand %ymm13,%ymm9,%ymm14 .byte 196,193,13,114,246,2 // vpslld $0x2,%ymm14,%ymm14 .byte 196,65,61,219,237 // vpand %ymm13,%ymm8,%ymm13 @@ -9874,8 +9876,8 @@ _sk_dither_hsw: .byte 196,65,61,235,194 // vpor %ymm10,%ymm8,%ymm8 .byte 196,65,61,235,193 // vpor %ymm9,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,23,74,0,0 // vbroadcastss 0x4a17(%rip),%ymm9 # 4bcc <_sk_callback_hsw+0x156> - .byte 196,98,125,24,21,18,74,0,0 // vbroadcastss 0x4a12(%rip),%ymm10 # 4bd0 <_sk_callback_hsw+0x15a> + .byte 196,98,125,24,13,71,74,0,0 // vbroadcastss 0x4a47(%rip),%ymm9 # 4bfc <_sk_callback_hsw+0x158> + .byte 196,98,125,24,21,66,74,0,0 // vbroadcastss 0x4a42(%rip),%ymm10 # 4c00 <_sk_callback_hsw+0x15c> .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 @@ -9944,7 +9946,7 @@ HIDDEN _sk_srcatop_hsw FUNCTION(_sk_srcatop_hsw) _sk_srcatop_hsw: .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 - .byte 196,98,125,24,5,106,73,0,0 // vbroadcastss 0x496a(%rip),%ymm8 # 4bd4 <_sk_callback_hsw+0x15e> + .byte 196,98,125,24,5,154,73,0,0 // vbroadcastss 0x499a(%rip),%ymm8 # 4c04 <_sk_callback_hsw+0x160> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,226,61,184,196 // vfmadd231ps %ymm4,%ymm8,%ymm0 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 @@ -9960,7 +9962,7 @@ HIDDEN _sk_dstatop_hsw .globl _sk_dstatop_hsw FUNCTION(_sk_dstatop_hsw) _sk_dstatop_hsw: - .byte 196,98,125,24,5,61,73,0,0 // vbroadcastss 0x493d(%rip),%ymm8 # 4bd8 <_sk_callback_hsw+0x162> + .byte 196,98,125,24,5,109,73,0,0 // vbroadcastss 0x496d(%rip),%ymm8 # 4c08 <_sk_callback_hsw+0x164> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 196,226,101,184,196 // vfmadd231ps %ymm4,%ymm3,%ymm0 @@ -9999,7 +10001,7 @@ HIDDEN _sk_srcout_hsw .globl _sk_srcout_hsw FUNCTION(_sk_srcout_hsw) _sk_srcout_hsw: - .byte 196,98,125,24,5,228,72,0,0 // vbroadcastss 0x48e4(%rip),%ymm8 # 4bdc <_sk_callback_hsw+0x166> + .byte 196,98,125,24,5,20,73,0,0 // vbroadcastss 0x4914(%rip),%ymm8 # 4c0c <_sk_callback_hsw+0x168> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -10012,7 +10014,7 @@ HIDDEN _sk_dstout_hsw .globl _sk_dstout_hsw FUNCTION(_sk_dstout_hsw) _sk_dstout_hsw: - .byte 196,226,125,24,5,199,72,0,0 // vbroadcastss 0x48c7(%rip),%ymm0 # 4be0 <_sk_callback_hsw+0x16a> + .byte 196,226,125,24,5,247,72,0,0 // vbroadcastss 0x48f7(%rip),%ymm0 # 4c10 <_sk_callback_hsw+0x16c> .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 @@ -10025,7 +10027,7 @@ HIDDEN _sk_srcover_hsw .globl _sk_srcover_hsw FUNCTION(_sk_srcover_hsw) _sk_srcover_hsw: - .byte 196,98,125,24,5,170,72,0,0 // vbroadcastss 0x48aa(%rip),%ymm8 # 4be4 <_sk_callback_hsw+0x16e> + .byte 196,98,125,24,5,218,72,0,0 // vbroadcastss 0x48da(%rip),%ymm8 # 4c14 <_sk_callback_hsw+0x170> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0 .byte 196,194,85,184,200 // vfmadd231ps %ymm8,%ymm5,%ymm1 @@ -10038,7 +10040,7 @@ HIDDEN _sk_dstover_hsw .globl _sk_dstover_hsw FUNCTION(_sk_dstover_hsw) _sk_dstover_hsw: - .byte 196,98,125,24,5,137,72,0,0 // vbroadcastss 0x4889(%rip),%ymm8 # 4be8 <_sk_callback_hsw+0x172> + .byte 196,98,125,24,5,185,72,0,0 // vbroadcastss 0x48b9(%rip),%ymm8 # 4c18 <_sk_callback_hsw+0x174> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 @@ -10062,7 +10064,7 @@ HIDDEN _sk_multiply_hsw .globl _sk_multiply_hsw FUNCTION(_sk_multiply_hsw) _sk_multiply_hsw: - .byte 196,98,125,24,5,84,72,0,0 // vbroadcastss 0x4854(%rip),%ymm8 # 4bec <_sk_callback_hsw+0x176> + .byte 196,98,125,24,5,132,72,0,0 // vbroadcastss 0x4884(%rip),%ymm8 # 4c1c <_sk_callback_hsw+0x178> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -10110,7 +10112,7 @@ HIDDEN _sk_xor__hsw .globl _sk_xor__hsw FUNCTION(_sk_xor__hsw) _sk_xor__hsw: - .byte 196,98,125,24,5,207,71,0,0 // vbroadcastss 0x47cf(%rip),%ymm8 # 4bf0 <_sk_callback_hsw+0x17a> + .byte 196,98,125,24,5,255,71,0,0 // vbroadcastss 0x47ff(%rip),%ymm8 # 4c20 <_sk_callback_hsw+0x17c> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -10144,7 +10146,7 @@ _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 196,98,125,24,5,87,71,0,0 // vbroadcastss 0x4757(%rip),%ymm8 # 4bf4 <_sk_callback_hsw+0x17e> + .byte 196,98,125,24,5,135,71,0,0 // vbroadcastss 0x4787(%rip),%ymm8 # 4c24 <_sk_callback_hsw+0x180> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -10169,7 +10171,7 @@ _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 196,98,125,24,5,6,71,0,0 // vbroadcastss 0x4706(%rip),%ymm8 # 4bf8 <_sk_callback_hsw+0x182> + .byte 196,98,125,24,5,54,71,0,0 // vbroadcastss 0x4736(%rip),%ymm8 # 4c28 <_sk_callback_hsw+0x184> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -10197,7 +10199,7 @@ _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 196,98,125,24,5,169,70,0,0 // vbroadcastss 0x46a9(%rip),%ymm8 # 4bfc <_sk_callback_hsw+0x186> + .byte 196,98,125,24,5,217,70,0,0 // vbroadcastss 0x46d9(%rip),%ymm8 # 4c2c <_sk_callback_hsw+0x188> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -10219,7 +10221,7 @@ _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 196,98,125,24,5,103,70,0,0 // vbroadcastss 0x4667(%rip),%ymm8 # 4c00 <_sk_callback_hsw+0x18a> + .byte 196,98,125,24,5,151,70,0,0 // vbroadcastss 0x4697(%rip),%ymm8 # 4c30 <_sk_callback_hsw+0x18c> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -10229,7 +10231,7 @@ HIDDEN _sk_colorburn_hsw .globl _sk_colorburn_hsw FUNCTION(_sk_colorburn_hsw) _sk_colorburn_hsw: - .byte 196,98,125,24,5,85,70,0,0 // vbroadcastss 0x4655(%rip),%ymm8 # 4c04 <_sk_callback_hsw+0x18e> + .byte 196,98,125,24,5,133,70,0,0 // vbroadcastss 0x4685(%rip),%ymm8 # 4c34 <_sk_callback_hsw+0x190> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 @@ -10287,7 +10289,7 @@ HIDDEN _sk_colordodge_hsw FUNCTION(_sk_colordodge_hsw) _sk_colordodge_hsw: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 196,98,125,24,13,96,69,0,0 // vbroadcastss 0x4560(%rip),%ymm9 # 4c08 <_sk_callback_hsw+0x192> + .byte 196,98,125,24,13,144,69,0,0 // vbroadcastss 0x4590(%rip),%ymm9 # 4c38 <_sk_callback_hsw+0x194> .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 @@ -10340,7 +10342,7 @@ HIDDEN _sk_hardlight_hsw .globl _sk_hardlight_hsw FUNCTION(_sk_hardlight_hsw) _sk_hardlight_hsw: - .byte 196,98,125,24,5,129,68,0,0 // vbroadcastss 0x4481(%rip),%ymm8 # 4c0c <_sk_callback_hsw+0x196> + .byte 196,98,125,24,5,177,68,0,0 // vbroadcastss 0x44b1(%rip),%ymm8 # 4c3c <_sk_callback_hsw+0x198> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -10391,7 +10393,7 @@ HIDDEN _sk_overlay_hsw .globl _sk_overlay_hsw FUNCTION(_sk_overlay_hsw) _sk_overlay_hsw: - .byte 196,98,125,24,5,185,67,0,0 // vbroadcastss 0x43b9(%rip),%ymm8 # 4c10 <_sk_callback_hsw+0x19a> + .byte 196,98,125,24,5,233,67,0,0 // vbroadcastss 0x43e9(%rip),%ymm8 # 4c40 <_sk_callback_hsw+0x19c> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -10452,10 +10454,10 @@ _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 196,98,125,24,29,196,66,0,0 // vbroadcastss 0x42c4(%rip),%ymm11 # 4c18 <_sk_callback_hsw+0x1a2> + .byte 196,98,125,24,29,244,66,0,0 // vbroadcastss 0x42f4(%rip),%ymm11 # 4c48 <_sk_callback_hsw+0x1a4> .byte 196,65,20,88,227 // vaddps %ymm11,%ymm13,%ymm12 .byte 196,65,28,89,192 // vmulps %ymm8,%ymm12,%ymm8 - .byte 196,98,125,24,37,181,66,0,0 // vbroadcastss 0x42b5(%rip),%ymm12 # 4c1c <_sk_callback_hsw+0x1a6> + .byte 196,98,125,24,37,229,66,0,0 // vbroadcastss 0x42e5(%rip),%ymm12 # 4c4c <_sk_callback_hsw+0x1a8> .byte 196,66,21,184,196 // vfmadd231ps %ymm12,%ymm13,%ymm8 .byte 196,65,124,82,245 // vrsqrtps %ymm13,%ymm14 .byte 196,65,124,83,246 // vrcpps %ymm14,%ymm14 @@ -10465,7 +10467,7 @@ _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 196,98,125,24,5,120,66,0,0 // vbroadcastss 0x4278(%rip),%ymm8 # 4c14 <_sk_callback_hsw+0x19e> + .byte 196,98,125,24,5,168,66,0,0 // vbroadcastss 0x42a8(%rip),%ymm8 # 4c44 <_sk_callback_hsw+0x1a0> .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13 .byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0 .byte 196,98,125,168,235 // vfmadd213ps %ymm3,%ymm0,%ymm13 @@ -10578,11 +10580,11 @@ _sk_hue_hsw: .byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10 .byte 196,65,44,94,214 // vdivps %ymm14,%ymm10,%ymm10 .byte 196,67,45,74,224,240 // vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - .byte 196,98,125,24,53,124,64,0,0 // vbroadcastss 0x407c(%rip),%ymm14 # 4c20 <_sk_callback_hsw+0x1aa> - .byte 196,98,125,24,61,119,64,0,0 // vbroadcastss 0x4077(%rip),%ymm15 # 4c24 <_sk_callback_hsw+0x1ae> + .byte 196,98,125,24,53,172,64,0,0 // vbroadcastss 0x40ac(%rip),%ymm14 # 4c50 <_sk_callback_hsw+0x1ac> + .byte 196,98,125,24,61,167,64,0,0 // vbroadcastss 0x40a7(%rip),%ymm15 # 4c54 <_sk_callback_hsw+0x1b0> .byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13 .byte 196,66,93,184,238 // vfmadd231ps %ymm14,%ymm4,%ymm13 - .byte 196,226,125,24,5,104,64,0,0 // vbroadcastss 0x4068(%rip),%ymm0 # 4c28 <_sk_callback_hsw+0x1b2> + .byte 196,226,125,24,5,152,64,0,0 // vbroadcastss 0x4098(%rip),%ymm0 # 4c58 <_sk_callback_hsw+0x1b4> .byte 196,98,77,184,232 // vfmadd231ps %ymm0,%ymm6,%ymm13 .byte 196,65,116,89,215 // vmulps %ymm15,%ymm1,%ymm10 .byte 196,66,53,184,214 // vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -10637,7 +10639,7 @@ _sk_hue_hsw: .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 .byte 196,65,36,95,200 // vmaxps %ymm8,%ymm11,%ymm9 .byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8 - .byte 196,226,125,24,13,85,63,0,0 // vbroadcastss 0x3f55(%rip),%ymm1 # 4c2c <_sk_callback_hsw+0x1b6> + .byte 196,226,125,24,13,133,63,0,0 // vbroadcastss 0x3f85(%rip),%ymm1 # 4c5c <_sk_callback_hsw+0x1b8> .byte 197,116,92,215 // vsubps %ymm7,%ymm1,%ymm10 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 197,116,92,219 // vsubps %ymm3,%ymm1,%ymm11 @@ -10691,11 +10693,11 @@ _sk_saturation_hsw: .byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10 .byte 196,65,44,94,214 // vdivps %ymm14,%ymm10,%ymm10 .byte 196,67,45,74,224,240 // vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - .byte 196,98,125,24,53,108,62,0,0 // vbroadcastss 0x3e6c(%rip),%ymm14 # 4c30 <_sk_callback_hsw+0x1ba> - .byte 196,98,125,24,61,103,62,0,0 // vbroadcastss 0x3e67(%rip),%ymm15 # 4c34 <_sk_callback_hsw+0x1be> + .byte 196,98,125,24,53,156,62,0,0 // vbroadcastss 0x3e9c(%rip),%ymm14 # 4c60 <_sk_callback_hsw+0x1bc> + .byte 196,98,125,24,61,151,62,0,0 // vbroadcastss 0x3e97(%rip),%ymm15 # 4c64 <_sk_callback_hsw+0x1c0> .byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13 .byte 196,66,93,184,238 // vfmadd231ps %ymm14,%ymm4,%ymm13 - .byte 196,226,125,24,5,88,62,0,0 // vbroadcastss 0x3e58(%rip),%ymm0 # 4c38 <_sk_callback_hsw+0x1c2> + .byte 196,226,125,24,5,136,62,0,0 // vbroadcastss 0x3e88(%rip),%ymm0 # 4c68 <_sk_callback_hsw+0x1c4> .byte 196,98,77,184,232 // vfmadd231ps %ymm0,%ymm6,%ymm13 .byte 196,65,116,89,215 // vmulps %ymm15,%ymm1,%ymm10 .byte 196,66,53,184,214 // vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -10750,7 +10752,7 @@ _sk_saturation_hsw: .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 .byte 196,65,36,95,200 // vmaxps %ymm8,%ymm11,%ymm9 .byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8 - .byte 196,226,125,24,13,69,61,0,0 // vbroadcastss 0x3d45(%rip),%ymm1 # 4c3c <_sk_callback_hsw+0x1c6> + .byte 196,226,125,24,13,117,61,0,0 // vbroadcastss 0x3d75(%rip),%ymm1 # 4c6c <_sk_callback_hsw+0x1c8> .byte 197,116,92,215 // vsubps %ymm7,%ymm1,%ymm10 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 197,116,92,219 // vsubps %ymm3,%ymm1,%ymm11 @@ -10778,11 +10780,11 @@ _sk_color_hsw: .byte 197,108,89,199 // vmulps %ymm7,%ymm2,%ymm8 .byte 197,116,89,215 // vmulps %ymm7,%ymm1,%ymm10 .byte 197,52,89,223 // vmulps %ymm7,%ymm9,%ymm11 - .byte 196,98,125,24,45,222,60,0,0 // vbroadcastss 0x3cde(%rip),%ymm13 # 4c40 <_sk_callback_hsw+0x1ca> - .byte 196,98,125,24,53,217,60,0,0 // vbroadcastss 0x3cd9(%rip),%ymm14 # 4c44 <_sk_callback_hsw+0x1ce> + .byte 196,98,125,24,45,14,61,0,0 // vbroadcastss 0x3d0e(%rip),%ymm13 # 4c70 <_sk_callback_hsw+0x1cc> + .byte 196,98,125,24,53,9,61,0,0 // vbroadcastss 0x3d09(%rip),%ymm14 # 4c74 <_sk_callback_hsw+0x1d0> .byte 196,65,84,89,230 // vmulps %ymm14,%ymm5,%ymm12 .byte 196,66,93,184,229 // vfmadd231ps %ymm13,%ymm4,%ymm12 - .byte 196,98,125,24,61,202,60,0,0 // vbroadcastss 0x3cca(%rip),%ymm15 # 4c48 <_sk_callback_hsw+0x1d2> + .byte 196,98,125,24,61,250,60,0,0 // vbroadcastss 0x3cfa(%rip),%ymm15 # 4c78 <_sk_callback_hsw+0x1d4> .byte 196,66,77,184,231 // vfmadd231ps %ymm15,%ymm6,%ymm12 .byte 196,65,44,89,206 // vmulps %ymm14,%ymm10,%ymm9 .byte 196,66,61,184,205 // vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -10838,7 +10840,7 @@ _sk_color_hsw: .byte 196,193,116,95,206 // vmaxps %ymm14,%ymm1,%ymm1 .byte 196,65,44,95,198 // vmaxps %ymm14,%ymm10,%ymm8 .byte 196,65,124,95,206 // vmaxps %ymm14,%ymm0,%ymm9 - .byte 196,226,125,24,5,172,59,0,0 // vbroadcastss 0x3bac(%rip),%ymm0 # 4c4c <_sk_callback_hsw+0x1d6> + .byte 196,226,125,24,5,220,59,0,0 // vbroadcastss 0x3bdc(%rip),%ymm0 # 4c7c <_sk_callback_hsw+0x1d8> .byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11 @@ -10866,11 +10868,11 @@ _sk_luminosity_hsw: .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 .byte 197,100,89,213 // vmulps %ymm5,%ymm3,%ymm10 .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11 - .byte 196,98,125,24,45,69,59,0,0 // vbroadcastss 0x3b45(%rip),%ymm13 # 4c50 <_sk_callback_hsw+0x1da> - .byte 196,98,125,24,53,64,59,0,0 // vbroadcastss 0x3b40(%rip),%ymm14 # 4c54 <_sk_callback_hsw+0x1de> + .byte 196,98,125,24,45,117,59,0,0 // vbroadcastss 0x3b75(%rip),%ymm13 # 4c80 <_sk_callback_hsw+0x1dc> + .byte 196,98,125,24,53,112,59,0,0 // vbroadcastss 0x3b70(%rip),%ymm14 # 4c84 <_sk_callback_hsw+0x1e0> .byte 196,65,116,89,230 // vmulps %ymm14,%ymm1,%ymm12 .byte 196,66,109,184,229 // vfmadd231ps %ymm13,%ymm2,%ymm12 - .byte 196,98,125,24,61,49,59,0,0 // vbroadcastss 0x3b31(%rip),%ymm15 # 4c58 <_sk_callback_hsw+0x1e2> + .byte 196,98,125,24,61,97,59,0,0 // vbroadcastss 0x3b61(%rip),%ymm15 # 4c88 <_sk_callback_hsw+0x1e4> .byte 196,66,53,184,231 // vfmadd231ps %ymm15,%ymm9,%ymm12 .byte 196,65,44,89,206 // vmulps %ymm14,%ymm10,%ymm9 .byte 196,66,61,184,205 // vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -10926,7 +10928,7 @@ _sk_luminosity_hsw: .byte 196,193,116,95,206 // vmaxps %ymm14,%ymm1,%ymm1 .byte 196,65,44,95,198 // vmaxps %ymm14,%ymm10,%ymm8 .byte 196,65,124,95,206 // vmaxps %ymm14,%ymm0,%ymm9 - .byte 196,226,125,24,5,19,58,0,0 // vbroadcastss 0x3a13(%rip),%ymm0 # 4c5c <_sk_callback_hsw+0x1e6> + .byte 196,226,125,24,5,67,58,0,0 // vbroadcastss 0x3a43(%rip),%ymm0 # 4c8c <_sk_callback_hsw+0x1e8> .byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11 @@ -10954,17 +10956,17 @@ _sk_srcover_rgba_8888_hsw: .byte 77,133,192 // test %r8,%r8 .byte 15,133,180,0,0,0 // jne 1358 <_sk_srcover_rgba_8888_hsw+0xcd> .byte 196,193,124,16,58 // vmovups (%r10),%ymm7 - .byte 197,196,84,37,15,61,0,0 // vandps 0x3d0f(%rip),%ymm7,%ymm4 # 4fc0 <_sk_callback_hsw+0x54a> + .byte 197,196,84,37,79,61,0,0 // vandps 0x3d4f(%rip),%ymm7,%ymm4 # 5000 <_sk_callback_hsw+0x55c> .byte 197,252,91,228 // vcvtdq2ps %ymm4,%ymm4 - .byte 196,226,69,0,45,34,61,0,0 // vpshufb 0x3d22(%rip),%ymm7,%ymm5 # 4fe0 <_sk_callback_hsw+0x56a> + .byte 196,226,69,0,45,98,61,0,0 // vpshufb 0x3d62(%rip),%ymm7,%ymm5 # 5020 <_sk_callback_hsw+0x57c> .byte 197,252,91,237 // vcvtdq2ps %ymm5,%ymm5 - .byte 196,226,69,0,53,53,61,0,0 // vpshufb 0x3d35(%rip),%ymm7,%ymm6 # 5000 <_sk_callback_hsw+0x58a> + .byte 196,226,69,0,53,117,61,0,0 // vpshufb 0x3d75(%rip),%ymm7,%ymm6 # 5040 <_sk_callback_hsw+0x59c> .byte 197,252,91,246 // vcvtdq2ps %ymm6,%ymm6 .byte 197,197,114,215,24 // vpsrld $0x18,%ymm7,%ymm7 .byte 197,252,91,255 // vcvtdq2ps %ymm7,%ymm7 - .byte 196,98,125,24,5,127,57,0,0 // vbroadcastss 0x397f(%rip),%ymm8 # 4c60 <_sk_callback_hsw+0x1ea> + .byte 196,98,125,24,5,175,57,0,0 // vbroadcastss 0x39af(%rip),%ymm8 # 4c90 <_sk_callback_hsw+0x1ec> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 - .byte 196,98,125,24,13,118,57,0,0 // vbroadcastss 0x3976(%rip),%ymm9 # 4c64 <_sk_callback_hsw+0x1ee> + .byte 196,98,125,24,13,166,57,0,0 // vbroadcastss 0x39a6(%rip),%ymm9 # 4c94 <_sk_callback_hsw+0x1f0> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0 .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 @@ -11024,7 +11026,7 @@ HIDDEN _sk_clamp_1_hsw .globl _sk_clamp_1_hsw FUNCTION(_sk_clamp_1_hsw) _sk_clamp_1_hsw: - .byte 196,98,125,24,5,155,56,0,0 // vbroadcastss 0x389b(%rip),%ymm8 # 4c68 <_sk_callback_hsw+0x1f2> + .byte 196,98,125,24,5,203,56,0,0 // vbroadcastss 0x38cb(%rip),%ymm8 # 4c98 <_sk_callback_hsw+0x1f4> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 @@ -11036,7 +11038,7 @@ HIDDEN _sk_clamp_a_hsw .globl _sk_clamp_a_hsw FUNCTION(_sk_clamp_a_hsw) _sk_clamp_a_hsw: - .byte 196,98,125,24,5,126,56,0,0 // vbroadcastss 0x387e(%rip),%ymm8 # 4c6c <_sk_callback_hsw+0x1f6> + .byte 196,98,125,24,5,174,56,0,0 // vbroadcastss 0x38ae(%rip),%ymm8 # 4c9c <_sk_callback_hsw+0x1f8> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 @@ -11122,7 +11124,7 @@ 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 196,98,125,24,21,198,55,0,0 // vbroadcastss 0x37c6(%rip),%ymm10 # 4c70 <_sk_callback_hsw+0x1fa> + .byte 196,98,125,24,21,246,55,0,0 // vbroadcastss 0x37f6(%rip),%ymm10 # 4ca0 <_sk_callback_hsw+0x1fc> .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 @@ -11135,16 +11137,16 @@ HIDDEN _sk_from_srgb_hsw .globl _sk_from_srgb_hsw FUNCTION(_sk_from_srgb_hsw) _sk_from_srgb_hsw: - .byte 196,98,125,24,5,167,55,0,0 // vbroadcastss 0x37a7(%rip),%ymm8 # 4c74 <_sk_callback_hsw+0x1fe> + .byte 196,98,125,24,5,215,55,0,0 // vbroadcastss 0x37d7(%rip),%ymm8 # 4ca4 <_sk_callback_hsw+0x200> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 - .byte 196,98,125,24,29,153,55,0,0 // vbroadcastss 0x3799(%rip),%ymm11 # 4c78 <_sk_callback_hsw+0x202> - .byte 196,98,125,24,37,148,55,0,0 // vbroadcastss 0x3794(%rip),%ymm12 # 4c7c <_sk_callback_hsw+0x206> + .byte 196,98,125,24,29,201,55,0,0 // vbroadcastss 0x37c9(%rip),%ymm11 # 4ca8 <_sk_callback_hsw+0x204> + .byte 196,98,125,24,37,196,55,0,0 // vbroadcastss 0x37c4(%rip),%ymm12 # 4cac <_sk_callback_hsw+0x208> .byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13 .byte 196,66,125,168,235 // vfmadd213ps %ymm11,%ymm0,%ymm13 - .byte 196,98,125,24,53,133,55,0,0 // vbroadcastss 0x3785(%rip),%ymm14 # 4c80 <_sk_callback_hsw+0x20a> + .byte 196,98,125,24,53,181,55,0,0 // vbroadcastss 0x37b5(%rip),%ymm14 # 4cb0 <_sk_callback_hsw+0x20c> .byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13 - .byte 196,98,125,24,21,123,55,0,0 // vbroadcastss 0x377b(%rip),%ymm10 # 4c84 <_sk_callback_hsw+0x20e> + .byte 196,98,125,24,21,171,55,0,0 // vbroadcastss 0x37ab(%rip),%ymm10 # 4cb4 <_sk_callback_hsw+0x210> .byte 196,193,124,194,194,1 // vcmpltps %ymm10,%ymm0,%ymm0 .byte 196,195,21,74,193,0 // vblendvps %ymm0,%ymm9,%ymm13,%ymm0 .byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9 @@ -11168,19 +11170,19 @@ HIDDEN _sk_to_srgb_hsw FUNCTION(_sk_to_srgb_hsw) _sk_to_srgb_hsw: .byte 197,124,82,200 // vrsqrtps %ymm0,%ymm9 - .byte 196,98,125,24,5,31,55,0,0 // vbroadcastss 0x371f(%rip),%ymm8 # 4c88 <_sk_callback_hsw+0x212> + .byte 196,98,125,24,5,79,55,0,0 // vbroadcastss 0x374f(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x214> .byte 196,65,124,89,208 // vmulps %ymm8,%ymm0,%ymm10 - .byte 196,98,125,24,29,21,55,0,0 // vbroadcastss 0x3715(%rip),%ymm11 # 4c8c <_sk_callback_hsw+0x216> - .byte 196,98,125,24,37,16,55,0,0 // vbroadcastss 0x3710(%rip),%ymm12 # 4c90 <_sk_callback_hsw+0x21a> + .byte 196,98,125,24,29,69,55,0,0 // vbroadcastss 0x3745(%rip),%ymm11 # 4cbc <_sk_callback_hsw+0x218> + .byte 196,98,125,24,37,64,55,0,0 // vbroadcastss 0x3740(%rip),%ymm12 # 4cc0 <_sk_callback_hsw+0x21c> .byte 196,65,124,40,236 // vmovaps %ymm12,%ymm13 .byte 196,66,53,168,235 // vfmadd213ps %ymm11,%ymm9,%ymm13 - .byte 196,98,125,24,53,1,55,0,0 // vbroadcastss 0x3701(%rip),%ymm14 # 4c94 <_sk_callback_hsw+0x21e> + .byte 196,98,125,24,53,49,55,0,0 // vbroadcastss 0x3731(%rip),%ymm14 # 4cc4 <_sk_callback_hsw+0x220> .byte 196,66,53,168,238 // vfmadd213ps %ymm14,%ymm9,%ymm13 - .byte 196,98,125,24,61,247,54,0,0 // vbroadcastss 0x36f7(%rip),%ymm15 # 4c98 <_sk_callback_hsw+0x222> + .byte 196,98,125,24,61,39,55,0,0 // vbroadcastss 0x3727(%rip),%ymm15 # 4cc8 <_sk_callback_hsw+0x224> .byte 196,65,52,88,207 // vaddps %ymm15,%ymm9,%ymm9 .byte 196,65,124,83,201 // vrcpps %ymm9,%ymm9 .byte 196,65,20,89,201 // vmulps %ymm9,%ymm13,%ymm9 - .byte 196,98,125,24,45,227,54,0,0 // vbroadcastss 0x36e3(%rip),%ymm13 # 4c9c <_sk_callback_hsw+0x226> + .byte 196,98,125,24,45,19,55,0,0 // vbroadcastss 0x3713(%rip),%ymm13 # 4ccc <_sk_callback_hsw+0x228> .byte 196,193,124,194,197,1 // vcmpltps %ymm13,%ymm0,%ymm0 .byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0 .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 @@ -11214,26 +11216,26 @@ _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 196,98,125,24,29,88,54,0,0 // vbroadcastss 0x3658(%rip),%ymm11 # 4ca0 <_sk_callback_hsw+0x22a> + .byte 196,98,125,24,29,136,54,0,0 // vbroadcastss 0x3688(%rip),%ymm11 # 4cd0 <_sk_callback_hsw+0x22c> .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11 .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12 .byte 197,116,194,234,1 // vcmpltps %ymm2,%ymm1,%ymm13 - .byte 196,98,125,24,53,69,54,0,0 // vbroadcastss 0x3645(%rip),%ymm14 # 4ca4 <_sk_callback_hsw+0x22e> + .byte 196,98,125,24,53,117,54,0,0 // vbroadcastss 0x3675(%rip),%ymm14 # 4cd4 <_sk_callback_hsw+0x230> .byte 196,65,4,87,255 // vxorps %ymm15,%ymm15,%ymm15 .byte 196,67,5,74,238,208 // vblendvps %ymm13,%ymm14,%ymm15,%ymm13 .byte 196,66,37,168,229 // vfmadd213ps %ymm13,%ymm11,%ymm12 .byte 197,236,92,208 // vsubps %ymm0,%ymm2,%ymm2 .byte 197,124,92,233 // vsubps %ymm1,%ymm0,%ymm13 - .byte 196,98,125,24,53,44,54,0,0 // vbroadcastss 0x362c(%rip),%ymm14 # 4cac <_sk_callback_hsw+0x236> + .byte 196,98,125,24,53,92,54,0,0 // vbroadcastss 0x365c(%rip),%ymm14 # 4cdc <_sk_callback_hsw+0x238> .byte 196,66,37,168,238 // vfmadd213ps %ymm14,%ymm11,%ymm13 - .byte 196,98,125,24,53,26,54,0,0 // vbroadcastss 0x361a(%rip),%ymm14 # 4ca8 <_sk_callback_hsw+0x232> + .byte 196,98,125,24,53,74,54,0,0 // vbroadcastss 0x364a(%rip),%ymm14 # 4cd8 <_sk_callback_hsw+0x234> .byte 196,194,37,168,214 // vfmadd213ps %ymm14,%ymm11,%ymm2 .byte 197,188,194,201,0 // vcmpeqps %ymm1,%ymm8,%ymm1 .byte 196,227,21,74,202,16 // vblendvps %ymm1,%ymm2,%ymm13,%ymm1 .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0 .byte 196,195,117,74,196,0 // vblendvps %ymm0,%ymm12,%ymm1,%ymm0 .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1 - .byte 196,98,125,24,29,253,53,0,0 // vbroadcastss 0x35fd(%rip),%ymm11 # 4cb4 <_sk_callback_hsw+0x23e> + .byte 196,98,125,24,29,45,54,0,0 // vbroadcastss 0x362d(%rip),%ymm11 # 4ce4 <_sk_callback_hsw+0x240> .byte 196,193,116,89,211 // vmulps %ymm11,%ymm1,%ymm2 .byte 197,36,194,218,1 // vcmpltps %ymm2,%ymm11,%ymm11 .byte 196,65,12,92,224 // vsubps %ymm8,%ymm14,%ymm12 @@ -11243,7 +11245,7 @@ _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 196,98,125,24,5,192,53,0,0 // vbroadcastss 0x35c0(%rip),%ymm8 # 4cb0 <_sk_callback_hsw+0x23a> + .byte 196,98,125,24,5,240,53,0,0 // vbroadcastss 0x35f0(%rip),%ymm8 # 4ce0 <_sk_callback_hsw+0x23c> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -11260,30 +11262,30 @@ _sk_hsl_to_rgb_hsw: .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 .byte 197,252,40,224 // vmovaps %ymm0,%ymm4 - .byte 196,98,125,24,5,141,53,0,0 // vbroadcastss 0x358d(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x242> + .byte 196,98,125,24,5,189,53,0,0 // vbroadcastss 0x35bd(%rip),%ymm8 # 4ce8 <_sk_callback_hsw+0x244> .byte 197,60,194,202,2 // vcmpleps %ymm2,%ymm8,%ymm9 .byte 197,84,89,210 // vmulps %ymm2,%ymm5,%ymm10 .byte 196,65,84,92,218 // vsubps %ymm10,%ymm5,%ymm11 .byte 196,67,45,74,203,144 // vblendvps %ymm9,%ymm11,%ymm10,%ymm9 .byte 197,52,88,210 // vaddps %ymm2,%ymm9,%ymm10 - .byte 196,98,125,24,13,112,53,0,0 // vbroadcastss 0x3570(%rip),%ymm9 # 4cbc <_sk_callback_hsw+0x246> + .byte 196,98,125,24,13,160,53,0,0 // vbroadcastss 0x35a0(%rip),%ymm9 # 4cec <_sk_callback_hsw+0x248> .byte 196,66,109,170,202 // vfmsub213ps %ymm10,%ymm2,%ymm9 - .byte 196,98,125,24,29,102,53,0,0 // vbroadcastss 0x3566(%rip),%ymm11 # 4cc0 <_sk_callback_hsw+0x24a> + .byte 196,98,125,24,29,150,53,0,0 // vbroadcastss 0x3596(%rip),%ymm11 # 4cf0 <_sk_callback_hsw+0x24c> .byte 196,65,92,88,219 // vaddps %ymm11,%ymm4,%ymm11 .byte 196,67,125,8,227,1 // vroundps $0x1,%ymm11,%ymm12 .byte 196,65,36,92,252 // vsubps %ymm12,%ymm11,%ymm15 .byte 196,65,44,92,217 // vsubps %ymm9,%ymm10,%ymm11 - .byte 196,98,125,24,45,80,53,0,0 // vbroadcastss 0x3550(%rip),%ymm13 # 4cc8 <_sk_callback_hsw+0x252> + .byte 196,98,125,24,45,128,53,0,0 // vbroadcastss 0x3580(%rip),%ymm13 # 4cf8 <_sk_callback_hsw+0x254> .byte 196,193,4,89,197 // vmulps %ymm13,%ymm15,%ymm0 - .byte 196,98,125,24,53,70,53,0,0 // vbroadcastss 0x3546(%rip),%ymm14 # 4ccc <_sk_callback_hsw+0x256> + .byte 196,98,125,24,53,118,53,0,0 // vbroadcastss 0x3576(%rip),%ymm14 # 4cfc <_sk_callback_hsw+0x258> .byte 197,12,92,224 // vsubps %ymm0,%ymm14,%ymm12 .byte 196,66,37,168,225 // vfmadd213ps %ymm9,%ymm11,%ymm12 - .byte 196,226,125,24,29,44,53,0,0 // vbroadcastss 0x352c(%rip),%ymm3 # 4cc4 <_sk_callback_hsw+0x24e> + .byte 196,226,125,24,29,92,53,0,0 // vbroadcastss 0x355c(%rip),%ymm3 # 4cf4 <_sk_callback_hsw+0x250> .byte 196,193,100,194,255,2 // vcmpleps %ymm15,%ymm3,%ymm7 .byte 196,195,29,74,249,112 // vblendvps %ymm7,%ymm9,%ymm12,%ymm7 .byte 196,65,60,194,231,2 // vcmpleps %ymm15,%ymm8,%ymm12 .byte 196,227,45,74,255,192 // vblendvps %ymm12,%ymm7,%ymm10,%ymm7 - .byte 196,98,125,24,37,23,53,0,0 // vbroadcastss 0x3517(%rip),%ymm12 # 4cd0 <_sk_callback_hsw+0x25a> + .byte 196,98,125,24,37,71,53,0,0 // vbroadcastss 0x3547(%rip),%ymm12 # 4d00 <_sk_callback_hsw+0x25c> .byte 196,65,28,194,255,2 // vcmpleps %ymm15,%ymm12,%ymm15 .byte 196,194,37,168,193 // vfmadd213ps %ymm9,%ymm11,%ymm0 .byte 196,99,125,74,255,240 // vblendvps %ymm15,%ymm7,%ymm0,%ymm15 @@ -11299,7 +11301,7 @@ _sk_hsl_to_rgb_hsw: .byte 197,156,194,192,2 // vcmpleps %ymm0,%ymm12,%ymm0 .byte 196,194,37,168,249 // vfmadd213ps %ymm9,%ymm11,%ymm7 .byte 196,227,69,74,201,0 // vblendvps %ymm0,%ymm1,%ymm7,%ymm1 - .byte 196,226,125,24,5,195,52,0,0 // vbroadcastss 0x34c3(%rip),%ymm0 # 4cd4 <_sk_callback_hsw+0x25e> + .byte 196,226,125,24,5,243,52,0,0 // vbroadcastss 0x34f3(%rip),%ymm0 # 4d04 <_sk_callback_hsw+0x260> .byte 197,220,88,192 // vaddps %ymm0,%ymm4,%ymm0 .byte 196,227,125,8,224,1 // vroundps $0x1,%ymm0,%ymm4 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 @@ -11349,10 +11351,10 @@ _sk_scale_u8_hsw: .byte 77,133,192 // test %r8,%r8 .byte 117,58 // jne 18f2 <_sk_scale_u8_hsw+0x44> .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8 - .byte 197,57,219,5,218,56,0,0 // vpand 0x38da(%rip),%xmm8,%xmm8 # 51a0 <_sk_callback_hsw+0x72a> + .byte 197,57,219,5,26,57,0,0 // vpand 0x391a(%rip),%xmm8,%xmm8 # 51e0 <_sk_callback_hsw+0x73c> .byte 196,66,125,51,192 // vpmovzxwd %xmm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,255,51,0,0 // vbroadcastss 0x33ff(%rip),%ymm9 # 4cd8 <_sk_callback_hsw+0x262> + .byte 196,98,125,24,13,47,52,0,0 // vbroadcastss 0x342f(%rip),%ymm9 # 4d08 <_sk_callback_hsw+0x264> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -11434,10 +11436,10 @@ _sk_lerp_u8_hsw: .byte 77,133,192 // test %r8,%r8 .byte 117,78 // jne 1a17 <_sk_lerp_u8_hsw+0x58> .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8 - .byte 197,57,219,5,217,55,0,0 // vpand 0x37d9(%rip),%xmm8,%xmm8 # 51b0 <_sk_callback_hsw+0x73a> + .byte 197,57,219,5,25,56,0,0 // vpand 0x3819(%rip),%xmm8,%xmm8 # 51f0 <_sk_callback_hsw+0x74c> .byte 196,66,125,51,192 // vpmovzxwd %xmm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,242,50,0,0 // vbroadcastss 0x32f2(%rip),%ymm9 # 4cdc <_sk_callback_hsw+0x266> + .byte 196,98,125,24,13,34,51,0,0 // vbroadcastss 0x3322(%rip),%ymm9 # 4d0c <_sk_callback_hsw+0x268> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -11508,20 +11510,20 @@ _sk_lerp_565_hsw: .byte 15,133,169,0,0,0 // jne 1b6b <_sk_lerp_565_hsw+0xb7> .byte 196,65,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm8 .byte 196,66,125,51,192 // vpmovzxwd %xmm8,%ymm8 - .byte 196,98,125,88,13,10,50,0,0 // vpbroadcastd 0x320a(%rip),%ymm9 # 4ce0 <_sk_callback_hsw+0x26a> + .byte 196,98,125,88,13,58,50,0,0 // vpbroadcastd 0x323a(%rip),%ymm9 # 4d10 <_sk_callback_hsw+0x26c> .byte 196,65,61,219,201 // vpand %ymm9,%ymm8,%ymm9 .byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9 - .byte 196,98,125,24,21,251,49,0,0 // vbroadcastss 0x31fb(%rip),%ymm10 # 4ce4 <_sk_callback_hsw+0x26e> + .byte 196,98,125,24,21,43,50,0,0 // vbroadcastss 0x322b(%rip),%ymm10 # 4d14 <_sk_callback_hsw+0x270> .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,88,21,241,49,0,0 // vpbroadcastd 0x31f1(%rip),%ymm10 # 4ce8 <_sk_callback_hsw+0x272> + .byte 196,98,125,88,21,33,50,0,0 // vpbroadcastd 0x3221(%rip),%ymm10 # 4d18 <_sk_callback_hsw+0x274> .byte 196,65,61,219,210 // vpand %ymm10,%ymm8,%ymm10 .byte 196,65,124,91,210 // vcvtdq2ps %ymm10,%ymm10 - .byte 196,98,125,24,29,226,49,0,0 // vbroadcastss 0x31e2(%rip),%ymm11 # 4cec <_sk_callback_hsw+0x276> + .byte 196,98,125,24,29,18,50,0,0 // vbroadcastss 0x3212(%rip),%ymm11 # 4d1c <_sk_callback_hsw+0x278> .byte 196,65,44,89,211 // vmulps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,88,29,216,49,0,0 // vpbroadcastd 0x31d8(%rip),%ymm11 # 4cf0 <_sk_callback_hsw+0x27a> + .byte 196,98,125,88,29,8,50,0,0 // vpbroadcastd 0x3208(%rip),%ymm11 # 4d20 <_sk_callback_hsw+0x27c> .byte 196,65,61,219,195 // vpand %ymm11,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,29,201,49,0,0 // vbroadcastss 0x31c9(%rip),%ymm11 # 4cf4 <_sk_callback_hsw+0x27e> + .byte 196,98,125,24,29,249,49,0,0 // vbroadcastss 0x31f9(%rip),%ymm11 # 4d24 <_sk_callback_hsw+0x280> .byte 196,65,60,89,195 // vmulps %ymm11,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,226,53,168,196 // vfmadd213ps %ymm4,%ymm9,%ymm0 @@ -11562,7 +11564,7 @@ _sk_lerp_565_hsw: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,225 // jmpq ffffffffe2001be4 <_sk_callback_hsw+0xffffffffe1ffd16e> + .byte 233,255,255,255,225 // jmpq ffffffffe2001be4 <_sk_callback_hsw+0xffffffffe1ffd140> .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) @@ -11591,21 +11593,21 @@ _sk_load_tables_hsw: .byte 77,133,192 // test %r8,%r8 .byte 117,105 // jne 1c76 <_sk_load_tables_hsw+0x7e> .byte 196,193,124,16,26 // vmovups (%r10),%ymm3 - .byte 197,228,84,13,6,52,0,0 // vandps 0x3406(%rip),%ymm3,%ymm1 # 5020 <_sk_callback_hsw+0x5aa> + .byte 197,228,84,13,70,52,0,0 // vandps 0x3446(%rip),%ymm3,%ymm1 # 5060 <_sk_callback_hsw+0x5bc> .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 .byte 72,139,72,8 // mov 0x8(%rax),%rcx .byte 76,139,80,16 // mov 0x10(%rax),%r10 .byte 197,237,118,210 // vpcmpeqd %ymm2,%ymm2,%ymm2 .byte 196,226,109,146,4,137 // vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0 - .byte 196,226,101,0,21,6,52,0,0 // vpshufb 0x3406(%rip),%ymm3,%ymm2 # 5040 <_sk_callback_hsw+0x5ca> + .byte 196,226,101,0,21,70,52,0,0 // vpshufb 0x3446(%rip),%ymm3,%ymm2 # 5080 <_sk_callback_hsw+0x5dc> .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 .byte 196,194,53,146,12,146 // vgatherdps %ymm9,(%r10,%ymm2,4),%ymm1 .byte 72,139,64,24 // mov 0x18(%rax),%rax - .byte 196,98,101,0,13,14,52,0,0 // vpshufb 0x340e(%rip),%ymm3,%ymm9 # 5060 <_sk_callback_hsw+0x5ea> + .byte 196,98,101,0,13,78,52,0,0 // vpshufb 0x344e(%rip),%ymm3,%ymm9 # 50a0 <_sk_callback_hsw+0x5fc> .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 196,98,125,24,5,142,48,0,0 // vbroadcastss 0x308e(%rip),%ymm8 # 4cf8 <_sk_callback_hsw+0x282> + .byte 196,98,125,24,5,190,48,0,0 // vbroadcastss 0x30be(%rip),%ymm8 # 4d28 <_sk_callback_hsw+0x284> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,201 // mov %r9,%rcx @@ -11644,7 +11646,7 @@ _sk_load_tables_u16_be_hsw: .byte 197,185,108,200 // vpunpcklqdq %xmm0,%xmm8,%xmm1 .byte 197,185,109,208 // vpunpckhqdq %xmm0,%xmm8,%xmm2 .byte 197,49,108,195 // vpunpcklqdq %xmm3,%xmm9,%xmm8 - .byte 197,121,111,21,186,52,0,0 // vmovdqa 0x34ba(%rip),%xmm10 # 51c0 <_sk_callback_hsw+0x74a> + .byte 197,121,111,21,250,52,0,0 // vmovdqa 0x34fa(%rip),%xmm10 # 5200 <_sk_callback_hsw+0x75c> .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 @@ -11666,7 +11668,7 @@ _sk_load_tables_u16_be_hsw: .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3 .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,135,47,0,0 // vbroadcastss 0x2f87(%rip),%ymm8 # 4cfc <_sk_callback_hsw+0x286> + .byte 196,98,125,24,5,183,47,0,0 // vbroadcastss 0x2fb7(%rip),%ymm8 # 4d2c <_sk_callback_hsw+0x288> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -11726,7 +11728,7 @@ _sk_load_tables_rgb_u16_be_hsw: .byte 197,185,108,218 // vpunpcklqdq %xmm2,%xmm8,%xmm3 .byte 197,185,109,210 // vpunpckhqdq %xmm2,%xmm8,%xmm2 .byte 197,121,108,193 // vpunpcklqdq %xmm1,%xmm0,%xmm8 - .byte 197,121,111,13,90,51,0,0 // vmovdqa 0x335a(%rip),%xmm9 # 51d0 <_sk_callback_hsw+0x75a> + .byte 197,121,111,13,154,51,0,0 // vmovdqa 0x339a(%rip),%xmm9 # 5210 <_sk_callback_hsw+0x76c> .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 @@ -11743,7 +11745,7 @@ _sk_load_tables_rgb_u16_be_hsw: .byte 196,98,125,51,194 // vpmovzxwd %xmm2,%ymm8 .byte 196,162,101,146,20,128 // vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,53,46,0,0 // vbroadcastss 0x2e35(%rip),%ymm3 # 4d00 <_sk_callback_hsw+0x28a> + .byte 196,226,125,24,29,101,46,0,0 // vbroadcastss 0x2e65(%rip),%ymm3 # 4d30 <_sk_callback_hsw+0x28c> .byte 255,224 // jmpq *%rax .byte 196,129,121,110,4,81 // vmovd (%r9,%r10,2),%xmm0 .byte 196,129,121,196,68,81,4,2 // vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0 @@ -11784,7 +11786,7 @@ HIDDEN _sk_byte_tables_hsw FUNCTION(_sk_byte_tables_hsw) _sk_byte_tables_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,125,45,0,0 // vbroadcastss 0x2d7d(%rip),%ymm8 # 4d04 <_sk_callback_hsw+0x28e> + .byte 196,98,125,24,5,173,45,0,0 // vbroadcastss 0x2dad(%rip),%ymm8 # 4d34 <_sk_callback_hsw+0x290> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 197,125,91,200 // vcvtps2dq %ymm0,%ymm9 .byte 196,65,249,126,201 // vmovq %xmm9,%r9 @@ -11906,7 +11908,7 @@ _sk_byte_tables_hsw: .byte 67,15,182,4,26 // movzbl (%r10,%r11,1),%eax .byte 196,194,125,49,193 // vpmovzxbd %xmm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,34,43,0,0 // vbroadcastss 0x2b22(%rip),%ymm8 # 4d08 <_sk_callback_hsw+0x292> + .byte 196,98,125,24,5,82,43,0,0 // vbroadcastss 0x2b52(%rip),%ymm8 # 4d38 <_sk_callback_hsw+0x294> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 @@ -12022,7 +12024,7 @@ _sk_byte_tables_rgb_hsw: .byte 67,15,182,4,26 // movzbl (%r10,%r11,1),%eax .byte 196,194,125,49,193 // vpmovzxbd %xmm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,14,41,0,0 // vbroadcastss 0x290e(%rip),%ymm8 # 4d0c <_sk_callback_hsw+0x296> + .byte 196,98,125,24,5,62,41,0,0 // vbroadcastss 0x293e(%rip),%ymm8 # 4d3c <_sk_callback_hsw+0x298> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 @@ -12121,33 +12123,33 @@ _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 196,98,125,24,37,232,39,0,0 // vbroadcastss 0x27e8(%rip),%ymm12 # 4d10 <_sk_callback_hsw+0x29a> - .byte 196,98,125,24,45,227,39,0,0 // vbroadcastss 0x27e3(%rip),%ymm13 # 4d14 <_sk_callback_hsw+0x29e> + .byte 196,98,125,24,37,24,40,0,0 // vbroadcastss 0x2818(%rip),%ymm12 # 4d40 <_sk_callback_hsw+0x29c> + .byte 196,98,125,24,45,19,40,0,0 // vbroadcastss 0x2813(%rip),%ymm13 # 4d44 <_sk_callback_hsw+0x2a0> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,217,39,0,0 // vbroadcastss 0x27d9(%rip),%ymm13 # 4d18 <_sk_callback_hsw+0x2a2> + .byte 196,98,125,24,45,9,40,0,0 // vbroadcastss 0x2809(%rip),%ymm13 # 4d48 <_sk_callback_hsw+0x2a4> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,207,39,0,0 // vbroadcastss 0x27cf(%rip),%ymm13 # 4d1c <_sk_callback_hsw+0x2a6> + .byte 196,98,125,24,45,255,39,0,0 // vbroadcastss 0x27ff(%rip),%ymm13 # 4d4c <_sk_callback_hsw+0x2a8> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,197,39,0,0 // vbroadcastss 0x27c5(%rip),%ymm11 # 4d20 <_sk_callback_hsw+0x2aa> + .byte 196,98,125,24,29,245,39,0,0 // vbroadcastss 0x27f5(%rip),%ymm11 # 4d50 <_sk_callback_hsw+0x2ac> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,187,39,0,0 // vbroadcastss 0x27bb(%rip),%ymm12 # 4d24 <_sk_callback_hsw+0x2ae> + .byte 196,98,125,24,37,235,39,0,0 // vbroadcastss 0x27eb(%rip),%ymm12 # 4d54 <_sk_callback_hsw+0x2b0> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,177,39,0,0 // vbroadcastss 0x27b1(%rip),%ymm12 # 4d28 <_sk_callback_hsw+0x2b2> + .byte 196,98,125,24,37,225,39,0,0 // vbroadcastss 0x27e1(%rip),%ymm12 # 4d58 <_sk_callback_hsw+0x2b4> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10 - .byte 196,98,125,24,29,146,39,0,0 // vbroadcastss 0x2792(%rip),%ymm11 # 4d2c <_sk_callback_hsw+0x2b6> + .byte 196,98,125,24,29,194,39,0,0 // vbroadcastss 0x27c2(%rip),%ymm11 # 4d5c <_sk_callback_hsw+0x2b8> .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,136,39,0,0 // vbroadcastss 0x2788(%rip),%ymm11 # 4d30 <_sk_callback_hsw+0x2ba> + .byte 196,98,125,24,29,184,39,0,0 // vbroadcastss 0x27b8(%rip),%ymm11 # 4d60 <_sk_callback_hsw+0x2bc> .byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11 - .byte 196,226,125,24,5,126,39,0,0 // vbroadcastss 0x277e(%rip),%ymm0 # 4d34 <_sk_callback_hsw+0x2be> + .byte 196,226,125,24,5,174,39,0,0 // vbroadcastss 0x27ae(%rip),%ymm0 # 4d64 <_sk_callback_hsw+0x2c0> .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 - .byte 196,98,125,24,21,116,39,0,0 // vbroadcastss 0x2774(%rip),%ymm10 # 4d38 <_sk_callback_hsw+0x2c2> + .byte 196,98,125,24,21,164,39,0,0 // vbroadcastss 0x27a4(%rip),%ymm10 # 4d68 <_sk_callback_hsw+0x2c4> .byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0 .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 - .byte 196,98,125,24,21,103,39,0,0 // vbroadcastss 0x2767(%rip),%ymm10 # 4d3c <_sk_callback_hsw+0x2c6> + .byte 196,98,125,24,21,151,39,0,0 // vbroadcastss 0x2797(%rip),%ymm10 # 4d6c <_sk_callback_hsw+0x2c8> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -12155,7 +12157,7 @@ _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 196,98,125,24,5,62,39,0,0 // vbroadcastss 0x273e(%rip),%ymm8 # 4d40 <_sk_callback_hsw+0x2ca> + .byte 196,98,125,24,5,110,39,0,0 // vbroadcastss 0x276e(%rip),%ymm8 # 4d70 <_sk_callback_hsw+0x2cc> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -12175,33 +12177,33 @@ _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 196,98,125,24,37,246,38,0,0 // vbroadcastss 0x26f6(%rip),%ymm12 # 4d44 <_sk_callback_hsw+0x2ce> - .byte 196,98,125,24,45,241,38,0,0 // vbroadcastss 0x26f1(%rip),%ymm13 # 4d48 <_sk_callback_hsw+0x2d2> + .byte 196,98,125,24,37,38,39,0,0 // vbroadcastss 0x2726(%rip),%ymm12 # 4d74 <_sk_callback_hsw+0x2d0> + .byte 196,98,125,24,45,33,39,0,0 // vbroadcastss 0x2721(%rip),%ymm13 # 4d78 <_sk_callback_hsw+0x2d4> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,231,38,0,0 // vbroadcastss 0x26e7(%rip),%ymm13 # 4d4c <_sk_callback_hsw+0x2d6> + .byte 196,98,125,24,45,23,39,0,0 // vbroadcastss 0x2717(%rip),%ymm13 # 4d7c <_sk_callback_hsw+0x2d8> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,221,38,0,0 // vbroadcastss 0x26dd(%rip),%ymm13 # 4d50 <_sk_callback_hsw+0x2da> + .byte 196,98,125,24,45,13,39,0,0 // vbroadcastss 0x270d(%rip),%ymm13 # 4d80 <_sk_callback_hsw+0x2dc> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,211,38,0,0 // vbroadcastss 0x26d3(%rip),%ymm11 # 4d54 <_sk_callback_hsw+0x2de> + .byte 196,98,125,24,29,3,39,0,0 // vbroadcastss 0x2703(%rip),%ymm11 # 4d84 <_sk_callback_hsw+0x2e0> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,201,38,0,0 // vbroadcastss 0x26c9(%rip),%ymm12 # 4d58 <_sk_callback_hsw+0x2e2> + .byte 196,98,125,24,37,249,38,0,0 // vbroadcastss 0x26f9(%rip),%ymm12 # 4d88 <_sk_callback_hsw+0x2e4> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,191,38,0,0 // vbroadcastss 0x26bf(%rip),%ymm12 # 4d5c <_sk_callback_hsw+0x2e6> + .byte 196,98,125,24,37,239,38,0,0 // vbroadcastss 0x26ef(%rip),%ymm12 # 4d8c <_sk_callback_hsw+0x2e8> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10 - .byte 196,98,125,24,29,160,38,0,0 // vbroadcastss 0x26a0(%rip),%ymm11 # 4d60 <_sk_callback_hsw+0x2ea> + .byte 196,98,125,24,29,208,38,0,0 // vbroadcastss 0x26d0(%rip),%ymm11 # 4d90 <_sk_callback_hsw+0x2ec> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,150,38,0,0 // vbroadcastss 0x2696(%rip),%ymm11 # 4d64 <_sk_callback_hsw+0x2ee> + .byte 196,98,125,24,29,198,38,0,0 // vbroadcastss 0x26c6(%rip),%ymm11 # 4d94 <_sk_callback_hsw+0x2f0> .byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11 - .byte 196,226,125,24,13,140,38,0,0 // vbroadcastss 0x268c(%rip),%ymm1 # 4d68 <_sk_callback_hsw+0x2f2> + .byte 196,226,125,24,13,188,38,0,0 // vbroadcastss 0x26bc(%rip),%ymm1 # 4d98 <_sk_callback_hsw+0x2f4> .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 - .byte 196,98,125,24,21,130,38,0,0 // vbroadcastss 0x2682(%rip),%ymm10 # 4d6c <_sk_callback_hsw+0x2f6> + .byte 196,98,125,24,21,178,38,0,0 // vbroadcastss 0x26b2(%rip),%ymm10 # 4d9c <_sk_callback_hsw+0x2f8> .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 - .byte 196,98,125,24,21,117,38,0,0 // vbroadcastss 0x2675(%rip),%ymm10 # 4d70 <_sk_callback_hsw+0x2fa> + .byte 196,98,125,24,21,165,38,0,0 // vbroadcastss 0x26a5(%rip),%ymm10 # 4da0 <_sk_callback_hsw+0x2fc> .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -12209,7 +12211,7 @@ _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 196,98,125,24,5,76,38,0,0 // vbroadcastss 0x264c(%rip),%ymm8 # 4d74 <_sk_callback_hsw+0x2fe> + .byte 196,98,125,24,5,124,38,0,0 // vbroadcastss 0x267c(%rip),%ymm8 # 4da4 <_sk_callback_hsw+0x300> .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -12229,33 +12231,33 @@ _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 196,98,125,24,37,4,38,0,0 // vbroadcastss 0x2604(%rip),%ymm12 # 4d78 <_sk_callback_hsw+0x302> - .byte 196,98,125,24,45,255,37,0,0 // vbroadcastss 0x25ff(%rip),%ymm13 # 4d7c <_sk_callback_hsw+0x306> + .byte 196,98,125,24,37,52,38,0,0 // vbroadcastss 0x2634(%rip),%ymm12 # 4da8 <_sk_callback_hsw+0x304> + .byte 196,98,125,24,45,47,38,0,0 // vbroadcastss 0x262f(%rip),%ymm13 # 4dac <_sk_callback_hsw+0x308> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,245,37,0,0 // vbroadcastss 0x25f5(%rip),%ymm13 # 4d80 <_sk_callback_hsw+0x30a> + .byte 196,98,125,24,45,37,38,0,0 // vbroadcastss 0x2625(%rip),%ymm13 # 4db0 <_sk_callback_hsw+0x30c> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,235,37,0,0 // vbroadcastss 0x25eb(%rip),%ymm13 # 4d84 <_sk_callback_hsw+0x30e> + .byte 196,98,125,24,45,27,38,0,0 // vbroadcastss 0x261b(%rip),%ymm13 # 4db4 <_sk_callback_hsw+0x310> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,225,37,0,0 // vbroadcastss 0x25e1(%rip),%ymm11 # 4d88 <_sk_callback_hsw+0x312> + .byte 196,98,125,24,29,17,38,0,0 // vbroadcastss 0x2611(%rip),%ymm11 # 4db8 <_sk_callback_hsw+0x314> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,215,37,0,0 // vbroadcastss 0x25d7(%rip),%ymm12 # 4d8c <_sk_callback_hsw+0x316> + .byte 196,98,125,24,37,7,38,0,0 // vbroadcastss 0x2607(%rip),%ymm12 # 4dbc <_sk_callback_hsw+0x318> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,205,37,0,0 // vbroadcastss 0x25cd(%rip),%ymm12 # 4d90 <_sk_callback_hsw+0x31a> + .byte 196,98,125,24,37,253,37,0,0 // vbroadcastss 0x25fd(%rip),%ymm12 # 4dc0 <_sk_callback_hsw+0x31c> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10 - .byte 196,98,125,24,29,174,37,0,0 // vbroadcastss 0x25ae(%rip),%ymm11 # 4d94 <_sk_callback_hsw+0x31e> + .byte 196,98,125,24,29,222,37,0,0 // vbroadcastss 0x25de(%rip),%ymm11 # 4dc4 <_sk_callback_hsw+0x320> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,164,37,0,0 // vbroadcastss 0x25a4(%rip),%ymm11 # 4d98 <_sk_callback_hsw+0x322> + .byte 196,98,125,24,29,212,37,0,0 // vbroadcastss 0x25d4(%rip),%ymm11 # 4dc8 <_sk_callback_hsw+0x324> .byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11 - .byte 196,226,125,24,21,154,37,0,0 // vbroadcastss 0x259a(%rip),%ymm2 # 4d9c <_sk_callback_hsw+0x326> + .byte 196,226,125,24,21,202,37,0,0 // vbroadcastss 0x25ca(%rip),%ymm2 # 4dcc <_sk_callback_hsw+0x328> .byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2 - .byte 196,98,125,24,21,144,37,0,0 // vbroadcastss 0x2590(%rip),%ymm10 # 4da0 <_sk_callback_hsw+0x32a> + .byte 196,98,125,24,21,192,37,0,0 // vbroadcastss 0x25c0(%rip),%ymm10 # 4dd0 <_sk_callback_hsw+0x32c> .byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2 .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 - .byte 196,98,125,24,21,131,37,0,0 // vbroadcastss 0x2583(%rip),%ymm10 # 4da4 <_sk_callback_hsw+0x32e> + .byte 196,98,125,24,21,179,37,0,0 // vbroadcastss 0x25b3(%rip),%ymm10 # 4dd4 <_sk_callback_hsw+0x330> .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -12263,7 +12265,7 @@ _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 196,98,125,24,5,90,37,0,0 // vbroadcastss 0x255a(%rip),%ymm8 # 4da8 <_sk_callback_hsw+0x332> + .byte 196,98,125,24,5,138,37,0,0 // vbroadcastss 0x258a(%rip),%ymm8 # 4dd8 <_sk_callback_hsw+0x334> .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -12283,33 +12285,33 @@ _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 196,98,125,24,37,18,37,0,0 // vbroadcastss 0x2512(%rip),%ymm12 # 4dac <_sk_callback_hsw+0x336> - .byte 196,98,125,24,45,13,37,0,0 // vbroadcastss 0x250d(%rip),%ymm13 # 4db0 <_sk_callback_hsw+0x33a> + .byte 196,98,125,24,37,66,37,0,0 // vbroadcastss 0x2542(%rip),%ymm12 # 4ddc <_sk_callback_hsw+0x338> + .byte 196,98,125,24,45,61,37,0,0 // vbroadcastss 0x253d(%rip),%ymm13 # 4de0 <_sk_callback_hsw+0x33c> .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,3,37,0,0 // vbroadcastss 0x2503(%rip),%ymm13 # 4db4 <_sk_callback_hsw+0x33e> + .byte 196,98,125,24,45,51,37,0,0 // vbroadcastss 0x2533(%rip),%ymm13 # 4de4 <_sk_callback_hsw+0x340> .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10 - .byte 196,98,125,24,45,249,36,0,0 // vbroadcastss 0x24f9(%rip),%ymm13 # 4db8 <_sk_callback_hsw+0x342> + .byte 196,98,125,24,45,41,37,0,0 // vbroadcastss 0x2529(%rip),%ymm13 # 4de8 <_sk_callback_hsw+0x344> .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13 - .byte 196,98,125,24,29,239,36,0,0 // vbroadcastss 0x24ef(%rip),%ymm11 # 4dbc <_sk_callback_hsw+0x346> + .byte 196,98,125,24,29,31,37,0,0 // vbroadcastss 0x251f(%rip),%ymm11 # 4dec <_sk_callback_hsw+0x348> .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11 - .byte 196,98,125,24,37,229,36,0,0 // vbroadcastss 0x24e5(%rip),%ymm12 # 4dc0 <_sk_callback_hsw+0x34a> + .byte 196,98,125,24,37,21,37,0,0 // vbroadcastss 0x2515(%rip),%ymm12 # 4df0 <_sk_callback_hsw+0x34c> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,219,36,0,0 // vbroadcastss 0x24db(%rip),%ymm12 # 4dc4 <_sk_callback_hsw+0x34e> + .byte 196,98,125,24,37,11,37,0,0 // vbroadcastss 0x250b(%rip),%ymm12 # 4df4 <_sk_callback_hsw+0x350> .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10 - .byte 196,98,125,24,29,188,36,0,0 // vbroadcastss 0x24bc(%rip),%ymm11 # 4dc8 <_sk_callback_hsw+0x352> + .byte 196,98,125,24,29,236,36,0,0 // vbroadcastss 0x24ec(%rip),%ymm11 # 4df8 <_sk_callback_hsw+0x354> .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,178,36,0,0 // vbroadcastss 0x24b2(%rip),%ymm11 # 4dcc <_sk_callback_hsw+0x356> + .byte 196,98,125,24,29,226,36,0,0 // vbroadcastss 0x24e2(%rip),%ymm11 # 4dfc <_sk_callback_hsw+0x358> .byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11 - .byte 196,226,125,24,29,168,36,0,0 // vbroadcastss 0x24a8(%rip),%ymm3 # 4dd0 <_sk_callback_hsw+0x35a> + .byte 196,226,125,24,29,216,36,0,0 // vbroadcastss 0x24d8(%rip),%ymm3 # 4e00 <_sk_callback_hsw+0x35c> .byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3 - .byte 196,98,125,24,21,158,36,0,0 // vbroadcastss 0x249e(%rip),%ymm10 # 4dd4 <_sk_callback_hsw+0x35e> + .byte 196,98,125,24,21,206,36,0,0 // vbroadcastss 0x24ce(%rip),%ymm10 # 4e04 <_sk_callback_hsw+0x360> .byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3 .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 - .byte 196,98,125,24,21,145,36,0,0 // vbroadcastss 0x2491(%rip),%ymm10 # 4dd8 <_sk_callback_hsw+0x362> + .byte 196,98,125,24,21,193,36,0,0 // vbroadcastss 0x24c1(%rip),%ymm10 # 4e08 <_sk_callback_hsw+0x364> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -12317,7 +12319,7 @@ _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 196,98,125,24,5,104,36,0,0 // vbroadcastss 0x2468(%rip),%ymm8 # 4ddc <_sk_callback_hsw+0x366> + .byte 196,98,125,24,5,152,36,0,0 // vbroadcastss 0x2498(%rip),%ymm8 # 4e0c <_sk_callback_hsw+0x368> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -12326,26 +12328,26 @@ HIDDEN _sk_lab_to_xyz_hsw .globl _sk_lab_to_xyz_hsw FUNCTION(_sk_lab_to_xyz_hsw) _sk_lab_to_xyz_hsw: - .byte 196,98,125,24,5,90,36,0,0 // vbroadcastss 0x245a(%rip),%ymm8 # 4de0 <_sk_callback_hsw+0x36a> - .byte 196,98,125,24,13,85,36,0,0 // vbroadcastss 0x2455(%rip),%ymm9 # 4de4 <_sk_callback_hsw+0x36e> - .byte 196,98,125,24,21,80,36,0,0 // vbroadcastss 0x2450(%rip),%ymm10 # 4de8 <_sk_callback_hsw+0x372> + .byte 196,98,125,24,5,138,36,0,0 // vbroadcastss 0x248a(%rip),%ymm8 # 4e10 <_sk_callback_hsw+0x36c> + .byte 196,98,125,24,13,133,36,0,0 // vbroadcastss 0x2485(%rip),%ymm9 # 4e14 <_sk_callback_hsw+0x370> + .byte 196,98,125,24,21,128,36,0,0 // vbroadcastss 0x2480(%rip),%ymm10 # 4e18 <_sk_callback_hsw+0x374> .byte 196,194,53,168,202 // vfmadd213ps %ymm10,%ymm9,%ymm1 .byte 196,194,53,168,210 // vfmadd213ps %ymm10,%ymm9,%ymm2 - .byte 196,98,125,24,13,65,36,0,0 // vbroadcastss 0x2441(%rip),%ymm9 # 4dec <_sk_callback_hsw+0x376> + .byte 196,98,125,24,13,113,36,0,0 // vbroadcastss 0x2471(%rip),%ymm9 # 4e1c <_sk_callback_hsw+0x378> .byte 196,66,125,184,200 // vfmadd231ps %ymm8,%ymm0,%ymm9 - .byte 196,226,125,24,5,55,36,0,0 // vbroadcastss 0x2437(%rip),%ymm0 # 4df0 <_sk_callback_hsw+0x37a> + .byte 196,226,125,24,5,103,36,0,0 // vbroadcastss 0x2467(%rip),%ymm0 # 4e20 <_sk_callback_hsw+0x37c> .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 - .byte 196,98,125,24,5,46,36,0,0 // vbroadcastss 0x242e(%rip),%ymm8 # 4df4 <_sk_callback_hsw+0x37e> + .byte 196,98,125,24,5,94,36,0,0 // vbroadcastss 0x245e(%rip),%ymm8 # 4e24 <_sk_callback_hsw+0x380> .byte 196,98,117,168,192 // vfmadd213ps %ymm0,%ymm1,%ymm8 - .byte 196,98,125,24,13,36,36,0,0 // vbroadcastss 0x2424(%rip),%ymm9 # 4df8 <_sk_callback_hsw+0x382> + .byte 196,98,125,24,13,84,36,0,0 // vbroadcastss 0x2454(%rip),%ymm9 # 4e28 <_sk_callback_hsw+0x384> .byte 196,98,109,172,200 // vfnmadd213ps %ymm0,%ymm2,%ymm9 .byte 196,193,60,89,200 // vmulps %ymm8,%ymm8,%ymm1 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 - .byte 196,226,125,24,21,17,36,0,0 // vbroadcastss 0x2411(%rip),%ymm2 # 4dfc <_sk_callback_hsw+0x386> + .byte 196,226,125,24,21,65,36,0,0 // vbroadcastss 0x2441(%rip),%ymm2 # 4e2c <_sk_callback_hsw+0x388> .byte 197,108,194,209,1 // vcmpltps %ymm1,%ymm2,%ymm10 - .byte 196,98,125,24,29,7,36,0,0 // vbroadcastss 0x2407(%rip),%ymm11 # 4e00 <_sk_callback_hsw+0x38a> + .byte 196,98,125,24,29,55,36,0,0 // vbroadcastss 0x2437(%rip),%ymm11 # 4e30 <_sk_callback_hsw+0x38c> .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 - .byte 196,98,125,24,37,253,35,0,0 // vbroadcastss 0x23fd(%rip),%ymm12 # 4e04 <_sk_callback_hsw+0x38e> + .byte 196,98,125,24,37,45,36,0,0 // vbroadcastss 0x242d(%rip),%ymm12 # 4e34 <_sk_callback_hsw+0x390> .byte 196,65,60,89,196 // vmulps %ymm12,%ymm8,%ymm8 .byte 196,99,61,74,193,160 // vblendvps %ymm10,%ymm1,%ymm8,%ymm8 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1 @@ -12360,9 +12362,9 @@ _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 196,226,125,24,5,178,35,0,0 // vbroadcastss 0x23b2(%rip),%ymm0 # 4e08 <_sk_callback_hsw+0x392> + .byte 196,226,125,24,5,226,35,0,0 // vbroadcastss 0x23e2(%rip),%ymm0 # 4e38 <_sk_callback_hsw+0x394> .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,169,35,0,0 // vbroadcastss 0x23a9(%rip),%ymm8 # 4e0c <_sk_callback_hsw+0x396> + .byte 196,98,125,24,5,217,35,0,0 // vbroadcastss 0x23d9(%rip),%ymm8 # 4e3c <_sk_callback_hsw+0x398> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -12376,10 +12378,10 @@ _sk_load_a8_hsw: .byte 77,133,192 // test %r8,%r8 .byte 117,52 // jne 2aaa <_sk_load_a8_hsw+0x3e> .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0 - .byte 197,249,219,5,92,39,0,0 // vpand 0x275c(%rip),%xmm0,%xmm0 # 51e0 <_sk_callback_hsw+0x76a> + .byte 197,249,219,5,156,39,0,0 // vpand 0x279c(%rip),%xmm0,%xmm0 # 5220 <_sk_callback_hsw+0x77c> .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,122,35,0,0 // vbroadcastss 0x237a(%rip),%ymm1 # 4e10 <_sk_callback_hsw+0x39a> + .byte 196,226,125,24,13,170,35,0,0 // vbroadcastss 0x23aa(%rip),%ymm1 # 4e40 <_sk_callback_hsw+0x39c> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -12474,7 +12476,7 @@ _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 196,226,125,24,13,26,34,0,0 // vbroadcastss 0x221a(%rip),%ymm1 # 4e14 <_sk_callback_hsw+0x39e> + .byte 196,226,125,24,13,74,34,0,0 // vbroadcastss 0x224a(%rip),%ymm1 # 4e44 <_sk_callback_hsw+0x3a0> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -12488,7 +12490,7 @@ FUNCTION(_sk_store_a8_hsw) _sk_store_a8_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 - .byte 196,98,125,24,5,252,33,0,0 // vbroadcastss 0x21fc(%rip),%ymm8 # 4e18 <_sk_callback_hsw+0x3a2> + .byte 196,98,125,24,5,44,34,0,0 // vbroadcastss 0x222c(%rip),%ymm8 # 4e48 <_sk_callback_hsw+0x3a4> .byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 @@ -12551,13 +12553,13 @@ _sk_load_g8_hsw: .byte 77,133,192 // test %r8,%r8 .byte 117,57 // jne 2d07 <_sk_load_g8_hsw+0x43> .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0 - .byte 197,249,219,5,20,37,0,0 // vpand 0x2514(%rip),%xmm0,%xmm0 # 51f0 <_sk_callback_hsw+0x77a> + .byte 197,249,219,5,84,37,0,0 // vpand 0x2554(%rip),%xmm0,%xmm0 # 5230 <_sk_callback_hsw+0x78c> .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,46,33,0,0 // vbroadcastss 0x212e(%rip),%ymm1 # 4e1c <_sk_callback_hsw+0x3a6> + .byte 196,226,125,24,13,94,33,0,0 // vbroadcastss 0x215e(%rip),%ymm1 # 4e4c <_sk_callback_hsw+0x3a8> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,35,33,0,0 // vbroadcastss 0x2123(%rip),%ymm3 # 4e20 <_sk_callback_hsw+0x3aa> + .byte 196,226,125,24,29,83,33,0,0 // vbroadcastss 0x2153(%rip),%ymm3 # 4e50 <_sk_callback_hsw+0x3ac> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 255,224 // jmpq *%rax @@ -12649,10 +12651,10 @@ _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 196,226,125,24,13,206,31,0,0 // vbroadcastss 0x1fce(%rip),%ymm1 # 4e24 <_sk_callback_hsw+0x3ae> + .byte 196,226,125,24,13,254,31,0,0 // vbroadcastss 0x1ffe(%rip),%ymm1 # 4e54 <_sk_callback_hsw+0x3b0> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,195,31,0,0 // vbroadcastss 0x1fc3(%rip),%ymm3 # 4e28 <_sk_callback_hsw+0x3b2> + .byte 196,226,125,24,29,243,31,0,0 // vbroadcastss 0x1ff3(%rip),%ymm3 # 4e58 <_sk_callback_hsw+0x3b4> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 255,224 // jmpq *%rax @@ -12700,14 +12702,14 @@ _sk_gather_i8_hsw: .byte 73,139,65,8 // mov 0x8(%r9),%rax .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 .byte 196,226,117,144,28,128 // vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 - .byte 197,229,219,5,91,33,0,0 // vpand 0x215b(%rip),%ymm3,%ymm0 # 5080 <_sk_callback_hsw+0x60a> + .byte 197,229,219,5,155,33,0,0 // vpand 0x219b(%rip),%ymm3,%ymm0 # 50c0 <_sk_callback_hsw+0x61c> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,250,30,0,0 // vbroadcastss 0x1efa(%rip),%ymm8 # 4e2c <_sk_callback_hsw+0x3b6> + .byte 196,98,125,24,5,42,31,0,0 // vbroadcastss 0x1f2a(%rip),%ymm8 # 4e5c <_sk_callback_hsw+0x3b8> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,226,101,0,13,96,33,0,0 // vpshufb 0x2160(%rip),%ymm3,%ymm1 # 50a0 <_sk_callback_hsw+0x62a> + .byte 196,226,101,0,13,160,33,0,0 // vpshufb 0x21a0(%rip),%ymm3,%ymm1 # 50e0 <_sk_callback_hsw+0x63c> .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,226,101,0,21,110,33,0,0 // vpshufb 0x216e(%rip),%ymm3,%ymm2 # 50c0 <_sk_callback_hsw+0x64a> + .byte 196,226,101,0,21,174,33,0,0 // vpshufb 0x21ae(%rip),%ymm3,%ymm2 # 5100 <_sk_callback_hsw+0x65c> .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 @@ -12727,23 +12729,23 @@ _sk_load_565_hsw: .byte 117,114 // jne 2fea <_sk_load_565_hsw+0x7c> .byte 196,193,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm0 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 - .byte 196,226,125,88,5,164,30,0,0 // vpbroadcastd 0x1ea4(%rip),%ymm0 # 4e30 <_sk_callback_hsw+0x3ba> + .byte 196,226,125,88,5,212,30,0,0 // vpbroadcastd 0x1ed4(%rip),%ymm0 # 4e60 <_sk_callback_hsw+0x3bc> .byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,151,30,0,0 // vbroadcastss 0x1e97(%rip),%ymm1 # 4e34 <_sk_callback_hsw+0x3be> + .byte 196,226,125,24,13,199,30,0,0 // vbroadcastss 0x1ec7(%rip),%ymm1 # 4e64 <_sk_callback_hsw+0x3c0> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,142,30,0,0 // vpbroadcastd 0x1e8e(%rip),%ymm1 # 4e38 <_sk_callback_hsw+0x3c2> + .byte 196,226,125,88,13,190,30,0,0 // vpbroadcastd 0x1ebe(%rip),%ymm1 # 4e68 <_sk_callback_hsw+0x3c4> .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,129,30,0,0 // vbroadcastss 0x1e81(%rip),%ymm3 # 4e3c <_sk_callback_hsw+0x3c6> + .byte 196,226,125,24,29,177,30,0,0 // vbroadcastss 0x1eb1(%rip),%ymm3 # 4e6c <_sk_callback_hsw+0x3c8> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,88,29,120,30,0,0 // vpbroadcastd 0x1e78(%rip),%ymm3 # 4e40 <_sk_callback_hsw+0x3ca> + .byte 196,226,125,88,29,168,30,0,0 // vpbroadcastd 0x1ea8(%rip),%ymm3 # 4e70 <_sk_callback_hsw+0x3cc> .byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,107,30,0,0 // vbroadcastss 0x1e6b(%rip),%ymm3 # 4e44 <_sk_callback_hsw+0x3ce> + .byte 196,226,125,24,29,155,30,0,0 // vbroadcastss 0x1e9b(%rip),%ymm3 # 4e74 <_sk_callback_hsw+0x3d0> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,96,30,0,0 // vbroadcastss 0x1e60(%rip),%ymm3 # 4e48 <_sk_callback_hsw+0x3d2> + .byte 196,226,125,24,29,144,30,0,0 // vbroadcastss 0x1e90(%rip),%ymm3 # 4e78 <_sk_callback_hsw+0x3d4> .byte 255,224 // jmpq *%rax .byte 69,137,193 // mov %r8d,%r9d .byte 65,128,225,7 // and $0x7,%r9b @@ -12831,23 +12833,23 @@ _sk_gather_565_hsw: .byte 67,15,183,4,89 // movzwl (%r9,%r11,2),%eax .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 - .byte 196,226,125,88,5,43,29,0,0 // vpbroadcastd 0x1d2b(%rip),%ymm0 # 4e4c <_sk_callback_hsw+0x3d6> + .byte 196,226,125,88,5,91,29,0,0 // vpbroadcastd 0x1d5b(%rip),%ymm0 # 4e7c <_sk_callback_hsw+0x3d8> .byte 197,237,219,192 // vpand %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,30,29,0,0 // vbroadcastss 0x1d1e(%rip),%ymm1 # 4e50 <_sk_callback_hsw+0x3da> + .byte 196,226,125,24,13,78,29,0,0 // vbroadcastss 0x1d4e(%rip),%ymm1 # 4e80 <_sk_callback_hsw+0x3dc> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,21,29,0,0 // vpbroadcastd 0x1d15(%rip),%ymm1 # 4e54 <_sk_callback_hsw+0x3de> + .byte 196,226,125,88,13,69,29,0,0 // vpbroadcastd 0x1d45(%rip),%ymm1 # 4e84 <_sk_callback_hsw+0x3e0> .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,8,29,0,0 // vbroadcastss 0x1d08(%rip),%ymm3 # 4e58 <_sk_callback_hsw+0x3e2> + .byte 196,226,125,24,29,56,29,0,0 // vbroadcastss 0x1d38(%rip),%ymm3 # 4e88 <_sk_callback_hsw+0x3e4> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,88,29,255,28,0,0 // vpbroadcastd 0x1cff(%rip),%ymm3 # 4e5c <_sk_callback_hsw+0x3e6> + .byte 196,226,125,88,29,47,29,0,0 // vpbroadcastd 0x1d2f(%rip),%ymm3 # 4e8c <_sk_callback_hsw+0x3e8> .byte 197,237,219,211 // vpand %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,242,28,0,0 // vbroadcastss 0x1cf2(%rip),%ymm3 # 4e60 <_sk_callback_hsw+0x3ea> + .byte 196,226,125,24,29,34,29,0,0 // vbroadcastss 0x1d22(%rip),%ymm3 # 4e90 <_sk_callback_hsw+0x3ec> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,231,28,0,0 // vbroadcastss 0x1ce7(%rip),%ymm3 # 4e64 <_sk_callback_hsw+0x3ee> + .byte 196,226,125,24,29,23,29,0,0 // vbroadcastss 0x1d17(%rip),%ymm3 # 4e94 <_sk_callback_hsw+0x3f0> .byte 255,224 // jmpq *%rax HIDDEN _sk_store_565_hsw @@ -12856,11 +12858,11 @@ FUNCTION(_sk_store_565_hsw) _sk_store_565_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 - .byte 196,98,125,24,5,219,28,0,0 // vbroadcastss 0x1cdb(%rip),%ymm8 # 4e68 <_sk_callback_hsw+0x3f2> + .byte 196,98,125,24,5,11,29,0,0 // vbroadcastss 0x1d0b(%rip),%ymm8 # 4e98 <_sk_callback_hsw+0x3f4> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,53,114,241,11 // vpslld $0xb,%ymm9,%ymm9 - .byte 196,98,125,24,21,198,28,0,0 // vbroadcastss 0x1cc6(%rip),%ymm10 # 4e6c <_sk_callback_hsw+0x3f6> + .byte 196,98,125,24,21,246,28,0,0 // vbroadcastss 0x1cf6(%rip),%ymm10 # 4e9c <_sk_callback_hsw+0x3f8> .byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 .byte 196,193,45,114,242,5 // vpslld $0x5,%ymm10,%ymm10 @@ -12928,25 +12930,25 @@ _sk_load_4444_hsw: .byte 15,133,138,0,0,0 // jne 32f8 <_sk_load_4444_hsw+0x98> .byte 196,193,122,111,4,83 // vmovdqu (%r11,%rdx,2),%xmm0 .byte 196,226,125,51,216 // vpmovzxwd %xmm0,%ymm3 - .byte 196,226,125,88,5,238,27,0,0 // vpbroadcastd 0x1bee(%rip),%ymm0 # 4e70 <_sk_callback_hsw+0x3fa> + .byte 196,226,125,88,5,30,28,0,0 // vpbroadcastd 0x1c1e(%rip),%ymm0 # 4ea0 <_sk_callback_hsw+0x3fc> .byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,225,27,0,0 // vbroadcastss 0x1be1(%rip),%ymm1 # 4e74 <_sk_callback_hsw+0x3fe> + .byte 196,226,125,24,13,17,28,0,0 // vbroadcastss 0x1c11(%rip),%ymm1 # 4ea4 <_sk_callback_hsw+0x400> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,216,27,0,0 // vpbroadcastd 0x1bd8(%rip),%ymm1 # 4e78 <_sk_callback_hsw+0x402> + .byte 196,226,125,88,13,8,28,0,0 // vpbroadcastd 0x1c08(%rip),%ymm1 # 4ea8 <_sk_callback_hsw+0x404> .byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,203,27,0,0 // vbroadcastss 0x1bcb(%rip),%ymm2 # 4e7c <_sk_callback_hsw+0x406> + .byte 196,226,125,24,21,251,27,0,0 // vbroadcastss 0x1bfb(%rip),%ymm2 # 4eac <_sk_callback_hsw+0x408> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,88,21,194,27,0,0 // vpbroadcastd 0x1bc2(%rip),%ymm2 # 4e80 <_sk_callback_hsw+0x40a> + .byte 196,226,125,88,21,242,27,0,0 // vpbroadcastd 0x1bf2(%rip),%ymm2 # 4eb0 <_sk_callback_hsw+0x40c> .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,181,27,0,0 // vbroadcastss 0x1bb5(%rip),%ymm8 # 4e84 <_sk_callback_hsw+0x40e> + .byte 196,98,125,24,5,229,27,0,0 // vbroadcastss 0x1be5(%rip),%ymm8 # 4eb4 <_sk_callback_hsw+0x410> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,88,5,171,27,0,0 // vpbroadcastd 0x1bab(%rip),%ymm8 # 4e88 <_sk_callback_hsw+0x412> + .byte 196,98,125,88,5,219,27,0,0 // vpbroadcastd 0x1bdb(%rip),%ymm8 # 4eb8 <_sk_callback_hsw+0x414> .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,157,27,0,0 // vbroadcastss 0x1b9d(%rip),%ymm8 # 4e8c <_sk_callback_hsw+0x416> + .byte 196,98,125,24,5,205,27,0,0 // vbroadcastss 0x1bcd(%rip),%ymm8 # 4ebc <_sk_callback_hsw+0x418> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13035,25 +13037,25 @@ _sk_gather_4444_hsw: .byte 67,15,183,4,89 // movzwl (%r9,%r11,2),%eax .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0 .byte 196,226,125,51,216 // vpmovzxwd %xmm0,%ymm3 - .byte 196,226,125,88,5,95,26,0,0 // vpbroadcastd 0x1a5f(%rip),%ymm0 # 4e90 <_sk_callback_hsw+0x41a> + .byte 196,226,125,88,5,143,26,0,0 // vpbroadcastd 0x1a8f(%rip),%ymm0 # 4ec0 <_sk_callback_hsw+0x41c> .byte 197,229,219,192 // vpand %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,82,26,0,0 // vbroadcastss 0x1a52(%rip),%ymm1 # 4e94 <_sk_callback_hsw+0x41e> + .byte 196,226,125,24,13,130,26,0,0 // vbroadcastss 0x1a82(%rip),%ymm1 # 4ec4 <_sk_callback_hsw+0x420> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,88,13,73,26,0,0 // vpbroadcastd 0x1a49(%rip),%ymm1 # 4e98 <_sk_callback_hsw+0x422> + .byte 196,226,125,88,13,121,26,0,0 // vpbroadcastd 0x1a79(%rip),%ymm1 # 4ec8 <_sk_callback_hsw+0x424> .byte 197,229,219,201 // vpand %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,60,26,0,0 // vbroadcastss 0x1a3c(%rip),%ymm2 # 4e9c <_sk_callback_hsw+0x426> + .byte 196,226,125,24,21,108,26,0,0 // vbroadcastss 0x1a6c(%rip),%ymm2 # 4ecc <_sk_callback_hsw+0x428> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,88,21,51,26,0,0 // vpbroadcastd 0x1a33(%rip),%ymm2 # 4ea0 <_sk_callback_hsw+0x42a> + .byte 196,226,125,88,21,99,26,0,0 // vpbroadcastd 0x1a63(%rip),%ymm2 # 4ed0 <_sk_callback_hsw+0x42c> .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,38,26,0,0 // vbroadcastss 0x1a26(%rip),%ymm8 # 4ea4 <_sk_callback_hsw+0x42e> + .byte 196,98,125,24,5,86,26,0,0 // vbroadcastss 0x1a56(%rip),%ymm8 # 4ed4 <_sk_callback_hsw+0x430> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,88,5,28,26,0,0 // vpbroadcastd 0x1a1c(%rip),%ymm8 # 4ea8 <_sk_callback_hsw+0x432> + .byte 196,98,125,88,5,76,26,0,0 // vpbroadcastd 0x1a4c(%rip),%ymm8 # 4ed8 <_sk_callback_hsw+0x434> .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,14,26,0,0 // vbroadcastss 0x1a0e(%rip),%ymm8 # 4eac <_sk_callback_hsw+0x436> + .byte 196,98,125,24,5,62,26,0,0 // vbroadcastss 0x1a3e(%rip),%ymm8 # 4edc <_sk_callback_hsw+0x438> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13064,7 +13066,7 @@ FUNCTION(_sk_store_4444_hsw) _sk_store_4444_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 - .byte 196,98,125,24,5,251,25,0,0 // vbroadcastss 0x19fb(%rip),%ymm8 # 4eb0 <_sk_callback_hsw+0x43a> + .byte 196,98,125,24,5,43,26,0,0 // vbroadcastss 0x1a2b(%rip),%ymm8 # 4ee0 <_sk_callback_hsw+0x43c> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,53,114,241,12 // vpslld $0xc,%ymm9,%ymm9 @@ -13140,14 +13142,14 @@ _sk_load_8888_hsw: .byte 77,133,192 // test %r8,%r8 .byte 117,88 // jne 3601 <_sk_load_8888_hsw+0x6d> .byte 196,193,124,16,26 // vmovups (%r10),%ymm3 - .byte 197,228,84,5,42,27,0,0 // vandps 0x1b2a(%rip),%ymm3,%ymm0 # 50e0 <_sk_callback_hsw+0x66a> + .byte 197,228,84,5,106,27,0,0 // vandps 0x1b6a(%rip),%ymm3,%ymm0 # 5120 <_sk_callback_hsw+0x67c> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,241,24,0,0 // vbroadcastss 0x18f1(%rip),%ymm8 # 4eb4 <_sk_callback_hsw+0x43e> + .byte 196,98,125,24,5,33,25,0,0 // vbroadcastss 0x1921(%rip),%ymm8 # 4ee4 <_sk_callback_hsw+0x440> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,226,101,0,13,47,27,0,0 // vpshufb 0x1b2f(%rip),%ymm3,%ymm1 # 5100 <_sk_callback_hsw+0x68a> + .byte 196,226,101,0,13,111,27,0,0 // vpshufb 0x1b6f(%rip),%ymm3,%ymm1 # 5140 <_sk_callback_hsw+0x69c> .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,226,101,0,21,61,27,0,0 // vpshufb 0x1b3d(%rip),%ymm3,%ymm2 # 5120 <_sk_callback_hsw+0x6aa> + .byte 196,226,101,0,21,125,27,0,0 // vpshufb 0x1b7d(%rip),%ymm3,%ymm2 # 5160 <_sk_callback_hsw+0x6bc> .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 @@ -13179,14 +13181,14 @@ _sk_gather_8888_hsw: .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 .byte 196,194,117,144,28,129 // vpgatherdd %ymm1,(%r9,%ymm0,4),%ymm3 - .byte 197,229,219,5,235,26,0,0 // vpand 0x1aeb(%rip),%ymm3,%ymm0 # 5140 <_sk_callback_hsw+0x6ca> + .byte 197,229,219,5,43,27,0,0 // vpand 0x1b2b(%rip),%ymm3,%ymm0 # 5180 <_sk_callback_hsw+0x6dc> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,86,24,0,0 // vbroadcastss 0x1856(%rip),%ymm8 # 4eb8 <_sk_callback_hsw+0x442> + .byte 196,98,125,24,5,134,24,0,0 // vbroadcastss 0x1886(%rip),%ymm8 # 4ee8 <_sk_callback_hsw+0x444> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,226,101,0,13,240,26,0,0 // vpshufb 0x1af0(%rip),%ymm3,%ymm1 # 5160 <_sk_callback_hsw+0x6ea> + .byte 196,226,101,0,13,48,27,0,0 // vpshufb 0x1b30(%rip),%ymm3,%ymm1 # 51a0 <_sk_callback_hsw+0x6fc> .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,226,101,0,21,254,26,0,0 // vpshufb 0x1afe(%rip),%ymm3,%ymm2 # 5180 <_sk_callback_hsw+0x70a> + .byte 196,226,101,0,21,62,27,0,0 // vpshufb 0x1b3e(%rip),%ymm3,%ymm2 # 51c0 <_sk_callback_hsw+0x71c> .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 @@ -13203,7 +13205,7 @@ _sk_store_8888_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10 .byte 76,3,16 // add (%rax),%r10 - .byte 196,98,125,24,5,6,24,0,0 // vbroadcastss 0x1806(%rip),%ymm8 # 4ebc <_sk_callback_hsw+0x446> + .byte 196,98,125,24,5,54,24,0,0 // vbroadcastss 0x1836(%rip),%ymm8 # 4eec <_sk_callback_hsw+0x448> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,65,116,89,208 // vmulps %ymm8,%ymm1,%ymm10 @@ -13400,7 +13402,7 @@ _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 196,98,125,24,21,253,20,0,0 // vbroadcastss 0x14fd(%rip),%ymm10 # 4ec0 <_sk_callback_hsw+0x44a> + .byte 196,98,125,24,21,45,21,0,0 // vbroadcastss 0x152d(%rip),%ymm10 # 4ef0 <_sk_callback_hsw+0x44c> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -13484,7 +13486,7 @@ _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 196,98,125,24,21,142,19,0,0 // vbroadcastss 0x138e(%rip),%ymm10 # 4ec4 <_sk_callback_hsw+0x44e> + .byte 196,98,125,24,21,190,19,0,0 // vbroadcastss 0x13be(%rip),%ymm10 # 4ef4 <_sk_callback_hsw+0x450> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -13501,7 +13503,7 @@ _sk_load_rgb_u16_be_hsw: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,66,19,0,0 // vbroadcastss 0x1342(%rip),%ymm3 # 4ec8 <_sk_callback_hsw+0x452> + .byte 196,226,125,24,29,114,19,0,0 // vbroadcastss 0x1372(%rip),%ymm3 # 4ef8 <_sk_callback_hsw+0x454> .byte 255,224 // jmpq *%rax .byte 196,193,121,110,4,65 // vmovd (%r9,%rax,2),%xmm0 .byte 196,193,121,196,68,65,4,2 // vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0 @@ -13544,7 +13546,7 @@ _sk_store_u16_be_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,8 // mov (%rax),%r9 .byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax - .byte 196,98,125,24,5,127,18,0,0 // vbroadcastss 0x127f(%rip),%ymm8 # 4ecc <_sk_callback_hsw+0x456> + .byte 196,98,125,24,5,175,18,0,0 // vbroadcastss 0x12af(%rip),%ymm8 # 4efc <_sk_callback_hsw+0x458> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 @@ -13741,13 +13743,14 @@ HIDDEN _sk_repeat_x_hsw FUNCTION(_sk_repeat_x_hsw) _sk_repeat_x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 - .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9 - .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 - .byte 196,98,61,172,200 // vfnmadd213ps %ymm0,%ymm8,%ymm9 + .byte 196,98,125,24,64,4 // vbroadcastss 0x4(%rax),%ymm8 + .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 + .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 + .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 + .byte 196,98,53,172,192 // vfnmadd213ps %ymm0,%ymm9,%ymm8 .byte 197,253,118,192 // vpcmpeqd %ymm0,%ymm0,%ymm0 - .byte 197,189,254,192 // vpaddd %ymm0,%ymm8,%ymm0 - .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 + .byte 197,181,254,192 // vpaddd %ymm0,%ymm9,%ymm0 + .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13756,13 +13759,14 @@ HIDDEN _sk_repeat_y_hsw FUNCTION(_sk_repeat_y_hsw) _sk_repeat_y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 - .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9 - .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 - .byte 196,98,61,172,201 // vfnmadd213ps %ymm1,%ymm8,%ymm9 + .byte 196,98,125,24,64,4 // vbroadcastss 0x4(%rax),%ymm8 + .byte 197,60,89,193 // vmulps %ymm1,%ymm8,%ymm8 + .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 + .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 + .byte 196,98,53,172,193 // vfnmadd213ps %ymm1,%ymm9,%ymm8 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 - .byte 197,189,254,201 // vpaddd %ymm1,%ymm8,%ymm1 - .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 + .byte 197,181,254,201 // vpaddd %ymm1,%ymm9,%ymm1 + .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13772,19 +13776,22 @@ FUNCTION(_sk_mirror_x_hsw) _sk_mirror_x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,122,16,0 // vmovss (%rax),%xmm8 - .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9 - .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10 + .byte 197,122,16,72,4 // vmovss 0x4(%rax),%xmm9 + .byte 196,66,125,24,208 // vbroadcastss %xmm8,%ymm10 + .byte 196,65,124,92,218 // vsubps %ymm10,%ymm0,%ymm11 .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0 + .byte 196,98,125,24,192 // vbroadcastss %xmm0,%ymm8 + .byte 197,178,89,5,61,15,0,0 // vmulss 0xf3d(%rip),%xmm9,%xmm0 # 4f00 <_sk_callback_hsw+0x45c> .byte 196,226,125,24,192 // vbroadcastss %xmm0,%ymm0 - .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8 - .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 - .byte 196,66,125,172,194 // vfnmadd213ps %ymm10,%ymm0,%ymm8 - .byte 196,193,60,92,193 // vsubps %ymm9,%ymm8,%ymm0 + .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0 + .byte 196,227,125,8,192,1 // vroundps $0x1,%ymm0,%ymm0 + .byte 196,194,61,172,195 // vfnmadd213ps %ymm11,%ymm8,%ymm0 + .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8 .byte 197,188,84,192 // vandps %ymm0,%ymm8,%ymm0 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 - .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8 + .byte 196,65,45,254,192 // vpaddd %ymm8,%ymm10,%ymm8 .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13795,19 +13802,22 @@ FUNCTION(_sk_mirror_y_hsw) _sk_mirror_y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,122,16,0 // vmovss (%rax),%xmm8 - .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9 - .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10 + .byte 197,122,16,72,4 // vmovss 0x4(%rax),%xmm9 + .byte 196,66,125,24,208 // vbroadcastss %xmm8,%ymm10 + .byte 196,65,116,92,218 // vsubps %ymm10,%ymm1,%ymm11 .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1 + .byte 196,98,125,24,193 // vbroadcastss %xmm1,%ymm8 + .byte 197,178,89,13,225,14,0,0 // vmulss 0xee1(%rip),%xmm9,%xmm1 # 4f04 <_sk_callback_hsw+0x460> .byte 196,226,125,24,201 // vbroadcastss %xmm1,%ymm1 - .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8 - .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 - .byte 196,66,117,172,194 // vfnmadd213ps %ymm10,%ymm1,%ymm8 - .byte 196,193,60,92,201 // vsubps %ymm9,%ymm8,%ymm1 + .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 + .byte 196,227,125,8,201,1 // vroundps $0x1,%ymm1,%ymm1 + .byte 196,194,61,172,203 // vfnmadd213ps %ymm11,%ymm8,%ymm1 + .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8 .byte 197,188,84,201 // vandps %ymm1,%ymm8,%ymm1 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 - .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8 + .byte 196,65,45,254,192 // vpaddd %ymm8,%ymm10,%ymm8 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13818,7 +13828,7 @@ FUNCTION(_sk_clamp_x_1_hsw) _sk_clamp_x_1_hsw: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,144,14,0,0 // vbroadcastss 0xe90(%rip),%ymm8 # 4ed0 <_sk_callback_hsw+0x45a> + .byte 196,98,125,24,5,154,14,0,0 // vbroadcastss 0xe9a(%rip),%ymm8 # 4f08 <_sk_callback_hsw+0x464> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -13836,9 +13846,9 @@ HIDDEN _sk_mirror_x_1_hsw .globl _sk_mirror_x_1_hsw FUNCTION(_sk_mirror_x_1_hsw) _sk_mirror_x_1_hsw: - .byte 196,98,125,24,5,115,14,0,0 // vbroadcastss 0xe73(%rip),%ymm8 # 4ed4 <_sk_callback_hsw+0x45e> + .byte 196,98,125,24,5,125,14,0,0 // vbroadcastss 0xe7d(%rip),%ymm8 # 4f0c <_sk_callback_hsw+0x468> .byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,13,105,14,0,0 // vbroadcastss 0xe69(%rip),%ymm9 # 4ed8 <_sk_callback_hsw+0x462> + .byte 196,98,125,24,13,115,14,0,0 // vbroadcastss 0xe73(%rip),%ymm9 # 4f10 <_sk_callback_hsw+0x46c> .byte 196,65,124,89,201 // vmulps %ymm9,%ymm0,%ymm9 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 .byte 196,65,52,88,201 // vaddps %ymm9,%ymm9,%ymm9 @@ -13854,11 +13864,11 @@ HIDDEN _sk_luminance_to_alpha_hsw .globl _sk_luminance_to_alpha_hsw FUNCTION(_sk_luminance_to_alpha_hsw) _sk_luminance_to_alpha_hsw: - .byte 196,226,125,24,29,57,14,0,0 // vbroadcastss 0xe39(%rip),%ymm3 # 4edc <_sk_callback_hsw+0x466> - .byte 196,98,125,24,5,52,14,0,0 // vbroadcastss 0xe34(%rip),%ymm8 # 4ee0 <_sk_callback_hsw+0x46a> + .byte 196,226,125,24,29,67,14,0,0 // vbroadcastss 0xe43(%rip),%ymm3 # 4f14 <_sk_callback_hsw+0x470> + .byte 196,98,125,24,5,62,14,0,0 // vbroadcastss 0xe3e(%rip),%ymm8 # 4f18 <_sk_callback_hsw+0x474> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 196,226,125,184,203 // vfmadd231ps %ymm3,%ymm0,%ymm1 - .byte 196,226,125,24,29,37,14,0,0 // vbroadcastss 0xe25(%rip),%ymm3 # 4ee4 <_sk_callback_hsw+0x46e> + .byte 196,226,125,24,29,47,14,0,0 // vbroadcastss 0xe2f(%rip),%ymm3 # 4f1c <_sk_callback_hsw+0x478> .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 @@ -14032,9 +14042,9 @@ _sk_evenly_spaced_gradient_hsw: .byte 76,139,72,8 // mov 0x8(%rax),%r9 .byte 77,137,211 // mov %r10,%r11 .byte 73,255,203 // dec %r11 - .byte 120,7 // js 4378 <_sk_evenly_spaced_gradient_hsw+0x19> + .byte 120,7 // js 43a6 <_sk_evenly_spaced_gradient_hsw+0x19> .byte 196,193,242,42,203 // vcvtsi2ss %r11,%xmm1,%xmm1 - .byte 235,22 // jmp 438e <_sk_evenly_spaced_gradient_hsw+0x2f> + .byte 235,22 // jmp 43bc <_sk_evenly_spaced_gradient_hsw+0x2f> .byte 76,137,219 // mov %r11,%rbx .byte 72,209,235 // shr %rbx .byte 65,131,227,1 // and $0x1,%r11d @@ -14045,7 +14055,7 @@ _sk_evenly_spaced_gradient_hsw: .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1 .byte 197,126,91,217 // vcvttps2dq %ymm1,%ymm11 .byte 73,131,250,8 // cmp $0x8,%r10 - .byte 119,70 // ja 43e7 <_sk_evenly_spaced_gradient_hsw+0x88> + .byte 119,70 // ja 4415 <_sk_evenly_spaced_gradient_hsw+0x88> .byte 196,66,37,22,1 // vpermps (%r9),%ymm11,%ymm8 .byte 72,139,88,40 // mov 0x28(%rax),%rbx .byte 196,98,37,22,11 // vpermps (%rbx),%ymm11,%ymm9 @@ -14061,7 +14071,7 @@ _sk_evenly_spaced_gradient_hsw: .byte 196,226,37,22,27 // vpermps (%rbx),%ymm11,%ymm3 .byte 72,139,64,64 // mov 0x40(%rax),%rax .byte 196,98,37,22,40 // vpermps (%rax),%ymm11,%ymm13 - .byte 235,110 // jmp 4455 <_sk_evenly_spaced_gradient_hsw+0xf6> + .byte 235,110 // jmp 4483 <_sk_evenly_spaced_gradient_hsw+0xf6> .byte 196,65,13,118,246 // vpcmpeqd %ymm14,%ymm14,%ymm14 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 .byte 196,2,117,146,4,153 // vgatherdps %ymm1,(%r9,%ymm11,4),%ymm8 @@ -14098,14 +14108,14 @@ HIDDEN _sk_gauss_a_to_rgba_hsw .globl _sk_gauss_a_to_rgba_hsw FUNCTION(_sk_gauss_a_to_rgba_hsw) _sk_gauss_a_to_rgba_hsw: - .byte 196,226,125,24,5,109,10,0,0 // vbroadcastss 0xa6d(%rip),%ymm0 # 4ee8 <_sk_callback_hsw+0x472> - .byte 196,226,125,24,13,104,10,0,0 // vbroadcastss 0xa68(%rip),%ymm1 # 4eec <_sk_callback_hsw+0x476> + .byte 196,226,125,24,5,119,10,0,0 // vbroadcastss 0xa77(%rip),%ymm0 # 4f20 <_sk_callback_hsw+0x47c> + .byte 196,226,125,24,13,114,10,0,0 // vbroadcastss 0xa72(%rip),%ymm1 # 4f24 <_sk_callback_hsw+0x480> .byte 196,226,101,168,200 // vfmadd213ps %ymm0,%ymm3,%ymm1 - .byte 196,226,125,24,5,94,10,0,0 // vbroadcastss 0xa5e(%rip),%ymm0 # 4ef0 <_sk_callback_hsw+0x47a> + .byte 196,226,125,24,5,104,10,0,0 // vbroadcastss 0xa68(%rip),%ymm0 # 4f28 <_sk_callback_hsw+0x484> .byte 196,226,101,184,193 // vfmadd231ps %ymm1,%ymm3,%ymm0 - .byte 196,226,125,24,13,84,10,0,0 // vbroadcastss 0xa54(%rip),%ymm1 # 4ef4 <_sk_callback_hsw+0x47e> + .byte 196,226,125,24,13,94,10,0,0 // vbroadcastss 0xa5e(%rip),%ymm1 # 4f2c <_sk_callback_hsw+0x488> .byte 196,226,101,184,200 // vfmadd231ps %ymm0,%ymm3,%ymm1 - .byte 196,226,125,24,5,74,10,0,0 // vbroadcastss 0xa4a(%rip),%ymm0 # 4ef8 <_sk_callback_hsw+0x482> + .byte 196,226,125,24,5,84,10,0,0 // vbroadcastss 0xa54(%rip),%ymm0 # 4f30 <_sk_callback_hsw+0x48c> .byte 196,226,101,184,193 // vfmadd231ps %ymm1,%ymm3,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 @@ -14120,11 +14130,11 @@ _sk_gradient_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,8 // mov (%rax),%r9 .byte 73,131,249,1 // cmp $0x1,%r9 - .byte 15,134,180,0,0,0 // jbe 4586 <_sk_gradient_hsw+0xc3> + .byte 15,134,180,0,0,0 // jbe 45b4 <_sk_gradient_hsw+0xc3> .byte 76,139,80,72 // mov 0x48(%rax),%r10 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 65,187,1,0,0,0 // mov $0x1,%r11d - .byte 196,226,125,24,21,19,10,0,0 // vbroadcastss 0xa13(%rip),%ymm2 # 4efc <_sk_callback_hsw+0x486> + .byte 196,226,125,24,21,29,10,0,0 // vbroadcastss 0xa1d(%rip),%ymm2 # 4f34 <_sk_callback_hsw+0x490> .byte 196,65,53,239,201 // vpxor %ymm9,%ymm9,%ymm9 .byte 196,130,125,24,28,154 // vbroadcastss (%r10,%r11,4),%ymm3 .byte 197,228,194,216,2 // vcmpleps %ymm0,%ymm3,%ymm3 @@ -14132,10 +14142,10 @@ _sk_gradient_hsw: .byte 196,65,101,254,201 // vpaddd %ymm9,%ymm3,%ymm9 .byte 73,255,195 // inc %r11 .byte 77,57,217 // cmp %r11,%r9 - .byte 117,226 // jne 44ee <_sk_gradient_hsw+0x2b> + .byte 117,226 // jne 451c <_sk_gradient_hsw+0x2b> .byte 76,139,80,8 // mov 0x8(%rax),%r10 .byte 73,131,249,8 // cmp $0x8,%r9 - .byte 118,121 // jbe 458f <_sk_gradient_hsw+0xcc> + .byte 118,121 // jbe 45bd <_sk_gradient_hsw+0xcc> .byte 196,65,13,118,246 // vpcmpeqd %ymm14,%ymm14,%ymm14 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 .byte 196,2,117,146,4,138 // vgatherdps %ymm1,(%r10,%ymm9,4),%ymm8 @@ -14159,7 +14169,7 @@ _sk_gradient_hsw: .byte 196,130,21,146,28,137 // vgatherdps %ymm13,(%r9,%ymm9,4),%ymm3 .byte 72,139,64,64 // mov 0x40(%rax),%rax .byte 196,34,13,146,44,136 // vgatherdps %ymm14,(%rax,%ymm9,4),%ymm13 - .byte 235,77 // jmp 45d3 <_sk_gradient_hsw+0x110> + .byte 235,77 // jmp 4601 <_sk_gradient_hsw+0x110> .byte 76,139,80,8 // mov 0x8(%rax),%r10 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 .byte 196,66,53,22,2 // vpermps (%r10),%ymm9,%ymm8 @@ -14219,24 +14229,24 @@ _sk_xy_to_unit_angle_hsw: .byte 196,65,52,95,226 // vmaxps %ymm10,%ymm9,%ymm12 .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 .byte 196,65,36,89,227 // vmulps %ymm11,%ymm11,%ymm12 - .byte 196,98,125,24,45,146,8,0,0 // vbroadcastss 0x892(%rip),%ymm13 # 4f00 <_sk_callback_hsw+0x48a> - .byte 196,98,125,24,53,141,8,0,0 // vbroadcastss 0x88d(%rip),%ymm14 # 4f04 <_sk_callback_hsw+0x48e> + .byte 196,98,125,24,45,156,8,0,0 // vbroadcastss 0x89c(%rip),%ymm13 # 4f38 <_sk_callback_hsw+0x494> + .byte 196,98,125,24,53,151,8,0,0 // vbroadcastss 0x897(%rip),%ymm14 # 4f3c <_sk_callback_hsw+0x498> .byte 196,66,29,184,245 // vfmadd231ps %ymm13,%ymm12,%ymm14 - .byte 196,98,125,24,45,131,8,0,0 // vbroadcastss 0x883(%rip),%ymm13 # 4f08 <_sk_callback_hsw+0x492> + .byte 196,98,125,24,45,141,8,0,0 // vbroadcastss 0x88d(%rip),%ymm13 # 4f40 <_sk_callback_hsw+0x49c> .byte 196,66,29,184,238 // vfmadd231ps %ymm14,%ymm12,%ymm13 - .byte 196,98,125,24,53,121,8,0,0 // vbroadcastss 0x879(%rip),%ymm14 # 4f0c <_sk_callback_hsw+0x496> + .byte 196,98,125,24,53,131,8,0,0 // vbroadcastss 0x883(%rip),%ymm14 # 4f44 <_sk_callback_hsw+0x4a0> .byte 196,66,29,184,245 // vfmadd231ps %ymm13,%ymm12,%ymm14 .byte 196,65,36,89,222 // vmulps %ymm14,%ymm11,%ymm11 .byte 196,65,52,194,202,1 // vcmpltps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,21,100,8,0,0 // vbroadcastss 0x864(%rip),%ymm10 # 4f10 <_sk_callback_hsw+0x49a> + .byte 196,98,125,24,21,110,8,0,0 // vbroadcastss 0x86e(%rip),%ymm10 # 4f48 <_sk_callback_hsw+0x4a4> .byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10 .byte 196,67,37,74,202,144 // vblendvps %ymm9,%ymm10,%ymm11,%ymm9 .byte 196,193,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,21,78,8,0,0 // vbroadcastss 0x84e(%rip),%ymm10 # 4f14 <_sk_callback_hsw+0x49e> + .byte 196,98,125,24,21,88,8,0,0 // vbroadcastss 0x858(%rip),%ymm10 # 4f4c <_sk_callback_hsw+0x4a8> .byte 196,65,44,92,209 // vsubps %ymm9,%ymm10,%ymm10 .byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0 .byte 196,65,116,194,200,1 // vcmpltps %ymm8,%ymm1,%ymm9 - .byte 196,98,125,24,21,56,8,0,0 // vbroadcastss 0x838(%rip),%ymm10 # 4f18 <_sk_callback_hsw+0x4a2> + .byte 196,98,125,24,21,66,8,0,0 // vbroadcastss 0x842(%rip),%ymm10 # 4f50 <_sk_callback_hsw+0x4ac> .byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10 .byte 196,195,125,74,194,144 // vblendvps %ymm9,%ymm10,%ymm0,%ymm0 .byte 196,65,124,194,200,3 // vcmpunordps %ymm8,%ymm0,%ymm9 @@ -14259,7 +14269,7 @@ HIDDEN _sk_save_xy_hsw FUNCTION(_sk_save_xy_hsw) _sk_save_xy_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,5,8,0,0 // vbroadcastss 0x805(%rip),%ymm8 # 4f1c <_sk_callback_hsw+0x4a6> + .byte 196,98,125,24,5,15,8,0,0 // vbroadcastss 0x80f(%rip),%ymm8 # 4f54 <_sk_callback_hsw+0x4b0> .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 @@ -14293,9 +14303,9 @@ HIDDEN _sk_bilinear_nx_hsw FUNCTION(_sk_bilinear_nx_hsw) _sk_bilinear_nx_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,153,7,0,0 // vbroadcastss 0x799(%rip),%ymm0 # 4f20 <_sk_callback_hsw+0x4aa> + .byte 196,226,125,24,5,163,7,0,0 // vbroadcastss 0x7a3(%rip),%ymm0 # 4f58 <_sk_callback_hsw+0x4b4> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,144,7,0,0 // vbroadcastss 0x790(%rip),%ymm8 # 4f24 <_sk_callback_hsw+0x4ae> + .byte 196,98,125,24,5,154,7,0,0 // vbroadcastss 0x79a(%rip),%ymm8 # 4f5c <_sk_callback_hsw+0x4b8> .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 @@ -14306,7 +14316,7 @@ HIDDEN _sk_bilinear_px_hsw FUNCTION(_sk_bilinear_px_hsw) _sk_bilinear_px_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,120,7,0,0 // vbroadcastss 0x778(%rip),%ymm0 # 4f28 <_sk_callback_hsw+0x4b2> + .byte 196,226,125,24,5,130,7,0,0 // vbroadcastss 0x782(%rip),%ymm0 # 4f60 <_sk_callback_hsw+0x4bc> .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) @@ -14318,9 +14328,9 @@ HIDDEN _sk_bilinear_ny_hsw FUNCTION(_sk_bilinear_ny_hsw) _sk_bilinear_ny_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,92,7,0,0 // vbroadcastss 0x75c(%rip),%ymm1 # 4f2c <_sk_callback_hsw+0x4b6> + .byte 196,226,125,24,13,102,7,0,0 // vbroadcastss 0x766(%rip),%ymm1 # 4f64 <_sk_callback_hsw+0x4c0> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,82,7,0,0 // vbroadcastss 0x752(%rip),%ymm8 # 4f30 <_sk_callback_hsw+0x4ba> + .byte 196,98,125,24,5,92,7,0,0 // vbroadcastss 0x75c(%rip),%ymm8 # 4f68 <_sk_callback_hsw+0x4c4> .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 @@ -14331,7 +14341,7 @@ HIDDEN _sk_bilinear_py_hsw FUNCTION(_sk_bilinear_py_hsw) _sk_bilinear_py_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,58,7,0,0 // vbroadcastss 0x73a(%rip),%ymm1 # 4f34 <_sk_callback_hsw+0x4be> + .byte 196,226,125,24,13,68,7,0,0 // vbroadcastss 0x744(%rip),%ymm1 # 4f6c <_sk_callback_hsw+0x4c8> .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) @@ -14343,13 +14353,13 @@ HIDDEN _sk_bicubic_n3x_hsw FUNCTION(_sk_bicubic_n3x_hsw) _sk_bicubic_n3x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,29,7,0,0 // vbroadcastss 0x71d(%rip),%ymm0 # 4f38 <_sk_callback_hsw+0x4c2> + .byte 196,226,125,24,5,39,7,0,0 // vbroadcastss 0x727(%rip),%ymm0 # 4f70 <_sk_callback_hsw+0x4cc> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,20,7,0,0 // vbroadcastss 0x714(%rip),%ymm8 # 4f3c <_sk_callback_hsw+0x4c6> + .byte 196,98,125,24,5,30,7,0,0 // vbroadcastss 0x71e(%rip),%ymm8 # 4f74 <_sk_callback_hsw+0x4d0> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,5,7,0,0 // vbroadcastss 0x705(%rip),%ymm10 # 4f40 <_sk_callback_hsw+0x4ca> - .byte 196,98,125,24,29,0,7,0,0 // vbroadcastss 0x700(%rip),%ymm11 # 4f44 <_sk_callback_hsw+0x4ce> + .byte 196,98,125,24,21,15,7,0,0 // vbroadcastss 0x70f(%rip),%ymm10 # 4f78 <_sk_callback_hsw+0x4d4> + .byte 196,98,125,24,29,10,7,0,0 // vbroadcastss 0x70a(%rip),%ymm11 # 4f7c <_sk_callback_hsw+0x4d8> .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) @@ -14361,16 +14371,16 @@ HIDDEN _sk_bicubic_n1x_hsw FUNCTION(_sk_bicubic_n1x_hsw) _sk_bicubic_n1x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,227,6,0,0 // vbroadcastss 0x6e3(%rip),%ymm0 # 4f48 <_sk_callback_hsw+0x4d2> + .byte 196,226,125,24,5,237,6,0,0 // vbroadcastss 0x6ed(%rip),%ymm0 # 4f80 <_sk_callback_hsw+0x4dc> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,218,6,0,0 // vbroadcastss 0x6da(%rip),%ymm8 # 4f4c <_sk_callback_hsw+0x4d6> + .byte 196,98,125,24,5,228,6,0,0 // vbroadcastss 0x6e4(%rip),%ymm8 # 4f84 <_sk_callback_hsw+0x4e0> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,208,6,0,0 // vbroadcastss 0x6d0(%rip),%ymm9 # 4f50 <_sk_callback_hsw+0x4da> - .byte 196,98,125,24,21,203,6,0,0 // vbroadcastss 0x6cb(%rip),%ymm10 # 4f54 <_sk_callback_hsw+0x4de> + .byte 196,98,125,24,13,218,6,0,0 // vbroadcastss 0x6da(%rip),%ymm9 # 4f88 <_sk_callback_hsw+0x4e4> + .byte 196,98,125,24,21,213,6,0,0 // vbroadcastss 0x6d5(%rip),%ymm10 # 4f8c <_sk_callback_hsw+0x4e8> .byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,13,193,6,0,0 // vbroadcastss 0x6c1(%rip),%ymm9 # 4f58 <_sk_callback_hsw+0x4e2> + .byte 196,98,125,24,13,203,6,0,0 // vbroadcastss 0x6cb(%rip),%ymm9 # 4f90 <_sk_callback_hsw+0x4ec> .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9 - .byte 196,98,125,24,21,183,6,0,0 // vbroadcastss 0x6b7(%rip),%ymm10 # 4f5c <_sk_callback_hsw+0x4e6> + .byte 196,98,125,24,21,193,6,0,0 // vbroadcastss 0x6c1(%rip),%ymm10 # 4f94 <_sk_callback_hsw+0x4f0> .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 @@ -14381,14 +14391,14 @@ HIDDEN _sk_bicubic_p1x_hsw FUNCTION(_sk_bicubic_p1x_hsw) _sk_bicubic_p1x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,159,6,0,0 // vbroadcastss 0x69f(%rip),%ymm8 # 4f60 <_sk_callback_hsw+0x4ea> + .byte 196,98,125,24,5,169,6,0,0 // vbroadcastss 0x6a9(%rip),%ymm8 # 4f98 <_sk_callback_hsw+0x4f4> .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 - .byte 196,98,125,24,21,145,6,0,0 // vbroadcastss 0x691(%rip),%ymm10 # 4f64 <_sk_callback_hsw+0x4ee> - .byte 196,98,125,24,29,140,6,0,0 // vbroadcastss 0x68c(%rip),%ymm11 # 4f68 <_sk_callback_hsw+0x4f2> + .byte 196,98,125,24,21,155,6,0,0 // vbroadcastss 0x69b(%rip),%ymm10 # 4f9c <_sk_callback_hsw+0x4f8> + .byte 196,98,125,24,29,150,6,0,0 // vbroadcastss 0x696(%rip),%ymm11 # 4fa0 <_sk_callback_hsw+0x4fc> .byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11 .byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11 - .byte 196,98,125,24,5,125,6,0,0 // vbroadcastss 0x67d(%rip),%ymm8 # 4f6c <_sk_callback_hsw+0x4f6> + .byte 196,98,125,24,5,135,6,0,0 // vbroadcastss 0x687(%rip),%ymm8 # 4fa4 <_sk_callback_hsw+0x500> .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 @@ -14399,12 +14409,12 @@ HIDDEN _sk_bicubic_p3x_hsw FUNCTION(_sk_bicubic_p3x_hsw) _sk_bicubic_p3x_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,101,6,0,0 // vbroadcastss 0x665(%rip),%ymm0 # 4f70 <_sk_callback_hsw+0x4fa> + .byte 196,226,125,24,5,111,6,0,0 // vbroadcastss 0x66f(%rip),%ymm0 # 4fa8 <_sk_callback_hsw+0x504> .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 196,98,125,24,21,82,6,0,0 // vbroadcastss 0x652(%rip),%ymm10 # 4f74 <_sk_callback_hsw+0x4fe> - .byte 196,98,125,24,29,77,6,0,0 // vbroadcastss 0x64d(%rip),%ymm11 # 4f78 <_sk_callback_hsw+0x502> + .byte 196,98,125,24,21,92,6,0,0 // vbroadcastss 0x65c(%rip),%ymm10 # 4fac <_sk_callback_hsw+0x508> + .byte 196,98,125,24,29,87,6,0,0 // vbroadcastss 0x657(%rip),%ymm11 # 4fb0 <_sk_callback_hsw+0x50c> .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) @@ -14416,13 +14426,13 @@ HIDDEN _sk_bicubic_n3y_hsw FUNCTION(_sk_bicubic_n3y_hsw) _sk_bicubic_n3y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,48,6,0,0 // vbroadcastss 0x630(%rip),%ymm1 # 4f7c <_sk_callback_hsw+0x506> + .byte 196,226,125,24,13,58,6,0,0 // vbroadcastss 0x63a(%rip),%ymm1 # 4fb4 <_sk_callback_hsw+0x510> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,38,6,0,0 // vbroadcastss 0x626(%rip),%ymm8 # 4f80 <_sk_callback_hsw+0x50a> + .byte 196,98,125,24,5,48,6,0,0 // vbroadcastss 0x630(%rip),%ymm8 # 4fb8 <_sk_callback_hsw+0x514> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,23,6,0,0 // vbroadcastss 0x617(%rip),%ymm10 # 4f84 <_sk_callback_hsw+0x50e> - .byte 196,98,125,24,29,18,6,0,0 // vbroadcastss 0x612(%rip),%ymm11 # 4f88 <_sk_callback_hsw+0x512> + .byte 196,98,125,24,21,33,6,0,0 // vbroadcastss 0x621(%rip),%ymm10 # 4fbc <_sk_callback_hsw+0x518> + .byte 196,98,125,24,29,28,6,0,0 // vbroadcastss 0x61c(%rip),%ymm11 # 4fc0 <_sk_callback_hsw+0x51c> .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) @@ -14434,16 +14444,16 @@ HIDDEN _sk_bicubic_n1y_hsw FUNCTION(_sk_bicubic_n1y_hsw) _sk_bicubic_n1y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,245,5,0,0 // vbroadcastss 0x5f5(%rip),%ymm1 # 4f8c <_sk_callback_hsw+0x516> + .byte 196,226,125,24,13,255,5,0,0 // vbroadcastss 0x5ff(%rip),%ymm1 # 4fc4 <_sk_callback_hsw+0x520> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,235,5,0,0 // vbroadcastss 0x5eb(%rip),%ymm8 # 4f90 <_sk_callback_hsw+0x51a> + .byte 196,98,125,24,5,245,5,0,0 // vbroadcastss 0x5f5(%rip),%ymm8 # 4fc8 <_sk_callback_hsw+0x524> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,225,5,0,0 // vbroadcastss 0x5e1(%rip),%ymm9 # 4f94 <_sk_callback_hsw+0x51e> - .byte 196,98,125,24,21,220,5,0,0 // vbroadcastss 0x5dc(%rip),%ymm10 # 4f98 <_sk_callback_hsw+0x522> + .byte 196,98,125,24,13,235,5,0,0 // vbroadcastss 0x5eb(%rip),%ymm9 # 4fcc <_sk_callback_hsw+0x528> + .byte 196,98,125,24,21,230,5,0,0 // vbroadcastss 0x5e6(%rip),%ymm10 # 4fd0 <_sk_callback_hsw+0x52c> .byte 196,66,61,168,209 // vfmadd213ps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,13,210,5,0,0 // vbroadcastss 0x5d2(%rip),%ymm9 # 4f9c <_sk_callback_hsw+0x526> + .byte 196,98,125,24,13,220,5,0,0 // vbroadcastss 0x5dc(%rip),%ymm9 # 4fd4 <_sk_callback_hsw+0x530> .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9 - .byte 196,98,125,24,21,200,5,0,0 // vbroadcastss 0x5c8(%rip),%ymm10 # 4fa0 <_sk_callback_hsw+0x52a> + .byte 196,98,125,24,21,210,5,0,0 // vbroadcastss 0x5d2(%rip),%ymm10 # 4fd8 <_sk_callback_hsw+0x534> .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 @@ -14454,14 +14464,14 @@ HIDDEN _sk_bicubic_p1y_hsw FUNCTION(_sk_bicubic_p1y_hsw) _sk_bicubic_p1y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,176,5,0,0 // vbroadcastss 0x5b0(%rip),%ymm8 # 4fa4 <_sk_callback_hsw+0x52e> + .byte 196,98,125,24,5,186,5,0,0 // vbroadcastss 0x5ba(%rip),%ymm8 # 4fdc <_sk_callback_hsw+0x538> .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 - .byte 196,98,125,24,21,161,5,0,0 // vbroadcastss 0x5a1(%rip),%ymm10 # 4fa8 <_sk_callback_hsw+0x532> - .byte 196,98,125,24,29,156,5,0,0 // vbroadcastss 0x59c(%rip),%ymm11 # 4fac <_sk_callback_hsw+0x536> + .byte 196,98,125,24,21,171,5,0,0 // vbroadcastss 0x5ab(%rip),%ymm10 # 4fe0 <_sk_callback_hsw+0x53c> + .byte 196,98,125,24,29,166,5,0,0 // vbroadcastss 0x5a6(%rip),%ymm11 # 4fe4 <_sk_callback_hsw+0x540> .byte 196,66,53,168,218 // vfmadd213ps %ymm10,%ymm9,%ymm11 .byte 196,66,53,168,216 // vfmadd213ps %ymm8,%ymm9,%ymm11 - .byte 196,98,125,24,5,141,5,0,0 // vbroadcastss 0x58d(%rip),%ymm8 # 4fb0 <_sk_callback_hsw+0x53a> + .byte 196,98,125,24,5,151,5,0,0 // vbroadcastss 0x597(%rip),%ymm8 # 4fe8 <_sk_callback_hsw+0x544> .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 @@ -14472,12 +14482,12 @@ HIDDEN _sk_bicubic_p3y_hsw FUNCTION(_sk_bicubic_p3y_hsw) _sk_bicubic_p3y_hsw: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,117,5,0,0 // vbroadcastss 0x575(%rip),%ymm1 # 4fb4 <_sk_callback_hsw+0x53e> + .byte 196,226,125,24,13,127,5,0,0 // vbroadcastss 0x57f(%rip),%ymm1 # 4fec <_sk_callback_hsw+0x548> .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 196,98,125,24,21,97,5,0,0 // vbroadcastss 0x561(%rip),%ymm10 # 4fb8 <_sk_callback_hsw+0x542> - .byte 196,98,125,24,29,92,5,0,0 // vbroadcastss 0x55c(%rip),%ymm11 # 4fbc <_sk_callback_hsw+0x546> + .byte 196,98,125,24,21,107,5,0,0 // vbroadcastss 0x56b(%rip),%ymm10 # 4ff0 <_sk_callback_hsw+0x54c> + .byte 196,98,125,24,29,102,5,0,0 // vbroadcastss 0x566(%rip),%ymm11 # 4ff4 <_sk_callback_hsw+0x550> .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) @@ -14607,25 +14617,25 @@ BALIGN4 .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 4c69 <.literal4+0xb1> + .byte 71,225,61 // rex.RXB loope 4c99 <.literal4+0xb1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 4c79 <.literal4+0xc1> + .byte 71,225,61 // rex.RXB loope 4ca9 <.literal4+0xc1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 4c89 <.literal4+0xd1> + .byte 71,225,61 // rex.RXB loope 4cb9 <.literal4+0xd1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 4c99 <.literal4+0xe1> + .byte 71,225,61 // rex.RXB loope 4cc9 <.literal4+0xe1> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax) @@ -14675,7 +14685,7 @@ 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 4cf1 <.literal4+0x139> + .byte 224,7 // loopne 4d21 <.literal4+0x139> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -14691,10 +14701,10 @@ 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 4d18 <.literal4+0x160> + .byte 127,0 // jg 4d48 <.literal4+0x160> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4d91 <.literal4+0x1d9> + .byte 119,115 // ja 4dc1 <.literal4+0x1d9> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -14708,10 +14718,10 @@ 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 4d4c <.literal4+0x194> + .byte 127,0 // jg 4d7c <.literal4+0x194> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4dc5 <.literal4+0x20d> + .byte 119,115 // ja 4df5 <.literal4+0x20d> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -14725,10 +14735,10 @@ 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 4d80 <.literal4+0x1c8> + .byte 127,0 // jg 4db0 <.literal4+0x1c8> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4df9 <.literal4+0x241> + .byte 119,115 // ja 4e29 <.literal4+0x241> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -14742,10 +14752,10 @@ 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 4db4 <.literal4+0x1fc> + .byte 127,0 // jg 4de4 <.literal4+0x1fc> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 4e2d <.literal4+0x275> + .byte 119,115 // ja 4e5d <.literal4+0x275> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -14758,7 +14768,7 @@ 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 4e2b <.literal4+0x273> + .byte 127,67 // jg 4e5b <.literal4+0x273> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -14770,10 +14780,10 @@ BALIGN4 .byte 190,80,128,3,62 // mov $0x3e038050,%esi .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 4e4b <.literal4+0x293> + .byte 118,63 // jbe 4e7b <.literal4+0x293> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 4e5f <.literal4+0x2a7> + .byte 127,67 // jg 4e8f <.literal4+0x2a7> .byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax) .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax) @@ -14782,7 +14792,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 4e41 <.literal4+0x289> + .byte 224,7 // loopne 4e71 <.literal4+0x289> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -14794,7 +14804,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 4e5d <.literal4+0x2a5> + .byte 224,7 // loopne 4e8d <.literal4+0x2a5> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -14805,7 +14815,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 4eb2 <.literal4+0x2fa> + .byte 124,66 // jl 4ee2 <.literal4+0x2fa> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax) @@ -14823,9 +14833,9 @@ 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 4ef5 <.literal4+0x33d> + .byte 112,65 // jo 4f25 <.literal4+0x33d> .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 4f03 <.literal4+0x34b> + .byte 127,67 // jg 4f33 <.literal4+0x34b> .byte 128,0,128 // addb $0x80,(%rax) .byte 55 // (bad) .byte 128,0,128 // addb $0x80,(%rax) @@ -14833,7 +14843,11 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 255 // (bad) - .byte 127,71 // jg 4f17 <.literal4+0x35f> + .byte 127,71 // jg 4f47 <.literal4+0x35f> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,191,0,0,0 // add %al,0xbf(%rax) @@ -14941,16 +14955,16 @@ 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) # a004fe8 <_sk_callback_hsw+0xa000572> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005028 <_sk_callback_hsw+0xa000584> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12004ff0 <_sk_callback_hsw+0x1200057a> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005030 <_sk_callback_hsw+0x1200058c> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a004ff8 <_sk_callback_hsw+0x1a000582> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005038 <_sk_callback_hsw+0x1a000594> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005000 <_sk_callback_hsw+0x300058a> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005040 <_sk_callback_hsw+0x300059c> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -14993,16 +15007,16 @@ 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) # a005048 <_sk_callback_hsw+0xa0005d2> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005088 <_sk_callback_hsw+0xa0005e4> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005050 <_sk_callback_hsw+0x120005da> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005090 <_sk_callback_hsw+0x120005ec> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005058 <_sk_callback_hsw+0x1a0005e2> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005098 <_sk_callback_hsw+0x1a0005f4> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005060 <_sk_callback_hsw+0x30005ea> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30050a0 <_sk_callback_hsw+0x30005fc> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -15045,16 +15059,16 @@ 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) # a0050a8 <_sk_callback_hsw+0xa000632> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0050e8 <_sk_callback_hsw+0xa000644> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120050b0 <_sk_callback_hsw+0x1200063a> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120050f0 <_sk_callback_hsw+0x1200064c> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0050b8 <_sk_callback_hsw+0x1a000642> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0050f8 <_sk_callback_hsw+0x1a000654> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30050c0 <_sk_callback_hsw+0x300064a> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005100 <_sk_callback_hsw+0x300065c> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -15097,16 +15111,16 @@ 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) # a005108 <_sk_callback_hsw+0xa000692> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005148 <_sk_callback_hsw+0xa0006a4> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005110 <_sk_callback_hsw+0x1200069a> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005150 <_sk_callback_hsw+0x120006ac> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005118 <_sk_callback_hsw+0x1a0006a2> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005158 <_sk_callback_hsw+0x1a0006b4> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005120 <_sk_callback_hsw+0x30006aa> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005160 <_sk_callback_hsw+0x30006bc> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -15149,16 +15163,16 @@ 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) # a005168 <_sk_callback_hsw+0xa0006f2> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0051a8 <_sk_callback_hsw+0xa000704> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 12005170 <_sk_callback_hsw+0x120006fa> + .byte 255,13,255,255,255,17 // decl 0x11ffffff(%rip) # 120051b0 <_sk_callback_hsw+0x1200070c> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a005178 <_sk_callback_hsw+0x1a000702> + .byte 255,21,255,255,255,25 // callq *0x19ffffff(%rip) # 1a0051b8 <_sk_callback_hsw+0x1a000714> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 3005180 <_sk_callback_hsw+0x300070a> + .byte 255,29,255,255,255,2 // lcall *0x2ffffff(%rip) # 30051c0 <_sk_callback_hsw+0x300071c> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -15319,7 +15333,7 @@ _sk_seed_shader_avx: .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,224,102,0,0 // vbroadcastss 0x66e0(%rip),%ymm1 # 67c0 <_sk_callback_avx+0x144> + .byte 196,226,125,24,13,28,103,0,0 // vbroadcastss 0x671c(%rip),%ymm1 # 67fc <_sk_callback_avx+0x142> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,88,7 // vaddps (%rdi),%ymm0,%ymm0 .byte 197,249,110,209 // vmovd %ecx,%xmm2 @@ -15328,7 +15342,7 @@ _sk_seed_shader_avx: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,21,186,102,0,0 // vbroadcastss 0x66ba(%rip),%ymm2 # 67c4 <_sk_callback_avx+0x148> + .byte 196,226,125,24,21,246,102,0,0 // vbroadcastss 0x66f6(%rip),%ymm2 # 6800 <_sk_callback_avx+0x146> .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 @@ -15352,7 +15366,7 @@ _sk_dither_avx: .byte 196,65,121,112,201,0 // vpshufd $0x0,%xmm9,%xmm9 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 .byte 196,65,52,87,208 // vxorps %ymm8,%ymm9,%ymm10 - .byte 196,98,125,24,29,99,102,0,0 // vbroadcastss 0x6663(%rip),%ymm11 # 67c8 <_sk_callback_avx+0x14c> + .byte 196,98,125,24,29,159,102,0,0 // vbroadcastss 0x669f(%rip),%ymm11 # 6804 <_sk_callback_avx+0x14a> .byte 196,65,44,84,203 // vandps %ymm11,%ymm10,%ymm9 .byte 196,193,25,114,241,5 // vpslld $0x5,%xmm9,%xmm12 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 @@ -15363,8 +15377,8 @@ _sk_dither_avx: .byte 196,67,125,25,219,1 // vextractf128 $0x1,%ymm11,%xmm11 .byte 196,193,33,114,243,4 // vpslld $0x4,%xmm11,%xmm11 .byte 196,67,29,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm12,%ymm11 - .byte 196,98,125,24,37,36,102,0,0 // vbroadcastss 0x6624(%rip),%ymm12 # 67cc <_sk_callback_avx+0x150> - .byte 196,98,125,24,45,31,102,0,0 // vbroadcastss 0x661f(%rip),%ymm13 # 67d0 <_sk_callback_avx+0x154> + .byte 196,98,125,24,37,96,102,0,0 // vbroadcastss 0x6660(%rip),%ymm12 # 6808 <_sk_callback_avx+0x14e> + .byte 196,98,125,24,45,91,102,0,0 // vbroadcastss 0x665b(%rip),%ymm13 # 680c <_sk_callback_avx+0x152> .byte 196,65,44,84,245 // vandps %ymm13,%ymm10,%ymm14 .byte 196,193,1,114,246,2 // vpslld $0x2,%xmm14,%xmm15 .byte 196,67,125,25,246,1 // vextractf128 $0x1,%ymm14,%xmm14 @@ -15391,9 +15405,9 @@ _sk_dither_avx: .byte 196,65,60,86,193 // vorps %ymm9,%ymm8,%ymm8 .byte 196,65,60,86,194 // vorps %ymm10,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,138,101,0,0 // vbroadcastss 0x658a(%rip),%ymm9 # 67d4 <_sk_callback_avx+0x158> + .byte 196,98,125,24,13,198,101,0,0 // vbroadcastss 0x65c6(%rip),%ymm9 # 6810 <_sk_callback_avx+0x156> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,128,101,0,0 // vbroadcastss 0x6580(%rip),%ymm9 # 67d8 <_sk_callback_avx+0x15c> + .byte 196,98,125,24,13,188,101,0,0 // vbroadcastss 0x65bc(%rip),%ymm9 # 6814 <_sk_callback_avx+0x15a> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 @@ -15462,7 +15476,7 @@ HIDDEN _sk_srcatop_avx FUNCTION(_sk_srcatop_avx) _sk_srcatop_avx: .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 - .byte 196,98,125,24,5,216,100,0,0 // vbroadcastss 0x64d8(%rip),%ymm8 # 67dc <_sk_callback_avx+0x160> + .byte 196,98,125,24,5,20,101,0,0 // vbroadcastss 0x6514(%rip),%ymm8 # 6818 <_sk_callback_avx+0x15e> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -15483,7 +15497,7 @@ HIDDEN _sk_dstatop_avx FUNCTION(_sk_dstatop_avx) _sk_dstatop_avx: .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 - .byte 196,98,125,24,13,154,100,0,0 // vbroadcastss 0x649a(%rip),%ymm9 # 67e0 <_sk_callback_avx+0x164> + .byte 196,98,125,24,13,214,100,0,0 // vbroadcastss 0x64d6(%rip),%ymm9 # 681c <_sk_callback_avx+0x162> .byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 @@ -15525,7 +15539,7 @@ HIDDEN _sk_srcout_avx .globl _sk_srcout_avx FUNCTION(_sk_srcout_avx) _sk_srcout_avx: - .byte 196,98,125,24,5,57,100,0,0 // vbroadcastss 0x6439(%rip),%ymm8 # 67e4 <_sk_callback_avx+0x168> + .byte 196,98,125,24,5,117,100,0,0 // vbroadcastss 0x6475(%rip),%ymm8 # 6820 <_sk_callback_avx+0x166> .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 @@ -15538,7 +15552,7 @@ HIDDEN _sk_dstout_avx .globl _sk_dstout_avx FUNCTION(_sk_dstout_avx) _sk_dstout_avx: - .byte 196,226,125,24,5,28,100,0,0 // vbroadcastss 0x641c(%rip),%ymm0 # 67e8 <_sk_callback_avx+0x16c> + .byte 196,226,125,24,5,88,100,0,0 // vbroadcastss 0x6458(%rip),%ymm0 # 6824 <_sk_callback_avx+0x16a> .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 @@ -15551,7 +15565,7 @@ HIDDEN _sk_srcover_avx .globl _sk_srcover_avx FUNCTION(_sk_srcover_avx) _sk_srcover_avx: - .byte 196,98,125,24,5,255,99,0,0 // vbroadcastss 0x63ff(%rip),%ymm8 # 67ec <_sk_callback_avx+0x170> + .byte 196,98,125,24,5,59,100,0,0 // vbroadcastss 0x643b(%rip),%ymm8 # 6828 <_sk_callback_avx+0x16e> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -15568,7 +15582,7 @@ HIDDEN _sk_dstover_avx .globl _sk_dstover_avx FUNCTION(_sk_dstover_avx) _sk_dstover_avx: - .byte 196,98,125,24,5,210,99,0,0 // vbroadcastss 0x63d2(%rip),%ymm8 # 67f0 <_sk_callback_avx+0x174> + .byte 196,98,125,24,5,14,100,0,0 // vbroadcastss 0x640e(%rip),%ymm8 # 682c <_sk_callback_avx+0x172> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 @@ -15596,7 +15610,7 @@ HIDDEN _sk_multiply_avx .globl _sk_multiply_avx FUNCTION(_sk_multiply_avx) _sk_multiply_avx: - .byte 196,98,125,24,5,145,99,0,0 // vbroadcastss 0x6391(%rip),%ymm8 # 67f4 <_sk_callback_avx+0x178> + .byte 196,98,125,24,5,205,99,0,0 // vbroadcastss 0x63cd(%rip),%ymm8 # 6830 <_sk_callback_avx+0x176> .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 @@ -15656,7 +15670,7 @@ HIDDEN _sk_xor__avx .globl _sk_xor__avx FUNCTION(_sk_xor__avx) _sk_xor__avx: - .byte 196,98,125,24,5,224,98,0,0 // vbroadcastss 0x62e0(%rip),%ymm8 # 67f8 <_sk_callback_avx+0x17c> + .byte 196,98,125,24,5,28,99,0,0 // vbroadcastss 0x631c(%rip),%ymm8 # 6834 <_sk_callback_avx+0x17a> .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 @@ -15693,7 +15707,7 @@ _sk_darken_avx: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,96,98,0,0 // vbroadcastss 0x6260(%rip),%ymm8 # 67fc <_sk_callback_avx+0x180> + .byte 196,98,125,24,5,156,98,0,0 // vbroadcastss 0x629c(%rip),%ymm8 # 6838 <_sk_callback_avx+0x17e> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15719,7 +15733,7 @@ _sk_lighten_avx: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,12,98,0,0 // vbroadcastss 0x620c(%rip),%ymm8 # 6800 <_sk_callback_avx+0x184> + .byte 196,98,125,24,5,72,98,0,0 // vbroadcastss 0x6248(%rip),%ymm8 # 683c <_sk_callback_avx+0x182> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15748,7 +15762,7 @@ _sk_difference_avx: .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,172,97,0,0 // vbroadcastss 0x61ac(%rip),%ymm8 # 6804 <_sk_callback_avx+0x188> + .byte 196,98,125,24,5,232,97,0,0 // vbroadcastss 0x61e8(%rip),%ymm8 # 6840 <_sk_callback_avx+0x186> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15771,7 +15785,7 @@ _sk_exclusion_avx: .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,103,97,0,0 // vbroadcastss 0x6167(%rip),%ymm8 # 6808 <_sk_callback_avx+0x18c> + .byte 196,98,125,24,5,163,97,0,0 // vbroadcastss 0x61a3(%rip),%ymm8 # 6844 <_sk_callback_avx+0x18a> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15782,7 +15796,7 @@ HIDDEN _sk_colorburn_avx .globl _sk_colorburn_avx FUNCTION(_sk_colorburn_avx) _sk_colorburn_avx: - .byte 196,98,125,24,5,82,97,0,0 // vbroadcastss 0x6152(%rip),%ymm8 # 680c <_sk_callback_avx+0x190> + .byte 196,98,125,24,5,142,97,0,0 // vbroadcastss 0x618e(%rip),%ymm8 # 6848 <_sk_callback_avx+0x18e> .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 @@ -15844,7 +15858,7 @@ HIDDEN _sk_colordodge_avx FUNCTION(_sk_colordodge_avx) _sk_colordodge_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 196,98,125,24,13,78,96,0,0 // vbroadcastss 0x604e(%rip),%ymm9 # 6810 <_sk_callback_avx+0x194> + .byte 196,98,125,24,13,138,96,0,0 // vbroadcastss 0x608a(%rip),%ymm9 # 684c <_sk_callback_avx+0x192> .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 @@ -15901,7 +15915,7 @@ HIDDEN _sk_hardlight_avx .globl _sk_hardlight_avx FUNCTION(_sk_hardlight_avx) _sk_hardlight_avx: - .byte 196,98,125,24,5,96,95,0,0 // vbroadcastss 0x5f60(%rip),%ymm8 # 6814 <_sk_callback_avx+0x198> + .byte 196,98,125,24,5,156,95,0,0 // vbroadcastss 0x5f9c(%rip),%ymm8 # 6850 <_sk_callback_avx+0x196> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -15956,7 +15970,7 @@ HIDDEN _sk_overlay_avx .globl _sk_overlay_avx FUNCTION(_sk_overlay_avx) _sk_overlay_avx: - .byte 196,98,125,24,5,137,94,0,0 // vbroadcastss 0x5e89(%rip),%ymm8 # 6818 <_sk_callback_avx+0x19c> + .byte 196,98,125,24,5,197,94,0,0 // vbroadcastss 0x5ec5(%rip),%ymm8 # 6854 <_sk_callback_avx+0x19a> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -16022,10 +16036,10 @@ _sk_softlight_avx: .byte 196,65,60,88,192 // vaddps %ymm8,%ymm8,%ymm8 .byte 196,65,60,89,216 // vmulps %ymm8,%ymm8,%ymm11 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 - .byte 196,98,125,24,29,128,93,0,0 // vbroadcastss 0x5d80(%rip),%ymm11 # 6820 <_sk_callback_avx+0x1a4> + .byte 196,98,125,24,29,188,93,0,0 // vbroadcastss 0x5dbc(%rip),%ymm11 # 685c <_sk_callback_avx+0x1a2> .byte 196,65,28,88,235 // vaddps %ymm11,%ymm12,%ymm13 .byte 196,65,20,89,192 // vmulps %ymm8,%ymm13,%ymm8 - .byte 196,98,125,24,45,113,93,0,0 // vbroadcastss 0x5d71(%rip),%ymm13 # 6824 <_sk_callback_avx+0x1a8> + .byte 196,98,125,24,45,173,93,0,0 // vbroadcastss 0x5dad(%rip),%ymm13 # 6860 <_sk_callback_avx+0x1a6> .byte 196,65,28,89,245 // vmulps %ymm13,%ymm12,%ymm14 .byte 196,65,12,88,192 // vaddps %ymm8,%ymm14,%ymm8 .byte 196,65,124,82,244 // vrsqrtps %ymm12,%ymm14 @@ -16036,7 +16050,7 @@ _sk_softlight_avx: .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15 .byte 196,67,13,74,240,240 // vblendvps %ymm15,%ymm8,%ymm14,%ymm14 .byte 197,116,88,249 // vaddps %ymm1,%ymm1,%ymm15 - .byte 196,98,125,24,5,47,93,0,0 // vbroadcastss 0x5d2f(%rip),%ymm8 # 681c <_sk_callback_avx+0x1a0> + .byte 196,98,125,24,5,107,93,0,0 // vbroadcastss 0x5d6b(%rip),%ymm8 # 6858 <_sk_callback_avx+0x19e> .byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12 .byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0 .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 @@ -16163,12 +16177,12 @@ _sk_hue_avx: .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11 .byte 196,65,36,94,222 // vdivps %ymm14,%ymm11,%ymm11 .byte 196,67,37,74,224,240 // vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - .byte 196,98,125,24,53,254,90,0,0 // vbroadcastss 0x5afe(%rip),%ymm14 # 6828 <_sk_callback_avx+0x1ac> + .byte 196,98,125,24,53,58,91,0,0 // vbroadcastss 0x5b3a(%rip),%ymm14 # 6864 <_sk_callback_avx+0x1aa> .byte 196,65,92,89,222 // vmulps %ymm14,%ymm4,%ymm11 - .byte 196,98,125,24,61,244,90,0,0 // vbroadcastss 0x5af4(%rip),%ymm15 # 682c <_sk_callback_avx+0x1b0> + .byte 196,98,125,24,61,48,91,0,0 // vbroadcastss 0x5b30(%rip),%ymm15 # 6868 <_sk_callback_avx+0x1ae> .byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 - .byte 196,226,125,24,5,229,90,0,0 // vbroadcastss 0x5ae5(%rip),%ymm0 # 6830 <_sk_callback_avx+0x1b4> + .byte 196,226,125,24,5,33,91,0,0 // vbroadcastss 0x5b21(%rip),%ymm0 # 686c <_sk_callback_avx+0x1b2> .byte 197,76,89,232 // vmulps %ymm0,%ymm6,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 .byte 196,65,52,89,238 // vmulps %ymm14,%ymm9,%ymm13 @@ -16229,7 +16243,7 @@ _sk_hue_avx: .byte 196,65,36,95,208 // vmaxps %ymm8,%ymm11,%ymm10 .byte 196,195,109,74,209,240 // vblendvps %ymm15,%ymm9,%ymm2,%ymm2 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,190,89,0,0 // vbroadcastss 0x59be(%rip),%ymm8 # 6834 <_sk_callback_avx+0x1b8> + .byte 196,98,125,24,5,250,89,0,0 // vbroadcastss 0x59fa(%rip),%ymm8 # 6870 <_sk_callback_avx+0x1b6> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -16286,12 +16300,12 @@ _sk_saturation_avx: .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11 .byte 196,65,36,94,222 // vdivps %ymm14,%ymm11,%ymm11 .byte 196,67,37,74,224,240 // vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - .byte 196,98,125,24,53,204,88,0,0 // vbroadcastss 0x58cc(%rip),%ymm14 # 6838 <_sk_callback_avx+0x1bc> + .byte 196,98,125,24,53,8,89,0,0 // vbroadcastss 0x5908(%rip),%ymm14 # 6874 <_sk_callback_avx+0x1ba> .byte 196,65,92,89,222 // vmulps %ymm14,%ymm4,%ymm11 - .byte 196,98,125,24,61,194,88,0,0 // vbroadcastss 0x58c2(%rip),%ymm15 # 683c <_sk_callback_avx+0x1c0> + .byte 196,98,125,24,61,254,88,0,0 // vbroadcastss 0x58fe(%rip),%ymm15 # 6878 <_sk_callback_avx+0x1be> .byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 - .byte 196,226,125,24,5,179,88,0,0 // vbroadcastss 0x58b3(%rip),%ymm0 # 6840 <_sk_callback_avx+0x1c4> + .byte 196,226,125,24,5,239,88,0,0 // vbroadcastss 0x58ef(%rip),%ymm0 # 687c <_sk_callback_avx+0x1c2> .byte 197,76,89,232 // vmulps %ymm0,%ymm6,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 .byte 196,65,52,89,238 // vmulps %ymm14,%ymm9,%ymm13 @@ -16352,7 +16366,7 @@ _sk_saturation_avx: .byte 196,65,36,95,208 // vmaxps %ymm8,%ymm11,%ymm10 .byte 196,195,109,74,209,240 // vblendvps %ymm15,%ymm9,%ymm2,%ymm2 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,140,87,0,0 // vbroadcastss 0x578c(%rip),%ymm8 # 6844 <_sk_callback_avx+0x1c8> + .byte 196,98,125,24,5,200,87,0,0 // vbroadcastss 0x57c8(%rip),%ymm8 # 6880 <_sk_callback_avx+0x1c6> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -16381,12 +16395,12 @@ _sk_color_avx: .byte 197,252,17,68,36,168 // vmovups %ymm0,-0x58(%rsp) .byte 197,124,89,199 // vmulps %ymm7,%ymm0,%ymm8 .byte 197,116,89,207 // vmulps %ymm7,%ymm1,%ymm9 - .byte 196,98,125,24,45,34,87,0,0 // vbroadcastss 0x5722(%rip),%ymm13 # 6848 <_sk_callback_avx+0x1cc> + .byte 196,98,125,24,45,94,87,0,0 // vbroadcastss 0x575e(%rip),%ymm13 # 6884 <_sk_callback_avx+0x1ca> .byte 196,65,92,89,213 // vmulps %ymm13,%ymm4,%ymm10 - .byte 196,98,125,24,53,24,87,0,0 // vbroadcastss 0x5718(%rip),%ymm14 # 684c <_sk_callback_avx+0x1d0> + .byte 196,98,125,24,53,84,87,0,0 // vbroadcastss 0x5754(%rip),%ymm14 # 6888 <_sk_callback_avx+0x1ce> .byte 196,65,84,89,222 // vmulps %ymm14,%ymm5,%ymm11 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,24,61,9,87,0,0 // vbroadcastss 0x5709(%rip),%ymm15 # 6850 <_sk_callback_avx+0x1d4> + .byte 196,98,125,24,61,69,87,0,0 // vbroadcastss 0x5745(%rip),%ymm15 # 688c <_sk_callback_avx+0x1d2> .byte 196,65,76,89,223 // vmulps %ymm15,%ymm6,%ymm11 .byte 196,193,44,88,195 // vaddps %ymm11,%ymm10,%ymm0 .byte 196,65,60,89,221 // vmulps %ymm13,%ymm8,%ymm11 @@ -16449,7 +16463,7 @@ _sk_color_avx: .byte 196,65,44,95,207 // vmaxps %ymm15,%ymm10,%ymm9 .byte 196,195,37,74,192,0 // vblendvps %ymm0,%ymm8,%ymm11,%ymm0 .byte 196,65,124,95,199 // vmaxps %ymm15,%ymm0,%ymm8 - .byte 196,226,125,24,5,208,85,0,0 // vbroadcastss 0x55d0(%rip),%ymm0 # 6854 <_sk_callback_avx+0x1d8> + .byte 196,226,125,24,5,12,86,0,0 // vbroadcastss 0x560c(%rip),%ymm0 # 6890 <_sk_callback_avx+0x1d6> .byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10 .byte 197,172,89,84,36,168 // vmulps -0x58(%rsp),%ymm10,%ymm2 .byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11 @@ -16479,12 +16493,12 @@ _sk_luminosity_avx: .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 - .byte 196,98,125,24,45,98,85,0,0 // vbroadcastss 0x5562(%rip),%ymm13 # 6858 <_sk_callback_avx+0x1dc> + .byte 196,98,125,24,45,158,85,0,0 // vbroadcastss 0x559e(%rip),%ymm13 # 6894 <_sk_callback_avx+0x1da> .byte 196,65,108,89,213 // vmulps %ymm13,%ymm2,%ymm10 - .byte 196,98,125,24,53,88,85,0,0 // vbroadcastss 0x5558(%rip),%ymm14 # 685c <_sk_callback_avx+0x1e0> + .byte 196,98,125,24,53,148,85,0,0 // vbroadcastss 0x5594(%rip),%ymm14 # 6898 <_sk_callback_avx+0x1de> .byte 196,65,116,89,222 // vmulps %ymm14,%ymm1,%ymm11 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,24,61,73,85,0,0 // vbroadcastss 0x5549(%rip),%ymm15 # 6860 <_sk_callback_avx+0x1e4> + .byte 196,98,125,24,61,133,85,0,0 // vbroadcastss 0x5585(%rip),%ymm15 # 689c <_sk_callback_avx+0x1e2> .byte 196,65,28,89,223 // vmulps %ymm15,%ymm12,%ymm11 .byte 196,193,44,88,195 // vaddps %ymm11,%ymm10,%ymm0 .byte 196,65,60,89,221 // vmulps %ymm13,%ymm8,%ymm11 @@ -16547,7 +16561,7 @@ _sk_luminosity_avx: .byte 196,65,44,95,207 // vmaxps %ymm15,%ymm10,%ymm9 .byte 196,195,37,74,192,0 // vblendvps %ymm0,%ymm8,%ymm11,%ymm0 .byte 196,65,124,95,199 // vmaxps %ymm15,%ymm0,%ymm8 - .byte 196,226,125,24,5,16,84,0,0 // vbroadcastss 0x5410(%rip),%ymm0 # 6864 <_sk_callback_avx+0x1e8> + .byte 196,226,125,24,5,76,84,0,0 // vbroadcastss 0x544c(%rip),%ymm0 # 68a0 <_sk_callback_avx+0x1e6> .byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11 @@ -16578,7 +16592,7 @@ _sk_srcover_rgba_8888_avx: .byte 77,133,192 // test %r8,%r8 .byte 15,133,43,1,0,0 // jne 15e3 <_sk_srcover_rgba_8888_avx+0x144> .byte 196,193,124,16,58 // vmovups (%r10),%ymm7 - .byte 197,124,40,13,27,88,0,0 // vmovaps 0x581b(%rip),%ymm9 # 6ce0 <_sk_callback_avx+0x664> + .byte 197,124,40,13,91,88,0,0 // vmovaps 0x585b(%rip),%ymm9 # 6d20 <_sk_callback_avx+0x666> .byte 196,193,68,84,225 // vandps %ymm9,%ymm7,%ymm4 .byte 197,252,91,228 // vcvtdq2ps %ymm4,%ymm4 .byte 197,209,114,215,8 // vpsrld $0x8,%xmm7,%xmm5 @@ -16596,9 +16610,9 @@ _sk_srcover_rgba_8888_avx: .byte 196,193,65,114,208,24 // vpsrld $0x18,%xmm8,%xmm7 .byte 196,227,53,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm9,%ymm7 .byte 197,252,91,255 // vcvtdq2ps %ymm7,%ymm7 - .byte 196,98,125,24,5,66,83,0,0 // vbroadcastss 0x5342(%rip),%ymm8 # 6868 <_sk_callback_avx+0x1ec> + .byte 196,98,125,24,5,126,83,0,0 // vbroadcastss 0x537e(%rip),%ymm8 # 68a4 <_sk_callback_avx+0x1ea> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 - .byte 196,98,125,24,13,57,83,0,0 // vbroadcastss 0x5339(%rip),%ymm9 # 686c <_sk_callback_avx+0x1f0> + .byte 196,98,125,24,13,117,83,0,0 // vbroadcastss 0x5375(%rip),%ymm9 # 68a8 <_sk_callback_avx+0x1ee> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 @@ -16643,9 +16657,9 @@ _sk_srcover_rgba_8888_avx: .byte 72,211,232 // shr %cl,%rax .byte 196,225,249,110,224 // vmovq %rax,%xmm4 .byte 196,226,121,48,228 // vpmovzxbw %xmm4,%xmm4 - .byte 196,226,89,0,45,229,85,0,0 // vpshufb 0x55e5(%rip),%xmm4,%xmm5 # 6bf0 <_sk_callback_avx+0x574> + .byte 196,226,89,0,45,53,86,0,0 // vpshufb 0x5635(%rip),%xmm4,%xmm5 # 6c40 <_sk_callback_avx+0x586> .byte 196,226,121,33,237 // vpmovsxbd %xmm5,%xmm5 - .byte 196,226,89,0,37,231,85,0,0 // vpshufb 0x55e7(%rip),%xmm4,%xmm4 # 6c00 <_sk_callback_avx+0x584> + .byte 196,226,89,0,37,55,86,0,0 // vpshufb 0x5637(%rip),%xmm4,%xmm4 # 6c50 <_sk_callback_avx+0x596> .byte 196,226,121,33,228 // vpmovsxbd %xmm4,%xmm4 .byte 196,227,85,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 .byte 196,194,93,44,58 // vmaskmovps (%r10),%ymm4,%ymm7 @@ -16657,9 +16671,9 @@ _sk_srcover_rgba_8888_avx: .byte 72,211,232 // shr %cl,%rax .byte 196,97,249,110,200 // vmovq %rax,%xmm9 .byte 196,66,121,48,201 // vpmovzxbw %xmm9,%xmm9 - .byte 196,98,49,0,21,154,85,0,0 // vpshufb 0x559a(%rip),%xmm9,%xmm10 # 6bf0 <_sk_callback_avx+0x574> + .byte 196,98,49,0,21,234,85,0,0 // vpshufb 0x55ea(%rip),%xmm9,%xmm10 # 6c40 <_sk_callback_avx+0x586> .byte 196,66,121,33,210 // vpmovsxbd %xmm10,%xmm10 - .byte 196,98,49,0,13,156,85,0,0 // vpshufb 0x559c(%rip),%xmm9,%xmm9 # 6c00 <_sk_callback_avx+0x584> + .byte 196,98,49,0,13,236,85,0,0 // vpshufb 0x55ec(%rip),%xmm9,%xmm9 # 6c50 <_sk_callback_avx+0x596> .byte 196,66,121,33,201 // vpmovsxbd %xmm9,%xmm9 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 .byte 196,66,53,46,2 // vmaskmovps %ymm8,%ymm9,(%r10) @@ -16681,7 +16695,7 @@ HIDDEN _sk_clamp_1_avx .globl _sk_clamp_1_avx FUNCTION(_sk_clamp_1_avx) _sk_clamp_1_avx: - .byte 196,98,125,24,5,209,81,0,0 // vbroadcastss 0x51d1(%rip),%ymm8 # 6870 <_sk_callback_avx+0x1f4> + .byte 196,98,125,24,5,13,82,0,0 // vbroadcastss 0x520d(%rip),%ymm8 # 68ac <_sk_callback_avx+0x1f2> .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 @@ -16693,7 +16707,7 @@ HIDDEN _sk_clamp_a_avx .globl _sk_clamp_a_avx FUNCTION(_sk_clamp_a_avx) _sk_clamp_a_avx: - .byte 196,98,125,24,5,180,81,0,0 // vbroadcastss 0x51b4(%rip),%ymm8 # 6874 <_sk_callback_avx+0x1f8> + .byte 196,98,125,24,5,240,81,0,0 // vbroadcastss 0x51f0(%rip),%ymm8 # 68b0 <_sk_callback_avx+0x1f6> .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 @@ -16779,7 +16793,7 @@ FUNCTION(_sk_unpremul_avx) _sk_unpremul_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 - .byte 196,98,125,24,21,252,80,0,0 // vbroadcastss 0x50fc(%rip),%ymm10 # 6878 <_sk_callback_avx+0x1fc> + .byte 196,98,125,24,21,56,81,0,0 // vbroadcastss 0x5138(%rip),%ymm10 # 68b4 <_sk_callback_avx+0x1fa> .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 @@ -16792,17 +16806,17 @@ HIDDEN _sk_from_srgb_avx .globl _sk_from_srgb_avx FUNCTION(_sk_from_srgb_avx) _sk_from_srgb_avx: - .byte 196,98,125,24,5,221,80,0,0 // vbroadcastss 0x50dd(%rip),%ymm8 # 687c <_sk_callback_avx+0x200> + .byte 196,98,125,24,5,25,81,0,0 // vbroadcastss 0x5119(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x1fe> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 - .byte 196,98,125,24,29,207,80,0,0 // vbroadcastss 0x50cf(%rip),%ymm11 # 6880 <_sk_callback_avx+0x204> + .byte 196,98,125,24,29,11,81,0,0 // vbroadcastss 0x510b(%rip),%ymm11 # 68bc <_sk_callback_avx+0x202> .byte 196,65,124,89,227 // vmulps %ymm11,%ymm0,%ymm12 - .byte 196,98,125,24,45,197,80,0,0 // vbroadcastss 0x50c5(%rip),%ymm13 # 6884 <_sk_callback_avx+0x208> + .byte 196,98,125,24,45,1,81,0,0 // vbroadcastss 0x5101(%rip),%ymm13 # 68c0 <_sk_callback_avx+0x206> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,182,80,0,0 // vbroadcastss 0x50b6(%rip),%ymm12 # 6888 <_sk_callback_avx+0x20c> + .byte 196,98,125,24,37,242,80,0,0 // vbroadcastss 0x50f2(%rip),%ymm12 # 68c4 <_sk_callback_avx+0x20a> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,53,172,80,0,0 // vbroadcastss 0x50ac(%rip),%ymm14 # 688c <_sk_callback_avx+0x210> + .byte 196,98,125,24,53,232,80,0,0 // vbroadcastss 0x50e8(%rip),%ymm14 # 68c8 <_sk_callback_avx+0x20e> .byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0 .byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0 .byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9 @@ -16829,20 +16843,20 @@ HIDDEN _sk_to_srgb_avx FUNCTION(_sk_to_srgb_avx) _sk_to_srgb_avx: .byte 197,124,82,200 // vrsqrtps %ymm0,%ymm9 - .byte 196,98,125,24,5,65,80,0,0 // vbroadcastss 0x5041(%rip),%ymm8 # 6890 <_sk_callback_avx+0x214> + .byte 196,98,125,24,5,125,80,0,0 // vbroadcastss 0x507d(%rip),%ymm8 # 68cc <_sk_callback_avx+0x212> .byte 196,65,124,89,208 // vmulps %ymm8,%ymm0,%ymm10 - .byte 196,98,125,24,29,55,80,0,0 // vbroadcastss 0x5037(%rip),%ymm11 # 6894 <_sk_callback_avx+0x218> + .byte 196,98,125,24,29,115,80,0,0 // vbroadcastss 0x5073(%rip),%ymm11 # 68d0 <_sk_callback_avx+0x216> .byte 196,65,52,89,227 // vmulps %ymm11,%ymm9,%ymm12 - .byte 196,98,125,24,45,45,80,0,0 // vbroadcastss 0x502d(%rip),%ymm13 # 6898 <_sk_callback_avx+0x21c> + .byte 196,98,125,24,45,105,80,0,0 // vbroadcastss 0x5069(%rip),%ymm13 # 68d4 <_sk_callback_avx+0x21a> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,52,89,228 // vmulps %ymm12,%ymm9,%ymm12 - .byte 196,98,125,24,53,30,80,0,0 // vbroadcastss 0x501e(%rip),%ymm14 # 689c <_sk_callback_avx+0x220> + .byte 196,98,125,24,53,90,80,0,0 // vbroadcastss 0x505a(%rip),%ymm14 # 68d8 <_sk_callback_avx+0x21e> .byte 196,65,28,88,230 // vaddps %ymm14,%ymm12,%ymm12 - .byte 196,98,125,24,61,20,80,0,0 // vbroadcastss 0x5014(%rip),%ymm15 # 68a0 <_sk_callback_avx+0x224> + .byte 196,98,125,24,61,80,80,0,0 // vbroadcastss 0x5050(%rip),%ymm15 # 68dc <_sk_callback_avx+0x222> .byte 196,65,52,88,207 // vaddps %ymm15,%ymm9,%ymm9 .byte 196,65,124,83,201 // vrcpps %ymm9,%ymm9 .byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9 - .byte 196,98,125,24,37,0,80,0,0 // vbroadcastss 0x5000(%rip),%ymm12 # 68a4 <_sk_callback_avx+0x228> + .byte 196,98,125,24,37,60,80,0,0 // vbroadcastss 0x503c(%rip),%ymm12 # 68e0 <_sk_callback_avx+0x226> .byte 196,193,124,194,196,1 // vcmpltps %ymm12,%ymm0,%ymm0 .byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0 .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 @@ -16879,7 +16893,7 @@ _sk_rgb_to_hsl_avx: .byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9 .byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9 .byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,29,102,79,0,0 // vbroadcastss 0x4f66(%rip),%ymm11 # 68a8 <_sk_callback_avx+0x22c> + .byte 196,98,125,24,29,162,79,0,0 // vbroadcastss 0x4fa2(%rip),%ymm11 # 68e4 <_sk_callback_avx+0x22a> .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11 .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12 .byte 196,65,28,89,227 // vmulps %ymm11,%ymm12,%ymm12 @@ -16889,19 +16903,19 @@ _sk_rgb_to_hsl_avx: .byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2 .byte 197,252,92,201 // vsubps %ymm1,%ymm0,%ymm1 .byte 196,193,116,89,203 // vmulps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,63,79,0,0 // vbroadcastss 0x4f3f(%rip),%ymm11 # 68b4 <_sk_callback_avx+0x238> + .byte 196,98,125,24,29,123,79,0,0 // vbroadcastss 0x4f7b(%rip),%ymm11 # 68f0 <_sk_callback_avx+0x236> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,45,79,0,0 // vbroadcastss 0x4f2d(%rip),%ymm11 # 68b0 <_sk_callback_avx+0x234> + .byte 196,98,125,24,29,105,79,0,0 // vbroadcastss 0x4f69(%rip),%ymm11 # 68ec <_sk_callback_avx+0x232> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,227,117,74,202,224 // vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,21,79,0,0 // vbroadcastss 0x4f15(%rip),%ymm2 # 68ac <_sk_callback_avx+0x230> + .byte 196,226,125,24,21,81,79,0,0 // vbroadcastss 0x4f51(%rip),%ymm2 # 68e8 <_sk_callback_avx+0x22e> .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14 .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2 .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0 .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 .byte 196,227,117,74,194,0 // vblendvps %ymm0,%ymm2,%ymm1,%ymm0 .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1 - .byte 196,98,125,24,37,252,78,0,0 // vbroadcastss 0x4efc(%rip),%ymm12 # 68bc <_sk_callback_avx+0x240> + .byte 196,98,125,24,37,56,79,0,0 // vbroadcastss 0x4f38(%rip),%ymm12 # 68f8 <_sk_callback_avx+0x23e> .byte 196,193,116,89,212 // vmulps %ymm12,%ymm1,%ymm2 .byte 197,28,194,226,1 // vcmpltps %ymm2,%ymm12,%ymm12 .byte 196,65,36,92,216 // vsubps %ymm8,%ymm11,%ymm11 @@ -16911,7 +16925,7 @@ _sk_rgb_to_hsl_avx: .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 .byte 196,195,125,74,198,128 // vblendvps %ymm8,%ymm14,%ymm0,%ymm0 .byte 196,195,117,74,206,128 // vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - .byte 196,98,125,24,5,191,78,0,0 // vbroadcastss 0x4ebf(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x23c> + .byte 196,98,125,24,5,251,78,0,0 // vbroadcastss 0x4efb(%rip),%ymm8 # 68f4 <_sk_callback_avx+0x23a> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -16928,7 +16942,7 @@ _sk_hsl_to_rgb_avx: .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) .byte 197,252,40,225 // vmovaps %ymm1,%ymm4 .byte 197,252,40,216 // vmovaps %ymm0,%ymm3 - .byte 196,98,125,24,5,140,78,0,0 // vbroadcastss 0x4e8c(%rip),%ymm8 # 68c0 <_sk_callback_avx+0x244> + .byte 196,98,125,24,5,200,78,0,0 // vbroadcastss 0x4ec8(%rip),%ymm8 # 68fc <_sk_callback_avx+0x242> .byte 197,60,194,202,2 // vcmpleps %ymm2,%ymm8,%ymm9 .byte 197,92,89,210 // vmulps %ymm2,%ymm4,%ymm10 .byte 196,65,92,92,218 // vsubps %ymm10,%ymm4,%ymm11 @@ -16936,23 +16950,23 @@ _sk_hsl_to_rgb_avx: .byte 197,52,88,210 // vaddps %ymm2,%ymm9,%ymm10 .byte 197,108,88,202 // vaddps %ymm2,%ymm2,%ymm9 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,29,102,78,0,0 // vbroadcastss 0x4e66(%rip),%ymm11 # 68c4 <_sk_callback_avx+0x248> + .byte 196,98,125,24,29,162,78,0,0 // vbroadcastss 0x4ea2(%rip),%ymm11 # 6900 <_sk_callback_avx+0x246> .byte 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 .byte 196,67,125,8,227,1 // vroundps $0x1,%ymm11,%ymm12 .byte 196,65,36,92,252 // vsubps %ymm12,%ymm11,%ymm15 .byte 196,65,44,92,217 // vsubps %ymm9,%ymm10,%ymm11 - .byte 196,98,125,24,37,80,78,0,0 // vbroadcastss 0x4e50(%rip),%ymm12 # 68cc <_sk_callback_avx+0x250> + .byte 196,98,125,24,37,140,78,0,0 // vbroadcastss 0x4e8c(%rip),%ymm12 # 6908 <_sk_callback_avx+0x24e> .byte 196,193,4,89,196 // vmulps %ymm12,%ymm15,%ymm0 - .byte 196,98,125,24,45,70,78,0,0 // vbroadcastss 0x4e46(%rip),%ymm13 # 68d0 <_sk_callback_avx+0x254> + .byte 196,98,125,24,45,130,78,0,0 // vbroadcastss 0x4e82(%rip),%ymm13 # 690c <_sk_callback_avx+0x252> .byte 197,20,92,240 // vsubps %ymm0,%ymm13,%ymm14 .byte 196,65,36,89,246 // vmulps %ymm14,%ymm11,%ymm14 .byte 196,65,52,88,246 // vaddps %ymm14,%ymm9,%ymm14 - .byte 196,226,125,24,13,39,78,0,0 // vbroadcastss 0x4e27(%rip),%ymm1 # 68c8 <_sk_callback_avx+0x24c> + .byte 196,226,125,24,13,99,78,0,0 // vbroadcastss 0x4e63(%rip),%ymm1 # 6904 <_sk_callback_avx+0x24a> .byte 196,193,116,194,255,2 // vcmpleps %ymm15,%ymm1,%ymm7 .byte 196,195,13,74,249,112 // vblendvps %ymm7,%ymm9,%ymm14,%ymm7 .byte 196,65,60,194,247,2 // vcmpleps %ymm15,%ymm8,%ymm14 .byte 196,227,45,74,255,224 // vblendvps %ymm14,%ymm7,%ymm10,%ymm7 - .byte 196,98,125,24,53,18,78,0,0 // vbroadcastss 0x4e12(%rip),%ymm14 # 68d4 <_sk_callback_avx+0x258> + .byte 196,98,125,24,53,78,78,0,0 // vbroadcastss 0x4e4e(%rip),%ymm14 # 6910 <_sk_callback_avx+0x256> .byte 196,65,12,194,255,2 // vcmpleps %ymm15,%ymm14,%ymm15 .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -16971,7 +16985,7 @@ _sk_hsl_to_rgb_avx: .byte 197,164,89,247 // vmulps %ymm7,%ymm11,%ymm6 .byte 197,180,88,246 // vaddps %ymm6,%ymm9,%ymm6 .byte 196,227,77,74,237,0 // vblendvps %ymm0,%ymm5,%ymm6,%ymm5 - .byte 196,226,125,24,5,180,77,0,0 // vbroadcastss 0x4db4(%rip),%ymm0 # 68d8 <_sk_callback_avx+0x25c> + .byte 196,226,125,24,5,240,77,0,0 // vbroadcastss 0x4df0(%rip),%ymm0 # 6914 <_sk_callback_avx+0x25a> .byte 197,228,88,192 // vaddps %ymm0,%ymm3,%ymm0 .byte 196,227,125,8,216,1 // vroundps $0x1,%ymm0,%ymm3 .byte 197,252,92,195 // vsubps %ymm3,%ymm0,%ymm0 @@ -17023,13 +17037,13 @@ _sk_scale_u8_avx: .byte 77,133,192 // test %r8,%r8 .byte 117,74 // jne 1c1b <_sk_scale_u8_avx+0x54> .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8 - .byte 197,57,219,5,49,80,0,0 // vpand 0x5031(%rip),%xmm8,%xmm8 # 6c10 <_sk_callback_avx+0x594> + .byte 197,57,219,5,129,80,0,0 // vpand 0x5081(%rip),%xmm8,%xmm8 # 6c60 <_sk_callback_avx+0x5a6> .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9 .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 .byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,218,76,0,0 // vbroadcastss 0x4cda(%rip),%ymm9 # 68dc <_sk_callback_avx+0x260> + .byte 196,98,125,24,13,22,77,0,0 // vbroadcastss 0x4d16(%rip),%ymm9 # 6918 <_sk_callback_avx+0x25e> .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 @@ -17116,13 +17130,13 @@ _sk_lerp_u8_avx: .byte 77,133,192 // test %r8,%r8 .byte 117,110 // jne 1d6f <_sk_lerp_u8_avx+0x78> .byte 196,66,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm8 - .byte 197,57,219,5,17,79,0,0 // vpand 0x4f11(%rip),%xmm8,%xmm8 # 6c20 <_sk_callback_avx+0x5a4> + .byte 197,57,219,5,97,79,0,0 // vpand 0x4f61(%rip),%xmm8,%xmm8 # 6c70 <_sk_callback_avx+0x5b6> .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9 .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 .byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,174,75,0,0 // vbroadcastss 0x4bae(%rip),%ymm9 # 68e0 <_sk_callback_avx+0x264> + .byte 196,98,125,24,13,234,75,0,0 // vbroadcastss 0x4bea(%rip),%ymm9 # 691c <_sk_callback_avx+0x262> .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 @@ -17200,20 +17214,20 @@ _sk_lerp_565_avx: .byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9 .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 .byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,166,74,0,0 // vbroadcastss 0x4aa6(%rip),%ymm9 # 68e4 <_sk_callback_avx+0x268> + .byte 196,98,125,24,13,226,74,0,0 // vbroadcastss 0x4ae2(%rip),%ymm9 # 6920 <_sk_callback_avx+0x266> .byte 196,65,60,84,201 // vandps %ymm9,%ymm8,%ymm9 .byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9 - .byte 196,98,125,24,21,151,74,0,0 // vbroadcastss 0x4a97(%rip),%ymm10 # 68e8 <_sk_callback_avx+0x26c> + .byte 196,98,125,24,21,211,74,0,0 // vbroadcastss 0x4ad3(%rip),%ymm10 # 6924 <_sk_callback_avx+0x26a> .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,21,141,74,0,0 // vbroadcastss 0x4a8d(%rip),%ymm10 # 68ec <_sk_callback_avx+0x270> + .byte 196,98,125,24,21,201,74,0,0 // vbroadcastss 0x4ac9(%rip),%ymm10 # 6928 <_sk_callback_avx+0x26e> .byte 196,65,60,84,210 // vandps %ymm10,%ymm8,%ymm10 .byte 196,65,124,91,210 // vcvtdq2ps %ymm10,%ymm10 - .byte 196,98,125,24,29,126,74,0,0 // vbroadcastss 0x4a7e(%rip),%ymm11 # 68f0 <_sk_callback_avx+0x274> + .byte 196,98,125,24,29,186,74,0,0 // vbroadcastss 0x4aba(%rip),%ymm11 # 692c <_sk_callback_avx+0x272> .byte 196,65,44,89,211 // vmulps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,24,29,116,74,0,0 // vbroadcastss 0x4a74(%rip),%ymm11 # 68f4 <_sk_callback_avx+0x278> + .byte 196,98,125,24,29,176,74,0,0 // vbroadcastss 0x4ab0(%rip),%ymm11 # 6930 <_sk_callback_avx+0x276> .byte 196,65,60,84,195 // vandps %ymm11,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,29,101,74,0,0 // vbroadcastss 0x4a65(%rip),%ymm11 # 68f8 <_sk_callback_avx+0x27c> + .byte 196,98,125,24,29,161,74,0,0 // vbroadcastss 0x4aa1(%rip),%ymm11 # 6934 <_sk_callback_avx+0x27a> .byte 196,65,60,89,195 // vmulps %ymm11,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 @@ -17293,7 +17307,7 @@ _sk_load_tables_avx: .byte 77,133,192 // test %r8,%r8 .byte 15,133,248,1,0,0 // jne 218c <_sk_load_tables_avx+0x218> .byte 196,65,124,16,18 // vmovups (%r10),%ymm10 - .byte 197,124,40,13,95,77,0,0 // vmovaps 0x4d5f(%rip),%ymm9 # 6d00 <_sk_callback_avx+0x684> + .byte 197,124,40,13,159,77,0,0 // vmovaps 0x4d9f(%rip),%ymm9 # 6d40 <_sk_callback_avx+0x686> .byte 196,193,44,84,201 // vandps %ymm9,%ymm10,%ymm1 .byte 196,227,125,25,200,1 // vextractf128 $0x1,%ymm1,%xmm0 .byte 196,193,249,126,195 // vmovq %xmm0,%r11 @@ -17385,7 +17399,7 @@ _sk_load_tables_avx: .byte 196,193,65,114,208,24 // vpsrld $0x18,%xmm8,%xmm7 .byte 196,227,101,24,223,1 // vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,226,125,24,61,130,71,0,0 // vbroadcastss 0x4782(%rip),%ymm7 # 68fc <_sk_callback_avx+0x280> + .byte 196,226,125,24,61,190,71,0,0 // vbroadcastss 0x47be(%rip),%ymm7 # 6938 <_sk_callback_avx+0x27e> .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,201 // mov %r9,%rcx @@ -17399,9 +17413,9 @@ _sk_load_tables_avx: .byte 73,211,235 // shr %cl,%r11 .byte 196,193,249,110,195 // vmovq %r11,%xmm0 .byte 196,226,121,48,192 // vpmovzxbw %xmm0,%xmm0 - .byte 196,226,121,0,13,124,74,0,0 // vpshufb 0x4a7c(%rip),%xmm0,%xmm1 # 6c30 <_sk_callback_avx+0x5b4> + .byte 196,226,121,0,13,204,74,0,0 // vpshufb 0x4acc(%rip),%xmm0,%xmm1 # 6c80 <_sk_callback_avx+0x5c6> .byte 196,226,121,33,201 // vpmovsxbd %xmm1,%xmm1 - .byte 196,226,121,0,5,126,74,0,0 // vpshufb 0x4a7e(%rip),%xmm0,%xmm0 # 6c40 <_sk_callback_avx+0x5c4> + .byte 196,226,121,0,5,206,74,0,0 // vpshufb 0x4ace(%rip),%xmm0,%xmm0 # 6c90 <_sk_callback_avx+0x5d6> .byte 196,226,121,33,192 // vpmovsxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 196,66,125,44,18 // vmaskmovps (%r10),%ymm0,%ymm10 @@ -17431,7 +17445,7 @@ _sk_load_tables_u16_be_avx: .byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11 .byte 197,177,108,200 // vpunpcklqdq %xmm0,%xmm9,%xmm1 .byte 197,49,109,224 // vpunpckhqdq %xmm0,%xmm9,%xmm12 - .byte 197,121,111,21,16,74,0,0 // vmovdqa 0x4a10(%rip),%xmm10 # 6c50 <_sk_callback_avx+0x5d4> + .byte 197,121,111,21,96,74,0,0 // vmovdqa 0x4a60(%rip),%xmm10 # 6ca0 <_sk_callback_avx+0x5e6> .byte 196,193,113,219,202 // vpand %xmm10,%xmm1,%xmm1 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 196,193,113,105,209 // vpunpckhwd %xmm9,%xmm1,%xmm2 @@ -17529,7 +17543,7 @@ _sk_load_tables_u16_be_avx: .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3 .byte 196,227,101,24,223,1 // vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,226,125,24,61,199,68,0,0 // vbroadcastss 0x44c7(%rip),%ymm7 # 6900 <_sk_callback_avx+0x284> + .byte 196,226,125,24,61,3,69,0,0 // vbroadcastss 0x4503(%rip),%ymm7 # 693c <_sk_callback_avx+0x282> .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7 @@ -17592,7 +17606,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 197,185,108,218 // vpunpcklqdq %xmm2,%xmm8,%xmm3 .byte 197,57,109,218 // vpunpckhqdq %xmm2,%xmm8,%xmm11 .byte 197,121,108,193 // vpunpcklqdq %xmm1,%xmm0,%xmm8 - .byte 197,121,111,13,21,71,0,0 // vmovdqa 0x4715(%rip),%xmm9 # 6c60 <_sk_callback_avx+0x5e4> + .byte 197,121,111,13,101,71,0,0 // vmovdqa 0x4765(%rip),%xmm9 # 6cb0 <_sk_callback_avx+0x5f6> .byte 196,193,97,219,193 // vpand %xmm9,%xmm3,%xmm0 .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10 .byte 196,193,121,105,202 // vpunpckhwd %xmm10,%xmm0,%xmm1 @@ -17682,7 +17696,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 196,195,105,33,211,48 // vinsertps $0x30,%xmm11,%xmm2,%xmm2 .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,234,65,0,0 // vbroadcastss 0x41ea(%rip),%ymm3 # 6904 <_sk_callback_avx+0x288> + .byte 196,226,125,24,29,38,66,0,0 // vbroadcastss 0x4226(%rip),%ymm3 # 6940 <_sk_callback_avx+0x286> .byte 197,252,16,116,36,168 // vmovups -0x58(%rsp),%ymm6 .byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7 .byte 255,224 // jmpq *%rax @@ -17725,7 +17739,7 @@ HIDDEN _sk_byte_tables_avx FUNCTION(_sk_byte_tables_avx) _sk_byte_tables_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,38,65,0,0 // vbroadcastss 0x4126(%rip),%ymm8 # 6908 <_sk_callback_avx+0x28c> + .byte 196,98,125,24,5,98,65,0,0 // vbroadcastss 0x4162(%rip),%ymm8 # 6944 <_sk_callback_avx+0x28a> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 197,125,91,200 // vcvtps2dq %ymm0,%ymm9 .byte 196,65,249,126,201 // vmovq %xmm9,%r9 @@ -17844,7 +17858,7 @@ _sk_byte_tables_avx: .byte 196,194,121,49,204 // vpmovzxbd %xmm12,%xmm1 .byte 196,194,121,49,213 // vpmovzxbd %xmm13,%xmm2 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - .byte 196,98,125,24,13,208,62,0,0 // vbroadcastss 0x3ed0(%rip),%ymm9 # 690c <_sk_callback_avx+0x290> + .byte 196,98,125,24,13,12,63,0,0 // vbroadcastss 0x3f0c(%rip),%ymm9 # 6948 <_sk_callback_avx+0x28e> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 @@ -17960,7 +17974,7 @@ _sk_byte_tables_rgb_avx: .byte 196,194,121,49,203 // vpmovzxbd %xmm11,%xmm1 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,177,60,0,0 // vbroadcastss 0x3cb1(%rip),%ymm9 # 6910 <_sk_callback_avx+0x294> + .byte 196,98,125,24,13,237,60,0,0 // vbroadcastss 0x3ced(%rip),%ymm9 # 694c <_sk_callback_avx+0x292> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,194,121,49,202 // vpmovzxbd %xmm10,%xmm1 .byte 196,194,121,49,212 // vpmovzxbd %xmm12,%xmm2 @@ -18165,36 +18179,36 @@ _sk_parametric_r_avx: .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,216 // vcvtdq2ps %ymm0,%ymm11 - .byte 196,98,125,24,37,94,57,0,0 // vbroadcastss 0x395e(%rip),%ymm12 # 6914 <_sk_callback_avx+0x298> + .byte 196,98,125,24,37,154,57,0,0 // vbroadcastss 0x399a(%rip),%ymm12 # 6950 <_sk_callback_avx+0x296> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,84,57,0,0 // vbroadcastss 0x3954(%rip),%ymm12 # 6918 <_sk_callback_avx+0x29c> + .byte 196,98,125,24,37,144,57,0,0 // vbroadcastss 0x3990(%rip),%ymm12 # 6954 <_sk_callback_avx+0x29a> .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,74,57,0,0 // vbroadcastss 0x394a(%rip),%ymm12 # 691c <_sk_callback_avx+0x2a0> + .byte 196,98,125,24,37,134,57,0,0 // vbroadcastss 0x3986(%rip),%ymm12 # 6958 <_sk_callback_avx+0x29e> .byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,64,57,0,0 // vbroadcastss 0x3940(%rip),%ymm12 # 6920 <_sk_callback_avx+0x2a4> + .byte 196,98,125,24,37,124,57,0,0 // vbroadcastss 0x397c(%rip),%ymm12 # 695c <_sk_callback_avx+0x2a2> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,54,57,0,0 // vbroadcastss 0x3936(%rip),%ymm12 # 6924 <_sk_callback_avx+0x2a8> + .byte 196,98,125,24,37,114,57,0,0 // vbroadcastss 0x3972(%rip),%ymm12 # 6960 <_sk_callback_avx+0x2a6> .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,39,57,0,0 // vbroadcastss 0x3927(%rip),%ymm12 # 6928 <_sk_callback_avx+0x2ac> + .byte 196,98,125,24,37,99,57,0,0 // vbroadcastss 0x3963(%rip),%ymm12 # 6964 <_sk_callback_avx+0x2aa> .byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,29,57,0,0 // vbroadcastss 0x391d(%rip),%ymm12 # 692c <_sk_callback_avx+0x2b0> + .byte 196,98,125,24,37,89,57,0,0 // vbroadcastss 0x3959(%rip),%ymm12 # 6968 <_sk_callback_avx+0x2ae> .byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0 .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10 - .byte 196,98,125,24,29,1,57,0,0 // vbroadcastss 0x3901(%rip),%ymm11 # 6930 <_sk_callback_avx+0x2b4> + .byte 196,98,125,24,29,61,57,0,0 // vbroadcastss 0x393d(%rip),%ymm11 # 696c <_sk_callback_avx+0x2b2> .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,247,56,0,0 // vbroadcastss 0x38f7(%rip),%ymm11 # 6934 <_sk_callback_avx+0x2b8> + .byte 196,98,125,24,29,51,57,0,0 // vbroadcastss 0x3933(%rip),%ymm11 # 6970 <_sk_callback_avx+0x2b6> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,232,56,0,0 // vbroadcastss 0x38e8(%rip),%ymm11 # 6938 <_sk_callback_avx+0x2bc> + .byte 196,98,125,24,29,36,57,0,0 // vbroadcastss 0x3924(%rip),%ymm11 # 6974 <_sk_callback_avx+0x2ba> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,222,56,0,0 // vbroadcastss 0x38de(%rip),%ymm11 # 693c <_sk_callback_avx+0x2c0> + .byte 196,98,125,24,29,26,57,0,0 // vbroadcastss 0x391a(%rip),%ymm11 # 6978 <_sk_callback_avx+0x2be> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 - .byte 196,98,125,24,21,207,56,0,0 // vbroadcastss 0x38cf(%rip),%ymm10 # 6940 <_sk_callback_avx+0x2c4> + .byte 196,98,125,24,21,11,57,0,0 // vbroadcastss 0x390b(%rip),%ymm10 # 697c <_sk_callback_avx+0x2c2> .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 @@ -18202,7 +18216,7 @@ _sk_parametric_r_avx: .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,166,56,0,0 // vbroadcastss 0x38a6(%rip),%ymm8 # 6944 <_sk_callback_avx+0x2c8> + .byte 196,98,125,24,5,226,56,0,0 // vbroadcastss 0x38e2(%rip),%ymm8 # 6980 <_sk_callback_avx+0x2c6> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18224,36 +18238,36 @@ _sk_parametric_g_avx: .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,217 // vcvtdq2ps %ymm1,%ymm11 - .byte 196,98,125,24,37,87,56,0,0 // vbroadcastss 0x3857(%rip),%ymm12 # 6948 <_sk_callback_avx+0x2cc> + .byte 196,98,125,24,37,147,56,0,0 // vbroadcastss 0x3893(%rip),%ymm12 # 6984 <_sk_callback_avx+0x2ca> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,77,56,0,0 // vbroadcastss 0x384d(%rip),%ymm12 # 694c <_sk_callback_avx+0x2d0> + .byte 196,98,125,24,37,137,56,0,0 // vbroadcastss 0x3889(%rip),%ymm12 # 6988 <_sk_callback_avx+0x2ce> .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,67,56,0,0 // vbroadcastss 0x3843(%rip),%ymm12 # 6950 <_sk_callback_avx+0x2d4> + .byte 196,98,125,24,37,127,56,0,0 // vbroadcastss 0x387f(%rip),%ymm12 # 698c <_sk_callback_avx+0x2d2> .byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,57,56,0,0 // vbroadcastss 0x3839(%rip),%ymm12 # 6954 <_sk_callback_avx+0x2d8> + .byte 196,98,125,24,37,117,56,0,0 // vbroadcastss 0x3875(%rip),%ymm12 # 6990 <_sk_callback_avx+0x2d6> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,47,56,0,0 // vbroadcastss 0x382f(%rip),%ymm12 # 6958 <_sk_callback_avx+0x2dc> + .byte 196,98,125,24,37,107,56,0,0 // vbroadcastss 0x386b(%rip),%ymm12 # 6994 <_sk_callback_avx+0x2da> .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,32,56,0,0 // vbroadcastss 0x3820(%rip),%ymm12 # 695c <_sk_callback_avx+0x2e0> + .byte 196,98,125,24,37,92,56,0,0 // vbroadcastss 0x385c(%rip),%ymm12 # 6998 <_sk_callback_avx+0x2de> .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,22,56,0,0 // vbroadcastss 0x3816(%rip),%ymm12 # 6960 <_sk_callback_avx+0x2e4> + .byte 196,98,125,24,37,82,56,0,0 // vbroadcastss 0x3852(%rip),%ymm12 # 699c <_sk_callback_avx+0x2e2> .byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1 .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10 - .byte 196,98,125,24,29,250,55,0,0 // vbroadcastss 0x37fa(%rip),%ymm11 # 6964 <_sk_callback_avx+0x2e8> + .byte 196,98,125,24,29,54,56,0,0 // vbroadcastss 0x3836(%rip),%ymm11 # 69a0 <_sk_callback_avx+0x2e6> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,240,55,0,0 // vbroadcastss 0x37f0(%rip),%ymm11 # 6968 <_sk_callback_avx+0x2ec> + .byte 196,98,125,24,29,44,56,0,0 // vbroadcastss 0x382c(%rip),%ymm11 # 69a4 <_sk_callback_avx+0x2ea> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,225,55,0,0 // vbroadcastss 0x37e1(%rip),%ymm11 # 696c <_sk_callback_avx+0x2f0> + .byte 196,98,125,24,29,29,56,0,0 // vbroadcastss 0x381d(%rip),%ymm11 # 69a8 <_sk_callback_avx+0x2ee> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,215,55,0,0 // vbroadcastss 0x37d7(%rip),%ymm11 # 6970 <_sk_callback_avx+0x2f4> + .byte 196,98,125,24,29,19,56,0,0 // vbroadcastss 0x3813(%rip),%ymm11 # 69ac <_sk_callback_avx+0x2f2> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 - .byte 196,98,125,24,21,200,55,0,0 // vbroadcastss 0x37c8(%rip),%ymm10 # 6974 <_sk_callback_avx+0x2f8> + .byte 196,98,125,24,21,4,56,0,0 // vbroadcastss 0x3804(%rip),%ymm10 # 69b0 <_sk_callback_avx+0x2f6> .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 @@ -18261,7 +18275,7 @@ _sk_parametric_g_avx: .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,5,159,55,0,0 // vbroadcastss 0x379f(%rip),%ymm8 # 6978 <_sk_callback_avx+0x2fc> + .byte 196,98,125,24,5,219,55,0,0 // vbroadcastss 0x37db(%rip),%ymm8 # 69b4 <_sk_callback_avx+0x2fa> .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18283,36 +18297,36 @@ _sk_parametric_b_avx: .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,218 // vcvtdq2ps %ymm2,%ymm11 - .byte 196,98,125,24,37,80,55,0,0 // vbroadcastss 0x3750(%rip),%ymm12 # 697c <_sk_callback_avx+0x300> + .byte 196,98,125,24,37,140,55,0,0 // vbroadcastss 0x378c(%rip),%ymm12 # 69b8 <_sk_callback_avx+0x2fe> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,70,55,0,0 // vbroadcastss 0x3746(%rip),%ymm12 # 6980 <_sk_callback_avx+0x304> + .byte 196,98,125,24,37,130,55,0,0 // vbroadcastss 0x3782(%rip),%ymm12 # 69bc <_sk_callback_avx+0x302> .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,60,55,0,0 // vbroadcastss 0x373c(%rip),%ymm12 # 6984 <_sk_callback_avx+0x308> + .byte 196,98,125,24,37,120,55,0,0 // vbroadcastss 0x3778(%rip),%ymm12 # 69c0 <_sk_callback_avx+0x306> .byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,50,55,0,0 // vbroadcastss 0x3732(%rip),%ymm12 # 6988 <_sk_callback_avx+0x30c> + .byte 196,98,125,24,37,110,55,0,0 // vbroadcastss 0x376e(%rip),%ymm12 # 69c4 <_sk_callback_avx+0x30a> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,40,55,0,0 // vbroadcastss 0x3728(%rip),%ymm12 # 698c <_sk_callback_avx+0x310> + .byte 196,98,125,24,37,100,55,0,0 // vbroadcastss 0x3764(%rip),%ymm12 # 69c8 <_sk_callback_avx+0x30e> .byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,25,55,0,0 // vbroadcastss 0x3719(%rip),%ymm12 # 6990 <_sk_callback_avx+0x314> + .byte 196,98,125,24,37,85,55,0,0 // vbroadcastss 0x3755(%rip),%ymm12 # 69cc <_sk_callback_avx+0x312> .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,15,55,0,0 // vbroadcastss 0x370f(%rip),%ymm12 # 6994 <_sk_callback_avx+0x318> + .byte 196,98,125,24,37,75,55,0,0 // vbroadcastss 0x374b(%rip),%ymm12 # 69d0 <_sk_callback_avx+0x316> .byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2 .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10 - .byte 196,98,125,24,29,243,54,0,0 // vbroadcastss 0x36f3(%rip),%ymm11 # 6998 <_sk_callback_avx+0x31c> + .byte 196,98,125,24,29,47,55,0,0 // vbroadcastss 0x372f(%rip),%ymm11 # 69d4 <_sk_callback_avx+0x31a> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,233,54,0,0 // vbroadcastss 0x36e9(%rip),%ymm11 # 699c <_sk_callback_avx+0x320> + .byte 196,98,125,24,29,37,55,0,0 // vbroadcastss 0x3725(%rip),%ymm11 # 69d8 <_sk_callback_avx+0x31e> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,218,54,0,0 // vbroadcastss 0x36da(%rip),%ymm11 # 69a0 <_sk_callback_avx+0x324> + .byte 196,98,125,24,29,22,55,0,0 // vbroadcastss 0x3716(%rip),%ymm11 # 69dc <_sk_callback_avx+0x322> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,208,54,0,0 // vbroadcastss 0x36d0(%rip),%ymm11 # 69a4 <_sk_callback_avx+0x328> + .byte 196,98,125,24,29,12,55,0,0 // vbroadcastss 0x370c(%rip),%ymm11 # 69e0 <_sk_callback_avx+0x326> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 - .byte 196,98,125,24,21,193,54,0,0 // vbroadcastss 0x36c1(%rip),%ymm10 # 69a8 <_sk_callback_avx+0x32c> + .byte 196,98,125,24,21,253,54,0,0 // vbroadcastss 0x36fd(%rip),%ymm10 # 69e4 <_sk_callback_avx+0x32a> .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 @@ -18320,7 +18334,7 @@ _sk_parametric_b_avx: .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,152,54,0,0 // vbroadcastss 0x3698(%rip),%ymm8 # 69ac <_sk_callback_avx+0x330> + .byte 196,98,125,24,5,212,54,0,0 // vbroadcastss 0x36d4(%rip),%ymm8 # 69e8 <_sk_callback_avx+0x32e> .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18342,36 +18356,36 @@ _sk_parametric_a_avx: .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,219 // vcvtdq2ps %ymm3,%ymm11 - .byte 196,98,125,24,37,73,54,0,0 // vbroadcastss 0x3649(%rip),%ymm12 # 69b0 <_sk_callback_avx+0x334> + .byte 196,98,125,24,37,133,54,0,0 // vbroadcastss 0x3685(%rip),%ymm12 # 69ec <_sk_callback_avx+0x332> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,63,54,0,0 // vbroadcastss 0x363f(%rip),%ymm12 # 69b4 <_sk_callback_avx+0x338> + .byte 196,98,125,24,37,123,54,0,0 // vbroadcastss 0x367b(%rip),%ymm12 # 69f0 <_sk_callback_avx+0x336> .byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,53,54,0,0 // vbroadcastss 0x3635(%rip),%ymm12 # 69b8 <_sk_callback_avx+0x33c> + .byte 196,98,125,24,37,113,54,0,0 // vbroadcastss 0x3671(%rip),%ymm12 # 69f4 <_sk_callback_avx+0x33a> .byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,43,54,0,0 // vbroadcastss 0x362b(%rip),%ymm12 # 69bc <_sk_callback_avx+0x340> + .byte 196,98,125,24,37,103,54,0,0 // vbroadcastss 0x3667(%rip),%ymm12 # 69f8 <_sk_callback_avx+0x33e> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,33,54,0,0 // vbroadcastss 0x3621(%rip),%ymm12 # 69c0 <_sk_callback_avx+0x344> + .byte 196,98,125,24,37,93,54,0,0 // vbroadcastss 0x365d(%rip),%ymm12 # 69fc <_sk_callback_avx+0x342> .byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,18,54,0,0 // vbroadcastss 0x3612(%rip),%ymm12 # 69c4 <_sk_callback_avx+0x348> + .byte 196,98,125,24,37,78,54,0,0 // vbroadcastss 0x364e(%rip),%ymm12 # 6a00 <_sk_callback_avx+0x346> .byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,8,54,0,0 // vbroadcastss 0x3608(%rip),%ymm12 # 69c8 <_sk_callback_avx+0x34c> + .byte 196,98,125,24,37,68,54,0,0 // vbroadcastss 0x3644(%rip),%ymm12 # 6a04 <_sk_callback_avx+0x34a> .byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3 .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10 - .byte 196,98,125,24,29,236,53,0,0 // vbroadcastss 0x35ec(%rip),%ymm11 # 69cc <_sk_callback_avx+0x350> + .byte 196,98,125,24,29,40,54,0,0 // vbroadcastss 0x3628(%rip),%ymm11 # 6a08 <_sk_callback_avx+0x34e> .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,226,53,0,0 // vbroadcastss 0x35e2(%rip),%ymm11 # 69d0 <_sk_callback_avx+0x354> + .byte 196,98,125,24,29,30,54,0,0 // vbroadcastss 0x361e(%rip),%ymm11 # 6a0c <_sk_callback_avx+0x352> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,211,53,0,0 // vbroadcastss 0x35d3(%rip),%ymm11 # 69d4 <_sk_callback_avx+0x358> + .byte 196,98,125,24,29,15,54,0,0 // vbroadcastss 0x360f(%rip),%ymm11 # 6a10 <_sk_callback_avx+0x356> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,201,53,0,0 // vbroadcastss 0x35c9(%rip),%ymm11 # 69d8 <_sk_callback_avx+0x35c> + .byte 196,98,125,24,29,5,54,0,0 // vbroadcastss 0x3605(%rip),%ymm11 # 6a14 <_sk_callback_avx+0x35a> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 - .byte 196,98,125,24,21,186,53,0,0 // vbroadcastss 0x35ba(%rip),%ymm10 # 69dc <_sk_callback_avx+0x360> + .byte 196,98,125,24,21,246,53,0,0 // vbroadcastss 0x35f6(%rip),%ymm10 # 6a18 <_sk_callback_avx+0x35e> .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 @@ -18379,7 +18393,7 @@ _sk_parametric_a_avx: .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 - .byte 196,98,125,24,5,145,53,0,0 // vbroadcastss 0x3591(%rip),%ymm8 # 69e0 <_sk_callback_avx+0x364> + .byte 196,98,125,24,5,205,53,0,0 // vbroadcastss 0x35cd(%rip),%ymm8 # 6a1c <_sk_callback_avx+0x362> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18388,31 +18402,31 @@ HIDDEN _sk_lab_to_xyz_avx .globl _sk_lab_to_xyz_avx FUNCTION(_sk_lab_to_xyz_avx) _sk_lab_to_xyz_avx: - .byte 196,98,125,24,5,131,53,0,0 // vbroadcastss 0x3583(%rip),%ymm8 # 69e4 <_sk_callback_avx+0x368> + .byte 196,98,125,24,5,191,53,0,0 // vbroadcastss 0x35bf(%rip),%ymm8 # 6a20 <_sk_callback_avx+0x366> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,121,53,0,0 // vbroadcastss 0x3579(%rip),%ymm8 # 69e8 <_sk_callback_avx+0x36c> + .byte 196,98,125,24,5,181,53,0,0 // vbroadcastss 0x35b5(%rip),%ymm8 # 6a24 <_sk_callback_avx+0x36a> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,13,111,53,0,0 // vbroadcastss 0x356f(%rip),%ymm9 # 69ec <_sk_callback_avx+0x370> + .byte 196,98,125,24,13,171,53,0,0 // vbroadcastss 0x35ab(%rip),%ymm9 # 6a28 <_sk_callback_avx+0x36e> .byte 196,193,116,88,201 // vaddps %ymm9,%ymm1,%ymm1 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 - .byte 196,98,125,24,5,91,53,0,0 // vbroadcastss 0x355b(%rip),%ymm8 # 69f0 <_sk_callback_avx+0x374> + .byte 196,98,125,24,5,151,53,0,0 // vbroadcastss 0x3597(%rip),%ymm8 # 6a2c <_sk_callback_avx+0x372> .byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,81,53,0,0 // vbroadcastss 0x3551(%rip),%ymm8 # 69f4 <_sk_callback_avx+0x378> + .byte 196,98,125,24,5,141,53,0,0 // vbroadcastss 0x358d(%rip),%ymm8 # 6a30 <_sk_callback_avx+0x376> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,71,53,0,0 // vbroadcastss 0x3547(%rip),%ymm8 # 69f8 <_sk_callback_avx+0x37c> + .byte 196,98,125,24,5,131,53,0,0 // vbroadcastss 0x3583(%rip),%ymm8 # 6a34 <_sk_callback_avx+0x37a> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1 - .byte 196,98,125,24,5,57,53,0,0 // vbroadcastss 0x3539(%rip),%ymm8 # 69fc <_sk_callback_avx+0x380> + .byte 196,98,125,24,5,117,53,0,0 // vbroadcastss 0x3575(%rip),%ymm8 # 6a38 <_sk_callback_avx+0x37e> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 197,252,92,210 // vsubps %ymm2,%ymm0,%ymm2 .byte 197,116,89,193 // vmulps %ymm1,%ymm1,%ymm8 .byte 196,65,116,89,192 // vmulps %ymm8,%ymm1,%ymm8 - .byte 196,98,125,24,13,34,53,0,0 // vbroadcastss 0x3522(%rip),%ymm9 # 6a00 <_sk_callback_avx+0x384> + .byte 196,98,125,24,13,94,53,0,0 // vbroadcastss 0x355e(%rip),%ymm9 # 6a3c <_sk_callback_avx+0x382> .byte 196,65,52,194,208,1 // vcmpltps %ymm8,%ymm9,%ymm10 - .byte 196,98,125,24,29,23,53,0,0 // vbroadcastss 0x3517(%rip),%ymm11 # 6a04 <_sk_callback_avx+0x388> + .byte 196,98,125,24,29,83,53,0,0 // vbroadcastss 0x3553(%rip),%ymm11 # 6a40 <_sk_callback_avx+0x386> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,37,13,53,0,0 // vbroadcastss 0x350d(%rip),%ymm12 # 6a08 <_sk_callback_avx+0x38c> + .byte 196,98,125,24,37,73,53,0,0 // vbroadcastss 0x3549(%rip),%ymm12 # 6a44 <_sk_callback_avx+0x38a> .byte 196,193,116,89,204 // vmulps %ymm12,%ymm1,%ymm1 .byte 196,67,117,74,192,160 // vblendvps %ymm10,%ymm8,%ymm1,%ymm8 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1 @@ -18427,9 +18441,9 @@ _sk_lab_to_xyz_avx: .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,193,108,89,212 // vmulps %ymm12,%ymm2,%ymm2 .byte 196,227,109,74,208,144 // vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - .byte 196,226,125,24,5,195,52,0,0 // vbroadcastss 0x34c3(%rip),%ymm0 # 6a0c <_sk_callback_avx+0x390> + .byte 196,226,125,24,5,255,52,0,0 // vbroadcastss 0x34ff(%rip),%ymm0 # 6a48 <_sk_callback_avx+0x38e> .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,186,52,0,0 // vbroadcastss 0x34ba(%rip),%ymm8 # 6a10 <_sk_callback_avx+0x394> + .byte 196,98,125,24,5,246,52,0,0 // vbroadcastss 0x34f6(%rip),%ymm8 # 6a4c <_sk_callback_avx+0x392> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18443,13 +18457,13 @@ _sk_load_a8_avx: .byte 77,133,192 // test %r8,%r8 .byte 117,66 // jne 35ab <_sk_load_a8_avx+0x4c> .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0 - .byte 197,249,219,5,249,54,0,0 // vpand 0x36f9(%rip),%xmm0,%xmm0 # 6c70 <_sk_callback_avx+0x5f4> + .byte 197,249,219,5,73,55,0,0 // vpand 0x3749(%rip),%xmm0,%xmm0 # 6cc0 <_sk_callback_avx+0x606> .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,125,52,0,0 // vbroadcastss 0x347d(%rip),%ymm1 # 6a14 <_sk_callback_avx+0x398> + .byte 196,226,125,24,13,185,52,0,0 // vbroadcastss 0x34b9(%rip),%ymm1 # 6a50 <_sk_callback_avx+0x396> .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 @@ -18548,7 +18562,7 @@ _sk_gather_a8_avx: .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,8,51,0,0 // vbroadcastss 0x3308(%rip),%ymm1 # 6a18 <_sk_callback_avx+0x39c> + .byte 196,226,125,24,13,68,51,0,0 // vbroadcastss 0x3344(%rip),%ymm1 # 6a54 <_sk_callback_avx+0x39a> .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 @@ -18562,7 +18576,7 @@ FUNCTION(_sk_store_a8_avx) _sk_store_a8_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 - .byte 196,98,125,24,5,234,50,0,0 // vbroadcastss 0x32ea(%rip),%ymm8 # 6a1c <_sk_callback_avx+0x3a0> + .byte 196,98,125,24,5,38,51,0,0 // vbroadcastss 0x3326(%rip),%ymm8 # 6a58 <_sk_callback_avx+0x39e> .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 @@ -18626,16 +18640,16 @@ _sk_load_g8_avx: .byte 77,133,192 // test %r8,%r8 .byte 117,71 // jne 382d <_sk_load_g8_avx+0x51> .byte 196,194,121,48,4,19 // vpmovzxbw (%r11,%rdx,1),%xmm0 - .byte 197,249,219,5,140,52,0,0 // vpand 0x348c(%rip),%xmm0,%xmm0 # 6c80 <_sk_callback_avx+0x604> + .byte 197,249,219,5,220,52,0,0 // vpand 0x34dc(%rip),%xmm0,%xmm0 # 6cd0 <_sk_callback_avx+0x616> .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,12,50,0,0 // vbroadcastss 0x320c(%rip),%ymm1 # 6a20 <_sk_callback_avx+0x3a4> + .byte 196,226,125,24,13,72,50,0,0 // vbroadcastss 0x3248(%rip),%ymm1 # 6a5c <_sk_callback_avx+0x3a2> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,1,50,0,0 // vbroadcastss 0x3201(%rip),%ymm3 # 6a24 <_sk_callback_avx+0x3a8> + .byte 196,226,125,24,29,61,50,0,0 // vbroadcastss 0x323d(%rip),%ymm3 # 6a60 <_sk_callback_avx+0x3a6> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 255,224 // jmpq *%rax @@ -18732,10 +18746,10 @@ _sk_gather_g8_avx: .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,148,48,0,0 // vbroadcastss 0x3094(%rip),%ymm1 # 6a28 <_sk_callback_avx+0x3ac> + .byte 196,226,125,24,13,208,48,0,0 // vbroadcastss 0x30d0(%rip),%ymm1 # 6a64 <_sk_callback_avx+0x3aa> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,137,48,0,0 // vbroadcastss 0x3089(%rip),%ymm3 # 6a2c <_sk_callback_avx+0x3b0> + .byte 196,226,125,24,29,197,48,0,0 // vbroadcastss 0x30c5(%rip),%ymm3 # 6a68 <_sk_callback_avx+0x3ae> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 255,224 // jmpq *%rax @@ -18807,10 +18821,10 @@ _sk_gather_i8_avx: .byte 72,193,235,30 // shr $0x1e,%rbx .byte 196,195,121,34,28,27,3 // vpinsrd $0x3,(%r11,%rbx,1),%xmm0,%xmm3 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - .byte 197,124,40,21,68,50,0,0 // vmovaps 0x3244(%rip),%ymm10 # 6d20 <_sk_callback_avx+0x6a4> + .byte 197,124,40,21,132,50,0,0 // vmovaps 0x3284(%rip),%ymm10 # 6d60 <_sk_callback_avx+0x6a6> .byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,66,47,0,0 // vbroadcastss 0x2f42(%rip),%ymm9 # 6a30 <_sk_callback_avx+0x3b4> + .byte 196,98,125,24,13,126,47,0,0 // vbroadcastss 0x2f7e(%rip),%ymm9 # 6a6c <_sk_callback_avx+0x3b2> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 @@ -18846,23 +18860,23 @@ _sk_load_565_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - .byte 196,226,125,24,5,180,46,0,0 // vbroadcastss 0x2eb4(%rip),%ymm0 # 6a34 <_sk_callback_avx+0x3b8> + .byte 196,226,125,24,5,240,46,0,0 // vbroadcastss 0x2ef0(%rip),%ymm0 # 6a70 <_sk_callback_avx+0x3b6> .byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,167,46,0,0 // vbroadcastss 0x2ea7(%rip),%ymm1 # 6a38 <_sk_callback_avx+0x3bc> + .byte 196,226,125,24,13,227,46,0,0 // vbroadcastss 0x2ee3(%rip),%ymm1 # 6a74 <_sk_callback_avx+0x3ba> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,158,46,0,0 // vbroadcastss 0x2e9e(%rip),%ymm1 # 6a3c <_sk_callback_avx+0x3c0> + .byte 196,226,125,24,13,218,46,0,0 // vbroadcastss 0x2eda(%rip),%ymm1 # 6a78 <_sk_callback_avx+0x3be> .byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,145,46,0,0 // vbroadcastss 0x2e91(%rip),%ymm3 # 6a40 <_sk_callback_avx+0x3c4> + .byte 196,226,125,24,29,205,46,0,0 // vbroadcastss 0x2ecd(%rip),%ymm3 # 6a7c <_sk_callback_avx+0x3c2> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,24,29,136,46,0,0 // vbroadcastss 0x2e88(%rip),%ymm3 # 6a44 <_sk_callback_avx+0x3c8> + .byte 196,226,125,24,29,196,46,0,0 // vbroadcastss 0x2ec4(%rip),%ymm3 # 6a80 <_sk_callback_avx+0x3c6> .byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,123,46,0,0 // vbroadcastss 0x2e7b(%rip),%ymm3 # 6a48 <_sk_callback_avx+0x3cc> + .byte 196,226,125,24,29,183,46,0,0 // vbroadcastss 0x2eb7(%rip),%ymm3 # 6a84 <_sk_callback_avx+0x3ca> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,112,46,0,0 // vbroadcastss 0x2e70(%rip),%ymm3 # 6a4c <_sk_callback_avx+0x3d0> + .byte 196,226,125,24,29,172,46,0,0 // vbroadcastss 0x2eac(%rip),%ymm3 # 6a88 <_sk_callback_avx+0x3ce> .byte 255,224 // jmpq *%rax .byte 69,137,193 // mov %r8d,%r9d .byte 65,128,225,7 // and $0x7,%r9b @@ -18957,23 +18971,23 @@ _sk_gather_565_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - .byte 196,226,125,24,5,22,45,0,0 // vbroadcastss 0x2d16(%rip),%ymm0 # 6a50 <_sk_callback_avx+0x3d4> + .byte 196,226,125,24,5,82,45,0,0 // vbroadcastss 0x2d52(%rip),%ymm0 # 6a8c <_sk_callback_avx+0x3d2> .byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,9,45,0,0 // vbroadcastss 0x2d09(%rip),%ymm1 # 6a54 <_sk_callback_avx+0x3d8> + .byte 196,226,125,24,13,69,45,0,0 // vbroadcastss 0x2d45(%rip),%ymm1 # 6a90 <_sk_callback_avx+0x3d6> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,0,45,0,0 // vbroadcastss 0x2d00(%rip),%ymm1 # 6a58 <_sk_callback_avx+0x3dc> + .byte 196,226,125,24,13,60,45,0,0 // vbroadcastss 0x2d3c(%rip),%ymm1 # 6a94 <_sk_callback_avx+0x3da> .byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,243,44,0,0 // vbroadcastss 0x2cf3(%rip),%ymm3 # 6a5c <_sk_callback_avx+0x3e0> + .byte 196,226,125,24,29,47,45,0,0 // vbroadcastss 0x2d2f(%rip),%ymm3 # 6a98 <_sk_callback_avx+0x3de> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,24,29,234,44,0,0 // vbroadcastss 0x2cea(%rip),%ymm3 # 6a60 <_sk_callback_avx+0x3e4> + .byte 196,226,125,24,29,38,45,0,0 // vbroadcastss 0x2d26(%rip),%ymm3 # 6a9c <_sk_callback_avx+0x3e2> .byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,221,44,0,0 // vbroadcastss 0x2cdd(%rip),%ymm3 # 6a64 <_sk_callback_avx+0x3e8> + .byte 196,226,125,24,29,25,45,0,0 // vbroadcastss 0x2d19(%rip),%ymm3 # 6aa0 <_sk_callback_avx+0x3e6> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,210,44,0,0 // vbroadcastss 0x2cd2(%rip),%ymm3 # 6a68 <_sk_callback_avx+0x3ec> + .byte 196,226,125,24,29,14,45,0,0 // vbroadcastss 0x2d0e(%rip),%ymm3 # 6aa4 <_sk_callback_avx+0x3ea> .byte 255,224 // jmpq *%rax HIDDEN _sk_store_565_avx @@ -18982,14 +18996,14 @@ FUNCTION(_sk_store_565_avx) _sk_store_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 - .byte 196,98,125,24,5,198,44,0,0 // vbroadcastss 0x2cc6(%rip),%ymm8 # 6a6c <_sk_callback_avx+0x3f0> + .byte 196,98,125,24,5,2,45,0,0 // vbroadcastss 0x2d02(%rip),%ymm8 # 6aa8 <_sk_callback_avx+0x3ee> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 .byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - .byte 196,98,125,24,21,159,44,0,0 // vbroadcastss 0x2c9f(%rip),%ymm10 # 6a70 <_sk_callback_avx+0x3f4> + .byte 196,98,125,24,21,219,44,0,0 // vbroadcastss 0x2cdb(%rip),%ymm10 # 6aac <_sk_callback_avx+0x3f2> .byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 .byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11 @@ -19063,25 +19077,25 @@ _sk_load_4444_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - .byte 196,226,125,24,5,168,43,0,0 // vbroadcastss 0x2ba8(%rip),%ymm0 # 6a74 <_sk_callback_avx+0x3f8> + .byte 196,226,125,24,5,228,43,0,0 // vbroadcastss 0x2be4(%rip),%ymm0 # 6ab0 <_sk_callback_avx+0x3f6> .byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,155,43,0,0 // vbroadcastss 0x2b9b(%rip),%ymm1 # 6a78 <_sk_callback_avx+0x3fc> + .byte 196,226,125,24,13,215,43,0,0 // vbroadcastss 0x2bd7(%rip),%ymm1 # 6ab4 <_sk_callback_avx+0x3fa> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,146,43,0,0 // vbroadcastss 0x2b92(%rip),%ymm1 # 6a7c <_sk_callback_avx+0x400> + .byte 196,226,125,24,13,206,43,0,0 // vbroadcastss 0x2bce(%rip),%ymm1 # 6ab8 <_sk_callback_avx+0x3fe> .byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,133,43,0,0 // vbroadcastss 0x2b85(%rip),%ymm2 # 6a80 <_sk_callback_avx+0x404> + .byte 196,226,125,24,21,193,43,0,0 // vbroadcastss 0x2bc1(%rip),%ymm2 # 6abc <_sk_callback_avx+0x402> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,124,43,0,0 // vbroadcastss 0x2b7c(%rip),%ymm2 # 6a84 <_sk_callback_avx+0x408> + .byte 196,226,125,24,21,184,43,0,0 // vbroadcastss 0x2bb8(%rip),%ymm2 # 6ac0 <_sk_callback_avx+0x406> .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,111,43,0,0 // vbroadcastss 0x2b6f(%rip),%ymm8 # 6a88 <_sk_callback_avx+0x40c> + .byte 196,98,125,24,5,171,43,0,0 // vbroadcastss 0x2bab(%rip),%ymm8 # 6ac4 <_sk_callback_avx+0x40a> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,101,43,0,0 // vbroadcastss 0x2b65(%rip),%ymm8 # 6a8c <_sk_callback_avx+0x410> + .byte 196,98,125,24,5,161,43,0,0 // vbroadcastss 0x2ba1(%rip),%ymm8 # 6ac8 <_sk_callback_avx+0x40e> .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,87,43,0,0 // vbroadcastss 0x2b57(%rip),%ymm8 # 6a90 <_sk_callback_avx+0x414> + .byte 196,98,125,24,5,147,43,0,0 // vbroadcastss 0x2b93(%rip),%ymm8 # 6acc <_sk_callback_avx+0x412> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -19178,25 +19192,25 @@ _sk_gather_4444_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - .byte 196,226,125,24,5,246,41,0,0 // vbroadcastss 0x29f6(%rip),%ymm0 # 6a94 <_sk_callback_avx+0x418> + .byte 196,226,125,24,5,50,42,0,0 // vbroadcastss 0x2a32(%rip),%ymm0 # 6ad0 <_sk_callback_avx+0x416> .byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,233,41,0,0 // vbroadcastss 0x29e9(%rip),%ymm1 # 6a98 <_sk_callback_avx+0x41c> + .byte 196,226,125,24,13,37,42,0,0 // vbroadcastss 0x2a25(%rip),%ymm1 # 6ad4 <_sk_callback_avx+0x41a> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,224,41,0,0 // vbroadcastss 0x29e0(%rip),%ymm1 # 6a9c <_sk_callback_avx+0x420> + .byte 196,226,125,24,13,28,42,0,0 // vbroadcastss 0x2a1c(%rip),%ymm1 # 6ad8 <_sk_callback_avx+0x41e> .byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,211,41,0,0 // vbroadcastss 0x29d3(%rip),%ymm2 # 6aa0 <_sk_callback_avx+0x424> + .byte 196,226,125,24,21,15,42,0,0 // vbroadcastss 0x2a0f(%rip),%ymm2 # 6adc <_sk_callback_avx+0x422> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,202,41,0,0 // vbroadcastss 0x29ca(%rip),%ymm2 # 6aa4 <_sk_callback_avx+0x428> + .byte 196,226,125,24,21,6,42,0,0 // vbroadcastss 0x2a06(%rip),%ymm2 # 6ae0 <_sk_callback_avx+0x426> .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,189,41,0,0 // vbroadcastss 0x29bd(%rip),%ymm8 # 6aa8 <_sk_callback_avx+0x42c> + .byte 196,98,125,24,5,249,41,0,0 // vbroadcastss 0x29f9(%rip),%ymm8 # 6ae4 <_sk_callback_avx+0x42a> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,179,41,0,0 // vbroadcastss 0x29b3(%rip),%ymm8 # 6aac <_sk_callback_avx+0x430> + .byte 196,98,125,24,5,239,41,0,0 // vbroadcastss 0x29ef(%rip),%ymm8 # 6ae8 <_sk_callback_avx+0x42e> .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,165,41,0,0 // vbroadcastss 0x29a5(%rip),%ymm8 # 6ab0 <_sk_callback_avx+0x434> + .byte 196,98,125,24,5,225,41,0,0 // vbroadcastss 0x29e1(%rip),%ymm8 # 6aec <_sk_callback_avx+0x432> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -19207,7 +19221,7 @@ FUNCTION(_sk_store_4444_avx) _sk_store_4444_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 - .byte 196,98,125,24,5,146,41,0,0 // vbroadcastss 0x2992(%rip),%ymm8 # 6ab4 <_sk_callback_avx+0x438> + .byte 196,98,125,24,5,206,41,0,0 // vbroadcastss 0x29ce(%rip),%ymm8 # 6af0 <_sk_callback_avx+0x436> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10 @@ -19290,10 +19304,10 @@ _sk_load_8888_avx: .byte 77,133,192 // test %r8,%r8 .byte 15,133,137,0,0,0 // jne 42d6 <_sk_load_8888_avx+0xa2> .byte 196,193,124,16,26 // vmovups (%r10),%ymm3 - .byte 197,124,40,21,230,42,0,0 // vmovaps 0x2ae6(%rip),%ymm10 # 6d40 <_sk_callback_avx+0x6c4> + .byte 197,124,40,21,38,43,0,0 // vmovaps 0x2b26(%rip),%ymm10 # 6d80 <_sk_callback_avx+0x6c6> .byte 196,193,100,84,194 // vandps %ymm10,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,76,40,0,0 // vbroadcastss 0x284c(%rip),%ymm8 # 6ab8 <_sk_callback_avx+0x43c> + .byte 196,98,125,24,5,136,40,0,0 // vbroadcastss 0x2888(%rip),%ymm8 # 6af4 <_sk_callback_avx+0x43a> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1 .byte 196,195,125,25,217,1 // vextractf128 $0x1,%ymm3,%xmm9 @@ -19323,9 +19337,9 @@ _sk_load_8888_avx: .byte 72,211,232 // shr %cl,%rax .byte 196,225,249,110,192 // vmovq %rax,%xmm0 .byte 196,226,121,48,192 // vpmovzxbw %xmm0,%xmm0 - .byte 196,226,121,0,13,146,41,0,0 // vpshufb 0x2992(%rip),%xmm0,%xmm1 # 6c90 <_sk_callback_avx+0x614> + .byte 196,226,121,0,13,226,41,0,0 // vpshufb 0x29e2(%rip),%xmm0,%xmm1 # 6ce0 <_sk_callback_avx+0x626> .byte 196,226,121,33,201 // vpmovsxbd %xmm1,%xmm1 - .byte 196,226,121,0,5,148,41,0,0 // vpshufb 0x2994(%rip),%xmm0,%xmm0 # 6ca0 <_sk_callback_avx+0x624> + .byte 196,226,121,0,5,228,41,0,0 // vpshufb 0x29e4(%rip),%xmm0,%xmm0 # 6cf0 <_sk_callback_avx+0x636> .byte 196,226,121,33,192 // vpmovsxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 196,194,125,44,26 // vmaskmovps (%r10),%ymm0,%ymm3 @@ -19368,10 +19382,10 @@ _sk_gather_8888_avx: .byte 73,193,234,32 // shr $0x20,%r10 .byte 196,131,121,34,28,145,3 // vpinsrd $0x3,(%r9,%r10,4),%xmm0,%xmm3 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - .byte 197,124,40,21,148,41,0,0 // vmovaps 0x2994(%rip),%ymm10 # 6d60 <_sk_callback_avx+0x6e4> + .byte 197,124,40,21,212,41,0,0 // vmovaps 0x29d4(%rip),%ymm10 # 6da0 <_sk_callback_avx+0x6e6> .byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,222,38,0,0 // vbroadcastss 0x26de(%rip),%ymm9 # 6abc <_sk_callback_avx+0x440> + .byte 196,98,125,24,13,26,39,0,0 // vbroadcastss 0x271a(%rip),%ymm9 # 6af8 <_sk_callback_avx+0x43e> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 @@ -19401,7 +19415,7 @@ _sk_store_8888_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,141,20,149,0,0,0,0 // lea 0x0(,%rdx,4),%r10 .byte 76,3,16 // add (%rax),%r10 - .byte 196,98,125,24,5,104,38,0,0 // vbroadcastss 0x2668(%rip),%ymm8 # 6ac0 <_sk_callback_avx+0x444> + .byte 196,98,125,24,5,164,38,0,0 // vbroadcastss 0x26a4(%rip),%ymm8 # 6afc <_sk_callback_avx+0x442> .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 @@ -19438,9 +19452,9 @@ _sk_store_8888_avx: .byte 72,211,232 // shr %cl,%rax .byte 196,97,249,110,200 // vmovq %rax,%xmm9 .byte 196,66,121,48,201 // vpmovzxbw %xmm9,%xmm9 - .byte 196,98,49,0,21,160,39,0,0 // vpshufb 0x27a0(%rip),%xmm9,%xmm10 # 6cb0 <_sk_callback_avx+0x634> + .byte 196,98,49,0,21,240,39,0,0 // vpshufb 0x27f0(%rip),%xmm9,%xmm10 # 6d00 <_sk_callback_avx+0x646> .byte 196,66,121,33,210 // vpmovsxbd %xmm10,%xmm10 - .byte 196,98,49,0,13,162,39,0,0 // vpshufb 0x27a2(%rip),%xmm9,%xmm9 # 6cc0 <_sk_callback_avx+0x644> + .byte 196,98,49,0,13,242,39,0,0 // vpshufb 0x27f2(%rip),%xmm9,%xmm9 # 6d10 <_sk_callback_avx+0x656> .byte 196,66,121,33,201 // vpmovsxbd %xmm9,%xmm9 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 .byte 196,66,53,46,2 // vmaskmovps %ymm8,%ymm9,(%r10) @@ -19476,13 +19490,13 @@ _sk_load_f16_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - .byte 196,98,125,24,37,17,37,0,0 // vbroadcastss 0x2511(%rip),%ymm12 # 6ac4 <_sk_callback_avx+0x448> + .byte 196,98,125,24,37,77,37,0,0 // vbroadcastss 0x254d(%rip),%ymm12 # 6b00 <_sk_callback_avx+0x446> .byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1 .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 - .byte 196,98,121,24,29,253,36,0,0 // vbroadcastss 0x24fd(%rip),%xmm11 # 6ac8 <_sk_callback_avx+0x44c> + .byte 196,98,121,24,29,57,37,0,0 // vbroadcastss 0x2539(%rip),%xmm11 # 6b04 <_sk_callback_avx+0x44a> .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 - .byte 196,98,121,24,45,243,36,0,0 // vbroadcastss 0x24f3(%rip),%xmm13 # 6acc <_sk_callback_avx+0x450> + .byte 196,98,121,24,45,47,37,0,0 // vbroadcastss 0x252f(%rip),%xmm13 # 6b08 <_sk_callback_avx+0x44e> .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -19496,7 +19510,7 @@ _sk_load_f16_avx: .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 .byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0 .byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1 - .byte 196,226,121,24,29,169,36,0,0 // vbroadcastss 0x24a9(%rip),%xmm3 # 6ad0 <_sk_callback_avx+0x454> + .byte 196,226,121,24,29,229,36,0,0 // vbroadcastss 0x24e5(%rip),%xmm3 # 6b0c <_sk_callback_avx+0x452> .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -19672,13 +19686,13 @@ _sk_gather_f16_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - .byte 196,98,125,24,37,110,33,0,0 // vbroadcastss 0x216e(%rip),%ymm12 # 6ad4 <_sk_callback_avx+0x458> + .byte 196,98,125,24,37,170,33,0,0 // vbroadcastss 0x21aa(%rip),%ymm12 # 6b10 <_sk_callback_avx+0x456> .byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1 .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 - .byte 196,98,121,24,29,90,33,0,0 // vbroadcastss 0x215a(%rip),%xmm11 # 6ad8 <_sk_callback_avx+0x45c> + .byte 196,98,121,24,29,150,33,0,0 // vbroadcastss 0x2196(%rip),%xmm11 # 6b14 <_sk_callback_avx+0x45a> .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 - .byte 196,98,121,24,45,80,33,0,0 // vbroadcastss 0x2150(%rip),%xmm13 # 6adc <_sk_callback_avx+0x460> + .byte 196,98,121,24,45,140,33,0,0 // vbroadcastss 0x218c(%rip),%xmm13 # 6b18 <_sk_callback_avx+0x45e> .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -19692,7 +19706,7 @@ _sk_gather_f16_avx: .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 .byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0 .byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1 - .byte 196,226,121,24,29,6,33,0,0 // vbroadcastss 0x2106(%rip),%xmm3 # 6ae0 <_sk_callback_avx+0x464> + .byte 196,226,121,24,29,66,33,0,0 // vbroadcastss 0x2142(%rip),%xmm3 # 6b1c <_sk_callback_avx+0x462> .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -19792,12 +19806,12 @@ _sk_store_f16_avx: .byte 197,252,17,52,36 // vmovups %ymm6,(%rsp) .byte 197,252,17,108,36,224 // vmovups %ymm5,-0x20(%rsp) .byte 197,252,17,100,36,192 // vmovups %ymm4,-0x40(%rsp) - .byte 196,98,125,24,13,38,31,0,0 // vbroadcastss 0x1f26(%rip),%ymm9 # 6ae4 <_sk_callback_avx+0x468> + .byte 196,98,125,24,13,98,31,0,0 // vbroadcastss 0x1f62(%rip),%ymm9 # 6b20 <_sk_callback_avx+0x466> .byte 196,65,124,84,209 // vandps %ymm9,%ymm0,%ymm10 .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp) .byte 196,65,124,87,218 // vxorps %ymm10,%ymm0,%ymm11 .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 - .byte 196,98,121,24,5,11,31,0,0 // vbroadcastss 0x1f0b(%rip),%xmm8 # 6ae8 <_sk_callback_avx+0x46c> + .byte 196,98,121,24,5,71,31,0,0 // vbroadcastss 0x1f47(%rip),%xmm8 # 6b24 <_sk_callback_avx+0x46a> .byte 196,65,57,102,236 // vpcmpgtd %xmm12,%xmm8,%xmm13 .byte 196,65,57,102,243 // vpcmpgtd %xmm11,%xmm8,%xmm14 .byte 196,67,13,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 @@ -19807,7 +19821,7 @@ _sk_store_f16_avx: .byte 196,67,13,24,242,1 // vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11 .byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12 - .byte 196,98,125,24,21,210,30,0,0 // vbroadcastss 0x1ed2(%rip),%ymm10 # 6aec <_sk_callback_avx+0x470> + .byte 196,98,125,24,21,14,31,0,0 // vbroadcastss 0x1f0e(%rip),%ymm10 # 6b28 <_sk_callback_avx+0x46e> .byte 196,65,12,86,242 // vorps %ymm10,%ymm14,%ymm14 .byte 196,67,125,25,247,1 // vextractf128 $0x1,%ymm14,%xmm15 .byte 196,65,1,254,228 // vpaddd %xmm12,%xmm15,%xmm12 @@ -19952,7 +19966,7 @@ _sk_load_u16_be_avx: .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,29,42,28,0,0 // vbroadcastss 0x1c2a(%rip),%ymm11 # 6af0 <_sk_callback_avx+0x474> + .byte 196,98,125,24,29,102,28,0,0 // vbroadcastss 0x1c66(%rip),%ymm11 # 6b2c <_sk_callback_avx+0x472> .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,177,109,202 // vpunpckhqdq %xmm2,%xmm9,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -20045,7 +20059,7 @@ _sk_load_rgb_u16_be_avx: .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,29,138,26,0,0 // vbroadcastss 0x1a8a(%rip),%ymm11 # 6af4 <_sk_callback_avx+0x478> + .byte 196,98,125,24,29,198,26,0,0 // vbroadcastss 0x1ac6(%rip),%ymm11 # 6b30 <_sk_callback_avx+0x476> .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -20066,7 +20080,7 @@ _sk_load_rgb_u16_be_avx: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,39,26,0,0 // vbroadcastss 0x1a27(%rip),%ymm3 # 6af8 <_sk_callback_avx+0x47c> + .byte 196,226,125,24,29,99,26,0,0 // vbroadcastss 0x1a63(%rip),%ymm3 # 6b34 <_sk_callback_avx+0x47a> .byte 255,224 // jmpq *%rax .byte 196,193,121,110,4,65 // vmovd (%r9,%rax,2),%xmm0 .byte 196,193,121,196,68,65,4,2 // vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0 @@ -20109,7 +20123,7 @@ _sk_store_u16_be_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,8 // mov (%rax),%r9 .byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax - .byte 196,98,125,24,5,100,25,0,0 // vbroadcastss 0x1964(%rip),%ymm8 # 6afc <_sk_callback_avx+0x480> + .byte 196,98,125,24,5,160,25,0,0 // vbroadcastss 0x19a0(%rip),%ymm8 # 6b38 <_sk_callback_avx+0x47e> .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 @@ -20310,17 +20324,18 @@ HIDDEN _sk_repeat_x_avx FUNCTION(_sk_repeat_x_avx) _sk_repeat_x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 - .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9 - .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 - .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 - .byte 196,65,124,92,201 // vsubps %ymm9,%ymm0,%ymm9 - .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 + .byte 196,98,125,24,64,4 // vbroadcastss 0x4(%rax),%ymm8 + .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 + .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 + .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 + .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 + .byte 196,65,124,92,192 // vsubps %ymm8,%ymm0,%ymm8 + .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 - .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 - .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0 - .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 + .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 + .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 + .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20329,17 +20344,18 @@ HIDDEN _sk_repeat_y_avx FUNCTION(_sk_repeat_y_avx) _sk_repeat_y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 - .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9 - .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 - .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 - .byte 196,65,116,92,201 // vsubps %ymm9,%ymm1,%ymm9 - .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1 + .byte 196,98,125,24,64,4 // vbroadcastss 0x4(%rax),%ymm8 + .byte 197,60,89,193 // vmulps %ymm1,%ymm8,%ymm8 + .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 + .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 + .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 + .byte 196,65,116,92,192 // vsubps %ymm8,%ymm1,%ymm8 + .byte 196,99,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm1 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 - .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 - .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 - .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 + .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 + .byte 196,227,53,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm9,%ymm1 + .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20349,24 +20365,28 @@ FUNCTION(_sk_mirror_x_avx) _sk_mirror_x_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,121,110,0 // vmovd (%rax),%xmm8 - .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9 - .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 - .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10 + .byte 197,122,16,72,4 // vmovss 0x4(%rax),%xmm9 + .byte 196,65,121,112,208,0 // vpshufd $0x0,%xmm8,%xmm10 + .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 + .byte 196,65,124,92,218 // vsubps %ymm10,%ymm0,%ymm11 .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%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 197,178,89,5,211,21,0,0 // vmulss 0x15d3(%rip),%xmm9,%xmm0 # 6b3c <_sk_callback_avx+0x482> + .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8 - .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 - .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 - .byte 197,172,92,192 // vsubps %ymm0,%ymm10,%ymm0 - .byte 196,193,124,92,193 // vsubps %ymm9,%ymm0,%ymm0 + .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0 + .byte 196,227,125,8,192,1 // vroundps $0x1,%ymm0,%ymm0 + .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 + .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 + .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8 .byte 197,60,84,192 // vandps %ymm0,%ymm8,%ymm8 - .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 - .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 - .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 - .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 + .byte 196,99,125,25,208,1 // vextractf128 $0x1,%ymm10,%xmm0 + .byte 196,65,49,118,201 // vpcmpeqd %xmm9,%xmm9,%xmm9 + .byte 196,193,121,254,193 // vpaddd %xmm9,%xmm0,%xmm0 + .byte 196,65,41,254,201 // vpaddd %xmm9,%xmm10,%xmm9 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax @@ -20378,24 +20398,28 @@ FUNCTION(_sk_mirror_y_avx) _sk_mirror_y_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,121,110,0 // vmovd (%rax),%xmm8 - .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9 - .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 - .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10 + .byte 197,122,16,72,4 // vmovss 0x4(%rax),%xmm9 + .byte 196,65,121,112,208,0 // vpshufd $0x0,%xmm8,%xmm10 + .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 + .byte 196,65,116,92,218 // vsubps %ymm10,%ymm1,%ymm11 .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 + .byte 196,99,117,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm8 + .byte 197,178,89,13,78,21,0,0 // vmulss 0x154e(%rip),%xmm9,%xmm1 # 6b40 <_sk_callback_avx+0x486> + .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8 - .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 - .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 - .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1 - .byte 196,193,116,92,201 // vsubps %ymm9,%ymm1,%ymm1 + .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 + .byte 196,227,125,8,201,1 // vroundps $0x1,%ymm1,%ymm1 + .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 + .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 + .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8 .byte 197,60,84,193 // vandps %ymm1,%ymm8,%ymm8 - .byte 196,99,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm1 - .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 - .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 - .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 + .byte 196,99,125,25,209,1 // vextractf128 $0x1,%ymm10,%xmm1 + .byte 196,65,49,118,201 // vpcmpeqd %xmm9,%xmm9,%xmm9 + .byte 196,193,113,254,201 // vpaddd %xmm9,%xmm1,%xmm1 + .byte 196,65,41,254,201 // vpaddd %xmm9,%xmm10,%xmm9 .byte 196,227,53,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm9,%ymm1 .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -20407,7 +20431,7 @@ FUNCTION(_sk_clamp_x_1_avx) _sk_clamp_x_1_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,230,20,0,0 // vbroadcastss 0x14e6(%rip),%ymm8 # 6b00 <_sk_callback_avx+0x484> + .byte 196,98,125,24,5,236,20,0,0 // vbroadcastss 0x14ec(%rip),%ymm8 # 6b44 <_sk_callback_avx+0x48a> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20425,9 +20449,9 @@ HIDDEN _sk_mirror_x_1_avx .globl _sk_mirror_x_1_avx FUNCTION(_sk_mirror_x_1_avx) _sk_mirror_x_1_avx: - .byte 196,98,125,24,5,201,20,0,0 // vbroadcastss 0x14c9(%rip),%ymm8 # 6b04 <_sk_callback_avx+0x488> + .byte 196,98,125,24,5,207,20,0,0 // vbroadcastss 0x14cf(%rip),%ymm8 # 6b48 <_sk_callback_avx+0x48e> .byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,13,191,20,0,0 // vbroadcastss 0x14bf(%rip),%ymm9 # 6b08 <_sk_callback_avx+0x48c> + .byte 196,98,125,24,13,197,20,0,0 // vbroadcastss 0x14c5(%rip),%ymm9 # 6b4c <_sk_callback_avx+0x492> .byte 196,65,124,89,201 // vmulps %ymm9,%ymm0,%ymm9 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 .byte 196,65,52,88,201 // vaddps %ymm9,%ymm9,%ymm9 @@ -20443,12 +20467,12 @@ HIDDEN _sk_luminance_to_alpha_avx .globl _sk_luminance_to_alpha_avx FUNCTION(_sk_luminance_to_alpha_avx) _sk_luminance_to_alpha_avx: - .byte 196,226,125,24,29,143,20,0,0 // vbroadcastss 0x148f(%rip),%ymm3 # 6b0c <_sk_callback_avx+0x490> + .byte 196,226,125,24,29,149,20,0,0 // vbroadcastss 0x1495(%rip),%ymm3 # 6b50 <_sk_callback_avx+0x496> .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,29,134,20,0,0 // vbroadcastss 0x1486(%rip),%ymm3 # 6b10 <_sk_callback_avx+0x494> + .byte 196,226,125,24,29,140,20,0,0 // vbroadcastss 0x148c(%rip),%ymm3 # 6b54 <_sk_callback_avx+0x49a> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,121,20,0,0 // vbroadcastss 0x1479(%rip),%ymm1 # 6b14 <_sk_callback_avx+0x498> + .byte 196,226,125,24,13,127,20,0,0 // vbroadcastss 0x147f(%rip),%ymm1 # 6b58 <_sk_callback_avx+0x49e> .byte 197,236,89,201 // vmulps %ymm1,%ymm2,%ymm1 .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -20668,9 +20692,9 @@ _sk_evenly_spaced_gradient_avx: .byte 72,139,24 // mov (%rax),%rbx .byte 72,139,104,8 // mov 0x8(%rax),%rbp .byte 72,255,203 // dec %rbx - .byte 120,7 // js 59f7 <_sk_evenly_spaced_gradient_avx+0x25> + .byte 120,7 // js 5a35 <_sk_evenly_spaced_gradient_avx+0x25> .byte 196,225,242,42,203 // vcvtsi2ss %rbx,%xmm1,%xmm1 - .byte 235,21 // jmp 5a0c <_sk_evenly_spaced_gradient_avx+0x3a> + .byte 235,21 // jmp 5a4a <_sk_evenly_spaced_gradient_avx+0x3a> .byte 73,137,217 // mov %rbx,%r9 .byte 73,209,233 // shr %r9 .byte 131,227,1 // and $0x1,%ebx @@ -20828,18 +20852,18 @@ HIDDEN _sk_gauss_a_to_rgba_avx .globl _sk_gauss_a_to_rgba_avx FUNCTION(_sk_gauss_a_to_rgba_avx) _sk_gauss_a_to_rgba_avx: - .byte 196,226,125,24,5,168,13,0,0 // vbroadcastss 0xda8(%rip),%ymm0 # 6b18 <_sk_callback_avx+0x49c> + .byte 196,226,125,24,5,174,13,0,0 // vbroadcastss 0xdae(%rip),%ymm0 # 6b5c <_sk_callback_avx+0x4a2> .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 - .byte 196,226,125,24,13,159,13,0,0 // vbroadcastss 0xd9f(%rip),%ymm1 # 6b1c <_sk_callback_avx+0x4a0> + .byte 196,226,125,24,13,165,13,0,0 // vbroadcastss 0xda5(%rip),%ymm1 # 6b60 <_sk_callback_avx+0x4a6> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,13,146,13,0,0 // vbroadcastss 0xd92(%rip),%ymm1 # 6b20 <_sk_callback_avx+0x4a4> + .byte 196,226,125,24,13,152,13,0,0 // vbroadcastss 0xd98(%rip),%ymm1 # 6b64 <_sk_callback_avx+0x4aa> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,13,133,13,0,0 // vbroadcastss 0xd85(%rip),%ymm1 # 6b24 <_sk_callback_avx+0x4a8> + .byte 196,226,125,24,13,139,13,0,0 // vbroadcastss 0xd8b(%rip),%ymm1 # 6b68 <_sk_callback_avx+0x4ae> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,13,120,13,0,0 // vbroadcastss 0xd78(%rip),%ymm1 # 6b28 <_sk_callback_avx+0x4ac> + .byte 196,226,125,24,13,126,13,0,0 // vbroadcastss 0xd7e(%rip),%ymm1 # 6b6c <_sk_callback_avx+0x4b2> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 @@ -20862,12 +20886,12 @@ _sk_gradient_avx: .byte 76,139,8 // mov (%rax),%r9 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 73,131,249,2 // cmp $0x2,%r9 - .byte 114,80 // jb 5e33 <_sk_gradient_avx+0x6f> + .byte 114,80 // jb 5e71 <_sk_gradient_avx+0x6f> .byte 72,139,88,72 // mov 0x48(%rax),%rbx .byte 73,255,201 // dec %r9 .byte 72,131,195,4 // add $0x4,%rbx .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 - .byte 196,98,125,24,21,48,13,0,0 // vbroadcastss 0xd30(%rip),%ymm10 # 6b2c <_sk_callback_avx+0x4b0> + .byte 196,98,125,24,21,54,13,0,0 // vbroadcastss 0xd36(%rip),%ymm10 # 6b70 <_sk_callback_avx+0x4b6> .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 196,98,125,24,3 // vbroadcastss (%rbx),%ymm8 .byte 197,60,194,192,2 // vcmpleps %ymm0,%ymm8,%ymm8 @@ -20879,7 +20903,7 @@ _sk_gradient_avx: .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 .byte 72,131,195,4 // add $0x4,%rbx .byte 73,255,201 // dec %r9 - .byte 117,205 // jne 5e00 <_sk_gradient_avx+0x3c> + .byte 117,205 // jne 5e3e <_sk_gradient_avx+0x3c> .byte 196,195,249,22,201,1 // vpextrq $0x1,%xmm1,%r9 .byte 69,137,202 // mov %r9d,%r10d .byte 73,193,233,32 // shr $0x20,%r9 @@ -21062,27 +21086,27 @@ _sk_xy_to_unit_angle_avx: .byte 196,65,52,95,226 // vmaxps %ymm10,%ymm9,%ymm12 .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 .byte 196,65,36,89,227 // vmulps %ymm11,%ymm11,%ymm12 - .byte 196,98,125,24,45,38,9,0,0 // vbroadcastss 0x926(%rip),%ymm13 # 6b30 <_sk_callback_avx+0x4b4> + .byte 196,98,125,24,45,44,9,0,0 // vbroadcastss 0x92c(%rip),%ymm13 # 6b74 <_sk_callback_avx+0x4ba> .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13 - .byte 196,98,125,24,53,28,9,0,0 // vbroadcastss 0x91c(%rip),%ymm14 # 6b34 <_sk_callback_avx+0x4b8> + .byte 196,98,125,24,53,34,9,0,0 // vbroadcastss 0x922(%rip),%ymm14 # 6b78 <_sk_callback_avx+0x4be> .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13 - .byte 196,98,125,24,53,13,9,0,0 // vbroadcastss 0x90d(%rip),%ymm14 # 6b38 <_sk_callback_avx+0x4bc> + .byte 196,98,125,24,53,19,9,0,0 // vbroadcastss 0x913(%rip),%ymm14 # 6b7c <_sk_callback_avx+0x4c2> .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 .byte 196,65,28,89,229 // vmulps %ymm13,%ymm12,%ymm12 - .byte 196,98,125,24,45,254,8,0,0 // vbroadcastss 0x8fe(%rip),%ymm13 # 6b3c <_sk_callback_avx+0x4c0> + .byte 196,98,125,24,45,4,9,0,0 // vbroadcastss 0x904(%rip),%ymm13 # 6b80 <_sk_callback_avx+0x4c6> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 .byte 196,65,52,194,202,1 // vcmpltps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,21,233,8,0,0 // vbroadcastss 0x8e9(%rip),%ymm10 # 6b40 <_sk_callback_avx+0x4c4> + .byte 196,98,125,24,21,239,8,0,0 // vbroadcastss 0x8ef(%rip),%ymm10 # 6b84 <_sk_callback_avx+0x4ca> .byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10 .byte 196,67,37,74,202,144 // vblendvps %ymm9,%ymm10,%ymm11,%ymm9 .byte 196,193,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,21,211,8,0,0 // vbroadcastss 0x8d3(%rip),%ymm10 # 6b44 <_sk_callback_avx+0x4c8> + .byte 196,98,125,24,21,217,8,0,0 // vbroadcastss 0x8d9(%rip),%ymm10 # 6b88 <_sk_callback_avx+0x4ce> .byte 196,65,44,92,209 // vsubps %ymm9,%ymm10,%ymm10 .byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0 .byte 196,65,116,194,200,1 // vcmpltps %ymm8,%ymm1,%ymm9 - .byte 196,98,125,24,21,189,8,0,0 // vbroadcastss 0x8bd(%rip),%ymm10 # 6b48 <_sk_callback_avx+0x4cc> + .byte 196,98,125,24,21,195,8,0,0 // vbroadcastss 0x8c3(%rip),%ymm10 # 6b8c <_sk_callback_avx+0x4d2> .byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10 .byte 196,195,125,74,194,144 // vblendvps %ymm9,%ymm10,%ymm0,%ymm0 .byte 196,65,124,194,200,3 // vcmpunordps %ymm8,%ymm0,%ymm9 @@ -21106,7 +21130,7 @@ HIDDEN _sk_save_xy_avx FUNCTION(_sk_save_xy_avx) _sk_save_xy_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,135,8,0,0 // vbroadcastss 0x887(%rip),%ymm8 # 6b4c <_sk_callback_avx+0x4d0> + .byte 196,98,125,24,5,141,8,0,0 // vbroadcastss 0x88d(%rip),%ymm8 # 6b90 <_sk_callback_avx+0x4d6> .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 @@ -21143,9 +21167,9 @@ HIDDEN _sk_bilinear_nx_avx FUNCTION(_sk_bilinear_nx_avx) _sk_bilinear_nx_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,19,8,0,0 // vbroadcastss 0x813(%rip),%ymm0 # 6b50 <_sk_callback_avx+0x4d4> + .byte 196,226,125,24,5,25,8,0,0 // vbroadcastss 0x819(%rip),%ymm0 # 6b94 <_sk_callback_avx+0x4da> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,10,8,0,0 // vbroadcastss 0x80a(%rip),%ymm8 # 6b54 <_sk_callback_avx+0x4d8> + .byte 196,98,125,24,5,16,8,0,0 // vbroadcastss 0x810(%rip),%ymm8 # 6b98 <_sk_callback_avx+0x4de> .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 @@ -21156,7 +21180,7 @@ HIDDEN _sk_bilinear_px_avx FUNCTION(_sk_bilinear_px_avx) _sk_bilinear_px_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,242,7,0,0 // vbroadcastss 0x7f2(%rip),%ymm0 # 6b58 <_sk_callback_avx+0x4dc> + .byte 196,226,125,24,5,248,7,0,0 // vbroadcastss 0x7f8(%rip),%ymm0 # 6b9c <_sk_callback_avx+0x4e2> .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) @@ -21168,9 +21192,9 @@ HIDDEN _sk_bilinear_ny_avx FUNCTION(_sk_bilinear_ny_avx) _sk_bilinear_ny_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,214,7,0,0 // vbroadcastss 0x7d6(%rip),%ymm1 # 6b5c <_sk_callback_avx+0x4e0> + .byte 196,226,125,24,13,220,7,0,0 // vbroadcastss 0x7dc(%rip),%ymm1 # 6ba0 <_sk_callback_avx+0x4e6> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,204,7,0,0 // vbroadcastss 0x7cc(%rip),%ymm8 # 6b60 <_sk_callback_avx+0x4e4> + .byte 196,98,125,24,5,210,7,0,0 // vbroadcastss 0x7d2(%rip),%ymm8 # 6ba4 <_sk_callback_avx+0x4ea> .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 @@ -21181,7 +21205,7 @@ HIDDEN _sk_bilinear_py_avx FUNCTION(_sk_bilinear_py_avx) _sk_bilinear_py_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,180,7,0,0 // vbroadcastss 0x7b4(%rip),%ymm1 # 6b64 <_sk_callback_avx+0x4e8> + .byte 196,226,125,24,13,186,7,0,0 // vbroadcastss 0x7ba(%rip),%ymm1 # 6ba8 <_sk_callback_avx+0x4ee> .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) @@ -21193,14 +21217,14 @@ HIDDEN _sk_bicubic_n3x_avx FUNCTION(_sk_bicubic_n3x_avx) _sk_bicubic_n3x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,151,7,0,0 // vbroadcastss 0x797(%rip),%ymm0 # 6b68 <_sk_callback_avx+0x4ec> + .byte 196,226,125,24,5,157,7,0,0 // vbroadcastss 0x79d(%rip),%ymm0 # 6bac <_sk_callback_avx+0x4f2> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,142,7,0,0 // vbroadcastss 0x78e(%rip),%ymm8 # 6b6c <_sk_callback_avx+0x4f0> + .byte 196,98,125,24,5,148,7,0,0 // vbroadcastss 0x794(%rip),%ymm8 # 6bb0 <_sk_callback_avx+0x4f6> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,127,7,0,0 // vbroadcastss 0x77f(%rip),%ymm10 # 6b70 <_sk_callback_avx+0x4f4> + .byte 196,98,125,24,21,133,7,0,0 // vbroadcastss 0x785(%rip),%ymm10 # 6bb4 <_sk_callback_avx+0x4fa> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,117,7,0,0 // vbroadcastss 0x775(%rip),%ymm10 # 6b74 <_sk_callback_avx+0x4f8> + .byte 196,98,125,24,21,123,7,0,0 // vbroadcastss 0x77b(%rip),%ymm10 # 6bb8 <_sk_callback_avx+0x4fe> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -21212,19 +21236,19 @@ HIDDEN _sk_bicubic_n1x_avx FUNCTION(_sk_bicubic_n1x_avx) _sk_bicubic_n1x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,88,7,0,0 // vbroadcastss 0x758(%rip),%ymm0 # 6b78 <_sk_callback_avx+0x4fc> + .byte 196,226,125,24,5,94,7,0,0 // vbroadcastss 0x75e(%rip),%ymm0 # 6bbc <_sk_callback_avx+0x502> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,79,7,0,0 // vbroadcastss 0x74f(%rip),%ymm8 # 6b7c <_sk_callback_avx+0x500> + .byte 196,98,125,24,5,85,7,0,0 // vbroadcastss 0x755(%rip),%ymm8 # 6bc0 <_sk_callback_avx+0x506> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,69,7,0,0 // vbroadcastss 0x745(%rip),%ymm9 # 6b80 <_sk_callback_avx+0x504> + .byte 196,98,125,24,13,75,7,0,0 // vbroadcastss 0x74b(%rip),%ymm9 # 6bc4 <_sk_callback_avx+0x50a> .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,59,7,0,0 // vbroadcastss 0x73b(%rip),%ymm10 # 6b84 <_sk_callback_avx+0x508> + .byte 196,98,125,24,21,65,7,0,0 // vbroadcastss 0x741(%rip),%ymm10 # 6bc8 <_sk_callback_avx+0x50e> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,44,7,0,0 // vbroadcastss 0x72c(%rip),%ymm10 # 6b88 <_sk_callback_avx+0x50c> + .byte 196,98,125,24,21,50,7,0,0 // vbroadcastss 0x732(%rip),%ymm10 # 6bcc <_sk_callback_avx+0x512> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,29,7,0,0 // vbroadcastss 0x71d(%rip),%ymm9 # 6b8c <_sk_callback_avx+0x510> + .byte 196,98,125,24,13,35,7,0,0 // vbroadcastss 0x723(%rip),%ymm9 # 6bd0 <_sk_callback_avx+0x516> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21235,17 +21259,17 @@ HIDDEN _sk_bicubic_p1x_avx FUNCTION(_sk_bicubic_p1x_avx) _sk_bicubic_p1x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,5,7,0,0 // vbroadcastss 0x705(%rip),%ymm8 # 6b90 <_sk_callback_avx+0x514> + .byte 196,98,125,24,5,11,7,0,0 // vbroadcastss 0x70b(%rip),%ymm8 # 6bd4 <_sk_callback_avx+0x51a> .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 - .byte 196,98,125,24,21,247,6,0,0 // vbroadcastss 0x6f7(%rip),%ymm10 # 6b94 <_sk_callback_avx+0x518> + .byte 196,98,125,24,21,253,6,0,0 // vbroadcastss 0x6fd(%rip),%ymm10 # 6bd8 <_sk_callback_avx+0x51e> .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 - .byte 196,98,125,24,29,237,6,0,0 // vbroadcastss 0x6ed(%rip),%ymm11 # 6b98 <_sk_callback_avx+0x51c> + .byte 196,98,125,24,29,243,6,0,0 // vbroadcastss 0x6f3(%rip),%ymm11 # 6bdc <_sk_callback_avx+0x522> .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 - .byte 196,98,125,24,13,212,6,0,0 // vbroadcastss 0x6d4(%rip),%ymm9 # 6b9c <_sk_callback_avx+0x520> + .byte 196,98,125,24,13,218,6,0,0 // vbroadcastss 0x6da(%rip),%ymm9 # 6be0 <_sk_callback_avx+0x526> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21256,13 +21280,13 @@ HIDDEN _sk_bicubic_p3x_avx FUNCTION(_sk_bicubic_p3x_avx) _sk_bicubic_p3x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,188,6,0,0 // vbroadcastss 0x6bc(%rip),%ymm0 # 6ba0 <_sk_callback_avx+0x524> + .byte 196,226,125,24,5,194,6,0,0 // vbroadcastss 0x6c2(%rip),%ymm0 # 6be4 <_sk_callback_avx+0x52a> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,169,6,0,0 // vbroadcastss 0x6a9(%rip),%ymm10 # 6ba4 <_sk_callback_avx+0x528> + .byte 196,98,125,24,21,175,6,0,0 // vbroadcastss 0x6af(%rip),%ymm10 # 6be8 <_sk_callback_avx+0x52e> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,159,6,0,0 // vbroadcastss 0x69f(%rip),%ymm10 # 6ba8 <_sk_callback_avx+0x52c> + .byte 196,98,125,24,21,165,6,0,0 // vbroadcastss 0x6a5(%rip),%ymm10 # 6bec <_sk_callback_avx+0x532> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -21274,14 +21298,14 @@ HIDDEN _sk_bicubic_n3y_avx FUNCTION(_sk_bicubic_n3y_avx) _sk_bicubic_n3y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,130,6,0,0 // vbroadcastss 0x682(%rip),%ymm1 # 6bac <_sk_callback_avx+0x530> + .byte 196,226,125,24,13,136,6,0,0 // vbroadcastss 0x688(%rip),%ymm1 # 6bf0 <_sk_callback_avx+0x536> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,120,6,0,0 // vbroadcastss 0x678(%rip),%ymm8 # 6bb0 <_sk_callback_avx+0x534> + .byte 196,98,125,24,5,126,6,0,0 // vbroadcastss 0x67e(%rip),%ymm8 # 6bf4 <_sk_callback_avx+0x53a> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,105,6,0,0 // vbroadcastss 0x669(%rip),%ymm10 # 6bb4 <_sk_callback_avx+0x538> + .byte 196,98,125,24,21,111,6,0,0 // vbroadcastss 0x66f(%rip),%ymm10 # 6bf8 <_sk_callback_avx+0x53e> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,95,6,0,0 // vbroadcastss 0x65f(%rip),%ymm10 # 6bb8 <_sk_callback_avx+0x53c> + .byte 196,98,125,24,21,101,6,0,0 // vbroadcastss 0x665(%rip),%ymm10 # 6bfc <_sk_callback_avx+0x542> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -21293,19 +21317,19 @@ HIDDEN _sk_bicubic_n1y_avx FUNCTION(_sk_bicubic_n1y_avx) _sk_bicubic_n1y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,66,6,0,0 // vbroadcastss 0x642(%rip),%ymm1 # 6bbc <_sk_callback_avx+0x540> + .byte 196,226,125,24,13,72,6,0,0 // vbroadcastss 0x648(%rip),%ymm1 # 6c00 <_sk_callback_avx+0x546> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,56,6,0,0 // vbroadcastss 0x638(%rip),%ymm8 # 6bc0 <_sk_callback_avx+0x544> + .byte 196,98,125,24,5,62,6,0,0 // vbroadcastss 0x63e(%rip),%ymm8 # 6c04 <_sk_callback_avx+0x54a> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,46,6,0,0 // vbroadcastss 0x62e(%rip),%ymm9 # 6bc4 <_sk_callback_avx+0x548> + .byte 196,98,125,24,13,52,6,0,0 // vbroadcastss 0x634(%rip),%ymm9 # 6c08 <_sk_callback_avx+0x54e> .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,36,6,0,0 // vbroadcastss 0x624(%rip),%ymm10 # 6bc8 <_sk_callback_avx+0x54c> + .byte 196,98,125,24,21,42,6,0,0 // vbroadcastss 0x62a(%rip),%ymm10 # 6c0c <_sk_callback_avx+0x552> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,21,6,0,0 // vbroadcastss 0x615(%rip),%ymm10 # 6bcc <_sk_callback_avx+0x550> + .byte 196,98,125,24,21,27,6,0,0 // vbroadcastss 0x61b(%rip),%ymm10 # 6c10 <_sk_callback_avx+0x556> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,6,6,0,0 // vbroadcastss 0x606(%rip),%ymm9 # 6bd0 <_sk_callback_avx+0x554> + .byte 196,98,125,24,13,12,6,0,0 // vbroadcastss 0x60c(%rip),%ymm9 # 6c14 <_sk_callback_avx+0x55a> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21316,17 +21340,17 @@ HIDDEN _sk_bicubic_p1y_avx FUNCTION(_sk_bicubic_p1y_avx) _sk_bicubic_p1y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,238,5,0,0 // vbroadcastss 0x5ee(%rip),%ymm8 # 6bd4 <_sk_callback_avx+0x558> + .byte 196,98,125,24,5,244,5,0,0 // vbroadcastss 0x5f4(%rip),%ymm8 # 6c18 <_sk_callback_avx+0x55e> .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 - .byte 196,98,125,24,21,223,5,0,0 // vbroadcastss 0x5df(%rip),%ymm10 # 6bd8 <_sk_callback_avx+0x55c> + .byte 196,98,125,24,21,229,5,0,0 // vbroadcastss 0x5e5(%rip),%ymm10 # 6c1c <_sk_callback_avx+0x562> .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 - .byte 196,98,125,24,29,213,5,0,0 // vbroadcastss 0x5d5(%rip),%ymm11 # 6bdc <_sk_callback_avx+0x560> + .byte 196,98,125,24,29,219,5,0,0 // vbroadcastss 0x5db(%rip),%ymm11 # 6c20 <_sk_callback_avx+0x566> .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 - .byte 196,98,125,24,13,188,5,0,0 // vbroadcastss 0x5bc(%rip),%ymm9 # 6be0 <_sk_callback_avx+0x564> + .byte 196,98,125,24,13,194,5,0,0 // vbroadcastss 0x5c2(%rip),%ymm9 # 6c24 <_sk_callback_avx+0x56a> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21337,13 +21361,13 @@ HIDDEN _sk_bicubic_p3y_avx FUNCTION(_sk_bicubic_p3y_avx) _sk_bicubic_p3y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,164,5,0,0 // vbroadcastss 0x5a4(%rip),%ymm1 # 6be4 <_sk_callback_avx+0x568> + .byte 196,226,125,24,13,170,5,0,0 // vbroadcastss 0x5aa(%rip),%ymm1 # 6c28 <_sk_callback_avx+0x56e> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,144,5,0,0 // vbroadcastss 0x590(%rip),%ymm10 # 6be8 <_sk_callback_avx+0x56c> + .byte 196,98,125,24,21,150,5,0,0 // vbroadcastss 0x596(%rip),%ymm10 # 6c2c <_sk_callback_avx+0x572> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,134,5,0,0 // vbroadcastss 0x586(%rip),%ymm10 # 6bec <_sk_callback_avx+0x570> + .byte 196,98,125,24,21,140,5,0,0 // vbroadcastss 0x58c(%rip),%ymm10 # 6c30 <_sk_callback_avx+0x576> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -21473,25 +21497,25 @@ BALIGN4 .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 6871 <.literal4+0xb1> + .byte 71,225,61 // rex.RXB loope 68ad <.literal4+0xb1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 6881 <.literal4+0xc1> + .byte 71,225,61 // rex.RXB loope 68bd <.literal4+0xc1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 6891 <.literal4+0xd1> + .byte 71,225,61 // rex.RXB loope 68cd <.literal4+0xd1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 68a1 <.literal4+0xe1> + .byte 71,225,61 // rex.RXB loope 68dd <.literal4+0xe1> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax) @@ -21543,7 +21567,7 @@ 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 68f5 <.literal4+0x135> + .byte 224,7 // loopne 6931 <.literal4+0x135> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21559,10 +21583,10 @@ 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 691c <.literal4+0x15c> + .byte 127,0 // jg 6958 <.literal4+0x15c> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 6995 <.literal4+0x1d5> + .byte 119,115 // ja 69d1 <.literal4+0x1d5> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21576,10 +21600,10 @@ 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 6950 <.literal4+0x190> + .byte 127,0 // jg 698c <.literal4+0x190> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 69c9 <.literal4+0x209> + .byte 119,115 // ja 6a05 <.literal4+0x209> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21593,10 +21617,10 @@ 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 6984 <.literal4+0x1c4> + .byte 127,0 // jg 69c0 <.literal4+0x1c4> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 69fd <.literal4+0x23d> + .byte 119,115 // ja 6a39 <.literal4+0x23d> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21610,10 +21634,10 @@ 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 69b8 <.literal4+0x1f8> + .byte 127,0 // jg 69f4 <.literal4+0x1f8> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 6a31 <.literal4+0x271> + .byte 119,115 // ja 6a6d <.literal4+0x271> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21626,7 +21650,7 @@ 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 6a2f <.literal4+0x26f> + .byte 127,67 // jg 6a6b <.literal4+0x26f> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -21638,10 +21662,10 @@ BALIGN4 .byte 190,80,128,3,62 // mov $0x3e038050,%esi .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 6a4f <.literal4+0x28f> + .byte 118,63 // jbe 6a8b <.literal4+0x28f> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 6a63 <.literal4+0x2a3> + .byte 127,67 // jg 6a9f <.literal4+0x2a3> .byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax) .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax) @@ -21650,7 +21674,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 6a45 <.literal4+0x285> + .byte 224,7 // loopne 6a81 <.literal4+0x285> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21662,7 +21686,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 6a61 <.literal4+0x2a1> + .byte 224,7 // loopne 6a9d <.literal4+0x2a1> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21673,7 +21697,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 6ab6 <.literal4+0x2f6> + .byte 124,66 // jl 6af2 <.literal4+0x2f6> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax) @@ -21691,9 +21715,9 @@ 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 6af9 <.literal4+0x339> + .byte 112,65 // jo 6b35 <.literal4+0x339> .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 6b07 <.literal4+0x347> + .byte 127,67 // jg 6b43 <.literal4+0x347> .byte 0,128,0,0,0,0 // add %al,0x0(%rax) .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax) .byte 0,0 // add %al,(%rax) @@ -21709,7 +21733,11 @@ BALIGN4 .byte 0,128,55,0,0,128 // add %al,-0x7fffffc9(%rax) .byte 63 // (bad) .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 6b47 <.literal4+0x387> + .byte 127,71 // jg 6b83 <.literal4+0x387> + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,191,0,0,0 // add %al,0xbf(%rax) @@ -22092,7 +22120,7 @@ _sk_seed_shader_sse41: .byte 102,15,110,194 // movd %edx,%xmm0 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 - .byte 15,40,21,76,80,0,0 // movaps 0x504c(%rip),%xmm2 # 5120 <_sk_callback_sse41+0xf5> + .byte 15,40,21,124,80,0,0 // movaps 0x507c(%rip),%xmm2 # 5150 <_sk_callback_sse41+0xf7> .byte 15,88,202 // addps %xmm2,%xmm1 .byte 15,16,7 // movups (%rdi),%xmm0 .byte 15,88,193 // addps %xmm1,%xmm0 @@ -22101,7 +22129,7 @@ _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 15,40,21,59,80,0,0 // movaps 0x503b(%rip),%xmm2 # 5130 <_sk_callback_sse41+0x105> + .byte 15,40,21,107,80,0,0 // movaps 0x506b(%rip),%xmm2 # 5160 <_sk_callback_sse41+0x107> .byte 15,87,219 // xorps %xmm3,%xmm3 .byte 15,87,228 // xorps %xmm4,%xmm4 .byte 15,87,237 // xorps %xmm5,%xmm5 @@ -22121,14 +22149,14 @@ _sk_dither_sse41: .byte 102,68,15,110,193 // movd %ecx,%xmm8 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 .byte 102,69,15,239,193 // pxor %xmm9,%xmm8 - .byte 102,68,15,111,21,9,80,0,0 // movdqa 0x5009(%rip),%xmm10 # 5140 <_sk_callback_sse41+0x115> + .byte 102,68,15,111,21,57,80,0,0 // movdqa 0x5039(%rip),%xmm10 # 5170 <_sk_callback_sse41+0x117> .byte 102,69,15,111,216 // movdqa %xmm8,%xmm11 .byte 102,69,15,219,218 // pand %xmm10,%xmm11 .byte 102,65,15,114,243,5 // pslld $0x5,%xmm11 .byte 102,69,15,219,209 // pand %xmm9,%xmm10 .byte 102,65,15,114,242,4 // pslld $0x4,%xmm10 - .byte 102,68,15,111,37,245,79,0,0 // movdqa 0x4ff5(%rip),%xmm12 # 5150 <_sk_callback_sse41+0x125> - .byte 102,68,15,111,45,252,79,0,0 // movdqa 0x4ffc(%rip),%xmm13 # 5160 <_sk_callback_sse41+0x135> + .byte 102,68,15,111,37,37,80,0,0 // movdqa 0x5025(%rip),%xmm12 # 5180 <_sk_callback_sse41+0x127> + .byte 102,68,15,111,45,44,80,0,0 // movdqa 0x502c(%rip),%xmm13 # 5190 <_sk_callback_sse41+0x137> .byte 102,69,15,111,240 // movdqa %xmm8,%xmm14 .byte 102,69,15,219,245 // pand %xmm13,%xmm14 .byte 102,65,15,114,246,2 // pslld $0x2,%xmm14 @@ -22144,8 +22172,8 @@ _sk_dither_sse41: .byte 102,69,15,235,245 // por %xmm13,%xmm14 .byte 102,69,15,235,240 // por %xmm8,%xmm14 .byte 69,15,91,198 // cvtdq2ps %xmm14,%xmm8 - .byte 68,15,89,5,183,79,0,0 // mulps 0x4fb7(%rip),%xmm8 # 5170 <_sk_callback_sse41+0x145> - .byte 68,15,88,5,191,79,0,0 // addps 0x4fbf(%rip),%xmm8 # 5180 <_sk_callback_sse41+0x155> + .byte 68,15,89,5,231,79,0,0 // mulps 0x4fe7(%rip),%xmm8 # 51a0 <_sk_callback_sse41+0x147> + .byte 68,15,88,5,239,79,0,0 // addps 0x4fef(%rip),%xmm8 # 51b0 <_sk_callback_sse41+0x157> .byte 243,68,15,16,16 // movss (%rax),%xmm10 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 @@ -22222,7 +22250,7 @@ HIDDEN _sk_srcatop_sse41 FUNCTION(_sk_srcatop_sse41) _sk_srcatop_sse41: .byte 15,89,199 // mulps %xmm7,%xmm0 - .byte 68,15,40,5,25,79,0,0 // movaps 0x4f19(%rip),%xmm8 # 5190 <_sk_callback_sse41+0x165> + .byte 68,15,40,5,73,79,0,0 // movaps 0x4f49(%rip),%xmm8 # 51c0 <_sk_callback_sse41+0x167> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -22247,7 +22275,7 @@ FUNCTION(_sk_dstatop_sse41) _sk_dstatop_sse41: .byte 68,15,40,195 // movaps %xmm3,%xmm8 .byte 68,15,89,196 // mulps %xmm4,%xmm8 - .byte 68,15,40,13,220,78,0,0 // movaps 0x4edc(%rip),%xmm9 # 51a0 <_sk_callback_sse41+0x175> + .byte 68,15,40,13,12,79,0,0 // movaps 0x4f0c(%rip),%xmm9 # 51d0 <_sk_callback_sse41+0x177> .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,88,192 // addps %xmm8,%xmm0 @@ -22294,7 +22322,7 @@ HIDDEN _sk_srcout_sse41 .globl _sk_srcout_sse41 FUNCTION(_sk_srcout_sse41) _sk_srcout_sse41: - .byte 68,15,40,5,128,78,0,0 // movaps 0x4e80(%rip),%xmm8 # 51b0 <_sk_callback_sse41+0x185> + .byte 68,15,40,5,176,78,0,0 // movaps 0x4eb0(%rip),%xmm8 # 51e0 <_sk_callback_sse41+0x187> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 @@ -22307,7 +22335,7 @@ HIDDEN _sk_dstout_sse41 .globl _sk_dstout_sse41 FUNCTION(_sk_dstout_sse41) _sk_dstout_sse41: - .byte 68,15,40,5,112,78,0,0 // movaps 0x4e70(%rip),%xmm8 # 51c0 <_sk_callback_sse41+0x195> + .byte 68,15,40,5,160,78,0,0 // movaps 0x4ea0(%rip),%xmm8 # 51f0 <_sk_callback_sse41+0x197> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 15,89,196 // mulps %xmm4,%xmm0 @@ -22324,7 +22352,7 @@ HIDDEN _sk_srcover_sse41 .globl _sk_srcover_sse41 FUNCTION(_sk_srcover_sse41) _sk_srcover_sse41: - .byte 68,15,40,5,83,78,0,0 // movaps 0x4e53(%rip),%xmm8 # 51d0 <_sk_callback_sse41+0x1a5> + .byte 68,15,40,5,131,78,0,0 // movaps 0x4e83(%rip),%xmm8 # 5200 <_sk_callback_sse41+0x1a7> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -22344,7 +22372,7 @@ HIDDEN _sk_dstover_sse41 .globl _sk_dstover_sse41 FUNCTION(_sk_dstover_sse41) _sk_dstover_sse41: - .byte 68,15,40,5,39,78,0,0 // movaps 0x4e27(%rip),%xmm8 # 51e0 <_sk_callback_sse41+0x1b5> + .byte 68,15,40,5,87,78,0,0 // movaps 0x4e57(%rip),%xmm8 # 5210 <_sk_callback_sse41+0x1b7> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -22372,7 +22400,7 @@ HIDDEN _sk_multiply_sse41 .globl _sk_multiply_sse41 FUNCTION(_sk_multiply_sse41) _sk_multiply_sse41: - .byte 68,15,40,5,251,77,0,0 // movaps 0x4dfb(%rip),%xmm8 # 51f0 <_sk_callback_sse41+0x1c5> + .byte 68,15,40,5,43,78,0,0 // movaps 0x4e2b(%rip),%xmm8 # 5220 <_sk_callback_sse41+0x1c7> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 69,15,40,209 // movaps %xmm9,%xmm10 @@ -22447,7 +22475,7 @@ HIDDEN _sk_xor__sse41 FUNCTION(_sk_xor__sse41) _sk_xor__sse41: .byte 68,15,40,195 // movaps %xmm3,%xmm8 - .byte 15,40,29,48,77,0,0 // movaps 0x4d30(%rip),%xmm3 # 5200 <_sk_callback_sse41+0x1d5> + .byte 15,40,29,96,77,0,0 // movaps 0x4d60(%rip),%xmm3 # 5230 <_sk_callback_sse41+0x1d7> .byte 68,15,40,203 // movaps %xmm3,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 @@ -22495,7 +22523,7 @@ _sk_darken_sse41: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,95,209 // maxps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,155,76,0,0 // movaps 0x4c9b(%rip),%xmm2 # 5210 <_sk_callback_sse41+0x1e5> + .byte 15,40,21,203,76,0,0 // movaps 0x4ccb(%rip),%xmm2 # 5240 <_sk_callback_sse41+0x1e7> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -22529,7 +22557,7 @@ _sk_lighten_sse41: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,64,76,0,0 // movaps 0x4c40(%rip),%xmm2 # 5220 <_sk_callback_sse41+0x1f5> + .byte 15,40,21,112,76,0,0 // movaps 0x4c70(%rip),%xmm2 # 5250 <_sk_callback_sse41+0x1f7> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -22566,7 +22594,7 @@ _sk_difference_sse41: .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,218,75,0,0 // movaps 0x4bda(%rip),%xmm2 # 5230 <_sk_callback_sse41+0x205> + .byte 15,40,21,10,76,0,0 // movaps 0x4c0a(%rip),%xmm2 # 5260 <_sk_callback_sse41+0x207> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -22594,7 +22622,7 @@ _sk_exclusion_sse41: .byte 15,89,214 // mulps %xmm6,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,154,75,0,0 // movaps 0x4b9a(%rip),%xmm2 # 5240 <_sk_callback_sse41+0x215> + .byte 15,40,21,202,75,0,0 // movaps 0x4bca(%rip),%xmm2 # 5270 <_sk_callback_sse41+0x217> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -22607,7 +22635,7 @@ HIDDEN _sk_colorburn_sse41 FUNCTION(_sk_colorburn_sse41) _sk_colorburn_sse41: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,40,21,141,75,0,0 // movaps 0x4b8d(%rip),%xmm10 # 5250 <_sk_callback_sse41+0x225> + .byte 68,15,40,21,189,75,0,0 // movaps 0x4bbd(%rip),%xmm10 # 5280 <_sk_callback_sse41+0x227> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 69,15,40,203 // movaps %xmm11,%xmm9 @@ -22689,7 +22717,7 @@ HIDDEN _sk_colordodge_sse41 FUNCTION(_sk_colordodge_sse41) _sk_colordodge_sse41: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,40,21,107,74,0,0 // movaps 0x4a6b(%rip),%xmm10 # 5260 <_sk_callback_sse41+0x235> + .byte 68,15,40,21,155,74,0,0 // movaps 0x4a9b(%rip),%xmm10 # 5290 <_sk_callback_sse41+0x237> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 69,15,40,227 // movaps %xmm11,%xmm12 @@ -22771,7 +22799,7 @@ _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 68,15,40,21,68,73,0,0 // movaps 0x4944(%rip),%xmm10 # 5270 <_sk_callback_sse41+0x245> + .byte 68,15,40,21,116,73,0,0 // movaps 0x4974(%rip),%xmm10 # 52a0 <_sk_callback_sse41+0x247> .byte 65,15,40,234 // movaps %xmm10,%xmm5 .byte 15,92,239 // subps %xmm7,%xmm5 .byte 15,40,197 // movaps %xmm5,%xmm0 @@ -22854,7 +22882,7 @@ FUNCTION(_sk_overlay_sse41) _sk_overlay_sse41: .byte 68,15,40,201 // movaps %xmm1,%xmm9 .byte 68,15,40,240 // movaps %xmm0,%xmm14 - .byte 68,15,40,21,41,72,0,0 // movaps 0x4829(%rip),%xmm10 # 5280 <_sk_callback_sse41+0x255> + .byte 68,15,40,21,89,72,0,0 // movaps 0x4859(%rip),%xmm10 # 52b0 <_sk_callback_sse41+0x257> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 65,15,40,195 // movaps %xmm11,%xmm0 @@ -22939,7 +22967,7 @@ _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 15,40,13,0,71,0,0 // movaps 0x4700(%rip),%xmm1 # 5290 <_sk_callback_sse41+0x265> + .byte 15,40,13,48,71,0,0 // movaps 0x4730(%rip),%xmm1 # 52c0 <_sk_callback_sse41+0x267> .byte 68,15,40,209 // movaps %xmm1,%xmm10 .byte 68,15,92,208 // subps %xmm0,%xmm10 .byte 68,15,40,240 // movaps %xmm0,%xmm14 @@ -22952,10 +22980,10 @@ _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 68,15,40,45,222,70,0,0 // movaps 0x46de(%rip),%xmm13 # 52a0 <_sk_callback_sse41+0x275> + .byte 68,15,40,45,14,71,0,0 // movaps 0x470e(%rip),%xmm13 # 52d0 <_sk_callback_sse41+0x277> .byte 69,15,88,245 // addps %xmm13,%xmm14 .byte 68,15,89,242 // mulps %xmm2,%xmm14 - .byte 68,15,40,37,222,70,0,0 // movaps 0x46de(%rip),%xmm12 # 52b0 <_sk_callback_sse41+0x285> + .byte 68,15,40,37,14,71,0,0 // movaps 0x470e(%rip),%xmm12 # 52e0 <_sk_callback_sse41+0x287> .byte 69,15,89,252 // mulps %xmm12,%xmm15 .byte 69,15,88,254 // addps %xmm14,%xmm15 .byte 15,40,198 // movaps %xmm6,%xmm0 @@ -23141,12 +23169,12 @@ _sk_hue_sse41: .byte 68,15,84,208 // andps %xmm0,%xmm10 .byte 15,84,200 // andps %xmm0,%xmm1 .byte 68,15,84,232 // andps %xmm0,%xmm13 - .byte 15,40,5,73,68,0,0 // movaps 0x4449(%rip),%xmm0 # 52c0 <_sk_callback_sse41+0x295> + .byte 15,40,5,121,68,0,0 // movaps 0x4479(%rip),%xmm0 # 52f0 <_sk_callback_sse41+0x297> .byte 68,15,89,224 // mulps %xmm0,%xmm12 - .byte 15,40,21,78,68,0,0 // movaps 0x444e(%rip),%xmm2 # 52d0 <_sk_callback_sse41+0x2a5> + .byte 15,40,21,126,68,0,0 // movaps 0x447e(%rip),%xmm2 # 5300 <_sk_callback_sse41+0x2a7> .byte 15,89,250 // mulps %xmm2,%xmm7 .byte 65,15,88,252 // addps %xmm12,%xmm7 - .byte 68,15,40,53,79,68,0,0 // movaps 0x444f(%rip),%xmm14 # 52e0 <_sk_callback_sse41+0x2b5> + .byte 68,15,40,53,127,68,0,0 // movaps 0x447f(%rip),%xmm14 # 5310 <_sk_callback_sse41+0x2b7> .byte 68,15,40,252 // movaps %xmm4,%xmm15 .byte 69,15,89,254 // mulps %xmm14,%xmm15 .byte 68,15,88,255 // addps %xmm7,%xmm15 @@ -23229,7 +23257,7 @@ _sk_hue_sse41: .byte 65,15,88,214 // addps %xmm14,%xmm2 .byte 15,40,196 // movaps %xmm4,%xmm0 .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1 - .byte 68,15,40,13,19,67,0,0 // movaps 0x4313(%rip),%xmm9 # 52f0 <_sk_callback_sse41+0x2c5> + .byte 68,15,40,13,67,67,0,0 // movaps 0x4343(%rip),%xmm9 # 5320 <_sk_callback_sse41+0x2c7> .byte 65,15,40,225 // movaps %xmm9,%xmm4 .byte 15,92,229 // subps %xmm5,%xmm4 .byte 15,40,68,36,200 // movaps -0x38(%rsp),%xmm0 @@ -23323,14 +23351,14 @@ _sk_saturation_sse41: .byte 68,15,84,215 // andps %xmm7,%xmm10 .byte 68,15,84,223 // andps %xmm7,%xmm11 .byte 68,15,84,199 // andps %xmm7,%xmm8 - .byte 15,40,21,205,65,0,0 // movaps 0x41cd(%rip),%xmm2 # 5300 <_sk_callback_sse41+0x2d5> + .byte 15,40,21,253,65,0,0 // movaps 0x41fd(%rip),%xmm2 # 5330 <_sk_callback_sse41+0x2d7> .byte 15,40,221 // movaps %xmm5,%xmm3 .byte 15,89,218 // mulps %xmm2,%xmm3 - .byte 15,40,13,208,65,0,0 // movaps 0x41d0(%rip),%xmm1 # 5310 <_sk_callback_sse41+0x2e5> + .byte 15,40,13,0,66,0,0 // movaps 0x4200(%rip),%xmm1 # 5340 <_sk_callback_sse41+0x2e7> .byte 15,40,254 // movaps %xmm6,%xmm7 .byte 15,89,249 // mulps %xmm1,%xmm7 .byte 15,88,251 // addps %xmm3,%xmm7 - .byte 68,15,40,45,207,65,0,0 // movaps 0x41cf(%rip),%xmm13 # 5320 <_sk_callback_sse41+0x2f5> + .byte 68,15,40,45,255,65,0,0 // movaps 0x41ff(%rip),%xmm13 # 5350 <_sk_callback_sse41+0x2f7> .byte 69,15,89,245 // mulps %xmm13,%xmm14 .byte 68,15,88,247 // addps %xmm7,%xmm14 .byte 65,15,40,218 // movaps %xmm10,%xmm3 @@ -23411,7 +23439,7 @@ _sk_saturation_sse41: .byte 65,15,88,253 // addps %xmm13,%xmm7 .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 102,68,15,56,20,223 // blendvps %xmm0,%xmm7,%xmm11 - .byte 68,15,40,13,149,64,0,0 // movaps 0x4095(%rip),%xmm9 # 5330 <_sk_callback_sse41+0x305> + .byte 68,15,40,13,197,64,0,0 // movaps 0x40c5(%rip),%xmm9 # 5360 <_sk_callback_sse41+0x307> .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 68,15,92,204 // subps %xmm4,%xmm9 .byte 15,40,124,36,168 // movaps -0x58(%rsp),%xmm7 @@ -23466,14 +23494,14 @@ _sk_color_sse41: .byte 15,40,231 // movaps %xmm7,%xmm4 .byte 68,15,89,244 // mulps %xmm4,%xmm14 .byte 15,89,204 // mulps %xmm4,%xmm1 - .byte 68,15,40,13,224,63,0,0 // movaps 0x3fe0(%rip),%xmm9 # 5340 <_sk_callback_sse41+0x315> + .byte 68,15,40,13,16,64,0,0 // movaps 0x4010(%rip),%xmm9 # 5370 <_sk_callback_sse41+0x317> .byte 65,15,40,250 // movaps %xmm10,%xmm7 .byte 65,15,89,249 // mulps %xmm9,%xmm7 - .byte 68,15,40,21,224,63,0,0 // movaps 0x3fe0(%rip),%xmm10 # 5350 <_sk_callback_sse41+0x325> + .byte 68,15,40,21,16,64,0,0 // movaps 0x4010(%rip),%xmm10 # 5380 <_sk_callback_sse41+0x327> .byte 65,15,40,219 // movaps %xmm11,%xmm3 .byte 65,15,89,218 // mulps %xmm10,%xmm3 .byte 15,88,223 // addps %xmm7,%xmm3 - .byte 68,15,40,29,221,63,0,0 // movaps 0x3fdd(%rip),%xmm11 # 5360 <_sk_callback_sse41+0x335> + .byte 68,15,40,29,13,64,0,0 // movaps 0x400d(%rip),%xmm11 # 5390 <_sk_callback_sse41+0x337> .byte 69,15,40,236 // movaps %xmm12,%xmm13 .byte 69,15,89,235 // mulps %xmm11,%xmm13 .byte 68,15,88,235 // addps %xmm3,%xmm13 @@ -23558,7 +23586,7 @@ _sk_color_sse41: .byte 65,15,88,251 // addps %xmm11,%xmm7 .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 102,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm1 - .byte 68,15,40,13,153,62,0,0 // movaps 0x3e99(%rip),%xmm9 # 5370 <_sk_callback_sse41+0x345> + .byte 68,15,40,13,201,62,0,0 // movaps 0x3ec9(%rip),%xmm9 # 53a0 <_sk_callback_sse41+0x347> .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 15,92,196 // subps %xmm4,%xmm0 .byte 68,15,89,192 // mulps %xmm0,%xmm8 @@ -23610,13 +23638,13 @@ _sk_luminosity_sse41: .byte 69,15,89,216 // mulps %xmm8,%xmm11 .byte 68,15,40,203 // movaps %xmm3,%xmm9 .byte 68,15,89,205 // mulps %xmm5,%xmm9 - .byte 68,15,40,5,241,61,0,0 // movaps 0x3df1(%rip),%xmm8 # 5380 <_sk_callback_sse41+0x355> + .byte 68,15,40,5,33,62,0,0 // movaps 0x3e21(%rip),%xmm8 # 53b0 <_sk_callback_sse41+0x357> .byte 65,15,89,192 // mulps %xmm8,%xmm0 - .byte 68,15,40,21,245,61,0,0 // movaps 0x3df5(%rip),%xmm10 # 5390 <_sk_callback_sse41+0x365> + .byte 68,15,40,21,37,62,0,0 // movaps 0x3e25(%rip),%xmm10 # 53c0 <_sk_callback_sse41+0x367> .byte 15,40,233 // movaps %xmm1,%xmm5 .byte 65,15,89,234 // mulps %xmm10,%xmm5 .byte 15,88,232 // addps %xmm0,%xmm5 - .byte 68,15,40,37,243,61,0,0 // movaps 0x3df3(%rip),%xmm12 # 53a0 <_sk_callback_sse41+0x375> + .byte 68,15,40,37,35,62,0,0 // movaps 0x3e23(%rip),%xmm12 # 53d0 <_sk_callback_sse41+0x377> .byte 68,15,40,242 // movaps %xmm2,%xmm14 .byte 69,15,89,244 // mulps %xmm12,%xmm14 .byte 68,15,88,245 // addps %xmm5,%xmm14 @@ -23701,7 +23729,7 @@ _sk_luminosity_sse41: .byte 65,15,88,244 // addps %xmm12,%xmm6 .byte 65,15,40,195 // movaps %xmm11,%xmm0 .byte 102,68,15,56,20,206 // blendvps %xmm0,%xmm6,%xmm9 - .byte 15,40,5,169,60,0,0 // movaps 0x3ca9(%rip),%xmm0 # 53b0 <_sk_callback_sse41+0x385> + .byte 15,40,5,217,60,0,0 // movaps 0x3cd9(%rip),%xmm0 # 53e0 <_sk_callback_sse41+0x387> .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 15,92,215 // subps %xmm7,%xmm2 .byte 15,89,226 // mulps %xmm2,%xmm4 @@ -23744,20 +23772,20 @@ _sk_srcover_rgba_8888_sse41: .byte 15,133,217,0,0,0 // jne 185b <_sk_srcover_rgba_8888_sse41+0xe7> .byte 243,15,111,60,144 // movdqu (%rax,%rdx,4),%xmm7 .byte 77,133,192 // test %r8,%r8 - .byte 102,15,111,37,46,60,0,0 // movdqa 0x3c2e(%rip),%xmm4 # 53c0 <_sk_callback_sse41+0x395> + .byte 102,15,111,37,94,60,0,0 // movdqa 0x3c5e(%rip),%xmm4 # 53f0 <_sk_callback_sse41+0x397> .byte 102,15,219,231 // pand %xmm7,%xmm4 .byte 15,91,228 // cvtdq2ps %xmm4,%xmm4 .byte 102,15,111,239 // movdqa %xmm7,%xmm5 - .byte 102,15,56,0,45,42,60,0,0 // pshufb 0x3c2a(%rip),%xmm5 # 53d0 <_sk_callback_sse41+0x3a5> + .byte 102,15,56,0,45,90,60,0,0 // pshufb 0x3c5a(%rip),%xmm5 # 5400 <_sk_callback_sse41+0x3a7> .byte 15,91,237 // cvtdq2ps %xmm5,%xmm5 .byte 102,15,111,247 // movdqa %xmm7,%xmm6 - .byte 102,15,56,0,53,42,60,0,0 // pshufb 0x3c2a(%rip),%xmm6 # 53e0 <_sk_callback_sse41+0x3b5> + .byte 102,15,56,0,53,90,60,0,0 // pshufb 0x3c5a(%rip),%xmm6 # 5410 <_sk_callback_sse41+0x3b7> .byte 15,91,246 // cvtdq2ps %xmm6,%xmm6 .byte 102,15,114,215,24 // psrld $0x18,%xmm7 .byte 15,91,255 // cvtdq2ps %xmm7,%xmm7 - .byte 68,15,40,5,39,60,0,0 // movaps 0x3c27(%rip),%xmm8 # 53f0 <_sk_callback_sse41+0x3c5> + .byte 68,15,40,5,87,60,0,0 // movaps 0x3c57(%rip),%xmm8 # 5420 <_sk_callback_sse41+0x3c7> .byte 68,15,92,195 // subps %xmm3,%xmm8 - .byte 68,15,40,37,43,60,0,0 // movaps 0x3c2b(%rip),%xmm12 # 5400 <_sk_callback_sse41+0x3d5> + .byte 68,15,40,37,91,60,0,0 // movaps 0x3c5b(%rip),%xmm12 # 5430 <_sk_callback_sse41+0x3d7> .byte 65,15,89,196 // mulps %xmm12,%xmm0 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -23834,7 +23862,7 @@ HIDDEN _sk_clamp_1_sse41 .globl _sk_clamp_1_sse41 FUNCTION(_sk_clamp_1_sse41) _sk_clamp_1_sse41: - .byte 68,15,40,5,34,59,0,0 // movaps 0x3b22(%rip),%xmm8 # 5410 <_sk_callback_sse41+0x3e5> + .byte 68,15,40,5,82,59,0,0 // movaps 0x3b52(%rip),%xmm8 # 5440 <_sk_callback_sse41+0x3e7> .byte 65,15,93,192 // minps %xmm8,%xmm0 .byte 65,15,93,200 // minps %xmm8,%xmm1 .byte 65,15,93,208 // minps %xmm8,%xmm2 @@ -23846,7 +23874,7 @@ HIDDEN _sk_clamp_a_sse41 .globl _sk_clamp_a_sse41 FUNCTION(_sk_clamp_a_sse41) _sk_clamp_a_sse41: - .byte 15,93,29,23,59,0,0 // minps 0x3b17(%rip),%xmm3 # 5420 <_sk_callback_sse41+0x3f5> + .byte 15,93,29,71,59,0,0 // minps 0x3b47(%rip),%xmm3 # 5450 <_sk_callback_sse41+0x3f7> .byte 15,93,195 // minps %xmm3,%xmm0 .byte 15,93,203 // minps %xmm3,%xmm1 .byte 15,93,211 // minps %xmm3,%xmm2 @@ -23933,7 +23961,7 @@ HIDDEN _sk_unpremul_sse41 FUNCTION(_sk_unpremul_sse41) _sk_unpremul_sse41: .byte 69,15,87,192 // xorps %xmm8,%xmm8 - .byte 68,15,40,13,130,58,0,0 // movaps 0x3a82(%rip),%xmm9 # 5430 <_sk_callback_sse41+0x405> + .byte 68,15,40,13,178,58,0,0 // movaps 0x3ab2(%rip),%xmm9 # 5460 <_sk_callback_sse41+0x407> .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 @@ -23947,20 +23975,20 @@ HIDDEN _sk_from_srgb_sse41 .globl _sk_from_srgb_sse41 FUNCTION(_sk_from_srgb_sse41) _sk_from_srgb_sse41: - .byte 68,15,40,29,109,58,0,0 // movaps 0x3a6d(%rip),%xmm11 # 5440 <_sk_callback_sse41+0x415> + .byte 68,15,40,29,157,58,0,0 // movaps 0x3a9d(%rip),%xmm11 # 5470 <_sk_callback_sse41+0x417> .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 68,15,40,37,101,58,0,0 // movaps 0x3a65(%rip),%xmm12 # 5450 <_sk_callback_sse41+0x425> + .byte 68,15,40,37,149,58,0,0 // movaps 0x3a95(%rip),%xmm12 # 5480 <_sk_callback_sse41+0x427> .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 69,15,89,196 // mulps %xmm12,%xmm8 - .byte 68,15,40,45,101,58,0,0 // movaps 0x3a65(%rip),%xmm13 # 5460 <_sk_callback_sse41+0x435> + .byte 68,15,40,45,149,58,0,0 // movaps 0x3a95(%rip),%xmm13 # 5490 <_sk_callback_sse41+0x437> .byte 69,15,88,197 // addps %xmm13,%xmm8 .byte 69,15,89,194 // mulps %xmm10,%xmm8 - .byte 68,15,40,53,101,58,0,0 // movaps 0x3a65(%rip),%xmm14 # 5470 <_sk_callback_sse41+0x445> + .byte 68,15,40,53,149,58,0,0 // movaps 0x3a95(%rip),%xmm14 # 54a0 <_sk_callback_sse41+0x447> .byte 69,15,88,198 // addps %xmm14,%xmm8 - .byte 68,15,40,61,105,58,0,0 // movaps 0x3a69(%rip),%xmm15 # 5480 <_sk_callback_sse41+0x455> + .byte 68,15,40,61,153,58,0,0 // movaps 0x3a99(%rip),%xmm15 # 54b0 <_sk_callback_sse41+0x457> .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 @@ -24003,22 +24031,22 @@ _sk_to_srgb_sse41: .byte 15,40,218 // movaps %xmm2,%xmm3 .byte 15,40,209 // movaps %xmm1,%xmm2 .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 - .byte 68,15,40,29,225,57,0,0 // movaps 0x39e1(%rip),%xmm11 # 5490 <_sk_callback_sse41+0x465> + .byte 68,15,40,29,17,58,0,0 // movaps 0x3a11(%rip),%xmm11 # 54c0 <_sk_callback_sse41+0x467> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 - .byte 68,15,40,37,225,57,0,0 // movaps 0x39e1(%rip),%xmm12 # 54a0 <_sk_callback_sse41+0x475> + .byte 68,15,40,37,17,58,0,0 // movaps 0x3a11(%rip),%xmm12 # 54d0 <_sk_callback_sse41+0x477> .byte 69,15,40,248 // movaps %xmm8,%xmm15 .byte 69,15,89,252 // mulps %xmm12,%xmm15 - .byte 68,15,40,21,225,57,0,0 // movaps 0x39e1(%rip),%xmm10 # 54b0 <_sk_callback_sse41+0x485> + .byte 68,15,40,21,17,58,0,0 // movaps 0x3a11(%rip),%xmm10 # 54e0 <_sk_callback_sse41+0x487> .byte 69,15,88,250 // addps %xmm10,%xmm15 .byte 69,15,89,248 // mulps %xmm8,%xmm15 - .byte 68,15,40,45,225,57,0,0 // movaps 0x39e1(%rip),%xmm13 # 54c0 <_sk_callback_sse41+0x495> + .byte 68,15,40,45,17,58,0,0 // movaps 0x3a11(%rip),%xmm13 # 54f0 <_sk_callback_sse41+0x497> .byte 69,15,88,253 // addps %xmm13,%xmm15 - .byte 68,15,40,53,229,57,0,0 // movaps 0x39e5(%rip),%xmm14 # 54d0 <_sk_callback_sse41+0x4a5> + .byte 68,15,40,53,21,58,0,0 // movaps 0x3a15(%rip),%xmm14 # 5500 <_sk_callback_sse41+0x4a7> .byte 69,15,88,198 // addps %xmm14,%xmm8 .byte 69,15,83,192 // rcpps %xmm8,%xmm8 .byte 69,15,89,199 // mulps %xmm15,%xmm8 - .byte 68,15,40,61,225,57,0,0 // movaps 0x39e1(%rip),%xmm15 # 54e0 <_sk_callback_sse41+0x4b5> + .byte 68,15,40,61,17,58,0,0 // movaps 0x3a11(%rip),%xmm15 # 5510 <_sk_callback_sse41+0x4b7> .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 .byte 102,69,15,56,20,193 // blendvps %xmm0,%xmm9,%xmm8 .byte 68,15,82,202 // rsqrtps %xmm2,%xmm9 @@ -24071,7 +24099,7 @@ _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 68,15,40,53,46,57,0,0 // movaps 0x392e(%rip),%xmm14 # 54f0 <_sk_callback_sse41+0x4c5> + .byte 68,15,40,53,94,57,0,0 // movaps 0x395e(%rip),%xmm14 # 5520 <_sk_callback_sse41+0x4c7> .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 @@ -24080,27 +24108,27 @@ _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 68,15,84,61,21,57,0,0 // andps 0x3915(%rip),%xmm15 # 5500 <_sk_callback_sse41+0x4d5> + .byte 68,15,84,61,69,57,0,0 // andps 0x3945(%rip),%xmm15 # 5530 <_sk_callback_sse41+0x4d7> .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 68,15,40,45,8,57,0,0 // movaps 0x3908(%rip),%xmm13 # 5510 <_sk_callback_sse41+0x4e5> + .byte 68,15,40,45,56,57,0,0 // movaps 0x3938(%rip),%xmm13 # 5540 <_sk_callback_sse41+0x4e7> .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 68,15,88,5,4,57,0,0 // addps 0x3904(%rip),%xmm8 # 5520 <_sk_callback_sse41+0x4f5> + .byte 68,15,88,5,52,57,0,0 // addps 0x3934(%rip),%xmm8 # 5550 <_sk_callback_sse41+0x4f7> .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 68,15,89,5,252,56,0,0 // mulps 0x38fc(%rip),%xmm8 # 5530 <_sk_callback_sse41+0x505> + .byte 68,15,89,5,44,57,0,0 // mulps 0x392c(%rip),%xmm8 # 5560 <_sk_callback_sse41+0x507> .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 15,40,5,240,56,0,0 // movaps 0x38f0(%rip),%xmm0 # 5540 <_sk_callback_sse41+0x515> + .byte 15,40,5,32,57,0,0 // movaps 0x3920(%rip),%xmm0 # 5570 <_sk_callback_sse41+0x517> .byte 65,15,40,211 // movaps %xmm11,%xmm2 .byte 15,89,208 // mulps %xmm0,%xmm2 .byte 15,194,194,1 // cmpltps %xmm2,%xmm0 @@ -24122,7 +24150,7 @@ _sk_hsl_to_rgb_sse41: .byte 15,41,100,36,184 // movaps %xmm4,-0x48(%rsp) .byte 15,41,92,36,168 // movaps %xmm3,-0x58(%rsp) .byte 68,15,40,208 // movaps %xmm0,%xmm10 - .byte 68,15,40,13,182,56,0,0 // movaps 0x38b6(%rip),%xmm9 # 5550 <_sk_callback_sse41+0x525> + .byte 68,15,40,13,230,56,0,0 // movaps 0x38e6(%rip),%xmm9 # 5580 <_sk_callback_sse41+0x527> .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 15,194,194,2 // cmpleps %xmm2,%xmm0 .byte 15,40,217 // movaps %xmm1,%xmm3 @@ -24135,19 +24163,19 @@ _sk_hsl_to_rgb_sse41: .byte 15,41,84,36,152 // movaps %xmm2,-0x68(%rsp) .byte 69,15,88,192 // addps %xmm8,%xmm8 .byte 68,15,92,197 // subps %xmm5,%xmm8 - .byte 68,15,40,53,145,56,0,0 // movaps 0x3891(%rip),%xmm14 # 5560 <_sk_callback_sse41+0x535> + .byte 68,15,40,53,193,56,0,0 // movaps 0x38c1(%rip),%xmm14 # 5590 <_sk_callback_sse41+0x537> .byte 69,15,88,242 // addps %xmm10,%xmm14 .byte 102,65,15,58,8,198,1 // roundps $0x1,%xmm14,%xmm0 .byte 68,15,92,240 // subps %xmm0,%xmm14 - .byte 68,15,40,29,138,56,0,0 // movaps 0x388a(%rip),%xmm11 # 5570 <_sk_callback_sse41+0x545> + .byte 68,15,40,29,186,56,0,0 // movaps 0x38ba(%rip),%xmm11 # 55a0 <_sk_callback_sse41+0x547> .byte 65,15,40,195 // movaps %xmm11,%xmm0 .byte 65,15,194,198,2 // cmpleps %xmm14,%xmm0 .byte 15,40,245 // movaps %xmm5,%xmm6 .byte 65,15,92,240 // subps %xmm8,%xmm6 - .byte 15,40,61,131,56,0,0 // movaps 0x3883(%rip),%xmm7 # 5580 <_sk_callback_sse41+0x555> + .byte 15,40,61,179,56,0,0 // movaps 0x38b3(%rip),%xmm7 # 55b0 <_sk_callback_sse41+0x557> .byte 69,15,40,238 // movaps %xmm14,%xmm13 .byte 68,15,89,239 // mulps %xmm7,%xmm13 - .byte 15,40,29,132,56,0,0 // movaps 0x3884(%rip),%xmm3 # 5590 <_sk_callback_sse41+0x565> + .byte 15,40,29,180,56,0,0 // movaps 0x38b4(%rip),%xmm3 # 55c0 <_sk_callback_sse41+0x567> .byte 68,15,40,227 // movaps %xmm3,%xmm12 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 68,15,89,230 // mulps %xmm6,%xmm12 @@ -24157,7 +24185,7 @@ _sk_hsl_to_rgb_sse41: .byte 65,15,194,198,2 // cmpleps %xmm14,%xmm0 .byte 68,15,40,253 // movaps %xmm5,%xmm15 .byte 102,69,15,56,20,252 // blendvps %xmm0,%xmm12,%xmm15 - .byte 68,15,40,37,99,56,0,0 // movaps 0x3863(%rip),%xmm12 # 55a0 <_sk_callback_sse41+0x575> + .byte 68,15,40,37,147,56,0,0 // movaps 0x3893(%rip),%xmm12 # 55d0 <_sk_callback_sse41+0x577> .byte 65,15,40,196 // movaps %xmm12,%xmm0 .byte 65,15,194,198,2 // cmpleps %xmm14,%xmm0 .byte 68,15,89,238 // mulps %xmm6,%xmm13 @@ -24191,7 +24219,7 @@ _sk_hsl_to_rgb_sse41: .byte 65,15,40,198 // movaps %xmm14,%xmm0 .byte 15,40,84,36,152 // movaps -0x68(%rsp),%xmm2 .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1 - .byte 68,15,88,21,219,55,0,0 // addps 0x37db(%rip),%xmm10 # 55b0 <_sk_callback_sse41+0x585> + .byte 68,15,88,21,11,56,0,0 // addps 0x380b(%rip),%xmm10 # 55e0 <_sk_callback_sse41+0x587> .byte 102,65,15,58,8,194,1 // roundps $0x1,%xmm10,%xmm0 .byte 68,15,92,208 // subps %xmm0,%xmm10 .byte 69,15,194,218,2 // cmpleps %xmm10,%xmm11 @@ -24244,9 +24272,9 @@ _sk_scale_u8_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,52 // jne 1eae <_sk_scale_u8_sse41+0x3e> .byte 102,69,15,56,49,4,18 // pmovzxbd (%r10,%rdx,1),%xmm8 - .byte 102,68,15,219,5,54,55,0,0 // pand 0x3736(%rip),%xmm8 # 55c0 <_sk_callback_sse41+0x595> + .byte 102,68,15,219,5,102,55,0,0 // pand 0x3766(%rip),%xmm8 # 55f0 <_sk_callback_sse41+0x597> .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,58,55,0,0 // mulps 0x373a(%rip),%xmm8 # 55d0 <_sk_callback_sse41+0x5a5> + .byte 68,15,89,5,106,55,0,0 // mulps 0x376a(%rip),%xmm8 # 5600 <_sk_callback_sse41+0x5a7> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 65,15,89,208 // mulps %xmm8,%xmm2 @@ -24303,9 +24331,9 @@ _sk_lerp_u8_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,72 // jne 1f82 <_sk_lerp_u8_sse41+0x52> .byte 102,69,15,56,49,4,18 // pmovzxbd (%r10,%rdx,1),%xmm8 - .byte 102,68,15,219,5,150,54,0,0 // pand 0x3696(%rip),%xmm8 # 55e0 <_sk_callback_sse41+0x5b5> + .byte 102,68,15,219,5,198,54,0,0 // pand 0x36c6(%rip),%xmm8 # 5610 <_sk_callback_sse41+0x5b7> .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,154,54,0,0 // mulps 0x369a(%rip),%xmm8 # 55f0 <_sk_callback_sse41+0x5c5> + .byte 68,15,89,5,202,54,0,0 // mulps 0x36ca(%rip),%xmm8 # 5620 <_sk_callback_sse41+0x5c7> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -24347,17 +24375,17 @@ _sk_lerp_565_sse41: .byte 77,133,192 // test %r8,%r8 .byte 15,133,152,0,0,0 // jne 2075 <_sk_lerp_565_sse41+0xa6> .byte 102,69,15,56,51,12,82 // pmovzxwd (%r10,%rdx,2),%xmm9 - .byte 102,68,15,111,5,19,54,0,0 // movdqa 0x3613(%rip),%xmm8 # 5600 <_sk_callback_sse41+0x5d5> + .byte 102,68,15,111,5,67,54,0,0 // movdqa 0x3643(%rip),%xmm8 # 5630 <_sk_callback_sse41+0x5d7> .byte 102,69,15,219,193 // pand %xmm9,%xmm8 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,18,54,0,0 // mulps 0x3612(%rip),%xmm8 # 5610 <_sk_callback_sse41+0x5e5> - .byte 102,68,15,111,21,25,54,0,0 // movdqa 0x3619(%rip),%xmm10 # 5620 <_sk_callback_sse41+0x5f5> + .byte 68,15,89,5,66,54,0,0 // mulps 0x3642(%rip),%xmm8 # 5640 <_sk_callback_sse41+0x5e7> + .byte 102,68,15,111,21,73,54,0,0 // movdqa 0x3649(%rip),%xmm10 # 5650 <_sk_callback_sse41+0x5f7> .byte 102,69,15,219,209 // pand %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 - .byte 68,15,89,21,24,54,0,0 // mulps 0x3618(%rip),%xmm10 # 5630 <_sk_callback_sse41+0x605> - .byte 102,68,15,219,13,31,54,0,0 // pand 0x361f(%rip),%xmm9 # 5640 <_sk_callback_sse41+0x615> + .byte 68,15,89,21,72,54,0,0 // mulps 0x3648(%rip),%xmm10 # 5660 <_sk_callback_sse41+0x607> + .byte 102,68,15,219,13,79,54,0,0 // pand 0x364f(%rip),%xmm9 # 5670 <_sk_callback_sse41+0x617> .byte 69,15,91,201 // cvtdq2ps %xmm9,%xmm9 - .byte 68,15,89,13,35,54,0,0 // mulps 0x3623(%rip),%xmm9 # 5650 <_sk_callback_sse41+0x625> + .byte 68,15,89,13,83,54,0,0 // mulps 0x3653(%rip),%xmm9 # 5680 <_sk_callback_sse41+0x627> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -24409,7 +24437,7 @@ _sk_load_tables_sse41: .byte 65,87 // push %r15 .byte 65,86 // push %r14 .byte 83 // push %rbx - .byte 102,15,111,5,121,53,0,0 // movdqa 0x3579(%rip),%xmm0 # 5660 <_sk_callback_sse41+0x635> + .byte 102,15,111,5,169,53,0,0 // movdqa 0x35a9(%rip),%xmm0 # 5690 <_sk_callback_sse41+0x637> .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9 .byte 102,73,15,126,194 // movq %xmm0,%r10 @@ -24424,7 +24452,7 @@ _sk_load_tables_sse41: .byte 102,66,15,58,33,4,179,32 // insertps $0x20,(%rbx,%r14,4),%xmm0 .byte 102,66,15,58,33,4,11,48 // insertps $0x30,(%rbx,%r9,1),%xmm0 .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1 - .byte 102,15,56,0,13,52,53,0,0 // pshufb 0x3534(%rip),%xmm1 # 5670 <_sk_callback_sse41+0x645> + .byte 102,15,56,0,13,100,53,0,0 // pshufb 0x3564(%rip),%xmm1 # 56a0 <_sk_callback_sse41+0x647> .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 102,72,15,126,203 // movq %xmm1,%rbx .byte 68,15,182,211 // movzbl %bl,%r10d @@ -24439,7 +24467,7 @@ _sk_load_tables_sse41: .byte 102,15,58,33,202,48 // insertps $0x30,%xmm2,%xmm1 .byte 76,139,72,24 // mov 0x18(%rax),%r9 .byte 102,65,15,111,208 // movdqa %xmm8,%xmm2 - .byte 102,15,56,0,21,240,52,0,0 // pshufb 0x34f0(%rip),%xmm2 # 5680 <_sk_callback_sse41+0x655> + .byte 102,15,56,0,21,32,53,0,0 // pshufb 0x3520(%rip),%xmm2 # 56b0 <_sk_callback_sse41+0x657> .byte 102,72,15,58,22,211,1 // pextrq $0x1,%xmm2,%rbx .byte 102,72,15,126,208 // movq %xmm2,%rax .byte 68,15,182,208 // movzbl %al,%r10d @@ -24454,7 +24482,7 @@ _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 15,89,29,173,52,0,0 // mulps 0x34ad(%rip),%xmm3 # 5690 <_sk_callback_sse41+0x665> + .byte 15,89,29,221,52,0,0 // mulps 0x34dd(%rip),%xmm3 # 56c0 <_sk_callback_sse41+0x667> .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx .byte 65,94 // pop %r14 @@ -24495,7 +24523,7 @@ _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 102,68,15,111,5,34,52,0,0 // movdqa 0x3422(%rip),%xmm8 # 56a0 <_sk_callback_sse41+0x675> + .byte 102,68,15,111,5,82,52,0,0 // movdqa 0x3452(%rip),%xmm8 # 56d0 <_sk_callback_sse41+0x677> .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 @@ -24513,7 +24541,7 @@ _sk_load_tables_u16_be_sse41: .byte 102,15,58,33,194,32 // insertps $0x20,%xmm2,%xmm0 .byte 243,66,15,16,20,11 // movss (%rbx,%r9,1),%xmm2 .byte 102,15,58,33,194,48 // insertps $0x30,%xmm2,%xmm0 - .byte 102,15,56,0,13,209,51,0,0 // pshufb 0x33d1(%rip),%xmm1 # 56b0 <_sk_callback_sse41+0x685> + .byte 102,15,56,0,13,1,52,0,0 // pshufb 0x3401(%rip),%xmm1 # 56e0 <_sk_callback_sse41+0x687> .byte 102,15,56,51,201 // pmovzxwd %xmm1,%xmm1 .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 102,72,15,126,203 // movq %xmm1,%rbx @@ -24549,7 +24577,7 @@ _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 15,89,29,31,51,0,0 // mulps 0x331f(%rip),%xmm3 # 56c0 <_sk_callback_sse41+0x695> + .byte 15,89,29,79,51,0,0 // mulps 0x334f(%rip),%xmm3 # 56f0 <_sk_callback_sse41+0x697> .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx .byte 65,94 // pop %r14 @@ -24591,7 +24619,7 @@ _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 102,68,15,111,5,148,50,0,0 // movdqa 0x3294(%rip),%xmm8 # 56d0 <_sk_callback_sse41+0x6a5> + .byte 102,68,15,111,5,196,50,0,0 // movdqa 0x32c4(%rip),%xmm8 # 5700 <_sk_callback_sse41+0x6a7> .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 @@ -24609,7 +24637,7 @@ _sk_load_tables_rgb_u16_be_sse41: .byte 102,15,58,33,195,32 // insertps $0x20,%xmm3,%xmm0 .byte 243,66,15,16,28,11 // movss (%rbx,%r9,1),%xmm3 .byte 102,15,58,33,195,48 // insertps $0x30,%xmm3,%xmm0 - .byte 102,15,56,0,13,67,50,0,0 // pshufb 0x3243(%rip),%xmm1 # 56e0 <_sk_callback_sse41+0x6b5> + .byte 102,15,56,0,13,115,50,0,0 // pshufb 0x3273(%rip),%xmm1 # 5710 <_sk_callback_sse41+0x6b7> .byte 102,15,56,51,201 // pmovzxwd %xmm1,%xmm1 .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 102,72,15,126,203 // movq %xmm1,%rbx @@ -24640,7 +24668,7 @@ _sk_load_tables_rgb_u16_be_sse41: .byte 243,65,15,16,28,25 // movss (%r9,%rbx,1),%xmm3 .byte 102,15,58,33,211,48 // insertps $0x30,%xmm3,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,174,49,0,0 // movaps 0x31ae(%rip),%xmm3 # 56f0 <_sk_callback_sse41+0x6c5> + .byte 15,40,29,222,49,0,0 // movaps 0x31de(%rip),%xmm3 # 5720 <_sk_callback_sse41+0x6c7> .byte 91 // pop %rbx .byte 65,94 // pop %r14 .byte 65,95 // pop %r15 @@ -24674,7 +24702,7 @@ _sk_byte_tables_sse41: .byte 65,84 // push %r12 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,70,49,0,0 // movaps 0x3146(%rip),%xmm8 # 5700 <_sk_callback_sse41+0x6d5> + .byte 68,15,40,5,118,49,0,0 // movaps 0x3176(%rip),%xmm8 # 5730 <_sk_callback_sse41+0x6d7> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9 @@ -24693,7 +24721,7 @@ _sk_byte_tables_sse41: .byte 102,15,58,32,195,3 // pinsrb $0x3,%ebx,%xmm0 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,13,247,48,0,0 // movaps 0x30f7(%rip),%xmm9 # 5710 <_sk_callback_sse41+0x6e5> + .byte 68,15,40,13,39,49,0,0 // movaps 0x3127(%rip),%xmm9 # 5740 <_sk_callback_sse41+0x6e7> .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -24789,7 +24817,7 @@ _sk_byte_tables_rgb_sse41: .byte 102,15,58,32,195,3 // pinsrb $0x3,%ebx,%xmm0 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,13,111,47,0,0 // movaps 0x2f6f(%rip),%xmm9 # 5720 <_sk_callback_sse41+0x6f5> + .byte 68,15,40,13,159,47,0,0 // movaps 0x2f9f(%rip),%xmm9 # 5750 <_sk_callback_sse41+0x6f7> .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -24976,31 +25004,31 @@ _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 68,15,89,5,183,44,0,0 // mulps 0x2cb7(%rip),%xmm8 # 5730 <_sk_callback_sse41+0x705> - .byte 68,15,84,21,191,44,0,0 // andps 0x2cbf(%rip),%xmm10 # 5740 <_sk_callback_sse41+0x715> - .byte 68,15,86,21,199,44,0,0 // orps 0x2cc7(%rip),%xmm10 # 5750 <_sk_callback_sse41+0x725> - .byte 68,15,88,5,207,44,0,0 // addps 0x2ccf(%rip),%xmm8 # 5760 <_sk_callback_sse41+0x735> - .byte 68,15,40,37,215,44,0,0 // movaps 0x2cd7(%rip),%xmm12 # 5770 <_sk_callback_sse41+0x745> + .byte 68,15,89,5,231,44,0,0 // mulps 0x2ce7(%rip),%xmm8 # 5760 <_sk_callback_sse41+0x707> + .byte 68,15,84,21,239,44,0,0 // andps 0x2cef(%rip),%xmm10 # 5770 <_sk_callback_sse41+0x717> + .byte 68,15,86,21,247,44,0,0 // orps 0x2cf7(%rip),%xmm10 # 5780 <_sk_callback_sse41+0x727> + .byte 68,15,88,5,255,44,0,0 // addps 0x2cff(%rip),%xmm8 # 5790 <_sk_callback_sse41+0x737> + .byte 68,15,40,37,7,45,0,0 // movaps 0x2d07(%rip),%xmm12 # 57a0 <_sk_callback_sse41+0x747> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,196 // subps %xmm12,%xmm8 - .byte 68,15,88,21,215,44,0,0 // addps 0x2cd7(%rip),%xmm10 # 5780 <_sk_callback_sse41+0x755> - .byte 68,15,40,37,223,44,0,0 // movaps 0x2cdf(%rip),%xmm12 # 5790 <_sk_callback_sse41+0x765> + .byte 68,15,88,21,7,45,0,0 // addps 0x2d07(%rip),%xmm10 # 57b0 <_sk_callback_sse41+0x757> + .byte 68,15,40,37,15,45,0,0 // movaps 0x2d0f(%rip),%xmm12 # 57c0 <_sk_callback_sse41+0x767> .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 68,15,88,5,204,44,0,0 // addps 0x2ccc(%rip),%xmm8 # 57a0 <_sk_callback_sse41+0x775> - .byte 68,15,40,21,212,44,0,0 // movaps 0x2cd4(%rip),%xmm10 # 57b0 <_sk_callback_sse41+0x785> + .byte 68,15,88,5,252,44,0,0 // addps 0x2cfc(%rip),%xmm8 # 57d0 <_sk_callback_sse41+0x777> + .byte 68,15,40,21,4,45,0,0 // movaps 0x2d04(%rip),%xmm10 # 57e0 <_sk_callback_sse41+0x787> .byte 69,15,89,211 // mulps %xmm11,%xmm10 .byte 69,15,92,194 // subps %xmm10,%xmm8 - .byte 68,15,40,21,212,44,0,0 // movaps 0x2cd4(%rip),%xmm10 # 57c0 <_sk_callback_sse41+0x795> + .byte 68,15,40,21,4,45,0,0 // movaps 0x2d04(%rip),%xmm10 # 57f0 <_sk_callback_sse41+0x797> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 68,15,40,29,216,44,0,0 // movaps 0x2cd8(%rip),%xmm11 # 57d0 <_sk_callback_sse41+0x7a5> + .byte 68,15,40,29,8,45,0,0 // movaps 0x2d08(%rip),%xmm11 # 5800 <_sk_callback_sse41+0x7a7> .byte 69,15,94,218 // divps %xmm10,%xmm11 .byte 69,15,88,216 // addps %xmm8,%xmm11 - .byte 68,15,89,29,216,44,0,0 // mulps 0x2cd8(%rip),%xmm11 # 57e0 <_sk_callback_sse41+0x7b5> + .byte 68,15,89,29,8,45,0,0 // mulps 0x2d08(%rip),%xmm11 # 5810 <_sk_callback_sse41+0x7b7> .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 @@ -25008,7 +25036,7 @@ _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 68,15,93,5,191,44,0,0 // minps 0x2cbf(%rip),%xmm8 # 57f0 <_sk_callback_sse41+0x7c5> + .byte 68,15,93,5,239,44,0,0 // minps 0x2cef(%rip),%xmm8 # 5820 <_sk_callback_sse41+0x7c7> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -25038,31 +25066,31 @@ _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 68,15,89,37,96,44,0,0 // mulps 0x2c60(%rip),%xmm12 # 5800 <_sk_callback_sse41+0x7d5> - .byte 68,15,84,29,104,44,0,0 // andps 0x2c68(%rip),%xmm11 # 5810 <_sk_callback_sse41+0x7e5> - .byte 68,15,86,29,112,44,0,0 // orps 0x2c70(%rip),%xmm11 # 5820 <_sk_callback_sse41+0x7f5> - .byte 68,15,88,37,120,44,0,0 // addps 0x2c78(%rip),%xmm12 # 5830 <_sk_callback_sse41+0x805> - .byte 15,40,13,129,44,0,0 // movaps 0x2c81(%rip),%xmm1 # 5840 <_sk_callback_sse41+0x815> + .byte 68,15,89,37,144,44,0,0 // mulps 0x2c90(%rip),%xmm12 # 5830 <_sk_callback_sse41+0x7d7> + .byte 68,15,84,29,152,44,0,0 // andps 0x2c98(%rip),%xmm11 # 5840 <_sk_callback_sse41+0x7e7> + .byte 68,15,86,29,160,44,0,0 // orps 0x2ca0(%rip),%xmm11 # 5850 <_sk_callback_sse41+0x7f7> + .byte 68,15,88,37,168,44,0,0 // addps 0x2ca8(%rip),%xmm12 # 5860 <_sk_callback_sse41+0x807> + .byte 15,40,13,177,44,0,0 // movaps 0x2cb1(%rip),%xmm1 # 5870 <_sk_callback_sse41+0x817> .byte 65,15,89,203 // mulps %xmm11,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 - .byte 68,15,88,29,129,44,0,0 // addps 0x2c81(%rip),%xmm11 # 5850 <_sk_callback_sse41+0x825> - .byte 15,40,13,138,44,0,0 // movaps 0x2c8a(%rip),%xmm1 # 5860 <_sk_callback_sse41+0x835> + .byte 68,15,88,29,177,44,0,0 // addps 0x2cb1(%rip),%xmm11 # 5880 <_sk_callback_sse41+0x827> + .byte 15,40,13,186,44,0,0 // movaps 0x2cba(%rip),%xmm1 # 5890 <_sk_callback_sse41+0x837> .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 68,15,88,37,119,44,0,0 // addps 0x2c77(%rip),%xmm12 # 5870 <_sk_callback_sse41+0x845> - .byte 15,40,13,128,44,0,0 // movaps 0x2c80(%rip),%xmm1 # 5880 <_sk_callback_sse41+0x855> + .byte 68,15,88,37,167,44,0,0 // addps 0x2ca7(%rip),%xmm12 # 58a0 <_sk_callback_sse41+0x847> + .byte 15,40,13,176,44,0,0 // movaps 0x2cb0(%rip),%xmm1 # 58b0 <_sk_callback_sse41+0x857> .byte 65,15,89,203 // mulps %xmm11,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 - .byte 68,15,40,21,128,44,0,0 // movaps 0x2c80(%rip),%xmm10 # 5890 <_sk_callback_sse41+0x865> + .byte 68,15,40,21,176,44,0,0 // movaps 0x2cb0(%rip),%xmm10 # 58c0 <_sk_callback_sse41+0x867> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 15,40,13,133,44,0,0 // movaps 0x2c85(%rip),%xmm1 # 58a0 <_sk_callback_sse41+0x875> + .byte 15,40,13,181,44,0,0 // movaps 0x2cb5(%rip),%xmm1 # 58d0 <_sk_callback_sse41+0x877> .byte 65,15,94,202 // divps %xmm10,%xmm1 .byte 65,15,88,204 // addps %xmm12,%xmm1 - .byte 15,89,13,134,44,0,0 // mulps 0x2c86(%rip),%xmm1 # 58b0 <_sk_callback_sse41+0x885> + .byte 15,89,13,182,44,0,0 // mulps 0x2cb6(%rip),%xmm1 # 58e0 <_sk_callback_sse41+0x887> .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 @@ -25070,7 +25098,7 @@ _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 15,93,13,113,44,0,0 // minps 0x2c71(%rip),%xmm1 # 58c0 <_sk_callback_sse41+0x895> + .byte 15,93,13,161,44,0,0 // minps 0x2ca1(%rip),%xmm1 # 58f0 <_sk_callback_sse41+0x897> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -25100,31 +25128,31 @@ _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 68,15,89,37,18,44,0,0 // mulps 0x2c12(%rip),%xmm12 # 58d0 <_sk_callback_sse41+0x8a5> - .byte 68,15,84,29,26,44,0,0 // andps 0x2c1a(%rip),%xmm11 # 58e0 <_sk_callback_sse41+0x8b5> - .byte 68,15,86,29,34,44,0,0 // orps 0x2c22(%rip),%xmm11 # 58f0 <_sk_callback_sse41+0x8c5> - .byte 68,15,88,37,42,44,0,0 // addps 0x2c2a(%rip),%xmm12 # 5900 <_sk_callback_sse41+0x8d5> - .byte 15,40,21,51,44,0,0 // movaps 0x2c33(%rip),%xmm2 # 5910 <_sk_callback_sse41+0x8e5> + .byte 68,15,89,37,66,44,0,0 // mulps 0x2c42(%rip),%xmm12 # 5900 <_sk_callback_sse41+0x8a7> + .byte 68,15,84,29,74,44,0,0 // andps 0x2c4a(%rip),%xmm11 # 5910 <_sk_callback_sse41+0x8b7> + .byte 68,15,86,29,82,44,0,0 // orps 0x2c52(%rip),%xmm11 # 5920 <_sk_callback_sse41+0x8c7> + .byte 68,15,88,37,90,44,0,0 // addps 0x2c5a(%rip),%xmm12 # 5930 <_sk_callback_sse41+0x8d7> + .byte 15,40,21,99,44,0,0 // movaps 0x2c63(%rip),%xmm2 # 5940 <_sk_callback_sse41+0x8e7> .byte 65,15,89,211 // mulps %xmm11,%xmm2 .byte 68,15,92,226 // subps %xmm2,%xmm12 - .byte 68,15,88,29,51,44,0,0 // addps 0x2c33(%rip),%xmm11 # 5920 <_sk_callback_sse41+0x8f5> - .byte 15,40,21,60,44,0,0 // movaps 0x2c3c(%rip),%xmm2 # 5930 <_sk_callback_sse41+0x905> + .byte 68,15,88,29,99,44,0,0 // addps 0x2c63(%rip),%xmm11 # 5950 <_sk_callback_sse41+0x8f7> + .byte 15,40,21,108,44,0,0 // movaps 0x2c6c(%rip),%xmm2 # 5960 <_sk_callback_sse41+0x907> .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 68,15,88,37,41,44,0,0 // addps 0x2c29(%rip),%xmm12 # 5940 <_sk_callback_sse41+0x915> - .byte 15,40,21,50,44,0,0 // movaps 0x2c32(%rip),%xmm2 # 5950 <_sk_callback_sse41+0x925> + .byte 68,15,88,37,89,44,0,0 // addps 0x2c59(%rip),%xmm12 # 5970 <_sk_callback_sse41+0x917> + .byte 15,40,21,98,44,0,0 // movaps 0x2c62(%rip),%xmm2 # 5980 <_sk_callback_sse41+0x927> .byte 65,15,89,211 // mulps %xmm11,%xmm2 .byte 68,15,92,226 // subps %xmm2,%xmm12 - .byte 68,15,40,21,50,44,0,0 // movaps 0x2c32(%rip),%xmm10 # 5960 <_sk_callback_sse41+0x935> + .byte 68,15,40,21,98,44,0,0 // movaps 0x2c62(%rip),%xmm10 # 5990 <_sk_callback_sse41+0x937> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 15,40,21,55,44,0,0 // movaps 0x2c37(%rip),%xmm2 # 5970 <_sk_callback_sse41+0x945> + .byte 15,40,21,103,44,0,0 // movaps 0x2c67(%rip),%xmm2 # 59a0 <_sk_callback_sse41+0x947> .byte 65,15,94,210 // divps %xmm10,%xmm2 .byte 65,15,88,212 // addps %xmm12,%xmm2 - .byte 15,89,21,56,44,0,0 // mulps 0x2c38(%rip),%xmm2 # 5980 <_sk_callback_sse41+0x955> + .byte 15,89,21,104,44,0,0 // mulps 0x2c68(%rip),%xmm2 # 59b0 <_sk_callback_sse41+0x957> .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 @@ -25132,7 +25160,7 @@ _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 15,93,21,35,44,0,0 // minps 0x2c23(%rip),%xmm2 # 5990 <_sk_callback_sse41+0x965> + .byte 15,93,21,83,44,0,0 // minps 0x2c53(%rip),%xmm2 # 59c0 <_sk_callback_sse41+0x967> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -25162,31 +25190,31 @@ _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 68,15,89,37,196,43,0,0 // mulps 0x2bc4(%rip),%xmm12 # 59a0 <_sk_callback_sse41+0x975> - .byte 68,15,84,29,204,43,0,0 // andps 0x2bcc(%rip),%xmm11 # 59b0 <_sk_callback_sse41+0x985> - .byte 68,15,86,29,212,43,0,0 // orps 0x2bd4(%rip),%xmm11 # 59c0 <_sk_callback_sse41+0x995> - .byte 68,15,88,37,220,43,0,0 // addps 0x2bdc(%rip),%xmm12 # 59d0 <_sk_callback_sse41+0x9a5> - .byte 15,40,29,229,43,0,0 // movaps 0x2be5(%rip),%xmm3 # 59e0 <_sk_callback_sse41+0x9b5> + .byte 68,15,89,37,244,43,0,0 // mulps 0x2bf4(%rip),%xmm12 # 59d0 <_sk_callback_sse41+0x977> + .byte 68,15,84,29,252,43,0,0 // andps 0x2bfc(%rip),%xmm11 # 59e0 <_sk_callback_sse41+0x987> + .byte 68,15,86,29,4,44,0,0 // orps 0x2c04(%rip),%xmm11 # 59f0 <_sk_callback_sse41+0x997> + .byte 68,15,88,37,12,44,0,0 // addps 0x2c0c(%rip),%xmm12 # 5a00 <_sk_callback_sse41+0x9a7> + .byte 15,40,29,21,44,0,0 // movaps 0x2c15(%rip),%xmm3 # 5a10 <_sk_callback_sse41+0x9b7> .byte 65,15,89,219 // mulps %xmm11,%xmm3 .byte 68,15,92,227 // subps %xmm3,%xmm12 - .byte 68,15,88,29,229,43,0,0 // addps 0x2be5(%rip),%xmm11 # 59f0 <_sk_callback_sse41+0x9c5> - .byte 15,40,29,238,43,0,0 // movaps 0x2bee(%rip),%xmm3 # 5a00 <_sk_callback_sse41+0x9d5> + .byte 68,15,88,29,21,44,0,0 // addps 0x2c15(%rip),%xmm11 # 5a20 <_sk_callback_sse41+0x9c7> + .byte 15,40,29,30,44,0,0 // movaps 0x2c1e(%rip),%xmm3 # 5a30 <_sk_callback_sse41+0x9d7> .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 68,15,88,37,219,43,0,0 // addps 0x2bdb(%rip),%xmm12 # 5a10 <_sk_callback_sse41+0x9e5> - .byte 15,40,29,228,43,0,0 // movaps 0x2be4(%rip),%xmm3 # 5a20 <_sk_callback_sse41+0x9f5> + .byte 68,15,88,37,11,44,0,0 // addps 0x2c0b(%rip),%xmm12 # 5a40 <_sk_callback_sse41+0x9e7> + .byte 15,40,29,20,44,0,0 // movaps 0x2c14(%rip),%xmm3 # 5a50 <_sk_callback_sse41+0x9f7> .byte 65,15,89,219 // mulps %xmm11,%xmm3 .byte 68,15,92,227 // subps %xmm3,%xmm12 - .byte 68,15,40,21,228,43,0,0 // movaps 0x2be4(%rip),%xmm10 # 5a30 <_sk_callback_sse41+0xa05> + .byte 68,15,40,21,20,44,0,0 // movaps 0x2c14(%rip),%xmm10 # 5a60 <_sk_callback_sse41+0xa07> .byte 69,15,92,211 // subps %xmm11,%xmm10 - .byte 15,40,29,233,43,0,0 // movaps 0x2be9(%rip),%xmm3 # 5a40 <_sk_callback_sse41+0xa15> + .byte 15,40,29,25,44,0,0 // movaps 0x2c19(%rip),%xmm3 # 5a70 <_sk_callback_sse41+0xa17> .byte 65,15,94,218 // divps %xmm10,%xmm3 .byte 65,15,88,220 // addps %xmm12,%xmm3 - .byte 15,89,29,234,43,0,0 // mulps 0x2bea(%rip),%xmm3 # 5a50 <_sk_callback_sse41+0xa25> + .byte 15,89,29,26,44,0,0 // mulps 0x2c1a(%rip),%xmm3 # 5a80 <_sk_callback_sse41+0xa27> .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 @@ -25194,7 +25222,7 @@ _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 15,93,29,213,43,0,0 // minps 0x2bd5(%rip),%xmm3 # 5a60 <_sk_callback_sse41+0xa35> + .byte 15,93,29,5,44,0,0 // minps 0x2c05(%rip),%xmm3 # 5a90 <_sk_callback_sse41+0xa37> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -25204,29 +25232,29 @@ 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 68,15,89,5,209,43,0,0 // mulps 0x2bd1(%rip),%xmm8 # 5a70 <_sk_callback_sse41+0xa45> - .byte 68,15,40,13,217,43,0,0 // movaps 0x2bd9(%rip),%xmm9 # 5a80 <_sk_callback_sse41+0xa55> + .byte 68,15,89,5,1,44,0,0 // mulps 0x2c01(%rip),%xmm8 # 5aa0 <_sk_callback_sse41+0xa47> + .byte 68,15,40,13,9,44,0,0 // movaps 0x2c09(%rip),%xmm9 # 5ab0 <_sk_callback_sse41+0xa57> .byte 65,15,89,201 // mulps %xmm9,%xmm1 - .byte 15,40,5,222,43,0,0 // movaps 0x2bde(%rip),%xmm0 # 5a90 <_sk_callback_sse41+0xa65> + .byte 15,40,5,14,44,0,0 // movaps 0x2c0e(%rip),%xmm0 # 5ac0 <_sk_callback_sse41+0xa67> .byte 15,88,200 // addps %xmm0,%xmm1 .byte 65,15,89,209 // mulps %xmm9,%xmm2 .byte 15,88,208 // addps %xmm0,%xmm2 - .byte 68,15,88,5,220,43,0,0 // addps 0x2bdc(%rip),%xmm8 # 5aa0 <_sk_callback_sse41+0xa75> - .byte 68,15,89,5,228,43,0,0 // mulps 0x2be4(%rip),%xmm8 # 5ab0 <_sk_callback_sse41+0xa85> - .byte 15,89,13,237,43,0,0 // mulps 0x2bed(%rip),%xmm1 # 5ac0 <_sk_callback_sse41+0xa95> + .byte 68,15,88,5,12,44,0,0 // addps 0x2c0c(%rip),%xmm8 # 5ad0 <_sk_callback_sse41+0xa77> + .byte 68,15,89,5,20,44,0,0 // mulps 0x2c14(%rip),%xmm8 # 5ae0 <_sk_callback_sse41+0xa87> + .byte 15,89,13,29,44,0,0 // mulps 0x2c1d(%rip),%xmm1 # 5af0 <_sk_callback_sse41+0xa97> .byte 65,15,88,200 // addps %xmm8,%xmm1 - .byte 15,89,21,242,43,0,0 // mulps 0x2bf2(%rip),%xmm2 # 5ad0 <_sk_callback_sse41+0xaa5> + .byte 15,89,21,34,44,0,0 // mulps 0x2c22(%rip),%xmm2 # 5b00 <_sk_callback_sse41+0xaa7> .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 68,15,40,13,230,43,0,0 // movaps 0x2be6(%rip),%xmm9 # 5ae0 <_sk_callback_sse41+0xab5> + .byte 68,15,40,13,22,44,0,0 // movaps 0x2c16(%rip),%xmm9 # 5b10 <_sk_callback_sse41+0xab7> .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0 - .byte 15,40,21,230,43,0,0 // movaps 0x2be6(%rip),%xmm2 # 5af0 <_sk_callback_sse41+0xac5> + .byte 15,40,21,22,44,0,0 // movaps 0x2c16(%rip),%xmm2 # 5b20 <_sk_callback_sse41+0xac7> .byte 15,88,202 // addps %xmm2,%xmm1 - .byte 68,15,40,37,235,43,0,0 // movaps 0x2beb(%rip),%xmm12 # 5b00 <_sk_callback_sse41+0xad5> + .byte 68,15,40,37,27,44,0,0 // movaps 0x2c1b(%rip),%xmm12 # 5b30 <_sk_callback_sse41+0xad7> .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 @@ -25245,8 +25273,8 @@ _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 15,89,13,164,43,0,0 // mulps 0x2ba4(%rip),%xmm1 # 5b10 <_sk_callback_sse41+0xae5> - .byte 15,89,21,173,43,0,0 // mulps 0x2bad(%rip),%xmm2 # 5b20 <_sk_callback_sse41+0xaf5> + .byte 15,89,13,212,43,0,0 // mulps 0x2bd4(%rip),%xmm1 # 5b40 <_sk_callback_sse41+0xae7> + .byte 15,89,21,221,43,0,0 // mulps 0x2bdd(%rip),%xmm2 # 5b50 <_sk_callback_sse41+0xaf7> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,40,193 // movaps %xmm1,%xmm0 .byte 65,15,40,200 // movaps %xmm8,%xmm1 @@ -25261,9 +25289,9 @@ _sk_load_a8_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,38 // jne 2fae <_sk_load_a8_sse41+0x30> .byte 102,65,15,56,49,4,18 // pmovzxbd (%r10,%rdx,1),%xmm0 - .byte 102,15,219,5,153,43,0,0 // pand 0x2b99(%rip),%xmm0 # 5b30 <_sk_callback_sse41+0xb05> + .byte 102,15,219,5,201,43,0,0 // pand 0x2bc9(%rip),%xmm0 # 5b60 <_sk_callback_sse41+0xb07> .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,159,43,0,0 // mulps 0x2b9f(%rip),%xmm3 # 5b40 <_sk_callback_sse41+0xb15> + .byte 15,89,29,207,43,0,0 // mulps 0x2bcf(%rip),%xmm3 # 5b70 <_sk_callback_sse41+0xb17> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 15,87,201 // xorps %xmm1,%xmm1 @@ -25314,7 +25342,7 @@ _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 15,89,29,237,42,0,0 // mulps 0x2aed(%rip),%xmm3 # 5b50 <_sk_callback_sse41+0xb25> + .byte 15,89,29,29,43,0,0 // mulps 0x2b1d(%rip),%xmm3 # 5b80 <_sk_callback_sse41+0xb27> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 102,15,239,201 // pxor %xmm1,%xmm1 @@ -25328,7 +25356,7 @@ FUNCTION(_sk_store_a8_sse41) _sk_store_a8_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 68,15,40,5,224,42,0,0 // movaps 0x2ae0(%rip),%xmm8 # 5b60 <_sk_callback_sse41+0xb35> + .byte 68,15,40,5,16,43,0,0 // movaps 0x2b10(%rip),%xmm8 # 5b90 <_sk_callback_sse41+0xb37> .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 @@ -25362,11 +25390,11 @@ _sk_load_g8_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,42 // jne 3115 <_sk_load_g8_sse41+0x34> .byte 102,65,15,56,49,4,18 // pmovzxbd (%r10,%rdx,1),%xmm0 - .byte 102,15,219,5,118,42,0,0 // pand 0x2a76(%rip),%xmm0 # 5b70 <_sk_callback_sse41+0xb45> + .byte 102,15,219,5,166,42,0,0 // pand 0x2aa6(%rip),%xmm0 # 5ba0 <_sk_callback_sse41+0xb47> .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,124,42,0,0 // mulps 0x2a7c(%rip),%xmm0 # 5b80 <_sk_callback_sse41+0xb55> + .byte 15,89,5,172,42,0,0 // mulps 0x2aac(%rip),%xmm0 # 5bb0 <_sk_callback_sse41+0xb57> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,131,42,0,0 // movaps 0x2a83(%rip),%xmm3 # 5b90 <_sk_callback_sse41+0xb65> + .byte 15,40,29,179,42,0,0 // movaps 0x2ab3(%rip),%xmm3 # 5bc0 <_sk_callback_sse41+0xb67> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 255,224 // jmpq *%rax @@ -25415,9 +25443,9 @@ _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 15,89,5,214,41,0,0 // mulps 0x29d6(%rip),%xmm0 # 5ba0 <_sk_callback_sse41+0xb75> + .byte 15,89,5,6,42,0,0 // mulps 0x2a06(%rip),%xmm0 # 5bd0 <_sk_callback_sse41+0xb77> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,221,41,0,0 // movaps 0x29dd(%rip),%xmm3 # 5bb0 <_sk_callback_sse41+0xb85> + .byte 15,40,29,13,42,0,0 // movaps 0x2a0d(%rip),%xmm3 # 5be0 <_sk_callback_sse41+0xb87> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 91 // pop %rbx @@ -25465,17 +25493,17 @@ _sk_gather_i8_sse41: .byte 102,15,58,34,28,24,1 // pinsrd $0x1,(%rax,%rbx,1),%xmm3 .byte 102,66,15,58,34,28,152,2 // pinsrd $0x2,(%rax,%r11,4),%xmm3 .byte 102,66,15,58,34,28,16,3 // pinsrd $0x3,(%rax,%r10,1),%xmm3 - .byte 102,15,111,5,48,41,0,0 // movdqa 0x2930(%rip),%xmm0 # 5bc0 <_sk_callback_sse41+0xb95> + .byte 102,15,111,5,96,41,0,0 // movdqa 0x2960(%rip),%xmm0 # 5bf0 <_sk_callback_sse41+0xb97> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,49,41,0,0 // movaps 0x2931(%rip),%xmm8 # 5bd0 <_sk_callback_sse41+0xba5> + .byte 68,15,40,5,97,41,0,0 // movaps 0x2961(%rip),%xmm8 # 5c00 <_sk_callback_sse41+0xba7> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,15,56,0,13,48,41,0,0 // pshufb 0x2930(%rip),%xmm1 # 5be0 <_sk_callback_sse41+0xbb5> + .byte 102,15,56,0,13,96,41,0,0 // pshufb 0x2960(%rip),%xmm1 # 5c10 <_sk_callback_sse41+0xbb7> .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,15,56,0,21,44,41,0,0 // pshufb 0x292c(%rip),%xmm2 # 5bf0 <_sk_callback_sse41+0xbc5> + .byte 102,15,56,0,21,92,41,0,0 // pshufb 0x295c(%rip),%xmm2 # 5c20 <_sk_callback_sse41+0xbc7> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 102,15,114,211,24 // psrld $0x18,%xmm3 @@ -25495,19 +25523,19 @@ _sk_load_565_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,80 // jne 3338 <_sk_load_565_sse41+0x5a> .byte 102,65,15,56,51,20,82 // pmovzxwd (%r10,%rdx,2),%xmm2 - .byte 102,15,111,5,9,41,0,0 // movdqa 0x2909(%rip),%xmm0 # 5c00 <_sk_callback_sse41+0xbd5> + .byte 102,15,111,5,57,41,0,0 // movdqa 0x2939(%rip),%xmm0 # 5c30 <_sk_callback_sse41+0xbd7> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,11,41,0,0 // mulps 0x290b(%rip),%xmm0 # 5c10 <_sk_callback_sse41+0xbe5> - .byte 102,15,111,13,19,41,0,0 // movdqa 0x2913(%rip),%xmm1 # 5c20 <_sk_callback_sse41+0xbf5> + .byte 15,89,5,59,41,0,0 // mulps 0x293b(%rip),%xmm0 # 5c40 <_sk_callback_sse41+0xbe7> + .byte 102,15,111,13,67,41,0,0 // movdqa 0x2943(%rip),%xmm1 # 5c50 <_sk_callback_sse41+0xbf7> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,21,41,0,0 // mulps 0x2915(%rip),%xmm1 # 5c30 <_sk_callback_sse41+0xc05> - .byte 102,15,219,21,29,41,0,0 // pand 0x291d(%rip),%xmm2 # 5c40 <_sk_callback_sse41+0xc15> + .byte 15,89,13,69,41,0,0 // mulps 0x2945(%rip),%xmm1 # 5c60 <_sk_callback_sse41+0xc07> + .byte 102,15,219,21,77,41,0,0 // pand 0x294d(%rip),%xmm2 # 5c70 <_sk_callback_sse41+0xc17> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,35,41,0,0 // mulps 0x2923(%rip),%xmm2 # 5c50 <_sk_callback_sse41+0xc25> + .byte 15,89,21,83,41,0,0 // mulps 0x2953(%rip),%xmm2 # 5c80 <_sk_callback_sse41+0xc27> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,42,41,0,0 // movaps 0x292a(%rip),%xmm3 # 5c60 <_sk_callback_sse41+0xc35> + .byte 15,40,29,90,41,0,0 // movaps 0x295a(%rip),%xmm3 # 5c90 <_sk_callback_sse41+0xc37> .byte 255,224 // jmpq *%rax .byte 69,137,193 // mov %r8d,%r9d .byte 65,128,225,3 // and $0x3,%r9b @@ -25553,19 +25581,19 @@ _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 102,15,111,5,134,40,0,0 // movdqa 0x2886(%rip),%xmm0 # 5c70 <_sk_callback_sse41+0xc45> + .byte 102,15,111,5,182,40,0,0 // movdqa 0x28b6(%rip),%xmm0 # 5ca0 <_sk_callback_sse41+0xc47> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,136,40,0,0 // mulps 0x2888(%rip),%xmm0 # 5c80 <_sk_callback_sse41+0xc55> - .byte 102,15,111,13,144,40,0,0 // movdqa 0x2890(%rip),%xmm1 # 5c90 <_sk_callback_sse41+0xc65> + .byte 15,89,5,184,40,0,0 // mulps 0x28b8(%rip),%xmm0 # 5cb0 <_sk_callback_sse41+0xc57> + .byte 102,15,111,13,192,40,0,0 // movdqa 0x28c0(%rip),%xmm1 # 5cc0 <_sk_callback_sse41+0xc67> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,146,40,0,0 // mulps 0x2892(%rip),%xmm1 # 5ca0 <_sk_callback_sse41+0xc75> - .byte 102,15,219,21,154,40,0,0 // pand 0x289a(%rip),%xmm2 # 5cb0 <_sk_callback_sse41+0xc85> + .byte 15,89,13,194,40,0,0 // mulps 0x28c2(%rip),%xmm1 # 5cd0 <_sk_callback_sse41+0xc77> + .byte 102,15,219,21,202,40,0,0 // pand 0x28ca(%rip),%xmm2 # 5ce0 <_sk_callback_sse41+0xc87> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,160,40,0,0 // mulps 0x28a0(%rip),%xmm2 # 5cc0 <_sk_callback_sse41+0xc95> + .byte 15,89,21,208,40,0,0 // mulps 0x28d0(%rip),%xmm2 # 5cf0 <_sk_callback_sse41+0xc97> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,167,40,0,0 // movaps 0x28a7(%rip),%xmm3 # 5cd0 <_sk_callback_sse41+0xca5> + .byte 15,40,29,215,40,0,0 // movaps 0x28d7(%rip),%xmm3 # 5d00 <_sk_callback_sse41+0xca7> .byte 91 // pop %rbx .byte 255,224 // jmpq *%rax @@ -25575,12 +25603,12 @@ FUNCTION(_sk_store_565_sse41) _sk_store_565_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,167,40,0,0 // movaps 0x28a7(%rip),%xmm8 # 5ce0 <_sk_callback_sse41+0xcb5> + .byte 68,15,40,5,215,40,0,0 // movaps 0x28d7(%rip),%xmm8 # 5d10 <_sk_callback_sse41+0xcb7> .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 68,15,40,21,156,40,0,0 // movaps 0x289c(%rip),%xmm10 # 5cf0 <_sk_callback_sse41+0xcc5> + .byte 68,15,40,21,204,40,0,0 // movaps 0x28cc(%rip),%xmm10 # 5d20 <_sk_callback_sse41+0xcc7> .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 @@ -25617,21 +25645,21 @@ _sk_load_4444_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,95 // jne 352f <_sk_load_4444_sse41+0x69> .byte 102,65,15,56,51,28,82 // pmovzxwd (%r10,%rdx,2),%xmm3 - .byte 102,15,111,5,33,40,0,0 // movdqa 0x2821(%rip),%xmm0 # 5d00 <_sk_callback_sse41+0xcd5> + .byte 102,15,111,5,81,40,0,0 // movdqa 0x2851(%rip),%xmm0 # 5d30 <_sk_callback_sse41+0xcd7> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,35,40,0,0 // mulps 0x2823(%rip),%xmm0 # 5d10 <_sk_callback_sse41+0xce5> - .byte 102,15,111,13,43,40,0,0 // movdqa 0x282b(%rip),%xmm1 # 5d20 <_sk_callback_sse41+0xcf5> + .byte 15,89,5,83,40,0,0 // mulps 0x2853(%rip),%xmm0 # 5d40 <_sk_callback_sse41+0xce7> + .byte 102,15,111,13,91,40,0,0 // movdqa 0x285b(%rip),%xmm1 # 5d50 <_sk_callback_sse41+0xcf7> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,45,40,0,0 // mulps 0x282d(%rip),%xmm1 # 5d30 <_sk_callback_sse41+0xd05> - .byte 102,15,111,21,53,40,0,0 // movdqa 0x2835(%rip),%xmm2 # 5d40 <_sk_callback_sse41+0xd15> + .byte 15,89,13,93,40,0,0 // mulps 0x285d(%rip),%xmm1 # 5d60 <_sk_callback_sse41+0xd07> + .byte 102,15,111,21,101,40,0,0 // movdqa 0x2865(%rip),%xmm2 # 5d70 <_sk_callback_sse41+0xd17> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,55,40,0,0 // mulps 0x2837(%rip),%xmm2 # 5d50 <_sk_callback_sse41+0xd25> - .byte 102,15,219,29,63,40,0,0 // pand 0x283f(%rip),%xmm3 # 5d60 <_sk_callback_sse41+0xd35> + .byte 15,89,21,103,40,0,0 // mulps 0x2867(%rip),%xmm2 # 5d80 <_sk_callback_sse41+0xd27> + .byte 102,15,219,29,111,40,0,0 // pand 0x286f(%rip),%xmm3 # 5d90 <_sk_callback_sse41+0xd37> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,69,40,0,0 // mulps 0x2845(%rip),%xmm3 # 5d70 <_sk_callback_sse41+0xd45> + .byte 15,89,29,117,40,0,0 // mulps 0x2875(%rip),%xmm3 # 5da0 <_sk_callback_sse41+0xd47> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax .byte 69,137,193 // mov %r8d,%r9d @@ -25678,21 +25706,21 @@ _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 102,15,111,5,159,39,0,0 // movdqa 0x279f(%rip),%xmm0 # 5d80 <_sk_callback_sse41+0xd55> + .byte 102,15,111,5,207,39,0,0 // movdqa 0x27cf(%rip),%xmm0 # 5db0 <_sk_callback_sse41+0xd57> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,161,39,0,0 // mulps 0x27a1(%rip),%xmm0 # 5d90 <_sk_callback_sse41+0xd65> - .byte 102,15,111,13,169,39,0,0 // movdqa 0x27a9(%rip),%xmm1 # 5da0 <_sk_callback_sse41+0xd75> + .byte 15,89,5,209,39,0,0 // mulps 0x27d1(%rip),%xmm0 # 5dc0 <_sk_callback_sse41+0xd67> + .byte 102,15,111,13,217,39,0,0 // movdqa 0x27d9(%rip),%xmm1 # 5dd0 <_sk_callback_sse41+0xd77> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,171,39,0,0 // mulps 0x27ab(%rip),%xmm1 # 5db0 <_sk_callback_sse41+0xd85> - .byte 102,15,111,21,179,39,0,0 // movdqa 0x27b3(%rip),%xmm2 # 5dc0 <_sk_callback_sse41+0xd95> + .byte 15,89,13,219,39,0,0 // mulps 0x27db(%rip),%xmm1 # 5de0 <_sk_callback_sse41+0xd87> + .byte 102,15,111,21,227,39,0,0 // movdqa 0x27e3(%rip),%xmm2 # 5df0 <_sk_callback_sse41+0xd97> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,181,39,0,0 // mulps 0x27b5(%rip),%xmm2 # 5dd0 <_sk_callback_sse41+0xda5> - .byte 102,15,219,29,189,39,0,0 // pand 0x27bd(%rip),%xmm3 # 5de0 <_sk_callback_sse41+0xdb5> + .byte 15,89,21,229,39,0,0 // mulps 0x27e5(%rip),%xmm2 # 5e00 <_sk_callback_sse41+0xda7> + .byte 102,15,219,29,237,39,0,0 // pand 0x27ed(%rip),%xmm3 # 5e10 <_sk_callback_sse41+0xdb7> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,195,39,0,0 // mulps 0x27c3(%rip),%xmm3 # 5df0 <_sk_callback_sse41+0xdc5> + .byte 15,89,29,243,39,0,0 // mulps 0x27f3(%rip),%xmm3 # 5e20 <_sk_callback_sse41+0xdc7> .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx .byte 255,224 // jmpq *%rax @@ -25703,7 +25731,7 @@ FUNCTION(_sk_store_4444_sse41) _sk_store_4444_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,193,39,0,0 // movaps 0x27c1(%rip),%xmm8 # 5e00 <_sk_callback_sse41+0xdd5> + .byte 68,15,40,5,241,39,0,0 // movaps 0x27f1(%rip),%xmm8 # 5e30 <_sk_callback_sse41+0xdd7> .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 @@ -25750,17 +25778,17 @@ _sk_load_8888_sse41: .byte 77,133,192 // test %r8,%r8 .byte 117,88 // jne 3742 <_sk_load_8888_sse41+0x62> .byte 243,15,111,28,144 // movdqu (%rax,%rdx,4),%xmm3 - .byte 102,15,111,5,25,39,0,0 // movdqa 0x2719(%rip),%xmm0 # 5e10 <_sk_callback_sse41+0xde5> + .byte 102,15,111,5,73,39,0,0 // movdqa 0x2749(%rip),%xmm0 # 5e40 <_sk_callback_sse41+0xde7> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,26,39,0,0 // movaps 0x271a(%rip),%xmm8 # 5e20 <_sk_callback_sse41+0xdf5> + .byte 68,15,40,5,74,39,0,0 // movaps 0x274a(%rip),%xmm8 # 5e50 <_sk_callback_sse41+0xdf7> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,15,56,0,13,25,39,0,0 // pshufb 0x2719(%rip),%xmm1 # 5e30 <_sk_callback_sse41+0xe05> + .byte 102,15,56,0,13,73,39,0,0 // pshufb 0x2749(%rip),%xmm1 # 5e60 <_sk_callback_sse41+0xe07> .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,15,56,0,21,21,39,0,0 // pshufb 0x2715(%rip),%xmm2 # 5e40 <_sk_callback_sse41+0xe15> + .byte 102,15,56,0,21,69,39,0,0 // pshufb 0x2745(%rip),%xmm2 # 5e70 <_sk_callback_sse41+0xe17> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 102,15,114,211,24 // psrld $0x18,%xmm3 @@ -25806,17 +25834,17 @@ _sk_gather_8888_sse41: .byte 102,65,15,58,34,28,129,1 // pinsrd $0x1,(%r9,%rax,4),%xmm3 .byte 102,65,15,58,34,28,153,2 // pinsrd $0x2,(%r9,%rbx,4),%xmm3 .byte 102,67,15,58,34,28,153,3 // pinsrd $0x3,(%r9,%r11,4),%xmm3 - .byte 102,15,111,5,113,38,0,0 // movdqa 0x2671(%rip),%xmm0 # 5e50 <_sk_callback_sse41+0xe25> + .byte 102,15,111,5,161,38,0,0 // movdqa 0x26a1(%rip),%xmm0 # 5e80 <_sk_callback_sse41+0xe27> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,114,38,0,0 // movaps 0x2672(%rip),%xmm8 # 5e60 <_sk_callback_sse41+0xe35> + .byte 68,15,40,5,162,38,0,0 // movaps 0x26a2(%rip),%xmm8 # 5e90 <_sk_callback_sse41+0xe37> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 - .byte 102,15,56,0,13,113,38,0,0 // pshufb 0x2671(%rip),%xmm1 # 5e70 <_sk_callback_sse41+0xe45> + .byte 102,15,56,0,13,161,38,0,0 // pshufb 0x26a1(%rip),%xmm1 # 5ea0 <_sk_callback_sse41+0xe47> .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,111,211 // movdqa %xmm3,%xmm2 - .byte 102,15,56,0,21,109,38,0,0 // pshufb 0x266d(%rip),%xmm2 # 5e80 <_sk_callback_sse41+0xe55> + .byte 102,15,56,0,21,157,38,0,0 // pshufb 0x269d(%rip),%xmm2 # 5eb0 <_sk_callback_sse41+0xe57> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 102,15,114,211,24 // psrld $0x18,%xmm3 @@ -25832,7 +25860,7 @@ FUNCTION(_sk_store_8888_sse41) _sk_store_8888_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,88,38,0,0 // movaps 0x2658(%rip),%xmm8 # 5e90 <_sk_callback_sse41+0xe65> + .byte 68,15,40,5,136,38,0,0 // movaps 0x2688(%rip),%xmm8 # 5ec0 <_sk_callback_sse41+0xe67> .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 @@ -25885,18 +25913,18 @@ _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 102,68,15,111,5,144,37,0,0 // movdqa 0x2590(%rip),%xmm8 # 5ea0 <_sk_callback_sse41+0xe75> + .byte 102,68,15,111,5,192,37,0,0 // movdqa 0x25c0(%rip),%xmm8 # 5ed0 <_sk_callback_sse41+0xe77> .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 102,15,111,29,139,37,0,0 // movdqa 0x258b(%rip),%xmm3 # 5eb0 <_sk_callback_sse41+0xe85> + .byte 102,15,111,29,187,37,0,0 // movdqa 0x25bb(%rip),%xmm3 # 5ee0 <_sk_callback_sse41+0xe87> .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 102,68,15,111,21,119,37,0,0 // movdqa 0x2577(%rip),%xmm10 # 5ec0 <_sk_callback_sse41+0xe95> + .byte 102,68,15,111,21,167,37,0,0 // movdqa 0x25a7(%rip),%xmm10 # 5ef0 <_sk_callback_sse41+0xe97> .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 @@ -25982,18 +26010,18 @@ _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 102,68,15,111,5,253,35,0,0 // movdqa 0x23fd(%rip),%xmm8 # 5ed0 <_sk_callback_sse41+0xea5> + .byte 102,68,15,111,5,45,36,0,0 // movdqa 0x242d(%rip),%xmm8 # 5f00 <_sk_callback_sse41+0xea7> .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 102,15,111,29,248,35,0,0 // movdqa 0x23f8(%rip),%xmm3 # 5ee0 <_sk_callback_sse41+0xeb5> + .byte 102,15,111,29,40,36,0,0 // movdqa 0x2428(%rip),%xmm3 # 5f10 <_sk_callback_sse41+0xeb7> .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 102,68,15,111,21,228,35,0,0 // movdqa 0x23e4(%rip),%xmm10 # 5ef0 <_sk_callback_sse41+0xec5> + .byte 102,68,15,111,21,20,36,0,0 // movdqa 0x2414(%rip),%xmm10 # 5f20 <_sk_callback_sse41+0xec7> .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 @@ -26042,17 +26070,17 @@ FUNCTION(_sk_store_f16_sse41) _sk_store_f16_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 102,68,15,111,21,25,35,0,0 // movdqa 0x2319(%rip),%xmm10 # 5f00 <_sk_callback_sse41+0xed5> + .byte 102,68,15,111,21,73,35,0,0 // movdqa 0x2349(%rip),%xmm10 # 5f30 <_sk_callback_sse41+0xed7> .byte 102,68,15,111,216 // movdqa %xmm0,%xmm11 .byte 102,69,15,219,218 // pand %xmm10,%xmm11 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13 .byte 102,69,15,239,235 // pxor %xmm11,%xmm13 - .byte 102,68,15,111,13,12,35,0,0 // movdqa 0x230c(%rip),%xmm9 # 5f10 <_sk_callback_sse41+0xee5> + .byte 102,68,15,111,13,60,35,0,0 // movdqa 0x233c(%rip),%xmm9 # 5f40 <_sk_callback_sse41+0xee7> .byte 102,65,15,114,211,16 // psrld $0x10,%xmm11 .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8 .byte 102,69,15,102,197 // pcmpgtd %xmm13,%xmm8 .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 - .byte 102,68,15,111,37,253,34,0,0 // movdqa 0x22fd(%rip),%xmm12 # 5f20 <_sk_callback_sse41+0xef5> + .byte 102,68,15,111,37,45,35,0,0 // movdqa 0x232d(%rip),%xmm12 # 5f50 <_sk_callback_sse41+0xef7> .byte 102,69,15,235,220 // por %xmm12,%xmm11 .byte 102,69,15,254,221 // paddd %xmm13,%xmm11 .byte 102,69,15,223,195 // pandn %xmm11,%xmm8 @@ -26136,7 +26164,7 @@ _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 68,15,40,5,142,33,0,0 // movaps 0x218e(%rip),%xmm8 # 5f30 <_sk_callback_sse41+0xf05> + .byte 68,15,40,5,190,33,0,0 // movaps 0x21be(%rip),%xmm8 # 5f60 <_sk_callback_sse41+0xf07> .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 @@ -26203,7 +26231,7 @@ _sk_load_rgb_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 68,15,40,5,134,32,0,0 // movaps 0x2086(%rip),%xmm8 # 5f40 <_sk_callback_sse41+0xf15> + .byte 68,15,40,5,182,32,0,0 // movaps 0x20b6(%rip),%xmm8 # 5f70 <_sk_callback_sse41+0xf17> .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 @@ -26220,7 +26248,7 @@ _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 15,40,29,77,32,0,0 // movaps 0x204d(%rip),%xmm3 # 5f50 <_sk_callback_sse41+0xf25> + .byte 15,40,29,125,32,0,0 // movaps 0x207d(%rip),%xmm3 # 5f80 <_sk_callback_sse41+0xf27> .byte 255,224 // jmpq *%rax .byte 102,65,15,110,20,65 // movd (%r9,%rax,2),%xmm2 .byte 102,65,15,196,84,65,4,2 // pinsrw $0x2,0x4(%r9,%rax,2),%xmm2 @@ -26249,7 +26277,7 @@ _sk_store_u16_be_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,8 // mov (%rax),%r9 .byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax - .byte 68,15,40,21,231,31,0,0 // movaps 0x1fe7(%rip),%xmm10 # 5f60 <_sk_callback_sse41+0xf35> + .byte 68,15,40,21,23,32,0,0 // movaps 0x2017(%rip),%xmm10 # 5f90 <_sk_callback_sse41+0xf37> .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 69,15,89,194 // mulps %xmm10,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 @@ -26422,10 +26450,11 @@ FUNCTION(_sk_repeat_x_sse41) _sk_repeat_x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,200 // movaps %xmm0,%xmm9 - .byte 69,15,94,200 // divps %xmm8,%xmm9 + .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 68,15,89,200 // mulps %xmm0,%xmm9 .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 65,15,92,193 // subps %xmm9,%xmm0 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 @@ -26440,10 +26469,11 @@ FUNCTION(_sk_repeat_y_sse41) _sk_repeat_y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,201 // movaps %xmm1,%xmm9 - .byte 69,15,94,200 // divps %xmm8,%xmm9 + .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 68,15,89,201 // mulps %xmm1,%xmm9 .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 65,15,92,201 // subps %xmm9,%xmm1 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 @@ -26458,22 +26488,24 @@ FUNCTION(_sk_mirror_x_sse41) _sk_mirror_x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,40,200 // movaps %xmm8,%xmm9 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 65,15,92,193 // subps %xmm9,%xmm0 + .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 + .byte 69,15,40,208 // movaps %xmm8,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 65,15,92,194 // subps %xmm10,%xmm0 .byte 243,69,15,88,192 // addss %xmm8,%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,208 // movaps %xmm0,%xmm10 - .byte 69,15,94,208 // divps %xmm8,%xmm10 - .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10 - .byte 69,15,89,208 // mulps %xmm8,%xmm10 - .byte 65,15,92,194 // subps %xmm10,%xmm0 + .byte 243,68,15,89,13,227,32,0,0 // mulss 0x20e3(%rip),%xmm9 # 6360 <_sk_callback_sse41+0x1307> + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 68,15,89,200 // mulps %xmm0,%xmm9 + .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 + .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 65,15,92,193 // subps %xmm9,%xmm0 + .byte 65,15,92,194 // subps %xmm10,%xmm0 .byte 69,15,87,192 // xorps %xmm8,%xmm8 .byte 68,15,92,192 // subps %xmm0,%xmm8 .byte 65,15,84,192 // andps %xmm8,%xmm0 .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 - .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 + .byte 102,69,15,254,194 // paddd %xmm10,%xmm8 .byte 65,15,93,192 // minps %xmm8,%xmm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -26484,22 +26516,24 @@ FUNCTION(_sk_mirror_y_sse41) _sk_mirror_y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,40,200 // movaps %xmm8,%xmm9 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 65,15,92,201 // subps %xmm9,%xmm1 + .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 + .byte 69,15,40,208 // movaps %xmm8,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 65,15,92,202 // subps %xmm10,%xmm1 .byte 243,69,15,88,192 // addss %xmm8,%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,209 // movaps %xmm1,%xmm10 - .byte 69,15,94,208 // divps %xmm8,%xmm10 - .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10 - .byte 69,15,89,208 // mulps %xmm8,%xmm10 - .byte 65,15,92,202 // subps %xmm10,%xmm1 + .byte 243,68,15,89,13,128,32,0,0 // mulss 0x2080(%rip),%xmm9 # 6364 <_sk_callback_sse41+0x130b> + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 68,15,89,201 // mulps %xmm1,%xmm9 + .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 + .byte 69,15,89,200 // mulps %xmm8,%xmm9 .byte 65,15,92,201 // subps %xmm9,%xmm1 + .byte 65,15,92,202 // subps %xmm10,%xmm1 .byte 69,15,87,192 // xorps %xmm8,%xmm8 .byte 68,15,92,193 // subps %xmm1,%xmm8 .byte 65,15,84,200 // andps %xmm8,%xmm1 .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 - .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 + .byte 102,69,15,254,194 // paddd %xmm10,%xmm8 .byte 65,15,93,200 // minps %xmm8,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -26510,7 +26544,7 @@ FUNCTION(_sk_clamp_x_1_sse41) _sk_clamp_x_1_sse41: .byte 69,15,87,192 // xorps %xmm8,%xmm8 .byte 68,15,95,192 // maxps %xmm0,%xmm8 - .byte 68,15,93,5,112,28,0,0 // minps 0x1c70(%rip),%xmm8 # 5f70 <_sk_callback_sse41+0xf45> + .byte 68,15,93,5,114,28,0,0 // minps 0x1c72(%rip),%xmm8 # 5fa0 <_sk_callback_sse41+0xf47> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -26528,9 +26562,9 @@ HIDDEN _sk_mirror_x_1_sse41 .globl _sk_mirror_x_1_sse41 FUNCTION(_sk_mirror_x_1_sse41) _sk_mirror_x_1_sse41: - .byte 68,15,40,5,97,28,0,0 // movaps 0x1c61(%rip),%xmm8 # 5f80 <_sk_callback_sse41+0xf55> + .byte 68,15,40,5,99,28,0,0 // movaps 0x1c63(%rip),%xmm8 # 5fb0 <_sk_callback_sse41+0xf57> .byte 65,15,88,192 // addps %xmm8,%xmm0 - .byte 68,15,40,13,101,28,0,0 // movaps 0x1c65(%rip),%xmm9 # 5f90 <_sk_callback_sse41+0xf65> + .byte 68,15,40,13,103,28,0,0 // movaps 0x1c67(%rip),%xmm9 # 5fc0 <_sk_callback_sse41+0xf67> .byte 68,15,89,200 // mulps %xmm0,%xmm9 .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 .byte 69,15,88,201 // addps %xmm9,%xmm9 @@ -26547,10 +26581,10 @@ 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 15,89,5,68,28,0,0 // mulps 0x1c44(%rip),%xmm0 # 5fa0 <_sk_callback_sse41+0xf75> - .byte 15,89,13,77,28,0,0 // mulps 0x1c4d(%rip),%xmm1 # 5fb0 <_sk_callback_sse41+0xf85> + .byte 15,89,5,70,28,0,0 // mulps 0x1c46(%rip),%xmm0 # 5fd0 <_sk_callback_sse41+0xf77> + .byte 15,89,13,79,28,0,0 // mulps 0x1c4f(%rip),%xmm1 # 5fe0 <_sk_callback_sse41+0xf87> .byte 15,88,200 // addps %xmm0,%xmm1 - .byte 15,89,29,83,28,0,0 // mulps 0x1c53(%rip),%xmm3 # 5fc0 <_sk_callback_sse41+0xf95> + .byte 15,89,29,85,28,0,0 // mulps 0x1c55(%rip),%xmm3 # 5ff0 <_sk_callback_sse41+0xf97> .byte 15,88,217 // addps %xmm1,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 @@ -26829,9 +26863,9 @@ _sk_evenly_spaced_gradient_sse41: .byte 72,139,24 // mov (%rax),%rbx .byte 76,139,112,8 // mov 0x8(%rax),%r14 .byte 72,255,203 // dec %rbx - .byte 120,7 // js 47ab <_sk_evenly_spaced_gradient_sse41+0x1a> + .byte 120,7 // js 47d9 <_sk_evenly_spaced_gradient_sse41+0x1a> .byte 243,72,15,42,203 // cvtsi2ss %rbx,%xmm1 - .byte 235,21 // jmp 47c0 <_sk_evenly_spaced_gradient_sse41+0x2f> + .byte 235,21 // jmp 47ee <_sk_evenly_spaced_gradient_sse41+0x2f> .byte 73,137,217 // mov %rbx,%r9 .byte 73,209,233 // shr %r9 .byte 131,227,1 // and $0x1,%ebx @@ -26921,15 +26955,15 @@ HIDDEN _sk_gauss_a_to_rgba_sse41 .globl _sk_gauss_a_to_rgba_sse41 FUNCTION(_sk_gauss_a_to_rgba_sse41) _sk_gauss_a_to_rgba_sse41: - .byte 15,40,5,98,22,0,0 // movaps 0x1662(%rip),%xmm0 # 5fd0 <_sk_callback_sse41+0xfa5> + .byte 15,40,5,100,22,0,0 // movaps 0x1664(%rip),%xmm0 # 6000 <_sk_callback_sse41+0xfa7> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,104,22,0,0 // addps 0x1668(%rip),%xmm0 # 5fe0 <_sk_callback_sse41+0xfb5> + .byte 15,88,5,106,22,0,0 // addps 0x166a(%rip),%xmm0 # 6010 <_sk_callback_sse41+0xfb7> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,110,22,0,0 // addps 0x166e(%rip),%xmm0 # 5ff0 <_sk_callback_sse41+0xfc5> + .byte 15,88,5,112,22,0,0 // addps 0x1670(%rip),%xmm0 # 6020 <_sk_callback_sse41+0xfc7> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,116,22,0,0 // addps 0x1674(%rip),%xmm0 # 6000 <_sk_callback_sse41+0xfd5> + .byte 15,88,5,118,22,0,0 // addps 0x1676(%rip),%xmm0 # 6030 <_sk_callback_sse41+0xfd7> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,122,22,0,0 // addps 0x167a(%rip),%xmm0 # 6010 <_sk_callback_sse41+0xfe5> + .byte 15,88,5,124,22,0,0 // addps 0x167c(%rip),%xmm0 # 6040 <_sk_callback_sse41+0xfe7> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 @@ -26947,12 +26981,12 @@ _sk_gradient_sse41: .byte 76,139,8 // mov (%rax),%r9 .byte 102,15,239,201 // pxor %xmm1,%xmm1 .byte 73,131,249,2 // cmp $0x2,%r9 - .byte 114,50 // jb 49e9 <_sk_gradient_sse41+0x46> + .byte 114,50 // jb 4a17 <_sk_gradient_sse41+0x46> .byte 72,139,88,72 // mov 0x48(%rax),%rbx .byte 73,255,201 // dec %r9 .byte 72,131,195,4 // add $0x4,%rbx .byte 102,15,239,201 // pxor %xmm1,%xmm1 - .byte 15,40,21,83,22,0,0 // movaps 0x1653(%rip),%xmm2 # 6020 <_sk_callback_sse41+0xff5> + .byte 15,40,21,85,22,0,0 // movaps 0x1655(%rip),%xmm2 # 6050 <_sk_callback_sse41+0xff7> .byte 243,15,16,27 // movss (%rbx),%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 15,194,216,2 // cmpleps %xmm0,%xmm3 @@ -26960,7 +26994,7 @@ _sk_gradient_sse41: .byte 102,15,254,203 // paddd %xmm3,%xmm1 .byte 72,131,195,4 // add $0x4,%rbx .byte 73,255,201 // dec %r9 - .byte 117,228 // jne 49cd <_sk_gradient_sse41+0x2a> + .byte 117,228 // jne 49fb <_sk_gradient_sse41+0x2a> .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 .byte 69,137,202 // mov %r9d,%r10d .byte 73,193,233,32 // shr $0x20,%r9 @@ -27090,26 +27124,26 @@ _sk_xy_to_unit_angle_sse41: .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,40,236 // movaps %xmm12,%xmm13 .byte 69,15,89,237 // mulps %xmm13,%xmm13 - .byte 68,15,40,21,244,19,0,0 // movaps 0x13f4(%rip),%xmm10 # 6030 <_sk_callback_sse41+0x1005> + .byte 68,15,40,21,246,19,0,0 // movaps 0x13f6(%rip),%xmm10 # 6060 <_sk_callback_sse41+0x1007> .byte 69,15,89,213 // mulps %xmm13,%xmm10 - .byte 68,15,88,21,248,19,0,0 // addps 0x13f8(%rip),%xmm10 # 6040 <_sk_callback_sse41+0x1015> + .byte 68,15,88,21,250,19,0,0 // addps 0x13fa(%rip),%xmm10 # 6070 <_sk_callback_sse41+0x1017> .byte 69,15,89,213 // mulps %xmm13,%xmm10 - .byte 68,15,88,21,252,19,0,0 // addps 0x13fc(%rip),%xmm10 # 6050 <_sk_callback_sse41+0x1025> + .byte 68,15,88,21,254,19,0,0 // addps 0x13fe(%rip),%xmm10 # 6080 <_sk_callback_sse41+0x1027> .byte 69,15,89,213 // mulps %xmm13,%xmm10 - .byte 68,15,88,21,0,20,0,0 // addps 0x1400(%rip),%xmm10 # 6060 <_sk_callback_sse41+0x1035> + .byte 68,15,88,21,2,20,0,0 // addps 0x1402(%rip),%xmm10 # 6090 <_sk_callback_sse41+0x1037> .byte 69,15,89,212 // mulps %xmm12,%xmm10 .byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0 - .byte 68,15,40,29,255,19,0,0 // movaps 0x13ff(%rip),%xmm11 # 6070 <_sk_callback_sse41+0x1045> + .byte 68,15,40,29,1,20,0,0 // movaps 0x1401(%rip),%xmm11 # 60a0 <_sk_callback_sse41+0x1047> .byte 69,15,92,218 // subps %xmm10,%xmm11 .byte 102,69,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm10 .byte 69,15,194,200,1 // cmpltps %xmm8,%xmm9 - .byte 68,15,40,29,248,19,0,0 // movaps 0x13f8(%rip),%xmm11 # 6080 <_sk_callback_sse41+0x1055> + .byte 68,15,40,29,250,19,0,0 // movaps 0x13fa(%rip),%xmm11 # 60b0 <_sk_callback_sse41+0x1057> .byte 69,15,92,218 // subps %xmm10,%xmm11 .byte 65,15,40,193 // movaps %xmm9,%xmm0 .byte 102,69,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm10 .byte 15,40,193 // movaps %xmm1,%xmm0 .byte 65,15,194,192,1 // cmpltps %xmm8,%xmm0 - .byte 68,15,40,13,234,19,0,0 // movaps 0x13ea(%rip),%xmm9 # 6090 <_sk_callback_sse41+0x1065> + .byte 68,15,40,13,236,19,0,0 // movaps 0x13ec(%rip),%xmm9 # 60c0 <_sk_callback_sse41+0x1067> .byte 69,15,92,202 // subps %xmm10,%xmm9 .byte 102,69,15,56,20,209 // blendvps %xmm0,%xmm9,%xmm10 .byte 69,15,194,194,7 // cmpordps %xmm10,%xmm8 @@ -27135,7 +27169,7 @@ HIDDEN _sk_save_xy_sse41 FUNCTION(_sk_save_xy_sse41) _sk_save_xy_sse41: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,190,19,0,0 // movaps 0x13be(%rip),%xmm8 # 60a0 <_sk_callback_sse41+0x1075> + .byte 68,15,40,5,192,19,0,0 // movaps 0x13c0(%rip),%xmm8 # 60d0 <_sk_callback_sse41+0x1077> .byte 15,17,0 // movups %xmm0,(%rax) .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,88,200 // addps %xmm8,%xmm9 @@ -27179,8 +27213,8 @@ _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 15,88,5,64,19,0,0 // addps 0x1340(%rip),%xmm0 # 60b0 <_sk_callback_sse41+0x1085> - .byte 68,15,40,13,72,19,0,0 // movaps 0x1348(%rip),%xmm9 # 60c0 <_sk_callback_sse41+0x1095> + .byte 15,88,5,66,19,0,0 // addps 0x1342(%rip),%xmm0 # 60e0 <_sk_callback_sse41+0x1087> + .byte 68,15,40,13,74,19,0,0 // movaps 0x134a(%rip),%xmm9 # 60f0 <_sk_callback_sse41+0x1097> .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 @@ -27193,7 +27227,7 @@ _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 15,88,5,55,19,0,0 // addps 0x1337(%rip),%xmm0 # 60d0 <_sk_callback_sse41+0x10a5> + .byte 15,88,5,57,19,0,0 // addps 0x1339(%rip),%xmm0 # 6100 <_sk_callback_sse41+0x10a7> .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 @@ -27205,8 +27239,8 @@ _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 15,88,13,41,19,0,0 // addps 0x1329(%rip),%xmm1 # 60e0 <_sk_callback_sse41+0x10b5> - .byte 68,15,40,13,49,19,0,0 // movaps 0x1331(%rip),%xmm9 # 60f0 <_sk_callback_sse41+0x10c5> + .byte 15,88,13,43,19,0,0 // addps 0x132b(%rip),%xmm1 # 6110 <_sk_callback_sse41+0x10b7> + .byte 68,15,40,13,51,19,0,0 // movaps 0x1333(%rip),%xmm9 # 6120 <_sk_callback_sse41+0x10c7> .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 @@ -27219,7 +27253,7 @@ _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 15,88,13,31,19,0,0 // addps 0x131f(%rip),%xmm1 # 6100 <_sk_callback_sse41+0x10d5> + .byte 15,88,13,33,19,0,0 // addps 0x1321(%rip),%xmm1 # 6130 <_sk_callback_sse41+0x10d7> .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 @@ -27231,13 +27265,13 @@ _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 15,88,5,18,19,0,0 // addps 0x1312(%rip),%xmm0 # 6110 <_sk_callback_sse41+0x10e5> - .byte 68,15,40,13,26,19,0,0 // movaps 0x131a(%rip),%xmm9 # 6120 <_sk_callback_sse41+0x10f5> + .byte 15,88,5,20,19,0,0 // addps 0x1314(%rip),%xmm0 # 6140 <_sk_callback_sse41+0x10e7> + .byte 68,15,40,13,28,19,0,0 // movaps 0x131c(%rip),%xmm9 # 6150 <_sk_callback_sse41+0x10f7> .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 68,15,89,13,22,19,0,0 // mulps 0x1316(%rip),%xmm9 # 6130 <_sk_callback_sse41+0x1105> - .byte 68,15,88,13,30,19,0,0 // addps 0x131e(%rip),%xmm9 # 6140 <_sk_callback_sse41+0x1115> + .byte 68,15,89,13,24,19,0,0 // mulps 0x1318(%rip),%xmm9 # 6160 <_sk_callback_sse41+0x1107> + .byte 68,15,88,13,32,19,0,0 // addps 0x1320(%rip),%xmm9 # 6170 <_sk_callback_sse41+0x1117> .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 @@ -27250,16 +27284,16 @@ _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 15,88,5,13,19,0,0 // addps 0x130d(%rip),%xmm0 # 6150 <_sk_callback_sse41+0x1125> - .byte 68,15,40,13,21,19,0,0 // movaps 0x1315(%rip),%xmm9 # 6160 <_sk_callback_sse41+0x1135> + .byte 15,88,5,15,19,0,0 // addps 0x130f(%rip),%xmm0 # 6180 <_sk_callback_sse41+0x1127> + .byte 68,15,40,13,23,19,0,0 // movaps 0x1317(%rip),%xmm9 # 6190 <_sk_callback_sse41+0x1137> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,25,19,0,0 // movaps 0x1319(%rip),%xmm8 # 6170 <_sk_callback_sse41+0x1145> + .byte 68,15,40,5,27,19,0,0 // movaps 0x131b(%rip),%xmm8 # 61a0 <_sk_callback_sse41+0x1147> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,29,19,0,0 // addps 0x131d(%rip),%xmm8 # 6180 <_sk_callback_sse41+0x1155> + .byte 68,15,88,5,31,19,0,0 // addps 0x131f(%rip),%xmm8 # 61b0 <_sk_callback_sse41+0x1157> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,33,19,0,0 // addps 0x1321(%rip),%xmm8 # 6190 <_sk_callback_sse41+0x1165> + .byte 68,15,88,5,35,19,0,0 // addps 0x1323(%rip),%xmm8 # 61c0 <_sk_callback_sse41+0x1167> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,37,19,0,0 // addps 0x1325(%rip),%xmm8 # 61a0 <_sk_callback_sse41+0x1175> + .byte 68,15,88,5,39,19,0,0 // addps 0x1327(%rip),%xmm8 # 61d0 <_sk_callback_sse41+0x1177> .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 @@ -27269,17 +27303,17 @@ HIDDEN _sk_bicubic_p1x_sse41 FUNCTION(_sk_bicubic_p1x_sse41) _sk_bicubic_p1x_sse41: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,31,19,0,0 // movaps 0x131f(%rip),%xmm8 # 61b0 <_sk_callback_sse41+0x1185> + .byte 68,15,40,5,33,19,0,0 // movaps 0x1321(%rip),%xmm8 # 61e0 <_sk_callback_sse41+0x1187> .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 68,15,40,21,27,19,0,0 // movaps 0x131b(%rip),%xmm10 # 61c0 <_sk_callback_sse41+0x1195> + .byte 68,15,40,21,29,19,0,0 // movaps 0x131d(%rip),%xmm10 # 61f0 <_sk_callback_sse41+0x1197> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,31,19,0,0 // addps 0x131f(%rip),%xmm10 # 61d0 <_sk_callback_sse41+0x11a5> + .byte 68,15,88,21,33,19,0,0 // addps 0x1321(%rip),%xmm10 # 6200 <_sk_callback_sse41+0x11a7> .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 68,15,88,21,27,19,0,0 // addps 0x131b(%rip),%xmm10 # 61e0 <_sk_callback_sse41+0x11b5> + .byte 68,15,88,21,29,19,0,0 // addps 0x131d(%rip),%xmm10 # 6210 <_sk_callback_sse41+0x11b7> .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 @@ -27291,11 +27325,11 @@ _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 15,88,5,14,19,0,0 // addps 0x130e(%rip),%xmm0 # 61f0 <_sk_callback_sse41+0x11c5> + .byte 15,88,5,16,19,0,0 // addps 0x1310(%rip),%xmm0 # 6220 <_sk_callback_sse41+0x11c7> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,14,19,0,0 // mulps 0x130e(%rip),%xmm8 # 6200 <_sk_callback_sse41+0x11d5> - .byte 68,15,88,5,22,19,0,0 // addps 0x1316(%rip),%xmm8 # 6210 <_sk_callback_sse41+0x11e5> + .byte 68,15,89,5,16,19,0,0 // mulps 0x1310(%rip),%xmm8 # 6230 <_sk_callback_sse41+0x11d7> + .byte 68,15,88,5,24,19,0,0 // addps 0x1318(%rip),%xmm8 # 6240 <_sk_callback_sse41+0x11e7> .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 @@ -27308,13 +27342,13 @@ _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 15,88,13,4,19,0,0 // addps 0x1304(%rip),%xmm1 # 6220 <_sk_callback_sse41+0x11f5> - .byte 68,15,40,13,12,19,0,0 // movaps 0x130c(%rip),%xmm9 # 6230 <_sk_callback_sse41+0x1205> + .byte 15,88,13,6,19,0,0 // addps 0x1306(%rip),%xmm1 # 6250 <_sk_callback_sse41+0x11f7> + .byte 68,15,40,13,14,19,0,0 // movaps 0x130e(%rip),%xmm9 # 6260 <_sk_callback_sse41+0x1207> .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 68,15,89,13,8,19,0,0 // mulps 0x1308(%rip),%xmm9 # 6240 <_sk_callback_sse41+0x1215> - .byte 68,15,88,13,16,19,0,0 // addps 0x1310(%rip),%xmm9 # 6250 <_sk_callback_sse41+0x1225> + .byte 68,15,89,13,10,19,0,0 // mulps 0x130a(%rip),%xmm9 # 6270 <_sk_callback_sse41+0x1217> + .byte 68,15,88,13,18,19,0,0 // addps 0x1312(%rip),%xmm9 # 6280 <_sk_callback_sse41+0x1227> .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 @@ -27327,16 +27361,16 @@ _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 15,88,13,254,18,0,0 // addps 0x12fe(%rip),%xmm1 # 6260 <_sk_callback_sse41+0x1235> - .byte 68,15,40,13,6,19,0,0 // movaps 0x1306(%rip),%xmm9 # 6270 <_sk_callback_sse41+0x1245> + .byte 15,88,13,0,19,0,0 // addps 0x1300(%rip),%xmm1 # 6290 <_sk_callback_sse41+0x1237> + .byte 68,15,40,13,8,19,0,0 // movaps 0x1308(%rip),%xmm9 # 62a0 <_sk_callback_sse41+0x1247> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,10,19,0,0 // movaps 0x130a(%rip),%xmm8 # 6280 <_sk_callback_sse41+0x1255> + .byte 68,15,40,5,12,19,0,0 // movaps 0x130c(%rip),%xmm8 # 62b0 <_sk_callback_sse41+0x1257> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,14,19,0,0 // addps 0x130e(%rip),%xmm8 # 6290 <_sk_callback_sse41+0x1265> + .byte 68,15,88,5,16,19,0,0 // addps 0x1310(%rip),%xmm8 # 62c0 <_sk_callback_sse41+0x1267> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,18,19,0,0 // addps 0x1312(%rip),%xmm8 # 62a0 <_sk_callback_sse41+0x1275> + .byte 68,15,88,5,20,19,0,0 // addps 0x1314(%rip),%xmm8 # 62d0 <_sk_callback_sse41+0x1277> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,22,19,0,0 // addps 0x1316(%rip),%xmm8 # 62b0 <_sk_callback_sse41+0x1285> + .byte 68,15,88,5,24,19,0,0 // addps 0x1318(%rip),%xmm8 # 62e0 <_sk_callback_sse41+0x1287> .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 @@ -27346,17 +27380,17 @@ HIDDEN _sk_bicubic_p1y_sse41 FUNCTION(_sk_bicubic_p1y_sse41) _sk_bicubic_p1y_sse41: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,16,19,0,0 // movaps 0x1310(%rip),%xmm8 # 62c0 <_sk_callback_sse41+0x1295> + .byte 68,15,40,5,18,19,0,0 // movaps 0x1312(%rip),%xmm8 # 62f0 <_sk_callback_sse41+0x1297> .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 68,15,40,21,11,19,0,0 // movaps 0x130b(%rip),%xmm10 # 62d0 <_sk_callback_sse41+0x12a5> + .byte 68,15,40,21,13,19,0,0 // movaps 0x130d(%rip),%xmm10 # 6300 <_sk_callback_sse41+0x12a7> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,15,19,0,0 // addps 0x130f(%rip),%xmm10 # 62e0 <_sk_callback_sse41+0x12b5> + .byte 68,15,88,21,17,19,0,0 // addps 0x1311(%rip),%xmm10 # 6310 <_sk_callback_sse41+0x12b7> .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 68,15,88,21,11,19,0,0 // addps 0x130b(%rip),%xmm10 # 62f0 <_sk_callback_sse41+0x12c5> + .byte 68,15,88,21,13,19,0,0 // addps 0x130d(%rip),%xmm10 # 6320 <_sk_callback_sse41+0x12c7> .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 @@ -27368,11 +27402,11 @@ _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 15,88,13,253,18,0,0 // addps 0x12fd(%rip),%xmm1 # 6300 <_sk_callback_sse41+0x12d5> + .byte 15,88,13,255,18,0,0 // addps 0x12ff(%rip),%xmm1 # 6330 <_sk_callback_sse41+0x12d7> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,253,18,0,0 // mulps 0x12fd(%rip),%xmm8 # 6310 <_sk_callback_sse41+0x12e5> - .byte 68,15,88,5,5,19,0,0 // addps 0x1305(%rip),%xmm8 # 6320 <_sk_callback_sse41+0x12f5> + .byte 68,15,89,5,255,18,0,0 // mulps 0x12ff(%rip),%xmm8 # 6340 <_sk_callback_sse41+0x12e7> + .byte 68,15,88,5,7,19,0,0 // addps 0x1307(%rip),%xmm8 # 6350 <_sk_callback_sse41+0x12f7> .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 @@ -27602,11 +27636,11 @@ BALIGN16 .byte 128,191,0,0,128,191,0 // cmpb $0x0,-0x40800000(%rdi) .byte 0,224 // add %ah,%al .byte 64,0,0 // add %al,(%rax) - .byte 224,64 // loopne 52f8 <.literal16+0x1d8> + .byte 224,64 // loopne 5328 <.literal16+0x1d8> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 52fc <.literal16+0x1dc> + .byte 224,64 // loopne 532c <.literal16+0x1dc> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 5300 <.literal16+0x1e0> + .byte 224,64 // loopne 5330 <.literal16+0x1e0> .byte 154 // (bad) .byte 153 // cltd .byte 153 // cltd @@ -27626,13 +27660,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5321 <.literal16+0x201> + .byte 71,225,61 // rex.RXB loope 5351 <.literal16+0x201> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5325 <.literal16+0x205> + .byte 71,225,61 // rex.RXB loope 5355 <.literal16+0x205> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5329 <.literal16+0x209> + .byte 71,225,61 // rex.RXB loope 5359 <.literal16+0x209> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 532d <.literal16+0x20d> + .byte 71,225,61 // rex.RXB loope 535d <.literal16+0x20d> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -27657,13 +27691,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5361 <.literal16+0x241> + .byte 71,225,61 // rex.RXB loope 5391 <.literal16+0x241> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5365 <.literal16+0x245> + .byte 71,225,61 // rex.RXB loope 5395 <.literal16+0x245> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5369 <.literal16+0x249> + .byte 71,225,61 // rex.RXB loope 5399 <.literal16+0x249> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 536d <.literal16+0x24d> + .byte 71,225,61 // rex.RXB loope 539d <.literal16+0x24d> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -27688,13 +27722,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53a1 <.literal16+0x281> + .byte 71,225,61 // rex.RXB loope 53d1 <.literal16+0x281> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53a5 <.literal16+0x285> + .byte 71,225,61 // rex.RXB loope 53d5 <.literal16+0x285> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53a9 <.literal16+0x289> + .byte 71,225,61 // rex.RXB loope 53d9 <.literal16+0x289> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53ad <.literal16+0x28d> + .byte 71,225,61 // rex.RXB loope 53dd <.literal16+0x28d> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -27719,13 +27753,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53e1 <.literal16+0x2c1> + .byte 71,225,61 // rex.RXB loope 5411 <.literal16+0x2c1> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53e5 <.literal16+0x2c5> + .byte 71,225,61 // rex.RXB loope 5415 <.literal16+0x2c5> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53e9 <.literal16+0x2c9> + .byte 71,225,61 // rex.RXB loope 5419 <.literal16+0x2c9> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 53ed <.literal16+0x2cd> + .byte 71,225,61 // rex.RXB loope 541d <.literal16+0x2cd> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -27742,10 +27776,10 @@ BALIGN16 .byte 0,1 // add %al,(%rcx) .byte 255 // (bad) .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0053d8 <_sk_callback_sse41+0xa0003ad> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005408 <_sk_callback_sse41+0xa0003af> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 30053e0 <_sk_callback_sse41+0x30003b5> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3005410 <_sk_callback_sse41+0x30003b7> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -27764,11 +27798,11 @@ BALIGN16 .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax) .byte 67,0,0 // rex.XB add %al,(%r8) - .byte 127,67 // jg 544b <.literal16+0x32b> + .byte 127,67 // jg 547b <.literal16+0x32b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 544f <.literal16+0x32f> + .byte 127,67 // jg 547f <.literal16+0x32f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5453 <.literal16+0x333> + .byte 127,67 // jg 5483 <.literal16+0x333> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -28003,13 +28037,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 5629 <.literal16+0x509> + .byte 224,7 // loopne 5659 <.literal16+0x509> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 562d <.literal16+0x50d> + .byte 224,7 // loopne 565d <.literal16+0x50d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5631 <.literal16+0x511> + .byte 224,7 // loopne 5661 <.literal16+0x511> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5635 <.literal16+0x515> + .byte 224,7 // loopne 5665 <.literal16+0x515> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -28043,10 +28077,10 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 1,255 // add %edi,%edi .byte 255 // (bad) - .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a005678 <_sk_callback_sse41+0xa00064d> + .byte 255,5,255,255,255,9 // incl 0x9ffffff(%rip) # a0056a8 <_sk_callback_sse41+0xa00064f> .byte 255 // (bad) .byte 255 // (bad) - .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3005680 <_sk_callback_sse41+0x3000655> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 30056b0 <_sk_callback_sse41+0x3000657> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -28101,11 +28135,11 @@ 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 574b <.literal16+0x62b> + .byte 127,67 // jg 577b <.literal16+0x62b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 574f <.literal16+0x62f> + .byte 127,67 // jg 577f <.literal16+0x62f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5753 <.literal16+0x633> + .byte 127,67 // jg 5783 <.literal16+0x633> .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) @@ -28120,16 +28154,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5744 <.literal16+0x624> + .byte 127,0 // jg 5774 <.literal16+0x624> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5748 <.literal16+0x628> + .byte 127,0 // jg 5778 <.literal16+0x628> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 574c <.literal16+0x62c> + .byte 127,0 // jg 577c <.literal16+0x62c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5750 <.literal16+0x630> + .byte 127,0 // jg 5780 <.literal16+0x630> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -28138,7 +28172,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 57d5 <.literal16+0x6b5> + .byte 119,115 // ja 5805 <.literal16+0x6b5> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -28149,7 +28183,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 5739 <.literal16+0x619> + .byte 117,191 // jne 5769 <.literal16+0x619> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -28161,7 +28195,7 @@ 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 ffffffffe9a3977a <_sk_callback_sse41+0xffffffffe9a3474f> + .byte 233,220,63,163,233 // jmpq ffffffffe9a397aa <_sk_callback_sse41+0xffffffffe9a34751> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -28216,16 +28250,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5814 <.literal16+0x6f4> + .byte 127,0 // jg 5844 <.literal16+0x6f4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5818 <.literal16+0x6f8> + .byte 127,0 // jg 5848 <.literal16+0x6f8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 581c <.literal16+0x6fc> + .byte 127,0 // jg 584c <.literal16+0x6fc> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5820 <.literal16+0x700> + .byte 127,0 // jg 5850 <.literal16+0x700> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -28234,7 +28268,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 58a5 <.literal16+0x785> + .byte 119,115 // ja 58d5 <.literal16+0x785> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -28245,7 +28279,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 5809 <.literal16+0x6e9> + .byte 117,191 // jne 5839 <.literal16+0x6e9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -28257,7 +28291,7 @@ 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 ffffffffe9a3984a <_sk_callback_sse41+0xffffffffe9a3481f> + .byte 233,220,63,163,233 // jmpq ffffffffe9a3987a <_sk_callback_sse41+0xffffffffe9a34821> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -28312,16 +28346,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 58e4 <.literal16+0x7c4> + .byte 127,0 // jg 5914 <.literal16+0x7c4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 58e8 <.literal16+0x7c8> + .byte 127,0 // jg 5918 <.literal16+0x7c8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 58ec <.literal16+0x7cc> + .byte 127,0 // jg 591c <.literal16+0x7cc> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 58f0 <.literal16+0x7d0> + .byte 127,0 // jg 5920 <.literal16+0x7d0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -28330,7 +28364,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5975 <.literal16+0x855> + .byte 119,115 // ja 59a5 <.literal16+0x855> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -28341,7 +28375,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 58d9 <.literal16+0x7b9> + .byte 117,191 // jne 5909 <.literal16+0x7b9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -28353,7 +28387,7 @@ 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 ffffffffe9a3991a <_sk_callback_sse41+0xffffffffe9a348ef> + .byte 233,220,63,163,233 // jmpq ffffffffe9a3994a <_sk_callback_sse41+0xffffffffe9a348f1> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -28408,16 +28442,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 59b4 <.literal16+0x894> + .byte 127,0 // jg 59e4 <.literal16+0x894> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 59b8 <.literal16+0x898> + .byte 127,0 // jg 59e8 <.literal16+0x898> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 59bc <.literal16+0x89c> + .byte 127,0 // jg 59ec <.literal16+0x89c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 59c0 <.literal16+0x8a0> + .byte 127,0 // jg 59f0 <.literal16+0x8a0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -28426,7 +28460,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5a45 <.literal16+0x925> + .byte 119,115 // ja 5a75 <.literal16+0x925> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -28437,7 +28471,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 59a9 <.literal16+0x889> + .byte 117,191 // jne 59d9 <.literal16+0x889> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -28449,7 +28483,7 @@ 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 ffffffffe9a399ea <_sk_callback_sse41+0xffffffffe9a349bf> + .byte 233,220,63,163,233 // jmpq ffffffffe9a39a1a <_sk_callback_sse41+0xffffffffe9a349c1> .byte 220,63 // fdivrl (%rdi) .byte 81 // push %rcx .byte 140,242 // mov %?,%edx @@ -28500,13 +28534,13 @@ 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 5ac7 <.literal16+0x9a7> + .byte 127,67 // jg 5af7 <.literal16+0x9a7> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5acb <.literal16+0x9ab> + .byte 127,67 // jg 5afb <.literal16+0x9ab> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5acf <.literal16+0x9af> + .byte 127,67 // jg 5aff <.literal16+0x9af> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5ad3 <.literal16+0x9b3> + .byte 127,67 // jg 5b03 <.literal16+0x9b3> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -28553,16 +28587,16 @@ BALIGN16 .byte 128,3,62 // addb $0x3e,(%rbx) .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 5b53 <.literal16+0xa33> + .byte 118,63 // jbe 5b83 <.literal16+0xa33> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 5b57 <.literal16+0xa37> + .byte 118,63 // jbe 5b87 <.literal16+0xa37> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 5b5b <.literal16+0xa3b> + .byte 118,63 // jbe 5b8b <.literal16+0xa3b> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 5b5f <.literal16+0xa3f> + .byte 118,63 // jbe 5b8f <.literal16+0xa3f> .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) @@ -28582,11 +28616,11 @@ 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 5bab <.literal16+0xa8b> + .byte 127,67 // jg 5bdb <.literal16+0xa8b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5baf <.literal16+0xa8f> + .byte 127,67 // jg 5bdf <.literal16+0xa8f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5bb3 <.literal16+0xa93> + .byte 127,67 // jg 5be3 <.literal16+0xa93> .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -28623,7 +28657,7 @@ 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) # 3005bf0 <_sk_callback_sse41+0x3000bc5> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3005c20 <_sk_callback_sse41+0x3000bc7> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -28652,13 +28686,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 5c29 <.literal16+0xb09> + .byte 224,7 // loopne 5c59 <.literal16+0xb09> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5c2d <.literal16+0xb0d> + .byte 224,7 // loopne 5c5d <.literal16+0xb0d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5c31 <.literal16+0xb11> + .byte 224,7 // loopne 5c61 <.literal16+0xb11> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5c35 <.literal16+0xb15> + .byte 224,7 // loopne 5c65 <.literal16+0xb15> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -28704,13 +28738,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 5c99 <.literal16+0xb79> + .byte 224,7 // loopne 5cc9 <.literal16+0xb79> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5c9d <.literal16+0xb7d> + .byte 224,7 // loopne 5ccd <.literal16+0xb7d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5ca1 <.literal16+0xb81> + .byte 224,7 // loopne 5cd1 <.literal16+0xb81> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5ca5 <.literal16+0xb85> + .byte 224,7 // loopne 5cd5 <.literal16+0xb85> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -28748,13 +28782,13 @@ BALIGN16 .byte 65,0,0 // add %al,(%r8) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 5d36 <.literal16+0xc16> + .byte 124,66 // jl 5d66 <.literal16+0xc16> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 5d3a <.literal16+0xc1a> + .byte 124,66 // jl 5d6a <.literal16+0xc1a> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 5d3e <.literal16+0xc1e> + .byte 124,66 // jl 5d6e <.literal16+0xc1e> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 5d42 <.literal16+0xc22> + .byte 124,66 // jl 5d72 <.literal16+0xc22> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 0,240 // add %dh,%al @@ -28844,13 +28878,13 @@ 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 5e45 <.literal16+0xd25> + .byte 112,65 // jo 5e75 <.literal16+0xd25> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 5e49 <.literal16+0xd29> + .byte 112,65 // jo 5e79 <.literal16+0xd29> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 5e4d <.literal16+0xd2d> + .byte 112,65 // jo 5e7d <.literal16+0xd2d> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 5e51 <.literal16+0xd31> + .byte 112,65 // jo 5e81 <.literal16+0xd31> .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -28865,7 +28899,7 @@ 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) # 3005e40 <_sk_callback_sse41+0x3000e15> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3005e70 <_sk_callback_sse41+0x3000e17> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -28892,7 +28926,7 @@ 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) # 3005e80 <_sk_callback_sse41+0x3000e55> + .byte 255,13,255,255,255,2 // decl 0x2ffffff(%rip) # 3005eb0 <_sk_callback_sse41+0x3000e57> .byte 255 // (bad) .byte 255 // (bad) .byte 255,6 // incl (%rsi) @@ -28907,11 +28941,11 @@ BALIGN16 .byte 255,0 // incl (%rax) .byte 0,127,67 // add %bh,0x43(%rdi) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5edb <.literal16+0xdbb> + .byte 127,67 // jg 5f0b <.literal16+0xdbb> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5edf <.literal16+0xdbf> + .byte 127,67 // jg 5f0f <.literal16+0xdbf> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5ee3 <.literal16+0xdc3> + .byte 127,67 // jg 5f13 <.literal16+0xdc3> .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) @@ -28987,13 +29021,13 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 255 // (bad) - .byte 127,71 // jg 5fab <.literal16+0xe8b> + .byte 127,71 // jg 5fdb <.literal16+0xe8b> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 5faf <.literal16+0xe8f> + .byte 127,71 // jg 5fdf <.literal16+0xe8f> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 5fb3 <.literal16+0xe93> + .byte 127,71 // jg 5fe3 <.literal16+0xe93> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 5fb7 <.literal16+0xe97> + .byte 127,71 // jg 5fe7 <.literal16+0xe97> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -29039,10 +29073,10 @@ BALIGN16 .byte 61,152,221,147,61 // cmp $0x3d93dd98,%eax .byte 152 // cwtl .byte 221,147,61,45,16,17 // fstl 0x11102d3d(%rbx) - .byte 192,45,16,17,192,45,16 // shrb $0x10,0x2dc01110(%rip) # 2dc070ea <_sk_callback_sse41+0x2dc020bf> + .byte 192,45,16,17,192,45,16 // shrb $0x10,0x2dc01110(%rip) # 2dc0711a <_sk_callback_sse41+0x2dc020c1> .byte 17,192 // adc %eax,%eax .byte 45,16,17,192,18 // sub $0x12c01110,%eax - .byte 120,57 // js 601c <.literal16+0xefc> + .byte 120,57 // js 604c <.literal16+0xefc> .byte 64,18,120,57 // adc 0x39(%rax),%dil .byte 64,18,120,57 // adc 0x39(%rax),%dil .byte 64,18,120,57 // adc 0x39(%rax),%dil @@ -29164,11 +29198,11 @@ 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 6152 <.literal16+0x1032> + .byte 62,114,28 // jb,pt 6182 <.literal16+0x1032> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6156 <.literal16+0x1036> + .byte 62,114,28 // jb,pt 6186 <.literal16+0x1036> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 615a <.literal16+0x103a> + .byte 62,114,28 // jb,pt 618a <.literal16+0x103a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -29212,7 +29246,7 @@ 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 # 3d63efe5 <_sk_callback_sse41+0x3d639fba> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f015 <_sk_callback_sse41+0x3d639fbc> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -29238,7 +29272,7 @@ 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 # 3d63f025 <_sk_callback_sse41+0x3d639ffa> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f055 <_sk_callback_sse41+0x3d639ffc> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -29247,13 +29281,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 621e <.literal16+0x10fe> + .byte 114,28 // jb 624e <.literal16+0x10fe> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6222 <.literal16+0x1102> + .byte 62,114,28 // jb,pt 6252 <.literal16+0x1102> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6226 <.literal16+0x1106> + .byte 62,114,28 // jb,pt 6256 <.literal16+0x1106> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 622a <.literal16+0x110a> + .byte 62,114,28 // jb,pt 625a <.literal16+0x110a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -29274,11 +29308,11 @@ 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 6262 <.literal16+0x1142> + .byte 62,114,28 // jb,pt 6292 <.literal16+0x1142> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6266 <.literal16+0x1146> + .byte 62,114,28 // jb,pt 6296 <.literal16+0x1146> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 626a <.literal16+0x114a> + .byte 62,114,28 // jb,pt 629a <.literal16+0x114a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -29322,7 +29356,7 @@ 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 # 3d63f0f5 <_sk_callback_sse41+0x3d63a0ca> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f125 <_sk_callback_sse41+0x3d63a0cc> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -29348,7 +29382,7 @@ 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 # 3d63f135 <_sk_callback_sse41+0x3d63a10a> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f165 <_sk_callback_sse41+0x3d63a10c> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -29357,13 +29391,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 632e <.literal16+0x120e> + .byte 114,28 // jb 635e <.literal16+0x120e> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6332 <_sk_callback_sse41+0x1307> + .byte 62,114,28 // jb,pt 6362 <_sk_callback_sse41+0x1309> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6336 <_sk_callback_sse41+0x130b> + .byte 62,114,28 // jb,pt 6366 <_sk_callback_sse41+0x130d> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 633a <_sk_callback_sse41+0x130f> + .byte 62,114,28 // jb,pt 636a <_sk_callback_sse41+0x1311> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -29373,6 +29407,12 @@ BALIGN16 .byte 170 // stos %al,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) .byte 190,171,170,170,190 // mov $0xbeaaaaab,%esi + +BALIGN4 + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) BALIGN32 HIDDEN _sk_start_pipeline_sse2 @@ -29465,7 +29505,7 @@ _sk_seed_shader_sse2: .byte 102,15,110,194 // movd %edx,%xmm0 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 - .byte 15,40,21,172,85,0,0 // movaps 0x55ac(%rip),%xmm2 # 5680 <_sk_callback_sse2+0xfe> + .byte 15,40,21,220,85,0,0 // movaps 0x55dc(%rip),%xmm2 # 56b0 <_sk_callback_sse2+0x100> .byte 15,88,202 // addps %xmm2,%xmm1 .byte 15,16,7 // movups (%rdi),%xmm0 .byte 15,88,193 // addps %xmm1,%xmm0 @@ -29474,7 +29514,7 @@ _sk_seed_shader_sse2: .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 15,88,202 // addps %xmm2,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,21,155,85,0,0 // movaps 0x559b(%rip),%xmm2 # 5690 <_sk_callback_sse2+0x10e> + .byte 15,40,21,203,85,0,0 // movaps 0x55cb(%rip),%xmm2 # 56c0 <_sk_callback_sse2+0x110> .byte 15,87,219 // xorps %xmm3,%xmm3 .byte 15,87,228 // xorps %xmm4,%xmm4 .byte 15,87,237 // xorps %xmm5,%xmm5 @@ -29494,14 +29534,14 @@ _sk_dither_sse2: .byte 102,68,15,110,193 // movd %ecx,%xmm8 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 .byte 102,69,15,239,193 // pxor %xmm9,%xmm8 - .byte 102,68,15,111,21,105,85,0,0 // movdqa 0x5569(%rip),%xmm10 # 56a0 <_sk_callback_sse2+0x11e> + .byte 102,68,15,111,21,153,85,0,0 // movdqa 0x5599(%rip),%xmm10 # 56d0 <_sk_callback_sse2+0x120> .byte 102,69,15,111,216 // movdqa %xmm8,%xmm11 .byte 102,69,15,219,218 // pand %xmm10,%xmm11 .byte 102,65,15,114,243,5 // pslld $0x5,%xmm11 .byte 102,69,15,219,209 // pand %xmm9,%xmm10 .byte 102,65,15,114,242,4 // pslld $0x4,%xmm10 - .byte 102,68,15,111,37,85,85,0,0 // movdqa 0x5555(%rip),%xmm12 # 56b0 <_sk_callback_sse2+0x12e> - .byte 102,68,15,111,45,92,85,0,0 // movdqa 0x555c(%rip),%xmm13 # 56c0 <_sk_callback_sse2+0x13e> + .byte 102,68,15,111,37,133,85,0,0 // movdqa 0x5585(%rip),%xmm12 # 56e0 <_sk_callback_sse2+0x130> + .byte 102,68,15,111,45,140,85,0,0 // movdqa 0x558c(%rip),%xmm13 # 56f0 <_sk_callback_sse2+0x140> .byte 102,69,15,111,240 // movdqa %xmm8,%xmm14 .byte 102,69,15,219,245 // pand %xmm13,%xmm14 .byte 102,65,15,114,246,2 // pslld $0x2,%xmm14 @@ -29517,8 +29557,8 @@ _sk_dither_sse2: .byte 102,69,15,235,245 // por %xmm13,%xmm14 .byte 102,69,15,235,240 // por %xmm8,%xmm14 .byte 69,15,91,198 // cvtdq2ps %xmm14,%xmm8 - .byte 68,15,89,5,23,85,0,0 // mulps 0x5517(%rip),%xmm8 # 56d0 <_sk_callback_sse2+0x14e> - .byte 68,15,88,5,31,85,0,0 // addps 0x551f(%rip),%xmm8 # 56e0 <_sk_callback_sse2+0x15e> + .byte 68,15,89,5,71,85,0,0 // mulps 0x5547(%rip),%xmm8 # 5700 <_sk_callback_sse2+0x150> + .byte 68,15,88,5,79,85,0,0 // addps 0x554f(%rip),%xmm8 # 5710 <_sk_callback_sse2+0x160> .byte 243,68,15,16,16 // movss (%rax),%xmm10 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 @@ -29595,7 +29635,7 @@ HIDDEN _sk_srcatop_sse2 FUNCTION(_sk_srcatop_sse2) _sk_srcatop_sse2: .byte 15,89,199 // mulps %xmm7,%xmm0 - .byte 68,15,40,5,121,84,0,0 // movaps 0x5479(%rip),%xmm8 # 56f0 <_sk_callback_sse2+0x16e> + .byte 68,15,40,5,169,84,0,0 // movaps 0x54a9(%rip),%xmm8 # 5720 <_sk_callback_sse2+0x170> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -29620,7 +29660,7 @@ FUNCTION(_sk_dstatop_sse2) _sk_dstatop_sse2: .byte 68,15,40,195 // movaps %xmm3,%xmm8 .byte 68,15,89,196 // mulps %xmm4,%xmm8 - .byte 68,15,40,13,60,84,0,0 // movaps 0x543c(%rip),%xmm9 # 5700 <_sk_callback_sse2+0x17e> + .byte 68,15,40,13,108,84,0,0 // movaps 0x546c(%rip),%xmm9 # 5730 <_sk_callback_sse2+0x180> .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 65,15,88,192 // addps %xmm8,%xmm0 @@ -29667,7 +29707,7 @@ HIDDEN _sk_srcout_sse2 .globl _sk_srcout_sse2 FUNCTION(_sk_srcout_sse2) _sk_srcout_sse2: - .byte 68,15,40,5,224,83,0,0 // movaps 0x53e0(%rip),%xmm8 # 5710 <_sk_callback_sse2+0x18e> + .byte 68,15,40,5,16,84,0,0 // movaps 0x5410(%rip),%xmm8 # 5740 <_sk_callback_sse2+0x190> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 @@ -29680,7 +29720,7 @@ HIDDEN _sk_dstout_sse2 .globl _sk_dstout_sse2 FUNCTION(_sk_dstout_sse2) _sk_dstout_sse2: - .byte 68,15,40,5,208,83,0,0 // movaps 0x53d0(%rip),%xmm8 # 5720 <_sk_callback_sse2+0x19e> + .byte 68,15,40,5,0,84,0,0 // movaps 0x5400(%rip),%xmm8 # 5750 <_sk_callback_sse2+0x1a0> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 15,89,196 // mulps %xmm4,%xmm0 @@ -29697,7 +29737,7 @@ HIDDEN _sk_srcover_sse2 .globl _sk_srcover_sse2 FUNCTION(_sk_srcover_sse2) _sk_srcover_sse2: - .byte 68,15,40,5,179,83,0,0 // movaps 0x53b3(%rip),%xmm8 # 5730 <_sk_callback_sse2+0x1ae> + .byte 68,15,40,5,227,83,0,0 // movaps 0x53e3(%rip),%xmm8 # 5760 <_sk_callback_sse2+0x1b0> .byte 68,15,92,195 // subps %xmm3,%xmm8 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -29717,7 +29757,7 @@ HIDDEN _sk_dstover_sse2 .globl _sk_dstover_sse2 FUNCTION(_sk_dstover_sse2) _sk_dstover_sse2: - .byte 68,15,40,5,135,83,0,0 // movaps 0x5387(%rip),%xmm8 # 5740 <_sk_callback_sse2+0x1be> + .byte 68,15,40,5,183,83,0,0 // movaps 0x53b7(%rip),%xmm8 # 5770 <_sk_callback_sse2+0x1c0> .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -29745,7 +29785,7 @@ HIDDEN _sk_multiply_sse2 .globl _sk_multiply_sse2 FUNCTION(_sk_multiply_sse2) _sk_multiply_sse2: - .byte 68,15,40,5,91,83,0,0 // movaps 0x535b(%rip),%xmm8 # 5750 <_sk_callback_sse2+0x1ce> + .byte 68,15,40,5,139,83,0,0 // movaps 0x538b(%rip),%xmm8 # 5780 <_sk_callback_sse2+0x1d0> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 69,15,40,209 // movaps %xmm9,%xmm10 @@ -29820,7 +29860,7 @@ HIDDEN _sk_xor__sse2 FUNCTION(_sk_xor__sse2) _sk_xor__sse2: .byte 68,15,40,195 // movaps %xmm3,%xmm8 - .byte 15,40,29,144,82,0,0 // movaps 0x5290(%rip),%xmm3 # 5760 <_sk_callback_sse2+0x1de> + .byte 15,40,29,192,82,0,0 // movaps 0x52c0(%rip),%xmm3 # 5790 <_sk_callback_sse2+0x1e0> .byte 68,15,40,203 // movaps %xmm3,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 65,15,89,193 // mulps %xmm9,%xmm0 @@ -29868,7 +29908,7 @@ _sk_darken_sse2: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,95,209 // maxps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,251,81,0,0 // movaps 0x51fb(%rip),%xmm2 # 5770 <_sk_callback_sse2+0x1ee> + .byte 15,40,21,43,82,0,0 // movaps 0x522b(%rip),%xmm2 # 57a0 <_sk_callback_sse2+0x1f0> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -29902,7 +29942,7 @@ _sk_lighten_sse2: .byte 68,15,89,206 // mulps %xmm6,%xmm9 .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,160,81,0,0 // movaps 0x51a0(%rip),%xmm2 # 5780 <_sk_callback_sse2+0x1fe> + .byte 15,40,21,208,81,0,0 // movaps 0x51d0(%rip),%xmm2 # 57b0 <_sk_callback_sse2+0x200> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -29939,7 +29979,7 @@ _sk_difference_sse2: .byte 65,15,93,209 // minps %xmm9,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,58,81,0,0 // movaps 0x513a(%rip),%xmm2 # 5790 <_sk_callback_sse2+0x20e> + .byte 15,40,21,106,81,0,0 // movaps 0x516a(%rip),%xmm2 # 57c0 <_sk_callback_sse2+0x210> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -29967,7 +30007,7 @@ _sk_exclusion_sse2: .byte 15,89,214 // mulps %xmm6,%xmm2 .byte 15,88,210 // addps %xmm2,%xmm2 .byte 68,15,92,194 // subps %xmm2,%xmm8 - .byte 15,40,21,250,80,0,0 // movaps 0x50fa(%rip),%xmm2 # 57a0 <_sk_callback_sse2+0x21e> + .byte 15,40,21,42,81,0,0 // movaps 0x512a(%rip),%xmm2 # 57d0 <_sk_callback_sse2+0x220> .byte 15,92,211 // subps %xmm3,%xmm2 .byte 15,89,215 // mulps %xmm7,%xmm2 .byte 15,88,218 // addps %xmm2,%xmm3 @@ -29980,7 +30020,7 @@ HIDDEN _sk_colorburn_sse2 FUNCTION(_sk_colorburn_sse2) _sk_colorburn_sse2: .byte 68,15,40,192 // movaps %xmm0,%xmm8 - .byte 68,15,40,21,237,80,0,0 // movaps 0x50ed(%rip),%xmm10 # 57b0 <_sk_callback_sse2+0x22e> + .byte 68,15,40,21,29,81,0,0 // movaps 0x511d(%rip),%xmm10 # 57e0 <_sk_callback_sse2+0x230> .byte 69,15,40,202 // movaps %xmm10,%xmm9 .byte 68,15,92,207 // subps %xmm7,%xmm9 .byte 69,15,40,217 // movaps %xmm9,%xmm11 @@ -30074,7 +30114,7 @@ HIDDEN _sk_colordodge_sse2 FUNCTION(_sk_colordodge_sse2) _sk_colordodge_sse2: .byte 68,15,40,200 // movaps %xmm0,%xmm9 - .byte 68,15,40,21,163,79,0,0 // movaps 0x4fa3(%rip),%xmm10 # 57c0 <_sk_callback_sse2+0x23e> + .byte 68,15,40,21,211,79,0,0 // movaps 0x4fd3(%rip),%xmm10 # 57f0 <_sk_callback_sse2+0x240> .byte 69,15,40,218 // movaps %xmm10,%xmm11 .byte 68,15,92,223 // subps %xmm7,%xmm11 .byte 69,15,40,227 // movaps %xmm11,%xmm12 @@ -30168,7 +30208,7 @@ _sk_hardlight_sse2: .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp) .byte 15,40,245 // movaps %xmm5,%xmm6 .byte 15,40,236 // movaps %xmm4,%xmm5 - .byte 68,15,40,29,88,78,0,0 // movaps 0x4e58(%rip),%xmm11 # 57d0 <_sk_callback_sse2+0x24e> + .byte 68,15,40,29,136,78,0,0 // movaps 0x4e88(%rip),%xmm11 # 5800 <_sk_callback_sse2+0x250> .byte 69,15,40,211 // movaps %xmm11,%xmm10 .byte 68,15,92,215 // subps %xmm7,%xmm10 .byte 69,15,40,194 // movaps %xmm10,%xmm8 @@ -30256,7 +30296,7 @@ FUNCTION(_sk_overlay_sse2) _sk_overlay_sse2: .byte 68,15,40,193 // movaps %xmm1,%xmm8 .byte 68,15,40,232 // movaps %xmm0,%xmm13 - .byte 68,15,40,13,38,77,0,0 // movaps 0x4d26(%rip),%xmm9 # 57e0 <_sk_callback_sse2+0x25e> + .byte 68,15,40,13,86,77,0,0 // movaps 0x4d56(%rip),%xmm9 # 5810 <_sk_callback_sse2+0x260> .byte 69,15,40,209 // movaps %xmm9,%xmm10 .byte 68,15,92,215 // subps %xmm7,%xmm10 .byte 69,15,40,218 // movaps %xmm10,%xmm11 @@ -30347,7 +30387,7 @@ _sk_softlight_sse2: .byte 68,15,40,213 // movaps %xmm5,%xmm10 .byte 68,15,94,215 // divps %xmm7,%xmm10 .byte 69,15,84,212 // andps %xmm12,%xmm10 - .byte 68,15,40,13,227,75,0,0 // movaps 0x4be3(%rip),%xmm9 # 57f0 <_sk_callback_sse2+0x26e> + .byte 68,15,40,13,19,76,0,0 // movaps 0x4c13(%rip),%xmm9 # 5820 <_sk_callback_sse2+0x270> .byte 69,15,40,249 // movaps %xmm9,%xmm15 .byte 69,15,92,250 // subps %xmm10,%xmm15 .byte 69,15,40,218 // movaps %xmm10,%xmm11 @@ -30360,10 +30400,10 @@ _sk_softlight_sse2: .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 15,89,192 // mulps %xmm0,%xmm0 .byte 65,15,88,194 // addps %xmm10,%xmm0 - .byte 68,15,40,53,189,75,0,0 // movaps 0x4bbd(%rip),%xmm14 # 5800 <_sk_callback_sse2+0x27e> + .byte 68,15,40,53,237,75,0,0 // movaps 0x4bed(%rip),%xmm14 # 5830 <_sk_callback_sse2+0x280> .byte 69,15,88,222 // addps %xmm14,%xmm11 .byte 68,15,89,216 // mulps %xmm0,%xmm11 - .byte 68,15,40,21,189,75,0,0 // movaps 0x4bbd(%rip),%xmm10 # 5810 <_sk_callback_sse2+0x28e> + .byte 68,15,40,21,237,75,0,0 // movaps 0x4bed(%rip),%xmm10 # 5840 <_sk_callback_sse2+0x290> .byte 69,15,89,234 // mulps %xmm10,%xmm13 .byte 69,15,88,235 // addps %xmm11,%xmm13 .byte 15,88,228 // addps %xmm4,%xmm4 @@ -30508,7 +30548,7 @@ _sk_hue_sse2: .byte 68,15,40,209 // movaps %xmm1,%xmm10 .byte 68,15,40,225 // movaps %xmm1,%xmm12 .byte 68,15,89,211 // mulps %xmm3,%xmm10 - .byte 68,15,40,5,0,74,0,0 // movaps 0x4a00(%rip),%xmm8 # 5850 <_sk_callback_sse2+0x2ce> + .byte 68,15,40,5,48,74,0,0 // movaps 0x4a30(%rip),%xmm8 # 5880 <_sk_callback_sse2+0x2d0> .byte 69,15,40,216 // movaps %xmm8,%xmm11 .byte 15,40,207 // movaps %xmm7,%xmm1 .byte 68,15,92,217 // subps %xmm1,%xmm11 @@ -30556,12 +30596,12 @@ _sk_hue_sse2: .byte 69,15,84,206 // andps %xmm14,%xmm9 .byte 69,15,84,214 // andps %xmm14,%xmm10 .byte 65,15,84,214 // andps %xmm14,%xmm2 - .byte 68,15,40,61,13,73,0,0 // movaps 0x490d(%rip),%xmm15 # 5820 <_sk_callback_sse2+0x29e> + .byte 68,15,40,61,61,73,0,0 // movaps 0x493d(%rip),%xmm15 # 5850 <_sk_callback_sse2+0x2a0> .byte 65,15,89,231 // mulps %xmm15,%xmm4 - .byte 15,40,5,18,73,0,0 // movaps 0x4912(%rip),%xmm0 # 5830 <_sk_callback_sse2+0x2ae> + .byte 15,40,5,66,73,0,0 // movaps 0x4942(%rip),%xmm0 # 5860 <_sk_callback_sse2+0x2b0> .byte 15,89,240 // mulps %xmm0,%xmm6 .byte 15,88,244 // addps %xmm4,%xmm6 - .byte 68,15,40,53,20,73,0,0 // movaps 0x4914(%rip),%xmm14 # 5840 <_sk_callback_sse2+0x2be> + .byte 68,15,40,53,68,73,0,0 // movaps 0x4944(%rip),%xmm14 # 5870 <_sk_callback_sse2+0x2c0> .byte 68,15,40,239 // movaps %xmm7,%xmm13 .byte 69,15,89,238 // mulps %xmm14,%xmm13 .byte 68,15,88,238 // addps %xmm6,%xmm13 @@ -30739,14 +30779,14 @@ _sk_saturation_sse2: .byte 68,15,84,211 // andps %xmm3,%xmm10 .byte 68,15,84,203 // andps %xmm3,%xmm9 .byte 15,84,195 // andps %xmm3,%xmm0 - .byte 68,15,40,5,166,70,0,0 // movaps 0x46a6(%rip),%xmm8 # 5860 <_sk_callback_sse2+0x2de> + .byte 68,15,40,5,214,70,0,0 // movaps 0x46d6(%rip),%xmm8 # 5890 <_sk_callback_sse2+0x2e0> .byte 15,40,214 // movaps %xmm6,%xmm2 .byte 65,15,89,208 // mulps %xmm8,%xmm2 - .byte 15,40,13,168,70,0,0 // movaps 0x46a8(%rip),%xmm1 # 5870 <_sk_callback_sse2+0x2ee> + .byte 15,40,13,216,70,0,0 // movaps 0x46d8(%rip),%xmm1 # 58a0 <_sk_callback_sse2+0x2f0> .byte 15,40,221 // movaps %xmm5,%xmm3 .byte 15,89,217 // mulps %xmm1,%xmm3 .byte 15,88,218 // addps %xmm2,%xmm3 - .byte 68,15,40,37,167,70,0,0 // movaps 0x46a7(%rip),%xmm12 # 5880 <_sk_callback_sse2+0x2fe> + .byte 68,15,40,37,215,70,0,0 // movaps 0x46d7(%rip),%xmm12 # 58b0 <_sk_callback_sse2+0x300> .byte 69,15,89,236 // mulps %xmm12,%xmm13 .byte 68,15,88,235 // addps %xmm3,%xmm13 .byte 65,15,40,210 // movaps %xmm10,%xmm2 @@ -30791,7 +30831,7 @@ _sk_saturation_sse2: .byte 15,40,223 // movaps %xmm7,%xmm3 .byte 15,40,236 // movaps %xmm4,%xmm5 .byte 15,89,221 // mulps %xmm5,%xmm3 - .byte 68,15,40,5,12,70,0,0 // movaps 0x460c(%rip),%xmm8 # 5890 <_sk_callback_sse2+0x30e> + .byte 68,15,40,5,60,70,0,0 // movaps 0x463c(%rip),%xmm8 # 58c0 <_sk_callback_sse2+0x310> .byte 65,15,40,224 // movaps %xmm8,%xmm4 .byte 68,15,92,199 // subps %xmm7,%xmm8 .byte 15,88,253 // addps %xmm5,%xmm7 @@ -30892,14 +30932,14 @@ _sk_color_sse2: .byte 68,15,40,213 // movaps %xmm5,%xmm10 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,40,208 // movaps %xmm8,%xmm2 - .byte 68,15,40,45,170,68,0,0 // movaps 0x44aa(%rip),%xmm13 # 58a0 <_sk_callback_sse2+0x31e> + .byte 68,15,40,45,218,68,0,0 // movaps 0x44da(%rip),%xmm13 # 58d0 <_sk_callback_sse2+0x320> .byte 68,15,40,198 // movaps %xmm6,%xmm8 .byte 69,15,89,197 // mulps %xmm13,%xmm8 - .byte 68,15,40,53,170,68,0,0 // movaps 0x44aa(%rip),%xmm14 # 58b0 <_sk_callback_sse2+0x32e> + .byte 68,15,40,53,218,68,0,0 // movaps 0x44da(%rip),%xmm14 # 58e0 <_sk_callback_sse2+0x330> .byte 65,15,40,195 // movaps %xmm11,%xmm0 .byte 65,15,89,198 // mulps %xmm14,%xmm0 .byte 65,15,88,192 // addps %xmm8,%xmm0 - .byte 68,15,40,29,166,68,0,0 // movaps 0x44a6(%rip),%xmm11 # 58c0 <_sk_callback_sse2+0x33e> + .byte 68,15,40,29,214,68,0,0 // movaps 0x44d6(%rip),%xmm11 # 58f0 <_sk_callback_sse2+0x340> .byte 69,15,89,227 // mulps %xmm11,%xmm12 .byte 68,15,88,224 // addps %xmm0,%xmm12 .byte 65,15,40,193 // movaps %xmm9,%xmm0 @@ -30907,7 +30947,7 @@ _sk_color_sse2: .byte 69,15,40,250 // movaps %xmm10,%xmm15 .byte 69,15,89,254 // mulps %xmm14,%xmm15 .byte 68,15,88,248 // addps %xmm0,%xmm15 - .byte 68,15,40,5,146,68,0,0 // movaps 0x4492(%rip),%xmm8 # 58d0 <_sk_callback_sse2+0x34e> + .byte 68,15,40,5,194,68,0,0 // movaps 0x44c2(%rip),%xmm8 # 5900 <_sk_callback_sse2+0x350> .byte 65,15,40,224 // movaps %xmm8,%xmm4 .byte 15,92,226 // subps %xmm2,%xmm4 .byte 15,89,252 // mulps %xmm4,%xmm7 @@ -31043,15 +31083,15 @@ _sk_luminosity_sse2: .byte 68,15,40,205 // movaps %xmm5,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 .byte 15,89,222 // mulps %xmm6,%xmm3 - .byte 68,15,40,37,169,66,0,0 // movaps 0x42a9(%rip),%xmm12 # 58e0 <_sk_callback_sse2+0x35e> + .byte 68,15,40,37,217,66,0,0 // movaps 0x42d9(%rip),%xmm12 # 5910 <_sk_callback_sse2+0x360> .byte 68,15,40,199 // movaps %xmm7,%xmm8 .byte 69,15,89,196 // mulps %xmm12,%xmm8 - .byte 68,15,40,45,169,66,0,0 // movaps 0x42a9(%rip),%xmm13 # 58f0 <_sk_callback_sse2+0x36e> + .byte 68,15,40,45,217,66,0,0 // movaps 0x42d9(%rip),%xmm13 # 5920 <_sk_callback_sse2+0x370> .byte 68,15,40,241 // movaps %xmm1,%xmm14 .byte 69,15,89,245 // mulps %xmm13,%xmm14 .byte 69,15,88,240 // addps %xmm8,%xmm14 - .byte 68,15,40,29,165,66,0,0 // movaps 0x42a5(%rip),%xmm11 # 5900 <_sk_callback_sse2+0x37e> - .byte 68,15,40,5,173,66,0,0 // movaps 0x42ad(%rip),%xmm8 # 5910 <_sk_callback_sse2+0x38e> + .byte 68,15,40,29,213,66,0,0 // movaps 0x42d5(%rip),%xmm11 # 5930 <_sk_callback_sse2+0x380> + .byte 68,15,40,5,221,66,0,0 // movaps 0x42dd(%rip),%xmm8 # 5940 <_sk_callback_sse2+0x390> .byte 69,15,40,248 // movaps %xmm8,%xmm15 .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 68,15,92,248 // subps %xmm0,%xmm15 @@ -31190,7 +31230,7 @@ _sk_srcover_rgba_8888_sse2: .byte 15,133,227,0,0,0 // jne 193b <_sk_srcover_rgba_8888_sse2+0xf1> .byte 243,68,15,111,4,144 // movdqu (%rax,%rdx,4),%xmm8 .byte 77,133,192 // test %r8,%r8 - .byte 102,15,111,53,183,64,0,0 // movdqa 0x40b7(%rip),%xmm6 # 5920 <_sk_callback_sse2+0x39e> + .byte 102,15,111,53,231,64,0,0 // movdqa 0x40e7(%rip),%xmm6 # 5950 <_sk_callback_sse2+0x3a0> .byte 102,65,15,111,224 // movdqa %xmm8,%xmm4 .byte 102,15,219,230 // pand %xmm6,%xmm4 .byte 15,91,228 // cvtdq2ps %xmm4,%xmm4 @@ -31204,9 +31244,9 @@ _sk_srcover_rgba_8888_sse2: .byte 15,91,247 // cvtdq2ps %xmm7,%xmm6 .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8 .byte 65,15,91,248 // cvtdq2ps %xmm8,%xmm7 - .byte 68,15,40,5,135,64,0,0 // movaps 0x4087(%rip),%xmm8 # 5930 <_sk_callback_sse2+0x3ae> + .byte 68,15,40,5,183,64,0,0 // movaps 0x40b7(%rip),%xmm8 # 5960 <_sk_callback_sse2+0x3b0> .byte 68,15,92,195 // subps %xmm3,%xmm8 - .byte 68,15,40,37,139,64,0,0 // movaps 0x408b(%rip),%xmm12 # 5940 <_sk_callback_sse2+0x3be> + .byte 68,15,40,37,187,64,0,0 // movaps 0x40bb(%rip),%xmm12 # 5970 <_sk_callback_sse2+0x3c0> .byte 65,15,89,196 // mulps %xmm12,%xmm0 .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 68,15,89,204 // mulps %xmm4,%xmm9 @@ -31289,7 +31329,7 @@ HIDDEN _sk_clamp_1_sse2 .globl _sk_clamp_1_sse2 FUNCTION(_sk_clamp_1_sse2) _sk_clamp_1_sse2: - .byte 68,15,40,5,107,63,0,0 // movaps 0x3f6b(%rip),%xmm8 # 5950 <_sk_callback_sse2+0x3ce> + .byte 68,15,40,5,155,63,0,0 // movaps 0x3f9b(%rip),%xmm8 # 5980 <_sk_callback_sse2+0x3d0> .byte 65,15,93,192 // minps %xmm8,%xmm0 .byte 65,15,93,200 // minps %xmm8,%xmm1 .byte 65,15,93,208 // minps %xmm8,%xmm2 @@ -31301,7 +31341,7 @@ HIDDEN _sk_clamp_a_sse2 .globl _sk_clamp_a_sse2 FUNCTION(_sk_clamp_a_sse2) _sk_clamp_a_sse2: - .byte 15,93,29,96,63,0,0 // minps 0x3f60(%rip),%xmm3 # 5960 <_sk_callback_sse2+0x3de> + .byte 15,93,29,144,63,0,0 // minps 0x3f90(%rip),%xmm3 # 5990 <_sk_callback_sse2+0x3e0> .byte 15,93,195 // minps %xmm3,%xmm0 .byte 15,93,203 // minps %xmm3,%xmm1 .byte 15,93,211 // minps %xmm3,%xmm2 @@ -31388,7 +31428,7 @@ HIDDEN _sk_unpremul_sse2 FUNCTION(_sk_unpremul_sse2) _sk_unpremul_sse2: .byte 69,15,87,192 // xorps %xmm8,%xmm8 - .byte 68,15,40,13,203,62,0,0 // movaps 0x3ecb(%rip),%xmm9 # 5970 <_sk_callback_sse2+0x3ee> + .byte 68,15,40,13,251,62,0,0 // movaps 0x3efb(%rip),%xmm9 # 59a0 <_sk_callback_sse2+0x3f0> .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 @@ -31402,20 +31442,20 @@ HIDDEN _sk_from_srgb_sse2 .globl _sk_from_srgb_sse2 FUNCTION(_sk_from_srgb_sse2) _sk_from_srgb_sse2: - .byte 68,15,40,5,182,62,0,0 // movaps 0x3eb6(%rip),%xmm8 # 5980 <_sk_callback_sse2+0x3fe> + .byte 68,15,40,5,230,62,0,0 // movaps 0x3ee6(%rip),%xmm8 # 59b0 <_sk_callback_sse2+0x400> .byte 68,15,40,232 // movaps %xmm0,%xmm13 .byte 69,15,89,232 // mulps %xmm8,%xmm13 .byte 68,15,40,216 // movaps %xmm0,%xmm11 .byte 69,15,89,219 // mulps %xmm11,%xmm11 - .byte 68,15,40,13,174,62,0,0 // movaps 0x3eae(%rip),%xmm9 # 5990 <_sk_callback_sse2+0x40e> + .byte 68,15,40,13,222,62,0,0 // movaps 0x3ede(%rip),%xmm9 # 59c0 <_sk_callback_sse2+0x410> .byte 68,15,40,240 // movaps %xmm0,%xmm14 .byte 69,15,89,241 // mulps %xmm9,%xmm14 - .byte 68,15,40,21,174,62,0,0 // movaps 0x3eae(%rip),%xmm10 # 59a0 <_sk_callback_sse2+0x41e> + .byte 68,15,40,21,222,62,0,0 // movaps 0x3ede(%rip),%xmm10 # 59d0 <_sk_callback_sse2+0x420> .byte 69,15,88,242 // addps %xmm10,%xmm14 .byte 69,15,89,243 // mulps %xmm11,%xmm14 - .byte 68,15,40,29,174,62,0,0 // movaps 0x3eae(%rip),%xmm11 # 59b0 <_sk_callback_sse2+0x42e> + .byte 68,15,40,29,222,62,0,0 // movaps 0x3ede(%rip),%xmm11 # 59e0 <_sk_callback_sse2+0x430> .byte 69,15,88,243 // addps %xmm11,%xmm14 - .byte 68,15,40,37,178,62,0,0 // movaps 0x3eb2(%rip),%xmm12 # 59c0 <_sk_callback_sse2+0x43e> + .byte 68,15,40,37,226,62,0,0 // movaps 0x3ee2(%rip),%xmm12 # 59f0 <_sk_callback_sse2+0x440> .byte 65,15,194,196,1 // cmpltps %xmm12,%xmm0 .byte 68,15,84,232 // andps %xmm0,%xmm13 .byte 65,15,85,198 // andnps %xmm14,%xmm0 @@ -31452,22 +31492,22 @@ HIDDEN _sk_to_srgb_sse2 FUNCTION(_sk_to_srgb_sse2) _sk_to_srgb_sse2: .byte 68,15,82,232 // rsqrtps %xmm0,%xmm13 - .byte 68,15,40,5,63,62,0,0 // movaps 0x3e3f(%rip),%xmm8 # 59d0 <_sk_callback_sse2+0x44e> + .byte 68,15,40,5,111,62,0,0 // movaps 0x3e6f(%rip),%xmm8 # 5a00 <_sk_callback_sse2+0x450> .byte 68,15,40,240 // movaps %xmm0,%xmm14 .byte 69,15,89,240 // mulps %xmm8,%xmm14 - .byte 68,15,40,13,63,62,0,0 // movaps 0x3e3f(%rip),%xmm9 # 59e0 <_sk_callback_sse2+0x45e> + .byte 68,15,40,13,111,62,0,0 // movaps 0x3e6f(%rip),%xmm9 # 5a10 <_sk_callback_sse2+0x460> .byte 69,15,40,253 // movaps %xmm13,%xmm15 .byte 69,15,89,249 // mulps %xmm9,%xmm15 - .byte 68,15,40,21,63,62,0,0 // movaps 0x3e3f(%rip),%xmm10 # 59f0 <_sk_callback_sse2+0x46e> + .byte 68,15,40,21,111,62,0,0 // movaps 0x3e6f(%rip),%xmm10 # 5a20 <_sk_callback_sse2+0x470> .byte 69,15,88,250 // addps %xmm10,%xmm15 .byte 69,15,89,253 // mulps %xmm13,%xmm15 - .byte 68,15,40,29,63,62,0,0 // movaps 0x3e3f(%rip),%xmm11 # 5a00 <_sk_callback_sse2+0x47e> + .byte 68,15,40,29,111,62,0,0 // movaps 0x3e6f(%rip),%xmm11 # 5a30 <_sk_callback_sse2+0x480> .byte 69,15,88,251 // addps %xmm11,%xmm15 - .byte 68,15,40,37,67,62,0,0 // movaps 0x3e43(%rip),%xmm12 # 5a10 <_sk_callback_sse2+0x48e> + .byte 68,15,40,37,115,62,0,0 // movaps 0x3e73(%rip),%xmm12 # 5a40 <_sk_callback_sse2+0x490> .byte 69,15,88,236 // addps %xmm12,%xmm13 .byte 69,15,83,237 // rcpps %xmm13,%xmm13 .byte 69,15,89,239 // mulps %xmm15,%xmm13 - .byte 68,15,40,61,63,62,0,0 // movaps 0x3e3f(%rip),%xmm15 # 5a20 <_sk_callback_sse2+0x49e> + .byte 68,15,40,61,111,62,0,0 // movaps 0x3e6f(%rip),%xmm15 # 5a50 <_sk_callback_sse2+0x4a0> .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 .byte 68,15,84,240 // andps %xmm0,%xmm14 .byte 65,15,85,197 // andnps %xmm13,%xmm0 @@ -31517,7 +31557,7 @@ _sk_rgb_to_hsl_sse2: .byte 68,15,93,218 // minps %xmm2,%xmm11 .byte 65,15,40,202 // movaps %xmm10,%xmm1 .byte 65,15,92,203 // subps %xmm11,%xmm1 - .byte 68,15,40,45,152,61,0,0 // movaps 0x3d98(%rip),%xmm13 # 5a30 <_sk_callback_sse2+0x4ae> + .byte 68,15,40,45,200,61,0,0 // movaps 0x3dc8(%rip),%xmm13 # 5a60 <_sk_callback_sse2+0x4b0> .byte 68,15,94,233 // divps %xmm1,%xmm13 .byte 65,15,40,194 // movaps %xmm10,%xmm0 .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 @@ -31526,30 +31566,30 @@ _sk_rgb_to_hsl_sse2: .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,40,241 // movaps %xmm9,%xmm14 .byte 68,15,194,242,1 // cmpltps %xmm2,%xmm14 - .byte 68,15,84,53,126,61,0,0 // andps 0x3d7e(%rip),%xmm14 # 5a40 <_sk_callback_sse2+0x4be> + .byte 68,15,84,53,174,61,0,0 // andps 0x3dae(%rip),%xmm14 # 5a70 <_sk_callback_sse2+0x4c0> .byte 69,15,88,244 // addps %xmm12,%xmm14 .byte 69,15,40,250 // movaps %xmm10,%xmm15 .byte 69,15,194,249,0 // cmpeqps %xmm9,%xmm15 .byte 65,15,92,208 // subps %xmm8,%xmm2 .byte 65,15,89,213 // mulps %xmm13,%xmm2 - .byte 68,15,40,37,113,61,0,0 // movaps 0x3d71(%rip),%xmm12 # 5a50 <_sk_callback_sse2+0x4ce> + .byte 68,15,40,37,161,61,0,0 // movaps 0x3da1(%rip),%xmm12 # 5a80 <_sk_callback_sse2+0x4d0> .byte 65,15,88,212 // addps %xmm12,%xmm2 .byte 69,15,92,193 // subps %xmm9,%xmm8 .byte 69,15,89,197 // mulps %xmm13,%xmm8 - .byte 68,15,88,5,109,61,0,0 // addps 0x3d6d(%rip),%xmm8 # 5a60 <_sk_callback_sse2+0x4de> + .byte 68,15,88,5,157,61,0,0 // addps 0x3d9d(%rip),%xmm8 # 5a90 <_sk_callback_sse2+0x4e0> .byte 65,15,84,215 // andps %xmm15,%xmm2 .byte 69,15,85,248 // andnps %xmm8,%xmm15 .byte 68,15,86,250 // orps %xmm2,%xmm15 .byte 68,15,84,240 // andps %xmm0,%xmm14 .byte 65,15,85,199 // andnps %xmm15,%xmm0 .byte 65,15,86,198 // orps %xmm14,%xmm0 - .byte 15,89,5,94,61,0,0 // mulps 0x3d5e(%rip),%xmm0 # 5a70 <_sk_callback_sse2+0x4ee> + .byte 15,89,5,142,61,0,0 // mulps 0x3d8e(%rip),%xmm0 # 5aa0 <_sk_callback_sse2+0x4f0> .byte 69,15,40,194 // movaps %xmm10,%xmm8 .byte 69,15,194,195,4 // cmpneqps %xmm11,%xmm8 .byte 65,15,84,192 // andps %xmm8,%xmm0 .byte 69,15,92,226 // subps %xmm10,%xmm12 .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 68,15,40,13,81,61,0,0 // movaps 0x3d51(%rip),%xmm9 # 5a80 <_sk_callback_sse2+0x4fe> + .byte 68,15,40,13,129,61,0,0 // movaps 0x3d81(%rip),%xmm9 # 5ab0 <_sk_callback_sse2+0x500> .byte 65,15,40,210 // movaps %xmm10,%xmm2 .byte 65,15,89,209 // mulps %xmm9,%xmm2 .byte 68,15,194,202,1 // cmpltps %xmm2,%xmm9 @@ -31573,7 +31613,7 @@ _sk_hsl_to_rgb_sse2: .byte 15,41,92,36,168 // movaps %xmm3,-0x58(%rsp) .byte 68,15,40,218 // movaps %xmm2,%xmm11 .byte 15,40,240 // movaps %xmm0,%xmm6 - .byte 68,15,40,13,16,61,0,0 // movaps 0x3d10(%rip),%xmm9 # 5a90 <_sk_callback_sse2+0x50e> + .byte 68,15,40,13,64,61,0,0 // movaps 0x3d40(%rip),%xmm9 # 5ac0 <_sk_callback_sse2+0x510> .byte 69,15,40,209 // movaps %xmm9,%xmm10 .byte 69,15,194,211,2 // cmpleps %xmm11,%xmm10 .byte 15,40,193 // movaps %xmm1,%xmm0 @@ -31590,28 +31630,28 @@ _sk_hsl_to_rgb_sse2: .byte 69,15,88,211 // addps %xmm11,%xmm10 .byte 69,15,88,219 // addps %xmm11,%xmm11 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 15,40,5,217,60,0,0 // movaps 0x3cd9(%rip),%xmm0 # 5aa0 <_sk_callback_sse2+0x51e> + .byte 15,40,5,9,61,0,0 // movaps 0x3d09(%rip),%xmm0 # 5ad0 <_sk_callback_sse2+0x520> .byte 15,88,198 // addps %xmm6,%xmm0 .byte 243,15,91,200 // cvttps2dq %xmm0,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 .byte 15,40,216 // movaps %xmm0,%xmm3 .byte 15,194,217,1 // cmpltps %xmm1,%xmm3 - .byte 15,84,29,209,60,0,0 // andps 0x3cd1(%rip),%xmm3 # 5ab0 <_sk_callback_sse2+0x52e> + .byte 15,84,29,1,61,0,0 // andps 0x3d01(%rip),%xmm3 # 5ae0 <_sk_callback_sse2+0x530> .byte 15,92,203 // subps %xmm3,%xmm1 .byte 15,92,193 // subps %xmm1,%xmm0 - .byte 68,15,40,45,211,60,0,0 // movaps 0x3cd3(%rip),%xmm13 # 5ac0 <_sk_callback_sse2+0x53e> + .byte 68,15,40,45,3,61,0,0 // movaps 0x3d03(%rip),%xmm13 # 5af0 <_sk_callback_sse2+0x540> .byte 69,15,40,197 // movaps %xmm13,%xmm8 .byte 68,15,194,192,2 // cmpleps %xmm0,%xmm8 .byte 69,15,40,242 // movaps %xmm10,%xmm14 .byte 69,15,92,243 // subps %xmm11,%xmm14 .byte 65,15,40,217 // movaps %xmm9,%xmm3 .byte 15,194,216,2 // cmpleps %xmm0,%xmm3 - .byte 15,40,21,227,60,0,0 // movaps 0x3ce3(%rip),%xmm2 # 5af0 <_sk_callback_sse2+0x56e> + .byte 15,40,21,19,61,0,0 // movaps 0x3d13(%rip),%xmm2 # 5b20 <_sk_callback_sse2+0x570> .byte 68,15,40,250 // movaps %xmm2,%xmm15 .byte 68,15,194,248,2 // cmpleps %xmm0,%xmm15 - .byte 15,40,13,179,60,0,0 // movaps 0x3cb3(%rip),%xmm1 # 5ad0 <_sk_callback_sse2+0x54e> + .byte 15,40,13,227,60,0,0 // movaps 0x3ce3(%rip),%xmm1 # 5b00 <_sk_callback_sse2+0x550> .byte 15,89,193 // mulps %xmm1,%xmm0 - .byte 15,40,45,185,60,0,0 // movaps 0x3cb9(%rip),%xmm5 # 5ae0 <_sk_callback_sse2+0x55e> + .byte 15,40,45,233,60,0,0 // movaps 0x3ce9(%rip),%xmm5 # 5b10 <_sk_callback_sse2+0x560> .byte 15,40,229 // movaps %xmm5,%xmm4 .byte 15,92,224 // subps %xmm0,%xmm4 .byte 65,15,89,230 // mulps %xmm14,%xmm4 @@ -31634,7 +31674,7 @@ _sk_hsl_to_rgb_sse2: .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 .byte 15,40,222 // movaps %xmm6,%xmm3 .byte 15,194,216,1 // cmpltps %xmm0,%xmm3 - .byte 15,84,29,46,60,0,0 // andps 0x3c2e(%rip),%xmm3 # 5ab0 <_sk_callback_sse2+0x52e> + .byte 15,84,29,94,60,0,0 // andps 0x3c5e(%rip),%xmm3 # 5ae0 <_sk_callback_sse2+0x530> .byte 15,92,195 // subps %xmm3,%xmm0 .byte 68,15,40,230 // movaps %xmm6,%xmm12 .byte 68,15,92,224 // subps %xmm0,%xmm12 @@ -31664,12 +31704,12 @@ _sk_hsl_to_rgb_sse2: .byte 15,40,124,36,136 // movaps -0x78(%rsp),%xmm7 .byte 15,40,231 // movaps %xmm7,%xmm4 .byte 15,85,227 // andnps %xmm3,%xmm4 - .byte 15,88,53,6,60,0,0 // addps 0x3c06(%rip),%xmm6 # 5b00 <_sk_callback_sse2+0x57e> + .byte 15,88,53,54,60,0,0 // addps 0x3c36(%rip),%xmm6 # 5b30 <_sk_callback_sse2+0x580> .byte 243,15,91,198 // cvttps2dq %xmm6,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 .byte 15,40,222 // movaps %xmm6,%xmm3 .byte 15,194,216,1 // cmpltps %xmm0,%xmm3 - .byte 15,84,29,161,59,0,0 // andps 0x3ba1(%rip),%xmm3 # 5ab0 <_sk_callback_sse2+0x52e> + .byte 15,84,29,209,59,0,0 // andps 0x3bd1(%rip),%xmm3 # 5ae0 <_sk_callback_sse2+0x530> .byte 15,92,195 // subps %xmm3,%xmm0 .byte 15,92,240 // subps %xmm0,%xmm6 .byte 15,89,206 // mulps %xmm6,%xmm1 @@ -31733,9 +31773,9 @@ _sk_scale_u8_sse2: .byte 102,69,15,110,4,18 // movd (%r10,%rdx,1),%xmm8 .byte 102,68,15,96,192 // punpcklbw %xmm0,%xmm8 .byte 102,68,15,97,192 // punpcklwd %xmm0,%xmm8 - .byte 102,68,15,219,5,50,59,0,0 // pand 0x3b32(%rip),%xmm8 # 5b10 <_sk_callback_sse2+0x58e> + .byte 102,68,15,219,5,98,59,0,0 // pand 0x3b62(%rip),%xmm8 # 5b40 <_sk_callback_sse2+0x590> .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,54,59,0,0 // mulps 0x3b36(%rip),%xmm8 # 5b20 <_sk_callback_sse2+0x59e> + .byte 68,15,89,5,102,59,0,0 // mulps 0x3b66(%rip),%xmm8 # 5b50 <_sk_callback_sse2+0x5a0> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 65,15,89,208 // mulps %xmm8,%xmm2 @@ -31798,9 +31838,9 @@ _sk_lerp_u8_sse2: .byte 102,69,15,110,4,18 // movd (%r10,%rdx,1),%xmm8 .byte 102,68,15,96,192 // punpcklbw %xmm0,%xmm8 .byte 102,68,15,97,192 // punpcklwd %xmm0,%xmm8 - .byte 102,68,15,219,5,119,58,0,0 // pand 0x3a77(%rip),%xmm8 # 5b30 <_sk_callback_sse2+0x5ae> + .byte 102,68,15,219,5,167,58,0,0 // pand 0x3aa7(%rip),%xmm8 # 5b60 <_sk_callback_sse2+0x5b0> .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,123,58,0,0 // mulps 0x3a7b(%rip),%xmm8 # 5b40 <_sk_callback_sse2+0x5be> + .byte 68,15,89,5,171,58,0,0 // mulps 0x3aab(%rip),%xmm8 # 5b70 <_sk_callback_sse2+0x5c0> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -31847,17 +31887,17 @@ _sk_lerp_565_sse2: .byte 15,133,156,0,0,0 // jne 21f7 <_sk_lerp_565_sse2+0xaa> .byte 243,69,15,126,12,82 // movq (%r10,%rdx,2),%xmm9 .byte 102,68,15,97,200 // punpcklwd %xmm0,%xmm9 - .byte 102,68,15,111,5,225,57,0,0 // movdqa 0x39e1(%rip),%xmm8 # 5b50 <_sk_callback_sse2+0x5ce> + .byte 102,68,15,111,5,17,58,0,0 // movdqa 0x3a11(%rip),%xmm8 # 5b80 <_sk_callback_sse2+0x5d0> .byte 102,69,15,219,193 // pand %xmm9,%xmm8 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 - .byte 68,15,89,5,224,57,0,0 // mulps 0x39e0(%rip),%xmm8 # 5b60 <_sk_callback_sse2+0x5de> - .byte 102,68,15,111,21,231,57,0,0 // movdqa 0x39e7(%rip),%xmm10 # 5b70 <_sk_callback_sse2+0x5ee> + .byte 68,15,89,5,16,58,0,0 // mulps 0x3a10(%rip),%xmm8 # 5b90 <_sk_callback_sse2+0x5e0> + .byte 102,68,15,111,21,23,58,0,0 // movdqa 0x3a17(%rip),%xmm10 # 5ba0 <_sk_callback_sse2+0x5f0> .byte 102,69,15,219,209 // pand %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 - .byte 68,15,89,21,230,57,0,0 // mulps 0x39e6(%rip),%xmm10 # 5b80 <_sk_callback_sse2+0x5fe> - .byte 102,68,15,219,13,237,57,0,0 // pand 0x39ed(%rip),%xmm9 # 5b90 <_sk_callback_sse2+0x60e> + .byte 68,15,89,21,22,58,0,0 // mulps 0x3a16(%rip),%xmm10 # 5bb0 <_sk_callback_sse2+0x600> + .byte 102,68,15,219,13,29,58,0,0 // pand 0x3a1d(%rip),%xmm9 # 5bc0 <_sk_callback_sse2+0x610> .byte 69,15,91,201 // cvtdq2ps %xmm9,%xmm9 - .byte 68,15,89,13,241,57,0,0 // mulps 0x39f1(%rip),%xmm9 # 5ba0 <_sk_callback_sse2+0x61e> + .byte 68,15,89,13,33,58,0,0 // mulps 0x3a21(%rip),%xmm9 # 5bd0 <_sk_callback_sse2+0x620> .byte 15,92,196 // subps %xmm4,%xmm0 .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 15,88,196 // addps %xmm4,%xmm0 @@ -31913,7 +31953,7 @@ _sk_load_tables_sse2: .byte 65,87 // push %r15 .byte 65,86 // push %r14 .byte 83 // push %rbx - .byte 102,68,15,111,5,55,57,0,0 // movdqa 0x3937(%rip),%xmm8 # 5bb0 <_sk_callback_sse2+0x62e> + .byte 102,68,15,111,5,103,57,0,0 // movdqa 0x3967(%rip),%xmm8 # 5be0 <_sk_callback_sse2+0x630> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 @@ -31969,7 +32009,7 @@ _sk_load_tables_sse2: .byte 65,15,20,208 // unpcklps %xmm8,%xmm2 .byte 102,65,15,114,209,24 // psrld $0x18,%xmm9 .byte 65,15,91,217 // cvtdq2ps %xmm9,%xmm3 - .byte 15,89,29,64,56,0,0 // mulps 0x3840(%rip),%xmm3 # 5bc0 <_sk_callback_sse2+0x63e> + .byte 15,89,29,112,56,0,0 // mulps 0x3870(%rip),%xmm3 # 5bf0 <_sk_callback_sse2+0x640> .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx .byte 65,94 // pop %r14 @@ -32014,7 +32054,7 @@ _sk_load_tables_u16_be_sse2: .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9 - .byte 102,68,15,111,21,166,55,0,0 // movdqa 0x37a6(%rip),%xmm10 # 5bd0 <_sk_callback_sse2+0x64e> + .byte 102,68,15,111,21,214,55,0,0 // movdqa 0x37d6(%rip),%xmm10 # 5c00 <_sk_callback_sse2+0x650> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,194 // pand %xmm10,%xmm0 .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 @@ -32076,7 +32116,7 @@ _sk_load_tables_u16_be_sse2: .byte 102,65,15,235,217 // por %xmm9,%xmm3 .byte 102,65,15,97,216 // punpcklwd %xmm8,%xmm3 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,145,54,0,0 // mulps 0x3691(%rip),%xmm3 # 5be0 <_sk_callback_sse2+0x65e> + .byte 15,89,29,193,54,0,0 // mulps 0x36c1(%rip),%xmm3 # 5c10 <_sk_callback_sse2+0x660> .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx .byte 65,94 // pop %r14 @@ -32119,7 +32159,7 @@ _sk_load_tables_rgb_u16_be_sse2: .byte 102,68,15,97,208 // punpcklwd %xmm0,%xmm10 .byte 102,65,15,111,195 // movdqa %xmm11,%xmm0 .byte 102,65,15,97,194 // punpcklwd %xmm10,%xmm0 - .byte 102,68,15,111,5,0,54,0,0 // movdqa 0x3600(%rip),%xmm8 # 5bf0 <_sk_callback_sse2+0x66e> + .byte 102,68,15,111,5,48,54,0,0 // movdqa 0x3630(%rip),%xmm8 # 5c20 <_sk_callback_sse2+0x670> .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 .byte 102,65,15,219,192 // pand %xmm8,%xmm0 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 @@ -32175,7 +32215,7 @@ _sk_load_tables_rgb_u16_be_sse2: .byte 15,20,211 // unpcklps %xmm3,%xmm2 .byte 65,15,20,208 // unpcklps %xmm8,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,11,53,0,0 // movaps 0x350b(%rip),%xmm3 # 5c00 <_sk_callback_sse2+0x67e> + .byte 15,40,29,59,53,0,0 // movaps 0x353b(%rip),%xmm3 # 5c30 <_sk_callback_sse2+0x680> .byte 91 // pop %rbx .byte 65,94 // pop %r14 .byte 65,95 // pop %r15 @@ -32209,7 +32249,7 @@ _sk_byte_tables_sse2: .byte 65,86 // push %r14 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,164,52,0,0 // movaps 0x34a4(%rip),%xmm8 # 5c10 <_sk_callback_sse2+0x68e> + .byte 68,15,40,5,212,52,0,0 // movaps 0x34d4(%rip),%xmm8 # 5c40 <_sk_callback_sse2+0x690> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 .byte 102,73,15,126,193 // movq %xmm0,%r9 @@ -32237,7 +32277,7 @@ _sk_byte_tables_sse2: .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,21,62,52,0,0 // movaps 0x343e(%rip),%xmm10 # 5c20 <_sk_callback_sse2+0x69e> + .byte 68,15,40,21,110,52,0,0 // movaps 0x346e(%rip),%xmm10 # 5c50 <_sk_callback_sse2+0x6a0> .byte 65,15,89,194 // mulps %xmm10,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -32358,7 +32398,7 @@ _sk_byte_tables_rgb_sse2: .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,21,134,50,0,0 // movaps 0x3286(%rip),%xmm10 # 5c30 <_sk_callback_sse2+0x6ae> + .byte 68,15,40,21,182,50,0,0 // movaps 0x32b6(%rip),%xmm10 # 5c60 <_sk_callback_sse2+0x6b0> .byte 65,15,89,194 // mulps %xmm10,%xmm0 .byte 65,15,89,200 // mulps %xmm8,%xmm1 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 @@ -32565,15 +32605,15 @@ _sk_parametric_r_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,186,47,0,0 // mulps 0x2fba(%rip),%xmm9 # 5c40 <_sk_callback_sse2+0x6be> - .byte 68,15,84,21,194,47,0,0 // andps 0x2fc2(%rip),%xmm10 # 5c50 <_sk_callback_sse2+0x6ce> - .byte 68,15,86,21,202,47,0,0 // orps 0x2fca(%rip),%xmm10 # 5c60 <_sk_callback_sse2+0x6de> - .byte 68,15,88,13,210,47,0,0 // addps 0x2fd2(%rip),%xmm9 # 5c70 <_sk_callback_sse2+0x6ee> - .byte 68,15,40,37,218,47,0,0 // movaps 0x2fda(%rip),%xmm12 # 5c80 <_sk_callback_sse2+0x6fe> + .byte 68,15,89,13,234,47,0,0 // mulps 0x2fea(%rip),%xmm9 # 5c70 <_sk_callback_sse2+0x6c0> + .byte 68,15,84,21,242,47,0,0 // andps 0x2ff2(%rip),%xmm10 # 5c80 <_sk_callback_sse2+0x6d0> + .byte 68,15,86,21,250,47,0,0 // orps 0x2ffa(%rip),%xmm10 # 5c90 <_sk_callback_sse2+0x6e0> + .byte 68,15,88,13,2,48,0,0 // addps 0x3002(%rip),%xmm9 # 5ca0 <_sk_callback_sse2+0x6f0> + .byte 68,15,40,37,10,48,0,0 // movaps 0x300a(%rip),%xmm12 # 5cb0 <_sk_callback_sse2+0x700> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,218,47,0,0 // addps 0x2fda(%rip),%xmm10 # 5c90 <_sk_callback_sse2+0x70e> - .byte 68,15,40,37,226,47,0,0 // movaps 0x2fe2(%rip),%xmm12 # 5ca0 <_sk_callback_sse2+0x71e> + .byte 68,15,88,21,10,48,0,0 // addps 0x300a(%rip),%xmm10 # 5cc0 <_sk_callback_sse2+0x710> + .byte 68,15,40,37,18,48,0,0 // movaps 0x3012(%rip),%xmm12 # 5cd0 <_sk_callback_sse2+0x720> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -32581,22 +32621,22 @@ _sk_parametric_r_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,204,47,0,0 // movaps 0x2fcc(%rip),%xmm10 # 5cb0 <_sk_callback_sse2+0x72e> + .byte 68,15,40,21,252,47,0,0 // movaps 0x2ffc(%rip),%xmm10 # 5ce0 <_sk_callback_sse2+0x730> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,192,47,0,0 // addps 0x2fc0(%rip),%xmm9 # 5cc0 <_sk_callback_sse2+0x73e> - .byte 68,15,40,37,200,47,0,0 // movaps 0x2fc8(%rip),%xmm12 # 5cd0 <_sk_callback_sse2+0x74e> + .byte 68,15,88,13,240,47,0,0 // addps 0x2ff0(%rip),%xmm9 # 5cf0 <_sk_callback_sse2+0x740> + .byte 68,15,40,37,248,47,0,0 // movaps 0x2ff8(%rip),%xmm12 # 5d00 <_sk_callback_sse2+0x750> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,200,47,0,0 // movaps 0x2fc8(%rip),%xmm12 # 5ce0 <_sk_callback_sse2+0x75e> + .byte 68,15,40,37,248,47,0,0 // movaps 0x2ff8(%rip),%xmm12 # 5d10 <_sk_callback_sse2+0x760> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,204,47,0,0 // movaps 0x2fcc(%rip),%xmm13 # 5cf0 <_sk_callback_sse2+0x76e> + .byte 68,15,40,45,252,47,0,0 // movaps 0x2ffc(%rip),%xmm13 # 5d20 <_sk_callback_sse2+0x770> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,204,47,0,0 // mulps 0x2fcc(%rip),%xmm13 # 5d00 <_sk_callback_sse2+0x77e> + .byte 68,15,89,45,252,47,0,0 // mulps 0x2ffc(%rip),%xmm13 # 5d30 <_sk_callback_sse2+0x780> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -32632,15 +32672,15 @@ _sk_parametric_g_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,76,47,0,0 // mulps 0x2f4c(%rip),%xmm9 # 5d10 <_sk_callback_sse2+0x78e> - .byte 68,15,84,21,84,47,0,0 // andps 0x2f54(%rip),%xmm10 # 5d20 <_sk_callback_sse2+0x79e> - .byte 68,15,86,21,92,47,0,0 // orps 0x2f5c(%rip),%xmm10 # 5d30 <_sk_callback_sse2+0x7ae> - .byte 68,15,88,13,100,47,0,0 // addps 0x2f64(%rip),%xmm9 # 5d40 <_sk_callback_sse2+0x7be> - .byte 68,15,40,37,108,47,0,0 // movaps 0x2f6c(%rip),%xmm12 # 5d50 <_sk_callback_sse2+0x7ce> + .byte 68,15,89,13,124,47,0,0 // mulps 0x2f7c(%rip),%xmm9 # 5d40 <_sk_callback_sse2+0x790> + .byte 68,15,84,21,132,47,0,0 // andps 0x2f84(%rip),%xmm10 # 5d50 <_sk_callback_sse2+0x7a0> + .byte 68,15,86,21,140,47,0,0 // orps 0x2f8c(%rip),%xmm10 # 5d60 <_sk_callback_sse2+0x7b0> + .byte 68,15,88,13,148,47,0,0 // addps 0x2f94(%rip),%xmm9 # 5d70 <_sk_callback_sse2+0x7c0> + .byte 68,15,40,37,156,47,0,0 // movaps 0x2f9c(%rip),%xmm12 # 5d80 <_sk_callback_sse2+0x7d0> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,108,47,0,0 // addps 0x2f6c(%rip),%xmm10 # 5d60 <_sk_callback_sse2+0x7de> - .byte 68,15,40,37,116,47,0,0 // movaps 0x2f74(%rip),%xmm12 # 5d70 <_sk_callback_sse2+0x7ee> + .byte 68,15,88,21,156,47,0,0 // addps 0x2f9c(%rip),%xmm10 # 5d90 <_sk_callback_sse2+0x7e0> + .byte 68,15,40,37,164,47,0,0 // movaps 0x2fa4(%rip),%xmm12 # 5da0 <_sk_callback_sse2+0x7f0> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -32648,22 +32688,22 @@ _sk_parametric_g_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,94,47,0,0 // movaps 0x2f5e(%rip),%xmm10 # 5d80 <_sk_callback_sse2+0x7fe> + .byte 68,15,40,21,142,47,0,0 // movaps 0x2f8e(%rip),%xmm10 # 5db0 <_sk_callback_sse2+0x800> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,82,47,0,0 // addps 0x2f52(%rip),%xmm9 # 5d90 <_sk_callback_sse2+0x80e> - .byte 68,15,40,37,90,47,0,0 // movaps 0x2f5a(%rip),%xmm12 # 5da0 <_sk_callback_sse2+0x81e> + .byte 68,15,88,13,130,47,0,0 // addps 0x2f82(%rip),%xmm9 # 5dc0 <_sk_callback_sse2+0x810> + .byte 68,15,40,37,138,47,0,0 // movaps 0x2f8a(%rip),%xmm12 # 5dd0 <_sk_callback_sse2+0x820> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,90,47,0,0 // movaps 0x2f5a(%rip),%xmm12 # 5db0 <_sk_callback_sse2+0x82e> + .byte 68,15,40,37,138,47,0,0 // movaps 0x2f8a(%rip),%xmm12 # 5de0 <_sk_callback_sse2+0x830> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,94,47,0,0 // movaps 0x2f5e(%rip),%xmm13 # 5dc0 <_sk_callback_sse2+0x83e> + .byte 68,15,40,45,142,47,0,0 // movaps 0x2f8e(%rip),%xmm13 # 5df0 <_sk_callback_sse2+0x840> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,94,47,0,0 // mulps 0x2f5e(%rip),%xmm13 # 5dd0 <_sk_callback_sse2+0x84e> + .byte 68,15,89,45,142,47,0,0 // mulps 0x2f8e(%rip),%xmm13 # 5e00 <_sk_callback_sse2+0x850> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -32699,15 +32739,15 @@ _sk_parametric_b_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,222,46,0,0 // mulps 0x2ede(%rip),%xmm9 # 5de0 <_sk_callback_sse2+0x85e> - .byte 68,15,84,21,230,46,0,0 // andps 0x2ee6(%rip),%xmm10 # 5df0 <_sk_callback_sse2+0x86e> - .byte 68,15,86,21,238,46,0,0 // orps 0x2eee(%rip),%xmm10 # 5e00 <_sk_callback_sse2+0x87e> - .byte 68,15,88,13,246,46,0,0 // addps 0x2ef6(%rip),%xmm9 # 5e10 <_sk_callback_sse2+0x88e> - .byte 68,15,40,37,254,46,0,0 // movaps 0x2efe(%rip),%xmm12 # 5e20 <_sk_callback_sse2+0x89e> + .byte 68,15,89,13,14,47,0,0 // mulps 0x2f0e(%rip),%xmm9 # 5e10 <_sk_callback_sse2+0x860> + .byte 68,15,84,21,22,47,0,0 // andps 0x2f16(%rip),%xmm10 # 5e20 <_sk_callback_sse2+0x870> + .byte 68,15,86,21,30,47,0,0 // orps 0x2f1e(%rip),%xmm10 # 5e30 <_sk_callback_sse2+0x880> + .byte 68,15,88,13,38,47,0,0 // addps 0x2f26(%rip),%xmm9 # 5e40 <_sk_callback_sse2+0x890> + .byte 68,15,40,37,46,47,0,0 // movaps 0x2f2e(%rip),%xmm12 # 5e50 <_sk_callback_sse2+0x8a0> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,254,46,0,0 // addps 0x2efe(%rip),%xmm10 # 5e30 <_sk_callback_sse2+0x8ae> - .byte 68,15,40,37,6,47,0,0 // movaps 0x2f06(%rip),%xmm12 # 5e40 <_sk_callback_sse2+0x8be> + .byte 68,15,88,21,46,47,0,0 // addps 0x2f2e(%rip),%xmm10 # 5e60 <_sk_callback_sse2+0x8b0> + .byte 68,15,40,37,54,47,0,0 // movaps 0x2f36(%rip),%xmm12 # 5e70 <_sk_callback_sse2+0x8c0> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -32715,22 +32755,22 @@ _sk_parametric_b_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,240,46,0,0 // movaps 0x2ef0(%rip),%xmm10 # 5e50 <_sk_callback_sse2+0x8ce> + .byte 68,15,40,21,32,47,0,0 // movaps 0x2f20(%rip),%xmm10 # 5e80 <_sk_callback_sse2+0x8d0> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,228,46,0,0 // addps 0x2ee4(%rip),%xmm9 # 5e60 <_sk_callback_sse2+0x8de> - .byte 68,15,40,37,236,46,0,0 // movaps 0x2eec(%rip),%xmm12 # 5e70 <_sk_callback_sse2+0x8ee> + .byte 68,15,88,13,20,47,0,0 // addps 0x2f14(%rip),%xmm9 # 5e90 <_sk_callback_sse2+0x8e0> + .byte 68,15,40,37,28,47,0,0 // movaps 0x2f1c(%rip),%xmm12 # 5ea0 <_sk_callback_sse2+0x8f0> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,236,46,0,0 // movaps 0x2eec(%rip),%xmm12 # 5e80 <_sk_callback_sse2+0x8fe> + .byte 68,15,40,37,28,47,0,0 // movaps 0x2f1c(%rip),%xmm12 # 5eb0 <_sk_callback_sse2+0x900> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,240,46,0,0 // movaps 0x2ef0(%rip),%xmm13 # 5e90 <_sk_callback_sse2+0x90e> + .byte 68,15,40,45,32,47,0,0 // movaps 0x2f20(%rip),%xmm13 # 5ec0 <_sk_callback_sse2+0x910> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,240,46,0,0 // mulps 0x2ef0(%rip),%xmm13 # 5ea0 <_sk_callback_sse2+0x91e> + .byte 68,15,89,45,32,47,0,0 // mulps 0x2f20(%rip),%xmm13 # 5ed0 <_sk_callback_sse2+0x920> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -32766,15 +32806,15 @@ _sk_parametric_a_sse2: .byte 69,15,88,209 // addps %xmm9,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9 - .byte 68,15,89,13,112,46,0,0 // mulps 0x2e70(%rip),%xmm9 # 5eb0 <_sk_callback_sse2+0x92e> - .byte 68,15,84,21,120,46,0,0 // andps 0x2e78(%rip),%xmm10 # 5ec0 <_sk_callback_sse2+0x93e> - .byte 68,15,86,21,128,46,0,0 // orps 0x2e80(%rip),%xmm10 # 5ed0 <_sk_callback_sse2+0x94e> - .byte 68,15,88,13,136,46,0,0 // addps 0x2e88(%rip),%xmm9 # 5ee0 <_sk_callback_sse2+0x95e> - .byte 68,15,40,37,144,46,0,0 // movaps 0x2e90(%rip),%xmm12 # 5ef0 <_sk_callback_sse2+0x96e> + .byte 68,15,89,13,160,46,0,0 // mulps 0x2ea0(%rip),%xmm9 # 5ee0 <_sk_callback_sse2+0x930> + .byte 68,15,84,21,168,46,0,0 // andps 0x2ea8(%rip),%xmm10 # 5ef0 <_sk_callback_sse2+0x940> + .byte 68,15,86,21,176,46,0,0 // orps 0x2eb0(%rip),%xmm10 # 5f00 <_sk_callback_sse2+0x950> + .byte 68,15,88,13,184,46,0,0 // addps 0x2eb8(%rip),%xmm9 # 5f10 <_sk_callback_sse2+0x960> + .byte 68,15,40,37,192,46,0,0 // movaps 0x2ec0(%rip),%xmm12 # 5f20 <_sk_callback_sse2+0x970> .byte 69,15,89,226 // mulps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,88,21,144,46,0,0 // addps 0x2e90(%rip),%xmm10 # 5f00 <_sk_callback_sse2+0x97e> - .byte 68,15,40,37,152,46,0,0 // movaps 0x2e98(%rip),%xmm12 # 5f10 <_sk_callback_sse2+0x98e> + .byte 68,15,88,21,192,46,0,0 // addps 0x2ec0(%rip),%xmm10 # 5f30 <_sk_callback_sse2+0x980> + .byte 68,15,40,37,200,46,0,0 // movaps 0x2ec8(%rip),%xmm12 # 5f40 <_sk_callback_sse2+0x990> .byte 69,15,94,226 // divps %xmm10,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 .byte 69,15,89,203 // mulps %xmm11,%xmm9 @@ -32782,22 +32822,22 @@ _sk_parametric_a_sse2: .byte 69,15,91,226 // cvtdq2ps %xmm10,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,194,236,1 // cmpltps %xmm12,%xmm13 - .byte 68,15,40,21,130,46,0,0 // movaps 0x2e82(%rip),%xmm10 # 5f20 <_sk_callback_sse2+0x99e> + .byte 68,15,40,21,178,46,0,0 // movaps 0x2eb2(%rip),%xmm10 # 5f50 <_sk_callback_sse2+0x9a0> .byte 69,15,84,234 // andps %xmm10,%xmm13 .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,229 // subps %xmm13,%xmm12 .byte 69,15,40,233 // movaps %xmm9,%xmm13 .byte 69,15,92,236 // subps %xmm12,%xmm13 - .byte 68,15,88,13,118,46,0,0 // addps 0x2e76(%rip),%xmm9 # 5f30 <_sk_callback_sse2+0x9ae> - .byte 68,15,40,37,126,46,0,0 // movaps 0x2e7e(%rip),%xmm12 # 5f40 <_sk_callback_sse2+0x9be> + .byte 68,15,88,13,166,46,0,0 // addps 0x2ea6(%rip),%xmm9 # 5f60 <_sk_callback_sse2+0x9b0> + .byte 68,15,40,37,174,46,0,0 // movaps 0x2eae(%rip),%xmm12 # 5f70 <_sk_callback_sse2+0x9c0> .byte 69,15,89,229 // mulps %xmm13,%xmm12 .byte 69,15,92,204 // subps %xmm12,%xmm9 - .byte 68,15,40,37,126,46,0,0 // movaps 0x2e7e(%rip),%xmm12 # 5f50 <_sk_callback_sse2+0x9ce> + .byte 68,15,40,37,174,46,0,0 // movaps 0x2eae(%rip),%xmm12 # 5f80 <_sk_callback_sse2+0x9d0> .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 68,15,40,45,130,46,0,0 // movaps 0x2e82(%rip),%xmm13 # 5f60 <_sk_callback_sse2+0x9de> + .byte 68,15,40,45,178,46,0,0 // movaps 0x2eb2(%rip),%xmm13 # 5f90 <_sk_callback_sse2+0x9e0> .byte 69,15,94,236 // divps %xmm12,%xmm13 .byte 69,15,88,233 // addps %xmm9,%xmm13 - .byte 68,15,89,45,130,46,0,0 // mulps 0x2e82(%rip),%xmm13 # 5f70 <_sk_callback_sse2+0x9ee> + .byte 68,15,89,45,178,46,0,0 // mulps 0x2eb2(%rip),%xmm13 # 5fa0 <_sk_callback_sse2+0x9f0> .byte 102,69,15,91,205 // cvtps2dq %xmm13,%xmm9 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 @@ -32814,29 +32854,29 @@ HIDDEN _sk_lab_to_xyz_sse2 .globl _sk_lab_to_xyz_sse2 FUNCTION(_sk_lab_to_xyz_sse2) _sk_lab_to_xyz_sse2: - .byte 15,89,5,95,46,0,0 // mulps 0x2e5f(%rip),%xmm0 # 5f80 <_sk_callback_sse2+0x9fe> - .byte 68,15,40,5,103,46,0,0 // movaps 0x2e67(%rip),%xmm8 # 5f90 <_sk_callback_sse2+0xa0e> + .byte 15,89,5,143,46,0,0 // mulps 0x2e8f(%rip),%xmm0 # 5fb0 <_sk_callback_sse2+0xa00> + .byte 68,15,40,5,151,46,0,0 // movaps 0x2e97(%rip),%xmm8 # 5fc0 <_sk_callback_sse2+0xa10> .byte 65,15,89,200 // mulps %xmm8,%xmm1 - .byte 68,15,40,13,107,46,0,0 // movaps 0x2e6b(%rip),%xmm9 # 5fa0 <_sk_callback_sse2+0xa1e> + .byte 68,15,40,13,155,46,0,0 // movaps 0x2e9b(%rip),%xmm9 # 5fd0 <_sk_callback_sse2+0xa20> .byte 65,15,88,201 // addps %xmm9,%xmm1 .byte 65,15,89,208 // mulps %xmm8,%xmm2 .byte 65,15,88,209 // addps %xmm9,%xmm2 - .byte 15,88,5,104,46,0,0 // addps 0x2e68(%rip),%xmm0 # 5fb0 <_sk_callback_sse2+0xa2e> - .byte 15,89,5,113,46,0,0 // mulps 0x2e71(%rip),%xmm0 # 5fc0 <_sk_callback_sse2+0xa3e> - .byte 15,89,13,122,46,0,0 // mulps 0x2e7a(%rip),%xmm1 # 5fd0 <_sk_callback_sse2+0xa4e> + .byte 15,88,5,152,46,0,0 // addps 0x2e98(%rip),%xmm0 # 5fe0 <_sk_callback_sse2+0xa30> + .byte 15,89,5,161,46,0,0 // mulps 0x2ea1(%rip),%xmm0 # 5ff0 <_sk_callback_sse2+0xa40> + .byte 15,89,13,170,46,0,0 // mulps 0x2eaa(%rip),%xmm1 # 6000 <_sk_callback_sse2+0xa50> .byte 15,88,200 // addps %xmm0,%xmm1 - .byte 15,89,21,128,46,0,0 // mulps 0x2e80(%rip),%xmm2 # 5fe0 <_sk_callback_sse2+0xa5e> + .byte 15,89,21,176,46,0,0 // mulps 0x2eb0(%rip),%xmm2 # 6010 <_sk_callback_sse2+0xa60> .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 68,15,92,202 // subps %xmm2,%xmm9 .byte 68,15,40,225 // movaps %xmm1,%xmm12 .byte 69,15,89,228 // mulps %xmm12,%xmm12 .byte 68,15,89,225 // mulps %xmm1,%xmm12 - .byte 15,40,21,117,46,0,0 // movaps 0x2e75(%rip),%xmm2 # 5ff0 <_sk_callback_sse2+0xa6e> + .byte 15,40,21,165,46,0,0 // movaps 0x2ea5(%rip),%xmm2 # 6020 <_sk_callback_sse2+0xa70> .byte 68,15,40,194 // movaps %xmm2,%xmm8 .byte 69,15,194,196,1 // cmpltps %xmm12,%xmm8 - .byte 68,15,40,21,116,46,0,0 // movaps 0x2e74(%rip),%xmm10 # 6000 <_sk_callback_sse2+0xa7e> + .byte 68,15,40,21,164,46,0,0 // movaps 0x2ea4(%rip),%xmm10 # 6030 <_sk_callback_sse2+0xa80> .byte 65,15,88,202 // addps %xmm10,%xmm1 - .byte 68,15,40,29,120,46,0,0 // movaps 0x2e78(%rip),%xmm11 # 6010 <_sk_callback_sse2+0xa8e> + .byte 68,15,40,29,168,46,0,0 // movaps 0x2ea8(%rip),%xmm11 # 6040 <_sk_callback_sse2+0xa90> .byte 65,15,89,203 // mulps %xmm11,%xmm1 .byte 69,15,84,224 // andps %xmm8,%xmm12 .byte 68,15,85,193 // andnps %xmm1,%xmm8 @@ -32860,8 +32900,8 @@ _sk_lab_to_xyz_sse2: .byte 15,84,194 // andps %xmm2,%xmm0 .byte 65,15,85,209 // andnps %xmm9,%xmm2 .byte 15,86,208 // orps %xmm0,%xmm2 - .byte 68,15,89,5,40,46,0,0 // mulps 0x2e28(%rip),%xmm8 # 6020 <_sk_callback_sse2+0xa9e> - .byte 15,89,21,49,46,0,0 // mulps 0x2e31(%rip),%xmm2 # 6030 <_sk_callback_sse2+0xaae> + .byte 68,15,89,5,88,46,0,0 // mulps 0x2e58(%rip),%xmm8 # 6050 <_sk_callback_sse2+0xaa0> + .byte 15,89,21,97,46,0,0 // mulps 0x2e61(%rip),%xmm2 # 6060 <_sk_callback_sse2+0xab0> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -32877,9 +32917,9 @@ _sk_load_a8_sse2: .byte 102,65,15,110,4,18 // movd (%r10,%rdx,1),%xmm0 .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 .byte 102,15,97,192 // punpcklwd %xmm0,%xmm0 - .byte 102,15,219,5,25,46,0,0 // pand 0x2e19(%rip),%xmm0 # 6040 <_sk_callback_sse2+0xabe> + .byte 102,15,219,5,73,46,0,0 // pand 0x2e49(%rip),%xmm0 # 6070 <_sk_callback_sse2+0xac0> .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,31,46,0,0 // mulps 0x2e1f(%rip),%xmm3 # 6050 <_sk_callback_sse2+0xace> + .byte 15,89,29,79,46,0,0 // mulps 0x2e4f(%rip),%xmm3 # 6080 <_sk_callback_sse2+0xad0> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 15,87,201 // xorps %xmm1,%xmm1 @@ -32947,7 +32987,7 @@ _sk_gather_a8_sse2: .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 .byte 15,91,216 // cvtdq2ps %xmm0,%xmm3 - .byte 15,89,29,63,45,0,0 // mulps 0x2d3f(%rip),%xmm3 # 6060 <_sk_callback_sse2+0xade> + .byte 15,89,29,111,45,0,0 // mulps 0x2d6f(%rip),%xmm3 # 6090 <_sk_callback_sse2+0xae0> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 102,15,239,201 // pxor %xmm1,%xmm1 @@ -32962,7 +33002,7 @@ FUNCTION(_sk_store_a8_sse2) _sk_store_a8_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 68,15,40,5,49,45,0,0 // movaps 0x2d31(%rip),%xmm8 # 6070 <_sk_callback_sse2+0xaee> + .byte 68,15,40,5,97,45,0,0 // movaps 0x2d61(%rip),%xmm8 # 60a0 <_sk_callback_sse2+0xaf0> .byte 68,15,89,195 // mulps %xmm3,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 @@ -33007,11 +33047,11 @@ _sk_load_g8_sse2: .byte 102,65,15,110,4,18 // movd (%r10,%rdx,1),%xmm0 .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 .byte 102,15,97,192 // punpcklwd %xmm0,%xmm0 - .byte 102,15,219,5,156,44,0,0 // pand 0x2c9c(%rip),%xmm0 # 6080 <_sk_callback_sse2+0xafe> + .byte 102,15,219,5,204,44,0,0 // pand 0x2ccc(%rip),%xmm0 # 60b0 <_sk_callback_sse2+0xb00> .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,162,44,0,0 // mulps 0x2ca2(%rip),%xmm0 # 6090 <_sk_callback_sse2+0xb0e> + .byte 15,89,5,210,44,0,0 // mulps 0x2cd2(%rip),%xmm0 # 60c0 <_sk_callback_sse2+0xb10> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,169,44,0,0 // movaps 0x2ca9(%rip),%xmm3 # 60a0 <_sk_callback_sse2+0xb1e> + .byte 15,40,29,217,44,0,0 // movaps 0x2cd9(%rip),%xmm3 # 60d0 <_sk_callback_sse2+0xb20> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 255,224 // jmpq *%rax @@ -33077,9 +33117,9 @@ _sk_gather_g8_sse2: .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,206,43,0,0 // mulps 0x2bce(%rip),%xmm0 # 60b0 <_sk_callback_sse2+0xb2e> + .byte 15,89,5,254,43,0,0 // mulps 0x2bfe(%rip),%xmm0 # 60e0 <_sk_callback_sse2+0xb30> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,213,43,0,0 // movaps 0x2bd5(%rip),%xmm3 # 60c0 <_sk_callback_sse2+0xb3e> + .byte 15,40,29,5,44,0,0 // movaps 0x2c05(%rip),%xmm3 # 60f0 <_sk_callback_sse2+0xb40> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 .byte 91 // pop %rbx @@ -33147,11 +33187,11 @@ _sk_gather_i8_sse2: .byte 102,66,15,110,76,149,0 // movd 0x0(%rbp,%r10,4),%xmm1 .byte 102,68,15,98,201 // punpckldq %xmm1,%xmm9 .byte 102,68,15,98,200 // punpckldq %xmm0,%xmm9 - .byte 102,15,111,21,238,42,0,0 // movdqa 0x2aee(%rip),%xmm2 # 60d0 <_sk_callback_sse2+0xb4e> + .byte 102,15,111,21,30,43,0,0 // movdqa 0x2b1e(%rip),%xmm2 # 6100 <_sk_callback_sse2+0xb50> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,234,42,0,0 // movaps 0x2aea(%rip),%xmm8 # 60e0 <_sk_callback_sse2+0xb5e> + .byte 68,15,40,5,26,43,0,0 // movaps 0x2b1a(%rip),%xmm8 # 6110 <_sk_callback_sse2+0xb60> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,114,209,8 // psrld $0x8,%xmm1 @@ -33182,19 +33222,19 @@ _sk_load_565_sse2: .byte 117,83 // jne 3697 <_sk_load_565_sse2+0x5d> .byte 243,65,15,126,20,82 // movq (%r10,%rdx,2),%xmm2 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 - .byte 102,15,111,5,154,42,0,0 // movdqa 0x2a9a(%rip),%xmm0 # 60f0 <_sk_callback_sse2+0xb6e> + .byte 102,15,111,5,202,42,0,0 // movdqa 0x2aca(%rip),%xmm0 # 6120 <_sk_callback_sse2+0xb70> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,156,42,0,0 // mulps 0x2a9c(%rip),%xmm0 # 6100 <_sk_callback_sse2+0xb7e> - .byte 102,15,111,13,164,42,0,0 // movdqa 0x2aa4(%rip),%xmm1 # 6110 <_sk_callback_sse2+0xb8e> + .byte 15,89,5,204,42,0,0 // mulps 0x2acc(%rip),%xmm0 # 6130 <_sk_callback_sse2+0xb80> + .byte 102,15,111,13,212,42,0,0 // movdqa 0x2ad4(%rip),%xmm1 # 6140 <_sk_callback_sse2+0xb90> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,166,42,0,0 // mulps 0x2aa6(%rip),%xmm1 # 6120 <_sk_callback_sse2+0xb9e> - .byte 102,15,219,21,174,42,0,0 // pand 0x2aae(%rip),%xmm2 # 6130 <_sk_callback_sse2+0xbae> + .byte 15,89,13,214,42,0,0 // mulps 0x2ad6(%rip),%xmm1 # 6150 <_sk_callback_sse2+0xba0> + .byte 102,15,219,21,222,42,0,0 // pand 0x2ade(%rip),%xmm2 # 6160 <_sk_callback_sse2+0xbb0> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,180,42,0,0 // mulps 0x2ab4(%rip),%xmm2 # 6140 <_sk_callback_sse2+0xbbe> + .byte 15,89,21,228,42,0,0 // mulps 0x2ae4(%rip),%xmm2 # 6170 <_sk_callback_sse2+0xbc0> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,187,42,0,0 // movaps 0x2abb(%rip),%xmm3 # 6150 <_sk_callback_sse2+0xbce> + .byte 15,40,29,235,42,0,0 // movaps 0x2aeb(%rip),%xmm3 # 6180 <_sk_callback_sse2+0xbd0> .byte 255,224 // jmpq *%rax .byte 69,137,193 // mov %r8d,%r9d .byte 65,128,225,3 // and $0x3,%r9b @@ -33251,19 +33291,19 @@ _sk_gather_565_sse2: .byte 102,15,196,208,3 // pinsrw $0x3,%eax,%xmm2 .byte 102,15,239,192 // pxor %xmm0,%xmm0 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 - .byte 102,15,111,5,240,41,0,0 // movdqa 0x29f0(%rip),%xmm0 # 6160 <_sk_callback_sse2+0xbde> + .byte 102,15,111,5,32,42,0,0 // movdqa 0x2a20(%rip),%xmm0 # 6190 <_sk_callback_sse2+0xbe0> .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,242,41,0,0 // mulps 0x29f2(%rip),%xmm0 # 6170 <_sk_callback_sse2+0xbee> - .byte 102,15,111,13,250,41,0,0 // movdqa 0x29fa(%rip),%xmm1 # 6180 <_sk_callback_sse2+0xbfe> + .byte 15,89,5,34,42,0,0 // mulps 0x2a22(%rip),%xmm0 # 61a0 <_sk_callback_sse2+0xbf0> + .byte 102,15,111,13,42,42,0,0 // movdqa 0x2a2a(%rip),%xmm1 # 61b0 <_sk_callback_sse2+0xc00> .byte 102,15,219,202 // pand %xmm2,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,252,41,0,0 // mulps 0x29fc(%rip),%xmm1 # 6190 <_sk_callback_sse2+0xc0e> - .byte 102,15,219,21,4,42,0,0 // pand 0x2a04(%rip),%xmm2 # 61a0 <_sk_callback_sse2+0xc1e> + .byte 15,89,13,44,42,0,0 // mulps 0x2a2c(%rip),%xmm1 # 61c0 <_sk_callback_sse2+0xc10> + .byte 102,15,219,21,52,42,0,0 // pand 0x2a34(%rip),%xmm2 # 61d0 <_sk_callback_sse2+0xc20> .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,10,42,0,0 // mulps 0x2a0a(%rip),%xmm2 # 61b0 <_sk_callback_sse2+0xc2e> + .byte 15,89,21,58,42,0,0 // mulps 0x2a3a(%rip),%xmm2 # 61e0 <_sk_callback_sse2+0xc30> .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,17,42,0,0 // movaps 0x2a11(%rip),%xmm3 # 61c0 <_sk_callback_sse2+0xc3e> + .byte 15,40,29,65,42,0,0 // movaps 0x2a41(%rip),%xmm3 # 61f0 <_sk_callback_sse2+0xc40> .byte 91 // pop %rbx .byte 255,224 // jmpq *%rax @@ -33273,12 +33313,12 @@ FUNCTION(_sk_store_565_sse2) _sk_store_565_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 68,15,40,5,17,42,0,0 // movaps 0x2a11(%rip),%xmm8 # 61d0 <_sk_callback_sse2+0xc4e> + .byte 68,15,40,5,65,42,0,0 // movaps 0x2a41(%rip),%xmm8 # 6200 <_sk_callback_sse2+0xc50> .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 68,15,40,21,6,42,0,0 // movaps 0x2a06(%rip),%xmm10 # 61e0 <_sk_callback_sse2+0xc5e> + .byte 68,15,40,21,54,42,0,0 // movaps 0x2a36(%rip),%xmm10 # 6210 <_sk_callback_sse2+0xc60> .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 @@ -33321,21 +33361,21 @@ _sk_load_4444_sse2: .byte 117,98 // jne 38ca <_sk_load_4444_sse2+0x6c> .byte 243,65,15,126,28,82 // movq (%r10,%rdx,2),%xmm3 .byte 102,15,97,216 // punpcklwd %xmm0,%xmm3 - .byte 102,15,111,5,118,41,0,0 // movdqa 0x2976(%rip),%xmm0 # 61f0 <_sk_callback_sse2+0xc6e> + .byte 102,15,111,5,166,41,0,0 // movdqa 0x29a6(%rip),%xmm0 # 6220 <_sk_callback_sse2+0xc70> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,120,41,0,0 // mulps 0x2978(%rip),%xmm0 # 6200 <_sk_callback_sse2+0xc7e> - .byte 102,15,111,13,128,41,0,0 // movdqa 0x2980(%rip),%xmm1 # 6210 <_sk_callback_sse2+0xc8e> + .byte 15,89,5,168,41,0,0 // mulps 0x29a8(%rip),%xmm0 # 6230 <_sk_callback_sse2+0xc80> + .byte 102,15,111,13,176,41,0,0 // movdqa 0x29b0(%rip),%xmm1 # 6240 <_sk_callback_sse2+0xc90> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,130,41,0,0 // mulps 0x2982(%rip),%xmm1 # 6220 <_sk_callback_sse2+0xc9e> - .byte 102,15,111,21,138,41,0,0 // movdqa 0x298a(%rip),%xmm2 # 6230 <_sk_callback_sse2+0xcae> + .byte 15,89,13,178,41,0,0 // mulps 0x29b2(%rip),%xmm1 # 6250 <_sk_callback_sse2+0xca0> + .byte 102,15,111,21,186,41,0,0 // movdqa 0x29ba(%rip),%xmm2 # 6260 <_sk_callback_sse2+0xcb0> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,140,41,0,0 // mulps 0x298c(%rip),%xmm2 # 6240 <_sk_callback_sse2+0xcbe> - .byte 102,15,219,29,148,41,0,0 // pand 0x2994(%rip),%xmm3 # 6250 <_sk_callback_sse2+0xcce> + .byte 15,89,21,188,41,0,0 // mulps 0x29bc(%rip),%xmm2 # 6270 <_sk_callback_sse2+0xcc0> + .byte 102,15,219,29,196,41,0,0 // pand 0x29c4(%rip),%xmm3 # 6280 <_sk_callback_sse2+0xcd0> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,154,41,0,0 // mulps 0x299a(%rip),%xmm3 # 6260 <_sk_callback_sse2+0xcde> + .byte 15,89,29,202,41,0,0 // mulps 0x29ca(%rip),%xmm3 # 6290 <_sk_callback_sse2+0xce0> .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax .byte 69,137,193 // mov %r8d,%r9d @@ -33393,21 +33433,21 @@ _sk_gather_4444_sse2: .byte 102,15,196,216,3 // pinsrw $0x3,%eax,%xmm3 .byte 102,15,239,192 // pxor %xmm0,%xmm0 .byte 102,15,97,216 // punpcklwd %xmm0,%xmm3 - .byte 102,15,111,5,205,40,0,0 // movdqa 0x28cd(%rip),%xmm0 # 6270 <_sk_callback_sse2+0xcee> + .byte 102,15,111,5,253,40,0,0 // movdqa 0x28fd(%rip),%xmm0 # 62a0 <_sk_callback_sse2+0xcf0> .byte 102,15,219,195 // pand %xmm3,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 15,89,5,207,40,0,0 // mulps 0x28cf(%rip),%xmm0 # 6280 <_sk_callback_sse2+0xcfe> - .byte 102,15,111,13,215,40,0,0 // movdqa 0x28d7(%rip),%xmm1 # 6290 <_sk_callback_sse2+0xd0e> + .byte 15,89,5,255,40,0,0 // mulps 0x28ff(%rip),%xmm0 # 62b0 <_sk_callback_sse2+0xd00> + .byte 102,15,111,13,7,41,0,0 // movdqa 0x2907(%rip),%xmm1 # 62c0 <_sk_callback_sse2+0xd10> .byte 102,15,219,203 // pand %xmm3,%xmm1 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 - .byte 15,89,13,217,40,0,0 // mulps 0x28d9(%rip),%xmm1 # 62a0 <_sk_callback_sse2+0xd1e> - .byte 102,15,111,21,225,40,0,0 // movdqa 0x28e1(%rip),%xmm2 # 62b0 <_sk_callback_sse2+0xd2e> + .byte 15,89,13,9,41,0,0 // mulps 0x2909(%rip),%xmm1 # 62d0 <_sk_callback_sse2+0xd20> + .byte 102,15,111,21,17,41,0,0 // movdqa 0x2911(%rip),%xmm2 # 62e0 <_sk_callback_sse2+0xd30> .byte 102,15,219,211 // pand %xmm3,%xmm2 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 - .byte 15,89,21,227,40,0,0 // mulps 0x28e3(%rip),%xmm2 # 62c0 <_sk_callback_sse2+0xd3e> - .byte 102,15,219,29,235,40,0,0 // pand 0x28eb(%rip),%xmm3 # 62d0 <_sk_callback_sse2+0xd4e> + .byte 15,89,21,19,41,0,0 // mulps 0x2913(%rip),%xmm2 # 62f0 <_sk_callback_sse2+0xd40> + .byte 102,15,219,29,27,41,0,0 // pand 0x291b(%rip),%xmm3 # 6300 <_sk_callback_sse2+0xd50> .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 - .byte 15,89,29,241,40,0,0 // mulps 0x28f1(%rip),%xmm3 # 62e0 <_sk_callback_sse2+0xd5e> + .byte 15,89,29,33,41,0,0 // mulps 0x2921(%rip),%xmm3 # 6310 <_sk_callback_sse2+0xd60> .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx .byte 255,224 // jmpq *%rax @@ -33418,7 +33458,7 @@ FUNCTION(_sk_store_4444_sse2) _sk_store_4444_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 68,15,40,5,239,40,0,0 // movaps 0x28ef(%rip),%xmm8 # 62f0 <_sk_callback_sse2+0xd6e> + .byte 68,15,40,5,31,41,0,0 // movaps 0x291f(%rip),%xmm8 # 6320 <_sk_callback_sse2+0xd70> .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 @@ -33470,11 +33510,11 @@ _sk_load_8888_sse2: .byte 77,133,192 // test %r8,%r8 .byte 117,98 // jne 3b20 <_sk_load_8888_sse2+0x6c> .byte 243,68,15,111,12,144 // movdqu (%rax,%rdx,4),%xmm9 - .byte 102,15,111,21,52,40,0,0 // movdqa 0x2834(%rip),%xmm2 # 6300 <_sk_callback_sse2+0xd7e> + .byte 102,15,111,21,100,40,0,0 // movdqa 0x2864(%rip),%xmm2 # 6330 <_sk_callback_sse2+0xd80> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,48,40,0,0 // movaps 0x2830(%rip),%xmm8 # 6310 <_sk_callback_sse2+0xd8e> + .byte 68,15,40,5,96,40,0,0 // movaps 0x2860(%rip),%xmm8 # 6340 <_sk_callback_sse2+0xd90> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,114,209,8 // psrld $0x8,%xmm1 @@ -33542,11 +33582,11 @@ _sk_gather_8888_sse2: .byte 102,67,15,110,12,145 // movd (%r9,%r10,4),%xmm1 .byte 102,68,15,98,201 // punpckldq %xmm1,%xmm9 .byte 102,68,15,98,200 // punpckldq %xmm0,%xmm9 - .byte 102,15,111,21,49,39,0,0 // movdqa 0x2731(%rip),%xmm2 # 6320 <_sk_callback_sse2+0xd9e> + .byte 102,15,111,21,97,39,0,0 // movdqa 0x2761(%rip),%xmm2 # 6350 <_sk_callback_sse2+0xda0> .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 .byte 102,15,219,194 // pand %xmm2,%xmm0 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 - .byte 68,15,40,5,45,39,0,0 // movaps 0x272d(%rip),%xmm8 # 6330 <_sk_callback_sse2+0xdae> + .byte 68,15,40,5,93,39,0,0 // movaps 0x275d(%rip),%xmm8 # 6360 <_sk_callback_sse2+0xdb0> .byte 65,15,89,192 // mulps %xmm8,%xmm0 .byte 102,65,15,111,201 // movdqa %xmm9,%xmm1 .byte 102,15,114,209,8 // psrld $0x8,%xmm1 @@ -33571,7 +33611,7 @@ FUNCTION(_sk_store_8888_sse2) _sk_store_8888_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 68,15,40,5,239,38,0,0 // movaps 0x26ef(%rip),%xmm8 # 6340 <_sk_callback_sse2+0xdbe> + .byte 68,15,40,5,31,39,0,0 // movaps 0x271f(%rip),%xmm8 # 6370 <_sk_callback_sse2+0xdc0> .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 @@ -33628,7 +33668,7 @@ _sk_load_f16_sse2: .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1 - .byte 102,68,15,111,13,22,38,0,0 // movdqa 0x2616(%rip),%xmm9 # 6350 <_sk_callback_sse2+0xdce> + .byte 102,68,15,111,13,70,38,0,0 // movdqa 0x2646(%rip),%xmm9 # 6380 <_sk_callback_sse2+0xdd0> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,193 // pand %xmm9,%xmm0 .byte 102,15,239,200 // pxor %xmm0,%xmm1 @@ -33636,11 +33676,11 @@ _sk_load_f16_sse2: .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 .byte 102,68,15,235,232 // por %xmm0,%xmm13 - .byte 102,68,15,111,29,251,37,0,0 // movdqa 0x25fb(%rip),%xmm11 # 6360 <_sk_callback_sse2+0xdde> + .byte 102,68,15,111,29,43,38,0,0 // movdqa 0x262b(%rip),%xmm11 # 6390 <_sk_callback_sse2+0xde0> .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 - .byte 102,68,15,111,37,253,37,0,0 // movdqa 0x25fd(%rip),%xmm12 # 6370 <_sk_callback_sse2+0xdee> + .byte 102,68,15,111,37,45,38,0,0 // movdqa 0x262d(%rip),%xmm12 # 63a0 <_sk_callback_sse2+0xdf0> .byte 102,65,15,239,204 // pxor %xmm12,%xmm1 - .byte 102,15,111,29,0,38,0,0 // movdqa 0x2600(%rip),%xmm3 # 6380 <_sk_callback_sse2+0xdfe> + .byte 102,15,111,29,48,38,0,0 // movdqa 0x2630(%rip),%xmm3 # 63b0 <_sk_callback_sse2+0xe00> .byte 102,15,111,195 // movdqa %xmm3,%xmm0 .byte 102,15,102,193 // pcmpgtd %xmm1,%xmm0 .byte 102,65,15,223,197 // pandn %xmm13,%xmm0 @@ -33740,7 +33780,7 @@ _sk_gather_f16_sse2: .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1 .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1 - .byte 102,68,15,111,13,82,36,0,0 // movdqa 0x2452(%rip),%xmm9 # 6390 <_sk_callback_sse2+0xe0e> + .byte 102,68,15,111,13,130,36,0,0 // movdqa 0x2482(%rip),%xmm9 # 63c0 <_sk_callback_sse2+0xe10> .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,65,15,219,193 // pand %xmm9,%xmm0 .byte 102,15,239,200 // pxor %xmm0,%xmm1 @@ -33748,11 +33788,11 @@ _sk_gather_f16_sse2: .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 .byte 102,68,15,235,232 // por %xmm0,%xmm13 - .byte 102,68,15,111,29,55,36,0,0 // movdqa 0x2437(%rip),%xmm11 # 63a0 <_sk_callback_sse2+0xe1e> + .byte 102,68,15,111,29,103,36,0,0 // movdqa 0x2467(%rip),%xmm11 # 63d0 <_sk_callback_sse2+0xe20> .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 - .byte 102,68,15,111,37,57,36,0,0 // movdqa 0x2439(%rip),%xmm12 # 63b0 <_sk_callback_sse2+0xe2e> + .byte 102,68,15,111,37,105,36,0,0 // movdqa 0x2469(%rip),%xmm12 # 63e0 <_sk_callback_sse2+0xe30> .byte 102,65,15,239,204 // pxor %xmm12,%xmm1 - .byte 102,15,111,29,60,36,0,0 // movdqa 0x243c(%rip),%xmm3 # 63c0 <_sk_callback_sse2+0xe3e> + .byte 102,15,111,29,108,36,0,0 // movdqa 0x246c(%rip),%xmm3 # 63f0 <_sk_callback_sse2+0xe40> .byte 102,15,111,195 // movdqa %xmm3,%xmm0 .byte 102,15,102,193 // pcmpgtd %xmm1,%xmm0 .byte 102,65,15,223,197 // pandn %xmm13,%xmm0 @@ -33806,17 +33846,17 @@ FUNCTION(_sk_store_f16_sse2) _sk_store_f16_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax - .byte 102,68,15,111,21,99,35,0,0 // movdqa 0x2363(%rip),%xmm10 # 63d0 <_sk_callback_sse2+0xe4e> + .byte 102,68,15,111,21,147,35,0,0 // movdqa 0x2393(%rip),%xmm10 # 6400 <_sk_callback_sse2+0xe50> .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12 .byte 102,69,15,219,226 // pand %xmm10,%xmm12 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13 .byte 102,69,15,239,236 // pxor %xmm12,%xmm13 - .byte 102,68,15,111,13,86,35,0,0 // movdqa 0x2356(%rip),%xmm9 # 63e0 <_sk_callback_sse2+0xe5e> + .byte 102,68,15,111,13,134,35,0,0 // movdqa 0x2386(%rip),%xmm9 # 6410 <_sk_callback_sse2+0xe60> .byte 102,65,15,114,212,16 // psrld $0x10,%xmm12 .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8 .byte 102,69,15,102,197 // pcmpgtd %xmm13,%xmm8 .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 - .byte 102,68,15,111,29,71,35,0,0 // movdqa 0x2347(%rip),%xmm11 # 63f0 <_sk_callback_sse2+0xe6e> + .byte 102,68,15,111,29,119,35,0,0 // movdqa 0x2377(%rip),%xmm11 # 6420 <_sk_callback_sse2+0xe70> .byte 102,69,15,235,227 // por %xmm11,%xmm12 .byte 102,69,15,254,229 // paddd %xmm13,%xmm12 .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12 @@ -33909,7 +33949,7 @@ _sk_load_u16_be_sse2: .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 .byte 15,91,193 // cvtdq2ps %xmm1,%xmm0 - .byte 68,15,40,5,167,33,0,0 // movaps 0x21a7(%rip),%xmm8 # 6400 <_sk_callback_sse2+0xe7e> + .byte 68,15,40,5,215,33,0,0 // movaps 0x21d7(%rip),%xmm8 # 6430 <_sk_callback_sse2+0xe80> .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 @@ -33978,7 +34018,7 @@ _sk_load_rgb_u16_be_sse2: .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 102,65,15,97,200 // punpcklwd %xmm8,%xmm1 .byte 15,91,193 // cvtdq2ps %xmm1,%xmm0 - .byte 68,15,40,13,150,32,0,0 // movaps 0x2096(%rip),%xmm9 # 6410 <_sk_callback_sse2+0xe8e> + .byte 68,15,40,13,198,32,0,0 // movaps 0x20c6(%rip),%xmm9 # 6440 <_sk_callback_sse2+0xe90> .byte 65,15,89,193 // mulps %xmm9,%xmm0 .byte 102,15,111,203 // movdqa %xmm3,%xmm1 .byte 102,15,113,241,8 // psllw $0x8,%xmm1 @@ -33995,7 +34035,7 @@ _sk_load_rgb_u16_be_sse2: .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 .byte 65,15,89,209 // mulps %xmm9,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,29,93,32,0,0 // movaps 0x205d(%rip),%xmm3 # 6420 <_sk_callback_sse2+0xe9e> + .byte 15,40,29,141,32,0,0 // movaps 0x208d(%rip),%xmm3 # 6450 <_sk_callback_sse2+0xea0> .byte 255,224 // jmpq *%rax .byte 102,65,15,110,20,65 // movd (%r9,%rax,2),%xmm2 .byte 102,65,15,196,84,65,4,2 // pinsrw $0x2,0x4(%r9,%rax,2),%xmm2 @@ -34024,7 +34064,7 @@ _sk_store_u16_be_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,8 // mov (%rax),%r9 .byte 72,141,4,149,0,0,0,0 // lea 0x0(,%rdx,4),%rax - .byte 68,15,40,21,247,31,0,0 // movaps 0x1ff7(%rip),%xmm10 # 6430 <_sk_callback_sse2+0xeae> + .byte 68,15,40,21,39,32,0,0 // movaps 0x2027(%rip),%xmm10 # 6460 <_sk_callback_sse2+0xeb0> .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 69,15,89,194 // mulps %xmm10,%xmm8 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 @@ -34205,14 +34245,15 @@ FUNCTION(_sk_repeat_x_sse2) _sk_repeat_x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,200 // movaps %xmm0,%xmm9 - .byte 69,15,94,200 // divps %xmm8,%xmm9 + .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 68,15,89,200 // mulps %xmm0,%xmm9 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 - .byte 68,15,84,13,82,29,0,0 // andps 0x1d52(%rip),%xmm9 # 6440 <_sk_callback_sse2+0xebe> + .byte 68,15,84,13,128,29,0,0 // andps 0x1d80(%rip),%xmm9 # 6470 <_sk_callback_sse2+0xec0> .byte 69,15,92,209 // subps %xmm9,%xmm10 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,92,194 // subps %xmm10,%xmm0 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 @@ -34227,14 +34268,15 @@ FUNCTION(_sk_repeat_y_sse2) _sk_repeat_y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,201 // movaps %xmm1,%xmm9 - .byte 69,15,94,200 // divps %xmm8,%xmm9 + .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 68,15,89,201 // mulps %xmm1,%xmm9 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 - .byte 68,15,84,13,26,29,0,0 // andps 0x1d1a(%rip),%xmm9 # 6450 <_sk_callback_sse2+0xece> + .byte 68,15,84,13,65,29,0,0 // andps 0x1d41(%rip),%xmm9 # 6480 <_sk_callback_sse2+0xed0> .byte 69,15,92,209 // subps %xmm9,%xmm10 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 69,15,89,208 // mulps %xmm8,%xmm10 .byte 65,15,92,202 // subps %xmm10,%xmm1 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 @@ -34248,28 +34290,30 @@ HIDDEN _sk_mirror_x_sse2 FUNCTION(_sk_mirror_x_sse2) _sk_mirror_x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,40,200 // movaps %xmm8,%xmm9 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 65,15,92,193 // subps %xmm9,%xmm0 - .byte 243,69,15,88,192 // addss %xmm8,%xmm8 + .byte 243,68,15,16,8 // movss (%rax),%xmm9 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,208 // movaps %xmm0,%xmm10 - .byte 69,15,94,208 // divps %xmm8,%xmm10 + .byte 65,15,92,192 // subps %xmm8,%xmm0 + .byte 243,69,15,88,201 // addss %xmm9,%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 243,68,15,89,21,17,33,0,0 // mulss 0x2111(%rip),%xmm10 # 68a0 <_sk_callback_sse2+0x12f0> + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,208 // mulps %xmm0,%xmm10 .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 - .byte 68,15,84,21,208,28,0,0 // andps 0x1cd0(%rip),%xmm10 # 6460 <_sk_callback_sse2+0xede> + .byte 68,15,84,21,226,28,0,0 // andps 0x1ce2(%rip),%xmm10 # 6490 <_sk_callback_sse2+0xee0> .byte 69,15,87,228 // xorps %xmm12,%xmm12 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 69,15,89,216 // mulps %xmm8,%xmm11 + .byte 69,15,89,217 // mulps %xmm9,%xmm11 .byte 65,15,92,195 // subps %xmm11,%xmm0 - .byte 65,15,92,193 // subps %xmm9,%xmm0 + .byte 65,15,92,192 // subps %xmm8,%xmm0 .byte 68,15,92,224 // subps %xmm0,%xmm12 .byte 65,15,84,196 // andps %xmm12,%xmm0 - .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 - .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 - .byte 65,15,93,192 // minps %xmm8,%xmm0 + .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 + .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 + .byte 65,15,93,193 // minps %xmm9,%xmm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -34278,28 +34322,30 @@ HIDDEN _sk_mirror_y_sse2 FUNCTION(_sk_mirror_y_sse2) _sk_mirror_y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 69,15,40,200 // movaps %xmm8,%xmm9 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 65,15,92,201 // subps %xmm9,%xmm1 - .byte 243,69,15,88,192 // addss %xmm8,%xmm8 + .byte 243,68,15,16,8 // movss (%rax),%xmm9 + .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 + .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 68,15,40,209 // movaps %xmm1,%xmm10 - .byte 69,15,94,208 // divps %xmm8,%xmm10 + .byte 65,15,92,200 // subps %xmm8,%xmm1 + .byte 243,69,15,88,201 // addss %xmm9,%xmm9 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 + .byte 243,68,15,89,21,155,32,0,0 // mulss 0x209b(%rip),%xmm10 # 68a4 <_sk_callback_sse2+0x12f4> + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 68,15,89,209 // mulps %xmm1,%xmm10 .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 - .byte 68,15,84,21,118,28,0,0 // andps 0x1c76(%rip),%xmm10 # 6470 <_sk_callback_sse2+0xeee> + .byte 68,15,84,21,120,28,0,0 // andps 0x1c78(%rip),%xmm10 # 64a0 <_sk_callback_sse2+0xef0> .byte 69,15,87,228 // xorps %xmm12,%xmm12 .byte 69,15,92,218 // subps %xmm10,%xmm11 - .byte 69,15,89,216 // mulps %xmm8,%xmm11 + .byte 69,15,89,217 // mulps %xmm9,%xmm11 .byte 65,15,92,203 // subps %xmm11,%xmm1 - .byte 65,15,92,201 // subps %xmm9,%xmm1 + .byte 65,15,92,200 // subps %xmm8,%xmm1 .byte 68,15,92,225 // subps %xmm1,%xmm12 .byte 65,15,84,204 // andps %xmm12,%xmm1 - .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 - .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 - .byte 65,15,93,200 // minps %xmm8,%xmm1 + .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 + .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 + .byte 65,15,93,201 // minps %xmm9,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -34309,7 +34355,7 @@ FUNCTION(_sk_clamp_x_1_sse2) _sk_clamp_x_1_sse2: .byte 69,15,87,192 // xorps %xmm8,%xmm8 .byte 68,15,95,192 // maxps %xmm0,%xmm8 - .byte 68,15,93,5,72,28,0,0 // minps 0x1c48(%rip),%xmm8 # 6480 <_sk_callback_sse2+0xefe> + .byte 68,15,93,5,74,28,0,0 // minps 0x1c4a(%rip),%xmm8 # 64b0 <_sk_callback_sse2+0xf00> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,15,40,192 // movaps %xmm8,%xmm0 .byte 255,224 // jmpq *%rax @@ -34322,7 +34368,7 @@ _sk_repeat_x_1_sse2: .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,194,200,1 // cmpltps %xmm8,%xmm9 - .byte 68,15,84,13,54,28,0,0 // andps 0x1c36(%rip),%xmm9 # 6490 <_sk_callback_sse2+0xf0e> + .byte 68,15,84,13,56,28,0,0 // andps 0x1c38(%rip),%xmm9 # 64c0 <_sk_callback_sse2+0xf10> .byte 69,15,92,193 // subps %xmm9,%xmm8 .byte 65,15,92,192 // subps %xmm8,%xmm0 .byte 72,173 // lods %ds:(%rsi),%rax @@ -34332,14 +34378,14 @@ HIDDEN _sk_mirror_x_1_sse2 .globl _sk_mirror_x_1_sse2 FUNCTION(_sk_mirror_x_1_sse2) _sk_mirror_x_1_sse2: - .byte 68,15,40,5,50,28,0,0 // movaps 0x1c32(%rip),%xmm8 # 64a0 <_sk_callback_sse2+0xf1e> + .byte 68,15,40,5,52,28,0,0 // movaps 0x1c34(%rip),%xmm8 # 64d0 <_sk_callback_sse2+0xf20> .byte 65,15,88,192 // addps %xmm8,%xmm0 - .byte 68,15,40,13,54,28,0,0 // movaps 0x1c36(%rip),%xmm9 # 64b0 <_sk_callback_sse2+0xf2e> + .byte 68,15,40,13,56,28,0,0 // movaps 0x1c38(%rip),%xmm9 # 64e0 <_sk_callback_sse2+0xf30> .byte 68,15,89,200 // mulps %xmm0,%xmm9 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 - .byte 68,15,84,13,44,28,0,0 // andps 0x1c2c(%rip),%xmm9 # 64c0 <_sk_callback_sse2+0xf3e> + .byte 68,15,84,13,46,28,0,0 // andps 0x1c2e(%rip),%xmm9 # 64f0 <_sk_callback_sse2+0xf40> .byte 69,15,87,219 // xorps %xmm11,%xmm11 .byte 69,15,92,209 // subps %xmm9,%xmm10 .byte 69,15,88,210 // addps %xmm10,%xmm10 @@ -34355,10 +34401,10 @@ HIDDEN _sk_luminance_to_alpha_sse2 FUNCTION(_sk_luminance_to_alpha_sse2) _sk_luminance_to_alpha_sse2: .byte 15,40,218 // movaps %xmm2,%xmm3 - .byte 15,89,5,18,28,0,0 // mulps 0x1c12(%rip),%xmm0 # 64d0 <_sk_callback_sse2+0xf4e> - .byte 15,89,13,27,28,0,0 // mulps 0x1c1b(%rip),%xmm1 # 64e0 <_sk_callback_sse2+0xf5e> + .byte 15,89,5,20,28,0,0 // mulps 0x1c14(%rip),%xmm0 # 6500 <_sk_callback_sse2+0xf50> + .byte 15,89,13,29,28,0,0 // mulps 0x1c1d(%rip),%xmm1 # 6510 <_sk_callback_sse2+0xf60> .byte 15,88,200 // addps %xmm0,%xmm1 - .byte 15,89,29,33,28,0,0 // mulps 0x1c21(%rip),%xmm3 # 64f0 <_sk_callback_sse2+0xf6e> + .byte 15,89,29,35,28,0,0 // mulps 0x1c23(%rip),%xmm3 # 6520 <_sk_callback_sse2+0xf70> .byte 15,88,217 // addps %xmm1,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,87,192 // xorps %xmm0,%xmm0 @@ -34636,9 +34682,9 @@ _sk_evenly_spaced_gradient_sse2: .byte 72,139,24 // mov (%rax),%rbx .byte 76,139,112,8 // mov 0x8(%rax),%r14 .byte 72,255,203 // dec %rbx - .byte 120,7 // js 4d0b <_sk_evenly_spaced_gradient_sse2+0x18> + .byte 120,7 // js 4d39 <_sk_evenly_spaced_gradient_sse2+0x18> .byte 243,72,15,42,203 // cvtsi2ss %rbx,%xmm1 - .byte 235,21 // jmp 4d20 <_sk_evenly_spaced_gradient_sse2+0x2d> + .byte 235,21 // jmp 4d4e <_sk_evenly_spaced_gradient_sse2+0x2d> .byte 73,137,217 // mov %rbx,%r9 .byte 73,209,233 // shr %r9 .byte 131,227,1 // and $0x1,%ebx @@ -34736,15 +34782,15 @@ HIDDEN _sk_gauss_a_to_rgba_sse2 .globl _sk_gauss_a_to_rgba_sse2 FUNCTION(_sk_gauss_a_to_rgba_sse2) _sk_gauss_a_to_rgba_sse2: - .byte 15,40,5,84,22,0,0 // movaps 0x1654(%rip),%xmm0 # 6500 <_sk_callback_sse2+0xf7e> + .byte 15,40,5,86,22,0,0 // movaps 0x1656(%rip),%xmm0 # 6530 <_sk_callback_sse2+0xf80> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,90,22,0,0 // addps 0x165a(%rip),%xmm0 # 6510 <_sk_callback_sse2+0xf8e> + .byte 15,88,5,92,22,0,0 // addps 0x165c(%rip),%xmm0 # 6540 <_sk_callback_sse2+0xf90> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,96,22,0,0 // addps 0x1660(%rip),%xmm0 # 6520 <_sk_callback_sse2+0xf9e> + .byte 15,88,5,98,22,0,0 // addps 0x1662(%rip),%xmm0 # 6550 <_sk_callback_sse2+0xfa0> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,102,22,0,0 // addps 0x1666(%rip),%xmm0 # 6530 <_sk_callback_sse2+0xfae> + .byte 15,88,5,104,22,0,0 // addps 0x1668(%rip),%xmm0 # 6560 <_sk_callback_sse2+0xfb0> .byte 15,89,195 // mulps %xmm3,%xmm0 - .byte 15,88,5,108,22,0,0 // addps 0x166c(%rip),%xmm0 # 6540 <_sk_callback_sse2+0xfbe> + .byte 15,88,5,110,22,0,0 // addps 0x166e(%rip),%xmm0 # 6570 <_sk_callback_sse2+0xfc0> .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 15,40,208 // movaps %xmm0,%xmm2 @@ -34762,12 +34808,12 @@ _sk_gradient_sse2: .byte 76,139,8 // mov (%rax),%r9 .byte 102,15,239,201 // pxor %xmm1,%xmm1 .byte 73,131,249,2 // cmp $0x2,%r9 - .byte 114,50 // jb 4f27 <_sk_gradient_sse2+0x46> + .byte 114,50 // jb 4f55 <_sk_gradient_sse2+0x46> .byte 72,139,88,72 // mov 0x48(%rax),%rbx .byte 73,255,201 // dec %r9 .byte 72,131,195,4 // add $0x4,%rbx .byte 102,15,239,201 // pxor %xmm1,%xmm1 - .byte 15,40,21,69,22,0,0 // movaps 0x1645(%rip),%xmm2 # 6550 <_sk_callback_sse2+0xfce> + .byte 15,40,21,71,22,0,0 // movaps 0x1647(%rip),%xmm2 # 6580 <_sk_callback_sse2+0xfd0> .byte 243,15,16,27 // movss (%rbx),%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 15,194,216,2 // cmpleps %xmm0,%xmm3 @@ -34775,7 +34821,7 @@ _sk_gradient_sse2: .byte 102,15,254,203 // paddd %xmm3,%xmm1 .byte 72,131,195,4 // add $0x4,%rbx .byte 73,255,201 // dec %r9 - .byte 117,228 // jne 4f0b <_sk_gradient_sse2+0x2a> + .byte 117,228 // jne 4f39 <_sk_gradient_sse2+0x2a> .byte 102,15,112,209,78 // pshufd $0x4e,%xmm1,%xmm2 .byte 102,73,15,126,211 // movq %xmm2,%r11 .byte 69,137,217 // mov %r11d,%r9d @@ -34914,29 +34960,29 @@ _sk_xy_to_unit_angle_sse2: .byte 69,15,94,220 // divps %xmm12,%xmm11 .byte 69,15,40,227 // movaps %xmm11,%xmm12 .byte 69,15,89,228 // mulps %xmm12,%xmm12 - .byte 68,15,40,45,6,20,0,0 // movaps 0x1406(%rip),%xmm13 # 6560 <_sk_callback_sse2+0xfde> + .byte 68,15,40,45,8,20,0,0 // movaps 0x1408(%rip),%xmm13 # 6590 <_sk_callback_sse2+0xfe0> .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 68,15,88,45,10,20,0,0 // addps 0x140a(%rip),%xmm13 # 6570 <_sk_callback_sse2+0xfee> + .byte 68,15,88,45,12,20,0,0 // addps 0x140c(%rip),%xmm13 # 65a0 <_sk_callback_sse2+0xff0> .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 68,15,88,45,14,20,0,0 // addps 0x140e(%rip),%xmm13 # 6580 <_sk_callback_sse2+0xffe> + .byte 68,15,88,45,16,20,0,0 // addps 0x1410(%rip),%xmm13 # 65b0 <_sk_callback_sse2+0x1000> .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 68,15,88,45,18,20,0,0 // addps 0x1412(%rip),%xmm13 # 6590 <_sk_callback_sse2+0x100e> + .byte 68,15,88,45,20,20,0,0 // addps 0x1414(%rip),%xmm13 # 65c0 <_sk_callback_sse2+0x1010> .byte 69,15,89,235 // mulps %xmm11,%xmm13 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 - .byte 68,15,40,21,17,20,0,0 // movaps 0x1411(%rip),%xmm10 # 65a0 <_sk_callback_sse2+0x101e> + .byte 68,15,40,21,19,20,0,0 // movaps 0x1413(%rip),%xmm10 # 65d0 <_sk_callback_sse2+0x1020> .byte 69,15,92,213 // subps %xmm13,%xmm10 .byte 69,15,84,209 // andps %xmm9,%xmm10 .byte 69,15,85,205 // andnps %xmm13,%xmm9 .byte 69,15,86,202 // orps %xmm10,%xmm9 .byte 68,15,194,192,1 // cmpltps %xmm0,%xmm8 - .byte 68,15,40,21,4,20,0,0 // movaps 0x1404(%rip),%xmm10 # 65b0 <_sk_callback_sse2+0x102e> + .byte 68,15,40,21,6,20,0,0 // movaps 0x1406(%rip),%xmm10 # 65e0 <_sk_callback_sse2+0x1030> .byte 69,15,92,209 // subps %xmm9,%xmm10 .byte 69,15,84,208 // andps %xmm8,%xmm10 .byte 69,15,85,193 // andnps %xmm9,%xmm8 .byte 69,15,86,194 // orps %xmm10,%xmm8 .byte 68,15,40,201 // movaps %xmm1,%xmm9 .byte 68,15,194,200,1 // cmpltps %xmm0,%xmm9 - .byte 68,15,40,21,243,19,0,0 // movaps 0x13f3(%rip),%xmm10 # 65c0 <_sk_callback_sse2+0x103e> + .byte 68,15,40,21,245,19,0,0 // movaps 0x13f5(%rip),%xmm10 # 65f0 <_sk_callback_sse2+0x1040> .byte 69,15,92,208 // subps %xmm8,%xmm10 .byte 69,15,84,209 // andps %xmm9,%xmm10 .byte 69,15,85,200 // andnps %xmm8,%xmm9 @@ -34963,7 +35009,7 @@ HIDDEN _sk_save_xy_sse2 FUNCTION(_sk_save_xy_sse2) _sk_save_xy_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,197,19,0,0 // movaps 0x13c5(%rip),%xmm8 # 65d0 <_sk_callback_sse2+0x104e> + .byte 68,15,40,5,199,19,0,0 // movaps 0x13c7(%rip),%xmm8 # 6600 <_sk_callback_sse2+0x1050> .byte 15,17,0 // movups %xmm0,(%rax) .byte 68,15,40,200 // movaps %xmm0,%xmm9 .byte 69,15,88,200 // addps %xmm8,%xmm9 @@ -34971,7 +35017,7 @@ _sk_save_xy_sse2: .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 .byte 69,15,40,217 // movaps %xmm9,%xmm11 .byte 69,15,194,218,1 // cmpltps %xmm10,%xmm11 - .byte 68,15,40,37,176,19,0,0 // movaps 0x13b0(%rip),%xmm12 # 65e0 <_sk_callback_sse2+0x105e> + .byte 68,15,40,37,178,19,0,0 // movaps 0x13b2(%rip),%xmm12 # 6610 <_sk_callback_sse2+0x1060> .byte 69,15,84,220 // andps %xmm12,%xmm11 .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 69,15,92,202 // subps %xmm10,%xmm9 @@ -35018,8 +35064,8 @@ _sk_bilinear_nx_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,41,19,0,0 // addps 0x1329(%rip),%xmm0 # 65f0 <_sk_callback_sse2+0x106e> - .byte 68,15,40,13,49,19,0,0 // movaps 0x1331(%rip),%xmm9 # 6600 <_sk_callback_sse2+0x107e> + .byte 15,88,5,43,19,0,0 // addps 0x132b(%rip),%xmm0 # 6620 <_sk_callback_sse2+0x1070> + .byte 68,15,40,13,51,19,0,0 // movaps 0x1333(%rip),%xmm9 # 6630 <_sk_callback_sse2+0x1080> .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 @@ -35032,7 +35078,7 @@ _sk_bilinear_px_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,32,19,0,0 // addps 0x1320(%rip),%xmm0 # 6610 <_sk_callback_sse2+0x108e> + .byte 15,88,5,34,19,0,0 // addps 0x1322(%rip),%xmm0 # 6640 <_sk_callback_sse2+0x1090> .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 @@ -35044,8 +35090,8 @@ _sk_bilinear_ny_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,18,19,0,0 // addps 0x1312(%rip),%xmm1 # 6620 <_sk_callback_sse2+0x109e> - .byte 68,15,40,13,26,19,0,0 // movaps 0x131a(%rip),%xmm9 # 6630 <_sk_callback_sse2+0x10ae> + .byte 15,88,13,20,19,0,0 // addps 0x1314(%rip),%xmm1 # 6650 <_sk_callback_sse2+0x10a0> + .byte 68,15,40,13,28,19,0,0 // movaps 0x131c(%rip),%xmm9 # 6660 <_sk_callback_sse2+0x10b0> .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 @@ -35058,7 +35104,7 @@ _sk_bilinear_py_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,8,19,0,0 // addps 0x1308(%rip),%xmm1 # 6640 <_sk_callback_sse2+0x10be> + .byte 15,88,13,10,19,0,0 // addps 0x130a(%rip),%xmm1 # 6670 <_sk_callback_sse2+0x10c0> .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 @@ -35070,13 +35116,13 @@ _sk_bicubic_n3x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,251,18,0,0 // addps 0x12fb(%rip),%xmm0 # 6650 <_sk_callback_sse2+0x10ce> - .byte 68,15,40,13,3,19,0,0 // movaps 0x1303(%rip),%xmm9 # 6660 <_sk_callback_sse2+0x10de> + .byte 15,88,5,253,18,0,0 // addps 0x12fd(%rip),%xmm0 # 6680 <_sk_callback_sse2+0x10d0> + .byte 68,15,40,13,5,19,0,0 // movaps 0x1305(%rip),%xmm9 # 6690 <_sk_callback_sse2+0x10e0> .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 68,15,89,13,255,18,0,0 // mulps 0x12ff(%rip),%xmm9 # 6670 <_sk_callback_sse2+0x10ee> - .byte 68,15,88,13,7,19,0,0 // addps 0x1307(%rip),%xmm9 # 6680 <_sk_callback_sse2+0x10fe> + .byte 68,15,89,13,1,19,0,0 // mulps 0x1301(%rip),%xmm9 # 66a0 <_sk_callback_sse2+0x10f0> + .byte 68,15,88,13,9,19,0,0 // addps 0x1309(%rip),%xmm9 # 66b0 <_sk_callback_sse2+0x1100> .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 @@ -35089,16 +35135,16 @@ _sk_bicubic_n1x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,246,18,0,0 // addps 0x12f6(%rip),%xmm0 # 6690 <_sk_callback_sse2+0x110e> - .byte 68,15,40,13,254,18,0,0 // movaps 0x12fe(%rip),%xmm9 # 66a0 <_sk_callback_sse2+0x111e> + .byte 15,88,5,248,18,0,0 // addps 0x12f8(%rip),%xmm0 # 66c0 <_sk_callback_sse2+0x1110> + .byte 68,15,40,13,0,19,0,0 // movaps 0x1300(%rip),%xmm9 # 66d0 <_sk_callback_sse2+0x1120> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,2,19,0,0 // movaps 0x1302(%rip),%xmm8 # 66b0 <_sk_callback_sse2+0x112e> + .byte 68,15,40,5,4,19,0,0 // movaps 0x1304(%rip),%xmm8 # 66e0 <_sk_callback_sse2+0x1130> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,6,19,0,0 // addps 0x1306(%rip),%xmm8 # 66c0 <_sk_callback_sse2+0x113e> + .byte 68,15,88,5,8,19,0,0 // addps 0x1308(%rip),%xmm8 # 66f0 <_sk_callback_sse2+0x1140> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,10,19,0,0 // addps 0x130a(%rip),%xmm8 # 66d0 <_sk_callback_sse2+0x114e> + .byte 68,15,88,5,12,19,0,0 // addps 0x130c(%rip),%xmm8 # 6700 <_sk_callback_sse2+0x1150> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,14,19,0,0 // addps 0x130e(%rip),%xmm8 # 66e0 <_sk_callback_sse2+0x115e> + .byte 68,15,88,5,16,19,0,0 // addps 0x1310(%rip),%xmm8 # 6710 <_sk_callback_sse2+0x1160> .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 @@ -35108,17 +35154,17 @@ HIDDEN _sk_bicubic_p1x_sse2 FUNCTION(_sk_bicubic_p1x_sse2) _sk_bicubic_p1x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,8,19,0,0 // movaps 0x1308(%rip),%xmm8 # 66f0 <_sk_callback_sse2+0x116e> + .byte 68,15,40,5,10,19,0,0 // movaps 0x130a(%rip),%xmm8 # 6720 <_sk_callback_sse2+0x1170> .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 68,15,40,21,4,19,0,0 // movaps 0x1304(%rip),%xmm10 # 6700 <_sk_callback_sse2+0x117e> + .byte 68,15,40,21,6,19,0,0 // movaps 0x1306(%rip),%xmm10 # 6730 <_sk_callback_sse2+0x1180> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,8,19,0,0 // addps 0x1308(%rip),%xmm10 # 6710 <_sk_callback_sse2+0x118e> + .byte 68,15,88,21,10,19,0,0 // addps 0x130a(%rip),%xmm10 # 6740 <_sk_callback_sse2+0x1190> .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 68,15,88,21,4,19,0,0 // addps 0x1304(%rip),%xmm10 # 6720 <_sk_callback_sse2+0x119e> + .byte 68,15,88,21,6,19,0,0 // addps 0x1306(%rip),%xmm10 # 6750 <_sk_callback_sse2+0x11a0> .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 @@ -35130,11 +35176,11 @@ _sk_bicubic_p3x_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,0 // movups (%rax),%xmm0 .byte 68,15,16,64,64 // movups 0x40(%rax),%xmm8 - .byte 15,88,5,247,18,0,0 // addps 0x12f7(%rip),%xmm0 # 6730 <_sk_callback_sse2+0x11ae> + .byte 15,88,5,249,18,0,0 // addps 0x12f9(%rip),%xmm0 # 6760 <_sk_callback_sse2+0x11b0> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,247,18,0,0 // mulps 0x12f7(%rip),%xmm8 # 6740 <_sk_callback_sse2+0x11be> - .byte 68,15,88,5,255,18,0,0 // addps 0x12ff(%rip),%xmm8 # 6750 <_sk_callback_sse2+0x11ce> + .byte 68,15,89,5,249,18,0,0 // mulps 0x12f9(%rip),%xmm8 # 6770 <_sk_callback_sse2+0x11c0> + .byte 68,15,88,5,1,19,0,0 // addps 0x1301(%rip),%xmm8 # 6780 <_sk_callback_sse2+0x11d0> .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 @@ -35147,13 +35193,13 @@ _sk_bicubic_n3y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,237,18,0,0 // addps 0x12ed(%rip),%xmm1 # 6760 <_sk_callback_sse2+0x11de> - .byte 68,15,40,13,245,18,0,0 // movaps 0x12f5(%rip),%xmm9 # 6770 <_sk_callback_sse2+0x11ee> + .byte 15,88,13,239,18,0,0 // addps 0x12ef(%rip),%xmm1 # 6790 <_sk_callback_sse2+0x11e0> + .byte 68,15,40,13,247,18,0,0 // movaps 0x12f7(%rip),%xmm9 # 67a0 <_sk_callback_sse2+0x11f0> .byte 69,15,92,200 // subps %xmm8,%xmm9 .byte 69,15,40,193 // movaps %xmm9,%xmm8 .byte 69,15,89,192 // mulps %xmm8,%xmm8 - .byte 68,15,89,13,241,18,0,0 // mulps 0x12f1(%rip),%xmm9 # 6780 <_sk_callback_sse2+0x11fe> - .byte 68,15,88,13,249,18,0,0 // addps 0x12f9(%rip),%xmm9 # 6790 <_sk_callback_sse2+0x120e> + .byte 68,15,89,13,243,18,0,0 // mulps 0x12f3(%rip),%xmm9 # 67b0 <_sk_callback_sse2+0x1200> + .byte 68,15,88,13,251,18,0,0 // addps 0x12fb(%rip),%xmm9 # 67c0 <_sk_callback_sse2+0x1210> .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 @@ -35166,16 +35212,16 @@ _sk_bicubic_n1y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,231,18,0,0 // addps 0x12e7(%rip),%xmm1 # 67a0 <_sk_callback_sse2+0x121e> - .byte 68,15,40,13,239,18,0,0 // movaps 0x12ef(%rip),%xmm9 # 67b0 <_sk_callback_sse2+0x122e> + .byte 15,88,13,233,18,0,0 // addps 0x12e9(%rip),%xmm1 # 67d0 <_sk_callback_sse2+0x1220> + .byte 68,15,40,13,241,18,0,0 // movaps 0x12f1(%rip),%xmm9 # 67e0 <_sk_callback_sse2+0x1230> .byte 69,15,92,200 // subps %xmm8,%xmm9 - .byte 68,15,40,5,243,18,0,0 // movaps 0x12f3(%rip),%xmm8 # 67c0 <_sk_callback_sse2+0x123e> + .byte 68,15,40,5,245,18,0,0 // movaps 0x12f5(%rip),%xmm8 # 67f0 <_sk_callback_sse2+0x1240> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,247,18,0,0 // addps 0x12f7(%rip),%xmm8 # 67d0 <_sk_callback_sse2+0x124e> + .byte 68,15,88,5,249,18,0,0 // addps 0x12f9(%rip),%xmm8 # 6800 <_sk_callback_sse2+0x1250> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,251,18,0,0 // addps 0x12fb(%rip),%xmm8 # 67e0 <_sk_callback_sse2+0x125e> + .byte 68,15,88,5,253,18,0,0 // addps 0x12fd(%rip),%xmm8 # 6810 <_sk_callback_sse2+0x1260> .byte 69,15,89,193 // mulps %xmm9,%xmm8 - .byte 68,15,88,5,255,18,0,0 // addps 0x12ff(%rip),%xmm8 # 67f0 <_sk_callback_sse2+0x126e> + .byte 68,15,88,5,1,19,0,0 // addps 0x1301(%rip),%xmm8 # 6820 <_sk_callback_sse2+0x1270> .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 @@ -35185,17 +35231,17 @@ HIDDEN _sk_bicubic_p1y_sse2 FUNCTION(_sk_bicubic_p1y_sse2) _sk_bicubic_p1y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 68,15,40,5,249,18,0,0 // movaps 0x12f9(%rip),%xmm8 # 6800 <_sk_callback_sse2+0x127e> + .byte 68,15,40,5,251,18,0,0 // movaps 0x12fb(%rip),%xmm8 # 6830 <_sk_callback_sse2+0x1280> .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 68,15,40,21,244,18,0,0 // movaps 0x12f4(%rip),%xmm10 # 6810 <_sk_callback_sse2+0x128e> + .byte 68,15,40,21,246,18,0,0 // movaps 0x12f6(%rip),%xmm10 # 6840 <_sk_callback_sse2+0x1290> .byte 69,15,89,209 // mulps %xmm9,%xmm10 - .byte 68,15,88,21,248,18,0,0 // addps 0x12f8(%rip),%xmm10 # 6820 <_sk_callback_sse2+0x129e> + .byte 68,15,88,21,250,18,0,0 // addps 0x12fa(%rip),%xmm10 # 6850 <_sk_callback_sse2+0x12a0> .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 68,15,88,21,244,18,0,0 // addps 0x12f4(%rip),%xmm10 # 6830 <_sk_callback_sse2+0x12ae> + .byte 68,15,88,21,246,18,0,0 // addps 0x12f6(%rip),%xmm10 # 6860 <_sk_callback_sse2+0x12b0> .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 @@ -35207,11 +35253,11 @@ _sk_bicubic_p3y_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 .byte 68,15,16,64,96 // movups 0x60(%rax),%xmm8 - .byte 15,88,13,230,18,0,0 // addps 0x12e6(%rip),%xmm1 # 6840 <_sk_callback_sse2+0x12be> + .byte 15,88,13,232,18,0,0 // addps 0x12e8(%rip),%xmm1 # 6870 <_sk_callback_sse2+0x12c0> .byte 69,15,40,200 // movaps %xmm8,%xmm9 .byte 69,15,89,201 // mulps %xmm9,%xmm9 - .byte 68,15,89,5,230,18,0,0 // mulps 0x12e6(%rip),%xmm8 # 6850 <_sk_callback_sse2+0x12ce> - .byte 68,15,88,5,238,18,0,0 // addps 0x12ee(%rip),%xmm8 # 6860 <_sk_callback_sse2+0x12de> + .byte 68,15,89,5,232,18,0,0 // mulps 0x12e8(%rip),%xmm8 # 6880 <_sk_callback_sse2+0x12d0> + .byte 68,15,88,5,240,18,0,0 // addps 0x12f0(%rip),%xmm8 # 6890 <_sk_callback_sse2+0x12e0> .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 @@ -35441,11 +35487,11 @@ BALIGN16 .byte 128,191,0,0,128,191,0 // cmpb $0x0,-0x40800000(%rdi) .byte 0,224 // add %ah,%al .byte 64,0,0 // add %al,(%rax) - .byte 224,64 // loopne 5858 <.literal16+0x1d8> + .byte 224,64 // loopne 5888 <.literal16+0x1d8> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 585c <.literal16+0x1dc> + .byte 224,64 // loopne 588c <.literal16+0x1dc> .byte 0,0 // add %al,(%rax) - .byte 224,64 // loopne 5860 <.literal16+0x1e0> + .byte 224,64 // loopne 5890 <.literal16+0x1e0> .byte 154 // (bad) .byte 153 // cltd .byte 153 // cltd @@ -35465,13 +35511,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5881 <.literal16+0x201> + .byte 71,225,61 // rex.RXB loope 58b1 <.literal16+0x201> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5885 <.literal16+0x205> + .byte 71,225,61 // rex.RXB loope 58b5 <.literal16+0x205> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5889 <.literal16+0x209> + .byte 71,225,61 // rex.RXB loope 58b9 <.literal16+0x209> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 588d <.literal16+0x20d> + .byte 71,225,61 // rex.RXB loope 58bd <.literal16+0x20d> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -35496,13 +35542,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 58c1 <.literal16+0x241> + .byte 71,225,61 // rex.RXB loope 58f1 <.literal16+0x241> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 58c5 <.literal16+0x245> + .byte 71,225,61 // rex.RXB loope 58f5 <.literal16+0x245> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 58c9 <.literal16+0x249> + .byte 71,225,61 // rex.RXB loope 58f9 <.literal16+0x249> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 58cd <.literal16+0x24d> + .byte 71,225,61 // rex.RXB loope 58fd <.literal16+0x24d> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -35527,13 +35573,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5901 <.literal16+0x281> + .byte 71,225,61 // rex.RXB loope 5931 <.literal16+0x281> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5905 <.literal16+0x285> + .byte 71,225,61 // rex.RXB loope 5935 <.literal16+0x285> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5909 <.literal16+0x289> + .byte 71,225,61 // rex.RXB loope 5939 <.literal16+0x289> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 590d <.literal16+0x28d> + .byte 71,225,61 // rex.RXB loope 593d <.literal16+0x28d> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -35558,13 +35604,13 @@ BALIGN16 .byte 10,23 // or (%rdi),%dl .byte 63 // (bad) .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5941 <.literal16+0x2c1> + .byte 71,225,61 // rex.RXB loope 5971 <.literal16+0x2c1> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5945 <.literal16+0x2c5> + .byte 71,225,61 // rex.RXB loope 5975 <.literal16+0x2c5> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 5949 <.literal16+0x2c9> + .byte 71,225,61 // rex.RXB loope 5979 <.literal16+0x2c9> .byte 174 // scas %es:(%rdi),%al - .byte 71,225,61 // rex.RXB loope 594d <.literal16+0x2cd> + .byte 71,225,61 // rex.RXB loope 597d <.literal16+0x2cd> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -35585,11 +35631,11 @@ BALIGN16 .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax) .byte 67,0,0 // rex.XB add %al,(%r8) - .byte 127,67 // jg 598b <.literal16+0x30b> + .byte 127,67 // jg 59bb <.literal16+0x30b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 598f <.literal16+0x30f> + .byte 127,67 // jg 59bf <.literal16+0x30f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5993 <.literal16+0x313> + .byte 127,67 // jg 59c3 <.literal16+0x313> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -35829,13 +35875,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 5b79 <.literal16+0x4f9> + .byte 224,7 // loopne 5ba9 <.literal16+0x4f9> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5b7d <.literal16+0x4fd> + .byte 224,7 // loopne 5bad <.literal16+0x4fd> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5b81 <.literal16+0x501> + .byte 224,7 // loopne 5bb1 <.literal16+0x501> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 5b85 <.literal16+0x505> + .byte 224,7 // loopne 5bb5 <.literal16+0x505> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -35900,11 +35946,11 @@ 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 5c5b <.literal16+0x5db> + .byte 127,67 // jg 5c8b <.literal16+0x5db> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5c5f <.literal16+0x5df> + .byte 127,67 // jg 5c8f <.literal16+0x5df> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5c63 <.literal16+0x5e3> + .byte 127,67 // jg 5c93 <.literal16+0x5e3> .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) @@ -35919,16 +35965,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5c54 <.literal16+0x5d4> + .byte 127,0 // jg 5c84 <.literal16+0x5d4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5c58 <.literal16+0x5d8> + .byte 127,0 // jg 5c88 <.literal16+0x5d8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5c5c <.literal16+0x5dc> + .byte 127,0 // jg 5c8c <.literal16+0x5dc> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5c60 <.literal16+0x5e0> + .byte 127,0 // jg 5c90 <.literal16+0x5e0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -35937,7 +35983,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5ce5 <.literal16+0x665> + .byte 119,115 // ja 5d15 <.literal16+0x665> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -35948,7 +35994,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 5c49 <.literal16+0x5c9> + .byte 117,191 // jne 5c79 <.literal16+0x5c9> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -35960,7 +36006,7 @@ 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 ffffffffe9a39c8a <_sk_callback_sse2+0xffffffffe9a34708> + .byte 233,220,63,163,233 // jmpq ffffffffe9a39cba <_sk_callback_sse2+0xffffffffe9a3470a> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -36014,16 +36060,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5d24 <.literal16+0x6a4> + .byte 127,0 // jg 5d54 <.literal16+0x6a4> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5d28 <.literal16+0x6a8> + .byte 127,0 // jg 5d58 <.literal16+0x6a8> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5d2c <.literal16+0x6ac> + .byte 127,0 // jg 5d5c <.literal16+0x6ac> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5d30 <.literal16+0x6b0> + .byte 127,0 // jg 5d60 <.literal16+0x6b0> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -36032,7 +36078,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5db5 <.literal16+0x735> + .byte 119,115 // ja 5de5 <.literal16+0x735> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -36043,7 +36089,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 5d19 <.literal16+0x699> + .byte 117,191 // jne 5d49 <.literal16+0x699> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -36055,7 +36101,7 @@ 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 ffffffffe9a39d5a <_sk_callback_sse2+0xffffffffe9a347d8> + .byte 233,220,63,163,233 // jmpq ffffffffe9a39d8a <_sk_callback_sse2+0xffffffffe9a347da> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -36109,16 +36155,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5df4 <.literal16+0x774> + .byte 127,0 // jg 5e24 <.literal16+0x774> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5df8 <.literal16+0x778> + .byte 127,0 // jg 5e28 <.literal16+0x778> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5dfc <.literal16+0x77c> + .byte 127,0 // jg 5e2c <.literal16+0x77c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5e00 <.literal16+0x780> + .byte 127,0 // jg 5e30 <.literal16+0x780> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -36127,7 +36173,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5e85 <.literal16+0x805> + .byte 119,115 // ja 5eb5 <.literal16+0x805> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -36138,7 +36184,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 5de9 <.literal16+0x769> + .byte 117,191 // jne 5e19 <.literal16+0x769> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -36150,7 +36196,7 @@ 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 ffffffffe9a39e2a <_sk_callback_sse2+0xffffffffe9a348a8> + .byte 233,220,63,163,233 // jmpq ffffffffe9a39e5a <_sk_callback_sse2+0xffffffffe9a348aa> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -36204,16 +36250,16 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 5ec4 <.literal16+0x844> + .byte 127,0 // jg 5ef4 <.literal16+0x844> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5ec8 <.literal16+0x848> + .byte 127,0 // jg 5ef8 <.literal16+0x848> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5ecc <.literal16+0x84c> + .byte 127,0 // jg 5efc <.literal16+0x84c> .byte 255 // (bad) .byte 255 // (bad) - .byte 127,0 // jg 5ed0 <.literal16+0x850> + .byte 127,0 // jg 5f00 <.literal16+0x850> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -36222,7 +36268,7 @@ BALIGN16 .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 5f55 <.literal16+0x8d5> + .byte 119,115 // ja 5f85 <.literal16+0x8d5> .byte 248 // clc .byte 194,119,115 // retq $0x7377 .byte 248 // clc @@ -36233,7 +36279,7 @@ BALIGN16 .byte 194,117,191 // retq $0xbf75 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) - .byte 117,191 // jne 5eb9 <.literal16+0x839> + .byte 117,191 // jne 5ee9 <.literal16+0x839> .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi .byte 63 // (bad) .byte 249 // stc @@ -36245,7 +36291,7 @@ 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 ffffffffe9a39efa <_sk_callback_sse2+0xffffffffe9a34978> + .byte 233,220,63,163,233 // jmpq ffffffffe9a39f2a <_sk_callback_sse2+0xffffffffe9a3497a> .byte 220,63 // fdivrl (%rdi) .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) @@ -36295,13 +36341,13 @@ 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 5fd7 <.literal16+0x957> + .byte 127,67 // jg 6007 <.literal16+0x957> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5fdb <.literal16+0x95b> + .byte 127,67 // jg 600b <.literal16+0x95b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5fdf <.literal16+0x95f> + .byte 127,67 // jg 600f <.literal16+0x95f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 5fe3 <.literal16+0x963> + .byte 127,67 // jg 6013 <.literal16+0x963> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -36348,16 +36394,16 @@ BALIGN16 .byte 128,3,62 // addb $0x3e,(%rbx) .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 6063 <.literal16+0x9e3> + .byte 118,63 // jbe 6093 <.literal16+0x9e3> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 6067 <.literal16+0x9e7> + .byte 118,63 // jbe 6097 <.literal16+0x9e7> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 606b <.literal16+0x9eb> + .byte 118,63 // jbe 609b <.literal16+0x9eb> .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 606f <.literal16+0x9ef> + .byte 118,63 // jbe 609f <.literal16+0x9ef> .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) @@ -36377,11 +36423,11 @@ 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 60bb <.literal16+0xa3b> + .byte 127,67 // jg 60eb <.literal16+0xa3b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 60bf <.literal16+0xa3f> + .byte 127,67 // jg 60ef <.literal16+0xa3f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 60c3 <.literal16+0xa43> + .byte 127,67 // jg 60f3 <.literal16+0xa43> .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -36429,13 +36475,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 6119 <.literal16+0xa99> + .byte 224,7 // loopne 6149 <.literal16+0xa99> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 611d <.literal16+0xa9d> + .byte 224,7 // loopne 614d <.literal16+0xa9d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 6121 <.literal16+0xaa1> + .byte 224,7 // loopne 6151 <.literal16+0xaa1> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 6125 <.literal16+0xaa5> + .byte 224,7 // loopne 6155 <.literal16+0xaa5> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -36481,13 +36527,13 @@ BALIGN16 .byte 132,55 // test %dh,(%rdi) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 6189 <.literal16+0xb09> + .byte 224,7 // loopne 61b9 <.literal16+0xb09> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 618d <.literal16+0xb0d> + .byte 224,7 // loopne 61bd <.literal16+0xb0d> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 6191 <.literal16+0xb11> + .byte 224,7 // loopne 61c1 <.literal16+0xb11> .byte 0,0 // add %al,(%rax) - .byte 224,7 // loopne 6195 <.literal16+0xb15> + .byte 224,7 // loopne 61c5 <.literal16+0xb15> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -36525,13 +36571,13 @@ BALIGN16 .byte 65,0,0 // add %al,(%r8) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 6226 <.literal16+0xba6> + .byte 124,66 // jl 6256 <.literal16+0xba6> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 622a <.literal16+0xbaa> + .byte 124,66 // jl 625a <.literal16+0xbaa> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 622e <.literal16+0xbae> + .byte 124,66 // jl 625e <.literal16+0xbae> .byte 0,0 // add %al,(%rax) - .byte 124,66 // jl 6232 <.literal16+0xbb2> + .byte 124,66 // jl 6262 <.literal16+0xbb2> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 0,240 // add %dh,%al @@ -36621,13 +36667,13 @@ 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 6335 <.literal16+0xcb5> + .byte 112,65 // jo 6365 <.literal16+0xcb5> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 6339 <.literal16+0xcb9> + .byte 112,65 // jo 6369 <.literal16+0xcb9> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 633d <.literal16+0xcbd> + .byte 112,65 // jo 636d <.literal16+0xcbd> .byte 0,0 // add %al,(%rax) - .byte 112,65 // jo 6341 <.literal16+0xcc1> + .byte 112,65 // jo 6371 <.literal16+0xcc1> .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -36649,11 +36695,11 @@ BALIGN16 .byte 128,59,129 // cmpb $0x81,(%rbx) .byte 128,128,59,0,0,127,67 // addb $0x43,0x7f00003b(%rax) .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 638b <.literal16+0xd0b> + .byte 127,67 // jg 63bb <.literal16+0xd0b> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 638f <.literal16+0xd0f> + .byte 127,67 // jg 63bf <.literal16+0xd0f> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 6393 <.literal16+0xd13> + .byte 127,67 // jg 63c3 <.literal16+0xd13> .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) @@ -36729,13 +36775,13 @@ BALIGN16 .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 255 // (bad) - .byte 127,71 // jg 647b <.literal16+0xdfb> + .byte 127,71 // jg 64ab <.literal16+0xdfb> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 647f <.literal16+0xdff> + .byte 127,71 // jg 64af <.literal16+0xdff> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 6483 <.literal16+0xe03> + .byte 127,71 // jg 64b3 <.literal16+0xe03> .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 6487 <.literal16+0xe07> + .byte 127,71 // jg 64b7 <.literal16+0xe07> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax) @@ -36812,10 +36858,10 @@ BALIGN16 .byte 61,152,221,147,61 // cmp $0x3d93dd98,%eax .byte 152 // cwtl .byte 221,147,61,45,16,17 // fstl 0x11102d3d(%rbx) - .byte 192,45,16,17,192,45,16 // shrb $0x10,0x2dc01110(%rip) # 2dc0761a <_sk_callback_sse2+0x2dc02098> + .byte 192,45,16,17,192,45,16 // shrb $0x10,0x2dc01110(%rip) # 2dc0764a <_sk_callback_sse2+0x2dc0209a> .byte 17,192 // adc %eax,%eax .byte 45,16,17,192,18 // sub $0x12c01110,%eax - .byte 120,57 // js 654c <.literal16+0xecc> + .byte 120,57 // js 657c <.literal16+0xecc> .byte 64,18,120,57 // adc 0x39(%rax),%dil .byte 64,18,120,57 // adc 0x39(%rax),%dil .byte 64,18,120,57 // adc 0x39(%rax),%dil @@ -36943,11 +36989,11 @@ 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 6692 <.literal16+0x1012> + .byte 62,114,28 // jb,pt 66c2 <.literal16+0x1012> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6696 <.literal16+0x1016> + .byte 62,114,28 // jb,pt 66c6 <.literal16+0x1016> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 669a <.literal16+0x101a> + .byte 62,114,28 // jb,pt 66ca <.literal16+0x101a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -36991,7 +37037,7 @@ 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 # 3d63f525 <_sk_callback_sse2+0x3d639fa3> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f555 <_sk_callback_sse2+0x3d639fa5> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -37017,7 +37063,7 @@ 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 # 3d63f565 <_sk_callback_sse2+0x3d639fe3> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f595 <_sk_callback_sse2+0x3d639fe5> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -37026,13 +37072,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 675e <.literal16+0x10de> + .byte 114,28 // jb 678e <.literal16+0x10de> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6762 <.literal16+0x10e2> + .byte 62,114,28 // jb,pt 6792 <.literal16+0x10e2> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6766 <.literal16+0x10e6> + .byte 62,114,28 // jb,pt 6796 <.literal16+0x10e6> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 676a <.literal16+0x10ea> + .byte 62,114,28 // jb,pt 679a <.literal16+0x10ea> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -37053,11 +37099,11 @@ 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 67a2 <.literal16+0x1122> + .byte 62,114,28 // jb,pt 67d2 <.literal16+0x1122> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 67a6 <.literal16+0x1126> + .byte 62,114,28 // jb,pt 67d6 <.literal16+0x1126> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 67aa <.literal16+0x112a> + .byte 62,114,28 // jb,pt 67da <.literal16+0x112a> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -37101,7 +37147,7 @@ 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 # 3d63f635 <_sk_callback_sse2+0x3d63a0b3> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f665 <_sk_callback_sse2+0x3d63a0b5> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 0,63 // add %bh,(%rdi) .byte 0,0 // add %al,(%rax) @@ -37127,7 +37173,7 @@ 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 # 3d63f675 <_sk_callback_sse2+0x3d63a0f3> + .byte 99,61,57,142,99,61 // movslq 0x3d638e39(%rip),%edi # 3d63f6a5 <_sk_callback_sse2+0x3d63a0f5> .byte 57,142,99,61,0,0 // cmp %ecx,0x3d63(%rsi) .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al @@ -37136,13 +37182,13 @@ BALIGN16 .byte 192,63,0 // sarb $0x0,(%rdi) .byte 0,192 // add %al,%al .byte 63 // (bad) - .byte 114,28 // jb 686e <.literal16+0x11ee> + .byte 114,28 // jb 689e <.literal16+0x11ee> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6872 <_sk_callback_sse2+0x12f0> + .byte 62,114,28 // jb,pt 68a2 <_sk_callback_sse2+0x12f2> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 6876 <_sk_callback_sse2+0x12f4> + .byte 62,114,28 // jb,pt 68a6 <_sk_callback_sse2+0x12f6> .byte 199 // (bad) - .byte 62,114,28 // jb,pt 687a <_sk_callback_sse2+0x12f8> + .byte 62,114,28 // jb,pt 68aa <_sk_callback_sse2+0x12fa> .byte 199 // (bad) .byte 62,171 // ds stos %eax,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) @@ -37152,6 +37198,12 @@ BALIGN16 .byte 170 // stos %al,%es:(%rdi) .byte 170 // stos %al,%es:(%rdi) .byte 190,171,170,170,190 // mov $0xbeaaaaab,%esi + +BALIGN4 + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) + .byte 0,0 // add %al,(%rax) + .byte 0,63 // add %bh,(%rdi) BALIGN32 HIDDEN _sk_start_pipeline_ssse3_lowp diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index d38ce49117..dab73a0b6f 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -110,7 +110,7 @@ _sk_seed_shader_hsw LABEL PROC DB 197,249,110,194 ; vmovd %edx,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,237,74,0,0 ; vbroadcastss 0x4aed(%rip),%ymm1 # 4c60 <_sk_callback_hsw+0x12a> + DB 196,226,125,24,13,29,75,0,0 ; vbroadcastss 0x4b1d(%rip),%ymm1 # 4c90 <_sk_callback_hsw+0x12c> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,88,7 ; vaddps (%rdi),%ymm0,%ymm0 DB 197,249,110,209 ; vmovd %ecx,%xmm2 @@ -118,7 +118,7 @@ _sk_seed_shader_hsw LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,205,74,0,0 ; vbroadcastss 0x4acd(%rip),%ymm2 # 4c64 <_sk_callback_hsw+0x12e> + DB 196,226,125,24,21,253,74,0,0 ; vbroadcastss 0x4afd(%rip),%ymm2 # 4c94 <_sk_callback_hsw+0x130> DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 @@ -135,13 +135,13 @@ _sk_dither_hsw LABEL PROC DB 197,121,110,201 ; vmovd %ecx,%xmm9 DB 196,66,125,88,201 ; vpbroadcastd %xmm9,%ymm9 DB 196,65,53,239,200 ; vpxor %ymm8,%ymm9,%ymm9 - DB 196,98,125,88,21,148,74,0,0 ; vpbroadcastd 0x4a94(%rip),%ymm10 # 4c68 <_sk_callback_hsw+0x132> + DB 196,98,125,88,21,196,74,0,0 ; vpbroadcastd 0x4ac4(%rip),%ymm10 # 4c98 <_sk_callback_hsw+0x134> DB 196,65,53,219,218 ; vpand %ymm10,%ymm9,%ymm11 DB 196,193,37,114,243,5 ; vpslld $0x5,%ymm11,%ymm11 DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,193,45,114,242,4 ; vpslld $0x4,%ymm10,%ymm10 - DB 196,98,125,88,37,121,74,0,0 ; vpbroadcastd 0x4a79(%rip),%ymm12 # 4c6c <_sk_callback_hsw+0x136> - DB 196,98,125,88,45,116,74,0,0 ; vpbroadcastd 0x4a74(%rip),%ymm13 # 4c70 <_sk_callback_hsw+0x13a> + DB 196,98,125,88,37,169,74,0,0 ; vpbroadcastd 0x4aa9(%rip),%ymm12 # 4c9c <_sk_callback_hsw+0x138> + DB 196,98,125,88,45,164,74,0,0 ; vpbroadcastd 0x4aa4(%rip),%ymm13 # 4ca0 <_sk_callback_hsw+0x13c> DB 196,65,53,219,245 ; vpand %ymm13,%ymm9,%ymm14 DB 196,193,13,114,246,2 ; vpslld $0x2,%ymm14,%ymm14 DB 196,65,61,219,237 ; vpand %ymm13,%ymm8,%ymm13 @@ -156,8 +156,8 @@ _sk_dither_hsw LABEL PROC DB 196,65,61,235,194 ; vpor %ymm10,%ymm8,%ymm8 DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,38,74,0,0 ; vbroadcastss 0x4a26(%rip),%ymm9 # 4c74 <_sk_callback_hsw+0x13e> - DB 196,98,125,24,21,33,74,0,0 ; vbroadcastss 0x4a21(%rip),%ymm10 # 4c78 <_sk_callback_hsw+0x142> + DB 196,98,125,24,13,86,74,0,0 ; vbroadcastss 0x4a56(%rip),%ymm9 # 4ca4 <_sk_callback_hsw+0x140> + DB 196,98,125,24,21,81,74,0,0 ; vbroadcastss 0x4a51(%rip),%ymm10 # 4ca8 <_sk_callback_hsw+0x144> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 @@ -216,7 +216,7 @@ _sk_clear_hsw LABEL PROC PUBLIC _sk_srcatop_hsw _sk_srcatop_hsw LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,121,73,0,0 ; vbroadcastss 0x4979(%rip),%ymm8 # 4c7c <_sk_callback_hsw+0x146> + DB 196,98,125,24,5,169,73,0,0 ; vbroadcastss 0x49a9(%rip),%ymm8 # 4cac <_sk_callback_hsw+0x148> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,226,61,184,196 ; vfmadd231ps %ymm4,%ymm8,%ymm0 DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 @@ -230,7 +230,7 @@ _sk_srcatop_hsw LABEL PROC PUBLIC _sk_dstatop_hsw _sk_dstatop_hsw LABEL PROC - DB 196,98,125,24,5,76,73,0,0 ; vbroadcastss 0x494c(%rip),%ymm8 # 4c80 <_sk_callback_hsw+0x14a> + DB 196,98,125,24,5,124,73,0,0 ; vbroadcastss 0x497c(%rip),%ymm8 # 4cb0 <_sk_callback_hsw+0x14c> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,226,101,184,196 ; vfmadd231ps %ymm4,%ymm3,%ymm0 @@ -263,7 +263,7 @@ _sk_dstin_hsw LABEL PROC PUBLIC _sk_srcout_hsw _sk_srcout_hsw LABEL PROC - DB 196,98,125,24,5,243,72,0,0 ; vbroadcastss 0x48f3(%rip),%ymm8 # 4c84 <_sk_callback_hsw+0x14e> + DB 196,98,125,24,5,35,73,0,0 ; vbroadcastss 0x4923(%rip),%ymm8 # 4cb4 <_sk_callback_hsw+0x150> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -274,7 +274,7 @@ _sk_srcout_hsw LABEL PROC PUBLIC _sk_dstout_hsw _sk_dstout_hsw LABEL PROC - DB 196,226,125,24,5,214,72,0,0 ; vbroadcastss 0x48d6(%rip),%ymm0 # 4c88 <_sk_callback_hsw+0x152> + DB 196,226,125,24,5,6,73,0,0 ; vbroadcastss 0x4906(%rip),%ymm0 # 4cb8 <_sk_callback_hsw+0x154> DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3 DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0 DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1 @@ -285,7 +285,7 @@ _sk_dstout_hsw LABEL PROC PUBLIC _sk_srcover_hsw _sk_srcover_hsw LABEL PROC - DB 196,98,125,24,5,185,72,0,0 ; vbroadcastss 0x48b9(%rip),%ymm8 # 4c8c <_sk_callback_hsw+0x156> + DB 196,98,125,24,5,233,72,0,0 ; vbroadcastss 0x48e9(%rip),%ymm8 # 4cbc <_sk_callback_hsw+0x158> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0 DB 196,194,85,184,200 ; vfmadd231ps %ymm8,%ymm5,%ymm1 @@ -296,7 +296,7 @@ _sk_srcover_hsw LABEL PROC PUBLIC _sk_dstover_hsw _sk_dstover_hsw LABEL PROC - DB 196,98,125,24,5,152,72,0,0 ; vbroadcastss 0x4898(%rip),%ymm8 # 4c90 <_sk_callback_hsw+0x15a> + DB 196,98,125,24,5,200,72,0,0 ; vbroadcastss 0x48c8(%rip),%ymm8 # 4cc0 <_sk_callback_hsw+0x15c> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 DB 196,226,61,168,205 ; vfmadd213ps %ymm5,%ymm8,%ymm1 @@ -316,7 +316,7 @@ _sk_modulate_hsw LABEL PROC PUBLIC _sk_multiply_hsw _sk_multiply_hsw LABEL PROC - DB 196,98,125,24,5,99,72,0,0 ; vbroadcastss 0x4863(%rip),%ymm8 # 4c94 <_sk_callback_hsw+0x15e> + DB 196,98,125,24,5,147,72,0,0 ; vbroadcastss 0x4893(%rip),%ymm8 # 4cc4 <_sk_callback_hsw+0x160> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -358,7 +358,7 @@ _sk_screen_hsw LABEL PROC PUBLIC _sk_xor__hsw _sk_xor__hsw LABEL PROC - DB 196,98,125,24,5,222,71,0,0 ; vbroadcastss 0x47de(%rip),%ymm8 # 4c98 <_sk_callback_hsw+0x162> + DB 196,98,125,24,5,14,72,0,0 ; vbroadcastss 0x480e(%rip),%ymm8 # 4cc8 <_sk_callback_hsw+0x164> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -390,7 +390,7 @@ _sk_darken_hsw LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,102,71,0,0 ; vbroadcastss 0x4766(%rip),%ymm8 # 4c9c <_sk_callback_hsw+0x166> + DB 196,98,125,24,5,150,71,0,0 ; vbroadcastss 0x4796(%rip),%ymm8 # 4ccc <_sk_callback_hsw+0x168> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -413,7 +413,7 @@ _sk_lighten_hsw LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,21,71,0,0 ; vbroadcastss 0x4715(%rip),%ymm8 # 4ca0 <_sk_callback_hsw+0x16a> + DB 196,98,125,24,5,69,71,0,0 ; vbroadcastss 0x4745(%rip),%ymm8 # 4cd0 <_sk_callback_hsw+0x16c> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -439,7 +439,7 @@ _sk_difference_hsw LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,184,70,0,0 ; vbroadcastss 0x46b8(%rip),%ymm8 # 4ca4 <_sk_callback_hsw+0x16e> + DB 196,98,125,24,5,232,70,0,0 ; vbroadcastss 0x46e8(%rip),%ymm8 # 4cd4 <_sk_callback_hsw+0x170> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -459,7 +459,7 @@ _sk_exclusion_hsw LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,118,70,0,0 ; vbroadcastss 0x4676(%rip),%ymm8 # 4ca8 <_sk_callback_hsw+0x172> + DB 196,98,125,24,5,166,70,0,0 ; vbroadcastss 0x46a6(%rip),%ymm8 # 4cd8 <_sk_callback_hsw+0x174> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -467,7 +467,7 @@ _sk_exclusion_hsw LABEL PROC PUBLIC _sk_colorburn_hsw _sk_colorburn_hsw LABEL PROC - DB 196,98,125,24,5,100,70,0,0 ; vbroadcastss 0x4664(%rip),%ymm8 # 4cac <_sk_callback_hsw+0x176> + DB 196,98,125,24,5,148,70,0,0 ; vbroadcastss 0x4694(%rip),%ymm8 # 4cdc <_sk_callback_hsw+0x178> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 @@ -523,7 +523,7 @@ _sk_colorburn_hsw LABEL PROC PUBLIC _sk_colordodge_hsw _sk_colordodge_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,111,69,0,0 ; vbroadcastss 0x456f(%rip),%ymm9 # 4cb0 <_sk_callback_hsw+0x17a> + DB 196,98,125,24,13,159,69,0,0 ; vbroadcastss 0x459f(%rip),%ymm9 # 4ce0 <_sk_callback_hsw+0x17c> DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9 @@ -574,7 +574,7 @@ _sk_colordodge_hsw LABEL PROC PUBLIC _sk_hardlight_hsw _sk_hardlight_hsw LABEL PROC - DB 196,98,125,24,5,144,68,0,0 ; vbroadcastss 0x4490(%rip),%ymm8 # 4cb4 <_sk_callback_hsw+0x17e> + DB 196,98,125,24,5,192,68,0,0 ; vbroadcastss 0x44c0(%rip),%ymm8 # 4ce4 <_sk_callback_hsw+0x180> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -623,7 +623,7 @@ _sk_hardlight_hsw LABEL PROC PUBLIC _sk_overlay_hsw _sk_overlay_hsw LABEL PROC - DB 196,98,125,24,5,200,67,0,0 ; vbroadcastss 0x43c8(%rip),%ymm8 # 4cb8 <_sk_callback_hsw+0x182> + DB 196,98,125,24,5,248,67,0,0 ; vbroadcastss 0x43f8(%rip),%ymm8 # 4ce8 <_sk_callback_hsw+0x184> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -683,10 +683,10 @@ _sk_softlight_hsw LABEL PROC DB 196,65,20,88,197 ; vaddps %ymm13,%ymm13,%ymm8 DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,66,61,168,192 ; vfmadd213ps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,29,207,66,0,0 ; vbroadcastss 0x42cf(%rip),%ymm11 # 4cc0 <_sk_callback_hsw+0x18a> + DB 196,98,125,24,29,255,66,0,0 ; vbroadcastss 0x42ff(%rip),%ymm11 # 4cf0 <_sk_callback_hsw+0x18c> DB 196,65,20,88,227 ; vaddps %ymm11,%ymm13,%ymm12 DB 196,65,28,89,192 ; vmulps %ymm8,%ymm12,%ymm8 - DB 196,98,125,24,37,192,66,0,0 ; vbroadcastss 0x42c0(%rip),%ymm12 # 4cc4 <_sk_callback_hsw+0x18e> + DB 196,98,125,24,37,240,66,0,0 ; vbroadcastss 0x42f0(%rip),%ymm12 # 4cf4 <_sk_callback_hsw+0x190> DB 196,66,21,184,196 ; vfmadd231ps %ymm12,%ymm13,%ymm8 DB 196,65,124,82,245 ; vrsqrtps %ymm13,%ymm14 DB 196,65,124,83,246 ; vrcpps %ymm14,%ymm14 @@ -696,7 +696,7 @@ _sk_softlight_hsw LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,131,66,0,0 ; vbroadcastss 0x4283(%rip),%ymm8 # 4cbc <_sk_callback_hsw+0x186> + DB 196,98,125,24,5,179,66,0,0 ; vbroadcastss 0x42b3(%rip),%ymm8 # 4cec <_sk_callback_hsw+0x188> DB 196,65,60,92,237 ; vsubps %ymm13,%ymm8,%ymm13 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,98,125,168,235 ; vfmadd213ps %ymm3,%ymm0,%ymm13 @@ -809,11 +809,11 @@ _sk_hue_hsw LABEL PROC DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10 DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10 DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - DB 196,98,125,24,53,130,64,0,0 ; vbroadcastss 0x4082(%rip),%ymm14 # 4cc8 <_sk_callback_hsw+0x192> - DB 196,98,125,24,61,125,64,0,0 ; vbroadcastss 0x407d(%rip),%ymm15 # 4ccc <_sk_callback_hsw+0x196> + DB 196,98,125,24,53,178,64,0,0 ; vbroadcastss 0x40b2(%rip),%ymm14 # 4cf8 <_sk_callback_hsw+0x194> + DB 196,98,125,24,61,173,64,0,0 ; vbroadcastss 0x40ad(%rip),%ymm15 # 4cfc <_sk_callback_hsw+0x198> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13 - DB 196,226,125,24,5,110,64,0,0 ; vbroadcastss 0x406e(%rip),%ymm0 # 4cd0 <_sk_callback_hsw+0x19a> + DB 196,226,125,24,5,158,64,0,0 ; vbroadcastss 0x409e(%rip),%ymm0 # 4d00 <_sk_callback_hsw+0x19c> DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13 DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10 DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -868,7 +868,7 @@ _sk_hue_hsw LABEL PROC DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9 DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,13,91,63,0,0 ; vbroadcastss 0x3f5b(%rip),%ymm1 # 4cd4 <_sk_callback_hsw+0x19e> + DB 196,226,125,24,13,139,63,0,0 ; vbroadcastss 0x3f8b(%rip),%ymm1 # 4d04 <_sk_callback_hsw+0x1a0> DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11 @@ -922,11 +922,11 @@ _sk_saturation_hsw LABEL PROC DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10 DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10 DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - DB 196,98,125,24,53,108,62,0,0 ; vbroadcastss 0x3e6c(%rip),%ymm14 # 4cd8 <_sk_callback_hsw+0x1a2> - DB 196,98,125,24,61,103,62,0,0 ; vbroadcastss 0x3e67(%rip),%ymm15 # 4cdc <_sk_callback_hsw+0x1a6> + DB 196,98,125,24,53,156,62,0,0 ; vbroadcastss 0x3e9c(%rip),%ymm14 # 4d08 <_sk_callback_hsw+0x1a4> + DB 196,98,125,24,61,151,62,0,0 ; vbroadcastss 0x3e97(%rip),%ymm15 # 4d0c <_sk_callback_hsw+0x1a8> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13 - DB 196,226,125,24,5,88,62,0,0 ; vbroadcastss 0x3e58(%rip),%ymm0 # 4ce0 <_sk_callback_hsw+0x1aa> + DB 196,226,125,24,5,136,62,0,0 ; vbroadcastss 0x3e88(%rip),%ymm0 # 4d10 <_sk_callback_hsw+0x1ac> DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13 DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10 DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -981,7 +981,7 @@ _sk_saturation_hsw LABEL PROC DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9 DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,13,69,61,0,0 ; vbroadcastss 0x3d45(%rip),%ymm1 # 4ce4 <_sk_callback_hsw+0x1ae> + DB 196,226,125,24,13,117,61,0,0 ; vbroadcastss 0x3d75(%rip),%ymm1 # 4d14 <_sk_callback_hsw+0x1b0> DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11 @@ -1009,11 +1009,11 @@ _sk_color_hsw LABEL PROC DB 197,108,89,199 ; vmulps %ymm7,%ymm2,%ymm8 DB 197,116,89,215 ; vmulps %ymm7,%ymm1,%ymm10 DB 197,52,89,223 ; vmulps %ymm7,%ymm9,%ymm11 - DB 196,98,125,24,45,216,60,0,0 ; vbroadcastss 0x3cd8(%rip),%ymm13 # 4ce8 <_sk_callback_hsw+0x1b2> - DB 196,98,125,24,53,211,60,0,0 ; vbroadcastss 0x3cd3(%rip),%ymm14 # 4cec <_sk_callback_hsw+0x1b6> + DB 196,98,125,24,45,8,61,0,0 ; vbroadcastss 0x3d08(%rip),%ymm13 # 4d18 <_sk_callback_hsw+0x1b4> + DB 196,98,125,24,53,3,61,0,0 ; vbroadcastss 0x3d03(%rip),%ymm14 # 4d1c <_sk_callback_hsw+0x1b8> DB 196,65,84,89,230 ; vmulps %ymm14,%ymm5,%ymm12 DB 196,66,93,184,229 ; vfmadd231ps %ymm13,%ymm4,%ymm12 - DB 196,98,125,24,61,196,60,0,0 ; vbroadcastss 0x3cc4(%rip),%ymm15 # 4cf0 <_sk_callback_hsw+0x1ba> + DB 196,98,125,24,61,244,60,0,0 ; vbroadcastss 0x3cf4(%rip),%ymm15 # 4d20 <_sk_callback_hsw+0x1bc> DB 196,66,77,184,231 ; vfmadd231ps %ymm15,%ymm6,%ymm12 DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9 DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -1069,7 +1069,7 @@ _sk_color_hsw LABEL PROC DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1 DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8 DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9 - DB 196,226,125,24,5,166,59,0,0 ; vbroadcastss 0x3ba6(%rip),%ymm0 # 4cf4 <_sk_callback_hsw+0x1be> + DB 196,226,125,24,5,214,59,0,0 ; vbroadcastss 0x3bd6(%rip),%ymm0 # 4d24 <_sk_callback_hsw+0x1c0> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -1097,11 +1097,11 @@ _sk_luminosity_hsw LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,213 ; vmulps %ymm5,%ymm3,%ymm10 DB 197,100,89,222 ; vmulps %ymm6,%ymm3,%ymm11 - DB 196,98,125,24,45,57,59,0,0 ; vbroadcastss 0x3b39(%rip),%ymm13 # 4cf8 <_sk_callback_hsw+0x1c2> - DB 196,98,125,24,53,52,59,0,0 ; vbroadcastss 0x3b34(%rip),%ymm14 # 4cfc <_sk_callback_hsw+0x1c6> + DB 196,98,125,24,45,105,59,0,0 ; vbroadcastss 0x3b69(%rip),%ymm13 # 4d28 <_sk_callback_hsw+0x1c4> + DB 196,98,125,24,53,100,59,0,0 ; vbroadcastss 0x3b64(%rip),%ymm14 # 4d2c <_sk_callback_hsw+0x1c8> DB 196,65,116,89,230 ; vmulps %ymm14,%ymm1,%ymm12 DB 196,66,109,184,229 ; vfmadd231ps %ymm13,%ymm2,%ymm12 - DB 196,98,125,24,61,37,59,0,0 ; vbroadcastss 0x3b25(%rip),%ymm15 # 4d00 <_sk_callback_hsw+0x1ca> + DB 196,98,125,24,61,85,59,0,0 ; vbroadcastss 0x3b55(%rip),%ymm15 # 4d30 <_sk_callback_hsw+0x1cc> DB 196,66,53,184,231 ; vfmadd231ps %ymm15,%ymm9,%ymm12 DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9 DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -1157,7 +1157,7 @@ _sk_luminosity_hsw LABEL PROC DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1 DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8 DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9 - DB 196,226,125,24,5,7,58,0,0 ; vbroadcastss 0x3a07(%rip),%ymm0 # 4d04 <_sk_callback_hsw+0x1ce> + DB 196,226,125,24,5,55,58,0,0 ; vbroadcastss 0x3a37(%rip),%ymm0 # 4d34 <_sk_callback_hsw+0x1d0> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -1184,17 +1184,17 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 15,133,180,0,0,0 ; jne 140f <_sk_srcover_rgba_8888_hsw+0xcd> DB 196,193,124,16,58 ; vmovups (%r10),%ymm7 - DB 197,196,84,37,24,61,0,0 ; vandps 0x3d18(%rip),%ymm7,%ymm4 # 5080 <_sk_callback_hsw+0x54a> + DB 197,196,84,37,56,61,0,0 ; vandps 0x3d38(%rip),%ymm7,%ymm4 # 50a0 <_sk_callback_hsw+0x53c> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,69,0,45,43,61,0,0 ; vpshufb 0x3d2b(%rip),%ymm7,%ymm5 # 50a0 <_sk_callback_hsw+0x56a> + DB 196,226,69,0,45,75,61,0,0 ; vpshufb 0x3d4b(%rip),%ymm7,%ymm5 # 50c0 <_sk_callback_hsw+0x55c> DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,69,0,53,62,61,0,0 ; vpshufb 0x3d3e(%rip),%ymm7,%ymm6 # 50c0 <_sk_callback_hsw+0x58a> + DB 196,226,69,0,53,94,61,0,0 ; vpshufb 0x3d5e(%rip),%ymm7,%ymm6 # 50e0 <_sk_callback_hsw+0x57c> DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 DB 197,197,114,215,24 ; vpsrld $0x18,%ymm7,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,112,57,0,0 ; vbroadcastss 0x3970(%rip),%ymm8 # 4d08 <_sk_callback_hsw+0x1d2> + DB 196,98,125,24,5,160,57,0,0 ; vbroadcastss 0x39a0(%rip),%ymm8 # 4d38 <_sk_callback_hsw+0x1d4> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,103,57,0,0 ; vbroadcastss 0x3967(%rip),%ymm9 # 4d0c <_sk_callback_hsw+0x1d6> + DB 196,98,125,24,13,151,57,0,0 ; vbroadcastss 0x3997(%rip),%ymm9 # 4d3c <_sk_callback_hsw+0x1d8> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 @@ -1250,7 +1250,7 @@ _sk_clamp_0_hsw LABEL PROC PUBLIC _sk_clamp_1_hsw _sk_clamp_1_hsw LABEL PROC - DB 196,98,125,24,5,140,56,0,0 ; vbroadcastss 0x388c(%rip),%ymm8 # 4d10 <_sk_callback_hsw+0x1da> + DB 196,98,125,24,5,188,56,0,0 ; vbroadcastss 0x38bc(%rip),%ymm8 # 4d40 <_sk_callback_hsw+0x1dc> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 @@ -1260,7 +1260,7 @@ _sk_clamp_1_hsw LABEL PROC PUBLIC _sk_clamp_a_hsw _sk_clamp_a_hsw LABEL PROC - DB 196,98,125,24,5,111,56,0,0 ; vbroadcastss 0x386f(%rip),%ymm8 # 4d14 <_sk_callback_hsw+0x1de> + DB 196,98,125,24,5,159,56,0,0 ; vbroadcastss 0x389f(%rip),%ymm8 # 4d44 <_sk_callback_hsw+0x1e0> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0 DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1 @@ -1332,7 +1332,7 @@ PUBLIC _sk_unpremul_hsw _sk_unpremul_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,183,55,0,0 ; vbroadcastss 0x37b7(%rip),%ymm10 # 4d18 <_sk_callback_hsw+0x1e2> + DB 196,98,125,24,21,231,55,0,0 ; vbroadcastss 0x37e7(%rip),%ymm10 # 4d48 <_sk_callback_hsw+0x1e4> DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10 DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 @@ -1343,16 +1343,16 @@ _sk_unpremul_hsw LABEL PROC PUBLIC _sk_from_srgb_hsw _sk_from_srgb_hsw LABEL PROC - DB 196,98,125,24,5,152,55,0,0 ; vbroadcastss 0x3798(%rip),%ymm8 # 4d1c <_sk_callback_hsw+0x1e6> + DB 196,98,125,24,5,200,55,0,0 ; vbroadcastss 0x37c8(%rip),%ymm8 # 4d4c <_sk_callback_hsw+0x1e8> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,138,55,0,0 ; vbroadcastss 0x378a(%rip),%ymm11 # 4d20 <_sk_callback_hsw+0x1ea> - DB 196,98,125,24,37,133,55,0,0 ; vbroadcastss 0x3785(%rip),%ymm12 # 4d24 <_sk_callback_hsw+0x1ee> + DB 196,98,125,24,29,186,55,0,0 ; vbroadcastss 0x37ba(%rip),%ymm11 # 4d50 <_sk_callback_hsw+0x1ec> + DB 196,98,125,24,37,181,55,0,0 ; vbroadcastss 0x37b5(%rip),%ymm12 # 4d54 <_sk_callback_hsw+0x1f0> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13 - DB 196,98,125,24,53,118,55,0,0 ; vbroadcastss 0x3776(%rip),%ymm14 # 4d28 <_sk_callback_hsw+0x1f2> + DB 196,98,125,24,53,166,55,0,0 ; vbroadcastss 0x37a6(%rip),%ymm14 # 4d58 <_sk_callback_hsw+0x1f4> DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13 - DB 196,98,125,24,21,108,55,0,0 ; vbroadcastss 0x376c(%rip),%ymm10 # 4d2c <_sk_callback_hsw+0x1f6> + DB 196,98,125,24,21,156,55,0,0 ; vbroadcastss 0x379c(%rip),%ymm10 # 4d5c <_sk_callback_hsw+0x1f8> DB 196,193,124,194,194,1 ; vcmpltps %ymm10,%ymm0,%ymm0 DB 196,195,21,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm13,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -1374,19 +1374,19 @@ _sk_from_srgb_hsw LABEL PROC PUBLIC _sk_to_srgb_hsw _sk_to_srgb_hsw LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,16,55,0,0 ; vbroadcastss 0x3710(%rip),%ymm8 # 4d30 <_sk_callback_hsw+0x1fa> + DB 196,98,125,24,5,64,55,0,0 ; vbroadcastss 0x3740(%rip),%ymm8 # 4d60 <_sk_callback_hsw+0x1fc> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,6,55,0,0 ; vbroadcastss 0x3706(%rip),%ymm11 # 4d34 <_sk_callback_hsw+0x1fe> - DB 196,98,125,24,37,1,55,0,0 ; vbroadcastss 0x3701(%rip),%ymm12 # 4d38 <_sk_callback_hsw+0x202> + DB 196,98,125,24,29,54,55,0,0 ; vbroadcastss 0x3736(%rip),%ymm11 # 4d64 <_sk_callback_hsw+0x200> + DB 196,98,125,24,37,49,55,0,0 ; vbroadcastss 0x3731(%rip),%ymm12 # 4d68 <_sk_callback_hsw+0x204> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,53,168,235 ; vfmadd213ps %ymm11,%ymm9,%ymm13 - DB 196,98,125,24,53,242,54,0,0 ; vbroadcastss 0x36f2(%rip),%ymm14 # 4d3c <_sk_callback_hsw+0x206> + DB 196,98,125,24,53,34,55,0,0 ; vbroadcastss 0x3722(%rip),%ymm14 # 4d6c <_sk_callback_hsw+0x208> DB 196,66,53,168,238 ; vfmadd213ps %ymm14,%ymm9,%ymm13 - DB 196,98,125,24,61,232,54,0,0 ; vbroadcastss 0x36e8(%rip),%ymm15 # 4d40 <_sk_callback_hsw+0x20a> + DB 196,98,125,24,61,24,55,0,0 ; vbroadcastss 0x3718(%rip),%ymm15 # 4d70 <_sk_callback_hsw+0x20c> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,20,89,201 ; vmulps %ymm9,%ymm13,%ymm9 - DB 196,98,125,24,45,212,54,0,0 ; vbroadcastss 0x36d4(%rip),%ymm13 # 4d44 <_sk_callback_hsw+0x20e> + DB 196,98,125,24,45,4,55,0,0 ; vbroadcastss 0x3704(%rip),%ymm13 # 4d74 <_sk_callback_hsw+0x210> DB 196,193,124,194,197,1 ; vcmpltps %ymm13,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -1418,26 +1418,26 @@ _sk_rgb_to_hsl_hsw LABEL PROC DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9 DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,73,54,0,0 ; vbroadcastss 0x3649(%rip),%ymm11 # 4d48 <_sk_callback_hsw+0x212> + DB 196,98,125,24,29,121,54,0,0 ; vbroadcastss 0x3679(%rip),%ymm11 # 4d78 <_sk_callback_hsw+0x214> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 197,116,194,234,1 ; vcmpltps %ymm2,%ymm1,%ymm13 - DB 196,98,125,24,53,54,54,0,0 ; vbroadcastss 0x3636(%rip),%ymm14 # 4d4c <_sk_callback_hsw+0x216> + DB 196,98,125,24,53,102,54,0,0 ; vbroadcastss 0x3666(%rip),%ymm14 # 4d7c <_sk_callback_hsw+0x218> DB 196,65,4,87,255 ; vxorps %ymm15,%ymm15,%ymm15 DB 196,67,5,74,238,208 ; vblendvps %ymm13,%ymm14,%ymm15,%ymm13 DB 196,66,37,168,229 ; vfmadd213ps %ymm13,%ymm11,%ymm12 DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 DB 197,124,92,233 ; vsubps %ymm1,%ymm0,%ymm13 - DB 196,98,125,24,53,29,54,0,0 ; vbroadcastss 0x361d(%rip),%ymm14 # 4d54 <_sk_callback_hsw+0x21e> + DB 196,98,125,24,53,77,54,0,0 ; vbroadcastss 0x364d(%rip),%ymm14 # 4d84 <_sk_callback_hsw+0x220> DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13 - DB 196,98,125,24,53,11,54,0,0 ; vbroadcastss 0x360b(%rip),%ymm14 # 4d50 <_sk_callback_hsw+0x21a> + DB 196,98,125,24,53,59,54,0,0 ; vbroadcastss 0x363b(%rip),%ymm14 # 4d80 <_sk_callback_hsw+0x21c> DB 196,194,37,168,214 ; vfmadd213ps %ymm14,%ymm11,%ymm2 DB 197,188,194,201,0 ; vcmpeqps %ymm1,%ymm8,%ymm1 DB 196,227,21,74,202,16 ; vblendvps %ymm1,%ymm2,%ymm13,%ymm1 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,195,117,74,196,0 ; vblendvps %ymm0,%ymm12,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,29,238,53,0,0 ; vbroadcastss 0x35ee(%rip),%ymm11 # 4d5c <_sk_callback_hsw+0x226> + DB 196,98,125,24,29,30,54,0,0 ; vbroadcastss 0x361e(%rip),%ymm11 # 4d8c <_sk_callback_hsw+0x228> DB 196,193,116,89,211 ; vmulps %ymm11,%ymm1,%ymm2 DB 197,36,194,218,1 ; vcmpltps %ymm2,%ymm11,%ymm11 DB 196,65,12,92,224 ; vsubps %ymm8,%ymm14,%ymm12 @@ -1447,7 +1447,7 @@ _sk_rgb_to_hsl_hsw LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,199,128 ; vblendvps %ymm8,%ymm15,%ymm0,%ymm0 DB 196,195,117,74,207,128 ; vblendvps %ymm8,%ymm15,%ymm1,%ymm1 - DB 196,98,125,24,5,177,53,0,0 ; vbroadcastss 0x35b1(%rip),%ymm8 # 4d58 <_sk_callback_hsw+0x222> + DB 196,98,125,24,5,225,53,0,0 ; vbroadcastss 0x35e1(%rip),%ymm8 # 4d88 <_sk_callback_hsw+0x224> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1462,30 +1462,30 @@ _sk_hsl_to_rgb_hsw LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,233 ; vmovaps %ymm1,%ymm5 DB 197,252,40,224 ; vmovaps %ymm0,%ymm4 - DB 196,98,125,24,5,120,53,0,0 ; vbroadcastss 0x3578(%rip),%ymm8 # 4d60 <_sk_callback_hsw+0x22a> + DB 196,98,125,24,5,168,53,0,0 ; vbroadcastss 0x35a8(%rip),%ymm8 # 4d90 <_sk_callback_hsw+0x22c> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,84,89,210 ; vmulps %ymm2,%ymm5,%ymm10 DB 196,65,84,92,218 ; vsubps %ymm10,%ymm5,%ymm11 DB 196,67,45,74,203,144 ; vblendvps %ymm9,%ymm11,%ymm10,%ymm9 DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 - DB 196,98,125,24,13,91,53,0,0 ; vbroadcastss 0x355b(%rip),%ymm9 # 4d64 <_sk_callback_hsw+0x22e> + DB 196,98,125,24,13,139,53,0,0 ; vbroadcastss 0x358b(%rip),%ymm9 # 4d94 <_sk_callback_hsw+0x230> DB 196,66,109,170,202 ; vfmsub213ps %ymm10,%ymm2,%ymm9 - DB 196,98,125,24,29,81,53,0,0 ; vbroadcastss 0x3551(%rip),%ymm11 # 4d68 <_sk_callback_hsw+0x232> + DB 196,98,125,24,29,129,53,0,0 ; vbroadcastss 0x3581(%rip),%ymm11 # 4d98 <_sk_callback_hsw+0x234> DB 196,65,92,88,219 ; vaddps %ymm11,%ymm4,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,45,59,53,0,0 ; vbroadcastss 0x353b(%rip),%ymm13 # 4d70 <_sk_callback_hsw+0x23a> + DB 196,98,125,24,45,107,53,0,0 ; vbroadcastss 0x356b(%rip),%ymm13 # 4da0 <_sk_callback_hsw+0x23c> DB 196,193,4,89,197 ; vmulps %ymm13,%ymm15,%ymm0 - DB 196,98,125,24,53,49,53,0,0 ; vbroadcastss 0x3531(%rip),%ymm14 # 4d74 <_sk_callback_hsw+0x23e> + DB 196,98,125,24,53,97,53,0,0 ; vbroadcastss 0x3561(%rip),%ymm14 # 4da4 <_sk_callback_hsw+0x240> DB 197,12,92,224 ; vsubps %ymm0,%ymm14,%ymm12 DB 196,66,37,168,225 ; vfmadd213ps %ymm9,%ymm11,%ymm12 - DB 196,226,125,24,29,23,53,0,0 ; vbroadcastss 0x3517(%rip),%ymm3 # 4d6c <_sk_callback_hsw+0x236> + DB 196,226,125,24,29,71,53,0,0 ; vbroadcastss 0x3547(%rip),%ymm3 # 4d9c <_sk_callback_hsw+0x238> DB 196,193,100,194,255,2 ; vcmpleps %ymm15,%ymm3,%ymm7 DB 196,195,29,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm12,%ymm7 DB 196,65,60,194,231,2 ; vcmpleps %ymm15,%ymm8,%ymm12 DB 196,227,45,74,255,192 ; vblendvps %ymm12,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,37,2,53,0,0 ; vbroadcastss 0x3502(%rip),%ymm12 # 4d78 <_sk_callback_hsw+0x242> + DB 196,98,125,24,37,50,53,0,0 ; vbroadcastss 0x3532(%rip),%ymm12 # 4da8 <_sk_callback_hsw+0x244> DB 196,65,28,194,255,2 ; vcmpleps %ymm15,%ymm12,%ymm15 DB 196,194,37,168,193 ; vfmadd213ps %ymm9,%ymm11,%ymm0 DB 196,99,125,74,255,240 ; vblendvps %ymm15,%ymm7,%ymm0,%ymm15 @@ -1501,7 +1501,7 @@ _sk_hsl_to_rgb_hsw LABEL PROC DB 197,156,194,192,2 ; vcmpleps %ymm0,%ymm12,%ymm0 DB 196,194,37,168,249 ; vfmadd213ps %ymm9,%ymm11,%ymm7 DB 196,227,69,74,201,0 ; vblendvps %ymm0,%ymm1,%ymm7,%ymm1 - DB 196,226,125,24,5,174,52,0,0 ; vbroadcastss 0x34ae(%rip),%ymm0 # 4d7c <_sk_callback_hsw+0x246> + DB 196,226,125,24,5,222,52,0,0 ; vbroadcastss 0x34de(%rip),%ymm0 # 4dac <_sk_callback_hsw+0x248> DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 DB 196,227,125,8,224,1 ; vroundps $0x1,%ymm0,%ymm4 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 @@ -1547,10 +1547,10 @@ _sk_scale_u8_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,58 ; jne 19b5 <_sk_scale_u8_hsw+0x44> DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8 - DB 197,57,219,5,215,56,0,0 ; vpand 0x38d7(%rip),%xmm8,%xmm8 # 5260 <_sk_callback_hsw+0x72a> + DB 197,57,219,5,247,56,0,0 ; vpand 0x38f7(%rip),%xmm8,%xmm8 # 5280 <_sk_callback_hsw+0x71c> DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,228,51,0,0 ; vbroadcastss 0x33e4(%rip),%ymm9 # 4d80 <_sk_callback_hsw+0x24a> + DB 196,98,125,24,13,20,52,0,0 ; vbroadcastss 0x3414(%rip),%ymm9 # 4db0 <_sk_callback_hsw+0x24c> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -1628,10 +1628,10 @@ _sk_lerp_u8_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,78 ; jne 1ad7 <_sk_lerp_u8_hsw+0x58> DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8 - DB 197,57,219,5,217,55,0,0 ; vpand 0x37d9(%rip),%xmm8,%xmm8 # 5270 <_sk_callback_hsw+0x73a> + DB 197,57,219,5,249,55,0,0 ; vpand 0x37f9(%rip),%xmm8,%xmm8 # 5290 <_sk_callback_hsw+0x72c> DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,218,50,0,0 ; vbroadcastss 0x32da(%rip),%ymm9 # 4d84 <_sk_callback_hsw+0x24e> + DB 196,98,125,24,13,10,51,0,0 ; vbroadcastss 0x330a(%rip),%ymm9 # 4db4 <_sk_callback_hsw+0x250> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -1700,20 +1700,20 @@ _sk_lerp_565_hsw LABEL PROC DB 15,133,169,0,0,0 ; jne 1c2b <_sk_lerp_565_hsw+0xb7> DB 196,65,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm8 DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 - DB 196,98,125,88,13,242,49,0,0 ; vpbroadcastd 0x31f2(%rip),%ymm9 # 4d88 <_sk_callback_hsw+0x252> + DB 196,98,125,88,13,34,50,0,0 ; vpbroadcastd 0x3222(%rip),%ymm9 # 4db8 <_sk_callback_hsw+0x254> DB 196,65,61,219,201 ; vpand %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,227,49,0,0 ; vbroadcastss 0x31e3(%rip),%ymm10 # 4d8c <_sk_callback_hsw+0x256> + DB 196,98,125,24,21,19,50,0,0 ; vbroadcastss 0x3213(%rip),%ymm10 # 4dbc <_sk_callback_hsw+0x258> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,88,21,217,49,0,0 ; vpbroadcastd 0x31d9(%rip),%ymm10 # 4d90 <_sk_callback_hsw+0x25a> + DB 196,98,125,88,21,9,50,0,0 ; vpbroadcastd 0x3209(%rip),%ymm10 # 4dc0 <_sk_callback_hsw+0x25c> DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,202,49,0,0 ; vbroadcastss 0x31ca(%rip),%ymm11 # 4d94 <_sk_callback_hsw+0x25e> + DB 196,98,125,24,29,250,49,0,0 ; vbroadcastss 0x31fa(%rip),%ymm11 # 4dc4 <_sk_callback_hsw+0x260> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,88,29,192,49,0,0 ; vpbroadcastd 0x31c0(%rip),%ymm11 # 4d98 <_sk_callback_hsw+0x262> + DB 196,98,125,88,29,240,49,0,0 ; vpbroadcastd 0x31f0(%rip),%ymm11 # 4dc8 <_sk_callback_hsw+0x264> DB 196,65,61,219,195 ; vpand %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,177,49,0,0 ; vbroadcastss 0x31b1(%rip),%ymm11 # 4d9c <_sk_callback_hsw+0x266> + DB 196,98,125,24,29,225,49,0,0 ; vbroadcastss 0x31e1(%rip),%ymm11 # 4dcc <_sk_callback_hsw+0x268> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,226,53,168,196 ; vfmadd213ps %ymm4,%ymm9,%ymm0 @@ -1754,7 +1754,7 @@ _sk_lerp_565_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2001ca4 <_sk_callback_hsw+0xffffffffe1ffd16e> + DB 233,255,255,255,225 ; jmpq ffffffffe2001ca4 <_sk_callback_hsw+0xffffffffe1ffd140> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -1781,21 +1781,21 @@ _sk_load_tables_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,105 ; jne 1d36 <_sk_load_tables_hsw+0x7e> DB 196,193,124,16,26 ; vmovups (%r10),%ymm3 - DB 197,228,84,13,6,52,0,0 ; vandps 0x3406(%rip),%ymm3,%ymm1 # 50e0 <_sk_callback_hsw+0x5aa> + DB 197,228,84,13,38,52,0,0 ; vandps 0x3426(%rip),%ymm3,%ymm1 # 5100 <_sk_callback_hsw+0x59c> DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 DB 72,139,72,8 ; mov 0x8(%rax),%rcx DB 76,139,80,16 ; mov 0x10(%rax),%r10 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 DB 196,226,109,146,4,137 ; vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0 - DB 196,226,101,0,21,6,52,0,0 ; vpshufb 0x3406(%rip),%ymm3,%ymm2 # 5100 <_sk_callback_hsw+0x5ca> + DB 196,226,101,0,21,38,52,0,0 ; vpshufb 0x3426(%rip),%ymm3,%ymm2 # 5120 <_sk_callback_hsw+0x5bc> DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 DB 196,194,53,146,12,146 ; vgatherdps %ymm9,(%r10,%ymm2,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,98,101,0,13,14,52,0,0 ; vpshufb 0x340e(%rip),%ymm3,%ymm9 # 5120 <_sk_callback_hsw+0x5ea> + DB 196,98,101,0,13,46,52,0,0 ; vpshufb 0x342e(%rip),%ymm3,%ymm9 # 5140 <_sk_callback_hsw+0x5dc> 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 196,98,125,24,5,118,48,0,0 ; vbroadcastss 0x3076(%rip),%ymm8 # 4da0 <_sk_callback_hsw+0x26a> + DB 196,98,125,24,5,166,48,0,0 ; vbroadcastss 0x30a6(%rip),%ymm8 # 4dd0 <_sk_callback_hsw+0x26c> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,201 ; mov %r9,%rcx @@ -1832,7 +1832,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC DB 197,185,108,200 ; vpunpcklqdq %xmm0,%xmm8,%xmm1 DB 197,185,109,208 ; vpunpckhqdq %xmm0,%xmm8,%xmm2 DB 197,49,108,195 ; vpunpcklqdq %xmm3,%xmm9,%xmm8 - DB 197,121,111,21,186,52,0,0 ; vmovdqa 0x34ba(%rip),%xmm10 # 5280 <_sk_callback_hsw+0x74a> + DB 197,121,111,21,218,52,0,0 ; vmovdqa 0x34da(%rip),%xmm10 # 52a0 <_sk_callback_hsw+0x73c> 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 @@ -1854,7 +1854,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3 DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,111,47,0,0 ; vbroadcastss 0x2f6f(%rip),%ymm8 # 4da4 <_sk_callback_hsw+0x26e> + DB 196,98,125,24,5,159,47,0,0 ; vbroadcastss 0x2f9f(%rip),%ymm8 # 4dd4 <_sk_callback_hsw+0x270> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1912,7 +1912,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3 DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2 DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8 - DB 197,121,111,13,90,51,0,0 ; vmovdqa 0x335a(%rip),%xmm9 # 5290 <_sk_callback_hsw+0x75a> + DB 197,121,111,13,122,51,0,0 ; vmovdqa 0x337a(%rip),%xmm9 # 52b0 <_sk_callback_hsw+0x74c> 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 @@ -1929,7 +1929,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8 DB 196,162,101,146,20,128 ; vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,29,46,0,0 ; vbroadcastss 0x2e1d(%rip),%ymm3 # 4da8 <_sk_callback_hsw+0x272> + DB 196,226,125,24,29,77,46,0,0 ; vbroadcastss 0x2e4d(%rip),%ymm3 # 4dd8 <_sk_callback_hsw+0x274> DB 255,224 ; jmpq *%rax DB 196,129,121,110,4,81 ; vmovd (%r9,%r10,2),%xmm0 DB 196,129,121,196,68,81,4,2 ; vpinsrw $0x2,0x4(%r9,%r10,2),%xmm0,%xmm0 @@ -1968,7 +1968,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC PUBLIC _sk_byte_tables_hsw _sk_byte_tables_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,101,45,0,0 ; vbroadcastss 0x2d65(%rip),%ymm8 # 4dac <_sk_callback_hsw+0x276> + DB 196,98,125,24,5,149,45,0,0 ; vbroadcastss 0x2d95(%rip),%ymm8 # 4ddc <_sk_callback_hsw+0x278> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 DB 196,65,249,126,201 ; vmovq %xmm9,%r9 @@ -2090,7 +2090,7 @@ _sk_byte_tables_hsw LABEL PROC DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,10,43,0,0 ; vbroadcastss 0x2b0a(%rip),%ymm8 # 4db0 <_sk_callback_hsw+0x27a> + DB 196,98,125,24,5,58,43,0,0 ; vbroadcastss 0x2b3a(%rip),%ymm8 # 4de0 <_sk_callback_hsw+0x27c> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 @@ -2204,7 +2204,7 @@ _sk_byte_tables_rgb_hsw LABEL PROC DB 67,15,182,4,26 ; movzbl (%r10,%r11,1),%eax DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,246,40,0,0 ; vbroadcastss 0x28f6(%rip),%ymm8 # 4db4 <_sk_callback_hsw+0x27e> + DB 196,98,125,24,5,38,41,0,0 ; vbroadcastss 0x2926(%rip),%ymm8 # 4de4 <_sk_callback_hsw+0x280> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 @@ -2293,33 +2293,33 @@ _sk_parametric_r_hsw LABEL PROC DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10 DB 196,226,125,24,0 ; vbroadcastss (%rax),%ymm0 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,208,39,0,0 ; vbroadcastss 0x27d0(%rip),%ymm12 # 4db8 <_sk_callback_hsw+0x282> - DB 196,98,125,24,45,203,39,0,0 ; vbroadcastss 0x27cb(%rip),%ymm13 # 4dbc <_sk_callback_hsw+0x286> + DB 196,98,125,24,37,0,40,0,0 ; vbroadcastss 0x2800(%rip),%ymm12 # 4de8 <_sk_callback_hsw+0x284> + DB 196,98,125,24,45,251,39,0,0 ; vbroadcastss 0x27fb(%rip),%ymm13 # 4dec <_sk_callback_hsw+0x288> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,193,39,0,0 ; vbroadcastss 0x27c1(%rip),%ymm13 # 4dc0 <_sk_callback_hsw+0x28a> + DB 196,98,125,24,45,241,39,0,0 ; vbroadcastss 0x27f1(%rip),%ymm13 # 4df0 <_sk_callback_hsw+0x28c> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,183,39,0,0 ; vbroadcastss 0x27b7(%rip),%ymm13 # 4dc4 <_sk_callback_hsw+0x28e> + DB 196,98,125,24,45,231,39,0,0 ; vbroadcastss 0x27e7(%rip),%ymm13 # 4df4 <_sk_callback_hsw+0x290> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,173,39,0,0 ; vbroadcastss 0x27ad(%rip),%ymm11 # 4dc8 <_sk_callback_hsw+0x292> + DB 196,98,125,24,29,221,39,0,0 ; vbroadcastss 0x27dd(%rip),%ymm11 # 4df8 <_sk_callback_hsw+0x294> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,163,39,0,0 ; vbroadcastss 0x27a3(%rip),%ymm12 # 4dcc <_sk_callback_hsw+0x296> + DB 196,98,125,24,37,211,39,0,0 ; vbroadcastss 0x27d3(%rip),%ymm12 # 4dfc <_sk_callback_hsw+0x298> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,153,39,0,0 ; vbroadcastss 0x2799(%rip),%ymm12 # 4dd0 <_sk_callback_hsw+0x29a> + DB 196,98,125,24,37,201,39,0,0 ; vbroadcastss 0x27c9(%rip),%ymm12 # 4e00 <_sk_callback_hsw+0x29c> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,122,39,0,0 ; vbroadcastss 0x277a(%rip),%ymm11 # 4dd4 <_sk_callback_hsw+0x29e> + DB 196,98,125,24,29,170,39,0,0 ; vbroadcastss 0x27aa(%rip),%ymm11 # 4e04 <_sk_callback_hsw+0x2a0> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,112,39,0,0 ; vbroadcastss 0x2770(%rip),%ymm11 # 4dd8 <_sk_callback_hsw+0x2a2> + DB 196,98,125,24,29,160,39,0,0 ; vbroadcastss 0x27a0(%rip),%ymm11 # 4e08 <_sk_callback_hsw+0x2a4> DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11 - DB 196,226,125,24,5,102,39,0,0 ; vbroadcastss 0x2766(%rip),%ymm0 # 4ddc <_sk_callback_hsw+0x2a6> + DB 196,226,125,24,5,150,39,0,0 ; vbroadcastss 0x2796(%rip),%ymm0 # 4e0c <_sk_callback_hsw+0x2a8> DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,92,39,0,0 ; vbroadcastss 0x275c(%rip),%ymm10 # 4de0 <_sk_callback_hsw+0x2aa> + DB 196,98,125,24,21,140,39,0,0 ; vbroadcastss 0x278c(%rip),%ymm10 # 4e10 <_sk_callback_hsw+0x2ac> DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0 DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,21,79,39,0,0 ; vbroadcastss 0x274f(%rip),%ymm10 # 4de4 <_sk_callback_hsw+0x2ae> + DB 196,98,125,24,21,127,39,0,0 ; vbroadcastss 0x277f(%rip),%ymm10 # 4e14 <_sk_callback_hsw+0x2b0> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2327,7 +2327,7 @@ _sk_parametric_r_hsw LABEL PROC DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,38,39,0,0 ; vbroadcastss 0x2726(%rip),%ymm8 # 4de8 <_sk_callback_hsw+0x2b2> + DB 196,98,125,24,5,86,39,0,0 ; vbroadcastss 0x2756(%rip),%ymm8 # 4e18 <_sk_callback_hsw+0x2b4> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2345,33 +2345,33 @@ _sk_parametric_g_hsw LABEL PROC DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10 DB 196,226,125,24,8 ; vbroadcastss (%rax),%ymm1 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,222,38,0,0 ; vbroadcastss 0x26de(%rip),%ymm12 # 4dec <_sk_callback_hsw+0x2b6> - DB 196,98,125,24,45,217,38,0,0 ; vbroadcastss 0x26d9(%rip),%ymm13 # 4df0 <_sk_callback_hsw+0x2ba> + DB 196,98,125,24,37,14,39,0,0 ; vbroadcastss 0x270e(%rip),%ymm12 # 4e1c <_sk_callback_hsw+0x2b8> + DB 196,98,125,24,45,9,39,0,0 ; vbroadcastss 0x2709(%rip),%ymm13 # 4e20 <_sk_callback_hsw+0x2bc> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,207,38,0,0 ; vbroadcastss 0x26cf(%rip),%ymm13 # 4df4 <_sk_callback_hsw+0x2be> + DB 196,98,125,24,45,255,38,0,0 ; vbroadcastss 0x26ff(%rip),%ymm13 # 4e24 <_sk_callback_hsw+0x2c0> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,197,38,0,0 ; vbroadcastss 0x26c5(%rip),%ymm13 # 4df8 <_sk_callback_hsw+0x2c2> + DB 196,98,125,24,45,245,38,0,0 ; vbroadcastss 0x26f5(%rip),%ymm13 # 4e28 <_sk_callback_hsw+0x2c4> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,187,38,0,0 ; vbroadcastss 0x26bb(%rip),%ymm11 # 4dfc <_sk_callback_hsw+0x2c6> + DB 196,98,125,24,29,235,38,0,0 ; vbroadcastss 0x26eb(%rip),%ymm11 # 4e2c <_sk_callback_hsw+0x2c8> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,177,38,0,0 ; vbroadcastss 0x26b1(%rip),%ymm12 # 4e00 <_sk_callback_hsw+0x2ca> + DB 196,98,125,24,37,225,38,0,0 ; vbroadcastss 0x26e1(%rip),%ymm12 # 4e30 <_sk_callback_hsw+0x2cc> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,167,38,0,0 ; vbroadcastss 0x26a7(%rip),%ymm12 # 4e04 <_sk_callback_hsw+0x2ce> + DB 196,98,125,24,37,215,38,0,0 ; vbroadcastss 0x26d7(%rip),%ymm12 # 4e34 <_sk_callback_hsw+0x2d0> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,136,38,0,0 ; vbroadcastss 0x2688(%rip),%ymm11 # 4e08 <_sk_callback_hsw+0x2d2> + DB 196,98,125,24,29,184,38,0,0 ; vbroadcastss 0x26b8(%rip),%ymm11 # 4e38 <_sk_callback_hsw+0x2d4> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,126,38,0,0 ; vbroadcastss 0x267e(%rip),%ymm11 # 4e0c <_sk_callback_hsw+0x2d6> + DB 196,98,125,24,29,174,38,0,0 ; vbroadcastss 0x26ae(%rip),%ymm11 # 4e3c <_sk_callback_hsw+0x2d8> DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11 - DB 196,226,125,24,13,116,38,0,0 ; vbroadcastss 0x2674(%rip),%ymm1 # 4e10 <_sk_callback_hsw+0x2da> + DB 196,226,125,24,13,164,38,0,0 ; vbroadcastss 0x26a4(%rip),%ymm1 # 4e40 <_sk_callback_hsw+0x2dc> DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,106,38,0,0 ; vbroadcastss 0x266a(%rip),%ymm10 # 4e14 <_sk_callback_hsw+0x2de> + DB 196,98,125,24,21,154,38,0,0 ; vbroadcastss 0x269a(%rip),%ymm10 # 4e44 <_sk_callback_hsw+0x2e0> DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 196,98,125,24,21,93,38,0,0 ; vbroadcastss 0x265d(%rip),%ymm10 # 4e18 <_sk_callback_hsw+0x2e2> + DB 196,98,125,24,21,141,38,0,0 ; vbroadcastss 0x268d(%rip),%ymm10 # 4e48 <_sk_callback_hsw+0x2e4> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2379,7 +2379,7 @@ _sk_parametric_g_hsw LABEL PROC DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,52,38,0,0 ; vbroadcastss 0x2634(%rip),%ymm8 # 4e1c <_sk_callback_hsw+0x2e6> + DB 196,98,125,24,5,100,38,0,0 ; vbroadcastss 0x2664(%rip),%ymm8 # 4e4c <_sk_callback_hsw+0x2e8> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2397,33 +2397,33 @@ _sk_parametric_b_hsw LABEL PROC DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10 DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,236,37,0,0 ; vbroadcastss 0x25ec(%rip),%ymm12 # 4e20 <_sk_callback_hsw+0x2ea> - DB 196,98,125,24,45,231,37,0,0 ; vbroadcastss 0x25e7(%rip),%ymm13 # 4e24 <_sk_callback_hsw+0x2ee> + DB 196,98,125,24,37,28,38,0,0 ; vbroadcastss 0x261c(%rip),%ymm12 # 4e50 <_sk_callback_hsw+0x2ec> + DB 196,98,125,24,45,23,38,0,0 ; vbroadcastss 0x2617(%rip),%ymm13 # 4e54 <_sk_callback_hsw+0x2f0> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,221,37,0,0 ; vbroadcastss 0x25dd(%rip),%ymm13 # 4e28 <_sk_callback_hsw+0x2f2> + DB 196,98,125,24,45,13,38,0,0 ; vbroadcastss 0x260d(%rip),%ymm13 # 4e58 <_sk_callback_hsw+0x2f4> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,211,37,0,0 ; vbroadcastss 0x25d3(%rip),%ymm13 # 4e2c <_sk_callback_hsw+0x2f6> + DB 196,98,125,24,45,3,38,0,0 ; vbroadcastss 0x2603(%rip),%ymm13 # 4e5c <_sk_callback_hsw+0x2f8> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,201,37,0,0 ; vbroadcastss 0x25c9(%rip),%ymm11 # 4e30 <_sk_callback_hsw+0x2fa> + DB 196,98,125,24,29,249,37,0,0 ; vbroadcastss 0x25f9(%rip),%ymm11 # 4e60 <_sk_callback_hsw+0x2fc> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,191,37,0,0 ; vbroadcastss 0x25bf(%rip),%ymm12 # 4e34 <_sk_callback_hsw+0x2fe> + DB 196,98,125,24,37,239,37,0,0 ; vbroadcastss 0x25ef(%rip),%ymm12 # 4e64 <_sk_callback_hsw+0x300> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,181,37,0,0 ; vbroadcastss 0x25b5(%rip),%ymm12 # 4e38 <_sk_callback_hsw+0x302> + DB 196,98,125,24,37,229,37,0,0 ; vbroadcastss 0x25e5(%rip),%ymm12 # 4e68 <_sk_callback_hsw+0x304> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,150,37,0,0 ; vbroadcastss 0x2596(%rip),%ymm11 # 4e3c <_sk_callback_hsw+0x306> + DB 196,98,125,24,29,198,37,0,0 ; vbroadcastss 0x25c6(%rip),%ymm11 # 4e6c <_sk_callback_hsw+0x308> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,140,37,0,0 ; vbroadcastss 0x258c(%rip),%ymm11 # 4e40 <_sk_callback_hsw+0x30a> + DB 196,98,125,24,29,188,37,0,0 ; vbroadcastss 0x25bc(%rip),%ymm11 # 4e70 <_sk_callback_hsw+0x30c> DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11 - DB 196,226,125,24,21,130,37,0,0 ; vbroadcastss 0x2582(%rip),%ymm2 # 4e44 <_sk_callback_hsw+0x30e> + DB 196,226,125,24,21,178,37,0,0 ; vbroadcastss 0x25b2(%rip),%ymm2 # 4e74 <_sk_callback_hsw+0x310> DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,120,37,0,0 ; vbroadcastss 0x2578(%rip),%ymm10 # 4e48 <_sk_callback_hsw+0x312> + DB 196,98,125,24,21,168,37,0,0 ; vbroadcastss 0x25a8(%rip),%ymm10 # 4e78 <_sk_callback_hsw+0x314> DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2 DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 196,98,125,24,21,107,37,0,0 ; vbroadcastss 0x256b(%rip),%ymm10 # 4e4c <_sk_callback_hsw+0x316> + DB 196,98,125,24,21,155,37,0,0 ; vbroadcastss 0x259b(%rip),%ymm10 # 4e7c <_sk_callback_hsw+0x318> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2431,7 +2431,7 @@ _sk_parametric_b_hsw LABEL PROC DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,66,37,0,0 ; vbroadcastss 0x2542(%rip),%ymm8 # 4e50 <_sk_callback_hsw+0x31a> + DB 196,98,125,24,5,114,37,0,0 ; vbroadcastss 0x2572(%rip),%ymm8 # 4e80 <_sk_callback_hsw+0x31c> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2449,33 +2449,33 @@ _sk_parametric_a_hsw LABEL PROC DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10 DB 196,226,125,24,24 ; vbroadcastss (%rax),%ymm3 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,250,36,0,0 ; vbroadcastss 0x24fa(%rip),%ymm12 # 4e54 <_sk_callback_hsw+0x31e> - DB 196,98,125,24,45,245,36,0,0 ; vbroadcastss 0x24f5(%rip),%ymm13 # 4e58 <_sk_callback_hsw+0x322> + DB 196,98,125,24,37,42,37,0,0 ; vbroadcastss 0x252a(%rip),%ymm12 # 4e84 <_sk_callback_hsw+0x320> + DB 196,98,125,24,45,37,37,0,0 ; vbroadcastss 0x2525(%rip),%ymm13 # 4e88 <_sk_callback_hsw+0x324> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,235,36,0,0 ; vbroadcastss 0x24eb(%rip),%ymm13 # 4e5c <_sk_callback_hsw+0x326> + DB 196,98,125,24,45,27,37,0,0 ; vbroadcastss 0x251b(%rip),%ymm13 # 4e8c <_sk_callback_hsw+0x328> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,225,36,0,0 ; vbroadcastss 0x24e1(%rip),%ymm13 # 4e60 <_sk_callback_hsw+0x32a> + DB 196,98,125,24,45,17,37,0,0 ; vbroadcastss 0x2511(%rip),%ymm13 # 4e90 <_sk_callback_hsw+0x32c> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,215,36,0,0 ; vbroadcastss 0x24d7(%rip),%ymm11 # 4e64 <_sk_callback_hsw+0x32e> + DB 196,98,125,24,29,7,37,0,0 ; vbroadcastss 0x2507(%rip),%ymm11 # 4e94 <_sk_callback_hsw+0x330> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,205,36,0,0 ; vbroadcastss 0x24cd(%rip),%ymm12 # 4e68 <_sk_callback_hsw+0x332> + DB 196,98,125,24,37,253,36,0,0 ; vbroadcastss 0x24fd(%rip),%ymm12 # 4e98 <_sk_callback_hsw+0x334> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,195,36,0,0 ; vbroadcastss 0x24c3(%rip),%ymm12 # 4e6c <_sk_callback_hsw+0x336> + DB 196,98,125,24,37,243,36,0,0 ; vbroadcastss 0x24f3(%rip),%ymm12 # 4e9c <_sk_callback_hsw+0x338> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,164,36,0,0 ; vbroadcastss 0x24a4(%rip),%ymm11 # 4e70 <_sk_callback_hsw+0x33a> + DB 196,98,125,24,29,212,36,0,0 ; vbroadcastss 0x24d4(%rip),%ymm11 # 4ea0 <_sk_callback_hsw+0x33c> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,154,36,0,0 ; vbroadcastss 0x249a(%rip),%ymm11 # 4e74 <_sk_callback_hsw+0x33e> + DB 196,98,125,24,29,202,36,0,0 ; vbroadcastss 0x24ca(%rip),%ymm11 # 4ea4 <_sk_callback_hsw+0x340> DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11 - DB 196,226,125,24,29,144,36,0,0 ; vbroadcastss 0x2490(%rip),%ymm3 # 4e78 <_sk_callback_hsw+0x342> + DB 196,226,125,24,29,192,36,0,0 ; vbroadcastss 0x24c0(%rip),%ymm3 # 4ea8 <_sk_callback_hsw+0x344> DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,134,36,0,0 ; vbroadcastss 0x2486(%rip),%ymm10 # 4e7c <_sk_callback_hsw+0x346> + DB 196,98,125,24,21,182,36,0,0 ; vbroadcastss 0x24b6(%rip),%ymm10 # 4eac <_sk_callback_hsw+0x348> DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3 DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 196,98,125,24,21,121,36,0,0 ; vbroadcastss 0x2479(%rip),%ymm10 # 4e80 <_sk_callback_hsw+0x34a> + DB 196,98,125,24,21,169,36,0,0 ; vbroadcastss 0x24a9(%rip),%ymm10 # 4eb0 <_sk_callback_hsw+0x34c> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -2483,33 +2483,33 @@ _sk_parametric_a_hsw LABEL PROC DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,80,36,0,0 ; vbroadcastss 0x2450(%rip),%ymm8 # 4e84 <_sk_callback_hsw+0x34e> + DB 196,98,125,24,5,128,36,0,0 ; vbroadcastss 0x2480(%rip),%ymm8 # 4eb4 <_sk_callback_hsw+0x350> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_lab_to_xyz_hsw _sk_lab_to_xyz_hsw LABEL PROC - DB 196,98,125,24,5,66,36,0,0 ; vbroadcastss 0x2442(%rip),%ymm8 # 4e88 <_sk_callback_hsw+0x352> - DB 196,98,125,24,13,61,36,0,0 ; vbroadcastss 0x243d(%rip),%ymm9 # 4e8c <_sk_callback_hsw+0x356> - DB 196,98,125,24,21,56,36,0,0 ; vbroadcastss 0x2438(%rip),%ymm10 # 4e90 <_sk_callback_hsw+0x35a> + DB 196,98,125,24,5,114,36,0,0 ; vbroadcastss 0x2472(%rip),%ymm8 # 4eb8 <_sk_callback_hsw+0x354> + DB 196,98,125,24,13,109,36,0,0 ; vbroadcastss 0x246d(%rip),%ymm9 # 4ebc <_sk_callback_hsw+0x358> + DB 196,98,125,24,21,104,36,0,0 ; vbroadcastss 0x2468(%rip),%ymm10 # 4ec0 <_sk_callback_hsw+0x35c> DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1 DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2 - DB 196,98,125,24,13,41,36,0,0 ; vbroadcastss 0x2429(%rip),%ymm9 # 4e94 <_sk_callback_hsw+0x35e> + DB 196,98,125,24,13,89,36,0,0 ; vbroadcastss 0x2459(%rip),%ymm9 # 4ec4 <_sk_callback_hsw+0x360> DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9 - DB 196,226,125,24,5,31,36,0,0 ; vbroadcastss 0x241f(%rip),%ymm0 # 4e98 <_sk_callback_hsw+0x362> + DB 196,226,125,24,5,79,36,0,0 ; vbroadcastss 0x244f(%rip),%ymm0 # 4ec8 <_sk_callback_hsw+0x364> DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 - DB 196,98,125,24,5,22,36,0,0 ; vbroadcastss 0x2416(%rip),%ymm8 # 4e9c <_sk_callback_hsw+0x366> + DB 196,98,125,24,5,70,36,0,0 ; vbroadcastss 0x2446(%rip),%ymm8 # 4ecc <_sk_callback_hsw+0x368> DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8 - DB 196,98,125,24,13,12,36,0,0 ; vbroadcastss 0x240c(%rip),%ymm9 # 4ea0 <_sk_callback_hsw+0x36a> + DB 196,98,125,24,13,60,36,0,0 ; vbroadcastss 0x243c(%rip),%ymm9 # 4ed0 <_sk_callback_hsw+0x36c> DB 196,98,109,172,200 ; vfnmadd213ps %ymm0,%ymm2,%ymm9 DB 196,193,60,89,200 ; vmulps %ymm8,%ymm8,%ymm1 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 - DB 196,226,125,24,21,249,35,0,0 ; vbroadcastss 0x23f9(%rip),%ymm2 # 4ea4 <_sk_callback_hsw+0x36e> + DB 196,226,125,24,21,41,36,0,0 ; vbroadcastss 0x2429(%rip),%ymm2 # 4ed4 <_sk_callback_hsw+0x370> DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10 - DB 196,98,125,24,29,239,35,0,0 ; vbroadcastss 0x23ef(%rip),%ymm11 # 4ea8 <_sk_callback_hsw+0x372> + DB 196,98,125,24,29,31,36,0,0 ; vbroadcastss 0x241f(%rip),%ymm11 # 4ed8 <_sk_callback_hsw+0x374> DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,37,229,35,0,0 ; vbroadcastss 0x23e5(%rip),%ymm12 # 4eac <_sk_callback_hsw+0x376> + DB 196,98,125,24,37,21,36,0,0 ; vbroadcastss 0x2415(%rip),%ymm12 # 4edc <_sk_callback_hsw+0x378> DB 196,65,60,89,196 ; vmulps %ymm12,%ymm8,%ymm8 DB 196,99,61,74,193,160 ; vblendvps %ymm10,%ymm1,%ymm8,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -2524,9 +2524,9 @@ _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 196,226,125,24,5,154,35,0,0 ; vbroadcastss 0x239a(%rip),%ymm0 # 4eb0 <_sk_callback_hsw+0x37a> + DB 196,226,125,24,5,202,35,0,0 ; vbroadcastss 0x23ca(%rip),%ymm0 # 4ee0 <_sk_callback_hsw+0x37c> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,145,35,0,0 ; vbroadcastss 0x2391(%rip),%ymm8 # 4eb4 <_sk_callback_hsw+0x37e> + DB 196,98,125,24,5,193,35,0,0 ; vbroadcastss 0x23c1(%rip),%ymm8 # 4ee4 <_sk_callback_hsw+0x380> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2538,10 +2538,10 @@ _sk_load_a8_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,52 ; jne 2b6a <_sk_load_a8_hsw+0x3e> DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0 - DB 197,249,219,5,92,39,0,0 ; vpand 0x275c(%rip),%xmm0,%xmm0 # 52a0 <_sk_callback_hsw+0x76a> + DB 197,249,219,5,124,39,0,0 ; vpand 0x277c(%rip),%xmm0,%xmm0 # 52c0 <_sk_callback_hsw+0x75c> DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,98,35,0,0 ; vbroadcastss 0x2362(%rip),%ymm1 # 4eb8 <_sk_callback_hsw+0x382> + DB 196,226,125,24,13,146,35,0,0 ; vbroadcastss 0x2392(%rip),%ymm1 # 4ee8 <_sk_callback_hsw+0x384> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -2634,7 +2634,7 @@ _sk_gather_a8_hsw LABEL PROC DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,2,34,0,0 ; vbroadcastss 0x2202(%rip),%ymm1 # 4ebc <_sk_callback_hsw+0x386> + DB 196,226,125,24,13,50,34,0,0 ; vbroadcastss 0x2232(%rip),%ymm1 # 4eec <_sk_callback_hsw+0x388> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -2646,7 +2646,7 @@ PUBLIC _sk_store_a8_hsw _sk_store_a8_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 - DB 196,98,125,24,5,228,33,0,0 ; vbroadcastss 0x21e4(%rip),%ymm8 # 4ec0 <_sk_callback_hsw+0x38a> + DB 196,98,125,24,5,20,34,0,0 ; vbroadcastss 0x2214(%rip),%ymm8 # 4ef0 <_sk_callback_hsw+0x38c> DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8 DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 @@ -2707,13 +2707,13 @@ _sk_load_g8_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,57 ; jne 2dc7 <_sk_load_g8_hsw+0x43> DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0 - DB 197,249,219,5,20,37,0,0 ; vpand 0x2514(%rip),%xmm0,%xmm0 # 52b0 <_sk_callback_hsw+0x77a> + DB 197,249,219,5,52,37,0,0 ; vpand 0x2534(%rip),%xmm0,%xmm0 # 52d0 <_sk_callback_hsw+0x76c> DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,22,33,0,0 ; vbroadcastss 0x2116(%rip),%ymm1 # 4ec4 <_sk_callback_hsw+0x38e> + DB 196,226,125,24,13,70,33,0,0 ; vbroadcastss 0x2146(%rip),%ymm1 # 4ef4 <_sk_callback_hsw+0x390> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,11,33,0,0 ; vbroadcastss 0x210b(%rip),%ymm3 # 4ec8 <_sk_callback_hsw+0x392> + DB 196,226,125,24,29,59,33,0,0 ; vbroadcastss 0x213b(%rip),%ymm3 # 4ef8 <_sk_callback_hsw+0x394> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax @@ -2803,10 +2803,10 @@ _sk_gather_g8_hsw LABEL PROC DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,182,31,0,0 ; vbroadcastss 0x1fb6(%rip),%ymm1 # 4ecc <_sk_callback_hsw+0x396> + DB 196,226,125,24,13,230,31,0,0 ; vbroadcastss 0x1fe6(%rip),%ymm1 # 4efc <_sk_callback_hsw+0x398> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,171,31,0,0 ; vbroadcastss 0x1fab(%rip),%ymm3 # 4ed0 <_sk_callback_hsw+0x39a> + DB 196,226,125,24,29,219,31,0,0 ; vbroadcastss 0x1fdb(%rip),%ymm3 # 4f00 <_sk_callback_hsw+0x39c> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax @@ -2852,14 +2852,14 @@ _sk_gather_i8_hsw LABEL PROC DB 73,139,65,8 ; mov 0x8(%r9),%rax DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,226,117,144,28,128 ; vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 - DB 197,229,219,5,91,33,0,0 ; vpand 0x215b(%rip),%ymm3,%ymm0 # 5140 <_sk_callback_hsw+0x60a> + DB 197,229,219,5,123,33,0,0 ; vpand 0x217b(%rip),%ymm3,%ymm0 # 5160 <_sk_callback_hsw+0x5fc> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,226,30,0,0 ; vbroadcastss 0x1ee2(%rip),%ymm8 # 4ed4 <_sk_callback_hsw+0x39e> + DB 196,98,125,24,5,18,31,0,0 ; vbroadcastss 0x1f12(%rip),%ymm8 # 4f04 <_sk_callback_hsw+0x3a0> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,96,33,0,0 ; vpshufb 0x2160(%rip),%ymm3,%ymm1 # 5160 <_sk_callback_hsw+0x62a> + DB 196,226,101,0,13,128,33,0,0 ; vpshufb 0x2180(%rip),%ymm3,%ymm1 # 5180 <_sk_callback_hsw+0x61c> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,110,33,0,0 ; vpshufb 0x216e(%rip),%ymm3,%ymm2 # 5180 <_sk_callback_hsw+0x64a> + DB 196,226,101,0,21,142,33,0,0 ; vpshufb 0x218e(%rip),%ymm3,%ymm2 # 51a0 <_sk_callback_hsw+0x63c> 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 @@ -2877,23 +2877,23 @@ _sk_load_565_hsw LABEL PROC DB 117,114 ; jne 30aa <_sk_load_565_hsw+0x7c> DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,140,30,0,0 ; vpbroadcastd 0x1e8c(%rip),%ymm0 # 4ed8 <_sk_callback_hsw+0x3a2> + DB 196,226,125,88,5,188,30,0,0 ; vpbroadcastd 0x1ebc(%rip),%ymm0 # 4f08 <_sk_callback_hsw+0x3a4> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,127,30,0,0 ; vbroadcastss 0x1e7f(%rip),%ymm1 # 4edc <_sk_callback_hsw+0x3a6> + DB 196,226,125,24,13,175,30,0,0 ; vbroadcastss 0x1eaf(%rip),%ymm1 # 4f0c <_sk_callback_hsw+0x3a8> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,118,30,0,0 ; vpbroadcastd 0x1e76(%rip),%ymm1 # 4ee0 <_sk_callback_hsw+0x3aa> + DB 196,226,125,88,13,166,30,0,0 ; vpbroadcastd 0x1ea6(%rip),%ymm1 # 4f10 <_sk_callback_hsw+0x3ac> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,105,30,0,0 ; vbroadcastss 0x1e69(%rip),%ymm3 # 4ee4 <_sk_callback_hsw+0x3ae> + DB 196,226,125,24,29,153,30,0,0 ; vbroadcastss 0x1e99(%rip),%ymm3 # 4f14 <_sk_callback_hsw+0x3b0> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,96,30,0,0 ; vpbroadcastd 0x1e60(%rip),%ymm3 # 4ee8 <_sk_callback_hsw+0x3b2> + DB 196,226,125,88,29,144,30,0,0 ; vpbroadcastd 0x1e90(%rip),%ymm3 # 4f18 <_sk_callback_hsw+0x3b4> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,83,30,0,0 ; vbroadcastss 0x1e53(%rip),%ymm3 # 4eec <_sk_callback_hsw+0x3b6> + DB 196,226,125,24,29,131,30,0,0 ; vbroadcastss 0x1e83(%rip),%ymm3 # 4f1c <_sk_callback_hsw+0x3b8> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,72,30,0,0 ; vbroadcastss 0x1e48(%rip),%ymm3 # 4ef0 <_sk_callback_hsw+0x3ba> + DB 196,226,125,24,29,120,30,0,0 ; vbroadcastss 0x1e78(%rip),%ymm3 # 4f20 <_sk_callback_hsw+0x3bc> DB 255,224 ; jmpq *%rax DB 69,137,193 ; mov %r8d,%r9d DB 65,128,225,7 ; and $0x7,%r9b @@ -2979,34 +2979,34 @@ _sk_gather_565_hsw LABEL PROC DB 67,15,183,4,89 ; movzwl (%r9,%r11,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,19,29,0,0 ; vpbroadcastd 0x1d13(%rip),%ymm0 # 4ef4 <_sk_callback_hsw+0x3be> + DB 196,226,125,88,5,67,29,0,0 ; vpbroadcastd 0x1d43(%rip),%ymm0 # 4f24 <_sk_callback_hsw+0x3c0> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,6,29,0,0 ; vbroadcastss 0x1d06(%rip),%ymm1 # 4ef8 <_sk_callback_hsw+0x3c2> + DB 196,226,125,24,13,54,29,0,0 ; vbroadcastss 0x1d36(%rip),%ymm1 # 4f28 <_sk_callback_hsw+0x3c4> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,253,28,0,0 ; vpbroadcastd 0x1cfd(%rip),%ymm1 # 4efc <_sk_callback_hsw+0x3c6> + DB 196,226,125,88,13,45,29,0,0 ; vpbroadcastd 0x1d2d(%rip),%ymm1 # 4f2c <_sk_callback_hsw+0x3c8> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,240,28,0,0 ; vbroadcastss 0x1cf0(%rip),%ymm3 # 4f00 <_sk_callback_hsw+0x3ca> + DB 196,226,125,24,29,32,29,0,0 ; vbroadcastss 0x1d20(%rip),%ymm3 # 4f30 <_sk_callback_hsw+0x3cc> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,231,28,0,0 ; vpbroadcastd 0x1ce7(%rip),%ymm3 # 4f04 <_sk_callback_hsw+0x3ce> + DB 196,226,125,88,29,23,29,0,0 ; vpbroadcastd 0x1d17(%rip),%ymm3 # 4f34 <_sk_callback_hsw+0x3d0> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,218,28,0,0 ; vbroadcastss 0x1cda(%rip),%ymm3 # 4f08 <_sk_callback_hsw+0x3d2> + DB 196,226,125,24,29,10,29,0,0 ; vbroadcastss 0x1d0a(%rip),%ymm3 # 4f38 <_sk_callback_hsw+0x3d4> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,207,28,0,0 ; vbroadcastss 0x1ccf(%rip),%ymm3 # 4f0c <_sk_callback_hsw+0x3d6> + DB 196,226,125,24,29,255,28,0,0 ; vbroadcastss 0x1cff(%rip),%ymm3 # 4f3c <_sk_callback_hsw+0x3d8> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_hsw _sk_store_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 - DB 196,98,125,24,5,195,28,0,0 ; vbroadcastss 0x1cc3(%rip),%ymm8 # 4f10 <_sk_callback_hsw+0x3da> + DB 196,98,125,24,5,243,28,0,0 ; vbroadcastss 0x1cf3(%rip),%ymm8 # 4f40 <_sk_callback_hsw+0x3dc> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,53,114,241,11 ; vpslld $0xb,%ymm9,%ymm9 - DB 196,98,125,24,21,174,28,0,0 ; vbroadcastss 0x1cae(%rip),%ymm10 # 4f14 <_sk_callback_hsw+0x3de> + DB 196,98,125,24,21,222,28,0,0 ; vbroadcastss 0x1cde(%rip),%ymm10 # 4f44 <_sk_callback_hsw+0x3e0> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,45,114,242,5 ; vpslld $0x5,%ymm10,%ymm10 @@ -3072,25 +3072,25 @@ _sk_load_4444_hsw LABEL PROC DB 15,133,138,0,0,0 ; jne 33b8 <_sk_load_4444_hsw+0x98> DB 196,193,122,111,4,83 ; vmovdqu (%r11,%rdx,2),%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,214,27,0,0 ; vpbroadcastd 0x1bd6(%rip),%ymm0 # 4f18 <_sk_callback_hsw+0x3e2> + DB 196,226,125,88,5,6,28,0,0 ; vpbroadcastd 0x1c06(%rip),%ymm0 # 4f48 <_sk_callback_hsw+0x3e4> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,201,27,0,0 ; vbroadcastss 0x1bc9(%rip),%ymm1 # 4f1c <_sk_callback_hsw+0x3e6> + DB 196,226,125,24,13,249,27,0,0 ; vbroadcastss 0x1bf9(%rip),%ymm1 # 4f4c <_sk_callback_hsw+0x3e8> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,192,27,0,0 ; vpbroadcastd 0x1bc0(%rip),%ymm1 # 4f20 <_sk_callback_hsw+0x3ea> + DB 196,226,125,88,13,240,27,0,0 ; vpbroadcastd 0x1bf0(%rip),%ymm1 # 4f50 <_sk_callback_hsw+0x3ec> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,179,27,0,0 ; vbroadcastss 0x1bb3(%rip),%ymm2 # 4f24 <_sk_callback_hsw+0x3ee> + DB 196,226,125,24,21,227,27,0,0 ; vbroadcastss 0x1be3(%rip),%ymm2 # 4f54 <_sk_callback_hsw+0x3f0> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,170,27,0,0 ; vpbroadcastd 0x1baa(%rip),%ymm2 # 4f28 <_sk_callback_hsw+0x3f2> + DB 196,226,125,88,21,218,27,0,0 ; vpbroadcastd 0x1bda(%rip),%ymm2 # 4f58 <_sk_callback_hsw+0x3f4> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,157,27,0,0 ; vbroadcastss 0x1b9d(%rip),%ymm8 # 4f2c <_sk_callback_hsw+0x3f6> + DB 196,98,125,24,5,205,27,0,0 ; vbroadcastss 0x1bcd(%rip),%ymm8 # 4f5c <_sk_callback_hsw+0x3f8> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,147,27,0,0 ; vpbroadcastd 0x1b93(%rip),%ymm8 # 4f30 <_sk_callback_hsw+0x3fa> + DB 196,98,125,88,5,195,27,0,0 ; vpbroadcastd 0x1bc3(%rip),%ymm8 # 4f60 <_sk_callback_hsw+0x3fc> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,133,27,0,0 ; vbroadcastss 0x1b85(%rip),%ymm8 # 4f34 <_sk_callback_hsw+0x3fe> + DB 196,98,125,24,5,181,27,0,0 ; vbroadcastss 0x1bb5(%rip),%ymm8 # 4f64 <_sk_callback_hsw+0x400> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3177,25 +3177,25 @@ _sk_gather_4444_hsw LABEL PROC DB 67,15,183,4,89 ; movzwl (%r9,%r11,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,71,26,0,0 ; vpbroadcastd 0x1a47(%rip),%ymm0 # 4f38 <_sk_callback_hsw+0x402> + DB 196,226,125,88,5,119,26,0,0 ; vpbroadcastd 0x1a77(%rip),%ymm0 # 4f68 <_sk_callback_hsw+0x404> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,58,26,0,0 ; vbroadcastss 0x1a3a(%rip),%ymm1 # 4f3c <_sk_callback_hsw+0x406> + DB 196,226,125,24,13,106,26,0,0 ; vbroadcastss 0x1a6a(%rip),%ymm1 # 4f6c <_sk_callback_hsw+0x408> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,49,26,0,0 ; vpbroadcastd 0x1a31(%rip),%ymm1 # 4f40 <_sk_callback_hsw+0x40a> + DB 196,226,125,88,13,97,26,0,0 ; vpbroadcastd 0x1a61(%rip),%ymm1 # 4f70 <_sk_callback_hsw+0x40c> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,36,26,0,0 ; vbroadcastss 0x1a24(%rip),%ymm2 # 4f44 <_sk_callback_hsw+0x40e> + DB 196,226,125,24,21,84,26,0,0 ; vbroadcastss 0x1a54(%rip),%ymm2 # 4f74 <_sk_callback_hsw+0x410> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,27,26,0,0 ; vpbroadcastd 0x1a1b(%rip),%ymm2 # 4f48 <_sk_callback_hsw+0x412> + DB 196,226,125,88,21,75,26,0,0 ; vpbroadcastd 0x1a4b(%rip),%ymm2 # 4f78 <_sk_callback_hsw+0x414> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,14,26,0,0 ; vbroadcastss 0x1a0e(%rip),%ymm8 # 4f4c <_sk_callback_hsw+0x416> + DB 196,98,125,24,5,62,26,0,0 ; vbroadcastss 0x1a3e(%rip),%ymm8 # 4f7c <_sk_callback_hsw+0x418> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,4,26,0,0 ; vpbroadcastd 0x1a04(%rip),%ymm8 # 4f50 <_sk_callback_hsw+0x41a> + DB 196,98,125,88,5,52,26,0,0 ; vpbroadcastd 0x1a34(%rip),%ymm8 # 4f80 <_sk_callback_hsw+0x41c> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,246,25,0,0 ; vbroadcastss 0x19f6(%rip),%ymm8 # 4f54 <_sk_callback_hsw+0x41e> + DB 196,98,125,24,5,38,26,0,0 ; vbroadcastss 0x1a26(%rip),%ymm8 # 4f84 <_sk_callback_hsw+0x420> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3204,7 +3204,7 @@ PUBLIC _sk_store_4444_hsw _sk_store_4444_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 - DB 196,98,125,24,5,227,25,0,0 ; vbroadcastss 0x19e3(%rip),%ymm8 # 4f58 <_sk_callback_hsw+0x422> + DB 196,98,125,24,5,19,26,0,0 ; vbroadcastss 0x1a13(%rip),%ymm8 # 4f88 <_sk_callback_hsw+0x424> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,53,114,241,12 ; vpslld $0xc,%ymm9,%ymm9 @@ -3278,14 +3278,14 @@ _sk_load_8888_hsw LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,88 ; jne 36c1 <_sk_load_8888_hsw+0x6d> DB 196,193,124,16,26 ; vmovups (%r10),%ymm3 - DB 197,228,84,5,42,27,0,0 ; vandps 0x1b2a(%rip),%ymm3,%ymm0 # 51a0 <_sk_callback_hsw+0x66a> + DB 197,228,84,5,74,27,0,0 ; vandps 0x1b4a(%rip),%ymm3,%ymm0 # 51c0 <_sk_callback_hsw+0x65c> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,217,24,0,0 ; vbroadcastss 0x18d9(%rip),%ymm8 # 4f5c <_sk_callback_hsw+0x426> + DB 196,98,125,24,5,9,25,0,0 ; vbroadcastss 0x1909(%rip),%ymm8 # 4f8c <_sk_callback_hsw+0x428> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,47,27,0,0 ; vpshufb 0x1b2f(%rip),%ymm3,%ymm1 # 51c0 <_sk_callback_hsw+0x68a> + DB 196,226,101,0,13,79,27,0,0 ; vpshufb 0x1b4f(%rip),%ymm3,%ymm1 # 51e0 <_sk_callback_hsw+0x67c> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,61,27,0,0 ; vpshufb 0x1b3d(%rip),%ymm3,%ymm2 # 51e0 <_sk_callback_hsw+0x6aa> + DB 196,226,101,0,21,93,27,0,0 ; vpshufb 0x1b5d(%rip),%ymm3,%ymm2 # 5200 <_sk_callback_hsw+0x69c> 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 @@ -3315,14 +3315,14 @@ _sk_gather_8888_hsw LABEL PROC DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,194,117,144,28,129 ; vpgatherdd %ymm1,(%r9,%ymm0,4),%ymm3 - DB 197,229,219,5,235,26,0,0 ; vpand 0x1aeb(%rip),%ymm3,%ymm0 # 5200 <_sk_callback_hsw+0x6ca> + DB 197,229,219,5,11,27,0,0 ; vpand 0x1b0b(%rip),%ymm3,%ymm0 # 5220 <_sk_callback_hsw+0x6bc> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,62,24,0,0 ; vbroadcastss 0x183e(%rip),%ymm8 # 4f60 <_sk_callback_hsw+0x42a> + DB 196,98,125,24,5,110,24,0,0 ; vbroadcastss 0x186e(%rip),%ymm8 # 4f90 <_sk_callback_hsw+0x42c> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,240,26,0,0 ; vpshufb 0x1af0(%rip),%ymm3,%ymm1 # 5220 <_sk_callback_hsw+0x6ea> + DB 196,226,101,0,13,16,27,0,0 ; vpshufb 0x1b10(%rip),%ymm3,%ymm1 # 5240 <_sk_callback_hsw+0x6dc> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,254,26,0,0 ; vpshufb 0x1afe(%rip),%ymm3,%ymm2 # 5240 <_sk_callback_hsw+0x70a> + DB 196,226,101,0,21,30,27,0,0 ; vpshufb 0x1b1e(%rip),%ymm3,%ymm2 # 5260 <_sk_callback_hsw+0x6fc> 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 @@ -3337,7 +3337,7 @@ _sk_store_8888_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 DB 76,3,16 ; add (%rax),%r10 - DB 196,98,125,24,5,238,23,0,0 ; vbroadcastss 0x17ee(%rip),%ymm8 # 4f64 <_sk_callback_hsw+0x42e> + DB 196,98,125,24,5,30,24,0,0 ; vbroadcastss 0x181e(%rip),%ymm8 # 4f94 <_sk_callback_hsw+0x430> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -3526,7 +3526,7 @@ _sk_load_u16_be_hsw LABEL PROC DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,21,229,20,0,0 ; vbroadcastss 0x14e5(%rip),%ymm10 # 4f68 <_sk_callback_hsw+0x432> + DB 196,98,125,24,21,21,21,0,0 ; vbroadcastss 0x1515(%rip),%ymm10 # 4f98 <_sk_callback_hsw+0x434> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -3608,7 +3608,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,21,118,19,0,0 ; vbroadcastss 0x1376(%rip),%ymm10 # 4f6c <_sk_callback_hsw+0x436> + DB 196,98,125,24,21,166,19,0,0 ; vbroadcastss 0x13a6(%rip),%ymm10 # 4f9c <_sk_callback_hsw+0x438> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -3625,7 +3625,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,42,19,0,0 ; vbroadcastss 0x132a(%rip),%ymm3 # 4f70 <_sk_callback_hsw+0x43a> + DB 196,226,125,24,29,90,19,0,0 ; vbroadcastss 0x135a(%rip),%ymm3 # 4fa0 <_sk_callback_hsw+0x43c> DB 255,224 ; jmpq *%rax DB 196,193,121,110,4,65 ; vmovd (%r9,%rax,2),%xmm0 DB 196,193,121,196,68,65,4,2 ; vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0 @@ -3666,7 +3666,7 @@ _sk_store_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,8 ; mov (%rax),%r9 DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax - DB 196,98,125,24,5,103,18,0,0 ; vbroadcastss 0x1267(%rip),%ymm8 # 4f74 <_sk_callback_hsw+0x43e> + DB 196,98,125,24,5,151,18,0,0 ; vbroadcastss 0x1297(%rip),%ymm8 # 4fa4 <_sk_callback_hsw+0x440> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10 @@ -3853,26 +3853,28 @@ _sk_clamp_y_hsw LABEL PROC PUBLIC _sk_repeat_x_hsw _sk_repeat_x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 - DB 196,65,124,94,200 ; vdivps %ymm8,%ymm0,%ymm9 - DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 - DB 196,98,61,172,200 ; vfnmadd213ps %ymm0,%ymm8,%ymm9 + DB 196,98,125,24,64,4 ; vbroadcastss 0x4(%rax),%ymm8 + DB 197,60,89,192 ; vmulps %ymm0,%ymm8,%ymm8 + DB 196,67,125,8,192,1 ; vroundps $0x1,%ymm8,%ymm8 + DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 + DB 196,98,53,172,192 ; vfnmadd213ps %ymm0,%ymm9,%ymm8 DB 197,253,118,192 ; vpcmpeqd %ymm0,%ymm0,%ymm0 - DB 197,189,254,192 ; vpaddd %ymm0,%ymm8,%ymm0 - DB 197,180,93,192 ; vminps %ymm0,%ymm9,%ymm0 + DB 197,181,254,192 ; vpaddd %ymm0,%ymm9,%ymm0 + DB 197,188,93,192 ; vminps %ymm0,%ymm8,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_repeat_y_hsw _sk_repeat_y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 - DB 196,65,116,94,200 ; vdivps %ymm8,%ymm1,%ymm9 - DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 - DB 196,98,61,172,201 ; vfnmadd213ps %ymm1,%ymm8,%ymm9 + DB 196,98,125,24,64,4 ; vbroadcastss 0x4(%rax),%ymm8 + DB 197,60,89,193 ; vmulps %ymm1,%ymm8,%ymm8 + DB 196,67,125,8,192,1 ; vroundps $0x1,%ymm8,%ymm8 + DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 + DB 196,98,53,172,193 ; vfnmadd213ps %ymm1,%ymm9,%ymm8 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 197,189,254,201 ; vpaddd %ymm1,%ymm8,%ymm1 - DB 197,180,93,201 ; vminps %ymm1,%ymm9,%ymm1 + DB 197,181,254,201 ; vpaddd %ymm1,%ymm9,%ymm1 + DB 197,188,93,201 ; vminps %ymm1,%ymm8,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3880,19 +3882,22 @@ PUBLIC _sk_mirror_x_hsw _sk_mirror_x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,122,16,0 ; vmovss (%rax),%xmm8 - DB 196,66,125,24,200 ; vbroadcastss %xmm8,%ymm9 - DB 196,65,124,92,209 ; vsubps %ymm9,%ymm0,%ymm10 + DB 197,122,16,72,4 ; vmovss 0x4(%rax),%xmm9 + DB 196,66,125,24,208 ; vbroadcastss %xmm8,%ymm10 + DB 196,65,124,92,218 ; vsubps %ymm10,%ymm0,%ymm11 DB 196,193,58,88,192 ; vaddss %xmm8,%xmm8,%xmm0 + DB 196,98,125,24,192 ; vbroadcastss %xmm0,%ymm8 + DB 197,178,89,5,37,15,0,0 ; vmulss 0xf25(%rip),%xmm9,%xmm0 # 4fa8 <_sk_callback_hsw+0x444> DB 196,226,125,24,192 ; vbroadcastss %xmm0,%ymm0 - DB 197,44,94,192 ; vdivps %ymm0,%ymm10,%ymm8 - DB 196,67,125,8,192,1 ; vroundps $0x1,%ymm8,%ymm8 - DB 196,66,125,172,194 ; vfnmadd213ps %ymm10,%ymm0,%ymm8 - DB 196,193,60,92,193 ; vsubps %ymm9,%ymm8,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 196,227,125,8,192,1 ; vroundps $0x1,%ymm0,%ymm0 + DB 196,194,61,172,195 ; vfnmadd213ps %ymm11,%ymm8,%ymm0 + DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,60,92,192 ; vsubps %ymm0,%ymm8,%ymm8 DB 197,188,84,192 ; vandps %ymm0,%ymm8,%ymm0 DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,65,53,254,192 ; vpaddd %ymm8,%ymm9,%ymm8 + DB 196,65,45,254,192 ; vpaddd %ymm8,%ymm10,%ymm8 DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3901,19 +3906,22 @@ PUBLIC _sk_mirror_y_hsw _sk_mirror_y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,122,16,0 ; vmovss (%rax),%xmm8 - DB 196,66,125,24,200 ; vbroadcastss %xmm8,%ymm9 - DB 196,65,116,92,209 ; vsubps %ymm9,%ymm1,%ymm10 + DB 197,122,16,72,4 ; vmovss 0x4(%rax),%xmm9 + DB 196,66,125,24,208 ; vbroadcastss %xmm8,%ymm10 + DB 196,65,116,92,218 ; vsubps %ymm10,%ymm1,%ymm11 DB 196,193,58,88,200 ; vaddss %xmm8,%xmm8,%xmm1 + DB 196,98,125,24,193 ; vbroadcastss %xmm1,%ymm8 + DB 197,178,89,13,201,14,0,0 ; vmulss 0xec9(%rip),%xmm9,%xmm1 # 4fac <_sk_callback_hsw+0x448> DB 196,226,125,24,201 ; vbroadcastss %xmm1,%ymm1 - DB 197,44,94,193 ; vdivps %ymm1,%ymm10,%ymm8 - DB 196,67,125,8,192,1 ; vroundps $0x1,%ymm8,%ymm8 - DB 196,66,117,172,194 ; vfnmadd213ps %ymm10,%ymm1,%ymm8 - DB 196,193,60,92,201 ; vsubps %ymm9,%ymm8,%ymm1 + DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 + DB 196,227,125,8,201,1 ; vroundps $0x1,%ymm1,%ymm1 + DB 196,194,61,172,203 ; vfnmadd213ps %ymm11,%ymm8,%ymm1 + DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,60,92,193 ; vsubps %ymm1,%ymm8,%ymm8 DB 197,188,84,201 ; vandps %ymm1,%ymm8,%ymm1 DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,65,53,254,192 ; vpaddd %ymm8,%ymm9,%ymm8 + DB 196,65,45,254,192 ; vpaddd %ymm8,%ymm10,%ymm8 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3922,7 +3930,7 @@ PUBLIC _sk_clamp_x_1_hsw _sk_clamp_x_1_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,120,14,0,0 ; vbroadcastss 0xe78(%rip),%ymm8 # 4f78 <_sk_callback_hsw+0x442> + DB 196,98,125,24,5,130,14,0,0 ; vbroadcastss 0xe82(%rip),%ymm8 # 4fb0 <_sk_callback_hsw+0x44c> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3936,9 +3944,9 @@ _sk_repeat_x_1_hsw LABEL PROC PUBLIC _sk_mirror_x_1_hsw _sk_mirror_x_1_hsw LABEL PROC - DB 196,98,125,24,5,91,14,0,0 ; vbroadcastss 0xe5b(%rip),%ymm8 # 4f7c <_sk_callback_hsw+0x446> + DB 196,98,125,24,5,101,14,0,0 ; vbroadcastss 0xe65(%rip),%ymm8 # 4fb4 <_sk_callback_hsw+0x450> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,81,14,0,0 ; vbroadcastss 0xe51(%rip),%ymm9 # 4f80 <_sk_callback_hsw+0x44a> + DB 196,98,125,24,13,91,14,0,0 ; vbroadcastss 0xe5b(%rip),%ymm9 # 4fb8 <_sk_callback_hsw+0x454> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -3952,11 +3960,11 @@ _sk_mirror_x_1_hsw LABEL PROC PUBLIC _sk_luminance_to_alpha_hsw _sk_luminance_to_alpha_hsw LABEL PROC - DB 196,226,125,24,29,33,14,0,0 ; vbroadcastss 0xe21(%rip),%ymm3 # 4f84 <_sk_callback_hsw+0x44e> - DB 196,98,125,24,5,28,14,0,0 ; vbroadcastss 0xe1c(%rip),%ymm8 # 4f88 <_sk_callback_hsw+0x452> + DB 196,226,125,24,29,43,14,0,0 ; vbroadcastss 0xe2b(%rip),%ymm3 # 4fbc <_sk_callback_hsw+0x458> + DB 196,98,125,24,5,38,14,0,0 ; vbroadcastss 0xe26(%rip),%ymm8 # 4fc0 <_sk_callback_hsw+0x45c> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1 - DB 196,226,125,24,29,13,14,0,0 ; vbroadcastss 0xe0d(%rip),%ymm3 # 4f8c <_sk_callback_hsw+0x456> + DB 196,226,125,24,29,23,14,0,0 ; vbroadcastss 0xe17(%rip),%ymm3 # 4fc4 <_sk_callback_hsw+0x460> 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 @@ -4118,9 +4126,9 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC DB 76,139,72,8 ; mov 0x8(%rax),%r9 DB 77,137,211 ; mov %r10,%r11 DB 73,255,203 ; dec %r11 - DB 120,7 ; js 4438 <_sk_evenly_spaced_gradient_hsw+0x19> + DB 120,7 ; js 4466 <_sk_evenly_spaced_gradient_hsw+0x19> DB 196,193,242,42,203 ; vcvtsi2ss %r11,%xmm1,%xmm1 - DB 235,22 ; jmp 444e <_sk_evenly_spaced_gradient_hsw+0x2f> + DB 235,22 ; jmp 447c <_sk_evenly_spaced_gradient_hsw+0x2f> DB 76,137,219 ; mov %r11,%rbx DB 72,209,235 ; shr %rbx DB 65,131,227,1 ; and $0x1,%r11d @@ -4131,7 +4139,7 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC DB 197,244,89,200 ; vmulps %ymm0,%ymm1,%ymm1 DB 197,126,91,217 ; vcvttps2dq %ymm1,%ymm11 DB 73,131,250,8 ; cmp $0x8,%r10 - DB 119,70 ; ja 44a7 <_sk_evenly_spaced_gradient_hsw+0x88> + DB 119,70 ; ja 44d5 <_sk_evenly_spaced_gradient_hsw+0x88> DB 196,66,37,22,1 ; vpermps (%r9),%ymm11,%ymm8 DB 72,139,88,40 ; mov 0x28(%rax),%rbx DB 196,98,37,22,11 ; vpermps (%rbx),%ymm11,%ymm9 @@ -4147,7 +4155,7 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC DB 196,226,37,22,27 ; vpermps (%rbx),%ymm11,%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 196,98,37,22,40 ; vpermps (%rax),%ymm11,%ymm13 - DB 235,110 ; jmp 4515 <_sk_evenly_spaced_gradient_hsw+0xf6> + DB 235,110 ; jmp 4543 <_sk_evenly_spaced_gradient_hsw+0xf6> DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,2,117,146,4,153 ; vgatherdps %ymm1,(%r9,%ymm11,4),%ymm8 @@ -4182,14 +4190,14 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC PUBLIC _sk_gauss_a_to_rgba_hsw _sk_gauss_a_to_rgba_hsw LABEL PROC - DB 196,226,125,24,5,85,10,0,0 ; vbroadcastss 0xa55(%rip),%ymm0 # 4f90 <_sk_callback_hsw+0x45a> - DB 196,226,125,24,13,80,10,0,0 ; vbroadcastss 0xa50(%rip),%ymm1 # 4f94 <_sk_callback_hsw+0x45e> + DB 196,226,125,24,5,95,10,0,0 ; vbroadcastss 0xa5f(%rip),%ymm0 # 4fc8 <_sk_callback_hsw+0x464> + DB 196,226,125,24,13,90,10,0,0 ; vbroadcastss 0xa5a(%rip),%ymm1 # 4fcc <_sk_callback_hsw+0x468> DB 196,226,101,168,200 ; vfmadd213ps %ymm0,%ymm3,%ymm1 - DB 196,226,125,24,5,70,10,0,0 ; vbroadcastss 0xa46(%rip),%ymm0 # 4f98 <_sk_callback_hsw+0x462> + DB 196,226,125,24,5,80,10,0,0 ; vbroadcastss 0xa50(%rip),%ymm0 # 4fd0 <_sk_callback_hsw+0x46c> DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0 - DB 196,226,125,24,13,60,10,0,0 ; vbroadcastss 0xa3c(%rip),%ymm1 # 4f9c <_sk_callback_hsw+0x466> + DB 196,226,125,24,13,70,10,0,0 ; vbroadcastss 0xa46(%rip),%ymm1 # 4fd4 <_sk_callback_hsw+0x470> DB 196,226,101,184,200 ; vfmadd231ps %ymm0,%ymm3,%ymm1 - DB 196,226,125,24,5,50,10,0,0 ; vbroadcastss 0xa32(%rip),%ymm0 # 4fa0 <_sk_callback_hsw+0x46a> + DB 196,226,125,24,5,60,10,0,0 ; vbroadcastss 0xa3c(%rip),%ymm0 # 4fd8 <_sk_callback_hsw+0x474> DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -4202,11 +4210,11 @@ _sk_gradient_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,8 ; mov (%rax),%r9 DB 73,131,249,1 ; cmp $0x1,%r9 - DB 15,134,180,0,0,0 ; jbe 4646 <_sk_gradient_hsw+0xc3> + DB 15,134,180,0,0,0 ; jbe 4674 <_sk_gradient_hsw+0xc3> DB 76,139,80,72 ; mov 0x48(%rax),%r10 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 65,187,1,0,0,0 ; mov $0x1,%r11d - DB 196,226,125,24,21,251,9,0,0 ; vbroadcastss 0x9fb(%rip),%ymm2 # 4fa4 <_sk_callback_hsw+0x46e> + DB 196,226,125,24,21,5,10,0,0 ; vbroadcastss 0xa05(%rip),%ymm2 # 4fdc <_sk_callback_hsw+0x478> DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9 DB 196,130,125,24,28,154 ; vbroadcastss (%r10,%r11,4),%ymm3 DB 197,228,194,216,2 ; vcmpleps %ymm0,%ymm3,%ymm3 @@ -4214,10 +4222,10 @@ _sk_gradient_hsw LABEL PROC DB 196,65,101,254,201 ; vpaddd %ymm9,%ymm3,%ymm9 DB 73,255,195 ; inc %r11 DB 77,57,217 ; cmp %r11,%r9 - DB 117,226 ; jne 45ae <_sk_gradient_hsw+0x2b> + DB 117,226 ; jne 45dc <_sk_gradient_hsw+0x2b> DB 76,139,80,8 ; mov 0x8(%rax),%r10 DB 73,131,249,8 ; cmp $0x8,%r9 - DB 118,121 ; jbe 464f <_sk_gradient_hsw+0xcc> + DB 118,121 ; jbe 467d <_sk_gradient_hsw+0xcc> DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,2,117,146,4,138 ; vgatherdps %ymm1,(%r10,%ymm9,4),%ymm8 @@ -4241,7 +4249,7 @@ _sk_gradient_hsw LABEL PROC DB 196,130,21,146,28,137 ; vgatherdps %ymm13,(%r9,%ymm9,4),%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 196,34,13,146,44,136 ; vgatherdps %ymm14,(%rax,%ymm9,4),%ymm13 - DB 235,77 ; jmp 4693 <_sk_gradient_hsw+0x110> + DB 235,77 ; jmp 46c1 <_sk_gradient_hsw+0x110> DB 76,139,80,8 ; mov 0x8(%rax),%r10 DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 196,66,53,22,2 ; vpermps (%r10),%ymm9,%ymm8 @@ -4297,24 +4305,24 @@ _sk_xy_to_unit_angle_hsw LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,122,8,0,0 ; vbroadcastss 0x87a(%rip),%ymm13 # 4fa8 <_sk_callback_hsw+0x472> - DB 196,98,125,24,53,117,8,0,0 ; vbroadcastss 0x875(%rip),%ymm14 # 4fac <_sk_callback_hsw+0x476> + DB 196,98,125,24,45,132,8,0,0 ; vbroadcastss 0x884(%rip),%ymm13 # 4fe0 <_sk_callback_hsw+0x47c> + DB 196,98,125,24,53,127,8,0,0 ; vbroadcastss 0x87f(%rip),%ymm14 # 4fe4 <_sk_callback_hsw+0x480> DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14 - DB 196,98,125,24,45,107,8,0,0 ; vbroadcastss 0x86b(%rip),%ymm13 # 4fb0 <_sk_callback_hsw+0x47a> + DB 196,98,125,24,45,117,8,0,0 ; vbroadcastss 0x875(%rip),%ymm13 # 4fe8 <_sk_callback_hsw+0x484> DB 196,66,29,184,238 ; vfmadd231ps %ymm14,%ymm12,%ymm13 - DB 196,98,125,24,53,97,8,0,0 ; vbroadcastss 0x861(%rip),%ymm14 # 4fb4 <_sk_callback_hsw+0x47e> + DB 196,98,125,24,53,107,8,0,0 ; vbroadcastss 0x86b(%rip),%ymm14 # 4fec <_sk_callback_hsw+0x488> DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14 DB 196,65,36,89,222 ; vmulps %ymm14,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,76,8,0,0 ; vbroadcastss 0x84c(%rip),%ymm10 # 4fb8 <_sk_callback_hsw+0x482> + DB 196,98,125,24,21,86,8,0,0 ; vbroadcastss 0x856(%rip),%ymm10 # 4ff0 <_sk_callback_hsw+0x48c> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,54,8,0,0 ; vbroadcastss 0x836(%rip),%ymm10 # 4fbc <_sk_callback_hsw+0x486> + DB 196,98,125,24,21,64,8,0,0 ; vbroadcastss 0x840(%rip),%ymm10 # 4ff4 <_sk_callback_hsw+0x490> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,32,8,0,0 ; vbroadcastss 0x820(%rip),%ymm10 # 4fc0 <_sk_callback_hsw+0x48a> + DB 196,98,125,24,21,42,8,0,0 ; vbroadcastss 0x82a(%rip),%ymm10 # 4ff8 <_sk_callback_hsw+0x494> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -4333,7 +4341,7 @@ _sk_xy_to_radius_hsw LABEL PROC PUBLIC _sk_save_xy_hsw _sk_save_xy_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,237,7,0,0 ; vbroadcastss 0x7ed(%rip),%ymm8 # 4fc4 <_sk_callback_hsw+0x48e> + DB 196,98,125,24,5,247,7,0,0 ; vbroadcastss 0x7f7(%rip),%ymm8 # 4ffc <_sk_callback_hsw+0x498> 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 @@ -4363,9 +4371,9 @@ _sk_accumulate_hsw LABEL PROC PUBLIC _sk_bilinear_nx_hsw _sk_bilinear_nx_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,129,7,0,0 ; vbroadcastss 0x781(%rip),%ymm0 # 4fc8 <_sk_callback_hsw+0x492> + DB 196,226,125,24,5,139,7,0,0 ; vbroadcastss 0x78b(%rip),%ymm0 # 5000 <_sk_callback_hsw+0x49c> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,120,7,0,0 ; vbroadcastss 0x778(%rip),%ymm8 # 4fcc <_sk_callback_hsw+0x496> + DB 196,98,125,24,5,130,7,0,0 ; vbroadcastss 0x782(%rip),%ymm8 # 5004 <_sk_callback_hsw+0x4a0> 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 @@ -4374,7 +4382,7 @@ _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 196,226,125,24,5,96,7,0,0 ; vbroadcastss 0x760(%rip),%ymm0 # 4fd0 <_sk_callback_hsw+0x49a> + DB 196,226,125,24,5,106,7,0,0 ; vbroadcastss 0x76a(%rip),%ymm0 # 5008 <_sk_callback_hsw+0x4a4> 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) @@ -4384,9 +4392,9 @@ _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 196,226,125,24,13,68,7,0,0 ; vbroadcastss 0x744(%rip),%ymm1 # 4fd4 <_sk_callback_hsw+0x49e> + DB 196,226,125,24,13,78,7,0,0 ; vbroadcastss 0x74e(%rip),%ymm1 # 500c <_sk_callback_hsw+0x4a8> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,58,7,0,0 ; vbroadcastss 0x73a(%rip),%ymm8 # 4fd8 <_sk_callback_hsw+0x4a2> + DB 196,98,125,24,5,68,7,0,0 ; vbroadcastss 0x744(%rip),%ymm8 # 5010 <_sk_callback_hsw+0x4ac> 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 @@ -4395,7 +4403,7 @@ _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 196,226,125,24,13,34,7,0,0 ; vbroadcastss 0x722(%rip),%ymm1 # 4fdc <_sk_callback_hsw+0x4a6> + DB 196,226,125,24,13,44,7,0,0 ; vbroadcastss 0x72c(%rip),%ymm1 # 5014 <_sk_callback_hsw+0x4b0> 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) @@ -4405,13 +4413,13 @@ _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 196,226,125,24,5,5,7,0,0 ; vbroadcastss 0x705(%rip),%ymm0 # 4fe0 <_sk_callback_hsw+0x4aa> + DB 196,226,125,24,5,15,7,0,0 ; vbroadcastss 0x70f(%rip),%ymm0 # 5018 <_sk_callback_hsw+0x4b4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,252,6,0,0 ; vbroadcastss 0x6fc(%rip),%ymm8 # 4fe4 <_sk_callback_hsw+0x4ae> + DB 196,98,125,24,5,6,7,0,0 ; vbroadcastss 0x706(%rip),%ymm8 # 501c <_sk_callback_hsw+0x4b8> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,237,6,0,0 ; vbroadcastss 0x6ed(%rip),%ymm10 # 4fe8 <_sk_callback_hsw+0x4b2> - DB 196,98,125,24,29,232,6,0,0 ; vbroadcastss 0x6e8(%rip),%ymm11 # 4fec <_sk_callback_hsw+0x4b6> + DB 196,98,125,24,21,247,6,0,0 ; vbroadcastss 0x6f7(%rip),%ymm10 # 5020 <_sk_callback_hsw+0x4bc> + DB 196,98,125,24,29,242,6,0,0 ; vbroadcastss 0x6f2(%rip),%ymm11 # 5024 <_sk_callback_hsw+0x4c0> 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) @@ -4421,16 +4429,16 @@ _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 196,226,125,24,5,203,6,0,0 ; vbroadcastss 0x6cb(%rip),%ymm0 # 4ff0 <_sk_callback_hsw+0x4ba> + DB 196,226,125,24,5,213,6,0,0 ; vbroadcastss 0x6d5(%rip),%ymm0 # 5028 <_sk_callback_hsw+0x4c4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,194,6,0,0 ; vbroadcastss 0x6c2(%rip),%ymm8 # 4ff4 <_sk_callback_hsw+0x4be> + DB 196,98,125,24,5,204,6,0,0 ; vbroadcastss 0x6cc(%rip),%ymm8 # 502c <_sk_callback_hsw+0x4c8> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,184,6,0,0 ; vbroadcastss 0x6b8(%rip),%ymm9 # 4ff8 <_sk_callback_hsw+0x4c2> - DB 196,98,125,24,21,179,6,0,0 ; vbroadcastss 0x6b3(%rip),%ymm10 # 4ffc <_sk_callback_hsw+0x4c6> + DB 196,98,125,24,13,194,6,0,0 ; vbroadcastss 0x6c2(%rip),%ymm9 # 5030 <_sk_callback_hsw+0x4cc> + DB 196,98,125,24,21,189,6,0,0 ; vbroadcastss 0x6bd(%rip),%ymm10 # 5034 <_sk_callback_hsw+0x4d0> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,169,6,0,0 ; vbroadcastss 0x6a9(%rip),%ymm9 # 5000 <_sk_callback_hsw+0x4ca> + DB 196,98,125,24,13,179,6,0,0 ; vbroadcastss 0x6b3(%rip),%ymm9 # 5038 <_sk_callback_hsw+0x4d4> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,159,6,0,0 ; vbroadcastss 0x69f(%rip),%ymm10 # 5004 <_sk_callback_hsw+0x4ce> + DB 196,98,125,24,21,169,6,0,0 ; vbroadcastss 0x6a9(%rip),%ymm10 # 503c <_sk_callback_hsw+0x4d8> 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 @@ -4439,14 +4447,14 @@ _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 196,98,125,24,5,135,6,0,0 ; vbroadcastss 0x687(%rip),%ymm8 # 5008 <_sk_callback_hsw+0x4d2> + DB 196,98,125,24,5,145,6,0,0 ; vbroadcastss 0x691(%rip),%ymm8 # 5040 <_sk_callback_hsw+0x4dc> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,121,6,0,0 ; vbroadcastss 0x679(%rip),%ymm10 # 500c <_sk_callback_hsw+0x4d6> - DB 196,98,125,24,29,116,6,0,0 ; vbroadcastss 0x674(%rip),%ymm11 # 5010 <_sk_callback_hsw+0x4da> + DB 196,98,125,24,21,131,6,0,0 ; vbroadcastss 0x683(%rip),%ymm10 # 5044 <_sk_callback_hsw+0x4e0> + DB 196,98,125,24,29,126,6,0,0 ; vbroadcastss 0x67e(%rip),%ymm11 # 5048 <_sk_callback_hsw+0x4e4> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,101,6,0,0 ; vbroadcastss 0x665(%rip),%ymm8 # 5014 <_sk_callback_hsw+0x4de> + DB 196,98,125,24,5,111,6,0,0 ; vbroadcastss 0x66f(%rip),%ymm8 # 504c <_sk_callback_hsw+0x4e8> 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 @@ -4455,12 +4463,12 @@ _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 196,226,125,24,5,77,6,0,0 ; vbroadcastss 0x64d(%rip),%ymm0 # 5018 <_sk_callback_hsw+0x4e2> + DB 196,226,125,24,5,87,6,0,0 ; vbroadcastss 0x657(%rip),%ymm0 # 5050 <_sk_callback_hsw+0x4ec> 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 196,98,125,24,21,58,6,0,0 ; vbroadcastss 0x63a(%rip),%ymm10 # 501c <_sk_callback_hsw+0x4e6> - DB 196,98,125,24,29,53,6,0,0 ; vbroadcastss 0x635(%rip),%ymm11 # 5020 <_sk_callback_hsw+0x4ea> + DB 196,98,125,24,21,68,6,0,0 ; vbroadcastss 0x644(%rip),%ymm10 # 5054 <_sk_callback_hsw+0x4f0> + DB 196,98,125,24,29,63,6,0,0 ; vbroadcastss 0x63f(%rip),%ymm11 # 5058 <_sk_callback_hsw+0x4f4> 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) @@ -4470,13 +4478,13 @@ _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 196,226,125,24,13,24,6,0,0 ; vbroadcastss 0x618(%rip),%ymm1 # 5024 <_sk_callback_hsw+0x4ee> + DB 196,226,125,24,13,34,6,0,0 ; vbroadcastss 0x622(%rip),%ymm1 # 505c <_sk_callback_hsw+0x4f8> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,14,6,0,0 ; vbroadcastss 0x60e(%rip),%ymm8 # 5028 <_sk_callback_hsw+0x4f2> + DB 196,98,125,24,5,24,6,0,0 ; vbroadcastss 0x618(%rip),%ymm8 # 5060 <_sk_callback_hsw+0x4fc> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,255,5,0,0 ; vbroadcastss 0x5ff(%rip),%ymm10 # 502c <_sk_callback_hsw+0x4f6> - DB 196,98,125,24,29,250,5,0,0 ; vbroadcastss 0x5fa(%rip),%ymm11 # 5030 <_sk_callback_hsw+0x4fa> + DB 196,98,125,24,21,9,6,0,0 ; vbroadcastss 0x609(%rip),%ymm10 # 5064 <_sk_callback_hsw+0x500> + DB 196,98,125,24,29,4,6,0,0 ; vbroadcastss 0x604(%rip),%ymm11 # 5068 <_sk_callback_hsw+0x504> 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) @@ -4486,16 +4494,16 @@ _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 196,226,125,24,13,221,5,0,0 ; vbroadcastss 0x5dd(%rip),%ymm1 # 5034 <_sk_callback_hsw+0x4fe> + DB 196,226,125,24,13,231,5,0,0 ; vbroadcastss 0x5e7(%rip),%ymm1 # 506c <_sk_callback_hsw+0x508> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,211,5,0,0 ; vbroadcastss 0x5d3(%rip),%ymm8 # 5038 <_sk_callback_hsw+0x502> + DB 196,98,125,24,5,221,5,0,0 ; vbroadcastss 0x5dd(%rip),%ymm8 # 5070 <_sk_callback_hsw+0x50c> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,201,5,0,0 ; vbroadcastss 0x5c9(%rip),%ymm9 # 503c <_sk_callback_hsw+0x506> - DB 196,98,125,24,21,196,5,0,0 ; vbroadcastss 0x5c4(%rip),%ymm10 # 5040 <_sk_callback_hsw+0x50a> + DB 196,98,125,24,13,211,5,0,0 ; vbroadcastss 0x5d3(%rip),%ymm9 # 5074 <_sk_callback_hsw+0x510> + DB 196,98,125,24,21,206,5,0,0 ; vbroadcastss 0x5ce(%rip),%ymm10 # 5078 <_sk_callback_hsw+0x514> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,186,5,0,0 ; vbroadcastss 0x5ba(%rip),%ymm9 # 5044 <_sk_callback_hsw+0x50e> + DB 196,98,125,24,13,196,5,0,0 ; vbroadcastss 0x5c4(%rip),%ymm9 # 507c <_sk_callback_hsw+0x518> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,176,5,0,0 ; vbroadcastss 0x5b0(%rip),%ymm10 # 5048 <_sk_callback_hsw+0x512> + DB 196,98,125,24,21,186,5,0,0 ; vbroadcastss 0x5ba(%rip),%ymm10 # 5080 <_sk_callback_hsw+0x51c> 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 @@ -4504,14 +4512,14 @@ _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 196,98,125,24,5,152,5,0,0 ; vbroadcastss 0x598(%rip),%ymm8 # 504c <_sk_callback_hsw+0x516> + DB 196,98,125,24,5,162,5,0,0 ; vbroadcastss 0x5a2(%rip),%ymm8 # 5084 <_sk_callback_hsw+0x520> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,137,5,0,0 ; vbroadcastss 0x589(%rip),%ymm10 # 5050 <_sk_callback_hsw+0x51a> - DB 196,98,125,24,29,132,5,0,0 ; vbroadcastss 0x584(%rip),%ymm11 # 5054 <_sk_callback_hsw+0x51e> + DB 196,98,125,24,21,147,5,0,0 ; vbroadcastss 0x593(%rip),%ymm10 # 5088 <_sk_callback_hsw+0x524> + DB 196,98,125,24,29,142,5,0,0 ; vbroadcastss 0x58e(%rip),%ymm11 # 508c <_sk_callback_hsw+0x528> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,117,5,0,0 ; vbroadcastss 0x575(%rip),%ymm8 # 5058 <_sk_callback_hsw+0x522> + DB 196,98,125,24,5,127,5,0,0 ; vbroadcastss 0x57f(%rip),%ymm8 # 5090 <_sk_callback_hsw+0x52c> 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 @@ -4520,12 +4528,12 @@ _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 196,226,125,24,13,93,5,0,0 ; vbroadcastss 0x55d(%rip),%ymm1 # 505c <_sk_callback_hsw+0x526> + DB 196,226,125,24,13,103,5,0,0 ; vbroadcastss 0x567(%rip),%ymm1 # 5094 <_sk_callback_hsw+0x530> 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 196,98,125,24,21,73,5,0,0 ; vbroadcastss 0x549(%rip),%ymm10 # 5060 <_sk_callback_hsw+0x52a> - DB 196,98,125,24,29,68,5,0,0 ; vbroadcastss 0x544(%rip),%ymm11 # 5064 <_sk_callback_hsw+0x52e> + DB 196,98,125,24,21,83,5,0,0 ; vbroadcastss 0x553(%rip),%ymm10 # 5098 <_sk_callback_hsw+0x534> + DB 196,98,125,24,29,78,5,0,0 ; vbroadcastss 0x54e(%rip),%ymm11 # 509c <_sk_callback_hsw+0x538> 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) @@ -4647,25 +4655,25 @@ ALIGN 4 DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4d11 <.literal4+0xb1> + DB 71,225,61 ; rex.RXB loope 4d41 <.literal4+0xb1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4d21 <.literal4+0xc1> + DB 71,225,61 ; rex.RXB loope 4d51 <.literal4+0xc1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4d31 <.literal4+0xd1> + DB 71,225,61 ; rex.RXB loope 4d61 <.literal4+0xd1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 4d41 <.literal4+0xe1> + DB 71,225,61 ; rex.RXB loope 4d71 <.literal4+0xe1> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) @@ -4715,7 +4723,7 @@ 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 4d99 <.literal4+0x139> + DB 224,7 ; loopne 4dc9 <.literal4+0x139> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -4731,10 +4739,10 @@ 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 4dc0 <.literal4+0x160> + DB 127,0 ; jg 4df0 <.literal4+0x160> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4e39 <.literal4+0x1d9> + DB 119,115 ; ja 4e69 <.literal4+0x1d9> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4748,10 +4756,10 @@ 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 4df4 <.literal4+0x194> + DB 127,0 ; jg 4e24 <.literal4+0x194> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4e6d <.literal4+0x20d> + DB 119,115 ; ja 4e9d <.literal4+0x20d> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4765,10 +4773,10 @@ 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 4e28 <.literal4+0x1c8> + DB 127,0 ; jg 4e58 <.literal4+0x1c8> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4ea1 <.literal4+0x241> + DB 119,115 ; ja 4ed1 <.literal4+0x241> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4782,10 +4790,10 @@ 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 4e5c <.literal4+0x1fc> + DB 127,0 ; jg 4e8c <.literal4+0x1fc> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 4ed5 <.literal4+0x275> + DB 119,115 ; ja 4f05 <.literal4+0x275> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -4798,7 +4806,7 @@ 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 4ed3 <.literal4+0x273> + DB 127,67 ; jg 4f03 <.literal4+0x273> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -4810,10 +4818,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 4ef3 <.literal4+0x293> + DB 118,63 ; jbe 4f23 <.literal4+0x293> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 4f07 <.literal4+0x2a7> + DB 127,67 ; jg 4f37 <.literal4+0x2a7> DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax) DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax) @@ -4822,7 +4830,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4ee9 <.literal4+0x289> + DB 224,7 ; loopne 4f19 <.literal4+0x289> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -4834,7 +4842,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 4f05 <.literal4+0x2a5> + DB 224,7 ; loopne 4f35 <.literal4+0x2a5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -4845,7 +4853,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 4f5a <.literal4+0x2fa> + DB 124,66 ; jl 4f8a <.literal4+0x2fa> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -4863,9 +4871,9 @@ 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 4f9d <.literal4+0x33d> + DB 112,65 ; jo 4fcd <.literal4+0x33d> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 4fab <.literal4+0x34b> + DB 127,67 ; jg 4fdb <.literal4+0x34b> DB 128,0,128 ; addb $0x80,(%rax) DB 55 ; (bad) DB 128,0,128 ; addb $0x80,(%rax) @@ -4873,7 +4881,11 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 4fbf <.literal4+0x35f> + DB 127,71 ; jg 4fef <.literal4+0x35f> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,191,0,0,0 ; add %al,0xbf(%rax) @@ -4981,16 +4993,16 @@ 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) # a0050a8 <_sk_callback_hsw+0xa000572> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0050c8 <_sk_callback_hsw+0xa000564> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120050b0 <_sk_callback_hsw+0x1200057a> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120050d0 <_sk_callback_hsw+0x1200056c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0050b8 <_sk_callback_hsw+0x1a000582> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0050d8 <_sk_callback_hsw+0x1a000574> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30050c0 <_sk_callback_hsw+0x300058a> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30050e0 <_sk_callback_hsw+0x300057c> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5033,16 +5045,16 @@ 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) # a005108 <_sk_callback_hsw+0xa0005d2> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005128 <_sk_callback_hsw+0xa0005c4> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005110 <_sk_callback_hsw+0x120005da> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005130 <_sk_callback_hsw+0x120005cc> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005118 <_sk_callback_hsw+0x1a0005e2> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005138 <_sk_callback_hsw+0x1a0005d4> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005120 <_sk_callback_hsw+0x30005ea> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005140 <_sk_callback_hsw+0x30005dc> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5085,16 +5097,16 @@ 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) # a005168 <_sk_callback_hsw+0xa000632> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005188 <_sk_callback_hsw+0xa000624> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005170 <_sk_callback_hsw+0x1200063a> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005190 <_sk_callback_hsw+0x1200062c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005178 <_sk_callback_hsw+0x1a000642> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005198 <_sk_callback_hsw+0x1a000634> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005180 <_sk_callback_hsw+0x300064a> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30051a0 <_sk_callback_hsw+0x300063c> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5137,16 +5149,16 @@ 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) # a0051c8 <_sk_callback_hsw+0xa000692> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0051e8 <_sk_callback_hsw+0xa000684> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120051d0 <_sk_callback_hsw+0x1200069a> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120051f0 <_sk_callback_hsw+0x1200068c> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0051d8 <_sk_callback_hsw+0x1a0006a2> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0051f8 <_sk_callback_hsw+0x1a000694> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30051e0 <_sk_callback_hsw+0x30006aa> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005200 <_sk_callback_hsw+0x300069c> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5189,16 +5201,16 @@ 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) # a005228 <_sk_callback_hsw+0xa0006f2> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005248 <_sk_callback_hsw+0xa0006e4> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005230 <_sk_callback_hsw+0x120006fa> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12005250 <_sk_callback_hsw+0x120006ec> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005238 <_sk_callback_hsw+0x1a000702> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a005258 <_sk_callback_hsw+0x1a0006f4> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005240 <_sk_callback_hsw+0x300070a> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3005260 <_sk_callback_hsw+0x30006fc> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -5376,7 +5388,7 @@ _sk_seed_shader_avx LABEL PROC DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,59,103,0,0 ; vbroadcastss 0x673b(%rip),%ymm1 # 68b4 <_sk_callback_avx+0x12c> + DB 196,226,125,24,13,119,103,0,0 ; vbroadcastss 0x6777(%rip),%ymm1 # 68f0 <_sk_callback_avx+0x12a> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,88,7 ; vaddps (%rdi),%ymm0,%ymm0 DB 197,249,110,209 ; vmovd %ecx,%xmm2 @@ -5385,7 +5397,7 @@ _sk_seed_shader_avx LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,21,103,0,0 ; vbroadcastss 0x6715(%rip),%ymm2 # 68b8 <_sk_callback_avx+0x130> + DB 196,226,125,24,21,81,103,0,0 ; vbroadcastss 0x6751(%rip),%ymm2 # 68f4 <_sk_callback_avx+0x12e> 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 @@ -5407,7 +5419,7 @@ _sk_dither_avx LABEL PROC DB 196,65,121,112,201,0 ; vpshufd $0x0,%xmm9,%xmm9 DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 DB 196,65,52,87,208 ; vxorps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,190,102,0,0 ; vbroadcastss 0x66be(%rip),%ymm11 # 68bc <_sk_callback_avx+0x134> + DB 196,98,125,24,29,250,102,0,0 ; vbroadcastss 0x66fa(%rip),%ymm11 # 68f8 <_sk_callback_avx+0x132> DB 196,65,44,84,203 ; vandps %ymm11,%ymm10,%ymm9 DB 196,193,25,114,241,5 ; vpslld $0x5,%xmm9,%xmm12 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 @@ -5418,8 +5430,8 @@ _sk_dither_avx LABEL PROC DB 196,67,125,25,219,1 ; vextractf128 $0x1,%ymm11,%xmm11 DB 196,193,33,114,243,4 ; vpslld $0x4,%xmm11,%xmm11 DB 196,67,29,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm12,%ymm11 - DB 196,98,125,24,37,127,102,0,0 ; vbroadcastss 0x667f(%rip),%ymm12 # 68c0 <_sk_callback_avx+0x138> - DB 196,98,125,24,45,122,102,0,0 ; vbroadcastss 0x667a(%rip),%ymm13 # 68c4 <_sk_callback_avx+0x13c> + DB 196,98,125,24,37,187,102,0,0 ; vbroadcastss 0x66bb(%rip),%ymm12 # 68fc <_sk_callback_avx+0x136> + DB 196,98,125,24,45,182,102,0,0 ; vbroadcastss 0x66b6(%rip),%ymm13 # 6900 <_sk_callback_avx+0x13a> DB 196,65,44,84,245 ; vandps %ymm13,%ymm10,%ymm14 DB 196,193,1,114,246,2 ; vpslld $0x2,%xmm14,%xmm15 DB 196,67,125,25,246,1 ; vextractf128 $0x1,%ymm14,%xmm14 @@ -5446,9 +5458,9 @@ _sk_dither_avx LABEL PROC DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8 DB 196,65,60,86,194 ; vorps %ymm10,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,229,101,0,0 ; vbroadcastss 0x65e5(%rip),%ymm9 # 68c8 <_sk_callback_avx+0x140> + DB 196,98,125,24,13,33,102,0,0 ; vbroadcastss 0x6621(%rip),%ymm9 # 6904 <_sk_callback_avx+0x13e> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,219,101,0,0 ; vbroadcastss 0x65db(%rip),%ymm9 # 68cc <_sk_callback_avx+0x144> + DB 196,98,125,24,13,23,102,0,0 ; vbroadcastss 0x6617(%rip),%ymm9 # 6908 <_sk_callback_avx+0x142> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 @@ -5507,7 +5519,7 @@ _sk_clear_avx LABEL PROC PUBLIC _sk_srcatop_avx _sk_srcatop_avx LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,51,101,0,0 ; vbroadcastss 0x6533(%rip),%ymm8 # 68d0 <_sk_callback_avx+0x148> + DB 196,98,125,24,5,111,101,0,0 ; vbroadcastss 0x656f(%rip),%ymm8 # 690c <_sk_callback_avx+0x146> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -5526,7 +5538,7 @@ _sk_srcatop_avx LABEL PROC PUBLIC _sk_dstatop_avx _sk_dstatop_avx LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 - DB 196,98,125,24,13,245,100,0,0 ; vbroadcastss 0x64f5(%rip),%ymm9 # 68d4 <_sk_callback_avx+0x14c> + DB 196,98,125,24,13,49,101,0,0 ; vbroadcastss 0x6531(%rip),%ymm9 # 6910 <_sk_callback_avx+0x14a> DB 197,52,92,207 ; vsubps %ymm7,%ymm9,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 @@ -5562,7 +5574,7 @@ _sk_dstin_avx LABEL PROC PUBLIC _sk_srcout_avx _sk_srcout_avx LABEL PROC - DB 196,98,125,24,5,148,100,0,0 ; vbroadcastss 0x6494(%rip),%ymm8 # 68d8 <_sk_callback_avx+0x150> + DB 196,98,125,24,5,208,100,0,0 ; vbroadcastss 0x64d0(%rip),%ymm8 # 6914 <_sk_callback_avx+0x14e> 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 @@ -5573,7 +5585,7 @@ _sk_srcout_avx LABEL PROC PUBLIC _sk_dstout_avx _sk_dstout_avx LABEL PROC - DB 196,226,125,24,5,119,100,0,0 ; vbroadcastss 0x6477(%rip),%ymm0 # 68dc <_sk_callback_avx+0x154> + DB 196,226,125,24,5,179,100,0,0 ; vbroadcastss 0x64b3(%rip),%ymm0 # 6918 <_sk_callback_avx+0x152> 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 @@ -5584,7 +5596,7 @@ _sk_dstout_avx LABEL PROC PUBLIC _sk_srcover_avx _sk_srcover_avx LABEL PROC - DB 196,98,125,24,5,90,100,0,0 ; vbroadcastss 0x645a(%rip),%ymm8 # 68e0 <_sk_callback_avx+0x158> + DB 196,98,125,24,5,150,100,0,0 ; vbroadcastss 0x6496(%rip),%ymm8 # 691c <_sk_callback_avx+0x156> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -5599,7 +5611,7 @@ _sk_srcover_avx LABEL PROC PUBLIC _sk_dstover_avx _sk_dstover_avx LABEL PROC - DB 196,98,125,24,5,45,100,0,0 ; vbroadcastss 0x642d(%rip),%ymm8 # 68e4 <_sk_callback_avx+0x15c> + DB 196,98,125,24,5,105,100,0,0 ; vbroadcastss 0x6469(%rip),%ymm8 # 6920 <_sk_callback_avx+0x15a> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0 @@ -5623,7 +5635,7 @@ _sk_modulate_avx LABEL PROC PUBLIC _sk_multiply_avx _sk_multiply_avx LABEL PROC - DB 196,98,125,24,5,236,99,0,0 ; vbroadcastss 0x63ec(%rip),%ymm8 # 68e8 <_sk_callback_avx+0x160> + DB 196,98,125,24,5,40,100,0,0 ; vbroadcastss 0x6428(%rip),%ymm8 # 6924 <_sk_callback_avx+0x15e> 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 @@ -5677,7 +5689,7 @@ _sk_screen_avx LABEL PROC PUBLIC _sk_xor__avx _sk_xor__avx LABEL PROC - DB 196,98,125,24,5,59,99,0,0 ; vbroadcastss 0x633b(%rip),%ymm8 # 68ec <_sk_callback_avx+0x164> + DB 196,98,125,24,5,119,99,0,0 ; vbroadcastss 0x6377(%rip),%ymm8 # 6928 <_sk_callback_avx+0x162> 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 @@ -5712,7 +5724,7 @@ _sk_darken_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,187,98,0,0 ; vbroadcastss 0x62bb(%rip),%ymm8 # 68f0 <_sk_callback_avx+0x168> + DB 196,98,125,24,5,247,98,0,0 ; vbroadcastss 0x62f7(%rip),%ymm8 # 692c <_sk_callback_avx+0x166> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5736,7 +5748,7 @@ _sk_lighten_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,103,98,0,0 ; vbroadcastss 0x6267(%rip),%ymm8 # 68f4 <_sk_callback_avx+0x16c> + DB 196,98,125,24,5,163,98,0,0 ; vbroadcastss 0x62a3(%rip),%ymm8 # 6930 <_sk_callback_avx+0x16a> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5763,7 +5775,7 @@ _sk_difference_avx LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,7,98,0,0 ; vbroadcastss 0x6207(%rip),%ymm8 # 68f8 <_sk_callback_avx+0x170> + DB 196,98,125,24,5,67,98,0,0 ; vbroadcastss 0x6243(%rip),%ymm8 # 6934 <_sk_callback_avx+0x16e> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5784,7 +5796,7 @@ _sk_exclusion_avx LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,194,97,0,0 ; vbroadcastss 0x61c2(%rip),%ymm8 # 68fc <_sk_callback_avx+0x174> + DB 196,98,125,24,5,254,97,0,0 ; vbroadcastss 0x61fe(%rip),%ymm8 # 6938 <_sk_callback_avx+0x172> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -5793,7 +5805,7 @@ _sk_exclusion_avx LABEL PROC PUBLIC _sk_colorburn_avx _sk_colorburn_avx LABEL PROC - DB 196,98,125,24,5,173,97,0,0 ; vbroadcastss 0x61ad(%rip),%ymm8 # 6900 <_sk_callback_avx+0x178> + DB 196,98,125,24,5,233,97,0,0 ; vbroadcastss 0x61e9(%rip),%ymm8 # 693c <_sk_callback_avx+0x176> 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 @@ -5853,7 +5865,7 @@ _sk_colorburn_avx LABEL PROC PUBLIC _sk_colordodge_avx _sk_colordodge_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,169,96,0,0 ; vbroadcastss 0x60a9(%rip),%ymm9 # 6904 <_sk_callback_avx+0x17c> + DB 196,98,125,24,13,229,96,0,0 ; vbroadcastss 0x60e5(%rip),%ymm9 # 6940 <_sk_callback_avx+0x17a> 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 @@ -5908,7 +5920,7 @@ _sk_colordodge_avx LABEL PROC PUBLIC _sk_hardlight_avx _sk_hardlight_avx LABEL PROC - DB 196,98,125,24,5,187,95,0,0 ; vbroadcastss 0x5fbb(%rip),%ymm8 # 6908 <_sk_callback_avx+0x180> + DB 196,98,125,24,5,247,95,0,0 ; vbroadcastss 0x5ff7(%rip),%ymm8 # 6944 <_sk_callback_avx+0x17e> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5961,7 +5973,7 @@ _sk_hardlight_avx LABEL PROC PUBLIC _sk_overlay_avx _sk_overlay_avx LABEL PROC - DB 196,98,125,24,5,228,94,0,0 ; vbroadcastss 0x5ee4(%rip),%ymm8 # 690c <_sk_callback_avx+0x184> + DB 196,98,125,24,5,32,95,0,0 ; vbroadcastss 0x5f20(%rip),%ymm8 # 6948 <_sk_callback_avx+0x182> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6026,10 +6038,10 @@ _sk_softlight_avx LABEL PROC DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,65,60,89,216 ; vmulps %ymm8,%ymm8,%ymm11 DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,29,215,93,0,0 ; vbroadcastss 0x5dd7(%rip),%ymm11 # 6914 <_sk_callback_avx+0x18c> + DB 196,98,125,24,29,19,94,0,0 ; vbroadcastss 0x5e13(%rip),%ymm11 # 6950 <_sk_callback_avx+0x18a> DB 196,65,28,88,235 ; vaddps %ymm11,%ymm12,%ymm13 DB 196,65,20,89,192 ; vmulps %ymm8,%ymm13,%ymm8 - DB 196,98,125,24,45,200,93,0,0 ; vbroadcastss 0x5dc8(%rip),%ymm13 # 6918 <_sk_callback_avx+0x190> + DB 196,98,125,24,45,4,94,0,0 ; vbroadcastss 0x5e04(%rip),%ymm13 # 6954 <_sk_callback_avx+0x18e> DB 196,65,28,89,245 ; vmulps %ymm13,%ymm12,%ymm14 DB 196,65,12,88,192 ; vaddps %ymm8,%ymm14,%ymm8 DB 196,65,124,82,244 ; vrsqrtps %ymm12,%ymm14 @@ -6040,7 +6052,7 @@ _sk_softlight_avx LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,134,93,0,0 ; vbroadcastss 0x5d86(%rip),%ymm8 # 6910 <_sk_callback_avx+0x188> + DB 196,98,125,24,5,194,93,0,0 ; vbroadcastss 0x5dc2(%rip),%ymm8 # 694c <_sk_callback_avx+0x186> DB 196,65,60,92,228 ; vsubps %ymm12,%ymm8,%ymm12 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 @@ -6167,12 +6179,12 @@ _sk_hue_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,80,91,0,0 ; vbroadcastss 0x5b50(%rip),%ymm14 # 691c <_sk_callback_avx+0x194> + DB 196,98,125,24,53,140,91,0,0 ; vbroadcastss 0x5b8c(%rip),%ymm14 # 6958 <_sk_callback_avx+0x192> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,70,91,0,0 ; vbroadcastss 0x5b46(%rip),%ymm15 # 6920 <_sk_callback_avx+0x198> + DB 196,98,125,24,61,130,91,0,0 ; vbroadcastss 0x5b82(%rip),%ymm15 # 695c <_sk_callback_avx+0x196> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,55,91,0,0 ; vbroadcastss 0x5b37(%rip),%ymm0 # 6924 <_sk_callback_avx+0x19c> + DB 196,226,125,24,5,115,91,0,0 ; vbroadcastss 0x5b73(%rip),%ymm0 # 6960 <_sk_callback_avx+0x19a> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -6233,7 +6245,7 @@ _sk_hue_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,16,90,0,0 ; vbroadcastss 0x5a10(%rip),%ymm8 # 6928 <_sk_callback_avx+0x1a0> + DB 196,98,125,24,5,76,90,0,0 ; vbroadcastss 0x5a4c(%rip),%ymm8 # 6964 <_sk_callback_avx+0x19e> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6290,12 +6302,12 @@ _sk_saturation_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,24,89,0,0 ; vbroadcastss 0x5918(%rip),%ymm14 # 692c <_sk_callback_avx+0x1a4> + DB 196,98,125,24,53,84,89,0,0 ; vbroadcastss 0x5954(%rip),%ymm14 # 6968 <_sk_callback_avx+0x1a2> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,14,89,0,0 ; vbroadcastss 0x590e(%rip),%ymm15 # 6930 <_sk_callback_avx+0x1a8> + DB 196,98,125,24,61,74,89,0,0 ; vbroadcastss 0x594a(%rip),%ymm15 # 696c <_sk_callback_avx+0x1a6> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,255,88,0,0 ; vbroadcastss 0x58ff(%rip),%ymm0 # 6934 <_sk_callback_avx+0x1ac> + DB 196,226,125,24,5,59,89,0,0 ; vbroadcastss 0x593b(%rip),%ymm0 # 6970 <_sk_callback_avx+0x1aa> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -6356,7 +6368,7 @@ _sk_saturation_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,216,87,0,0 ; vbroadcastss 0x57d8(%rip),%ymm8 # 6938 <_sk_callback_avx+0x1b0> + DB 196,98,125,24,5,20,88,0,0 ; vbroadcastss 0x5814(%rip),%ymm8 # 6974 <_sk_callback_avx+0x1ae> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6385,12 +6397,12 @@ _sk_color_avx LABEL PROC DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp) DB 197,124,89,199 ; vmulps %ymm7,%ymm0,%ymm8 DB 197,116,89,207 ; vmulps %ymm7,%ymm1,%ymm9 - DB 196,98,125,24,45,104,87,0,0 ; vbroadcastss 0x5768(%rip),%ymm13 # 693c <_sk_callback_avx+0x1b4> + DB 196,98,125,24,45,164,87,0,0 ; vbroadcastss 0x57a4(%rip),%ymm13 # 6978 <_sk_callback_avx+0x1b2> DB 196,65,92,89,213 ; vmulps %ymm13,%ymm4,%ymm10 - DB 196,98,125,24,53,94,87,0,0 ; vbroadcastss 0x575e(%rip),%ymm14 # 6940 <_sk_callback_avx+0x1b8> + DB 196,98,125,24,53,154,87,0,0 ; vbroadcastss 0x579a(%rip),%ymm14 # 697c <_sk_callback_avx+0x1b6> DB 196,65,84,89,222 ; vmulps %ymm14,%ymm5,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,79,87,0,0 ; vbroadcastss 0x574f(%rip),%ymm15 # 6944 <_sk_callback_avx+0x1bc> + DB 196,98,125,24,61,139,87,0,0 ; vbroadcastss 0x578b(%rip),%ymm15 # 6980 <_sk_callback_avx+0x1ba> DB 196,65,76,89,223 ; vmulps %ymm15,%ymm6,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -6453,7 +6465,7 @@ _sk_color_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,22,86,0,0 ; vbroadcastss 0x5616(%rip),%ymm0 # 6948 <_sk_callback_avx+0x1c0> + DB 196,226,125,24,5,82,86,0,0 ; vbroadcastss 0x5652(%rip),%ymm0 # 6984 <_sk_callback_avx+0x1be> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,84,36,32 ; vmulps 0x20(%rsp),%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -6483,12 +6495,12 @@ _sk_luminosity_avx LABEL PROC DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 - DB 196,98,125,24,45,162,85,0,0 ; vbroadcastss 0x55a2(%rip),%ymm13 # 694c <_sk_callback_avx+0x1c4> + DB 196,98,125,24,45,222,85,0,0 ; vbroadcastss 0x55de(%rip),%ymm13 # 6988 <_sk_callback_avx+0x1c2> DB 196,65,108,89,213 ; vmulps %ymm13,%ymm2,%ymm10 - DB 196,98,125,24,53,152,85,0,0 ; vbroadcastss 0x5598(%rip),%ymm14 # 6950 <_sk_callback_avx+0x1c8> + DB 196,98,125,24,53,212,85,0,0 ; vbroadcastss 0x55d4(%rip),%ymm14 # 698c <_sk_callback_avx+0x1c6> DB 196,65,116,89,222 ; vmulps %ymm14,%ymm1,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,137,85,0,0 ; vbroadcastss 0x5589(%rip),%ymm15 # 6954 <_sk_callback_avx+0x1cc> + DB 196,98,125,24,61,197,85,0,0 ; vbroadcastss 0x55c5(%rip),%ymm15 # 6990 <_sk_callback_avx+0x1ca> DB 196,65,28,89,223 ; vmulps %ymm15,%ymm12,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -6551,7 +6563,7 @@ _sk_luminosity_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,80,84,0,0 ; vbroadcastss 0x5450(%rip),%ymm0 # 6958 <_sk_callback_avx+0x1d0> + DB 196,226,125,24,5,140,84,0,0 ; vbroadcastss 0x548c(%rip),%ymm0 # 6994 <_sk_callback_avx+0x1ce> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -6582,7 +6594,7 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 15,133,47,1,0,0 ; jne 16a2 <_sk_srcover_rgba_8888_avx+0x14c> DB 196,193,124,16,58 ; vmovups (%r10),%ymm7 - DB 197,124,40,13,96,88,0,0 ; vmovaps 0x5860(%rip),%ymm9 # 6de0 <_sk_callback_avx+0x658> + DB 197,124,40,13,160,88,0,0 ; vmovaps 0x58a0(%rip),%ymm9 # 6e20 <_sk_callback_avx+0x65a> DB 196,193,68,84,225 ; vandps %ymm9,%ymm7,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 197,209,114,215,8 ; vpsrld $0x8,%xmm7,%xmm5 @@ -6600,9 +6612,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7 DB 196,227,53,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm9,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,123,83,0,0 ; vbroadcastss 0x537b(%rip),%ymm8 # 695c <_sk_callback_avx+0x1d4> + DB 196,98,125,24,5,183,83,0,0 ; vbroadcastss 0x53b7(%rip),%ymm8 # 6998 <_sk_callback_avx+0x1d2> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,114,83,0,0 ; vbroadcastss 0x5372(%rip),%ymm9 # 6960 <_sk_callback_avx+0x1d8> + DB 196,98,125,24,13,174,83,0,0 ; vbroadcastss 0x53ae(%rip),%ymm9 # 699c <_sk_callback_avx+0x1d6> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,60,89,212 ; vmulps %ymm4,%ymm8,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 @@ -6648,9 +6660,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,224 ; vmovq %rax,%xmm4 DB 196,226,121,48,228 ; vpmovzxbw %xmm4,%xmm4 - DB 196,226,89,0,45,38,86,0,0 ; vpshufb 0x5626(%rip),%xmm4,%xmm5 # 6cf0 <_sk_callback_avx+0x568> + DB 196,226,89,0,45,102,86,0,0 ; vpshufb 0x5666(%rip),%xmm4,%xmm5 # 6d30 <_sk_callback_avx+0x56a> DB 196,226,121,33,237 ; vpmovsxbd %xmm5,%xmm5 - DB 196,226,89,0,37,40,86,0,0 ; vpshufb 0x5628(%rip),%xmm4,%xmm4 # 6d00 <_sk_callback_avx+0x578> + DB 196,226,89,0,37,104,86,0,0 ; vpshufb 0x5668(%rip),%xmm4,%xmm4 # 6d40 <_sk_callback_avx+0x57a> DB 196,226,121,33,228 ; vpmovsxbd %xmm4,%xmm4 DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 DB 196,194,93,44,58 ; vmaskmovps (%r10),%ymm4,%ymm7 @@ -6662,9 +6674,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 - DB 196,98,49,0,21,219,85,0,0 ; vpshufb 0x55db(%rip),%xmm9,%xmm10 # 6cf0 <_sk_callback_avx+0x568> + DB 196,98,49,0,21,27,86,0,0 ; vpshufb 0x561b(%rip),%xmm9,%xmm10 # 6d30 <_sk_callback_avx+0x56a> DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10 - DB 196,98,49,0,13,221,85,0,0 ; vpshufb 0x55dd(%rip),%xmm9,%xmm9 # 6d00 <_sk_callback_avx+0x578> + DB 196,98,49,0,13,29,86,0,0 ; vpshufb 0x561d(%rip),%xmm9,%xmm9 # 6d40 <_sk_callback_avx+0x57a> DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10) @@ -6682,7 +6694,7 @@ _sk_clamp_0_avx LABEL PROC PUBLIC _sk_clamp_1_avx _sk_clamp_1_avx LABEL PROC - DB 196,98,125,24,5,6,82,0,0 ; vbroadcastss 0x5206(%rip),%ymm8 # 6964 <_sk_callback_avx+0x1dc> + DB 196,98,125,24,5,66,82,0,0 ; vbroadcastss 0x5242(%rip),%ymm8 # 69a0 <_sk_callback_avx+0x1da> 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 @@ -6692,7 +6704,7 @@ _sk_clamp_1_avx LABEL PROC PUBLIC _sk_clamp_a_avx _sk_clamp_a_avx LABEL PROC - DB 196,98,125,24,5,233,81,0,0 ; vbroadcastss 0x51e9(%rip),%ymm8 # 6968 <_sk_callback_avx+0x1e0> + DB 196,98,125,24,5,37,82,0,0 ; vbroadcastss 0x5225(%rip),%ymm8 # 69a4 <_sk_callback_avx+0x1de> 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 @@ -6764,7 +6776,7 @@ PUBLIC _sk_unpremul_avx _sk_unpremul_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,49,81,0,0 ; vbroadcastss 0x5131(%rip),%ymm10 # 696c <_sk_callback_avx+0x1e4> + DB 196,98,125,24,21,109,81,0,0 ; vbroadcastss 0x516d(%rip),%ymm10 # 69a8 <_sk_callback_avx+0x1e2> 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 @@ -6775,17 +6787,17 @@ _sk_unpremul_avx LABEL PROC PUBLIC _sk_from_srgb_avx _sk_from_srgb_avx LABEL PROC - DB 196,98,125,24,5,18,81,0,0 ; vbroadcastss 0x5112(%rip),%ymm8 # 6970 <_sk_callback_avx+0x1e8> + DB 196,98,125,24,5,78,81,0,0 ; vbroadcastss 0x514e(%rip),%ymm8 # 69ac <_sk_callback_avx+0x1e6> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,4,81,0,0 ; vbroadcastss 0x5104(%rip),%ymm11 # 6974 <_sk_callback_avx+0x1ec> + DB 196,98,125,24,29,64,81,0,0 ; vbroadcastss 0x5140(%rip),%ymm11 # 69b0 <_sk_callback_avx+0x1ea> DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12 - DB 196,98,125,24,45,250,80,0,0 ; vbroadcastss 0x50fa(%rip),%ymm13 # 6978 <_sk_callback_avx+0x1f0> + DB 196,98,125,24,45,54,81,0,0 ; vbroadcastss 0x5136(%rip),%ymm13 # 69b4 <_sk_callback_avx+0x1ee> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,235,80,0,0 ; vbroadcastss 0x50eb(%rip),%ymm12 # 697c <_sk_callback_avx+0x1f4> + DB 196,98,125,24,37,39,81,0,0 ; vbroadcastss 0x5127(%rip),%ymm12 # 69b8 <_sk_callback_avx+0x1f2> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,53,225,80,0,0 ; vbroadcastss 0x50e1(%rip),%ymm14 # 6980 <_sk_callback_avx+0x1f8> + DB 196,98,125,24,53,29,81,0,0 ; vbroadcastss 0x511d(%rip),%ymm14 # 69bc <_sk_callback_avx+0x1f6> DB 196,193,124,194,198,1 ; vcmpltps %ymm14,%ymm0,%ymm0 DB 196,195,45,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm10,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -6810,20 +6822,20 @@ _sk_from_srgb_avx LABEL PROC PUBLIC _sk_to_srgb_avx _sk_to_srgb_avx LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,118,80,0,0 ; vbroadcastss 0x5076(%rip),%ymm8 # 6984 <_sk_callback_avx+0x1fc> + DB 196,98,125,24,5,178,80,0,0 ; vbroadcastss 0x50b2(%rip),%ymm8 # 69c0 <_sk_callback_avx+0x1fa> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,108,80,0,0 ; vbroadcastss 0x506c(%rip),%ymm11 # 6988 <_sk_callback_avx+0x200> + DB 196,98,125,24,29,168,80,0,0 ; vbroadcastss 0x50a8(%rip),%ymm11 # 69c4 <_sk_callback_avx+0x1fe> DB 196,65,52,89,227 ; vmulps %ymm11,%ymm9,%ymm12 - DB 196,98,125,24,45,98,80,0,0 ; vbroadcastss 0x5062(%rip),%ymm13 # 698c <_sk_callback_avx+0x204> + DB 196,98,125,24,45,158,80,0,0 ; vbroadcastss 0x509e(%rip),%ymm13 # 69c8 <_sk_callback_avx+0x202> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,52,89,228 ; vmulps %ymm12,%ymm9,%ymm12 - DB 196,98,125,24,53,83,80,0,0 ; vbroadcastss 0x5053(%rip),%ymm14 # 6990 <_sk_callback_avx+0x208> + DB 196,98,125,24,53,143,80,0,0 ; vbroadcastss 0x508f(%rip),%ymm14 # 69cc <_sk_callback_avx+0x206> DB 196,65,28,88,230 ; vaddps %ymm14,%ymm12,%ymm12 - DB 196,98,125,24,61,73,80,0,0 ; vbroadcastss 0x5049(%rip),%ymm15 # 6994 <_sk_callback_avx+0x20c> + DB 196,98,125,24,61,133,80,0,0 ; vbroadcastss 0x5085(%rip),%ymm15 # 69d0 <_sk_callback_avx+0x20a> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 - DB 196,98,125,24,37,53,80,0,0 ; vbroadcastss 0x5035(%rip),%ymm12 # 6998 <_sk_callback_avx+0x210> + DB 196,98,125,24,37,113,80,0,0 ; vbroadcastss 0x5071(%rip),%ymm12 # 69d4 <_sk_callback_avx+0x20e> DB 196,193,124,194,196,1 ; vcmpltps %ymm12,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -6858,7 +6870,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9 DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,155,79,0,0 ; vbroadcastss 0x4f9b(%rip),%ymm11 # 699c <_sk_callback_avx+0x214> + DB 196,98,125,24,29,215,79,0,0 ; vbroadcastss 0x4fd7(%rip),%ymm11 # 69d8 <_sk_callback_avx+0x212> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 196,65,28,89,227 ; vmulps %ymm11,%ymm12,%ymm12 @@ -6868,19 +6880,19 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 DB 196,193,116,89,203 ; vmulps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,116,79,0,0 ; vbroadcastss 0x4f74(%rip),%ymm11 # 69a8 <_sk_callback_avx+0x220> + DB 196,98,125,24,29,176,79,0,0 ; vbroadcastss 0x4fb0(%rip),%ymm11 # 69e4 <_sk_callback_avx+0x21e> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,98,79,0,0 ; vbroadcastss 0x4f62(%rip),%ymm11 # 69a4 <_sk_callback_avx+0x21c> + DB 196,98,125,24,29,158,79,0,0 ; vbroadcastss 0x4f9e(%rip),%ymm11 # 69e0 <_sk_callback_avx+0x21a> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,227,117,74,202,224 ; vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,74,79,0,0 ; vbroadcastss 0x4f4a(%rip),%ymm2 # 69a0 <_sk_callback_avx+0x218> + DB 196,226,125,24,21,134,79,0,0 ; vbroadcastss 0x4f86(%rip),%ymm2 # 69dc <_sk_callback_avx+0x216> DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 DB 196,227,117,74,194,0 ; vblendvps %ymm0,%ymm2,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,37,49,79,0,0 ; vbroadcastss 0x4f31(%rip),%ymm12 # 69b0 <_sk_callback_avx+0x228> + DB 196,98,125,24,37,109,79,0,0 ; vbroadcastss 0x4f6d(%rip),%ymm12 # 69ec <_sk_callback_avx+0x226> DB 196,193,116,89,212 ; vmulps %ymm12,%ymm1,%ymm2 DB 197,28,194,226,1 ; vcmpltps %ymm2,%ymm12,%ymm12 DB 196,65,36,92,216 ; vsubps %ymm8,%ymm11,%ymm11 @@ -6890,7 +6902,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,198,128 ; vblendvps %ymm8,%ymm14,%ymm0,%ymm0 DB 196,195,117,74,206,128 ; vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - DB 196,98,125,24,5,244,78,0,0 ; vbroadcastss 0x4ef4(%rip),%ymm8 # 69ac <_sk_callback_avx+0x224> + DB 196,98,125,24,5,48,79,0,0 ; vbroadcastss 0x4f30(%rip),%ymm8 # 69e8 <_sk_callback_avx+0x222> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6905,7 +6917,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,225 ; vmovaps %ymm1,%ymm4 DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 - DB 196,98,125,24,5,187,78,0,0 ; vbroadcastss 0x4ebb(%rip),%ymm8 # 69b4 <_sk_callback_avx+0x22c> + DB 196,98,125,24,5,247,78,0,0 ; vbroadcastss 0x4ef7(%rip),%ymm8 # 69f0 <_sk_callback_avx+0x22a> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,92,89,210 ; vmulps %ymm2,%ymm4,%ymm10 DB 196,65,92,92,218 ; vsubps %ymm10,%ymm4,%ymm11 @@ -6913,23 +6925,23 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 DB 197,108,88,202 ; vaddps %ymm2,%ymm2,%ymm9 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,29,149,78,0,0 ; vbroadcastss 0x4e95(%rip),%ymm11 # 69b8 <_sk_callback_avx+0x230> + DB 196,98,125,24,29,209,78,0,0 ; vbroadcastss 0x4ed1(%rip),%ymm11 # 69f4 <_sk_callback_avx+0x22e> DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,37,127,78,0,0 ; vbroadcastss 0x4e7f(%rip),%ymm12 # 69c0 <_sk_callback_avx+0x238> + DB 196,98,125,24,37,187,78,0,0 ; vbroadcastss 0x4ebb(%rip),%ymm12 # 69fc <_sk_callback_avx+0x236> DB 196,193,4,89,196 ; vmulps %ymm12,%ymm15,%ymm0 - DB 196,98,125,24,45,117,78,0,0 ; vbroadcastss 0x4e75(%rip),%ymm13 # 69c4 <_sk_callback_avx+0x23c> + DB 196,98,125,24,45,177,78,0,0 ; vbroadcastss 0x4eb1(%rip),%ymm13 # 6a00 <_sk_callback_avx+0x23a> DB 197,20,92,240 ; vsubps %ymm0,%ymm13,%ymm14 DB 196,65,36,89,246 ; vmulps %ymm14,%ymm11,%ymm14 DB 196,65,52,88,246 ; vaddps %ymm14,%ymm9,%ymm14 - DB 196,226,125,24,13,86,78,0,0 ; vbroadcastss 0x4e56(%rip),%ymm1 # 69bc <_sk_callback_avx+0x234> + DB 196,226,125,24,13,146,78,0,0 ; vbroadcastss 0x4e92(%rip),%ymm1 # 69f8 <_sk_callback_avx+0x232> DB 196,193,116,194,255,2 ; vcmpleps %ymm15,%ymm1,%ymm7 DB 196,195,13,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm14,%ymm7 DB 196,65,60,194,247,2 ; vcmpleps %ymm15,%ymm8,%ymm14 DB 196,227,45,74,255,224 ; vblendvps %ymm14,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,53,65,78,0,0 ; vbroadcastss 0x4e41(%rip),%ymm14 # 69c8 <_sk_callback_avx+0x240> + DB 196,98,125,24,53,125,78,0,0 ; vbroadcastss 0x4e7d(%rip),%ymm14 # 6a04 <_sk_callback_avx+0x23e> DB 196,65,12,194,255,2 ; vcmpleps %ymm15,%ymm14,%ymm15 DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -6948,7 +6960,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,164,89,247 ; vmulps %ymm7,%ymm11,%ymm6 DB 197,180,88,246 ; vaddps %ymm6,%ymm9,%ymm6 DB 196,227,77,74,237,0 ; vblendvps %ymm0,%ymm5,%ymm6,%ymm5 - DB 196,226,125,24,5,227,77,0,0 ; vbroadcastss 0x4de3(%rip),%ymm0 # 69cc <_sk_callback_avx+0x244> + DB 196,226,125,24,5,31,78,0,0 ; vbroadcastss 0x4e1f(%rip),%ymm0 # 6a08 <_sk_callback_avx+0x242> DB 197,228,88,192 ; vaddps %ymm0,%ymm3,%ymm0 DB 196,227,125,8,216,1 ; vroundps $0x1,%ymm0,%ymm3 DB 197,252,92,195 ; vsubps %ymm3,%ymm0,%ymm0 @@ -6996,13 +7008,13 @@ _sk_scale_u8_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,74 ; jne 1ce6 <_sk_scale_u8_avx+0x54> DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8 - DB 197,57,219,5,102,80,0,0 ; vpand 0x5066(%rip),%xmm8,%xmm8 # 6d10 <_sk_callback_avx+0x588> + DB 197,57,219,5,166,80,0,0 ; vpand 0x50a6(%rip),%xmm8,%xmm8 # 6d50 <_sk_callback_avx+0x58a> DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,3,77,0,0 ; vbroadcastss 0x4d03(%rip),%ymm9 # 69d0 <_sk_callback_avx+0x248> + DB 196,98,125,24,13,63,77,0,0 ; vbroadcastss 0x4d3f(%rip),%ymm9 # 6a0c <_sk_callback_avx+0x246> 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 @@ -7084,13 +7096,13 @@ _sk_lerp_u8_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,110 ; jne 1e3b <_sk_lerp_u8_avx+0x78> DB 196,66,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm8 - DB 197,57,219,5,69,79,0,0 ; vpand 0x4f45(%rip),%xmm8,%xmm8 # 6d20 <_sk_callback_avx+0x598> + DB 197,57,219,5,133,79,0,0 ; vpand 0x4f85(%rip),%xmm8,%xmm8 # 6d60 <_sk_callback_avx+0x59a> DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,214,75,0,0 ; vbroadcastss 0x4bd6(%rip),%ymm9 # 69d4 <_sk_callback_avx+0x24c> + DB 196,98,125,24,13,18,76,0,0 ; vbroadcastss 0x4c12(%rip),%ymm9 # 6a10 <_sk_callback_avx+0x24a> 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 @@ -7166,20 +7178,20 @@ _sk_lerp_565_avx LABEL PROC DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - DB 196,98,125,24,13,206,74,0,0 ; vbroadcastss 0x4ace(%rip),%ymm9 # 69d8 <_sk_callback_avx+0x250> + DB 196,98,125,24,13,10,75,0,0 ; vbroadcastss 0x4b0a(%rip),%ymm9 # 6a14 <_sk_callback_avx+0x24e> DB 196,65,60,84,201 ; vandps %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,191,74,0,0 ; vbroadcastss 0x4abf(%rip),%ymm10 # 69dc <_sk_callback_avx+0x254> + DB 196,98,125,24,21,251,74,0,0 ; vbroadcastss 0x4afb(%rip),%ymm10 # 6a18 <_sk_callback_avx+0x252> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,181,74,0,0 ; vbroadcastss 0x4ab5(%rip),%ymm10 # 69e0 <_sk_callback_avx+0x258> + DB 196,98,125,24,21,241,74,0,0 ; vbroadcastss 0x4af1(%rip),%ymm10 # 6a1c <_sk_callback_avx+0x256> DB 196,65,60,84,210 ; vandps %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,166,74,0,0 ; vbroadcastss 0x4aa6(%rip),%ymm11 # 69e4 <_sk_callback_avx+0x25c> + DB 196,98,125,24,29,226,74,0,0 ; vbroadcastss 0x4ae2(%rip),%ymm11 # 6a20 <_sk_callback_avx+0x25a> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,29,156,74,0,0 ; vbroadcastss 0x4a9c(%rip),%ymm11 # 69e8 <_sk_callback_avx+0x260> + DB 196,98,125,24,29,216,74,0,0 ; vbroadcastss 0x4ad8(%rip),%ymm11 # 6a24 <_sk_callback_avx+0x25e> DB 196,65,60,84,195 ; vandps %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,141,74,0,0 ; vbroadcastss 0x4a8d(%rip),%ymm11 # 69ec <_sk_callback_avx+0x264> + DB 196,98,125,24,29,201,74,0,0 ; vbroadcastss 0x4ac9(%rip),%ymm11 # 6a28 <_sk_callback_avx+0x262> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 @@ -7258,7 +7270,7 @@ _sk_load_tables_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 15,133,251,1,0,0 ; jne 225e <_sk_load_tables_avx+0x21e> DB 196,65,124,16,18 ; vmovups (%r10),%ymm10 - DB 197,124,40,13,144,77,0,0 ; vmovaps 0x4d90(%rip),%ymm9 # 6e00 <_sk_callback_avx+0x678> + DB 197,124,40,13,208,77,0,0 ; vmovaps 0x4dd0(%rip),%ymm9 # 6e40 <_sk_callback_avx+0x67a> DB 196,193,44,84,201 ; vandps %ymm9,%ymm10,%ymm1 DB 196,227,125,25,200,1 ; vextractf128 $0x1,%ymm1,%xmm0 DB 196,193,249,126,195 ; vmovq %xmm0,%r11 @@ -7350,7 +7362,7 @@ _sk_load_tables_avx LABEL PROC DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7 DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,226,125,24,61,167,71,0,0 ; vbroadcastss 0x47a7(%rip),%ymm7 # 69f0 <_sk_callback_avx+0x268> + DB 196,226,125,24,61,227,71,0,0 ; vbroadcastss 0x47e3(%rip),%ymm7 # 6a2c <_sk_callback_avx+0x266> DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,201 ; mov %r9,%rcx @@ -7365,9 +7377,9 @@ _sk_load_tables_avx LABEL PROC DB 73,211,235 ; shr %cl,%r11 DB 196,193,249,110,195 ; vmovq %r11,%xmm0 DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0 - DB 196,226,121,0,13,170,74,0,0 ; vpshufb 0x4aaa(%rip),%xmm0,%xmm1 # 6d30 <_sk_callback_avx+0x5a8> + DB 196,226,121,0,13,234,74,0,0 ; vpshufb 0x4aea(%rip),%xmm0,%xmm1 # 6d70 <_sk_callback_avx+0x5aa> DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1 - DB 196,226,121,0,5,172,74,0,0 ; vpshufb 0x4aac(%rip),%xmm0,%xmm0 # 6d40 <_sk_callback_avx+0x5b8> + DB 196,226,121,0,5,236,74,0,0 ; vpshufb 0x4aec(%rip),%xmm0,%xmm0 # 6d80 <_sk_callback_avx+0x5ba> DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 196,66,125,44,18 ; vmaskmovps (%r10),%ymm0,%ymm10 @@ -7396,7 +7408,7 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 197,113,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm11 DB 197,177,108,200 ; vpunpcklqdq %xmm0,%xmm9,%xmm1 DB 197,49,109,224 ; vpunpckhqdq %xmm0,%xmm9,%xmm12 - DB 197,121,111,21,59,74,0,0 ; vmovdqa 0x4a3b(%rip),%xmm10 # 6d50 <_sk_callback_avx+0x5c8> + DB 197,121,111,21,123,74,0,0 ; vmovdqa 0x4a7b(%rip),%xmm10 # 6d90 <_sk_callback_avx+0x5ca> DB 196,193,113,219,202 ; vpand %xmm10,%xmm1,%xmm1 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2 @@ -7494,7 +7506,7 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,226,125,24,61,230,68,0,0 ; vbroadcastss 0x44e6(%rip),%ymm7 # 69f4 <_sk_callback_avx+0x26c> + DB 196,226,125,24,61,34,69,0,0 ; vbroadcastss 0x4522(%rip),%ymm7 # 6a30 <_sk_callback_avx+0x26a> DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7 @@ -7557,7 +7569,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3 DB 197,57,109,218 ; vpunpckhqdq %xmm2,%xmm8,%xmm11 DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8 - DB 197,121,111,13,58,71,0,0 ; vmovdqa 0x473a(%rip),%xmm9 # 6d60 <_sk_callback_avx+0x5d8> + DB 197,121,111,13,122,71,0,0 ; vmovdqa 0x477a(%rip),%xmm9 # 6da0 <_sk_callback_avx+0x5da> DB 196,193,97,219,193 ; vpand %xmm9,%xmm3,%xmm0 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 @@ -7647,7 +7659,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 196,195,105,33,211,48 ; vinsertps $0x30,%xmm11,%xmm2,%xmm2 DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,3,66,0,0 ; vbroadcastss 0x4203(%rip),%ymm3 # 69f8 <_sk_callback_avx+0x270> + DB 196,226,125,24,29,63,66,0,0 ; vbroadcastss 0x423f(%rip),%ymm3 # 6a34 <_sk_callback_avx+0x26e> DB 197,252,16,52,36 ; vmovups (%rsp),%ymm6 DB 197,252,16,124,36,32 ; vmovups 0x20(%rsp),%ymm7 DB 72,131,196,88 ; add $0x58,%rsp @@ -7689,7 +7701,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC PUBLIC _sk_byte_tables_avx _sk_byte_tables_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,60,65,0,0 ; vbroadcastss 0x413c(%rip),%ymm8 # 69fc <_sk_callback_avx+0x274> + DB 196,98,125,24,5,120,65,0,0 ; vbroadcastss 0x4178(%rip),%ymm8 # 6a38 <_sk_callback_avx+0x272> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 DB 196,65,249,126,201 ; vmovq %xmm9,%r9 @@ -7808,7 +7820,7 @@ _sk_byte_tables_avx LABEL PROC DB 196,194,121,49,204 ; vpmovzxbd %xmm12,%xmm1 DB 196,194,121,49,213 ; vpmovzxbd %xmm13,%xmm2 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,98,125,24,13,230,62,0,0 ; vbroadcastss 0x3ee6(%rip),%ymm9 # 6a00 <_sk_callback_avx+0x278> + DB 196,98,125,24,13,34,63,0,0 ; vbroadcastss 0x3f22(%rip),%ymm9 # 6a3c <_sk_callback_avx+0x276> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 @@ -7922,7 +7934,7 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 196,194,121,49,203 ; vpmovzxbd %xmm11,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,199,60,0,0 ; vbroadcastss 0x3cc7(%rip),%ymm9 # 6a04 <_sk_callback_avx+0x27c> + DB 196,98,125,24,13,3,61,0,0 ; vbroadcastss 0x3d03(%rip),%ymm9 # 6a40 <_sk_callback_avx+0x27a> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,194,121,49,202 ; vpmovzxbd %xmm10,%xmm1 DB 196,194,121,49,212 ; vpmovzxbd %xmm12,%xmm2 @@ -8117,36 +8129,36 @@ _sk_parametric_r_avx LABEL PROC DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,216 ; vcvtdq2ps %ymm0,%ymm11 - DB 196,98,125,24,37,116,57,0,0 ; vbroadcastss 0x3974(%rip),%ymm12 # 6a08 <_sk_callback_avx+0x280> + DB 196,98,125,24,37,176,57,0,0 ; vbroadcastss 0x39b0(%rip),%ymm12 # 6a44 <_sk_callback_avx+0x27e> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,106,57,0,0 ; vbroadcastss 0x396a(%rip),%ymm12 # 6a0c <_sk_callback_avx+0x284> + DB 196,98,125,24,37,166,57,0,0 ; vbroadcastss 0x39a6(%rip),%ymm12 # 6a48 <_sk_callback_avx+0x282> DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,96,57,0,0 ; vbroadcastss 0x3960(%rip),%ymm12 # 6a10 <_sk_callback_avx+0x288> + DB 196,98,125,24,37,156,57,0,0 ; vbroadcastss 0x399c(%rip),%ymm12 # 6a4c <_sk_callback_avx+0x286> DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,86,57,0,0 ; vbroadcastss 0x3956(%rip),%ymm12 # 6a14 <_sk_callback_avx+0x28c> + DB 196,98,125,24,37,146,57,0,0 ; vbroadcastss 0x3992(%rip),%ymm12 # 6a50 <_sk_callback_avx+0x28a> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,76,57,0,0 ; vbroadcastss 0x394c(%rip),%ymm12 # 6a18 <_sk_callback_avx+0x290> + DB 196,98,125,24,37,136,57,0,0 ; vbroadcastss 0x3988(%rip),%ymm12 # 6a54 <_sk_callback_avx+0x28e> DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,61,57,0,0 ; vbroadcastss 0x393d(%rip),%ymm12 # 6a1c <_sk_callback_avx+0x294> + DB 196,98,125,24,37,121,57,0,0 ; vbroadcastss 0x3979(%rip),%ymm12 # 6a58 <_sk_callback_avx+0x292> DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,51,57,0,0 ; vbroadcastss 0x3933(%rip),%ymm12 # 6a20 <_sk_callback_avx+0x298> + DB 196,98,125,24,37,111,57,0,0 ; vbroadcastss 0x396f(%rip),%ymm12 # 6a5c <_sk_callback_avx+0x296> DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,23,57,0,0 ; vbroadcastss 0x3917(%rip),%ymm11 # 6a24 <_sk_callback_avx+0x29c> + DB 196,98,125,24,29,83,57,0,0 ; vbroadcastss 0x3953(%rip),%ymm11 # 6a60 <_sk_callback_avx+0x29a> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,13,57,0,0 ; vbroadcastss 0x390d(%rip),%ymm11 # 6a28 <_sk_callback_avx+0x2a0> + DB 196,98,125,24,29,73,57,0,0 ; vbroadcastss 0x3949(%rip),%ymm11 # 6a64 <_sk_callback_avx+0x29e> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,254,56,0,0 ; vbroadcastss 0x38fe(%rip),%ymm11 # 6a2c <_sk_callback_avx+0x2a4> + DB 196,98,125,24,29,58,57,0,0 ; vbroadcastss 0x393a(%rip),%ymm11 # 6a68 <_sk_callback_avx+0x2a2> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,244,56,0,0 ; vbroadcastss 0x38f4(%rip),%ymm11 # 6a30 <_sk_callback_avx+0x2a8> + DB 196,98,125,24,29,48,57,0,0 ; vbroadcastss 0x3930(%rip),%ymm11 # 6a6c <_sk_callback_avx+0x2a6> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,229,56,0,0 ; vbroadcastss 0x38e5(%rip),%ymm10 # 6a34 <_sk_callback_avx+0x2ac> + DB 196,98,125,24,21,33,57,0,0 ; vbroadcastss 0x3921(%rip),%ymm10 # 6a70 <_sk_callback_avx+0x2aa> 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 @@ -8154,7 +8166,7 @@ _sk_parametric_r_avx LABEL PROC DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,188,56,0,0 ; vbroadcastss 0x38bc(%rip),%ymm8 # 6a38 <_sk_callback_avx+0x2b0> + DB 196,98,125,24,5,248,56,0,0 ; vbroadcastss 0x38f8(%rip),%ymm8 # 6a74 <_sk_callback_avx+0x2ae> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8174,36 +8186,36 @@ _sk_parametric_g_avx LABEL PROC DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,217 ; vcvtdq2ps %ymm1,%ymm11 - DB 196,98,125,24,37,109,56,0,0 ; vbroadcastss 0x386d(%rip),%ymm12 # 6a3c <_sk_callback_avx+0x2b4> + DB 196,98,125,24,37,169,56,0,0 ; vbroadcastss 0x38a9(%rip),%ymm12 # 6a78 <_sk_callback_avx+0x2b2> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,99,56,0,0 ; vbroadcastss 0x3863(%rip),%ymm12 # 6a40 <_sk_callback_avx+0x2b8> + DB 196,98,125,24,37,159,56,0,0 ; vbroadcastss 0x389f(%rip),%ymm12 # 6a7c <_sk_callback_avx+0x2b6> DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,89,56,0,0 ; vbroadcastss 0x3859(%rip),%ymm12 # 6a44 <_sk_callback_avx+0x2bc> + DB 196,98,125,24,37,149,56,0,0 ; vbroadcastss 0x3895(%rip),%ymm12 # 6a80 <_sk_callback_avx+0x2ba> DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,79,56,0,0 ; vbroadcastss 0x384f(%rip),%ymm12 # 6a48 <_sk_callback_avx+0x2c0> + DB 196,98,125,24,37,139,56,0,0 ; vbroadcastss 0x388b(%rip),%ymm12 # 6a84 <_sk_callback_avx+0x2be> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,69,56,0,0 ; vbroadcastss 0x3845(%rip),%ymm12 # 6a4c <_sk_callback_avx+0x2c4> + DB 196,98,125,24,37,129,56,0,0 ; vbroadcastss 0x3881(%rip),%ymm12 # 6a88 <_sk_callback_avx+0x2c2> DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,54,56,0,0 ; vbroadcastss 0x3836(%rip),%ymm12 # 6a50 <_sk_callback_avx+0x2c8> + DB 196,98,125,24,37,114,56,0,0 ; vbroadcastss 0x3872(%rip),%ymm12 # 6a8c <_sk_callback_avx+0x2c6> DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,44,56,0,0 ; vbroadcastss 0x382c(%rip),%ymm12 # 6a54 <_sk_callback_avx+0x2cc> + DB 196,98,125,24,37,104,56,0,0 ; vbroadcastss 0x3868(%rip),%ymm12 # 6a90 <_sk_callback_avx+0x2ca> DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1 DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,16,56,0,0 ; vbroadcastss 0x3810(%rip),%ymm11 # 6a58 <_sk_callback_avx+0x2d0> + DB 196,98,125,24,29,76,56,0,0 ; vbroadcastss 0x384c(%rip),%ymm11 # 6a94 <_sk_callback_avx+0x2ce> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,6,56,0,0 ; vbroadcastss 0x3806(%rip),%ymm11 # 6a5c <_sk_callback_avx+0x2d4> + DB 196,98,125,24,29,66,56,0,0 ; vbroadcastss 0x3842(%rip),%ymm11 # 6a98 <_sk_callback_avx+0x2d2> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,247,55,0,0 ; vbroadcastss 0x37f7(%rip),%ymm11 # 6a60 <_sk_callback_avx+0x2d8> + DB 196,98,125,24,29,51,56,0,0 ; vbroadcastss 0x3833(%rip),%ymm11 # 6a9c <_sk_callback_avx+0x2d6> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,237,55,0,0 ; vbroadcastss 0x37ed(%rip),%ymm11 # 6a64 <_sk_callback_avx+0x2dc> + DB 196,98,125,24,29,41,56,0,0 ; vbroadcastss 0x3829(%rip),%ymm11 # 6aa0 <_sk_callback_avx+0x2da> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,222,55,0,0 ; vbroadcastss 0x37de(%rip),%ymm10 # 6a68 <_sk_callback_avx+0x2e0> + DB 196,98,125,24,21,26,56,0,0 ; vbroadcastss 0x381a(%rip),%ymm10 # 6aa4 <_sk_callback_avx+0x2de> 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 @@ -8211,7 +8223,7 @@ _sk_parametric_g_avx LABEL PROC DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,181,55,0,0 ; vbroadcastss 0x37b5(%rip),%ymm8 # 6a6c <_sk_callback_avx+0x2e4> + DB 196,98,125,24,5,241,55,0,0 ; vbroadcastss 0x37f1(%rip),%ymm8 # 6aa8 <_sk_callback_avx+0x2e2> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8231,36 +8243,36 @@ _sk_parametric_b_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,218 ; vcvtdq2ps %ymm2,%ymm11 - DB 196,98,125,24,37,102,55,0,0 ; vbroadcastss 0x3766(%rip),%ymm12 # 6a70 <_sk_callback_avx+0x2e8> + DB 196,98,125,24,37,162,55,0,0 ; vbroadcastss 0x37a2(%rip),%ymm12 # 6aac <_sk_callback_avx+0x2e6> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,92,55,0,0 ; vbroadcastss 0x375c(%rip),%ymm12 # 6a74 <_sk_callback_avx+0x2ec> + DB 196,98,125,24,37,152,55,0,0 ; vbroadcastss 0x3798(%rip),%ymm12 # 6ab0 <_sk_callback_avx+0x2ea> DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,82,55,0,0 ; vbroadcastss 0x3752(%rip),%ymm12 # 6a78 <_sk_callback_avx+0x2f0> + DB 196,98,125,24,37,142,55,0,0 ; vbroadcastss 0x378e(%rip),%ymm12 # 6ab4 <_sk_callback_avx+0x2ee> DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,72,55,0,0 ; vbroadcastss 0x3748(%rip),%ymm12 # 6a7c <_sk_callback_avx+0x2f4> + DB 196,98,125,24,37,132,55,0,0 ; vbroadcastss 0x3784(%rip),%ymm12 # 6ab8 <_sk_callback_avx+0x2f2> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,62,55,0,0 ; vbroadcastss 0x373e(%rip),%ymm12 # 6a80 <_sk_callback_avx+0x2f8> + DB 196,98,125,24,37,122,55,0,0 ; vbroadcastss 0x377a(%rip),%ymm12 # 6abc <_sk_callback_avx+0x2f6> DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,47,55,0,0 ; vbroadcastss 0x372f(%rip),%ymm12 # 6a84 <_sk_callback_avx+0x2fc> + DB 196,98,125,24,37,107,55,0,0 ; vbroadcastss 0x376b(%rip),%ymm12 # 6ac0 <_sk_callback_avx+0x2fa> DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,37,55,0,0 ; vbroadcastss 0x3725(%rip),%ymm12 # 6a88 <_sk_callback_avx+0x300> + DB 196,98,125,24,37,97,55,0,0 ; vbroadcastss 0x3761(%rip),%ymm12 # 6ac4 <_sk_callback_avx+0x2fe> DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2 DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,9,55,0,0 ; vbroadcastss 0x3709(%rip),%ymm11 # 6a8c <_sk_callback_avx+0x304> + DB 196,98,125,24,29,69,55,0,0 ; vbroadcastss 0x3745(%rip),%ymm11 # 6ac8 <_sk_callback_avx+0x302> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,255,54,0,0 ; vbroadcastss 0x36ff(%rip),%ymm11 # 6a90 <_sk_callback_avx+0x308> + DB 196,98,125,24,29,59,55,0,0 ; vbroadcastss 0x373b(%rip),%ymm11 # 6acc <_sk_callback_avx+0x306> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,240,54,0,0 ; vbroadcastss 0x36f0(%rip),%ymm11 # 6a94 <_sk_callback_avx+0x30c> + DB 196,98,125,24,29,44,55,0,0 ; vbroadcastss 0x372c(%rip),%ymm11 # 6ad0 <_sk_callback_avx+0x30a> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,230,54,0,0 ; vbroadcastss 0x36e6(%rip),%ymm11 # 6a98 <_sk_callback_avx+0x310> + DB 196,98,125,24,29,34,55,0,0 ; vbroadcastss 0x3722(%rip),%ymm11 # 6ad4 <_sk_callback_avx+0x30e> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,215,54,0,0 ; vbroadcastss 0x36d7(%rip),%ymm10 # 6a9c <_sk_callback_avx+0x314> + DB 196,98,125,24,21,19,55,0,0 ; vbroadcastss 0x3713(%rip),%ymm10 # 6ad8 <_sk_callback_avx+0x312> 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 @@ -8268,7 +8280,7 @@ _sk_parametric_b_avx LABEL PROC DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,174,54,0,0 ; vbroadcastss 0x36ae(%rip),%ymm8 # 6aa0 <_sk_callback_avx+0x318> + DB 196,98,125,24,5,234,54,0,0 ; vbroadcastss 0x36ea(%rip),%ymm8 # 6adc <_sk_callback_avx+0x316> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8288,36 +8300,36 @@ _sk_parametric_a_avx LABEL PROC DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,219 ; vcvtdq2ps %ymm3,%ymm11 - DB 196,98,125,24,37,95,54,0,0 ; vbroadcastss 0x365f(%rip),%ymm12 # 6aa4 <_sk_callback_avx+0x31c> + DB 196,98,125,24,37,155,54,0,0 ; vbroadcastss 0x369b(%rip),%ymm12 # 6ae0 <_sk_callback_avx+0x31a> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,85,54,0,0 ; vbroadcastss 0x3655(%rip),%ymm12 # 6aa8 <_sk_callback_avx+0x320> + DB 196,98,125,24,37,145,54,0,0 ; vbroadcastss 0x3691(%rip),%ymm12 # 6ae4 <_sk_callback_avx+0x31e> DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,75,54,0,0 ; vbroadcastss 0x364b(%rip),%ymm12 # 6aac <_sk_callback_avx+0x324> + DB 196,98,125,24,37,135,54,0,0 ; vbroadcastss 0x3687(%rip),%ymm12 # 6ae8 <_sk_callback_avx+0x322> DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,65,54,0,0 ; vbroadcastss 0x3641(%rip),%ymm12 # 6ab0 <_sk_callback_avx+0x328> + DB 196,98,125,24,37,125,54,0,0 ; vbroadcastss 0x367d(%rip),%ymm12 # 6aec <_sk_callback_avx+0x326> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,55,54,0,0 ; vbroadcastss 0x3637(%rip),%ymm12 # 6ab4 <_sk_callback_avx+0x32c> + DB 196,98,125,24,37,115,54,0,0 ; vbroadcastss 0x3673(%rip),%ymm12 # 6af0 <_sk_callback_avx+0x32a> DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,40,54,0,0 ; vbroadcastss 0x3628(%rip),%ymm12 # 6ab8 <_sk_callback_avx+0x330> + DB 196,98,125,24,37,100,54,0,0 ; vbroadcastss 0x3664(%rip),%ymm12 # 6af4 <_sk_callback_avx+0x32e> DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,30,54,0,0 ; vbroadcastss 0x361e(%rip),%ymm12 # 6abc <_sk_callback_avx+0x334> + DB 196,98,125,24,37,90,54,0,0 ; vbroadcastss 0x365a(%rip),%ymm12 # 6af8 <_sk_callback_avx+0x332> DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3 DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,2,54,0,0 ; vbroadcastss 0x3602(%rip),%ymm11 # 6ac0 <_sk_callback_avx+0x338> + DB 196,98,125,24,29,62,54,0,0 ; vbroadcastss 0x363e(%rip),%ymm11 # 6afc <_sk_callback_avx+0x336> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,248,53,0,0 ; vbroadcastss 0x35f8(%rip),%ymm11 # 6ac4 <_sk_callback_avx+0x33c> + DB 196,98,125,24,29,52,54,0,0 ; vbroadcastss 0x3634(%rip),%ymm11 # 6b00 <_sk_callback_avx+0x33a> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,233,53,0,0 ; vbroadcastss 0x35e9(%rip),%ymm11 # 6ac8 <_sk_callback_avx+0x340> + DB 196,98,125,24,29,37,54,0,0 ; vbroadcastss 0x3625(%rip),%ymm11 # 6b04 <_sk_callback_avx+0x33e> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,223,53,0,0 ; vbroadcastss 0x35df(%rip),%ymm11 # 6acc <_sk_callback_avx+0x344> + DB 196,98,125,24,29,27,54,0,0 ; vbroadcastss 0x361b(%rip),%ymm11 # 6b08 <_sk_callback_avx+0x342> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,208,53,0,0 ; vbroadcastss 0x35d0(%rip),%ymm10 # 6ad0 <_sk_callback_avx+0x348> + DB 196,98,125,24,21,12,54,0,0 ; vbroadcastss 0x360c(%rip),%ymm10 # 6b0c <_sk_callback_avx+0x346> 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 @@ -8325,38 +8337,38 @@ _sk_parametric_a_avx LABEL PROC DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,167,53,0,0 ; vbroadcastss 0x35a7(%rip),%ymm8 # 6ad4 <_sk_callback_avx+0x34c> + DB 196,98,125,24,5,227,53,0,0 ; vbroadcastss 0x35e3(%rip),%ymm8 # 6b10 <_sk_callback_avx+0x34a> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_lab_to_xyz_avx _sk_lab_to_xyz_avx LABEL PROC - DB 196,98,125,24,5,153,53,0,0 ; vbroadcastss 0x3599(%rip),%ymm8 # 6ad8 <_sk_callback_avx+0x350> + DB 196,98,125,24,5,213,53,0,0 ; vbroadcastss 0x35d5(%rip),%ymm8 # 6b14 <_sk_callback_avx+0x34e> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,143,53,0,0 ; vbroadcastss 0x358f(%rip),%ymm8 # 6adc <_sk_callback_avx+0x354> + DB 196,98,125,24,5,203,53,0,0 ; vbroadcastss 0x35cb(%rip),%ymm8 # 6b18 <_sk_callback_avx+0x352> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,13,133,53,0,0 ; vbroadcastss 0x3585(%rip),%ymm9 # 6ae0 <_sk_callback_avx+0x358> + DB 196,98,125,24,13,193,53,0,0 ; vbroadcastss 0x35c1(%rip),%ymm9 # 6b1c <_sk_callback_avx+0x356> DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 196,193,108,88,209 ; vaddps %ymm9,%ymm2,%ymm2 - DB 196,98,125,24,5,113,53,0,0 ; vbroadcastss 0x3571(%rip),%ymm8 # 6ae4 <_sk_callback_avx+0x35c> + DB 196,98,125,24,5,173,53,0,0 ; vbroadcastss 0x35ad(%rip),%ymm8 # 6b20 <_sk_callback_avx+0x35a> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,103,53,0,0 ; vbroadcastss 0x3567(%rip),%ymm8 # 6ae8 <_sk_callback_avx+0x360> + DB 196,98,125,24,5,163,53,0,0 ; vbroadcastss 0x35a3(%rip),%ymm8 # 6b24 <_sk_callback_avx+0x35e> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,93,53,0,0 ; vbroadcastss 0x355d(%rip),%ymm8 # 6aec <_sk_callback_avx+0x364> + DB 196,98,125,24,5,153,53,0,0 ; vbroadcastss 0x3599(%rip),%ymm8 # 6b28 <_sk_callback_avx+0x362> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1 - DB 196,98,125,24,5,79,53,0,0 ; vbroadcastss 0x354f(%rip),%ymm8 # 6af0 <_sk_callback_avx+0x368> + DB 196,98,125,24,5,139,53,0,0 ; vbroadcastss 0x358b(%rip),%ymm8 # 6b2c <_sk_callback_avx+0x366> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,252,92,210 ; vsubps %ymm2,%ymm0,%ymm2 DB 197,116,89,193 ; vmulps %ymm1,%ymm1,%ymm8 DB 196,65,116,89,192 ; vmulps %ymm8,%ymm1,%ymm8 - DB 196,98,125,24,13,56,53,0,0 ; vbroadcastss 0x3538(%rip),%ymm9 # 6af4 <_sk_callback_avx+0x36c> + DB 196,98,125,24,13,116,53,0,0 ; vbroadcastss 0x3574(%rip),%ymm9 # 6b30 <_sk_callback_avx+0x36a> DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,45,53,0,0 ; vbroadcastss 0x352d(%rip),%ymm11 # 6af8 <_sk_callback_avx+0x370> + DB 196,98,125,24,29,105,53,0,0 ; vbroadcastss 0x3569(%rip),%ymm11 # 6b34 <_sk_callback_avx+0x36e> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,37,35,53,0,0 ; vbroadcastss 0x3523(%rip),%ymm12 # 6afc <_sk_callback_avx+0x374> + DB 196,98,125,24,37,95,53,0,0 ; vbroadcastss 0x355f(%rip),%ymm12 # 6b38 <_sk_callback_avx+0x372> DB 196,193,116,89,204 ; vmulps %ymm12,%ymm1,%ymm1 DB 196,67,117,74,192,160 ; vblendvps %ymm10,%ymm8,%ymm1,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -8371,9 +8383,9 @@ _sk_lab_to_xyz_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,193,108,89,212 ; vmulps %ymm12,%ymm2,%ymm2 DB 196,227,109,74,208,144 ; vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - DB 196,226,125,24,5,217,52,0,0 ; vbroadcastss 0x34d9(%rip),%ymm0 # 6b00 <_sk_callback_avx+0x378> + DB 196,226,125,24,5,21,53,0,0 ; vbroadcastss 0x3515(%rip),%ymm0 # 6b3c <_sk_callback_avx+0x376> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,208,52,0,0 ; vbroadcastss 0x34d0(%rip),%ymm8 # 6b04 <_sk_callback_avx+0x37c> + DB 196,98,125,24,5,12,53,0,0 ; vbroadcastss 0x350c(%rip),%ymm8 # 6b40 <_sk_callback_avx+0x37a> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8385,13 +8397,13 @@ _sk_load_a8_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,66 ; jne 3689 <_sk_load_a8_avx+0x4c> DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0 - DB 197,249,219,5,27,55,0,0 ; vpand 0x371b(%rip),%xmm0,%xmm0 # 6d70 <_sk_callback_avx+0x5e8> + DB 197,249,219,5,91,55,0,0 ; vpand 0x375b(%rip),%xmm0,%xmm0 # 6db0 <_sk_callback_avx+0x5ea> DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,147,52,0,0 ; vbroadcastss 0x3493(%rip),%ymm1 # 6b08 <_sk_callback_avx+0x380> + DB 196,226,125,24,13,207,52,0,0 ; vbroadcastss 0x34cf(%rip),%ymm1 # 6b44 <_sk_callback_avx+0x37e> 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 @@ -8489,7 +8501,7 @@ _sk_gather_a8_avx LABEL PROC DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,28,51,0,0 ; vbroadcastss 0x331c(%rip),%ymm1 # 6b0c <_sk_callback_avx+0x384> + DB 196,226,125,24,13,88,51,0,0 ; vbroadcastss 0x3358(%rip),%ymm1 # 6b48 <_sk_callback_avx+0x382> 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 @@ -8501,7 +8513,7 @@ PUBLIC _sk_store_a8_avx _sk_store_a8_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 - DB 196,98,125,24,5,254,50,0,0 ; vbroadcastss 0x32fe(%rip),%ymm8 # 6b10 <_sk_callback_avx+0x388> + DB 196,98,125,24,5,58,51,0,0 ; vbroadcastss 0x333a(%rip),%ymm8 # 6b4c <_sk_callback_avx+0x386> 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 @@ -8563,16 +8575,16 @@ _sk_load_g8_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,71 ; jne 390d <_sk_load_g8_avx+0x51> DB 196,194,121,48,4,19 ; vpmovzxbw (%r11,%rdx,1),%xmm0 - DB 197,249,219,5,172,52,0,0 ; vpand 0x34ac(%rip),%xmm0,%xmm0 # 6d80 <_sk_callback_avx+0x5f8> + DB 197,249,219,5,236,52,0,0 ; vpand 0x34ec(%rip),%xmm0,%xmm0 # 6dc0 <_sk_callback_avx+0x5fa> DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,32,50,0,0 ; vbroadcastss 0x3220(%rip),%ymm1 # 6b14 <_sk_callback_avx+0x38c> + DB 196,226,125,24,13,92,50,0,0 ; vbroadcastss 0x325c(%rip),%ymm1 # 6b50 <_sk_callback_avx+0x38a> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,21,50,0,0 ; vbroadcastss 0x3215(%rip),%ymm3 # 6b18 <_sk_callback_avx+0x390> + DB 196,226,125,24,29,81,50,0,0 ; vbroadcastss 0x3251(%rip),%ymm3 # 6b54 <_sk_callback_avx+0x38e> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax @@ -8667,10 +8679,10 @@ _sk_gather_g8_avx LABEL PROC DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,168,48,0,0 ; vbroadcastss 0x30a8(%rip),%ymm1 # 6b1c <_sk_callback_avx+0x394> + DB 196,226,125,24,13,228,48,0,0 ; vbroadcastss 0x30e4(%rip),%ymm1 # 6b58 <_sk_callback_avx+0x392> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,157,48,0,0 ; vbroadcastss 0x309d(%rip),%ymm3 # 6b20 <_sk_callback_avx+0x398> + DB 196,226,125,24,29,217,48,0,0 ; vbroadcastss 0x30d9(%rip),%ymm3 # 6b5c <_sk_callback_avx+0x396> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax @@ -8740,10 +8752,10 @@ _sk_gather_i8_avx LABEL PROC DB 72,193,235,30 ; shr $0x1e,%rbx DB 196,195,121,34,28,27,3 ; vpinsrd $0x3,(%r11,%rbx,1),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,100,50,0,0 ; vmovaps 0x3264(%rip),%ymm10 # 6e20 <_sk_callback_avx+0x698> + DB 197,124,40,21,164,50,0,0 ; vmovaps 0x32a4(%rip),%ymm10 # 6e60 <_sk_callback_avx+0x69a> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,86,47,0,0 ; vbroadcastss 0x2f56(%rip),%ymm9 # 6b24 <_sk_callback_avx+0x39c> + DB 196,98,125,24,13,146,47,0,0 ; vbroadcastss 0x2f92(%rip),%ymm9 # 6b60 <_sk_callback_avx+0x39a> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -8777,23 +8789,23 @@ _sk_load_565_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,200,46,0,0 ; vbroadcastss 0x2ec8(%rip),%ymm0 # 6b28 <_sk_callback_avx+0x3a0> + DB 196,226,125,24,5,4,47,0,0 ; vbroadcastss 0x2f04(%rip),%ymm0 # 6b64 <_sk_callback_avx+0x39e> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,187,46,0,0 ; vbroadcastss 0x2ebb(%rip),%ymm1 # 6b2c <_sk_callback_avx+0x3a4> + DB 196,226,125,24,13,247,46,0,0 ; vbroadcastss 0x2ef7(%rip),%ymm1 # 6b68 <_sk_callback_avx+0x3a2> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,178,46,0,0 ; vbroadcastss 0x2eb2(%rip),%ymm1 # 6b30 <_sk_callback_avx+0x3a8> + DB 196,226,125,24,13,238,46,0,0 ; vbroadcastss 0x2eee(%rip),%ymm1 # 6b6c <_sk_callback_avx+0x3a6> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,165,46,0,0 ; vbroadcastss 0x2ea5(%rip),%ymm3 # 6b34 <_sk_callback_avx+0x3ac> + DB 196,226,125,24,29,225,46,0,0 ; vbroadcastss 0x2ee1(%rip),%ymm3 # 6b70 <_sk_callback_avx+0x3aa> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,156,46,0,0 ; vbroadcastss 0x2e9c(%rip),%ymm3 # 6b38 <_sk_callback_avx+0x3b0> + DB 196,226,125,24,29,216,46,0,0 ; vbroadcastss 0x2ed8(%rip),%ymm3 # 6b74 <_sk_callback_avx+0x3ae> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,143,46,0,0 ; vbroadcastss 0x2e8f(%rip),%ymm3 # 6b3c <_sk_callback_avx+0x3b4> + DB 196,226,125,24,29,203,46,0,0 ; vbroadcastss 0x2ecb(%rip),%ymm3 # 6b78 <_sk_callback_avx+0x3b2> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,132,46,0,0 ; vbroadcastss 0x2e84(%rip),%ymm3 # 6b40 <_sk_callback_avx+0x3b8> + DB 196,226,125,24,29,192,46,0,0 ; vbroadcastss 0x2ec0(%rip),%ymm3 # 6b7c <_sk_callback_avx+0x3b6> DB 255,224 ; jmpq *%rax DB 69,137,193 ; mov %r8d,%r9d DB 65,128,225,7 ; and $0x7,%r9b @@ -8886,37 +8898,37 @@ _sk_gather_565_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,42,45,0,0 ; vbroadcastss 0x2d2a(%rip),%ymm0 # 6b44 <_sk_callback_avx+0x3bc> + DB 196,226,125,24,5,102,45,0,0 ; vbroadcastss 0x2d66(%rip),%ymm0 # 6b80 <_sk_callback_avx+0x3ba> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,29,45,0,0 ; vbroadcastss 0x2d1d(%rip),%ymm1 # 6b48 <_sk_callback_avx+0x3c0> + DB 196,226,125,24,13,89,45,0,0 ; vbroadcastss 0x2d59(%rip),%ymm1 # 6b84 <_sk_callback_avx+0x3be> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,20,45,0,0 ; vbroadcastss 0x2d14(%rip),%ymm1 # 6b4c <_sk_callback_avx+0x3c4> + DB 196,226,125,24,13,80,45,0,0 ; vbroadcastss 0x2d50(%rip),%ymm1 # 6b88 <_sk_callback_avx+0x3c2> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,7,45,0,0 ; vbroadcastss 0x2d07(%rip),%ymm3 # 6b50 <_sk_callback_avx+0x3c8> + DB 196,226,125,24,29,67,45,0,0 ; vbroadcastss 0x2d43(%rip),%ymm3 # 6b8c <_sk_callback_avx+0x3c6> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,254,44,0,0 ; vbroadcastss 0x2cfe(%rip),%ymm3 # 6b54 <_sk_callback_avx+0x3cc> + DB 196,226,125,24,29,58,45,0,0 ; vbroadcastss 0x2d3a(%rip),%ymm3 # 6b90 <_sk_callback_avx+0x3ca> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,241,44,0,0 ; vbroadcastss 0x2cf1(%rip),%ymm3 # 6b58 <_sk_callback_avx+0x3d0> + DB 196,226,125,24,29,45,45,0,0 ; vbroadcastss 0x2d2d(%rip),%ymm3 # 6b94 <_sk_callback_avx+0x3ce> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,230,44,0,0 ; vbroadcastss 0x2ce6(%rip),%ymm3 # 6b5c <_sk_callback_avx+0x3d4> + DB 196,226,125,24,29,34,45,0,0 ; vbroadcastss 0x2d22(%rip),%ymm3 # 6b98 <_sk_callback_avx+0x3d2> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_avx _sk_store_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 - DB 196,98,125,24,5,218,44,0,0 ; vbroadcastss 0x2cda(%rip),%ymm8 # 6b60 <_sk_callback_avx+0x3d8> + DB 196,98,125,24,5,22,45,0,0 ; vbroadcastss 0x2d16(%rip),%ymm8 # 6b9c <_sk_callback_avx+0x3d6> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,11 ; vpslld $0xb,%xmm9,%xmm10 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 DB 196,193,49,114,241,11 ; vpslld $0xb,%xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,98,125,24,21,179,44,0,0 ; vbroadcastss 0x2cb3(%rip),%ymm10 # 6b64 <_sk_callback_avx+0x3dc> + DB 196,98,125,24,21,239,44,0,0 ; vbroadcastss 0x2cef(%rip),%ymm10 # 6ba0 <_sk_callback_avx+0x3da> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,33,114,242,5 ; vpslld $0x5,%xmm10,%xmm11 @@ -8988,25 +9000,25 @@ _sk_load_4444_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,188,43,0,0 ; vbroadcastss 0x2bbc(%rip),%ymm0 # 6b68 <_sk_callback_avx+0x3e0> + DB 196,226,125,24,5,248,43,0,0 ; vbroadcastss 0x2bf8(%rip),%ymm0 # 6ba4 <_sk_callback_avx+0x3de> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,175,43,0,0 ; vbroadcastss 0x2baf(%rip),%ymm1 # 6b6c <_sk_callback_avx+0x3e4> + DB 196,226,125,24,13,235,43,0,0 ; vbroadcastss 0x2beb(%rip),%ymm1 # 6ba8 <_sk_callback_avx+0x3e2> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,166,43,0,0 ; vbroadcastss 0x2ba6(%rip),%ymm1 # 6b70 <_sk_callback_avx+0x3e8> + DB 196,226,125,24,13,226,43,0,0 ; vbroadcastss 0x2be2(%rip),%ymm1 # 6bac <_sk_callback_avx+0x3e6> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,153,43,0,0 ; vbroadcastss 0x2b99(%rip),%ymm2 # 6b74 <_sk_callback_avx+0x3ec> + DB 196,226,125,24,21,213,43,0,0 ; vbroadcastss 0x2bd5(%rip),%ymm2 # 6bb0 <_sk_callback_avx+0x3ea> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,144,43,0,0 ; vbroadcastss 0x2b90(%rip),%ymm2 # 6b78 <_sk_callback_avx+0x3f0> + DB 196,226,125,24,21,204,43,0,0 ; vbroadcastss 0x2bcc(%rip),%ymm2 # 6bb4 <_sk_callback_avx+0x3ee> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,131,43,0,0 ; vbroadcastss 0x2b83(%rip),%ymm8 # 6b7c <_sk_callback_avx+0x3f4> + DB 196,98,125,24,5,191,43,0,0 ; vbroadcastss 0x2bbf(%rip),%ymm8 # 6bb8 <_sk_callback_avx+0x3f2> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,121,43,0,0 ; vbroadcastss 0x2b79(%rip),%ymm8 # 6b80 <_sk_callback_avx+0x3f8> + DB 196,98,125,24,5,181,43,0,0 ; vbroadcastss 0x2bb5(%rip),%ymm8 # 6bbc <_sk_callback_avx+0x3f6> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,107,43,0,0 ; vbroadcastss 0x2b6b(%rip),%ymm8 # 6b84 <_sk_callback_avx+0x3fc> + DB 196,98,125,24,5,167,43,0,0 ; vbroadcastss 0x2ba7(%rip),%ymm8 # 6bc0 <_sk_callback_avx+0x3fa> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -9101,25 +9113,25 @@ _sk_gather_4444_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,10,42,0,0 ; vbroadcastss 0x2a0a(%rip),%ymm0 # 6b88 <_sk_callback_avx+0x400> + DB 196,226,125,24,5,70,42,0,0 ; vbroadcastss 0x2a46(%rip),%ymm0 # 6bc4 <_sk_callback_avx+0x3fe> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,253,41,0,0 ; vbroadcastss 0x29fd(%rip),%ymm1 # 6b8c <_sk_callback_avx+0x404> + DB 196,226,125,24,13,57,42,0,0 ; vbroadcastss 0x2a39(%rip),%ymm1 # 6bc8 <_sk_callback_avx+0x402> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,244,41,0,0 ; vbroadcastss 0x29f4(%rip),%ymm1 # 6b90 <_sk_callback_avx+0x408> + DB 196,226,125,24,13,48,42,0,0 ; vbroadcastss 0x2a30(%rip),%ymm1 # 6bcc <_sk_callback_avx+0x406> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,231,41,0,0 ; vbroadcastss 0x29e7(%rip),%ymm2 # 6b94 <_sk_callback_avx+0x40c> + DB 196,226,125,24,21,35,42,0,0 ; vbroadcastss 0x2a23(%rip),%ymm2 # 6bd0 <_sk_callback_avx+0x40a> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,222,41,0,0 ; vbroadcastss 0x29de(%rip),%ymm2 # 6b98 <_sk_callback_avx+0x410> + DB 196,226,125,24,21,26,42,0,0 ; vbroadcastss 0x2a1a(%rip),%ymm2 # 6bd4 <_sk_callback_avx+0x40e> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,209,41,0,0 ; vbroadcastss 0x29d1(%rip),%ymm8 # 6b9c <_sk_callback_avx+0x414> + DB 196,98,125,24,5,13,42,0,0 ; vbroadcastss 0x2a0d(%rip),%ymm8 # 6bd8 <_sk_callback_avx+0x412> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,199,41,0,0 ; vbroadcastss 0x29c7(%rip),%ymm8 # 6ba0 <_sk_callback_avx+0x418> + DB 196,98,125,24,5,3,42,0,0 ; vbroadcastss 0x2a03(%rip),%ymm8 # 6bdc <_sk_callback_avx+0x416> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,185,41,0,0 ; vbroadcastss 0x29b9(%rip),%ymm8 # 6ba4 <_sk_callback_avx+0x41c> + DB 196,98,125,24,5,245,41,0,0 ; vbroadcastss 0x29f5(%rip),%ymm8 # 6be0 <_sk_callback_avx+0x41a> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -9128,7 +9140,7 @@ PUBLIC _sk_store_4444_avx _sk_store_4444_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 - DB 196,98,125,24,5,166,41,0,0 ; vbroadcastss 0x29a6(%rip),%ymm8 # 6ba8 <_sk_callback_avx+0x420> + DB 196,98,125,24,5,226,41,0,0 ; vbroadcastss 0x29e2(%rip),%ymm8 # 6be4 <_sk_callback_avx+0x41e> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,12 ; vpslld $0xc,%xmm9,%xmm10 @@ -9210,10 +9222,10 @@ _sk_load_8888_avx LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 15,133,139,0,0,0 ; jne 43b9 <_sk_load_8888_avx+0xa5> DB 196,193,124,16,26 ; vmovups (%r10),%ymm3 - DB 197,124,40,21,5,43,0,0 ; vmovaps 0x2b05(%rip),%ymm10 # 6e40 <_sk_callback_avx+0x6b8> + DB 197,124,40,21,69,43,0,0 ; vmovaps 0x2b45(%rip),%ymm10 # 6e80 <_sk_callback_avx+0x6ba> DB 196,193,100,84,194 ; vandps %ymm10,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,95,40,0,0 ; vbroadcastss 0x285f(%rip),%ymm8 # 6bac <_sk_callback_avx+0x424> + DB 196,98,125,24,5,155,40,0,0 ; vbroadcastss 0x289b(%rip),%ymm8 # 6be8 <_sk_callback_avx+0x422> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,241,114,211,8 ; vpsrld $0x8,%xmm3,%xmm1 DB 196,195,125,25,217,1 ; vextractf128 $0x1,%ymm3,%xmm9 @@ -9244,9 +9256,9 @@ _sk_load_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0 - DB 196,226,121,0,13,175,41,0,0 ; vpshufb 0x29af(%rip),%xmm0,%xmm1 # 6d90 <_sk_callback_avx+0x608> + DB 196,226,121,0,13,239,41,0,0 ; vpshufb 0x29ef(%rip),%xmm0,%xmm1 # 6dd0 <_sk_callback_avx+0x60a> DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1 - DB 196,226,121,0,5,177,41,0,0 ; vpshufb 0x29b1(%rip),%xmm0,%xmm0 # 6da0 <_sk_callback_avx+0x618> + DB 196,226,121,0,5,241,41,0,0 ; vpshufb 0x29f1(%rip),%xmm0,%xmm0 # 6de0 <_sk_callback_avx+0x61a> DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 196,194,125,44,26 ; vmaskmovps (%r10),%ymm0,%ymm3 @@ -9287,10 +9299,10 @@ _sk_gather_8888_avx LABEL PROC DB 73,193,234,32 ; shr $0x20,%r10 DB 196,131,121,34,28,145,3 ; vpinsrd $0x3,(%r9,%r10,4),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,177,41,0,0 ; vmovaps 0x29b1(%rip),%ymm10 # 6e60 <_sk_callback_avx+0x6d8> + DB 197,124,40,21,241,41,0,0 ; vmovaps 0x29f1(%rip),%ymm10 # 6ea0 <_sk_callback_avx+0x6da> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,239,38,0,0 ; vbroadcastss 0x26ef(%rip),%ymm9 # 6bb0 <_sk_callback_avx+0x428> + DB 196,98,125,24,13,43,39,0,0 ; vbroadcastss 0x272b(%rip),%ymm9 # 6bec <_sk_callback_avx+0x426> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -9319,7 +9331,7 @@ _sk_store_8888_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,141,20,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r10 DB 76,3,16 ; add (%rax),%r10 - DB 196,98,125,24,5,120,38,0,0 ; vbroadcastss 0x2678(%rip),%ymm8 # 6bb4 <_sk_callback_avx+0x42c> + DB 196,98,125,24,5,180,38,0,0 ; vbroadcastss 0x26b4(%rip),%ymm8 # 6bf0 <_sk_callback_avx+0x42a> 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 @@ -9357,9 +9369,9 @@ _sk_store_8888_avx LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 - DB 196,98,49,0,21,186,39,0,0 ; vpshufb 0x27ba(%rip),%xmm9,%xmm10 # 6db0 <_sk_callback_avx+0x628> + DB 196,98,49,0,21,250,39,0,0 ; vpshufb 0x27fa(%rip),%xmm9,%xmm10 # 6df0 <_sk_callback_avx+0x62a> DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10 - DB 196,98,49,0,13,188,39,0,0 ; vpshufb 0x27bc(%rip),%xmm9,%xmm9 # 6dc0 <_sk_callback_avx+0x638> + DB 196,98,49,0,13,252,39,0,0 ; vpshufb 0x27fc(%rip),%xmm9,%xmm9 # 6e00 <_sk_callback_avx+0x63a> DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 DB 196,66,53,46,2 ; vmaskmovps %ymm8,%ymm9,(%r10) @@ -9393,13 +9405,13 @@ _sk_load_f16_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,98,125,24,37,29,37,0,0 ; vbroadcastss 0x251d(%rip),%ymm12 # 6bb8 <_sk_callback_avx+0x430> + DB 196,98,125,24,37,89,37,0,0 ; vbroadcastss 0x2559(%rip),%ymm12 # 6bf4 <_sk_callback_avx+0x42e> DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,9,37,0,0 ; vbroadcastss 0x2509(%rip),%xmm11 # 6bbc <_sk_callback_avx+0x434> + DB 196,98,121,24,29,69,37,0,0 ; vbroadcastss 0x2545(%rip),%xmm11 # 6bf8 <_sk_callback_avx+0x432> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,255,36,0,0 ; vbroadcastss 0x24ff(%rip),%xmm13 # 6bc0 <_sk_callback_avx+0x438> + DB 196,98,121,24,45,59,37,0,0 ; vbroadcastss 0x253b(%rip),%xmm13 # 6bfc <_sk_callback_avx+0x436> DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -9413,7 +9425,7 @@ _sk_load_f16_avx LABEL PROC DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0 DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 - DB 196,226,121,24,29,181,36,0,0 ; vbroadcastss 0x24b5(%rip),%xmm3 # 6bc4 <_sk_callback_avx+0x43c> + DB 196,226,121,24,29,241,36,0,0 ; vbroadcastss 0x24f1(%rip),%xmm3 # 6c00 <_sk_callback_avx+0x43a> DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -9587,13 +9599,13 @@ _sk_gather_f16_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,98,125,24,37,118,33,0,0 ; vbroadcastss 0x2176(%rip),%ymm12 # 6bc8 <_sk_callback_avx+0x440> + DB 196,98,125,24,37,178,33,0,0 ; vbroadcastss 0x21b2(%rip),%ymm12 # 6c04 <_sk_callback_avx+0x43e> DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,98,33,0,0 ; vbroadcastss 0x2162(%rip),%xmm11 # 6bcc <_sk_callback_avx+0x444> + DB 196,98,121,24,29,158,33,0,0 ; vbroadcastss 0x219e(%rip),%xmm11 # 6c08 <_sk_callback_avx+0x442> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,88,33,0,0 ; vbroadcastss 0x2158(%rip),%xmm13 # 6bd0 <_sk_callback_avx+0x448> + DB 196,98,121,24,45,148,33,0,0 ; vbroadcastss 0x2194(%rip),%xmm13 # 6c0c <_sk_callback_avx+0x446> DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -9607,7 +9619,7 @@ _sk_gather_f16_avx LABEL PROC DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0 DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 - DB 196,226,121,24,29,14,33,0,0 ; vbroadcastss 0x210e(%rip),%xmm3 # 6bd4 <_sk_callback_avx+0x44c> + DB 196,226,121,24,29,74,33,0,0 ; vbroadcastss 0x214a(%rip),%xmm3 # 6c10 <_sk_callback_avx+0x44a> DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -9705,12 +9717,12 @@ _sk_store_f16_avx LABEL PROC DB 197,252,17,180,36,128,0,0,0 ; vmovups %ymm6,0x80(%rsp) DB 197,252,17,108,36,96 ; vmovups %ymm5,0x60(%rsp) DB 197,252,17,100,36,64 ; vmovups %ymm4,0x40(%rsp) - DB 196,98,125,24,13,34,31,0,0 ; vbroadcastss 0x1f22(%rip),%ymm9 # 6bd8 <_sk_callback_avx+0x450> + DB 196,98,125,24,13,94,31,0,0 ; vbroadcastss 0x1f5e(%rip),%ymm9 # 6c14 <_sk_callback_avx+0x44e> DB 196,65,124,84,209 ; vandps %ymm9,%ymm0,%ymm10 DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) DB 196,65,124,87,218 ; vxorps %ymm10,%ymm0,%ymm11 DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12 - DB 196,98,121,24,5,8,31,0,0 ; vbroadcastss 0x1f08(%rip),%xmm8 # 6bdc <_sk_callback_avx+0x454> + DB 196,98,121,24,5,68,31,0,0 ; vbroadcastss 0x1f44(%rip),%xmm8 # 6c18 <_sk_callback_avx+0x452> DB 196,65,57,102,236 ; vpcmpgtd %xmm12,%xmm8,%xmm13 DB 196,65,57,102,243 ; vpcmpgtd %xmm11,%xmm8,%xmm14 DB 196,67,13,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 @@ -9720,7 +9732,7 @@ _sk_store_f16_avx LABEL PROC DB 196,67,13,24,242,1 ; vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11 DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12 - DB 196,98,125,24,21,207,30,0,0 ; vbroadcastss 0x1ecf(%rip),%ymm10 # 6be0 <_sk_callback_avx+0x458> + DB 196,98,125,24,21,11,31,0,0 ; vbroadcastss 0x1f0b(%rip),%ymm10 # 6c1c <_sk_callback_avx+0x456> DB 196,65,12,86,242 ; vorps %ymm10,%ymm14,%ymm14 DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15 DB 196,65,1,254,228 ; vpaddd %xmm12,%xmm15,%xmm12 @@ -9863,7 +9875,7 @@ _sk_load_u16_be_avx LABEL PROC DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,30,28,0,0 ; vbroadcastss 0x1c1e(%rip),%ymm11 # 6be4 <_sk_callback_avx+0x45c> + DB 196,98,125,24,29,90,28,0,0 ; vbroadcastss 0x1c5a(%rip),%ymm11 # 6c20 <_sk_callback_avx+0x45a> DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,177,109,202 ; vpunpckhqdq %xmm2,%xmm9,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -9954,7 +9966,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,126,26,0,0 ; vbroadcastss 0x1a7e(%rip),%ymm11 # 6be8 <_sk_callback_avx+0x460> + DB 196,98,125,24,29,186,26,0,0 ; vbroadcastss 0x1aba(%rip),%ymm11 # 6c24 <_sk_callback_avx+0x45e> DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -9975,7 +9987,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,27,26,0,0 ; vbroadcastss 0x1a1b(%rip),%ymm3 # 6bec <_sk_callback_avx+0x464> + DB 196,226,125,24,29,87,26,0,0 ; vbroadcastss 0x1a57(%rip),%ymm3 # 6c28 <_sk_callback_avx+0x462> DB 255,224 ; jmpq *%rax DB 196,193,121,110,4,65 ; vmovd (%r9,%rax,2),%xmm0 DB 196,193,121,196,68,65,4,2 ; vpinsrw $0x2,0x4(%r9,%rax,2),%xmm0,%xmm0 @@ -10016,7 +10028,7 @@ _sk_store_u16_be_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,8 ; mov (%rax),%r9 DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax - DB 196,98,125,24,5,88,25,0,0 ; vbroadcastss 0x1958(%rip),%ymm8 # 6bf0 <_sk_callback_avx+0x468> + DB 196,98,125,24,5,148,25,0,0 ; vbroadcastss 0x1994(%rip),%ymm8 # 6c2c <_sk_callback_avx+0x466> 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 @@ -10207,55 +10219,12 @@ _sk_clamp_y_avx LABEL PROC PUBLIC _sk_repeat_x_avx _sk_repeat_x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 - DB 196,65,124,94,200 ; vdivps %ymm8,%ymm0,%ymm9 - DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 - DB 196,65,52,89,200 ; vmulps %ymm8,%ymm9,%ymm9 - DB 196,65,124,92,201 ; vsubps %ymm9,%ymm0,%ymm9 - DB 196,99,125,25,192,1 ; vextractf128 $0x1,%ymm8,%xmm0 - DB 196,65,41,118,210 ; vpcmpeqd %xmm10,%xmm10,%xmm10 - DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 - DB 196,65,57,254,194 ; vpaddd %xmm10,%xmm8,%xmm8 - DB 196,227,61,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm8,%ymm0 - DB 197,180,93,192 ; vminps %ymm0,%ymm9,%ymm0 - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - -PUBLIC _sk_repeat_y_avx -_sk_repeat_y_avx LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 - DB 196,65,116,94,200 ; vdivps %ymm8,%ymm1,%ymm9 - DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 - DB 196,65,52,89,200 ; vmulps %ymm8,%ymm9,%ymm9 - DB 196,65,116,92,201 ; vsubps %ymm9,%ymm1,%ymm9 - DB 196,99,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm1 - DB 196,65,41,118,210 ; vpcmpeqd %xmm10,%xmm10,%xmm10 - DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 - DB 196,65,57,254,194 ; vpaddd %xmm10,%xmm8,%xmm8 - DB 196,227,61,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 - DB 197,180,93,201 ; vminps %ymm1,%ymm9,%ymm1 - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - -PUBLIC _sk_mirror_x_avx -_sk_mirror_x_avx LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,121,110,0 ; vmovd (%rax),%xmm8 - DB 196,65,121,112,200,0 ; vpshufd $0x0,%xmm8,%xmm9 - DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 - DB 196,65,124,92,209 ; vsubps %ymm9,%ymm0,%ymm10 - DB 196,193,58,88,192 ; vaddss %xmm8,%xmm8,%xmm0 - DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,44,94,192 ; vdivps %ymm0,%ymm10,%ymm8 + DB 196,98,125,24,64,4 ; vbroadcastss 0x4(%rax),%ymm8 + DB 197,60,89,192 ; vmulps %ymm0,%ymm8,%ymm8 DB 196,67,125,8,192,1 ; vroundps $0x1,%ymm8,%ymm8 - DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 197,172,92,192 ; vsubps %ymm0,%ymm10,%ymm0 - DB 196,193,124,92,193 ; vsubps %ymm9,%ymm0,%ymm0 - DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 197,60,92,192 ; vsubps %ymm0,%ymm8,%ymm8 - DB 197,60,84,192 ; vandps %ymm0,%ymm8,%ymm8 + DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 + DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 + DB 196,65,124,92,192 ; vsubps %ymm8,%ymm0,%ymm8 DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 DB 196,65,41,118,210 ; vpcmpeqd %xmm10,%xmm10,%xmm10 DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 @@ -10265,24 +10234,15 @@ _sk_mirror_x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax -PUBLIC _sk_mirror_y_avx -_sk_mirror_y_avx LABEL PROC +PUBLIC _sk_repeat_y_avx +_sk_repeat_y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,121,110,0 ; vmovd (%rax),%xmm8 - DB 196,65,121,112,200,0 ; vpshufd $0x0,%xmm8,%xmm9 - DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 - DB 196,65,116,92,209 ; vsubps %ymm9,%ymm1,%ymm10 - DB 196,193,58,88,200 ; vaddss %xmm8,%xmm8,%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 197,44,94,193 ; vdivps %ymm1,%ymm10,%ymm8 + DB 196,98,125,24,64,4 ; vbroadcastss 0x4(%rax),%ymm8 + DB 197,60,89,193 ; vmulps %ymm1,%ymm8,%ymm8 DB 196,67,125,8,192,1 ; vroundps $0x1,%ymm8,%ymm8 - DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 - DB 197,172,92,201 ; vsubps %ymm1,%ymm10,%ymm1 - DB 196,193,116,92,201 ; vsubps %ymm9,%ymm1,%ymm1 - DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 197,60,92,193 ; vsubps %ymm1,%ymm8,%ymm8 - DB 197,60,84,193 ; vandps %ymm1,%ymm8,%ymm8 + DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 + DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 + DB 196,65,116,92,192 ; vsubps %ymm8,%ymm1,%ymm8 DB 196,99,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm1 DB 196,65,41,118,210 ; vpcmpeqd %xmm10,%xmm10,%xmm10 DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 @@ -10292,11 +10252,73 @@ _sk_mirror_y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_mirror_x_avx +_sk_mirror_x_avx LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 197,121,110,0 ; vmovd (%rax),%xmm8 + DB 197,122,16,72,4 ; vmovss 0x4(%rax),%xmm9 + DB 196,65,121,112,208,0 ; vpshufd $0x0,%xmm8,%xmm10 + DB 196,67,45,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 + DB 196,65,124,92,218 ; vsubps %ymm10,%ymm0,%ymm11 + DB 196,193,58,88,192 ; vaddss %xmm8,%xmm8,%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 197,178,89,5,199,21,0,0 ; vmulss 0x15c7(%rip),%xmm9,%xmm0 # 6c30 <_sk_callback_avx+0x46a> + DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 + DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 196,227,125,8,192,1 ; vroundps $0x1,%ymm0,%ymm0 + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 + DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 197,60,92,192 ; vsubps %ymm0,%ymm8,%ymm8 + DB 197,60,84,192 ; vandps %ymm0,%ymm8,%ymm8 + DB 196,99,125,25,208,1 ; vextractf128 $0x1,%ymm10,%xmm0 + DB 196,65,49,118,201 ; vpcmpeqd %xmm9,%xmm9,%xmm9 + DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 + DB 196,65,41,254,201 ; vpaddd %xmm9,%xmm10,%xmm9 + DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 + DB 197,188,93,192 ; vminps %ymm0,%ymm8,%ymm0 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_mirror_y_avx +_sk_mirror_y_avx LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 197,121,110,0 ; vmovd (%rax),%xmm8 + DB 197,122,16,72,4 ; vmovss 0x4(%rax),%xmm9 + DB 196,65,121,112,208,0 ; vpshufd $0x0,%xmm8,%xmm10 + DB 196,67,45,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 + DB 196,65,116,92,218 ; vsubps %ymm10,%ymm1,%ymm11 + DB 196,193,58,88,200 ; vaddss %xmm8,%xmm8,%xmm1 + DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 + DB 196,99,117,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm8 + DB 197,178,89,13,66,21,0,0 ; vmulss 0x1542(%rip),%xmm9,%xmm1 # 6c34 <_sk_callback_avx+0x46e> + DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 + DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 + DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 + DB 196,227,125,8,201,1 ; vroundps $0x1,%ymm1,%ymm1 + DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 + DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 + DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 197,60,92,193 ; vsubps %ymm1,%ymm8,%ymm8 + DB 197,60,84,193 ; vandps %ymm1,%ymm8,%ymm8 + DB 196,99,125,25,209,1 ; vextractf128 $0x1,%ymm10,%xmm1 + DB 196,65,49,118,201 ; vpcmpeqd %xmm9,%xmm9,%xmm9 + DB 196,193,113,254,201 ; vpaddd %xmm9,%xmm1,%xmm1 + DB 196,65,41,254,201 ; vpaddd %xmm9,%xmm10,%xmm9 + DB 196,227,53,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm9,%ymm1 + DB 197,188,93,201 ; vminps %ymm1,%ymm8,%ymm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_clamp_x_1_avx _sk_clamp_x_1_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,218,20,0,0 ; vbroadcastss 0x14da(%rip),%ymm8 # 6bf4 <_sk_callback_avx+0x46c> + DB 196,98,125,24,5,224,20,0,0 ; vbroadcastss 0x14e0(%rip),%ymm8 # 6c38 <_sk_callback_avx+0x472> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10310,9 +10332,9 @@ _sk_repeat_x_1_avx LABEL PROC PUBLIC _sk_mirror_x_1_avx _sk_mirror_x_1_avx LABEL PROC - DB 196,98,125,24,5,189,20,0,0 ; vbroadcastss 0x14bd(%rip),%ymm8 # 6bf8 <_sk_callback_avx+0x470> + DB 196,98,125,24,5,195,20,0,0 ; vbroadcastss 0x14c3(%rip),%ymm8 # 6c3c <_sk_callback_avx+0x476> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,179,20,0,0 ; vbroadcastss 0x14b3(%rip),%ymm9 # 6bfc <_sk_callback_avx+0x474> + DB 196,98,125,24,13,185,20,0,0 ; vbroadcastss 0x14b9(%rip),%ymm9 # 6c40 <_sk_callback_avx+0x47a> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -10326,12 +10348,12 @@ _sk_mirror_x_1_avx LABEL PROC PUBLIC _sk_luminance_to_alpha_avx _sk_luminance_to_alpha_avx LABEL PROC - DB 196,226,125,24,29,131,20,0,0 ; vbroadcastss 0x1483(%rip),%ymm3 # 6c00 <_sk_callback_avx+0x478> + DB 196,226,125,24,29,137,20,0,0 ; vbroadcastss 0x1489(%rip),%ymm3 # 6c44 <_sk_callback_avx+0x47e> DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,29,122,20,0,0 ; vbroadcastss 0x147a(%rip),%ymm3 # 6c04 <_sk_callback_avx+0x47c> + DB 196,226,125,24,29,128,20,0,0 ; vbroadcastss 0x1480(%rip),%ymm3 # 6c48 <_sk_callback_avx+0x482> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,109,20,0,0 ; vbroadcastss 0x146d(%rip),%ymm1 # 6c08 <_sk_callback_avx+0x480> + DB 196,226,125,24,13,115,20,0,0 ; vbroadcastss 0x1473(%rip),%ymm1 # 6c4c <_sk_callback_avx+0x486> DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 DB 197,252,88,217 ; vaddps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -10540,9 +10562,9 @@ _sk_evenly_spaced_gradient_avx LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 72,139,104,8 ; mov 0x8(%rax),%rbp DB 72,255,203 ; dec %rbx - DB 120,7 ; js 5afa <_sk_evenly_spaced_gradient_avx+0x28> + DB 120,7 ; js 5b38 <_sk_evenly_spaced_gradient_avx+0x28> DB 196,225,242,42,203 ; vcvtsi2ss %rbx,%xmm1,%xmm1 - DB 235,21 ; jmp 5b0f <_sk_evenly_spaced_gradient_avx+0x3d> + DB 235,21 ; jmp 5b4d <_sk_evenly_spaced_gradient_avx+0x3d> DB 73,137,217 ; mov %rbx,%r9 DB 73,209,233 ; shr %r9 DB 131,227,1 ; and $0x1,%ebx @@ -10699,18 +10721,18 @@ _sk_evenly_spaced_gradient_avx LABEL PROC PUBLIC _sk_gauss_a_to_rgba_avx _sk_gauss_a_to_rgba_avx LABEL PROC - DB 196,226,125,24,5,150,13,0,0 ; vbroadcastss 0xd96(%rip),%ymm0 # 6c0c <_sk_callback_avx+0x484> + DB 196,226,125,24,5,156,13,0,0 ; vbroadcastss 0xd9c(%rip),%ymm0 # 6c50 <_sk_callback_avx+0x48a> DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 196,226,125,24,13,141,13,0,0 ; vbroadcastss 0xd8d(%rip),%ymm1 # 6c10 <_sk_callback_avx+0x488> + DB 196,226,125,24,13,147,13,0,0 ; vbroadcastss 0xd93(%rip),%ymm1 # 6c54 <_sk_callback_avx+0x48e> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,128,13,0,0 ; vbroadcastss 0xd80(%rip),%ymm1 # 6c14 <_sk_callback_avx+0x48c> + DB 196,226,125,24,13,134,13,0,0 ; vbroadcastss 0xd86(%rip),%ymm1 # 6c58 <_sk_callback_avx+0x492> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,115,13,0,0 ; vbroadcastss 0xd73(%rip),%ymm1 # 6c18 <_sk_callback_avx+0x490> + DB 196,226,125,24,13,121,13,0,0 ; vbroadcastss 0xd79(%rip),%ymm1 # 6c5c <_sk_callback_avx+0x496> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,102,13,0,0 ; vbroadcastss 0xd66(%rip),%ymm1 # 6c1c <_sk_callback_avx+0x494> + DB 196,226,125,24,13,108,13,0,0 ; vbroadcastss 0xd6c(%rip),%ymm1 # 6c60 <_sk_callback_avx+0x49a> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -10732,12 +10754,12 @@ _sk_gradient_avx LABEL PROC DB 76,139,8 ; mov (%rax),%r9 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 73,131,249,2 ; cmp $0x2,%r9 - DB 114,80 ; jb 5f3c <_sk_gradient_avx+0x72> + DB 114,80 ; jb 5f7a <_sk_gradient_avx+0x72> DB 72,139,88,72 ; mov 0x48(%rax),%rbx DB 73,255,201 ; dec %r9 DB 72,131,195,4 ; add $0x4,%rbx DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 - DB 196,98,125,24,21,27,13,0,0 ; vbroadcastss 0xd1b(%rip),%ymm10 # 6c20 <_sk_callback_avx+0x498> + DB 196,98,125,24,21,33,13,0,0 ; vbroadcastss 0xd21(%rip),%ymm10 # 6c64 <_sk_callback_avx+0x49e> DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,98,125,24,3 ; vbroadcastss (%rbx),%ymm8 DB 197,60,194,192,2 ; vcmpleps %ymm0,%ymm8,%ymm8 @@ -10749,7 +10771,7 @@ _sk_gradient_avx LABEL PROC DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 72,131,195,4 ; add $0x4,%rbx DB 73,255,201 ; dec %r9 - DB 117,205 ; jne 5f09 <_sk_gradient_avx+0x3f> + DB 117,205 ; jne 5f47 <_sk_gradient_avx+0x3f> DB 196,195,249,22,201,1 ; vpextrq $0x1,%xmm1,%r9 DB 69,137,202 ; mov %r9d,%r10d DB 73,193,233,32 ; shr $0x20,%r9 @@ -10929,27 +10951,27 @@ _sk_xy_to_unit_angle_avx LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,14,9,0,0 ; vbroadcastss 0x90e(%rip),%ymm13 # 6c24 <_sk_callback_avx+0x49c> + DB 196,98,125,24,45,20,9,0,0 ; vbroadcastss 0x914(%rip),%ymm13 # 6c68 <_sk_callback_avx+0x4a2> DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,4,9,0,0 ; vbroadcastss 0x904(%rip),%ymm14 # 6c28 <_sk_callback_avx+0x4a0> + DB 196,98,125,24,53,10,9,0,0 ; vbroadcastss 0x90a(%rip),%ymm14 # 6c6c <_sk_callback_avx+0x4a6> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,245,8,0,0 ; vbroadcastss 0x8f5(%rip),%ymm14 # 6c2c <_sk_callback_avx+0x4a4> + DB 196,98,125,24,53,251,8,0,0 ; vbroadcastss 0x8fb(%rip),%ymm14 # 6c70 <_sk_callback_avx+0x4aa> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,229 ; vmulps %ymm13,%ymm12,%ymm12 - DB 196,98,125,24,45,230,8,0,0 ; vbroadcastss 0x8e6(%rip),%ymm13 # 6c30 <_sk_callback_avx+0x4a8> + DB 196,98,125,24,45,236,8,0,0 ; vbroadcastss 0x8ec(%rip),%ymm13 # 6c74 <_sk_callback_avx+0x4ae> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,209,8,0,0 ; vbroadcastss 0x8d1(%rip),%ymm10 # 6c34 <_sk_callback_avx+0x4ac> + DB 196,98,125,24,21,215,8,0,0 ; vbroadcastss 0x8d7(%rip),%ymm10 # 6c78 <_sk_callback_avx+0x4b2> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,187,8,0,0 ; vbroadcastss 0x8bb(%rip),%ymm10 # 6c38 <_sk_callback_avx+0x4b0> + DB 196,98,125,24,21,193,8,0,0 ; vbroadcastss 0x8c1(%rip),%ymm10 # 6c7c <_sk_callback_avx+0x4b6> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,165,8,0,0 ; vbroadcastss 0x8a5(%rip),%ymm10 # 6c3c <_sk_callback_avx+0x4b4> + DB 196,98,125,24,21,171,8,0,0 ; vbroadcastss 0x8ab(%rip),%ymm10 # 6c80 <_sk_callback_avx+0x4ba> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -10969,7 +10991,7 @@ _sk_xy_to_radius_avx LABEL PROC PUBLIC _sk_save_xy_avx _sk_save_xy_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,111,8,0,0 ; vbroadcastss 0x86f(%rip),%ymm8 # 6c40 <_sk_callback_avx+0x4b8> + DB 196,98,125,24,5,117,8,0,0 ; vbroadcastss 0x875(%rip),%ymm8 # 6c84 <_sk_callback_avx+0x4be> 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 @@ -11002,9 +11024,9 @@ _sk_accumulate_avx LABEL PROC PUBLIC _sk_bilinear_nx_avx _sk_bilinear_nx_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,251,7,0,0 ; vbroadcastss 0x7fb(%rip),%ymm0 # 6c44 <_sk_callback_avx+0x4bc> + DB 196,226,125,24,5,1,8,0,0 ; vbroadcastss 0x801(%rip),%ymm0 # 6c88 <_sk_callback_avx+0x4c2> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,242,7,0,0 ; vbroadcastss 0x7f2(%rip),%ymm8 # 6c48 <_sk_callback_avx+0x4c0> + DB 196,98,125,24,5,248,7,0,0 ; vbroadcastss 0x7f8(%rip),%ymm8 # 6c8c <_sk_callback_avx+0x4c6> 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 @@ -11013,7 +11035,7 @@ _sk_bilinear_nx_avx LABEL PROC PUBLIC _sk_bilinear_px_avx _sk_bilinear_px_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,218,7,0,0 ; vbroadcastss 0x7da(%rip),%ymm0 # 6c4c <_sk_callback_avx+0x4c4> + DB 196,226,125,24,5,224,7,0,0 ; vbroadcastss 0x7e0(%rip),%ymm0 # 6c90 <_sk_callback_avx+0x4ca> 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) @@ -11023,9 +11045,9 @@ _sk_bilinear_px_avx LABEL PROC PUBLIC _sk_bilinear_ny_avx _sk_bilinear_ny_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,190,7,0,0 ; vbroadcastss 0x7be(%rip),%ymm1 # 6c50 <_sk_callback_avx+0x4c8> + DB 196,226,125,24,13,196,7,0,0 ; vbroadcastss 0x7c4(%rip),%ymm1 # 6c94 <_sk_callback_avx+0x4ce> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,180,7,0,0 ; vbroadcastss 0x7b4(%rip),%ymm8 # 6c54 <_sk_callback_avx+0x4cc> + DB 196,98,125,24,5,186,7,0,0 ; vbroadcastss 0x7ba(%rip),%ymm8 # 6c98 <_sk_callback_avx+0x4d2> 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 @@ -11034,7 +11056,7 @@ _sk_bilinear_ny_avx LABEL PROC PUBLIC _sk_bilinear_py_avx _sk_bilinear_py_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,156,7,0,0 ; vbroadcastss 0x79c(%rip),%ymm1 # 6c58 <_sk_callback_avx+0x4d0> + DB 196,226,125,24,13,162,7,0,0 ; vbroadcastss 0x7a2(%rip),%ymm1 # 6c9c <_sk_callback_avx+0x4d6> 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) @@ -11044,14 +11066,14 @@ _sk_bilinear_py_avx LABEL PROC PUBLIC _sk_bicubic_n3x_avx _sk_bicubic_n3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,127,7,0,0 ; vbroadcastss 0x77f(%rip),%ymm0 # 6c5c <_sk_callback_avx+0x4d4> + DB 196,226,125,24,5,133,7,0,0 ; vbroadcastss 0x785(%rip),%ymm0 # 6ca0 <_sk_callback_avx+0x4da> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,118,7,0,0 ; vbroadcastss 0x776(%rip),%ymm8 # 6c60 <_sk_callback_avx+0x4d8> + DB 196,98,125,24,5,124,7,0,0 ; vbroadcastss 0x77c(%rip),%ymm8 # 6ca4 <_sk_callback_avx+0x4de> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,103,7,0,0 ; vbroadcastss 0x767(%rip),%ymm10 # 6c64 <_sk_callback_avx+0x4dc> + DB 196,98,125,24,21,109,7,0,0 ; vbroadcastss 0x76d(%rip),%ymm10 # 6ca8 <_sk_callback_avx+0x4e2> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,93,7,0,0 ; vbroadcastss 0x75d(%rip),%ymm10 # 6c68 <_sk_callback_avx+0x4e0> + DB 196,98,125,24,21,99,7,0,0 ; vbroadcastss 0x763(%rip),%ymm10 # 6cac <_sk_callback_avx+0x4e6> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -11061,19 +11083,19 @@ _sk_bicubic_n3x_avx LABEL PROC PUBLIC _sk_bicubic_n1x_avx _sk_bicubic_n1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,64,7,0,0 ; vbroadcastss 0x740(%rip),%ymm0 # 6c6c <_sk_callback_avx+0x4e4> + DB 196,226,125,24,5,70,7,0,0 ; vbroadcastss 0x746(%rip),%ymm0 # 6cb0 <_sk_callback_avx+0x4ea> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,55,7,0,0 ; vbroadcastss 0x737(%rip),%ymm8 # 6c70 <_sk_callback_avx+0x4e8> + DB 196,98,125,24,5,61,7,0,0 ; vbroadcastss 0x73d(%rip),%ymm8 # 6cb4 <_sk_callback_avx+0x4ee> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,45,7,0,0 ; vbroadcastss 0x72d(%rip),%ymm9 # 6c74 <_sk_callback_avx+0x4ec> + DB 196,98,125,24,13,51,7,0,0 ; vbroadcastss 0x733(%rip),%ymm9 # 6cb8 <_sk_callback_avx+0x4f2> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,35,7,0,0 ; vbroadcastss 0x723(%rip),%ymm10 # 6c78 <_sk_callback_avx+0x4f0> + DB 196,98,125,24,21,41,7,0,0 ; vbroadcastss 0x729(%rip),%ymm10 # 6cbc <_sk_callback_avx+0x4f6> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,20,7,0,0 ; vbroadcastss 0x714(%rip),%ymm10 # 6c7c <_sk_callback_avx+0x4f4> + DB 196,98,125,24,21,26,7,0,0 ; vbroadcastss 0x71a(%rip),%ymm10 # 6cc0 <_sk_callback_avx+0x4fa> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,5,7,0,0 ; vbroadcastss 0x705(%rip),%ymm9 # 6c80 <_sk_callback_avx+0x4f8> + DB 196,98,125,24,13,11,7,0,0 ; vbroadcastss 0x70b(%rip),%ymm9 # 6cc4 <_sk_callback_avx+0x4fe> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11082,17 +11104,17 @@ _sk_bicubic_n1x_avx LABEL PROC PUBLIC _sk_bicubic_p1x_avx _sk_bicubic_p1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,237,6,0,0 ; vbroadcastss 0x6ed(%rip),%ymm8 # 6c84 <_sk_callback_avx+0x4fc> + DB 196,98,125,24,5,243,6,0,0 ; vbroadcastss 0x6f3(%rip),%ymm8 # 6cc8 <_sk_callback_avx+0x502> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,223,6,0,0 ; vbroadcastss 0x6df(%rip),%ymm10 # 6c88 <_sk_callback_avx+0x500> + DB 196,98,125,24,21,229,6,0,0 ; vbroadcastss 0x6e5(%rip),%ymm10 # 6ccc <_sk_callback_avx+0x506> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,213,6,0,0 ; vbroadcastss 0x6d5(%rip),%ymm11 # 6c8c <_sk_callback_avx+0x504> + DB 196,98,125,24,29,219,6,0,0 ; vbroadcastss 0x6db(%rip),%ymm11 # 6cd0 <_sk_callback_avx+0x50a> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,188,6,0,0 ; vbroadcastss 0x6bc(%rip),%ymm9 # 6c90 <_sk_callback_avx+0x508> + DB 196,98,125,24,13,194,6,0,0 ; vbroadcastss 0x6c2(%rip),%ymm9 # 6cd4 <_sk_callback_avx+0x50e> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11101,13 +11123,13 @@ _sk_bicubic_p1x_avx LABEL PROC PUBLIC _sk_bicubic_p3x_avx _sk_bicubic_p3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,164,6,0,0 ; vbroadcastss 0x6a4(%rip),%ymm0 # 6c94 <_sk_callback_avx+0x50c> + DB 196,226,125,24,5,170,6,0,0 ; vbroadcastss 0x6aa(%rip),%ymm0 # 6cd8 <_sk_callback_avx+0x512> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,145,6,0,0 ; vbroadcastss 0x691(%rip),%ymm10 # 6c98 <_sk_callback_avx+0x510> + DB 196,98,125,24,21,151,6,0,0 ; vbroadcastss 0x697(%rip),%ymm10 # 6cdc <_sk_callback_avx+0x516> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,135,6,0,0 ; vbroadcastss 0x687(%rip),%ymm10 # 6c9c <_sk_callback_avx+0x514> + DB 196,98,125,24,21,141,6,0,0 ; vbroadcastss 0x68d(%rip),%ymm10 # 6ce0 <_sk_callback_avx+0x51a> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -11117,14 +11139,14 @@ _sk_bicubic_p3x_avx LABEL PROC PUBLIC _sk_bicubic_n3y_avx _sk_bicubic_n3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,106,6,0,0 ; vbroadcastss 0x66a(%rip),%ymm1 # 6ca0 <_sk_callback_avx+0x518> + DB 196,226,125,24,13,112,6,0,0 ; vbroadcastss 0x670(%rip),%ymm1 # 6ce4 <_sk_callback_avx+0x51e> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,96,6,0,0 ; vbroadcastss 0x660(%rip),%ymm8 # 6ca4 <_sk_callback_avx+0x51c> + DB 196,98,125,24,5,102,6,0,0 ; vbroadcastss 0x666(%rip),%ymm8 # 6ce8 <_sk_callback_avx+0x522> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,81,6,0,0 ; vbroadcastss 0x651(%rip),%ymm10 # 6ca8 <_sk_callback_avx+0x520> + DB 196,98,125,24,21,87,6,0,0 ; vbroadcastss 0x657(%rip),%ymm10 # 6cec <_sk_callback_avx+0x526> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,71,6,0,0 ; vbroadcastss 0x647(%rip),%ymm10 # 6cac <_sk_callback_avx+0x524> + DB 196,98,125,24,21,77,6,0,0 ; vbroadcastss 0x64d(%rip),%ymm10 # 6cf0 <_sk_callback_avx+0x52a> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -11134,19 +11156,19 @@ _sk_bicubic_n3y_avx LABEL PROC PUBLIC _sk_bicubic_n1y_avx _sk_bicubic_n1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,42,6,0,0 ; vbroadcastss 0x62a(%rip),%ymm1 # 6cb0 <_sk_callback_avx+0x528> + DB 196,226,125,24,13,48,6,0,0 ; vbroadcastss 0x630(%rip),%ymm1 # 6cf4 <_sk_callback_avx+0x52e> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,32,6,0,0 ; vbroadcastss 0x620(%rip),%ymm8 # 6cb4 <_sk_callback_avx+0x52c> + DB 196,98,125,24,5,38,6,0,0 ; vbroadcastss 0x626(%rip),%ymm8 # 6cf8 <_sk_callback_avx+0x532> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,22,6,0,0 ; vbroadcastss 0x616(%rip),%ymm9 # 6cb8 <_sk_callback_avx+0x530> + DB 196,98,125,24,13,28,6,0,0 ; vbroadcastss 0x61c(%rip),%ymm9 # 6cfc <_sk_callback_avx+0x536> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,12,6,0,0 ; vbroadcastss 0x60c(%rip),%ymm10 # 6cbc <_sk_callback_avx+0x534> + DB 196,98,125,24,21,18,6,0,0 ; vbroadcastss 0x612(%rip),%ymm10 # 6d00 <_sk_callback_avx+0x53a> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,253,5,0,0 ; vbroadcastss 0x5fd(%rip),%ymm10 # 6cc0 <_sk_callback_avx+0x538> + DB 196,98,125,24,21,3,6,0,0 ; vbroadcastss 0x603(%rip),%ymm10 # 6d04 <_sk_callback_avx+0x53e> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,238,5,0,0 ; vbroadcastss 0x5ee(%rip),%ymm9 # 6cc4 <_sk_callback_avx+0x53c> + DB 196,98,125,24,13,244,5,0,0 ; vbroadcastss 0x5f4(%rip),%ymm9 # 6d08 <_sk_callback_avx+0x542> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11155,17 +11177,17 @@ _sk_bicubic_n1y_avx LABEL PROC PUBLIC _sk_bicubic_p1y_avx _sk_bicubic_p1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,214,5,0,0 ; vbroadcastss 0x5d6(%rip),%ymm8 # 6cc8 <_sk_callback_avx+0x540> + DB 196,98,125,24,5,220,5,0,0 ; vbroadcastss 0x5dc(%rip),%ymm8 # 6d0c <_sk_callback_avx+0x546> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,199,5,0,0 ; vbroadcastss 0x5c7(%rip),%ymm10 # 6ccc <_sk_callback_avx+0x544> + DB 196,98,125,24,21,205,5,0,0 ; vbroadcastss 0x5cd(%rip),%ymm10 # 6d10 <_sk_callback_avx+0x54a> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,189,5,0,0 ; vbroadcastss 0x5bd(%rip),%ymm11 # 6cd0 <_sk_callback_avx+0x548> + DB 196,98,125,24,29,195,5,0,0 ; vbroadcastss 0x5c3(%rip),%ymm11 # 6d14 <_sk_callback_avx+0x54e> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,164,5,0,0 ; vbroadcastss 0x5a4(%rip),%ymm9 # 6cd4 <_sk_callback_avx+0x54c> + DB 196,98,125,24,13,170,5,0,0 ; vbroadcastss 0x5aa(%rip),%ymm9 # 6d18 <_sk_callback_avx+0x552> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11174,13 +11196,13 @@ _sk_bicubic_p1y_avx LABEL PROC PUBLIC _sk_bicubic_p3y_avx _sk_bicubic_p3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,140,5,0,0 ; vbroadcastss 0x58c(%rip),%ymm1 # 6cd8 <_sk_callback_avx+0x550> + DB 196,226,125,24,13,146,5,0,0 ; vbroadcastss 0x592(%rip),%ymm1 # 6d1c <_sk_callback_avx+0x556> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,120,5,0,0 ; vbroadcastss 0x578(%rip),%ymm10 # 6cdc <_sk_callback_avx+0x554> + DB 196,98,125,24,21,126,5,0,0 ; vbroadcastss 0x57e(%rip),%ymm10 # 6d20 <_sk_callback_avx+0x55a> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,110,5,0,0 ; vbroadcastss 0x56e(%rip),%ymm10 # 6ce0 <_sk_callback_avx+0x558> + DB 196,98,125,24,21,116,5,0,0 ; vbroadcastss 0x574(%rip),%ymm10 # 6d24 <_sk_callback_avx+0x55e> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -11302,25 +11324,25 @@ ALIGN 4 DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6965 <.literal4+0xb1> + DB 71,225,61 ; rex.RXB loope 69a1 <.literal4+0xb1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6975 <.literal4+0xc1> + DB 71,225,61 ; rex.RXB loope 69b1 <.literal4+0xc1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6985 <.literal4+0xd1> + DB 71,225,61 ; rex.RXB loope 69c1 <.literal4+0xd1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6995 <.literal4+0xe1> + DB 71,225,61 ; rex.RXB loope 69d1 <.literal4+0xe1> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) @@ -11372,7 +11394,7 @@ 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 69e9 <.literal4+0x135> + DB 224,7 ; loopne 6a25 <.literal4+0x135> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11388,10 +11410,10 @@ 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 6a10 <.literal4+0x15c> + DB 127,0 ; jg 6a4c <.literal4+0x15c> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6a89 <.literal4+0x1d5> + DB 119,115 ; ja 6ac5 <.literal4+0x1d5> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11405,10 +11427,10 @@ 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 6a44 <.literal4+0x190> + DB 127,0 ; jg 6a80 <.literal4+0x190> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6abd <.literal4+0x209> + DB 119,115 ; ja 6af9 <.literal4+0x209> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11422,10 +11444,10 @@ 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 6a78 <.literal4+0x1c4> + DB 127,0 ; jg 6ab4 <.literal4+0x1c4> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6af1 <.literal4+0x23d> + DB 119,115 ; ja 6b2d <.literal4+0x23d> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11439,10 +11461,10 @@ 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 6aac <.literal4+0x1f8> + DB 127,0 ; jg 6ae8 <.literal4+0x1f8> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6b25 <.literal4+0x271> + DB 119,115 ; ja 6b61 <.literal4+0x271> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11455,7 +11477,7 @@ 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 6b23 <.literal4+0x26f> + DB 127,67 ; jg 6b5f <.literal4+0x26f> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -11467,10 +11489,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 6b43 <.literal4+0x28f> + DB 118,63 ; jbe 6b7f <.literal4+0x28f> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 6b57 <.literal4+0x2a3> + DB 127,67 ; jg 6b93 <.literal4+0x2a3> DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax) DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax) @@ -11479,7 +11501,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6b39 <.literal4+0x285> + DB 224,7 ; loopne 6b75 <.literal4+0x285> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11491,7 +11513,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6b55 <.literal4+0x2a1> + DB 224,7 ; loopne 6b91 <.literal4+0x2a1> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11502,7 +11524,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 6baa <.literal4+0x2f6> + DB 124,66 ; jl 6be6 <.literal4+0x2f6> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -11520,9 +11542,9 @@ 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 6bed <.literal4+0x339> + DB 112,65 ; jo 6c29 <.literal4+0x339> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 6bfb <.literal4+0x347> + DB 127,67 ; jg 6c37 <.literal4+0x347> DB 0,128,0,0,0,0 ; add %al,0x0(%rax) DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) DB 0,0 ; add %al,(%rax) @@ -11538,7 +11560,11 @@ ALIGN 4 DB 0,128,55,0,0,128 ; add %al,-0x7fffffc9(%rax) DB 63 ; (bad) DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6c3b <.literal4+0x387> + DB 127,71 ; jg 6c77 <.literal4+0x387> + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,191,0,0,0 ; add %al,0xbf(%rax) @@ -11930,7 +11956,7 @@ _sk_seed_shader_sse41 LABEL PROC DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 - DB 15,40,21,103,80,0,0 ; movaps 0x5067(%rip),%xmm2 # 51c0 <_sk_callback_sse41+0xd8> + DB 15,40,21,151,80,0,0 ; movaps 0x5097(%rip),%xmm2 # 51f0 <_sk_callback_sse41+0xda> DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,16,7 ; movups (%rdi),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 @@ -11939,7 +11965,7 @@ _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 15,40,21,86,80,0,0 ; movaps 0x5056(%rip),%xmm2 # 51d0 <_sk_callback_sse41+0xe8> + DB 15,40,21,134,80,0,0 ; movaps 0x5086(%rip),%xmm2 # 5200 <_sk_callback_sse41+0xea> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -11957,14 +11983,14 @@ _sk_dither_sse41 LABEL PROC DB 102,68,15,110,193 ; movd %ecx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 - DB 102,68,15,111,21,36,80,0,0 ; movdqa 0x5024(%rip),%xmm10 # 51e0 <_sk_callback_sse41+0xf8> + DB 102,68,15,111,21,84,80,0,0 ; movdqa 0x5054(%rip),%xmm10 # 5210 <_sk_callback_sse41+0xfa> DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,65,15,114,243,5 ; pslld $0x5,%xmm11 DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 102,65,15,114,242,4 ; pslld $0x4,%xmm10 - DB 102,68,15,111,37,16,80,0,0 ; movdqa 0x5010(%rip),%xmm12 # 51f0 <_sk_callback_sse41+0x108> - DB 102,68,15,111,45,23,80,0,0 ; movdqa 0x5017(%rip),%xmm13 # 5200 <_sk_callback_sse41+0x118> + DB 102,68,15,111,37,64,80,0,0 ; movdqa 0x5040(%rip),%xmm12 # 5220 <_sk_callback_sse41+0x10a> + DB 102,68,15,111,45,71,80,0,0 ; movdqa 0x5047(%rip),%xmm13 # 5230 <_sk_callback_sse41+0x11a> DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 DB 102,69,15,219,245 ; pand %xmm13,%xmm14 DB 102,65,15,114,246,2 ; pslld $0x2,%xmm14 @@ -11980,8 +12006,8 @@ _sk_dither_sse41 LABEL PROC DB 102,69,15,235,245 ; por %xmm13,%xmm14 DB 102,69,15,235,240 ; por %xmm8,%xmm14 DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8 - DB 68,15,89,5,210,79,0,0 ; mulps 0x4fd2(%rip),%xmm8 # 5210 <_sk_callback_sse41+0x128> - DB 68,15,88,5,218,79,0,0 ; addps 0x4fda(%rip),%xmm8 # 5220 <_sk_callback_sse41+0x138> + DB 68,15,89,5,2,80,0,0 ; mulps 0x5002(%rip),%xmm8 # 5240 <_sk_callback_sse41+0x12a> + DB 68,15,88,5,10,80,0,0 ; addps 0x500a(%rip),%xmm8 # 5250 <_sk_callback_sse41+0x13a> DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 @@ -12048,7 +12074,7 @@ _sk_clear_sse41 LABEL PROC PUBLIC _sk_srcatop_sse41 _sk_srcatop_sse41 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,52,79,0,0 ; movaps 0x4f34(%rip),%xmm8 # 5230 <_sk_callback_sse41+0x148> + DB 68,15,40,5,100,79,0,0 ; movaps 0x4f64(%rip),%xmm8 # 5260 <_sk_callback_sse41+0x14a> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -12071,7 +12097,7 @@ 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 68,15,40,13,247,78,0,0 ; movaps 0x4ef7(%rip),%xmm9 # 5240 <_sk_callback_sse41+0x158> + DB 68,15,40,13,39,79,0,0 ; movaps 0x4f27(%rip),%xmm9 # 5270 <_sk_callback_sse41+0x15a> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -12112,7 +12138,7 @@ _sk_dstin_sse41 LABEL PROC PUBLIC _sk_srcout_sse41 _sk_srcout_sse41 LABEL PROC - DB 68,15,40,5,155,78,0,0 ; movaps 0x4e9b(%rip),%xmm8 # 5250 <_sk_callback_sse41+0x168> + DB 68,15,40,5,203,78,0,0 ; movaps 0x4ecb(%rip),%xmm8 # 5280 <_sk_callback_sse41+0x16a> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -12123,7 +12149,7 @@ _sk_srcout_sse41 LABEL PROC PUBLIC _sk_dstout_sse41 _sk_dstout_sse41 LABEL PROC - DB 68,15,40,5,139,78,0,0 ; movaps 0x4e8b(%rip),%xmm8 # 5260 <_sk_callback_sse41+0x178> + DB 68,15,40,5,187,78,0,0 ; movaps 0x4ebb(%rip),%xmm8 # 5290 <_sk_callback_sse41+0x17a> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -12138,7 +12164,7 @@ _sk_dstout_sse41 LABEL PROC PUBLIC _sk_srcover_sse41 _sk_srcover_sse41 LABEL PROC - DB 68,15,40,5,110,78,0,0 ; movaps 0x4e6e(%rip),%xmm8 # 5270 <_sk_callback_sse41+0x188> + DB 68,15,40,5,158,78,0,0 ; movaps 0x4e9e(%rip),%xmm8 # 52a0 <_sk_callback_sse41+0x18a> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -12156,7 +12182,7 @@ _sk_srcover_sse41 LABEL PROC PUBLIC _sk_dstover_sse41 _sk_dstover_sse41 LABEL PROC - DB 68,15,40,5,66,78,0,0 ; movaps 0x4e42(%rip),%xmm8 # 5280 <_sk_callback_sse41+0x198> + DB 68,15,40,5,114,78,0,0 ; movaps 0x4e72(%rip),%xmm8 # 52b0 <_sk_callback_sse41+0x19a> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -12180,7 +12206,7 @@ _sk_modulate_sse41 LABEL PROC PUBLIC _sk_multiply_sse41 _sk_multiply_sse41 LABEL PROC - DB 68,15,40,5,22,78,0,0 ; movaps 0x4e16(%rip),%xmm8 # 5290 <_sk_callback_sse41+0x1a8> + DB 68,15,40,5,70,78,0,0 ; movaps 0x4e46(%rip),%xmm8 # 52c0 <_sk_callback_sse41+0x1aa> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -12249,7 +12275,7 @@ _sk_screen_sse41 LABEL PROC PUBLIC _sk_xor__sse41 _sk_xor__sse41 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,75,77,0,0 ; movaps 0x4d4b(%rip),%xmm3 # 52a0 <_sk_callback_sse41+0x1b8> + DB 15,40,29,123,77,0,0 ; movaps 0x4d7b(%rip),%xmm3 # 52d0 <_sk_callback_sse41+0x1ba> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -12295,7 +12321,7 @@ _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 15,40,21,182,76,0,0 ; movaps 0x4cb6(%rip),%xmm2 # 52b0 <_sk_callback_sse41+0x1c8> + DB 15,40,21,230,76,0,0 ; movaps 0x4ce6(%rip),%xmm2 # 52e0 <_sk_callback_sse41+0x1ca> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12327,7 +12353,7 @@ _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 15,40,21,91,76,0,0 ; movaps 0x4c5b(%rip),%xmm2 # 52c0 <_sk_callback_sse41+0x1d8> + DB 15,40,21,139,76,0,0 ; movaps 0x4c8b(%rip),%xmm2 # 52f0 <_sk_callback_sse41+0x1da> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12362,7 +12388,7 @@ _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 15,40,21,245,75,0,0 ; movaps 0x4bf5(%rip),%xmm2 # 52d0 <_sk_callback_sse41+0x1e8> + DB 15,40,21,37,76,0,0 ; movaps 0x4c25(%rip),%xmm2 # 5300 <_sk_callback_sse41+0x1ea> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12388,7 +12414,7 @@ _sk_exclusion_sse41 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,181,75,0,0 ; movaps 0x4bb5(%rip),%xmm2 # 52e0 <_sk_callback_sse41+0x1f8> + DB 15,40,21,229,75,0,0 ; movaps 0x4be5(%rip),%xmm2 # 5310 <_sk_callback_sse41+0x1fa> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -12399,7 +12425,7 @@ _sk_exclusion_sse41 LABEL PROC PUBLIC _sk_colorburn_sse41 _sk_colorburn_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,168,75,0,0 ; movaps 0x4ba8(%rip),%xmm10 # 52f0 <_sk_callback_sse41+0x208> + DB 68,15,40,21,216,75,0,0 ; movaps 0x4bd8(%rip),%xmm10 # 5320 <_sk_callback_sse41+0x20a> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,203 ; movaps %xmm11,%xmm9 @@ -12479,7 +12505,7 @@ _sk_colorburn_sse41 LABEL PROC PUBLIC _sk_colordodge_sse41 _sk_colordodge_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,134,74,0,0 ; movaps 0x4a86(%rip),%xmm10 # 5300 <_sk_callback_sse41+0x218> + DB 68,15,40,21,182,74,0,0 ; movaps 0x4ab6(%rip),%xmm10 # 5330 <_sk_callback_sse41+0x21a> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -12560,7 +12586,7 @@ _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 68,15,40,21,92,73,0,0 ; movaps 0x495c(%rip),%xmm10 # 5310 <_sk_callback_sse41+0x228> + DB 68,15,40,21,140,73,0,0 ; movaps 0x498c(%rip),%xmm10 # 5340 <_sk_callback_sse41+0x22a> DB 65,15,40,234 ; movaps %xmm10,%xmm5 DB 15,92,239 ; subps %xmm7,%xmm5 DB 15,40,197 ; movaps %xmm5,%xmm0 @@ -12642,7 +12668,7 @@ 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 68,15,40,21,62,72,0,0 ; movaps 0x483e(%rip),%xmm10 # 5320 <_sk_callback_sse41+0x238> + DB 68,15,40,21,110,72,0,0 ; movaps 0x486e(%rip),%xmm10 # 5350 <_sk_callback_sse41+0x23a> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 65,15,40,195 ; movaps %xmm11,%xmm0 @@ -12726,7 +12752,7 @@ _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 15,40,13,17,71,0,0 ; movaps 0x4711(%rip),%xmm1 # 5330 <_sk_callback_sse41+0x248> + DB 15,40,13,65,71,0,0 ; movaps 0x4741(%rip),%xmm1 # 5360 <_sk_callback_sse41+0x24a> DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 68,15,40,240 ; movaps %xmm0,%xmm14 @@ -12739,10 +12765,10 @@ _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 68,15,40,45,239,70,0,0 ; movaps 0x46ef(%rip),%xmm13 # 5340 <_sk_callback_sse41+0x258> + DB 68,15,40,45,31,71,0,0 ; movaps 0x471f(%rip),%xmm13 # 5370 <_sk_callback_sse41+0x25a> DB 69,15,88,245 ; addps %xmm13,%xmm14 DB 68,15,89,242 ; mulps %xmm2,%xmm14 - DB 68,15,40,37,239,70,0,0 ; movaps 0x46ef(%rip),%xmm12 # 5350 <_sk_callback_sse41+0x268> + DB 68,15,40,37,31,71,0,0 ; movaps 0x471f(%rip),%xmm12 # 5380 <_sk_callback_sse41+0x26a> DB 69,15,89,252 ; mulps %xmm12,%xmm15 DB 69,15,88,254 ; addps %xmm14,%xmm15 DB 15,40,198 ; movaps %xmm6,%xmm0 @@ -12928,12 +12954,12 @@ _sk_hue_sse41 LABEL PROC DB 68,15,84,208 ; andps %xmm0,%xmm10 DB 15,84,200 ; andps %xmm0,%xmm1 DB 68,15,84,232 ; andps %xmm0,%xmm13 - DB 15,40,5,85,68,0,0 ; movaps 0x4455(%rip),%xmm0 # 5360 <_sk_callback_sse41+0x278> + DB 15,40,5,133,68,0,0 ; movaps 0x4485(%rip),%xmm0 # 5390 <_sk_callback_sse41+0x27a> DB 68,15,89,224 ; mulps %xmm0,%xmm12 - DB 15,40,21,90,68,0,0 ; movaps 0x445a(%rip),%xmm2 # 5370 <_sk_callback_sse41+0x288> + DB 15,40,21,138,68,0,0 ; movaps 0x448a(%rip),%xmm2 # 53a0 <_sk_callback_sse41+0x28a> DB 15,89,250 ; mulps %xmm2,%xmm7 DB 65,15,88,252 ; addps %xmm12,%xmm7 - DB 68,15,40,53,91,68,0,0 ; movaps 0x445b(%rip),%xmm14 # 5380 <_sk_callback_sse41+0x298> + DB 68,15,40,53,139,68,0,0 ; movaps 0x448b(%rip),%xmm14 # 53b0 <_sk_callback_sse41+0x29a> DB 68,15,40,252 ; movaps %xmm4,%xmm15 DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 68,15,88,255 ; addps %xmm7,%xmm15 @@ -13016,7 +13042,7 @@ _sk_hue_sse41 LABEL PROC DB 65,15,88,214 ; addps %xmm14,%xmm2 DB 15,40,196 ; movaps %xmm4,%xmm0 DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 - DB 68,15,40,13,32,67,0,0 ; movaps 0x4320(%rip),%xmm9 # 5390 <_sk_callback_sse41+0x2a8> + DB 68,15,40,13,80,67,0,0 ; movaps 0x4350(%rip),%xmm9 # 53c0 <_sk_callback_sse41+0x2aa> DB 65,15,40,225 ; movaps %xmm9,%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 15,40,68,36,48 ; movaps 0x30(%rsp),%xmm0 @@ -13110,14 +13136,14 @@ _sk_saturation_sse41 LABEL PROC DB 68,15,84,215 ; andps %xmm7,%xmm10 DB 68,15,84,223 ; andps %xmm7,%xmm11 DB 68,15,84,199 ; andps %xmm7,%xmm8 - DB 15,40,21,211,65,0,0 ; movaps 0x41d3(%rip),%xmm2 # 53a0 <_sk_callback_sse41+0x2b8> + DB 15,40,21,3,66,0,0 ; movaps 0x4203(%rip),%xmm2 # 53d0 <_sk_callback_sse41+0x2ba> DB 15,40,221 ; movaps %xmm5,%xmm3 DB 15,89,218 ; mulps %xmm2,%xmm3 - DB 15,40,13,214,65,0,0 ; movaps 0x41d6(%rip),%xmm1 # 53b0 <_sk_callback_sse41+0x2c8> + DB 15,40,13,6,66,0,0 ; movaps 0x4206(%rip),%xmm1 # 53e0 <_sk_callback_sse41+0x2ca> DB 15,40,254 ; movaps %xmm6,%xmm7 DB 15,89,249 ; mulps %xmm1,%xmm7 DB 15,88,251 ; addps %xmm3,%xmm7 - DB 68,15,40,45,213,65,0,0 ; movaps 0x41d5(%rip),%xmm13 # 53c0 <_sk_callback_sse41+0x2d8> + DB 68,15,40,45,5,66,0,0 ; movaps 0x4205(%rip),%xmm13 # 53f0 <_sk_callback_sse41+0x2da> DB 69,15,89,245 ; mulps %xmm13,%xmm14 DB 68,15,88,247 ; addps %xmm7,%xmm14 DB 65,15,40,218 ; movaps %xmm10,%xmm3 @@ -13198,7 +13224,7 @@ _sk_saturation_sse41 LABEL PROC DB 65,15,88,253 ; addps %xmm13,%xmm7 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 102,68,15,56,20,223 ; blendvps %xmm0,%xmm7,%xmm11 - DB 68,15,40,13,155,64,0,0 ; movaps 0x409b(%rip),%xmm9 # 53d0 <_sk_callback_sse41+0x2e8> + DB 68,15,40,13,203,64,0,0 ; movaps 0x40cb(%rip),%xmm9 # 5400 <_sk_callback_sse41+0x2ea> DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 68,15,92,204 ; subps %xmm4,%xmm9 DB 15,40,60,36 ; movaps (%rsp),%xmm7 @@ -13253,14 +13279,14 @@ _sk_color_sse41 LABEL PROC DB 15,40,231 ; movaps %xmm7,%xmm4 DB 68,15,89,244 ; mulps %xmm4,%xmm14 DB 15,89,204 ; mulps %xmm4,%xmm1 - DB 68,15,40,13,224,63,0,0 ; movaps 0x3fe0(%rip),%xmm9 # 53e0 <_sk_callback_sse41+0x2f8> + DB 68,15,40,13,16,64,0,0 ; movaps 0x4010(%rip),%xmm9 # 5410 <_sk_callback_sse41+0x2fa> DB 65,15,40,250 ; movaps %xmm10,%xmm7 DB 65,15,89,249 ; mulps %xmm9,%xmm7 - DB 68,15,40,21,224,63,0,0 ; movaps 0x3fe0(%rip),%xmm10 # 53f0 <_sk_callback_sse41+0x308> + DB 68,15,40,21,16,64,0,0 ; movaps 0x4010(%rip),%xmm10 # 5420 <_sk_callback_sse41+0x30a> DB 65,15,40,219 ; movaps %xmm11,%xmm3 DB 65,15,89,218 ; mulps %xmm10,%xmm3 DB 15,88,223 ; addps %xmm7,%xmm3 - DB 68,15,40,29,221,63,0,0 ; movaps 0x3fdd(%rip),%xmm11 # 5400 <_sk_callback_sse41+0x318> + DB 68,15,40,29,13,64,0,0 ; movaps 0x400d(%rip),%xmm11 # 5430 <_sk_callback_sse41+0x31a> DB 69,15,40,236 ; movaps %xmm12,%xmm13 DB 69,15,89,235 ; mulps %xmm11,%xmm13 DB 68,15,88,235 ; addps %xmm3,%xmm13 @@ -13345,7 +13371,7 @@ _sk_color_sse41 LABEL PROC DB 65,15,88,251 ; addps %xmm11,%xmm7 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 102,15,56,20,207 ; blendvps %xmm0,%xmm7,%xmm1 - DB 68,15,40,13,153,62,0,0 ; movaps 0x3e99(%rip),%xmm9 # 5410 <_sk_callback_sse41+0x328> + DB 68,15,40,13,201,62,0,0 ; movaps 0x3ec9(%rip),%xmm9 # 5440 <_sk_callback_sse41+0x32a> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,92,196 ; subps %xmm4,%xmm0 DB 68,15,89,192 ; mulps %xmm0,%xmm8 @@ -13397,13 +13423,13 @@ _sk_luminosity_sse41 LABEL PROC DB 69,15,89,216 ; mulps %xmm8,%xmm11 DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,89,205 ; mulps %xmm5,%xmm9 - DB 68,15,40,5,235,61,0,0 ; movaps 0x3deb(%rip),%xmm8 # 5420 <_sk_callback_sse41+0x338> + DB 68,15,40,5,27,62,0,0 ; movaps 0x3e1b(%rip),%xmm8 # 5450 <_sk_callback_sse41+0x33a> DB 65,15,89,192 ; mulps %xmm8,%xmm0 - DB 68,15,40,21,239,61,0,0 ; movaps 0x3def(%rip),%xmm10 # 5430 <_sk_callback_sse41+0x348> + DB 68,15,40,21,31,62,0,0 ; movaps 0x3e1f(%rip),%xmm10 # 5460 <_sk_callback_sse41+0x34a> DB 15,40,233 ; movaps %xmm1,%xmm5 DB 65,15,89,234 ; mulps %xmm10,%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 - DB 68,15,40,37,237,61,0,0 ; movaps 0x3ded(%rip),%xmm12 # 5440 <_sk_callback_sse41+0x358> + DB 68,15,40,37,29,62,0,0 ; movaps 0x3e1d(%rip),%xmm12 # 5470 <_sk_callback_sse41+0x35a> DB 68,15,40,242 ; movaps %xmm2,%xmm14 DB 69,15,89,244 ; mulps %xmm12,%xmm14 DB 68,15,88,245 ; addps %xmm5,%xmm14 @@ -13488,7 +13514,7 @@ _sk_luminosity_sse41 LABEL PROC DB 65,15,88,244 ; addps %xmm12,%xmm6 DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 102,68,15,56,20,206 ; blendvps %xmm0,%xmm6,%xmm9 - DB 15,40,5,163,60,0,0 ; movaps 0x3ca3(%rip),%xmm0 # 5450 <_sk_callback_sse41+0x368> + DB 15,40,5,211,60,0,0 ; movaps 0x3cd3(%rip),%xmm0 # 5480 <_sk_callback_sse41+0x36a> DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,92,215 ; subps %xmm7,%xmm2 DB 15,89,226 ; mulps %xmm2,%xmm4 @@ -13530,20 +13556,20 @@ _sk_srcover_rgba_8888_sse41 LABEL PROC DB 15,133,217,0,0,0 ; jne 1904 <_sk_srcover_rgba_8888_sse41+0xe7> DB 243,15,111,60,144 ; movdqu (%rax,%rdx,4),%xmm7 DB 77,133,192 ; test %r8,%r8 - DB 102,15,111,37,37,60,0,0 ; movdqa 0x3c25(%rip),%xmm4 # 5460 <_sk_callback_sse41+0x378> + DB 102,15,111,37,85,60,0,0 ; movdqa 0x3c55(%rip),%xmm4 # 5490 <_sk_callback_sse41+0x37a> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,56,0,45,33,60,0,0 ; pshufb 0x3c21(%rip),%xmm5 # 5470 <_sk_callback_sse41+0x388> + DB 102,15,56,0,45,81,60,0,0 ; pshufb 0x3c51(%rip),%xmm5 # 54a0 <_sk_callback_sse41+0x38a> DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,56,0,53,33,60,0,0 ; pshufb 0x3c21(%rip),%xmm6 # 5480 <_sk_callback_sse41+0x398> + DB 102,15,56,0,53,81,60,0,0 ; pshufb 0x3c51(%rip),%xmm6 # 54b0 <_sk_callback_sse41+0x39a> DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 DB 102,15,114,215,24 ; psrld $0x18,%xmm7 DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 - DB 68,15,40,5,30,60,0,0 ; movaps 0x3c1e(%rip),%xmm8 # 5490 <_sk_callback_sse41+0x3a8> + DB 68,15,40,5,78,60,0,0 ; movaps 0x3c4e(%rip),%xmm8 # 54c0 <_sk_callback_sse41+0x3aa> DB 68,15,92,195 ; subps %xmm3,%xmm8 - DB 68,15,40,37,34,60,0,0 ; movaps 0x3c22(%rip),%xmm12 # 54a0 <_sk_callback_sse41+0x3b8> + DB 68,15,40,37,82,60,0,0 ; movaps 0x3c52(%rip),%xmm12 # 54d0 <_sk_callback_sse41+0x3ba> DB 65,15,89,196 ; mulps %xmm12,%xmm0 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -13616,7 +13642,7 @@ _sk_clamp_0_sse41 LABEL PROC PUBLIC _sk_clamp_1_sse41 _sk_clamp_1_sse41 LABEL PROC - DB 68,15,40,5,25,59,0,0 ; movaps 0x3b19(%rip),%xmm8 # 54b0 <_sk_callback_sse41+0x3c8> + DB 68,15,40,5,73,59,0,0 ; movaps 0x3b49(%rip),%xmm8 # 54e0 <_sk_callback_sse41+0x3ca> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -13626,7 +13652,7 @@ _sk_clamp_1_sse41 LABEL PROC PUBLIC _sk_clamp_a_sse41 _sk_clamp_a_sse41 LABEL PROC - DB 15,93,29,14,59,0,0 ; minps 0x3b0e(%rip),%xmm3 # 54c0 <_sk_callback_sse41+0x3d8> + DB 15,93,29,62,59,0,0 ; minps 0x3b3e(%rip),%xmm3 # 54f0 <_sk_callback_sse41+0x3da> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -13699,7 +13725,7 @@ _sk_premul_sse41 LABEL PROC PUBLIC _sk_unpremul_sse41 _sk_unpremul_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,121,58,0,0 ; movaps 0x3a79(%rip),%xmm9 # 54d0 <_sk_callback_sse41+0x3e8> + DB 68,15,40,13,169,58,0,0 ; movaps 0x3aa9(%rip),%xmm9 # 5500 <_sk_callback_sse41+0x3ea> 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 @@ -13711,20 +13737,20 @@ _sk_unpremul_sse41 LABEL PROC PUBLIC _sk_from_srgb_sse41 _sk_from_srgb_sse41 LABEL PROC - DB 68,15,40,29,100,58,0,0 ; movaps 0x3a64(%rip),%xmm11 # 54e0 <_sk_callback_sse41+0x3f8> + DB 68,15,40,29,148,58,0,0 ; movaps 0x3a94(%rip),%xmm11 # 5510 <_sk_callback_sse41+0x3fa> 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 68,15,40,37,92,58,0,0 ; movaps 0x3a5c(%rip),%xmm12 # 54f0 <_sk_callback_sse41+0x408> + DB 68,15,40,37,140,58,0,0 ; movaps 0x3a8c(%rip),%xmm12 # 5520 <_sk_callback_sse41+0x40a> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,92,58,0,0 ; movaps 0x3a5c(%rip),%xmm13 # 5500 <_sk_callback_sse41+0x418> + DB 68,15,40,45,140,58,0,0 ; movaps 0x3a8c(%rip),%xmm13 # 5530 <_sk_callback_sse41+0x41a> DB 69,15,88,197 ; addps %xmm13,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 - DB 68,15,40,53,92,58,0,0 ; movaps 0x3a5c(%rip),%xmm14 # 5510 <_sk_callback_sse41+0x428> + DB 68,15,40,53,140,58,0,0 ; movaps 0x3a8c(%rip),%xmm14 # 5540 <_sk_callback_sse41+0x42a> DB 69,15,88,198 ; addps %xmm14,%xmm8 - DB 68,15,40,61,96,58,0,0 ; movaps 0x3a60(%rip),%xmm15 # 5520 <_sk_callback_sse41+0x438> + DB 68,15,40,61,144,58,0,0 ; movaps 0x3a90(%rip),%xmm15 # 5550 <_sk_callback_sse41+0x43a> 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 @@ -13766,22 +13792,22 @@ _sk_to_srgb_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,209 ; movaps %xmm1,%xmm2 DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 - DB 68,15,40,29,213,57,0,0 ; movaps 0x39d5(%rip),%xmm11 # 5530 <_sk_callback_sse41+0x448> + DB 68,15,40,29,5,58,0,0 ; movaps 0x3a05(%rip),%xmm11 # 5560 <_sk_callback_sse41+0x44a> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 - DB 68,15,40,37,213,57,0,0 ; movaps 0x39d5(%rip),%xmm12 # 5540 <_sk_callback_sse41+0x458> + DB 68,15,40,37,5,58,0,0 ; movaps 0x3a05(%rip),%xmm12 # 5570 <_sk_callback_sse41+0x45a> DB 69,15,40,248 ; movaps %xmm8,%xmm15 DB 69,15,89,252 ; mulps %xmm12,%xmm15 - DB 68,15,40,21,213,57,0,0 ; movaps 0x39d5(%rip),%xmm10 # 5550 <_sk_callback_sse41+0x468> + DB 68,15,40,21,5,58,0,0 ; movaps 0x3a05(%rip),%xmm10 # 5580 <_sk_callback_sse41+0x46a> DB 69,15,88,250 ; addps %xmm10,%xmm15 DB 69,15,89,248 ; mulps %xmm8,%xmm15 - DB 68,15,40,45,213,57,0,0 ; movaps 0x39d5(%rip),%xmm13 # 5560 <_sk_callback_sse41+0x478> + DB 68,15,40,45,5,58,0,0 ; movaps 0x3a05(%rip),%xmm13 # 5590 <_sk_callback_sse41+0x47a> DB 69,15,88,253 ; addps %xmm13,%xmm15 - DB 68,15,40,53,217,57,0,0 ; movaps 0x39d9(%rip),%xmm14 # 5570 <_sk_callback_sse41+0x488> + DB 68,15,40,53,9,58,0,0 ; movaps 0x3a09(%rip),%xmm14 # 55a0 <_sk_callback_sse41+0x48a> DB 69,15,88,198 ; addps %xmm14,%xmm8 DB 69,15,83,192 ; rcpps %xmm8,%xmm8 DB 69,15,89,199 ; mulps %xmm15,%xmm8 - DB 68,15,40,61,213,57,0,0 ; movaps 0x39d5(%rip),%xmm15 # 5580 <_sk_callback_sse41+0x498> + DB 68,15,40,61,5,58,0,0 ; movaps 0x3a05(%rip),%xmm15 # 55b0 <_sk_callback_sse41+0x49a> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 68,15,82,202 ; rsqrtps %xmm2,%xmm9 @@ -13833,7 +13859,7 @@ _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 68,15,40,53,31,57,0,0 ; movaps 0x391f(%rip),%xmm14 # 5590 <_sk_callback_sse41+0x4a8> + DB 68,15,40,53,79,57,0,0 ; movaps 0x394f(%rip),%xmm14 # 55c0 <_sk_callback_sse41+0x4aa> 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 @@ -13842,27 +13868,27 @@ _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 68,15,84,61,6,57,0,0 ; andps 0x3906(%rip),%xmm15 # 55a0 <_sk_callback_sse41+0x4b8> + DB 68,15,84,61,54,57,0,0 ; andps 0x3936(%rip),%xmm15 # 55d0 <_sk_callback_sse41+0x4ba> 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 68,15,40,45,249,56,0,0 ; movaps 0x38f9(%rip),%xmm13 # 55b0 <_sk_callback_sse41+0x4c8> + DB 68,15,40,45,41,57,0,0 ; movaps 0x3929(%rip),%xmm13 # 55e0 <_sk_callback_sse41+0x4ca> 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 68,15,88,5,245,56,0,0 ; addps 0x38f5(%rip),%xmm8 # 55c0 <_sk_callback_sse41+0x4d8> + DB 68,15,88,5,37,57,0,0 ; addps 0x3925(%rip),%xmm8 # 55f0 <_sk_callback_sse41+0x4da> 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 68,15,89,5,237,56,0,0 ; mulps 0x38ed(%rip),%xmm8 # 55d0 <_sk_callback_sse41+0x4e8> + DB 68,15,89,5,29,57,0,0 ; mulps 0x391d(%rip),%xmm8 # 5600 <_sk_callback_sse41+0x4ea> 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 15,40,5,225,56,0,0 ; movaps 0x38e1(%rip),%xmm0 # 55e0 <_sk_callback_sse41+0x4f8> + DB 15,40,5,17,57,0,0 ; movaps 0x3911(%rip),%xmm0 # 5610 <_sk_callback_sse41+0x4fa> DB 65,15,40,211 ; movaps %xmm11,%xmm2 DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,194,194,1 ; cmpltps %xmm2,%xmm0 @@ -13883,7 +13909,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,100,36,32 ; movaps %xmm4,0x20(%rsp) DB 15,41,92,36,16 ; movaps %xmm3,0x10(%rsp) DB 68,15,40,208 ; movaps %xmm0,%xmm10 - DB 68,15,40,13,163,56,0,0 ; movaps 0x38a3(%rip),%xmm9 # 55f0 <_sk_callback_sse41+0x508> + DB 68,15,40,13,211,56,0,0 ; movaps 0x38d3(%rip),%xmm9 # 5620 <_sk_callback_sse41+0x50a> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,194,194,2 ; cmpleps %xmm2,%xmm0 DB 15,40,217 ; movaps %xmm1,%xmm3 @@ -13896,19 +13922,19 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,20,36 ; movaps %xmm2,(%rsp) DB 69,15,88,192 ; addps %xmm8,%xmm8 DB 68,15,92,197 ; subps %xmm5,%xmm8 - DB 68,15,40,53,127,56,0,0 ; movaps 0x387f(%rip),%xmm14 # 5600 <_sk_callback_sse41+0x518> + DB 68,15,40,53,175,56,0,0 ; movaps 0x38af(%rip),%xmm14 # 5630 <_sk_callback_sse41+0x51a> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 102,65,15,58,8,198,1 ; roundps $0x1,%xmm14,%xmm0 DB 68,15,92,240 ; subps %xmm0,%xmm14 - DB 68,15,40,29,120,56,0,0 ; movaps 0x3878(%rip),%xmm11 # 5610 <_sk_callback_sse41+0x528> + DB 68,15,40,29,168,56,0,0 ; movaps 0x38a8(%rip),%xmm11 # 5640 <_sk_callback_sse41+0x52a> DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 15,40,245 ; movaps %xmm5,%xmm6 DB 65,15,92,240 ; subps %xmm8,%xmm6 - DB 15,40,61,113,56,0,0 ; movaps 0x3871(%rip),%xmm7 # 5620 <_sk_callback_sse41+0x538> + DB 15,40,61,161,56,0,0 ; movaps 0x38a1(%rip),%xmm7 # 5650 <_sk_callback_sse41+0x53a> DB 69,15,40,238 ; movaps %xmm14,%xmm13 DB 68,15,89,239 ; mulps %xmm7,%xmm13 - DB 15,40,29,114,56,0,0 ; movaps 0x3872(%rip),%xmm3 # 5630 <_sk_callback_sse41+0x548> + DB 15,40,29,162,56,0,0 ; movaps 0x38a2(%rip),%xmm3 # 5660 <_sk_callback_sse41+0x54a> DB 68,15,40,227 ; movaps %xmm3,%xmm12 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 68,15,89,230 ; mulps %xmm6,%xmm12 @@ -13918,7 +13944,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 68,15,40,253 ; movaps %xmm5,%xmm15 DB 102,69,15,56,20,252 ; blendvps %xmm0,%xmm12,%xmm15 - DB 68,15,40,37,81,56,0,0 ; movaps 0x3851(%rip),%xmm12 # 5640 <_sk_callback_sse41+0x558> + DB 68,15,40,37,129,56,0,0 ; movaps 0x3881(%rip),%xmm12 # 5670 <_sk_callback_sse41+0x55a> DB 65,15,40,196 ; movaps %xmm12,%xmm0 DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 68,15,89,238 ; mulps %xmm6,%xmm13 @@ -13952,7 +13978,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 65,15,40,198 ; movaps %xmm14,%xmm0 DB 15,40,20,36 ; movaps (%rsp),%xmm2 DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 - DB 68,15,88,21,202,55,0,0 ; addps 0x37ca(%rip),%xmm10 # 5650 <_sk_callback_sse41+0x568> + DB 68,15,88,21,250,55,0,0 ; addps 0x37fa(%rip),%xmm10 # 5680 <_sk_callback_sse41+0x56a> DB 102,65,15,58,8,194,1 ; roundps $0x1,%xmm10,%xmm0 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 69,15,194,218,2 ; cmpleps %xmm10,%xmm11 @@ -14002,9 +14028,9 @@ _sk_scale_u8_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,52 ; jne 1f63 <_sk_scale_u8_sse41+0x3e> DB 102,69,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm8 - DB 102,68,15,219,5,33,55,0,0 ; pand 0x3721(%rip),%xmm8 # 5660 <_sk_callback_sse41+0x578> + DB 102,68,15,219,5,81,55,0,0 ; pand 0x3751(%rip),%xmm8 # 5690 <_sk_callback_sse41+0x57a> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,37,55,0,0 ; mulps 0x3725(%rip),%xmm8 # 5670 <_sk_callback_sse41+0x588> + DB 68,15,89,5,85,55,0,0 ; mulps 0x3755(%rip),%xmm8 # 56a0 <_sk_callback_sse41+0x58a> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -14057,9 +14083,9 @@ _sk_lerp_u8_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,72 ; jne 2037 <_sk_lerp_u8_sse41+0x52> DB 102,69,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm8 - DB 102,68,15,219,5,129,54,0,0 ; pand 0x3681(%rip),%xmm8 # 5680 <_sk_callback_sse41+0x598> + DB 102,68,15,219,5,177,54,0,0 ; pand 0x36b1(%rip),%xmm8 # 56b0 <_sk_callback_sse41+0x59a> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,133,54,0,0 ; mulps 0x3685(%rip),%xmm8 # 5690 <_sk_callback_sse41+0x5a8> + DB 68,15,89,5,181,54,0,0 ; mulps 0x36b5(%rip),%xmm8 # 56c0 <_sk_callback_sse41+0x5aa> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -14099,17 +14125,17 @@ _sk_lerp_565_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 15,133,152,0,0,0 ; jne 212a <_sk_lerp_565_sse41+0xa6> DB 102,69,15,56,51,12,82 ; pmovzxwd (%r10,%rdx,2),%xmm9 - DB 102,68,15,111,5,254,53,0,0 ; movdqa 0x35fe(%rip),%xmm8 # 56a0 <_sk_callback_sse41+0x5b8> + DB 102,68,15,111,5,46,54,0,0 ; movdqa 0x362e(%rip),%xmm8 # 56d0 <_sk_callback_sse41+0x5ba> DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,253,53,0,0 ; mulps 0x35fd(%rip),%xmm8 # 56b0 <_sk_callback_sse41+0x5c8> - DB 102,68,15,111,21,4,54,0,0 ; movdqa 0x3604(%rip),%xmm10 # 56c0 <_sk_callback_sse41+0x5d8> + DB 68,15,89,5,45,54,0,0 ; mulps 0x362d(%rip),%xmm8 # 56e0 <_sk_callback_sse41+0x5ca> + DB 102,68,15,111,21,52,54,0,0 ; movdqa 0x3634(%rip),%xmm10 # 56f0 <_sk_callback_sse41+0x5da> DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 - DB 68,15,89,21,3,54,0,0 ; mulps 0x3603(%rip),%xmm10 # 56d0 <_sk_callback_sse41+0x5e8> - DB 102,68,15,219,13,10,54,0,0 ; pand 0x360a(%rip),%xmm9 # 56e0 <_sk_callback_sse41+0x5f8> + DB 68,15,89,21,51,54,0,0 ; mulps 0x3633(%rip),%xmm10 # 5700 <_sk_callback_sse41+0x5ea> + DB 102,68,15,219,13,58,54,0,0 ; pand 0x363a(%rip),%xmm9 # 5710 <_sk_callback_sse41+0x5fa> DB 69,15,91,201 ; cvtdq2ps %xmm9,%xmm9 - DB 68,15,89,13,14,54,0,0 ; mulps 0x360e(%rip),%xmm9 # 56f0 <_sk_callback_sse41+0x608> + DB 68,15,89,13,62,54,0,0 ; mulps 0x363e(%rip),%xmm9 # 5720 <_sk_callback_sse41+0x60a> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -14159,7 +14185,7 @@ _sk_load_tables_sse41 LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx - DB 102,15,111,5,100,53,0,0 ; movdqa 0x3564(%rip),%xmm0 # 5700 <_sk_callback_sse41+0x618> + DB 102,15,111,5,148,53,0,0 ; movdqa 0x3594(%rip),%xmm0 # 5730 <_sk_callback_sse41+0x61a> DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 DB 102,73,15,126,194 ; movq %xmm0,%r10 @@ -14174,7 +14200,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,66,15,58,33,4,179,32 ; insertps $0x20,(%rbx,%r14,4),%xmm0 DB 102,66,15,58,33,4,11,48 ; insertps $0x30,(%rbx,%r9,1),%xmm0 DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 - DB 102,15,56,0,13,31,53,0,0 ; pshufb 0x351f(%rip),%xmm1 # 5710 <_sk_callback_sse41+0x628> + DB 102,15,56,0,13,79,53,0,0 ; pshufb 0x354f(%rip),%xmm1 # 5740 <_sk_callback_sse41+0x62a> DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx DB 68,15,182,211 ; movzbl %bl,%r10d @@ -14189,7 +14215,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 - DB 102,15,56,0,21,219,52,0,0 ; pshufb 0x34db(%rip),%xmm2 # 5720 <_sk_callback_sse41+0x638> + DB 102,15,56,0,21,11,53,0,0 ; pshufb 0x350b(%rip),%xmm2 # 5750 <_sk_callback_sse41+0x63a> DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax DB 68,15,182,208 ; movzbl %al,%r10d @@ -14204,7 +14230,7 @@ _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 15,89,29,152,52,0,0 ; mulps 0x3498(%rip),%xmm3 # 5730 <_sk_callback_sse41+0x648> + DB 15,89,29,200,52,0,0 ; mulps 0x34c8(%rip),%xmm3 # 5760 <_sk_callback_sse41+0x64a> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 @@ -14243,7 +14269,7 @@ _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 102,68,15,111,5,13,52,0,0 ; movdqa 0x340d(%rip),%xmm8 # 5740 <_sk_callback_sse41+0x658> + DB 102,68,15,111,5,61,52,0,0 ; movdqa 0x343d(%rip),%xmm8 # 5770 <_sk_callback_sse41+0x65a> 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 @@ -14261,7 +14287,7 @@ _sk_load_tables_u16_be_sse41 LABEL PROC DB 102,15,58,33,194,32 ; insertps $0x20,%xmm2,%xmm0 DB 243,66,15,16,20,11 ; movss (%rbx,%r9,1),%xmm2 DB 102,15,58,33,194,48 ; insertps $0x30,%xmm2,%xmm0 - DB 102,15,56,0,13,188,51,0,0 ; pshufb 0x33bc(%rip),%xmm1 # 5750 <_sk_callback_sse41+0x668> + DB 102,15,56,0,13,236,51,0,0 ; pshufb 0x33ec(%rip),%xmm1 # 5780 <_sk_callback_sse41+0x66a> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx @@ -14297,7 +14323,7 @@ _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 15,89,29,10,51,0,0 ; mulps 0x330a(%rip),%xmm3 # 5760 <_sk_callback_sse41+0x678> + DB 15,89,29,58,51,0,0 ; mulps 0x333a(%rip),%xmm3 # 5790 <_sk_callback_sse41+0x67a> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 @@ -14337,7 +14363,7 @@ _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 102,68,15,111,5,127,50,0,0 ; movdqa 0x327f(%rip),%xmm8 # 5770 <_sk_callback_sse41+0x688> + DB 102,68,15,111,5,175,50,0,0 ; movdqa 0x32af(%rip),%xmm8 # 57a0 <_sk_callback_sse41+0x68a> 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 @@ -14355,7 +14381,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 102,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm0 DB 243,66,15,16,28,11 ; movss (%rbx,%r9,1),%xmm3 DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0 - DB 102,15,56,0,13,46,50,0,0 ; pshufb 0x322e(%rip),%xmm1 # 5780 <_sk_callback_sse41+0x698> + DB 102,15,56,0,13,94,50,0,0 ; pshufb 0x325e(%rip),%xmm1 # 57b0 <_sk_callback_sse41+0x69a> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 102,72,15,126,203 ; movq %xmm1,%rbx @@ -14386,7 +14412,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 243,65,15,16,28,25 ; movss (%r9,%rbx,1),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,153,49,0,0 ; movaps 0x3199(%rip),%xmm3 # 5790 <_sk_callback_sse41+0x6a8> + DB 15,40,29,201,49,0,0 ; movaps 0x31c9(%rip),%xmm3 # 57c0 <_sk_callback_sse41+0x6aa> DB 91 ; pop %rbx DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 @@ -14418,7 +14444,7 @@ _sk_byte_tables_sse41 LABEL PROC DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,49,49,0,0 ; movaps 0x3131(%rip),%xmm8 # 57a0 <_sk_callback_sse41+0x6b8> + DB 68,15,40,5,97,49,0,0 ; movaps 0x3161(%rip),%xmm8 # 57d0 <_sk_callback_sse41+0x6ba> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 @@ -14437,7 +14463,7 @@ _sk_byte_tables_sse41 LABEL PROC DB 102,15,58,32,195,3 ; pinsrb $0x3,%ebx,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,226,48,0,0 ; movaps 0x30e2(%rip),%xmm9 # 57b0 <_sk_callback_sse41+0x6c8> + DB 68,15,40,13,18,49,0,0 ; movaps 0x3112(%rip),%xmm9 # 57e0 <_sk_callback_sse41+0x6ca> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -14531,7 +14557,7 @@ _sk_byte_tables_rgb_sse41 LABEL PROC DB 102,15,58,32,195,3 ; pinsrb $0x3,%ebx,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,90,47,0,0 ; movaps 0x2f5a(%rip),%xmm9 # 57c0 <_sk_callback_sse41+0x6d8> + DB 68,15,40,13,138,47,0,0 ; movaps 0x2f8a(%rip),%xmm9 # 57f0 <_sk_callback_sse41+0x6da> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -14708,31 +14734,31 @@ _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 68,15,89,5,162,44,0,0 ; mulps 0x2ca2(%rip),%xmm8 # 57d0 <_sk_callback_sse41+0x6e8> - DB 68,15,84,21,170,44,0,0 ; andps 0x2caa(%rip),%xmm10 # 57e0 <_sk_callback_sse41+0x6f8> - DB 68,15,86,21,178,44,0,0 ; orps 0x2cb2(%rip),%xmm10 # 57f0 <_sk_callback_sse41+0x708> - DB 68,15,88,5,186,44,0,0 ; addps 0x2cba(%rip),%xmm8 # 5800 <_sk_callback_sse41+0x718> - DB 68,15,40,37,194,44,0,0 ; movaps 0x2cc2(%rip),%xmm12 # 5810 <_sk_callback_sse41+0x728> + DB 68,15,89,5,210,44,0,0 ; mulps 0x2cd2(%rip),%xmm8 # 5800 <_sk_callback_sse41+0x6ea> + DB 68,15,84,21,218,44,0,0 ; andps 0x2cda(%rip),%xmm10 # 5810 <_sk_callback_sse41+0x6fa> + DB 68,15,86,21,226,44,0,0 ; orps 0x2ce2(%rip),%xmm10 # 5820 <_sk_callback_sse41+0x70a> + DB 68,15,88,5,234,44,0,0 ; addps 0x2cea(%rip),%xmm8 # 5830 <_sk_callback_sse41+0x71a> + DB 68,15,40,37,242,44,0,0 ; movaps 0x2cf2(%rip),%xmm12 # 5840 <_sk_callback_sse41+0x72a> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 68,15,88,21,194,44,0,0 ; addps 0x2cc2(%rip),%xmm10 # 5820 <_sk_callback_sse41+0x738> - DB 68,15,40,37,202,44,0,0 ; movaps 0x2cca(%rip),%xmm12 # 5830 <_sk_callback_sse41+0x748> + DB 68,15,88,21,242,44,0,0 ; addps 0x2cf2(%rip),%xmm10 # 5850 <_sk_callback_sse41+0x73a> + DB 68,15,40,37,250,44,0,0 ; movaps 0x2cfa(%rip),%xmm12 # 5860 <_sk_callback_sse41+0x74a> 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 68,15,88,5,183,44,0,0 ; addps 0x2cb7(%rip),%xmm8 # 5840 <_sk_callback_sse41+0x758> - DB 68,15,40,21,191,44,0,0 ; movaps 0x2cbf(%rip),%xmm10 # 5850 <_sk_callback_sse41+0x768> + DB 68,15,88,5,231,44,0,0 ; addps 0x2ce7(%rip),%xmm8 # 5870 <_sk_callback_sse41+0x75a> + DB 68,15,40,21,239,44,0,0 ; movaps 0x2cef(%rip),%xmm10 # 5880 <_sk_callback_sse41+0x76a> DB 69,15,89,211 ; mulps %xmm11,%xmm10 DB 69,15,92,194 ; subps %xmm10,%xmm8 - DB 68,15,40,21,191,44,0,0 ; movaps 0x2cbf(%rip),%xmm10 # 5860 <_sk_callback_sse41+0x778> + DB 68,15,40,21,239,44,0,0 ; movaps 0x2cef(%rip),%xmm10 # 5890 <_sk_callback_sse41+0x77a> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 68,15,40,29,195,44,0,0 ; movaps 0x2cc3(%rip),%xmm11 # 5870 <_sk_callback_sse41+0x788> + DB 68,15,40,29,243,44,0,0 ; movaps 0x2cf3(%rip),%xmm11 # 58a0 <_sk_callback_sse41+0x78a> DB 69,15,94,218 ; divps %xmm10,%xmm11 DB 69,15,88,216 ; addps %xmm8,%xmm11 - DB 68,15,89,29,195,44,0,0 ; mulps 0x2cc3(%rip),%xmm11 # 5880 <_sk_callback_sse41+0x798> + DB 68,15,89,29,243,44,0,0 ; mulps 0x2cf3(%rip),%xmm11 # 58b0 <_sk_callback_sse41+0x79a> 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 @@ -14740,7 +14766,7 @@ _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 68,15,93,5,170,44,0,0 ; minps 0x2caa(%rip),%xmm8 # 5890 <_sk_callback_sse41+0x7a8> + DB 68,15,93,5,218,44,0,0 ; minps 0x2cda(%rip),%xmm8 # 58c0 <_sk_callback_sse41+0x7aa> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14768,31 +14794,31 @@ _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 68,15,89,37,75,44,0,0 ; mulps 0x2c4b(%rip),%xmm12 # 58a0 <_sk_callback_sse41+0x7b8> - DB 68,15,84,29,83,44,0,0 ; andps 0x2c53(%rip),%xmm11 # 58b0 <_sk_callback_sse41+0x7c8> - DB 68,15,86,29,91,44,0,0 ; orps 0x2c5b(%rip),%xmm11 # 58c0 <_sk_callback_sse41+0x7d8> - DB 68,15,88,37,99,44,0,0 ; addps 0x2c63(%rip),%xmm12 # 58d0 <_sk_callback_sse41+0x7e8> - DB 15,40,13,108,44,0,0 ; movaps 0x2c6c(%rip),%xmm1 # 58e0 <_sk_callback_sse41+0x7f8> + DB 68,15,89,37,123,44,0,0 ; mulps 0x2c7b(%rip),%xmm12 # 58d0 <_sk_callback_sse41+0x7ba> + DB 68,15,84,29,131,44,0,0 ; andps 0x2c83(%rip),%xmm11 # 58e0 <_sk_callback_sse41+0x7ca> + DB 68,15,86,29,139,44,0,0 ; orps 0x2c8b(%rip),%xmm11 # 58f0 <_sk_callback_sse41+0x7da> + DB 68,15,88,37,147,44,0,0 ; addps 0x2c93(%rip),%xmm12 # 5900 <_sk_callback_sse41+0x7ea> + DB 15,40,13,156,44,0,0 ; movaps 0x2c9c(%rip),%xmm1 # 5910 <_sk_callback_sse41+0x7fa> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,88,29,108,44,0,0 ; addps 0x2c6c(%rip),%xmm11 # 58f0 <_sk_callback_sse41+0x808> - DB 15,40,13,117,44,0,0 ; movaps 0x2c75(%rip),%xmm1 # 5900 <_sk_callback_sse41+0x818> + DB 68,15,88,29,156,44,0,0 ; addps 0x2c9c(%rip),%xmm11 # 5920 <_sk_callback_sse41+0x80a> + DB 15,40,13,165,44,0,0 ; movaps 0x2ca5(%rip),%xmm1 # 5930 <_sk_callback_sse41+0x81a> 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 68,15,88,37,98,44,0,0 ; addps 0x2c62(%rip),%xmm12 # 5910 <_sk_callback_sse41+0x828> - DB 15,40,13,107,44,0,0 ; movaps 0x2c6b(%rip),%xmm1 # 5920 <_sk_callback_sse41+0x838> + DB 68,15,88,37,146,44,0,0 ; addps 0x2c92(%rip),%xmm12 # 5940 <_sk_callback_sse41+0x82a> + DB 15,40,13,155,44,0,0 ; movaps 0x2c9b(%rip),%xmm1 # 5950 <_sk_callback_sse41+0x83a> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,40,21,107,44,0,0 ; movaps 0x2c6b(%rip),%xmm10 # 5930 <_sk_callback_sse41+0x848> + DB 68,15,40,21,155,44,0,0 ; movaps 0x2c9b(%rip),%xmm10 # 5960 <_sk_callback_sse41+0x84a> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,13,112,44,0,0 ; movaps 0x2c70(%rip),%xmm1 # 5940 <_sk_callback_sse41+0x858> + DB 15,40,13,160,44,0,0 ; movaps 0x2ca0(%rip),%xmm1 # 5970 <_sk_callback_sse41+0x85a> DB 65,15,94,202 ; divps %xmm10,%xmm1 DB 65,15,88,204 ; addps %xmm12,%xmm1 - DB 15,89,13,113,44,0,0 ; mulps 0x2c71(%rip),%xmm1 # 5950 <_sk_callback_sse41+0x868> + DB 15,89,13,161,44,0,0 ; mulps 0x2ca1(%rip),%xmm1 # 5980 <_sk_callback_sse41+0x86a> 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 @@ -14800,7 +14826,7 @@ _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 15,93,13,92,44,0,0 ; minps 0x2c5c(%rip),%xmm1 # 5960 <_sk_callback_sse41+0x878> + DB 15,93,13,140,44,0,0 ; minps 0x2c8c(%rip),%xmm1 # 5990 <_sk_callback_sse41+0x87a> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14828,31 +14854,31 @@ _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 68,15,89,37,253,43,0,0 ; mulps 0x2bfd(%rip),%xmm12 # 5970 <_sk_callback_sse41+0x888> - DB 68,15,84,29,5,44,0,0 ; andps 0x2c05(%rip),%xmm11 # 5980 <_sk_callback_sse41+0x898> - DB 68,15,86,29,13,44,0,0 ; orps 0x2c0d(%rip),%xmm11 # 5990 <_sk_callback_sse41+0x8a8> - DB 68,15,88,37,21,44,0,0 ; addps 0x2c15(%rip),%xmm12 # 59a0 <_sk_callback_sse41+0x8b8> - DB 15,40,21,30,44,0,0 ; movaps 0x2c1e(%rip),%xmm2 # 59b0 <_sk_callback_sse41+0x8c8> + DB 68,15,89,37,45,44,0,0 ; mulps 0x2c2d(%rip),%xmm12 # 59a0 <_sk_callback_sse41+0x88a> + DB 68,15,84,29,53,44,0,0 ; andps 0x2c35(%rip),%xmm11 # 59b0 <_sk_callback_sse41+0x89a> + DB 68,15,86,29,61,44,0,0 ; orps 0x2c3d(%rip),%xmm11 # 59c0 <_sk_callback_sse41+0x8aa> + DB 68,15,88,37,69,44,0,0 ; addps 0x2c45(%rip),%xmm12 # 59d0 <_sk_callback_sse41+0x8ba> + DB 15,40,21,78,44,0,0 ; movaps 0x2c4e(%rip),%xmm2 # 59e0 <_sk_callback_sse41+0x8ca> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,88,29,30,44,0,0 ; addps 0x2c1e(%rip),%xmm11 # 59c0 <_sk_callback_sse41+0x8d8> - DB 15,40,21,39,44,0,0 ; movaps 0x2c27(%rip),%xmm2 # 59d0 <_sk_callback_sse41+0x8e8> + DB 68,15,88,29,78,44,0,0 ; addps 0x2c4e(%rip),%xmm11 # 59f0 <_sk_callback_sse41+0x8da> + DB 15,40,21,87,44,0,0 ; movaps 0x2c57(%rip),%xmm2 # 5a00 <_sk_callback_sse41+0x8ea> 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 68,15,88,37,20,44,0,0 ; addps 0x2c14(%rip),%xmm12 # 59e0 <_sk_callback_sse41+0x8f8> - DB 15,40,21,29,44,0,0 ; movaps 0x2c1d(%rip),%xmm2 # 59f0 <_sk_callback_sse41+0x908> + DB 68,15,88,37,68,44,0,0 ; addps 0x2c44(%rip),%xmm12 # 5a10 <_sk_callback_sse41+0x8fa> + DB 15,40,21,77,44,0,0 ; movaps 0x2c4d(%rip),%xmm2 # 5a20 <_sk_callback_sse41+0x90a> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,40,21,29,44,0,0 ; movaps 0x2c1d(%rip),%xmm10 # 5a00 <_sk_callback_sse41+0x918> + DB 68,15,40,21,77,44,0,0 ; movaps 0x2c4d(%rip),%xmm10 # 5a30 <_sk_callback_sse41+0x91a> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,21,34,44,0,0 ; movaps 0x2c22(%rip),%xmm2 # 5a10 <_sk_callback_sse41+0x928> + DB 15,40,21,82,44,0,0 ; movaps 0x2c52(%rip),%xmm2 # 5a40 <_sk_callback_sse41+0x92a> DB 65,15,94,210 ; divps %xmm10,%xmm2 DB 65,15,88,212 ; addps %xmm12,%xmm2 - DB 15,89,21,35,44,0,0 ; mulps 0x2c23(%rip),%xmm2 # 5a20 <_sk_callback_sse41+0x938> + DB 15,89,21,83,44,0,0 ; mulps 0x2c53(%rip),%xmm2 # 5a50 <_sk_callback_sse41+0x93a> 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 @@ -14860,7 +14886,7 @@ _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 15,93,21,14,44,0,0 ; minps 0x2c0e(%rip),%xmm2 # 5a30 <_sk_callback_sse41+0x948> + DB 15,93,21,62,44,0,0 ; minps 0x2c3e(%rip),%xmm2 # 5a60 <_sk_callback_sse41+0x94a> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14888,31 +14914,31 @@ _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 68,15,89,37,175,43,0,0 ; mulps 0x2baf(%rip),%xmm12 # 5a40 <_sk_callback_sse41+0x958> - DB 68,15,84,29,183,43,0,0 ; andps 0x2bb7(%rip),%xmm11 # 5a50 <_sk_callback_sse41+0x968> - DB 68,15,86,29,191,43,0,0 ; orps 0x2bbf(%rip),%xmm11 # 5a60 <_sk_callback_sse41+0x978> - DB 68,15,88,37,199,43,0,0 ; addps 0x2bc7(%rip),%xmm12 # 5a70 <_sk_callback_sse41+0x988> - DB 15,40,29,208,43,0,0 ; movaps 0x2bd0(%rip),%xmm3 # 5a80 <_sk_callback_sse41+0x998> + DB 68,15,89,37,223,43,0,0 ; mulps 0x2bdf(%rip),%xmm12 # 5a70 <_sk_callback_sse41+0x95a> + DB 68,15,84,29,231,43,0,0 ; andps 0x2be7(%rip),%xmm11 # 5a80 <_sk_callback_sse41+0x96a> + DB 68,15,86,29,239,43,0,0 ; orps 0x2bef(%rip),%xmm11 # 5a90 <_sk_callback_sse41+0x97a> + DB 68,15,88,37,247,43,0,0 ; addps 0x2bf7(%rip),%xmm12 # 5aa0 <_sk_callback_sse41+0x98a> + DB 15,40,29,0,44,0,0 ; movaps 0x2c00(%rip),%xmm3 # 5ab0 <_sk_callback_sse41+0x99a> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,88,29,208,43,0,0 ; addps 0x2bd0(%rip),%xmm11 # 5a90 <_sk_callback_sse41+0x9a8> - DB 15,40,29,217,43,0,0 ; movaps 0x2bd9(%rip),%xmm3 # 5aa0 <_sk_callback_sse41+0x9b8> + DB 68,15,88,29,0,44,0,0 ; addps 0x2c00(%rip),%xmm11 # 5ac0 <_sk_callback_sse41+0x9aa> + DB 15,40,29,9,44,0,0 ; movaps 0x2c09(%rip),%xmm3 # 5ad0 <_sk_callback_sse41+0x9ba> 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 68,15,88,37,198,43,0,0 ; addps 0x2bc6(%rip),%xmm12 # 5ab0 <_sk_callback_sse41+0x9c8> - DB 15,40,29,207,43,0,0 ; movaps 0x2bcf(%rip),%xmm3 # 5ac0 <_sk_callback_sse41+0x9d8> + DB 68,15,88,37,246,43,0,0 ; addps 0x2bf6(%rip),%xmm12 # 5ae0 <_sk_callback_sse41+0x9ca> + DB 15,40,29,255,43,0,0 ; movaps 0x2bff(%rip),%xmm3 # 5af0 <_sk_callback_sse41+0x9da> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,40,21,207,43,0,0 ; movaps 0x2bcf(%rip),%xmm10 # 5ad0 <_sk_callback_sse41+0x9e8> + DB 68,15,40,21,255,43,0,0 ; movaps 0x2bff(%rip),%xmm10 # 5b00 <_sk_callback_sse41+0x9ea> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,29,212,43,0,0 ; movaps 0x2bd4(%rip),%xmm3 # 5ae0 <_sk_callback_sse41+0x9f8> + DB 15,40,29,4,44,0,0 ; movaps 0x2c04(%rip),%xmm3 # 5b10 <_sk_callback_sse41+0x9fa> DB 65,15,94,218 ; divps %xmm10,%xmm3 DB 65,15,88,220 ; addps %xmm12,%xmm3 - DB 15,89,29,213,43,0,0 ; mulps 0x2bd5(%rip),%xmm3 # 5af0 <_sk_callback_sse41+0xa08> + DB 15,89,29,5,44,0,0 ; mulps 0x2c05(%rip),%xmm3 # 5b20 <_sk_callback_sse41+0xa0a> 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 @@ -14920,7 +14946,7 @@ _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 15,93,29,192,43,0,0 ; minps 0x2bc0(%rip),%xmm3 # 5b00 <_sk_callback_sse41+0xa18> + DB 15,93,29,240,43,0,0 ; minps 0x2bf0(%rip),%xmm3 # 5b30 <_sk_callback_sse41+0xa1a> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -14928,29 +14954,29 @@ _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 68,15,89,5,188,43,0,0 ; mulps 0x2bbc(%rip),%xmm8 # 5b10 <_sk_callback_sse41+0xa28> - DB 68,15,40,13,196,43,0,0 ; movaps 0x2bc4(%rip),%xmm9 # 5b20 <_sk_callback_sse41+0xa38> + DB 68,15,89,5,236,43,0,0 ; mulps 0x2bec(%rip),%xmm8 # 5b40 <_sk_callback_sse41+0xa2a> + DB 68,15,40,13,244,43,0,0 ; movaps 0x2bf4(%rip),%xmm9 # 5b50 <_sk_callback_sse41+0xa3a> DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 15,40,5,201,43,0,0 ; movaps 0x2bc9(%rip),%xmm0 # 5b30 <_sk_callback_sse41+0xa48> + DB 15,40,5,249,43,0,0 ; movaps 0x2bf9(%rip),%xmm0 # 5b60 <_sk_callback_sse41+0xa4a> DB 15,88,200 ; addps %xmm0,%xmm1 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 68,15,88,5,199,43,0,0 ; addps 0x2bc7(%rip),%xmm8 # 5b40 <_sk_callback_sse41+0xa58> - DB 68,15,89,5,207,43,0,0 ; mulps 0x2bcf(%rip),%xmm8 # 5b50 <_sk_callback_sse41+0xa68> - DB 15,89,13,216,43,0,0 ; mulps 0x2bd8(%rip),%xmm1 # 5b60 <_sk_callback_sse41+0xa78> + DB 68,15,88,5,247,43,0,0 ; addps 0x2bf7(%rip),%xmm8 # 5b70 <_sk_callback_sse41+0xa5a> + DB 68,15,89,5,255,43,0,0 ; mulps 0x2bff(%rip),%xmm8 # 5b80 <_sk_callback_sse41+0xa6a> + DB 15,89,13,8,44,0,0 ; mulps 0x2c08(%rip),%xmm1 # 5b90 <_sk_callback_sse41+0xa7a> DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 15,89,21,221,43,0,0 ; mulps 0x2bdd(%rip),%xmm2 # 5b70 <_sk_callback_sse41+0xa88> + DB 15,89,21,13,44,0,0 ; mulps 0x2c0d(%rip),%xmm2 # 5ba0 <_sk_callback_sse41+0xa8a> 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 68,15,40,13,209,43,0,0 ; movaps 0x2bd1(%rip),%xmm9 # 5b80 <_sk_callback_sse41+0xa98> + DB 68,15,40,13,1,44,0,0 ; movaps 0x2c01(%rip),%xmm9 # 5bb0 <_sk_callback_sse41+0xa9a> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 15,40,21,209,43,0,0 ; movaps 0x2bd1(%rip),%xmm2 # 5b90 <_sk_callback_sse41+0xaa8> + DB 15,40,21,1,44,0,0 ; movaps 0x2c01(%rip),%xmm2 # 5bc0 <_sk_callback_sse41+0xaaa> DB 15,88,202 ; addps %xmm2,%xmm1 - DB 68,15,40,37,214,43,0,0 ; movaps 0x2bd6(%rip),%xmm12 # 5ba0 <_sk_callback_sse41+0xab8> + DB 68,15,40,37,6,44,0,0 ; movaps 0x2c06(%rip),%xmm12 # 5bd0 <_sk_callback_sse41+0xaba> 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 @@ -14969,8 +14995,8 @@ _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 15,89,13,143,43,0,0 ; mulps 0x2b8f(%rip),%xmm1 # 5bb0 <_sk_callback_sse41+0xac8> - DB 15,89,21,152,43,0,0 ; mulps 0x2b98(%rip),%xmm2 # 5bc0 <_sk_callback_sse41+0xad8> + DB 15,89,13,191,43,0,0 ; mulps 0x2bbf(%rip),%xmm1 # 5be0 <_sk_callback_sse41+0xaca> + DB 15,89,21,200,43,0,0 ; mulps 0x2bc8(%rip),%xmm2 # 5bf0 <_sk_callback_sse41+0xada> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,40,200 ; movaps %xmm8,%xmm1 @@ -14983,9 +15009,9 @@ _sk_load_a8_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,38 ; jne 3063 <_sk_load_a8_sse41+0x30> DB 102,65,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm0 - DB 102,15,219,5,132,43,0,0 ; pand 0x2b84(%rip),%xmm0 # 5bd0 <_sk_callback_sse41+0xae8> + DB 102,15,219,5,180,43,0,0 ; pand 0x2bb4(%rip),%xmm0 # 5c00 <_sk_callback_sse41+0xaea> DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,138,43,0,0 ; mulps 0x2b8a(%rip),%xmm3 # 5be0 <_sk_callback_sse41+0xaf8> + DB 15,89,29,186,43,0,0 ; mulps 0x2bba(%rip),%xmm3 # 5c10 <_sk_callback_sse41+0xafa> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 @@ -15034,7 +15060,7 @@ _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 15,89,29,216,42,0,0 ; mulps 0x2ad8(%rip),%xmm3 # 5bf0 <_sk_callback_sse41+0xb08> + DB 15,89,29,8,43,0,0 ; mulps 0x2b08(%rip),%xmm3 # 5c20 <_sk_callback_sse41+0xb0a> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -15047,7 +15073,7 @@ _sk_store_a8_sse41 LABEL PROC DB 72,131,236,4 ; sub $0x4,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 68,15,40,5,199,42,0,0 ; movaps 0x2ac7(%rip),%xmm8 # 5c00 <_sk_callback_sse41+0xb18> + DB 68,15,40,5,247,42,0,0 ; movaps 0x2af7(%rip),%xmm8 # 5c30 <_sk_callback_sse41+0xb1a> 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 @@ -15080,11 +15106,11 @@ _sk_load_g8_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,42 ; jne 31d2 <_sk_load_g8_sse41+0x34> DB 102,65,15,56,49,4,18 ; pmovzxbd (%r10,%rdx,1),%xmm0 - DB 102,15,219,5,89,42,0,0 ; pand 0x2a59(%rip),%xmm0 # 5c10 <_sk_callback_sse41+0xb28> + DB 102,15,219,5,137,42,0,0 ; pand 0x2a89(%rip),%xmm0 # 5c40 <_sk_callback_sse41+0xb2a> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,95,42,0,0 ; mulps 0x2a5f(%rip),%xmm0 # 5c20 <_sk_callback_sse41+0xb38> + DB 15,89,5,143,42,0,0 ; mulps 0x2a8f(%rip),%xmm0 # 5c50 <_sk_callback_sse41+0xb3a> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,102,42,0,0 ; movaps 0x2a66(%rip),%xmm3 # 5c30 <_sk_callback_sse41+0xb48> + DB 15,40,29,150,42,0,0 ; movaps 0x2a96(%rip),%xmm3 # 5c60 <_sk_callback_sse41+0xb4a> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -15131,9 +15157,9 @@ _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 15,89,5,185,41,0,0 ; mulps 0x29b9(%rip),%xmm0 # 5c40 <_sk_callback_sse41+0xb58> + DB 15,89,5,233,41,0,0 ; mulps 0x29e9(%rip),%xmm0 # 5c70 <_sk_callback_sse41+0xb5a> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,192,41,0,0 ; movaps 0x29c0(%rip),%xmm3 # 5c50 <_sk_callback_sse41+0xb68> + DB 15,40,29,240,41,0,0 ; movaps 0x29f0(%rip),%xmm3 # 5c80 <_sk_callback_sse41+0xb6a> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 91 ; pop %rbx @@ -15179,17 +15205,17 @@ _sk_gather_i8_sse41 LABEL PROC DB 102,15,58,34,28,24,1 ; pinsrd $0x1,(%rax,%rbx,1),%xmm3 DB 102,66,15,58,34,28,152,2 ; pinsrd $0x2,(%rax,%r11,4),%xmm3 DB 102,66,15,58,34,28,16,3 ; pinsrd $0x3,(%rax,%r10,1),%xmm3 - DB 102,15,111,5,19,41,0,0 ; movdqa 0x2913(%rip),%xmm0 # 5c60 <_sk_callback_sse41+0xb78> + DB 102,15,111,5,67,41,0,0 ; movdqa 0x2943(%rip),%xmm0 # 5c90 <_sk_callback_sse41+0xb7a> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,20,41,0,0 ; movaps 0x2914(%rip),%xmm8 # 5c70 <_sk_callback_sse41+0xb88> + DB 68,15,40,5,68,41,0,0 ; movaps 0x2944(%rip),%xmm8 # 5ca0 <_sk_callback_sse41+0xb8a> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,19,41,0,0 ; pshufb 0x2913(%rip),%xmm1 # 5c80 <_sk_callback_sse41+0xb98> + DB 102,15,56,0,13,67,41,0,0 ; pshufb 0x2943(%rip),%xmm1 # 5cb0 <_sk_callback_sse41+0xb9a> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,15,41,0,0 ; pshufb 0x290f(%rip),%xmm2 # 5c90 <_sk_callback_sse41+0xba8> + DB 102,15,56,0,21,63,41,0,0 ; pshufb 0x293f(%rip),%xmm2 # 5cc0 <_sk_callback_sse41+0xbaa> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -15207,19 +15233,19 @@ _sk_load_565_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,80 ; jne 33f5 <_sk_load_565_sse41+0x5a> DB 102,65,15,56,51,20,82 ; pmovzxwd (%r10,%rdx,2),%xmm2 - DB 102,15,111,5,236,40,0,0 ; movdqa 0x28ec(%rip),%xmm0 # 5ca0 <_sk_callback_sse41+0xbb8> + DB 102,15,111,5,28,41,0,0 ; movdqa 0x291c(%rip),%xmm0 # 5cd0 <_sk_callback_sse41+0xbba> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,238,40,0,0 ; mulps 0x28ee(%rip),%xmm0 # 5cb0 <_sk_callback_sse41+0xbc8> - DB 102,15,111,13,246,40,0,0 ; movdqa 0x28f6(%rip),%xmm1 # 5cc0 <_sk_callback_sse41+0xbd8> + DB 15,89,5,30,41,0,0 ; mulps 0x291e(%rip),%xmm0 # 5ce0 <_sk_callback_sse41+0xbca> + DB 102,15,111,13,38,41,0,0 ; movdqa 0x2926(%rip),%xmm1 # 5cf0 <_sk_callback_sse41+0xbda> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,248,40,0,0 ; mulps 0x28f8(%rip),%xmm1 # 5cd0 <_sk_callback_sse41+0xbe8> - DB 102,15,219,21,0,41,0,0 ; pand 0x2900(%rip),%xmm2 # 5ce0 <_sk_callback_sse41+0xbf8> + DB 15,89,13,40,41,0,0 ; mulps 0x2928(%rip),%xmm1 # 5d00 <_sk_callback_sse41+0xbea> + DB 102,15,219,21,48,41,0,0 ; pand 0x2930(%rip),%xmm2 # 5d10 <_sk_callback_sse41+0xbfa> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,6,41,0,0 ; mulps 0x2906(%rip),%xmm2 # 5cf0 <_sk_callback_sse41+0xc08> + DB 15,89,21,54,41,0,0 ; mulps 0x2936(%rip),%xmm2 # 5d20 <_sk_callback_sse41+0xc0a> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,13,41,0,0 ; movaps 0x290d(%rip),%xmm3 # 5d00 <_sk_callback_sse41+0xc18> + DB 15,40,29,61,41,0,0 ; movaps 0x293d(%rip),%xmm3 # 5d30 <_sk_callback_sse41+0xc1a> DB 255,224 ; jmpq *%rax DB 69,137,193 ; mov %r8d,%r9d DB 65,128,225,3 ; and $0x3,%r9b @@ -15263,19 +15289,19 @@ _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 102,15,111,5,105,40,0,0 ; movdqa 0x2869(%rip),%xmm0 # 5d10 <_sk_callback_sse41+0xc28> + DB 102,15,111,5,153,40,0,0 ; movdqa 0x2899(%rip),%xmm0 # 5d40 <_sk_callback_sse41+0xc2a> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,107,40,0,0 ; mulps 0x286b(%rip),%xmm0 # 5d20 <_sk_callback_sse41+0xc38> - DB 102,15,111,13,115,40,0,0 ; movdqa 0x2873(%rip),%xmm1 # 5d30 <_sk_callback_sse41+0xc48> + DB 15,89,5,155,40,0,0 ; mulps 0x289b(%rip),%xmm0 # 5d50 <_sk_callback_sse41+0xc3a> + DB 102,15,111,13,163,40,0,0 ; movdqa 0x28a3(%rip),%xmm1 # 5d60 <_sk_callback_sse41+0xc4a> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,117,40,0,0 ; mulps 0x2875(%rip),%xmm1 # 5d40 <_sk_callback_sse41+0xc58> - DB 102,15,219,21,125,40,0,0 ; pand 0x287d(%rip),%xmm2 # 5d50 <_sk_callback_sse41+0xc68> + DB 15,89,13,165,40,0,0 ; mulps 0x28a5(%rip),%xmm1 # 5d70 <_sk_callback_sse41+0xc5a> + DB 102,15,219,21,173,40,0,0 ; pand 0x28ad(%rip),%xmm2 # 5d80 <_sk_callback_sse41+0xc6a> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,131,40,0,0 ; mulps 0x2883(%rip),%xmm2 # 5d60 <_sk_callback_sse41+0xc78> + DB 15,89,21,179,40,0,0 ; mulps 0x28b3(%rip),%xmm2 # 5d90 <_sk_callback_sse41+0xc7a> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,138,40,0,0 ; movaps 0x288a(%rip),%xmm3 # 5d70 <_sk_callback_sse41+0xc88> + DB 15,40,29,186,40,0,0 ; movaps 0x28ba(%rip),%xmm3 # 5da0 <_sk_callback_sse41+0xc8a> DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax @@ -15283,12 +15309,12 @@ 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 68,15,40,5,138,40,0,0 ; movaps 0x288a(%rip),%xmm8 # 5d80 <_sk_callback_sse41+0xc98> + DB 68,15,40,5,186,40,0,0 ; movaps 0x28ba(%rip),%xmm8 # 5db0 <_sk_callback_sse41+0xc9a> 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 68,15,40,21,127,40,0,0 ; movaps 0x287f(%rip),%xmm10 # 5d90 <_sk_callback_sse41+0xca8> + DB 68,15,40,21,175,40,0,0 ; movaps 0x28af(%rip),%xmm10 # 5dc0 <_sk_callback_sse41+0xcaa> 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 @@ -15323,21 +15349,21 @@ _sk_load_4444_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,95 ; jne 35ec <_sk_load_4444_sse41+0x69> DB 102,65,15,56,51,28,82 ; pmovzxwd (%r10,%rdx,2),%xmm3 - DB 102,15,111,5,4,40,0,0 ; movdqa 0x2804(%rip),%xmm0 # 5da0 <_sk_callback_sse41+0xcb8> + DB 102,15,111,5,52,40,0,0 ; movdqa 0x2834(%rip),%xmm0 # 5dd0 <_sk_callback_sse41+0xcba> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,6,40,0,0 ; mulps 0x2806(%rip),%xmm0 # 5db0 <_sk_callback_sse41+0xcc8> - DB 102,15,111,13,14,40,0,0 ; movdqa 0x280e(%rip),%xmm1 # 5dc0 <_sk_callback_sse41+0xcd8> + DB 15,89,5,54,40,0,0 ; mulps 0x2836(%rip),%xmm0 # 5de0 <_sk_callback_sse41+0xcca> + DB 102,15,111,13,62,40,0,0 ; movdqa 0x283e(%rip),%xmm1 # 5df0 <_sk_callback_sse41+0xcda> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,16,40,0,0 ; mulps 0x2810(%rip),%xmm1 # 5dd0 <_sk_callback_sse41+0xce8> - DB 102,15,111,21,24,40,0,0 ; movdqa 0x2818(%rip),%xmm2 # 5de0 <_sk_callback_sse41+0xcf8> + DB 15,89,13,64,40,0,0 ; mulps 0x2840(%rip),%xmm1 # 5e00 <_sk_callback_sse41+0xcea> + DB 102,15,111,21,72,40,0,0 ; movdqa 0x2848(%rip),%xmm2 # 5e10 <_sk_callback_sse41+0xcfa> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,26,40,0,0 ; mulps 0x281a(%rip),%xmm2 # 5df0 <_sk_callback_sse41+0xd08> - DB 102,15,219,29,34,40,0,0 ; pand 0x2822(%rip),%xmm3 # 5e00 <_sk_callback_sse41+0xd18> + DB 15,89,21,74,40,0,0 ; mulps 0x284a(%rip),%xmm2 # 5e20 <_sk_callback_sse41+0xd0a> + DB 102,15,219,29,82,40,0,0 ; pand 0x2852(%rip),%xmm3 # 5e30 <_sk_callback_sse41+0xd1a> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,40,40,0,0 ; mulps 0x2828(%rip),%xmm3 # 5e10 <_sk_callback_sse41+0xd28> + DB 15,89,29,88,40,0,0 ; mulps 0x2858(%rip),%xmm3 # 5e40 <_sk_callback_sse41+0xd2a> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 69,137,193 ; mov %r8d,%r9d @@ -15382,21 +15408,21 @@ _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 102,15,111,5,130,39,0,0 ; movdqa 0x2782(%rip),%xmm0 # 5e20 <_sk_callback_sse41+0xd38> + DB 102,15,111,5,178,39,0,0 ; movdqa 0x27b2(%rip),%xmm0 # 5e50 <_sk_callback_sse41+0xd3a> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,132,39,0,0 ; mulps 0x2784(%rip),%xmm0 # 5e30 <_sk_callback_sse41+0xd48> - DB 102,15,111,13,140,39,0,0 ; movdqa 0x278c(%rip),%xmm1 # 5e40 <_sk_callback_sse41+0xd58> + DB 15,89,5,180,39,0,0 ; mulps 0x27b4(%rip),%xmm0 # 5e60 <_sk_callback_sse41+0xd4a> + DB 102,15,111,13,188,39,0,0 ; movdqa 0x27bc(%rip),%xmm1 # 5e70 <_sk_callback_sse41+0xd5a> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,142,39,0,0 ; mulps 0x278e(%rip),%xmm1 # 5e50 <_sk_callback_sse41+0xd68> - DB 102,15,111,21,150,39,0,0 ; movdqa 0x2796(%rip),%xmm2 # 5e60 <_sk_callback_sse41+0xd78> + DB 15,89,13,190,39,0,0 ; mulps 0x27be(%rip),%xmm1 # 5e80 <_sk_callback_sse41+0xd6a> + DB 102,15,111,21,198,39,0,0 ; movdqa 0x27c6(%rip),%xmm2 # 5e90 <_sk_callback_sse41+0xd7a> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,152,39,0,0 ; mulps 0x2798(%rip),%xmm2 # 5e70 <_sk_callback_sse41+0xd88> - DB 102,15,219,29,160,39,0,0 ; pand 0x27a0(%rip),%xmm3 # 5e80 <_sk_callback_sse41+0xd98> + DB 15,89,21,200,39,0,0 ; mulps 0x27c8(%rip),%xmm2 # 5ea0 <_sk_callback_sse41+0xd8a> + DB 102,15,219,29,208,39,0,0 ; pand 0x27d0(%rip),%xmm3 # 5eb0 <_sk_callback_sse41+0xd9a> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,166,39,0,0 ; mulps 0x27a6(%rip),%xmm3 # 5e90 <_sk_callback_sse41+0xda8> + DB 15,89,29,214,39,0,0 ; mulps 0x27d6(%rip),%xmm3 # 5ec0 <_sk_callback_sse41+0xdaa> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax @@ -15405,7 +15431,7 @@ 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 68,15,40,5,164,39,0,0 ; movaps 0x27a4(%rip),%xmm8 # 5ea0 <_sk_callback_sse41+0xdb8> + DB 68,15,40,5,212,39,0,0 ; movaps 0x27d4(%rip),%xmm8 # 5ed0 <_sk_callback_sse41+0xdba> 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 @@ -15450,17 +15476,17 @@ _sk_load_8888_sse41 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,88 ; jne 37ff <_sk_load_8888_sse41+0x62> DB 243,15,111,28,144 ; movdqu (%rax,%rdx,4),%xmm3 - DB 102,15,111,5,252,38,0,0 ; movdqa 0x26fc(%rip),%xmm0 # 5eb0 <_sk_callback_sse41+0xdc8> + DB 102,15,111,5,44,39,0,0 ; movdqa 0x272c(%rip),%xmm0 # 5ee0 <_sk_callback_sse41+0xdca> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,253,38,0,0 ; movaps 0x26fd(%rip),%xmm8 # 5ec0 <_sk_callback_sse41+0xdd8> + DB 68,15,40,5,45,39,0,0 ; movaps 0x272d(%rip),%xmm8 # 5ef0 <_sk_callback_sse41+0xdda> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,252,38,0,0 ; pshufb 0x26fc(%rip),%xmm1 # 5ed0 <_sk_callback_sse41+0xde8> + DB 102,15,56,0,13,44,39,0,0 ; pshufb 0x272c(%rip),%xmm1 # 5f00 <_sk_callback_sse41+0xdea> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,248,38,0,0 ; pshufb 0x26f8(%rip),%xmm2 # 5ee0 <_sk_callback_sse41+0xdf8> + DB 102,15,56,0,21,40,39,0,0 ; pshufb 0x2728(%rip),%xmm2 # 5f10 <_sk_callback_sse41+0xdfa> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -15504,17 +15530,17 @@ _sk_gather_8888_sse41 LABEL PROC DB 102,65,15,58,34,28,129,1 ; pinsrd $0x1,(%r9,%rax,4),%xmm3 DB 102,65,15,58,34,28,153,2 ; pinsrd $0x2,(%r9,%rbx,4),%xmm3 DB 102,67,15,58,34,28,153,3 ; pinsrd $0x3,(%r9,%r11,4),%xmm3 - DB 102,15,111,5,84,38,0,0 ; movdqa 0x2654(%rip),%xmm0 # 5ef0 <_sk_callback_sse41+0xe08> + DB 102,15,111,5,132,38,0,0 ; movdqa 0x2684(%rip),%xmm0 # 5f20 <_sk_callback_sse41+0xe0a> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,85,38,0,0 ; movaps 0x2655(%rip),%xmm8 # 5f00 <_sk_callback_sse41+0xe18> + DB 68,15,40,5,133,38,0,0 ; movaps 0x2685(%rip),%xmm8 # 5f30 <_sk_callback_sse41+0xe1a> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,84,38,0,0 ; pshufb 0x2654(%rip),%xmm1 # 5f10 <_sk_callback_sse41+0xe28> + DB 102,15,56,0,13,132,38,0,0 ; pshufb 0x2684(%rip),%xmm1 # 5f40 <_sk_callback_sse41+0xe2a> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,80,38,0,0 ; pshufb 0x2650(%rip),%xmm2 # 5f20 <_sk_callback_sse41+0xe38> + DB 102,15,56,0,21,128,38,0,0 ; pshufb 0x2680(%rip),%xmm2 # 5f50 <_sk_callback_sse41+0xe3a> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -15528,7 +15554,7 @@ 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 68,15,40,5,59,38,0,0 ; movaps 0x263b(%rip),%xmm8 # 5f30 <_sk_callback_sse41+0xe48> + DB 68,15,40,5,107,38,0,0 ; movaps 0x266b(%rip),%xmm8 # 5f60 <_sk_callback_sse41+0xe4a> 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 @@ -15579,18 +15605,18 @@ _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 102,68,15,111,5,115,37,0,0 ; movdqa 0x2573(%rip),%xmm8 # 5f40 <_sk_callback_sse41+0xe58> + DB 102,68,15,111,5,163,37,0,0 ; movdqa 0x25a3(%rip),%xmm8 # 5f70 <_sk_callback_sse41+0xe5a> 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 102,15,111,29,110,37,0,0 ; movdqa 0x256e(%rip),%xmm3 # 5f50 <_sk_callback_sse41+0xe68> + DB 102,15,111,29,158,37,0,0 ; movdqa 0x259e(%rip),%xmm3 # 5f80 <_sk_callback_sse41+0xe6a> 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 102,68,15,111,21,90,37,0,0 ; movdqa 0x255a(%rip),%xmm10 # 5f60 <_sk_callback_sse41+0xe78> + DB 102,68,15,111,21,138,37,0,0 ; movdqa 0x258a(%rip),%xmm10 # 5f90 <_sk_callback_sse41+0xe7a> 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 @@ -15674,18 +15700,18 @@ _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 102,68,15,111,5,224,35,0,0 ; movdqa 0x23e0(%rip),%xmm8 # 5f70 <_sk_callback_sse41+0xe88> + DB 102,68,15,111,5,16,36,0,0 ; movdqa 0x2410(%rip),%xmm8 # 5fa0 <_sk_callback_sse41+0xe8a> 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 102,15,111,29,219,35,0,0 ; movdqa 0x23db(%rip),%xmm3 # 5f80 <_sk_callback_sse41+0xe98> + DB 102,15,111,29,11,36,0,0 ; movdqa 0x240b(%rip),%xmm3 # 5fb0 <_sk_callback_sse41+0xe9a> 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 102,68,15,111,21,199,35,0,0 ; movdqa 0x23c7(%rip),%xmm10 # 5f90 <_sk_callback_sse41+0xea8> + DB 102,68,15,111,21,247,35,0,0 ; movdqa 0x23f7(%rip),%xmm10 # 5fc0 <_sk_callback_sse41+0xeaa> 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 @@ -15732,17 +15758,17 @@ 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 102,68,15,111,21,252,34,0,0 ; movdqa 0x22fc(%rip),%xmm10 # 5fa0 <_sk_callback_sse41+0xeb8> + DB 102,68,15,111,21,44,35,0,0 ; movdqa 0x232c(%rip),%xmm10 # 5fd0 <_sk_callback_sse41+0xeba> DB 102,68,15,111,216 ; movdqa %xmm0,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 DB 102,69,15,239,235 ; pxor %xmm11,%xmm13 - DB 102,68,15,111,13,239,34,0,0 ; movdqa 0x22ef(%rip),%xmm9 # 5fb0 <_sk_callback_sse41+0xec8> + DB 102,68,15,111,13,31,35,0,0 ; movdqa 0x231f(%rip),%xmm9 # 5fe0 <_sk_callback_sse41+0xeca> DB 102,65,15,114,211,16 ; psrld $0x10,%xmm11 DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 102,68,15,111,37,224,34,0,0 ; movdqa 0x22e0(%rip),%xmm12 # 5fc0 <_sk_callback_sse41+0xed8> + DB 102,68,15,111,37,16,35,0,0 ; movdqa 0x2310(%rip),%xmm12 # 5ff0 <_sk_callback_sse41+0xeda> DB 102,69,15,235,220 ; por %xmm12,%xmm11 DB 102,69,15,254,221 ; paddd %xmm13,%xmm11 DB 102,69,15,223,195 ; pandn %xmm11,%xmm8 @@ -15824,7 +15850,7 @@ _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 68,15,40,5,113,33,0,0 ; movaps 0x2171(%rip),%xmm8 # 5fd0 <_sk_callback_sse41+0xee8> + DB 68,15,40,5,161,33,0,0 ; movaps 0x21a1(%rip),%xmm8 # 6000 <_sk_callback_sse41+0xeea> 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 @@ -15889,7 +15915,7 @@ _sk_load_rgb_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 68,15,40,5,105,32,0,0 ; movaps 0x2069(%rip),%xmm8 # 5fe0 <_sk_callback_sse41+0xef8> + DB 68,15,40,5,153,32,0,0 ; movaps 0x2099(%rip),%xmm8 # 6010 <_sk_callback_sse41+0xefa> 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 @@ -15906,7 +15932,7 @@ _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 15,40,29,48,32,0,0 ; movaps 0x2030(%rip),%xmm3 # 5ff0 <_sk_callback_sse41+0xf08> + DB 15,40,29,96,32,0,0 ; movaps 0x2060(%rip),%xmm3 # 6020 <_sk_callback_sse41+0xf0a> DB 255,224 ; jmpq *%rax DB 102,65,15,110,20,65 ; movd (%r9,%rax,2),%xmm2 DB 102,65,15,196,84,65,4,2 ; pinsrw $0x2,0x4(%r9,%rax,2),%xmm2 @@ -15933,7 +15959,7 @@ _sk_store_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,8 ; mov (%rax),%r9 DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax - DB 68,15,40,21,202,31,0,0 ; movaps 0x1fca(%rip),%xmm10 # 6000 <_sk_callback_sse41+0xf18> + DB 68,15,40,21,250,31,0,0 ; movaps 0x1ffa(%rip),%xmm10 # 6030 <_sk_callback_sse41+0xf1a> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 @@ -16096,10 +16122,11 @@ PUBLIC _sk_repeat_x_sse41 _sk_repeat_x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 69,15,94,200 ; divps %xmm8,%xmm9 + DB 243,68,15,16,72,4 ; movss 0x4(%rax),%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 65,15,92,193 ; subps %xmm9,%xmm0 DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 @@ -16112,10 +16139,11 @@ PUBLIC _sk_repeat_y_sse41 _sk_repeat_y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,201 ; movaps %xmm1,%xmm9 - DB 69,15,94,200 ; divps %xmm8,%xmm9 + DB 243,68,15,16,72,4 ; movss 0x4(%rax),%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 65,15,92,201 ; subps %xmm9,%xmm1 DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 @@ -16128,22 +16156,24 @@ PUBLIC _sk_mirror_x_sse41 _sk_mirror_x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,40,200 ; movaps %xmm8,%xmm9 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 243,68,15,16,72,4 ; movss 0x4(%rax),%xmm9 + DB 69,15,40,208 ; movaps %xmm8,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 65,15,92,194 ; subps %xmm10,%xmm0 DB 243,69,15,88,192 ; addss %xmm8,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,208 ; movaps %xmm0,%xmm10 - DB 69,15,94,208 ; divps %xmm8,%xmm10 - DB 102,69,15,58,8,210,1 ; roundps $0x1,%xmm10,%xmm10 - DB 69,15,89,208 ; mulps %xmm8,%xmm10 - DB 65,15,92,194 ; subps %xmm10,%xmm0 + DB 243,68,15,89,13,198,32,0,0 ; mulss 0x20c6(%rip),%xmm9 # 6400 <_sk_callback_sse41+0x12ea> + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 68,15,89,200 ; mulps %xmm0,%xmm9 + DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 + DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,92,194 ; subps %xmm10,%xmm0 DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,92,192 ; subps %xmm0,%xmm8 DB 65,15,84,192 ; andps %xmm8,%xmm0 DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 - DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 + DB 102,69,15,254,194 ; paddd %xmm10,%xmm8 DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16152,22 +16182,24 @@ PUBLIC _sk_mirror_y_sse41 _sk_mirror_y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,40,200 ; movaps %xmm8,%xmm9 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 65,15,92,201 ; subps %xmm9,%xmm1 + DB 243,68,15,16,72,4 ; movss 0x4(%rax),%xmm9 + DB 69,15,40,208 ; movaps %xmm8,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 65,15,92,202 ; subps %xmm10,%xmm1 DB 243,69,15,88,192 ; addss %xmm8,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,209 ; movaps %xmm1,%xmm10 - DB 69,15,94,208 ; divps %xmm8,%xmm10 - DB 102,69,15,58,8,210,1 ; roundps $0x1,%xmm10,%xmm10 - DB 69,15,89,208 ; mulps %xmm8,%xmm10 - DB 65,15,92,202 ; subps %xmm10,%xmm1 + DB 243,68,15,89,13,99,32,0,0 ; mulss 0x2063(%rip),%xmm9 # 6404 <_sk_callback_sse41+0x12ee> + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 + DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 65,15,92,201 ; subps %xmm9,%xmm1 + DB 65,15,92,202 ; subps %xmm10,%xmm1 DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,92,193 ; subps %xmm1,%xmm8 DB 65,15,84,200 ; andps %xmm8,%xmm1 DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 - DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 + DB 102,69,15,254,194 ; paddd %xmm10,%xmm8 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16176,7 +16208,7 @@ PUBLIC _sk_clamp_x_1_sse41 _sk_clamp_x_1_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,83,28,0,0 ; minps 0x1c53(%rip),%xmm8 # 6010 <_sk_callback_sse41+0xf28> + DB 68,15,93,5,85,28,0,0 ; minps 0x1c55(%rip),%xmm8 # 6040 <_sk_callback_sse41+0xf2a> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -16190,9 +16222,9 @@ _sk_repeat_x_1_sse41 LABEL PROC PUBLIC _sk_mirror_x_1_sse41 _sk_mirror_x_1_sse41 LABEL PROC - DB 68,15,40,5,68,28,0,0 ; movaps 0x1c44(%rip),%xmm8 # 6020 <_sk_callback_sse41+0xf38> + DB 68,15,40,5,70,28,0,0 ; movaps 0x1c46(%rip),%xmm8 # 6050 <_sk_callback_sse41+0xf3a> DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,13,72,28,0,0 ; movaps 0x1c48(%rip),%xmm9 # 6030 <_sk_callback_sse41+0xf48> + DB 68,15,40,13,74,28,0,0 ; movaps 0x1c4a(%rip),%xmm9 # 6060 <_sk_callback_sse41+0xf4a> DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 DB 69,15,88,201 ; addps %xmm9,%xmm9 @@ -16207,10 +16239,10 @@ _sk_mirror_x_1_sse41 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse41 _sk_luminance_to_alpha_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,39,28,0,0 ; mulps 0x1c27(%rip),%xmm0 # 6040 <_sk_callback_sse41+0xf58> - DB 15,89,13,48,28,0,0 ; mulps 0x1c30(%rip),%xmm1 # 6050 <_sk_callback_sse41+0xf68> + DB 15,89,5,41,28,0,0 ; mulps 0x1c29(%rip),%xmm0 # 6070 <_sk_callback_sse41+0xf5a> + DB 15,89,13,50,28,0,0 ; mulps 0x1c32(%rip),%xmm1 # 6080 <_sk_callback_sse41+0xf6a> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,54,28,0,0 ; mulps 0x1c36(%rip),%xmm3 # 6060 <_sk_callback_sse41+0xf78> + DB 15,89,29,56,28,0,0 ; mulps 0x1c38(%rip),%xmm3 # 6090 <_sk_callback_sse41+0xf7a> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -16477,9 +16509,9 @@ _sk_evenly_spaced_gradient_sse41 LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 76,139,112,8 ; mov 0x8(%rax),%r14 DB 72,255,203 ; dec %rbx - DB 120,7 ; js 4868 <_sk_evenly_spaced_gradient_sse41+0x1a> + DB 120,7 ; js 4896 <_sk_evenly_spaced_gradient_sse41+0x1a> DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 - DB 235,21 ; jmp 487d <_sk_evenly_spaced_gradient_sse41+0x2f> + DB 235,21 ; jmp 48ab <_sk_evenly_spaced_gradient_sse41+0x2f> DB 73,137,217 ; mov %rbx,%r9 DB 73,209,233 ; shr %r9 DB 131,227,1 ; and $0x1,%ebx @@ -16567,15 +16599,15 @@ _sk_evenly_spaced_gradient_sse41 LABEL PROC PUBLIC _sk_gauss_a_to_rgba_sse41 _sk_gauss_a_to_rgba_sse41 LABEL PROC - DB 15,40,5,69,22,0,0 ; movaps 0x1645(%rip),%xmm0 # 6070 <_sk_callback_sse41+0xf88> + DB 15,40,5,71,22,0,0 ; movaps 0x1647(%rip),%xmm0 # 60a0 <_sk_callback_sse41+0xf8a> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,75,22,0,0 ; addps 0x164b(%rip),%xmm0 # 6080 <_sk_callback_sse41+0xf98> + DB 15,88,5,77,22,0,0 ; addps 0x164d(%rip),%xmm0 # 60b0 <_sk_callback_sse41+0xf9a> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,81,22,0,0 ; addps 0x1651(%rip),%xmm0 # 6090 <_sk_callback_sse41+0xfa8> + DB 15,88,5,83,22,0,0 ; addps 0x1653(%rip),%xmm0 # 60c0 <_sk_callback_sse41+0xfaa> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,87,22,0,0 ; addps 0x1657(%rip),%xmm0 # 60a0 <_sk_callback_sse41+0xfb8> + DB 15,88,5,89,22,0,0 ; addps 0x1659(%rip),%xmm0 # 60d0 <_sk_callback_sse41+0xfba> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,93,22,0,0 ; addps 0x165d(%rip),%xmm0 # 60b0 <_sk_callback_sse41+0xfc8> + DB 15,88,5,95,22,0,0 ; addps 0x165f(%rip),%xmm0 # 60e0 <_sk_callback_sse41+0xfca> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -16591,12 +16623,12 @@ _sk_gradient_sse41 LABEL PROC DB 76,139,8 ; mov (%rax),%r9 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 73,131,249,2 ; cmp $0x2,%r9 - DB 114,50 ; jb 4aa6 <_sk_gradient_sse41+0x46> + DB 114,50 ; jb 4ad4 <_sk_gradient_sse41+0x46> DB 72,139,88,72 ; mov 0x48(%rax),%rbx DB 73,255,201 ; dec %r9 DB 72,131,195,4 ; add $0x4,%rbx DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,21,54,22,0,0 ; movaps 0x1636(%rip),%xmm2 # 60c0 <_sk_callback_sse41+0xfd8> + DB 15,40,21,56,22,0,0 ; movaps 0x1638(%rip),%xmm2 # 60f0 <_sk_callback_sse41+0xfda> DB 243,15,16,27 ; movss (%rbx),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 @@ -16604,7 +16636,7 @@ _sk_gradient_sse41 LABEL PROC DB 102,15,254,203 ; paddd %xmm3,%xmm1 DB 72,131,195,4 ; add $0x4,%rbx DB 73,255,201 ; dec %r9 - DB 117,228 ; jne 4a8a <_sk_gradient_sse41+0x2a> + DB 117,228 ; jne 4ab8 <_sk_gradient_sse41+0x2a> DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9 DB 69,137,202 ; mov %r9d,%r10d DB 73,193,233,32 ; shr $0x20,%r9 @@ -16730,26 +16762,26 @@ _sk_xy_to_unit_angle_sse41 LABEL PROC DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,40,236 ; movaps %xmm12,%xmm13 DB 69,15,89,237 ; mulps %xmm13,%xmm13 - DB 68,15,40,21,215,19,0,0 ; movaps 0x13d7(%rip),%xmm10 # 60d0 <_sk_callback_sse41+0xfe8> + DB 68,15,40,21,217,19,0,0 ; movaps 0x13d9(%rip),%xmm10 # 6100 <_sk_callback_sse41+0xfea> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,219,19,0,0 ; addps 0x13db(%rip),%xmm10 # 60e0 <_sk_callback_sse41+0xff8> + DB 68,15,88,21,221,19,0,0 ; addps 0x13dd(%rip),%xmm10 # 6110 <_sk_callback_sse41+0xffa> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,223,19,0,0 ; addps 0x13df(%rip),%xmm10 # 60f0 <_sk_callback_sse41+0x1008> + DB 68,15,88,21,225,19,0,0 ; addps 0x13e1(%rip),%xmm10 # 6120 <_sk_callback_sse41+0x100a> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,227,19,0,0 ; addps 0x13e3(%rip),%xmm10 # 6100 <_sk_callback_sse41+0x1018> + DB 68,15,88,21,229,19,0,0 ; addps 0x13e5(%rip),%xmm10 # 6130 <_sk_callback_sse41+0x101a> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 68,15,40,29,226,19,0,0 ; movaps 0x13e2(%rip),%xmm11 # 6110 <_sk_callback_sse41+0x1028> + DB 68,15,40,29,228,19,0,0 ; movaps 0x13e4(%rip),%xmm11 # 6140 <_sk_callback_sse41+0x102a> DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 102,69,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm10 DB 69,15,194,200,1 ; cmpltps %xmm8,%xmm9 - DB 68,15,40,29,219,19,0,0 ; movaps 0x13db(%rip),%xmm11 # 6120 <_sk_callback_sse41+0x1038> + DB 68,15,40,29,221,19,0,0 ; movaps 0x13dd(%rip),%xmm11 # 6150 <_sk_callback_sse41+0x103a> DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 102,69,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm10 DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,194,192,1 ; cmpltps %xmm8,%xmm0 - DB 68,15,40,13,205,19,0,0 ; movaps 0x13cd(%rip),%xmm9 # 6130 <_sk_callback_sse41+0x1048> + DB 68,15,40,13,207,19,0,0 ; movaps 0x13cf(%rip),%xmm9 # 6160 <_sk_callback_sse41+0x104a> DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 102,69,15,56,20,209 ; blendvps %xmm0,%xmm9,%xmm10 DB 69,15,194,194,7 ; cmpordps %xmm10,%xmm8 @@ -16771,7 +16803,7 @@ _sk_xy_to_radius_sse41 LABEL PROC PUBLIC _sk_save_xy_sse41 _sk_save_xy_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,161,19,0,0 ; movaps 0x13a1(%rip),%xmm8 # 6140 <_sk_callback_sse41+0x1058> + DB 68,15,40,5,163,19,0,0 ; movaps 0x13a3(%rip),%xmm8 # 6170 <_sk_callback_sse41+0x105a> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -16811,8 +16843,8 @@ _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 15,88,5,35,19,0,0 ; addps 0x1323(%rip),%xmm0 # 6150 <_sk_callback_sse41+0x1068> - DB 68,15,40,13,43,19,0,0 ; movaps 0x132b(%rip),%xmm9 # 6160 <_sk_callback_sse41+0x1078> + DB 15,88,5,37,19,0,0 ; addps 0x1325(%rip),%xmm0 # 6180 <_sk_callback_sse41+0x106a> + DB 68,15,40,13,45,19,0,0 ; movaps 0x132d(%rip),%xmm9 # 6190 <_sk_callback_sse41+0x107a> 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 @@ -16823,7 +16855,7 @@ _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 15,88,5,26,19,0,0 ; addps 0x131a(%rip),%xmm0 # 6170 <_sk_callback_sse41+0x1088> + DB 15,88,5,28,19,0,0 ; addps 0x131c(%rip),%xmm0 # 61a0 <_sk_callback_sse41+0x108a> 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 @@ -16833,8 +16865,8 @@ _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 15,88,13,12,19,0,0 ; addps 0x130c(%rip),%xmm1 # 6180 <_sk_callback_sse41+0x1098> - DB 68,15,40,13,20,19,0,0 ; movaps 0x1314(%rip),%xmm9 # 6190 <_sk_callback_sse41+0x10a8> + DB 15,88,13,14,19,0,0 ; addps 0x130e(%rip),%xmm1 # 61b0 <_sk_callback_sse41+0x109a> + DB 68,15,40,13,22,19,0,0 ; movaps 0x1316(%rip),%xmm9 # 61c0 <_sk_callback_sse41+0x10aa> 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 @@ -16845,7 +16877,7 @@ _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 15,88,13,2,19,0,0 ; addps 0x1302(%rip),%xmm1 # 61a0 <_sk_callback_sse41+0x10b8> + DB 15,88,13,4,19,0,0 ; addps 0x1304(%rip),%xmm1 # 61d0 <_sk_callback_sse41+0x10ba> 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 @@ -16855,13 +16887,13 @@ _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 15,88,5,245,18,0,0 ; addps 0x12f5(%rip),%xmm0 # 61b0 <_sk_callback_sse41+0x10c8> - DB 68,15,40,13,253,18,0,0 ; movaps 0x12fd(%rip),%xmm9 # 61c0 <_sk_callback_sse41+0x10d8> + DB 15,88,5,247,18,0,0 ; addps 0x12f7(%rip),%xmm0 # 61e0 <_sk_callback_sse41+0x10ca> + DB 68,15,40,13,255,18,0,0 ; movaps 0x12ff(%rip),%xmm9 # 61f0 <_sk_callback_sse41+0x10da> 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 68,15,89,13,249,18,0,0 ; mulps 0x12f9(%rip),%xmm9 # 61d0 <_sk_callback_sse41+0x10e8> - DB 68,15,88,13,1,19,0,0 ; addps 0x1301(%rip),%xmm9 # 61e0 <_sk_callback_sse41+0x10f8> + DB 68,15,89,13,251,18,0,0 ; mulps 0x12fb(%rip),%xmm9 # 6200 <_sk_callback_sse41+0x10ea> + DB 68,15,88,13,3,19,0,0 ; addps 0x1303(%rip),%xmm9 # 6210 <_sk_callback_sse41+0x10fa> 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 @@ -16872,16 +16904,16 @@ _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 15,88,5,240,18,0,0 ; addps 0x12f0(%rip),%xmm0 # 61f0 <_sk_callback_sse41+0x1108> - DB 68,15,40,13,248,18,0,0 ; movaps 0x12f8(%rip),%xmm9 # 6200 <_sk_callback_sse41+0x1118> + DB 15,88,5,242,18,0,0 ; addps 0x12f2(%rip),%xmm0 # 6220 <_sk_callback_sse41+0x110a> + DB 68,15,40,13,250,18,0,0 ; movaps 0x12fa(%rip),%xmm9 # 6230 <_sk_callback_sse41+0x111a> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,252,18,0,0 ; movaps 0x12fc(%rip),%xmm8 # 6210 <_sk_callback_sse41+0x1128> + DB 68,15,40,5,254,18,0,0 ; movaps 0x12fe(%rip),%xmm8 # 6240 <_sk_callback_sse41+0x112a> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,0,19,0,0 ; addps 0x1300(%rip),%xmm8 # 6220 <_sk_callback_sse41+0x1138> + DB 68,15,88,5,2,19,0,0 ; addps 0x1302(%rip),%xmm8 # 6250 <_sk_callback_sse41+0x113a> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,4,19,0,0 ; addps 0x1304(%rip),%xmm8 # 6230 <_sk_callback_sse41+0x1148> + DB 68,15,88,5,6,19,0,0 ; addps 0x1306(%rip),%xmm8 # 6260 <_sk_callback_sse41+0x114a> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,8,19,0,0 ; addps 0x1308(%rip),%xmm8 # 6240 <_sk_callback_sse41+0x1158> + DB 68,15,88,5,10,19,0,0 ; addps 0x130a(%rip),%xmm8 # 6270 <_sk_callback_sse41+0x115a> 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 @@ -16889,17 +16921,17 @@ _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 68,15,40,5,2,19,0,0 ; movaps 0x1302(%rip),%xmm8 # 6250 <_sk_callback_sse41+0x1168> + DB 68,15,40,5,4,19,0,0 ; movaps 0x1304(%rip),%xmm8 # 6280 <_sk_callback_sse41+0x116a> 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 68,15,40,21,254,18,0,0 ; movaps 0x12fe(%rip),%xmm10 # 6260 <_sk_callback_sse41+0x1178> + DB 68,15,40,21,0,19,0,0 ; movaps 0x1300(%rip),%xmm10 # 6290 <_sk_callback_sse41+0x117a> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,2,19,0,0 ; addps 0x1302(%rip),%xmm10 # 6270 <_sk_callback_sse41+0x1188> + DB 68,15,88,21,4,19,0,0 ; addps 0x1304(%rip),%xmm10 # 62a0 <_sk_callback_sse41+0x118a> 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 68,15,88,21,254,18,0,0 ; addps 0x12fe(%rip),%xmm10 # 6280 <_sk_callback_sse41+0x1198> + DB 68,15,88,21,0,19,0,0 ; addps 0x1300(%rip),%xmm10 # 62b0 <_sk_callback_sse41+0x119a> 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 @@ -16909,11 +16941,11 @@ _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 15,88,5,241,18,0,0 ; addps 0x12f1(%rip),%xmm0 # 6290 <_sk_callback_sse41+0x11a8> + DB 15,88,5,243,18,0,0 ; addps 0x12f3(%rip),%xmm0 # 62c0 <_sk_callback_sse41+0x11aa> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,241,18,0,0 ; mulps 0x12f1(%rip),%xmm8 # 62a0 <_sk_callback_sse41+0x11b8> - DB 68,15,88,5,249,18,0,0 ; addps 0x12f9(%rip),%xmm8 # 62b0 <_sk_callback_sse41+0x11c8> + DB 68,15,89,5,243,18,0,0 ; mulps 0x12f3(%rip),%xmm8 # 62d0 <_sk_callback_sse41+0x11ba> + DB 68,15,88,5,251,18,0,0 ; addps 0x12fb(%rip),%xmm8 # 62e0 <_sk_callback_sse41+0x11ca> 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 @@ -16924,13 +16956,13 @@ _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 15,88,13,231,18,0,0 ; addps 0x12e7(%rip),%xmm1 # 62c0 <_sk_callback_sse41+0x11d8> - DB 68,15,40,13,239,18,0,0 ; movaps 0x12ef(%rip),%xmm9 # 62d0 <_sk_callback_sse41+0x11e8> + DB 15,88,13,233,18,0,0 ; addps 0x12e9(%rip),%xmm1 # 62f0 <_sk_callback_sse41+0x11da> + DB 68,15,40,13,241,18,0,0 ; movaps 0x12f1(%rip),%xmm9 # 6300 <_sk_callback_sse41+0x11ea> 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 68,15,89,13,235,18,0,0 ; mulps 0x12eb(%rip),%xmm9 # 62e0 <_sk_callback_sse41+0x11f8> - DB 68,15,88,13,243,18,0,0 ; addps 0x12f3(%rip),%xmm9 # 62f0 <_sk_callback_sse41+0x1208> + DB 68,15,89,13,237,18,0,0 ; mulps 0x12ed(%rip),%xmm9 # 6310 <_sk_callback_sse41+0x11fa> + DB 68,15,88,13,245,18,0,0 ; addps 0x12f5(%rip),%xmm9 # 6320 <_sk_callback_sse41+0x120a> 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 @@ -16941,16 +16973,16 @@ _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 15,88,13,225,18,0,0 ; addps 0x12e1(%rip),%xmm1 # 6300 <_sk_callback_sse41+0x1218> - DB 68,15,40,13,233,18,0,0 ; movaps 0x12e9(%rip),%xmm9 # 6310 <_sk_callback_sse41+0x1228> + DB 15,88,13,227,18,0,0 ; addps 0x12e3(%rip),%xmm1 # 6330 <_sk_callback_sse41+0x121a> + DB 68,15,40,13,235,18,0,0 ; movaps 0x12eb(%rip),%xmm9 # 6340 <_sk_callback_sse41+0x122a> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,237,18,0,0 ; movaps 0x12ed(%rip),%xmm8 # 6320 <_sk_callback_sse41+0x1238> + DB 68,15,40,5,239,18,0,0 ; movaps 0x12ef(%rip),%xmm8 # 6350 <_sk_callback_sse41+0x123a> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,241,18,0,0 ; addps 0x12f1(%rip),%xmm8 # 6330 <_sk_callback_sse41+0x1248> + DB 68,15,88,5,243,18,0,0 ; addps 0x12f3(%rip),%xmm8 # 6360 <_sk_callback_sse41+0x124a> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,245,18,0,0 ; addps 0x12f5(%rip),%xmm8 # 6340 <_sk_callback_sse41+0x1258> + DB 68,15,88,5,247,18,0,0 ; addps 0x12f7(%rip),%xmm8 # 6370 <_sk_callback_sse41+0x125a> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,249,18,0,0 ; addps 0x12f9(%rip),%xmm8 # 6350 <_sk_callback_sse41+0x1268> + DB 68,15,88,5,251,18,0,0 ; addps 0x12fb(%rip),%xmm8 # 6380 <_sk_callback_sse41+0x126a> 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 @@ -16958,17 +16990,17 @@ _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 68,15,40,5,243,18,0,0 ; movaps 0x12f3(%rip),%xmm8 # 6360 <_sk_callback_sse41+0x1278> + DB 68,15,40,5,245,18,0,0 ; movaps 0x12f5(%rip),%xmm8 # 6390 <_sk_callback_sse41+0x127a> 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 68,15,40,21,238,18,0,0 ; movaps 0x12ee(%rip),%xmm10 # 6370 <_sk_callback_sse41+0x1288> + DB 68,15,40,21,240,18,0,0 ; movaps 0x12f0(%rip),%xmm10 # 63a0 <_sk_callback_sse41+0x128a> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,242,18,0,0 ; addps 0x12f2(%rip),%xmm10 # 6380 <_sk_callback_sse41+0x1298> + DB 68,15,88,21,244,18,0,0 ; addps 0x12f4(%rip),%xmm10 # 63b0 <_sk_callback_sse41+0x129a> 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 68,15,88,21,238,18,0,0 ; addps 0x12ee(%rip),%xmm10 # 6390 <_sk_callback_sse41+0x12a8> + DB 68,15,88,21,240,18,0,0 ; addps 0x12f0(%rip),%xmm10 # 63c0 <_sk_callback_sse41+0x12aa> 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 @@ -16978,11 +17010,11 @@ _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 15,88,13,224,18,0,0 ; addps 0x12e0(%rip),%xmm1 # 63a0 <_sk_callback_sse41+0x12b8> + DB 15,88,13,226,18,0,0 ; addps 0x12e2(%rip),%xmm1 # 63d0 <_sk_callback_sse41+0x12ba> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,224,18,0,0 ; mulps 0x12e0(%rip),%xmm8 # 63b0 <_sk_callback_sse41+0x12c8> - DB 68,15,88,5,232,18,0,0 ; addps 0x12e8(%rip),%xmm8 # 63c0 <_sk_callback_sse41+0x12d8> + DB 68,15,89,5,226,18,0,0 ; mulps 0x12e2(%rip),%xmm8 # 63e0 <_sk_callback_sse41+0x12ca> + DB 68,15,88,5,234,18,0,0 ; addps 0x12ea(%rip),%xmm8 # 63f0 <_sk_callback_sse41+0x12da> 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 @@ -17200,11 +17232,11 @@ ALIGN 16 DB 128,191,0,0,128,191,0 ; cmpb $0x0,-0x40800000(%rdi) DB 0,224 ; add %ah,%al DB 64,0,0 ; add %al,(%rax) - DB 224,64 ; loopne 5398 <.literal16+0x1d8> + DB 224,64 ; loopne 53c8 <.literal16+0x1d8> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 539c <.literal16+0x1dc> + DB 224,64 ; loopne 53cc <.literal16+0x1dc> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 53a0 <.literal16+0x1e0> + DB 224,64 ; loopne 53d0 <.literal16+0x1e0> DB 154 ; (bad) DB 153 ; cltd DB 153 ; cltd @@ -17224,13 +17256,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53c1 <.literal16+0x201> + DB 71,225,61 ; rex.RXB loope 53f1 <.literal16+0x201> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53c5 <.literal16+0x205> + DB 71,225,61 ; rex.RXB loope 53f5 <.literal16+0x205> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53c9 <.literal16+0x209> + DB 71,225,61 ; rex.RXB loope 53f9 <.literal16+0x209> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 53cd <.literal16+0x20d> + DB 71,225,61 ; rex.RXB loope 53fd <.literal16+0x20d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17255,13 +17287,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5401 <.literal16+0x241> + DB 71,225,61 ; rex.RXB loope 5431 <.literal16+0x241> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5405 <.literal16+0x245> + DB 71,225,61 ; rex.RXB loope 5435 <.literal16+0x245> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5409 <.literal16+0x249> + DB 71,225,61 ; rex.RXB loope 5439 <.literal16+0x249> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 540d <.literal16+0x24d> + DB 71,225,61 ; rex.RXB loope 543d <.literal16+0x24d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17286,13 +17318,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5441 <.literal16+0x281> + DB 71,225,61 ; rex.RXB loope 5471 <.literal16+0x281> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5445 <.literal16+0x285> + DB 71,225,61 ; rex.RXB loope 5475 <.literal16+0x285> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5449 <.literal16+0x289> + DB 71,225,61 ; rex.RXB loope 5479 <.literal16+0x289> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 544d <.literal16+0x28d> + DB 71,225,61 ; rex.RXB loope 547d <.literal16+0x28d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17317,13 +17349,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5481 <.literal16+0x2c1> + DB 71,225,61 ; rex.RXB loope 54b1 <.literal16+0x2c1> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5485 <.literal16+0x2c5> + DB 71,225,61 ; rex.RXB loope 54b5 <.literal16+0x2c5> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5489 <.literal16+0x2c9> + DB 71,225,61 ; rex.RXB loope 54b9 <.literal16+0x2c9> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 548d <.literal16+0x2cd> + DB 71,225,61 ; rex.RXB loope 54bd <.literal16+0x2cd> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17340,10 +17372,10 @@ ALIGN 16 DB 0,1 ; add %al,(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005478 <_sk_callback_sse41+0xa000390> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0054a8 <_sk_callback_sse41+0xa000392> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005480 <_sk_callback_sse41+0x3000398> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30054b0 <_sk_callback_sse41+0x300039a> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -17362,11 +17394,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) DB 67,0,0 ; rex.XB add %al,(%r8) - DB 127,67 ; jg 54eb <.literal16+0x32b> + DB 127,67 ; jg 551b <.literal16+0x32b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 54ef <.literal16+0x32f> + DB 127,67 ; jg 551f <.literal16+0x32f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 54f3 <.literal16+0x333> + DB 127,67 ; jg 5523 <.literal16+0x333> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -17601,13 +17633,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 56c9 <.literal16+0x509> + DB 224,7 ; loopne 56f9 <.literal16+0x509> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 56cd <.literal16+0x50d> + DB 224,7 ; loopne 56fd <.literal16+0x50d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 56d1 <.literal16+0x511> + DB 224,7 ; loopne 5701 <.literal16+0x511> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 56d5 <.literal16+0x515> + DB 224,7 ; loopne 5705 <.literal16+0x515> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -17641,10 +17673,10 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005718 <_sk_callback_sse41+0xa000630> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a005748 <_sk_callback_sse41+0xa000632> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005720 <_sk_callback_sse41+0x3000638> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005750 <_sk_callback_sse41+0x300063a> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -17699,11 +17731,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 57eb <.literal16+0x62b> + DB 127,67 ; jg 581b <.literal16+0x62b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 57ef <.literal16+0x62f> + DB 127,67 ; jg 581f <.literal16+0x62f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 57f3 <.literal16+0x633> + DB 127,67 ; jg 5823 <.literal16+0x633> DB 129,128,128,59,129,128,128,59,129,128; addl $0x80813b80,-0x7f7ec480(%rax) DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,129,128,128,59 ; addb $0x3b,-0x7f7f7ec5(%rax) @@ -17718,16 +17750,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 57e4 <.literal16+0x624> + DB 127,0 ; jg 5814 <.literal16+0x624> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 57e8 <.literal16+0x628> + DB 127,0 ; jg 5818 <.literal16+0x628> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 57ec <.literal16+0x62c> + DB 127,0 ; jg 581c <.literal16+0x62c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 57f0 <.literal16+0x630> + DB 127,0 ; jg 5820 <.literal16+0x630> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17736,7 +17768,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5875 <.literal16+0x6b5> + DB 119,115 ; ja 58a5 <.literal16+0x6b5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17747,7 +17779,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 57d9 <.literal16+0x619> + DB 117,191 ; jne 5809 <.literal16+0x619> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17759,7 +17791,7 @@ 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 ffffffffe9a3981a <_sk_callback_sse41+0xffffffffe9a34732> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3984a <_sk_callback_sse41+0xffffffffe9a34734> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -17814,16 +17846,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 58b4 <.literal16+0x6f4> + DB 127,0 ; jg 58e4 <.literal16+0x6f4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 58b8 <.literal16+0x6f8> + DB 127,0 ; jg 58e8 <.literal16+0x6f8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 58bc <.literal16+0x6fc> + DB 127,0 ; jg 58ec <.literal16+0x6fc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 58c0 <.literal16+0x700> + DB 127,0 ; jg 58f0 <.literal16+0x700> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17832,7 +17864,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5945 <.literal16+0x785> + DB 119,115 ; ja 5975 <.literal16+0x785> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17843,7 +17875,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 58a9 <.literal16+0x6e9> + DB 117,191 ; jne 58d9 <.literal16+0x6e9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17855,7 +17887,7 @@ 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 ffffffffe9a398ea <_sk_callback_sse41+0xffffffffe9a34802> + DB 233,220,63,163,233 ; jmpq ffffffffe9a3991a <_sk_callback_sse41+0xffffffffe9a34804> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -17910,16 +17942,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5984 <.literal16+0x7c4> + DB 127,0 ; jg 59b4 <.literal16+0x7c4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5988 <.literal16+0x7c8> + DB 127,0 ; jg 59b8 <.literal16+0x7c8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 598c <.literal16+0x7cc> + DB 127,0 ; jg 59bc <.literal16+0x7cc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5990 <.literal16+0x7d0> + DB 127,0 ; jg 59c0 <.literal16+0x7d0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -17928,7 +17960,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5a15 <.literal16+0x855> + DB 119,115 ; ja 5a45 <.literal16+0x855> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -17939,7 +17971,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5979 <.literal16+0x7b9> + DB 117,191 ; jne 59a9 <.literal16+0x7b9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -17951,7 +17983,7 @@ 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 ffffffffe9a399ba <_sk_callback_sse41+0xffffffffe9a348d2> + DB 233,220,63,163,233 ; jmpq ffffffffe9a399ea <_sk_callback_sse41+0xffffffffe9a348d4> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -18006,16 +18038,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5a54 <.literal16+0x894> + DB 127,0 ; jg 5a84 <.literal16+0x894> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5a58 <.literal16+0x898> + DB 127,0 ; jg 5a88 <.literal16+0x898> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5a5c <.literal16+0x89c> + DB 127,0 ; jg 5a8c <.literal16+0x89c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5a60 <.literal16+0x8a0> + DB 127,0 ; jg 5a90 <.literal16+0x8a0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -18024,7 +18056,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5ae5 <.literal16+0x925> + DB 119,115 ; ja 5b15 <.literal16+0x925> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -18035,7 +18067,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5a49 <.literal16+0x889> + DB 117,191 ; jne 5a79 <.literal16+0x889> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -18047,7 +18079,7 @@ 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 ffffffffe9a39a8a <_sk_callback_sse41+0xffffffffe9a349a2> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39aba <_sk_callback_sse41+0xffffffffe9a349a4> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -18098,13 +18130,13 @@ ALIGN 16 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 - DB 127,67 ; jg 5b67 <.literal16+0x9a7> + DB 127,67 ; jg 5b97 <.literal16+0x9a7> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5b6b <.literal16+0x9ab> + DB 127,67 ; jg 5b9b <.literal16+0x9ab> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5b6f <.literal16+0x9af> + DB 127,67 ; jg 5b9f <.literal16+0x9af> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5b73 <.literal16+0x9b3> + DB 127,67 ; jg 5ba3 <.literal16+0x9b3> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -18151,16 +18183,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5bf3 <.literal16+0xa33> + DB 118,63 ; jbe 5c23 <.literal16+0xa33> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5bf7 <.literal16+0xa37> + DB 118,63 ; jbe 5c27 <.literal16+0xa37> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5bfb <.literal16+0xa3b> + DB 118,63 ; jbe 5c2b <.literal16+0xa3b> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 5bff <.literal16+0xa3f> + DB 118,63 ; jbe 5c2f <.literal16+0xa3f> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) @@ -18180,11 +18212,11 @@ ALIGN 16 DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5c4b <.literal16+0xa8b> + DB 127,67 ; jg 5c7b <.literal16+0xa8b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5c4f <.literal16+0xa8f> + DB 127,67 ; jg 5c7f <.literal16+0xa8f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5c53 <.literal16+0xa93> + DB 127,67 ; jg 5c83 <.literal16+0xa93> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -18221,7 +18253,7 @@ 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) # 3005c90 <_sk_callback_sse41+0x3000ba8> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005cc0 <_sk_callback_sse41+0x3000baa> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -18250,13 +18282,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5cc9 <.literal16+0xb09> + DB 224,7 ; loopne 5cf9 <.literal16+0xb09> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5ccd <.literal16+0xb0d> + DB 224,7 ; loopne 5cfd <.literal16+0xb0d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5cd1 <.literal16+0xb11> + DB 224,7 ; loopne 5d01 <.literal16+0xb11> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5cd5 <.literal16+0xb15> + DB 224,7 ; loopne 5d05 <.literal16+0xb15> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -18302,13 +18334,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5d39 <.literal16+0xb79> + DB 224,7 ; loopne 5d69 <.literal16+0xb79> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5d3d <.literal16+0xb7d> + DB 224,7 ; loopne 5d6d <.literal16+0xb7d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5d41 <.literal16+0xb81> + DB 224,7 ; loopne 5d71 <.literal16+0xb81> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5d45 <.literal16+0xb85> + DB 224,7 ; loopne 5d75 <.literal16+0xb85> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -18346,13 +18378,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 5dd6 <.literal16+0xc16> + DB 124,66 ; jl 5e06 <.literal16+0xc16> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 5dda <.literal16+0xc1a> + DB 124,66 ; jl 5e0a <.literal16+0xc1a> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 5dde <.literal16+0xc1e> + DB 124,66 ; jl 5e0e <.literal16+0xc1e> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 5de2 <.literal16+0xc22> + DB 124,66 ; jl 5e12 <.literal16+0xc22> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -18442,13 +18474,13 @@ ALIGN 16 DB 136,136,61,137,136,136 ; mov %cl,-0x777776c3(%rax) DB 61,137,136,136,61 ; cmp $0x3d888889,%eax DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5ee5 <.literal16+0xd25> + DB 112,65 ; jo 5f15 <.literal16+0xd25> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5ee9 <.literal16+0xd29> + DB 112,65 ; jo 5f19 <.literal16+0xd29> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5eed <.literal16+0xd2d> + DB 112,65 ; jo 5f1d <.literal16+0xd2d> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 5ef1 <.literal16+0xd31> + DB 112,65 ; jo 5f21 <.literal16+0xd31> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -18463,7 +18495,7 @@ 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) # 3005ee0 <_sk_callback_sse41+0x3000df8> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005f10 <_sk_callback_sse41+0x3000dfa> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -18490,7 +18522,7 @@ 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) # 3005f20 <_sk_callback_sse41+0x3000e38> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3005f50 <_sk_callback_sse41+0x3000e3a> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -18505,11 +18537,11 @@ ALIGN 16 DB 255,0 ; incl (%rax) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5f7b <.literal16+0xdbb> + DB 127,67 ; jg 5fab <.literal16+0xdbb> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5f7f <.literal16+0xdbf> + DB 127,67 ; jg 5faf <.literal16+0xdbf> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5f83 <.literal16+0xdc3> + DB 127,67 ; jg 5fb3 <.literal16+0xdc3> DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) @@ -18585,13 +18617,13 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 604b <.literal16+0xe8b> + DB 127,71 ; jg 607b <.literal16+0xe8b> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 604f <.literal16+0xe8f> + DB 127,71 ; jg 607f <.literal16+0xe8f> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6053 <.literal16+0xe93> + DB 127,71 ; jg 6083 <.literal16+0xe93> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6057 <.literal16+0xe97> + DB 127,71 ; jg 6087 <.literal16+0xe97> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -18637,10 +18669,10 @@ ALIGN 16 DB 61,152,221,147,61 ; cmp $0x3d93dd98,%eax DB 152 ; cwtl DB 221,147,61,45,16,17 ; fstl 0x11102d3d(%rbx) - DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc0718a <_sk_callback_sse41+0x2dc020a2> + DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc071ba <_sk_callback_sse41+0x2dc020a4> DB 17,192 ; adc %eax,%eax DB 45,16,17,192,18 ; sub $0x12c01110,%eax - DB 120,57 ; js 60bc <.literal16+0xefc> + DB 120,57 ; js 60ec <.literal16+0xefc> DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil @@ -18762,11 +18794,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 61f2 <.literal16+0x1032> + DB 62,114,28 ; jb,pt 6222 <.literal16+0x1032> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 61f6 <.literal16+0x1036> + DB 62,114,28 ; jb,pt 6226 <.literal16+0x1036> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 61fa <.literal16+0x103a> + DB 62,114,28 ; jb,pt 622a <.literal16+0x103a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18810,7 +18842,7 @@ 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 # 3d63f085 <_sk_callback_sse41+0x3d639f9d> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f0b5 <_sk_callback_sse41+0x3d639f9f> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -18836,7 +18868,7 @@ 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 # 3d63f0c5 <_sk_callback_sse41+0x3d639fdd> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f0f5 <_sk_callback_sse41+0x3d639fdf> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -18845,13 +18877,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 62be <.literal16+0x10fe> + DB 114,28 ; jb 62ee <.literal16+0x10fe> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 62c2 <.literal16+0x1102> + DB 62,114,28 ; jb,pt 62f2 <.literal16+0x1102> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 62c6 <.literal16+0x1106> + DB 62,114,28 ; jb,pt 62f6 <.literal16+0x1106> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 62ca <.literal16+0x110a> + DB 62,114,28 ; jb,pt 62fa <.literal16+0x110a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18872,11 +18904,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6302 <.literal16+0x1142> + DB 62,114,28 ; jb,pt 6332 <.literal16+0x1142> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6306 <.literal16+0x1146> + DB 62,114,28 ; jb,pt 6336 <.literal16+0x1146> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 630a <.literal16+0x114a> + DB 62,114,28 ; jb,pt 633a <.literal16+0x114a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18920,7 +18952,7 @@ 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 # 3d63f195 <_sk_callback_sse41+0x3d63a0ad> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f1c5 <_sk_callback_sse41+0x3d63a0af> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -18946,7 +18978,7 @@ 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 # 3d63f1d5 <_sk_callback_sse41+0x3d63a0ed> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f205 <_sk_callback_sse41+0x3d63a0ef> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -18955,13 +18987,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 63ce <.literal16+0x120e> + DB 114,28 ; jb 63fe <.literal16+0x120e> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 63d2 <_sk_callback_sse41+0x12ea> + DB 62,114,28 ; jb,pt 6402 <_sk_callback_sse41+0x12ec> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 63d6 <_sk_callback_sse41+0x12ee> + DB 62,114,28 ; jb,pt 6406 <_sk_callback_sse41+0x12f0> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 63da <_sk_callback_sse41+0x12f2> + DB 62,114,28 ; jb,pt 640a <_sk_callback_sse41+0x12f4> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -18971,6 +19003,12 @@ ALIGN 16 DB 170 ; stos %al,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) DB 190,171,170,170,190 ; mov $0xbeaaaaab,%esi + +ALIGN 4 + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) ALIGN 32 PUBLIC _sk_start_pipeline_sse2 @@ -19072,7 +19110,7 @@ _sk_seed_shader_sse2 LABEL PROC DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 - DB 15,40,21,167,85,0,0 ; movaps 0x55a7(%rip),%xmm2 # 5700 <_sk_callback_sse2+0xcb> + DB 15,40,21,215,85,0,0 ; movaps 0x55d7(%rip),%xmm2 # 5730 <_sk_callback_sse2+0xcd> DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,16,7 ; movups (%rdi),%xmm0 DB 15,88,193 ; addps %xmm1,%xmm0 @@ -19081,7 +19119,7 @@ _sk_seed_shader_sse2 LABEL PROC DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,150,85,0,0 ; movaps 0x5596(%rip),%xmm2 # 5710 <_sk_callback_sse2+0xdb> + DB 15,40,21,198,85,0,0 ; movaps 0x55c6(%rip),%xmm2 # 5740 <_sk_callback_sse2+0xdd> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -19099,14 +19137,14 @@ _sk_dither_sse2 LABEL PROC DB 102,68,15,110,193 ; movd %ecx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 - DB 102,68,15,111,21,100,85,0,0 ; movdqa 0x5564(%rip),%xmm10 # 5720 <_sk_callback_sse2+0xeb> + DB 102,68,15,111,21,148,85,0,0 ; movdqa 0x5594(%rip),%xmm10 # 5750 <_sk_callback_sse2+0xed> DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,65,15,114,243,5 ; pslld $0x5,%xmm11 DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 102,65,15,114,242,4 ; pslld $0x4,%xmm10 - DB 102,68,15,111,37,80,85,0,0 ; movdqa 0x5550(%rip),%xmm12 # 5730 <_sk_callback_sse2+0xfb> - DB 102,68,15,111,45,87,85,0,0 ; movdqa 0x5557(%rip),%xmm13 # 5740 <_sk_callback_sse2+0x10b> + DB 102,68,15,111,37,128,85,0,0 ; movdqa 0x5580(%rip),%xmm12 # 5760 <_sk_callback_sse2+0xfd> + DB 102,68,15,111,45,135,85,0,0 ; movdqa 0x5587(%rip),%xmm13 # 5770 <_sk_callback_sse2+0x10d> DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 DB 102,69,15,219,245 ; pand %xmm13,%xmm14 DB 102,65,15,114,246,2 ; pslld $0x2,%xmm14 @@ -19122,8 +19160,8 @@ _sk_dither_sse2 LABEL PROC DB 102,69,15,235,245 ; por %xmm13,%xmm14 DB 102,69,15,235,240 ; por %xmm8,%xmm14 DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8 - DB 68,15,89,5,18,85,0,0 ; mulps 0x5512(%rip),%xmm8 # 5750 <_sk_callback_sse2+0x11b> - DB 68,15,88,5,26,85,0,0 ; addps 0x551a(%rip),%xmm8 # 5760 <_sk_callback_sse2+0x12b> + DB 68,15,89,5,66,85,0,0 ; mulps 0x5542(%rip),%xmm8 # 5780 <_sk_callback_sse2+0x11d> + DB 68,15,88,5,74,85,0,0 ; addps 0x554a(%rip),%xmm8 # 5790 <_sk_callback_sse2+0x12d> DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 @@ -19190,7 +19228,7 @@ _sk_clear_sse2 LABEL PROC PUBLIC _sk_srcatop_sse2 _sk_srcatop_sse2 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,116,84,0,0 ; movaps 0x5474(%rip),%xmm8 # 5770 <_sk_callback_sse2+0x13b> + DB 68,15,40,5,164,84,0,0 ; movaps 0x54a4(%rip),%xmm8 # 57a0 <_sk_callback_sse2+0x13d> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -19213,7 +19251,7 @@ PUBLIC _sk_dstatop_sse2 _sk_dstatop_sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 DB 68,15,89,196 ; mulps %xmm4,%xmm8 - DB 68,15,40,13,55,84,0,0 ; movaps 0x5437(%rip),%xmm9 # 5780 <_sk_callback_sse2+0x14b> + DB 68,15,40,13,103,84,0,0 ; movaps 0x5467(%rip),%xmm9 # 57b0 <_sk_callback_sse2+0x14d> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -19254,7 +19292,7 @@ _sk_dstin_sse2 LABEL PROC PUBLIC _sk_srcout_sse2 _sk_srcout_sse2 LABEL PROC - DB 68,15,40,5,219,83,0,0 ; movaps 0x53db(%rip),%xmm8 # 5790 <_sk_callback_sse2+0x15b> + DB 68,15,40,5,11,84,0,0 ; movaps 0x540b(%rip),%xmm8 # 57c0 <_sk_callback_sse2+0x15d> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -19265,7 +19303,7 @@ _sk_srcout_sse2 LABEL PROC PUBLIC _sk_dstout_sse2 _sk_dstout_sse2 LABEL PROC - DB 68,15,40,5,203,83,0,0 ; movaps 0x53cb(%rip),%xmm8 # 57a0 <_sk_callback_sse2+0x16b> + DB 68,15,40,5,251,83,0,0 ; movaps 0x53fb(%rip),%xmm8 # 57d0 <_sk_callback_sse2+0x16d> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -19280,7 +19318,7 @@ _sk_dstout_sse2 LABEL PROC PUBLIC _sk_srcover_sse2 _sk_srcover_sse2 LABEL PROC - DB 68,15,40,5,174,83,0,0 ; movaps 0x53ae(%rip),%xmm8 # 57b0 <_sk_callback_sse2+0x17b> + DB 68,15,40,5,222,83,0,0 ; movaps 0x53de(%rip),%xmm8 # 57e0 <_sk_callback_sse2+0x17d> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -19298,7 +19336,7 @@ _sk_srcover_sse2 LABEL PROC PUBLIC _sk_dstover_sse2 _sk_dstover_sse2 LABEL PROC - DB 68,15,40,5,130,83,0,0 ; movaps 0x5382(%rip),%xmm8 # 57c0 <_sk_callback_sse2+0x18b> + DB 68,15,40,5,178,83,0,0 ; movaps 0x53b2(%rip),%xmm8 # 57f0 <_sk_callback_sse2+0x18d> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -19322,7 +19360,7 @@ _sk_modulate_sse2 LABEL PROC PUBLIC _sk_multiply_sse2 _sk_multiply_sse2 LABEL PROC - DB 68,15,40,5,86,83,0,0 ; movaps 0x5356(%rip),%xmm8 # 57d0 <_sk_callback_sse2+0x19b> + DB 68,15,40,5,134,83,0,0 ; movaps 0x5386(%rip),%xmm8 # 5800 <_sk_callback_sse2+0x19d> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -19391,7 +19429,7 @@ _sk_screen_sse2 LABEL PROC PUBLIC _sk_xor__sse2 _sk_xor__sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,139,82,0,0 ; movaps 0x528b(%rip),%xmm3 # 57e0 <_sk_callback_sse2+0x1ab> + DB 15,40,29,187,82,0,0 ; movaps 0x52bb(%rip),%xmm3 # 5810 <_sk_callback_sse2+0x1ad> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -19437,7 +19475,7 @@ _sk_darken_sse2 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,95,209 ; maxps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,246,81,0,0 ; movaps 0x51f6(%rip),%xmm2 # 57f0 <_sk_callback_sse2+0x1bb> + DB 15,40,21,38,82,0,0 ; movaps 0x5226(%rip),%xmm2 # 5820 <_sk_callback_sse2+0x1bd> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19469,7 +19507,7 @@ _sk_lighten_sse2 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,155,81,0,0 ; movaps 0x519b(%rip),%xmm2 # 5800 <_sk_callback_sse2+0x1cb> + DB 15,40,21,203,81,0,0 ; movaps 0x51cb(%rip),%xmm2 # 5830 <_sk_callback_sse2+0x1cd> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19504,7 +19542,7 @@ _sk_difference_sse2 LABEL PROC DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,53,81,0,0 ; movaps 0x5135(%rip),%xmm2 # 5810 <_sk_callback_sse2+0x1db> + DB 15,40,21,101,81,0,0 ; movaps 0x5165(%rip),%xmm2 # 5840 <_sk_callback_sse2+0x1dd> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19530,7 +19568,7 @@ _sk_exclusion_sse2 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,245,80,0,0 ; movaps 0x50f5(%rip),%xmm2 # 5820 <_sk_callback_sse2+0x1eb> + DB 15,40,21,37,81,0,0 ; movaps 0x5125(%rip),%xmm2 # 5850 <_sk_callback_sse2+0x1ed> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -19541,7 +19579,7 @@ _sk_exclusion_sse2 LABEL PROC PUBLIC _sk_colorburn_sse2 _sk_colorburn_sse2 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,232,80,0,0 ; movaps 0x50e8(%rip),%xmm10 # 5830 <_sk_callback_sse2+0x1fb> + DB 68,15,40,21,24,81,0,0 ; movaps 0x5118(%rip),%xmm10 # 5860 <_sk_callback_sse2+0x1fd> DB 69,15,40,202 ; movaps %xmm10,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,217 ; movaps %xmm9,%xmm11 @@ -19633,7 +19671,7 @@ _sk_colorburn_sse2 LABEL PROC PUBLIC _sk_colordodge_sse2 _sk_colordodge_sse2 LABEL PROC DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 68,15,40,21,158,79,0,0 ; movaps 0x4f9e(%rip),%xmm10 # 5840 <_sk_callback_sse2+0x20b> + DB 68,15,40,21,206,79,0,0 ; movaps 0x4fce(%rip),%xmm10 # 5870 <_sk_callback_sse2+0x20d> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -19726,7 +19764,7 @@ _sk_hardlight_sse2 LABEL PROC DB 15,41,52,36 ; movaps %xmm6,(%rsp) DB 15,40,245 ; movaps %xmm5,%xmm6 DB 15,40,236 ; movaps %xmm4,%xmm5 - DB 68,15,40,29,80,78,0,0 ; movaps 0x4e50(%rip),%xmm11 # 5850 <_sk_callback_sse2+0x21b> + DB 68,15,40,29,128,78,0,0 ; movaps 0x4e80(%rip),%xmm11 # 5880 <_sk_callback_sse2+0x21d> DB 69,15,40,211 ; movaps %xmm11,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,194 ; movaps %xmm10,%xmm8 @@ -19813,7 +19851,7 @@ PUBLIC _sk_overlay_sse2 _sk_overlay_sse2 LABEL PROC DB 68,15,40,193 ; movaps %xmm1,%xmm8 DB 68,15,40,232 ; movaps %xmm0,%xmm13 - DB 68,15,40,13,27,77,0,0 ; movaps 0x4d1b(%rip),%xmm9 # 5860 <_sk_callback_sse2+0x22b> + DB 68,15,40,13,75,77,0,0 ; movaps 0x4d4b(%rip),%xmm9 # 5890 <_sk_callback_sse2+0x22d> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -19903,7 +19941,7 @@ _sk_softlight_sse2 LABEL PROC DB 68,15,40,213 ; movaps %xmm5,%xmm10 DB 68,15,94,215 ; divps %xmm7,%xmm10 DB 69,15,84,212 ; andps %xmm12,%xmm10 - DB 68,15,40,13,213,75,0,0 ; movaps 0x4bd5(%rip),%xmm9 # 5870 <_sk_callback_sse2+0x23b> + DB 68,15,40,13,5,76,0,0 ; movaps 0x4c05(%rip),%xmm9 # 58a0 <_sk_callback_sse2+0x23d> DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 69,15,92,250 ; subps %xmm10,%xmm15 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -19916,10 +19954,10 @@ _sk_softlight_sse2 LABEL PROC DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 DB 65,15,88,194 ; addps %xmm10,%xmm0 - DB 68,15,40,53,175,75,0,0 ; movaps 0x4baf(%rip),%xmm14 # 5880 <_sk_callback_sse2+0x24b> + DB 68,15,40,53,223,75,0,0 ; movaps 0x4bdf(%rip),%xmm14 # 58b0 <_sk_callback_sse2+0x24d> DB 69,15,88,222 ; addps %xmm14,%xmm11 DB 68,15,89,216 ; mulps %xmm0,%xmm11 - DB 68,15,40,21,175,75,0,0 ; movaps 0x4baf(%rip),%xmm10 # 5890 <_sk_callback_sse2+0x25b> + DB 68,15,40,21,223,75,0,0 ; movaps 0x4bdf(%rip),%xmm10 # 58c0 <_sk_callback_sse2+0x25d> DB 69,15,89,234 ; mulps %xmm10,%xmm13 DB 69,15,88,235 ; addps %xmm11,%xmm13 DB 15,88,228 ; addps %xmm4,%xmm4 @@ -20064,7 +20102,7 @@ _sk_hue_sse2 LABEL PROC DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 68,15,89,211 ; mulps %xmm3,%xmm10 - DB 68,15,40,5,235,73,0,0 ; movaps 0x49eb(%rip),%xmm8 # 58d0 <_sk_callback_sse2+0x29b> + DB 68,15,40,5,27,74,0,0 ; movaps 0x4a1b(%rip),%xmm8 # 5900 <_sk_callback_sse2+0x29d> DB 69,15,40,216 ; movaps %xmm8,%xmm11 DB 15,40,207 ; movaps %xmm7,%xmm1 DB 68,15,92,217 ; subps %xmm1,%xmm11 @@ -20112,12 +20150,12 @@ _sk_hue_sse2 LABEL PROC DB 69,15,84,206 ; andps %xmm14,%xmm9 DB 69,15,84,214 ; andps %xmm14,%xmm10 DB 65,15,84,214 ; andps %xmm14,%xmm2 - DB 68,15,40,61,248,72,0,0 ; movaps 0x48f8(%rip),%xmm15 # 58a0 <_sk_callback_sse2+0x26b> + DB 68,15,40,61,40,73,0,0 ; movaps 0x4928(%rip),%xmm15 # 58d0 <_sk_callback_sse2+0x26d> DB 65,15,89,231 ; mulps %xmm15,%xmm4 - DB 15,40,5,253,72,0,0 ; movaps 0x48fd(%rip),%xmm0 # 58b0 <_sk_callback_sse2+0x27b> + DB 15,40,5,45,73,0,0 ; movaps 0x492d(%rip),%xmm0 # 58e0 <_sk_callback_sse2+0x27d> DB 15,89,240 ; mulps %xmm0,%xmm6 DB 15,88,244 ; addps %xmm4,%xmm6 - DB 68,15,40,53,255,72,0,0 ; movaps 0x48ff(%rip),%xmm14 # 58c0 <_sk_callback_sse2+0x28b> + DB 68,15,40,53,47,73,0,0 ; movaps 0x492f(%rip),%xmm14 # 58f0 <_sk_callback_sse2+0x28d> DB 68,15,40,239 ; movaps %xmm7,%xmm13 DB 69,15,89,238 ; mulps %xmm14,%xmm13 DB 68,15,88,238 ; addps %xmm6,%xmm13 @@ -20295,14 +20333,14 @@ _sk_saturation_sse2 LABEL PROC DB 68,15,84,211 ; andps %xmm3,%xmm10 DB 68,15,84,203 ; andps %xmm3,%xmm9 DB 15,84,195 ; andps %xmm3,%xmm0 - DB 68,15,40,5,140,70,0,0 ; movaps 0x468c(%rip),%xmm8 # 58e0 <_sk_callback_sse2+0x2ab> + DB 68,15,40,5,188,70,0,0 ; movaps 0x46bc(%rip),%xmm8 # 5910 <_sk_callback_sse2+0x2ad> DB 15,40,214 ; movaps %xmm6,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 - DB 15,40,13,142,70,0,0 ; movaps 0x468e(%rip),%xmm1 # 58f0 <_sk_callback_sse2+0x2bb> + DB 15,40,13,190,70,0,0 ; movaps 0x46be(%rip),%xmm1 # 5920 <_sk_callback_sse2+0x2bd> DB 15,40,221 ; movaps %xmm5,%xmm3 DB 15,89,217 ; mulps %xmm1,%xmm3 DB 15,88,218 ; addps %xmm2,%xmm3 - DB 68,15,40,37,141,70,0,0 ; movaps 0x468d(%rip),%xmm12 # 5900 <_sk_callback_sse2+0x2cb> + DB 68,15,40,37,189,70,0,0 ; movaps 0x46bd(%rip),%xmm12 # 5930 <_sk_callback_sse2+0x2cd> DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 68,15,88,235 ; addps %xmm3,%xmm13 DB 65,15,40,210 ; movaps %xmm10,%xmm2 @@ -20347,7 +20385,7 @@ _sk_saturation_sse2 LABEL PROC DB 15,40,223 ; movaps %xmm7,%xmm3 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,221 ; mulps %xmm5,%xmm3 - DB 68,15,40,5,242,69,0,0 ; movaps 0x45f2(%rip),%xmm8 # 5910 <_sk_callback_sse2+0x2db> + DB 68,15,40,5,34,70,0,0 ; movaps 0x4622(%rip),%xmm8 # 5940 <_sk_callback_sse2+0x2dd> DB 65,15,40,224 ; movaps %xmm8,%xmm4 DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 15,88,253 ; addps %xmm5,%xmm7 @@ -20448,14 +20486,14 @@ _sk_color_sse2 LABEL PROC DB 68,15,40,213 ; movaps %xmm5,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,40,208 ; movaps %xmm8,%xmm2 - DB 68,15,40,45,138,68,0,0 ; movaps 0x448a(%rip),%xmm13 # 5920 <_sk_callback_sse2+0x2eb> + DB 68,15,40,45,186,68,0,0 ; movaps 0x44ba(%rip),%xmm13 # 5950 <_sk_callback_sse2+0x2ed> DB 68,15,40,198 ; movaps %xmm6,%xmm8 DB 69,15,89,197 ; mulps %xmm13,%xmm8 - DB 68,15,40,53,138,68,0,0 ; movaps 0x448a(%rip),%xmm14 # 5930 <_sk_callback_sse2+0x2fb> + DB 68,15,40,53,186,68,0,0 ; movaps 0x44ba(%rip),%xmm14 # 5960 <_sk_callback_sse2+0x2fd> DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,89,198 ; mulps %xmm14,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,29,134,68,0,0 ; movaps 0x4486(%rip),%xmm11 # 5940 <_sk_callback_sse2+0x30b> + DB 68,15,40,29,182,68,0,0 ; movaps 0x44b6(%rip),%xmm11 # 5970 <_sk_callback_sse2+0x30d> DB 69,15,89,227 ; mulps %xmm11,%xmm12 DB 68,15,88,224 ; addps %xmm0,%xmm12 DB 65,15,40,193 ; movaps %xmm9,%xmm0 @@ -20463,7 +20501,7 @@ _sk_color_sse2 LABEL PROC DB 69,15,40,250 ; movaps %xmm10,%xmm15 DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 68,15,88,248 ; addps %xmm0,%xmm15 - DB 68,15,40,5,114,68,0,0 ; movaps 0x4472(%rip),%xmm8 # 5950 <_sk_callback_sse2+0x31b> + DB 68,15,40,5,162,68,0,0 ; movaps 0x44a2(%rip),%xmm8 # 5980 <_sk_callback_sse2+0x31d> DB 65,15,40,224 ; movaps %xmm8,%xmm4 DB 15,92,226 ; subps %xmm2,%xmm4 DB 15,89,252 ; mulps %xmm4,%xmm7 @@ -20599,15 +20637,15 @@ _sk_luminosity_sse2 LABEL PROC DB 68,15,40,205 ; movaps %xmm5,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 DB 15,89,222 ; mulps %xmm6,%xmm3 - DB 68,15,40,37,132,66,0,0 ; movaps 0x4284(%rip),%xmm12 # 5960 <_sk_callback_sse2+0x32b> + DB 68,15,40,37,180,66,0,0 ; movaps 0x42b4(%rip),%xmm12 # 5990 <_sk_callback_sse2+0x32d> DB 68,15,40,199 ; movaps %xmm7,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,132,66,0,0 ; movaps 0x4284(%rip),%xmm13 # 5970 <_sk_callback_sse2+0x33b> + DB 68,15,40,45,180,66,0,0 ; movaps 0x42b4(%rip),%xmm13 # 59a0 <_sk_callback_sse2+0x33d> DB 68,15,40,241 ; movaps %xmm1,%xmm14 DB 69,15,89,245 ; mulps %xmm13,%xmm14 DB 69,15,88,240 ; addps %xmm8,%xmm14 - DB 68,15,40,29,128,66,0,0 ; movaps 0x4280(%rip),%xmm11 # 5980 <_sk_callback_sse2+0x34b> - DB 68,15,40,5,136,66,0,0 ; movaps 0x4288(%rip),%xmm8 # 5990 <_sk_callback_sse2+0x35b> + DB 68,15,40,29,176,66,0,0 ; movaps 0x42b0(%rip),%xmm11 # 59b0 <_sk_callback_sse2+0x34d> + DB 68,15,40,5,184,66,0,0 ; movaps 0x42b8(%rip),%xmm8 # 59c0 <_sk_callback_sse2+0x35d> DB 69,15,40,248 ; movaps %xmm8,%xmm15 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 68,15,92,248 ; subps %xmm0,%xmm15 @@ -20745,7 +20783,7 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 15,133,227,0,0,0 ; jne 19e2 <_sk_srcover_rgba_8888_sse2+0xf1> DB 243,68,15,111,4,144 ; movdqu (%rax,%rdx,4),%xmm8 DB 77,133,192 ; test %r8,%r8 - DB 102,15,111,53,144,64,0,0 ; movdqa 0x4090(%rip),%xmm6 # 59a0 <_sk_callback_sse2+0x36b> + DB 102,15,111,53,192,64,0,0 ; movdqa 0x40c0(%rip),%xmm6 # 59d0 <_sk_callback_sse2+0x36d> DB 102,65,15,111,224 ; movdqa %xmm8,%xmm4 DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 @@ -20759,9 +20797,9 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 15,91,247 ; cvtdq2ps %xmm7,%xmm6 DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8 DB 65,15,91,248 ; cvtdq2ps %xmm8,%xmm7 - DB 68,15,40,5,96,64,0,0 ; movaps 0x4060(%rip),%xmm8 # 59b0 <_sk_callback_sse2+0x37b> + DB 68,15,40,5,144,64,0,0 ; movaps 0x4090(%rip),%xmm8 # 59e0 <_sk_callback_sse2+0x37d> DB 68,15,92,195 ; subps %xmm3,%xmm8 - DB 68,15,40,37,100,64,0,0 ; movaps 0x4064(%rip),%xmm12 # 59c0 <_sk_callback_sse2+0x38b> + DB 68,15,40,37,148,64,0,0 ; movaps 0x4094(%rip),%xmm12 # 59f0 <_sk_callback_sse2+0x38d> DB 65,15,89,196 ; mulps %xmm12,%xmm0 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -20840,7 +20878,7 @@ _sk_clamp_0_sse2 LABEL PROC PUBLIC _sk_clamp_1_sse2 _sk_clamp_1_sse2 LABEL PROC - DB 68,15,40,5,68,63,0,0 ; movaps 0x3f44(%rip),%xmm8 # 59d0 <_sk_callback_sse2+0x39b> + DB 68,15,40,5,116,63,0,0 ; movaps 0x3f74(%rip),%xmm8 # 5a00 <_sk_callback_sse2+0x39d> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -20850,7 +20888,7 @@ _sk_clamp_1_sse2 LABEL PROC PUBLIC _sk_clamp_a_sse2 _sk_clamp_a_sse2 LABEL PROC - DB 15,93,29,57,63,0,0 ; minps 0x3f39(%rip),%xmm3 # 59e0 <_sk_callback_sse2+0x3ab> + DB 15,93,29,105,63,0,0 ; minps 0x3f69(%rip),%xmm3 # 5a10 <_sk_callback_sse2+0x3ad> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -20923,7 +20961,7 @@ _sk_premul_sse2 LABEL PROC PUBLIC _sk_unpremul_sse2 _sk_unpremul_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,164,62,0,0 ; movaps 0x3ea4(%rip),%xmm9 # 59f0 <_sk_callback_sse2+0x3bb> + DB 68,15,40,13,212,62,0,0 ; movaps 0x3ed4(%rip),%xmm9 # 5a20 <_sk_callback_sse2+0x3bd> 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 @@ -20935,20 +20973,20 @@ _sk_unpremul_sse2 LABEL PROC PUBLIC _sk_from_srgb_sse2 _sk_from_srgb_sse2 LABEL PROC - DB 68,15,40,5,143,62,0,0 ; movaps 0x3e8f(%rip),%xmm8 # 5a00 <_sk_callback_sse2+0x3cb> + DB 68,15,40,5,191,62,0,0 ; movaps 0x3ebf(%rip),%xmm8 # 5a30 <_sk_callback_sse2+0x3cd> DB 68,15,40,232 ; movaps %xmm0,%xmm13 DB 69,15,89,232 ; mulps %xmm8,%xmm13 DB 68,15,40,216 ; movaps %xmm0,%xmm11 DB 69,15,89,219 ; mulps %xmm11,%xmm11 - DB 68,15,40,13,135,62,0,0 ; movaps 0x3e87(%rip),%xmm9 # 5a10 <_sk_callback_sse2+0x3db> + DB 68,15,40,13,183,62,0,0 ; movaps 0x3eb7(%rip),%xmm9 # 5a40 <_sk_callback_sse2+0x3dd> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,241 ; mulps %xmm9,%xmm14 - DB 68,15,40,21,135,62,0,0 ; movaps 0x3e87(%rip),%xmm10 # 5a20 <_sk_callback_sse2+0x3eb> + DB 68,15,40,21,183,62,0,0 ; movaps 0x3eb7(%rip),%xmm10 # 5a50 <_sk_callback_sse2+0x3ed> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 69,15,89,243 ; mulps %xmm11,%xmm14 - DB 68,15,40,29,135,62,0,0 ; movaps 0x3e87(%rip),%xmm11 # 5a30 <_sk_callback_sse2+0x3fb> + DB 68,15,40,29,183,62,0,0 ; movaps 0x3eb7(%rip),%xmm11 # 5a60 <_sk_callback_sse2+0x3fd> DB 69,15,88,243 ; addps %xmm11,%xmm14 - DB 68,15,40,37,139,62,0,0 ; movaps 0x3e8b(%rip),%xmm12 # 5a40 <_sk_callback_sse2+0x40b> + DB 68,15,40,37,187,62,0,0 ; movaps 0x3ebb(%rip),%xmm12 # 5a70 <_sk_callback_sse2+0x40d> DB 65,15,194,196,1 ; cmpltps %xmm12,%xmm0 DB 68,15,84,232 ; andps %xmm0,%xmm13 DB 65,15,85,198 ; andnps %xmm14,%xmm0 @@ -20983,22 +21021,22 @@ _sk_from_srgb_sse2 LABEL PROC PUBLIC _sk_to_srgb_sse2 _sk_to_srgb_sse2 LABEL PROC DB 68,15,82,232 ; rsqrtps %xmm0,%xmm13 - DB 68,15,40,5,24,62,0,0 ; movaps 0x3e18(%rip),%xmm8 # 5a50 <_sk_callback_sse2+0x41b> + DB 68,15,40,5,72,62,0,0 ; movaps 0x3e48(%rip),%xmm8 # 5a80 <_sk_callback_sse2+0x41d> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,240 ; mulps %xmm8,%xmm14 - DB 68,15,40,13,24,62,0,0 ; movaps 0x3e18(%rip),%xmm9 # 5a60 <_sk_callback_sse2+0x42b> + DB 68,15,40,13,72,62,0,0 ; movaps 0x3e48(%rip),%xmm9 # 5a90 <_sk_callback_sse2+0x42d> DB 69,15,40,253 ; movaps %xmm13,%xmm15 DB 69,15,89,249 ; mulps %xmm9,%xmm15 - DB 68,15,40,21,24,62,0,0 ; movaps 0x3e18(%rip),%xmm10 # 5a70 <_sk_callback_sse2+0x43b> + DB 68,15,40,21,72,62,0,0 ; movaps 0x3e48(%rip),%xmm10 # 5aa0 <_sk_callback_sse2+0x43d> DB 69,15,88,250 ; addps %xmm10,%xmm15 DB 69,15,89,253 ; mulps %xmm13,%xmm15 - DB 68,15,40,29,24,62,0,0 ; movaps 0x3e18(%rip),%xmm11 # 5a80 <_sk_callback_sse2+0x44b> + DB 68,15,40,29,72,62,0,0 ; movaps 0x3e48(%rip),%xmm11 # 5ab0 <_sk_callback_sse2+0x44d> DB 69,15,88,251 ; addps %xmm11,%xmm15 - DB 68,15,40,37,28,62,0,0 ; movaps 0x3e1c(%rip),%xmm12 # 5a90 <_sk_callback_sse2+0x45b> + DB 68,15,40,37,76,62,0,0 ; movaps 0x3e4c(%rip),%xmm12 # 5ac0 <_sk_callback_sse2+0x45d> DB 69,15,88,236 ; addps %xmm12,%xmm13 DB 69,15,83,237 ; rcpps %xmm13,%xmm13 DB 69,15,89,239 ; mulps %xmm15,%xmm13 - DB 68,15,40,61,24,62,0,0 ; movaps 0x3e18(%rip),%xmm15 # 5aa0 <_sk_callback_sse2+0x46b> + DB 68,15,40,61,72,62,0,0 ; movaps 0x3e48(%rip),%xmm15 # 5ad0 <_sk_callback_sse2+0x46d> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,197 ; andnps %xmm13,%xmm0 @@ -21046,7 +21084,7 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 68,15,93,218 ; minps %xmm2,%xmm11 DB 65,15,40,202 ; movaps %xmm10,%xmm1 DB 65,15,92,203 ; subps %xmm11,%xmm1 - DB 68,15,40,45,113,61,0,0 ; movaps 0x3d71(%rip),%xmm13 # 5ab0 <_sk_callback_sse2+0x47b> + DB 68,15,40,45,161,61,0,0 ; movaps 0x3da1(%rip),%xmm13 # 5ae0 <_sk_callback_sse2+0x47d> DB 68,15,94,233 ; divps %xmm1,%xmm13 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,194,192,0 ; cmpeqps %xmm8,%xmm0 @@ -21055,30 +21093,30 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,40,241 ; movaps %xmm9,%xmm14 DB 68,15,194,242,1 ; cmpltps %xmm2,%xmm14 - DB 68,15,84,53,87,61,0,0 ; andps 0x3d57(%rip),%xmm14 # 5ac0 <_sk_callback_sse2+0x48b> + DB 68,15,84,53,135,61,0,0 ; andps 0x3d87(%rip),%xmm14 # 5af0 <_sk_callback_sse2+0x48d> DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,40,250 ; movaps %xmm10,%xmm15 DB 69,15,194,249,0 ; cmpeqps %xmm9,%xmm15 DB 65,15,92,208 ; subps %xmm8,%xmm2 DB 65,15,89,213 ; mulps %xmm13,%xmm2 - DB 68,15,40,37,74,61,0,0 ; movaps 0x3d4a(%rip),%xmm12 # 5ad0 <_sk_callback_sse2+0x49b> + DB 68,15,40,37,122,61,0,0 ; movaps 0x3d7a(%rip),%xmm12 # 5b00 <_sk_callback_sse2+0x49d> DB 65,15,88,212 ; addps %xmm12,%xmm2 DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 69,15,89,197 ; mulps %xmm13,%xmm8 - DB 68,15,88,5,70,61,0,0 ; addps 0x3d46(%rip),%xmm8 # 5ae0 <_sk_callback_sse2+0x4ab> + DB 68,15,88,5,118,61,0,0 ; addps 0x3d76(%rip),%xmm8 # 5b10 <_sk_callback_sse2+0x4ad> DB 65,15,84,215 ; andps %xmm15,%xmm2 DB 69,15,85,248 ; andnps %xmm8,%xmm15 DB 68,15,86,250 ; orps %xmm2,%xmm15 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,199 ; andnps %xmm15,%xmm0 DB 65,15,86,198 ; orps %xmm14,%xmm0 - DB 15,89,5,55,61,0,0 ; mulps 0x3d37(%rip),%xmm0 # 5af0 <_sk_callback_sse2+0x4bb> + DB 15,89,5,103,61,0,0 ; mulps 0x3d67(%rip),%xmm0 # 5b20 <_sk_callback_sse2+0x4bd> DB 69,15,40,194 ; movaps %xmm10,%xmm8 DB 69,15,194,195,4 ; cmpneqps %xmm11,%xmm8 DB 65,15,84,192 ; andps %xmm8,%xmm0 DB 69,15,92,226 ; subps %xmm10,%xmm12 DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 68,15,40,13,42,61,0,0 ; movaps 0x3d2a(%rip),%xmm9 # 5b00 <_sk_callback_sse2+0x4cb> + DB 68,15,40,13,90,61,0,0 ; movaps 0x3d5a(%rip),%xmm9 # 5b30 <_sk_callback_sse2+0x4cd> DB 65,15,40,210 ; movaps %xmm10,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 68,15,194,202,1 ; cmpltps %xmm2,%xmm9 @@ -21101,7 +21139,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,41,92,36,32 ; movaps %xmm3,0x20(%rsp) DB 68,15,40,218 ; movaps %xmm2,%xmm11 DB 15,40,240 ; movaps %xmm0,%xmm6 - DB 68,15,40,13,229,60,0,0 ; movaps 0x3ce5(%rip),%xmm9 # 5b10 <_sk_callback_sse2+0x4db> + DB 68,15,40,13,21,61,0,0 ; movaps 0x3d15(%rip),%xmm9 # 5b40 <_sk_callback_sse2+0x4dd> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 69,15,194,211,2 ; cmpleps %xmm11,%xmm10 DB 15,40,193 ; movaps %xmm1,%xmm0 @@ -21118,28 +21156,28 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 69,15,88,211 ; addps %xmm11,%xmm10 DB 69,15,88,219 ; addps %xmm11,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 15,40,5,175,60,0,0 ; movaps 0x3caf(%rip),%xmm0 # 5b20 <_sk_callback_sse2+0x4eb> + DB 15,40,5,223,60,0,0 ; movaps 0x3cdf(%rip),%xmm0 # 5b50 <_sk_callback_sse2+0x4ed> DB 15,88,198 ; addps %xmm6,%xmm0 DB 243,15,91,200 ; cvttps2dq %xmm0,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,40,216 ; movaps %xmm0,%xmm3 DB 15,194,217,1 ; cmpltps %xmm1,%xmm3 - DB 15,84,29,167,60,0,0 ; andps 0x3ca7(%rip),%xmm3 # 5b30 <_sk_callback_sse2+0x4fb> + DB 15,84,29,215,60,0,0 ; andps 0x3cd7(%rip),%xmm3 # 5b60 <_sk_callback_sse2+0x4fd> DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,92,193 ; subps %xmm1,%xmm0 - DB 68,15,40,45,169,60,0,0 ; movaps 0x3ca9(%rip),%xmm13 # 5b40 <_sk_callback_sse2+0x50b> + DB 68,15,40,45,217,60,0,0 ; movaps 0x3cd9(%rip),%xmm13 # 5b70 <_sk_callback_sse2+0x50d> DB 69,15,40,197 ; movaps %xmm13,%xmm8 DB 68,15,194,192,2 ; cmpleps %xmm0,%xmm8 DB 69,15,40,242 ; movaps %xmm10,%xmm14 DB 69,15,92,243 ; subps %xmm11,%xmm14 DB 65,15,40,217 ; movaps %xmm9,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 - DB 15,40,21,185,60,0,0 ; movaps 0x3cb9(%rip),%xmm2 # 5b70 <_sk_callback_sse2+0x53b> + DB 15,40,21,233,60,0,0 ; movaps 0x3ce9(%rip),%xmm2 # 5ba0 <_sk_callback_sse2+0x53d> DB 68,15,40,250 ; movaps %xmm2,%xmm15 DB 68,15,194,248,2 ; cmpleps %xmm0,%xmm15 - DB 15,40,13,137,60,0,0 ; movaps 0x3c89(%rip),%xmm1 # 5b50 <_sk_callback_sse2+0x51b> + DB 15,40,13,185,60,0,0 ; movaps 0x3cb9(%rip),%xmm1 # 5b80 <_sk_callback_sse2+0x51d> DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,40,45,143,60,0,0 ; movaps 0x3c8f(%rip),%xmm5 # 5b60 <_sk_callback_sse2+0x52b> + DB 15,40,45,191,60,0,0 ; movaps 0x3cbf(%rip),%xmm5 # 5b90 <_sk_callback_sse2+0x52d> DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,92,224 ; subps %xmm0,%xmm4 DB 65,15,89,230 ; mulps %xmm14,%xmm4 @@ -21162,7 +21200,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,222 ; movaps %xmm6,%xmm3 DB 15,194,216,1 ; cmpltps %xmm0,%xmm3 - DB 15,84,29,4,60,0,0 ; andps 0x3c04(%rip),%xmm3 # 5b30 <_sk_callback_sse2+0x4fb> + DB 15,84,29,52,60,0,0 ; andps 0x3c34(%rip),%xmm3 # 5b60 <_sk_callback_sse2+0x4fd> DB 15,92,195 ; subps %xmm3,%xmm0 DB 68,15,40,230 ; movaps %xmm6,%xmm12 DB 68,15,92,224 ; subps %xmm0,%xmm12 @@ -21192,12 +21230,12 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,40,60,36 ; movaps (%rsp),%xmm7 DB 15,40,231 ; movaps %xmm7,%xmm4 DB 15,85,227 ; andnps %xmm3,%xmm4 - DB 15,88,53,221,59,0,0 ; addps 0x3bdd(%rip),%xmm6 # 5b80 <_sk_callback_sse2+0x54b> + DB 15,88,53,13,60,0,0 ; addps 0x3c0d(%rip),%xmm6 # 5bb0 <_sk_callback_sse2+0x54d> DB 243,15,91,198 ; cvttps2dq %xmm6,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,222 ; movaps %xmm6,%xmm3 DB 15,194,216,1 ; cmpltps %xmm0,%xmm3 - DB 15,84,29,120,59,0,0 ; andps 0x3b78(%rip),%xmm3 # 5b30 <_sk_callback_sse2+0x4fb> + DB 15,84,29,168,59,0,0 ; andps 0x3ba8(%rip),%xmm3 # 5b60 <_sk_callback_sse2+0x4fd> DB 15,92,195 ; subps %xmm3,%xmm0 DB 15,92,240 ; subps %xmm0,%xmm6 DB 15,89,206 ; mulps %xmm6,%xmm1 @@ -21258,9 +21296,9 @@ _sk_scale_u8_sse2 LABEL PROC DB 102,69,15,110,4,18 ; movd (%r10,%rdx,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,219,5,5,59,0,0 ; pand 0x3b05(%rip),%xmm8 # 5b90 <_sk_callback_sse2+0x55b> + DB 102,68,15,219,5,53,59,0,0 ; pand 0x3b35(%rip),%xmm8 # 5bc0 <_sk_callback_sse2+0x55d> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,9,59,0,0 ; mulps 0x3b09(%rip),%xmm8 # 5ba0 <_sk_callback_sse2+0x56b> + DB 68,15,89,5,57,59,0,0 ; mulps 0x3b39(%rip),%xmm8 # 5bd0 <_sk_callback_sse2+0x56d> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -21319,9 +21357,9 @@ _sk_lerp_u8_sse2 LABEL PROC DB 102,69,15,110,4,18 ; movd (%r10,%rdx,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,219,5,74,58,0,0 ; pand 0x3a4a(%rip),%xmm8 # 5bb0 <_sk_callback_sse2+0x57b> + DB 102,68,15,219,5,122,58,0,0 ; pand 0x3a7a(%rip),%xmm8 # 5be0 <_sk_callback_sse2+0x57d> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,78,58,0,0 ; mulps 0x3a4e(%rip),%xmm8 # 5bc0 <_sk_callback_sse2+0x58b> + DB 68,15,89,5,126,58,0,0 ; mulps 0x3a7e(%rip),%xmm8 # 5bf0 <_sk_callback_sse2+0x58d> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -21366,17 +21404,17 @@ _sk_lerp_565_sse2 LABEL PROC DB 15,133,156,0,0,0 ; jne 22a4 <_sk_lerp_565_sse2+0xaa> DB 243,69,15,126,12,82 ; movq (%r10,%rdx,2),%xmm9 DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 - DB 102,68,15,111,5,180,57,0,0 ; movdqa 0x39b4(%rip),%xmm8 # 5bd0 <_sk_callback_sse2+0x59b> + DB 102,68,15,111,5,228,57,0,0 ; movdqa 0x39e4(%rip),%xmm8 # 5c00 <_sk_callback_sse2+0x59d> DB 102,69,15,219,193 ; pand %xmm9,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,179,57,0,0 ; mulps 0x39b3(%rip),%xmm8 # 5be0 <_sk_callback_sse2+0x5ab> - DB 102,68,15,111,21,186,57,0,0 ; movdqa 0x39ba(%rip),%xmm10 # 5bf0 <_sk_callback_sse2+0x5bb> + DB 68,15,89,5,227,57,0,0 ; mulps 0x39e3(%rip),%xmm8 # 5c10 <_sk_callback_sse2+0x5ad> + DB 102,68,15,111,21,234,57,0,0 ; movdqa 0x39ea(%rip),%xmm10 # 5c20 <_sk_callback_sse2+0x5bd> DB 102,69,15,219,209 ; pand %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 - DB 68,15,89,21,185,57,0,0 ; mulps 0x39b9(%rip),%xmm10 # 5c00 <_sk_callback_sse2+0x5cb> - DB 102,68,15,219,13,192,57,0,0 ; pand 0x39c0(%rip),%xmm9 # 5c10 <_sk_callback_sse2+0x5db> + DB 68,15,89,21,233,57,0,0 ; mulps 0x39e9(%rip),%xmm10 # 5c30 <_sk_callback_sse2+0x5cd> + DB 102,68,15,219,13,240,57,0,0 ; pand 0x39f0(%rip),%xmm9 # 5c40 <_sk_callback_sse2+0x5dd> DB 69,15,91,201 ; cvtdq2ps %xmm9,%xmm9 - DB 68,15,89,13,196,57,0,0 ; mulps 0x39c4(%rip),%xmm9 # 5c20 <_sk_callback_sse2+0x5eb> + DB 68,15,89,13,244,57,0,0 ; mulps 0x39f4(%rip),%xmm9 # 5c50 <_sk_callback_sse2+0x5ed> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -21430,7 +21468,7 @@ _sk_load_tables_sse2 LABEL PROC DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 83 ; push %rbx - DB 102,68,15,111,5,10,57,0,0 ; movdqa 0x390a(%rip),%xmm8 # 5c30 <_sk_callback_sse2+0x5fb> + DB 102,68,15,111,5,58,57,0,0 ; movdqa 0x393a(%rip),%xmm8 # 5c60 <_sk_callback_sse2+0x5fd> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 @@ -21486,7 +21524,7 @@ _sk_load_tables_sse2 LABEL PROC DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 DB 102,65,15,114,209,24 ; psrld $0x18,%xmm9 DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3 - DB 15,89,29,19,56,0,0 ; mulps 0x3813(%rip),%xmm3 # 5c40 <_sk_callback_sse2+0x60b> + DB 15,89,29,67,56,0,0 ; mulps 0x3843(%rip),%xmm3 # 5c70 <_sk_callback_sse2+0x60d> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 @@ -21529,7 +21567,7 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,68,15,111,21,121,55,0,0 ; movdqa 0x3779(%rip),%xmm10 # 5c50 <_sk_callback_sse2+0x61b> + DB 102,68,15,111,21,169,55,0,0 ; movdqa 0x37a9(%rip),%xmm10 # 5c80 <_sk_callback_sse2+0x61d> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,194 ; pand %xmm10,%xmm0 DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 @@ -21591,7 +21629,7 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 102,65,15,235,217 ; por %xmm9,%xmm3 DB 102,65,15,97,216 ; punpcklwd %xmm8,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,100,54,0,0 ; mulps 0x3664(%rip),%xmm3 # 5c60 <_sk_callback_sse2+0x62b> + DB 15,89,29,148,54,0,0 ; mulps 0x3694(%rip),%xmm3 # 5c90 <_sk_callback_sse2+0x62d> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 @@ -21632,7 +21670,7 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC DB 102,68,15,97,208 ; punpcklwd %xmm0,%xmm10 DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 DB 102,65,15,97,194 ; punpcklwd %xmm10,%xmm0 - DB 102,68,15,111,5,211,53,0,0 ; movdqa 0x35d3(%rip),%xmm8 # 5c70 <_sk_callback_sse2+0x63b> + DB 102,68,15,111,5,3,54,0,0 ; movdqa 0x3603(%rip),%xmm8 # 5ca0 <_sk_callback_sse2+0x63d> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 @@ -21688,7 +21726,7 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC DB 15,20,211 ; unpcklps %xmm3,%xmm2 DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,222,52,0,0 ; movaps 0x34de(%rip),%xmm3 # 5c80 <_sk_callback_sse2+0x64b> + DB 15,40,29,14,53,0,0 ; movaps 0x350e(%rip),%xmm3 # 5cb0 <_sk_callback_sse2+0x64d> DB 91 ; pop %rbx DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 @@ -21720,7 +21758,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,119,52,0,0 ; movaps 0x3477(%rip),%xmm8 # 5c90 <_sk_callback_sse2+0x65b> + DB 68,15,40,5,167,52,0,0 ; movaps 0x34a7(%rip),%xmm8 # 5cc0 <_sk_callback_sse2+0x65d> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,73,15,126,193 ; movq %xmm0,%r9 @@ -21748,7 +21786,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,21,17,52,0,0 ; movaps 0x3411(%rip),%xmm10 # 5ca0 <_sk_callback_sse2+0x66b> + DB 68,15,40,21,65,52,0,0 ; movaps 0x3441(%rip),%xmm10 # 5cd0 <_sk_callback_sse2+0x66d> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -21867,7 +21905,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,21,89,50,0,0 ; movaps 0x3259(%rip),%xmm10 # 5cb0 <_sk_callback_sse2+0x67b> + DB 68,15,40,21,137,50,0,0 ; movaps 0x3289(%rip),%xmm10 # 5ce0 <_sk_callback_sse2+0x67d> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -22064,15 +22102,15 @@ _sk_parametric_r_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,141,47,0,0 ; mulps 0x2f8d(%rip),%xmm9 # 5cc0 <_sk_callback_sse2+0x68b> - DB 68,15,84,21,149,47,0,0 ; andps 0x2f95(%rip),%xmm10 # 5cd0 <_sk_callback_sse2+0x69b> - DB 68,15,86,21,157,47,0,0 ; orps 0x2f9d(%rip),%xmm10 # 5ce0 <_sk_callback_sse2+0x6ab> - DB 68,15,88,13,165,47,0,0 ; addps 0x2fa5(%rip),%xmm9 # 5cf0 <_sk_callback_sse2+0x6bb> - DB 68,15,40,37,173,47,0,0 ; movaps 0x2fad(%rip),%xmm12 # 5d00 <_sk_callback_sse2+0x6cb> + DB 68,15,89,13,189,47,0,0 ; mulps 0x2fbd(%rip),%xmm9 # 5cf0 <_sk_callback_sse2+0x68d> + DB 68,15,84,21,197,47,0,0 ; andps 0x2fc5(%rip),%xmm10 # 5d00 <_sk_callback_sse2+0x69d> + DB 68,15,86,21,205,47,0,0 ; orps 0x2fcd(%rip),%xmm10 # 5d10 <_sk_callback_sse2+0x6ad> + DB 68,15,88,13,213,47,0,0 ; addps 0x2fd5(%rip),%xmm9 # 5d20 <_sk_callback_sse2+0x6bd> + DB 68,15,40,37,221,47,0,0 ; movaps 0x2fdd(%rip),%xmm12 # 5d30 <_sk_callback_sse2+0x6cd> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,173,47,0,0 ; addps 0x2fad(%rip),%xmm10 # 5d10 <_sk_callback_sse2+0x6db> - DB 68,15,40,37,181,47,0,0 ; movaps 0x2fb5(%rip),%xmm12 # 5d20 <_sk_callback_sse2+0x6eb> + DB 68,15,88,21,221,47,0,0 ; addps 0x2fdd(%rip),%xmm10 # 5d40 <_sk_callback_sse2+0x6dd> + DB 68,15,40,37,229,47,0,0 ; movaps 0x2fe5(%rip),%xmm12 # 5d50 <_sk_callback_sse2+0x6ed> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -22080,22 +22118,22 @@ _sk_parametric_r_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,159,47,0,0 ; movaps 0x2f9f(%rip),%xmm10 # 5d30 <_sk_callback_sse2+0x6fb> + DB 68,15,40,21,207,47,0,0 ; movaps 0x2fcf(%rip),%xmm10 # 5d60 <_sk_callback_sse2+0x6fd> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,147,47,0,0 ; addps 0x2f93(%rip),%xmm9 # 5d40 <_sk_callback_sse2+0x70b> - DB 68,15,40,37,155,47,0,0 ; movaps 0x2f9b(%rip),%xmm12 # 5d50 <_sk_callback_sse2+0x71b> + DB 68,15,88,13,195,47,0,0 ; addps 0x2fc3(%rip),%xmm9 # 5d70 <_sk_callback_sse2+0x70d> + DB 68,15,40,37,203,47,0,0 ; movaps 0x2fcb(%rip),%xmm12 # 5d80 <_sk_callback_sse2+0x71d> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,155,47,0,0 ; movaps 0x2f9b(%rip),%xmm12 # 5d60 <_sk_callback_sse2+0x72b> + DB 68,15,40,37,203,47,0,0 ; movaps 0x2fcb(%rip),%xmm12 # 5d90 <_sk_callback_sse2+0x72d> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,159,47,0,0 ; movaps 0x2f9f(%rip),%xmm13 # 5d70 <_sk_callback_sse2+0x73b> + DB 68,15,40,45,207,47,0,0 ; movaps 0x2fcf(%rip),%xmm13 # 5da0 <_sk_callback_sse2+0x73d> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,159,47,0,0 ; mulps 0x2f9f(%rip),%xmm13 # 5d80 <_sk_callback_sse2+0x74b> + DB 68,15,89,45,207,47,0,0 ; mulps 0x2fcf(%rip),%xmm13 # 5db0 <_sk_callback_sse2+0x74d> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -22129,15 +22167,15 @@ _sk_parametric_g_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,31,47,0,0 ; mulps 0x2f1f(%rip),%xmm9 # 5d90 <_sk_callback_sse2+0x75b> - DB 68,15,84,21,39,47,0,0 ; andps 0x2f27(%rip),%xmm10 # 5da0 <_sk_callback_sse2+0x76b> - DB 68,15,86,21,47,47,0,0 ; orps 0x2f2f(%rip),%xmm10 # 5db0 <_sk_callback_sse2+0x77b> - DB 68,15,88,13,55,47,0,0 ; addps 0x2f37(%rip),%xmm9 # 5dc0 <_sk_callback_sse2+0x78b> - DB 68,15,40,37,63,47,0,0 ; movaps 0x2f3f(%rip),%xmm12 # 5dd0 <_sk_callback_sse2+0x79b> + DB 68,15,89,13,79,47,0,0 ; mulps 0x2f4f(%rip),%xmm9 # 5dc0 <_sk_callback_sse2+0x75d> + DB 68,15,84,21,87,47,0,0 ; andps 0x2f57(%rip),%xmm10 # 5dd0 <_sk_callback_sse2+0x76d> + DB 68,15,86,21,95,47,0,0 ; orps 0x2f5f(%rip),%xmm10 # 5de0 <_sk_callback_sse2+0x77d> + DB 68,15,88,13,103,47,0,0 ; addps 0x2f67(%rip),%xmm9 # 5df0 <_sk_callback_sse2+0x78d> + DB 68,15,40,37,111,47,0,0 ; movaps 0x2f6f(%rip),%xmm12 # 5e00 <_sk_callback_sse2+0x79d> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,63,47,0,0 ; addps 0x2f3f(%rip),%xmm10 # 5de0 <_sk_callback_sse2+0x7ab> - DB 68,15,40,37,71,47,0,0 ; movaps 0x2f47(%rip),%xmm12 # 5df0 <_sk_callback_sse2+0x7bb> + DB 68,15,88,21,111,47,0,0 ; addps 0x2f6f(%rip),%xmm10 # 5e10 <_sk_callback_sse2+0x7ad> + DB 68,15,40,37,119,47,0,0 ; movaps 0x2f77(%rip),%xmm12 # 5e20 <_sk_callback_sse2+0x7bd> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -22145,22 +22183,22 @@ _sk_parametric_g_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,49,47,0,0 ; movaps 0x2f31(%rip),%xmm10 # 5e00 <_sk_callback_sse2+0x7cb> + DB 68,15,40,21,97,47,0,0 ; movaps 0x2f61(%rip),%xmm10 # 5e30 <_sk_callback_sse2+0x7cd> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,37,47,0,0 ; addps 0x2f25(%rip),%xmm9 # 5e10 <_sk_callback_sse2+0x7db> - DB 68,15,40,37,45,47,0,0 ; movaps 0x2f2d(%rip),%xmm12 # 5e20 <_sk_callback_sse2+0x7eb> + DB 68,15,88,13,85,47,0,0 ; addps 0x2f55(%rip),%xmm9 # 5e40 <_sk_callback_sse2+0x7dd> + DB 68,15,40,37,93,47,0,0 ; movaps 0x2f5d(%rip),%xmm12 # 5e50 <_sk_callback_sse2+0x7ed> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,45,47,0,0 ; movaps 0x2f2d(%rip),%xmm12 # 5e30 <_sk_callback_sse2+0x7fb> + DB 68,15,40,37,93,47,0,0 ; movaps 0x2f5d(%rip),%xmm12 # 5e60 <_sk_callback_sse2+0x7fd> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,49,47,0,0 ; movaps 0x2f31(%rip),%xmm13 # 5e40 <_sk_callback_sse2+0x80b> + DB 68,15,40,45,97,47,0,0 ; movaps 0x2f61(%rip),%xmm13 # 5e70 <_sk_callback_sse2+0x80d> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,49,47,0,0 ; mulps 0x2f31(%rip),%xmm13 # 5e50 <_sk_callback_sse2+0x81b> + DB 68,15,89,45,97,47,0,0 ; mulps 0x2f61(%rip),%xmm13 # 5e80 <_sk_callback_sse2+0x81d> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -22194,15 +22232,15 @@ _sk_parametric_b_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,177,46,0,0 ; mulps 0x2eb1(%rip),%xmm9 # 5e60 <_sk_callback_sse2+0x82b> - DB 68,15,84,21,185,46,0,0 ; andps 0x2eb9(%rip),%xmm10 # 5e70 <_sk_callback_sse2+0x83b> - DB 68,15,86,21,193,46,0,0 ; orps 0x2ec1(%rip),%xmm10 # 5e80 <_sk_callback_sse2+0x84b> - DB 68,15,88,13,201,46,0,0 ; addps 0x2ec9(%rip),%xmm9 # 5e90 <_sk_callback_sse2+0x85b> - DB 68,15,40,37,209,46,0,0 ; movaps 0x2ed1(%rip),%xmm12 # 5ea0 <_sk_callback_sse2+0x86b> + DB 68,15,89,13,225,46,0,0 ; mulps 0x2ee1(%rip),%xmm9 # 5e90 <_sk_callback_sse2+0x82d> + DB 68,15,84,21,233,46,0,0 ; andps 0x2ee9(%rip),%xmm10 # 5ea0 <_sk_callback_sse2+0x83d> + DB 68,15,86,21,241,46,0,0 ; orps 0x2ef1(%rip),%xmm10 # 5eb0 <_sk_callback_sse2+0x84d> + DB 68,15,88,13,249,46,0,0 ; addps 0x2ef9(%rip),%xmm9 # 5ec0 <_sk_callback_sse2+0x85d> + DB 68,15,40,37,1,47,0,0 ; movaps 0x2f01(%rip),%xmm12 # 5ed0 <_sk_callback_sse2+0x86d> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,209,46,0,0 ; addps 0x2ed1(%rip),%xmm10 # 5eb0 <_sk_callback_sse2+0x87b> - DB 68,15,40,37,217,46,0,0 ; movaps 0x2ed9(%rip),%xmm12 # 5ec0 <_sk_callback_sse2+0x88b> + DB 68,15,88,21,1,47,0,0 ; addps 0x2f01(%rip),%xmm10 # 5ee0 <_sk_callback_sse2+0x87d> + DB 68,15,40,37,9,47,0,0 ; movaps 0x2f09(%rip),%xmm12 # 5ef0 <_sk_callback_sse2+0x88d> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -22210,22 +22248,22 @@ _sk_parametric_b_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,195,46,0,0 ; movaps 0x2ec3(%rip),%xmm10 # 5ed0 <_sk_callback_sse2+0x89b> + DB 68,15,40,21,243,46,0,0 ; movaps 0x2ef3(%rip),%xmm10 # 5f00 <_sk_callback_sse2+0x89d> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,183,46,0,0 ; addps 0x2eb7(%rip),%xmm9 # 5ee0 <_sk_callback_sse2+0x8ab> - DB 68,15,40,37,191,46,0,0 ; movaps 0x2ebf(%rip),%xmm12 # 5ef0 <_sk_callback_sse2+0x8bb> + DB 68,15,88,13,231,46,0,0 ; addps 0x2ee7(%rip),%xmm9 # 5f10 <_sk_callback_sse2+0x8ad> + DB 68,15,40,37,239,46,0,0 ; movaps 0x2eef(%rip),%xmm12 # 5f20 <_sk_callback_sse2+0x8bd> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,191,46,0,0 ; movaps 0x2ebf(%rip),%xmm12 # 5f00 <_sk_callback_sse2+0x8cb> + DB 68,15,40,37,239,46,0,0 ; movaps 0x2eef(%rip),%xmm12 # 5f30 <_sk_callback_sse2+0x8cd> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,195,46,0,0 ; movaps 0x2ec3(%rip),%xmm13 # 5f10 <_sk_callback_sse2+0x8db> + DB 68,15,40,45,243,46,0,0 ; movaps 0x2ef3(%rip),%xmm13 # 5f40 <_sk_callback_sse2+0x8dd> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,195,46,0,0 ; mulps 0x2ec3(%rip),%xmm13 # 5f20 <_sk_callback_sse2+0x8eb> + DB 68,15,89,45,243,46,0,0 ; mulps 0x2ef3(%rip),%xmm13 # 5f50 <_sk_callback_sse2+0x8ed> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -22259,15 +22297,15 @@ _sk_parametric_a_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,67,46,0,0 ; mulps 0x2e43(%rip),%xmm9 # 5f30 <_sk_callback_sse2+0x8fb> - DB 68,15,84,21,75,46,0,0 ; andps 0x2e4b(%rip),%xmm10 # 5f40 <_sk_callback_sse2+0x90b> - DB 68,15,86,21,83,46,0,0 ; orps 0x2e53(%rip),%xmm10 # 5f50 <_sk_callback_sse2+0x91b> - DB 68,15,88,13,91,46,0,0 ; addps 0x2e5b(%rip),%xmm9 # 5f60 <_sk_callback_sse2+0x92b> - DB 68,15,40,37,99,46,0,0 ; movaps 0x2e63(%rip),%xmm12 # 5f70 <_sk_callback_sse2+0x93b> + DB 68,15,89,13,115,46,0,0 ; mulps 0x2e73(%rip),%xmm9 # 5f60 <_sk_callback_sse2+0x8fd> + DB 68,15,84,21,123,46,0,0 ; andps 0x2e7b(%rip),%xmm10 # 5f70 <_sk_callback_sse2+0x90d> + DB 68,15,86,21,131,46,0,0 ; orps 0x2e83(%rip),%xmm10 # 5f80 <_sk_callback_sse2+0x91d> + DB 68,15,88,13,139,46,0,0 ; addps 0x2e8b(%rip),%xmm9 # 5f90 <_sk_callback_sse2+0x92d> + DB 68,15,40,37,147,46,0,0 ; movaps 0x2e93(%rip),%xmm12 # 5fa0 <_sk_callback_sse2+0x93d> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,99,46,0,0 ; addps 0x2e63(%rip),%xmm10 # 5f80 <_sk_callback_sse2+0x94b> - DB 68,15,40,37,107,46,0,0 ; movaps 0x2e6b(%rip),%xmm12 # 5f90 <_sk_callback_sse2+0x95b> + DB 68,15,88,21,147,46,0,0 ; addps 0x2e93(%rip),%xmm10 # 5fb0 <_sk_callback_sse2+0x94d> + DB 68,15,40,37,155,46,0,0 ; movaps 0x2e9b(%rip),%xmm12 # 5fc0 <_sk_callback_sse2+0x95d> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -22275,22 +22313,22 @@ _sk_parametric_a_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,85,46,0,0 ; movaps 0x2e55(%rip),%xmm10 # 5fa0 <_sk_callback_sse2+0x96b> + DB 68,15,40,21,133,46,0,0 ; movaps 0x2e85(%rip),%xmm10 # 5fd0 <_sk_callback_sse2+0x96d> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,73,46,0,0 ; addps 0x2e49(%rip),%xmm9 # 5fb0 <_sk_callback_sse2+0x97b> - DB 68,15,40,37,81,46,0,0 ; movaps 0x2e51(%rip),%xmm12 # 5fc0 <_sk_callback_sse2+0x98b> + DB 68,15,88,13,121,46,0,0 ; addps 0x2e79(%rip),%xmm9 # 5fe0 <_sk_callback_sse2+0x97d> + DB 68,15,40,37,129,46,0,0 ; movaps 0x2e81(%rip),%xmm12 # 5ff0 <_sk_callback_sse2+0x98d> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,81,46,0,0 ; movaps 0x2e51(%rip),%xmm12 # 5fd0 <_sk_callback_sse2+0x99b> + DB 68,15,40,37,129,46,0,0 ; movaps 0x2e81(%rip),%xmm12 # 6000 <_sk_callback_sse2+0x99d> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,85,46,0,0 ; movaps 0x2e55(%rip),%xmm13 # 5fe0 <_sk_callback_sse2+0x9ab> + DB 68,15,40,45,133,46,0,0 ; movaps 0x2e85(%rip),%xmm13 # 6010 <_sk_callback_sse2+0x9ad> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,85,46,0,0 ; mulps 0x2e55(%rip),%xmm13 # 5ff0 <_sk_callback_sse2+0x9bb> + DB 68,15,89,45,133,46,0,0 ; mulps 0x2e85(%rip),%xmm13 # 6020 <_sk_callback_sse2+0x9bd> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -22305,29 +22343,29 @@ _sk_parametric_a_sse2 LABEL PROC PUBLIC _sk_lab_to_xyz_sse2 _sk_lab_to_xyz_sse2 LABEL PROC - DB 15,89,5,50,46,0,0 ; mulps 0x2e32(%rip),%xmm0 # 6000 <_sk_callback_sse2+0x9cb> - DB 68,15,40,5,58,46,0,0 ; movaps 0x2e3a(%rip),%xmm8 # 6010 <_sk_callback_sse2+0x9db> + DB 15,89,5,98,46,0,0 ; mulps 0x2e62(%rip),%xmm0 # 6030 <_sk_callback_sse2+0x9cd> + DB 68,15,40,5,106,46,0,0 ; movaps 0x2e6a(%rip),%xmm8 # 6040 <_sk_callback_sse2+0x9dd> DB 65,15,89,200 ; mulps %xmm8,%xmm1 - DB 68,15,40,13,62,46,0,0 ; movaps 0x2e3e(%rip),%xmm9 # 6020 <_sk_callback_sse2+0x9eb> + DB 68,15,40,13,110,46,0,0 ; movaps 0x2e6e(%rip),%xmm9 # 6050 <_sk_callback_sse2+0x9ed> DB 65,15,88,201 ; addps %xmm9,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 65,15,88,209 ; addps %xmm9,%xmm2 - DB 15,88,5,59,46,0,0 ; addps 0x2e3b(%rip),%xmm0 # 6030 <_sk_callback_sse2+0x9fb> - DB 15,89,5,68,46,0,0 ; mulps 0x2e44(%rip),%xmm0 # 6040 <_sk_callback_sse2+0xa0b> - DB 15,89,13,77,46,0,0 ; mulps 0x2e4d(%rip),%xmm1 # 6050 <_sk_callback_sse2+0xa1b> + DB 15,88,5,107,46,0,0 ; addps 0x2e6b(%rip),%xmm0 # 6060 <_sk_callback_sse2+0x9fd> + DB 15,89,5,116,46,0,0 ; mulps 0x2e74(%rip),%xmm0 # 6070 <_sk_callback_sse2+0xa0d> + DB 15,89,13,125,46,0,0 ; mulps 0x2e7d(%rip),%xmm1 # 6080 <_sk_callback_sse2+0xa1d> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,21,83,46,0,0 ; mulps 0x2e53(%rip),%xmm2 # 6060 <_sk_callback_sse2+0xa2b> + DB 15,89,21,131,46,0,0 ; mulps 0x2e83(%rip),%xmm2 # 6090 <_sk_callback_sse2+0xa2d> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 68,15,92,202 ; subps %xmm2,%xmm9 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 DB 68,15,89,225 ; mulps %xmm1,%xmm12 - DB 15,40,21,72,46,0,0 ; movaps 0x2e48(%rip),%xmm2 # 6070 <_sk_callback_sse2+0xa3b> + DB 15,40,21,120,46,0,0 ; movaps 0x2e78(%rip),%xmm2 # 60a0 <_sk_callback_sse2+0xa3d> DB 68,15,40,194 ; movaps %xmm2,%xmm8 DB 69,15,194,196,1 ; cmpltps %xmm12,%xmm8 - DB 68,15,40,21,71,46,0,0 ; movaps 0x2e47(%rip),%xmm10 # 6080 <_sk_callback_sse2+0xa4b> + DB 68,15,40,21,119,46,0,0 ; movaps 0x2e77(%rip),%xmm10 # 60b0 <_sk_callback_sse2+0xa4d> DB 65,15,88,202 ; addps %xmm10,%xmm1 - DB 68,15,40,29,75,46,0,0 ; movaps 0x2e4b(%rip),%xmm11 # 6090 <_sk_callback_sse2+0xa5b> + DB 68,15,40,29,123,46,0,0 ; movaps 0x2e7b(%rip),%xmm11 # 60c0 <_sk_callback_sse2+0xa5d> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 69,15,84,224 ; andps %xmm8,%xmm12 DB 68,15,85,193 ; andnps %xmm1,%xmm8 @@ -22351,8 +22389,8 @@ _sk_lab_to_xyz_sse2 LABEL PROC DB 15,84,194 ; andps %xmm2,%xmm0 DB 65,15,85,209 ; andnps %xmm9,%xmm2 DB 15,86,208 ; orps %xmm0,%xmm2 - DB 68,15,89,5,251,45,0,0 ; mulps 0x2dfb(%rip),%xmm8 # 60a0 <_sk_callback_sse2+0xa6b> - DB 15,89,21,4,46,0,0 ; mulps 0x2e04(%rip),%xmm2 # 60b0 <_sk_callback_sse2+0xa7b> + DB 68,15,89,5,43,46,0,0 ; mulps 0x2e2b(%rip),%xmm8 # 60d0 <_sk_callback_sse2+0xa6d> + DB 15,89,21,52,46,0,0 ; mulps 0x2e34(%rip),%xmm2 # 60e0 <_sk_callback_sse2+0xa7d> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -22366,9 +22404,9 @@ _sk_load_a8_sse2 LABEL PROC DB 102,65,15,110,4,18 ; movd (%r10,%rdx,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,5,236,45,0,0 ; pand 0x2dec(%rip),%xmm0 # 60c0 <_sk_callback_sse2+0xa8b> + DB 102,15,219,5,28,46,0,0 ; pand 0x2e1c(%rip),%xmm0 # 60f0 <_sk_callback_sse2+0xa8d> DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,242,45,0,0 ; mulps 0x2df2(%rip),%xmm3 # 60d0 <_sk_callback_sse2+0xa9b> + DB 15,89,29,34,46,0,0 ; mulps 0x2e22(%rip),%xmm3 # 6100 <_sk_callback_sse2+0xa9d> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 @@ -22434,7 +22472,7 @@ _sk_gather_a8_sse2 LABEL PROC DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,18,45,0,0 ; mulps 0x2d12(%rip),%xmm3 # 60e0 <_sk_callback_sse2+0xaab> + DB 15,89,29,66,45,0,0 ; mulps 0x2d42(%rip),%xmm3 # 6110 <_sk_callback_sse2+0xaad> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -22448,7 +22486,7 @@ _sk_store_a8_sse2 LABEL PROC DB 72,131,236,56 ; sub $0x38,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 68,15,40,5,0,45,0,0 ; movaps 0x2d00(%rip),%xmm8 # 60f0 <_sk_callback_sse2+0xabb> + DB 68,15,40,5,48,45,0,0 ; movaps 0x2d30(%rip),%xmm8 # 6120 <_sk_callback_sse2+0xabd> DB 68,15,89,195 ; mulps %xmm3,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 @@ -22492,11 +22530,11 @@ _sk_load_g8_sse2 LABEL PROC DB 102,65,15,110,4,18 ; movd (%r10,%rdx,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,5,105,44,0,0 ; pand 0x2c69(%rip),%xmm0 # 6100 <_sk_callback_sse2+0xacb> + DB 102,15,219,5,153,44,0,0 ; pand 0x2c99(%rip),%xmm0 # 6130 <_sk_callback_sse2+0xacd> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,111,44,0,0 ; mulps 0x2c6f(%rip),%xmm0 # 6110 <_sk_callback_sse2+0xadb> + DB 15,89,5,159,44,0,0 ; mulps 0x2c9f(%rip),%xmm0 # 6140 <_sk_callback_sse2+0xadd> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,118,44,0,0 ; movaps 0x2c76(%rip),%xmm3 # 6120 <_sk_callback_sse2+0xaeb> + DB 15,40,29,166,44,0,0 ; movaps 0x2ca6(%rip),%xmm3 # 6150 <_sk_callback_sse2+0xaed> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -22560,9 +22598,9 @@ _sk_gather_g8_sse2 LABEL PROC DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,155,43,0,0 ; mulps 0x2b9b(%rip),%xmm0 # 6130 <_sk_callback_sse2+0xafb> + DB 15,89,5,203,43,0,0 ; mulps 0x2bcb(%rip),%xmm0 # 6160 <_sk_callback_sse2+0xafd> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,162,43,0,0 ; movaps 0x2ba2(%rip),%xmm3 # 6140 <_sk_callback_sse2+0xb0b> + DB 15,40,29,210,43,0,0 ; movaps 0x2bd2(%rip),%xmm3 # 6170 <_sk_callback_sse2+0xb0d> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 91 ; pop %rbx @@ -22628,11 +22666,11 @@ _sk_gather_i8_sse2 LABEL PROC DB 102,66,15,110,76,149,0 ; movd 0x0(%rbp,%r10,4),%xmm1 DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,187,42,0,0 ; movdqa 0x2abb(%rip),%xmm2 # 6150 <_sk_callback_sse2+0xb1b> + DB 102,15,111,21,235,42,0,0 ; movdqa 0x2aeb(%rip),%xmm2 # 6180 <_sk_callback_sse2+0xb1d> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,183,42,0,0 ; movaps 0x2ab7(%rip),%xmm8 # 6160 <_sk_callback_sse2+0xb2b> + DB 68,15,40,5,231,42,0,0 ; movaps 0x2ae7(%rip),%xmm8 # 6190 <_sk_callback_sse2+0xb2d> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -22661,19 +22699,19 @@ _sk_load_565_sse2 LABEL PROC DB 117,83 ; jne 374a <_sk_load_565_sse2+0x5d> DB 243,65,15,126,20,82 ; movq (%r10,%rdx,2),%xmm2 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,103,42,0,0 ; movdqa 0x2a67(%rip),%xmm0 # 6170 <_sk_callback_sse2+0xb3b> + DB 102,15,111,5,151,42,0,0 ; movdqa 0x2a97(%rip),%xmm0 # 61a0 <_sk_callback_sse2+0xb3d> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,105,42,0,0 ; mulps 0x2a69(%rip),%xmm0 # 6180 <_sk_callback_sse2+0xb4b> - DB 102,15,111,13,113,42,0,0 ; movdqa 0x2a71(%rip),%xmm1 # 6190 <_sk_callback_sse2+0xb5b> + DB 15,89,5,153,42,0,0 ; mulps 0x2a99(%rip),%xmm0 # 61b0 <_sk_callback_sse2+0xb4d> + DB 102,15,111,13,161,42,0,0 ; movdqa 0x2aa1(%rip),%xmm1 # 61c0 <_sk_callback_sse2+0xb5d> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,115,42,0,0 ; mulps 0x2a73(%rip),%xmm1 # 61a0 <_sk_callback_sse2+0xb6b> - DB 102,15,219,21,123,42,0,0 ; pand 0x2a7b(%rip),%xmm2 # 61b0 <_sk_callback_sse2+0xb7b> + DB 15,89,13,163,42,0,0 ; mulps 0x2aa3(%rip),%xmm1 # 61d0 <_sk_callback_sse2+0xb6d> + DB 102,15,219,21,171,42,0,0 ; pand 0x2aab(%rip),%xmm2 # 61e0 <_sk_callback_sse2+0xb7d> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,129,42,0,0 ; mulps 0x2a81(%rip),%xmm2 # 61c0 <_sk_callback_sse2+0xb8b> + DB 15,89,21,177,42,0,0 ; mulps 0x2ab1(%rip),%xmm2 # 61f0 <_sk_callback_sse2+0xb8d> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,136,42,0,0 ; movaps 0x2a88(%rip),%xmm3 # 61d0 <_sk_callback_sse2+0xb9b> + DB 15,40,29,184,42,0,0 ; movaps 0x2ab8(%rip),%xmm3 # 6200 <_sk_callback_sse2+0xb9d> DB 255,224 ; jmpq *%rax DB 69,137,193 ; mov %r8d,%r9d DB 65,128,225,3 ; and $0x3,%r9b @@ -22728,19 +22766,19 @@ _sk_gather_565_sse2 LABEL PROC DB 102,15,196,208,3 ; pinsrw $0x3,%eax,%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,189,41,0,0 ; movdqa 0x29bd(%rip),%xmm0 # 61e0 <_sk_callback_sse2+0xbab> + DB 102,15,111,5,237,41,0,0 ; movdqa 0x29ed(%rip),%xmm0 # 6210 <_sk_callback_sse2+0xbad> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,191,41,0,0 ; mulps 0x29bf(%rip),%xmm0 # 61f0 <_sk_callback_sse2+0xbbb> - DB 102,15,111,13,199,41,0,0 ; movdqa 0x29c7(%rip),%xmm1 # 6200 <_sk_callback_sse2+0xbcb> + DB 15,89,5,239,41,0,0 ; mulps 0x29ef(%rip),%xmm0 # 6220 <_sk_callback_sse2+0xbbd> + DB 102,15,111,13,247,41,0,0 ; movdqa 0x29f7(%rip),%xmm1 # 6230 <_sk_callback_sse2+0xbcd> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,201,41,0,0 ; mulps 0x29c9(%rip),%xmm1 # 6210 <_sk_callback_sse2+0xbdb> - DB 102,15,219,21,209,41,0,0 ; pand 0x29d1(%rip),%xmm2 # 6220 <_sk_callback_sse2+0xbeb> + DB 15,89,13,249,41,0,0 ; mulps 0x29f9(%rip),%xmm1 # 6240 <_sk_callback_sse2+0xbdd> + DB 102,15,219,21,1,42,0,0 ; pand 0x2a01(%rip),%xmm2 # 6250 <_sk_callback_sse2+0xbed> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,215,41,0,0 ; mulps 0x29d7(%rip),%xmm2 # 6230 <_sk_callback_sse2+0xbfb> + DB 15,89,21,7,42,0,0 ; mulps 0x2a07(%rip),%xmm2 # 6260 <_sk_callback_sse2+0xbfd> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,222,41,0,0 ; movaps 0x29de(%rip),%xmm3 # 6240 <_sk_callback_sse2+0xc0b> + DB 15,40,29,14,42,0,0 ; movaps 0x2a0e(%rip),%xmm3 # 6270 <_sk_callback_sse2+0xc0d> DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax @@ -22748,12 +22786,12 @@ PUBLIC _sk_store_565_sse2 _sk_store_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 68,15,40,5,222,41,0,0 ; movaps 0x29de(%rip),%xmm8 # 6250 <_sk_callback_sse2+0xc1b> + DB 68,15,40,5,14,42,0,0 ; movaps 0x2a0e(%rip),%xmm8 # 6280 <_sk_callback_sse2+0xc1d> 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 68,15,40,21,211,41,0,0 ; movaps 0x29d3(%rip),%xmm10 # 6260 <_sk_callback_sse2+0xc2b> + DB 68,15,40,21,3,42,0,0 ; movaps 0x2a03(%rip),%xmm10 # 6290 <_sk_callback_sse2+0xc2d> 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 @@ -22794,21 +22832,21 @@ _sk_load_4444_sse2 LABEL PROC DB 117,98 ; jne 397d <_sk_load_4444_sse2+0x6c> DB 243,65,15,126,28,82 ; movq (%r10,%rdx,2),%xmm3 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,67,41,0,0 ; movdqa 0x2943(%rip),%xmm0 # 6270 <_sk_callback_sse2+0xc3b> + DB 102,15,111,5,115,41,0,0 ; movdqa 0x2973(%rip),%xmm0 # 62a0 <_sk_callback_sse2+0xc3d> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,69,41,0,0 ; mulps 0x2945(%rip),%xmm0 # 6280 <_sk_callback_sse2+0xc4b> - DB 102,15,111,13,77,41,0,0 ; movdqa 0x294d(%rip),%xmm1 # 6290 <_sk_callback_sse2+0xc5b> + DB 15,89,5,117,41,0,0 ; mulps 0x2975(%rip),%xmm0 # 62b0 <_sk_callback_sse2+0xc4d> + DB 102,15,111,13,125,41,0,0 ; movdqa 0x297d(%rip),%xmm1 # 62c0 <_sk_callback_sse2+0xc5d> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,79,41,0,0 ; mulps 0x294f(%rip),%xmm1 # 62a0 <_sk_callback_sse2+0xc6b> - DB 102,15,111,21,87,41,0,0 ; movdqa 0x2957(%rip),%xmm2 # 62b0 <_sk_callback_sse2+0xc7b> + DB 15,89,13,127,41,0,0 ; mulps 0x297f(%rip),%xmm1 # 62d0 <_sk_callback_sse2+0xc6d> + DB 102,15,111,21,135,41,0,0 ; movdqa 0x2987(%rip),%xmm2 # 62e0 <_sk_callback_sse2+0xc7d> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,89,41,0,0 ; mulps 0x2959(%rip),%xmm2 # 62c0 <_sk_callback_sse2+0xc8b> - DB 102,15,219,29,97,41,0,0 ; pand 0x2961(%rip),%xmm3 # 62d0 <_sk_callback_sse2+0xc9b> + DB 15,89,21,137,41,0,0 ; mulps 0x2989(%rip),%xmm2 # 62f0 <_sk_callback_sse2+0xc8d> + DB 102,15,219,29,145,41,0,0 ; pand 0x2991(%rip),%xmm3 # 6300 <_sk_callback_sse2+0xc9d> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,103,41,0,0 ; mulps 0x2967(%rip),%xmm3 # 62e0 <_sk_callback_sse2+0xcab> + DB 15,89,29,151,41,0,0 ; mulps 0x2997(%rip),%xmm3 # 6310 <_sk_callback_sse2+0xcad> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 69,137,193 ; mov %r8d,%r9d @@ -22864,21 +22902,21 @@ _sk_gather_4444_sse2 LABEL PROC DB 102,15,196,216,3 ; pinsrw $0x3,%eax,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,154,40,0,0 ; movdqa 0x289a(%rip),%xmm0 # 62f0 <_sk_callback_sse2+0xcbb> + DB 102,15,111,5,202,40,0,0 ; movdqa 0x28ca(%rip),%xmm0 # 6320 <_sk_callback_sse2+0xcbd> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,156,40,0,0 ; mulps 0x289c(%rip),%xmm0 # 6300 <_sk_callback_sse2+0xccb> - DB 102,15,111,13,164,40,0,0 ; movdqa 0x28a4(%rip),%xmm1 # 6310 <_sk_callback_sse2+0xcdb> + DB 15,89,5,204,40,0,0 ; mulps 0x28cc(%rip),%xmm0 # 6330 <_sk_callback_sse2+0xccd> + DB 102,15,111,13,212,40,0,0 ; movdqa 0x28d4(%rip),%xmm1 # 6340 <_sk_callback_sse2+0xcdd> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,166,40,0,0 ; mulps 0x28a6(%rip),%xmm1 # 6320 <_sk_callback_sse2+0xceb> - DB 102,15,111,21,174,40,0,0 ; movdqa 0x28ae(%rip),%xmm2 # 6330 <_sk_callback_sse2+0xcfb> + DB 15,89,13,214,40,0,0 ; mulps 0x28d6(%rip),%xmm1 # 6350 <_sk_callback_sse2+0xced> + DB 102,15,111,21,222,40,0,0 ; movdqa 0x28de(%rip),%xmm2 # 6360 <_sk_callback_sse2+0xcfd> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,176,40,0,0 ; mulps 0x28b0(%rip),%xmm2 # 6340 <_sk_callback_sse2+0xd0b> - DB 102,15,219,29,184,40,0,0 ; pand 0x28b8(%rip),%xmm3 # 6350 <_sk_callback_sse2+0xd1b> + DB 15,89,21,224,40,0,0 ; mulps 0x28e0(%rip),%xmm2 # 6370 <_sk_callback_sse2+0xd0d> + DB 102,15,219,29,232,40,0,0 ; pand 0x28e8(%rip),%xmm3 # 6380 <_sk_callback_sse2+0xd1d> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,190,40,0,0 ; mulps 0x28be(%rip),%xmm3 # 6360 <_sk_callback_sse2+0xd2b> + DB 15,89,29,238,40,0,0 ; mulps 0x28ee(%rip),%xmm3 # 6390 <_sk_callback_sse2+0xd2d> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax @@ -22887,7 +22925,7 @@ PUBLIC _sk_store_4444_sse2 _sk_store_4444_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 68,15,40,5,188,40,0,0 ; movaps 0x28bc(%rip),%xmm8 # 6370 <_sk_callback_sse2+0xd3b> + DB 68,15,40,5,236,40,0,0 ; movaps 0x28ec(%rip),%xmm8 # 63a0 <_sk_callback_sse2+0xd3d> 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 @@ -22937,11 +22975,11 @@ _sk_load_8888_sse2 LABEL PROC DB 77,133,192 ; test %r8,%r8 DB 117,98 ; jne 3bd3 <_sk_load_8888_sse2+0x6c> DB 243,68,15,111,12,144 ; movdqu (%rax,%rdx,4),%xmm9 - DB 102,15,111,21,1,40,0,0 ; movdqa 0x2801(%rip),%xmm2 # 6380 <_sk_callback_sse2+0xd4b> + DB 102,15,111,21,49,40,0,0 ; movdqa 0x2831(%rip),%xmm2 # 63b0 <_sk_callback_sse2+0xd4d> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,253,39,0,0 ; movaps 0x27fd(%rip),%xmm8 # 6390 <_sk_callback_sse2+0xd5b> + DB 68,15,40,5,45,40,0,0 ; movaps 0x282d(%rip),%xmm8 # 63c0 <_sk_callback_sse2+0xd5d> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -23007,11 +23045,11 @@ _sk_gather_8888_sse2 LABEL PROC DB 102,67,15,110,12,145 ; movd (%r9,%r10,4),%xmm1 DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,254,38,0,0 ; movdqa 0x26fe(%rip),%xmm2 # 63a0 <_sk_callback_sse2+0xd6b> + DB 102,15,111,21,46,39,0,0 ; movdqa 0x272e(%rip),%xmm2 # 63d0 <_sk_callback_sse2+0xd6d> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,250,38,0,0 ; movaps 0x26fa(%rip),%xmm8 # 63b0 <_sk_callback_sse2+0xd7b> + DB 68,15,40,5,42,39,0,0 ; movaps 0x272a(%rip),%xmm8 # 63e0 <_sk_callback_sse2+0xd7d> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -23034,7 +23072,7 @@ PUBLIC _sk_store_8888_sse2 _sk_store_8888_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 68,15,40,5,188,38,0,0 ; movaps 0x26bc(%rip),%xmm8 # 63c0 <_sk_callback_sse2+0xd8b> + DB 68,15,40,5,236,38,0,0 ; movaps 0x26ec(%rip),%xmm8 # 63f0 <_sk_callback_sse2+0xd8d> 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 @@ -23089,7 +23127,7 @@ _sk_load_f16_sse2 LABEL PROC DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 - DB 102,68,15,111,13,227,37,0,0 ; movdqa 0x25e3(%rip),%xmm9 # 63d0 <_sk_callback_sse2+0xd9b> + DB 102,68,15,111,13,19,38,0,0 ; movdqa 0x2613(%rip),%xmm9 # 6400 <_sk_callback_sse2+0xd9d> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,193 ; pand %xmm9,%xmm0 DB 102,15,239,200 ; pxor %xmm0,%xmm1 @@ -23097,11 +23135,11 @@ _sk_load_f16_sse2 LABEL PROC DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 DB 102,68,15,235,232 ; por %xmm0,%xmm13 - DB 102,68,15,111,29,200,37,0,0 ; movdqa 0x25c8(%rip),%xmm11 # 63e0 <_sk_callback_sse2+0xdab> + DB 102,68,15,111,29,248,37,0,0 ; movdqa 0x25f8(%rip),%xmm11 # 6410 <_sk_callback_sse2+0xdad> DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,202,37,0,0 ; movdqa 0x25ca(%rip),%xmm12 # 63f0 <_sk_callback_sse2+0xdbb> + DB 102,68,15,111,37,250,37,0,0 ; movdqa 0x25fa(%rip),%xmm12 # 6420 <_sk_callback_sse2+0xdbd> DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,205,37,0,0 ; movdqa 0x25cd(%rip),%xmm3 # 6400 <_sk_callback_sse2+0xdcb> + DB 102,15,111,29,253,37,0,0 ; movdqa 0x25fd(%rip),%xmm3 # 6430 <_sk_callback_sse2+0xdcd> DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0 DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 @@ -23199,7 +23237,7 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 - DB 102,68,15,111,13,31,36,0,0 ; movdqa 0x241f(%rip),%xmm9 # 6410 <_sk_callback_sse2+0xddb> + DB 102,68,15,111,13,79,36,0,0 ; movdqa 0x244f(%rip),%xmm9 # 6440 <_sk_callback_sse2+0xddd> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,193 ; pand %xmm9,%xmm0 DB 102,15,239,200 ; pxor %xmm0,%xmm1 @@ -23207,11 +23245,11 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 DB 102,68,15,235,232 ; por %xmm0,%xmm13 - DB 102,68,15,111,29,4,36,0,0 ; movdqa 0x2404(%rip),%xmm11 # 6420 <_sk_callback_sse2+0xdeb> + DB 102,68,15,111,29,52,36,0,0 ; movdqa 0x2434(%rip),%xmm11 # 6450 <_sk_callback_sse2+0xded> DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,6,36,0,0 ; movdqa 0x2406(%rip),%xmm12 # 6430 <_sk_callback_sse2+0xdfb> + DB 102,68,15,111,37,54,36,0,0 ; movdqa 0x2436(%rip),%xmm12 # 6460 <_sk_callback_sse2+0xdfd> DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,9,36,0,0 ; movdqa 0x2409(%rip),%xmm3 # 6440 <_sk_callback_sse2+0xe0b> + DB 102,15,111,29,57,36,0,0 ; movdqa 0x2439(%rip),%xmm3 # 6470 <_sk_callback_sse2+0xe0d> DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0 DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 @@ -23263,17 +23301,17 @@ PUBLIC _sk_store_f16_sse2 _sk_store_f16_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax - DB 102,68,15,111,21,48,35,0,0 ; movdqa 0x2330(%rip),%xmm10 # 6450 <_sk_callback_sse2+0xe1b> + DB 102,68,15,111,21,96,35,0,0 ; movdqa 0x2360(%rip),%xmm10 # 6480 <_sk_callback_sse2+0xe1d> DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 DB 102,69,15,219,226 ; pand %xmm10,%xmm12 DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 - DB 102,68,15,111,13,35,35,0,0 ; movdqa 0x2323(%rip),%xmm9 # 6460 <_sk_callback_sse2+0xe2b> + DB 102,68,15,111,13,83,35,0,0 ; movdqa 0x2353(%rip),%xmm9 # 6490 <_sk_callback_sse2+0xe2d> DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 102,68,15,111,29,20,35,0,0 ; movdqa 0x2314(%rip),%xmm11 # 6470 <_sk_callback_sse2+0xe3b> + DB 102,68,15,111,29,68,35,0,0 ; movdqa 0x2344(%rip),%xmm11 # 64a0 <_sk_callback_sse2+0xe3d> DB 102,69,15,235,227 ; por %xmm11,%xmm12 DB 102,69,15,254,229 ; paddd %xmm13,%xmm12 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 @@ -23364,7 +23402,7 @@ _sk_load_u16_be_sse2 LABEL PROC DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 68,15,40,5,116,33,0,0 ; movaps 0x2174(%rip),%xmm8 # 6480 <_sk_callback_sse2+0xe4b> + DB 68,15,40,5,164,33,0,0 ; movaps 0x21a4(%rip),%xmm8 # 64b0 <_sk_callback_sse2+0xe4d> 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 @@ -23431,7 +23469,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,65,15,97,200 ; punpcklwd %xmm8,%xmm1 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 68,15,40,13,99,32,0,0 ; movaps 0x2063(%rip),%xmm9 # 6490 <_sk_callback_sse2+0xe5b> + DB 68,15,40,13,147,32,0,0 ; movaps 0x2093(%rip),%xmm9 # 64c0 <_sk_callback_sse2+0xe5d> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -23448,7 +23486,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,42,32,0,0 ; movaps 0x202a(%rip),%xmm3 # 64a0 <_sk_callback_sse2+0xe6b> + DB 15,40,29,90,32,0,0 ; movaps 0x205a(%rip),%xmm3 # 64d0 <_sk_callback_sse2+0xe6d> DB 255,224 ; jmpq *%rax DB 102,65,15,110,20,65 ; movd (%r9,%rax,2),%xmm2 DB 102,65,15,196,84,65,4,2 ; pinsrw $0x2,0x4(%r9,%rax,2),%xmm2 @@ -23475,7 +23513,7 @@ _sk_store_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,8 ; mov (%rax),%r9 DB 72,141,4,149,0,0,0,0 ; lea 0x0(,%rdx,4),%rax - DB 68,15,40,21,196,31,0,0 ; movaps 0x1fc4(%rip),%xmm10 # 64b0 <_sk_callback_sse2+0xe7b> + DB 68,15,40,21,244,31,0,0 ; movaps 0x1ff4(%rip),%xmm10 # 64e0 <_sk_callback_sse2+0xe7d> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 @@ -23646,14 +23684,15 @@ PUBLIC _sk_repeat_x_sse2 _sk_repeat_x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 69,15,94,200 ; divps %xmm8,%xmm9 + DB 243,68,15,16,72,4 ; movss 0x4(%rax),%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,31,29,0,0 ; andps 0x1d1f(%rip),%xmm9 # 64c0 <_sk_callback_sse2+0xe8b> + DB 68,15,84,13,77,29,0,0 ; andps 0x1d4d(%rip),%xmm9 # 64f0 <_sk_callback_sse2+0xe8d> DB 69,15,92,209 ; subps %xmm9,%xmm10 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,194 ; subps %xmm10,%xmm0 DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 @@ -23666,14 +23705,15 @@ PUBLIC _sk_repeat_y_sse2 _sk_repeat_y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,201 ; movaps %xmm1,%xmm9 - DB 69,15,94,200 ; divps %xmm8,%xmm9 + DB 243,68,15,16,72,4 ; movss 0x4(%rax),%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,231,28,0,0 ; andps 0x1ce7(%rip),%xmm9 # 64d0 <_sk_callback_sse2+0xe9b> + DB 68,15,84,13,14,29,0,0 ; andps 0x1d0e(%rip),%xmm9 # 6500 <_sk_callback_sse2+0xe9d> DB 69,15,92,209 ; subps %xmm9,%xmm10 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,92,202 ; subps %xmm10,%xmm1 DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 @@ -23685,56 +23725,60 @@ _sk_repeat_y_sse2 LABEL PROC PUBLIC _sk_mirror_x_sse2 _sk_mirror_x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,40,200 ; movaps %xmm8,%xmm9 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 243,69,15,88,192 ; addss %xmm8,%xmm8 + DB 243,68,15,16,8 ; movss (%rax),%xmm9 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,208 ; movaps %xmm0,%xmm10 - DB 69,15,94,208 ; divps %xmm8,%xmm10 + DB 65,15,92,192 ; subps %xmm8,%xmm0 + DB 243,69,15,88,201 ; addss %xmm9,%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 243,68,15,89,21,222,32,0,0 ; mulss 0x20de(%rip),%xmm10 # 6920 <_sk_callback_sse2+0x12bd> + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,208 ; mulps %xmm0,%xmm10 DB 243,69,15,91,218 ; cvttps2dq %xmm10,%xmm11 DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10 - DB 68,15,84,21,157,28,0,0 ; andps 0x1c9d(%rip),%xmm10 # 64e0 <_sk_callback_sse2+0xeab> + DB 68,15,84,21,175,28,0,0 ; andps 0x1caf(%rip),%xmm10 # 6510 <_sk_callback_sse2+0xead> DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 69,15,89,216 ; mulps %xmm8,%xmm11 + DB 69,15,89,217 ; mulps %xmm9,%xmm11 DB 65,15,92,195 ; subps %xmm11,%xmm0 - DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,92,192 ; subps %xmm8,%xmm0 DB 68,15,92,224 ; subps %xmm0,%xmm12 DB 65,15,84,196 ; andps %xmm12,%xmm0 - DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 - DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 - DB 65,15,93,192 ; minps %xmm8,%xmm0 + DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 + DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 + DB 65,15,93,193 ; minps %xmm9,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_mirror_y_sse2 _sk_mirror_y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 69,15,40,200 ; movaps %xmm8,%xmm9 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 65,15,92,201 ; subps %xmm9,%xmm1 - DB 243,69,15,88,192 ; addss %xmm8,%xmm8 + DB 243,68,15,16,8 ; movss (%rax),%xmm9 + DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 + DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 68,15,40,209 ; movaps %xmm1,%xmm10 - DB 69,15,94,208 ; divps %xmm8,%xmm10 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 243,69,15,88,201 ; addss %xmm9,%xmm9 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 + DB 243,68,15,89,21,104,32,0,0 ; mulss 0x2068(%rip),%xmm10 # 6924 <_sk_callback_sse2+0x12c1> + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 68,15,89,209 ; mulps %xmm1,%xmm10 DB 243,69,15,91,218 ; cvttps2dq %xmm10,%xmm11 DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10 - DB 68,15,84,21,67,28,0,0 ; andps 0x1c43(%rip),%xmm10 # 64f0 <_sk_callback_sse2+0xebb> + DB 68,15,84,21,69,28,0,0 ; andps 0x1c45(%rip),%xmm10 # 6520 <_sk_callback_sse2+0xebd> DB 69,15,87,228 ; xorps %xmm12,%xmm12 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 69,15,89,216 ; mulps %xmm8,%xmm11 + DB 69,15,89,217 ; mulps %xmm9,%xmm11 DB 65,15,92,203 ; subps %xmm11,%xmm1 - DB 65,15,92,201 ; subps %xmm9,%xmm1 + DB 65,15,92,200 ; subps %xmm8,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 DB 65,15,84,204 ; andps %xmm12,%xmm1 - DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 - DB 102,69,15,254,193 ; paddd %xmm9,%xmm8 - DB 65,15,93,200 ; minps %xmm8,%xmm1 + DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 + DB 102,69,15,254,200 ; paddd %xmm8,%xmm9 + DB 65,15,93,201 ; minps %xmm9,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -23742,7 +23786,7 @@ PUBLIC _sk_clamp_x_1_sse2 _sk_clamp_x_1_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,21,28,0,0 ; minps 0x1c15(%rip),%xmm8 # 6500 <_sk_callback_sse2+0xecb> + DB 68,15,93,5,23,28,0,0 ; minps 0x1c17(%rip),%xmm8 # 6530 <_sk_callback_sse2+0xecd> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -23753,7 +23797,7 @@ _sk_repeat_x_1_sse2 LABEL PROC DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,194,200,1 ; cmpltps %xmm8,%xmm9 - DB 68,15,84,13,3,28,0,0 ; andps 0x1c03(%rip),%xmm9 # 6510 <_sk_callback_sse2+0xedb> + DB 68,15,84,13,5,28,0,0 ; andps 0x1c05(%rip),%xmm9 # 6540 <_sk_callback_sse2+0xedd> DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 65,15,92,192 ; subps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax @@ -23761,14 +23805,14 @@ _sk_repeat_x_1_sse2 LABEL PROC PUBLIC _sk_mirror_x_1_sse2 _sk_mirror_x_1_sse2 LABEL PROC - DB 68,15,40,5,255,27,0,0 ; movaps 0x1bff(%rip),%xmm8 # 6520 <_sk_callback_sse2+0xeeb> + DB 68,15,40,5,1,28,0,0 ; movaps 0x1c01(%rip),%xmm8 # 6550 <_sk_callback_sse2+0xeed> DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,13,3,28,0,0 ; movaps 0x1c03(%rip),%xmm9 # 6530 <_sk_callback_sse2+0xefb> + DB 68,15,40,13,5,28,0,0 ; movaps 0x1c05(%rip),%xmm9 # 6560 <_sk_callback_sse2+0xefd> DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,249,27,0,0 ; andps 0x1bf9(%rip),%xmm9 # 6540 <_sk_callback_sse2+0xf0b> + DB 68,15,84,13,251,27,0,0 ; andps 0x1bfb(%rip),%xmm9 # 6570 <_sk_callback_sse2+0xf0d> DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,88,210 ; addps %xmm10,%xmm10 @@ -23782,10 +23826,10 @@ _sk_mirror_x_1_sse2 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse2 _sk_luminance_to_alpha_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,223,27,0,0 ; mulps 0x1bdf(%rip),%xmm0 # 6550 <_sk_callback_sse2+0xf1b> - DB 15,89,13,232,27,0,0 ; mulps 0x1be8(%rip),%xmm1 # 6560 <_sk_callback_sse2+0xf2b> + DB 15,89,5,225,27,0,0 ; mulps 0x1be1(%rip),%xmm0 # 6580 <_sk_callback_sse2+0xf1d> + DB 15,89,13,234,27,0,0 ; mulps 0x1bea(%rip),%xmm1 # 6590 <_sk_callback_sse2+0xf2d> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,238,27,0,0 ; mulps 0x1bee(%rip),%xmm3 # 6570 <_sk_callback_sse2+0xf3b> + DB 15,89,29,240,27,0,0 ; mulps 0x1bf0(%rip),%xmm3 # 65a0 <_sk_callback_sse2+0xf3d> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -24051,9 +24095,9 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 76,139,112,8 ; mov 0x8(%rax),%r14 DB 72,255,203 ; dec %rbx - DB 120,7 ; js 4dbe <_sk_evenly_spaced_gradient_sse2+0x18> + DB 120,7 ; js 4dec <_sk_evenly_spaced_gradient_sse2+0x18> DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 - DB 235,21 ; jmp 4dd3 <_sk_evenly_spaced_gradient_sse2+0x2d> + DB 235,21 ; jmp 4e01 <_sk_evenly_spaced_gradient_sse2+0x2d> DB 73,137,217 ; mov %rbx,%r9 DB 73,209,233 ; shr %r9 DB 131,227,1 ; and $0x1,%ebx @@ -24149,15 +24193,15 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC PUBLIC _sk_gauss_a_to_rgba_sse2 _sk_gauss_a_to_rgba_sse2 LABEL PROC - DB 15,40,5,33,22,0,0 ; movaps 0x1621(%rip),%xmm0 # 6580 <_sk_callback_sse2+0xf4b> + DB 15,40,5,35,22,0,0 ; movaps 0x1623(%rip),%xmm0 # 65b0 <_sk_callback_sse2+0xf4d> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,39,22,0,0 ; addps 0x1627(%rip),%xmm0 # 6590 <_sk_callback_sse2+0xf5b> + DB 15,88,5,41,22,0,0 ; addps 0x1629(%rip),%xmm0 # 65c0 <_sk_callback_sse2+0xf5d> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,45,22,0,0 ; addps 0x162d(%rip),%xmm0 # 65a0 <_sk_callback_sse2+0xf6b> + DB 15,88,5,47,22,0,0 ; addps 0x162f(%rip),%xmm0 # 65d0 <_sk_callback_sse2+0xf6d> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,51,22,0,0 ; addps 0x1633(%rip),%xmm0 # 65b0 <_sk_callback_sse2+0xf7b> + DB 15,88,5,53,22,0,0 ; addps 0x1635(%rip),%xmm0 # 65e0 <_sk_callback_sse2+0xf7d> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,57,22,0,0 ; addps 0x1639(%rip),%xmm0 # 65c0 <_sk_callback_sse2+0xf8b> + DB 15,88,5,59,22,0,0 ; addps 0x163b(%rip),%xmm0 # 65f0 <_sk_callback_sse2+0xf8d> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -24173,12 +24217,12 @@ _sk_gradient_sse2 LABEL PROC DB 76,139,8 ; mov (%rax),%r9 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 73,131,249,2 ; cmp $0x2,%r9 - DB 114,50 ; jb 4fda <_sk_gradient_sse2+0x46> + DB 114,50 ; jb 5008 <_sk_gradient_sse2+0x46> DB 72,139,88,72 ; mov 0x48(%rax),%rbx DB 73,255,201 ; dec %r9 DB 72,131,195,4 ; add $0x4,%rbx DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,21,18,22,0,0 ; movaps 0x1612(%rip),%xmm2 # 65d0 <_sk_callback_sse2+0xf9b> + DB 15,40,21,20,22,0,0 ; movaps 0x1614(%rip),%xmm2 # 6600 <_sk_callback_sse2+0xf9d> DB 243,15,16,27 ; movss (%rbx),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 @@ -24186,7 +24230,7 @@ _sk_gradient_sse2 LABEL PROC DB 102,15,254,203 ; paddd %xmm3,%xmm1 DB 72,131,195,4 ; add $0x4,%rbx DB 73,255,201 ; dec %r9 - DB 117,228 ; jne 4fbe <_sk_gradient_sse2+0x2a> + DB 117,228 ; jne 4fec <_sk_gradient_sse2+0x2a> DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 DB 102,73,15,126,211 ; movq %xmm2,%r11 DB 69,137,217 ; mov %r11d,%r9d @@ -24321,29 +24365,29 @@ _sk_xy_to_unit_angle_sse2 LABEL PROC DB 69,15,94,220 ; divps %xmm12,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 - DB 68,15,40,45,211,19,0,0 ; movaps 0x13d3(%rip),%xmm13 # 65e0 <_sk_callback_sse2+0xfab> + DB 68,15,40,45,213,19,0,0 ; movaps 0x13d5(%rip),%xmm13 # 6610 <_sk_callback_sse2+0xfad> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,215,19,0,0 ; addps 0x13d7(%rip),%xmm13 # 65f0 <_sk_callback_sse2+0xfbb> + DB 68,15,88,45,217,19,0,0 ; addps 0x13d9(%rip),%xmm13 # 6620 <_sk_callback_sse2+0xfbd> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,219,19,0,0 ; addps 0x13db(%rip),%xmm13 # 6600 <_sk_callback_sse2+0xfcb> + DB 68,15,88,45,221,19,0,0 ; addps 0x13dd(%rip),%xmm13 # 6630 <_sk_callback_sse2+0xfcd> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,223,19,0,0 ; addps 0x13df(%rip),%xmm13 # 6610 <_sk_callback_sse2+0xfdb> + DB 68,15,88,45,225,19,0,0 ; addps 0x13e1(%rip),%xmm13 # 6640 <_sk_callback_sse2+0xfdd> DB 69,15,89,235 ; mulps %xmm11,%xmm13 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,40,21,222,19,0,0 ; movaps 0x13de(%rip),%xmm10 # 6620 <_sk_callback_sse2+0xfeb> + DB 68,15,40,21,224,19,0,0 ; movaps 0x13e0(%rip),%xmm10 # 6650 <_sk_callback_sse2+0xfed> DB 69,15,92,213 ; subps %xmm13,%xmm10 DB 69,15,84,209 ; andps %xmm9,%xmm10 DB 69,15,85,205 ; andnps %xmm13,%xmm9 DB 69,15,86,202 ; orps %xmm10,%xmm9 DB 68,15,194,192,1 ; cmpltps %xmm0,%xmm8 - DB 68,15,40,21,209,19,0,0 ; movaps 0x13d1(%rip),%xmm10 # 6630 <_sk_callback_sse2+0xffb> + DB 68,15,40,21,211,19,0,0 ; movaps 0x13d3(%rip),%xmm10 # 6660 <_sk_callback_sse2+0xffd> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,84,208 ; andps %xmm8,%xmm10 DB 69,15,85,193 ; andnps %xmm9,%xmm8 DB 69,15,86,194 ; orps %xmm10,%xmm8 DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 68,15,194,200,1 ; cmpltps %xmm0,%xmm9 - DB 68,15,40,21,192,19,0,0 ; movaps 0x13c0(%rip),%xmm10 # 6640 <_sk_callback_sse2+0x100b> + DB 68,15,40,21,194,19,0,0 ; movaps 0x13c2(%rip),%xmm10 # 6670 <_sk_callback_sse2+0x100d> DB 69,15,92,208 ; subps %xmm8,%xmm10 DB 69,15,84,209 ; andps %xmm9,%xmm10 DB 69,15,85,200 ; andnps %xmm8,%xmm9 @@ -24366,7 +24410,7 @@ _sk_xy_to_radius_sse2 LABEL PROC PUBLIC _sk_save_xy_sse2 _sk_save_xy_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,146,19,0,0 ; movaps 0x1392(%rip),%xmm8 # 6650 <_sk_callback_sse2+0x101b> + DB 68,15,40,5,148,19,0,0 ; movaps 0x1394(%rip),%xmm8 # 6680 <_sk_callback_sse2+0x101d> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -24374,7 +24418,7 @@ _sk_save_xy_sse2 LABEL PROC DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,40,217 ; movaps %xmm9,%xmm11 DB 69,15,194,218,1 ; cmpltps %xmm10,%xmm11 - DB 68,15,40,37,125,19,0,0 ; movaps 0x137d(%rip),%xmm12 # 6660 <_sk_callback_sse2+0x102b> + DB 68,15,40,37,127,19,0,0 ; movaps 0x137f(%rip),%xmm12 # 6690 <_sk_callback_sse2+0x102d> DB 69,15,84,220 ; andps %xmm12,%xmm11 DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 69,15,92,202 ; subps %xmm10,%xmm9 @@ -24417,8 +24461,8 @@ _sk_bilinear_nx_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,246,18,0,0 ; addps 0x12f6(%rip),%xmm0 # 6670 <_sk_callback_sse2+0x103b> - DB 68,15,40,13,254,18,0,0 ; movaps 0x12fe(%rip),%xmm9 # 6680 <_sk_callback_sse2+0x104b> + DB 15,88,5,248,18,0,0 ; addps 0x12f8(%rip),%xmm0 # 66a0 <_sk_callback_sse2+0x103d> + DB 68,15,40,13,0,19,0,0 ; movaps 0x1300(%rip),%xmm9 # 66b0 <_sk_callback_sse2+0x104d> 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 @@ -24429,7 +24473,7 @@ _sk_bilinear_px_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,237,18,0,0 ; addps 0x12ed(%rip),%xmm0 # 6690 <_sk_callback_sse2+0x105b> + DB 15,88,5,239,18,0,0 ; addps 0x12ef(%rip),%xmm0 # 66c0 <_sk_callback_sse2+0x105d> 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 @@ -24439,8 +24483,8 @@ _sk_bilinear_ny_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,223,18,0,0 ; addps 0x12df(%rip),%xmm1 # 66a0 <_sk_callback_sse2+0x106b> - DB 68,15,40,13,231,18,0,0 ; movaps 0x12e7(%rip),%xmm9 # 66b0 <_sk_callback_sse2+0x107b> + DB 15,88,13,225,18,0,0 ; addps 0x12e1(%rip),%xmm1 # 66d0 <_sk_callback_sse2+0x106d> + DB 68,15,40,13,233,18,0,0 ; movaps 0x12e9(%rip),%xmm9 # 66e0 <_sk_callback_sse2+0x107d> 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 @@ -24451,7 +24495,7 @@ _sk_bilinear_py_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,213,18,0,0 ; addps 0x12d5(%rip),%xmm1 # 66c0 <_sk_callback_sse2+0x108b> + DB 15,88,13,215,18,0,0 ; addps 0x12d7(%rip),%xmm1 # 66f0 <_sk_callback_sse2+0x108d> 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 @@ -24461,13 +24505,13 @@ _sk_bicubic_n3x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,200,18,0,0 ; addps 0x12c8(%rip),%xmm0 # 66d0 <_sk_callback_sse2+0x109b> - DB 68,15,40,13,208,18,0,0 ; movaps 0x12d0(%rip),%xmm9 # 66e0 <_sk_callback_sse2+0x10ab> + DB 15,88,5,202,18,0,0 ; addps 0x12ca(%rip),%xmm0 # 6700 <_sk_callback_sse2+0x109d> + DB 68,15,40,13,210,18,0,0 ; movaps 0x12d2(%rip),%xmm9 # 6710 <_sk_callback_sse2+0x10ad> 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 68,15,89,13,204,18,0,0 ; mulps 0x12cc(%rip),%xmm9 # 66f0 <_sk_callback_sse2+0x10bb> - DB 68,15,88,13,212,18,0,0 ; addps 0x12d4(%rip),%xmm9 # 6700 <_sk_callback_sse2+0x10cb> + DB 68,15,89,13,206,18,0,0 ; mulps 0x12ce(%rip),%xmm9 # 6720 <_sk_callback_sse2+0x10bd> + DB 68,15,88,13,214,18,0,0 ; addps 0x12d6(%rip),%xmm9 # 6730 <_sk_callback_sse2+0x10cd> 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 @@ -24478,16 +24522,16 @@ _sk_bicubic_n1x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,195,18,0,0 ; addps 0x12c3(%rip),%xmm0 # 6710 <_sk_callback_sse2+0x10db> - DB 68,15,40,13,203,18,0,0 ; movaps 0x12cb(%rip),%xmm9 # 6720 <_sk_callback_sse2+0x10eb> + DB 15,88,5,197,18,0,0 ; addps 0x12c5(%rip),%xmm0 # 6740 <_sk_callback_sse2+0x10dd> + DB 68,15,40,13,205,18,0,0 ; movaps 0x12cd(%rip),%xmm9 # 6750 <_sk_callback_sse2+0x10ed> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,207,18,0,0 ; movaps 0x12cf(%rip),%xmm8 # 6730 <_sk_callback_sse2+0x10fb> + DB 68,15,40,5,209,18,0,0 ; movaps 0x12d1(%rip),%xmm8 # 6760 <_sk_callback_sse2+0x10fd> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,211,18,0,0 ; addps 0x12d3(%rip),%xmm8 # 6740 <_sk_callback_sse2+0x110b> + DB 68,15,88,5,213,18,0,0 ; addps 0x12d5(%rip),%xmm8 # 6770 <_sk_callback_sse2+0x110d> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,215,18,0,0 ; addps 0x12d7(%rip),%xmm8 # 6750 <_sk_callback_sse2+0x111b> + DB 68,15,88,5,217,18,0,0 ; addps 0x12d9(%rip),%xmm8 # 6780 <_sk_callback_sse2+0x111d> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,219,18,0,0 ; addps 0x12db(%rip),%xmm8 # 6760 <_sk_callback_sse2+0x112b> + DB 68,15,88,5,221,18,0,0 ; addps 0x12dd(%rip),%xmm8 # 6790 <_sk_callback_sse2+0x112d> 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 @@ -24495,17 +24539,17 @@ _sk_bicubic_n1x_sse2 LABEL PROC PUBLIC _sk_bicubic_p1x_sse2 _sk_bicubic_p1x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,213,18,0,0 ; movaps 0x12d5(%rip),%xmm8 # 6770 <_sk_callback_sse2+0x113b> + DB 68,15,40,5,215,18,0,0 ; movaps 0x12d7(%rip),%xmm8 # 67a0 <_sk_callback_sse2+0x113d> 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 68,15,40,21,209,18,0,0 ; movaps 0x12d1(%rip),%xmm10 # 6780 <_sk_callback_sse2+0x114b> + DB 68,15,40,21,211,18,0,0 ; movaps 0x12d3(%rip),%xmm10 # 67b0 <_sk_callback_sse2+0x114d> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,213,18,0,0 ; addps 0x12d5(%rip),%xmm10 # 6790 <_sk_callback_sse2+0x115b> + DB 68,15,88,21,215,18,0,0 ; addps 0x12d7(%rip),%xmm10 # 67c0 <_sk_callback_sse2+0x115d> 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 68,15,88,21,209,18,0,0 ; addps 0x12d1(%rip),%xmm10 # 67a0 <_sk_callback_sse2+0x116b> + DB 68,15,88,21,211,18,0,0 ; addps 0x12d3(%rip),%xmm10 # 67d0 <_sk_callback_sse2+0x116d> 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 @@ -24515,11 +24559,11 @@ _sk_bicubic_p3x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,196,18,0,0 ; addps 0x12c4(%rip),%xmm0 # 67b0 <_sk_callback_sse2+0x117b> + DB 15,88,5,198,18,0,0 ; addps 0x12c6(%rip),%xmm0 # 67e0 <_sk_callback_sse2+0x117d> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,196,18,0,0 ; mulps 0x12c4(%rip),%xmm8 # 67c0 <_sk_callback_sse2+0x118b> - DB 68,15,88,5,204,18,0,0 ; addps 0x12cc(%rip),%xmm8 # 67d0 <_sk_callback_sse2+0x119b> + DB 68,15,89,5,198,18,0,0 ; mulps 0x12c6(%rip),%xmm8 # 67f0 <_sk_callback_sse2+0x118d> + DB 68,15,88,5,206,18,0,0 ; addps 0x12ce(%rip),%xmm8 # 6800 <_sk_callback_sse2+0x119d> 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 @@ -24530,13 +24574,13 @@ _sk_bicubic_n3y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,186,18,0,0 ; addps 0x12ba(%rip),%xmm1 # 67e0 <_sk_callback_sse2+0x11ab> - DB 68,15,40,13,194,18,0,0 ; movaps 0x12c2(%rip),%xmm9 # 67f0 <_sk_callback_sse2+0x11bb> + DB 15,88,13,188,18,0,0 ; addps 0x12bc(%rip),%xmm1 # 6810 <_sk_callback_sse2+0x11ad> + DB 68,15,40,13,196,18,0,0 ; movaps 0x12c4(%rip),%xmm9 # 6820 <_sk_callback_sse2+0x11bd> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,190,18,0,0 ; mulps 0x12be(%rip),%xmm9 # 6800 <_sk_callback_sse2+0x11cb> - DB 68,15,88,13,198,18,0,0 ; addps 0x12c6(%rip),%xmm9 # 6810 <_sk_callback_sse2+0x11db> + DB 68,15,89,13,192,18,0,0 ; mulps 0x12c0(%rip),%xmm9 # 6830 <_sk_callback_sse2+0x11cd> + DB 68,15,88,13,200,18,0,0 ; addps 0x12c8(%rip),%xmm9 # 6840 <_sk_callback_sse2+0x11dd> 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 @@ -24547,16 +24591,16 @@ _sk_bicubic_n1y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,180,18,0,0 ; addps 0x12b4(%rip),%xmm1 # 6820 <_sk_callback_sse2+0x11eb> - DB 68,15,40,13,188,18,0,0 ; movaps 0x12bc(%rip),%xmm9 # 6830 <_sk_callback_sse2+0x11fb> + DB 15,88,13,182,18,0,0 ; addps 0x12b6(%rip),%xmm1 # 6850 <_sk_callback_sse2+0x11ed> + DB 68,15,40,13,190,18,0,0 ; movaps 0x12be(%rip),%xmm9 # 6860 <_sk_callback_sse2+0x11fd> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,192,18,0,0 ; movaps 0x12c0(%rip),%xmm8 # 6840 <_sk_callback_sse2+0x120b> + DB 68,15,40,5,194,18,0,0 ; movaps 0x12c2(%rip),%xmm8 # 6870 <_sk_callback_sse2+0x120d> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,196,18,0,0 ; addps 0x12c4(%rip),%xmm8 # 6850 <_sk_callback_sse2+0x121b> + DB 68,15,88,5,198,18,0,0 ; addps 0x12c6(%rip),%xmm8 # 6880 <_sk_callback_sse2+0x121d> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,200,18,0,0 ; addps 0x12c8(%rip),%xmm8 # 6860 <_sk_callback_sse2+0x122b> + DB 68,15,88,5,202,18,0,0 ; addps 0x12ca(%rip),%xmm8 # 6890 <_sk_callback_sse2+0x122d> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,204,18,0,0 ; addps 0x12cc(%rip),%xmm8 # 6870 <_sk_callback_sse2+0x123b> + DB 68,15,88,5,206,18,0,0 ; addps 0x12ce(%rip),%xmm8 # 68a0 <_sk_callback_sse2+0x123d> 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 @@ -24564,17 +24608,17 @@ _sk_bicubic_n1y_sse2 LABEL PROC PUBLIC _sk_bicubic_p1y_sse2 _sk_bicubic_p1y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,198,18,0,0 ; movaps 0x12c6(%rip),%xmm8 # 6880 <_sk_callback_sse2+0x124b> + DB 68,15,40,5,200,18,0,0 ; movaps 0x12c8(%rip),%xmm8 # 68b0 <_sk_callback_sse2+0x124d> 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 68,15,40,21,193,18,0,0 ; movaps 0x12c1(%rip),%xmm10 # 6890 <_sk_callback_sse2+0x125b> + DB 68,15,40,21,195,18,0,0 ; movaps 0x12c3(%rip),%xmm10 # 68c0 <_sk_callback_sse2+0x125d> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,197,18,0,0 ; addps 0x12c5(%rip),%xmm10 # 68a0 <_sk_callback_sse2+0x126b> + DB 68,15,88,21,199,18,0,0 ; addps 0x12c7(%rip),%xmm10 # 68d0 <_sk_callback_sse2+0x126d> 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 68,15,88,21,193,18,0,0 ; addps 0x12c1(%rip),%xmm10 # 68b0 <_sk_callback_sse2+0x127b> + DB 68,15,88,21,195,18,0,0 ; addps 0x12c3(%rip),%xmm10 # 68e0 <_sk_callback_sse2+0x127d> 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 @@ -24584,11 +24628,11 @@ _sk_bicubic_p3y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,179,18,0,0 ; addps 0x12b3(%rip),%xmm1 # 68c0 <_sk_callback_sse2+0x128b> + DB 15,88,13,181,18,0,0 ; addps 0x12b5(%rip),%xmm1 # 68f0 <_sk_callback_sse2+0x128d> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,179,18,0,0 ; mulps 0x12b3(%rip),%xmm8 # 68d0 <_sk_callback_sse2+0x129b> - DB 68,15,88,5,187,18,0,0 ; addps 0x12bb(%rip),%xmm8 # 68e0 <_sk_callback_sse2+0x12ab> + DB 68,15,89,5,181,18,0,0 ; mulps 0x12b5(%rip),%xmm8 # 6900 <_sk_callback_sse2+0x129d> + DB 68,15,88,5,189,18,0,0 ; addps 0x12bd(%rip),%xmm8 # 6910 <_sk_callback_sse2+0x12ad> 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 @@ -24806,11 +24850,11 @@ ALIGN 16 DB 128,191,0,0,128,191,0 ; cmpb $0x0,-0x40800000(%rdi) DB 0,224 ; add %ah,%al DB 64,0,0 ; add %al,(%rax) - DB 224,64 ; loopne 58d8 <.literal16+0x1d8> + DB 224,64 ; loopne 5908 <.literal16+0x1d8> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 58dc <.literal16+0x1dc> + DB 224,64 ; loopne 590c <.literal16+0x1dc> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 58e0 <.literal16+0x1e0> + DB 224,64 ; loopne 5910 <.literal16+0x1e0> DB 154 ; (bad) DB 153 ; cltd DB 153 ; cltd @@ -24830,13 +24874,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5901 <.literal16+0x201> + DB 71,225,61 ; rex.RXB loope 5931 <.literal16+0x201> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5905 <.literal16+0x205> + DB 71,225,61 ; rex.RXB loope 5935 <.literal16+0x205> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5909 <.literal16+0x209> + DB 71,225,61 ; rex.RXB loope 5939 <.literal16+0x209> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 590d <.literal16+0x20d> + DB 71,225,61 ; rex.RXB loope 593d <.literal16+0x20d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24861,13 +24905,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5941 <.literal16+0x241> + DB 71,225,61 ; rex.RXB loope 5971 <.literal16+0x241> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5945 <.literal16+0x245> + DB 71,225,61 ; rex.RXB loope 5975 <.literal16+0x245> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5949 <.literal16+0x249> + DB 71,225,61 ; rex.RXB loope 5979 <.literal16+0x249> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 594d <.literal16+0x24d> + DB 71,225,61 ; rex.RXB loope 597d <.literal16+0x24d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24892,13 +24936,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5981 <.literal16+0x281> + DB 71,225,61 ; rex.RXB loope 59b1 <.literal16+0x281> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5985 <.literal16+0x285> + DB 71,225,61 ; rex.RXB loope 59b5 <.literal16+0x285> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 5989 <.literal16+0x289> + DB 71,225,61 ; rex.RXB loope 59b9 <.literal16+0x289> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 598d <.literal16+0x28d> + DB 71,225,61 ; rex.RXB loope 59bd <.literal16+0x28d> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24923,13 +24967,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 59c1 <.literal16+0x2c1> + DB 71,225,61 ; rex.RXB loope 59f1 <.literal16+0x2c1> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 59c5 <.literal16+0x2c5> + DB 71,225,61 ; rex.RXB loope 59f5 <.literal16+0x2c5> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 59c9 <.literal16+0x2c9> + DB 71,225,61 ; rex.RXB loope 59f9 <.literal16+0x2c9> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 59cd <.literal16+0x2cd> + DB 71,225,61 ; rex.RXB loope 59fd <.literal16+0x2cd> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -24950,11 +24994,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) DB 67,0,0 ; rex.XB add %al,(%r8) - DB 127,67 ; jg 5a0b <.literal16+0x30b> + DB 127,67 ; jg 5a3b <.literal16+0x30b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5a0f <.literal16+0x30f> + DB 127,67 ; jg 5a3f <.literal16+0x30f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5a13 <.literal16+0x313> + DB 127,67 ; jg 5a43 <.literal16+0x313> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -25194,13 +25238,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 5bf9 <.literal16+0x4f9> + DB 224,7 ; loopne 5c29 <.literal16+0x4f9> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5bfd <.literal16+0x4fd> + DB 224,7 ; loopne 5c2d <.literal16+0x4fd> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c01 <.literal16+0x501> + DB 224,7 ; loopne 5c31 <.literal16+0x501> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 5c05 <.literal16+0x505> + DB 224,7 ; loopne 5c35 <.literal16+0x505> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -25265,11 +25309,11 @@ ALIGN 16 DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5cdb <.literal16+0x5db> + DB 127,67 ; jg 5d0b <.literal16+0x5db> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5cdf <.literal16+0x5df> + DB 127,67 ; jg 5d0f <.literal16+0x5df> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 5ce3 <.literal16+0x5e3> + DB 127,67 ; jg 5d13 <.literal16+0x5e3> DB 129,128,128,59,129,128,128,59,129,128; addl $0x80813b80,-0x7f7ec480(%rax) DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,129,128,128,59 ; addb $0x3b,-0x7f7f7ec5(%rax) @@ -25284,16 +25328,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5cd4 <.literal16+0x5d4> + DB 127,0 ; jg 5d04 <.literal16+0x5d4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5cd8 <.literal16+0x5d8> + DB 127,0 ; jg 5d08 <.literal16+0x5d8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5cdc <.literal16+0x5dc> + DB 127,0 ; jg 5d0c <.literal16+0x5dc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5ce0 <.literal16+0x5e0> + DB 127,0 ; jg 5d10 <.literal16+0x5e0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25302,7 +25346,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5d65 <.literal16+0x665> + DB 119,115 ; ja 5d95 <.literal16+0x665> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25313,7 +25357,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5cc9 <.literal16+0x5c9> + DB 117,191 ; jne 5cf9 <.literal16+0x5c9> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25325,7 +25369,7 @@ 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 ffffffffe9a39d0a <_sk_callback_sse2+0xffffffffe9a346d5> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39d3a <_sk_callback_sse2+0xffffffffe9a346d7> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25379,16 +25423,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5da4 <.literal16+0x6a4> + DB 127,0 ; jg 5dd4 <.literal16+0x6a4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5da8 <.literal16+0x6a8> + DB 127,0 ; jg 5dd8 <.literal16+0x6a8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5dac <.literal16+0x6ac> + DB 127,0 ; jg 5ddc <.literal16+0x6ac> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5db0 <.literal16+0x6b0> + DB 127,0 ; jg 5de0 <.literal16+0x6b0> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25397,7 +25441,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5e35 <.literal16+0x735> + DB 119,115 ; ja 5e65 <.literal16+0x735> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25408,7 +25452,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5d99 <.literal16+0x699> + DB 117,191 ; jne 5dc9 <.literal16+0x699> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25420,7 +25464,7 @@ 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 ffffffffe9a39dda <_sk_callback_sse2+0xffffffffe9a347a5> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39e0a <_sk_callback_sse2+0xffffffffe9a347a7> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25474,16 +25518,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5e74 <.literal16+0x774> + DB 127,0 ; jg 5ea4 <.literal16+0x774> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5e78 <.literal16+0x778> + DB 127,0 ; jg 5ea8 <.literal16+0x778> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5e7c <.literal16+0x77c> + DB 127,0 ; jg 5eac <.literal16+0x77c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5e80 <.literal16+0x780> + DB 127,0 ; jg 5eb0 <.literal16+0x780> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25492,7 +25536,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5f05 <.literal16+0x805> + DB 119,115 ; ja 5f35 <.literal16+0x805> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25503,7 +25547,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5e69 <.literal16+0x769> + DB 117,191 ; jne 5e99 <.literal16+0x769> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25515,7 +25559,7 @@ 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 ffffffffe9a39eaa <_sk_callback_sse2+0xffffffffe9a34875> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39eda <_sk_callback_sse2+0xffffffffe9a34877> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25569,16 +25613,16 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 5f44 <.literal16+0x844> + DB 127,0 ; jg 5f74 <.literal16+0x844> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5f48 <.literal16+0x848> + DB 127,0 ; jg 5f78 <.literal16+0x848> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5f4c <.literal16+0x84c> + DB 127,0 ; jg 5f7c <.literal16+0x84c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 5f50 <.literal16+0x850> + DB 127,0 ; jg 5f80 <.literal16+0x850> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -25587,7 +25631,7 @@ ALIGN 16 DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 5fd5 <.literal16+0x8d5> + DB 119,115 ; ja 6005 <.literal16+0x8d5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -25598,7 +25642,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 5f39 <.literal16+0x839> + DB 117,191 ; jne 5f69 <.literal16+0x839> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -25610,7 +25654,7 @@ 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 ffffffffe9a39f7a <_sk_callback_sse2+0xffffffffe9a34945> + DB 233,220,63,163,233 ; jmpq ffffffffe9a39faa <_sk_callback_sse2+0xffffffffe9a34947> DB 220,63 ; fdivrl (%rdi) DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) @@ -25660,13 +25704,13 @@ ALIGN 16 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 DB 200,66,0,0 ; enterq $0x42,$0x0 - DB 127,67 ; jg 6057 <.literal16+0x957> + DB 127,67 ; jg 6087 <.literal16+0x957> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 605b <.literal16+0x95b> + DB 127,67 ; jg 608b <.literal16+0x95b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 605f <.literal16+0x95f> + DB 127,67 ; jg 608f <.literal16+0x95f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 6063 <.literal16+0x963> + DB 127,67 ; jg 6093 <.literal16+0x963> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -25713,16 +25757,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 60e3 <.literal16+0x9e3> + DB 118,63 ; jbe 6113 <.literal16+0x9e3> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 60e7 <.literal16+0x9e7> + DB 118,63 ; jbe 6117 <.literal16+0x9e7> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 60eb <.literal16+0x9eb> + DB 118,63 ; jbe 611b <.literal16+0x9eb> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 60ef <.literal16+0x9ef> + DB 118,63 ; jbe 611f <.literal16+0x9ef> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) @@ -25742,11 +25786,11 @@ ALIGN 16 DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 613b <.literal16+0xa3b> + DB 127,67 ; jg 616b <.literal16+0xa3b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 613f <.literal16+0xa3f> + DB 127,67 ; jg 616f <.literal16+0xa3f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 6143 <.literal16+0xa43> + DB 127,67 ; jg 6173 <.literal16+0xa43> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -25794,13 +25838,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6199 <.literal16+0xa99> + DB 224,7 ; loopne 61c9 <.literal16+0xa99> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 619d <.literal16+0xa9d> + DB 224,7 ; loopne 61cd <.literal16+0xa9d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 61a1 <.literal16+0xaa1> + DB 224,7 ; loopne 61d1 <.literal16+0xaa1> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 61a5 <.literal16+0xaa5> + DB 224,7 ; loopne 61d5 <.literal16+0xaa5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -25846,13 +25890,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6209 <.literal16+0xb09> + DB 224,7 ; loopne 6239 <.literal16+0xb09> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 620d <.literal16+0xb0d> + DB 224,7 ; loopne 623d <.literal16+0xb0d> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 6211 <.literal16+0xb11> + DB 224,7 ; loopne 6241 <.literal16+0xb11> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 6215 <.literal16+0xb15> + DB 224,7 ; loopne 6245 <.literal16+0xb15> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -25890,13 +25934,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 62a6 <.literal16+0xba6> + DB 124,66 ; jl 62d6 <.literal16+0xba6> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 62aa <.literal16+0xbaa> + DB 124,66 ; jl 62da <.literal16+0xbaa> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 62ae <.literal16+0xbae> + DB 124,66 ; jl 62de <.literal16+0xbae> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 62b2 <.literal16+0xbb2> + DB 124,66 ; jl 62e2 <.literal16+0xbb2> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -25986,13 +26030,13 @@ ALIGN 16 DB 136,136,61,137,136,136 ; mov %cl,-0x777776c3(%rax) DB 61,137,136,136,61 ; cmp $0x3d888889,%eax DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 63b5 <.literal16+0xcb5> + DB 112,65 ; jo 63e5 <.literal16+0xcb5> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 63b9 <.literal16+0xcb9> + DB 112,65 ; jo 63e9 <.literal16+0xcb9> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 63bd <.literal16+0xcbd> + DB 112,65 ; jo 63ed <.literal16+0xcbd> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 63c1 <.literal16+0xcc1> + DB 112,65 ; jo 63f1 <.literal16+0xcc1> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -26014,11 +26058,11 @@ ALIGN 16 DB 128,59,129 ; cmpb $0x81,(%rbx) DB 128,128,59,0,0,127,67 ; addb $0x43,0x7f00003b(%rax) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 640b <.literal16+0xd0b> + DB 127,67 ; jg 643b <.literal16+0xd0b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 640f <.literal16+0xd0f> + DB 127,67 ; jg 643f <.literal16+0xd0f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 6413 <.literal16+0xd13> + DB 127,67 ; jg 6443 <.literal16+0xd13> DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) @@ -26094,13 +26138,13 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 255 ; (bad) - DB 127,71 ; jg 64fb <.literal16+0xdfb> + DB 127,71 ; jg 652b <.literal16+0xdfb> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 64ff <.literal16+0xdff> + DB 127,71 ; jg 652f <.literal16+0xdff> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6503 <.literal16+0xe03> + DB 127,71 ; jg 6533 <.literal16+0xe03> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6507 <.literal16+0xe07> + DB 127,71 ; jg 6537 <.literal16+0xe07> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax) @@ -26177,10 +26221,10 @@ ALIGN 16 DB 61,152,221,147,61 ; cmp $0x3d93dd98,%eax DB 152 ; cwtl DB 221,147,61,45,16,17 ; fstl 0x11102d3d(%rbx) - DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc0769a <_sk_callback_sse2+0x2dc02065> + DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc076ca <_sk_callback_sse2+0x2dc02067> DB 17,192 ; adc %eax,%eax DB 45,16,17,192,18 ; sub $0x12c01110,%eax - DB 120,57 ; js 65cc <.literal16+0xecc> + DB 120,57 ; js 65fc <.literal16+0xecc> DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil @@ -26308,11 +26352,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6712 <.literal16+0x1012> + DB 62,114,28 ; jb,pt 6742 <.literal16+0x1012> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6716 <.literal16+0x1016> + DB 62,114,28 ; jb,pt 6746 <.literal16+0x1016> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 671a <.literal16+0x101a> + DB 62,114,28 ; jb,pt 674a <.literal16+0x101a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26356,7 +26400,7 @@ 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 # 3d63f5a5 <_sk_callback_sse2+0x3d639f70> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f5d5 <_sk_callback_sse2+0x3d639f72> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -26382,7 +26426,7 @@ 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 # 3d63f5e5 <_sk_callback_sse2+0x3d639fb0> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f615 <_sk_callback_sse2+0x3d639fb2> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -26391,13 +26435,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 67de <.literal16+0x10de> + DB 114,28 ; jb 680e <.literal16+0x10de> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 67e2 <.literal16+0x10e2> + DB 62,114,28 ; jb,pt 6812 <.literal16+0x10e2> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 67e6 <.literal16+0x10e6> + DB 62,114,28 ; jb,pt 6816 <.literal16+0x10e6> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 67ea <.literal16+0x10ea> + DB 62,114,28 ; jb,pt 681a <.literal16+0x10ea> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26418,11 +26462,11 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 128,63,114 ; cmpb $0x72,(%rdi) DB 28,199 ; sbb $0xc7,%al - DB 62,114,28 ; jb,pt 6822 <.literal16+0x1122> + DB 62,114,28 ; jb,pt 6852 <.literal16+0x1122> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 6826 <.literal16+0x1126> + DB 62,114,28 ; jb,pt 6856 <.literal16+0x1126> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 682a <.literal16+0x112a> + DB 62,114,28 ; jb,pt 685a <.literal16+0x112a> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26466,7 +26510,7 @@ 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 # 3d63f6b5 <_sk_callback_sse2+0x3d63a080> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f6e5 <_sk_callback_sse2+0x3d63a082> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 0,63 ; add %bh,(%rdi) DB 0,0 ; add %al,(%rax) @@ -26492,7 +26536,7 @@ 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 # 3d63f6f5 <_sk_callback_sse2+0x3d63a0c0> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d63f725 <_sk_callback_sse2+0x3d63a0c2> DB 57,142,99,61,0,0 ; cmp %ecx,0x3d63(%rsi) DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al @@ -26501,13 +26545,13 @@ ALIGN 16 DB 192,63,0 ; sarb $0x0,(%rdi) DB 0,192 ; add %al,%al DB 63 ; (bad) - DB 114,28 ; jb 68ee <.literal16+0x11ee> + DB 114,28 ; jb 691e <.literal16+0x11ee> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 68f2 <_sk_callback_sse2+0x12bd> + DB 62,114,28 ; jb,pt 6922 <_sk_callback_sse2+0x12bf> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 68f6 <_sk_callback_sse2+0x12c1> + DB 62,114,28 ; jb,pt 6926 <_sk_callback_sse2+0x12c3> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 68fa <_sk_callback_sse2+0x12c5> + DB 62,114,28 ; jb,pt 692a <_sk_callback_sse2+0x12c7> DB 199 ; (bad) DB 62,171 ; ds stos %eax,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) @@ -26517,6 +26561,12 @@ ALIGN 16 DB 170 ; stos %al,%es:(%rdi) DB 170 ; stos %al,%es:(%rdi) DB 190,171,170,170,190 ; mov $0xbeaaaaab,%esi + +ALIGN 4 + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) + DB 0,0 ; add %al,(%rax) + DB 0,63 ; add %bh,(%rdi) ALIGN 32 PUBLIC _sk_start_pipeline_ssse3_lowp diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp index 44af125449..5993ec03db 100644 --- a/src/jumper/SkJumper_stages.cpp +++ b/src/jumper/SkJumper_stages.cpp @@ -945,25 +945,27 @@ SI F ulp_before(F f) { return unaligned_load(&bits); } -SI F exclusive_clamp(F v, float limit) { +SI F exclusive_clamp(F v, const SkJumper_TileCtx* ctx) { v = max(0,v); - return min(v, ulp_before(limit)); + return min(v, ulp_before(ctx->scale)); } -SI F exclusive_repeat(F v, float limit) { - v = v - floor_(v/limit)*limit; - return min(v, ulp_before(limit)); +SI F exclusive_repeat(F v, const SkJumper_TileCtx* ctx) { + v = v - floor_(v*ctx->invScale)*ctx->scale; + return min(v, ulp_before(ctx->scale)); } -SI F exclusive_mirror(F v, float limit) { - v = abs_( (v-limit) - (limit+limit)*floor_((v-limit)/(limit+limit)) - limit ); +SI F exclusive_mirror(F v, const SkJumper_TileCtx* ctx) { + auto limit = ctx->scale; + auto invLimit = ctx->invScale; + v = abs_( (v-limit) - (limit+limit)*floor_((v-limit)*(invLimit*0.5f)) - limit ); return min(v, ulp_before(limit)); } // Clamp x or y to [0,limit) == [0,limit - 1 ulp] (think, sampling from images). -STAGE(clamp_x) { r = exclusive_clamp (r, *(const float*)ctx); } -STAGE(clamp_y) { g = exclusive_clamp (g, *(const float*)ctx); } -STAGE(repeat_x) { r = exclusive_repeat(r, *(const float*)ctx); } -STAGE(repeat_y) { g = exclusive_repeat(g, *(const float*)ctx); } -STAGE(mirror_x) { r = exclusive_mirror(r, *(const float*)ctx); } -STAGE(mirror_y) { g = exclusive_mirror(g, *(const float*)ctx); } +STAGE(clamp_x) { r = exclusive_clamp (r, (const SkJumper_TileCtx*)ctx); } +STAGE(clamp_y) { g = exclusive_clamp (g, (const SkJumper_TileCtx*)ctx); } +STAGE(repeat_x) { r = exclusive_repeat(r, (const SkJumper_TileCtx*)ctx); } +STAGE(repeat_y) { g = exclusive_repeat(g, (const SkJumper_TileCtx*)ctx); } +STAGE(mirror_x) { r = exclusive_mirror(r, (const SkJumper_TileCtx*)ctx); } +STAGE(mirror_y) { g = exclusive_mirror(g, (const SkJumper_TileCtx*)ctx); } // Clamp x to [0,1], both sides exclusive (think, gradients). STAGE( clamp_x_1) { r = min(max(0, r), 1.0f); } diff --git a/src/shaders/SkImageShader.cpp b/src/shaders/SkImageShader.cpp index ed032fa65f..61d24837d1 100644 --- a/src/shaders/SkImageShader.cpp +++ b/src/shaders/SkImageShader.cpp @@ -298,8 +298,12 @@ bool SkImageShader::onAppendStages(SkRasterPipeline* p, SkColorSpace* dstCS, SkA gather->ctable = pm.ctable() ? pm.ctable()->readColors() : nullptr; gather->stride = pm.rowBytesAsPixels(); - auto limit_x = alloc->make(pm. width()), - limit_y = alloc->make(pm.height()); + auto limit_x = alloc->make(), + limit_y = alloc->make(); + limit_x->scale = pm.width(); + limit_x->invScale = 1.0f / pm.width(); + limit_y->scale = pm.height(); + limit_y->invScale = 1.0f / pm.height(); auto append_tiling_and_gather = [&] { switch (fTileModeX) {