From: Mike Klein Date: Wed, 19 Apr 2017 18:33:58 +0000 (-0400) Subject: refactor approx_{log2,pow2,powf} X-Git-Tag: submit/tizen/20180928.044319~36^2~697 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=da16434928b4c721b462d780d5140f9fc68d1413;p=platform%2Fupstream%2FlibSkiaSharp.git refactor approx_{log2,pow2,powf} - Move to SkJumper_vectors.h - Fold the -127 and +2.774485010. - approx_powf(F,F) instead of approx_powf(F,float) for consistency. - A little layout reformatting. Change-Id: If9cb3d62a097cb6ecf89f157a1dde672c1516371 Reviewed-on: https://skia-review.googlesource.com/13865 Reviewed-by: Mike Klein Commit-Queue: Mike Klein --- diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index f12e5e252d..9aa29d015e 100644 --- a/src/jumper/SkJumper_generated.S +++ b/src/jumper/SkJumper_generated.S @@ -1905,73 +1905,70 @@ _sk_parametric_r_aarch64: .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 .long 0x91004109 // add x9, x8, #0x10 .long 0x9100610a // add x10, x8, #0x18 - .long 0x4d40c933 // ld1r {v19.4s}, [x9] + .long 0x4d40c932 // ld1r {v18.4s}, [x9] .long 0xaa0803e9 // mov x9, x8 - .long 0xbd400d12 // ldr s18, [x8, #12] + .long 0xbd400d11 // ldr s17, [x8, #12] .long 0x4d40c950 // ld1r {v16.4s}, [x10] - .long 0x4ddfc931 // ld1r {v17.4s}, [x9], #4 + .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 .long 0x9100210a // add x10, x8, #0x8 .long 0x4d40c954 // ld1r {v20.4s}, [x10] - .long 0x4f921010 // fmla v16.4s, v0.4s, v18.s[0] + .long 0x4f911010 // fmla v16.4s, v0.4s, v17.s[0] .long 0xbd400135 // ldr s21, [x9] - .long 0x52b85fc9 // mov w9, #0xc2fe0000 + .long 0x52b85f09 // mov w9, #0xc2f80000 + .long 0x728e6ee9 // movk w9, #0x7377 .long 0x4e040d37 // dup v23.4s, w9 - .long 0x52a80629 // mov w9, #0x40310000 - .long 0x72922549 // movk w9, #0x912a - .long 0x4f951014 // fmla v20.4s, v0.4s, v21.s[0] - .long 0x6e20e660 // fcmge v0.4s, v19.4s, v0.4s - .long 0x4e040d33 // dup v19.4s, w9 .long 0x52a7f7e9 // mov w9, #0x3fbf0000 - .long 0x4f03d7f2 // movi v18.4s, #0x7f, msl #16 .long 0x7297eea9 // movk w9, #0xbf75 - .long 0x4e21da95 // scvtf v21.4s, v20.4s - .long 0x4e321e92 // and v18.16b, v20.16b, v18.16b - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f951014 // fmla v20.4s, v0.4s, v21.s[0] + .long 0x6e20e640 // fcmge v0.4s, v18.4s, v0.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a7d689 // mov w9, #0x3eb40000 + .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 .long 0x72889f29 // movk w9, #0x44f9 - .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e21da95 // scvtf v21.4s, v20.4s + .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7fb89 // mov w9, #0x3fdc0000 - .long 0x4e33d6f3 // fadd v19.4s, v23.4s, v19.4s + .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s .long 0x729d3469 // movk w9, #0xe9a3 - .long 0x4f0177f2 // orr v18.4s, #0x3f, lsl #24 - .long 0x4eb4ce53 // fmls v19.4s, v18.4s, v20.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 + .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a85e49 // mov w9, #0x42f20000 .long 0x72918a29 // movk w9, #0x8c51 - .long 0x4e35d652 // fadd v18.4s, v18.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7f7c9 // mov w9, #0x3fbe0000 .long 0x729791a9 // movk w9, #0xbc8d - .long 0x6e32fe92 // fdiv v18.4s, v20.4s, v18.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a81349 // mov w9, #0x409a0000 - .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s + .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s .long 0x729ebf09 // movk w9, #0xf5f8 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x52a83ba9 // mov w9, #0x41dd0000 - .long 0x4e219a32 // frintm v18.4s, v17.4s + .long 0x4e219a33 // frintm v19.4s, v17.4s .long 0x729a5fc9 // movk w9, #0xd2fe - .long 0x4e35d635 // fadd v21.4s, v17.4s, v21.4s - .long 0x4eb2d631 // fsub v17.4s, v17.4s, v18.4s - .long 0x4eb4ce35 // fmls v21.4s, v17.4s, v20.4s - .long 0x4eb1d671 // fsub v17.4s, v19.4s, v17.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s + .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s + .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s + .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x91005108 // add x8, x8, #0x14 - .long 0x6e31fe71 // fdiv v17.4s, v19.4s, v17.4s - .long 0x4e31d6b1 // fadd v17.4s, v21.4s, v17.4s - .long 0x4d40c915 // ld1r {v21.4s}, [x8] - .long 0x4f026572 // movi v18.4s, #0x4b, lsl #24 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s + .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s + .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s + .long 0x4d40c914 // ld1r {v20.4s}, [x8] + .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 + .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s .long 0xf9400423 // ldr x3, [x1, #8] - .long 0x4e35d631 // fadd v17.4s, v17.4s, v21.4s - .long 0x6f00e414 // movi v20.2d, #0x0 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x6f00e412 // movi v18.2d, #0x0 .long 0x6e711e00 // bsl v0.16b, v16.16b, v17.16b - .long 0x4f03f613 // fmov v19.4s, #1.000000000000000000e+00 - .long 0x4e34f400 // fmax v0.4s, v0.4s, v20.4s - .long 0x4eb3f400 // fmin v0.4s, v0.4s, v19.4s + .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 + .long 0x4e32f400 // fmax v0.4s, v0.4s, v18.4s + .long 0x4eb5f400 // fmin v0.4s, v0.4s, v21.4s .long 0x91004021 // add x1, x1, #0x10 .long 0xd61f0060 // br x3 @@ -1983,73 +1980,70 @@ _sk_parametric_g_aarch64: .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 .long 0x91004109 // add x9, x8, #0x10 .long 0x9100610a // add x10, x8, #0x18 - .long 0x4d40c933 // ld1r {v19.4s}, [x9] + .long 0x4d40c932 // ld1r {v18.4s}, [x9] .long 0xaa0803e9 // mov x9, x8 - .long 0xbd400d12 // ldr s18, [x8, #12] + .long 0xbd400d11 // ldr s17, [x8, #12] .long 0x4d40c950 // ld1r {v16.4s}, [x10] - .long 0x4ddfc931 // ld1r {v17.4s}, [x9], #4 + .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 .long 0x9100210a // add x10, x8, #0x8 .long 0x4d40c954 // ld1r {v20.4s}, [x10] - .long 0x4f921030 // fmla v16.4s, v1.4s, v18.s[0] + .long 0x4f911030 // fmla v16.4s, v1.4s, v17.s[0] .long 0xbd400135 // ldr s21, [x9] - .long 0x52b85fc9 // mov w9, #0xc2fe0000 + .long 0x52b85f09 // mov w9, #0xc2f80000 + .long 0x728e6ee9 // movk w9, #0x7377 .long 0x4e040d37 // dup v23.4s, w9 - .long 0x52a80629 // mov w9, #0x40310000 - .long 0x72922549 // movk w9, #0x912a - .long 0x4f951034 // fmla v20.4s, v1.4s, v21.s[0] - .long 0x6e21e661 // fcmge v1.4s, v19.4s, v1.4s - .long 0x4e040d33 // dup v19.4s, w9 .long 0x52a7f7e9 // mov w9, #0x3fbf0000 - .long 0x4f03d7f2 // movi v18.4s, #0x7f, msl #16 .long 0x7297eea9 // movk w9, #0xbf75 - .long 0x4e21da95 // scvtf v21.4s, v20.4s - .long 0x4e321e92 // and v18.16b, v20.16b, v18.16b - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f951034 // fmla v20.4s, v1.4s, v21.s[0] + .long 0x6e21e641 // fcmge v1.4s, v18.4s, v1.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a7d689 // mov w9, #0x3eb40000 + .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 .long 0x72889f29 // movk w9, #0x44f9 - .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e21da95 // scvtf v21.4s, v20.4s + .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7fb89 // mov w9, #0x3fdc0000 - .long 0x4e33d6f3 // fadd v19.4s, v23.4s, v19.4s + .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s .long 0x729d3469 // movk w9, #0xe9a3 - .long 0x4f0177f2 // orr v18.4s, #0x3f, lsl #24 - .long 0x4eb4ce53 // fmls v19.4s, v18.4s, v20.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 + .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a85e49 // mov w9, #0x42f20000 .long 0x72918a29 // movk w9, #0x8c51 - .long 0x4e35d652 // fadd v18.4s, v18.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7f7c9 // mov w9, #0x3fbe0000 .long 0x729791a9 // movk w9, #0xbc8d - .long 0x6e32fe92 // fdiv v18.4s, v20.4s, v18.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a81349 // mov w9, #0x409a0000 - .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s + .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s .long 0x729ebf09 // movk w9, #0xf5f8 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x52a83ba9 // mov w9, #0x41dd0000 - .long 0x4e219a32 // frintm v18.4s, v17.4s + .long 0x4e219a33 // frintm v19.4s, v17.4s .long 0x729a5fc9 // movk w9, #0xd2fe - .long 0x4e35d635 // fadd v21.4s, v17.4s, v21.4s - .long 0x4eb2d631 // fsub v17.4s, v17.4s, v18.4s - .long 0x4eb4ce35 // fmls v21.4s, v17.4s, v20.4s - .long 0x4eb1d671 // fsub v17.4s, v19.4s, v17.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s + .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s + .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s + .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x91005108 // add x8, x8, #0x14 - .long 0x6e31fe71 // fdiv v17.4s, v19.4s, v17.4s - .long 0x4e31d6b1 // fadd v17.4s, v21.4s, v17.4s - .long 0x4d40c915 // ld1r {v21.4s}, [x8] - .long 0x4f026572 // movi v18.4s, #0x4b, lsl #24 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s + .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s + .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s + .long 0x4d40c914 // ld1r {v20.4s}, [x8] + .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 + .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s .long 0xf9400423 // ldr x3, [x1, #8] - .long 0x4e35d631 // fadd v17.4s, v17.4s, v21.4s - .long 0x6f00e414 // movi v20.2d, #0x0 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x6f00e412 // movi v18.2d, #0x0 .long 0x6e711e01 // bsl v1.16b, v16.16b, v17.16b - .long 0x4f03f613 // fmov v19.4s, #1.000000000000000000e+00 - .long 0x4e34f421 // fmax v1.4s, v1.4s, v20.4s - .long 0x4eb3f421 // fmin v1.4s, v1.4s, v19.4s + .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 + .long 0x4e32f421 // fmax v1.4s, v1.4s, v18.4s + .long 0x4eb5f421 // fmin v1.4s, v1.4s, v21.4s .long 0x91004021 // add x1, x1, #0x10 .long 0xd61f0060 // br x3 @@ -2061,73 +2055,70 @@ _sk_parametric_b_aarch64: .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 .long 0x91004109 // add x9, x8, #0x10 .long 0x9100610a // add x10, x8, #0x18 - .long 0x4d40c933 // ld1r {v19.4s}, [x9] + .long 0x4d40c932 // ld1r {v18.4s}, [x9] .long 0xaa0803e9 // mov x9, x8 - .long 0xbd400d12 // ldr s18, [x8, #12] + .long 0xbd400d11 // ldr s17, [x8, #12] .long 0x4d40c950 // ld1r {v16.4s}, [x10] - .long 0x4ddfc931 // ld1r {v17.4s}, [x9], #4 + .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 .long 0x9100210a // add x10, x8, #0x8 .long 0x4d40c954 // ld1r {v20.4s}, [x10] - .long 0x4f921050 // fmla v16.4s, v2.4s, v18.s[0] + .long 0x4f911050 // fmla v16.4s, v2.4s, v17.s[0] .long 0xbd400135 // ldr s21, [x9] - .long 0x52b85fc9 // mov w9, #0xc2fe0000 + .long 0x52b85f09 // mov w9, #0xc2f80000 + .long 0x728e6ee9 // movk w9, #0x7377 .long 0x4e040d37 // dup v23.4s, w9 - .long 0x52a80629 // mov w9, #0x40310000 - .long 0x72922549 // movk w9, #0x912a - .long 0x4f951054 // fmla v20.4s, v2.4s, v21.s[0] - .long 0x6e22e662 // fcmge v2.4s, v19.4s, v2.4s - .long 0x4e040d33 // dup v19.4s, w9 .long 0x52a7f7e9 // mov w9, #0x3fbf0000 - .long 0x4f03d7f2 // movi v18.4s, #0x7f, msl #16 .long 0x7297eea9 // movk w9, #0xbf75 - .long 0x4e21da95 // scvtf v21.4s, v20.4s - .long 0x4e321e92 // and v18.16b, v20.16b, v18.16b - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f951054 // fmla v20.4s, v2.4s, v21.s[0] + .long 0x6e22e642 // fcmge v2.4s, v18.4s, v2.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a7d689 // mov w9, #0x3eb40000 + .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 .long 0x72889f29 // movk w9, #0x44f9 - .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e21da95 // scvtf v21.4s, v20.4s + .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7fb89 // mov w9, #0x3fdc0000 - .long 0x4e33d6f3 // fadd v19.4s, v23.4s, v19.4s + .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s .long 0x729d3469 // movk w9, #0xe9a3 - .long 0x4f0177f2 // orr v18.4s, #0x3f, lsl #24 - .long 0x4eb4ce53 // fmls v19.4s, v18.4s, v20.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 + .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a85e49 // mov w9, #0x42f20000 .long 0x72918a29 // movk w9, #0x8c51 - .long 0x4e35d652 // fadd v18.4s, v18.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7f7c9 // mov w9, #0x3fbe0000 .long 0x729791a9 // movk w9, #0xbc8d - .long 0x6e32fe92 // fdiv v18.4s, v20.4s, v18.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a81349 // mov w9, #0x409a0000 - .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s + .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s .long 0x729ebf09 // movk w9, #0xf5f8 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x52a83ba9 // mov w9, #0x41dd0000 - .long 0x4e219a32 // frintm v18.4s, v17.4s + .long 0x4e219a33 // frintm v19.4s, v17.4s .long 0x729a5fc9 // movk w9, #0xd2fe - .long 0x4e35d635 // fadd v21.4s, v17.4s, v21.4s - .long 0x4eb2d631 // fsub v17.4s, v17.4s, v18.4s - .long 0x4eb4ce35 // fmls v21.4s, v17.4s, v20.4s - .long 0x4eb1d671 // fsub v17.4s, v19.4s, v17.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s + .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s + .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s + .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x91005108 // add x8, x8, #0x14 - .long 0x6e31fe71 // fdiv v17.4s, v19.4s, v17.4s - .long 0x4e31d6b1 // fadd v17.4s, v21.4s, v17.4s - .long 0x4d40c915 // ld1r {v21.4s}, [x8] - .long 0x4f026572 // movi v18.4s, #0x4b, lsl #24 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s + .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s + .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s + .long 0x4d40c914 // ld1r {v20.4s}, [x8] + .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 + .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s .long 0xf9400423 // ldr x3, [x1, #8] - .long 0x4e35d631 // fadd v17.4s, v17.4s, v21.4s - .long 0x6f00e414 // movi v20.2d, #0x0 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x6f00e412 // movi v18.2d, #0x0 .long 0x6e711e02 // bsl v2.16b, v16.16b, v17.16b - .long 0x4f03f613 // fmov v19.4s, #1.000000000000000000e+00 - .long 0x4e34f442 // fmax v2.4s, v2.4s, v20.4s - .long 0x4eb3f442 // fmin v2.4s, v2.4s, v19.4s + .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 + .long 0x4e32f442 // fmax v2.4s, v2.4s, v18.4s + .long 0x4eb5f442 // fmin v2.4s, v2.4s, v21.4s .long 0x91004021 // add x1, x1, #0x10 .long 0xd61f0060 // br x3 @@ -2139,73 +2130,70 @@ _sk_parametric_a_aarch64: .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 .long 0x91004109 // add x9, x8, #0x10 .long 0x9100610a // add x10, x8, #0x18 - .long 0x4d40c933 // ld1r {v19.4s}, [x9] + .long 0x4d40c932 // ld1r {v18.4s}, [x9] .long 0xaa0803e9 // mov x9, x8 - .long 0xbd400d12 // ldr s18, [x8, #12] + .long 0xbd400d11 // ldr s17, [x8, #12] .long 0x4d40c950 // ld1r {v16.4s}, [x10] - .long 0x4ddfc931 // ld1r {v17.4s}, [x9], #4 + .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 .long 0x9100210a // add x10, x8, #0x8 .long 0x4d40c954 // ld1r {v20.4s}, [x10] - .long 0x4f921070 // fmla v16.4s, v3.4s, v18.s[0] + .long 0x4f911070 // fmla v16.4s, v3.4s, v17.s[0] .long 0xbd400135 // ldr s21, [x9] - .long 0x52b85fc9 // mov w9, #0xc2fe0000 + .long 0x52b85f09 // mov w9, #0xc2f80000 + .long 0x728e6ee9 // movk w9, #0x7377 .long 0x4e040d37 // dup v23.4s, w9 - .long 0x52a80629 // mov w9, #0x40310000 - .long 0x72922549 // movk w9, #0x912a - .long 0x4f951074 // fmla v20.4s, v3.4s, v21.s[0] - .long 0x6e23e663 // fcmge v3.4s, v19.4s, v3.4s - .long 0x4e040d33 // dup v19.4s, w9 .long 0x52a7f7e9 // mov w9, #0x3fbf0000 - .long 0x4f03d7f2 // movi v18.4s, #0x7f, msl #16 .long 0x7297eea9 // movk w9, #0xbf75 - .long 0x4e21da95 // scvtf v21.4s, v20.4s - .long 0x4e321e92 // and v18.16b, v20.16b, v18.16b - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f951074 // fmla v20.4s, v3.4s, v21.s[0] + .long 0x6e23e643 // fcmge v3.4s, v18.4s, v3.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a7d689 // mov w9, #0x3eb40000 + .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 .long 0x72889f29 // movk w9, #0x44f9 - .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e21da95 // scvtf v21.4s, v20.4s + .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7fb89 // mov w9, #0x3fdc0000 - .long 0x4e33d6f3 // fadd v19.4s, v23.4s, v19.4s + .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s .long 0x729d3469 // movk w9, #0xe9a3 - .long 0x4f0177f2 // orr v18.4s, #0x3f, lsl #24 - .long 0x4eb4ce53 // fmls v19.4s, v18.4s, v20.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 + .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a85e49 // mov w9, #0x42f20000 .long 0x72918a29 // movk w9, #0x8c51 - .long 0x4e35d652 // fadd v18.4s, v18.4s, v21.4s - .long 0x4e040d35 // dup v21.4s, w9 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x4e040d34 // dup v20.4s, w9 .long 0x52a7f7c9 // mov w9, #0x3fbe0000 .long 0x729791a9 // movk w9, #0xbc8d - .long 0x6e32fe92 // fdiv v18.4s, v20.4s, v18.4s - .long 0x4e040d34 // dup v20.4s, w9 + .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s + .long 0x4e040d32 // dup v18.4s, w9 .long 0x52a81349 // mov w9, #0x409a0000 - .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s + .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s .long 0x729ebf09 // movk w9, #0xf5f8 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x52a83ba9 // mov w9, #0x41dd0000 - .long 0x4e219a32 // frintm v18.4s, v17.4s + .long 0x4e219a33 // frintm v19.4s, v17.4s .long 0x729a5fc9 // movk w9, #0xd2fe - .long 0x4e35d635 // fadd v21.4s, v17.4s, v21.4s - .long 0x4eb2d631 // fsub v17.4s, v17.4s, v18.4s - .long 0x4eb4ce35 // fmls v21.4s, v17.4s, v20.4s - .long 0x4eb1d671 // fsub v17.4s, v19.4s, v17.4s - .long 0x4e040d33 // dup v19.4s, w9 + .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s + .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s + .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s + .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s + .long 0x4e040d35 // dup v21.4s, w9 .long 0x91005108 // add x8, x8, #0x14 - .long 0x6e31fe71 // fdiv v17.4s, v19.4s, v17.4s - .long 0x4e31d6b1 // fadd v17.4s, v21.4s, v17.4s - .long 0x4d40c915 // ld1r {v21.4s}, [x8] - .long 0x4f026572 // movi v18.4s, #0x4b, lsl #24 - .long 0x6e32de31 // fmul v17.4s, v17.4s, v18.4s + .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s + .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s + .long 0x4d40c914 // ld1r {v20.4s}, [x8] + .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 + .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s .long 0xf9400423 // ldr x3, [x1, #8] - .long 0x4e35d631 // fadd v17.4s, v17.4s, v21.4s - .long 0x6f00e414 // movi v20.2d, #0x0 + .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s + .long 0x6f00e412 // movi v18.2d, #0x0 .long 0x6e711e03 // bsl v3.16b, v16.16b, v17.16b - .long 0x4f03f613 // fmov v19.4s, #1.000000000000000000e+00 - .long 0x4e34f463 // fmax v3.4s, v3.4s, v20.4s - .long 0x4eb3f463 // fmin v3.4s, v3.4s, v19.4s + .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 + .long 0x4e32f463 // fmax v3.4s, v3.4s, v18.4s + .long 0x4eb5f463 // fmin v3.4s, v3.4s, v21.4s .long 0x91004021 // add x1, x1, #0x10 .long 0xd61f0060 // br x3 @@ -2361,9 +2349,9 @@ FUNCTION(_sk_gather_i8_aarch64) _sk_gather_i8_aarch64: .long 0xaa0103e8 // mov x8, x1 .long 0xf8408429 // ldr x9, [x1], #8 - .long 0xb4000069 // cbz x9, 1f70 + .long 0xb4000069 // cbz x9, 1f40 .long 0xaa0903ea // mov x10, x9 - .long 0x14000003 // b 1f78 + .long 0x14000003 // b 1f48 .long 0xf940050a // ldr x10, [x8, #8] .long 0x91004101 // add x1, x8, #0x10 .long 0xf8410548 // ldr x8, [x10], #16 @@ -3212,7 +3200,7 @@ _sk_linear_gradient_aarch64: .long 0x4d40c902 // ld1r {v2.4s}, [x8] .long 0xf9400128 // ldr x8, [x9] .long 0x4d40c943 // ld1r {v3.4s}, [x10] - .long 0xb40006c8 // cbz x8, 2b44 + .long 0xb40006c8 // cbz x8, 2b14 .long 0x6dbf23e9 // stp d9, d8, [sp, #-16]! .long 0xf9400529 // ldr x9, [x9, #8] .long 0x6f00e413 // movi v19.2d, #0x0 @@ -3263,9 +3251,9 @@ _sk_linear_gradient_aarch64: .long 0xd1000508 // sub x8, x8, #0x1 .long 0x6e771fd0 // bsl v16.16b, v30.16b, v23.16b .long 0x91009129 // add x9, x9, #0x24 - .long 0xb5fffaa8 // cbnz x8, 2a8c + .long 0xb5fffaa8 // cbnz x8, 2a5c .long 0x6cc123e9 // ldp d9, d8, [sp], #16 - .long 0x14000005 // b 2b54 + .long 0x14000005 // b 2b24 .long 0x6f00e414 // movi v20.2d, #0x0 .long 0x6f00e412 // movi v18.2d, #0x0 .long 0x6f00e411 // movi v17.2d, #0x0 @@ -5553,10 +5541,10 @@ _sk_parametric_r_vfp4: .long 0xe92d4800 // push {fp, lr} .long 0xed2d8b06 // vpush {d8-d10} .long 0xe591e000 // ldr lr, [r1] - .long 0xeddf3b43 // vldr d19, [pc, #268] - .long 0xed9f8a52 // vldr s16, [pc, #328] + .long 0xeddf3b41 // vldr d19, [pc, #260] + .long 0xed9f8a4e // vldr s16, [pc, #312] .long 0xe1a0300e // mov r3, lr - .long 0xeddf4b46 // vldr d20, [pc, #280] + .long 0xeddf4b42 // vldr d20, [pc, #264] .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 @@ -5572,16 +5560,14 @@ _sk_parametric_r_vfp4: .long 0xf2019da3 // vadd.f32 d9, d17, d19 .long 0xf2c33614 // vmov.i32 d19, #872415232 .long 0xf3422db3 // vmul.f32 d18, d18, d19 - .long 0xeddf3b32 // vldr d19, [pc, #200] + .long 0xeddf3b30 // vldr d19, [pc, #192] .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 .long 0xee88aa09 // vdiv.f32 s20, s16, s18 - .long 0xf2422da3 // vadd.f32 d18, d18, d19 - .long 0xeddf3b30 // vldr d19, [pc, #192] - .long 0xed9f8a3c // vldr s16, [pc, #240] .long 0xf3411db3 // vmul.f32 d17, d17, d19 - .long 0xf2c03010 // vmov.i32 d19, #0 + .long 0xed9f8a39 // vldr s16, [pc, #228] .long 0xf2422da4 // vadd.f32 d18, d18, d20 - .long 0xeddf4b2f // vldr d20, [pc, #188] + .long 0xeddf4b2e // vldr d20, [pc, #184] + .long 0xf2c03010 // vmov.i32 d19, #0 .long 0xf2621da1 // vsub.f32 d17, d18, d17 .long 0xf2611d8a // vsub.f32 d17, d17, d10 .long 0xf3400db1 // vmul.f32 d16, d16, d17 @@ -5589,13 +5575,13 @@ _sk_parametric_r_vfp4: .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 .long 0xf35421b3 // vbsl d18, d20, d19 - .long 0xeddf4b2d // vldr d20, [pc, #180] + .long 0xeddf4b2b // vldr d20, [pc, #172] .long 0xf2611da2 // vsub.f32 d17, d17, d18 - .long 0xeddf2b27 // vldr d18, [pc, #156] + .long 0xeddf2b25 // vldr d18, [pc, #148] .long 0xf2601da1 // vsub.f32 d17, d16, d17 .long 0xf2400da4 // vadd.f32 d16, d16, d20 .long 0xf2229da1 // vsub.f32 d9, d18, d17 - .long 0xeddf2b25 // vldr d18, [pc, #148] + .long 0xeddf2b23 // vldr d18, [pc, #140] .long 0xf3411db2 // vmul.f32 d17, d17, d18 .long 0xf2c3261f // vmov.i32 d18, #1056964608 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 @@ -5624,12 +5610,10 @@ _sk_parametric_r_vfp4: .long 0xe12fff1c // bx ip .long 0x3eb444f9 // .word 0x3eb444f9 .long 0x3eb444f9 // .word 0x3eb444f9 - .long 0xc2fe0000 // .word 0xc2fe0000 - .long 0xc2fe0000 // .word 0xc2fe0000 .long 0x3fbfbf75 // .word 0x3fbfbf75 .long 0x3fbfbf75 // .word 0x3fbfbf75 - .long 0x4031912a // .word 0x4031912a - .long 0x4031912a // .word 0x4031912a + .long 0xc2f87377 // .word 0xc2f87377 + .long 0xc2f87377 // .word 0xc2f87377 .long 0x3f800000 // .word 0x3f800000 .long 0x3f800000 // .word 0x3f800000 .long 0x409af5f8 // .word 0x409af5f8 @@ -5648,10 +5632,10 @@ _sk_parametric_g_vfp4: .long 0xe92d4800 // push {fp, lr} .long 0xed2d8b06 // vpush {d8-d10} .long 0xe591e000 // ldr lr, [r1] - .long 0xeddf3b43 // vldr d19, [pc, #268] - .long 0xed9f8a52 // vldr s16, [pc, #328] + .long 0xeddf3b41 // vldr d19, [pc, #260] + .long 0xed9f8a4e // vldr s16, [pc, #312] .long 0xe1a0300e // mov r3, lr - .long 0xeddf4b46 // vldr d20, [pc, #280] + .long 0xeddf4b42 // vldr d20, [pc, #264] .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 @@ -5667,16 +5651,14 @@ _sk_parametric_g_vfp4: .long 0xf2019da3 // vadd.f32 d9, d17, d19 .long 0xf2c33614 // vmov.i32 d19, #872415232 .long 0xf3422db3 // vmul.f32 d18, d18, d19 - .long 0xeddf3b32 // vldr d19, [pc, #200] + .long 0xeddf3b30 // vldr d19, [pc, #192] .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 .long 0xee88aa09 // vdiv.f32 s20, s16, s18 - .long 0xf2422da3 // vadd.f32 d18, d18, d19 - .long 0xeddf3b30 // vldr d19, [pc, #192] - .long 0xed9f8a3c // vldr s16, [pc, #240] .long 0xf3411db3 // vmul.f32 d17, d17, d19 - .long 0xf2c03010 // vmov.i32 d19, #0 + .long 0xed9f8a39 // vldr s16, [pc, #228] .long 0xf2422da4 // vadd.f32 d18, d18, d20 - .long 0xeddf4b2f // vldr d20, [pc, #188] + .long 0xeddf4b2e // vldr d20, [pc, #184] + .long 0xf2c03010 // vmov.i32 d19, #0 .long 0xf2621da1 // vsub.f32 d17, d18, d17 .long 0xf2611d8a // vsub.f32 d17, d17, d10 .long 0xf3400db1 // vmul.f32 d16, d16, d17 @@ -5684,13 +5666,13 @@ _sk_parametric_g_vfp4: .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 .long 0xf35421b3 // vbsl d18, d20, d19 - .long 0xeddf4b2d // vldr d20, [pc, #180] + .long 0xeddf4b2b // vldr d20, [pc, #172] .long 0xf2611da2 // vsub.f32 d17, d17, d18 - .long 0xeddf2b27 // vldr d18, [pc, #156] + .long 0xeddf2b25 // vldr d18, [pc, #148] .long 0xf2601da1 // vsub.f32 d17, d16, d17 .long 0xf2400da4 // vadd.f32 d16, d16, d20 .long 0xf2229da1 // vsub.f32 d9, d18, d17 - .long 0xeddf2b25 // vldr d18, [pc, #148] + .long 0xeddf2b23 // vldr d18, [pc, #140] .long 0xf3411db2 // vmul.f32 d17, d17, d18 .long 0xf2c3261f // vmov.i32 d18, #1056964608 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 @@ -5719,12 +5701,10 @@ _sk_parametric_g_vfp4: .long 0xe12fff1c // bx ip .long 0x3eb444f9 // .word 0x3eb444f9 .long 0x3eb444f9 // .word 0x3eb444f9 - .long 0xc2fe0000 // .word 0xc2fe0000 - .long 0xc2fe0000 // .word 0xc2fe0000 .long 0x3fbfbf75 // .word 0x3fbfbf75 .long 0x3fbfbf75 // .word 0x3fbfbf75 - .long 0x4031912a // .word 0x4031912a - .long 0x4031912a // .word 0x4031912a + .long 0xc2f87377 // .word 0xc2f87377 + .long 0xc2f87377 // .word 0xc2f87377 .long 0x3f800000 // .word 0x3f800000 .long 0x3f800000 // .word 0x3f800000 .long 0x409af5f8 // .word 0x409af5f8 @@ -5743,10 +5723,10 @@ _sk_parametric_b_vfp4: .long 0xe92d4800 // push {fp, lr} .long 0xed2d8b06 // vpush {d8-d10} .long 0xe591e000 // ldr lr, [r1] - .long 0xeddf3b43 // vldr d19, [pc, #268] - .long 0xed9f8a52 // vldr s16, [pc, #328] + .long 0xeddf3b41 // vldr d19, [pc, #260] + .long 0xed9f8a4e // vldr s16, [pc, #312] .long 0xe1a0300e // mov r3, lr - .long 0xeddf4b46 // vldr d20, [pc, #280] + .long 0xeddf4b42 // vldr d20, [pc, #264] .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 @@ -5762,16 +5742,14 @@ _sk_parametric_b_vfp4: .long 0xf2019da3 // vadd.f32 d9, d17, d19 .long 0xf2c33614 // vmov.i32 d19, #872415232 .long 0xf3422db3 // vmul.f32 d18, d18, d19 - .long 0xeddf3b32 // vldr d19, [pc, #200] + .long 0xeddf3b30 // vldr d19, [pc, #192] .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 .long 0xee88aa09 // vdiv.f32 s20, s16, s18 - .long 0xf2422da3 // vadd.f32 d18, d18, d19 - .long 0xeddf3b30 // vldr d19, [pc, #192] - .long 0xed9f8a3c // vldr s16, [pc, #240] .long 0xf3411db3 // vmul.f32 d17, d17, d19 - .long 0xf2c03010 // vmov.i32 d19, #0 + .long 0xed9f8a39 // vldr s16, [pc, #228] .long 0xf2422da4 // vadd.f32 d18, d18, d20 - .long 0xeddf4b2f // vldr d20, [pc, #188] + .long 0xeddf4b2e // vldr d20, [pc, #184] + .long 0xf2c03010 // vmov.i32 d19, #0 .long 0xf2621da1 // vsub.f32 d17, d18, d17 .long 0xf2611d8a // vsub.f32 d17, d17, d10 .long 0xf3400db1 // vmul.f32 d16, d16, d17 @@ -5779,13 +5757,13 @@ _sk_parametric_b_vfp4: .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 .long 0xf35421b3 // vbsl d18, d20, d19 - .long 0xeddf4b2d // vldr d20, [pc, #180] + .long 0xeddf4b2b // vldr d20, [pc, #172] .long 0xf2611da2 // vsub.f32 d17, d17, d18 - .long 0xeddf2b27 // vldr d18, [pc, #156] + .long 0xeddf2b25 // vldr d18, [pc, #148] .long 0xf2601da1 // vsub.f32 d17, d16, d17 .long 0xf2400da4 // vadd.f32 d16, d16, d20 .long 0xf2229da1 // vsub.f32 d9, d18, d17 - .long 0xeddf2b25 // vldr d18, [pc, #148] + .long 0xeddf2b23 // vldr d18, [pc, #140] .long 0xf3411db2 // vmul.f32 d17, d17, d18 .long 0xf2c3261f // vmov.i32 d18, #1056964608 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 @@ -5814,12 +5792,10 @@ _sk_parametric_b_vfp4: .long 0xe12fff1c // bx ip .long 0x3eb444f9 // .word 0x3eb444f9 .long 0x3eb444f9 // .word 0x3eb444f9 - .long 0xc2fe0000 // .word 0xc2fe0000 - .long 0xc2fe0000 // .word 0xc2fe0000 .long 0x3fbfbf75 // .word 0x3fbfbf75 .long 0x3fbfbf75 // .word 0x3fbfbf75 - .long 0x4031912a // .word 0x4031912a - .long 0x4031912a // .word 0x4031912a + .long 0xc2f87377 // .word 0xc2f87377 + .long 0xc2f87377 // .word 0xc2f87377 .long 0x3f800000 // .word 0x3f800000 .long 0x3f800000 // .word 0x3f800000 .long 0x409af5f8 // .word 0x409af5f8 @@ -5838,10 +5814,10 @@ _sk_parametric_a_vfp4: .long 0xe92d4800 // push {fp, lr} .long 0xed2d8b06 // vpush {d8-d10} .long 0xe591e000 // ldr lr, [r1] - .long 0xeddf3b43 // vldr d19, [pc, #268] - .long 0xed9f8a52 // vldr s16, [pc, #328] + .long 0xeddf3b41 // vldr d19, [pc, #260] + .long 0xed9f8a4e // vldr s16, [pc, #312] .long 0xe1a0300e // mov r3, lr - .long 0xeddf4b46 // vldr d20, [pc, #280] + .long 0xeddf4b42 // vldr d20, [pc, #264] .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! .long 0xe591c004 // ldr ip, [r1, #4] .long 0xe2811008 // add r1, r1, #8 @@ -5857,16 +5833,14 @@ _sk_parametric_a_vfp4: .long 0xf2019da3 // vadd.f32 d9, d17, d19 .long 0xf2c33614 // vmov.i32 d19, #872415232 .long 0xf3422db3 // vmul.f32 d18, d18, d19 - .long 0xeddf3b32 // vldr d19, [pc, #200] + .long 0xeddf3b30 // vldr d19, [pc, #192] .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 .long 0xee88aa09 // vdiv.f32 s20, s16, s18 - .long 0xf2422da3 // vadd.f32 d18, d18, d19 - .long 0xeddf3b30 // vldr d19, [pc, #192] - .long 0xed9f8a3c // vldr s16, [pc, #240] .long 0xf3411db3 // vmul.f32 d17, d17, d19 - .long 0xf2c03010 // vmov.i32 d19, #0 + .long 0xed9f8a39 // vldr s16, [pc, #228] .long 0xf2422da4 // vadd.f32 d18, d18, d20 - .long 0xeddf4b2f // vldr d20, [pc, #188] + .long 0xeddf4b2e // vldr d20, [pc, #184] + .long 0xf2c03010 // vmov.i32 d19, #0 .long 0xf2621da1 // vsub.f32 d17, d18, d17 .long 0xf2611d8a // vsub.f32 d17, d17, d10 .long 0xf3400db1 // vmul.f32 d16, d16, d17 @@ -5874,13 +5848,13 @@ _sk_parametric_a_vfp4: .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 .long 0xf35421b3 // vbsl d18, d20, d19 - .long 0xeddf4b2d // vldr d20, [pc, #180] + .long 0xeddf4b2b // vldr d20, [pc, #172] .long 0xf2611da2 // vsub.f32 d17, d17, d18 - .long 0xeddf2b27 // vldr d18, [pc, #156] + .long 0xeddf2b25 // vldr d18, [pc, #148] .long 0xf2601da1 // vsub.f32 d17, d16, d17 .long 0xf2400da4 // vadd.f32 d16, d16, d20 .long 0xf2229da1 // vsub.f32 d9, d18, d17 - .long 0xeddf2b25 // vldr d18, [pc, #148] + .long 0xeddf2b23 // vldr d18, [pc, #140] .long 0xf3411db2 // vmul.f32 d17, d17, d18 .long 0xf2c3261f // vmov.i32 d18, #1056964608 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 @@ -5909,12 +5883,10 @@ _sk_parametric_a_vfp4: .long 0xe12fff1c // bx ip .long 0x3eb444f9 // .word 0x3eb444f9 .long 0x3eb444f9 // .word 0x3eb444f9 - .long 0xc2fe0000 // .word 0xc2fe0000 - .long 0xc2fe0000 // .word 0xc2fe0000 .long 0x3fbfbf75 // .word 0x3fbfbf75 .long 0x3fbfbf75 // .word 0x3fbfbf75 - .long 0x4031912a // .word 0x4031912a - .long 0x4031912a // .word 0x4031912a + .long 0xc2f87377 // .word 0xc2f87377 + .long 0xc2f87377 // .word 0xc2f87377 .long 0x3f800000 // .word 0x3f800000 .long 0x3f800000 // .word 0x3f800000 .long 0x409af5f8 // .word 0x409af5f8 @@ -7106,7 +7078,7 @@ _sk_linear_gradient_vfp4: .long 0xe494c00c // ldr ip, [r4], #12 .long 0xf4a41c9f // vld1.32 {d1[]}, [r4 :32] .long 0xe35c0000 // cmp ip, #0 - .long 0x0a000036 // beq 2fb8 + .long 0x0a000036 // beq 2f78 .long 0xe59e3004 // ldr r3, [lr, #4] .long 0xf2c01010 // vmov.i32 d17, #0 .long 0xf2c07010 // vmov.i32 d23, #0 @@ -7156,12 +7128,12 @@ _sk_linear_gradient_vfp4: .long 0xf26371b3 // vorr d23, d19, d19 .long 0xf26481b4 // vorr d24, d20, d20 .long 0xf26561b5 // vorr d22, d21, d21 - .long 0x1affffd3 // bne 2ef4 + .long 0x1affffd3 // bne 2eb4 .long 0xf26c01bc // vorr d16, d28, d28 .long 0xf22b11bb // vorr d1, d27, d27 .long 0xf22a21ba // vorr d2, d26, d26 .long 0xf22931b9 // vorr d3, d25, d25 - .long 0xea000003 // b 2fc8 + .long 0xea000003 // b 2f88 .long 0xf2c05010 // vmov.i32 d21, #0 .long 0xf2c04010 // vmov.i32 d20, #0 .long 0xf2c03010 // vmov.i32 d19, #0 @@ -9592,14 +9564,11 @@ _sk_parametric_r_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,125,168,211 // vfmadd213ps %ymm11,%ymm0,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,224 // vpbroadcastd %xmm0,%ymm12 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 - .byte 196,66,37,186,236 // vfmsub231ps %ymm12,%ymm11,%ymm13 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 @@ -9608,25 +9577,24 @@ _sk_parametric_r_hsw: .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 .byte 197,45,235,208 // vpor %ymm0,%ymm10,%ymm10 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 - .byte 197,20,88,216 // vaddps %ymm0,%ymm13,%ymm11 + .byte 196,98,125,88,240 // vpbroadcastd %xmm0,%ymm14 + .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,224 // vpbroadcastd %xmm0,%ymm12 - .byte 196,66,45,172,227 // vfnmadd213ps %ymm11,%ymm10,%ymm12 + .byte 196,98,125,88,216 // vpbroadcastd %xmm0,%ymm11 + .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,98,125,88,216 // vpbroadcastd %xmm0,%ymm11 + .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0 - .byte 197,164,94,192 // vdivps %ymm0,%ymm11,%ymm0 - .byte 197,156,92,192 // vsubps %ymm0,%ymm12,%ymm0 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 + .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 + .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 + .byte 197,28,89,216 // vmulps %ymm0,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -9676,14 +9644,11 @@ _sk_parametric_g_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,117,168,211 // vfmadd213ps %ymm11,%ymm1,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 - .byte 196,66,37,186,236 // vfmsub231ps %ymm12,%ymm11,%ymm13 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 @@ -9692,25 +9657,24 @@ _sk_parametric_g_hsw: .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 .byte 197,45,235,209 // vpor %ymm1,%ymm10,%ymm10 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 - .byte 197,20,88,217 // vaddps %ymm1,%ymm13,%ymm11 + .byte 196,98,125,88,241 // vpbroadcastd %xmm1,%ymm14 + .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12 - .byte 196,66,45,172,227 // vfnmadd213ps %ymm11,%ymm10,%ymm12 + .byte 196,98,125,88,217 // vpbroadcastd %xmm1,%ymm11 + .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,98,125,88,217 // vpbroadcastd %xmm1,%ymm11 + .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 - .byte 197,164,94,201 // vdivps %ymm1,%ymm11,%ymm1 - .byte 197,156,92,201 // vsubps %ymm1,%ymm12,%ymm1 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 + .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 + .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 + .byte 197,28,89,217 // vmulps %ymm1,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -9760,14 +9724,11 @@ _sk_parametric_b_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,109,168,211 // vfmadd213ps %ymm11,%ymm2,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,226 // vpbroadcastd %xmm2,%ymm12 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 - .byte 196,66,37,186,236 // vfmsub231ps %ymm12,%ymm11,%ymm13 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 @@ -9776,25 +9737,24 @@ _sk_parametric_b_hsw: .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 .byte 197,45,235,210 // vpor %ymm2,%ymm10,%ymm10 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 - .byte 197,20,88,218 // vaddps %ymm2,%ymm13,%ymm11 + .byte 196,98,125,88,242 // vpbroadcastd %xmm2,%ymm14 + .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,226 // vpbroadcastd %xmm2,%ymm12 - .byte 196,66,45,172,227 // vfnmadd213ps %ymm11,%ymm10,%ymm12 + .byte 196,98,125,88,218 // vpbroadcastd %xmm2,%ymm11 + .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,98,125,88,218 // vpbroadcastd %xmm2,%ymm11 + .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 - .byte 197,164,94,210 // vdivps %ymm2,%ymm11,%ymm2 - .byte 197,156,92,210 // vsubps %ymm2,%ymm12,%ymm2 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,218 // vmulps %ymm2,%ymm10,%ymm11 + .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 + .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 + .byte 197,28,89,218 // vmulps %ymm2,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -9844,14 +9804,11 @@ _sk_parametric_a_hsw: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 196,66,101,168,211 // vfmadd213ps %ymm11,%ymm3,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,227 // vpbroadcastd %xmm3,%ymm12 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 - .byte 196,66,37,186,236 // vfmsub231ps %ymm12,%ymm11,%ymm13 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 @@ -9860,25 +9817,24 @@ _sk_parametric_a_hsw: .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 .byte 197,45,235,211 // vpor %ymm3,%ymm10,%ymm10 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 - .byte 197,20,88,219 // vaddps %ymm3,%ymm13,%ymm11 + .byte 196,98,125,88,243 // vpbroadcastd %xmm3,%ymm14 + .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,227 // vpbroadcastd %xmm3,%ymm12 - .byte 196,66,45,172,227 // vfnmadd213ps %ymm11,%ymm10,%ymm12 + .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 + .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 + .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 .byte 197,172,88,219 // vaddps %ymm3,%ymm10,%ymm3 - .byte 197,164,94,219 // vdivps %ymm3,%ymm11,%ymm3 - .byte 197,156,92,219 // vsubps %ymm3,%ymm12,%ymm3 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,219 // vmulps %ymm3,%ymm10,%ymm11 + .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 + .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 + .byte 197,28,89,219 // vmulps %ymm3,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -9924,7 +9880,7 @@ _sk_load_a8_hsw: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,50 // jne 23a8 <_sk_load_a8_hsw+0x42> + .byte 117,50 // jne 2358 <_sk_load_a8_hsw+0x42> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 @@ -9947,9 +9903,9 @@ _sk_load_a8_hsw: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 23b0 <_sk_load_a8_hsw+0x4a> + .byte 117,234 // jne 2360 <_sk_load_a8_hsw+0x4a> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,173 // jmp 237a <_sk_load_a8_hsw+0x14> + .byte 235,173 // jmp 232a <_sk_load_a8_hsw+0x14> HIDDEN _sk_gather_a8_hsw .globl _sk_gather_a8_hsw @@ -10024,7 +9980,7 @@ _sk_store_a8_hsw: .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 24e5 <_sk_store_a8_hsw+0x3b> + .byte 117,10 // jne 2495 <_sk_store_a8_hsw+0x3b> .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -10032,10 +9988,10 @@ _sk_store_a8_hsw: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 24e1 <_sk_store_a8_hsw+0x37> + .byte 119,236 // ja 2491 <_sk_store_a8_hsw+0x37> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 2548 <_sk_store_a8_hsw+0x9e> + .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 24f8 <_sk_store_a8_hsw+0x9e> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -10046,7 +10002,7 @@ _sk_store_a8_hsw: .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) - .byte 235,154 // jmp 24e1 <_sk_store_a8_hsw+0x37> + .byte 235,154 // jmp 2491 <_sk_store_a8_hsw+0x37> .byte 144 // nop .byte 246,255 // idiv %bh .byte 255 // (bad) @@ -10080,7 +10036,7 @@ _sk_load_g8_hsw: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,60 // jne 25b0 <_sk_load_g8_hsw+0x4c> + .byte 117,60 // jne 2560 <_sk_load_g8_hsw+0x4c> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 @@ -10105,9 +10061,9 @@ _sk_load_g8_hsw: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 25b8 <_sk_load_g8_hsw+0x54> + .byte 117,234 // jne 2568 <_sk_load_g8_hsw+0x54> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,163 // jmp 2578 <_sk_load_g8_hsw+0x14> + .byte 235,163 // jmp 2528 <_sk_load_g8_hsw+0x14> HIDDEN _sk_gather_g8_hsw .globl _sk_gather_g8_hsw @@ -10176,9 +10132,9 @@ _sk_gather_i8_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 26cb <_sk_gather_i8_hsw+0xf> + .byte 116,5 // je 267b <_sk_gather_i8_hsw+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 26cd <_sk_gather_i8_hsw+0x11> + .byte 235,2 // jmp 267d <_sk_gather_i8_hsw+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -10251,7 +10207,7 @@ _sk_load_565_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,149,0,0,0 // jne 287f <_sk_load_565_hsw+0xa3> + .byte 15,133,149,0,0,0 // jne 282f <_sk_load_565_hsw+0xa3> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 .byte 184,0,248,0,0 // mov $0xf800,%eax @@ -10291,9 +10247,9 @@ _sk_load_565_hsw: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,89,255,255,255 // ja 27f0 <_sk_load_565_hsw+0x14> + .byte 15,135,89,255,255,255 // ja 27a0 <_sk_load_565_hsw+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 28ec <_sk_load_565_hsw+0x110> + .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 289c <_sk_load_565_hsw+0x110> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -10305,12 +10261,12 @@ _sk_load_565_hsw: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,5,255,255,255 // jmpq 27f0 <_sk_load_565_hsw+0x14> + .byte 233,5,255,255,255 // jmpq 27a0 <_sk_load_565_hsw+0x14> .byte 144 // nop .byte 243,255 // repz (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 235,255 // jmp 28f1 <_sk_load_565_hsw+0x115> + .byte 235,255 // jmp 28a1 <_sk_load_565_hsw+0x115> .byte 255 // (bad) .byte 255,227 // jmpq *%rbx .byte 255 // (bad) @@ -10437,7 +10393,7 @@ _sk_store_565_hsw: .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 2ab7 <_sk_store_565_hsw+0x6c> + .byte 117,10 // jne 2a67 <_sk_store_565_hsw+0x6c> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -10445,9 +10401,9 @@ _sk_store_565_hsw: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 2ab3 <_sk_store_565_hsw+0x68> + .byte 119,236 // ja 2a63 <_sk_store_565_hsw+0x68> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2b14 <_sk_store_565_hsw+0xc9> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2ac4 <_sk_store_565_hsw+0xc9> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -10458,7 +10414,7 @@ _sk_store_565_hsw: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 2ab3 <_sk_store_565_hsw+0x68> + .byte 235,159 // jmp 2a63 <_sk_store_565_hsw+0x68> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -10489,7 +10445,7 @@ _sk_load_4444_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,179,0,0,0 // jne 2bf1 <_sk_load_4444_hsw+0xc1> + .byte 15,133,179,0,0,0 // jne 2ba1 <_sk_load_4444_hsw+0xc1> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 196,98,125,51,200 // vpmovzxwd %xmm0,%ymm9 .byte 184,0,240,0,0 // mov $0xf000,%eax @@ -10535,9 +10491,9 @@ _sk_load_4444_hsw: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,59,255,255,255 // ja 2b44 <_sk_load_4444_hsw+0x14> + .byte 15,135,59,255,255,255 // ja 2af4 <_sk_load_4444_hsw+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 2c60 <_sk_load_4444_hsw+0x130> + .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 2c10 <_sk_load_4444_hsw+0x130> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -10549,13 +10505,13 @@ _sk_load_4444_hsw: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,231,254,255,255 // jmpq 2b44 <_sk_load_4444_hsw+0x14> + .byte 233,231,254,255,255 // jmpq 2af4 <_sk_load_4444_hsw+0x14> .byte 15,31,0 // nopl (%rax) .byte 241 // icebp .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,225 // jmpq ffffffffe2002c68 <_sk_callback_hsw+0xffffffffe1ffeace> + .byte 233,255,255,255,225 // jmpq ffffffffe2002c18 <_sk_callback_hsw+0xffffffffe1ffeace> .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) @@ -10687,7 +10643,7 @@ _sk_store_4444_hsw: .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 2e4f <_sk_store_4444_hsw+0x72> + .byte 117,10 // jne 2dff <_sk_store_4444_hsw+0x72> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -10695,9 +10651,9 @@ _sk_store_4444_hsw: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 2e4b <_sk_store_4444_hsw+0x6e> + .byte 119,236 // ja 2dfb <_sk_store_4444_hsw+0x6e> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2eac <_sk_store_4444_hsw+0xcf> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2e5c <_sk_store_4444_hsw+0xcf> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -10708,7 +10664,7 @@ _sk_store_4444_hsw: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 2e4b <_sk_store_4444_hsw+0x6e> + .byte 235,159 // jmp 2dfb <_sk_store_4444_hsw+0x6e> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -10741,7 +10697,7 @@ _sk_load_8888_hsw: .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 76,3,8 // add (%rax),%r9 .byte 77,133,192 // test %r8,%r8 - .byte 117,104 // jne 2f45 <_sk_load_8888_hsw+0x7d> + .byte 117,104 // jne 2ef5 <_sk_load_8888_hsw+0x7d> .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 .byte 184,255,0,0,0 // mov $0xff,%eax .byte 197,249,110,192 // vmovd %eax,%xmm0 @@ -10774,7 +10730,7 @@ _sk_load_8888_hsw: .byte 196,225,249,110,192 // vmovq %rax,%xmm0 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 - .byte 233,116,255,255,255 // jmpq 2ee2 <_sk_load_8888_hsw+0x1a> + .byte 233,116,255,255,255 // jmpq 2e92 <_sk_load_8888_hsw+0x1a> HIDDEN _sk_gather_8888_hsw .globl _sk_gather_8888_hsw @@ -10838,7 +10794,7 @@ _sk_store_8888_hsw: .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8 .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8 .byte 77,133,192 // test %r8,%r8 - .byte 117,12 // jne 3068 <_sk_store_8888_hsw+0x74> + .byte 117,12 // jne 3018 <_sk_store_8888_hsw+0x74> .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,193 // mov %r8,%rcx @@ -10851,7 +10807,7 @@ _sk_store_8888_hsw: .byte 196,97,249,110,200 // vmovq %rax,%xmm9 .byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9 .byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9) - .byte 235,211 // jmp 3061 <_sk_store_8888_hsw+0x6d> + .byte 235,211 // jmp 3011 <_sk_store_8888_hsw+0x6d> HIDDEN _sk_load_f16_hsw .globl _sk_load_f16_hsw @@ -10860,7 +10816,7 @@ _sk_load_f16_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 117,97 // jne 30f9 <_sk_load_f16_hsw+0x6b> + .byte 117,97 // jne 30a9 <_sk_load_f16_hsw+0x6b> .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -10886,29 +10842,29 @@ _sk_load_f16_hsw: .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,79 // je 3158 <_sk_load_f16_hsw+0xca> + .byte 116,79 // je 3108 <_sk_load_f16_hsw+0xca> .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,67 // jb 3158 <_sk_load_f16_hsw+0xca> + .byte 114,67 // jb 3108 <_sk_load_f16_hsw+0xca> .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,68 // je 3165 <_sk_load_f16_hsw+0xd7> + .byte 116,68 // je 3115 <_sk_load_f16_hsw+0xd7> .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,56 // jb 3165 <_sk_load_f16_hsw+0xd7> + .byte 114,56 // jb 3115 <_sk_load_f16_hsw+0xd7> .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,114,255,255,255 // je 30af <_sk_load_f16_hsw+0x21> + .byte 15,132,114,255,255,255 // je 305f <_sk_load_f16_hsw+0x21> .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,98,255,255,255 // jb 30af <_sk_load_f16_hsw+0x21> + .byte 15,130,98,255,255,255 // jb 305f <_sk_load_f16_hsw+0x21> .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 - .byte 233,87,255,255,255 // jmpq 30af <_sk_load_f16_hsw+0x21> + .byte 233,87,255,255,255 // jmpq 305f <_sk_load_f16_hsw+0x21> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,74,255,255,255 // jmpq 30af <_sk_load_f16_hsw+0x21> + .byte 233,74,255,255,255 // jmpq 305f <_sk_load_f16_hsw+0x21> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,65,255,255,255 // jmpq 30af <_sk_load_f16_hsw+0x21> + .byte 233,65,255,255,255 // jmpq 305f <_sk_load_f16_hsw+0x21> HIDDEN _sk_gather_f16_hsw .globl _sk_gather_f16_hsw @@ -10966,7 +10922,7 @@ _sk_store_f16_hsw: .byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9 .byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,27 // jne 325d <_sk_store_f16_hsw+0x65> + .byte 117,27 // jne 320d <_sk_store_f16_hsw+0x65> .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8) .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8) .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8) @@ -10975,22 +10931,22 @@ _sk_store_f16_hsw: .byte 255,224 // jmpq *%rax .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,241 // je 3259 <_sk_store_f16_hsw+0x61> + .byte 116,241 // je 3209 <_sk_store_f16_hsw+0x61> .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,229 // jb 3259 <_sk_store_f16_hsw+0x61> + .byte 114,229 // jb 3209 <_sk_store_f16_hsw+0x61> .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) - .byte 116,221 // je 3259 <_sk_store_f16_hsw+0x61> + .byte 116,221 // je 3209 <_sk_store_f16_hsw+0x61> .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,209 // jb 3259 <_sk_store_f16_hsw+0x61> + .byte 114,209 // jb 3209 <_sk_store_f16_hsw+0x61> .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) - .byte 116,201 // je 3259 <_sk_store_f16_hsw+0x61> + .byte 116,201 // je 3209 <_sk_store_f16_hsw+0x61> .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,189 // jb 3259 <_sk_store_f16_hsw+0x61> + .byte 114,189 // jb 3209 <_sk_store_f16_hsw+0x61> .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) - .byte 235,181 // jmp 3259 <_sk_store_f16_hsw+0x61> + .byte 235,181 // jmp 3209 <_sk_store_f16_hsw+0x61> HIDDEN _sk_load_u16_be_hsw .globl _sk_load_u16_be_hsw @@ -11000,7 +10956,7 @@ _sk_load_u16_be_hsw: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,205,0,0,0 // jne 3387 <_sk_load_u16_be_hsw+0xe3> + .byte 15,133,205,0,0,0 // jne 3337 <_sk_load_u16_be_hsw+0xe3> .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -11049,29 +11005,29 @@ _sk_load_u16_be_hsw: .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 33ed <_sk_load_u16_be_hsw+0x149> + .byte 116,85 // je 339d <_sk_load_u16_be_hsw+0x149> .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 33ed <_sk_load_u16_be_hsw+0x149> + .byte 114,72 // jb 339d <_sk_load_u16_be_hsw+0x149> .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 33fa <_sk_load_u16_be_hsw+0x156> + .byte 116,72 // je 33aa <_sk_load_u16_be_hsw+0x156> .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 33fa <_sk_load_u16_be_hsw+0x156> + .byte 114,59 // jb 33aa <_sk_load_u16_be_hsw+0x156> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,5,255,255,255 // je 32d5 <_sk_load_u16_be_hsw+0x31> + .byte 15,132,5,255,255,255 // je 3285 <_sk_load_u16_be_hsw+0x31> .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,244,254,255,255 // jb 32d5 <_sk_load_u16_be_hsw+0x31> + .byte 15,130,244,254,255,255 // jb 3285 <_sk_load_u16_be_hsw+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,232,254,255,255 // jmpq 32d5 <_sk_load_u16_be_hsw+0x31> + .byte 233,232,254,255,255 // jmpq 3285 <_sk_load_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,219,254,255,255 // jmpq 32d5 <_sk_load_u16_be_hsw+0x31> + .byte 233,219,254,255,255 // jmpq 3285 <_sk_load_u16_be_hsw+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,210,254,255,255 // jmpq 32d5 <_sk_load_u16_be_hsw+0x31> + .byte 233,210,254,255,255 // jmpq 3285 <_sk_load_u16_be_hsw+0x31> HIDDEN _sk_load_rgb_u16_be_hsw .globl _sk_load_rgb_u16_be_hsw @@ -11081,7 +11037,7 @@ _sk_load_rgb_u16_be_hsw: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,211,0,0,0 // jne 34e8 <_sk_load_rgb_u16_be_hsw+0xe5> + .byte 15,133,211,0,0,0 // jne 3498 <_sk_load_rgb_u16_be_hsw+0xe5> .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -11131,36 +11087,36 @@ _sk_load_rgb_u16_be_hsw: .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 3501 <_sk_load_rgb_u16_be_hsw+0xfe> - .byte 233,72,255,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,5 // jne 34b1 <_sk_load_rgb_u16_be_hsw+0xfe> + .byte 233,72,255,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 3530 <_sk_load_rgb_u16_be_hsw+0x12d> + .byte 114,26 // jb 34e0 <_sk_load_rgb_u16_be_hsw+0x12d> .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 3535 <_sk_load_rgb_u16_be_hsw+0x132> - .byte 233,25,255,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,20,255,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 34e5 <_sk_load_rgb_u16_be_hsw+0x132> + .byte 233,25,255,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,20,255,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 3564 <_sk_load_rgb_u16_be_hsw+0x161> + .byte 114,26 // jb 3514 <_sk_load_rgb_u16_be_hsw+0x161> .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 3569 <_sk_load_rgb_u16_be_hsw+0x166> - .byte 233,229,254,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,224,254,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 117,10 // jne 3519 <_sk_load_rgb_u16_be_hsw+0x166> + .byte 233,229,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,224,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 3592 <_sk_load_rgb_u16_be_hsw+0x18f> + .byte 114,20 // jb 3542 <_sk_load_rgb_u16_be_hsw+0x18f> .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - .byte 233,183,254,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> - .byte 233,178,254,255,255 // jmpq 3449 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,183,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> + .byte 233,178,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> HIDDEN _sk_store_u16_be_hsw .globl _sk_store_u16_be_hsw @@ -11209,7 +11165,7 @@ _sk_store_u16_be_hsw: .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,31 // jne 3692 <_sk_store_u16_be_hsw+0xfb> + .byte 117,31 // jne 3642 <_sk_store_u16_be_hsw+0xfb> .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2) .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2) .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2) @@ -11218,22 +11174,22 @@ _sk_store_u16_be_hsw: .byte 255,224 // jmpq *%rax .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 368e <_sk_store_u16_be_hsw+0xf7> + .byte 116,240 // je 363e <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 368e <_sk_store_u16_be_hsw+0xf7> + .byte 114,227 // jb 363e <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2) - .byte 116,218 // je 368e <_sk_store_u16_be_hsw+0xf7> + .byte 116,218 // je 363e <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 368e <_sk_store_u16_be_hsw+0xf7> + .byte 114,205 // jb 363e <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2) - .byte 116,196 // je 368e <_sk_store_u16_be_hsw+0xf7> + .byte 116,196 // je 363e <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 368e <_sk_store_u16_be_hsw+0xf7> + .byte 114,183 // jb 363e <_sk_store_u16_be_hsw+0xf7> .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2) - .byte 235,174 // jmp 368e <_sk_store_u16_be_hsw+0xf7> + .byte 235,174 // jmp 363e <_sk_store_u16_be_hsw+0xf7> HIDDEN _sk_load_f32_hsw .globl _sk_load_f32_hsw @@ -11241,10 +11197,10 @@ FUNCTION(_sk_load_f32_hsw) _sk_load_f32_hsw: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 119,110 // ja 3756 <_sk_load_f32_hsw+0x76> + .byte 119,110 // ja 3706 <_sk_load_f32_hsw+0x76> .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 - .byte 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 3780 <_sk_load_f32_hsw+0xa0> + .byte 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 3730 <_sk_load_f32_hsw+0xa0> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -11303,7 +11259,7 @@ _sk_store_f32_hsw: .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 .byte 72,133,201 // test %rcx,%rcx - .byte 117,55 // jne 380d <_sk_store_f32_hsw+0x6d> + .byte 117,55 // jne 37bd <_sk_store_f32_hsw+0x6d> .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -11316,22 +11272,22 @@ _sk_store_f32_hsw: .byte 255,224 // jmpq *%rax .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 3809 <_sk_store_f32_hsw+0x69> + .byte 116,240 // je 37b9 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 3809 <_sk_store_f32_hsw+0x69> + .byte 114,227 // jb 37b9 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 3809 <_sk_store_f32_hsw+0x69> + .byte 116,218 // je 37b9 <_sk_store_f32_hsw+0x69> .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 3809 <_sk_store_f32_hsw+0x69> + .byte 114,205 // jb 37b9 <_sk_store_f32_hsw+0x69> .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - .byte 116,195 // je 3809 <_sk_store_f32_hsw+0x69> + .byte 116,195 // je 37b9 <_sk_store_f32_hsw+0x69> .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,181 // jb 3809 <_sk_store_f32_hsw+0x69> + .byte 114,181 // jb 37b9 <_sk_store_f32_hsw+0x69> .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - .byte 235,171 // jmp 3809 <_sk_store_f32_hsw+0x69> + .byte 235,171 // jmp 37b9 <_sk_store_f32_hsw+0x69> HIDDEN _sk_clamp_x_hsw .globl _sk_clamp_x_hsw @@ -11596,7 +11552,7 @@ _sk_linear_gradient_hsw: .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 .byte 76,139,0 // mov (%rax),%r8 .byte 77,133,192 // test %r8,%r8 - .byte 15,132,143,0,0,0 // je 3c99 <_sk_linear_gradient_hsw+0xb5> + .byte 15,132,143,0,0,0 // je 3c49 <_sk_linear_gradient_hsw+0xb5> .byte 72,139,64,8 // mov 0x8(%rax),%rax .byte 72,131,192,32 // add $0x20,%rax .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 @@ -11623,8 +11579,8 @@ _sk_linear_gradient_hsw: .byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9 .byte 72,131,192,36 // add $0x24,%rax .byte 73,255,200 // dec %r8 - .byte 117,140 // jne 3c23 <_sk_linear_gradient_hsw+0x3f> - .byte 235,17 // jmp 3caa <_sk_linear_gradient_hsw+0xc6> + .byte 117,140 // jne 3bd3 <_sk_linear_gradient_hsw+0x3f> + .byte 235,17 // jmp 3c5a <_sk_linear_gradient_hsw+0xc6> .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 @@ -14653,52 +14609,47 @@ _sk_parametric_r_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 - .byte 196,65,124,88,219 // vaddps %ymm11,%ymm0,%ymm11 - .byte 196,65,124,91,211 // vcvtdq2ps %ymm11,%ymm10 + .byte 196,65,124,88,211 // vaddps %ymm11,%ymm0,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,44,89,208 // vmulps %ymm0,%ymm10,%ymm10 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 - .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10 + .byte 197,36,89,216 // vmulps %ymm0,%ymm11,%ymm11 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 196,65,124,84,219 // vandps %ymm11,%ymm0,%ymm11 + .byte 196,65,124,84,210 // vandps %ymm10,%ymm0,%ymm10 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,36,86,216 // vorps %ymm0,%ymm11,%ymm11 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 197,44,86,208 // vorps %ymm0,%ymm10,%ymm10 + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,44,88,208 // vaddps %ymm0,%ymm10,%ymm10 + .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0 - .byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10 + .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 + .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 - .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 + .byte 196,99,125,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 - .byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0 - .byte 197,172,92,192 // vsubps %ymm0,%ymm10,%ymm0 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 + .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0 + .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 + .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 + .byte 197,28,89,216 // vmulps %ymm0,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -14756,52 +14707,47 @@ _sk_parametric_g_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 - .byte 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11 - .byte 196,65,124,91,211 // vcvtdq2ps %ymm11,%ymm10 + .byte 196,65,116,88,211 // vaddps %ymm11,%ymm1,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 - .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,44,92,209 // vsubps %ymm1,%ymm10,%ymm10 + .byte 197,36,89,217 // vmulps %ymm1,%ymm11,%ymm11 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 196,65,116,84,219 // vandps %ymm11,%ymm1,%ymm11 + .byte 196,65,116,84,210 // vandps %ymm10,%ymm1,%ymm10 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,36,86,217 // vorps %ymm1,%ymm11,%ymm11 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 197,44,86,209 // vorps %ymm1,%ymm10,%ymm10 + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,44,88,209 // vaddps %ymm1,%ymm10,%ymm10 + .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 - .byte 197,44,92,209 // vsubps %ymm1,%ymm10,%ymm10 + .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 + .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 - .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 + .byte 196,99,117,24,233,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 - .byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1 - .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 + .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 + .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 + .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 + .byte 197,28,89,217 // vmulps %ymm1,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -14859,52 +14805,47 @@ _sk_parametric_b_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 - .byte 196,65,108,88,219 // vaddps %ymm11,%ymm2,%ymm11 - .byte 196,65,124,91,211 // vcvtdq2ps %ymm11,%ymm10 + .byte 196,65,108,88,211 // vaddps %ymm11,%ymm2,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 - .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,44,92,210 // vsubps %ymm2,%ymm10,%ymm10 + .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 196,65,108,84,219 // vandps %ymm11,%ymm2,%ymm11 + .byte 196,65,108,84,210 // vandps %ymm10,%ymm2,%ymm10 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,36,86,218 // vorps %ymm2,%ymm11,%ymm11 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 197,44,86,210 // vorps %ymm2,%ymm10,%ymm10 + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,44,88,210 // vaddps %ymm2,%ymm10,%ymm10 + .byte 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,164,89,210 // vmulps %ymm2,%ymm11,%ymm2 - .byte 197,44,92,210 // vsubps %ymm2,%ymm10,%ymm10 + .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 + .byte 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 - .byte 196,99,109,24,226,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm12 + .byte 196,99,109,24,234,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 - .byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2 - .byte 197,172,92,210 // vsubps %ymm2,%ymm10,%ymm2 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,218 // vmulps %ymm2,%ymm10,%ymm11 + .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 + .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 + .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 + .byte 197,28,89,218 // vmulps %ymm2,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -14962,52 +14903,47 @@ _sk_parametric_a_avx: .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 - .byte 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 - .byte 196,65,124,91,211 // vcvtdq2ps %ymm11,%ymm10 + .byte 196,65,100,88,211 // vaddps %ymm11,%ymm3,%ymm10 + .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 + .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10 - .byte 65,184,0,0,254,66 // mov $0x42fe0000,%r8d - .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 - .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 - .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,44,92,211 // vsubps %ymm3,%ymm10,%ymm10 + .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 196,65,100,84,219 // vandps %ymm11,%ymm3,%ymm11 + .byte 196,65,100,84,210 // vandps %ymm10,%ymm3,%ymm10 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,36,86,219 // vorps %ymm3,%ymm11,%ymm11 - .byte 65,184,42,145,49,64 // mov $0x4031912a,%r8d + .byte 197,44,86,211 // vorps %ymm3,%ymm10,%ymm10 + .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,44,88,211 // vaddps %ymm3,%ymm10,%ymm10 + .byte 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,164,89,219 // vmulps %ymm3,%ymm11,%ymm3 - .byte 197,44,92,211 // vsubps %ymm3,%ymm10,%ymm10 + .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 + .byte 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 - .byte 196,99,101,24,227,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm12 + .byte 196,99,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 - .byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3 - .byte 197,172,92,219 // vsubps %ymm3,%ymm10,%ymm3 - .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 - .byte 197,44,89,219 // vmulps %ymm3,%ymm10,%ymm11 + .byte 197,172,88,219 // vaddps %ymm3,%ymm10,%ymm3 + .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 + .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 + .byte 197,28,89,219 // vmulps %ymm3,%ymm12,%ymm11 .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d @@ -15060,7 +14996,7 @@ _sk_load_a8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,74 // jne 310a <_sk_load_a8_avx+0x5a> + .byte 117,74 // jne 309e <_sk_load_a8_avx+0x5a> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 @@ -15087,9 +15023,9 @@ _sk_load_a8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 3112 <_sk_load_a8_avx+0x62> + .byte 117,234 // jne 30a6 <_sk_load_a8_avx+0x62> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,149 // jmp 30c4 <_sk_load_a8_avx+0x14> + .byte 235,149 // jmp 3058 <_sk_load_a8_avx+0x14> HIDDEN _sk_gather_a8_avx .globl _sk_gather_a8_avx @@ -15170,7 +15106,7 @@ _sk_store_a8_avx: .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 326b <_sk_store_a8_avx+0x42> + .byte 117,10 // jne 31ff <_sk_store_a8_avx+0x42> .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15178,10 +15114,10 @@ _sk_store_a8_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 3267 <_sk_store_a8_avx+0x3e> + .byte 119,236 // ja 31fb <_sk_store_a8_avx+0x3e> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 32d0 <_sk_store_a8_avx+0xa7> + .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3264 <_sk_store_a8_avx+0xa7> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -15192,7 +15128,7 @@ _sk_store_a8_avx: .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) - .byte 235,154 // jmp 3267 <_sk_store_a8_avx+0x3e> + .byte 235,154 // jmp 31fb <_sk_store_a8_avx+0x3e> .byte 15,31,0 // nopl (%rax) .byte 244 // hlt .byte 255 // (bad) @@ -15227,7 +15163,7 @@ _sk_load_g8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,91 // jne 3357 <_sk_load_g8_avx+0x6b> + .byte 117,91 // jne 32eb <_sk_load_g8_avx+0x6b> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 @@ -15257,9 +15193,9 @@ _sk_load_g8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 335f <_sk_load_g8_avx+0x73> + .byte 117,234 // jne 32f3 <_sk_load_g8_avx+0x73> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,132 // jmp 3300 <_sk_load_g8_avx+0x14> + .byte 235,132 // jmp 3294 <_sk_load_g8_avx+0x14> HIDDEN _sk_gather_g8_avx .globl _sk_gather_g8_avx @@ -15334,9 +15270,9 @@ _sk_gather_i8_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 3496 <_sk_gather_i8_avx+0xf> + .byte 116,5 // je 342a <_sk_gather_i8_avx+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 3498 <_sk_gather_i8_avx+0x11> + .byte 235,2 // jmp 342c <_sk_gather_i8_avx+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -15441,7 +15377,7 @@ _sk_load_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,209,0,0,0 // jne 3732 <_sk_load_565_avx+0xdf> + .byte 15,133,209,0,0,0 // jne 36c6 <_sk_load_565_avx+0xdf> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -15491,9 +15427,9 @@ _sk_load_565_avx: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,29,255,255,255 // ja 3667 <_sk_load_565_avx+0x14> + .byte 15,135,29,255,255,255 // ja 35fb <_sk_load_565_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 37a0 <_sk_load_565_avx+0x14d> + .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3734 <_sk_load_565_avx+0x14d> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15505,7 +15441,7 @@ _sk_load_565_avx: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,201,254,255,255 // jmpq 3667 <_sk_load_565_avx+0x14> + .byte 233,201,254,255,255 // jmpq 35fb <_sk_load_565_avx+0x14> .byte 102,144 // xchg %ax,%ax .byte 242,255 // repnz (bad) .byte 255 // (bad) @@ -15662,7 +15598,7 @@ _sk_store_565_avx: .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 39eb <_sk_store_565_avx+0x9e> + .byte 117,10 // jne 397f <_sk_store_565_avx+0x9e> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15670,9 +15606,9 @@ _sk_store_565_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 39e7 <_sk_store_565_avx+0x9a> + .byte 119,236 // ja 397b <_sk_store_565_avx+0x9a> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 3a48 <_sk_store_565_avx+0xfb> + .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 39dc <_sk_store_565_avx+0xfb> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -15683,7 +15619,7 @@ _sk_store_565_avx: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 39e7 <_sk_store_565_avx+0x9a> + .byte 235,159 // jmp 397b <_sk_store_565_avx+0x9a> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -15714,7 +15650,7 @@ _sk_load_4444_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,245,0,0,0 // jne 3b67 <_sk_load_4444_avx+0x103> + .byte 15,133,245,0,0,0 // jne 3afb <_sk_load_4444_avx+0x103> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -15771,9 +15707,9 @@ _sk_load_4444_avx: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,249,254,255,255 // ja 3a78 <_sk_load_4444_avx+0x14> + .byte 15,135,249,254,255,255 // ja 3a0c <_sk_load_4444_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 3bd4 <_sk_load_4444_avx+0x170> + .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 3b68 <_sk_load_4444_avx+0x170> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -15785,12 +15721,12 @@ _sk_load_4444_avx: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,165,254,255,255 // jmpq 3a78 <_sk_load_4444_avx+0x14> + .byte 233,165,254,255,255 // jmpq 3a0c <_sk_load_4444_avx+0x14> .byte 144 // nop .byte 243,255 // repz (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 235,255 // jmp 3bd9 <_sk_load_4444_avx+0x175> + .byte 235,255 // jmp 3b6d <_sk_load_4444_avx+0x175> .byte 255 // (bad) .byte 255,227 // jmpq *%rbx .byte 255 // (bad) @@ -15951,7 +15887,7 @@ _sk_store_4444_avx: .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 3e54 <_sk_store_4444_avx+0xaf> + .byte 117,10 // jne 3de8 <_sk_store_4444_avx+0xaf> .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -15959,9 +15895,9 @@ _sk_store_4444_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 3e50 <_sk_store_4444_avx+0xab> + .byte 119,236 // ja 3de4 <_sk_store_4444_avx+0xab> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3eb4 <_sk_store_4444_avx+0x10f> + .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3e48 <_sk_store_4444_avx+0x10f> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -15972,7 +15908,7 @@ _sk_store_4444_avx: .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) - .byte 235,159 // jmp 3e50 <_sk_store_4444_avx+0xab> + .byte 235,159 // jmp 3de4 <_sk_store_4444_avx+0xab> .byte 15,31,0 // nopl (%rax) .byte 244 // hlt .byte 255 // (bad) @@ -16005,7 +15941,7 @@ _sk_load_8888_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,157,0,0,0 // jne 3f7b <_sk_load_8888_avx+0xab> + .byte 15,133,157,0,0,0 // jne 3f0f <_sk_load_8888_avx+0xab> .byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9 .byte 184,255,0,0,0 // mov $0xff,%eax .byte 197,249,110,192 // vmovd %eax,%xmm0 @@ -16043,9 +15979,9 @@ _sk_load_8888_avx: .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,80,255,255,255 // ja 3ee4 <_sk_load_8888_avx+0x14> + .byte 15,135,80,255,255,255 // ja 3e78 <_sk_load_8888_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 4028 <_sk_load_8888_avx+0x158> + .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 3fbc <_sk_load_8888_avx+0x158> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -16068,7 +16004,7 @@ _sk_load_8888_avx: .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 .byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 - .byte 233,188,254,255,255 // jmpq 3ee4 <_sk_load_8888_avx+0x14> + .byte 233,188,254,255,255 // jmpq 3e78 <_sk_load_8888_avx+0x14> .byte 238 // out %al,(%dx) .byte 255 // (bad) .byte 255 // (bad) @@ -16198,7 +16134,7 @@ _sk_store_8888_avx: .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8 .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 4229 <_sk_store_8888_avx+0xa4> + .byte 117,10 // jne 41bd <_sk_store_8888_avx+0xa4> .byte 196,65,124,17,4,185 // vmovups %ymm8,(%r9,%rdi,4) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -16206,9 +16142,9 @@ _sk_store_8888_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 4225 <_sk_store_8888_avx+0xa0> + .byte 119,236 // ja 41b9 <_sk_store_8888_avx+0xa0> .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,84,0,0,0 // lea 0x54(%rip),%r8 # 4298 <_sk_store_8888_avx+0x113> + .byte 76,141,5,84,0,0,0 // lea 0x54(%rip),%r8 # 422c <_sk_store_8888_avx+0x113> .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax .byte 76,1,192 // add %r8,%rax .byte 255,224 // jmpq *%rax @@ -16222,7 +16158,7 @@ _sk_store_8888_avx: .byte 196,67,121,22,68,185,8,2 // vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) .byte 196,67,121,22,68,185,4,1 // vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) .byte 196,65,121,126,4,185 // vmovd %xmm8,(%r9,%rdi,4) - .byte 235,143 // jmp 4225 <_sk_store_8888_avx+0xa0> + .byte 235,143 // jmp 41b9 <_sk_store_8888_avx+0xa0> .byte 102,144 // xchg %ax,%ax .byte 246,255 // idiv %bh .byte 255 // (bad) @@ -16254,7 +16190,7 @@ _sk_load_f16_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,17,1,0,0 // jne 43d3 <_sk_load_f16_avx+0x11f> + .byte 15,133,17,1,0,0 // jne 4367 <_sk_load_f16_avx+0x11f> .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -16316,29 +16252,29 @@ _sk_load_f16_avx: .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,79 // je 4432 <_sk_load_f16_avx+0x17e> + .byte 116,79 // je 43c6 <_sk_load_f16_avx+0x17e> .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,67 // jb 4432 <_sk_load_f16_avx+0x17e> + .byte 114,67 // jb 43c6 <_sk_load_f16_avx+0x17e> .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,68 // je 443f <_sk_load_f16_avx+0x18b> + .byte 116,68 // je 43d3 <_sk_load_f16_avx+0x18b> .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,56 // jb 443f <_sk_load_f16_avx+0x18b> + .byte 114,56 // jb 43d3 <_sk_load_f16_avx+0x18b> .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,194,254,255,255 // je 42d9 <_sk_load_f16_avx+0x25> + .byte 15,132,194,254,255,255 // je 426d <_sk_load_f16_avx+0x25> .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,178,254,255,255 // jb 42d9 <_sk_load_f16_avx+0x25> + .byte 15,130,178,254,255,255 // jb 426d <_sk_load_f16_avx+0x25> .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 - .byte 233,167,254,255,255 // jmpq 42d9 <_sk_load_f16_avx+0x25> + .byte 233,167,254,255,255 // jmpq 426d <_sk_load_f16_avx+0x25> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,154,254,255,255 // jmpq 42d9 <_sk_load_f16_avx+0x25> + .byte 233,154,254,255,255 // jmpq 426d <_sk_load_f16_avx+0x25> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,145,254,255,255 // jmpq 42d9 <_sk_load_f16_avx+0x25> + .byte 233,145,254,255,255 // jmpq 426d <_sk_load_f16_avx+0x25> HIDDEN _sk_gather_f16_avx .globl _sk_gather_f16_avx @@ -16482,7 +16418,7 @@ _sk_store_f16_avx: .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,31 // jne 46c8 <_sk_store_f16_avx+0xd2> + .byte 117,31 // jne 465c <_sk_store_f16_avx+0xd2> .byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8) .byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8) .byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8) @@ -16491,22 +16427,22 @@ _sk_store_f16_avx: .byte 255,224 // jmpq *%rax .byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 46c4 <_sk_store_f16_avx+0xce> + .byte 116,240 // je 4658 <_sk_store_f16_avx+0xce> .byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 46c4 <_sk_store_f16_avx+0xce> + .byte 114,227 // jb 4658 <_sk_store_f16_avx+0xce> .byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8) - .byte 116,218 // je 46c4 <_sk_store_f16_avx+0xce> + .byte 116,218 // je 4658 <_sk_store_f16_avx+0xce> .byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 46c4 <_sk_store_f16_avx+0xce> + .byte 114,205 // jb 4658 <_sk_store_f16_avx+0xce> .byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8) - .byte 116,196 // je 46c4 <_sk_store_f16_avx+0xce> + .byte 116,196 // je 4658 <_sk_store_f16_avx+0xce> .byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 46c4 <_sk_store_f16_avx+0xce> + .byte 114,183 // jb 4658 <_sk_store_f16_avx+0xce> .byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8) - .byte 235,174 // jmp 46c4 <_sk_store_f16_avx+0xce> + .byte 235,174 // jmp 4658 <_sk_store_f16_avx+0xce> HIDDEN _sk_load_u16_be_avx .globl _sk_load_u16_be_avx @@ -16516,7 +16452,7 @@ _sk_load_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,5,1,0,0 // jne 4831 <_sk_load_u16_be_avx+0x11b> + .byte 15,133,5,1,0,0 // jne 47c5 <_sk_load_u16_be_avx+0x11b> .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -16575,29 +16511,29 @@ _sk_load_u16_be_avx: .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 4897 <_sk_load_u16_be_avx+0x181> + .byte 116,85 // je 482b <_sk_load_u16_be_avx+0x181> .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 4897 <_sk_load_u16_be_avx+0x181> + .byte 114,72 // jb 482b <_sk_load_u16_be_avx+0x181> .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 48a4 <_sk_load_u16_be_avx+0x18e> + .byte 116,72 // je 4838 <_sk_load_u16_be_avx+0x18e> .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 48a4 <_sk_load_u16_be_avx+0x18e> + .byte 114,59 // jb 4838 <_sk_load_u16_be_avx+0x18e> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,205,254,255,255 // je 4747 <_sk_load_u16_be_avx+0x31> + .byte 15,132,205,254,255,255 // je 46db <_sk_load_u16_be_avx+0x31> .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,188,254,255,255 // jb 4747 <_sk_load_u16_be_avx+0x31> + .byte 15,130,188,254,255,255 // jb 46db <_sk_load_u16_be_avx+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,176,254,255,255 // jmpq 4747 <_sk_load_u16_be_avx+0x31> + .byte 233,176,254,255,255 // jmpq 46db <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,163,254,255,255 // jmpq 4747 <_sk_load_u16_be_avx+0x31> + .byte 233,163,254,255,255 // jmpq 46db <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,154,254,255,255 // jmpq 4747 <_sk_load_u16_be_avx+0x31> + .byte 233,154,254,255,255 // jmpq 46db <_sk_load_u16_be_avx+0x31> HIDDEN _sk_load_rgb_u16_be_avx .globl _sk_load_rgb_u16_be_avx @@ -16607,7 +16543,7 @@ _sk_load_rgb_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,8,1,0,0 // jne 49c7 <_sk_load_rgb_u16_be_avx+0x11a> + .byte 15,133,8,1,0,0 // jne 495b <_sk_load_rgb_u16_be_avx+0x11a> .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -16666,36 +16602,36 @@ _sk_load_rgb_u16_be_avx: .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 49e0 <_sk_load_rgb_u16_be_avx+0x133> - .byte 233,19,255,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,5 // jne 4974 <_sk_load_rgb_u16_be_avx+0x133> + .byte 233,19,255,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 4a0f <_sk_load_rgb_u16_be_avx+0x162> + .byte 114,26 // jb 49a3 <_sk_load_rgb_u16_be_avx+0x162> .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 4a14 <_sk_load_rgb_u16_be_avx+0x167> - .byte 233,228,254,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,223,254,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 49a8 <_sk_load_rgb_u16_be_avx+0x167> + .byte 233,228,254,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,223,254,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 4a43 <_sk_load_rgb_u16_be_avx+0x196> + .byte 114,26 // jb 49d7 <_sk_load_rgb_u16_be_avx+0x196> .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 4a48 <_sk_load_rgb_u16_be_avx+0x19b> - .byte 233,176,254,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,171,254,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 49dc <_sk_load_rgb_u16_be_avx+0x19b> + .byte 233,176,254,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,171,254,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 4a71 <_sk_load_rgb_u16_be_avx+0x1c4> + .byte 114,20 // jb 4a05 <_sk_load_rgb_u16_be_avx+0x1c4> .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - .byte 233,130,254,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,125,254,255,255 // jmpq 48f3 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,130,254,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,125,254,255,255 // jmpq 4887 <_sk_load_rgb_u16_be_avx+0x46> HIDDEN _sk_store_u16_be_avx .globl _sk_store_u16_be_avx @@ -16745,7 +16681,7 @@ _sk_store_u16_be_avx: .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,31 // jne 4b78 <_sk_store_u16_be_avx+0x102> + .byte 117,31 // jne 4b0c <_sk_store_u16_be_avx+0x102> .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2) .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2) .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2) @@ -16754,22 +16690,22 @@ _sk_store_u16_be_avx: .byte 255,224 // jmpq *%rax .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 116,240 // je 4b08 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 114,227 // jb 4b08 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2) - .byte 116,218 // je 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 116,218 // je 4b08 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 114,205 // jb 4b08 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2) - .byte 116,196 // je 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 116,196 // je 4b08 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 114,183 // jb 4b08 <_sk_store_u16_be_avx+0xfe> .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2) - .byte 235,174 // jmp 4b74 <_sk_store_u16_be_avx+0xfe> + .byte 235,174 // jmp 4b08 <_sk_store_u16_be_avx+0xfe> HIDDEN _sk_load_f32_avx .globl _sk_load_f32_avx @@ -16777,10 +16713,10 @@ FUNCTION(_sk_load_f32_avx) _sk_load_f32_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 119,110 // ja 4c3c <_sk_load_f32_avx+0x76> + .byte 119,110 // ja 4bd0 <_sk_load_f32_avx+0x76> .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 - .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 4c64 <_sk_load_f32_avx+0x9e> + .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 4bf8 <_sk_load_f32_avx+0x9e> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -16839,7 +16775,7 @@ _sk_store_f32_avx: .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 .byte 72,133,201 // test %rcx,%rcx - .byte 117,55 // jne 4cf1 <_sk_store_f32_avx+0x6d> + .byte 117,55 // jne 4c85 <_sk_store_f32_avx+0x6d> .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -16852,22 +16788,22 @@ _sk_store_f32_avx: .byte 255,224 // jmpq *%rax .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 4ced <_sk_store_f32_avx+0x69> + .byte 116,240 // je 4c81 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 4ced <_sk_store_f32_avx+0x69> + .byte 114,227 // jb 4c81 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 4ced <_sk_store_f32_avx+0x69> + .byte 116,218 // je 4c81 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 4ced <_sk_store_f32_avx+0x69> + .byte 114,205 // jb 4c81 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - .byte 116,195 // je 4ced <_sk_store_f32_avx+0x69> + .byte 116,195 // je 4c81 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,181 // jb 4ced <_sk_store_f32_avx+0x69> + .byte 114,181 // jb 4c81 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - .byte 235,171 // jmp 4ced <_sk_store_f32_avx+0x69> + .byte 235,171 // jmp 4c81 <_sk_store_f32_avx+0x69> HIDDEN _sk_clamp_x_avx .globl _sk_clamp_x_avx @@ -17195,7 +17131,7 @@ _sk_linear_gradient_avx: .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3 .byte 76,139,0 // mov (%rax),%r8 .byte 77,133,192 // test %r8,%r8 - .byte 15,132,146,0,0,0 // je 52a5 <_sk_linear_gradient_avx+0xb8> + .byte 15,132,146,0,0,0 // je 5239 <_sk_linear_gradient_avx+0xb8> .byte 72,139,64,8 // mov 0x8(%rax),%rax .byte 72,131,192,32 // add $0x20,%rax .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 @@ -17222,8 +17158,8 @@ _sk_linear_gradient_avx: .byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3 .byte 72,131,192,36 // add $0x24,%rax .byte 73,255,200 // dec %r8 - .byte 117,140 // jne 522f <_sk_linear_gradient_avx+0x42> - .byte 235,20 // jmp 52b9 <_sk_linear_gradient_avx+0xcc> + .byte 117,140 // jne 51c3 <_sk_linear_gradient_avx+0x42> + .byte 235,20 // jmp 524d <_sk_linear_gradient_avx+0xcc> .byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 @@ -19929,54 +19865,50 @@ _sk_parametric_r_sse41: .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 68,15,89,200 // mulps %xmm0,%xmm9 - .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 68,15,89,208 // mulps %xmm0,%xmm10 - .byte 65,15,194,192,2 // cmpleps %xmm8,%xmm0 - .byte 243,68,15,16,64,24 // movss 0x18(%rax),%xmm8 + .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 68,15,89,216 // mulps %xmm0,%xmm11 + .byte 65,15,194,192,2 // cmpleps %xmm8,%xmm0 + .byte 243,68,15,16,64,24 // movss 0x18(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 69,15,88,200 // addps %xmm8,%xmm9 - .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 243,68,15,16,88,8 // movss 0x8(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 243,68,15,16,16 // movss (%rax),%xmm10 + .byte 243,68,15,16,64,8 // movss 0x8(%rax),%xmm8 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 + .byte 69,15,88,216 // addps %xmm8,%xmm11 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,227 // mulps %xmm11,%xmm12 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 102,68,15,110,193 // movd %ecx,%xmm8 + .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 + .byte 69,15,89,196 // mulps %xmm12,%xmm8 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 - .byte 102,69,15,219,218 // pand %xmm10,%xmm11 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 102,69,15,112,228,0 // pshufd $0x0,%xmm12,%xmm12 + .byte 102,69,15,219,227 // pand %xmm11,%xmm12 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 102,69,15,112,210,0 // pshufd $0x0,%xmm10,%xmm10 - .byte 102,69,15,235,211 // por %xmm11,%xmm10 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,88,220 // addps %xmm12,%xmm11 + .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 + .byte 102,69,15,235,220 // por %xmm12,%xmm11 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,92,196 // subps %xmm12,%xmm8 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx .byte 102,68,15,110,225 // movd %ecx,%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 69,15,92,220 // subps %xmm12,%xmm11 + .byte 69,15,89,227 // mulps %xmm11,%xmm12 + .byte 69,15,92,196 // subps %xmm12,%xmm8 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx .byte 102,68,15,110,225 // movd %ecx,%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,88,234 // addps %xmm10,%xmm13 + .byte 69,15,88,235 // addps %xmm11,%xmm13 .byte 69,15,94,229 // divps %xmm13,%xmm12 - .byte 69,15,92,220 // subps %xmm12,%xmm11 - .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 - .byte 69,15,89,195 // mulps %xmm11,%xmm8 + .byte 69,15,92,196 // subps %xmm12,%xmm8 + .byte 69,15,89,194 // mulps %xmm10,%xmm8 .byte 102,69,15,58,8,216,1 // roundps $0x1,%xmm8,%xmm11 .byte 185,0,0,0,75 // mov $0x4b000000,%ecx .byte 102,68,15,110,209 // movd %ecx,%xmm10 @@ -20035,70 +19967,66 @@ _sk_parametric_g_sse41: .byte 243,15,16,72,24 // movss 0x18(%rax),%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 .byte 68,15,88,201 // addps %xmm1,%xmm9 - .byte 243,68,15,16,32 // movss (%rax),%xmm12 + .byte 243,68,15,16,16 // movss (%rax),%xmm10 .byte 243,15,16,72,8 // movss 0x8(%rax),%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 .byte 68,15,88,217 // addps %xmm1,%xmm11 - .byte 69,15,91,211 // cvtdq2ps %xmm11,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 68,15,92,233 // subps %xmm1,%xmm13 + .byte 69,15,89,236 // mulps %xmm12,%xmm13 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx .byte 102,15,110,201 // movd %ecx,%xmm1 - .byte 102,68,15,112,209,0 // pshufd $0x0,%xmm1,%xmm10 - .byte 102,69,15,219,211 // pand %xmm11,%xmm10 + .byte 102,68,15,112,225,0 // pshufd $0x0,%xmm1,%xmm12 + .byte 102,69,15,219,227 // pand %xmm11,%xmm12 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,15,110,201 // movd %ecx,%xmm1 .byte 102,68,15,112,217,0 // pshufd $0x0,%xmm1,%xmm11 - .byte 102,69,15,235,218 // por %xmm10,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,213 // addps %xmm13,%xmm10 + .byte 102,69,15,235,220 // por %xmm12,%xmm11 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,15,110,201 // movd %ecx,%xmm1 + .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 + .byte 68,15,92,233 // subps %xmm1,%xmm13 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx .byte 102,15,110,201 // movd %ecx,%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 .byte 65,15,89,203 // mulps %xmm11,%xmm1 - .byte 68,15,92,209 // subps %xmm1,%xmm10 + .byte 68,15,92,233 // subps %xmm1,%xmm13 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,15,110,201 // movd %ecx,%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 .byte 65,15,88,203 // addps %xmm11,%xmm1 - .byte 68,15,94,233 // divps %xmm1,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,58,8,220,1 // roundps $0x1,%xmm12,%xmm11 + .byte 68,15,94,225 // divps %xmm1,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 69,15,89,234 // mulps %xmm10,%xmm13 + .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 .byte 185,0,0,0,75 // mov $0x4b000000,%ecx .byte 102,68,15,110,209 // movd %ecx,%xmm10 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,88,236 // addps %xmm12,%xmm13 - .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 69,15,92,235 // subps %xmm11,%xmm13 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx .byte 102,15,110,201 // movd %ecx,%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 65,15,89,204 // mulps %xmm12,%xmm1 - .byte 68,15,92,233 // subps %xmm1,%xmm13 + .byte 65,15,89,205 // mulps %xmm13,%xmm1 + .byte 68,15,92,225 // subps %xmm1,%xmm12 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,15,110,201 // movd %ecx,%xmm1 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 - .byte 65,15,92,204 // subps %xmm12,%xmm1 + .byte 65,15,92,205 // subps %xmm13,%xmm1 .byte 68,15,94,217 // divps %xmm1,%xmm11 - .byte 69,15,88,221 // addps %xmm13,%xmm11 + .byte 69,15,88,220 // addps %xmm12,%xmm11 .byte 69,15,89,218 // mulps %xmm10,%xmm11 .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 .byte 243,15,16,72,20 // movss 0x14(%rax),%xmm1 @@ -20134,70 +20062,66 @@ _sk_parametric_b_sse41: .byte 243,15,16,80,24 // movss 0x18(%rax),%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 .byte 68,15,88,202 // addps %xmm2,%xmm9 - .byte 243,68,15,16,32 // movss (%rax),%xmm12 + .byte 243,68,15,16,16 // movss (%rax),%xmm10 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 .byte 68,15,88,218 // addps %xmm2,%xmm11 - .byte 69,15,91,211 // cvtdq2ps %xmm11,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 68,15,92,234 // subps %xmm2,%xmm13 + .byte 69,15,89,236 // mulps %xmm12,%xmm13 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx .byte 102,15,110,209 // movd %ecx,%xmm2 - .byte 102,68,15,112,210,0 // pshufd $0x0,%xmm2,%xmm10 - .byte 102,69,15,219,211 // pand %xmm11,%xmm10 + .byte 102,68,15,112,226,0 // pshufd $0x0,%xmm2,%xmm12 + .byte 102,69,15,219,227 // pand %xmm11,%xmm12 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,15,110,209 // movd %ecx,%xmm2 .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 - .byte 102,69,15,235,218 // por %xmm10,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,213 // addps %xmm13,%xmm10 + .byte 102,69,15,235,220 // por %xmm12,%xmm11 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,15,110,209 // movd %ecx,%xmm2 + .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 + .byte 68,15,92,234 // subps %xmm2,%xmm13 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx .byte 102,15,110,209 // movd %ecx,%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 .byte 65,15,89,211 // mulps %xmm11,%xmm2 - .byte 68,15,92,210 // subps %xmm2,%xmm10 + .byte 68,15,92,234 // subps %xmm2,%xmm13 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,15,110,209 // movd %ecx,%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 .byte 65,15,88,211 // addps %xmm11,%xmm2 - .byte 68,15,94,234 // divps %xmm2,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,58,8,220,1 // roundps $0x1,%xmm12,%xmm11 + .byte 68,15,94,226 // divps %xmm2,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 69,15,89,234 // mulps %xmm10,%xmm13 + .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 .byte 185,0,0,0,75 // mov $0x4b000000,%ecx .byte 102,68,15,110,209 // movd %ecx,%xmm10 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,88,236 // addps %xmm12,%xmm13 - .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 69,15,92,235 // subps %xmm11,%xmm13 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx .byte 102,15,110,209 // movd %ecx,%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 65,15,89,212 // mulps %xmm12,%xmm2 - .byte 68,15,92,234 // subps %xmm2,%xmm13 + .byte 65,15,89,213 // mulps %xmm13,%xmm2 + .byte 68,15,92,226 // subps %xmm2,%xmm12 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,15,110,209 // movd %ecx,%xmm2 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 - .byte 65,15,92,212 // subps %xmm12,%xmm2 + .byte 65,15,92,213 // subps %xmm13,%xmm2 .byte 68,15,94,218 // divps %xmm2,%xmm11 - .byte 69,15,88,221 // addps %xmm13,%xmm11 + .byte 69,15,88,220 // addps %xmm12,%xmm11 .byte 69,15,89,218 // mulps %xmm10,%xmm11 .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2 @@ -20233,70 +20157,66 @@ _sk_parametric_a_sse41: .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 68,15,88,203 // addps %xmm3,%xmm9 - .byte 243,68,15,16,32 // movss (%rax),%xmm12 + .byte 243,68,15,16,16 // movss (%rax),%xmm10 .byte 243,15,16,88,8 // movss 0x8(%rax),%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 68,15,88,219 // addps %xmm3,%xmm11 - .byte 69,15,91,211 // cvtdq2ps %xmm11,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,234 // mulps %xmm10,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 68,15,92,235 // subps %xmm3,%xmm13 + .byte 69,15,89,236 // mulps %xmm12,%xmm13 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx .byte 102,15,110,217 // movd %ecx,%xmm3 - .byte 102,68,15,112,211,0 // pshufd $0x0,%xmm3,%xmm10 - .byte 102,69,15,219,211 // pand %xmm11,%xmm10 + .byte 102,68,15,112,227,0 // pshufd $0x0,%xmm3,%xmm12 + .byte 102,69,15,219,227 // pand %xmm11,%xmm12 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,15,110,217 // movd %ecx,%xmm3 .byte 102,68,15,112,219,0 // pshufd $0x0,%xmm3,%xmm11 - .byte 102,69,15,235,218 // por %xmm10,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,88,213 // addps %xmm13,%xmm10 + .byte 102,69,15,235,220 // por %xmm12,%xmm11 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,15,110,217 // movd %ecx,%xmm3 + .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 + .byte 68,15,92,235 // subps %xmm3,%xmm13 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx .byte 102,15,110,217 // movd %ecx,%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 65,15,89,219 // mulps %xmm11,%xmm3 - .byte 68,15,92,211 // subps %xmm3,%xmm10 + .byte 68,15,92,235 // subps %xmm3,%xmm13 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,15,110,217 // movd %ecx,%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 .byte 65,15,88,219 // addps %xmm11,%xmm3 - .byte 68,15,94,235 // divps %xmm3,%xmm13 - .byte 69,15,92,213 // subps %xmm13,%xmm10 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,89,226 // mulps %xmm10,%xmm12 - .byte 102,69,15,58,8,220,1 // roundps $0x1,%xmm12,%xmm11 + .byte 68,15,94,227 // divps %xmm3,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 69,15,89,234 // mulps %xmm10,%xmm13 + .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 .byte 185,0,0,0,75 // mov $0x4b000000,%ecx .byte 102,68,15,110,209 // movd %ecx,%xmm10 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,88,236 // addps %xmm12,%xmm13 - .byte 69,15,92,227 // subps %xmm11,%xmm12 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 69,15,92,235 // subps %xmm11,%xmm13 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx .byte 102,15,110,217 // movd %ecx,%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 65,15,89,220 // mulps %xmm12,%xmm3 - .byte 68,15,92,235 // subps %xmm3,%xmm13 + .byte 65,15,89,221 // mulps %xmm13,%xmm3 + .byte 68,15,92,227 // subps %xmm3,%xmm12 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,15,110,217 // movd %ecx,%xmm3 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 - .byte 65,15,92,220 // subps %xmm12,%xmm3 + .byte 65,15,92,221 // subps %xmm13,%xmm3 .byte 68,15,94,219 // divps %xmm3,%xmm11 - .byte 69,15,88,221 // addps %xmm13,%xmm11 + .byte 69,15,88,220 // addps %xmm12,%xmm11 .byte 69,15,89,218 // mulps %xmm10,%xmm11 .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 .byte 243,15,16,88,20 // movss 0x14(%rax),%xmm3 @@ -20449,9 +20369,9 @@ _sk_gather_i8_sse41: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 26ed <_sk_gather_i8_sse41+0xf> + .byte 116,5 // je 26a1 <_sk_gather_i8_sse41+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 26ef <_sk_gather_i8_sse41+0x11> + .byte 235,2 // jmp 26a3 <_sk_gather_i8_sse41+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 @@ -21603,7 +21523,7 @@ _sk_linear_gradient_sse41: .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 .byte 72,139,8 // mov (%rax),%rcx .byte 72,133,201 // test %rcx,%rcx - .byte 15,132,254,0,0,0 // je 3915 <_sk_linear_gradient_sse41+0x138> + .byte 15,132,254,0,0,0 // je 38c9 <_sk_linear_gradient_sse41+0x138> .byte 15,41,100,36,168 // movaps %xmm4,-0x58(%rsp) .byte 15,41,108,36,184 // movaps %xmm5,-0x48(%rsp) .byte 15,41,116,36,200 // movaps %xmm6,-0x38(%rsp) @@ -21653,12 +21573,12 @@ _sk_linear_gradient_sse41: .byte 15,40,196 // movaps %xmm4,%xmm0 .byte 72,131,192,36 // add $0x24,%rax .byte 72,255,201 // dec %rcx - .byte 15,133,65,255,255,255 // jne 3840 <_sk_linear_gradient_sse41+0x63> + .byte 15,133,65,255,255,255 // jne 37f4 <_sk_linear_gradient_sse41+0x63> .byte 15,40,124,36,216 // movaps -0x28(%rsp),%xmm7 .byte 15,40,116,36,200 // movaps -0x38(%rsp),%xmm6 .byte 15,40,108,36,184 // movaps -0x48(%rsp),%xmm5 .byte 15,40,100,36,168 // movaps -0x58(%rsp),%xmm4 - .byte 235,13 // jmp 3922 <_sk_linear_gradient_sse41+0x145> + .byte 235,13 // jmp 38d6 <_sk_linear_gradient_sse41+0x145> .byte 15,87,201 // xorps %xmm1,%xmm1 .byte 15,87,210 // xorps %xmm2,%xmm2 .byte 15,87,219 // xorps %xmm3,%xmm3 @@ -24472,97 +24392,92 @@ _sk_parametric_r_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,192 // mulps %xmm0,%xmm8 - .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 68,15,89,208 // mulps %xmm0,%xmm10 + .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 68,15,89,216 // mulps %xmm0,%xmm11 .byte 65,15,194,193,2 // cmpleps %xmm9,%xmm0 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,88,8 // movss 0x8(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,88,218 // addps %xmm10,%xmm11 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,235 // subps %xmm11,%xmm13 + .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,89,212 // mulps %xmm12,%xmm10 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 + .byte 102,69,15,219,235 // pand %xmm11,%xmm13 + .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,219,226 // pand %xmm10,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 102,69,15,112,218,0 // pshufd $0x0,%xmm10,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 102,69,15,235,229 // por %xmm13,%xmm12 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,211 // mulps %xmm11,%xmm10 - .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,243 // addps %xmm11,%xmm14 + .byte 69,15,88,244 // addps %xmm12,%xmm14 .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,89,204 // mulps %xmm12,%xmm9 - .byte 243,69,15,91,217 // cvttps2dq %xmm9,%xmm11 - .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 - .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,213 // subps %xmm13,%xmm10 + .byte 69,15,89,209 // mulps %xmm9,%xmm10 + .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 + .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 + .byte 69,15,40,234 // movaps %xmm10,%xmm13 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,249 // addps %xmm9,%xmm15 - .byte 69,15,40,233 // movaps %xmm9,%xmm13 - .byte 69,15,194,235,1 // cmpltps %xmm11,%xmm13 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,84,234 // andps %xmm10,%xmm13 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,221 // subps %xmm13,%xmm11 - .byte 69,15,92,227 // subps %xmm11,%xmm12 - .byte 102,69,15,110,216 // movd %r8d,%xmm11 + .byte 69,15,88,250 // addps %xmm10,%xmm15 + .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,84,211 // andps %xmm11,%xmm10 + .byte 69,15,87,201 // xorps %xmm9,%xmm9 + .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 102,69,15,110,208 // movd %r8d,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 69,15,92,253 // subps %xmm13,%xmm15 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,89,229 // mulps %xmm13,%xmm12 + .byte 69,15,92,252 // subps %xmm12,%xmm15 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,244 // subps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,88,239 // addps %xmm15,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 102,69,15,91,221 // cvtps2dq %xmm13,%xmm11 + .byte 69,15,92,245 // subps %xmm13,%xmm14 + .byte 69,15,94,230 // divps %xmm14,%xmm12 + .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,89,226 // mulps %xmm10,%xmm12 + .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,227 // addps %xmm11,%xmm12 + .byte 69,15,88,226 // addps %xmm10,%xmm12 .byte 68,15,84,192 // andps %xmm0,%xmm8 .byte 65,15,85,196 // andnps %xmm12,%xmm0 .byte 65,15,86,192 // orps %xmm8,%xmm0 .byte 65,15,95,193 // maxps %xmm9,%xmm0 - .byte 65,15,93,194 // minps %xmm10,%xmm0 + .byte 65,15,93,195 // minps %xmm11,%xmm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24576,97 +24491,92 @@ _sk_parametric_g_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,193 // mulps %xmm1,%xmm8 - .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 68,15,89,209 // mulps %xmm1,%xmm10 + .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 68,15,89,217 // mulps %xmm1,%xmm11 .byte 65,15,194,201,2 // cmpleps %xmm9,%xmm1 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,88,8 // movss 0x8(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,88,218 // addps %xmm10,%xmm11 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,235 // subps %xmm11,%xmm13 + .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,89,212 // mulps %xmm12,%xmm10 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 + .byte 102,69,15,219,235 // pand %xmm11,%xmm13 + .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,219,226 // pand %xmm10,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 102,69,15,112,218,0 // pshufd $0x0,%xmm10,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 102,69,15,235,229 // por %xmm13,%xmm12 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,211 // mulps %xmm11,%xmm10 - .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,243 // addps %xmm11,%xmm14 + .byte 69,15,88,244 // addps %xmm12,%xmm14 .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,89,204 // mulps %xmm12,%xmm9 - .byte 243,69,15,91,217 // cvttps2dq %xmm9,%xmm11 - .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 - .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,213 // subps %xmm13,%xmm10 + .byte 69,15,89,209 // mulps %xmm9,%xmm10 + .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 + .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 + .byte 69,15,40,234 // movaps %xmm10,%xmm13 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,249 // addps %xmm9,%xmm15 - .byte 69,15,40,233 // movaps %xmm9,%xmm13 - .byte 69,15,194,235,1 // cmpltps %xmm11,%xmm13 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,84,234 // andps %xmm10,%xmm13 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,221 // subps %xmm13,%xmm11 - .byte 69,15,92,227 // subps %xmm11,%xmm12 - .byte 102,69,15,110,216 // movd %r8d,%xmm11 + .byte 69,15,88,250 // addps %xmm10,%xmm15 + .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,84,211 // andps %xmm11,%xmm10 + .byte 69,15,87,201 // xorps %xmm9,%xmm9 + .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 102,69,15,110,208 // movd %r8d,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 69,15,92,253 // subps %xmm13,%xmm15 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,89,229 // mulps %xmm13,%xmm12 + .byte 69,15,92,252 // subps %xmm12,%xmm15 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,244 // subps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,88,239 // addps %xmm15,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 102,69,15,91,221 // cvtps2dq %xmm13,%xmm11 + .byte 69,15,92,245 // subps %xmm13,%xmm14 + .byte 69,15,94,230 // divps %xmm14,%xmm12 + .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,89,226 // mulps %xmm10,%xmm12 + .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,227 // addps %xmm11,%xmm12 + .byte 69,15,88,226 // addps %xmm10,%xmm12 .byte 68,15,84,193 // andps %xmm1,%xmm8 .byte 65,15,85,204 // andnps %xmm12,%xmm1 .byte 65,15,86,200 // orps %xmm8,%xmm1 .byte 65,15,95,201 // maxps %xmm9,%xmm1 - .byte 65,15,93,202 // minps %xmm10,%xmm1 + .byte 65,15,93,203 // minps %xmm11,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24680,97 +24590,92 @@ _sk_parametric_b_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,194 // mulps %xmm2,%xmm8 - .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 68,15,89,210 // mulps %xmm2,%xmm10 + .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 68,15,89,218 // mulps %xmm2,%xmm11 .byte 65,15,194,209,2 // cmpleps %xmm9,%xmm2 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,88,8 // movss 0x8(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,88,218 // addps %xmm10,%xmm11 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,235 // subps %xmm11,%xmm13 + .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,89,212 // mulps %xmm12,%xmm10 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 + .byte 102,69,15,219,235 // pand %xmm11,%xmm13 + .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,219,226 // pand %xmm10,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 102,69,15,112,218,0 // pshufd $0x0,%xmm10,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 102,69,15,235,229 // por %xmm13,%xmm12 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,211 // mulps %xmm11,%xmm10 - .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,243 // addps %xmm11,%xmm14 + .byte 69,15,88,244 // addps %xmm12,%xmm14 .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,89,204 // mulps %xmm12,%xmm9 - .byte 243,69,15,91,217 // cvttps2dq %xmm9,%xmm11 - .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 - .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,213 // subps %xmm13,%xmm10 + .byte 69,15,89,209 // mulps %xmm9,%xmm10 + .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 + .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 + .byte 69,15,40,234 // movaps %xmm10,%xmm13 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,249 // addps %xmm9,%xmm15 - .byte 69,15,40,233 // movaps %xmm9,%xmm13 - .byte 69,15,194,235,1 // cmpltps %xmm11,%xmm13 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,84,234 // andps %xmm10,%xmm13 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,221 // subps %xmm13,%xmm11 - .byte 69,15,92,227 // subps %xmm11,%xmm12 - .byte 102,69,15,110,216 // movd %r8d,%xmm11 + .byte 69,15,88,250 // addps %xmm10,%xmm15 + .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,84,211 // andps %xmm11,%xmm10 + .byte 69,15,87,201 // xorps %xmm9,%xmm9 + .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 102,69,15,110,208 // movd %r8d,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 69,15,92,253 // subps %xmm13,%xmm15 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,89,229 // mulps %xmm13,%xmm12 + .byte 69,15,92,252 // subps %xmm12,%xmm15 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,244 // subps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,88,239 // addps %xmm15,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 102,69,15,91,221 // cvtps2dq %xmm13,%xmm11 + .byte 69,15,92,245 // subps %xmm13,%xmm14 + .byte 69,15,94,230 // divps %xmm14,%xmm12 + .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,89,226 // mulps %xmm10,%xmm12 + .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,227 // addps %xmm11,%xmm12 + .byte 69,15,88,226 // addps %xmm10,%xmm12 .byte 68,15,84,194 // andps %xmm2,%xmm8 .byte 65,15,85,212 // andnps %xmm12,%xmm2 .byte 65,15,86,208 // orps %xmm8,%xmm2 .byte 65,15,95,209 // maxps %xmm9,%xmm2 - .byte 65,15,93,210 // minps %xmm10,%xmm2 + .byte 65,15,93,211 // minps %xmm11,%xmm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -24784,97 +24689,92 @@ _sk_parametric_a_sse2: .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 .byte 68,15,89,195 // mulps %xmm3,%xmm8 - .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 68,15,89,211 // mulps %xmm3,%xmm10 + .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 68,15,89,219 // mulps %xmm3,%xmm11 .byte 65,15,194,217,2 // cmpleps %xmm9,%xmm3 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,88,193 // addps %xmm9,%xmm8 .byte 243,68,15,16,8 // movss (%rax),%xmm9 - .byte 243,68,15,16,88,8 // movss 0x8(%rax),%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,88,211 // addps %xmm11,%xmm10 - .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11 + .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,88,218 // addps %xmm10,%xmm11 + .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 .byte 185,0,0,0,52 // mov $0x34000000,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 185,0,0,254,66 // mov $0x42fe0000,%ecx - .byte 102,68,15,110,217 // movd %ecx,%xmm11 - .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 - .byte 69,15,92,235 // subps %xmm11,%xmm13 + .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 + .byte 69,15,89,212 // mulps %xmm12,%xmm10 .byte 185,255,255,127,0 // mov $0x7fffff,%ecx + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 + .byte 102,69,15,219,235 // pand %xmm11,%xmm13 + .byte 185,0,0,0,63 // mov $0x3f000000,%ecx .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 - .byte 102,69,15,219,226 // pand %xmm10,%xmm12 - .byte 185,0,0,0,63 // mov $0x3f000000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 102,69,15,112,218,0 // pshufd $0x0,%xmm10,%xmm11 - .byte 102,69,15,235,220 // por %xmm12,%xmm11 - .byte 185,42,145,49,64 // mov $0x4031912a,%ecx - .byte 102,68,15,110,225 // movd %ecx,%xmm12 - .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,229 // addps %xmm13,%xmm12 + .byte 102,69,15,235,229 // por %xmm13,%xmm12 + .byte 185,119,115,248,66 // mov $0x42f87377,%ecx + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,89,211 // mulps %xmm11,%xmm10 - .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 + .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,89,220 // mulps %xmm12,%xmm11 + .byte 69,15,92,211 // subps %xmm11,%xmm10 .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx .byte 102,68,15,110,233 // movd %ecx,%xmm13 .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx - .byte 102,68,15,110,209 // movd %ecx,%xmm10 + .byte 102,68,15,110,217 // movd %ecx,%xmm11 .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx .byte 102,68,15,110,249 // movd %ecx,%xmm15 + .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,88,243 // addps %xmm11,%xmm14 + .byte 69,15,88,244 // addps %xmm12,%xmm14 .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,92,229 // subps %xmm13,%xmm12 - .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 - .byte 69,15,89,204 // mulps %xmm12,%xmm9 - .byte 243,69,15,91,217 // cvttps2dq %xmm9,%xmm11 - .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 - .byte 69,15,40,225 // movaps %xmm9,%xmm12 + .byte 69,15,92,213 // subps %xmm13,%xmm10 + .byte 69,15,89,209 // mulps %xmm9,%xmm10 + .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 + .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 + .byte 69,15,40,234 // movaps %xmm10,%xmm13 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 - .byte 69,15,88,249 // addps %xmm9,%xmm15 - .byte 69,15,40,233 // movaps %xmm9,%xmm13 - .byte 69,15,194,235,1 // cmpltps %xmm11,%xmm13 - .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 - .byte 69,15,84,234 // andps %xmm10,%xmm13 - .byte 69,15,87,201 // xorps %xmm9,%xmm9 - .byte 69,15,92,221 // subps %xmm13,%xmm11 - .byte 69,15,92,227 // subps %xmm11,%xmm12 - .byte 102,69,15,110,216 // movd %r8d,%xmm11 + .byte 69,15,88,250 // addps %xmm10,%xmm15 + .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 + .byte 69,15,84,211 // andps %xmm11,%xmm10 + .byte 69,15,87,201 // xorps %xmm9,%xmm9 + .byte 69,15,92,226 // subps %xmm10,%xmm12 + .byte 69,15,92,236 // subps %xmm12,%xmm13 + .byte 102,69,15,110,208 // movd %r8d,%xmm10 + .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 - .byte 69,15,89,236 // mulps %xmm12,%xmm13 - .byte 69,15,92,253 // subps %xmm13,%xmm15 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 + .byte 69,15,89,229 // mulps %xmm13,%xmm12 + .byte 69,15,92,252 // subps %xmm12,%xmm15 .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx - .byte 102,68,15,110,233 // movd %ecx,%xmm13 - .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 + .byte 102,68,15,110,225 // movd %ecx,%xmm12 + .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx .byte 102,68,15,110,241 // movd %ecx,%xmm14 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 - .byte 69,15,92,244 // subps %xmm12,%xmm14 - .byte 69,15,94,238 // divps %xmm14,%xmm13 - .byte 69,15,88,239 // addps %xmm15,%xmm13 - .byte 69,15,89,235 // mulps %xmm11,%xmm13 - .byte 102,69,15,91,221 // cvtps2dq %xmm13,%xmm11 + .byte 69,15,92,245 // subps %xmm13,%xmm14 + .byte 69,15,94,230 // divps %xmm14,%xmm12 + .byte 69,15,88,231 // addps %xmm15,%xmm12 + .byte 69,15,89,226 // mulps %xmm10,%xmm12 + .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 - .byte 69,15,88,227 // addps %xmm11,%xmm12 + .byte 69,15,88,226 // addps %xmm10,%xmm12 .byte 68,15,84,195 // andps %xmm3,%xmm8 .byte 65,15,85,220 // andnps %xmm12,%xmm3 .byte 65,15,86,216 // orps %xmm8,%xmm3 .byte 65,15,95,217 // maxps %xmm9,%xmm3 - .byte 65,15,93,218 // minps %xmm10,%xmm3 + .byte 65,15,93,219 // minps %xmm11,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -25046,9 +24946,9 @@ _sk_gather_i8_sse2: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 2914 <_sk_gather_i8_sse2+0xf> + .byte 116,5 // je 28b8 <_sk_gather_i8_sse2+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 2916 <_sk_gather_i8_sse2+0x11> + .byte 235,2 // jmp 28ba <_sk_gather_i8_sse2+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 @@ -26303,7 +26203,7 @@ _sk_linear_gradient_sse2: .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 .byte 72,139,8 // mov (%rax),%rcx .byte 72,133,201 // test %rcx,%rcx - .byte 15,132,15,1,0,0 // je 3cfe <_sk_linear_gradient_sse2+0x149> + .byte 15,132,15,1,0,0 // je 3ca2 <_sk_linear_gradient_sse2+0x149> .byte 72,139,64,8 // mov 0x8(%rax),%rax .byte 72,131,192,32 // add $0x20,%rax .byte 69,15,87,192 // xorps %xmm8,%xmm8 @@ -26364,8 +26264,8 @@ _sk_linear_gradient_sse2: .byte 69,15,86,231 // orps %xmm15,%xmm12 .byte 72,131,192,36 // add $0x24,%rax .byte 72,255,201 // dec %rcx - .byte 15,133,8,255,255,255 // jne 3c04 <_sk_linear_gradient_sse2+0x4f> - .byte 235,13 // jmp 3d0b <_sk_linear_gradient_sse2+0x156> + .byte 15,133,8,255,255,255 // jne 3ba8 <_sk_linear_gradient_sse2+0x4f> + .byte 235,13 // jmp 3caf <_sk_linear_gradient_sse2+0x156> .byte 15,87,201 // xorps %xmm1,%xmm1 .byte 15,87,210 // xorps %xmm2,%xmm2 .byte 15,87,219 // xorps %xmm3,%xmm3 diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index d8f82efef0..1a3bb5ed43 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -1357,7 +1357,7 @@ _sk_lerp_565_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2001478 <_sk_callback_hsw+0xffffffffe1ffd22e> + DB 233,255,255,255,225 ; jmpq ffffffffe2001478 <_sk_callback_hsw+0xffffffffe1ffd27e> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -1945,14 +1945,11 @@ _sk_parametric_r_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,224 ; vpbroadcastd %xmm0,%ymm12 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 196,98,125,88,232 ; vpbroadcastd %xmm0,%ymm13 - DB 196,66,37,186,236 ; vfmsub231ps %ymm12,%ymm11,%ymm13 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 @@ -1961,25 +1958,24 @@ _sk_parametric_r_hsw LABEL PROC DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 DB 197,45,235,208 ; vpor %ymm0,%ymm10,%ymm10 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 - DB 197,20,88,216 ; vaddps %ymm0,%ymm13,%ymm11 + DB 196,98,125,88,240 ; vpbroadcastd %xmm0,%ymm14 + DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,224 ; vpbroadcastd %xmm0,%ymm12 - DB 196,66,45,172,227 ; vfnmadd213ps %ymm11,%ymm10,%ymm12 + DB 196,98,125,88,216 ; vpbroadcastd %xmm0,%ymm11 + DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 - DB 196,98,125,88,216 ; vpbroadcastd %xmm0,%ymm11 + DB 196,98,125,88,232 ; vpbroadcastd %xmm0,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 - DB 197,164,94,192 ; vdivps %ymm0,%ymm11,%ymm0 - DB 197,156,92,192 ; vsubps %ymm0,%ymm12,%ymm0 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 + DB 197,148,94,192 ; vdivps %ymm0,%ymm13,%ymm0 + DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 + DB 197,28,89,216 ; vmulps %ymm0,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -2027,14 +2023,11 @@ _sk_parametric_g_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,225 ; vpbroadcastd %xmm1,%ymm12 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,98,125,88,233 ; vpbroadcastd %xmm1,%ymm13 - DB 196,66,37,186,236 ; vfmsub231ps %ymm12,%ymm11,%ymm13 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 @@ -2043,25 +2036,24 @@ _sk_parametric_g_hsw LABEL PROC DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 DB 197,45,235,209 ; vpor %ymm1,%ymm10,%ymm10 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 - DB 197,20,88,217 ; vaddps %ymm1,%ymm13,%ymm11 + DB 196,98,125,88,241 ; vpbroadcastd %xmm1,%ymm14 + DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,225 ; vpbroadcastd %xmm1,%ymm12 - DB 196,66,45,172,227 ; vfnmadd213ps %ymm11,%ymm10,%ymm12 + DB 196,98,125,88,217 ; vpbroadcastd %xmm1,%ymm11 + DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,98,125,88,217 ; vpbroadcastd %xmm1,%ymm11 + DB 196,98,125,88,233 ; vpbroadcastd %xmm1,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1 - DB 197,164,94,201 ; vdivps %ymm1,%ymm11,%ymm1 - DB 197,156,92,201 ; vsubps %ymm1,%ymm12,%ymm1 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,217 ; vmulps %ymm1,%ymm10,%ymm11 + DB 197,148,94,201 ; vdivps %ymm1,%ymm13,%ymm1 + DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 + DB 197,28,89,217 ; vmulps %ymm1,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -2109,14 +2101,11 @@ _sk_parametric_b_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,226 ; vpbroadcastd %xmm2,%ymm12 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,98,125,88,234 ; vpbroadcastd %xmm2,%ymm13 - DB 196,66,37,186,236 ; vfmsub231ps %ymm12,%ymm11,%ymm13 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 @@ -2125,25 +2114,24 @@ _sk_parametric_b_hsw LABEL PROC DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 DB 197,45,235,210 ; vpor %ymm2,%ymm10,%ymm10 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 - DB 197,20,88,218 ; vaddps %ymm2,%ymm13,%ymm11 + DB 196,98,125,88,242 ; vpbroadcastd %xmm2,%ymm14 + DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,226 ; vpbroadcastd %xmm2,%ymm12 - DB 196,66,45,172,227 ; vfnmadd213ps %ymm11,%ymm10,%ymm12 + DB 196,98,125,88,218 ; vpbroadcastd %xmm2,%ymm11 + DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,98,125,88,218 ; vpbroadcastd %xmm2,%ymm11 + DB 196,98,125,88,234 ; vpbroadcastd %xmm2,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 DB 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 - DB 197,164,94,210 ; vdivps %ymm2,%ymm11,%ymm2 - DB 197,156,92,210 ; vsubps %ymm2,%ymm12,%ymm2 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,218 ; vmulps %ymm2,%ymm10,%ymm11 + DB 197,148,94,210 ; vdivps %ymm2,%ymm13,%ymm2 + DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 + DB 197,28,89,218 ; vmulps %ymm2,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -2191,14 +2179,11 @@ _sk_parametric_a_hsw LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,227 ; vpbroadcastd %xmm3,%ymm12 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,98,125,88,235 ; vpbroadcastd %xmm3,%ymm13 - DB 196,66,37,186,236 ; vfmsub231ps %ymm12,%ymm11,%ymm13 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 @@ -2207,25 +2192,24 @@ _sk_parametric_a_hsw LABEL PROC DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 DB 197,45,235,211 ; vpor %ymm3,%ymm10,%ymm10 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 197,20,88,219 ; vaddps %ymm3,%ymm13,%ymm11 + DB 196,98,125,88,243 ; vpbroadcastd %xmm3,%ymm14 + DB 196,66,37,186,245 ; vfmsub231ps %ymm13,%ymm11,%ymm14 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,227 ; vpbroadcastd %xmm3,%ymm12 - DB 196,66,45,172,227 ; vfnmadd213ps %ymm11,%ymm10,%ymm12 + DB 196,98,125,88,219 ; vpbroadcastd %xmm3,%ymm11 + DB 196,66,45,172,222 ; vfnmadd213ps %ymm14,%ymm10,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,98,125,88,219 ; vpbroadcastd %xmm3,%ymm11 + DB 196,98,125,88,235 ; vpbroadcastd %xmm3,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 DB 197,172,88,219 ; vaddps %ymm3,%ymm10,%ymm3 - DB 197,164,94,219 ; vdivps %ymm3,%ymm11,%ymm3 - DB 197,156,92,219 ; vsubps %ymm3,%ymm12,%ymm3 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,219 ; vmulps %ymm3,%ymm10,%ymm11 + DB 197,148,94,219 ; vdivps %ymm3,%ymm13,%ymm3 + DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 + DB 197,28,89,219 ; vmulps %ymm3,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -2269,7 +2253,7 @@ _sk_load_a8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,50 ; jne 2458 <_sk_load_a8_hsw+0x42> + DB 117,50 ; jne 2408 <_sk_load_a8_hsw+0x42> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 @@ -2292,9 +2276,9 @@ _sk_load_a8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2460 <_sk_load_a8_hsw+0x4a> + DB 117,234 ; jne 2410 <_sk_load_a8_hsw+0x4a> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,173 ; jmp 242a <_sk_load_a8_hsw+0x14> + DB 235,173 ; jmp 23da <_sk_load_a8_hsw+0x14> PUBLIC _sk_gather_a8_hsw _sk_gather_a8_hsw LABEL PROC @@ -2365,7 +2349,7 @@ _sk_store_a8_hsw LABEL PROC DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 2595 <_sk_store_a8_hsw+0x3b> + DB 117,10 ; jne 2545 <_sk_store_a8_hsw+0x3b> DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2373,10 +2357,10 @@ _sk_store_a8_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 2591 <_sk_store_a8_hsw+0x37> + DB 119,236 ; ja 2541 <_sk_store_a8_hsw+0x37> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 25f8 <_sk_store_a8_hsw+0x9e> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 25a8 <_sk_store_a8_hsw+0x9e> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2387,7 +2371,7 @@ _sk_store_a8_hsw LABEL PROC DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,154 ; jmp 2591 <_sk_store_a8_hsw+0x37> + DB 235,154 ; jmp 2541 <_sk_store_a8_hsw+0x37> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -2419,7 +2403,7 @@ _sk_load_g8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,60 ; jne 2660 <_sk_load_g8_hsw+0x4c> + DB 117,60 ; jne 2610 <_sk_load_g8_hsw+0x4c> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 @@ -2444,9 +2428,9 @@ _sk_load_g8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 2668 <_sk_load_g8_hsw+0x54> + DB 117,234 ; jne 2618 <_sk_load_g8_hsw+0x54> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,163 ; jmp 2628 <_sk_load_g8_hsw+0x14> + DB 235,163 ; jmp 25d8 <_sk_load_g8_hsw+0x14> PUBLIC _sk_gather_g8_hsw _sk_gather_g8_hsw LABEL PROC @@ -2511,9 +2495,9 @@ _sk_gather_i8_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 73,137,192 ; mov %rax,%r8 DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 277b <_sk_gather_i8_hsw+0xf> + DB 116,5 ; je 272b <_sk_gather_i8_hsw+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 277d <_sk_gather_i8_hsw+0x11> + DB 235,2 ; jmp 272d <_sk_gather_i8_hsw+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -2584,7 +2568,7 @@ _sk_load_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,149,0,0,0 ; jne 292f <_sk_load_565_hsw+0xa3> + DB 15,133,149,0,0,0 ; jne 28df <_sk_load_565_hsw+0xa3> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 DB 184,0,248,0,0 ; mov $0xf800,%eax @@ -2624,9 +2608,9 @@ _sk_load_565_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,89,255,255,255 ; ja 28a0 <_sk_load_565_hsw+0x14> + DB 15,135,89,255,255,255 ; ja 2850 <_sk_load_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 299c <_sk_load_565_hsw+0x110> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 294c <_sk_load_565_hsw+0x110> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2638,12 +2622,12 @@ _sk_load_565_hsw LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,5,255,255,255 ; jmpq 28a0 <_sk_load_565_hsw+0x14> + DB 233,5,255,255,255 ; jmpq 2850 <_sk_load_565_hsw+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 29a1 <_sk_load_565_hsw+0x115> + DB 235,255 ; jmp 2951 <_sk_load_565_hsw+0x115> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -2766,7 +2750,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 2b67 <_sk_store_565_hsw+0x6c> + DB 117,10 ; jne 2b17 <_sk_store_565_hsw+0x6c> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2774,9 +2758,9 @@ _sk_store_565_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 2b63 <_sk_store_565_hsw+0x68> + DB 119,236 ; ja 2b13 <_sk_store_565_hsw+0x68> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2bc4 <_sk_store_565_hsw+0xc9> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2b74 <_sk_store_565_hsw+0xc9> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2787,7 +2771,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 2b63 <_sk_store_565_hsw+0x68> + DB 235,159 ; jmp 2b13 <_sk_store_565_hsw+0x68> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -2816,7 +2800,7 @@ _sk_load_4444_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,179,0,0,0 ; jne 2ca1 <_sk_load_4444_hsw+0xc1> + DB 15,133,179,0,0,0 ; jne 2c51 <_sk_load_4444_hsw+0xc1> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,98,125,51,200 ; vpmovzxwd %xmm0,%ymm9 DB 184,0,240,0,0 ; mov $0xf000,%eax @@ -2862,9 +2846,9 @@ _sk_load_4444_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,59,255,255,255 ; ja 2bf4 <_sk_load_4444_hsw+0x14> + DB 15,135,59,255,255,255 ; ja 2ba4 <_sk_load_4444_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 2d10 <_sk_load_4444_hsw+0x130> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 2cc0 <_sk_load_4444_hsw+0x130> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2876,13 +2860,13 @@ _sk_load_4444_hsw LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,231,254,255,255 ; jmpq 2bf4 <_sk_load_4444_hsw+0x14> + DB 233,231,254,255,255 ; jmpq 2ba4 <_sk_load_4444_hsw+0x14> DB 15,31,0 ; nopl (%rax) DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2002d18 <_sk_callback_hsw+0xffffffffe1ffeace> + DB 233,255,255,255,225 ; jmpq ffffffffe2002cc8 <_sk_callback_hsw+0xffffffffe1ffeace> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -3010,7 +2994,7 @@ _sk_store_4444_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 2eff <_sk_store_4444_hsw+0x72> + DB 117,10 ; jne 2eaf <_sk_store_4444_hsw+0x72> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3018,9 +3002,9 @@ _sk_store_4444_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 2efb <_sk_store_4444_hsw+0x6e> + DB 119,236 ; ja 2eab <_sk_store_4444_hsw+0x6e> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2f5c <_sk_store_4444_hsw+0xcf> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2f0c <_sk_store_4444_hsw+0xcf> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -3031,7 +3015,7 @@ _sk_store_4444_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 2efb <_sk_store_4444_hsw+0x6e> + DB 235,159 ; jmp 2eab <_sk_store_4444_hsw+0x6e> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -3062,7 +3046,7 @@ _sk_load_8888_hsw LABEL PROC DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 - DB 117,104 ; jne 2ff5 <_sk_load_8888_hsw+0x7d> + DB 117,104 ; jne 2fa5 <_sk_load_8888_hsw+0x7d> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -3095,7 +3079,7 @@ _sk_load_8888_hsw LABEL PROC DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3 - DB 233,116,255,255,255 ; jmpq 2f92 <_sk_load_8888_hsw+0x1a> + DB 233,116,255,255,255 ; jmpq 2f42 <_sk_load_8888_hsw+0x1a> PUBLIC _sk_gather_8888_hsw _sk_gather_8888_hsw LABEL PROC @@ -3155,7 +3139,7 @@ _sk_store_8888_hsw LABEL PROC DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 77,133,192 ; test %r8,%r8 - DB 117,12 ; jne 3118 <_sk_store_8888_hsw+0x74> + DB 117,12 ; jne 30c8 <_sk_store_8888_hsw+0x74> DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx @@ -3168,14 +3152,14 @@ _sk_store_8888_hsw LABEL PROC DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9 DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9) - DB 235,211 ; jmp 3111 <_sk_store_8888_hsw+0x6d> + DB 235,211 ; jmp 30c1 <_sk_store_8888_hsw+0x6d> PUBLIC _sk_load_f16_hsw _sk_load_f16_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 117,97 ; jne 31a9 <_sk_load_f16_hsw+0x6b> + DB 117,97 ; jne 3159 <_sk_load_f16_hsw+0x6b> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -3201,29 +3185,29 @@ _sk_load_f16_hsw LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 3208 <_sk_load_f16_hsw+0xca> + DB 116,79 ; je 31b8 <_sk_load_f16_hsw+0xca> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 3208 <_sk_load_f16_hsw+0xca> + DB 114,67 ; jb 31b8 <_sk_load_f16_hsw+0xca> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 3215 <_sk_load_f16_hsw+0xd7> + DB 116,68 ; je 31c5 <_sk_load_f16_hsw+0xd7> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 3215 <_sk_load_f16_hsw+0xd7> + DB 114,56 ; jb 31c5 <_sk_load_f16_hsw+0xd7> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,114,255,255,255 ; je 315f <_sk_load_f16_hsw+0x21> + DB 15,132,114,255,255,255 ; je 310f <_sk_load_f16_hsw+0x21> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,98,255,255,255 ; jb 315f <_sk_load_f16_hsw+0x21> + DB 15,130,98,255,255,255 ; jb 310f <_sk_load_f16_hsw+0x21> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,87,255,255,255 ; jmpq 315f <_sk_load_f16_hsw+0x21> + DB 233,87,255,255,255 ; jmpq 310f <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,74,255,255,255 ; jmpq 315f <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq 310f <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq 315f <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 310f <_sk_load_f16_hsw+0x21> PUBLIC _sk_gather_f16_hsw _sk_gather_f16_hsw LABEL PROC @@ -3277,7 +3261,7 @@ _sk_store_f16_hsw LABEL PROC DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9 DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,27 ; jne 330d <_sk_store_f16_hsw+0x65> + DB 117,27 ; jne 32bd <_sk_store_f16_hsw+0x65> DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8) DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8) DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8) @@ -3286,22 +3270,22 @@ _sk_store_f16_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,241 ; je 3309 <_sk_store_f16_hsw+0x61> + DB 116,241 ; je 32b9 <_sk_store_f16_hsw+0x61> DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,229 ; jb 3309 <_sk_store_f16_hsw+0x61> + DB 114,229 ; jb 32b9 <_sk_store_f16_hsw+0x61> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je 3309 <_sk_store_f16_hsw+0x61> + DB 116,221 ; je 32b9 <_sk_store_f16_hsw+0x61> DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,209 ; jb 3309 <_sk_store_f16_hsw+0x61> + DB 114,209 ; jb 32b9 <_sk_store_f16_hsw+0x61> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je 3309 <_sk_store_f16_hsw+0x61> + DB 116,201 ; je 32b9 <_sk_store_f16_hsw+0x61> DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,189 ; jb 3309 <_sk_store_f16_hsw+0x61> + DB 114,189 ; jb 32b9 <_sk_store_f16_hsw+0x61> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp 3309 <_sk_store_f16_hsw+0x61> + DB 235,181 ; jmp 32b9 <_sk_store_f16_hsw+0x61> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_hsw LABEL PROC @@ -3309,7 +3293,7 @@ _sk_load_u16_be_hsw LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,205,0,0,0 ; jne 3437 <_sk_load_u16_be_hsw+0xe3> + DB 15,133,205,0,0,0 ; jne 33e7 <_sk_load_u16_be_hsw+0xe3> DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -3358,29 +3342,29 @@ _sk_load_u16_be_hsw LABEL PROC DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 349d <_sk_load_u16_be_hsw+0x149> + DB 116,85 ; je 344d <_sk_load_u16_be_hsw+0x149> DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 349d <_sk_load_u16_be_hsw+0x149> + DB 114,72 ; jb 344d <_sk_load_u16_be_hsw+0x149> DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 34aa <_sk_load_u16_be_hsw+0x156> + DB 116,72 ; je 345a <_sk_load_u16_be_hsw+0x156> DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 34aa <_sk_load_u16_be_hsw+0x156> + DB 114,59 ; jb 345a <_sk_load_u16_be_hsw+0x156> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,5,255,255,255 ; je 3385 <_sk_load_u16_be_hsw+0x31> + DB 15,132,5,255,255,255 ; je 3335 <_sk_load_u16_be_hsw+0x31> DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,244,254,255,255 ; jb 3385 <_sk_load_u16_be_hsw+0x31> + DB 15,130,244,254,255,255 ; jb 3335 <_sk_load_u16_be_hsw+0x31> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,232,254,255,255 ; jmpq 3385 <_sk_load_u16_be_hsw+0x31> + DB 233,232,254,255,255 ; jmpq 3335 <_sk_load_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,219,254,255,255 ; jmpq 3385 <_sk_load_u16_be_hsw+0x31> + DB 233,219,254,255,255 ; jmpq 3335 <_sk_load_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,210,254,255,255 ; jmpq 3385 <_sk_load_u16_be_hsw+0x31> + DB 233,210,254,255,255 ; jmpq 3335 <_sk_load_u16_be_hsw+0x31> PUBLIC _sk_load_rgb_u16_be_hsw _sk_load_rgb_u16_be_hsw LABEL PROC @@ -3388,7 +3372,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,211,0,0,0 ; jne 3598 <_sk_load_rgb_u16_be_hsw+0xe5> + DB 15,133,211,0,0,0 ; jne 3548 <_sk_load_rgb_u16_be_hsw+0xe5> DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0 DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2 DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -3438,36 +3422,36 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 35b1 <_sk_load_rgb_u16_be_hsw+0xfe> - DB 233,72,255,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,5 ; jne 3561 <_sk_load_rgb_u16_be_hsw+0xfe> + DB 233,72,255,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1 DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 35e0 <_sk_load_rgb_u16_be_hsw+0x12d> + DB 114,26 ; jb 3590 <_sk_load_rgb_u16_be_hsw+0x12d> DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1 DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 35e5 <_sk_load_rgb_u16_be_hsw+0x132> - DB 233,25,255,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,20,255,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 3595 <_sk_load_rgb_u16_be_hsw+0x132> + DB 233,25,255,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,20,255,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1 DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 3614 <_sk_load_rgb_u16_be_hsw+0x161> + DB 114,26 ; jb 35c4 <_sk_load_rgb_u16_be_hsw+0x161> DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1 DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 3619 <_sk_load_rgb_u16_be_hsw+0x166> - DB 233,229,254,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,224,254,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 117,10 ; jne 35c9 <_sk_load_rgb_u16_be_hsw+0x166> + DB 233,229,254,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,224,254,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3 DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 3642 <_sk_load_rgb_u16_be_hsw+0x18f> + DB 114,20 ; jb 35f2 <_sk_load_rgb_u16_be_hsw+0x18f> DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3 DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 233,183,254,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> - DB 233,178,254,255,255 ; jmpq 34f9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,183,254,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> + DB 233,178,254,255,255 ; jmpq 34a9 <_sk_load_rgb_u16_be_hsw+0x46> PUBLIC _sk_store_u16_be_hsw _sk_store_u16_be_hsw LABEL PROC @@ -3514,7 +3498,7 @@ _sk_store_u16_be_hsw LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 3742 <_sk_store_u16_be_hsw+0xfb> + DB 117,31 ; jne 36f2 <_sk_store_u16_be_hsw+0xfb> DB 196,1,120,17,28,72 ; vmovups %xmm11,(%r8,%r9,2) DB 196,1,120,17,84,72,16 ; vmovups %xmm10,0x10(%r8,%r9,2) DB 196,1,120,17,76,72,32 ; vmovups %xmm9,0x20(%r8,%r9,2) @@ -3523,31 +3507,31 @@ _sk_store_u16_be_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,1,121,214,28,72 ; vmovq %xmm11,(%r8,%r9,2) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 373e <_sk_store_u16_be_hsw+0xf7> + DB 116,240 ; je 36ee <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,23,92,72,8 ; vmovhpd %xmm11,0x8(%r8,%r9,2) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 373e <_sk_store_u16_be_hsw+0xf7> + DB 114,227 ; jb 36ee <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,214,84,72,16 ; vmovq %xmm10,0x10(%r8,%r9,2) - DB 116,218 ; je 373e <_sk_store_u16_be_hsw+0xf7> + DB 116,218 ; je 36ee <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,23,84,72,24 ; vmovhpd %xmm10,0x18(%r8,%r9,2) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 373e <_sk_store_u16_be_hsw+0xf7> + DB 114,205 ; jb 36ee <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,214,76,72,32 ; vmovq %xmm9,0x20(%r8,%r9,2) - DB 116,196 ; je 373e <_sk_store_u16_be_hsw+0xf7> + DB 116,196 ; je 36ee <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,23,76,72,40 ; vmovhpd %xmm9,0x28(%r8,%r9,2) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 373e <_sk_store_u16_be_hsw+0xf7> + DB 114,183 ; jb 36ee <_sk_store_u16_be_hsw+0xf7> DB 196,1,121,214,68,72,48 ; vmovq %xmm8,0x30(%r8,%r9,2) - DB 235,174 ; jmp 373e <_sk_store_u16_be_hsw+0xf7> + DB 235,174 ; jmp 36ee <_sk_store_u16_be_hsw+0xf7> PUBLIC _sk_load_f32_hsw _sk_load_f32_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,131,249,7 ; cmp $0x7,%rcx - DB 119,110 ; ja 3806 <_sk_load_f32_hsw+0x76> + DB 119,110 ; ja 37b6 <_sk_load_f32_hsw+0x76> DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 3830 <_sk_load_f32_hsw+0xa0> + DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 37e0 <_sk_load_f32_hsw+0xa0> DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -3604,7 +3588,7 @@ _sk_store_f32_hsw LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 38bd <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 386d <_sk_store_f32_hsw+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -3617,22 +3601,22 @@ _sk_store_f32_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 38b9 <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 3869 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 38b9 <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 3869 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 38b9 <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 3869 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 38b9 <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 3869 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 38b9 <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 3869 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 38b9 <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 3869 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 38b9 <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 3869 <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -3873,7 +3857,7 @@ _sk_linear_gradient_hsw LABEL PROC DB 196,98,125,24,72,28 ; vbroadcastss 0x1c(%rax),%ymm9 DB 76,139,0 ; mov (%rax),%r8 DB 77,133,192 ; test %r8,%r8 - DB 15,132,143,0,0,0 ; je 3d49 <_sk_linear_gradient_hsw+0xb5> + DB 15,132,143,0,0,0 ; je 3cf9 <_sk_linear_gradient_hsw+0xb5> DB 72,139,64,8 ; mov 0x8(%rax),%rax DB 72,131,192,32 ; add $0x20,%rax DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12 @@ -3900,8 +3884,8 @@ _sk_linear_gradient_hsw LABEL PROC DB 196,67,13,74,201,208 ; vblendvps %ymm13,%ymm9,%ymm14,%ymm9 DB 72,131,192,36 ; add $0x24,%rax DB 73,255,200 ; dec %r8 - DB 117,140 ; jne 3cd3 <_sk_linear_gradient_hsw+0x3f> - DB 235,17 ; jmp 3d5a <_sk_linear_gradient_hsw+0xc6> + DB 117,140 ; jne 3c83 <_sk_linear_gradient_hsw+0x3f> + DB 235,17 ; jmp 3d0a <_sk_linear_gradient_hsw+0xc6> DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 @@ -6803,52 +6787,47 @@ _sk_parametric_r_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 - DB 196,65,124,88,219 ; vaddps %ymm11,%ymm0,%ymm11 - DB 196,65,124,91,211 ; vcvtdq2ps %ymm11,%ymm10 + DB 196,65,124,88,211 ; vaddps %ymm11,%ymm0,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%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 197,44,89,208 ; vmulps %ymm0,%ymm10,%ymm10 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,192 ; vmovd %r8d,%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 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 + DB 197,36,89,216 ; vmulps %ymm0,%ymm11,%ymm11 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 196,65,124,84,219 ; vandps %ymm11,%ymm0,%ymm11 + DB 196,65,124,84,210 ; vandps %ymm10,%ymm0,%ymm10 DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,36,86,216 ; vorps %ymm0,%ymm11,%ymm11 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 197,44,86,208 ; vorps %ymm0,%ymm10,%ymm10 + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%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 197,44,88,208 ; vaddps %ymm0,%ymm10,%ymm10 + DB 197,36,92,216 ; vsubps %ymm0,%ymm11,%ymm11 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%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 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 - DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 + DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 + DB 197,36,92,216 ; vsubps %ymm0,%ymm11,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%xmm0 DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 - DB 196,99,125,24,224,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 + DB 196,99,125,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,192 ; vmovd %r8d,%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 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0 - DB 197,172,92,192 ; vsubps %ymm0,%ymm10,%ymm0 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 + DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 + DB 197,148,94,192 ; vdivps %ymm0,%ymm13,%ymm0 + DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 + DB 197,28,89,216 ; vmulps %ymm0,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -6904,52 +6883,47 @@ _sk_parametric_g_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 - DB 196,65,116,88,219 ; vaddps %ymm11,%ymm1,%ymm11 - DB 196,65,124,91,211 ; vcvtdq2ps %ymm11,%ymm10 + DB 196,65,116,88,211 ; vaddps %ymm11,%ymm1,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,44,89,209 ; vmulps %ymm1,%ymm10,%ymm10 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,44,92,209 ; vsubps %ymm1,%ymm10,%ymm10 + DB 197,36,89,217 ; vmulps %ymm1,%ymm11,%ymm11 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 196,65,116,84,219 ; vandps %ymm11,%ymm1,%ymm11 + DB 196,65,116,84,210 ; vandps %ymm10,%ymm1,%ymm10 DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,36,86,217 ; vorps %ymm1,%ymm11,%ymm11 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 197,44,86,209 ; vorps %ymm1,%ymm10,%ymm10 + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,44,88,209 ; vaddps %ymm1,%ymm10,%ymm10 + DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 - DB 197,44,92,209 ; vsubps %ymm1,%ymm10,%ymm10 + DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 + DB 197,36,92,217 ; vsubps %ymm1,%ymm11,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 - DB 196,99,117,24,225,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 + DB 196,99,117,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 - DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1 - DB 197,172,92,201 ; vsubps %ymm1,%ymm10,%ymm1 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,217 ; vmulps %ymm1,%ymm10,%ymm11 + DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1 + DB 197,148,94,201 ; vdivps %ymm1,%ymm13,%ymm1 + DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 + DB 197,28,89,217 ; vmulps %ymm1,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -7005,52 +6979,47 @@ _sk_parametric_b_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 - DB 196,65,108,88,219 ; vaddps %ymm11,%ymm2,%ymm11 - DB 196,65,124,91,211 ; vcvtdq2ps %ymm11,%ymm10 + DB 196,65,108,88,211 ; vaddps %ymm11,%ymm2,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,44,89,210 ; vmulps %ymm2,%ymm10,%ymm10 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,44,92,210 ; vsubps %ymm2,%ymm10,%ymm10 + DB 197,36,89,218 ; vmulps %ymm2,%ymm11,%ymm11 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 196,65,108,84,219 ; vandps %ymm11,%ymm2,%ymm11 + DB 196,65,108,84,210 ; vandps %ymm10,%ymm2,%ymm10 DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,36,86,218 ; vorps %ymm2,%ymm11,%ymm11 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 197,44,86,210 ; vorps %ymm2,%ymm10,%ymm10 + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,44,88,210 ; vaddps %ymm2,%ymm10,%ymm10 + DB 197,36,92,218 ; vsubps %ymm2,%ymm11,%ymm11 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,164,89,210 ; vmulps %ymm2,%ymm11,%ymm2 - DB 197,44,92,210 ; vsubps %ymm2,%ymm10,%ymm10 + DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 + DB 197,36,92,218 ; vsubps %ymm2,%ymm11,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 - DB 196,99,109,24,226,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm12 + DB 196,99,109,24,234,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2 DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2 - DB 197,172,92,210 ; vsubps %ymm2,%ymm10,%ymm2 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,218 ; vmulps %ymm2,%ymm10,%ymm11 + DB 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 + DB 197,148,94,210 ; vdivps %ymm2,%ymm13,%ymm2 + DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 + DB 197,28,89,218 ; vmulps %ymm2,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -7106,52 +7075,47 @@ _sk_parametric_a_avx LABEL PROC DB 196,98,125,24,80,4 ; vbroadcastss 0x4(%rax),%ymm10 DB 196,98,125,24,88,8 ; vbroadcastss 0x8(%rax),%ymm11 DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 - DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 - DB 196,65,124,91,211 ; vcvtdq2ps %ymm11,%ymm10 + DB 196,65,100,88,211 ; vaddps %ymm11,%ymm3,%ymm10 + DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 + DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 DB 65,184,0,0,0,52 ; mov $0x34000000,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,44,89,211 ; vmulps %ymm3,%ymm10,%ymm10 - DB 65,184,0,0,254,66 ; mov $0x42fe0000,%r8d - DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 - DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,44,92,211 ; vsubps %ymm3,%ymm10,%ymm10 + DB 197,36,89,219 ; vmulps %ymm3,%ymm11,%ymm11 DB 65,184,255,255,127,0 ; mov $0x7fffff,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 196,65,100,84,219 ; vandps %ymm11,%ymm3,%ymm11 + DB 196,65,100,84,210 ; vandps %ymm10,%ymm3,%ymm10 DB 65,184,0,0,0,63 ; mov $0x3f000000,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,36,86,219 ; vorps %ymm3,%ymm11,%ymm11 - DB 65,184,42,145,49,64 ; mov $0x4031912a,%r8d + DB 197,44,86,211 ; vorps %ymm3,%ymm10,%ymm10 + DB 65,184,119,115,248,66 ; mov $0x42f87377,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,44,88,211 ; vaddps %ymm3,%ymm10,%ymm10 + DB 197,36,92,219 ; vsubps %ymm3,%ymm11,%ymm11 DB 65,184,117,191,191,63 ; mov $0x3fbfbf75,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,164,89,219 ; vmulps %ymm3,%ymm11,%ymm3 - DB 197,44,92,211 ; vsubps %ymm3,%ymm10,%ymm10 + DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 + DB 197,36,92,219 ; vsubps %ymm3,%ymm11,%ymm11 DB 65,184,163,233,220,63 ; mov $0x3fdce9a3,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 - DB 196,99,101,24,227,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm12 + DB 196,99,101,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 DB 65,184,249,68,180,62 ; mov $0x3eb444f9,%r8d DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 - DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3 - DB 197,172,92,219 ; vsubps %ymm3,%ymm10,%ymm3 - DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 - DB 197,44,89,219 ; vmulps %ymm3,%ymm10,%ymm11 + DB 197,172,88,219 ; vaddps %ymm3,%ymm10,%ymm3 + DB 197,148,94,219 ; vdivps %ymm3,%ymm13,%ymm3 + DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 + DB 197,28,89,219 ; vmulps %ymm3,%ymm12,%ymm11 DB 196,67,125,8,211,1 ; vroundps $0x1,%ymm11,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d @@ -7202,7 +7166,7 @@ _sk_load_a8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,74 ; jne 31b2 <_sk_load_a8_avx+0x5a> + DB 117,74 ; jne 3146 <_sk_load_a8_avx+0x5a> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1 DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0 @@ -7229,9 +7193,9 @@ _sk_load_a8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 31ba <_sk_load_a8_avx+0x62> + DB 117,234 ; jne 314e <_sk_load_a8_avx+0x62> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,149 ; jmp 316c <_sk_load_a8_avx+0x14> + DB 235,149 ; jmp 3100 <_sk_load_a8_avx+0x14> PUBLIC _sk_gather_a8_avx _sk_gather_a8_avx LABEL PROC @@ -7308,7 +7272,7 @@ _sk_store_a8_avx LABEL PROC DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3313 <_sk_store_a8_avx+0x42> + DB 117,10 ; jne 32a7 <_sk_store_a8_avx+0x42> DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -7316,10 +7280,10 @@ _sk_store_a8_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 330f <_sk_store_a8_avx+0x3e> + DB 119,236 ; ja 32a3 <_sk_store_a8_avx+0x3e> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 3378 <_sk_store_a8_avx+0xa7> + DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 330c <_sk_store_a8_avx+0xa7> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -7330,7 +7294,7 @@ _sk_store_a8_avx LABEL PROC DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,154 ; jmp 330f <_sk_store_a8_avx+0x3e> + DB 235,154 ; jmp 32a3 <_sk_store_a8_avx+0x3e> DB 15,31,0 ; nopl (%rax) DB 244 ; hlt DB 255 ; (bad) @@ -7363,7 +7327,7 @@ _sk_load_g8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,91 ; jne 33ff <_sk_load_g8_avx+0x6b> + DB 117,91 ; jne 3393 <_sk_load_g8_avx+0x6b> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1 DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0 @@ -7393,9 +7357,9 @@ _sk_load_g8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 3407 <_sk_load_g8_avx+0x73> + DB 117,234 ; jne 339b <_sk_load_g8_avx+0x73> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,132 ; jmp 33a8 <_sk_load_g8_avx+0x14> + DB 235,132 ; jmp 333c <_sk_load_g8_avx+0x14> PUBLIC _sk_gather_g8_avx _sk_gather_g8_avx LABEL PROC @@ -7466,9 +7430,9 @@ _sk_gather_i8_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 73,137,192 ; mov %rax,%r8 DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 353e <_sk_gather_i8_avx+0xf> + DB 116,5 ; je 34d2 <_sk_gather_i8_avx+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 3540 <_sk_gather_i8_avx+0x11> + DB 235,2 ; jmp 34d4 <_sk_gather_i8_avx+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -7571,7 +7535,7 @@ _sk_load_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,209,0,0,0 ; jne 37da <_sk_load_565_avx+0xdf> + DB 15,133,209,0,0,0 ; jne 376e <_sk_load_565_avx+0xdf> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -7621,9 +7585,9 @@ _sk_load_565_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,29,255,255,255 ; ja 370f <_sk_load_565_avx+0x14> + DB 15,135,29,255,255,255 ; ja 36a3 <_sk_load_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 3848 <_sk_load_565_avx+0x14d> + DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 37dc <_sk_load_565_avx+0x14d> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7635,7 +7599,7 @@ _sk_load_565_avx LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,201,254,255,255 ; jmpq 370f <_sk_load_565_avx+0x14> + DB 233,201,254,255,255 ; jmpq 36a3 <_sk_load_565_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 242,255 ; repnz (bad) DB 255 ; (bad) @@ -7788,7 +7752,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3a93 <_sk_store_565_avx+0x9e> + DB 117,10 ; jne 3a27 <_sk_store_565_avx+0x9e> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -7796,9 +7760,9 @@ _sk_store_565_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 3a8f <_sk_store_565_avx+0x9a> + DB 119,236 ; ja 3a23 <_sk_store_565_avx+0x9a> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 3af0 <_sk_store_565_avx+0xfb> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 3a84 <_sk_store_565_avx+0xfb> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -7809,7 +7773,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 3a8f <_sk_store_565_avx+0x9a> + DB 235,159 ; jmp 3a23 <_sk_store_565_avx+0x9a> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -7838,7 +7802,7 @@ _sk_load_4444_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,245,0,0,0 ; jne 3c0f <_sk_load_4444_avx+0x103> + DB 15,133,245,0,0,0 ; jne 3ba3 <_sk_load_4444_avx+0x103> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -7895,9 +7859,9 @@ _sk_load_4444_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,249,254,255,255 ; ja 3b20 <_sk_load_4444_avx+0x14> + DB 15,135,249,254,255,255 ; ja 3ab4 <_sk_load_4444_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 3c7c <_sk_load_4444_avx+0x170> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 3c10 <_sk_load_4444_avx+0x170> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7909,12 +7873,12 @@ _sk_load_4444_avx LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,165,254,255,255 ; jmpq 3b20 <_sk_load_4444_avx+0x14> + DB 233,165,254,255,255 ; jmpq 3ab4 <_sk_load_4444_avx+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 3c81 <_sk_load_4444_avx+0x175> + DB 235,255 ; jmp 3c15 <_sk_load_4444_avx+0x175> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -8071,7 +8035,7 @@ _sk_store_4444_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3efc <_sk_store_4444_avx+0xaf> + DB 117,10 ; jne 3e90 <_sk_store_4444_avx+0xaf> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8079,9 +8043,9 @@ _sk_store_4444_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 3ef8 <_sk_store_4444_avx+0xab> + DB 119,236 ; ja 3e8c <_sk_store_4444_avx+0xab> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 3f5c <_sk_store_4444_avx+0x10f> + DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 3ef0 <_sk_store_4444_avx+0x10f> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -8092,7 +8056,7 @@ _sk_store_4444_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 3ef8 <_sk_store_4444_avx+0xab> + DB 235,159 ; jmp 3e8c <_sk_store_4444_avx+0xab> DB 15,31,0 ; nopl (%rax) DB 244 ; hlt DB 255 ; (bad) @@ -8123,7 +8087,7 @@ _sk_load_8888_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,157,0,0,0 ; jne 4023 <_sk_load_8888_avx+0xab> + DB 15,133,157,0,0,0 ; jne 3fb7 <_sk_load_8888_avx+0xab> DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -8161,9 +8125,9 @@ _sk_load_8888_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,80,255,255,255 ; ja 3f8c <_sk_load_8888_avx+0x14> + DB 15,135,80,255,255,255 ; ja 3f20 <_sk_load_8888_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 40d0 <_sk_load_8888_avx+0x158> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 4064 <_sk_load_8888_avx+0x158> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -8186,7 +8150,7 @@ _sk_load_8888_avx LABEL PROC DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 DB 196,195,49,34,4,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 - DB 233,188,254,255,255 ; jmpq 3f8c <_sk_load_8888_avx+0x14> + DB 233,188,254,255,255 ; jmpq 3f20 <_sk_load_8888_avx+0x14> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -8312,7 +8276,7 @@ _sk_store_8888_avx LABEL PROC DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8 DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 42d1 <_sk_store_8888_avx+0xa4> + DB 117,10 ; jne 4265 <_sk_store_8888_avx+0xa4> DB 196,65,124,17,4,185 ; vmovups %ymm8,(%r9,%rdi,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8320,9 +8284,9 @@ _sk_store_8888_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 42cd <_sk_store_8888_avx+0xa0> + DB 119,236 ; ja 4261 <_sk_store_8888_avx+0xa0> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 4340 <_sk_store_8888_avx+0x113> + DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 42d4 <_sk_store_8888_avx+0x113> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -8336,7 +8300,7 @@ _sk_store_8888_avx LABEL PROC DB 196,67,121,22,68,185,8,2 ; vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) DB 196,67,121,22,68,185,4,1 ; vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) DB 196,65,121,126,4,185 ; vmovd %xmm8,(%r9,%rdi,4) - DB 235,143 ; jmp 42cd <_sk_store_8888_avx+0xa0> + DB 235,143 ; jmp 4261 <_sk_store_8888_avx+0xa0> DB 102,144 ; xchg %ax,%ax DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -8366,7 +8330,7 @@ _sk_load_f16_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,17,1,0,0 ; jne 447b <_sk_load_f16_avx+0x11f> + DB 15,133,17,1,0,0 ; jne 440f <_sk_load_f16_avx+0x11f> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -8428,29 +8392,29 @@ _sk_load_f16_avx LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 44da <_sk_load_f16_avx+0x17e> + DB 116,79 ; je 446e <_sk_load_f16_avx+0x17e> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 44da <_sk_load_f16_avx+0x17e> + DB 114,67 ; jb 446e <_sk_load_f16_avx+0x17e> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 44e7 <_sk_load_f16_avx+0x18b> + DB 116,68 ; je 447b <_sk_load_f16_avx+0x18b> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 44e7 <_sk_load_f16_avx+0x18b> + DB 114,56 ; jb 447b <_sk_load_f16_avx+0x18b> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,194,254,255,255 ; je 4381 <_sk_load_f16_avx+0x25> + DB 15,132,194,254,255,255 ; je 4315 <_sk_load_f16_avx+0x25> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,178,254,255,255 ; jb 4381 <_sk_load_f16_avx+0x25> + DB 15,130,178,254,255,255 ; jb 4315 <_sk_load_f16_avx+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,167,254,255,255 ; jmpq 4381 <_sk_load_f16_avx+0x25> + DB 233,167,254,255,255 ; jmpq 4315 <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,154,254,255,255 ; jmpq 4381 <_sk_load_f16_avx+0x25> + DB 233,154,254,255,255 ; jmpq 4315 <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,145,254,255,255 ; jmpq 4381 <_sk_load_f16_avx+0x25> + DB 233,145,254,255,255 ; jmpq 4315 <_sk_load_f16_avx+0x25> PUBLIC _sk_gather_f16_avx _sk_gather_f16_avx LABEL PROC @@ -8590,7 +8554,7 @@ _sk_store_f16_avx LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 4770 <_sk_store_f16_avx+0xd2> + DB 117,31 ; jne 4704 <_sk_store_f16_avx+0xd2> DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8) DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8) DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8) @@ -8599,22 +8563,22 @@ _sk_store_f16_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 476c <_sk_store_f16_avx+0xce> + DB 116,240 ; je 4700 <_sk_store_f16_avx+0xce> DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 476c <_sk_store_f16_avx+0xce> + DB 114,227 ; jb 4700 <_sk_store_f16_avx+0xce> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 476c <_sk_store_f16_avx+0xce> + DB 116,218 ; je 4700 <_sk_store_f16_avx+0xce> DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 476c <_sk_store_f16_avx+0xce> + DB 114,205 ; jb 4700 <_sk_store_f16_avx+0xce> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 476c <_sk_store_f16_avx+0xce> + DB 116,196 ; je 4700 <_sk_store_f16_avx+0xce> DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 476c <_sk_store_f16_avx+0xce> + DB 114,183 ; jb 4700 <_sk_store_f16_avx+0xce> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 476c <_sk_store_f16_avx+0xce> + DB 235,174 ; jmp 4700 <_sk_store_f16_avx+0xce> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC @@ -8622,7 +8586,7 @@ _sk_load_u16_be_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,5,1,0,0 ; jne 48d9 <_sk_load_u16_be_avx+0x11b> + DB 15,133,5,1,0,0 ; jne 486d <_sk_load_u16_be_avx+0x11b> DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -8681,29 +8645,29 @@ _sk_load_u16_be_avx LABEL PROC DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 493f <_sk_load_u16_be_avx+0x181> + DB 116,85 ; je 48d3 <_sk_load_u16_be_avx+0x181> DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 493f <_sk_load_u16_be_avx+0x181> + DB 114,72 ; jb 48d3 <_sk_load_u16_be_avx+0x181> DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 494c <_sk_load_u16_be_avx+0x18e> + DB 116,72 ; je 48e0 <_sk_load_u16_be_avx+0x18e> DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 494c <_sk_load_u16_be_avx+0x18e> + DB 114,59 ; jb 48e0 <_sk_load_u16_be_avx+0x18e> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,205,254,255,255 ; je 47ef <_sk_load_u16_be_avx+0x31> + DB 15,132,205,254,255,255 ; je 4783 <_sk_load_u16_be_avx+0x31> DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,188,254,255,255 ; jb 47ef <_sk_load_u16_be_avx+0x31> + DB 15,130,188,254,255,255 ; jb 4783 <_sk_load_u16_be_avx+0x31> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,176,254,255,255 ; jmpq 47ef <_sk_load_u16_be_avx+0x31> + DB 233,176,254,255,255 ; jmpq 4783 <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,163,254,255,255 ; jmpq 47ef <_sk_load_u16_be_avx+0x31> + DB 233,163,254,255,255 ; jmpq 4783 <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,154,254,255,255 ; jmpq 47ef <_sk_load_u16_be_avx+0x31> + DB 233,154,254,255,255 ; jmpq 4783 <_sk_load_u16_be_avx+0x31> PUBLIC _sk_load_rgb_u16_be_avx _sk_load_rgb_u16_be_avx LABEL PROC @@ -8711,7 +8675,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,8,1,0,0 ; jne 4a6f <_sk_load_rgb_u16_be_avx+0x11a> + DB 15,133,8,1,0,0 ; jne 4a03 <_sk_load_rgb_u16_be_avx+0x11a> DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0 DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2 DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -8770,36 +8734,36 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 4a88 <_sk_load_rgb_u16_be_avx+0x133> - DB 233,19,255,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> + DB 117,5 ; jne 4a1c <_sk_load_rgb_u16_be_avx+0x133> + DB 233,19,255,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1 DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 4ab7 <_sk_load_rgb_u16_be_avx+0x162> + DB 114,26 ; jb 4a4b <_sk_load_rgb_u16_be_avx+0x162> DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1 DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 4abc <_sk_load_rgb_u16_be_avx+0x167> - DB 233,228,254,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> - DB 233,223,254,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 4a50 <_sk_load_rgb_u16_be_avx+0x167> + DB 233,228,254,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> + DB 233,223,254,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1 DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 4aeb <_sk_load_rgb_u16_be_avx+0x196> + DB 114,26 ; jb 4a7f <_sk_load_rgb_u16_be_avx+0x196> DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1 DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 4af0 <_sk_load_rgb_u16_be_avx+0x19b> - DB 233,176,254,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> - DB 233,171,254,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 4a84 <_sk_load_rgb_u16_be_avx+0x19b> + DB 233,176,254,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> + DB 233,171,254,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3 DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 4b19 <_sk_load_rgb_u16_be_avx+0x1c4> + DB 114,20 ; jb 4aad <_sk_load_rgb_u16_be_avx+0x1c4> DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3 DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 233,130,254,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> - DB 233,125,254,255,255 ; jmpq 499b <_sk_load_rgb_u16_be_avx+0x46> + DB 233,130,254,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> + DB 233,125,254,255,255 ; jmpq 492f <_sk_load_rgb_u16_be_avx+0x46> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC @@ -8847,7 +8811,7 @@ _sk_store_u16_be_avx LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 4c20 <_sk_store_u16_be_avx+0x102> + DB 117,31 ; jne 4bb4 <_sk_store_u16_be_avx+0x102> DB 196,1,120,17,28,72 ; vmovups %xmm11,(%r8,%r9,2) DB 196,1,120,17,84,72,16 ; vmovups %xmm10,0x10(%r8,%r9,2) DB 196,1,120,17,76,72,32 ; vmovups %xmm9,0x20(%r8,%r9,2) @@ -8856,31 +8820,31 @@ _sk_store_u16_be_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,1,121,214,28,72 ; vmovq %xmm11,(%r8,%r9,2) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 4c1c <_sk_store_u16_be_avx+0xfe> + DB 116,240 ; je 4bb0 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,23,92,72,8 ; vmovhpd %xmm11,0x8(%r8,%r9,2) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 4c1c <_sk_store_u16_be_avx+0xfe> + DB 114,227 ; jb 4bb0 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,214,84,72,16 ; vmovq %xmm10,0x10(%r8,%r9,2) - DB 116,218 ; je 4c1c <_sk_store_u16_be_avx+0xfe> + DB 116,218 ; je 4bb0 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,23,84,72,24 ; vmovhpd %xmm10,0x18(%r8,%r9,2) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 4c1c <_sk_store_u16_be_avx+0xfe> + DB 114,205 ; jb 4bb0 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,214,76,72,32 ; vmovq %xmm9,0x20(%r8,%r9,2) - DB 116,196 ; je 4c1c <_sk_store_u16_be_avx+0xfe> + DB 116,196 ; je 4bb0 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,23,76,72,40 ; vmovhpd %xmm9,0x28(%r8,%r9,2) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 4c1c <_sk_store_u16_be_avx+0xfe> + DB 114,183 ; jb 4bb0 <_sk_store_u16_be_avx+0xfe> DB 196,1,121,214,68,72,48 ; vmovq %xmm8,0x30(%r8,%r9,2) - DB 235,174 ; jmp 4c1c <_sk_store_u16_be_avx+0xfe> + DB 235,174 ; jmp 4bb0 <_sk_store_u16_be_avx+0xfe> PUBLIC _sk_load_f32_avx _sk_load_f32_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,131,249,7 ; cmp $0x7,%rcx - DB 119,110 ; ja 4ce4 <_sk_load_f32_avx+0x76> + DB 119,110 ; ja 4c78 <_sk_load_f32_avx+0x76> DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,141,21,132,0,0,0 ; lea 0x84(%rip),%r10 # 4d0c <_sk_load_f32_avx+0x9e> + DB 76,141,21,132,0,0,0 ; lea 0x84(%rip),%r10 # 4ca0 <_sk_load_f32_avx+0x9e> DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -8937,7 +8901,7 @@ _sk_store_f32_avx LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 4d99 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 4d2d <_sk_store_f32_avx+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -8950,22 +8914,22 @@ _sk_store_f32_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 4d95 <_sk_store_f32_avx+0x69> + DB 116,240 ; je 4d29 <_sk_store_f32_avx+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 4d95 <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 4d29 <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 4d95 <_sk_store_f32_avx+0x69> + DB 116,218 ; je 4d29 <_sk_store_f32_avx+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 4d95 <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 4d29 <_sk_store_f32_avx+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 4d95 <_sk_store_f32_avx+0x69> + DB 116,195 ; je 4d29 <_sk_store_f32_avx+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 4d95 <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 4d29 <_sk_store_f32_avx+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 4d95 <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 4d29 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -9269,7 +9233,7 @@ _sk_linear_gradient_avx LABEL PROC DB 196,226,125,24,88,28 ; vbroadcastss 0x1c(%rax),%ymm3 DB 76,139,0 ; mov (%rax),%r8 DB 77,133,192 ; test %r8,%r8 - DB 15,132,146,0,0,0 ; je 534d <_sk_linear_gradient_avx+0xb8> + DB 15,132,146,0,0,0 ; je 52e1 <_sk_linear_gradient_avx+0xb8> DB 72,139,64,8 ; mov 0x8(%rax),%rax DB 72,131,192,32 ; add $0x20,%rax DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12 @@ -9296,8 +9260,8 @@ _sk_linear_gradient_avx LABEL PROC DB 196,227,13,74,219,208 ; vblendvps %ymm13,%ymm3,%ymm14,%ymm3 DB 72,131,192,36 ; add $0x24,%rax DB 73,255,200 ; dec %r8 - DB 117,140 ; jne 52d7 <_sk_linear_gradient_avx+0x42> - DB 235,20 ; jmp 5361 <_sk_linear_gradient_avx+0xcc> + DB 117,140 ; jne 526b <_sk_linear_gradient_avx+0x42> + DB 235,20 ; jmp 52f5 <_sk_linear_gradient_avx+0xcc> DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 @@ -11882,54 +11846,50 @@ _sk_parametric_r_sse41 LABEL PROC DB 243,68,15,16,72,12 ; movss 0xc(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,89,200 ; mulps %xmm0,%xmm9 - DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 68,15,89,208 ; mulps %xmm0,%xmm10 + DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 68,15,89,216 ; mulps %xmm0,%xmm11 DB 65,15,194,192,2 ; cmpleps %xmm8,%xmm0 DB 243,68,15,16,64,24 ; movss 0x18(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 69,15,88,200 ; addps %xmm8,%xmm9 - DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 243,68,15,16,88,8 ; movss 0x8(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 243,68,15,16,16 ; movss (%rax),%xmm10 + DB 243,68,15,16,64,8 ; movss 0x8(%rax),%xmm8 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 + DB 69,15,88,216 ; addps %xmm8,%xmm11 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,227 ; mulps %xmm11,%xmm12 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 102,68,15,110,193 ; movd %ecx,%xmm8 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 + DB 69,15,89,196 ; mulps %xmm12,%xmm8 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 102,69,15,112,219,0 ; pshufd $0x0,%xmm11,%xmm11 - DB 102,69,15,219,218 ; pand %xmm10,%xmm11 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 102,69,15,112,228,0 ; pshufd $0x0,%xmm12,%xmm12 + DB 102,69,15,219,227 ; pand %xmm11,%xmm12 DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 102,69,15,112,210,0 ; pshufd $0x0,%xmm10,%xmm10 - DB 102,69,15,235,211 ; por %xmm11,%xmm10 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,88,220 ; addps %xmm12,%xmm11 + DB 102,69,15,112,219,0 ; pshufd $0x0,%xmm11,%xmm11 + DB 102,69,15,235,220 ; por %xmm12,%xmm11 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,92,196 ; subps %xmm12,%xmm8 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx DB 102,68,15,110,225 ; movd %ecx,%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 69,15,92,220 ; subps %xmm12,%xmm11 + DB 69,15,89,227 ; mulps %xmm11,%xmm12 + DB 69,15,92,196 ; subps %xmm12,%xmm8 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx DB 102,68,15,110,225 ; movd %ecx,%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,88,234 ; addps %xmm10,%xmm13 + DB 69,15,88,235 ; addps %xmm11,%xmm13 DB 69,15,94,229 ; divps %xmm13,%xmm12 - DB 69,15,92,220 ; subps %xmm12,%xmm11 - DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 69,15,89,195 ; mulps %xmm11,%xmm8 + DB 69,15,92,196 ; subps %xmm12,%xmm8 + DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,58,8,216,1 ; roundps $0x1,%xmm8,%xmm11 DB 185,0,0,0,75 ; mov $0x4b000000,%ecx DB 102,68,15,110,209 ; movd %ecx,%xmm10 @@ -11986,70 +11946,66 @@ _sk_parametric_g_sse41 LABEL PROC DB 243,15,16,72,24 ; movss 0x18(%rax),%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 68,15,88,201 ; addps %xmm1,%xmm9 - DB 243,68,15,16,32 ; movss (%rax),%xmm12 + DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 243,15,16,72,8 ; movss 0x8(%rax),%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 68,15,88,217 ; addps %xmm1,%xmm11 - DB 69,15,91,211 ; cvtdq2ps %xmm11,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 68,15,92,233 ; subps %xmm1,%xmm13 + DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx DB 102,15,110,201 ; movd %ecx,%xmm1 - DB 102,68,15,112,209,0 ; pshufd $0x0,%xmm1,%xmm10 - DB 102,69,15,219,211 ; pand %xmm11,%xmm10 + DB 102,68,15,112,225,0 ; pshufd $0x0,%xmm1,%xmm12 + DB 102,69,15,219,227 ; pand %xmm11,%xmm12 DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,68,15,112,217,0 ; pshufd $0x0,%xmm1,%xmm11 - DB 102,69,15,235,218 ; por %xmm10,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,213 ; addps %xmm13,%xmm10 + DB 102,69,15,235,220 ; por %xmm12,%xmm11 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,15,110,201 ; movd %ecx,%xmm1 + DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 + DB 68,15,92,233 ; subps %xmm1,%xmm13 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx DB 102,15,110,201 ; movd %ecx,%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 65,15,89,203 ; mulps %xmm11,%xmm1 - DB 68,15,92,209 ; subps %xmm1,%xmm10 + DB 68,15,92,233 ; subps %xmm1,%xmm13 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,15,110,201 ; movd %ecx,%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 65,15,88,203 ; addps %xmm11,%xmm1 - DB 68,15,94,233 ; divps %xmm1,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,58,8,220,1 ; roundps $0x1,%xmm12,%xmm11 + DB 68,15,94,225 ; divps %xmm1,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 69,15,89,234 ; mulps %xmm10,%xmm13 + DB 102,69,15,58,8,221,1 ; roundps $0x1,%xmm13,%xmm11 DB 185,0,0,0,75 ; mov $0x4b000000,%ecx DB 102,68,15,110,209 ; movd %ecx,%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,88,236 ; addps %xmm12,%xmm13 - DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 69,15,92,235 ; subps %xmm11,%xmm13 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx DB 102,15,110,201 ; movd %ecx,%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 65,15,89,204 ; mulps %xmm12,%xmm1 - DB 68,15,92,233 ; subps %xmm1,%xmm13 + DB 65,15,89,205 ; mulps %xmm13,%xmm1 + DB 68,15,92,225 ; subps %xmm1,%xmm12 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,15,110,201 ; movd %ecx,%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 - DB 65,15,92,204 ; subps %xmm12,%xmm1 + DB 65,15,92,205 ; subps %xmm13,%xmm1 DB 68,15,94,217 ; divps %xmm1,%xmm11 - DB 69,15,88,221 ; addps %xmm13,%xmm11 + DB 69,15,88,220 ; addps %xmm12,%xmm11 DB 69,15,89,218 ; mulps %xmm10,%xmm11 DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 DB 243,15,16,72,20 ; movss 0x14(%rax),%xmm1 @@ -12083,70 +12039,66 @@ _sk_parametric_b_sse41 LABEL PROC DB 243,15,16,80,24 ; movss 0x18(%rax),%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 DB 68,15,88,202 ; addps %xmm2,%xmm9 - DB 243,68,15,16,32 ; movss (%rax),%xmm12 + DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 243,15,16,80,8 ; movss 0x8(%rax),%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 DB 68,15,88,218 ; addps %xmm2,%xmm11 - DB 69,15,91,211 ; cvtdq2ps %xmm11,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 68,15,92,234 ; subps %xmm2,%xmm13 + DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 102,68,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm10 - DB 102,69,15,219,211 ; pand %xmm11,%xmm10 + DB 102,68,15,112,226,0 ; pshufd $0x0,%xmm2,%xmm12 + DB 102,69,15,219,227 ; pand %xmm11,%xmm12 DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 DB 102,68,15,112,218,0 ; pshufd $0x0,%xmm2,%xmm11 - DB 102,69,15,235,218 ; por %xmm10,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,213 ; addps %xmm13,%xmm10 + DB 102,69,15,235,220 ; por %xmm12,%xmm11 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,15,110,209 ; movd %ecx,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 68,15,92,234 ; subps %xmm2,%xmm13 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 DB 65,15,89,211 ; mulps %xmm11,%xmm2 - DB 68,15,92,210 ; subps %xmm2,%xmm10 + DB 68,15,92,234 ; subps %xmm2,%xmm13 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 DB 65,15,88,211 ; addps %xmm11,%xmm2 - DB 68,15,94,234 ; divps %xmm2,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,58,8,220,1 ; roundps $0x1,%xmm12,%xmm11 + DB 68,15,94,226 ; divps %xmm2,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 69,15,89,234 ; mulps %xmm10,%xmm13 + DB 102,69,15,58,8,221,1 ; roundps $0x1,%xmm13,%xmm11 DB 185,0,0,0,75 ; mov $0x4b000000,%ecx DB 102,68,15,110,209 ; movd %ecx,%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,88,236 ; addps %xmm12,%xmm13 - DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 69,15,92,235 ; subps %xmm11,%xmm13 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 65,15,89,212 ; mulps %xmm12,%xmm2 - DB 68,15,92,234 ; subps %xmm2,%xmm13 + DB 65,15,89,213 ; mulps %xmm13,%xmm2 + DB 68,15,92,226 ; subps %xmm2,%xmm12 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 - DB 65,15,92,212 ; subps %xmm12,%xmm2 + DB 65,15,92,213 ; subps %xmm13,%xmm2 DB 68,15,94,218 ; divps %xmm2,%xmm11 - DB 69,15,88,221 ; addps %xmm13,%xmm11 + DB 69,15,88,220 ; addps %xmm12,%xmm11 DB 69,15,89,218 ; mulps %xmm10,%xmm11 DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 DB 243,15,16,80,20 ; movss 0x14(%rax),%xmm2 @@ -12180,70 +12132,66 @@ _sk_parametric_a_sse41 LABEL PROC DB 243,15,16,88,24 ; movss 0x18(%rax),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 68,15,88,203 ; addps %xmm3,%xmm9 - DB 243,68,15,16,32 ; movss (%rax),%xmm12 + DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 243,15,16,88,8 ; movss 0x8(%rax),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 68,15,88,219 ; addps %xmm3,%xmm11 - DB 69,15,91,211 ; cvtdq2ps %xmm11,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,234 ; mulps %xmm10,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 68,15,92,235 ; subps %xmm3,%xmm13 + DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 - DB 102,68,15,112,211,0 ; pshufd $0x0,%xmm3,%xmm10 - DB 102,69,15,219,211 ; pand %xmm11,%xmm10 + DB 102,68,15,112,227,0 ; pshufd $0x0,%xmm3,%xmm12 + DB 102,69,15,219,227 ; pand %xmm11,%xmm12 DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,68,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm11 - DB 102,69,15,235,218 ; por %xmm10,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,88,213 ; addps %xmm13,%xmm10 + DB 102,69,15,235,220 ; por %xmm12,%xmm11 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,15,110,217 ; movd %ecx,%xmm3 + DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 + DB 68,15,92,235 ; subps %xmm3,%xmm13 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 65,15,89,219 ; mulps %xmm11,%xmm3 - DB 68,15,92,211 ; subps %xmm3,%xmm10 + DB 68,15,92,235 ; subps %xmm3,%xmm13 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 DB 65,15,88,219 ; addps %xmm11,%xmm3 - DB 68,15,94,235 ; divps %xmm3,%xmm13 - DB 69,15,92,213 ; subps %xmm13,%xmm10 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,89,226 ; mulps %xmm10,%xmm12 - DB 102,69,15,58,8,220,1 ; roundps $0x1,%xmm12,%xmm11 + DB 68,15,94,227 ; divps %xmm3,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 69,15,89,234 ; mulps %xmm10,%xmm13 + DB 102,69,15,58,8,221,1 ; roundps $0x1,%xmm13,%xmm11 DB 185,0,0,0,75 ; mov $0x4b000000,%ecx DB 102,68,15,110,209 ; movd %ecx,%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,88,236 ; addps %xmm12,%xmm13 - DB 69,15,92,227 ; subps %xmm11,%xmm12 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 69,15,92,235 ; subps %xmm11,%xmm13 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 65,15,89,220 ; mulps %xmm12,%xmm3 - DB 68,15,92,235 ; subps %xmm3,%xmm13 + DB 65,15,89,221 ; mulps %xmm13,%xmm3 + DB 68,15,92,227 ; subps %xmm3,%xmm12 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 65,15,92,220 ; subps %xmm12,%xmm3 + DB 65,15,92,221 ; subps %xmm13,%xmm3 DB 68,15,94,219 ; divps %xmm3,%xmm11 - DB 69,15,88,221 ; addps %xmm13,%xmm11 + DB 69,15,88,220 ; addps %xmm12,%xmm11 DB 69,15,89,218 ; mulps %xmm10,%xmm11 DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 DB 243,15,16,88,20 ; movss 0x14(%rax),%xmm3 @@ -12384,9 +12332,9 @@ _sk_gather_i8_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 73,137,192 ; mov %rax,%r8 DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 27a3 <_sk_gather_i8_sse41+0xf> + DB 116,5 ; je 2757 <_sk_gather_i8_sse41+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 27a5 <_sk_gather_i8_sse41+0x11> + DB 235,2 ; jmp 2759 <_sk_gather_i8_sse41+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 @@ -13480,7 +13428,7 @@ _sk_linear_gradient_sse41 LABEL PROC DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 DB 72,139,8 ; mov (%rax),%rcx DB 72,133,201 ; test %rcx,%rcx - DB 15,132,4,1,0,0 ; je 39d1 <_sk_linear_gradient_sse41+0x13e> + DB 15,132,4,1,0,0 ; je 3985 <_sk_linear_gradient_sse41+0x13e> DB 72,131,236,88 ; sub $0x58,%rsp DB 15,41,36,36 ; movaps %xmm4,(%rsp) DB 15,41,108,36,16 ; movaps %xmm5,0x10(%rsp) @@ -13531,13 +13479,13 @@ _sk_linear_gradient_sse41 LABEL PROC DB 15,40,196 ; movaps %xmm4,%xmm0 DB 72,131,192,36 ; add $0x24,%rax DB 72,255,201 ; dec %rcx - DB 15,133,65,255,255,255 ; jne 38f9 <_sk_linear_gradient_sse41+0x66> + DB 15,133,65,255,255,255 ; jne 38ad <_sk_linear_gradient_sse41+0x66> DB 15,40,124,36,48 ; movaps 0x30(%rsp),%xmm7 DB 15,40,116,36,32 ; movaps 0x20(%rsp),%xmm6 DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5 DB 15,40,36,36 ; movaps (%rsp),%xmm4 DB 72,131,196,88 ; add $0x58,%rsp - DB 235,13 ; jmp 39de <_sk_linear_gradient_sse41+0x14b> + DB 235,13 ; jmp 3992 <_sk_linear_gradient_sse41+0x14b> DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 DB 15,87,219 ; xorps %xmm3,%xmm3 @@ -16224,97 +16172,92 @@ _sk_parametric_r_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,192 ; mulps %xmm0,%xmm8 - DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 68,15,89,208 ; mulps %xmm0,%xmm10 + DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 68,15,89,216 ; mulps %xmm0,%xmm11 DB 65,15,194,193,2 ; cmpleps %xmm9,%xmm0 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,88,8 ; movss 0x8(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,88,218 ; addps %xmm10,%xmm11 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,235 ; subps %xmm11,%xmm13 + DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,219,226 ; pand %xmm10,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 102,69,15,112,218,0 ; pshufd $0x0,%xmm10,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,211 ; mulps %xmm11,%xmm10 - DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,243 ; addps %xmm11,%xmm14 + DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,89,204 ; mulps %xmm12,%xmm9 - DB 243,69,15,91,217 ; cvttps2dq %xmm9,%xmm11 - DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 - DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,213 ; subps %xmm13,%xmm10 + DB 69,15,89,209 ; mulps %xmm9,%xmm10 + DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 + DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 + DB 69,15,40,234 ; movaps %xmm10,%xmm13 DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,249 ; addps %xmm9,%xmm15 - DB 69,15,40,233 ; movaps %xmm9,%xmm13 - DB 69,15,194,235,1 ; cmpltps %xmm11,%xmm13 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,84,234 ; andps %xmm10,%xmm13 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,221 ; subps %xmm13,%xmm11 - DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 102,69,15,110,216 ; movd %r8d,%xmm11 + DB 69,15,88,250 ; addps %xmm10,%xmm15 + DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,84,211 ; andps %xmm11,%xmm10 + DB 69,15,87,201 ; xorps %xmm9,%xmm9 + DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 102,69,15,110,208 ; movd %r8d,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 69,15,92,253 ; subps %xmm13,%xmm15 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 69,15,92,252 ; subps %xmm12,%xmm15 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,244 ; subps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,88,239 ; addps %xmm15,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 102,69,15,91,221 ; cvtps2dq %xmm13,%xmm11 + DB 69,15,92,245 ; subps %xmm13,%xmm14 + DB 69,15,94,230 ; divps %xmm14,%xmm12 + DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,89,226 ; mulps %xmm10,%xmm12 + DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,227 ; addps %xmm11,%xmm12 + DB 69,15,88,226 ; addps %xmm10,%xmm12 DB 68,15,84,192 ; andps %xmm0,%xmm8 DB 65,15,85,196 ; andnps %xmm12,%xmm0 DB 65,15,86,192 ; orps %xmm8,%xmm0 DB 65,15,95,193 ; maxps %xmm9,%xmm0 - DB 65,15,93,194 ; minps %xmm10,%xmm0 + DB 65,15,93,195 ; minps %xmm11,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16326,97 +16269,92 @@ _sk_parametric_g_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,193 ; mulps %xmm1,%xmm8 - DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 68,15,89,209 ; mulps %xmm1,%xmm10 + DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 68,15,89,217 ; mulps %xmm1,%xmm11 DB 65,15,194,201,2 ; cmpleps %xmm9,%xmm1 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,88,8 ; movss 0x8(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,88,218 ; addps %xmm10,%xmm11 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,235 ; subps %xmm11,%xmm13 + DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,219,226 ; pand %xmm10,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 102,69,15,112,218,0 ; pshufd $0x0,%xmm10,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,211 ; mulps %xmm11,%xmm10 - DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,243 ; addps %xmm11,%xmm14 + DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,89,204 ; mulps %xmm12,%xmm9 - DB 243,69,15,91,217 ; cvttps2dq %xmm9,%xmm11 - DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 - DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,213 ; subps %xmm13,%xmm10 + DB 69,15,89,209 ; mulps %xmm9,%xmm10 + DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 + DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 + DB 69,15,40,234 ; movaps %xmm10,%xmm13 DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,249 ; addps %xmm9,%xmm15 - DB 69,15,40,233 ; movaps %xmm9,%xmm13 - DB 69,15,194,235,1 ; cmpltps %xmm11,%xmm13 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,84,234 ; andps %xmm10,%xmm13 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,221 ; subps %xmm13,%xmm11 - DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 102,69,15,110,216 ; movd %r8d,%xmm11 + DB 69,15,88,250 ; addps %xmm10,%xmm15 + DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,84,211 ; andps %xmm11,%xmm10 + DB 69,15,87,201 ; xorps %xmm9,%xmm9 + DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 102,69,15,110,208 ; movd %r8d,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 69,15,92,253 ; subps %xmm13,%xmm15 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 69,15,92,252 ; subps %xmm12,%xmm15 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,244 ; subps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,88,239 ; addps %xmm15,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 102,69,15,91,221 ; cvtps2dq %xmm13,%xmm11 + DB 69,15,92,245 ; subps %xmm13,%xmm14 + DB 69,15,94,230 ; divps %xmm14,%xmm12 + DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,89,226 ; mulps %xmm10,%xmm12 + DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,227 ; addps %xmm11,%xmm12 + DB 69,15,88,226 ; addps %xmm10,%xmm12 DB 68,15,84,193 ; andps %xmm1,%xmm8 DB 65,15,85,204 ; andnps %xmm12,%xmm1 DB 65,15,86,200 ; orps %xmm8,%xmm1 DB 65,15,95,201 ; maxps %xmm9,%xmm1 - DB 65,15,93,202 ; minps %xmm10,%xmm1 + DB 65,15,93,203 ; minps %xmm11,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16428,97 +16366,92 @@ _sk_parametric_b_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,194 ; mulps %xmm2,%xmm8 - DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 68,15,89,210 ; mulps %xmm2,%xmm10 + DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 68,15,89,218 ; mulps %xmm2,%xmm11 DB 65,15,194,209,2 ; cmpleps %xmm9,%xmm2 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,88,8 ; movss 0x8(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,88,218 ; addps %xmm10,%xmm11 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,235 ; subps %xmm11,%xmm13 + DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,219,226 ; pand %xmm10,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 102,69,15,112,218,0 ; pshufd $0x0,%xmm10,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,211 ; mulps %xmm11,%xmm10 - DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,243 ; addps %xmm11,%xmm14 + DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,89,204 ; mulps %xmm12,%xmm9 - DB 243,69,15,91,217 ; cvttps2dq %xmm9,%xmm11 - DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 - DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,213 ; subps %xmm13,%xmm10 + DB 69,15,89,209 ; mulps %xmm9,%xmm10 + DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 + DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 + DB 69,15,40,234 ; movaps %xmm10,%xmm13 DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,249 ; addps %xmm9,%xmm15 - DB 69,15,40,233 ; movaps %xmm9,%xmm13 - DB 69,15,194,235,1 ; cmpltps %xmm11,%xmm13 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,84,234 ; andps %xmm10,%xmm13 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,221 ; subps %xmm13,%xmm11 - DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 102,69,15,110,216 ; movd %r8d,%xmm11 + DB 69,15,88,250 ; addps %xmm10,%xmm15 + DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,84,211 ; andps %xmm11,%xmm10 + DB 69,15,87,201 ; xorps %xmm9,%xmm9 + DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 102,69,15,110,208 ; movd %r8d,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 69,15,92,253 ; subps %xmm13,%xmm15 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 69,15,92,252 ; subps %xmm12,%xmm15 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,244 ; subps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,88,239 ; addps %xmm15,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 102,69,15,91,221 ; cvtps2dq %xmm13,%xmm11 + DB 69,15,92,245 ; subps %xmm13,%xmm14 + DB 69,15,94,230 ; divps %xmm14,%xmm12 + DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,89,226 ; mulps %xmm10,%xmm12 + DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,227 ; addps %xmm11,%xmm12 + DB 69,15,88,226 ; addps %xmm10,%xmm12 DB 68,15,84,194 ; andps %xmm2,%xmm8 DB 65,15,85,212 ; andnps %xmm12,%xmm2 DB 65,15,86,208 ; orps %xmm8,%xmm2 DB 65,15,95,209 ; maxps %xmm9,%xmm2 - DB 65,15,93,210 ; minps %xmm10,%xmm2 + DB 65,15,93,211 ; minps %xmm11,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16530,97 +16463,92 @@ _sk_parametric_a_sse2 LABEL PROC DB 243,68,15,16,64,12 ; movss 0xc(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 68,15,89,195 ; mulps %xmm3,%xmm8 - DB 243,68,15,16,80,4 ; movss 0x4(%rax),%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 68,15,89,211 ; mulps %xmm3,%xmm10 + DB 243,68,15,16,88,4 ; movss 0x4(%rax),%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 68,15,89,219 ; mulps %xmm3,%xmm11 DB 65,15,194,217,2 ; cmpleps %xmm9,%xmm3 DB 243,68,15,16,72,24 ; movss 0x18(%rax),%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,88,193 ; addps %xmm9,%xmm8 DB 243,68,15,16,8 ; movss (%rax),%xmm9 - DB 243,68,15,16,88,8 ; movss 0x8(%rax),%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 + DB 243,68,15,16,80,8 ; movss 0x8(%rax),%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,88,218 ; addps %xmm10,%xmm11 + DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 DB 185,0,0,0,52 ; mov $0x34000000,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 185,0,0,254,66 ; mov $0x42fe0000,%ecx - DB 102,68,15,110,217 ; movd %ecx,%xmm11 - DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 - DB 69,15,92,235 ; subps %xmm11,%xmm13 + DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 + DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 185,255,255,127,0 ; mov $0x7fffff,%ecx + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 102,69,15,112,236,0 ; pshufd $0x0,%xmm12,%xmm13 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 185,0,0,0,63 ; mov $0x3f000000,%ecx DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 102,69,15,112,227,0 ; pshufd $0x0,%xmm11,%xmm12 - DB 102,69,15,219,226 ; pand %xmm10,%xmm12 - DB 185,0,0,0,63 ; mov $0x3f000000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 102,69,15,112,218,0 ; pshufd $0x0,%xmm10,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 185,42,145,49,64 ; mov $0x4031912a,%ecx - DB 102,68,15,110,225 ; movd %ecx,%xmm12 - DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,229 ; addps %xmm13,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 185,119,115,248,66 ; mov $0x42f87377,%ecx + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,117,191,191,63 ; mov $0x3fbfbf75,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,89,211 ; mulps %xmm11,%xmm10 - DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 + DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,89,220 ; mulps %xmm12,%xmm11 + DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 185,163,233,220,63 ; mov $0x3fdce9a3,%ecx DB 102,68,15,110,233 ; movd %ecx,%xmm13 DB 185,249,68,180,62 ; mov $0x3eb444f9,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 185,0,0,128,63 ; mov $0x3f800000,%ecx - DB 102,68,15,110,209 ; movd %ecx,%xmm10 + DB 102,68,15,110,217 ; movd %ecx,%xmm11 DB 65,184,0,0,0,75 ; mov $0x4b000000,%r8d DB 185,81,140,242,66 ; mov $0x42f28c51,%ecx DB 102,68,15,110,249 ; movd %ecx,%xmm15 + DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,88,243 ; addps %xmm11,%xmm14 + DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 - DB 69,15,89,204 ; mulps %xmm12,%xmm9 - DB 243,69,15,91,217 ; cvttps2dq %xmm9,%xmm11 - DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 - DB 69,15,40,225 ; movaps %xmm9,%xmm12 + DB 69,15,92,213 ; subps %xmm13,%xmm10 + DB 69,15,89,209 ; mulps %xmm9,%xmm10 + DB 243,69,15,91,202 ; cvttps2dq %xmm10,%xmm9 + DB 69,15,91,225 ; cvtdq2ps %xmm9,%xmm12 + DB 69,15,40,234 ; movaps %xmm10,%xmm13 DB 69,15,198,255,0 ; shufps $0x0,%xmm15,%xmm15 - DB 69,15,88,249 ; addps %xmm9,%xmm15 - DB 69,15,40,233 ; movaps %xmm9,%xmm13 - DB 69,15,194,235,1 ; cmpltps %xmm11,%xmm13 - DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 69,15,84,234 ; andps %xmm10,%xmm13 - DB 69,15,87,201 ; xorps %xmm9,%xmm9 - DB 69,15,92,221 ; subps %xmm13,%xmm11 - DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 102,69,15,110,216 ; movd %r8d,%xmm11 + DB 69,15,88,250 ; addps %xmm10,%xmm15 + DB 69,15,194,212,1 ; cmpltps %xmm12,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 + DB 69,15,84,211 ; andps %xmm11,%xmm10 + DB 69,15,87,201 ; xorps %xmm9,%xmm9 + DB 69,15,92,226 ; subps %xmm10,%xmm12 + DB 69,15,92,236 ; subps %xmm12,%xmm13 + DB 102,69,15,110,208 ; movd %r8d,%xmm10 + DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 185,141,188,190,63 ; mov $0x3fbebc8d,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 69,15,92,253 ; subps %xmm13,%xmm15 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 69,15,92,252 ; subps %xmm12,%xmm15 DB 185,254,210,221,65 ; mov $0x41ddd2fe,%ecx - DB 102,68,15,110,233 ; movd %ecx,%xmm13 - DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13 + DB 102,68,15,110,225 ; movd %ecx,%xmm12 + DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 185,248,245,154,64 ; mov $0x409af5f8,%ecx DB 102,68,15,110,241 ; movd %ecx,%xmm14 DB 69,15,198,246,0 ; shufps $0x0,%xmm14,%xmm14 - DB 69,15,92,244 ; subps %xmm12,%xmm14 - DB 69,15,94,238 ; divps %xmm14,%xmm13 - DB 69,15,88,239 ; addps %xmm15,%xmm13 - DB 69,15,89,235 ; mulps %xmm11,%xmm13 - DB 102,69,15,91,221 ; cvtps2dq %xmm13,%xmm11 + DB 69,15,92,245 ; subps %xmm13,%xmm14 + DB 69,15,94,230 ; divps %xmm14,%xmm12 + DB 69,15,88,231 ; addps %xmm15,%xmm12 + DB 69,15,89,226 ; mulps %xmm10,%xmm12 + DB 102,69,15,91,212 ; cvtps2dq %xmm12,%xmm10 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 - DB 69,15,88,227 ; addps %xmm11,%xmm12 + DB 69,15,88,226 ; addps %xmm10,%xmm12 DB 68,15,84,195 ; andps %xmm3,%xmm8 DB 65,15,85,220 ; andnps %xmm12,%xmm3 DB 65,15,86,216 ; orps %xmm8,%xmm3 DB 65,15,95,217 ; maxps %xmm9,%xmm3 - DB 65,15,93,218 ; minps %xmm10,%xmm3 + DB 65,15,93,219 ; minps %xmm11,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -16780,9 +16708,9 @@ _sk_gather_i8_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 73,137,192 ; mov %rax,%r8 DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 29b9 <_sk_gather_i8_sse2+0xf> + DB 116,5 ; je 295d <_sk_gather_i8_sse2+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 29bb <_sk_gather_i8_sse2+0x11> + DB 235,2 ; jmp 295f <_sk_gather_i8_sse2+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 @@ -17979,7 +17907,7 @@ _sk_linear_gradient_sse2 LABEL PROC DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 DB 72,139,8 ; mov (%rax),%rcx DB 72,133,201 ; test %rcx,%rcx - DB 15,132,15,1,0,0 ; je 3da3 <_sk_linear_gradient_sse2+0x149> + DB 15,132,15,1,0,0 ; je 3d47 <_sk_linear_gradient_sse2+0x149> DB 72,139,64,8 ; mov 0x8(%rax),%rax DB 72,131,192,32 ; add $0x20,%rax DB 69,15,87,192 ; xorps %xmm8,%xmm8 @@ -18040,8 +17968,8 @@ _sk_linear_gradient_sse2 LABEL PROC DB 69,15,86,231 ; orps %xmm15,%xmm12 DB 72,131,192,36 ; add $0x24,%rax DB 72,255,201 ; dec %rcx - DB 15,133,8,255,255,255 ; jne 3ca9 <_sk_linear_gradient_sse2+0x4f> - DB 235,13 ; jmp 3db0 <_sk_linear_gradient_sse2+0x156> + DB 15,133,8,255,255,255 ; jne 3c4d <_sk_linear_gradient_sse2+0x4f> + DB 235,13 ; jmp 3d54 <_sk_linear_gradient_sse2+0x156> DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 DB 15,87,219 ; xorps %xmm3,%xmm3 diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp index a7f6d8036e..76ea648b5a 100644 --- a/src/jumper/SkJumper_stages.cpp +++ b/src/jumper/SkJumper_stages.cpp @@ -671,30 +671,6 @@ STAGE(table_g) { g = table(g, ctx); } STAGE(table_b) { b = table(b, ctx); } STAGE(table_a) { a = table(a, ctx); } -// See http://www.machinedlearnings.com/2011/06/fast-approximate-logarithm-exponential.html. -SI F approx_log2(F x) { - // e is a fair approximation of log2(x) in its own right... - F e = cast(bit_cast(x)) * C(1.0f / (1<<23)) - 127.0_f; - - // ... but using the mantissa to refine its error is _much_ better. - F m = bit_cast((bit_cast(x) & 0x007fffff_i) | 0x3f000000_i); - return e - + 2.774485010_f - - 1.498030302_f * m - - 1.725879990_f / (0.3520887068_f + m); -} -SI F approx_pow2(F x) { - F f = fract(x); - return bit_cast(round(C(1.0f * (1<<23)), - x + 121.2740575_f - - 1.490129070_f * f - + 27.72802330_f / (4.84252568_f - f))); -} - -SI F approx_powf(F x, float g) { - return approx_pow2(approx_log2(x) * g); -} - SI F parametric(F v, const SkJumper_ParametricTransferFunction* ctx) { F r = if_then_else(v <= ctx->D, mad(ctx->C, v, ctx->F) , approx_powf(mad(ctx->A, v, ctx->B), ctx->G) + ctx->E); diff --git a/src/jumper/SkJumper_vectors.h b/src/jumper/SkJumper_vectors.h index 590fe9c077..bd8ad40262 100644 --- a/src/jumper/SkJumper_vectors.h +++ b/src/jumper/SkJumper_vectors.h @@ -628,4 +628,28 @@ SI U16 bswap(U16 x) { SI F fract(F v) { return v - floor_(v); } +// See http://www.machinedlearnings.com/2011/06/fast-approximate-logarithm-exponential.html. +SI F approx_log2(F x) { + // e - 127 is a fair approximation of log2(x) in its own right... + F e = cast(bit_cast(x)) * C(1.0f / (1<<23)); + + // ... but using the mantissa to refine its error is _much_ better. + F m = bit_cast((bit_cast(x) & 0x007fffff_i) | 0x3f000000_i); + return e + - 124.225514990_f + - 1.498030302_f * m + - 1.725879990_f / (0.3520887068_f + m); +} +SI F approx_pow2(F x) { + F f = fract(x); + return bit_cast(round(C(1.0f * (1<<23)), + x + 121.274057500_f + - 1.490129070_f * f + + 27.728023300_f / (4.84252568_f - f))); +} + +SI F approx_powf(F x, F y) { + return approx_pow2(approx_log2(x) * y); +} + #endif//SkJumper_vectors_DEFINED