Revert of SSE4 opaque blend using intrinsics instead of assembly. (patchset #16 id:300001 of https://codereview.chromium.org/874863002/)

Reason for revert:
This causes a bug on the 'hittestpath' GM on MacMini 4,1

See:

https://gold.skia.org/#/triage/hittestpath?head=0

for details.

Original issue's description:
> SSE4 opaque blend using intrinsics instead of assembly.
>
> Since we had such a hard time with the assembly versions of this blit (to the
> point that we have them completely disabled everywhere), I thought I'd take
> a shot at writing a version of the blit using intrinsics.
>
> The key feature of SSE4 we're exploiting is that we can use ptest (_mm_test*)
> to skip the blend when the 16 src pixels we consider each loop are all opaque
> or all transparent.  _mm_shuffle_epi8 from SSSE3 also lends a hand to extract
> all those alphas.
>
> It's worth looking to see if we can backport this type of logic to SSE2 using
> _mm_movemask_epi8, or up to 32 pixels at a time using AVX.
>
> My local performance testing doesn't show this to be an unambiguous win
> (there are probably microbenchmarks and SKPs where we'd be better off just
> powering through the blend rather than looking at alphas), but the potential
> does seem tantalizing enough to let skiaperf vet it on the bots.  (< 1.0x is a win.)
>
> DM says it draws pixel perfect compare to the old code.
>
> Microbenchmarks:
>                bitmap_RGBA_8888_A_source_stripes_two	  14us -> 14.4us	1.03x
>              bitmap_RGBA_8888_A_source_stripes_three	14.3us -> 14.5us	1.01x
>                        bitmap_RGBA_8888_scale_bilerp	61.9us -> 62.2us	1.01x
> bitmap_RGBA_8888_update_volatile_scale_rotate_bilerp	 102us ->  101us	0.99x
>                 bitmap_RGBA_8888_scale_rotate_bilerp	 103us ->  101us	0.99x
>                               bitmap_RGBA_8888_scale	18.4us -> 18.2us	0.99x
>              bitmap_RGBA_8888_A_scale_rotate_bicubic	  71us ->   70us	0.99x
>          bitmap_RGBA_8888_update_scale_rotate_bilerp	 103us ->  101us	0.99x
>               bitmap_RGBA_8888_A_scale_rotate_bilerp	 112us ->  109us	0.98x
>                     bitmap_RGBA_8888_update_volatile	5.72us -> 5.58us	0.98x
>                                     bitmap_RGBA_8888	5.73us -> 5.58us	0.97x
>                              bitmap_RGBA_8888_update	5.78us ->  5.6us	0.97x
>                      bitmap_RGBA_8888_A_scale_bilerp	70.7us ->   68us	0.96x
>                     bitmap_RGBA_8888_A_scale_bicubic	23.7us -> 21.8us	0.92x
>                                   bitmap_RGBA_8888_A	13.9us -> 10.9us	0.78x
>                     bitmap_RGBA_8888_A_source_opaque	  14us -> 6.29us	0.45x
>                bitmap_RGBA_8888_A_source_transparent	  14us -> 3.65us	0.26x
>
> Running over our ~70 SKP web page captures, this looks like we spend 0.7x
> the time in S32A_Opaque_BlitRow compared to the SSE2 version, which should
> be a decent predictor of real-world impact.
>
> BUG=chromium:399842
>
> Committed: https://skia.googlesource.com/skia/+/04bc91b972417038fecfa87c484771eac2b9b785
>
> CQ_EXTRA_TRYBOTS=client.skia:Test-Mac10.6-MacMini4.1-GeForce320M-x86_64-Release-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/6dbfb21a6c88af6d94e8c823c3ad559f1a41b493

TBR=henrik.smiding@intel.com,mtklein@google.com,herb@google.com,reed@google.com,thakis@chromium.org,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:399842

Review URL: https://codereview.chromium.org/873553003

7 files changed, 972 insertions, 81 deletions

diff --git a/gyp/opts.gypi b/gyp/opts.gypi
index dfcf434aff..f6257a97b4 100644
--- a/gyp/opts.gypi
+++ b/gyp/opts.gypi
@@ -79,6 +79,5 @@
         ],
         'sse41_sources': [
             '<(skia_src_path)/opts/SkBlurImage_opts_SSE4.cpp',
-            '<(skia_src_path)/opts/SkBlitRow_opts_SSE4.cpp',
         ],
 }
diff --git a/src/opts/SkBlitRow_opts_SSE4.cpp b/src/opts/SkBlitRow_opts_SSE4.cpp
deleted file mode 100644
index ae92a77eb2..0000000000
--- a/src/opts/SkBlitRow_opts_SSE4.cpp
+++ /dev/null
@@ -1,67 +0,0 @@
-#include "SkBlitRow_opts_SSE4.h"
-
-// Some compilers can't compile SSSE3 or SSE4 intrinsics.  We give them stub methods.
-// The stubs should never be called, so we make them crash just to confirm that.
-#if SK_CPU_SSE_LEVEL < SK_CPU_SSE_LEVEL_SSE41
-void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT, const SkPMColor* SK_RESTRICT, int, U8CPU) {
-    sk_throw();
-}
-
-#else
-
-#include <emmintrin.h>  // SSE2:   Most _mm_foo() in this file.
-#include <smmintrin.h>  // SSE4.1: _mm_testz_si128 and _mm_testc_si128.
-
-#include "SkColorPriv.h"
-#include "SkColor_opts_SSE2.h"
-
-void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst,
-                                const SkPMColor* SK_RESTRICT src,
-                                int count,
-                                U8CPU alpha) {
-    SkASSERT(alpha == 255);
-    // As long as we can, we'll work on 16 pixel pairs at once.
-    int count16 = count / 16;
-    __m128i* dst4 = (__m128i*)dst;
-    const __m128i* src4 = (const __m128i*)src;
-
-    for (int i = 0; i < count16 * 4; i += 4) {
-        // Load 16 source pixels.
-        __m128i s0 = _mm_loadu_si128(src4+i+0),
-                s1 = _mm_loadu_si128(src4+i+1),
-                s2 = _mm_loadu_si128(src4+i+2),
-                s3 = _mm_loadu_si128(src4+i+3);
-
-        const __m128i alphaMask = _mm_set1_epi32(0xFF << SK_A32_SHIFT);
-        const __m128i ORed = _mm_or_si128(s3, _mm_or_si128(s2, _mm_or_si128(s1, s0)));
-        if (_mm_testz_si128(ORed, alphaMask)) {
-            // All 16 source pixels are fully transparent.  There's nothing to do!
-            continue;
-        }
-        const __m128i ANDed = _mm_and_si128(s3, _mm_and_si128(s2, _mm_and_si128(s1, s0)));
-        if (_mm_testc_si128(ANDed, alphaMask)) {
-            // All 16 source pixels are fully opaque.  There's no need to read dst or blend it.
-            _mm_storeu_si128(dst4+i+0, s0);
-            _mm_storeu_si128(dst4+i+1, s1);
-            _mm_storeu_si128(dst4+i+2, s2);
-            _mm_storeu_si128(dst4+i+3, s3);
-            continue;
-        }
-        // The general slow case: do the blend for all 16 pixels.
-        _mm_storeu_si128(dst4+i+0, SkPMSrcOver_SSE2(s0, _mm_loadu_si128(dst4+i+0)));
-        _mm_storeu_si128(dst4+i+1, SkPMSrcOver_SSE2(s1, _mm_loadu_si128(dst4+i+1)));
-        _mm_storeu_si128(dst4+i+2, SkPMSrcOver_SSE2(s2, _mm_loadu_si128(dst4+i+2)));
-        _mm_storeu_si128(dst4+i+3, SkPMSrcOver_SSE2(s3, _mm_loadu_si128(dst4+i+3)));
-    }
-
-    // Wrap up the last <= 15 pixels.
-    SkASSERT(count - (count16*16) <= 15);
-    for (int i = count16*16; i < count; i++) {
-        // This check is not really necessarily, but it prevents pointless autovectorization.
-        if (src[i] & 0xFF000000) {
-            dst[i] = SkPMSrcOver(src[i], dst[i]);
-        }
-    }
-}
-
-#endif
diff --git a/src/opts/SkBlitRow_opts_SSE4.h b/src/opts/SkBlitRow_opts_SSE4.h
index 577ace6f8f..600e669893 100644
--- a/src/opts/SkBlitRow_opts_SSE4.h
+++ b/src/opts/SkBlitRow_opts_SSE4.h
@@ -10,9 +10,24 @@
 
 #include "SkBlitRow.h"
 
-void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT,
-                                const SkPMColor* SK_RESTRICT,
-                                int count,
-                                U8CPU alpha);
+#ifdef CRBUG_399842_FIXED
+
+/* Check if we are able to build assembly code, GCC/AT&T syntax:
+ *  1) Clang and GCC are generally OK.  OS X's old LLVM-GCC 4.2 can't handle it;
+ *  2) We're intentionally not linking this in even when supported (Clang) on Windows;
+ *  3) MemorySanitizer cannot instrument assembly at all.
+ */
+#if /* 1)*/ (defined(__clang__) || (defined(__GNUC__) && !defined(SK_BUILD_FOR_MAC))) \
+    /* 2)*/ && !defined(SK_BUILD_FOR_WIN)                                             \
+    /* 3)*/ && !defined(MEMORY_SANITIZER)
+extern "C" void S32A_Opaque_BlitRow32_SSE4_asm(SkPMColor* SK_RESTRICT dst,
+                                               const SkPMColor* SK_RESTRICT src,
+                                               int count, U8CPU alpha);
+
+#define SK_ATT_ASM_SUPPORTED
+#endif
+
+#endif // CRBUG_399842_FIXED
+
 #endif
 
diff --git a/src/opts/SkBlitRow_opts_SSE4_asm.S b/src/opts/SkBlitRow_opts_SSE4_asm.S
new file mode 100644
index 0000000000..0f5281713d
--- /dev/null
+++ b/src/opts/SkBlitRow_opts_SSE4_asm.S
@@ -0,0 +1,475 @@
+/*
+ * Copyright 2014 The Android Open Source Project
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifdef CRBUG_399842_FIXED
+
+#if defined(__clang__) || (defined(__GNUC__) && !defined(SK_BUILD_FOR_MAC))
+
+#define CFI_PUSH(REG) \
+    .cfi_adjust_cfa_offset 4; \
+    .cfi_rel_offset REG, 0
+
+#define CFI_POP(REG) \
+    .cfi_adjust_cfa_offset -4; \
+    .cfi_restore REG
+
+#define PUSH(REG) pushl REG; CFI_PUSH (REG)
+#define POP(REG)  popl REG; CFI_POP (REG)
+#define RETURN    POP(%edi); ret
+
+#define EXTRACT_ALPHA(var1, var2) \
+    movdqa      %var1, %var2;           /* Clone source pixels to extract alpha */\
+    psrlw       $8, %var2;              /* Discard red and blue, leaving alpha and green */\
+    pshufhw     $0xF5, %var2, %var2;    /* Repeat alpha for scaling (high) */\
+    movdqa      %xmm6, %xmm4;           \
+    pshuflw     $0xF5, %var2, %var2;    /* Repeat alpha for scaling (low) */\
+    movdqa      %xmm5, %xmm3;           \
+    psubw       %var2, %xmm4            /* Finalize alpha calculations */
+
+#define SCALE_PIXELS \
+    psllw       $8, %xmm5;              /* Filter out red and blue components */\
+    pmulhuw     %xmm4, %xmm5;           /* Scale red and blue */\
+    psrlw       $8, %xmm3;              /* Filter out alpha and green components */\
+    pmullw      %xmm4, %xmm3            /* Scale alpha and green */
+
+
+/*
+ * void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst,
+ *                                 const SkPMColor* SK_RESTRICT src,
+ *                                 int count, U8CPU alpha)
+ *
+ * This function is divided into six blocks: initialization, blit 4-15 pixels,
+ * blit 0-3 pixels, align destination for 16+ pixel blits,
+ * blit 16+ pixels with source unaligned, blit 16+ pixels with source aligned.
+ * There are some code reuse between the blocks.
+ *
+ * The primary optimization comes from checking the source pixels' alpha value.
+ * If the alpha is zero, the pixel can be skipped entirely.
+ * If the alpha is fully opaque, the pixel can be copied directly to the destination.
+ * According to collected statistics, these two cases are the most common.
+ * The main loop(s) uses pre-loading and unrolling in an attempt to reduce the
+ * memory latency worse-case.
+ */
+
+#ifdef __clang__
+    .text
+#else
+    .section .text.sse4.2,"ax",@progbits
+    .type S32A_Opaque_BlitRow32_SSE4_asm, @function
+#endif
+    .p2align 4
+#if defined(SK_BUILD_FOR_MAC)
+    .global _S32A_Opaque_BlitRow32_SSE4_asm
+    .private_extern _S32A_Opaque_BlitRow32_SSE4_asm
+_S32A_Opaque_BlitRow32_SSE4_asm:
+#else
+    .global S32A_Opaque_BlitRow32_SSE4_asm
+    .hidden S32A_Opaque_BlitRow32_SSE4_asm
+S32A_Opaque_BlitRow32_SSE4_asm:
+#endif
+    .cfi_startproc
+    movl        8(%esp), %eax           // Source pointer
+    movl        12(%esp), %ecx          // Pixel count
+    movl        4(%esp), %edx           // Destination pointer
+    prefetcht0  (%eax)
+
+    // Setup SSE constants
+    pcmpeqd     %xmm7, %xmm7            // 0xFF000000 mask to check alpha
+    pslld       $24, %xmm7
+    pcmpeqw     %xmm6, %xmm6            // 16-bit 256 to calculate inv. alpha
+    psrlw       $15, %xmm6
+    psllw       $8, %xmm6
+    pcmpeqw     %xmm0, %xmm0            // 0x00FF00FF mask (Must be in xmm0 because of pblendvb)
+    psrlw       $8, %xmm0
+    subl        $4, %ecx                // Check if we have only 0-3 pixels
+    js          .LReallySmall
+    PUSH(%edi)
+    cmpl        $11, %ecx               // Do we have enough pixels to run the main loop?
+    ja          .LBigBlit
+
+    // Handle small blits (4-15 pixels)
+    ////////////////////////////////////////////////////////////////////////////////
+    xorl        %edi, %edi              // Reset offset to zero
+
+.LSmallLoop:
+    lddqu       (%eax, %edi), %xmm1     // Load four source pixels
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LSmallAlphaNotOpaqueOrZero
+    jz          .LSmallAlphaZero        // If all alphas are zero, skip the pixels completely
+    movdqu      %xmm1, (%edx, %edi)     // Store four destination pixels
+.LSmallAlphaZero:
+    addl        $16, %edi
+    subl        $4, %ecx                // Check if there are four additional pixels, at least
+    jns         .LSmallLoop
+    jmp         .LSmallRemaining
+
+    // Handle mixed alphas (calculate and scale)
+    .p2align 4
+.LSmallAlphaNotOpaqueOrZero:
+    lddqu       (%edx, %edi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    addl        $16, %edi
+    subl        $4, %ecx                // Check if there are four additional pixels, at least
+    pblendvb    %xmm5, %xmm3            // Mask in %xmm0, implicitly
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqu      %xmm1, -16(%edx, %edi)  // Store four destination pixels
+    jns         .LSmallLoop
+
+    // Handle the last 0-3 pixels (also used by the main loops)
+.LSmallRemaining:
+    cmpl        $-4, %ecx               // Check if we are done
+    je          .LSmallExit
+    sall        $2, %ecx                // Calculate offset for last pixels
+    addl        %ecx, %edi
+
+    lddqu       (%eax, %edi), %xmm1     // Load last four source pixels (overlapping)
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    jc          .LSmallRemainingStoreAll// If all alphas are opaque, just store (overlapping)
+    jz          .LSmallExit             // If all alphas are zero, skip the pixels completely
+
+    // Handle mixed alphas (calculate and scale)
+    lddqu       (%edx, %edi), %xmm5     // Load last four destination pixels (overlapping)
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+
+    psllw       $8, %xmm3               // Filter out red and blue components
+    pmulhuw     %xmm4, %xmm3            // Scale red and blue
+    movdqa      %xmm5, %xmm2
+    psrlw       $8, %xmm2               // Filter out alpha and green components
+    pmullw      %xmm4, %xmm2            // Scale alpha and green
+
+    cmpl        $-8, %ecx               // Check how many pixels should be written
+    pblendvb    %xmm3, %xmm2            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm2, %xmm1            // Add source and destination pixels together
+    jb          .LSmallPixelsLeft1
+    ja          .LSmallPixelsLeft3      // To avoid double-blending the overlapping pixels...
+    pblendw     $0xF0, %xmm1, %xmm5     // Merge only the final two pixels to the destination
+    movdqu      %xmm5, (%edx, %edi)     // Store last two destination pixels
+.LSmallExit:
+    RETURN
+
+.LSmallPixelsLeft1:
+    pblendw     $0xC0, %xmm1, %xmm5     // Merge only the final pixel to the destination
+    movdqu      %xmm5, (%edx, %edi)     // Store last destination pixel
+    RETURN
+
+.LSmallPixelsLeft3:
+    pblendw     $0xFC, %xmm1, %xmm5     // Merge only the final three pixels to the destination
+    movdqu      %xmm5, (%edx, %edi)     // Store last three destination pixels
+    RETURN
+
+.LSmallRemainingStoreAll:
+    movdqu      %xmm1, (%edx, %edi)     // Store last destination pixels (overwrite)
+    RETURN
+
+    // Handle really small blits (0-3 pixels)
+    ////////////////////////////////////////////////////////////////////////////////
+.LReallySmall:
+    addl        $4, %ecx
+    jle         .LReallySmallExit
+    pcmpeqd     %xmm1, %xmm1
+    cmp         $2, %ecx                // Check how many pixels should be read
+    pinsrd      $0x0, (%eax), %xmm1     // Load one source pixel
+    pinsrd      $0x0, (%edx), %xmm5     // Load one destination pixel
+    jb          .LReallySmallCalc
+    pinsrd      $0x1, 4(%eax), %xmm1    // Load second source pixel
+    pinsrd      $0x1, 4(%edx), %xmm5    // Load second destination pixel
+    je          .LReallySmallCalc
+    pinsrd      $0x2, 8(%eax), %xmm1    // Load third source pixel
+    pinsrd      $0x2, 8(%edx), %xmm5    // Load third destination pixel
+
+.LReallySmallCalc:
+    ptest       %xmm7, %xmm1            // Check if all alphas are opaque
+    jc          .LReallySmallStore      // If all alphas are opaque, just store
+
+    // Handle mixed alphas (calculate and scale)
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+
+    pand        %xmm0, %xmm5            // Filter out red and blue components
+    pmullw      %xmm4, %xmm5            // Scale red and blue
+    psrlw       $8, %xmm3               // Filter out alpha and green components
+    pmullw      %xmm4, %xmm3            // Scale alpha and green
+
+    psrlw       $8, %xmm5               // Combine results
+    pblendvb    %xmm5, %xmm3            // Mask in %xmm0, implicitly
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+
+.LReallySmallStore:
+    cmp         $2, %ecx                // Check how many pixels should be written
+    pextrd      $0x0, %xmm1, (%edx)     // Store one destination pixel
+    jb          .LReallySmallExit
+    pextrd      $0x1, %xmm1, 4(%edx)    // Store second destination pixel
+    je          .LReallySmallExit
+    pextrd      $0x2, %xmm1, 8(%edx)    // Store third destination pixel
+.LReallySmallExit:
+    ret
+
+    // Handle bigger blit operations (16+ pixels)
+    ////////////////////////////////////////////////////////////////////////////////
+    .p2align 4
+.LBigBlit:
+    // Align destination?
+    testl       $0xF, %edx
+    lddqu       (%eax), %xmm1           // Pre-load four source pixels
+    jz          .LAligned
+
+    movl        %edx, %edi              // Calculate alignment of destination pointer
+    negl        %edi
+    andl        $0xF, %edi
+
+    // Handle 1-3 pixels to align destination
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    jz          .LAlignDone             // If all alphas are zero, just skip
+    lddqu       (%edx), %xmm5           // Load four destination pixels
+    jc          .LAlignStore            // If all alphas are opaque, just store
+
+    // Handle mixed alphas (calculate and scale)
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+
+    psllw       $8, %xmm3               // Filter out red and blue components
+    pmulhuw     %xmm4, %xmm3            // Scale red and blue
+    movdqa      %xmm5, %xmm2
+    psrlw       $8, %xmm2               // Filter out alpha and green components
+    pmullw      %xmm4, %xmm2            // Scale alpha and green
+
+    pblendvb    %xmm3, %xmm2            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm2, %xmm1            // Add source and destination pixels together
+
+.LAlignStore:
+    cmp         $8, %edi                // Check how many pixels should be written
+    jb          .LAlignPixelsLeft1
+    ja          .LAlignPixelsLeft3
+    pblendw     $0x0F, %xmm1, %xmm5     // Blend two pixels
+    jmp .LAlignStorePixels
+
+.LAlignPixelsLeft1:
+    pblendw     $0x03, %xmm1, %xmm5     // Blend one pixel
+    jmp .LAlignStorePixels
+
+.LAlignPixelsLeft3:
+    pblendw     $0x3F, %xmm1, %xmm5     // Blend three pixels
+
+.LAlignStorePixels:
+    movdqu      %xmm5, (%edx)           // Store destination pixels
+
+.LAlignDone:
+    addl        %edi, %eax              // Adjust pointers and pixel count
+    addl        %edi, %edx
+    shrl        $2, %edi
+    lddqu       (%eax), %xmm1           // Pre-load new source pixels (after alignment)
+    subl        %edi, %ecx
+
+.LAligned:                              // Destination is guaranteed to be 16 byte aligned
+    xorl        %edi, %edi              // Reset offset to zero
+    subl        $8, %ecx                // Decrease counter (Reserve four pixels for the cleanup)
+    testl       $0xF, %eax              // Check alignment of source pointer
+    jz          .LAlignedLoop
+
+    // Source not aligned to destination
+    ////////////////////////////////////////////////////////////////////////////////
+    .p2align 4
+.LUnalignedLoop:                        // Main loop for unaligned, handles eight pixels per iteration
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero00
+    lddqu       16(%eax, %edi), %xmm2   // Pre-load four source pixels
+    jz          .LAlphaZero00
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+
+.LAlphaZero00:
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero01
+    lddqu       32(%eax, %edi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero01
+    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels
+
+.LAlphaZero01:
+    addl        $32, %edi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LUnalignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup0
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero00:
+    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    lddqu       16(%eax, %edi), %xmm2   // Pre-load four source pixels
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+
+    // Handle next four pixels
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero01
+    lddqu       32(%eax, %edi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero02
+    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels
+.LAlphaZero02:
+    addl        $32, %edi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LUnalignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup0
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero01:
+    movdqa      16(%edx, %edi), %xmm5   // Load four destination pixels
+    EXTRACT_ALPHA(xmm2, xmm1)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    lddqu       32(%eax, %edi), %xmm1   // Pre-load four source pixels
+    addl        $32, %edi
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm2            // Add source and destination pixels together
+    subl        $8, %ecx
+    movdqa      %xmm2, -16(%edx, %edi)  // Store four destination pixels
+    jae         .LUnalignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+
+    // Cleanup - handle pending pixels from loop
+.LLoopCleanup0:
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero02
+    jz          .LAlphaZero03
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+.LAlphaZero03:
+    addl        $16, %edi
+    subl        $4, %ecx
+    js          .LSmallRemaining        // Reuse code from small loop
+
+.LRemain0:
+    lddqu       (%eax, %edi), %xmm1     // Load four source pixels
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero02
+    jz          .LAlphaZero04
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+.LAlphaZero04:
+    addl        $16, %edi
+    subl        $4, %ecx
+    jmp         .LSmallRemaining        // Reuse code from small loop
+
+.LAlphaNotOpaqueOrZero02:
+    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    addl        $16, %edi
+    subl        $4, %ecx
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, -16(%edx, %edi)  // Store four destination pixels
+    js          .LSmallRemaining        // Reuse code from small loop
+    jmp         .LRemain0
+
+    // Source aligned to destination
+    ////////////////////////////////////////////////////////////////////////////////
+    .p2align 4
+.LAlignedLoop:                          // Main loop for aligned, handles eight pixels per iteration
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero10
+    movdqa      16(%eax, %edi), %xmm2   // Pre-load four source pixels
+    jz          .LAlphaZero10
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+
+.LAlphaZero10:
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero11
+    movdqa      32(%eax, %edi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero11
+    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels
+
+.LAlphaZero11:
+    addl        $32, %edi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LAlignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup1
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero10:
+    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    movdqa      16(%eax, %edi), %xmm2   // Pre-load four source pixels
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+
+    // Handle next four pixels
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero11
+    movdqa      32(%eax, %edi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero12
+    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels
+.LAlphaZero12:
+    addl        $32, %edi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LAlignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup1
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero11:
+    movdqa      16(%edx, %edi), %xmm5   // Load four destination pixels
+    EXTRACT_ALPHA(xmm2, xmm1)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+    movdqa      32(%eax, %edi), %xmm1   // Pre-load four source pixels
+
+    addl        $32, %edi
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm2            // Add source and destination pixels together
+    subl        $8, %ecx
+    movdqa      %xmm2, -16(%edx, %edi)  // Store four destination pixels
+    jae         .LAlignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+
+    // Cleanup - handle pending pixels from loop
+.LLoopCleanup1:
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero12
+    jz          .LAlphaZero13
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+.LAlphaZero13:
+    addl        $16, %edi
+    subl        $4, %ecx
+    js          .LSmallRemaining        // Reuse code from small loop
+
+.LRemain1:
+    movdqa      (%eax, %edi), %xmm1     // Load four source pixels
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero12
+    jz          .LAlphaZero14
+    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
+.LAlphaZero14:
+    addl        $16, %edi
+    subl        $4, %ecx
+    jmp         .LSmallRemaining        // Reuse code from small loop
+
+.LAlphaNotOpaqueOrZero12:
+    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    addl        $16, %edi
+    subl        $4, %ecx
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, -16(%edx, %edi)  // Store four destination pixels
+    js          .LSmallRemaining        // Reuse code from small loop
+    jmp         .LRemain1
+
+    .cfi_endproc
+#ifndef __clang__
+    .size S32A_Opaque_BlitRow32_SSE4_asm, .-S32A_Opaque_BlitRow32_SSE4_asm
+#endif
+#endif
+
+#endif // CRBUG_399842_FIXED
diff --git a/src/opts/SkBlitRow_opts_SSE4_x64_asm.S b/src/opts/SkBlitRow_opts_SSE4_x64_asm.S
new file mode 100644
index 0000000000..9a754a635b
--- /dev/null
+++ b/src/opts/SkBlitRow_opts_SSE4_x64_asm.S
@@ -0,0 +1,472 @@
+/*
+ * Copyright 2014 The Android Open Source Project
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifdef CRBUG_399842_FIXED
+
+#if defined(__clang__) || (defined(__GNUC__) && !defined(SK_BUILD_FOR_MAC))
+
+#define EXTRACT_ALPHA(var1, var2) \
+    movdqa      %var1, %var2;           /* Clone source pixels to extract alpha */\
+    psrlw       $8, %var2;              /* Discard red and blue, leaving alpha and green */\
+    pshufhw     $0xF5, %var2, %var2;    /* Repeat alpha for scaling (high) */\
+    movdqa      %xmm6, %xmm4;           \
+    pshuflw     $0xF5, %var2, %var2;    /* Repeat alpha for scaling (low) */\
+    movdqa      %xmm5, %xmm3;           \
+    psubw       %var2, %xmm4            /* Finalize alpha calculations */
+
+#define SCALE_PIXELS \
+    psllw       $8, %xmm5;              /* Filter out red and blue components */\
+    pmulhuw     %xmm4, %xmm5;           /* Scale red and blue */\
+    psrlw       $8, %xmm3;              /* Filter out alpha and green components */\
+    pmullw      %xmm4, %xmm3            /* Scale alpha and green */
+
+
+/*
+ * void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst,
+ *                                 const SkPMColor* SK_RESTRICT src,
+ *                                 int count, U8CPU alpha)
+ *
+ * This function is divided into six blocks: initialization, blit 4-15 pixels,
+ * blit 0-3 pixels, align destination for 16+ pixel blits,
+ * blit 16+ pixels with source unaligned, blit 16+ pixels with source aligned.
+ * There are some code reuse between the blocks.
+ *
+ * The primary optimization comes from checking the source pixels' alpha value.
+ * If the alpha is zero, the pixel can be skipped entirely.
+ * If the alpha is fully opaque, the pixel can be copied directly to the destination.
+ * According to collected statistics, these two cases are the most common.
+ * The main loop(s) uses pre-loading and unrolling in an attempt to reduce the
+ * memory latency worse-case.
+ */
+
+#ifdef __clang__
+    .text
+#else
+    .section .text.sse4.2,"ax",@progbits
+    .type S32A_Opaque_BlitRow32_SSE4_asm, @function
+#endif
+    .p2align 4
+#if defined(SK_BUILD_FOR_MAC)
+    .global _S32A_Opaque_BlitRow32_SSE4_asm
+    .private_extern _S32A_Opaque_BlitRow32_SSE4_asm
+_S32A_Opaque_BlitRow32_SSE4_asm:
+#else
+    .global S32A_Opaque_BlitRow32_SSE4_asm
+    .hidden S32A_Opaque_BlitRow32_SSE4_asm
+S32A_Opaque_BlitRow32_SSE4_asm:
+#endif
+    .cfi_startproc
+    prefetcht0  (%rsi)
+    movl        %edx, %ecx              // Pixel count
+    movq        %rdi, %rdx              // Destination pointer
+    movq        %rsi, %rax              // Source pointer
+
+    // Setup SSE constants
+    movdqa      .LAlphaCheckMask(%rip), %xmm7  // 0xFF000000 mask to check alpha
+    movdqa      .LInverseAlphaCalc(%rip), %xmm6// 16-bit 256 to calculate inv. alpha
+    movdqa      .LResultMergeMask(%rip), %xmm0 // 0x00FF00FF mask (Must be in xmm0 because of pblendvb)
+
+    subl        $4, %ecx                // Check if we have only 0-3 pixels
+    js          .LReallySmall
+    cmpl        $11, %ecx               // Do we have enough pixels to run the main loop?
+    ja          .LBigBlit
+
+    // Handle small blits (4-15 pixels)
+    ////////////////////////////////////////////////////////////////////////////////
+    xorq        %rdi, %rdi              // Reset offset to zero
+
+.LSmallLoop:
+    lddqu       (%rax, %rdi), %xmm1     // Load four source pixels
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LSmallAlphaNotOpaqueOrZero
+    jz          .LSmallAlphaZero
+    movdqu      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+.LSmallAlphaZero:
+    addq        $16, %rdi
+    subl        $4, %ecx                // Check if there are four additional pixels, at least
+    jns         .LSmallLoop
+    jmp         .LSmallRemaining
+
+    // Handle mixed alphas (calculate and scale)
+    .p2align 4
+.LSmallAlphaNotOpaqueOrZero:
+    lddqu       (%rdx, %rdi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    addq        $16, %rdi
+    subl        $4, %ecx                // Check if there are four additional pixels, at least
+    pblendvb    %xmm5, %xmm3            // Mask in %xmm0, implicitly
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqu      %xmm1, -16(%rdx, %rdi)  // Store four destination pixels
+    jns         .LSmallLoop
+
+    // Handle the last 0-3 pixels (also used by the main loops)
+.LSmallRemaining:
+    cmpl        $-4, %ecx               // Check if we are done
+    je          .LSmallExit
+    sall        $2, %ecx                // Calculate offset for last pixels
+    movslq      %ecx, %rcx
+    addq        %rcx, %rdi
+
+    lddqu       (%rax, %rdi), %xmm1     // Load last four source pixels (overlapping)
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    jc          .LSmallRemainingStoreAll// If all alphas are opaque, just store (overlapping)
+    jz          .LSmallExit             // If all alphas are zero, skip the pixels completely
+
+    // Handle mixed alphas (calculate and scale)
+    lddqu       (%rdx, %rdi), %xmm5     // Load last four destination pixels (overlapping)
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+
+    psllw       $8, %xmm3               // Filter out red and blue components
+    pmulhuw     %xmm4, %xmm3            // Scale red and blue
+    movdqa      %xmm5, %xmm2
+    psrlw       $8, %xmm2               // Filter out alpha and green components
+    pmullw      %xmm4, %xmm2            // Scale alpha and green
+
+    cmpl        $-8, %ecx               // Check how many pixels should be written
+    pblendvb    %xmm3, %xmm2            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm2, %xmm1            // Add source and destination pixels together
+    jb          .LSmallPixelsLeft1
+    ja          .LSmallPixelsLeft3      // To avoid double-blending the overlapping pixels...
+    pblendw     $0xF0, %xmm1, %xmm5     // Merge only the final two pixels to the destination
+    movdqu      %xmm5, (%rdx, %rdi)     // Store last two destination pixels
+.LSmallExit:
+    ret
+
+.LSmallPixelsLeft1:
+    pblendw     $0xC0, %xmm1, %xmm5     // Merge only the final pixel to the destination
+    movdqu      %xmm5, (%rdx, %rdi)     // Store last destination pixel
+    ret
+
+.LSmallPixelsLeft3:
+    pblendw     $0xFC, %xmm1, %xmm5     // Merge only the final three pixels to the destination
+    movdqu      %xmm5, (%rdx, %rdi)     // Store last three destination pixels
+    ret
+
+.LSmallRemainingStoreAll:
+    movdqu      %xmm1, (%rdx, %rdi)     // Store last destination pixels (overwrite)
+    ret
+
+    // Handle really small blits (0-3 pixels)
+    ////////////////////////////////////////////////////////////////////////////////
+.LReallySmall:
+    addl        $4, %ecx
+    jle         .LReallySmallExit
+    pcmpeqd     %xmm1, %xmm1
+    cmpl        $2, %ecx                // Check how many pixels should be read
+    pinsrd      $0x0, (%rax), %xmm1     // Load one source pixel
+    pinsrd      $0x0, (%rdx), %xmm5     // Load one destination pixel
+    jb          .LReallySmallCalc
+    pinsrd      $0x1, 4(%rax), %xmm1    // Load second source pixel
+    pinsrd      $0x1, 4(%rdx), %xmm5    // Load second destination pixel
+    je          .LReallySmallCalc
+    pinsrd      $0x2, 8(%rax), %xmm1    // Load third source pixel
+    pinsrd      $0x2, 8(%rdx), %xmm5    // Load third destination pixel
+
+.LReallySmallCalc:
+    ptest       %xmm7, %xmm1            // Check if all alphas are opaque
+    jc          .LReallySmallStore      // If all alphas are opaque, just store
+
+    // Handle mixed alphas (calculate and scale)
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+
+    pand        %xmm0, %xmm5            // Filter out red and blue components
+    pmullw      %xmm4, %xmm5            // Scale red and blue
+    psrlw       $8, %xmm3               // Filter out alpha and green components
+    pmullw      %xmm4, %xmm3            // Scale alpha and green
+
+    psrlw       $8, %xmm5               // Combine results
+    pblendvb    %xmm5, %xmm3            // Mask in %xmm0, implicitly
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+
+.LReallySmallStore:
+    cmpl        $2, %ecx                // Check how many pixels should be written
+    pextrd      $0x0, %xmm1, (%rdx)     // Store one destination pixel
+    jb          .LReallySmallExit
+    pextrd      $0x1, %xmm1, 4(%rdx)    // Store second destination pixel
+    je          .LReallySmallExit
+    pextrd      $0x2, %xmm1, 8(%rdx)    // Store third destination pixel
+.LReallySmallExit:
+    ret
+
+    // Handle bigger blit operations (16+ pixels)
+    ////////////////////////////////////////////////////////////////////////////////
+    .p2align 4
+.LBigBlit:
+    // Align destination?
+    testl       $0xF, %edx
+    lddqu       (%rax), %xmm1           // Pre-load four source pixels
+    jz          .LAligned
+
+    movq        %rdx, %rdi              // Calculate alignment of destination pointer
+    negq        %rdi
+    andl        $0xF, %edi
+
+    // Handle 1-3 pixels to align destination
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    jz          .LAlignDone             // If all alphas are zero, just skip
+    lddqu       (%rdx), %xmm5           // Load four destination pixels
+    jc          .LAlignStore            // If all alphas are opaque, just store
+
+    // Handle mixed alphas (calculate and scale)
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+
+    psllw       $8, %xmm3               // Filter out red and blue components
+    pmulhuw     %xmm4, %xmm3            // Scale red and blue
+    movdqa      %xmm5, %xmm2
+    psrlw       $8, %xmm2               // Filter out alpha and green components
+    pmullw      %xmm4, %xmm2            // Scale alpha and green
+
+    pblendvb    %xmm3, %xmm2            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm2, %xmm1            // Add source and destination pixels together
+
+.LAlignStore:
+    cmpl        $8, %edi                // Check how many pixels should be written
+    jb          .LAlignPixelsLeft1
+    ja          .LAlignPixelsLeft3
+    pblendw     $0x0F, %xmm1, %xmm5     // Blend two pixels
+    jmp .LAlignStorePixels
+
+.LAlignPixelsLeft1:
+    pblendw     $0x03, %xmm1, %xmm5     // Blend one pixel
+    jmp .LAlignStorePixels
+
+.LAlignPixelsLeft3:
+    pblendw     $0x3F, %xmm1, %xmm5     // Blend three pixels
+
+.LAlignStorePixels:
+    movdqu      %xmm5, (%rdx)           // Store destination pixels
+
+.LAlignDone:
+    addq        %rdi, %rax              // Adjust pointers and pixel count
+    addq        %rdi, %rdx
+    shrq        $2, %rdi
+    lddqu       (%rax), %xmm1           // Pre-load new source pixels (after alignment)
+    subl        %edi, %ecx
+
+.LAligned:                              // Destination is guaranteed to be 16 byte aligned
+    xorq        %rdi, %rdi              // Reset offset to zero
+    subl        $8, %ecx                // Decrease counter (Reserve four pixels for the cleanup)
+    testl       $0xF, %eax              // Check alignment of source pointer
+    jz          .LAlignedLoop
+
+    // Source not aligned to destination
+    ////////////////////////////////////////////////////////////////////////////////
+    .p2align 4
+.LUnalignedLoop:                        // Main loop for unaligned, handles eight pixels per iteration
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero00
+    lddqu       16(%rax, %rdi), %xmm2   // Pre-load four source pixels
+    jz          .LAlphaZero00
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+
+.LAlphaZero00:
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero01
+    lddqu       32(%rax, %rdi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero01
+    movdqa      %xmm2, 16(%rdx, %rdi)   // Store four destination pixels
+
+.LAlphaZero01:
+    addq        $32, %rdi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LUnalignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup0
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero00:
+    movdqa      (%rdx, %rdi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    lddqu       16(%rax, %rdi), %xmm2   // Pre-load four source pixels
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+
+    // Handle next four pixels
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero01
+    lddqu       32(%rax, %rdi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero02
+    movdqa      %xmm2, 16(%rdx, %rdi)   // Store four destination pixels
+.LAlphaZero02:
+    addq        $32, %rdi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LUnalignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup0
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero01:
+    movdqa      16(%rdx, %rdi), %xmm5   // Load four destination pixels
+    EXTRACT_ALPHA(xmm2, xmm1)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    lddqu       32(%rax, %rdi), %xmm1   // Pre-load four source pixels
+    addq        $32, %rdi
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm2            // Add source and destination pixels together
+    subl        $8, %ecx
+    movdqa      %xmm2, -16(%rdx, %rdi)  // Store four destination pixels
+    jae         .LUnalignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+
+    // Cleanup - handle pending pixels from loop
+.LLoopCleanup0:
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero02
+    jz          .LAlphaZero03
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+.LAlphaZero03:
+    addq        $16, %rdi
+    subl        $4, %ecx
+    js          .LSmallRemaining        // Reuse code from small loop
+
+.LRemain0:
+    lddqu       (%rax, %rdi), %xmm1     // Load four source pixels
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero02
+    jz          .LAlphaZero04
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+.LAlphaZero04:
+    addq        $16, %rdi
+    subl        $4, %ecx
+    jmp         .LSmallRemaining        // Reuse code from small loop
+
+.LAlphaNotOpaqueOrZero02:
+    movdqa      (%rdx, %rdi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    addq        $16, %rdi
+    subl        $4, %ecx
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, -16(%rdx, %rdi)  // Store four destination pixels
+    js          .LSmallRemaining        // Reuse code from small loop
+    jmp         .LRemain0
+
+    // Source aligned to destination
+    ////////////////////////////////////////////////////////////////////////////////
+    .p2align 4
+.LAlignedLoop:                          // Main loop for aligned, handles eight pixels per iteration
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero10
+    movdqa      16(%rax, %rdi), %xmm2   // Pre-load four source pixels
+    jz          .LAlphaZero10
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+
+.LAlphaZero10:
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero11
+    movdqa      32(%rax, %rdi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero11
+    movdqa      %xmm2, 16(%rdx, %rdi)   // Store four destination pixels
+
+.LAlphaZero11:
+    addq        $32, %rdi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LAlignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup1
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero10:
+    movdqa      (%rdx, %rdi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    movdqa      16(%rax, %rdi), %xmm2   // Pre-load four source pixels
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+
+    // Handle next four pixels
+    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero11
+    movdqa      32(%rax, %rdi), %xmm1   // Pre-load four source pixels
+    jz          .LAlphaZero12
+    movdqa      %xmm2, 16(%rdx, %rdi)   // Store four destination pixels
+.LAlphaZero12:
+    addq        $32, %rdi               // Adjust offset and pixel count
+    subl        $8, %ecx
+    jae         .LAlignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+    jmp         .LLoopCleanup1
+
+    .p2align 4
+.LAlphaNotOpaqueOrZero11:
+    movdqa      16(%rdx, %rdi), %xmm5   // Load four destination pixels
+    EXTRACT_ALPHA(xmm2, xmm1)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+    movdqa      32(%rax, %rdi), %xmm1   // Pre-load four source pixels
+
+    addq        $32, %rdi
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm2            // Add source and destination pixels together
+    subl        $8, %ecx
+    movdqa      %xmm2, -16(%rdx, %rdi)  // Store four destination pixels
+    jae         .LAlignedLoop
+    addl        $8, %ecx                // Adjust pixel count
+
+    // Cleanup - handle four pending pixels from loop
+.LLoopCleanup1:
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero12
+    jz          .LAlphaZero13
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+.LAlphaZero13:
+    addq        $16, %rdi
+    subl        $4, %ecx
+    js          .LSmallRemaining        // Reuse code from small loop
+
+.LRemain1:
+    movdqa      (%rax, %rdi), %xmm1     // Pre-load four source pixels
+    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
+    ja          .LAlphaNotOpaqueOrZero12
+    jz          .LAlphaZero14
+    movdqa      %xmm1, (%rdx, %rdi)     // Store four destination pixels
+.LAlphaZero14:
+    addq        $16, %rdi
+    subl        $4, %ecx
+    jmp         .LSmallRemaining        // Reuse code from small loop
+
+.LAlphaNotOpaqueOrZero12:
+    movdqa      (%rdx, %rdi), %xmm5     // Load four destination pixels
+    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
+    SCALE_PIXELS                        // Scale pixels using alpha
+
+    addq        $16, %rdi
+    subl        $4, %ecx
+    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
+    paddb       %xmm3, %xmm1            // Add source and destination pixels together
+    movdqa      %xmm1, -16(%rdx, %rdi)  // Store four destination pixels
+    js          .LSmallRemaining        // Reuse code from small loop
+    jmp         .LRemain1
+
+    .cfi_endproc
+#ifndef __clang__
+    .size S32A_Opaque_BlitRow32_SSE4_asm, .-S32A_Opaque_BlitRow32_SSE4_asm
+#endif
+
+    // Constants for SSE code
+#ifndef __clang__
+    .section .rodata
+#endif
+    .p2align 4
+.LAlphaCheckMask:
+    .long   0xFF000000, 0xFF000000, 0xFF000000, 0xFF000000
+.LInverseAlphaCalc:
+    .word   256, 256, 256, 256, 256, 256, 256, 256
+.LResultMergeMask:
+    .long   0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF
+#endif
+
+#endif // CRBUG_399842_FIXED
diff --git a/src/opts/SkColor_opts_SSE2.h b/src/opts/SkColor_opts_SSE2.h
index feb1d98f8d..970abb859b 100644
--- a/src/opts/SkColor_opts_SSE2.h
+++ b/src/opts/SkColor_opts_SSE2.h
@@ -206,14 +206,7 @@ static inline __m128i SkPixel32ToPixel16_ToU16_SSE2(const __m128i& src_pixel1,
     return d_pixel;
 }
 
-// Portable version is SkPMSrcOver in SkColorPriv.h.
-static inline __m128i SkPMSrcOver_SSE2(const __m128i& src, const __m128i& dst) {
-    return _mm_add_epi32(src,
-                         SkAlphaMulQ_SSE2(dst, _mm_sub_epi32(_mm_set1_epi32(256),
-                                                             SkGetPackedA32_SSE2(src))));
-}
-
-// Portable version is SkBlendARGB32 in SkColorPriv.h.
+// Portable version SkBlendARGB32 is in SkColorPriv.h.
 static inline __m128i SkBlendARGB32_SSE2(const __m128i& src, const __m128i& dst,
                                          const __m128i& aa) {
     __m128i src_scale = SkAlpha255To256_SSE2(aa);
diff --git a/src/opts/opts_check_x86.cpp b/src/opts/opts_check_x86.cpp
index 84a4913021..71107d8756 100644
--- a/src/opts/opts_check_x86.cpp
+++ b/src/opts/opts_check_x86.cpp
@@ -227,17 +227,21 @@ static SkBlitRow::Proc32 platform_32_procs_SSE2[] = {
     S32A_Blend_BlitRow32_SSE2,          // S32A_Blend,
 };
 
+#if defined(SK_ATT_ASM_SUPPORTED)
 static SkBlitRow::Proc32 platform_32_procs_SSE4[] = {
     NULL,                               // S32_Opaque,
     S32_Blend_BlitRow32_SSE2,           // S32_Blend,
-    S32A_Opaque_BlitRow32_SSE4,         // S32A_Opaque
+    S32A_Opaque_BlitRow32_SSE4_asm,     // S32A_Opaque
     S32A_Blend_BlitRow32_SSE2,          // S32A_Blend,
 };
+#endif
 
 SkBlitRow::Proc32 SkBlitRow::PlatformProcs32(unsigned flags) {
+#if defined(SK_ATT_ASM_SUPPORTED)
     if (supports_simd(SK_CPU_SSE_LEVEL_SSE41)) {
         return platform_32_procs_SSE4[flags];
     } else
+#endif
     if (supports_simd(SK_CPU_SSE_LEVEL_SSE2)) {
         return platform_32_procs_SSE2[flags];
     } else {