aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/opts/SkBlitRow_opts_SSE4_asm.S
diff options
context:
space:
mode:
authorGravatar stephana <stephana@google.com>2015-02-02 09:52:43 -0800
committerGravatar Commit bot <commit-bot@chromium.org>2015-02-02 09:52:43 -0800
commit4988891a1173cd405bf1c1dd3a3668c451f45e4c (patch)
tree0fd4b535b7e9b57fb0b6248f69b04339bc5c7b69 /src/opts/SkBlitRow_opts_SSE4_asm.S
parentdb204e301b1320af242e1be5e477cd9453b126a6 (diff)
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
Diffstat (limited to 'src/opts/SkBlitRow_opts_SSE4_asm.S')
-rw-r--r--src/opts/SkBlitRow_opts_SSE4_asm.S475
1 files changed, 475 insertions, 0 deletions
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