Add SSE4 optimization of S32A_Opaque_Blitrow

src/opts/opts_check_x86.cpp

Index: src/opts/opts_check_x86.cpp
diff --git a/src/opts/opts_check_x86.cpp b/src/opts/opts_check_x86.cpp
index e3b6b470786f9825f04857bcf9b535f1a194a052..272d75944a915bb066844aa633a3cdb78f1a7d57 100644
--- a/src/opts/opts_check_x86.cpp
+++ b/src/opts/opts_check_x86.cpp
@@ -12,6 +12,7 @@
 #include "SkBlitRect_opts_SSE2.h"
 #include "SkBlitRow.h"
 #include "SkBlitRow_opts_SSE2.h"
+#include "SkBlitRow_opts_SSE4.h"
 #include "SkBlurImage_opts_SSE2.h"
 #include "SkMorphology_opts.h"
 #include "SkMorphology_opts_SSE2.h"
@@ -206,16 +207,30 @@ SkBlitRow::Proc SkBlitRow::PlatformProcs565(unsigned flags) {
     }
 }
 
-static SkBlitRow::Proc32 platform_32_procs[] = {
+static SkBlitRow::Proc32 platform_32_procs_SSE2[] = {
     NULL,                               // S32_Opaque,
     S32_Blend_BlitRow32_SSE2,           // S32_Blend,
     S32A_Opaque_BlitRow32_SSE2,         // S32A_Opaque
     S32A_Blend_BlitRow32_SSE2,          // S32A_Blend,
 };
 
+#if !defined(_MSC_VER)
+static SkBlitRow::Proc32 platform_32_procs_SSE4[] = {
+    NULL,                               // S32_Opaque,
+    S32_Blend_BlitRow32_SSE2,           // S32_Blend,
+    S32A_Opaque_BlitRow32_SSE4_asm,     // S32A_Opaque
+    S32A_Blend_BlitRow32_SSE2,          // S32A_Blend,
+};
+#endif
+
 SkBlitRow::Proc32 SkBlitRow::PlatformProcs32(unsigned flags) {
+#if !defined(_MSC_VER)
+    if (supports_simd(SK_CPU_SSE_LEVEL_SSE42)) {
+        return platform_32_procs_SSE4[flags];
+    } else
+#endif
     if (supports_simd(SK_CPU_SSE_LEVEL_SSE2)) {
-        return platform_32_procs[flags];
+        return platform_32_procs_SSE2[flags];
     } else {
         return NULL;
     }