Replace SSE optimization of Color32A_D565

src/opts/SkBlitRow_opts_SSE4.cpp

Index: src/opts/SkBlitRow_opts_SSE4.cpp
diff --git a/src/opts/SkBlitRow_opts_SSE4.cpp b/src/opts/SkBlitRow_opts_SSE4.cpp
index f4273d27b4b0bfaf81d1df0c5d86f8f25f5a8164..82fc70b4d349667ccaa9b137e5aea1ad120654de 100644
--- a/src/opts/SkBlitRow_opts_SSE4.cpp
+++ b/src/opts/SkBlitRow_opts_SSE4.cpp
@@ -66,11 +66,6 @@ void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst,
     }
 }
 
-static inline uint16_t Color32A_D565_1x(uint16_t dst, unsigned scale, uint32_t src_expand) {
-    uint32_t dst_expand = SkExpand_rgb_16(dst) * scale;
-    return SkCompact_rgb_16((src_expand + dst_expand) >> 5);
-}
-
 void Color32A_D565_SSE4(uint16_t dst[], SkPMColor src, int count, int x, int y) {
     SkASSERT(count > 0);
 
@@ -90,7 +85,7 @@ void Color32A_D565_SSE4(uint16_t dst[], SkPMColor src, int count, int x, int y)
 
         // Align dst to an even 16 byte address (0-7 pixels)
         while (((((size_t)dst) & 0x0F) != 0) && (count > 0)) {
-            *dst = Color32A_D565_1x(*dst, scale, src_expand);
+            *dst = SkBlend32A_D565(*dst, scale, src_expand);
             dst += 1;
             count--;
         }
@@ -132,7 +127,7 @@ void Color32A_D565_SSE4(uint16_t dst[], SkPMColor src, int count, int x, int y)
 
     // Small loop to handle remaining pixels.
     while (count > 0) {
-        *dst = Color32A_D565_1x(*dst, scale, src_expand);
+        *dst = SkBlend32A_D565(*dst, scale, src_expand);
         dst += 1;
         count--;
     }