Minor sk_memset{16|32}_SSE2 optimization.

src/opts/SkUtils_opts_SSE2.cpp

Index: src/opts/SkUtils_opts_SSE2.cpp
diff --git a/src/opts/SkUtils_opts_SSE2.cpp b/src/opts/SkUtils_opts_SSE2.cpp
index 08e4f669483b0a6c66931cdd916c781d45bec7a9..e22044d39d36821726cb95ff3e1605e8976ec29c 100644
--- a/src/opts/SkUtils_opts_SSE2.cpp
+++ b/src/opts/SkUtils_opts_SSE2.cpp
@@ -25,10 +25,11 @@ void sk_memset16_SSE2(uint16_t *dst, uint16_t value, int count)
         __m128i *d = reinterpret_cast<__m128i*>(dst);
         __m128i value_wide = _mm_set1_epi16(value);
         while (count >= 32) {
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
+            _mm_store_si128(d    , value_wide);
+            _mm_store_si128(d + 1, value_wide);
+            _mm_store_si128(d + 2, value_wide);
+            _mm_store_si128(d + 3, value_wide);
+            d += 4;
             count -= 32;
         }
         dst = reinterpret_cast<uint16_t*>(d);
@@ -54,10 +55,11 @@ void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count)
         __m128i *d = reinterpret_cast<__m128i*>(dst);
         __m128i value_wide = _mm_set1_epi32(value);
         while (count >= 16) {
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
+            _mm_store_si128(d    , value_wide);
+            _mm_store_si128(d + 1, value_wide);
+            _mm_store_si128(d + 2, value_wide);
+            _mm_store_si128(d + 3, value_wide);
+            d += 4;
             count -= 16;
         }
         dst = reinterpret_cast<uint32_t*>(d);