Fix memcpy32_sse2_unalign.

bench/MemcpyBench.cpp

 /*
  * Copyright 2014 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkBenchmark.h"
 #include "SkRandom.h"
 #include "SkTemplates.h"
@@ skipping 72 matching lines @@
 // Align dst to 16 bytes, then use aligned stores.  src isn't algined, so use unaligned loads.
 static void memcpy32_sse2_align(uint32_t* dst, const uint32_t* src, int count) {
     if (count >= 16) {
         while (uintptr_t(dst) & 0xF) {
             *dst++ = *src++;
             count--;
         }
 
         __m128i* dst128 = reinterpret_cast<__m128i*>(dst);
         const __m128i* src128 = reinterpret_cast<const __m128i*>(src);
+        dst += 16 * (count / 16);
+        src += 16 * (count / 16);
         while (count >= 16) {
             __m128i a = _mm_loadu_si128(src128++);
             __m128i b = _mm_loadu_si128(src128++);
             __m128i c = _mm_loadu_si128(src128++);
             __m128i d = _mm_loadu_si128(src128++);
 
             _mm_store_si128(dst128++, a);
             _mm_store_si128(dst128++, b);
             _mm_store_si128(dst128++, c);
             _mm_store_si128(dst128++, d);
 
             count -= 16;
         }
-
-        dst = reinterpret_cast<uint32_t*>(dst128);
-        src = reinterpret_cast<const uint32_t*>(src128);
     }
 
     while (count --> 0) {
         *dst++ = *src++;
     }
 }
 BENCH(memcpy32_sse2_align, 10)
 BENCH(memcpy32_sse2_align, 100)
 BENCH(memcpy32_sse2_align, 1000)
 BENCH(memcpy32_sse2_align, 10000)
 BENCH(memcpy32_sse2_align, 100000)
 
 // Leave both dst and src unaliged, and so use unaligned stores for dst and unaligned loads for src.
 static void memcpy32_sse2_unalign(uint32_t* dst, const uint32_t* src, int count) {
     __m128i* dst128 = reinterpret_cast<__m128i*>(dst);
     const __m128i* src128 = reinterpret_cast<const __m128i*>(src);
+    dst += 16 * (count / 16);
+    src += 16 * (count / 16);
     while (count >= 16) {
         __m128i a = _mm_loadu_si128(src128++);
         __m128i b = _mm_loadu_si128(src128++);
         __m128i c = _mm_loadu_si128(src128++);
         __m128i d = _mm_loadu_si128(src128++);
 
         _mm_storeu_si128(dst128++, a);
         _mm_storeu_si128(dst128++, b);
         _mm_storeu_si128(dst128++, c);
         _mm_storeu_si128(dst128++, d);
 
         count -= 16;
     }
 
-    dst = reinterpret_cast<uint32_t*>(dst128);
-    src = reinterpret_cast<const uint32_t*>(src128);
     while (count --> 0) {
         *dst++ = *src++;
     }
 }
-// skia:2589: Crashing on ChromeOS Alex bot.  TODO(mtklein): why?
-//BENCH(memcpy32_sse2_unalign, 10)
+BENCH(memcpy32_sse2_unalign, 10)
 BENCH(memcpy32_sse2_unalign, 100)
 BENCH(memcpy32_sse2_unalign, 1000)
 BENCH(memcpy32_sse2_unalign, 10000)
 BENCH(memcpy32_sse2_unalign, 100000)
 
 #endif // SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
 
 #undef BENCH