libvpx: Pull from upstream

source/libvpx/third_party/libyuv/source/compare_posix.cc

Index: source/libvpx/third_party/libyuv/source/compare_posix.cc
diff --git a/source/libvpx/third_party/libyuv/source/compare_posix.cc b/source/libvpx/third_party/libyuv/source/compare_posix.cc
index ac361190e8867da6deea1c3c0d991138ca754af6..247cb33bbaab6f0b6b6eab7bb1eb8e3f554ce353 100644
--- a/source/libvpx/third_party/libyuv/source/compare_posix.cc
+++ b/source/libvpx/third_party/libyuv/source/compare_posix.cc
@@ -25,11 +25,10 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     "pxor      %%xmm5,%%xmm5                   \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
     "lea       " MEMLEA(0x10, 0) ",%0          \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
     "lea       " MEMLEA(0x10, 1) ",%1          \n"
-    "sub       $0x10,%2                        \n"
     "movdqa    %%xmm1,%%xmm3                   \n"
     "psubusb   %%xmm2,%%xmm1                   \n"
     "psubusb   %%xmm3,%%xmm2                   \n"
@@ -41,6 +40,7 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     "pmaddwd   %%xmm2,%%xmm2                   \n"
     "paddd     %%xmm1,%%xmm0                   \n"
     "paddd     %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
 
     "pshufd    $0xee,%%xmm0,%%xmm1             \n"
@@ -53,11 +53,7 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     "+r"(src_b),      // %1
     "+r"(count),      // %2
     "=g"(sse)         // %3
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );  // NOLINT
   return sse;
 }
@@ -124,13 +120,13 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     "pmulld    %%xmm5,%%xmm1                   \n"
     "paddd     %%xmm4,%%xmm3                   \n"
     "paddd     %%xmm2,%%xmm1                   \n"
-    "sub       $0x10,%1                        \n"
     "paddd     %%xmm3,%%xmm1                   \n"
     "pshufd    $0xe,%%xmm1,%%xmm2              \n"
     "paddd     %%xmm2,%%xmm1                   \n"
     "pshufd    $0x1,%%xmm1,%%xmm2              \n"
     "paddd     %%xmm2,%%xmm1                   \n"
     "paddd     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%1                        \n"
     "jg        1b                              \n"
     "movd      %%xmm0,%3                       \n"
   : "+r"(src),        // %0
@@ -143,9 +139,7 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     "m"(kHashMul2),   // %7
     "m"(kHashMul3)    // %8
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
   );  // NOLINT
   return hash;
 }