pavgb with memory op requires alignment. This CL disables conversions that use pavgb, and resolves scale by 3/8 unittest for checking alignment works. The 3/8 code used a pavgb with a memory operand. tests are added for scaling and allow unaligning on purpose.

BUG=365
TESTED=local change to force unaligned memory fails on some conversions and scaling code.
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/29699004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1114 16f28f9a-4ce2-e073-06de-1de4eb20be90

11 files changed, 27 insertions, 156 deletions

diff --git a/README.chromium b/README.chromium
index c4d4a4f..c1647af 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1112
+Version: 1114
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index f5abf1e..186dcbd 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -87,7 +87,6 @@ extern "C" {
 #define HAS_ARGBSEPIAROW_SSSE3
 #define HAS_ARGBSHADEROW_SSE2
 #define HAS_ARGBSUBTRACTROW_SSE2
-#define HAS_ARGBTOUVROW_SSSE3
 #define HAS_ARGBUNATTENUATEROW_SSE2
 #define HAS_COMPUTECUMULATIVESUMROW_SSE2
 #define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
@@ -101,7 +100,8 @@ extern "C" {
 #define HAS_SOBELYROW_SSE2
 
 // Conversions:
-#define HAS_ABGRTOUVROW_SSSE3
+//#define HAS_ARGBTOUVROW_SSSE3
+//#define HAS_ABGRTOUVROW_SSSE3
 #define HAS_ABGRTOYROW_SSSE3
 #define HAS_ARGB1555TOARGBROW_SSE2
 #define HAS_ARGB4444TOARGBROW_SSE2
@@ -116,10 +116,10 @@ extern "C" {
 #define HAS_ARGBTORGB565ROW_SSE2
 #define HAS_ARGBTOUV422ROW_SSSE3
 #define HAS_ARGBTOUV444ROW_SSSE3
-#define HAS_ARGBTOUVJROW_SSSE3
+//#define HAS_ARGBTOUVJROW_SSSE3
 #define HAS_ARGBTOYJROW_SSSE3
 #define HAS_ARGBTOYROW_SSSE3
-#define HAS_BGRATOUVROW_SSSE3
+//#define HAS_BGRATOUVROW_SSSE3
 #define HAS_BGRATOYROW_SSSE3
 #define HAS_COPYROW_ERMS
 #define HAS_COPYROW_SSE2
@@ -153,7 +153,7 @@ extern "C" {
 #define HAS_RGB24TOARGBROW_SSSE3
 #define HAS_RGB24TOYROW_SSSE3
 #define HAS_RGB565TOARGBROW_SSE2
-#define HAS_RGBATOUVROW_SSSE3
+//#define HAS_RGBATOUVROW_SSSE3
 #define HAS_RGBATOYROW_SSSE3
 #define HAS_SETROW_X86
 #define HAS_SPLITUVROW_SSE2
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index 59aed1a..fb0e396 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1112
+#define LIBYUV_VERSION 1114
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/source/convert.cc b/source/convert.cc
index 19ad659..da3d4c6 100644
--- a/source/convert.cc
+++ b/source/convert.cc
@@ -735,7 +735,7 @@ int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
     src_bgra = src_bgra + (height - 1) * src_stride_bgra;
     src_stride_bgra = -src_stride_bgra;
   }
-#if defined(HAS_BGRATOYROW_SSSE3)
+#if defined(HAS_BGRATOYROW_SSSE3) && defined(HAS_BGRATOUVROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
     BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
     BGRAToYRow = BGRAToYRow_Any_SSSE3;
@@ -800,7 +800,7 @@ int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
     src_abgr = src_abgr + (height - 1) * src_stride_abgr;
     src_stride_abgr = -src_stride_abgr;
   }
-#if defined(HAS_ABGRTOYROW_SSSE3)
+#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
     ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
     ABGRToYRow = ABGRToYRow_Any_SSSE3;
@@ -865,7 +865,7 @@ int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
     src_rgba = src_rgba + (height - 1) * src_stride_rgba;
     src_stride_rgba = -src_stride_rgba;
   }
-#if defined(HAS_RGBATOYROW_SSSE3)
+#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
     RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
     RGBAToYRow = RGBAToYRow_Any_SSSE3;
diff --git a/source/row_any.cc b/source/row_any.cc
index 0d58f5d..9d8a5e5 100644
--- a/source/row_any.cc
+++ b/source/row_any.cc
@@ -350,11 +350,12 @@ UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
 #endif
 #ifdef HAS_ARGBTOUVROW_SSSE3
 UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, ARGBToUVRow_C, 4, 15)
-UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, ARGBToUVJRow_C,
-      4, 15)
+UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, ARGBToUVJRow_C, 4, 15)
 UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, BGRAToUVRow_C, 4, 15)
 UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, ABGRToUVRow_C, 4, 15)
 UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, RGBAToUVRow_C, 4, 15)
+#endif
+#ifdef HAS_YUY2TOUVROW_SSE2
 UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, YUY2ToUVRow_C, 2, 15)
 UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, UYVYToUVRow_C, 2, 15)
 #endif
@@ -417,9 +418,11 @@ UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
 UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
          UYVYToUV422Row_C, 2, 31, 1)
 #endif
-#ifdef HAS_ARGBTOUVROW_SSSE3
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
 UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3,
          ARGBToUV422Row_C, 4, 15, 1)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_SSE2
 UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2,
          YUY2ToUV422Row_C, 2, 15, 1)
 UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2,
diff --git a/source/row_posix.cc b/source/row_posix.cc
index 94b24a6..15a8321 100644
--- a/source/row_posix.cc
+++ b/source/row_posix.cc
@@ -743,43 +743,6 @@ void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
 #endif
   );
 }
-
-void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(pix)        // %2
-  : "m"(kARGBToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
 #endif  // HAS_ARGBTOYROW_SSSE3
 
 #ifdef HAS_ARGBTOYJROW_SSSE3
@@ -1027,71 +990,6 @@ void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
   );
 }
 
-void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u,
-                                    uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),  // %0
-    "m"(kARGBToV),  // %1
-    "m"(kAddUV128)  // %2
-  );
-  asm volatile (
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm6                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm2                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "packsswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm3,%%xmm0                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm2                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "packsswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),        // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6"
-#endif
-  );
-}
-
 void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
                           uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
@@ -1330,43 +1228,6 @@ void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
   );
 }
 
-void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_rgba),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(pix)        // %2
-  : "m"(kRGBAToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
 void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
                        uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
diff --git a/source/scale_posix.cc b/source/scale_posix.cc
index 6320e50..92e3354 100644
--- a/source/scale_posix.cc
+++ b/source/scale_posix.cc
@@ -526,8 +526,9 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb   (%0,%3,1),%%xmm0
+    MEMOPREG(movdqu,0x00,0,3,1,xmm1)           //  movdqu  (%0,%3,1),%%xmm1
     "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "pshufb    %%xmm2,%%xmm1                   \n"
     "movdqa    %%xmm0,%%xmm6                   \n"
diff --git a/source/scale_win.cc b/source/scale_win.cc
index 091587e..8370ef4 100644
--- a/source/scale_win.cc
+++ b/source/scale_win.cc
@@ -589,8 +589,9 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
     align      4
   xloop:
     movdqu     xmm0, [eax]           // average 2 rows into xmm0
-    pavgb      xmm0, [eax + esi]
+    movdqu     xmm1, [eax + esi]
     lea        eax, [eax + 16]
+    pavgb      xmm0, xmm1
 
     movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
     pshufb     xmm1, xmm2
diff --git a/unit_test/scale_argb_test.cc b/unit_test/scale_argb_test.cc
index 6a2bc79..bbeb4f8 100644
--- a/unit_test/scale_argb_test.cc
+++ b/unit_test/scale_argb_test.cc
@@ -223,6 +223,7 @@ TEST_FACTOR(2, 1 / 2, 1 / 2)
 TEST_FACTOR(4, 1 / 4, 1 / 4)
 TEST_FACTOR(8, 1 / 8, 1 / 8)
 TEST_FACTOR(3by4, 3 / 4, 3 / 4)
+TEST_FACTOR(3by8, 3 / 8, 3 / 8)
 #undef TEST_FACTOR1
 #undef TEST_FACTOR
 
diff --git a/unit_test/scale_test.cc b/unit_test/scale_test.cc
index 00f0707..5d08365 100644
--- a/unit_test/scale_test.cc
+++ b/unit_test/scale_test.cc
@@ -288,6 +288,7 @@ TEST_FACTOR(2, 1 / 2, 1 / 2)
 TEST_FACTOR(4, 1 / 4, 1 / 4)
 TEST_FACTOR(8, 1 / 8, 1 / 8)
 TEST_FACTOR(3by4, 3 / 4, 3 / 4)
+TEST_FACTOR(3by8, 3 / 8, 3 / 8)
 #undef TEST_FACTOR1
 #undef TEST_FACTOR
 
diff --git a/unit_test/unit_test.h b/unit_test/unit_test.h
index cfce548..0151796 100644
--- a/unit_test/unit_test.h
+++ b/unit_test/unit_test.h
@@ -26,11 +26,14 @@ static __inline int Abs(int v) {
   return v >= 0 ? v : -v;
 }
 
+#define OFFBY 0
+
 #define align_buffer_page_end(var, size)                                       \
   uint8* var;                                                                  \
   uint8* var##_mem;                                                            \
-  var##_mem = reinterpret_cast<uint8*>(malloc(((size) + 4095) & ~4095));       \
-  var = var##_mem + (-(size) & 4095);
+  var##_mem = reinterpret_cast<uint8*>(malloc((((size) + 4095) & ~4095) +      \
+      OFFBY));                                                                 \
+  var = var##_mem + (-(size) & 4095) + OFFBY;
 
 #define free_aligned_buffer_page_end(var) \
   free(var##_mem);  \