libyuv: update to b8ddb5a2

Fixes color issue when scaling without breaking mingw.

BUG=https://bugs.chromium.org/p/libyuv/issues/detail?id=605
BUG=https://bugs.chromium.org/p/webm/issues/detail?id=1252

Change-Id: I3920c5664def7ae7a23f60fb160d26d23bc86a27

1 files changed, 75 insertions, 82 deletions

diff --git a/third_party/libyuv/source/cpu_id.cc b/third_party/libyuv/source/cpu_id.cc
index 8a10b0083..84927ebc3 100644
--- a/third_party/libyuv/source/cpu_id.cc
+++ b/third_party/libyuv/source/cpu_id.cc
@@ -10,12 +10,12 @@
 
 #include "libyuv/cpu_id.h"
 
-#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+#if defined(_MSC_VER)
 #include <intrin.h>  // For __cpuidex()
 #endif
 #if !defined(__pnacl__) && !defined(__CLR_VER) && \
     !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \
-    defined(_MSC_VER) && !defined(__clang__) && (_MSC_FULL_VER >= 160040219)
+    defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
 #include <immintrin.h>  // For _xgetbv()
 #endif
 
@@ -36,7 +36,8 @@ extern "C" {
 
 // For functions that use the stack and have runtime checks for overflow,
 // use SAFEBUFFERS to avoid additional check.
-#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) && \
+    !defined(__clang__)
 #define SAFEBUFFERS __declspec(safebuffers)
 #else
 #define SAFEBUFFERS
@@ -48,9 +49,9 @@ extern "C" {
     !defined(__pnacl__) && !defined(__CLR_VER)
 LIBYUV_API
 void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
-#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+#if defined(_MSC_VER)
 // Visual C version uses intrinsic or inline x86 assembly.
-#if (_MSC_FULL_VER >= 160040219)
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
   __cpuidex((int*)(cpu_info), info_eax, info_ecx);
 #elif defined(_M_IX86)
   __asm {
@@ -63,7 +64,7 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
     mov        [edi + 8], ecx
     mov        [edi + 12], edx
   }
-#else
+#else  // Visual C but not x86
   if (info_ecx == 0) {
     __cpuid((int*)(cpu_info), info_eax);
   } else {
@@ -71,9 +72,9 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
   }
 #endif
 // GCC version uses inline x86 assembly.
-#else  // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+#else  // defined(_MSC_VER)
   uint32 info_ebx, info_edx;
-  asm volatile (  // NOLINT
+  asm volatile (
 #if defined( __i386__) && defined(__PIC__)
     // Preserve ebx for fpic 32 bit.
     "mov %%ebx, %%edi                          \n"
@@ -89,7 +90,7 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
   cpu_info[1] = info_ebx;
   cpu_info[2] = info_ecx;
   cpu_info[3] = info_edx;
-#endif  // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+#endif  // defined(_MSC_VER)
 }
 #else  // (defined(_M_IX86) || defined(_M_X64) ...
 LIBYUV_API
@@ -98,28 +99,37 @@ void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
 }
 #endif
 
-// TODO(fbarchard): Enable xgetbv when validator supports it.
+// For VS2010 and earlier emit can be used:
+//   _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+//  __asm {
+//    xor        ecx, ecx    // xcr 0
+//    xgetbv
+//    mov        xcr0, eax
+//  }
+// For VS2013 and earlier 32 bit, the _xgetbv(0) optimizer produces bad code.
+// https://code.google.com/p/libyuv/issues/detail?id=529
+#if defined(_M_IX86) && (_MSC_VER < 1900)
+#pragma optimize("g", off)
+#endif
 #if (defined(_M_IX86) || defined(_M_X64) || \
     defined(__i386__) || defined(__x86_64__)) && \
     !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__)
 #define HAS_XGETBV
 // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
-int TestOsSaveYmm() {
+int GetXCR0() {
   uint32 xcr0 = 0u;
-#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
   xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
-#elif defined(_M_IX86) && defined(_MSC_VER) && !defined(__clang__)
-  __asm {
-    xor        ecx, ecx    // xcr 0
-    _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
-    mov        xcr0, eax
-  }
 #elif defined(__i386__) || defined(__x86_64__)
   asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
 #endif  // defined(__i386__) || defined(__x86_64__)
-  return((xcr0 & 6) == 6);  // Is ymm saved?
+  return xcr0;
 }
 #endif  // defined(_M_IX86) || defined(_M_X64) ..
+// Return optimization to previous setting.
+#if defined(_M_IX86) && (_MSC_VER < 1900)
+#pragma optimize("g", on)
+#endif
 
 // based on libvpx arm_cpudetect.c
 // For Arm, but public to allow testing on any CPU
@@ -151,30 +161,9 @@ int ArmCpuCaps(const char* cpuinfo_name) {
   return 0;
 }
 
-#if defined(__mips__) && defined(__linux__)
-static int MipsCpuCaps(const char* search_string) {
-  char cpuinfo_line[512];
-  const char* file_name = "/proc/cpuinfo";
-  FILE* f = fopen(file_name, "r");
-  if (!f) {
-    // Assume DSP if /proc/cpuinfo is unavailable.
-    // This will occur for Chrome sandbox for Pepper or Render process.
-    return kCpuHasMIPS_DSP;
-  }
-  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
-    if (strstr(cpuinfo_line, search_string) != NULL) {
-      fclose(f);
-      return kCpuHasMIPS_DSP;
-    }
-  }
-  fclose(f);
-  return 0;
-}
-#endif
-
 // CPU detect function for SIMD instruction sets.
 LIBYUV_API
-int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
+int cpu_info_ = 0;  // cpu_info is not initialized yet.
 
 // Test environment variable for disabling CPU features. Any non-zero value
 // to disable. Zero ignored to make it easy to set the variable on/off.
@@ -197,8 +186,9 @@ static LIBYUV_BOOL TestEnv(const char*) {
 
 LIBYUV_API SAFEBUFFERS
 int InitCpuFlags(void) {
+  // TODO(fbarchard): swap kCpuInit logic so 0 means uninitialized.
+  int cpu_info = 0;
 #if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
-
   uint32 cpu_info0[4] = { 0, 0, 0, 0 };
   uint32 cpu_info1[4] = { 0, 0, 0, 0 };
   uint32 cpu_info7[4] = { 0, 0, 0, 0 };
@@ -207,66 +197,66 @@ int InitCpuFlags(void) {
   if (cpu_info0[0] >= 7) {
     CpuId(7, 0, cpu_info7);
   }
-  cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
-              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
-              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
-              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
-              ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
-              ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
-              kCpuHasX86;
+  cpu_info = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+             ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+             ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+             ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+             ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
+             ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+             kCpuHasX86;
 
 #ifdef HAS_XGETBV
-  if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
-      TestOsSaveYmm()) {  // Saves YMM.
-    cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
-                 kCpuHasAVX;
+  // AVX requires CPU has AVX, XSAVE and OSXSave for xgetbv
+  if (((cpu_info1[2] & 0x1c000000) == 0x1c000000) &&  // AVX and OSXSave
+      ((GetXCR0() & 6) == 6)) {  // Test OS saves YMM registers
+    cpu_info |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | kCpuHasAVX;
+
+    // Detect AVX512bw
+    if ((GetXCR0() & 0xe0) == 0xe0) {
+      cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX3 : 0;
+    }
   }
 #endif
+
   // Environment variable overrides for testing.
   if (TestEnv("LIBYUV_DISABLE_X86")) {
-    cpu_info_ &= ~kCpuHasX86;
+    cpu_info &= ~kCpuHasX86;
   }
   if (TestEnv("LIBYUV_DISABLE_SSE2")) {
-    cpu_info_ &= ~kCpuHasSSE2;
+    cpu_info &= ~kCpuHasSSE2;
   }
   if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
-    cpu_info_ &= ~kCpuHasSSSE3;
+    cpu_info &= ~kCpuHasSSSE3;
   }
   if (TestEnv("LIBYUV_DISABLE_SSE41")) {
-    cpu_info_ &= ~kCpuHasSSE41;
+    cpu_info &= ~kCpuHasSSE41;
   }
   if (TestEnv("LIBYUV_DISABLE_SSE42")) {
-    cpu_info_ &= ~kCpuHasSSE42;
+    cpu_info &= ~kCpuHasSSE42;
   }
   if (TestEnv("LIBYUV_DISABLE_AVX")) {
-    cpu_info_ &= ~kCpuHasAVX;
+    cpu_info &= ~kCpuHasAVX;
   }
   if (TestEnv("LIBYUV_DISABLE_AVX2")) {
-    cpu_info_ &= ~kCpuHasAVX2;
+    cpu_info &= ~kCpuHasAVX2;
   }
   if (TestEnv("LIBYUV_DISABLE_ERMS")) {
-    cpu_info_ &= ~kCpuHasERMS;
+    cpu_info &= ~kCpuHasERMS;
   }
   if (TestEnv("LIBYUV_DISABLE_FMA3")) {
-    cpu_info_ &= ~kCpuHasFMA3;
+    cpu_info &= ~kCpuHasFMA3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX3")) {
+    cpu_info &= ~kCpuHasAVX3;
   }
 #endif
 #if defined(__mips__) && defined(__linux__)
-  // Linux mips parse text file for dsp detect.
-  cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
 #if defined(__mips_dspr2)
-  cpu_info_ |= kCpuHasMIPS_DSPR2;
+  cpu_info |= kCpuHasDSPR2;
 #endif
-  cpu_info_ |= kCpuHasMIPS;
-
-  if (getenv("LIBYUV_DISABLE_MIPS")) {
-    cpu_info_ &= ~kCpuHasMIPS;
-  }
-  if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
-    cpu_info_ &= ~kCpuHasMIPS_DSP;
-  }
-  if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
-    cpu_info_ &= ~kCpuHasMIPS_DSPR2;
+  cpu_info |= kCpuHasMIPS;
+  if (getenv("LIBYUV_DISABLE_DSPR2")) {
+    cpu_info &= ~kCpuHasDSPR2;
   }
 #endif
 #if defined(__arm__) || defined(__aarch64__)
@@ -274,28 +264,31 @@ int InitCpuFlags(void) {
 // __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
 // For Linux, /proc/cpuinfo can be tested but without that assume Neon.
 #if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
-  cpu_info_ = kCpuHasNEON;
+  cpu_info = kCpuHasNEON;
 // For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon
 // flag in it.
 // So for aarch64, neon enabling is hard coded here.
 #endif
 #if defined(__aarch64__)
-  cpu_info_ = kCpuHasNEON;
+  cpu_info = kCpuHasNEON;
 #else
   // Linux arm parse text file for neon detect.
-  cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
+  cpu_info = ArmCpuCaps("/proc/cpuinfo");
 #endif
-  cpu_info_ |= kCpuHasARM;
+  cpu_info |= kCpuHasARM;
   if (TestEnv("LIBYUV_DISABLE_NEON")) {
-    cpu_info_ &= ~kCpuHasNEON;
+    cpu_info &= ~kCpuHasNEON;
   }
 #endif  // __arm__
   if (TestEnv("LIBYUV_DISABLE_ASM")) {
-    cpu_info_ = 0;
+    cpu_info = 0;
   }
-  return cpu_info_;
+  cpu_info  |= kCpuInitialized;
+  cpu_info_ = cpu_info;
+  return cpu_info;
 }
 
+// Note that use of this function is not thread safe.
 LIBYUV_API
 void MaskCpuFlags(int enable_flags) {
   cpu_info_ = InitCpuFlags() & enable_flags;