[x64] Introduce FMA3 instructions on scalar data elements.

src/base/cpu.cc

Index: src/base/cpu.cc
diff --git a/src/base/cpu.cc b/src/base/cpu.cc
index 5bc8b13a97ba2fb9ca2094e32b43d288cf3b0157..7fde50cccd7f994456292300c176553f3bd663f9 100644
--- a/src/base/cpu.cc
+++ b/src/base/cpu.cc
@@ -64,6 +64,19 @@ static V8_INLINE void __cpuid(int cpu_info[4], int info_type) {
 
 #endif  // !V8_LIBC_MSVCRT
 
+static V8_INLINE bool avx_os_support() {

[Benedikt Meurer, 2014/11/25 17:32:42]

That does not belong here. The CPU class should only be concerned with detecting
AVX support in the CPU. Also I don't think that we actually care about the AVX
OS support unless we make use of the ymm registers. We should really separate
that, and move the AVX OS detection to platform.h.

+#if !V8_LIBC_MSVCRT
+  unsigned int eax, edx;
+  unsigned int ecx = 0;
+  __asm__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(ecx));
+  return (eax & 0x6) == 0x6;
+#else
+  uint64_t xcrFeatureMask = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
+  return (xcrFeatureMask & 0x6) == 0x6;
+#endif
+}
+
+
 #elif V8_HOST_ARCH_ARM || V8_HOST_ARCH_ARM64 \
     || V8_HOST_ARCH_MIPS || V8_HOST_ARCH_MIPS64
 
@@ -291,32 +304,35 @@ static bool HasListItem(const char* list, const char* item) {
 
 #endif  // V8_HOST_ARCH_IA32 || V8_HOST_ARCH_X64
 
-CPU::CPU() : stepping_(0),
-             model_(0),
-             ext_model_(0),
-             family_(0),
-             ext_family_(0),
-             type_(0),
-             implementer_(0),
-             architecture_(0),
-             part_(0),
-             has_fpu_(false),
-             has_cmov_(false),
-             has_sahf_(false),
-             has_mmx_(false),
-             has_sse_(false),
-             has_sse2_(false),
-             has_sse3_(false),
-             has_ssse3_(false),
-             has_sse41_(false),
-             has_sse42_(false),
-             has_idiva_(false),
-             has_neon_(false),
-             has_thumb2_(false),
-             has_vfp_(false),
-             has_vfp3_(false),
-             has_vfp3_d32_(false),
-             is_fp64_mode_(false) {
+CPU::CPU()
+    : stepping_(0),
+      model_(0),
+      ext_model_(0),
+      family_(0),
+      ext_family_(0),
+      type_(0),
+      implementer_(0),
+      architecture_(0),
+      part_(0),
+      has_fpu_(false),
+      has_cmov_(false),
+      has_sahf_(false),
+      has_mmx_(false),
+      has_sse_(false),
+      has_sse2_(false),
+      has_sse3_(false),
+      has_ssse3_(false),
+      has_sse41_(false),
+      has_sse42_(false),
+      has_avx_(false),
+      has_fma3_(false),
+      has_idiva_(false),
+      has_neon_(false),
+      has_thumb2_(false),
+      has_vfp_(false),
+      has_vfp3_(false),
+      has_vfp3_d32_(false),
+      is_fp64_mode_(false) {
   memcpy(vendor_, "Unknown", 8);
 #if V8_OS_NACL
 // Portable host shouldn't do feature detection.
@@ -356,6 +372,8 @@ CPU::CPU() : stepping_(0),
     has_ssse3_ = (cpu_info[2] & 0x00000200) != 0;
     has_sse41_ = (cpu_info[2] & 0x00080000) != 0;
     has_sse42_ = (cpu_info[2] & 0x00100000) != 0;
+    has_avx_ = ((cpu_info[2] & 0x18000000) == 0x18000000) && avx_os_support();

[Benedikt Meurer, 2014/11/25 17:32:42]

As mentioned above, remove the avx_os_support here.

+    if (has_avx_) has_fma3_ = (cpu_info[2] & 0x00001000) != 0;
   }
 
 #if V8_HOST_ARCH_IA32