[x64] Introduce FMA3 instructions on scalar data elements.

src/base/cpu.cc

 // Copyright 2013 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/base/cpu.h"
 
 #if V8_LIBC_MSVCRT
 #include <intrin.h>  // __cpuid()
 #endif
 #if V8_OS_LINUX
@@ skipping 46 matching lines @@
   __asm__ volatile (
     "cpuid \n\t"
     : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
     : "a"(info_type)
   );
 #endif  // defined(__i386__) && defined(__pic__)
 }
 
 #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
 
 #if V8_OS_LINUX
 
 #if V8_HOST_ARCH_ARM
 
 // See <uapi/asm/hwcap.h> kernel header.
 /*
  * HWCAP flags - for elf_hwcap (in kernel) and AT_HWCAP
@@ skipping 207 matching lines @@
   }
   return false;
 }
 
 #endif  // V8_HOST_ARCH_ARM || V8_HOST_ARCH_MIPS || V8_HOST_ARCH_MIPS64
 
 #endif  // V8_OS_LINUX
 
 #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.
 // TODO(jfb): Remove the hardcoded ARM simulator flags in the build, and
 // hardcode them here instead.
 #elif V8_HOST_ARCH_IA32 || V8_HOST_ARCH_X64
   int cpu_info[4];
 
   // __cpuid with an InfoType argument of 0 returns the number of
   // valid Ids in CPUInfo[0] and the CPU identification string in
@@ skipping 19 matching lines @@
     ext_family_ = (cpu_info[0] >> 20) & 0xff;
     has_fpu_ = (cpu_info[3] & 0x00000001) != 0;
     has_cmov_ = (cpu_info[3] & 0x00008000) != 0;
     has_mmx_ = (cpu_info[3] & 0x00800000) != 0;
     has_sse_ = (cpu_info[3] & 0x02000000) != 0;
     has_sse2_ = (cpu_info[3] & 0x04000000) != 0;
     has_sse3_ = (cpu_info[2] & 0x00000001) != 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
   // SAHF is always available in compat/legacy mode,
   has_sahf_ = true;
 #else
   // Query extended IDs.
   __cpuid(cpu_info, 0x80000000);
   unsigned num_ext_ids = cpu_info[0];
 
@@ skipping 182 matching lines @@
     if (end == part) {
       part_ = 0;
     }
     delete[] part;
   }
 
 #endif
 }
 
 } }  // namespace v8::base