Merge pull request #140 from shuffle2/master

Implement workaround for Windows versions which do not support XSAVE.

Source/Core/Common/Common.h

@@ -101,20 +101,6 @@ private:
 
 #endif
 
-// Architecture detection for Windows
-// Architecture detection is done in cmake on all other platforms
-// Windows is built on only x86/x86_64
-#if _WIN32 || _WIN64
-#define _M_X86 1
-#if _WIN64
-#define _ARCH_64 1
-#define _M_X86_64 1
-#else
-#define _ARCH_32 1
-#define _M_X86_32 1
-#endif
-#endif
-
 // Windows compatibility
 #ifndef _WIN32
 #include <limits.h>

Source/Core/Common/Common.vcxproj

@@ -125,6 +125,7 @@
     <ClCompile Include="x64CPUDetect.cpp" />
     <ClCompile Include="x64Emitter.cpp" />
     <ClCompile Include="x64FPURoundMode.cpp" />
+    <ClCompile Include="XSaveWorkaround.cpp" />
   </ItemGroup>
   <ItemGroup>
     <Text Include="CMakeLists.txt" />

Source/Core/Common/Common.vcxproj.filters

@@ -103,6 +103,7 @@
       <Filter>Logging</Filter>
     </ClCompile>
     <ClCompile Include="stdafx.cpp" />
+    <ClCompile Include="XSaveWorkaround.cpp" />
   </ItemGroup>
   <ItemGroup>
     <Text Include="CMakeLists.txt" />

Source/Core/Common/XSaveWorkaround.cpp

@@ -0,0 +1,40 @@
+// Copyright 2013 Dolphin Emulator Project
+// Licensed under GPLv2
+// Refer to the license.txt file included.
+
+#if defined(_WIN32) && defined(_ARCH_64)
+
+#include <math.h>
+#include <Windows.h>
+
+// This puts the rest of this translation unit into a segment which is
+// initialized by the CRT *before* any of the other code (including C++
+// static initializers).
+#pragma warning(disable : 4075)
+#pragma init_seg(".CRT$XCB")
+
+struct EnableXSaveWorkaround
+{
+	EnableXSaveWorkaround()
+	{
+		// Some Windows environments may have hardware support for AVX/FMA,
+		// but the OS does not support it. The CRT math library does not support
+		// this scenario, so we have to manually tell it not to use FMA3
+		// instructions.
+
+		// The API name is somewhat misleading - we're testing for OS support
+		// here.
+		if (!IsProcessorFeaturePresent(PF_XSAVE_ENABLED))
+		{
+			_set_FMA3_enable(0);
+		}
+	}
+};
+
+static EnableXSaveWorkaround enableXSaveWorkaround;
+
+// N.B. Any code after this will still be in the .CRT$XCB segment. Please just
+// do not append any code here unless it is intended to be executed before
+// static initializers.
+
+#endif

Source/VSProps/Base.props

@@ -57,7 +57,9 @@
       <RuntimeTypeInfo>false</RuntimeTypeInfo>
       <MinimalRebuild>false</MinimalRebuild>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
-      <AdditionalOptions>/d2Zi+ /volatile:iso  /D PSAPI_VERSION=1 %(AdditionalOptions)</AdditionalOptions>
+      <AdditionalOptions>/d2Zi+ /volatile:iso /D PSAPI_VERSION=1 /D _M_X86=1 %(AdditionalOptions)</AdditionalOptions>
+      <AdditionalOptions Condition="'$(Platform)'=='Win32'">/D _ARCH_32=1 /D _M_X86_32=1 %(AdditionalOptions)</AdditionalOptions>
+      <AdditionalOptions Condition="'$(Platform)'=='x64'">/D _ARCH_64=1 /D _M_X86_64=1 %(AdditionalOptions)</AdditionalOptions>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
       <!--
       This is for GetVersionEx being marked as depreciated - which is idiotic and there's