Merge pull request #776 from shuffle2/improve-xsaveworkaround

Windows: Improve XSaveWorkaround to behave correctly when XSAVE processor feature is enabled, but AVX support isn't available for whatever reason.

1 files changed, 20 insertions, 0 deletions

diff --git a/Source/Core/Common/XSaveWorkaround.cpp b/Source/Core/Common/XSaveWorkaround.cpp
index 78f92b9bcf..d73b4a6ee0 100644
--- a/Source/Core/Common/XSaveWorkaround.cpp
+++ b/Source/Core/Common/XSaveWorkaround.cpp
@@ -7,6 +7,8 @@
 #include <math.h>
 #include <Windows.h>
 
+typedef decltype(&GetEnabledXStateFeatures) GetEnabledXStateFeatures_t;
+
 int __cdecl EnableXSaveWorkaround()
 {
 	// Some Windows environments may have hardware support for AVX/FMA,
@@ -19,6 +21,24 @@ int __cdecl EnableXSaveWorkaround()
 	if (!IsProcessorFeaturePresent(PF_XSAVE_ENABLED))
 	{
 		_set_FMA3_enable(0);
+		return 0;
+	}
+
+	// Even if XSAVE feature is enabled, we have to see if
+	// GetEnabledXStateFeatures function is present, and see what it says about
+	// AVX state.
+	auto kernel32Handle = GetModuleHandle(TEXT("kernel32.dll"));
+	if (kernel32Handle == nullptr)
+	{
+		std::abort();
+	}
+
+	auto pGetEnabledXStateFeatures = (GetEnabledXStateFeatures_t)GetProcAddress(
+		kernel32Handle, "GetEnabledXStateFeatures");
+	if (pGetEnabledXStateFeatures == nullptr ||
+		(pGetEnabledXStateFeatures() & XSTATE_MASK_AVX) == 0)
+	{
+		_set_FMA3_enable(0);
 	}
 
 	return 0;