Merge pull request #3161 from shuffle2/ucrt-fread-workaround

exchange XSaveWorkaround for ucrtFreadWorkaround

Source/Core/Common/Common.vcxproj

@@ -162,13 +162,13 @@
     <ClCompile Include="Thread.cpp" />
     <ClCompile Include="Timer.cpp" />
     <ClCompile Include="TraversalClient.cpp" />
+    <ClCompile Include="ucrtFreadWorkaround.cpp" />
     <ClCompile Include="Version.cpp" />
     <ClCompile Include="x64ABI.cpp" />
     <ClCompile Include="x64Analyzer.cpp" />
     <ClCompile Include="x64CPUDetect.cpp" />
     <ClCompile Include="x64Emitter.cpp" />
     <ClCompile Include="x64FPURoundMode.cpp" />
-    <ClCompile Include="XSaveWorkaround.cpp" />
     <ClCompile Include="Crypto\bn.cpp" />
     <ClCompile Include="Crypto\ec.cpp" />
     <ClCompile Include="Logging\LogManager.cpp" />

Source/Core/Common/Common.vcxproj.filters

@@ -230,7 +230,6 @@
     <ClCompile Include="Logging\LogManager.cpp">
       <Filter>Logging</Filter>
     </ClCompile>
-    <ClCompile Include="XSaveWorkaround.cpp" />
     <ClCompile Include="GekkoDisassembler.cpp" />
     <ClCompile Include="JitRegister.cpp" />
     <ClCompile Include="TraversalClient.cpp" />
@@ -249,6 +248,7 @@
     <ClCompile Include="GL\GLInterface\GLInterface.cpp">
       <Filter>GL\GLInterface</Filter>
     </ClCompile>
+    <ClCompile Include="ucrtFreadWorkaround.cpp" />
   </ItemGroup>
   <ItemGroup>
     <Text Include="CMakeLists.txt" />

Source/Core/Common/XSaveWorkaround.cpp

@@ -1,61 +0,0 @@
-// Copyright 2014 Dolphin Emulator Project
-// Licensed under GPLv2+
-// Refer to the license.txt file included.
-
-#if defined(_WIN32)
-
-#include <math.h>
-#include <Windows.h>
-
-typedef decltype(&GetEnabledXStateFeatures) GetEnabledXStateFeatures_t;
-
-int __cdecl 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);
-		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;
-}
-
-// Create a segment which is recognized by the linker to be part of the CRT
-// initialization. XI* = C startup, XC* = C++ startup. "A" placement is reserved
-// for system use. Thus, the earliest we can get is XIB (C startup is before
-// C++).
-#pragma section(".CRT$XIB", read)
-
-// Place a symbol in the special segment, make it have C linkage so that
-// referencing it doesn't require ugly decorated names.
-// Use /include:XSaveWorkaround linker flag to enable this.
-extern "C" {
-	__declspec(allocate(".CRT$XIB"))
-	decltype(&EnableXSaveWorkaround) XSaveWorkaround = EnableXSaveWorkaround;
-};
-
-#endif

Source/Core/Common/ucrtFreadWorkaround.cpp

@@ -0,0 +1,85 @@
+// Copyright 2014 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#if defined(_WIN32)
+
+#include "CommonTypes.h"
+#include <Windows.h>
+
+struct PatchInfo {
+	const wchar_t *module_name;
+	u32 checksum;
+	u32 rva;
+	u32 length;
+} static const s_patches[] = {
+	{ L"ucrtbase.dll", 0xF61ED, 0x6AE7B, 5 },
+	{ L"ucrtbased.dll", 0x1C1915 , 0x91905, 5 },
+};
+
+bool ApplyPatch(const PatchInfo &patch) {
+	auto module = GetModuleHandleW(patch.module_name);
+	if (module == nullptr)
+	{
+		return false;
+	}
+
+	auto ucrtbase_pe = (PIMAGE_NT_HEADERS)((uintptr_t)module + ((PIMAGE_DOS_HEADER)module)->e_lfanew);
+	if (ucrtbase_pe->OptionalHeader.CheckSum != patch.checksum) {
+		return false;
+	}
+
+	void *patch_addr = (void *)((uintptr_t)module + patch.rva);
+	size_t patch_size = patch.length;
+
+	DWORD old_protect;
+	if (!VirtualProtect(patch_addr, patch_size, PAGE_EXECUTE_READWRITE, &old_protect))
+	{
+		return false;
+	}
+
+	memset(patch_addr, 0x90, patch_size);
+
+	VirtualProtect(patch_addr, patch_size, old_protect, &old_protect);
+
+	FlushInstructionCache(GetCurrentProcess(), patch_addr, patch_size);
+
+	return true;
+}
+
+int __cdecl EnableucrtFreadWorkaround()
+{
+	// This patches ucrtbase such that fseek will always
+	// synchronize the file object's internal buffer.
+
+	bool applied_at_least_one = false;
+	for (const auto &patch : s_patches) {
+		if (ApplyPatch(patch)) {
+			applied_at_least_one = true;
+		}
+	}
+
+	/* For forward compat, do not fail if patches don't apply (e.g. version mismatch)
+	if (!applied_at_least_one) {
+		std::abort();
+	}
+	//*/
+
+	return 0;
+}
+
+// Create a segment which is recognized by the linker to be part of the CRT
+// initialization. XI* = C startup, XC* = C++ startup. "A" placement is reserved
+// for system use. Thus, the earliest we can get is XIB (C startup is before
+// C++).
+#pragma section(".CRT$XIB", read)
+
+// Place a symbol in the special segment, make it have C linkage so that
+// referencing it doesn't require ugly decorated names.
+// Use /include:EnableucrtFreadWorkaround linker flag to enable this.
+extern "C" {
+	__declspec(allocate(".CRT$XIB"))
+	decltype(&EnableucrtFreadWorkaround) ucrtFreadWorkaround = EnableucrtFreadWorkaround;
+};
+
+#endif

Source/VSProps/Base.props

@@ -122,8 +122,8 @@
     </Link>
     <!--Link Base:Application-->
     <Link Condition="'$(ConfigurationType)'=='Application'">
-      <!--See Common/EnableXSaveWorkaround.cpp-->
+      <!--See Common/ucrtFreadWorkaround.cpp-->
-      <ForceSymbolReferences>XSaveWorkaround</ForceSymbolReferences>
+      <ForceSymbolReferences>ucrtFreadWorkaround</ForceSymbolReferences>
     </Link>
     <Lib>
       <TreatLibWarningAsErrors>true</TreatLibWarningAsErrors>