Set the locale per-thread instead of globally when generating shaders. Add cross-compatible versions of newlocale, uselocale and freelocale.

This commit fixes a rare race condition when generating shaders because setlocale is global.

Source/Core/Common/Src/CommonFuncs.h

@@ -87,6 +87,45 @@ inline u64 _rotr64(u64 x, unsigned int shift){
 	#define snprintf _snprintf
 	#define vscprintf _vscprintf
 	
+// Locale Cross-Compatibility
+	#define locale_t _locale_t
+	#define freelocale _free_locale
+	#define newlocale(mask, locale, base) _create_locale(mask, locale)
+	
+	#define LC_GLOBAL_LOCALE	((locale_t)-1)
+	#define LC_ALL_MASK			LC_ALL
+	#define LC_COLLATE_MASK		LC_COLLATE
+	#define LC_CTYPE_MASK		LC_CTYPE
+	#define LC_MONETARY_MASK	LC_MONETARY
+	#define LC_NUMERIC_MASK		LC_NUMERIC
+	#define LC_TIME_MASK		LC_TIME
+	
+	inline locale_t uselocale(locale_t new_locale)
+	{
+		// Retrieve the current per thread locale setting
+		bool bIsPerThread = (_configthreadlocale(0) == _ENABLE_PER_THREAD_LOCALE);
+
+		// Retrieve the current thread-specific locale
+		locale_t old_locale = bIsPerThread ? _get_current_locale() : LC_GLOBAL_LOCALE;
+
+		if(new_locale == LC_GLOBAL_LOCALE)
+		{
+			// Restore the global locale
+			_configthreadlocale(_DISABLE_PER_THREAD_LOCALE);
+		}
+		else if(new_locale != NULL)
+		{
+			// Configure the thread to set the locale only for this thread
+			_configthreadlocale(_ENABLE_PER_THREAD_LOCALE);
+
+			// Set all locale categories
+			for(int i = LC_MIN; i <= LC_MAX; i++)
+				setlocale(i, new_locale->locinfo->lc_category[i].locale);
+		}
+
+		return old_locale;
+	}
+
 // 64 bit offsets for windows
 	#define fseeko _fseeki64
 	#define ftello _ftelli64

Source/Core/VideoCommon/Src/PixelShaderGen.cpp

@@ -6,6 +6,9 @@
 #include <cmath>
 #include <assert.h>
 #include <locale.h>
+#ifdef __APPLE__
+	#include <xlocale.h>
+#endif
 
 #include "LightingShaderGen.h"
 #include "PixelShaderGen.h"
@@ -513,7 +516,8 @@ const char *WriteLocation(API_TYPE ApiType)
 
 const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components)
 {
-	setlocale(LC_NUMERIC, "C"); // Reset locale for compilation
+	locale_t locale = newlocale(LC_NUMERIC_MASK, "C", NULL); // New locale for compilation
+	locale_t old_locale = uselocale(locale); // Apply the locale for this thread
 	text[sizeof(text) - 1] = 0x7C;  // canary
 
 	BuildSwapModeTable(); // Needed for WriteStage
@@ -886,7 +890,8 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 	if (text[sizeof(text) - 1] != 0x7C)
 		PanicAlert("PixelShader generator - buffer too small, canary has been eaten!");
 
-	setlocale(LC_NUMERIC, ""); // restore locale
+	uselocale(old_locale); // restore locale
+	freelocale(locale);
 	return text;
 }
 

Source/Core/VideoCommon/Src/TextureConversionShader.cpp

@@ -6,6 +6,9 @@
 #include <stdio.h>
 #include <math.h>
 #include <locale.h>
+#ifdef __APPLE__
+	#include <xlocale.h>
+#endif
 
 #include "TextureConversionShader.h"
 #include "TextureDecoder.h"
@@ -804,7 +807,8 @@ void WriteZ24Encoder(char* p, API_TYPE ApiType)
 
 const char *GenerateEncodingShader(u32 format,API_TYPE ApiType)
 {
-	setlocale(LC_NUMERIC, "C"); // Reset locale for compilation
+	locale_t locale = newlocale(LC_NUMERIC_MASK, "C", NULL); // New locale for compilation
+	locale_t old_locale = uselocale(locale); // Apply the locale for this thread
 	text[sizeof(text) - 1] = 0x7C;  // canary
 
 	char *p = text;
@@ -888,7 +892,8 @@ const char *GenerateEncodingShader(u32 format,API_TYPE ApiType)
 	if (text[sizeof(text) - 1] != 0x7C)
 		PanicAlert("TextureConversionShader generator - buffer too small, canary has been eaten!");
 
-	setlocale(LC_NUMERIC, ""); // restore locale
+	uselocale(old_locale); // restore locale
+	freelocale(locale);
 	return text;
 }
 

Source/Core/VideoCommon/Src/VertexShaderGen.cpp

@@ -4,6 +4,9 @@
 
 #include <math.h>
 #include <locale.h>
+#ifdef __APPLE__
+	#include <xlocale.h>
+#endif
 
 #include "NativeVertexFormat.h"
 
@@ -173,7 +176,8 @@ extern const char *WriteLocation(API_TYPE ApiType);
 
 const char *GenerateVertexShaderCode(u32 components, API_TYPE ApiType)
 {
-	setlocale(LC_NUMERIC, "C"); // Reset locale for compilation
+	locale_t locale = newlocale(LC_NUMERIC_MASK, "C", NULL); // New locale for compilation
+	locale_t old_locale = uselocale(locale); // Apply the locale for this thread
 	text[sizeof(text) - 1] = 0x7C;  // canary
 
 	_assert_(bpmem.genMode.numtexgens == xfregs.numTexGen.numTexGens);
@@ -640,6 +644,7 @@ const char *GenerateVertexShaderCode(u32 components, API_TYPE ApiType)
 
 	if (text[sizeof(text) - 1] != 0x7C)
 		PanicAlert("VertexShader generator - buffer too small, canary has been eaten!");
-	setlocale(LC_NUMERIC, ""); // restore locale
+	uselocale(old_locale); // restore locale
+	freelocale(locale);
 	return text;
 }

Source/Plugins/Plugin_VideoDX9/Src/PixelShaderCache.cpp

@@ -4,6 +4,10 @@
 
 #include <map>
 #include <set>
+#include <locale.h>
+#ifdef __APPLE__
+	#include <xlocale.h>
+#endif
 
 #include "Common.h"
 #include "Hash.h"
@@ -155,7 +159,8 @@ static LPDIRECT3DPIXELSHADER9 CreateCopyShader(int copyMatrixType, int depthConv
 	// this should create the same shaders as before (plus some extras added for DF16), just... more manageably than listing the full program for each combination
 	char text[3072];
 
-	setlocale(LC_NUMERIC, "C"); // Reset locale for compilation
+	locale_t locale = newlocale(LC_NUMERIC_MASK, "C", NULL); // New locale for compilation
+	locale_t old_locale = uselocale(locale); // Apply the locale for this thread
 	text[sizeof(text) - 1] = 0x7C;  // canary
 
 	char* p = text;
@@ -215,7 +220,8 @@ static LPDIRECT3DPIXELSHADER9 CreateCopyShader(int copyMatrixType, int depthConv
 	if (text[sizeof(text) - 1] != 0x7C)
 		PanicAlert("PixelShaderCache copy shader generator - buffer too small, canary has been eaten!");
 
-	setlocale(LC_NUMERIC, ""); // restore locale
+	uselocale(old_locale); // restore locale
+	freelocale(locale);
 	return D3D::CompileAndCreatePixelShader(text, (int)strlen(text));
 }