Merge branch 'master' of https://code.google.com/p/dolphin-emu into dx9-ssaa-fix

8 files changed, 123 insertions, 142 deletions

diff --git a/Source/Core/VideoCommon/Src/BPMemory.cpp b/Source/Core/VideoCommon/Src/BPMemory.cpp
index 77f2f6bab3..7272ed0993 100644
--- a/Source/Core/VideoCommon/Src/BPMemory.cpp
+++ b/Source/Core/VideoCommon/Src/BPMemory.cpp
@@ -299,38 +299,3 @@ void GetBPRegInfo(const u8* data, char* name, size_t name_size, char* desc, size
 #undef SetRegName
 	}
 }
-
-AlphaTest::TEST_RESULT AlphaTest::TestResult()
-{
-	switch(logic)
-	{
-	case 0: // AND
-		if (comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_ALWAYS)
-			return PASS;
-		if (comp0 == ALPHACMP_NEVER || comp1 == ALPHACMP_NEVER)
-			return FAIL;
-		break;
-
-	case 1: // OR
-		if (comp0 == ALPHACMP_ALWAYS || comp1 == ALPHACMP_ALWAYS)
-			return PASS;
-		if (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_NEVER)
-			return FAIL;
-		break;
-
-	case 2: // XOR
-		if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_NEVER) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_ALWAYS))
-			return PASS;
-		if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_ALWAYS) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_NEVER))
-			return FAIL;
-		break;
-
-	case 3: // XNOR
-		if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_NEVER) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_ALWAYS))
-			return FAIL;
-		if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_ALWAYS) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_NEVER))
-			return PASS;
-		break;
-	}
-	return UNDETERMINED;
-}
diff --git a/Source/Core/VideoCommon/Src/BPMemory.h b/Source/Core/VideoCommon/Src/BPMemory.h
index 7ce19376f5..83dbefed50 100644
--- a/Source/Core/VideoCommon/Src/BPMemory.h
+++ b/Source/Core/VideoCommon/Src/BPMemory.h
@@ -885,7 +885,40 @@ union AlphaTest
 		PASS = 2,
 	};
 
-	TEST_RESULT TestResult();
+	inline TEST_RESULT TestResult() const
+	{
+		switch(logic)
+		{
+		case 0: // AND
+			if (comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_ALWAYS)
+				return PASS;
+			if (comp0 == ALPHACMP_NEVER || comp1 == ALPHACMP_NEVER)
+				return FAIL;
+			break;
+
+		case 1: // OR
+			if (comp0 == ALPHACMP_ALWAYS || comp1 == ALPHACMP_ALWAYS)
+				return PASS;
+			if (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_NEVER)
+				return FAIL;
+			break;
+
+		case 2: // XOR
+			if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_NEVER) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_ALWAYS))
+				return PASS;
+			if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_ALWAYS) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_NEVER))
+				return FAIL;
+			break;
+
+		case 3: // XNOR
+			if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_NEVER) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_ALWAYS))
+				return FAIL;
+			if ((comp0 == ALPHACMP_ALWAYS && comp1 == ALPHACMP_ALWAYS) || (comp0 == ALPHACMP_NEVER && comp1 == ALPHACMP_NEVER))
+				return PASS;
+			break;
+		}
+		return UNDETERMINED;
+	}
 };
 
 union UPE_Copy
diff --git a/Source/Core/VideoCommon/Src/LightingShaderGen.h b/Source/Core/VideoCommon/Src/LightingShaderGen.h
index b7f596bdd3..fc107a2e2b 100644
--- a/Source/Core/VideoCommon/Src/LightingShaderGen.h
+++ b/Source/Core/VideoCommon/Src/LightingShaderGen.h
@@ -9,40 +9,22 @@
 #include "NativeVertexFormat.h"
 #include "XFMemory.h"
 
-static const char* LightCol(const char* lightsName, unsigned int index, const char* swizzle)
-{
-	static char result[32];
-	snprintf(result, sizeof(result), "%s[5*%d].%s", lightsName, index, swizzle);
-	return result;
-}
 
-static const char* LightCosAtt(const char* lightsName, unsigned int index)
-{
-	static char result[32];
-	snprintf(result, sizeof(result), "%s[5*%d+1]", lightsName, index);
-	return result;
-}
+#define LIGHT_COL "%s[5*%d].%s"
+#define LIGHT_COL_PARAMS(lightsName, index, swizzle) (lightsName), (index), (swizzle)
 
-static const char* LightDistAtt(const char* lightsName, unsigned int index)
-{
-	static char result[32];
-	snprintf(result, sizeof(result), "%s[5*%d+2]", lightsName, index);
-	return result;
-}
+#define LIGHT_COSATT "%s[5*%d+1]"
+#define LIGHT_COSATT_PARAMS(lightsName, index) (lightsName), (index)
 
-static const char* LightPos(const char* lightsName, unsigned int index)
-{
-	static char result[32];
-	snprintf(result, sizeof(result), "%s[5*%d+3]", lightsName, index);
-	return result;
-}
+#define LIGHT_DISTATT "%s[5*%d+2]"
+#define LIGHT_DISTATT_PARAMS(lightsName, index) (lightsName), (index)
+
+#define LIGHT_POS "%s[5*%d+3]"
+#define LIGHT_POS_PARAMS(lightsName, index) (lightsName), (index)
+
+#define LIGHT_DIR "%s[5*%d+4]"
+#define LIGHT_DIR_PARAMS(lightsName, index) (lightsName), (index)
 
-static const char* LightDir(const char* lightsName, unsigned int index)
-{
-	static char result[32];
-	snprintf(result, sizeof(result), "%s[5*%d+4]", lightsName, index);
-	return result;
-}
 
 template<class T>
 static void GenerateLightShader(T& object, LightingUidData& uid_data, int index, int litchan_index, const char* lightsName, int coloralpha)
@@ -62,13 +44,13 @@ static void GenerateLightShader(T& object, LightingUidData& uid_data, int index,
 		switch (chan.diffusefunc)
 		{
 			case LIGHTDIF_NONE:
-				object.Write("lacc.%s += %s;\n", swizzle, LightCol(lightsName, index, swizzle));
+				object.Write("lacc.%s += " LIGHT_COL";\n", swizzle, LIGHT_COL_PARAMS(lightsName, index, swizzle));
 				break;
 			case LIGHTDIF_SIGN:
 			case LIGHTDIF_CLAMP:
-				object.Write("ldir = normalize(%s.xyz - pos.xyz);\n", LightPos(lightsName, index));
-				object.Write("lacc.%s += %sdot(ldir, _norm0)) * %s;\n",
-					swizzle, chan.diffusefunc != LIGHTDIF_SIGN ? "max(0.0f," :"(", LightCol(lightsName, index, swizzle));
+				object.Write("ldir = normalize(" LIGHT_POS".xyz - pos.xyz);\n", LIGHT_POS_PARAMS(lightsName, index));
+				object.Write("lacc.%s += %sdot(ldir, _norm0)) * " LIGHT_COL";\n",
+					swizzle, chan.diffusefunc != LIGHTDIF_SIGN ? "max(0.0f," :"(", LIGHT_COL_PARAMS(lightsName, index, swizzle));
 				break;
 			default: _assert_(0);
 		}
@@ -77,31 +59,34 @@ static void GenerateLightShader(T& object, LightingUidData& uid_data, int index,
 	{
 		if (chan.attnfunc == 3)
 		{ // spot
-			object.Write("ldir = %s.xyz - pos.xyz;\n", LightPos(lightsName, index));
+			object.Write("ldir = " LIGHT_POS".xyz - pos.xyz;\n", LIGHT_POS_PARAMS(lightsName, index));
 			object.Write("dist2 = dot(ldir, ldir);\n"
 						"dist = sqrt(dist2);\n"
 						"ldir = ldir / dist;\n"
-						"attn = max(0.0f, dot(ldir, %s.xyz));\n", LightDir(lightsName, index));
-			object.Write("attn = max(0.0f, dot(%s.xyz, float3(1.0f, attn, attn*attn))) / dot(%s.xyz, float3(1.0f,dist,dist2));\n", LightCosAtt(lightsName, index), LightDistAtt(lightsName, index));
+						"attn = max(0.0f, dot(ldir, " LIGHT_DIR".xyz));\n",
+						LIGHT_DIR_PARAMS(lightsName, index));
+			object.Write("attn = max(0.0f, dot(" LIGHT_COSATT".xyz, float3(1.0f, attn, attn*attn))) / dot(" LIGHT_DISTATT".xyz, float3(1.0f,dist,dist2));\n",
+						LIGHT_COSATT_PARAMS(lightsName, index), LIGHT_DISTATT_PARAMS(lightsName, index));
 		}
 		else if (chan.attnfunc == 1)
 		{ // specular
-			object.Write("ldir = normalize(%s.xyz);\n", LightPos(lightsName, index));
-			object.Write("attn = (dot(_norm0,ldir) >= 0.0f) ? max(0.0f, dot(_norm0, %s.xyz)) : 0.0f;\n", LightDir(lightsName, index));
-			object.Write("attn = max(0.0f, dot(%s.xyz, float3(1,attn,attn*attn))) / dot(%s.xyz, float3(1,attn,attn*attn));\n", LightCosAtt(lightsName, index), LightDistAtt(lightsName, index));
+			object.Write("ldir = normalize(" LIGHT_POS".xyz);\n", LIGHT_POS_PARAMS(lightsName, index));
+			object.Write("attn = (dot(_norm0,ldir) >= 0.0f) ? max(0.0f, dot(_norm0, " LIGHT_DIR".xyz)) : 0.0f;\n", LIGHT_DIR_PARAMS(lightsName, index));
+			object.Write("attn = max(0.0f, dot(" LIGHT_COSATT".xyz, float3(1,attn,attn*attn))) / dot(" LIGHT_DISTATT".xyz, float3(1,attn,attn*attn));\n",
+						LIGHT_COSATT_PARAMS(lightsName, index), LIGHT_DISTATT_PARAMS(lightsName, index));
 		}
 
 		switch (chan.diffusefunc)
 		{
 			case LIGHTDIF_NONE:
-				object.Write("lacc.%s += attn * %s;\n", swizzle, LightCol(lightsName, index, swizzle));
+				object.Write("lacc.%s += attn * " LIGHT_COL";\n", swizzle, LIGHT_COL_PARAMS(lightsName, index, swizzle));
 				break;
 			case LIGHTDIF_SIGN:
 			case LIGHTDIF_CLAMP:
-				object.Write("lacc.%s += attn * %sdot(ldir, _norm0)) * %s;\n",
+				object.Write("lacc.%s += attn * %sdot(ldir, _norm0)) * " LIGHT_COL";\n",
 					swizzle,
 					chan.diffusefunc != LIGHTDIF_SIGN ? "max(0.0f," :"(",
-					LightCol(lightsName, index, swizzle));
+					LIGHT_COL_PARAMS(lightsName, index, swizzle));
 				break;
 			default: _assert_(0);
 		}
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
index 28e609112e..7de5d12638 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
@@ -216,7 +216,7 @@ static char swapModeTable[4][5];
 
 static char text[16384];
 
-static void BuildSwapModeTable()
+static inline void BuildSwapModeTable()
 {
 	static const char *swapColors = "rgba";
 	for (int i = 0; i < 4; i++)
@@ -229,13 +229,13 @@ static void BuildSwapModeTable()
 	}
 }
 
-template<class T> static void WriteStage(T& out, pixel_shader_uid_data& uid_data, int n, API_TYPE ApiType, RegisterState RegisterStates[4]);
-template<class T> static void SampleTexture(T& out, const char *texcoords, const char *texswap, int texmap, API_TYPE ApiType);
-template<class T> static void WriteAlphaTest(T& out, pixel_shader_uid_data& uid_data, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth);
-template<class T> static void WriteFog(T& out, pixel_shader_uid_data& uid_data);
+template<class T> static inline void WriteStage(T& out, pixel_shader_uid_data& uid_data, int n, API_TYPE ApiType, RegisterState RegisterStates[4]);
+template<class T> static inline void SampleTexture(T& out, const char *texcoords, const char *texswap, int texmap, API_TYPE ApiType);
+template<class T> static inline void WriteAlphaTest(T& out, pixel_shader_uid_data& uid_data, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth);
+template<class T> static inline void WriteFog(T& out, pixel_shader_uid_data& uid_data);
 
 template<class T>
-static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components)
+static inline void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components)
 {
 	// Non-uid template parameters will write to the dummy data (=> gets optimized out)
 	pixel_shader_uid_data dummy_data;
@@ -243,17 +243,19 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 										? out.template GetUidData<pixel_shader_uid_data>() : dummy_data;
 
 	out.SetBuffer(text);
+	const bool is_writing_shadercode = (out.GetBuffer() != NULL);
 #ifndef ANDROID
 	locale_t locale;
 	locale_t old_locale;
-	if (out.GetBuffer() != NULL)
+	if (is_writing_shadercode)
 	{
 		locale = newlocale(LC_NUMERIC_MASK, "C", NULL); // New locale for compilation
 		old_locale = uselocale(locale); // Apply the locale for this thread
 	}
 #endif
 
-	text[sizeof(text) - 1] = 0x7C;  // canary
+	if (is_writing_shadercode)
+		text[sizeof(text) - 1] = 0x7C;  // canary
 
 	unsigned int numStages = bpmem.genMode.numtevstages + 1;
 	unsigned int numTexgen = bpmem.genMode.numtexgens;
@@ -331,7 +333,7 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 
 		out.Write("VARYIN float4 colors_02;\n");
 		out.Write("VARYIN float4 colors_12;\n");
-		
+
 		// compute window position if needed because binding semantic WPOS is not widely supported
 		// Let's set up attributes
 		if (xfregs.numTexGen.numTexGens < 7)
@@ -372,7 +374,7 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 			// It just allows it, but it seems that all drivers do.
 			out.Write("layout(early_fragment_tests) in;\n");
 		}
-		else if (bpmem.UseEarlyDepthTest() && (g_ActiveConfig.bFastDepthCalc || bpmem.alpha_test.TestResult() == AlphaTest::UNDETERMINED))
+		else if (bpmem.UseEarlyDepthTest() && (g_ActiveConfig.bFastDepthCalc || bpmem.alpha_test.TestResult() == AlphaTest::UNDETERMINED) && is_writing_shadercode)
 		{
 			static bool warn_once = true;
 			if (warn_once)
@@ -388,7 +390,7 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 		{
 			out.Write("[earlydepthstencil]\n");
 		}
-		else if (bpmem.UseEarlyDepthTest() && (g_ActiveConfig.bFastDepthCalc || bpmem.alpha_test.TestResult() == AlphaTest::UNDETERMINED))
+		else if (bpmem.UseEarlyDepthTest() && (g_ActiveConfig.bFastDepthCalc || bpmem.alpha_test.TestResult() == AlphaTest::UNDETERMINED) && is_writing_shadercode)
 		{
 			static bool warn_once = true;
 			if (warn_once)
@@ -500,7 +502,6 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 
 	if (g_ActiveConfig.bEnablePixelLighting && g_ActiveConfig.backend_info.bSupportsPixelLighting)
 	{
-		uid_data.xfregs_numTexGen_numTexGens = xfregs.numTexGen.numTexGens;
 		if (xfregs.numTexGen.numTexGens < 7)
 		{
 			out.Write("\tfloat3 _norm0 = normalize(Normal.xyz);\n\n");
@@ -706,16 +707,16 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 
 	out.Write("}\n");
 
-	if (text[sizeof(text) - 1] != 0x7C)
-		PanicAlert("PixelShader generator - buffer too small, canary has been eaten!");
+	if (is_writing_shadercode)
+	{
+		if (text[sizeof(text) - 1] != 0x7C)
+			PanicAlert("PixelShader generator - buffer too small, canary has been eaten!");
 
 #ifndef ANDROID
-	if (out.GetBuffer() != NULL)
-	{
 		uselocale(old_locale); // restore locale
 		freelocale(locale);
-	}
 #endif
+	}
 }
 
 
@@ -763,7 +764,7 @@ static const char *TEVCMPAlphaOPTable[16] =
 };
 
 template<class T>
-static void WriteStage(T& out, pixel_shader_uid_data& uid_data, int n, API_TYPE ApiType, RegisterState RegisterStates[4])
+static inline void WriteStage(T& out, pixel_shader_uid_data& uid_data, int n, API_TYPE ApiType, RegisterState RegisterStates[4])
 {
 	int texcoord = bpmem.tevorders[n/2].getTexCoord(n&1);
 	bool bHasTexCoord = (u32)texcoord < bpmem.genMode.numtexgens;
@@ -906,7 +907,7 @@ static void WriteStage(T& out, pixel_shader_uid_data& uid_data, int n, API_TYPE
 		char *texswap = swapModeTable[bpmem.combiners[n].alphaC.tswap];
 		int texmap = bpmem.tevorders[n/2].getTexMap(n&1);
 		uid_data.SetTevindrefTexmap(i, texmap);
-		
+
 		out.Write("textemp = ");
 		SampleTexture<T>(out, "tevcoord", texswap, texmap, ApiType);
 	}
@@ -1133,7 +1134,7 @@ template<class T>
 void SampleTexture(T& out, const char *texcoords, const char *texswap, int texmap, API_TYPE ApiType)
 {
 	out.SetConstantsUsed(C_TEXDIMS+texmap,C_TEXDIMS+texmap);
-	
+
 	if (ApiType == API_D3D11)
 		out.Write("Tex%d.Sample(samp%d,%s.xy * " I_TEXDIMS"[%d].xy).%s;\n", texmap,texmap, texcoords, texmap, texswap);
 	else
@@ -1161,7 +1162,7 @@ static const char *tevAlphaFunclogicTable[] =
 };
 
 template<class T>
-static void WriteAlphaTest(T& out, pixel_shader_uid_data& uid_data, API_TYPE ApiType, DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth)
+static inline void WriteAlphaTest(T& out, pixel_shader_uid_data& uid_data, API_TYPE ApiType, DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth)
 {
 	static const char *alphaRef[2] =
 	{
@@ -1232,7 +1233,7 @@ static const char *tevFogFuncsTable[] =
 };
 
 template<class T>
-static void WriteFog(T& out, pixel_shader_uid_data& uid_data)
+static inline void WriteFog(T& out, pixel_shader_uid_data& uid_data)
 {
 	uid_data.fog_fsel = bpmem.fog.c_proj_fsel.fsel;
 	if(bpmem.fog.c_proj_fsel.fsel == 0)
@@ -1274,7 +1275,7 @@ static void WriteFog(T& out, pixel_shader_uid_data& uid_data)
 	}
 	else
 	{
-		if (bpmem.fog.c_proj_fsel.fsel != 2)
+		if (bpmem.fog.c_proj_fsel.fsel != 2 && out.GetBuffer() != NULL)
 			WARN_LOG(VIDEO, "Unknown Fog Type! %08x", bpmem.fog.c_proj_fsel.fsel);
 	}
 
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.h b/Source/Core/VideoCommon/Src/PixelShaderGen.h
index c1ca009438..7763cbd574 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.h
@@ -63,18 +63,30 @@ struct pixel_shader_uid_data
 	u32 num_values; // TODO: Shouldn't be a u32
 	u32 NumValues() const { return num_values; }
 
-	u32 components;
+	u32 components : 23;
 	u32 dstAlphaMode : 2;
 	u32 Pretest : 2;
+	u32 nIndirectStagesUsed : 4;
+	u32 pad0 : 1;
 
 	u32 genMode_numtexgens : 4;
 	u32 genMode_numtevstages : 4;
 	u32 genMode_numindstages : 3;
-
-	u32 nIndirectStagesUsed : 8;
+	u32 alpha_test_comp0 : 3;
+	u32 alpha_test_comp1 : 3;
+	u32 alpha_test_logic : 2;
+	u32 alpha_test_use_zcomploc_hack : 1;
+	u32 fog_proj : 1;
+	u32 fog_fsel : 3;
+	u32 fog_RangeBaseEnabled : 1;
+	u32 ztex_op : 2;
+	u32 fast_depth_calc : 1;
+	u32 per_pixel_depth : 1;
+	u32 forced_early_z : 1;
+	u32 early_ztest : 1;
+	u32 pad1 : 1;
 
 	u32 texMtxInfo_n_projection : 8; // 8x1 bit
-
 	u32 tevindref_bi0 : 3;
 	u32 tevindref_bc0 : 3;
 	u32 tevindref_bi1 : 3;
@@ -83,6 +95,7 @@ struct pixel_shader_uid_data
 	u32 tevindref_bc3 : 3;
 	u32 tevindref_bi4 : 3;
 	u32 tevindref_bc4 : 3;
+
 	inline void SetTevindrefValues(int index, u32 texcoord, u32 texmap)
 	{
 		if (index == 0) { tevindref_bc0 = texcoord; tevindref_bi0 = texmap; }
@@ -98,25 +111,6 @@ struct pixel_shader_uid_data
 		else if (index == 3) { tevindref_bi4 = texmap; }
 	}
 
-	u32 alpha_test_comp0 : 3;
-	u32 alpha_test_comp1 : 3;
-	u32 alpha_test_logic : 2;
-
-	u32 alpha_test_use_zcomploc_hack : 1;
-
-	u32 fog_proj : 1;
-	u32 fog_fsel : 3;
-	u32 fog_RangeBaseEnabled : 1;
-
-	u32 ztex_op : 2;
-
-	u32 fast_depth_calc : 1;
-	u32 per_pixel_depth : 1;
-	u32 forced_early_z : 1;
-	u32 early_ztest : 1;
-
-	u32 xfregs_numTexGen_numTexGens : 4;
-
 	struct {
 		// TODO: Can save a lot space by removing the padding bits
 		u32 cc : 24;
diff --git a/Source/Core/VideoCommon/Src/ShaderGenCommon.h b/Source/Core/VideoCommon/Src/ShaderGenCommon.h
index ec4fe27416..a804c08f13 100644
--- a/Source/Core/VideoCommon/Src/ShaderGenCommon.h
+++ b/Source/Core/VideoCommon/Src/ShaderGenCommon.h
@@ -173,7 +173,7 @@ private:
 };
 
 template<class T>
-static void WriteRegister(T& object, API_TYPE ApiType, const char *prefix, const u32 num)
+static inline void WriteRegister(T& object, API_TYPE ApiType, const char *prefix, const u32 num)
 {
 	if (ApiType == API_OPENGL)
 		return; // Nothing to do here
@@ -182,7 +182,7 @@ static void WriteRegister(T& object, API_TYPE ApiType, const char *prefix, const
 }
 
 template<class T>
-static void WriteLocation(T& object, API_TYPE ApiType, bool using_ubos)
+static inline void WriteLocation(T& object, API_TYPE ApiType, bool using_ubos)
 {
 	if (using_ubos)
 		return;
@@ -191,7 +191,7 @@ static void WriteLocation(T& object, API_TYPE ApiType, bool using_ubos)
 }
 
 template<class T>
-static void DeclareUniform(T& object, API_TYPE api_type, bool using_ubos, const u32 num, const char* type, const char* name)
+static inline void DeclareUniform(T& object, API_TYPE api_type, bool using_ubos, const u32 num, const char* type, const char* name)
 {
 	WriteLocation(object, api_type, using_ubos);
 	object.Write("%s %s ", type, name);
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
index a682b4ea21..e44e01a13b 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
@@ -38,7 +38,7 @@ static void DefineVSOutputStructMember(T& object, API_TYPE api_type, const char*
 }
 
 template<class T>
-static void GenerateVSOutputStruct(T& object, u32 components, API_TYPE api_type)
+static inline void GenerateVSOutputStruct(T& object, u32 components, API_TYPE api_type)
 {
 	object.Write("struct VS_OUTPUT {\n");
 	DefineVSOutputStructMember(object, api_type, "float4", "pos", -1, "POSITION");
@@ -67,7 +67,7 @@ static void GenerateVSOutputStruct(T& object, u32 components, API_TYPE api_type)
 }
 
 template<class T>
-static void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
+static inline void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
 {
 	// Non-uid template parameters will write to the dummy data (=> gets optimized out)
 	vertex_shader_uid_data dummy_data;
@@ -75,16 +75,19 @@ static void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
 											? out.template GetUidData<vertex_shader_uid_data>() : dummy_data;
 
 	out.SetBuffer(text);
+	const bool is_writing_shadercode = (out.GetBuffer() != NULL);
 #ifndef ANDROID
 	locale_t locale;
 	locale_t old_locale;
-	if (out.GetBuffer() != NULL)
+	if (is_writing_shadercode)
 	{
 		locale = newlocale(LC_NUMERIC_MASK, "C", NULL); // New locale for compilation
 		old_locale = uselocale(locale); // Apply the locale for this thread
 	}
 #endif
-	text[sizeof(text) - 1] = 0x7C;  // canary
+
+	if (is_writing_shadercode)
+		text[sizeof(text) - 1] = 0x7C;  // canary
 
 	_assert_(bpmem.genMode.numtexgens == xfregs.numTexGen.numTexGens);
 	_assert_(bpmem.genMode.numcolchans == xfregs.numChan.numColorChans);
@@ -225,7 +228,7 @@ static void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
 			out.Write("int posmtx = int(fposmtx);\n");
 		}
 
-		if (DriverDetails::HasBug(DriverDetails::BUG_NODYNUBOACCESS))
+		if (is_writing_shadercode && DriverDetails::HasBug(DriverDetails::BUG_NODYNUBOACCESS))
 		{
 			// This'll cause issues, but  it can't be helped
 			out.Write("float4 pos = float4(dot(" I_TRANSFORMMATRICES"[0], rawpos), dot(" I_TRANSFORMMATRICES"[1], rawpos), dot(" I_TRANSFORMMATRICES"[2], rawpos), 1);\n");
@@ -353,7 +356,7 @@ static void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
 					// transform the light dir into tangent space
 					uid_data.texMtxInfo[i].embosslightshift = xfregs.texMtxInfo[i].embosslightshift;
 					uid_data.texMtxInfo[i].embosssourceshift = xfregs.texMtxInfo[i].embosssourceshift;
-					out.Write("ldir = normalize(%s.xyz - pos.xyz);\n", LightPos(I_LIGHTS, texinfo.embosslightshift));
+					out.Write("ldir = normalize(" LIGHT_POS".xyz - pos.xyz);\n", LIGHT_POS_PARAMS(I_LIGHTS, texinfo.embosslightshift));
 					out.Write("o.tex%d.xyz = o.tex%d.xyz + float3(dot(ldir, _norm1), dot(ldir, _norm2), 0.0f);\n", i, texinfo.embosssourceshift);
 				}
 				else
@@ -547,16 +550,16 @@ static void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
 		out.Write("return o;\n}\n");
 	}
 
-	if (text[sizeof(text) - 1] != 0x7C)
-		PanicAlert("VertexShader generator - buffer too small, canary has been eaten!");
+	if (is_writing_shadercode)
+	{
+		if (text[sizeof(text) - 1] != 0x7C)
+			PanicAlert("VertexShader generator - buffer too small, canary has been eaten!");
 
 #ifndef ANDROID
-	if (out.GetBuffer() != NULL)
-	{
 		uselocale(old_locale); // restore locale
 		freelocale(locale);
-	}
 #endif
+	}
 }
 
 void GetVertexShaderUid(VertexShaderUid& object, u32 components, API_TYPE api_type)
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.h b/Source/Core/VideoCommon/Src/VertexShaderGen.h
index eef0646e6f..ab9a378107 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.h
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.h
@@ -69,12 +69,12 @@ struct vertex_shader_uid_data
 {
 	u32 NumValues() const { return sizeof(vertex_shader_uid_data); }
 
-	u32 components;
-	u32 numColorChans : 2;
+	u32 components : 23;
 	u32 numTexGens : 4;
-
+	u32 numColorChans : 2;
 	u32 dualTexTrans_enabled : 1;
 	u32 pixel_lighting : 1;
+	u32 pad0 : 1;
 
 	u32 texMtxInfo_n_projection : 16; // Stored separately to guarantee that the texMtxInfo struct is 8 bits wide
 	struct {