Optimize shader uid checks by checking the number of uid values which are actually used.

5 files changed, 42 insertions, 27 deletions

diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
index dce8bfbe96..dd284ced7d 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
@@ -535,6 +535,11 @@ static void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE Api
 	for (unsigned int i = 0; i < numStages; i++)
 		WriteStage<T>(out, uid_data, i, ApiType, RegisterStates); // build the equation for this stage
 
+#define MY_STRUCT_OFFSET(str,elem) ((u32)((u64)&(str).elem-(u64)&(str)))
+	bool enable_pl = g_ActiveConfig.bEnablePixelLighting && g_ActiveConfig.backend_info.bSupportsPixelLighting;
+	uid_data.num_values = (enable_pl) ? sizeof(uid_data)/sizeof(32) : MY_STRUCT_OFFSET(uid_data,stagehash[numStages])/sizeof(u32);
+
+
 	if (numStages)
 	{
 		// The results of the last texenv stage are put onto the screen,
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.h b/Source/Core/VideoCommon/Src/PixelShaderGen.h
index 01b0ad3489..210dd9dd32 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.h
@@ -61,6 +61,9 @@ struct pixel_shader_uid_data
 {
 	// TODO: Optimize field order for easy access!
 
+	u32 num_values; // TODO: Shouldn't be a u32
+	u32 NumValues() const { return num_values; } // TODO: Can be optimized :)
+
 	u32 components;
 	u32 dstAlphaMode : 2;
 	u32 Pretest : 2;
@@ -96,6 +99,23 @@ 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 xfregs_numTexGen_numTexGens : 4;
+
 	struct {
 		u32 cc : 24;
 		u32 ac : 24;
@@ -106,41 +126,24 @@ struct pixel_shader_uid_data
 		u32 tevorders_colorchan : 3;
 		u32 pad1 : 6;
 
+		// TODO: Clean up the swapXY mess
 		u32 hasindstage : 1;
 		u32 tevind : 21;
-		u32 tevksel_swap1a : 2; // TODO: Doesn't fit here..
-		u32 tevksel_swap2a : 2; // TODO: Doesn't fit here..
-		u32 tevksel_swap1b : 2; // TODO: Doesn't fit here..
-		u32 tevksel_swap2b : 2; // TODO: Doesn't fit here..
+		u32 tevksel_swap1a : 2;
+		u32 tevksel_swap2a : 2;
+		u32 tevksel_swap1b : 2;
+		u32 tevksel_swap2b : 2;
 		u32 pad2 : 2;
 
-		u32 tevksel_swap1c : 2; // TODO: Doesn't fit here..
-		u32 tevksel_swap2c : 2; // TODO: Doesn't fit here..
-		u32 tevksel_swap1d : 2; // TODO: Doesn't fit here..
-		u32 tevksel_swap2d : 2; // TODO: Doesn't fit here..
+		u32 tevksel_swap1c : 2;
+		u32 tevksel_swap2c : 2;
+		u32 tevksel_swap1d : 2;
+		u32 tevksel_swap2d : 2;
 		u32 tevksel_kc : 5;
 		u32 tevksel_ka : 5;
 		u32 pad3 : 14;
 	} stagehash[16];
 
-
-	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 xfregs_numTexGen_numTexGens : 4;
-
 	// TODO: I think we're fine without an enablePixelLighting field, should probably double check, though..
 	LightingUidData lighting;
 };
diff --git a/Source/Core/VideoCommon/Src/ShaderGenCommon.h b/Source/Core/VideoCommon/Src/ShaderGenCommon.h
index b8820e2e98..53f2405f8a 100644
--- a/Source/Core/VideoCommon/Src/ShaderGenCommon.h
+++ b/Source/Core/VideoCommon/Src/ShaderGenCommon.h
@@ -100,7 +100,7 @@ public:
 	bool operator < (const ShaderUid& obj) const
 	{
 		// TODO: Store last frame used and order by that? makes much more sense anyway...
-		for (unsigned int i = 0; i < sizeof(uid_data) / sizeof(u32); ++i)
+		for (unsigned int i = 0; i < data.NumValues(); ++i)
 		{
 			if (this->values[i] < obj.values[i])
 				return true;
@@ -212,6 +212,8 @@ struct LightingUidData
 	u32 diffusefunc : 8; // 4x2 bits
 	u32 attnfunc : 8; // 4x2 bits
 	u32 light_mask : 32; // 4x8 bits
+
+	u32 NumValues() const { return sizeof(LightingUidData) / sizeof(u32); }
 };
 #pragma pack()
 
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
index 3edb51c505..8c157dc030 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
@@ -74,6 +74,8 @@ static void GenerateVertexShader(T& out, u32 components, API_TYPE api_type)
 	vertex_shader_uid_data& uid_data = (&out.template GetUidData<vertex_shader_uid_data>() != NULL)
 											? out.template GetUidData<vertex_shader_uid_data>() : dummy_data;
 
+	uid_data.num_values = sizeof(uid_data)/sizeof(u32);
+
 	out.SetBuffer(text);
 #ifndef ANDROID
 	locale_t locale;
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.h b/Source/Core/VideoCommon/Src/VertexShaderGen.h
index eb7236678a..3f3087783d 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.h
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.h
@@ -67,6 +67,9 @@ const s_svar VSVar_Loc[] = {  {I_POSNORMALMATRIX, C_POSNORMALMATRIX, 6 },
 
 struct vertex_shader_uid_data
 {
+	u32 num_values; // TODO: Shouldn't be a u32
+
+	u32 NumValues() const { return num_values; }
 	u32 components;
 	u32 numColorChans : 2;
 	u32 numTexGens : 4;