rsx/fp: Detect broken/NOP shaders automatically

- Do not compile body if the shader is of no consequence, leave as a passthrough shader

rpcs3/Emu/RSX/Common/FragmentProgramDecompiler.cpp

@@ -482,6 +482,59 @@ template<typename T> std::string FragmentProgramDecompiler::GetSRC(T src)
 
 std::string FragmentProgramDecompiler::BuildCode()
 {
+	// Shader validation
+	// Shader must at least write to one output for the body to be considered valid
+
+	const std::string vec4_type = getFloatTypeName(4);
+	const std::string init_value = vec4_type + "(0., 0., 0., 0.)";
+	std::array<std::string, 4> output_register_names;
+	std::array<u32, 4> ouput_register_indices = { 0, 2, 3, 4 };
+	bool shader_is_valid = false;
+
+	// Check depth export
+	if (m_ctrl & CELL_GCM_SHADER_CONTROL_DEPTH_EXPORT)
+	{
+		if (shader_is_valid = !!temp_registers[1].h0_writes; !shader_is_valid)
+		{
+			LOG_WARNING(RSX, "Fragment shader fails to write the depth value!");
+		}
+	}
+
+	// Add the color output registers. They are statically written to and have guaranteed initialization (except r1.z which == wpos.z)
+	// This can be used instead of an explicit clear pass in some games (Motorstorm)
+	if (m_ctrl & CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS)
+	{
+		output_register_names = { "r0", "r2", "r3", "r4" };
+	}
+	else
+	{
+		output_register_names = { "h0", "h4", "h6", "h8" };
+	}
+
+	for (int n = 0; n < 4; ++n)
+	{
+		if (!m_parr.HasParam(PF_PARAM_NONE, vec4_type, output_register_names[n]))
+		{
+			m_parr.AddParam(PF_PARAM_NONE, vec4_type, output_register_names[n], init_value);
+			continue;
+		}
+
+		const auto block_index = ouput_register_indices[n];
+		shader_is_valid |= (!!temp_registers[block_index].h0_writes);
+	}
+
+	if (!shader_is_valid)
+	{
+		properties.has_no_output = true;
+
+		if (!properties.has_discard_op)
+		{
+			// NOTE: Discard operation overrides output
+			LOG_WARNING(RSX, "Shader does not write to any output register and will be NOPed");
+			main = "/*" + main + "*/";
+		}
+	}
+
 	std::stringstream OS;
 	insertHeader(OS);
 	OS << "\n";
@@ -765,23 +818,6 @@ std::string FragmentProgramDecompiler::Decompile()
 
 	int forced_unit = FORCE_NONE;
 
-	//Add the output registers. They are statically written to and have guaranteed initialization (except r1.z which == wpos.z)
-	//This can be used instead of an explicit clear pass in some games (Motorstorm)
-	if (m_ctrl & CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS)
-	{
-		AddReg(0, CELL_GCM_FALSE);
-		AddReg(2, CELL_GCM_FALSE);
-		AddReg(3, CELL_GCM_FALSE);
-		AddReg(4, CELL_GCM_FALSE);
-	}
-	else
-	{
-		AddReg(0, CELL_GCM_TRUE);
-		AddReg(4, CELL_GCM_TRUE);
-		AddReg(6, CELL_GCM_TRUE);
-		AddReg(8, CELL_GCM_TRUE);
-	}
-
 	while (true)
 	{
 		for (auto found = std::find(m_end_offsets.begin(), m_end_offsets.end(), m_size);
@@ -888,7 +924,10 @@ std::string FragmentProgramDecompiler::Decompile()
 		switch (opcode)
 		{
 		case RSX_FP_OPCODE_NOP: break;
-		case RSX_FP_OPCODE_KIL: AddFlowOp("discard"); break;
+		case RSX_FP_OPCODE_KIL:
+			properties.has_discard_op = true;
+			AddFlowOp("discard");
+			break;
 
 		default:
 			int prev_force_unit = forced_unit;

rpcs3/Emu/RSX/Common/FragmentProgramDecompiler.h

@@ -1,4 +1,4 @@
-#pragma once
+#pragma once
 #include "ShaderParam.h"
 #include "Emu/RSX/RSXFragmentProgram.h"
 #include <sstream>
@@ -28,6 +28,9 @@ class FragmentProgramDecompiler
 
 		u32 real_index = UINT32_MAX;
 
+		u32 h0_writes = 0u;   // Number of writes to the first 64-bits of the register
+		u32 h1_writes = 0u;   // Number of writes to the last 64-bits of the register
+
 		void tag(u32 index, bool half_register, bool x, bool y, bool z, bool w)
 		{
 			if (half_register)
@@ -40,6 +43,7 @@ class FragmentProgramDecompiler
 					if (w) last_write_half[3] = true;
 
 					aliased_h1 = true;
+					h1_writes++;
 				}
 				else
 				{
@@ -49,6 +53,7 @@ class FragmentProgramDecompiler
 					if (w) last_write_half[1] = true;
 
 					aliased_h0 = true;
+					h0_writes++;
 				}
 			}
 			else
@@ -59,6 +64,9 @@ class FragmentProgramDecompiler
 				if (w) last_write_half[3] = false;
 
 				aliased_r0 = true;
+
+				h0_writes++;
+				h1_writes++;
 			}
 
 			if (real_index == UINT32_MAX)
@@ -247,6 +255,7 @@ public:
 		bool has_gather_op = false;
 		bool has_wpos_input = false;
 		bool has_no_output = false;
+		bool has_discard_op = false;
 	}
 	properties;
 

rpcs3/Emu/RSX/Common/ShaderParam.h

@@ -67,8 +67,8 @@ enum ParamFlag
 
 struct ParamItem
 {
-	std::string name;
-	std::string value;
+	const std::string name;
+	const std::string value;
 	int location;
 
 	ParamItem(const std::string& _name, int _location, const std::string& _value = "")
@@ -81,7 +81,7 @@ struct ParamItem
 struct ParamType
 {
 	const ParamFlag flag;
-	std::string type;
+	const std::string type;
 	std::vector<ParamItem> items;
 
 	ParamType(const ParamFlag _flag, const std::string& _type)
@@ -127,13 +127,13 @@ struct ParamArray
 		return false;
 	}
 
-	bool HasParam(const ParamFlag flag, std::string type, const std::string& name)
+	bool HasParam(const ParamFlag flag, const std::string& type, const std::string& name)
 	{
 		ParamType* t = SearchParam(flag, type);
 		return t && t->SearchName(name);
 	}
 
-	std::string AddParam(const ParamFlag flag, std::string type, const std::string& name, const std::string& value)
+	std::string AddParam(const ParamFlag flag, const std::string& type, const std::string& name, const std::string& value)
 	{
 		ParamType* t = SearchParam(flag, type);
 
@@ -150,7 +150,7 @@ struct ParamArray
 		return name;
 	}
 
-	std::string AddParam(const ParamFlag flag, std::string type, const std::string& name, int location = -1)
+	std::string AddParam(const ParamFlag flag, const std::string& type, const std::string& name, int location = -1)
 	{
 		ParamType* t = SearchParam(flag, type);