rsx: Decompiler fixups and improvements

- Fix 2D coordinate sampling of W coordinate.
  W is actually HPOS.w and not 1. Z is however always 0.
- Optimize register usage a bit
  Disassembling compiled SPV shows that global declaration results in less ops than using inout modifiers. Modifiers generate extra mov instructions.

rpcs3/Emu/RSX/Common/FragmentProgramDecompiler.cpp

@@ -524,7 +524,7 @@ template<typename T> std::string FragmentProgramDecompiler::GetSRC(T src)
 		// NOTE: Hw testing showed the following:
 		// 1. Reading from registers 1 and 2 (COL0 and COL1) is clamped to (0, 1)
 		// 2. Reading from registers 4-12 (inclusive) is not clamped, but..
-		// 3. If the texcoord control mask is enabled, the last 2 values are always 0 and 1!
+		// 3. If the texcoord control mask is enabled, the last 2 values are always 0 and hpos.w!
 		const std::string reg_var = (dst.src_attr_reg_num < std::size(reg_table))? reg_table[dst.src_attr_reg_num] : "unk";
 		bool insert = true;
 
@@ -541,15 +541,30 @@ template<typename T> std::string FragmentProgramDecompiler::GetSRC(T src)
 		case 0x02:
 		{
 			// COL0, COL1
-			ret += "_saturate(" + reg_var + ")";
+			if (!src2.use_index_reg)
-			apply_precision_modifier = false;
+			{
+				ret += "_saturate(" + reg_var + ")";
+				apply_precision_modifier = false;
+			}
+			else
+			{
+				// Raw access
+				ret += reg_var;
+			}
 			break;
 		}
 		case 0x03:
 		{
 			// FOGC
-			// TODO: Confirm if precision modifiers affect this one
+			if (!src2.use_index_reg)
-			ret += reg_var;
+			{
+				ret += reg_var;
+			}
+			else
+			{
+				// Raw access
+				ret += "fog_c";
+			}
 			break;
 		}
 		case 0x4:
@@ -567,7 +582,8 @@ template<typename T> std::string FragmentProgramDecompiler::GetSRC(T src)
 			// Texcoord mask seems to reset the last 2 arguments to 0 and 1 if set
 			if (m_prog.texcoord_is_2d(dst.src_attr_reg_num - 4))
 			{
-				ret += getFloatTypeName(4) + "(" + reg_var + ".x, " + reg_var + ".y, 0., 1.)";
+				ret += getFloatTypeName(4) + "(" + reg_var + ".x, " + reg_var + ".y, 0., in_w)";
+				properties.has_w_access = true;
 			}
 			else
 			{

rpcs3/Emu/RSX/Common/FragmentProgramDecompiler.h

@@ -281,6 +281,7 @@ public:
 		bool has_tex_op = false;
 		bool has_divsq = false;
 		bool has_clamp = false;
+		bool has_w_access = false;
 	}
 	properties;
 

rpcs3/Emu/RSX/GL/GLFragmentProgram.cpp

@@ -202,34 +202,55 @@ void GLFragmentDecompilerThread::insertMainStart(std::stringstream & OS)
 	if (properties.in_register_mask & in_fogc)
 		glsl::insert_fog_declaration(OS);
 
-	const std::set<std::string> output_values =
+	std::set<std::string> output_registers;
+	if (m_ctrl & CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS)
 	{
-		"r0", "r1", "r2", "r3", "r4",
+		output_registers = { "r0", "r2", "r3", "r4" };
-		"h0", "h2", "h4", "h6", "h8"
+	}
-	};
+	else
-
-	std::string parameters;
-	const auto half4 = getHalfTypeName(4);
-	for (auto &reg_name : output_values)
 	{
-		const auto type = (reg_name[0] == 'r' || !device_props.has_native_half_support)? "vec4" : half4;
+		output_registers = { "h0", "h4", "h6", "h8" };
-		if (m_parr.HasParam(PF_PARAM_NONE, type, reg_name))
-		{
-			if (parameters.length())
-				parameters += ", ";
-
-			parameters += "inout " + type + " " + reg_name;
-		}
 	}
 
-	OS << "void fs_main(" << parameters << ")\n";
+	if (m_ctrl & CELL_GCM_SHADER_CONTROL_DEPTH_EXPORT)
+	{
+		output_registers.insert("r1");
+	}
+
+	std::string registers;
+	std::string reg_type;
+	const auto half4 = getHalfTypeName(4);
+	for (auto &reg_name : output_registers)
+	{
+		const auto type = (reg_name[0] == 'r' || !device_props.has_native_half_support)? "vec4" : half4;
+		if (LIKELY(reg_type == type))
+		{
+			registers += ", " + reg_name + " = " + type + "(0.)";
+		}
+		else
+		{
+			if (!registers.empty())
+				registers += ";\n";
+
+			registers += type + " " + reg_name + " = " + type + "(0.)";
+		}
+
+		reg_type = type;
+	}
+
+	if (!registers.empty())
+	{
+		OS << registers << ";\n";
+	}
+
+	OS << "void fs_main()\n";
 	OS << "{\n";
 
 	for (const ParamType& PT : m_parr.params[PF_PARAM_NONE])
 	{
-		for (const ParamItem& PI : PT.items)
+		for (const auto& PI : PT.items)
 		{
-			if (output_values.find(PI.name) != output_values.end())
+			if (output_registers.find(PI.name) != output_registers.end())
 				continue;
 
 			OS << "	" << PT.type << " " << PI.name;
@@ -240,11 +261,8 @@ void GLFragmentDecompilerThread::insertMainStart(std::stringstream & OS)
 		}
 	}
 
-	if (m_parr.HasParam(PF_PARAM_IN, "vec4", "ssa"))
+	if (properties.has_w_access)
-		OS << "	vec4 ssa = gl_FrontFacing ? vec4(1.) : vec4(-1.);\n";
+		OS << "	float in_w = (1. / gl_FragCoord.w);\n";
-
-	if (properties.in_register_mask & in_wpos)
-		OS << "	vec4 wpos = get_wpos();\n";
 
 	if (properties.in_register_mask & in_ssa)
 		OS << "	vec4 ssa = gl_FrontFacing ? vec4(1.) : vec4(-1.);\n";
@@ -267,34 +285,12 @@ void GLFragmentDecompilerThread::insertMainStart(std::stringstream & OS)
 
 void GLFragmentDecompilerThread::insertMainEnd(std::stringstream & OS)
 {
-	const std::set<std::string> output_values =
-	{
-		"r0", "r1", "r2", "r3", "r4",
-		"h0", "h2", "h4", "h6", "h8"
-	};
-
 	OS << "}\n\n";
 
 	OS << "void main()\n";
 	OS << "{\n";
 
-	std::string parameters;
+	OS << "\n" << "	fs_main();\n\n";
-	const auto half4 = getHalfTypeName(4);
-
-	for (auto &reg_name : output_values)
-	{
-		const std::string type = (reg_name[0] == 'r' || !device_props.has_native_half_support)? "vec4" : half4;
-		if (m_parr.HasParam(PF_PARAM_NONE, type, reg_name))
-		{
-			if (parameters.length())
-				parameters += ", ";
-
-			parameters += reg_name;
-			OS << "	" << type << " " << reg_name << " = " << type << "(0.);\n";
-		}
-	}
-
-	OS << "\n" << "	fs_main(" + parameters + ");\n\n";
 
 	glsl::insert_rop(
 		OS,

rpcs3/Emu/RSX/GL/GLVertexProgram.cpp

@@ -129,20 +129,37 @@ void GLVertexDecompilerThread::insertMainStart(std::stringstream & OS)
 	insert_glsl_legacy_function(OS, properties2);
 	glsl::insert_vertex_input_fetch(OS, glsl::glsl_rules_opengl4, dev_caps.vendor_INTEL == false);
 
-	std::string parameters;
+	// Declare global registers with optional initialization
-	for (int i = 0; i < 16; ++i)
+	std::string registers;
+	if (ParamType *vec4Types = m_parr.SearchParam(PF_PARAM_OUT, "vec4"))
 	{
-		std::string reg_name = "dst_reg" + std::to_string(i);
+		for (auto &PI : vec4Types->items)
-		if (m_parr.HasParam(PF_PARAM_OUT, "vec4", reg_name))
 		{
-			if (parameters.length())
+			if (registers.length())
-				parameters += ", ";
+				registers += ", ";
+			else
+				registers = "vec4 ";
 
-			parameters += "inout vec4 " + reg_name;
+			registers += PI.name;
+
+			if (!PI.value.empty())
+			{
+				printf("Value=%s\n", PI.value.c_str());
+				// Simplify default initialization
+				if (PI.value == "vec4(0.0, 0.0, 0.0, 0.0)")
+					registers += " = vec4(0.)";
+				else
+					registers += " = " + PI.value;
+			}
 		}
 	}
 
-	OS << "void vs_main(" << parameters << ")\n";
+	if (!registers.empty())
+	{
+		OS << registers << ";\n";
+	}
+
+	OS << "void vs_main()\n";
 	OS << "{\n";
 
 	//Declare temporary registers, ignoring those mapped to outputs
@@ -177,33 +194,7 @@ void GLVertexDecompilerThread::insertMainEnd(std::stringstream & OS)
 	OS << "void main ()\n";
 	OS << "{\n";
 
-	std::string parameters;
+	OS << "\n" << "	vs_main();\n\n";
-
-	if (ParamType *vec4Types = m_parr.SearchParam(PF_PARAM_OUT, "vec4"))
-	{
-		for (int i = 0; i < 16; ++i)
-		{
-			std::string reg_name = "dst_reg" + std::to_string(i);
-			for (auto &PI : vec4Types->items)
-			{
-				if (reg_name == PI.name)
-				{
-					if (parameters.length())
-						parameters += ", ";
-
-					parameters += reg_name;
-					OS << "	vec4 " << reg_name;
-
-					if (!PI.value.empty())
-						OS << "= " << PI.value;
-
-					OS << ";\n";
-				}
-			}
-		}
-	}
-
-	OS << "\n" << "	vs_main(" << parameters << ");\n\n";
 
 	for (auto &i : reg_table)
 	{

rpcs3/Emu/RSX/VK/VKFragmentProgram.cpp

@@ -246,34 +246,55 @@ void VKFragmentDecompilerThread::insertMainStart(std::stringstream & OS)
 	if (properties.in_register_mask & in_fogc)
 		glsl::insert_fog_declaration(OS);
 
-	const std::set<std::string> output_values =
+	std::set<std::string> output_registers;
+	if (m_ctrl & CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS)
 	{
-		"r0", "r1", "r2", "r3", "r4",
+		output_registers = { "r0", "r2", "r3", "r4" };
-		"h0", "h2", "h4", "h6", "h8"
+	}
-	};
+	else
-
-	std::string parameters;
-	const auto half4 = getHalfTypeName(4);
-	for (auto &reg_name : output_values)
 	{
-		const auto type = (reg_name[0] == 'r' || !device_props.has_native_half_support)? "vec4" : half4;
+		output_registers = { "h0", "h4", "h6", "h8" };
-		if (m_parr.HasParam(PF_PARAM_NONE, type, reg_name))
-		{
-			if (parameters.length())
-				parameters += ", ";
-
-			parameters += "inout " + type + " " + reg_name;
-		}
 	}
 
-	OS << "void fs_main(" << parameters << ")\n";
+	if (m_ctrl & CELL_GCM_SHADER_CONTROL_DEPTH_EXPORT)
+	{
+		output_registers.insert("r1");
+	}
+
+	std::string registers;
+	std::string reg_type;
+	const auto half4 = getHalfTypeName(4);
+	for (auto &reg_name : output_registers)
+	{
+		const auto type = (reg_name[0] == 'r' || !device_props.has_native_half_support)? "vec4" : half4;
+		if (LIKELY(reg_type == type))
+		{
+			registers += ", " + reg_name + " = " + type + "(0.)";
+		}
+		else
+		{
+			if (!registers.empty())
+				registers += ";\n";
+
+			registers += type + " " + reg_name + " = " + type + "(0.)";
+		}
+
+		reg_type = type;
+	}
+
+	if (!registers.empty())
+	{
+		OS << registers << ";\n";
+	}
+
+	OS << "void fs_main()\n";
 	OS << "{\n";
 
 	for (const ParamType& PT : m_parr.params[PF_PARAM_NONE])
 	{
 		for (const ParamItem& PI : PT.items)
 		{
-			if (output_values.find(PI.name) != output_values.end())
+			if (output_registers.find(PI.name) != output_registers.end())
 				continue;
 
 			OS << "	" << PT.type << " " << PI.name;
@@ -284,6 +305,9 @@ void VKFragmentDecompilerThread::insertMainStart(std::stringstream & OS)
 		}
 	}
 
+	if (properties.has_w_access)
+		OS << "	float in_w = (1. / gl_FragCoord.w);\n";
+
 	if (properties.in_register_mask & in_ssa)
 		OS << "	vec4 ssa = gl_FrontFacing ? vec4(1.) : vec4(-1.);\n";
 
@@ -305,34 +329,12 @@ void VKFragmentDecompilerThread::insertMainStart(std::stringstream & OS)
 
 void VKFragmentDecompilerThread::insertMainEnd(std::stringstream & OS)
 {
-	const std::set<std::string> output_values =
-	{
-		"r0", "r1", "r2", "r3", "r4",
-		"h0", "h2", "h4", "h6", "h8"
-	};
-
 	OS << "}\n\n";
 
 	OS << "void main()\n";
 	OS << "{\n";
 
-	std::string parameters;
+	OS << "\n" << "	fs_main();\n\n";
-	const auto half4 = getHalfTypeName(4);
-
-	for (auto &reg_name : output_values)
-	{
-		const std::string type = (reg_name[0] == 'r' || !device_props.has_native_half_support)? "vec4" : half4;
-		if (m_parr.HasParam(PF_PARAM_NONE, type, reg_name))
-		{
-			if (parameters.length())
-				parameters += ", ";
-
-			parameters += reg_name;
-			OS << "	" << type << " " << reg_name << " = " << type << "(0.);\n";
-		}
-	}
-
-	OS << "\n" << "	fs_main(" + parameters + ");\n\n";
 
 	glsl::insert_rop(
 		OS,

rpcs3/Emu/RSX/VK/VKVertexProgram.cpp

@@ -177,20 +177,36 @@ void VKVertexDecompilerThread::insertMainStart(std::stringstream & OS)
 	glsl::insert_glsl_legacy_function(OS, properties2);
 	glsl::insert_vertex_input_fetch(OS, glsl::glsl_rules_spirv);
 
-	std::string parameters;
+	// Declare global registers with optional initialization
-	for (int i = 0; i < 16; ++i)
+	std::string registers;
+	if (ParamType *vec4Types = m_parr.SearchParam(PF_PARAM_OUT, "vec4"))
 	{
-		std::string reg_name = "dst_reg" + std::to_string(i);
+		for (auto &PI : vec4Types->items)
-		if (m_parr.HasParam(PF_PARAM_OUT, "vec4", reg_name))
 		{
-			if (parameters.length())
+			if (registers.length())
-				parameters += ", ";
+				registers += ", ";
+			else
+				registers = "vec4 ";
 
-			parameters += "inout vec4 " + reg_name;
+			registers += PI.name;
+
+			if (!PI.value.empty())
+			{
+				// Simplify default initialization
+				if (PI.value == "vec4(0.0, 0.0, 0.0, 0.0)")
+					registers += " = vec4(0.)";
+				else
+					registers += " = " + PI.value;
+			}
 		}
 	}
 
-	OS << "void vs_main(" << parameters << ")\n";
+	if (!registers.empty())
+	{
+		OS << registers << ";\n";
+	}
+
+	OS << "void vs_main()\n";
 	OS << "{\n";
 
 	//Declare temporary registers, ignoring those mapped to outputs
@@ -225,33 +241,7 @@ void VKVertexDecompilerThread::insertMainEnd(std::stringstream & OS)
 	OS << "void main ()\n";
 	OS << "{\n";
 
-	std::string parameters;
+	OS << "\n" << "	vs_main();\n\n";
-
-	if (ParamType *vec4Types = m_parr.SearchParam(PF_PARAM_OUT, "vec4"))
-	{
-		for (int i = 0; i < 16; ++i)
-		{
-			std::string reg_name = "dst_reg" + std::to_string(i);
-			for (auto &PI : vec4Types->items)
-			{
-				if (reg_name == PI.name)
-				{
-					if (parameters.length())
-						parameters += ", ";
-
-					parameters += reg_name;
-					OS << "	vec4 " << reg_name;
-
-					if (!PI.value.empty())
-						OS << "= " << PI.value;
-
-					OS << ";\n";
-				}
-			}
-		}
-	}
-
-	OS << "\n" << "	vs_main(" << parameters << ");\n\n";
 
 	for (auto &i : reg_table)
 	{