rsx: Reimplement conditional execution in shaders

- Per-channel conditional execution introduces RAW hazards all over the place
- Its cheaper to process both branches and select between the two
- Also improves ShaderVariable functionality to allow functionality such as match_size and taking complex variables as inputs

rpcs3/Emu/RSX/Common/FragmentProgramDecompiler.cpp

@@ -1,4 +1,4 @@
-#include "stdafx.h"
+#include "stdafx.h"
 #include "Emu/Memory/vm.h"
 #include "Emu/System.h"
 
@@ -372,25 +372,9 @@ void FragmentProgramDecompiler::AddCodeCond(const std::string& dst, const std::s
 		return;
 	}
 
-	static const char f[4] = { 'x', 'y', 'z', 'w' };
-	std::string cond = GetRawCond();
-
-	ShaderVariable dst_var(dst);
-	dst_var.simplify();
-
-	//const char *c_mask = f;
-
-	if (dst_var.swizzles[0].length() == 1)
-	{
-		AddCode("if (" + cond + ".x) " + dst + " = " + src + ";");
-	}
-	else
-	{
-		for (int i = 0; i < dst_var.swizzles[0].length(); ++i)
-		{
-			AddCode("if (" + cond + "." + f[i] + ") " + dst + "." + f[i] + " = " + src + "." + f[i] + ";");
-		}
-	}
+	// NOTE: dst = _select(dst, src, cond) is equivalent to dst = cond? src : dst;
+	const auto cond = ShaderVariable(dst).match_size(GetRawCond());
+	AddCode(dst + " = _select(" + dst + ", " + src + ", " + cond + ");");
 }
 
 template<typename T> std::string FragmentProgramDecompiler::GetSRC(T src)

rpcs3/Emu/RSX/Common/GLSLCommon.h

@@ -415,6 +415,8 @@ namespace glsl
 
 	static void insert_glsl_legacy_function(std::ostream& OS, glsl::program_domain domain, bool require_lit_emulation, bool require_depth_conversion = false, bool require_wpos = false, bool require_texture_ops = true)
 	{
+		OS << "#define _select mix\n\n";
+
 		if (require_lit_emulation)
 		{
 			OS <<

rpcs3/Emu/RSX/Common/ShaderParam.h

@@ -1,4 +1,4 @@
-#pragma once
+#pragma once
 
 #include <string>
 #include <vector>
@@ -179,22 +179,30 @@ public:
 	{
 		// Separate 'double destination' variables 'X=Y=SRC'
 		std::string simple_var;
-		const auto pos = var.find('=');
+		const auto eq_pos = var.find('=');
 
-		if (pos != std::string::npos)
+		if (eq_pos != std::string::npos)
 		{
-			simple_var = var.substr(0, pos - 1);
+			simple_var = var.substr(0, eq_pos - 1);
 		}
 		else
 		{
 			simple_var = var;
 		}
 
+		const auto brace_pos = var.find_last_of(")");
+		std::string prefix;
+		if (brace_pos != std::string::npos)
+		{
+			prefix = simple_var.substr(0, brace_pos);
+			simple_var = simple_var.substr(brace_pos);
+		}
+
 		auto var_blocks = fmt::split(simple_var, { "." });
 
 		verify(HERE), (var_blocks.size() != 0);
 
-		name = var_blocks[0];
+		name = prefix + var_blocks[0];
 
 		if (var_blocks.size() == 1)
 		{
@@ -263,6 +271,45 @@ public:
 
 		return name + "." + fmt::merge({ swizzles }, ".");
 	}
+
+	std::string match_size(const std::string& other_var) const
+	{
+		// Make other_var the same vector length as this var
+		ShaderVariable other(other_var);
+		const auto this_size = get_vector_size();
+		const auto other_size = other.get_vector_size();
+
+		if (LIKELY(this_size == other_size))
+		{
+			return other_var;
+		}
+
+		if (LIKELY(this_size < other_size))
+		{
+			switch (this_size)
+			{
+			case 0:
+			case 4:
+				return other_var;
+			case 1:
+				return other_var + ".x";
+			case 2:
+				return other_var + ".xy";
+			case 3:
+				return other_var + ".xyz";
+			default:
+				fmt::throw_exception("Unreachable" HERE);
+			}
+		}
+		else
+		{
+			auto remaining = this_size - other_size;
+			std::string result = other_var;
+
+			while (remaining--) result += "x";
+			return result;
+		}
+	}
 };
 
 struct vertex_reg_info

rpcs3/Emu/RSX/Common/VertexProgramDecompiler.cpp

@@ -1,4 +1,4 @@
-#include "stdafx.h"
+#include "stdafx.h"
 #include "Emu/System.h"
 
 #include "VertexProgramDecompiler.h"
@@ -240,18 +240,8 @@ std::string VertexProgramDecompiler::Format(const std::string& code)
 	return fmt::replace_all(code, repl_list);
 }
 
-std::string VertexProgramDecompiler::GetCond()
+std::string VertexProgramDecompiler::GetRawCond()
 {
-	enum
-	{
-		lt = 0x1,
-		eq = 0x2,
-		gt = 0x4,
-	};
-
-	if (d0.cond == 0) return "false";
-	if (d0.cond == (lt | gt | eq)) return "true";
-
 	static const COMPARE cond_string_table[(lt | gt | eq) + 1] =
 	{
 		COMPARE::FUNCTION_SLT, // "error"
@@ -272,7 +262,15 @@ std::string VertexProgramDecompiler::GetCond()
 	swizzle += f[d0.mask_w];
 
 	swizzle = swizzle == "xyzw" ? "" : "." + swizzle;
-	return "any(" + compareFunction(cond_string_table[d0.cond], AddCondReg() + swizzle, getFloatTypeName(4) + "(0., 0., 0., 0.)" + swizzle) + ")";
+	return compareFunction(cond_string_table[d0.cond], AddCondReg() + swizzle, getFloatTypeName(4) + "(0., 0., 0., 0.)" + swizzle);
+}
+
+std::string VertexProgramDecompiler::GetCond()
+{
+	if (d0.cond == 0) return "false";
+	if (d0.cond == (lt | gt | eq)) return "true";
+
+	return "any(" + GetRawCond() + ")";
 }
 
 std::string VertexProgramDecompiler::GetOptionalBranchCond()
@@ -292,7 +290,6 @@ void VertexProgramDecompiler::AddCodeCond(const std::string& dst, const std::str
 		gt = 0x4,
 	};
 
-
 	if (!d0.cond_test_enable || d0.cond == (lt | gt | eq))
 	{
 		AddCode(dst + " = " + src + ";");
@@ -305,58 +302,9 @@ void VertexProgramDecompiler::AddCodeCond(const std::string& dst, const std::str
 		return;
 	}
 
-	static const COMPARE cond_string_table[(lt | gt | eq) + 1] =
-	{
-		COMPARE::FUNCTION_SLT, // "error"
-		COMPARE::FUNCTION_SLT,
-		COMPARE::FUNCTION_SEQ,
-		COMPARE::FUNCTION_SLE,
-		COMPARE::FUNCTION_SGT,
-		COMPARE::FUNCTION_SNE,
-		COMPARE::FUNCTION_SGE,
-	};
-
-	ShaderVariable dst_var(dst);
-	dst_var.simplify();
-
-	static const char f[4] = { 'x', 'y', 'z', 'w' };
-	const u32 mask_index[4] = { d0.mask_x, d0.mask_y, d0.mask_z, d0.mask_w };
-
-	auto get_masked_dst = [](const std::string& dest, const char mask)
-	{
-		const auto selector = std::string(".") + mask;
-		const auto pos = dest.find('=');
-
-		std::string result = dest + selector;
-
-		if (pos != std::string::npos)
-		{
-			result.insert(pos - 1, selector);
-		}
-
-		return result;
-	};
-
-	auto get_cond_func = [this, &mask_index](COMPARE op, int index)
-	{
-		// Condition reg check for single element (x,y,z,w)
-		const auto cond_mask = f[mask_index[index]];
-		return compareFunction(op, AddCondReg() + "." + cond_mask, "0.", true);
-	};
-
-	if (dst_var.swizzles[0].length() == 1)
-	{
-		const std::string cond = get_cond_func(cond_string_table[d0.cond], 0);
-		AddCode("if (" + cond + ") " + dst + " = " + src + ";");
-	}
-	else
-	{
-		for (int i = 0; i < dst_var.swizzles[0].length(); ++i)
-		{
-			const std::string cond = get_cond_func(cond_string_table[d0.cond], i);
-			AddCode("if (" + cond + ") " + get_masked_dst(dst, f[i]) + " = " + src + "." + f[i] + ";");
-		}
-	}
+	// NOTE: dst = _select(dst, src, cond) is equivalent to dst = cond? src : dst;
+	const auto cond = ShaderVariable(dst).match_size(GetRawCond());
+	AddCode(dst + " = _select(" + dst + ", " + src + ", " + cond + ");");
 }
 
 std::string VertexProgramDecompiler::AddAddrReg()

rpcs3/Emu/RSX/Common/VertexProgramDecompiler.h

@@ -1,4 +1,4 @@
-#pragma once
+#pragma once
 #include "Emu/RSX/RSXVertexProgram.h"
 #include <vector>
 #include <set>
@@ -27,6 +27,13 @@ struct VertexProgramDecompiler
 	D3 d3;
 	SRC src[3];
 
+	enum
+	{
+		lt = 0x1,
+		eq = 0x2,
+		gt = 0x4,
+	};
+
 	struct FuncInfo
 	{
 		u32 offset;
@@ -66,9 +73,9 @@ struct VertexProgramDecompiler
 	std::string GetDST(bool is_sca = false);
 	std::string GetSRC(const u32 n);
 	std::string GetTex();
+	std::string GetRawCond();
 	std::string GetCond();
 	std::string GetOptionalBranchCond();	//Conditional branch expression modified externally at runtime
-	std::string AddAddrMask();
 	std::string AddAddrReg();
 	std::string AddCondReg();
 	u32 GetAddr();

rpcs3/Emu/RSX/D3D12/D3D12CommonDecompiler.cpp

@@ -1,4 +1,4 @@
-#ifdef _MSC_VER
+#ifdef _MSC_VER
 #include "stdafx.h"
 #include "stdafx_d3d12.h"
 #include "D3D12CommonDecompiler.h"
@@ -114,6 +114,8 @@ std::string compareFunctionImp(COMPARE f, const std::string &Op0, const std::str
 
 void insert_d3d12_legacy_function(std::ostream& OS, bool is_fragment_program)
 {
+	OS << "#define _select lerp\n\n";
+
 	OS << "float4 lit_legacy(float4 val)";
 	OS << "{\n";
 	OS << "	float4 clamped_val = val;\n";