[Shader JIT] Moar

CMakeLists.txt

@@ -94,7 +94,7 @@ set(SERVICE_SOURCE_FILES src/core/services/service_manager.cpp src/core/services
 )
 set(PICA_SOURCE_FILES src/core/PICA/gpu.cpp src/core/PICA/regs.cpp src/core/PICA/shader_unit.cpp
                       src/core/PICA/shader_interpreter.cpp src/core/PICA/dynapica/shader_rec.cpp
-                      src/core/PICA/dynapica/shader_rec_instructions_x64.cpp
+                      src/core/PICA/dynapica/shader_rec_emitter_x64.cpp 
 )
 
 set(RENDERER_GL_SOURCE_FILES src/core/renderer_gl/renderer_gl.cpp src/core/renderer_gl/textures.cpp src/core/renderer_gl/etc1.cpp)
@@ -124,6 +124,7 @@ set(HEADER_FILES include/emulator.hpp include/helpers.hpp include/opengl.hpp inc
                  include/services/ldr_ro.hpp include/ipc.hpp include/services/act.hpp include/services/nfc.hpp
                  include/system_models.hpp include/services/dlp_srvr.hpp include/PICA/dynapica/pica_recs.hpp
                  include/PICA/dynapica/x64_regs.hpp include/PICA/dynapica/vertex_loader_rec.hpp include/PICA/dynapica/shader_rec.hpp
+                 include/PICA/dynapica/shader_rec_emitter_x64.hpp
 )
 
 set(THIRD_PARTY_SOURCE_FILES third_party/imgui/imgui.cpp

include/PICA/dynapica/shader_rec.hpp

@@ -7,8 +7,7 @@
 #include <unordered_map>
 
 #ifdef PANDA3DS_X64_HOST
-#include "xbyak/xbyak.h"
-using ShaderEmitter = Xbyak::CodeGenerator;
+#include "shader_rec_emitter_x64.hpp"
 #endif
 #endif
 
@@ -16,6 +15,7 @@ class ShaderJIT {
 #ifdef PANDA3DS_SHADER_JIT_SUPPORTED
 	using Hash = PICAShader::Hash;
 	using ShaderCache = std::unordered_map<Hash, std::unique_ptr<ShaderEmitter>>;
+	ShaderEmitter::Callback activeShaderCallback;
 
 	ShaderCache cache;
 	void compileShader(PICAShader& shaderUnit);
@@ -35,8 +35,13 @@ public:
 		Helpers::panic("Vertex Loader JIT: Tried to load vertices with JIT on platform that does not support vertex loader jit");
 	}
 
+	// Define dummy callback. This should never be called if the shader JIT is not supported
+	using Callback = void(*)(PICAShader& shaderUnit);
+	Callback activeShaderCallback = nullptr;
+
 	void reset() {}
 	static constexpr bool isAvailable() { return false; }
 #endif
 
+	auto getCallback() { return activeShaderCallback; }
 };

include/PICA/dynapica/shader_rec_emitter_x64.hpp

@@ -0,0 +1,23 @@
+#pragma once
+
+// Only do anything if we're on an x64 target with JIT support enabled
+#if defined(PANDA3DS_DYNAPICA_SUPPORTED) && defined(PANDA3DS_X64_HOST)
+#include "helpers.hpp"
+#include "PICA/shader.hpp"
+#include "xbyak/xbyak.h"
+#include "x64_regs.hpp"
+
+class ShaderEmitter : public Xbyak::CodeGenerator {
+	static constexpr size_t executableMemorySize = PICAShader::maxInstructionCount * 96; // How much executable memory to alloc for each shader
+	// Allocate some extra space as padding for security purposes in the extremely unlikely occasion we manage to overflow the above size
+	static constexpr size_t allocSize = executableMemorySize + 0x1000;
+
+public:
+	using Callback = void(*)(const PICAShader& shaderUnit);
+
+	// Initialize our emitter with "allocSize" bytes of RWX memory
+	ShaderEmitter() : Xbyak::CodeGenerator(allocSize) {}
+	void compile(const PICAShader& shaderUnit);
+};
+
+#endif // x64 recompiler check

include/PICA/shader.hpp

@@ -100,7 +100,9 @@ protected:
 	bool codeHashDirty = false;
 	bool opdescHashDirty = false;
 
+	// Add these as friend classes for the JIT so it has access to all important state
 	friend class ShaderJIT;
+	friend class ShaderEmitter;
 
 private:
 	vec4f getSource(u32 source);
@@ -184,8 +186,9 @@ private:
 	bool isCondTrue(u32 instruction);
 
 public:
-	std::array<u32, 4096> loadedShader; // Currently loaded & active shader
-	std::array<u32, 4096> bufferedShader; // Shader to be transferred when the SH_CODETRANSFER_END reg gets written to
+	static constexpr size_t maxInstructionCount = 4096;
+	std::array<u32, maxInstructionCount> loadedShader; // Currently loaded & active shader
+	std::array<u32, maxInstructionCount> bufferedShader; // Shader to be transferred when the SH_CODETRANSFER_END reg gets written to
 
 	u32 entrypoint = 0; // Initial shader PC
 	u32 boolUniform;

src/core/PICA/dynapica/shader_rec.cpp

@@ -7,6 +7,7 @@ void ShaderJIT::reset() {
 }
 
 void ShaderJIT::prepare(PICAShader& shaderUnit) {
+	shaderUnit.pc = shaderUnit.entrypoint;
 	// We construct a shader hash from both the code and operand descriptor hashes
 	// This is so that if only one of them changes, we still properly recompile the shader
 	// This code is inspired from how Citra solves this problem
@@ -15,11 +16,10 @@ void ShaderJIT::prepare(PICAShader& shaderUnit) {
 
 	if (it == cache.end()) { // Block has not been compiled yet
 		auto emitter = std::make_unique<ShaderEmitter>();
+		emitter->compile(shaderUnit);
 		cache.emplace_hint(it, hash, std::move(emitter));
 	} else { // Block has been compiled and found, use it
-
+		auto emitter = it->second.get();
 	}
-
-	shaderUnit.pc = shaderUnit.entrypoint;
 }
 #endif // PANDA3DS_SHADER_JIT_SUPPORTED

src/core/PICA/dynapica/shader_rec_emitter_x64.cpp

@@ -0,0 +1,11 @@
+#if defined(PANDA3DS_DYNAPICA_SUPPORTED) && defined(PANDA3DS_X64_HOST)
+#include "PICA/dynapica/shader_rec_emitter_x64.hpp"
+
+using namespace Xbyak;
+using namespace Xbyak::util;
+
+void ShaderEmitter::compile(const PICAShader& shaderUnit) {
+
+}
+
+#endif