Merge pull request #114 from delroth/mmio-optimization

MMIO reads JIT optimization

2 files changed, 74 insertions, 3 deletions

diff --git a/Source/Core/Common/x64ABI.cpp b/Source/Core/Common/x64ABI.cpp
index 6f423eaa6a..0116212485 100644
--- a/Source/Core/Common/x64ABI.cpp
+++ b/Source/Core/Common/x64ABI.cpp
@@ -176,6 +176,14 @@ void XEmitter::ABI_CallFunctionCC(void *func, u32 param1, u32 param2) {
 	ABI_RestoreStack(2 * 4);
 }
 
+void XEmitter::ABI_CallFunctionCP(void *func, u32 param1, void *param2) {
+	ABI_AlignStack(2 * 4);
+	PUSH(32, Imm32((u32)param2));
+	PUSH(32, Imm32(param1));
+	CALL(func);
+	ABI_RestoreStack(2 * 4);
+}
+
 void XEmitter::ABI_CallFunctionCCC(void *func, u32 param1, u32 param2, u32 param3) {
 	ABI_AlignStack(3 * 4);
 	PUSH(32, Imm32(param3));
@@ -204,6 +212,14 @@ void XEmitter::ABI_CallFunctionCCCP(void *func, u32 param1, u32 param2,u32 param
 	ABI_RestoreStack(4 * 4);
 }
 
+void XEmitter::ABI_CallFunctionPC(void *func, void *param1, u32 param2) {
+	ABI_AlignStack(3 * 4);
+	PUSH(32, Imm32(param2));
+	PUSH(32, Imm32((u32)param1));
+	CALL(func);
+	ABI_RestoreStack(3 * 4);
+}
+
 void XEmitter::ABI_CallFunctionPPC(void *func, void *param1, void *param2,u32 param3) {
 	ABI_AlignStack(3 * 4);
 	PUSH(32, Imm32(param3));
@@ -344,6 +360,22 @@ void XEmitter::ABI_CallFunctionCC(void *func, u32 param1, u32 param2) {
 	ABI_RestoreStack(0);
 }
 
+void XEmitter::ABI_CallFunctionCP(void *func, u32 param1, void *param2) {
+	ABI_AlignStack(0);
+	MOV(32, R(ABI_PARAM1), Imm32(param1));
+	MOV(64, R(ABI_PARAM2), Imm64((u64)param2));
+	u64 distance = u64(func) - (u64(code) + 5);
+	if (distance >= 0x0000000080000000ULL
+	 && distance <  0xFFFFFFFF80000000ULL) {
+		// Far call
+		MOV(64, R(RAX), Imm64((u64)func));
+		CALLptr(R(RAX));
+	} else {
+		CALL(func);
+	}
+	ABI_RestoreStack(0);
+}
+
 void XEmitter::ABI_CallFunctionCCC(void *func, u32 param1, u32 param2, u32 param3) {
 	ABI_AlignStack(0);
 	MOV(32, R(ABI_PARAM1), Imm32(param1));
@@ -396,6 +428,22 @@ void XEmitter::ABI_CallFunctionCCCP(void *func, u32 param1, u32 param2, u32 para
 	ABI_RestoreStack(0);
 }
 
+void XEmitter::ABI_CallFunctionPC(void *func, void *param1, u32 param2) {
+	ABI_AlignStack(0);
+	MOV(64, R(ABI_PARAM1), Imm64((u64)param1));
+	MOV(32, R(ABI_PARAM2), Imm32(param2));
+	u64 distance = u64(func) - (u64(code) + 5);
+	if (distance >= 0x0000000080000000ULL
+	 && distance <  0xFFFFFFFF80000000ULL) {
+		// Far call
+		MOV(64, R(RAX), Imm64((u64)func));
+		CALLptr(R(RAX));
+	} else {
+		CALL(func);
+	}
+	ABI_RestoreStack(0);
+}
+
 void XEmitter::ABI_CallFunctionPPC(void *func, void *param1, void *param2, u32 param3) {
 	ABI_AlignStack(0);
 	MOV(64, R(ABI_PARAM1), Imm64((u64)param1));
diff --git a/Source/Core/Common/x64Emitter.h b/Source/Core/Common/x64Emitter.h
index f8acb6fe95..b35fff3319 100644
--- a/Source/Core/Common/x64Emitter.h
+++ b/Source/Core/Common/x64Emitter.h
@@ -8,6 +8,7 @@
 
 #include <cstddef>
 #include <cstring>
+#include <functional>
 
 #include "Common/Common.h"
 #include "Common/MemoryUtil.h"
@@ -171,7 +172,7 @@ private:
 	u16 indexReg;
 };
 
-inline OpArg M(void *ptr)       {return OpArg((u64)ptr, (int)SCALE_RIP);}
+inline OpArg M(const void *ptr) {return OpArg((u64)ptr, (int)SCALE_RIP);}
 inline OpArg R(X64Reg value)    {return OpArg(0, SCALE_NONE, value);}
 inline OpArg MatR(X64Reg value) {return OpArg(0, SCALE_ATREG, value);}
 inline OpArg MDisp(X64Reg value, int offset) {
@@ -194,9 +195,9 @@ inline OpArg Imm16(u16 imm) {return OpArg(imm, SCALE_IMM16);} //rarely used
 inline OpArg Imm32(u32 imm) {return OpArg(imm, SCALE_IMM32);}
 inline OpArg Imm64(u64 imm) {return OpArg(imm, SCALE_IMM64);}
 #ifdef _ARCH_64
-inline OpArg ImmPtr(void* imm) {return Imm64((u64)imm);}
+inline OpArg ImmPtr(const void* imm) {return Imm64((u64)imm);}
 #else
-inline OpArg ImmPtr(void* imm) {return Imm32((u32)imm);}
+inline OpArg ImmPtr(const void* imm) {return Imm32((u32)imm);}
 #endif
 inline u32 PtrOffset(void* ptr, void* base) {
 #ifdef _ARCH_64 
@@ -671,9 +672,11 @@ public:
 	// These will destroy the 1 or 2 first "parameter regs".
 	void ABI_CallFunctionC(void *func, u32 param1);
 	void ABI_CallFunctionCC(void *func, u32 param1, u32 param2);
+	void ABI_CallFunctionCP(void *func, u32 param1, void *param2);
 	void ABI_CallFunctionCCC(void *func, u32 param1, u32 param2, u32 param3);
 	void ABI_CallFunctionCCP(void *func, u32 param1, u32 param2, void *param3);
 	void ABI_CallFunctionCCCP(void *func, u32 param1, u32 param2,u32 param3, void *param4);
+	void ABI_CallFunctionPC(void *func, void *param1, u32 param2);
 	void ABI_CallFunctionPPC(void *func, void *param1, void *param2,u32 param3);
 	void ABI_CallFunctionAC(void *func, const Gen::OpArg &arg1, u32 param2);
 	void ABI_CallFunctionA(void *func, const Gen::OpArg &arg1);
@@ -737,6 +740,26 @@ public:
 	#define DECLARE_IMPORT(x) extern "C" void *__imp_##x
 
 #endif
+
+	// Utility to generate a call to a std::function object.
+	//
+	// Unfortunately, calling operator() directly is undefined behavior in C++
+	// (this method might be a thunk in the case of multi-inheritance) so we
+	// have to go through a trampoline function.
+	template <typename T, typename... Args>
+	static void CallLambdaTrampoline(const std::function<T(Args...)>* f,
+	                                 Args... args)
+	{
+		(*f)(args...);
+	}
+
+	template <typename T, typename... Args>
+	void ABI_CallLambdaC(const std::function<T(Args...)>* f, u32 p1)
+	{
+		// Double casting is required by VC++ for some reason.
+		auto trampoline = (void(*)())&XEmitter::CallLambdaTrampoline<T, Args...>;
+		ABI_CallFunctionPC((void*)trampoline, const_cast<void*>((const void*)f), p1);
+	}
 };  // class XEmitter