Optimize vertex loader with a mini JIT (only first step, more optimizations may follow). Some various error message and warning fixes.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@1276 8ced0084-cf51-0410-be5f-012b33b47a6e

4 files changed, 41 insertions, 49 deletions

diff --git a/Source/Core/Common/Src/ABI.cpp b/Source/Core/Common/Src/ABI.cpp
index 097d6fc45b..2e29e2bc6c 100644
--- a/Source/Core/Common/Src/ABI.cpp
+++ b/Source/Core/Common/Src/ABI.cpp
@@ -4,6 +4,40 @@
 

 using namespace Gen;

 

+// Shared code between Win64 and Unix64

+// ====================================

+

+// Sets up a __cdecl function.

+void ABI_EmitPrologue(int maxCallParams)

+{

+#ifdef _M_IX86

+	// Don't really need to do anything

+#elif defined(_M_X64)

+#if _WIN32

+	int stacksize = ((maxCallParams + 1) & ~1)*8 + 8;

+	// Set up a stack frame so that we can call functions

+	// TODO: use maxCallParams

+    SUB(64, R(RSP), Imm8(stacksize));

+#endif

+#else

+#error Arch not supported

+#endif

+}

+void ABI_EmitEpilogue(int maxCallParams)

+{

+#ifdef _M_IX86

+	RET();

+#elif defined(_M_X64)

+#ifdef _WIN32

+	int stacksize = ((maxCallParams+1)&~1)*8 + 8;

+	ADD(64, R(RSP), Imm8(stacksize));

+#endif

+	RET();

+#else

+#error Arch not supported

+#endif

+}

+

 #ifdef _M_IX86 // All32

 

 // Shared code between Win32 and Unix32

@@ -76,6 +110,7 @@ unsigned int ABI_GetAlignedFrameSize(unsigned int frameSize) {
 	return alignedSize;

 }

 

+

 void ABI_AlignStack(unsigned int frameSize) {

 // Mac OS X requires the stack to be 16-byte aligned before every call.

 // Linux requires the stack to be 16-byte aligned before calls that put SSE

@@ -103,9 +138,6 @@ void ABI_RestoreStack(unsigned int frameSize) {
 

 #else

 

-// Shared code between Win64 and Unix64

-// ====================================

-

 void ABI_CallFunctionC(void *func, u32 param1) {

 	MOV(32, R(ABI_PARAM1), Imm32(param1));

 	CALL(func);

diff --git a/Source/Core/Common/Src/ABI.h b/Source/Core/Common/Src/ABI.h
index 2bbd169d00..2dfbf5aa19 100644
--- a/Source/Core/Common/Src/ABI.h
+++ b/Source/Core/Common/Src/ABI.h
@@ -107,6 +107,11 @@ unsigned int ABI_GetAlignedFrameSize(unsigned int frameSize);
 void ABI_AlignStack(unsigned int frameSize);

 void ABI_RestoreStack(unsigned int frameSize);

 

+// Sets up a __cdecl function.

+// Only x64 really needs the parameter.

+void ABI_EmitPrologue(int maxCallParams);

+void ABI_EmitEpilogue(int maxCallParams);

+

 #ifdef _M_IX86

 inline int ABI_GetNumXMMRegs() { return 8; }

 #else

diff --git a/Source/Core/Common/Src/x64Emitter.cpp b/Source/Core/Common/Src/x64Emitter.cpp
index fe839d8537..e3adb2e54b 100644
--- a/Source/Core/Common/Src/x64Emitter.cpp
+++ b/Source/Core/Common/Src/x64Emitter.cpp
@@ -1316,43 +1316,6 @@ namespace Gen
 	}

 

 	void RTDSC() { Write8(0x0F); Write8(0x31); }

-

-	namespace Util

-	{

-

-	// Sets up a __cdecl function.

-	void EmitPrologue(int maxCallParams)

-	{

-#ifdef _M_IX86

-		// Don't really need to do anything

-#elif defined(_M_X64)

-#if _WIN32

-		int stacksize = ((maxCallParams + 1) & ~1)*8 + 8;

-		// Set up a stack frame so that we can call functions

-		// TODO: use maxCallParams

-	    SUB(64, R(RSP), Imm8(stacksize));

-#endif

-#else

-#error Arch not supported

-#endif

-	}

-	void EmitEpilogue(int maxCallParams)

-	{

-#ifdef _M_IX86

-		RET();

-#elif defined(_M_X64)

-#ifdef _WIN32

-		int stacksize = ((maxCallParams+1)&~1)*8 + 8;

-		ADD(64, R(RSP), Imm8(stacksize));

-#endif

-		RET();

-#else

-#error Arch not supported

-#endif

-	}

-

-	}  // namespace

-

 	

 // helper routines for setting pointers

 void CallCdeclFunction3(void* fnptr, u32 arg0, u32 arg1, u32 arg2)

diff --git a/Source/Core/Common/Src/x64Emitter.h b/Source/Core/Common/Src/x64Emitter.h
index 995143b91a..0238569a28 100644
--- a/Source/Core/Common/Src/x64Emitter.h
+++ b/Source/Core/Common/Src/x64Emitter.h
@@ -520,16 +520,8 @@ namespace Gen
 	void PMOVMSKB(X64Reg dest, OpArg arg);

 	void PSHUFB(X64Reg dest, OpArg arg);

 

-        void RTDSC();

+	void RTDSC();

 

-	namespace Util

-	{

-		// Sets up a __cdecl function.

-		// Only x64 really needs the parameter.

-		void EmitPrologue(int maxCallParams);

-		void EmitEpilogue(int maxCallParams);

-	}

-	

 void CallCdeclFunction3(void* fnptr, u32 arg0, u32 arg1, u32 arg2);

 void CallCdeclFunction4(void* fnptr, u32 arg0, u32 arg1, u32 arg2, u32 arg3);

 void CallCdeclFunction5(void* fnptr, u32 arg0, u32 arg1, u32 arg2, u32 arg3, u32 arg4);