Merge branch '6663'

Should fix some fastmem-related bugs and possibly improve performance a bit.
dolphin-emu · Oct 5, 2013 · f6c0fb7 · f6c0fb7
2 parents 8c103a8 + a51eb5f
commit f6c0fb7
Show file tree

Hide file tree

Showing 45 changed files with 276 additions and 603 deletions.
diff --git a/Source/Core/Common/CMakeLists.txt b/Source/Core/Common/CMakeLists.txt
@@ -37,7 +37,6 @@ else()
 	if(NOT _M_GENERIC) #X86
 		set(SRCS	${SRCS}
 					Src/x64FPURoundMode.cpp
-					Src/x64Thunk.cpp
 					)
 	endif()
 	set(SRCS ${SRCS} Src/x64CPUDetect.cpp)

diff --git a/Source/Core/Common/Common.vcxproj b/Source/Core/Common/Common.vcxproj
@@ -217,7 +217,6 @@
     <ClCompile Include="Src\x64CPUDetect.cpp" />
     <ClCompile Include="Src\x64Emitter.cpp" />
     <ClCompile Include="Src\x64FPURoundMode.cpp" />
-    <ClCompile Include="Src\x64Thunk.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Src\Atomic.h" />
@@ -263,7 +262,6 @@
     <ClInclude Include="Src\SymbolDB.h" />
     <ClInclude Include="Src\SysConf.h" />
     <ClInclude Include="Src\Thread.h" />
-    <ClInclude Include="Src\Thunk.h" />
     <ClInclude Include="Src\Timer.h" />
     <ClInclude Include="Src\x64ABI.h" />
     <ClInclude Include="Src\x64Analyzer.h" />

diff --git a/Source/Core/Common/Common.vcxproj.filters b/Source/Core/Common/Common.vcxproj.filters
@@ -47,7 +47,6 @@
     <ClCompile Include="Src\x64ABI.cpp" />
     <ClCompile Include="Src\x64CPUDetect.cpp" />
     <ClCompile Include="Src\x64FPURoundMode.cpp" />
-    <ClCompile Include="Src\x64Thunk.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Src\Atomic.h" />
@@ -84,7 +83,6 @@
     <ClInclude Include="Src\SymbolDB.h" />
     <ClInclude Include="Src\SysConf.h" />
     <ClInclude Include="Src\Thread.h" />
-    <ClInclude Include="Src\Thunk.h" />
     <ClInclude Include="Src\Timer.h" />
     <ClInclude Include="Src\x64Analyzer.h" />
     <ClInclude Include="Src\x64Emitter.h" />

diff --git a/Source/Core/Common/Src/Thunk.h b/Source/Core/Common/Src/Thunk.h
diff --git a/Source/Core/Common/Src/x64ABI.cpp b/Source/Core/Common/Src/x64ABI.cpp
@@ -57,6 +57,86 @@ void XEmitter::ABI_RestoreStack(unsigned int frameSize, bool noProlog) {
 	}
 }
 
+void XEmitter::ABI_PushRegistersAndAdjustStack(u32 mask, bool noProlog)
+{
+	int regSize =
+#ifdef _M_X64
+		8;
+#else
+		4;
+#endif
+	int shadow = 0;
+#if defined(_WIN32) && defined(_M_X64)
+	shadow = 0x20;
+#endif
+	int count = 0;
+	for (int r = 0; r < 16; r++)
+	{
+		if (mask & (1 << r))
+		{
+			PUSH((X64Reg) r);
+			count++;
+		}
+	}
+	int size = ((noProlog ? -regSize : 0) - (count * regSize)) & 0xf;
+	for (int x = 0; x < 16; x++)
+	{
+		if (mask & (1 << (16 + x)))
+			size += 16;
+	}
+	size += shadow;
+	if (size)
+		SUB(regSize * 8, R(RSP), size >= 0x80 ? Imm32(size) : Imm8(size));
+	int offset = shadow;
+	for (int x = 0; x < 16; x++)
+	{
+		if (mask & (1 << (16 + x)))
+		{
+			MOVAPD(MDisp(RSP, offset), (X64Reg) x);
+			offset += 16;
+		}
+	}
+}
+
+void XEmitter::ABI_PopRegistersAndAdjustStack(u32 mask, bool noProlog)
+{
+	int regSize =
+#ifdef _M_X64
+		8;
+#else
+		4;
+#endif
+	int size = 0;
+#if defined(_WIN32) && defined(_M_X64)
+	size += 0x20;
+#endif
+	for (int x = 0; x < 16; x++)
+	{
+		if (mask & (1 << (16 + x)))
+		{
+			MOVAPD((X64Reg) x, MDisp(RSP, size));
+			size += 16;
+		}
+	}
+	int count = 0;
+	for (int r = 0; r < 16; r++)
+	{
+		if (mask & (1 << r))
+			count++;
+	}
+	size += ((noProlog ? -regSize : 0) - (count * regSize)) & 0xf;
+
+	if (size)
+		ADD(regSize * 8, R(RSP), size >= 0x80 ? Imm32(size) : Imm8(size));
+	for (int r = 15; r >= 0; r--)
+	{
+		if (mask & (1 << r))
+		{
+			POP((X64Reg) r);
+		}
+	}
+}
+
 #ifdef _M_IX86 // All32
 
 // Shared code between Win32 and Unix32

diff --git a/Source/Core/Common/Src/x64ABI.h b/Source/Core/Common/Src/x64ABI.h
@@ -43,6 +43,8 @@
 // 32-bit bog standard cdecl, shared between linux and windows
 // MacOSX 32-bit is same as System V with a few exceptions that we probably don't care much about.
 
+#define ALL_CALLEE_SAVED ((1 << EAX) | (1 << ECX) | (1 << EDX))
+
 #else // 64 bit calling convention
 
 #ifdef _WIN32 // 64-bit Windows - the really exotic calling convention 
@@ -52,7 +54,12 @@
 #define ABI_PARAM3 R8
 #define ABI_PARAM4 R9
 
-#else  //64-bit Unix (hopefully MacOSX too) 
+#define ABI_ALL_CALLEE_SAVED ((1 << RAX) | (1 << RCX) | (1 << RDX) | (1 << R8) | \
+                              (1 << R9) | (1 << R10) | (1 << R11) | \
+                              (1 << XMM0) | (1 << XMM1) | (1 << XMM2) | (1 << XMM3) | \
+                              (1 << XMM4) | (1 << XMM5))
+
+#else  //64-bit Unix / OS X
 
 #define ABI_PARAM1 RDI
 #define ABI_PARAM2 RSI
@@ -61,6 +68,10 @@
 #define ABI_PARAM5 R8
 #define ABI_PARAM6 R9
 
+#define ABI_ALL_CALLEE_SAVED ((1 << RAX) | (1 << RCX) | (1 << RDX) | (1 << RDI) | \
+                              (1 << RSI) | (1 << R8) | (1 << R9) | (1 << R10) | (1 << R11) | \
+                              0xffff0000 /* xmm0..15 */)
+
 #endif // WIN32
 
 #endif // X86

diff --git a/Source/Core/Common/Src/x64Emitter.cpp b/Source/Core/Common/Src/x64Emitter.cpp
@@ -1634,74 +1634,6 @@ void XEmitter::___CallCdeclImport6(void* impptr, u32 arg0, u32 arg1, u32 arg2, u
 	CALLptr(M(impptr));
 }
 
-void XEmitter::PushRegistersAndAlignStack(u32 mask)
-{
-	int shadow = 0;
-#ifdef _WIN32
-	shadow = 0x20;
-#endif
-	int count = 0;
-	for (int r = 0; r < 16; r++)
-	{
-		if (mask & (1 << r))
-		{
-			PUSH((X64Reg) r);
-			count++;
-		}
-	}
-	int size = (count & 1) ? 0 : 8;
-	for (int x = 0; x < 16; x++)
-	{
-		if (mask & (1 << (16 + x)))
-			size += 16;
-	}
-	size += shadow;
-	if (size)
-		SUB(64, R(RSP), size >= 0x80 ? Imm32(size) : Imm8(size));
-	int offset = shadow;
-	for (int x = 0; x < 16; x++)
-	{
-		if (mask & (1 << (16 + x)))
-		{
-			MOVAPD(MDisp(RSP, offset), (X64Reg) x);
-			offset += 16;
-		}
-	}
-}
-
-void XEmitter::PopRegistersAndAlignStack(u32 mask)
-{
-	int size = 0;
-#ifdef _WIN32
-	size += 0x20;
-#endif
-	for (int x = 0; x < 16; x++)
-	{
-		if (mask & (1 << (16 + x)))
-		{
-			MOVAPD((X64Reg) x, MDisp(RSP, size));
-			size += 16;
-		}
-	}
-	int count = 0;
-	for (int r = 0; r < 16; r++)
-	{
-		if (mask & (1 << r))
-			count++;
-	}
-	size += (count & 1) ? 0 : 8;
-
-	if (size)
-		ADD(64, R(RSP), size >= 0x80 ? Imm32(size) : Imm8(size));
-	for (int r = 15; r >= 0; r--)
-	{
-		if (mask & (1 << r))
-		{
-			POP((X64Reg) r);
-		}
-	}
-}
-
 #endif
 
 }
diff --git a/Source/Core/Common/Src/x64Emitter.h b/Source/Core/Common/Src/x64Emitter.h
@@ -646,6 +646,10 @@ class XEmitter
 	void ABI_PushAllCalleeSavedRegsAndAdjustStack();
 	void ABI_PopAllCalleeSavedRegsAndAdjustStack();
 
+	// A more flexible version of the above.
+	void ABI_PushRegistersAndAdjustStack(u32 mask, bool noProlog);
+	void ABI_PopRegistersAndAdjustStack(u32 mask, bool noProlog);
+
 	unsigned int ABI_GetAlignedFrameSize(unsigned int frameSize, bool noProlog = false);
 	void ABI_AlignStack(unsigned int frameSize, bool noProlog = false);
 	void ABI_RestoreStack(unsigned int frameSize, bool noProlog = false);
@@ -691,9 +695,6 @@ class XEmitter
 
 	#define DECLARE_IMPORT(x) extern "C" void *__imp_##x
 
-	void PushRegistersAndAlignStack(u32 mask);
-	void PopRegistersAndAlignStack(u32 mask);
-
 #endif
 };  // class XEmitter