A few small JitIL changes that I've had sitting in my tree for a while;

nothing really significant to note.



git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@2106 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Core/Src/PowerPC/Jit64IL/IR.cpp

@@ -585,6 +585,18 @@ InstLoc IRBuilder::FoldInterpreterFallback(InstLoc Op1, InstLoc Op2) {
 	return EmitBiOp(InterpreterFallback, Op1, Op2);
 }
 
+InstLoc IRBuilder::FoldDoubleBiOp(unsigned Opcode, InstLoc Op1, InstLoc Op2) {
+	if (getOpcode(*Op1) == InsertDoubleInMReg) {
+		return FoldDoubleBiOp(Opcode, getOp1(Op1), Op2);
+	}
+
+	if (getOpcode(*Op2) == InsertDoubleInMReg) {
+		return FoldDoubleBiOp(Opcode, Op1, getOp1(Op2));
+	}
+
+	return EmitBiOp(Opcode, Op1, Op2);
+}
+
 InstLoc IRBuilder::FoldBiOp(unsigned Opcode, InstLoc Op1, InstLoc Op2, unsigned extra) {
 	switch (Opcode) {
 		case Add: return FoldAdd(Op1, Op2);
@@ -601,6 +613,7 @@ InstLoc IRBuilder::FoldBiOp(unsigned Opcode, InstLoc Op1, InstLoc Op2, unsigned
 		case ICmpSgt: case ICmpSlt: case ICmpSge: case ICmpSle:
 			return FoldICmp(Opcode, Op1, Op2);
 		case InterpreterFallback: return FoldInterpreterFallback(Op1, Op2);
+		case FDMul: case FDAdd: case FDSub: return FoldDoubleBiOp(Opcode, Op1, Op2);
 		default: return EmitBiOp(Opcode, Op1, Op2, extra);
 	}
 }

Source/Core/Core/Src/PowerPC/Jit64IL/IR.h

@@ -88,7 +88,6 @@ namespace IREmitter {
 		Store16,
 		Store32,
 		BranchCond,
-#if 0
 		// Floating-point
 		// There are three floating-point formats: single, double,
 		// and packed.  For any operation where the format of the
@@ -98,59 +97,15 @@ namespace IREmitter {
 		// The "mreg" format is a pair of doubles; this is the
 		// most general possible represenation which is used
 		// in the register state.
-		// This might seem like overkill, but it's a huge advantage
-		// to keep operands in the right format because extra
-		// precision can screw up games.
-		// FIXME: Does the slight loss of precision due to not
-		// having a madd instruction matter?  It would be a
-		// performance loss for singles because the operations
-		// would have to be done in double precision, and a completely 
-		// accurate double madd would require an extremely expensive
-		// fallback.
-		FDAdd,
-		FDSub,
-		FDMul,
-		FDDiv,
-		FDNeg,
-		FSAdd,
-		FSSub,
-		FSMul,
-		FSDiv,
-		FSNeg,
-		FPSAdd,
-		FPSSub,
-		FPSMul,
-		FPSDiv,
-		FPSNeg,
-		// FP Loads
-		LoadSingle,
-		LoadDouble,
-		// LoadPacked, // FIXME: Work out how this instruction should
-				// be implemented
-		// FP Stores
-		StoreSingle,
-		StoreDouble,
-		// StorePacked, // FIXME: Work out how this instruction should
-				// be implemented
-		PackedToSingle, // Extract PS0 from packed (type-pun)
-		// PackedToDouble == PackedToSingle+SingleToDouble
-		PackedToMReg,   // Convert from packed format to mreg format (CVTPS2PD)
-		SingleToDouble, // Widen single to double (CVTSS2SD)
-		SingleToPacked, // Duplicate single to packed
-		// SingleToMReg == SingleToPacked+PackedToMReg
-		MRegToPacked,   // Convert from mreg format to packed format (CVTPD2PS)
-		MRegToDouble,   // Extract bottom half from mreg format. (type-pun)
-		// MRegToSingle == MRegToDouble + DoubleToSingle
-		DoubleToMReg,   // Convert from double format to mreg format
-		DoubleToSingle, // Convert from double to single format (CVTSD2SS)
-		// DoubleToPacked should never be needed
-
-		ForceToPacked,  // ForceTo* are "virtual"; they should be
-				// folded into the above conversions.
-		ForceToSingle,
-		ForceToDouble,
-		ForceToMReg,
-#endif
+		// This might seem like overkill, but the semantics require
+		// having the different formats.
+		// FIXME: Check the accuracy of the mapping:
+		// 1. Is paired arithmetic always rounded to single-precision
+		//    first, or does it do double-to-single like the
+		//    single-precision instructions?
+		// 2. The implementation of madd is slightly off, and
+		//    the implementation of fmuls is very slightly off;
+		//    likely nothing cares, though.
 		FResult_Start,
 		LoadSingle,
 		LoadDouble,
@@ -263,6 +218,7 @@ namespace IREmitter {
 		InstLoc FoldXor(InstLoc Op1, InstLoc Op2);
 		InstLoc FoldBranchCond(InstLoc Op1, InstLoc Op2);
 		InstLoc FoldICmp(unsigned Opcode, InstLoc Op1, InstLoc Op2);
+		InstLoc FoldDoubleBiOp(unsigned Opcode, InstLoc Op1, InstLoc Op2);
 
 		InstLoc FoldInterpreterFallback(InstLoc Op1, InstLoc Op2);
 

Source/Core/Core/Src/PowerPC/Jit64IL/Jit.cpp

@@ -178,7 +178,7 @@ namespace CPUCompare
 			CODE_SIZE = 1024*1024*8*8;
 
 		jo.optimizeStack = true;
-		jo.enableBlocklink = true;  // Speed boost, but not 100% safe
+		jo.enableBlocklink = false;  // Speed boost, but not 100% safe
 #ifdef _M_X64
 		jo.enableFastMem = Core::GetStartupParameter().bUseFastMem;
 #else
@@ -200,6 +200,13 @@ namespace CPUCompare
 		asm_routines.Init();
 	}
 
+	void Jit64::ClearCache()
+	{
+		blocks.Clear();
+		trampolines.ClearCodeSpace();
+		ClearCodeSpace();
+	}
+
 	void Jit64::Shutdown()
 	{
 		FreeCodeSpace();

Source/Core/Core/Src/PowerPC/Jit64IL/Jit.h

@@ -146,11 +146,7 @@ public:
 
 	void NotifyBreakpoint(u32 em_address, bool set);
 
-	void ClearCache() 
-	{
-		blocks.Clear();
-		trampolines.ClearCodeSpace();
-	}
+	void ClearCache();
 
 	// Run!