Implemented correct idle skipping in JIT IL

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@3300 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Core/Src/PowerPC/Jit64/Jit_LoadStore.cpp

@@ -137,7 +137,6 @@
 			Memory::ReadUnchecked_U32(js.compilerPC + 8) == 0x4182fff8)
 		{
 			// TODO(LinesPrower): 			
-			// - Fix idle skipping in JITIL!
 			// - Rewrite this!
 			// It seems to be ugly and unefficient, but I don't know JIT stuff enough to make it right
 			// It only demonstrates the idea
@@ -164,7 +163,7 @@
 			ABI_CallFunctionC((void *)&PowerPC::OnIdle, PowerPC::ppcState.gpr[a] + (s32)(s16)inst.SIMM_16);
 
 			// ! we must continue executing of the loop after exception handling, maybe there is still 0 in r0
-			//MOV(32, M(&PowerPC::ppcState.pc), Imm32(js.compilerPC + 12));
+			//MOV(32, M(&PowerPC::ppcState.pc), Imm32(js.compilerPC));
 			JMP(asm_routines.testExceptions, true);			
 
 			SetJumpTarget(noIdle);

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

@@ -514,6 +514,14 @@ InstLoc IRBuilder::FoldBranchCond(InstLoc Op1, InstLoc Op2) {
 	return EmitBiOp(BranchCond, Op1, Op2);
 }
 
+InstLoc IRBuilder::FoldIdleBranch(InstLoc Op1, InstLoc Op2) {
+	return EmitBiOp(
+		IdleBranch,
+		EmitICmpEq(getOp1(getOp1(Op1)),
+		getOp2(getOp1(Op1))), Op2
+		);
+}
+
 InstLoc IRBuilder::FoldICmp(unsigned Opcode, InstLoc Op1, InstLoc Op2) {
 	if (isImm(*Op1)) {
 		if (isImm(*Op2)) {
@@ -659,6 +667,7 @@ InstLoc IRBuilder::FoldBiOp(unsigned Opcode, InstLoc Op1, InstLoc Op2, unsigned
 		case Shrl: return FoldShrl(Op1, Op2);
 		case Rol: return FoldRol(Op1, Op2);
 		case BranchCond: return FoldBranchCond(Op1, Op2);
+		case IdleBranch: return FoldIdleBranch(Op1, Op2);
 		case ICmpEq: case ICmpNe:
 		case ICmpUgt: case ICmpUlt: case ICmpUge: case ICmpUle:
 		case ICmpSgt: case ICmpSlt: case ICmpSge: case ICmpSle:
@@ -1311,7 +1320,6 @@ static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile) {
 		case SystemCall:
 		case RFIExit:
 		case InterpreterBranch:		
-		case IdleLoop:
 		case ShortIdleLoop:		
 		case Tramp:
 			// No liveness effects
@@ -1419,6 +1427,9 @@ static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile) {
 			if (!isImm(*getOp1(I)))
 				regMarkUse(RI, I, getOp1(I), 1);
 			break;
+		case IdleBranch:
+			regMarkUse(RI, I, getOp1(getOp1(I)), 1);
+			break;						 
 		case BranchCond: {
 			if (isICmp(*getOp1(I)) &&
 			    isImm(*getOp2(getOp1(I)))) {
@@ -2112,6 +2123,26 @@ static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile) {
 		case BlockStart:
 		case BlockEnd:
 			break;
+
+		case IdleBranch: {			
+			Jit->CMP(32, regLocForInst(RI, getOp1(getOp1(I))),
+					 Imm32(RI.Build->GetImmValue(getOp2(getOp1(I)))));			
+			FixupBranch cont = Jit->J_CC(CC_NE);
+
+			RI.Jit->Cleanup(); // is it needed?			
+			Jit->ABI_CallFunction((void *)&PowerPC::OnIdleIL);
+			
+			Jit->MOV(32, M(&PC), Imm32(ibuild->GetImmValue( getOp2(I) )));
+			Jit->JMP(asm_routines.testExceptions, true);
+
+			Jit->SetJumpTarget(cont);
+			if (RI.IInfo[I - RI.FirstI] & 4)
+					regClearInst(RI, getOp1(getOp1(I)));
+			if (RI.IInfo[I - RI.FirstI] & 8)
+				regClearInst(RI, getOp2(I));
+			break;
+		}
+
 		case BranchCond: {
 			if (isICmp(*getOp1(I)) &&
 			    isImm(*getOp2(getOp1(I)))) {
@@ -2152,14 +2183,6 @@ static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile) {
 			regNormalRegClear(RI, I);
 			break;
 		}		
-		case IdleLoop: {
-			unsigned IdleParam = ibuild->GetImmValue(getOp1(I));
-			unsigned InstLoc = ibuild->GetImmValue(getOp2(I));
-			Jit->ABI_CallFunctionC((void *)&PowerPC::OnIdle, IdleParam);
-			Jit->MOV(32, M(&PC), Imm32(InstLoc + 12));
-			Jit->JMP(asm_routines.testExceptions, true);
-			break;
-		}
 		case ShortIdleLoop: {
 			unsigned InstLoc = ibuild->GetImmValue(getOp1(I));
 			Jit->ABI_CallFunction((void *)&CoreTiming::Idle);

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

@@ -156,7 +156,9 @@ namespace IREmitter {
 		SystemCall,
 		RFIExit,
 		InterpreterBranch,
-		IdleLoop,      // The "usual" idle loop, load+compare+branch
+
+		//IdleLoop,      // The "usual" idle loop, load+compare+branch
+		IdleBranch,	   // branch operation belonging to idle loop
 		ShortIdleLoop, // Idle loop seen in homebrew like wii mahjong,
 			       // just a branch
 
@@ -220,6 +222,7 @@ namespace IREmitter {
 		InstLoc FoldShrl(InstLoc Op1, InstLoc Op2);
 		InstLoc FoldXor(InstLoc Op1, InstLoc Op2);
 		InstLoc FoldBranchCond(InstLoc Op1, InstLoc Op2);
+		InstLoc FoldIdleBranch(InstLoc Op1, InstLoc Op2);
 		InstLoc FoldICmp(unsigned Opcode, InstLoc Op1, InstLoc Op2);
 		InstLoc FoldICmpCRSigned(InstLoc Op1, InstLoc Op2);
 		InstLoc FoldICmpCRUnsigned(InstLoc Op1, InstLoc Op2);
@@ -246,6 +249,9 @@ namespace IREmitter {
 		InstLoc EmitBranchCond(InstLoc check, InstLoc dest) {
 			return FoldBiOp(BranchCond, check, dest);
 		}
+		InstLoc EmitIdleBranch(InstLoc check, InstLoc dest) {
+			return FoldBiOp(IdleBranch, check, dest);
+		}
 		InstLoc EmitLoadCR(unsigned crreg) {
 			return FoldZeroOp(LoadCR, crreg);
 		}
@@ -381,9 +387,6 @@ namespace IREmitter {
 		InstLoc EmitRFIExit() {
 			return FoldZeroOp(RFIExit, 0);
 		}		
-		InstLoc EmitIdleLoop(InstLoc idleParam, InstLoc pc) {
-			return FoldBiOp(IdleLoop, idleParam, pc);
-		}
 		InstLoc EmitShortIdleLoop(InstLoc pc) {
 			return FoldUOp(ShortIdleLoop, pc);
 		}

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

@@ -27,6 +27,8 @@
 #include "JitRegCache.h"
 #include "JitAsm.h"
 
+#include "../../HW/Memmap.h"
+
 // The branches are known good, or at least reasonably good.
 // No need for a disable-mechanism.
 
@@ -127,7 +129,19 @@ using namespace Gen;
 		else
 			destination = js.compilerPC + SignExt16(inst.BD << 2);
 
+		if (Core::GetStartupParameter().bSkipIdle &&
+			inst.hex == 0x4182fff8 &&			
+			(Memory::ReadUnchecked_U32(js.compilerPC - 8) & 0xFFFF0000) == 0x800D0000 &&
+			(Memory::ReadUnchecked_U32(js.compilerPC - 4) == 0x28000000 ||
+			(Core::GetStartupParameter().bWii && Memory::ReadUnchecked_U32(js.compilerPC - 4) == 0x2C000000)) 
+			)
+		{			
+			ibuild.EmitIdleBranch(Test, ibuild.EmitIntConst(destination));
+		}
+		else
+		{
 			ibuild.EmitBranchCond(Test, ibuild.EmitIntConst(destination));
+		}
 		ibuild.EmitBranchUncond(ibuild.EmitIntConst(js.compilerPC + 4));
 	}
 

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

@@ -54,19 +54,6 @@ void Jit64::lXz(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
 	JITDISABLE(LoadStore)	
-	if (Core::GetStartupParameter().bSkipIdle &&
-		inst.OPCD == 32 && 
-		(inst.hex & 0xFFFF0000) == 0x800D0000 &&
-		(Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x28000000 ||
-		(Core::GetStartupParameter().bWii && Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x2C000000)) &&
-		Memory::ReadUnchecked_U32(js.compilerPC + 8) == 0x4182fff8)
-	{
-		ibuild.EmitIdleLoop(ibuild.EmitIntConst(PowerPC::ppcState.gpr[inst.RA] + (s32)(s16)inst.SIMM_16),
-				    ibuild.EmitIntConst(js.compilerPC));
-		js.compilerPC += 8;
-		return;
-	}
-
 	IREmitter::InstLoc addr = ibuild.EmitIntConst(inst.SIMM_16);
 	if (inst.RA)
 		addr = ibuild.EmitAdd(addr, ibuild.EmitLoadGReg(inst.RA));

Source/Core/Core/Src/PowerPC/PowerPC.cpp

@@ -348,4 +348,9 @@ void OnIdle(u32 _uThreadAddr)
 		CoreTiming::Idle();
 }
 
+void OnIdleIL()
+{
+	CoreTiming::Idle();
+}
+
 }  // namespace

Source/Core/Core/Src/PowerPC/PowerPC.h

@@ -92,6 +92,7 @@ void CompactCR();
 void ExpandCR();
 
 void OnIdle(u32 _uThreadAddr);
+void OnIdleIL();
 
 	// Easy register access macros.
 #define HID2 ((UReg_HID2&)PowerPC::ppcState.spr[SPR_HID2])