jit: Avoid applying/restoring the rounding mode.

If the game never sets it, we can skip around syscalls, interpreter,
replacements, etc.

Core/MIPS/ARM/ArmAsm.cpp

@@ -114,9 +114,9 @@ void Jit::GenerateFixedCode()
 	MovToPC(R0);
 	outerLoop = GetCodePtr();
 		SaveDowncount();
-		RestoreRoundingMode();
+		RestoreRoundingMode(true);
 		QuickCallFunction(R0, &CoreTiming::Advance);
-		ApplyRoundingMode();
+		ApplyRoundingMode(true);
 		RestoreDowncount();
 		FixupBranch skipToRealDispatch = B(); //skip the sync and compare first time
 
@@ -175,9 +175,9 @@ void Jit::GenerateFixedCode()
 
 			// No block found, let's jit
 			SaveDowncount();
-			RestoreRoundingMode();
+			RestoreRoundingMode(true);
 			QuickCallFunction(R2, (void *)&JitAt);
-			ApplyRoundingMode();
+			ApplyRoundingMode(true);
 			RestoreDowncount();
 
 			B(dispatcherNoCheck); // no point in special casing this
@@ -199,7 +199,7 @@ void Jit::GenerateFixedCode()
 	}
 
 	SaveDowncount();
-	RestoreRoundingMode();
+	RestoreRoundingMode(true);
 
 	ADD(R_SP, R_SP, 4);
 

Core/MIPS/ARM/ArmCompFPU.cpp

@@ -420,6 +420,7 @@ void Jit::Comp_mxc1(MIPSOpcode op)
 				AND(gpr.R(MIPS_REG_FPCOND), SCRATCHREG1, Operand2(1));
 #endif
 			}
+			UpdateRoundingMode();
 			ApplyRoundingMode();
 		} else {
 			Comp_Generic(op);

Core/MIPS/ARM/ArmJit.cpp

@@ -94,22 +94,32 @@ Jit::Jit(MIPSState *mips) : blocks(mips, this), gpr(mips, &jo), fpr(mips), mips_
 
 void Jit::DoState(PointerWrap &p)
 {
-	auto s = p.Section("Jit", 1);
+	auto s = p.Section("Jit", 1, 2);
 	if (!s)
 		return;
 
 	p.Do(js.startDefaultPrefix);
+	if (s >= 2) {
+		p.Do(js.hasSetRounding);
+		js.lastSetRounding = 0;
+	} else {
+		js.hasSetRounding = 1;
+	}
 }
 
 // This is here so the savestate matches between jit and non-jit.
 void Jit::DoDummyState(PointerWrap &p)
 {
-	auto s = p.Section("Jit", 1);
+	auto s = p.Section("Jit", 1, 2);
 	if (!s)
 		return;
 
 	bool dummy = false;
 	p.Do(dummy);
+	if (s >= 2) {
+		dummy = true;
+		p.Do(dummy);
+	}
 }
 
 void Jit::FlushAll()
@@ -201,17 +211,28 @@ void Jit::Compile(u32 em_address) {
 	DoJit(em_address, b);
 	blocks.FinalizeBlock(block_num, jo.enableBlocklink);
 
+	bool cleanSlate = false;
+
+	if (js.hasSetRounding && !js.lastSetRounding) {
+		WARN_LOG(JIT, "Detected rounding mode usage, rebuilding jit with checks");
+		// Won't loop, since hasSetRounding is only ever set to 1.
+		js.lastSetRounding = js.hasSetRounding;
+		cleanSlate = true;
+	}
+
 	// Drat.  The VFPU hit an uneaten prefix at the end of a block.
 	if (js.startDefaultPrefix && js.MayHavePrefix()) {
 		WARN_LOG(JIT, "An uneaten prefix at end of block: %08x", js.compilerPC - 4);
 		js.LogPrefix();
 
+		// Let's try that one more time.  We won't get back here because we toggled the value.
 		js.startDefaultPrefix = false;
+		cleanSlate = true;
+	}
 
+	if (cleanSlate) {
 		// Our assumptions are all wrong so it's clean-slate time.
 		ClearCache();
-
-		// Let's try that one more time.  We won't get back here because we toggled the value.
 		Compile(em_address);
 	}
 }
@@ -547,8 +568,9 @@ void Jit::WriteDownCountR(ARMReg reg) {
 	}
 }
 
-void Jit::RestoreRoundingMode() {
+void Jit::RestoreRoundingMode(bool force) {
-	if (g_Config.bSetRoundingMode) {
+	// If the game has never set an interesting rounding mode, we can safely skip this.
+	if (g_Config.bSetRoundingMode && (force || !g_Config.bForceFlushToZero || js.hasSetRounding)) {
 		VMRS(SCRATCHREG2);
 		// Assume we're always in round-to-nearest mode beforehand.
 		// Also on ARM, we're always in flush-to-zero in C++, so stay that way.
@@ -560,9 +582,10 @@ void Jit::RestoreRoundingMode() {
 	}
 }
 
-void Jit::ApplyRoundingMode() {
+void Jit::ApplyRoundingMode(bool force) {
 	// NOTE: Must not destory R0.
-	if (g_Config.bSetRoundingMode) {
+	// If the game has never set an interesting rounding mode, we can safely skip this.
+	if (g_Config.bSetRoundingMode && (force || !g_Config.bForceFlushToZero || js.hasSetRounding)) {
 		LDR(SCRATCHREG2, CTXREG, offsetof(MIPSState, fcr31));
 		if (!g_Config.bForceFlushToZero) {
 			TST(SCRATCHREG2, AssumeMakeOperand2(1 << 24));
@@ -609,6 +632,32 @@ void Jit::ApplyRoundingMode() {
 	}
 }
 
+void Jit::UpdateRoundingMode() {
+	// NOTE: Must not destory R0.
+	if (g_Config.bSetRoundingMode) {
+		LDR(SCRATCHREG2, CTXREG, offsetof(MIPSState, fcr31));
+		if (!g_Config.bForceFlushToZero) {
+			TST(SCRATCHREG2, AssumeMakeOperand2(1 << 24));
+			AND(SCRATCHREG2, SCRATCHREG2, Operand2(3));
+			SetCC(CC_NEQ);
+			ADD(SCRATCHREG2, SCRATCHREG2, Operand2(4));
+			SetCC(CC_AL);
+			// We can only skip if the rounding mode is zero and flush is set.
+			CMP(SCRATCHREG2, Operand2(4));
+		} else {
+			ANDS(SCRATCHREG2, SCRATCHREG2, Operand2(3));
+		}
+
+		FixupBranch skip = B_CC(CC_EQ);
+		PUSH(1, SCRATCHREG1);
+		MOVI2R(SCRATCHREG2, 1);
+		MOVP2R(SCRATCHREG1, &js.hasSetRounding);
+		STRB(SCRATCHREG2, SCRATCHREG1, 0);
+		POP(1, SCRATCHREG1);
+		SetJumpTarget(skip);
+	}
+}
+
 // IDEA - could have a WriteDualExit that takes two destinations and two condition flags,
 // and just have conditional that set PC "twice". This only works when we fall back to dispatcher
 // though, as we need to have the SUBS flag set in the end. So with block linking in the mix,

Core/MIPS/ARM/ArmJit.h

@@ -192,8 +192,9 @@ private:
 
 	void WriteDownCount(int offset = 0);
 	void WriteDownCountR(ARMReg reg);
-	void RestoreRoundingMode();
+	void RestoreRoundingMode(bool force = false);
-	void ApplyRoundingMode();
+	void ApplyRoundingMode(bool force = false);
+	void UpdateRoundingMode();
 	void MovFromPC(ARMReg r);
 	void MovToPC(ARMReg r);
 

Core/MIPS/JitCommon/JitState.h

@@ -55,7 +55,9 @@ namespace MIPSComp {
 		};
 
 		JitState()
-			: startDefaultPrefix(true),
+			: hasSetRounding(0),
+			lastSetRounding(0),
+			startDefaultPrefix(true),
 			prefixSFlag(PREFIX_UNKNOWN),
 			prefixTFlag(PREFIX_UNKNOWN),
 			prefixDFlag(PREFIX_UNKNOWN) {}
@@ -72,6 +74,9 @@ namespace MIPSComp {
 		bool compiling;	// TODO: get rid of this in favor of using analysis results to determine end of block
 		JitBlock *curBlock;
 
+		u8 hasSetRounding;
+		u8 lastSetRounding;
+
 		// VFPU prefix magic
 		bool startDefaultPrefix;
 		u32 prefixS;

Core/MIPS/x86/Asm.cpp

@@ -77,9 +77,9 @@ void AsmRoutineManager::Generate(MIPSState *mips, MIPSComp::Jit *jit)
 #endif
 
 	outerLoop = GetCodePtr();
-		jit->RestoreRoundingMode(this);
+		jit->RestoreRoundingMode(true, this);
 		ABI_CallFunction(reinterpret_cast<void *>(&CoreTiming::Advance));
-		jit->ApplyRoundingMode(this);
+		jit->ApplyRoundingMode(true, this);
 		FixupBranch skipToRealDispatch = J(); //skip the sync and compare first time
 
 		dispatcherCheckCoreState = GetCodePtr();
@@ -134,9 +134,9 @@ void AsmRoutineManager::Generate(MIPSState *mips, MIPSComp::Jit *jit)
 			SetJumpTarget(notfound);
 
 			//Ok, no block, let's jit
-			jit->RestoreRoundingMode(this);
+			jit->RestoreRoundingMode(true, this);
 			ABI_CallFunction(&Jit);
-			jit->ApplyRoundingMode(this);
+			jit->ApplyRoundingMode(true, this);
 			JMP(dispatcherNoCheck, true); // Let's just dispatch again, we'll enter the block since we know it's there.
 
 		SetJumpTarget(bail);
@@ -146,12 +146,12 @@ void AsmRoutineManager::Generate(MIPSState *mips, MIPSComp::Jit *jit)
 		J_CC(CC_Z, outerLoop, true);
 
 	SetJumpTarget(badCoreState);
-	jit->RestoreRoundingMode(this);
+	jit->RestoreRoundingMode(true, this);
 	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	RET();
 
 	breakpointBailout = GetCodePtr();
-	jit->RestoreRoundingMode(this);
+	jit->RestoreRoundingMode(true, this);
 	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	RET();
 }

Core/MIPS/x86/CompFPU.cpp

@@ -387,6 +387,7 @@ void Jit::Comp_mxc1(MIPSOpcode op)
 				if ((gpr.GetImm(rt) & 0x1000003) == 0) {
 					// Default nearest / no-flush mode, just leave it cleared.
 				} else {
+					UpdateRoundingMode();
 					ApplyRoundingMode();
 				}
 			} else {
@@ -399,6 +400,7 @@ void Jit::Comp_mxc1(MIPSOpcode op)
 				MOV(32, M(&mips_->fcr31), gpr.R(rt));
 				AND(32, M(&mips_->fcr31), Imm32(0x0181FFFF));
 				gpr.UnlockAll();
+				UpdateRoundingMode();
 				ApplyRoundingMode();
 			}
 		} else {

Core/MIPS/x86/Jit.cpp

@@ -145,22 +145,32 @@ Jit::~Jit() {
 
 void Jit::DoState(PointerWrap &p)
 {
-	auto s = p.Section("Jit", 1);
+	auto s = p.Section("Jit", 1, 2);
 	if (!s)
 		return;
 
 	p.Do(js.startDefaultPrefix);
+	if (s >= 2) {
+		p.Do(js.hasSetRounding);
+		js.lastSetRounding = 0;
+	} else {
+		js.hasSetRounding = 1;
+	}
 }
 
 // This is here so the savestate matches between jit and non-jit.
 void Jit::DoDummyState(PointerWrap &p)
 {
-	auto s = p.Section("Jit", 1);
+	auto s = p.Section("Jit", 1, 2);
 	if (!s)
 		return;
 
 	bool dummy = false;
 	p.Do(dummy);
+	if (s >= 2) {
+		dummy = true;
+		p.Do(dummy);
+	}
 }
 
 
@@ -211,9 +221,10 @@ void Jit::WriteDowncount(int offset)
 	SUB(32, M(&currentMIPS->downcount), downcount > 127 ? Imm32(downcount) : Imm8(downcount));
 }
 
-void Jit::RestoreRoundingMode(XEmitter *emitter)
+void Jit::RestoreRoundingMode(bool force, XEmitter *emitter)
 {
-	if (g_Config.bSetRoundingMode)
+	// If the game has never set an interesting rounding mode, we can safely skip this.
+	if (g_Config.bSetRoundingMode && (force || g_Config.bForceFlushToZero || js.hasSetRounding))
 	{
 		if (emitter == NULL)
 			emitter = this;
@@ -224,9 +235,10 @@ void Jit::RestoreRoundingMode(XEmitter *emitter)
 	}
 }
 
-void Jit::ApplyRoundingMode(XEmitter *emitter)
+void Jit::ApplyRoundingMode(bool force, XEmitter *emitter)
 {
-	if (g_Config.bSetRoundingMode)
+	// If the game has never set an interesting rounding mode, we can safely skip this.
+	if (g_Config.bSetRoundingMode && (force || g_Config.bForceFlushToZero || js.hasSetRounding))
 	{
 		if (emitter == NULL)
 			emitter = this;
@@ -265,6 +277,22 @@ void Jit::ApplyRoundingMode(XEmitter *emitter)
 	}
 }
 
+void Jit::UpdateRoundingMode(XEmitter *emitter)
+{
+	if (g_Config.bSetRoundingMode)
+	{
+		if (emitter == NULL)
+			emitter = this;
+
+		// If it's only ever 0, we don't actually bother applying or restoring it.
+		// This is the most common situation.
+		emitter->TEST(32, M(&mips_->fcr31), Imm32(0x01000003));
+		FixupBranch skip = emitter->J_CC(CC_Z);
+		emitter->MOV(8, M(&js.hasSetRounding), Imm8(1));
+		emitter->SetJumpTarget(skip);
+	}
+}
+
 void Jit::ClearCache()
 {
 	blocks.Clear();
@@ -330,14 +358,28 @@ void Jit::Compile(u32 em_address)
 	DoJit(em_address, b);
 	blocks.FinalizeBlock(block_num, jo.enableBlocklink);
 
+	bool cleanSlate = false;
+
+	if (js.hasSetRounding && !js.lastSetRounding) {
+		WARN_LOG(JIT, "Detected rounding mode usage, rebuilding jit with checks");
+		// Won't loop, since hasSetRounding is only ever set to 1.
+		js.lastSetRounding = js.hasSetRounding;
+		cleanSlate = true;
+	}
+
 	// Drat.  The VFPU hit an uneaten prefix at the end of a block.
 	if (js.startDefaultPrefix && js.MayHavePrefix()) {
-		WARN_LOG(JIT, "Uneaten prefix at end of block: %08x", js.compilerPC - 4);
+		WARN_LOG(JIT, "An uneaten prefix at end of block: %08x", js.compilerPC - 4);
-		js.startDefaultPrefix = false;
+		js.LogPrefix();
-		// Our assumptions are all wrong so it's clean-slate time.
-		ClearCache();
 
 		// Let's try that one more time.  We won't get back here because we toggled the value.
+		js.startDefaultPrefix = false;
+		cleanSlate = true;
+	}
+
+	if (cleanSlate) {
+		// Our assumptions are all wrong so it's clean-slate time.
+		ClearCache();
 		Compile(em_address);
 	}
 }

Core/MIPS/x86/Jit.h

@@ -165,8 +165,9 @@ public:
 	void GetVectorRegsPrefixD(u8 *regs, VectorSize sz, int vectorReg);
 	void EatPrefix() { js.EatPrefix(); }
 
-	void RestoreRoundingMode(XEmitter *emitter = NULL);
+	void RestoreRoundingMode(bool force = false, XEmitter *emitter = NULL);
-	void ApplyRoundingMode(XEmitter *emitter = NULL);
+	void ApplyRoundingMode(bool force = false, XEmitter *emitter = NULL);
+	void UpdateRoundingMode(XEmitter *emitter = NULL);
 
 	JitBlockCache *GetBlockCache() { return &blocks; }
 	AsmRoutineManager &Asm() { return asm_; }