Support conditional register cache flushing on ARMv7.

Similar to PR #451, but for ARM as well.

Source/Core/Core/PowerPC/JitArm32/JitArm_Branch.cpp

@@ -150,9 +150,6 @@ void JitArm::bcx(UGeckoInstruction inst)
 	JITDISABLE(bJITBranchOff)
 	// USES_CR
 
-	gpr.Flush();
-	fpr.Flush();
-
 	ARMReg rA = gpr.GetReg();
 	ARMReg rB = gpr.GetReg();
 	FixupBranch pCTRDontBranch;
@@ -195,6 +192,9 @@ void JitArm::bcx(UGeckoInstruction inst)
 		destination = SignExt16(inst.BD << 2);
 	else
 		destination = js.compilerPC + SignExt16(inst.BD << 2);
+
+	gpr.Flush(FLUSH_MAINTAIN_STATE);
+	fpr.Flush(FLUSH_MAINTAIN_STATE);
 	WriteExit(destination);
 
 	if ((inst.BO & BO_DONT_CHECK_CONDITION) == 0)
@@ -203,16 +203,17 @@ void JitArm::bcx(UGeckoInstruction inst)
 		SetJumpTarget( pCTRDontBranch );
 
 	if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
+	{
+		gpr.Flush();
+		fpr.Flush();
 		WriteExit(js.compilerPC + 4);
+	}
 }
 void JitArm::bcctrx(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
 	JITDISABLE(bJITBranchOff)
 
-	gpr.Flush();
-	fpr.Flush();
-
 	// bcctrx doesn't decrement and/or test CTR
 	_dbg_assert_msg_(POWERPC, inst.BO_2 & BO_DONT_DECREMENT_FLAG, "bcctrx with decrement and test CTR option is invalid!");
 
@@ -221,6 +222,9 @@ void JitArm::bcctrx(UGeckoInstruction inst)
 		// BO_2 == 1z1zz -> b always
 
 		//NPC = CTR & 0xfffffffc;
+		gpr.Flush();
+		fpr.Flush();
+
 		ARMReg rA = gpr.GetReg();
 
 		if (inst.LK_3)
@@ -262,12 +266,19 @@ void JitArm::bcctrx(UGeckoInstruction inst)
 			//ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
 		}
 		gpr.Unlock(rB); // rA gets unlocked in WriteExitDestInR
+		gpr.Flush(FLUSH_MAINTAIN_STATE);
+		fpr.Flush(FLUSH_MAINTAIN_STATE);
+
 		WriteExitDestInR(rA);
 
 		SetJumpTarget(b);
 
 		if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
+		{
+			gpr.Flush();
+			fpr.Flush();
 			WriteExit(js.compilerPC + 4);
+		}
 	}
 }
 void JitArm::bclrx(UGeckoInstruction inst)
@@ -275,9 +286,6 @@ void JitArm::bclrx(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(bJITBranchOff)
 
-	gpr.Flush();
-	fpr.Flush();
-
 	ARMReg rA = gpr.GetReg();
 	ARMReg rB = gpr.GetReg();
 	FixupBranch pCTRDontBranch;
@@ -327,6 +335,9 @@ void JitArm::bclrx(UGeckoInstruction inst)
 		//ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
 	}
 	gpr.Unlock(rB); // rA gets unlocked in WriteExitDestInR
+
+	gpr.Flush(FLUSH_MAINTAIN_STATE);
+	fpr.Flush(FLUSH_MAINTAIN_STATE);
 	WriteExitDestInR(rA);
 
 	if ((inst.BO & BO_DONT_CHECK_CONDITION) == 0)
@@ -335,5 +346,9 @@ void JitArm::bclrx(UGeckoInstruction inst)
 		SetJumpTarget( pCTRDontBranch );
 
 	if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
+	{
+		gpr.Flush();
+		fpr.Flush();
 		WriteExit(js.compilerPC + 4);
+	}
 }

Source/Core/Core/PowerPC/JitArm32/JitArm_Integer.cpp

@@ -954,9 +954,6 @@ void JitArm::twx(UGeckoInstruction inst)
 
 	s32 a = inst.RA;
 
-	gpr.Flush();
-	fpr.Flush();
-
 	ARMReg RA = gpr.GetReg();
 	ARMReg RB = gpr.GetReg();
 	MOV(RA, inst.TO);
@@ -1003,6 +1000,9 @@ void JitArm::twx(UGeckoInstruction inst)
 	SetJumpTarget(take4);
 	SetJumpTarget(take5);
 
+	gpr.Flush(FLUSH_MAINTAIN_STATE);
+	fpr.Flush(FLUSH_MAINTAIN_STATE);
+
 	LDR(RA, R9, PPCSTATE_OFF(Exceptions));
 	MOVI2R(RB, EXCEPTION_PROGRAM); // XXX: Can be optimized
 	ORR(RA, RA, RB);
@@ -1016,7 +1016,12 @@ void JitArm::twx(UGeckoInstruction inst)
 	SetJumpTarget(exit5);
 
 	if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
+	{
+		gpr.Flush();
+		fpr.Flush();
+
 		WriteExit(js.compilerPC + 4);
+	}
 
 	gpr.Unlock(RA, RB);
 }

Source/Core/Core/PowerPC/JitArm32/JitArm_LoadStore.cpp

@@ -446,12 +446,14 @@ void JitArm::lXX(UGeckoInstruction inst)
 		Memory::ReadUnchecked_U32(js.compilerPC + 8) == 0x4182fff8)
 	{
 		ARMReg RD = gpr.R(d);
-		gpr.Flush();
-		fpr.Flush();
 
 		// if it's still 0, we can wait until the next event
 		TST(RD, RD);
 		FixupBranch noIdle = B_CC(CC_NEQ);
+
+		gpr.Flush(FLUSH_MAINTAIN_STATE);
+		fpr.Flush(FLUSH_MAINTAIN_STATE);
+
 		rA = gpr.GetReg();
 
 		MOVI2R(rA, (u32)&PowerPC::OnIdle);

Source/Core/Core/PowerPC/JitArm32/JitFPRCache.cpp

@@ -32,6 +32,27 @@ void ArmFPRCache::Init(ARMXEmitter *emitter)
 
 void ArmFPRCache::Start(PPCAnalyst::BlockRegStats &stats)
 {
+	// Make sure the state is wiped on Start
+	// There is a potential for the state remaining dirty from the previous block
+	// This is due to conditional branches not clearing the register cache state
+
+	for (u8 a = 0; a < 32; ++a)
+	{
+		if (_regs[a][0].GetType() != REG_NOTLOADED)
+		{
+			u32 regindex = _regs[a][0].GetRegIndex();
+			ArmCRegs[regindex].PPCReg = 33;
+			ArmCRegs[regindex].LastLoad = 0;
+			_regs[a][0].Flush();
+		}
+		if (_regs[a][1].GetType() != REG_NOTLOADED)
+		{
+			u32 regindex = _regs[a][1].GetRegIndex();
+			ArmCRegs[regindex].PPCReg = 33;
+			ArmCRegs[regindex].LastLoad = 0;
+			_regs[a][1].Flush();
+		}
+	}
 }
 
 ARMReg *ArmFPRCache::GetPPCAllocationOrder(int &count)
@@ -162,7 +183,7 @@ ARMReg ArmFPRCache::R1(u32 preg, bool preLoad)
 	return GetPPCReg(preg, true, preLoad);
 }
 
-void ArmFPRCache::Flush()
+void ArmFPRCache::Flush(FlushMode mode)
 {
 	for (u8 a = 0; a < 32; ++a)
 	{
@@ -172,9 +193,12 @@ void ArmFPRCache::Flush()
 			u32 regindex = _regs[a][0].GetRegIndex();
 			emit->VSTR(ArmCRegs[regindex].Reg, R9, offset);
 
-			ArmCRegs[regindex].PPCReg = 33;
+			if (mode == FLUSH_ALL)
-			ArmCRegs[regindex].LastLoad = 0;
+			{
-			_regs[a][0].Flush();
+				ArmCRegs[regindex].PPCReg = 33;
+				ArmCRegs[regindex].LastLoad = 0;
+				_regs[a][0].Flush();
+			}
 		}
 		if (_regs[a][1].GetType() != REG_NOTLOADED)
 		{
@@ -182,9 +206,12 @@ void ArmFPRCache::Flush()
 			u32 regindex = _regs[a][1].GetRegIndex();
 			emit->VSTR(ArmCRegs[regindex].Reg, R9, offset);
 
-			ArmCRegs[regindex].PPCReg = 33;
+			if (mode == FLUSH_ALL)
-			ArmCRegs[regindex].LastLoad = 0;
+			{
-			_regs[a][1].Flush();
+				ArmCRegs[regindex].PPCReg = 33;
+				ArmCRegs[regindex].LastLoad = 0;
+				_regs[a][1].Flush();
+			}
 		}
 	}
 }

Source/Core/Core/PowerPC/JitArm32/JitFPRCache.h

@@ -43,7 +43,7 @@ public:
 
 	ARMReg GetReg(bool AutoLock = true); // Return a ARM register we can use.
 	void Unlock(ARMReg V0);
-	void Flush();
+	void Flush(FlushMode mode = FLUSH_ALL);
 	ARMReg R0(u32 preg, bool preLoad = true); // Returns a cached register
 	ARMReg R1(u32 preg, bool preLoad = true);
 };

Source/Core/Core/PowerPC/JitArm32/JitRegCache.cpp

@@ -30,6 +30,19 @@ void ArmRegCache::Init(ARMXEmitter *emitter)
 }
 void ArmRegCache::Start(PPCAnalyst::BlockRegStats &stats)
 {
+	// Make sure the state is wiped on Start
+	// There is a potential for the state remaining dirty from the previous block
+	// This is due to conditional branches not clearing the register cache state
+	for (u8 a = 0; a < 32; ++a)
+	{
+		if (regs[a].GetType() == REG_REG)
+		{
+			u32 regindex = regs[a].GetRegIndex();
+			ArmCRegs[regindex].PPCReg = 33;
+			ArmCRegs[regindex].LastLoad = 0;
+		}
+		regs[a].Flush();
+	}
 }
 
 ARMReg *ArmRegCache::GetPPCAllocationOrder(int &count)
@@ -192,21 +205,36 @@ void ArmRegCache::SetImmediate(u32 preg, u32 imm)
 	regs[preg].LoadToImm(imm);
 }
 
-void ArmRegCache::Flush()
+void ArmRegCache::Flush(FlushMode mode)
 {
 	for (u8 a = 0; a < 32; ++a)
 	{
 		if (regs[a].GetType() == REG_IMM)
-			BindToRegister(a);
+		{
+			if (mode == FLUSH_ALL)
+			{
+				// This changes the type over to a REG_REG and gets caught below.
+				BindToRegister(a);
+			}
+			else
+			{
+				ARMReg tmp = GetReg();
+				emit->MOVI2R(tmp, regs[a].GetImm());
+				emit->STR(tmp, R9, PPCSTATE_OFF(gpr) + a * 4);
+				Unlock(tmp);
+			}
+		}
 		if (regs[a].GetType() == REG_REG)
 		{
 			u32 regindex = regs[a].GetRegIndex();
 			emit->STR(ArmCRegs[regindex].Reg, R9, PPCSTATE_OFF(gpr) + a * 4);
-			ArmCRegs[regindex].PPCReg = 33;
+			if (mode == FLUSH_ALL)
-			ArmCRegs[regindex].LastLoad = 0;
+			{
+				ArmCRegs[regindex].PPCReg = 33;
+				ArmCRegs[regindex].LastLoad = 0;
+				regs[a].Flush();
+			}
 		}
-
-		regs[a].Flush();
 	}
 }
 

Source/Core/Core/PowerPC/JitArm32/JitRegCache.h

@@ -29,6 +29,12 @@ enum RegType
 	REG_AWAY, // Bound to a register, but not preloaded
 };
 
+enum FlushMode
+{
+	FLUSH_ALL = 0,
+	FLUSH_MAINTAIN_STATE,
+};
+
 class OpArg
 {
 	private:
@@ -116,9 +122,8 @@ public:
 	void Start(PPCAnalyst::BlockRegStats &stats);
 
 	ARMReg GetReg(bool AutoLock = true); // Return a ARM register we can use.
-	void Unlock(ARMReg R0, ARMReg R1 = INVALID_REG, ARMReg R2 = INVALID_REG, ARMReg R3 =
+	void Unlock(ARMReg R0, ARMReg R1 = INVALID_REG, ARMReg R2 = INVALID_REG, ARMReg R3 = INVALID_REG);
-	INVALID_REG);
+	void Flush(FlushMode mode = FLUSH_ALL);
-	void Flush();
 	ARMReg R(u32 preg); // Returns a cached register
 	bool IsImm(u32 preg) { return regs[preg].GetType() == REG_IMM; }
 	u32 GetImm(u32 preg) { return regs[preg].GetImm(); }