[ARM] Fastmem is back. Still disabled for Android.

Source/Core/Core/Src/PowerPC/JitArm32/Jit.h

@@ -197,6 +197,7 @@ class JitArm : public JitBase, public ArmGen::ARMXCodeBlock
 	void sthu(UGeckoInstruction _inst);
 	void stw(UGeckoInstruction _inst);
 	void stwu(UGeckoInstruction _inst);
+	void stwx(UGeckoInstruction _inst);
 
 	// Floating point
 	void fabsx(UGeckoInstruction _inst);
@@ -211,6 +212,7 @@ class JitArm : public JitBase, public ArmGen::ARMXCodeBlock
 	// Floating point loadStore
 	void lfs(UGeckoInstruction _inst);
 	void lfd(UGeckoInstruction _inst);
+	void stfs(UGeckoInstruction _inst);
 
 	// Paired Singles
 	void ps_add(UGeckoInstruction _inst);

Source/Core/Core/Src/PowerPC/JitArm32/JitArm_FloatingPoint.cpp

@@ -107,21 +107,19 @@ void JitArm::fsubx(UGeckoInstruction inst)
 	if (inst.Rc) Helper_UpdateCR1(vD);
 }
 
-// Breaks Animal Crossing
 void JitArm::fmulsx(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
 	JITDISABLE(FloatingPoint)
 
-	Default(inst); return;
-
 	ARMReg vA = fpr.R0(inst.FA);
 	ARMReg vC = fpr.R0(inst.FC);
 	ARMReg vD0 = fpr.R0(inst.FD);
 	ARMReg vD1 = fpr.R1(inst.FD);
 
 	VMUL(vD0, vA, vC);
 	VMOV(vD1, vD0);
+	fpr.Flush(); // Shouldn't be needed. Missing a flush somewhere
 	if (inst.Rc) Helper_UpdateCR1(vD0);
 }
 void JitArm::fmulx(UGeckoInstruction inst)

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

@@ -34,7 +34,7 @@
 #ifdef ANDROID
 #define FASTMEM 0
 #else
-#define FASTMEM 0
+#define FASTMEM 1
 #endif
 void JitArm::stb(UGeckoInstruction inst)
 {
@@ -282,6 +282,40 @@ void JitArm::stwu(UGeckoInstruction inst)
 
 	gpr.Unlock(ValueReg, Addr, Function);
 }
+void JitArm::stwx(UGeckoInstruction inst)
+{
+	INSTRUCTION_START
+	JITDISABLE(LoadStore)
+	u32 a = inst.RA, b = inst.RB, s = inst.RB;	
+
+	if (a) {
+		// Doesn't work
+		Default(inst); return;
+	}
+
+	ARMReg RB = gpr.R(b);
+	ARMReg RS = gpr.R(s);
+	ARMReg ValueReg = gpr.GetReg();
+	ARMReg Addr = gpr.GetReg();
+	ARMReg Function = gpr.GetReg();
+
+	if (a)
+		ADD(Addr, gpr.R(a), RB);
+	else
+		MOV(Addr, RB);
+
+	MOV(ValueReg, RS);
+	fpr.Flush();	
+	MOVI2R(Function, (u32)&Memory::Write_U32);	
+	PUSH(4, R0, R1, R2, R3);
+	MOV(R0, ValueReg);
+	MOV(R1, Addr);
+	BL(Function);
+	POP(4, R0, R1, R2, R3);
+
+	gpr.Unlock(ValueReg, Addr, Function);
+}
+
 void JitArm::StoreFromReg(ARMReg dest, ARMReg value, int accessSize, s32 offset)
 {
 	ARMReg rA = gpr.GetReg();
@@ -417,7 +451,7 @@ void JitArm::lhz(UGeckoInstruction inst)
 	LDR(rA, R9, PPCSTATE_OFF(Exceptions));
 	CMP(rA, EXCEPTION_DSI);
 	FixupBranch DoNotLoad = B_CC(CC_EQ);
-#if 0 // FASTMEM
+#if FASTMEM
 	// Backpatch route
 	// Gets loaded in to RD
 	// Address is in R10

Source/Core/Core/Src/PowerPC/JitArm32/JitArm_LoadStoreFloating.cpp

@@ -39,6 +39,9 @@ void JitArm::lfs(UGeckoInstruction inst)
 
 	ARMReg rA = gpr.GetReg();
 	ARMReg rB = gpr.GetReg();
+
+	fpr.Flush();
+
 	LDR(rA, R9, PPCSTATE_OFF(Exceptions));
 	CMP(rA, EXCEPTION_DSI);
 	FixupBranch DoNotLoad = B_CC(CC_EQ);
@@ -52,7 +55,6 @@ void JitArm::lfs(UGeckoInstruction inst)
 	else
 		MOVI2R(rB, (u32)inst.SIMM_16);
 
-	fpr.Flush();
 
 	MOVI2R(rA, (u32)&Memory::Read_F32);	
 	PUSH(4, R0, R1, R2, R3);
@@ -80,6 +82,9 @@ void JitArm::lfd(UGeckoInstruction inst)
 
 	ARMReg rA = gpr.GetReg();
 	ARMReg rB = gpr.GetReg();
+
+	fpr.Flush();
+
 	LDR(rA, R9, PPCSTATE_OFF(Exceptions));
 	CMP(rA, EXCEPTION_DSI);
 	FixupBranch DoNotLoad = B_CC(CC_EQ);
@@ -93,7 +98,6 @@ void JitArm::lfd(UGeckoInstruction inst)
 	else
 		MOVI2R(rB, (u32)inst.SIMM_16);
 
-	fpr.Flush();
 
 	MOVI2R(rA, (u32)&Memory::Read_F64);	
 	PUSH(4, R0, R1, R2, R3);
@@ -112,3 +116,37 @@ void JitArm::lfd(UGeckoInstruction inst)
 	gpr.Unlock(rA, rB);
 	SetJumpTarget(DoNotLoad);
 }
+
+void JitArm::stfs(UGeckoInstruction inst)
+{
+	INSTRUCTION_START
+	JITDISABLE(LoadStoreFloating)
+
+	ARMReg rA = gpr.GetReg();
+	ARMReg rB = gpr.GetReg();
+	ARMReg v0 = fpr.R0(inst.FS);
+	VCVT(S0, v0, 0);
+	fpr.Flush();
+
+	if (inst.RA)
+	{
+		MOVI2R(rB, inst.SIMM_16);
+		ARMReg RA = gpr.R(inst.RA);
+		ADD(rB, rB, RA);
+	}
+	else
+		MOVI2R(rB, (u32)inst.SIMM_16);
+
+
+	MOVI2R(rA, (u32)&Memory::Write_U32);	
+	PUSH(4, R0, R1, R2, R3);
+	VMOV(R0, S0);
+	MOV(R1, rB);
+
+	BL(rA);
+
+	POP(4, R0, R1, R2, R3);
+
+	gpr.Unlock(rA, rB);
+}
+

Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp

@@ -102,7 +102,7 @@ static GekkoOPTemplate primarytable[] =
 	{50, &JitArm::lfd}, //"lfd",  OPTYPE_LOADFP, FL_IN_A}},
 	{51, &JitArm::Default}, //"lfdu", OPTYPE_LOADFP, FL_OUT_A | FL_IN_A}},
 
-	{52, &JitArm::Default}, //"stfs",  OPTYPE_STOREFP, FL_IN_A}},
+	{52, &JitArm::stfs}, //"stfs",  OPTYPE_STOREFP, FL_IN_A}},
 	{53, &JitArm::Default}, //"stfsu", OPTYPE_STOREFP, FL_OUT_A | FL_IN_A}},
 	{54, &JitArm::Default}, //"stfd",  OPTYPE_STOREFP, FL_IN_A}},
 	{55, &JitArm::Default}, //"stfdu", OPTYPE_STOREFP, FL_OUT_A | FL_IN_A}},
@@ -249,7 +249,7 @@ static GekkoOPTemplate table31[] =
 	{597, &JitArm::Default}, //"lswi",  OPTYPE_LOAD, FL_EVIL | FL_IN_AB | FL_OUT_D}},
 
 	//store word
-	{151, &JitArm::Default}, //"stwx",   OPTYPE_STORE, FL_IN_A0 | FL_IN_B}},
+	{151, &JitArm::stwx}, //"stwx",   OPTYPE_STORE, FL_IN_A0 | FL_IN_B}},
 	{183, &JitArm::Default}, //"stwux",  OPTYPE_STORE, FL_OUT_A | FL_IN_A | FL_IN_B}},
 
 	//store halfword
@@ -293,7 +293,7 @@ static GekkoOPTemplate table31[] =
 	{659, &JitArm::Default}, //"mfsrin", OPTYPE_SYSTEM, FL_OUT_D, 2}},
 
 	{4,   &JitArm::Break}, //"tw",     OPTYPE_SYSTEM, FL_ENDBLOCK, 1}},
-	{598, &JitArm::Default}, //"sync",   OPTYPE_SYSTEM, 0, 2}},
+	{598, &JitArm::DoNothing}, //"sync",   OPTYPE_SYSTEM, 0, 2}},
 	{982, &JitArm::icbi}, //"icbi",   OPTYPE_SYSTEM, FL_ENDBLOCK, 3}},
 
 	// Unused instructions on GC