make literal optimisation more reliable

fixes spanish Pokemon HeartGold

4 files changed, 77 insertions, 14 deletions

diff --git a/src/ARMJIT.cpp b/src/ARMJIT.cpp
index 0695b85..c7387c9 100644
--- a/src/ARMJIT.cpp
+++ b/src/ARMJIT.cpp
@@ -161,6 +161,27 @@ void FloodFillSetFlags(FetchedInstr instrs[], int start, u8 flags)
 	}
 }
 
+bool DecodeLiteral(const FetchedInstr& instr, u32& addr)
+{
+	switch (instr.Info.Kind)
+	{
+	case ARMInstrInfo::ak_STR_IMM:
+	case ARMInstrInfo::ak_STRB_IMM:
+		addr = (instr.Addr + 8) + ((instr.Instr & 0xFFF) * (instr.Instr & (1 << 23) ? 1 : -1));
+		return true;
+	case ARMInstrInfo::ak_STRD_IMM:
+	case ARMInstrInfo::ak_STRH_IMM:
+		addr = (instr.Addr + 8) + (((instr.Instr & 0xF00) >> 4 | (instr.Instr & 0xF)) * (instr.Instr & (1 << 23) ? 1 : -1));
+		return true;
+	case ARMInstrInfo::ak_STM: // I honestly hope noone was ever crazy enough to do stm pc, {whatever}
+		addr = instr.Addr + 8;
+		return true;
+	default:
+		JIT_DEBUGPRINT("Literal %08x %x not recognised\n", instr.Instr, instr.Addr);
+		return false;
+	}
+}
+
 bool DecodeBranch(bool thumb, const FetchedInstr& instr, u32& cond, bool hasLink, u32 lr, bool& link, 
 	u32& linkAddr, u32& targetAddr)
 {
@@ -463,6 +484,23 @@ void CompileBlock(ARM* cpu)
 		instrs[i].DataCycles = cpu->DataCycles;
 		instrs[i].DataRegion = cpu->DataRegion;
 
+		if (instrs[i].Info.SpecialKind == ARMInstrInfo::special_WriteMem 
+			&& instrs[i].Info.SrcRegs == (1 << 15)
+			&& instrs[i].Info.DstRegs == 0)
+		{
+			assert (!thumb);
+
+			u32 addr;
+			if (DecodeLiteral(instrs[i], addr))
+			{
+				JIT_DEBUGPRINT("pc relative write detected\n");
+				u32 translatedAddr = cpu->Num == 0 ? TranslateAddr<0>(addr) : TranslateAddr<1>(addr);
+
+				ARMJIT::InvalidateByAddr(translatedAddr, false);
+				CodeRanges[translatedAddr / 512].InvalidLiterals |= (1 << ((translatedAddr & 0x1FF) / 16));
+			}
+		}
+
 		if (thumb && instrs[i].Info.Kind == ARMInstrInfo::tk_BL_LONG_2 && i > 0
 			&& instrs[i - 1].Info.Kind == ARMInstrInfo::tk_BL_LONG_1)
 		{
@@ -631,7 +669,7 @@ void CompileBlock(ARM* cpu)
 	JitBlocks.Add(block);
 }
 
-void InvalidateByAddr(u32 pseudoPhysical)
+void InvalidateByAddr(u32 pseudoPhysical, bool mayRestore)
 {
 	JIT_DEBUGPRINT("invalidating by addr %x\n", pseudoPhysical);
 	AddressRange* range = &CodeRanges[pseudoPhysical / 512];
@@ -657,11 +695,14 @@ void InvalidateByAddr(u32 pseudoPhysical)
 
 		FastBlockAccess[block->PseudoPhysicalAddr / 2] = NULL;
 
-		u32 slot = HashRestoreCandidate(block->PseudoPhysicalAddr);
-		if (RestoreCandidates[slot] && RestoreCandidates[slot] != block)
-			delete RestoreCandidates[slot];
+		if (mayRestore)
+		{
+			u32 slot = HashRestoreCandidate(block->PseudoPhysicalAddr);
+			if (RestoreCandidates[slot] && RestoreCandidates[slot] != block)
+				delete RestoreCandidates[slot];
 
-		RestoreCandidates[slot] = block;
+			RestoreCandidates[slot] = block;
+		}
 	}
 	if ((range->TimesInvalidated + 1) > range->TimesInvalidated)
 		range->TimesInvalidated++;
@@ -732,6 +773,7 @@ void ResetBlockCache()
 			u32 addr = block->AddressRanges()[j];
 			CodeRanges[addr / 512].Blocks.Clear();
 			CodeRanges[addr / 512].TimesInvalidated = 0;
+			CodeRanges[addr / 512].InvalidLiterals = 0;
 		}
 		delete block;
 	}
diff --git a/src/ARMJIT.h b/src/ARMJIT.h
index 1db4d66..09cc463 100644
--- a/src/ARMJIT.h
+++ b/src/ARMJIT.h
@@ -61,7 +61,7 @@ inline JitBlockEntry LookUpBlock(u32 addr)
 void Init();
 void DeInit();
 
-void InvalidateByAddr(u32 pseudoPhysical);
+void InvalidateByAddr(u32 pseudoPhysical, bool mayRestore = true);
 void InvalidateAll();
 
 void InvalidateITCM(u32 addr);
diff --git a/src/ARMJIT_Internal.h b/src/ARMJIT_Internal.h
index 9e6713d..fb05f75 100644
--- a/src/ARMJIT_Internal.h
+++ b/src/ARMJIT_Internal.h
@@ -63,7 +63,7 @@ struct __attribute__((packed)) TinyVector
 {
 	T* Data = NULL;
 	u16 Capacity = 0;
-	u32 Length = 0; // make it 32 bit so we don't need movzx
+	u16 Length = 0;
 
 	~TinyVector()
 	{
@@ -181,6 +181,7 @@ private:
 struct __attribute__((packed)) AddressRange
 {
 	TinyVector<JitBlock*> Blocks;
+	u16 InvalidLiterals;
 	u16 TimesInvalidated;
 };
 
diff --git a/src/ARMJIT_x64/ARMJIT_LoadStore.cpp b/src/ARMJIT_x64/ARMJIT_LoadStore.cpp
index 3799774..82f80a7 100644
--- a/src/ARMJIT_x64/ARMJIT_LoadStore.cpp
+++ b/src/ARMJIT_x64/ARMJIT_LoadStore.cpp
@@ -108,7 +108,7 @@ void* Compiler::Gen_MemoryRoutine9(bool store, int size)
         MOV(32, R(RSCRATCH), R(ABI_PARAM1));
         SHR(32, R(RSCRATCH), Imm8(9));
         SHL(32, R(RSCRATCH), Imm8(4));
-        CMP(32, MDisp(RSCRATCH, squeezePointer(CodeRanges) + offsetof(AddressRange, Blocks.Length)), Imm8(0));
+        CMP(16, MDisp(RSCRATCH, squeezePointer(CodeRanges) + offsetof(AddressRange, Blocks.Length)), Imm8(0));
         FixupBranch noCode = J_CC(CC_Z);
         JMP((u8*)InvalidateByAddr, true);
         SetJumpTarget(noCode);
@@ -206,7 +206,7 @@ void* Compiler::Gen_MemoryRoutineSeq9(bool store, bool preinc)
         MOV(32, R(ABI_PARAM4), R(RSCRATCH));
         SHR(32, R(RSCRATCH), Imm8(9));
         SHL(32, R(RSCRATCH), Imm8(4));
-        CMP(32, MDisp(RSCRATCH, squeezePointer(CodeRanges) + offsetof(AddressRange, Blocks.Length)), Imm8(0));
+        CMP(16, MDisp(RSCRATCH, squeezePointer(CodeRanges) + offsetof(AddressRange, Blocks.Length)), Imm8(0));
         FixupBranch noCode = J_CC(CC_Z);
         ABI_PushRegistersAndAdjustStack({ABI_PARAM1, ABI_PARAM2, ABI_PARAM3}, 8);
         MOV(32, R(ABI_PARAM1), R(ABI_PARAM4));
@@ -278,10 +278,10 @@ void Compiler::Comp_MemLoadLiteral(int size, int rd, u32 addr)
     Comp_AddCycles_CDI();
 }
 
-void fault(u32 a, u32 b)
+/*void fault(u32 a, u32 b, u32 c, u32 d)
 {
-    printf("actually not static! %x %x\n", a, b);
-}
+    printf("actually not static! %x %x %x %x\n", a, b, c, d);
+}*/
 
 void Compiler::Comp_MemAccess(int rd, int rn, const ComplexOperand& op2, int size, int flags)
 {
@@ -291,11 +291,17 @@ void Compiler::Comp_MemAccess(int rd, int rn, const ComplexOperand& op2, int siz
     if (size == 16)
         addressMask = ~1;
 
+    //bool check = false;
     if (Config::JIT_LiteralOptimisations && rn == 15 && rd != 15 && op2.IsImm && !(flags & (memop_SignExtend|memop_Post|memop_Store|memop_Writeback)))
     {
         u32 addr = R15 + op2.Imm * ((flags & memop_SubtractOffset) ? -1 : 1);
-        Comp_MemLoadLiteral(size, rd, addr);
-        return;
+        u32 translatedAddr = Num == 0 ? TranslateAddr<0>(addr) : TranslateAddr<1>(addr);
+
+        if (!(CodeRanges[translatedAddr / 512].InvalidLiterals & (1 << ((translatedAddr & 0x1FF) / 16))))
+        {
+            Comp_MemLoadLiteral(size, rd, addr);
+            return;
+        }
     }
 
     {
@@ -438,6 +444,20 @@ void Compiler::Comp_MemAccess(int rd, int rn, const ComplexOperand& op2, int siz
 
         CALL(memoryFunc);
 
+        /*if (Num == 0 && check)
+        {
+            CMP(32, R(EAX), rdMapped);
+            FixupBranch notEqual = J_CC(CC_E);
+            ABI_PushRegistersAndAdjustStack({RSCRATCH}, 0);
+            MOV(32, R(ABI_PARAM1), Imm32(R15 - (Thumb ? 4 : 8)));
+            MOV(32, R(ABI_PARAM2), R(EAX));
+            MOV(32, R(ABI_PARAM3), rdMapped);
+            MOV(32, R(ABI_PARAM4), Imm32(CurInstr.Instr));
+            CALL((u8*)fault);
+            ABI_PopRegistersAndAdjustStack({RSCRATCH}, 0);
+            SetJumpTarget(notEqual);
+        }*/
+
         if (!(flags & memop_Store))
         {
             if (inlinePreparation && size == 32)