JIT: handle STR post with rd == rn

fixes Zelda Four Swords

src/ARMJIT_A64/ARMJIT_Compiler.cpp

@@ -174,10 +174,13 @@ void Compiler::PopRegs(bool saveHiRegs)
 {
     if (saveHiRegs)
     {
-        BitSet16 hiRegsLoaded(RegCache.LoadedRegs & 0x7F00);
+        if (!Thumb && CurInstr.Cond() != 0xE)
+        {
+            BitSet16 hiRegsLoaded(RegCache.LoadedRegs & 0x7F00);
 
-        for (int reg : hiRegsLoaded)
+            for (int reg : hiRegsLoaded)
-            LoadReg(reg, RegCache.Mapping[reg]);
+                LoadReg(reg, RegCache.Mapping[reg]);
+        }
     }
 }
 
@@ -326,9 +329,10 @@ Compiler::Compiler()
         {
             for (int size = 0; size < 3; size++)
             {
-                for (int reg = 0; reg < 8; reg++)
+                for (int reg = 0; reg < 32; reg++)
                 {
-                    ARM64Reg rdMapped = (ARM64Reg)(W19 + reg);
+                    if (!(reg == W4 || (reg >= W19 && reg <= W26)))
+                        continue;
                     PatchedStoreFuncs[consoleType][num][size][reg] = GetRXPtr();
                     if (num == 0)
                     {
@@ -711,7 +715,9 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                 QuickCallFunction(X1, InterpretTHUMB[CurInstr.Info.Kind]);
             }
             else
+            {
                 (this->*comp)();
+            }
         }
         else
         {
@@ -727,10 +733,12 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                 }
             }
             else if (cond == 0xF)
+            {
                 Comp_AddCycles_C();
+            }
             else
             {
-                IrregularCycles = false;
+                IrregularCycles = comp == NULL;
 
                 FixupBranch skipExecute;
                 if (cond < 0xE)
@@ -763,7 +771,9 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                         SetJumpTarget(skipNop);
                     }
                     else
+                    {
                         SetJumpTarget(skipExecute);
+                    }
                 }
 
             }

src/ARMJIT_A64/ARMJIT_Compiler.h

@@ -241,17 +241,8 @@ public:
     u32 JitMemSecondarySize;
     u32 JitMemMainSize;
 
-    void* ReadBanked, *WriteBanked;
-
-    void* JumpToFuncs9[3];
-    void* JumpToFuncs7[3];
-
     std::unordered_map<ptrdiff_t, LoadStorePatch> LoadStorePatches; 
 
-    // [Console Type][Num][Size][Sign Extend][Output register]
-    void* PatchedLoadFuncs[2][2][3][2][8];
-    void* PatchedStoreFuncs[2][2][3][8];
-
     RegisterCache<Compiler, Arm64Gen::ARM64Reg> RegCache;
 
     bool CPSRDirty = false;
@@ -263,6 +254,15 @@ public:
     void* JitRWStart;
     void* JitRXStart;
 #endif
+
+    void* ReadBanked, *WriteBanked;
+
+    void* JumpToFuncs9[3];
+    void* JumpToFuncs7[3];
+
+    // [Console Type][Num][Size][Sign Extend][Output register]
+    void* PatchedLoadFuncs[2][2][3][2][32];
+    void* PatchedStoreFuncs[2][2][3][32];
 };
 
 }

src/ARMJIT_A64/ARMJIT_LoadStore.cpp

@@ -116,6 +116,12 @@ void Compiler::Comp_MemAccess(int rd, int rn, Op2 offset, int size, int flags)
         rnMapped = W3;
     }
 
+    if (flags & memop_Store && flags & (memop_Post|memop_Writeback) && rd == rn)
+    {
+        MOV(W4, rdMapped);
+        rdMapped = W4;
+    }
+
     ARM64Reg finalAddr = W0;
     if (flags & memop_Post)
     {
@@ -170,10 +176,10 @@ void Compiler::Comp_MemAccess(int rd, int rn, Op2 offset, int size, int flags)
         ptrdiff_t memopStart = GetCodeOffset();
         LoadStorePatch patch;
 
+        assert((rdMapped >= W19 && rdMapped <= W26) || rdMapped == W4);
         patch.PatchFunc = flags & memop_Store
-            ? PatchedStoreFuncs[NDS::ConsoleType][Num][__builtin_ctz(size) - 3][rdMapped - W19]
+            ? PatchedStoreFuncs[NDS::ConsoleType][Num][__builtin_ctz(size) - 3][rdMapped]
-            : PatchedLoadFuncs[NDS::ConsoleType][Num][__builtin_ctz(size) - 3][!!(flags & memop_SignExtend)][rdMapped - W19];
+            : PatchedLoadFuncs[NDS::ConsoleType][Num][__builtin_ctz(size) - 3][!!(flags & memop_SignExtend)][rdMapped];
-        assert(rdMapped - W19 >= 0 && rdMapped - W19 < 8);
 
         MOVP2R(X7, Num == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start);
 

src/ARMJIT_x64/ARMJIT_Compiler.cpp

@@ -98,7 +98,9 @@ void Compiler::A_Comp_MRS()
         MOV(32, rd, R(RSCRATCH3));
     }
     else
+    {
         MOV(32, rd, R(RCPSR));
+    }
 }
 
 void UpdateModeTrampoline(ARM* arm, u32 oldmode, u32 newmode)
@@ -703,7 +705,9 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                 ABI_CallFunction(InterpretTHUMB[CurInstr.Info.Kind]);
             }
             else
+            {
                 (this->*comp)();
+            }
         }
         else
         {
@@ -724,7 +728,7 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
             }
             else
             {
-                IrregularCycles = false;
+                IrregularCycles = comp == NULL;
 
                 FixupBranch skipExecute;
                 if (cond < 0xE)
@@ -737,7 +741,9 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                     ABI_CallFunction(InterpretARM[CurInstr.Info.Kind]);
                 }
                 else
+                {
                     (this->*comp)();
+                }
 
                 Comp_SpecialBranchBehaviour(true);
 
@@ -755,7 +761,9 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                         SetJumpTarget(skipFailed);
                     }
                     else
+                    {
                         SetJumpTarget(skipExecute);
+                    }
                 }
                 
             }

src/ARMJIT_x64/ARMJIT_LoadStore.cpp

@@ -130,6 +130,12 @@ void Compiler::Comp_MemAccess(int rd, int rn, const Op2& op2, int size, int flag
     if (Thumb && rn == 15)
         rnMapped = Imm32(R15 & ~0x2);
 
+    if (flags & memop_Store && flags & (memop_Post|memop_Writeback) && rd == rn)
+    {
+        MOV(32, R(RSCRATCH4), rdMapped);
+        rdMapped = R(RSCRATCH4);
+    }
+
     X64Reg finalAddr = RSCRATCH3;
     if (flags & memop_Post)
     {
@@ -282,13 +288,15 @@ void Compiler::Comp_MemAccess(int rd, int rn, const Op2& op2, int size, int flag
         {
             if (Num == 0)
             {
+                // on Windows param 3 is R8 which is also scratch 4 which can be used for rd
+                if (flags & memop_Store)
+                    MOV(32, R(ABI_PARAM3), rdMapped);
+
                 MOV(64, R(ABI_PARAM2), R(RCPU));
                 if (ABI_PARAM1 != RSCRATCH3)
                     MOV(32, R(ABI_PARAM1), R(RSCRATCH3));
                 if (flags & memop_Store)
                 {
-                    MOV(32, R(ABI_PARAM3), rdMapped);
-
                     switch (size | NDS::ConsoleType)
                     {
                     case 32: CALL((void*)&SlowWrite9<u32, 0>); break;