JIT: handle STR post with rd == rn

fixes Zelda Four Swords

2 files changed, 19 insertions, 3 deletions

diff --git a/src/ARMJIT_x64/ARMJIT_Compiler.cpp b/src/ARMJIT_x64/ARMJIT_Compiler.cpp
index cc4ad80..1eb5a2a 100644
--- a/src/ARMJIT_x64/ARMJIT_Compiler.cpp
+++ b/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);
+                    }
                 }
                 
             }
diff --git a/src/ARMJIT_x64/ARMJIT_LoadStore.cpp b/src/ARMJIT_x64/ARMJIT_LoadStore.cpp
index d80b25b..c044e69 100644
--- a/src/ARMJIT_x64/ARMJIT_LoadStore.cpp
+++ b/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;