it all makes sense now...

src/ARM.cpp

@@ -286,47 +286,6 @@ void ARM::SetupCodeMem(u32 addr)
     }
 }
 
-void ARMv5::BuggedJumpTo32(const u32 addr)
-{
-    // ldrd to pc
-    // behavior seems to be related to if a bugged 8/16 bit write has prefetch aborted (does any p.abort work?)
-    // switching to thumb mode only seems to work the first time an ldrd pc is executed after one of the above aborts?
-    // also it can restore cpsr but only if the PU is disabled (?????????????????????????????????????)
-    if (BuggyJump == 1)
-    {
-        BuggyJump = 2;
-
-        if (CP15Control & (1<<15))
-            JumpTo(addr & ~0x1, !(CP15Control & 1));
-        else
-            JumpTo(addr, !(CP15Control & 1));
-    }
-    else
-    {
-        JumpTo(addr & ~0x1, !(CP15Control & 1));
-    }
-}
-
-void ARMv5::BuggedJumpTo(const u32 addr)
-{
-    // 16 and 8 bit loads (signed instructions included) to pc
-    // if they're misaligned they'll prefetch abort
-    // but they can only prefetch abort once, every time afterwards will succeed (more testing needed)
-    // if the lsb is set they will try to switch to thumb state, though it'll fail if they haven't prefetch aborted yet
-    if ((BuggyJump == 0) && (addr & 0x3))
-    {
-        if (addr & 0x1) CPSR |= 0x20;
-        BuggyJump = 1;
-        PrefetchAbort();
-        return;
-    }
-
-    if (CP15Control & (1<<15))
-        JumpTo(addr & ~0x1);
-    else
-        JumpTo(addr);
-}
-
 void ARMv5::JumpTo(u32 addr, bool restorecpsr)
 {
     if (restorecpsr)
@@ -395,14 +354,25 @@ void ARMv5::JumpTo(u32 addr, bool restorecpsr)
     NDS.MonitorARM9Jump(addr);
 }
 
-void ARMv4::BuggedJumpTo32(const u32 addr)
+void ARMv5::JumpTo8_16Bit(const u32 addr)
 {
-    JumpTo(addr & ~1); // todo
+    // 8 and 16 loads (signed included) to pc
+    if (!(CP15Control & 0x1))
+    {
+        // if the pu is disabled it behaves like a normal jump
+        JumpTo((CP15Control & (1<<15)) ? (addr & ~0x1) : addr);
     }
-
+    else
-void ARMv4::BuggedJumpTo(const u32 addr)
     {
-    JumpTo(addr & ~1); // todo
+        if (addr & 0x3)
+        {
+            // if the pu is enabled it will always prefetch abort if not word aligned
+            // although it will still attempt (and fail) to enter thumb mode if enabled
+            if ((addr & 0x1) && !(CP15Control & (1<<15))) CPSR |= 0x20;
+            PrefetchAbort();
+        }
+        else JumpTo(addr);
+    }
 }
 
 void ARMv4::JumpTo(u32 addr, bool restorecpsr)
@@ -449,6 +419,11 @@ void ARMv4::JumpTo(u32 addr, bool restorecpsr)
     }
 }
 
+void ARMv4::JumpTo8_16Bit(const u32 addr)
+{
+    JumpTo(addr & ~1); // checkme?
+}
+
 void ARM::RestoreCPSR()
 {
     u32 oldcpsr = CPSR;

src/ARM.h

@@ -65,9 +65,8 @@ public:
 
     virtual void FillPipeline() = 0;
     
-    virtual void BuggedJumpTo32(const u32 addr) = 0;
-    virtual void BuggedJumpTo(const u32 addr) = 0;
     virtual void JumpTo(u32 addr, bool restorecpsr = false) = 0;
+    virtual void JumpTo8_16Bit(u32 addr) = 0;
     void RestoreCPSR();
 
     void Halt(u32 halt)
@@ -239,9 +238,8 @@ public:
 
     void FillPipeline() override;
     
-    void BuggedJumpTo32(const u32 addr) override;
-    void BuggedJumpTo(const u32 addr) override;
     void JumpTo(u32 addr, bool restorecpsr = false) override;
+    void JumpTo8_16Bit(const u32 addr) override;
 
     void PrefetchAbort();
     void DataAbort();
@@ -386,9 +384,8 @@ public:
 
     void FillPipeline() override;
     
-    void BuggedJumpTo32(const u32 addr) override;
-    void BuggedJumpTo(const u32 addr) override;
     void JumpTo(u32 addr, bool restorecpsr = false) override;
+    void JumpTo8_16Bit(const u32 addr) override;
 
     void Execute() override;
 #ifdef JIT_ENABLED

src/ARMInterpreter_LoadStore.cpp

@@ -141,7 +141,7 @@ namespace melonDS::ARMInterpreter
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
     if (cpu->CurInstr & (1<<21)) cpu->R[(cpu->CurInstr>>16) & 0xF] = offset; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val;
 
 // TODO: user mode
@@ -151,7 +151,7 @@ namespace melonDS::ARMInterpreter
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
     cpu->R[(cpu->CurInstr>>16) & 0xF] += offset; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val;
 
 
@@ -262,7 +262,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     if (r&1) { A_UNK(cpu); return; } /* checkme */ \
     if (!cpu->DataRead32 (offset  , &cpu->R[r  ])) {cpu->AddCycles_CDI(); return;} \
     u32 val; if (!cpu->DataRead32S(offset+4, &val)) {cpu->AddCycles_CDI(); return;} \
-    if (r == 14) cpu->BuggedJumpTo32(val); \
+    if (r == 14) cpu->JumpTo(((((ARMv5*)cpu)->CP15Control & (1<<15)) ? (val & ~0x1) : val), true); /* restores cpsr for some reason? */ \
     else cpu->R[r+1] = val; \
     cpu->AddCycles_CDI(); \
     if (cpu->CurInstr & (1<<21)) cpu->R[(cpu->CurInstr>>16) & 0xF] = offset;
@@ -274,7 +274,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     if (r&1) { A_UNK(cpu); return; } /* checkme */ \
     if (!cpu->DataRead32 (addr  , &cpu->R[r  ])) {cpu->AddCycles_CDI(); return;} \
     u32 val; if (!cpu->DataRead32S(addr+4, &val)) {cpu->AddCycles_CDI(); return;} \
-    if (r == 14) cpu->BuggedJumpTo32(val); \
+    if (r == 14) cpu->JumpTo(((((ARMv5*)cpu)->CP15Control & (1<<15)) ? (val & ~0x1) : val), true); /* restores cpsr for some reason? */ \
     else cpu->R[r+1] = val; \
     cpu->AddCycles_CDI(); \
     cpu->R[(cpu->CurInstr>>16) & 0xF] += offset;
@@ -308,7 +308,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     u32 val; bool dataabort = !cpu->DataRead16(offset, &val); \
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val; \
     if (cpu->CurInstr & (1<<21)) cpu->R[(cpu->CurInstr>>16) & 0xF] = offset;
 
@@ -317,7 +317,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     u32 val; bool dataabort = !cpu->DataRead16(addr, &val); \
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val; \
     cpu->R[(cpu->CurInstr>>16) & 0xF] += offset;
 
@@ -327,7 +327,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
     val = (s32)(s8)val; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val; \
     if (cpu->CurInstr & (1<<21)) cpu->R[(cpu->CurInstr>>16) & 0xF] = offset;
 
@@ -337,7 +337,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
     val = (s32)(s8)val; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val; \
     cpu->R[(cpu->CurInstr>>16) & 0xF] += offset;
 
@@ -347,7 +347,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
     val = (s32)(s16)val; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val; \
     if (cpu->CurInstr & (1<<21)) cpu->R[(cpu->CurInstr>>16) & 0xF] = offset;
 
@@ -357,7 +357,7 @@ A_IMPLEMENT_WB_LDRSTR(LDRB)
     cpu->AddCycles_CDI(); \
     if (dataabort) return; \
     val = (s32)(s16)val; \
-    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->BuggedJumpTo(val); \
+    if (((cpu->CurInstr>>12) & 0xF) == 15) cpu->JumpTo8_16Bit(val); \
     else cpu->R[(cpu->CurInstr>>12) & 0xF] = val; \
     cpu->R[(cpu->CurInstr>>16) & 0xF] += offset;