From 9ed4c665920bc1ed5ea1ccf8452020d292f62ee2 Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Fri, 18 Oct 2024 17:56:56 -0400
Subject: [PATCH] add support for 1 reg ldm/stm a9 timings and fix a bug

---
 src/ARMInterpreter_LoadStore.cpp | 78 +++++++++++++++++++++++++-------
 src/CP15.cpp                     |  4 +-
 2 files changed, 63 insertions(+), 19 deletions(-)

diff --git a/src/ARMInterpreter_LoadStore.cpp b/src/ARMInterpreter_LoadStore.cpp
index 93ea54dc..b965c078 100644
--- a/src/ARMInterpreter_LoadStore.cpp
+++ b/src/ARMInterpreter_LoadStore.cpp
@@ -617,6 +617,12 @@ void A_LDM(ARM* cpu)
         
     cpu->AddCycles_CDI();
 
+    if (__builtin_popcount(cpu->CurInstr & 0xFFFF) == 1) [[unlikely]] // single reg
+    {
+        if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+        else; // CHECKME: ARM7 timing behavior?
+    }
+
     // handle data aborts
     if (dabort) [[unlikely]]
     {
@@ -735,6 +741,12 @@ void A_STM(ARM* cpu)
 
     cpu->AddCycles_CD();
 
+    if (__builtin_popcount(cpu->CurInstr & 0xFFFF) == 1) [[unlikely]] // single reg
+    {
+        if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+        else; // CHECKME: ARM7 timing behavior?
+    }
+
     // handle data aborts
     if (dabort) [[unlikely]]
     {
@@ -909,6 +921,12 @@ void T_PUSH(ARM* cpu)
 
     cpu->AddCycles_CD();
 
+    if (__builtin_popcount(cpu->CurInstr & 0x1FF) == 1) [[unlikely]] // single reg
+    {
+        if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+        else; // CHECKME: ARM7 timing behavior?
+    }
+
     if (dabort) [[unlikely]]
     {
         ((ARMv5*)cpu)->DataAbort();
@@ -952,33 +970,47 @@ void T_POP(ARM* cpu)
         dabort |= !(first ? cpu->DataRead32 (base, &pc)
                           : cpu->DataRead32S(base, &pc));
 
-        if (dabort) [[unlikely]] goto dataabort;
+        if (!dabort) [[likely]]
+        {
+            cpu->AddCycles_CDI();
 
-        cpu->AddCycles_CDI();
-        if (cpu->Num==1 || (((ARMv5*)cpu)->CP15Control & (1<<15))) pc |= 0x1;
-        cpu->JumpTo(pc);
-        base += 4;
+            if (__builtin_popcount(cpu->CurInstr & 0x1FF) == 1) [[unlikely]] // single reg
+            {
+                if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+                else; // CHECKME: ARM7 timing behavior?
+            }
+
+            if (cpu->Num==1 || (((ARMv5*)cpu)->CP15Control & (1<<15))) pc |= 0x1;
+            cpu->JumpTo(pc);
+            base += 4;
+        }
     }
     else
     {
         cpu->AddCycles_CDI();
 
+        if (__builtin_popcount(cpu->CurInstr & 0x1FF) == 1) [[unlikely]] // single reg
+        {
+            if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+            else; // CHECKME: ARM7 timing behavior?
+        }
+
         if (cpu->Num == 0)
         {
-            u8 lastreg = 31 - __builtin_clz(cpu->CurInstr & 0xFF);
-            ((ARMv5*)cpu)->ILCurrReg = lastreg;
-            ((ARMv5*)cpu)->ILCurrTime = ((ARMv5*)cpu)->TimestampActual;
+            if (dabort) [[unlikely]]
+            {
+                ((ARMv5*)cpu)->DataAbort();
+                return;
+            }
+            else
+            {
+                u8 lastreg = 31 - __builtin_clz(cpu->CurInstr & 0xFF);
+                ((ARMv5*)cpu)->ILCurrReg = lastreg;
+                ((ARMv5*)cpu)->ILCurrTime = ((ARMv5*)cpu)->TimestampActual;
+            }
         }
     }
 
-    if (dabort) [[unlikely]]
-    {
-        dataabort:
-        cpu->AddCycles_CDI();
-        ((ARMv5*)cpu)->DataAbort();
-        return;
-    }
-
     cpu->R[13] = base;
 }
 
@@ -1015,6 +1047,12 @@ void T_STMIA(ARM* cpu)
 
     cpu->AddCycles_CD();
 
+    if (__builtin_popcount(cpu->CurInstr & 0xFF) == 1) [[unlikely]] // single reg
+    {
+        if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+        else; // CHECKME: ARM7 timing behavior?
+    }
+
     if (dabort) [[unlikely]]
     {
         ((ARMv5*)cpu)->DataAbort();
@@ -1054,12 +1092,18 @@ void T_LDMIA(ARM* cpu)
 
     cpu->AddCycles_CDI();
 
+    if (__builtin_popcount(cpu->CurInstr & 0xFF) == 1) [[unlikely]] // single reg
+    {
+        if (cpu->Num == 0) cpu->NDS.ARM9Timestamp = ((ARMv5*)cpu)->TimestampActual; // on arm9 single reg ldm/stm cannot overlap memory and fetch stages
+        else; // CHECKME: ARM7 timing behavior?
+    }
+
     if (dabort) [[unlikely]]
     {
         ((ARMv5*)cpu)->DataAbort();
         return;
     }
-    
+
     if (cpu->Num == 0)
     {
         u8 lastreg = 31 - __builtin_clz(cpu->CurInstr & 0xFF);
diff --git a/src/CP15.cpp b/src/CP15.cpp
index 44b7e88a..6c9c7951 100644
--- a/src/CP15.cpp
+++ b/src/CP15.cpp
@@ -321,7 +321,7 @@ void ARMv5::UpdateRegionTimings(u32 addrstart, u32 addrend)
         {
             MemTimings[i][1] = ((bustimings[0] - 1) << NDS.ARM9ClockShift) + 1;
             MemTimings[i][2] = ((bustimings[2] - 1) << NDS.ARM9ClockShift) + 1;
-            MemTimings[i][3] = bustimings[3] << NDS.ARM9ClockShift; // inaccurate but ehgh
+            MemTimings[i][3] = ((bustimings[3] - 1) << NDS.ARM9ClockShift) + 1;; // inaccurate but ehgh
         }
     }
 }
@@ -1367,7 +1367,7 @@ bool ARMv5::DataWrite32S(u32 addr, u32 val)
 
         if ((addr >> 24) == 0x02)
         {
-            if ((DataRegion != Mem9_MainRAM) && ((NDS.ARM9Timestamp + DataCycles) < MainRAMTimestamp)) NDS.ARM9Timestamp = MainRAMTimestamp - DataCycles;
+            if ((DataRegion != Mem9_MainRAM) && ((NDS.ARM9Timestamp + DataCycles) < MainRAMTimestamp)) DataCycles += MainRAMTimestamp - NDS.ARM9Timestamp;
             MainRAMTimestamp = NDS.ARM9Timestamp + DataCycles;
             DataRegion = Mem9_MainRAM;
         }