diff --git a/src/ARM.cpp b/src/ARM.cpp
index b2b10c63..96e97ce3 100644
--- a/src/ARM.cpp
+++ b/src/ARM.cpp
@@ -331,7 +331,7 @@ void ARMv5::JumpTo(u32 addr, bool restorecpsr, u8 R15)
 {
     //printf("JUMP! %08X %i %i\n", addr, restorecpsr, R15);
     NDS.MonitorARM9Jump(addr);
-    
+
     BranchRestore = restorecpsr;
     BranchUpdate = R15;
     BranchAddr = addr;
@@ -382,15 +382,13 @@ void ARMv5::JumpTo_2()
         // doesn't matter if we put garbage in the MSbs there
         if (BranchAddr & 0x2)
         {
+            DelayedQueue = &ARMv5::JumpTo_3A;
             CodeRead32(BranchAddr-2);
-
-            QueueFunction(&ARMv5::JumpTo_3A);
         }
         else
         {
+            DelayedQueue = &ARMv5::JumpTo_3B;
             CodeRead32(BranchAddr);
-
-            QueueFunction(&ARMv5::JumpTo_3B);
         }
     }
     else
@@ -399,19 +397,17 @@ void ARMv5::JumpTo_2()
         R[15] = BranchAddr+4;
 
         CPSR &= ~0x20;
-
+        
+        DelayedQueue = &ARMv5::JumpTo_3C;
         CodeRead32(BranchAddr);
-
-        QueueFunction(&ARMv5::JumpTo_3C);
     }
 }
 
 void ARMv5::JumpTo_3A()
 {
     NextInstr[0] = RetVal >> 16;
+    DelayedQueue = &ARMv5::JumpTo_4;
     CodeRead32(BranchAddr+2);
-
-    QueueFunction(&ARMv5::JumpTo_4);
 }
 
 void ARMv5::JumpTo_3B()
@@ -423,9 +419,8 @@ void ARMv5::JumpTo_3B()
 void ARMv5::JumpTo_3C()
 {
     NextInstr[0] = RetVal;
+    DelayedQueue = &ARMv5::JumpTo_4;
     CodeRead32(BranchAddr+4);
-
-    QueueFunction(&ARMv5::JumpTo_4);
 }
 
 void ARMv5::JumpTo_4()
@@ -1377,12 +1372,13 @@ void ARMv5::CodeFetch()
         if (NDS.ARM9Timestamp < TimestampMemory) NDS.ARM9Timestamp = TimestampMemory;
         Store = false;
         DataRegion = Mem9_Null;
+        QueueFunction(&ARMv5::AddExecute);
     }
     else
     {
+        DelayedQueue = &ARMv5::AddExecute;
         CodeRead32(PC);
     }
-    QueueFunction(&ARMv5::AddExecute);
 }
 
 void ARMv5::AddExecute()
diff --git a/src/ARM.h b/src/ARM.h
index 80b492cb..c23c2254 100644
--- a/src/ARM.h
+++ b/src/ARM.h
@@ -839,7 +839,7 @@ public:
     u64 ITCMTimestamp;
     u64 TimestampMemory;
     void (ARMv5::*FuncQueue[32])(void);
-    void (ARMv5::*DelayedQueue)(void);
+    void (ARMv5::*DelayedQueue)(void); // adding more than one new entry to the queue while it's already active does not work. so uh. we use this to work around that. it's less than ideal...
     u32 PC;
     bool NullFetch;
     bool Store;
diff --git a/src/CP15.cpp b/src/CP15.cpp
index c19c32e9..c7b076cb 100644
--- a/src/CP15.cpp
+++ b/src/CP15.cpp
@@ -443,6 +443,7 @@ bool ARMv5::ICacheLookup(const u32 addr)
             Store = false;
 
             RetVal = cacheLine[(addr & (ICACHE_LINELENGTH -1)) / 4];
+            QueueFunction(DelayedQueue);
             return true;
         }
     }
@@ -554,6 +555,7 @@ void ARMv5::ICacheLookup_2()
     }
     Store = false;
     DataRegion = Mem9_Null;
+    QueueFunction(DelayedQueue);
 }
 
 void ARMv5::ICacheInvalidateByAddr(const u32 addr)
@@ -2181,6 +2183,7 @@ void ARMv5::CodeRead32(u32 addr)
         DataRegion = Mem9_Null;
         Store = false;
         RetVal = ((u64)1<<63);
+        QueueFunction(DelayedQueue);
         return;
     }
 
@@ -2192,6 +2195,7 @@ void ARMv5::CodeRead32(u32 addr)
         DataRegion = Mem9_Null;
         Store = false;
         RetVal = *(u32*)&ITCM[addr & (ITCMPhysicalSize - 1)];
+        QueueFunction(DelayedQueue);
         return;
     }
     
@@ -2253,6 +2257,7 @@ void ARMv5::CodeRead32_2()
 
     Store = false;
     DataRegion = Mem9_Null;
+    QueueFunction(DelayedQueue);
     return;
 }