diff --git a/desmume/src/GPU.h b/desmume/src/GPU.h
index 654a29b54..65960d307 100644
--- a/desmume/src/GPU.h
+++ b/desmume/src/GPU.h
@@ -139,11 +139,11 @@ static INLINE void GPU_ligne(Screen * screen, u16 l)
       * I'm really not sure it's correct.
       */
      if (gpu->lcd == 0) {
-        unsigned long mainlcdcnt = ((unsigned long *)ARM9Mem.ARM9_REG)[0];
+        u32 mainlcdcnt = ((u32 *)ARM9Mem.ARM9_REG)[0];
         int ii = l*256;
         if ((mainlcdcnt&0x10000)==0) {
            for (i=0; i<256; i++) {
-               ((unsigned short*)dst)[i] = ((unsigned short*)ARM9Mem.ARM9_LCD)[ii];
+               ((u16 *)dst)[i] = ((u16 *)ARM9Mem.ARM9_LCD)[ii];
                ii++;
            }
            return;
diff --git a/desmume/src/NDSSystem.c b/desmume/src/NDSSystem.c
index 8b54edce1..22667df44 100644
--- a/desmume/src/NDSSystem.c
+++ b/desmume/src/NDSSystem.c
@@ -85,7 +85,7 @@ BOOL NDS_SetROM(u8 * rom, u32 mask)
 
      MMU_setRom(rom, mask);
      
-     NDS_header * header = (NDS_header *)MMU.CART_ROM;
+     NDS_header * header = NDS_getROMHeader();
      
      u32 src = header->ARM9src;
      u32 dst = header->ARM9cpy;
@@ -193,7 +193,51 @@ BOOL NDS_SetROM(u8 * rom, u32 mask)
 
 NDS_header * NDS_getROMHeader(void)
 {
-     return (NDS_header *)MMU.CART_ROM;
+	NDS_header * header = malloc(sizeof(NDS_header));
+
+	memcpy(header->gameTile, MMU.CART_ROM, 12);
+	memcpy(header->gameCode, MMU.CART_ROM + 12, 4);
+	header->makerCode = T1ReadWord(MMU.CART_ROM, 16);
+	header->unitCode = MMU.CART_ROM[18];
+	header->deviceCode = MMU.CART_ROM[19];
+	header->cardSize = MMU.CART_ROM[20];
+	memcpy(header->cardInfo, MMU.CART_ROM + 21, 8);
+	header->flags = MMU.CART_ROM[29];
+	header->ARM9src = T1ReadLong(MMU.CART_ROM, 32);
+	header->ARM9exe = T1ReadLong(MMU.CART_ROM, 36);
+	header->ARM9cpy = T1ReadLong(MMU.CART_ROM, 40);
+	header->ARM9binSize = T1ReadLong(MMU.CART_ROM, 44);
+	header->ARM7src = T1ReadLong(MMU.CART_ROM, 48);
+	header->ARM7exe = T1ReadLong(MMU.CART_ROM, 52);
+	header->ARM7cpy = T1ReadLong(MMU.CART_ROM, 56);
+	header->ARM7binSize = T1ReadLong(MMU.CART_ROM, 60);
+	header->FNameTblOff = T1ReadLong(MMU.CART_ROM, 64);
+	header->FNameTblSize = T1ReadLong(MMU.CART_ROM, 68);
+	header->FATOff = T1ReadLong(MMU.CART_ROM, 72);
+	header->FATSize = T1ReadLong(MMU.CART_ROM, 76);
+	header->ARM9OverlayOff = T1ReadLong(MMU.CART_ROM, 80);
+	header->ARM9OverlaySize = T1ReadLong(MMU.CART_ROM, 84);
+	header->ARM7OverlayOff = T1ReadLong(MMU.CART_ROM, 88);
+	header->ARM7OverlaySize = T1ReadLong(MMU.CART_ROM, 92);
+	header->unknown2a = T1ReadLong(MMU.CART_ROM, 96);
+	header->unknown2b = T1ReadLong(MMU.CART_ROM, 100);
+	header->IconOff = T1ReadLong(MMU.CART_ROM, 104);
+	header->CRC16 = T1ReadWord(MMU.CART_ROM, 108);
+	header->ROMtimeout = T1ReadWord(MMU.CART_ROM, 110);
+	header->ARM9unk = T1ReadLong(MMU.CART_ROM, 112);
+	header->ARM7unk = T1ReadLong(MMU.CART_ROM, 116);
+	memcpy(header->unknown3c, MMU.CART_ROM + 120, 8);
+	header->ROMSize = T1ReadLong(MMU.CART_ROM, 128);
+	header->HeaderSize = T1ReadLong(MMU.CART_ROM, 132);
+	memcpy(header->unknown5, MMU.CART_ROM + 136, 56);
+	memcpy(header->logo, MMU.CART_ROM + 192, 156);
+	header->logoCRC16 = T1ReadWord(MMU.CART_ROM, 348);
+	header->headerCRC16 = T1ReadWord(MMU.CART_ROM, 350);
+	memcpy(header->reserved, MMU.CART_ROM + 352, 160);
+
+	return header;
+
+     //return (NDS_header *)MMU.CART_ROM;
 } 
 
 void NDS_setTouchPos(u16 x, u16 y)
diff --git a/desmume/src/NDSSystem.h b/desmume/src/NDSSystem.h
index b457934c7..a4e135dab 100644
--- a/desmume/src/NDSSystem.h
+++ b/desmume/src/NDSSystem.h
@@ -55,58 +55,6 @@ extern BOOL click;
 #define reg_IPCSYNC         (0x180>>1)
 #define reg_IPCFIFOCNT      (0x184>>1)
 
-#ifdef WORDS_BIGENDIAN
-typedef struct
-{
-       u8      reserved[160];
-       u16     headerCRC16;
-       u16     logoCRC16;
-       u8      logo[156];
-       u8      unknown5[56];
-       u32     HeaderSize;
-       u32     ROMSize;
-       u8      unknown3c[8];
-
-       u32     ARM7unk;
-       u32     ARM9unk;
-       u16     ROMtimeout;
-       u16     CRC16;
-       u32     IconOff;
-       
-       u32     unknown2b;
-       u32     unknown2a;
-       
-       u32     ARM7OverlaySize;
-       u32     ARM7OverlayOff;
-       u32     ARM9OverlaySize;
-       u32     ARM9OverlayOff;
-       
-       u32      FATSize;
-       u32      FATOff;
-       
-       u32      FNameTblSize;
-       u32      FNameTblOff;
-       
-       u32      ARM7binSize;
-       u32      ARM7cpy;
-       u32      ARM7exe;
-       u32      ARM7src;
-       
-       u32      ARM9binSize;
-       u32      ARM9cpy;
-       u32      ARM9exe;
-       u32      ARM9src;
-       
-       u8       flags;
-       u8       cardInfo[8];
-       u8       cardSize;
-       u8       deviceCode;
-       u8       unitCode;
-       u16      makerCode;
-       char     gameCode[4];
-       char     gameTile[12];
-} NDS_header;
-#else
 typedef struct
 {
        char     gameTile[12];
@@ -157,7 +105,6 @@ typedef struct
        u16     headerCRC16;
        u8      reserved[160];
 } NDS_header;
-#endif
 
 extern void debug();
 
@@ -805,6 +752,7 @@ int NDS_LoadFirmware(const char *filename);
                  if((MMU.reg_IF[1]&MMU.reg_IE[1]) && (MMU.reg_IME[1]))
                       if (armcpu_irqExeption(&NDS_ARM7))
                            nds.ARM7Cycle = nds.cycles;
+
             }
             return nds.cycles;
        }
diff --git a/desmume/src/armcpu.h b/desmume/src/armcpu.h
index 450f2fc1b..a1556cb7d 100644
--- a/desmume/src/armcpu.h
+++ b/desmume/src/armcpu.h
@@ -96,6 +96,25 @@ enum Mode
 	SYS = 0x1F
 };
 
+#ifdef WORDS_BIGENDIAN
+typedef union 
+{
+	struct 
+	{
+		u32 N : 1,
+		Z : 1, 
+		C : 1, 
+		V : 1,
+		Q : 1,
+		RAZ : 19,
+		I : 1,
+		F : 1,
+		T : 1,
+                mode : 5;
+	} bits;
+        u32 val;
+} Status_Reg;
+#else
 typedef union 
 {
 	struct 
@@ -113,6 +132,7 @@ typedef union
 	} bits;
         u32 val;
 } Status_Reg;
+#endif
 
 typedef void* armcp_t;
 
diff --git a/desmume/src/cli/main.c b/desmume/src/cli/main.c
index 4c4052010..07b3975f9 100644
--- a/desmume/src/cli/main.c
+++ b/desmume/src/cli/main.c
@@ -59,8 +59,6 @@ int main(int argc, char ** argv) {
 
 	surface = SDL_SetVideoMode(256, 384, 32, SDL_SWSURFACE);
 
-	
-	
 	while(!end) {
 		
 		
diff --git a/desmume/src/mem.h b/desmume/src/mem.h
index a428009c2..9d279e4c6 100644
--- a/desmume/src/mem.h
+++ b/desmume/src/mem.h
@@ -34,7 +34,7 @@ static INLINE u8 T1ReadByte(u8 * mem, u32 addr)
 static INLINE u16 T1ReadWord(u8 * mem, u32 addr)
 {
 #ifdef WORDS_BIGENDIAN
-   return (mem[addr] << 8) | mem[addr + 1];
+   return (mem[addr + 1] << 8) | mem[addr];
 #else
    return *((u16 *) (mem + addr));
 #endif
@@ -43,8 +43,8 @@ static INLINE u16 T1ReadWord(u8 * mem, u32 addr)
 static INLINE u32 T1ReadLong(u8 * mem, u32 addr)
 {
 #ifdef WORDS_BIGENDIAN
-   return (mem[addr] << 24 | mem[addr + 1] << 16 |
-           mem[addr + 2] << 8 | mem[addr + 3]);
+   return (mem[addr + 3] << 24 | mem[addr + 2] << 16 |
+           mem[addr + 1] << 8 | mem[addr]);
 #else
    return *((u32 *) (mem + addr));
 #endif
@@ -53,10 +53,10 @@ static INLINE u32 T1ReadLong(u8 * mem, u32 addr)
 static INLINE u64 T1ReadQuad(u8 * mem, u32 addr)
 {
 #ifdef WORDS_BIGENDIAN
-   return (mem[addr] << 56 | mem[addr + 1] << 48 |
-	   mem[addr + 2] << 40 | mem[addr + 3] << 32 |
-           mem[addr + 4] << 24 | mem[addr + 5] << 16 |
-           mem[addr + 6] << 8 | mem[addr + 7]);
+   return (mem[addr + 7] << 56 | mem[addr + 6] << 48 |
+	   mem[addr + 5] << 40 | mem[addr + 4] << 32 |
+           mem[addr + 3] << 24 | mem[addr + 2] << 16 |
+           mem[addr + 1] << 8 | mem[addr]);
 #else
    return *((u64 *) (mem + addr));
 #endif
@@ -70,8 +70,8 @@ static INLINE void T1WriteByte(u8 * mem, u32 addr, u8 val)
 static INLINE void T1WriteWord(u8 * mem, u32 addr, u16 val)
 {
 #ifdef WORDS_BIGENDIAN
-   mem[addr] = val >> 8;
-   mem[addr + 1] = val & 0xFF;
+   mem[addr + 1] = val >> 8;
+   mem[addr] = val & 0xFF;
 #else
    *((u16 *) (mem + addr)) = val;
 #endif
@@ -80,10 +80,10 @@ static INLINE void T1WriteWord(u8 * mem, u32 addr, u16 val)
 static INLINE void T1WriteLong(u8 * mem, u32 addr, u32 val)
 {
 #ifdef WORDS_BIGENDIAN
-   mem[addr] = val >> 24;
-   mem[addr + 1] = (val >> 16) & 0xFF;
-   mem[addr + 2] = (val >> 8) & 0xFF;
-   mem[addr + 3] = val & 0xFF;
+   mem[addr + 3] = val >> 24;
+   mem[addr + 2] = (val >> 16) & 0xFF;
+   mem[addr + 1] = (val >> 8) & 0xFF;
+   mem[addr] = val & 0xFF;
 #else
    *((u32 *) (mem + addr)) = val;
 #endif