diff --git a/NINTV-DS.nds b/NINTV-DS.nds
index f25c0aa..998c8bc 100644
Binary files a/NINTV-DS.nds and b/NINTV-DS.nds differ
diff --git a/arm9/Makefile b/arm9/Makefile
index a760fee..e4baa4a 100644
--- a/arm9/Makefile
+++ b/arm9/Makefile
@@ -26,7 +26,7 @@ GRAPHICS    :=  gfx
 #ARCH	:=	-mthumb -mthumb-interwork
 ARCH	:=	
 
-CFLAGS	:=	-Ofast -march=armv5te -mtune=arm946e-s -fomit-frame-pointer -ffast-math $(ARCH) -falign-functions=16 -frename-registers
+CFLAGS	:=	-Ofast -march=armv5te -mtune=arm946e-s -fomit-frame-pointer -ffast-math $(ARCH) -falign-functions=4 -frename-registers
 
 CFLAGS	+=	$(INCLUDE) -DARM9
 CXXFLAGS	:=	$(CFLAGS) -fno-rtti -fno-exceptions -Wno-conversion-null
diff --git a/arm9/source/debugger.cpp b/arm9/source/debugger.cpp
index f3f3314..a5c431b 100644
--- a/arm9/source/debugger.cpp
+++ b/arm9/source/debugger.cpp
@@ -32,7 +32,7 @@
 #include "AudioMixer.h"
 #include "printf.h"
 
-INT32 debug[8] __attribute__((section(".dtcm")))  = {0};
+INT32 debug[DEBUG_SIZE] __attribute__((section(".dtcm")))  = {0};
 
 int getMemUsed(void)      // returns the amount of used memory in bytes
 {
diff --git a/arm9/source/debugger.h b/arm9/source/debugger.h
index f472324..3b2a95a 100644
--- a/arm9/source/debugger.h
+++ b/arm9/source/debugger.h
@@ -32,6 +32,7 @@ extern AY38900 *debug_stic;
 extern AY38914 *debug_psg;
 extern AY38914 *debug_psg2;
 
+#define DEBUG_SIZE  8
 extern INT32 debug[];
 
 extern void show_debug_overlay(void);
diff --git a/arm9/source/emucore/SP0256.cpp b/arm9/source/emucore/SP0256.cpp
index 55726b3..2d662ab 100644
--- a/arm9/source/emucore/SP0256.cpp
+++ b/arm9/source/emucore/SP0256.cpp
@@ -27,7 +27,7 @@ INT32 bitsLeft                  __attribute__((section(".dtcm")));
 INT32 currentBits               __attribute__((section(".dtcm")));
 UINT16 pc                       __attribute__((section(".dtcm")));
 UINT16 stack                    __attribute__((section(".dtcm")));
-INT32 mode                      __attribute__((section(".dtcm")));
+UINT8  mode                     __attribute__((section(".dtcm")));
 INT32 repeatPrefix              __attribute__((section(".dtcm")));
 
 UINT16 fifoBytes[64] __attribute__((section(".dtcm")));
@@ -166,8 +166,7 @@ ITCM_CODE INT32 SP0256::tick(INT32 minimum)
             if (periodCounter == 0) {
                 periodCounter = 64;
                 repeat--;
-                for (UINT8 j = 0; j < 6; j++)
-                    y[j][0] = y[j][1] = 0;
+                memset(y, 0x00, sizeof(y));
             }
             else
                 periodCounter--;
@@ -183,9 +182,7 @@ ITCM_CODE INT32 SP0256::tick(INT32 minimum)
                 periodCounter = period;
                 repeat--;
                 sample = ((amplitude & 0x1F) << ((amplitude & 0xE0) >> 5));
-                for (INT32 j = 0; j < 6; j++)
-                    y[j][0] = y[j][1] = 0;
-
+                memset(y, 0x00, sizeof(y));
             }
             else
                 periodCounter--;
@@ -213,7 +210,7 @@ ITCM_CODE INT32 SP0256::tick(INT32 minimum)
     return totalTicks;
 }
 
-INT8 SP0256::readDelta(INT32 numBits) {
+ITCM_CODE INT8 SP0256::readDelta(INT32 numBits) {
     INT32 value = readBits(numBits);
     if ((value & (1 << (numBits - 1))) != 0)
         value |= -1 << numBits;
@@ -270,7 +267,6 @@ ITCM_CODE INT32 SP0256::readBits(INT32 numBits)
             bitsLeft += 10;
             pc++;
         }
-
     }
 
     INT32 output = currentBits & bitMasks[numBits-1];
@@ -800,7 +796,7 @@ void SP0256::PAUSE(INT32 immed4) {
     periodInterpolation = 0;
 }
 
-void SP0256::decode() {
+ITCM_CODE void SP0256::decode() {
     INT32 immed4 = readBits(4);
     INT32 nextInstruction = readBitsReverse(4);
     switch (nextInstruction) {
diff --git a/arm9/source/emucore/SP0256.h b/arm9/source/emucore/SP0256.h
index dabb184..1896c93 100644
--- a/arm9/source/emucore/SP0256.h
+++ b/arm9/source/emucore/SP0256.h
@@ -67,7 +67,7 @@ extern  INT32 currentBits;
 //registers
 extern  UINT16 pc;
 extern  UINT16 stack;
-extern  INT32 mode;
+extern  UINT8  mode;
 extern  INT32 repeatPrefix;
 extern  INT32 command;
 
diff --git a/arm9/source/loadgame.cpp b/arm9/source/loadgame.cpp
index a8663c3..fe34262 100644
--- a/arm9/source/loadgame.cpp
+++ b/arm9/source/loadgame.cpp
@@ -61,7 +61,8 @@ BOOL LoadCart(const CHAR* filename)
     bIsFatalError = false;
     bGameLoaded = FALSE;
     
-    memset(debug, 0x00, 8 * (sizeof(UINT32)));
+    // Clear out the debug array with every new game loaded
+    memset(debug, 0x00, DEBUG_SIZE * (sizeof(UINT32)));
     
     // Load up the configuration based on the CRC32 of the game. Do this early since we need some of those properties to load the RIP
     FindAndLoadConfig(CRC32::getCrc(filename));
@@ -84,6 +85,11 @@ BOOL LoadCart(const CHAR* filename)
 
     const CHAR* extStart = filename + strlen(filename) - 4;
     
+    // -------------------------------------------------------------------------------------------------------------------------
+    // A .bin is always assumed to be a flat binary file (and may or may not have a .cfg file to go with it).
+    // A .rom is always assumed to be a file with extra meta-data that lets us know where to load it in memory.
+    // A .int file can switch-hit and might be a .bin or a .rom -- we look for the signature 0xA8 byte to see if it's a .rom
+    // -------------------------------------------------------------------------------------------------------------------------
     UINT8 bIsROM = ((strcmpi(extStart, ".rom") == 0) ? true:false);
     if (strcmpi(extStart, ".int") == 0)
     {
@@ -93,7 +99,7 @@ BOOL LoadCart(const CHAR* filename)
         if (file == NULL) 
         {
             FatalError("BIN FILE DOES NOT EXIST");
-            return NULL;
+            return FALSE;
         }
         if (fgetc(file) == 0xA8) bIsROM = true;
         fclose(file);
@@ -101,7 +107,7 @@ BOOL LoadCart(const CHAR* filename)
     
     if (bIsROM)
     {
-        //load the binary file as a Rip
+        //load the .rom file as a Rip with meta-data parsed
         currentRip = Rip::LoadRom(filename);
         if (currentRip == NULL)
         {