18 files changed, 293 insertions, 139 deletions
diff --git a/backends/audiocd/audiocd.h b/backends/audiocd/audiocd.h
index 70eabf650b..5ea5bfcfd4 100644
--- a/backends/audiocd/audiocd.h
+++ b/backends/audiocd/audiocd.h
@@ -27,6 +27,7 @@
 #define BACKENDS_AUDIOCD_ABSTRACT_H
 
 #include "common/noncopyable.h"
+#include "common/scummsys.h"
 
 /**
 * Abstract Audio CD manager class. Subclasses implement the actual
@@ -45,6 +46,8 @@ public:
 		int start;
 		int duration;
 		int numLoops;
+		int volume;
+		int balance;
 	};
 
 	/**
@@ -72,6 +75,16 @@ public:
 	virtual bool isPlaying() const = 0;
 
 	/**
+	 * Set the audio volume
+	 */
+	virtual void setVolume(byte volume) = 0;
+
+	/**
+	 * Set the speakers balance
+	 */
+	virtual void setBalance(int8 balance) = 0;
+
+	/**
 	 * Stop CD or emulated audio playback.
 	 */
 	virtual void stop() = 0;
diff --git a/backends/audiocd/default/default-audiocd.cpp b/backends/audiocd/default/default-audiocd.cpp
index 0e7dc66ba1..56e737d359 100644
--- a/backends/audiocd/default/default-audiocd.cpp
+++ b/backends/audiocd/default/default-audiocd.cpp
@@ -33,6 +33,8 @@ DefaultAudioCDManager::DefaultAudioCDManager() {
 	_cd.start = 0;
 	_cd.duration = 0;
 	_cd.numLoops = 0;
+	_cd.volume = Audio::Mixer::kMaxChannelVolume;
+	_cd.balance = 0;
 	_mixer = g_system->getMixer();
 	_emulating = false;
 	assert(_mixer);
@@ -70,7 +72,7 @@ void DefaultAudioCDManager::play(int track, int numLoops, int startFrame, int du
 			*/
 			_emulating = true;
 			_mixer->playStream(Audio::Mixer::kMusicSoundType, &_handle,
-			                        Audio::makeLoopingAudioStream(stream, start, end, (numLoops < 1) ? numLoops + 1 : numLoops));
+			                        Audio::makeLoopingAudioStream(stream, start, end, (numLoops < 1) ? numLoops + 1 : numLoops), -1, _cd.volume, _cd.balance);
 		} else {
 			_emulating = false;
 			if (!only_emulate)
@@ -100,6 +102,38 @@ bool DefaultAudioCDManager::isPlaying() const {
 	}
 }
 
+void DefaultAudioCDManager::setVolume(byte volume) {
+	_cd.volume = volume;
+	if (_emulating) {
+		// Audio CD emulation
+		if (_mixer->isSoundHandleActive(_handle))
+			_mixer->setChannelVolume(_handle, _cd.volume);
+	} else {
+		// Real Audio CD
+		
+		// Unfortunately I can't implement this atm
+		// since SDL doesn't seem to offer an interface method for this.
+
+		// g_system->setVolumeCD(_cd.volume);
+	}
+}
+
+void DefaultAudioCDManager::setBalance(int8 balance) {
+	_cd.balance = balance;
+	if (_emulating) {
+		// Audio CD emulation
+		if (isPlaying())
+			_mixer->setChannelBalance(_handle, _cd.balance);
+	} else {
+		// Real Audio CD
+		
+		// Unfortunately I can't implement this atm
+		// since SDL doesn't seem to offer an interface method for this.
+
+		// g_system->setBalanceCD(_cd.balance);
+	}
+}
+
 void DefaultAudioCDManager::update() {
 	if (_emulating) {
 		// Check whether the audio track stopped playback
diff --git a/backends/audiocd/default/default-audiocd.h b/backends/audiocd/default/default-audiocd.h
index ec209d9eb5..3f9df0bbbb 100644
--- a/backends/audiocd/default/default-audiocd.h
+++ b/backends/audiocd/default/default-audiocd.h
@@ -40,6 +40,8 @@ public:
 	void play(int track, int numLoops, int startFrame, int duration, bool only_emulate = false);
 	void stop();
 	bool isPlaying() const;
+	void setVolume(byte volume);
+	void setBalance(int8 balance);
 	void update();
 	virtual Status getStatus() const; // Subclasses should override for better status results
 
diff --git a/backends/platform/ds/arm9/makefile b/backends/platform/ds/arm9/makefile
index fac50178b4..eedf75c256 100644
--- a/backends/platform/ds/arm9/makefile
+++ b/backends/platform/ds/arm9/makefile
@@ -205,7 +205,6 @@ CXX     = arm-eabi-g++
 LD	= arm-eabi-g++
 
 CFLAGS	=	-Wno-multichar -Wall\
-		-Wno-multichar -mcpu=arm9tdmi -mtune=arm9tdmi \
 		-mcpu=arm9tdmi -mtune=arm9tdmi -fomit-frame-pointer\
 		-mthumb-interwork -DUSE_ARM_COSTUME_ASM=1 -DDISABLE_SID
 
@@ -281,7 +280,6 @@ PORT_OBJS := \
 	$(portdir)/source/blitters_arm.o \
 	$(portdir)/source/cdaudio.o \
 	$(portdir)/source/dsmain.o \
-	$(portdir)/../../../fs/ds/ds-fs.o \
 	$(portdir)/source/gbampsave.o \
 	$(portdir)/source/scummhelp.o \
 	$(portdir)/source/osystem_ds.o \
@@ -297,14 +295,14 @@ ifdef USE_PROFILER
 endif
 
 
-DATA_OBJS :=
+DATA_OBJS := \
 	$(portdir)/data/icons.o \
 	$(portdir)/data/keyboard.o \
 	$(portdir)/data/keyboard_pal.o \
 	$(portdir)/data/default_font.o \
 	$(portdir)/data/8x8font_tga.o
 
-FAT_OBJS :=
+FAT_OBJS := \
 	$(portdir)/source/fat/disc_io.o \
 	$(portdir)/source/fat/gba_nds_fat.o \
 	$(portdir)/source/fat/io_fcsr.o \
diff --git a/backends/platform/ds/arm9/source/portdefs.h b/backends/platform/ds/arm9/source/portdefs.h
index bdb42993d1..cc38d66a73 100644
--- a/backends/platform/ds/arm9/source/portdefs.h
+++ b/backends/platform/ds/arm9/source/portdefs.h
@@ -26,6 +26,9 @@
 #ifndef _PORTDEFS_H_
 #define _PORTDEFS_H_
 
+// Include ndstypes.h for uint16 etc. typedefs
+#include "nds/ndstypes.h"
+
 // Somebody removed these from scummsys.h, but they're still required, so I'm
 // adding them here in the hope that they'll stay.
 #include <stdio.h>
@@ -49,6 +52,7 @@
 #define STREAM_AUDIO_FROM_DISK
 #endif
 
+// FIXME: What is "NO_DEBUG_MSGS" good for?
 #define NO_DEBUG_MSGS
 
 // This is defined in dsmain.cpp
@@ -65,16 +69,22 @@ void consolePrintf(const char *format, ...);
 #undef assert
 #endif
 
+#ifdef NDEBUG
+
+#define	assert(e)	((void)0)
+
+#else
+
+// FIXME: Shouldn't assert() also bail out / exit / halt the program? Right now we just
+// print an error message...
 #define assert(s) \
 	do { \
-		if (!(s)) \
+		if (!(s)) { \
 			consolePrintf("Assertion failed: '##s##' at file %s, line %d\n", __FILE__, __LINE__); \
+		} \
 	} while (0)
 
-//#include "ds-fs.h"
-
-//#define debug(fmt, ...) consolePrintf(fmt, ##__VA_ARGS__)
-//#define debug(fmt, ...) debug(0, fmt, ##__VA_ARGS__)
+#endif
 
 // FIXME: Since I can't change the engine at the moment (post lockdown) this define can go here.
 // This define changes the mouse-relative motion which doesn't make sense on a touch screen to
diff --git a/backends/platform/ds/ds.mk b/backends/platform/ds/ds.mk
index 97f104c391..a9ed3e0096 100644
--- a/backends/platform/ds/ds.mk
+++ b/backends/platform/ds/ds.mk
@@ -75,7 +75,7 @@ endif
 
 
 # Compiler options for files which should be optimised for speed
-OPT_SPEED := -O3
+OPT_SPEED := -O3 -mno-thumb
 
 # Compiler options for files which should be optimised for space
 OPT_SIZE := -Os -mthumb
@@ -83,31 +83,46 @@ OPT_SIZE := -Os -mthumb
 # By default optimize for size
 CXXFLAGS += $(OPT_SIZE)
 
-# Files listed below will be optimisied for speed, otherwise they will be optimised for space
-# TODO: speed original list contained three more files that should be optimized
-# for speed: actor.cpp gfx.cpp sprite.cpp -- but there are many files with these
-# names, which are the "right" ones?
+# Files listed below will be optimisied for speed, otherwise they will be optimised for space.
+# TODO: Several of these files probably should not be optimized for speed, but for now
+# we replicate the *precise* list from the old DS makefile, to ensure full compatibility.
+# Eventually, we should tune this list.
+$(ndsdir)/arm9/source/blitters.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 $(ndsdir)/arm9/source/dsmain.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 $(ndsdir)/arm9/source/osystem_ds.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-$(ndsdir)/arm9/source/blitters.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-$(ndsdir)/arm9/source/ds_main.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/saga/isomap.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+base/main.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 sound/rate.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/saga/actor_walk.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+sound/softsynth/opl/mame.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/agi/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/agos/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/agos/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/cine/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/cruise/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/cruise/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/draci/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/draci/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/gob/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/groovie/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/kyra/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/m4/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/m4/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/m4/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/made/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 engines/saga/actor_path.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/saga/image.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-sound/fmopl.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/saga/actor_walk.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 engines/saga/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/scumm/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/m4/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/scumm/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/cine/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/agos/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 engines/saga/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/saga/image.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/saga/isomap.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/saga/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 engines/saga/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/m4/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-engines/agi/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
-# TODO: Fingolfin says: optimizing staticres for size would
+engines/sci/engine/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/scumm/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/scumm/gfx.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/scumm/script.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/sword2/sprite.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+engines/teenagent/actor.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
+# TODO: Fingolfin says: optimizing kyra/staticres.o for size would
 # save about 30k, so maybe consider that?
 #engines/kyra/staticres.o: CXXFLAGS:=$(CXXFLAGS) $(OPT_SPEED)
 
diff --git a/backends/platform/ds/setup-builddirs.sh b/backends/platform/ds/setup-builddirs.sh
new file mode 100755
index 0000000000..11aacc1233
--- /dev/null
+++ b/backends/platform/ds/setup-builddirs.sh
@@ -0,0 +1,57 @@
+#!/bin/sh
+
+# TODO: Allow specifying (a list of) build ids (a, b, ...) on the command line.
+# TODO: Allow overriding the LOGO variable
+
+# TODO: Computer srcdir in a better way
+srcdir=../../../`dirname $0`/..
+
+BASE_PARAM="--host=ds --disable-translation --disable-debug --disable-all-engines"
+
+for build in a b c d e f g h i k; do
+	echo "--- Setting up build $build ---"
+	mkdir -p build-$build && cd build-$build
+	case $build in
+	a)
+		EXTRA_PARAM="--enable-scumm"
+		;;
+	b)
+		EXTRA_PARAM="--enable-sky --enable-queen"
+		;;
+	c)
+		EXTRA_PARAM="--enable-agos"
+		;;
+	d)
+		EXTRA_PARAM="--enable-gob --enable-cine --enable-agi"
+		;;
+	e)
+		EXTRA_PARAM="--enable-saga --disable-mad"
+		;;
+	f)
+		EXTRA_PARAM="--enable-kyra --disable-mad"
+		;;
+	g)
+		EXTRA_PARAM="--enable-lure"
+		;;
+	h)
+		EXTRA_PARAM="--enable-parallaction"
+		;;
+	i)
+		EXTRA_PARAM="--enable-made --disable-mad"
+		;;
+	k)
+		EXTRA_PARAM="--enable-cruise --disable-mad"
+		;;
+	*)
+		echo "Invalid build $build selected"
+		exit 1
+		;;
+	esac
+
+	defname="DS_BUILD_`echo $build | tr '[a-z]' '[A-Z]'`"
+	CPPFLAGS="$CPPFLAGS -D$defname"
+	$srcdir/configure $BASE_PARAM $EXTRA_PARAM
+	cd ..
+	echo DONE
+	echo
+done
diff --git a/backends/platform/iphone/iphone_video.m b/backends/platform/iphone/iphone_video.m
index faa0719b6c..cd8b38acb3 100644
--- a/backends/platform/iphone/iphone_video.m
+++ b/backends/platform/iphone/iphone_video.m
@@ -276,7 +276,7 @@ uint getSizeNextPOT(uint size) {
 
 	glGenTextures(1, &_screenTexture);
 	glBindTexture(GL_TEXTURE_2D, _screenTexture);
-	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 
 	if (_textureBuffer) {
 		free(_textureBuffer);
diff --git a/backends/platform/n64/README.N64 b/backends/platform/n64/README.N64
index 1cdd82e1e9..276436b676 100644
--- a/backends/platform/n64/README.N64
+++ b/backends/platform/n64/README.N64
@@ -13,6 +13,8 @@ Requirements
 
 Build cart images from sources
 ==============================
+You can download hkz-libn64 sources from here: http://hkzlab.ipv7.net/consoles.html
+hkz-libn64 is a library to control Nintendo64 hardware (es, video, audio, input, etc.).
 * TODO *
 
 Build cart images from binaries
diff --git a/backends/platform/n64/module.mk b/backends/platform/n64/module.mk
index 8fb6ba49ab..429b63802e 100644
--- a/backends/platform/n64/module.mk
+++ b/backends/platform/n64/module.mk
@@ -1,7 +1,12 @@
 MODULE := backends/platform/n64
 
 MODULE_OBJS := \
-	nintendo64.o
+	nintendo64.o \
+	osys_n64_base.o \
+	osys_n64_events.o \
+	osys_n64_utilities.o \
+	pakfs_save_manager.o \
+	framfs_save_manager.o 
 
 # We don't use rules.mk but rather manually update OBJS and MODULE_DIRS.
 MODULE_OBJS := $(addprefix $(MODULE)/, $(MODULE_OBJS))
diff --git a/backends/platform/psp/audio.cpp b/backends/platform/psp/audio.cpp
index bf1fb9ab41..e540733162 100644
--- a/backends/platform/psp/audio.cpp
+++ b/backends/platform/psp/audio.cpp
@@ -28,7 +28,6 @@
  
 #include "common/scummsys.h" 
 #include "backends/platform/psp/audio.h"
-#include "backends/platform/psp/thread.h"
 
 //#define __PSP_DEBUG_FUNCS__	/* For debugging function calls */
 //#define __PSP_DEBUG_PRINT__	/* For debug printouts */
@@ -85,43 +84,13 @@ bool PspAudio::open(uint32 freq, uint32 numOfChannels, uint32 numOfSamples, call
 	_init = true;
 	_paused = true;	// start in paused mode
 	
-	createThread();
+	threadCreateAndStart("audioThread", PRIORITY_AUDIO_THREAD, STACK_AUDIO_THREAD);	// start the consumer thread
 	
 	return true;
 }
 
-bool PspAudio::createThread() {
-	DEBUG_ENTER_FUNC();
-	int threadId = sceKernelCreateThread("audioThread", thread, PRIORITY_AUDIO_THREAD, STACK_AUDIO_THREAD, THREAD_ATTR_USER, 0); 
-	
-	if (threadId < 0) {	// error
-		PSP_ERROR("failed to create audio thread. Error code %d\n", threadId);
-		return false;
-	}
-	
-	PspAudio *_this = this;	// trick to get into context when the thread starts
-	
-	if (sceKernelStartThread(threadId, sizeof(uint32 *), &_this) < 0) {
-		PSP_ERROR("failed to start thread %d\n", threadId);
-		return false;
-	}
-	
-	PSP_DEBUG_PRINT("created audio thread[%x]\n", threadId);
-	
-	return true;
-}
-
-// Static function to be called upon thread startup. Will call a non-static function	
-int PspAudio::thread(SceSize, void *__this) {
-	DEBUG_ENTER_FUNC();
-	PspAudio *_this = *(PspAudio **)__this;		// get our this for the context
-	
-	_this->audioThread();
-	return 0;
-};
-
 // The real thread function
-void PspAudio::audioThread() {	
+void PspAudio::threadFunction() {	
 	assert(_callback);
 	PSP_DEBUG_PRINT_FUNC("audio thread started\n");
 
@@ -129,15 +98,15 @@ void PspAudio::audioThread() {
 		if (_paused)
 			PSP_DEBUG_PRINT("audio thread paused\n");
 		while (_paused) {	// delay until we stop pausing
-			sceKernelDelayThread(100000);	// 100ms
+			PspThread::delayMicros(100000);	// 100ms
 			if (!_paused)
 				PSP_DEBUG_PRINT("audio thread unpaused\n");
 		}
 
-		PSP_DEBUG_PRINT("remaining samples[%d]\n", remainingSamples);
+		PSP_DEBUG_PRINT("remaining samples[%d]\n", _remainingSamples);
 
 		PSP_DEBUG_PRINT("filling buffer[%d]\n", _bufferToFill);
-		_callback(_userData, _buffers[_bufferToFill], _bufferSize); // ask mixer to fill in
+		_callback(_userData, _buffers[_bufferToFill], _bufferSize); // ask mixer to fill in data
 		nextBuffer(_bufferToFill);
 		
 		PSP_DEBUG_PRINT("playing buffer[%d].\n", _bufferToPlay);
@@ -151,7 +120,7 @@ void PspAudio::audioThread() {
 	PSP_DEBUG_PRINT("audio thread exiting. ****************************\n");
 }
 
-// Much faster than using %
+// Much faster than using %, especially with conditional moves (MIPS)
 inline void PspAudio::nextBuffer(int &bufferIdx) {
 	DEBUG_ENTER_FUNC();
 	bufferIdx++;
@@ -176,6 +145,6 @@ inline bool PspAudio::playBuffer() {
 }
 
 void PspAudio::close() {
-	PSP_DEBUG_PRINT("close had been called ***************\n");
+	PSP_DEBUG_PRINT("close has been called ***************\n");
 	_init = false; 
 }
diff --git a/backends/platform/psp/audio.h b/backends/platform/psp/audio.h
index 603f8f6bfc..eeba598fed 100644
--- a/backends/platform/psp/audio.h
+++ b/backends/platform/psp/audio.h
@@ -26,13 +26,15 @@
 #ifndef PSP_AUDIO_H
 #define PSP_AUDIO_H
 
-class PspAudio {
+#include "backends/platform/psp/thread.h"
+
+class PspAudio : public PspThreadable {
 public:
 	enum {
 		NUM_BUFFERS = 2,
 		FREQUENCY = 44100	/* only frequency we allow */
 	};
-	typedef void (* callbackFunc)(void *userData, byte *samples, int len);
+	typedef void (* callbackFunc)(void *userData, byte *samples, int len);	// audio callback to call
 	PspAudio() : _pspChannel(0), 
 			_numOfChannels(0), _numOfSamples(0), _callback(0), 
 			_bufferToPlay(0), _bufferToFill(0), 
@@ -43,14 +45,12 @@ public:
 	~PspAudio() { close(); }
 	bool playBuffer();
 	void nextBuffer(int &bufferIdx);
-	static int thread(SceSize, void *);
-	void audioThread();
 	bool open(uint32 freq, uint32 numOfChannels, uint32 numOfSamples, callbackFunc callback, void *userData);
-	bool createThread();
 	void close();
 	uint32 getFrequency() { return FREQUENCY; }
 	void pause() { _paused = true; }
 	void unpause() { _paused = false; }
+	virtual void threadFunction();	// actual audio thread
 	
 private:
 	int _pspChannel;				// chosen hardware output channel
diff --git a/backends/platform/psp/display_manager.cpp b/backends/platform/psp/display_manager.cpp
index a9f33f6091..5037543f12 100644
--- a/backends/platform/psp/display_manager.cpp
+++ b/backends/platform/psp/display_manager.cpp
@@ -34,7 +34,6 @@
 #include "backends/platform/psp/default_display_client.h"
 #include "backends/platform/psp/cursor.h"
 #include "backends/platform/psp/pspkeyboard.h"
-#include "backends/platform/psp/thread.h"
 
 #define USE_DISPLAY_CALLBACK	// to use callback for finishing the render
 #include "backends/platform/psp/display_manager.h"
@@ -65,37 +64,24 @@ const OSystem::GraphicsMode DisplayManager::_supportedModes[] = {
 
 void MasterGuRenderer::setupCallbackThread() {
 	DEBUG_ENTER_FUNC();
-	int thid = sceKernelCreateThread("displayCbThread", guCallbackThread, PRIORITY_DISPLAY_THREAD, STACK_DISPLAY_THREAD, THREAD_ATTR_USER, 0);
 	
-	PSP_DEBUG_PRINT("Display CB thread id is %x\n", thid);
-	
-	// We want to pass the pointer to this, but we'll have to take address of this so use a little trick
-	MasterGuRenderer *_this = this;
-	
-	if (thid >= 0) {
-		sceKernelStartThread(thid, sizeof(uint32 *), &_this);
-	} else 
-		PSP_ERROR("failed to create display callback thread\n");
+	// start the thread that updates the display
+	threadCreateAndStart("DisplayCbThread", PRIORITY_DISPLAY_THREAD, STACK_DISPLAY_THREAD);		
 }
 
-// thread that reacts to the callback
-int MasterGuRenderer::guCallbackThread(SceSize, void *__this) {
+// this function gets called by PspThread when starting the new thread
+void MasterGuRenderer::threadFunction() {
 	DEBUG_ENTER_FUNC();
 	
-	// Dereferenced the copied value which was this
-	MasterGuRenderer *_this = *(MasterGuRenderer **)__this;
-	
 	// Create the callback. It should always get the pointer to MasterGuRenderer
-	_this->_callbackId = sceKernelCreateCallback("Display Callback", guCallback, _this);
-	if (_this->_callbackId < 0) {
-		PSP_ERROR("failed to create display callback\n");
-		return -1;
+	_callbackId = sceKernelCreateCallback("Display Callback", guCallback, this);
+	if (_callbackId < 0) {
+		PSP_ERROR("failed to create display callback\n");		
 	}
 	
 	PSP_DEBUG_PRINT("created callback. Going to sleep\n");
 
-	sceKernelSleepThreadCB();	// sleep until we get a callback
-	return 0;
+	sceKernelSleepThreadCB();	// sleep until we get a callback	
 }
 
 // This callback is called when the render is finished. It swaps the buffers
diff --git a/backends/platform/psp/display_manager.h b/backends/platform/psp/display_manager.h
index dbbdf2022c..1f7320902c 100644
--- a/backends/platform/psp/display_manager.h
+++ b/backends/platform/psp/display_manager.h
@@ -26,10 +26,12 @@
 #ifndef PSP_DISPLAY_MAN_H
 #define PSP_DISPLAY_MAN_H
 
+#include "backends/platform/psp/thread.h"
+
 /**
  *	Class used only by DisplayManager to start/stop GU rendering
  */
-class MasterGuRenderer {
+class MasterGuRenderer : public PspThreadable {
 public:
 	MasterGuRenderer() : _lastRenderTime(0), _renderFinished(true), _callbackId(-1) {}
 	void guInit();
@@ -37,15 +39,15 @@ public:
 	void guPostRender();
 	void guShutDown();
 	bool isRenderFinished() { return _renderFinished; }
-	void setupCallbackThread();
+	void setupCallbackThread();	
 private:
+	virtual void threadFunction();			// for the display callback thread
 	static uint32 _displayList[];
 	uint32 _lastRenderTime;					// For measuring rendering time
 	void guProgramDisplayBufferSizes();
-	static int guCallbackThread(SceSize, void *);	// for the graphics callbacks
-	static int guCallback(int, int, void *__this);
+	static int guCallback(int, int, void *__this);	// for the display callback
 	bool _renderFinished;					// for sync with render callback
-	int _callbackId;						// to keep track of render callback
+	int _callbackId;						// to keep track of render callback	
 };
 
 class Screen;
diff --git a/backends/platform/psp/memory.cpp b/backends/platform/psp/memory.cpp
index 8eef223f8c..29d0482d9a 100644
--- a/backends/platform/psp/memory.cpp
+++ b/backends/platform/psp/memory.cpp
@@ -39,9 +39,17 @@
 
 extern "C" {
 
+#ifdef TEST_MEMORY_COPY		/* we won't be able to run in this case b/c of printouts */
+extern void *__real_memcpy(void *dst, void *src, size_t bytes);
+#endif
+
 void *__wrap_memcpy(void *dst, void *src, size_t bytes) {
+#ifdef TEST_MEMORY_COPY		/* we won't be able to run in this case */
+	return __real_memcpy(dst, src, bytes);
+#else
 	PspMemory::fastCopy((byte *)dst, (byte *)src, bytes);
 	return dst;
+#endif
 }
 
 }
@@ -291,43 +299,31 @@ void PspMemory::copy32Misaligned(uint32 *dst32, const byte *src, uint32 bytes, u
 	PSP_DEBUG_PRINT("copy32Misaligned: dst32[%p], src[%p], bytes[%d], alignSrc[%d]\n", dst32, src, bytes, alignSrc);
 	
 	uint32 *src32 = (uint32 *)(((uint32)src) & 0xFFFFFFFC);	// remove misalignment
-	uint32 offset;
+	uint32 shiftValue, lastShiftValue;
 	
 	switch (alignSrc) {
 	case 1:
-		offset = misaligned32Detail(dst32, src32, bytes, alignSrc, 8, 24);
+		shiftValue = 8;
+		lastShiftValue = 24;
 		break;
 	case 2:
-		offset = misaligned32Detail(dst32, src32, bytes, alignSrc, 16, 16);
+		shiftValue = 16;
+		lastShiftValue = 16;
 		break;
 	default: /* 3 */
-		offset = misaligned32Detail(dst32, src32, bytes, alignSrc, 24, 8);
+		shiftValue = 24;
+		lastShiftValue = 8;
 		break;
 	}
-	
-	uint32 remainingBytes = bytes & 3;
-	
-	if (remainingBytes) {
-		byte *dst = (byte *)dst32;
-		src += offset;
-		dst += offset;
-		copy8(dst, src, remainingBytes);
-	}	
-}
 
-// returns offset in dst
-uint32 PspMemory::misaligned32Detail(uint32 *dst32, uint32 *src32, uint32 bytes, uint32 alignSrc, const uint32 shiftValue, const uint32 lastShiftValue) {
-	uint32 *origDst32 = dst32;
-	register uint32 dstWord, srcWord;
-	
-	PSP_DEBUG_PRINT("misaligned32Detail(): alignSrc[%d], dst32[%p], src32[%p], bytes[%d]\n", alignSrc, dst32, src32, bytes);
-	
+	uint32 dstWord, srcWord;
+
 	// Try to do groups of 4 words
 	uint32 words4 = bytes >> 4;
 	
-	srcWord = src32[0];
-
-	while (words4--) {
+	srcWord = *src32;		// preload 1st word so we read ahead
+	
+	for (; words4; words4--) {
 		dstWord = srcWord >> shiftValue;
 		srcWord = src32[1];
 		dstWord |= srcWord << lastShiftValue;
@@ -348,22 +344,29 @@ uint32 PspMemory::misaligned32Detail(uint32 *dst32, uint32 *src32, uint32 bytes,
 		dst32 += 4;
 	}
 	
-	uint32 words = (bytes & 0xF) >> 2;
+	uint32 words = (bytes & 0xF) >> 2;	// now get remaining words
 	
 	// we read one word ahead of what we write
 	// setup the first read
-	if (words) {
-		src32++;	// we already loaded the value, so just increment
+	
+	for (; words ;words--) {
+		dstWord = srcWord >> shiftValue;
+		srcWord = src32[1];				// we still go one ahead
+		src32++;
+		dstWord |= srcWord << lastShiftValue;
+		*dst32++ = dstWord;
+	}
+	
+	uint32 bytesLeft = bytes & 3;	// and remaining bytes
 		
-		while (words--) {
-			dstWord = srcWord >> shiftValue;
-			srcWord = *src32++;
-			dstWord |= srcWord << lastShiftValue;
-			*dst32++ = dstWord;
+	if (bytesLeft) {
+		byte *dst8 = (byte *)dst32;
+		byte *src8 = ((byte *)src32) + ((uint32)src & 0x3);	// get exact location we should be at
+
+		for(; bytesLeft; bytesLeft--) {
+			*dst8++ = *src8++;
 		}
 	}
-	
-	return (byte *)dst32 - (byte *)origDst32;
 }
 
 // More challenging -- need to shift
diff --git a/backends/platform/psp/memory.h b/backends/platform/psp/memory.h
index c9b3d21cb6..793bc94888 100644
--- a/backends/platform/psp/memory.h
+++ b/backends/platform/psp/memory.h
@@ -52,7 +52,6 @@ private:
 	static void copy32Aligned(uint32 *dst32, const uint32 *src32, uint32 bytes);
 	static void swap32Aligned(uint32 *dst32, const uint32 *src32, uint32 bytes, PSPPixelFormat &format);
 	static void copy32Misaligned(uint32 *dst32, const byte *src, uint32 bytes, uint32 alignSrc);
-	static uint32 misaligned32Detail(uint32 *dst32, uint32 *src32, uint32 bytes, uint32 alignSrc, const uint32 shiftValue, const uint32 lastShiftValue);
 	static void swap32Misaligned(uint32 *dst32, const uint16 *src16, uint32 bytes, PSPPixelFormat &format);
 	static void copy16(uint16 *dst, const uint16 *src, uint32 bytes);
 
diff --git a/backends/platform/psp/thread.cpp b/backends/platform/psp/thread.cpp
index c19ff5f9e3..916b1e553b 100644
--- a/backends/platform/psp/thread.cpp
+++ b/backends/platform/psp/thread.cpp
@@ -28,7 +28,50 @@
 #include "backends/platform/psp/thread.h"
 #include "backends/platform/psp/trace.h"
  
-// Class PspThread -------------------------------------------------- 
+// Class PspThreadable -------------------------------------------------- 
+// Inherit this to create C++ threads easily
+
+bool PspThreadable::threadCreateAndStart(const char *threadName, int priority, int stackSize, bool useVfpu /*= false*/) {
+	DEBUG_ENTER_FUNC();
+
+	if (_threadId != -1) {
+		PSP_ERROR("thread already created!\n");
+		return false;
+	}
+		
+	_threadId = sceKernelCreateThread(threadName, __threadCallback, priority, stackSize, THREAD_ATTR_USER, 0);	// add VFPU support
+
+	if (_threadId < 0) {
+		PSP_ERROR("failed to create %s thread. Error code %d\n", threadName, _threadId);
+		return false;
+	}
+	
+	// We want to pass the pointer to this, but we'll have to take address of this so use a little trick
+	PspThreadable *_this = this;
+	
+	if (sceKernelStartThread(_threadId, sizeof(uint32 *), &_this) < 0) {
+		PSP_ERROR("failed to start %s thread id[%d]\n", threadName, _threadId);
+		return false;
+	}
+	
+	PSP_DEBUG_PRINT("Started %s thread with id[%x]\n", threadName, _threadId);	
+	
+	return true;
+}
+
+// Callback function to be called by PSP kernel
+int PspThreadable::__threadCallback(SceSize, void *__this) {
+	DEBUG_ENTER_FUNC();
+	
+	PspThreadable *_this = *(PspThreadable **)__this;	// Dereference the copied value which was 'this'
+	
+	_this->threadFunction();	// call the virtual function
+	
+	return 0;
+}
+
+// PspThread class
+// Utilities to access general thread functions
  
 void PspThread::delayMillis(uint32 ms) {
 	sceKernelDelayThread(ms * 1000);
diff --git a/backends/platform/psp/thread.h b/backends/platform/psp/thread.h
index 27d53903d6..de1c10a2aa 100644
--- a/backends/platform/psp/thread.h
+++ b/backends/platform/psp/thread.h
@@ -26,11 +26,26 @@
 #ifndef PSP_THREAD_H
 #define PSP_THREAD_H
 
+#include <pspthreadman.h>
 #include "common/scummsys.h"
 
+// class to inherit for creating threads
+class PspThreadable {
+protected:
+	int _threadId;
+	virtual void threadFunction() = 0;	// this function will be called when the thread starts
+public:	
+	PspThreadable() : _threadId(-1) {}					// constructor
+	virtual ~PspThreadable() {}							// destructor
+	static int __threadCallback(SceSize, void *__this);	// used to get called by sceKernelStartThread() Don't override
+	bool threadCreateAndStart(const char *threadName, int priority, int stackSize, bool useVfpu = false);
+};
+
+// class for thread utils
 class PspThread {
-public:
-	static void delayMillis(uint32 ms);
+public:	
+	// static functions
+	static void delayMillis(uint32 ms);	// delay the current thread
 	static void delayMicros(uint32 us);
 };
 
@@ -85,6 +100,7 @@ enum ThreadPriority {
 };
 
 enum StackSizes {
+	STACK_DEFAULT = 4 * 1024,
 	STACK_AUDIO_THREAD = 16 * 1024,
 	STACK_TIMER_THREAD = 32 * 1024,
 	STACK_DISPLAY_THREAD = 2 * 1024,