Misc. libpsn00b bugfixes, new critical section macros

15 files changed, 124 insertions, 134 deletions

diff --git a/libpsn00b/include/dlfcn.h b/libpsn00b/include/dlfcn.h
index 3c5260d..f6a5baf 100644
--- a/libpsn00b/include/dlfcn.h
+++ b/libpsn00b/include/dlfcn.h
@@ -24,7 +24,7 @@
  * GCC will generate code to set $t9 appropriately.
  */
 #define DL_PRE_CALL(func) \
-	__asm__ volatile("move $t9, %0;" :: "r"(func) : "$t9");
+	__asm__ volatile("move $t9, %0;" :: "r"(func) : "$t9")
 
 /* Structure and enum definitions */
 
diff --git a/libpsn00b/include/hwregs_a.inc b/libpsn00b/include/hwregs_a.inc
index ca38542..b0c6954 100644
--- a/libpsn00b/include/hwregs_a.inc
+++ b/libpsn00b/include/hwregs_a.inc
@@ -3,7 +3,7 @@
 
 ## Constants
 
-.set IOBASE,			0x1f80
+.set IOBASE,			0xbf80
 .set F_CPU,				33868800
 .set F_GPU,				53222400
 
diff --git a/libpsn00b/include/psxapi.h b/libpsn00b/include/psxapi.h
index 5d1097e..7353ed2 100644
--- a/libpsn00b/include/psxapi.h
+++ b/libpsn00b/include/psxapi.h
@@ -8,6 +8,7 @@
 
 #include <stdint.h>
 #include <stddef.h>
+#include <hwregs_c.h>
 
 /* Definitions */
 
@@ -143,6 +144,43 @@ typedef struct {
 	int			_reserved;
 } INT_RP;
 
+/* Fast interrupt disabling macros */
+
+// Clearing the IRQ_MASK register is faster than manipulating cop0r12, even
+// though it requires declaring a "hidden" local variable to save its state to;
+// it's also resilient to race conditions as there's no read-modify-write
+// operation during which an interrupt can occur. Note that interrupt flags in
+// the IRQ_STAT register will get set even if the respective enable bits in
+// IRQ_MASK are cleared, so doing this will properly defer IRQs rather than
+// dropping them.
+#define FastEnterCriticalSection() \
+	uint16_t __saved_irq_mask = IRQ_MASK; (IRQ_MASK = 0)
+#define FastExitCriticalSection() \
+	(IRQ_MASK = __saved_irq_mask)
+
+/*#define FastEnterCriticalSection() { \
+	uint32_t r0, r1; \
+	__asm__ volatile( \
+		"mfc0 %0, $12;" \
+		"li   %1, -1026;" \
+		"and  %1, %0;" \
+		"mtc0 %1, $12;" \
+		"nop;" \
+		: "=r"(r0), "=r"(r1) :: \
+	); \
+}
+#define FastExitCriticalSection() { \
+	uint32_t r0; \
+	__asm__ volatile( \
+		"mfc0 %0, $12;" \
+		"nop;" \
+		"ori  %0, 0x0401;" \
+		"mtc0 %0, $12;" \
+		"nop;" \
+		: "=r"(r0) :: \
+	); \
+}*/
+
 /* API */
 
 #ifdef __cplusplus
@@ -179,9 +217,9 @@ int DelDev(const char *name);
 void ListDev(void);
 void AddDummyTty(void);
 
-void EnterCriticalSection(void);
+int EnterCriticalSection(void);
 void ExitCriticalSection(void);
-void SwEnterCriticalSection(void);
+int SwEnterCriticalSection(void);
 void SwExitCriticalSection(void);
 
 void _InitCd(void);
diff --git a/libpsn00b/include/psxgpu.h b/libpsn00b/include/psxgpu.h
index 68e3bff..b0c5302 100644
--- a/libpsn00b/include/psxgpu.h
+++ b/libpsn00b/include/psxgpu.h
@@ -525,10 +525,10 @@ void *DrawSyncCallback(void (*func)(void));
 
 int LoadImage(const RECT *rect, const uint32_t *data);
 int StoreImage(const RECT *rect, uint32_t *data);
-int MoveImage(const RECT *rect, int x, int y);
+//int MoveImage(const RECT *rect, int x, int y);
 void LoadImage2(const RECT *rect, const uint32_t *data);
 void StoreImage2(const RECT *rect, uint32_t *data);
-void MoveImage2(const RECT *rect, int x, int y);
+//void MoveImage2(const RECT *rect, int x, int y);
 
 void ClearOTagR(uint32_t *ot, size_t length);
 void ClearOTag(uint32_t *ot, size_t length);
diff --git a/libpsn00b/psxapi/_syscalls.s b/libpsn00b/psxapi/_syscalls.s
index 24864f3..6eaed72 100644
--- a/libpsn00b/psxapi/_syscalls.s
+++ b/libpsn00b/psxapi/_syscalls.s
@@ -27,20 +27,21 @@ ExitCriticalSection:
 .global SwEnterCriticalSection
 .type SwEnterCriticalSection, @function
 SwEnterCriticalSection:
-	mfc0 $a0, $12 # cop0r12 &= ~0x00000401
-	li   $a1, -1026
-	and  $a0, $a1
-	mtc0 $a0, $12
-	nop
+	mfc0  $a0, $12 # cop0r12 &= ~0x401
+	li    $a1, -1026
+	and   $a1, $a0
+	mtc0  $a1, $12
+	andi  $a0, 0x0401 # return ((cop0r12_prev & 0x401) == 0x401)
+	sltiu $v0, $a0, 0x0401
 
-	jr   $ra
-	nop
+	jr    $ra
+	xori  $v0, 1
 
 .section .text.SwExitCriticalSection
 .global SwExitCriticalSection
 .type SwExitCriticalSection, @function
 SwExitCriticalSection:
-	mfc0 $a0, $12 # cop0r12 |= 0x00000401
+	mfc0 $a0, $12 # cop0r12 |= 0x401
 	nop
 	ori  $a0, 0x0401
 	mtc0 $a0, $12
diff --git a/libpsn00b/psxetc/dl.c b/libpsn00b/psxetc/dl.c
index ccf7a7c..1d73eb6 100644
--- a/libpsn00b/psxetc/dl.c
+++ b/libpsn00b/psxetc/dl.c
@@ -450,9 +450,9 @@ DLL *DL_CreateDLL(DLL *dll, void *ptr, size_t size, DL_ResolveMode mode) {
 		}
 	}
 
-	EnterCriticalSection();
+	FastEnterCriticalSection();
 	FlushCache();
-	ExitCriticalSection();
+	FastExitCriticalSection();
 
 	// Call the DLL's global constructors. This is the same thing we'd do in
 	// _start() for regular executables, but we have to do it outside of the
diff --git a/libpsn00b/psxetc/interrupts.c b/libpsn00b/psxetc/interrupts.c
index 0d926c4..2dacaed 100644
--- a/libpsn00b/psxetc/interrupts.c
+++ b/libpsn00b/psxetc/interrupts.c
@@ -137,13 +137,13 @@ void *DMACallback(DMA_Channel dma, void (*func)(void)) {
 		DMA_DICR |= (0x10000 << dma) | (1 << 23);
 
 		if (!(_num_dma_handlers++))
-			InterruptCallback(3, &_global_dma_handler);
+			InterruptCallback(IRQ_DMA, &_global_dma_handler);
 	} else if (old_callback && !func) {
 		if (--_num_dma_handlers) {
 			DMA_DICR &= ~(0x10000 << dma);
 		} else {
 			DMA_DICR = 0;
-			InterruptCallback(3, 0);
+			InterruptCallback(IRQ_DMA, 0);
 		}
 	}
 
diff --git a/libpsn00b/psxetc/logging.c b/libpsn00b/psxetc/logging.c
deleted file mode 100644
index 5199190..0000000
--- a/libpsn00b/psxetc/logging.c
+++ /dev/null
@@ -1,50 +0,0 @@
-/*
- * PSn00bSDK internal debug logger
- * (C) 2022 spicyjpeg - MPL licensed
- *
- * This file provides the (admittedly minimal) logging system used by all
- * PSn00bSDK libraries. Log messages and warnings are issued using the
- * _sdk_log() macro and collected into a buffer, whose contents can be flushed
- * by calling _sdk_dump_log() (by default this is done by VSync()). Logging is
- * only enabled in debug builds of libpsn00b.
- */
-
-#include <stddef.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <psxapi.h>
-#include <psxetc.h>
-
-#define LOG_BUFFER_SIZE		256
-
-#ifndef NDEBUG
-
-/* Internal globals */
-
-static char   _log_buffer[LOG_BUFFER_SIZE];
-static size_t _log_buffer_length = 0;
-
-/* Internal logging API */
-
-void _sdk_log_inner(const char *fmt, ...) {
-	va_list ap;
-
-	va_start(ap, fmt);
-	_log_buffer_length += vsnprintf(
-		&_log_buffer[_log_buffer_length],
-		LOG_BUFFER_SIZE - _log_buffer_length,
-		fmt,
-		ap
-	);
-	va_end(ap);
-}
-
-void _sdk_dump_log_inner(void) {
-	if (!_log_buffer_length)
-		return;
-
-	write(1, _log_buffer, _log_buffer_length);
-	_log_buffer_length = 0;
-}
-
-#endif
diff --git a/libpsn00b/psxgpu/common.c b/libpsn00b/psxgpu/common.c
index 9f45f10..93fdb52 100644
--- a/libpsn00b/psxgpu/common.c
+++ b/libpsn00b/psxgpu/common.c
@@ -38,9 +38,6 @@ static volatile uint16_t   _last_hblank;
 
 /* Private interrupt handlers */
 
-#define _ENTER_CRITICAL()	uint16_t mask = IRQ_MASK; IRQ_MASK = 0;
-#define _EXIT_CRITICAL()	IRQ_MASK = mask;
-
 static void _vblank_handler(void) {
 	_vblank_counter++;
 
@@ -74,8 +71,8 @@ void ResetGraph(int mode) {
 	// the first time.
 	if (!ResetCallback()) {
 		EnterCriticalSection();
-		InterruptCallback(0, &_vblank_handler);
-		DMACallback(2, &_gpu_dma_handler);
+		InterruptCallback(IRQ_VBLANK, &_vblank_handler);
+		DMACallback(DMA_GPU, &_gpu_dma_handler);
 
 		_gpu_video_mode = (GPU_GP1 >> 20) & 1;
 		ExitCriticalSection();
@@ -149,22 +146,22 @@ int VSync(int mode) {
 }
 
 void *VSyncHaltFunction(void (*func)(void)) {
-	//_ENTER_CRITICAL();
+	//FastEnterCriticalSection();
 
 	void *old_callback  = _vsync_halt_func;
 	_vsync_halt_func    = func;
 
-	//_EXIT_CRITICAL();
+	//FastExitCriticalSection();
 	return old_callback;
 }
 
 void *VSyncCallback(void (*func)(void)) {
-	_ENTER_CRITICAL();
+	FastEnterCriticalSection();
 
 	void *old_callback  = _vsync_callback;
 	_vsync_callback     = func;
 
-	_EXIT_CRITICAL();
+	FastExitCriticalSection();
 	return old_callback;
 }
 
@@ -184,18 +181,18 @@ int EnqueueDrawOp(
 	// race condition where the DMA transfer could end while interrupts are
 	// being disabled. Interrupts are disabled through the IRQ_MASK register
 	// rather than via syscalls for performance reasons.
-	_ENTER_CRITICAL();
+	FastEnterCriticalSection();
 	int length = _queue_length;
 
 	if (!length) {
 		_queue_length = 1;
-		_EXIT_CRITICAL();
+		FastExitCriticalSection();
 
 		func(arg1, arg2, arg3);
 		return 0;
 	}
 	if (length >= QUEUE_LENGTH) {
-		_EXIT_CRITICAL();
+		FastExitCriticalSection();
 
 		_sdk_log("draw queue overflow, dropping commands\n");
 		return -1;
@@ -211,7 +208,7 @@ int EnqueueDrawOp(
 	entry->arg2 = arg2;
 	entry->arg3 = arg3;
 
-	_EXIT_CRITICAL();
+	FastExitCriticalSection();
 	return length;
 }
 
@@ -242,12 +239,12 @@ int DrawSync(int mode) {
 }
 
 void *DrawSyncCallback(void (*func)(void)) {
-	_ENTER_CRITICAL();
+	FastEnterCriticalSection();
 
 	void *old_callback = _drawsync_callback;
 	_drawsync_callback = func;
 
-	_EXIT_CRITICAL();
+	FastExitCriticalSection();
 	return old_callback;
 }
 
diff --git a/libpsn00b/psxgpu/env.c b/libpsn00b/psxgpu/env.c
index f513727..07edacf 100644
--- a/libpsn00b/psxgpu/env.c
+++ b/libpsn00b/psxgpu/env.c
@@ -63,7 +63,7 @@ int DrawOTagEnv(const uint32_t *ot, DRAWENV *env) {
 	prim->code[2] |= ((env->clip.y + env->ofs[1]) & 0x7ff) << 11;
 
 	// Texture page (reset active page and set dither/mask bits)
-	prim->code[3]  = 0xe1000000;
+	prim->code[3]  = 0xe1000000 | env->tpage;
 	prim->code[3] |= (env->dtd & 1) << 9;
 	prim->code[3] |= (env->dfe & 1) << 10;
 
diff --git a/libpsn00b/psxgpu/image.c b/libpsn00b/psxgpu/image.c
index bbdb7c8..e73505f 100644
--- a/libpsn00b/psxgpu/image.c
+++ b/libpsn00b/psxgpu/image.c
@@ -1,6 +1,10 @@
 /*
  * PSn00bSDK GPU library (image and VRAM transfer functions)
  * (C) 2022 spicyjpeg - MPL licensed
+ *
+ * TODO: MoveImage() is currently commented out as it won't trigger a DMA IRQ,
+ * making it unusable as a draw queue command. A way around this (perhaps using
+ * the GPU IRQ?) shall be found.
  */
 
 #include <stdint.h>
@@ -58,9 +62,9 @@ int StoreImage(const RECT *rect, uint32_t *data) {
 	);
 }
 
-int MoveImage(const RECT *rect, int x, int y) {
+/*int MoveImage(const RECT *rect, int x, int y) {
 	return EnqueueDrawOp((void *) &MoveImage2, (uint32_t) rect, x, y);
-}
+}*/
 
 void LoadImage2(const RECT *rect, const uint32_t *data) {
 	_dma_transfer(rect, (uint32_t *) data, 1);
@@ -70,14 +74,14 @@ void StoreImage2(const RECT *rect, uint32_t *data) {
 	_dma_transfer(rect, data, 0);
 }
 
-void MoveImage2(const RECT *rect, int x, int y) {
+/*void MoveImage2(const RECT *rect, int x, int y) {
 	GPU_GP0 = 0x80000000;
 	//GPU_GP0 = rect->x | (rect->y << 16);
 	GPU_GP0 = *((const uint32_t *) &(rect->x));
 	GPU_GP0 = (x & 0xffff) | (y << 16);
 	//GPU_GP0 = rect->w | (rect->h << 16);
 	GPU_GP0 = *((const uint32_t *) &(rect->w));
-}
+}*/
 
 /* .TIM image parsers */
 
diff --git a/libpsn00b/psxgte/initgeom.s b/libpsn00b/psxgte/initgeom.s
index 6d1e36a..ccda7f1 100644
--- a/libpsn00b/psxgte/initgeom.s
+++ b/libpsn00b/psxgte/initgeom.s
@@ -1,43 +1,44 @@
 .set noreorder
 
+.include "hwregs_a.inc"
 .include "gtereg.inc"
 
 .section .text.InitGeom
 .global InitGeom
 .type InitGeom, @function
 InitGeom:
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
+	# Disable interrupts and make sure the GTE is enabled in COP0.
+	lui   $v0, IOBASE
+	lhu   $v1, IRQ_MASK($v0)
+	nop
+	sh    $0,  IRQ_MASK($v0)
 
-	jal		EnterCriticalSection
+	mfc0  $a0, $12
+	lui   $a1, 0x4000
+	or    $a1, $a0
+	mtc0  $a1, $12
 	nop
+	#nop
 
-	mfc0	$v0, $12				# Get SR
-	lui		$v1, 0x4000				# Set bit to enable cop2
-	or		$v0, $v1
-	mtc0	$v0, $12				# Set new SR
+	# Re-enable interrupts, then load default values into some GTE registers.
+	sh    $v1, IRQ_MASK($v0)
 
-	jal		ExitCriticalSection
+	ctc2  $0, C2_OFX
 	nop
+	ctc2  $0, C2_OFY
 
-	ctc2	$0 , $24				# Reset GTE offset
-	ctc2	$0 , $25
-
-	li		$v0, 320				# Set default projection plane
-	ctc2	$v0, $26
+	li    $a0, 320
+	ctc2  $a0, C2_H
 
-	li		$v0, 0x155				# Set ZSF3 and ZSF4 defaults
-	ctc2	$v0, $29
-	li		$v0, 0x100
-	ctc2	$v0, $30
+	li    $a0, 0x155
+	ctc2  $a0, C2_ZSF3
+	li    $a0, 0x100
+	ctc2  $a0, C2_ZSF4
 
-	li		$v0, 0xef9e				# DQA and DQB defaults
-	lui		$v1, 0x0140
-	ctc2	$v0, C2_DQA
-	ctc2	$v1, C2_DQB
+	li    $a0, 0xef9e
+	ctc2  $a0, C2_DQA
+	lui   $a0, 0x0140
+	ctc2  $a0, C2_DQB
 
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
+	jr    $ra
 	nop
-
diff --git a/libpsn00b/psxpress/mdec.c b/libpsn00b/psxpress/mdec.c
index d55dbbb..b430f2c 100644
--- a/libpsn00b/psxpress/mdec.c
+++ b/libpsn00b/psxpress/mdec.c
@@ -84,7 +84,7 @@ static const DECDCTENV _default_mdec_env = {
 /* Public API */
 
 void DecDCTReset(int mode) {
-	EnterCriticalSection();
+	FastEnterCriticalSection();
 
 	DMA_DPCR   |= 0x000000bb; // Enable DMA0 and DMA1
 	DMA_CHCR(0) = 0x00000201; // Stop DMA0
@@ -92,7 +92,7 @@ void DecDCTReset(int mode) {
 	MDEC1       = 0x80000000; // Reset MDEC
 	MDEC1       = 0x60000000; // Enable DMA in/out requests
 
-	ExitCriticalSection();
+	FastExitCriticalSection();
 	if (!mode)
 		DecDCTPutEnv(0, 0);
 }
diff --git a/libpsn00b/psxsio/sio.c b/libpsn00b/psxsio/sio.c
index 6b80352..cdcef0e 100644
--- a/libpsn00b/psxsio/sio.c
+++ b/libpsn00b/psxsio/sio.c
@@ -34,9 +34,6 @@ static volatile RingBuffer _tx_buffer, _rx_buffer;
 
 /* Private interrupt handler */
 
-#define _ENTER_CRITICAL()	uint16_t mask = IRQ_MASK; IRQ_MASK = 0;
-#define _EXIT_CRITICAL()	IRQ_MASK = mask;
-
 static void _sio_handler(void) {
 	// Handle any incoming bytes.
 	while (SIO_STAT & SR_RXRDY) {
@@ -92,7 +89,7 @@ static void _sio_handler(void) {
 
 void SIO_Init(int baud, uint16_t mode) {
 	EnterCriticalSection();
-	_old_sio_handler = InterruptCallback(8, &_sio_handler);
+	_old_sio_handler = InterruptCallback(IRQ_SIO1, &_sio_handler);
 
 	SIO_CTRL = CR_ERRRST;
 	SIO_MODE = (mode & 0xfffc) | MR_BR_16;
@@ -114,7 +111,7 @@ void SIO_Init(int baud, uint16_t mode) {
 
 void SIO_Quit(void) {
 	EnterCriticalSection();
-	InterruptCallback(8, _old_sio_handler);
+	InterruptCallback(IRQ_SIO1, _old_sio_handler);
 
 	SIO_CTRL = CR_ERRRST;
 
@@ -122,7 +119,7 @@ void SIO_Quit(void) {
 }
 
 void SIO_SetFlowControl(SIO_FlowControl mode) {
-	_ENTER_CRITICAL();
+	FastEnterCriticalSection();
 
 	switch (mode) {
 		case SIO_FC_NONE:
@@ -147,7 +144,7 @@ void SIO_SetFlowControl(SIO_FlowControl mode) {
 			break;*/
 	}
 
-	_EXIT_CRITICAL();
+	FastExitCriticalSection();
 }
 
 /* Reading API */
@@ -167,13 +164,13 @@ int SIO_ReadByte2(void) {
 	if (!_rx_buffer.length)
 		return -1;
 
-	_ENTER_CRITICAL();
+	FastEnterCriticalSection();
 
 	int head        = _rx_buffer.head;
 	_rx_buffer.head = (head + 1) % BUFFER_LENGTH;
 	_rx_buffer.length--;
 
-	_EXIT_CRITICAL();
+	FastExitCriticalSection();
 	return _rx_buffer.data[head];
 }
 
@@ -192,12 +189,12 @@ int SIO_ReadSync(int mode) {
 }
 
 void *SIO_ReadCallback(int (*func)(uint8_t)) {
-	EnterCriticalSection();
+	FastEnterCriticalSection();
 
 	void *old_callback  = _read_callback;
 	_read_callback      = func;
 
-	ExitCriticalSection();
+	FastExitCriticalSection();
 }
 
 /* Writing API */
@@ -219,18 +216,18 @@ int SIO_WriteByte2(uint8_t value) {
 	// condition where the transfer could end while interrupts are being
 	// disabled. Interrupts are disabled through the IRQ_MASK register rather
 	// than via syscalls for performance reasons.
-	_ENTER_CRITICAL();
+	FastEnterCriticalSection();
 
 	if (SIO_STAT & (SR_TXRDY | SR_TXU)) {
 		SIO_TXRX = value;
-		_EXIT_CRITICAL();
+		FastExitCriticalSection();
 		return 0;
 	}
 
 	int length = _tx_buffer.length;
 
 	if (length >= BUFFER_LENGTH) {
-		_EXIT_CRITICAL();
+		FastExitCriticalSection();
 
 		//_sdk_log("TX overrun, dropping bytes\n");
 		return -1;
@@ -243,7 +240,7 @@ int SIO_WriteByte2(uint8_t value) {
 	_tx_buffer.data[tail] = value;
 	SIO_CTRL |= CR_TXIEN;
 
-	_EXIT_CRITICAL();
+	FastExitCriticalSection();
 	return length;
 }
 
diff --git a/libpsn00b/psxspu/common.c b/libpsn00b/psxspu/common.c
index d1dabfe..a340af6 100644
--- a/libpsn00b/psxspu/common.c
+++ b/libpsn00b/psxspu/common.c
@@ -14,6 +14,10 @@
 #define DMA_CHUNK_LENGTH	16
 #define STATUS_TIMEOUT		0x100000
 
+static const uint32_t _dummy_block[4] = {
+	0x00000500, 0x00000000, 0x00000000, 0x00000000
+};
+
 /* Internal globals */
 
 static SPU_TransferMode	_transfer_mode = SPU_TRANSFER_BY_DMA;
@@ -128,10 +132,8 @@ void SpuInit(void) {
 
 	// Upload a dummy looping ADPCM block to the first 16 bytes of SPU RAM.
 	// This may be freely used or overwritten.
-	uint32_t block[4] = { 0x0500, 0, 0, 0 };
-
 	_transfer_addr = WRITABLE_AREA_ADDR;
-	_manual_write((const uint16_t *) block, 16);
+	_manual_write((const uint16_t *) _dummy_block, 16);
 
 	// "Play" the dummy block on all channels. This will reset the start
 	// address and ADSR envelope status of each channel.
@@ -183,7 +185,7 @@ uint32_t SpuSetTransferStartAddr(uint32_t addr) {
 	if (addr > 0x7ffff)
 		return 0;
 
-	_transfer_addr = (addr + 7) / 8;
+	_transfer_addr = getSPUAddr(addr);
 	return addr;
 }