Fix delay function, logs in 480i mode, clangd config

.clangd

@@ -0,0 +1,9 @@
+# As clang/clangd's MIPS-I support is still experimental, some minor changes to
+# the GCC arguments it picks up from CMake are required in order to prevent it
+# from erroring out. Additionally, specifying the target architecture manually
+# fixes some edge cases (such as CMake emitting 8.3 format paths on Windows and
+# breaking clangd's target autodetection).
+
+CompileFlags:
+  Add:    [ --target=mipsel-none-elf, -march=mips1 ]
+  Remove: [ -march, -mno-llsc, -mdivide-breaks ]

.gitignore

@@ -5,6 +5,7 @@ desktop.ini
 
 # Do not include any built or cached files.
 build/
+.cache/
 __pycache__/
 *.pyc
 *.pyo

src/cartdata.cpp

@@ -336,6 +336,7 @@ Parser *newCartParser(Dump &dump, FormatType formatType, uint8_t flags) {
 
 Parser *newCartParser(Dump &dump) {
 	// Try all formats from the most complex one to the simplest.
+	//for (auto &format : _KNOWN_FORMATS) {
 	for (int i = util::countOf(_KNOWN_FORMATS) - 1; i >= 0; i--) {
 		auto   &format = _KNOWN_FORMATS[i];
 		Parser *parser = newCartParser(dump, format.format, format.flags);

src/libc/crt0.c

@@ -56,7 +56,7 @@ int _start(int argc, const char **argv) {
 	// Set $gp to point to the middle of the .sdata/.sbss sections, ensuring
 	// variables placed in those sections can be quickly accessed. See the
 	// linker script for more details.
-	__asm__ volatile("la $gp, _gp;");
+	__asm__ volatile("la $gp, _gp\n");
 
 	// Set all uninitialized variables to zero by clearing the BSS section.
 	__builtin_memset(_bssStart, 0, _bssEnd - _bssStart);

src/ps1/cop0gte.h

@@ -19,24 +19,24 @@
 #include <stdint.h>
 
 #define COP0_GET(reg, output) \
-	__asm__ volatile("mfc0 %0, $%1;" :  "=r"(output) : "i"(reg))
+	__asm__ volatile("mfc0 %0, $%1\n" :  "=r"(output) : "i"(reg))
 #define COP0_SET(reg, input) \
-	__asm__ volatile("mtc0 %0, $%1;" :: "r"(input), "i"(reg))
+	__asm__ volatile("mtc0 %0, $%1\n" :: "r"(input), "i"(reg))
 
 #define GTE_GET(reg, output) \
-	__asm__ volatile("mfc2 %0, $%1;" :  "=r"(output) : "i"(reg))
+	__asm__ volatile("mfc2 %0, $%1\n" :  "=r"(output) : "i"(reg))
 #define GTE_SET(reg, input) \
-	__asm__ volatile("mtc2 %0, $%1;" :: "r"(input), "i"(reg))
+	__asm__ volatile("mtc2 %0, $%1\n" :: "r"(input), "i"(reg))
 
 #define GTE_GETC(reg, output) \
-	__asm__ volatile("cfc2 %0, $%1;" :  "=r"(output) : "i"(reg))
+	__asm__ volatile("cfc2 %0, $%1\n" :  "=r"(output) : "i"(reg))
 #define GTE_SETC(reg, input) \
-	__asm__ volatile("ctc2 %0, $%1;" :: "r"(input), "i"(reg))
+	__asm__ volatile("ctc2 %0, $%1\n" :: "r"(input), "i"(reg))
 
 #define GTE_LOAD(reg, offset, ptr) \
-	__asm__ volatile("lwc2 $%0, %1(%2);" :: "i"(reg), "i"(offset), "r"(ptr))
+	__asm__ volatile("lwc2 $%0, %1(%2)\n" :: "i"(reg), "i"(offset), "r"(ptr))
 #define GTE_STORE(reg, offset, ptr) \
-	__asm__ volatile("swc2 $%0, %1(%2);" :: "i"(reg), "i"(offset), "r"(ptr) : "memory")
+	__asm__ volatile("swc2 $%0, %1(%2)\n" :: "i"(reg), "i"(offset), "r"(ptr) : "memory")
 
 /* Coprocessor 0 */
 
@@ -179,7 +179,7 @@ typedef struct {
 } GTEMatrix;
 
 #define gte_command(cmd) \
-	__asm__ volatile("nop; nop; cop2 %0;" :: "i"(cmd))
+	__asm__ volatile("nop\n" "nop\n" "cop2 %0\n" :: "i"(cmd))
 
 /* GTE control registers */
 

src/ps1/registers.h

@@ -241,9 +241,9 @@ typedef enum {
 	TIMER_CTRL_OVERFLOWED      = 1 << 12
 } TimerControlFlag;
 
-#define TIMER_VALUE(N)  _MMIO32((IO_BASE | 0x100) + (16 * (N)))
+#define TIMER_VALUE(N)  _MMIO16((IO_BASE | 0x100) + (16 * (N)))
-#define TIMER_CTRL(N)   _MMIO32((IO_BASE | 0x104) + (16 * (N)))
+#define TIMER_CTRL(N)   _MMIO16((IO_BASE | 0x104) + (16 * (N)))
-#define TIMER_RELOAD(N) _MMIO32((IO_BASE | 0x108) + (16 * (N)))
+#define TIMER_RELOAD(N) _MMIO16((IO_BASE | 0x108) + (16 * (N)))
 
 /* CD-ROM drive */
 

src/ps1/system.c

@@ -89,22 +89,10 @@ void flushCache(void) {
 void softReset(void) {
 	disableInterrupts();
 	BIOS_ENTRY_POINT();
+	__builtin_unreachable();
 }
 
-/* IRQ acknowledgement and blocking delay */
+/* IRQ acknowledgement */
-
-void delayMicroseconds(int us) {
-	// 1 us = 33.8688 cycles = 17 loop iterations (2 cycles per iteration)
-	us *= (F_CPU + 1000000) / 2000000;
-
-	__asm__ volatile(
-		".set noreorder;"
-		"bgtz  %0, .;"
-		"addiu %0, -1;"
-		".set reorder;"
-		: "=r"(us) : "r"(us)
-	);
-}
 
 bool acknowledgeInterrupt(IRQChannel irq) {
 	if (IRQ_STAT & (1 << irq)) {
@@ -116,22 +104,22 @@ bool acknowledgeInterrupt(IRQChannel irq) {
 }
 
 bool waitForInterrupt(IRQChannel irq, int timeout) {
-	for (; timeout > 0; timeout--) {
+	for (; timeout > 0; timeout -= 10) {
 		if (acknowledgeInterrupt(irq))
 			return true;
 
-		delayMicroseconds(1);
+		delayMicroseconds(10);
 	}
 
 	return false;
 }
 
 bool waitForDMATransfer(DMAChannel dma, int timeout) {
-	for (; timeout > 0; timeout--) {
+	for (; timeout > 0; timeout -= 10) {
 		if (!(DMA_CHCR(dma) & DMA_CHCR_ENABLE))
 			return true;
 
-		delayMicroseconds(1);
+		delayMicroseconds(10);
 	}
 
 	return false;

src/ps1/system.h

@@ -136,13 +136,24 @@ void softReset(void);
 
 /**
  * @brief Blocks for (roughly) the specified number of microseconds. This
- * function does not rely on a hardware timer, so interrupts may throw off
+ * function will reset hardware timer 2 and use it for timing. Disabling
- * timings if not explicitly disabled prior to calling delayMicroseconds().
+ * interrupts prior to calling delayMicroseconds() is highly recommended to
+ * prevent jitter, but not strictly necessary unless the interrupt handler
+ * accesses timer 2.
  *
  * @param time
  */
 void delayMicroseconds(int time);
 
+/**
+ * @brief Blocks for (roughly) the specified number of microseconds. This
+ * function does not rely on a hardware timer, so interrupts may throw off
+ * timings if not explicitly disabled prior to calling delayMicrosecondsBusy().
+ *
+ * @param time
+ */
+void delayMicrosecondsBusy(int time);
+
 /**
  * @brief Checks if the specified interrupt was fired but not yet acknowledged;
  * if so, acknowledges it and returns true. This function can be used in a
@@ -162,9 +173,9 @@ bool acknowledgeInterrupt(IRQChannel irq);
 
 /**
  * @brief Waits for the specified interrupt to be fired for up to the specified
- * number of microseconds. This function will work with interrupts that are not
+ * number of microseconds (with 10 us granularity). This function will work with
- * explicitly enabled in the IRQ_MASK register, but will *not* work with
+ * interrupts that are not explicitly enabled in the IRQ_MASK register, but will
- * interrupts that have been enabled if any callback set using
+ * *not* work with interrupts that have been enabled if any callback set using
  * setInterruptHandler() acknowledges them.
  *
  * @param irq
@@ -175,7 +186,7 @@ bool waitForInterrupt(IRQChannel irq, int timeout);
 
 /**
  * @brief Waits for the specified DMA channel to finish any ongoing transfer for
- * up to the specified number of microseconds.
+ * up to the specified number of microseconds (with 10 us granularity).
  *
  * @param dma
  * @param timeout
@@ -207,7 +218,7 @@ static inline void switchThreadImmediate(Thread *thread) {
 	switchThread(thread);
 
 	// Execute a syscall to force the switch to happen.
-	__asm__ volatile("syscall 0;" ::: "memory");
+	__asm__ volatile("syscall 0\n" ::: "memory");
 }
 
 #ifdef __cplusplus

src/ps1/system.s

@@ -15,6 +15,8 @@
 .set noreorder
 .set noat
 
+## Exception handler
+
 .set BADV,  $8
 .set SR,    $12
 .set CAUSE, $13
@@ -179,3 +181,96 @@ _exceptionHandler:
 
 	jr    $k1
 	rfe
+
+## Delay functions
+
+.set IO_BASE, 0xbf801000
+
+.set TIMER2_VALUE,  IO_BASE | 0x120
+.set TIMER2_CTRL,   IO_BASE | 0x124
+.set TIMER2_RELOAD, IO_BASE | 0x128
+
+.section .text.delayMicroseconds, "ax", @progbits
+.global delayMicroseconds
+.type delayMicroseconds, @function
+
+delayMicroseconds:
+	# Calculate the approximate number of CPU cycles that need to be burned,
+	# assuming a 33.8688 MHz clock (1 us = 33.8688 = ~33.875 cycles).
+	sll   $a1, $a0, 8 # cycles = ((us * 271) + 4) / 8
+	sll   $a2, $a0, 4
+	addu  $a1, $a2
+	subu  $a1, $a0
+	addiu $a1, 4
+	sra   $a0, $a1, 3
+
+	# Compensate for the overhead of calculating the cycle count, entering the
+	# loop and returning.
+	addiu $a0, -(6 + 1 + 2 + 3 + 2)
+
+	# Reset timer 2 to its default setting of counting system clock edges.
+	lui   $v1, %hi(IO_BASE)
+	sh    $0,  %lo(TIMER2_CTRL)($v1) # TIMER2_CTRL = 0
+
+	# Wait for up to 0xff00 cycles at a time (resetting the timer and waiting
+	# for it to count up each time), as the counter is only 16 bits wide. We
+	# have to wait 0xff00 cycles rather than 0x10000 since the counter wraps
+	# around rather than saturating on overflow.
+	li    $a1, 0xff00
+	slt   $v0, $a1, $a0
+	#beqz  $v0, .LshortDelay
+	#nop
+	beqz  $v0, .LskipLongDelay
+
+.LlongDelay: # for (; cycles > 0xff00; cycles -= 0xff00)
+	sh    $0,  %lo(TIMER2_VALUE)($v1) # TIMER2_VALUE = 0
+	li    $v0, 0
+
+.LlongDelayLoop: # while (TIMER2_VALUE < 0xff00);
+	nop
+	slt   $v0, $v0, $a1
+	bnez  $v0, .LlongDelayLoop
+	lhu   $v0, %lo(TIMER2_VALUE)($v1)
+
+	slt   $v0, $a1, $a0
+	bnez  $v0, .LlongDelay
+	subu  $a0, $a1
+
+.LshortDelay:
+	# Run the last busy loop once less than 0xff00 cycles are remaining.
+	sh    $0,  %lo(TIMER2_VALUE)($v1) # TIMER2_VALUE = 0
+.LskipLongDelay:
+	li    $v0, 0
+
+.LshortDelayLoop: # while (TIMER2_VALUE < cycles);
+	nop
+	slt   $v0, $v0, $a0
+	bnez  $v0, .LshortDelayLoop
+	lhu   $v0, %lo(TIMER2_VALUE)($v1)
+
+	jr    $ra
+	nop
+
+.section .text.delayMicrosecondsBusy, "ax", @progbits
+.global delayMicrosecondsBusy
+.type delayMicrosecondsBusy, @function
+
+delayMicrosecondsBusy:
+	# Calculate the approximate number of CPU cycles that need to be burned,
+	# assuming a 33.8688 MHz clock (1 us = 33.8688 = ~33.875 cycles).
+	sll   $a1, $a0, 8 # cycles = ((us * 271) + 4) / 8
+	sll   $a2, $a0, 4
+	addu  $a1, $a2
+	subu  $a1, $a0
+	addiu $a1, 4
+	sra   $a0, $a1, 3
+
+	# Compensate for the overhead of calculating the cycle count and returning.
+	addiu $a0, -(6 + 1 + 2)
+
+.Lloop: # while (cycles > 0) cycles -= 2
+	bgtz  $a0, .Lloop
+	addiu $a0, -2
+
+	jr    $ra
+	nop

src/util.cpp

@@ -42,8 +42,8 @@ Logger::Logger(void)
 }
 
 void Logger::clear(void) {
-	for (int i = 0; i < MAX_LOG_LINES; i++)
+	for (auto line : _lines)
-		_lines[i][0] = 0;
+		line[0] = 0;
 }
 
 void Logger::log(const char *format, ...) {
@@ -114,8 +114,10 @@ uint32_t zipCRC32(const uint8_t *data, size_t length, uint32_t crc) {
 	crc        = ~crc;
 
 	for (; length; length--) {
+		uint32_t temp = crc;
+
 		crc >>= 8;
-		crc  ^= table[(crc ^ *(data++)) & 0xff];
+		crc  ^= table[(temp ^ *(data++)) & 0xff];
 	}
 
 	return ~crc;

src/util.hpp

@@ -224,7 +224,7 @@ public:
 /* Logger (basically a ring buffer of lines) */
 
 static constexpr int MAX_LOG_LINE_LENGTH = 128;
-static constexpr int MAX_LOG_LINES       = 32;
+static constexpr int MAX_LOG_LINES       = 64;
 
 class Logger {
 private: