Merge pull request #70 from Lameguy64/v0.23-wip

Header cleanups, PCDRV, more safety checks, libc and mkpsxiso fixes (v0.23)

35 files changed, 927 insertions, 585 deletions

diff --git a/libpsn00b/include/assert.h b/libpsn00b/include/assert.h
index 1b2bda2..8f8df74 100644
--- a/libpsn00b/include/assert.h
+++ b/libpsn00b/include/assert.h
@@ -1,13 +1,12 @@
 /*
  * PSn00bSDK assert macro and internal logging
- * (C) 2022 spicyjpeg - MPL licensed
+ * (C) 2022-2023 spicyjpeg - MPL licensed
  *
- * Note that the _sdk_log() macro is used internally by PSn00bSDK to output
- * debug messages and warnings.
+ * The _sdk_*() macros are used internally by PSn00bSDK to output messages when
+ * building in debug mode.
  */
 
-#ifndef __ASSERT_H
-#define __ASSERT_H
+#pragma once
 
 #include <stdio.h>
 
@@ -25,6 +24,9 @@ void _assert_abort(const char *file, int line, const char *expr);
 
 #define assert(expr)
 #define _sdk_log(fmt, ...)
+#define _sdk_assert(expr, fmt, ...)
+#define _sdk_validate_args_void(expr)
+#define _sdk_validate_args(expr, ret)
 
 #else
 
@@ -32,11 +34,27 @@ void _assert_abort(const char *file, int line, const char *expr);
 	((expr) ? ((void) 0) : _assert_abort(__FILE__, __LINE__, #expr))
 
 #ifdef SDK_LIBRARY_NAME
-#define _sdk_log(fmt, ...) printf(SDK_LIBRARY_NAME ": " fmt, ##__VA_ARGS__)
+#define _sdk_log(fmt, ...) \
+	printf(SDK_LIBRARY_NAME ": " fmt __VA_OPT__(,) __VA_ARGS__)
 #else
-#define _sdk_log(fmt, ...) printf(fmt, ##__VA_ARGS__)
+#define _sdk_log(fmt, ...) \
+	printf(fmt __VA_OPT__(,) __VA_ARGS__)
 #endif
 
-#endif
+#define _sdk_assert(expr, ret, fmt, ...) \
+	if (!(expr)) { \
+		_sdk_log(fmt, __VA_ARGS__); \
+		return ret; \
+	}
+#define _sdk_validate_args_void(expr) \
+	if (!(expr)) { \
+		_sdk_log("invalid args to %s() (%s)\n", __func__, #expr); \
+		return; \
+	}
+#define _sdk_validate_args(expr, ret) \
+	if (!(expr)) { \
+		_sdk_log("invalid args to %s() (%s)\n", __func__, #expr); \
+		return ret; \
+	}
 
 #endif
diff --git a/libpsn00b/include/cassert b/libpsn00b/include/cassert
new file mode 100644
index 0000000..0923486
--- /dev/null
+++ b/libpsn00b/include/cassert
@@ -0,0 +1,8 @@
+/*
+ * PSn00bSDK assert macro and internal logging
+ * (C) 2022-2023 spicyjpeg - MPL licensed
+ */
+
+#pragma once
+
+#include <assert.h>
diff --git a/libpsn00b/include/cctype b/libpsn00b/include/cctype
new file mode 100644
index 0000000..b73ad34
--- /dev/null
+++ b/libpsn00b/include/cctype
@@ -0,0 +1,22 @@
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#pragma once
+
+namespace std {
+extern "C" {
+
+int isprint(int ch);
+int isgraph(int ch);
+int isspace(int ch);
+int isblank(int ch);
+int isalpha(int ch);
+int isdigit(int ch);
+
+int tolower(int ch);
+int toupper(int ch);
+
+}
+}
diff --git a/libpsn00b/include/cstdint b/libpsn00b/include/cstdint
new file mode 100644
index 0000000..3b1bc4a
--- /dev/null
+++ b/libpsn00b/include/cstdint
@@ -0,0 +1,34 @@
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
+ *
+ * This is a replacement for the <cstdint> header included with GCC, which seems
+ * to be broken (at least in GCC 12.2.0) as it requires some macros to be set.
+ */
+
+#pragma once
+
+#include <stdint.h>
+
+namespace std {
+
+#define _DEF_TYPE(bits, prefix) \
+	using ::prefix##bits##_t; \
+	using ::prefix##_fast##bits##_t; \
+	using ::prefix##_least##bits##_t;
+
+_DEF_TYPE( 8, int)
+_DEF_TYPE( 8, uint)
+_DEF_TYPE(16, int)
+_DEF_TYPE(16, uint)
+_DEF_TYPE(32, int)
+_DEF_TYPE(32, uint)
+
+#undef _DEF_TYPE
+
+using ::intmax_t;
+using ::uintmax_t;
+using ::intptr_t;
+using ::uintptr_t;
+
+}
diff --git a/libpsn00b/include/cstdio b/libpsn00b/include/cstdio
new file mode 100644
index 0000000..800d1a2
--- /dev/null
+++ b/libpsn00b/include/cstdio
@@ -0,0 +1,32 @@
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#pragma once
+
+#include <cstdarg>
+
+namespace std {
+extern "C" {
+
+/* String I/O API (provided by BIOS) */
+
+int printf(const char *fmt, ...);
+char *gets(char *str);
+void puts(const char *str);
+int getchar(void);
+void putchar(int ch);
+
+/* String formatting API (built-in) */
+
+int vsnprintf(char *string, unsigned int size, const char *fmt, va_list ap);
+int vsprintf(char *string, const char *fmt, va_list ap);
+int sprintf(char *string, const char *fmt, ...);
+int snprintf(char *string, unsigned int size, const char *fmt, ...);
+
+int vsscanf(const char *str, const char *format, va_list ap);
+int sscanf(const char *str, const char *fmt, ...);
+
+}
+}
diff --git a/libpsn00b/include/cstdlib b/libpsn00b/include/cstdlib
new file mode 100644
index 0000000..4fa859d
--- /dev/null
+++ b/libpsn00b/include/cstdlib
@@ -0,0 +1,59 @@
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#pragma once
+
+#include <cstddef>
+
+namespace std {
+
+/* Definitions */
+
+static constexpr int RAND_MAX = 0x7fff;
+
+/* Structure definitions */
+
+struct HeapUsage {
+	size_t total;		// Total size of heap + stack
+	size_t heap;		// Amount of memory currently reserved for heap
+	size_t stack;		// Amount of memory currently reserved for stack
+	size_t alloc;		// Amount of memory currently allocated
+	size_t alloc_max;	// Maximum amount of memory ever allocated
+};
+
+/* API */
+
+extern "C" {
+
+extern int			__argc;
+extern const char	**__argv;
+
+void abort(void);
+
+int rand(void);
+void srand(int seed);
+
+int abs(int j);
+long labs(long i);
+
+long strtol(const char *str, char **str_end, int base);
+long long strtoll(const char *str, char **str_end, int base);
+//float strtof(const char *str, char **str_end);
+//double strtod(const char *str, char **str_end);
+//long double strtold(const char *str, char **str_end);
+
+void InitHeap(void *addr, size_t size);
+void *sbrk(ptrdiff_t incr);
+
+void TrackHeapUsage(ptrdiff_t alloc_incr);
+void GetHeapUsage(HeapUsage *usage);
+
+void *malloc(size_t size);
+void *calloc(size_t num, size_t size);
+void *realloc(void *ptr, size_t size);
+void free(void *ptr);
+
+}
+}
diff --git a/libpsn00b/include/cstring b/libpsn00b/include/cstring
new file mode 100644
index 0000000..1ce7246
--- /dev/null
+++ b/libpsn00b/include/cstring
@@ -0,0 +1,38 @@
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#pragma once
+
+#include <cstddef>
+
+namespace std {
+extern "C" {
+
+void *memset(void *dest, int ch, size_t count);
+void *memcpy(void *dest, const void *src, size_t count);
+void *memccpy(void *dest, const void *src, int ch, size_t count);
+void *memmove(void *dest, const void *src, size_t count);
+int memcmp(const void *lhs, const void *rhs, size_t count);
+void *memchr(const void *ptr, int ch, size_t count);
+
+char *strcpy(char *dest, const char *src);
+char *strncpy(char *dest, const char *src, size_t count);
+int strcmp(const char *lhs, const char *rhs);
+int strncmp(const char *lhs, const char *rhs, size_t count);
+char *strchr(const char *str, int ch);
+char *strrchr(const char *str, int ch);
+char *strpbrk(const char *str, const char *breakset);
+char *strstr(const char *str, const char *substr);
+
+size_t strlen(const char *str);
+char *strcat(char *dest, const char *src);
+char *strncat(char *dest, const char *src, size_t count);
+char *strdup(const char *str);
+char *strndup(const char *str, size_t count);
+
+char *strtok(char *str, const char *delim);
+
+}
+}
diff --git a/libpsn00b/include/ctype.h b/libpsn00b/include/ctype.h
index 24ee9d9..2fe0a42 100644
--- a/libpsn00b/include/ctype.h
+++ b/libpsn00b/include/ctype.h
@@ -1,20 +1,24 @@
 /*
  * PSn00bSDK standard library
- * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __CTYPE_H
-#define __CTYPE_H
+#pragma once
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-int tolower(int chr);
-int toupper(int chr);
+int isprint(int ch);
+int isgraph(int ch);
+int isspace(int ch);
+int isblank(int ch);
+int isalpha(int ch);
+int isdigit(int ch);
+
+int tolower(int ch);
+int toupper(int ch);
 
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/dlfcn.h b/libpsn00b/include/dlfcn.h
index 6192430..5e1e3b6 100644
--- a/libpsn00b/include/dlfcn.h
+++ b/libpsn00b/include/dlfcn.h
@@ -3,8 +3,7 @@
  * (C) 2021-2022 spicyjpeg - MPL licensed
  */
 
-#ifndef __DLFCN_H
-#define __DLFCN_H 
+#pragma once
 
 #include <stdint.h>
 #include <stddef.h>
@@ -215,5 +214,3 @@ void *DL_GetDLLSymbol(const DLL *dll, const char *name);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/elf.h b/libpsn00b/include/elf.h
index abfb3d5..b0ddf71 100644
--- a/libpsn00b/include/elf.h
+++ b/libpsn00b/include/elf.h
@@ -9,8 +9,7 @@
  * converted to enums.
  */
 
-#ifndef __ELF_H
-#define __ELF_H 
+#pragma once
 
 #include <stdint.h>
 
@@ -121,18 +120,3 @@ typedef enum {
 	STT_LOPROC	= 13,		/* Start of processor-specific */
 	STT_HIPROC	= 15		/* End of processor-specific */
 } Elf32_st_type;
-
-// If you need to add more constants, you may use the following Python snippet
-// to quickly convert #defines to enums:
-/*
-import re
-t = """<paste #defines here>"""
-t = re.sub(
-	r"(0x[0-9a-f]+|0b[01]+|[0-9]+)",
-	lambda m: f"= {m.group(1)},",
-	t.replace("#define ",  "\t").replace("#define\t", "\t")
-)
-print("typedef enum {\n" + t + "\n} NAME;")
-*/
-
-#endif
diff --git a/libpsn00b/include/hwregs_c.h b/libpsn00b/include/hwregs_c.h
index 7015101..2152986 100644
--- a/libpsn00b/include/hwregs_c.h
+++ b/libpsn00b/include/hwregs_c.h
@@ -3,8 +3,7 @@
  * (C) 2022 spicyjpeg - MPL licensed
  */
 
-#ifndef __HWREGS_C_H
-#define __HWREGS_C_H
+#pragma once
 
 #include <stdint.h>
 
@@ -35,7 +34,7 @@
 #define CD_DATA				_MMIO8(IOBASE | 0x1802)
 #define CD_IRQ				_MMIO8(IOBASE | 0x1803)
 
-#define CD_REG(N)			_MMIO8(IOBASE | 0x1800 + (N))
+#define CD_REG(N)			_MMIO8((IOBASE | 0x1800) + (N))
 
 /* SPU */
 
@@ -74,13 +73,13 @@
 
 // These are not named SPU_VOICE_* to avoid name clashes with SPU attribute
 // flags defined in psxspu.h.
-#define SPU_CH_VOL_L(N)		_MMIO16(IOBASE | 0x1c00 + 16 * (N))
-#define SPU_CH_VOL_R(N)		_MMIO16(IOBASE | 0x1c02 + 16 * (N))
-#define SPU_CH_FREQ(N)		_MMIO16(IOBASE | 0x1c04 + 16 * (N))
-#define SPU_CH_ADDR(N)		_MMIO16(IOBASE | 0x1c06 + 16 * (N))
-#define SPU_CH_ADSR1(N)		_MMIO16(IOBASE | 0x1c08 + 16 * (N))
-#define SPU_CH_ADSR2(N)		_MMIO16(IOBASE | 0x1c0a + 16 * (N))
-#define SPU_CH_LOOP_ADDR(N)	_MMIO16(IOBASE | 0x1c0e + 16 * (N))
+#define SPU_CH_VOL_L(N)		_MMIO16((IOBASE | 0x1c00) + (16 * (N)))
+#define SPU_CH_VOL_R(N)		_MMIO16((IOBASE | 0x1c02) + (16 * (N)))
+#define SPU_CH_FREQ(N)		_MMIO16((IOBASE | 0x1c04) + (16 * (N)))
+#define SPU_CH_ADDR(N)		_MMIO16((IOBASE | 0x1c06) + (16 * (N)))
+#define SPU_CH_ADSR1(N)		_MMIO16((IOBASE | 0x1c08) + (16 * (N)))
+#define SPU_CH_ADSR2(N)		_MMIO16((IOBASE | 0x1c0a) + (16 * (N)))
+#define SPU_CH_LOOP_ADDR(N)	_MMIO16((IOBASE | 0x1c0e) + (16 * (N)))
 
 /* MDEC */
 
@@ -92,11 +91,11 @@
 // IMPORTANT: even though SIO_DATA is a 32-bit register, it should only be
 // accessed as 8-bit. Reading it as 16 or 32-bit works fine on real hardware,
 // but leads to problems in some emulators.
-#define SIO_DATA(N)			_MMIO8 (IOBASE | 0x1040 + 16 * (N))
-#define SIO_STAT(N)			_MMIO16(IOBASE | 0x1044 + 16 * (N))
-#define SIO_MODE(N)			_MMIO16(IOBASE | 0x1048 + 16 * (N))
-#define SIO_CTRL(N)			_MMIO16(IOBASE | 0x104a + 16 * (N))
-#define SIO_BAUD(N)			_MMIO16(IOBASE | 0x104e + 16 * (N))
+#define SIO_DATA(N)			_MMIO8 ((IOBASE | 0x1040) + (16 * (N)))
+#define SIO_STAT(N)			_MMIO16((IOBASE | 0x1044) + (16 * (N)))
+#define SIO_MODE(N)			_MMIO16((IOBASE | 0x1048) + (16 * (N)))
+#define SIO_CTRL(N)			_MMIO16((IOBASE | 0x104a) + (16 * (N)))
+#define SIO_BAUD(N)			_MMIO16((IOBASE | 0x104e) + (16 * (N)))
 
 /* IRQ controller */
 
@@ -108,15 +107,15 @@
 #define DMA_DPCR			_MMIO32(IOBASE | 0x10f0)
 #define DMA_DICR			_MMIO32(IOBASE | 0x10f4)
 
-#define DMA_MADR(N)			_MMIO32(IOBASE | 0x1080 + 16 * (N))
-#define DMA_BCR(N)			_MMIO32(IOBASE | 0x1084 + 16 * (N))
-#define DMA_CHCR(N)			_MMIO32(IOBASE | 0x1088 + 16 * (N))
+#define DMA_MADR(N)			_MMIO32((IOBASE | 0x1080) + (16 * (N)))
+#define DMA_BCR(N)			_MMIO32((IOBASE | 0x1084) + (16 * (N)))
+#define DMA_CHCR(N)			_MMIO32((IOBASE | 0x1088) + (16 * (N)))
 
 /* Timers */
 
-#define TIMER_VALUE(N)		_MMIO32(IOBASE | 0x1100 + 16 * (N))
-#define TIMER_CTRL(N)		_MMIO32(IOBASE | 0x1104 + 16 * (N))
-#define TIMER_RELOAD(N)		_MMIO32(IOBASE | 0x1108 + 16 * (N))
+#define TIMER_VALUE(N)		_MMIO32((IOBASE | 0x1100) + (16 * (N)))
+#define TIMER_CTRL(N)		_MMIO32((IOBASE | 0x1104) + (16 * (N)))
+#define TIMER_RELOAD(N)		_MMIO32((IOBASE | 0x1108) + (16 * (N)))
 
 /* Memory/bus control */
 
@@ -130,5 +129,3 @@
 #define BUS_EXP2_CFG		_MMIO32(IOBASE | 0x101c)
 #define BUS_COM_DELAY		_MMIO32(IOBASE | 0x1020)
 #define BUS_RAM_SIZE		_MMIO32(IOBASE | 0x1060)
-
-#endif
diff --git a/libpsn00b/include/inline_c.h b/libpsn00b/include/inline_c.h
index 5facc1c..cb550b9 100644
--- a/libpsn00b/include/inline_c.h
+++ b/libpsn00b/include/inline_c.h
@@ -16,8 +16,7 @@
  * compiled object files.
  */
 
-#ifndef _INLINE_C_H
-#define _INLINE_C_H
+#pragma once
 
 /* GTE load macros */
 
@@ -1612,5 +1611,3 @@
 	: "g"( r0 ) )
 #define gte_mvmva_b(sf, mx, v, cv, lm) gte_mvmva_core_b( 0x0400012 | \
 	((sf)<<19) | ((mx)<<17) | ((v)<<15) | ((cv)<<13) | ((lm)<<10) )
-
-#endif // _INLINE_C_H
diff --git a/libpsn00b/include/ioctl.h b/libpsn00b/include/ioctl.h
deleted file mode 100644
index 5c56422..0000000
--- a/libpsn00b/include/ioctl.h
+++ /dev/null
@@ -1,19 +0,0 @@
-#ifndef _IOCTL_H
-#define _IOCTL_H
-
-#ifndef NULL
-#define NULL	0
-#endif
-
-#ifndef EOF
-#define EOF		-1
-#endif
-
-// General
-#define FIONBLOCK	(('f'<<8)|1)
-#define FIOCSCAN	(('f'<<8)|2)
-
-// disk
-#define DIO_FORMAT	(('d'<<8)|1)
-
-#endif
\ No newline at end of file
diff --git a/libpsn00b/include/lzconfig.h b/libpsn00b/include/lzconfig.h
deleted file mode 100644
index cb8a830..0000000
--- a/libpsn00b/include/lzconfig.h
+++ /dev/null
@@ -1,68 +0,0 @@
-/*!	\file lzconfig.h
- *	\brief Library configuration header
- *	\details Define settings will only take effect when you recompile the library.
- */
-
-#ifndef _LZP_CONFIG_H
-#define _LZP_CONFIG_H
-
-
-#ifndef TRUE
-#define TRUE	1
-#endif
-#ifndef FALSE
-#define FALSE	0
-#endif
-
-
-/* Set to TRUE to compile without data compression routines useful if you
- * plan to use this library on a program that does not require said routines
- * especially on a platform with limited memory (such as the PlayStation).
- *
- * This define will rule out lzCompress(), lzSetHashSizes() and
- * lzResetHashSizes() functions and their associated functions.
- */
-#define LZP_NO_COMPRESS		TRUE
-
-
-/* Set to TRUE to make default compression table sizes to maximum and works best
- * when compressing large amounts of data. LZP_USE_MALLOC must be set to TRUE to
- * prevent stack overflow errors.
- *
- * Do not enable this if you plan to compile for a platform with limited memory
- * otherwise, the library will consume all memory and crash the system.
- *
- * This define only affects lzCompress().
- */
-#define LZP_MAX_COMPRESS	FALSE
-
-
-/* Uncomment to make the library use malloc() instead of array initializers to
- * allocate hash tables. Enabling this is a must if you plan to use large hash
- * and window table sizes.
- */
-#define LZP_USE_MALLOC		FALSE
-
-
-/* Hash table sizes (in power-of-two multiple units)
- *
- * These define only affect lzCompress().
- */
-#if LZP_MAX_COMPRESS == TRUE
-
-// Minimal defaults
-#define LZP_WINDOW_SIZE	17
-#define LZP_HASH1_SIZE	8
-#define LZP_HASH2_SIZE	10
-
-#else
-
-// Maximum defaults
-#define LZP_WINDOW_SIZE	17
-#define LZP_HASH1_SIZE	22
-#define LZP_HASH2_SIZE	24
-
-#endif
-
-
-#endif // _LZP_CONFIG_H
diff --git a/libpsn00b/include/lzp/lzp.h b/libpsn00b/include/lzp/lzp.h
index 456de02..1aeea30 100644
--- a/libpsn00b/include/lzp/lzp.h
+++ b/libpsn00b/include/lzp/lzp.h
@@ -1,20 +1,29 @@
-/*!	\file lzp.h
- *	\brief Main library header
+/*
+ * liblzp data compression library
+ * (C) 2019 Lameguy64 - MPL licensed
  */
 
-/*!	\mainpage
- *	\version	0.20b
- *	\author		John Wilbert 'Lameguy64' Villamor
+/**
+ * @file lzp.h
+ * @brief Utility library for file bundling and compression
  *
- * 	\section	creditsSection Credits
- * 	- LZ77 data compression/decompression routines based from Ilya Muravyov's
- *	  crush.cpp released under public domain. Refined and ported to C by Lameguy64.
- *	- CRC calculation routines based from Lammert Bies' lib_crc routines.
+ * @details This library implements a simple in-memory archive format which
+ * can be used to package and compress assets for faster loading, as well as a
+ * generic LZ77 compressor and matching decompressor. Two archive formats are
+ * supported, one uncompressed (.QLP) and one with individually compressed
+ * entries (.LZP).
  *
+ * This header provides the LZ77 compression API and functions to parse and
+ * decompress .LZP archives after they have been loaded into memory.
+ *
+ * @section creditsSection Credits
+ * - LZ77 data compression/decompression routines based from Ilya Muravyov's
+ *   crush.cpp released under public domain. Refined and ported to C by
+ *   Lameguy64.
+ * - CRC calculation routines based from Lammert Bies' lib_crc routines.
  */
 
-#ifndef _LZPACK_H
-#define _LZPACK_H
+#pragma once
 
 #include <stdint.h>
 #ifdef _WIN32
@@ -218,6 +227,3 @@ int lzpUnpackFile(void* buff, const LZP_HEAD* lzpack, int fileNum);
 #ifdef __cplusplus
 }
 #endif
-
-
-#endif // _LZPACK_H
diff --git a/libpsn00b/include/lzp/lzqlp.h b/libpsn00b/include/lzp/lzqlp.h
index 32ce0d7..127f263 100644
--- a/libpsn00b/include/lzp/lzqlp.h
+++ b/libpsn00b/include/lzp/lzqlp.h
@@ -1,5 +1,23 @@
-#ifndef _QLP_H
-#define _QLP_H
+/*
+ * liblzp data compression library
+ * (C) 2019 Lameguy64 - MPL licensed
+ */
+
+/**
+ * @file lzqlp.h
+ * @brief Utility library for file bundling
+ *
+ * @details This library implements a simple in-memory archive format which
+ * can be used to package and compress assets for faster loading, as well as a
+ * generic LZ77 compressor and matching decompressor. Two archive formats are
+ * supported, one uncompressed (.QLP) and one with individually compressed
+ * entries (.LZP).
+ *
+ * This header provides functions to parse .QLP archives and retrieve pointers
+ * to their contents after they have been loaded into memory.
+ */
+
+#pragma once
 
 #include <stdint.h>
 #ifdef _WIN32
@@ -23,9 +41,17 @@ typedef struct {
 	uint32_t	offs;
 } QLP_FILE;
 
+
+// Function prototypes
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 int qlpFileCount(const QLP_HEAD* qlpfile);
 const QLP_FILE* qlpFileEntry(int index, const QLP_HEAD* qlpfile);
 const void* qlpFileAddr(int index, const QLP_HEAD* qlpfile);
 int qlpFindFile(char* fileName, const QLP_HEAD* qlpfile);
 
-#endif // _QLP_H
\ No newline at end of file
+#ifdef __cplusplus
+}
+#endif
diff --git a/libpsn00b/include/malloc.h b/libpsn00b/include/malloc.h
deleted file mode 100644
index 75c3711..0000000
--- a/libpsn00b/include/malloc.h
+++ /dev/null
@@ -1,8 +0,0 @@
-#ifndef _MALLOC_H
-#define _MALLOC_H
-
-#warning "<malloc.h> is deprecated, include <stdlib.h> instead"
-
-#include <stdlib.h>
-
-#endif // _MALLOC_H
\ No newline at end of file
diff --git a/libpsn00b/include/psxapi.h b/libpsn00b/include/psxapi.h
index 7353ed2..35ee040 100644
--- a/libpsn00b/include/psxapi.h
+++ b/libpsn00b/include/psxapi.h
@@ -1,10 +1,21 @@
 /*
  * PSn00bSDK kernel API library
- * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __PSXAPI_H
-#define __PSXAPI_H
+/**
+ * @file psxapi.h
+ * @brief Kernel API library header
+ *
+ * @details This header provides access to most of the APIs made available by
+ * the system's BIOS, including basic file I/O, TTY output, controller and
+ * memory card drivers, threads, events as well as kernel memory allocation.
+ *
+ * For more information and up-to-date documentation on kernel APIs, see:
+ * https://psx-spx.consoledev.net/kernelbios/
+ */
+
+#pragma once
 
 #include <stdint.h>
 #include <stddef.h>
@@ -12,13 +23,38 @@
 
 /* Definitions */
 
-#define DescHW			0xf0000000
-#define DescSW			0xf4000000
-
-#define HwCARD			(DescHW|0x11)
-#define HwCARD_1		(DescHW|0x12)
-#define HwCARD_0		(DescHW|0x13)
-#define SwCARD			(DescHW|0x02)
+// TODO: these desperately need to be cleaned up
+
+#define SEEK_SET	0
+#define SEEK_CUR	1
+#define SEEK_END	2
+
+#define DescMask	0xff000000		// Event descriptor mask
+#define DescTH		DescMask
+#define DescHW		0xf0000000		// Hardware event (IRQ)
+#define DescEV		0xf1000000		// Event event
+#define DescRC		0xf2000000		// Root counter event
+#define DescUEV		0xf3000000		// User event
+#define DescSW		0xf4000000		// BIOS event
+
+#define	HwVBLANK	(DescHW|0x01)	// VBlank
+#define HwGPU		(DescHW|0x02)	// GPU
+#define HwCdRom		(DescHW|0x03)	// CDROM
+#define HwDMAC		(DescHW|0x04)	// DMA
+#define HwRTC0		(DescHW|0x05)	// Timer 0
+#define HwRTC1		(DescHW|0x06)	// Timer 1
+#define HwRTC2		(DescHW|0x07)	// Timer 2
+#define HwCNTL		(DescHW|0x08)	// Controller
+#define HwSPU		(DescHW|0x09)	// SPU
+#define HwPIO		(DescHW|0x0a)	// PIO & lightgun
+#define HwSIO		(DescHW|0x0b)	// Serial
+
+#define HwCPU		(DescHW|0x10)	// Processor exception
+#define HwCARD		(DescHW|0x11)	// Memory card (lower level BIOS functions)
+#define HwCard_0	(DescHW|0x12)
+#define HwCard_1	(DescHW|0x13)
+#define SwCARD		(DescSW|0x01)	// Memory card (higher level BIOS functions)
+#define SwMATH		(DescSW|0x02)
 
 #define EvSpIOE			0x0004
 #define EvSpERROR		0x8000
@@ -135,8 +171,6 @@ struct JMP_BUF {
 	uint32_t gp;
 };
 
-// Not recommended to use these functions to install IRQ handlers
-
 typedef struct {
 	uint32_t	*next;
 	uint32_t	*func2;
@@ -158,7 +192,8 @@ typedef struct {
 #define FastExitCriticalSection() \
 	(IRQ_MASK = __saved_irq_mask)
 
-/*#define FastEnterCriticalSection() { \
+#if 0
+#define FastEnterCriticalSection() { \
 	uint32_t r0, r1; \
 	__asm__ volatile( \
 		"mfc0 %0, $12;" \
@@ -179,9 +214,10 @@ typedef struct {
 		"nop;" \
 		: "=r"(r0) :: \
 	); \
-}*/
+}
+#endif
 
-/* API */
+/* BIOS API */
 
 #ifdef __cplusplus
 extern "C" {
@@ -199,23 +235,28 @@ int DisableEvent(int event);
 void DeliverEvent(uint32_t cl, uint32_t spec);
 void UnDeliverEvent(uint32_t cl, uint32_t spec);
 
-int open(const char *name, int mode);
+int open(const char *path, int mode);
 int close(int fd);
-int seek(int fd, uint32_t offset, int mode);
-int read(int fd, uint8_t *buff, size_t len);
-int write(int fd, const uint8_t *buff, size_t len);
+int lseek(int fd, uint32_t offset, int mode);
+int read(int fd, void *buff, size_t len);
+int write(int fd, const void *buff, size_t len);
+int getc(int fd);
+int putc(int ch, int fd);
 int ioctl(int fd, int cmd, int arg);
+int isatty(int fd);
 struct DIRENTRY *firstfile(const char *wildcard, struct DIRENTRY *entry);
 struct DIRENTRY *nextfile(struct DIRENTRY *entry);
-int erase(const char *name);
-int chdir(const char *path);
+int erase(const char *path);
+int undelete(const char *path);
+int cd(const char *path);
 
-//#define cd(p) chdir(p)
+int _get_errno(void);
+int _get_error(int fd);
 
-int AddDev(DCB *dcb);
-int DelDev(const char *name);
-void ListDev(void);
-void AddDummyTty(void);
+int AddDrv(DCB *dcb);
+int DelDrv(const char *name);
+void ListDrv(void);
+void add_nullcon_driver(void);
 
 int EnterCriticalSection(void);
 void ExitCriticalSection(void);
@@ -254,30 +295,33 @@ int ResetRCnt(int spec);
 void ChangeClearPAD(int mode);
 void ChangeClearRCnt(int t, int m);
 
-uint32_t OpenTh(uint32_t (*func)(), uint32_t sp, uint32_t gp);
-int CloseTh(uint32_t thread);
-int ChangeTh(uint32_t thread);
+int OpenTh(uint32_t (*func)(), uint32_t sp, uint32_t gp);
+int CloseTh(int thread);
+int ChangeTh(int thread);
 
-int Exec(struct EXEC *exec, int argc, char **argv);
+int Exec(struct EXEC *exec, int argc, const char **argv);
+int LoadExec(const char *path, int argc, const char **argv);
 void FlushCache(void);
 
 void b_setjmp(struct JMP_BUF *buf);
 void b_longjmp(const struct JMP_BUF *buf, int param);
-void SetDefaultExitFromException(void);
-void SetCustomExitFromException(const struct JMP_BUF *buf);
+void ResetEntryInt(void);
+void HookEntryInt(const struct JMP_BUF *buf);
 void ReturnFromException(void);
 
+int SetConf(int evcb, int tcb, uint32_t sp);
+void GetConf(int *evcb, int *tcb, uint32_t *sp);
+void SetMem(int size);
+
 int GetSystemInfo(int index);
 void *GetB0Table(void);
 void *GetC0Table(void);
 
-void *_kernel_malloc(int size);
-void _kernel_free(void *ptr);
+void *alloc_kernel_memory(int size);
+void free_kernel_memory(void *ptr);
 
 void _boot(void);
 
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxcd.h b/libpsn00b/include/psxcd.h
index 503bc83..fc9c391 100644
--- a/libpsn00b/include/psxcd.h
+++ b/libpsn00b/include/psxcd.h
@@ -21,8 +21,7 @@
  * library extension is considered for future development.
  */
 
-#ifndef __PSXCD_H
-#define __PSXCD_H
+#pragma once
 
 #include <stdint.h>
 
@@ -811,6 +810,48 @@ int CdMode(void);
 int CdStatus(void);
 
 /**
+ * @brief Returns the CD-ROM controller's region code.
+ *
+ * @details Reads region information from the drive using a CdlTest command.
+ * This can be used to reliably determine the system's region without having to
+ * resort to workarounds like probing the BIOS ROM.
+ *
+ * This function may return incorrect results and trigger error callbacks on
+ * emulators or consoles equipped with CD-ROM drive emulation devices such as
+ * the PSIO. It is not affected by modchips.
+ *
+ * @return Region code or 0 if the region cannot be determined
+ */
+CdlRegionCode CdGetRegion(void);
+
+/**
+ * @brief Attempts to disable the CD-ROM controller's region check.
+ *
+ * @details Sends undocumented commands to the drive in an attempt to disable
+ * the region string check, in order to allow reading data from non-PS1 discs
+ * as well as CD-Rs without needing a modchip. As unlocking commands are region
+ * specific, the drive's region must be obtained beforehand using CdGetRegion()
+ * and passed to this function. The unlock persists even if the lid is opened,
+ * but not if a CdlReset command is issued.
+ *
+ * Unlocking is only supported on US, European and Net Yaroze consoles (not on
+ * Japanese models, devkits and most emulators). This function will return 1
+ * without doing anything if CdlRegionDebug is passed as region, as debug
+ * consoles can already read unlicensed discs.
+ *
+ * NOTE: if any callbacks were set using CdReadyCallback() or CdSyncCallback()
+ * prior to calling CdUnlock(), they will be called with an error code as part
+ * of the unlocking sequence, even if the unlock was successful. It is thus
+ * recommended to call this function before setting any callbacks.
+ *
+ * @param region
+ * @return 1 if the drive was successfully unlocked, 0 otherwise
+ *
+ * @see CdGetRegion()
+ */
+int CdUnlock(CdlRegionCode region);
+
+/**
  * @brief Retrieves the disc's table of contents.
  *
  * @details Retrieves the track entries from a CD's table of contents (TOC). The
@@ -832,21 +873,6 @@ int CdStatus(void);
 int CdGetToc(CdlLOC *toc);
 
 /**
- * @brief Returns the CD-ROM controller's region code.
- *
- * @details Attempts to fetch region information from the drive using a CdlTest
- * command. This can be used to reliably determine the system's region without
- * having to resort to workarounds like probing the BIOS ROM.
- *
- * This function may return incorrect results on emulators or consoles equipped
- * with CD-ROM drive emulation devices such as the PSIO. It is not affected by
- * modchips.
- *
- * @return Region code or 0 if the region cannot be determined
- */
-CdlRegionCode CdGetRegion(void);
-
-/**
  * @brief Sets the CD-ROM volume mixing matrix.
  *
  * @details Sets the volume levels of the CD-ROM drive's audio output (used for
@@ -1046,5 +1072,3 @@ int CdLoadSession(int session);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxetc.h b/libpsn00b/include/psxetc.h
index ae4611e..8dd1dd5 100644
--- a/libpsn00b/include/psxetc.h
+++ b/libpsn00b/include/psxetc.h
@@ -1,6 +1,6 @@
 /*
  * PSn00bSDK interrupt management library
- * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
  */
 
 /**
@@ -13,8 +13,7 @@
  * separate header).
  */
 
-#ifndef __PSXETC_H
-#define __PSXETC_H
+#pragma once
 
 /* IRQ and DMA channel definitions */
 
@@ -68,7 +67,7 @@ extern "C" {
  * | ID  | Channel          | Used by                                 |
  * | --: | :--------------- | :-------------------------------------- |
  * |   0 | IRQ_VBLANK       | psxgpu (use VSyncCallback() instead)    |
- * |   1 | IRQ_GPU          |                                         |
+ * |   1 | IRQ_GPU          | psxgpu (use DrawSyncCallback() instead) |
  * |   2 | IRQ_CD           | psxcd (use CdReadyCallback() instead)   |
  * |   3 | IRQ_DMA          | psxetc (use DMACallback() instead)      |
  * |   4 | IRQ_TIMER0       |                                         |
@@ -156,19 +155,50 @@ void *DMACallback(DMA_Channel dma, void (*func)(void));
 void *GetDMACallback(DMA_Channel dma);
 
 /**
- * @brief Initializes the interrupt dispatcher.
+ * @brief Enables, disables or sets the priority of a DMA channel.
+ *
+ * @details Enables the specified DMA channel and configures its priority (if
+ * priority >= 0) or disables it (if priority = -1). The priority value must be
+ * in 0-7 range, with 0 being the highest priority and 7 the lowest.
+ *
+ * All channels are disabled upon calling ResetCallback(); most libraries will
+ * re-enable them as needed. By default the priority is set to 3 for all
+ * channels.
+ *
+ * @param dma
+ * @param priority Priority in 0-7 range or -1 to disable the channel
+ * @return Previously set priority in 0-7 range, -1 if the channel was disabled
+ */
+int SetDMAPriority(DMA_Channel dma, int priority);
+
+/**
+ * @brief Gets the priority of a DMA channel.
+ *
+ * @details Returns the currently set priority value for the specified DMA
+ * channel in 0-7 range, with 0 being the highest priority and 7 the lowest.
+ * Returns -1 if the channel is not enabled.
+ *
+ * @param dma
+ * @return Priority in 0-7 range, -1 if the channel is disabled
+ *
+ * @see SetDMAPriority()
+ */
+int GetDMAPriority(DMA_Channel dma);
+
+/**
+ * @brief Initializes the interrupt dispatcher and DMA controller.
  *
  * @details Sets up the interrupt handling system, hooks the BIOS to dispatch
- * interrupts to the library and clears all registered callbacks. This function
- * must be called once at the beginning of the program, prior to registering
- * any IRQ or DMA callbacks.
+ * interrupts to the library, clears all registered callbacks and disables all
+ * DMA channels. This function must be called once at the beginning of the
+ * program, prior to registering any IRQ or DMA callbacks.
  *
  * ResetCallback() is called by psxgpu's ResetGraph(), so invoking it manually
  * is usually not required. Calling ResetCallback() after ResetGraph() will
  * actually result in improper initialization, as ResetGraph() registers
  * several callbacks used internally by psxgpu.
  *
- * @return 0 or -1 if the was already initialized
+ * @return 0 or -1 if the dispatcher was already initialized
  */
 int ResetCallback(void);
 
@@ -196,6 +226,11 @@ void RestartCallback(void);
  * Note that interrupts are (obviously) disabled until RestartCallback() is
  * called.
  *
+ * WARNING: any ongoing background processing or DMA transfer must be stopped
+ * before calling StopCallback(), otherwise crashes may occur. This includes
+ * flushing psxgpu's command queue using DrawSync(), stopping CD-ROM reading
+ * and calling StopPAD() to disable the BIOS controller driver if used.
+ *
  * @see RestartCallback()
  */
 void StopCallback(void);
@@ -203,5 +238,3 @@ void StopCallback(void);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxgpu.h b/libpsn00b/include/psxgpu.h
index 26e560f..d7f1ad5 100644
--- a/libpsn00b/include/psxgpu.h
+++ b/libpsn00b/include/psxgpu.h
@@ -1,10 +1,26 @@
 /*
  * PSn00bSDK GPU library
- * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __PSXGPU_H
-#define __PSXGPU_H
+/**
+ * @file psxgpu.h
+ * @brief GPU library header
+ *
+ * @details This library provides access to the PS1's GPU through a fully
+ * asynchronous command queue, which allows GPU commands to be batched and sent
+ * efficiently in the background without stalling the CPU. Helper structures
+ * and macros to initialize, generate and link GPU display lists in memory are
+ * also provided, in addition to support for asynchronous VRAM data transfers
+ * and a debug font API that can be used to easily draw text overlays for
+ * debugging purposes.
+ *
+ * This library is for the most part a drop-in replacement for the official
+ * SDK's GPU library and is only missing a handful of functions, mainly related
+ * to Kanji debug fonts and command queue pausing.
+ */
+
+#pragma once
 
 #include <stdint.h>
 #include <stddef.h>
@@ -28,6 +44,11 @@ typedef enum _GPU_VideoMode {
 	MODE_PAL	= 1
 } GPU_VideoMode;
 
+typedef enum _GPU_DrawOpType {
+	DRAWOP_TYPE_DMA		= 1,
+	DRAWOP_TYPE_GPU_IRQ	= 2
+} GPU_DrawOpType;
+
 /* Structure macros */
 
 #define setVector(v, _x, _y, _z) \
@@ -83,7 +104,7 @@ typedef enum _GPU_VideoMode {
 	(p)->u0 = (_u0), (p)->v0 = (_v0), \
 	(p)->u1 = (_u1), (p)->v1 = (_v1), \
 	(p)->u2 = (_u2), (p)->v2 = (_v2)
-	
+
 #define setUV4(p, _u0, _v0, _u1, _v1, _u2, _v2, _u3, _v3) \
 	(p)->u0 = (_u0), (p)->v0 = (_v0), \
 	(p)->u1 = (_u1), (p)->v1 = (_v1), \
@@ -101,9 +122,12 @@ typedef enum _GPU_VideoMode {
 #define setlen(p, _len)		(((P_TAG *) (p))->len = (uint8_t) (_len))
 #define setaddr(p, _addr)	(((P_TAG *) (p))->addr = (uint32_t) (_addr))
 #define setcode(p, _code)	(((P_TAG *) (p))->code = (uint8_t) (_code))
+#define setcode_T(p, _code)	(((P_TAG_T *) (p))->code = (uint8_t) (_code))
+
 #define getlen(p)			(((P_TAG *) (p))->len)
 #define getaddr(p)			(((P_TAG *) (p))->addr)
 #define getcode(p)			(((P_TAG *) (p))->code)
+#define getcode_T(p)		(((P_TAG_T *) (p))->code)
 
 #define nextPrim(p)			(void *) (0x80000000 | (((P_TAG *) (p))->addr))
 #define isendprim(p)		((((P_TAG *) (p))->addr) == 0xffffff)
@@ -114,16 +138,20 @@ typedef enum _GPU_VideoMode {
 
 #define setSemiTrans(p, abe) \
 	((abe) ? (getcode(p) |= 2) : (getcode(p) &= ~2))
+#define setSemiTrans_T(p, abe) \
+	((abe) ? (getcode_T(p) |= 2) : (getcode_T(p) &= ~2))
 
 #define setShadeTex(p, tge) \
 	((tge) ? (getcode(p) |= 1) : (getcode(p) &= ~1))
+#define setShadeTex_T(p, tge) \
+	((tge) ? (getcode_T(p) |= 1) : (getcode_T(p) &= ~1))
 
 #define getTPage(tp, abr, x, y) ( \
-	(((x)  /  64) & 15) | \
-	((((y) / 256) &  1) <<  4) | \
-	(((abr)       &  3) <<  5) | \
-	(((tp)        &  3) <<  7) | \
-	((((y) / 512) &  1) << 11) \
+	(((x) & 0x3c0) >> 6) | \
+	(((y) & 0x100) >> 4) | \
+	(((y) & 0x200) << 2) | \
+	(((abr) & 3) << 5) | \
+	(((tp)  & 3) << 7) \
 )
 
 #define getClut(x, y) (((y) << 6) | (((x) >> 4) & 0x3f))
@@ -147,59 +175,109 @@ typedef enum _GPU_VideoMode {
 #define setTile(p)		setlen(p,  3), setcode(p, 0x60)
 #define setLineF2(p)	setlen(p,  3), setcode(p, 0x40)
 #define setLineG2(p)	setlen(p,  4), setcode(p, 0x50)
-#define setLineF3(p)	setlen(p,  5), setcode(p, 0x48), \
-	(p)->pad = 0x55555555
-#define setLineG3(p)	setlen(p,  7), setcode(p, 0x58), \
-	(p)->pad = 0x55555555, (p)->p1 = 0, (p)->p2 = 0
-#define setLineF4(p)	setlen(p,  6), setcode(p, 0x4c), \
-	(p)->pad = 0x55555555
-#define setLineG4(p)	setlen(p,  9), setcode(p, 0x5c), \
-	(p)->pad = 0x55555555, (p)->p1 = 0, (p)->p2 = 0, (p)->p3 = 0
-#define setFill(p) 		setlen(p,  3), setcode(p, 0x02)
-#define setVram2Vram(p)	setlen(p,  8), setcode(p, 0x80), \
+#define setLineF3(p)	setlen(p,  5), setcode(p, 0x48), (p)->pad = 0x55555555
+#define setLineG3(p)	setlen(p,  7), setcode(p, 0x58), (p)->pad = 0x55555555, \
+	(p)->p1 = 0, (p)->p2 = 0
+#define setLineF4(p)	setlen(p,  6), setcode(p, 0x4c), (p)->pad = 0x55555555
+#define setLineG4(p)	setlen(p,  9), setcode(p, 0x5c), (p)->pad = 0x55555555, \
+	(p)->p1 = 0, (p)->p2 = 0, (p)->p3 = 0
+#define setFill(p)		setlen(p,  3), setcode(p, 0x02)
+#define setBlit(p)		setlen(p,  8), setcode(p, 0x80), \
 	(p)->pad[0] = 0, (p)->pad[1] = 0, (p)->pad[2] = 0, (p)->pad[3] = 0
 
-#define setDrawTPage(p, dfe, dtd, tpage) \
-	setlen(p, 1), \
+#define setPolyF3_T(p)	setcode_T(p, 0x20)
+#define setPolyFT3_T(p)	setcode_T(p, 0x24)
+#define setPolyG3_T(p)	setcode_T(p, 0x30)
+#define setPolyGT3_T(p)	setcode_T(p, 0x34)
+#define setPolyF4_T(p)	setcode_T(p, 0x28)
+#define setPolyFT4_T(p)	setcode_T(p, 0x2c)
+#define setPolyG4_T(p)	setcode_T(p, 0x38)
+#define setPolyGT4_T(p)	setcode_T(p, 0x3c)
+#define setSprt8_T(p)	setcode_T(p, 0x74)
+#define setSprt16_T(p)	setcode_T(p, 0x7c)
+#define setSprt_T(p)	setcode_T(p, 0x64)
+#define setTile1_T(p)	setcode_T(p, 0x68)
+#define setTile8_T(p)	setcode_T(p, 0x70)
+#define setTile16_T(p)	setcode_T(p, 0x78)
+#define setTile_T(p)	setcode_T(p, 0x60)
+#define setLineF2_T(p)	setcode_T(p, 0x40)
+#define setLineG2_T(p)	setcode_T(p, 0x50)
+#define setLineF3_T(p)	setcode_T(p, 0x48), (p)->pad = 0x55555555
+#define setLineG3_T(p)	setcode_T(p, 0x58), (p)->pad = 0x55555555, \
+	(p)->p1 = 0, (p)->p2 = 0
+#define setLineF4_T(p)	setcode_T(p, 0x4c), (p)->pad = 0x55555555
+#define setLineG4_T(p)	setcode_T(p, 0x5c), (p)->pad = 0x55555555, \
+	(p)->p1 = 0, (p)->p2 = 0, (p)->p3 = 0
+#define setFill_T(p)	setcode_T(p, 0x02)
+#define setBlit_T(p)	setcode_T(p, 0x80), \
+	(p)->pad[0] = 0, (p)->pad[1] = 0, (p)->pad[2] = 0, (p)->pad[3] = 0
+
+#define setDrawTPage_T(p, dfe, dtd, tpage) \
 	(p)->code[0] = (0xe1000000 | \
 		(tpage) | \
-		((dtd) <<  9) | \
-		((dfe) << 10) \
+		(((dtd) & 1) <<  9) | \
+		(((dfe) & 1) << 10) \
 	)
+#define setDrawTPage(p, dfe, dtd, tpage) \
+	setlen(p, 1), setDrawTPage_T(p, dfe, dtd, tpage)
 
-#define setDrawOffset(p, _x, _y) \
-	setlen(p, 1), \
-	(p)->code[0] = (0xe5000000 | \
-		((_x)  % 1024) | \
-		(((_y) % 1024) << 11) \
+#define setTexWindow_T(p, r) \
+	(p)->code[0] = (0xe2000000 | \
+		((r)->w  & 0x1f) | \
+		(((r)->h & 0x1f) <<  5) | \
+		(((r)->x & 0x1f) << 10) | \
+		(((r)->y & 0x1f) << 15) \
 	)
+#define setTexWindow(p, r) \
+	setlen(p, 1), setTexWindow_T(p, r)
 
-#define setDrawMask(p, sb, mt) \
-	setlen(p, 1), \
-	(p)->code[0] = (0xe6000000 | (sb) | ((mt) << 1))
-
-#define setDrawArea(p, r) \
-	setlen(p, 2), \
+#define setDrawAreaXY_T(p, _x0, _y0, _x1, _y1) \
 	(p)->code[0] = (0xe3000000 | \
-		((r)->x  % 1024) | \
-		(((r)->y % 1024) << 10) \
+		((_x0)  & 0x3ff) | \
+		(((_y0) & 0x3ff) << 10) \
 	), \
 	(p)->code[1] = (0xe4000000 | \
-		(((r)->x  + (r)->w - 1) % 1024) | \
-		((((r)->y + (r)->h - 1) % 1024) << 10) \
+		((_x1)  & 0x3ff) | \
+		(((_y1) & 0x3ff) << 10) \
 	)
+#define setDrawAreaXY(p, _x0, _y0, _x1, _y1) \
+	setlen(p, 2), setDrawAreaXY_T(p, _x0, _y0, _x1, _y1)
+
+#define setDrawArea_T(p, r) \
+	setDrawAreaXY_T(p, \
+		(r)->x, \
+		(r)->y, \
+		(r)->x + (r)->w - 1, \
+		(r)->y + (r)->h - 1 \
+	)
+#define setDrawArea(p, r) \
+	setlen(p, 2), setDrawArea_T(p, r)
 
-#define setTexWindow(p, r) \
-	setlen(p, 1), \
-	(p)->code[0] = (0xe2000000 | \
-		((r)->w  % 32) | \
-		(((r)->h % 32) <<  5) | \
-		(((r)->x % 32) << 10) | \
-		(((r)->y % 32) << 15) \
+#define setDrawOffset_T(p, _x, _y) \
+	(p)->code[0] = (0xe5000000 | \
+		((_x)  & 0x7ff) | \
+		(((_y) & 0x7ff) << 11) \
 	)
+#define setDrawOffset(p, _x, _y) \
+	setlen(p, 1), setDrawOffset_T(p, _x, _y)
+
+#define setDrawStp_T(p, pbw, mt) \
+	(p)->code[0] = (0xe6000000 | (pbw) | ((mt) << 1))
+#define setDrawStp(p, pbw, mt) \
+	setlen(p, 1), setDrawStp_T(p, pbw, mt)
+
+#define setDrawIRQ_T(p) \
+	(p)->code[0] = 0x1f000000
+#define setDrawIRQ(p) \
+	setlen(p, 1), setDrawIRQ_T(p)
 
 /* Primitive structure definitions */
 
+typedef struct _P_TAG_T {
+	uint32_t	color:24;
+	uint32_t	code:8;
+} P_TAG_T;
+
 typedef struct _P_TAG {
 	uint32_t	addr:24;
 	uint32_t	len:8;
@@ -212,25 +290,31 @@ typedef struct _P_COLOR {
 	uint32_t	pad:8;
 } P_COLOR;
 
-typedef struct _POLY_F3 {
-	uint32_t	tag;
+// These macros are used to define two variants of each primitive, a regular one
+// and a "tagless" one (_T suffix) without the OT/display list header.
+#define _DEF_PRIM(name, ...) \
+	typedef struct _##name##_T { __VA_ARGS__ } name##_T; \
+	typedef struct _##name { uint32_t tag; __VA_ARGS__ } name;
+#define _DEF_ALIAS(name, target) \
+	typedef struct _##target##_T name##_T; \
+	typedef struct _##target name;
+
+_DEF_PRIM(POLY_F3,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	int16_t		x1, y1;
 	int16_t		x2, y2;
-} POLY_F3;
+)
 
-typedef struct _POLY_F4 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_F4,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	int16_t		x1, y1;
 	int16_t		x2, y2;
 	int16_t		x3, y3;
-} POLY_F4;
+)
 
-typedef struct _POLY_FT3 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_FT3,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		u0, v0;
@@ -241,10 +325,9 @@ typedef struct _POLY_FT3 {
 	int16_t		x2, y2;
 	uint8_t		u2, v2;
 	uint16_t	pad;
-} POLY_FT3;
+)
 
-typedef struct _POLY_FT4 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_FT4,
 	uint8_t		r0, g0, b0, code;
 	uint16_t	x0, y0;
 	uint8_t		u0, v0;
@@ -258,20 +341,18 @@ typedef struct _POLY_FT4 {
 	int16_t		x3, y3;
 	uint8_t		u3, v3;
 	uint16_t	pad1;
-} POLY_FT4;
+)
 
-typedef struct _POLY_G3 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_G3,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		r1, g1, b1, pad0;
 	int16_t		x1, y1;
 	uint8_t		r2, g2, b2, pad1;
 	int16_t		x2, y2;
-} POLY_G3;
+)
 
-typedef struct _POLY_G4 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_G4,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		r1, g1, b1, pad0;
@@ -280,10 +361,9 @@ typedef struct _POLY_G4 {
 	int16_t		x2, y2;
 	uint8_t		r3, g3, b3, pad2;
 	int16_t		x3, y3;
-} POLY_G4;
+)
 
-typedef struct _POLY_GT3 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_GT3,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		u0, v0;
@@ -296,10 +376,9 @@ typedef struct _POLY_GT3 {
 	int16_t		x2, y2;
 	uint8_t		u2, v2;
 	uint16_t	pad2;
-} POLY_GT3;
+)
 
-typedef struct _POLY_GT4 {
-	uint32_t	tag;
+_DEF_PRIM(POLY_GT4,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		u0, v0;
@@ -316,34 +395,30 @@ typedef struct _POLY_GT4 {
 	int16_t		x3, y3;
 	uint8_t		u3, v3;
 	uint16_t	pad4;
-} POLY_GT4;
+)
 
-typedef struct _LINE_F2 {
-	uint32_t	tag;
+_DEF_PRIM(LINE_F2,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	int16_t		x1, y1;
-} LINE_F2;
+)
 
-typedef struct _LINE_G2 {
-	uint32_t	tag;
+_DEF_PRIM(LINE_G2,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		r1, g1, b1, p1;
 	int16_t		x1, y1;
-} LINE_G2;
+)
 
-typedef struct _LINE_F3 {
-	uint32_t	tag;
+_DEF_PRIM(LINE_F3,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	int16_t		x1, y1;
 	int16_t		x2, y2;
 	uint32_t	pad;
-} LINE_F3;
+)
 
-typedef struct _LINE_G3 {
-	uint32_t	tag;
+_DEF_PRIM(LINE_G3,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		r1, g1, b1, p1;
@@ -351,20 +426,18 @@ typedef struct _LINE_G3 {
 	uint8_t		r2, g2, b2, p2;
 	int16_t		x2, y2;
 	uint32_t	pad;
-} LINE_G3;
+)
 
-typedef struct _LINE_F4 {
-	uint32_t	tag;
+_DEF_PRIM(LINE_F4,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	int16_t		x1, y1;
 	int16_t		x2, y2;
 	int16_t		x3, y3;
 	uint32_t	pad;
-} LINE_F4;
+)
 
-typedef struct _LINE_G4 {
-	uint32_t	tag;
+_DEF_PRIM(LINE_G4,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		r1, g1, b1, p1;
@@ -374,88 +447,80 @@ typedef struct _LINE_G4 {
 	uint8_t		r3, g3, b3, p3;
 	int16_t		x3, y3;
 	uint32_t	pad;
-} LINE_G4;
+)
 
-typedef struct _TILE {
-	uint32_t	tag;
+_DEF_PRIM(TILE,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	int16_t		w, h;
-} TILE;
+)
 
-struct _TILE_FIXED {
-	uint32_t	tag;
+_DEF_PRIM(TILE_1,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
-};
-typedef struct _TILE_FIXED TILE_1;
-typedef struct _TILE_FIXED TILE_8;
-typedef struct _TILE_FIXED TILE_16;
+)
+_DEF_ALIAS(TILE_8,  TILE_1)
+_DEF_ALIAS(TILE_16, TILE_1)
 
-typedef struct _SPRT {
-	uint32_t	tag;
+_DEF_PRIM(SPRT,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		u0, v0;
 	uint16_t	clut;
 	uint16_t	w, h;
-} SPRT;
+)
 
-struct _SPRT_FIXED {
-	uint32_t	tag;
+_DEF_PRIM(SPRT_1,
 	uint8_t		r0, g0, b0, code;
 	int16_t		x0, y0;
 	uint8_t		u0, v0;
 	uint16_t	clut;
-};
-typedef struct _SPRT_FIXED SPRT_8;
-typedef struct _SPRT_FIXED SPRT_16;
-
-typedef struct _DR_ENV {
-	uint32_t tag;
-	uint32_t code[8];
-} DR_ENV;
-
-typedef struct _DR_AREA {
-	uint32_t tag;
-	uint32_t code[2];
-} DR_AREA;
-
-typedef struct _DR_OFFSET {
-	uint32_t tag;
-	uint32_t code[1];
-} DR_OFFSET;
-
-typedef struct _DR_TWIN {
-	uint32_t tag;
-	uint32_t code[2];
-} DR_TWIN;
-
-typedef struct _DR_TPAGE {
-	uint32_t tag;
-	uint32_t code[1];
-} DR_TPAGE;
-
-typedef struct _DR_MASK {
-	uint32_t tag;
-	uint32_t code[1];
-} DR_MASK;
+)
+_DEF_ALIAS(SPRT_8,  SPRT_1)
+_DEF_ALIAS(SPRT_16, SPRT_1)
 
-typedef struct _FILL {
-	uint32_t	tag;
+_DEF_PRIM(FILL,
 	uint8_t		r0, g0, b0, code;
-	uint16_t	x0, y0; // Note: coordinates must be in 16 pixel steps
+	uint16_t	x0, y0;
 	uint16_t	w, h;
-} FILL;
+)
 
-typedef struct _VRAM2VRAM {
-	uint32_t	tag;
+_DEF_PRIM(DR_MOVE,
 	uint8_t		p0, p1, p2, code;
 	uint16_t	x0, y0;
 	uint16_t	x1, y1;
 	uint16_t	w, h;
-	uint32_t	pad[4];
-} VRAM2VRAM;
+)
+
+_DEF_PRIM(DR_AREA,
+	uint32_t code[2];
+)
+_DEF_PRIM(DR_OFFSET,
+	uint32_t code[1];
+)
+_DEF_PRIM(DR_TWIN,
+	uint32_t code[1];
+)
+_DEF_PRIM(DR_TPAGE,
+	uint32_t code[1];
+)
+_DEF_PRIM(DR_STP,
+	uint32_t code[1];
+)
+_DEF_PRIM(DR_IRQ,
+	uint32_t code[1];
+)
+
+_DEF_PRIM(DR_ENV,
+	DR_TPAGE_T	tpage;
+	DR_TWIN_T	twin;
+	DR_AREA_T	area;
+	DR_OFFSET_T	offset;
+	FILL_T		fill;
+)
+
+#undef _DEF_PRIM
+#undef _DEF_ALIAS
 
 /* Structure definitions */
 
@@ -478,13 +543,13 @@ typedef struct _DISPENV {
 typedef struct _DRAWENV {
 	RECT		clip;		// Drawing area
 	int16_t		ofs[2];		// GPU draw offset (relative to draw area)
-	RECT		tw;			// Texture window (doesn't do anything atm)
+	RECT		tw;			// Texture window
 	uint16_t	tpage;		// Initial tpage value
 	uint8_t		dtd;		// Dither processing flag (simply OR'ed to tpage)
 	uint8_t		dfe;		// Drawing to display area blocked/allowed (simply OR'ed to tpage)
 	uint8_t		isbg;		// Clear draw area if non-zero
 	uint8_t		r0, g0, b0;	// Draw area clear color (if isbg iz nonzero)
-	DR_ENV		dr_env;		// Draw mode packet area (used by PutDrawEnv)
+	DR_ENV		dr_env;		// GPU primitive cache area (used internally)
 } DRAWENV;
 
 typedef struct _TIM_IMAGE {
@@ -521,31 +586,35 @@ void PutDrawEnv(DRAWENV *env);
 void PutDrawEnvFast(DRAWENV *env);
 
 int GetODE(void);
+int IsIdleGPU(int timeout);
 int VSync(int mode);
 void *VSyncHaltFunction(void (*func)(void));
 void *VSyncCallback(void (*func)(void));
 
-int EnqueueDrawOp(
-	void		(*func)(uint32_t, uint32_t, uint32_t),
-	uint32_t	arg1,
-	uint32_t	arg2,
-	uint32_t	arg3
-);
+void SetDrawOpType(GPU_DrawOpType type);
+int EnqueueDrawOp(void (*func)(), uint32_t arg1, uint32_t arg2, uint32_t arg3);
 int DrawSync(int mode);
 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);
 int DrawOTag(const uint32_t *ot);
+int DrawOTagIRQ(const uint32_t *ot);
 int DrawOTagEnv(const uint32_t *ot, DRAWENV *env);
+int DrawOTagEnvIRQ(const uint32_t *ot, DRAWENV *env);
 void DrawOTag2(const uint32_t *ot);
+void DrawOTagIRQ2(const uint32_t *ot);
+int DrawBuffer(const uint32_t *buf, size_t length);
+int DrawBufferIRQ(const uint32_t *buf, size_t length);
+void DrawBuffer2(const uint32_t *buf, size_t length);
+void DrawBufferIRQ2(const uint32_t *buf, size_t length);
 void DrawPrim(const uint32_t *pri);
 
 void AddPrim(uint32_t *ot, const void *pri);
@@ -565,5 +634,3 @@ char *FntFlush(int id);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxgte.h b/libpsn00b/include/psxgte.h
index 91dfd6a..2200a55 100644
--- a/libpsn00b/include/psxgte.h
+++ b/libpsn00b/include/psxgte.h
@@ -14,8 +14,7 @@
  * registers and issue commands to the GTE.
  */
 
-#ifndef __PSXGTE_H
-#define __PSXGTE_H
+#pragma once
 
 #include <stdint.h>
 
@@ -259,5 +258,3 @@ void Square0(VECTOR *v0, VECTOR *v1);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxkernel.h b/libpsn00b/include/psxkernel.h
deleted file mode 100644
index 0c55bcb..0000000
--- a/libpsn00b/include/psxkernel.h
+++ /dev/null
@@ -1,45 +0,0 @@
-#ifndef _PSXKERNEL_H
-#define _PSXKERNEL_H
-
-// Event descriptors
-#define DescMask	0xff000000		// Event descriptor mask
-#define DescTH		DescMask
-#define DescHW		0xf0000000		// Hardware event (IRQ)
-#define DescEV		0xf1000000		// Event event
-#define DescRC		0xf2000000		// Root counter event
-#define DescUEV		0xf3000000		// User event
-#define DescSW		0xf4000000		// BIOS event
-
-// Hardware events
-#define	HwVBLANK	(DescHW|0x01)	// VBlank
-#define HwGPU		(DescHW|0x02)	// GPU
-#define HwCdRom		(DescHW|0x03)	// CDROM
-#define HwDMAC		(DescHW|0x04)	// DMA
-#define HwRTC0		(DescHW|0x05)	// Timer 0
-#define HwRTC1		(DescHW|0x06)	// Timer 1
-#define HwRTC2		(DescHW|0x07)	// Timer 2
-#define HwCNTL		(DescHW|0x08)	// Controller
-#define HwSPU		(DescHW|0x09)	// SPU
-#define HwPIO		(DescHW|0x0a)	// PIO & lightgun
-#define HwSIO		(DescHW|0x0b)	// Serial
-
-#define HwCPU		(DescHW|0x10)	// Processor exception
-#define HwCARD		(DescHW|0x11)	// Memory card (lower level BIOS functions)
-#define HwCard_0	(DescHW|0x12)	// Memory card (unused)
-#define HwCard_1	(DescHW|0x13)	// Memory card (unused)
-#define SwCARD		(DescSW|0x01)	// Memory card (higher level BIOS functions)
-#define SwMATH		(DescSW|0x02)	// Libmath related apparently, unknown purpose
-
-#define RCntCNT0	(DescRC|0x00)	// Root counter 0 (dot clock)
-#define RCntCNT1	(DescRC|0x01)	// Horizontal sync
-#define RCntCNT2	(DescRC|0x02)	// 1/8 of system clock
-#define RCntCNT3	(DescRC|0x03)	// Vertical blank
-
-#define RCntMdINTR		0x1000		// General interrupt
-#define RCntMdNOINTR	0x2000		// New device
-#define RCntMdSC		0x0001		// Counter becomes zero
-#define RCntMdSP		0x0000		// Unknown purpose
-#define RCntMdFR		0x0000
-#define RCntMdGATE		0x0010		// Command acknowledged
-
-#endif // _PSXKERNEL_H
\ No newline at end of file
diff --git a/libpsn00b/include/psxpad.h b/libpsn00b/include/psxpad.h
index 32f7f8a..09f28c4 100644
--- a/libpsn00b/include/psxpad.h
+++ b/libpsn00b/include/psxpad.h
@@ -11,8 +11,7 @@
  * Reference: https://gist.github.com/scanlime/5042071
  */
  
-#ifndef _PSXPAD_H
-#define _PSXPAD_H
+#pragma once
 
 #include <stdint.h>
 
@@ -234,5 +233,3 @@ typedef struct __attribute__((packed)) _MemCardRequest {
 	uint8_t checksum;	// = lba_h ^ lba_l ^ data (CMD_WRITE only)
 	uint8_t dummy2[3];
 } MemCardRequest;
-
-#endif
\ No newline at end of file
diff --git a/libpsn00b/include/psxpress.h b/libpsn00b/include/psxpress.h
index dc1d52c..f26e030 100644
--- a/libpsn00b/include/psxpress.h
+++ b/libpsn00b/include/psxpress.h
@@ -1,6 +1,6 @@
 /*
  * PSn00bSDK MDEC library
- * (C) 2022 spicyjpeg - MPL licensed
+ * (C) 2022-2023 spicyjpeg - MPL licensed
  */
 
 /**
@@ -17,11 +17,12 @@
  * FMV playback is not part of this library per se, but can implemented using
  * the APIs defined here alongside some code to stream data from the CD drive.
  *
- * Currently only version 1 and 2 .BS files are supported.
+ * Currently bitstream versions 1, 2 and 3 are supported. Version 0 and .IKI
+ * bitstreams are not supported, but no encoder is publicly available for those
+ * anyway.
  */
 
-#ifndef __PSXPRESS_H
-#define __PSXPRESS_H
+#pragma once
 
 #include <stdint.h>
 #include <stddef.h>
@@ -34,28 +35,26 @@ typedef struct _DECDCTENV {
 	int16_t dct[64];	// Inverse DCT matrix (2.14 fixed-point)
 } DECDCTENV;
 
-// This is the "small" lookup table used by DecDCTvlc(). It can be copied to
-// the scratchpad.
+typedef struct _VLC_TableV2 {
+	uint16_t ac0[2];
+	uint32_t ac2[8], ac3[64];
+	uint16_t ac4[8], ac5[8], ac7[16], ac8[32];
+	uint16_t ac9[32], ac10[32], ac11[32], ac12[32];
+} VLC_TableV2;
+
+typedef struct _VLC_TableV3 {
+	uint16_t ac0[2];
+	uint32_t ac2[8], ac3[64];
+	uint16_t ac4[8], ac5[8], ac7[16], ac8[32];
+	uint16_t ac9[32], ac10[32], ac11[32], ac12[32];
+	uint8_t  dc[128], dc_len[9];
+	uint8_t  _reserved[3];
+} VLC_TableV3;
+
 typedef struct _DECDCTTAB {
-	uint16_t	lut0[2];
-	uint32_t	lut2[8];
-	uint32_t	lut3[64];
-	uint16_t	lut4[8];
-	uint16_t	lut5[8];
-	uint16_t	lut7[16];
-	uint16_t	lut8[32];
-	uint16_t	lut9[32];
-	uint16_t	lut10[32];
-	uint16_t	lut11[32];
-	uint16_t	lut12[32];
+	uint32_t ac[8192], ac00[512];
 } DECDCTTAB;
 
-// This is the "large" table used by DecDCTvlc2().
-typedef struct _DECDCTTAB2 {
-	uint32_t	lut[8192];
-	uint32_t	lut00[512];
-} DECDCTTAB2;
-
 typedef enum _DECDCTMODE {
 	DECDCT_MODE_24BPP		= 1,
 	DECDCT_MODE_16BPP		= 0,
@@ -66,8 +65,9 @@ typedef enum _DECDCTMODE {
 typedef struct _VLC_Context {
 	const uint32_t	*input;
 	uint32_t		window, next_window, remaining;
-	uint16_t		quant_scale;
 	int8_t			is_v3, bit_offset, block_index, coeff_index;
+	uint16_t		quant_scale;
+	int16_t			last_y, last_cr, last_cb;
 } VLC_Context;
 
 // Despite what some docs claim, the "number of 32-byte blocks" and "always
@@ -233,8 +233,9 @@ int DecDCToutSync(int mode);
  * frame) into a buffer that can be passed to DecDCTin(). This function uses a
  * small (<1 KB) lookup table combined with the GTE to accelerate the process;
  * performance is roughly on par with DecDCTvlcStart2() if the lookup table
- * is copied to the scratchpad beforehand by calling DecDCTvlcCopyTable(). The
- * contents of the GTE's LZCR register, if any, will be destroyed.
+ * is copied to the scratchpad beforehand by calling DecDCTvlcCopyTableV2() or
+ * DecDCTvlcCopyTableV3(). The contents of the GTE's LZCS and LZCR registers,
+ * if any, will be destroyed.
  *
  * A VLC_Context object must be created and passed to this function, which will
  * then proceed to initialize its fields. The max_size argument sets the
@@ -244,8 +245,6 @@ int DecDCToutSync(int mode);
  * can be different). If max_size = 0, the entire frame will always be decoded
  * in one shot.
  *
- * Only bitstream version 2 is currently supported.
- *
  * WARNING: InitGeom() must be called prior to using DecDCTvlcStart() for the
  * first time. Attempting to call this function with the GTE disabled will
  * result in a crash.
@@ -256,7 +255,7 @@ int DecDCToutSync(int mode);
  * @param bs
  * @return 0, 1 if more data needs to be output or -1 in case of failure
  *
- * @see DecDCTvlcContinue(), DecDCTvlcCopyTable()
+ * @see DecDCTvlcContinue(), DecDCTvlcCopyTableV2(), DecDCTvlcCopyTableV3()
  */
 int DecDCTvlcStart(VLC_Context *ctx, uint32_t *buf, size_t max_size, const uint32_t *bs);
 
@@ -275,7 +274,8 @@ int DecDCTvlcStart(VLC_Context *ctx, uint32_t *buf, size_t max_size, const uint3
  * context returned 0; in that case the context shall be discarded or reused to
  * decode another bitstream.
  *
- * The contents of the GTE's LZCR register, if any, will be destroyed.
+ * The contents of the GTE's LZCS and LZCR registers, if any, will be
+ * destroyed.
  *
  * See DecDCTvlcStart() for more details.
  *
@@ -309,7 +309,7 @@ int DecDCTvlcContinue(VLC_Context *ctx, uint32_t *buf, size_t max_size);
  * @param buf
  * @return 0, 1 if more data needs to be output or -1 in case of failure
  *
- * @see DecDCTvlcSize(), DecDCTvlcCopyTable()
+ * @see DecDCTvlcSize(), DecDCTvlcCopyTableV2(), DecDCTvlcCopyTableV3()
  */
 int DecDCTvlc(const uint32_t *bs, uint32_t *buf);
 
@@ -332,23 +332,60 @@ int DecDCTvlc(const uint32_t *bs, uint32_t *buf);
 size_t DecDCTvlcSize(size_t size);
 
 /**
- * @brief Moves the lookup table used by the .BS decompressor to the scratchpad
- * region.
+ * @brief Copies the lookup tables used by the .BS decompressor (v1/v2) to the
+ * scratchpad region.
  *
- * @details Copies the small (<1 KB) lookup table used by DecDCTvlcContinue(),
- * DecDCTvlcStart() and DecDCTvlc() (a DECDCTTAB structure) to the specified
- * address. A copy of this table is always present in main RAM, however this
- * function can be used to copy it to the scratchpad region to boost
- * decompression performance.
+ * @details Copies the lookup table used by DecDCTvlcContinue(),
+ * DecDCTvlcStart() and DecDCTvlc() to the specified address. A copy of this
+ * table is always present in main RAM, however this function can be used to
+ * copy it to the scratchpad region to boost decompression performance.
+ *
+ * This function copies a 676-byte table (VLC_TableV2 structure) containing
+ * only the data necessary for decoding version 1 and 2 bitstreams, to help
+ * save scratchpad space. If support for version 3 is required,
+ * DecDCTvlcCopyTableV3() can be used instead to copy the full 816-byte table.
  *
  * The address passed to this function is saved. Calls to DecDCTvlcStart(),
  * DecDCTvlcContinue() and DecDCTvlc() will automatically use the last table
- * copied. Call DecDCTvlcCopyTable(0) to revert to using the library's internal
- * table in main RAM.
+ * copied. Call DecDCTvlcCopyTableV2(0) or DecDCTvlcCopyTableV3(0) to revert to
+ * using the library's internal table in main RAM.
+ *
+ * WARNING: attempting to decode a version 3 .BS file or .STR frame after
+ * calling this function will result in undefined behavior and potentially a
+ * crash. To re-enable version 3 decoding, use DecDCTvlcCopyTableV3() to copy
+ * the full table to the scratchpad or revert to using the built-in table in
+ * main RAM.
+ *
+ * @param addr Pointer to free 676-byte area in scratchpad region or 0 to reset
  *
- * @param addr Pointer to free area in scratchpad region or 0 to reset
+ * @see DecDCTvlcCopyTableV3()
  */
-void DecDCTvlcCopyTable(DECDCTTAB *addr);
+void DecDCTvlcCopyTableV2(VLC_TableV2 *addr);
+
+/**
+ * @brief Copies the lookup tables used by the .BS decompressor (v1/v2/v3) to
+ * the scratchpad region.
+ *
+ * @details Copies the lookup table used by DecDCTvlcContinue(),
+ * DecDCTvlcStart() and DecDCTvlc() to the specified address. A copy of this
+ * table is always present in main RAM, however this function can be used to
+ * copy it to the scratchpad region to boost decompression performance.
+ *
+ * This function copies the full 816-byte table (VLC_TableV3 structure),
+ * including the data used to decode version 3 bitstreams. If support for
+ * version 3 is not required, DecDCTvlcCopyTableV2() can be used instead to
+ * save scratchpad space by only copying the first 676 bytes of the table.
+ *
+ * The address passed to this function is saved. Calls to DecDCTvlcStart(),
+ * DecDCTvlcContinue() and DecDCTvlc() will automatically use the last table
+ * copied. Call DecDCTvlcCopyTableV2(0) or DecDCTvlcCopyTableV3(0) to revert to
+ * using the library's internal table in main RAM.
+ *
+ * @param addr Pointer to free 816-byte area in scratchpad region or 0 to reset
+ *
+ * @see DecDCTvlcCopyTableV2()
+ */
+void DecDCTvlcCopyTableV3(VLC_TableV3 *addr);
 
 /**
  * @brief Decompresses or begins decompressing a .BS file into MDEC codes
@@ -360,8 +397,8 @@ void DecDCTvlcCopyTable(DECDCTTAB *addr);
  * calling DecDCTvlcBuild(), but does not use the GTE nor the scratchpad.
  * Depending on the specific bitstream being decoded DecDCTvlcStart2() might be
  * slightly faster or slower than DecDCTvlcStart() with its lookup table copied
- * to the scratchpad (see DecDCTvlcCopyTable()). DecDCTvlcStart() with the
- * table in main RAM tends to be much slower.
+ * to the scratchpad (see DecDCTvlcCopyTableV2() and DecDCTvlcCopyTableV3()).
+ * DecDCTvlcStart() with the table in main RAM tends to be much slower.
  *
  * A VLC_Context object must be created and passed to this function, which will
  * then proceed to initialize its fields. The max_size argument sets the
@@ -371,7 +408,8 @@ void DecDCTvlcCopyTable(DECDCTTAB *addr);
  * buffer can be different). If max_size = 0, the entire frame will always be
  * decoded in one shot.
  *
- * Only bitstream version 2 is currently supported.
+ * This function only supports decoding version 1 and 2 bitstreams. Use
+ * DecDCTvlcStart() to decode a version 3 bitstream.
  *
  * @param ctx Pointer to VLC_Context structure (which will be initialized)
  * @param buf
@@ -432,7 +470,7 @@ int DecDCTvlcContinue2(VLC_Context *ctx, uint32_t *buf, size_t max_size);
  *
  * @see DecDCTvlcSize2(), DecDCTvlcBuild()
  */
-int DecDCTvlc2(const uint32_t *bs, uint32_t *buf, DECDCTTAB2 *table);
+int DecDCTvlc2(const uint32_t *bs, uint32_t *buf, DECDCTTAB *table);
 
 /**
  * @brief Sets the maximum amount of data to be decompressed (alternate
@@ -458,7 +496,7 @@ size_t DecDCTvlcSize2(size_t size);
  * the .BS decompressor.
  *
  * @details Generates the lookup table required by DecDCTvlcStart2(),
- * DecDCTvlcContinue2() and DecDCTvlc2() (a DECDCTTAB2 structure) into the
+ * DecDCTvlcContinue2() and DecDCTvlc2() (a DECDCTTAB structure) into the
  * specified buffer. Since the table is relatively large (34 KB), it is
  * recommended to only generate it in a dynamically-allocated buffer when
  * needed and deallocate the buffer afterwards.
@@ -468,10 +506,8 @@ size_t DecDCTvlcSize2(size_t size);
  *
  * @param table
  */
-void DecDCTvlcBuild(DECDCTTAB2 *table);
+void DecDCTvlcBuild(DECDCTTAB *table);
 
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxsio.h b/libpsn00b/include/psxsio.h
index 449e43a..8932830 100644
--- a/libpsn00b/include/psxsio.h
+++ b/libpsn00b/include/psxsio.h
@@ -18,8 +18,7 @@
  * debugging purposes.
  */
 
-#ifndef __PSXSIO_H
-#define __PSXSIO_H
+#pragma once
 
 #include <stdint.h>
 
@@ -280,5 +279,3 @@ void DelSIO(void);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/psxsn.h b/libpsn00b/include/psxsn.h
new file mode 100644
index 0000000..1acbc18
--- /dev/null
+++ b/libpsn00b/include/psxsn.h
@@ -0,0 +1,51 @@
+/*
+ * PSn00bSDK kernel API library (host file access)
+ * (C) 2023 spicyjpeg - MPL licensed
+ */
+
+/**
+ * @file psxsn.h
+ * @brief Host file access API header
+ *
+ * @details This header provides stubs for the PCDRV API, which grants read and
+ * write access to a directory on the host's filesystem when the executable is
+ * running on an emulator or through a debugger that supports the PCDRV
+ * protocol, such as Unirom or pcsx-redux. These functions are completely
+ * separate and independent from the BIOS file API and do not register any
+ * device drivers.
+ *
+ * Note that in the official SDK these functions are provided by libsn, while
+ * in PSn00bSDK they are part of libpsxapi.
+ */
+
+#pragma once
+
+#include <stddef.h>
+
+typedef enum _PCDRV_OpenMode {
+	PCDRV_MODE_READ       = 0,
+	PCDRV_MODE_WRITE      = 1,
+	PCDRV_MODE_READ_WRITE = 2
+} PCDRV_OpenMode;
+
+typedef enum _PCDRV_SeekMode {
+	PCDRV_SEEK_SET = 0,
+	PCDRV_SEEK_CUR = 1,
+	PCDRV_SEEK_END = 2
+} PCDRV_SeekMode;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int PCinit(void);
+int PCcreat(const char *path);
+int PCopen(const char *path, PCDRV_OpenMode mode);
+int PCclose(int fd);
+int PCread(int fd, void *data, size_t length);
+int PCwrite(int fd, const void *data, size_t length);
+int PClseek(int fd, int offset, PCDRV_SeekMode mode);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/libpsn00b/include/psxspu.h b/libpsn00b/include/psxspu.h
index cdc3ac7..b544952 100644
--- a/libpsn00b/include/psxspu.h
+++ b/libpsn00b/include/psxspu.h
@@ -1,10 +1,25 @@
 /*
  * PSn00bSDK SPU library
- * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __PSXSPU_H
-#define __PSXSPU_H
+/**
+ * @file psxspu.h
+ * @brief SPU library header
+ *
+ * @details The PSn00bSDK SPU library allows for SPU initialization, DMA
+ * transfers (both sample data uploads and capture buffer reads) and provides
+ * helper macros for accessing SPU control registers, which can be used to
+ * control sample playback on each channel, configure reverb and enable more
+ * advanced features such as interrupts.
+ *
+ * This library currently has fewer functions than its Sony SDK counterpart, in
+ * part because it is not yet complete but also since the vast majority of the
+ * Sony library's functions are redundant, inefficient and can be replaced with
+ * simple SPU register writes.
+ */
+
+#pragma once
 
 #include <stdint.h>
 #include <stddef.h>
@@ -12,6 +27,7 @@
 
 /* Definitions */
 
+#if 0
 typedef enum _SPU_AttrMask {
 	SPU_VOICE_VOLL			= 1 << 0,	// Left volume
 	SPU_VOICE_VOLR			= 1 << 1,	// Right volume
@@ -33,6 +49,7 @@ typedef enum _SPU_AttrMask {
 	SPU_VOICE_ADSR_ADSR1	= 1 << 17,
 	SPU_VOICE_ADSR_ADSR2	= 1 << 18
 } SPU_AttrMask;
+#endif
 
 typedef enum _SPU_TransferMode {
 	SPU_TRANSFER_BY_DMA	= 0,
@@ -46,6 +63,7 @@ typedef enum _SPU_WaitMode {
 
 /* Structure definitions */
 
+#if 0
 typedef struct _SpuVolume {
 	int16_t left, right;
 } SpuVolume;
@@ -72,6 +90,7 @@ typedef struct _SpuCommonAttr {
 	SpuVolume	mvol, mvolmode, mvolx;
 	SpuExtAttr	cd, ext;
 } SpuCommonAttr;
+#endif
 
 /* Macros */
 
@@ -137,11 +156,11 @@ size_t SpuRead(uint32_t *data, size_t size);
 size_t SpuWrite(const uint32_t *data, size_t size);
 size_t SpuWritePartly(const uint32_t *data, size_t size);
 SPU_TransferMode SpuSetTransferMode(SPU_TransferMode mode);
+SPU_TransferMode SpuGetTransferMode(void);
 uint32_t SpuSetTransferStartAddr(uint32_t addr);
+uint32_t SpuGetTransferStartAddr(void);
 int SpuIsTransferCompleted(int mode);
 
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/stdio.h b/libpsn00b/include/stdio.h
index 8aaf4c7..1bb5b74 100644
--- a/libpsn00b/include/stdio.h
+++ b/libpsn00b/include/stdio.h
@@ -1,39 +1,26 @@
-#ifndef _STDIO_H
-#define _STDIO_H
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
+ */
 
-#include <stdarg.h>
+#pragma once
 
-// BIOS seek modes
-#ifndef SEEK_SET
-#define SEEK_SET	0
-#endif
-#ifndef SEEK_CUR
-#define SEEK_CUR	1
-#endif
-#ifndef SEEK_END
-#define SEEK_END	2		/* warning: reportedly buggy */
-#endif
+#include <stdarg.h>
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-// The following functions use the BIOS
-extern void printf (const char *__format, ...);
-
-extern int getc(int __fd);
-extern int putc(int __char, int __fd);
+/* String I/O API (provided by BIOS) */
 
-#define fputc(__char, __fd) 	putc(__char, __fd)
-#define fgetc(__char, __fd) 	getc(__char, __fd)
+int printf(const char *fmt, ...);
+char *gets(char *str);
+void puts(const char *str);
+int getchar(void);
+void putchar(int ch);
 
-// Console TTY
-extern void gets(char *__s);
-extern void puts(const char *__s);
-extern int getchar(void);
-extern void putchar(int __c);
+/* String formatting API (built-in) */
 
-// The following functions do not use the BIOS
 int vsnprintf(char *string, unsigned int size, const char *fmt, va_list ap);
 int vsprintf(char *string, const char *fmt, va_list ap);
 int sprintf(char *string, const char *fmt, ...);
@@ -45,5 +32,3 @@ int sscanf(const char *str, const char *fmt, ...);
 #ifdef __cplusplus
 }
 #endif
-
-#endif	// _STDIO_H
\ No newline at end of file
diff --git a/libpsn00b/include/stdlib.h b/libpsn00b/include/stdlib.h
index 049d067..19761df 100644
--- a/libpsn00b/include/stdlib.h
+++ b/libpsn00b/include/stdlib.h
@@ -1,10 +1,9 @@
 /*
  * PSn00bSDK standard library
- * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __STDLIB_H
-#define __STDLIB_H
+#pragma once
 
 #include <stddef.h>
 
@@ -39,11 +38,11 @@ void srand(int seed);
 int abs(int j);
 long labs(long i);
 
-long strtol(const char *nptr, char **endptr, int base);
-long long strtoll(const char *nptr, char **endptr, int base);
-float strtof(const char *nptr, char **endptr);
-double strtod(const char *nptr, char **endptr);
-long double strtold(const char *nptr, char **endptr);
+long strtol(const char *str, char **str_end, int base);
+long long strtoll(const char *str, char **str_end, int base);
+//float strtof(const char *str, char **str_end);
+//double strtod(const char *str, char **str_end);
+//long double strtold(const char *str, char **str_end);
 
 void InitHeap(void *addr, size_t size);
 void *sbrk(ptrdiff_t incr);
@@ -59,5 +58,3 @@ void free(void *ptr);
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/string.h b/libpsn00b/include/string.h
index ceee066..6310b1a 100644
--- a/libpsn00b/include/string.h
+++ b/libpsn00b/include/string.h
@@ -1,37 +1,40 @@
 /*
  * PSn00bSDK standard library
- * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ * (C) 2019-2023 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __STRING_H
-#define __STRING_H
+#pragma once
+
+#include <stddef.h>
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-int strcmp(const char *dst , const char *src);
-int strncmp(const char *dst , const char *src , int len);
-char *strpbrk(const char *dst , const char *src);
-char *strtok(char *s , char *set);
-char *strstr(const char *big , const char *little);
-
-char *strcat(char *s , const char *append);
-char *strncat(char *s , const char *append, int n);
-char *strcpy(char *dst , const char *src);
-char *strncpy(char *dst , const char *src , int n);
-int strlen(const char *s);
-char *strchr(const char *s , int c);
-char *strrchr(const char *s , int c);
-
-void *memmove(void *dst , const void *src , int n);
-void *memchr(void *s , int c , int n);
-void *memcpy(void *dst , const void *src , int n);
-void *memset(void *dst , char c , int n);
-int memcmp(const void *b1 , const void *b2 , int n);
+void *memset(void *dest, int ch, size_t count);
+void *memcpy(void *dest, const void *src, size_t count);
+void *memccpy(void *dest, const void *src, int ch, size_t count);
+void *memmove(void *dest, const void *src, size_t count);
+int memcmp(const void *lhs, const void *rhs, size_t count);
+void *memchr(const void *ptr, int ch, size_t count);
+
+char *strcpy(char *dest, const char *src);
+char *strncpy(char *dest, const char *src, size_t count);
+int strcmp(const char *lhs, const char *rhs);
+int strncmp(const char *lhs, const char *rhs, size_t count);
+char *strchr(const char *str, int ch);
+char *strrchr(const char *str, int ch);
+char *strpbrk(const char *str, const char *breakset);
+char *strstr(const char *str, const char *substr);
+
+size_t strlen(const char *str);
+char *strcat(char *dest, const char *src);
+char *strncat(char *dest, const char *src, size_t count);
+char *strdup(const char *str);
+char *strndup(const char *str, size_t count);
+
+char *strtok(char *str, const char *delim);
 
 #ifdef __cplusplus
 }
 #endif
-
-#endif
diff --git a/libpsn00b/include/strings.h b/libpsn00b/include/strings.h
index 7223ab9..0595637 100644
--- a/libpsn00b/include/strings.h
+++ b/libpsn00b/include/strings.h
@@ -3,8 +3,7 @@
  * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef __STRINGS_H
-#define __STRINGS_H
+#pragma once
 
 #include <string.h>
 
@@ -15,5 +14,3 @@
 #define bcmp(b1, b2, len)		memcmp(b1, b2, len)
 #define index(s, c)				strchr(s, c)
 #define rindex(s, c)			strrchr(s, c)
-
-#endif
diff --git a/libpsn00b/include/sys/fcntl.h b/libpsn00b/include/sys/fcntl.h
index dfbf5b2..54c2d05 100644
--- a/libpsn00b/include/sys/fcntl.h
+++ b/libpsn00b/include/sys/fcntl.h
@@ -1,8 +1,10 @@
-#ifndef _SYS_FCNTL_H
-#define _SYS_FCNTL_H
+/*
+ * PSn00bSDK kernel API library
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#pragma once
 
-// File control mode flags for BIOS file functions
-// (many weren't documented in nocash docs)
 #define FREAD			0x1			// Read
 #define FWRITE			0x2			// Write
 #define FNBLOCK			0x4			// Non-blocking read access
@@ -16,5 +18,3 @@
 #define FNBUF			0x4000		// No ring buffer and terminal interrupt
 #define FASYNC			0x8000		// Asynchronous I/O
 #define FNBLOCKS(a)		(a<<16)		// Number of blocks? (from nocash docs)
-
-#endif
\ No newline at end of file
diff --git a/libpsn00b/include/sys/ioctl.h b/libpsn00b/include/sys/ioctl.h
new file mode 100644
index 0000000..af65e5d
--- /dev/null
+++ b/libpsn00b/include/sys/ioctl.h
@@ -0,0 +1,13 @@
+/*
+ * PSn00bSDK kernel API library
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#pragma once
+
+#define EOF			-1
+
+#define FIONBLOCK	(('f'<<8)|1)
+#define FIOCSCAN	(('f'<<8)|2)
+
+#define DIOFORMAT	(('d'<<8)|1)
diff --git a/libpsn00b/include/sys/types.h b/libpsn00b/include/sys/types.h
index da43590..9075b5e 100644
--- a/libpsn00b/include/sys/types.h
+++ b/libpsn00b/include/sys/types.h
@@ -1,13 +1,13 @@
-#ifndef _TYPES_H
-#define _TYPES_H
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2023 Lameguy64, spicyjpeg - MPL licensed
+ */
 
-//#warning "<sys/types.h> and u_* types are deprecated, include <stdint.h> instead"
+#pragma once
 
-//#include <stdint.h>
+//#warning "<sys/types.h> and u_* types are deprecated, use <stdint.h> instead"
 
 typedef unsigned char		u_char;
 typedef unsigned short		u_short;
 typedef unsigned int		u_int;
 typedef unsigned long		u_long;
-
-#endif // _TYPES_H
\ No newline at end of file