gpu: start doing some basic gpu timing

[pcsx_rearmed.git] / deps / lightrec / tlsf / tlsf.c

#include <assert.h>\r
#include <limits.h>\r
#include <stddef.h>\r
#include <stdio.h>\r
#include <stdlib.h>\r
#include <string.h>\r
\r
#include "tlsf.h"\r
\r
#if defined(__cplusplus)\r
#define tlsf_decl inline\r
#else\r
#define tlsf_decl static\r
#endif\r
\r
/*\r
** Architecture-specific bit manipulation routines.\r
**\r
** TLSF achieves O(1) cost for malloc and free operations by limiting\r
** the search for a free block to a free list of guaranteed size\r
** adequate to fulfill the request, combined with efficient free list\r
** queries using bitmasks and architecture-specific bit-manipulation\r
** routines.\r
**\r
** Most modern processors provide instructions to count leading zeroes\r
** in a word, find the lowest and highest set bit, etc. These\r
** specific implementations will be used when available, falling back\r
** to a reasonably efficient generic implementation.\r
**\r
** NOTE: TLSF spec relies on ffs/fls returning value 0..31.\r
** ffs/fls return 1-32 by default, returning 0 for error.\r
*/\r
\r
/*\r
** Detect whether or not we are building for a 32- or 64-bit (LP/LLP)\r
** architecture. There is no reliable portable method at compile-time.\r
*/\r
#if defined (__alpha__) || defined (__ia64__) || defined (__x86_64__) \\r
	|| defined (_WIN64) || defined (__LP64__) || defined (__LLP64__)\r
#define TLSF_64BIT\r
#endif\r
\r
/*\r
** gcc 3.4 and above have builtin support, specialized for architecture.\r
** Some compilers masquerade as gcc; patchlevel test filters them out.\r
*/\r
#if defined (__GNUC__) && (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4)) \\r
	&& defined (__GNUC_PATCHLEVEL__)\r
\r
#if defined (__SNC__)\r
/* SNC for Playstation 3. */\r
\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	const unsigned int reverse = word & (~word + 1);\r
	const int bit = 32 - __builtin_clz(reverse);\r
	return bit - 1;\r
}\r
\r
#else\r
\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	return __builtin_ffs(word) - 1;\r
}\r
\r
#endif\r
\r
tlsf_decl int tlsf_fls(unsigned int word)\r
{\r
	const int bit = word ? 32 - __builtin_clz(word) : 0;\r
	return bit - 1;\r
}\r
\r
#elif defined (_MSC_VER) && (_MSC_VER >= 1400) && (defined (_M_IX86) || defined (_M_X64))\r
/* Microsoft Visual C++ support on x86/X64 architectures. */\r
\r
#include <intrin.h>\r
\r
#pragma intrinsic(_BitScanReverse)\r
#pragma intrinsic(_BitScanForward)\r
\r
tlsf_decl int tlsf_fls(unsigned int word)\r
{\r
	unsigned long index;\r
	return _BitScanReverse(&index, word) ? index : -1;\r
}\r
\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	unsigned long index;\r
	return _BitScanForward(&index, word) ? index : -1;\r
}\r
\r
#elif defined (_MSC_VER) && defined (_M_PPC)\r
/* Microsoft Visual C++ support on PowerPC architectures. */\r
\r
#include <ppcintrinsics.h>\r
\r
tlsf_decl int tlsf_fls(unsigned int word)\r
{\r
	const int bit = 32 - _CountLeadingZeros(word);\r
	return bit - 1;\r
}\r
\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	const unsigned int reverse = word & (~word + 1);\r
	const int bit = 32 - _CountLeadingZeros(reverse);\r
	return bit - 1;\r
}\r
\r
#elif defined (__ARMCC_VERSION)\r
/* RealView Compilation Tools for ARM */\r
\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	const unsigned int reverse = word & (~word + 1);\r
	const int bit = 32 - __clz(reverse);\r
	return bit - 1;\r
}\r
\r
tlsf_decl int tlsf_fls(unsigned int word)\r
{\r
	const int bit = word ? 32 - __clz(word) : 0;\r
	return bit - 1;\r
}\r
\r
#elif defined (__ghs__)\r
/* Green Hills support for PowerPC */\r
\r
#include <ppc_ghs.h>\r
\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	const unsigned int reverse = word & (~word + 1);\r
	const int bit = 32 - __CLZ32(reverse);\r
	return bit - 1;\r
}\r
\r
tlsf_decl int tlsf_fls(unsigned int word)\r
{\r
	const int bit = word ? 32 - __CLZ32(word) : 0;\r
	return bit - 1;\r
}\r
\r
#else\r
/* Fall back to generic implementation. */\r
\r
tlsf_decl int tlsf_fls_generic(unsigned int word)\r
{\r
	int bit = 32;\r
\r
	if (!word) bit -= 1;\r
	if (!(word & 0xffff0000)) { word <<= 16; bit -= 16; }\r
	if (!(word & 0xff000000)) { word <<= 8; bit -= 8; }\r
	if (!(word & 0xf0000000)) { word <<= 4; bit -= 4; }\r
	if (!(word & 0xc0000000)) { word <<= 2; bit -= 2; }\r
	if (!(word & 0x80000000)) { word <<= 1; bit -= 1; }\r
\r
	return bit;\r
}\r
\r
/* Implement ffs in terms of fls. */\r
tlsf_decl int tlsf_ffs(unsigned int word)\r
{\r
	return tlsf_fls_generic(word & (~word + 1)) - 1;\r
}\r
\r
tlsf_decl int tlsf_fls(unsigned int word)\r
{\r
	return tlsf_fls_generic(word) - 1;\r
}\r
\r
#endif\r
\r
/* Possibly 64-bit version of tlsf_fls. */\r
#if defined (TLSF_64BIT)\r
tlsf_decl int tlsf_fls_sizet(size_t size)\r
{\r
	int high = (int)(size >> 32);\r
	int bits = 0;\r
	if (high)\r
	{\r
		bits = 32 + tlsf_fls(high);\r
	}\r
	else\r
	{\r
		bits = tlsf_fls((int)size & 0xffffffff);\r
\r
	}\r
	return bits;\r
}\r
#else\r
#define tlsf_fls_sizet tlsf_fls\r
#endif\r
\r
#undef tlsf_decl\r
\r
/*\r
** Constants.\r
*/\r
\r
/* Public constants: may be modified. */\r
enum tlsf_public\r
{\r
	/* log2 of number of linear subdivisions of block sizes. Larger\r
	** values require more memory in the control structure. Values of\r
	** 4 or 5 are typical.\r
	*/\r
	SL_INDEX_COUNT_LOG2 = 5,\r
};\r
\r
/* Private constants: do not modify. */\r
enum tlsf_private\r
{\r
#if defined (TLSF_64BIT)\r
	/* All allocation sizes and addresses are aligned to 8 bytes. */\r
	ALIGN_SIZE_LOG2 = 3,\r
#else\r
	/* All allocation sizes and addresses are aligned to 4 bytes. */\r
	ALIGN_SIZE_LOG2 = 2,\r
#endif\r
	ALIGN_SIZE = (1 << ALIGN_SIZE_LOG2),\r
\r
	/*\r
	** We support allocations of sizes up to (1 << FL_INDEX_MAX) bits.\r
	** However, because we linearly subdivide the second-level lists, and\r
	** our minimum size granularity is 4 bytes, it doesn't make sense to\r
	** create first-level lists for sizes smaller than SL_INDEX_COUNT * 4,\r
	** or (1 << (SL_INDEX_COUNT_LOG2 + 2)) bytes, as there we will be\r
	** trying to split size ranges into more slots than we have available.\r
	** Instead, we calculate the minimum threshold size, and place all\r
	** blocks below that size into the 0th first-level list.\r
	*/\r
\r
#if defined (TLSF_64BIT)\r
	/*\r
	** TODO: We can increase this to support larger sizes, at the expense\r
	** of more overhead in the TLSF structure.\r
	*/\r
	FL_INDEX_MAX = 32,\r
#else\r
	FL_INDEX_MAX = 30,\r
#endif\r
	SL_INDEX_COUNT = (1 << SL_INDEX_COUNT_LOG2),\r
	FL_INDEX_SHIFT = (SL_INDEX_COUNT_LOG2 + ALIGN_SIZE_LOG2),\r
	FL_INDEX_COUNT = (FL_INDEX_MAX - FL_INDEX_SHIFT + 1),\r
\r
	SMALL_BLOCK_SIZE = (1 << FL_INDEX_SHIFT),\r
};\r
\r
/*\r
** Cast and min/max macros.\r
*/\r
\r
#define tlsf_cast(t, exp)	((t) (exp))\r
#define tlsf_min(a, b)		((a) < (b) ? (a) : (b))\r
#define tlsf_max(a, b)		((a) > (b) ? (a) : (b))\r
\r
/*\r
** Set assert macro, if it has not been provided by the user.\r
*/\r
#if !defined (tlsf_assert)\r
#define tlsf_assert assert\r
#endif\r
\r
/*\r
** Static assertion mechanism.\r
*/\r
\r
#define _tlsf_glue2(x, y) x ## y\r
#define _tlsf_glue(x, y) _tlsf_glue2(x, y)\r
#define tlsf_static_assert(exp) \\r
	typedef char _tlsf_glue(static_assert, __LINE__) [(exp) ? 1 : -1]\r
\r
/* This code has been tested on 32- and 64-bit (LP/LLP) architectures. */\r
tlsf_static_assert(sizeof(int) * CHAR_BIT == 32);\r
tlsf_static_assert(sizeof(size_t) * CHAR_BIT >= 32);\r
tlsf_static_assert(sizeof(size_t) * CHAR_BIT <= 64);\r
\r
/* SL_INDEX_COUNT must be <= number of bits in sl_bitmap's storage type. */\r
tlsf_static_assert(sizeof(unsigned int) * CHAR_BIT >= SL_INDEX_COUNT);\r
\r
/* Ensure we've properly tuned our sizes. */\r
tlsf_static_assert(ALIGN_SIZE == SMALL_BLOCK_SIZE / SL_INDEX_COUNT);\r
\r
/*\r
** Data structures and associated constants.\r
*/\r
\r
/*\r
** Block header structure.\r
**\r
** There are several implementation subtleties involved:\r
** - The prev_phys_block field is only valid if the previous block is free.\r
** - The prev_phys_block field is actually stored at the end of the\r
**   previous block. It appears at the beginning of this structure only to\r
**   simplify the implementation.\r
** - The next_free / prev_free fields are only valid if the block is free.\r
*/\r
typedef struct block_header_t\r
{\r
	/* Points to the previous physical block. */\r
	struct block_header_t* prev_phys_block;\r
\r
	/* The size of this block, excluding the block header. */\r
	size_t size;\r
\r
	/* Next and previous free blocks. */\r
	struct block_header_t* next_free;\r
	struct block_header_t* prev_free;\r
} block_header_t;\r
\r
/*\r
** Since block sizes are always at least a multiple of 4, the two least\r
** significant bits of the size field are used to store the block status:\r
** - bit 0: whether block is busy or free\r
** - bit 1: whether previous block is busy or free\r
*/\r
static const size_t block_header_free_bit = 1 << 0;\r
static const size_t block_header_prev_free_bit = 1 << 1;\r
\r
/*\r
** The size of the block header exposed to used blocks is the size field.\r
** The prev_phys_block field is stored *inside* the previous free block.\r
*/\r
static const size_t block_header_overhead = sizeof(size_t);\r
\r
/* User data starts directly after the size field in a used block. */\r
static const size_t block_start_offset =\r
	offsetof(block_header_t, size) + sizeof(size_t);\r
\r
/*\r
** A free block must be large enough to store its header minus the size of\r
** the prev_phys_block field, and no larger than the number of addressable\r
** bits for FL_INDEX.\r
*/\r
static const size_t block_size_min = \r
	sizeof(block_header_t) - sizeof(block_header_t*);\r
static const size_t block_size_max = tlsf_cast(size_t, 1) << FL_INDEX_MAX;\r
\r
\r
/* The TLSF control structure. */\r
typedef struct control_t\r
{\r
	/* Empty lists point at this block to indicate they are free. */\r
	block_header_t block_null;\r
\r
	/* Bitmaps for free lists. */\r
	unsigned int fl_bitmap;\r
	unsigned int sl_bitmap[FL_INDEX_COUNT];\r
\r
	/* Head of free lists. */\r
	block_header_t* blocks[FL_INDEX_COUNT][SL_INDEX_COUNT];\r
} control_t;\r
\r
/* A type used for casting when doing pointer arithmetic. */\r
typedef ptrdiff_t tlsfptr_t;\r
\r
/*\r
** block_header_t member functions.\r
*/\r
\r
static size_t block_size(const block_header_t* block)\r
{\r
	return block->size & ~(block_header_free_bit | block_header_prev_free_bit);\r
}\r
\r
static void block_set_size(block_header_t* block, size_t size)\r
{\r
	const size_t oldsize = block->size;\r
	block->size = size | (oldsize & (block_header_free_bit | block_header_prev_free_bit));\r
}\r
\r
static int block_is_last(const block_header_t* block)\r
{\r
	return block_size(block) == 0;\r
}\r
\r
static int block_is_free(const block_header_t* block)\r
{\r
	return tlsf_cast(int, block->size & block_header_free_bit);\r
}\r
\r
static void block_set_free(block_header_t* block)\r
{\r
	block->size |= block_header_free_bit;\r
}\r
\r
static void block_set_used(block_header_t* block)\r
{\r
	block->size &= ~block_header_free_bit;\r
}\r
\r
static int block_is_prev_free(const block_header_t* block)\r
{\r
	return tlsf_cast(int, block->size & block_header_prev_free_bit);\r
}\r
\r
static void block_set_prev_free(block_header_t* block)\r
{\r
	block->size |= block_header_prev_free_bit;\r
}\r
\r
static void block_set_prev_used(block_header_t* block)\r
{\r
	block->size &= ~block_header_prev_free_bit;\r
}\r
\r
static block_header_t* block_from_ptr(const void* ptr)\r
{\r
	return tlsf_cast(block_header_t*,\r
		tlsf_cast(unsigned char*, ptr) - block_start_offset);\r
}\r
\r
static void* block_to_ptr(const block_header_t* block)\r
{\r
	return tlsf_cast(void*,\r
		tlsf_cast(unsigned char*, block) + block_start_offset);\r
}\r
\r
/* Return location of next block after block of given size. */\r
static block_header_t* offset_to_block(const void* ptr, size_t size)\r
{\r
	return tlsf_cast(block_header_t*, tlsf_cast(tlsfptr_t, ptr) + size);\r
}\r
\r
/* Return location of previous block. */\r
static block_header_t* block_prev(const block_header_t* block)\r
{\r
	tlsf_assert(block_is_prev_free(block) && "previous block must be free");\r
	return block->prev_phys_block;\r
}\r
\r
/* Return location of next existing block. */\r
static block_header_t* block_next(const block_header_t* block)\r
{\r
	block_header_t* next = offset_to_block(block_to_ptr(block),\r
		block_size(block) - block_header_overhead);\r
	tlsf_assert(!block_is_last(block));\r
	return next;\r
}\r
\r
/* Link a new block with its physical neighbor, return the neighbor. */\r
static block_header_t* block_link_next(block_header_t* block)\r
{\r
	block_header_t* next = block_next(block);\r
	next->prev_phys_block = block;\r
	return next;\r
}\r
\r
static void block_mark_as_free(block_header_t* block)\r
{\r
	/* Link the block to the next block, first. */\r
	block_header_t* next = block_link_next(block);\r
	block_set_prev_free(next);\r
	block_set_free(block);\r
}\r
\r
static void block_mark_as_used(block_header_t* block)\r
{\r
	block_header_t* next = block_next(block);\r
	block_set_prev_used(next);\r
	block_set_used(block);\r
}\r
\r
static size_t align_up(size_t x, size_t align)\r
{\r
	tlsf_assert(0 == (align & (align - 1)) && "must align to a power of two");\r
	return (x + (align - 1)) & ~(align - 1);\r
}\r
\r
static size_t align_down(size_t x, size_t align)\r
{\r
	tlsf_assert(0 == (align & (align - 1)) && "must align to a power of two");\r
	return x - (x & (align - 1));\r
}\r
\r
static void* align_ptr(const void* ptr, size_t align)\r
{\r
	const tlsfptr_t aligned =\r
		(tlsf_cast(tlsfptr_t, ptr) + (align - 1)) & ~(align - 1);\r
	tlsf_assert(0 == (align & (align - 1)) && "must align to a power of two");\r
	return tlsf_cast(void*, aligned);\r
}\r
\r
/*\r
** Adjust an allocation size to be aligned to word size, and no smaller\r
** than internal minimum.\r
*/\r
static size_t adjust_request_size(size_t size, size_t align)\r
{\r
	size_t adjust = 0;\r
	if (size)\r
	{\r
		const size_t aligned = align_up(size, align);\r
\r
		/* aligned sized must not exceed block_size_max or we'll go out of bounds on sl_bitmap */\r
		if (aligned < block_size_max) \r
		{\r
			adjust = tlsf_max(aligned, block_size_min);\r
		}\r
	}\r
	return adjust;\r
}\r
\r
/*\r
** TLSF utility functions. In most cases, these are direct translations of\r
** the documentation found in the white paper.\r
*/\r
\r
static void mapping_insert(size_t size, int* fli, int* sli)\r
{\r
	int fl, sl;\r
	if (size < SMALL_BLOCK_SIZE)\r
	{\r
		/* Store small blocks in first list. */\r
		fl = 0;\r
		sl = tlsf_cast(int, size) / (SMALL_BLOCK_SIZE / SL_INDEX_COUNT);\r
	}\r
	else\r
	{\r
		fl = tlsf_fls_sizet(size);\r
		sl = tlsf_cast(int, size >> (fl - SL_INDEX_COUNT_LOG2)) ^ (1 << SL_INDEX_COUNT_LOG2);\r
		fl -= (FL_INDEX_SHIFT - 1);\r
	}\r
	*fli = fl;\r
	*sli = sl;\r
}\r
\r
/* This version rounds up to the next block size (for allocations) */\r
static void mapping_search(size_t size, int* fli, int* sli)\r
{\r
	if (size >= SMALL_BLOCK_SIZE)\r
	{\r
		const size_t round = (1 << (tlsf_fls_sizet(size) - SL_INDEX_COUNT_LOG2)) - 1;\r
		size += round;\r
	}\r
	mapping_insert(size, fli, sli);\r
}\r
\r
static block_header_t* search_suitable_block(control_t* control, int* fli, int* sli)\r
{\r
	int fl = *fli;\r
	int sl = *sli;\r
\r
	/*\r
	** First, search for a block in the list associated with the given\r
	** fl/sl index.\r
	*/\r
	unsigned int sl_map = control->sl_bitmap[fl] & (~0U << sl);\r
	if (!sl_map)\r
	{\r
		/* No block exists. Search in the next largest first-level list. */\r
		const unsigned int fl_map = control->fl_bitmap & (~0U << (fl + 1));\r
		if (!fl_map)\r
		{\r
			/* No free blocks available, memory has been exhausted. */\r
			return 0;\r
		}\r
\r
		fl = tlsf_ffs(fl_map);\r
		*fli = fl;\r
		sl_map = control->sl_bitmap[fl];\r
	}\r
	tlsf_assert(sl_map && "internal error - second level bitmap is null");\r
	sl = tlsf_ffs(sl_map);\r
	*sli = sl;\r
\r
	/* Return the first block in the free list. */\r
	return control->blocks[fl][sl];\r
}\r
\r
/* Remove a free block from the free list.*/\r
static void remove_free_block(control_t* control, block_header_t* block, int fl, int sl)\r
{\r
	block_header_t* prev = block->prev_free;\r
	block_header_t* next = block->next_free;\r
	tlsf_assert(prev && "prev_free field can not be null");\r
	tlsf_assert(next && "next_free field can not be null");\r
	next->prev_free = prev;\r
	prev->next_free = next;\r
\r
	/* If this block is the head of the free list, set new head. */\r
	if (control->blocks[fl][sl] == block)\r
	{\r
		control->blocks[fl][sl] = next;\r
\r
		/* If the new head is null, clear the bitmap. */\r
		if (next == &control->block_null)\r
		{\r
			control->sl_bitmap[fl] &= ~(1U << sl);\r
\r
			/* If the second bitmap is now empty, clear the fl bitmap. */\r
			if (!control->sl_bitmap[fl])\r
			{\r
				control->fl_bitmap &= ~(1U << fl);\r
			}\r
		}\r
	}\r
}\r
\r
/* Insert a free block into the free block list. */\r
static void insert_free_block(control_t* control, block_header_t* block, int fl, int sl)\r
{\r
	block_header_t* current = control->blocks[fl][sl];\r
	tlsf_assert(current && "free list cannot have a null entry");\r
	tlsf_assert(block && "cannot insert a null entry into the free list");\r
	block->next_free = current;\r
	block->prev_free = &control->block_null;\r
	current->prev_free = block;\r
\r
	tlsf_assert(block_to_ptr(block) == align_ptr(block_to_ptr(block), ALIGN_SIZE)\r
		&& "block not aligned properly");\r
	/*\r
	** Insert the new block at the head of the list, and mark the first-\r
	** and second-level bitmaps appropriately.\r
	*/\r
	control->blocks[fl][sl] = block;\r
	control->fl_bitmap |= (1U << fl);\r
	control->sl_bitmap[fl] |= (1U << sl);\r
}\r
\r
/* Remove a given block from the free list. */\r
static void block_remove(control_t* control, block_header_t* block)\r
{\r
	int fl, sl;\r
	mapping_insert(block_size(block), &fl, &sl);\r
	remove_free_block(control, block, fl, sl);\r
}\r
\r
/* Insert a given block into the free list. */\r
static void block_insert(control_t* control, block_header_t* block)\r
{\r
	int fl, sl;\r
	mapping_insert(block_size(block), &fl, &sl);\r
	insert_free_block(control, block, fl, sl);\r
}\r
\r
static int block_can_split(block_header_t* block, size_t size)\r
{\r
	return block_size(block) >= sizeof(block_header_t) + size;\r
}\r
\r
/* Split a block into two, the second of which is free. */\r
static block_header_t* block_split(block_header_t* block, size_t size)\r
{\r
	/* Calculate the amount of space left in the remaining block. */\r
	block_header_t* remaining =\r
		offset_to_block(block_to_ptr(block), size - block_header_overhead);\r
\r
	const size_t remain_size = block_size(block) - (size + block_header_overhead);\r
\r
	tlsf_assert(block_to_ptr(remaining) == align_ptr(block_to_ptr(remaining), ALIGN_SIZE)\r
		&& "remaining block not aligned properly");\r
\r
	tlsf_assert(block_size(block) == remain_size + size + block_header_overhead);\r
	block_set_size(remaining, remain_size);\r
	tlsf_assert(block_size(remaining) >= block_size_min && "block split with invalid size");\r
\r
	block_set_size(block, size);\r
	block_mark_as_free(remaining);\r
\r
	return remaining;\r
}\r
\r
/* Absorb a free block's storage into an adjacent previous free block. */\r
static block_header_t* block_absorb(block_header_t* prev, block_header_t* block)\r
{\r
	tlsf_assert(!block_is_last(prev) && "previous block can't be last");\r
	/* Note: Leaves flags untouched. */\r
	prev->size += block_size(block) + block_header_overhead;\r
	block_link_next(prev);\r
	return prev;\r
}\r
\r
/* Merge a just-freed block with an adjacent previous free block. */\r
static block_header_t* block_merge_prev(control_t* control, block_header_t* block)\r
{\r
	if (block_is_prev_free(block))\r
	{\r
		block_header_t* prev = block_prev(block);\r
		tlsf_assert(prev && "prev physical block can't be null");\r
		tlsf_assert(block_is_free(prev) && "prev block is not free though marked as such");\r
		block_remove(control, prev);\r
		block = block_absorb(prev, block);\r
	}\r
\r
	return block;\r
}\r
\r
/* Merge a just-freed block with an adjacent free block. */\r
static block_header_t* block_merge_next(control_t* control, block_header_t* block)\r
{\r
	block_header_t* next = block_next(block);\r
	tlsf_assert(next && "next physical block can't be null");\r
\r
	if (block_is_free(next))\r
	{\r
		tlsf_assert(!block_is_last(block) && "previous block can't be last");\r
		block_remove(control, next);\r
		block = block_absorb(block, next);\r
	}\r
\r
	return block;\r
}\r
\r
/* Trim any trailing block space off the end of a block, return to pool. */\r
static void block_trim_free(control_t* control, block_header_t* block, size_t size)\r
{\r
	tlsf_assert(block_is_free(block) && "block must be free");\r
	if (block_can_split(block, size))\r
	{\r
		block_header_t* remaining_block = block_split(block, size);\r
		block_link_next(block);\r
		block_set_prev_free(remaining_block);\r
		block_insert(control, remaining_block);\r
	}\r
}\r
\r
/* Trim any trailing block space off the end of a used block, return to pool. */\r
static void block_trim_used(control_t* control, block_header_t* block, size_t size)\r
{\r
	tlsf_assert(!block_is_free(block) && "block must be used");\r
	if (block_can_split(block, size))\r
	{\r
		/* If the next block is free, we must coalesce. */\r
		block_header_t* remaining_block = block_split(block, size);\r
		block_set_prev_used(remaining_block);\r
\r
		remaining_block = block_merge_next(control, remaining_block);\r
		block_insert(control, remaining_block);\r
	}\r
}\r
\r
static block_header_t* block_trim_free_leading(control_t* control, block_header_t* block, size_t size)\r
{\r
	block_header_t* remaining_block = block;\r
	if (block_can_split(block, size))\r
	{\r
		/* We want the 2nd block. */\r
		remaining_block = block_split(block, size - block_header_overhead);\r
		block_set_prev_free(remaining_block);\r
\r
		block_link_next(block);\r
		block_insert(control, block);\r
	}\r
\r
	return remaining_block;\r
}\r
\r
static block_header_t* block_locate_free(control_t* control, size_t size)\r
{\r
	int fl = 0, sl = 0;\r
	block_header_t* block = 0;\r
\r
	if (size)\r
	{\r
		mapping_search(size, &fl, &sl);\r
		\r
		/*\r
		** mapping_search can futz with the size, so for excessively large sizes it can sometimes wind up \r
		** with indices that are off the end of the block array.\r
		** So, we protect against that here, since this is the only callsite of mapping_search.\r
		** Note that we don't need to check sl, since it comes from a modulo operation that guarantees it's always in range.\r
		*/\r
		if (fl < FL_INDEX_COUNT)\r
		{\r
			block = search_suitable_block(control, &fl, &sl);\r
		}\r
	}\r
\r
	if (block)\r
	{\r
		tlsf_assert(block_size(block) >= size);\r
		remove_free_block(control, block, fl, sl);\r
	}\r
\r
	return block;\r
}\r
\r
static void* block_prepare_used(control_t* control, block_header_t* block, size_t size)\r
{\r
	void* p = 0;\r
	if (block)\r
	{\r
		tlsf_assert(size && "size must be non-zero");\r
		block_trim_free(control, block, size);\r
		block_mark_as_used(block);\r
		p = block_to_ptr(block);\r
	}\r
	return p;\r
}\r
\r
/* Clear structure and point all empty lists at the null block. */\r
static void control_construct(control_t* control)\r
{\r
	int i, j;\r
\r
	control->block_null.next_free = &control->block_null;\r
	control->block_null.prev_free = &control->block_null;\r
\r
	control->fl_bitmap = 0;\r
	for (i = 0; i < FL_INDEX_COUNT; ++i)\r
	{\r
		control->sl_bitmap[i] = 0;\r
		for (j = 0; j < SL_INDEX_COUNT; ++j)\r
		{\r
			control->blocks[i][j] = &control->block_null;\r
		}\r
	}\r
}\r
\r
/*\r
** Debugging utilities.\r
*/\r
\r
typedef struct integrity_t\r
{\r
	int prev_status;\r
	int status;\r
} integrity_t;\r
\r
#define tlsf_insist(x) { tlsf_assert(x); if (!(x)) { status--; } }\r
\r
static void integrity_walker(void* ptr, size_t size, int used, void* user)\r
{\r
	block_header_t* block = block_from_ptr(ptr);\r
	integrity_t* integ = tlsf_cast(integrity_t*, user);\r
	const int this_prev_status = block_is_prev_free(block) ? 1 : 0;\r
	const int this_status = block_is_free(block) ? 1 : 0;\r
	const size_t this_block_size = block_size(block);\r
\r
	int status = 0;\r
	(void)used;\r
	tlsf_insist(integ->prev_status == this_prev_status && "prev status incorrect");\r
	tlsf_insist(size == this_block_size && "block size incorrect");\r
\r
	integ->prev_status = this_status;\r
	integ->status += status;\r
}\r
\r
int tlsf_check(tlsf_t tlsf)\r
{\r
	int i, j;\r
\r
	control_t* control = tlsf_cast(control_t*, tlsf);\r
	int status = 0;\r
\r
	/* Check that the free lists and bitmaps are accurate. */\r
	for (i = 0; i < FL_INDEX_COUNT; ++i)\r
	{\r
		for (j = 0; j < SL_INDEX_COUNT; ++j)\r
		{\r
			const int fl_map = control->fl_bitmap & (1U << i);\r
			const int sl_list = control->sl_bitmap[i];\r
			const int sl_map = sl_list & (1U << j);\r
			const block_header_t* block = control->blocks[i][j];\r
\r
			/* Check that first- and second-level lists agree. */\r
			if (!fl_map)\r
			{\r
				tlsf_insist(!sl_map && "second-level map must be null");\r
			}\r
\r
			if (!sl_map)\r
			{\r
				tlsf_insist(block == &control->block_null && "block list must be null");\r
				continue;\r
			}\r
\r
			/* Check that there is at least one free block. */\r
			tlsf_insist(sl_list && "no free blocks in second-level map");\r
			tlsf_insist(block != &control->block_null && "block should not be null");\r
\r
			while (block != &control->block_null)\r
			{\r
				int fli, sli;\r
				tlsf_insist(block_is_free(block) && "block should be free");\r
				tlsf_insist(!block_is_prev_free(block) && "blocks should have coalesced");\r
				tlsf_insist(!block_is_free(block_next(block)) && "blocks should have coalesced");\r
				tlsf_insist(block_is_prev_free(block_next(block)) && "block should be free");\r
				tlsf_insist(block_size(block) >= block_size_min && "block not minimum size");\r
\r
				mapping_insert(block_size(block), &fli, &sli);\r
				tlsf_insist(fli == i && sli == j && "block size indexed in wrong list");\r
				block = block->next_free;\r
			}\r
		}\r
	}\r
\r
	return status;\r
}\r
\r
#undef tlsf_insist\r
\r
static void default_walker(void* ptr, size_t size, int used, void* user)\r
{\r
	(void)user;\r
	printf("\t%p %s size: %x (%p)\n", ptr, used ? "used" : "free", (unsigned int)size, block_from_ptr(ptr));\r
}\r
\r
void tlsf_walk_pool(pool_t pool, tlsf_walker walker, void* user)\r
{\r
	tlsf_walker pool_walker = walker ? walker : default_walker;\r
	block_header_t* block =\r
		offset_to_block(pool, -(int)block_header_overhead);\r
\r
	while (block && !block_is_last(block))\r
	{\r
		pool_walker(\r
			block_to_ptr(block),\r
			block_size(block),\r
			!block_is_free(block),\r
			user);\r
		block = block_next(block);\r
	}\r
}\r
\r
size_t tlsf_block_size(void* ptr)\r
{\r
	size_t size = 0;\r
	if (ptr)\r
	{\r
		const block_header_t* block = block_from_ptr(ptr);\r
		size = block_size(block);\r
	}\r
	return size;\r
}\r
\r
int tlsf_check_pool(pool_t pool)\r
{\r
	/* Check that the blocks are physically correct. */\r
	integrity_t integ = { 0, 0 };\r
	tlsf_walk_pool(pool, integrity_walker, &integ);\r
\r
	return integ.status;\r
}\r
\r
/*\r
** Size of the TLSF structures in a given memory block passed to\r
** tlsf_create, equal to the size of a control_t\r
*/\r
size_t tlsf_size(void)\r
{\r
	return sizeof(control_t);\r
}\r
\r
size_t tlsf_align_size(void)\r
{\r
	return ALIGN_SIZE;\r
}\r
\r
size_t tlsf_block_size_min(void)\r
{\r
	return block_size_min;\r
}\r
\r
size_t tlsf_block_size_max(void)\r
{\r
	return block_size_max;\r
}\r
\r
/*\r
** Overhead of the TLSF structures in a given memory block passed to\r
** tlsf_add_pool, equal to the overhead of a free block and the\r
** sentinel block.\r
*/\r
size_t tlsf_pool_overhead(void)\r
{\r
	return 2 * block_header_overhead;\r
}\r
\r
size_t tlsf_alloc_overhead(void)\r
{\r
	return block_header_overhead;\r
}\r
\r
pool_t tlsf_add_pool(tlsf_t tlsf, void* mem, size_t bytes)\r
{\r
	block_header_t* block;\r
	block_header_t* next;\r
\r
	const size_t pool_overhead = tlsf_pool_overhead();\r
	const size_t pool_bytes = align_down(bytes - pool_overhead, ALIGN_SIZE);\r
\r
	if (((ptrdiff_t)mem % ALIGN_SIZE) != 0)\r
	{\r
		printf("tlsf_add_pool: Memory must be aligned by %u bytes.\n",\r
			(unsigned int)ALIGN_SIZE);\r
		return 0;\r
	}\r
\r
	if (pool_bytes < block_size_min || pool_bytes > block_size_max)\r
	{\r
#if defined (TLSF_64BIT)\r
		printf("tlsf_add_pool: Memory size must be between 0x%x and 0x%x00 bytes.\n", \r
			(unsigned int)(pool_overhead + block_size_min),\r
			(unsigned int)((pool_overhead + block_size_max) / 256));\r
#else\r
		printf("tlsf_add_pool: Memory size must be between %u and %u bytes.\n", \r
			(unsigned int)(pool_overhead + block_size_min),\r
			(unsigned int)(pool_overhead + block_size_max));\r
#endif\r
		return 0;\r
	}\r
\r
	/*\r
	** Create the main free block. Offset the start of the block slightly\r
	** so that the prev_phys_block field falls outside of the pool -\r
	** it will never be used.\r
	*/\r
	block = offset_to_block(mem, -(tlsfptr_t)block_header_overhead);\r
	block_set_size(block, pool_bytes);\r
	block_set_free(block);\r
	block_set_prev_used(block);\r
	block_insert(tlsf_cast(control_t*, tlsf), block);\r
\r
	/* Split the block to create a zero-size sentinel block. */\r
	next = block_link_next(block);\r
	block_set_size(next, 0);\r
	block_set_used(next);\r
	block_set_prev_free(next);\r
\r
	return mem;\r
}\r
\r
void tlsf_remove_pool(tlsf_t tlsf, pool_t pool)\r
{\r
	control_t* control = tlsf_cast(control_t*, tlsf);\r
	block_header_t* block = offset_to_block(pool, -(int)block_header_overhead);\r
\r
	int fl = 0, sl = 0;\r
\r
	tlsf_assert(block_is_free(block) && "block should be free");\r
	tlsf_assert(!block_is_free(block_next(block)) && "next block should not be free");\r
	tlsf_assert(block_size(block_next(block)) == 0 && "next block size should be zero");\r
\r
	mapping_insert(block_size(block), &fl, &sl);\r
	remove_free_block(control, block, fl, sl);\r
}\r
\r
/*\r
** TLSF main interface.\r
*/\r
\r
#if _DEBUG\r
int test_ffs_fls()\r
{\r
	/* Verify ffs/fls work properly. */\r
	int rv = 0;\r
	rv += (tlsf_ffs(0) == -1) ? 0 : 0x1;\r
	rv += (tlsf_fls(0) == -1) ? 0 : 0x2;\r
	rv += (tlsf_ffs(1) == 0) ? 0 : 0x4;\r
	rv += (tlsf_fls(1) == 0) ? 0 : 0x8;\r
	rv += (tlsf_ffs(0x80000000) == 31) ? 0 : 0x10;\r
	rv += (tlsf_ffs(0x80008000) == 15) ? 0 : 0x20;\r
	rv += (tlsf_fls(0x80000008) == 31) ? 0 : 0x40;\r
	rv += (tlsf_fls(0x7FFFFFFF) == 30) ? 0 : 0x80;\r
\r
#if defined (TLSF_64BIT)\r
	rv += (tlsf_fls_sizet(0x80000000) == 31) ? 0 : 0x100;\r
	rv += (tlsf_fls_sizet(0x100000000) == 32) ? 0 : 0x200;\r
	rv += (tlsf_fls_sizet(0xffffffffffffffff) == 63) ? 0 : 0x400;\r
#endif\r
\r
	if (rv)\r
	{\r
		printf("test_ffs_fls: %x ffs/fls tests failed.\n", rv);\r
	}\r
	return rv;\r
}\r
#endif\r
\r
tlsf_t tlsf_create(void* mem)\r
{\r
#if _DEBUG\r
	if (test_ffs_fls())\r
	{\r
		return 0;\r
	}\r
#endif\r
\r
	if (((tlsfptr_t)mem % ALIGN_SIZE) != 0)\r
	{\r
		printf("tlsf_create: Memory must be aligned to %u bytes.\n",\r
			(unsigned int)ALIGN_SIZE);\r
		return 0;\r
	}\r
\r
	control_construct(tlsf_cast(control_t*, mem));\r
\r
	return tlsf_cast(tlsf_t, mem);\r
}\r
\r
tlsf_t tlsf_create_with_pool(void* mem, size_t bytes)\r
{\r
	tlsf_t tlsf = tlsf_create(mem);\r
	tlsf_add_pool(tlsf, (char*)mem + tlsf_size(), bytes - tlsf_size());\r
	return tlsf;\r
}\r
\r
void tlsf_destroy(tlsf_t tlsf)\r
{\r
	/* Nothing to do. */\r
	(void)tlsf;\r
}\r
\r
pool_t tlsf_get_pool(tlsf_t tlsf)\r
{\r
	return tlsf_cast(pool_t, (char*)tlsf + tlsf_size());\r
}\r
\r
void* tlsf_malloc(tlsf_t tlsf, size_t size)\r
{\r
	control_t* control = tlsf_cast(control_t*, tlsf);\r
	const size_t adjust = adjust_request_size(size, ALIGN_SIZE);\r
	block_header_t* block = block_locate_free(control, adjust);\r
	return block_prepare_used(control, block, adjust);\r
}\r
\r
void* tlsf_memalign(tlsf_t tlsf, size_t align, size_t size)\r
{\r
	control_t* control = tlsf_cast(control_t*, tlsf);\r
	const size_t adjust = adjust_request_size(size, ALIGN_SIZE);\r
\r
	/*\r
	** We must allocate an additional minimum block size bytes so that if\r
	** our free block will leave an alignment gap which is smaller, we can\r
	** trim a leading free block and release it back to the pool. We must\r
	** do this because the previous physical block is in use, therefore\r
	** the prev_phys_block field is not valid, and we can't simply adjust\r
	** the size of that block.\r
	*/\r
	const size_t gap_minimum = sizeof(block_header_t);\r
	const size_t size_with_gap = adjust_request_size(adjust + align + gap_minimum, align);\r
\r
	/*\r
	** If alignment is less than or equals base alignment, we're done.\r
	** If we requested 0 bytes, return null, as tlsf_malloc(0) does.\r
	*/\r
	const size_t aligned_size = (adjust && align > ALIGN_SIZE) ? size_with_gap : adjust;\r
\r
	block_header_t* block = block_locate_free(control, aligned_size);\r
\r
	/* This can't be a static assert. */\r
	tlsf_assert(sizeof(block_header_t) == block_size_min + block_header_overhead);\r
\r
	if (block)\r
	{\r
		void* ptr = block_to_ptr(block);\r
		void* aligned = align_ptr(ptr, align);\r
		size_t gap = tlsf_cast(size_t,\r
			tlsf_cast(tlsfptr_t, aligned) - tlsf_cast(tlsfptr_t, ptr));\r
\r
		/* If gap size is too small, offset to next aligned boundary. */\r
		if (gap && gap < gap_minimum)\r
		{\r
			const size_t gap_remain = gap_minimum - gap;\r
			const size_t offset = tlsf_max(gap_remain, align);\r
			const void* next_aligned = tlsf_cast(void*,\r
				tlsf_cast(tlsfptr_t, aligned) + offset);\r
\r
			aligned = align_ptr(next_aligned, align);\r
			gap = tlsf_cast(size_t,\r
				tlsf_cast(tlsfptr_t, aligned) - tlsf_cast(tlsfptr_t, ptr));\r
		}\r
\r
		if (gap)\r
		{\r
			tlsf_assert(gap >= gap_minimum && "gap size too small");\r
			block = block_trim_free_leading(control, block, gap);\r
		}\r
	}\r
\r
	return block_prepare_used(control, block, adjust);\r
}\r
\r
void tlsf_free(tlsf_t tlsf, void* ptr)\r
{\r
	/* Don't attempt to free a NULL pointer. */\r
	if (ptr)\r
	{\r
		control_t* control = tlsf_cast(control_t*, tlsf);\r
		block_header_t* block = block_from_ptr(ptr);\r
		tlsf_assert(!block_is_free(block) && "block already marked as free");\r
		block_mark_as_free(block);\r
		block = block_merge_prev(control, block);\r
		block = block_merge_next(control, block);\r
		block_insert(control, block);\r
	}\r
}\r
\r
/*\r
** The TLSF block information provides us with enough information to\r
** provide a reasonably intelligent implementation of realloc, growing or\r
** shrinking the currently allocated block as required.\r
**\r
** This routine handles the somewhat esoteric edge cases of realloc:\r
** - a non-zero size with a null pointer will behave like malloc\r
** - a zero size with a non-null pointer will behave like free\r
** - a request that cannot be satisfied will leave the original buffer\r
**   untouched\r
** - an extended buffer size will leave the newly-allocated area with\r
**   contents undefined\r
*/\r
void* tlsf_realloc(tlsf_t tlsf, void* ptr, size_t size)\r
{\r
	control_t* control = tlsf_cast(control_t*, tlsf);\r
	void* p = 0;\r
\r
	/* Zero-size requests are treated as free. */\r
	if (ptr && size == 0)\r
	{\r
		tlsf_free(tlsf, ptr);\r
	}\r
	/* Requests with NULL pointers are treated as malloc. */\r
	else if (!ptr)\r
	{\r
		p = tlsf_malloc(tlsf, size);\r
	}\r
	else\r
	{\r
		block_header_t* block = block_from_ptr(ptr);\r
		block_header_t* next = block_next(block);\r
\r
		const size_t cursize = block_size(block);\r
		const size_t combined = cursize + block_size(next) + block_header_overhead;\r
		const size_t adjust = adjust_request_size(size, ALIGN_SIZE);\r
\r
		tlsf_assert(!block_is_free(block) && "block already marked as free");\r
\r
		/*\r
		** If the next block is used, or when combined with the current\r
		** block, does not offer enough space, we must reallocate and copy.\r
		*/\r
		if (adjust > cursize && (!block_is_free(next) || adjust > combined))\r
		{\r
			p = tlsf_malloc(tlsf, size);\r
			if (p)\r
			{\r
				const size_t minsize = tlsf_min(cursize, size);\r
				memcpy(p, ptr, minsize);\r
				tlsf_free(tlsf, ptr);\r
			}\r
		}\r
		else\r
		{\r
			/* Do we need to expand to the next block? */\r
			if (adjust > cursize)\r
			{\r
				block_merge_next(control, block);\r
				block_mark_as_used(block);\r
			}\r
\r
			/* Trim the resulting block and return the original pointer. */\r
			block_trim_used(control, block, adjust);\r
			p = ptr;\r
		}\r
	}\r
\r
	return p;\r
}\r