Context Navigation

heap.cfa @ b8ab91a

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since b8ab91a was 7dd98b6, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Moved cfa_main_returned to libcfa so it works when the main is written in C.
Property mode set to `100644`
File size: 54.9 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2017 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// heap.cfa --
8	//
9	// Author : Peter A. Buhr
10	// Created On : Tue Dec 19 21:58:35 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Mon Aug 9 19:03:02 2021
13	// Update Count : 1040
14	//
15
16	#include <unistd.h> // sbrk, sysconf
17	#include <stdlib.h> // EXIT_FAILURE
18	#include <stdbool.h> // true, false
19	#include <stdio.h> // snprintf, fileno
20	#include <errno.h> // errno
21	#include <string.h> // memset, memcpy
22	#include <limits.h> // ULONG_MAX
23	#include <malloc.h> // memalign, malloc_usable_size
24	#include <sys/mman.h> // mmap, munmap
25
26	#include "bits/align.hfa" // libAlign
27	#include "bits/defs.hfa" // likely, unlikely
28	#include "bits/locks.hfa" // __spinlock_t
29	#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
30	#include "math.hfa" // min
31	#include "bitmanip.hfa" // is_pow2, ceiling2
32
33	static bool traceHeap = false;
34
35	inline bool traceHeap() { return traceHeap; }
36
37	bool traceHeapOn() {
38	bool temp = traceHeap;
39	traceHeap = true;
40	return temp;
41	} // traceHeapOn
42
43	bool traceHeapOff() {
44	bool temp = traceHeap;
45	traceHeap = false;
46	return temp;
47	} // traceHeapOff
48
49	bool traceHeapTerm() { return false; }
50
51
52	static bool prtFree = false;
53
54	bool prtFree() {
55	return prtFree;
56	} // prtFree
57
58	bool prtFreeOn() {
59	bool temp = prtFree;
60	prtFree = true;
61	return temp;
62	} // prtFreeOn
63
64	bool prtFreeOff() {
65	bool temp = prtFree;
66	prtFree = false;
67	return temp;
68	} // prtFreeOff
69
70
71	enum {
72	// Define the default extension heap amount in units of bytes. When the uC++ supplied heap reaches the brk address,
73	// the brk address is extended by the extension amount.
74	__CFA_DEFAULT_HEAP_EXPANSION__ = (10 * 1024 * 1024),
75
76	// Define the mmap crossover point during allocation. Allocations less than this amount are allocated from buckets;
77	// values greater than or equal to this value are mmap from the operating system.
78	__CFA_DEFAULT_MMAP_START__ = (512 * 1024 + 1),
79	};
80
81	size_t default_mmap_start() __attribute__(( weak )) {
82	return __CFA_DEFAULT_MMAP_START__;
83	} // default_mmap_start
84
85	size_t default_heap_expansion() __attribute__(( weak )) {
86	return __CFA_DEFAULT_HEAP_EXPANSION__;
87	} // default_heap_expansion
88
89
90	#ifdef __CFA_DEBUG__
91	static size_t allocUnfreed; // running total of allocations minus frees
92
93	static void prtUnfreed() {
94	if ( allocUnfreed != 0 ) {
95	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
96	char helpText[512];
97	int len = snprintf( helpText, sizeof(helpText), "CFA warning (UNIX pid:%ld) : program terminating with %zu(0x%zx) bytes of storage allocated but not freed.\n"
98	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
99	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
100	__cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
101	} // if
102	} // prtUnfreed
103
104	extern int cfa_main_returned; // from interpose.cfa
105	extern "C" {
106	void heapAppStart() { // called by __cfaabi_appready_startup
107	allocUnfreed = 0;
108	} // heapAppStart
109
110	void heapAppStop() { // called by __cfaabi_appready_startdown
111	fclose( stdin ); fclose( stdout );
112	if ( cfa_main_returned ) prtUnfreed(); // do not check unfreed storage if exit called
113	} // heapAppStop
114	} // extern "C"
115	#endif // __CFA_DEBUG__
116
117
118	// statically allocated variables => zero filled.
119	size_t __page_size; // architecture pagesize
120	int __map_prot; // common mmap/mprotect protection
121	static size_t heapExpand; // sbrk advance
122	static size_t mmapStart; // cross over point for mmap
123	static unsigned int maxBucketsUsed; // maximum number of buckets in use
124
125
126	#define SPINLOCK 0
127	#define LOCKFREE 1
128	#define BUCKETLOCK SPINLOCK
129	#if BUCKETLOCK == SPINLOCK
130	#elif BUCKETLOCK == LOCKFREE
131	#include <stackLockFree.hfa>
132	#else
133	#error undefined lock type for bucket lock
134	#endif // LOCKFREE
135
136	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
137	// Break recusion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
138	enum { NoBucketSizes = 91 }; // number of buckets sizes
139
140	struct HeapManager {
141	struct Storage {
142	struct Header { // header
143	union Kind {
144	struct RealHeader {
145	union {
146	struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
147	#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
148	uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
149	#endif // __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
150
151	union {
152	// FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
153	// 2nd low-order bit => zero filled
154	void * home; // allocated block points back to home locations (must overlay alignment)
155	size_t blockSize; // size for munmap (must overlay alignment)
156	#if BUCKETLOCK == SPINLOCK
157	Storage * next; // freed block points next freed block of same size
158	#endif // SPINLOCK
159	};
160	size_t size; // allocation size in bytes
161
162	#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
163	uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
164	#endif // __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
165	};
166	#if BUCKETLOCK == LOCKFREE
167	Link(Storage) next; // freed block points next freed block of same size (double-wide)
168	#endif // LOCKFREE
169	};
170	} real; // RealHeader
171
172	struct FakeHeader {
173	#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
174	uint32_t alignment; // 1st low-order bit => fake header & alignment
175	#endif // __ORDER_LITTLE_ENDIAN__
176
177	uint32_t offset;
178
179	#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
180	uint32_t alignment; // low-order bits of home/blockSize used for tricks
181	#endif // __ORDER_BIG_ENDIAN__
182	} fake; // FakeHeader
183	} kind; // Kind
184	} header; // Header
185	char pad[libAlign() - sizeof( Header )];
186	char data[0]; // storage
187	}; // Storage
188
189	static_assert( libAlign() >= sizeof( Storage ), "libAlign() < sizeof( Storage )" );
190
191	struct FreeHeader {
192	#if BUCKETLOCK == SPINLOCK
193	__spinlock_t lock; // must be first field for alignment
194	Storage * freeList;
195	#else
196	StackLF(Storage) freeList;
197	#endif // BUCKETLOCK
198	size_t blockSize; // size of allocations on this list
199	}; // FreeHeader
200
201	// must be first fields for alignment
202	__spinlock_t extlock; // protects allocation-buffer extension
203	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
204
205	void * heapBegin; // start of heap
206	void * heapEnd; // logical end of heap
207	size_t heapRemaining; // amount of storage not allocated in the current chunk
208	}; // HeapManager
209
210	#if BUCKETLOCK == LOCKFREE
211	static inline {
212	Link(HeapManager.Storage) * ?`next( HeapManager.Storage * this ) { return &this->header.kind.real.next; }
213	void ?{}( HeapManager.FreeHeader & ) {}
214	void ^?{}( HeapManager.FreeHeader & ) {}
215	} // distribution
216	#endif // LOCKFREE
217
218	static inline size_t getKey( const HeapManager.FreeHeader & freeheader ) { return freeheader.blockSize; }
219
220
221	#define FASTLOOKUP
222	#define __STATISTICS__
223
224	// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
225	// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
226	// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
227	static const unsigned int bucketSizes[] @= { // different bucket sizes
228	16 + sizeof(HeapManager.Storage), 32 + sizeof(HeapManager.Storage), 48 + sizeof(HeapManager.Storage), 64 + sizeof(HeapManager.Storage), // 4
229	96 + sizeof(HeapManager.Storage), 112 + sizeof(HeapManager.Storage), 128 + sizeof(HeapManager.Storage), // 3
230	160, 192, 224, 256 + sizeof(HeapManager.Storage), // 4
231	320, 384, 448, 512 + sizeof(HeapManager.Storage), // 4
232	640, 768, 896, 1_024 + sizeof(HeapManager.Storage), // 4
233	1_536, 2_048 + sizeof(HeapManager.Storage), // 2
234	2_560, 3_072, 3_584, 4_096 + sizeof(HeapManager.Storage), // 4
235	6_144, 8_192 + sizeof(HeapManager.Storage), // 2
236	9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(HeapManager.Storage), // 8
237	18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(HeapManager.Storage), // 8
238	36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(HeapManager.Storage), // 8
239	73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(HeapManager.Storage), // 8
240	147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(HeapManager.Storage), // 8
241	294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(HeapManager.Storage), // 8
242	655_360, 786_432, 917_504, 1_048_576 + sizeof(HeapManager.Storage), // 4
243	1_179_648, 1_310_720, 1_441_792, 1_572_864, 1_703_936, 1_835_008, 1_966_080, 2_097_152 + sizeof(HeapManager.Storage), // 8
244	2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(HeapManager.Storage), // 4
245	};
246
247	static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
248
249	#ifdef FASTLOOKUP
250	enum { LookupSizes = 65_536 + sizeof(HeapManager.Storage) }; // number of fast lookup sizes
251	static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
252	#endif // FASTLOOKUP
253
254	static const off_t mmapFd = -1; // fake or actual fd for anonymous file
255	#ifdef __CFA_DEBUG__
256	static bool heapBoot = 0; // detect recursion during boot
257	#endif // __CFA_DEBUG__
258
259	// The constructor for heapManager is called explicitly in memory_startup.
260	static HeapManager heapManager __attribute__(( aligned (128) )) @= {}; // size of cache line to prevent false sharing
261
262
263	#ifdef __STATISTICS__
264	// Heap statistics counters.
265	static unsigned int malloc_zero_calls, malloc_calls;
266	static unsigned long long int malloc_storage;
267	static unsigned int aalloc_zero_calls, aalloc_calls;
268	static unsigned long long int aalloc_storage;
269	static unsigned int calloc_zero_calls, calloc_calls;
270	static unsigned long long int calloc_storage;
271	static unsigned int memalign_zero_calls, memalign_calls;
272	static unsigned long long int memalign_storage;
273	static unsigned int amemalign_zero_calls, amemalign_calls;
274	static unsigned long long int amemalign_storage;
275	static unsigned int cmemalign_zero_calls, cmemalign_calls;
276	static unsigned long long int cmemalign_storage;
277	static unsigned int resize_zero_calls, resize_calls;
278	static unsigned long long int resize_storage;
279	static unsigned int realloc_zero_calls, realloc_calls;
280	static unsigned long long int realloc_storage;
281	static unsigned int free_zero_calls, free_calls;
282	static unsigned long long int free_storage;
283	static unsigned int mmap_calls;
284	static unsigned long long int mmap_storage;
285	static unsigned int munmap_calls;
286	static unsigned long long int munmap_storage;
287	static unsigned int sbrk_calls;
288	static unsigned long long int sbrk_storage;
289	// Statistics file descriptor (changed by malloc_stats_fd).
290	static int stats_fd = STDERR_FILENO; // default stderr
291
292	// Use "write" because streams may be shutdown when calls are made.
293	static void printStats() {
294	char helpText[1024];
295	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
296	"\nHeap statistics:\n"
297	" malloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
298	" aalloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
299	" calloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
300	" memalign 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
301	" amemalign 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
302	" cmemalign 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
303	" resize 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
304	" realloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
305	" free 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
306	" mmap calls %'u; storage %'llu bytes\n"
307	" munmap calls %'u; storage %'llu bytes\n"
308	" sbrk calls %'u; storage %'llu bytes\n",
309	malloc_zero_calls, malloc_calls, malloc_storage,
310	aalloc_zero_calls, aalloc_calls, aalloc_storage,
311	calloc_zero_calls, calloc_calls, calloc_storage,
312	memalign_zero_calls, memalign_calls, memalign_storage,
313	amemalign_zero_calls, amemalign_calls, amemalign_storage,
314	cmemalign_zero_calls, cmemalign_calls, cmemalign_storage,
315	resize_zero_calls, resize_calls, resize_storage,
316	realloc_zero_calls, realloc_calls, realloc_storage,
317	free_zero_calls, free_calls, free_storage,
318	mmap_calls, mmap_storage,
319	munmap_calls, munmap_storage,
320	sbrk_calls, sbrk_storage
321	);
322	} // printStats
323
324	static int printStatsXML( FILE * stream ) { // see malloc_info
325	char helpText[1024];
326	int len = snprintf( helpText, sizeof(helpText),
327	"<malloc version=\"1\">\n"
328	"<heap nr=\"0\">\n"
329	"<sizes>\n"
330	"</sizes>\n"
331	"<total type=\"malloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
332	"<total type=\"aalloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
333	"<total type=\"calloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
334	"<total type=\"memalign\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
335	"<total type=\"amemalign\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
336	"<total type=\"cmemalign\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
337	"<total type=\"resize\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
338	"<total type=\"realloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
339	"<total type=\"free\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
340	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
341	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
342	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
343	"</malloc>",
344	malloc_zero_calls, malloc_calls, malloc_storage,
345	aalloc_zero_calls, aalloc_calls, aalloc_storage,
346	calloc_zero_calls, calloc_calls, calloc_storage,
347	memalign_zero_calls, memalign_calls, memalign_storage,
348	amemalign_zero_calls, amemalign_calls, amemalign_storage,
349	cmemalign_zero_calls, cmemalign_calls, cmemalign_storage,
350	resize_zero_calls, resize_calls, resize_storage,
351	realloc_zero_calls, realloc_calls, realloc_storage,
352	free_zero_calls, free_calls, free_storage,
353	mmap_calls, mmap_storage,
354	munmap_calls, munmap_storage,
355	sbrk_calls, sbrk_storage
356	);
357	__cfaabi_bits_write( fileno( stream ), helpText, len ); // ensures all bytes written or exit
358	return len;
359	} // printStatsXML
360	#endif // __STATISTICS__
361
362
363	// thunk problem
364	size_t Bsearchl( unsigned int key, const unsigned int * vals, size_t dim ) {
365	size_t l = 0, m, h = dim;
366	while ( l < h ) {
367	m = (l + h) / 2;
368	if ( (unsigned int &)(vals[m]) < key ) { // cast away const
369	l = m + 1;
370	} else {
371	h = m;
372	} // if
373	} // while
374	return l;
375	} // Bsearchl
376
377
378	static inline bool setMmapStart( size_t value ) { // true => mmapped, false => sbrk
379	if ( value < __page_size \|\| bucketSizes[NoBucketSizes - 1] < value ) return false;
380	mmapStart = value; // set global
381
382	// find the closest bucket size less than or equal to the mmapStart size
383	maxBucketsUsed = Bsearchl( (unsigned int)mmapStart, bucketSizes, NoBucketSizes ); // binary search
384	assert( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
385	assert( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
386	return true;
387	} // setMmapStart
388
389
390	// <-------+----------------------------------------------------> bsize (bucket size)
391	// \|header \|addr
392	//==================================================================================
393	// align/offset \|
394	// <-----------------<------------+-----------------------------> bsize (bucket size)
395	// \|fake-header \| addr
396	#define headerAddr( addr ) ((HeapManager.Storage.Header )( (char )addr - sizeof(HeapManager.Storage) ))
397	#define realHeader( header ) ((HeapManager.Storage.Header )((char )header - header->kind.fake.offset))
398
399	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
400	// \|header \|addr
401	//==================================================================================
402	// align/offset \|
403	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
404	// \|fake-header \|addr
405	#define dataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
406
407
408	static inline void checkAlign( size_t alignment ) {
409	if ( alignment < libAlign() \|\| ! is_pow2( alignment ) ) {
410	abort( "Alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
411	} // if
412	} // checkAlign
413
414
415	static inline void checkHeader( bool check, const char name[], void * addr ) {
416	if ( unlikely( check ) ) { // bad address ?
417	abort( "Attempt to %s storage %p with address outside the heap.\n"
418	"Possible cause is duplicate free on same block or overwriting of memory.",
419	name, addr );
420	} // if
421	} // checkHeader
422
423
424	static inline void fakeHeader( HeapManager.Storage.Header *& header, size_t & alignment ) {
425	if ( unlikely( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
426	alignment = header->kind.fake.alignment & -2; // remove flag from value
427	#ifdef __CFA_DEBUG__
428	checkAlign( alignment ); // check alignment
429	#endif // __CFA_DEBUG__
430	header = realHeader( header ); // backup from fake to real header
431	} else {
432	alignment = libAlign(); // => no fake header
433	} // if
434	} // fakeHeader
435
436
437	static inline bool headers( const char name[] __attribute__(( unused )), void * addr, HeapManager.Storage.Header & header, HeapManager.FreeHeader & freeElem,
438	size_t & size, size_t & alignment ) with( heapManager ) {
439	header = headerAddr( addr );
440
441	if ( unlikely( addr < heapBegin \|\| heapEnd < addr ) ) { // mmapped ?
442	fakeHeader( header, alignment );
443	size = header->kind.real.blockSize & -3; // mmap size
444	return true;
445	} // if
446
447	#ifdef __CFA_DEBUG__
448	checkHeader( header < (HeapManager.Storage.Header *)heapBegin, name, addr ); // bad low address ?
449	#endif // __CFA_DEBUG__
450
451	// header may be safe to dereference
452	fakeHeader( header, alignment );
453	#ifdef __CFA_DEBUG__
454	checkHeader( header < (HeapManager.Storage.Header )heapBegin \|\| (HeapManager.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
455	#endif // __CFA_DEBUG__
456
457	freeElem = (HeapManager.FreeHeader *)((size_t)header->kind.real.home & -3);
458	#ifdef __CFA_DEBUG__
459	if ( freeElem < &freeLists[0] \|\| &freeLists[NoBucketSizes] <= freeElem ) {
460	abort( "Attempt to %s storage %p with corrupted header.\n"
461	"Possible cause is duplicate free on same block or overwriting of header information.",
462	name, addr );
463	} // if
464	#endif // __CFA_DEBUG__
465	size = freeElem->blockSize;
466	return false;
467	} // headers
468
469	// #ifdef __CFA_DEBUG__
470	// #if __SIZEOF_POINTER__ == 4
471	// #define MASK 0xdeadbeef
472	// #else
473	// #define MASK 0xdeadbeefdeadbeef
474	// #endif
475	// #define STRIDE size_t
476
477	// static void * Memset( void * addr, STRIDE size ) { // debug only
478	// if ( size % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, size %zd not multiple of %zd.", size, sizeof(STRIDE) );
479	// if ( (STRIDE)addr % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, addr %p not multiple of %zd.", addr, sizeof(STRIDE) );
480
481	// STRIDE * end = (STRIDE *)addr + size / sizeof(STRIDE);
482	// for ( STRIDE * p = (STRIDE )addr; p < end; p += 1 ) p = MASK;
483	// return addr;
484	// } // Memset
485	// #endif // __CFA_DEBUG__
486
487
488	#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."
489
490	static inline void * extend( size_t size ) with( heapManager ) {
491	lock( extlock __cfaabi_dbg_ctx2 );
492	ptrdiff_t rem = heapRemaining - size;
493	if ( rem < 0 ) {
494	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
495
496	size_t increase = ceiling2( size > heapExpand ? size : heapExpand, __page_size );
497	// Do not call abort or strerror( errno ) as they may call malloc.
498	if ( sbrk( increase ) == (void *)-1 ) { // failed, no memory ?
499	unlock( extlock );
500	__cfaabi_bits_print_nolock( STDERR_FILENO, NO_MEMORY_MSG, size );
501	_exit( EXIT_FAILURE ); // give up
502	} // if
503	// Make storage executable for thunks.
504	if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
505	unlock( extlock );
506	__cfaabi_bits_print_nolock( STDERR_FILENO, "extend() : internal error, mprotect failure, heapEnd:%p size:%zd, errno:%d.\n", heapEnd, increase, errno );
507	_exit( EXIT_FAILURE );
508	} // if
509	#ifdef __STATISTICS__
510	sbrk_calls += 1;
511	sbrk_storage += increase;
512	#endif // __STATISTICS__
513	#ifdef __CFA_DEBUG__
514	// Set new memory to garbage so subsequent uninitialized usages might fail.
515	memset( (char *)heapEnd + heapRemaining, '\xde', increase );
516	//Memset( (char *)heapEnd + heapRemaining, increase );
517	#endif // __CFA_DEBUG__
518	rem = heapRemaining + increase - size;
519	} // if
520
521	HeapManager.Storage * block = (HeapManager.Storage *)heapEnd;
522	heapRemaining = rem;
523	heapEnd = (char *)heapEnd + size;
524	unlock( extlock );
525	return block;
526	} // extend
527
528
529	static inline void * doMalloc( size_t size ) with( heapManager ) {
530	HeapManager.Storage * block; // pointer to new block of storage
531
532	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
533	// along with the block and is a multiple of the alignment size.
534
535	if ( unlikely( size > ULONG_MAX - sizeof(HeapManager.Storage) ) ) return 0p;
536	size_t tsize = size + sizeof(HeapManager.Storage);
537	if ( likely( tsize < mmapStart ) ) { // small size => sbrk
538	size_t posn;
539	#ifdef FASTLOOKUP
540	if ( tsize < LookupSizes ) posn = lookup[tsize];
541	else
542	#endif // FASTLOOKUP
543	posn = Bsearchl( (unsigned int)tsize, bucketSizes, (size_t)maxBucketsUsed );
544	HeapManager.FreeHeader * freeElem = &freeLists[posn];
545	verify( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
546	verify( tsize <= freeElem->blockSize ); // search failure ?
547	tsize = freeElem->blockSize; // total space needed for request
548
549	// Spin until the lock is acquired for this particular size of block.
550
551	#if BUCKETLOCK == SPINLOCK
552	lock( freeElem->lock __cfaabi_dbg_ctx2 );
553	block = freeElem->freeList; // remove node from stack
554	#else
555	block = pop( freeElem->freeList );
556	#endif // BUCKETLOCK
557	if ( unlikely( block == 0p ) ) { // no free block ?
558	#if BUCKETLOCK == SPINLOCK
559	unlock( freeElem->lock );
560	#endif // BUCKETLOCK
561
562	// Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
563	// and then carve it off.
564
565	block = (HeapManager.Storage *)extend( tsize ); // mutual exclusion on call
566	#if BUCKETLOCK == SPINLOCK
567	} else {
568	freeElem->freeList = block->header.kind.real.next;
569	unlock( freeElem->lock );
570	#endif // BUCKETLOCK
571	} // if
572
573	block->header.kind.real.home = freeElem; // pointer back to free list of apropriate size
574	} else { // large size => mmap
575	if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
576	tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
577	#ifdef __STATISTICS__
578	__atomic_add_fetch( &mmap_calls, 1, __ATOMIC_SEQ_CST );
579	__atomic_add_fetch( &mmap_storage, tsize, __ATOMIC_SEQ_CST );
580	#endif // __STATISTICS__
581
582	block = (HeapManager.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, mmapFd, 0 );
583	if ( block == (HeapManager.Storage *)MAP_FAILED ) { // failed ?
584	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
585	// Do not call strerror( errno ) as it may call malloc.
586	abort( "(HeapManager &)0x%p.doMalloc() : internal error, mmap failure, size:%zu errno:%d.", &heapManager, tsize, errno );
587	} //if
588	#ifdef __CFA_DEBUG__
589	// Set new memory to garbage so subsequent uninitialized usages might fail.
590	memset( block, '\xde', tsize );
591	//Memset( block, tsize );
592	#endif // __CFA_DEBUG__
593	block->header.kind.real.blockSize = tsize; // storage size for munmap
594	} // if
595
596	block->header.kind.real.size = size; // store allocation size
597	void * addr = &(block->data); // adjust off header to user bytes
598	verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
599
600	#ifdef __CFA_DEBUG__
601	__atomic_add_fetch( &allocUnfreed, tsize, __ATOMIC_SEQ_CST );
602	if ( traceHeap() ) {
603	enum { BufferSize = 64 };
604	char helpText[BufferSize];
605	int len = snprintf( helpText, BufferSize, "%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize );
606	__cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
607	} // if
608	#endif // __CFA_DEBUG__
609
610	return addr;
611	} // doMalloc
612
613
614	static inline void doFree( void * addr ) with( heapManager ) {
615	#ifdef __CFA_DEBUG__
616	if ( unlikely( heapManager.heapBegin == 0p ) ) {
617	abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
618	} // if
619	#endif // __CFA_DEBUG__
620
621	HeapManager.Storage.Header * header;
622	HeapManager.FreeHeader * freeElem;
623	size_t size, alignment; // not used (see realloc)
624
625	if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
626	#ifdef __STATISTICS__
627	__atomic_add_fetch( &munmap_calls, 1, __ATOMIC_SEQ_CST );
628	__atomic_add_fetch( &munmap_storage, size, __ATOMIC_SEQ_CST );
629	#endif // __STATISTICS__
630	if ( munmap( header, size ) == -1 ) {
631	abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
632	"Possible cause is invalid pointer.",
633	addr );
634	} // if
635	} else {
636	#ifdef __CFA_DEBUG__
637	// Set free memory to garbage so subsequent usages might fail.
638	memset( ((HeapManager.Storage *)header)->data, '\xde', freeElem->blockSize - sizeof( HeapManager.Storage ) );
639	//Memset( ((HeapManager.Storage *)header)->data, freeElem->blockSize - sizeof( HeapManager.Storage ) );
640	#endif // __CFA_DEBUG__
641
642	#ifdef __STATISTICS__
643	__atomic_add_fetch( &free_calls, 1, __ATOMIC_SEQ_CST );
644	__atomic_add_fetch( &free_storage, size, __ATOMIC_SEQ_CST );
645	#endif // __STATISTICS__
646
647	#if BUCKETLOCK == SPINLOCK
648	lock( freeElem->lock __cfaabi_dbg_ctx2 ); // acquire spin lock
649	header->kind.real.next = freeElem->freeList; // push on stack
650	freeElem->freeList = (HeapManager.Storage *)header;
651	unlock( freeElem->lock ); // release spin lock
652	#else
653	push( freeElem->freeList, (HeapManager.Storage )header );
654	#endif // BUCKETLOCK
655	} // if
656
657	#ifdef __CFA_DEBUG__
658	__atomic_add_fetch( &allocUnfreed, -size, __ATOMIC_SEQ_CST );
659	if ( traceHeap() ) {
660	char helpText[64];
661	int len = snprintf( helpText, sizeof(helpText), "Free( %p ) size:%zu\n", addr, size );
662	__cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
663	} // if
664	#endif // __CFA_DEBUG__
665	} // doFree
666
667
668	size_t prtFree( HeapManager & manager ) with( manager ) {
669	size_t total = 0;
670	#ifdef __STATISTICS__
671	__cfaabi_bits_acquire();
672	__cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
673	#endif // __STATISTICS__
674	for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
675	size_t size = freeLists[i].blockSize;
676	#ifdef __STATISTICS__
677	unsigned int N = 0;
678	#endif // __STATISTICS__
679
680	#if BUCKETLOCK == SPINLOCK
681	for ( HeapManager.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
682	#else
683	for(;;) {
684	// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
685	// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
686	// HeapManager.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
687	// typeof(p) temp = (( p )`next)->top; // FIX ME: direct assignent fails, initialization works`
688	// p = temp;
689	#endif // BUCKETLOCK
690	total += size;
691	#ifdef __STATISTICS__
692	N += 1;
693	#endif // __STATISTICS__
694	} // for
695
696	#ifdef __STATISTICS__
697	__cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
698	if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
699	#endif // __STATISTICS__
700	} // for
701	#ifdef __STATISTICS__
702	__cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
703	__cfaabi_bits_release();
704	#endif // __STATISTICS__
705	return (char )heapEnd - (char )heapBegin - total;
706	} // prtFree
707
708
709	static void ?{}( HeapManager & manager ) with( manager ) {
710	__page_size = sysconf( _SC_PAGESIZE );
711	__map_prot = PROT_READ \| PROT_WRITE \| PROT_EXEC;
712
713	for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
714	freeLists[i].blockSize = bucketSizes[i];
715	} // for
716
717	#ifdef FASTLOOKUP
718	unsigned int idx = 0;
719	for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
720	if ( i > bucketSizes[idx] ) idx += 1;
721	lookup[i] = idx;
722	} // for
723	#endif // FASTLOOKUP
724
725	if ( ! setMmapStart( default_mmap_start() ) ) {
726	abort( "HeapManager : internal error, mmap start initialization failure." );
727	} // if
728	heapExpand = default_heap_expansion();
729
730	char * end = (char *)sbrk( 0 );
731	heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, __page_size ) - end ); // move start of heap to multiple of alignment
732	} // HeapManager
733
734
735	static void ^?{}( HeapManager & ) {
736	#ifdef __STATISTICS__
737	if ( traceHeapTerm() ) {
738	printStats();
739	// prtUnfreed() called in heapAppStop()
740	} // if
741	#endif // __STATISTICS__
742	} // ~HeapManager
743
744
745	static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
746	void memory_startup( void ) {
747	#ifdef __CFA_DEBUG__
748	if ( heapBoot ) { // check for recursion during system boot
749	abort( "boot() : internal error, recursively invoked during system boot." );
750	} // if
751	heapBoot = true;
752	#endif // __CFA_DEBUG__
753
754	//verify( heapManager.heapBegin != 0 );
755	//heapManager{};
756	if ( heapManager.heapBegin == 0p ) heapManager{}; // sanity check
757	} // memory_startup
758
759	static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
760	void memory_shutdown( void ) {
761	^heapManager{};
762	} // memory_shutdown
763
764
765	static inline void * mallocNoStats( size_t size ) { // necessary for malloc statistics
766	verify( heapManager.heapBegin != 0p ); // called before memory_startup ?
767	if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
768
769	#if __SIZEOF_POINTER__ == 8
770	verify( size < ((typeof(size_t))1 << 48) );
771	#endif // __SIZEOF_POINTER__ == 8
772	return doMalloc( size );
773	} // mallocNoStats
774
775
776	static inline void * memalignNoStats( size_t alignment, size_t size ) {
777	if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
778
779	#ifdef __CFA_DEBUG__
780	checkAlign( alignment ); // check alignment
781	#endif // __CFA_DEBUG__
782
783	// if alignment <= default alignment, do normal malloc as two headers are unnecessary
784	if ( unlikely( alignment <= libAlign() ) ) return mallocNoStats( size );
785
786	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
787	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
788	// .-------------v-----------------v----------------v----------,
789	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
790	// `-------------^-----------------^-+--------------^----------'
791	// \|<--------------------------------' offset/align \|<-- alignment boundary
792
793	// subtract libAlign() because it is already the minimum alignment
794	// add sizeof(Storage) for fake header
795	char * addr = (char *)mallocNoStats( size + alignment - libAlign() + sizeof(HeapManager.Storage) );
796
797	// address in the block of the "next" alignment address
798	char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(HeapManager.Storage)), alignment );
799
800	// address of header from malloc
801	HeapManager.Storage.Header * realHeader = headerAddr( addr );
802	realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
803	// address of fake header * before* the alignment location
804	HeapManager.Storage.Header * fakeHeader = headerAddr( user );
805	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
806	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
807	// SKULLDUGGERY: odd alignment imples fake header
808	fakeHeader->kind.fake.alignment = alignment \| 1;
809
810	return user;
811	} // memalignNoStats
812
813
814	extern "C" {
815	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
816	// then malloc() returns a unique pointer value that can later be successfully passed to free().
817	void * malloc( size_t size ) {
818	#ifdef __STATISTICS__
819	if ( likely( size > 0 ) ) {
820	__atomic_add_fetch( &malloc_calls, 1, __ATOMIC_SEQ_CST );
821	__atomic_add_fetch( &malloc_storage, size, __ATOMIC_SEQ_CST );
822	} else {
823	__atomic_add_fetch( &malloc_zero_calls, 1, __ATOMIC_SEQ_CST );
824	} // if
825	#endif // __STATISTICS__
826
827	return mallocNoStats( size );
828	} // malloc
829
830
831	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
832	void * aalloc( size_t dim, size_t elemSize ) {
833	size_t size = dim * elemSize;
834	#ifdef __STATISTICS__
835	if ( likely( size > 0 ) ) {
836	__atomic_add_fetch( &aalloc_calls, 1, __ATOMIC_SEQ_CST );
837	__atomic_add_fetch( &aalloc_storage, size, __ATOMIC_SEQ_CST );
838	} else {
839	__atomic_add_fetch( &aalloc_zero_calls, 1, __ATOMIC_SEQ_CST );
840	} // if
841	#endif // __STATISTICS__
842
843	return mallocNoStats( size );
844	} // aalloc
845
846
847	// Same as aalloc() with memory set to zero.
848	void * calloc( size_t dim, size_t elemSize ) {
849	size_t size = dim * elemSize;
850	if ( unlikely( size ) == 0 ) { // 0 BYTE ALLOCATION RETURNS NULL POINTER
851	#ifdef __STATISTICS__
852	__atomic_add_fetch( &calloc_zero_calls, 1, __ATOMIC_SEQ_CST );
853	#endif // __STATISTICS__
854	return 0p;
855	} // if
856	#ifdef __STATISTICS__
857	__atomic_add_fetch( &calloc_calls, 1, __ATOMIC_SEQ_CST );
858	__atomic_add_fetch( &calloc_storage, dim * elemSize, __ATOMIC_SEQ_CST );
859	#endif // __STATISTICS__
860
861	char * addr = (char *)mallocNoStats( size );
862
863	HeapManager.Storage.Header * header;
864	HeapManager.FreeHeader * freeElem;
865	size_t bsize, alignment;
866
867	#ifndef __CFA_DEBUG__
868	bool mapped =
869	#endif // __CFA_DEBUG__
870	headers( "calloc", addr, header, freeElem, bsize, alignment );
871
872	#ifndef __CFA_DEBUG__
873	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
874	if ( ! mapped )
875	#endif // __CFA_DEBUG__
876	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
877	// `-header`-addr `-size
878	memset( addr, '\0', size ); // set to zeros
879
880	header->kind.real.blockSize \|= 2; // mark as zero filled
881	return addr;
882	} // calloc
883
884
885	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
886	// 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
887	// not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
888	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
889	void * resize( void * oaddr, size_t size ) {
890	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
891	if ( unlikely( size == 0 ) ) { // special cases
892	#ifdef __STATISTICS__
893	__atomic_add_fetch( &resize_zero_calls, 1, __ATOMIC_SEQ_CST );
894	#endif // __STATISTICS__
895	free( oaddr );
896	return 0p;
897	} // if
898	#ifdef __STATISTICS__
899	__atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
900	#endif // __STATISTICS__
901
902	if ( unlikely( oaddr == 0p ) ) {
903	#ifdef __STATISTICS__
904	__atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
905	#endif // __STATISTICS__
906	return mallocNoStats( size );
907	} // if
908
909	HeapManager.Storage.Header * header;
910	HeapManager.FreeHeader * freeElem;
911	size_t bsize, oalign;
912	headers( "resize", oaddr, header, freeElem, bsize, oalign );
913
914	size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
915	// same size, DO NOT preserve STICKY PROPERTIES.
916	if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
917	header->kind.real.blockSize &= -2; // no alignment and turn off 0 fill
918	header->kind.real.size = size; // reset allocation size
919	return oaddr;
920	} // if
921
922	#ifdef __STATISTICS__
923	__atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
924	#endif // __STATISTICS__
925
926	// change size, DO NOT preserve STICKY PROPERTIES.
927	free( oaddr );
928	return mallocNoStats( size ); // create new area
929	} // resize
930
931
932	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
933	// the old and new sizes.
934	void * realloc( void * oaddr, size_t size ) {
935	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
936	if ( unlikely( size == 0 ) ) { // special cases
937	#ifdef __STATISTICS__
938	__atomic_add_fetch( &realloc_zero_calls, 1, __ATOMIC_SEQ_CST );
939	#endif // __STATISTICS__
940	free( oaddr );
941	return 0p;
942	} // if
943	#ifdef __STATISTICS__
944	__atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
945	#endif // __STATISTICS__
946
947	if ( unlikely( oaddr == 0p ) ) {
948	#ifdef __STATISTICS__
949	__atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
950	#endif // __STATISTICS__
951	return mallocNoStats( size );
952	} // if
953
954	HeapManager.Storage.Header * header;
955	HeapManager.FreeHeader * freeElem;
956	size_t bsize, oalign;
957	headers( "realloc", oaddr, header, freeElem, bsize, oalign );
958
959	size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
960	size_t osize = header->kind.real.size; // old allocation size
961	bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
962	if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
963	header->kind.real.size = size; // reset allocation size
964	if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
965	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
966	} // if
967	return oaddr;
968	} // if
969
970	#ifdef __STATISTICS__
971	__atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
972	#endif // __STATISTICS__
973
974	// change size and copy old content to new storage
975
976	void * naddr;
977	if ( likely( oalign == libAlign() ) ) { // previous request not aligned ?
978	naddr = mallocNoStats( size ); // create new area
979	} else {
980	naddr = memalignNoStats( oalign, size ); // create new aligned area
981	} // if
982
983	headers( "realloc", naddr, header, freeElem, bsize, oalign );
984	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
985	free( oaddr );
986
987	if ( unlikely( ozfill ) ) { // previous request zero fill ?
988	header->kind.real.blockSize \|= 2; // mark new request as zero filled
989	if ( size > osize ) { // previous request larger ?
990	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
991	} // if
992	} // if
993	return naddr;
994	} // realloc
995
996
997	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
998	void * memalign( size_t alignment, size_t size ) {
999	#ifdef __STATISTICS__
1000	if ( likely( size > 0 ) ) {
1001	__atomic_add_fetch( &memalign_calls, 1, __ATOMIC_SEQ_CST );
1002	__atomic_add_fetch( &memalign_storage, size, __ATOMIC_SEQ_CST );
1003	} else {
1004	__atomic_add_fetch( &memalign_zero_calls, 1, __ATOMIC_SEQ_CST );
1005	} // if
1006	#endif // __STATISTICS__
1007
1008	return memalignNoStats( alignment, size );
1009	} // memalign
1010
1011
1012	// Same as aalloc() with memory alignment.
1013	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
1014	size_t size = dim * elemSize;
1015	#ifdef __STATISTICS__
1016	if ( likely( size > 0 ) ) {
1017	__atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1018	__atomic_add_fetch( &cmemalign_storage, size, __ATOMIC_SEQ_CST );
1019	} else {
1020	__atomic_add_fetch( &cmemalign_zero_calls, 1, __ATOMIC_SEQ_CST );
1021	} // if
1022	#endif // __STATISTICS__
1023
1024	return memalignNoStats( alignment, size );
1025	} // amemalign
1026
1027
1028	// Same as calloc() with memory alignment.
1029	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
1030	size_t size = dim * elemSize;
1031	if ( unlikely( size ) == 0 ) { // 0 BYTE ALLOCATION RETURNS NULL POINTER
1032	#ifdef __STATISTICS__
1033	__atomic_add_fetch( &cmemalign_zero_calls, 1, __ATOMIC_SEQ_CST );
1034	#endif // __STATISTICS__
1035	return 0p;
1036	} // if
1037	#ifdef __STATISTICS__
1038	__atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1039	__atomic_add_fetch( &cmemalign_storage, dim * elemSize, __ATOMIC_SEQ_CST );
1040	#endif // __STATISTICS__
1041
1042	char * addr = (char *)memalignNoStats( alignment, size );
1043
1044	HeapManager.Storage.Header * header;
1045	HeapManager.FreeHeader * freeElem;
1046	size_t bsize;
1047
1048	#ifndef __CFA_DEBUG__
1049	bool mapped =
1050	#endif // __CFA_DEBUG__
1051	headers( "cmemalign", addr, header, freeElem, bsize, alignment );
1052
1053	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1054	#ifndef __CFA_DEBUG__
1055	if ( ! mapped )
1056	#endif // __CFA_DEBUG__
1057	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1058	// `-header`-addr `-size
1059	memset( addr, '\0', size ); // set to zeros
1060
1061	header->kind.real.blockSize \|= 2; // mark as zero filled
1062	return addr;
1063	} // cmemalign
1064
1065
1066	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1067	// of alignment. This requirement is universally ignored.
1068	void * aligned_alloc( size_t alignment, size_t size ) {
1069	return memalign( alignment, size );
1070	} // aligned_alloc
1071
1072
1073	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1074	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1075	// is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1076	// free(3).
1077	int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
1078	if ( alignment < libAlign() \|\| ! is_pow2( alignment ) ) return EINVAL; // check alignment
1079	* memptr = memalign( alignment, size );
1080	return 0;
1081	} // posix_memalign
1082
1083
1084	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1085	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1086	void * valloc( size_t size ) {
1087	return memalign( __page_size, size );
1088	} // valloc
1089
1090
1091	// Same as valloc but rounds size to multiple of page size.
1092	void * pvalloc( size_t size ) {
1093	return memalign( __page_size, ceiling2( size, __page_size ) );
1094	} // pvalloc
1095
1096
1097	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1098	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1099	// 0p, no operation is performed.
1100	void free( void * addr ) {
1101	if ( unlikely( addr == 0p ) ) { // special case
1102	#ifdef __STATISTICS__
1103	__atomic_add_fetch( &free_zero_calls, 1, __ATOMIC_SEQ_CST );
1104	#endif // __STATISTICS__
1105
1106	// #ifdef __CFA_DEBUG__
1107	// if ( traceHeap() ) {
1108	// #define nullmsg "Free( 0x0 ) size:0\n"
1109	// // Do not debug print free( 0p ), as it can cause recursive entry from sprintf.
1110	// __cfaabi_dbg_write( nullmsg, sizeof(nullmsg) - 1 );
1111	// } // if
1112	// #endif // __CFA_DEBUG__
1113	return;
1114	} // exit
1115
1116	doFree( addr );
1117	} // free
1118
1119
1120	// Returns the alignment of an allocation.
1121	size_t malloc_alignment( void * addr ) {
1122	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1123	HeapManager.Storage.Header * header = headerAddr( addr );
1124	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1125	return header->kind.fake.alignment & -2; // remove flag from value
1126	} else {
1127	return libAlign(); // minimum alignment
1128	} // if
1129	} // malloc_alignment
1130
1131
1132	// Set the alignment for an the allocation and return previous alignment or 0 if no alignment.
1133	size_t malloc_alignment_set$( void * addr, size_t alignment ) {
1134	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1135	size_t ret;
1136	HeapManager.Storage.Header * header = headerAddr( addr );
1137	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1138	ret = header->kind.fake.alignment & -2; // remove flag from old value
1139	header->kind.fake.alignment = alignment \| 1; // add flag to new value
1140	} else {
1141	ret = 0; // => no alignment to change
1142	} // if
1143	return ret;
1144	} // malloc_alignment_set$
1145
1146
1147	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1148	bool malloc_zero_fill( void * addr ) {
1149	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1150	HeapManager.Storage.Header * header = headerAddr( addr );
1151	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1152	header = realHeader( header ); // backup from fake to real header
1153	} // if
1154	return (header->kind.real.blockSize & 2) != 0; // zero filled ?
1155	} // malloc_zero_fill
1156
1157	// Set allocation is zero filled and return previous zero filled.
1158	bool malloc_zero_fill_set$( void * addr ) {
1159	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1160	HeapManager.Storage.Header * header = headerAddr( addr );
1161	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1162	header = realHeader( header ); // backup from fake to real header
1163	} // if
1164	bool ret = (header->kind.real.blockSize & 2) != 0; // zero filled ?
1165	header->kind.real.blockSize \|= 2; // mark as zero filled
1166	return ret;
1167	} // malloc_zero_fill_set$
1168
1169
1170	// Returns original total allocation size (not bucket size) => array size is dimension * sizeif(T).
1171	size_t malloc_size( void * addr ) {
1172	if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1173	HeapManager.Storage.Header * header = headerAddr( addr );
1174	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1175	header = realHeader( header ); // backup from fake to real header
1176	} // if
1177	return header->kind.real.size;
1178	} // malloc_size
1179
1180	// Set allocation size and return previous size.
1181	size_t malloc_size_set$( void * addr, size_t size ) {
1182	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1183	HeapManager.Storage.Header * header = headerAddr( addr );
1184	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1185	header = realHeader( header ); // backup from fake to real header
1186	} // if
1187	size_t ret = header->kind.real.size;
1188	header->kind.real.size = size;
1189	return ret;
1190	} // malloc_size_set$
1191
1192
1193	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1194	// malloc or a related function.
1195	size_t malloc_usable_size( void * addr ) {
1196	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1197	HeapManager.Storage.Header * header;
1198	HeapManager.FreeHeader * freeElem;
1199	size_t bsize, alignment;
1200
1201	headers( "malloc_usable_size", addr, header, freeElem, bsize, alignment );
1202	return dataStorage( bsize, addr, header ); // data storage in bucket
1203	} // malloc_usable_size
1204
1205
1206	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1207	void malloc_stats( void ) {
1208	#ifdef __STATISTICS__
1209	printStats();
1210	if ( prtFree() ) prtFree( heapManager );
1211	#endif // __STATISTICS__
1212	} // malloc_stats
1213
1214
1215	// Changes the file descripter where malloc_stats() writes statistics.
1216	int malloc_stats_fd( int fd __attribute__(( unused )) ) {
1217	#ifdef __STATISTICS__
1218	int temp = stats_fd;
1219	stats_fd = fd;
1220	return temp;
1221	#else
1222	return -1;
1223	#endif // __STATISTICS__
1224	} // malloc_stats_fd
1225
1226
1227	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1228	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1229	int mallopt( int option, int value ) {
1230	choose( option ) {
1231	case M_TOP_PAD:
1232	heapExpand = ceiling2( value, __page_size ); return 1;
1233	case M_MMAP_THRESHOLD:
1234	if ( setMmapStart( value ) ) return 1;
1235	break;
1236	} // switch
1237	return 0; // error, unsupported
1238	} // mallopt
1239
1240
1241	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1242	int malloc_trim( size_t ) {
1243	return 0; // => impossible to release memory
1244	} // malloc_trim
1245
1246
1247	// Exports an XML string that describes the current state of the memory-allocation implementation in the caller.
1248	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1249	// malloc).
1250	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) {
1251	if ( options != 0 ) { errno = EINVAL; return -1; }
1252	#ifdef __STATISTICS__
1253	return printStatsXML( stream );
1254	#else
1255	return 0; // unsupported
1256	#endif // __STATISTICS__
1257	} // malloc_info
1258
1259
1260	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1261	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1262	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1263	// result. (The caller must free this memory.)
1264	void * malloc_get_state( void ) {
1265	return 0p; // unsupported
1266	} // malloc_get_state
1267
1268
1269	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1270	// structure pointed to by state.
1271	int malloc_set_state( void * ) {
1272	return 0; // unsupported
1273	} // malloc_set_state
1274	} // extern "C"
1275
1276
1277	// Must have CFA linkage to overload with C linkage realloc.
1278	void * resize( void * oaddr, size_t nalign, size_t size ) {
1279	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1280	if ( unlikely( size == 0 ) ) { // special cases
1281	#ifdef __STATISTICS__
1282	__atomic_add_fetch( &resize_zero_calls, 1, __ATOMIC_SEQ_CST );
1283	#endif // __STATISTICS__
1284	free( oaddr );
1285	return 0p;
1286	} // if
1287
1288	if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1289	#ifdef __CFA_DEBUG__
1290	else checkAlign( nalign ); // check alignment
1291	#endif // __CFA_DEBUG__
1292
1293	if ( unlikely( oaddr == 0p ) ) {
1294	#ifdef __STATISTICS__
1295	__atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
1296	__atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1297	#endif // __STATISTICS__
1298	return memalignNoStats( nalign, size );
1299	} // if
1300
1301	// Attempt to reuse existing alignment.
1302	HeapManager.Storage.Header * header = headerAddr( oaddr );
1303	bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1304	size_t oalign;
1305	if ( isFakeHeader ) {
1306	oalign = header->kind.fake.alignment & -2; // old alignment
1307	if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1308	&& ( oalign <= nalign // going down
1309	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1310	) {
1311	headerAddr( oaddr )->kind.fake.alignment = nalign \| 1; // update alignment (could be the same)
1312	HeapManager.FreeHeader * freeElem;
1313	size_t bsize, oalign;
1314	headers( "resize", oaddr, header, freeElem, bsize, oalign );
1315	size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1316
1317	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1318	headerAddr( oaddr )->kind.fake.alignment = nalign \| 1; // update alignment (could be the same)
1319
1320	header->kind.real.blockSize &= -2; // turn off 0 fill
1321	header->kind.real.size = size; // reset allocation size
1322	return oaddr;
1323	} // if
1324	} // if
1325	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1326	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1327	return resize( oaddr, size ); // duplicate special case checks
1328	} // if
1329
1330	#ifdef __STATISTICS__
1331	__atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1332	#endif // __STATISTICS__
1333
1334	// change size, DO NOT preserve STICKY PROPERTIES.
1335	free( oaddr );
1336	return memalignNoStats( nalign, size ); // create new aligned area
1337	} // resize
1338
1339
1340	void * realloc( void * oaddr, size_t nalign, size_t size ) {
1341	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1342	if ( unlikely( size == 0 ) ) { // special cases
1343	#ifdef __STATISTICS__
1344	__atomic_add_fetch( &realloc_zero_calls, 1, __ATOMIC_SEQ_CST );
1345	#endif // __STATISTICS__
1346	free( oaddr );
1347	return 0p;
1348	} // if
1349
1350	if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1351	#ifdef __CFA_DEBUG__
1352	else checkAlign( nalign ); // check alignment
1353	#endif // __CFA_DEBUG__
1354
1355	if ( unlikely( oaddr == 0p ) ) {
1356	#ifdef __STATISTICS__
1357	__atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1358	__atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1359	#endif // __STATISTICS__
1360	return memalignNoStats( nalign, size );
1361	} // if
1362
1363	// Attempt to reuse existing alignment.
1364	HeapManager.Storage.Header * header = headerAddr( oaddr );
1365	bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1366	size_t oalign;
1367	if ( isFakeHeader ) {
1368	oalign = header->kind.fake.alignment & -2; // old alignment
1369	if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1370	&& ( oalign <= nalign // going down
1371	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1372	) {
1373	headerAddr( oaddr )->kind.fake.alignment = nalign \| 1; // update alignment (could be the same)
1374	return realloc( oaddr, size ); // duplicate alignment and special case checks
1375	} // if
1376	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1377	&& nalign == libAlign() ) // new alignment also on libAlign => no fake header needed
1378	return realloc( oaddr, size ); // duplicate alignment and special case checks
1379
1380	#ifdef __STATISTICS__
1381	__atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1382	__atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1383	#endif // __STATISTICS__
1384
1385	HeapManager.FreeHeader * freeElem;
1386	size_t bsize;
1387	headers( "realloc", oaddr, header, freeElem, bsize, oalign );
1388
1389	// change size and copy old content to new storage
1390
1391	size_t osize = header->kind.real.size; // old allocation size
1392	bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
1393
1394	void * naddr = memalignNoStats( nalign, size ); // create new aligned area
1395
1396	headers( "realloc", naddr, header, freeElem, bsize, oalign );
1397	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1398	free( oaddr );
1399
1400	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1401	header->kind.real.blockSize \|= 2; // mark new request as zero filled
1402	if ( size > osize ) { // previous request larger ?
1403	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1404	} // if
1405	} // if
1406	return naddr;
1407	} // realloc
1408
1409
1410	// Local Variables: //
1411	// tab-width: 4 //
1412	// compile-command: "cfa -nodebug -O2 heap.cfa" //
1413	// End: //

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format