Context Navigation

heap.cfa @ c3b9d639

ADTast-experimentalpthread-emulationqualifiedEnum

Last change on this file since c3b9d639 was 032234bd, checked in by Thierry Delisle <tdelisle@…>, 2 years ago
Visibility of the core libcfa files.
Property mode set to `100644`
File size: 58.1 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 : Fri Apr 29 19:05:03 2022
13	// Update Count : 1167
14	//
15
16	#include <string.h> // memset, memcpy
17	#include <limits.h> // ULONG_MAX
18	#include <stdlib.h> // EXIT_FAILURE
19	#include <errno.h> // errno, ENOMEM, EINVAL
20	#include <unistd.h> // STDERR_FILENO, sbrk, sysconf
21	#include <malloc.h> // memalign, malloc_usable_size
22	#include <sys/mman.h> // mmap, munmap
23	#include <sys/sysinfo.h> // get_nprocs
24
25	#include "bits/align.hfa" // libAlign
26	#include "bits/defs.hfa" // likely, unlikely
27	#include "bits/locks.hfa" // __spinlock_t
28	#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
29	#include "math.hfa" // min
30	#include "bitmanip.hfa" // is_pow2, ceiling2
31
32	#define FASTLOOKUP
33	#define __STATISTICS__
34
35
36	static bool traceHeap = false;
37
38	inline bool traceHeap() libcfa_public { return traceHeap; }
39
40	bool traceHeapOn() libcfa_public {
41	bool temp = traceHeap;
42	traceHeap = true;
43	return temp;
44	} // traceHeapOn
45
46	bool traceHeapOff() libcfa_public {
47	bool temp = traceHeap;
48	traceHeap = false;
49	return temp;
50	} // traceHeapOff
51
52	bool traceHeapTerm() libcfa_public { return false; }
53
54
55	static bool prtFree = false;
56
57	static bool prtFree() {
58	return prtFree;
59	} // prtFree
60
61	static bool prtFreeOn() {
62	bool temp = prtFree;
63	prtFree = true;
64	return temp;
65	} // prtFreeOn
66
67	static bool prtFreeOff() {
68	bool temp = prtFree;
69	prtFree = false;
70	return temp;
71	} // prtFreeOff
72
73
74	enum {
75	// The default extension heap amount in units of bytes. When the current heap reaches the brk address, the brk
76	// address is extended by the extension amount.
77	__CFA_DEFAULT_HEAP_EXPANSION__ = 10 * 1024 * 1024,
78
79	// The mmap crossover point during allocation. Allocations less than this amount are allocated from buckets; values
80	// greater than or equal to this value are mmap from the operating system.
81	__CFA_DEFAULT_MMAP_START__ = 512 * 1024 + 1,
82
83	// The default unfreed storage amount in units of bytes. When the uC++ program ends it subtracts this amount from
84	// the malloc/free counter to adjust for storage the program does not free.
85	__CFA_DEFAULT_HEAP_UNFREED__ = 0
86	}; // enum
87
88
89	//####################### Heap Statistics ####################
90
91
92	#ifdef __STATISTICS__
93	enum { CntTriples = 12 }; // number of counter triples
94	enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC, FREE };
95
96	struct StatsOverlay { // overlay for iteration
97	unsigned int calls, calls_0;
98	unsigned long long int request, alloc;
99	};
100
101	// Heap statistics counters.
102	union HeapStatistics {
103	struct { // minimum qualification
104	unsigned int malloc_calls, malloc_0_calls;
105	unsigned long long int malloc_storage_request, malloc_storage_alloc;
106	unsigned int aalloc_calls, aalloc_0_calls;
107	unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
108	unsigned int calloc_calls, calloc_0_calls;
109	unsigned long long int calloc_storage_request, calloc_storage_alloc;
110	unsigned int memalign_calls, memalign_0_calls;
111	unsigned long long int memalign_storage_request, memalign_storage_alloc;
112	unsigned int amemalign_calls, amemalign_0_calls;
113	unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
114	unsigned int cmemalign_calls, cmemalign_0_calls;
115	unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
116	unsigned int resize_calls, resize_0_calls;
117	unsigned long long int resize_storage_request, resize_storage_alloc;
118	unsigned int realloc_calls, realloc_0_calls;
119	unsigned long long int realloc_storage_request, realloc_storage_alloc;
120	unsigned int free_calls, free_null_calls;
121	unsigned long long int free_storage_request, free_storage_alloc;
122	unsigned int away_pulls, away_pushes;
123	unsigned long long int away_storage_request, away_storage_alloc;
124	unsigned int mmap_calls, mmap_0_calls; // no zero calls
125	unsigned long long int mmap_storage_request, mmap_storage_alloc;
126	unsigned int munmap_calls, munmap_0_calls; // no zero calls
127	unsigned long long int munmap_storage_request, munmap_storage_alloc;
128	};
129	struct StatsOverlay counters[CntTriples]; // overlay for iteration
130	}; // HeapStatistics
131
132	static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(StatsOverlay),
133	"Heap statistics counter-triplets does not match with array size" );
134
135	static void HeapStatisticsCtor( HeapStatistics & stats ) {
136	memset( &stats, '\0', sizeof(stats) ); // very fast
137	// for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
138	// stats.counters[i].calls = stats.counters[i].calls_0 = stats.counters[i].request = stats.counters[i].alloc = 0;
139	// } // for
140	} // HeapStatisticsCtor
141
142	static HeapStatistics & ?+=?( HeapStatistics & lhs, const HeapStatistics & rhs ) {
143	for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
144	lhs.counters[i].calls += rhs.counters[i].calls;
145	lhs.counters[i].calls_0 += rhs.counters[i].calls_0;
146	lhs.counters[i].request += rhs.counters[i].request;
147	lhs.counters[i].alloc += rhs.counters[i].alloc;
148	} // for
149	return lhs;
150	} // ?+=?
151	#endif // __STATISTICS__
152
153
154	#define SPINLOCK 0
155	#define LOCKFREE 1
156	#define BUCKETLOCK SPINLOCK
157	#if BUCKETLOCK == SPINLOCK
158	#elif BUCKETLOCK == LOCKFREE
159	#include <stackLockFree.hfa>
160	#else
161	#error undefined lock type for bucket lock
162	#endif // LOCKFREE
163
164	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
165	// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
166	enum { NoBucketSizes = 91 }; // number of buckets sizes
167
168	struct Heap {
169	struct Storage {
170	struct Header { // header
171	union Kind {
172	struct RealHeader {
173	union {
174	struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
175	union {
176	// 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
177	// FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
178	void * home; // allocated block points back to home locations (must overlay alignment)
179	size_t blockSize; // size for munmap (must overlay alignment)
180	#if BUCKETLOCK == SPINLOCK
181	Storage * next; // freed block points to next freed block of same size
182	#endif // SPINLOCK
183	};
184	size_t size; // allocation size in bytes
185	};
186	#if BUCKETLOCK == LOCKFREE
187	Link(Storage) next; // freed block points next freed block of same size (double-wide)
188	#endif // LOCKFREE
189	};
190	} real; // RealHeader
191
192	struct FakeHeader {
193	uintptr_t alignment; // 1st low-order bit => fake header & alignment
194	uintptr_t offset;
195	} fake; // FakeHeader
196	} kind; // Kind
197	} header; // Header
198
199	char pad[libAlign() - sizeof( Header )];
200	char data[0]; // storage
201	}; // Storage
202
203	static_assert( libAlign() >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );
204
205	struct FreeHeader {
206	size_t blockSize __attribute__(( aligned (8) )); // size of allocations on this list
207	#if BUCKETLOCK == SPINLOCK
208	__spinlock_t lock;
209	Storage * freeList;
210	#else
211	StackLF(Storage) freeList;
212	#endif // BUCKETLOCK
213	} __attribute__(( aligned (8) )); // FreeHeader
214
215	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
216
217	__spinlock_t extlock; // protects allocation-buffer extension
218	void * heapBegin; // start of heap
219	void * heapEnd; // logical end of heap
220	size_t heapRemaining; // amount of storage not allocated in the current chunk
221	}; // Heap
222
223	#if BUCKETLOCK == LOCKFREE
224	static inline {
225	Link(Heap.Storage) * ?`next( Heap.Storage * this ) { return &this->header.kind.real.next; }
226	void ?{}( Heap.FreeHeader & ) {}
227	void ^?{}( Heap.FreeHeader & ) {}
228	} // distribution
229	#endif // LOCKFREE
230
231	static inline size_t getKey( const Heap.FreeHeader & freeheader ) { return freeheader.blockSize; }
232
233
234	#ifdef FASTLOOKUP
235	enum { LookupSizes = 65_536 + sizeof(Heap.Storage) }; // number of fast lookup sizes
236	static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
237	#endif // FASTLOOKUP
238
239	static const off_t mmapFd = -1; // fake or actual fd for anonymous file
240	#ifdef __CFA_DEBUG__
241	static bool heapBoot = 0; // detect recursion during boot
242	#endif // __CFA_DEBUG__
243
244
245	// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
246	// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
247	// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
248	static const unsigned int bucketSizes[] @= { // different bucket sizes
249	16 + sizeof(Heap.Storage), 32 + sizeof(Heap.Storage), 48 + sizeof(Heap.Storage), 64 + sizeof(Heap.Storage), // 4
250	96 + sizeof(Heap.Storage), 112 + sizeof(Heap.Storage), 128 + sizeof(Heap.Storage), // 3
251	160, 192, 224, 256 + sizeof(Heap.Storage), // 4
252	320, 384, 448, 512 + sizeof(Heap.Storage), // 4
253	640, 768, 896, 1_024 + sizeof(Heap.Storage), // 4
254	1_536, 2_048 + sizeof(Heap.Storage), // 2
255	2_560, 3_072, 3_584, 4_096 + sizeof(Heap.Storage), // 4
256	6_144, 8_192 + sizeof(Heap.Storage), // 2
257	9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(Heap.Storage), // 8
258	18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(Heap.Storage), // 8
259	36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(Heap.Storage), // 8
260	73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(Heap.Storage), // 8
261	147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(Heap.Storage), // 8
262	294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(Heap.Storage), // 8
263	655_360, 786_432, 917_504, 1_048_576 + sizeof(Heap.Storage), // 4
264	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(Heap.Storage), // 8
265	2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(Heap.Storage), // 4
266	};
267
268	static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
269
270	// The constructor for heapManager is called explicitly in memory_startup.
271	static Heap heapManager __attribute__(( aligned (128) )) @= {}; // size of cache line to prevent false sharing
272
273
274	//####################### Memory Allocation Routines Helpers ####################
275
276
277	#ifdef __CFA_DEBUG__
278	static size_t allocUnfreed; // running total of allocations minus frees
279
280	static void prtUnfreed() {
281	if ( allocUnfreed != 0 ) {
282	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
283	char helpText[512];
284	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
285	"CFA warning (UNIX pid:%ld) : program terminating with %zu(0x%zx) bytes of storage allocated but not freed.\n"
286	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
287	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
288	} // if
289	} // prtUnfreed
290
291	extern int cfa_main_returned; // from interpose.cfa
292	extern "C" {
293	void heapAppStart() { // called by __cfaabi_appready_startup
294	allocUnfreed = 0;
295	} // heapAppStart
296
297	void heapAppStop() { // called by __cfaabi_appready_startdown
298	fclose( stdin ); fclose( stdout );
299	if ( cfa_main_returned ) prtUnfreed(); // do not check unfreed storage if exit called
300	} // heapAppStop
301	} // extern "C"
302	#endif // __CFA_DEBUG__
303
304
305	#ifdef __STATISTICS__
306	static HeapStatistics stats; // zero filled
307	static unsigned int sbrk_calls;
308	static unsigned long long int sbrk_storage;
309	// Statistics file descriptor (changed by malloc_stats_fd).
310	static int stats_fd = STDERR_FILENO; // default stderr
311
312	#define prtFmt \
313	"\nHeap statistics: (storage request / allocation)\n" \
314	" malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
315	" aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
316	" calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
317	" memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
318	" amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
319	" cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
320	" resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
321	" realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
322	" free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n" \
323	" sbrk calls %'u; storage %'llu bytes\n" \
324	" mmap calls %'u; storage %'llu / %'llu bytes\n" \
325	" munmap calls %'u; storage %'llu / %'llu bytes\n" \
326
327	// Use "write" because streams may be shutdown when calls are made.
328	static int printStats() { // see malloc_stats
329	char helpText[sizeof(prtFmt) + 1024]; // space for message and values
330	return __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText), prtFmt,
331	stats.malloc_calls, stats.malloc_0_calls, stats.malloc_storage_request, stats.malloc_storage_alloc,
332	stats.aalloc_calls, stats.aalloc_0_calls, stats.aalloc_storage_request, stats.aalloc_storage_alloc,
333	stats.calloc_calls, stats.calloc_0_calls, stats.calloc_storage_request, stats.calloc_storage_alloc,
334	stats.memalign_calls, stats.memalign_0_calls, stats.memalign_storage_request, stats.memalign_storage_alloc,
335	stats.amemalign_calls, stats.amemalign_0_calls, stats.amemalign_storage_request, stats.amemalign_storage_alloc,
336	stats.cmemalign_calls, stats.cmemalign_0_calls, stats.cmemalign_storage_request, stats.cmemalign_storage_alloc,
337	stats.resize_calls, stats.resize_0_calls, stats.resize_storage_request, stats.resize_storage_alloc,
338	stats.realloc_calls, stats.realloc_0_calls, stats.realloc_storage_request, stats.realloc_storage_alloc,
339	stats.free_calls, stats.free_null_calls, stats.free_storage_request, stats.free_storage_alloc,
340	sbrk_calls, sbrk_storage,
341	stats.mmap_calls, stats.mmap_storage_request, stats.mmap_storage_alloc,
342	stats.munmap_calls, stats.munmap_storage_request, stats.munmap_storage_alloc
343	);
344	} // printStats
345
346	#define prtFmtXML \
347	"<malloc version=\"1\">\n" \
348	"<heap nr=\"0\">\n" \
349	"<sizes>\n" \
350	"</sizes>\n" \
351	"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
352	"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
353	"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
354	"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
355	"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
356	"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
357	"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
358	"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
359	"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
360	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n" \
361	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n" \
362	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
363	"</malloc>"
364
365	static int printStatsXML( FILE * stream ) { // see malloc_info
366	char helpText[sizeof(prtFmtXML) + 1024]; // space for message and values
367	return __cfaabi_bits_print_buffer( fileno( stream ), helpText, sizeof(helpText), prtFmtXML,
368	stats.malloc_calls, stats.malloc_0_calls, stats.malloc_storage_request, stats.malloc_storage_alloc,
369	stats.aalloc_calls, stats.aalloc_0_calls, stats.aalloc_storage_request, stats.aalloc_storage_alloc,
370	stats.calloc_calls, stats.calloc_0_calls, stats.calloc_storage_request, stats.calloc_storage_alloc,
371	stats.memalign_calls, stats.memalign_0_calls, stats.memalign_storage_request, stats.memalign_storage_alloc,
372	stats.amemalign_calls, stats.amemalign_0_calls, stats.amemalign_storage_request, stats.amemalign_storage_alloc,
373	stats.cmemalign_calls, stats.cmemalign_0_calls, stats.cmemalign_storage_request, stats.cmemalign_storage_alloc,
374	stats.resize_calls, stats.resize_0_calls, stats.resize_storage_request, stats.resize_storage_alloc,
375	stats.realloc_calls, stats.realloc_0_calls, stats.realloc_storage_request, stats.realloc_storage_alloc,
376	stats.free_calls, stats.free_null_calls, stats.free_storage_request, stats.free_storage_alloc,
377	sbrk_calls, sbrk_storage,
378	stats.mmap_calls, stats.mmap_storage_request, stats.mmap_storage_alloc,
379	stats.munmap_calls, stats.munmap_storage_request, stats.munmap_storage_alloc
380	);
381	} // printStatsXML
382	#endif // __STATISTICS__
383
384
385	// statically allocated variables => zero filled.
386	static size_t heapExpand; // sbrk advance
387	static size_t mmapStart; // cross over point for mmap
388	static unsigned int maxBucketsUsed; // maximum number of buckets in use
389	// extern visibility, used by runtime kernel
390	// would be cool to remove libcfa_public but it's needed for libcfathread
391	libcfa_public size_t __page_size; // architecture pagesize
392	libcfa_public int __map_prot; // common mmap/mprotect protection
393
394
395	// thunk problem
396	size_t Bsearchl( unsigned int key, const unsigned int * vals, size_t dim ) {
397	size_t l = 0, m, h = dim;
398	while ( l < h ) {
399	m = (l + h) / 2;
400	if ( (unsigned int &)(vals[m]) < key ) { // cast away const
401	l = m + 1;
402	} else {
403	h = m;
404	} // if
405	} // while
406	return l;
407	} // Bsearchl
408
409
410	static inline bool setMmapStart( size_t value ) { // true => mmapped, false => sbrk
411	if ( value < __page_size \|\| bucketSizes[NoBucketSizes - 1] < value ) return false;
412	mmapStart = value; // set global
413
414	// find the closest bucket size less than or equal to the mmapStart size
415	maxBucketsUsed = Bsearchl( (unsigned int)mmapStart, bucketSizes, NoBucketSizes ); // binary search
416	assert( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
417	assert( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
418	return true;
419	} // setMmapStart
420
421
422	// <-------+----------------------------------------------------> bsize (bucket size)
423	// \|header \|addr
424	//==================================================================================
425	// align/offset \|
426	// <-----------------<------------+-----------------------------> bsize (bucket size)
427	// \|fake-header \| addr
428	#define HeaderAddr( addr ) ((Heap.Storage.Header )( (char )addr - sizeof(Heap.Storage) ))
429	#define RealHeader( header ) ((Heap.Storage.Header )((char )header - header->kind.fake.offset))
430
431	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
432	// \|header \|addr
433	//==================================================================================
434	// align/offset \|
435	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
436	// \|fake-header \|addr
437	#define DataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
438
439
440	static inline void checkAlign( size_t alignment ) {
441	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) {
442	abort( "** Error ** alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
443	} // if
444	} // checkAlign
445
446
447	static inline void checkHeader( bool check, const char name[], void * addr ) {
448	if ( unlikely( check ) ) { // bad address ?
449	abort( "** Error ** attempt to %s storage %p with address outside the heap.\n"
450	"Possible cause is duplicate free on same block or overwriting of memory.",
451	name, addr );
452	} // if
453	} // checkHeader
454
455
456	// Manipulate sticky bits stored in unused 3 low-order bits of an address.
457	// bit0 => alignment => fake header
458	// bit1 => zero filled (calloc)
459	// bit2 => mapped allocation versus sbrk
460	#define StickyBits( header ) (((header)->kind.real.blockSize & 0x7))
461	#define ClearStickyBits( addr ) (typeof(addr))((uintptr_t)(addr) & ~7)
462	#define MarkAlignmentBit( align ) ((align) \| 1)
463	#define AlignmentBit( header ) ((((header)->kind.fake.alignment) & 1))
464	#define ClearAlignmentBit( header ) (((header)->kind.fake.alignment) & ~1)
465	#define ZeroFillBit( header ) ((((header)->kind.real.blockSize) & 2))
466	#define ClearZeroFillBit( header ) ((((header)->kind.real.blockSize) &= ~2))
467	#define MarkZeroFilledBit( header ) ((header)->kind.real.blockSize \|= 2)
468	#define MmappedBit( header ) ((((header)->kind.real.blockSize) & 4))
469	#define MarkMmappedBit( size ) ((size) \| 4)
470
471
472	static inline void fakeHeader( Heap.Storage.Header *& header, size_t & alignment ) {
473	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
474	alignment = ClearAlignmentBit( header ); // clear flag from value
475	#ifdef __CFA_DEBUG__
476	checkAlign( alignment ); // check alignment
477	#endif // __CFA_DEBUG__
478	header = RealHeader( header ); // backup from fake to real header
479	} else {
480	alignment = libAlign(); // => no fake header
481	} // if
482	} // fakeHeader
483
484
485	static inline bool headers( const char name[] __attribute__(( unused )), void * addr, Heap.Storage.Header *& header,
486	Heap.FreeHeader *& freeHead, size_t & size, size_t & alignment ) with( heapManager ) {
487	header = HeaderAddr( addr );
488
489	#ifdef __CFA_DEBUG__
490	checkHeader( header < (Heap.Storage.Header *)heapBegin, name, addr ); // bad low address ?
491	#endif // __CFA_DEBUG__
492
493	if ( likely( ! StickyBits( header ) ) ) { // no sticky bits ?
494	freeHead = (Heap.FreeHeader *)(header->kind.real.home);
495	alignment = libAlign();
496	} else {
497	fakeHeader( header, alignment );
498	if ( unlikely( MmappedBit( header ) ) ) { // mmapped ?
499	verify( addr < heapBegin \|\| heapEnd < addr );
500	size = ClearStickyBits( header->kind.real.blockSize ); // mmap size
501	return true;
502	} // if
503
504	freeHead = (Heap.FreeHeader *)(ClearStickyBits( header->kind.real.home ));
505	} // if
506	size = freeHead->blockSize;
507
508	#ifdef __CFA_DEBUG__
509	checkHeader( header < (Heap.Storage.Header )heapBegin \|\| (Heap.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
510
511	Heap * homeManager;
512	if ( unlikely( freeHead == 0p \|\| // freed and only free-list node => null link
513	// freed and link points at another free block not to a bucket in the bucket array.
514	freeHead < &freeLists[0] \|\| &freeLists[NoBucketSizes] <= freeHead ) ) {
515	abort( "** Error ** attempt to %s storage %p with corrupted header.\n"
516	"Possible cause is duplicate free on same block or overwriting of header information.",
517	name, addr );
518	} // if
519	#endif // __CFA_DEBUG__
520
521	return false;
522	} // headers
523
524	// #ifdef __CFA_DEBUG__
525	// #if __SIZEOF_POINTER__ == 4
526	// #define MASK 0xdeadbeef
527	// #else
528	// #define MASK 0xdeadbeefdeadbeef
529	// #endif
530	// #define STRIDE size_t
531
532	// static void * Memset( void * addr, STRIDE size ) { // debug only
533	// if ( size % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, size %zd not multiple of %zd.", size, sizeof(STRIDE) );
534	// if ( (STRIDE)addr % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, addr %p not multiple of %zd.", addr, sizeof(STRIDE) );
535
536	// STRIDE * end = (STRIDE *)addr + size / sizeof(STRIDE);
537	// for ( STRIDE * p = (STRIDE )addr; p < end; p += 1 ) p = MASK;
538	// return addr;
539	// } // Memset
540	// #endif // __CFA_DEBUG__
541
542
543	#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."
544
545	static inline void * extend( size_t size ) with( heapManager ) {
546	lock( extlock __cfaabi_dbg_ctx2 );
547
548	ptrdiff_t rem = heapRemaining - size;
549	if ( unlikely( rem < 0 ) ) {
550	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
551
552	size_t increase = ceiling2( size > heapExpand ? size : heapExpand, __page_size );
553	// Do not call abort or strerror( errno ) as they may call malloc.
554	if ( sbrk( increase ) == (void *)-1 ) { // failed, no memory ?
555	unlock( extlock );
556	__cfaabi_bits_print_nolock( STDERR_FILENO, NO_MEMORY_MSG, size );
557	_exit( EXIT_FAILURE ); // give up
558	} // if
559
560	// Make storage executable for thunks.
561	if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
562	unlock( extlock );
563	__cfaabi_bits_print_nolock( STDERR_FILENO, "extend() : internal error, mprotect failure, heapEnd:%p size:%zd, errno:%d.\n", heapEnd, increase, errno );
564	_exit( EXIT_FAILURE );
565	} // if
566
567	#ifdef __STATISTICS__
568	sbrk_calls += 1;
569	sbrk_storage += increase;
570	#endif // __STATISTICS__
571
572	#ifdef __CFA_DEBUG__
573	// Set new memory to garbage so subsequent uninitialized usages might fail.
574	memset( (char *)heapEnd + heapRemaining, '\xde', increase );
575	//Memset( (char *)heapEnd + heapRemaining, increase );
576	#endif // __CFA_DEBUG__
577
578	rem = heapRemaining + increase - size;
579	} // if
580
581	Heap.Storage * block = (Heap.Storage *)heapEnd;
582	heapRemaining = rem;
583	heapEnd = (char *)heapEnd + size;
584	unlock( extlock );
585	return block;
586	} // extend
587
588
589	static inline void * doMalloc( size_t size ) with( heapManager ) {
590	Heap.Storage * block; // pointer to new block of storage
591
592	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
593	// along with the block and is a multiple of the alignment size.
594
595	size_t tsize = size + sizeof(Heap.Storage);
596
597	if ( likely( tsize < mmapStart ) ) { // small size => sbrk
598	size_t posn;
599	#ifdef FASTLOOKUP
600	if ( tsize < LookupSizes ) posn = lookup[tsize];
601	else
602	#endif // FASTLOOKUP
603	posn = Bsearchl( (unsigned int)tsize, bucketSizes, (size_t)maxBucketsUsed );
604	Heap.FreeHeader * freeElem = &freeLists[posn];
605	verify( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
606	verify( tsize <= freeElem->blockSize ); // search failure ?
607	tsize = freeElem->blockSize; // total space needed for request
608
609	// Spin until the lock is acquired for this particular size of block.
610
611	#if BUCKETLOCK == SPINLOCK
612	lock( freeElem->lock __cfaabi_dbg_ctx2 );
613	block = freeElem->freeList; // remove node from stack
614	#else
615	block = pop( freeElem->freeList );
616	#endif // BUCKETLOCK
617	if ( unlikely( block == 0p ) ) { // no free block ?
618	#if BUCKETLOCK == SPINLOCK
619	unlock( freeElem->lock );
620	#endif // BUCKETLOCK
621
622	// Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
623	// and then carve it off.
624
625	block = (Heap.Storage *)extend( tsize ); // mutual exclusion on call
626	#if BUCKETLOCK == SPINLOCK
627	} else {
628	freeElem->freeList = block->header.kind.real.next;
629	unlock( freeElem->lock );
630	#endif // BUCKETLOCK
631	} // if
632
633	block->header.kind.real.home = freeElem; // pointer back to free list of apropriate size
634	} else { // large size => mmap
635	if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
636	tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
637	#ifdef __STATISTICS__
638	__atomic_add_fetch( &stats.mmap_calls, 1, __ATOMIC_SEQ_CST );
639	__atomic_add_fetch( &stats.mmap_storage_request, size, __ATOMIC_SEQ_CST );
640	__atomic_add_fetch( &stats.mmap_storage_alloc, tsize, __ATOMIC_SEQ_CST );
641	#endif // __STATISTICS__
642
643	block = (Heap.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, mmapFd, 0 );
644	if ( block == (Heap.Storage *)MAP_FAILED ) { // failed ?
645	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
646	// Do not call strerror( errno ) as it may call malloc.
647	abort( "(Heap &)0x%p.doMalloc() : internal error, mmap failure, size:%zu errno:%d.", &heapManager, tsize, errno );
648	} //if
649	#ifdef __CFA_DEBUG__
650	// Set new memory to garbage so subsequent uninitialized usages might fail.
651	memset( block, '\xde', tsize );
652	//Memset( block, tsize );
653	#endif // __CFA_DEBUG__
654	block->header.kind.real.blockSize = MarkMmappedBit( tsize ); // storage size for munmap
655	} // if
656
657	block->header.kind.real.size = size; // store allocation size
658	void * addr = &(block->data); // adjust off header to user bytes
659	verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
660
661	#ifdef __CFA_DEBUG__
662	__atomic_add_fetch( &allocUnfreed, tsize, __ATOMIC_SEQ_CST );
663	if ( traceHeap() ) {
664	char helpText[64];
665	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
666	"%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize ); // print debug/nodebug
667	} // if
668	#endif // __CFA_DEBUG__
669
670	return addr;
671	} // doMalloc
672
673
674	static inline void doFree( void * addr ) with( heapManager ) {
675	#ifdef __CFA_DEBUG__
676	if ( unlikely( heapManager.heapBegin == 0p ) ) {
677	abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
678	} // if
679	#endif // __CFA_DEBUG__
680
681	Heap.Storage.Header * header;
682	Heap.FreeHeader * freeElem;
683	size_t size, alignment; // not used (see realloc)
684
685	if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
686	#ifdef __STATISTICS__
687	__atomic_add_fetch( &stats.munmap_calls, 1, __ATOMIC_SEQ_CST );
688	__atomic_add_fetch( &stats.munmap_storage_request, header->kind.real.size, __ATOMIC_SEQ_CST );
689	__atomic_add_fetch( &stats.munmap_storage_alloc, size, __ATOMIC_SEQ_CST );
690	#endif // __STATISTICS__
691	if ( munmap( header, size ) == -1 ) {
692	abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
693	"Possible cause is invalid pointer.",
694	addr );
695	} // if
696	} else {
697	#ifdef __CFA_DEBUG__
698	// Set free memory to garbage so subsequent usages might fail.
699	memset( ((Heap.Storage *)header)->data, '\xde', freeElem->blockSize - sizeof( Heap.Storage ) );
700	//Memset( ((Heap.Storage *)header)->data, freeElem->blockSize - sizeof( Heap.Storage ) );
701	#endif // __CFA_DEBUG__
702
703	#ifdef __STATISTICS__
704	__atomic_add_fetch( &stats.free_calls, 1, __ATOMIC_SEQ_CST );
705	__atomic_add_fetch( &stats.free_storage_request, header->kind.real.size, __ATOMIC_SEQ_CST );
706	__atomic_add_fetch( &stats.free_storage_alloc, size, __ATOMIC_SEQ_CST );
707	#endif // __STATISTICS__
708
709	#if BUCKETLOCK == SPINLOCK
710	lock( freeElem->lock __cfaabi_dbg_ctx2 ); // acquire spin lock
711	header->kind.real.next = freeElem->freeList; // push on stack
712	freeElem->freeList = (Heap.Storage *)header;
713	unlock( freeElem->lock ); // release spin lock
714	#else
715	push( freeElem->freeList, (Heap.Storage )header );
716	#endif // BUCKETLOCK
717	} // if
718
719	#ifdef __CFA_DEBUG__
720	__atomic_add_fetch( &allocUnfreed, -size, __ATOMIC_SEQ_CST );
721	if ( traceHeap() ) {
722	char helpText[64];
723	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
724	"Free( %p ) size:%zu\n", addr, size ); // print debug/nodebug
725	} // if
726	#endif // __CFA_DEBUG__
727	} // doFree
728
729
730	static size_t prtFree( Heap & manager ) with( manager ) {
731	size_t total = 0;
732	#ifdef __STATISTICS__
733	__cfaabi_bits_acquire();
734	__cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
735	#endif // __STATISTICS__
736	for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
737	size_t size = freeLists[i].blockSize;
738	#ifdef __STATISTICS__
739	unsigned int N = 0;
740	#endif // __STATISTICS__
741
742	#if BUCKETLOCK == SPINLOCK
743	for ( Heap.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
744	#else
745	for(;;) {
746	// for ( Heap.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
747	// for ( Heap.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
748	// Heap.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
749	// typeof(p) temp = (( p )`next)->top; // FIX ME: direct assignent fails, initialization works`
750	// p = temp;
751	#endif // BUCKETLOCK
752	total += size;
753	#ifdef __STATISTICS__
754	N += 1;
755	#endif // __STATISTICS__
756	} // for
757
758	#ifdef __STATISTICS__
759	__cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
760	if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
761	#endif // __STATISTICS__
762	} // for
763	#ifdef __STATISTICS__
764	__cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
765	__cfaabi_bits_release();
766	#endif // __STATISTICS__
767	return (char )heapEnd - (char )heapBegin - total;
768	} // prtFree
769
770
771	static void ?{}( Heap & manager ) with( manager ) {
772	__page_size = sysconf( _SC_PAGESIZE );
773	__map_prot = PROT_READ \| PROT_WRITE \| PROT_EXEC;
774
775	for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
776	freeLists[i].blockSize = bucketSizes[i];
777	} // for
778
779	#ifdef FASTLOOKUP
780	unsigned int idx = 0;
781	for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
782	if ( i > bucketSizes[idx] ) idx += 1;
783	lookup[i] = idx;
784	} // for
785	#endif // FASTLOOKUP
786
787	if ( ! setMmapStart( malloc_mmap_start() ) ) {
788	abort( "Heap : internal error, mmap start initialization failure." );
789	} // if
790	heapExpand = malloc_expansion();
791
792	char * end = (char *)sbrk( 0 );
793	heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, __page_size ) - end ); // move start of heap to multiple of alignment
794	} // Heap
795
796
797	static void ^?{}( Heap & ) {
798	#ifdef __STATISTICS__
799	if ( traceHeapTerm() ) {
800	printStats();
801	// prtUnfreed() called in heapAppStop()
802	} // if
803	#endif // __STATISTICS__
804	} // ~Heap
805
806
807	static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
808	void memory_startup( void ) {
809	#ifdef __CFA_DEBUG__
810	if ( heapBoot ) { // check for recursion during system boot
811	abort( "boot() : internal error, recursively invoked during system boot." );
812	} // if
813	heapBoot = true;
814	#endif // __CFA_DEBUG__
815
816	//verify( heapManager.heapBegin != 0 );
817	//heapManager{};
818	if ( heapManager.heapBegin == 0p ) heapManager{}; // sanity check
819	} // memory_startup
820
821	static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
822	void memory_shutdown( void ) {
823	^heapManager{};
824	} // memory_shutdown
825
826
827	static inline void * mallocNoStats( size_t size ) { // necessary for malloc statistics
828	verify( heapManager.heapBegin != 0p ); // called before memory_startup ?
829	if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
830
831	#if __SIZEOF_POINTER__ == 8
832	verify( size < ((typeof(size_t))1 << 48) );
833	#endif // __SIZEOF_POINTER__ == 8
834	return doMalloc( size );
835	} // mallocNoStats
836
837
838	static inline void * memalignNoStats( size_t alignment, size_t size ) {
839	if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
840
841	#ifdef __CFA_DEBUG__
842	checkAlign( alignment ); // check alignment
843	#endif // __CFA_DEBUG__
844
845	// if alignment <= default alignment, do normal malloc as two headers are unnecessary
846	if ( unlikely( alignment <= libAlign() ) ) return mallocNoStats( size );
847
848	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
849	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
850	// .-------------v-----------------v----------------v----------,
851	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
852	// `-------------^-----------------^-+--------------^----------'
853	// \|<--------------------------------' offset/align \|<-- alignment boundary
854
855	// subtract libAlign() because it is already the minimum alignment
856	// add sizeof(Storage) for fake header
857	char * addr = (char *)mallocNoStats( size + alignment - libAlign() + sizeof(Heap.Storage) );
858
859	// address in the block of the "next" alignment address
860	char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(Heap.Storage)), alignment );
861
862	// address of header from malloc
863	Heap.Storage.Header * RealHeader = HeaderAddr( addr );
864	RealHeader->kind.real.size = size; // correct size to eliminate above alignment offset
865	// address of fake header * before* the alignment location
866	Heap.Storage.Header * fakeHeader = HeaderAddr( user );
867	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
868	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )RealHeader;
869	// SKULLDUGGERY: odd alignment implies fake header
870	fakeHeader->kind.fake.alignment = MarkAlignmentBit( alignment );
871
872	return user;
873	} // memalignNoStats
874
875
876	//####################### Memory Allocation Routines ####################
877
878
879	extern "C" {
880	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
881	// then malloc() returns a unique pointer value that can later be successfully passed to free().
882	void * malloc( size_t size ) libcfa_public {
883	#ifdef __STATISTICS__
884	if ( likely( size > 0 ) ) {
885	__atomic_add_fetch( &stats.malloc_calls, 1, __ATOMIC_SEQ_CST );
886	__atomic_add_fetch( &stats.malloc_storage_request, size, __ATOMIC_SEQ_CST );
887	} else {
888	__atomic_add_fetch( &stats.malloc_0_calls, 1, __ATOMIC_SEQ_CST );
889	} // if
890	#endif // __STATISTICS__
891
892	return mallocNoStats( size );
893	} // malloc
894
895
896	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
897	void * aalloc( size_t dim, size_t elemSize ) libcfa_public {
898	size_t size = dim * elemSize;
899	#ifdef __STATISTICS__
900	if ( likely( size > 0 ) ) {
901	__atomic_add_fetch( &stats.aalloc_calls, 1, __ATOMIC_SEQ_CST );
902	__atomic_add_fetch( &stats.aalloc_storage_request, size, __ATOMIC_SEQ_CST );
903	} else {
904	__atomic_add_fetch( &stats.aalloc_0_calls, 1, __ATOMIC_SEQ_CST );
905	} // if
906	#endif // __STATISTICS__
907
908	return mallocNoStats( size );
909	} // aalloc
910
911
912	// Same as aalloc() with memory set to zero.
913	void * calloc( size_t dim, size_t elemSize ) libcfa_public {
914	size_t size = dim * elemSize;
915	if ( unlikely( size ) == 0 ) { // 0 BYTE ALLOCATION RETURNS NULL POINTER
916	#ifdef __STATISTICS__
917	__atomic_add_fetch( &stats.calloc_0_calls, 1, __ATOMIC_SEQ_CST );
918	#endif // __STATISTICS__
919	return 0p;
920	} // if
921	#ifdef __STATISTICS__
922	__atomic_add_fetch( &stats.calloc_calls, 1, __ATOMIC_SEQ_CST );
923	__atomic_add_fetch( &stats.calloc_storage_request, dim * elemSize, __ATOMIC_SEQ_CST );
924	#endif // __STATISTICS__
925
926	char * addr = (char *)mallocNoStats( size );
927
928	Heap.Storage.Header * header;
929	Heap.FreeHeader * freeElem;
930	size_t bsize, alignment;
931
932	#ifndef __CFA_DEBUG__
933	bool mapped =
934	#endif // __CFA_DEBUG__
935	headers( "calloc", addr, header, freeElem, bsize, alignment );
936
937	#ifndef __CFA_DEBUG__
938	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
939	if ( ! mapped )
940	#endif // __CFA_DEBUG__
941	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
942	// `-header`-addr `-size
943	memset( addr, '\0', size ); // set to zeros
944
945	MarkZeroFilledBit( header ); // mark as zero fill
946	return addr;
947	} // calloc
948
949
950	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
951	// 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
952	// not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
953	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
954	void * resize( void * oaddr, size_t size ) libcfa_public {
955	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
956	if ( unlikely( size == 0 ) ) { // special cases
957	#ifdef __STATISTICS__
958	__atomic_add_fetch( &stats.resize_0_calls, 1, __ATOMIC_SEQ_CST );
959	#endif // __STATISTICS__
960	free( oaddr );
961	return 0p;
962	} // if
963	#ifdef __STATISTICS__
964	__atomic_add_fetch( &stats.resize_calls, 1, __ATOMIC_SEQ_CST );
965	#endif // __STATISTICS__
966
967	if ( unlikely( oaddr == 0p ) ) {
968	#ifdef __STATISTICS__
969	__atomic_add_fetch( &stats.resize_storage_request, size, __ATOMIC_SEQ_CST );
970	#endif // __STATISTICS__
971	return mallocNoStats( size );
972	} // if
973
974	Heap.Storage.Header * header;
975	Heap.FreeHeader * freeElem;
976	size_t bsize, oalign;
977	headers( "resize", oaddr, header, freeElem, bsize, oalign );
978
979	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
980	// same size, DO NOT preserve STICKY PROPERTIES.
981	if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
982	ClearZeroFillBit( header ); // no alignment and turn off 0 fill
983	header->kind.real.size = size; // reset allocation size
984	return oaddr;
985	} // if
986
987	#ifdef __STATISTICS__
988	__atomic_add_fetch( &stats.resize_storage_request, size, __ATOMIC_SEQ_CST );
989	#endif // __STATISTICS__
990
991	// change size, DO NOT preserve STICKY PROPERTIES.
992	free( oaddr );
993	return mallocNoStats( size ); // create new area
994	} // resize
995
996
997	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
998	// the old and new sizes.
999	void * realloc( void * oaddr, size_t size ) libcfa_public {
1000	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1001	if ( unlikely( size == 0 ) ) { // special cases
1002	#ifdef __STATISTICS__
1003	__atomic_add_fetch( &stats.realloc_0_calls, 1, __ATOMIC_SEQ_CST );
1004	#endif // __STATISTICS__
1005	free( oaddr );
1006	return 0p;
1007	} // if
1008	#ifdef __STATISTICS__
1009	__atomic_add_fetch( &stats.realloc_calls, 1, __ATOMIC_SEQ_CST );
1010	#endif // __STATISTICS__
1011
1012	if ( unlikely( oaddr == 0p ) ) {
1013	#ifdef __STATISTICS__
1014	__atomic_add_fetch( &stats.realloc_storage_request, size, __ATOMIC_SEQ_CST );
1015	#endif // __STATISTICS__
1016	return mallocNoStats( size );
1017	} // if
1018
1019	Heap.Storage.Header * header;
1020	Heap.FreeHeader * freeElem;
1021	size_t bsize, oalign;
1022	headers( "realloc", oaddr, header, freeElem, bsize, oalign );
1023
1024	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1025	size_t osize = header->kind.real.size; // old allocation size
1026	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1027	if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1028	header->kind.real.size = size; // reset allocation size
1029	if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1030	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1031	} // if
1032	return oaddr;
1033	} // if
1034
1035	#ifdef __STATISTICS__
1036	__atomic_add_fetch( &stats.realloc_storage_request, size, __ATOMIC_SEQ_CST );
1037	#endif // __STATISTICS__
1038
1039	// change size and copy old content to new storage
1040
1041	void * naddr;
1042	if ( likely( oalign == libAlign() ) ) { // previous request not aligned ?
1043	naddr = mallocNoStats( size ); // create new area
1044	} else {
1045	naddr = memalignNoStats( oalign, size ); // create new aligned area
1046	} // if
1047
1048	headers( "realloc", naddr, header, freeElem, bsize, oalign );
1049	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1050	free( oaddr );
1051
1052	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1053	MarkZeroFilledBit( header ); // mark new request as zero filled
1054	if ( size > osize ) { // previous request larger ?
1055	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1056	} // if
1057	} // if
1058	return naddr;
1059	} // realloc
1060
1061
1062	// Same as realloc() except the new allocation size is large enough for an array of nelem elements of size elsize.
1063	void * reallocarray( void * oaddr, size_t dim, size_t elemSize ) libcfa_public {
1064	return realloc( oaddr, dim * elemSize );
1065	} // reallocarray
1066
1067
1068	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1069	void * memalign( size_t alignment, size_t size ) libcfa_public {
1070	#ifdef __STATISTICS__
1071	if ( likely( size > 0 ) ) {
1072	__atomic_add_fetch( &stats.memalign_calls, 1, __ATOMIC_SEQ_CST );
1073	__atomic_add_fetch( &stats.memalign_storage_request, size, __ATOMIC_SEQ_CST );
1074	} else {
1075	__atomic_add_fetch( &stats.memalign_0_calls, 1, __ATOMIC_SEQ_CST );
1076	} // if
1077	#endif // __STATISTICS__
1078
1079	return memalignNoStats( alignment, size );
1080	} // memalign
1081
1082
1083	// Same as aalloc() with memory alignment.
1084	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1085	size_t size = dim * elemSize;
1086	#ifdef __STATISTICS__
1087	if ( likely( size > 0 ) ) {
1088	__atomic_add_fetch( &stats.cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1089	__atomic_add_fetch( &stats.cmemalign_storage_request, size, __ATOMIC_SEQ_CST );
1090	} else {
1091	__atomic_add_fetch( &stats.cmemalign_0_calls, 1, __ATOMIC_SEQ_CST );
1092	} // if
1093	#endif // __STATISTICS__
1094
1095	return memalignNoStats( alignment, size );
1096	} // amemalign
1097
1098
1099	// Same as calloc() with memory alignment.
1100	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1101	size_t size = dim * elemSize;
1102	if ( unlikely( size ) == 0 ) { // 0 BYTE ALLOCATION RETURNS NULL POINTER
1103	#ifdef __STATISTICS__
1104	__atomic_add_fetch( &stats.cmemalign_0_calls, 1, __ATOMIC_SEQ_CST );
1105	#endif // __STATISTICS__
1106	return 0p;
1107	} // if
1108	#ifdef __STATISTICS__
1109	__atomic_add_fetch( &stats.cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1110	__atomic_add_fetch( &stats.cmemalign_storage_request, dim * elemSize, __ATOMIC_SEQ_CST );
1111	#endif // __STATISTICS__
1112
1113	char * addr = (char *)memalignNoStats( alignment, size );
1114
1115	Heap.Storage.Header * header;
1116	Heap.FreeHeader * freeElem;
1117	size_t bsize;
1118
1119	#ifndef __CFA_DEBUG__
1120	bool mapped =
1121	#endif // __CFA_DEBUG__
1122	headers( "cmemalign", addr, header, freeElem, bsize, alignment );
1123
1124	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1125	#ifndef __CFA_DEBUG__
1126	if ( ! mapped )
1127	#endif // __CFA_DEBUG__
1128	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1129	// `-header`-addr `-size
1130	memset( addr, '\0', size ); // set to zeros
1131
1132	MarkZeroFilledBit( header ); // mark as zero filled
1133	return addr;
1134	} // cmemalign
1135
1136
1137	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1138	// of alignment. This requirement is universally ignored.
1139	void * aligned_alloc( size_t alignment, size_t size ) libcfa_public {
1140	return memalign( alignment, size );
1141	} // aligned_alloc
1142
1143
1144	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1145	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1146	// is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1147	// free(3).
1148	int posix_memalign( void ** memptr, size_t alignment, size_t size ) libcfa_public {
1149	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) return EINVAL; // check alignment
1150	*memptr = memalign( alignment, size );
1151	return 0;
1152	} // posix_memalign
1153
1154
1155	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1156	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1157	void * valloc( size_t size ) libcfa_public {
1158	return memalign( __page_size, size );
1159	} // valloc
1160
1161
1162	// Same as valloc but rounds size to multiple of page size.
1163	void * pvalloc( size_t size ) libcfa_public {
1164	return memalign( __page_size, ceiling2( size, __page_size ) ); // round size to multiple of page size
1165	} // pvalloc
1166
1167
1168	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1169	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1170	// 0p, no operation is performed.
1171	void free( void * addr ) libcfa_public {
1172	if ( unlikely( addr == 0p ) ) { // special case
1173	#ifdef __STATISTICS__
1174	__atomic_add_fetch( &stats.free_null_calls, 1, __ATOMIC_SEQ_CST );
1175	#endif // __STATISTICS__
1176
1177	// #ifdef __CFA_DEBUG__
1178	// if ( traceHeap() ) {
1179	// #define nullmsg "Free( 0x0 ) size:0\n"
1180	// // Do not debug print free( 0p ), as it can cause recursive entry from sprintf.
1181	// __cfaabi_dbg_write( nullmsg, sizeof(nullmsg) - 1 );
1182	// } // if
1183	// #endif // __CFA_DEBUG__
1184	return;
1185	} // exit
1186
1187	doFree( addr );
1188	} // free
1189
1190
1191	// Returns the alignment of an allocation.
1192	size_t malloc_alignment( void * addr ) libcfa_public {
1193	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1194	Heap.Storage.Header * header = HeaderAddr( addr );
1195	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1196	return ClearAlignmentBit( header ); // clear flag from value
1197	} else {
1198	return libAlign(); // minimum alignment
1199	} // if
1200	} // malloc_alignment
1201
1202
1203	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1204	bool malloc_zero_fill( void * addr ) libcfa_public {
1205	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1206	Heap.Storage.Header * header = HeaderAddr( addr );
1207	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1208	header = RealHeader( header ); // backup from fake to real header
1209	} // if
1210	return ZeroFillBit( header ); // zero filled ?
1211	} // malloc_zero_fill
1212
1213
1214	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1215	size_t malloc_size( void * addr ) libcfa_public {
1216	if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1217	Heap.Storage.Header * header = HeaderAddr( addr );
1218	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1219	header = RealHeader( header ); // backup from fake to real header
1220	} // if
1221	return header->kind.real.size;
1222	} // malloc_size
1223
1224
1225	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1226	// malloc or a related function.
1227	size_t malloc_usable_size( void * addr ) libcfa_public {
1228	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1229	Heap.Storage.Header * header;
1230	Heap.FreeHeader * freeElem;
1231	size_t bsize, alignment;
1232
1233	headers( "malloc_usable_size", addr, header, freeElem, bsize, alignment );
1234	return DataStorage( bsize, addr, header ); // data storage in bucket
1235	} // malloc_usable_size
1236
1237
1238	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1239	void malloc_stats( void ) libcfa_public {
1240	#ifdef __STATISTICS__
1241	printStats();
1242	if ( prtFree() ) prtFree( heapManager );
1243	#endif // __STATISTICS__
1244	} // malloc_stats
1245
1246
1247	// Changes the file descriptor where malloc_stats() writes statistics.
1248	int malloc_stats_fd( int fd __attribute__(( unused )) ) libcfa_public {
1249	#ifdef __STATISTICS__
1250	int temp = stats_fd;
1251	stats_fd = fd;
1252	return temp;
1253	#else
1254	return -1; // unsupported
1255	#endif // __STATISTICS__
1256	} // malloc_stats_fd
1257
1258
1259	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1260	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1261	// malloc).
1262	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) libcfa_public {
1263	if ( options != 0 ) { errno = EINVAL; return -1; }
1264	#ifdef __STATISTICS__
1265	return printStatsXML( stream );
1266	#else
1267	return 0; // unsupported
1268	#endif // __STATISTICS__
1269	} // malloc_info
1270
1271
1272	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1273	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1274	int mallopt( int option, int value ) libcfa_public {
1275	if ( value < 0 ) return 0;
1276	choose( option ) {
1277	case M_TOP_PAD:
1278	heapExpand = ceiling2( value, __page_size );
1279	return 1;
1280	case M_MMAP_THRESHOLD:
1281	if ( setMmapStart( value ) ) return 1;
1282	} // choose
1283	return 0; // error, unsupported
1284	} // mallopt
1285
1286
1287	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1288	int malloc_trim( size_t ) libcfa_public {
1289	return 0; // => impossible to release memory
1290	} // malloc_trim
1291
1292
1293	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1294	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1295	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1296	// result. (The caller must free this memory.)
1297	void * malloc_get_state( void ) libcfa_public {
1298	return 0p; // unsupported
1299	} // malloc_get_state
1300
1301
1302	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1303	// structure pointed to by state.
1304	int malloc_set_state( void * ) libcfa_public {
1305	return 0; // unsupported
1306	} // malloc_set_state
1307
1308
1309	// Sets the amount (bytes) to extend the heap when there is insufficent free storage to service an allocation.
1310	__attribute__((weak)) size_t malloc_expansion() libcfa_public { return __CFA_DEFAULT_HEAP_EXPANSION__; }
1311
1312	// Sets the crossover point between allocations occuring in the sbrk area or separately mmapped.
1313	__attribute__((weak)) size_t malloc_mmap_start() libcfa_public { return __CFA_DEFAULT_MMAP_START__; }
1314
1315	// Amount subtracted to adjust for unfreed program storage (debug only).
1316	__attribute__((weak)) size_t malloc_unfreed() libcfa_public { return __CFA_DEFAULT_HEAP_UNFREED__; }
1317	} // extern "C"
1318
1319
1320	// Must have CFA linkage to overload with C linkage realloc.
1321	void * resize( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1322	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1323	if ( unlikely( size == 0 ) ) { // special cases
1324	#ifdef __STATISTICS__
1325	__atomic_add_fetch( &stats.resize_0_calls, 1, __ATOMIC_SEQ_CST );
1326	#endif // __STATISTICS__
1327	free( oaddr );
1328	return 0p;
1329	} // if
1330
1331	if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1332	#ifdef __CFA_DEBUG__
1333	else checkAlign( nalign ); // check alignment
1334	#endif // __CFA_DEBUG__
1335
1336	if ( unlikely( oaddr == 0p ) ) {
1337	#ifdef __STATISTICS__
1338	__atomic_add_fetch( &stats.resize_calls, 1, __ATOMIC_SEQ_CST );
1339	__atomic_add_fetch( &stats.resize_storage_request, size, __ATOMIC_SEQ_CST );
1340	#endif // __STATISTICS__
1341	return memalignNoStats( nalign, size );
1342	} // if
1343
1344	// Attempt to reuse existing alignment.
1345	Heap.Storage.Header * header = HeaderAddr( oaddr );
1346	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1347	size_t oalign;
1348
1349	if ( unlikely( isFakeHeader ) ) {
1350	oalign = ClearAlignmentBit( header ); // old alignment
1351	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1352	&& ( oalign <= nalign // going down
1353	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1354	) ) {
1355	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1356	Heap.FreeHeader * freeElem;
1357	size_t bsize, oalign;
1358	headers( "resize", oaddr, header, freeElem, bsize, oalign );
1359	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1360
1361	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1362	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1363	ClearZeroFillBit( header ); // turn off 0 fill
1364	header->kind.real.size = size; // reset allocation size
1365	return oaddr;
1366	} // if
1367	} // if
1368	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1369	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1370	return resize( oaddr, size ); // duplicate special case checks
1371	} // if
1372
1373	#ifdef __STATISTICS__
1374	__atomic_add_fetch( &stats.resize_storage_request, size, __ATOMIC_SEQ_CST );
1375	#endif // __STATISTICS__
1376
1377	// change size, DO NOT preserve STICKY PROPERTIES.
1378	free( oaddr );
1379	return memalignNoStats( nalign, size ); // create new aligned area
1380	} // resize
1381
1382
1383	void * realloc( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1384	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1385	if ( unlikely( size == 0 ) ) { // special cases
1386	#ifdef __STATISTICS__
1387	__atomic_add_fetch( &stats.realloc_0_calls, 1, __ATOMIC_SEQ_CST );
1388	#endif // __STATISTICS__
1389	free( oaddr );
1390	return 0p;
1391	} // if
1392
1393	if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1394	#ifdef __CFA_DEBUG__
1395	else checkAlign( nalign ); // check alignment
1396	#endif // __CFA_DEBUG__
1397
1398	if ( unlikely( oaddr == 0p ) ) {
1399	#ifdef __STATISTICS__
1400	__atomic_add_fetch( &stats.realloc_calls, 1, __ATOMIC_SEQ_CST );
1401	__atomic_add_fetch( &stats.realloc_storage_request, size, __ATOMIC_SEQ_CST );
1402	#endif // __STATISTICS__
1403	return memalignNoStats( nalign, size );
1404	} // if
1405
1406	// Attempt to reuse existing alignment.
1407	Heap.Storage.Header * header = HeaderAddr( oaddr );
1408	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1409	size_t oalign;
1410	if ( unlikely( isFakeHeader ) ) {
1411	oalign = ClearAlignmentBit( header ); // old alignment
1412	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1413	&& ( oalign <= nalign // going down
1414	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1415	) ) {
1416	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1417	return realloc( oaddr, size ); // duplicate special case checks
1418	} // if
1419	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1420	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1421	return realloc( oaddr, size ); // duplicate special case checks
1422	} // if
1423
1424	#ifdef __STATISTICS__
1425	__atomic_add_fetch( &stats.realloc_calls, 1, __ATOMIC_SEQ_CST );
1426	__atomic_add_fetch( &stats.realloc_storage_request, size, __ATOMIC_SEQ_CST );
1427	#endif // __STATISTICS__
1428
1429	Heap.FreeHeader * freeElem;
1430	size_t bsize;
1431	headers( "realloc", oaddr, header, freeElem, bsize, oalign );
1432
1433	// change size and copy old content to new storage
1434
1435	size_t osize = header->kind.real.size; // old allocation size
1436	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1437
1438	void * naddr = memalignNoStats( nalign, size ); // create new aligned area
1439
1440	headers( "realloc", naddr, header, freeElem, bsize, oalign );
1441	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1442	free( oaddr );
1443
1444	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1445	MarkZeroFilledBit( header ); // mark new request as zero filled
1446	if ( size > osize ) { // previous request larger ?
1447	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1448	} // if
1449	} // if
1450	return naddr;
1451	} // realloc
1452
1453
1454	// Local Variables: //
1455	// tab-width: 4 //
1456	// compile-command: "cfa -nodebug -O2 heap.cfa" //
1457	// End: //

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

Download in other formats:

Original Format