]> gcc.gnu.org Git - gcc.git/blob - gcc/ggc-page.c
gcse.c (INSN_CUID, [...]): Use gcc_assert and gcc_unreachable.
[gcc.git] / gcc / ggc-page.c
1 /* "Bag-of-pages" garbage collector for the GNU compiler.
2 Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004
3 Free Software Foundation, Inc.
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
11
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
21
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "rtl.h"
28 #include "tm_p.h"
29 #include "toplev.h"
30 #include "flags.h"
31 #include "ggc.h"
32 #include "timevar.h"
33 #include "params.h"
34 #include "tree-flow.h"
35 #ifdef ENABLE_VALGRIND_CHECKING
36 # ifdef HAVE_VALGRIND_MEMCHECK_H
37 # include <valgrind/memcheck.h>
38 # elif defined HAVE_MEMCHECK_H
39 # include <memcheck.h>
40 # else
41 # include <valgrind.h>
42 # endif
43 #else
44 /* Avoid #ifdef:s when we can help it. */
45 #define VALGRIND_DISCARD(x)
46 #endif
47
48 /* Prefer MAP_ANON(YMOUS) to /dev/zero, since we don't need to keep a
49 file open. Prefer either to valloc. */
50 #ifdef HAVE_MMAP_ANON
51 # undef HAVE_MMAP_DEV_ZERO
52
53 # include <sys/mman.h>
54 # ifndef MAP_FAILED
55 # define MAP_FAILED -1
56 # endif
57 # if !defined (MAP_ANONYMOUS) && defined (MAP_ANON)
58 # define MAP_ANONYMOUS MAP_ANON
59 # endif
60 # define USING_MMAP
61
62 #endif
63
64 #ifdef HAVE_MMAP_DEV_ZERO
65
66 # include <sys/mman.h>
67 # ifndef MAP_FAILED
68 # define MAP_FAILED -1
69 # endif
70 # define USING_MMAP
71
72 #endif
73
74 #ifndef USING_MMAP
75 #define USING_MALLOC_PAGE_GROUPS
76 #endif
77
78 /* Stategy:
79
80 This garbage-collecting allocator allocates objects on one of a set
81 of pages. Each page can allocate objects of a single size only;
82 available sizes are powers of two starting at four bytes. The size
83 of an allocation request is rounded up to the next power of two
84 (`order'), and satisfied from the appropriate page.
85
86 Each page is recorded in a page-entry, which also maintains an
87 in-use bitmap of object positions on the page. This allows the
88 allocation state of a particular object to be flipped without
89 touching the page itself.
90
91 Each page-entry also has a context depth, which is used to track
92 pushing and popping of allocation contexts. Only objects allocated
93 in the current (highest-numbered) context may be collected.
94
95 Page entries are arranged in an array of singly-linked lists. The
96 array is indexed by the allocation size, in bits, of the pages on
97 it; i.e. all pages on a list allocate objects of the same size.
98 Pages are ordered on the list such that all non-full pages precede
99 all full pages, with non-full pages arranged in order of decreasing
100 context depth.
101
102 Empty pages (of all orders) are kept on a single page cache list,
103 and are considered first when new pages are required; they are
104 deallocated at the start of the next collection if they haven't
105 been recycled by then. */
106
107 /* Define GGC_DEBUG_LEVEL to print debugging information.
108 0: No debugging output.
109 1: GC statistics only.
110 2: Page-entry allocations/deallocations as well.
111 3: Object allocations as well.
112 4: Object marks as well. */
113 #define GGC_DEBUG_LEVEL (0)
114 \f
115 #ifndef HOST_BITS_PER_PTR
116 #define HOST_BITS_PER_PTR HOST_BITS_PER_LONG
117 #endif
118
119 \f
120 /* A two-level tree is used to look up the page-entry for a given
121 pointer. Two chunks of the pointer's bits are extracted to index
122 the first and second levels of the tree, as follows:
123
124 HOST_PAGE_SIZE_BITS
125 32 | |
126 msb +----------------+----+------+------+ lsb
127 | | |
128 PAGE_L1_BITS |
129 | |
130 PAGE_L2_BITS
131
132 The bottommost HOST_PAGE_SIZE_BITS are ignored, since page-entry
133 pages are aligned on system page boundaries. The next most
134 significant PAGE_L2_BITS and PAGE_L1_BITS are the second and first
135 index values in the lookup table, respectively.
136
137 For 32-bit architectures and the settings below, there are no
138 leftover bits. For architectures with wider pointers, the lookup
139 tree points to a list of pages, which must be scanned to find the
140 correct one. */
141
142 #define PAGE_L1_BITS (8)
143 #define PAGE_L2_BITS (32 - PAGE_L1_BITS - G.lg_pagesize)
144 #define PAGE_L1_SIZE ((size_t) 1 << PAGE_L1_BITS)
145 #define PAGE_L2_SIZE ((size_t) 1 << PAGE_L2_BITS)
146
147 #define LOOKUP_L1(p) \
148 (((size_t) (p) >> (32 - PAGE_L1_BITS)) & ((1 << PAGE_L1_BITS) - 1))
149
150 #define LOOKUP_L2(p) \
151 (((size_t) (p) >> G.lg_pagesize) & ((1 << PAGE_L2_BITS) - 1))
152
153 /* The number of objects per allocation page, for objects on a page of
154 the indicated ORDER. */
155 #define OBJECTS_PER_PAGE(ORDER) objects_per_page_table[ORDER]
156
157 /* The number of objects in P. */
158 #define OBJECTS_IN_PAGE(P) ((P)->bytes / OBJECT_SIZE ((P)->order))
159
160 /* The size of an object on a page of the indicated ORDER. */
161 #define OBJECT_SIZE(ORDER) object_size_table[ORDER]
162
163 /* For speed, we avoid doing a general integer divide to locate the
164 offset in the allocation bitmap, by precalculating numbers M, S
165 such that (O * M) >> S == O / Z (modulo 2^32), for any offset O
166 within the page which is evenly divisible by the object size Z. */
167 #define DIV_MULT(ORDER) inverse_table[ORDER].mult
168 #define DIV_SHIFT(ORDER) inverse_table[ORDER].shift
169 #define OFFSET_TO_BIT(OFFSET, ORDER) \
170 (((OFFSET) * DIV_MULT (ORDER)) >> DIV_SHIFT (ORDER))
171
172 /* The number of extra orders, not corresponding to power-of-two sized
173 objects. */
174
175 #define NUM_EXTRA_ORDERS ARRAY_SIZE (extra_order_size_table)
176
177 #define RTL_SIZE(NSLOTS) \
178 (RTX_HDR_SIZE + (NSLOTS) * sizeof (rtunion))
179
180 #define TREE_EXP_SIZE(OPS) \
181 (sizeof (struct tree_exp) + ((OPS) - 1) * sizeof (tree))
182
183 /* The Ith entry is the maximum size of an object to be stored in the
184 Ith extra order. Adding a new entry to this array is the *only*
185 thing you need to do to add a new special allocation size. */
186
187 static const size_t extra_order_size_table[] = {
188 sizeof (struct stmt_ann_d),
189 sizeof (struct tree_decl),
190 sizeof (struct tree_list),
191 TREE_EXP_SIZE (2),
192 RTL_SIZE (2), /* MEM, PLUS, etc. */
193 RTL_SIZE (9), /* INSN */
194 };
195
196 /* The total number of orders. */
197
198 #define NUM_ORDERS (HOST_BITS_PER_PTR + NUM_EXTRA_ORDERS)
199
200 /* We use this structure to determine the alignment required for
201 allocations. For power-of-two sized allocations, that's not a
202 problem, but it does matter for odd-sized allocations. */
203
204 struct max_alignment {
205 char c;
206 union {
207 HOST_WIDEST_INT i;
208 long double d;
209 } u;
210 };
211
212 /* The biggest alignment required. */
213
214 #define MAX_ALIGNMENT (offsetof (struct max_alignment, u))
215
216 /* Compute the smallest nonnegative number which when added to X gives
217 a multiple of F. */
218
219 #define ROUND_UP_VALUE(x, f) ((f) - 1 - ((f) - 1 + (x)) % (f))
220
221 /* Compute the smallest multiple of F that is >= X. */
222
223 #define ROUND_UP(x, f) (CEIL (x, f) * (f))
224
225 /* The Ith entry is the number of objects on a page or order I. */
226
227 static unsigned objects_per_page_table[NUM_ORDERS];
228
229 /* The Ith entry is the size of an object on a page of order I. */
230
231 static size_t object_size_table[NUM_ORDERS];
232
233 /* The Ith entry is a pair of numbers (mult, shift) such that
234 ((k * mult) >> shift) mod 2^32 == (k / OBJECT_SIZE(I)) mod 2^32,
235 for all k evenly divisible by OBJECT_SIZE(I). */
236
237 static struct
238 {
239 size_t mult;
240 unsigned int shift;
241 }
242 inverse_table[NUM_ORDERS];
243
244 /* A page_entry records the status of an allocation page. This
245 structure is dynamically sized to fit the bitmap in_use_p. */
246 typedef struct page_entry
247 {
248 /* The next page-entry with objects of the same size, or NULL if
249 this is the last page-entry. */
250 struct page_entry *next;
251
252 /* The previous page-entry with objects of the same size, or NULL if
253 this is the first page-entry. The PREV pointer exists solely to
254 keep the cost of ggc_free manageable. */
255 struct page_entry *prev;
256
257 /* The number of bytes allocated. (This will always be a multiple
258 of the host system page size.) */
259 size_t bytes;
260
261 /* The address at which the memory is allocated. */
262 char *page;
263
264 #ifdef USING_MALLOC_PAGE_GROUPS
265 /* Back pointer to the page group this page came from. */
266 struct page_group *group;
267 #endif
268
269 /* This is the index in the by_depth varray where this page table
270 can be found. */
271 unsigned long index_by_depth;
272
273 /* Context depth of this page. */
274 unsigned short context_depth;
275
276 /* The number of free objects remaining on this page. */
277 unsigned short num_free_objects;
278
279 /* A likely candidate for the bit position of a free object for the
280 next allocation from this page. */
281 unsigned short next_bit_hint;
282
283 /* The lg of size of objects allocated from this page. */
284 unsigned char order;
285
286 /* A bit vector indicating whether or not objects are in use. The
287 Nth bit is one if the Nth object on this page is allocated. This
288 array is dynamically sized. */
289 unsigned long in_use_p[1];
290 } page_entry;
291
292 #ifdef USING_MALLOC_PAGE_GROUPS
293 /* A page_group describes a large allocation from malloc, from which
294 we parcel out aligned pages. */
295 typedef struct page_group
296 {
297 /* A linked list of all extant page groups. */
298 struct page_group *next;
299
300 /* The address we received from malloc. */
301 char *allocation;
302
303 /* The size of the block. */
304 size_t alloc_size;
305
306 /* A bitmask of pages in use. */
307 unsigned int in_use;
308 } page_group;
309 #endif
310
311 #if HOST_BITS_PER_PTR <= 32
312
313 /* On 32-bit hosts, we use a two level page table, as pictured above. */
314 typedef page_entry **page_table[PAGE_L1_SIZE];
315
316 #else
317
318 /* On 64-bit hosts, we use the same two level page tables plus a linked
319 list that disambiguates the top 32-bits. There will almost always be
320 exactly one entry in the list. */
321 typedef struct page_table_chain
322 {
323 struct page_table_chain *next;
324 size_t high_bits;
325 page_entry **table[PAGE_L1_SIZE];
326 } *page_table;
327
328 #endif
329
330 /* The rest of the global variables. */
331 static struct globals
332 {
333 /* The Nth element in this array is a page with objects of size 2^N.
334 If there are any pages with free objects, they will be at the
335 head of the list. NULL if there are no page-entries for this
336 object size. */
337 page_entry *pages[NUM_ORDERS];
338
339 /* The Nth element in this array is the last page with objects of
340 size 2^N. NULL if there are no page-entries for this object
341 size. */
342 page_entry *page_tails[NUM_ORDERS];
343
344 /* Lookup table for associating allocation pages with object addresses. */
345 page_table lookup;
346
347 /* The system's page size. */
348 size_t pagesize;
349 size_t lg_pagesize;
350
351 /* Bytes currently allocated. */
352 size_t allocated;
353
354 /* Bytes currently allocated at the end of the last collection. */
355 size_t allocated_last_gc;
356
357 /* Total amount of memory mapped. */
358 size_t bytes_mapped;
359
360 /* Bit N set if any allocations have been done at context depth N. */
361 unsigned long context_depth_allocations;
362
363 /* Bit N set if any collections have been done at context depth N. */
364 unsigned long context_depth_collections;
365
366 /* The current depth in the context stack. */
367 unsigned short context_depth;
368
369 /* A file descriptor open to /dev/zero for reading. */
370 #if defined (HAVE_MMAP_DEV_ZERO)
371 int dev_zero_fd;
372 #endif
373
374 /* A cache of free system pages. */
375 page_entry *free_pages;
376
377 #ifdef USING_MALLOC_PAGE_GROUPS
378 page_group *page_groups;
379 #endif
380
381 /* The file descriptor for debugging output. */
382 FILE *debug_file;
383
384 /* Current number of elements in use in depth below. */
385 unsigned int depth_in_use;
386
387 /* Maximum number of elements that can be used before resizing. */
388 unsigned int depth_max;
389
390 /* Each element of this arry is an index in by_depth where the given
391 depth starts. This structure is indexed by that given depth we
392 are interested in. */
393 unsigned int *depth;
394
395 /* Current number of elements in use in by_depth below. */
396 unsigned int by_depth_in_use;
397
398 /* Maximum number of elements that can be used before resizing. */
399 unsigned int by_depth_max;
400
401 /* Each element of this array is a pointer to a page_entry, all
402 page_entries can be found in here by increasing depth.
403 index_by_depth in the page_entry is the index into this data
404 structure where that page_entry can be found. This is used to
405 speed up finding all page_entries at a particular depth. */
406 page_entry **by_depth;
407
408 /* Each element is a pointer to the saved in_use_p bits, if any,
409 zero otherwise. We allocate them all together, to enable a
410 better runtime data access pattern. */
411 unsigned long **save_in_use;
412
413 #ifdef ENABLE_GC_ALWAYS_COLLECT
414 /* List of free objects to be verified as actually free on the
415 next collection. */
416 struct free_object
417 {
418 void *object;
419 struct free_object *next;
420 } *free_object_list;
421 #endif
422
423 #ifdef GATHER_STATISTICS
424 struct
425 {
426 /* Total memory allocated with ggc_alloc. */
427 unsigned long long total_allocated;
428 /* Total overhead for memory to be allocated with ggc_alloc. */
429 unsigned long long total_overhead;
430
431 /* Total allocations and overhead for sizes less than 32, 64 and 128.
432 These sizes are interesting because they are typical cache line
433 sizes. */
434
435 unsigned long long total_allocated_under32;
436 unsigned long long total_overhead_under32;
437
438 unsigned long long total_allocated_under64;
439 unsigned long long total_overhead_under64;
440
441 unsigned long long total_allocated_under128;
442 unsigned long long total_overhead_under128;
443
444 /* The allocations for each of the allocation orders. */
445 unsigned long long total_allocated_per_order[NUM_ORDERS];
446
447 /* The overhead for each of the allocation orders. */
448 unsigned long long total_overhead_per_order[NUM_ORDERS];
449 } stats;
450 #endif
451 } G;
452
453 /* The size in bytes required to maintain a bitmap for the objects
454 on a page-entry. */
455 #define BITMAP_SIZE(Num_objects) \
456 (CEIL ((Num_objects), HOST_BITS_PER_LONG) * sizeof(long))
457
458 /* Allocate pages in chunks of this size, to throttle calls to memory
459 allocation routines. The first page is used, the rest go onto the
460 free list. This cannot be larger than HOST_BITS_PER_INT for the
461 in_use bitmask for page_group. */
462 #define GGC_QUIRE_SIZE 16
463
464 /* Initial guess as to how many page table entries we might need. */
465 #define INITIAL_PTE_COUNT 128
466 \f
467 static int ggc_allocated_p (const void *);
468 static page_entry *lookup_page_table_entry (const void *);
469 static void set_page_table_entry (void *, page_entry *);
470 #ifdef USING_MMAP
471 static char *alloc_anon (char *, size_t);
472 #endif
473 #ifdef USING_MALLOC_PAGE_GROUPS
474 static size_t page_group_index (char *, char *);
475 static void set_page_group_in_use (page_group *, char *);
476 static void clear_page_group_in_use (page_group *, char *);
477 #endif
478 static struct page_entry * alloc_page (unsigned);
479 static void free_page (struct page_entry *);
480 static void release_pages (void);
481 static void clear_marks (void);
482 static void sweep_pages (void);
483 static void ggc_recalculate_in_use_p (page_entry *);
484 static void compute_inverse (unsigned);
485 static inline void adjust_depth (void);
486 static void move_ptes_to_front (int, int);
487
488 void debug_print_page_list (int);
489 static void push_depth (unsigned int);
490 static void push_by_depth (page_entry *, unsigned long *);
491 struct alloc_zone *rtl_zone = NULL;
492 struct alloc_zone *tree_zone = NULL;
493 struct alloc_zone *garbage_zone = NULL;
494
495 /* Push an entry onto G.depth. */
496
497 inline static void
498 push_depth (unsigned int i)
499 {
500 if (G.depth_in_use >= G.depth_max)
501 {
502 G.depth_max *= 2;
503 G.depth = xrealloc (G.depth, G.depth_max * sizeof (unsigned int));
504 }
505 G.depth[G.depth_in_use++] = i;
506 }
507
508 /* Push an entry onto G.by_depth and G.save_in_use. */
509
510 inline static void
511 push_by_depth (page_entry *p, unsigned long *s)
512 {
513 if (G.by_depth_in_use >= G.by_depth_max)
514 {
515 G.by_depth_max *= 2;
516 G.by_depth = xrealloc (G.by_depth,
517 G.by_depth_max * sizeof (page_entry *));
518 G.save_in_use = xrealloc (G.save_in_use,
519 G.by_depth_max * sizeof (unsigned long *));
520 }
521 G.by_depth[G.by_depth_in_use] = p;
522 G.save_in_use[G.by_depth_in_use++] = s;
523 }
524
525 #if (GCC_VERSION < 3001)
526 #define prefetch(X) ((void) X)
527 #else
528 #define prefetch(X) __builtin_prefetch (X)
529 #endif
530
531 #define save_in_use_p_i(__i) \
532 (G.save_in_use[__i])
533 #define save_in_use_p(__p) \
534 (save_in_use_p_i (__p->index_by_depth))
535
536 /* Returns nonzero if P was allocated in GC'able memory. */
537
538 static inline int
539 ggc_allocated_p (const void *p)
540 {
541 page_entry ***base;
542 size_t L1, L2;
543
544 #if HOST_BITS_PER_PTR <= 32
545 base = &G.lookup[0];
546 #else
547 page_table table = G.lookup;
548 size_t high_bits = (size_t) p & ~ (size_t) 0xffffffff;
549 while (1)
550 {
551 if (table == NULL)
552 return 0;
553 if (table->high_bits == high_bits)
554 break;
555 table = table->next;
556 }
557 base = &table->table[0];
558 #endif
559
560 /* Extract the level 1 and 2 indices. */
561 L1 = LOOKUP_L1 (p);
562 L2 = LOOKUP_L2 (p);
563
564 return base[L1] && base[L1][L2];
565 }
566
567 /* Traverse the page table and find the entry for a page.
568 Die (probably) if the object wasn't allocated via GC. */
569
570 static inline page_entry *
571 lookup_page_table_entry (const void *p)
572 {
573 page_entry ***base;
574 size_t L1, L2;
575
576 #if HOST_BITS_PER_PTR <= 32
577 base = &G.lookup[0];
578 #else
579 page_table table = G.lookup;
580 size_t high_bits = (size_t) p & ~ (size_t) 0xffffffff;
581 while (table->high_bits != high_bits)
582 table = table->next;
583 base = &table->table[0];
584 #endif
585
586 /* Extract the level 1 and 2 indices. */
587 L1 = LOOKUP_L1 (p);
588 L2 = LOOKUP_L2 (p);
589
590 return base[L1][L2];
591 }
592
593 /* Set the page table entry for a page. */
594
595 static void
596 set_page_table_entry (void *p, page_entry *entry)
597 {
598 page_entry ***base;
599 size_t L1, L2;
600
601 #if HOST_BITS_PER_PTR <= 32
602 base = &G.lookup[0];
603 #else
604 page_table table;
605 size_t high_bits = (size_t) p & ~ (size_t) 0xffffffff;
606 for (table = G.lookup; table; table = table->next)
607 if (table->high_bits == high_bits)
608 goto found;
609
610 /* Not found -- allocate a new table. */
611 table = xcalloc (1, sizeof(*table));
612 table->next = G.lookup;
613 table->high_bits = high_bits;
614 G.lookup = table;
615 found:
616 base = &table->table[0];
617 #endif
618
619 /* Extract the level 1 and 2 indices. */
620 L1 = LOOKUP_L1 (p);
621 L2 = LOOKUP_L2 (p);
622
623 if (base[L1] == NULL)
624 base[L1] = xcalloc (PAGE_L2_SIZE, sizeof (page_entry *));
625
626 base[L1][L2] = entry;
627 }
628
629 /* Prints the page-entry for object size ORDER, for debugging. */
630
631 void
632 debug_print_page_list (int order)
633 {
634 page_entry *p;
635 printf ("Head=%p, Tail=%p:\n", (void *) G.pages[order],
636 (void *) G.page_tails[order]);
637 p = G.pages[order];
638 while (p != NULL)
639 {
640 printf ("%p(%1d|%3d) -> ", (void *) p, p->context_depth,
641 p->num_free_objects);
642 p = p->next;
643 }
644 printf ("NULL\n");
645 fflush (stdout);
646 }
647
648 #ifdef USING_MMAP
649 /* Allocate SIZE bytes of anonymous memory, preferably near PREF,
650 (if non-null). The ifdef structure here is intended to cause a
651 compile error unless exactly one of the HAVE_* is defined. */
652
653 static inline char *
654 alloc_anon (char *pref ATTRIBUTE_UNUSED, size_t size)
655 {
656 #ifdef HAVE_MMAP_ANON
657 char *page = mmap (pref, size, PROT_READ | PROT_WRITE,
658 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
659 #endif
660 #ifdef HAVE_MMAP_DEV_ZERO
661 char *page = mmap (pref, size, PROT_READ | PROT_WRITE,
662 MAP_PRIVATE, G.dev_zero_fd, 0);
663 #endif
664
665 if (page == (char *) MAP_FAILED)
666 {
667 perror ("virtual memory exhausted");
668 exit (FATAL_EXIT_CODE);
669 }
670
671 /* Remember that we allocated this memory. */
672 G.bytes_mapped += size;
673
674 /* Pretend we don't have access to the allocated pages. We'll enable
675 access to smaller pieces of the area in ggc_alloc. Discard the
676 handle to avoid handle leak. */
677 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (page, size));
678
679 return page;
680 }
681 #endif
682 #ifdef USING_MALLOC_PAGE_GROUPS
683 /* Compute the index for this page into the page group. */
684
685 static inline size_t
686 page_group_index (char *allocation, char *page)
687 {
688 return (size_t) (page - allocation) >> G.lg_pagesize;
689 }
690
691 /* Set and clear the in_use bit for this page in the page group. */
692
693 static inline void
694 set_page_group_in_use (page_group *group, char *page)
695 {
696 group->in_use |= 1 << page_group_index (group->allocation, page);
697 }
698
699 static inline void
700 clear_page_group_in_use (page_group *group, char *page)
701 {
702 group->in_use &= ~(1 << page_group_index (group->allocation, page));
703 }
704 #endif
705
706 /* Allocate a new page for allocating objects of size 2^ORDER,
707 and return an entry for it. The entry is not added to the
708 appropriate page_table list. */
709
710 static inline struct page_entry *
711 alloc_page (unsigned order)
712 {
713 struct page_entry *entry, *p, **pp;
714 char *page;
715 size_t num_objects;
716 size_t bitmap_size;
717 size_t page_entry_size;
718 size_t entry_size;
719 #ifdef USING_MALLOC_PAGE_GROUPS
720 page_group *group;
721 #endif
722
723 num_objects = OBJECTS_PER_PAGE (order);
724 bitmap_size = BITMAP_SIZE (num_objects + 1);
725 page_entry_size = sizeof (page_entry) - sizeof (long) + bitmap_size;
726 entry_size = num_objects * OBJECT_SIZE (order);
727 if (entry_size < G.pagesize)
728 entry_size = G.pagesize;
729
730 entry = NULL;
731 page = NULL;
732
733 /* Check the list of free pages for one we can use. */
734 for (pp = &G.free_pages, p = *pp; p; pp = &p->next, p = *pp)
735 if (p->bytes == entry_size)
736 break;
737
738 if (p != NULL)
739 {
740 /* Recycle the allocated memory from this page ... */
741 *pp = p->next;
742 page = p->page;
743
744 #ifdef USING_MALLOC_PAGE_GROUPS
745 group = p->group;
746 #endif
747
748 /* ... and, if possible, the page entry itself. */
749 if (p->order == order)
750 {
751 entry = p;
752 memset (entry, 0, page_entry_size);
753 }
754 else
755 free (p);
756 }
757 #ifdef USING_MMAP
758 else if (entry_size == G.pagesize)
759 {
760 /* We want just one page. Allocate a bunch of them and put the
761 extras on the freelist. (Can only do this optimization with
762 mmap for backing store.) */
763 struct page_entry *e, *f = G.free_pages;
764 int i;
765
766 page = alloc_anon (NULL, G.pagesize * GGC_QUIRE_SIZE);
767
768 /* This loop counts down so that the chain will be in ascending
769 memory order. */
770 for (i = GGC_QUIRE_SIZE - 1; i >= 1; i--)
771 {
772 e = xcalloc (1, page_entry_size);
773 e->order = order;
774 e->bytes = G.pagesize;
775 e->page = page + (i << G.lg_pagesize);
776 e->next = f;
777 f = e;
778 }
779
780 G.free_pages = f;
781 }
782 else
783 page = alloc_anon (NULL, entry_size);
784 #endif
785 #ifdef USING_MALLOC_PAGE_GROUPS
786 else
787 {
788 /* Allocate a large block of memory and serve out the aligned
789 pages therein. This results in much less memory wastage
790 than the traditional implementation of valloc. */
791
792 char *allocation, *a, *enda;
793 size_t alloc_size, head_slop, tail_slop;
794 int multiple_pages = (entry_size == G.pagesize);
795
796 if (multiple_pages)
797 alloc_size = GGC_QUIRE_SIZE * G.pagesize;
798 else
799 alloc_size = entry_size + G.pagesize - 1;
800 allocation = xmalloc (alloc_size);
801
802 page = (char *) (((size_t) allocation + G.pagesize - 1) & -G.pagesize);
803 head_slop = page - allocation;
804 if (multiple_pages)
805 tail_slop = ((size_t) allocation + alloc_size) & (G.pagesize - 1);
806 else
807 tail_slop = alloc_size - entry_size - head_slop;
808 enda = allocation + alloc_size - tail_slop;
809
810 /* We allocated N pages, which are likely not aligned, leaving
811 us with N-1 usable pages. We plan to place the page_group
812 structure somewhere in the slop. */
813 if (head_slop >= sizeof (page_group))
814 group = (page_group *)page - 1;
815 else
816 {
817 /* We magically got an aligned allocation. Too bad, we have
818 to waste a page anyway. */
819 if (tail_slop == 0)
820 {
821 enda -= G.pagesize;
822 tail_slop += G.pagesize;
823 }
824 gcc_assert (tail_slop >= sizeof (page_group));
825 group = (page_group *)enda;
826 tail_slop -= sizeof (page_group);
827 }
828
829 /* Remember that we allocated this memory. */
830 group->next = G.page_groups;
831 group->allocation = allocation;
832 group->alloc_size = alloc_size;
833 group->in_use = 0;
834 G.page_groups = group;
835 G.bytes_mapped += alloc_size;
836
837 /* If we allocated multiple pages, put the rest on the free list. */
838 if (multiple_pages)
839 {
840 struct page_entry *e, *f = G.free_pages;
841 for (a = enda - G.pagesize; a != page; a -= G.pagesize)
842 {
843 e = xcalloc (1, page_entry_size);
844 e->order = order;
845 e->bytes = G.pagesize;
846 e->page = a;
847 e->group = group;
848 e->next = f;
849 f = e;
850 }
851 G.free_pages = f;
852 }
853 }
854 #endif
855
856 if (entry == NULL)
857 entry = xcalloc (1, page_entry_size);
858
859 entry->bytes = entry_size;
860 entry->page = page;
861 entry->context_depth = G.context_depth;
862 entry->order = order;
863 entry->num_free_objects = num_objects;
864 entry->next_bit_hint = 1;
865
866 G.context_depth_allocations |= (unsigned long)1 << G.context_depth;
867
868 #ifdef USING_MALLOC_PAGE_GROUPS
869 entry->group = group;
870 set_page_group_in_use (group, page);
871 #endif
872
873 /* Set the one-past-the-end in-use bit. This acts as a sentry as we
874 increment the hint. */
875 entry->in_use_p[num_objects / HOST_BITS_PER_LONG]
876 = (unsigned long) 1 << (num_objects % HOST_BITS_PER_LONG);
877
878 set_page_table_entry (page, entry);
879
880 if (GGC_DEBUG_LEVEL >= 2)
881 fprintf (G.debug_file,
882 "Allocating page at %p, object size=%lu, data %p-%p\n",
883 (void *) entry, (unsigned long) OBJECT_SIZE (order), page,
884 page + entry_size - 1);
885
886 return entry;
887 }
888
889 /* Adjust the size of G.depth so that no index greater than the one
890 used by the top of the G.by_depth is used. */
891
892 static inline void
893 adjust_depth (void)
894 {
895 page_entry *top;
896
897 if (G.by_depth_in_use)
898 {
899 top = G.by_depth[G.by_depth_in_use-1];
900
901 /* Peel back indices in depth that index into by_depth, so that
902 as new elements are added to by_depth, we note the indices
903 of those elements, if they are for new context depths. */
904 while (G.depth_in_use > (size_t)top->context_depth+1)
905 --G.depth_in_use;
906 }
907 }
908
909 /* For a page that is no longer needed, put it on the free page list. */
910
911 static void
912 free_page (page_entry *entry)
913 {
914 if (GGC_DEBUG_LEVEL >= 2)
915 fprintf (G.debug_file,
916 "Deallocating page at %p, data %p-%p\n", (void *) entry,
917 entry->page, entry->page + entry->bytes - 1);
918
919 /* Mark the page as inaccessible. Discard the handle to avoid handle
920 leak. */
921 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (entry->page, entry->bytes));
922
923 set_page_table_entry (entry->page, NULL);
924
925 #ifdef USING_MALLOC_PAGE_GROUPS
926 clear_page_group_in_use (entry->group, entry->page);
927 #endif
928
929 if (G.by_depth_in_use > 1)
930 {
931 page_entry *top = G.by_depth[G.by_depth_in_use-1];
932 int i = entry->index_by_depth;
933
934 /* We cannot free a page from a context deeper than the current
935 one. */
936 gcc_assert (entry->context_depth == top->context_depth);
937
938 /* Put top element into freed slot. */
939 G.by_depth[i] = top;
940 G.save_in_use[i] = G.save_in_use[G.by_depth_in_use-1];
941 top->index_by_depth = i;
942 }
943 --G.by_depth_in_use;
944
945 adjust_depth ();
946
947 entry->next = G.free_pages;
948 G.free_pages = entry;
949 }
950
951 /* Release the free page cache to the system. */
952
953 static void
954 release_pages (void)
955 {
956 #ifdef USING_MMAP
957 page_entry *p, *next;
958 char *start;
959 size_t len;
960
961 /* Gather up adjacent pages so they are unmapped together. */
962 p = G.free_pages;
963
964 while (p)
965 {
966 start = p->page;
967 next = p->next;
968 len = p->bytes;
969 free (p);
970 p = next;
971
972 while (p && p->page == start + len)
973 {
974 next = p->next;
975 len += p->bytes;
976 free (p);
977 p = next;
978 }
979
980 munmap (start, len);
981 G.bytes_mapped -= len;
982 }
983
984 G.free_pages = NULL;
985 #endif
986 #ifdef USING_MALLOC_PAGE_GROUPS
987 page_entry **pp, *p;
988 page_group **gp, *g;
989
990 /* Remove all pages from free page groups from the list. */
991 pp = &G.free_pages;
992 while ((p = *pp) != NULL)
993 if (p->group->in_use == 0)
994 {
995 *pp = p->next;
996 free (p);
997 }
998 else
999 pp = &p->next;
1000
1001 /* Remove all free page groups, and release the storage. */
1002 gp = &G.page_groups;
1003 while ((g = *gp) != NULL)
1004 if (g->in_use == 0)
1005 {
1006 *gp = g->next;
1007 G.bytes_mapped -= g->alloc_size;
1008 free (g->allocation);
1009 }
1010 else
1011 gp = &g->next;
1012 #endif
1013 }
1014
1015 /* This table provides a fast way to determine ceil(log_2(size)) for
1016 allocation requests. The minimum allocation size is eight bytes. */
1017
1018 static unsigned char size_lookup[257] =
1019 {
1020 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4,
1021 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
1022 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
1023 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
1024 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1025 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1026 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1027 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1028 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1029 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1030 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1031 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1032 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1033 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1034 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1035 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1036 8
1037 };
1038
1039 /* Typed allocation function. Does nothing special in this collector. */
1040
1041 void *
1042 ggc_alloc_typed_stat (enum gt_types_enum type ATTRIBUTE_UNUSED, size_t size
1043 MEM_STAT_DECL)
1044 {
1045 return ggc_alloc_stat (size PASS_MEM_STAT);
1046 }
1047
1048 /* Zone allocation function. Does nothing special in this collector. */
1049
1050 void *
1051 ggc_alloc_zone_stat (size_t size, struct alloc_zone *zone ATTRIBUTE_UNUSED
1052 MEM_STAT_DECL)
1053 {
1054 return ggc_alloc_stat (size PASS_MEM_STAT);
1055 }
1056
1057 /* Allocate a chunk of memory of SIZE bytes. Its contents are undefined. */
1058
1059 void *
1060 ggc_alloc_stat (size_t size MEM_STAT_DECL)
1061 {
1062 size_t order, word, bit, object_offset, object_size;
1063 struct page_entry *entry;
1064 void *result;
1065
1066 if (size <= 256)
1067 {
1068 order = size_lookup[size];
1069 object_size = OBJECT_SIZE (order);
1070 }
1071 else
1072 {
1073 order = 9;
1074 while (size > (object_size = OBJECT_SIZE (order)))
1075 order++;
1076 }
1077
1078 /* If there are non-full pages for this size allocation, they are at
1079 the head of the list. */
1080 entry = G.pages[order];
1081
1082 /* If there is no page for this object size, or all pages in this
1083 context are full, allocate a new page. */
1084 if (entry == NULL || entry->num_free_objects == 0)
1085 {
1086 struct page_entry *new_entry;
1087 new_entry = alloc_page (order);
1088
1089 new_entry->index_by_depth = G.by_depth_in_use;
1090 push_by_depth (new_entry, 0);
1091
1092 /* We can skip context depths, if we do, make sure we go all the
1093 way to the new depth. */
1094 while (new_entry->context_depth >= G.depth_in_use)
1095 push_depth (G.by_depth_in_use-1);
1096
1097 /* If this is the only entry, it's also the tail. If it is not
1098 the only entry, then we must update the PREV pointer of the
1099 ENTRY (G.pages[order]) to point to our new page entry. */
1100 if (entry == NULL)
1101 G.page_tails[order] = new_entry;
1102 else
1103 entry->prev = new_entry;
1104
1105 /* Put new pages at the head of the page list. By definition the
1106 entry at the head of the list always has a NULL pointer. */
1107 new_entry->next = entry;
1108 new_entry->prev = NULL;
1109 entry = new_entry;
1110 G.pages[order] = new_entry;
1111
1112 /* For a new page, we know the word and bit positions (in the
1113 in_use bitmap) of the first available object -- they're zero. */
1114 new_entry->next_bit_hint = 1;
1115 word = 0;
1116 bit = 0;
1117 object_offset = 0;
1118 }
1119 else
1120 {
1121 /* First try to use the hint left from the previous allocation
1122 to locate a clear bit in the in-use bitmap. We've made sure
1123 that the one-past-the-end bit is always set, so if the hint
1124 has run over, this test will fail. */
1125 unsigned hint = entry->next_bit_hint;
1126 word = hint / HOST_BITS_PER_LONG;
1127 bit = hint % HOST_BITS_PER_LONG;
1128
1129 /* If the hint didn't work, scan the bitmap from the beginning. */
1130 if ((entry->in_use_p[word] >> bit) & 1)
1131 {
1132 word = bit = 0;
1133 while (~entry->in_use_p[word] == 0)
1134 ++word;
1135 while ((entry->in_use_p[word] >> bit) & 1)
1136 ++bit;
1137 hint = word * HOST_BITS_PER_LONG + bit;
1138 }
1139
1140 /* Next time, try the next bit. */
1141 entry->next_bit_hint = hint + 1;
1142
1143 object_offset = hint * object_size;
1144 }
1145
1146 /* Set the in-use bit. */
1147 entry->in_use_p[word] |= ((unsigned long) 1 << bit);
1148
1149 /* Keep a running total of the number of free objects. If this page
1150 fills up, we may have to move it to the end of the list if the
1151 next page isn't full. If the next page is full, all subsequent
1152 pages are full, so there's no need to move it. */
1153 if (--entry->num_free_objects == 0
1154 && entry->next != NULL
1155 && entry->next->num_free_objects > 0)
1156 {
1157 /* We have a new head for the list. */
1158 G.pages[order] = entry->next;
1159
1160 /* We are moving ENTRY to the end of the page table list.
1161 The new page at the head of the list will have NULL in
1162 its PREV field and ENTRY will have NULL in its NEXT field. */
1163 entry->next->prev = NULL;
1164 entry->next = NULL;
1165
1166 /* Append ENTRY to the tail of the list. */
1167 entry->prev = G.page_tails[order];
1168 G.page_tails[order]->next = entry;
1169 G.page_tails[order] = entry;
1170 }
1171
1172 /* Calculate the object's address. */
1173 result = entry->page + object_offset;
1174 #ifdef GATHER_STATISTICS
1175 ggc_record_overhead (OBJECT_SIZE (order), OBJECT_SIZE (order) - size,
1176 result PASS_MEM_STAT);
1177 #endif
1178
1179 #ifdef ENABLE_GC_CHECKING
1180 /* Keep poisoning-by-writing-0xaf the object, in an attempt to keep the
1181 exact same semantics in presence of memory bugs, regardless of
1182 ENABLE_VALGRIND_CHECKING. We override this request below. Drop the
1183 handle to avoid handle leak. */
1184 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (result, object_size));
1185
1186 /* `Poison' the entire allocated object, including any padding at
1187 the end. */
1188 memset (result, 0xaf, object_size);
1189
1190 /* Make the bytes after the end of the object unaccessible. Discard the
1191 handle to avoid handle leak. */
1192 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS ((char *) result + size,
1193 object_size - size));
1194 #endif
1195
1196 /* Tell Valgrind that the memory is there, but its content isn't
1197 defined. The bytes at the end of the object are still marked
1198 unaccessible. */
1199 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (result, size));
1200
1201 /* Keep track of how many bytes are being allocated. This
1202 information is used in deciding when to collect. */
1203 G.allocated += object_size;
1204
1205 #ifdef GATHER_STATISTICS
1206 {
1207 size_t overhead = object_size - size;
1208
1209 G.stats.total_overhead += overhead;
1210 G.stats.total_allocated += object_size;
1211 G.stats.total_overhead_per_order[order] += overhead;
1212 G.stats.total_allocated_per_order[order] += object_size;
1213
1214 if (size <= 32)
1215 {
1216 G.stats.total_overhead_under32 += overhead;
1217 G.stats.total_allocated_under32 += object_size;
1218 }
1219 if (size <= 64)
1220 {
1221 G.stats.total_overhead_under64 += overhead;
1222 G.stats.total_allocated_under64 += object_size;
1223 }
1224 if (size <= 128)
1225 {
1226 G.stats.total_overhead_under128 += overhead;
1227 G.stats.total_allocated_under128 += object_size;
1228 }
1229 }
1230 #endif
1231
1232 if (GGC_DEBUG_LEVEL >= 3)
1233 fprintf (G.debug_file,
1234 "Allocating object, requested size=%lu, actual=%lu at %p on %p\n",
1235 (unsigned long) size, (unsigned long) object_size, result,
1236 (void *) entry);
1237
1238 return result;
1239 }
1240
1241 /* If P is not marked, marks it and return false. Otherwise return true.
1242 P must have been allocated by the GC allocator; it mustn't point to
1243 static objects, stack variables, or memory allocated with malloc. */
1244
1245 int
1246 ggc_set_mark (const void *p)
1247 {
1248 page_entry *entry;
1249 unsigned bit, word;
1250 unsigned long mask;
1251
1252 /* Look up the page on which the object is alloced. If the object
1253 wasn't allocated by the collector, we'll probably die. */
1254 entry = lookup_page_table_entry (p);
1255 gcc_assert (entry);
1256
1257 /* Calculate the index of the object on the page; this is its bit
1258 position in the in_use_p bitmap. */
1259 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order);
1260 word = bit / HOST_BITS_PER_LONG;
1261 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG);
1262
1263 /* If the bit was previously set, skip it. */
1264 if (entry->in_use_p[word] & mask)
1265 return 1;
1266
1267 /* Otherwise set it, and decrement the free object count. */
1268 entry->in_use_p[word] |= mask;
1269 entry->num_free_objects -= 1;
1270
1271 if (GGC_DEBUG_LEVEL >= 4)
1272 fprintf (G.debug_file, "Marking %p\n", p);
1273
1274 return 0;
1275 }
1276
1277 /* Return 1 if P has been marked, zero otherwise.
1278 P must have been allocated by the GC allocator; it mustn't point to
1279 static objects, stack variables, or memory allocated with malloc. */
1280
1281 int
1282 ggc_marked_p (const void *p)
1283 {
1284 page_entry *entry;
1285 unsigned bit, word;
1286 unsigned long mask;
1287
1288 /* Look up the page on which the object is alloced. If the object
1289 wasn't allocated by the collector, we'll probably die. */
1290 entry = lookup_page_table_entry (p);
1291 gcc_assert (entry);
1292
1293 /* Calculate the index of the object on the page; this is its bit
1294 position in the in_use_p bitmap. */
1295 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order);
1296 word = bit / HOST_BITS_PER_LONG;
1297 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG);
1298
1299 return (entry->in_use_p[word] & mask) != 0;
1300 }
1301
1302 /* Return the size of the gc-able object P. */
1303
1304 size_t
1305 ggc_get_size (const void *p)
1306 {
1307 page_entry *pe = lookup_page_table_entry (p);
1308 return OBJECT_SIZE (pe->order);
1309 }
1310
1311 /* Release the memory for object P. */
1312
1313 void
1314 ggc_free (void *p)
1315 {
1316 page_entry *pe = lookup_page_table_entry (p);
1317 size_t order = pe->order;
1318 size_t size = OBJECT_SIZE (order);
1319
1320 #ifdef GATHER_STATISTICS
1321 ggc_free_overhead (p);
1322 #endif
1323
1324 if (GGC_DEBUG_LEVEL >= 3)
1325 fprintf (G.debug_file,
1326 "Freeing object, actual size=%lu, at %p on %p\n",
1327 (unsigned long) size, p, (void *) pe);
1328
1329 #ifdef ENABLE_GC_CHECKING
1330 /* Poison the data, to indicate the data is garbage. */
1331 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (p, size));
1332 memset (p, 0xa5, size);
1333 #endif
1334 /* Let valgrind know the object is free. */
1335 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (p, size));
1336
1337 #ifdef ENABLE_GC_ALWAYS_COLLECT
1338 /* In the completely-anal-checking mode, we do *not* immediately free
1339 the data, but instead verify that the data is *actually* not
1340 reachable the next time we collect. */
1341 {
1342 struct free_object *fo = xmalloc (sizeof (struct free_object));
1343 fo->object = p;
1344 fo->next = G.free_object_list;
1345 G.free_object_list = fo;
1346 }
1347 #else
1348 {
1349 unsigned int bit_offset, word, bit;
1350
1351 G.allocated -= size;
1352
1353 /* Mark the object not-in-use. */
1354 bit_offset = OFFSET_TO_BIT (((const char *) p) - pe->page, order);
1355 word = bit_offset / HOST_BITS_PER_LONG;
1356 bit = bit_offset % HOST_BITS_PER_LONG;
1357 pe->in_use_p[word] &= ~(1UL << bit);
1358
1359 if (pe->num_free_objects++ == 0)
1360 {
1361 page_entry *p, *q;
1362
1363 /* If the page is completely full, then it's supposed to
1364 be after all pages that aren't. Since we've freed one
1365 object from a page that was full, we need to move the
1366 page to the head of the list.
1367
1368 PE is the node we want to move. Q is the previous node
1369 and P is the next node in the list. */
1370 q = pe->prev;
1371 if (q && q->num_free_objects == 0)
1372 {
1373 p = pe->next;
1374
1375 q->next = p;
1376
1377 /* If PE was at the end of the list, then Q becomes the
1378 new end of the list. If PE was not the end of the
1379 list, then we need to update the PREV field for P. */
1380 if (!p)
1381 G.page_tails[order] = q;
1382 else
1383 p->prev = q;
1384
1385 /* Move PE to the head of the list. */
1386 pe->next = G.pages[order];
1387 pe->prev = NULL;
1388 G.pages[order]->prev = pe;
1389 G.pages[order] = pe;
1390 }
1391
1392 /* Reset the hint bit to point to the only free object. */
1393 pe->next_bit_hint = bit_offset;
1394 }
1395 }
1396 #endif
1397 }
1398 \f
1399 /* Subroutine of init_ggc which computes the pair of numbers used to
1400 perform division by OBJECT_SIZE (order) and fills in inverse_table[].
1401
1402 This algorithm is taken from Granlund and Montgomery's paper
1403 "Division by Invariant Integers using Multiplication"
1404 (Proc. SIGPLAN PLDI, 1994), section 9 (Exact division by
1405 constants). */
1406
1407 static void
1408 compute_inverse (unsigned order)
1409 {
1410 size_t size, inv;
1411 unsigned int e;
1412
1413 size = OBJECT_SIZE (order);
1414 e = 0;
1415 while (size % 2 == 0)
1416 {
1417 e++;
1418 size >>= 1;
1419 }
1420
1421 inv = size;
1422 while (inv * size != 1)
1423 inv = inv * (2 - inv*size);
1424
1425 DIV_MULT (order) = inv;
1426 DIV_SHIFT (order) = e;
1427 }
1428
1429 /* Initialize the ggc-mmap allocator. */
1430 void
1431 init_ggc (void)
1432 {
1433 unsigned order;
1434
1435 G.pagesize = getpagesize();
1436 G.lg_pagesize = exact_log2 (G.pagesize);
1437
1438 #ifdef HAVE_MMAP_DEV_ZERO
1439 G.dev_zero_fd = open ("/dev/zero", O_RDONLY);
1440 if (G.dev_zero_fd == -1)
1441 internal_error ("open /dev/zero: %m");
1442 #endif
1443
1444 #if 0
1445 G.debug_file = fopen ("ggc-mmap.debug", "w");
1446 #else
1447 G.debug_file = stdout;
1448 #endif
1449
1450 #ifdef USING_MMAP
1451 /* StunOS has an amazing off-by-one error for the first mmap allocation
1452 after fiddling with RLIMIT_STACK. The result, as hard as it is to
1453 believe, is an unaligned page allocation, which would cause us to
1454 hork badly if we tried to use it. */
1455 {
1456 char *p = alloc_anon (NULL, G.pagesize);
1457 struct page_entry *e;
1458 if ((size_t)p & (G.pagesize - 1))
1459 {
1460 /* How losing. Discard this one and try another. If we still
1461 can't get something useful, give up. */
1462
1463 p = alloc_anon (NULL, G.pagesize);
1464 gcc_assert (!((size_t)p & (G.pagesize - 1)));
1465 }
1466
1467 /* We have a good page, might as well hold onto it... */
1468 e = xcalloc (1, sizeof (struct page_entry));
1469 e->bytes = G.pagesize;
1470 e->page = p;
1471 e->next = G.free_pages;
1472 G.free_pages = e;
1473 }
1474 #endif
1475
1476 /* Initialize the object size table. */
1477 for (order = 0; order < HOST_BITS_PER_PTR; ++order)
1478 object_size_table[order] = (size_t) 1 << order;
1479 for (order = HOST_BITS_PER_PTR; order < NUM_ORDERS; ++order)
1480 {
1481 size_t s = extra_order_size_table[order - HOST_BITS_PER_PTR];
1482
1483 /* If S is not a multiple of the MAX_ALIGNMENT, then round it up
1484 so that we're sure of getting aligned memory. */
1485 s = ROUND_UP (s, MAX_ALIGNMENT);
1486 object_size_table[order] = s;
1487 }
1488
1489 /* Initialize the objects-per-page and inverse tables. */
1490 for (order = 0; order < NUM_ORDERS; ++order)
1491 {
1492 objects_per_page_table[order] = G.pagesize / OBJECT_SIZE (order);
1493 if (objects_per_page_table[order] == 0)
1494 objects_per_page_table[order] = 1;
1495 compute_inverse (order);
1496 }
1497
1498 /* Reset the size_lookup array to put appropriately sized objects in
1499 the special orders. All objects bigger than the previous power
1500 of two, but no greater than the special size, should go in the
1501 new order. */
1502 for (order = HOST_BITS_PER_PTR; order < NUM_ORDERS; ++order)
1503 {
1504 int o;
1505 int i;
1506
1507 o = size_lookup[OBJECT_SIZE (order)];
1508 for (i = OBJECT_SIZE (order); size_lookup [i] == o; --i)
1509 size_lookup[i] = order;
1510 }
1511
1512 G.depth_in_use = 0;
1513 G.depth_max = 10;
1514 G.depth = xmalloc (G.depth_max * sizeof (unsigned int));
1515
1516 G.by_depth_in_use = 0;
1517 G.by_depth_max = INITIAL_PTE_COUNT;
1518 G.by_depth = xmalloc (G.by_depth_max * sizeof (page_entry *));
1519 G.save_in_use = xmalloc (G.by_depth_max * sizeof (unsigned long *));
1520 }
1521
1522 /* Start a new GGC zone. */
1523
1524 struct alloc_zone *
1525 new_ggc_zone (const char *name ATTRIBUTE_UNUSED)
1526 {
1527 return NULL;
1528 }
1529
1530 /* Destroy a GGC zone. */
1531 void
1532 destroy_ggc_zone (struct alloc_zone *zone ATTRIBUTE_UNUSED)
1533 {
1534 }
1535
1536 /* Increment the `GC context'. Objects allocated in an outer context
1537 are never freed, eliminating the need to register their roots. */
1538
1539 void
1540 ggc_push_context (void)
1541 {
1542 ++G.context_depth;
1543
1544 /* Die on wrap. */
1545 gcc_assert (G.context_depth < HOST_BITS_PER_LONG);
1546 }
1547
1548 /* Merge the SAVE_IN_USE_P and IN_USE_P arrays in P so that IN_USE_P
1549 reflects reality. Recalculate NUM_FREE_OBJECTS as well. */
1550
1551 static void
1552 ggc_recalculate_in_use_p (page_entry *p)
1553 {
1554 unsigned int i;
1555 size_t num_objects;
1556
1557 /* Because the past-the-end bit in in_use_p is always set, we
1558 pretend there is one additional object. */
1559 num_objects = OBJECTS_IN_PAGE (p) + 1;
1560
1561 /* Reset the free object count. */
1562 p->num_free_objects = num_objects;
1563
1564 /* Combine the IN_USE_P and SAVE_IN_USE_P arrays. */
1565 for (i = 0;
1566 i < CEIL (BITMAP_SIZE (num_objects),
1567 sizeof (*p->in_use_p));
1568 ++i)
1569 {
1570 unsigned long j;
1571
1572 /* Something is in use if it is marked, or if it was in use in a
1573 context further down the context stack. */
1574 p->in_use_p[i] |= save_in_use_p (p)[i];
1575
1576 /* Decrement the free object count for every object allocated. */
1577 for (j = p->in_use_p[i]; j; j >>= 1)
1578 p->num_free_objects -= (j & 1);
1579 }
1580
1581 gcc_assert (p->num_free_objects < num_objects);
1582 }
1583
1584 /* Decrement the `GC context'. All objects allocated since the
1585 previous ggc_push_context are migrated to the outer context. */
1586
1587 void
1588 ggc_pop_context (void)
1589 {
1590 unsigned long omask;
1591 unsigned int depth, i, e;
1592 #ifdef ENABLE_CHECKING
1593 unsigned int order;
1594 #endif
1595
1596 depth = --G.context_depth;
1597 omask = (unsigned long)1 << (depth + 1);
1598
1599 if (!((G.context_depth_allocations | G.context_depth_collections) & omask))
1600 return;
1601
1602 G.context_depth_allocations |= (G.context_depth_allocations & omask) >> 1;
1603 G.context_depth_allocations &= omask - 1;
1604 G.context_depth_collections &= omask - 1;
1605
1606 /* The G.depth array is shortened so that the last index is the
1607 context_depth of the top element of by_depth. */
1608 if (depth+1 < G.depth_in_use)
1609 e = G.depth[depth+1];
1610 else
1611 e = G.by_depth_in_use;
1612
1613 /* We might not have any PTEs of depth depth. */
1614 if (depth < G.depth_in_use)
1615 {
1616
1617 /* First we go through all the pages at depth depth to
1618 recalculate the in use bits. */
1619 for (i = G.depth[depth]; i < e; ++i)
1620 {
1621 page_entry *p = G.by_depth[i];
1622
1623 /* Check that all of the pages really are at the depth that
1624 we expect. */
1625 gcc_assert (p->context_depth == depth);
1626 gcc_assert (p->index_by_depth == i);
1627
1628 prefetch (&save_in_use_p_i (i+8));
1629 prefetch (&save_in_use_p_i (i+16));
1630 if (save_in_use_p_i (i))
1631 {
1632 p = G.by_depth[i];
1633 ggc_recalculate_in_use_p (p);
1634 free (save_in_use_p_i (i));
1635 save_in_use_p_i (i) = 0;
1636 }
1637 }
1638 }
1639
1640 /* Then, we reset all page_entries with a depth greater than depth
1641 to be at depth. */
1642 for (i = e; i < G.by_depth_in_use; ++i)
1643 {
1644 page_entry *p = G.by_depth[i];
1645
1646 /* Check that all of the pages really are at the depth we
1647 expect. */
1648 gcc_assert (p->context_depth > depth);
1649 gcc_assert (p->index_by_depth == i);
1650 p->context_depth = depth;
1651 }
1652
1653 adjust_depth ();
1654
1655 #ifdef ENABLE_CHECKING
1656 for (order = 2; order < NUM_ORDERS; order++)
1657 {
1658 page_entry *p;
1659
1660 for (p = G.pages[order]; p != NULL; p = p->next)
1661 gcc_assert (p->context_depth < depth ||
1662 (p->context_depth == depth && !save_in_use_p (p)));
1663 }
1664 #endif
1665 }
1666 \f
1667 /* Unmark all objects. */
1668
1669 static void
1670 clear_marks (void)
1671 {
1672 unsigned order;
1673
1674 for (order = 2; order < NUM_ORDERS; order++)
1675 {
1676 page_entry *p;
1677
1678 for (p = G.pages[order]; p != NULL; p = p->next)
1679 {
1680 size_t num_objects = OBJECTS_IN_PAGE (p);
1681 size_t bitmap_size = BITMAP_SIZE (num_objects + 1);
1682
1683 /* The data should be page-aligned. */
1684 gcc_assert (!((size_t) p->page & (G.pagesize - 1)));
1685
1686 /* Pages that aren't in the topmost context are not collected;
1687 nevertheless, we need their in-use bit vectors to store GC
1688 marks. So, back them up first. */
1689 if (p->context_depth < G.context_depth)
1690 {
1691 if (! save_in_use_p (p))
1692 save_in_use_p (p) = xmalloc (bitmap_size);
1693 memcpy (save_in_use_p (p), p->in_use_p, bitmap_size);
1694 }
1695
1696 /* Reset reset the number of free objects and clear the
1697 in-use bits. These will be adjusted by mark_obj. */
1698 p->num_free_objects = num_objects;
1699 memset (p->in_use_p, 0, bitmap_size);
1700
1701 /* Make sure the one-past-the-end bit is always set. */
1702 p->in_use_p[num_objects / HOST_BITS_PER_LONG]
1703 = ((unsigned long) 1 << (num_objects % HOST_BITS_PER_LONG));
1704 }
1705 }
1706 }
1707
1708 /* Free all empty pages. Partially empty pages need no attention
1709 because the `mark' bit doubles as an `unused' bit. */
1710
1711 static void
1712 sweep_pages (void)
1713 {
1714 unsigned order;
1715
1716 for (order = 2; order < NUM_ORDERS; order++)
1717 {
1718 /* The last page-entry to consider, regardless of entries
1719 placed at the end of the list. */
1720 page_entry * const last = G.page_tails[order];
1721
1722 size_t num_objects;
1723 size_t live_objects;
1724 page_entry *p, *previous;
1725 int done;
1726
1727 p = G.pages[order];
1728 if (p == NULL)
1729 continue;
1730
1731 previous = NULL;
1732 do
1733 {
1734 page_entry *next = p->next;
1735
1736 /* Loop until all entries have been examined. */
1737 done = (p == last);
1738
1739 num_objects = OBJECTS_IN_PAGE (p);
1740
1741 /* Add all live objects on this page to the count of
1742 allocated memory. */
1743 live_objects = num_objects - p->num_free_objects;
1744
1745 G.allocated += OBJECT_SIZE (order) * live_objects;
1746
1747 /* Only objects on pages in the topmost context should get
1748 collected. */
1749 if (p->context_depth < G.context_depth)
1750 ;
1751
1752 /* Remove the page if it's empty. */
1753 else if (live_objects == 0)
1754 {
1755 /* If P was the first page in the list, then NEXT
1756 becomes the new first page in the list, otherwise
1757 splice P out of the forward pointers. */
1758 if (! previous)
1759 G.pages[order] = next;
1760 else
1761 previous->next = next;
1762
1763 /* Splice P out of the back pointers too. */
1764 if (next)
1765 next->prev = previous;
1766
1767 /* Are we removing the last element? */
1768 if (p == G.page_tails[order])
1769 G.page_tails[order] = previous;
1770 free_page (p);
1771 p = previous;
1772 }
1773
1774 /* If the page is full, move it to the end. */
1775 else if (p->num_free_objects == 0)
1776 {
1777 /* Don't move it if it's already at the end. */
1778 if (p != G.page_tails[order])
1779 {
1780 /* Move p to the end of the list. */
1781 p->next = NULL;
1782 p->prev = G.page_tails[order];
1783 G.page_tails[order]->next = p;
1784
1785 /* Update the tail pointer... */
1786 G.page_tails[order] = p;
1787
1788 /* ... and the head pointer, if necessary. */
1789 if (! previous)
1790 G.pages[order] = next;
1791 else
1792 previous->next = next;
1793
1794 /* And update the backpointer in NEXT if necessary. */
1795 if (next)
1796 next->prev = previous;
1797
1798 p = previous;
1799 }
1800 }
1801
1802 /* If we've fallen through to here, it's a page in the
1803 topmost context that is neither full nor empty. Such a
1804 page must precede pages at lesser context depth in the
1805 list, so move it to the head. */
1806 else if (p != G.pages[order])
1807 {
1808 previous->next = p->next;
1809
1810 /* Update the backchain in the next node if it exists. */
1811 if (p->next)
1812 p->next->prev = previous;
1813
1814 /* Move P to the head of the list. */
1815 p->next = G.pages[order];
1816 p->prev = NULL;
1817 G.pages[order]->prev = p;
1818
1819 /* Update the head pointer. */
1820 G.pages[order] = p;
1821
1822 /* Are we moving the last element? */
1823 if (G.page_tails[order] == p)
1824 G.page_tails[order] = previous;
1825 p = previous;
1826 }
1827
1828 previous = p;
1829 p = next;
1830 }
1831 while (! done);
1832
1833 /* Now, restore the in_use_p vectors for any pages from contexts
1834 other than the current one. */
1835 for (p = G.pages[order]; p; p = p->next)
1836 if (p->context_depth != G.context_depth)
1837 ggc_recalculate_in_use_p (p);
1838 }
1839 }
1840
1841 #ifdef ENABLE_GC_CHECKING
1842 /* Clobber all free objects. */
1843
1844 static void
1845 poison_pages (void)
1846 {
1847 unsigned order;
1848
1849 for (order = 2; order < NUM_ORDERS; order++)
1850 {
1851 size_t size = OBJECT_SIZE (order);
1852 page_entry *p;
1853
1854 for (p = G.pages[order]; p != NULL; p = p->next)
1855 {
1856 size_t num_objects;
1857 size_t i;
1858
1859 if (p->context_depth != G.context_depth)
1860 /* Since we don't do any collection for pages in pushed
1861 contexts, there's no need to do any poisoning. And
1862 besides, the IN_USE_P array isn't valid until we pop
1863 contexts. */
1864 continue;
1865
1866 num_objects = OBJECTS_IN_PAGE (p);
1867 for (i = 0; i < num_objects; i++)
1868 {
1869 size_t word, bit;
1870 word = i / HOST_BITS_PER_LONG;
1871 bit = i % HOST_BITS_PER_LONG;
1872 if (((p->in_use_p[word] >> bit) & 1) == 0)
1873 {
1874 char *object = p->page + i * size;
1875
1876 /* Keep poison-by-write when we expect to use Valgrind,
1877 so the exact same memory semantics is kept, in case
1878 there are memory errors. We override this request
1879 below. */
1880 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (object, size));
1881 memset (object, 0xa5, size);
1882
1883 /* Drop the handle to avoid handle leak. */
1884 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (object, size));
1885 }
1886 }
1887 }
1888 }
1889 }
1890 #else
1891 #define poison_pages()
1892 #endif
1893
1894 #ifdef ENABLE_GC_ALWAYS_COLLECT
1895 /* Validate that the reportedly free objects actually are. */
1896
1897 static void
1898 validate_free_objects (void)
1899 {
1900 struct free_object *f, *next, *still_free = NULL;
1901
1902 for (f = G.free_object_list; f ; f = next)
1903 {
1904 page_entry *pe = lookup_page_table_entry (f->object);
1905 size_t bit, word;
1906
1907 bit = OFFSET_TO_BIT ((char *)f->object - pe->page, pe->order);
1908 word = bit / HOST_BITS_PER_LONG;
1909 bit = bit % HOST_BITS_PER_LONG;
1910 next = f->next;
1911
1912 /* Make certain it isn't visible from any root. Notice that we
1913 do this check before sweep_pages merges save_in_use_p. */
1914 gcc_assert (!(pe->in_use_p[word] & (1UL << bit)));
1915
1916 /* If the object comes from an outer context, then retain the
1917 free_object entry, so that we can verify that the address
1918 isn't live on the stack in some outer context. */
1919 if (pe->context_depth != G.context_depth)
1920 {
1921 f->next = still_free;
1922 still_free = f;
1923 }
1924 else
1925 free (f);
1926 }
1927
1928 G.free_object_list = still_free;
1929 }
1930 #else
1931 #define validate_free_objects()
1932 #endif
1933
1934 /* Top level mark-and-sweep routine. */
1935
1936 void
1937 ggc_collect (void)
1938 {
1939 /* Avoid frequent unnecessary work by skipping collection if the
1940 total allocations haven't expanded much since the last
1941 collection. */
1942 float allocated_last_gc =
1943 MAX (G.allocated_last_gc, (size_t)PARAM_VALUE (GGC_MIN_HEAPSIZE) * 1024);
1944
1945 float min_expand = allocated_last_gc * PARAM_VALUE (GGC_MIN_EXPAND) / 100;
1946
1947 if (G.allocated < allocated_last_gc + min_expand && !ggc_force_collect)
1948 return;
1949
1950 timevar_push (TV_GC);
1951 if (!quiet_flag)
1952 fprintf (stderr, " {GC %luk -> ", (unsigned long) G.allocated / 1024);
1953 if (GGC_DEBUG_LEVEL >= 2)
1954 fprintf (G.debug_file, "BEGIN COLLECTING\n");
1955
1956 /* Zero the total allocated bytes. This will be recalculated in the
1957 sweep phase. */
1958 G.allocated = 0;
1959
1960 /* Release the pages we freed the last time we collected, but didn't
1961 reuse in the interim. */
1962 release_pages ();
1963
1964 /* Indicate that we've seen collections at this context depth. */
1965 G.context_depth_collections = ((unsigned long)1 << (G.context_depth + 1)) - 1;
1966
1967 clear_marks ();
1968 ggc_mark_roots ();
1969 #ifdef GATHER_STATISTICS
1970 ggc_prune_overhead_list ();
1971 #endif
1972 poison_pages ();
1973 validate_free_objects ();
1974 sweep_pages ();
1975
1976 G.allocated_last_gc = G.allocated;
1977
1978 timevar_pop (TV_GC);
1979
1980 if (!quiet_flag)
1981 fprintf (stderr, "%luk}", (unsigned long) G.allocated / 1024);
1982 if (GGC_DEBUG_LEVEL >= 2)
1983 fprintf (G.debug_file, "END COLLECTING\n");
1984 }
1985
1986 /* Print allocation statistics. */
1987 #define SCALE(x) ((unsigned long) ((x) < 1024*10 \
1988 ? (x) \
1989 : ((x) < 1024*1024*10 \
1990 ? (x) / 1024 \
1991 : (x) / (1024*1024))))
1992 #define STAT_LABEL(x) ((x) < 1024*10 ? ' ' : ((x) < 1024*1024*10 ? 'k' : 'M'))
1993
1994 void
1995 ggc_print_statistics (void)
1996 {
1997 struct ggc_statistics stats;
1998 unsigned int i;
1999 size_t total_overhead = 0;
2000
2001 /* Clear the statistics. */
2002 memset (&stats, 0, sizeof (stats));
2003
2004 /* Make sure collection will really occur. */
2005 G.allocated_last_gc = 0;
2006
2007 /* Collect and print the statistics common across collectors. */
2008 ggc_print_common_statistics (stderr, &stats);
2009
2010 /* Release free pages so that we will not count the bytes allocated
2011 there as part of the total allocated memory. */
2012 release_pages ();
2013
2014 /* Collect some information about the various sizes of
2015 allocation. */
2016 fprintf (stderr,
2017 "Memory still allocated at the end of the compilation process\n");
2018 fprintf (stderr, "%-5s %10s %10s %10s\n",
2019 "Size", "Allocated", "Used", "Overhead");
2020 for (i = 0; i < NUM_ORDERS; ++i)
2021 {
2022 page_entry *p;
2023 size_t allocated;
2024 size_t in_use;
2025 size_t overhead;
2026
2027 /* Skip empty entries. */
2028 if (!G.pages[i])
2029 continue;
2030
2031 overhead = allocated = in_use = 0;
2032
2033 /* Figure out the total number of bytes allocated for objects of
2034 this size, and how many of them are actually in use. Also figure
2035 out how much memory the page table is using. */
2036 for (p = G.pages[i]; p; p = p->next)
2037 {
2038 allocated += p->bytes;
2039 in_use +=
2040 (OBJECTS_IN_PAGE (p) - p->num_free_objects) * OBJECT_SIZE (i);
2041
2042 overhead += (sizeof (page_entry) - sizeof (long)
2043 + BITMAP_SIZE (OBJECTS_IN_PAGE (p) + 1));
2044 }
2045 fprintf (stderr, "%-5lu %10lu%c %10lu%c %10lu%c\n",
2046 (unsigned long) OBJECT_SIZE (i),
2047 SCALE (allocated), STAT_LABEL (allocated),
2048 SCALE (in_use), STAT_LABEL (in_use),
2049 SCALE (overhead), STAT_LABEL (overhead));
2050 total_overhead += overhead;
2051 }
2052 fprintf (stderr, "%-5s %10lu%c %10lu%c %10lu%c\n", "Total",
2053 SCALE (G.bytes_mapped), STAT_LABEL (G.bytes_mapped),
2054 SCALE (G.allocated), STAT_LABEL(G.allocated),
2055 SCALE (total_overhead), STAT_LABEL (total_overhead));
2056
2057 #ifdef GATHER_STATISTICS
2058 {
2059 fprintf (stderr, "\nTotal allocations and overheads during the compilation process\n");
2060
2061 fprintf (stderr, "Total Overhead: %10lld\n",
2062 G.stats.total_overhead);
2063 fprintf (stderr, "Total Allocated: %10lld\n",
2064 G.stats.total_allocated);
2065
2066 fprintf (stderr, "Total Overhead under 32B: %10lld\n",
2067 G.stats.total_overhead_under32);
2068 fprintf (stderr, "Total Allocated under 32B: %10lld\n",
2069 G.stats.total_allocated_under32);
2070 fprintf (stderr, "Total Overhead under 64B: %10lld\n",
2071 G.stats.total_overhead_under64);
2072 fprintf (stderr, "Total Allocated under 64B: %10lld\n",
2073 G.stats.total_allocated_under64);
2074 fprintf (stderr, "Total Overhead under 128B: %10lld\n",
2075 G.stats.total_overhead_under128);
2076 fprintf (stderr, "Total Allocated under 128B: %10lld\n",
2077 G.stats.total_allocated_under128);
2078
2079 for (i = 0; i < NUM_ORDERS; i++)
2080 if (G.stats.total_allocated_per_order[i])
2081 {
2082 fprintf (stderr, "Total Overhead page size %7d: %10lld\n",
2083 OBJECT_SIZE (i), G.stats.total_overhead_per_order[i]);
2084 fprintf (stderr, "Total Allocated page size %7d: %10lld\n",
2085 OBJECT_SIZE (i), G.stats.total_allocated_per_order[i]);
2086 }
2087 }
2088 #endif
2089 }
2090 \f
2091 struct ggc_pch_data
2092 {
2093 struct ggc_pch_ondisk
2094 {
2095 unsigned totals[NUM_ORDERS];
2096 } d;
2097 size_t base[NUM_ORDERS];
2098 size_t written[NUM_ORDERS];
2099 };
2100
2101 struct ggc_pch_data *
2102 init_ggc_pch (void)
2103 {
2104 return xcalloc (sizeof (struct ggc_pch_data), 1);
2105 }
2106
2107 void
2108 ggc_pch_count_object (struct ggc_pch_data *d, void *x ATTRIBUTE_UNUSED,
2109 size_t size, bool is_string ATTRIBUTE_UNUSED)
2110 {
2111 unsigned order;
2112
2113 if (size <= 256)
2114 order = size_lookup[size];
2115 else
2116 {
2117 order = 9;
2118 while (size > OBJECT_SIZE (order))
2119 order++;
2120 }
2121
2122 d->d.totals[order]++;
2123 }
2124
2125 size_t
2126 ggc_pch_total_size (struct ggc_pch_data *d)
2127 {
2128 size_t a = 0;
2129 unsigned i;
2130
2131 for (i = 0; i < NUM_ORDERS; i++)
2132 a += ROUND_UP (d->d.totals[i] * OBJECT_SIZE (i), G.pagesize);
2133 return a;
2134 }
2135
2136 void
2137 ggc_pch_this_base (struct ggc_pch_data *d, void *base)
2138 {
2139 size_t a = (size_t) base;
2140 unsigned i;
2141
2142 for (i = 0; i < NUM_ORDERS; i++)
2143 {
2144 d->base[i] = a;
2145 a += ROUND_UP (d->d.totals[i] * OBJECT_SIZE (i), G.pagesize);
2146 }
2147 }
2148
2149
2150 char *
2151 ggc_pch_alloc_object (struct ggc_pch_data *d, void *x ATTRIBUTE_UNUSED,
2152 size_t size, bool is_string ATTRIBUTE_UNUSED)
2153 {
2154 unsigned order;
2155 char *result;
2156
2157 if (size <= 256)
2158 order = size_lookup[size];
2159 else
2160 {
2161 order = 9;
2162 while (size > OBJECT_SIZE (order))
2163 order++;
2164 }
2165
2166 result = (char *) d->base[order];
2167 d->base[order] += OBJECT_SIZE (order);
2168 return result;
2169 }
2170
2171 void
2172 ggc_pch_prepare_write (struct ggc_pch_data *d ATTRIBUTE_UNUSED,
2173 FILE *f ATTRIBUTE_UNUSED)
2174 {
2175 /* Nothing to do. */
2176 }
2177
2178 void
2179 ggc_pch_write_object (struct ggc_pch_data *d ATTRIBUTE_UNUSED,
2180 FILE *f, void *x, void *newx ATTRIBUTE_UNUSED,
2181 size_t size, bool is_string ATTRIBUTE_UNUSED)
2182 {
2183 unsigned order;
2184 static const char emptyBytes[256];
2185
2186 if (size <= 256)
2187 order = size_lookup[size];
2188 else
2189 {
2190 order = 9;
2191 while (size > OBJECT_SIZE (order))
2192 order++;
2193 }
2194
2195 if (fwrite (x, size, 1, f) != 1)
2196 fatal_error ("can't write PCH file: %m");
2197
2198 /* If SIZE is not the same as OBJECT_SIZE(order), then we need to pad the
2199 object out to OBJECT_SIZE(order). This happens for strings. */
2200
2201 if (size != OBJECT_SIZE (order))
2202 {
2203 unsigned padding = OBJECT_SIZE(order) - size;
2204
2205 /* To speed small writes, we use a nulled-out array that's larger
2206 than most padding requests as the source for our null bytes. This
2207 permits us to do the padding with fwrite() rather than fseek(), and
2208 limits the chance the the OS may try to flush any outstanding
2209 writes. */
2210 if (padding <= sizeof(emptyBytes))
2211 {
2212 if (fwrite (emptyBytes, 1, padding, f) != padding)
2213 fatal_error ("can't write PCH file");
2214 }
2215 else
2216 {
2217 /* Larger than our buffer? Just default to fseek. */
2218 if (fseek (f, padding, SEEK_CUR) != 0)
2219 fatal_error ("can't write PCH file");
2220 }
2221 }
2222
2223 d->written[order]++;
2224 if (d->written[order] == d->d.totals[order]
2225 && fseek (f, ROUND_UP_VALUE (d->d.totals[order] * OBJECT_SIZE (order),
2226 G.pagesize),
2227 SEEK_CUR) != 0)
2228 fatal_error ("can't write PCH file: %m");
2229 }
2230
2231 void
2232 ggc_pch_finish (struct ggc_pch_data *d, FILE *f)
2233 {
2234 if (fwrite (&d->d, sizeof (d->d), 1, f) != 1)
2235 fatal_error ("can't write PCH file: %m");
2236 free (d);
2237 }
2238
2239 /* Move the PCH PTE entries just added to the end of by_depth, to the
2240 front. */
2241
2242 static void
2243 move_ptes_to_front (int count_old_page_tables, int count_new_page_tables)
2244 {
2245 unsigned i;
2246
2247 /* First, we swap the new entries to the front of the varrays. */
2248 page_entry **new_by_depth;
2249 unsigned long **new_save_in_use;
2250
2251 new_by_depth = xmalloc (G.by_depth_max * sizeof (page_entry *));
2252 new_save_in_use = xmalloc (G.by_depth_max * sizeof (unsigned long *));
2253
2254 memcpy (&new_by_depth[0],
2255 &G.by_depth[count_old_page_tables],
2256 count_new_page_tables * sizeof (void *));
2257 memcpy (&new_by_depth[count_new_page_tables],
2258 &G.by_depth[0],
2259 count_old_page_tables * sizeof (void *));
2260 memcpy (&new_save_in_use[0],
2261 &G.save_in_use[count_old_page_tables],
2262 count_new_page_tables * sizeof (void *));
2263 memcpy (&new_save_in_use[count_new_page_tables],
2264 &G.save_in_use[0],
2265 count_old_page_tables * sizeof (void *));
2266
2267 free (G.by_depth);
2268 free (G.save_in_use);
2269
2270 G.by_depth = new_by_depth;
2271 G.save_in_use = new_save_in_use;
2272
2273 /* Now update all the index_by_depth fields. */
2274 for (i = G.by_depth_in_use; i > 0; --i)
2275 {
2276 page_entry *p = G.by_depth[i-1];
2277 p->index_by_depth = i-1;
2278 }
2279
2280 /* And last, we update the depth pointers in G.depth. The first
2281 entry is already 0, and context 0 entries always start at index
2282 0, so there is nothing to update in the first slot. We need a
2283 second slot, only if we have old ptes, and if we do, they start
2284 at index count_new_page_tables. */
2285 if (count_old_page_tables)
2286 push_depth (count_new_page_tables);
2287 }
2288
2289 void
2290 ggc_pch_read (FILE *f, void *addr)
2291 {
2292 struct ggc_pch_ondisk d;
2293 unsigned i;
2294 char *offs = addr;
2295 unsigned long count_old_page_tables;
2296 unsigned long count_new_page_tables;
2297
2298 count_old_page_tables = G.by_depth_in_use;
2299
2300 /* We've just read in a PCH file. So, every object that used to be
2301 allocated is now free. */
2302 clear_marks ();
2303 #ifdef ENABLE_GC_CHECKING
2304 poison_pages ();
2305 #endif
2306
2307 /* No object read from a PCH file should ever be freed. So, set the
2308 context depth to 1, and set the depth of all the currently-allocated
2309 pages to be 1 too. PCH pages will have depth 0. */
2310 gcc_assert (!G.context_depth);
2311 G.context_depth = 1;
2312 for (i = 0; i < NUM_ORDERS; i++)
2313 {
2314 page_entry *p;
2315 for (p = G.pages[i]; p != NULL; p = p->next)
2316 p->context_depth = G.context_depth;
2317 }
2318
2319 /* Allocate the appropriate page-table entries for the pages read from
2320 the PCH file. */
2321 if (fread (&d, sizeof (d), 1, f) != 1)
2322 fatal_error ("can't read PCH file: %m");
2323
2324 for (i = 0; i < NUM_ORDERS; i++)
2325 {
2326 struct page_entry *entry;
2327 char *pte;
2328 size_t bytes;
2329 size_t num_objs;
2330 size_t j;
2331
2332 if (d.totals[i] == 0)
2333 continue;
2334
2335 bytes = ROUND_UP (d.totals[i] * OBJECT_SIZE (i), G.pagesize);
2336 num_objs = bytes / OBJECT_SIZE (i);
2337 entry = xcalloc (1, (sizeof (struct page_entry)
2338 - sizeof (long)
2339 + BITMAP_SIZE (num_objs + 1)));
2340 entry->bytes = bytes;
2341 entry->page = offs;
2342 entry->context_depth = 0;
2343 offs += bytes;
2344 entry->num_free_objects = 0;
2345 entry->order = i;
2346
2347 for (j = 0;
2348 j + HOST_BITS_PER_LONG <= num_objs + 1;
2349 j += HOST_BITS_PER_LONG)
2350 entry->in_use_p[j / HOST_BITS_PER_LONG] = -1;
2351 for (; j < num_objs + 1; j++)
2352 entry->in_use_p[j / HOST_BITS_PER_LONG]
2353 |= 1L << (j % HOST_BITS_PER_LONG);
2354
2355 for (pte = entry->page;
2356 pte < entry->page + entry->bytes;
2357 pte += G.pagesize)
2358 set_page_table_entry (pte, entry);
2359
2360 if (G.page_tails[i] != NULL)
2361 G.page_tails[i]->next = entry;
2362 else
2363 G.pages[i] = entry;
2364 G.page_tails[i] = entry;
2365
2366 /* We start off by just adding all the new information to the
2367 end of the varrays, later, we will move the new information
2368 to the front of the varrays, as the PCH page tables are at
2369 context 0. */
2370 push_by_depth (entry, 0);
2371 }
2372
2373 /* Now, we update the various data structures that speed page table
2374 handling. */
2375 count_new_page_tables = G.by_depth_in_use - count_old_page_tables;
2376
2377 move_ptes_to_front (count_old_page_tables, count_new_page_tables);
2378
2379 /* Update the statistics. */
2380 G.allocated = G.allocated_last_gc = offs - (char *)addr;
2381 }
This page took 0.146745 seconds and 6 git commands to generate.