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