1 /* "Bag-of-pages" garbage collector for the GNU compiler.
2 Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005
3 Free Software Foundation, Inc.
5 This file is part of GCC.
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
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
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, 51 Franklin Street, Fifth Floor, Boston, MA
24 #include "coretypes.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>
41 # include <valgrind.h>
44 /* Avoid #ifdef:s when we can help it. */
45 #define VALGRIND_DISCARD(x)
48 /* Prefer MAP_ANON(YMOUS) to /dev/zero, since we don't need to keep a
49 file open. Prefer either to valloc. */
51 # undef HAVE_MMAP_DEV_ZERO
53 # include <sys/mman.h>
55 # define MAP_FAILED -1
57 # if !defined (MAP_ANONYMOUS) && defined (MAP_ANON)
58 # define MAP_ANONYMOUS MAP_ANON
64 #ifdef HAVE_MMAP_DEV_ZERO
66 # include <sys/mman.h>
68 # define MAP_FAILED -1
75 #define USING_MALLOC_PAGE_GROUPS
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.
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.
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.
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
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. */
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)
115 #ifndef HOST_BITS_PER_PTR
116 #define HOST_BITS_PER_PTR HOST_BITS_PER_LONG
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:
126 msb +----------------+----+------+------+ lsb
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.
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
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)
147 #define LOOKUP_L1(p) \
148 (((size_t) (p) >> (32 - PAGE_L1_BITS)) & ((1 << PAGE_L1_BITS) - 1))
150 #define LOOKUP_L2(p) \
151 (((size_t) (p) >> G.lg_pagesize) & ((1 << PAGE_L2_BITS) - 1))
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]
157 /* The number of objects in P. */
158 #define OBJECTS_IN_PAGE(P) ((P)->bytes / OBJECT_SIZE ((P)->order))
160 /* The size of an object on a page of the indicated ORDER. */
161 #define OBJECT_SIZE(ORDER) object_size_table[ORDER]
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))
172 /* The number of extra orders, not corresponding to power-of-two sized
175 #define NUM_EXTRA_ORDERS ARRAY_SIZE (extra_order_size_table)
177 #define RTL_SIZE(NSLOTS) \
178 (RTX_HDR_SIZE + (NSLOTS) * sizeof (rtunion))
180 #define TREE_EXP_SIZE(OPS) \
181 (sizeof (struct tree_exp) + ((OPS) - 1) * sizeof (tree))
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. */
187 static const size_t extra_order_size_table
[] = {
188 sizeof (struct stmt_ann_d
),
189 sizeof (struct var_ann_d
),
190 sizeof (struct tree_decl_non_common
),
191 sizeof (struct tree_field_decl
),
192 sizeof (struct tree_parm_decl
),
193 sizeof (struct tree_var_decl
),
194 sizeof (struct tree_list
),
195 sizeof (struct tree_ssa_name
),
196 sizeof (struct tree_function_decl
),
197 sizeof (struct tree_binfo
),
198 sizeof (struct function
),
199 sizeof (struct basic_block_def
),
200 sizeof (bitmap_element
),
201 /* PHI nodes with one to three arguments are already covered by the
203 sizeof (struct tree_phi_node
) + sizeof (struct phi_arg_d
) * 3,
205 RTL_SIZE (2), /* MEM, PLUS, etc. */
206 RTL_SIZE (9), /* INSN */
209 /* The total number of orders. */
211 #define NUM_ORDERS (HOST_BITS_PER_PTR + NUM_EXTRA_ORDERS)
213 /* We use this structure to determine the alignment required for
214 allocations. For power-of-two sized allocations, that's not a
215 problem, but it does matter for odd-sized allocations. */
217 struct max_alignment
{
225 /* The biggest alignment required. */
227 #define MAX_ALIGNMENT (offsetof (struct max_alignment, u))
229 /* Compute the smallest nonnegative number which when added to X gives
232 #define ROUND_UP_VALUE(x, f) ((f) - 1 - ((f) - 1 + (x)) % (f))
234 /* Compute the smallest multiple of F that is >= X. */
236 #define ROUND_UP(x, f) (CEIL (x, f) * (f))
238 /* The Ith entry is the number of objects on a page or order I. */
240 static unsigned objects_per_page_table
[NUM_ORDERS
];
242 /* The Ith entry is the size of an object on a page of order I. */
244 static size_t object_size_table
[NUM_ORDERS
];
246 /* The Ith entry is a pair of numbers (mult, shift) such that
247 ((k * mult) >> shift) mod 2^32 == (k / OBJECT_SIZE(I)) mod 2^32,
248 for all k evenly divisible by OBJECT_SIZE(I). */
255 inverse_table
[NUM_ORDERS
];
257 /* A page_entry records the status of an allocation page. This
258 structure is dynamically sized to fit the bitmap in_use_p. */
259 typedef struct page_entry
261 /* The next page-entry with objects of the same size, or NULL if
262 this is the last page-entry. */
263 struct page_entry
*next
;
265 /* The previous page-entry with objects of the same size, or NULL if
266 this is the first page-entry. The PREV pointer exists solely to
267 keep the cost of ggc_free manageable. */
268 struct page_entry
*prev
;
270 /* The number of bytes allocated. (This will always be a multiple
271 of the host system page size.) */
274 /* The address at which the memory is allocated. */
277 #ifdef USING_MALLOC_PAGE_GROUPS
278 /* Back pointer to the page group this page came from. */
279 struct page_group
*group
;
282 /* This is the index in the by_depth varray where this page table
284 unsigned long index_by_depth
;
286 /* Context depth of this page. */
287 unsigned short context_depth
;
289 /* The number of free objects remaining on this page. */
290 unsigned short num_free_objects
;
292 /* A likely candidate for the bit position of a free object for the
293 next allocation from this page. */
294 unsigned short next_bit_hint
;
296 /* The lg of size of objects allocated from this page. */
299 /* A bit vector indicating whether or not objects are in use. The
300 Nth bit is one if the Nth object on this page is allocated. This
301 array is dynamically sized. */
302 unsigned long in_use_p
[1];
305 #ifdef USING_MALLOC_PAGE_GROUPS
306 /* A page_group describes a large allocation from malloc, from which
307 we parcel out aligned pages. */
308 typedef struct page_group
310 /* A linked list of all extant page groups. */
311 struct page_group
*next
;
313 /* The address we received from malloc. */
316 /* The size of the block. */
319 /* A bitmask of pages in use. */
324 #if HOST_BITS_PER_PTR <= 32
326 /* On 32-bit hosts, we use a two level page table, as pictured above. */
327 typedef page_entry
**page_table
[PAGE_L1_SIZE
];
331 /* On 64-bit hosts, we use the same two level page tables plus a linked
332 list that disambiguates the top 32-bits. There will almost always be
333 exactly one entry in the list. */
334 typedef struct page_table_chain
336 struct page_table_chain
*next
;
338 page_entry
**table
[PAGE_L1_SIZE
];
343 /* The rest of the global variables. */
344 static struct globals
346 /* The Nth element in this array is a page with objects of size 2^N.
347 If there are any pages with free objects, they will be at the
348 head of the list. NULL if there are no page-entries for this
350 page_entry
*pages
[NUM_ORDERS
];
352 /* The Nth element in this array is the last page with objects of
353 size 2^N. NULL if there are no page-entries for this object
355 page_entry
*page_tails
[NUM_ORDERS
];
357 /* Lookup table for associating allocation pages with object addresses. */
360 /* The system's page size. */
364 /* Bytes currently allocated. */
367 /* Bytes currently allocated at the end of the last collection. */
368 size_t allocated_last_gc
;
370 /* Total amount of memory mapped. */
373 /* Bit N set if any allocations have been done at context depth N. */
374 unsigned long context_depth_allocations
;
376 /* Bit N set if any collections have been done at context depth N. */
377 unsigned long context_depth_collections
;
379 /* The current depth in the context stack. */
380 unsigned short context_depth
;
382 /* A file descriptor open to /dev/zero for reading. */
383 #if defined (HAVE_MMAP_DEV_ZERO)
387 /* A cache of free system pages. */
388 page_entry
*free_pages
;
390 #ifdef USING_MALLOC_PAGE_GROUPS
391 page_group
*page_groups
;
394 /* The file descriptor for debugging output. */
397 /* Current number of elements in use in depth below. */
398 unsigned int depth_in_use
;
400 /* Maximum number of elements that can be used before resizing. */
401 unsigned int depth_max
;
403 /* Each element of this arry is an index in by_depth where the given
404 depth starts. This structure is indexed by that given depth we
405 are interested in. */
408 /* Current number of elements in use in by_depth below. */
409 unsigned int by_depth_in_use
;
411 /* Maximum number of elements that can be used before resizing. */
412 unsigned int by_depth_max
;
414 /* Each element of this array is a pointer to a page_entry, all
415 page_entries can be found in here by increasing depth.
416 index_by_depth in the page_entry is the index into this data
417 structure where that page_entry can be found. This is used to
418 speed up finding all page_entries at a particular depth. */
419 page_entry
**by_depth
;
421 /* Each element is a pointer to the saved in_use_p bits, if any,
422 zero otherwise. We allocate them all together, to enable a
423 better runtime data access pattern. */
424 unsigned long **save_in_use
;
426 #ifdef ENABLE_GC_ALWAYS_COLLECT
427 /* List of free objects to be verified as actually free on the
432 struct free_object
*next
;
436 #ifdef GATHER_STATISTICS
439 /* Total memory allocated with ggc_alloc. */
440 unsigned long long total_allocated
;
441 /* Total overhead for memory to be allocated with ggc_alloc. */
442 unsigned long long total_overhead
;
444 /* Total allocations and overhead for sizes less than 32, 64 and 128.
445 These sizes are interesting because they are typical cache line
448 unsigned long long total_allocated_under32
;
449 unsigned long long total_overhead_under32
;
451 unsigned long long total_allocated_under64
;
452 unsigned long long total_overhead_under64
;
454 unsigned long long total_allocated_under128
;
455 unsigned long long total_overhead_under128
;
457 /* The allocations for each of the allocation orders. */
458 unsigned long long total_allocated_per_order
[NUM_ORDERS
];
460 /* The overhead for each of the allocation orders. */
461 unsigned long long total_overhead_per_order
[NUM_ORDERS
];
466 /* The size in bytes required to maintain a bitmap for the objects
468 #define BITMAP_SIZE(Num_objects) \
469 (CEIL ((Num_objects), HOST_BITS_PER_LONG) * sizeof(long))
471 /* Allocate pages in chunks of this size, to throttle calls to memory
472 allocation routines. The first page is used, the rest go onto the
473 free list. This cannot be larger than HOST_BITS_PER_INT for the
474 in_use bitmask for page_group. Hosts that need a different value
475 can override this by defining GGC_QUIRE_SIZE explicitly. */
476 #ifndef GGC_QUIRE_SIZE
478 # define GGC_QUIRE_SIZE 256
480 # define GGC_QUIRE_SIZE 16
484 /* Initial guess as to how many page table entries we might need. */
485 #define INITIAL_PTE_COUNT 128
487 static int ggc_allocated_p (const void *);
488 static page_entry
*lookup_page_table_entry (const void *);
489 static void set_page_table_entry (void *, page_entry
*);
491 static char *alloc_anon (char *, size_t);
493 #ifdef USING_MALLOC_PAGE_GROUPS
494 static size_t page_group_index (char *, char *);
495 static void set_page_group_in_use (page_group
*, char *);
496 static void clear_page_group_in_use (page_group
*, char *);
498 static struct page_entry
* alloc_page (unsigned);
499 static void free_page (struct page_entry
*);
500 static void release_pages (void);
501 static void clear_marks (void);
502 static void sweep_pages (void);
503 static void ggc_recalculate_in_use_p (page_entry
*);
504 static void compute_inverse (unsigned);
505 static inline void adjust_depth (void);
506 static void move_ptes_to_front (int, int);
508 void debug_print_page_list (int);
509 static void push_depth (unsigned int);
510 static void push_by_depth (page_entry
*, unsigned long *);
512 /* Push an entry onto G.depth. */
515 push_depth (unsigned int i
)
517 if (G
.depth_in_use
>= G
.depth_max
)
520 G
.depth
= xrealloc (G
.depth
, G
.depth_max
* sizeof (unsigned int));
522 G
.depth
[G
.depth_in_use
++] = i
;
525 /* Push an entry onto G.by_depth and G.save_in_use. */
528 push_by_depth (page_entry
*p
, unsigned long *s
)
530 if (G
.by_depth_in_use
>= G
.by_depth_max
)
533 G
.by_depth
= xrealloc (G
.by_depth
,
534 G
.by_depth_max
* sizeof (page_entry
*));
535 G
.save_in_use
= xrealloc (G
.save_in_use
,
536 G
.by_depth_max
* sizeof (unsigned long *));
538 G
.by_depth
[G
.by_depth_in_use
] = p
;
539 G
.save_in_use
[G
.by_depth_in_use
++] = s
;
542 #if (GCC_VERSION < 3001)
543 #define prefetch(X) ((void) X)
545 #define prefetch(X) __builtin_prefetch (X)
548 #define save_in_use_p_i(__i) \
550 #define save_in_use_p(__p) \
551 (save_in_use_p_i (__p->index_by_depth))
553 /* Returns nonzero if P was allocated in GC'able memory. */
556 ggc_allocated_p (const void *p
)
561 #if HOST_BITS_PER_PTR <= 32
564 page_table table
= G
.lookup
;
565 size_t high_bits
= (size_t) p
& ~ (size_t) 0xffffffff;
570 if (table
->high_bits
== high_bits
)
574 base
= &table
->table
[0];
577 /* Extract the level 1 and 2 indices. */
581 return base
[L1
] && base
[L1
][L2
];
584 /* Traverse the page table and find the entry for a page.
585 Die (probably) if the object wasn't allocated via GC. */
587 static inline page_entry
*
588 lookup_page_table_entry (const void *p
)
593 #if HOST_BITS_PER_PTR <= 32
596 page_table table
= G
.lookup
;
597 size_t high_bits
= (size_t) p
& ~ (size_t) 0xffffffff;
598 while (table
->high_bits
!= high_bits
)
600 base
= &table
->table
[0];
603 /* Extract the level 1 and 2 indices. */
610 /* Set the page table entry for a page. */
613 set_page_table_entry (void *p
, page_entry
*entry
)
618 #if HOST_BITS_PER_PTR <= 32
622 size_t high_bits
= (size_t) p
& ~ (size_t) 0xffffffff;
623 for (table
= G
.lookup
; table
; table
= table
->next
)
624 if (table
->high_bits
== high_bits
)
627 /* Not found -- allocate a new table. */
628 table
= xcalloc (1, sizeof(*table
));
629 table
->next
= G
.lookup
;
630 table
->high_bits
= high_bits
;
633 base
= &table
->table
[0];
636 /* Extract the level 1 and 2 indices. */
640 if (base
[L1
] == NULL
)
641 base
[L1
] = XCNEWVEC (page_entry
*, PAGE_L2_SIZE
);
643 base
[L1
][L2
] = entry
;
646 /* Prints the page-entry for object size ORDER, for debugging. */
649 debug_print_page_list (int order
)
652 printf ("Head=%p, Tail=%p:\n", (void *) G
.pages
[order
],
653 (void *) G
.page_tails
[order
]);
657 printf ("%p(%1d|%3d) -> ", (void *) p
, p
->context_depth
,
658 p
->num_free_objects
);
666 /* Allocate SIZE bytes of anonymous memory, preferably near PREF,
667 (if non-null). The ifdef structure here is intended to cause a
668 compile error unless exactly one of the HAVE_* is defined. */
671 alloc_anon (char *pref ATTRIBUTE_UNUSED
, size_t size
)
673 #ifdef HAVE_MMAP_ANON
674 char *page
= mmap (pref
, size
, PROT_READ
| PROT_WRITE
,
675 MAP_PRIVATE
| MAP_ANONYMOUS
, -1, 0);
677 #ifdef HAVE_MMAP_DEV_ZERO
678 char *page
= mmap (pref
, size
, PROT_READ
| PROT_WRITE
,
679 MAP_PRIVATE
, G
.dev_zero_fd
, 0);
682 if (page
== (char *) MAP_FAILED
)
684 perror ("virtual memory exhausted");
685 exit (FATAL_EXIT_CODE
);
688 /* Remember that we allocated this memory. */
689 G
.bytes_mapped
+= size
;
691 /* Pretend we don't have access to the allocated pages. We'll enable
692 access to smaller pieces of the area in ggc_alloc. Discard the
693 handle to avoid handle leak. */
694 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (page
, size
));
699 #ifdef USING_MALLOC_PAGE_GROUPS
700 /* Compute the index for this page into the page group. */
703 page_group_index (char *allocation
, char *page
)
705 return (size_t) (page
- allocation
) >> G
.lg_pagesize
;
708 /* Set and clear the in_use bit for this page in the page group. */
711 set_page_group_in_use (page_group
*group
, char *page
)
713 group
->in_use
|= 1 << page_group_index (group
->allocation
, page
);
717 clear_page_group_in_use (page_group
*group
, char *page
)
719 group
->in_use
&= ~(1 << page_group_index (group
->allocation
, page
));
723 /* Allocate a new page for allocating objects of size 2^ORDER,
724 and return an entry for it. The entry is not added to the
725 appropriate page_table list. */
727 static inline struct page_entry
*
728 alloc_page (unsigned order
)
730 struct page_entry
*entry
, *p
, **pp
;
734 size_t page_entry_size
;
736 #ifdef USING_MALLOC_PAGE_GROUPS
740 num_objects
= OBJECTS_PER_PAGE (order
);
741 bitmap_size
= BITMAP_SIZE (num_objects
+ 1);
742 page_entry_size
= sizeof (page_entry
) - sizeof (long) + bitmap_size
;
743 entry_size
= num_objects
* OBJECT_SIZE (order
);
744 if (entry_size
< G
.pagesize
)
745 entry_size
= G
.pagesize
;
750 /* Check the list of free pages for one we can use. */
751 for (pp
= &G
.free_pages
, p
= *pp
; p
; pp
= &p
->next
, p
= *pp
)
752 if (p
->bytes
== entry_size
)
757 /* Recycle the allocated memory from this page ... */
761 #ifdef USING_MALLOC_PAGE_GROUPS
765 /* ... and, if possible, the page entry itself. */
766 if (p
->order
== order
)
769 memset (entry
, 0, page_entry_size
);
775 else if (entry_size
== G
.pagesize
)
777 /* We want just one page. Allocate a bunch of them and put the
778 extras on the freelist. (Can only do this optimization with
779 mmap for backing store.) */
780 struct page_entry
*e
, *f
= G
.free_pages
;
783 page
= alloc_anon (NULL
, G
.pagesize
* GGC_QUIRE_SIZE
);
785 /* This loop counts down so that the chain will be in ascending
787 for (i
= GGC_QUIRE_SIZE
- 1; i
>= 1; i
--)
789 e
= xcalloc (1, page_entry_size
);
791 e
->bytes
= G
.pagesize
;
792 e
->page
= page
+ (i
<< G
.lg_pagesize
);
800 page
= alloc_anon (NULL
, entry_size
);
802 #ifdef USING_MALLOC_PAGE_GROUPS
805 /* Allocate a large block of memory and serve out the aligned
806 pages therein. This results in much less memory wastage
807 than the traditional implementation of valloc. */
809 char *allocation
, *a
, *enda
;
810 size_t alloc_size
, head_slop
, tail_slop
;
811 int multiple_pages
= (entry_size
== G
.pagesize
);
814 alloc_size
= GGC_QUIRE_SIZE
* G
.pagesize
;
816 alloc_size
= entry_size
+ G
.pagesize
- 1;
817 allocation
= xmalloc (alloc_size
);
819 page
= (char *) (((size_t) allocation
+ G
.pagesize
- 1) & -G
.pagesize
);
820 head_slop
= page
- allocation
;
822 tail_slop
= ((size_t) allocation
+ alloc_size
) & (G
.pagesize
- 1);
824 tail_slop
= alloc_size
- entry_size
- head_slop
;
825 enda
= allocation
+ alloc_size
- tail_slop
;
827 /* We allocated N pages, which are likely not aligned, leaving
828 us with N-1 usable pages. We plan to place the page_group
829 structure somewhere in the slop. */
830 if (head_slop
>= sizeof (page_group
))
831 group
= (page_group
*)page
- 1;
834 /* We magically got an aligned allocation. Too bad, we have
835 to waste a page anyway. */
839 tail_slop
+= G
.pagesize
;
841 gcc_assert (tail_slop
>= sizeof (page_group
));
842 group
= (page_group
*)enda
;
843 tail_slop
-= sizeof (page_group
);
846 /* Remember that we allocated this memory. */
847 group
->next
= G
.page_groups
;
848 group
->allocation
= allocation
;
849 group
->alloc_size
= alloc_size
;
851 G
.page_groups
= group
;
852 G
.bytes_mapped
+= alloc_size
;
854 /* If we allocated multiple pages, put the rest on the free list. */
857 struct page_entry
*e
, *f
= G
.free_pages
;
858 for (a
= enda
- G
.pagesize
; a
!= page
; a
-= G
.pagesize
)
860 e
= xcalloc (1, page_entry_size
);
862 e
->bytes
= G
.pagesize
;
874 entry
= xcalloc (1, page_entry_size
);
876 entry
->bytes
= entry_size
;
878 entry
->context_depth
= G
.context_depth
;
879 entry
->order
= order
;
880 entry
->num_free_objects
= num_objects
;
881 entry
->next_bit_hint
= 1;
883 G
.context_depth_allocations
|= (unsigned long)1 << G
.context_depth
;
885 #ifdef USING_MALLOC_PAGE_GROUPS
886 entry
->group
= group
;
887 set_page_group_in_use (group
, page
);
890 /* Set the one-past-the-end in-use bit. This acts as a sentry as we
891 increment the hint. */
892 entry
->in_use_p
[num_objects
/ HOST_BITS_PER_LONG
]
893 = (unsigned long) 1 << (num_objects
% HOST_BITS_PER_LONG
);
895 set_page_table_entry (page
, entry
);
897 if (GGC_DEBUG_LEVEL
>= 2)
898 fprintf (G
.debug_file
,
899 "Allocating page at %p, object size=%lu, data %p-%p\n",
900 (void *) entry
, (unsigned long) OBJECT_SIZE (order
), page
,
901 page
+ entry_size
- 1);
906 /* Adjust the size of G.depth so that no index greater than the one
907 used by the top of the G.by_depth is used. */
914 if (G
.by_depth_in_use
)
916 top
= G
.by_depth
[G
.by_depth_in_use
-1];
918 /* Peel back indices in depth that index into by_depth, so that
919 as new elements are added to by_depth, we note the indices
920 of those elements, if they are for new context depths. */
921 while (G
.depth_in_use
> (size_t)top
->context_depth
+1)
926 /* For a page that is no longer needed, put it on the free page list. */
929 free_page (page_entry
*entry
)
931 if (GGC_DEBUG_LEVEL
>= 2)
932 fprintf (G
.debug_file
,
933 "Deallocating page at %p, data %p-%p\n", (void *) entry
,
934 entry
->page
, entry
->page
+ entry
->bytes
- 1);
936 /* Mark the page as inaccessible. Discard the handle to avoid handle
938 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (entry
->page
, entry
->bytes
));
940 set_page_table_entry (entry
->page
, NULL
);
942 #ifdef USING_MALLOC_PAGE_GROUPS
943 clear_page_group_in_use (entry
->group
, entry
->page
);
946 if (G
.by_depth_in_use
> 1)
948 page_entry
*top
= G
.by_depth
[G
.by_depth_in_use
-1];
949 int i
= entry
->index_by_depth
;
951 /* We cannot free a page from a context deeper than the current
953 gcc_assert (entry
->context_depth
== top
->context_depth
);
955 /* Put top element into freed slot. */
957 G
.save_in_use
[i
] = G
.save_in_use
[G
.by_depth_in_use
-1];
958 top
->index_by_depth
= i
;
964 entry
->next
= G
.free_pages
;
965 G
.free_pages
= entry
;
968 /* Release the free page cache to the system. */
974 page_entry
*p
, *next
;
978 /* Gather up adjacent pages so they are unmapped together. */
989 while (p
&& p
->page
== start
+ len
)
998 G
.bytes_mapped
-= len
;
1001 G
.free_pages
= NULL
;
1003 #ifdef USING_MALLOC_PAGE_GROUPS
1004 page_entry
**pp
, *p
;
1005 page_group
**gp
, *g
;
1007 /* Remove all pages from free page groups from the list. */
1009 while ((p
= *pp
) != NULL
)
1010 if (p
->group
->in_use
== 0)
1018 /* Remove all free page groups, and release the storage. */
1019 gp
= &G
.page_groups
;
1020 while ((g
= *gp
) != NULL
)
1024 G
.bytes_mapped
-= g
->alloc_size
;
1025 free (g
->allocation
);
1032 /* This table provides a fast way to determine ceil(log_2(size)) for
1033 allocation requests. The minimum allocation size is eight bytes. */
1035 static unsigned char size_lookup
[512] =
1037 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4,
1038 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
1039 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
1040 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
1041 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1042 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1043 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1044 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
1045 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1046 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1047 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1048 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1049 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1050 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1051 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1052 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1053 8, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1054 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1055 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1056 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1057 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1058 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1059 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1060 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1061 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1062 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1063 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1064 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1065 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1066 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1067 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
1068 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9
1071 /* Typed allocation function. Does nothing special in this collector. */
1074 ggc_alloc_typed_stat (enum gt_types_enum type ATTRIBUTE_UNUSED
, size_t size
1077 return ggc_alloc_stat (size PASS_MEM_STAT
);
1080 /* Allocate a chunk of memory of SIZE bytes. Its contents are undefined. */
1083 ggc_alloc_stat (size_t size MEM_STAT_DECL
)
1085 size_t order
, word
, bit
, object_offset
, object_size
;
1086 struct page_entry
*entry
;
1091 order
= size_lookup
[size
];
1092 object_size
= OBJECT_SIZE (order
);
1097 while (size
> (object_size
= OBJECT_SIZE (order
)))
1101 /* If there are non-full pages for this size allocation, they are at
1102 the head of the list. */
1103 entry
= G
.pages
[order
];
1105 /* If there is no page for this object size, or all pages in this
1106 context are full, allocate a new page. */
1107 if (entry
== NULL
|| entry
->num_free_objects
== 0)
1109 struct page_entry
*new_entry
;
1110 new_entry
= alloc_page (order
);
1112 new_entry
->index_by_depth
= G
.by_depth_in_use
;
1113 push_by_depth (new_entry
, 0);
1115 /* We can skip context depths, if we do, make sure we go all the
1116 way to the new depth. */
1117 while (new_entry
->context_depth
>= G
.depth_in_use
)
1118 push_depth (G
.by_depth_in_use
-1);
1120 /* If this is the only entry, it's also the tail. If it is not
1121 the only entry, then we must update the PREV pointer of the
1122 ENTRY (G.pages[order]) to point to our new page entry. */
1124 G
.page_tails
[order
] = new_entry
;
1126 entry
->prev
= new_entry
;
1128 /* Put new pages at the head of the page list. By definition the
1129 entry at the head of the list always has a NULL pointer. */
1130 new_entry
->next
= entry
;
1131 new_entry
->prev
= NULL
;
1133 G
.pages
[order
] = new_entry
;
1135 /* For a new page, we know the word and bit positions (in the
1136 in_use bitmap) of the first available object -- they're zero. */
1137 new_entry
->next_bit_hint
= 1;
1144 /* First try to use the hint left from the previous allocation
1145 to locate a clear bit in the in-use bitmap. We've made sure
1146 that the one-past-the-end bit is always set, so if the hint
1147 has run over, this test will fail. */
1148 unsigned hint
= entry
->next_bit_hint
;
1149 word
= hint
/ HOST_BITS_PER_LONG
;
1150 bit
= hint
% HOST_BITS_PER_LONG
;
1152 /* If the hint didn't work, scan the bitmap from the beginning. */
1153 if ((entry
->in_use_p
[word
] >> bit
) & 1)
1156 while (~entry
->in_use_p
[word
] == 0)
1159 #if GCC_VERSION >= 3004
1160 bit
= __builtin_ctzl (~entry
->in_use_p
[word
]);
1162 while ((entry
->in_use_p
[word
] >> bit
) & 1)
1166 hint
= word
* HOST_BITS_PER_LONG
+ bit
;
1169 /* Next time, try the next bit. */
1170 entry
->next_bit_hint
= hint
+ 1;
1172 object_offset
= hint
* object_size
;
1175 /* Set the in-use bit. */
1176 entry
->in_use_p
[word
] |= ((unsigned long) 1 << bit
);
1178 /* Keep a running total of the number of free objects. If this page
1179 fills up, we may have to move it to the end of the list if the
1180 next page isn't full. If the next page is full, all subsequent
1181 pages are full, so there's no need to move it. */
1182 if (--entry
->num_free_objects
== 0
1183 && entry
->next
!= NULL
1184 && entry
->next
->num_free_objects
> 0)
1186 /* We have a new head for the list. */
1187 G
.pages
[order
] = entry
->next
;
1189 /* We are moving ENTRY to the end of the page table list.
1190 The new page at the head of the list will have NULL in
1191 its PREV field and ENTRY will have NULL in its NEXT field. */
1192 entry
->next
->prev
= NULL
;
1195 /* Append ENTRY to the tail of the list. */
1196 entry
->prev
= G
.page_tails
[order
];
1197 G
.page_tails
[order
]->next
= entry
;
1198 G
.page_tails
[order
] = entry
;
1201 /* Calculate the object's address. */
1202 result
= entry
->page
+ object_offset
;
1203 #ifdef GATHER_STATISTICS
1204 ggc_record_overhead (OBJECT_SIZE (order
), OBJECT_SIZE (order
) - size
,
1205 result PASS_MEM_STAT
);
1208 #ifdef ENABLE_GC_CHECKING
1209 /* Keep poisoning-by-writing-0xaf the object, in an attempt to keep the
1210 exact same semantics in presence of memory bugs, regardless of
1211 ENABLE_VALGRIND_CHECKING. We override this request below. Drop the
1212 handle to avoid handle leak. */
1213 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (result
, object_size
));
1215 /* `Poison' the entire allocated object, including any padding at
1217 memset (result
, 0xaf, object_size
);
1219 /* Make the bytes after the end of the object unaccessible. Discard the
1220 handle to avoid handle leak. */
1221 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS ((char *) result
+ size
,
1222 object_size
- size
));
1225 /* Tell Valgrind that the memory is there, but its content isn't
1226 defined. The bytes at the end of the object are still marked
1228 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (result
, size
));
1230 /* Keep track of how many bytes are being allocated. This
1231 information is used in deciding when to collect. */
1232 G
.allocated
+= object_size
;
1234 /* For timevar statistics. */
1235 timevar_ggc_mem_total
+= object_size
;
1237 #ifdef GATHER_STATISTICS
1239 size_t overhead
= object_size
- size
;
1241 G
.stats
.total_overhead
+= overhead
;
1242 G
.stats
.total_allocated
+= object_size
;
1243 G
.stats
.total_overhead_per_order
[order
] += overhead
;
1244 G
.stats
.total_allocated_per_order
[order
] += object_size
;
1248 G
.stats
.total_overhead_under32
+= overhead
;
1249 G
.stats
.total_allocated_under32
+= object_size
;
1253 G
.stats
.total_overhead_under64
+= overhead
;
1254 G
.stats
.total_allocated_under64
+= object_size
;
1258 G
.stats
.total_overhead_under128
+= overhead
;
1259 G
.stats
.total_allocated_under128
+= object_size
;
1264 if (GGC_DEBUG_LEVEL
>= 3)
1265 fprintf (G
.debug_file
,
1266 "Allocating object, requested size=%lu, actual=%lu at %p on %p\n",
1267 (unsigned long) size
, (unsigned long) object_size
, result
,
1273 /* If P is not marked, marks it and return false. Otherwise return true.
1274 P must have been allocated by the GC allocator; it mustn't point to
1275 static objects, stack variables, or memory allocated with malloc. */
1278 ggc_set_mark (const void *p
)
1284 /* Look up the page on which the object is alloced. If the object
1285 wasn't allocated by the collector, we'll probably die. */
1286 entry
= lookup_page_table_entry (p
);
1289 /* Calculate the index of the object on the page; this is its bit
1290 position in the in_use_p bitmap. */
1291 bit
= OFFSET_TO_BIT (((const char *) p
) - entry
->page
, entry
->order
);
1292 word
= bit
/ HOST_BITS_PER_LONG
;
1293 mask
= (unsigned long) 1 << (bit
% HOST_BITS_PER_LONG
);
1295 /* If the bit was previously set, skip it. */
1296 if (entry
->in_use_p
[word
] & mask
)
1299 /* Otherwise set it, and decrement the free object count. */
1300 entry
->in_use_p
[word
] |= mask
;
1301 entry
->num_free_objects
-= 1;
1303 if (GGC_DEBUG_LEVEL
>= 4)
1304 fprintf (G
.debug_file
, "Marking %p\n", p
);
1309 /* Return 1 if P has been marked, zero otherwise.
1310 P must have been allocated by the GC allocator; it mustn't point to
1311 static objects, stack variables, or memory allocated with malloc. */
1314 ggc_marked_p (const void *p
)
1320 /* Look up the page on which the object is alloced. If the object
1321 wasn't allocated by the collector, we'll probably die. */
1322 entry
= lookup_page_table_entry (p
);
1325 /* Calculate the index of the object on the page; this is its bit
1326 position in the in_use_p bitmap. */
1327 bit
= OFFSET_TO_BIT (((const char *) p
) - entry
->page
, entry
->order
);
1328 word
= bit
/ HOST_BITS_PER_LONG
;
1329 mask
= (unsigned long) 1 << (bit
% HOST_BITS_PER_LONG
);
1331 return (entry
->in_use_p
[word
] & mask
) != 0;
1334 /* Return the size of the gc-able object P. */
1337 ggc_get_size (const void *p
)
1339 page_entry
*pe
= lookup_page_table_entry (p
);
1340 return OBJECT_SIZE (pe
->order
);
1343 /* Release the memory for object P. */
1348 page_entry
*pe
= lookup_page_table_entry (p
);
1349 size_t order
= pe
->order
;
1350 size_t size
= OBJECT_SIZE (order
);
1352 #ifdef GATHER_STATISTICS
1353 ggc_free_overhead (p
);
1356 if (GGC_DEBUG_LEVEL
>= 3)
1357 fprintf (G
.debug_file
,
1358 "Freeing object, actual size=%lu, at %p on %p\n",
1359 (unsigned long) size
, p
, (void *) pe
);
1361 #ifdef ENABLE_GC_CHECKING
1362 /* Poison the data, to indicate the data is garbage. */
1363 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (p
, size
));
1364 memset (p
, 0xa5, size
);
1366 /* Let valgrind know the object is free. */
1367 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (p
, size
));
1369 #ifdef ENABLE_GC_ALWAYS_COLLECT
1370 /* In the completely-anal-checking mode, we do *not* immediately free
1371 the data, but instead verify that the data is *actually* not
1372 reachable the next time we collect. */
1374 struct free_object
*fo
= XNEW (struct free_object
);
1376 fo
->next
= G
.free_object_list
;
1377 G
.free_object_list
= fo
;
1381 unsigned int bit_offset
, word
, bit
;
1383 G
.allocated
-= size
;
1385 /* Mark the object not-in-use. */
1386 bit_offset
= OFFSET_TO_BIT (((const char *) p
) - pe
->page
, order
);
1387 word
= bit_offset
/ HOST_BITS_PER_LONG
;
1388 bit
= bit_offset
% HOST_BITS_PER_LONG
;
1389 pe
->in_use_p
[word
] &= ~(1UL << bit
);
1391 if (pe
->num_free_objects
++ == 0)
1395 /* If the page is completely full, then it's supposed to
1396 be after all pages that aren't. Since we've freed one
1397 object from a page that was full, we need to move the
1398 page to the head of the list.
1400 PE is the node we want to move. Q is the previous node
1401 and P is the next node in the list. */
1403 if (q
&& q
->num_free_objects
== 0)
1409 /* If PE was at the end of the list, then Q becomes the
1410 new end of the list. If PE was not the end of the
1411 list, then we need to update the PREV field for P. */
1413 G
.page_tails
[order
] = q
;
1417 /* Move PE to the head of the list. */
1418 pe
->next
= G
.pages
[order
];
1420 G
.pages
[order
]->prev
= pe
;
1421 G
.pages
[order
] = pe
;
1424 /* Reset the hint bit to point to the only free object. */
1425 pe
->next_bit_hint
= bit_offset
;
1431 /* Subroutine of init_ggc which computes the pair of numbers used to
1432 perform division by OBJECT_SIZE (order) and fills in inverse_table[].
1434 This algorithm is taken from Granlund and Montgomery's paper
1435 "Division by Invariant Integers using Multiplication"
1436 (Proc. SIGPLAN PLDI, 1994), section 9 (Exact division by
1440 compute_inverse (unsigned order
)
1445 size
= OBJECT_SIZE (order
);
1447 while (size
% 2 == 0)
1454 while (inv
* size
!= 1)
1455 inv
= inv
* (2 - inv
*size
);
1457 DIV_MULT (order
) = inv
;
1458 DIV_SHIFT (order
) = e
;
1461 /* Initialize the ggc-mmap allocator. */
1467 G
.pagesize
= getpagesize();
1468 G
.lg_pagesize
= exact_log2 (G
.pagesize
);
1470 #ifdef HAVE_MMAP_DEV_ZERO
1471 G
.dev_zero_fd
= open ("/dev/zero", O_RDONLY
);
1472 if (G
.dev_zero_fd
== -1)
1473 internal_error ("open /dev/zero: %m");
1477 G
.debug_file
= fopen ("ggc-mmap.debug", "w");
1479 G
.debug_file
= stdout
;
1483 /* StunOS has an amazing off-by-one error for the first mmap allocation
1484 after fiddling with RLIMIT_STACK. The result, as hard as it is to
1485 believe, is an unaligned page allocation, which would cause us to
1486 hork badly if we tried to use it. */
1488 char *p
= alloc_anon (NULL
, G
.pagesize
);
1489 struct page_entry
*e
;
1490 if ((size_t)p
& (G
.pagesize
- 1))
1492 /* How losing. Discard this one and try another. If we still
1493 can't get something useful, give up. */
1495 p
= alloc_anon (NULL
, G
.pagesize
);
1496 gcc_assert (!((size_t)p
& (G
.pagesize
- 1)));
1499 /* We have a good page, might as well hold onto it... */
1500 e
= XCNEW (struct page_entry
);
1501 e
->bytes
= G
.pagesize
;
1503 e
->next
= G
.free_pages
;
1508 /* Initialize the object size table. */
1509 for (order
= 0; order
< HOST_BITS_PER_PTR
; ++order
)
1510 object_size_table
[order
] = (size_t) 1 << order
;
1511 for (order
= HOST_BITS_PER_PTR
; order
< NUM_ORDERS
; ++order
)
1513 size_t s
= extra_order_size_table
[order
- HOST_BITS_PER_PTR
];
1514 object_size_table
[order
] = s
;
1517 /* Initialize the objects-per-page and inverse tables. */
1518 for (order
= 0; order
< NUM_ORDERS
; ++order
)
1520 objects_per_page_table
[order
] = G
.pagesize
/ OBJECT_SIZE (order
);
1521 if (objects_per_page_table
[order
] == 0)
1522 objects_per_page_table
[order
] = 1;
1523 compute_inverse (order
);
1526 /* Reset the size_lookup array to put appropriately sized objects in
1527 the special orders. All objects bigger than the previous power
1528 of two, but no greater than the special size, should go in the
1530 Enforce alignment during lookup. The resulting bin size must
1531 have the same or bigger alignment than the apparent alignment
1532 requirement from the size request (but not bigger alignment
1533 than MAX_ALIGNMENT). Consider an extra bin of size 76 (in
1534 addition to the 64 and 128 byte sized bins). A request of
1535 allocation size of 72 bytes must be served from the 128 bytes
1536 bin, because 72 bytes looks like a request for 8 byte aligned
1537 memory, while the 76 byte bin can only serve chunks with a
1538 guaranteed alignment of 4 bytes. */
1539 for (order
= HOST_BITS_PER_PTR
; order
< NUM_ORDERS
; ++order
)
1543 /* Build an alignment mask that can be used for testing
1544 size % 2*align. If (size | MAX_ALIGNMENT) & mask is non-zero
1545 then the requested size apparent alignment requirement
1546 (which is at most MAX_ALIGNMENT) is less or equal than what
1547 the OBJECT_SIZE bin can guarantee. */
1548 mask
= ~(((unsigned)-1) << ffs (OBJECT_SIZE (order
)));
1549 mask
&= 2 * MAX_ALIGNMENT
- 1;
1551 /* All objects smaller than the OBJECT_SIZE for this ORDER could go
1552 into ORDER. Determine the cases for which that is profitable
1553 and fulfilling the alignment requirements. Stop searching
1554 once a smaller bin with same or better alignment guarantee is
1556 for (i
= OBJECT_SIZE (order
); ; --i
)
1558 unsigned int old_sz
= OBJECT_SIZE (size_lookup
[i
]);
1559 if (!(old_sz
& (mask
>> 1))
1560 && old_sz
< OBJECT_SIZE (order
))
1563 /* If object of size I are presently using a larger bin, we would
1564 like to move them to ORDER. However, we can only do that if we
1565 can be sure they will be properly aligned. They will be properly
1566 aligned if either the ORDER bin is maximally aligned, or if
1567 objects of size I cannot be more strictly aligned than the
1568 alignment of this order. */
1569 if ((i
| MAX_ALIGNMENT
) & mask
1570 && old_sz
> OBJECT_SIZE (order
))
1571 size_lookup
[i
] = order
;
1575 /* Verify we got everything right with respect to alignment requests. */
1576 for (order
= 1; order
< 512; ++order
)
1577 gcc_assert (ffs (OBJECT_SIZE (size_lookup
[order
]))
1578 >= ffs (order
| MAX_ALIGNMENT
));
1582 G
.depth
= XNEWVEC (unsigned int, G
.depth_max
);
1584 G
.by_depth_in_use
= 0;
1585 G
.by_depth_max
= INITIAL_PTE_COUNT
;
1586 G
.by_depth
= XNEWVEC (page_entry
*, G
.by_depth_max
);
1587 G
.save_in_use
= XNEWVEC (unsigned long *, G
.by_depth_max
);
1590 /* Start a new GGC zone. */
1593 new_ggc_zone (const char *name ATTRIBUTE_UNUSED
)
1598 /* Destroy a GGC zone. */
1600 destroy_ggc_zone (struct alloc_zone
*zone ATTRIBUTE_UNUSED
)
1604 /* Merge the SAVE_IN_USE_P and IN_USE_P arrays in P so that IN_USE_P
1605 reflects reality. Recalculate NUM_FREE_OBJECTS as well. */
1608 ggc_recalculate_in_use_p (page_entry
*p
)
1613 /* Because the past-the-end bit in in_use_p is always set, we
1614 pretend there is one additional object. */
1615 num_objects
= OBJECTS_IN_PAGE (p
) + 1;
1617 /* Reset the free object count. */
1618 p
->num_free_objects
= num_objects
;
1620 /* Combine the IN_USE_P and SAVE_IN_USE_P arrays. */
1622 i
< CEIL (BITMAP_SIZE (num_objects
),
1623 sizeof (*p
->in_use_p
));
1628 /* Something is in use if it is marked, or if it was in use in a
1629 context further down the context stack. */
1630 p
->in_use_p
[i
] |= save_in_use_p (p
)[i
];
1632 /* Decrement the free object count for every object allocated. */
1633 for (j
= p
->in_use_p
[i
]; j
; j
>>= 1)
1634 p
->num_free_objects
-= (j
& 1);
1637 gcc_assert (p
->num_free_objects
< num_objects
);
1640 /* Unmark all objects. */
1647 for (order
= 2; order
< NUM_ORDERS
; order
++)
1651 for (p
= G
.pages
[order
]; p
!= NULL
; p
= p
->next
)
1653 size_t num_objects
= OBJECTS_IN_PAGE (p
);
1654 size_t bitmap_size
= BITMAP_SIZE (num_objects
+ 1);
1656 /* The data should be page-aligned. */
1657 gcc_assert (!((size_t) p
->page
& (G
.pagesize
- 1)));
1659 /* Pages that aren't in the topmost context are not collected;
1660 nevertheless, we need their in-use bit vectors to store GC
1661 marks. So, back them up first. */
1662 if (p
->context_depth
< G
.context_depth
)
1664 if (! save_in_use_p (p
))
1665 save_in_use_p (p
) = xmalloc (bitmap_size
);
1666 memcpy (save_in_use_p (p
), p
->in_use_p
, bitmap_size
);
1669 /* Reset reset the number of free objects and clear the
1670 in-use bits. These will be adjusted by mark_obj. */
1671 p
->num_free_objects
= num_objects
;
1672 memset (p
->in_use_p
, 0, bitmap_size
);
1674 /* Make sure the one-past-the-end bit is always set. */
1675 p
->in_use_p
[num_objects
/ HOST_BITS_PER_LONG
]
1676 = ((unsigned long) 1 << (num_objects
% HOST_BITS_PER_LONG
));
1681 /* Free all empty pages. Partially empty pages need no attention
1682 because the `mark' bit doubles as an `unused' bit. */
1689 for (order
= 2; order
< NUM_ORDERS
; order
++)
1691 /* The last page-entry to consider, regardless of entries
1692 placed at the end of the list. */
1693 page_entry
* const last
= G
.page_tails
[order
];
1696 size_t live_objects
;
1697 page_entry
*p
, *previous
;
1707 page_entry
*next
= p
->next
;
1709 /* Loop until all entries have been examined. */
1712 num_objects
= OBJECTS_IN_PAGE (p
);
1714 /* Add all live objects on this page to the count of
1715 allocated memory. */
1716 live_objects
= num_objects
- p
->num_free_objects
;
1718 G
.allocated
+= OBJECT_SIZE (order
) * live_objects
;
1720 /* Only objects on pages in the topmost context should get
1722 if (p
->context_depth
< G
.context_depth
)
1725 /* Remove the page if it's empty. */
1726 else if (live_objects
== 0)
1728 /* If P was the first page in the list, then NEXT
1729 becomes the new first page in the list, otherwise
1730 splice P out of the forward pointers. */
1732 G
.pages
[order
] = next
;
1734 previous
->next
= next
;
1736 /* Splice P out of the back pointers too. */
1738 next
->prev
= previous
;
1740 /* Are we removing the last element? */
1741 if (p
== G
.page_tails
[order
])
1742 G
.page_tails
[order
] = previous
;
1747 /* If the page is full, move it to the end. */
1748 else if (p
->num_free_objects
== 0)
1750 /* Don't move it if it's already at the end. */
1751 if (p
!= G
.page_tails
[order
])
1753 /* Move p to the end of the list. */
1755 p
->prev
= G
.page_tails
[order
];
1756 G
.page_tails
[order
]->next
= p
;
1758 /* Update the tail pointer... */
1759 G
.page_tails
[order
] = p
;
1761 /* ... and the head pointer, if necessary. */
1763 G
.pages
[order
] = next
;
1765 previous
->next
= next
;
1767 /* And update the backpointer in NEXT if necessary. */
1769 next
->prev
= previous
;
1775 /* If we've fallen through to here, it's a page in the
1776 topmost context that is neither full nor empty. Such a
1777 page must precede pages at lesser context depth in the
1778 list, so move it to the head. */
1779 else if (p
!= G
.pages
[order
])
1781 previous
->next
= p
->next
;
1783 /* Update the backchain in the next node if it exists. */
1785 p
->next
->prev
= previous
;
1787 /* Move P to the head of the list. */
1788 p
->next
= G
.pages
[order
];
1790 G
.pages
[order
]->prev
= p
;
1792 /* Update the head pointer. */
1795 /* Are we moving the last element? */
1796 if (G
.page_tails
[order
] == p
)
1797 G
.page_tails
[order
] = previous
;
1806 /* Now, restore the in_use_p vectors for any pages from contexts
1807 other than the current one. */
1808 for (p
= G
.pages
[order
]; p
; p
= p
->next
)
1809 if (p
->context_depth
!= G
.context_depth
)
1810 ggc_recalculate_in_use_p (p
);
1814 #ifdef ENABLE_GC_CHECKING
1815 /* Clobber all free objects. */
1822 for (order
= 2; order
< NUM_ORDERS
; order
++)
1824 size_t size
= OBJECT_SIZE (order
);
1827 for (p
= G
.pages
[order
]; p
!= NULL
; p
= p
->next
)
1832 if (p
->context_depth
!= G
.context_depth
)
1833 /* Since we don't do any collection for pages in pushed
1834 contexts, there's no need to do any poisoning. And
1835 besides, the IN_USE_P array isn't valid until we pop
1839 num_objects
= OBJECTS_IN_PAGE (p
);
1840 for (i
= 0; i
< num_objects
; i
++)
1843 word
= i
/ HOST_BITS_PER_LONG
;
1844 bit
= i
% HOST_BITS_PER_LONG
;
1845 if (((p
->in_use_p
[word
] >> bit
) & 1) == 0)
1847 char *object
= p
->page
+ i
* size
;
1849 /* Keep poison-by-write when we expect to use Valgrind,
1850 so the exact same memory semantics is kept, in case
1851 there are memory errors. We override this request
1853 VALGRIND_DISCARD (VALGRIND_MAKE_WRITABLE (object
, size
));
1854 memset (object
, 0xa5, size
);
1856 /* Drop the handle to avoid handle leak. */
1857 VALGRIND_DISCARD (VALGRIND_MAKE_NOACCESS (object
, size
));
1864 #define poison_pages()
1867 #ifdef ENABLE_GC_ALWAYS_COLLECT
1868 /* Validate that the reportedly free objects actually are. */
1871 validate_free_objects (void)
1873 struct free_object
*f
, *next
, *still_free
= NULL
;
1875 for (f
= G
.free_object_list
; f
; f
= next
)
1877 page_entry
*pe
= lookup_page_table_entry (f
->object
);
1880 bit
= OFFSET_TO_BIT ((char *)f
->object
- pe
->page
, pe
->order
);
1881 word
= bit
/ HOST_BITS_PER_LONG
;
1882 bit
= bit
% HOST_BITS_PER_LONG
;
1885 /* Make certain it isn't visible from any root. Notice that we
1886 do this check before sweep_pages merges save_in_use_p. */
1887 gcc_assert (!(pe
->in_use_p
[word
] & (1UL << bit
)));
1889 /* If the object comes from an outer context, then retain the
1890 free_object entry, so that we can verify that the address
1891 isn't live on the stack in some outer context. */
1892 if (pe
->context_depth
!= G
.context_depth
)
1894 f
->next
= still_free
;
1901 G
.free_object_list
= still_free
;
1904 #define validate_free_objects()
1907 /* Top level mark-and-sweep routine. */
1912 /* Avoid frequent unnecessary work by skipping collection if the
1913 total allocations haven't expanded much since the last
1915 float allocated_last_gc
=
1916 MAX (G
.allocated_last_gc
, (size_t)PARAM_VALUE (GGC_MIN_HEAPSIZE
) * 1024);
1918 float min_expand
= allocated_last_gc
* PARAM_VALUE (GGC_MIN_EXPAND
) / 100;
1920 if (G
.allocated
< allocated_last_gc
+ min_expand
&& !ggc_force_collect
)
1923 timevar_push (TV_GC
);
1925 fprintf (stderr
, " {GC %luk -> ", (unsigned long) G
.allocated
/ 1024);
1926 if (GGC_DEBUG_LEVEL
>= 2)
1927 fprintf (G
.debug_file
, "BEGIN COLLECTING\n");
1929 /* Zero the total allocated bytes. This will be recalculated in the
1933 /* Release the pages we freed the last time we collected, but didn't
1934 reuse in the interim. */
1937 /* Indicate that we've seen collections at this context depth. */
1938 G
.context_depth_collections
= ((unsigned long)1 << (G
.context_depth
+ 1)) - 1;
1942 #ifdef GATHER_STATISTICS
1943 ggc_prune_overhead_list ();
1946 validate_free_objects ();
1949 G
.allocated_last_gc
= G
.allocated
;
1951 timevar_pop (TV_GC
);
1954 fprintf (stderr
, "%luk}", (unsigned long) G
.allocated
/ 1024);
1955 if (GGC_DEBUG_LEVEL
>= 2)
1956 fprintf (G
.debug_file
, "END COLLECTING\n");
1959 /* Print allocation statistics. */
1960 #define SCALE(x) ((unsigned long) ((x) < 1024*10 \
1962 : ((x) < 1024*1024*10 \
1964 : (x) / (1024*1024))))
1965 #define STAT_LABEL(x) ((x) < 1024*10 ? ' ' : ((x) < 1024*1024*10 ? 'k' : 'M'))
1968 ggc_print_statistics (void)
1970 struct ggc_statistics stats
;
1972 size_t total_overhead
= 0;
1974 /* Clear the statistics. */
1975 memset (&stats
, 0, sizeof (stats
));
1977 /* Make sure collection will really occur. */
1978 G
.allocated_last_gc
= 0;
1980 /* Collect and print the statistics common across collectors. */
1981 ggc_print_common_statistics (stderr
, &stats
);
1983 /* Release free pages so that we will not count the bytes allocated
1984 there as part of the total allocated memory. */
1987 /* Collect some information about the various sizes of
1990 "Memory still allocated at the end of the compilation process\n");
1991 fprintf (stderr
, "%-5s %10s %10s %10s\n",
1992 "Size", "Allocated", "Used", "Overhead");
1993 for (i
= 0; i
< NUM_ORDERS
; ++i
)
2000 /* Skip empty entries. */
2004 overhead
= allocated
= in_use
= 0;
2006 /* Figure out the total number of bytes allocated for objects of
2007 this size, and how many of them are actually in use. Also figure
2008 out how much memory the page table is using. */
2009 for (p
= G
.pages
[i
]; p
; p
= p
->next
)
2011 allocated
+= p
->bytes
;
2013 (OBJECTS_IN_PAGE (p
) - p
->num_free_objects
) * OBJECT_SIZE (i
);
2015 overhead
+= (sizeof (page_entry
) - sizeof (long)
2016 + BITMAP_SIZE (OBJECTS_IN_PAGE (p
) + 1));
2018 fprintf (stderr
, "%-5lu %10lu%c %10lu%c %10lu%c\n",
2019 (unsigned long) OBJECT_SIZE (i
),
2020 SCALE (allocated
), STAT_LABEL (allocated
),
2021 SCALE (in_use
), STAT_LABEL (in_use
),
2022 SCALE (overhead
), STAT_LABEL (overhead
));
2023 total_overhead
+= overhead
;
2025 fprintf (stderr
, "%-5s %10lu%c %10lu%c %10lu%c\n", "Total",
2026 SCALE (G
.bytes_mapped
), STAT_LABEL (G
.bytes_mapped
),
2027 SCALE (G
.allocated
), STAT_LABEL(G
.allocated
),
2028 SCALE (total_overhead
), STAT_LABEL (total_overhead
));
2030 #ifdef GATHER_STATISTICS
2032 fprintf (stderr
, "\nTotal allocations and overheads during the compilation process\n");
2034 fprintf (stderr
, "Total Overhead: %10lld\n",
2035 G
.stats
.total_overhead
);
2036 fprintf (stderr
, "Total Allocated: %10lld\n",
2037 G
.stats
.total_allocated
);
2039 fprintf (stderr
, "Total Overhead under 32B: %10lld\n",
2040 G
.stats
.total_overhead_under32
);
2041 fprintf (stderr
, "Total Allocated under 32B: %10lld\n",
2042 G
.stats
.total_allocated_under32
);
2043 fprintf (stderr
, "Total Overhead under 64B: %10lld\n",
2044 G
.stats
.total_overhead_under64
);
2045 fprintf (stderr
, "Total Allocated under 64B: %10lld\n",
2046 G
.stats
.total_allocated_under64
);
2047 fprintf (stderr
, "Total Overhead under 128B: %10lld\n",
2048 G
.stats
.total_overhead_under128
);
2049 fprintf (stderr
, "Total Allocated under 128B: %10lld\n",
2050 G
.stats
.total_allocated_under128
);
2052 for (i
= 0; i
< NUM_ORDERS
; i
++)
2053 if (G
.stats
.total_allocated_per_order
[i
])
2055 fprintf (stderr
, "Total Overhead page size %7d: %10lld\n",
2056 OBJECT_SIZE (i
), G
.stats
.total_overhead_per_order
[i
]);
2057 fprintf (stderr
, "Total Allocated page size %7d: %10lld\n",
2058 OBJECT_SIZE (i
), G
.stats
.total_allocated_per_order
[i
]);
2066 struct ggc_pch_ondisk
2068 unsigned totals
[NUM_ORDERS
];
2070 size_t base
[NUM_ORDERS
];
2071 size_t written
[NUM_ORDERS
];
2074 struct ggc_pch_data
*
2077 return XCNEW (struct ggc_pch_data
);
2081 ggc_pch_count_object (struct ggc_pch_data
*d
, void *x ATTRIBUTE_UNUSED
,
2082 size_t size
, bool is_string ATTRIBUTE_UNUSED
,
2083 enum gt_types_enum type ATTRIBUTE_UNUSED
)
2088 order
= size_lookup
[size
];
2092 while (size
> OBJECT_SIZE (order
))
2096 d
->d
.totals
[order
]++;
2100 ggc_pch_total_size (struct ggc_pch_data
*d
)
2105 for (i
= 0; i
< NUM_ORDERS
; i
++)
2106 a
+= ROUND_UP (d
->d
.totals
[i
] * OBJECT_SIZE (i
), G
.pagesize
);
2111 ggc_pch_this_base (struct ggc_pch_data
*d
, void *base
)
2113 size_t a
= (size_t) base
;
2116 for (i
= 0; i
< NUM_ORDERS
; i
++)
2119 a
+= ROUND_UP (d
->d
.totals
[i
] * OBJECT_SIZE (i
), G
.pagesize
);
2125 ggc_pch_alloc_object (struct ggc_pch_data
*d
, void *x ATTRIBUTE_UNUSED
,
2126 size_t size
, bool is_string ATTRIBUTE_UNUSED
,
2127 enum gt_types_enum type ATTRIBUTE_UNUSED
)
2133 order
= size_lookup
[size
];
2137 while (size
> OBJECT_SIZE (order
))
2141 result
= (char *) d
->base
[order
];
2142 d
->base
[order
] += OBJECT_SIZE (order
);
2147 ggc_pch_prepare_write (struct ggc_pch_data
*d ATTRIBUTE_UNUSED
,
2148 FILE *f ATTRIBUTE_UNUSED
)
2150 /* Nothing to do. */
2154 ggc_pch_write_object (struct ggc_pch_data
*d ATTRIBUTE_UNUSED
,
2155 FILE *f
, void *x
, void *newx ATTRIBUTE_UNUSED
,
2156 size_t size
, bool is_string ATTRIBUTE_UNUSED
)
2159 static const char emptyBytes
[256];
2162 order
= size_lookup
[size
];
2166 while (size
> OBJECT_SIZE (order
))
2170 if (fwrite (x
, size
, 1, f
) != 1)
2171 fatal_error ("can't write PCH file: %m");
2173 /* If SIZE is not the same as OBJECT_SIZE(order), then we need to pad the
2174 object out to OBJECT_SIZE(order). This happens for strings. */
2176 if (size
!= OBJECT_SIZE (order
))
2178 unsigned padding
= OBJECT_SIZE(order
) - size
;
2180 /* To speed small writes, we use a nulled-out array that's larger
2181 than most padding requests as the source for our null bytes. This
2182 permits us to do the padding with fwrite() rather than fseek(), and
2183 limits the chance the OS may try to flush any outstanding writes. */
2184 if (padding
<= sizeof(emptyBytes
))
2186 if (fwrite (emptyBytes
, 1, padding
, f
) != padding
)
2187 fatal_error ("can't write PCH file");
2191 /* Larger than our buffer? Just default to fseek. */
2192 if (fseek (f
, padding
, SEEK_CUR
) != 0)
2193 fatal_error ("can't write PCH file");
2197 d
->written
[order
]++;
2198 if (d
->written
[order
] == d
->d
.totals
[order
]
2199 && fseek (f
, ROUND_UP_VALUE (d
->d
.totals
[order
] * OBJECT_SIZE (order
),
2202 fatal_error ("can't write PCH file: %m");
2206 ggc_pch_finish (struct ggc_pch_data
*d
, FILE *f
)
2208 if (fwrite (&d
->d
, sizeof (d
->d
), 1, f
) != 1)
2209 fatal_error ("can't write PCH file: %m");
2213 /* Move the PCH PTE entries just added to the end of by_depth, to the
2217 move_ptes_to_front (int count_old_page_tables
, int count_new_page_tables
)
2221 /* First, we swap the new entries to the front of the varrays. */
2222 page_entry
**new_by_depth
;
2223 unsigned long **new_save_in_use
;
2225 new_by_depth
= XNEWVEC (page_entry
*, G
.by_depth_max
);
2226 new_save_in_use
= XNEWVEC (unsigned long *, G
.by_depth_max
);
2228 memcpy (&new_by_depth
[0],
2229 &G
.by_depth
[count_old_page_tables
],
2230 count_new_page_tables
* sizeof (void *));
2231 memcpy (&new_by_depth
[count_new_page_tables
],
2233 count_old_page_tables
* sizeof (void *));
2234 memcpy (&new_save_in_use
[0],
2235 &G
.save_in_use
[count_old_page_tables
],
2236 count_new_page_tables
* sizeof (void *));
2237 memcpy (&new_save_in_use
[count_new_page_tables
],
2239 count_old_page_tables
* sizeof (void *));
2242 free (G
.save_in_use
);
2244 G
.by_depth
= new_by_depth
;
2245 G
.save_in_use
= new_save_in_use
;
2247 /* Now update all the index_by_depth fields. */
2248 for (i
= G
.by_depth_in_use
; i
> 0; --i
)
2250 page_entry
*p
= G
.by_depth
[i
-1];
2251 p
->index_by_depth
= i
-1;
2254 /* And last, we update the depth pointers in G.depth. The first
2255 entry is already 0, and context 0 entries always start at index
2256 0, so there is nothing to update in the first slot. We need a
2257 second slot, only if we have old ptes, and if we do, they start
2258 at index count_new_page_tables. */
2259 if (count_old_page_tables
)
2260 push_depth (count_new_page_tables
);
2264 ggc_pch_read (FILE *f
, void *addr
)
2266 struct ggc_pch_ondisk d
;
2269 unsigned long count_old_page_tables
;
2270 unsigned long count_new_page_tables
;
2272 count_old_page_tables
= G
.by_depth_in_use
;
2274 /* We've just read in a PCH file. So, every object that used to be
2275 allocated is now free. */
2277 #ifdef ENABLE_GC_CHECKING
2281 /* No object read from a PCH file should ever be freed. So, set the
2282 context depth to 1, and set the depth of all the currently-allocated
2283 pages to be 1 too. PCH pages will have depth 0. */
2284 gcc_assert (!G
.context_depth
);
2285 G
.context_depth
= 1;
2286 for (i
= 0; i
< NUM_ORDERS
; i
++)
2289 for (p
= G
.pages
[i
]; p
!= NULL
; p
= p
->next
)
2290 p
->context_depth
= G
.context_depth
;
2293 /* Allocate the appropriate page-table entries for the pages read from
2295 if (fread (&d
, sizeof (d
), 1, f
) != 1)
2296 fatal_error ("can't read PCH file: %m");
2298 for (i
= 0; i
< NUM_ORDERS
; i
++)
2300 struct page_entry
*entry
;
2306 if (d
.totals
[i
] == 0)
2309 bytes
= ROUND_UP (d
.totals
[i
] * OBJECT_SIZE (i
), G
.pagesize
);
2310 num_objs
= bytes
/ OBJECT_SIZE (i
);
2311 entry
= xcalloc (1, (sizeof (struct page_entry
)
2313 + BITMAP_SIZE (num_objs
+ 1)));
2314 entry
->bytes
= bytes
;
2316 entry
->context_depth
= 0;
2318 entry
->num_free_objects
= 0;
2322 j
+ HOST_BITS_PER_LONG
<= num_objs
+ 1;
2323 j
+= HOST_BITS_PER_LONG
)
2324 entry
->in_use_p
[j
/ HOST_BITS_PER_LONG
] = -1;
2325 for (; j
< num_objs
+ 1; j
++)
2326 entry
->in_use_p
[j
/ HOST_BITS_PER_LONG
]
2327 |= 1L << (j
% HOST_BITS_PER_LONG
);
2329 for (pte
= entry
->page
;
2330 pte
< entry
->page
+ entry
->bytes
;
2332 set_page_table_entry (pte
, entry
);
2334 if (G
.page_tails
[i
] != NULL
)
2335 G
.page_tails
[i
]->next
= entry
;
2338 G
.page_tails
[i
] = entry
;
2340 /* We start off by just adding all the new information to the
2341 end of the varrays, later, we will move the new information
2342 to the front of the varrays, as the PCH page tables are at
2344 push_by_depth (entry
, 0);
2347 /* Now, we update the various data structures that speed page table
2349 count_new_page_tables
= G
.by_depth_in_use
- count_old_page_tables
;
2351 move_ptes_to_front (count_old_page_tables
, count_new_page_tables
);
2353 /* Update the statistics. */
2354 G
.allocated
= G
.allocated_last_gc
= offs
- (char *)addr
;