Mercurial > hg > CbC > CbC_gcc
annotate gcc/ggc-page.c @ 67:f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
author | nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Tue, 22 Mar 2011 17:18:12 +0900 |
parents | 77e2b8dfacca |
children | 04ced10e8804 |
rev | line source |
---|---|
0 | 1 /* "Bag-of-pages" garbage collector for the GNU compiler. |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2 Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008, 2009, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
3 2010 Free Software Foundation, Inc. |
0 | 4 |
5 This file is part of GCC. | |
6 | |
7 GCC is free software; you can redistribute it and/or modify it under | |
8 the terms of the GNU General Public License as published by the Free | |
9 Software Foundation; either version 3, or (at your option) any later | |
10 version. | |
11 | |
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 for more details. | |
16 | |
17 You should have received a copy of the GNU General Public License | |
18 along with GCC; see the file COPYING3. If not see | |
19 <http://www.gnu.org/licenses/>. */ | |
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" | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
28 #include "diagnostic-core.h" |
0 | 29 #include "flags.h" |
30 #include "ggc.h" | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
31 #include "ggc-internal.h" |
0 | 32 #include "timevar.h" |
33 #include "params.h" | |
34 #include "tree-flow.h" | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
35 #include "cfgloop.h" |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
36 #include "plugin.h" |
0 | 37 |
38 /* Prefer MAP_ANON(YMOUS) to /dev/zero, since we don't need to keep a | |
39 file open. Prefer either to valloc. */ | |
40 #ifdef HAVE_MMAP_ANON | |
41 # undef HAVE_MMAP_DEV_ZERO | |
42 # define USING_MMAP | |
43 #endif | |
44 | |
45 #ifdef HAVE_MMAP_DEV_ZERO | |
46 # define USING_MMAP | |
47 #endif | |
48 | |
49 #ifndef USING_MMAP | |
50 #define USING_MALLOC_PAGE_GROUPS | |
51 #endif | |
52 | |
53 /* Strategy: | |
54 | |
55 This garbage-collecting allocator allocates objects on one of a set | |
56 of pages. Each page can allocate objects of a single size only; | |
57 available sizes are powers of two starting at four bytes. The size | |
58 of an allocation request is rounded up to the next power of two | |
59 (`order'), and satisfied from the appropriate page. | |
60 | |
61 Each page is recorded in a page-entry, which also maintains an | |
62 in-use bitmap of object positions on the page. This allows the | |
63 allocation state of a particular object to be flipped without | |
64 touching the page itself. | |
65 | |
66 Each page-entry also has a context depth, which is used to track | |
67 pushing and popping of allocation contexts. Only objects allocated | |
68 in the current (highest-numbered) context may be collected. | |
69 | |
70 Page entries are arranged in an array of singly-linked lists. The | |
71 array is indexed by the allocation size, in bits, of the pages on | |
72 it; i.e. all pages on a list allocate objects of the same size. | |
73 Pages are ordered on the list such that all non-full pages precede | |
74 all full pages, with non-full pages arranged in order of decreasing | |
75 context depth. | |
76 | |
77 Empty pages (of all orders) are kept on a single page cache list, | |
78 and are considered first when new pages are required; they are | |
79 deallocated at the start of the next collection if they haven't | |
80 been recycled by then. */ | |
81 | |
82 /* Define GGC_DEBUG_LEVEL to print debugging information. | |
83 0: No debugging output. | |
84 1: GC statistics only. | |
85 2: Page-entry allocations/deallocations as well. | |
86 3: Object allocations as well. | |
87 4: Object marks as well. */ | |
88 #define GGC_DEBUG_LEVEL (0) | |
89 | |
90 #ifndef HOST_BITS_PER_PTR | |
91 #define HOST_BITS_PER_PTR HOST_BITS_PER_LONG | |
92 #endif | |
93 | |
94 | |
95 /* A two-level tree is used to look up the page-entry for a given | |
96 pointer. Two chunks of the pointer's bits are extracted to index | |
97 the first and second levels of the tree, as follows: | |
98 | |
99 HOST_PAGE_SIZE_BITS | |
100 32 | | | |
101 msb +----------------+----+------+------+ lsb | |
102 | | | | |
103 PAGE_L1_BITS | | |
104 | | | |
105 PAGE_L2_BITS | |
106 | |
107 The bottommost HOST_PAGE_SIZE_BITS are ignored, since page-entry | |
108 pages are aligned on system page boundaries. The next most | |
109 significant PAGE_L2_BITS and PAGE_L1_BITS are the second and first | |
110 index values in the lookup table, respectively. | |
111 | |
112 For 32-bit architectures and the settings below, there are no | |
113 leftover bits. For architectures with wider pointers, the lookup | |
114 tree points to a list of pages, which must be scanned to find the | |
115 correct one. */ | |
116 | |
117 #define PAGE_L1_BITS (8) | |
118 #define PAGE_L2_BITS (32 - PAGE_L1_BITS - G.lg_pagesize) | |
119 #define PAGE_L1_SIZE ((size_t) 1 << PAGE_L1_BITS) | |
120 #define PAGE_L2_SIZE ((size_t) 1 << PAGE_L2_BITS) | |
121 | |
122 #define LOOKUP_L1(p) \ | |
123 (((size_t) (p) >> (32 - PAGE_L1_BITS)) & ((1 << PAGE_L1_BITS) - 1)) | |
124 | |
125 #define LOOKUP_L2(p) \ | |
126 (((size_t) (p) >> G.lg_pagesize) & ((1 << PAGE_L2_BITS) - 1)) | |
127 | |
128 /* The number of objects per allocation page, for objects on a page of | |
129 the indicated ORDER. */ | |
130 #define OBJECTS_PER_PAGE(ORDER) objects_per_page_table[ORDER] | |
131 | |
132 /* The number of objects in P. */ | |
133 #define OBJECTS_IN_PAGE(P) ((P)->bytes / OBJECT_SIZE ((P)->order)) | |
134 | |
135 /* The size of an object on a page of the indicated ORDER. */ | |
136 #define OBJECT_SIZE(ORDER) object_size_table[ORDER] | |
137 | |
138 /* For speed, we avoid doing a general integer divide to locate the | |
139 offset in the allocation bitmap, by precalculating numbers M, S | |
140 such that (O * M) >> S == O / Z (modulo 2^32), for any offset O | |
141 within the page which is evenly divisible by the object size Z. */ | |
142 #define DIV_MULT(ORDER) inverse_table[ORDER].mult | |
143 #define DIV_SHIFT(ORDER) inverse_table[ORDER].shift | |
144 #define OFFSET_TO_BIT(OFFSET, ORDER) \ | |
145 (((OFFSET) * DIV_MULT (ORDER)) >> DIV_SHIFT (ORDER)) | |
146 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
147 /* We use this structure to determine the alignment required for |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
148 allocations. For power-of-two sized allocations, that's not a |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
149 problem, but it does matter for odd-sized allocations. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
150 We do not care about alignment for floating-point types. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
151 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
152 struct max_alignment { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
153 char c; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
154 union { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
155 HOST_WIDEST_INT i; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
156 void *p; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
157 } u; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
158 }; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
159 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
160 /* The biggest alignment required. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
161 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
162 #define MAX_ALIGNMENT (offsetof (struct max_alignment, u)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
163 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
164 |
0 | 165 /* The number of extra orders, not corresponding to power-of-two sized |
166 objects. */ | |
167 | |
168 #define NUM_EXTRA_ORDERS ARRAY_SIZE (extra_order_size_table) | |
169 | |
170 #define RTL_SIZE(NSLOTS) \ | |
171 (RTX_HDR_SIZE + (NSLOTS) * sizeof (rtunion)) | |
172 | |
173 #define TREE_EXP_SIZE(OPS) \ | |
174 (sizeof (struct tree_exp) + ((OPS) - 1) * sizeof (tree)) | |
175 | |
176 /* The Ith entry is the maximum size of an object to be stored in the | |
177 Ith extra order. Adding a new entry to this array is the *only* | |
178 thing you need to do to add a new special allocation size. */ | |
179 | |
180 static const size_t extra_order_size_table[] = { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
181 /* Extra orders for small non-power-of-two multiples of MAX_ALIGNMENT. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
182 There are a lot of structures with these sizes and explicitly |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
183 listing them risks orders being dropped because they changed size. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
184 MAX_ALIGNMENT * 3, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
185 MAX_ALIGNMENT * 5, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
186 MAX_ALIGNMENT * 6, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
187 MAX_ALIGNMENT * 7, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
188 MAX_ALIGNMENT * 9, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
189 MAX_ALIGNMENT * 10, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
190 MAX_ALIGNMENT * 11, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
191 MAX_ALIGNMENT * 12, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
192 MAX_ALIGNMENT * 13, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
193 MAX_ALIGNMENT * 14, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
194 MAX_ALIGNMENT * 15, |
0 | 195 sizeof (struct tree_decl_non_common), |
196 sizeof (struct tree_field_decl), | |
197 sizeof (struct tree_parm_decl), | |
198 sizeof (struct tree_var_decl), | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
199 sizeof (struct tree_type), |
0 | 200 sizeof (struct function), |
201 sizeof (struct basic_block_def), | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
202 sizeof (struct cgraph_node), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
203 sizeof (struct loop), |
0 | 204 }; |
205 | |
206 /* The total number of orders. */ | |
207 | |
208 #define NUM_ORDERS (HOST_BITS_PER_PTR + NUM_EXTRA_ORDERS) | |
209 | |
210 /* Compute the smallest nonnegative number which when added to X gives | |
211 a multiple of F. */ | |
212 | |
213 #define ROUND_UP_VALUE(x, f) ((f) - 1 - ((f) - 1 + (x)) % (f)) | |
214 | |
215 /* Compute the smallest multiple of F that is >= X. */ | |
216 | |
217 #define ROUND_UP(x, f) (CEIL (x, f) * (f)) | |
218 | |
219 /* The Ith entry is the number of objects on a page or order I. */ | |
220 | |
221 static unsigned objects_per_page_table[NUM_ORDERS]; | |
222 | |
223 /* The Ith entry is the size of an object on a page of order I. */ | |
224 | |
225 static size_t object_size_table[NUM_ORDERS]; | |
226 | |
227 /* The Ith entry is a pair of numbers (mult, shift) such that | |
228 ((k * mult) >> shift) mod 2^32 == (k / OBJECT_SIZE(I)) mod 2^32, | |
229 for all k evenly divisible by OBJECT_SIZE(I). */ | |
230 | |
231 static struct | |
232 { | |
233 size_t mult; | |
234 unsigned int shift; | |
235 } | |
236 inverse_table[NUM_ORDERS]; | |
237 | |
238 /* A page_entry records the status of an allocation page. This | |
239 structure is dynamically sized to fit the bitmap in_use_p. */ | |
240 typedef struct page_entry | |
241 { | |
242 /* The next page-entry with objects of the same size, or NULL if | |
243 this is the last page-entry. */ | |
244 struct page_entry *next; | |
245 | |
246 /* The previous page-entry with objects of the same size, or NULL if | |
247 this is the first page-entry. The PREV pointer exists solely to | |
248 keep the cost of ggc_free manageable. */ | |
249 struct page_entry *prev; | |
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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
324 #ifdef ENABLE_GC_ALWAYS_COLLECT |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
325 /* List of free objects to be verified as actually free on the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
326 next collection. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
327 struct free_object |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
328 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
329 void *object; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
330 struct free_object *next; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
331 }; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
332 #endif |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
333 |
0 | 334 /* The rest of the global variables. */ |
335 static struct globals | |
336 { | |
337 /* The Nth element in this array is a page with objects of size 2^N. | |
338 If there are any pages with free objects, they will be at the | |
339 head of the list. NULL if there are no page-entries for this | |
340 object size. */ | |
341 page_entry *pages[NUM_ORDERS]; | |
342 | |
343 /* The Nth element in this array is the last page with objects of | |
344 size 2^N. NULL if there are no page-entries for this object | |
345 size. */ | |
346 page_entry *page_tails[NUM_ORDERS]; | |
347 | |
348 /* Lookup table for associating allocation pages with object addresses. */ | |
349 page_table lookup; | |
350 | |
351 /* The system's page size. */ | |
352 size_t pagesize; | |
353 size_t lg_pagesize; | |
354 | |
355 /* Bytes currently allocated. */ | |
356 size_t allocated; | |
357 | |
358 /* Bytes currently allocated at the end of the last collection. */ | |
359 size_t allocated_last_gc; | |
360 | |
361 /* Total amount of memory mapped. */ | |
362 size_t bytes_mapped; | |
363 | |
364 /* Bit N set if any allocations have been done at context depth N. */ | |
365 unsigned long context_depth_allocations; | |
366 | |
367 /* Bit N set if any collections have been done at context depth N. */ | |
368 unsigned long context_depth_collections; | |
369 | |
370 /* The current depth in the context stack. */ | |
371 unsigned short context_depth; | |
372 | |
373 /* A file descriptor open to /dev/zero for reading. */ | |
374 #if defined (HAVE_MMAP_DEV_ZERO) | |
375 int dev_zero_fd; | |
376 #endif | |
377 | |
378 /* A cache of free system pages. */ | |
379 page_entry *free_pages; | |
380 | |
381 #ifdef USING_MALLOC_PAGE_GROUPS | |
382 page_group *page_groups; | |
383 #endif | |
384 | |
385 /* The file descriptor for debugging output. */ | |
386 FILE *debug_file; | |
387 | |
388 /* Current number of elements in use in depth below. */ | |
389 unsigned int depth_in_use; | |
390 | |
391 /* Maximum number of elements that can be used before resizing. */ | |
392 unsigned int depth_max; | |
393 | |
394 /* Each element of this array is an index in by_depth where the given | |
395 depth starts. This structure is indexed by that given depth we | |
396 are interested in. */ | |
397 unsigned int *depth; | |
398 | |
399 /* Current number of elements in use in by_depth below. */ | |
400 unsigned int by_depth_in_use; | |
401 | |
402 /* Maximum number of elements that can be used before resizing. */ | |
403 unsigned int by_depth_max; | |
404 | |
405 /* Each element of this array is a pointer to a page_entry, all | |
406 page_entries can be found in here by increasing depth. | |
407 index_by_depth in the page_entry is the index into this data | |
408 structure where that page_entry can be found. This is used to | |
409 speed up finding all page_entries at a particular depth. */ | |
410 page_entry **by_depth; | |
411 | |
412 /* Each element is a pointer to the saved in_use_p bits, if any, | |
413 zero otherwise. We allocate them all together, to enable a | |
414 better runtime data access pattern. */ | |
415 unsigned long **save_in_use; | |
416 | |
417 #ifdef ENABLE_GC_ALWAYS_COLLECT | |
418 /* List of free objects to be verified as actually free on the | |
419 next collection. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
420 struct free_object *free_object_list; |
0 | 421 #endif |
422 | |
423 #ifdef GATHER_STATISTICS | |
424 struct | |
425 { | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
426 /* Total GC-allocated memory. */ |
0 | 427 unsigned long long total_allocated; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
428 /* Total overhead for GC-allocated memory. */ |
0 | 429 unsigned long long total_overhead; |
430 | |
431 /* Total allocations and overhead for sizes less than 32, 64 and 128. | |
432 These sizes are interesting because they are typical cache line | |
433 sizes. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
434 |
0 | 435 unsigned long long total_allocated_under32; |
436 unsigned long long total_overhead_under32; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
437 |
0 | 438 unsigned long long total_allocated_under64; |
439 unsigned long long total_overhead_under64; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
440 |
0 | 441 unsigned long long total_allocated_under128; |
442 unsigned long long total_overhead_under128; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
443 |
0 | 444 /* The allocations for each of the allocation orders. */ |
445 unsigned long long total_allocated_per_order[NUM_ORDERS]; | |
446 | |
447 /* The overhead for each of the allocation orders. */ | |
448 unsigned long long total_overhead_per_order[NUM_ORDERS]; | |
449 } stats; | |
450 #endif | |
451 } G; | |
452 | |
453 /* The size in bytes required to maintain a bitmap for the objects | |
454 on a page-entry. */ | |
455 #define BITMAP_SIZE(Num_objects) \ | |
456 (CEIL ((Num_objects), HOST_BITS_PER_LONG) * sizeof(long)) | |
457 | |
458 /* Allocate pages in chunks of this size, to throttle calls to memory | |
459 allocation routines. The first page is used, the rest go onto the | |
460 free list. This cannot be larger than HOST_BITS_PER_INT for the | |
461 in_use bitmask for page_group. Hosts that need a different value | |
462 can override this by defining GGC_QUIRE_SIZE explicitly. */ | |
463 #ifndef GGC_QUIRE_SIZE | |
464 # ifdef USING_MMAP | |
465 # define GGC_QUIRE_SIZE 256 | |
466 # else | |
467 # define GGC_QUIRE_SIZE 16 | |
468 # endif | |
469 #endif | |
470 | |
471 /* Initial guess as to how many page table entries we might need. */ | |
472 #define INITIAL_PTE_COUNT 128 | |
473 | |
474 static int ggc_allocated_p (const void *); | |
475 static page_entry *lookup_page_table_entry (const void *); | |
476 static void set_page_table_entry (void *, page_entry *); | |
477 #ifdef USING_MMAP | |
478 static char *alloc_anon (char *, size_t); | |
479 #endif | |
480 #ifdef USING_MALLOC_PAGE_GROUPS | |
481 static size_t page_group_index (char *, char *); | |
482 static void set_page_group_in_use (page_group *, char *); | |
483 static void clear_page_group_in_use (page_group *, char *); | |
484 #endif | |
485 static struct page_entry * alloc_page (unsigned); | |
486 static void free_page (struct page_entry *); | |
487 static void release_pages (void); | |
488 static void clear_marks (void); | |
489 static void sweep_pages (void); | |
490 static void ggc_recalculate_in_use_p (page_entry *); | |
491 static void compute_inverse (unsigned); | |
492 static inline void adjust_depth (void); | |
493 static void move_ptes_to_front (int, int); | |
494 | |
495 void debug_print_page_list (int); | |
496 static void push_depth (unsigned int); | |
497 static void push_by_depth (page_entry *, unsigned long *); | |
498 | |
499 /* Push an entry onto G.depth. */ | |
500 | |
501 inline static void | |
502 push_depth (unsigned int i) | |
503 { | |
504 if (G.depth_in_use >= G.depth_max) | |
505 { | |
506 G.depth_max *= 2; | |
507 G.depth = XRESIZEVEC (unsigned int, G.depth, G.depth_max); | |
508 } | |
509 G.depth[G.depth_in_use++] = i; | |
510 } | |
511 | |
512 /* Push an entry onto G.by_depth and G.save_in_use. */ | |
513 | |
514 inline static void | |
515 push_by_depth (page_entry *p, unsigned long *s) | |
516 { | |
517 if (G.by_depth_in_use >= G.by_depth_max) | |
518 { | |
519 G.by_depth_max *= 2; | |
520 G.by_depth = XRESIZEVEC (page_entry *, G.by_depth, G.by_depth_max); | |
521 G.save_in_use = XRESIZEVEC (unsigned long *, G.save_in_use, | |
522 G.by_depth_max); | |
523 } | |
524 G.by_depth[G.by_depth_in_use] = p; | |
525 G.save_in_use[G.by_depth_in_use++] = s; | |
526 } | |
527 | |
528 #if (GCC_VERSION < 3001) | |
529 #define prefetch(X) ((void) X) | |
530 #else | |
531 #define prefetch(X) __builtin_prefetch (X) | |
532 #endif | |
533 | |
534 #define save_in_use_p_i(__i) \ | |
535 (G.save_in_use[__i]) | |
536 #define save_in_use_p(__p) \ | |
537 (save_in_use_p_i (__p->index_by_depth)) | |
538 | |
539 /* Returns nonzero if P was allocated in GC'able memory. */ | |
540 | |
541 static inline int | |
542 ggc_allocated_p (const void *p) | |
543 { | |
544 page_entry ***base; | |
545 size_t L1, L2; | |
546 | |
547 #if HOST_BITS_PER_PTR <= 32 | |
548 base = &G.lookup[0]; | |
549 #else | |
550 page_table table = G.lookup; | |
551 size_t high_bits = (size_t) p & ~ (size_t) 0xffffffff; | |
552 while (1) | |
553 { | |
554 if (table == NULL) | |
555 return 0; | |
556 if (table->high_bits == high_bits) | |
557 break; | |
558 table = table->next; | |
559 } | |
560 base = &table->table[0]; | |
561 #endif | |
562 | |
563 /* Extract the level 1 and 2 indices. */ | |
564 L1 = LOOKUP_L1 (p); | |
565 L2 = LOOKUP_L2 (p); | |
566 | |
567 return base[L1] && base[L1][L2]; | |
568 } | |
569 | |
570 /* Traverse the page table and find the entry for a page. | |
571 Die (probably) if the object wasn't allocated via GC. */ | |
572 | |
573 static inline page_entry * | |
574 lookup_page_table_entry (const void *p) | |
575 { | |
576 page_entry ***base; | |
577 size_t L1, L2; | |
578 | |
579 #if HOST_BITS_PER_PTR <= 32 | |
580 base = &G.lookup[0]; | |
581 #else | |
582 page_table table = G.lookup; | |
583 size_t high_bits = (size_t) p & ~ (size_t) 0xffffffff; | |
584 while (table->high_bits != high_bits) | |
585 table = table->next; | |
586 base = &table->table[0]; | |
587 #endif | |
588 | |
589 /* Extract the level 1 and 2 indices. */ | |
590 L1 = LOOKUP_L1 (p); | |
591 L2 = LOOKUP_L2 (p); | |
592 | |
593 return base[L1][L2]; | |
594 } | |
595 | |
596 /* Set the page table entry for a page. */ | |
597 | |
598 static void | |
599 set_page_table_entry (void *p, page_entry *entry) | |
600 { | |
601 page_entry ***base; | |
602 size_t L1, L2; | |
603 | |
604 #if HOST_BITS_PER_PTR <= 32 | |
605 base = &G.lookup[0]; | |
606 #else | |
607 page_table table; | |
608 size_t high_bits = (size_t) p & ~ (size_t) 0xffffffff; | |
609 for (table = G.lookup; table; table = table->next) | |
610 if (table->high_bits == high_bits) | |
611 goto found; | |
612 | |
613 /* Not found -- allocate a new table. */ | |
614 table = XCNEW (struct page_table_chain); | |
615 table->next = G.lookup; | |
616 table->high_bits = high_bits; | |
617 G.lookup = table; | |
618 found: | |
619 base = &table->table[0]; | |
620 #endif | |
621 | |
622 /* Extract the level 1 and 2 indices. */ | |
623 L1 = LOOKUP_L1 (p); | |
624 L2 = LOOKUP_L2 (p); | |
625 | |
626 if (base[L1] == NULL) | |
627 base[L1] = XCNEWVEC (page_entry *, PAGE_L2_SIZE); | |
628 | |
629 base[L1][L2] = entry; | |
630 } | |
631 | |
632 /* Prints the page-entry for object size ORDER, for debugging. */ | |
633 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
634 DEBUG_FUNCTION void |
0 | 635 debug_print_page_list (int order) |
636 { | |
637 page_entry *p; | |
638 printf ("Head=%p, Tail=%p:\n", (void *) G.pages[order], | |
639 (void *) G.page_tails[order]); | |
640 p = G.pages[order]; | |
641 while (p != NULL) | |
642 { | |
643 printf ("%p(%1d|%3d) -> ", (void *) p, p->context_depth, | |
644 p->num_free_objects); | |
645 p = p->next; | |
646 } | |
647 printf ("NULL\n"); | |
648 fflush (stdout); | |
649 } | |
650 | |
651 #ifdef USING_MMAP | |
652 /* Allocate SIZE bytes of anonymous memory, preferably near PREF, | |
653 (if non-null). The ifdef structure here is intended to cause a | |
654 compile error unless exactly one of the HAVE_* is defined. */ | |
655 | |
656 static inline char * | |
657 alloc_anon (char *pref ATTRIBUTE_UNUSED, size_t size) | |
658 { | |
659 #ifdef HAVE_MMAP_ANON | |
660 char *page = (char *) mmap (pref, size, PROT_READ | PROT_WRITE, | |
661 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); | |
662 #endif | |
663 #ifdef HAVE_MMAP_DEV_ZERO | |
664 char *page = (char *) mmap (pref, size, PROT_READ | PROT_WRITE, | |
665 MAP_PRIVATE, G.dev_zero_fd, 0); | |
666 #endif | |
667 | |
668 if (page == (char *) MAP_FAILED) | |
669 { | |
670 perror ("virtual memory exhausted"); | |
671 exit (FATAL_EXIT_CODE); | |
672 } | |
673 | |
674 /* Remember that we allocated this memory. */ | |
675 G.bytes_mapped += size; | |
676 | |
677 /* Pretend we don't have access to the allocated pages. We'll enable | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
678 access to smaller pieces of the area in ggc_internal_alloc. Discard the |
0 | 679 handle to avoid handle leak. */ |
680 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (page, size)); | |
681 | |
682 return page; | |
683 } | |
684 #endif | |
685 #ifdef USING_MALLOC_PAGE_GROUPS | |
686 /* Compute the index for this page into the page group. */ | |
687 | |
688 static inline size_t | |
689 page_group_index (char *allocation, char *page) | |
690 { | |
691 return (size_t) (page - allocation) >> G.lg_pagesize; | |
692 } | |
693 | |
694 /* Set and clear the in_use bit for this page in the page group. */ | |
695 | |
696 static inline void | |
697 set_page_group_in_use (page_group *group, char *page) | |
698 { | |
699 group->in_use |= 1 << page_group_index (group->allocation, page); | |
700 } | |
701 | |
702 static inline void | |
703 clear_page_group_in_use (page_group *group, char *page) | |
704 { | |
705 group->in_use &= ~(1 << page_group_index (group->allocation, page)); | |
706 } | |
707 #endif | |
708 | |
709 /* Allocate a new page for allocating objects of size 2^ORDER, | |
710 and return an entry for it. The entry is not added to the | |
711 appropriate page_table list. */ | |
712 | |
713 static inline struct page_entry * | |
714 alloc_page (unsigned order) | |
715 { | |
716 struct page_entry *entry, *p, **pp; | |
717 char *page; | |
718 size_t num_objects; | |
719 size_t bitmap_size; | |
720 size_t page_entry_size; | |
721 size_t entry_size; | |
722 #ifdef USING_MALLOC_PAGE_GROUPS | |
723 page_group *group; | |
724 #endif | |
725 | |
726 num_objects = OBJECTS_PER_PAGE (order); | |
727 bitmap_size = BITMAP_SIZE (num_objects + 1); | |
728 page_entry_size = sizeof (page_entry) - sizeof (long) + bitmap_size; | |
729 entry_size = num_objects * OBJECT_SIZE (order); | |
730 if (entry_size < G.pagesize) | |
731 entry_size = G.pagesize; | |
732 | |
733 entry = NULL; | |
734 page = NULL; | |
735 | |
736 /* Check the list of free pages for one we can use. */ | |
737 for (pp = &G.free_pages, p = *pp; p; pp = &p->next, p = *pp) | |
738 if (p->bytes == entry_size) | |
739 break; | |
740 | |
741 if (p != NULL) | |
742 { | |
743 /* Recycle the allocated memory from this page ... */ | |
744 *pp = p->next; | |
745 page = p->page; | |
746 | |
747 #ifdef USING_MALLOC_PAGE_GROUPS | |
748 group = p->group; | |
749 #endif | |
750 | |
751 /* ... and, if possible, the page entry itself. */ | |
752 if (p->order == order) | |
753 { | |
754 entry = p; | |
755 memset (entry, 0, page_entry_size); | |
756 } | |
757 else | |
758 free (p); | |
759 } | |
760 #ifdef USING_MMAP | |
761 else if (entry_size == G.pagesize) | |
762 { | |
763 /* We want just one page. Allocate a bunch of them and put the | |
764 extras on the freelist. (Can only do this optimization with | |
765 mmap for backing store.) */ | |
766 struct page_entry *e, *f = G.free_pages; | |
767 int i; | |
768 | |
769 page = alloc_anon (NULL, G.pagesize * GGC_QUIRE_SIZE); | |
770 | |
771 /* This loop counts down so that the chain will be in ascending | |
772 memory order. */ | |
773 for (i = GGC_QUIRE_SIZE - 1; i >= 1; i--) | |
774 { | |
775 e = XCNEWVAR (struct page_entry, page_entry_size); | |
776 e->order = order; | |
777 e->bytes = G.pagesize; | |
778 e->page = page + (i << G.lg_pagesize); | |
779 e->next = f; | |
780 f = e; | |
781 } | |
782 | |
783 G.free_pages = f; | |
784 } | |
785 else | |
786 page = alloc_anon (NULL, entry_size); | |
787 #endif | |
788 #ifdef USING_MALLOC_PAGE_GROUPS | |
789 else | |
790 { | |
791 /* Allocate a large block of memory and serve out the aligned | |
792 pages therein. This results in much less memory wastage | |
793 than the traditional implementation of valloc. */ | |
794 | |
795 char *allocation, *a, *enda; | |
796 size_t alloc_size, head_slop, tail_slop; | |
797 int multiple_pages = (entry_size == G.pagesize); | |
798 | |
799 if (multiple_pages) | |
800 alloc_size = GGC_QUIRE_SIZE * G.pagesize; | |
801 else | |
802 alloc_size = entry_size + G.pagesize - 1; | |
803 allocation = XNEWVEC (char, alloc_size); | |
804 | |
805 page = (char *) (((size_t) allocation + G.pagesize - 1) & -G.pagesize); | |
806 head_slop = page - allocation; | |
807 if (multiple_pages) | |
808 tail_slop = ((size_t) allocation + alloc_size) & (G.pagesize - 1); | |
809 else | |
810 tail_slop = alloc_size - entry_size - head_slop; | |
811 enda = allocation + alloc_size - tail_slop; | |
812 | |
813 /* We allocated N pages, which are likely not aligned, leaving | |
814 us with N-1 usable pages. We plan to place the page_group | |
815 structure somewhere in the slop. */ | |
816 if (head_slop >= sizeof (page_group)) | |
817 group = (page_group *)page - 1; | |
818 else | |
819 { | |
820 /* We magically got an aligned allocation. Too bad, we have | |
821 to waste a page anyway. */ | |
822 if (tail_slop == 0) | |
823 { | |
824 enda -= G.pagesize; | |
825 tail_slop += G.pagesize; | |
826 } | |
827 gcc_assert (tail_slop >= sizeof (page_group)); | |
828 group = (page_group *)enda; | |
829 tail_slop -= sizeof (page_group); | |
830 } | |
831 | |
832 /* Remember that we allocated this memory. */ | |
833 group->next = G.page_groups; | |
834 group->allocation = allocation; | |
835 group->alloc_size = alloc_size; | |
836 group->in_use = 0; | |
837 G.page_groups = group; | |
838 G.bytes_mapped += alloc_size; | |
839 | |
840 /* If we allocated multiple pages, put the rest on the free list. */ | |
841 if (multiple_pages) | |
842 { | |
843 struct page_entry *e, *f = G.free_pages; | |
844 for (a = enda - G.pagesize; a != page; a -= G.pagesize) | |
845 { | |
846 e = XCNEWVAR (struct page_entry, page_entry_size); | |
847 e->order = order; | |
848 e->bytes = G.pagesize; | |
849 e->page = a; | |
850 e->group = group; | |
851 e->next = f; | |
852 f = e; | |
853 } | |
854 G.free_pages = f; | |
855 } | |
856 } | |
857 #endif | |
858 | |
859 if (entry == NULL) | |
860 entry = XCNEWVAR (struct page_entry, page_entry_size); | |
861 | |
862 entry->bytes = entry_size; | |
863 entry->page = page; | |
864 entry->context_depth = G.context_depth; | |
865 entry->order = order; | |
866 entry->num_free_objects = num_objects; | |
867 entry->next_bit_hint = 1; | |
868 | |
869 G.context_depth_allocations |= (unsigned long)1 << G.context_depth; | |
870 | |
871 #ifdef USING_MALLOC_PAGE_GROUPS | |
872 entry->group = group; | |
873 set_page_group_in_use (group, page); | |
874 #endif | |
875 | |
876 /* Set the one-past-the-end in-use bit. This acts as a sentry as we | |
877 increment the hint. */ | |
878 entry->in_use_p[num_objects / HOST_BITS_PER_LONG] | |
879 = (unsigned long) 1 << (num_objects % HOST_BITS_PER_LONG); | |
880 | |
881 set_page_table_entry (page, entry); | |
882 | |
883 if (GGC_DEBUG_LEVEL >= 2) | |
884 fprintf (G.debug_file, | |
885 "Allocating page at %p, object size=%lu, data %p-%p\n", | |
886 (void *) entry, (unsigned long) OBJECT_SIZE (order), page, | |
887 page + entry_size - 1); | |
888 | |
889 return entry; | |
890 } | |
891 | |
892 /* Adjust the size of G.depth so that no index greater than the one | |
893 used by the top of the G.by_depth is used. */ | |
894 | |
895 static inline void | |
896 adjust_depth (void) | |
897 { | |
898 page_entry *top; | |
899 | |
900 if (G.by_depth_in_use) | |
901 { | |
902 top = G.by_depth[G.by_depth_in_use-1]; | |
903 | |
904 /* Peel back indices in depth that index into by_depth, so that | |
905 as new elements are added to by_depth, we note the indices | |
906 of those elements, if they are for new context depths. */ | |
907 while (G.depth_in_use > (size_t)top->context_depth+1) | |
908 --G.depth_in_use; | |
909 } | |
910 } | |
911 | |
912 /* For a page that is no longer needed, put it on the free page list. */ | |
913 | |
914 static void | |
915 free_page (page_entry *entry) | |
916 { | |
917 if (GGC_DEBUG_LEVEL >= 2) | |
918 fprintf (G.debug_file, | |
919 "Deallocating page at %p, data %p-%p\n", (void *) entry, | |
920 entry->page, entry->page + entry->bytes - 1); | |
921 | |
922 /* Mark the page as inaccessible. Discard the handle to avoid handle | |
923 leak. */ | |
924 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (entry->page, entry->bytes)); | |
925 | |
926 set_page_table_entry (entry->page, NULL); | |
927 | |
928 #ifdef USING_MALLOC_PAGE_GROUPS | |
929 clear_page_group_in_use (entry->group, entry->page); | |
930 #endif | |
931 | |
932 if (G.by_depth_in_use > 1) | |
933 { | |
934 page_entry *top = G.by_depth[G.by_depth_in_use-1]; | |
935 int i = entry->index_by_depth; | |
936 | |
937 /* We cannot free a page from a context deeper than the current | |
938 one. */ | |
939 gcc_assert (entry->context_depth == top->context_depth); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
940 |
0 | 941 /* Put top element into freed slot. */ |
942 G.by_depth[i] = top; | |
943 G.save_in_use[i] = G.save_in_use[G.by_depth_in_use-1]; | |
944 top->index_by_depth = i; | |
945 } | |
946 --G.by_depth_in_use; | |
947 | |
948 adjust_depth (); | |
949 | |
950 entry->next = G.free_pages; | |
951 G.free_pages = entry; | |
952 } | |
953 | |
954 /* Release the free page cache to the system. */ | |
955 | |
956 static void | |
957 release_pages (void) | |
958 { | |
959 #ifdef USING_MMAP | |
960 page_entry *p, *next; | |
961 char *start; | |
962 size_t len; | |
963 | |
964 /* Gather up adjacent pages so they are unmapped together. */ | |
965 p = G.free_pages; | |
966 | |
967 while (p) | |
968 { | |
969 start = p->page; | |
970 next = p->next; | |
971 len = p->bytes; | |
972 free (p); | |
973 p = next; | |
974 | |
975 while (p && p->page == start + len) | |
976 { | |
977 next = p->next; | |
978 len += p->bytes; | |
979 free (p); | |
980 p = next; | |
981 } | |
982 | |
983 munmap (start, len); | |
984 G.bytes_mapped -= len; | |
985 } | |
986 | |
987 G.free_pages = NULL; | |
988 #endif | |
989 #ifdef USING_MALLOC_PAGE_GROUPS | |
990 page_entry **pp, *p; | |
991 page_group **gp, *g; | |
992 | |
993 /* Remove all pages from free page groups from the list. */ | |
994 pp = &G.free_pages; | |
995 while ((p = *pp) != NULL) | |
996 if (p->group->in_use == 0) | |
997 { | |
998 *pp = p->next; | |
999 free (p); | |
1000 } | |
1001 else | |
1002 pp = &p->next; | |
1003 | |
1004 /* Remove all free page groups, and release the storage. */ | |
1005 gp = &G.page_groups; | |
1006 while ((g = *gp) != NULL) | |
1007 if (g->in_use == 0) | |
1008 { | |
1009 *gp = g->next; | |
1010 G.bytes_mapped -= g->alloc_size; | |
1011 free (g->allocation); | |
1012 } | |
1013 else | |
1014 gp = &g->next; | |
1015 #endif | |
1016 } | |
1017 | |
1018 /* This table provides a fast way to determine ceil(log_2(size)) for | |
1019 allocation requests. The minimum allocation size is eight bytes. */ | |
1020 #define NUM_SIZE_LOOKUP 512 | |
1021 static unsigned char size_lookup[NUM_SIZE_LOOKUP] = | |
1022 { | |
1023 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, | |
1024 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, | |
1025 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, | |
1026 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, | |
1027 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1028 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1029 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1030 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1031 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1032 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1033 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1034 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1035 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1036 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1037 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1038 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1039 8, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1040 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1041 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1042 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1043 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1044 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1045 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1046 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1047 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1048 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1049 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1050 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1051 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1052 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1053 9, 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 }; | |
1056 | |
1057 /* Typed allocation function. Does nothing special in this collector. */ | |
1058 | |
1059 void * | |
1060 ggc_alloc_typed_stat (enum gt_types_enum type ATTRIBUTE_UNUSED, size_t size | |
1061 MEM_STAT_DECL) | |
1062 { | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1063 return ggc_internal_alloc_stat (size PASS_MEM_STAT); |
0 | 1064 } |
1065 | |
1066 /* Allocate a chunk of memory of SIZE bytes. Its contents are undefined. */ | |
1067 | |
1068 void * | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1069 ggc_internal_alloc_stat (size_t size MEM_STAT_DECL) |
0 | 1070 { |
1071 size_t order, word, bit, object_offset, object_size; | |
1072 struct page_entry *entry; | |
1073 void *result; | |
1074 | |
1075 if (size < NUM_SIZE_LOOKUP) | |
1076 { | |
1077 order = size_lookup[size]; | |
1078 object_size = OBJECT_SIZE (order); | |
1079 } | |
1080 else | |
1081 { | |
1082 order = 10; | |
1083 while (size > (object_size = OBJECT_SIZE (order))) | |
1084 order++; | |
1085 } | |
1086 | |
1087 /* If there are non-full pages for this size allocation, they are at | |
1088 the head of the list. */ | |
1089 entry = G.pages[order]; | |
1090 | |
1091 /* If there is no page for this object size, or all pages in this | |
1092 context are full, allocate a new page. */ | |
1093 if (entry == NULL || entry->num_free_objects == 0) | |
1094 { | |
1095 struct page_entry *new_entry; | |
1096 new_entry = alloc_page (order); | |
1097 | |
1098 new_entry->index_by_depth = G.by_depth_in_use; | |
1099 push_by_depth (new_entry, 0); | |
1100 | |
1101 /* We can skip context depths, if we do, make sure we go all the | |
1102 way to the new depth. */ | |
1103 while (new_entry->context_depth >= G.depth_in_use) | |
1104 push_depth (G.by_depth_in_use-1); | |
1105 | |
1106 /* If this is the only entry, it's also the tail. If it is not | |
1107 the only entry, then we must update the PREV pointer of the | |
1108 ENTRY (G.pages[order]) to point to our new page entry. */ | |
1109 if (entry == NULL) | |
1110 G.page_tails[order] = new_entry; | |
1111 else | |
1112 entry->prev = new_entry; | |
1113 | |
1114 /* Put new pages at the head of the page list. By definition the | |
1115 entry at the head of the list always has a NULL pointer. */ | |
1116 new_entry->next = entry; | |
1117 new_entry->prev = NULL; | |
1118 entry = new_entry; | |
1119 G.pages[order] = new_entry; | |
1120 | |
1121 /* For a new page, we know the word and bit positions (in the | |
1122 in_use bitmap) of the first available object -- they're zero. */ | |
1123 new_entry->next_bit_hint = 1; | |
1124 word = 0; | |
1125 bit = 0; | |
1126 object_offset = 0; | |
1127 } | |
1128 else | |
1129 { | |
1130 /* First try to use the hint left from the previous allocation | |
1131 to locate a clear bit in the in-use bitmap. We've made sure | |
1132 that the one-past-the-end bit is always set, so if the hint | |
1133 has run over, this test will fail. */ | |
1134 unsigned hint = entry->next_bit_hint; | |
1135 word = hint / HOST_BITS_PER_LONG; | |
1136 bit = hint % HOST_BITS_PER_LONG; | |
1137 | |
1138 /* If the hint didn't work, scan the bitmap from the beginning. */ | |
1139 if ((entry->in_use_p[word] >> bit) & 1) | |
1140 { | |
1141 word = bit = 0; | |
1142 while (~entry->in_use_p[word] == 0) | |
1143 ++word; | |
1144 | |
1145 #if GCC_VERSION >= 3004 | |
1146 bit = __builtin_ctzl (~entry->in_use_p[word]); | |
1147 #else | |
1148 while ((entry->in_use_p[word] >> bit) & 1) | |
1149 ++bit; | |
1150 #endif | |
1151 | |
1152 hint = word * HOST_BITS_PER_LONG + bit; | |
1153 } | |
1154 | |
1155 /* Next time, try the next bit. */ | |
1156 entry->next_bit_hint = hint + 1; | |
1157 | |
1158 object_offset = hint * object_size; | |
1159 } | |
1160 | |
1161 /* Set the in-use bit. */ | |
1162 entry->in_use_p[word] |= ((unsigned long) 1 << bit); | |
1163 | |
1164 /* Keep a running total of the number of free objects. If this page | |
1165 fills up, we may have to move it to the end of the list if the | |
1166 next page isn't full. If the next page is full, all subsequent | |
1167 pages are full, so there's no need to move it. */ | |
1168 if (--entry->num_free_objects == 0 | |
1169 && entry->next != NULL | |
1170 && entry->next->num_free_objects > 0) | |
1171 { | |
1172 /* We have a new head for the list. */ | |
1173 G.pages[order] = entry->next; | |
1174 | |
1175 /* We are moving ENTRY to the end of the page table list. | |
1176 The new page at the head of the list will have NULL in | |
1177 its PREV field and ENTRY will have NULL in its NEXT field. */ | |
1178 entry->next->prev = NULL; | |
1179 entry->next = NULL; | |
1180 | |
1181 /* Append ENTRY to the tail of the list. */ | |
1182 entry->prev = G.page_tails[order]; | |
1183 G.page_tails[order]->next = entry; | |
1184 G.page_tails[order] = entry; | |
1185 } | |
1186 | |
1187 /* Calculate the object's address. */ | |
1188 result = entry->page + object_offset; | |
1189 #ifdef GATHER_STATISTICS | |
1190 ggc_record_overhead (OBJECT_SIZE (order), OBJECT_SIZE (order) - size, | |
1191 result PASS_MEM_STAT); | |
1192 #endif | |
1193 | |
1194 #ifdef ENABLE_GC_CHECKING | |
1195 /* Keep poisoning-by-writing-0xaf the object, in an attempt to keep the | |
1196 exact same semantics in presence of memory bugs, regardless of | |
1197 ENABLE_VALGRIND_CHECKING. We override this request below. Drop the | |
1198 handle to avoid handle leak. */ | |
1199 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (result, object_size)); | |
1200 | |
1201 /* `Poison' the entire allocated object, including any padding at | |
1202 the end. */ | |
1203 memset (result, 0xaf, object_size); | |
1204 | |
1205 /* Make the bytes after the end of the object unaccessible. Discard the | |
1206 handle to avoid handle leak. */ | |
1207 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS ((char *) result + size, | |
1208 object_size - size)); | |
1209 #endif | |
1210 | |
1211 /* Tell Valgrind that the memory is there, but its content isn't | |
1212 defined. The bytes at the end of the object are still marked | |
1213 unaccessible. */ | |
1214 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (result, size)); | |
1215 | |
1216 /* Keep track of how many bytes are being allocated. This | |
1217 information is used in deciding when to collect. */ | |
1218 G.allocated += object_size; | |
1219 | |
1220 /* For timevar statistics. */ | |
1221 timevar_ggc_mem_total += object_size; | |
1222 | |
1223 #ifdef GATHER_STATISTICS | |
1224 { | |
1225 size_t overhead = object_size - size; | |
1226 | |
1227 G.stats.total_overhead += overhead; | |
1228 G.stats.total_allocated += object_size; | |
1229 G.stats.total_overhead_per_order[order] += overhead; | |
1230 G.stats.total_allocated_per_order[order] += object_size; | |
1231 | |
1232 if (size <= 32) | |
1233 { | |
1234 G.stats.total_overhead_under32 += overhead; | |
1235 G.stats.total_allocated_under32 += object_size; | |
1236 } | |
1237 if (size <= 64) | |
1238 { | |
1239 G.stats.total_overhead_under64 += overhead; | |
1240 G.stats.total_allocated_under64 += object_size; | |
1241 } | |
1242 if (size <= 128) | |
1243 { | |
1244 G.stats.total_overhead_under128 += overhead; | |
1245 G.stats.total_allocated_under128 += object_size; | |
1246 } | |
1247 } | |
1248 #endif | |
1249 | |
1250 if (GGC_DEBUG_LEVEL >= 3) | |
1251 fprintf (G.debug_file, | |
1252 "Allocating object, requested size=%lu, actual=%lu at %p on %p\n", | |
1253 (unsigned long) size, (unsigned long) object_size, result, | |
1254 (void *) entry); | |
1255 | |
1256 return result; | |
1257 } | |
1258 | |
1259 /* Mark function for strings. */ | |
1260 | |
1261 void | |
1262 gt_ggc_m_S (const void *p) | |
1263 { | |
1264 page_entry *entry; | |
1265 unsigned bit, word; | |
1266 unsigned long mask; | |
1267 unsigned long offset; | |
1268 | |
1269 if (!p || !ggc_allocated_p (p)) | |
1270 return; | |
1271 | |
1272 /* Look up the page on which the object is alloced. . */ | |
1273 entry = lookup_page_table_entry (p); | |
1274 gcc_assert (entry); | |
1275 | |
1276 /* Calculate the index of the object on the page; this is its bit | |
1277 position in the in_use_p bitmap. Note that because a char* might | |
1278 point to the middle of an object, we need special code here to | |
1279 make sure P points to the start of an object. */ | |
1280 offset = ((const char *) p - entry->page) % object_size_table[entry->order]; | |
1281 if (offset) | |
1282 { | |
1283 /* Here we've seen a char* which does not point to the beginning | |
1284 of an allocated object. We assume it points to the middle of | |
1285 a STRING_CST. */ | |
1286 gcc_assert (offset == offsetof (struct tree_string, str)); | |
1287 p = ((const char *) p) - offset; | |
1288 gt_ggc_mx_lang_tree_node (CONST_CAST (void *, p)); | |
1289 return; | |
1290 } | |
1291 | |
1292 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order); | |
1293 word = bit / HOST_BITS_PER_LONG; | |
1294 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG); | |
1295 | |
1296 /* If the bit was previously set, skip it. */ | |
1297 if (entry->in_use_p[word] & mask) | |
1298 return; | |
1299 | |
1300 /* Otherwise set it, and decrement the free object count. */ | |
1301 entry->in_use_p[word] |= mask; | |
1302 entry->num_free_objects -= 1; | |
1303 | |
1304 if (GGC_DEBUG_LEVEL >= 4) | |
1305 fprintf (G.debug_file, "Marking %p\n", p); | |
1306 | |
1307 return; | |
1308 } | |
1309 | |
1310 /* If P is not marked, marks it and return false. Otherwise return true. | |
1311 P must have been allocated by the GC allocator; it mustn't point to | |
1312 static objects, stack variables, or memory allocated with malloc. */ | |
1313 | |
1314 int | |
1315 ggc_set_mark (const void *p) | |
1316 { | |
1317 page_entry *entry; | |
1318 unsigned bit, word; | |
1319 unsigned long mask; | |
1320 | |
1321 /* Look up the page on which the object is alloced. If the object | |
1322 wasn't allocated by the collector, we'll probably die. */ | |
1323 entry = lookup_page_table_entry (p); | |
1324 gcc_assert (entry); | |
1325 | |
1326 /* Calculate the index of the object on the page; this is its bit | |
1327 position in the in_use_p bitmap. */ | |
1328 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order); | |
1329 word = bit / HOST_BITS_PER_LONG; | |
1330 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG); | |
1331 | |
1332 /* If the bit was previously set, skip it. */ | |
1333 if (entry->in_use_p[word] & mask) | |
1334 return 1; | |
1335 | |
1336 /* Otherwise set it, and decrement the free object count. */ | |
1337 entry->in_use_p[word] |= mask; | |
1338 entry->num_free_objects -= 1; | |
1339 | |
1340 if (GGC_DEBUG_LEVEL >= 4) | |
1341 fprintf (G.debug_file, "Marking %p\n", p); | |
1342 | |
1343 return 0; | |
1344 } | |
1345 | |
1346 /* Return 1 if P has been marked, zero otherwise. | |
1347 P must have been allocated by the GC allocator; it mustn't point to | |
1348 static objects, stack variables, or memory allocated with malloc. */ | |
1349 | |
1350 int | |
1351 ggc_marked_p (const void *p) | |
1352 { | |
1353 page_entry *entry; | |
1354 unsigned bit, word; | |
1355 unsigned long mask; | |
1356 | |
1357 /* Look up the page on which the object is alloced. If the object | |
1358 wasn't allocated by the collector, we'll probably die. */ | |
1359 entry = lookup_page_table_entry (p); | |
1360 gcc_assert (entry); | |
1361 | |
1362 /* Calculate the index of the object on the page; this is its bit | |
1363 position in the in_use_p bitmap. */ | |
1364 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order); | |
1365 word = bit / HOST_BITS_PER_LONG; | |
1366 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG); | |
1367 | |
1368 return (entry->in_use_p[word] & mask) != 0; | |
1369 } | |
1370 | |
1371 /* Return the size of the gc-able object P. */ | |
1372 | |
1373 size_t | |
1374 ggc_get_size (const void *p) | |
1375 { | |
1376 page_entry *pe = lookup_page_table_entry (p); | |
1377 return OBJECT_SIZE (pe->order); | |
1378 } | |
1379 | |
1380 /* Release the memory for object P. */ | |
1381 | |
1382 void | |
1383 ggc_free (void *p) | |
1384 { | |
1385 page_entry *pe = lookup_page_table_entry (p); | |
1386 size_t order = pe->order; | |
1387 size_t size = OBJECT_SIZE (order); | |
1388 | |
1389 #ifdef GATHER_STATISTICS | |
1390 ggc_free_overhead (p); | |
1391 #endif | |
1392 | |
1393 if (GGC_DEBUG_LEVEL >= 3) | |
1394 fprintf (G.debug_file, | |
1395 "Freeing object, actual size=%lu, at %p on %p\n", | |
1396 (unsigned long) size, p, (void *) pe); | |
1397 | |
1398 #ifdef ENABLE_GC_CHECKING | |
1399 /* Poison the data, to indicate the data is garbage. */ | |
1400 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (p, size)); | |
1401 memset (p, 0xa5, size); | |
1402 #endif | |
1403 /* Let valgrind know the object is free. */ | |
1404 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (p, size)); | |
1405 | |
1406 #ifdef ENABLE_GC_ALWAYS_COLLECT | |
1407 /* In the completely-anal-checking mode, we do *not* immediately free | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1408 the data, but instead verify that the data is *actually* not |
0 | 1409 reachable the next time we collect. */ |
1410 { | |
1411 struct free_object *fo = XNEW (struct free_object); | |
1412 fo->object = p; | |
1413 fo->next = G.free_object_list; | |
1414 G.free_object_list = fo; | |
1415 } | |
1416 #else | |
1417 { | |
1418 unsigned int bit_offset, word, bit; | |
1419 | |
1420 G.allocated -= size; | |
1421 | |
1422 /* Mark the object not-in-use. */ | |
1423 bit_offset = OFFSET_TO_BIT (((const char *) p) - pe->page, order); | |
1424 word = bit_offset / HOST_BITS_PER_LONG; | |
1425 bit = bit_offset % HOST_BITS_PER_LONG; | |
1426 pe->in_use_p[word] &= ~(1UL << bit); | |
1427 | |
1428 if (pe->num_free_objects++ == 0) | |
1429 { | |
1430 page_entry *p, *q; | |
1431 | |
1432 /* If the page is completely full, then it's supposed to | |
1433 be after all pages that aren't. Since we've freed one | |
1434 object from a page that was full, we need to move the | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1435 page to the head of the list. |
0 | 1436 |
1437 PE is the node we want to move. Q is the previous node | |
1438 and P is the next node in the list. */ | |
1439 q = pe->prev; | |
1440 if (q && q->num_free_objects == 0) | |
1441 { | |
1442 p = pe->next; | |
1443 | |
1444 q->next = p; | |
1445 | |
1446 /* If PE was at the end of the list, then Q becomes the | |
1447 new end of the list. If PE was not the end of the | |
1448 list, then we need to update the PREV field for P. */ | |
1449 if (!p) | |
1450 G.page_tails[order] = q; | |
1451 else | |
1452 p->prev = q; | |
1453 | |
1454 /* Move PE to the head of the list. */ | |
1455 pe->next = G.pages[order]; | |
1456 pe->prev = NULL; | |
1457 G.pages[order]->prev = pe; | |
1458 G.pages[order] = pe; | |
1459 } | |
1460 | |
1461 /* Reset the hint bit to point to the only free object. */ | |
1462 pe->next_bit_hint = bit_offset; | |
1463 } | |
1464 } | |
1465 #endif | |
1466 } | |
1467 | |
1468 /* Subroutine of init_ggc which computes the pair of numbers used to | |
1469 perform division by OBJECT_SIZE (order) and fills in inverse_table[]. | |
1470 | |
1471 This algorithm is taken from Granlund and Montgomery's paper | |
1472 "Division by Invariant Integers using Multiplication" | |
1473 (Proc. SIGPLAN PLDI, 1994), section 9 (Exact division by | |
1474 constants). */ | |
1475 | |
1476 static void | |
1477 compute_inverse (unsigned order) | |
1478 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1479 size_t size, inv; |
0 | 1480 unsigned int e; |
1481 | |
1482 size = OBJECT_SIZE (order); | |
1483 e = 0; | |
1484 while (size % 2 == 0) | |
1485 { | |
1486 e++; | |
1487 size >>= 1; | |
1488 } | |
1489 | |
1490 inv = size; | |
1491 while (inv * size != 1) | |
1492 inv = inv * (2 - inv*size); | |
1493 | |
1494 DIV_MULT (order) = inv; | |
1495 DIV_SHIFT (order) = e; | |
1496 } | |
1497 | |
1498 /* Initialize the ggc-mmap allocator. */ | |
1499 void | |
1500 init_ggc (void) | |
1501 { | |
1502 unsigned order; | |
1503 | |
1504 G.pagesize = getpagesize(); | |
1505 G.lg_pagesize = exact_log2 (G.pagesize); | |
1506 | |
1507 #ifdef HAVE_MMAP_DEV_ZERO | |
1508 G.dev_zero_fd = open ("/dev/zero", O_RDONLY); | |
1509 if (G.dev_zero_fd == -1) | |
1510 internal_error ("open /dev/zero: %m"); | |
1511 #endif | |
1512 | |
1513 #if 0 | |
1514 G.debug_file = fopen ("ggc-mmap.debug", "w"); | |
1515 #else | |
1516 G.debug_file = stdout; | |
1517 #endif | |
1518 | |
1519 #ifdef USING_MMAP | |
1520 /* StunOS has an amazing off-by-one error for the first mmap allocation | |
1521 after fiddling with RLIMIT_STACK. The result, as hard as it is to | |
1522 believe, is an unaligned page allocation, which would cause us to | |
1523 hork badly if we tried to use it. */ | |
1524 { | |
1525 char *p = alloc_anon (NULL, G.pagesize); | |
1526 struct page_entry *e; | |
1527 if ((size_t)p & (G.pagesize - 1)) | |
1528 { | |
1529 /* How losing. Discard this one and try another. If we still | |
1530 can't get something useful, give up. */ | |
1531 | |
1532 p = alloc_anon (NULL, G.pagesize); | |
1533 gcc_assert (!((size_t)p & (G.pagesize - 1))); | |
1534 } | |
1535 | |
1536 /* We have a good page, might as well hold onto it... */ | |
1537 e = XCNEW (struct page_entry); | |
1538 e->bytes = G.pagesize; | |
1539 e->page = p; | |
1540 e->next = G.free_pages; | |
1541 G.free_pages = e; | |
1542 } | |
1543 #endif | |
1544 | |
1545 /* Initialize the object size table. */ | |
1546 for (order = 0; order < HOST_BITS_PER_PTR; ++order) | |
1547 object_size_table[order] = (size_t) 1 << order; | |
1548 for (order = HOST_BITS_PER_PTR; order < NUM_ORDERS; ++order) | |
1549 { | |
1550 size_t s = extra_order_size_table[order - HOST_BITS_PER_PTR]; | |
1551 | |
1552 /* If S is not a multiple of the MAX_ALIGNMENT, then round it up | |
1553 so that we're sure of getting aligned memory. */ | |
1554 s = ROUND_UP (s, MAX_ALIGNMENT); | |
1555 object_size_table[order] = s; | |
1556 } | |
1557 | |
1558 /* Initialize the objects-per-page and inverse tables. */ | |
1559 for (order = 0; order < NUM_ORDERS; ++order) | |
1560 { | |
1561 objects_per_page_table[order] = G.pagesize / OBJECT_SIZE (order); | |
1562 if (objects_per_page_table[order] == 0) | |
1563 objects_per_page_table[order] = 1; | |
1564 compute_inverse (order); | |
1565 } | |
1566 | |
1567 /* Reset the size_lookup array to put appropriately sized objects in | |
1568 the special orders. All objects bigger than the previous power | |
1569 of two, but no greater than the special size, should go in the | |
1570 new order. */ | |
1571 for (order = HOST_BITS_PER_PTR; order < NUM_ORDERS; ++order) | |
1572 { | |
1573 int o; | |
1574 int i; | |
1575 | |
1576 i = OBJECT_SIZE (order); | |
1577 if (i >= NUM_SIZE_LOOKUP) | |
1578 continue; | |
1579 | |
1580 for (o = size_lookup[i]; o == size_lookup [i]; --i) | |
1581 size_lookup[i] = order; | |
1582 } | |
1583 | |
1584 G.depth_in_use = 0; | |
1585 G.depth_max = 10; | |
1586 G.depth = XNEWVEC (unsigned int, G.depth_max); | |
1587 | |
1588 G.by_depth_in_use = 0; | |
1589 G.by_depth_max = INITIAL_PTE_COUNT; | |
1590 G.by_depth = XNEWVEC (page_entry *, G.by_depth_max); | |
1591 G.save_in_use = XNEWVEC (unsigned long *, G.by_depth_max); | |
1592 } | |
1593 | |
1594 /* Merge the SAVE_IN_USE_P and IN_USE_P arrays in P so that IN_USE_P | |
1595 reflects reality. Recalculate NUM_FREE_OBJECTS as well. */ | |
1596 | |
1597 static void | |
1598 ggc_recalculate_in_use_p (page_entry *p) | |
1599 { | |
1600 unsigned int i; | |
1601 size_t num_objects; | |
1602 | |
1603 /* Because the past-the-end bit in in_use_p is always set, we | |
1604 pretend there is one additional object. */ | |
1605 num_objects = OBJECTS_IN_PAGE (p) + 1; | |
1606 | |
1607 /* Reset the free object count. */ | |
1608 p->num_free_objects = num_objects; | |
1609 | |
1610 /* Combine the IN_USE_P and SAVE_IN_USE_P arrays. */ | |
1611 for (i = 0; | |
1612 i < CEIL (BITMAP_SIZE (num_objects), | |
1613 sizeof (*p->in_use_p)); | |
1614 ++i) | |
1615 { | |
1616 unsigned long j; | |
1617 | |
1618 /* Something is in use if it is marked, or if it was in use in a | |
1619 context further down the context stack. */ | |
1620 p->in_use_p[i] |= save_in_use_p (p)[i]; | |
1621 | |
1622 /* Decrement the free object count for every object allocated. */ | |
1623 for (j = p->in_use_p[i]; j; j >>= 1) | |
1624 p->num_free_objects -= (j & 1); | |
1625 } | |
1626 | |
1627 gcc_assert (p->num_free_objects < num_objects); | |
1628 } | |
1629 | |
1630 /* Unmark all objects. */ | |
1631 | |
1632 static void | |
1633 clear_marks (void) | |
1634 { | |
1635 unsigned order; | |
1636 | |
1637 for (order = 2; order < NUM_ORDERS; order++) | |
1638 { | |
1639 page_entry *p; | |
1640 | |
1641 for (p = G.pages[order]; p != NULL; p = p->next) | |
1642 { | |
1643 size_t num_objects = OBJECTS_IN_PAGE (p); | |
1644 size_t bitmap_size = BITMAP_SIZE (num_objects + 1); | |
1645 | |
1646 /* The data should be page-aligned. */ | |
1647 gcc_assert (!((size_t) p->page & (G.pagesize - 1))); | |
1648 | |
1649 /* Pages that aren't in the topmost context are not collected; | |
1650 nevertheless, we need their in-use bit vectors to store GC | |
1651 marks. So, back them up first. */ | |
1652 if (p->context_depth < G.context_depth) | |
1653 { | |
1654 if (! save_in_use_p (p)) | |
1655 save_in_use_p (p) = XNEWVAR (unsigned long, bitmap_size); | |
1656 memcpy (save_in_use_p (p), p->in_use_p, bitmap_size); | |
1657 } | |
1658 | |
1659 /* Reset reset the number of free objects and clear the | |
1660 in-use bits. These will be adjusted by mark_obj. */ | |
1661 p->num_free_objects = num_objects; | |
1662 memset (p->in_use_p, 0, bitmap_size); | |
1663 | |
1664 /* Make sure the one-past-the-end bit is always set. */ | |
1665 p->in_use_p[num_objects / HOST_BITS_PER_LONG] | |
1666 = ((unsigned long) 1 << (num_objects % HOST_BITS_PER_LONG)); | |
1667 } | |
1668 } | |
1669 } | |
1670 | |
1671 /* Free all empty pages. Partially empty pages need no attention | |
1672 because the `mark' bit doubles as an `unused' bit. */ | |
1673 | |
1674 static void | |
1675 sweep_pages (void) | |
1676 { | |
1677 unsigned order; | |
1678 | |
1679 for (order = 2; order < NUM_ORDERS; order++) | |
1680 { | |
1681 /* The last page-entry to consider, regardless of entries | |
1682 placed at the end of the list. */ | |
1683 page_entry * const last = G.page_tails[order]; | |
1684 | |
1685 size_t num_objects; | |
1686 size_t live_objects; | |
1687 page_entry *p, *previous; | |
1688 int done; | |
1689 | |
1690 p = G.pages[order]; | |
1691 if (p == NULL) | |
1692 continue; | |
1693 | |
1694 previous = NULL; | |
1695 do | |
1696 { | |
1697 page_entry *next = p->next; | |
1698 | |
1699 /* Loop until all entries have been examined. */ | |
1700 done = (p == last); | |
1701 | |
1702 num_objects = OBJECTS_IN_PAGE (p); | |
1703 | |
1704 /* Add all live objects on this page to the count of | |
1705 allocated memory. */ | |
1706 live_objects = num_objects - p->num_free_objects; | |
1707 | |
1708 G.allocated += OBJECT_SIZE (order) * live_objects; | |
1709 | |
1710 /* Only objects on pages in the topmost context should get | |
1711 collected. */ | |
1712 if (p->context_depth < G.context_depth) | |
1713 ; | |
1714 | |
1715 /* Remove the page if it's empty. */ | |
1716 else if (live_objects == 0) | |
1717 { | |
1718 /* If P was the first page in the list, then NEXT | |
1719 becomes the new first page in the list, otherwise | |
1720 splice P out of the forward pointers. */ | |
1721 if (! previous) | |
1722 G.pages[order] = next; | |
1723 else | |
1724 previous->next = next; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1725 |
0 | 1726 /* Splice P out of the back pointers too. */ |
1727 if (next) | |
1728 next->prev = previous; | |
1729 | |
1730 /* Are we removing the last element? */ | |
1731 if (p == G.page_tails[order]) | |
1732 G.page_tails[order] = previous; | |
1733 free_page (p); | |
1734 p = previous; | |
1735 } | |
1736 | |
1737 /* If the page is full, move it to the end. */ | |
1738 else if (p->num_free_objects == 0) | |
1739 { | |
1740 /* Don't move it if it's already at the end. */ | |
1741 if (p != G.page_tails[order]) | |
1742 { | |
1743 /* Move p to the end of the list. */ | |
1744 p->next = NULL; | |
1745 p->prev = G.page_tails[order]; | |
1746 G.page_tails[order]->next = p; | |
1747 | |
1748 /* Update the tail pointer... */ | |
1749 G.page_tails[order] = p; | |
1750 | |
1751 /* ... and the head pointer, if necessary. */ | |
1752 if (! previous) | |
1753 G.pages[order] = next; | |
1754 else | |
1755 previous->next = next; | |
1756 | |
1757 /* And update the backpointer in NEXT if necessary. */ | |
1758 if (next) | |
1759 next->prev = previous; | |
1760 | |
1761 p = previous; | |
1762 } | |
1763 } | |
1764 | |
1765 /* If we've fallen through to here, it's a page in the | |
1766 topmost context that is neither full nor empty. Such a | |
1767 page must precede pages at lesser context depth in the | |
1768 list, so move it to the head. */ | |
1769 else if (p != G.pages[order]) | |
1770 { | |
1771 previous->next = p->next; | |
1772 | |
1773 /* Update the backchain in the next node if it exists. */ | |
1774 if (p->next) | |
1775 p->next->prev = previous; | |
1776 | |
1777 /* Move P to the head of the list. */ | |
1778 p->next = G.pages[order]; | |
1779 p->prev = NULL; | |
1780 G.pages[order]->prev = p; | |
1781 | |
1782 /* Update the head pointer. */ | |
1783 G.pages[order] = p; | |
1784 | |
1785 /* Are we moving the last element? */ | |
1786 if (G.page_tails[order] == p) | |
1787 G.page_tails[order] = previous; | |
1788 p = previous; | |
1789 } | |
1790 | |
1791 previous = p; | |
1792 p = next; | |
1793 } | |
1794 while (! done); | |
1795 | |
1796 /* Now, restore the in_use_p vectors for any pages from contexts | |
1797 other than the current one. */ | |
1798 for (p = G.pages[order]; p; p = p->next) | |
1799 if (p->context_depth != G.context_depth) | |
1800 ggc_recalculate_in_use_p (p); | |
1801 } | |
1802 } | |
1803 | |
1804 #ifdef ENABLE_GC_CHECKING | |
1805 /* Clobber all free objects. */ | |
1806 | |
1807 static void | |
1808 poison_pages (void) | |
1809 { | |
1810 unsigned order; | |
1811 | |
1812 for (order = 2; order < NUM_ORDERS; order++) | |
1813 { | |
1814 size_t size = OBJECT_SIZE (order); | |
1815 page_entry *p; | |
1816 | |
1817 for (p = G.pages[order]; p != NULL; p = p->next) | |
1818 { | |
1819 size_t num_objects; | |
1820 size_t i; | |
1821 | |
1822 if (p->context_depth != G.context_depth) | |
1823 /* Since we don't do any collection for pages in pushed | |
1824 contexts, there's no need to do any poisoning. And | |
1825 besides, the IN_USE_P array isn't valid until we pop | |
1826 contexts. */ | |
1827 continue; | |
1828 | |
1829 num_objects = OBJECTS_IN_PAGE (p); | |
1830 for (i = 0; i < num_objects; i++) | |
1831 { | |
1832 size_t word, bit; | |
1833 word = i / HOST_BITS_PER_LONG; | |
1834 bit = i % HOST_BITS_PER_LONG; | |
1835 if (((p->in_use_p[word] >> bit) & 1) == 0) | |
1836 { | |
1837 char *object = p->page + i * size; | |
1838 | |
1839 /* Keep poison-by-write when we expect to use Valgrind, | |
1840 so the exact same memory semantics is kept, in case | |
1841 there are memory errors. We override this request | |
1842 below. */ | |
1843 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (object, | |
1844 size)); | |
1845 memset (object, 0xa5, size); | |
1846 | |
1847 /* Drop the handle to avoid handle leak. */ | |
1848 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (object, size)); | |
1849 } | |
1850 } | |
1851 } | |
1852 } | |
1853 } | |
1854 #else | |
1855 #define poison_pages() | |
1856 #endif | |
1857 | |
1858 #ifdef ENABLE_GC_ALWAYS_COLLECT | |
1859 /* Validate that the reportedly free objects actually are. */ | |
1860 | |
1861 static void | |
1862 validate_free_objects (void) | |
1863 { | |
1864 struct free_object *f, *next, *still_free = NULL; | |
1865 | |
1866 for (f = G.free_object_list; f ; f = next) | |
1867 { | |
1868 page_entry *pe = lookup_page_table_entry (f->object); | |
1869 size_t bit, word; | |
1870 | |
1871 bit = OFFSET_TO_BIT ((char *)f->object - pe->page, pe->order); | |
1872 word = bit / HOST_BITS_PER_LONG; | |
1873 bit = bit % HOST_BITS_PER_LONG; | |
1874 next = f->next; | |
1875 | |
1876 /* Make certain it isn't visible from any root. Notice that we | |
1877 do this check before sweep_pages merges save_in_use_p. */ | |
1878 gcc_assert (!(pe->in_use_p[word] & (1UL << bit))); | |
1879 | |
1880 /* If the object comes from an outer context, then retain the | |
1881 free_object entry, so that we can verify that the address | |
1882 isn't live on the stack in some outer context. */ | |
1883 if (pe->context_depth != G.context_depth) | |
1884 { | |
1885 f->next = still_free; | |
1886 still_free = f; | |
1887 } | |
1888 else | |
1889 free (f); | |
1890 } | |
1891 | |
1892 G.free_object_list = still_free; | |
1893 } | |
1894 #else | |
1895 #define validate_free_objects() | |
1896 #endif | |
1897 | |
1898 /* Top level mark-and-sweep routine. */ | |
1899 | |
1900 void | |
1901 ggc_collect (void) | |
1902 { | |
1903 /* Avoid frequent unnecessary work by skipping collection if the | |
1904 total allocations haven't expanded much since the last | |
1905 collection. */ | |
1906 float allocated_last_gc = | |
1907 MAX (G.allocated_last_gc, (size_t)PARAM_VALUE (GGC_MIN_HEAPSIZE) * 1024); | |
1908 | |
1909 float min_expand = allocated_last_gc * PARAM_VALUE (GGC_MIN_EXPAND) / 100; | |
1910 | |
1911 if (G.allocated < allocated_last_gc + min_expand && !ggc_force_collect) | |
1912 return; | |
1913 | |
1914 timevar_push (TV_GC); | |
1915 if (!quiet_flag) | |
1916 fprintf (stderr, " {GC %luk -> ", (unsigned long) G.allocated / 1024); | |
1917 if (GGC_DEBUG_LEVEL >= 2) | |
1918 fprintf (G.debug_file, "BEGIN COLLECTING\n"); | |
1919 | |
1920 /* Zero the total allocated bytes. This will be recalculated in the | |
1921 sweep phase. */ | |
1922 G.allocated = 0; | |
1923 | |
1924 /* Release the pages we freed the last time we collected, but didn't | |
1925 reuse in the interim. */ | |
1926 release_pages (); | |
1927 | |
1928 /* Indicate that we've seen collections at this context depth. */ | |
1929 G.context_depth_collections = ((unsigned long)1 << (G.context_depth + 1)) - 1; | |
1930 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1931 invoke_plugin_callbacks (PLUGIN_GGC_START, NULL); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1932 |
0 | 1933 clear_marks (); |
1934 ggc_mark_roots (); | |
1935 #ifdef GATHER_STATISTICS | |
1936 ggc_prune_overhead_list (); | |
1937 #endif | |
1938 poison_pages (); | |
1939 validate_free_objects (); | |
1940 sweep_pages (); | |
1941 | |
1942 G.allocated_last_gc = G.allocated; | |
1943 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1944 invoke_plugin_callbacks (PLUGIN_GGC_END, NULL); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1945 |
0 | 1946 timevar_pop (TV_GC); |
1947 | |
1948 if (!quiet_flag) | |
1949 fprintf (stderr, "%luk}", (unsigned long) G.allocated / 1024); | |
1950 if (GGC_DEBUG_LEVEL >= 2) | |
1951 fprintf (G.debug_file, "END COLLECTING\n"); | |
1952 } | |
1953 | |
1954 /* Print allocation statistics. */ | |
1955 #define SCALE(x) ((unsigned long) ((x) < 1024*10 \ | |
1956 ? (x) \ | |
1957 : ((x) < 1024*1024*10 \ | |
1958 ? (x) / 1024 \ | |
1959 : (x) / (1024*1024)))) | |
1960 #define STAT_LABEL(x) ((x) < 1024*10 ? ' ' : ((x) < 1024*1024*10 ? 'k' : 'M')) | |
1961 | |
1962 void | |
1963 ggc_print_statistics (void) | |
1964 { | |
1965 struct ggc_statistics stats; | |
1966 unsigned int i; | |
1967 size_t total_overhead = 0; | |
1968 | |
1969 /* Clear the statistics. */ | |
1970 memset (&stats, 0, sizeof (stats)); | |
1971 | |
1972 /* Make sure collection will really occur. */ | |
1973 G.allocated_last_gc = 0; | |
1974 | |
1975 /* Collect and print the statistics common across collectors. */ | |
1976 ggc_print_common_statistics (stderr, &stats); | |
1977 | |
1978 /* Release free pages so that we will not count the bytes allocated | |
1979 there as part of the total allocated memory. */ | |
1980 release_pages (); | |
1981 | |
1982 /* Collect some information about the various sizes of | |
1983 allocation. */ | |
1984 fprintf (stderr, | |
1985 "Memory still allocated at the end of the compilation process\n"); | |
1986 fprintf (stderr, "%-5s %10s %10s %10s\n", | |
1987 "Size", "Allocated", "Used", "Overhead"); | |
1988 for (i = 0; i < NUM_ORDERS; ++i) | |
1989 { | |
1990 page_entry *p; | |
1991 size_t allocated; | |
1992 size_t in_use; | |
1993 size_t overhead; | |
1994 | |
1995 /* Skip empty entries. */ | |
1996 if (!G.pages[i]) | |
1997 continue; | |
1998 | |
1999 overhead = allocated = in_use = 0; | |
2000 | |
2001 /* Figure out the total number of bytes allocated for objects of | |
2002 this size, and how many of them are actually in use. Also figure | |
2003 out how much memory the page table is using. */ | |
2004 for (p = G.pages[i]; p; p = p->next) | |
2005 { | |
2006 allocated += p->bytes; | |
2007 in_use += | |
2008 (OBJECTS_IN_PAGE (p) - p->num_free_objects) * OBJECT_SIZE (i); | |
2009 | |
2010 overhead += (sizeof (page_entry) - sizeof (long) | |
2011 + BITMAP_SIZE (OBJECTS_IN_PAGE (p) + 1)); | |
2012 } | |
2013 fprintf (stderr, "%-5lu %10lu%c %10lu%c %10lu%c\n", | |
2014 (unsigned long) OBJECT_SIZE (i), | |
2015 SCALE (allocated), STAT_LABEL (allocated), | |
2016 SCALE (in_use), STAT_LABEL (in_use), | |
2017 SCALE (overhead), STAT_LABEL (overhead)); | |
2018 total_overhead += overhead; | |
2019 } | |
2020 fprintf (stderr, "%-5s %10lu%c %10lu%c %10lu%c\n", "Total", | |
2021 SCALE (G.bytes_mapped), STAT_LABEL (G.bytes_mapped), | |
2022 SCALE (G.allocated), STAT_LABEL(G.allocated), | |
2023 SCALE (total_overhead), STAT_LABEL (total_overhead)); | |
2024 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2025 #ifdef GATHER_STATISTICS |
0 | 2026 { |
2027 fprintf (stderr, "\nTotal allocations and overheads during the compilation process\n"); | |
2028 | |
2029 fprintf (stderr, "Total Overhead: %10lld\n", | |
2030 G.stats.total_overhead); | |
2031 fprintf (stderr, "Total Allocated: %10lld\n", | |
2032 G.stats.total_allocated); | |
2033 | |
2034 fprintf (stderr, "Total Overhead under 32B: %10lld\n", | |
2035 G.stats.total_overhead_under32); | |
2036 fprintf (stderr, "Total Allocated under 32B: %10lld\n", | |
2037 G.stats.total_allocated_under32); | |
2038 fprintf (stderr, "Total Overhead under 64B: %10lld\n", | |
2039 G.stats.total_overhead_under64); | |
2040 fprintf (stderr, "Total Allocated under 64B: %10lld\n", | |
2041 G.stats.total_allocated_under64); | |
2042 fprintf (stderr, "Total Overhead under 128B: %10lld\n", | |
2043 G.stats.total_overhead_under128); | |
2044 fprintf (stderr, "Total Allocated under 128B: %10lld\n", | |
2045 G.stats.total_allocated_under128); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2046 |
0 | 2047 for (i = 0; i < NUM_ORDERS; i++) |
2048 if (G.stats.total_allocated_per_order[i]) | |
2049 { | |
2050 fprintf (stderr, "Total Overhead page size %7lu: %10lld\n", | |
2051 (unsigned long) OBJECT_SIZE (i), | |
2052 G.stats.total_overhead_per_order[i]); | |
2053 fprintf (stderr, "Total Allocated page size %7lu: %10lld\n", | |
2054 (unsigned long) OBJECT_SIZE (i), | |
2055 G.stats.total_allocated_per_order[i]); | |
2056 } | |
2057 } | |
2058 #endif | |
2059 } | |
2060 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2061 struct ggc_pch_ondisk |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2062 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2063 unsigned totals[NUM_ORDERS]; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2064 }; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2065 |
0 | 2066 struct ggc_pch_data |
2067 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2068 struct ggc_pch_ondisk d; |
0 | 2069 size_t base[NUM_ORDERS]; |
2070 size_t written[NUM_ORDERS]; | |
2071 }; | |
2072 | |
2073 struct ggc_pch_data * | |
2074 init_ggc_pch (void) | |
2075 { | |
2076 return XCNEW (struct ggc_pch_data); | |
2077 } | |
2078 | |
2079 void | |
2080 ggc_pch_count_object (struct ggc_pch_data *d, void *x ATTRIBUTE_UNUSED, | |
2081 size_t size, bool is_string ATTRIBUTE_UNUSED, | |
2082 enum gt_types_enum type ATTRIBUTE_UNUSED) | |
2083 { | |
2084 unsigned order; | |
2085 | |
2086 if (size < NUM_SIZE_LOOKUP) | |
2087 order = size_lookup[size]; | |
2088 else | |
2089 { | |
2090 order = 10; | |
2091 while (size > OBJECT_SIZE (order)) | |
2092 order++; | |
2093 } | |
2094 | |
2095 d->d.totals[order]++; | |
2096 } | |
2097 | |
2098 size_t | |
2099 ggc_pch_total_size (struct ggc_pch_data *d) | |
2100 { | |
2101 size_t a = 0; | |
2102 unsigned i; | |
2103 | |
2104 for (i = 0; i < NUM_ORDERS; i++) | |
2105 a += ROUND_UP (d->d.totals[i] * OBJECT_SIZE (i), G.pagesize); | |
2106 return a; | |
2107 } | |
2108 | |
2109 void | |
2110 ggc_pch_this_base (struct ggc_pch_data *d, void *base) | |
2111 { | |
2112 size_t a = (size_t) base; | |
2113 unsigned i; | |
2114 | |
2115 for (i = 0; i < NUM_ORDERS; i++) | |
2116 { | |
2117 d->base[i] = a; | |
2118 a += ROUND_UP (d->d.totals[i] * OBJECT_SIZE (i), G.pagesize); | |
2119 } | |
2120 } | |
2121 | |
2122 | |
2123 char * | |
2124 ggc_pch_alloc_object (struct ggc_pch_data *d, void *x ATTRIBUTE_UNUSED, | |
2125 size_t size, bool is_string ATTRIBUTE_UNUSED, | |
2126 enum gt_types_enum type ATTRIBUTE_UNUSED) | |
2127 { | |
2128 unsigned order; | |
2129 char *result; | |
2130 | |
2131 if (size < NUM_SIZE_LOOKUP) | |
2132 order = size_lookup[size]; | |
2133 else | |
2134 { | |
2135 order = 10; | |
2136 while (size > OBJECT_SIZE (order)) | |
2137 order++; | |
2138 } | |
2139 | |
2140 result = (char *) d->base[order]; | |
2141 d->base[order] += OBJECT_SIZE (order); | |
2142 return result; | |
2143 } | |
2144 | |
2145 void | |
2146 ggc_pch_prepare_write (struct ggc_pch_data *d ATTRIBUTE_UNUSED, | |
2147 FILE *f ATTRIBUTE_UNUSED) | |
2148 { | |
2149 /* Nothing to do. */ | |
2150 } | |
2151 | |
2152 void | |
2153 ggc_pch_write_object (struct ggc_pch_data *d ATTRIBUTE_UNUSED, | |
2154 FILE *f, void *x, void *newx ATTRIBUTE_UNUSED, | |
2155 size_t size, bool is_string ATTRIBUTE_UNUSED) | |
2156 { | |
2157 unsigned order; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2158 static const char emptyBytes[256] = { 0 }; |
0 | 2159 |
2160 if (size < NUM_SIZE_LOOKUP) | |
2161 order = size_lookup[size]; | |
2162 else | |
2163 { | |
2164 order = 10; | |
2165 while (size > OBJECT_SIZE (order)) | |
2166 order++; | |
2167 } | |
2168 | |
2169 if (fwrite (x, size, 1, f) != 1) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2170 fatal_error ("can%'t write PCH file: %m"); |
0 | 2171 |
2172 /* If SIZE is not the same as OBJECT_SIZE(order), then we need to pad the | |
2173 object out to OBJECT_SIZE(order). This happens for strings. */ | |
2174 | |
2175 if (size != OBJECT_SIZE (order)) | |
2176 { | |
2177 unsigned padding = OBJECT_SIZE(order) - size; | |
2178 | |
2179 /* To speed small writes, we use a nulled-out array that's larger | |
2180 than most padding requests as the source for our null bytes. This | |
2181 permits us to do the padding with fwrite() rather than fseek(), and | |
2182 limits the chance the OS may try to flush any outstanding writes. */ | |
2183 if (padding <= sizeof(emptyBytes)) | |
2184 { | |
2185 if (fwrite (emptyBytes, 1, padding, f) != padding) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2186 fatal_error ("can%'t write PCH file"); |
0 | 2187 } |
2188 else | |
2189 { | |
2190 /* Larger than our buffer? Just default to fseek. */ | |
2191 if (fseek (f, padding, SEEK_CUR) != 0) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2192 fatal_error ("can%'t write PCH file"); |
0 | 2193 } |
2194 } | |
2195 | |
2196 d->written[order]++; | |
2197 if (d->written[order] == d->d.totals[order] | |
2198 && fseek (f, ROUND_UP_VALUE (d->d.totals[order] * OBJECT_SIZE (order), | |
2199 G.pagesize), | |
2200 SEEK_CUR) != 0) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2201 fatal_error ("can%'t write PCH file: %m"); |
0 | 2202 } |
2203 | |
2204 void | |
2205 ggc_pch_finish (struct ggc_pch_data *d, FILE *f) | |
2206 { | |
2207 if (fwrite (&d->d, sizeof (d->d), 1, f) != 1) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2208 fatal_error ("can%'t write PCH file: %m"); |
0 | 2209 free (d); |
2210 } | |
2211 | |
2212 /* Move the PCH PTE entries just added to the end of by_depth, to the | |
2213 front. */ | |
2214 | |
2215 static void | |
2216 move_ptes_to_front (int count_old_page_tables, int count_new_page_tables) | |
2217 { | |
2218 unsigned i; | |
2219 | |
2220 /* First, we swap the new entries to the front of the varrays. */ | |
2221 page_entry **new_by_depth; | |
2222 unsigned long **new_save_in_use; | |
2223 | |
2224 new_by_depth = XNEWVEC (page_entry *, G.by_depth_max); | |
2225 new_save_in_use = XNEWVEC (unsigned long *, G.by_depth_max); | |
2226 | |
2227 memcpy (&new_by_depth[0], | |
2228 &G.by_depth[count_old_page_tables], | |
2229 count_new_page_tables * sizeof (void *)); | |
2230 memcpy (&new_by_depth[count_new_page_tables], | |
2231 &G.by_depth[0], | |
2232 count_old_page_tables * sizeof (void *)); | |
2233 memcpy (&new_save_in_use[0], | |
2234 &G.save_in_use[count_old_page_tables], | |
2235 count_new_page_tables * sizeof (void *)); | |
2236 memcpy (&new_save_in_use[count_new_page_tables], | |
2237 &G.save_in_use[0], | |
2238 count_old_page_tables * sizeof (void *)); | |
2239 | |
2240 free (G.by_depth); | |
2241 free (G.save_in_use); | |
2242 | |
2243 G.by_depth = new_by_depth; | |
2244 G.save_in_use = new_save_in_use; | |
2245 | |
2246 /* Now update all the index_by_depth fields. */ | |
2247 for (i = G.by_depth_in_use; i > 0; --i) | |
2248 { | |
2249 page_entry *p = G.by_depth[i-1]; | |
2250 p->index_by_depth = i-1; | |
2251 } | |
2252 | |
2253 /* And last, we update the depth pointers in G.depth. The first | |
2254 entry is already 0, and context 0 entries always start at index | |
2255 0, so there is nothing to update in the first slot. We need a | |
2256 second slot, only if we have old ptes, and if we do, they start | |
2257 at index count_new_page_tables. */ | |
2258 if (count_old_page_tables) | |
2259 push_depth (count_new_page_tables); | |
2260 } | |
2261 | |
2262 void | |
2263 ggc_pch_read (FILE *f, void *addr) | |
2264 { | |
2265 struct ggc_pch_ondisk d; | |
2266 unsigned i; | |
2267 char *offs = (char *) addr; | |
2268 unsigned long count_old_page_tables; | |
2269 unsigned long count_new_page_tables; | |
2270 | |
2271 count_old_page_tables = G.by_depth_in_use; | |
2272 | |
2273 /* We've just read in a PCH file. So, every object that used to be | |
2274 allocated is now free. */ | |
2275 clear_marks (); | |
2276 #ifdef ENABLE_GC_CHECKING | |
2277 poison_pages (); | |
2278 #endif | |
2279 /* Since we free all the allocated objects, the free list becomes | |
2280 useless. Validate it now, which will also clear it. */ | |
2281 validate_free_objects(); | |
2282 | |
2283 /* No object read from a PCH file should ever be freed. So, set the | |
2284 context depth to 1, and set the depth of all the currently-allocated | |
2285 pages to be 1 too. PCH pages will have depth 0. */ | |
2286 gcc_assert (!G.context_depth); | |
2287 G.context_depth = 1; | |
2288 for (i = 0; i < NUM_ORDERS; i++) | |
2289 { | |
2290 page_entry *p; | |
2291 for (p = G.pages[i]; p != NULL; p = p->next) | |
2292 p->context_depth = G.context_depth; | |
2293 } | |
2294 | |
2295 /* Allocate the appropriate page-table entries for the pages read from | |
2296 the PCH file. */ | |
2297 if (fread (&d, sizeof (d), 1, f) != 1) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2298 fatal_error ("can%'t read PCH file: %m"); |
0 | 2299 |
2300 for (i = 0; i < NUM_ORDERS; i++) | |
2301 { | |
2302 struct page_entry *entry; | |
2303 char *pte; | |
2304 size_t bytes; | |
2305 size_t num_objs; | |
2306 size_t j; | |
2307 | |
2308 if (d.totals[i] == 0) | |
2309 continue; | |
2310 | |
2311 bytes = ROUND_UP (d.totals[i] * OBJECT_SIZE (i), G.pagesize); | |
2312 num_objs = bytes / OBJECT_SIZE (i); | |
2313 entry = XCNEWVAR (struct page_entry, (sizeof (struct page_entry) | |
2314 - sizeof (long) | |
2315 + BITMAP_SIZE (num_objs + 1))); | |
2316 entry->bytes = bytes; | |
2317 entry->page = offs; | |
2318 entry->context_depth = 0; | |
2319 offs += bytes; | |
2320 entry->num_free_objects = 0; | |
2321 entry->order = i; | |
2322 | |
2323 for (j = 0; | |
2324 j + HOST_BITS_PER_LONG <= num_objs + 1; | |
2325 j += HOST_BITS_PER_LONG) | |
2326 entry->in_use_p[j / HOST_BITS_PER_LONG] = -1; | |
2327 for (; j < num_objs + 1; j++) | |
2328 entry->in_use_p[j / HOST_BITS_PER_LONG] | |
2329 |= 1L << (j % HOST_BITS_PER_LONG); | |
2330 | |
2331 for (pte = entry->page; | |
2332 pte < entry->page + entry->bytes; | |
2333 pte += G.pagesize) | |
2334 set_page_table_entry (pte, entry); | |
2335 | |
2336 if (G.page_tails[i] != NULL) | |
2337 G.page_tails[i]->next = entry; | |
2338 else | |
2339 G.pages[i] = entry; | |
2340 G.page_tails[i] = entry; | |
2341 | |
2342 /* We start off by just adding all the new information to the | |
2343 end of the varrays, later, we will move the new information | |
2344 to the front of the varrays, as the PCH page tables are at | |
2345 context 0. */ | |
2346 push_by_depth (entry, 0); | |
2347 } | |
2348 | |
2349 /* Now, we update the various data structures that speed page table | |
2350 handling. */ | |
2351 count_new_page_tables = G.by_depth_in_use - count_old_page_tables; | |
2352 | |
2353 move_ptes_to_front (count_old_page_tables, count_new_page_tables); | |
2354 | |
2355 /* Update the statistics. */ | |
2356 G.allocated = G.allocated_last_gc = offs - (char *)addr; | |
2357 } | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2358 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2359 struct alloc_zone |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2360 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2361 int dummy; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2362 }; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2363 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2364 struct alloc_zone rtl_zone; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2365 struct alloc_zone tree_zone; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2366 struct alloc_zone tree_id_zone; |