Mercurial > hg > CbC > CbC_gcc
annotate gcc/ggc-page.c @ 111:04ced10e8804
gcc 7
author | kono |
---|---|
date | Fri, 27 Oct 2017 22:46:09 +0900 |
parents | f6334be47118 |
children | 84e7813d76e9 |
rev | line source |
---|---|
0 | 1 /* "Bag-of-pages" garbage collector for the GNU compiler. |
111 | 2 Copyright (C) 1999-2017 Free Software Foundation, Inc. |
0 | 3 |
4 This file is part of GCC. | |
5 | |
6 GCC is free software; you can redistribute it and/or modify it under | |
7 the terms of the GNU General Public License as published by the Free | |
8 Software Foundation; either version 3, or (at your option) any later | |
9 version. | |
10 | |
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 for more details. | |
15 | |
16 You should have received a copy of the GNU General Public License | |
17 along with GCC; see the file COPYING3. If not see | |
18 <http://www.gnu.org/licenses/>. */ | |
19 | |
20 #include "config.h" | |
21 #include "system.h" | |
22 #include "coretypes.h" | |
111 | 23 #include "backend.h" |
24 #include "alias.h" | |
0 | 25 #include "tree.h" |
26 #include "rtl.h" | |
111 | 27 #include "memmodel.h" |
0 | 28 #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
|
29 #include "diagnostic-core.h" |
0 | 30 #include "flags.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" | |
111 | 34 #include "cgraph.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 | |
111 | 53 #if defined(HAVE_MADVISE) && HAVE_DECL_MADVISE && defined(MADV_DONTNEED) \ |
54 && defined(USING_MMAP) | |
55 # define USING_MADVISE | |
56 #endif | |
57 | |
0 | 58 /* Strategy: |
59 | |
60 This garbage-collecting allocator allocates objects on one of a set | |
61 of pages. Each page can allocate objects of a single size only; | |
62 available sizes are powers of two starting at four bytes. The size | |
63 of an allocation request is rounded up to the next power of two | |
64 (`order'), and satisfied from the appropriate page. | |
65 | |
66 Each page is recorded in a page-entry, which also maintains an | |
67 in-use bitmap of object positions on the page. This allows the | |
68 allocation state of a particular object to be flipped without | |
69 touching the page itself. | |
70 | |
71 Each page-entry also has a context depth, which is used to track | |
72 pushing and popping of allocation contexts. Only objects allocated | |
73 in the current (highest-numbered) context may be collected. | |
74 | |
75 Page entries are arranged in an array of singly-linked lists. The | |
76 array is indexed by the allocation size, in bits, of the pages on | |
77 it; i.e. all pages on a list allocate objects of the same size. | |
78 Pages are ordered on the list such that all non-full pages precede | |
79 all full pages, with non-full pages arranged in order of decreasing | |
80 context depth. | |
81 | |
82 Empty pages (of all orders) are kept on a single page cache list, | |
83 and are considered first when new pages are required; they are | |
84 deallocated at the start of the next collection if they haven't | |
85 been recycled by then. */ | |
86 | |
87 /* Define GGC_DEBUG_LEVEL to print debugging information. | |
88 0: No debugging output. | |
89 1: GC statistics only. | |
90 2: Page-entry allocations/deallocations as well. | |
91 3: Object allocations as well. | |
92 4: Object marks as well. */ | |
93 #define GGC_DEBUG_LEVEL (0) | |
94 | |
95 #ifndef HOST_BITS_PER_PTR | |
96 #define HOST_BITS_PER_PTR HOST_BITS_PER_LONG | |
97 #endif | |
98 | |
99 | |
100 /* A two-level tree is used to look up the page-entry for a given | |
101 pointer. Two chunks of the pointer's bits are extracted to index | |
102 the first and second levels of the tree, as follows: | |
103 | |
104 HOST_PAGE_SIZE_BITS | |
105 32 | | | |
106 msb +----------------+----+------+------+ lsb | |
107 | | | | |
108 PAGE_L1_BITS | | |
109 | | | |
110 PAGE_L2_BITS | |
111 | |
112 The bottommost HOST_PAGE_SIZE_BITS are ignored, since page-entry | |
113 pages are aligned on system page boundaries. The next most | |
114 significant PAGE_L2_BITS and PAGE_L1_BITS are the second and first | |
115 index values in the lookup table, respectively. | |
116 | |
117 For 32-bit architectures and the settings below, there are no | |
118 leftover bits. For architectures with wider pointers, the lookup | |
119 tree points to a list of pages, which must be scanned to find the | |
120 correct one. */ | |
121 | |
122 #define PAGE_L1_BITS (8) | |
123 #define PAGE_L2_BITS (32 - PAGE_L1_BITS - G.lg_pagesize) | |
111 | 124 #define PAGE_L1_SIZE ((uintptr_t) 1 << PAGE_L1_BITS) |
125 #define PAGE_L2_SIZE ((uintptr_t) 1 << PAGE_L2_BITS) | |
0 | 126 |
127 #define LOOKUP_L1(p) \ | |
111 | 128 (((uintptr_t) (p) >> (32 - PAGE_L1_BITS)) & ((1 << PAGE_L1_BITS) - 1)) |
0 | 129 |
130 #define LOOKUP_L2(p) \ | |
111 | 131 (((uintptr_t) (p) >> G.lg_pagesize) & ((1 << PAGE_L2_BITS) - 1)) |
0 | 132 |
133 /* The number of objects per allocation page, for objects on a page of | |
134 the indicated ORDER. */ | |
135 #define OBJECTS_PER_PAGE(ORDER) objects_per_page_table[ORDER] | |
136 | |
137 /* The number of objects in P. */ | |
138 #define OBJECTS_IN_PAGE(P) ((P)->bytes / OBJECT_SIZE ((P)->order)) | |
139 | |
140 /* The size of an object on a page of the indicated ORDER. */ | |
141 #define OBJECT_SIZE(ORDER) object_size_table[ORDER] | |
142 | |
143 /* For speed, we avoid doing a general integer divide to locate the | |
144 offset in the allocation bitmap, by precalculating numbers M, S | |
145 such that (O * M) >> S == O / Z (modulo 2^32), for any offset O | |
146 within the page which is evenly divisible by the object size Z. */ | |
147 #define DIV_MULT(ORDER) inverse_table[ORDER].mult | |
148 #define DIV_SHIFT(ORDER) inverse_table[ORDER].shift | |
149 #define OFFSET_TO_BIT(OFFSET, ORDER) \ | |
150 (((OFFSET) * DIV_MULT (ORDER)) >> DIV_SHIFT (ORDER)) | |
151 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
152 /* 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
|
153 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
|
154 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
|
155 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
|
156 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
157 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
|
158 char c; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
159 union { |
111 | 160 int64_t i; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
161 void *p; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
162 } 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 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
165 /* 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
|
166 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
167 #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
|
168 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
169 |
0 | 170 /* The number of extra orders, not corresponding to power-of-two sized |
171 objects. */ | |
172 | |
173 #define NUM_EXTRA_ORDERS ARRAY_SIZE (extra_order_size_table) | |
174 | |
175 #define RTL_SIZE(NSLOTS) \ | |
176 (RTX_HDR_SIZE + (NSLOTS) * sizeof (rtunion)) | |
177 | |
178 #define TREE_EXP_SIZE(OPS) \ | |
179 (sizeof (struct tree_exp) + ((OPS) - 1) * sizeof (tree)) | |
180 | |
181 /* The Ith entry is the maximum size of an object to be stored in the | |
182 Ith extra order. Adding a new entry to this array is the *only* | |
183 thing you need to do to add a new special allocation size. */ | |
184 | |
185 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
|
186 /* 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
|
187 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
|
188 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
|
189 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
|
190 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
|
191 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
|
192 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
|
193 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
|
194 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
|
195 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
|
196 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
|
197 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
|
198 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
|
199 MAX_ALIGNMENT * 15, |
0 | 200 sizeof (struct tree_decl_non_common), |
201 sizeof (struct tree_field_decl), | |
202 sizeof (struct tree_parm_decl), | |
203 sizeof (struct tree_var_decl), | |
111 | 204 sizeof (struct tree_type_non_common), |
0 | 205 sizeof (struct function), |
206 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
|
207 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
|
208 sizeof (struct loop), |
0 | 209 }; |
210 | |
211 /* The total number of orders. */ | |
212 | |
213 #define NUM_ORDERS (HOST_BITS_PER_PTR + NUM_EXTRA_ORDERS) | |
214 | |
215 /* Compute the smallest nonnegative number which when added to X gives | |
216 a multiple of F. */ | |
217 | |
218 #define ROUND_UP_VALUE(x, f) ((f) - 1 - ((f) - 1 + (x)) % (f)) | |
219 | |
111 | 220 /* Round X to next multiple of the page size */ |
221 | |
222 #define PAGE_ALIGN(x) ROUND_UP ((x), G.pagesize) | |
0 | 223 |
224 /* The Ith entry is the number of objects on a page or order I. */ | |
225 | |
226 static unsigned objects_per_page_table[NUM_ORDERS]; | |
227 | |
228 /* The Ith entry is the size of an object on a page of order I. */ | |
229 | |
230 static size_t object_size_table[NUM_ORDERS]; | |
231 | |
232 /* The Ith entry is a pair of numbers (mult, shift) such that | |
233 ((k * mult) >> shift) mod 2^32 == (k / OBJECT_SIZE(I)) mod 2^32, | |
234 for all k evenly divisible by OBJECT_SIZE(I). */ | |
235 | |
236 static struct | |
237 { | |
238 size_t mult; | |
239 unsigned int shift; | |
240 } | |
241 inverse_table[NUM_ORDERS]; | |
242 | |
243 /* A page_entry records the status of an allocation page. This | |
244 structure is dynamically sized to fit the bitmap in_use_p. */ | |
111 | 245 struct page_entry |
0 | 246 { |
247 /* The next page-entry with objects of the same size, or NULL if | |
248 this is the last page-entry. */ | |
249 struct page_entry *next; | |
250 | |
251 /* The previous page-entry with objects of the same size, or NULL if | |
252 this is the first page-entry. The PREV pointer exists solely to | |
253 keep the cost of ggc_free manageable. */ | |
254 struct page_entry *prev; | |
255 | |
256 /* The number of bytes allocated. (This will always be a multiple | |
257 of the host system page size.) */ | |
258 size_t bytes; | |
259 | |
260 /* The address at which the memory is allocated. */ | |
261 char *page; | |
262 | |
263 #ifdef USING_MALLOC_PAGE_GROUPS | |
264 /* Back pointer to the page group this page came from. */ | |
265 struct page_group *group; | |
266 #endif | |
267 | |
268 /* This is the index in the by_depth varray where this page table | |
269 can be found. */ | |
270 unsigned long index_by_depth; | |
271 | |
272 /* Context depth of this page. */ | |
273 unsigned short context_depth; | |
274 | |
275 /* The number of free objects remaining on this page. */ | |
276 unsigned short num_free_objects; | |
277 | |
278 /* A likely candidate for the bit position of a free object for the | |
279 next allocation from this page. */ | |
280 unsigned short next_bit_hint; | |
281 | |
282 /* The lg of size of objects allocated from this page. */ | |
283 unsigned char order; | |
284 | |
111 | 285 /* Discarded page? */ |
286 bool discarded; | |
287 | |
0 | 288 /* A bit vector indicating whether or not objects are in use. The |
289 Nth bit is one if the Nth object on this page is allocated. This | |
290 array is dynamically sized. */ | |
291 unsigned long in_use_p[1]; | |
111 | 292 }; |
0 | 293 |
294 #ifdef USING_MALLOC_PAGE_GROUPS | |
295 /* A page_group describes a large allocation from malloc, from which | |
296 we parcel out aligned pages. */ | |
111 | 297 struct page_group |
0 | 298 { |
299 /* A linked list of all extant page groups. */ | |
300 struct page_group *next; | |
301 | |
302 /* The address we received from malloc. */ | |
303 char *allocation; | |
304 | |
305 /* The size of the block. */ | |
306 size_t alloc_size; | |
307 | |
308 /* A bitmask of pages in use. */ | |
309 unsigned int in_use; | |
111 | 310 }; |
0 | 311 #endif |
312 | |
313 #if HOST_BITS_PER_PTR <= 32 | |
314 | |
315 /* On 32-bit hosts, we use a two level page table, as pictured above. */ | |
316 typedef page_entry **page_table[PAGE_L1_SIZE]; | |
317 | |
318 #else | |
319 | |
320 /* On 64-bit hosts, we use the same two level page tables plus a linked | |
321 list that disambiguates the top 32-bits. There will almost always be | |
322 exactly one entry in the list. */ | |
323 typedef struct page_table_chain | |
324 { | |
325 struct page_table_chain *next; | |
326 size_t high_bits; | |
327 page_entry **table[PAGE_L1_SIZE]; | |
328 } *page_table; | |
329 | |
330 #endif | |
331 | |
111 | 332 class finalizer |
333 { | |
334 public: | |
335 finalizer (void *addr, void (*f)(void *)) : m_addr (addr), m_function (f) {} | |
336 | |
337 void *addr () const { return m_addr; } | |
338 | |
339 void call () const { m_function (m_addr); } | |
340 | |
341 private: | |
342 void *m_addr; | |
343 void (*m_function)(void *); | |
344 }; | |
345 | |
346 class vec_finalizer | |
347 { | |
348 public: | |
349 vec_finalizer (uintptr_t addr, void (*f)(void *), size_t s, size_t n) : | |
350 m_addr (addr), m_function (f), m_object_size (s), m_n_objects (n) {} | |
351 | |
352 void call () const | |
353 { | |
354 for (size_t i = 0; i < m_n_objects; i++) | |
355 m_function (reinterpret_cast<void *> (m_addr + (i * m_object_size))); | |
356 } | |
357 | |
358 void *addr () const { return reinterpret_cast<void *> (m_addr); } | |
359 | |
360 private: | |
361 uintptr_t m_addr; | |
362 void (*m_function)(void *); | |
363 size_t m_object_size; | |
364 size_t m_n_objects; | |
365 }; | |
366 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
367 #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
|
368 /* 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
|
369 next collection. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
370 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
|
371 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
372 void *object; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
373 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
|
374 }; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
375 #endif |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
376 |
0 | 377 /* The rest of the global variables. */ |
111 | 378 static struct ggc_globals |
0 | 379 { |
380 /* The Nth element in this array is a page with objects of size 2^N. | |
381 If there are any pages with free objects, they will be at the | |
382 head of the list. NULL if there are no page-entries for this | |
383 object size. */ | |
384 page_entry *pages[NUM_ORDERS]; | |
385 | |
386 /* The Nth element in this array is the last page with objects of | |
387 size 2^N. NULL if there are no page-entries for this object | |
388 size. */ | |
389 page_entry *page_tails[NUM_ORDERS]; | |
390 | |
391 /* Lookup table for associating allocation pages with object addresses. */ | |
392 page_table lookup; | |
393 | |
394 /* The system's page size. */ | |
395 size_t pagesize; | |
396 size_t lg_pagesize; | |
397 | |
398 /* Bytes currently allocated. */ | |
399 size_t allocated; | |
400 | |
401 /* Bytes currently allocated at the end of the last collection. */ | |
402 size_t allocated_last_gc; | |
403 | |
404 /* Total amount of memory mapped. */ | |
405 size_t bytes_mapped; | |
406 | |
407 /* Bit N set if any allocations have been done at context depth N. */ | |
408 unsigned long context_depth_allocations; | |
409 | |
410 /* Bit N set if any collections have been done at context depth N. */ | |
411 unsigned long context_depth_collections; | |
412 | |
413 /* The current depth in the context stack. */ | |
414 unsigned short context_depth; | |
415 | |
416 /* A file descriptor open to /dev/zero for reading. */ | |
417 #if defined (HAVE_MMAP_DEV_ZERO) | |
418 int dev_zero_fd; | |
419 #endif | |
420 | |
421 /* A cache of free system pages. */ | |
422 page_entry *free_pages; | |
423 | |
424 #ifdef USING_MALLOC_PAGE_GROUPS | |
425 page_group *page_groups; | |
426 #endif | |
427 | |
428 /* The file descriptor for debugging output. */ | |
429 FILE *debug_file; | |
430 | |
431 /* Current number of elements in use in depth below. */ | |
432 unsigned int depth_in_use; | |
433 | |
434 /* Maximum number of elements that can be used before resizing. */ | |
435 unsigned int depth_max; | |
436 | |
437 /* Each element of this array is an index in by_depth where the given | |
438 depth starts. This structure is indexed by that given depth we | |
439 are interested in. */ | |
440 unsigned int *depth; | |
441 | |
442 /* Current number of elements in use in by_depth below. */ | |
443 unsigned int by_depth_in_use; | |
444 | |
445 /* Maximum number of elements that can be used before resizing. */ | |
446 unsigned int by_depth_max; | |
447 | |
448 /* Each element of this array is a pointer to a page_entry, all | |
449 page_entries can be found in here by increasing depth. | |
450 index_by_depth in the page_entry is the index into this data | |
451 structure where that page_entry can be found. This is used to | |
452 speed up finding all page_entries at a particular depth. */ | |
453 page_entry **by_depth; | |
454 | |
455 /* Each element is a pointer to the saved in_use_p bits, if any, | |
456 zero otherwise. We allocate them all together, to enable a | |
457 better runtime data access pattern. */ | |
458 unsigned long **save_in_use; | |
459 | |
111 | 460 /* Finalizers for single objects. The first index is collection_depth. */ |
461 vec<vec<finalizer> > finalizers; | |
462 | |
463 /* Finalizers for vectors of objects. */ | |
464 vec<vec<vec_finalizer> > vec_finalizers; | |
465 | |
0 | 466 #ifdef ENABLE_GC_ALWAYS_COLLECT |
467 /* List of free objects to be verified as actually free on the | |
468 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
|
469 struct free_object *free_object_list; |
0 | 470 #endif |
471 | |
472 struct | |
473 { | |
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
|
474 /* Total GC-allocated memory. */ |
0 | 475 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
|
476 /* Total overhead for GC-allocated memory. */ |
0 | 477 unsigned long long total_overhead; |
478 | |
479 /* Total allocations and overhead for sizes less than 32, 64 and 128. | |
480 These sizes are interesting because they are typical cache line | |
481 sizes. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
482 |
0 | 483 unsigned long long total_allocated_under32; |
484 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
|
485 |
0 | 486 unsigned long long total_allocated_under64; |
487 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
|
488 |
0 | 489 unsigned long long total_allocated_under128; |
490 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
|
491 |
0 | 492 /* The allocations for each of the allocation orders. */ |
493 unsigned long long total_allocated_per_order[NUM_ORDERS]; | |
494 | |
495 /* The overhead for each of the allocation orders. */ | |
496 unsigned long long total_overhead_per_order[NUM_ORDERS]; | |
497 } stats; | |
498 } G; | |
499 | |
111 | 500 /* True if a gc is currently taking place. */ |
501 | |
502 static bool in_gc = false; | |
503 | |
0 | 504 /* The size in bytes required to maintain a bitmap for the objects |
505 on a page-entry. */ | |
506 #define BITMAP_SIZE(Num_objects) \ | |
111 | 507 (CEIL ((Num_objects), HOST_BITS_PER_LONG) * sizeof (long)) |
0 | 508 |
509 /* Allocate pages in chunks of this size, to throttle calls to memory | |
510 allocation routines. The first page is used, the rest go onto the | |
511 free list. This cannot be larger than HOST_BITS_PER_INT for the | |
512 in_use bitmask for page_group. Hosts that need a different value | |
513 can override this by defining GGC_QUIRE_SIZE explicitly. */ | |
514 #ifndef GGC_QUIRE_SIZE | |
515 # ifdef USING_MMAP | |
111 | 516 # define GGC_QUIRE_SIZE 512 /* 2MB for 4K pages */ |
0 | 517 # else |
518 # define GGC_QUIRE_SIZE 16 | |
519 # endif | |
520 #endif | |
521 | |
522 /* Initial guess as to how many page table entries we might need. */ | |
523 #define INITIAL_PTE_COUNT 128 | |
524 | |
525 static page_entry *lookup_page_table_entry (const void *); | |
526 static void set_page_table_entry (void *, page_entry *); | |
527 #ifdef USING_MMAP | |
111 | 528 static char *alloc_anon (char *, size_t, bool check); |
0 | 529 #endif |
530 #ifdef USING_MALLOC_PAGE_GROUPS | |
531 static size_t page_group_index (char *, char *); | |
532 static void set_page_group_in_use (page_group *, char *); | |
533 static void clear_page_group_in_use (page_group *, char *); | |
534 #endif | |
535 static struct page_entry * alloc_page (unsigned); | |
536 static void free_page (struct page_entry *); | |
537 static void release_pages (void); | |
538 static void clear_marks (void); | |
539 static void sweep_pages (void); | |
540 static void ggc_recalculate_in_use_p (page_entry *); | |
541 static void compute_inverse (unsigned); | |
542 static inline void adjust_depth (void); | |
543 static void move_ptes_to_front (int, int); | |
544 | |
545 void debug_print_page_list (int); | |
546 static void push_depth (unsigned int); | |
547 static void push_by_depth (page_entry *, unsigned long *); | |
548 | |
549 /* Push an entry onto G.depth. */ | |
550 | |
551 inline static void | |
552 push_depth (unsigned int i) | |
553 { | |
554 if (G.depth_in_use >= G.depth_max) | |
555 { | |
556 G.depth_max *= 2; | |
557 G.depth = XRESIZEVEC (unsigned int, G.depth, G.depth_max); | |
558 } | |
559 G.depth[G.depth_in_use++] = i; | |
560 } | |
561 | |
562 /* Push an entry onto G.by_depth and G.save_in_use. */ | |
563 | |
564 inline static void | |
565 push_by_depth (page_entry *p, unsigned long *s) | |
566 { | |
567 if (G.by_depth_in_use >= G.by_depth_max) | |
568 { | |
569 G.by_depth_max *= 2; | |
570 G.by_depth = XRESIZEVEC (page_entry *, G.by_depth, G.by_depth_max); | |
571 G.save_in_use = XRESIZEVEC (unsigned long *, G.save_in_use, | |
572 G.by_depth_max); | |
573 } | |
574 G.by_depth[G.by_depth_in_use] = p; | |
575 G.save_in_use[G.by_depth_in_use++] = s; | |
576 } | |
577 | |
578 #if (GCC_VERSION < 3001) | |
579 #define prefetch(X) ((void) X) | |
580 #else | |
581 #define prefetch(X) __builtin_prefetch (X) | |
582 #endif | |
583 | |
584 #define save_in_use_p_i(__i) \ | |
585 (G.save_in_use[__i]) | |
586 #define save_in_use_p(__p) \ | |
587 (save_in_use_p_i (__p->index_by_depth)) | |
588 | |
111 | 589 /* Traverse the page table and find the entry for a page. |
590 If the object wasn't allocated in GC return NULL. */ | |
591 | |
592 static inline page_entry * | |
593 safe_lookup_page_table_entry (const void *p) | |
0 | 594 { |
595 page_entry ***base; | |
596 size_t L1, L2; | |
597 | |
598 #if HOST_BITS_PER_PTR <= 32 | |
599 base = &G.lookup[0]; | |
600 #else | |
601 page_table table = G.lookup; | |
111 | 602 uintptr_t high_bits = (uintptr_t) p & ~ (uintptr_t) 0xffffffff; |
0 | 603 while (1) |
604 { | |
605 if (table == NULL) | |
111 | 606 return NULL; |
0 | 607 if (table->high_bits == high_bits) |
608 break; | |
609 table = table->next; | |
610 } | |
611 base = &table->table[0]; | |
612 #endif | |
613 | |
614 /* Extract the level 1 and 2 indices. */ | |
615 L1 = LOOKUP_L1 (p); | |
616 L2 = LOOKUP_L2 (p); | |
111 | 617 if (! base[L1]) |
618 return NULL; | |
619 | |
620 return base[L1][L2]; | |
0 | 621 } |
622 | |
623 /* Traverse the page table and find the entry for a page. | |
624 Die (probably) if the object wasn't allocated via GC. */ | |
625 | |
626 static inline page_entry * | |
627 lookup_page_table_entry (const void *p) | |
628 { | |
629 page_entry ***base; | |
630 size_t L1, L2; | |
631 | |
632 #if HOST_BITS_PER_PTR <= 32 | |
633 base = &G.lookup[0]; | |
634 #else | |
635 page_table table = G.lookup; | |
111 | 636 uintptr_t high_bits = (uintptr_t) p & ~ (uintptr_t) 0xffffffff; |
0 | 637 while (table->high_bits != high_bits) |
638 table = table->next; | |
639 base = &table->table[0]; | |
640 #endif | |
641 | |
642 /* Extract the level 1 and 2 indices. */ | |
643 L1 = LOOKUP_L1 (p); | |
644 L2 = LOOKUP_L2 (p); | |
645 | |
646 return base[L1][L2]; | |
647 } | |
648 | |
649 /* Set the page table entry for a page. */ | |
650 | |
651 static void | |
652 set_page_table_entry (void *p, page_entry *entry) | |
653 { | |
654 page_entry ***base; | |
655 size_t L1, L2; | |
656 | |
657 #if HOST_BITS_PER_PTR <= 32 | |
658 base = &G.lookup[0]; | |
659 #else | |
660 page_table table; | |
111 | 661 uintptr_t high_bits = (uintptr_t) p & ~ (uintptr_t) 0xffffffff; |
0 | 662 for (table = G.lookup; table; table = table->next) |
663 if (table->high_bits == high_bits) | |
664 goto found; | |
665 | |
666 /* Not found -- allocate a new table. */ | |
667 table = XCNEW (struct page_table_chain); | |
668 table->next = G.lookup; | |
669 table->high_bits = high_bits; | |
670 G.lookup = table; | |
671 found: | |
672 base = &table->table[0]; | |
673 #endif | |
674 | |
675 /* Extract the level 1 and 2 indices. */ | |
676 L1 = LOOKUP_L1 (p); | |
677 L2 = LOOKUP_L2 (p); | |
678 | |
679 if (base[L1] == NULL) | |
680 base[L1] = XCNEWVEC (page_entry *, PAGE_L2_SIZE); | |
681 | |
682 base[L1][L2] = entry; | |
683 } | |
684 | |
685 /* Prints the page-entry for object size ORDER, for debugging. */ | |
686 | |
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
|
687 DEBUG_FUNCTION void |
0 | 688 debug_print_page_list (int order) |
689 { | |
690 page_entry *p; | |
691 printf ("Head=%p, Tail=%p:\n", (void *) G.pages[order], | |
692 (void *) G.page_tails[order]); | |
693 p = G.pages[order]; | |
694 while (p != NULL) | |
695 { | |
696 printf ("%p(%1d|%3d) -> ", (void *) p, p->context_depth, | |
697 p->num_free_objects); | |
698 p = p->next; | |
699 } | |
700 printf ("NULL\n"); | |
701 fflush (stdout); | |
702 } | |
703 | |
704 #ifdef USING_MMAP | |
705 /* Allocate SIZE bytes of anonymous memory, preferably near PREF, | |
706 (if non-null). The ifdef structure here is intended to cause a | |
707 compile error unless exactly one of the HAVE_* is defined. */ | |
708 | |
709 static inline char * | |
111 | 710 alloc_anon (char *pref ATTRIBUTE_UNUSED, size_t size, bool check) |
0 | 711 { |
712 #ifdef HAVE_MMAP_ANON | |
713 char *page = (char *) mmap (pref, size, PROT_READ | PROT_WRITE, | |
714 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); | |
715 #endif | |
716 #ifdef HAVE_MMAP_DEV_ZERO | |
717 char *page = (char *) mmap (pref, size, PROT_READ | PROT_WRITE, | |
718 MAP_PRIVATE, G.dev_zero_fd, 0); | |
719 #endif | |
720 | |
721 if (page == (char *) MAP_FAILED) | |
722 { | |
111 | 723 if (!check) |
724 return NULL; | |
0 | 725 perror ("virtual memory exhausted"); |
726 exit (FATAL_EXIT_CODE); | |
727 } | |
728 | |
729 /* Remember that we allocated this memory. */ | |
730 G.bytes_mapped += size; | |
731 | |
732 /* 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
|
733 access to smaller pieces of the area in ggc_internal_alloc. Discard the |
0 | 734 handle to avoid handle leak. */ |
735 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (page, size)); | |
736 | |
737 return page; | |
738 } | |
739 #endif | |
740 #ifdef USING_MALLOC_PAGE_GROUPS | |
741 /* Compute the index for this page into the page group. */ | |
742 | |
743 static inline size_t | |
744 page_group_index (char *allocation, char *page) | |
745 { | |
746 return (size_t) (page - allocation) >> G.lg_pagesize; | |
747 } | |
748 | |
749 /* Set and clear the in_use bit for this page in the page group. */ | |
750 | |
751 static inline void | |
752 set_page_group_in_use (page_group *group, char *page) | |
753 { | |
754 group->in_use |= 1 << page_group_index (group->allocation, page); | |
755 } | |
756 | |
757 static inline void | |
758 clear_page_group_in_use (page_group *group, char *page) | |
759 { | |
760 group->in_use &= ~(1 << page_group_index (group->allocation, page)); | |
761 } | |
762 #endif | |
763 | |
764 /* Allocate a new page for allocating objects of size 2^ORDER, | |
765 and return an entry for it. The entry is not added to the | |
766 appropriate page_table list. */ | |
767 | |
768 static inline struct page_entry * | |
769 alloc_page (unsigned order) | |
770 { | |
771 struct page_entry *entry, *p, **pp; | |
772 char *page; | |
773 size_t num_objects; | |
774 size_t bitmap_size; | |
775 size_t page_entry_size; | |
776 size_t entry_size; | |
777 #ifdef USING_MALLOC_PAGE_GROUPS | |
778 page_group *group; | |
779 #endif | |
780 | |
781 num_objects = OBJECTS_PER_PAGE (order); | |
782 bitmap_size = BITMAP_SIZE (num_objects + 1); | |
783 page_entry_size = sizeof (page_entry) - sizeof (long) + bitmap_size; | |
784 entry_size = num_objects * OBJECT_SIZE (order); | |
785 if (entry_size < G.pagesize) | |
786 entry_size = G.pagesize; | |
111 | 787 entry_size = PAGE_ALIGN (entry_size); |
0 | 788 |
789 entry = NULL; | |
790 page = NULL; | |
791 | |
792 /* Check the list of free pages for one we can use. */ | |
793 for (pp = &G.free_pages, p = *pp; p; pp = &p->next, p = *pp) | |
794 if (p->bytes == entry_size) | |
795 break; | |
796 | |
797 if (p != NULL) | |
798 { | |
111 | 799 if (p->discarded) |
800 G.bytes_mapped += p->bytes; | |
801 p->discarded = false; | |
802 | |
0 | 803 /* Recycle the allocated memory from this page ... */ |
804 *pp = p->next; | |
805 page = p->page; | |
806 | |
807 #ifdef USING_MALLOC_PAGE_GROUPS | |
808 group = p->group; | |
809 #endif | |
810 | |
811 /* ... and, if possible, the page entry itself. */ | |
812 if (p->order == order) | |
813 { | |
814 entry = p; | |
815 memset (entry, 0, page_entry_size); | |
816 } | |
817 else | |
818 free (p); | |
819 } | |
820 #ifdef USING_MMAP | |
821 else if (entry_size == G.pagesize) | |
822 { | |
823 /* We want just one page. Allocate a bunch of them and put the | |
824 extras on the freelist. (Can only do this optimization with | |
825 mmap for backing store.) */ | |
826 struct page_entry *e, *f = G.free_pages; | |
111 | 827 int i, entries = GGC_QUIRE_SIZE; |
828 | |
829 page = alloc_anon (NULL, G.pagesize * GGC_QUIRE_SIZE, false); | |
830 if (page == NULL) | |
831 { | |
832 page = alloc_anon (NULL, G.pagesize, true); | |
833 entries = 1; | |
834 } | |
0 | 835 |
836 /* This loop counts down so that the chain will be in ascending | |
837 memory order. */ | |
111 | 838 for (i = entries - 1; i >= 1; i--) |
0 | 839 { |
840 e = XCNEWVAR (struct page_entry, page_entry_size); | |
841 e->order = order; | |
842 e->bytes = G.pagesize; | |
843 e->page = page + (i << G.lg_pagesize); | |
844 e->next = f; | |
845 f = e; | |
846 } | |
847 | |
848 G.free_pages = f; | |
849 } | |
850 else | |
111 | 851 page = alloc_anon (NULL, entry_size, true); |
0 | 852 #endif |
853 #ifdef USING_MALLOC_PAGE_GROUPS | |
854 else | |
855 { | |
856 /* Allocate a large block of memory and serve out the aligned | |
857 pages therein. This results in much less memory wastage | |
858 than the traditional implementation of valloc. */ | |
859 | |
860 char *allocation, *a, *enda; | |
861 size_t alloc_size, head_slop, tail_slop; | |
862 int multiple_pages = (entry_size == G.pagesize); | |
863 | |
864 if (multiple_pages) | |
865 alloc_size = GGC_QUIRE_SIZE * G.pagesize; | |
866 else | |
867 alloc_size = entry_size + G.pagesize - 1; | |
868 allocation = XNEWVEC (char, alloc_size); | |
869 | |
111 | 870 page = (char *) (((uintptr_t) allocation + G.pagesize - 1) & -G.pagesize); |
0 | 871 head_slop = page - allocation; |
872 if (multiple_pages) | |
873 tail_slop = ((size_t) allocation + alloc_size) & (G.pagesize - 1); | |
874 else | |
875 tail_slop = alloc_size - entry_size - head_slop; | |
876 enda = allocation + alloc_size - tail_slop; | |
877 | |
878 /* We allocated N pages, which are likely not aligned, leaving | |
879 us with N-1 usable pages. We plan to place the page_group | |
880 structure somewhere in the slop. */ | |
881 if (head_slop >= sizeof (page_group)) | |
882 group = (page_group *)page - 1; | |
883 else | |
884 { | |
885 /* We magically got an aligned allocation. Too bad, we have | |
886 to waste a page anyway. */ | |
887 if (tail_slop == 0) | |
888 { | |
889 enda -= G.pagesize; | |
890 tail_slop += G.pagesize; | |
891 } | |
892 gcc_assert (tail_slop >= sizeof (page_group)); | |
893 group = (page_group *)enda; | |
894 tail_slop -= sizeof (page_group); | |
895 } | |
896 | |
897 /* Remember that we allocated this memory. */ | |
898 group->next = G.page_groups; | |
899 group->allocation = allocation; | |
900 group->alloc_size = alloc_size; | |
901 group->in_use = 0; | |
902 G.page_groups = group; | |
903 G.bytes_mapped += alloc_size; | |
904 | |
905 /* If we allocated multiple pages, put the rest on the free list. */ | |
906 if (multiple_pages) | |
907 { | |
908 struct page_entry *e, *f = G.free_pages; | |
909 for (a = enda - G.pagesize; a != page; a -= G.pagesize) | |
910 { | |
911 e = XCNEWVAR (struct page_entry, page_entry_size); | |
912 e->order = order; | |
913 e->bytes = G.pagesize; | |
914 e->page = a; | |
915 e->group = group; | |
916 e->next = f; | |
917 f = e; | |
918 } | |
919 G.free_pages = f; | |
920 } | |
921 } | |
922 #endif | |
923 | |
924 if (entry == NULL) | |
925 entry = XCNEWVAR (struct page_entry, page_entry_size); | |
926 | |
927 entry->bytes = entry_size; | |
928 entry->page = page; | |
929 entry->context_depth = G.context_depth; | |
930 entry->order = order; | |
931 entry->num_free_objects = num_objects; | |
932 entry->next_bit_hint = 1; | |
933 | |
934 G.context_depth_allocations |= (unsigned long)1 << G.context_depth; | |
935 | |
936 #ifdef USING_MALLOC_PAGE_GROUPS | |
937 entry->group = group; | |
938 set_page_group_in_use (group, page); | |
939 #endif | |
940 | |
941 /* Set the one-past-the-end in-use bit. This acts as a sentry as we | |
942 increment the hint. */ | |
943 entry->in_use_p[num_objects / HOST_BITS_PER_LONG] | |
944 = (unsigned long) 1 << (num_objects % HOST_BITS_PER_LONG); | |
945 | |
946 set_page_table_entry (page, entry); | |
947 | |
948 if (GGC_DEBUG_LEVEL >= 2) | |
949 fprintf (G.debug_file, | |
950 "Allocating page at %p, object size=%lu, data %p-%p\n", | |
951 (void *) entry, (unsigned long) OBJECT_SIZE (order), page, | |
952 page + entry_size - 1); | |
953 | |
954 return entry; | |
955 } | |
956 | |
957 /* Adjust the size of G.depth so that no index greater than the one | |
958 used by the top of the G.by_depth is used. */ | |
959 | |
960 static inline void | |
961 adjust_depth (void) | |
962 { | |
963 page_entry *top; | |
964 | |
965 if (G.by_depth_in_use) | |
966 { | |
967 top = G.by_depth[G.by_depth_in_use-1]; | |
968 | |
969 /* Peel back indices in depth that index into by_depth, so that | |
970 as new elements are added to by_depth, we note the indices | |
971 of those elements, if they are for new context depths. */ | |
972 while (G.depth_in_use > (size_t)top->context_depth+1) | |
973 --G.depth_in_use; | |
974 } | |
975 } | |
976 | |
977 /* For a page that is no longer needed, put it on the free page list. */ | |
978 | |
979 static void | |
980 free_page (page_entry *entry) | |
981 { | |
982 if (GGC_DEBUG_LEVEL >= 2) | |
983 fprintf (G.debug_file, | |
984 "Deallocating page at %p, data %p-%p\n", (void *) entry, | |
985 entry->page, entry->page + entry->bytes - 1); | |
986 | |
987 /* Mark the page as inaccessible. Discard the handle to avoid handle | |
988 leak. */ | |
989 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (entry->page, entry->bytes)); | |
990 | |
991 set_page_table_entry (entry->page, NULL); | |
992 | |
993 #ifdef USING_MALLOC_PAGE_GROUPS | |
994 clear_page_group_in_use (entry->group, entry->page); | |
995 #endif | |
996 | |
997 if (G.by_depth_in_use > 1) | |
998 { | |
999 page_entry *top = G.by_depth[G.by_depth_in_use-1]; | |
1000 int i = entry->index_by_depth; | |
1001 | |
1002 /* We cannot free a page from a context deeper than the current | |
1003 one. */ | |
1004 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
|
1005 |
0 | 1006 /* Put top element into freed slot. */ |
1007 G.by_depth[i] = top; | |
1008 G.save_in_use[i] = G.save_in_use[G.by_depth_in_use-1]; | |
1009 top->index_by_depth = i; | |
1010 } | |
1011 --G.by_depth_in_use; | |
1012 | |
1013 adjust_depth (); | |
1014 | |
1015 entry->next = G.free_pages; | |
1016 G.free_pages = entry; | |
1017 } | |
1018 | |
1019 /* Release the free page cache to the system. */ | |
1020 | |
1021 static void | |
1022 release_pages (void) | |
1023 { | |
111 | 1024 #ifdef USING_MADVISE |
1025 page_entry *p, *start_p; | |
1026 char *start; | |
1027 size_t len; | |
1028 size_t mapped_len; | |
1029 page_entry *next, *prev, *newprev; | |
1030 size_t free_unit = (GGC_QUIRE_SIZE/2) * G.pagesize; | |
1031 | |
1032 /* First free larger continuous areas to the OS. | |
1033 This allows other allocators to grab these areas if needed. | |
1034 This is only done on larger chunks to avoid fragmentation. | |
1035 This does not always work because the free_pages list is only | |
1036 approximately sorted. */ | |
1037 | |
1038 p = G.free_pages; | |
1039 prev = NULL; | |
1040 while (p) | |
1041 { | |
1042 start = p->page; | |
1043 start_p = p; | |
1044 len = 0; | |
1045 mapped_len = 0; | |
1046 newprev = prev; | |
1047 while (p && p->page == start + len) | |
1048 { | |
1049 len += p->bytes; | |
1050 if (!p->discarded) | |
1051 mapped_len += p->bytes; | |
1052 newprev = p; | |
1053 p = p->next; | |
1054 } | |
1055 if (len >= free_unit) | |
1056 { | |
1057 while (start_p != p) | |
1058 { | |
1059 next = start_p->next; | |
1060 free (start_p); | |
1061 start_p = next; | |
1062 } | |
1063 munmap (start, len); | |
1064 if (prev) | |
1065 prev->next = p; | |
1066 else | |
1067 G.free_pages = p; | |
1068 G.bytes_mapped -= mapped_len; | |
1069 continue; | |
1070 } | |
1071 prev = newprev; | |
1072 } | |
1073 | |
1074 /* Now give back the fragmented pages to the OS, but keep the address | |
1075 space to reuse it next time. */ | |
1076 | |
1077 for (p = G.free_pages; p; ) | |
1078 { | |
1079 if (p->discarded) | |
1080 { | |
1081 p = p->next; | |
1082 continue; | |
1083 } | |
1084 start = p->page; | |
1085 len = p->bytes; | |
1086 start_p = p; | |
1087 p = p->next; | |
1088 while (p && p->page == start + len) | |
1089 { | |
1090 len += p->bytes; | |
1091 p = p->next; | |
1092 } | |
1093 /* Give the page back to the kernel, but don't free the mapping. | |
1094 This avoids fragmentation in the virtual memory map of the | |
1095 process. Next time we can reuse it by just touching it. */ | |
1096 madvise (start, len, MADV_DONTNEED); | |
1097 /* Don't count those pages as mapped to not touch the garbage collector | |
1098 unnecessarily. */ | |
1099 G.bytes_mapped -= len; | |
1100 while (start_p != p) | |
1101 { | |
1102 start_p->discarded = true; | |
1103 start_p = start_p->next; | |
1104 } | |
1105 } | |
1106 #endif | |
1107 #if defined(USING_MMAP) && !defined(USING_MADVISE) | |
0 | 1108 page_entry *p, *next; |
1109 char *start; | |
1110 size_t len; | |
1111 | |
1112 /* Gather up adjacent pages so they are unmapped together. */ | |
1113 p = G.free_pages; | |
1114 | |
1115 while (p) | |
1116 { | |
1117 start = p->page; | |
1118 next = p->next; | |
1119 len = p->bytes; | |
1120 free (p); | |
1121 p = next; | |
1122 | |
1123 while (p && p->page == start + len) | |
1124 { | |
1125 next = p->next; | |
1126 len += p->bytes; | |
1127 free (p); | |
1128 p = next; | |
1129 } | |
1130 | |
1131 munmap (start, len); | |
1132 G.bytes_mapped -= len; | |
1133 } | |
1134 | |
1135 G.free_pages = NULL; | |
1136 #endif | |
1137 #ifdef USING_MALLOC_PAGE_GROUPS | |
1138 page_entry **pp, *p; | |
1139 page_group **gp, *g; | |
1140 | |
1141 /* Remove all pages from free page groups from the list. */ | |
1142 pp = &G.free_pages; | |
1143 while ((p = *pp) != NULL) | |
1144 if (p->group->in_use == 0) | |
1145 { | |
1146 *pp = p->next; | |
1147 free (p); | |
1148 } | |
1149 else | |
1150 pp = &p->next; | |
1151 | |
1152 /* Remove all free page groups, and release the storage. */ | |
1153 gp = &G.page_groups; | |
1154 while ((g = *gp) != NULL) | |
1155 if (g->in_use == 0) | |
1156 { | |
1157 *gp = g->next; | |
1158 G.bytes_mapped -= g->alloc_size; | |
1159 free (g->allocation); | |
1160 } | |
1161 else | |
1162 gp = &g->next; | |
1163 #endif | |
1164 } | |
1165 | |
1166 /* This table provides a fast way to determine ceil(log_2(size)) for | |
1167 allocation requests. The minimum allocation size is eight bytes. */ | |
1168 #define NUM_SIZE_LOOKUP 512 | |
1169 static unsigned char size_lookup[NUM_SIZE_LOOKUP] = | |
1170 { | |
1171 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, | |
1172 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, | |
1173 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, | |
1174 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, | |
1175 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1176 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1177 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1178 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, | |
1179 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1180 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1181 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1182 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1183 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1184 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1185 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1186 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, | |
1187 8, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1188 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1189 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1190 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1191 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1192 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1193 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1194 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1195 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1196 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1197 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1198 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1199 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1200 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1201 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, | |
1202 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9 | |
1203 }; | |
1204 | |
111 | 1205 /* For a given size of memory requested for allocation, return the |
1206 actual size that is going to be allocated, as well as the size | |
1207 order. */ | |
1208 | |
1209 static void | |
1210 ggc_round_alloc_size_1 (size_t requested_size, | |
1211 size_t *size_order, | |
1212 size_t *alloced_size) | |
0 | 1213 { |
111 | 1214 size_t order, object_size; |
1215 | |
1216 if (requested_size < NUM_SIZE_LOOKUP) | |
1217 { | |
1218 order = size_lookup[requested_size]; | |
1219 object_size = OBJECT_SIZE (order); | |
1220 } | |
1221 else | |
1222 { | |
1223 order = 10; | |
1224 while (requested_size > (object_size = OBJECT_SIZE (order))) | |
1225 order++; | |
1226 } | |
1227 | |
1228 if (size_order) | |
1229 *size_order = order; | |
1230 if (alloced_size) | |
1231 *alloced_size = object_size; | |
1232 } | |
1233 | |
1234 /* For a given size of memory requested for allocation, return the | |
1235 actual size that is going to be allocated. */ | |
1236 | |
1237 size_t | |
1238 ggc_round_alloc_size (size_t requested_size) | |
1239 { | |
1240 size_t size = 0; | |
1241 | |
1242 ggc_round_alloc_size_1 (requested_size, NULL, &size); | |
1243 return size; | |
1244 } | |
1245 | |
1246 /* Push a finalizer onto the appropriate vec. */ | |
1247 | |
1248 static void | |
1249 add_finalizer (void *result, void (*f)(void *), size_t s, size_t n) | |
1250 { | |
1251 if (f == NULL) | |
1252 /* No finalizer. */; | |
1253 else if (n == 1) | |
1254 { | |
1255 finalizer fin (result, f); | |
1256 G.finalizers[G.context_depth].safe_push (fin); | |
1257 } | |
1258 else | |
1259 { | |
1260 vec_finalizer fin (reinterpret_cast<uintptr_t> (result), f, s, n); | |
1261 G.vec_finalizers[G.context_depth].safe_push (fin); | |
1262 } | |
0 | 1263 } |
1264 | |
1265 /* Allocate a chunk of memory of SIZE bytes. Its contents are undefined. */ | |
1266 | |
1267 void * | |
111 | 1268 ggc_internal_alloc (size_t size, void (*f)(void *), size_t s, size_t n |
1269 MEM_STAT_DECL) | |
0 | 1270 { |
1271 size_t order, word, bit, object_offset, object_size; | |
1272 struct page_entry *entry; | |
1273 void *result; | |
1274 | |
111 | 1275 ggc_round_alloc_size_1 (size, &order, &object_size); |
0 | 1276 |
1277 /* If there are non-full pages for this size allocation, they are at | |
1278 the head of the list. */ | |
1279 entry = G.pages[order]; | |
1280 | |
1281 /* If there is no page for this object size, or all pages in this | |
1282 context are full, allocate a new page. */ | |
1283 if (entry == NULL || entry->num_free_objects == 0) | |
1284 { | |
1285 struct page_entry *new_entry; | |
1286 new_entry = alloc_page (order); | |
1287 | |
1288 new_entry->index_by_depth = G.by_depth_in_use; | |
1289 push_by_depth (new_entry, 0); | |
1290 | |
1291 /* We can skip context depths, if we do, make sure we go all the | |
1292 way to the new depth. */ | |
1293 while (new_entry->context_depth >= G.depth_in_use) | |
1294 push_depth (G.by_depth_in_use-1); | |
1295 | |
1296 /* If this is the only entry, it's also the tail. If it is not | |
1297 the only entry, then we must update the PREV pointer of the | |
1298 ENTRY (G.pages[order]) to point to our new page entry. */ | |
1299 if (entry == NULL) | |
1300 G.page_tails[order] = new_entry; | |
1301 else | |
1302 entry->prev = new_entry; | |
1303 | |
1304 /* Put new pages at the head of the page list. By definition the | |
1305 entry at the head of the list always has a NULL pointer. */ | |
1306 new_entry->next = entry; | |
1307 new_entry->prev = NULL; | |
1308 entry = new_entry; | |
1309 G.pages[order] = new_entry; | |
1310 | |
1311 /* For a new page, we know the word and bit positions (in the | |
1312 in_use bitmap) of the first available object -- they're zero. */ | |
1313 new_entry->next_bit_hint = 1; | |
1314 word = 0; | |
1315 bit = 0; | |
1316 object_offset = 0; | |
1317 } | |
1318 else | |
1319 { | |
1320 /* First try to use the hint left from the previous allocation | |
1321 to locate a clear bit in the in-use bitmap. We've made sure | |
1322 that the one-past-the-end bit is always set, so if the hint | |
1323 has run over, this test will fail. */ | |
1324 unsigned hint = entry->next_bit_hint; | |
1325 word = hint / HOST_BITS_PER_LONG; | |
1326 bit = hint % HOST_BITS_PER_LONG; | |
1327 | |
1328 /* If the hint didn't work, scan the bitmap from the beginning. */ | |
1329 if ((entry->in_use_p[word] >> bit) & 1) | |
1330 { | |
1331 word = bit = 0; | |
1332 while (~entry->in_use_p[word] == 0) | |
1333 ++word; | |
1334 | |
1335 #if GCC_VERSION >= 3004 | |
1336 bit = __builtin_ctzl (~entry->in_use_p[word]); | |
1337 #else | |
1338 while ((entry->in_use_p[word] >> bit) & 1) | |
1339 ++bit; | |
1340 #endif | |
1341 | |
1342 hint = word * HOST_BITS_PER_LONG + bit; | |
1343 } | |
1344 | |
1345 /* Next time, try the next bit. */ | |
1346 entry->next_bit_hint = hint + 1; | |
1347 | |
1348 object_offset = hint * object_size; | |
1349 } | |
1350 | |
1351 /* Set the in-use bit. */ | |
1352 entry->in_use_p[word] |= ((unsigned long) 1 << bit); | |
1353 | |
1354 /* Keep a running total of the number of free objects. If this page | |
1355 fills up, we may have to move it to the end of the list if the | |
1356 next page isn't full. If the next page is full, all subsequent | |
1357 pages are full, so there's no need to move it. */ | |
1358 if (--entry->num_free_objects == 0 | |
1359 && entry->next != NULL | |
1360 && entry->next->num_free_objects > 0) | |
1361 { | |
1362 /* We have a new head for the list. */ | |
1363 G.pages[order] = entry->next; | |
1364 | |
1365 /* We are moving ENTRY to the end of the page table list. | |
1366 The new page at the head of the list will have NULL in | |
1367 its PREV field and ENTRY will have NULL in its NEXT field. */ | |
1368 entry->next->prev = NULL; | |
1369 entry->next = NULL; | |
1370 | |
1371 /* Append ENTRY to the tail of the list. */ | |
1372 entry->prev = G.page_tails[order]; | |
1373 G.page_tails[order]->next = entry; | |
1374 G.page_tails[order] = entry; | |
1375 } | |
1376 | |
1377 /* Calculate the object's address. */ | |
1378 result = entry->page + object_offset; | |
111 | 1379 if (GATHER_STATISTICS) |
1380 ggc_record_overhead (OBJECT_SIZE (order), OBJECT_SIZE (order) - size, | |
1381 result FINAL_PASS_MEM_STAT); | |
0 | 1382 |
1383 #ifdef ENABLE_GC_CHECKING | |
1384 /* Keep poisoning-by-writing-0xaf the object, in an attempt to keep the | |
1385 exact same semantics in presence of memory bugs, regardless of | |
1386 ENABLE_VALGRIND_CHECKING. We override this request below. Drop the | |
1387 handle to avoid handle leak. */ | |
1388 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (result, object_size)); | |
1389 | |
1390 /* `Poison' the entire allocated object, including any padding at | |
1391 the end. */ | |
1392 memset (result, 0xaf, object_size); | |
1393 | |
1394 /* Make the bytes after the end of the object unaccessible. Discard the | |
1395 handle to avoid handle leak. */ | |
1396 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS ((char *) result + size, | |
1397 object_size - size)); | |
1398 #endif | |
1399 | |
1400 /* Tell Valgrind that the memory is there, but its content isn't | |
1401 defined. The bytes at the end of the object are still marked | |
1402 unaccessible. */ | |
1403 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (result, size)); | |
1404 | |
1405 /* Keep track of how many bytes are being allocated. This | |
1406 information is used in deciding when to collect. */ | |
1407 G.allocated += object_size; | |
1408 | |
1409 /* For timevar statistics. */ | |
1410 timevar_ggc_mem_total += object_size; | |
1411 | |
111 | 1412 if (f) |
1413 add_finalizer (result, f, s, n); | |
1414 | |
1415 if (GATHER_STATISTICS) | |
1416 { | |
1417 size_t overhead = object_size - size; | |
1418 | |
1419 G.stats.total_overhead += overhead; | |
1420 G.stats.total_allocated += object_size; | |
1421 G.stats.total_overhead_per_order[order] += overhead; | |
1422 G.stats.total_allocated_per_order[order] += object_size; | |
1423 | |
1424 if (size <= 32) | |
1425 { | |
1426 G.stats.total_overhead_under32 += overhead; | |
1427 G.stats.total_allocated_under32 += object_size; | |
1428 } | |
1429 if (size <= 64) | |
1430 { | |
1431 G.stats.total_overhead_under64 += overhead; | |
1432 G.stats.total_allocated_under64 += object_size; | |
1433 } | |
1434 if (size <= 128) | |
1435 { | |
1436 G.stats.total_overhead_under128 += overhead; | |
1437 G.stats.total_allocated_under128 += object_size; | |
1438 } | |
1439 } | |
0 | 1440 |
1441 if (GGC_DEBUG_LEVEL >= 3) | |
1442 fprintf (G.debug_file, | |
1443 "Allocating object, requested size=%lu, actual=%lu at %p on %p\n", | |
1444 (unsigned long) size, (unsigned long) object_size, result, | |
1445 (void *) entry); | |
1446 | |
1447 return result; | |
1448 } | |
1449 | |
1450 /* Mark function for strings. */ | |
1451 | |
1452 void | |
1453 gt_ggc_m_S (const void *p) | |
1454 { | |
1455 page_entry *entry; | |
1456 unsigned bit, word; | |
1457 unsigned long mask; | |
1458 unsigned long offset; | |
1459 | |
111 | 1460 if (!p) |
0 | 1461 return; |
1462 | |
111 | 1463 /* Look up the page on which the object is alloced. If it was not |
1464 GC allocated, gracefully bail out. */ | |
1465 entry = safe_lookup_page_table_entry (p); | |
1466 if (!entry) | |
1467 return; | |
0 | 1468 |
1469 /* Calculate the index of the object on the page; this is its bit | |
1470 position in the in_use_p bitmap. Note that because a char* might | |
1471 point to the middle of an object, we need special code here to | |
1472 make sure P points to the start of an object. */ | |
1473 offset = ((const char *) p - entry->page) % object_size_table[entry->order]; | |
1474 if (offset) | |
1475 { | |
1476 /* Here we've seen a char* which does not point to the beginning | |
1477 of an allocated object. We assume it points to the middle of | |
1478 a STRING_CST. */ | |
1479 gcc_assert (offset == offsetof (struct tree_string, str)); | |
1480 p = ((const char *) p) - offset; | |
1481 gt_ggc_mx_lang_tree_node (CONST_CAST (void *, p)); | |
1482 return; | |
1483 } | |
1484 | |
1485 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order); | |
1486 word = bit / HOST_BITS_PER_LONG; | |
1487 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG); | |
1488 | |
1489 /* If the bit was previously set, skip it. */ | |
1490 if (entry->in_use_p[word] & mask) | |
1491 return; | |
1492 | |
1493 /* Otherwise set it, and decrement the free object count. */ | |
1494 entry->in_use_p[word] |= mask; | |
1495 entry->num_free_objects -= 1; | |
1496 | |
1497 if (GGC_DEBUG_LEVEL >= 4) | |
1498 fprintf (G.debug_file, "Marking %p\n", p); | |
1499 | |
1500 return; | |
1501 } | |
1502 | |
111 | 1503 |
1504 /* User-callable entry points for marking string X. */ | |
1505 | |
1506 void | |
1507 gt_ggc_mx (const char *& x) | |
1508 { | |
1509 gt_ggc_m_S (x); | |
1510 } | |
1511 | |
1512 void | |
1513 gt_ggc_mx (unsigned char *& x) | |
1514 { | |
1515 gt_ggc_m_S (x); | |
1516 } | |
1517 | |
1518 void | |
1519 gt_ggc_mx (unsigned char& x ATTRIBUTE_UNUSED) | |
1520 { | |
1521 } | |
1522 | |
0 | 1523 /* If P is not marked, marks it and return false. Otherwise return true. |
1524 P must have been allocated by the GC allocator; it mustn't point to | |
1525 static objects, stack variables, or memory allocated with malloc. */ | |
1526 | |
1527 int | |
1528 ggc_set_mark (const void *p) | |
1529 { | |
1530 page_entry *entry; | |
1531 unsigned bit, word; | |
1532 unsigned long mask; | |
1533 | |
1534 /* Look up the page on which the object is alloced. If the object | |
1535 wasn't allocated by the collector, we'll probably die. */ | |
1536 entry = lookup_page_table_entry (p); | |
1537 gcc_assert (entry); | |
1538 | |
1539 /* Calculate the index of the object on the page; this is its bit | |
1540 position in the in_use_p bitmap. */ | |
1541 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order); | |
1542 word = bit / HOST_BITS_PER_LONG; | |
1543 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG); | |
1544 | |
1545 /* If the bit was previously set, skip it. */ | |
1546 if (entry->in_use_p[word] & mask) | |
1547 return 1; | |
1548 | |
1549 /* Otherwise set it, and decrement the free object count. */ | |
1550 entry->in_use_p[word] |= mask; | |
1551 entry->num_free_objects -= 1; | |
1552 | |
1553 if (GGC_DEBUG_LEVEL >= 4) | |
1554 fprintf (G.debug_file, "Marking %p\n", p); | |
1555 | |
1556 return 0; | |
1557 } | |
1558 | |
1559 /* Return 1 if P has been marked, zero otherwise. | |
1560 P must have been allocated by the GC allocator; it mustn't point to | |
1561 static objects, stack variables, or memory allocated with malloc. */ | |
1562 | |
1563 int | |
1564 ggc_marked_p (const void *p) | |
1565 { | |
1566 page_entry *entry; | |
1567 unsigned bit, word; | |
1568 unsigned long mask; | |
1569 | |
1570 /* Look up the page on which the object is alloced. If the object | |
1571 wasn't allocated by the collector, we'll probably die. */ | |
1572 entry = lookup_page_table_entry (p); | |
1573 gcc_assert (entry); | |
1574 | |
1575 /* Calculate the index of the object on the page; this is its bit | |
1576 position in the in_use_p bitmap. */ | |
1577 bit = OFFSET_TO_BIT (((const char *) p) - entry->page, entry->order); | |
1578 word = bit / HOST_BITS_PER_LONG; | |
1579 mask = (unsigned long) 1 << (bit % HOST_BITS_PER_LONG); | |
1580 | |
1581 return (entry->in_use_p[word] & mask) != 0; | |
1582 } | |
1583 | |
1584 /* Return the size of the gc-able object P. */ | |
1585 | |
1586 size_t | |
1587 ggc_get_size (const void *p) | |
1588 { | |
1589 page_entry *pe = lookup_page_table_entry (p); | |
1590 return OBJECT_SIZE (pe->order); | |
1591 } | |
1592 | |
1593 /* Release the memory for object P. */ | |
1594 | |
1595 void | |
1596 ggc_free (void *p) | |
1597 { | |
111 | 1598 if (in_gc) |
1599 return; | |
1600 | |
0 | 1601 page_entry *pe = lookup_page_table_entry (p); |
1602 size_t order = pe->order; | |
1603 size_t size = OBJECT_SIZE (order); | |
1604 | |
111 | 1605 if (GATHER_STATISTICS) |
1606 ggc_free_overhead (p); | |
0 | 1607 |
1608 if (GGC_DEBUG_LEVEL >= 3) | |
1609 fprintf (G.debug_file, | |
1610 "Freeing object, actual size=%lu, at %p on %p\n", | |
1611 (unsigned long) size, p, (void *) pe); | |
1612 | |
1613 #ifdef ENABLE_GC_CHECKING | |
1614 /* Poison the data, to indicate the data is garbage. */ | |
1615 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (p, size)); | |
1616 memset (p, 0xa5, size); | |
1617 #endif | |
1618 /* Let valgrind know the object is free. */ | |
1619 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (p, size)); | |
1620 | |
1621 #ifdef ENABLE_GC_ALWAYS_COLLECT | |
1622 /* 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
|
1623 the data, but instead verify that the data is *actually* not |
0 | 1624 reachable the next time we collect. */ |
1625 { | |
1626 struct free_object *fo = XNEW (struct free_object); | |
1627 fo->object = p; | |
1628 fo->next = G.free_object_list; | |
1629 G.free_object_list = fo; | |
1630 } | |
1631 #else | |
1632 { | |
1633 unsigned int bit_offset, word, bit; | |
1634 | |
1635 G.allocated -= size; | |
1636 | |
1637 /* Mark the object not-in-use. */ | |
1638 bit_offset = OFFSET_TO_BIT (((const char *) p) - pe->page, order); | |
1639 word = bit_offset / HOST_BITS_PER_LONG; | |
1640 bit = bit_offset % HOST_BITS_PER_LONG; | |
1641 pe->in_use_p[word] &= ~(1UL << bit); | |
1642 | |
1643 if (pe->num_free_objects++ == 0) | |
1644 { | |
1645 page_entry *p, *q; | |
1646 | |
1647 /* If the page is completely full, then it's supposed to | |
1648 be after all pages that aren't. Since we've freed one | |
1649 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
|
1650 page to the head of the list. |
0 | 1651 |
1652 PE is the node we want to move. Q is the previous node | |
1653 and P is the next node in the list. */ | |
1654 q = pe->prev; | |
1655 if (q && q->num_free_objects == 0) | |
1656 { | |
1657 p = pe->next; | |
1658 | |
1659 q->next = p; | |
1660 | |
1661 /* If PE was at the end of the list, then Q becomes the | |
1662 new end of the list. If PE was not the end of the | |
1663 list, then we need to update the PREV field for P. */ | |
1664 if (!p) | |
1665 G.page_tails[order] = q; | |
1666 else | |
1667 p->prev = q; | |
1668 | |
1669 /* Move PE to the head of the list. */ | |
1670 pe->next = G.pages[order]; | |
1671 pe->prev = NULL; | |
1672 G.pages[order]->prev = pe; | |
1673 G.pages[order] = pe; | |
1674 } | |
1675 | |
1676 /* Reset the hint bit to point to the only free object. */ | |
1677 pe->next_bit_hint = bit_offset; | |
1678 } | |
1679 } | |
1680 #endif | |
1681 } | |
1682 | |
1683 /* Subroutine of init_ggc which computes the pair of numbers used to | |
1684 perform division by OBJECT_SIZE (order) and fills in inverse_table[]. | |
1685 | |
1686 This algorithm is taken from Granlund and Montgomery's paper | |
1687 "Division by Invariant Integers using Multiplication" | |
1688 (Proc. SIGPLAN PLDI, 1994), section 9 (Exact division by | |
1689 constants). */ | |
1690 | |
1691 static void | |
1692 compute_inverse (unsigned order) | |
1693 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1694 size_t size, inv; |
0 | 1695 unsigned int e; |
1696 | |
1697 size = OBJECT_SIZE (order); | |
1698 e = 0; | |
1699 while (size % 2 == 0) | |
1700 { | |
1701 e++; | |
1702 size >>= 1; | |
1703 } | |
1704 | |
1705 inv = size; | |
1706 while (inv * size != 1) | |
1707 inv = inv * (2 - inv*size); | |
1708 | |
1709 DIV_MULT (order) = inv; | |
1710 DIV_SHIFT (order) = e; | |
1711 } | |
1712 | |
1713 /* Initialize the ggc-mmap allocator. */ | |
1714 void | |
1715 init_ggc (void) | |
1716 { | |
111 | 1717 static bool init_p = false; |
0 | 1718 unsigned order; |
1719 | |
111 | 1720 if (init_p) |
1721 return; | |
1722 init_p = true; | |
1723 | |
1724 G.pagesize = getpagesize (); | |
0 | 1725 G.lg_pagesize = exact_log2 (G.pagesize); |
1726 | |
1727 #ifdef HAVE_MMAP_DEV_ZERO | |
1728 G.dev_zero_fd = open ("/dev/zero", O_RDONLY); | |
1729 if (G.dev_zero_fd == -1) | |
1730 internal_error ("open /dev/zero: %m"); | |
1731 #endif | |
1732 | |
1733 #if 0 | |
1734 G.debug_file = fopen ("ggc-mmap.debug", "w"); | |
1735 #else | |
1736 G.debug_file = stdout; | |
1737 #endif | |
1738 | |
1739 #ifdef USING_MMAP | |
1740 /* StunOS has an amazing off-by-one error for the first mmap allocation | |
1741 after fiddling with RLIMIT_STACK. The result, as hard as it is to | |
1742 believe, is an unaligned page allocation, which would cause us to | |
1743 hork badly if we tried to use it. */ | |
1744 { | |
111 | 1745 char *p = alloc_anon (NULL, G.pagesize, true); |
0 | 1746 struct page_entry *e; |
111 | 1747 if ((uintptr_t)p & (G.pagesize - 1)) |
0 | 1748 { |
1749 /* How losing. Discard this one and try another. If we still | |
1750 can't get something useful, give up. */ | |
1751 | |
111 | 1752 p = alloc_anon (NULL, G.pagesize, true); |
1753 gcc_assert (!((uintptr_t)p & (G.pagesize - 1))); | |
0 | 1754 } |
1755 | |
1756 /* We have a good page, might as well hold onto it... */ | |
1757 e = XCNEW (struct page_entry); | |
1758 e->bytes = G.pagesize; | |
1759 e->page = p; | |
1760 e->next = G.free_pages; | |
1761 G.free_pages = e; | |
1762 } | |
1763 #endif | |
1764 | |
1765 /* Initialize the object size table. */ | |
1766 for (order = 0; order < HOST_BITS_PER_PTR; ++order) | |
1767 object_size_table[order] = (size_t) 1 << order; | |
1768 for (order = HOST_BITS_PER_PTR; order < NUM_ORDERS; ++order) | |
1769 { | |
1770 size_t s = extra_order_size_table[order - HOST_BITS_PER_PTR]; | |
1771 | |
1772 /* If S is not a multiple of the MAX_ALIGNMENT, then round it up | |
1773 so that we're sure of getting aligned memory. */ | |
1774 s = ROUND_UP (s, MAX_ALIGNMENT); | |
1775 object_size_table[order] = s; | |
1776 } | |
1777 | |
1778 /* Initialize the objects-per-page and inverse tables. */ | |
1779 for (order = 0; order < NUM_ORDERS; ++order) | |
1780 { | |
1781 objects_per_page_table[order] = G.pagesize / OBJECT_SIZE (order); | |
1782 if (objects_per_page_table[order] == 0) | |
1783 objects_per_page_table[order] = 1; | |
1784 compute_inverse (order); | |
1785 } | |
1786 | |
1787 /* Reset the size_lookup array to put appropriately sized objects in | |
1788 the special orders. All objects bigger than the previous power | |
1789 of two, but no greater than the special size, should go in the | |
1790 new order. */ | |
1791 for (order = HOST_BITS_PER_PTR; order < NUM_ORDERS; ++order) | |
1792 { | |
1793 int o; | |
1794 int i; | |
1795 | |
1796 i = OBJECT_SIZE (order); | |
1797 if (i >= NUM_SIZE_LOOKUP) | |
1798 continue; | |
1799 | |
1800 for (o = size_lookup[i]; o == size_lookup [i]; --i) | |
1801 size_lookup[i] = order; | |
1802 } | |
1803 | |
1804 G.depth_in_use = 0; | |
1805 G.depth_max = 10; | |
1806 G.depth = XNEWVEC (unsigned int, G.depth_max); | |
1807 | |
1808 G.by_depth_in_use = 0; | |
1809 G.by_depth_max = INITIAL_PTE_COUNT; | |
1810 G.by_depth = XNEWVEC (page_entry *, G.by_depth_max); | |
1811 G.save_in_use = XNEWVEC (unsigned long *, G.by_depth_max); | |
111 | 1812 |
1813 /* Allocate space for the depth 0 finalizers. */ | |
1814 G.finalizers.safe_push (vNULL); | |
1815 G.vec_finalizers.safe_push (vNULL); | |
1816 gcc_assert (G.finalizers.length() == 1); | |
0 | 1817 } |
1818 | |
1819 /* Merge the SAVE_IN_USE_P and IN_USE_P arrays in P so that IN_USE_P | |
1820 reflects reality. Recalculate NUM_FREE_OBJECTS as well. */ | |
1821 | |
1822 static void | |
1823 ggc_recalculate_in_use_p (page_entry *p) | |
1824 { | |
1825 unsigned int i; | |
1826 size_t num_objects; | |
1827 | |
1828 /* Because the past-the-end bit in in_use_p is always set, we | |
1829 pretend there is one additional object. */ | |
1830 num_objects = OBJECTS_IN_PAGE (p) + 1; | |
1831 | |
1832 /* Reset the free object count. */ | |
1833 p->num_free_objects = num_objects; | |
1834 | |
1835 /* Combine the IN_USE_P and SAVE_IN_USE_P arrays. */ | |
1836 for (i = 0; | |
1837 i < CEIL (BITMAP_SIZE (num_objects), | |
1838 sizeof (*p->in_use_p)); | |
1839 ++i) | |
1840 { | |
1841 unsigned long j; | |
1842 | |
1843 /* Something is in use if it is marked, or if it was in use in a | |
1844 context further down the context stack. */ | |
1845 p->in_use_p[i] |= save_in_use_p (p)[i]; | |
1846 | |
1847 /* Decrement the free object count for every object allocated. */ | |
1848 for (j = p->in_use_p[i]; j; j >>= 1) | |
1849 p->num_free_objects -= (j & 1); | |
1850 } | |
1851 | |
1852 gcc_assert (p->num_free_objects < num_objects); | |
1853 } | |
1854 | |
1855 /* Unmark all objects. */ | |
1856 | |
1857 static void | |
1858 clear_marks (void) | |
1859 { | |
1860 unsigned order; | |
1861 | |
1862 for (order = 2; order < NUM_ORDERS; order++) | |
1863 { | |
1864 page_entry *p; | |
1865 | |
1866 for (p = G.pages[order]; p != NULL; p = p->next) | |
1867 { | |
1868 size_t num_objects = OBJECTS_IN_PAGE (p); | |
1869 size_t bitmap_size = BITMAP_SIZE (num_objects + 1); | |
1870 | |
1871 /* The data should be page-aligned. */ | |
111 | 1872 gcc_assert (!((uintptr_t) p->page & (G.pagesize - 1))); |
0 | 1873 |
1874 /* Pages that aren't in the topmost context are not collected; | |
1875 nevertheless, we need their in-use bit vectors to store GC | |
1876 marks. So, back them up first. */ | |
1877 if (p->context_depth < G.context_depth) | |
1878 { | |
1879 if (! save_in_use_p (p)) | |
1880 save_in_use_p (p) = XNEWVAR (unsigned long, bitmap_size); | |
1881 memcpy (save_in_use_p (p), p->in_use_p, bitmap_size); | |
1882 } | |
1883 | |
1884 /* Reset reset the number of free objects and clear the | |
1885 in-use bits. These will be adjusted by mark_obj. */ | |
1886 p->num_free_objects = num_objects; | |
1887 memset (p->in_use_p, 0, bitmap_size); | |
1888 | |
1889 /* Make sure the one-past-the-end bit is always set. */ | |
1890 p->in_use_p[num_objects / HOST_BITS_PER_LONG] | |
1891 = ((unsigned long) 1 << (num_objects % HOST_BITS_PER_LONG)); | |
1892 } | |
1893 } | |
1894 } | |
1895 | |
111 | 1896 /* Check if any blocks with a registered finalizer have become unmarked. If so |
1897 run the finalizer and unregister it because the block is about to be freed. | |
1898 Note that no garantee is made about what order finalizers will run in so | |
1899 touching other objects in gc memory is extremely unwise. */ | |
1900 | |
1901 static void | |
1902 ggc_handle_finalizers () | |
1903 { | |
1904 unsigned dlen = G.finalizers.length(); | |
1905 for (unsigned d = G.context_depth; d < dlen; ++d) | |
1906 { | |
1907 vec<finalizer> &v = G.finalizers[d]; | |
1908 unsigned length = v.length (); | |
1909 for (unsigned int i = 0; i < length;) | |
1910 { | |
1911 finalizer &f = v[i]; | |
1912 if (!ggc_marked_p (f.addr ())) | |
1913 { | |
1914 f.call (); | |
1915 v.unordered_remove (i); | |
1916 length--; | |
1917 } | |
1918 else | |
1919 i++; | |
1920 } | |
1921 } | |
1922 | |
1923 gcc_assert (dlen == G.vec_finalizers.length()); | |
1924 for (unsigned d = G.context_depth; d < dlen; ++d) | |
1925 { | |
1926 vec<vec_finalizer> &vv = G.vec_finalizers[d]; | |
1927 unsigned length = vv.length (); | |
1928 for (unsigned int i = 0; i < length;) | |
1929 { | |
1930 vec_finalizer &f = vv[i]; | |
1931 if (!ggc_marked_p (f.addr ())) | |
1932 { | |
1933 f.call (); | |
1934 vv.unordered_remove (i); | |
1935 length--; | |
1936 } | |
1937 else | |
1938 i++; | |
1939 } | |
1940 } | |
1941 } | |
1942 | |
0 | 1943 /* Free all empty pages. Partially empty pages need no attention |
1944 because the `mark' bit doubles as an `unused' bit. */ | |
1945 | |
1946 static void | |
1947 sweep_pages (void) | |
1948 { | |
1949 unsigned order; | |
1950 | |
1951 for (order = 2; order < NUM_ORDERS; order++) | |
1952 { | |
1953 /* The last page-entry to consider, regardless of entries | |
1954 placed at the end of the list. */ | |
1955 page_entry * const last = G.page_tails[order]; | |
1956 | |
1957 size_t num_objects; | |
1958 size_t live_objects; | |
1959 page_entry *p, *previous; | |
1960 int done; | |
1961 | |
1962 p = G.pages[order]; | |
1963 if (p == NULL) | |
1964 continue; | |
1965 | |
1966 previous = NULL; | |
1967 do | |
1968 { | |
1969 page_entry *next = p->next; | |
1970 | |
1971 /* Loop until all entries have been examined. */ | |
1972 done = (p == last); | |
1973 | |
1974 num_objects = OBJECTS_IN_PAGE (p); | |
1975 | |
1976 /* Add all live objects on this page to the count of | |
1977 allocated memory. */ | |
1978 live_objects = num_objects - p->num_free_objects; | |
1979 | |
1980 G.allocated += OBJECT_SIZE (order) * live_objects; | |
1981 | |
1982 /* Only objects on pages in the topmost context should get | |
1983 collected. */ | |
1984 if (p->context_depth < G.context_depth) | |
1985 ; | |
1986 | |
1987 /* Remove the page if it's empty. */ | |
1988 else if (live_objects == 0) | |
1989 { | |
1990 /* If P was the first page in the list, then NEXT | |
1991 becomes the new first page in the list, otherwise | |
1992 splice P out of the forward pointers. */ | |
1993 if (! previous) | |
1994 G.pages[order] = next; | |
1995 else | |
1996 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
|
1997 |
0 | 1998 /* Splice P out of the back pointers too. */ |
1999 if (next) | |
2000 next->prev = previous; | |
2001 | |
2002 /* Are we removing the last element? */ | |
2003 if (p == G.page_tails[order]) | |
2004 G.page_tails[order] = previous; | |
2005 free_page (p); | |
2006 p = previous; | |
2007 } | |
2008 | |
2009 /* If the page is full, move it to the end. */ | |
2010 else if (p->num_free_objects == 0) | |
2011 { | |
2012 /* Don't move it if it's already at the end. */ | |
2013 if (p != G.page_tails[order]) | |
2014 { | |
2015 /* Move p to the end of the list. */ | |
2016 p->next = NULL; | |
2017 p->prev = G.page_tails[order]; | |
2018 G.page_tails[order]->next = p; | |
2019 | |
2020 /* Update the tail pointer... */ | |
2021 G.page_tails[order] = p; | |
2022 | |
2023 /* ... and the head pointer, if necessary. */ | |
2024 if (! previous) | |
2025 G.pages[order] = next; | |
2026 else | |
2027 previous->next = next; | |
2028 | |
2029 /* And update the backpointer in NEXT if necessary. */ | |
2030 if (next) | |
2031 next->prev = previous; | |
2032 | |
2033 p = previous; | |
2034 } | |
2035 } | |
2036 | |
2037 /* If we've fallen through to here, it's a page in the | |
2038 topmost context that is neither full nor empty. Such a | |
2039 page must precede pages at lesser context depth in the | |
2040 list, so move it to the head. */ | |
2041 else if (p != G.pages[order]) | |
2042 { | |
2043 previous->next = p->next; | |
2044 | |
2045 /* Update the backchain in the next node if it exists. */ | |
2046 if (p->next) | |
2047 p->next->prev = previous; | |
2048 | |
2049 /* Move P to the head of the list. */ | |
2050 p->next = G.pages[order]; | |
2051 p->prev = NULL; | |
2052 G.pages[order]->prev = p; | |
2053 | |
2054 /* Update the head pointer. */ | |
2055 G.pages[order] = p; | |
2056 | |
2057 /* Are we moving the last element? */ | |
2058 if (G.page_tails[order] == p) | |
2059 G.page_tails[order] = previous; | |
2060 p = previous; | |
2061 } | |
2062 | |
2063 previous = p; | |
2064 p = next; | |
2065 } | |
2066 while (! done); | |
2067 | |
2068 /* Now, restore the in_use_p vectors for any pages from contexts | |
2069 other than the current one. */ | |
2070 for (p = G.pages[order]; p; p = p->next) | |
2071 if (p->context_depth != G.context_depth) | |
2072 ggc_recalculate_in_use_p (p); | |
2073 } | |
2074 } | |
2075 | |
2076 #ifdef ENABLE_GC_CHECKING | |
2077 /* Clobber all free objects. */ | |
2078 | |
2079 static void | |
2080 poison_pages (void) | |
2081 { | |
2082 unsigned order; | |
2083 | |
2084 for (order = 2; order < NUM_ORDERS; order++) | |
2085 { | |
2086 size_t size = OBJECT_SIZE (order); | |
2087 page_entry *p; | |
2088 | |
2089 for (p = G.pages[order]; p != NULL; p = p->next) | |
2090 { | |
2091 size_t num_objects; | |
2092 size_t i; | |
2093 | |
2094 if (p->context_depth != G.context_depth) | |
2095 /* Since we don't do any collection for pages in pushed | |
2096 contexts, there's no need to do any poisoning. And | |
2097 besides, the IN_USE_P array isn't valid until we pop | |
2098 contexts. */ | |
2099 continue; | |
2100 | |
2101 num_objects = OBJECTS_IN_PAGE (p); | |
2102 for (i = 0; i < num_objects; i++) | |
2103 { | |
2104 size_t word, bit; | |
2105 word = i / HOST_BITS_PER_LONG; | |
2106 bit = i % HOST_BITS_PER_LONG; | |
2107 if (((p->in_use_p[word] >> bit) & 1) == 0) | |
2108 { | |
2109 char *object = p->page + i * size; | |
2110 | |
2111 /* Keep poison-by-write when we expect to use Valgrind, | |
2112 so the exact same memory semantics is kept, in case | |
2113 there are memory errors. We override this request | |
2114 below. */ | |
2115 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_UNDEFINED (object, | |
2116 size)); | |
2117 memset (object, 0xa5, size); | |
2118 | |
2119 /* Drop the handle to avoid handle leak. */ | |
2120 VALGRIND_DISCARD (VALGRIND_MAKE_MEM_NOACCESS (object, size)); | |
2121 } | |
2122 } | |
2123 } | |
2124 } | |
2125 } | |
2126 #else | |
2127 #define poison_pages() | |
2128 #endif | |
2129 | |
2130 #ifdef ENABLE_GC_ALWAYS_COLLECT | |
2131 /* Validate that the reportedly free objects actually are. */ | |
2132 | |
2133 static void | |
2134 validate_free_objects (void) | |
2135 { | |
2136 struct free_object *f, *next, *still_free = NULL; | |
2137 | |
2138 for (f = G.free_object_list; f ; f = next) | |
2139 { | |
2140 page_entry *pe = lookup_page_table_entry (f->object); | |
2141 size_t bit, word; | |
2142 | |
2143 bit = OFFSET_TO_BIT ((char *)f->object - pe->page, pe->order); | |
2144 word = bit / HOST_BITS_PER_LONG; | |
2145 bit = bit % HOST_BITS_PER_LONG; | |
2146 next = f->next; | |
2147 | |
2148 /* Make certain it isn't visible from any root. Notice that we | |
2149 do this check before sweep_pages merges save_in_use_p. */ | |
2150 gcc_assert (!(pe->in_use_p[word] & (1UL << bit))); | |
2151 | |
2152 /* If the object comes from an outer context, then retain the | |
2153 free_object entry, so that we can verify that the address | |
2154 isn't live on the stack in some outer context. */ | |
2155 if (pe->context_depth != G.context_depth) | |
2156 { | |
2157 f->next = still_free; | |
2158 still_free = f; | |
2159 } | |
2160 else | |
2161 free (f); | |
2162 } | |
2163 | |
2164 G.free_object_list = still_free; | |
2165 } | |
2166 #else | |
2167 #define validate_free_objects() | |
2168 #endif | |
2169 | |
2170 /* Top level mark-and-sweep routine. */ | |
2171 | |
2172 void | |
2173 ggc_collect (void) | |
2174 { | |
2175 /* Avoid frequent unnecessary work by skipping collection if the | |
2176 total allocations haven't expanded much since the last | |
2177 collection. */ | |
2178 float allocated_last_gc = | |
2179 MAX (G.allocated_last_gc, (size_t)PARAM_VALUE (GGC_MIN_HEAPSIZE) * 1024); | |
2180 | |
2181 float min_expand = allocated_last_gc * PARAM_VALUE (GGC_MIN_EXPAND) / 100; | |
2182 if (G.allocated < allocated_last_gc + min_expand && !ggc_force_collect) | |
2183 return; | |
2184 | |
2185 timevar_push (TV_GC); | |
2186 if (!quiet_flag) | |
2187 fprintf (stderr, " {GC %luk -> ", (unsigned long) G.allocated / 1024); | |
2188 if (GGC_DEBUG_LEVEL >= 2) | |
2189 fprintf (G.debug_file, "BEGIN COLLECTING\n"); | |
2190 | |
2191 /* Zero the total allocated bytes. This will be recalculated in the | |
2192 sweep phase. */ | |
2193 G.allocated = 0; | |
2194 | |
2195 /* Release the pages we freed the last time we collected, but didn't | |
2196 reuse in the interim. */ | |
2197 release_pages (); | |
2198 | |
2199 /* Indicate that we've seen collections at this context depth. */ | |
2200 G.context_depth_collections = ((unsigned long)1 << (G.context_depth + 1)) - 1; | |
2201 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2202 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
|
2203 |
111 | 2204 in_gc = true; |
0 | 2205 clear_marks (); |
2206 ggc_mark_roots (); | |
111 | 2207 ggc_handle_finalizers (); |
2208 | |
2209 if (GATHER_STATISTICS) | |
2210 ggc_prune_overhead_list (); | |
2211 | |
0 | 2212 poison_pages (); |
2213 validate_free_objects (); | |
2214 sweep_pages (); | |
2215 | |
111 | 2216 in_gc = false; |
0 | 2217 G.allocated_last_gc = G.allocated; |
2218 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2219 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
|
2220 |
0 | 2221 timevar_pop (TV_GC); |
2222 | |
2223 if (!quiet_flag) | |
2224 fprintf (stderr, "%luk}", (unsigned long) G.allocated / 1024); | |
2225 if (GGC_DEBUG_LEVEL >= 2) | |
2226 fprintf (G.debug_file, "END COLLECTING\n"); | |
2227 } | |
2228 | |
111 | 2229 /* Assume that all GGC memory is reachable and grow the limits for next collection. |
2230 With checking, trigger GGC so -Q compilation outputs how much of memory really is | |
2231 reachable. */ | |
2232 | |
2233 void | |
2234 ggc_grow (void) | |
2235 { | |
2236 if (!flag_checking) | |
2237 G.allocated_last_gc = MAX (G.allocated_last_gc, | |
2238 G.allocated); | |
2239 else | |
2240 ggc_collect (); | |
2241 if (!quiet_flag) | |
2242 fprintf (stderr, " {GC start %luk} ", (unsigned long) G.allocated / 1024); | |
2243 } | |
2244 | |
0 | 2245 /* Print allocation statistics. */ |
2246 #define SCALE(x) ((unsigned long) ((x) < 1024*10 \ | |
2247 ? (x) \ | |
2248 : ((x) < 1024*1024*10 \ | |
2249 ? (x) / 1024 \ | |
2250 : (x) / (1024*1024)))) | |
2251 #define STAT_LABEL(x) ((x) < 1024*10 ? ' ' : ((x) < 1024*1024*10 ? 'k' : 'M')) | |
2252 | |
2253 void | |
2254 ggc_print_statistics (void) | |
2255 { | |
2256 struct ggc_statistics stats; | |
2257 unsigned int i; | |
2258 size_t total_overhead = 0; | |
2259 | |
2260 /* Clear the statistics. */ | |
2261 memset (&stats, 0, sizeof (stats)); | |
2262 | |
2263 /* Make sure collection will really occur. */ | |
2264 G.allocated_last_gc = 0; | |
2265 | |
2266 /* Collect and print the statistics common across collectors. */ | |
2267 ggc_print_common_statistics (stderr, &stats); | |
2268 | |
2269 /* Release free pages so that we will not count the bytes allocated | |
2270 there as part of the total allocated memory. */ | |
2271 release_pages (); | |
2272 | |
2273 /* Collect some information about the various sizes of | |
2274 allocation. */ | |
2275 fprintf (stderr, | |
2276 "Memory still allocated at the end of the compilation process\n"); | |
111 | 2277 fprintf (stderr, "%-8s %10s %10s %10s\n", |
0 | 2278 "Size", "Allocated", "Used", "Overhead"); |
2279 for (i = 0; i < NUM_ORDERS; ++i) | |
2280 { | |
2281 page_entry *p; | |
2282 size_t allocated; | |
2283 size_t in_use; | |
2284 size_t overhead; | |
2285 | |
2286 /* Skip empty entries. */ | |
2287 if (!G.pages[i]) | |
2288 continue; | |
2289 | |
2290 overhead = allocated = in_use = 0; | |
2291 | |
2292 /* Figure out the total number of bytes allocated for objects of | |
2293 this size, and how many of them are actually in use. Also figure | |
2294 out how much memory the page table is using. */ | |
2295 for (p = G.pages[i]; p; p = p->next) | |
2296 { | |
2297 allocated += p->bytes; | |
2298 in_use += | |
2299 (OBJECTS_IN_PAGE (p) - p->num_free_objects) * OBJECT_SIZE (i); | |
2300 | |
2301 overhead += (sizeof (page_entry) - sizeof (long) | |
2302 + BITMAP_SIZE (OBJECTS_IN_PAGE (p) + 1)); | |
2303 } | |
111 | 2304 fprintf (stderr, "%-8lu %10lu%c %10lu%c %10lu%c\n", |
0 | 2305 (unsigned long) OBJECT_SIZE (i), |
2306 SCALE (allocated), STAT_LABEL (allocated), | |
2307 SCALE (in_use), STAT_LABEL (in_use), | |
2308 SCALE (overhead), STAT_LABEL (overhead)); | |
2309 total_overhead += overhead; | |
2310 } | |
111 | 2311 fprintf (stderr, "%-8s %10lu%c %10lu%c %10lu%c\n", "Total", |
0 | 2312 SCALE (G.bytes_mapped), STAT_LABEL (G.bytes_mapped), |
111 | 2313 SCALE (G.allocated), STAT_LABEL (G.allocated), |
0 | 2314 SCALE (total_overhead), STAT_LABEL (total_overhead)); |
2315 | |
111 | 2316 if (GATHER_STATISTICS) |
2317 { | |
2318 fprintf (stderr, "\nTotal allocations and overheads during " | |
2319 "the compilation process\n"); | |
2320 | |
2321 fprintf (stderr, "Total Overhead: %10" | |
2322 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_overhead); | |
2323 fprintf (stderr, "Total Allocated: %10" | |
2324 HOST_LONG_LONG_FORMAT "d\n", | |
2325 G.stats.total_allocated); | |
2326 | |
2327 fprintf (stderr, "Total Overhead under 32B: %10" | |
2328 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_overhead_under32); | |
2329 fprintf (stderr, "Total Allocated under 32B: %10" | |
2330 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_allocated_under32); | |
2331 fprintf (stderr, "Total Overhead under 64B: %10" | |
2332 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_overhead_under64); | |
2333 fprintf (stderr, "Total Allocated under 64B: %10" | |
2334 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_allocated_under64); | |
2335 fprintf (stderr, "Total Overhead under 128B: %10" | |
2336 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_overhead_under128); | |
2337 fprintf (stderr, "Total Allocated under 128B: %10" | |
2338 HOST_LONG_LONG_FORMAT "d\n", G.stats.total_allocated_under128); | |
2339 | |
2340 for (i = 0; i < NUM_ORDERS; i++) | |
2341 if (G.stats.total_allocated_per_order[i]) | |
2342 { | |
2343 fprintf (stderr, "Total Overhead page size %9lu: %10" | |
2344 HOST_LONG_LONG_FORMAT "d\n", | |
2345 (unsigned long) OBJECT_SIZE (i), | |
2346 G.stats.total_overhead_per_order[i]); | |
2347 fprintf (stderr, "Total Allocated page size %9lu: %10" | |
2348 HOST_LONG_LONG_FORMAT "d\n", | |
2349 (unsigned long) OBJECT_SIZE (i), | |
2350 G.stats.total_allocated_per_order[i]); | |
2351 } | |
0 | 2352 } |
2353 } | |
2354 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2355 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
|
2356 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2357 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
|
2358 }; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2359 |
0 | 2360 struct ggc_pch_data |
2361 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2362 struct ggc_pch_ondisk d; |
111 | 2363 uintptr_t base[NUM_ORDERS]; |
0 | 2364 size_t written[NUM_ORDERS]; |
2365 }; | |
2366 | |
2367 struct ggc_pch_data * | |
2368 init_ggc_pch (void) | |
2369 { | |
2370 return XCNEW (struct ggc_pch_data); | |
2371 } | |
2372 | |
2373 void | |
2374 ggc_pch_count_object (struct ggc_pch_data *d, void *x ATTRIBUTE_UNUSED, | |
111 | 2375 size_t size, bool is_string ATTRIBUTE_UNUSED) |
0 | 2376 { |
2377 unsigned order; | |
2378 | |
2379 if (size < NUM_SIZE_LOOKUP) | |
2380 order = size_lookup[size]; | |
2381 else | |
2382 { | |
2383 order = 10; | |
2384 while (size > OBJECT_SIZE (order)) | |
2385 order++; | |
2386 } | |
2387 | |
2388 d->d.totals[order]++; | |
2389 } | |
2390 | |
2391 size_t | |
2392 ggc_pch_total_size (struct ggc_pch_data *d) | |
2393 { | |
2394 size_t a = 0; | |
2395 unsigned i; | |
2396 | |
2397 for (i = 0; i < NUM_ORDERS; i++) | |
111 | 2398 a += PAGE_ALIGN (d->d.totals[i] * OBJECT_SIZE (i)); |
0 | 2399 return a; |
2400 } | |
2401 | |
2402 void | |
2403 ggc_pch_this_base (struct ggc_pch_data *d, void *base) | |
2404 { | |
111 | 2405 uintptr_t a = (uintptr_t) base; |
0 | 2406 unsigned i; |
2407 | |
2408 for (i = 0; i < NUM_ORDERS; i++) | |
2409 { | |
2410 d->base[i] = a; | |
111 | 2411 a += PAGE_ALIGN (d->d.totals[i] * OBJECT_SIZE (i)); |
0 | 2412 } |
2413 } | |
2414 | |
2415 | |
2416 char * | |
2417 ggc_pch_alloc_object (struct ggc_pch_data *d, void *x ATTRIBUTE_UNUSED, | |
111 | 2418 size_t size, bool is_string ATTRIBUTE_UNUSED) |
0 | 2419 { |
2420 unsigned order; | |
2421 char *result; | |
2422 | |
2423 if (size < NUM_SIZE_LOOKUP) | |
2424 order = size_lookup[size]; | |
2425 else | |
2426 { | |
2427 order = 10; | |
2428 while (size > OBJECT_SIZE (order)) | |
2429 order++; | |
2430 } | |
2431 | |
2432 result = (char *) d->base[order]; | |
2433 d->base[order] += OBJECT_SIZE (order); | |
2434 return result; | |
2435 } | |
2436 | |
2437 void | |
2438 ggc_pch_prepare_write (struct ggc_pch_data *d ATTRIBUTE_UNUSED, | |
2439 FILE *f ATTRIBUTE_UNUSED) | |
2440 { | |
2441 /* Nothing to do. */ | |
2442 } | |
2443 | |
2444 void | |
111 | 2445 ggc_pch_write_object (struct ggc_pch_data *d, |
0 | 2446 FILE *f, void *x, void *newx ATTRIBUTE_UNUSED, |
2447 size_t size, bool is_string ATTRIBUTE_UNUSED) | |
2448 { | |
2449 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
|
2450 static const char emptyBytes[256] = { 0 }; |
0 | 2451 |
2452 if (size < NUM_SIZE_LOOKUP) | |
2453 order = size_lookup[size]; | |
2454 else | |
2455 { | |
2456 order = 10; | |
2457 while (size > OBJECT_SIZE (order)) | |
2458 order++; | |
2459 } | |
2460 | |
2461 if (fwrite (x, size, 1, f) != 1) | |
111 | 2462 fatal_error (input_location, "can%'t write PCH file: %m"); |
0 | 2463 |
2464 /* If SIZE is not the same as OBJECT_SIZE(order), then we need to pad the | |
2465 object out to OBJECT_SIZE(order). This happens for strings. */ | |
2466 | |
2467 if (size != OBJECT_SIZE (order)) | |
2468 { | |
111 | 2469 unsigned padding = OBJECT_SIZE (order) - size; |
0 | 2470 |
2471 /* To speed small writes, we use a nulled-out array that's larger | |
2472 than most padding requests as the source for our null bytes. This | |
2473 permits us to do the padding with fwrite() rather than fseek(), and | |
2474 limits the chance the OS may try to flush any outstanding writes. */ | |
111 | 2475 if (padding <= sizeof (emptyBytes)) |
0 | 2476 { |
2477 if (fwrite (emptyBytes, 1, padding, f) != padding) | |
111 | 2478 fatal_error (input_location, "can%'t write PCH file"); |
0 | 2479 } |
2480 else | |
2481 { | |
2482 /* Larger than our buffer? Just default to fseek. */ | |
2483 if (fseek (f, padding, SEEK_CUR) != 0) | |
111 | 2484 fatal_error (input_location, "can%'t write PCH file"); |
0 | 2485 } |
2486 } | |
2487 | |
2488 d->written[order]++; | |
2489 if (d->written[order] == d->d.totals[order] | |
2490 && fseek (f, ROUND_UP_VALUE (d->d.totals[order] * OBJECT_SIZE (order), | |
2491 G.pagesize), | |
2492 SEEK_CUR) != 0) | |
111 | 2493 fatal_error (input_location, "can%'t write PCH file: %m"); |
0 | 2494 } |
2495 | |
2496 void | |
2497 ggc_pch_finish (struct ggc_pch_data *d, FILE *f) | |
2498 { | |
2499 if (fwrite (&d->d, sizeof (d->d), 1, f) != 1) | |
111 | 2500 fatal_error (input_location, "can%'t write PCH file: %m"); |
0 | 2501 free (d); |
2502 } | |
2503 | |
2504 /* Move the PCH PTE entries just added to the end of by_depth, to the | |
2505 front. */ | |
2506 | |
2507 static void | |
2508 move_ptes_to_front (int count_old_page_tables, int count_new_page_tables) | |
2509 { | |
2510 /* First, we swap the new entries to the front of the varrays. */ | |
2511 page_entry **new_by_depth; | |
2512 unsigned long **new_save_in_use; | |
2513 | |
2514 new_by_depth = XNEWVEC (page_entry *, G.by_depth_max); | |
2515 new_save_in_use = XNEWVEC (unsigned long *, G.by_depth_max); | |
2516 | |
2517 memcpy (&new_by_depth[0], | |
2518 &G.by_depth[count_old_page_tables], | |
2519 count_new_page_tables * sizeof (void *)); | |
2520 memcpy (&new_by_depth[count_new_page_tables], | |
2521 &G.by_depth[0], | |
2522 count_old_page_tables * sizeof (void *)); | |
2523 memcpy (&new_save_in_use[0], | |
2524 &G.save_in_use[count_old_page_tables], | |
2525 count_new_page_tables * sizeof (void *)); | |
2526 memcpy (&new_save_in_use[count_new_page_tables], | |
2527 &G.save_in_use[0], | |
2528 count_old_page_tables * sizeof (void *)); | |
2529 | |
2530 free (G.by_depth); | |
2531 free (G.save_in_use); | |
2532 | |
2533 G.by_depth = new_by_depth; | |
2534 G.save_in_use = new_save_in_use; | |
2535 | |
2536 /* Now update all the index_by_depth fields. */ | |
111 | 2537 for (unsigned i = G.by_depth_in_use; i--;) |
0 | 2538 { |
111 | 2539 page_entry *p = G.by_depth[i]; |
2540 p->index_by_depth = i; | |
0 | 2541 } |
2542 | |
2543 /* And last, we update the depth pointers in G.depth. The first | |
2544 entry is already 0, and context 0 entries always start at index | |
2545 0, so there is nothing to update in the first slot. We need a | |
2546 second slot, only if we have old ptes, and if we do, they start | |
2547 at index count_new_page_tables. */ | |
2548 if (count_old_page_tables) | |
2549 push_depth (count_new_page_tables); | |
2550 } | |
2551 | |
2552 void | |
2553 ggc_pch_read (FILE *f, void *addr) | |
2554 { | |
2555 struct ggc_pch_ondisk d; | |
2556 unsigned i; | |
2557 char *offs = (char *) addr; | |
2558 unsigned long count_old_page_tables; | |
2559 unsigned long count_new_page_tables; | |
2560 | |
2561 count_old_page_tables = G.by_depth_in_use; | |
2562 | |
2563 /* We've just read in a PCH file. So, every object that used to be | |
2564 allocated is now free. */ | |
2565 clear_marks (); | |
2566 #ifdef ENABLE_GC_CHECKING | |
2567 poison_pages (); | |
2568 #endif | |
2569 /* Since we free all the allocated objects, the free list becomes | |
2570 useless. Validate it now, which will also clear it. */ | |
111 | 2571 validate_free_objects (); |
0 | 2572 |
2573 /* No object read from a PCH file should ever be freed. So, set the | |
2574 context depth to 1, and set the depth of all the currently-allocated | |
2575 pages to be 1 too. PCH pages will have depth 0. */ | |
2576 gcc_assert (!G.context_depth); | |
2577 G.context_depth = 1; | |
111 | 2578 /* Allocate space for the depth 1 finalizers. */ |
2579 G.finalizers.safe_push (vNULL); | |
2580 G.vec_finalizers.safe_push (vNULL); | |
2581 gcc_assert (G.finalizers.length() == 2); | |
0 | 2582 for (i = 0; i < NUM_ORDERS; i++) |
2583 { | |
2584 page_entry *p; | |
2585 for (p = G.pages[i]; p != NULL; p = p->next) | |
2586 p->context_depth = G.context_depth; | |
2587 } | |
2588 | |
2589 /* Allocate the appropriate page-table entries for the pages read from | |
2590 the PCH file. */ | |
2591 if (fread (&d, sizeof (d), 1, f) != 1) | |
111 | 2592 fatal_error (input_location, "can%'t read PCH file: %m"); |
0 | 2593 |
2594 for (i = 0; i < NUM_ORDERS; i++) | |
2595 { | |
2596 struct page_entry *entry; | |
2597 char *pte; | |
2598 size_t bytes; | |
2599 size_t num_objs; | |
2600 size_t j; | |
2601 | |
2602 if (d.totals[i] == 0) | |
2603 continue; | |
2604 | |
111 | 2605 bytes = PAGE_ALIGN (d.totals[i] * OBJECT_SIZE (i)); |
0 | 2606 num_objs = bytes / OBJECT_SIZE (i); |
2607 entry = XCNEWVAR (struct page_entry, (sizeof (struct page_entry) | |
2608 - sizeof (long) | |
2609 + BITMAP_SIZE (num_objs + 1))); | |
2610 entry->bytes = bytes; | |
2611 entry->page = offs; | |
2612 entry->context_depth = 0; | |
2613 offs += bytes; | |
2614 entry->num_free_objects = 0; | |
2615 entry->order = i; | |
2616 | |
2617 for (j = 0; | |
2618 j + HOST_BITS_PER_LONG <= num_objs + 1; | |
2619 j += HOST_BITS_PER_LONG) | |
2620 entry->in_use_p[j / HOST_BITS_PER_LONG] = -1; | |
2621 for (; j < num_objs + 1; j++) | |
2622 entry->in_use_p[j / HOST_BITS_PER_LONG] | |
2623 |= 1L << (j % HOST_BITS_PER_LONG); | |
2624 | |
2625 for (pte = entry->page; | |
2626 pte < entry->page + entry->bytes; | |
2627 pte += G.pagesize) | |
2628 set_page_table_entry (pte, entry); | |
2629 | |
2630 if (G.page_tails[i] != NULL) | |
2631 G.page_tails[i]->next = entry; | |
2632 else | |
2633 G.pages[i] = entry; | |
2634 G.page_tails[i] = entry; | |
2635 | |
2636 /* We start off by just adding all the new information to the | |
2637 end of the varrays, later, we will move the new information | |
2638 to the front of the varrays, as the PCH page tables are at | |
2639 context 0. */ | |
2640 push_by_depth (entry, 0); | |
2641 } | |
2642 | |
2643 /* Now, we update the various data structures that speed page table | |
2644 handling. */ | |
2645 count_new_page_tables = G.by_depth_in_use - count_old_page_tables; | |
2646 | |
2647 move_ptes_to_front (count_old_page_tables, count_new_page_tables); | |
2648 | |
2649 /* Update the statistics. */ | |
2650 G.allocated = G.allocated_last_gc = offs - (char *)addr; | |
2651 } |