Print this page
rev 2691 : [mq]: g1-reference-processing
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
+++ new/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
1 1 /*
2 2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/symbolTable.hpp"
27 27 #include "classfile/systemDictionary.hpp"
28 28 #include "code/codeCache.hpp"
29 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
30 30 #include "gc_implementation/parallelScavenge/generationSizer.hpp"
31 31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
32 32 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
33 33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
34 34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
35 35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
36 36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
37 37 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
38 38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
39 39 #include "gc_implementation/parallelScavenge/psPermGen.hpp"
40 40 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
41 41 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
42 42 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
43 43 #include "gc_implementation/shared/isGCActiveMark.hpp"
44 44 #include "gc_interface/gcCause.hpp"
45 45 #include "memory/gcLocker.inline.hpp"
46 46 #include "memory/referencePolicy.hpp"
47 47 #include "memory/referenceProcessor.hpp"
48 48 #include "oops/methodDataOop.hpp"
49 49 #include "oops/oop.inline.hpp"
50 50 #include "oops/oop.pcgc.inline.hpp"
51 51 #include "runtime/fprofiler.hpp"
52 52 #include "runtime/safepoint.hpp"
53 53 #include "runtime/vmThread.hpp"
54 54 #include "services/management.hpp"
55 55 #include "services/memoryService.hpp"
56 56 #include "utilities/events.hpp"
57 57 #include "utilities/stack.inline.hpp"
58 58
59 59 #include <math.h>
60 60
61 61 // All sizes are in HeapWords.
62 62 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words
63 63 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
64 64 const size_t ParallelCompactData::RegionSizeBytes =
65 65 RegionSize << LogHeapWordSize;
66 66 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
67 67 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
68 68 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
69 69
70 70 const ParallelCompactData::RegionData::region_sz_t
71 71 ParallelCompactData::RegionData::dc_shift = 27;
72 72
73 73 const ParallelCompactData::RegionData::region_sz_t
74 74 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
75 75
76 76 const ParallelCompactData::RegionData::region_sz_t
77 77 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
78 78
79 79 const ParallelCompactData::RegionData::region_sz_t
80 80 ParallelCompactData::RegionData::los_mask = ~dc_mask;
81 81
82 82 const ParallelCompactData::RegionData::region_sz_t
83 83 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
84 84
85 85 const ParallelCompactData::RegionData::region_sz_t
86 86 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
87 87
88 88 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
89 89 bool PSParallelCompact::_print_phases = false;
90 90
91 91 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
92 92 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL;
93 93
94 94 double PSParallelCompact::_dwl_mean;
95 95 double PSParallelCompact::_dwl_std_dev;
96 96 double PSParallelCompact::_dwl_first_term;
97 97 double PSParallelCompact::_dwl_adjustment;
98 98 #ifdef ASSERT
99 99 bool PSParallelCompact::_dwl_initialized = false;
100 100 #endif // #ifdef ASSERT
101 101
102 102 #ifdef VALIDATE_MARK_SWEEP
103 103 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL;
104 104 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL;
105 105 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL;
106 106 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL;
107 107 size_t PSParallelCompact::_live_oops_index = 0;
108 108 size_t PSParallelCompact::_live_oops_index_at_perm = 0;
109 109 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL;
110 110 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL;
111 111 bool PSParallelCompact::_pointer_tracking = false;
112 112 bool PSParallelCompact::_root_tracking = true;
113 113
114 114 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
115 115 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
116 116 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL;
117 117 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
118 118 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
119 119 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL;
120 120 #endif
121 121
122 122 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
123 123 HeapWord* destination)
124 124 {
125 125 assert(src_region_idx != 0, "invalid src_region_idx");
126 126 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
127 127 assert(destination != NULL, "invalid destination argument");
128 128
129 129 _src_region_idx = src_region_idx;
130 130 _partial_obj_size = partial_obj_size;
131 131 _destination = destination;
132 132
133 133 // These fields may not be updated below, so make sure they're clear.
134 134 assert(_dest_region_addr == NULL, "should have been cleared");
135 135 assert(_first_src_addr == NULL, "should have been cleared");
136 136
137 137 // Determine the number of destination regions for the partial object.
138 138 HeapWord* const last_word = destination + partial_obj_size - 1;
139 139 const ParallelCompactData& sd = PSParallelCompact::summary_data();
140 140 HeapWord* const beg_region_addr = sd.region_align_down(destination);
141 141 HeapWord* const end_region_addr = sd.region_align_down(last_word);
142 142
143 143 if (beg_region_addr == end_region_addr) {
144 144 // One destination region.
145 145 _destination_count = 1;
146 146 if (end_region_addr == destination) {
147 147 // The destination falls on a region boundary, thus the first word of the
148 148 // partial object will be the first word copied to the destination region.
149 149 _dest_region_addr = end_region_addr;
150 150 _first_src_addr = sd.region_to_addr(src_region_idx);
151 151 }
152 152 } else {
153 153 // Two destination regions. When copied, the partial object will cross a
154 154 // destination region boundary, so a word somewhere within the partial
155 155 // object will be the first word copied to the second destination region.
156 156 _destination_count = 2;
157 157 _dest_region_addr = end_region_addr;
158 158 const size_t ofs = pointer_delta(end_region_addr, destination);
159 159 assert(ofs < _partial_obj_size, "sanity");
160 160 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
161 161 }
162 162 }
163 163
164 164 void SplitInfo::clear()
165 165 {
166 166 _src_region_idx = 0;
167 167 _partial_obj_size = 0;
168 168 _destination = NULL;
169 169 _destination_count = 0;
170 170 _dest_region_addr = NULL;
171 171 _first_src_addr = NULL;
172 172 assert(!is_valid(), "sanity");
173 173 }
174 174
175 175 #ifdef ASSERT
176 176 void SplitInfo::verify_clear()
177 177 {
178 178 assert(_src_region_idx == 0, "not clear");
179 179 assert(_partial_obj_size == 0, "not clear");
180 180 assert(_destination == NULL, "not clear");
181 181 assert(_destination_count == 0, "not clear");
182 182 assert(_dest_region_addr == NULL, "not clear");
183 183 assert(_first_src_addr == NULL, "not clear");
184 184 }
185 185 #endif // #ifdef ASSERT
186 186
187 187
188 188 #ifndef PRODUCT
189 189 const char* PSParallelCompact::space_names[] = {
190 190 "perm", "old ", "eden", "from", "to "
191 191 };
192 192
193 193 void PSParallelCompact::print_region_ranges()
194 194 {
195 195 tty->print_cr("space bottom top end new_top");
196 196 tty->print_cr("------ ---------- ---------- ---------- ----------");
197 197
198 198 for (unsigned int id = 0; id < last_space_id; ++id) {
199 199 const MutableSpace* space = _space_info[id].space();
200 200 tty->print_cr("%u %s "
201 201 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
202 202 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
203 203 id, space_names[id],
204 204 summary_data().addr_to_region_idx(space->bottom()),
205 205 summary_data().addr_to_region_idx(space->top()),
206 206 summary_data().addr_to_region_idx(space->end()),
207 207 summary_data().addr_to_region_idx(_space_info[id].new_top()));
208 208 }
209 209 }
210 210
211 211 void
212 212 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
213 213 {
214 214 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
215 215 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
216 216
217 217 ParallelCompactData& sd = PSParallelCompact::summary_data();
218 218 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
219 219 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
220 220 REGION_IDX_FORMAT " " PTR_FORMAT " "
221 221 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
222 222 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
223 223 i, c->data_location(), dci, c->destination(),
224 224 c->partial_obj_size(), c->live_obj_size(),
225 225 c->data_size(), c->source_region(), c->destination_count());
226 226
227 227 #undef REGION_IDX_FORMAT
228 228 #undef REGION_DATA_FORMAT
229 229 }
230 230
231 231 void
232 232 print_generic_summary_data(ParallelCompactData& summary_data,
233 233 HeapWord* const beg_addr,
234 234 HeapWord* const end_addr)
235 235 {
236 236 size_t total_words = 0;
237 237 size_t i = summary_data.addr_to_region_idx(beg_addr);
238 238 const size_t last = summary_data.addr_to_region_idx(end_addr);
239 239 HeapWord* pdest = 0;
240 240
241 241 while (i <= last) {
242 242 ParallelCompactData::RegionData* c = summary_data.region(i);
243 243 if (c->data_size() != 0 || c->destination() != pdest) {
244 244 print_generic_summary_region(i, c);
245 245 total_words += c->data_size();
246 246 pdest = c->destination();
247 247 }
248 248 ++i;
249 249 }
250 250
251 251 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
252 252 }
253 253
254 254 void
255 255 print_generic_summary_data(ParallelCompactData& summary_data,
256 256 SpaceInfo* space_info)
257 257 {
258 258 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
259 259 const MutableSpace* space = space_info[id].space();
260 260 print_generic_summary_data(summary_data, space->bottom(),
261 261 MAX2(space->top(), space_info[id].new_top()));
262 262 }
263 263 }
264 264
265 265 void
266 266 print_initial_summary_region(size_t i,
267 267 const ParallelCompactData::RegionData* c,
268 268 bool newline = true)
269 269 {
270 270 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
271 271 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
272 272 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
273 273 i, c->destination(),
274 274 c->partial_obj_size(), c->live_obj_size(),
275 275 c->data_size(), c->source_region(), c->destination_count());
276 276 if (newline) tty->cr();
277 277 }
278 278
279 279 void
280 280 print_initial_summary_data(ParallelCompactData& summary_data,
281 281 const MutableSpace* space) {
282 282 if (space->top() == space->bottom()) {
283 283 return;
284 284 }
285 285
286 286 const size_t region_size = ParallelCompactData::RegionSize;
287 287 typedef ParallelCompactData::RegionData RegionData;
288 288 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
289 289 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
290 290 const RegionData* c = summary_data.region(end_region - 1);
291 291 HeapWord* end_addr = c->destination() + c->data_size();
292 292 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
293 293
294 294 // Print (and count) the full regions at the beginning of the space.
295 295 size_t full_region_count = 0;
296 296 size_t i = summary_data.addr_to_region_idx(space->bottom());
297 297 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
298 298 print_initial_summary_region(i, summary_data.region(i));
299 299 ++full_region_count;
300 300 ++i;
301 301 }
302 302
303 303 size_t live_to_right = live_in_space - full_region_count * region_size;
304 304
305 305 double max_reclaimed_ratio = 0.0;
306 306 size_t max_reclaimed_ratio_region = 0;
307 307 size_t max_dead_to_right = 0;
308 308 size_t max_live_to_right = 0;
309 309
310 310 // Print the 'reclaimed ratio' for regions while there is something live in
311 311 // the region or to the right of it. The remaining regions are empty (and
312 312 // uninteresting), and computing the ratio will result in division by 0.
313 313 while (i < end_region && live_to_right > 0) {
314 314 c = summary_data.region(i);
315 315 HeapWord* const region_addr = summary_data.region_to_addr(i);
316 316 const size_t used_to_right = pointer_delta(space->top(), region_addr);
317 317 const size_t dead_to_right = used_to_right - live_to_right;
318 318 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
319 319
320 320 if (reclaimed_ratio > max_reclaimed_ratio) {
321 321 max_reclaimed_ratio = reclaimed_ratio;
322 322 max_reclaimed_ratio_region = i;
323 323 max_dead_to_right = dead_to_right;
324 324 max_live_to_right = live_to_right;
325 325 }
326 326
327 327 print_initial_summary_region(i, c, false);
328 328 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
329 329 reclaimed_ratio, dead_to_right, live_to_right);
330 330
331 331 live_to_right -= c->data_size();
332 332 ++i;
333 333 }
334 334
335 335 // Any remaining regions are empty. Print one more if there is one.
336 336 if (i < end_region) {
337 337 print_initial_summary_region(i, summary_data.region(i));
338 338 }
339 339
340 340 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
341 341 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
342 342 max_reclaimed_ratio_region, max_dead_to_right,
343 343 max_live_to_right, max_reclaimed_ratio);
344 344 }
345 345
346 346 void
347 347 print_initial_summary_data(ParallelCompactData& summary_data,
348 348 SpaceInfo* space_info) {
349 349 unsigned int id = PSParallelCompact::perm_space_id;
350 350 const MutableSpace* space;
351 351 do {
352 352 space = space_info[id].space();
353 353 print_initial_summary_data(summary_data, space);
354 354 } while (++id < PSParallelCompact::eden_space_id);
355 355
356 356 do {
357 357 space = space_info[id].space();
358 358 print_generic_summary_data(summary_data, space->bottom(), space->top());
359 359 } while (++id < PSParallelCompact::last_space_id);
360 360 }
361 361 #endif // #ifndef PRODUCT
362 362
363 363 #ifdef ASSERT
364 364 size_t add_obj_count;
365 365 size_t add_obj_size;
366 366 size_t mark_bitmap_count;
367 367 size_t mark_bitmap_size;
368 368 #endif // #ifdef ASSERT
369 369
370 370 ParallelCompactData::ParallelCompactData()
371 371 {
372 372 _region_start = 0;
373 373
374 374 _region_vspace = 0;
375 375 _region_data = 0;
376 376 _region_count = 0;
377 377 }
378 378
379 379 bool ParallelCompactData::initialize(MemRegion covered_region)
380 380 {
381 381 _region_start = covered_region.start();
382 382 const size_t region_size = covered_region.word_size();
383 383 DEBUG_ONLY(_region_end = _region_start + region_size;)
384 384
385 385 assert(region_align_down(_region_start) == _region_start,
386 386 "region start not aligned");
387 387 assert((region_size & RegionSizeOffsetMask) == 0,
388 388 "region size not a multiple of RegionSize");
389 389
390 390 bool result = initialize_region_data(region_size);
391 391
392 392 return result;
393 393 }
394 394
395 395 PSVirtualSpace*
396 396 ParallelCompactData::create_vspace(size_t count, size_t element_size)
397 397 {
398 398 const size_t raw_bytes = count * element_size;
399 399 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
400 400 const size_t granularity = os::vm_allocation_granularity();
401 401 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
402 402
403 403 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
404 404 MAX2(page_sz, granularity);
405 405 ReservedSpace rs(bytes, rs_align, rs_align > 0);
406 406 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
407 407 rs.size());
408 408 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
409 409 if (vspace != 0) {
410 410 if (vspace->expand_by(bytes)) {
411 411 return vspace;
412 412 }
413 413 delete vspace;
414 414 // Release memory reserved in the space.
415 415 rs.release();
416 416 }
417 417
418 418 return 0;
419 419 }
420 420
421 421 bool ParallelCompactData::initialize_region_data(size_t region_size)
422 422 {
423 423 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
424 424 _region_vspace = create_vspace(count, sizeof(RegionData));
425 425 if (_region_vspace != 0) {
426 426 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
427 427 _region_count = count;
428 428 return true;
429 429 }
430 430 return false;
431 431 }
432 432
433 433 void ParallelCompactData::clear()
434 434 {
435 435 memset(_region_data, 0, _region_vspace->committed_size());
436 436 }
437 437
438 438 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
439 439 assert(beg_region <= _region_count, "beg_region out of range");
440 440 assert(end_region <= _region_count, "end_region out of range");
441 441
442 442 const size_t region_cnt = end_region - beg_region;
443 443 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
444 444 }
445 445
446 446 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
447 447 {
448 448 const RegionData* cur_cp = region(region_idx);
449 449 const RegionData* const end_cp = region(region_count() - 1);
450 450
451 451 HeapWord* result = region_to_addr(region_idx);
452 452 if (cur_cp < end_cp) {
453 453 do {
454 454 result += cur_cp->partial_obj_size();
455 455 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
456 456 }
457 457 return result;
458 458 }
459 459
460 460 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
461 461 {
462 462 const size_t obj_ofs = pointer_delta(addr, _region_start);
463 463 const size_t beg_region = obj_ofs >> Log2RegionSize;
464 464 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
465 465
466 466 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
467 467 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
468 468
469 469 if (beg_region == end_region) {
470 470 // All in one region.
471 471 _region_data[beg_region].add_live_obj(len);
472 472 return;
473 473 }
474 474
475 475 // First region.
476 476 const size_t beg_ofs = region_offset(addr);
477 477 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
478 478
479 479 klassOop klass = ((oop)addr)->klass();
480 480 // Middle regions--completely spanned by this object.
481 481 for (size_t region = beg_region + 1; region < end_region; ++region) {
482 482 _region_data[region].set_partial_obj_size(RegionSize);
483 483 _region_data[region].set_partial_obj_addr(addr);
484 484 }
485 485
486 486 // Last region.
487 487 const size_t end_ofs = region_offset(addr + len - 1);
488 488 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
489 489 _region_data[end_region].set_partial_obj_addr(addr);
490 490 }
491 491
492 492 void
493 493 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
494 494 {
495 495 assert(region_offset(beg) == 0, "not RegionSize aligned");
496 496 assert(region_offset(end) == 0, "not RegionSize aligned");
497 497
498 498 size_t cur_region = addr_to_region_idx(beg);
499 499 const size_t end_region = addr_to_region_idx(end);
500 500 HeapWord* addr = beg;
501 501 while (cur_region < end_region) {
502 502 _region_data[cur_region].set_destination(addr);
503 503 _region_data[cur_region].set_destination_count(0);
504 504 _region_data[cur_region].set_source_region(cur_region);
505 505 _region_data[cur_region].set_data_location(addr);
506 506
507 507 // Update live_obj_size so the region appears completely full.
508 508 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
509 509 _region_data[cur_region].set_live_obj_size(live_size);
510 510
511 511 ++cur_region;
512 512 addr += RegionSize;
513 513 }
514 514 }
515 515
516 516 // Find the point at which a space can be split and, if necessary, record the
517 517 // split point.
518 518 //
519 519 // If the current src region (which overflowed the destination space) doesn't
520 520 // have a partial object, the split point is at the beginning of the current src
521 521 // region (an "easy" split, no extra bookkeeping required).
522 522 //
523 523 // If the current src region has a partial object, the split point is in the
524 524 // region where that partial object starts (call it the split_region). If
525 525 // split_region has a partial object, then the split point is just after that
526 526 // partial object (a "hard" split where we have to record the split data and
527 527 // zero the partial_obj_size field). With a "hard" split, we know that the
528 528 // partial_obj ends within split_region because the partial object that caused
529 529 // the overflow starts in split_region. If split_region doesn't have a partial
530 530 // obj, then the split is at the beginning of split_region (another "easy"
531 531 // split).
532 532 HeapWord*
533 533 ParallelCompactData::summarize_split_space(size_t src_region,
534 534 SplitInfo& split_info,
535 535 HeapWord* destination,
536 536 HeapWord* target_end,
537 537 HeapWord** target_next)
538 538 {
539 539 assert(destination <= target_end, "sanity");
540 540 assert(destination + _region_data[src_region].data_size() > target_end,
541 541 "region should not fit into target space");
542 542 assert(is_region_aligned(target_end), "sanity");
543 543
544 544 size_t split_region = src_region;
545 545 HeapWord* split_destination = destination;
546 546 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
547 547
548 548 if (destination + partial_obj_size > target_end) {
549 549 // The split point is just after the partial object (if any) in the
550 550 // src_region that contains the start of the object that overflowed the
551 551 // destination space.
552 552 //
553 553 // Find the start of the "overflow" object and set split_region to the
554 554 // region containing it.
555 555 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
556 556 split_region = addr_to_region_idx(overflow_obj);
557 557
558 558 // Clear the source_region field of all destination regions whose first word
559 559 // came from data after the split point (a non-null source_region field
560 560 // implies a region must be filled).
561 561 //
562 562 // An alternative to the simple loop below: clear during post_compact(),
563 563 // which uses memcpy instead of individual stores, and is easy to
564 564 // parallelize. (The downside is that it clears the entire RegionData
565 565 // object as opposed to just one field.)
566 566 //
567 567 // post_compact() would have to clear the summary data up to the highest
568 568 // address that was written during the summary phase, which would be
569 569 //
570 570 // max(top, max(new_top, clear_top))
571 571 //
572 572 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
573 573 // to target_end.
574 574 const RegionData* const sr = region(split_region);
575 575 const size_t beg_idx =
576 576 addr_to_region_idx(region_align_up(sr->destination() +
577 577 sr->partial_obj_size()));
578 578 const size_t end_idx = addr_to_region_idx(target_end);
579 579
580 580 if (TraceParallelOldGCSummaryPhase) {
581 581 gclog_or_tty->print_cr("split: clearing source_region field in ["
582 582 SIZE_FORMAT ", " SIZE_FORMAT ")",
583 583 beg_idx, end_idx);
584 584 }
585 585 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
586 586 _region_data[idx].set_source_region(0);
587 587 }
588 588
589 589 // Set split_destination and partial_obj_size to reflect the split region.
590 590 split_destination = sr->destination();
591 591 partial_obj_size = sr->partial_obj_size();
592 592 }
593 593
594 594 // The split is recorded only if a partial object extends onto the region.
595 595 if (partial_obj_size != 0) {
596 596 _region_data[split_region].set_partial_obj_size(0);
597 597 split_info.record(split_region, partial_obj_size, split_destination);
598 598 }
599 599
600 600 // Setup the continuation addresses.
601 601 *target_next = split_destination + partial_obj_size;
602 602 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
603 603
604 604 if (TraceParallelOldGCSummaryPhase) {
605 605 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
606 606 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
607 607 " pos=" SIZE_FORMAT,
608 608 split_type, source_next, split_region,
609 609 partial_obj_size);
610 610 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
611 611 " tn=" PTR_FORMAT,
612 612 split_type, split_destination,
613 613 addr_to_region_idx(split_destination),
614 614 *target_next);
615 615
616 616 if (partial_obj_size != 0) {
617 617 HeapWord* const po_beg = split_info.destination();
618 618 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
619 619 gclog_or_tty->print_cr("%s split: "
620 620 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
621 621 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
622 622 split_type,
623 623 po_beg, addr_to_region_idx(po_beg),
624 624 po_end, addr_to_region_idx(po_end));
625 625 }
626 626 }
627 627
628 628 return source_next;
629 629 }
630 630
631 631 bool ParallelCompactData::summarize(SplitInfo& split_info,
632 632 HeapWord* source_beg, HeapWord* source_end,
633 633 HeapWord** source_next,
634 634 HeapWord* target_beg, HeapWord* target_end,
635 635 HeapWord** target_next)
636 636 {
637 637 if (TraceParallelOldGCSummaryPhase) {
638 638 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
639 639 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
640 640 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
641 641 source_beg, source_end, source_next_val,
642 642 target_beg, target_end, *target_next);
643 643 }
644 644
645 645 size_t cur_region = addr_to_region_idx(source_beg);
646 646 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
647 647
648 648 HeapWord *dest_addr = target_beg;
649 649 while (cur_region < end_region) {
650 650 // The destination must be set even if the region has no data.
651 651 _region_data[cur_region].set_destination(dest_addr);
652 652
653 653 size_t words = _region_data[cur_region].data_size();
654 654 if (words > 0) {
655 655 // If cur_region does not fit entirely into the target space, find a point
656 656 // at which the source space can be 'split' so that part is copied to the
657 657 // target space and the rest is copied elsewhere.
658 658 if (dest_addr + words > target_end) {
659 659 assert(source_next != NULL, "source_next is NULL when splitting");
660 660 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
661 661 target_end, target_next);
662 662 return false;
663 663 }
664 664
665 665 // Compute the destination_count for cur_region, and if necessary, update
666 666 // source_region for a destination region. The source_region field is
667 667 // updated if cur_region is the first (left-most) region to be copied to a
668 668 // destination region.
669 669 //
670 670 // The destination_count calculation is a bit subtle. A region that has
671 671 // data that compacts into itself does not count itself as a destination.
672 672 // This maintains the invariant that a zero count means the region is
673 673 // available and can be claimed and then filled.
674 674 uint destination_count = 0;
675 675 if (split_info.is_split(cur_region)) {
676 676 // The current region has been split: the partial object will be copied
677 677 // to one destination space and the remaining data will be copied to
678 678 // another destination space. Adjust the initial destination_count and,
679 679 // if necessary, set the source_region field if the partial object will
680 680 // cross a destination region boundary.
681 681 destination_count = split_info.destination_count();
682 682 if (destination_count == 2) {
683 683 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
684 684 _region_data[dest_idx].set_source_region(cur_region);
685 685 }
686 686 }
687 687
688 688 HeapWord* const last_addr = dest_addr + words - 1;
689 689 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
690 690 const size_t dest_region_2 = addr_to_region_idx(last_addr);
691 691
692 692 // Initially assume that the destination regions will be the same and
693 693 // adjust the value below if necessary. Under this assumption, if
694 694 // cur_region == dest_region_2, then cur_region will be compacted
695 695 // completely into itself.
696 696 destination_count += cur_region == dest_region_2 ? 0 : 1;
697 697 if (dest_region_1 != dest_region_2) {
698 698 // Destination regions differ; adjust destination_count.
699 699 destination_count += 1;
700 700 // Data from cur_region will be copied to the start of dest_region_2.
701 701 _region_data[dest_region_2].set_source_region(cur_region);
702 702 } else if (region_offset(dest_addr) == 0) {
703 703 // Data from cur_region will be copied to the start of the destination
704 704 // region.
705 705 _region_data[dest_region_1].set_source_region(cur_region);
706 706 }
707 707
708 708 _region_data[cur_region].set_destination_count(destination_count);
709 709 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
710 710 dest_addr += words;
711 711 }
712 712
713 713 ++cur_region;
714 714 }
715 715
716 716 *target_next = dest_addr;
717 717 return true;
718 718 }
719 719
720 720 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
721 721 assert(addr != NULL, "Should detect NULL oop earlier");
722 722 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
723 723 #ifdef ASSERT
724 724 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
725 725 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
726 726 }
727 727 #endif
728 728 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
729 729
730 730 // Region covering the object.
731 731 size_t region_index = addr_to_region_idx(addr);
732 732 const RegionData* const region_ptr = region(region_index);
733 733 HeapWord* const region_addr = region_align_down(addr);
734 734
735 735 assert(addr < region_addr + RegionSize, "Region does not cover object");
736 736 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
737 737
738 738 HeapWord* result = region_ptr->destination();
739 739
740 740 // If all the data in the region is live, then the new location of the object
741 741 // can be calculated from the destination of the region plus the offset of the
742 742 // object in the region.
743 743 if (region_ptr->data_size() == RegionSize) {
744 744 result += pointer_delta(addr, region_addr);
745 745 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
746 746 return result;
747 747 }
748 748
749 749 // The new location of the object is
750 750 // region destination +
751 751 // size of the partial object extending onto the region +
752 752 // sizes of the live objects in the Region that are to the left of addr
753 753 const size_t partial_obj_size = region_ptr->partial_obj_size();
754 754 HeapWord* const search_start = region_addr + partial_obj_size;
755 755
756 756 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
757 757 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
758 758
759 759 result += partial_obj_size + live_to_left;
760 760 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
761 761 return result;
762 762 }
763 763
764 764 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
765 765 klassOop updated_klass;
766 766 if (PSParallelCompact::should_update_klass(old_klass)) {
767 767 updated_klass = (klassOop) calc_new_pointer(old_klass);
768 768 } else {
769 769 updated_klass = old_klass;
770 770 }
771 771
772 772 return updated_klass;
773 773 }
774 774
775 775 #ifdef ASSERT
776 776 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
777 777 {
778 778 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
779 779 const size_t* const end = (const size_t*)vspace->committed_high_addr();
780 780 for (const size_t* p = beg; p < end; ++p) {
781 781 assert(*p == 0, "not zero");
782 782 }
783 783 }
784 784
785 785 void ParallelCompactData::verify_clear()
786 786 {
787 787 verify_clear(_region_vspace);
788 788 }
789 789 #endif // #ifdef ASSERT
790 790
791 791 #ifdef NOT_PRODUCT
792 792 ParallelCompactData::RegionData* debug_region(size_t region_index) {
793 793 ParallelCompactData& sd = PSParallelCompact::summary_data();
794 794 return sd.region(region_index);
795 795 }
796 796 #endif
797 797
798 798 elapsedTimer PSParallelCompact::_accumulated_time;
799 799 unsigned int PSParallelCompact::_total_invocations = 0;
800 800 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
801 801 jlong PSParallelCompact::_time_of_last_gc = 0;
802 802 CollectorCounters* PSParallelCompact::_counters = NULL;
803 803 ParMarkBitMap PSParallelCompact::_mark_bitmap;
804 804 ParallelCompactData PSParallelCompact::_summary_data;
805 805
806 806 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
807 807
808 808 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
809 809 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
810 810
811 811 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
812 812 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
813 813
814 814 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
815 815 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
816 816
817 817 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); }
818 818 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
819 819
820 820 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
821 821
822 822 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); }
823 823 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
824 824
825 825 void PSParallelCompact::post_initialize() {
826 826 ParallelScavengeHeap* heap = gc_heap();
827 827 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
828 828
829 829 MemRegion mr = heap->reserved_region();
830 830 _ref_processor =
831 831 new ReferenceProcessor(mr, // span
832 832 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
833 833 (int) ParallelGCThreads, // mt processing degree
834 834 true, // mt discovery
835 835 (int) ParallelGCThreads, // mt discovery degree
836 836 true, // atomic_discovery
837 837 &_is_alive_closure, // non-header is alive closure
838 838 false); // write barrier for next field updates
839 839 _counters = new CollectorCounters("PSParallelCompact", 1);
840 840
841 841 // Initialize static fields in ParCompactionManager.
842 842 ParCompactionManager::initialize(mark_bitmap());
843 843 }
844 844
845 845 bool PSParallelCompact::initialize() {
846 846 ParallelScavengeHeap* heap = gc_heap();
847 847 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
848 848 MemRegion mr = heap->reserved_region();
849 849
850 850 // Was the old gen get allocated successfully?
851 851 if (!heap->old_gen()->is_allocated()) {
852 852 return false;
853 853 }
854 854
855 855 initialize_space_info();
856 856 initialize_dead_wood_limiter();
857 857
858 858 if (!_mark_bitmap.initialize(mr)) {
859 859 vm_shutdown_during_initialization("Unable to allocate bit map for "
860 860 "parallel garbage collection for the requested heap size.");
861 861 return false;
862 862 }
863 863
864 864 if (!_summary_data.initialize(mr)) {
865 865 vm_shutdown_during_initialization("Unable to allocate tables for "
866 866 "parallel garbage collection for the requested heap size.");
867 867 return false;
868 868 }
869 869
870 870 return true;
871 871 }
872 872
873 873 void PSParallelCompact::initialize_space_info()
874 874 {
875 875 memset(&_space_info, 0, sizeof(_space_info));
876 876
877 877 ParallelScavengeHeap* heap = gc_heap();
878 878 PSYoungGen* young_gen = heap->young_gen();
879 879 MutableSpace* perm_space = heap->perm_gen()->object_space();
880 880
881 881 _space_info[perm_space_id].set_space(perm_space);
882 882 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
883 883 _space_info[eden_space_id].set_space(young_gen->eden_space());
884 884 _space_info[from_space_id].set_space(young_gen->from_space());
885 885 _space_info[to_space_id].set_space(young_gen->to_space());
886 886
887 887 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
888 888 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
889 889
890 890 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
891 891 if (TraceParallelOldGCDensePrefix) {
892 892 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
893 893 _space_info[perm_space_id].min_dense_prefix());
894 894 }
895 895 }
896 896
897 897 void PSParallelCompact::initialize_dead_wood_limiter()
898 898 {
899 899 const size_t max = 100;
900 900 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
901 901 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
902 902 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
903 903 DEBUG_ONLY(_dwl_initialized = true;)
904 904 _dwl_adjustment = normal_distribution(1.0);
905 905 }
906 906
907 907 // Simple class for storing info about the heap at the start of GC, to be used
908 908 // after GC for comparison/printing.
909 909 class PreGCValues {
910 910 public:
911 911 PreGCValues() { }
912 912 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
913 913
914 914 void fill(ParallelScavengeHeap* heap) {
915 915 _heap_used = heap->used();
916 916 _young_gen_used = heap->young_gen()->used_in_bytes();
917 917 _old_gen_used = heap->old_gen()->used_in_bytes();
918 918 _perm_gen_used = heap->perm_gen()->used_in_bytes();
919 919 };
920 920
921 921 size_t heap_used() const { return _heap_used; }
922 922 size_t young_gen_used() const { return _young_gen_used; }
923 923 size_t old_gen_used() const { return _old_gen_used; }
924 924 size_t perm_gen_used() const { return _perm_gen_used; }
925 925
926 926 private:
927 927 size_t _heap_used;
928 928 size_t _young_gen_used;
929 929 size_t _old_gen_used;
930 930 size_t _perm_gen_used;
931 931 };
932 932
933 933 void
934 934 PSParallelCompact::clear_data_covering_space(SpaceId id)
935 935 {
936 936 // At this point, top is the value before GC, new_top() is the value that will
937 937 // be set at the end of GC. The marking bitmap is cleared to top; nothing
938 938 // should be marked above top. The summary data is cleared to the larger of
939 939 // top & new_top.
940 940 MutableSpace* const space = _space_info[id].space();
941 941 HeapWord* const bot = space->bottom();
942 942 HeapWord* const top = space->top();
943 943 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
944 944
945 945 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
946 946 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
947 947 _mark_bitmap.clear_range(beg_bit, end_bit);
948 948
949 949 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
950 950 const size_t end_region =
951 951 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
952 952 _summary_data.clear_range(beg_region, end_region);
953 953
954 954 // Clear the data used to 'split' regions.
955 955 SplitInfo& split_info = _space_info[id].split_info();
956 956 if (split_info.is_valid()) {
957 957 split_info.clear();
958 958 }
959 959 DEBUG_ONLY(split_info.verify_clear();)
960 960 }
961 961
962 962 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
963 963 {
964 964 // Update the from & to space pointers in space_info, since they are swapped
965 965 // at each young gen gc. Do the update unconditionally (even though a
966 966 // promotion failure does not swap spaces) because an unknown number of minor
967 967 // collections will have swapped the spaces an unknown number of times.
968 968 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
969 969 ParallelScavengeHeap* heap = gc_heap();
970 970 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
971 971 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
972 972
973 973 pre_gc_values->fill(heap);
974 974
975 975 ParCompactionManager::reset();
976 976 NOT_PRODUCT(_mark_bitmap.reset_counters());
977 977 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
978 978 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
979 979
980 980 // Increment the invocation count
981 981 heap->increment_total_collections(true);
982 982
983 983 // We need to track unique mark sweep invocations as well.
984 984 _total_invocations++;
985 985
986 986 if (PrintHeapAtGC) {
987 987 Universe::print_heap_before_gc();
988 988 }
989 989
990 990 // Fill in TLABs
991 991 heap->accumulate_statistics_all_tlabs();
992 992 heap->ensure_parsability(true); // retire TLABs
993 993
994 994 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
995 995 HandleMark hm; // Discard invalid handles created during verification
996 996 gclog_or_tty->print(" VerifyBeforeGC:");
997 997 Universe::verify(true);
998 998 }
999 999
1000 1000 // Verify object start arrays
1001 1001 if (VerifyObjectStartArray &&
1002 1002 VerifyBeforeGC) {
1003 1003 heap->old_gen()->verify_object_start_array();
1004 1004 heap->perm_gen()->verify_object_start_array();
1005 1005 }
1006 1006
1007 1007 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1008 1008 DEBUG_ONLY(summary_data().verify_clear();)
1009 1009
1010 1010 // Have worker threads release resources the next time they run a task.
1011 1011 gc_task_manager()->release_all_resources();
1012 1012 }
1013 1013
1014 1014 void PSParallelCompact::post_compact()
1015 1015 {
1016 1016 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
1017 1017
1018 1018 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1019 1019 // Clear the marking bitmap, summary data and split info.
1020 1020 clear_data_covering_space(SpaceId(id));
1021 1021 // Update top(). Must be done after clearing the bitmap and summary data.
1022 1022 _space_info[id].publish_new_top();
1023 1023 }
1024 1024
1025 1025 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1026 1026 MutableSpace* const from_space = _space_info[from_space_id].space();
1027 1027 MutableSpace* const to_space = _space_info[to_space_id].space();
1028 1028
1029 1029 ParallelScavengeHeap* heap = gc_heap();
1030 1030 bool eden_empty = eden_space->is_empty();
1031 1031 if (!eden_empty) {
1032 1032 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1033 1033 heap->young_gen(), heap->old_gen());
1034 1034 }
1035 1035
1036 1036 // Update heap occupancy information which is used as input to the soft ref
1037 1037 // clearing policy at the next gc.
1038 1038 Universe::update_heap_info_at_gc();
1039 1039
1040 1040 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1041 1041 to_space->is_empty();
1042 1042
1043 1043 BarrierSet* bs = heap->barrier_set();
1044 1044 if (bs->is_a(BarrierSet::ModRef)) {
1045 1045 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1046 1046 MemRegion old_mr = heap->old_gen()->reserved();
1047 1047 MemRegion perm_mr = heap->perm_gen()->reserved();
1048 1048 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1049 1049
1050 1050 if (young_gen_empty) {
1051 1051 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1052 1052 } else {
1053 1053 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1054 1054 }
1055 1055 }
1056 1056
1057 1057 Threads::gc_epilogue();
1058 1058 CodeCache::gc_epilogue();
1059 1059 JvmtiExport::gc_epilogue();
1060 1060
1061 1061 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1062 1062
1063 1063 ref_processor()->enqueue_discovered_references(NULL);
1064 1064
1065 1065 if (ZapUnusedHeapArea) {
1066 1066 heap->gen_mangle_unused_area();
1067 1067 }
1068 1068
1069 1069 // Update time of last GC
1070 1070 reset_millis_since_last_gc();
1071 1071 }
1072 1072
1073 1073 HeapWord*
1074 1074 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1075 1075 bool maximum_compaction)
1076 1076 {
1077 1077 const size_t region_size = ParallelCompactData::RegionSize;
1078 1078 const ParallelCompactData& sd = summary_data();
1079 1079
1080 1080 const MutableSpace* const space = _space_info[id].space();
1081 1081 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1082 1082 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1083 1083 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1084 1084
1085 1085 // Skip full regions at the beginning of the space--they are necessarily part
1086 1086 // of the dense prefix.
1087 1087 size_t full_count = 0;
1088 1088 const RegionData* cp;
1089 1089 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1090 1090 ++full_count;
1091 1091 }
1092 1092
1093 1093 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1094 1094 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1095 1095 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1096 1096 if (maximum_compaction || cp == end_cp || interval_ended) {
1097 1097 _maximum_compaction_gc_num = total_invocations();
1098 1098 return sd.region_to_addr(cp);
1099 1099 }
1100 1100
1101 1101 HeapWord* const new_top = _space_info[id].new_top();
1102 1102 const size_t space_live = pointer_delta(new_top, space->bottom());
1103 1103 const size_t space_used = space->used_in_words();
1104 1104 const size_t space_capacity = space->capacity_in_words();
1105 1105
1106 1106 const double cur_density = double(space_live) / space_capacity;
1107 1107 const double deadwood_density =
1108 1108 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1109 1109 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1110 1110
1111 1111 if (TraceParallelOldGCDensePrefix) {
1112 1112 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1113 1113 cur_density, deadwood_density, deadwood_goal);
1114 1114 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1115 1115 "space_cap=" SIZE_FORMAT,
1116 1116 space_live, space_used,
1117 1117 space_capacity);
1118 1118 }
1119 1119
1120 1120 // XXX - Use binary search?
1121 1121 HeapWord* dense_prefix = sd.region_to_addr(cp);
1122 1122 const RegionData* full_cp = cp;
1123 1123 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1124 1124 while (cp < end_cp) {
1125 1125 HeapWord* region_destination = cp->destination();
1126 1126 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1127 1127 if (TraceParallelOldGCDensePrefix && Verbose) {
1128 1128 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1129 1129 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1130 1130 sd.region(cp), region_destination,
1131 1131 dense_prefix, cur_deadwood);
1132 1132 }
1133 1133
1134 1134 if (cur_deadwood >= deadwood_goal) {
1135 1135 // Found the region that has the correct amount of deadwood to the left.
1136 1136 // This typically occurs after crossing a fairly sparse set of regions, so
1137 1137 // iterate backwards over those sparse regions, looking for the region
1138 1138 // that has the lowest density of live objects 'to the right.'
1139 1139 size_t space_to_left = sd.region(cp) * region_size;
1140 1140 size_t live_to_left = space_to_left - cur_deadwood;
1141 1141 size_t space_to_right = space_capacity - space_to_left;
1142 1142 size_t live_to_right = space_live - live_to_left;
1143 1143 double density_to_right = double(live_to_right) / space_to_right;
1144 1144 while (cp > full_cp) {
1145 1145 --cp;
1146 1146 const size_t prev_region_live_to_right = live_to_right -
1147 1147 cp->data_size();
1148 1148 const size_t prev_region_space_to_right = space_to_right + region_size;
1149 1149 double prev_region_density_to_right =
1150 1150 double(prev_region_live_to_right) / prev_region_space_to_right;
1151 1151 if (density_to_right <= prev_region_density_to_right) {
1152 1152 return dense_prefix;
1153 1153 }
1154 1154 if (TraceParallelOldGCDensePrefix && Verbose) {
1155 1155 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1156 1156 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1157 1157 prev_region_density_to_right);
1158 1158 }
1159 1159 dense_prefix -= region_size;
1160 1160 live_to_right = prev_region_live_to_right;
1161 1161 space_to_right = prev_region_space_to_right;
1162 1162 density_to_right = prev_region_density_to_right;
1163 1163 }
1164 1164 return dense_prefix;
1165 1165 }
1166 1166
1167 1167 dense_prefix += region_size;
1168 1168 ++cp;
1169 1169 }
1170 1170
1171 1171 return dense_prefix;
1172 1172 }
1173 1173
1174 1174 #ifndef PRODUCT
1175 1175 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1176 1176 const SpaceId id,
1177 1177 const bool maximum_compaction,
1178 1178 HeapWord* const addr)
1179 1179 {
1180 1180 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1181 1181 RegionData* const cp = summary_data().region(region_idx);
1182 1182 const MutableSpace* const space = _space_info[id].space();
1183 1183 HeapWord* const new_top = _space_info[id].new_top();
1184 1184
1185 1185 const size_t space_live = pointer_delta(new_top, space->bottom());
1186 1186 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1187 1187 const size_t space_cap = space->capacity_in_words();
1188 1188 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1189 1189 const size_t live_to_right = new_top - cp->destination();
1190 1190 const size_t dead_to_right = space->top() - addr - live_to_right;
1191 1191
1192 1192 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1193 1193 "spl=" SIZE_FORMAT " "
1194 1194 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1195 1195 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1196 1196 " ratio=%10.8f",
1197 1197 algorithm, addr, region_idx,
1198 1198 space_live,
1199 1199 dead_to_left, dead_to_left_pct,
1200 1200 dead_to_right, live_to_right,
1201 1201 double(dead_to_right) / live_to_right);
1202 1202 }
1203 1203 #endif // #ifndef PRODUCT
1204 1204
1205 1205 // Return a fraction indicating how much of the generation can be treated as
1206 1206 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1207 1207 // based on the density of live objects in the generation to determine a limit,
1208 1208 // which is then adjusted so the return value is min_percent when the density is
1209 1209 // 1.
1210 1210 //
1211 1211 // The following table shows some return values for a different values of the
1212 1212 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1213 1213 // min_percent is 1.
1214 1214 //
1215 1215 // fraction allowed as dead wood
1216 1216 // -----------------------------------------------------------------
1217 1217 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1218 1218 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1219 1219 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1220 1220 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1221 1221 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1222 1222 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1223 1223 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1224 1224 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1225 1225 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1226 1226 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1227 1227 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1228 1228 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1229 1229 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1230 1230 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1231 1231 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1232 1232 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1233 1233 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1234 1234 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1235 1235 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1236 1236 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1237 1237 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1238 1238 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1239 1239 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1240 1240
1241 1241 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1242 1242 {
1243 1243 assert(_dwl_initialized, "uninitialized");
1244 1244
1245 1245 // The raw limit is the value of the normal distribution at x = density.
1246 1246 const double raw_limit = normal_distribution(density);
1247 1247
1248 1248 // Adjust the raw limit so it becomes the minimum when the density is 1.
1249 1249 //
1250 1250 // First subtract the adjustment value (which is simply the precomputed value
1251 1251 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1252 1252 // Then add the minimum value, so the minimum is returned when the density is
1253 1253 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1254 1254 const double min = double(min_percent) / 100.0;
1255 1255 const double limit = raw_limit - _dwl_adjustment + min;
1256 1256 return MAX2(limit, 0.0);
1257 1257 }
1258 1258
1259 1259 ParallelCompactData::RegionData*
1260 1260 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1261 1261 const RegionData* end)
1262 1262 {
1263 1263 const size_t region_size = ParallelCompactData::RegionSize;
1264 1264 ParallelCompactData& sd = summary_data();
1265 1265 size_t left = sd.region(beg);
1266 1266 size_t right = end > beg ? sd.region(end) - 1 : left;
1267 1267
1268 1268 // Binary search.
1269 1269 while (left < right) {
1270 1270 // Equivalent to (left + right) / 2, but does not overflow.
1271 1271 const size_t middle = left + (right - left) / 2;
1272 1272 RegionData* const middle_ptr = sd.region(middle);
1273 1273 HeapWord* const dest = middle_ptr->destination();
1274 1274 HeapWord* const addr = sd.region_to_addr(middle);
1275 1275 assert(dest != NULL, "sanity");
1276 1276 assert(dest <= addr, "must move left");
1277 1277
1278 1278 if (middle > left && dest < addr) {
1279 1279 right = middle - 1;
1280 1280 } else if (middle < right && middle_ptr->data_size() == region_size) {
1281 1281 left = middle + 1;
1282 1282 } else {
1283 1283 return middle_ptr;
1284 1284 }
1285 1285 }
1286 1286 return sd.region(left);
1287 1287 }
1288 1288
1289 1289 ParallelCompactData::RegionData*
1290 1290 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1291 1291 const RegionData* end,
1292 1292 size_t dead_words)
1293 1293 {
1294 1294 ParallelCompactData& sd = summary_data();
1295 1295 size_t left = sd.region(beg);
1296 1296 size_t right = end > beg ? sd.region(end) - 1 : left;
1297 1297
1298 1298 // Binary search.
1299 1299 while (left < right) {
1300 1300 // Equivalent to (left + right) / 2, but does not overflow.
1301 1301 const size_t middle = left + (right - left) / 2;
1302 1302 RegionData* const middle_ptr = sd.region(middle);
1303 1303 HeapWord* const dest = middle_ptr->destination();
1304 1304 HeapWord* const addr = sd.region_to_addr(middle);
1305 1305 assert(dest != NULL, "sanity");
1306 1306 assert(dest <= addr, "must move left");
1307 1307
1308 1308 const size_t dead_to_left = pointer_delta(addr, dest);
1309 1309 if (middle > left && dead_to_left > dead_words) {
1310 1310 right = middle - 1;
1311 1311 } else if (middle < right && dead_to_left < dead_words) {
1312 1312 left = middle + 1;
1313 1313 } else {
1314 1314 return middle_ptr;
1315 1315 }
1316 1316 }
1317 1317 return sd.region(left);
1318 1318 }
1319 1319
1320 1320 // The result is valid during the summary phase, after the initial summarization
1321 1321 // of each space into itself, and before final summarization.
1322 1322 inline double
1323 1323 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1324 1324 HeapWord* const bottom,
1325 1325 HeapWord* const top,
1326 1326 HeapWord* const new_top)
1327 1327 {
1328 1328 ParallelCompactData& sd = summary_data();
1329 1329
1330 1330 assert(cp != NULL, "sanity");
1331 1331 assert(bottom != NULL, "sanity");
1332 1332 assert(top != NULL, "sanity");
1333 1333 assert(new_top != NULL, "sanity");
1334 1334 assert(top >= new_top, "summary data problem?");
1335 1335 assert(new_top > bottom, "space is empty; should not be here");
1336 1336 assert(new_top >= cp->destination(), "sanity");
1337 1337 assert(top >= sd.region_to_addr(cp), "sanity");
1338 1338
1339 1339 HeapWord* const destination = cp->destination();
1340 1340 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1341 1341 const size_t compacted_region_live = pointer_delta(new_top, destination);
1342 1342 const size_t compacted_region_used = pointer_delta(top,
1343 1343 sd.region_to_addr(cp));
1344 1344 const size_t reclaimable = compacted_region_used - compacted_region_live;
1345 1345
1346 1346 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1347 1347 return double(reclaimable) / divisor;
1348 1348 }
1349 1349
1350 1350 // Return the address of the end of the dense prefix, a.k.a. the start of the
1351 1351 // compacted region. The address is always on a region boundary.
1352 1352 //
1353 1353 // Completely full regions at the left are skipped, since no compaction can
1354 1354 // occur in those regions. Then the maximum amount of dead wood to allow is
1355 1355 // computed, based on the density (amount live / capacity) of the generation;
1356 1356 // the region with approximately that amount of dead space to the left is
1357 1357 // identified as the limit region. Regions between the last completely full
1358 1358 // region and the limit region are scanned and the one that has the best
1359 1359 // (maximum) reclaimed_ratio() is selected.
1360 1360 HeapWord*
1361 1361 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1362 1362 bool maximum_compaction)
1363 1363 {
1364 1364 if (ParallelOldGCSplitALot) {
1365 1365 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1366 1366 // The value was chosen to provoke splitting a young gen space; use it.
1367 1367 return _space_info[id].dense_prefix();
1368 1368 }
1369 1369 }
1370 1370
1371 1371 const size_t region_size = ParallelCompactData::RegionSize;
1372 1372 const ParallelCompactData& sd = summary_data();
1373 1373
1374 1374 const MutableSpace* const space = _space_info[id].space();
1375 1375 HeapWord* const top = space->top();
1376 1376 HeapWord* const top_aligned_up = sd.region_align_up(top);
1377 1377 HeapWord* const new_top = _space_info[id].new_top();
1378 1378 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1379 1379 HeapWord* const bottom = space->bottom();
1380 1380 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1381 1381 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1382 1382 const RegionData* const new_top_cp =
1383 1383 sd.addr_to_region_ptr(new_top_aligned_up);
1384 1384
1385 1385 // Skip full regions at the beginning of the space--they are necessarily part
1386 1386 // of the dense prefix.
1387 1387 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1388 1388 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1389 1389 space->is_empty(), "no dead space allowed to the left");
1390 1390 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1391 1391 "region must have dead space");
1392 1392
1393 1393 // The gc number is saved whenever a maximum compaction is done, and used to
1394 1394 // determine when the maximum compaction interval has expired. This avoids
1395 1395 // successive max compactions for different reasons.
1396 1396 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1397 1397 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1398 1398 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1399 1399 total_invocations() == HeapFirstMaximumCompactionCount;
1400 1400 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1401 1401 _maximum_compaction_gc_num = total_invocations();
1402 1402 return sd.region_to_addr(full_cp);
1403 1403 }
1404 1404
1405 1405 const size_t space_live = pointer_delta(new_top, bottom);
1406 1406 const size_t space_used = space->used_in_words();
1407 1407 const size_t space_capacity = space->capacity_in_words();
1408 1408
1409 1409 const double density = double(space_live) / double(space_capacity);
1410 1410 const size_t min_percent_free =
1411 1411 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1412 1412 const double limiter = dead_wood_limiter(density, min_percent_free);
1413 1413 const size_t dead_wood_max = space_used - space_live;
1414 1414 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1415 1415 dead_wood_max);
1416 1416
1417 1417 if (TraceParallelOldGCDensePrefix) {
1418 1418 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1419 1419 "space_cap=" SIZE_FORMAT,
1420 1420 space_live, space_used,
1421 1421 space_capacity);
1422 1422 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1423 1423 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1424 1424 density, min_percent_free, limiter,
1425 1425 dead_wood_max, dead_wood_limit);
1426 1426 }
1427 1427
1428 1428 // Locate the region with the desired amount of dead space to the left.
1429 1429 const RegionData* const limit_cp =
1430 1430 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1431 1431
1432 1432 // Scan from the first region with dead space to the limit region and find the
1433 1433 // one with the best (largest) reclaimed ratio.
1434 1434 double best_ratio = 0.0;
1435 1435 const RegionData* best_cp = full_cp;
1436 1436 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1437 1437 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1438 1438 if (tmp_ratio > best_ratio) {
1439 1439 best_cp = cp;
1440 1440 best_ratio = tmp_ratio;
1441 1441 }
1442 1442 }
1443 1443
1444 1444 #if 0
1445 1445 // Something to consider: if the region with the best ratio is 'close to' the
1446 1446 // first region w/free space, choose the first region with free space
1447 1447 // ("first-free"). The first-free region is usually near the start of the
1448 1448 // heap, which means we are copying most of the heap already, so copy a bit
1449 1449 // more to get complete compaction.
1450 1450 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1451 1451 _maximum_compaction_gc_num = total_invocations();
1452 1452 best_cp = full_cp;
1453 1453 }
1454 1454 #endif // #if 0
1455 1455
1456 1456 return sd.region_to_addr(best_cp);
1457 1457 }
1458 1458
1459 1459 #ifndef PRODUCT
1460 1460 void
1461 1461 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1462 1462 size_t words)
1463 1463 {
1464 1464 if (TraceParallelOldGCSummaryPhase) {
1465 1465 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1466 1466 SIZE_FORMAT, start, start + words, words);
1467 1467 }
1468 1468
1469 1469 ObjectStartArray* const start_array = _space_info[id].start_array();
1470 1470 CollectedHeap::fill_with_objects(start, words);
1471 1471 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1472 1472 _mark_bitmap.mark_obj(p, words);
1473 1473 _summary_data.add_obj(p, words);
1474 1474 start_array->allocate_block(p);
1475 1475 }
1476 1476 }
1477 1477
1478 1478 void
1479 1479 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1480 1480 {
1481 1481 ParallelCompactData& sd = summary_data();
1482 1482 MutableSpace* space = _space_info[id].space();
1483 1483
1484 1484 // Find the source and destination start addresses.
1485 1485 HeapWord* const src_addr = sd.region_align_down(start);
1486 1486 HeapWord* dst_addr;
1487 1487 if (src_addr < start) {
1488 1488 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1489 1489 } else if (src_addr > space->bottom()) {
1490 1490 // The start (the original top() value) is aligned to a region boundary so
1491 1491 // the associated region does not have a destination. Compute the
1492 1492 // destination from the previous region.
1493 1493 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1494 1494 dst_addr = cp->destination() + cp->data_size();
1495 1495 } else {
1496 1496 // Filling the entire space.
1497 1497 dst_addr = space->bottom();
1498 1498 }
1499 1499 assert(dst_addr != NULL, "sanity");
1500 1500
1501 1501 // Update the summary data.
1502 1502 bool result = _summary_data.summarize(_space_info[id].split_info(),
1503 1503 src_addr, space->top(), NULL,
1504 1504 dst_addr, space->end(),
1505 1505 _space_info[id].new_top_addr());
1506 1506 assert(result, "should not fail: bad filler object size");
1507 1507 }
1508 1508
1509 1509 void
1510 1510 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1511 1511 {
1512 1512 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1513 1513 return;
1514 1514 }
1515 1515
1516 1516 MutableSpace* const space = _space_info[id].space();
1517 1517 if (space->is_empty()) {
1518 1518 HeapWord* b = space->bottom();
1519 1519 HeapWord* t = b + space->capacity_in_words() / 2;
1520 1520 space->set_top(t);
1521 1521 if (ZapUnusedHeapArea) {
1522 1522 space->set_top_for_allocations();
1523 1523 }
1524 1524
1525 1525 size_t min_size = CollectedHeap::min_fill_size();
1526 1526 size_t obj_len = min_size;
1527 1527 while (b + obj_len <= t) {
1528 1528 CollectedHeap::fill_with_object(b, obj_len);
1529 1529 mark_bitmap()->mark_obj(b, obj_len);
1530 1530 summary_data().add_obj(b, obj_len);
1531 1531 b += obj_len;
1532 1532 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1533 1533 }
1534 1534 if (b < t) {
1535 1535 // The loop didn't completely fill to t (top); adjust top downward.
1536 1536 space->set_top(b);
1537 1537 if (ZapUnusedHeapArea) {
1538 1538 space->set_top_for_allocations();
1539 1539 }
1540 1540 }
1541 1541
1542 1542 HeapWord** nta = _space_info[id].new_top_addr();
1543 1543 bool result = summary_data().summarize(_space_info[id].split_info(),
1544 1544 space->bottom(), space->top(), NULL,
1545 1545 space->bottom(), space->end(), nta);
1546 1546 assert(result, "space must fit into itself");
1547 1547 }
1548 1548 }
1549 1549
1550 1550 void
1551 1551 PSParallelCompact::provoke_split(bool & max_compaction)
1552 1552 {
1553 1553 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1554 1554 return;
1555 1555 }
1556 1556
1557 1557 const size_t region_size = ParallelCompactData::RegionSize;
1558 1558 ParallelCompactData& sd = summary_data();
1559 1559
1560 1560 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1561 1561 MutableSpace* const from_space = _space_info[from_space_id].space();
1562 1562 const size_t eden_live = pointer_delta(eden_space->top(),
1563 1563 _space_info[eden_space_id].new_top());
1564 1564 const size_t from_live = pointer_delta(from_space->top(),
1565 1565 _space_info[from_space_id].new_top());
1566 1566
1567 1567 const size_t min_fill_size = CollectedHeap::min_fill_size();
1568 1568 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1569 1569 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1570 1570 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1571 1571 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1572 1572
1573 1573 // Choose the space to split; need at least 2 regions live (or fillable).
1574 1574 SpaceId id;
1575 1575 MutableSpace* space;
1576 1576 size_t live_words;
1577 1577 size_t fill_words;
1578 1578 if (eden_live + eden_fillable >= region_size * 2) {
1579 1579 id = eden_space_id;
1580 1580 space = eden_space;
1581 1581 live_words = eden_live;
1582 1582 fill_words = eden_fillable;
1583 1583 } else if (from_live + from_fillable >= region_size * 2) {
1584 1584 id = from_space_id;
1585 1585 space = from_space;
1586 1586 live_words = from_live;
1587 1587 fill_words = from_fillable;
1588 1588 } else {
1589 1589 return; // Give up.
1590 1590 }
1591 1591 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1592 1592
1593 1593 if (live_words < region_size * 2) {
1594 1594 // Fill from top() to end() w/live objects of mixed sizes.
1595 1595 HeapWord* const fill_start = space->top();
1596 1596 live_words += fill_words;
1597 1597
1598 1598 space->set_top(fill_start + fill_words);
1599 1599 if (ZapUnusedHeapArea) {
1600 1600 space->set_top_for_allocations();
1601 1601 }
1602 1602
1603 1603 HeapWord* cur_addr = fill_start;
1604 1604 while (fill_words > 0) {
1605 1605 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1606 1606 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1607 1607 if (fill_words - cur_size < min_fill_size) {
1608 1608 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1609 1609 }
1610 1610
1611 1611 CollectedHeap::fill_with_object(cur_addr, cur_size);
1612 1612 mark_bitmap()->mark_obj(cur_addr, cur_size);
1613 1613 sd.add_obj(cur_addr, cur_size);
1614 1614
1615 1615 cur_addr += cur_size;
1616 1616 fill_words -= cur_size;
1617 1617 }
1618 1618
1619 1619 summarize_new_objects(id, fill_start);
1620 1620 }
1621 1621
1622 1622 max_compaction = false;
1623 1623
1624 1624 // Manipulate the old gen so that it has room for about half of the live data
1625 1625 // in the target young gen space (live_words / 2).
1626 1626 id = old_space_id;
1627 1627 space = _space_info[id].space();
1628 1628 const size_t free_at_end = space->free_in_words();
1629 1629 const size_t free_target = align_object_size(live_words / 2);
1630 1630 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1631 1631
1632 1632 if (free_at_end >= free_target + min_fill_size) {
1633 1633 // Fill space above top() and set the dense prefix so everything survives.
1634 1634 HeapWord* const fill_start = space->top();
1635 1635 const size_t fill_size = free_at_end - free_target;
1636 1636 space->set_top(space->top() + fill_size);
1637 1637 if (ZapUnusedHeapArea) {
1638 1638 space->set_top_for_allocations();
1639 1639 }
1640 1640 fill_with_live_objects(id, fill_start, fill_size);
1641 1641 summarize_new_objects(id, fill_start);
1642 1642 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1643 1643 } else if (dead + free_at_end > free_target) {
1644 1644 // Find a dense prefix that makes the right amount of space available.
1645 1645 HeapWord* cur = sd.region_align_down(space->top());
1646 1646 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1647 1647 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1648 1648 while (dead_to_right < free_target) {
1649 1649 cur -= region_size;
1650 1650 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1651 1651 dead_to_right = pointer_delta(space->end(), cur_destination);
1652 1652 }
1653 1653 _space_info[id].set_dense_prefix(cur);
1654 1654 }
1655 1655 }
1656 1656 #endif // #ifndef PRODUCT
1657 1657
1658 1658 void PSParallelCompact::summarize_spaces_quick()
1659 1659 {
1660 1660 for (unsigned int i = 0; i < last_space_id; ++i) {
1661 1661 const MutableSpace* space = _space_info[i].space();
1662 1662 HeapWord** nta = _space_info[i].new_top_addr();
1663 1663 bool result = _summary_data.summarize(_space_info[i].split_info(),
1664 1664 space->bottom(), space->top(), NULL,
1665 1665 space->bottom(), space->end(), nta);
1666 1666 assert(result, "space must fit into itself");
1667 1667 _space_info[i].set_dense_prefix(space->bottom());
1668 1668 }
1669 1669
1670 1670 #ifndef PRODUCT
1671 1671 if (ParallelOldGCSplitALot) {
1672 1672 provoke_split_fill_survivor(to_space_id);
1673 1673 }
1674 1674 #endif // #ifndef PRODUCT
1675 1675 }
1676 1676
1677 1677 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1678 1678 {
1679 1679 HeapWord* const dense_prefix_end = dense_prefix(id);
1680 1680 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1681 1681 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1682 1682 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1683 1683 // Only enough dead space is filled so that any remaining dead space to the
1684 1684 // left is larger than the minimum filler object. (The remainder is filled
1685 1685 // during the copy/update phase.)
1686 1686 //
1687 1687 // The size of the dead space to the right of the boundary is not a
1688 1688 // concern, since compaction will be able to use whatever space is
1689 1689 // available.
1690 1690 //
1691 1691 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1692 1692 // surrounds the space to be filled with an object.
1693 1693 //
1694 1694 // In the 32-bit VM, each bit represents two 32-bit words:
1695 1695 // +---+
1696 1696 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1697 1697 // end_bits: ... x x x | 0 | || 0 x x ...
1698 1698 // +---+
1699 1699 //
1700 1700 // In the 64-bit VM, each bit represents one 64-bit word:
1701 1701 // +------------+
1702 1702 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1703 1703 // end_bits: ... x x 1 | 0 || 0 | x x ...
1704 1704 // +------------+
1705 1705 // +-------+
1706 1706 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1707 1707 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1708 1708 // +-------+
1709 1709 // +-----------+
1710 1710 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1711 1711 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1712 1712 // +-----------+
1713 1713 // +-------+
1714 1714 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1715 1715 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1716 1716 // +-------+
1717 1717
1718 1718 // Initially assume case a, c or e will apply.
1719 1719 size_t obj_len = CollectedHeap::min_fill_size();
1720 1720 HeapWord* obj_beg = dense_prefix_end - obj_len;
1721 1721
1722 1722 #ifdef _LP64
1723 1723 if (MinObjAlignment > 1) { // object alignment > heap word size
1724 1724 // Cases a, c or e.
1725 1725 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1726 1726 // Case b above.
1727 1727 obj_beg = dense_prefix_end - 1;
1728 1728 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1729 1729 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1730 1730 // Case d above.
1731 1731 obj_beg = dense_prefix_end - 3;
1732 1732 obj_len = 3;
1733 1733 }
1734 1734 #endif // #ifdef _LP64
1735 1735
1736 1736 CollectedHeap::fill_with_object(obj_beg, obj_len);
1737 1737 _mark_bitmap.mark_obj(obj_beg, obj_len);
1738 1738 _summary_data.add_obj(obj_beg, obj_len);
1739 1739 assert(start_array(id) != NULL, "sanity");
1740 1740 start_array(id)->allocate_block(obj_beg);
1741 1741 }
1742 1742 }
1743 1743
1744 1744 void
1745 1745 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1746 1746 {
1747 1747 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1748 1748 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1749 1749 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1750 1750 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1751 1751 cur->set_source_region(0);
1752 1752 }
1753 1753 }
1754 1754
1755 1755 void
1756 1756 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1757 1757 {
1758 1758 assert(id < last_space_id, "id out of range");
1759 1759 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1760 1760 ParallelOldGCSplitALot && id == old_space_id,
1761 1761 "should have been reset in summarize_spaces_quick()");
1762 1762
1763 1763 const MutableSpace* space = _space_info[id].space();
1764 1764 if (_space_info[id].new_top() != space->bottom()) {
1765 1765 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1766 1766 _space_info[id].set_dense_prefix(dense_prefix_end);
1767 1767
1768 1768 #ifndef PRODUCT
1769 1769 if (TraceParallelOldGCDensePrefix) {
1770 1770 print_dense_prefix_stats("ratio", id, maximum_compaction,
1771 1771 dense_prefix_end);
1772 1772 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1773 1773 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1774 1774 }
1775 1775 #endif // #ifndef PRODUCT
1776 1776
1777 1777 // Recompute the summary data, taking into account the dense prefix. If
1778 1778 // every last byte will be reclaimed, then the existing summary data which
1779 1779 // compacts everything can be left in place.
1780 1780 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1781 1781 // If dead space crosses the dense prefix boundary, it is (at least
1782 1782 // partially) filled with a dummy object, marked live and added to the
1783 1783 // summary data. This simplifies the copy/update phase and must be done
1784 1784 // before the final locations of objects are determined, to prevent
1785 1785 // leaving a fragment of dead space that is too small to fill.
1786 1786 fill_dense_prefix_end(id);
1787 1787
1788 1788 // Compute the destination of each Region, and thus each object.
1789 1789 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1790 1790 _summary_data.summarize(_space_info[id].split_info(),
1791 1791 dense_prefix_end, space->top(), NULL,
1792 1792 dense_prefix_end, space->end(),
1793 1793 _space_info[id].new_top_addr());
1794 1794 }
1795 1795 }
1796 1796
1797 1797 if (TraceParallelOldGCSummaryPhase) {
1798 1798 const size_t region_size = ParallelCompactData::RegionSize;
1799 1799 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1800 1800 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1801 1801 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1802 1802 HeapWord* const new_top = _space_info[id].new_top();
1803 1803 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1804 1804 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1805 1805 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1806 1806 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1807 1807 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1808 1808 id, space->capacity_in_words(), dense_prefix_end,
1809 1809 dp_region, dp_words / region_size,
1810 1810 cr_words / region_size, new_top);
1811 1811 }
1812 1812 }
1813 1813
1814 1814 #ifndef PRODUCT
1815 1815 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1816 1816 HeapWord* dst_beg, HeapWord* dst_end,
1817 1817 SpaceId src_space_id,
1818 1818 HeapWord* src_beg, HeapWord* src_end)
1819 1819 {
1820 1820 if (TraceParallelOldGCSummaryPhase) {
1821 1821 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1822 1822 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1823 1823 SIZE_FORMAT "-" SIZE_FORMAT " "
1824 1824 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1825 1825 SIZE_FORMAT "-" SIZE_FORMAT,
1826 1826 src_space_id, space_names[src_space_id],
1827 1827 dst_space_id, space_names[dst_space_id],
1828 1828 src_beg, src_end,
1829 1829 _summary_data.addr_to_region_idx(src_beg),
1830 1830 _summary_data.addr_to_region_idx(src_end),
1831 1831 dst_beg, dst_end,
1832 1832 _summary_data.addr_to_region_idx(dst_beg),
1833 1833 _summary_data.addr_to_region_idx(dst_end));
1834 1834 }
1835 1835 }
1836 1836 #endif // #ifndef PRODUCT
1837 1837
1838 1838 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1839 1839 bool maximum_compaction)
1840 1840 {
1841 1841 EventMark m("2 summarize");
1842 1842 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1843 1843 // trace("2");
1844 1844
1845 1845 #ifdef ASSERT
1846 1846 if (TraceParallelOldGCMarkingPhase) {
1847 1847 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1848 1848 "add_obj_bytes=" SIZE_FORMAT,
1849 1849 add_obj_count, add_obj_size * HeapWordSize);
1850 1850 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1851 1851 "mark_bitmap_bytes=" SIZE_FORMAT,
1852 1852 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1853 1853 }
1854 1854 #endif // #ifdef ASSERT
1855 1855
1856 1856 // Quick summarization of each space into itself, to see how much is live.
1857 1857 summarize_spaces_quick();
1858 1858
1859 1859 if (TraceParallelOldGCSummaryPhase) {
1860 1860 tty->print_cr("summary_phase: after summarizing each space to self");
1861 1861 Universe::print();
1862 1862 NOT_PRODUCT(print_region_ranges());
1863 1863 if (Verbose) {
1864 1864 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1865 1865 }
1866 1866 }
1867 1867
1868 1868 // The amount of live data that will end up in old space (assuming it fits).
1869 1869 size_t old_space_total_live = 0;
1870 1870 assert(perm_space_id < old_space_id, "should not count perm data here");
1871 1871 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1872 1872 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1873 1873 _space_info[id].space()->bottom());
1874 1874 }
1875 1875
1876 1876 MutableSpace* const old_space = _space_info[old_space_id].space();
1877 1877 const size_t old_capacity = old_space->capacity_in_words();
1878 1878 if (old_space_total_live > old_capacity) {
1879 1879 // XXX - should also try to expand
1880 1880 maximum_compaction = true;
1881 1881 }
1882 1882 #ifndef PRODUCT
1883 1883 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1884 1884 provoke_split(maximum_compaction);
1885 1885 }
1886 1886 #endif // #ifndef PRODUCT
1887 1887
1888 1888 // Permanent and Old generations.
1889 1889 summarize_space(perm_space_id, maximum_compaction);
1890 1890 summarize_space(old_space_id, maximum_compaction);
1891 1891
1892 1892 // Summarize the remaining spaces in the young gen. The initial target space
1893 1893 // is the old gen. If a space does not fit entirely into the target, then the
1894 1894 // remainder is compacted into the space itself and that space becomes the new
1895 1895 // target.
1896 1896 SpaceId dst_space_id = old_space_id;
1897 1897 HeapWord* dst_space_end = old_space->end();
1898 1898 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1899 1899 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1900 1900 const MutableSpace* space = _space_info[id].space();
1901 1901 const size_t live = pointer_delta(_space_info[id].new_top(),
1902 1902 space->bottom());
1903 1903 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1904 1904
1905 1905 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1906 1906 SpaceId(id), space->bottom(), space->top());)
1907 1907 if (live > 0 && live <= available) {
1908 1908 // All the live data will fit.
1909 1909 bool done = _summary_data.summarize(_space_info[id].split_info(),
1910 1910 space->bottom(), space->top(),
1911 1911 NULL,
1912 1912 *new_top_addr, dst_space_end,
1913 1913 new_top_addr);
1914 1914 assert(done, "space must fit into old gen");
1915 1915
1916 1916 // Reset the new_top value for the space.
1917 1917 _space_info[id].set_new_top(space->bottom());
1918 1918 } else if (live > 0) {
1919 1919 // Attempt to fit part of the source space into the target space.
1920 1920 HeapWord* next_src_addr = NULL;
1921 1921 bool done = _summary_data.summarize(_space_info[id].split_info(),
1922 1922 space->bottom(), space->top(),
1923 1923 &next_src_addr,
1924 1924 *new_top_addr, dst_space_end,
1925 1925 new_top_addr);
1926 1926 assert(!done, "space should not fit into old gen");
1927 1927 assert(next_src_addr != NULL, "sanity");
1928 1928
1929 1929 // The source space becomes the new target, so the remainder is compacted
1930 1930 // within the space itself.
1931 1931 dst_space_id = SpaceId(id);
1932 1932 dst_space_end = space->end();
1933 1933 new_top_addr = _space_info[id].new_top_addr();
1934 1934 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1935 1935 space->bottom(), dst_space_end,
1936 1936 SpaceId(id), next_src_addr, space->top());)
1937 1937 done = _summary_data.summarize(_space_info[id].split_info(),
1938 1938 next_src_addr, space->top(),
1939 1939 NULL,
1940 1940 space->bottom(), dst_space_end,
1941 1941 new_top_addr);
1942 1942 assert(done, "space must fit when compacted into itself");
1943 1943 assert(*new_top_addr <= space->top(), "usage should not grow");
1944 1944 }
1945 1945 }
1946 1946
1947 1947 if (TraceParallelOldGCSummaryPhase) {
1948 1948 tty->print_cr("summary_phase: after final summarization");
1949 1949 Universe::print();
1950 1950 NOT_PRODUCT(print_region_ranges());
1951 1951 if (Verbose) {
1952 1952 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1953 1953 }
1954 1954 }
1955 1955 }
1956 1956
1957 1957 // This method should contain all heap-specific policy for invoking a full
1958 1958 // collection. invoke_no_policy() will only attempt to compact the heap; it
1959 1959 // will do nothing further. If we need to bail out for policy reasons, scavenge
1960 1960 // before full gc, or any other specialized behavior, it needs to be added here.
1961 1961 //
1962 1962 // Note that this method should only be called from the vm_thread while at a
1963 1963 // safepoint.
1964 1964 //
1965 1965 // Note that the all_soft_refs_clear flag in the collector policy
1966 1966 // may be true because this method can be called without intervening
1967 1967 // activity. For example when the heap space is tight and full measure
1968 1968 // are being taken to free space.
1969 1969 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1970 1970 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1971 1971 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1972 1972 "should be in vm thread");
1973 1973
1974 1974 ParallelScavengeHeap* heap = gc_heap();
1975 1975 GCCause::Cause gc_cause = heap->gc_cause();
1976 1976 assert(!heap->is_gc_active(), "not reentrant");
1977 1977
1978 1978 PSAdaptiveSizePolicy* policy = heap->size_policy();
1979 1979 IsGCActiveMark mark;
1980 1980
1981 1981 if (ScavengeBeforeFullGC) {
1982 1982 PSScavenge::invoke_no_policy();
1983 1983 }
1984 1984
1985 1985 const bool clear_all_soft_refs =
1986 1986 heap->collector_policy()->should_clear_all_soft_refs();
1987 1987
1988 1988 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1989 1989 maximum_heap_compaction);
1990 1990 }
1991 1991
1992 1992 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1993 1993 size_t addr_region_index = addr_to_region_idx(addr);
1994 1994 return region_index == addr_region_index;
1995 1995 }
1996 1996
1997 1997 // This method contains no policy. You should probably
1998 1998 // be calling invoke() instead.
1999 1999 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
2000 2000 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
2001 2001 assert(ref_processor() != NULL, "Sanity");
2002 2002
2003 2003 if (GC_locker::check_active_before_gc()) {
2004 2004 return;
2005 2005 }
2006 2006
2007 2007 TimeStamp marking_start;
2008 2008 TimeStamp compaction_start;
2009 2009 TimeStamp collection_exit;
2010 2010
2011 2011 ParallelScavengeHeap* heap = gc_heap();
2012 2012 GCCause::Cause gc_cause = heap->gc_cause();
2013 2013 PSYoungGen* young_gen = heap->young_gen();
2014 2014 PSOldGen* old_gen = heap->old_gen();
2015 2015 PSPermGen* perm_gen = heap->perm_gen();
2016 2016 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2017 2017
2018 2018 // The scope of casr should end after code that can change
2019 2019 // CollectorPolicy::_should_clear_all_soft_refs.
2020 2020 ClearedAllSoftRefs casr(maximum_heap_compaction,
2021 2021 heap->collector_policy());
2022 2022
2023 2023 if (ZapUnusedHeapArea) {
2024 2024 // Save information needed to minimize mangling
2025 2025 heap->record_gen_tops_before_GC();
2026 2026 }
2027 2027
2028 2028 heap->pre_full_gc_dump();
2029 2029
2030 2030 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2031 2031
2032 2032 // Make sure data structures are sane, make the heap parsable, and do other
2033 2033 // miscellaneous bookkeeping.
2034 2034 PreGCValues pre_gc_values;
2035 2035 pre_compact(&pre_gc_values);
2036 2036
2037 2037 // Get the compaction manager reserved for the VM thread.
2038 2038 ParCompactionManager* const vmthread_cm =
2039 2039 ParCompactionManager::manager_array(gc_task_manager()->workers());
2040 2040
2041 2041 // Place after pre_compact() where the number of invocations is incremented.
2042 2042 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2043 2043
2044 2044 {
2045 2045 ResourceMark rm;
2046 2046 HandleMark hm;
2047 2047
2048 2048 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
2049 2049
2050 2050 // This is useful for debugging but don't change the output the
2051 2051 // the customer sees.
2052 2052 const char* gc_cause_str = "Full GC";
2053 2053 if (is_system_gc && PrintGCDetails) {
2054 2054 gc_cause_str = "Full GC (System)";
2055 2055 }
2056 2056 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2057 2057 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2058 2058 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
2059 2059 TraceCollectorStats tcs(counters());
2060 2060 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2061 2061
↓ open down ↓ |
2061 lines elided |
↑ open up ↑ |
2062 2062 if (TraceGen1Time) accumulated_time()->start();
2063 2063
2064 2064 // Let the size policy know we're starting
2065 2065 size_policy->major_collection_begin();
2066 2066
2067 2067 // When collecting the permanent generation methodOops may be moving,
2068 2068 // so we either have to flush all bcp data or convert it into bci.
2069 2069 CodeCache::gc_prologue();
2070 2070 Threads::gc_prologue();
2071 2071
2072 - NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2073 2072 COMPILER2_PRESENT(DerivedPointerTable::clear());
2074 2073
2075 - ref_processor()->enable_discovery();
2074 + ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2076 2075 ref_processor()->setup_policy(maximum_heap_compaction);
2077 2076
2078 2077 bool marked_for_unloading = false;
2079 2078
2080 2079 marking_start.update();
2081 2080 marking_phase(vmthread_cm, maximum_heap_compaction);
2082 2081
2083 2082 #ifndef PRODUCT
2084 2083 if (TraceParallelOldGCMarkingPhase) {
2085 2084 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2086 2085 "cas_by_another %d",
2087 2086 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2088 2087 mark_bitmap()->cas_by_another());
2089 2088 }
2090 2089 #endif // #ifndef PRODUCT
2091 2090
2092 2091 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
2093 2092 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2094 2093
2095 2094 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2096 2095 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2097 2096
2098 2097 // adjust_roots() updates Universe::_intArrayKlassObj which is
2099 2098 // needed by the compaction for filling holes in the dense prefix.
2100 2099 adjust_roots();
2101 2100
2102 2101 compaction_start.update();
2103 2102 // Does the perm gen always have to be done serially because
2104 2103 // klasses are used in the update of an object?
2105 2104 compact_perm(vmthread_cm);
2106 2105
2107 2106 compact();
2108 2107
2109 2108 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2110 2109 // done before resizing.
2111 2110 post_compact();
2112 2111
2113 2112 // Let the size policy know we're done
2114 2113 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2115 2114
2116 2115 if (UseAdaptiveSizePolicy) {
2117 2116 if (PrintAdaptiveSizePolicy) {
2118 2117 gclog_or_tty->print("AdaptiveSizeStart: ");
2119 2118 gclog_or_tty->stamp();
2120 2119 gclog_or_tty->print_cr(" collection: %d ",
2121 2120 heap->total_collections());
2122 2121 if (Verbose) {
2123 2122 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2124 2123 " perm_gen_capacity: %d ",
2125 2124 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2126 2125 perm_gen->capacity_in_bytes());
2127 2126 }
2128 2127 }
2129 2128
2130 2129 // Don't check if the size_policy is ready here. Let
2131 2130 // the size_policy check that internally.
2132 2131 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2133 2132 ((gc_cause != GCCause::_java_lang_system_gc) ||
2134 2133 UseAdaptiveSizePolicyWithSystemGC)) {
2135 2134 // Calculate optimal free space amounts
2136 2135 assert(young_gen->max_size() >
2137 2136 young_gen->from_space()->capacity_in_bytes() +
2138 2137 young_gen->to_space()->capacity_in_bytes(),
2139 2138 "Sizes of space in young gen are out-of-bounds");
2140 2139 size_t max_eden_size = young_gen->max_size() -
2141 2140 young_gen->from_space()->capacity_in_bytes() -
2142 2141 young_gen->to_space()->capacity_in_bytes();
2143 2142 size_policy->compute_generation_free_space(
2144 2143 young_gen->used_in_bytes(),
2145 2144 young_gen->eden_space()->used_in_bytes(),
2146 2145 old_gen->used_in_bytes(),
2147 2146 perm_gen->used_in_bytes(),
2148 2147 young_gen->eden_space()->capacity_in_bytes(),
2149 2148 old_gen->max_gen_size(),
2150 2149 max_eden_size,
2151 2150 true /* full gc*/,
2152 2151 gc_cause,
2153 2152 heap->collector_policy());
2154 2153
2155 2154 heap->resize_old_gen(
2156 2155 size_policy->calculated_old_free_size_in_bytes());
2157 2156
2158 2157 // Don't resize the young generation at an major collection. A
2159 2158 // desired young generation size may have been calculated but
2160 2159 // resizing the young generation complicates the code because the
2161 2160 // resizing of the old generation may have moved the boundary
2162 2161 // between the young generation and the old generation. Let the
2163 2162 // young generation resizing happen at the minor collections.
2164 2163 }
2165 2164 if (PrintAdaptiveSizePolicy) {
2166 2165 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2167 2166 heap->total_collections());
2168 2167 }
2169 2168 }
2170 2169
2171 2170 if (UsePerfData) {
2172 2171 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2173 2172 counters->update_counters();
2174 2173 counters->update_old_capacity(old_gen->capacity_in_bytes());
2175 2174 counters->update_young_capacity(young_gen->capacity_in_bytes());
2176 2175 }
2177 2176
2178 2177 heap->resize_all_tlabs();
2179 2178
2180 2179 // We collected the perm gen, so we'll resize it here.
2181 2180 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2182 2181
2183 2182 if (TraceGen1Time) accumulated_time()->stop();
2184 2183
2185 2184 if (PrintGC) {
2186 2185 if (PrintGCDetails) {
2187 2186 // No GC timestamp here. This is after GC so it would be confusing.
2188 2187 young_gen->print_used_change(pre_gc_values.young_gen_used());
2189 2188 old_gen->print_used_change(pre_gc_values.old_gen_used());
2190 2189 heap->print_heap_change(pre_gc_values.heap_used());
2191 2190 // Print perm gen last (print_heap_change() excludes the perm gen).
2192 2191 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2193 2192 } else {
2194 2193 heap->print_heap_change(pre_gc_values.heap_used());
2195 2194 }
2196 2195 }
2197 2196
2198 2197 // Track memory usage and detect low memory
2199 2198 MemoryService::track_memory_usage();
2200 2199 heap->update_counters();
2201 2200 }
2202 2201
2203 2202 #ifdef ASSERT
2204 2203 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2205 2204 ParCompactionManager* const cm =
2206 2205 ParCompactionManager::manager_array(int(i));
2207 2206 assert(cm->marking_stack()->is_empty(), "should be empty");
2208 2207 assert(cm->region_stack()->is_empty(), "should be empty");
2209 2208 assert(cm->revisit_klass_stack()->is_empty(), "should be empty");
2210 2209 }
2211 2210 #endif // ASSERT
2212 2211
2213 2212 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2214 2213 HandleMark hm; // Discard invalid handles created during verification
2215 2214 gclog_or_tty->print(" VerifyAfterGC:");
2216 2215 Universe::verify(false);
2217 2216 }
2218 2217
2219 2218 // Re-verify object start arrays
2220 2219 if (VerifyObjectStartArray &&
2221 2220 VerifyAfterGC) {
2222 2221 old_gen->verify_object_start_array();
2223 2222 perm_gen->verify_object_start_array();
2224 2223 }
2225 2224
2226 2225 if (ZapUnusedHeapArea) {
2227 2226 old_gen->object_space()->check_mangled_unused_area_complete();
2228 2227 perm_gen->object_space()->check_mangled_unused_area_complete();
2229 2228 }
2230 2229
2231 2230 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2232 2231
2233 2232 collection_exit.update();
2234 2233
2235 2234 if (PrintHeapAtGC) {
2236 2235 Universe::print_heap_after_gc();
2237 2236 }
2238 2237 if (PrintGCTaskTimeStamps) {
2239 2238 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2240 2239 INT64_FORMAT,
2241 2240 marking_start.ticks(), compaction_start.ticks(),
2242 2241 collection_exit.ticks());
2243 2242 gc_task_manager()->print_task_time_stamps();
2244 2243 }
2245 2244
2246 2245 heap->post_full_gc_dump();
2247 2246
2248 2247 #ifdef TRACESPINNING
2249 2248 ParallelTaskTerminator::print_termination_counts();
2250 2249 #endif
2251 2250 }
2252 2251
2253 2252 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2254 2253 PSYoungGen* young_gen,
2255 2254 PSOldGen* old_gen) {
2256 2255 MutableSpace* const eden_space = young_gen->eden_space();
2257 2256 assert(!eden_space->is_empty(), "eden must be non-empty");
2258 2257 assert(young_gen->virtual_space()->alignment() ==
2259 2258 old_gen->virtual_space()->alignment(), "alignments do not match");
2260 2259
2261 2260 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2262 2261 return false;
2263 2262 }
2264 2263
2265 2264 // Both generations must be completely committed.
2266 2265 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2267 2266 return false;
2268 2267 }
2269 2268 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2270 2269 return false;
2271 2270 }
2272 2271
2273 2272 // Figure out how much to take from eden. Include the average amount promoted
2274 2273 // in the total; otherwise the next young gen GC will simply bail out to a
2275 2274 // full GC.
2276 2275 const size_t alignment = old_gen->virtual_space()->alignment();
2277 2276 const size_t eden_used = eden_space->used_in_bytes();
2278 2277 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2279 2278 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2280 2279 const size_t eden_capacity = eden_space->capacity_in_bytes();
2281 2280
2282 2281 if (absorb_size >= eden_capacity) {
2283 2282 return false; // Must leave some space in eden.
2284 2283 }
2285 2284
2286 2285 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2287 2286 if (new_young_size < young_gen->min_gen_size()) {
2288 2287 return false; // Respect young gen minimum size.
2289 2288 }
2290 2289
2291 2290 if (TraceAdaptiveGCBoundary && Verbose) {
2292 2291 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2293 2292 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2294 2293 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2295 2294 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2296 2295 absorb_size / K,
2297 2296 eden_capacity / K, (eden_capacity - absorb_size) / K,
2298 2297 young_gen->from_space()->used_in_bytes() / K,
2299 2298 young_gen->to_space()->used_in_bytes() / K,
2300 2299 young_gen->capacity_in_bytes() / K, new_young_size / K);
2301 2300 }
2302 2301
2303 2302 // Fill the unused part of the old gen.
2304 2303 MutableSpace* const old_space = old_gen->object_space();
2305 2304 HeapWord* const unused_start = old_space->top();
2306 2305 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2307 2306
2308 2307 if (unused_words > 0) {
2309 2308 if (unused_words < CollectedHeap::min_fill_size()) {
2310 2309 return false; // If the old gen cannot be filled, must give up.
2311 2310 }
2312 2311 CollectedHeap::fill_with_objects(unused_start, unused_words);
2313 2312 }
2314 2313
2315 2314 // Take the live data from eden and set both top and end in the old gen to
2316 2315 // eden top. (Need to set end because reset_after_change() mangles the region
2317 2316 // from end to virtual_space->high() in debug builds).
2318 2317 HeapWord* const new_top = eden_space->top();
2319 2318 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2320 2319 absorb_size);
2321 2320 young_gen->reset_after_change();
2322 2321 old_space->set_top(new_top);
2323 2322 old_space->set_end(new_top);
2324 2323 old_gen->reset_after_change();
2325 2324
2326 2325 // Update the object start array for the filler object and the data from eden.
2327 2326 ObjectStartArray* const start_array = old_gen->start_array();
2328 2327 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2329 2328 start_array->allocate_block(p);
2330 2329 }
2331 2330
2332 2331 // Could update the promoted average here, but it is not typically updated at
2333 2332 // full GCs and the value to use is unclear. Something like
2334 2333 //
2335 2334 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2336 2335
2337 2336 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2338 2337 return true;
2339 2338 }
2340 2339
2341 2340 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2342 2341 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2343 2342 "shouldn't return NULL");
2344 2343 return ParallelScavengeHeap::gc_task_manager();
2345 2344 }
2346 2345
2347 2346 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2348 2347 bool maximum_heap_compaction) {
2349 2348 // Recursively traverse all live objects and mark them
2350 2349 EventMark m("1 mark object");
2351 2350 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2352 2351
2353 2352 ParallelScavengeHeap* heap = gc_heap();
2354 2353 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2355 2354 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2356 2355 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2357 2356
2358 2357 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2359 2358 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2360 2359
2361 2360 {
2362 2361 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2363 2362 ParallelScavengeHeap::ParStrongRootsScope psrs;
2364 2363
2365 2364 GCTaskQueue* q = GCTaskQueue::create();
2366 2365
2367 2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2368 2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2369 2368 // We scan the thread roots in parallel
2370 2369 Threads::create_thread_roots_marking_tasks(q);
2371 2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2372 2371 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2373 2372 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2374 2373 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2375 2374 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2376 2375 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2377 2376
2378 2377 if (parallel_gc_threads > 1) {
2379 2378 for (uint j = 0; j < parallel_gc_threads; j++) {
2380 2379 q->enqueue(new StealMarkingTask(&terminator));
2381 2380 }
2382 2381 }
2383 2382
2384 2383 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2385 2384 q->enqueue(fin);
2386 2385
2387 2386 gc_task_manager()->add_list(q);
2388 2387
2389 2388 fin->wait_for();
2390 2389
2391 2390 // We have to release the barrier tasks!
2392 2391 WaitForBarrierGCTask::destroy(fin);
2393 2392 }
2394 2393
2395 2394 // Process reference objects found during marking
2396 2395 {
2397 2396 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2398 2397 if (ref_processor()->processing_is_mt()) {
2399 2398 RefProcTaskExecutor task_executor;
2400 2399 ref_processor()->process_discovered_references(
2401 2400 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2402 2401 &task_executor);
2403 2402 } else {
2404 2403 ref_processor()->process_discovered_references(
2405 2404 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2406 2405 }
2407 2406 }
2408 2407
2409 2408 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2410 2409 // Follow system dictionary roots and unload classes.
2411 2410 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2412 2411
2413 2412 // Follow code cache roots.
2414 2413 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2415 2414 purged_class);
2416 2415 cm->follow_marking_stacks(); // Flush marking stack.
2417 2416
2418 2417 // Update subklass/sibling/implementor links of live klasses
2419 2418 // revisit_klass_stack is used in follow_weak_klass_links().
2420 2419 follow_weak_klass_links();
2421 2420
2422 2421 // Revisit memoized MDO's and clear any unmarked weak refs
2423 2422 follow_mdo_weak_refs();
2424 2423
2425 2424 // Visit interned string tables and delete unmarked oops
2426 2425 StringTable::unlink(is_alive_closure());
2427 2426 // Clean up unreferenced symbols in symbol table.
2428 2427 SymbolTable::unlink();
2429 2428
2430 2429 assert(cm->marking_stacks_empty(), "marking stacks should be empty");
2431 2430 }
2432 2431
2433 2432 // This should be moved to the shared markSweep code!
2434 2433 class PSAlwaysTrueClosure: public BoolObjectClosure {
2435 2434 public:
2436 2435 void do_object(oop p) { ShouldNotReachHere(); }
2437 2436 bool do_object_b(oop p) { return true; }
2438 2437 };
2439 2438 static PSAlwaysTrueClosure always_true;
2440 2439
2441 2440 void PSParallelCompact::adjust_roots() {
2442 2441 // Adjust the pointers to reflect the new locations
2443 2442 EventMark m("3 adjust roots");
2444 2443 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2445 2444
2446 2445 // General strong roots.
2447 2446 Universe::oops_do(adjust_root_pointer_closure());
2448 2447 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
2449 2448 Threads::oops_do(adjust_root_pointer_closure(), NULL);
2450 2449 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2451 2450 FlatProfiler::oops_do(adjust_root_pointer_closure());
2452 2451 Management::oops_do(adjust_root_pointer_closure());
2453 2452 JvmtiExport::oops_do(adjust_root_pointer_closure());
2454 2453 // SO_AllClasses
2455 2454 SystemDictionary::oops_do(adjust_root_pointer_closure());
2456 2455
2457 2456 // Now adjust pointers in remaining weak roots. (All of which should
2458 2457 // have been cleared if they pointed to non-surviving objects.)
2459 2458 // Global (weak) JNI handles
2460 2459 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2461 2460
2462 2461 CodeCache::oops_do(adjust_pointer_closure());
2463 2462 StringTable::oops_do(adjust_root_pointer_closure());
2464 2463 ref_processor()->weak_oops_do(adjust_root_pointer_closure());
2465 2464 // Roots were visited so references into the young gen in roots
2466 2465 // may have been scanned. Process them also.
2467 2466 // Should the reference processor have a span that excludes
2468 2467 // young gen objects?
2469 2468 PSScavenge::reference_processor()->weak_oops_do(
2470 2469 adjust_root_pointer_closure());
2471 2470 }
2472 2471
2473 2472 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2474 2473 EventMark m("4 compact perm");
2475 2474 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
2476 2475 // trace("4");
2477 2476
2478 2477 gc_heap()->perm_gen()->start_array()->reset();
2479 2478 move_and_update(cm, perm_space_id);
2480 2479 }
2481 2480
2482 2481 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2483 2482 uint parallel_gc_threads)
2484 2483 {
2485 2484 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2486 2485
2487 2486 const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
2488 2487 for (unsigned int j = 0; j < task_count; j++) {
2489 2488 q->enqueue(new DrainStacksCompactionTask(j));
2490 2489 }
2491 2490
2492 2491 // Find all regions that are available (can be filled immediately) and
2493 2492 // distribute them to the thread stacks. The iteration is done in reverse
2494 2493 // order (high to low) so the regions will be removed in ascending order.
2495 2494
2496 2495 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2497 2496
2498 2497 size_t fillable_regions = 0; // A count for diagnostic purposes.
2499 2498 unsigned int which = 0; // The worker thread number.
2500 2499
2501 2500 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2502 2501 SpaceInfo* const space_info = _space_info + id;
2503 2502 MutableSpace* const space = space_info->space();
2504 2503 HeapWord* const new_top = space_info->new_top();
2505 2504
2506 2505 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2507 2506 const size_t end_region =
2508 2507 sd.addr_to_region_idx(sd.region_align_up(new_top));
2509 2508 assert(end_region > 0, "perm gen cannot be empty");
2510 2509
2511 2510 for (size_t cur = end_region - 1; cur >= beg_region; --cur) {
2512 2511 if (sd.region(cur)->claim_unsafe()) {
2513 2512 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2514 2513 cm->push_region(cur);
2515 2514
2516 2515 if (TraceParallelOldGCCompactionPhase && Verbose) {
2517 2516 const size_t count_mod_8 = fillable_regions & 7;
2518 2517 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2519 2518 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2520 2519 if (count_mod_8 == 7) gclog_or_tty->cr();
2521 2520 }
2522 2521
2523 2522 NOT_PRODUCT(++fillable_regions;)
2524 2523
2525 2524 // Assign regions to threads in round-robin fashion.
2526 2525 if (++which == task_count) {
2527 2526 which = 0;
2528 2527 }
2529 2528 }
2530 2529 }
2531 2530 }
2532 2531
2533 2532 if (TraceParallelOldGCCompactionPhase) {
2534 2533 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2535 2534 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2536 2535 }
2537 2536 }
2538 2537
2539 2538 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2540 2539
2541 2540 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2542 2541 uint parallel_gc_threads) {
2543 2542 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2544 2543
2545 2544 ParallelCompactData& sd = PSParallelCompact::summary_data();
2546 2545
2547 2546 // Iterate over all the spaces adding tasks for updating
2548 2547 // regions in the dense prefix. Assume that 1 gc thread
2549 2548 // will work on opening the gaps and the remaining gc threads
2550 2549 // will work on the dense prefix.
2551 2550 unsigned int space_id;
2552 2551 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2553 2552 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2554 2553 const MutableSpace* const space = _space_info[space_id].space();
2555 2554
2556 2555 if (dense_prefix_end == space->bottom()) {
2557 2556 // There is no dense prefix for this space.
2558 2557 continue;
2559 2558 }
2560 2559
2561 2560 // The dense prefix is before this region.
2562 2561 size_t region_index_end_dense_prefix =
2563 2562 sd.addr_to_region_idx(dense_prefix_end);
2564 2563 RegionData* const dense_prefix_cp =
2565 2564 sd.region(region_index_end_dense_prefix);
2566 2565 assert(dense_prefix_end == space->end() ||
2567 2566 dense_prefix_cp->available() ||
2568 2567 dense_prefix_cp->claimed(),
2569 2568 "The region after the dense prefix should always be ready to fill");
2570 2569
2571 2570 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2572 2571
2573 2572 // Is there dense prefix work?
2574 2573 size_t total_dense_prefix_regions =
2575 2574 region_index_end_dense_prefix - region_index_start;
2576 2575 // How many regions of the dense prefix should be given to
2577 2576 // each thread?
2578 2577 if (total_dense_prefix_regions > 0) {
2579 2578 uint tasks_for_dense_prefix = 1;
2580 2579 if (total_dense_prefix_regions <=
2581 2580 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2582 2581 // Don't over partition. This assumes that
2583 2582 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2584 2583 // so there are not many regions to process.
2585 2584 tasks_for_dense_prefix = parallel_gc_threads;
2586 2585 } else {
2587 2586 // Over partition
2588 2587 tasks_for_dense_prefix = parallel_gc_threads *
2589 2588 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2590 2589 }
2591 2590 size_t regions_per_thread = total_dense_prefix_regions /
2592 2591 tasks_for_dense_prefix;
2593 2592 // Give each thread at least 1 region.
2594 2593 if (regions_per_thread == 0) {
2595 2594 regions_per_thread = 1;
2596 2595 }
2597 2596
2598 2597 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2599 2598 if (region_index_start >= region_index_end_dense_prefix) {
2600 2599 break;
2601 2600 }
2602 2601 // region_index_end is not processed
2603 2602 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2604 2603 region_index_end_dense_prefix);
2605 2604 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2606 2605 region_index_start,
2607 2606 region_index_end));
2608 2607 region_index_start = region_index_end;
2609 2608 }
2610 2609 }
2611 2610 // This gets any part of the dense prefix that did not
2612 2611 // fit evenly.
2613 2612 if (region_index_start < region_index_end_dense_prefix) {
2614 2613 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2615 2614 region_index_start,
2616 2615 region_index_end_dense_prefix));
2617 2616 }
2618 2617 }
2619 2618 }
2620 2619
2621 2620 void PSParallelCompact::enqueue_region_stealing_tasks(
2622 2621 GCTaskQueue* q,
2623 2622 ParallelTaskTerminator* terminator_ptr,
2624 2623 uint parallel_gc_threads) {
2625 2624 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2626 2625
2627 2626 // Once a thread has drained it's stack, it should try to steal regions from
2628 2627 // other threads.
2629 2628 if (parallel_gc_threads > 1) {
2630 2629 for (uint j = 0; j < parallel_gc_threads; j++) {
2631 2630 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2632 2631 }
2633 2632 }
2634 2633 }
2635 2634
2636 2635 void PSParallelCompact::compact() {
2637 2636 EventMark m("5 compact");
2638 2637 // trace("5");
2639 2638 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2640 2639
2641 2640 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2642 2641 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2643 2642 PSOldGen* old_gen = heap->old_gen();
2644 2643 old_gen->start_array()->reset();
2645 2644 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2646 2645 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2647 2646 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2648 2647
2649 2648 GCTaskQueue* q = GCTaskQueue::create();
2650 2649 enqueue_region_draining_tasks(q, parallel_gc_threads);
2651 2650 enqueue_dense_prefix_tasks(q, parallel_gc_threads);
2652 2651 enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads);
2653 2652
2654 2653 {
2655 2654 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2656 2655
2657 2656 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2658 2657 q->enqueue(fin);
2659 2658
2660 2659 gc_task_manager()->add_list(q);
2661 2660
2662 2661 fin->wait_for();
2663 2662
2664 2663 // We have to release the barrier tasks!
2665 2664 WaitForBarrierGCTask::destroy(fin);
2666 2665
2667 2666 #ifdef ASSERT
2668 2667 // Verify that all regions have been processed before the deferred updates.
2669 2668 // Note that perm_space_id is skipped; this type of verification is not
2670 2669 // valid until the perm gen is compacted by regions.
2671 2670 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2672 2671 verify_complete(SpaceId(id));
2673 2672 }
2674 2673 #endif
2675 2674 }
2676 2675
2677 2676 {
2678 2677 // Update the deferred objects, if any. Any compaction manager can be used.
2679 2678 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2680 2679 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2681 2680 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2682 2681 update_deferred_objects(cm, SpaceId(id));
2683 2682 }
2684 2683 }
2685 2684 }
2686 2685
2687 2686 #ifdef ASSERT
2688 2687 void PSParallelCompact::verify_complete(SpaceId space_id) {
2689 2688 // All Regions between space bottom() to new_top() should be marked as filled
2690 2689 // and all Regions between new_top() and top() should be available (i.e.,
2691 2690 // should have been emptied).
2692 2691 ParallelCompactData& sd = summary_data();
2693 2692 SpaceInfo si = _space_info[space_id];
2694 2693 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2695 2694 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2696 2695 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2697 2696 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2698 2697 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2699 2698
2700 2699 bool issued_a_warning = false;
2701 2700
2702 2701 size_t cur_region;
2703 2702 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2704 2703 const RegionData* const c = sd.region(cur_region);
2705 2704 if (!c->completed()) {
2706 2705 warning("region " SIZE_FORMAT " not filled: "
2707 2706 "destination_count=" SIZE_FORMAT,
2708 2707 cur_region, c->destination_count());
2709 2708 issued_a_warning = true;
2710 2709 }
2711 2710 }
2712 2711
2713 2712 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2714 2713 const RegionData* const c = sd.region(cur_region);
2715 2714 if (!c->available()) {
2716 2715 warning("region " SIZE_FORMAT " not empty: "
2717 2716 "destination_count=" SIZE_FORMAT,
2718 2717 cur_region, c->destination_count());
2719 2718 issued_a_warning = true;
2720 2719 }
2721 2720 }
2722 2721
2723 2722 if (issued_a_warning) {
2724 2723 print_region_ranges();
2725 2724 }
2726 2725 }
2727 2726 #endif // #ifdef ASSERT
2728 2727
2729 2728 void
2730 2729 PSParallelCompact::follow_weak_klass_links() {
2731 2730 // All klasses on the revisit stack are marked at this point.
2732 2731 // Update and follow all subklass, sibling and implementor links.
2733 2732 if (PrintRevisitStats) {
2734 2733 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2735 2734 SystemDictionary::number_of_classes());
2736 2735 }
2737 2736 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2738 2737 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2739 2738 KeepAliveClosure keep_alive_closure(cm);
2740 2739 Stack<Klass*>* const rks = cm->revisit_klass_stack();
2741 2740 if (PrintRevisitStats) {
2742 2741 gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT,
2743 2742 i, rks->size());
2744 2743 }
2745 2744 while (!rks->is_empty()) {
2746 2745 Klass* const k = rks->pop();
2747 2746 k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure);
2748 2747 }
2749 2748
2750 2749 cm->follow_marking_stacks();
2751 2750 }
2752 2751 }
2753 2752
2754 2753 void
2755 2754 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2756 2755 cm->revisit_klass_stack()->push(k);
2757 2756 }
2758 2757
2759 2758 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) {
2760 2759 cm->revisit_mdo_stack()->push(p);
2761 2760 }
2762 2761
2763 2762 void PSParallelCompact::follow_mdo_weak_refs() {
2764 2763 // All strongly reachable oops have been marked at this point;
2765 2764 // we can visit and clear any weak references from MDO's which
2766 2765 // we memoized during the strong marking phase.
2767 2766 if (PrintRevisitStats) {
2768 2767 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2769 2768 SystemDictionary::number_of_classes());
2770 2769 }
2771 2770 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2772 2771 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2773 2772 Stack<DataLayout*>* rms = cm->revisit_mdo_stack();
2774 2773 if (PrintRevisitStats) {
2775 2774 gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT,
2776 2775 i, rms->size());
2777 2776 }
2778 2777 while (!rms->is_empty()) {
2779 2778 rms->pop()->follow_weak_refs(is_alive_closure());
2780 2779 }
2781 2780
2782 2781 cm->follow_marking_stacks();
2783 2782 }
2784 2783 }
2785 2784
2786 2785
2787 2786 #ifdef VALIDATE_MARK_SWEEP
2788 2787
2789 2788 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2790 2789 if (!ValidateMarkSweep)
2791 2790 return;
2792 2791
2793 2792 if (!isroot) {
2794 2793 if (_pointer_tracking) {
2795 2794 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2796 2795 _adjusted_pointers->remove(p);
2797 2796 }
2798 2797 } else {
2799 2798 ptrdiff_t index = _root_refs_stack->find(p);
2800 2799 if (index != -1) {
2801 2800 int l = _root_refs_stack->length();
2802 2801 if (l > 0 && l - 1 != index) {
2803 2802 void* last = _root_refs_stack->pop();
2804 2803 assert(last != p, "should be different");
2805 2804 _root_refs_stack->at_put(index, last);
2806 2805 } else {
2807 2806 _root_refs_stack->remove(p);
2808 2807 }
2809 2808 }
2810 2809 }
2811 2810 }
2812 2811
2813 2812
2814 2813 void PSParallelCompact::check_adjust_pointer(void* p) {
2815 2814 _adjusted_pointers->push(p);
2816 2815 }
2817 2816
2818 2817
2819 2818 class AdjusterTracker: public OopClosure {
2820 2819 public:
2821 2820 AdjusterTracker() {};
2822 2821 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
2823 2822 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); }
2824 2823 };
2825 2824
2826 2825
2827 2826 void PSParallelCompact::track_interior_pointers(oop obj) {
2828 2827 if (ValidateMarkSweep) {
2829 2828 _adjusted_pointers->clear();
2830 2829 _pointer_tracking = true;
2831 2830
2832 2831 AdjusterTracker checker;
2833 2832 obj->oop_iterate(&checker);
2834 2833 }
2835 2834 }
2836 2835
2837 2836
2838 2837 void PSParallelCompact::check_interior_pointers() {
2839 2838 if (ValidateMarkSweep) {
2840 2839 _pointer_tracking = false;
2841 2840 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2842 2841 }
2843 2842 }
2844 2843
2845 2844
2846 2845 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2847 2846 if (ValidateMarkSweep) {
2848 2847 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2849 2848 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2850 2849 }
2851 2850 }
2852 2851
2853 2852
2854 2853 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2855 2854 if (ValidateMarkSweep) {
2856 2855 _live_oops->push(p);
2857 2856 _live_oops_size->push(size);
2858 2857 _live_oops_index++;
2859 2858 }
2860 2859 }
2861 2860
2862 2861 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2863 2862 if (ValidateMarkSweep) {
2864 2863 oop obj = _live_oops->at((int)_live_oops_index);
2865 2864 guarantee(obj == p, "should be the same object");
2866 2865 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2867 2866 _live_oops_index++;
2868 2867 }
2869 2868 }
2870 2869
2871 2870 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2872 2871 HeapWord* compaction_top) {
2873 2872 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2874 2873 "should be moved to forwarded location");
2875 2874 if (ValidateMarkSweep) {
2876 2875 PSParallelCompact::validate_live_oop(oop(q), size);
2877 2876 _live_oops_moved_to->push(oop(compaction_top));
2878 2877 }
2879 2878 if (RecordMarkSweepCompaction) {
2880 2879 _cur_gc_live_oops->push(q);
2881 2880 _cur_gc_live_oops_moved_to->push(compaction_top);
2882 2881 _cur_gc_live_oops_size->push(size);
2883 2882 }
2884 2883 }
2885 2884
2886 2885
2887 2886 void PSParallelCompact::compaction_complete() {
2888 2887 if (RecordMarkSweepCompaction) {
2889 2888 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
2890 2889 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2891 2890 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
2892 2891
2893 2892 _cur_gc_live_oops = _last_gc_live_oops;
2894 2893 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
2895 2894 _cur_gc_live_oops_size = _last_gc_live_oops_size;
2896 2895 _last_gc_live_oops = _tmp_live_oops;
2897 2896 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2898 2897 _last_gc_live_oops_size = _tmp_live_oops_size;
2899 2898 }
2900 2899 }
2901 2900
2902 2901
2903 2902 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2904 2903 if (!RecordMarkSweepCompaction) {
2905 2904 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2906 2905 return;
2907 2906 }
2908 2907
2909 2908 if (_last_gc_live_oops == NULL) {
2910 2909 tty->print_cr("No compaction information gathered yet");
2911 2910 return;
2912 2911 }
2913 2912
2914 2913 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2915 2914 HeapWord* old_oop = _last_gc_live_oops->at(i);
2916 2915 size_t sz = _last_gc_live_oops_size->at(i);
2917 2916 if (old_oop <= q && q < (old_oop + sz)) {
2918 2917 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2919 2918 size_t offset = (q - old_oop);
2920 2919 tty->print_cr("Address " PTR_FORMAT, q);
2921 2920 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2922 2921 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2923 2922 return;
2924 2923 }
2925 2924 }
2926 2925
2927 2926 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2928 2927 }
2929 2928 #endif //VALIDATE_MARK_SWEEP
2930 2929
2931 2930 // Update interior oops in the ranges of regions [beg_region, end_region).
2932 2931 void
2933 2932 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2934 2933 SpaceId space_id,
2935 2934 size_t beg_region,
2936 2935 size_t end_region) {
2937 2936 ParallelCompactData& sd = summary_data();
2938 2937 ParMarkBitMap* const mbm = mark_bitmap();
2939 2938
2940 2939 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2941 2940 HeapWord* const end_addr = sd.region_to_addr(end_region);
2942 2941 assert(beg_region <= end_region, "bad region range");
2943 2942 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2944 2943
2945 2944 #ifdef ASSERT
2946 2945 // Claim the regions to avoid triggering an assert when they are marked as
2947 2946 // filled.
2948 2947 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2949 2948 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2950 2949 }
2951 2950 #endif // #ifdef ASSERT
2952 2951
2953 2952 if (beg_addr != space(space_id)->bottom()) {
2954 2953 // Find the first live object or block of dead space that *starts* in this
2955 2954 // range of regions. If a partial object crosses onto the region, skip it;
2956 2955 // it will be marked for 'deferred update' when the object head is
2957 2956 // processed. If dead space crosses onto the region, it is also skipped; it
2958 2957 // will be filled when the prior region is processed. If neither of those
2959 2958 // apply, the first word in the region is the start of a live object or dead
2960 2959 // space.
2961 2960 assert(beg_addr > space(space_id)->bottom(), "sanity");
2962 2961 const RegionData* const cp = sd.region(beg_region);
2963 2962 if (cp->partial_obj_size() != 0) {
2964 2963 beg_addr = sd.partial_obj_end(beg_region);
2965 2964 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2966 2965 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2967 2966 }
2968 2967 }
2969 2968
2970 2969 if (beg_addr < end_addr) {
2971 2970 // A live object or block of dead space starts in this range of Regions.
2972 2971 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2973 2972
2974 2973 // Create closures and iterate.
2975 2974 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2976 2975 FillClosure fill_closure(cm, space_id);
2977 2976 ParMarkBitMap::IterationStatus status;
2978 2977 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2979 2978 dense_prefix_end);
2980 2979 if (status == ParMarkBitMap::incomplete) {
2981 2980 update_closure.do_addr(update_closure.source());
2982 2981 }
2983 2982 }
2984 2983
2985 2984 // Mark the regions as filled.
2986 2985 RegionData* const beg_cp = sd.region(beg_region);
2987 2986 RegionData* const end_cp = sd.region(end_region);
2988 2987 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2989 2988 cp->set_completed();
2990 2989 }
2991 2990 }
2992 2991
2993 2992 // Return the SpaceId for the space containing addr. If addr is not in the
2994 2993 // heap, last_space_id is returned. In debug mode it expects the address to be
2995 2994 // in the heap and asserts such.
2996 2995 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2997 2996 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2998 2997
2999 2998 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
3000 2999 if (_space_info[id].space()->contains(addr)) {
3001 3000 return SpaceId(id);
3002 3001 }
3003 3002 }
3004 3003
3005 3004 assert(false, "no space contains the addr");
3006 3005 return last_space_id;
3007 3006 }
3008 3007
3009 3008 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
3010 3009 SpaceId id) {
3011 3010 assert(id < last_space_id, "bad space id");
3012 3011
3013 3012 ParallelCompactData& sd = summary_data();
3014 3013 const SpaceInfo* const space_info = _space_info + id;
3015 3014 ObjectStartArray* const start_array = space_info->start_array();
3016 3015
3017 3016 const MutableSpace* const space = space_info->space();
3018 3017 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
3019 3018 HeapWord* const beg_addr = space_info->dense_prefix();
3020 3019 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
3021 3020
3022 3021 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
3023 3022 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
3024 3023 const RegionData* cur_region;
3025 3024 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
3026 3025 HeapWord* const addr = cur_region->deferred_obj_addr();
3027 3026 if (addr != NULL) {
3028 3027 if (start_array != NULL) {
3029 3028 start_array->allocate_block(addr);
3030 3029 }
3031 3030 oop(addr)->update_contents(cm);
3032 3031 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3033 3032 }
3034 3033 }
3035 3034 }
3036 3035
3037 3036 // Skip over count live words starting from beg, and return the address of the
3038 3037 // next live word. Unless marked, the word corresponding to beg is assumed to
3039 3038 // be dead. Callers must either ensure beg does not correspond to the middle of
3040 3039 // an object, or account for those live words in some other way. Callers must
3041 3040 // also ensure that there are enough live words in the range [beg, end) to skip.
3042 3041 HeapWord*
3043 3042 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3044 3043 {
3045 3044 assert(count > 0, "sanity");
3046 3045
3047 3046 ParMarkBitMap* m = mark_bitmap();
3048 3047 idx_t bits_to_skip = m->words_to_bits(count);
3049 3048 idx_t cur_beg = m->addr_to_bit(beg);
3050 3049 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3051 3050
3052 3051 do {
3053 3052 cur_beg = m->find_obj_beg(cur_beg, search_end);
3054 3053 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3055 3054 const size_t obj_bits = cur_end - cur_beg + 1;
3056 3055 if (obj_bits > bits_to_skip) {
3057 3056 return m->bit_to_addr(cur_beg + bits_to_skip);
3058 3057 }
3059 3058 bits_to_skip -= obj_bits;
3060 3059 cur_beg = cur_end + 1;
3061 3060 } while (bits_to_skip > 0);
3062 3061
3063 3062 // Skipping the desired number of words landed just past the end of an object.
3064 3063 // Find the start of the next object.
3065 3064 cur_beg = m->find_obj_beg(cur_beg, search_end);
3066 3065 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3067 3066 return m->bit_to_addr(cur_beg);
3068 3067 }
3069 3068
3070 3069 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3071 3070 SpaceId src_space_id,
3072 3071 size_t src_region_idx)
3073 3072 {
3074 3073 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
3075 3074
3076 3075 const SplitInfo& split_info = _space_info[src_space_id].split_info();
3077 3076 if (split_info.dest_region_addr() == dest_addr) {
3078 3077 // The partial object ending at the split point contains the first word to
3079 3078 // be copied to dest_addr.
3080 3079 return split_info.first_src_addr();
3081 3080 }
3082 3081
3083 3082 const ParallelCompactData& sd = summary_data();
3084 3083 ParMarkBitMap* const bitmap = mark_bitmap();
3085 3084 const size_t RegionSize = ParallelCompactData::RegionSize;
3086 3085
3087 3086 assert(sd.is_region_aligned(dest_addr), "not aligned");
3088 3087 const RegionData* const src_region_ptr = sd.region(src_region_idx);
3089 3088 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
3090 3089 HeapWord* const src_region_destination = src_region_ptr->destination();
3091 3090
3092 3091 assert(dest_addr >= src_region_destination, "wrong src region");
3093 3092 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
3094 3093
3095 3094 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
3096 3095 HeapWord* const src_region_end = src_region_beg + RegionSize;
3097 3096
3098 3097 HeapWord* addr = src_region_beg;
3099 3098 if (dest_addr == src_region_destination) {
3100 3099 // Return the first live word in the source region.
3101 3100 if (partial_obj_size == 0) {
3102 3101 addr = bitmap->find_obj_beg(addr, src_region_end);
3103 3102 assert(addr < src_region_end, "no objects start in src region");
3104 3103 }
3105 3104 return addr;
3106 3105 }
3107 3106
3108 3107 // Must skip some live data.
3109 3108 size_t words_to_skip = dest_addr - src_region_destination;
3110 3109 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
3111 3110
3112 3111 if (partial_obj_size >= words_to_skip) {
3113 3112 // All the live words to skip are part of the partial object.
3114 3113 addr += words_to_skip;
3115 3114 if (partial_obj_size == words_to_skip) {
3116 3115 // Find the first live word past the partial object.
3117 3116 addr = bitmap->find_obj_beg(addr, src_region_end);
3118 3117 assert(addr < src_region_end, "wrong src region");
3119 3118 }
3120 3119 return addr;
3121 3120 }
3122 3121
3123 3122 // Skip over the partial object (if any).
3124 3123 if (partial_obj_size != 0) {
3125 3124 words_to_skip -= partial_obj_size;
3126 3125 addr += partial_obj_size;
3127 3126 }
3128 3127
3129 3128 // Skip over live words due to objects that start in the region.
3130 3129 addr = skip_live_words(addr, src_region_end, words_to_skip);
3131 3130 assert(addr < src_region_end, "wrong src region");
3132 3131 return addr;
3133 3132 }
3134 3133
3135 3134 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3136 3135 SpaceId src_space_id,
3137 3136 size_t beg_region,
3138 3137 HeapWord* end_addr)
3139 3138 {
3140 3139 ParallelCompactData& sd = summary_data();
3141 3140
3142 3141 #ifdef ASSERT
3143 3142 MutableSpace* const src_space = _space_info[src_space_id].space();
3144 3143 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3145 3144 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3146 3145 "src_space_id does not match beg_addr");
3147 3146 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3148 3147 "src_space_id does not match end_addr");
3149 3148 #endif // #ifdef ASSERT
3150 3149
3151 3150 RegionData* const beg = sd.region(beg_region);
3152 3151 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3153 3152
3154 3153 // Regions up to new_top() are enqueued if they become available.
3155 3154 HeapWord* const new_top = _space_info[src_space_id].new_top();
3156 3155 RegionData* const enqueue_end =
3157 3156 sd.addr_to_region_ptr(sd.region_align_up(new_top));
3158 3157
3159 3158 for (RegionData* cur = beg; cur < end; ++cur) {
3160 3159 assert(cur->data_size() > 0, "region must have live data");
3161 3160 cur->decrement_destination_count();
3162 3161 if (cur < enqueue_end && cur->available() && cur->claim()) {
3163 3162 cm->push_region(sd.region(cur));
3164 3163 }
3165 3164 }
3166 3165 }
3167 3166
3168 3167 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3169 3168 SpaceId& src_space_id,
3170 3169 HeapWord*& src_space_top,
3171 3170 HeapWord* end_addr)
3172 3171 {
3173 3172 typedef ParallelCompactData::RegionData RegionData;
3174 3173
3175 3174 ParallelCompactData& sd = PSParallelCompact::summary_data();
3176 3175 const size_t region_size = ParallelCompactData::RegionSize;
3177 3176
3178 3177 size_t src_region_idx = 0;
3179 3178
3180 3179 // Skip empty regions (if any) up to the top of the space.
3181 3180 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3182 3181 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3183 3182 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3184 3183 const RegionData* const top_region_ptr =
3185 3184 sd.addr_to_region_ptr(top_aligned_up);
3186 3185 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3187 3186 ++src_region_ptr;
3188 3187 }
3189 3188
3190 3189 if (src_region_ptr < top_region_ptr) {
3191 3190 // The next source region is in the current space. Update src_region_idx
3192 3191 // and the source address to match src_region_ptr.
3193 3192 src_region_idx = sd.region(src_region_ptr);
3194 3193 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3195 3194 if (src_region_addr > closure.source()) {
3196 3195 closure.set_source(src_region_addr);
3197 3196 }
3198 3197 return src_region_idx;
3199 3198 }
3200 3199
3201 3200 // Switch to a new source space and find the first non-empty region.
3202 3201 unsigned int space_id = src_space_id + 1;
3203 3202 assert(space_id < last_space_id, "not enough spaces");
3204 3203
3205 3204 HeapWord* const destination = closure.destination();
3206 3205
3207 3206 do {
3208 3207 MutableSpace* space = _space_info[space_id].space();
3209 3208 HeapWord* const bottom = space->bottom();
3210 3209 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3211 3210
3212 3211 // Iterate over the spaces that do not compact into themselves.
3213 3212 if (bottom_cp->destination() != bottom) {
3214 3213 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3215 3214 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3216 3215
3217 3216 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3218 3217 if (src_cp->live_obj_size() > 0) {
3219 3218 // Found it.
3220 3219 assert(src_cp->destination() == destination,
3221 3220 "first live obj in the space must match the destination");
3222 3221 assert(src_cp->partial_obj_size() == 0,
3223 3222 "a space cannot begin with a partial obj");
3224 3223
3225 3224 src_space_id = SpaceId(space_id);
3226 3225 src_space_top = space->top();
3227 3226 const size_t src_region_idx = sd.region(src_cp);
3228 3227 closure.set_source(sd.region_to_addr(src_region_idx));
3229 3228 return src_region_idx;
3230 3229 } else {
3231 3230 assert(src_cp->data_size() == 0, "sanity");
3232 3231 }
3233 3232 }
3234 3233 }
3235 3234 } while (++space_id < last_space_id);
3236 3235
3237 3236 assert(false, "no source region was found");
3238 3237 return 0;
3239 3238 }
3240 3239
3241 3240 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3242 3241 {
3243 3242 typedef ParMarkBitMap::IterationStatus IterationStatus;
3244 3243 const size_t RegionSize = ParallelCompactData::RegionSize;
3245 3244 ParMarkBitMap* const bitmap = mark_bitmap();
3246 3245 ParallelCompactData& sd = summary_data();
3247 3246 RegionData* const region_ptr = sd.region(region_idx);
3248 3247
3249 3248 // Get the items needed to construct the closure.
3250 3249 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3251 3250 SpaceId dest_space_id = space_id(dest_addr);
3252 3251 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3253 3252 HeapWord* new_top = _space_info[dest_space_id].new_top();
3254 3253 assert(dest_addr < new_top, "sanity");
3255 3254 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3256 3255
3257 3256 // Get the source region and related info.
3258 3257 size_t src_region_idx = region_ptr->source_region();
3259 3258 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3260 3259 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3261 3260
3262 3261 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3263 3262 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3264 3263
3265 3264 // Adjust src_region_idx to prepare for decrementing destination counts (the
3266 3265 // destination count is not decremented when a region is copied to itself).
3267 3266 if (src_region_idx == region_idx) {
3268 3267 src_region_idx += 1;
3269 3268 }
3270 3269
3271 3270 if (bitmap->is_unmarked(closure.source())) {
3272 3271 // The first source word is in the middle of an object; copy the remainder
3273 3272 // of the object or as much as will fit. The fact that pointer updates were
3274 3273 // deferred will be noted when the object header is processed.
3275 3274 HeapWord* const old_src_addr = closure.source();
3276 3275 closure.copy_partial_obj();
3277 3276 if (closure.is_full()) {
3278 3277 decrement_destination_counts(cm, src_space_id, src_region_idx,
3279 3278 closure.source());
3280 3279 region_ptr->set_deferred_obj_addr(NULL);
3281 3280 region_ptr->set_completed();
3282 3281 return;
3283 3282 }
3284 3283
3285 3284 HeapWord* const end_addr = sd.region_align_down(closure.source());
3286 3285 if (sd.region_align_down(old_src_addr) != end_addr) {
3287 3286 // The partial object was copied from more than one source region.
3288 3287 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3289 3288
3290 3289 // Move to the next source region, possibly switching spaces as well. All
3291 3290 // args except end_addr may be modified.
3292 3291 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3293 3292 end_addr);
3294 3293 }
3295 3294 }
3296 3295
3297 3296 do {
3298 3297 HeapWord* const cur_addr = closure.source();
3299 3298 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3300 3299 src_space_top);
3301 3300 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3302 3301
3303 3302 if (status == ParMarkBitMap::incomplete) {
3304 3303 // The last obj that starts in the source region does not end in the
3305 3304 // region.
3306 3305 assert(closure.source() < end_addr, "sanity");
3307 3306 HeapWord* const obj_beg = closure.source();
3308 3307 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3309 3308 src_space_top);
3310 3309 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3311 3310 if (obj_end < range_end) {
3312 3311 // The end was found; the entire object will fit.
3313 3312 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3314 3313 assert(status != ParMarkBitMap::would_overflow, "sanity");
3315 3314 } else {
3316 3315 // The end was not found; the object will not fit.
3317 3316 assert(range_end < src_space_top, "obj cannot cross space boundary");
3318 3317 status = ParMarkBitMap::would_overflow;
3319 3318 }
3320 3319 }
3321 3320
3322 3321 if (status == ParMarkBitMap::would_overflow) {
3323 3322 // The last object did not fit. Note that interior oop updates were
3324 3323 // deferred, then copy enough of the object to fill the region.
3325 3324 region_ptr->set_deferred_obj_addr(closure.destination());
3326 3325 status = closure.copy_until_full(); // copies from closure.source()
3327 3326
3328 3327 decrement_destination_counts(cm, src_space_id, src_region_idx,
3329 3328 closure.source());
3330 3329 region_ptr->set_completed();
3331 3330 return;
3332 3331 }
3333 3332
3334 3333 if (status == ParMarkBitMap::full) {
3335 3334 decrement_destination_counts(cm, src_space_id, src_region_idx,
3336 3335 closure.source());
3337 3336 region_ptr->set_deferred_obj_addr(NULL);
3338 3337 region_ptr->set_completed();
3339 3338 return;
3340 3339 }
3341 3340
3342 3341 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3343 3342
3344 3343 // Move to the next source region, possibly switching spaces as well. All
3345 3344 // args except end_addr may be modified.
3346 3345 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3347 3346 end_addr);
3348 3347 } while (true);
3349 3348 }
3350 3349
3351 3350 void
3352 3351 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3353 3352 const MutableSpace* sp = space(space_id);
3354 3353 if (sp->is_empty()) {
3355 3354 return;
3356 3355 }
3357 3356
3358 3357 ParallelCompactData& sd = PSParallelCompact::summary_data();
3359 3358 ParMarkBitMap* const bitmap = mark_bitmap();
3360 3359 HeapWord* const dp_addr = dense_prefix(space_id);
3361 3360 HeapWord* beg_addr = sp->bottom();
3362 3361 HeapWord* end_addr = sp->top();
3363 3362
3364 3363 #ifdef ASSERT
3365 3364 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3366 3365 if (cm->should_verify_only()) {
3367 3366 VerifyUpdateClosure verify_update(cm, sp);
3368 3367 bitmap->iterate(&verify_update, beg_addr, end_addr);
3369 3368 return;
3370 3369 }
3371 3370
3372 3371 if (cm->should_reset_only()) {
3373 3372 ResetObjectsClosure reset_objects(cm);
3374 3373 bitmap->iterate(&reset_objects, beg_addr, end_addr);
3375 3374 return;
3376 3375 }
3377 3376 #endif
3378 3377
3379 3378 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3380 3379 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3381 3380 if (beg_region < dp_region) {
3382 3381 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3383 3382 }
3384 3383
3385 3384 // The destination of the first live object that starts in the region is one
3386 3385 // past the end of the partial object entering the region (if any).
3387 3386 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3388 3387 HeapWord* const new_top = _space_info[space_id].new_top();
3389 3388 assert(new_top >= dest_addr, "bad new_top value");
3390 3389 const size_t words = pointer_delta(new_top, dest_addr);
3391 3390
3392 3391 if (words > 0) {
3393 3392 ObjectStartArray* start_array = _space_info[space_id].start_array();
3394 3393 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3395 3394
3396 3395 ParMarkBitMap::IterationStatus status;
3397 3396 status = bitmap->iterate(&closure, dest_addr, end_addr);
3398 3397 assert(status == ParMarkBitMap::full, "iteration not complete");
3399 3398 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3400 3399 "live objects skipped because closure is full");
3401 3400 }
3402 3401 }
3403 3402
3404 3403 jlong PSParallelCompact::millis_since_last_gc() {
3405 3404 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
3406 3405 // XXX See note in genCollectedHeap::millis_since_last_gc().
3407 3406 if (ret_val < 0) {
3408 3407 NOT_PRODUCT(warning("time warp: %d", ret_val);)
3409 3408 return 0;
3410 3409 }
3411 3410 return ret_val;
3412 3411 }
3413 3412
3414 3413 void PSParallelCompact::reset_millis_since_last_gc() {
3415 3414 _time_of_last_gc = os::javaTimeMillis();
3416 3415 }
3417 3416
3418 3417 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3419 3418 {
3420 3419 if (source() != destination()) {
3421 3420 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3422 3421 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3423 3422 }
3424 3423 update_state(words_remaining());
3425 3424 assert(is_full(), "sanity");
3426 3425 return ParMarkBitMap::full;
3427 3426 }
3428 3427
3429 3428 void MoveAndUpdateClosure::copy_partial_obj()
3430 3429 {
3431 3430 size_t words = words_remaining();
3432 3431
3433 3432 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3434 3433 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3435 3434 if (end_addr < range_end) {
3436 3435 words = bitmap()->obj_size(source(), end_addr);
3437 3436 }
3438 3437
3439 3438 // This test is necessary; if omitted, the pointer updates to a partial object
3440 3439 // that crosses the dense prefix boundary could be overwritten.
3441 3440 if (source() != destination()) {
3442 3441 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3443 3442 Copy::aligned_conjoint_words(source(), destination(), words);
3444 3443 }
3445 3444 update_state(words);
3446 3445 }
3447 3446
3448 3447 ParMarkBitMapClosure::IterationStatus
3449 3448 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3450 3449 assert(destination() != NULL, "sanity");
3451 3450 assert(bitmap()->obj_size(addr) == words, "bad size");
3452 3451
3453 3452 _source = addr;
3454 3453 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3455 3454 destination(), "wrong destination");
3456 3455
3457 3456 if (words > words_remaining()) {
3458 3457 return ParMarkBitMap::would_overflow;
3459 3458 }
3460 3459
3461 3460 // The start_array must be updated even if the object is not moving.
3462 3461 if (_start_array != NULL) {
3463 3462 _start_array->allocate_block(destination());
3464 3463 }
3465 3464
3466 3465 if (destination() != source()) {
3467 3466 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3468 3467 Copy::aligned_conjoint_words(source(), destination(), words);
3469 3468 }
3470 3469
3471 3470 oop moved_oop = (oop) destination();
3472 3471 moved_oop->update_contents(compaction_manager());
3473 3472 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3474 3473
3475 3474 update_state(words);
3476 3475 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3477 3476 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3478 3477 }
3479 3478
3480 3479 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3481 3480 ParCompactionManager* cm,
3482 3481 PSParallelCompact::SpaceId space_id) :
3483 3482 ParMarkBitMapClosure(mbm, cm),
3484 3483 _space_id(space_id),
3485 3484 _start_array(PSParallelCompact::start_array(space_id))
3486 3485 {
3487 3486 }
3488 3487
3489 3488 // Updates the references in the object to their new values.
3490 3489 ParMarkBitMapClosure::IterationStatus
3491 3490 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3492 3491 do_addr(addr);
3493 3492 return ParMarkBitMap::incomplete;
3494 3493 }
3495 3494
3496 3495 // Verify the new location using the forwarding pointer
3497 3496 // from MarkSweep::mark_sweep_phase2(). Set the mark_word
3498 3497 // to the initial value.
3499 3498 ParMarkBitMapClosure::IterationStatus
3500 3499 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3501 3500 // The second arg (words) is not used.
3502 3501 oop obj = (oop) addr;
3503 3502 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
3504 3503 HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
3505 3504 if (forwarding_ptr == NULL) {
3506 3505 // The object is dead or not moving.
3507 3506 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
3508 3507 "Object liveness is wrong.");
3509 3508 return ParMarkBitMap::incomplete;
3510 3509 }
3511 3510 assert(HeapMaximumCompactionInterval > 1 || MarkSweepAlwaysCompactCount > 1 ||
3512 3511 forwarding_ptr == new_pointer, "new location is incorrect");
3513 3512 return ParMarkBitMap::incomplete;
3514 3513 }
3515 3514
3516 3515 // Reset objects modified for debug checking.
3517 3516 ParMarkBitMapClosure::IterationStatus
3518 3517 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
3519 3518 // The second arg (words) is not used.
3520 3519 oop obj = (oop) addr;
3521 3520 obj->init_mark();
3522 3521 return ParMarkBitMap::incomplete;
3523 3522 }
3524 3523
3525 3524 // Prepare for compaction. This method is executed once
3526 3525 // (i.e., by a single thread) before compaction.
3527 3526 // Save the updated location of the intArrayKlassObj for
3528 3527 // filling holes in the dense prefix.
3529 3528 void PSParallelCompact::compact_prologue() {
3530 3529 _updated_int_array_klass_obj = (klassOop)
3531 3530 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3532 3531 }
3533 3532
↓ open down ↓ |
1448 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX