1 /*
2 * Copyright (c) 2005, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/stringTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc/parallel/gcTaskManager.hpp"
30 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
31 #include "gc/parallel/pcTasks.hpp"
32 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
33 #include "gc/parallel/psCompactionManager.inline.hpp"
34 #include "gc/parallel/psMarkSweep.hpp"
35 #include "gc/parallel/psMarkSweepDecorator.hpp"
36 #include "gc/parallel/psOldGen.hpp"
37 #include "gc/parallel/psParallelCompact.inline.hpp"
38 #include "gc/parallel/psPromotionManager.inline.hpp"
39 #include "gc/parallel/psScavenge.hpp"
40 #include "gc/parallel/psYoungGen.hpp"
41 #include "gc/shared/gcCause.hpp"
42 #include "gc/shared/gcHeapSummary.hpp"
43 #include "gc/shared/gcLocker.inline.hpp"
44 #include "gc/shared/gcTimer.hpp"
45 #include "gc/shared/gcTrace.hpp"
46 #include "gc/shared/gcTraceTime.hpp"
47 #include "gc/shared/isGCActiveMark.hpp"
48 #include "gc/shared/referencePolicy.hpp"
49 #include "gc/shared/referenceProcessor.hpp"
50 #include "gc/shared/spaceDecorator.hpp"
51 #include "oops/instanceKlass.inline.hpp"
52 #include "oops/instanceMirrorKlass.inline.hpp"
53 #include "oops/methodData.hpp"
54 #include "oops/objArrayKlass.inline.hpp"
55 #include "oops/oop.inline.hpp"
56 #include "runtime/atomic.inline.hpp"
57 #include "runtime/fprofiler.hpp"
58 #include "runtime/safepoint.hpp"
59 #include "runtime/vmThread.hpp"
60 #include "services/management.hpp"
61 #include "services/memTracker.hpp"
62 #include "services/memoryService.hpp"
63 #include "utilities/events.hpp"
64 #include "utilities/stack.inline.hpp"
65
66 #include <math.h>
67
68 // All sizes are in HeapWords.
69 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
70 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
71 const size_t ParallelCompactData::RegionSizeBytes =
72 RegionSize << LogHeapWordSize;
73 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
74 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
75 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
76
77 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
78 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
79 const size_t ParallelCompactData::BlockSizeBytes =
80 BlockSize << LogHeapWordSize;
81 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
82 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
83 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
84
85 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
86 const size_t ParallelCompactData::Log2BlocksPerRegion =
87 Log2RegionSize - Log2BlockSize;
88
89 const ParallelCompactData::RegionData::region_sz_t
90 ParallelCompactData::RegionData::dc_shift = 27;
91
92 const ParallelCompactData::RegionData::region_sz_t
93 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
94
95 const ParallelCompactData::RegionData::region_sz_t
96 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
97
98 const ParallelCompactData::RegionData::region_sz_t
99 ParallelCompactData::RegionData::los_mask = ~dc_mask;
100
101 const ParallelCompactData::RegionData::region_sz_t
102 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
103
104 const ParallelCompactData::RegionData::region_sz_t
105 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
106
107 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
108 bool PSParallelCompact::_print_phases = false;
109
110 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
111
112 double PSParallelCompact::_dwl_mean;
113 double PSParallelCompact::_dwl_std_dev;
114 double PSParallelCompact::_dwl_first_term;
115 double PSParallelCompact::_dwl_adjustment;
116 #ifdef ASSERT
117 bool PSParallelCompact::_dwl_initialized = false;
118 #endif // #ifdef ASSERT
119
120 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
121 HeapWord* destination)
122 {
123 assert(src_region_idx != 0, "invalid src_region_idx");
124 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
125 assert(destination != NULL, "invalid destination argument");
126
127 _src_region_idx = src_region_idx;
128 _partial_obj_size = partial_obj_size;
129 _destination = destination;
130
131 // These fields may not be updated below, so make sure they're clear.
132 assert(_dest_region_addr == NULL, "should have been cleared");
133 assert(_first_src_addr == NULL, "should have been cleared");
134
135 // Determine the number of destination regions for the partial object.
136 HeapWord* const last_word = destination + partial_obj_size - 1;
137 const ParallelCompactData& sd = PSParallelCompact::summary_data();
138 HeapWord* const beg_region_addr = sd.region_align_down(destination);
139 HeapWord* const end_region_addr = sd.region_align_down(last_word);
140
141 if (beg_region_addr == end_region_addr) {
142 // One destination region.
143 _destination_count = 1;
144 if (end_region_addr == destination) {
145 // The destination falls on a region boundary, thus the first word of the
146 // partial object will be the first word copied to the destination region.
147 _dest_region_addr = end_region_addr;
148 _first_src_addr = sd.region_to_addr(src_region_idx);
149 }
150 } else {
151 // Two destination regions. When copied, the partial object will cross a
152 // destination region boundary, so a word somewhere within the partial
153 // object will be the first word copied to the second destination region.
154 _destination_count = 2;
155 _dest_region_addr = end_region_addr;
156 const size_t ofs = pointer_delta(end_region_addr, destination);
157 assert(ofs < _partial_obj_size, "sanity");
158 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
159 }
160 }
161
162 void SplitInfo::clear()
163 {
164 _src_region_idx = 0;
165 _partial_obj_size = 0;
166 _destination = NULL;
167 _destination_count = 0;
168 _dest_region_addr = NULL;
169 _first_src_addr = NULL;
170 assert(!is_valid(), "sanity");
171 }
172
173 #ifdef ASSERT
174 void SplitInfo::verify_clear()
175 {
176 assert(_src_region_idx == 0, "not clear");
177 assert(_partial_obj_size == 0, "not clear");
178 assert(_destination == NULL, "not clear");
179 assert(_destination_count == 0, "not clear");
180 assert(_dest_region_addr == NULL, "not clear");
181 assert(_first_src_addr == NULL, "not clear");
182 }
183 #endif // #ifdef ASSERT
184
185
186 void PSParallelCompact::print_on_error(outputStream* st) {
187 _mark_bitmap.print_on_error(st);
188 }
189
190 #ifndef PRODUCT
191 const char* PSParallelCompact::space_names[] = {
192 "old ", "eden", "from", "to "
193 };
194
195 void PSParallelCompact::print_region_ranges()
196 {
197 tty->print_cr("space bottom top end new_top");
198 tty->print_cr("------ ---------- ---------- ---------- ----------");
199
200 for (unsigned int id = 0; id < last_space_id; ++id) {
201 const MutableSpace* space = _space_info[id].space();
202 tty->print_cr("%u %s "
203 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
205 id, space_names[id],
206 summary_data().addr_to_region_idx(space->bottom()),
207 summary_data().addr_to_region_idx(space->top()),
208 summary_data().addr_to_region_idx(space->end()),
209 summary_data().addr_to_region_idx(_space_info[id].new_top()));
210 }
211 }
212
213 void
214 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
215 {
216 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
217 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
218
219 ParallelCompactData& sd = PSParallelCompact::summary_data();
220 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
221 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
222 REGION_IDX_FORMAT " " PTR_FORMAT " "
223 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
224 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
225 i, p2i(c->data_location()), dci, p2i(c->destination()),
226 c->partial_obj_size(), c->live_obj_size(),
227 c->data_size(), c->source_region(), c->destination_count());
228
229 #undef REGION_IDX_FORMAT
230 #undef REGION_DATA_FORMAT
231 }
232
233 void
234 print_generic_summary_data(ParallelCompactData& summary_data,
235 HeapWord* const beg_addr,
236 HeapWord* const end_addr)
237 {
238 size_t total_words = 0;
239 size_t i = summary_data.addr_to_region_idx(beg_addr);
240 const size_t last = summary_data.addr_to_region_idx(end_addr);
241 HeapWord* pdest = 0;
242
243 while (i <= last) {
244 ParallelCompactData::RegionData* c = summary_data.region(i);
245 if (c->data_size() != 0 || c->destination() != pdest) {
246 print_generic_summary_region(i, c);
247 total_words += c->data_size();
248 pdest = c->destination();
249 }
250 ++i;
251 }
252
253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
254 }
255
256 void
257 print_generic_summary_data(ParallelCompactData& summary_data,
258 SpaceInfo* space_info)
259 {
260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
261 const MutableSpace* space = space_info[id].space();
262 print_generic_summary_data(summary_data, space->bottom(),
263 MAX2(space->top(), space_info[id].new_top()));
264 }
265 }
266
267 void
268 print_initial_summary_region(size_t i,
269 const ParallelCompactData::RegionData* c,
270 bool newline = true)
271 {
272 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
273 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
275 i, p2i(c->destination()),
276 c->partial_obj_size(), c->live_obj_size(),
277 c->data_size(), c->source_region(), c->destination_count());
278 if (newline) tty->cr();
279 }
280
281 void
282 print_initial_summary_data(ParallelCompactData& summary_data,
283 const MutableSpace* space) {
284 if (space->top() == space->bottom()) {
285 return;
286 }
287
288 const size_t region_size = ParallelCompactData::RegionSize;
289 typedef ParallelCompactData::RegionData RegionData;
290 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
291 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
292 const RegionData* c = summary_data.region(end_region - 1);
293 HeapWord* end_addr = c->destination() + c->data_size();
294 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
295
296 // Print (and count) the full regions at the beginning of the space.
297 size_t full_region_count = 0;
298 size_t i = summary_data.addr_to_region_idx(space->bottom());
299 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
300 print_initial_summary_region(i, summary_data.region(i));
301 ++full_region_count;
302 ++i;
303 }
304
305 size_t live_to_right = live_in_space - full_region_count * region_size;
306
307 double max_reclaimed_ratio = 0.0;
308 size_t max_reclaimed_ratio_region = 0;
309 size_t max_dead_to_right = 0;
310 size_t max_live_to_right = 0;
311
312 // Print the 'reclaimed ratio' for regions while there is something live in
313 // the region or to the right of it. The remaining regions are empty (and
314 // uninteresting), and computing the ratio will result in division by 0.
315 while (i < end_region && live_to_right > 0) {
316 c = summary_data.region(i);
317 HeapWord* const region_addr = summary_data.region_to_addr(i);
318 const size_t used_to_right = pointer_delta(space->top(), region_addr);
319 const size_t dead_to_right = used_to_right - live_to_right;
320 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
321
322 if (reclaimed_ratio > max_reclaimed_ratio) {
323 max_reclaimed_ratio = reclaimed_ratio;
324 max_reclaimed_ratio_region = i;
325 max_dead_to_right = dead_to_right;
326 max_live_to_right = live_to_right;
327 }
328
329 print_initial_summary_region(i, c, false);
330 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
331 reclaimed_ratio, dead_to_right, live_to_right);
332
333 live_to_right -= c->data_size();
334 ++i;
335 }
336
337 // Any remaining regions are empty. Print one more if there is one.
338 if (i < end_region) {
339 print_initial_summary_region(i, summary_data.region(i));
340 }
341
342 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
343 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
344 max_reclaimed_ratio_region, max_dead_to_right,
345 max_live_to_right, max_reclaimed_ratio);
346 }
347
348 void
349 print_initial_summary_data(ParallelCompactData& summary_data,
350 SpaceInfo* space_info) {
351 unsigned int id = PSParallelCompact::old_space_id;
352 const MutableSpace* space;
353 do {
354 space = space_info[id].space();
355 print_initial_summary_data(summary_data, space);
356 } while (++id < PSParallelCompact::eden_space_id);
357
358 do {
359 space = space_info[id].space();
360 print_generic_summary_data(summary_data, space->bottom(), space->top());
361 } while (++id < PSParallelCompact::last_space_id);
362 }
363 #endif // #ifndef PRODUCT
364
365 #ifdef ASSERT
366 size_t add_obj_count;
367 size_t add_obj_size;
368 size_t mark_bitmap_count;
369 size_t mark_bitmap_size;
370 #endif // #ifdef ASSERT
371
372 ParallelCompactData::ParallelCompactData()
373 {
374 _region_start = 0;
375
376 _region_vspace = 0;
377 _reserved_byte_size = 0;
378 _region_data = 0;
379 _region_count = 0;
380
381 _block_vspace = 0;
382 _block_data = 0;
383 _block_count = 0;
384 }
385
386 bool ParallelCompactData::initialize(MemRegion covered_region)
387 {
388 _region_start = covered_region.start();
389 const size_t region_size = covered_region.word_size();
390 DEBUG_ONLY(_region_end = _region_start + region_size;)
391
392 assert(region_align_down(_region_start) == _region_start,
393 "region start not aligned");
394 assert((region_size & RegionSizeOffsetMask) == 0,
395 "region size not a multiple of RegionSize");
396
397 bool result = initialize_region_data(region_size) && initialize_block_data();
398 return result;
399 }
400
401 PSVirtualSpace*
402 ParallelCompactData::create_vspace(size_t count, size_t element_size)
403 {
404 const size_t raw_bytes = count * element_size;
405 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
406 const size_t granularity = os::vm_allocation_granularity();
407 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
408
409 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
410 MAX2(page_sz, granularity);
411 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
412 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
413 rs.size());
414
415 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
416
417 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
418 if (vspace != 0) {
419 if (vspace->expand_by(_reserved_byte_size)) {
420 return vspace;
421 }
422 delete vspace;
423 // Release memory reserved in the space.
424 rs.release();
425 }
426
427 return 0;
428 }
429
430 bool ParallelCompactData::initialize_region_data(size_t region_size)
431 {
432 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
433 _region_vspace = create_vspace(count, sizeof(RegionData));
434 if (_region_vspace != 0) {
435 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
436 _region_count = count;
437 return true;
438 }
439 return false;
440 }
441
442 bool ParallelCompactData::initialize_block_data()
443 {
444 assert(_region_count != 0, "region data must be initialized first");
445 const size_t count = _region_count << Log2BlocksPerRegion;
446 _block_vspace = create_vspace(count, sizeof(BlockData));
447 if (_block_vspace != 0) {
448 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
449 _block_count = count;
450 return true;
451 }
452 return false;
453 }
454
455 void ParallelCompactData::clear()
456 {
457 memset(_region_data, 0, _region_vspace->committed_size());
458 memset(_block_data, 0, _block_vspace->committed_size());
459 }
460
461 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
462 assert(beg_region <= _region_count, "beg_region out of range");
463 assert(end_region <= _region_count, "end_region out of range");
464 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
465
466 const size_t region_cnt = end_region - beg_region;
467 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
468
469 const size_t beg_block = beg_region * BlocksPerRegion;
470 const size_t block_cnt = region_cnt * BlocksPerRegion;
471 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
472 }
473
474 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
475 {
476 const RegionData* cur_cp = region(region_idx);
477 const RegionData* const end_cp = region(region_count() - 1);
478
479 HeapWord* result = region_to_addr(region_idx);
480 if (cur_cp < end_cp) {
481 do {
482 result += cur_cp->partial_obj_size();
483 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
484 }
485 return result;
486 }
487
488 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
489 {
490 const size_t obj_ofs = pointer_delta(addr, _region_start);
491 const size_t beg_region = obj_ofs >> Log2RegionSize;
492 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
493
494 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
495 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
496
497 if (beg_region == end_region) {
498 // All in one region.
499 _region_data[beg_region].add_live_obj(len);
500 return;
501 }
502
503 // First region.
504 const size_t beg_ofs = region_offset(addr);
505 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
506
507 Klass* klass = ((oop)addr)->klass();
508 // Middle regions--completely spanned by this object.
509 for (size_t region = beg_region + 1; region < end_region; ++region) {
510 _region_data[region].set_partial_obj_size(RegionSize);
511 _region_data[region].set_partial_obj_addr(addr);
512 }
513
514 // Last region.
515 const size_t end_ofs = region_offset(addr + len - 1);
516 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
517 _region_data[end_region].set_partial_obj_addr(addr);
518 }
519
520 void
521 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
522 {
523 assert(region_offset(beg) == 0, "not RegionSize aligned");
524 assert(region_offset(end) == 0, "not RegionSize aligned");
525
526 size_t cur_region = addr_to_region_idx(beg);
527 const size_t end_region = addr_to_region_idx(end);
528 HeapWord* addr = beg;
529 while (cur_region < end_region) {
530 _region_data[cur_region].set_destination(addr);
531 _region_data[cur_region].set_destination_count(0);
532 _region_data[cur_region].set_source_region(cur_region);
533 _region_data[cur_region].set_data_location(addr);
534
535 // Update live_obj_size so the region appears completely full.
536 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
537 _region_data[cur_region].set_live_obj_size(live_size);
538
539 ++cur_region;
540 addr += RegionSize;
541 }
542 }
543
544 // Find the point at which a space can be split and, if necessary, record the
545 // split point.
546 //
547 // If the current src region (which overflowed the destination space) doesn't
548 // have a partial object, the split point is at the beginning of the current src
549 // region (an "easy" split, no extra bookkeeping required).
550 //
551 // If the current src region has a partial object, the split point is in the
552 // region where that partial object starts (call it the split_region). If
553 // split_region has a partial object, then the split point is just after that
554 // partial object (a "hard" split where we have to record the split data and
555 // zero the partial_obj_size field). With a "hard" split, we know that the
556 // partial_obj ends within split_region because the partial object that caused
557 // the overflow starts in split_region. If split_region doesn't have a partial
558 // obj, then the split is at the beginning of split_region (another "easy"
559 // split).
560 HeapWord*
561 ParallelCompactData::summarize_split_space(size_t src_region,
562 SplitInfo& split_info,
563 HeapWord* destination,
564 HeapWord* target_end,
565 HeapWord** target_next)
566 {
567 assert(destination <= target_end, "sanity");
568 assert(destination + _region_data[src_region].data_size() > target_end,
569 "region should not fit into target space");
570 assert(is_region_aligned(target_end), "sanity");
571
572 size_t split_region = src_region;
573 HeapWord* split_destination = destination;
574 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
575
576 if (destination + partial_obj_size > target_end) {
577 // The split point is just after the partial object (if any) in the
578 // src_region that contains the start of the object that overflowed the
579 // destination space.
580 //
581 // Find the start of the "overflow" object and set split_region to the
582 // region containing it.
583 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
584 split_region = addr_to_region_idx(overflow_obj);
585
586 // Clear the source_region field of all destination regions whose first word
587 // came from data after the split point (a non-null source_region field
588 // implies a region must be filled).
589 //
590 // An alternative to the simple loop below: clear during post_compact(),
591 // which uses memcpy instead of individual stores, and is easy to
592 // parallelize. (The downside is that it clears the entire RegionData
593 // object as opposed to just one field.)
594 //
595 // post_compact() would have to clear the summary data up to the highest
596 // address that was written during the summary phase, which would be
597 //
598 // max(top, max(new_top, clear_top))
599 //
600 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
601 // to target_end.
602 const RegionData* const sr = region(split_region);
603 const size_t beg_idx =
604 addr_to_region_idx(region_align_up(sr->destination() +
605 sr->partial_obj_size()));
606 const size_t end_idx = addr_to_region_idx(target_end);
607
608 if (TraceParallelOldGCSummaryPhase) {
609 gclog_or_tty->print_cr("split: clearing source_region field in ["
610 SIZE_FORMAT ", " SIZE_FORMAT ")",
611 beg_idx, end_idx);
612 }
613 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
614 _region_data[idx].set_source_region(0);
615 }
616
617 // Set split_destination and partial_obj_size to reflect the split region.
618 split_destination = sr->destination();
619 partial_obj_size = sr->partial_obj_size();
620 }
621
622 // The split is recorded only if a partial object extends onto the region.
623 if (partial_obj_size != 0) {
624 _region_data[split_region].set_partial_obj_size(0);
625 split_info.record(split_region, partial_obj_size, split_destination);
626 }
627
628 // Setup the continuation addresses.
629 *target_next = split_destination + partial_obj_size;
630 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
631
632 if (TraceParallelOldGCSummaryPhase) {
633 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
634 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
635 " pos=" SIZE_FORMAT,
636 split_type, p2i(source_next), split_region,
637 partial_obj_size);
638 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
639 " tn=" PTR_FORMAT,
640 split_type, p2i(split_destination),
641 addr_to_region_idx(split_destination),
642 p2i(*target_next));
643
644 if (partial_obj_size != 0) {
645 HeapWord* const po_beg = split_info.destination();
646 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
647 gclog_or_tty->print_cr("%s split: "
648 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
649 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
650 split_type,
651 p2i(po_beg), addr_to_region_idx(po_beg),
652 p2i(po_end), addr_to_region_idx(po_end));
653 }
654 }
655
656 return source_next;
657 }
658
659 bool ParallelCompactData::summarize(SplitInfo& split_info,
660 HeapWord* source_beg, HeapWord* source_end,
661 HeapWord** source_next,
662 HeapWord* target_beg, HeapWord* target_end,
663 HeapWord** target_next)
664 {
665 if (TraceParallelOldGCSummaryPhase) {
666 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
667 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
668 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
669 p2i(source_beg), p2i(source_end), p2i(source_next_val),
670 p2i(target_beg), p2i(target_end), p2i(*target_next));
671 }
672
673 size_t cur_region = addr_to_region_idx(source_beg);
674 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
675
676 HeapWord *dest_addr = target_beg;
677 while (cur_region < end_region) {
678 // The destination must be set even if the region has no data.
679 _region_data[cur_region].set_destination(dest_addr);
680
681 size_t words = _region_data[cur_region].data_size();
682 if (words > 0) {
683 // If cur_region does not fit entirely into the target space, find a point
684 // at which the source space can be 'split' so that part is copied to the
685 // target space and the rest is copied elsewhere.
686 if (dest_addr + words > target_end) {
687 assert(source_next != NULL, "source_next is NULL when splitting");
688 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
689 target_end, target_next);
690 return false;
691 }
692
693 // Compute the destination_count for cur_region, and if necessary, update
694 // source_region for a destination region. The source_region field is
695 // updated if cur_region is the first (left-most) region to be copied to a
696 // destination region.
697 //
698 // The destination_count calculation is a bit subtle. A region that has
699 // data that compacts into itself does not count itself as a destination.
700 // This maintains the invariant that a zero count means the region is
701 // available and can be claimed and then filled.
702 uint destination_count = 0;
703 if (split_info.is_split(cur_region)) {
704 // The current region has been split: the partial object will be copied
705 // to one destination space and the remaining data will be copied to
706 // another destination space. Adjust the initial destination_count and,
707 // if necessary, set the source_region field if the partial object will
708 // cross a destination region boundary.
709 destination_count = split_info.destination_count();
710 if (destination_count == 2) {
711 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
712 _region_data[dest_idx].set_source_region(cur_region);
713 }
714 }
715
716 HeapWord* const last_addr = dest_addr + words - 1;
717 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
718 const size_t dest_region_2 = addr_to_region_idx(last_addr);
719
720 // Initially assume that the destination regions will be the same and
721 // adjust the value below if necessary. Under this assumption, if
722 // cur_region == dest_region_2, then cur_region will be compacted
723 // completely into itself.
724 destination_count += cur_region == dest_region_2 ? 0 : 1;
725 if (dest_region_1 != dest_region_2) {
726 // Destination regions differ; adjust destination_count.
727 destination_count += 1;
728 // Data from cur_region will be copied to the start of dest_region_2.
729 _region_data[dest_region_2].set_source_region(cur_region);
730 } else if (region_offset(dest_addr) == 0) {
731 // Data from cur_region will be copied to the start of the destination
732 // region.
733 _region_data[dest_region_1].set_source_region(cur_region);
734 }
735
736 _region_data[cur_region].set_destination_count(destination_count);
737 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
738 dest_addr += words;
739 }
740
741 ++cur_region;
742 }
743
744 *target_next = dest_addr;
745 return true;
746 }
747
748 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
749 assert(addr != NULL, "Should detect NULL oop earlier");
750 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
751 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
752
753 // Region covering the object.
754 RegionData* const region_ptr = addr_to_region_ptr(addr);
755 HeapWord* result = region_ptr->destination();
756
757 // If the entire Region is live, the new location is region->destination + the
758 // offset of the object within in the Region.
759
760 // Run some performance tests to determine if this special case pays off. It
761 // is worth it for pointers into the dense prefix. If the optimization to
762 // avoid pointer updates in regions that only point to the dense prefix is
763 // ever implemented, this should be revisited.
764 if (region_ptr->data_size() == RegionSize) {
765 result += region_offset(addr);
766 return result;
767 }
768
769 // Otherwise, the new location is region->destination + block offset + the
770 // number of live words in the Block that are (a) to the left of addr and (b)
771 // due to objects that start in the Block.
772
773 // Fill in the block table if necessary. This is unsynchronized, so multiple
774 // threads may fill the block table for a region (harmless, since it is
775 // idempotent).
776 if (!region_ptr->blocks_filled()) {
777 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
778 region_ptr->set_blocks_filled();
779 }
780
781 HeapWord* const search_start = block_align_down(addr);
782 const size_t block_offset = addr_to_block_ptr(addr)->offset();
783
784 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
785 const size_t live = bitmap->live_words_in_range(search_start, oop(addr));
786 result += block_offset + live;
787 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
788 return result;
789 }
790
791 #ifdef ASSERT
792 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
793 {
794 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
795 const size_t* const end = (const size_t*)vspace->committed_high_addr();
796 for (const size_t* p = beg; p < end; ++p) {
797 assert(*p == 0, "not zero");
798 }
799 }
800
801 void ParallelCompactData::verify_clear()
802 {
803 verify_clear(_region_vspace);
804 verify_clear(_block_vspace);
805 }
806 #endif // #ifdef ASSERT
807
808 STWGCTimer PSParallelCompact::_gc_timer;
809 ParallelOldTracer PSParallelCompact::_gc_tracer;
810 elapsedTimer PSParallelCompact::_accumulated_time;
811 unsigned int PSParallelCompact::_total_invocations = 0;
812 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
813 jlong PSParallelCompact::_time_of_last_gc = 0;
814 CollectorCounters* PSParallelCompact::_counters = NULL;
815 ParMarkBitMap PSParallelCompact::_mark_bitmap;
816 ParallelCompactData PSParallelCompact::_summary_data;
817
818 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
819
820 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
821
822 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
823 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
824
825 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
826 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
827 }
828
829 void PSParallelCompact::post_initialize() {
830 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
831 MemRegion mr = heap->reserved_region();
832 _ref_processor =
833 new ReferenceProcessor(mr, // span
834 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
835 ParallelGCThreads, // mt processing degree
836 true, // mt discovery
837 ParallelGCThreads, // mt discovery degree
838 true, // atomic_discovery
839 &_is_alive_closure); // non-header is alive closure
840 _counters = new CollectorCounters("PSParallelCompact", 1);
841
842 // Initialize static fields in ParCompactionManager.
843 ParCompactionManager::initialize(mark_bitmap());
844 }
845
846 bool PSParallelCompact::initialize() {
847 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
848 MemRegion mr = heap->reserved_region();
849
850 // Was the old gen get allocated successfully?
851 if (!heap->old_gen()->is_allocated()) {
852 return false;
853 }
854
855 initialize_space_info();
856 initialize_dead_wood_limiter();
857
858 if (!_mark_bitmap.initialize(mr)) {
859 vm_shutdown_during_initialization(
860 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
861 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
862 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
863 return false;
864 }
865
866 if (!_summary_data.initialize(mr)) {
867 vm_shutdown_during_initialization(
868 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
869 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
870 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
871 return false;
872 }
873
874 return true;
875 }
876
877 void PSParallelCompact::initialize_space_info()
878 {
879 memset(&_space_info, 0, sizeof(_space_info));
880
881 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
882 PSYoungGen* young_gen = heap->young_gen();
883
884 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
885 _space_info[eden_space_id].set_space(young_gen->eden_space());
886 _space_info[from_space_id].set_space(young_gen->from_space());
887 _space_info[to_space_id].set_space(young_gen->to_space());
888
889 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
890 }
891
892 void PSParallelCompact::initialize_dead_wood_limiter()
893 {
894 const size_t max = 100;
895 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
896 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
897 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
898 DEBUG_ONLY(_dwl_initialized = true;)
899 _dwl_adjustment = normal_distribution(1.0);
900 }
901
902 // Simple class for storing info about the heap at the start of GC, to be used
903 // after GC for comparison/printing.
904 class PreGCValues {
905 public:
906 PreGCValues() { }
907 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
908
909 void fill(ParallelScavengeHeap* heap) {
910 _heap_used = heap->used();
911 _young_gen_used = heap->young_gen()->used_in_bytes();
912 _old_gen_used = heap->old_gen()->used_in_bytes();
913 _metadata_used = MetaspaceAux::used_bytes();
914 };
915
916 size_t heap_used() const { return _heap_used; }
917 size_t young_gen_used() const { return _young_gen_used; }
918 size_t old_gen_used() const { return _old_gen_used; }
919 size_t metadata_used() const { return _metadata_used; }
920
921 private:
922 size_t _heap_used;
923 size_t _young_gen_used;
924 size_t _old_gen_used;
925 size_t _metadata_used;
926 };
927
928 void
929 PSParallelCompact::clear_data_covering_space(SpaceId id)
930 {
931 // At this point, top is the value before GC, new_top() is the value that will
932 // be set at the end of GC. The marking bitmap is cleared to top; nothing
933 // should be marked above top. The summary data is cleared to the larger of
934 // top & new_top.
935 MutableSpace* const space = _space_info[id].space();
936 HeapWord* const bot = space->bottom();
937 HeapWord* const top = space->top();
938 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
939
940 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
941 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
942 _mark_bitmap.clear_range(beg_bit, end_bit);
943
944 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
945 const size_t end_region =
946 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
947 _summary_data.clear_range(beg_region, end_region);
948
949 // Clear the data used to 'split' regions.
950 SplitInfo& split_info = _space_info[id].split_info();
951 if (split_info.is_valid()) {
952 split_info.clear();
953 }
954 DEBUG_ONLY(split_info.verify_clear();)
955 }
956
957 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
958 {
959 // Update the from & to space pointers in space_info, since they are swapped
960 // at each young gen gc. Do the update unconditionally (even though a
961 // promotion failure does not swap spaces) because an unknown number of young
962 // collections will have swapped the spaces an unknown number of times.
963 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
964 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
965 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
966 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
967
968 pre_gc_values->fill(heap);
969
970 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
971 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
972
973 // Increment the invocation count
974 heap->increment_total_collections(true);
975
976 // We need to track unique mark sweep invocations as well.
977 _total_invocations++;
978
979 heap->print_heap_before_gc();
980 heap->trace_heap_before_gc(&_gc_tracer);
981
982 // Fill in TLABs
983 heap->accumulate_statistics_all_tlabs();
984 heap->ensure_parsability(true); // retire TLABs
985
986 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
987 HandleMark hm; // Discard invalid handles created during verification
988 Universe::verify(" VerifyBeforeGC:");
989 }
990
991 // Verify object start arrays
992 if (VerifyObjectStartArray &&
993 VerifyBeforeGC) {
994 heap->old_gen()->verify_object_start_array();
995 }
996
997 DEBUG_ONLY(mark_bitmap()->verify_clear();)
998 DEBUG_ONLY(summary_data().verify_clear();)
999
1000 // Have worker threads release resources the next time they run a task.
1001 gc_task_manager()->release_all_resources();
1002 }
1003
1004 void PSParallelCompact::post_compact()
1005 {
1006 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
1007
1008 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1009 // Clear the marking bitmap, summary data and split info.
1010 clear_data_covering_space(SpaceId(id));
1011 // Update top(). Must be done after clearing the bitmap and summary data.
1012 _space_info[id].publish_new_top();
1013 }
1014
1015 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1016 MutableSpace* const from_space = _space_info[from_space_id].space();
1017 MutableSpace* const to_space = _space_info[to_space_id].space();
1018
1019 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1020 bool eden_empty = eden_space->is_empty();
1021 if (!eden_empty) {
1022 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1023 heap->young_gen(), heap->old_gen());
1024 }
1025
1026 // Update heap occupancy information which is used as input to the soft ref
1027 // clearing policy at the next gc.
1028 Universe::update_heap_info_at_gc();
1029
1030 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1031 to_space->is_empty();
1032
1033 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set());
1034 MemRegion old_mr = heap->old_gen()->reserved();
1035 if (young_gen_empty) {
1036 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1037 } else {
1038 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1039 }
1040
1041 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1042 ClassLoaderDataGraph::purge();
1043 MetaspaceAux::verify_metrics();
1044
1045 CodeCache::gc_epilogue();
1046 JvmtiExport::gc_epilogue();
1047
1048 #if defined(COMPILER2) || INCLUDE_JVMCI
1049 DerivedPointerTable::update_pointers();
1050 #endif
1051
1052 ref_processor()->enqueue_discovered_references(NULL);
1053
1054 if (ZapUnusedHeapArea) {
1055 heap->gen_mangle_unused_area();
1056 }
1057
1058 // Update time of last GC
1059 reset_millis_since_last_gc();
1060 }
1061
1062 HeapWord*
1063 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1064 bool maximum_compaction)
1065 {
1066 const size_t region_size = ParallelCompactData::RegionSize;
1067 const ParallelCompactData& sd = summary_data();
1068
1069 const MutableSpace* const space = _space_info[id].space();
1070 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1071 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1072 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1073
1074 // Skip full regions at the beginning of the space--they are necessarily part
1075 // of the dense prefix.
1076 size_t full_count = 0;
1077 const RegionData* cp;
1078 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1079 ++full_count;
1080 }
1081
1082 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1083 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1084 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1085 if (maximum_compaction || cp == end_cp || interval_ended) {
1086 _maximum_compaction_gc_num = total_invocations();
1087 return sd.region_to_addr(cp);
1088 }
1089
1090 HeapWord* const new_top = _space_info[id].new_top();
1091 const size_t space_live = pointer_delta(new_top, space->bottom());
1092 const size_t space_used = space->used_in_words();
1093 const size_t space_capacity = space->capacity_in_words();
1094
1095 const double cur_density = double(space_live) / space_capacity;
1096 const double deadwood_density =
1097 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1098 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1099
1100 if (TraceParallelOldGCDensePrefix) {
1101 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1102 cur_density, deadwood_density, deadwood_goal);
1103 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1104 "space_cap=" SIZE_FORMAT,
1105 space_live, space_used,
1106 space_capacity);
1107 }
1108
1109 // XXX - Use binary search?
1110 HeapWord* dense_prefix = sd.region_to_addr(cp);
1111 const RegionData* full_cp = cp;
1112 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1113 while (cp < end_cp) {
1114 HeapWord* region_destination = cp->destination();
1115 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1116 if (TraceParallelOldGCDensePrefix && Verbose) {
1117 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1118 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1119 sd.region(cp), p2i(region_destination),
1120 p2i(dense_prefix), cur_deadwood);
1121 }
1122
1123 if (cur_deadwood >= deadwood_goal) {
1124 // Found the region that has the correct amount of deadwood to the left.
1125 // This typically occurs after crossing a fairly sparse set of regions, so
1126 // iterate backwards over those sparse regions, looking for the region
1127 // that has the lowest density of live objects 'to the right.'
1128 size_t space_to_left = sd.region(cp) * region_size;
1129 size_t live_to_left = space_to_left - cur_deadwood;
1130 size_t space_to_right = space_capacity - space_to_left;
1131 size_t live_to_right = space_live - live_to_left;
1132 double density_to_right = double(live_to_right) / space_to_right;
1133 while (cp > full_cp) {
1134 --cp;
1135 const size_t prev_region_live_to_right = live_to_right -
1136 cp->data_size();
1137 const size_t prev_region_space_to_right = space_to_right + region_size;
1138 double prev_region_density_to_right =
1139 double(prev_region_live_to_right) / prev_region_space_to_right;
1140 if (density_to_right <= prev_region_density_to_right) {
1141 return dense_prefix;
1142 }
1143 if (TraceParallelOldGCDensePrefix && Verbose) {
1144 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1145 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1146 prev_region_density_to_right);
1147 }
1148 dense_prefix -= region_size;
1149 live_to_right = prev_region_live_to_right;
1150 space_to_right = prev_region_space_to_right;
1151 density_to_right = prev_region_density_to_right;
1152 }
1153 return dense_prefix;
1154 }
1155
1156 dense_prefix += region_size;
1157 ++cp;
1158 }
1159
1160 return dense_prefix;
1161 }
1162
1163 #ifndef PRODUCT
1164 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1165 const SpaceId id,
1166 const bool maximum_compaction,
1167 HeapWord* const addr)
1168 {
1169 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1170 RegionData* const cp = summary_data().region(region_idx);
1171 const MutableSpace* const space = _space_info[id].space();
1172 HeapWord* const new_top = _space_info[id].new_top();
1173
1174 const size_t space_live = pointer_delta(new_top, space->bottom());
1175 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1176 const size_t space_cap = space->capacity_in_words();
1177 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1178 const size_t live_to_right = new_top - cp->destination();
1179 const size_t dead_to_right = space->top() - addr - live_to_right;
1180
1181 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1182 "spl=" SIZE_FORMAT " "
1183 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1184 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1185 " ratio=%10.8f",
1186 algorithm, p2i(addr), region_idx,
1187 space_live,
1188 dead_to_left, dead_to_left_pct,
1189 dead_to_right, live_to_right,
1190 double(dead_to_right) / live_to_right);
1191 }
1192 #endif // #ifndef PRODUCT
1193
1194 // Return a fraction indicating how much of the generation can be treated as
1195 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1196 // based on the density of live objects in the generation to determine a limit,
1197 // which is then adjusted so the return value is min_percent when the density is
1198 // 1.
1199 //
1200 // The following table shows some return values for a different values of the
1201 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1202 // min_percent is 1.
1203 //
1204 // fraction allowed as dead wood
1205 // -----------------------------------------------------------------
1206 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1207 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1208 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1209 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1210 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1211 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1212 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1213 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1214 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1215 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1216 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1217 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1218 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1219 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1220 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1221 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1222 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1223 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1224 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1225 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1226 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1227 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1228 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1229
1230 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1231 {
1232 assert(_dwl_initialized, "uninitialized");
1233
1234 // The raw limit is the value of the normal distribution at x = density.
1235 const double raw_limit = normal_distribution(density);
1236
1237 // Adjust the raw limit so it becomes the minimum when the density is 1.
1238 //
1239 // First subtract the adjustment value (which is simply the precomputed value
1240 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1241 // Then add the minimum value, so the minimum is returned when the density is
1242 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1243 const double min = double(min_percent) / 100.0;
1244 const double limit = raw_limit - _dwl_adjustment + min;
1245 return MAX2(limit, 0.0);
1246 }
1247
1248 ParallelCompactData::RegionData*
1249 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1250 const RegionData* end)
1251 {
1252 const size_t region_size = ParallelCompactData::RegionSize;
1253 ParallelCompactData& sd = summary_data();
1254 size_t left = sd.region(beg);
1255 size_t right = end > beg ? sd.region(end) - 1 : left;
1256
1257 // Binary search.
1258 while (left < right) {
1259 // Equivalent to (left + right) / 2, but does not overflow.
1260 const size_t middle = left + (right - left) / 2;
1261 RegionData* const middle_ptr = sd.region(middle);
1262 HeapWord* const dest = middle_ptr->destination();
1263 HeapWord* const addr = sd.region_to_addr(middle);
1264 assert(dest != NULL, "sanity");
1265 assert(dest <= addr, "must move left");
1266
1267 if (middle > left && dest < addr) {
1268 right = middle - 1;
1269 } else if (middle < right && middle_ptr->data_size() == region_size) {
1270 left = middle + 1;
1271 } else {
1272 return middle_ptr;
1273 }
1274 }
1275 return sd.region(left);
1276 }
1277
1278 ParallelCompactData::RegionData*
1279 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1280 const RegionData* end,
1281 size_t dead_words)
1282 {
1283 ParallelCompactData& sd = summary_data();
1284 size_t left = sd.region(beg);
1285 size_t right = end > beg ? sd.region(end) - 1 : left;
1286
1287 // Binary search.
1288 while (left < right) {
1289 // Equivalent to (left + right) / 2, but does not overflow.
1290 const size_t middle = left + (right - left) / 2;
1291 RegionData* const middle_ptr = sd.region(middle);
1292 HeapWord* const dest = middle_ptr->destination();
1293 HeapWord* const addr = sd.region_to_addr(middle);
1294 assert(dest != NULL, "sanity");
1295 assert(dest <= addr, "must move left");
1296
1297 const size_t dead_to_left = pointer_delta(addr, dest);
1298 if (middle > left && dead_to_left > dead_words) {
1299 right = middle - 1;
1300 } else if (middle < right && dead_to_left < dead_words) {
1301 left = middle + 1;
1302 } else {
1303 return middle_ptr;
1304 }
1305 }
1306 return sd.region(left);
1307 }
1308
1309 // The result is valid during the summary phase, after the initial summarization
1310 // of each space into itself, and before final summarization.
1311 inline double
1312 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1313 HeapWord* const bottom,
1314 HeapWord* const top,
1315 HeapWord* const new_top)
1316 {
1317 ParallelCompactData& sd = summary_data();
1318
1319 assert(cp != NULL, "sanity");
1320 assert(bottom != NULL, "sanity");
1321 assert(top != NULL, "sanity");
1322 assert(new_top != NULL, "sanity");
1323 assert(top >= new_top, "summary data problem?");
1324 assert(new_top > bottom, "space is empty; should not be here");
1325 assert(new_top >= cp->destination(), "sanity");
1326 assert(top >= sd.region_to_addr(cp), "sanity");
1327
1328 HeapWord* const destination = cp->destination();
1329 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1330 const size_t compacted_region_live = pointer_delta(new_top, destination);
1331 const size_t compacted_region_used = pointer_delta(top,
1332 sd.region_to_addr(cp));
1333 const size_t reclaimable = compacted_region_used - compacted_region_live;
1334
1335 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1336 return double(reclaimable) / divisor;
1337 }
1338
1339 // Return the address of the end of the dense prefix, a.k.a. the start of the
1340 // compacted region. The address is always on a region boundary.
1341 //
1342 // Completely full regions at the left are skipped, since no compaction can
1343 // occur in those regions. Then the maximum amount of dead wood to allow is
1344 // computed, based on the density (amount live / capacity) of the generation;
1345 // the region with approximately that amount of dead space to the left is
1346 // identified as the limit region. Regions between the last completely full
1347 // region and the limit region are scanned and the one that has the best
1348 // (maximum) reclaimed_ratio() is selected.
1349 HeapWord*
1350 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1351 bool maximum_compaction)
1352 {
1353 if (ParallelOldGCSplitALot) {
1354 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1355 // The value was chosen to provoke splitting a young gen space; use it.
1356 return _space_info[id].dense_prefix();
1357 }
1358 }
1359
1360 const size_t region_size = ParallelCompactData::RegionSize;
1361 const ParallelCompactData& sd = summary_data();
1362
1363 const MutableSpace* const space = _space_info[id].space();
1364 HeapWord* const top = space->top();
1365 HeapWord* const top_aligned_up = sd.region_align_up(top);
1366 HeapWord* const new_top = _space_info[id].new_top();
1367 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1368 HeapWord* const bottom = space->bottom();
1369 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1370 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1371 const RegionData* const new_top_cp =
1372 sd.addr_to_region_ptr(new_top_aligned_up);
1373
1374 // Skip full regions at the beginning of the space--they are necessarily part
1375 // of the dense prefix.
1376 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1377 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1378 space->is_empty(), "no dead space allowed to the left");
1379 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1380 "region must have dead space");
1381
1382 // The gc number is saved whenever a maximum compaction is done, and used to
1383 // determine when the maximum compaction interval has expired. This avoids
1384 // successive max compactions for different reasons.
1385 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1386 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1387 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1388 total_invocations() == HeapFirstMaximumCompactionCount;
1389 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1390 _maximum_compaction_gc_num = total_invocations();
1391 return sd.region_to_addr(full_cp);
1392 }
1393
1394 const size_t space_live = pointer_delta(new_top, bottom);
1395 const size_t space_used = space->used_in_words();
1396 const size_t space_capacity = space->capacity_in_words();
1397
1398 const double density = double(space_live) / double(space_capacity);
1399 const size_t min_percent_free = MarkSweepDeadRatio;
1400 const double limiter = dead_wood_limiter(density, min_percent_free);
1401 const size_t dead_wood_max = space_used - space_live;
1402 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1403 dead_wood_max);
1404
1405 if (TraceParallelOldGCDensePrefix) {
1406 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1407 "space_cap=" SIZE_FORMAT,
1408 space_live, space_used,
1409 space_capacity);
1410 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1411 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1412 density, min_percent_free, limiter,
1413 dead_wood_max, dead_wood_limit);
1414 }
1415
1416 // Locate the region with the desired amount of dead space to the left.
1417 const RegionData* const limit_cp =
1418 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1419
1420 // Scan from the first region with dead space to the limit region and find the
1421 // one with the best (largest) reclaimed ratio.
1422 double best_ratio = 0.0;
1423 const RegionData* best_cp = full_cp;
1424 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1425 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1426 if (tmp_ratio > best_ratio) {
1427 best_cp = cp;
1428 best_ratio = tmp_ratio;
1429 }
1430 }
1431
1432 #if 0
1433 // Something to consider: if the region with the best ratio is 'close to' the
1434 // first region w/free space, choose the first region with free space
1435 // ("first-free"). The first-free region is usually near the start of the
1436 // heap, which means we are copying most of the heap already, so copy a bit
1437 // more to get complete compaction.
1438 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1439 _maximum_compaction_gc_num = total_invocations();
1440 best_cp = full_cp;
1441 }
1442 #endif // #if 0
1443
1444 return sd.region_to_addr(best_cp);
1445 }
1446
1447 #ifndef PRODUCT
1448 void
1449 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1450 size_t words)
1451 {
1452 if (TraceParallelOldGCSummaryPhase) {
1453 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1454 SIZE_FORMAT, p2i(start), p2i(start + words), words);
1455 }
1456
1457 ObjectStartArray* const start_array = _space_info[id].start_array();
1458 CollectedHeap::fill_with_objects(start, words);
1459 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1460 _mark_bitmap.mark_obj(p, words);
1461 _summary_data.add_obj(p, words);
1462 start_array->allocate_block(p);
1463 }
1464 }
1465
1466 void
1467 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1468 {
1469 ParallelCompactData& sd = summary_data();
1470 MutableSpace* space = _space_info[id].space();
1471
1472 // Find the source and destination start addresses.
1473 HeapWord* const src_addr = sd.region_align_down(start);
1474 HeapWord* dst_addr;
1475 if (src_addr < start) {
1476 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1477 } else if (src_addr > space->bottom()) {
1478 // The start (the original top() value) is aligned to a region boundary so
1479 // the associated region does not have a destination. Compute the
1480 // destination from the previous region.
1481 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1482 dst_addr = cp->destination() + cp->data_size();
1483 } else {
1484 // Filling the entire space.
1485 dst_addr = space->bottom();
1486 }
1487 assert(dst_addr != NULL, "sanity");
1488
1489 // Update the summary data.
1490 bool result = _summary_data.summarize(_space_info[id].split_info(),
1491 src_addr, space->top(), NULL,
1492 dst_addr, space->end(),
1493 _space_info[id].new_top_addr());
1494 assert(result, "should not fail: bad filler object size");
1495 }
1496
1497 void
1498 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1499 {
1500 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1501 return;
1502 }
1503
1504 MutableSpace* const space = _space_info[id].space();
1505 if (space->is_empty()) {
1506 HeapWord* b = space->bottom();
1507 HeapWord* t = b + space->capacity_in_words() / 2;
1508 space->set_top(t);
1509 if (ZapUnusedHeapArea) {
1510 space->set_top_for_allocations();
1511 }
1512
1513 size_t min_size = CollectedHeap::min_fill_size();
1514 size_t obj_len = min_size;
1515 while (b + obj_len <= t) {
1516 CollectedHeap::fill_with_object(b, obj_len);
1517 mark_bitmap()->mark_obj(b, obj_len);
1518 summary_data().add_obj(b, obj_len);
1519 b += obj_len;
1520 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1521 }
1522 if (b < t) {
1523 // The loop didn't completely fill to t (top); adjust top downward.
1524 space->set_top(b);
1525 if (ZapUnusedHeapArea) {
1526 space->set_top_for_allocations();
1527 }
1528 }
1529
1530 HeapWord** nta = _space_info[id].new_top_addr();
1531 bool result = summary_data().summarize(_space_info[id].split_info(),
1532 space->bottom(), space->top(), NULL,
1533 space->bottom(), space->end(), nta);
1534 assert(result, "space must fit into itself");
1535 }
1536 }
1537
1538 void
1539 PSParallelCompact::provoke_split(bool & max_compaction)
1540 {
1541 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1542 return;
1543 }
1544
1545 const size_t region_size = ParallelCompactData::RegionSize;
1546 ParallelCompactData& sd = summary_data();
1547
1548 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1549 MutableSpace* const from_space = _space_info[from_space_id].space();
1550 const size_t eden_live = pointer_delta(eden_space->top(),
1551 _space_info[eden_space_id].new_top());
1552 const size_t from_live = pointer_delta(from_space->top(),
1553 _space_info[from_space_id].new_top());
1554
1555 const size_t min_fill_size = CollectedHeap::min_fill_size();
1556 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1557 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1558 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1559 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1560
1561 // Choose the space to split; need at least 2 regions live (or fillable).
1562 SpaceId id;
1563 MutableSpace* space;
1564 size_t live_words;
1565 size_t fill_words;
1566 if (eden_live + eden_fillable >= region_size * 2) {
1567 id = eden_space_id;
1568 space = eden_space;
1569 live_words = eden_live;
1570 fill_words = eden_fillable;
1571 } else if (from_live + from_fillable >= region_size * 2) {
1572 id = from_space_id;
1573 space = from_space;
1574 live_words = from_live;
1575 fill_words = from_fillable;
1576 } else {
1577 return; // Give up.
1578 }
1579 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1580
1581 if (live_words < region_size * 2) {
1582 // Fill from top() to end() w/live objects of mixed sizes.
1583 HeapWord* const fill_start = space->top();
1584 live_words += fill_words;
1585
1586 space->set_top(fill_start + fill_words);
1587 if (ZapUnusedHeapArea) {
1588 space->set_top_for_allocations();
1589 }
1590
1591 HeapWord* cur_addr = fill_start;
1592 while (fill_words > 0) {
1593 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1594 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1595 if (fill_words - cur_size < min_fill_size) {
1596 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1597 }
1598
1599 CollectedHeap::fill_with_object(cur_addr, cur_size);
1600 mark_bitmap()->mark_obj(cur_addr, cur_size);
1601 sd.add_obj(cur_addr, cur_size);
1602
1603 cur_addr += cur_size;
1604 fill_words -= cur_size;
1605 }
1606
1607 summarize_new_objects(id, fill_start);
1608 }
1609
1610 max_compaction = false;
1611
1612 // Manipulate the old gen so that it has room for about half of the live data
1613 // in the target young gen space (live_words / 2).
1614 id = old_space_id;
1615 space = _space_info[id].space();
1616 const size_t free_at_end = space->free_in_words();
1617 const size_t free_target = align_object_size(live_words / 2);
1618 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1619
1620 if (free_at_end >= free_target + min_fill_size) {
1621 // Fill space above top() and set the dense prefix so everything survives.
1622 HeapWord* const fill_start = space->top();
1623 const size_t fill_size = free_at_end - free_target;
1624 space->set_top(space->top() + fill_size);
1625 if (ZapUnusedHeapArea) {
1626 space->set_top_for_allocations();
1627 }
1628 fill_with_live_objects(id, fill_start, fill_size);
1629 summarize_new_objects(id, fill_start);
1630 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1631 } else if (dead + free_at_end > free_target) {
1632 // Find a dense prefix that makes the right amount of space available.
1633 HeapWord* cur = sd.region_align_down(space->top());
1634 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1635 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1636 while (dead_to_right < free_target) {
1637 cur -= region_size;
1638 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1639 dead_to_right = pointer_delta(space->end(), cur_destination);
1640 }
1641 _space_info[id].set_dense_prefix(cur);
1642 }
1643 }
1644 #endif // #ifndef PRODUCT
1645
1646 void PSParallelCompact::summarize_spaces_quick()
1647 {
1648 for (unsigned int i = 0; i < last_space_id; ++i) {
1649 const MutableSpace* space = _space_info[i].space();
1650 HeapWord** nta = _space_info[i].new_top_addr();
1651 bool result = _summary_data.summarize(_space_info[i].split_info(),
1652 space->bottom(), space->top(), NULL,
1653 space->bottom(), space->end(), nta);
1654 assert(result, "space must fit into itself");
1655 _space_info[i].set_dense_prefix(space->bottom());
1656 }
1657
1658 #ifndef PRODUCT
1659 if (ParallelOldGCSplitALot) {
1660 provoke_split_fill_survivor(to_space_id);
1661 }
1662 #endif // #ifndef PRODUCT
1663 }
1664
1665 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1666 {
1667 HeapWord* const dense_prefix_end = dense_prefix(id);
1668 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1669 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1670 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1671 // Only enough dead space is filled so that any remaining dead space to the
1672 // left is larger than the minimum filler object. (The remainder is filled
1673 // during the copy/update phase.)
1674 //
1675 // The size of the dead space to the right of the boundary is not a
1676 // concern, since compaction will be able to use whatever space is
1677 // available.
1678 //
1679 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1680 // surrounds the space to be filled with an object.
1681 //
1682 // In the 32-bit VM, each bit represents two 32-bit words:
1683 // +---+
1684 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1685 // end_bits: ... x x x | 0 | || 0 x x ...
1686 // +---+
1687 //
1688 // In the 64-bit VM, each bit represents one 64-bit word:
1689 // +------------+
1690 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1691 // end_bits: ... x x 1 | 0 || 0 | x x ...
1692 // +------------+
1693 // +-------+
1694 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1695 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1696 // +-------+
1697 // +-----------+
1698 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1699 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1700 // +-----------+
1701 // +-------+
1702 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1703 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1704 // +-------+
1705
1706 // Initially assume case a, c or e will apply.
1707 size_t obj_len = CollectedHeap::min_fill_size();
1708 HeapWord* obj_beg = dense_prefix_end - obj_len;
1709
1710 #ifdef _LP64
1711 if (MinObjAlignment > 1) { // object alignment > heap word size
1712 // Cases a, c or e.
1713 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1714 // Case b above.
1715 obj_beg = dense_prefix_end - 1;
1716 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1717 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1718 // Case d above.
1719 obj_beg = dense_prefix_end - 3;
1720 obj_len = 3;
1721 }
1722 #endif // #ifdef _LP64
1723
1724 CollectedHeap::fill_with_object(obj_beg, obj_len);
1725 _mark_bitmap.mark_obj(obj_beg, obj_len);
1726 _summary_data.add_obj(obj_beg, obj_len);
1727 assert(start_array(id) != NULL, "sanity");
1728 start_array(id)->allocate_block(obj_beg);
1729 }
1730 }
1731
1732 void
1733 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1734 {
1735 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1736 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1737 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1738 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1739 cur->set_source_region(0);
1740 }
1741 }
1742
1743 void
1744 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1745 {
1746 assert(id < last_space_id, "id out of range");
1747 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1748 ParallelOldGCSplitALot && id == old_space_id,
1749 "should have been reset in summarize_spaces_quick()");
1750
1751 const MutableSpace* space = _space_info[id].space();
1752 if (_space_info[id].new_top() != space->bottom()) {
1753 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1754 _space_info[id].set_dense_prefix(dense_prefix_end);
1755
1756 #ifndef PRODUCT
1757 if (TraceParallelOldGCDensePrefix) {
1758 print_dense_prefix_stats("ratio", id, maximum_compaction,
1759 dense_prefix_end);
1760 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1761 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1762 }
1763 #endif // #ifndef PRODUCT
1764
1765 // Recompute the summary data, taking into account the dense prefix. If
1766 // every last byte will be reclaimed, then the existing summary data which
1767 // compacts everything can be left in place.
1768 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1769 // If dead space crosses the dense prefix boundary, it is (at least
1770 // partially) filled with a dummy object, marked live and added to the
1771 // summary data. This simplifies the copy/update phase and must be done
1772 // before the final locations of objects are determined, to prevent
1773 // leaving a fragment of dead space that is too small to fill.
1774 fill_dense_prefix_end(id);
1775
1776 // Compute the destination of each Region, and thus each object.
1777 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1778 _summary_data.summarize(_space_info[id].split_info(),
1779 dense_prefix_end, space->top(), NULL,
1780 dense_prefix_end, space->end(),
1781 _space_info[id].new_top_addr());
1782 }
1783 }
1784
1785 if (TraceParallelOldGCSummaryPhase) {
1786 const size_t region_size = ParallelCompactData::RegionSize;
1787 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1788 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1789 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1790 HeapWord* const new_top = _space_info[id].new_top();
1791 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1792 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1793 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1794 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1795 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1796 id, space->capacity_in_words(), p2i(dense_prefix_end),
1797 dp_region, dp_words / region_size,
1798 cr_words / region_size, p2i(new_top));
1799 }
1800 }
1801
1802 #ifndef PRODUCT
1803 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1804 HeapWord* dst_beg, HeapWord* dst_end,
1805 SpaceId src_space_id,
1806 HeapWord* src_beg, HeapWord* src_end)
1807 {
1808 if (TraceParallelOldGCSummaryPhase) {
1809 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1810 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1811 SIZE_FORMAT "-" SIZE_FORMAT " "
1812 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1813 SIZE_FORMAT "-" SIZE_FORMAT,
1814 src_space_id, space_names[src_space_id],
1815 dst_space_id, space_names[dst_space_id],
1816 p2i(src_beg), p2i(src_end),
1817 _summary_data.addr_to_region_idx(src_beg),
1818 _summary_data.addr_to_region_idx(src_end),
1819 p2i(dst_beg), p2i(dst_end),
1820 _summary_data.addr_to_region_idx(dst_beg),
1821 _summary_data.addr_to_region_idx(dst_end));
1822 }
1823 }
1824 #endif // #ifndef PRODUCT
1825
1826 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1827 bool maximum_compaction)
1828 {
1829 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
1830 // trace("2");
1831
1832 #ifdef ASSERT
1833 if (TraceParallelOldGCMarkingPhase) {
1834 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1835 "add_obj_bytes=" SIZE_FORMAT,
1836 add_obj_count, add_obj_size * HeapWordSize);
1837 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1838 "mark_bitmap_bytes=" SIZE_FORMAT,
1839 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1840 }
1841 #endif // #ifdef ASSERT
1842
1843 // Quick summarization of each space into itself, to see how much is live.
1844 summarize_spaces_quick();
1845
1846 if (TraceParallelOldGCSummaryPhase) {
1847 tty->print_cr("summary_phase: after summarizing each space to self");
1848 Universe::print();
1849 NOT_PRODUCT(print_region_ranges());
1850 if (Verbose) {
1851 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1852 }
1853 }
1854
1855 // The amount of live data that will end up in old space (assuming it fits).
1856 size_t old_space_total_live = 0;
1857 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1858 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1859 _space_info[id].space()->bottom());
1860 }
1861
1862 MutableSpace* const old_space = _space_info[old_space_id].space();
1863 const size_t old_capacity = old_space->capacity_in_words();
1864 if (old_space_total_live > old_capacity) {
1865 // XXX - should also try to expand
1866 maximum_compaction = true;
1867 }
1868 #ifndef PRODUCT
1869 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1870 provoke_split(maximum_compaction);
1871 }
1872 #endif // #ifndef PRODUCT
1873
1874 // Old generations.
1875 summarize_space(old_space_id, maximum_compaction);
1876
1877 // Summarize the remaining spaces in the young gen. The initial target space
1878 // is the old gen. If a space does not fit entirely into the target, then the
1879 // remainder is compacted into the space itself and that space becomes the new
1880 // target.
1881 SpaceId dst_space_id = old_space_id;
1882 HeapWord* dst_space_end = old_space->end();
1883 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1884 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1885 const MutableSpace* space = _space_info[id].space();
1886 const size_t live = pointer_delta(_space_info[id].new_top(),
1887 space->bottom());
1888 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1889
1890 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1891 SpaceId(id), space->bottom(), space->top());)
1892 if (live > 0 && live <= available) {
1893 // All the live data will fit.
1894 bool done = _summary_data.summarize(_space_info[id].split_info(),
1895 space->bottom(), space->top(),
1896 NULL,
1897 *new_top_addr, dst_space_end,
1898 new_top_addr);
1899 assert(done, "space must fit into old gen");
1900
1901 // Reset the new_top value for the space.
1902 _space_info[id].set_new_top(space->bottom());
1903 } else if (live > 0) {
1904 // Attempt to fit part of the source space into the target space.
1905 HeapWord* next_src_addr = NULL;
1906 bool done = _summary_data.summarize(_space_info[id].split_info(),
1907 space->bottom(), space->top(),
1908 &next_src_addr,
1909 *new_top_addr, dst_space_end,
1910 new_top_addr);
1911 assert(!done, "space should not fit into old gen");
1912 assert(next_src_addr != NULL, "sanity");
1913
1914 // The source space becomes the new target, so the remainder is compacted
1915 // within the space itself.
1916 dst_space_id = SpaceId(id);
1917 dst_space_end = space->end();
1918 new_top_addr = _space_info[id].new_top_addr();
1919 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1920 space->bottom(), dst_space_end,
1921 SpaceId(id), next_src_addr, space->top());)
1922 done = _summary_data.summarize(_space_info[id].split_info(),
1923 next_src_addr, space->top(),
1924 NULL,
1925 space->bottom(), dst_space_end,
1926 new_top_addr);
1927 assert(done, "space must fit when compacted into itself");
1928 assert(*new_top_addr <= space->top(), "usage should not grow");
1929 }
1930 }
1931
1932 if (TraceParallelOldGCSummaryPhase) {
1933 tty->print_cr("summary_phase: after final summarization");
1934 Universe::print();
1935 NOT_PRODUCT(print_region_ranges());
1936 if (Verbose) {
1937 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1938 }
1939 }
1940 }
1941
1942 // This method should contain all heap-specific policy for invoking a full
1943 // collection. invoke_no_policy() will only attempt to compact the heap; it
1944 // will do nothing further. If we need to bail out for policy reasons, scavenge
1945 // before full gc, or any other specialized behavior, it needs to be added here.
1946 //
1947 // Note that this method should only be called from the vm_thread while at a
1948 // safepoint.
1949 //
1950 // Note that the all_soft_refs_clear flag in the collector policy
1951 // may be true because this method can be called without intervening
1952 // activity. For example when the heap space is tight and full measure
1953 // are being taken to free space.
1954 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1955 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1956 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1957 "should be in vm thread");
1958
1959 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1960 GCCause::Cause gc_cause = heap->gc_cause();
1961 assert(!heap->is_gc_active(), "not reentrant");
1962
1963 PSAdaptiveSizePolicy* policy = heap->size_policy();
1964 IsGCActiveMark mark;
1965
1966 if (ScavengeBeforeFullGC) {
1967 PSScavenge::invoke_no_policy();
1968 }
1969
1970 const bool clear_all_soft_refs =
1971 heap->collector_policy()->should_clear_all_soft_refs();
1972
1973 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1974 maximum_heap_compaction);
1975 }
1976
1977 // This method contains no policy. You should probably
1978 // be calling invoke() instead.
1979 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1980 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1981 assert(ref_processor() != NULL, "Sanity");
1982
1983 if (GC_locker::check_active_before_gc()) {
1984 return false;
1985 }
1986
1987 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1988
1989 _gc_timer.register_gc_start();
1990 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1991
1992 TimeStamp marking_start;
1993 TimeStamp compaction_start;
1994 TimeStamp collection_exit;
1995
1996 GCCause::Cause gc_cause = heap->gc_cause();
1997 PSYoungGen* young_gen = heap->young_gen();
1998 PSOldGen* old_gen = heap->old_gen();
1999 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2000
2001 // The scope of casr should end after code that can change
2002 // CollectorPolicy::_should_clear_all_soft_refs.
2003 ClearedAllSoftRefs casr(maximum_heap_compaction,
2004 heap->collector_policy());
2005
2006 if (ZapUnusedHeapArea) {
2007 // Save information needed to minimize mangling
2008 heap->record_gen_tops_before_GC();
2009 }
2010
2011 heap->pre_full_gc_dump(&_gc_timer);
2012
2013 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2014
2015 // Make sure data structures are sane, make the heap parsable, and do other
2016 // miscellaneous bookkeeping.
2017 PreGCValues pre_gc_values;
2018 pre_compact(&pre_gc_values);
2019
2020 // Get the compaction manager reserved for the VM thread.
2021 ParCompactionManager* const vmthread_cm =
2022 ParCompactionManager::manager_array(gc_task_manager()->workers());
2023
2024 // Place after pre_compact() where the number of invocations is incremented.
2025 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2026
2027 {
2028 ResourceMark rm;
2029 HandleMark hm;
2030
2031 // Set the number of GC threads to be used in this collection
2032 gc_task_manager()->set_active_gang();
2033 gc_task_manager()->task_idle_workers();
2034
2035 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2036 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id());
2037 TraceCollectorStats tcs(counters());
2038 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2039
2040 if (TraceOldGenTime) accumulated_time()->start();
2041
2042 // Let the size policy know we're starting
2043 size_policy->major_collection_begin();
2044
2045 CodeCache::gc_prologue();
2046
2047 #if defined(COMPILER2) || INCLUDE_JVMCI
2048 DerivedPointerTable::clear();
2049 #endif
2050
2051 ref_processor()->enable_discovery();
2052 ref_processor()->setup_policy(maximum_heap_compaction);
2053
2054 bool marked_for_unloading = false;
2055
2056 marking_start.update();
2057 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
2058
2059 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2060 && GCCause::is_user_requested_gc(gc_cause);
2061 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2062
2063 #if defined(COMPILER2) || INCLUDE_JVMCI
2064 assert(DerivedPointerTable::is_active(), "Sanity");
2065 DerivedPointerTable::set_active(false);
2066 #endif
2067
2068 // adjust_roots() updates Universe::_intArrayKlassObj which is
2069 // needed by the compaction for filling holes in the dense prefix.
2070 adjust_roots();
2071
2072 compaction_start.update();
2073 compact();
2074
2075 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2076 // done before resizing.
2077 post_compact();
2078
2079 // Let the size policy know we're done
2080 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2081
2082 if (UseAdaptiveSizePolicy) {
2083 if (PrintAdaptiveSizePolicy) {
2084 gclog_or_tty->print("AdaptiveSizeStart: ");
2085 gclog_or_tty->stamp();
2086 gclog_or_tty->print_cr(" collection: %d ",
2087 heap->total_collections());
2088 if (Verbose) {
2089 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT
2090 " young_gen_capacity: " SIZE_FORMAT,
2091 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2092 }
2093 }
2094
2095 // Don't check if the size_policy is ready here. Let
2096 // the size_policy check that internally.
2097 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2098 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
2099 // Swap the survivor spaces if from_space is empty. The
2100 // resize_young_gen() called below is normally used after
2101 // a successful young GC and swapping of survivor spaces;
2102 // otherwise, it will fail to resize the young gen with
2103 // the current implementation.
2104 if (young_gen->from_space()->is_empty()) {
2105 young_gen->from_space()->clear(SpaceDecorator::Mangle);
2106 young_gen->swap_spaces();
2107 }
2108
2109 // Calculate optimal free space amounts
2110 assert(young_gen->max_size() >
2111 young_gen->from_space()->capacity_in_bytes() +
2112 young_gen->to_space()->capacity_in_bytes(),
2113 "Sizes of space in young gen are out-of-bounds");
2114
2115 size_t young_live = young_gen->used_in_bytes();
2116 size_t eden_live = young_gen->eden_space()->used_in_bytes();
2117 size_t old_live = old_gen->used_in_bytes();
2118 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2119 size_t max_old_gen_size = old_gen->max_gen_size();
2120 size_t max_eden_size = young_gen->max_size() -
2121 young_gen->from_space()->capacity_in_bytes() -
2122 young_gen->to_space()->capacity_in_bytes();
2123
2124 // Used for diagnostics
2125 size_policy->clear_generation_free_space_flags();
2126
2127 size_policy->compute_generations_free_space(young_live,
2128 eden_live,
2129 old_live,
2130 cur_eden,
2131 max_old_gen_size,
2132 max_eden_size,
2133 true /* full gc*/);
2134
2135 size_policy->check_gc_overhead_limit(young_live,
2136 eden_live,
2137 max_old_gen_size,
2138 max_eden_size,
2139 true /* full gc*/,
2140 gc_cause,
2141 heap->collector_policy());
2142
2143 size_policy->decay_supplemental_growth(true /* full gc*/);
2144
2145 heap->resize_old_gen(
2146 size_policy->calculated_old_free_size_in_bytes());
2147
2148 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
2149 size_policy->calculated_survivor_size_in_bytes());
2150 }
2151 if (PrintAdaptiveSizePolicy) {
2152 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2153 heap->total_collections());
2154 }
2155 }
2156
2157 if (UsePerfData) {
2158 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2159 counters->update_counters();
2160 counters->update_old_capacity(old_gen->capacity_in_bytes());
2161 counters->update_young_capacity(young_gen->capacity_in_bytes());
2162 }
2163
2164 heap->resize_all_tlabs();
2165
2166 // Resize the metaspace capacity after a collection
2167 MetaspaceGC::compute_new_size();
2168
2169 if (TraceOldGenTime) accumulated_time()->stop();
2170
2171 if (PrintGC) {
2172 if (PrintGCDetails) {
2173 // No GC timestamp here. This is after GC so it would be confusing.
2174 young_gen->print_used_change(pre_gc_values.young_gen_used());
2175 old_gen->print_used_change(pre_gc_values.old_gen_used());
2176 heap->print_heap_change(pre_gc_values.heap_used());
2177 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2178 } else {
2179 heap->print_heap_change(pre_gc_values.heap_used());
2180 }
2181 }
2182
2183 // Track memory usage and detect low memory
2184 MemoryService::track_memory_usage();
2185 heap->update_counters();
2186 gc_task_manager()->release_idle_workers();
2187 }
2188
2189 #ifdef ASSERT
2190 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2191 ParCompactionManager* const cm =
2192 ParCompactionManager::manager_array(int(i));
2193 assert(cm->marking_stack()->is_empty(), "should be empty");
2194 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2195 }
2196 #endif // ASSERT
2197
2198 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2199 HandleMark hm; // Discard invalid handles created during verification
2200 Universe::verify(" VerifyAfterGC:");
2201 }
2202
2203 // Re-verify object start arrays
2204 if (VerifyObjectStartArray &&
2205 VerifyAfterGC) {
2206 old_gen->verify_object_start_array();
2207 }
2208
2209 if (ZapUnusedHeapArea) {
2210 old_gen->object_space()->check_mangled_unused_area_complete();
2211 }
2212
2213 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2214
2215 collection_exit.update();
2216
2217 heap->print_heap_after_gc();
2218 heap->trace_heap_after_gc(&_gc_tracer);
2219
2220 if (PrintGCTaskTimeStamps) {
2221 gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " "
2222 JLONG_FORMAT,
2223 marking_start.ticks(), compaction_start.ticks(),
2224 collection_exit.ticks());
2225 gc_task_manager()->print_task_time_stamps();
2226 }
2227
2228 heap->post_full_gc_dump(&_gc_timer);
2229
2230 #ifdef TRACESPINNING
2231 ParallelTaskTerminator::print_termination_counts();
2232 #endif
2233
2234 _gc_timer.register_gc_end();
2235
2236 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2237 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2238
2239 return true;
2240 }
2241
2242 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2243 PSYoungGen* young_gen,
2244 PSOldGen* old_gen) {
2245 MutableSpace* const eden_space = young_gen->eden_space();
2246 assert(!eden_space->is_empty(), "eden must be non-empty");
2247 assert(young_gen->virtual_space()->alignment() ==
2248 old_gen->virtual_space()->alignment(), "alignments do not match");
2249
2250 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2251 return false;
2252 }
2253
2254 // Both generations must be completely committed.
2255 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2256 return false;
2257 }
2258 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2259 return false;
2260 }
2261
2262 // Figure out how much to take from eden. Include the average amount promoted
2263 // in the total; otherwise the next young gen GC will simply bail out to a
2264 // full GC.
2265 const size_t alignment = old_gen->virtual_space()->alignment();
2266 const size_t eden_used = eden_space->used_in_bytes();
2267 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2268 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2269 const size_t eden_capacity = eden_space->capacity_in_bytes();
2270
2271 if (absorb_size >= eden_capacity) {
2272 return false; // Must leave some space in eden.
2273 }
2274
2275 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2276 if (new_young_size < young_gen->min_gen_size()) {
2277 return false; // Respect young gen minimum size.
2278 }
2279
2280 if (TraceAdaptiveGCBoundary && Verbose) {
2281 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2282 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2283 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2284 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2285 absorb_size / K,
2286 eden_capacity / K, (eden_capacity - absorb_size) / K,
2287 young_gen->from_space()->used_in_bytes() / K,
2288 young_gen->to_space()->used_in_bytes() / K,
2289 young_gen->capacity_in_bytes() / K, new_young_size / K);
2290 }
2291
2292 // Fill the unused part of the old gen.
2293 MutableSpace* const old_space = old_gen->object_space();
2294 HeapWord* const unused_start = old_space->top();
2295 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2296
2297 if (unused_words > 0) {
2298 if (unused_words < CollectedHeap::min_fill_size()) {
2299 return false; // If the old gen cannot be filled, must give up.
2300 }
2301 CollectedHeap::fill_with_objects(unused_start, unused_words);
2302 }
2303
2304 // Take the live data from eden and set both top and end in the old gen to
2305 // eden top. (Need to set end because reset_after_change() mangles the region
2306 // from end to virtual_space->high() in debug builds).
2307 HeapWord* const new_top = eden_space->top();
2308 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2309 absorb_size);
2310 young_gen->reset_after_change();
2311 old_space->set_top(new_top);
2312 old_space->set_end(new_top);
2313 old_gen->reset_after_change();
2314
2315 // Update the object start array for the filler object and the data from eden.
2316 ObjectStartArray* const start_array = old_gen->start_array();
2317 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2318 start_array->allocate_block(p);
2319 }
2320
2321 // Could update the promoted average here, but it is not typically updated at
2322 // full GCs and the value to use is unclear. Something like
2323 //
2324 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2325
2326 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2327 return true;
2328 }
2329
2330 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2331 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2332 "shouldn't return NULL");
2333 return ParallelScavengeHeap::gc_task_manager();
2334 }
2335
2336 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2337 bool maximum_heap_compaction,
2338 ParallelOldTracer *gc_tracer) {
2339 // Recursively traverse all live objects and mark them
2340 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2341
2342 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2343 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2344 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2345 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2346 ParallelTaskTerminator terminator(active_gc_threads, qset);
2347
2348 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm);
2349 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2350
2351 // Need new claim bits before marking starts.
2352 ClassLoaderDataGraph::clear_claimed_marks();
2353
2354 {
2355 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2356
2357 ParallelScavengeHeap::ParStrongRootsScope psrs;
2358
2359 GCTaskQueue* q = GCTaskQueue::create();
2360
2361 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2362 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2363 // We scan the thread roots in parallel
2364 Threads::create_thread_roots_marking_tasks(q);
2365 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2368 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2369 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2371 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2372
2373 if (active_gc_threads > 1) {
2374 for (uint j = 0; j < active_gc_threads; j++) {
2375 q->enqueue(new StealMarkingTask(&terminator));
2376 }
2377 }
2378
2379 gc_task_manager()->execute_and_wait(q);
2380 }
2381
2382 // Process reference objects found during marking
2383 {
2384 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2385
2386 ReferenceProcessorStats stats;
2387 if (ref_processor()->processing_is_mt()) {
2388 RefProcTaskExecutor task_executor;
2389 stats = ref_processor()->process_discovered_references(
2390 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2391 &task_executor, &_gc_timer, _gc_tracer.gc_id());
2392 } else {
2393 stats = ref_processor()->process_discovered_references(
2394 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2395 &_gc_timer, _gc_tracer.gc_id());
2396 }
2397
2398 gc_tracer->report_gc_reference_stats(stats);
2399 }
2400
2401 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2402
2403 // This is the point where the entire marking should have completed.
2404 assert(cm->marking_stacks_empty(), "Marking should have completed");
2405
2406 // Follow system dictionary roots and unload classes.
2407 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2408
2409 // Unload nmethods.
2410 CodeCache::do_unloading(is_alive_closure(), purged_class);
2411
2412 // Prune dead klasses from subklass/sibling/implementor lists.
2413 Klass::clean_weak_klass_links(is_alive_closure());
2414
2415 // Delete entries for dead interned strings.
2416 StringTable::unlink(is_alive_closure());
2417
2418 // Clean up unreferenced symbols in symbol table.
2419 SymbolTable::unlink();
2420 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2421 }
2422
2423 // This should be moved to the shared markSweep code!
2424 class PSAlwaysTrueClosure: public BoolObjectClosure {
2425 public:
2426 bool do_object_b(oop p) { return true; }
2427 };
2428 static PSAlwaysTrueClosure always_true;
2429
2430 void PSParallelCompact::adjust_roots() {
2431 // Adjust the pointers to reflect the new locations
2432 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2433
2434 // Need new claim bits when tracing through and adjusting pointers.
2435 ClassLoaderDataGraph::clear_claimed_marks();
2436
2437 // General strong roots.
2438 Universe::oops_do(adjust_pointer_closure());
2439 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles
2440 CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2441 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2442 ObjectSynchronizer::oops_do(adjust_pointer_closure());
2443 FlatProfiler::oops_do(adjust_pointer_closure());
2444 Management::oops_do(adjust_pointer_closure());
2445 JvmtiExport::oops_do(adjust_pointer_closure());
2446 SystemDictionary::oops_do(adjust_pointer_closure());
2447 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2448
2449 // Now adjust pointers in remaining weak roots. (All of which should
2450 // have been cleared if they pointed to non-surviving objects.)
2451 // Global (weak) JNI handles
2452 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2453
2454 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations);
2455 CodeCache::blobs_do(&adjust_from_blobs);
2456 StringTable::oops_do(adjust_pointer_closure());
2457 ref_processor()->weak_oops_do(adjust_pointer_closure());
2458 // Roots were visited so references into the young gen in roots
2459 // may have been scanned. Process them also.
2460 // Should the reference processor have a span that excludes
2461 // young gen objects?
2462 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2463 }
2464
2465 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2466 uint parallel_gc_threads)
2467 {
2468 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2469
2470 // Find the threads that are active
2471 unsigned int which = 0;
2472
2473 const uint task_count = MAX2(parallel_gc_threads, 1U);
2474 for (uint j = 0; j < task_count; j++) {
2475 q->enqueue(new DrainStacksCompactionTask(j));
2476 ParCompactionManager::verify_region_list_empty(j);
2477 // Set the region stacks variables to "no" region stack values
2478 // so that they will be recognized and needing a region stack
2479 // in the stealing tasks if they do not get one by executing
2480 // a draining stack.
2481 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2482 cm->set_region_stack(NULL);
2483 cm->set_region_stack_index((uint)max_uintx);
2484 }
2485 ParCompactionManager::reset_recycled_stack_index();
2486
2487 // Find all regions that are available (can be filled immediately) and
2488 // distribute them to the thread stacks. The iteration is done in reverse
2489 // order (high to low) so the regions will be removed in ascending order.
2490
2491 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2492
2493 size_t fillable_regions = 0; // A count for diagnostic purposes.
2494 // A region index which corresponds to the tasks created above.
2495 // "which" must be 0 <= which < task_count
2496
2497 which = 0;
2498 // id + 1 is used to test termination so unsigned can
2499 // be used with an old_space_id == 0.
2500 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2501 SpaceInfo* const space_info = _space_info + id;
2502 MutableSpace* const space = space_info->space();
2503 HeapWord* const new_top = space_info->new_top();
2504
2505 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2506 const size_t end_region =
2507 sd.addr_to_region_idx(sd.region_align_up(new_top));
2508
2509 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2510 if (sd.region(cur)->claim_unsafe()) {
2511 ParCompactionManager::region_list_push(which, cur);
2512
2513 if (TraceParallelOldGCCompactionPhase && Verbose) {
2514 const size_t count_mod_8 = fillable_regions & 7;
2515 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2516 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2517 if (count_mod_8 == 7) gclog_or_tty->cr();
2518 }
2519
2520 NOT_PRODUCT(++fillable_regions;)
2521
2522 // Assign regions to tasks in round-robin fashion.
2523 if (++which == task_count) {
2524 assert(which <= parallel_gc_threads,
2525 "Inconsistent number of workers");
2526 which = 0;
2527 }
2528 }
2529 }
2530 }
2531
2532 if (TraceParallelOldGCCompactionPhase) {
2533 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2534 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions);
2535 }
2536 }
2537
2538 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2539
2540 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2541 uint parallel_gc_threads) {
2542 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2543
2544 ParallelCompactData& sd = PSParallelCompact::summary_data();
2545
2546 // Iterate over all the spaces adding tasks for updating
2547 // regions in the dense prefix. Assume that 1 gc thread
2548 // will work on opening the gaps and the remaining gc threads
2549 // will work on the dense prefix.
2550 unsigned int space_id;
2551 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2552 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2553 const MutableSpace* const space = _space_info[space_id].space();
2554
2555 if (dense_prefix_end == space->bottom()) {
2556 // There is no dense prefix for this space.
2557 continue;
2558 }
2559
2560 // The dense prefix is before this region.
2561 size_t region_index_end_dense_prefix =
2562 sd.addr_to_region_idx(dense_prefix_end);
2563 RegionData* const dense_prefix_cp =
2564 sd.region(region_index_end_dense_prefix);
2565 assert(dense_prefix_end == space->end() ||
2566 dense_prefix_cp->available() ||
2567 dense_prefix_cp->claimed(),
2568 "The region after the dense prefix should always be ready to fill");
2569
2570 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2571
2572 // Is there dense prefix work?
2573 size_t total_dense_prefix_regions =
2574 region_index_end_dense_prefix - region_index_start;
2575 // How many regions of the dense prefix should be given to
2576 // each thread?
2577 if (total_dense_prefix_regions > 0) {
2578 uint tasks_for_dense_prefix = 1;
2579 if (total_dense_prefix_regions <=
2580 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2581 // Don't over partition. This assumes that
2582 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2583 // so there are not many regions to process.
2584 tasks_for_dense_prefix = parallel_gc_threads;
2585 } else {
2586 // Over partition
2587 tasks_for_dense_prefix = parallel_gc_threads *
2588 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2589 }
2590 size_t regions_per_thread = total_dense_prefix_regions /
2591 tasks_for_dense_prefix;
2592 // Give each thread at least 1 region.
2593 if (regions_per_thread == 0) {
2594 regions_per_thread = 1;
2595 }
2596
2597 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2598 if (region_index_start >= region_index_end_dense_prefix) {
2599 break;
2600 }
2601 // region_index_end is not processed
2602 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2603 region_index_end_dense_prefix);
2604 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2605 region_index_start,
2606 region_index_end));
2607 region_index_start = region_index_end;
2608 }
2609 }
2610 // This gets any part of the dense prefix that did not
2611 // fit evenly.
2612 if (region_index_start < region_index_end_dense_prefix) {
2613 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2614 region_index_start,
2615 region_index_end_dense_prefix));
2616 }
2617 }
2618 }
2619
2620 void PSParallelCompact::enqueue_region_stealing_tasks(
2621 GCTaskQueue* q,
2622 ParallelTaskTerminator* terminator_ptr,
2623 uint parallel_gc_threads) {
2624 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2625
2626 // Once a thread has drained it's stack, it should try to steal regions from
2627 // other threads.
2628 if (parallel_gc_threads > 1) {
2629 for (uint j = 0; j < parallel_gc_threads; j++) {
2630 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2631 }
2632 }
2633 }
2634
2635 #ifdef ASSERT
2636 // Write a histogram of the number of times the block table was filled for a
2637 // region.
2638 void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
2639 {
2640 if (!TraceParallelOldGCCompactionPhase) return;
2641
2642 typedef ParallelCompactData::RegionData rd_t;
2643 ParallelCompactData& sd = summary_data();
2644
2645 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2646 MutableSpace* const spc = _space_info[id].space();
2647 if (spc->bottom() != spc->top()) {
2648 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2649 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2650 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2651
2652 size_t histo[5] = { 0, 0, 0, 0, 0 };
2653 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2654 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2655
2656 for (const rd_t* cur = beg; cur < end; ++cur) {
2657 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2658 }
2659 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2660 for (size_t i = 0; i < histo_len; ++i) {
2661 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2662 histo[i], 100.0 * histo[i] / region_cnt);
2663 }
2664 out->cr();
2665 }
2666 }
2667 }
2668 #endif // #ifdef ASSERT
2669
2670 void PSParallelCompact::compact() {
2671 // trace("5");
2672 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2673
2674 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2675 PSOldGen* old_gen = heap->old_gen();
2676 old_gen->start_array()->reset();
2677 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2678 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2679 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2680 ParallelTaskTerminator terminator(active_gc_threads, qset);
2681
2682 GCTaskQueue* q = GCTaskQueue::create();
2683 enqueue_region_draining_tasks(q, active_gc_threads);
2684 enqueue_dense_prefix_tasks(q, active_gc_threads);
2685 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2686
2687 {
2688 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2689
2690 gc_task_manager()->execute_and_wait(q);
2691
2692 #ifdef ASSERT
2693 // Verify that all regions have been processed before the deferred updates.
2694 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2695 verify_complete(SpaceId(id));
2696 }
2697 #endif
2698 }
2699
2700 {
2701 // Update the deferred objects, if any. Any compaction manager can be used.
2702 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2703 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2704 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2705 update_deferred_objects(cm, SpaceId(id));
2706 }
2707 }
2708
2709 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
2710 }
2711
2712 #ifdef ASSERT
2713 void PSParallelCompact::verify_complete(SpaceId space_id) {
2714 // All Regions between space bottom() to new_top() should be marked as filled
2715 // and all Regions between new_top() and top() should be available (i.e.,
2716 // should have been emptied).
2717 ParallelCompactData& sd = summary_data();
2718 SpaceInfo si = _space_info[space_id];
2719 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2720 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2721 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2722 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2723 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2724
2725 bool issued_a_warning = false;
2726
2727 size_t cur_region;
2728 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2729 const RegionData* const c = sd.region(cur_region);
2730 if (!c->completed()) {
2731 warning("region " SIZE_FORMAT " not filled: "
2732 "destination_count=%u",
2733 cur_region, c->destination_count());
2734 issued_a_warning = true;
2735 }
2736 }
2737
2738 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2739 const RegionData* const c = sd.region(cur_region);
2740 if (!c->available()) {
2741 warning("region " SIZE_FORMAT " not empty: "
2742 "destination_count=%u",
2743 cur_region, c->destination_count());
2744 issued_a_warning = true;
2745 }
2746 }
2747
2748 if (issued_a_warning) {
2749 print_region_ranges();
2750 }
2751 }
2752 #endif // #ifdef ASSERT
2753
2754 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2755 _start_array->allocate_block(addr);
2756 compaction_manager()->update_contents(oop(addr));
2757 }
2758
2759 // Update interior oops in the ranges of regions [beg_region, end_region).
2760 void
2761 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2762 SpaceId space_id,
2763 size_t beg_region,
2764 size_t end_region) {
2765 ParallelCompactData& sd = summary_data();
2766 ParMarkBitMap* const mbm = mark_bitmap();
2767
2768 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2769 HeapWord* const end_addr = sd.region_to_addr(end_region);
2770 assert(beg_region <= end_region, "bad region range");
2771 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2772
2773 #ifdef ASSERT
2774 // Claim the regions to avoid triggering an assert when they are marked as
2775 // filled.
2776 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2777 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2778 }
2779 #endif // #ifdef ASSERT
2780
2781 if (beg_addr != space(space_id)->bottom()) {
2782 // Find the first live object or block of dead space that *starts* in this
2783 // range of regions. If a partial object crosses onto the region, skip it;
2784 // it will be marked for 'deferred update' when the object head is
2785 // processed. If dead space crosses onto the region, it is also skipped; it
2786 // will be filled when the prior region is processed. If neither of those
2787 // apply, the first word in the region is the start of a live object or dead
2788 // space.
2789 assert(beg_addr > space(space_id)->bottom(), "sanity");
2790 const RegionData* const cp = sd.region(beg_region);
2791 if (cp->partial_obj_size() != 0) {
2792 beg_addr = sd.partial_obj_end(beg_region);
2793 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2794 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2795 }
2796 }
2797
2798 if (beg_addr < end_addr) {
2799 // A live object or block of dead space starts in this range of Regions.
2800 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2801
2802 // Create closures and iterate.
2803 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2804 FillClosure fill_closure(cm, space_id);
2805 ParMarkBitMap::IterationStatus status;
2806 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2807 dense_prefix_end);
2808 if (status == ParMarkBitMap::incomplete) {
2809 update_closure.do_addr(update_closure.source());
2810 }
2811 }
2812
2813 // Mark the regions as filled.
2814 RegionData* const beg_cp = sd.region(beg_region);
2815 RegionData* const end_cp = sd.region(end_region);
2816 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2817 cp->set_completed();
2818 }
2819 }
2820
2821 // Return the SpaceId for the space containing addr. If addr is not in the
2822 // heap, last_space_id is returned. In debug mode it expects the address to be
2823 // in the heap and asserts such.
2824 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2825 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2826
2827 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2828 if (_space_info[id].space()->contains(addr)) {
2829 return SpaceId(id);
2830 }
2831 }
2832
2833 assert(false, "no space contains the addr");
2834 return last_space_id;
2835 }
2836
2837 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2838 SpaceId id) {
2839 assert(id < last_space_id, "bad space id");
2840
2841 ParallelCompactData& sd = summary_data();
2842 const SpaceInfo* const space_info = _space_info + id;
2843 ObjectStartArray* const start_array = space_info->start_array();
2844
2845 const MutableSpace* const space = space_info->space();
2846 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2847 HeapWord* const beg_addr = space_info->dense_prefix();
2848 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2849
2850 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2851 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2852 const RegionData* cur_region;
2853 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2854 HeapWord* const addr = cur_region->deferred_obj_addr();
2855 if (addr != NULL) {
2856 if (start_array != NULL) {
2857 start_array->allocate_block(addr);
2858 }
2859 cm->update_contents(oop(addr));
2860 assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))));
2861 }
2862 }
2863 }
2864
2865 // Skip over count live words starting from beg, and return the address of the
2866 // next live word. Unless marked, the word corresponding to beg is assumed to
2867 // be dead. Callers must either ensure beg does not correspond to the middle of
2868 // an object, or account for those live words in some other way. Callers must
2869 // also ensure that there are enough live words in the range [beg, end) to skip.
2870 HeapWord*
2871 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2872 {
2873 assert(count > 0, "sanity");
2874
2875 ParMarkBitMap* m = mark_bitmap();
2876 idx_t bits_to_skip = m->words_to_bits(count);
2877 idx_t cur_beg = m->addr_to_bit(beg);
2878 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2879
2880 do {
2881 cur_beg = m->find_obj_beg(cur_beg, search_end);
2882 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2883 const size_t obj_bits = cur_end - cur_beg + 1;
2884 if (obj_bits > bits_to_skip) {
2885 return m->bit_to_addr(cur_beg + bits_to_skip);
2886 }
2887 bits_to_skip -= obj_bits;
2888 cur_beg = cur_end + 1;
2889 } while (bits_to_skip > 0);
2890
2891 // Skipping the desired number of words landed just past the end of an object.
2892 // Find the start of the next object.
2893 cur_beg = m->find_obj_beg(cur_beg, search_end);
2894 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2895 return m->bit_to_addr(cur_beg);
2896 }
2897
2898 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2899 SpaceId src_space_id,
2900 size_t src_region_idx)
2901 {
2902 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2903
2904 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2905 if (split_info.dest_region_addr() == dest_addr) {
2906 // The partial object ending at the split point contains the first word to
2907 // be copied to dest_addr.
2908 return split_info.first_src_addr();
2909 }
2910
2911 const ParallelCompactData& sd = summary_data();
2912 ParMarkBitMap* const bitmap = mark_bitmap();
2913 const size_t RegionSize = ParallelCompactData::RegionSize;
2914
2915 assert(sd.is_region_aligned(dest_addr), "not aligned");
2916 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2917 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2918 HeapWord* const src_region_destination = src_region_ptr->destination();
2919
2920 assert(dest_addr >= src_region_destination, "wrong src region");
2921 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2922
2923 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2924 HeapWord* const src_region_end = src_region_beg + RegionSize;
2925
2926 HeapWord* addr = src_region_beg;
2927 if (dest_addr == src_region_destination) {
2928 // Return the first live word in the source region.
2929 if (partial_obj_size == 0) {
2930 addr = bitmap->find_obj_beg(addr, src_region_end);
2931 assert(addr < src_region_end, "no objects start in src region");
2932 }
2933 return addr;
2934 }
2935
2936 // Must skip some live data.
2937 size_t words_to_skip = dest_addr - src_region_destination;
2938 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2939
2940 if (partial_obj_size >= words_to_skip) {
2941 // All the live words to skip are part of the partial object.
2942 addr += words_to_skip;
2943 if (partial_obj_size == words_to_skip) {
2944 // Find the first live word past the partial object.
2945 addr = bitmap->find_obj_beg(addr, src_region_end);
2946 assert(addr < src_region_end, "wrong src region");
2947 }
2948 return addr;
2949 }
2950
2951 // Skip over the partial object (if any).
2952 if (partial_obj_size != 0) {
2953 words_to_skip -= partial_obj_size;
2954 addr += partial_obj_size;
2955 }
2956
2957 // Skip over live words due to objects that start in the region.
2958 addr = skip_live_words(addr, src_region_end, words_to_skip);
2959 assert(addr < src_region_end, "wrong src region");
2960 return addr;
2961 }
2962
2963 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2964 SpaceId src_space_id,
2965 size_t beg_region,
2966 HeapWord* end_addr)
2967 {
2968 ParallelCompactData& sd = summary_data();
2969
2970 #ifdef ASSERT
2971 MutableSpace* const src_space = _space_info[src_space_id].space();
2972 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2973 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2974 "src_space_id does not match beg_addr");
2975 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2976 "src_space_id does not match end_addr");
2977 #endif // #ifdef ASSERT
2978
2979 RegionData* const beg = sd.region(beg_region);
2980 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2981
2982 // Regions up to new_top() are enqueued if they become available.
2983 HeapWord* const new_top = _space_info[src_space_id].new_top();
2984 RegionData* const enqueue_end =
2985 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2986
2987 for (RegionData* cur = beg; cur < end; ++cur) {
2988 assert(cur->data_size() > 0, "region must have live data");
2989 cur->decrement_destination_count();
2990 if (cur < enqueue_end && cur->available() && cur->claim()) {
2991 cm->push_region(sd.region(cur));
2992 }
2993 }
2994 }
2995
2996 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2997 SpaceId& src_space_id,
2998 HeapWord*& src_space_top,
2999 HeapWord* end_addr)
3000 {
3001 typedef ParallelCompactData::RegionData RegionData;
3002
3003 ParallelCompactData& sd = PSParallelCompact::summary_data();
3004 const size_t region_size = ParallelCompactData::RegionSize;
3005
3006 size_t src_region_idx = 0;
3007
3008 // Skip empty regions (if any) up to the top of the space.
3009 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3010 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3011 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3012 const RegionData* const top_region_ptr =
3013 sd.addr_to_region_ptr(top_aligned_up);
3014 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3015 ++src_region_ptr;
3016 }
3017
3018 if (src_region_ptr < top_region_ptr) {
3019 // The next source region is in the current space. Update src_region_idx
3020 // and the source address to match src_region_ptr.
3021 src_region_idx = sd.region(src_region_ptr);
3022 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3023 if (src_region_addr > closure.source()) {
3024 closure.set_source(src_region_addr);
3025 }
3026 return src_region_idx;
3027 }
3028
3029 // Switch to a new source space and find the first non-empty region.
3030 unsigned int space_id = src_space_id + 1;
3031 assert(space_id < last_space_id, "not enough spaces");
3032
3033 HeapWord* const destination = closure.destination();
3034
3035 do {
3036 MutableSpace* space = _space_info[space_id].space();
3037 HeapWord* const bottom = space->bottom();
3038 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3039
3040 // Iterate over the spaces that do not compact into themselves.
3041 if (bottom_cp->destination() != bottom) {
3042 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3043 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3044
3045 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3046 if (src_cp->live_obj_size() > 0) {
3047 // Found it.
3048 assert(src_cp->destination() == destination,
3049 "first live obj in the space must match the destination");
3050 assert(src_cp->partial_obj_size() == 0,
3051 "a space cannot begin with a partial obj");
3052
3053 src_space_id = SpaceId(space_id);
3054 src_space_top = space->top();
3055 const size_t src_region_idx = sd.region(src_cp);
3056 closure.set_source(sd.region_to_addr(src_region_idx));
3057 return src_region_idx;
3058 } else {
3059 assert(src_cp->data_size() == 0, "sanity");
3060 }
3061 }
3062 }
3063 } while (++space_id < last_space_id);
3064
3065 assert(false, "no source region was found");
3066 return 0;
3067 }
3068
3069 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3070 {
3071 typedef ParMarkBitMap::IterationStatus IterationStatus;
3072 const size_t RegionSize = ParallelCompactData::RegionSize;
3073 ParMarkBitMap* const bitmap = mark_bitmap();
3074 ParallelCompactData& sd = summary_data();
3075 RegionData* const region_ptr = sd.region(region_idx);
3076
3077 // Get the items needed to construct the closure.
3078 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3079 SpaceId dest_space_id = space_id(dest_addr);
3080 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3081 HeapWord* new_top = _space_info[dest_space_id].new_top();
3082 assert(dest_addr < new_top, "sanity");
3083 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3084
3085 // Get the source region and related info.
3086 size_t src_region_idx = region_ptr->source_region();
3087 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3088 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3089
3090 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3091 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3092
3093 // Adjust src_region_idx to prepare for decrementing destination counts (the
3094 // destination count is not decremented when a region is copied to itself).
3095 if (src_region_idx == region_idx) {
3096 src_region_idx += 1;
3097 }
3098
3099 if (bitmap->is_unmarked(closure.source())) {
3100 // The first source word is in the middle of an object; copy the remainder
3101 // of the object or as much as will fit. The fact that pointer updates were
3102 // deferred will be noted when the object header is processed.
3103 HeapWord* const old_src_addr = closure.source();
3104 closure.copy_partial_obj();
3105 if (closure.is_full()) {
3106 decrement_destination_counts(cm, src_space_id, src_region_idx,
3107 closure.source());
3108 region_ptr->set_deferred_obj_addr(NULL);
3109 region_ptr->set_completed();
3110 return;
3111 }
3112
3113 HeapWord* const end_addr = sd.region_align_down(closure.source());
3114 if (sd.region_align_down(old_src_addr) != end_addr) {
3115 // The partial object was copied from more than one source region.
3116 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3117
3118 // Move to the next source region, possibly switching spaces as well. All
3119 // args except end_addr may be modified.
3120 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3121 end_addr);
3122 }
3123 }
3124
3125 do {
3126 HeapWord* const cur_addr = closure.source();
3127 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3128 src_space_top);
3129 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3130
3131 if (status == ParMarkBitMap::incomplete) {
3132 // The last obj that starts in the source region does not end in the
3133 // region.
3134 assert(closure.source() < end_addr, "sanity");
3135 HeapWord* const obj_beg = closure.source();
3136 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3137 src_space_top);
3138 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3139 if (obj_end < range_end) {
3140 // The end was found; the entire object will fit.
3141 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3142 assert(status != ParMarkBitMap::would_overflow, "sanity");
3143 } else {
3144 // The end was not found; the object will not fit.
3145 assert(range_end < src_space_top, "obj cannot cross space boundary");
3146 status = ParMarkBitMap::would_overflow;
3147 }
3148 }
3149
3150 if (status == ParMarkBitMap::would_overflow) {
3151 // The last object did not fit. Note that interior oop updates were
3152 // deferred, then copy enough of the object to fill the region.
3153 region_ptr->set_deferred_obj_addr(closure.destination());
3154 status = closure.copy_until_full(); // copies from closure.source()
3155
3156 decrement_destination_counts(cm, src_space_id, src_region_idx,
3157 closure.source());
3158 region_ptr->set_completed();
3159 return;
3160 }
3161
3162 if (status == ParMarkBitMap::full) {
3163 decrement_destination_counts(cm, src_space_id, src_region_idx,
3164 closure.source());
3165 region_ptr->set_deferred_obj_addr(NULL);
3166 region_ptr->set_completed();
3167 return;
3168 }
3169
3170 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3171
3172 // Move to the next source region, possibly switching spaces as well. All
3173 // args except end_addr may be modified.
3174 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3175 end_addr);
3176 } while (true);
3177 }
3178
3179 void PSParallelCompact::fill_blocks(size_t region_idx)
3180 {
3181 // Fill in the block table elements for the specified region. Each block
3182 // table element holds the number of live words in the region that are to the
3183 // left of the first object that starts in the block. Thus only blocks in
3184 // which an object starts need to be filled.
3185 //
3186 // The algorithm scans the section of the bitmap that corresponds to the
3187 // region, keeping a running total of the live words. When an object start is
3188 // found, if it's the first to start in the block that contains it, the
3189 // current total is written to the block table element.
3190 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3191 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3192 const size_t RegionSize = ParallelCompactData::RegionSize;
3193
3194 ParallelCompactData& sd = summary_data();
3195 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3196 if (partial_obj_size >= RegionSize) {
3197 return; // No objects start in this region.
3198 }
3199
3200 // Ensure the first loop iteration decides that the block has changed.
3201 size_t cur_block = sd.block_count();
3202
3203 const ParMarkBitMap* const bitmap = mark_bitmap();
3204
3205 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3206 assert((size_t)1 << Log2BitsPerBlock ==
3207 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3208
3209 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3210 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3211 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3212 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3213 while (beg_bit < range_end) {
3214 const size_t new_block = beg_bit >> Log2BitsPerBlock;
3215 if (new_block != cur_block) {
3216 cur_block = new_block;
3217 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3218 }
3219
3220 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3221 if (end_bit < range_end - 1) {
3222 live_bits += end_bit - beg_bit + 1;
3223 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3224 } else {
3225 return;
3226 }
3227 }
3228 }
3229
3230 void
3231 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3232 const MutableSpace* sp = space(space_id);
3233 if (sp->is_empty()) {
3234 return;
3235 }
3236
3237 ParallelCompactData& sd = PSParallelCompact::summary_data();
3238 ParMarkBitMap* const bitmap = mark_bitmap();
3239 HeapWord* const dp_addr = dense_prefix(space_id);
3240 HeapWord* beg_addr = sp->bottom();
3241 HeapWord* end_addr = sp->top();
3242
3243 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3244
3245 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3246 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3247 if (beg_region < dp_region) {
3248 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3249 }
3250
3251 // The destination of the first live object that starts in the region is one
3252 // past the end of the partial object entering the region (if any).
3253 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3254 HeapWord* const new_top = _space_info[space_id].new_top();
3255 assert(new_top >= dest_addr, "bad new_top value");
3256 const size_t words = pointer_delta(new_top, dest_addr);
3257
3258 if (words > 0) {
3259 ObjectStartArray* start_array = _space_info[space_id].start_array();
3260 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3261
3262 ParMarkBitMap::IterationStatus status;
3263 status = bitmap->iterate(&closure, dest_addr, end_addr);
3264 assert(status == ParMarkBitMap::full, "iteration not complete");
3265 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3266 "live objects skipped because closure is full");
3267 }
3268 }
3269
3270 jlong PSParallelCompact::millis_since_last_gc() {
3271 // We need a monotonically non-decreasing time in ms but
3272 // os::javaTimeMillis() does not guarantee monotonicity.
3273 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3274 jlong ret_val = now - _time_of_last_gc;
3275 // XXX See note in genCollectedHeap::millis_since_last_gc().
3276 if (ret_val < 0) {
3277 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);)
3278 return 0;
3279 }
3280 return ret_val;
3281 }
3282
3283 void PSParallelCompact::reset_millis_since_last_gc() {
3284 // We need a monotonically non-decreasing time in ms but
3285 // os::javaTimeMillis() does not guarantee monotonicity.
3286 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3287 }
3288
3289 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3290 {
3291 if (source() != destination()) {
3292 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3293 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3294 }
3295 update_state(words_remaining());
3296 assert(is_full(), "sanity");
3297 return ParMarkBitMap::full;
3298 }
3299
3300 void MoveAndUpdateClosure::copy_partial_obj()
3301 {
3302 size_t words = words_remaining();
3303
3304 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3305 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3306 if (end_addr < range_end) {
3307 words = bitmap()->obj_size(source(), end_addr);
3308 }
3309
3310 // This test is necessary; if omitted, the pointer updates to a partial object
3311 // that crosses the dense prefix boundary could be overwritten.
3312 if (source() != destination()) {
3313 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3314 Copy::aligned_conjoint_words(source(), destination(), words);
3315 }
3316 update_state(words);
3317 }
3318
3319 void InstanceKlass::oop_pc_update_pointers(oop obj) {
3320 oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure());
3321 }
3322
3323 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) {
3324 InstanceKlass::oop_pc_update_pointers(obj);
3325
3326 oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure());
3327 }
3328
3329 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) {
3330 InstanceKlass::oop_pc_update_pointers(obj);
3331 }
3332
3333 #ifdef ASSERT
3334 template <class T> static void trace_reference_gc(const char *s, oop obj,
3335 T* referent_addr,
3336 T* next_addr,
3337 T* discovered_addr) {
3338 if(TraceReferenceGC && PrintGCDetails) {
3339 gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj));
3340 gclog_or_tty->print_cr(" referent_addr/* " PTR_FORMAT " / "
3341 PTR_FORMAT, p2i(referent_addr),
3342 referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL);
3343 gclog_or_tty->print_cr(" next_addr/* " PTR_FORMAT " / "
3344 PTR_FORMAT, p2i(next_addr),
3345 next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL);
3346 gclog_or_tty->print_cr(" discovered_addr/* " PTR_FORMAT " / "
3347 PTR_FORMAT, p2i(discovered_addr),
3348 discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL);
3349 }
3350 }
3351 #endif
3352
3353 template <class T>
3354 static void oop_pc_update_pointers_specialized(oop obj) {
3355 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj);
3356 PSParallelCompact::adjust_pointer(referent_addr);
3357 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj);
3358 PSParallelCompact::adjust_pointer(next_addr);
3359 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj);
3360 PSParallelCompact::adjust_pointer(discovered_addr);
3361 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj,
3362 referent_addr, next_addr, discovered_addr);)
3363 }
3364
3365 void InstanceRefKlass::oop_pc_update_pointers(oop obj) {
3366 InstanceKlass::oop_pc_update_pointers(obj);
3367
3368 if (UseCompressedOops) {
3369 oop_pc_update_pointers_specialized<narrowOop>(obj);
3370 } else {
3371 oop_pc_update_pointers_specialized<oop>(obj);
3372 }
3373 }
3374
3375 void ObjArrayKlass::oop_pc_update_pointers(oop obj) {
3376 assert(obj->is_objArray(), "obj must be obj array");
3377 oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure());
3378 }
3379
3380 void TypeArrayKlass::oop_pc_update_pointers(oop obj) {
3381 assert(obj->is_typeArray(),"must be a type array");
3382 }
3383
3384 ParMarkBitMapClosure::IterationStatus
3385 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3386 assert(destination() != NULL, "sanity");
3387 assert(bitmap()->obj_size(addr) == words, "bad size");
3388
3389 _source = addr;
3390 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3391 destination(), "wrong destination");
3392
3393 if (words > words_remaining()) {
3394 return ParMarkBitMap::would_overflow;
3395 }
3396
3397 // The start_array must be updated even if the object is not moving.
3398 if (_start_array != NULL) {
3399 _start_array->allocate_block(destination());
3400 }
3401
3402 if (destination() != source()) {
3403 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3404 Copy::aligned_conjoint_words(source(), destination(), words);
3405 }
3406
3407 oop moved_oop = (oop) destination();
3408 compaction_manager()->update_contents(moved_oop);
3409 assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)));
3410
3411 update_state(words);
3412 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3413 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3414 }
3415
3416 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3417 ParCompactionManager* cm,
3418 PSParallelCompact::SpaceId space_id) :
3419 ParMarkBitMapClosure(mbm, cm),
3420 _space_id(space_id),
3421 _start_array(PSParallelCompact::start_array(space_id))
3422 {
3423 }
3424
3425 // Updates the references in the object to their new values.
3426 ParMarkBitMapClosure::IterationStatus
3427 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3428 do_addr(addr);
3429 return ParMarkBitMap::incomplete;
3430 }